xref: /netbsd-src/sys/uvm/uvm_page.c (revision 1ad9454efb13a65cd7535ccf867508cb14d9d30e)
1 /*	$NetBSD: uvm_page.c,v 1.113 2006/09/15 15:51:13 yamt 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 1.113 2006/09/15 15:51:13 yamt Exp $");
75 
76 #include "opt_uvmhist.h"
77 #include "opt_readahead.h"
78 
79 #include <sys/param.h>
80 #include <sys/systm.h>
81 #include <sys/malloc.h>
82 #include <sys/sched.h>
83 #include <sys/kernel.h>
84 #include <sys/vnode.h>
85 #include <sys/proc.h>
86 
87 #include <uvm/uvm.h>
88 #include <uvm/uvm_pdpolicy.h>
89 
90 /*
91  * global vars... XXXCDC: move to uvm. structure.
92  */
93 
94 /*
95  * physical memory config is stored in vm_physmem.
96  */
97 
98 struct vm_physseg vm_physmem[VM_PHYSSEG_MAX];	/* XXXCDC: uvm.physmem */
99 int vm_nphysseg = 0;				/* XXXCDC: uvm.nphysseg */
100 
101 /*
102  * Some supported CPUs in a given architecture don't support all
103  * of the things necessary to do idle page zero'ing efficiently.
104  * We therefore provide a way to disable it from machdep code here.
105  */
106 /*
107  * XXX disabled until we can find a way to do this without causing
108  * problems for either CPU caches or DMA latency.
109  */
110 boolean_t vm_page_zero_enable = FALSE;
111 
112 /*
113  * local variables
114  */
115 
116 /*
117  * these variables record the values returned by vm_page_bootstrap,
118  * for debugging purposes.  The implementation of uvm_pageboot_alloc
119  * and pmap_startup here also uses them internally.
120  */
121 
122 static vaddr_t      virtual_space_start;
123 static vaddr_t      virtual_space_end;
124 
125 /*
126  * we use a hash table with only one bucket during bootup.  we will
127  * later rehash (resize) the hash table once the allocator is ready.
128  * we static allocate the one bootstrap bucket below...
129  */
130 
131 static struct pglist uvm_bootbucket;
132 
133 /*
134  * we allocate an initial number of page colors in uvm_page_init(),
135  * and remember them.  We may re-color pages as cache sizes are
136  * discovered during the autoconfiguration phase.  But we can never
137  * free the initial set of buckets, since they are allocated using
138  * uvm_pageboot_alloc().
139  */
140 
141 static boolean_t have_recolored_pages /* = FALSE */;
142 
143 MALLOC_DEFINE(M_VMPAGE, "VM page", "VM page");
144 
145 #ifdef DEBUG
146 vaddr_t uvm_zerocheckkva;
147 #endif /* DEBUG */
148 
149 /*
150  * local prototypes
151  */
152 
153 static void uvm_pageinsert(struct vm_page *);
154 static void uvm_pageinsert_after(struct vm_page *, struct vm_page *);
155 static void uvm_pageremove(struct vm_page *);
156 
157 /*
158  * inline functions
159  */
160 
161 /*
162  * uvm_pageinsert: insert a page in the object and the hash table
163  * uvm_pageinsert_after: insert a page into the specified place in listq
164  *
165  * => caller must lock object
166  * => caller must lock page queues
167  * => call should have already set pg's object and offset pointers
168  *    and bumped the version counter
169  */
170 
171 inline static void
172 uvm_pageinsert_after(struct vm_page *pg, struct vm_page *where)
173 {
174 	struct pglist *buck;
175 	struct uvm_object *uobj = pg->uobject;
176 
177 	KASSERT((pg->flags & PG_TABLED) == 0);
178 	KASSERT(where == NULL || (where->flags & PG_TABLED));
179 	KASSERT(where == NULL || (where->uobject == uobj));
180 	buck = &uvm.page_hash[uvm_pagehash(uobj, pg->offset)];
181 	simple_lock(&uvm.hashlock);
182 	TAILQ_INSERT_TAIL(buck, pg, hashq);
183 	simple_unlock(&uvm.hashlock);
184 
185 	if (UVM_OBJ_IS_VNODE(uobj)) {
186 		if (uobj->uo_npages == 0) {
187 			struct vnode *vp = (struct vnode *)uobj;
188 
189 			vholdl(vp);
190 		}
191 		if (UVM_OBJ_IS_VTEXT(uobj)) {
192 			uvmexp.execpages++;
193 		} else {
194 			uvmexp.filepages++;
195 		}
196 	} else if (UVM_OBJ_IS_AOBJ(uobj)) {
197 		uvmexp.anonpages++;
198 	}
199 
200 	if (where)
201 		TAILQ_INSERT_AFTER(&uobj->memq, where, pg, listq);
202 	else
203 		TAILQ_INSERT_TAIL(&uobj->memq, pg, listq);
204 	pg->flags |= PG_TABLED;
205 	uobj->uo_npages++;
206 }
207 
208 inline static void
209 uvm_pageinsert(struct vm_page *pg)
210 {
211 
212 	uvm_pageinsert_after(pg, NULL);
213 }
214 
215 /*
216  * uvm_page_remove: remove page from object and hash
217  *
218  * => caller must lock object
219  * => caller must lock page queues
220  */
221 
222 static inline void
223 uvm_pageremove(struct vm_page *pg)
224 {
225 	struct pglist *buck;
226 	struct uvm_object *uobj = pg->uobject;
227 
228 	KASSERT(pg->flags & PG_TABLED);
229 	buck = &uvm.page_hash[uvm_pagehash(uobj, pg->offset)];
230 	simple_lock(&uvm.hashlock);
231 	TAILQ_REMOVE(buck, pg, hashq);
232 	simple_unlock(&uvm.hashlock);
233 
234 	if (UVM_OBJ_IS_VNODE(uobj)) {
235 		if (uobj->uo_npages == 1) {
236 			struct vnode *vp = (struct vnode *)uobj;
237 
238 			holdrelel(vp);
239 		}
240 		if (UVM_OBJ_IS_VTEXT(uobj)) {
241 			uvmexp.execpages--;
242 		} else {
243 			uvmexp.filepages--;
244 		}
245 	} else if (UVM_OBJ_IS_AOBJ(uobj)) {
246 		uvmexp.anonpages--;
247 	}
248 
249 	/* object should be locked */
250 	uobj->uo_npages--;
251 	TAILQ_REMOVE(&uobj->memq, pg, listq);
252 	pg->flags &= ~PG_TABLED;
253 	pg->uobject = NULL;
254 }
255 
256 static void
257 uvm_page_init_buckets(struct pgfreelist *pgfl)
258 {
259 	int color, i;
260 
261 	for (color = 0; color < uvmexp.ncolors; color++) {
262 		for (i = 0; i < PGFL_NQUEUES; i++) {
263 			TAILQ_INIT(&pgfl->pgfl_buckets[color].pgfl_queues[i]);
264 		}
265 	}
266 }
267 
268 /*
269  * uvm_page_init: init the page system.   called from uvm_init().
270  *
271  * => we return the range of kernel virtual memory in kvm_startp/kvm_endp
272  */
273 
274 void
275 uvm_page_init(vaddr_t *kvm_startp, vaddr_t *kvm_endp)
276 {
277 	vsize_t freepages, pagecount, bucketcount, n;
278 	struct pgflbucket *bucketarray;
279 	struct vm_page *pagearray;
280 	int lcv;
281 	u_int i;
282 	paddr_t paddr;
283 
284 	/*
285 	 * init the page queues and page queue locks, except the free
286 	 * list; we allocate that later (with the initial vm_page
287 	 * structures).
288 	 */
289 
290 	uvmpdpol_init();
291 	simple_lock_init(&uvm.pageqlock);
292 	simple_lock_init(&uvm.fpageqlock);
293 
294 	/*
295 	 * init the <obj,offset> => <page> hash table.  for now
296 	 * we just have one bucket (the bootstrap bucket).  later on we
297 	 * will allocate new buckets as we dynamically resize the hash table.
298 	 */
299 
300 	uvm.page_nhash = 1;			/* 1 bucket */
301 	uvm.page_hashmask = 0;			/* mask for hash function */
302 	uvm.page_hash = &uvm_bootbucket;	/* install bootstrap bucket */
303 	TAILQ_INIT(uvm.page_hash);		/* init hash table */
304 	simple_lock_init(&uvm.hashlock);	/* init hash table lock */
305 
306 	/*
307 	 * allocate vm_page structures.
308 	 */
309 
310 	/*
311 	 * sanity check:
312 	 * before calling this function the MD code is expected to register
313 	 * some free RAM with the uvm_page_physload() function.   our job
314 	 * now is to allocate vm_page structures for this memory.
315 	 */
316 
317 	if (vm_nphysseg == 0)
318 		panic("uvm_page_bootstrap: no memory pre-allocated");
319 
320 	/*
321 	 * first calculate the number of free pages...
322 	 *
323 	 * note that we use start/end rather than avail_start/avail_end.
324 	 * this allows us to allocate extra vm_page structures in case we
325 	 * want to return some memory to the pool after booting.
326 	 */
327 
328 	freepages = 0;
329 	for (lcv = 0 ; lcv < vm_nphysseg ; lcv++)
330 		freepages += (vm_physmem[lcv].end - vm_physmem[lcv].start);
331 
332 	/*
333 	 * Let MD code initialize the number of colors, or default
334 	 * to 1 color if MD code doesn't care.
335 	 */
336 	if (uvmexp.ncolors == 0)
337 		uvmexp.ncolors = 1;
338 	uvmexp.colormask = uvmexp.ncolors - 1;
339 
340 	/*
341 	 * we now know we have (PAGE_SIZE * freepages) bytes of memory we can
342 	 * use.   for each page of memory we use we need a vm_page structure.
343 	 * thus, the total number of pages we can use is the total size of
344 	 * the memory divided by the PAGE_SIZE plus the size of the vm_page
345 	 * structure.   we add one to freepages as a fudge factor to avoid
346 	 * truncation errors (since we can only allocate in terms of whole
347 	 * pages).
348 	 */
349 
350 	bucketcount = uvmexp.ncolors * VM_NFREELIST;
351 	pagecount = ((freepages + 1) << PAGE_SHIFT) /
352 	    (PAGE_SIZE + sizeof(struct vm_page));
353 
354 	bucketarray = (void *)uvm_pageboot_alloc((bucketcount *
355 	    sizeof(struct pgflbucket)) + (pagecount *
356 	    sizeof(struct vm_page)));
357 	pagearray = (struct vm_page *)(bucketarray + bucketcount);
358 
359 	for (lcv = 0; lcv < VM_NFREELIST; lcv++) {
360 		uvm.page_free[lcv].pgfl_buckets =
361 		    (bucketarray + (lcv * uvmexp.ncolors));
362 		uvm_page_init_buckets(&uvm.page_free[lcv]);
363 	}
364 	memset(pagearray, 0, pagecount * sizeof(struct vm_page));
365 
366 	/*
367 	 * init the vm_page structures and put them in the correct place.
368 	 */
369 
370 	for (lcv = 0 ; lcv < vm_nphysseg ; lcv++) {
371 		n = vm_physmem[lcv].end - vm_physmem[lcv].start;
372 
373 		/* set up page array pointers */
374 		vm_physmem[lcv].pgs = pagearray;
375 		pagearray += n;
376 		pagecount -= n;
377 		vm_physmem[lcv].lastpg = vm_physmem[lcv].pgs + (n - 1);
378 
379 		/* init and free vm_pages (we've already zeroed them) */
380 		paddr = ptoa(vm_physmem[lcv].start);
381 		for (i = 0 ; i < n ; i++, paddr += PAGE_SIZE) {
382 			vm_physmem[lcv].pgs[i].phys_addr = paddr;
383 #ifdef __HAVE_VM_PAGE_MD
384 			VM_MDPAGE_INIT(&vm_physmem[lcv].pgs[i]);
385 #endif
386 			if (atop(paddr) >= vm_physmem[lcv].avail_start &&
387 			    atop(paddr) <= vm_physmem[lcv].avail_end) {
388 				uvmexp.npages++;
389 				/* add page to free pool */
390 				uvm_pagefree(&vm_physmem[lcv].pgs[i]);
391 			}
392 		}
393 	}
394 
395 	/*
396 	 * pass up the values of virtual_space_start and
397 	 * virtual_space_end (obtained by uvm_pageboot_alloc) to the upper
398 	 * layers of the VM.
399 	 */
400 
401 	*kvm_startp = round_page(virtual_space_start);
402 	*kvm_endp = trunc_page(virtual_space_end);
403 #ifdef DEBUG
404 	/*
405 	 * steal kva for uvm_pagezerocheck().
406 	 */
407 	uvm_zerocheckkva = *kvm_startp;
408 	*kvm_startp += PAGE_SIZE;
409 #endif /* DEBUG */
410 
411 	/*
412 	 * init locks for kernel threads
413 	 */
414 
415 	simple_lock_init(&uvm.pagedaemon_lock);
416 	simple_lock_init(&uvm.aiodoned_lock);
417 
418 	/*
419 	 * init various thresholds.
420 	 */
421 
422 	uvmexp.reserve_pagedaemon = 1;
423 	uvmexp.reserve_kernel = 5;
424 
425 	/*
426 	 * determine if we should zero pages in the idle loop.
427 	 */
428 
429 	uvm.page_idle_zero = vm_page_zero_enable;
430 
431 	/*
432 	 * done!
433 	 */
434 
435 	uvm.page_init_done = TRUE;
436 }
437 
438 /*
439  * uvm_setpagesize: set the page size
440  *
441  * => sets page_shift and page_mask from uvmexp.pagesize.
442  */
443 
444 void
445 uvm_setpagesize(void)
446 {
447 
448 	/*
449 	 * If uvmexp.pagesize is 0 at this point, we expect PAGE_SIZE
450 	 * to be a constant (indicated by being a non-zero value).
451 	 */
452 	if (uvmexp.pagesize == 0) {
453 		if (PAGE_SIZE == 0)
454 			panic("uvm_setpagesize: uvmexp.pagesize not set");
455 		uvmexp.pagesize = PAGE_SIZE;
456 	}
457 	uvmexp.pagemask = uvmexp.pagesize - 1;
458 	if ((uvmexp.pagemask & uvmexp.pagesize) != 0)
459 		panic("uvm_setpagesize: page size not a power of two");
460 	for (uvmexp.pageshift = 0; ; uvmexp.pageshift++)
461 		if ((1 << uvmexp.pageshift) == uvmexp.pagesize)
462 			break;
463 }
464 
465 /*
466  * uvm_pageboot_alloc: steal memory from physmem for bootstrapping
467  */
468 
469 vaddr_t
470 uvm_pageboot_alloc(vsize_t size)
471 {
472 	static boolean_t initialized = FALSE;
473 	vaddr_t addr;
474 #if !defined(PMAP_STEAL_MEMORY)
475 	vaddr_t vaddr;
476 	paddr_t paddr;
477 #endif
478 
479 	/*
480 	 * on first call to this function, initialize ourselves.
481 	 */
482 	if (initialized == FALSE) {
483 		pmap_virtual_space(&virtual_space_start, &virtual_space_end);
484 
485 		/* round it the way we like it */
486 		virtual_space_start = round_page(virtual_space_start);
487 		virtual_space_end = trunc_page(virtual_space_end);
488 
489 		initialized = TRUE;
490 	}
491 
492 	/* round to page size */
493 	size = round_page(size);
494 
495 #if defined(PMAP_STEAL_MEMORY)
496 
497 	/*
498 	 * defer bootstrap allocation to MD code (it may want to allocate
499 	 * from a direct-mapped segment).  pmap_steal_memory should adjust
500 	 * virtual_space_start/virtual_space_end if necessary.
501 	 */
502 
503 	addr = pmap_steal_memory(size, &virtual_space_start,
504 	    &virtual_space_end);
505 
506 	return(addr);
507 
508 #else /* !PMAP_STEAL_MEMORY */
509 
510 	/*
511 	 * allocate virtual memory for this request
512 	 */
513 	if (virtual_space_start == virtual_space_end ||
514 	    (virtual_space_end - virtual_space_start) < size)
515 		panic("uvm_pageboot_alloc: out of virtual space");
516 
517 	addr = virtual_space_start;
518 
519 #ifdef PMAP_GROWKERNEL
520 	/*
521 	 * If the kernel pmap can't map the requested space,
522 	 * then allocate more resources for it.
523 	 */
524 	if (uvm_maxkaddr < (addr + size)) {
525 		uvm_maxkaddr = pmap_growkernel(addr + size);
526 		if (uvm_maxkaddr < (addr + size))
527 			panic("uvm_pageboot_alloc: pmap_growkernel() failed");
528 	}
529 #endif
530 
531 	virtual_space_start += size;
532 
533 	/*
534 	 * allocate and mapin physical pages to back new virtual pages
535 	 */
536 
537 	for (vaddr = round_page(addr) ; vaddr < addr + size ;
538 	    vaddr += PAGE_SIZE) {
539 
540 		if (!uvm_page_physget(&paddr))
541 			panic("uvm_pageboot_alloc: out of memory");
542 
543 		/*
544 		 * Note this memory is no longer managed, so using
545 		 * pmap_kenter is safe.
546 		 */
547 		pmap_kenter_pa(vaddr, paddr, VM_PROT_READ|VM_PROT_WRITE);
548 	}
549 	pmap_update(pmap_kernel());
550 	return(addr);
551 #endif	/* PMAP_STEAL_MEMORY */
552 }
553 
554 #if !defined(PMAP_STEAL_MEMORY)
555 /*
556  * uvm_page_physget: "steal" one page from the vm_physmem structure.
557  *
558  * => attempt to allocate it off the end of a segment in which the "avail"
559  *    values match the start/end values.   if we can't do that, then we
560  *    will advance both values (making them equal, and removing some
561  *    vm_page structures from the non-avail area).
562  * => return false if out of memory.
563  */
564 
565 /* subroutine: try to allocate from memory chunks on the specified freelist */
566 static boolean_t uvm_page_physget_freelist(paddr_t *, int);
567 
568 static boolean_t
569 uvm_page_physget_freelist(paddr_t *paddrp, int freelist)
570 {
571 	int lcv, x;
572 
573 	/* pass 1: try allocating from a matching end */
574 #if (VM_PHYSSEG_STRAT == VM_PSTRAT_BIGFIRST)
575 	for (lcv = vm_nphysseg - 1 ; lcv >= 0 ; lcv--)
576 #else
577 	for (lcv = 0 ; lcv < vm_nphysseg ; lcv++)
578 #endif
579 	{
580 
581 		if (uvm.page_init_done == TRUE)
582 			panic("uvm_page_physget: called _after_ bootstrap");
583 
584 		if (vm_physmem[lcv].free_list != freelist)
585 			continue;
586 
587 		/* try from front */
588 		if (vm_physmem[lcv].avail_start == vm_physmem[lcv].start &&
589 		    vm_physmem[lcv].avail_start < vm_physmem[lcv].avail_end) {
590 			*paddrp = ptoa(vm_physmem[lcv].avail_start);
591 			vm_physmem[lcv].avail_start++;
592 			vm_physmem[lcv].start++;
593 			/* nothing left?   nuke it */
594 			if (vm_physmem[lcv].avail_start ==
595 			    vm_physmem[lcv].end) {
596 				if (vm_nphysseg == 1)
597 				    panic("uvm_page_physget: out of memory!");
598 				vm_nphysseg--;
599 				for (x = lcv ; x < vm_nphysseg ; x++)
600 					/* structure copy */
601 					vm_physmem[x] = vm_physmem[x+1];
602 			}
603 			return (TRUE);
604 		}
605 
606 		/* try from rear */
607 		if (vm_physmem[lcv].avail_end == vm_physmem[lcv].end &&
608 		    vm_physmem[lcv].avail_start < vm_physmem[lcv].avail_end) {
609 			*paddrp = ptoa(vm_physmem[lcv].avail_end - 1);
610 			vm_physmem[lcv].avail_end--;
611 			vm_physmem[lcv].end--;
612 			/* nothing left?   nuke it */
613 			if (vm_physmem[lcv].avail_end ==
614 			    vm_physmem[lcv].start) {
615 				if (vm_nphysseg == 1)
616 				    panic("uvm_page_physget: out of memory!");
617 				vm_nphysseg--;
618 				for (x = lcv ; x < vm_nphysseg ; x++)
619 					/* structure copy */
620 					vm_physmem[x] = vm_physmem[x+1];
621 			}
622 			return (TRUE);
623 		}
624 	}
625 
626 	/* pass2: forget about matching ends, just allocate something */
627 #if (VM_PHYSSEG_STRAT == VM_PSTRAT_BIGFIRST)
628 	for (lcv = vm_nphysseg - 1 ; lcv >= 0 ; lcv--)
629 #else
630 	for (lcv = 0 ; lcv < vm_nphysseg ; lcv++)
631 #endif
632 	{
633 
634 		/* any room in this bank? */
635 		if (vm_physmem[lcv].avail_start >= vm_physmem[lcv].avail_end)
636 			continue;  /* nope */
637 
638 		*paddrp = ptoa(vm_physmem[lcv].avail_start);
639 		vm_physmem[lcv].avail_start++;
640 		/* truncate! */
641 		vm_physmem[lcv].start = vm_physmem[lcv].avail_start;
642 
643 		/* nothing left?   nuke it */
644 		if (vm_physmem[lcv].avail_start == vm_physmem[lcv].end) {
645 			if (vm_nphysseg == 1)
646 				panic("uvm_page_physget: out of memory!");
647 			vm_nphysseg--;
648 			for (x = lcv ; x < vm_nphysseg ; x++)
649 				/* structure copy */
650 				vm_physmem[x] = vm_physmem[x+1];
651 		}
652 		return (TRUE);
653 	}
654 
655 	return (FALSE);        /* whoops! */
656 }
657 
658 boolean_t
659 uvm_page_physget(paddr_t *paddrp)
660 {
661 	int i;
662 
663 	/* try in the order of freelist preference */
664 	for (i = 0; i < VM_NFREELIST; i++)
665 		if (uvm_page_physget_freelist(paddrp, i) == TRUE)
666 			return (TRUE);
667 	return (FALSE);
668 }
669 #endif /* PMAP_STEAL_MEMORY */
670 
671 /*
672  * uvm_page_physload: load physical memory into VM system
673  *
674  * => all args are PFs
675  * => all pages in start/end get vm_page structures
676  * => areas marked by avail_start/avail_end get added to the free page pool
677  * => we are limited to VM_PHYSSEG_MAX physical memory segments
678  */
679 
680 void
681 uvm_page_physload(paddr_t start, paddr_t end, paddr_t avail_start,
682     paddr_t avail_end, int free_list)
683 {
684 	int preload, lcv;
685 	psize_t npages;
686 	struct vm_page *pgs;
687 	struct vm_physseg *ps;
688 
689 	if (uvmexp.pagesize == 0)
690 		panic("uvm_page_physload: page size not set!");
691 	if (free_list >= VM_NFREELIST || free_list < VM_FREELIST_DEFAULT)
692 		panic("uvm_page_physload: bad free list %d", free_list);
693 	if (start >= end)
694 		panic("uvm_page_physload: start >= end");
695 
696 	/*
697 	 * do we have room?
698 	 */
699 
700 	if (vm_nphysseg == VM_PHYSSEG_MAX) {
701 		printf("uvm_page_physload: unable to load physical memory "
702 		    "segment\n");
703 		printf("\t%d segments allocated, ignoring 0x%llx -> 0x%llx\n",
704 		    VM_PHYSSEG_MAX, (long long)start, (long long)end);
705 		printf("\tincrease VM_PHYSSEG_MAX\n");
706 		return;
707 	}
708 
709 	/*
710 	 * check to see if this is a "preload" (i.e. uvm_mem_init hasn't been
711 	 * called yet, so malloc is not available).
712 	 */
713 
714 	for (lcv = 0 ; lcv < vm_nphysseg ; lcv++) {
715 		if (vm_physmem[lcv].pgs)
716 			break;
717 	}
718 	preload = (lcv == vm_nphysseg);
719 
720 	/*
721 	 * if VM is already running, attempt to malloc() vm_page structures
722 	 */
723 
724 	if (!preload) {
725 #if defined(VM_PHYSSEG_NOADD)
726 		panic("uvm_page_physload: tried to add RAM after vm_mem_init");
727 #else
728 		/* XXXCDC: need some sort of lockout for this case */
729 		paddr_t paddr;
730 		npages = end - start;  /* # of pages */
731 		pgs = malloc(sizeof(struct vm_page) * npages,
732 		    M_VMPAGE, M_NOWAIT);
733 		if (pgs == NULL) {
734 			printf("uvm_page_physload: can not malloc vm_page "
735 			    "structs for segment\n");
736 			printf("\tignoring 0x%lx -> 0x%lx\n", start, end);
737 			return;
738 		}
739 		/* zero data, init phys_addr and free_list, and free pages */
740 		memset(pgs, 0, sizeof(struct vm_page) * npages);
741 		for (lcv = 0, paddr = ptoa(start) ;
742 				 lcv < npages ; lcv++, paddr += PAGE_SIZE) {
743 			pgs[lcv].phys_addr = paddr;
744 			pgs[lcv].free_list = free_list;
745 			if (atop(paddr) >= avail_start &&
746 			    atop(paddr) <= avail_end)
747 				uvm_pagefree(&pgs[lcv]);
748 		}
749 		/* XXXCDC: incomplete: need to update uvmexp.free, what else? */
750 		/* XXXCDC: need hook to tell pmap to rebuild pv_list, etc... */
751 #endif
752 	} else {
753 		pgs = NULL;
754 		npages = 0;
755 	}
756 
757 	/*
758 	 * now insert us in the proper place in vm_physmem[]
759 	 */
760 
761 #if (VM_PHYSSEG_STRAT == VM_PSTRAT_RANDOM)
762 	/* random: put it at the end (easy!) */
763 	ps = &vm_physmem[vm_nphysseg];
764 #elif (VM_PHYSSEG_STRAT == VM_PSTRAT_BSEARCH)
765 	{
766 		int x;
767 		/* sort by address for binary search */
768 		for (lcv = 0 ; lcv < vm_nphysseg ; lcv++)
769 			if (start < vm_physmem[lcv].start)
770 				break;
771 		ps = &vm_physmem[lcv];
772 		/* move back other entries, if necessary ... */
773 		for (x = vm_nphysseg ; x > lcv ; x--)
774 			/* structure copy */
775 			vm_physmem[x] = vm_physmem[x - 1];
776 	}
777 #elif (VM_PHYSSEG_STRAT == VM_PSTRAT_BIGFIRST)
778 	{
779 		int x;
780 		/* sort by largest segment first */
781 		for (lcv = 0 ; lcv < vm_nphysseg ; lcv++)
782 			if ((end - start) >
783 			    (vm_physmem[lcv].end - vm_physmem[lcv].start))
784 				break;
785 		ps = &vm_physmem[lcv];
786 		/* move back other entries, if necessary ... */
787 		for (x = vm_nphysseg ; x > lcv ; x--)
788 			/* structure copy */
789 			vm_physmem[x] = vm_physmem[x - 1];
790 	}
791 #else
792 	panic("uvm_page_physload: unknown physseg strategy selected!");
793 #endif
794 
795 	ps->start = start;
796 	ps->end = end;
797 	ps->avail_start = avail_start;
798 	ps->avail_end = avail_end;
799 	if (preload) {
800 		ps->pgs = NULL;
801 	} else {
802 		ps->pgs = pgs;
803 		ps->lastpg = pgs + npages - 1;
804 	}
805 	ps->free_list = free_list;
806 	vm_nphysseg++;
807 
808 	if (!preload) {
809 		uvm_page_rehash();
810 		uvmpdpol_reinit();
811 	}
812 }
813 
814 /*
815  * uvm_page_rehash: reallocate hash table based on number of free pages.
816  */
817 
818 void
819 uvm_page_rehash(void)
820 {
821 	int freepages, lcv, bucketcount, oldcount;
822 	struct pglist *newbuckets, *oldbuckets;
823 	struct vm_page *pg;
824 	size_t newsize, oldsize;
825 
826 	/*
827 	 * compute number of pages that can go in the free pool
828 	 */
829 
830 	freepages = 0;
831 	for (lcv = 0 ; lcv < vm_nphysseg ; lcv++)
832 		freepages +=
833 		    (vm_physmem[lcv].avail_end - vm_physmem[lcv].avail_start);
834 
835 	/*
836 	 * compute number of buckets needed for this number of pages
837 	 */
838 
839 	bucketcount = 1;
840 	while (bucketcount < freepages)
841 		bucketcount = bucketcount * 2;
842 
843 	/*
844 	 * compute the size of the current table and new table.
845 	 */
846 
847 	oldbuckets = uvm.page_hash;
848 	oldcount = uvm.page_nhash;
849 	oldsize = round_page(sizeof(struct pglist) * oldcount);
850 	newsize = round_page(sizeof(struct pglist) * bucketcount);
851 
852 	/*
853 	 * allocate the new buckets
854 	 */
855 
856 	newbuckets = (struct pglist *) uvm_km_alloc(kernel_map, newsize,
857 	    0, UVM_KMF_WIRED);
858 	if (newbuckets == NULL) {
859 		printf("uvm_page_physrehash: WARNING: could not grow page "
860 		    "hash table\n");
861 		return;
862 	}
863 	for (lcv = 0 ; lcv < bucketcount ; lcv++)
864 		TAILQ_INIT(&newbuckets[lcv]);
865 
866 	/*
867 	 * now replace the old buckets with the new ones and rehash everything
868 	 */
869 
870 	simple_lock(&uvm.hashlock);
871 	uvm.page_hash = newbuckets;
872 	uvm.page_nhash = bucketcount;
873 	uvm.page_hashmask = bucketcount - 1;  /* power of 2 */
874 
875 	/* ... and rehash */
876 	for (lcv = 0 ; lcv < oldcount ; lcv++) {
877 		while ((pg = oldbuckets[lcv].tqh_first) != NULL) {
878 			TAILQ_REMOVE(&oldbuckets[lcv], pg, hashq);
879 			TAILQ_INSERT_TAIL(
880 			  &uvm.page_hash[uvm_pagehash(pg->uobject, pg->offset)],
881 			  pg, hashq);
882 		}
883 	}
884 	simple_unlock(&uvm.hashlock);
885 
886 	/*
887 	 * free old bucket array if is not the boot-time table
888 	 */
889 
890 	if (oldbuckets != &uvm_bootbucket)
891 		uvm_km_free(kernel_map, (vaddr_t) oldbuckets, oldsize,
892 		    UVM_KMF_WIRED);
893 }
894 
895 /*
896  * uvm_page_recolor: Recolor the pages if the new bucket count is
897  * larger than the old one.
898  */
899 
900 void
901 uvm_page_recolor(int newncolors)
902 {
903 	struct pgflbucket *bucketarray, *oldbucketarray;
904 	struct pgfreelist pgfl;
905 	struct vm_page *pg;
906 	vsize_t bucketcount;
907 	int s, lcv, color, i, ocolors;
908 
909 	if (newncolors <= uvmexp.ncolors)
910 		return;
911 
912 	if (uvm.page_init_done == FALSE) {
913 		uvmexp.ncolors = newncolors;
914 		return;
915 	}
916 
917 	bucketcount = newncolors * VM_NFREELIST;
918 	bucketarray = malloc(bucketcount * sizeof(struct pgflbucket),
919 	    M_VMPAGE, M_NOWAIT);
920 	if (bucketarray == NULL) {
921 		printf("WARNING: unable to allocate %ld page color buckets\n",
922 		    (long) bucketcount);
923 		return;
924 	}
925 
926 	s = uvm_lock_fpageq();
927 
928 	/* Make sure we should still do this. */
929 	if (newncolors <= uvmexp.ncolors) {
930 		uvm_unlock_fpageq(s);
931 		free(bucketarray, M_VMPAGE);
932 		return;
933 	}
934 
935 	oldbucketarray = uvm.page_free[0].pgfl_buckets;
936 	ocolors = uvmexp.ncolors;
937 
938 	uvmexp.ncolors = newncolors;
939 	uvmexp.colormask = uvmexp.ncolors - 1;
940 
941 	for (lcv = 0; lcv < VM_NFREELIST; lcv++) {
942 		pgfl.pgfl_buckets = (bucketarray + (lcv * newncolors));
943 		uvm_page_init_buckets(&pgfl);
944 		for (color = 0; color < ocolors; color++) {
945 			for (i = 0; i < PGFL_NQUEUES; i++) {
946 				while ((pg = TAILQ_FIRST(&uvm.page_free[
947 				    lcv].pgfl_buckets[color].pgfl_queues[i]))
948 				    != NULL) {
949 					TAILQ_REMOVE(&uvm.page_free[
950 					    lcv].pgfl_buckets[
951 					    color].pgfl_queues[i], pg, pageq);
952 					TAILQ_INSERT_TAIL(&pgfl.pgfl_buckets[
953 					    VM_PGCOLOR_BUCKET(pg)].pgfl_queues[
954 					    i], pg, pageq);
955 				}
956 			}
957 		}
958 		uvm.page_free[lcv].pgfl_buckets = pgfl.pgfl_buckets;
959 	}
960 
961 	if (have_recolored_pages) {
962 		uvm_unlock_fpageq(s);
963 		free(oldbucketarray, M_VMPAGE);
964 		return;
965 	}
966 
967 	have_recolored_pages = TRUE;
968 	uvm_unlock_fpageq(s);
969 }
970 
971 /*
972  * uvm_pagealloc_pgfl: helper routine for uvm_pagealloc_strat
973  */
974 
975 static inline struct vm_page *
976 uvm_pagealloc_pgfl(struct pgfreelist *pgfl, int try1, int try2,
977     int *trycolorp)
978 {
979 	struct pglist *freeq;
980 	struct vm_page *pg;
981 	int color, trycolor = *trycolorp;
982 
983 	color = trycolor;
984 	do {
985 		if ((pg = TAILQ_FIRST((freeq =
986 		    &pgfl->pgfl_buckets[color].pgfl_queues[try1]))) != NULL)
987 			goto gotit;
988 		if ((pg = TAILQ_FIRST((freeq =
989 		    &pgfl->pgfl_buckets[color].pgfl_queues[try2]))) != NULL)
990 			goto gotit;
991 		color = (color + 1) & uvmexp.colormask;
992 	} while (color != trycolor);
993 
994 	return (NULL);
995 
996  gotit:
997 	TAILQ_REMOVE(freeq, pg, pageq);
998 	uvmexp.free--;
999 
1000 	/* update zero'd page count */
1001 	if (pg->flags & PG_ZERO)
1002 		uvmexp.zeropages--;
1003 
1004 	if (color == trycolor)
1005 		uvmexp.colorhit++;
1006 	else {
1007 		uvmexp.colormiss++;
1008 		*trycolorp = color;
1009 	}
1010 
1011 	return (pg);
1012 }
1013 
1014 /*
1015  * uvm_pagealloc_strat: allocate vm_page from a particular free list.
1016  *
1017  * => return null if no pages free
1018  * => wake up pagedaemon if number of free pages drops below low water mark
1019  * => if obj != NULL, obj must be locked (to put in hash)
1020  * => if anon != NULL, anon must be locked (to put in anon)
1021  * => only one of obj or anon can be non-null
1022  * => caller must activate/deactivate page if it is not wired.
1023  * => free_list is ignored if strat == UVM_PGA_STRAT_NORMAL.
1024  * => policy decision: it is more important to pull a page off of the
1025  *	appropriate priority free list than it is to get a zero'd or
1026  *	unknown contents page.  This is because we live with the
1027  *	consequences of a bad free list decision for the entire
1028  *	lifetime of the page, e.g. if the page comes from memory that
1029  *	is slower to access.
1030  */
1031 
1032 struct vm_page *
1033 uvm_pagealloc_strat(struct uvm_object *obj, voff_t off, struct vm_anon *anon,
1034     int flags, int strat, int free_list)
1035 {
1036 	int lcv, try1, try2, s, zeroit = 0, color;
1037 	struct vm_page *pg;
1038 	boolean_t use_reserve;
1039 
1040 	KASSERT(obj == NULL || anon == NULL);
1041 	KASSERT(anon == NULL || off == 0);
1042 	KASSERT(off == trunc_page(off));
1043 	LOCK_ASSERT(obj == NULL || simple_lock_held(&obj->vmobjlock));
1044 	LOCK_ASSERT(anon == NULL || simple_lock_held(&anon->an_lock));
1045 
1046 	s = uvm_lock_fpageq();
1047 
1048 	/*
1049 	 * This implements a global round-robin page coloring
1050 	 * algorithm.
1051 	 *
1052 	 * XXXJRT: Should we make the `nextcolor' per-CPU?
1053 	 * XXXJRT: What about virtually-indexed caches?
1054 	 */
1055 
1056 	color = uvm.page_free_nextcolor;
1057 
1058 	/*
1059 	 * check to see if we need to generate some free pages waking
1060 	 * the pagedaemon.
1061 	 */
1062 
1063 	uvm_kick_pdaemon();
1064 
1065 	/*
1066 	 * fail if any of these conditions is true:
1067 	 * [1]  there really are no free pages, or
1068 	 * [2]  only kernel "reserved" pages remain and
1069 	 *        the page isn't being allocated to a kernel object.
1070 	 * [3]  only pagedaemon "reserved" pages remain and
1071 	 *        the requestor isn't the pagedaemon.
1072 	 */
1073 
1074 	use_reserve = (flags & UVM_PGA_USERESERVE) ||
1075 		(obj && UVM_OBJ_IS_KERN_OBJECT(obj));
1076 	if ((uvmexp.free <= uvmexp.reserve_kernel && !use_reserve) ||
1077 	    (uvmexp.free <= uvmexp.reserve_pagedaemon &&
1078 	     !(use_reserve && curproc == uvm.pagedaemon_proc)))
1079 		goto fail;
1080 
1081 #if PGFL_NQUEUES != 2
1082 #error uvm_pagealloc_strat needs to be updated
1083 #endif
1084 
1085 	/*
1086 	 * If we want a zero'd page, try the ZEROS queue first, otherwise
1087 	 * we try the UNKNOWN queue first.
1088 	 */
1089 	if (flags & UVM_PGA_ZERO) {
1090 		try1 = PGFL_ZEROS;
1091 		try2 = PGFL_UNKNOWN;
1092 	} else {
1093 		try1 = PGFL_UNKNOWN;
1094 		try2 = PGFL_ZEROS;
1095 	}
1096 
1097  again:
1098 	switch (strat) {
1099 	case UVM_PGA_STRAT_NORMAL:
1100 		/* Check all freelists in descending priority order. */
1101 		for (lcv = 0; lcv < VM_NFREELIST; lcv++) {
1102 			pg = uvm_pagealloc_pgfl(&uvm.page_free[lcv],
1103 			    try1, try2, &color);
1104 			if (pg != NULL)
1105 				goto gotit;
1106 		}
1107 
1108 		/* No pages free! */
1109 		goto fail;
1110 
1111 	case UVM_PGA_STRAT_ONLY:
1112 	case UVM_PGA_STRAT_FALLBACK:
1113 		/* Attempt to allocate from the specified free list. */
1114 		KASSERT(free_list >= 0 && free_list < VM_NFREELIST);
1115 		pg = uvm_pagealloc_pgfl(&uvm.page_free[free_list],
1116 		    try1, try2, &color);
1117 		if (pg != NULL)
1118 			goto gotit;
1119 
1120 		/* Fall back, if possible. */
1121 		if (strat == UVM_PGA_STRAT_FALLBACK) {
1122 			strat = UVM_PGA_STRAT_NORMAL;
1123 			goto again;
1124 		}
1125 
1126 		/* No pages free! */
1127 		goto fail;
1128 
1129 	default:
1130 		panic("uvm_pagealloc_strat: bad strat %d", strat);
1131 		/* NOTREACHED */
1132 	}
1133 
1134  gotit:
1135 	/*
1136 	 * We now know which color we actually allocated from; set
1137 	 * the next color accordingly.
1138 	 */
1139 
1140 	uvm.page_free_nextcolor = (color + 1) & uvmexp.colormask;
1141 
1142 	/*
1143 	 * update allocation statistics and remember if we have to
1144 	 * zero the page
1145 	 */
1146 
1147 	if (flags & UVM_PGA_ZERO) {
1148 		if (pg->flags & PG_ZERO) {
1149 			uvmexp.pga_zerohit++;
1150 			zeroit = 0;
1151 		} else {
1152 			uvmexp.pga_zeromiss++;
1153 			zeroit = 1;
1154 		}
1155 	}
1156 	uvm_unlock_fpageq(s);
1157 
1158 	pg->offset = off;
1159 	pg->uobject = obj;
1160 	pg->uanon = anon;
1161 	pg->flags = PG_BUSY|PG_CLEAN|PG_FAKE;
1162 	if (anon) {
1163 		anon->an_page = pg;
1164 		pg->pqflags = PQ_ANON;
1165 		uvmexp.anonpages++;
1166 	} else {
1167 		if (obj) {
1168 			uvm_pageinsert(pg);
1169 		}
1170 		pg->pqflags = 0;
1171 	}
1172 #if defined(UVM_PAGE_TRKOWN)
1173 	pg->owner_tag = NULL;
1174 #endif
1175 	UVM_PAGE_OWN(pg, "new alloc");
1176 
1177 	if (flags & UVM_PGA_ZERO) {
1178 		/*
1179 		 * A zero'd page is not clean.  If we got a page not already
1180 		 * zero'd, then we have to zero it ourselves.
1181 		 */
1182 		pg->flags &= ~PG_CLEAN;
1183 		if (zeroit)
1184 			pmap_zero_page(VM_PAGE_TO_PHYS(pg));
1185 	}
1186 
1187 	return(pg);
1188 
1189  fail:
1190 	uvm_unlock_fpageq(s);
1191 	return (NULL);
1192 }
1193 
1194 /*
1195  * uvm_pagereplace: replace a page with another
1196  *
1197  * => object must be locked
1198  */
1199 
1200 void
1201 uvm_pagereplace(struct vm_page *oldpg, struct vm_page *newpg)
1202 {
1203 
1204 	KASSERT((oldpg->flags & PG_TABLED) != 0);
1205 	KASSERT(oldpg->uobject != NULL);
1206 	KASSERT((newpg->flags & PG_TABLED) == 0);
1207 	KASSERT(newpg->uobject == NULL);
1208 	LOCK_ASSERT(simple_lock_held(&oldpg->uobject->vmobjlock));
1209 
1210 	newpg->uobject = oldpg->uobject;
1211 	newpg->offset = oldpg->offset;
1212 
1213 	uvm_pageinsert_after(newpg, oldpg);
1214 	uvm_pageremove(oldpg);
1215 }
1216 
1217 /*
1218  * uvm_pagerealloc: reallocate a page from one object to another
1219  *
1220  * => both objects must be locked
1221  */
1222 
1223 void
1224 uvm_pagerealloc(struct vm_page *pg, struct uvm_object *newobj, voff_t newoff)
1225 {
1226 	/*
1227 	 * remove it from the old object
1228 	 */
1229 
1230 	if (pg->uobject) {
1231 		uvm_pageremove(pg);
1232 	}
1233 
1234 	/*
1235 	 * put it in the new object
1236 	 */
1237 
1238 	if (newobj) {
1239 		pg->uobject = newobj;
1240 		pg->offset = newoff;
1241 		uvm_pageinsert(pg);
1242 	}
1243 }
1244 
1245 #ifdef DEBUG
1246 /*
1247  * check if page is zero-filled
1248  *
1249  *  - called with free page queue lock held.
1250  */
1251 void
1252 uvm_pagezerocheck(struct vm_page *pg)
1253 {
1254 	int *p, *ep;
1255 
1256 	KASSERT(uvm_zerocheckkva != 0);
1257 	LOCK_ASSERT(simple_lock_held(&uvm.fpageqlock));
1258 
1259 	/*
1260 	 * XXX assuming pmap_kenter_pa and pmap_kremove never call
1261 	 * uvm page allocator.
1262 	 *
1263 	 * it might be better to have "CPU-local temporary map" pmap interface.
1264 	 */
1265 	pmap_kenter_pa(uvm_zerocheckkva, VM_PAGE_TO_PHYS(pg), VM_PROT_READ);
1266 	p = (int *)uvm_zerocheckkva;
1267 	ep = (int *)((char *)p + PAGE_SIZE);
1268 	pmap_update(pmap_kernel());
1269 	while (p < ep) {
1270 		if (*p != 0)
1271 			panic("PG_ZERO page isn't zero-filled");
1272 		p++;
1273 	}
1274 	pmap_kremove(uvm_zerocheckkva, PAGE_SIZE);
1275 }
1276 #endif /* DEBUG */
1277 
1278 /*
1279  * uvm_pagefree: free page
1280  *
1281  * => erase page's identity (i.e. remove from hash/object)
1282  * => put page on free list
1283  * => caller must lock owning object (either anon or uvm_object)
1284  * => caller must lock page queues
1285  * => assumes all valid mappings of pg are gone
1286  */
1287 
1288 void
1289 uvm_pagefree(struct vm_page *pg)
1290 {
1291 	int s;
1292 	struct pglist *pgfl;
1293 	boolean_t iszero;
1294 
1295 	KASSERT((pg->flags & PG_PAGEOUT) == 0);
1296 	LOCK_ASSERT(simple_lock_held(&uvm.pageqlock) ||
1297 		    !uvmpdpol_pageisqueued_p(pg));
1298 	LOCK_ASSERT(pg->uobject == NULL ||
1299 		    simple_lock_held(&pg->uobject->vmobjlock));
1300 	LOCK_ASSERT(pg->uobject != NULL || pg->uanon == NULL ||
1301 		    simple_lock_held(&pg->uanon->an_lock));
1302 
1303 #ifdef DEBUG
1304 	if (pg->uobject == (void *)0xdeadbeef &&
1305 	    pg->uanon == (void *)0xdeadbeef) {
1306 		panic("uvm_pagefree: freeing free page %p", pg);
1307 	}
1308 #endif /* DEBUG */
1309 
1310 	/*
1311 	 * if the page is loaned, resolve the loan instead of freeing.
1312 	 */
1313 
1314 	if (pg->loan_count) {
1315 		KASSERT(pg->wire_count == 0);
1316 
1317 		/*
1318 		 * if the page is owned by an anon then we just want to
1319 		 * drop anon ownership.  the kernel will free the page when
1320 		 * it is done with it.  if the page is owned by an object,
1321 		 * remove it from the object and mark it dirty for the benefit
1322 		 * of possible anon owners.
1323 		 *
1324 		 * regardless of previous ownership, wakeup any waiters,
1325 		 * unbusy the page, and we're done.
1326 		 */
1327 
1328 		if (pg->uobject != NULL) {
1329 			uvm_pageremove(pg);
1330 			pg->flags &= ~PG_CLEAN;
1331 		} else if (pg->uanon != NULL) {
1332 			if ((pg->pqflags & PQ_ANON) == 0) {
1333 				pg->loan_count--;
1334 			} else {
1335 				pg->pqflags &= ~PQ_ANON;
1336 				uvmexp.anonpages--;
1337 			}
1338 			pg->uanon->an_page = NULL;
1339 			pg->uanon = NULL;
1340 		}
1341 		if (pg->flags & PG_WANTED) {
1342 			wakeup(pg);
1343 		}
1344 		pg->flags &= ~(PG_WANTED|PG_BUSY|PG_RELEASED|PG_PAGER1);
1345 #ifdef UVM_PAGE_TRKOWN
1346 		pg->owner_tag = NULL;
1347 #endif
1348 		if (pg->loan_count) {
1349 			uvm_pagedequeue(pg);
1350 			return;
1351 		}
1352 	}
1353 
1354 	/*
1355 	 * remove page from its object or anon.
1356 	 */
1357 
1358 	if (pg->uobject != NULL) {
1359 		uvm_pageremove(pg);
1360 	} else if (pg->uanon != NULL) {
1361 		pg->uanon->an_page = NULL;
1362 		uvmexp.anonpages--;
1363 	}
1364 
1365 	/*
1366 	 * now remove the page from the queues.
1367 	 */
1368 
1369 	uvm_pagedequeue(pg);
1370 
1371 	/*
1372 	 * if the page was wired, unwire it now.
1373 	 */
1374 
1375 	if (pg->wire_count) {
1376 		pg->wire_count = 0;
1377 		uvmexp.wired--;
1378 	}
1379 
1380 	/*
1381 	 * and put on free queue
1382 	 */
1383 
1384 	iszero = (pg->flags & PG_ZERO);
1385 	pgfl = &uvm.page_free[uvm_page_lookup_freelist(pg)].
1386 	    pgfl_buckets[VM_PGCOLOR_BUCKET(pg)].
1387 	    pgfl_queues[iszero ? PGFL_ZEROS : PGFL_UNKNOWN];
1388 
1389 	pg->pqflags = PQ_FREE;
1390 #ifdef DEBUG
1391 	pg->uobject = (void *)0xdeadbeef;
1392 	pg->offset = 0xdeadbeef;
1393 	pg->uanon = (void *)0xdeadbeef;
1394 #endif
1395 
1396 	s = uvm_lock_fpageq();
1397 
1398 #ifdef DEBUG
1399 	if (iszero)
1400 		uvm_pagezerocheck(pg);
1401 #endif /* DEBUG */
1402 
1403 	TAILQ_INSERT_HEAD(pgfl, pg, pageq);
1404 	uvmexp.free++;
1405 	if (iszero)
1406 		uvmexp.zeropages++;
1407 
1408 	if (uvmexp.zeropages < UVM_PAGEZERO_TARGET)
1409 		uvm.page_idle_zero = vm_page_zero_enable;
1410 
1411 	uvm_unlock_fpageq(s);
1412 }
1413 
1414 /*
1415  * uvm_page_unbusy: unbusy an array of pages.
1416  *
1417  * => pages must either all belong to the same object, or all belong to anons.
1418  * => if pages are object-owned, object must be locked.
1419  * => if pages are anon-owned, anons must be locked.
1420  * => caller must lock page queues if pages may be released.
1421  * => caller must make sure that anon-owned pages are not PG_RELEASED.
1422  */
1423 
1424 void
1425 uvm_page_unbusy(struct vm_page **pgs, int npgs)
1426 {
1427 	struct vm_page *pg;
1428 	int i;
1429 	UVMHIST_FUNC("uvm_page_unbusy"); UVMHIST_CALLED(ubchist);
1430 
1431 	for (i = 0; i < npgs; i++) {
1432 		pg = pgs[i];
1433 		if (pg == NULL || pg == PGO_DONTCARE) {
1434 			continue;
1435 		}
1436 
1437 		LOCK_ASSERT(pg->uobject == NULL ||
1438 		    simple_lock_held(&pg->uobject->vmobjlock));
1439 		LOCK_ASSERT(pg->uobject != NULL ||
1440 		    (pg->uanon != NULL &&
1441 		    simple_lock_held(&pg->uanon->an_lock)));
1442 
1443 		KASSERT(pg->flags & PG_BUSY);
1444 		KASSERT((pg->flags & PG_PAGEOUT) == 0);
1445 		if (pg->flags & PG_WANTED) {
1446 			wakeup(pg);
1447 		}
1448 		if (pg->flags & PG_RELEASED) {
1449 			UVMHIST_LOG(ubchist, "releasing pg %p", pg,0,0,0);
1450 			KASSERT(pg->uobject != NULL ||
1451 			    (pg->uanon != NULL && pg->uanon->an_ref > 0));
1452 			pg->flags &= ~PG_RELEASED;
1453 			uvm_pagefree(pg);
1454 		} else {
1455 			UVMHIST_LOG(ubchist, "unbusying pg %p", pg,0,0,0);
1456 			pg->flags &= ~(PG_WANTED|PG_BUSY);
1457 			UVM_PAGE_OWN(pg, NULL);
1458 		}
1459 	}
1460 }
1461 
1462 #if defined(UVM_PAGE_TRKOWN)
1463 /*
1464  * uvm_page_own: set or release page ownership
1465  *
1466  * => this is a debugging function that keeps track of who sets PG_BUSY
1467  *	and where they do it.   it can be used to track down problems
1468  *	such a process setting "PG_BUSY" and never releasing it.
1469  * => page's object [if any] must be locked
1470  * => if "tag" is NULL then we are releasing page ownership
1471  */
1472 void
1473 uvm_page_own(struct vm_page *pg, const char *tag)
1474 {
1475 	struct uvm_object *uobj;
1476 	struct vm_anon *anon;
1477 
1478 	KASSERT((pg->flags & (PG_PAGEOUT|PG_RELEASED)) == 0);
1479 
1480 	uobj = pg->uobject;
1481 	anon = pg->uanon;
1482 	if (uobj != NULL) {
1483 		LOCK_ASSERT(simple_lock_held(&uobj->vmobjlock));
1484 	} else if (anon != NULL) {
1485 		LOCK_ASSERT(simple_lock_held(&anon->an_lock));
1486 	}
1487 
1488 	KASSERT((pg->flags & PG_WANTED) == 0);
1489 
1490 	/* gain ownership? */
1491 	if (tag) {
1492 		KASSERT((pg->flags & PG_BUSY) != 0);
1493 		if (pg->owner_tag) {
1494 			printf("uvm_page_own: page %p already owned "
1495 			    "by proc %d [%s]\n", pg,
1496 			    pg->owner, pg->owner_tag);
1497 			panic("uvm_page_own");
1498 		}
1499 		pg->owner = (curproc) ? curproc->p_pid :  (pid_t) -1;
1500 		pg->owner_tag = tag;
1501 		return;
1502 	}
1503 
1504 	/* drop ownership */
1505 	KASSERT((pg->flags & PG_BUSY) == 0);
1506 	if (pg->owner_tag == NULL) {
1507 		printf("uvm_page_own: dropping ownership of an non-owned "
1508 		    "page (%p)\n", pg);
1509 		panic("uvm_page_own");
1510 	}
1511 	KASSERT(uvmpdpol_pageisqueued_p(pg) ||
1512 	    (pg->uanon == NULL && pg->uobject == NULL) ||
1513 	    pg->uobject == uvm.kernel_object ||
1514 	    pg->wire_count > 0 ||
1515 	    (pg->loan_count == 1 && pg->uanon == NULL) ||
1516 	    pg->loan_count > 1);
1517 	pg->owner_tag = NULL;
1518 }
1519 #endif
1520 
1521 /*
1522  * uvm_pageidlezero: zero free pages while the system is idle.
1523  *
1524  * => try to complete one color bucket at a time, to reduce our impact
1525  *	on the CPU cache.
1526  * => we loop until we either reach the target or whichqs indicates that
1527  *	there is a process ready to run.
1528  */
1529 void
1530 uvm_pageidlezero(void)
1531 {
1532 	struct vm_page *pg;
1533 	struct pgfreelist *pgfl;
1534 	int free_list, s, firstbucket;
1535 	static int nextbucket;
1536 
1537 	KERNEL_LOCK(LK_EXCLUSIVE | LK_CANRECURSE);
1538 	s = uvm_lock_fpageq();
1539 	firstbucket = nextbucket;
1540 	do {
1541 		if (sched_whichqs != 0)
1542 			goto quit;
1543 		if (uvmexp.zeropages >= UVM_PAGEZERO_TARGET) {
1544 			uvm.page_idle_zero = FALSE;
1545 			goto quit;
1546 		}
1547 		for (free_list = 0; free_list < VM_NFREELIST; free_list++) {
1548 			pgfl = &uvm.page_free[free_list];
1549 			while ((pg = TAILQ_FIRST(&pgfl->pgfl_buckets[
1550 			    nextbucket].pgfl_queues[PGFL_UNKNOWN])) != NULL) {
1551 				if (sched_whichqs != 0)
1552 					goto quit;
1553 
1554 				TAILQ_REMOVE(&pgfl->pgfl_buckets[
1555 				    nextbucket].pgfl_queues[PGFL_UNKNOWN],
1556 				    pg, pageq);
1557 				uvmexp.free--;
1558 				uvm_unlock_fpageq(s);
1559 				KERNEL_UNLOCK();
1560 #ifdef PMAP_PAGEIDLEZERO
1561 				if (!PMAP_PAGEIDLEZERO(VM_PAGE_TO_PHYS(pg))) {
1562 
1563 					/*
1564 					 * The machine-dependent code detected
1565 					 * some reason for us to abort zeroing
1566 					 * pages, probably because there is a
1567 					 * process now ready to run.
1568 					 */
1569 
1570 					KERNEL_LOCK(
1571 					    LK_EXCLUSIVE | LK_CANRECURSE);
1572 					s = uvm_lock_fpageq();
1573 					TAILQ_INSERT_HEAD(&pgfl->pgfl_buckets[
1574 					    nextbucket].pgfl_queues[
1575 					    PGFL_UNKNOWN], pg, pageq);
1576 					uvmexp.free++;
1577 					uvmexp.zeroaborts++;
1578 					goto quit;
1579 				}
1580 #else
1581 				pmap_zero_page(VM_PAGE_TO_PHYS(pg));
1582 #endif /* PMAP_PAGEIDLEZERO */
1583 				pg->flags |= PG_ZERO;
1584 
1585 				KERNEL_LOCK(LK_EXCLUSIVE | LK_CANRECURSE);
1586 				s = uvm_lock_fpageq();
1587 				TAILQ_INSERT_HEAD(&pgfl->pgfl_buckets[
1588 				    nextbucket].pgfl_queues[PGFL_ZEROS],
1589 				    pg, pageq);
1590 				uvmexp.free++;
1591 				uvmexp.zeropages++;
1592 			}
1593 		}
1594 		nextbucket = (nextbucket + 1) & uvmexp.colormask;
1595 	} while (nextbucket != firstbucket);
1596 quit:
1597 	uvm_unlock_fpageq(s);
1598 	KERNEL_UNLOCK();
1599 }
1600 
1601 /*
1602  * uvm_lock_fpageq: lock the free page queue
1603  *
1604  * => free page queue can be accessed in interrupt context, so this
1605  *	blocks all interrupts that can cause memory allocation, and
1606  *	returns the previous interrupt level.
1607  */
1608 
1609 int
1610 uvm_lock_fpageq(void)
1611 {
1612 	int s;
1613 
1614 	s = splvm();
1615 	simple_lock(&uvm.fpageqlock);
1616 	return (s);
1617 }
1618 
1619 /*
1620  * uvm_unlock_fpageq: unlock the free page queue
1621  *
1622  * => caller must supply interrupt level returned by uvm_lock_fpageq()
1623  *	so that it may be restored.
1624  */
1625 
1626 void
1627 uvm_unlock_fpageq(int s)
1628 {
1629 
1630 	simple_unlock(&uvm.fpageqlock);
1631 	splx(s);
1632 }
1633 
1634 /*
1635  * uvm_pagelookup: look up a page
1636  *
1637  * => caller should lock object to keep someone from pulling the page
1638  *	out from under it
1639  */
1640 
1641 struct vm_page *
1642 uvm_pagelookup(struct uvm_object *obj, voff_t off)
1643 {
1644 	struct vm_page *pg;
1645 	struct pglist *buck;
1646 
1647 	buck = &uvm.page_hash[uvm_pagehash(obj,off)];
1648 	simple_lock(&uvm.hashlock);
1649 	TAILQ_FOREACH(pg, buck, hashq) {
1650 		if (pg->uobject == obj && pg->offset == off) {
1651 			break;
1652 		}
1653 	}
1654 	simple_unlock(&uvm.hashlock);
1655 	KASSERT(pg == NULL || obj->uo_npages != 0);
1656 	KASSERT(pg == NULL || (pg->flags & (PG_RELEASED|PG_PAGEOUT)) == 0 ||
1657 		(pg->flags & PG_BUSY) != 0);
1658 	return(pg);
1659 }
1660 
1661 /*
1662  * uvm_pagewire: wire the page, thus removing it from the daemon's grasp
1663  *
1664  * => caller must lock page queues
1665  */
1666 
1667 void
1668 uvm_pagewire(struct vm_page *pg)
1669 {
1670 	UVM_LOCK_ASSERT_PAGEQ();
1671 #if defined(READAHEAD_STATS)
1672 	if ((pg->pqflags & PQ_READAHEAD) != 0) {
1673 		uvm_ra_hit.ev_count++;
1674 		pg->pqflags &= ~PQ_READAHEAD;
1675 	}
1676 #endif /* defined(READAHEAD_STATS) */
1677 	if (pg->wire_count == 0) {
1678 		uvm_pagedequeue(pg);
1679 		uvmexp.wired++;
1680 	}
1681 	pg->wire_count++;
1682 }
1683 
1684 /*
1685  * uvm_pageunwire: unwire the page.
1686  *
1687  * => activate if wire count goes to zero.
1688  * => caller must lock page queues
1689  */
1690 
1691 void
1692 uvm_pageunwire(struct vm_page *pg)
1693 {
1694 	UVM_LOCK_ASSERT_PAGEQ();
1695 	pg->wire_count--;
1696 	if (pg->wire_count == 0) {
1697 		uvm_pageactivate(pg);
1698 		uvmexp.wired--;
1699 	}
1700 }
1701 
1702 /*
1703  * uvm_pagedeactivate: deactivate page
1704  *
1705  * => caller must lock page queues
1706  * => caller must check to make sure page is not wired
1707  * => object that page belongs to must be locked (so we can adjust pg->flags)
1708  * => caller must clear the reference on the page before calling
1709  */
1710 
1711 void
1712 uvm_pagedeactivate(struct vm_page *pg)
1713 {
1714 
1715 	UVM_LOCK_ASSERT_PAGEQ();
1716 	KASSERT(pg->wire_count != 0 || uvmpdpol_pageisqueued_p(pg));
1717 	uvmpdpol_pagedeactivate(pg);
1718 }
1719 
1720 /*
1721  * uvm_pageactivate: activate page
1722  *
1723  * => caller must lock page queues
1724  */
1725 
1726 void
1727 uvm_pageactivate(struct vm_page *pg)
1728 {
1729 
1730 	UVM_LOCK_ASSERT_PAGEQ();
1731 #if defined(READAHEAD_STATS)
1732 	if ((pg->pqflags & PQ_READAHEAD) != 0) {
1733 		uvm_ra_hit.ev_count++;
1734 		pg->pqflags &= ~PQ_READAHEAD;
1735 	}
1736 #endif /* defined(READAHEAD_STATS) */
1737 	if (pg->wire_count != 0) {
1738 		return;
1739 	}
1740 	uvmpdpol_pageactivate(pg);
1741 }
1742 
1743 /*
1744  * uvm_pagedequeue: remove a page from any paging queue
1745  */
1746 
1747 void
1748 uvm_pagedequeue(struct vm_page *pg)
1749 {
1750 
1751 #if defined(LOCKDEBUG)
1752 	if (uvmpdpol_pageisqueued_p(pg)) {
1753 		UVM_LOCK_ASSERT_PAGEQ();
1754 	}
1755 #endif /* defined(LOCKDEBUG) */
1756 	uvmpdpol_pagedequeue(pg);
1757 }
1758 
1759 /*
1760  * uvm_pageenqueue: add a page to a paging queue without activating.
1761  * used where a page is not really demanded (yet).  eg. read-ahead
1762  */
1763 
1764 void
1765 uvm_pageenqueue(struct vm_page *pg)
1766 {
1767 
1768 	UVM_LOCK_ASSERT_PAGEQ();
1769 	if (pg->wire_count != 0) {
1770 		return;
1771 	}
1772 	uvmpdpol_pageenqueue(pg);
1773 }
1774 
1775 /*
1776  * uvm_pagezero: zero fill a page
1777  *
1778  * => if page is part of an object then the object should be locked
1779  *	to protect pg->flags.
1780  */
1781 
1782 void
1783 uvm_pagezero(struct vm_page *pg)
1784 {
1785 	pg->flags &= ~PG_CLEAN;
1786 	pmap_zero_page(VM_PAGE_TO_PHYS(pg));
1787 }
1788 
1789 /*
1790  * uvm_pagecopy: copy a page
1791  *
1792  * => if page is part of an object then the object should be locked
1793  *	to protect pg->flags.
1794  */
1795 
1796 void
1797 uvm_pagecopy(struct vm_page *src, struct vm_page *dst)
1798 {
1799 
1800 	dst->flags &= ~PG_CLEAN;
1801 	pmap_copy_page(VM_PAGE_TO_PHYS(src), VM_PAGE_TO_PHYS(dst));
1802 }
1803 
1804 /*
1805  * uvm_page_lookup_freelist: look up the free list for the specified page
1806  */
1807 
1808 int
1809 uvm_page_lookup_freelist(struct vm_page *pg)
1810 {
1811 	int lcv;
1812 
1813 	lcv = vm_physseg_find(atop(VM_PAGE_TO_PHYS(pg)), NULL);
1814 	KASSERT(lcv != -1);
1815 	return (vm_physmem[lcv].free_list);
1816 }
1817