xref: /netbsd-src/sys/uvm/uvm_page.c (revision 27578b9aac214cc7796ead81dcc5427e79d5f2a0)
1 /*	$NetBSD: uvm_page.c,v 1.67 2001/09/15 20:36:46 chs 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 "opt_uvmhist.h"
74 
75 #include <sys/param.h>
76 #include <sys/systm.h>
77 #include <sys/malloc.h>
78 #include <sys/sched.h>
79 #include <sys/kernel.h>
80 #include <sys/vnode.h>
81 
82 #define UVM_PAGE                /* pull in uvm_page.h functions */
83 #include <uvm/uvm.h>
84 
85 /*
86  * global vars... XXXCDC: move to uvm. structure.
87  */
88 
89 /*
90  * physical memory config is stored in vm_physmem.
91  */
92 
93 struct vm_physseg vm_physmem[VM_PHYSSEG_MAX];	/* XXXCDC: uvm.physmem */
94 int vm_nphysseg = 0;				/* XXXCDC: uvm.nphysseg */
95 
96 /*
97  * Some supported CPUs in a given architecture don't support all
98  * of the things necessary to do idle page zero'ing efficiently.
99  * We therefore provide a way to disable it from machdep code here.
100  */
101 /*
102  * XXX disabled until we can find a way to do this without causing
103  * problems for either cpu caches or DMA latency.
104  */
105 boolean_t vm_page_zero_enable = FALSE;
106 
107 /*
108  * local variables
109  */
110 
111 /*
112  * these variables record the values returned by vm_page_bootstrap,
113  * for debugging purposes.  The implementation of uvm_pageboot_alloc
114  * and pmap_startup here also uses them internally.
115  */
116 
117 static vaddr_t      virtual_space_start;
118 static vaddr_t      virtual_space_end;
119 
120 /*
121  * we use a hash table with only one bucket during bootup.  we will
122  * later rehash (resize) the hash table once the allocator is ready.
123  * we static allocate the one bootstrap bucket below...
124  */
125 
126 static struct pglist uvm_bootbucket;
127 
128 /*
129  * we allocate an initial number of page colors in uvm_page_init(),
130  * and remember them.  We may re-color pages as cache sizes are
131  * discovered during the autoconfiguration phase.  But we can never
132  * free the initial set of buckets, since they are allocated using
133  * uvm_pageboot_alloc().
134  */
135 
136 static boolean_t have_recolored_pages /* = FALSE */;
137 
138 /*
139  * local prototypes
140  */
141 
142 static void uvm_pageinsert __P((struct vm_page *));
143 static void uvm_pageremove __P((struct vm_page *));
144 
145 /*
146  * inline functions
147  */
148 
149 /*
150  * uvm_pageinsert: insert a page in the object and the hash table
151  *
152  * => caller must lock object
153  * => caller must lock page queues
154  * => call should have already set pg's object and offset pointers
155  *    and bumped the version counter
156  */
157 
158 __inline static void
159 uvm_pageinsert(pg)
160 	struct vm_page *pg;
161 {
162 	struct pglist *buck;
163 	struct uvm_object *uobj = pg->uobject;
164 
165 	KASSERT((pg->flags & PG_TABLED) == 0);
166 	buck = &uvm.page_hash[uvm_pagehash(uobj, pg->offset)];
167 	simple_lock(&uvm.hashlock);
168 	TAILQ_INSERT_TAIL(buck, pg, hashq);
169 	simple_unlock(&uvm.hashlock);
170 
171 	TAILQ_INSERT_TAIL(&uobj->memq, pg, listq);
172 	pg->flags |= PG_TABLED;
173 	uobj->uo_npages++;
174 }
175 
176 /*
177  * uvm_page_remove: remove page from object and hash
178  *
179  * => caller must lock object
180  * => caller must lock page queues
181  */
182 
183 static __inline void
184 uvm_pageremove(pg)
185 	struct vm_page *pg;
186 {
187 	struct pglist *buck;
188 	struct uvm_object *uobj = pg->uobject;
189 
190 	KASSERT(pg->flags & PG_TABLED);
191 	buck = &uvm.page_hash[uvm_pagehash(uobj ,pg->offset)];
192 	simple_lock(&uvm.hashlock);
193 	TAILQ_REMOVE(buck, pg, hashq);
194 	simple_unlock(&uvm.hashlock);
195 
196 	if (UVM_OBJ_IS_VTEXT(uobj)) {
197 		uvmexp.vtextpages--;
198 	} else if (UVM_OBJ_IS_VNODE(uobj)) {
199 		uvmexp.vnodepages--;
200 	}
201 
202 	/* object should be locked */
203 	uobj->uo_npages--;
204 	TAILQ_REMOVE(&uobj->memq, pg, listq);
205 	pg->flags &= ~PG_TABLED;
206 	pg->uobject = NULL;
207 }
208 
209 static void
210 uvm_page_init_buckets(struct pgfreelist *pgfl)
211 {
212 	int color, i;
213 
214 	for (color = 0; color < uvmexp.ncolors; color++) {
215 		for (i = 0; i < PGFL_NQUEUES; i++) {
216 			TAILQ_INIT(&pgfl->pgfl_buckets[
217 			    color].pgfl_queues[i]);
218 		}
219 	}
220 }
221 
222 /*
223  * uvm_page_init: init the page system.   called from uvm_init().
224  *
225  * => we return the range of kernel virtual memory in kvm_startp/kvm_endp
226  */
227 
228 void
229 uvm_page_init(kvm_startp, kvm_endp)
230 	vaddr_t *kvm_startp, *kvm_endp;
231 {
232 	vsize_t freepages, pagecount, bucketcount, n;
233 	struct pgflbucket *bucketarray;
234 	struct vm_page *pagearray;
235 	int lcv, i;
236 	paddr_t paddr;
237 
238 	/*
239 	 * init the page queues and page queue locks, except the free
240 	 * list; we allocate that later (with the initial vm_page
241 	 * structures).
242 	 */
243 
244 	TAILQ_INIT(&uvm.page_active);
245 	TAILQ_INIT(&uvm.page_inactive);
246 	simple_lock_init(&uvm.pageqlock);
247 	simple_lock_init(&uvm.fpageqlock);
248 
249 	/*
250 	 * init the <obj,offset> => <page> hash table.  for now
251 	 * we just have one bucket (the bootstrap bucket).  later on we
252 	 * will allocate new buckets as we dynamically resize the hash table.
253 	 */
254 
255 	uvm.page_nhash = 1;			/* 1 bucket */
256 	uvm.page_hashmask = 0;			/* mask for hash function */
257 	uvm.page_hash = &uvm_bootbucket;	/* install bootstrap bucket */
258 	TAILQ_INIT(uvm.page_hash);		/* init hash table */
259 	simple_lock_init(&uvm.hashlock);	/* init hash table lock */
260 
261 	/*
262 	 * allocate vm_page structures.
263 	 */
264 
265 	/*
266 	 * sanity check:
267 	 * before calling this function the MD code is expected to register
268 	 * some free RAM with the uvm_page_physload() function.   our job
269 	 * now is to allocate vm_page structures for this memory.
270 	 */
271 
272 	if (vm_nphysseg == 0)
273 		panic("uvm_page_bootstrap: no memory pre-allocated");
274 
275 	/*
276 	 * first calculate the number of free pages...
277 	 *
278 	 * note that we use start/end rather than avail_start/avail_end.
279 	 * this allows us to allocate extra vm_page structures in case we
280 	 * want to return some memory to the pool after booting.
281 	 */
282 
283 	freepages = 0;
284 	for (lcv = 0 ; lcv < vm_nphysseg ; lcv++)
285 		freepages += (vm_physmem[lcv].end - vm_physmem[lcv].start);
286 
287 	/*
288 	 * Let MD code initialize the number of colors, or default
289 	 * to 1 color if MD code doesn't care.
290 	 */
291 	if (uvmexp.ncolors == 0)
292 		uvmexp.ncolors = 1;
293 	uvmexp.colormask = uvmexp.ncolors - 1;
294 
295 	/*
296 	 * we now know we have (PAGE_SIZE * freepages) bytes of memory we can
297 	 * use.   for each page of memory we use we need a vm_page structure.
298 	 * thus, the total number of pages we can use is the total size of
299 	 * the memory divided by the PAGE_SIZE plus the size of the vm_page
300 	 * structure.   we add one to freepages as a fudge factor to avoid
301 	 * truncation errors (since we can only allocate in terms of whole
302 	 * pages).
303 	 */
304 
305 	bucketcount = uvmexp.ncolors * VM_NFREELIST;
306 	pagecount = ((freepages + 1) << PAGE_SHIFT) /
307 	    (PAGE_SIZE + sizeof(struct vm_page));
308 
309 	bucketarray = (void *)uvm_pageboot_alloc((bucketcount *
310 	    sizeof(struct pgflbucket)) + (pagecount *
311 	    sizeof(struct vm_page)));
312 	pagearray = (struct vm_page *)(bucketarray + bucketcount);
313 
314 	for (lcv = 0; lcv < VM_NFREELIST; lcv++) {
315 		uvm.page_free[lcv].pgfl_buckets =
316 		    (bucketarray + (lcv * uvmexp.ncolors));
317 		uvm_page_init_buckets(&uvm.page_free[lcv]);
318 	}
319 	memset(pagearray, 0, pagecount * sizeof(struct vm_page));
320 
321 	/*
322 	 * init the vm_page structures and put them in the correct place.
323 	 */
324 
325 	for (lcv = 0 ; lcv < vm_nphysseg ; lcv++) {
326 		n = vm_physmem[lcv].end - vm_physmem[lcv].start;
327 
328 		/* set up page array pointers */
329 		vm_physmem[lcv].pgs = pagearray;
330 		pagearray += n;
331 		pagecount -= n;
332 		vm_physmem[lcv].lastpg = vm_physmem[lcv].pgs + (n - 1);
333 
334 		/* init and free vm_pages (we've already zeroed them) */
335 		paddr = ptoa(vm_physmem[lcv].start);
336 		for (i = 0 ; i < n ; i++, paddr += PAGE_SIZE) {
337 			vm_physmem[lcv].pgs[i].phys_addr = paddr;
338 #ifdef __HAVE_VM_PAGE_MD
339 			VM_MDPAGE_INIT(&vm_physmem[lcv].pgs[i]);
340 #endif
341 			if (atop(paddr) >= vm_physmem[lcv].avail_start &&
342 			    atop(paddr) <= vm_physmem[lcv].avail_end) {
343 				uvmexp.npages++;
344 				/* add page to free pool */
345 				uvm_pagefree(&vm_physmem[lcv].pgs[i]);
346 			}
347 		}
348 	}
349 
350 	/*
351 	 * pass up the values of virtual_space_start and
352 	 * virtual_space_end (obtained by uvm_pageboot_alloc) to the upper
353 	 * layers of the VM.
354 	 */
355 
356 	*kvm_startp = round_page(virtual_space_start);
357 	*kvm_endp = trunc_page(virtual_space_end);
358 
359 	/*
360 	 * init locks for kernel threads
361 	 */
362 
363 	simple_lock_init(&uvm.pagedaemon_lock);
364 	simple_lock_init(&uvm.aiodoned_lock);
365 
366 	/*
367 	 * init various thresholds.
368 	 * XXXCDC - values may need adjusting
369 	 */
370 
371 	uvmexp.reserve_pagedaemon = 1;
372 	uvmexp.reserve_kernel = 5;
373 	uvmexp.anonminpct = 10;
374 	uvmexp.vnodeminpct = 10;
375 	uvmexp.vtextminpct = 5;
376 	uvmexp.anonmin = uvmexp.anonminpct * 256 / 100;
377 	uvmexp.vnodemin = uvmexp.vnodeminpct * 256 / 100;
378 	uvmexp.vtextmin = uvmexp.vtextminpct * 256 / 100;
379 
380 	/*
381 	 * determine if we should zero pages in the idle loop.
382 	 */
383 
384 	uvm.page_idle_zero = vm_page_zero_enable;
385 
386 	/*
387 	 * done!
388 	 */
389 
390 	uvm.page_init_done = TRUE;
391 }
392 
393 /*
394  * uvm_setpagesize: set the page size
395  *
396  * => sets page_shift and page_mask from uvmexp.pagesize.
397  */
398 
399 void
400 uvm_setpagesize()
401 {
402 	if (uvmexp.pagesize == 0)
403 		uvmexp.pagesize = DEFAULT_PAGE_SIZE;
404 	uvmexp.pagemask = uvmexp.pagesize - 1;
405 	if ((uvmexp.pagemask & uvmexp.pagesize) != 0)
406 		panic("uvm_setpagesize: page size not a power of two");
407 	for (uvmexp.pageshift = 0; ; uvmexp.pageshift++)
408 		if ((1 << uvmexp.pageshift) == uvmexp.pagesize)
409 			break;
410 }
411 
412 /*
413  * uvm_pageboot_alloc: steal memory from physmem for bootstrapping
414  */
415 
416 vaddr_t
417 uvm_pageboot_alloc(size)
418 	vsize_t size;
419 {
420 	static boolean_t initialized = FALSE;
421 	vaddr_t addr;
422 #if !defined(PMAP_STEAL_MEMORY)
423 	vaddr_t vaddr;
424 	paddr_t paddr;
425 #endif
426 
427 	/*
428 	 * on first call to this function, initialize ourselves.
429 	 */
430 	if (initialized == FALSE) {
431 		pmap_virtual_space(&virtual_space_start, &virtual_space_end);
432 
433 		/* round it the way we like it */
434 		virtual_space_start = round_page(virtual_space_start);
435 		virtual_space_end = trunc_page(virtual_space_end);
436 
437 		initialized = TRUE;
438 	}
439 
440 	/* round to page size */
441 	size = round_page(size);
442 
443 #if defined(PMAP_STEAL_MEMORY)
444 
445 	/*
446 	 * defer bootstrap allocation to MD code (it may want to allocate
447 	 * from a direct-mapped segment).  pmap_steal_memory should adjust
448 	 * virtual_space_start/virtual_space_end if necessary.
449 	 */
450 
451 	addr = pmap_steal_memory(size, &virtual_space_start,
452 	    &virtual_space_end);
453 
454 	return(addr);
455 
456 #else /* !PMAP_STEAL_MEMORY */
457 
458 	/*
459 	 * allocate virtual memory for this request
460 	 */
461 	if (virtual_space_start == virtual_space_end ||
462 	    (virtual_space_end - virtual_space_start) < size)
463 		panic("uvm_pageboot_alloc: out of virtual space");
464 
465 	addr = virtual_space_start;
466 
467 #ifdef PMAP_GROWKERNEL
468 	/*
469 	 * If the kernel pmap can't map the requested space,
470 	 * then allocate more resources for it.
471 	 */
472 	if (uvm_maxkaddr < (addr + size)) {
473 		uvm_maxkaddr = pmap_growkernel(addr + size);
474 		if (uvm_maxkaddr < (addr + size))
475 			panic("uvm_pageboot_alloc: pmap_growkernel() failed");
476 	}
477 #endif
478 
479 	virtual_space_start += size;
480 
481 	/*
482 	 * allocate and mapin physical pages to back new virtual pages
483 	 */
484 
485 	for (vaddr = round_page(addr) ; vaddr < addr + size ;
486 	    vaddr += PAGE_SIZE) {
487 
488 		if (!uvm_page_physget(&paddr))
489 			panic("uvm_pageboot_alloc: out of memory");
490 
491 		/*
492 		 * Note this memory is no longer managed, so using
493 		 * pmap_kenter is safe.
494 		 */
495 		pmap_kenter_pa(vaddr, paddr, VM_PROT_READ|VM_PROT_WRITE);
496 	}
497 	pmap_update(pmap_kernel());
498 	return(addr);
499 #endif	/* PMAP_STEAL_MEMORY */
500 }
501 
502 #if !defined(PMAP_STEAL_MEMORY)
503 /*
504  * uvm_page_physget: "steal" one page from the vm_physmem structure.
505  *
506  * => attempt to allocate it off the end of a segment in which the "avail"
507  *    values match the start/end values.   if we can't do that, then we
508  *    will advance both values (making them equal, and removing some
509  *    vm_page structures from the non-avail area).
510  * => return false if out of memory.
511  */
512 
513 /* subroutine: try to allocate from memory chunks on the specified freelist */
514 static boolean_t uvm_page_physget_freelist __P((paddr_t *, int));
515 
516 static boolean_t
517 uvm_page_physget_freelist(paddrp, freelist)
518 	paddr_t *paddrp;
519 	int freelist;
520 {
521 	int lcv, x;
522 
523 	/* pass 1: try allocating from a matching end */
524 #if (VM_PHYSSEG_STRAT == VM_PSTRAT_BIGFIRST)
525 	for (lcv = vm_nphysseg - 1 ; lcv >= 0 ; lcv--)
526 #else
527 	for (lcv = 0 ; lcv < vm_nphysseg ; lcv++)
528 #endif
529 	{
530 
531 		if (uvm.page_init_done == TRUE)
532 			panic("uvm_page_physget: called _after_ bootstrap");
533 
534 		if (vm_physmem[lcv].free_list != freelist)
535 			continue;
536 
537 		/* try from front */
538 		if (vm_physmem[lcv].avail_start == vm_physmem[lcv].start &&
539 		    vm_physmem[lcv].avail_start < vm_physmem[lcv].avail_end) {
540 			*paddrp = ptoa(vm_physmem[lcv].avail_start);
541 			vm_physmem[lcv].avail_start++;
542 			vm_physmem[lcv].start++;
543 			/* nothing left?   nuke it */
544 			if (vm_physmem[lcv].avail_start ==
545 			    vm_physmem[lcv].end) {
546 				if (vm_nphysseg == 1)
547 				    panic("vum_page_physget: out of memory!");
548 				vm_nphysseg--;
549 				for (x = lcv ; x < vm_nphysseg ; x++)
550 					/* structure copy */
551 					vm_physmem[x] = vm_physmem[x+1];
552 			}
553 			return (TRUE);
554 		}
555 
556 		/* try from rear */
557 		if (vm_physmem[lcv].avail_end == vm_physmem[lcv].end &&
558 		    vm_physmem[lcv].avail_start < vm_physmem[lcv].avail_end) {
559 			*paddrp = ptoa(vm_physmem[lcv].avail_end - 1);
560 			vm_physmem[lcv].avail_end--;
561 			vm_physmem[lcv].end--;
562 			/* nothing left?   nuke it */
563 			if (vm_physmem[lcv].avail_end ==
564 			    vm_physmem[lcv].start) {
565 				if (vm_nphysseg == 1)
566 				    panic("uvm_page_physget: out of memory!");
567 				vm_nphysseg--;
568 				for (x = lcv ; x < vm_nphysseg ; x++)
569 					/* structure copy */
570 					vm_physmem[x] = vm_physmem[x+1];
571 			}
572 			return (TRUE);
573 		}
574 	}
575 
576 	/* pass2: forget about matching ends, just allocate something */
577 #if (VM_PHYSSEG_STRAT == VM_PSTRAT_BIGFIRST)
578 	for (lcv = vm_nphysseg - 1 ; lcv >= 0 ; lcv--)
579 #else
580 	for (lcv = 0 ; lcv < vm_nphysseg ; lcv++)
581 #endif
582 	{
583 
584 		/* any room in this bank? */
585 		if (vm_physmem[lcv].avail_start >= vm_physmem[lcv].avail_end)
586 			continue;  /* nope */
587 
588 		*paddrp = ptoa(vm_physmem[lcv].avail_start);
589 		vm_physmem[lcv].avail_start++;
590 		/* truncate! */
591 		vm_physmem[lcv].start = vm_physmem[lcv].avail_start;
592 
593 		/* nothing left?   nuke it */
594 		if (vm_physmem[lcv].avail_start == vm_physmem[lcv].end) {
595 			if (vm_nphysseg == 1)
596 				panic("uvm_page_physget: out of memory!");
597 			vm_nphysseg--;
598 			for (x = lcv ; x < vm_nphysseg ; x++)
599 				/* structure copy */
600 				vm_physmem[x] = vm_physmem[x+1];
601 		}
602 		return (TRUE);
603 	}
604 
605 	return (FALSE);        /* whoops! */
606 }
607 
608 boolean_t
609 uvm_page_physget(paddrp)
610 	paddr_t *paddrp;
611 {
612 	int i;
613 
614 	/* try in the order of freelist preference */
615 	for (i = 0; i < VM_NFREELIST; i++)
616 		if (uvm_page_physget_freelist(paddrp, i) == TRUE)
617 			return (TRUE);
618 	return (FALSE);
619 }
620 #endif /* PMAP_STEAL_MEMORY */
621 
622 /*
623  * uvm_page_physload: load physical memory into VM system
624  *
625  * => all args are PFs
626  * => all pages in start/end get vm_page structures
627  * => areas marked by avail_start/avail_end get added to the free page pool
628  * => we are limited to VM_PHYSSEG_MAX physical memory segments
629  */
630 
631 void
632 uvm_page_physload(start, end, avail_start, avail_end, free_list)
633 	paddr_t start, end, avail_start, avail_end;
634 	int free_list;
635 {
636 	int preload, lcv;
637 	psize_t npages;
638 	struct vm_page *pgs;
639 	struct vm_physseg *ps;
640 
641 	if (uvmexp.pagesize == 0)
642 		panic("uvm_page_physload: page size not set!");
643 	if (free_list >= VM_NFREELIST || free_list < VM_FREELIST_DEFAULT)
644 		panic("uvm_page_physload: bad free list %d\n", free_list);
645 	if (start >= end)
646 		panic("uvm_page_physload: start >= end");
647 
648 	/*
649 	 * do we have room?
650 	 */
651 
652 	if (vm_nphysseg == VM_PHYSSEG_MAX) {
653 		printf("uvm_page_physload: unable to load physical memory "
654 		    "segment\n");
655 		printf("\t%d segments allocated, ignoring 0x%llx -> 0x%llx\n",
656 		    VM_PHYSSEG_MAX, (long long)start, (long long)end);
657 		printf("\tincrease VM_PHYSSEG_MAX\n");
658 		return;
659 	}
660 
661 	/*
662 	 * check to see if this is a "preload" (i.e. uvm_mem_init hasn't been
663 	 * called yet, so malloc is not available).
664 	 */
665 
666 	for (lcv = 0 ; lcv < vm_nphysseg ; lcv++) {
667 		if (vm_physmem[lcv].pgs)
668 			break;
669 	}
670 	preload = (lcv == vm_nphysseg);
671 
672 	/*
673 	 * if VM is already running, attempt to malloc() vm_page structures
674 	 */
675 
676 	if (!preload) {
677 #if defined(VM_PHYSSEG_NOADD)
678 		panic("uvm_page_physload: tried to add RAM after vm_mem_init");
679 #else
680 		/* XXXCDC: need some sort of lockout for this case */
681 		paddr_t paddr;
682 		npages = end - start;  /* # of pages */
683 		pgs = malloc(sizeof(struct vm_page) * npages,
684 		    M_VMPAGE, M_NOWAIT);
685 		if (pgs == NULL) {
686 			printf("uvm_page_physload: can not malloc vm_page "
687 			    "structs for segment\n");
688 			printf("\tignoring 0x%lx -> 0x%lx\n", start, end);
689 			return;
690 		}
691 		/* zero data, init phys_addr and free_list, and free pages */
692 		memset(pgs, 0, sizeof(struct vm_page) * npages);
693 		for (lcv = 0, paddr = ptoa(start) ;
694 				 lcv < npages ; lcv++, paddr += PAGE_SIZE) {
695 			pgs[lcv].phys_addr = paddr;
696 			pgs[lcv].free_list = free_list;
697 			if (atop(paddr) >= avail_start &&
698 			    atop(paddr) <= avail_end)
699 				uvm_pagefree(&pgs[lcv]);
700 		}
701 		/* XXXCDC: incomplete: need to update uvmexp.free, what else? */
702 		/* XXXCDC: need hook to tell pmap to rebuild pv_list, etc... */
703 #endif
704 	} else {
705 		pgs = NULL;
706 		npages = 0;
707 	}
708 
709 	/*
710 	 * now insert us in the proper place in vm_physmem[]
711 	 */
712 
713 #if (VM_PHYSSEG_STRAT == VM_PSTRAT_RANDOM)
714 	/* random: put it at the end (easy!) */
715 	ps = &vm_physmem[vm_nphysseg];
716 #elif (VM_PHYSSEG_STRAT == VM_PSTRAT_BSEARCH)
717 	{
718 		int x;
719 		/* sort by address for binary search */
720 		for (lcv = 0 ; lcv < vm_nphysseg ; lcv++)
721 			if (start < vm_physmem[lcv].start)
722 				break;
723 		ps = &vm_physmem[lcv];
724 		/* move back other entries, if necessary ... */
725 		for (x = vm_nphysseg ; x > lcv ; x--)
726 			/* structure copy */
727 			vm_physmem[x] = vm_physmem[x - 1];
728 	}
729 #elif (VM_PHYSSEG_STRAT == VM_PSTRAT_BIGFIRST)
730 	{
731 		int x;
732 		/* sort by largest segment first */
733 		for (lcv = 0 ; lcv < vm_nphysseg ; lcv++)
734 			if ((end - start) >
735 			    (vm_physmem[lcv].end - vm_physmem[lcv].start))
736 				break;
737 		ps = &vm_physmem[lcv];
738 		/* move back other entries, if necessary ... */
739 		for (x = vm_nphysseg ; x > lcv ; x--)
740 			/* structure copy */
741 			vm_physmem[x] = vm_physmem[x - 1];
742 	}
743 #else
744 	panic("uvm_page_physload: unknown physseg strategy selected!");
745 #endif
746 
747 	ps->start = start;
748 	ps->end = end;
749 	ps->avail_start = avail_start;
750 	ps->avail_end = avail_end;
751 	if (preload) {
752 		ps->pgs = NULL;
753 	} else {
754 		ps->pgs = pgs;
755 		ps->lastpg = pgs + npages - 1;
756 	}
757 	ps->free_list = free_list;
758 	vm_nphysseg++;
759 
760 	if (!preload)
761 		uvm_page_rehash();
762 }
763 
764 /*
765  * uvm_page_rehash: reallocate hash table based on number of free pages.
766  */
767 
768 void
769 uvm_page_rehash()
770 {
771 	int freepages, lcv, bucketcount, oldcount;
772 	struct pglist *newbuckets, *oldbuckets;
773 	struct vm_page *pg;
774 	size_t newsize, oldsize;
775 
776 	/*
777 	 * compute number of pages that can go in the free pool
778 	 */
779 
780 	freepages = 0;
781 	for (lcv = 0 ; lcv < vm_nphysseg ; lcv++)
782 		freepages +=
783 		    (vm_physmem[lcv].avail_end - vm_physmem[lcv].avail_start);
784 
785 	/*
786 	 * compute number of buckets needed for this number of pages
787 	 */
788 
789 	bucketcount = 1;
790 	while (bucketcount < freepages)
791 		bucketcount = bucketcount * 2;
792 
793 	/*
794 	 * compute the size of the current table and new table.
795 	 */
796 
797 	oldbuckets = uvm.page_hash;
798 	oldcount = uvm.page_nhash;
799 	oldsize = round_page(sizeof(struct pglist) * oldcount);
800 	newsize = round_page(sizeof(struct pglist) * bucketcount);
801 
802 	/*
803 	 * allocate the new buckets
804 	 */
805 
806 	newbuckets = (struct pglist *) uvm_km_alloc(kernel_map, newsize);
807 	if (newbuckets == NULL) {
808 		printf("uvm_page_physrehash: WARNING: could not grow page "
809 		    "hash table\n");
810 		return;
811 	}
812 	for (lcv = 0 ; lcv < bucketcount ; lcv++)
813 		TAILQ_INIT(&newbuckets[lcv]);
814 
815 	/*
816 	 * now replace the old buckets with the new ones and rehash everything
817 	 */
818 
819 	simple_lock(&uvm.hashlock);
820 	uvm.page_hash = newbuckets;
821 	uvm.page_nhash = bucketcount;
822 	uvm.page_hashmask = bucketcount - 1;  /* power of 2 */
823 
824 	/* ... and rehash */
825 	for (lcv = 0 ; lcv < oldcount ; lcv++) {
826 		while ((pg = oldbuckets[lcv].tqh_first) != NULL) {
827 			TAILQ_REMOVE(&oldbuckets[lcv], pg, hashq);
828 			TAILQ_INSERT_TAIL(
829 			  &uvm.page_hash[uvm_pagehash(pg->uobject, pg->offset)],
830 			  pg, hashq);
831 		}
832 	}
833 	simple_unlock(&uvm.hashlock);
834 
835 	/*
836 	 * free old bucket array if is not the boot-time table
837 	 */
838 
839 	if (oldbuckets != &uvm_bootbucket)
840 		uvm_km_free(kernel_map, (vaddr_t) oldbuckets, oldsize);
841 }
842 
843 /*
844  * uvm_page_recolor: Recolor the pages if the new bucket count is
845  * larger than the old one.
846  */
847 
848 void
849 uvm_page_recolor(int newncolors)
850 {
851 	struct pgflbucket *bucketarray, *oldbucketarray;
852 	struct pgfreelist pgfl;
853 	struct vm_page *pg;
854 	vsize_t bucketcount;
855 	int s, lcv, color, i, ocolors;
856 
857 	if (newncolors <= uvmexp.ncolors)
858 		return;
859 
860 	bucketcount = newncolors * VM_NFREELIST;
861 	bucketarray = malloc(bucketcount * sizeof(struct pgflbucket),
862 	    M_VMPAGE, M_NOWAIT);
863 	if (bucketarray == NULL) {
864 		printf("WARNING: unable to allocate %ld page color buckets\n",
865 		    (long) bucketcount);
866 		return;
867 	}
868 
869 	s = uvm_lock_fpageq();
870 
871 	/* Make sure we should still do this. */
872 	if (newncolors <= uvmexp.ncolors) {
873 		uvm_unlock_fpageq(s);
874 		free(bucketarray, M_VMPAGE);
875 		return;
876 	}
877 
878 	oldbucketarray = uvm.page_free[0].pgfl_buckets;
879 	ocolors = uvmexp.ncolors;
880 
881 	uvmexp.ncolors = newncolors;
882 	uvmexp.colormask = uvmexp.ncolors - 1;
883 
884 	for (lcv = 0; lcv < VM_NFREELIST; lcv++) {
885 		pgfl.pgfl_buckets = (bucketarray + (lcv * newncolors));
886 		uvm_page_init_buckets(&pgfl);
887 		for (color = 0; color < ocolors; color++) {
888 			for (i = 0; i < PGFL_NQUEUES; i++) {
889 				while ((pg = TAILQ_FIRST(&uvm.page_free[
890 				    lcv].pgfl_buckets[color].pgfl_queues[i]))
891 				    != NULL) {
892 					TAILQ_REMOVE(&uvm.page_free[
893 					    lcv].pgfl_buckets[
894 					    color].pgfl_queues[i], pg, pageq);
895 					TAILQ_INSERT_TAIL(&pgfl.pgfl_buckets[
896 					    VM_PGCOLOR_BUCKET(pg)].pgfl_queues[
897 					    i], pg, pageq);
898 				}
899 			}
900 		}
901 		uvm.page_free[lcv].pgfl_buckets = pgfl.pgfl_buckets;
902 	}
903 
904 	if (have_recolored_pages) {
905 		uvm_unlock_fpageq(s);
906 		free(oldbucketarray, M_VMPAGE);
907 		return;
908 	}
909 
910 	have_recolored_pages = TRUE;
911 	uvm_unlock_fpageq(s);
912 }
913 
914 /*
915  * uvm_pagealloc_pgfl: helper routine for uvm_pagealloc_strat
916  */
917 
918 static __inline struct vm_page *
919 uvm_pagealloc_pgfl(struct pgfreelist *pgfl, int try1, int try2,
920     unsigned int *trycolorp)
921 {
922 	struct pglist *freeq;
923 	struct vm_page *pg;
924 	int color, trycolor = *trycolorp;
925 
926 	color = trycolor;
927 	do {
928 		if ((pg = TAILQ_FIRST((freeq =
929 		    &pgfl->pgfl_buckets[color].pgfl_queues[try1]))) != NULL)
930 			goto gotit;
931 		if ((pg = TAILQ_FIRST((freeq =
932 		    &pgfl->pgfl_buckets[color].pgfl_queues[try2]))) != NULL)
933 			goto gotit;
934 		color = (color + 1) & uvmexp.colormask;
935 	} while (color != trycolor);
936 
937 	return (NULL);
938 
939  gotit:
940 	TAILQ_REMOVE(freeq, pg, pageq);
941 	uvmexp.free--;
942 
943 	/* update zero'd page count */
944 	if (pg->flags & PG_ZERO)
945 		uvmexp.zeropages--;
946 
947 	if (color == trycolor)
948 		uvmexp.colorhit++;
949 	else {
950 		uvmexp.colormiss++;
951 		*trycolorp = color;
952 	}
953 
954 	return (pg);
955 }
956 
957 /*
958  * uvm_pagealloc_strat: allocate vm_page from a particular free list.
959  *
960  * => return null if no pages free
961  * => wake up pagedaemon if number of free pages drops below low water mark
962  * => if obj != NULL, obj must be locked (to put in hash)
963  * => if anon != NULL, anon must be locked (to put in anon)
964  * => only one of obj or anon can be non-null
965  * => caller must activate/deactivate page if it is not wired.
966  * => free_list is ignored if strat == UVM_PGA_STRAT_NORMAL.
967  * => policy decision: it is more important to pull a page off of the
968  *	appropriate priority free list than it is to get a zero'd or
969  *	unknown contents page.  This is because we live with the
970  *	consequences of a bad free list decision for the entire
971  *	lifetime of the page, e.g. if the page comes from memory that
972  *	is slower to access.
973  */
974 
975 struct vm_page *
976 uvm_pagealloc_strat(obj, off, anon, flags, strat, free_list)
977 	struct uvm_object *obj;
978 	voff_t off;
979 	int flags;
980 	struct vm_anon *anon;
981 	int strat, free_list;
982 {
983 	int lcv, try1, try2, s, zeroit = 0, color;
984 	struct vm_page *pg;
985 	boolean_t use_reserve;
986 
987 	KASSERT(obj == NULL || anon == NULL);
988 	KASSERT(off == trunc_page(off));
989 	LOCK_ASSERT(obj == NULL || simple_lock_held(&obj->vmobjlock));
990 	LOCK_ASSERT(anon == NULL || simple_lock_held(&anon->an_lock));
991 
992 	s = uvm_lock_fpageq();
993 
994 	/*
995 	 * This implements a global round-robin page coloring
996 	 * algorithm.
997 	 *
998 	 * XXXJRT: Should we make the `nextcolor' per-cpu?
999 	 * XXXJRT: What about virtually-indexed caches?
1000 	 */
1001 
1002 	color = uvm.page_free_nextcolor;
1003 
1004 	/*
1005 	 * check to see if we need to generate some free pages waking
1006 	 * the pagedaemon.
1007 	 */
1008 
1009 	UVM_KICK_PDAEMON();
1010 
1011 	/*
1012 	 * fail if any of these conditions is true:
1013 	 * [1]  there really are no free pages, or
1014 	 * [2]  only kernel "reserved" pages remain and
1015 	 *        the page isn't being allocated to a kernel object.
1016 	 * [3]  only pagedaemon "reserved" pages remain and
1017 	 *        the requestor isn't the pagedaemon.
1018 	 */
1019 
1020 	use_reserve = (flags & UVM_PGA_USERESERVE) ||
1021 		(obj && UVM_OBJ_IS_KERN_OBJECT(obj));
1022 	if ((uvmexp.free <= uvmexp.reserve_kernel && !use_reserve) ||
1023 	    (uvmexp.free <= uvmexp.reserve_pagedaemon &&
1024 	     !(use_reserve && curproc == uvm.pagedaemon_proc)))
1025 		goto fail;
1026 
1027 #if PGFL_NQUEUES != 2
1028 #error uvm_pagealloc_strat needs to be updated
1029 #endif
1030 
1031 	/*
1032 	 * If we want a zero'd page, try the ZEROS queue first, otherwise
1033 	 * we try the UNKNOWN queue first.
1034 	 */
1035 	if (flags & UVM_PGA_ZERO) {
1036 		try1 = PGFL_ZEROS;
1037 		try2 = PGFL_UNKNOWN;
1038 	} else {
1039 		try1 = PGFL_UNKNOWN;
1040 		try2 = PGFL_ZEROS;
1041 	}
1042 
1043  again:
1044 	switch (strat) {
1045 	case UVM_PGA_STRAT_NORMAL:
1046 		/* Check all freelists in descending priority order. */
1047 		for (lcv = 0; lcv < VM_NFREELIST; lcv++) {
1048 			pg = uvm_pagealloc_pgfl(&uvm.page_free[lcv],
1049 			    try1, try2, &color);
1050 			if (pg != NULL)
1051 				goto gotit;
1052 		}
1053 
1054 		/* No pages free! */
1055 		goto fail;
1056 
1057 	case UVM_PGA_STRAT_ONLY:
1058 	case UVM_PGA_STRAT_FALLBACK:
1059 		/* Attempt to allocate from the specified free list. */
1060 		KASSERT(free_list >= 0 && free_list < VM_NFREELIST);
1061 		pg = uvm_pagealloc_pgfl(&uvm.page_free[free_list],
1062 		    try1, try2, &color);
1063 		if (pg != NULL)
1064 			goto gotit;
1065 
1066 		/* Fall back, if possible. */
1067 		if (strat == UVM_PGA_STRAT_FALLBACK) {
1068 			strat = UVM_PGA_STRAT_NORMAL;
1069 			goto again;
1070 		}
1071 
1072 		/* No pages free! */
1073 		goto fail;
1074 
1075 	default:
1076 		panic("uvm_pagealloc_strat: bad strat %d", strat);
1077 		/* NOTREACHED */
1078 	}
1079 
1080  gotit:
1081 	/*
1082 	 * We now know which color we actually allocated from; set
1083 	 * the next color accordingly.
1084 	 */
1085 
1086 	uvm.page_free_nextcolor = (color + 1) & uvmexp.colormask;
1087 
1088 	/*
1089 	 * update allocation statistics and remember if we have to
1090 	 * zero the page
1091 	 */
1092 
1093 	if (flags & UVM_PGA_ZERO) {
1094 		if (pg->flags & PG_ZERO) {
1095 			uvmexp.pga_zerohit++;
1096 			zeroit = 0;
1097 		} else {
1098 			uvmexp.pga_zeromiss++;
1099 			zeroit = 1;
1100 		}
1101 	}
1102 	uvm_unlock_fpageq(s);
1103 
1104 	pg->offset = off;
1105 	pg->uobject = obj;
1106 	pg->uanon = anon;
1107 	pg->flags = PG_BUSY|PG_CLEAN|PG_FAKE;
1108 	if (anon) {
1109 		anon->u.an_page = pg;
1110 		pg->pqflags = PQ_ANON;
1111 		uvmexp.anonpages++;
1112 	} else {
1113 		if (obj) {
1114 			uvm_pageinsert(pg);
1115 		}
1116 		pg->pqflags = 0;
1117 	}
1118 #if defined(UVM_PAGE_TRKOWN)
1119 	pg->owner_tag = NULL;
1120 #endif
1121 	UVM_PAGE_OWN(pg, "new alloc");
1122 
1123 	if (flags & UVM_PGA_ZERO) {
1124 		/*
1125 		 * A zero'd page is not clean.  If we got a page not already
1126 		 * zero'd, then we have to zero it ourselves.
1127 		 */
1128 		pg->flags &= ~PG_CLEAN;
1129 		if (zeroit)
1130 			pmap_zero_page(VM_PAGE_TO_PHYS(pg));
1131 	}
1132 
1133 	return(pg);
1134 
1135  fail:
1136 	uvm_unlock_fpageq(s);
1137 	return (NULL);
1138 }
1139 
1140 /*
1141  * uvm_pagerealloc: reallocate a page from one object to another
1142  *
1143  * => both objects must be locked
1144  */
1145 
1146 void
1147 uvm_pagerealloc(pg, newobj, newoff)
1148 	struct vm_page *pg;
1149 	struct uvm_object *newobj;
1150 	voff_t newoff;
1151 {
1152 	/*
1153 	 * remove it from the old object
1154 	 */
1155 
1156 	if (pg->uobject) {
1157 		uvm_pageremove(pg);
1158 	}
1159 
1160 	/*
1161 	 * put it in the new object
1162 	 */
1163 
1164 	if (newobj) {
1165 		pg->uobject = newobj;
1166 		pg->offset = newoff;
1167 		uvm_pageinsert(pg);
1168 	}
1169 }
1170 
1171 /*
1172  * uvm_pagefree: free page
1173  *
1174  * => erase page's identity (i.e. remove from hash/object)
1175  * => put page on free list
1176  * => caller must lock owning object (either anon or uvm_object)
1177  * => caller must lock page queues
1178  * => assumes all valid mappings of pg are gone
1179  */
1180 
1181 void
1182 uvm_pagefree(pg)
1183 	struct vm_page *pg;
1184 {
1185 	int s;
1186 
1187 	KASSERT((pg->flags & PG_PAGEOUT) == 0);
1188 	LOCK_ASSERT(simple_lock_held(&uvm.pageqlock) ||
1189 		    (pg->pqflags & (PQ_ACTIVE|PQ_INACTIVE)) == 0);
1190 
1191 #ifdef DEBUG
1192 	if (pg->uobject == (void *)0xdeadbeef &&
1193 	    pg->uanon == (void *)0xdeadbeef) {
1194 		panic("uvm_pagefree: freeing free page %p\n", pg);
1195 	}
1196 #endif
1197 
1198 	/*
1199 	 * if the page is loaned, resolve the loan instead of freeing.
1200 	 */
1201 
1202 	if (pg->loan_count) {
1203 
1204 		/*
1205 		 * if the page is owned by an anon then we just want to
1206 		 * drop ownership and return.  the kernel will free the page
1207 		 * when it is done with it.  if the page is owned by an object,
1208 		 * mark it clean for the benefit of possible anon owners
1209 		 * and mark it as anon-owned if there current anon loanees.
1210 		 */
1211 
1212 		if (pg->pqflags & PQ_ANON) {
1213 			pg->pqflags &= ~PQ_ANON;
1214 			pg->uanon = NULL;
1215 		} else if (pg->flags & PG_TABLED) {
1216 			pg->flags &= ~PG_CLEAN;
1217 			if (pg->uanon) {
1218 				pg->pqflags |= PQ_ANON;
1219 			}
1220 		}
1221 		return;
1222 	}
1223 
1224 	/*
1225 	 * remove page from its object or anon.
1226 	 * adjust swpgonly if the page is swap-backed.
1227 	 */
1228 
1229 	if (pg->flags & PG_TABLED) {
1230 		if (pg->pqflags & PQ_AOBJ &&
1231 		    uao_find_swslot(pg->uobject,
1232 				    pg->offset >> PAGE_SHIFT) != 0) {
1233 			simple_lock(&uvm.swap_data_lock);
1234 			uvmexp.swpgonly++;
1235 			simple_unlock(&uvm.swap_data_lock);
1236 		}
1237 		uvm_pageremove(pg);
1238 	} else if (pg->pqflags & PQ_ANON) {
1239 		if (pg->uanon->an_swslot) {
1240 			simple_lock(&uvm.swap_data_lock);
1241 			uvmexp.swpgonly++;
1242 			simple_unlock(&uvm.swap_data_lock);
1243 		}
1244 		pg->uanon->u.an_page = NULL;
1245 	}
1246 
1247 	/*
1248 	 * now remove the page from the queues
1249 	 */
1250 
1251 	uvm_pagedequeue(pg);
1252 
1253 	/*
1254 	 * if the page was wired, unwire it now.
1255 	 */
1256 
1257 	if (pg->wire_count) {
1258 		pg->wire_count = 0;
1259 		uvmexp.wired--;
1260 	}
1261 	if (pg->pqflags & PQ_ANON) {
1262 		uvmexp.anonpages--;
1263 	}
1264 
1265 	/*
1266 	 * and put on free queue
1267 	 */
1268 
1269 	pg->flags &= ~PG_ZERO;
1270 
1271 	s = uvm_lock_fpageq();
1272 	TAILQ_INSERT_TAIL(&uvm.page_free[
1273 	    uvm_page_lookup_freelist(pg)].pgfl_buckets[
1274 	    VM_PGCOLOR_BUCKET(pg)].pgfl_queues[PGFL_UNKNOWN], pg, pageq);
1275 	pg->pqflags = PQ_FREE;
1276 #ifdef DEBUG
1277 	pg->uobject = (void *)0xdeadbeef;
1278 	pg->offset = 0xdeadbeef;
1279 	pg->uanon = (void *)0xdeadbeef;
1280 #endif
1281 	uvmexp.free++;
1282 
1283 	if (uvmexp.zeropages < UVM_PAGEZERO_TARGET)
1284 		uvm.page_idle_zero = vm_page_zero_enable;
1285 
1286 	uvm_unlock_fpageq(s);
1287 }
1288 
1289 /*
1290  * uvm_page_unbusy: unbusy an array of pages.
1291  *
1292  * => pages must either all belong to the same object, or all belong to anons.
1293  * => if pages are object-owned, object must be locked.
1294  * => if pages are anon-owned, anons must be locked.
1295  */
1296 
1297 void
1298 uvm_page_unbusy(pgs, npgs)
1299 	struct vm_page **pgs;
1300 	int npgs;
1301 {
1302 	struct vm_page *pg;
1303 	int i;
1304 	UVMHIST_FUNC("uvm_page_unbusy"); UVMHIST_CALLED(ubchist);
1305 
1306 	for (i = 0; i < npgs; i++) {
1307 		pg = pgs[i];
1308 		if (pg == NULL) {
1309 			continue;
1310 		}
1311 		if (pg->flags & PG_WANTED) {
1312 			wakeup(pg);
1313 		}
1314 		if (pg->flags & PG_RELEASED) {
1315 			UVMHIST_LOG(ubchist, "releasing pg %p", pg,0,0,0);
1316 			pg->flags &= ~PG_RELEASED;
1317 			uvm_pagefree(pg);
1318 		} else {
1319 			UVMHIST_LOG(ubchist, "unbusying pg %p", pg,0,0,0);
1320 			pg->flags &= ~(PG_WANTED|PG_BUSY);
1321 			UVM_PAGE_OWN(pg, NULL);
1322 		}
1323 	}
1324 }
1325 
1326 #if defined(UVM_PAGE_TRKOWN)
1327 /*
1328  * uvm_page_own: set or release page ownership
1329  *
1330  * => this is a debugging function that keeps track of who sets PG_BUSY
1331  *	and where they do it.   it can be used to track down problems
1332  *	such a process setting "PG_BUSY" and never releasing it.
1333  * => page's object [if any] must be locked
1334  * => if "tag" is NULL then we are releasing page ownership
1335  */
1336 void
1337 uvm_page_own(pg, tag)
1338 	struct vm_page *pg;
1339 	char *tag;
1340 {
1341 	KASSERT((pg->flags & (PG_PAGEOUT|PG_RELEASED)) == 0);
1342 
1343 	/* gain ownership? */
1344 	if (tag) {
1345 		if (pg->owner_tag) {
1346 			printf("uvm_page_own: page %p already owned "
1347 			    "by proc %d [%s]\n", pg,
1348 			     pg->owner, pg->owner_tag);
1349 			panic("uvm_page_own");
1350 		}
1351 		pg->owner = (curproc) ? curproc->p_pid :  (pid_t) -1;
1352 		pg->owner_tag = tag;
1353 		return;
1354 	}
1355 
1356 	/* drop ownership */
1357 	if (pg->owner_tag == NULL) {
1358 		printf("uvm_page_own: dropping ownership of an non-owned "
1359 		    "page (%p)\n", pg);
1360 		panic("uvm_page_own");
1361 	}
1362 	pg->owner_tag = NULL;
1363 	KASSERT((pg->uanon == NULL && pg->uobject == NULL) ||
1364 		pg->uobject == uvm.kernel_object ||
1365 		pg->wire_count || (pg->pqflags & (PQ_ACTIVE|PQ_INACTIVE)));
1366 	return;
1367 }
1368 #endif
1369 
1370 /*
1371  * uvm_pageidlezero: zero free pages while the system is idle.
1372  *
1373  * => try to complete one color bucket at a time, to reduce our impact
1374  *	on the CPU cache.
1375  * => we loop until we either reach the target or whichqs indicates that
1376  *	there is a process ready to run.
1377  */
1378 void
1379 uvm_pageidlezero()
1380 {
1381 	struct vm_page *pg;
1382 	struct pgfreelist *pgfl;
1383 	int free_list, s, firstbucket;
1384 	static int nextbucket;
1385 
1386 	s = uvm_lock_fpageq();
1387 	firstbucket = nextbucket;
1388 	do {
1389 		if (sched_whichqs != 0) {
1390 			uvm_unlock_fpageq(s);
1391 			return;
1392 		}
1393 		if (uvmexp.zeropages >= UVM_PAGEZERO_TARGET) {
1394 			uvm.page_idle_zero = FALSE;
1395 			uvm_unlock_fpageq(s);
1396 			return;
1397 		}
1398 		for (free_list = 0; free_list < VM_NFREELIST; free_list++) {
1399 			pgfl = &uvm.page_free[free_list];
1400 			while ((pg = TAILQ_FIRST(&pgfl->pgfl_buckets[
1401 			    nextbucket].pgfl_queues[PGFL_UNKNOWN])) != NULL) {
1402 				if (sched_whichqs != 0) {
1403 					uvm_unlock_fpageq(s);
1404 					return;
1405 				}
1406 
1407 				TAILQ_REMOVE(&pgfl->pgfl_buckets[
1408 				    nextbucket].pgfl_queues[PGFL_UNKNOWN],
1409 				    pg, pageq);
1410 				uvmexp.free--;
1411 				uvm_unlock_fpageq(s);
1412 #ifdef PMAP_PAGEIDLEZERO
1413 				if (!PMAP_PAGEIDLEZERO(VM_PAGE_TO_PHYS(pg))) {
1414 
1415 					/*
1416 					 * The machine-dependent code detected
1417 					 * some reason for us to abort zeroing
1418 					 * pages, probably because there is a
1419 					 * process now ready to run.
1420 					 */
1421 
1422 					s = uvm_lock_fpageq();
1423 					TAILQ_INSERT_HEAD(&pgfl->pgfl_buckets[
1424 					    nextbucket].pgfl_queues[
1425 					    PGFL_UNKNOWN], pg, pageq);
1426 					uvmexp.free++;
1427 					uvmexp.zeroaborts++;
1428 					uvm_unlock_fpageq(s);
1429 					return;
1430 				}
1431 #else
1432 				pmap_zero_page(VM_PAGE_TO_PHYS(pg));
1433 #endif /* PMAP_PAGEIDLEZERO */
1434 				pg->flags |= PG_ZERO;
1435 
1436 				s = uvm_lock_fpageq();
1437 				TAILQ_INSERT_HEAD(&pgfl->pgfl_buckets[
1438 				    nextbucket].pgfl_queues[PGFL_ZEROS],
1439 				    pg, pageq);
1440 				uvmexp.free++;
1441 				uvmexp.zeropages++;
1442 			}
1443 		}
1444 		nextbucket = (nextbucket + 1) & uvmexp.colormask;
1445 	} while (nextbucket != firstbucket);
1446 	uvm_unlock_fpageq(s);
1447 }
1448