xref: /netbsd-src/sys/uvm/uvm_page.c (revision e89934bbf778a6d6d6894877c4da59d0c7835b0f)
1 /*	$NetBSD: uvm_page.c,v 1.192 2017/02/05 07:25:49 maya 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 1.192 2017/02/05 07:25:49 maya 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 	if (bucketarray == NULL) {
724 		printf("WARNING: unable to allocate %ld page color buckets\n",
725 		    (long) bucketcount);
726 		return;
727 	}
728 
729 	mutex_spin_enter(&uvm_fpageqlock);
730 
731 	/* Make sure we should still do this. */
732 	if (newncolors <= uvmexp.ncolors) {
733 		mutex_spin_exit(&uvm_fpageqlock);
734 		kmem_free(bucketarray, bucketmemsize);
735 		return;
736 	}
737 
738 	oldbucketarray = uvm.page_free[0].pgfl_buckets;
739 	ocolors = uvmexp.ncolors;
740 
741 	uvmexp.ncolors = newncolors;
742 	uvmexp.colormask = uvmexp.ncolors - 1;
743 
744 	ucpu = curcpu()->ci_data.cpu_uvm;
745 	for (lcv = 0; lcv < VM_NFREELIST; lcv++) {
746 		gpgfl.pgfl_buckets = (bucketarray + (lcv * newncolors));
747 		pgfl.pgfl_buckets = (cpuarray + (lcv * uvmexp.ncolors));
748 		uvm_page_init_buckets(&gpgfl);
749 		uvm_page_init_buckets(&pgfl);
750 		for (color = 0; color < ocolors; color++) {
751 			for (i = 0; i < PGFL_NQUEUES; i++) {
752 				while ((pg = LIST_FIRST(&uvm.page_free[
753 				    lcv].pgfl_buckets[color].pgfl_queues[i]))
754 				    != NULL) {
755 					LIST_REMOVE(pg, pageq.list); /* global */
756 					LIST_REMOVE(pg, listq.list); /* cpu */
757 					LIST_INSERT_HEAD(&gpgfl.pgfl_buckets[
758 					    VM_PGCOLOR_BUCKET(pg)].pgfl_queues[
759 					    i], pg, pageq.list);
760 					LIST_INSERT_HEAD(&pgfl.pgfl_buckets[
761 					    VM_PGCOLOR_BUCKET(pg)].pgfl_queues[
762 					    i], pg, listq.list);
763 				}
764 			}
765 		}
766 		uvm.page_free[lcv].pgfl_buckets = gpgfl.pgfl_buckets;
767 		ucpu->page_free[lcv].pgfl_buckets = pgfl.pgfl_buckets;
768 	}
769 
770 	oldbucketmemsize = recolored_pages_memsize;
771 
772 	recolored_pages_memsize = bucketmemsize;
773 	mutex_spin_exit(&uvm_fpageqlock);
774 
775 	if (oldbucketmemsize) {
776 		kmem_free(oldbucketarray, recolored_pages_memsize);
777 	}
778 
779 	/*
780 	 * this calls uvm_km_alloc() which may want to hold
781 	 * uvm_fpageqlock.
782 	 */
783 	uvm_pager_realloc_emerg();
784 }
785 
786 /*
787  * uvm_cpu_attach: initialize per-CPU data structures.
788  */
789 
790 void
791 uvm_cpu_attach(struct cpu_info *ci)
792 {
793 	struct pgflbucket *bucketarray;
794 	struct pgfreelist pgfl;
795 	struct uvm_cpu *ucpu;
796 	vsize_t bucketcount;
797 	int lcv;
798 
799 	if (CPU_IS_PRIMARY(ci)) {
800 		/* Already done in uvm_page_init(). */
801 		goto attachrnd;
802 	}
803 
804 	/* Add more reserve pages for this CPU. */
805 	uvmexp.reserve_kernel += vm_page_reserve_kernel;
806 
807 	/* Configure this CPU's free lists. */
808 	bucketcount = uvmexp.ncolors * VM_NFREELIST;
809 	bucketarray = kmem_alloc(bucketcount * sizeof(struct pgflbucket),
810 	    KM_SLEEP);
811 	ucpu = kmem_zalloc(sizeof(*ucpu), KM_SLEEP);
812 	uvm.cpus[cpu_index(ci)] = ucpu;
813 	ci->ci_data.cpu_uvm = ucpu;
814 	for (lcv = 0; lcv < VM_NFREELIST; lcv++) {
815 		pgfl.pgfl_buckets = (bucketarray + (lcv * uvmexp.ncolors));
816 		uvm_page_init_buckets(&pgfl);
817 		ucpu->page_free[lcv].pgfl_buckets = pgfl.pgfl_buckets;
818 	}
819 
820 attachrnd:
821 	/*
822 	 * Attach RNG source for this CPU's VM events
823 	 */
824         rnd_attach_source(&uvm.cpus[cpu_index(ci)]->rs,
825 			  ci->ci_data.cpu_name, RND_TYPE_VM,
826 			  RND_FLAG_COLLECT_TIME|RND_FLAG_COLLECT_VALUE|
827 			  RND_FLAG_ESTIMATE_VALUE);
828 
829 }
830 
831 /*
832  * uvm_pagealloc_pgfl: helper routine for uvm_pagealloc_strat
833  */
834 
835 static struct vm_page *
836 uvm_pagealloc_pgfl(struct uvm_cpu *ucpu, int flist, int try1, int try2,
837     int *trycolorp)
838 {
839 	struct pgflist *freeq;
840 	struct vm_page *pg;
841 	int color, trycolor = *trycolorp;
842 	struct pgfreelist *gpgfl, *pgfl;
843 
844 	KASSERT(mutex_owned(&uvm_fpageqlock));
845 
846 	color = trycolor;
847 	pgfl = &ucpu->page_free[flist];
848 	gpgfl = &uvm.page_free[flist];
849 	do {
850 		/* cpu, try1 */
851 		if ((pg = LIST_FIRST((freeq =
852 		    &pgfl->pgfl_buckets[color].pgfl_queues[try1]))) != NULL) {
853 			KASSERT(pg->pqflags & PQ_FREE);
854 			KASSERT(try1 == PGFL_ZEROS || !(pg->flags & PG_ZERO));
855 			KASSERT(try1 == PGFL_UNKNOWN || (pg->flags & PG_ZERO));
856 			KASSERT(ucpu == VM_FREE_PAGE_TO_CPU(pg));
857 			VM_FREE_PAGE_TO_CPU(pg)->pages[try1]--;
858 		    	uvmexp.cpuhit++;
859 			goto gotit;
860 		}
861 		/* global, try1 */
862 		if ((pg = LIST_FIRST((freeq =
863 		    &gpgfl->pgfl_buckets[color].pgfl_queues[try1]))) != NULL) {
864 			KASSERT(pg->pqflags & PQ_FREE);
865 			KASSERT(try1 == PGFL_ZEROS || !(pg->flags & PG_ZERO));
866 			KASSERT(try1 == PGFL_UNKNOWN || (pg->flags & PG_ZERO));
867 			KASSERT(ucpu != VM_FREE_PAGE_TO_CPU(pg));
868 			VM_FREE_PAGE_TO_CPU(pg)->pages[try1]--;
869 		    	uvmexp.cpumiss++;
870 			goto gotit;
871 		}
872 		/* cpu, try2 */
873 		if ((pg = LIST_FIRST((freeq =
874 		    &pgfl->pgfl_buckets[color].pgfl_queues[try2]))) != NULL) {
875 			KASSERT(pg->pqflags & PQ_FREE);
876 			KASSERT(try2 == PGFL_ZEROS || !(pg->flags & PG_ZERO));
877 			KASSERT(try2 == PGFL_UNKNOWN || (pg->flags & PG_ZERO));
878 			KASSERT(ucpu == VM_FREE_PAGE_TO_CPU(pg));
879 			VM_FREE_PAGE_TO_CPU(pg)->pages[try2]--;
880 		    	uvmexp.cpuhit++;
881 			goto gotit;
882 		}
883 		/* global, try2 */
884 		if ((pg = LIST_FIRST((freeq =
885 		    &gpgfl->pgfl_buckets[color].pgfl_queues[try2]))) != NULL) {
886 			KASSERT(pg->pqflags & PQ_FREE);
887 			KASSERT(try2 == PGFL_ZEROS || !(pg->flags & PG_ZERO));
888 			KASSERT(try2 == PGFL_UNKNOWN || (pg->flags & PG_ZERO));
889 			KASSERT(ucpu != VM_FREE_PAGE_TO_CPU(pg));
890 			VM_FREE_PAGE_TO_CPU(pg)->pages[try2]--;
891 		    	uvmexp.cpumiss++;
892 			goto gotit;
893 		}
894 		color = (color + 1) & uvmexp.colormask;
895 	} while (color != trycolor);
896 
897 	return (NULL);
898 
899  gotit:
900 	LIST_REMOVE(pg, pageq.list);	/* global list */
901 	LIST_REMOVE(pg, listq.list);	/* per-cpu list */
902 	uvmexp.free--;
903 
904 	/* update zero'd page count */
905 	if (pg->flags & PG_ZERO)
906 		uvmexp.zeropages--;
907 
908 	if (color == trycolor)
909 		uvmexp.colorhit++;
910 	else {
911 		uvmexp.colormiss++;
912 		*trycolorp = color;
913 	}
914 
915 	return (pg);
916 }
917 
918 /*
919  * uvm_pagealloc_strat: allocate vm_page from a particular free list.
920  *
921  * => return null if no pages free
922  * => wake up pagedaemon if number of free pages drops below low water mark
923  * => if obj != NULL, obj must be locked (to put in obj's tree)
924  * => if anon != NULL, anon must be locked (to put in anon)
925  * => only one of obj or anon can be non-null
926  * => caller must activate/deactivate page if it is not wired.
927  * => free_list is ignored if strat == UVM_PGA_STRAT_NORMAL.
928  * => policy decision: it is more important to pull a page off of the
929  *	appropriate priority free list than it is to get a zero'd or
930  *	unknown contents page.  This is because we live with the
931  *	consequences of a bad free list decision for the entire
932  *	lifetime of the page, e.g. if the page comes from memory that
933  *	is slower to access.
934  */
935 
936 struct vm_page *
937 uvm_pagealloc_strat(struct uvm_object *obj, voff_t off, struct vm_anon *anon,
938     int flags, int strat, int free_list)
939 {
940 	int try1, try2, zeroit = 0, color;
941 	int lcv;
942 	struct uvm_cpu *ucpu;
943 	struct vm_page *pg;
944 	lwp_t *l;
945 
946 	KASSERT(obj == NULL || anon == NULL);
947 	KASSERT(anon == NULL || (flags & UVM_FLAG_COLORMATCH) || off == 0);
948 	KASSERT(off == trunc_page(off));
949 	KASSERT(obj == NULL || mutex_owned(obj->vmobjlock));
950 	KASSERT(anon == NULL || anon->an_lock == NULL ||
951 	    mutex_owned(anon->an_lock));
952 
953 	mutex_spin_enter(&uvm_fpageqlock);
954 
955 	/*
956 	 * This implements a global round-robin page coloring
957 	 * algorithm.
958 	 */
959 
960 	ucpu = curcpu()->ci_data.cpu_uvm;
961 	if (flags & UVM_FLAG_COLORMATCH) {
962 		color = atop(off) & uvmexp.colormask;
963 	} else {
964 		color = ucpu->page_free_nextcolor;
965 	}
966 
967 	/*
968 	 * check to see if we need to generate some free pages waking
969 	 * the pagedaemon.
970 	 */
971 
972 	uvm_kick_pdaemon();
973 
974 	/*
975 	 * fail if any of these conditions is true:
976 	 * [1]  there really are no free pages, or
977 	 * [2]  only kernel "reserved" pages remain and
978 	 *        reserved pages have not been requested.
979 	 * [3]  only pagedaemon "reserved" pages remain and
980 	 *        the requestor isn't the pagedaemon.
981 	 * we make kernel reserve pages available if called by a
982 	 * kernel thread or a realtime thread.
983 	 */
984 	l = curlwp;
985 	if (__predict_true(l != NULL) && lwp_eprio(l) >= PRI_KTHREAD) {
986 		flags |= UVM_PGA_USERESERVE;
987 	}
988 	if ((uvmexp.free <= uvmexp.reserve_kernel &&
989 	    (flags & UVM_PGA_USERESERVE) == 0) ||
990 	    (uvmexp.free <= uvmexp.reserve_pagedaemon &&
991 	     curlwp != uvm.pagedaemon_lwp))
992 		goto fail;
993 
994 #if PGFL_NQUEUES != 2
995 #error uvm_pagealloc_strat needs to be updated
996 #endif
997 
998 	/*
999 	 * If we want a zero'd page, try the ZEROS queue first, otherwise
1000 	 * we try the UNKNOWN queue first.
1001 	 */
1002 	if (flags & UVM_PGA_ZERO) {
1003 		try1 = PGFL_ZEROS;
1004 		try2 = PGFL_UNKNOWN;
1005 	} else {
1006 		try1 = PGFL_UNKNOWN;
1007 		try2 = PGFL_ZEROS;
1008 	}
1009 
1010  again:
1011 	switch (strat) {
1012 	case UVM_PGA_STRAT_NORMAL:
1013 		/* Check freelists: descending priority (ascending id) order */
1014 		for (lcv = 0; lcv < VM_NFREELIST; lcv++) {
1015 			pg = uvm_pagealloc_pgfl(ucpu, lcv,
1016 			    try1, try2, &color);
1017 			if (pg != NULL)
1018 				goto gotit;
1019 		}
1020 
1021 		/* No pages free! */
1022 		goto fail;
1023 
1024 	case UVM_PGA_STRAT_ONLY:
1025 	case UVM_PGA_STRAT_FALLBACK:
1026 		/* Attempt to allocate from the specified free list. */
1027 		KASSERT(free_list >= 0 && free_list < VM_NFREELIST);
1028 		pg = uvm_pagealloc_pgfl(ucpu, free_list,
1029 		    try1, try2, &color);
1030 		if (pg != NULL)
1031 			goto gotit;
1032 
1033 		/* Fall back, if possible. */
1034 		if (strat == UVM_PGA_STRAT_FALLBACK) {
1035 			strat = UVM_PGA_STRAT_NORMAL;
1036 			goto again;
1037 		}
1038 
1039 		/* No pages free! */
1040 		goto fail;
1041 
1042 	default:
1043 		panic("uvm_pagealloc_strat: bad strat %d", strat);
1044 		/* NOTREACHED */
1045 	}
1046 
1047  gotit:
1048 	/*
1049 	 * We now know which color we actually allocated from; set
1050 	 * the next color accordingly.
1051 	 */
1052 
1053 	ucpu->page_free_nextcolor = (color + 1) & uvmexp.colormask;
1054 
1055 	/*
1056 	 * update allocation statistics and remember if we have to
1057 	 * zero the page
1058 	 */
1059 
1060 	if (flags & UVM_PGA_ZERO) {
1061 		if (pg->flags & PG_ZERO) {
1062 			uvmexp.pga_zerohit++;
1063 			zeroit = 0;
1064 		} else {
1065 			uvmexp.pga_zeromiss++;
1066 			zeroit = 1;
1067 		}
1068 		if (ucpu->pages[PGFL_ZEROS] < ucpu->pages[PGFL_UNKNOWN]) {
1069 			ucpu->page_idle_zero = vm_page_zero_enable;
1070 		}
1071 	}
1072 	KASSERT(pg->pqflags == PQ_FREE);
1073 
1074 	pg->offset = off;
1075 	pg->uobject = obj;
1076 	pg->uanon = anon;
1077 	pg->flags = PG_BUSY|PG_CLEAN|PG_FAKE;
1078 	if (anon) {
1079 		anon->an_page = pg;
1080 		pg->pqflags = PQ_ANON;
1081 		atomic_inc_uint(&uvmexp.anonpages);
1082 	} else {
1083 		if (obj) {
1084 			uvm_pageinsert(obj, pg);
1085 		}
1086 		pg->pqflags = 0;
1087 	}
1088 	mutex_spin_exit(&uvm_fpageqlock);
1089 
1090 #if defined(UVM_PAGE_TRKOWN)
1091 	pg->owner_tag = NULL;
1092 #endif
1093 	UVM_PAGE_OWN(pg, "new alloc");
1094 
1095 	if (flags & UVM_PGA_ZERO) {
1096 		/*
1097 		 * A zero'd page is not clean.  If we got a page not already
1098 		 * zero'd, then we have to zero it ourselves.
1099 		 */
1100 		pg->flags &= ~PG_CLEAN;
1101 		if (zeroit)
1102 			pmap_zero_page(VM_PAGE_TO_PHYS(pg));
1103 	}
1104 
1105 	return(pg);
1106 
1107  fail:
1108 	mutex_spin_exit(&uvm_fpageqlock);
1109 	return (NULL);
1110 }
1111 
1112 /*
1113  * uvm_pagereplace: replace a page with another
1114  *
1115  * => object must be locked
1116  */
1117 
1118 void
1119 uvm_pagereplace(struct vm_page *oldpg, struct vm_page *newpg)
1120 {
1121 	struct uvm_object *uobj = oldpg->uobject;
1122 
1123 	KASSERT((oldpg->flags & PG_TABLED) != 0);
1124 	KASSERT(uobj != NULL);
1125 	KASSERT((newpg->flags & PG_TABLED) == 0);
1126 	KASSERT(newpg->uobject == NULL);
1127 	KASSERT(mutex_owned(uobj->vmobjlock));
1128 
1129 	newpg->uobject = uobj;
1130 	newpg->offset = oldpg->offset;
1131 
1132 	uvm_pageremove_tree(uobj, oldpg);
1133 	uvm_pageinsert_tree(uobj, newpg);
1134 	uvm_pageinsert_list(uobj, newpg, oldpg);
1135 	uvm_pageremove_list(uobj, oldpg);
1136 }
1137 
1138 /*
1139  * uvm_pagerealloc: reallocate a page from one object to another
1140  *
1141  * => both objects must be locked
1142  */
1143 
1144 void
1145 uvm_pagerealloc(struct vm_page *pg, struct uvm_object *newobj, voff_t newoff)
1146 {
1147 	/*
1148 	 * remove it from the old object
1149 	 */
1150 
1151 	if (pg->uobject) {
1152 		uvm_pageremove(pg->uobject, pg);
1153 	}
1154 
1155 	/*
1156 	 * put it in the new object
1157 	 */
1158 
1159 	if (newobj) {
1160 		pg->uobject = newobj;
1161 		pg->offset = newoff;
1162 		uvm_pageinsert(newobj, pg);
1163 	}
1164 }
1165 
1166 #ifdef DEBUG
1167 /*
1168  * check if page is zero-filled
1169  *
1170  *  - called with free page queue lock held.
1171  */
1172 void
1173 uvm_pagezerocheck(struct vm_page *pg)
1174 {
1175 	int *p, *ep;
1176 
1177 	KASSERT(uvm_zerocheckkva != 0);
1178 	KASSERT(mutex_owned(&uvm_fpageqlock));
1179 
1180 	/*
1181 	 * XXX assuming pmap_kenter_pa and pmap_kremove never call
1182 	 * uvm page allocator.
1183 	 *
1184 	 * it might be better to have "CPU-local temporary map" pmap interface.
1185 	 */
1186 	pmap_kenter_pa(uvm_zerocheckkva, VM_PAGE_TO_PHYS(pg), VM_PROT_READ, 0);
1187 	p = (int *)uvm_zerocheckkva;
1188 	ep = (int *)((char *)p + PAGE_SIZE);
1189 	pmap_update(pmap_kernel());
1190 	while (p < ep) {
1191 		if (*p != 0)
1192 			panic("PG_ZERO page isn't zero-filled");
1193 		p++;
1194 	}
1195 	pmap_kremove(uvm_zerocheckkva, PAGE_SIZE);
1196 	/*
1197 	 * pmap_update() is not necessary here because no one except us
1198 	 * uses this VA.
1199 	 */
1200 }
1201 #endif /* DEBUG */
1202 
1203 /*
1204  * uvm_pagefree: free page
1205  *
1206  * => erase page's identity (i.e. remove from object)
1207  * => put page on free list
1208  * => caller must lock owning object (either anon or uvm_object)
1209  * => caller must lock page queues
1210  * => assumes all valid mappings of pg are gone
1211  */
1212 
1213 void
1214 uvm_pagefree(struct vm_page *pg)
1215 {
1216 	struct pgflist *pgfl;
1217 	struct uvm_cpu *ucpu;
1218 	int index, color, queue;
1219 	bool iszero;
1220 
1221 #ifdef DEBUG
1222 	if (pg->uobject == (void *)0xdeadbeef &&
1223 	    pg->uanon == (void *)0xdeadbeef) {
1224 		panic("uvm_pagefree: freeing free page %p", pg);
1225 	}
1226 #endif /* DEBUG */
1227 
1228 	KASSERT((pg->flags & PG_PAGEOUT) == 0);
1229 	KASSERT(!(pg->pqflags & PQ_FREE));
1230 	//KASSERT(mutex_owned(&uvm_pageqlock) || !uvmpdpol_pageisqueued_p(pg));
1231 	KASSERT(pg->uobject == NULL || mutex_owned(pg->uobject->vmobjlock));
1232 	KASSERT(pg->uobject != NULL || pg->uanon == NULL ||
1233 		mutex_owned(pg->uanon->an_lock));
1234 
1235 	/*
1236 	 * if the page is loaned, resolve the loan instead of freeing.
1237 	 */
1238 
1239 	if (pg->loan_count) {
1240 		KASSERT(pg->wire_count == 0);
1241 
1242 		/*
1243 		 * if the page is owned by an anon then we just want to
1244 		 * drop anon ownership.  the kernel will free the page when
1245 		 * it is done with it.  if the page is owned by an object,
1246 		 * remove it from the object and mark it dirty for the benefit
1247 		 * of possible anon owners.
1248 		 *
1249 		 * regardless of previous ownership, wakeup any waiters,
1250 		 * unbusy the page, and we're done.
1251 		 */
1252 
1253 		if (pg->uobject != NULL) {
1254 			uvm_pageremove(pg->uobject, pg);
1255 			pg->flags &= ~PG_CLEAN;
1256 		} else if (pg->uanon != NULL) {
1257 			if ((pg->pqflags & PQ_ANON) == 0) {
1258 				pg->loan_count--;
1259 			} else {
1260 				pg->pqflags &= ~PQ_ANON;
1261 				atomic_dec_uint(&uvmexp.anonpages);
1262 			}
1263 			pg->uanon->an_page = NULL;
1264 			pg->uanon = NULL;
1265 		}
1266 		if (pg->flags & PG_WANTED) {
1267 			wakeup(pg);
1268 		}
1269 		pg->flags &= ~(PG_WANTED|PG_BUSY|PG_RELEASED|PG_PAGER1);
1270 #ifdef UVM_PAGE_TRKOWN
1271 		pg->owner_tag = NULL;
1272 #endif
1273 		if (pg->loan_count) {
1274 			KASSERT(pg->uobject == NULL);
1275 			if (pg->uanon == NULL) {
1276 				KASSERT(mutex_owned(&uvm_pageqlock));
1277 				uvm_pagedequeue(pg);
1278 			}
1279 			return;
1280 		}
1281 	}
1282 
1283 	/*
1284 	 * remove page from its object or anon.
1285 	 */
1286 
1287 	if (pg->uobject != NULL) {
1288 		uvm_pageremove(pg->uobject, pg);
1289 	} else if (pg->uanon != NULL) {
1290 		pg->uanon->an_page = NULL;
1291 		atomic_dec_uint(&uvmexp.anonpages);
1292 	}
1293 
1294 	/*
1295 	 * now remove the page from the queues.
1296 	 */
1297 	if (uvmpdpol_pageisqueued_p(pg)) {
1298 		KASSERT(mutex_owned(&uvm_pageqlock));
1299 		uvm_pagedequeue(pg);
1300 	}
1301 
1302 	/*
1303 	 * if the page was wired, unwire it now.
1304 	 */
1305 
1306 	if (pg->wire_count) {
1307 		pg->wire_count = 0;
1308 		uvmexp.wired--;
1309 	}
1310 
1311 	/*
1312 	 * and put on free queue
1313 	 */
1314 
1315 	iszero = (pg->flags & PG_ZERO);
1316 	index = uvm_page_lookup_freelist(pg);
1317 	color = VM_PGCOLOR_BUCKET(pg);
1318 	queue = (iszero ? PGFL_ZEROS : PGFL_UNKNOWN);
1319 
1320 #ifdef DEBUG
1321 	pg->uobject = (void *)0xdeadbeef;
1322 	pg->uanon = (void *)0xdeadbeef;
1323 #endif
1324 
1325 	mutex_spin_enter(&uvm_fpageqlock);
1326 	pg->pqflags = PQ_FREE;
1327 
1328 #ifdef DEBUG
1329 	if (iszero)
1330 		uvm_pagezerocheck(pg);
1331 #endif /* DEBUG */
1332 
1333 
1334 	/* global list */
1335 	pgfl = &uvm.page_free[index].pgfl_buckets[color].pgfl_queues[queue];
1336 	LIST_INSERT_HEAD(pgfl, pg, pageq.list);
1337 	uvmexp.free++;
1338 	if (iszero) {
1339 		uvmexp.zeropages++;
1340 	}
1341 
1342 	/* per-cpu list */
1343 	ucpu = curcpu()->ci_data.cpu_uvm;
1344 	pg->offset = (uintptr_t)ucpu;
1345 	pgfl = &ucpu->page_free[index].pgfl_buckets[color].pgfl_queues[queue];
1346 	LIST_INSERT_HEAD(pgfl, pg, listq.list);
1347 	ucpu->pages[queue]++;
1348 	if (ucpu->pages[PGFL_ZEROS] < ucpu->pages[PGFL_UNKNOWN]) {
1349 		ucpu->page_idle_zero = vm_page_zero_enable;
1350 	}
1351 
1352 	mutex_spin_exit(&uvm_fpageqlock);
1353 }
1354 
1355 /*
1356  * uvm_page_unbusy: unbusy an array of pages.
1357  *
1358  * => pages must either all belong to the same object, or all belong to anons.
1359  * => if pages are object-owned, object must be locked.
1360  * => if pages are anon-owned, anons must be locked.
1361  * => caller must lock page queues if pages may be released.
1362  * => caller must make sure that anon-owned pages are not PG_RELEASED.
1363  */
1364 
1365 void
1366 uvm_page_unbusy(struct vm_page **pgs, int npgs)
1367 {
1368 	struct vm_page *pg;
1369 	int i;
1370 	UVMHIST_FUNC("uvm_page_unbusy"); UVMHIST_CALLED(ubchist);
1371 
1372 	for (i = 0; i < npgs; i++) {
1373 		pg = pgs[i];
1374 		if (pg == NULL || pg == PGO_DONTCARE) {
1375 			continue;
1376 		}
1377 
1378 		KASSERT(uvm_page_locked_p(pg));
1379 		KASSERT(pg->flags & PG_BUSY);
1380 		KASSERT((pg->flags & PG_PAGEOUT) == 0);
1381 		if (pg->flags & PG_WANTED) {
1382 			wakeup(pg);
1383 		}
1384 		if (pg->flags & PG_RELEASED) {
1385 			UVMHIST_LOG(ubchist, "releasing pg %p", pg,0,0,0);
1386 			KASSERT(pg->uobject != NULL ||
1387 			    (pg->uanon != NULL && pg->uanon->an_ref > 0));
1388 			pg->flags &= ~PG_RELEASED;
1389 			uvm_pagefree(pg);
1390 		} else {
1391 			UVMHIST_LOG(ubchist, "unbusying pg %p", pg,0,0,0);
1392 			KASSERT((pg->flags & PG_FAKE) == 0);
1393 			pg->flags &= ~(PG_WANTED|PG_BUSY);
1394 			UVM_PAGE_OWN(pg, NULL);
1395 		}
1396 	}
1397 }
1398 
1399 #if defined(UVM_PAGE_TRKOWN)
1400 /*
1401  * uvm_page_own: set or release page ownership
1402  *
1403  * => this is a debugging function that keeps track of who sets PG_BUSY
1404  *	and where they do it.   it can be used to track down problems
1405  *	such a process setting "PG_BUSY" and never releasing it.
1406  * => page's object [if any] must be locked
1407  * => if "tag" is NULL then we are releasing page ownership
1408  */
1409 void
1410 uvm_page_own(struct vm_page *pg, const char *tag)
1411 {
1412 
1413 	KASSERT((pg->flags & (PG_PAGEOUT|PG_RELEASED)) == 0);
1414 	KASSERT((pg->flags & PG_WANTED) == 0);
1415 	KASSERT(uvm_page_locked_p(pg));
1416 
1417 	/* gain ownership? */
1418 	if (tag) {
1419 		KASSERT((pg->flags & PG_BUSY) != 0);
1420 		if (pg->owner_tag) {
1421 			printf("uvm_page_own: page %p already owned "
1422 			    "by proc %d [%s]\n", pg,
1423 			    pg->owner, pg->owner_tag);
1424 			panic("uvm_page_own");
1425 		}
1426 		pg->owner = curproc->p_pid;
1427 		pg->lowner = curlwp->l_lid;
1428 		pg->owner_tag = tag;
1429 		return;
1430 	}
1431 
1432 	/* drop ownership */
1433 	KASSERT((pg->flags & PG_BUSY) == 0);
1434 	if (pg->owner_tag == NULL) {
1435 		printf("uvm_page_own: dropping ownership of an non-owned "
1436 		    "page (%p)\n", pg);
1437 		panic("uvm_page_own");
1438 	}
1439 	if (!uvmpdpol_pageisqueued_p(pg)) {
1440 		KASSERT((pg->uanon == NULL && pg->uobject == NULL) ||
1441 		    pg->wire_count > 0);
1442 	} else {
1443 		KASSERT(pg->wire_count == 0);
1444 	}
1445 	pg->owner_tag = NULL;
1446 }
1447 #endif
1448 
1449 /*
1450  * uvm_pageidlezero: zero free pages while the system is idle.
1451  *
1452  * => try to complete one color bucket at a time, to reduce our impact
1453  *	on the CPU cache.
1454  * => we loop until we either reach the target or there is a lwp ready
1455  *      to run, or MD code detects a reason to break early.
1456  */
1457 void
1458 uvm_pageidlezero(void)
1459 {
1460 	struct vm_page *pg;
1461 	struct pgfreelist *pgfl, *gpgfl;
1462 	struct uvm_cpu *ucpu;
1463 	int free_list, firstbucket, nextbucket;
1464 	bool lcont = false;
1465 
1466 	ucpu = curcpu()->ci_data.cpu_uvm;
1467 	if (!ucpu->page_idle_zero ||
1468 	    ucpu->pages[PGFL_UNKNOWN] < uvmexp.ncolors) {
1469 	    	ucpu->page_idle_zero = false;
1470 		return;
1471 	}
1472 	if (!mutex_tryenter(&uvm_fpageqlock)) {
1473 		/* Contention: let other CPUs to use the lock. */
1474 		return;
1475 	}
1476 	firstbucket = ucpu->page_free_nextcolor;
1477 	nextbucket = firstbucket;
1478 	do {
1479 		for (free_list = 0; free_list < VM_NFREELIST; free_list++) {
1480 			if (sched_curcpu_runnable_p()) {
1481 				goto quit;
1482 			}
1483 			pgfl = &ucpu->page_free[free_list];
1484 			gpgfl = &uvm.page_free[free_list];
1485 			while ((pg = LIST_FIRST(&pgfl->pgfl_buckets[
1486 			    nextbucket].pgfl_queues[PGFL_UNKNOWN])) != NULL) {
1487 				if (lcont || sched_curcpu_runnable_p()) {
1488 					goto quit;
1489 				}
1490 				LIST_REMOVE(pg, pageq.list); /* global list */
1491 				LIST_REMOVE(pg, listq.list); /* per-cpu list */
1492 				ucpu->pages[PGFL_UNKNOWN]--;
1493 				uvmexp.free--;
1494 				KASSERT(pg->pqflags == PQ_FREE);
1495 				pg->pqflags = 0;
1496 				mutex_spin_exit(&uvm_fpageqlock);
1497 #ifdef PMAP_PAGEIDLEZERO
1498 				if (!PMAP_PAGEIDLEZERO(VM_PAGE_TO_PHYS(pg))) {
1499 
1500 					/*
1501 					 * The machine-dependent code detected
1502 					 * some reason for us to abort zeroing
1503 					 * pages, probably because there is a
1504 					 * process now ready to run.
1505 					 */
1506 
1507 					mutex_spin_enter(&uvm_fpageqlock);
1508 					pg->pqflags = PQ_FREE;
1509 					LIST_INSERT_HEAD(&gpgfl->pgfl_buckets[
1510 					    nextbucket].pgfl_queues[
1511 					    PGFL_UNKNOWN], pg, pageq.list);
1512 					LIST_INSERT_HEAD(&pgfl->pgfl_buckets[
1513 					    nextbucket].pgfl_queues[
1514 					    PGFL_UNKNOWN], pg, listq.list);
1515 					ucpu->pages[PGFL_UNKNOWN]++;
1516 					uvmexp.free++;
1517 					uvmexp.zeroaborts++;
1518 					goto quit;
1519 				}
1520 #else
1521 				pmap_zero_page(VM_PAGE_TO_PHYS(pg));
1522 #endif /* PMAP_PAGEIDLEZERO */
1523 				pg->flags |= PG_ZERO;
1524 
1525 				if (!mutex_tryenter(&uvm_fpageqlock)) {
1526 					lcont = true;
1527 					mutex_spin_enter(&uvm_fpageqlock);
1528 				} else {
1529 					lcont = false;
1530 				}
1531 				pg->pqflags = PQ_FREE;
1532 				LIST_INSERT_HEAD(&gpgfl->pgfl_buckets[
1533 				    nextbucket].pgfl_queues[PGFL_ZEROS],
1534 				    pg, pageq.list);
1535 				LIST_INSERT_HEAD(&pgfl->pgfl_buckets[
1536 				    nextbucket].pgfl_queues[PGFL_ZEROS],
1537 				    pg, listq.list);
1538 				ucpu->pages[PGFL_ZEROS]++;
1539 				uvmexp.free++;
1540 				uvmexp.zeropages++;
1541 			}
1542 		}
1543 		if (ucpu->pages[PGFL_UNKNOWN] < uvmexp.ncolors) {
1544 			break;
1545 		}
1546 		nextbucket = (nextbucket + 1) & uvmexp.colormask;
1547 	} while (nextbucket != firstbucket);
1548 	ucpu->page_idle_zero = false;
1549  quit:
1550 	mutex_spin_exit(&uvm_fpageqlock);
1551 }
1552 
1553 /*
1554  * uvm_pagelookup: look up a page
1555  *
1556  * => caller should lock object to keep someone from pulling the page
1557  *	out from under it
1558  */
1559 
1560 struct vm_page *
1561 uvm_pagelookup(struct uvm_object *obj, voff_t off)
1562 {
1563 	struct vm_page *pg;
1564 
1565 	KASSERT(mutex_owned(obj->vmobjlock));
1566 
1567 	pg = rb_tree_find_node(&obj->rb_tree, &off);
1568 
1569 	KASSERT(pg == NULL || obj->uo_npages != 0);
1570 	KASSERT(pg == NULL || (pg->flags & (PG_RELEASED|PG_PAGEOUT)) == 0 ||
1571 		(pg->flags & PG_BUSY) != 0);
1572 	return pg;
1573 }
1574 
1575 /*
1576  * uvm_pagewire: wire the page, thus removing it from the daemon's grasp
1577  *
1578  * => caller must lock page queues
1579  */
1580 
1581 void
1582 uvm_pagewire(struct vm_page *pg)
1583 {
1584 	KASSERT(mutex_owned(&uvm_pageqlock));
1585 #if defined(READAHEAD_STATS)
1586 	if ((pg->pqflags & PQ_READAHEAD) != 0) {
1587 		uvm_ra_hit.ev_count++;
1588 		pg->pqflags &= ~PQ_READAHEAD;
1589 	}
1590 #endif /* defined(READAHEAD_STATS) */
1591 	if (pg->wire_count == 0) {
1592 		uvm_pagedequeue(pg);
1593 		uvmexp.wired++;
1594 	}
1595 	pg->wire_count++;
1596 }
1597 
1598 /*
1599  * uvm_pageunwire: unwire the page.
1600  *
1601  * => activate if wire count goes to zero.
1602  * => caller must lock page queues
1603  */
1604 
1605 void
1606 uvm_pageunwire(struct vm_page *pg)
1607 {
1608 	KASSERT(mutex_owned(&uvm_pageqlock));
1609 	pg->wire_count--;
1610 	if (pg->wire_count == 0) {
1611 		uvm_pageactivate(pg);
1612 		uvmexp.wired--;
1613 	}
1614 }
1615 
1616 /*
1617  * uvm_pagedeactivate: deactivate page
1618  *
1619  * => caller must lock page queues
1620  * => caller must check to make sure page is not wired
1621  * => object that page belongs to must be locked (so we can adjust pg->flags)
1622  * => caller must clear the reference on the page before calling
1623  */
1624 
1625 void
1626 uvm_pagedeactivate(struct vm_page *pg)
1627 {
1628 
1629 	KASSERT(mutex_owned(&uvm_pageqlock));
1630 	KASSERT(uvm_page_locked_p(pg));
1631 	KASSERT(pg->wire_count != 0 || uvmpdpol_pageisqueued_p(pg));
1632 	uvmpdpol_pagedeactivate(pg);
1633 }
1634 
1635 /*
1636  * uvm_pageactivate: activate page
1637  *
1638  * => caller must lock page queues
1639  */
1640 
1641 void
1642 uvm_pageactivate(struct vm_page *pg)
1643 {
1644 
1645 	KASSERT(mutex_owned(&uvm_pageqlock));
1646 	KASSERT(uvm_page_locked_p(pg));
1647 #if defined(READAHEAD_STATS)
1648 	if ((pg->pqflags & PQ_READAHEAD) != 0) {
1649 		uvm_ra_hit.ev_count++;
1650 		pg->pqflags &= ~PQ_READAHEAD;
1651 	}
1652 #endif /* defined(READAHEAD_STATS) */
1653 	if (pg->wire_count != 0) {
1654 		return;
1655 	}
1656 	uvmpdpol_pageactivate(pg);
1657 }
1658 
1659 /*
1660  * uvm_pagedequeue: remove a page from any paging queue
1661  */
1662 
1663 void
1664 uvm_pagedequeue(struct vm_page *pg)
1665 {
1666 
1667 	if (uvmpdpol_pageisqueued_p(pg)) {
1668 		KASSERT(mutex_owned(&uvm_pageqlock));
1669 	}
1670 
1671 	uvmpdpol_pagedequeue(pg);
1672 }
1673 
1674 /*
1675  * uvm_pageenqueue: add a page to a paging queue without activating.
1676  * used where a page is not really demanded (yet).  eg. read-ahead
1677  */
1678 
1679 void
1680 uvm_pageenqueue(struct vm_page *pg)
1681 {
1682 
1683 	KASSERT(mutex_owned(&uvm_pageqlock));
1684 	if (pg->wire_count != 0) {
1685 		return;
1686 	}
1687 	uvmpdpol_pageenqueue(pg);
1688 }
1689 
1690 /*
1691  * uvm_pagezero: zero fill a page
1692  *
1693  * => if page is part of an object then the object should be locked
1694  *	to protect pg->flags.
1695  */
1696 
1697 void
1698 uvm_pagezero(struct vm_page *pg)
1699 {
1700 	pg->flags &= ~PG_CLEAN;
1701 	pmap_zero_page(VM_PAGE_TO_PHYS(pg));
1702 }
1703 
1704 /*
1705  * uvm_pagecopy: copy a page
1706  *
1707  * => if page is part of an object then the object should be locked
1708  *	to protect pg->flags.
1709  */
1710 
1711 void
1712 uvm_pagecopy(struct vm_page *src, struct vm_page *dst)
1713 {
1714 
1715 	dst->flags &= ~PG_CLEAN;
1716 	pmap_copy_page(VM_PAGE_TO_PHYS(src), VM_PAGE_TO_PHYS(dst));
1717 }
1718 
1719 /*
1720  * uvm_pageismanaged: test it see that a page (specified by PA) is managed.
1721  */
1722 
1723 bool
1724 uvm_pageismanaged(paddr_t pa)
1725 {
1726 
1727 	return (uvm_physseg_find(atop(pa), NULL) != UVM_PHYSSEG_TYPE_INVALID);
1728 }
1729 
1730 /*
1731  * uvm_page_lookup_freelist: look up the free list for the specified page
1732  */
1733 
1734 int
1735 uvm_page_lookup_freelist(struct vm_page *pg)
1736 {
1737 	uvm_physseg_t upm;
1738 
1739 	upm = uvm_physseg_find(atop(VM_PAGE_TO_PHYS(pg)), NULL);
1740 	KASSERT(upm != UVM_PHYSSEG_TYPE_INVALID);
1741 	return uvm_physseg_get_free_list(upm);
1742 }
1743 
1744 /*
1745  * uvm_page_locked_p: return true if object associated with page is
1746  * locked.  this is a weak check for runtime assertions only.
1747  */
1748 
1749 bool
1750 uvm_page_locked_p(struct vm_page *pg)
1751 {
1752 
1753 	if (pg->uobject != NULL) {
1754 		return mutex_owned(pg->uobject->vmobjlock);
1755 	}
1756 	if (pg->uanon != NULL) {
1757 		return mutex_owned(pg->uanon->an_lock);
1758 	}
1759 	return true;
1760 }
1761 
1762 #if defined(DDB) || defined(DEBUGPRINT)
1763 
1764 /*
1765  * uvm_page_printit: actually print the page
1766  */
1767 
1768 static const char page_flagbits[] = UVM_PGFLAGBITS;
1769 static const char page_pqflagbits[] = UVM_PQFLAGBITS;
1770 
1771 void
1772 uvm_page_printit(struct vm_page *pg, bool full,
1773     void (*pr)(const char *, ...))
1774 {
1775 	struct vm_page *tpg;
1776 	struct uvm_object *uobj;
1777 	struct pgflist *pgl;
1778 	char pgbuf[128];
1779 	char pqbuf[128];
1780 
1781 	(*pr)("PAGE %p:\n", pg);
1782 	snprintb(pgbuf, sizeof(pgbuf), page_flagbits, pg->flags);
1783 	snprintb(pqbuf, sizeof(pqbuf), page_pqflagbits, pg->pqflags);
1784 	(*pr)("  flags=%s, pqflags=%s, wire_count=%d, pa=0x%lx\n",
1785 	    pgbuf, pqbuf, pg->wire_count, (long)VM_PAGE_TO_PHYS(pg));
1786 	(*pr)("  uobject=%p, uanon=%p, offset=0x%llx loan_count=%d\n",
1787 	    pg->uobject, pg->uanon, (long long)pg->offset, pg->loan_count);
1788 #if defined(UVM_PAGE_TRKOWN)
1789 	if (pg->flags & PG_BUSY)
1790 		(*pr)("  owning process = %d, tag=%s\n",
1791 		    pg->owner, pg->owner_tag);
1792 	else
1793 		(*pr)("  page not busy, no owner\n");
1794 #else
1795 	(*pr)("  [page ownership tracking disabled]\n");
1796 #endif
1797 
1798 	if (!full)
1799 		return;
1800 
1801 	/* cross-verify object/anon */
1802 	if ((pg->pqflags & PQ_FREE) == 0) {
1803 		if (pg->pqflags & PQ_ANON) {
1804 			if (pg->uanon == NULL || pg->uanon->an_page != pg)
1805 			    (*pr)("  >>> ANON DOES NOT POINT HERE <<< (%p)\n",
1806 				(pg->uanon) ? pg->uanon->an_page : NULL);
1807 			else
1808 				(*pr)("  anon backpointer is OK\n");
1809 		} else {
1810 			uobj = pg->uobject;
1811 			if (uobj) {
1812 				(*pr)("  checking object list\n");
1813 				TAILQ_FOREACH(tpg, &uobj->memq, listq.queue) {
1814 					if (tpg == pg) {
1815 						break;
1816 					}
1817 				}
1818 				if (tpg)
1819 					(*pr)("  page found on object list\n");
1820 				else
1821 			(*pr)("  >>> PAGE NOT FOUND ON OBJECT LIST! <<<\n");
1822 			}
1823 		}
1824 	}
1825 
1826 	/* cross-verify page queue */
1827 	if (pg->pqflags & PQ_FREE) {
1828 		int fl = uvm_page_lookup_freelist(pg);
1829 		int color = VM_PGCOLOR_BUCKET(pg);
1830 		pgl = &uvm.page_free[fl].pgfl_buckets[color].pgfl_queues[
1831 		    ((pg)->flags & PG_ZERO) ? PGFL_ZEROS : PGFL_UNKNOWN];
1832 	} else {
1833 		pgl = NULL;
1834 	}
1835 
1836 	if (pgl) {
1837 		(*pr)("  checking pageq list\n");
1838 		LIST_FOREACH(tpg, pgl, pageq.list) {
1839 			if (tpg == pg) {
1840 				break;
1841 			}
1842 		}
1843 		if (tpg)
1844 			(*pr)("  page found on pageq list\n");
1845 		else
1846 			(*pr)("  >>> PAGE NOT FOUND ON PAGEQ LIST! <<<\n");
1847 	}
1848 }
1849 
1850 /*
1851  * uvm_pages_printthem - print a summary of all managed pages
1852  */
1853 
1854 void
1855 uvm_page_printall(void (*pr)(const char *, ...))
1856 {
1857 	uvm_physseg_t i;
1858 	paddr_t pfn;
1859 	struct vm_page *pg;
1860 
1861 	(*pr)("%18s %4s %4s %18s %18s"
1862 #ifdef UVM_PAGE_TRKOWN
1863 	    " OWNER"
1864 #endif
1865 	    "\n", "PAGE", "FLAG", "PQ", "UOBJECT", "UANON");
1866 	for (i = uvm_physseg_get_first();
1867 	     uvm_physseg_valid_p(i);
1868 	     i = uvm_physseg_get_next(i)) {
1869 		for (pfn = uvm_physseg_get_start(i);
1870 		     pfn < uvm_physseg_get_end(i);
1871 		     pfn++) {
1872 			pg = PHYS_TO_VM_PAGE(ptoa(pfn));
1873 
1874 			(*pr)("%18p %04x %04x %18p %18p",
1875 			    pg, pg->flags, pg->pqflags, pg->uobject,
1876 			    pg->uanon);
1877 #ifdef UVM_PAGE_TRKOWN
1878 			if (pg->flags & PG_BUSY)
1879 				(*pr)(" %d [%s]", pg->owner, pg->owner_tag);
1880 #endif
1881 			(*pr)("\n");
1882 		}
1883 	}
1884 }
1885 
1886 #endif /* DDB || DEBUGPRINT */
1887