xref: /netbsd-src/sys/uvm/uvm_page.c (revision d20841bb642898112fe68f0ad3f7b26dddf56f07)
1 /*	$NetBSD: uvm_page.c,v 1.95 2004/02/13 11:36:23 wiz 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 1.95 2004/02/13 11:36:23 wiz 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 MALLOC_DEFINE(M_VMPAGE, "VM page", "VM page");
143 
144 #ifdef DEBUG
145 vaddr_t uvm_zerocheckkva;
146 #endif /* DEBUG */
147 
148 /*
149  * local prototypes
150  */
151 
152 static void uvm_pageinsert __P((struct vm_page *));
153 static void uvm_pageremove __P((struct vm_page *));
154 
155 /*
156  * inline functions
157  */
158 
159 /*
160  * uvm_pageinsert: insert a page in the object and the hash table
161  *
162  * => caller must lock object
163  * => caller must lock page queues
164  * => call should have already set pg's object and offset pointers
165  *    and bumped the version counter
166  */
167 
168 __inline static void
169 uvm_pageinsert(pg)
170 	struct vm_page *pg;
171 {
172 	struct pglist *buck;
173 	struct uvm_object *uobj = pg->uobject;
174 
175 	KASSERT((pg->flags & PG_TABLED) == 0);
176 	buck = &uvm.page_hash[uvm_pagehash(uobj, pg->offset)];
177 	simple_lock(&uvm.hashlock);
178 	TAILQ_INSERT_TAIL(buck, pg, hashq);
179 	simple_unlock(&uvm.hashlock);
180 
181 	if (UVM_OBJ_IS_VNODE(uobj)) {
182 		if (uobj->uo_npages == 0) {
183 			struct vnode *vp = (struct vnode *)uobj;
184 
185 			vholdl(vp);
186 		}
187 		if (UVM_OBJ_IS_VTEXT(uobj)) {
188 			uvmexp.execpages++;
189 		} else {
190 			uvmexp.filepages++;
191 		}
192 	} else if (UVM_OBJ_IS_AOBJ(uobj)) {
193 		uvmexp.anonpages++;
194 	}
195 
196 	TAILQ_INSERT_TAIL(&uobj->memq, pg, listq);
197 	pg->flags |= PG_TABLED;
198 	uobj->uo_npages++;
199 }
200 
201 /*
202  * uvm_page_remove: remove page from object and hash
203  *
204  * => caller must lock object
205  * => caller must lock page queues
206  */
207 
208 static __inline void
209 uvm_pageremove(pg)
210 	struct vm_page *pg;
211 {
212 	struct pglist *buck;
213 	struct uvm_object *uobj = pg->uobject;
214 
215 	KASSERT(pg->flags & PG_TABLED);
216 	buck = &uvm.page_hash[uvm_pagehash(uobj, pg->offset)];
217 	simple_lock(&uvm.hashlock);
218 	TAILQ_REMOVE(buck, pg, hashq);
219 	simple_unlock(&uvm.hashlock);
220 
221 	if (UVM_OBJ_IS_VNODE(uobj)) {
222 		if (uobj->uo_npages == 1) {
223 			struct vnode *vp = (struct vnode *)uobj;
224 
225 			holdrelel(vp);
226 		}
227 		if (UVM_OBJ_IS_VTEXT(uobj)) {
228 			uvmexp.execpages--;
229 		} else {
230 			uvmexp.filepages--;
231 		}
232 	} else if (UVM_OBJ_IS_AOBJ(uobj)) {
233 		uvmexp.anonpages--;
234 	}
235 
236 	/* object should be locked */
237 	uobj->uo_npages--;
238 	TAILQ_REMOVE(&uobj->memq, pg, listq);
239 	pg->flags &= ~PG_TABLED;
240 	pg->uobject = NULL;
241 }
242 
243 static void
244 uvm_page_init_buckets(struct pgfreelist *pgfl)
245 {
246 	int color, i;
247 
248 	for (color = 0; color < uvmexp.ncolors; color++) {
249 		for (i = 0; i < PGFL_NQUEUES; i++) {
250 			TAILQ_INIT(&pgfl->pgfl_buckets[color].pgfl_queues[i]);
251 		}
252 	}
253 }
254 
255 /*
256  * uvm_page_init: init the page system.   called from uvm_init().
257  *
258  * => we return the range of kernel virtual memory in kvm_startp/kvm_endp
259  */
260 
261 void
262 uvm_page_init(kvm_startp, kvm_endp)
263 	vaddr_t *kvm_startp, *kvm_endp;
264 {
265 	vsize_t freepages, pagecount, bucketcount, n;
266 	struct pgflbucket *bucketarray;
267 	struct vm_page *pagearray;
268 	int lcv;
269 	u_int i;
270 	paddr_t paddr;
271 
272 	/*
273 	 * init the page queues and page queue locks, except the free
274 	 * list; we allocate that later (with the initial vm_page
275 	 * structures).
276 	 */
277 
278 	TAILQ_INIT(&uvm.page_active);
279 	TAILQ_INIT(&uvm.page_inactive);
280 	simple_lock_init(&uvm.pageqlock);
281 	simple_lock_init(&uvm.fpageqlock);
282 
283 	/*
284 	 * init the <obj,offset> => <page> hash table.  for now
285 	 * we just have one bucket (the bootstrap bucket).  later on we
286 	 * will allocate new buckets as we dynamically resize the hash table.
287 	 */
288 
289 	uvm.page_nhash = 1;			/* 1 bucket */
290 	uvm.page_hashmask = 0;			/* mask for hash function */
291 	uvm.page_hash = &uvm_bootbucket;	/* install bootstrap bucket */
292 	TAILQ_INIT(uvm.page_hash);		/* init hash table */
293 	simple_lock_init(&uvm.hashlock);	/* init hash table lock */
294 
295 	/*
296 	 * allocate vm_page structures.
297 	 */
298 
299 	/*
300 	 * sanity check:
301 	 * before calling this function the MD code is expected to register
302 	 * some free RAM with the uvm_page_physload() function.   our job
303 	 * now is to allocate vm_page structures for this memory.
304 	 */
305 
306 	if (vm_nphysseg == 0)
307 		panic("uvm_page_bootstrap: no memory pre-allocated");
308 
309 	/*
310 	 * first calculate the number of free pages...
311 	 *
312 	 * note that we use start/end rather than avail_start/avail_end.
313 	 * this allows us to allocate extra vm_page structures in case we
314 	 * want to return some memory to the pool after booting.
315 	 */
316 
317 	freepages = 0;
318 	for (lcv = 0 ; lcv < vm_nphysseg ; lcv++)
319 		freepages += (vm_physmem[lcv].end - vm_physmem[lcv].start);
320 
321 	/*
322 	 * Let MD code initialize the number of colors, or default
323 	 * to 1 color if MD code doesn't care.
324 	 */
325 	if (uvmexp.ncolors == 0)
326 		uvmexp.ncolors = 1;
327 	uvmexp.colormask = uvmexp.ncolors - 1;
328 
329 	/*
330 	 * we now know we have (PAGE_SIZE * freepages) bytes of memory we can
331 	 * use.   for each page of memory we use we need a vm_page structure.
332 	 * thus, the total number of pages we can use is the total size of
333 	 * the memory divided by the PAGE_SIZE plus the size of the vm_page
334 	 * structure.   we add one to freepages as a fudge factor to avoid
335 	 * truncation errors (since we can only allocate in terms of whole
336 	 * pages).
337 	 */
338 
339 	bucketcount = uvmexp.ncolors * VM_NFREELIST;
340 	pagecount = ((freepages + 1) << PAGE_SHIFT) /
341 	    (PAGE_SIZE + sizeof(struct vm_page));
342 
343 	bucketarray = (void *)uvm_pageboot_alloc((bucketcount *
344 	    sizeof(struct pgflbucket)) + (pagecount *
345 	    sizeof(struct vm_page)));
346 	pagearray = (struct vm_page *)(bucketarray + bucketcount);
347 
348 	for (lcv = 0; lcv < VM_NFREELIST; lcv++) {
349 		uvm.page_free[lcv].pgfl_buckets =
350 		    (bucketarray + (lcv * uvmexp.ncolors));
351 		uvm_page_init_buckets(&uvm.page_free[lcv]);
352 	}
353 	memset(pagearray, 0, pagecount * sizeof(struct vm_page));
354 
355 	/*
356 	 * init the vm_page structures and put them in the correct place.
357 	 */
358 
359 	for (lcv = 0 ; lcv < vm_nphysseg ; lcv++) {
360 		n = vm_physmem[lcv].end - vm_physmem[lcv].start;
361 
362 		/* set up page array pointers */
363 		vm_physmem[lcv].pgs = pagearray;
364 		pagearray += n;
365 		pagecount -= n;
366 		vm_physmem[lcv].lastpg = vm_physmem[lcv].pgs + (n - 1);
367 
368 		/* init and free vm_pages (we've already zeroed them) */
369 		paddr = ptoa(vm_physmem[lcv].start);
370 		for (i = 0 ; i < n ; i++, paddr += PAGE_SIZE) {
371 			vm_physmem[lcv].pgs[i].phys_addr = paddr;
372 #ifdef __HAVE_VM_PAGE_MD
373 			VM_MDPAGE_INIT(&vm_physmem[lcv].pgs[i]);
374 #endif
375 			if (atop(paddr) >= vm_physmem[lcv].avail_start &&
376 			    atop(paddr) <= vm_physmem[lcv].avail_end) {
377 				uvmexp.npages++;
378 				/* add page to free pool */
379 				uvm_pagefree(&vm_physmem[lcv].pgs[i]);
380 			}
381 		}
382 	}
383 
384 	/*
385 	 * pass up the values of virtual_space_start and
386 	 * virtual_space_end (obtained by uvm_pageboot_alloc) to the upper
387 	 * layers of the VM.
388 	 */
389 
390 	*kvm_startp = round_page(virtual_space_start);
391 	*kvm_endp = trunc_page(virtual_space_end);
392 #ifdef DEBUG
393 	/*
394 	 * steal kva for uvm_pagezerocheck().
395 	 */
396 	uvm_zerocheckkva = *kvm_startp;
397 	*kvm_startp += PAGE_SIZE;
398 #endif /* DEBUG */
399 
400 	/*
401 	 * init locks for kernel threads
402 	 */
403 
404 	simple_lock_init(&uvm.pagedaemon_lock);
405 	simple_lock_init(&uvm.aiodoned_lock);
406 
407 	/*
408 	 * init various thresholds.
409 	 */
410 
411 	uvmexp.reserve_pagedaemon = 1;
412 	uvmexp.reserve_kernel = 5;
413 	uvmexp.anonminpct = 10;
414 	uvmexp.fileminpct = 10;
415 	uvmexp.execminpct = 5;
416 	uvmexp.anonmaxpct = 80;
417 	uvmexp.filemaxpct = 50;
418 	uvmexp.execmaxpct = 30;
419 	uvmexp.anonmin = uvmexp.anonminpct * 256 / 100;
420 	uvmexp.filemin = uvmexp.fileminpct * 256 / 100;
421 	uvmexp.execmin = uvmexp.execminpct * 256 / 100;
422 	uvmexp.anonmax = uvmexp.anonmaxpct * 256 / 100;
423 	uvmexp.filemax = uvmexp.filemaxpct * 256 / 100;
424 	uvmexp.execmax = uvmexp.execmaxpct * 256 / 100;
425 
426 	/*
427 	 * determine if we should zero pages in the idle loop.
428 	 */
429 
430 	uvm.page_idle_zero = vm_page_zero_enable;
431 
432 	/*
433 	 * done!
434 	 */
435 
436 	uvm.page_init_done = TRUE;
437 }
438 
439 /*
440  * uvm_setpagesize: set the page size
441  *
442  * => sets page_shift and page_mask from uvmexp.pagesize.
443  */
444 
445 void
446 uvm_setpagesize()
447 {
448 
449 	/*
450 	 * If uvmexp.pagesize is 0 at this point, we expect PAGE_SIZE
451 	 * to be a constant (indicated by being a non-zero value).
452 	 */
453 	if (uvmexp.pagesize == 0) {
454 		if (PAGE_SIZE == 0)
455 			panic("uvm_setpagesize: uvmexp.pagesize not set");
456 		uvmexp.pagesize = PAGE_SIZE;
457 	}
458 	uvmexp.pagemask = uvmexp.pagesize - 1;
459 	if ((uvmexp.pagemask & uvmexp.pagesize) != 0)
460 		panic("uvm_setpagesize: page size not a power of two");
461 	for (uvmexp.pageshift = 0; ; uvmexp.pageshift++)
462 		if ((1 << uvmexp.pageshift) == uvmexp.pagesize)
463 			break;
464 }
465 
466 /*
467  * uvm_pageboot_alloc: steal memory from physmem for bootstrapping
468  */
469 
470 vaddr_t
471 uvm_pageboot_alloc(size)
472 	vsize_t size;
473 {
474 	static boolean_t initialized = FALSE;
475 	vaddr_t addr;
476 #if !defined(PMAP_STEAL_MEMORY)
477 	vaddr_t vaddr;
478 	paddr_t paddr;
479 #endif
480 
481 	/*
482 	 * on first call to this function, initialize ourselves.
483 	 */
484 	if (initialized == FALSE) {
485 		pmap_virtual_space(&virtual_space_start, &virtual_space_end);
486 
487 		/* round it the way we like it */
488 		virtual_space_start = round_page(virtual_space_start);
489 		virtual_space_end = trunc_page(virtual_space_end);
490 
491 		initialized = TRUE;
492 	}
493 
494 	/* round to page size */
495 	size = round_page(size);
496 
497 #if defined(PMAP_STEAL_MEMORY)
498 
499 	/*
500 	 * defer bootstrap allocation to MD code (it may want to allocate
501 	 * from a direct-mapped segment).  pmap_steal_memory should adjust
502 	 * virtual_space_start/virtual_space_end if necessary.
503 	 */
504 
505 	addr = pmap_steal_memory(size, &virtual_space_start,
506 	    &virtual_space_end);
507 
508 	return(addr);
509 
510 #else /* !PMAP_STEAL_MEMORY */
511 
512 	/*
513 	 * allocate virtual memory for this request
514 	 */
515 	if (virtual_space_start == virtual_space_end ||
516 	    (virtual_space_end - virtual_space_start) < size)
517 		panic("uvm_pageboot_alloc: out of virtual space");
518 
519 	addr = virtual_space_start;
520 
521 #ifdef PMAP_GROWKERNEL
522 	/*
523 	 * If the kernel pmap can't map the requested space,
524 	 * then allocate more resources for it.
525 	 */
526 	if (uvm_maxkaddr < (addr + size)) {
527 		uvm_maxkaddr = pmap_growkernel(addr + size);
528 		if (uvm_maxkaddr < (addr + size))
529 			panic("uvm_pageboot_alloc: pmap_growkernel() failed");
530 	}
531 #endif
532 
533 	virtual_space_start += size;
534 
535 	/*
536 	 * allocate and mapin physical pages to back new virtual pages
537 	 */
538 
539 	for (vaddr = round_page(addr) ; vaddr < addr + size ;
540 	    vaddr += PAGE_SIZE) {
541 
542 		if (!uvm_page_physget(&paddr))
543 			panic("uvm_pageboot_alloc: out of memory");
544 
545 		/*
546 		 * Note this memory is no longer managed, so using
547 		 * pmap_kenter is safe.
548 		 */
549 		pmap_kenter_pa(vaddr, paddr, VM_PROT_READ|VM_PROT_WRITE);
550 	}
551 	pmap_update(pmap_kernel());
552 	return(addr);
553 #endif	/* PMAP_STEAL_MEMORY */
554 }
555 
556 #if !defined(PMAP_STEAL_MEMORY)
557 /*
558  * uvm_page_physget: "steal" one page from the vm_physmem structure.
559  *
560  * => attempt to allocate it off the end of a segment in which the "avail"
561  *    values match the start/end values.   if we can't do that, then we
562  *    will advance both values (making them equal, and removing some
563  *    vm_page structures from the non-avail area).
564  * => return false if out of memory.
565  */
566 
567 /* subroutine: try to allocate from memory chunks on the specified freelist */
568 static boolean_t uvm_page_physget_freelist __P((paddr_t *, int));
569 
570 static boolean_t
571 uvm_page_physget_freelist(paddrp, freelist)
572 	paddr_t *paddrp;
573 	int freelist;
574 {
575 	int lcv, x;
576 
577 	/* pass 1: try allocating from a matching end */
578 #if (VM_PHYSSEG_STRAT == VM_PSTRAT_BIGFIRST)
579 	for (lcv = vm_nphysseg - 1 ; lcv >= 0 ; lcv--)
580 #else
581 	for (lcv = 0 ; lcv < vm_nphysseg ; lcv++)
582 #endif
583 	{
584 
585 		if (uvm.page_init_done == TRUE)
586 			panic("uvm_page_physget: called _after_ bootstrap");
587 
588 		if (vm_physmem[lcv].free_list != freelist)
589 			continue;
590 
591 		/* try from front */
592 		if (vm_physmem[lcv].avail_start == vm_physmem[lcv].start &&
593 		    vm_physmem[lcv].avail_start < vm_physmem[lcv].avail_end) {
594 			*paddrp = ptoa(vm_physmem[lcv].avail_start);
595 			vm_physmem[lcv].avail_start++;
596 			vm_physmem[lcv].start++;
597 			/* nothing left?   nuke it */
598 			if (vm_physmem[lcv].avail_start ==
599 			    vm_physmem[lcv].end) {
600 				if (vm_nphysseg == 1)
601 				    panic("uvm_page_physget: out of memory!");
602 				vm_nphysseg--;
603 				for (x = lcv ; x < vm_nphysseg ; x++)
604 					/* structure copy */
605 					vm_physmem[x] = vm_physmem[x+1];
606 			}
607 			return (TRUE);
608 		}
609 
610 		/* try from rear */
611 		if (vm_physmem[lcv].avail_end == vm_physmem[lcv].end &&
612 		    vm_physmem[lcv].avail_start < vm_physmem[lcv].avail_end) {
613 			*paddrp = ptoa(vm_physmem[lcv].avail_end - 1);
614 			vm_physmem[lcv].avail_end--;
615 			vm_physmem[lcv].end--;
616 			/* nothing left?   nuke it */
617 			if (vm_physmem[lcv].avail_end ==
618 			    vm_physmem[lcv].start) {
619 				if (vm_nphysseg == 1)
620 				    panic("uvm_page_physget: out of memory!");
621 				vm_nphysseg--;
622 				for (x = lcv ; x < vm_nphysseg ; x++)
623 					/* structure copy */
624 					vm_physmem[x] = vm_physmem[x+1];
625 			}
626 			return (TRUE);
627 		}
628 	}
629 
630 	/* pass2: forget about matching ends, just allocate something */
631 #if (VM_PHYSSEG_STRAT == VM_PSTRAT_BIGFIRST)
632 	for (lcv = vm_nphysseg - 1 ; lcv >= 0 ; lcv--)
633 #else
634 	for (lcv = 0 ; lcv < vm_nphysseg ; lcv++)
635 #endif
636 	{
637 
638 		/* any room in this bank? */
639 		if (vm_physmem[lcv].avail_start >= vm_physmem[lcv].avail_end)
640 			continue;  /* nope */
641 
642 		*paddrp = ptoa(vm_physmem[lcv].avail_start);
643 		vm_physmem[lcv].avail_start++;
644 		/* truncate! */
645 		vm_physmem[lcv].start = vm_physmem[lcv].avail_start;
646 
647 		/* nothing left?   nuke it */
648 		if (vm_physmem[lcv].avail_start == vm_physmem[lcv].end) {
649 			if (vm_nphysseg == 1)
650 				panic("uvm_page_physget: out of memory!");
651 			vm_nphysseg--;
652 			for (x = lcv ; x < vm_nphysseg ; x++)
653 				/* structure copy */
654 				vm_physmem[x] = vm_physmem[x+1];
655 		}
656 		return (TRUE);
657 	}
658 
659 	return (FALSE);        /* whoops! */
660 }
661 
662 boolean_t
663 uvm_page_physget(paddrp)
664 	paddr_t *paddrp;
665 {
666 	int i;
667 
668 	/* try in the order of freelist preference */
669 	for (i = 0; i < VM_NFREELIST; i++)
670 		if (uvm_page_physget_freelist(paddrp, i) == TRUE)
671 			return (TRUE);
672 	return (FALSE);
673 }
674 #endif /* PMAP_STEAL_MEMORY */
675 
676 /*
677  * uvm_page_physload: load physical memory into VM system
678  *
679  * => all args are PFs
680  * => all pages in start/end get vm_page structures
681  * => areas marked by avail_start/avail_end get added to the free page pool
682  * => we are limited to VM_PHYSSEG_MAX physical memory segments
683  */
684 
685 void
686 uvm_page_physload(start, end, avail_start, avail_end, free_list)
687 	paddr_t start, end, avail_start, avail_end;
688 	int free_list;
689 {
690 	int preload, lcv;
691 	psize_t npages;
692 	struct vm_page *pgs;
693 	struct vm_physseg *ps;
694 
695 	if (uvmexp.pagesize == 0)
696 		panic("uvm_page_physload: page size not set!");
697 	if (free_list >= VM_NFREELIST || free_list < VM_FREELIST_DEFAULT)
698 		panic("uvm_page_physload: bad free list %d", free_list);
699 	if (start >= end)
700 		panic("uvm_page_physload: start >= end");
701 
702 	/*
703 	 * do we have room?
704 	 */
705 
706 	if (vm_nphysseg == VM_PHYSSEG_MAX) {
707 		printf("uvm_page_physload: unable to load physical memory "
708 		    "segment\n");
709 		printf("\t%d segments allocated, ignoring 0x%llx -> 0x%llx\n",
710 		    VM_PHYSSEG_MAX, (long long)start, (long long)end);
711 		printf("\tincrease VM_PHYSSEG_MAX\n");
712 		return;
713 	}
714 
715 	/*
716 	 * check to see if this is a "preload" (i.e. uvm_mem_init hasn't been
717 	 * called yet, so malloc is not available).
718 	 */
719 
720 	for (lcv = 0 ; lcv < vm_nphysseg ; lcv++) {
721 		if (vm_physmem[lcv].pgs)
722 			break;
723 	}
724 	preload = (lcv == vm_nphysseg);
725 
726 	/*
727 	 * if VM is already running, attempt to malloc() vm_page structures
728 	 */
729 
730 	if (!preload) {
731 #if defined(VM_PHYSSEG_NOADD)
732 		panic("uvm_page_physload: tried to add RAM after vm_mem_init");
733 #else
734 		/* XXXCDC: need some sort of lockout for this case */
735 		paddr_t paddr;
736 		npages = end - start;  /* # of pages */
737 		pgs = malloc(sizeof(struct vm_page) * npages,
738 		    M_VMPAGE, M_NOWAIT);
739 		if (pgs == NULL) {
740 			printf("uvm_page_physload: can not malloc vm_page "
741 			    "structs for segment\n");
742 			printf("\tignoring 0x%lx -> 0x%lx\n", start, end);
743 			return;
744 		}
745 		/* zero data, init phys_addr and free_list, and free pages */
746 		memset(pgs, 0, sizeof(struct vm_page) * npages);
747 		for (lcv = 0, paddr = ptoa(start) ;
748 				 lcv < npages ; lcv++, paddr += PAGE_SIZE) {
749 			pgs[lcv].phys_addr = paddr;
750 			pgs[lcv].free_list = free_list;
751 			if (atop(paddr) >= avail_start &&
752 			    atop(paddr) <= avail_end)
753 				uvm_pagefree(&pgs[lcv]);
754 		}
755 		/* XXXCDC: incomplete: need to update uvmexp.free, what else? */
756 		/* XXXCDC: need hook to tell pmap to rebuild pv_list, etc... */
757 #endif
758 	} else {
759 		pgs = NULL;
760 		npages = 0;
761 	}
762 
763 	/*
764 	 * now insert us in the proper place in vm_physmem[]
765 	 */
766 
767 #if (VM_PHYSSEG_STRAT == VM_PSTRAT_RANDOM)
768 	/* random: put it at the end (easy!) */
769 	ps = &vm_physmem[vm_nphysseg];
770 #elif (VM_PHYSSEG_STRAT == VM_PSTRAT_BSEARCH)
771 	{
772 		int x;
773 		/* sort by address for binary search */
774 		for (lcv = 0 ; lcv < vm_nphysseg ; lcv++)
775 			if (start < vm_physmem[lcv].start)
776 				break;
777 		ps = &vm_physmem[lcv];
778 		/* move back other entries, if necessary ... */
779 		for (x = vm_nphysseg ; x > lcv ; x--)
780 			/* structure copy */
781 			vm_physmem[x] = vm_physmem[x - 1];
782 	}
783 #elif (VM_PHYSSEG_STRAT == VM_PSTRAT_BIGFIRST)
784 	{
785 		int x;
786 		/* sort by largest segment first */
787 		for (lcv = 0 ; lcv < vm_nphysseg ; lcv++)
788 			if ((end - start) >
789 			    (vm_physmem[lcv].end - vm_physmem[lcv].start))
790 				break;
791 		ps = &vm_physmem[lcv];
792 		/* move back other entries, if necessary ... */
793 		for (x = vm_nphysseg ; x > lcv ; x--)
794 			/* structure copy */
795 			vm_physmem[x] = vm_physmem[x - 1];
796 	}
797 #else
798 	panic("uvm_page_physload: unknown physseg strategy selected!");
799 #endif
800 
801 	ps->start = start;
802 	ps->end = end;
803 	ps->avail_start = avail_start;
804 	ps->avail_end = avail_end;
805 	if (preload) {
806 		ps->pgs = NULL;
807 	} else {
808 		ps->pgs = pgs;
809 		ps->lastpg = pgs + npages - 1;
810 	}
811 	ps->free_list = free_list;
812 	vm_nphysseg++;
813 
814 	if (!preload)
815 		uvm_page_rehash();
816 }
817 
818 /*
819  * uvm_page_rehash: reallocate hash table based on number of free pages.
820  */
821 
822 void
823 uvm_page_rehash()
824 {
825 	int freepages, lcv, bucketcount, oldcount;
826 	struct pglist *newbuckets, *oldbuckets;
827 	struct vm_page *pg;
828 	size_t newsize, oldsize;
829 
830 	/*
831 	 * compute number of pages that can go in the free pool
832 	 */
833 
834 	freepages = 0;
835 	for (lcv = 0 ; lcv < vm_nphysseg ; lcv++)
836 		freepages +=
837 		    (vm_physmem[lcv].avail_end - vm_physmem[lcv].avail_start);
838 
839 	/*
840 	 * compute number of buckets needed for this number of pages
841 	 */
842 
843 	bucketcount = 1;
844 	while (bucketcount < freepages)
845 		bucketcount = bucketcount * 2;
846 
847 	/*
848 	 * compute the size of the current table and new table.
849 	 */
850 
851 	oldbuckets = uvm.page_hash;
852 	oldcount = uvm.page_nhash;
853 	oldsize = round_page(sizeof(struct pglist) * oldcount);
854 	newsize = round_page(sizeof(struct pglist) * bucketcount);
855 
856 	/*
857 	 * allocate the new buckets
858 	 */
859 
860 	newbuckets = (struct pglist *) uvm_km_alloc(kernel_map, newsize);
861 	if (newbuckets == NULL) {
862 		printf("uvm_page_physrehash: WARNING: could not grow page "
863 		    "hash table\n");
864 		return;
865 	}
866 	for (lcv = 0 ; lcv < bucketcount ; lcv++)
867 		TAILQ_INIT(&newbuckets[lcv]);
868 
869 	/*
870 	 * now replace the old buckets with the new ones and rehash everything
871 	 */
872 
873 	simple_lock(&uvm.hashlock);
874 	uvm.page_hash = newbuckets;
875 	uvm.page_nhash = bucketcount;
876 	uvm.page_hashmask = bucketcount - 1;  /* power of 2 */
877 
878 	/* ... and rehash */
879 	for (lcv = 0 ; lcv < oldcount ; lcv++) {
880 		while ((pg = oldbuckets[lcv].tqh_first) != NULL) {
881 			TAILQ_REMOVE(&oldbuckets[lcv], pg, hashq);
882 			TAILQ_INSERT_TAIL(
883 			  &uvm.page_hash[uvm_pagehash(pg->uobject, pg->offset)],
884 			  pg, hashq);
885 		}
886 	}
887 	simple_unlock(&uvm.hashlock);
888 
889 	/*
890 	 * free old bucket array if is not the boot-time table
891 	 */
892 
893 	if (oldbuckets != &uvm_bootbucket)
894 		uvm_km_free(kernel_map, (vaddr_t) oldbuckets, oldsize);
895 }
896 
897 /*
898  * uvm_page_recolor: Recolor the pages if the new bucket count is
899  * larger than the old one.
900  */
901 
902 void
903 uvm_page_recolor(int newncolors)
904 {
905 	struct pgflbucket *bucketarray, *oldbucketarray;
906 	struct pgfreelist pgfl;
907 	struct vm_page *pg;
908 	vsize_t bucketcount;
909 	int s, lcv, color, i, ocolors;
910 
911 	if (newncolors <= uvmexp.ncolors)
912 		return;
913 
914 	if (uvm.page_init_done == FALSE) {
915 		uvmexp.ncolors = newncolors;
916 		return;
917 	}
918 
919 	bucketcount = newncolors * VM_NFREELIST;
920 	bucketarray = malloc(bucketcount * sizeof(struct pgflbucket),
921 	    M_VMPAGE, M_NOWAIT);
922 	if (bucketarray == NULL) {
923 		printf("WARNING: unable to allocate %ld page color buckets\n",
924 		    (long) bucketcount);
925 		return;
926 	}
927 
928 	s = uvm_lock_fpageq();
929 
930 	/* Make sure we should still do this. */
931 	if (newncolors <= uvmexp.ncolors) {
932 		uvm_unlock_fpageq(s);
933 		free(bucketarray, M_VMPAGE);
934 		return;
935 	}
936 
937 	oldbucketarray = uvm.page_free[0].pgfl_buckets;
938 	ocolors = uvmexp.ncolors;
939 
940 	uvmexp.ncolors = newncolors;
941 	uvmexp.colormask = uvmexp.ncolors - 1;
942 
943 	for (lcv = 0; lcv < VM_NFREELIST; lcv++) {
944 		pgfl.pgfl_buckets = (bucketarray + (lcv * newncolors));
945 		uvm_page_init_buckets(&pgfl);
946 		for (color = 0; color < ocolors; color++) {
947 			for (i = 0; i < PGFL_NQUEUES; i++) {
948 				while ((pg = TAILQ_FIRST(&uvm.page_free[
949 				    lcv].pgfl_buckets[color].pgfl_queues[i]))
950 				    != NULL) {
951 					TAILQ_REMOVE(&uvm.page_free[
952 					    lcv].pgfl_buckets[
953 					    color].pgfl_queues[i], pg, pageq);
954 					TAILQ_INSERT_TAIL(&pgfl.pgfl_buckets[
955 					    VM_PGCOLOR_BUCKET(pg)].pgfl_queues[
956 					    i], pg, pageq);
957 				}
958 			}
959 		}
960 		uvm.page_free[lcv].pgfl_buckets = pgfl.pgfl_buckets;
961 	}
962 
963 	if (have_recolored_pages) {
964 		uvm_unlock_fpageq(s);
965 		free(oldbucketarray, M_VMPAGE);
966 		return;
967 	}
968 
969 	have_recolored_pages = TRUE;
970 	uvm_unlock_fpageq(s);
971 }
972 
973 /*
974  * uvm_pagealloc_pgfl: helper routine for uvm_pagealloc_strat
975  */
976 
977 static __inline struct vm_page *
978 uvm_pagealloc_pgfl(struct pgfreelist *pgfl, int try1, int try2,
979     int *trycolorp)
980 {
981 	struct pglist *freeq;
982 	struct vm_page *pg;
983 	int color, trycolor = *trycolorp;
984 
985 	color = trycolor;
986 	do {
987 		if ((pg = TAILQ_FIRST((freeq =
988 		    &pgfl->pgfl_buckets[color].pgfl_queues[try1]))) != NULL)
989 			goto gotit;
990 		if ((pg = TAILQ_FIRST((freeq =
991 		    &pgfl->pgfl_buckets[color].pgfl_queues[try2]))) != NULL)
992 			goto gotit;
993 		color = (color + 1) & uvmexp.colormask;
994 	} while (color != trycolor);
995 
996 	return (NULL);
997 
998  gotit:
999 	TAILQ_REMOVE(freeq, pg, pageq);
1000 	uvmexp.free--;
1001 
1002 	/* update zero'd page count */
1003 	if (pg->flags & PG_ZERO)
1004 		uvmexp.zeropages--;
1005 
1006 	if (color == trycolor)
1007 		uvmexp.colorhit++;
1008 	else {
1009 		uvmexp.colormiss++;
1010 		*trycolorp = color;
1011 	}
1012 
1013 	return (pg);
1014 }
1015 
1016 /*
1017  * uvm_pagealloc_strat: allocate vm_page from a particular free list.
1018  *
1019  * => return null if no pages free
1020  * => wake up pagedaemon if number of free pages drops below low water mark
1021  * => if obj != NULL, obj must be locked (to put in hash)
1022  * => if anon != NULL, anon must be locked (to put in anon)
1023  * => only one of obj or anon can be non-null
1024  * => caller must activate/deactivate page if it is not wired.
1025  * => free_list is ignored if strat == UVM_PGA_STRAT_NORMAL.
1026  * => policy decision: it is more important to pull a page off of the
1027  *	appropriate priority free list than it is to get a zero'd or
1028  *	unknown contents page.  This is because we live with the
1029  *	consequences of a bad free list decision for the entire
1030  *	lifetime of the page, e.g. if the page comes from memory that
1031  *	is slower to access.
1032  */
1033 
1034 struct vm_page *
1035 uvm_pagealloc_strat(obj, off, anon, flags, strat, free_list)
1036 	struct uvm_object *obj;
1037 	voff_t off;
1038 	int flags;
1039 	struct vm_anon *anon;
1040 	int strat, free_list;
1041 {
1042 	int lcv, try1, try2, s, zeroit = 0, color;
1043 	struct vm_page *pg;
1044 	boolean_t use_reserve;
1045 
1046 	KASSERT(obj == NULL || anon == NULL);
1047 	KASSERT(off == trunc_page(off));
1048 	LOCK_ASSERT(obj == NULL || simple_lock_held(&obj->vmobjlock));
1049 	LOCK_ASSERT(anon == NULL || simple_lock_held(&anon->an_lock));
1050 
1051 	s = uvm_lock_fpageq();
1052 
1053 	/*
1054 	 * This implements a global round-robin page coloring
1055 	 * algorithm.
1056 	 *
1057 	 * XXXJRT: Should we make the `nextcolor' per-CPU?
1058 	 * XXXJRT: What about virtually-indexed caches?
1059 	 */
1060 
1061 	color = uvm.page_free_nextcolor;
1062 
1063 	/*
1064 	 * check to see if we need to generate some free pages waking
1065 	 * the pagedaemon.
1066 	 */
1067 
1068 	UVM_KICK_PDAEMON();
1069 
1070 	/*
1071 	 * fail if any of these conditions is true:
1072 	 * [1]  there really are no free pages, or
1073 	 * [2]  only kernel "reserved" pages remain and
1074 	 *        the page isn't being allocated to a kernel object.
1075 	 * [3]  only pagedaemon "reserved" pages remain and
1076 	 *        the requestor isn't the pagedaemon.
1077 	 */
1078 
1079 	use_reserve = (flags & UVM_PGA_USERESERVE) ||
1080 		(obj && UVM_OBJ_IS_KERN_OBJECT(obj));
1081 	if ((uvmexp.free <= uvmexp.reserve_kernel && !use_reserve) ||
1082 	    (uvmexp.free <= uvmexp.reserve_pagedaemon &&
1083 	     !(use_reserve && curproc == uvm.pagedaemon_proc)))
1084 		goto fail;
1085 
1086 #if PGFL_NQUEUES != 2
1087 #error uvm_pagealloc_strat needs to be updated
1088 #endif
1089 
1090 	/*
1091 	 * If we want a zero'd page, try the ZEROS queue first, otherwise
1092 	 * we try the UNKNOWN queue first.
1093 	 */
1094 	if (flags & UVM_PGA_ZERO) {
1095 		try1 = PGFL_ZEROS;
1096 		try2 = PGFL_UNKNOWN;
1097 	} else {
1098 		try1 = PGFL_UNKNOWN;
1099 		try2 = PGFL_ZEROS;
1100 	}
1101 
1102  again:
1103 	switch (strat) {
1104 	case UVM_PGA_STRAT_NORMAL:
1105 		/* Check all freelists in descending priority order. */
1106 		for (lcv = 0; lcv < VM_NFREELIST; lcv++) {
1107 			pg = uvm_pagealloc_pgfl(&uvm.page_free[lcv],
1108 			    try1, try2, &color);
1109 			if (pg != NULL)
1110 				goto gotit;
1111 		}
1112 
1113 		/* No pages free! */
1114 		goto fail;
1115 
1116 	case UVM_PGA_STRAT_ONLY:
1117 	case UVM_PGA_STRAT_FALLBACK:
1118 		/* Attempt to allocate from the specified free list. */
1119 		KASSERT(free_list >= 0 && free_list < VM_NFREELIST);
1120 		pg = uvm_pagealloc_pgfl(&uvm.page_free[free_list],
1121 		    try1, try2, &color);
1122 		if (pg != NULL)
1123 			goto gotit;
1124 
1125 		/* Fall back, if possible. */
1126 		if (strat == UVM_PGA_STRAT_FALLBACK) {
1127 			strat = UVM_PGA_STRAT_NORMAL;
1128 			goto again;
1129 		}
1130 
1131 		/* No pages free! */
1132 		goto fail;
1133 
1134 	default:
1135 		panic("uvm_pagealloc_strat: bad strat %d", strat);
1136 		/* NOTREACHED */
1137 	}
1138 
1139  gotit:
1140 	/*
1141 	 * We now know which color we actually allocated from; set
1142 	 * the next color accordingly.
1143 	 */
1144 
1145 	uvm.page_free_nextcolor = (color + 1) & uvmexp.colormask;
1146 
1147 	/*
1148 	 * update allocation statistics and remember if we have to
1149 	 * zero the page
1150 	 */
1151 
1152 	if (flags & UVM_PGA_ZERO) {
1153 		if (pg->flags & PG_ZERO) {
1154 			uvmexp.pga_zerohit++;
1155 			zeroit = 0;
1156 		} else {
1157 			uvmexp.pga_zeromiss++;
1158 			zeroit = 1;
1159 		}
1160 	}
1161 	uvm_unlock_fpageq(s);
1162 
1163 	pg->offset = off;
1164 	pg->uobject = obj;
1165 	pg->uanon = anon;
1166 	pg->flags = PG_BUSY|PG_CLEAN|PG_FAKE;
1167 	if (anon) {
1168 		anon->u.an_page = pg;
1169 		pg->pqflags = PQ_ANON;
1170 		uvmexp.anonpages++;
1171 	} else {
1172 		if (obj) {
1173 			uvm_pageinsert(pg);
1174 		}
1175 		pg->pqflags = 0;
1176 	}
1177 #if defined(UVM_PAGE_TRKOWN)
1178 	pg->owner_tag = NULL;
1179 #endif
1180 	UVM_PAGE_OWN(pg, "new alloc");
1181 
1182 	if (flags & UVM_PGA_ZERO) {
1183 		/*
1184 		 * A zero'd page is not clean.  If we got a page not already
1185 		 * zero'd, then we have to zero it ourselves.
1186 		 */
1187 		pg->flags &= ~PG_CLEAN;
1188 		if (zeroit)
1189 			pmap_zero_page(VM_PAGE_TO_PHYS(pg));
1190 	}
1191 
1192 	return(pg);
1193 
1194  fail:
1195 	uvm_unlock_fpageq(s);
1196 	return (NULL);
1197 }
1198 
1199 /*
1200  * uvm_pagerealloc: reallocate a page from one object to another
1201  *
1202  * => both objects must be locked
1203  */
1204 
1205 void
1206 uvm_pagerealloc(pg, newobj, newoff)
1207 	struct vm_page *pg;
1208 	struct uvm_object *newobj;
1209 	voff_t newoff;
1210 {
1211 	/*
1212 	 * remove it from the old object
1213 	 */
1214 
1215 	if (pg->uobject) {
1216 		uvm_pageremove(pg);
1217 	}
1218 
1219 	/*
1220 	 * put it in the new object
1221 	 */
1222 
1223 	if (newobj) {
1224 		pg->uobject = newobj;
1225 		pg->offset = newoff;
1226 		uvm_pageinsert(pg);
1227 	}
1228 }
1229 
1230 #ifdef DEBUG
1231 /*
1232  * check if page is zero-filled
1233  *
1234  *  - called with free page queue lock held.
1235  */
1236 void
1237 uvm_pagezerocheck(struct vm_page *pg)
1238 {
1239 	int *p, *ep;
1240 
1241 	KASSERT(uvm_zerocheckkva != 0);
1242 	LOCK_ASSERT(simple_lock_held(&uvm.fpageqlock));
1243 
1244 	/*
1245 	 * XXX assuming pmap_kenter_pa and pmap_kremove never call
1246 	 * uvm page allocator.
1247 	 *
1248 	 * it might be better to have "CPU-local temporary map" pmap interface.
1249 	 */
1250 	pmap_kenter_pa(uvm_zerocheckkva, VM_PAGE_TO_PHYS(pg), VM_PROT_READ);
1251 	p = (int *)uvm_zerocheckkva;
1252 	ep = (int *)((char *)p + PAGE_SIZE);
1253 	pmap_update(pmap_kernel());
1254 	while (p < ep) {
1255 		if (*p != 0)
1256 			panic("PG_ZERO page isn't zero-filled");
1257 		p++;
1258 	}
1259 	pmap_kremove(uvm_zerocheckkva, PAGE_SIZE);
1260 }
1261 #endif /* DEBUG */
1262 
1263 /*
1264  * uvm_pagefree: free page
1265  *
1266  * => erase page's identity (i.e. remove from hash/object)
1267  * => put page on free list
1268  * => caller must lock owning object (either anon or uvm_object)
1269  * => caller must lock page queues
1270  * => assumes all valid mappings of pg are gone
1271  */
1272 
1273 void
1274 uvm_pagefree(pg)
1275 	struct vm_page *pg;
1276 {
1277 	int s;
1278 	struct pglist *pgfl;
1279 	boolean_t iszero;
1280 
1281 	KASSERT((pg->flags & PG_PAGEOUT) == 0);
1282 	LOCK_ASSERT(simple_lock_held(&uvm.pageqlock) ||
1283 		    (pg->pqflags & (PQ_ACTIVE|PQ_INACTIVE)) == 0);
1284 	LOCK_ASSERT(pg->uobject == NULL ||
1285 		    simple_lock_held(&pg->uobject->vmobjlock));
1286 	LOCK_ASSERT(pg->uobject != NULL || pg->uanon == NULL ||
1287 		    simple_lock_held(&pg->uanon->an_lock));
1288 
1289 #ifdef DEBUG
1290 	if (pg->uobject == (void *)0xdeadbeef &&
1291 	    pg->uanon == (void *)0xdeadbeef) {
1292 		panic("uvm_pagefree: freeing free page %p", pg);
1293 	}
1294 #endif /* DEBUG */
1295 
1296 	/*
1297 	 * if the page is loaned, resolve the loan instead of freeing.
1298 	 */
1299 
1300 	if (pg->loan_count) {
1301 		KASSERT(pg->wire_count == 0);
1302 
1303 		/*
1304 		 * if the page is owned by an anon then we just want to
1305 		 * drop anon ownership.  the kernel will free the page when
1306 		 * it is done with it.  if the page is owned by an object,
1307 		 * remove it from the object and mark it dirty for the benefit
1308 		 * of possible anon owners.
1309 		 *
1310 		 * regardless of previous ownership, wakeup any waiters,
1311 		 * unbusy the page, and we're done.
1312 		 */
1313 
1314 		if (pg->uobject != NULL) {
1315 			uvm_pageremove(pg);
1316 			pg->flags &= ~PG_CLEAN;
1317 		} else if (pg->uanon != NULL) {
1318 			if ((pg->pqflags & PQ_ANON) == 0) {
1319 				pg->loan_count--;
1320 			} else {
1321 				pg->pqflags &= ~PQ_ANON;
1322 			}
1323 			pg->uanon = NULL;
1324 		}
1325 		if (pg->flags & PG_WANTED) {
1326 			wakeup(pg);
1327 		}
1328 		pg->flags &= ~(PG_WANTED|PG_BUSY|PG_RELEASED|PG_PAGER1);
1329 #ifdef UVM_PAGE_TRKOWN
1330 		pg->owner_tag = NULL;
1331 #endif
1332 		if (pg->loan_count) {
1333 			uvm_pagedequeue(pg);
1334 			return;
1335 		}
1336 	}
1337 
1338 	/*
1339 	 * remove page from its object or anon.
1340 	 */
1341 
1342 	if (pg->uobject != NULL) {
1343 		uvm_pageremove(pg);
1344 	} else if (pg->uanon != NULL) {
1345 		pg->uanon->u.an_page = NULL;
1346 		uvmexp.anonpages--;
1347 	}
1348 
1349 	/*
1350 	 * now remove the page from the queues.
1351 	 */
1352 
1353 	uvm_pagedequeue(pg);
1354 
1355 	/*
1356 	 * if the page was wired, unwire it now.
1357 	 */
1358 
1359 	if (pg->wire_count) {
1360 		pg->wire_count = 0;
1361 		uvmexp.wired--;
1362 	}
1363 
1364 	/*
1365 	 * and put on free queue
1366 	 */
1367 
1368 	iszero = (pg->flags & PG_ZERO);
1369 	pgfl = &uvm.page_free[uvm_page_lookup_freelist(pg)].
1370 	    pgfl_buckets[VM_PGCOLOR_BUCKET(pg)].
1371 	    pgfl_queues[iszero ? PGFL_ZEROS : PGFL_UNKNOWN];
1372 
1373 	pg->pqflags = PQ_FREE;
1374 #ifdef DEBUG
1375 	pg->uobject = (void *)0xdeadbeef;
1376 	pg->offset = 0xdeadbeef;
1377 	pg->uanon = (void *)0xdeadbeef;
1378 #endif
1379 
1380 	s = uvm_lock_fpageq();
1381 
1382 #ifdef DEBUG
1383 	if (iszero)
1384 		uvm_pagezerocheck(pg);
1385 #endif /* DEBUG */
1386 
1387 	TAILQ_INSERT_TAIL(pgfl, pg, pageq);
1388 	uvmexp.free++;
1389 	if (iszero)
1390 		uvmexp.zeropages++;
1391 
1392 	if (uvmexp.zeropages < UVM_PAGEZERO_TARGET)
1393 		uvm.page_idle_zero = vm_page_zero_enable;
1394 
1395 	uvm_unlock_fpageq(s);
1396 }
1397 
1398 /*
1399  * uvm_page_unbusy: unbusy an array of pages.
1400  *
1401  * => pages must either all belong to the same object, or all belong to anons.
1402  * => if pages are object-owned, object must be locked.
1403  * => if pages are anon-owned, anons must be locked.
1404  * => caller must lock page queues if pages may be released.
1405  */
1406 
1407 void
1408 uvm_page_unbusy(pgs, npgs)
1409 	struct vm_page **pgs;
1410 	int npgs;
1411 {
1412 	struct vm_page *pg;
1413 	int i;
1414 	UVMHIST_FUNC("uvm_page_unbusy"); UVMHIST_CALLED(ubchist);
1415 
1416 	for (i = 0; i < npgs; i++) {
1417 		pg = pgs[i];
1418 		if (pg == NULL || pg == PGO_DONTCARE) {
1419 			continue;
1420 		}
1421 		if (pg->flags & PG_WANTED) {
1422 			wakeup(pg);
1423 		}
1424 		if (pg->flags & PG_RELEASED) {
1425 			UVMHIST_LOG(ubchist, "releasing pg %p", pg,0,0,0);
1426 			pg->flags &= ~PG_RELEASED;
1427 			uvm_pagefree(pg);
1428 		} else {
1429 			UVMHIST_LOG(ubchist, "unbusying pg %p", pg,0,0,0);
1430 			pg->flags &= ~(PG_WANTED|PG_BUSY);
1431 			UVM_PAGE_OWN(pg, NULL);
1432 		}
1433 	}
1434 }
1435 
1436 #if defined(UVM_PAGE_TRKOWN)
1437 /*
1438  * uvm_page_own: set or release page ownership
1439  *
1440  * => this is a debugging function that keeps track of who sets PG_BUSY
1441  *	and where they do it.   it can be used to track down problems
1442  *	such a process setting "PG_BUSY" and never releasing it.
1443  * => page's object [if any] must be locked
1444  * => if "tag" is NULL then we are releasing page ownership
1445  */
1446 void
1447 uvm_page_own(pg, tag)
1448 	struct vm_page *pg;
1449 	char *tag;
1450 {
1451 	KASSERT((pg->flags & (PG_PAGEOUT|PG_RELEASED)) == 0);
1452 
1453 	/* gain ownership? */
1454 	if (tag) {
1455 		if (pg->owner_tag) {
1456 			printf("uvm_page_own: page %p already owned "
1457 			    "by proc %d [%s]\n", pg,
1458 			    pg->owner, pg->owner_tag);
1459 			panic("uvm_page_own");
1460 		}
1461 		pg->owner = (curproc) ? curproc->p_pid :  (pid_t) -1;
1462 		pg->owner_tag = tag;
1463 		return;
1464 	}
1465 
1466 	/* drop ownership */
1467 	if (pg->owner_tag == NULL) {
1468 		printf("uvm_page_own: dropping ownership of an non-owned "
1469 		    "page (%p)\n", pg);
1470 		panic("uvm_page_own");
1471 	}
1472 	KASSERT((pg->pqflags & (PQ_ACTIVE|PQ_INACTIVE)) ||
1473 	    (pg->uanon == NULL && pg->uobject == NULL) ||
1474 	    pg->uobject == uvm.kernel_object ||
1475 	    pg->wire_count > 0 ||
1476 	    (pg->loan_count == 1 && pg->uanon == NULL) ||
1477 	    pg->loan_count > 1);
1478 	pg->owner_tag = NULL;
1479 }
1480 #endif
1481 
1482 /*
1483  * uvm_pageidlezero: zero free pages while the system is idle.
1484  *
1485  * => try to complete one color bucket at a time, to reduce our impact
1486  *	on the CPU cache.
1487  * => we loop until we either reach the target or whichqs indicates that
1488  *	there is a process ready to run.
1489  */
1490 void
1491 uvm_pageidlezero()
1492 {
1493 	struct vm_page *pg;
1494 	struct pgfreelist *pgfl;
1495 	int free_list, s, firstbucket;
1496 	static int nextbucket;
1497 
1498 	s = uvm_lock_fpageq();
1499 	firstbucket = nextbucket;
1500 	do {
1501 		if (sched_whichqs != 0) {
1502 			uvm_unlock_fpageq(s);
1503 			return;
1504 		}
1505 		if (uvmexp.zeropages >= UVM_PAGEZERO_TARGET) {
1506 			uvm.page_idle_zero = FALSE;
1507 			uvm_unlock_fpageq(s);
1508 			return;
1509 		}
1510 		for (free_list = 0; free_list < VM_NFREELIST; free_list++) {
1511 			pgfl = &uvm.page_free[free_list];
1512 			while ((pg = TAILQ_FIRST(&pgfl->pgfl_buckets[
1513 			    nextbucket].pgfl_queues[PGFL_UNKNOWN])) != NULL) {
1514 				if (sched_whichqs != 0) {
1515 					uvm_unlock_fpageq(s);
1516 					return;
1517 				}
1518 
1519 				TAILQ_REMOVE(&pgfl->pgfl_buckets[
1520 				    nextbucket].pgfl_queues[PGFL_UNKNOWN],
1521 				    pg, pageq);
1522 				uvmexp.free--;
1523 				uvm_unlock_fpageq(s);
1524 #ifdef PMAP_PAGEIDLEZERO
1525 				if (!PMAP_PAGEIDLEZERO(VM_PAGE_TO_PHYS(pg))) {
1526 
1527 					/*
1528 					 * The machine-dependent code detected
1529 					 * some reason for us to abort zeroing
1530 					 * pages, probably because there is a
1531 					 * process now ready to run.
1532 					 */
1533 
1534 					s = uvm_lock_fpageq();
1535 					TAILQ_INSERT_HEAD(&pgfl->pgfl_buckets[
1536 					    nextbucket].pgfl_queues[
1537 					    PGFL_UNKNOWN], pg, pageq);
1538 					uvmexp.free++;
1539 					uvmexp.zeroaborts++;
1540 					uvm_unlock_fpageq(s);
1541 					return;
1542 				}
1543 #else
1544 				pmap_zero_page(VM_PAGE_TO_PHYS(pg));
1545 #endif /* PMAP_PAGEIDLEZERO */
1546 				pg->flags |= PG_ZERO;
1547 
1548 				s = uvm_lock_fpageq();
1549 				TAILQ_INSERT_HEAD(&pgfl->pgfl_buckets[
1550 				    nextbucket].pgfl_queues[PGFL_ZEROS],
1551 				    pg, pageq);
1552 				uvmexp.free++;
1553 				uvmexp.zeropages++;
1554 			}
1555 		}
1556 		nextbucket = (nextbucket + 1) & uvmexp.colormask;
1557 	} while (nextbucket != firstbucket);
1558 	uvm_unlock_fpageq(s);
1559 }
1560