xref: /netbsd-src/sys/uvm/uvm_page.c (revision 796c32c94f6e154afc9de0f63da35c91bb739b45)
1 /*	$NetBSD: uvm_page.c,v 1.194 2017/10/28 00:37:13 pgoyette 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. Neither the name of the University nor the names of its contributors
21  *    may be used to endorse or promote products derived from this software
22  *    without specific prior written permission.
23  *
24  * THIS SOFTWARE IS PROVIDED BY THE REGENTS AND CONTRIBUTORS ``AS IS'' AND
25  * ANY EXPRESS OR IMPLIED WARRANTIES, INCLUDING, BUT NOT LIMITED TO, THE
26  * IMPLIED WARRANTIES OF MERCHANTABILITY AND FITNESS FOR A PARTICULAR PURPOSE
27  * ARE DISCLAIMED.  IN NO EVENT SHALL THE REGENTS OR CONTRIBUTORS BE LIABLE
28  * FOR ANY DIRECT, INDIRECT, INCIDENTAL, SPECIAL, EXEMPLARY, OR CONSEQUENTIAL
29  * DAMAGES (INCLUDING, BUT NOT LIMITED TO, PROCUREMENT OF SUBSTITUTE GOODS
30  * OR SERVICES; LOSS OF USE, DATA, OR PROFITS; OR BUSINESS INTERRUPTION)
31  * HOWEVER CAUSED AND ON ANY THEORY OF LIABILITY, WHETHER IN CONTRACT, STRICT
32  * LIABILITY, OR TORT (INCLUDING NEGLIGENCE OR OTHERWISE) ARISING IN ANY WAY
33  * OUT OF THE USE OF THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF
34  * SUCH DAMAGE.
35  *
36  *	@(#)vm_page.c   8.3 (Berkeley) 3/21/94
37  * from: Id: uvm_page.c,v 1.1.2.18 1998/02/06 05:24:42 chs Exp
38  *
39  *
40  * Copyright (c) 1987, 1990 Carnegie-Mellon University.
41  * All rights reserved.
42  *
43  * Permission to use, copy, modify and distribute this software and
44  * its documentation is hereby granted, provided that both the copyright
45  * notice and this permission notice appear in all copies of the
46  * software, derivative works or modified versions, and any portions
47  * thereof, and that both notices appear in supporting documentation.
48  *
49  * CARNEGIE MELLON ALLOWS FREE USE OF THIS SOFTWARE IN ITS "AS IS"
50  * CONDITION.  CARNEGIE MELLON DISCLAIMS ANY LIABILITY OF ANY KIND
51  * FOR ANY DAMAGES WHATSOEVER RESULTING FROM THE USE OF THIS SOFTWARE.
52  *
53  * Carnegie Mellon requests users of this software to return to
54  *
55  *  Software Distribution Coordinator  or  Software.Distribution@CS.CMU.EDU
56  *  School of Computer Science
57  *  Carnegie Mellon University
58  *  Pittsburgh PA 15213-3890
59  *
60  * any improvements or extensions that they make and grant Carnegie the
61  * rights to redistribute these changes.
62  */
63 
64 /*
65  * uvm_page.c: page ops.
66  */
67 
68 #include <sys/cdefs.h>
69 __KERNEL_RCSID(0, "$NetBSD: uvm_page.c,v 1.194 2017/10/28 00:37:13 pgoyette Exp $");
70 
71 #include "opt_ddb.h"
72 #include "opt_uvm.h"
73 #include "opt_uvmhist.h"
74 #include "opt_readahead.h"
75 
76 #include <sys/param.h>
77 #include <sys/systm.h>
78 #include <sys/sched.h>
79 #include <sys/kernel.h>
80 #include <sys/vnode.h>
81 #include <sys/proc.h>
82 #include <sys/atomic.h>
83 #include <sys/cpu.h>
84 #include <sys/extent.h>
85 
86 #include <uvm/uvm.h>
87 #include <uvm/uvm_ddb.h>
88 #include <uvm/uvm_pdpolicy.h>
89 
90 /*
91  * Some supported CPUs in a given architecture don't support all
92  * of the things necessary to do idle page zero'ing efficiently.
93  * We therefore provide a way to enable it from machdep code here.
94  */
95 bool vm_page_zero_enable = false;
96 
97 /*
98  * number of pages per-CPU to reserve for the kernel.
99  */
100 #ifndef	UVM_RESERVED_PAGES_PER_CPU
101 #define	UVM_RESERVED_PAGES_PER_CPU	5
102 #endif
103 int vm_page_reserve_kernel = UVM_RESERVED_PAGES_PER_CPU;
104 
105 /*
106  * physical memory size;
107  */
108 psize_t physmem;
109 
110 /*
111  * local variables
112  */
113 
114 /*
115  * these variables record the values returned by vm_page_bootstrap,
116  * for debugging purposes.  The implementation of uvm_pageboot_alloc
117  * and pmap_startup here also uses them internally.
118  */
119 
120 static vaddr_t      virtual_space_start;
121 static vaddr_t      virtual_space_end;
122 
123 /*
124  * we allocate an initial number of page colors in uvm_page_init(),
125  * and remember them.  We may re-color pages as cache sizes are
126  * discovered during the autoconfiguration phase.  But we can never
127  * free the initial set of buckets, since they are allocated using
128  * uvm_pageboot_alloc().
129  */
130 
131 static size_t recolored_pages_memsize /* = 0 */;
132 
133 #ifdef DEBUG
134 vaddr_t uvm_zerocheckkva;
135 #endif /* DEBUG */
136 
137 /*
138  * These functions are reserved for uvm(9) internal use and are not
139  * exported in the header file uvm_physseg.h
140  *
141  * Thus they are redefined here.
142  */
143 void uvm_physseg_init_seg(uvm_physseg_t, struct vm_page *);
144 void uvm_physseg_seg_chomp_slab(uvm_physseg_t, struct vm_page *, size_t);
145 
146 /* returns a pgs array */
147 struct vm_page *uvm_physseg_seg_alloc_from_slab(uvm_physseg_t, size_t);
148 
149 /*
150  * local prototypes
151  */
152 
153 static void uvm_pageinsert(struct uvm_object *, struct vm_page *);
154 static void uvm_pageremove(struct uvm_object *, struct vm_page *);
155 
156 /*
157  * per-object tree of pages
158  */
159 
160 static signed int
161 uvm_page_compare_nodes(void *ctx, const void *n1, const void *n2)
162 {
163 	const struct vm_page *pg1 = n1;
164 	const struct vm_page *pg2 = n2;
165 	const voff_t a = pg1->offset;
166 	const voff_t b = pg2->offset;
167 
168 	if (a < b)
169 		return -1;
170 	if (a > b)
171 		return 1;
172 	return 0;
173 }
174 
175 static signed int
176 uvm_page_compare_key(void *ctx, const void *n, const void *key)
177 {
178 	const struct vm_page *pg = n;
179 	const voff_t a = pg->offset;
180 	const voff_t b = *(const voff_t *)key;
181 
182 	if (a < b)
183 		return -1;
184 	if (a > b)
185 		return 1;
186 	return 0;
187 }
188 
189 const rb_tree_ops_t uvm_page_tree_ops = {
190 	.rbto_compare_nodes = uvm_page_compare_nodes,
191 	.rbto_compare_key = uvm_page_compare_key,
192 	.rbto_node_offset = offsetof(struct vm_page, rb_node),
193 	.rbto_context = NULL
194 };
195 
196 /*
197  * inline functions
198  */
199 
200 /*
201  * uvm_pageinsert: insert a page in the object.
202  *
203  * => caller must lock object
204  * => caller must lock page queues
205  * => call should have already set pg's object and offset pointers
206  *    and bumped the version counter
207  */
208 
209 static inline void
210 uvm_pageinsert_list(struct uvm_object *uobj, struct vm_page *pg,
211     struct vm_page *where)
212 {
213 
214 	KASSERT(uobj == pg->uobject);
215 	KASSERT(mutex_owned(uobj->vmobjlock));
216 	KASSERT((pg->flags & PG_TABLED) == 0);
217 	KASSERT(where == NULL || (where->flags & PG_TABLED));
218 	KASSERT(where == NULL || (where->uobject == uobj));
219 
220 	if (UVM_OBJ_IS_VNODE(uobj)) {
221 		if (uobj->uo_npages == 0) {
222 			struct vnode *vp = (struct vnode *)uobj;
223 
224 			vholdl(vp);
225 		}
226 		if (UVM_OBJ_IS_VTEXT(uobj)) {
227 			atomic_inc_uint(&uvmexp.execpages);
228 		} else {
229 			atomic_inc_uint(&uvmexp.filepages);
230 		}
231 	} else if (UVM_OBJ_IS_AOBJ(uobj)) {
232 		atomic_inc_uint(&uvmexp.anonpages);
233 	}
234 
235 	if (where)
236 		TAILQ_INSERT_AFTER(&uobj->memq, where, pg, listq.queue);
237 	else
238 		TAILQ_INSERT_TAIL(&uobj->memq, pg, listq.queue);
239 	pg->flags |= PG_TABLED;
240 	uobj->uo_npages++;
241 }
242 
243 
244 static inline void
245 uvm_pageinsert_tree(struct uvm_object *uobj, struct vm_page *pg)
246 {
247 	struct vm_page *ret __diagused;
248 
249 	KASSERT(uobj == pg->uobject);
250 	ret = rb_tree_insert_node(&uobj->rb_tree, pg);
251 	KASSERT(ret == pg);
252 }
253 
254 static inline void
255 uvm_pageinsert(struct uvm_object *uobj, struct vm_page *pg)
256 {
257 
258 	KDASSERT(uobj != NULL);
259 	uvm_pageinsert_tree(uobj, pg);
260 	uvm_pageinsert_list(uobj, pg, NULL);
261 }
262 
263 /*
264  * uvm_page_remove: remove page from object.
265  *
266  * => caller must lock object
267  * => caller must lock page queues
268  */
269 
270 static inline void
271 uvm_pageremove_list(struct uvm_object *uobj, struct vm_page *pg)
272 {
273 
274 	KASSERT(uobj == pg->uobject);
275 	KASSERT(mutex_owned(uobj->vmobjlock));
276 	KASSERT(pg->flags & PG_TABLED);
277 
278 	if (UVM_OBJ_IS_VNODE(uobj)) {
279 		if (uobj->uo_npages == 1) {
280 			struct vnode *vp = (struct vnode *)uobj;
281 
282 			holdrelel(vp);
283 		}
284 		if (UVM_OBJ_IS_VTEXT(uobj)) {
285 			atomic_dec_uint(&uvmexp.execpages);
286 		} else {
287 			atomic_dec_uint(&uvmexp.filepages);
288 		}
289 	} else if (UVM_OBJ_IS_AOBJ(uobj)) {
290 		atomic_dec_uint(&uvmexp.anonpages);
291 	}
292 
293 	/* object should be locked */
294 	uobj->uo_npages--;
295 	TAILQ_REMOVE(&uobj->memq, pg, listq.queue);
296 	pg->flags &= ~PG_TABLED;
297 	pg->uobject = NULL;
298 }
299 
300 static inline void
301 uvm_pageremove_tree(struct uvm_object *uobj, struct vm_page *pg)
302 {
303 
304 	KASSERT(uobj == pg->uobject);
305 	rb_tree_remove_node(&uobj->rb_tree, pg);
306 }
307 
308 static inline void
309 uvm_pageremove(struct uvm_object *uobj, struct vm_page *pg)
310 {
311 
312 	KDASSERT(uobj != NULL);
313 	uvm_pageremove_tree(uobj, pg);
314 	uvm_pageremove_list(uobj, pg);
315 }
316 
317 static void
318 uvm_page_init_buckets(struct pgfreelist *pgfl)
319 {
320 	int color, i;
321 
322 	for (color = 0; color < uvmexp.ncolors; color++) {
323 		for (i = 0; i < PGFL_NQUEUES; i++) {
324 			LIST_INIT(&pgfl->pgfl_buckets[color].pgfl_queues[i]);
325 		}
326 	}
327 }
328 
329 /*
330  * uvm_page_init: init the page system.   called from uvm_init().
331  *
332  * => we return the range of kernel virtual memory in kvm_startp/kvm_endp
333  */
334 
335 void
336 uvm_page_init(vaddr_t *kvm_startp, vaddr_t *kvm_endp)
337 {
338 	static struct uvm_cpu boot_cpu;
339 	psize_t freepages, pagecount, bucketcount, n;
340 	struct pgflbucket *bucketarray, *cpuarray;
341 	struct vm_page *pagearray;
342 	uvm_physseg_t bank;
343 	int lcv;
344 
345 	KASSERT(ncpu <= 1);
346 	CTASSERT(sizeof(pagearray->offset) >= sizeof(struct uvm_cpu *));
347 
348 	/*
349 	 * init the page queues and page queue locks, except the free
350 	 * list; we allocate that later (with the initial vm_page
351 	 * structures).
352 	 */
353 
354 	uvm.cpus[0] = &boot_cpu;
355 	curcpu()->ci_data.cpu_uvm = &boot_cpu;
356 	uvmpdpol_init();
357 	mutex_init(&uvm_pageqlock, MUTEX_DRIVER, IPL_NONE);
358 	mutex_init(&uvm_fpageqlock, MUTEX_DRIVER, IPL_VM);
359 
360 	/*
361 	 * allocate vm_page structures.
362 	 */
363 
364 	/*
365 	 * sanity check:
366 	 * before calling this function the MD code is expected to register
367 	 * some free RAM with the uvm_page_physload() function.   our job
368 	 * now is to allocate vm_page structures for this memory.
369 	 */
370 
371 	if (uvm_physseg_get_last() == UVM_PHYSSEG_TYPE_INVALID)
372 		panic("uvm_page_bootstrap: no memory pre-allocated");
373 
374 	/*
375 	 * first calculate the number of free pages...
376 	 *
377 	 * note that we use start/end rather than avail_start/avail_end.
378 	 * this allows us to allocate extra vm_page structures in case we
379 	 * want to return some memory to the pool after booting.
380 	 */
381 
382 	freepages = 0;
383 
384 	for (bank = uvm_physseg_get_first();
385 	     uvm_physseg_valid_p(bank) ;
386 	     bank = uvm_physseg_get_next(bank)) {
387 		freepages += (uvm_physseg_get_end(bank) - uvm_physseg_get_start(bank));
388 	}
389 
390 	/*
391 	 * Let MD code initialize the number of colors, or default
392 	 * to 1 color if MD code doesn't care.
393 	 */
394 	if (uvmexp.ncolors == 0)
395 		uvmexp.ncolors = 1;
396 	uvmexp.colormask = uvmexp.ncolors - 1;
397 	KASSERT((uvmexp.colormask & uvmexp.ncolors) == 0);
398 
399 	/*
400 	 * we now know we have (PAGE_SIZE * freepages) bytes of memory we can
401 	 * use.   for each page of memory we use we need a vm_page structure.
402 	 * thus, the total number of pages we can use is the total size of
403 	 * the memory divided by the PAGE_SIZE plus the size of the vm_page
404 	 * structure.   we add one to freepages as a fudge factor to avoid
405 	 * truncation errors (since we can only allocate in terms of whole
406 	 * pages).
407 	 */
408 
409 	bucketcount = uvmexp.ncolors * VM_NFREELIST;
410 	pagecount = ((freepages + 1) << PAGE_SHIFT) /
411 	    (PAGE_SIZE + sizeof(struct vm_page));
412 
413 	bucketarray = (void *)uvm_pageboot_alloc((bucketcount *
414 	    sizeof(struct pgflbucket) * 2) + (pagecount *
415 	    sizeof(struct vm_page)));
416 	cpuarray = bucketarray + bucketcount;
417 	pagearray = (struct vm_page *)(bucketarray + bucketcount * 2);
418 
419 	for (lcv = 0; lcv < VM_NFREELIST; lcv++) {
420 		uvm.page_free[lcv].pgfl_buckets =
421 		    (bucketarray + (lcv * uvmexp.ncolors));
422 		uvm_page_init_buckets(&uvm.page_free[lcv]);
423 		uvm.cpus[0]->page_free[lcv].pgfl_buckets =
424 		    (cpuarray + (lcv * uvmexp.ncolors));
425 		uvm_page_init_buckets(&uvm.cpus[0]->page_free[lcv]);
426 	}
427 	memset(pagearray, 0, pagecount * sizeof(struct vm_page));
428 
429 	/*
430 	 * init the vm_page structures and put them in the correct place.
431 	 */
432 	/* First init the extent */
433 
434 	for (bank = uvm_physseg_get_first(),
435 		 uvm_physseg_seg_chomp_slab(bank, pagearray, pagecount);
436 	     uvm_physseg_valid_p(bank);
437 	     bank = uvm_physseg_get_next(bank)) {
438 
439 		n = uvm_physseg_get_end(bank) - uvm_physseg_get_start(bank);
440 		uvm_physseg_seg_alloc_from_slab(bank, n);
441 		uvm_physseg_init_seg(bank, pagearray);
442 
443 		/* set up page array pointers */
444 		pagearray += n;
445 		pagecount -= n;
446 	}
447 
448 	/*
449 	 * pass up the values of virtual_space_start and
450 	 * virtual_space_end (obtained by uvm_pageboot_alloc) to the upper
451 	 * layers of the VM.
452 	 */
453 
454 	*kvm_startp = round_page(virtual_space_start);
455 	*kvm_endp = trunc_page(virtual_space_end);
456 #ifdef DEBUG
457 	/*
458 	 * steal kva for uvm_pagezerocheck().
459 	 */
460 	uvm_zerocheckkva = *kvm_startp;
461 	*kvm_startp += PAGE_SIZE;
462 #endif /* DEBUG */
463 
464 	/*
465 	 * init various thresholds.
466 	 */
467 
468 	uvmexp.reserve_pagedaemon = 1;
469 	uvmexp.reserve_kernel = vm_page_reserve_kernel;
470 
471 	/*
472 	 * determine if we should zero pages in the idle loop.
473 	 */
474 
475 	uvm.cpus[0]->page_idle_zero = vm_page_zero_enable;
476 
477 	/*
478 	 * done!
479 	 */
480 
481 	uvm.page_init_done = true;
482 }
483 
484 /*
485  * uvm_setpagesize: set the page size
486  *
487  * => sets page_shift and page_mask from uvmexp.pagesize.
488  */
489 
490 void
491 uvm_setpagesize(void)
492 {
493 
494 	/*
495 	 * If uvmexp.pagesize is 0 at this point, we expect PAGE_SIZE
496 	 * to be a constant (indicated by being a non-zero value).
497 	 */
498 	if (uvmexp.pagesize == 0) {
499 		if (PAGE_SIZE == 0)
500 			panic("uvm_setpagesize: uvmexp.pagesize not set");
501 		uvmexp.pagesize = PAGE_SIZE;
502 	}
503 	uvmexp.pagemask = uvmexp.pagesize - 1;
504 	if ((uvmexp.pagemask & uvmexp.pagesize) != 0)
505 		panic("uvm_setpagesize: page size %u (%#x) not a power of two",
506 		    uvmexp.pagesize, uvmexp.pagesize);
507 	for (uvmexp.pageshift = 0; ; uvmexp.pageshift++)
508 		if ((1 << uvmexp.pageshift) == uvmexp.pagesize)
509 			break;
510 }
511 
512 /*
513  * uvm_pageboot_alloc: steal memory from physmem for bootstrapping
514  */
515 
516 vaddr_t
517 uvm_pageboot_alloc(vsize_t size)
518 {
519 	static bool initialized = false;
520 	vaddr_t addr;
521 #if !defined(PMAP_STEAL_MEMORY)
522 	vaddr_t vaddr;
523 	paddr_t paddr;
524 #endif
525 
526 	/*
527 	 * on first call to this function, initialize ourselves.
528 	 */
529 	if (initialized == false) {
530 		pmap_virtual_space(&virtual_space_start, &virtual_space_end);
531 
532 		/* round it the way we like it */
533 		virtual_space_start = round_page(virtual_space_start);
534 		virtual_space_end = trunc_page(virtual_space_end);
535 
536 		initialized = true;
537 	}
538 
539 	/* round to page size */
540 	size = round_page(size);
541 
542 #if defined(PMAP_STEAL_MEMORY)
543 
544 	/*
545 	 * defer bootstrap allocation to MD code (it may want to allocate
546 	 * from a direct-mapped segment).  pmap_steal_memory should adjust
547 	 * virtual_space_start/virtual_space_end if necessary.
548 	 */
549 
550 	addr = pmap_steal_memory(size, &virtual_space_start,
551 	    &virtual_space_end);
552 
553 	return(addr);
554 
555 #else /* !PMAP_STEAL_MEMORY */
556 
557 	/*
558 	 * allocate virtual memory for this request
559 	 */
560 	if (virtual_space_start == virtual_space_end ||
561 	    (virtual_space_end - virtual_space_start) < size)
562 		panic("uvm_pageboot_alloc: out of virtual space");
563 
564 	addr = virtual_space_start;
565 
566 #ifdef PMAP_GROWKERNEL
567 	/*
568 	 * If the kernel pmap can't map the requested space,
569 	 * then allocate more resources for it.
570 	 */
571 	if (uvm_maxkaddr < (addr + size)) {
572 		uvm_maxkaddr = pmap_growkernel(addr + size);
573 		if (uvm_maxkaddr < (addr + size))
574 			panic("uvm_pageboot_alloc: pmap_growkernel() failed");
575 	}
576 #endif
577 
578 	virtual_space_start += size;
579 
580 	/*
581 	 * allocate and mapin physical pages to back new virtual pages
582 	 */
583 
584 	for (vaddr = round_page(addr) ; vaddr < addr + size ;
585 	    vaddr += PAGE_SIZE) {
586 
587 		if (!uvm_page_physget(&paddr))
588 			panic("uvm_pageboot_alloc: out of memory");
589 
590 		/*
591 		 * Note this memory is no longer managed, so using
592 		 * pmap_kenter is safe.
593 		 */
594 		pmap_kenter_pa(vaddr, paddr, VM_PROT_READ|VM_PROT_WRITE, 0);
595 	}
596 	pmap_update(pmap_kernel());
597 	return(addr);
598 #endif	/* PMAP_STEAL_MEMORY */
599 }
600 
601 #if !defined(PMAP_STEAL_MEMORY)
602 /*
603  * uvm_page_physget: "steal" one page from the vm_physmem structure.
604  *
605  * => attempt to allocate it off the end of a segment in which the "avail"
606  *    values match the start/end values.   if we can't do that, then we
607  *    will advance both values (making them equal, and removing some
608  *    vm_page structures from the non-avail area).
609  * => return false if out of memory.
610  */
611 
612 /* subroutine: try to allocate from memory chunks on the specified freelist */
613 static bool uvm_page_physget_freelist(paddr_t *, int);
614 
615 static bool
616 uvm_page_physget_freelist(paddr_t *paddrp, int freelist)
617 {
618 	uvm_physseg_t lcv;
619 
620 	/* pass 1: try allocating from a matching end */
621 #if (VM_PHYSSEG_STRAT == VM_PSTRAT_BIGFIRST)
622 	for (lcv = uvm_physseg_get_last(); uvm_physseg_valid_p(lcv); lcv = uvm_physseg_get_prev(lcv))
623 #else
624 	for (lcv = uvm_physseg_get_first(); uvm_physseg_valid_p(lcv); lcv = uvm_physseg_get_next(lcv))
625 #endif
626 	{
627 		if (uvm.page_init_done == true)
628 			panic("uvm_page_physget: called _after_ bootstrap");
629 
630 		/* Try to match at front or back on unused segment */
631 		if (uvm_page_physunload(lcv, freelist, paddrp) == false) {
632 			if (paddrp == NULL) /* freelist fail, try next */
633 				continue;
634 		} else
635 			return true;
636 	}
637 
638 	/* pass2: forget about matching ends, just allocate something */
639 #if (VM_PHYSSEG_STRAT == VM_PSTRAT_BIGFIRST)
640 	for (lcv = uvm_physseg_get_last(); uvm_physseg_valid_p(lcv); lcv = uvm_physseg_get_prev(lcv))
641 #else
642 	for (lcv = uvm_physseg_get_first(); uvm_physseg_valid_p(lcv); lcv = uvm_physseg_get_next(lcv))
643 #endif
644 	{
645 		/* Try the front regardless. */
646 		if (uvm_page_physunload_force(lcv, freelist, paddrp) == false) {
647 			if (paddrp == NULL) /* freelist fail, try next */
648 				continue;
649 		} else
650 			return true;
651 	}
652 	return false;
653 }
654 
655 bool
656 uvm_page_physget(paddr_t *paddrp)
657 {
658 	int i;
659 
660 	/* try in the order of freelist preference */
661 	for (i = 0; i < VM_NFREELIST; i++)
662 		if (uvm_page_physget_freelist(paddrp, i) == true)
663 			return (true);
664 	return (false);
665 }
666 #endif /* PMAP_STEAL_MEMORY */
667 
668 /*
669  * PHYS_TO_VM_PAGE: find vm_page for a PA.   used by MI code to get vm_pages
670  * back from an I/O mapping (ugh!).   used in some MD code as well.
671  */
672 struct vm_page *
673 uvm_phys_to_vm_page(paddr_t pa)
674 {
675 	paddr_t pf = atop(pa);
676 	paddr_t	off;
677 	uvm_physseg_t	upm;
678 
679 	upm = uvm_physseg_find(pf, &off);
680 	if (upm != UVM_PHYSSEG_TYPE_INVALID)
681 		return uvm_physseg_get_pg(upm, off);
682 	return(NULL);
683 }
684 
685 paddr_t
686 uvm_vm_page_to_phys(const struct vm_page *pg)
687 {
688 
689 	return pg->phys_addr;
690 }
691 
692 /*
693  * uvm_page_recolor: Recolor the pages if the new bucket count is
694  * larger than the old one.
695  */
696 
697 void
698 uvm_page_recolor(int newncolors)
699 {
700 	struct pgflbucket *bucketarray, *cpuarray, *oldbucketarray;
701 	struct pgfreelist gpgfl, pgfl;
702 	struct vm_page *pg;
703 	vsize_t bucketcount;
704 	size_t bucketmemsize, oldbucketmemsize;
705 	int color, i, ocolors;
706 	int lcv;
707 	struct uvm_cpu *ucpu;
708 
709 	KASSERT(((newncolors - 1) & newncolors) == 0);
710 
711 	if (newncolors <= uvmexp.ncolors)
712 		return;
713 
714 	if (uvm.page_init_done == false) {
715 		uvmexp.ncolors = newncolors;
716 		return;
717 	}
718 
719 	bucketcount = newncolors * VM_NFREELIST;
720 	bucketmemsize = bucketcount * sizeof(struct pgflbucket) * 2;
721 	bucketarray = kmem_alloc(bucketmemsize, KM_SLEEP);
722 	cpuarray = bucketarray + bucketcount;
723 
724 	mutex_spin_enter(&uvm_fpageqlock);
725 
726 	/* Make sure we should still do this. */
727 	if (newncolors <= uvmexp.ncolors) {
728 		mutex_spin_exit(&uvm_fpageqlock);
729 		kmem_free(bucketarray, bucketmemsize);
730 		return;
731 	}
732 
733 	oldbucketarray = uvm.page_free[0].pgfl_buckets;
734 	ocolors = uvmexp.ncolors;
735 
736 	uvmexp.ncolors = newncolors;
737 	uvmexp.colormask = uvmexp.ncolors - 1;
738 
739 	ucpu = curcpu()->ci_data.cpu_uvm;
740 	for (lcv = 0; lcv < VM_NFREELIST; lcv++) {
741 		gpgfl.pgfl_buckets = (bucketarray + (lcv * newncolors));
742 		pgfl.pgfl_buckets = (cpuarray + (lcv * uvmexp.ncolors));
743 		uvm_page_init_buckets(&gpgfl);
744 		uvm_page_init_buckets(&pgfl);
745 		for (color = 0; color < ocolors; color++) {
746 			for (i = 0; i < PGFL_NQUEUES; i++) {
747 				while ((pg = LIST_FIRST(&uvm.page_free[
748 				    lcv].pgfl_buckets[color].pgfl_queues[i]))
749 				    != NULL) {
750 					LIST_REMOVE(pg, pageq.list); /* global */
751 					LIST_REMOVE(pg, listq.list); /* cpu */
752 					LIST_INSERT_HEAD(&gpgfl.pgfl_buckets[
753 					    VM_PGCOLOR_BUCKET(pg)].pgfl_queues[
754 					    i], pg, pageq.list);
755 					LIST_INSERT_HEAD(&pgfl.pgfl_buckets[
756 					    VM_PGCOLOR_BUCKET(pg)].pgfl_queues[
757 					    i], pg, listq.list);
758 				}
759 			}
760 		}
761 		uvm.page_free[lcv].pgfl_buckets = gpgfl.pgfl_buckets;
762 		ucpu->page_free[lcv].pgfl_buckets = pgfl.pgfl_buckets;
763 	}
764 
765 	oldbucketmemsize = recolored_pages_memsize;
766 
767 	recolored_pages_memsize = bucketmemsize;
768 	mutex_spin_exit(&uvm_fpageqlock);
769 
770 	if (oldbucketmemsize) {
771 		kmem_free(oldbucketarray, recolored_pages_memsize);
772 	}
773 
774 	/*
775 	 * this calls uvm_km_alloc() which may want to hold
776 	 * uvm_fpageqlock.
777 	 */
778 	uvm_pager_realloc_emerg();
779 }
780 
781 /*
782  * uvm_cpu_attach: initialize per-CPU data structures.
783  */
784 
785 void
786 uvm_cpu_attach(struct cpu_info *ci)
787 {
788 	struct pgflbucket *bucketarray;
789 	struct pgfreelist pgfl;
790 	struct uvm_cpu *ucpu;
791 	vsize_t bucketcount;
792 	int lcv;
793 
794 	if (CPU_IS_PRIMARY(ci)) {
795 		/* Already done in uvm_page_init(). */
796 		goto attachrnd;
797 	}
798 
799 	/* Add more reserve pages for this CPU. */
800 	uvmexp.reserve_kernel += vm_page_reserve_kernel;
801 
802 	/* Configure this CPU's free lists. */
803 	bucketcount = uvmexp.ncolors * VM_NFREELIST;
804 	bucketarray = kmem_alloc(bucketcount * sizeof(struct pgflbucket),
805 	    KM_SLEEP);
806 	ucpu = kmem_zalloc(sizeof(*ucpu), KM_SLEEP);
807 	uvm.cpus[cpu_index(ci)] = ucpu;
808 	ci->ci_data.cpu_uvm = ucpu;
809 	for (lcv = 0; lcv < VM_NFREELIST; lcv++) {
810 		pgfl.pgfl_buckets = (bucketarray + (lcv * uvmexp.ncolors));
811 		uvm_page_init_buckets(&pgfl);
812 		ucpu->page_free[lcv].pgfl_buckets = pgfl.pgfl_buckets;
813 	}
814 
815 attachrnd:
816 	/*
817 	 * Attach RNG source for this CPU's VM events
818 	 */
819         rnd_attach_source(&uvm.cpus[cpu_index(ci)]->rs,
820 			  ci->ci_data.cpu_name, RND_TYPE_VM,
821 			  RND_FLAG_COLLECT_TIME|RND_FLAG_COLLECT_VALUE|
822 			  RND_FLAG_ESTIMATE_VALUE);
823 
824 }
825 
826 /*
827  * uvm_pagealloc_pgfl: helper routine for uvm_pagealloc_strat
828  */
829 
830 static struct vm_page *
831 uvm_pagealloc_pgfl(struct uvm_cpu *ucpu, int flist, int try1, int try2,
832     int *trycolorp)
833 {
834 	struct pgflist *freeq;
835 	struct vm_page *pg;
836 	int color, trycolor = *trycolorp;
837 	struct pgfreelist *gpgfl, *pgfl;
838 
839 	KASSERT(mutex_owned(&uvm_fpageqlock));
840 
841 	color = trycolor;
842 	pgfl = &ucpu->page_free[flist];
843 	gpgfl = &uvm.page_free[flist];
844 	do {
845 		/* cpu, try1 */
846 		if ((pg = LIST_FIRST((freeq =
847 		    &pgfl->pgfl_buckets[color].pgfl_queues[try1]))) != NULL) {
848 			KASSERT(pg->pqflags & PQ_FREE);
849 			KASSERT(try1 == PGFL_ZEROS || !(pg->flags & PG_ZERO));
850 			KASSERT(try1 == PGFL_UNKNOWN || (pg->flags & PG_ZERO));
851 			KASSERT(ucpu == VM_FREE_PAGE_TO_CPU(pg));
852 			VM_FREE_PAGE_TO_CPU(pg)->pages[try1]--;
853 		    	uvmexp.cpuhit++;
854 			goto gotit;
855 		}
856 		/* global, try1 */
857 		if ((pg = LIST_FIRST((freeq =
858 		    &gpgfl->pgfl_buckets[color].pgfl_queues[try1]))) != NULL) {
859 			KASSERT(pg->pqflags & PQ_FREE);
860 			KASSERT(try1 == PGFL_ZEROS || !(pg->flags & PG_ZERO));
861 			KASSERT(try1 == PGFL_UNKNOWN || (pg->flags & PG_ZERO));
862 			KASSERT(ucpu != VM_FREE_PAGE_TO_CPU(pg));
863 			VM_FREE_PAGE_TO_CPU(pg)->pages[try1]--;
864 		    	uvmexp.cpumiss++;
865 			goto gotit;
866 		}
867 		/* cpu, try2 */
868 		if ((pg = LIST_FIRST((freeq =
869 		    &pgfl->pgfl_buckets[color].pgfl_queues[try2]))) != NULL) {
870 			KASSERT(pg->pqflags & PQ_FREE);
871 			KASSERT(try2 == PGFL_ZEROS || !(pg->flags & PG_ZERO));
872 			KASSERT(try2 == PGFL_UNKNOWN || (pg->flags & PG_ZERO));
873 			KASSERT(ucpu == VM_FREE_PAGE_TO_CPU(pg));
874 			VM_FREE_PAGE_TO_CPU(pg)->pages[try2]--;
875 		    	uvmexp.cpuhit++;
876 			goto gotit;
877 		}
878 		/* global, try2 */
879 		if ((pg = LIST_FIRST((freeq =
880 		    &gpgfl->pgfl_buckets[color].pgfl_queues[try2]))) != NULL) {
881 			KASSERT(pg->pqflags & PQ_FREE);
882 			KASSERT(try2 == PGFL_ZEROS || !(pg->flags & PG_ZERO));
883 			KASSERT(try2 == PGFL_UNKNOWN || (pg->flags & PG_ZERO));
884 			KASSERT(ucpu != VM_FREE_PAGE_TO_CPU(pg));
885 			VM_FREE_PAGE_TO_CPU(pg)->pages[try2]--;
886 		    	uvmexp.cpumiss++;
887 			goto gotit;
888 		}
889 		color = (color + 1) & uvmexp.colormask;
890 	} while (color != trycolor);
891 
892 	return (NULL);
893 
894  gotit:
895 	LIST_REMOVE(pg, pageq.list);	/* global list */
896 	LIST_REMOVE(pg, listq.list);	/* per-cpu list */
897 	uvmexp.free--;
898 
899 	/* update zero'd page count */
900 	if (pg->flags & PG_ZERO)
901 		uvmexp.zeropages--;
902 
903 	if (color == trycolor)
904 		uvmexp.colorhit++;
905 	else {
906 		uvmexp.colormiss++;
907 		*trycolorp = color;
908 	}
909 
910 	return (pg);
911 }
912 
913 /*
914  * uvm_pagealloc_strat: allocate vm_page from a particular free list.
915  *
916  * => return null if no pages free
917  * => wake up pagedaemon if number of free pages drops below low water mark
918  * => if obj != NULL, obj must be locked (to put in obj's tree)
919  * => if anon != NULL, anon must be locked (to put in anon)
920  * => only one of obj or anon can be non-null
921  * => caller must activate/deactivate page if it is not wired.
922  * => free_list is ignored if strat == UVM_PGA_STRAT_NORMAL.
923  * => policy decision: it is more important to pull a page off of the
924  *	appropriate priority free list than it is to get a zero'd or
925  *	unknown contents page.  This is because we live with the
926  *	consequences of a bad free list decision for the entire
927  *	lifetime of the page, e.g. if the page comes from memory that
928  *	is slower to access.
929  */
930 
931 struct vm_page *
932 uvm_pagealloc_strat(struct uvm_object *obj, voff_t off, struct vm_anon *anon,
933     int flags, int strat, int free_list)
934 {
935 	int try1, try2, zeroit = 0, color;
936 	int lcv;
937 	struct uvm_cpu *ucpu;
938 	struct vm_page *pg;
939 	lwp_t *l;
940 
941 	KASSERT(obj == NULL || anon == NULL);
942 	KASSERT(anon == NULL || (flags & UVM_FLAG_COLORMATCH) || off == 0);
943 	KASSERT(off == trunc_page(off));
944 	KASSERT(obj == NULL || mutex_owned(obj->vmobjlock));
945 	KASSERT(anon == NULL || anon->an_lock == NULL ||
946 	    mutex_owned(anon->an_lock));
947 
948 	mutex_spin_enter(&uvm_fpageqlock);
949 
950 	/*
951 	 * This implements a global round-robin page coloring
952 	 * algorithm.
953 	 */
954 
955 	ucpu = curcpu()->ci_data.cpu_uvm;
956 	if (flags & UVM_FLAG_COLORMATCH) {
957 		color = atop(off) & uvmexp.colormask;
958 	} else {
959 		color = ucpu->page_free_nextcolor;
960 	}
961 
962 	/*
963 	 * check to see if we need to generate some free pages waking
964 	 * the pagedaemon.
965 	 */
966 
967 	uvm_kick_pdaemon();
968 
969 	/*
970 	 * fail if any of these conditions is true:
971 	 * [1]  there really are no free pages, or
972 	 * [2]  only kernel "reserved" pages remain and
973 	 *        reserved pages have not been requested.
974 	 * [3]  only pagedaemon "reserved" pages remain and
975 	 *        the requestor isn't the pagedaemon.
976 	 * we make kernel reserve pages available if called by a
977 	 * kernel thread or a realtime thread.
978 	 */
979 	l = curlwp;
980 	if (__predict_true(l != NULL) && lwp_eprio(l) >= PRI_KTHREAD) {
981 		flags |= UVM_PGA_USERESERVE;
982 	}
983 	if ((uvmexp.free <= uvmexp.reserve_kernel &&
984 	    (flags & UVM_PGA_USERESERVE) == 0) ||
985 	    (uvmexp.free <= uvmexp.reserve_pagedaemon &&
986 	     curlwp != uvm.pagedaemon_lwp))
987 		goto fail;
988 
989 #if PGFL_NQUEUES != 2
990 #error uvm_pagealloc_strat needs to be updated
991 #endif
992 
993 	/*
994 	 * If we want a zero'd page, try the ZEROS queue first, otherwise
995 	 * we try the UNKNOWN queue first.
996 	 */
997 	if (flags & UVM_PGA_ZERO) {
998 		try1 = PGFL_ZEROS;
999 		try2 = PGFL_UNKNOWN;
1000 	} else {
1001 		try1 = PGFL_UNKNOWN;
1002 		try2 = PGFL_ZEROS;
1003 	}
1004 
1005  again:
1006 	switch (strat) {
1007 	case UVM_PGA_STRAT_NORMAL:
1008 		/* Check freelists: descending priority (ascending id) order */
1009 		for (lcv = 0; lcv < VM_NFREELIST; lcv++) {
1010 			pg = uvm_pagealloc_pgfl(ucpu, lcv,
1011 			    try1, try2, &color);
1012 			if (pg != NULL)
1013 				goto gotit;
1014 		}
1015 
1016 		/* No pages free! */
1017 		goto fail;
1018 
1019 	case UVM_PGA_STRAT_ONLY:
1020 	case UVM_PGA_STRAT_FALLBACK:
1021 		/* Attempt to allocate from the specified free list. */
1022 		KASSERT(free_list >= 0 && free_list < VM_NFREELIST);
1023 		pg = uvm_pagealloc_pgfl(ucpu, free_list,
1024 		    try1, try2, &color);
1025 		if (pg != NULL)
1026 			goto gotit;
1027 
1028 		/* Fall back, if possible. */
1029 		if (strat == UVM_PGA_STRAT_FALLBACK) {
1030 			strat = UVM_PGA_STRAT_NORMAL;
1031 			goto again;
1032 		}
1033 
1034 		/* No pages free! */
1035 		goto fail;
1036 
1037 	default:
1038 		panic("uvm_pagealloc_strat: bad strat %d", strat);
1039 		/* NOTREACHED */
1040 	}
1041 
1042  gotit:
1043 	/*
1044 	 * We now know which color we actually allocated from; set
1045 	 * the next color accordingly.
1046 	 */
1047 
1048 	ucpu->page_free_nextcolor = (color + 1) & uvmexp.colormask;
1049 
1050 	/*
1051 	 * update allocation statistics and remember if we have to
1052 	 * zero the page
1053 	 */
1054 
1055 	if (flags & UVM_PGA_ZERO) {
1056 		if (pg->flags & PG_ZERO) {
1057 			uvmexp.pga_zerohit++;
1058 			zeroit = 0;
1059 		} else {
1060 			uvmexp.pga_zeromiss++;
1061 			zeroit = 1;
1062 		}
1063 		if (ucpu->pages[PGFL_ZEROS] < ucpu->pages[PGFL_UNKNOWN]) {
1064 			ucpu->page_idle_zero = vm_page_zero_enable;
1065 		}
1066 	}
1067 	KASSERT(pg->pqflags == PQ_FREE);
1068 
1069 	pg->offset = off;
1070 	pg->uobject = obj;
1071 	pg->uanon = anon;
1072 	pg->flags = PG_BUSY|PG_CLEAN|PG_FAKE;
1073 	if (anon) {
1074 		anon->an_page = pg;
1075 		pg->pqflags = PQ_ANON;
1076 		atomic_inc_uint(&uvmexp.anonpages);
1077 	} else {
1078 		if (obj) {
1079 			uvm_pageinsert(obj, pg);
1080 		}
1081 		pg->pqflags = 0;
1082 	}
1083 	mutex_spin_exit(&uvm_fpageqlock);
1084 
1085 #if defined(UVM_PAGE_TRKOWN)
1086 	pg->owner_tag = NULL;
1087 #endif
1088 	UVM_PAGE_OWN(pg, "new alloc");
1089 
1090 	if (flags & UVM_PGA_ZERO) {
1091 		/*
1092 		 * A zero'd page is not clean.  If we got a page not already
1093 		 * zero'd, then we have to zero it ourselves.
1094 		 */
1095 		pg->flags &= ~PG_CLEAN;
1096 		if (zeroit)
1097 			pmap_zero_page(VM_PAGE_TO_PHYS(pg));
1098 	}
1099 
1100 	return(pg);
1101 
1102  fail:
1103 	mutex_spin_exit(&uvm_fpageqlock);
1104 	return (NULL);
1105 }
1106 
1107 /*
1108  * uvm_pagereplace: replace a page with another
1109  *
1110  * => object must be locked
1111  */
1112 
1113 void
1114 uvm_pagereplace(struct vm_page *oldpg, struct vm_page *newpg)
1115 {
1116 	struct uvm_object *uobj = oldpg->uobject;
1117 
1118 	KASSERT((oldpg->flags & PG_TABLED) != 0);
1119 	KASSERT(uobj != NULL);
1120 	KASSERT((newpg->flags & PG_TABLED) == 0);
1121 	KASSERT(newpg->uobject == NULL);
1122 	KASSERT(mutex_owned(uobj->vmobjlock));
1123 
1124 	newpg->uobject = uobj;
1125 	newpg->offset = oldpg->offset;
1126 
1127 	uvm_pageremove_tree(uobj, oldpg);
1128 	uvm_pageinsert_tree(uobj, newpg);
1129 	uvm_pageinsert_list(uobj, newpg, oldpg);
1130 	uvm_pageremove_list(uobj, oldpg);
1131 }
1132 
1133 /*
1134  * uvm_pagerealloc: reallocate a page from one object to another
1135  *
1136  * => both objects must be locked
1137  */
1138 
1139 void
1140 uvm_pagerealloc(struct vm_page *pg, struct uvm_object *newobj, voff_t newoff)
1141 {
1142 	/*
1143 	 * remove it from the old object
1144 	 */
1145 
1146 	if (pg->uobject) {
1147 		uvm_pageremove(pg->uobject, pg);
1148 	}
1149 
1150 	/*
1151 	 * put it in the new object
1152 	 */
1153 
1154 	if (newobj) {
1155 		pg->uobject = newobj;
1156 		pg->offset = newoff;
1157 		uvm_pageinsert(newobj, pg);
1158 	}
1159 }
1160 
1161 #ifdef DEBUG
1162 /*
1163  * check if page is zero-filled
1164  *
1165  *  - called with free page queue lock held.
1166  */
1167 void
1168 uvm_pagezerocheck(struct vm_page *pg)
1169 {
1170 	int *p, *ep;
1171 
1172 	KASSERT(uvm_zerocheckkva != 0);
1173 	KASSERT(mutex_owned(&uvm_fpageqlock));
1174 
1175 	/*
1176 	 * XXX assuming pmap_kenter_pa and pmap_kremove never call
1177 	 * uvm page allocator.
1178 	 *
1179 	 * it might be better to have "CPU-local temporary map" pmap interface.
1180 	 */
1181 	pmap_kenter_pa(uvm_zerocheckkva, VM_PAGE_TO_PHYS(pg), VM_PROT_READ, 0);
1182 	p = (int *)uvm_zerocheckkva;
1183 	ep = (int *)((char *)p + PAGE_SIZE);
1184 	pmap_update(pmap_kernel());
1185 	while (p < ep) {
1186 		if (*p != 0)
1187 			panic("PG_ZERO page isn't zero-filled");
1188 		p++;
1189 	}
1190 	pmap_kremove(uvm_zerocheckkva, PAGE_SIZE);
1191 	/*
1192 	 * pmap_update() is not necessary here because no one except us
1193 	 * uses this VA.
1194 	 */
1195 }
1196 #endif /* DEBUG */
1197 
1198 /*
1199  * uvm_pagefree: free page
1200  *
1201  * => erase page's identity (i.e. remove from object)
1202  * => put page on free list
1203  * => caller must lock owning object (either anon or uvm_object)
1204  * => caller must lock page queues
1205  * => assumes all valid mappings of pg are gone
1206  */
1207 
1208 void
1209 uvm_pagefree(struct vm_page *pg)
1210 {
1211 	struct pgflist *pgfl;
1212 	struct uvm_cpu *ucpu;
1213 	int index, color, queue;
1214 	bool iszero;
1215 
1216 #ifdef DEBUG
1217 	if (pg->uobject == (void *)0xdeadbeef &&
1218 	    pg->uanon == (void *)0xdeadbeef) {
1219 		panic("uvm_pagefree: freeing free page %p", pg);
1220 	}
1221 #endif /* DEBUG */
1222 
1223 	KASSERT((pg->flags & PG_PAGEOUT) == 0);
1224 	KASSERT(!(pg->pqflags & PQ_FREE));
1225 	//KASSERT(mutex_owned(&uvm_pageqlock) || !uvmpdpol_pageisqueued_p(pg));
1226 	KASSERT(pg->uobject == NULL || mutex_owned(pg->uobject->vmobjlock));
1227 	KASSERT(pg->uobject != NULL || pg->uanon == NULL ||
1228 		mutex_owned(pg->uanon->an_lock));
1229 
1230 	/*
1231 	 * if the page is loaned, resolve the loan instead of freeing.
1232 	 */
1233 
1234 	if (pg->loan_count) {
1235 		KASSERT(pg->wire_count == 0);
1236 
1237 		/*
1238 		 * if the page is owned by an anon then we just want to
1239 		 * drop anon ownership.  the kernel will free the page when
1240 		 * it is done with it.  if the page is owned by an object,
1241 		 * remove it from the object and mark it dirty for the benefit
1242 		 * of possible anon owners.
1243 		 *
1244 		 * regardless of previous ownership, wakeup any waiters,
1245 		 * unbusy the page, and we're done.
1246 		 */
1247 
1248 		if (pg->uobject != NULL) {
1249 			uvm_pageremove(pg->uobject, pg);
1250 			pg->flags &= ~PG_CLEAN;
1251 		} else if (pg->uanon != NULL) {
1252 			if ((pg->pqflags & PQ_ANON) == 0) {
1253 				pg->loan_count--;
1254 			} else {
1255 				pg->pqflags &= ~PQ_ANON;
1256 				atomic_dec_uint(&uvmexp.anonpages);
1257 			}
1258 			pg->uanon->an_page = NULL;
1259 			pg->uanon = NULL;
1260 		}
1261 		if (pg->flags & PG_WANTED) {
1262 			wakeup(pg);
1263 		}
1264 		pg->flags &= ~(PG_WANTED|PG_BUSY|PG_RELEASED|PG_PAGER1);
1265 #ifdef UVM_PAGE_TRKOWN
1266 		pg->owner_tag = NULL;
1267 #endif
1268 		if (pg->loan_count) {
1269 			KASSERT(pg->uobject == NULL);
1270 			if (pg->uanon == NULL) {
1271 				KASSERT(mutex_owned(&uvm_pageqlock));
1272 				uvm_pagedequeue(pg);
1273 			}
1274 			return;
1275 		}
1276 	}
1277 
1278 	/*
1279 	 * remove page from its object or anon.
1280 	 */
1281 
1282 	if (pg->uobject != NULL) {
1283 		uvm_pageremove(pg->uobject, pg);
1284 	} else if (pg->uanon != NULL) {
1285 		pg->uanon->an_page = NULL;
1286 		atomic_dec_uint(&uvmexp.anonpages);
1287 	}
1288 
1289 	/*
1290 	 * now remove the page from the queues.
1291 	 */
1292 	if (uvmpdpol_pageisqueued_p(pg)) {
1293 		KASSERT(mutex_owned(&uvm_pageqlock));
1294 		uvm_pagedequeue(pg);
1295 	}
1296 
1297 	/*
1298 	 * if the page was wired, unwire it now.
1299 	 */
1300 
1301 	if (pg->wire_count) {
1302 		pg->wire_count = 0;
1303 		uvmexp.wired--;
1304 	}
1305 
1306 	/*
1307 	 * and put on free queue
1308 	 */
1309 
1310 	iszero = (pg->flags & PG_ZERO);
1311 	index = uvm_page_lookup_freelist(pg);
1312 	color = VM_PGCOLOR_BUCKET(pg);
1313 	queue = (iszero ? PGFL_ZEROS : PGFL_UNKNOWN);
1314 
1315 #ifdef DEBUG
1316 	pg->uobject = (void *)0xdeadbeef;
1317 	pg->uanon = (void *)0xdeadbeef;
1318 #endif
1319 
1320 	mutex_spin_enter(&uvm_fpageqlock);
1321 	pg->pqflags = PQ_FREE;
1322 
1323 #ifdef DEBUG
1324 	if (iszero)
1325 		uvm_pagezerocheck(pg);
1326 #endif /* DEBUG */
1327 
1328 
1329 	/* global list */
1330 	pgfl = &uvm.page_free[index].pgfl_buckets[color].pgfl_queues[queue];
1331 	LIST_INSERT_HEAD(pgfl, pg, pageq.list);
1332 	uvmexp.free++;
1333 	if (iszero) {
1334 		uvmexp.zeropages++;
1335 	}
1336 
1337 	/* per-cpu list */
1338 	ucpu = curcpu()->ci_data.cpu_uvm;
1339 	pg->offset = (uintptr_t)ucpu;
1340 	pgfl = &ucpu->page_free[index].pgfl_buckets[color].pgfl_queues[queue];
1341 	LIST_INSERT_HEAD(pgfl, pg, listq.list);
1342 	ucpu->pages[queue]++;
1343 	if (ucpu->pages[PGFL_ZEROS] < ucpu->pages[PGFL_UNKNOWN]) {
1344 		ucpu->page_idle_zero = vm_page_zero_enable;
1345 	}
1346 
1347 	mutex_spin_exit(&uvm_fpageqlock);
1348 }
1349 
1350 /*
1351  * uvm_page_unbusy: unbusy an array of pages.
1352  *
1353  * => pages must either all belong to the same object, or all belong to anons.
1354  * => if pages are object-owned, object must be locked.
1355  * => if pages are anon-owned, anons must be locked.
1356  * => caller must lock page queues if pages may be released.
1357  * => caller must make sure that anon-owned pages are not PG_RELEASED.
1358  */
1359 
1360 void
1361 uvm_page_unbusy(struct vm_page **pgs, int npgs)
1362 {
1363 	struct vm_page *pg;
1364 	int i;
1365 	UVMHIST_FUNC("uvm_page_unbusy"); UVMHIST_CALLED(ubchist);
1366 
1367 	for (i = 0; i < npgs; i++) {
1368 		pg = pgs[i];
1369 		if (pg == NULL || pg == PGO_DONTCARE) {
1370 			continue;
1371 		}
1372 
1373 		KASSERT(uvm_page_locked_p(pg));
1374 		KASSERT(pg->flags & PG_BUSY);
1375 		KASSERT((pg->flags & PG_PAGEOUT) == 0);
1376 		if (pg->flags & PG_WANTED) {
1377 			wakeup(pg);
1378 		}
1379 		if (pg->flags & PG_RELEASED) {
1380 			UVMHIST_LOG(ubchist, "releasing pg %#jx",
1381 			    (uintptr_t)pg, 0, 0, 0);
1382 			KASSERT(pg->uobject != NULL ||
1383 			    (pg->uanon != NULL && pg->uanon->an_ref > 0));
1384 			pg->flags &= ~PG_RELEASED;
1385 			uvm_pagefree(pg);
1386 		} else {
1387 			UVMHIST_LOG(ubchist, "unbusying pg %#jx",
1388 			    (uintptr_t)pg, 0, 0, 0);
1389 			KASSERT((pg->flags & PG_FAKE) == 0);
1390 			pg->flags &= ~(PG_WANTED|PG_BUSY);
1391 			UVM_PAGE_OWN(pg, NULL);
1392 		}
1393 	}
1394 }
1395 
1396 #if defined(UVM_PAGE_TRKOWN)
1397 /*
1398  * uvm_page_own: set or release page ownership
1399  *
1400  * => this is a debugging function that keeps track of who sets PG_BUSY
1401  *	and where they do it.   it can be used to track down problems
1402  *	such a process setting "PG_BUSY" and never releasing it.
1403  * => page's object [if any] must be locked
1404  * => if "tag" is NULL then we are releasing page ownership
1405  */
1406 void
1407 uvm_page_own(struct vm_page *pg, const char *tag)
1408 {
1409 
1410 	KASSERT((pg->flags & (PG_PAGEOUT|PG_RELEASED)) == 0);
1411 	KASSERT((pg->flags & PG_WANTED) == 0);
1412 	KASSERT(uvm_page_locked_p(pg));
1413 
1414 	/* gain ownership? */
1415 	if (tag) {
1416 		KASSERT((pg->flags & PG_BUSY) != 0);
1417 		if (pg->owner_tag) {
1418 			printf("uvm_page_own: page %p already owned "
1419 			    "by proc %d [%s]\n", pg,
1420 			    pg->owner, pg->owner_tag);
1421 			panic("uvm_page_own");
1422 		}
1423 		pg->owner = curproc->p_pid;
1424 		pg->lowner = curlwp->l_lid;
1425 		pg->owner_tag = tag;
1426 		return;
1427 	}
1428 
1429 	/* drop ownership */
1430 	KASSERT((pg->flags & PG_BUSY) == 0);
1431 	if (pg->owner_tag == NULL) {
1432 		printf("uvm_page_own: dropping ownership of an non-owned "
1433 		    "page (%p)\n", pg);
1434 		panic("uvm_page_own");
1435 	}
1436 	if (!uvmpdpol_pageisqueued_p(pg)) {
1437 		KASSERT((pg->uanon == NULL && pg->uobject == NULL) ||
1438 		    pg->wire_count > 0);
1439 	} else {
1440 		KASSERT(pg->wire_count == 0);
1441 	}
1442 	pg->owner_tag = NULL;
1443 }
1444 #endif
1445 
1446 /*
1447  * uvm_pageidlezero: zero free pages while the system is idle.
1448  *
1449  * => try to complete one color bucket at a time, to reduce our impact
1450  *	on the CPU cache.
1451  * => we loop until we either reach the target or there is a lwp ready
1452  *      to run, or MD code detects a reason to break early.
1453  */
1454 void
1455 uvm_pageidlezero(void)
1456 {
1457 	struct vm_page *pg;
1458 	struct pgfreelist *pgfl, *gpgfl;
1459 	struct uvm_cpu *ucpu;
1460 	int free_list, firstbucket, nextbucket;
1461 	bool lcont = false;
1462 
1463 	ucpu = curcpu()->ci_data.cpu_uvm;
1464 	if (!ucpu->page_idle_zero ||
1465 	    ucpu->pages[PGFL_UNKNOWN] < uvmexp.ncolors) {
1466 	    	ucpu->page_idle_zero = false;
1467 		return;
1468 	}
1469 	if (!mutex_tryenter(&uvm_fpageqlock)) {
1470 		/* Contention: let other CPUs to use the lock. */
1471 		return;
1472 	}
1473 	firstbucket = ucpu->page_free_nextcolor;
1474 	nextbucket = firstbucket;
1475 	do {
1476 		for (free_list = 0; free_list < VM_NFREELIST; free_list++) {
1477 			if (sched_curcpu_runnable_p()) {
1478 				goto quit;
1479 			}
1480 			pgfl = &ucpu->page_free[free_list];
1481 			gpgfl = &uvm.page_free[free_list];
1482 			while ((pg = LIST_FIRST(&pgfl->pgfl_buckets[
1483 			    nextbucket].pgfl_queues[PGFL_UNKNOWN])) != NULL) {
1484 				if (lcont || sched_curcpu_runnable_p()) {
1485 					goto quit;
1486 				}
1487 				LIST_REMOVE(pg, pageq.list); /* global list */
1488 				LIST_REMOVE(pg, listq.list); /* per-cpu list */
1489 				ucpu->pages[PGFL_UNKNOWN]--;
1490 				uvmexp.free--;
1491 				KASSERT(pg->pqflags == PQ_FREE);
1492 				pg->pqflags = 0;
1493 				mutex_spin_exit(&uvm_fpageqlock);
1494 #ifdef PMAP_PAGEIDLEZERO
1495 				if (!PMAP_PAGEIDLEZERO(VM_PAGE_TO_PHYS(pg))) {
1496 
1497 					/*
1498 					 * The machine-dependent code detected
1499 					 * some reason for us to abort zeroing
1500 					 * pages, probably because there is a
1501 					 * process now ready to run.
1502 					 */
1503 
1504 					mutex_spin_enter(&uvm_fpageqlock);
1505 					pg->pqflags = PQ_FREE;
1506 					LIST_INSERT_HEAD(&gpgfl->pgfl_buckets[
1507 					    nextbucket].pgfl_queues[
1508 					    PGFL_UNKNOWN], pg, pageq.list);
1509 					LIST_INSERT_HEAD(&pgfl->pgfl_buckets[
1510 					    nextbucket].pgfl_queues[
1511 					    PGFL_UNKNOWN], pg, listq.list);
1512 					ucpu->pages[PGFL_UNKNOWN]++;
1513 					uvmexp.free++;
1514 					uvmexp.zeroaborts++;
1515 					goto quit;
1516 				}
1517 #else
1518 				pmap_zero_page(VM_PAGE_TO_PHYS(pg));
1519 #endif /* PMAP_PAGEIDLEZERO */
1520 				pg->flags |= PG_ZERO;
1521 
1522 				if (!mutex_tryenter(&uvm_fpageqlock)) {
1523 					lcont = true;
1524 					mutex_spin_enter(&uvm_fpageqlock);
1525 				} else {
1526 					lcont = false;
1527 				}
1528 				pg->pqflags = PQ_FREE;
1529 				LIST_INSERT_HEAD(&gpgfl->pgfl_buckets[
1530 				    nextbucket].pgfl_queues[PGFL_ZEROS],
1531 				    pg, pageq.list);
1532 				LIST_INSERT_HEAD(&pgfl->pgfl_buckets[
1533 				    nextbucket].pgfl_queues[PGFL_ZEROS],
1534 				    pg, listq.list);
1535 				ucpu->pages[PGFL_ZEROS]++;
1536 				uvmexp.free++;
1537 				uvmexp.zeropages++;
1538 			}
1539 		}
1540 		if (ucpu->pages[PGFL_UNKNOWN] < uvmexp.ncolors) {
1541 			break;
1542 		}
1543 		nextbucket = (nextbucket + 1) & uvmexp.colormask;
1544 	} while (nextbucket != firstbucket);
1545 	ucpu->page_idle_zero = false;
1546  quit:
1547 	mutex_spin_exit(&uvm_fpageqlock);
1548 }
1549 
1550 /*
1551  * uvm_pagelookup: look up a page
1552  *
1553  * => caller should lock object to keep someone from pulling the page
1554  *	out from under it
1555  */
1556 
1557 struct vm_page *
1558 uvm_pagelookup(struct uvm_object *obj, voff_t off)
1559 {
1560 	struct vm_page *pg;
1561 
1562 	KASSERT(mutex_owned(obj->vmobjlock));
1563 
1564 	pg = rb_tree_find_node(&obj->rb_tree, &off);
1565 
1566 	KASSERT(pg == NULL || obj->uo_npages != 0);
1567 	KASSERT(pg == NULL || (pg->flags & (PG_RELEASED|PG_PAGEOUT)) == 0 ||
1568 		(pg->flags & PG_BUSY) != 0);
1569 	return pg;
1570 }
1571 
1572 /*
1573  * uvm_pagewire: wire the page, thus removing it from the daemon's grasp
1574  *
1575  * => caller must lock page queues
1576  */
1577 
1578 void
1579 uvm_pagewire(struct vm_page *pg)
1580 {
1581 	KASSERT(mutex_owned(&uvm_pageqlock));
1582 #if defined(READAHEAD_STATS)
1583 	if ((pg->pqflags & PQ_READAHEAD) != 0) {
1584 		uvm_ra_hit.ev_count++;
1585 		pg->pqflags &= ~PQ_READAHEAD;
1586 	}
1587 #endif /* defined(READAHEAD_STATS) */
1588 	if (pg->wire_count == 0) {
1589 		uvm_pagedequeue(pg);
1590 		uvmexp.wired++;
1591 	}
1592 	pg->wire_count++;
1593 }
1594 
1595 /*
1596  * uvm_pageunwire: unwire the page.
1597  *
1598  * => activate if wire count goes to zero.
1599  * => caller must lock page queues
1600  */
1601 
1602 void
1603 uvm_pageunwire(struct vm_page *pg)
1604 {
1605 	KASSERT(mutex_owned(&uvm_pageqlock));
1606 	pg->wire_count--;
1607 	if (pg->wire_count == 0) {
1608 		uvm_pageactivate(pg);
1609 		uvmexp.wired--;
1610 	}
1611 }
1612 
1613 /*
1614  * uvm_pagedeactivate: deactivate page
1615  *
1616  * => caller must lock page queues
1617  * => caller must check to make sure page is not wired
1618  * => object that page belongs to must be locked (so we can adjust pg->flags)
1619  * => caller must clear the reference on the page before calling
1620  */
1621 
1622 void
1623 uvm_pagedeactivate(struct vm_page *pg)
1624 {
1625 
1626 	KASSERT(mutex_owned(&uvm_pageqlock));
1627 	KASSERT(uvm_page_locked_p(pg));
1628 	KASSERT(pg->wire_count != 0 || uvmpdpol_pageisqueued_p(pg));
1629 	uvmpdpol_pagedeactivate(pg);
1630 }
1631 
1632 /*
1633  * uvm_pageactivate: activate page
1634  *
1635  * => caller must lock page queues
1636  */
1637 
1638 void
1639 uvm_pageactivate(struct vm_page *pg)
1640 {
1641 
1642 	KASSERT(mutex_owned(&uvm_pageqlock));
1643 	KASSERT(uvm_page_locked_p(pg));
1644 #if defined(READAHEAD_STATS)
1645 	if ((pg->pqflags & PQ_READAHEAD) != 0) {
1646 		uvm_ra_hit.ev_count++;
1647 		pg->pqflags &= ~PQ_READAHEAD;
1648 	}
1649 #endif /* defined(READAHEAD_STATS) */
1650 	if (pg->wire_count != 0) {
1651 		return;
1652 	}
1653 	uvmpdpol_pageactivate(pg);
1654 }
1655 
1656 /*
1657  * uvm_pagedequeue: remove a page from any paging queue
1658  */
1659 
1660 void
1661 uvm_pagedequeue(struct vm_page *pg)
1662 {
1663 
1664 	if (uvmpdpol_pageisqueued_p(pg)) {
1665 		KASSERT(mutex_owned(&uvm_pageqlock));
1666 	}
1667 
1668 	uvmpdpol_pagedequeue(pg);
1669 }
1670 
1671 /*
1672  * uvm_pageenqueue: add a page to a paging queue without activating.
1673  * used where a page is not really demanded (yet).  eg. read-ahead
1674  */
1675 
1676 void
1677 uvm_pageenqueue(struct vm_page *pg)
1678 {
1679 
1680 	KASSERT(mutex_owned(&uvm_pageqlock));
1681 	if (pg->wire_count != 0) {
1682 		return;
1683 	}
1684 	uvmpdpol_pageenqueue(pg);
1685 }
1686 
1687 /*
1688  * uvm_pagezero: zero fill a page
1689  *
1690  * => if page is part of an object then the object should be locked
1691  *	to protect pg->flags.
1692  */
1693 
1694 void
1695 uvm_pagezero(struct vm_page *pg)
1696 {
1697 	pg->flags &= ~PG_CLEAN;
1698 	pmap_zero_page(VM_PAGE_TO_PHYS(pg));
1699 }
1700 
1701 /*
1702  * uvm_pagecopy: copy a page
1703  *
1704  * => if page is part of an object then the object should be locked
1705  *	to protect pg->flags.
1706  */
1707 
1708 void
1709 uvm_pagecopy(struct vm_page *src, struct vm_page *dst)
1710 {
1711 
1712 	dst->flags &= ~PG_CLEAN;
1713 	pmap_copy_page(VM_PAGE_TO_PHYS(src), VM_PAGE_TO_PHYS(dst));
1714 }
1715 
1716 /*
1717  * uvm_pageismanaged: test it see that a page (specified by PA) is managed.
1718  */
1719 
1720 bool
1721 uvm_pageismanaged(paddr_t pa)
1722 {
1723 
1724 	return (uvm_physseg_find(atop(pa), NULL) != UVM_PHYSSEG_TYPE_INVALID);
1725 }
1726 
1727 /*
1728  * uvm_page_lookup_freelist: look up the free list for the specified page
1729  */
1730 
1731 int
1732 uvm_page_lookup_freelist(struct vm_page *pg)
1733 {
1734 	uvm_physseg_t upm;
1735 
1736 	upm = uvm_physseg_find(atop(VM_PAGE_TO_PHYS(pg)), NULL);
1737 	KASSERT(upm != UVM_PHYSSEG_TYPE_INVALID);
1738 	return uvm_physseg_get_free_list(upm);
1739 }
1740 
1741 /*
1742  * uvm_page_locked_p: return true if object associated with page is
1743  * locked.  this is a weak check for runtime assertions only.
1744  */
1745 
1746 bool
1747 uvm_page_locked_p(struct vm_page *pg)
1748 {
1749 
1750 	if (pg->uobject != NULL) {
1751 		return mutex_owned(pg->uobject->vmobjlock);
1752 	}
1753 	if (pg->uanon != NULL) {
1754 		return mutex_owned(pg->uanon->an_lock);
1755 	}
1756 	return true;
1757 }
1758 
1759 #if defined(DDB) || defined(DEBUGPRINT)
1760 
1761 /*
1762  * uvm_page_printit: actually print the page
1763  */
1764 
1765 static const char page_flagbits[] = UVM_PGFLAGBITS;
1766 static const char page_pqflagbits[] = UVM_PQFLAGBITS;
1767 
1768 void
1769 uvm_page_printit(struct vm_page *pg, bool full,
1770     void (*pr)(const char *, ...))
1771 {
1772 	struct vm_page *tpg;
1773 	struct uvm_object *uobj;
1774 	struct pgflist *pgl;
1775 	char pgbuf[128];
1776 	char pqbuf[128];
1777 
1778 	(*pr)("PAGE %p:\n", pg);
1779 	snprintb(pgbuf, sizeof(pgbuf), page_flagbits, pg->flags);
1780 	snprintb(pqbuf, sizeof(pqbuf), page_pqflagbits, pg->pqflags);
1781 	(*pr)("  flags=%s, pqflags=%s, wire_count=%d, pa=0x%lx\n",
1782 	    pgbuf, pqbuf, pg->wire_count, (long)VM_PAGE_TO_PHYS(pg));
1783 	(*pr)("  uobject=%p, uanon=%p, offset=0x%llx loan_count=%d\n",
1784 	    pg->uobject, pg->uanon, (long long)pg->offset, pg->loan_count);
1785 #if defined(UVM_PAGE_TRKOWN)
1786 	if (pg->flags & PG_BUSY)
1787 		(*pr)("  owning process = %d, tag=%s\n",
1788 		    pg->owner, pg->owner_tag);
1789 	else
1790 		(*pr)("  page not busy, no owner\n");
1791 #else
1792 	(*pr)("  [page ownership tracking disabled]\n");
1793 #endif
1794 
1795 	if (!full)
1796 		return;
1797 
1798 	/* cross-verify object/anon */
1799 	if ((pg->pqflags & PQ_FREE) == 0) {
1800 		if (pg->pqflags & PQ_ANON) {
1801 			if (pg->uanon == NULL || pg->uanon->an_page != pg)
1802 			    (*pr)("  >>> ANON DOES NOT POINT HERE <<< (%p)\n",
1803 				(pg->uanon) ? pg->uanon->an_page : NULL);
1804 			else
1805 				(*pr)("  anon backpointer is OK\n");
1806 		} else {
1807 			uobj = pg->uobject;
1808 			if (uobj) {
1809 				(*pr)("  checking object list\n");
1810 				TAILQ_FOREACH(tpg, &uobj->memq, listq.queue) {
1811 					if (tpg == pg) {
1812 						break;
1813 					}
1814 				}
1815 				if (tpg)
1816 					(*pr)("  page found on object list\n");
1817 				else
1818 			(*pr)("  >>> PAGE NOT FOUND ON OBJECT LIST! <<<\n");
1819 			}
1820 		}
1821 	}
1822 
1823 	/* cross-verify page queue */
1824 	if (pg->pqflags & PQ_FREE) {
1825 		int fl = uvm_page_lookup_freelist(pg);
1826 		int color = VM_PGCOLOR_BUCKET(pg);
1827 		pgl = &uvm.page_free[fl].pgfl_buckets[color].pgfl_queues[
1828 		    ((pg)->flags & PG_ZERO) ? PGFL_ZEROS : PGFL_UNKNOWN];
1829 	} else {
1830 		pgl = NULL;
1831 	}
1832 
1833 	if (pgl) {
1834 		(*pr)("  checking pageq list\n");
1835 		LIST_FOREACH(tpg, pgl, pageq.list) {
1836 			if (tpg == pg) {
1837 				break;
1838 			}
1839 		}
1840 		if (tpg)
1841 			(*pr)("  page found on pageq list\n");
1842 		else
1843 			(*pr)("  >>> PAGE NOT FOUND ON PAGEQ LIST! <<<\n");
1844 	}
1845 }
1846 
1847 /*
1848  * uvm_pages_printthem - print a summary of all managed pages
1849  */
1850 
1851 void
1852 uvm_page_printall(void (*pr)(const char *, ...))
1853 {
1854 	uvm_physseg_t i;
1855 	paddr_t pfn;
1856 	struct vm_page *pg;
1857 
1858 	(*pr)("%18s %4s %4s %18s %18s"
1859 #ifdef UVM_PAGE_TRKOWN
1860 	    " OWNER"
1861 #endif
1862 	    "\n", "PAGE", "FLAG", "PQ", "UOBJECT", "UANON");
1863 	for (i = uvm_physseg_get_first();
1864 	     uvm_physseg_valid_p(i);
1865 	     i = uvm_physseg_get_next(i)) {
1866 		for (pfn = uvm_physseg_get_start(i);
1867 		     pfn < uvm_physseg_get_end(i);
1868 		     pfn++) {
1869 			pg = PHYS_TO_VM_PAGE(ptoa(pfn));
1870 
1871 			(*pr)("%18p %04x %04x %18p %18p",
1872 			    pg, pg->flags, pg->pqflags, pg->uobject,
1873 			    pg->uanon);
1874 #ifdef UVM_PAGE_TRKOWN
1875 			if (pg->flags & PG_BUSY)
1876 				(*pr)(" %d [%s]", pg->owner, pg->owner_tag);
1877 #endif
1878 			(*pr)("\n");
1879 		}
1880 	}
1881 }
1882 
1883 #endif /* DDB || DEBUGPRINT */
1884