xref: /netbsd-src/sys/uvm/uvm_page.c (revision 8b0f9554ff8762542c4defc4f70e1eb76fb508fa)
1 /*	$NetBSD: uvm_page.c,v 1.126 2007/11/29 18:07:11 ad 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 1.126 2007/11/29 18:07:11 ad 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 #include <sys/atomic.h>
87 
88 #include <uvm/uvm.h>
89 #include <uvm/uvm_pdpolicy.h>
90 
91 /*
92  * global vars... XXXCDC: move to uvm. structure.
93  */
94 
95 /*
96  * physical memory config is stored in vm_physmem.
97  */
98 
99 struct vm_physseg vm_physmem[VM_PHYSSEG_MAX];	/* XXXCDC: uvm.physmem */
100 int vm_nphysseg = 0;				/* XXXCDC: uvm.nphysseg */
101 
102 /*
103  * Some supported CPUs in a given architecture don't support all
104  * of the things necessary to do idle page zero'ing efficiently.
105  * We therefore provide a way to disable it from machdep code here.
106  */
107 /*
108  * XXX disabled until we can find a way to do this without causing
109  * problems for either CPU caches or DMA latency.
110  */
111 bool vm_page_zero_enable = false;
112 
113 /*
114  * local variables
115  */
116 
117 /*
118  * these variables record the values returned by vm_page_bootstrap,
119  * for debugging purposes.  The implementation of uvm_pageboot_alloc
120  * and pmap_startup here also uses them internally.
121  */
122 
123 static vaddr_t      virtual_space_start;
124 static vaddr_t      virtual_space_end;
125 
126 /*
127  * we use a hash table with only one bucket during bootup.  we will
128  * later rehash (resize) the hash table once the allocator is ready.
129  * we static allocate the one bootstrap bucket below...
130  */
131 
132 static struct pglist uvm_bootbucket;
133 
134 /*
135  * we allocate an initial number of page colors in uvm_page_init(),
136  * and remember them.  We may re-color pages as cache sizes are
137  * discovered during the autoconfiguration phase.  But we can never
138  * free the initial set of buckets, since they are allocated using
139  * uvm_pageboot_alloc().
140  */
141 
142 static bool have_recolored_pages /* = false */;
143 
144 MALLOC_DEFINE(M_VMPAGE, "VM page", "VM page");
145 
146 #ifdef DEBUG
147 vaddr_t uvm_zerocheckkva;
148 #endif /* DEBUG */
149 
150 /*
151  * locks on the hash table.  allocated in 32 byte chunks to try
152  * and reduce cache traffic between CPUs.
153  */
154 
155 #define	UVM_HASHLOCK_CNT	32
156 #define	uvm_hashlock(hash)	\
157     (&uvm_hashlocks[(hash) & (UVM_HASHLOCK_CNT - 1)].lock)
158 
159 static union {
160 	kmutex_t	lock;
161 	uint8_t		pad[32];
162 } uvm_hashlocks[UVM_HASHLOCK_CNT] __aligned(32);
163 
164 /*
165  * local prototypes
166  */
167 
168 static void uvm_pageinsert(struct vm_page *);
169 static void uvm_pageinsert_after(struct vm_page *, struct vm_page *);
170 static void uvm_pageremove(struct vm_page *);
171 
172 /*
173  * inline functions
174  */
175 
176 /*
177  * uvm_pageinsert: insert a page in the object and the hash table
178  * uvm_pageinsert_after: insert a page into the specified place in listq
179  *
180  * => caller must lock object
181  * => caller must lock page queues
182  * => call should have already set pg's object and offset pointers
183  *    and bumped the version counter
184  */
185 
186 inline static void
187 uvm_pageinsert_after(struct vm_page *pg, struct vm_page *where)
188 {
189 	struct pglist *buck;
190 	struct uvm_object *uobj = pg->uobject;
191 	kmutex_t *lock;
192 	u_int hash;
193 
194 	LOCK_ASSERT(simple_lock_held(&uobj->vmobjlock));
195 	KASSERT((pg->flags & PG_TABLED) == 0);
196 	KASSERT(where == NULL || (where->flags & PG_TABLED));
197 	KASSERT(where == NULL || (where->uobject == uobj));
198 
199 	hash = uvm_pagehash(uobj, pg->offset);
200 	buck = &uvm.page_hash[hash];
201 	lock = uvm_hashlock(hash);
202 	mutex_spin_enter(lock);
203 	TAILQ_INSERT_TAIL(buck, pg, hashq);
204 	mutex_spin_exit(lock);
205 
206 	if (UVM_OBJ_IS_VNODE(uobj)) {
207 		if (uobj->uo_npages == 0) {
208 			struct vnode *vp = (struct vnode *)uobj;
209 
210 			vholdl(vp);
211 		}
212 		if (UVM_OBJ_IS_VTEXT(uobj)) {
213 			atomic_inc_uint(&uvmexp.execpages);
214 		} else {
215 			atomic_inc_uint(&uvmexp.filepages);
216 		}
217 	} else if (UVM_OBJ_IS_AOBJ(uobj)) {
218 		atomic_inc_uint(&uvmexp.anonpages);
219 	}
220 
221 	if (where)
222 		TAILQ_INSERT_AFTER(&uobj->memq, where, pg, listq);
223 	else
224 		TAILQ_INSERT_TAIL(&uobj->memq, pg, listq);
225 	pg->flags |= PG_TABLED;
226 	uobj->uo_npages++;
227 }
228 
229 inline static void
230 uvm_pageinsert(struct vm_page *pg)
231 {
232 
233 	uvm_pageinsert_after(pg, NULL);
234 }
235 
236 /*
237  * uvm_page_remove: remove page from object and hash
238  *
239  * => caller must lock object
240  * => caller must lock page queues
241  */
242 
243 static inline void
244 uvm_pageremove(struct vm_page *pg)
245 {
246 	struct pglist *buck;
247 	struct uvm_object *uobj = pg->uobject;
248 	kmutex_t *lock;
249 	u_int hash;
250 
251 	LOCK_ASSERT(simple_lock_held(&uobj->vmobjlock));
252 	KASSERT(pg->flags & PG_TABLED);
253 
254 	hash = uvm_pagehash(uobj, pg->offset);
255 	buck = &uvm.page_hash[hash];
256 	lock = uvm_hashlock(hash);
257 	mutex_spin_enter(lock);
258 	TAILQ_REMOVE(buck, pg, hashq);
259 	mutex_spin_exit(lock);
260 
261 	if (UVM_OBJ_IS_VNODE(uobj)) {
262 		if (uobj->uo_npages == 1) {
263 			struct vnode *vp = (struct vnode *)uobj;
264 
265 			holdrelel(vp);
266 		}
267 		if (UVM_OBJ_IS_VTEXT(uobj)) {
268 			atomic_dec_uint(&uvmexp.execpages);
269 		} else {
270 			atomic_dec_uint(&uvmexp.filepages);
271 		}
272 	} else if (UVM_OBJ_IS_AOBJ(uobj)) {
273 		atomic_dec_uint(&uvmexp.anonpages);
274 	}
275 
276 	/* object should be locked */
277 	uobj->uo_npages--;
278 	TAILQ_REMOVE(&uobj->memq, pg, listq);
279 	pg->flags &= ~PG_TABLED;
280 	pg->uobject = NULL;
281 }
282 
283 static void
284 uvm_page_init_buckets(struct pgfreelist *pgfl)
285 {
286 	int color, i;
287 
288 	for (color = 0; color < uvmexp.ncolors; color++) {
289 		for (i = 0; i < PGFL_NQUEUES; i++) {
290 			TAILQ_INIT(&pgfl->pgfl_buckets[color].pgfl_queues[i]);
291 		}
292 	}
293 }
294 
295 /*
296  * uvm_page_init: init the page system.   called from uvm_init().
297  *
298  * => we return the range of kernel virtual memory in kvm_startp/kvm_endp
299  */
300 
301 void
302 uvm_page_init(vaddr_t *kvm_startp, vaddr_t *kvm_endp)
303 {
304 	vsize_t freepages, pagecount, bucketcount, n;
305 	struct pgflbucket *bucketarray;
306 	struct vm_page *pagearray;
307 	int lcv;
308 	u_int i;
309 	paddr_t paddr;
310 
311 	/*
312 	 * init the page queues and page queue locks, except the free
313 	 * list; we allocate that later (with the initial vm_page
314 	 * structures).
315 	 */
316 
317 	uvmpdpol_init();
318 	simple_lock_init(&uvm.pageqlock);
319 	mutex_init(&uvm_fpageqlock, MUTEX_DRIVER, IPL_VM);
320 
321 	/*
322 	 * init the <obj,offset> => <page> hash table.  for now
323 	 * we just have one bucket (the bootstrap bucket).  later on we
324 	 * will allocate new buckets as we dynamically resize the hash table.
325 	 */
326 
327 	uvm.page_nhash = 1;			/* 1 bucket */
328 	uvm.page_hashmask = 0;			/* mask for hash function */
329 	uvm.page_hash = &uvm_bootbucket;	/* install bootstrap bucket */
330 	TAILQ_INIT(uvm.page_hash);		/* init hash table */
331 
332 	/*
333 	 * init hashtable locks.  these must be spinlocks, as they are
334 	 * called from sites in the pmap modules where we cannot block.
335 	 * if taking multiple locks, the order is: low numbered first,
336 	 * high numbered second.
337 	 */
338 
339 	for (i = 0; i < UVM_HASHLOCK_CNT; i++)
340 		mutex_init(&uvm_hashlocks[i].lock, MUTEX_SPIN, IPL_VM);
341 
342 	/*
343 	 * allocate vm_page structures.
344 	 */
345 
346 	/*
347 	 * sanity check:
348 	 * before calling this function the MD code is expected to register
349 	 * some free RAM with the uvm_page_physload() function.   our job
350 	 * now is to allocate vm_page structures for this memory.
351 	 */
352 
353 	if (vm_nphysseg == 0)
354 		panic("uvm_page_bootstrap: no memory pre-allocated");
355 
356 	/*
357 	 * first calculate the number of free pages...
358 	 *
359 	 * note that we use start/end rather than avail_start/avail_end.
360 	 * this allows us to allocate extra vm_page structures in case we
361 	 * want to return some memory to the pool after booting.
362 	 */
363 
364 	freepages = 0;
365 	for (lcv = 0 ; lcv < vm_nphysseg ; lcv++)
366 		freepages += (vm_physmem[lcv].end - vm_physmem[lcv].start);
367 
368 	/*
369 	 * Let MD code initialize the number of colors, or default
370 	 * to 1 color if MD code doesn't care.
371 	 */
372 	if (uvmexp.ncolors == 0)
373 		uvmexp.ncolors = 1;
374 	uvmexp.colormask = uvmexp.ncolors - 1;
375 
376 	/*
377 	 * we now know we have (PAGE_SIZE * freepages) bytes of memory we can
378 	 * use.   for each page of memory we use we need a vm_page structure.
379 	 * thus, the total number of pages we can use is the total size of
380 	 * the memory divided by the PAGE_SIZE plus the size of the vm_page
381 	 * structure.   we add one to freepages as a fudge factor to avoid
382 	 * truncation errors (since we can only allocate in terms of whole
383 	 * pages).
384 	 */
385 
386 	bucketcount = uvmexp.ncolors * VM_NFREELIST;
387 	pagecount = ((freepages + 1) << PAGE_SHIFT) /
388 	    (PAGE_SIZE + sizeof(struct vm_page));
389 
390 	bucketarray = (void *)uvm_pageboot_alloc((bucketcount *
391 	    sizeof(struct pgflbucket)) + (pagecount *
392 	    sizeof(struct vm_page)));
393 	pagearray = (struct vm_page *)(bucketarray + bucketcount);
394 
395 	for (lcv = 0; lcv < VM_NFREELIST; lcv++) {
396 		uvm.page_free[lcv].pgfl_buckets =
397 		    (bucketarray + (lcv * uvmexp.ncolors));
398 		uvm_page_init_buckets(&uvm.page_free[lcv]);
399 	}
400 	memset(pagearray, 0, pagecount * sizeof(struct vm_page));
401 
402 	/*
403 	 * init the vm_page structures and put them in the correct place.
404 	 */
405 
406 	for (lcv = 0 ; lcv < vm_nphysseg ; lcv++) {
407 		n = vm_physmem[lcv].end - vm_physmem[lcv].start;
408 
409 		/* set up page array pointers */
410 		vm_physmem[lcv].pgs = pagearray;
411 		pagearray += n;
412 		pagecount -= n;
413 		vm_physmem[lcv].lastpg = vm_physmem[lcv].pgs + (n - 1);
414 
415 		/* init and free vm_pages (we've already zeroed them) */
416 		paddr = ptoa(vm_physmem[lcv].start);
417 		for (i = 0 ; i < n ; i++, paddr += PAGE_SIZE) {
418 			vm_physmem[lcv].pgs[i].phys_addr = paddr;
419 #ifdef __HAVE_VM_PAGE_MD
420 			VM_MDPAGE_INIT(&vm_physmem[lcv].pgs[i]);
421 #endif
422 			if (atop(paddr) >= vm_physmem[lcv].avail_start &&
423 			    atop(paddr) <= vm_physmem[lcv].avail_end) {
424 				uvmexp.npages++;
425 				/* add page to free pool */
426 				uvm_pagefree(&vm_physmem[lcv].pgs[i]);
427 			}
428 		}
429 	}
430 
431 	/*
432 	 * pass up the values of virtual_space_start and
433 	 * virtual_space_end (obtained by uvm_pageboot_alloc) to the upper
434 	 * layers of the VM.
435 	 */
436 
437 	*kvm_startp = round_page(virtual_space_start);
438 	*kvm_endp = trunc_page(virtual_space_end);
439 #ifdef DEBUG
440 	/*
441 	 * steal kva for uvm_pagezerocheck().
442 	 */
443 	uvm_zerocheckkva = *kvm_startp;
444 	*kvm_startp += PAGE_SIZE;
445 #endif /* DEBUG */
446 
447 	/*
448 	 * init locks for kernel threads
449 	 */
450 
451 	mutex_init(&uvm_pagedaemon_lock, MUTEX_DEFAULT, IPL_NONE);
452 
453 	/*
454 	 * init various thresholds.
455 	 */
456 
457 	uvmexp.reserve_pagedaemon = 1;
458 	uvmexp.reserve_kernel = 5;
459 
460 	/*
461 	 * determine if we should zero pages in the idle loop.
462 	 */
463 
464 	uvm.page_idle_zero = vm_page_zero_enable;
465 
466 	/*
467 	 * done!
468 	 */
469 
470 	uvm.page_init_done = true;
471 }
472 
473 /*
474  * uvm_setpagesize: set the page size
475  *
476  * => sets page_shift and page_mask from uvmexp.pagesize.
477  */
478 
479 void
480 uvm_setpagesize(void)
481 {
482 
483 	/*
484 	 * If uvmexp.pagesize is 0 at this point, we expect PAGE_SIZE
485 	 * to be a constant (indicated by being a non-zero value).
486 	 */
487 	if (uvmexp.pagesize == 0) {
488 		if (PAGE_SIZE == 0)
489 			panic("uvm_setpagesize: uvmexp.pagesize not set");
490 		uvmexp.pagesize = PAGE_SIZE;
491 	}
492 	uvmexp.pagemask = uvmexp.pagesize - 1;
493 	if ((uvmexp.pagemask & uvmexp.pagesize) != 0)
494 		panic("uvm_setpagesize: page size not a power of two");
495 	for (uvmexp.pageshift = 0; ; uvmexp.pageshift++)
496 		if ((1 << uvmexp.pageshift) == uvmexp.pagesize)
497 			break;
498 }
499 
500 /*
501  * uvm_pageboot_alloc: steal memory from physmem for bootstrapping
502  */
503 
504 vaddr_t
505 uvm_pageboot_alloc(vsize_t size)
506 {
507 	static bool initialized = false;
508 	vaddr_t addr;
509 #if !defined(PMAP_STEAL_MEMORY)
510 	vaddr_t vaddr;
511 	paddr_t paddr;
512 #endif
513 
514 	/*
515 	 * on first call to this function, initialize ourselves.
516 	 */
517 	if (initialized == false) {
518 		pmap_virtual_space(&virtual_space_start, &virtual_space_end);
519 
520 		/* round it the way we like it */
521 		virtual_space_start = round_page(virtual_space_start);
522 		virtual_space_end = trunc_page(virtual_space_end);
523 
524 		initialized = true;
525 	}
526 
527 	/* round to page size */
528 	size = round_page(size);
529 
530 #if defined(PMAP_STEAL_MEMORY)
531 
532 	/*
533 	 * defer bootstrap allocation to MD code (it may want to allocate
534 	 * from a direct-mapped segment).  pmap_steal_memory should adjust
535 	 * virtual_space_start/virtual_space_end if necessary.
536 	 */
537 
538 	addr = pmap_steal_memory(size, &virtual_space_start,
539 	    &virtual_space_end);
540 
541 	return(addr);
542 
543 #else /* !PMAP_STEAL_MEMORY */
544 
545 	/*
546 	 * allocate virtual memory for this request
547 	 */
548 	if (virtual_space_start == virtual_space_end ||
549 	    (virtual_space_end - virtual_space_start) < size)
550 		panic("uvm_pageboot_alloc: out of virtual space");
551 
552 	addr = virtual_space_start;
553 
554 #ifdef PMAP_GROWKERNEL
555 	/*
556 	 * If the kernel pmap can't map the requested space,
557 	 * then allocate more resources for it.
558 	 */
559 	if (uvm_maxkaddr < (addr + size)) {
560 		uvm_maxkaddr = pmap_growkernel(addr + size);
561 		if (uvm_maxkaddr < (addr + size))
562 			panic("uvm_pageboot_alloc: pmap_growkernel() failed");
563 	}
564 #endif
565 
566 	virtual_space_start += size;
567 
568 	/*
569 	 * allocate and mapin physical pages to back new virtual pages
570 	 */
571 
572 	for (vaddr = round_page(addr) ; vaddr < addr + size ;
573 	    vaddr += PAGE_SIZE) {
574 
575 		if (!uvm_page_physget(&paddr))
576 			panic("uvm_pageboot_alloc: out of memory");
577 
578 		/*
579 		 * Note this memory is no longer managed, so using
580 		 * pmap_kenter is safe.
581 		 */
582 		pmap_kenter_pa(vaddr, paddr, VM_PROT_READ|VM_PROT_WRITE);
583 	}
584 	pmap_update(pmap_kernel());
585 	return(addr);
586 #endif	/* PMAP_STEAL_MEMORY */
587 }
588 
589 #if !defined(PMAP_STEAL_MEMORY)
590 /*
591  * uvm_page_physget: "steal" one page from the vm_physmem structure.
592  *
593  * => attempt to allocate it off the end of a segment in which the "avail"
594  *    values match the start/end values.   if we can't do that, then we
595  *    will advance both values (making them equal, and removing some
596  *    vm_page structures from the non-avail area).
597  * => return false if out of memory.
598  */
599 
600 /* subroutine: try to allocate from memory chunks on the specified freelist */
601 static bool uvm_page_physget_freelist(paddr_t *, int);
602 
603 static bool
604 uvm_page_physget_freelist(paddr_t *paddrp, int freelist)
605 {
606 	int lcv, x;
607 
608 	/* pass 1: try allocating from a matching end */
609 #if (VM_PHYSSEG_STRAT == VM_PSTRAT_BIGFIRST)
610 	for (lcv = vm_nphysseg - 1 ; lcv >= 0 ; lcv--)
611 #else
612 	for (lcv = 0 ; lcv < vm_nphysseg ; lcv++)
613 #endif
614 	{
615 
616 		if (uvm.page_init_done == true)
617 			panic("uvm_page_physget: called _after_ bootstrap");
618 
619 		if (vm_physmem[lcv].free_list != freelist)
620 			continue;
621 
622 		/* try from front */
623 		if (vm_physmem[lcv].avail_start == vm_physmem[lcv].start &&
624 		    vm_physmem[lcv].avail_start < vm_physmem[lcv].avail_end) {
625 			*paddrp = ptoa(vm_physmem[lcv].avail_start);
626 			vm_physmem[lcv].avail_start++;
627 			vm_physmem[lcv].start++;
628 			/* nothing left?   nuke it */
629 			if (vm_physmem[lcv].avail_start ==
630 			    vm_physmem[lcv].end) {
631 				if (vm_nphysseg == 1)
632 				    panic("uvm_page_physget: out of memory!");
633 				vm_nphysseg--;
634 				for (x = lcv ; x < vm_nphysseg ; x++)
635 					/* structure copy */
636 					vm_physmem[x] = vm_physmem[x+1];
637 			}
638 			return (true);
639 		}
640 
641 		/* try from rear */
642 		if (vm_physmem[lcv].avail_end == vm_physmem[lcv].end &&
643 		    vm_physmem[lcv].avail_start < vm_physmem[lcv].avail_end) {
644 			*paddrp = ptoa(vm_physmem[lcv].avail_end - 1);
645 			vm_physmem[lcv].avail_end--;
646 			vm_physmem[lcv].end--;
647 			/* nothing left?   nuke it */
648 			if (vm_physmem[lcv].avail_end ==
649 			    vm_physmem[lcv].start) {
650 				if (vm_nphysseg == 1)
651 				    panic("uvm_page_physget: out of memory!");
652 				vm_nphysseg--;
653 				for (x = lcv ; x < vm_nphysseg ; x++)
654 					/* structure copy */
655 					vm_physmem[x] = vm_physmem[x+1];
656 			}
657 			return (true);
658 		}
659 	}
660 
661 	/* pass2: forget about matching ends, just allocate something */
662 #if (VM_PHYSSEG_STRAT == VM_PSTRAT_BIGFIRST)
663 	for (lcv = vm_nphysseg - 1 ; lcv >= 0 ; lcv--)
664 #else
665 	for (lcv = 0 ; lcv < vm_nphysseg ; lcv++)
666 #endif
667 	{
668 
669 		/* any room in this bank? */
670 		if (vm_physmem[lcv].avail_start >= vm_physmem[lcv].avail_end)
671 			continue;  /* nope */
672 
673 		*paddrp = ptoa(vm_physmem[lcv].avail_start);
674 		vm_physmem[lcv].avail_start++;
675 		/* truncate! */
676 		vm_physmem[lcv].start = vm_physmem[lcv].avail_start;
677 
678 		/* nothing left?   nuke it */
679 		if (vm_physmem[lcv].avail_start == vm_physmem[lcv].end) {
680 			if (vm_nphysseg == 1)
681 				panic("uvm_page_physget: out of memory!");
682 			vm_nphysseg--;
683 			for (x = lcv ; x < vm_nphysseg ; x++)
684 				/* structure copy */
685 				vm_physmem[x] = vm_physmem[x+1];
686 		}
687 		return (true);
688 	}
689 
690 	return (false);        /* whoops! */
691 }
692 
693 bool
694 uvm_page_physget(paddr_t *paddrp)
695 {
696 	int i;
697 
698 	/* try in the order of freelist preference */
699 	for (i = 0; i < VM_NFREELIST; i++)
700 		if (uvm_page_physget_freelist(paddrp, i) == true)
701 			return (true);
702 	return (false);
703 }
704 #endif /* PMAP_STEAL_MEMORY */
705 
706 /*
707  * uvm_page_physload: load physical memory into VM system
708  *
709  * => all args are PFs
710  * => all pages in start/end get vm_page structures
711  * => areas marked by avail_start/avail_end get added to the free page pool
712  * => we are limited to VM_PHYSSEG_MAX physical memory segments
713  */
714 
715 void
716 uvm_page_physload(paddr_t start, paddr_t end, paddr_t avail_start,
717     paddr_t avail_end, int free_list)
718 {
719 	int preload, lcv;
720 	psize_t npages;
721 	struct vm_page *pgs;
722 	struct vm_physseg *ps;
723 
724 	if (uvmexp.pagesize == 0)
725 		panic("uvm_page_physload: page size not set!");
726 	if (free_list >= VM_NFREELIST || free_list < VM_FREELIST_DEFAULT)
727 		panic("uvm_page_physload: bad free list %d", free_list);
728 	if (start >= end)
729 		panic("uvm_page_physload: start >= end");
730 
731 	/*
732 	 * do we have room?
733 	 */
734 
735 	if (vm_nphysseg == VM_PHYSSEG_MAX) {
736 		printf("uvm_page_physload: unable to load physical memory "
737 		    "segment\n");
738 		printf("\t%d segments allocated, ignoring 0x%llx -> 0x%llx\n",
739 		    VM_PHYSSEG_MAX, (long long)start, (long long)end);
740 		printf("\tincrease VM_PHYSSEG_MAX\n");
741 		return;
742 	}
743 
744 	/*
745 	 * check to see if this is a "preload" (i.e. uvm_mem_init hasn't been
746 	 * called yet, so malloc is not available).
747 	 */
748 
749 	for (lcv = 0 ; lcv < vm_nphysseg ; lcv++) {
750 		if (vm_physmem[lcv].pgs)
751 			break;
752 	}
753 	preload = (lcv == vm_nphysseg);
754 
755 	/*
756 	 * if VM is already running, attempt to malloc() vm_page structures
757 	 */
758 
759 	if (!preload) {
760 #if defined(VM_PHYSSEG_NOADD)
761 		panic("uvm_page_physload: tried to add RAM after vm_mem_init");
762 #else
763 		/* XXXCDC: need some sort of lockout for this case */
764 		paddr_t paddr;
765 		npages = end - start;  /* # of pages */
766 		pgs = malloc(sizeof(struct vm_page) * npages,
767 		    M_VMPAGE, M_NOWAIT);
768 		if (pgs == NULL) {
769 			printf("uvm_page_physload: can not malloc vm_page "
770 			    "structs for segment\n");
771 			printf("\tignoring 0x%lx -> 0x%lx\n", start, end);
772 			return;
773 		}
774 		/* zero data, init phys_addr and free_list, and free pages */
775 		memset(pgs, 0, sizeof(struct vm_page) * npages);
776 		for (lcv = 0, paddr = ptoa(start) ;
777 				 lcv < npages ; lcv++, paddr += PAGE_SIZE) {
778 			pgs[lcv].phys_addr = paddr;
779 			pgs[lcv].free_list = free_list;
780 			if (atop(paddr) >= avail_start &&
781 			    atop(paddr) <= avail_end)
782 				uvm_pagefree(&pgs[lcv]);
783 		}
784 		/* XXXCDC: incomplete: need to update uvmexp.free, what else? */
785 		/* XXXCDC: need hook to tell pmap to rebuild pv_list, etc... */
786 #endif
787 	} else {
788 		pgs = NULL;
789 		npages = 0;
790 	}
791 
792 	/*
793 	 * now insert us in the proper place in vm_physmem[]
794 	 */
795 
796 #if (VM_PHYSSEG_STRAT == VM_PSTRAT_RANDOM)
797 	/* random: put it at the end (easy!) */
798 	ps = &vm_physmem[vm_nphysseg];
799 #elif (VM_PHYSSEG_STRAT == VM_PSTRAT_BSEARCH)
800 	{
801 		int x;
802 		/* sort by address for binary search */
803 		for (lcv = 0 ; lcv < vm_nphysseg ; lcv++)
804 			if (start < vm_physmem[lcv].start)
805 				break;
806 		ps = &vm_physmem[lcv];
807 		/* move back other entries, if necessary ... */
808 		for (x = vm_nphysseg ; x > lcv ; x--)
809 			/* structure copy */
810 			vm_physmem[x] = vm_physmem[x - 1];
811 	}
812 #elif (VM_PHYSSEG_STRAT == VM_PSTRAT_BIGFIRST)
813 	{
814 		int x;
815 		/* sort by largest segment first */
816 		for (lcv = 0 ; lcv < vm_nphysseg ; lcv++)
817 			if ((end - start) >
818 			    (vm_physmem[lcv].end - vm_physmem[lcv].start))
819 				break;
820 		ps = &vm_physmem[lcv];
821 		/* move back other entries, if necessary ... */
822 		for (x = vm_nphysseg ; x > lcv ; x--)
823 			/* structure copy */
824 			vm_physmem[x] = vm_physmem[x - 1];
825 	}
826 #else
827 	panic("uvm_page_physload: unknown physseg strategy selected!");
828 #endif
829 
830 	ps->start = start;
831 	ps->end = end;
832 	ps->avail_start = avail_start;
833 	ps->avail_end = avail_end;
834 	if (preload) {
835 		ps->pgs = NULL;
836 	} else {
837 		ps->pgs = pgs;
838 		ps->lastpg = pgs + npages - 1;
839 	}
840 	ps->free_list = free_list;
841 	vm_nphysseg++;
842 
843 	if (!preload) {
844 		uvm_page_rehash();
845 		uvmpdpol_reinit();
846 	}
847 }
848 
849 /*
850  * uvm_page_rehash: reallocate hash table based on number of free pages.
851  */
852 
853 void
854 uvm_page_rehash(void)
855 {
856 	int freepages, lcv, bucketcount, oldcount, i;
857 	struct pglist *newbuckets, *oldbuckets;
858 	struct vm_page *pg;
859 	size_t newsize, oldsize;
860 
861 	/*
862 	 * compute number of pages that can go in the free pool
863 	 */
864 
865 	freepages = 0;
866 	for (lcv = 0 ; lcv < vm_nphysseg ; lcv++)
867 		freepages +=
868 		    (vm_physmem[lcv].avail_end - vm_physmem[lcv].avail_start);
869 
870 	/*
871 	 * compute number of buckets needed for this number of pages
872 	 */
873 
874 	bucketcount = 1;
875 	while (bucketcount < freepages)
876 		bucketcount = bucketcount * 2;
877 
878 	/*
879 	 * compute the size of the current table and new table.
880 	 */
881 
882 	oldbuckets = uvm.page_hash;
883 	oldcount = uvm.page_nhash;
884 	oldsize = round_page(sizeof(struct pglist) * oldcount);
885 	newsize = round_page(sizeof(struct pglist) * bucketcount);
886 
887 	/*
888 	 * allocate the new buckets
889 	 */
890 
891 	newbuckets = (struct pglist *) uvm_km_alloc(kernel_map, newsize,
892 	    0, UVM_KMF_WIRED);
893 	if (newbuckets == NULL) {
894 		printf("uvm_page_physrehash: WARNING: could not grow page "
895 		    "hash table\n");
896 		return;
897 	}
898 	for (lcv = 0 ; lcv < bucketcount ; lcv++)
899 		TAILQ_INIT(&newbuckets[lcv]);
900 
901 	/*
902 	 * now replace the old buckets with the new ones and rehash everything
903 	 */
904 
905 	for (i = 0; i < UVM_HASHLOCK_CNT; i++)
906 		mutex_spin_enter(&uvm_hashlocks[i].lock);
907 
908 	uvm.page_hash = newbuckets;
909 	uvm.page_nhash = bucketcount;
910 	uvm.page_hashmask = bucketcount - 1;  /* power of 2 */
911 
912 	/* ... and rehash */
913 	for (lcv = 0 ; lcv < oldcount ; lcv++) {
914 		while ((pg = oldbuckets[lcv].tqh_first) != NULL) {
915 			TAILQ_REMOVE(&oldbuckets[lcv], pg, hashq);
916 			TAILQ_INSERT_TAIL(
917 			  &uvm.page_hash[uvm_pagehash(pg->uobject, pg->offset)],
918 			  pg, hashq);
919 		}
920 	}
921 
922 	for (i = 0; i < UVM_HASHLOCK_CNT; i++)
923 		mutex_spin_exit(&uvm_hashlocks[i].lock);
924 
925 	/*
926 	 * free old bucket array if is not the boot-time table
927 	 */
928 
929 	if (oldbuckets != &uvm_bootbucket)
930 		uvm_km_free(kernel_map, (vaddr_t) oldbuckets, oldsize,
931 		    UVM_KMF_WIRED);
932 }
933 
934 /*
935  * uvm_page_recolor: Recolor the pages if the new bucket count is
936  * larger than the old one.
937  */
938 
939 void
940 uvm_page_recolor(int newncolors)
941 {
942 	struct pgflbucket *bucketarray, *oldbucketarray;
943 	struct pgfreelist pgfl;
944 	struct vm_page *pg;
945 	vsize_t bucketcount;
946 	int lcv, color, i, ocolors;
947 
948 	if (newncolors <= uvmexp.ncolors)
949 		return;
950 
951 	if (uvm.page_init_done == false) {
952 		uvmexp.ncolors = newncolors;
953 		return;
954 	}
955 
956 	bucketcount = newncolors * VM_NFREELIST;
957 	bucketarray = malloc(bucketcount * sizeof(struct pgflbucket),
958 	    M_VMPAGE, M_NOWAIT);
959 	if (bucketarray == NULL) {
960 		printf("WARNING: unable to allocate %ld page color buckets\n",
961 		    (long) bucketcount);
962 		return;
963 	}
964 
965 	mutex_spin_enter(&uvm_fpageqlock);
966 
967 	/* Make sure we should still do this. */
968 	if (newncolors <= uvmexp.ncolors) {
969 		mutex_spin_exit(&uvm_fpageqlock);
970 		free(bucketarray, M_VMPAGE);
971 		return;
972 	}
973 
974 	oldbucketarray = uvm.page_free[0].pgfl_buckets;
975 	ocolors = uvmexp.ncolors;
976 
977 	uvmexp.ncolors = newncolors;
978 	uvmexp.colormask = uvmexp.ncolors - 1;
979 
980 	for (lcv = 0; lcv < VM_NFREELIST; lcv++) {
981 		pgfl.pgfl_buckets = (bucketarray + (lcv * newncolors));
982 		uvm_page_init_buckets(&pgfl);
983 		for (color = 0; color < ocolors; color++) {
984 			for (i = 0; i < PGFL_NQUEUES; i++) {
985 				while ((pg = TAILQ_FIRST(&uvm.page_free[
986 				    lcv].pgfl_buckets[color].pgfl_queues[i]))
987 				    != NULL) {
988 					TAILQ_REMOVE(&uvm.page_free[
989 					    lcv].pgfl_buckets[
990 					    color].pgfl_queues[i], pg, pageq);
991 					TAILQ_INSERT_TAIL(&pgfl.pgfl_buckets[
992 					    VM_PGCOLOR_BUCKET(pg)].pgfl_queues[
993 					    i], pg, pageq);
994 				}
995 			}
996 		}
997 		uvm.page_free[lcv].pgfl_buckets = pgfl.pgfl_buckets;
998 	}
999 
1000 	if (have_recolored_pages) {
1001 		mutex_spin_exit(&uvm_fpageqlock);
1002 		free(oldbucketarray, M_VMPAGE);
1003 		return;
1004 	}
1005 
1006 	have_recolored_pages = true;
1007 	mutex_spin_exit(&uvm_fpageqlock);
1008 }
1009 
1010 /*
1011  * uvm_pagealloc_pgfl: helper routine for uvm_pagealloc_strat
1012  */
1013 
1014 static struct vm_page *
1015 uvm_pagealloc_pgfl(struct pgfreelist *pgfl, int try1, int try2,
1016     int *trycolorp)
1017 {
1018 	struct pglist *freeq;
1019 	struct vm_page *pg;
1020 	int color, trycolor = *trycolorp;
1021 
1022 	color = trycolor;
1023 	do {
1024 		if ((pg = TAILQ_FIRST((freeq =
1025 		    &pgfl->pgfl_buckets[color].pgfl_queues[try1]))) != NULL)
1026 			goto gotit;
1027 		if ((pg = TAILQ_FIRST((freeq =
1028 		    &pgfl->pgfl_buckets[color].pgfl_queues[try2]))) != NULL)
1029 			goto gotit;
1030 		color = (color + 1) & uvmexp.colormask;
1031 	} while (color != trycolor);
1032 
1033 	return (NULL);
1034 
1035  gotit:
1036 	TAILQ_REMOVE(freeq, pg, pageq);
1037 	uvmexp.free--;
1038 
1039 	/* update zero'd page count */
1040 	if (pg->flags & PG_ZERO)
1041 		uvmexp.zeropages--;
1042 
1043 	if (color == trycolor)
1044 		uvmexp.colorhit++;
1045 	else {
1046 		uvmexp.colormiss++;
1047 		*trycolorp = color;
1048 	}
1049 
1050 	return (pg);
1051 }
1052 
1053 /*
1054  * uvm_pagealloc_strat: allocate vm_page from a particular free list.
1055  *
1056  * => return null if no pages free
1057  * => wake up pagedaemon if number of free pages drops below low water mark
1058  * => if obj != NULL, obj must be locked (to put in hash)
1059  * => if anon != NULL, anon must be locked (to put in anon)
1060  * => only one of obj or anon can be non-null
1061  * => caller must activate/deactivate page if it is not wired.
1062  * => free_list is ignored if strat == UVM_PGA_STRAT_NORMAL.
1063  * => policy decision: it is more important to pull a page off of the
1064  *	appropriate priority free list than it is to get a zero'd or
1065  *	unknown contents page.  This is because we live with the
1066  *	consequences of a bad free list decision for the entire
1067  *	lifetime of the page, e.g. if the page comes from memory that
1068  *	is slower to access.
1069  */
1070 
1071 struct vm_page *
1072 uvm_pagealloc_strat(struct uvm_object *obj, voff_t off, struct vm_anon *anon,
1073     int flags, int strat, int free_list)
1074 {
1075 	int lcv, try1, try2, zeroit = 0, color;
1076 	struct vm_page *pg;
1077 	bool use_reserve;
1078 
1079 	KASSERT(obj == NULL || anon == NULL);
1080 	KASSERT(anon == NULL || off == 0);
1081 	KASSERT(off == trunc_page(off));
1082 	LOCK_ASSERT(obj == NULL || simple_lock_held(&obj->vmobjlock));
1083 	LOCK_ASSERT(anon == NULL || simple_lock_held(&anon->an_lock));
1084 
1085 	mutex_spin_enter(&uvm_fpageqlock);
1086 
1087 	/*
1088 	 * This implements a global round-robin page coloring
1089 	 * algorithm.
1090 	 *
1091 	 * XXXJRT: Should we make the `nextcolor' per-CPU?
1092 	 * XXXJRT: What about virtually-indexed caches?
1093 	 */
1094 
1095 	color = uvm.page_free_nextcolor;
1096 
1097 	/*
1098 	 * check to see if we need to generate some free pages waking
1099 	 * the pagedaemon.
1100 	 */
1101 
1102 	uvm_kick_pdaemon();
1103 
1104 	/*
1105 	 * fail if any of these conditions is true:
1106 	 * [1]  there really are no free pages, or
1107 	 * [2]  only kernel "reserved" pages remain and
1108 	 *        the page isn't being allocated to a kernel object.
1109 	 * [3]  only pagedaemon "reserved" pages remain and
1110 	 *        the requestor isn't the pagedaemon.
1111 	 */
1112 
1113 	use_reserve = (flags & UVM_PGA_USERESERVE) ||
1114 		(obj && UVM_OBJ_IS_KERN_OBJECT(obj));
1115 	if ((uvmexp.free <= uvmexp.reserve_kernel && !use_reserve) ||
1116 	    (uvmexp.free <= uvmexp.reserve_pagedaemon &&
1117 	     !(use_reserve && curlwp == uvm.pagedaemon_lwp)))
1118 		goto fail;
1119 
1120 #if PGFL_NQUEUES != 2
1121 #error uvm_pagealloc_strat needs to be updated
1122 #endif
1123 
1124 	/*
1125 	 * If we want a zero'd page, try the ZEROS queue first, otherwise
1126 	 * we try the UNKNOWN queue first.
1127 	 */
1128 	if (flags & UVM_PGA_ZERO) {
1129 		try1 = PGFL_ZEROS;
1130 		try2 = PGFL_UNKNOWN;
1131 	} else {
1132 		try1 = PGFL_UNKNOWN;
1133 		try2 = PGFL_ZEROS;
1134 	}
1135 
1136  again:
1137 	switch (strat) {
1138 	case UVM_PGA_STRAT_NORMAL:
1139 		/* Check all freelists in descending priority order. */
1140 		for (lcv = 0; lcv < VM_NFREELIST; lcv++) {
1141 			pg = uvm_pagealloc_pgfl(&uvm.page_free[lcv],
1142 			    try1, try2, &color);
1143 			if (pg != NULL)
1144 				goto gotit;
1145 		}
1146 
1147 		/* No pages free! */
1148 		goto fail;
1149 
1150 	case UVM_PGA_STRAT_ONLY:
1151 	case UVM_PGA_STRAT_FALLBACK:
1152 		/* Attempt to allocate from the specified free list. */
1153 		KASSERT(free_list >= 0 && free_list < VM_NFREELIST);
1154 		pg = uvm_pagealloc_pgfl(&uvm.page_free[free_list],
1155 		    try1, try2, &color);
1156 		if (pg != NULL)
1157 			goto gotit;
1158 
1159 		/* Fall back, if possible. */
1160 		if (strat == UVM_PGA_STRAT_FALLBACK) {
1161 			strat = UVM_PGA_STRAT_NORMAL;
1162 			goto again;
1163 		}
1164 
1165 		/* No pages free! */
1166 		goto fail;
1167 
1168 	default:
1169 		panic("uvm_pagealloc_strat: bad strat %d", strat);
1170 		/* NOTREACHED */
1171 	}
1172 
1173  gotit:
1174 	/*
1175 	 * We now know which color we actually allocated from; set
1176 	 * the next color accordingly.
1177 	 */
1178 
1179 	uvm.page_free_nextcolor = (color + 1) & uvmexp.colormask;
1180 
1181 	/*
1182 	 * update allocation statistics and remember if we have to
1183 	 * zero the page
1184 	 */
1185 
1186 	if (flags & UVM_PGA_ZERO) {
1187 		if (pg->flags & PG_ZERO) {
1188 			uvmexp.pga_zerohit++;
1189 			zeroit = 0;
1190 		} else {
1191 			uvmexp.pga_zeromiss++;
1192 			zeroit = 1;
1193 		}
1194 	}
1195 	mutex_spin_exit(&uvm_fpageqlock);
1196 
1197 	pg->offset = off;
1198 	pg->uobject = obj;
1199 	pg->uanon = anon;
1200 	pg->flags = PG_BUSY|PG_CLEAN|PG_FAKE;
1201 	if (anon) {
1202 		anon->an_page = pg;
1203 		pg->pqflags = PQ_ANON;
1204 		atomic_inc_uint(&uvmexp.anonpages);
1205 	} else {
1206 		if (obj) {
1207 			uvm_pageinsert(pg);
1208 		}
1209 		pg->pqflags = 0;
1210 	}
1211 #if defined(UVM_PAGE_TRKOWN)
1212 	pg->owner_tag = NULL;
1213 #endif
1214 	UVM_PAGE_OWN(pg, "new alloc");
1215 
1216 	if (flags & UVM_PGA_ZERO) {
1217 		/*
1218 		 * A zero'd page is not clean.  If we got a page not already
1219 		 * zero'd, then we have to zero it ourselves.
1220 		 */
1221 		pg->flags &= ~PG_CLEAN;
1222 		if (zeroit)
1223 			pmap_zero_page(VM_PAGE_TO_PHYS(pg));
1224 	}
1225 
1226 	return(pg);
1227 
1228  fail:
1229 	mutex_spin_exit(&uvm_fpageqlock);
1230 	return (NULL);
1231 }
1232 
1233 /*
1234  * uvm_pagereplace: replace a page with another
1235  *
1236  * => object must be locked
1237  */
1238 
1239 void
1240 uvm_pagereplace(struct vm_page *oldpg, struct vm_page *newpg)
1241 {
1242 
1243 	KASSERT((oldpg->flags & PG_TABLED) != 0);
1244 	KASSERT(oldpg->uobject != NULL);
1245 	KASSERT((newpg->flags & PG_TABLED) == 0);
1246 	KASSERT(newpg->uobject == NULL);
1247 	LOCK_ASSERT(simple_lock_held(&oldpg->uobject->vmobjlock));
1248 
1249 	newpg->uobject = oldpg->uobject;
1250 	newpg->offset = oldpg->offset;
1251 
1252 	uvm_pageinsert_after(newpg, oldpg);
1253 	uvm_pageremove(oldpg);
1254 }
1255 
1256 /*
1257  * uvm_pagerealloc: reallocate a page from one object to another
1258  *
1259  * => both objects must be locked
1260  */
1261 
1262 void
1263 uvm_pagerealloc(struct vm_page *pg, struct uvm_object *newobj, voff_t newoff)
1264 {
1265 	/*
1266 	 * remove it from the old object
1267 	 */
1268 
1269 	if (pg->uobject) {
1270 		uvm_pageremove(pg);
1271 	}
1272 
1273 	/*
1274 	 * put it in the new object
1275 	 */
1276 
1277 	if (newobj) {
1278 		pg->uobject = newobj;
1279 		pg->offset = newoff;
1280 		uvm_pageinsert(pg);
1281 	}
1282 }
1283 
1284 #ifdef DEBUG
1285 /*
1286  * check if page is zero-filled
1287  *
1288  *  - called with free page queue lock held.
1289  */
1290 void
1291 uvm_pagezerocheck(struct vm_page *pg)
1292 {
1293 	int *p, *ep;
1294 
1295 	KASSERT(uvm_zerocheckkva != 0);
1296 	KASSERT(mutex_owned(&uvm_fpageqlock));
1297 
1298 	/*
1299 	 * XXX assuming pmap_kenter_pa and pmap_kremove never call
1300 	 * uvm page allocator.
1301 	 *
1302 	 * it might be better to have "CPU-local temporary map" pmap interface.
1303 	 */
1304 	pmap_kenter_pa(uvm_zerocheckkva, VM_PAGE_TO_PHYS(pg), VM_PROT_READ);
1305 	p = (int *)uvm_zerocheckkva;
1306 	ep = (int *)((char *)p + PAGE_SIZE);
1307 	pmap_update(pmap_kernel());
1308 	while (p < ep) {
1309 		if (*p != 0)
1310 			panic("PG_ZERO page isn't zero-filled");
1311 		p++;
1312 	}
1313 	pmap_kremove(uvm_zerocheckkva, PAGE_SIZE);
1314 }
1315 #endif /* DEBUG */
1316 
1317 /*
1318  * uvm_pagefree: free page
1319  *
1320  * => erase page's identity (i.e. remove from hash/object)
1321  * => put page on free list
1322  * => caller must lock owning object (either anon or uvm_object)
1323  * => caller must lock page queues
1324  * => assumes all valid mappings of pg are gone
1325  */
1326 
1327 void
1328 uvm_pagefree(struct vm_page *pg)
1329 {
1330 	struct pglist *pgfl;
1331 	bool iszero;
1332 
1333 #ifdef DEBUG
1334 	if (pg->uobject == (void *)0xdeadbeef &&
1335 	    pg->uanon == (void *)0xdeadbeef) {
1336 		panic("uvm_pagefree: freeing free page %p", pg);
1337 	}
1338 #endif /* DEBUG */
1339 
1340 	KASSERT((pg->flags & PG_PAGEOUT) == 0);
1341 	LOCK_ASSERT(simple_lock_held(&uvm.pageqlock) ||
1342 		!uvmpdpol_pageisqueued_p(pg));
1343 	LOCK_ASSERT(pg->uobject == NULL ||
1344 		simple_lock_held(&pg->uobject->vmobjlock));
1345 	LOCK_ASSERT(pg->uobject != NULL || pg->uanon == NULL ||
1346 		simple_lock_held(&pg->uanon->an_lock));
1347 
1348 	/*
1349 	 * if the page is loaned, resolve the loan instead of freeing.
1350 	 */
1351 
1352 	if (pg->loan_count) {
1353 		KASSERT(pg->wire_count == 0);
1354 
1355 		/*
1356 		 * if the page is owned by an anon then we just want to
1357 		 * drop anon ownership.  the kernel will free the page when
1358 		 * it is done with it.  if the page is owned by an object,
1359 		 * remove it from the object and mark it dirty for the benefit
1360 		 * of possible anon owners.
1361 		 *
1362 		 * regardless of previous ownership, wakeup any waiters,
1363 		 * unbusy the page, and we're done.
1364 		 */
1365 
1366 		if (pg->uobject != NULL) {
1367 			uvm_pageremove(pg);
1368 			pg->flags &= ~PG_CLEAN;
1369 		} else if (pg->uanon != NULL) {
1370 			if ((pg->pqflags & PQ_ANON) == 0) {
1371 				pg->loan_count--;
1372 			} else {
1373 				pg->pqflags &= ~PQ_ANON;
1374 				atomic_dec_uint(&uvmexp.anonpages);
1375 			}
1376 			pg->uanon->an_page = NULL;
1377 			pg->uanon = NULL;
1378 		}
1379 		if (pg->flags & PG_WANTED) {
1380 			wakeup(pg);
1381 		}
1382 		pg->flags &= ~(PG_WANTED|PG_BUSY|PG_RELEASED|PG_PAGER1);
1383 #ifdef UVM_PAGE_TRKOWN
1384 		pg->owner_tag = NULL;
1385 #endif
1386 		if (pg->loan_count) {
1387 			KASSERT(pg->uobject == NULL);
1388 			if (pg->uanon == NULL) {
1389 				uvm_pagedequeue(pg);
1390 			}
1391 			return;
1392 		}
1393 	}
1394 
1395 	/*
1396 	 * remove page from its object or anon.
1397 	 */
1398 
1399 	if (pg->uobject != NULL) {
1400 		uvm_pageremove(pg);
1401 	} else if (pg->uanon != NULL) {
1402 		pg->uanon->an_page = NULL;
1403 		atomic_dec_uint(&uvmexp.anonpages);
1404 	}
1405 
1406 	/*
1407 	 * now remove the page from the queues.
1408 	 */
1409 
1410 	uvm_pagedequeue(pg);
1411 
1412 	/*
1413 	 * if the page was wired, unwire it now.
1414 	 */
1415 
1416 	if (pg->wire_count) {
1417 		pg->wire_count = 0;
1418 		uvmexp.wired--;
1419 	}
1420 
1421 	/*
1422 	 * and put on free queue
1423 	 */
1424 
1425 	iszero = (pg->flags & PG_ZERO);
1426 	pgfl = &uvm.page_free[uvm_page_lookup_freelist(pg)].
1427 	    pgfl_buckets[VM_PGCOLOR_BUCKET(pg)].
1428 	    pgfl_queues[iszero ? PGFL_ZEROS : PGFL_UNKNOWN];
1429 
1430 	pg->pqflags = PQ_FREE;
1431 #ifdef DEBUG
1432 	pg->uobject = (void *)0xdeadbeef;
1433 	pg->offset = 0xdeadbeef;
1434 	pg->uanon = (void *)0xdeadbeef;
1435 #endif
1436 
1437 	mutex_spin_enter(&uvm_fpageqlock);
1438 
1439 #ifdef DEBUG
1440 	if (iszero)
1441 		uvm_pagezerocheck(pg);
1442 #endif /* DEBUG */
1443 
1444 	TAILQ_INSERT_HEAD(pgfl, pg, pageq);
1445 	uvmexp.free++;
1446 	if (iszero)
1447 		uvmexp.zeropages++;
1448 
1449 	if (uvmexp.zeropages < UVM_PAGEZERO_TARGET)
1450 		uvm.page_idle_zero = vm_page_zero_enable;
1451 
1452 	mutex_spin_exit(&uvm_fpageqlock);
1453 }
1454 
1455 /*
1456  * uvm_page_unbusy: unbusy an array of pages.
1457  *
1458  * => pages must either all belong to the same object, or all belong to anons.
1459  * => if pages are object-owned, object must be locked.
1460  * => if pages are anon-owned, anons must be locked.
1461  * => caller must lock page queues if pages may be released.
1462  * => caller must make sure that anon-owned pages are not PG_RELEASED.
1463  */
1464 
1465 void
1466 uvm_page_unbusy(struct vm_page **pgs, int npgs)
1467 {
1468 	struct vm_page *pg;
1469 	int i;
1470 	UVMHIST_FUNC("uvm_page_unbusy"); UVMHIST_CALLED(ubchist);
1471 
1472 	for (i = 0; i < npgs; i++) {
1473 		pg = pgs[i];
1474 		if (pg == NULL || pg == PGO_DONTCARE) {
1475 			continue;
1476 		}
1477 
1478 		LOCK_ASSERT(pg->uobject == NULL ||
1479 		    simple_lock_held(&pg->uobject->vmobjlock));
1480 		LOCK_ASSERT(pg->uobject != NULL ||
1481 		    (pg->uanon != NULL &&
1482 		    simple_lock_held(&pg->uanon->an_lock)));
1483 
1484 		KASSERT(pg->flags & PG_BUSY);
1485 		KASSERT((pg->flags & PG_PAGEOUT) == 0);
1486 		if (pg->flags & PG_WANTED) {
1487 			wakeup(pg);
1488 		}
1489 		if (pg->flags & PG_RELEASED) {
1490 			UVMHIST_LOG(ubchist, "releasing pg %p", pg,0,0,0);
1491 			KASSERT(pg->uobject != NULL ||
1492 			    (pg->uanon != NULL && pg->uanon->an_ref > 0));
1493 			pg->flags &= ~PG_RELEASED;
1494 			uvm_pagefree(pg);
1495 		} else {
1496 			UVMHIST_LOG(ubchist, "unbusying pg %p", pg,0,0,0);
1497 			pg->flags &= ~(PG_WANTED|PG_BUSY);
1498 			UVM_PAGE_OWN(pg, NULL);
1499 		}
1500 	}
1501 }
1502 
1503 #if defined(UVM_PAGE_TRKOWN)
1504 /*
1505  * uvm_page_own: set or release page ownership
1506  *
1507  * => this is a debugging function that keeps track of who sets PG_BUSY
1508  *	and where they do it.   it can be used to track down problems
1509  *	such a process setting "PG_BUSY" and never releasing it.
1510  * => page's object [if any] must be locked
1511  * => if "tag" is NULL then we are releasing page ownership
1512  */
1513 void
1514 uvm_page_own(struct vm_page *pg, const char *tag)
1515 {
1516 	struct uvm_object *uobj;
1517 	struct vm_anon *anon;
1518 
1519 	KASSERT((pg->flags & (PG_PAGEOUT|PG_RELEASED)) == 0);
1520 
1521 	uobj = pg->uobject;
1522 	anon = pg->uanon;
1523 	if (uobj != NULL) {
1524 		LOCK_ASSERT(simple_lock_held(&uobj->vmobjlock));
1525 	} else if (anon != NULL) {
1526 		LOCK_ASSERT(simple_lock_held(&anon->an_lock));
1527 	}
1528 
1529 	KASSERT((pg->flags & PG_WANTED) == 0);
1530 
1531 	/* gain ownership? */
1532 	if (tag) {
1533 		KASSERT((pg->flags & PG_BUSY) != 0);
1534 		if (pg->owner_tag) {
1535 			printf("uvm_page_own: page %p already owned "
1536 			    "by proc %d [%s]\n", pg,
1537 			    pg->owner, pg->owner_tag);
1538 			panic("uvm_page_own");
1539 		}
1540 		pg->owner = (curproc) ? curproc->p_pid :  (pid_t) -1;
1541 		pg->lowner = (curlwp) ? curlwp->l_lid :  (lwpid_t) -1;
1542 		pg->owner_tag = tag;
1543 		return;
1544 	}
1545 
1546 	/* drop ownership */
1547 	KASSERT((pg->flags & PG_BUSY) == 0);
1548 	if (pg->owner_tag == NULL) {
1549 		printf("uvm_page_own: dropping ownership of an non-owned "
1550 		    "page (%p)\n", pg);
1551 		panic("uvm_page_own");
1552 	}
1553 	if (!uvmpdpol_pageisqueued_p(pg)) {
1554 		KASSERT((pg->uanon == NULL && pg->uobject == NULL) ||
1555 		    pg->wire_count > 0);
1556 	} else {
1557 		KASSERT(pg->wire_count == 0);
1558 	}
1559 	pg->owner_tag = NULL;
1560 }
1561 #endif
1562 
1563 /*
1564  * uvm_pageidlezero: zero free pages while the system is idle.
1565  *
1566  * => try to complete one color bucket at a time, to reduce our impact
1567  *	on the CPU cache.
1568  * => we loop until we either reach the target or there is a lwp ready to run.
1569  */
1570 void
1571 uvm_pageidlezero(void)
1572 {
1573 	struct vm_page *pg;
1574 	struct pgfreelist *pgfl;
1575 	int free_list, firstbucket;
1576 	static int nextbucket;
1577 
1578 	KERNEL_LOCK(1, NULL);
1579 	mutex_spin_enter(&uvm_fpageqlock);
1580 	firstbucket = nextbucket;
1581 	do {
1582 		if (sched_curcpu_runnable_p()) {
1583 			goto quit;
1584 		}
1585 		if (uvmexp.zeropages >= UVM_PAGEZERO_TARGET) {
1586 			uvm.page_idle_zero = false;
1587 			goto quit;
1588 		}
1589 		for (free_list = 0; free_list < VM_NFREELIST; free_list++) {
1590 			pgfl = &uvm.page_free[free_list];
1591 			while ((pg = TAILQ_FIRST(&pgfl->pgfl_buckets[
1592 			    nextbucket].pgfl_queues[PGFL_UNKNOWN])) != NULL) {
1593 				if (sched_curcpu_runnable_p())
1594 					goto quit;
1595 
1596 				TAILQ_REMOVE(&pgfl->pgfl_buckets[
1597 				    nextbucket].pgfl_queues[PGFL_UNKNOWN],
1598 				    pg, pageq);
1599 				uvmexp.free--;
1600 				mutex_spin_exit(&uvm_fpageqlock);
1601 				KERNEL_UNLOCK_LAST(NULL);
1602 #ifdef PMAP_PAGEIDLEZERO
1603 				if (!PMAP_PAGEIDLEZERO(VM_PAGE_TO_PHYS(pg))) {
1604 
1605 					/*
1606 					 * The machine-dependent code detected
1607 					 * some reason for us to abort zeroing
1608 					 * pages, probably because there is a
1609 					 * process now ready to run.
1610 					 */
1611 
1612 					KERNEL_LOCK(1, NULL);
1613 					mutex_spin_enter(&uvm_fpageqlock);
1614 					TAILQ_INSERT_HEAD(&pgfl->pgfl_buckets[
1615 					    nextbucket].pgfl_queues[
1616 					    PGFL_UNKNOWN], pg, pageq);
1617 					uvmexp.free++;
1618 					uvmexp.zeroaborts++;
1619 					goto quit;
1620 				}
1621 #else
1622 				pmap_zero_page(VM_PAGE_TO_PHYS(pg));
1623 #endif /* PMAP_PAGEIDLEZERO */
1624 				pg->flags |= PG_ZERO;
1625 
1626 				KERNEL_LOCK(1, NULL);
1627 				mutex_spin_enter(&uvm_fpageqlock);
1628 				TAILQ_INSERT_HEAD(&pgfl->pgfl_buckets[
1629 				    nextbucket].pgfl_queues[PGFL_ZEROS],
1630 				    pg, pageq);
1631 				uvmexp.free++;
1632 				uvmexp.zeropages++;
1633 			}
1634 		}
1635 		nextbucket = (nextbucket + 1) & uvmexp.colormask;
1636 	} while (nextbucket != firstbucket);
1637 quit:
1638 	mutex_spin_exit(&uvm_fpageqlock);
1639 	KERNEL_UNLOCK_LAST(NULL);
1640 }
1641 
1642 /*
1643  * uvm_pagelookup: look up a page
1644  *
1645  * => caller should lock object to keep someone from pulling the page
1646  *	out from under it
1647  */
1648 
1649 struct vm_page *
1650 uvm_pagelookup(struct uvm_object *obj, voff_t off)
1651 {
1652 	struct vm_page *pg;
1653 	struct pglist *buck;
1654 	kmutex_t *lock;
1655 	u_int hash;
1656 
1657 	LOCK_ASSERT(simple_lock_held(&obj->vmobjlock));
1658 
1659 	hash = uvm_pagehash(obj, off);
1660 	buck = &uvm.page_hash[hash];
1661 	lock = uvm_hashlock(hash);
1662 	mutex_spin_enter(lock);
1663 	TAILQ_FOREACH(pg, buck, hashq) {
1664 		if (pg->uobject == obj && pg->offset == off) {
1665 			break;
1666 		}
1667 	}
1668 	mutex_spin_exit(lock);
1669 	KASSERT(pg == NULL || obj->uo_npages != 0);
1670 	KASSERT(pg == NULL || (pg->flags & (PG_RELEASED|PG_PAGEOUT)) == 0 ||
1671 		(pg->flags & PG_BUSY) != 0);
1672 	return(pg);
1673 }
1674 
1675 /*
1676  * uvm_pagewire: wire the page, thus removing it from the daemon's grasp
1677  *
1678  * => caller must lock page queues
1679  */
1680 
1681 void
1682 uvm_pagewire(struct vm_page *pg)
1683 {
1684 	UVM_LOCK_ASSERT_PAGEQ();
1685 #if defined(READAHEAD_STATS)
1686 	if ((pg->pqflags & PQ_READAHEAD) != 0) {
1687 		uvm_ra_hit.ev_count++;
1688 		pg->pqflags &= ~PQ_READAHEAD;
1689 	}
1690 #endif /* defined(READAHEAD_STATS) */
1691 	if (pg->wire_count == 0) {
1692 		uvm_pagedequeue(pg);
1693 		uvmexp.wired++;
1694 	}
1695 	pg->wire_count++;
1696 }
1697 
1698 /*
1699  * uvm_pageunwire: unwire the page.
1700  *
1701  * => activate if wire count goes to zero.
1702  * => caller must lock page queues
1703  */
1704 
1705 void
1706 uvm_pageunwire(struct vm_page *pg)
1707 {
1708 	UVM_LOCK_ASSERT_PAGEQ();
1709 	pg->wire_count--;
1710 	if (pg->wire_count == 0) {
1711 		uvm_pageactivate(pg);
1712 		uvmexp.wired--;
1713 	}
1714 }
1715 
1716 /*
1717  * uvm_pagedeactivate: deactivate page
1718  *
1719  * => caller must lock page queues
1720  * => caller must check to make sure page is not wired
1721  * => object that page belongs to must be locked (so we can adjust pg->flags)
1722  * => caller must clear the reference on the page before calling
1723  */
1724 
1725 void
1726 uvm_pagedeactivate(struct vm_page *pg)
1727 {
1728 
1729 	UVM_LOCK_ASSERT_PAGEQ();
1730 	KASSERT(pg->wire_count != 0 || uvmpdpol_pageisqueued_p(pg));
1731 	uvmpdpol_pagedeactivate(pg);
1732 }
1733 
1734 /*
1735  * uvm_pageactivate: activate page
1736  *
1737  * => caller must lock page queues
1738  */
1739 
1740 void
1741 uvm_pageactivate(struct vm_page *pg)
1742 {
1743 
1744 	UVM_LOCK_ASSERT_PAGEQ();
1745 #if defined(READAHEAD_STATS)
1746 	if ((pg->pqflags & PQ_READAHEAD) != 0) {
1747 		uvm_ra_hit.ev_count++;
1748 		pg->pqflags &= ~PQ_READAHEAD;
1749 	}
1750 #endif /* defined(READAHEAD_STATS) */
1751 	if (pg->wire_count != 0) {
1752 		return;
1753 	}
1754 	uvmpdpol_pageactivate(pg);
1755 }
1756 
1757 /*
1758  * uvm_pagedequeue: remove a page from any paging queue
1759  */
1760 
1761 void
1762 uvm_pagedequeue(struct vm_page *pg)
1763 {
1764 
1765 	if (uvmpdpol_pageisqueued_p(pg)) {
1766 		UVM_LOCK_ASSERT_PAGEQ();
1767 	}
1768 
1769 	uvmpdpol_pagedequeue(pg);
1770 }
1771 
1772 /*
1773  * uvm_pageenqueue: add a page to a paging queue without activating.
1774  * used where a page is not really demanded (yet).  eg. read-ahead
1775  */
1776 
1777 void
1778 uvm_pageenqueue(struct vm_page *pg)
1779 {
1780 
1781 	UVM_LOCK_ASSERT_PAGEQ();
1782 	if (pg->wire_count != 0) {
1783 		return;
1784 	}
1785 	uvmpdpol_pageenqueue(pg);
1786 }
1787 
1788 /*
1789  * uvm_pagezero: zero fill a page
1790  *
1791  * => if page is part of an object then the object should be locked
1792  *	to protect pg->flags.
1793  */
1794 
1795 void
1796 uvm_pagezero(struct vm_page *pg)
1797 {
1798 	pg->flags &= ~PG_CLEAN;
1799 	pmap_zero_page(VM_PAGE_TO_PHYS(pg));
1800 }
1801 
1802 /*
1803  * uvm_pagecopy: copy a page
1804  *
1805  * => if page is part of an object then the object should be locked
1806  *	to protect pg->flags.
1807  */
1808 
1809 void
1810 uvm_pagecopy(struct vm_page *src, struct vm_page *dst)
1811 {
1812 
1813 	dst->flags &= ~PG_CLEAN;
1814 	pmap_copy_page(VM_PAGE_TO_PHYS(src), VM_PAGE_TO_PHYS(dst));
1815 }
1816 
1817 /*
1818  * uvm_page_lookup_freelist: look up the free list for the specified page
1819  */
1820 
1821 int
1822 uvm_page_lookup_freelist(struct vm_page *pg)
1823 {
1824 	int lcv;
1825 
1826 	lcv = vm_physseg_find(atop(VM_PAGE_TO_PHYS(pg)), NULL);
1827 	KASSERT(lcv != -1);
1828 	return (vm_physmem[lcv].free_list);
1829 }
1830