xref: /netbsd-src/sys/uvm/uvm_page.c (revision 1ca5c1b28139779176bd5c13ad7c5f25c0bcd5f8)
1 /*	$NetBSD: uvm_page.c,v 1.73 2001/12/31 19:21:37 chs Exp $	*/
2 
3 /*
4  * Copyright (c) 1997 Charles D. Cranor and Washington University.
5  * Copyright (c) 1991, 1993, The Regents of the University of California.
6  *
7  * All rights reserved.
8  *
9  * This code is derived from software contributed to Berkeley by
10  * The Mach Operating System project at Carnegie-Mellon University.
11  *
12  * Redistribution and use in source and binary forms, with or without
13  * modification, are permitted provided that the following conditions
14  * are met:
15  * 1. Redistributions of source code must retain the above copyright
16  *    notice, this list of conditions and the following disclaimer.
17  * 2. Redistributions in binary form must reproduce the above copyright
18  *    notice, this list of conditions and the following disclaimer in the
19  *    documentation and/or other materials provided with the distribution.
20  * 3. All advertising materials mentioning features or use of this software
21  *    must display the following acknowledgement:
22  *	This product includes software developed by Charles D. Cranor,
23  *      Washington University, the University of California, Berkeley and
24  *      its contributors.
25  * 4. Neither the name of the University nor the names of its contributors
26  *    may be used to endorse or promote products derived from this software
27  *    without specific prior written permission.
28  *
29  * THIS SOFTWARE IS PROVIDED BY THE REGENTS AND CONTRIBUTORS ``AS IS'' AND
30  * ANY EXPRESS OR IMPLIED WARRANTIES, INCLUDING, BUT NOT LIMITED TO, THE
31  * IMPLIED WARRANTIES OF MERCHANTABILITY AND FITNESS FOR A PARTICULAR PURPOSE
32  * ARE DISCLAIMED.  IN NO EVENT SHALL THE REGENTS OR CONTRIBUTORS BE LIABLE
33  * FOR ANY DIRECT, INDIRECT, INCIDENTAL, SPECIAL, EXEMPLARY, OR CONSEQUENTIAL
34  * DAMAGES (INCLUDING, BUT NOT LIMITED TO, PROCUREMENT OF SUBSTITUTE GOODS
35  * OR SERVICES; LOSS OF USE, DATA, OR PROFITS; OR BUSINESS INTERRUPTION)
36  * HOWEVER CAUSED AND ON ANY THEORY OF LIABILITY, WHETHER IN CONTRACT, STRICT
37  * LIABILITY, OR TORT (INCLUDING NEGLIGENCE OR OTHERWISE) ARISING IN ANY WAY
38  * OUT OF THE USE OF THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF
39  * SUCH DAMAGE.
40  *
41  *	@(#)vm_page.c   8.3 (Berkeley) 3/21/94
42  * from: Id: uvm_page.c,v 1.1.2.18 1998/02/06 05:24:42 chs Exp
43  *
44  *
45  * Copyright (c) 1987, 1990 Carnegie-Mellon University.
46  * All rights reserved.
47  *
48  * Permission to use, copy, modify and distribute this software and
49  * its documentation is hereby granted, provided that both the copyright
50  * notice and this permission notice appear in all copies of the
51  * software, derivative works or modified versions, and any portions
52  * thereof, and that both notices appear in supporting documentation.
53  *
54  * CARNEGIE MELLON ALLOWS FREE USE OF THIS SOFTWARE IN ITS "AS IS"
55  * CONDITION.  CARNEGIE MELLON DISCLAIMS ANY LIABILITY OF ANY KIND
56  * FOR ANY DAMAGES WHATSOEVER RESULTING FROM THE USE OF THIS SOFTWARE.
57  *
58  * Carnegie Mellon requests users of this software to return to
59  *
60  *  Software Distribution Coordinator  or  Software.Distribution@CS.CMU.EDU
61  *  School of Computer Science
62  *  Carnegie Mellon University
63  *  Pittsburgh PA 15213-3890
64  *
65  * any improvements or extensions that they make and grant Carnegie the
66  * rights to redistribute these changes.
67  */
68 
69 /*
70  * uvm_page.c: page ops.
71  */
72 
73 #include <sys/cdefs.h>
74 __KERNEL_RCSID(0, "$NetBSD: uvm_page.c,v 1.73 2001/12/31 19:21:37 chs Exp $");
75 
76 #include "opt_uvmhist.h"
77 
78 #include <sys/param.h>
79 #include <sys/systm.h>
80 #include <sys/malloc.h>
81 #include <sys/sched.h>
82 #include <sys/kernel.h>
83 #include <sys/vnode.h>
84 #include <sys/proc.h>
85 
86 #define UVM_PAGE                /* pull in uvm_page.h functions */
87 #include <uvm/uvm.h>
88 
89 /*
90  * global vars... XXXCDC: move to uvm. structure.
91  */
92 
93 /*
94  * physical memory config is stored in vm_physmem.
95  */
96 
97 struct vm_physseg vm_physmem[VM_PHYSSEG_MAX];	/* XXXCDC: uvm.physmem */
98 int vm_nphysseg = 0;				/* XXXCDC: uvm.nphysseg */
99 
100 /*
101  * Some supported CPUs in a given architecture don't support all
102  * of the things necessary to do idle page zero'ing efficiently.
103  * We therefore provide a way to disable it from machdep code here.
104  */
105 /*
106  * XXX disabled until we can find a way to do this without causing
107  * problems for either cpu caches or DMA latency.
108  */
109 boolean_t vm_page_zero_enable = FALSE;
110 
111 /*
112  * local variables
113  */
114 
115 /*
116  * these variables record the values returned by vm_page_bootstrap,
117  * for debugging purposes.  The implementation of uvm_pageboot_alloc
118  * and pmap_startup here also uses them internally.
119  */
120 
121 static vaddr_t      virtual_space_start;
122 static vaddr_t      virtual_space_end;
123 
124 /*
125  * we use a hash table with only one bucket during bootup.  we will
126  * later rehash (resize) the hash table once the allocator is ready.
127  * we static allocate the one bootstrap bucket below...
128  */
129 
130 static struct pglist uvm_bootbucket;
131 
132 /*
133  * we allocate an initial number of page colors in uvm_page_init(),
134  * and remember them.  We may re-color pages as cache sizes are
135  * discovered during the autoconfiguration phase.  But we can never
136  * free the initial set of buckets, since they are allocated using
137  * uvm_pageboot_alloc().
138  */
139 
140 static boolean_t have_recolored_pages /* = FALSE */;
141 
142 /*
143  * local prototypes
144  */
145 
146 static void uvm_pageinsert __P((struct vm_page *));
147 static void uvm_pageremove __P((struct vm_page *));
148 
149 /*
150  * inline functions
151  */
152 
153 /*
154  * uvm_pageinsert: insert a page in the object and the hash table
155  *
156  * => caller must lock object
157  * => caller must lock page queues
158  * => call should have already set pg's object and offset pointers
159  *    and bumped the version counter
160  */
161 
162 __inline static void
163 uvm_pageinsert(pg)
164 	struct vm_page *pg;
165 {
166 	struct pglist *buck;
167 	struct uvm_object *uobj = pg->uobject;
168 
169 	KASSERT((pg->flags & PG_TABLED) == 0);
170 	buck = &uvm.page_hash[uvm_pagehash(uobj, pg->offset)];
171 	simple_lock(&uvm.hashlock);
172 	TAILQ_INSERT_TAIL(buck, pg, hashq);
173 	simple_unlock(&uvm.hashlock);
174 
175 	TAILQ_INSERT_TAIL(&uobj->memq, pg, listq);
176 	pg->flags |= PG_TABLED;
177 	uobj->uo_npages++;
178 }
179 
180 /*
181  * uvm_page_remove: remove page from object and hash
182  *
183  * => caller must lock object
184  * => caller must lock page queues
185  */
186 
187 static __inline void
188 uvm_pageremove(pg)
189 	struct vm_page *pg;
190 {
191 	struct pglist *buck;
192 	struct uvm_object *uobj = pg->uobject;
193 
194 	KASSERT(pg->flags & PG_TABLED);
195 	buck = &uvm.page_hash[uvm_pagehash(uobj ,pg->offset)];
196 	simple_lock(&uvm.hashlock);
197 	TAILQ_REMOVE(buck, pg, hashq);
198 	simple_unlock(&uvm.hashlock);
199 
200 	if (UVM_OBJ_IS_VTEXT(uobj)) {
201 		uvmexp.execpages--;
202 	} else if (UVM_OBJ_IS_VNODE(uobj)) {
203 		uvmexp.filepages--;
204 	}
205 
206 	/* object should be locked */
207 	uobj->uo_npages--;
208 	TAILQ_REMOVE(&uobj->memq, pg, listq);
209 	pg->flags &= ~PG_TABLED;
210 	pg->uobject = NULL;
211 }
212 
213 static void
214 uvm_page_init_buckets(struct pgfreelist *pgfl)
215 {
216 	int color, i;
217 
218 	for (color = 0; color < uvmexp.ncolors; color++) {
219 		for (i = 0; i < PGFL_NQUEUES; i++) {
220 			TAILQ_INIT(&pgfl->pgfl_buckets[
221 			    color].pgfl_queues[i]);
222 		}
223 	}
224 }
225 
226 /*
227  * uvm_page_init: init the page system.   called from uvm_init().
228  *
229  * => we return the range of kernel virtual memory in kvm_startp/kvm_endp
230  */
231 
232 void
233 uvm_page_init(kvm_startp, kvm_endp)
234 	vaddr_t *kvm_startp, *kvm_endp;
235 {
236 	vsize_t freepages, pagecount, bucketcount, n;
237 	struct pgflbucket *bucketarray;
238 	struct vm_page *pagearray;
239 	int lcv, i;
240 	paddr_t paddr;
241 
242 	/*
243 	 * init the page queues and page queue locks, except the free
244 	 * list; we allocate that later (with the initial vm_page
245 	 * structures).
246 	 */
247 
248 	TAILQ_INIT(&uvm.page_active);
249 	TAILQ_INIT(&uvm.page_inactive);
250 	simple_lock_init(&uvm.pageqlock);
251 	simple_lock_init(&uvm.fpageqlock);
252 
253 	/*
254 	 * init the <obj,offset> => <page> hash table.  for now
255 	 * we just have one bucket (the bootstrap bucket).  later on we
256 	 * will allocate new buckets as we dynamically resize the hash table.
257 	 */
258 
259 	uvm.page_nhash = 1;			/* 1 bucket */
260 	uvm.page_hashmask = 0;			/* mask for hash function */
261 	uvm.page_hash = &uvm_bootbucket;	/* install bootstrap bucket */
262 	TAILQ_INIT(uvm.page_hash);		/* init hash table */
263 	simple_lock_init(&uvm.hashlock);	/* init hash table lock */
264 
265 	/*
266 	 * allocate vm_page structures.
267 	 */
268 
269 	/*
270 	 * sanity check:
271 	 * before calling this function the MD code is expected to register
272 	 * some free RAM with the uvm_page_physload() function.   our job
273 	 * now is to allocate vm_page structures for this memory.
274 	 */
275 
276 	if (vm_nphysseg == 0)
277 		panic("uvm_page_bootstrap: no memory pre-allocated");
278 
279 	/*
280 	 * first calculate the number of free pages...
281 	 *
282 	 * note that we use start/end rather than avail_start/avail_end.
283 	 * this allows us to allocate extra vm_page structures in case we
284 	 * want to return some memory to the pool after booting.
285 	 */
286 
287 	freepages = 0;
288 	for (lcv = 0 ; lcv < vm_nphysseg ; lcv++)
289 		freepages += (vm_physmem[lcv].end - vm_physmem[lcv].start);
290 
291 	/*
292 	 * Let MD code initialize the number of colors, or default
293 	 * to 1 color if MD code doesn't care.
294 	 */
295 	if (uvmexp.ncolors == 0)
296 		uvmexp.ncolors = 1;
297 	uvmexp.colormask = uvmexp.ncolors - 1;
298 
299 	/*
300 	 * we now know we have (PAGE_SIZE * freepages) bytes of memory we can
301 	 * use.   for each page of memory we use we need a vm_page structure.
302 	 * thus, the total number of pages we can use is the total size of
303 	 * the memory divided by the PAGE_SIZE plus the size of the vm_page
304 	 * structure.   we add one to freepages as a fudge factor to avoid
305 	 * truncation errors (since we can only allocate in terms of whole
306 	 * pages).
307 	 */
308 
309 	bucketcount = uvmexp.ncolors * VM_NFREELIST;
310 	pagecount = ((freepages + 1) << PAGE_SHIFT) /
311 	    (PAGE_SIZE + sizeof(struct vm_page));
312 
313 	bucketarray = (void *)uvm_pageboot_alloc((bucketcount *
314 	    sizeof(struct pgflbucket)) + (pagecount *
315 	    sizeof(struct vm_page)));
316 	pagearray = (struct vm_page *)(bucketarray + bucketcount);
317 
318 	for (lcv = 0; lcv < VM_NFREELIST; lcv++) {
319 		uvm.page_free[lcv].pgfl_buckets =
320 		    (bucketarray + (lcv * uvmexp.ncolors));
321 		uvm_page_init_buckets(&uvm.page_free[lcv]);
322 	}
323 	memset(pagearray, 0, pagecount * sizeof(struct vm_page));
324 
325 	/*
326 	 * init the vm_page structures and put them in the correct place.
327 	 */
328 
329 	for (lcv = 0 ; lcv < vm_nphysseg ; lcv++) {
330 		n = vm_physmem[lcv].end - vm_physmem[lcv].start;
331 
332 		/* set up page array pointers */
333 		vm_physmem[lcv].pgs = pagearray;
334 		pagearray += n;
335 		pagecount -= n;
336 		vm_physmem[lcv].lastpg = vm_physmem[lcv].pgs + (n - 1);
337 
338 		/* init and free vm_pages (we've already zeroed them) */
339 		paddr = ptoa(vm_physmem[lcv].start);
340 		for (i = 0 ; i < n ; i++, paddr += PAGE_SIZE) {
341 			vm_physmem[lcv].pgs[i].phys_addr = paddr;
342 #ifdef __HAVE_VM_PAGE_MD
343 			VM_MDPAGE_INIT(&vm_physmem[lcv].pgs[i]);
344 #endif
345 			if (atop(paddr) >= vm_physmem[lcv].avail_start &&
346 			    atop(paddr) <= vm_physmem[lcv].avail_end) {
347 				uvmexp.npages++;
348 				/* add page to free pool */
349 				uvm_pagefree(&vm_physmem[lcv].pgs[i]);
350 			}
351 		}
352 	}
353 
354 	/*
355 	 * pass up the values of virtual_space_start and
356 	 * virtual_space_end (obtained by uvm_pageboot_alloc) to the upper
357 	 * layers of the VM.
358 	 */
359 
360 	*kvm_startp = round_page(virtual_space_start);
361 	*kvm_endp = trunc_page(virtual_space_end);
362 
363 	/*
364 	 * init locks for kernel threads
365 	 */
366 
367 	simple_lock_init(&uvm.pagedaemon_lock);
368 	simple_lock_init(&uvm.aiodoned_lock);
369 
370 	/*
371 	 * init various thresholds.
372 	 */
373 
374 	uvmexp.reserve_pagedaemon = 1;
375 	uvmexp.reserve_kernel = 5;
376 	uvmexp.anonminpct = 10;
377 	uvmexp.fileminpct = 10;
378 	uvmexp.execminpct = 5;
379 	uvmexp.anonmaxpct = 80;
380 	uvmexp.filemaxpct = 50;
381 	uvmexp.execmaxpct = 30;
382 	uvmexp.anonmin = uvmexp.anonminpct * 256 / 100;
383 	uvmexp.filemin = uvmexp.fileminpct * 256 / 100;
384 	uvmexp.execmin = uvmexp.execminpct * 256 / 100;
385 	uvmexp.anonmax = uvmexp.anonmaxpct * 256 / 100;
386 	uvmexp.filemax = uvmexp.filemaxpct * 256 / 100;
387 	uvmexp.execmax = uvmexp.execmaxpct * 256 / 100;
388 
389 	/*
390 	 * determine if we should zero pages in the idle loop.
391 	 */
392 
393 	uvm.page_idle_zero = vm_page_zero_enable;
394 
395 	/*
396 	 * done!
397 	 */
398 
399 	uvm.page_init_done = TRUE;
400 }
401 
402 /*
403  * uvm_setpagesize: set the page size
404  *
405  * => sets page_shift and page_mask from uvmexp.pagesize.
406  */
407 
408 void
409 uvm_setpagesize()
410 {
411 	if (uvmexp.pagesize == 0)
412 		uvmexp.pagesize = DEFAULT_PAGE_SIZE;
413 	uvmexp.pagemask = uvmexp.pagesize - 1;
414 	if ((uvmexp.pagemask & uvmexp.pagesize) != 0)
415 		panic("uvm_setpagesize: page size not a power of two");
416 	for (uvmexp.pageshift = 0; ; uvmexp.pageshift++)
417 		if ((1 << uvmexp.pageshift) == uvmexp.pagesize)
418 			break;
419 }
420 
421 /*
422  * uvm_pageboot_alloc: steal memory from physmem for bootstrapping
423  */
424 
425 vaddr_t
426 uvm_pageboot_alloc(size)
427 	vsize_t size;
428 {
429 	static boolean_t initialized = FALSE;
430 	vaddr_t addr;
431 #if !defined(PMAP_STEAL_MEMORY)
432 	vaddr_t vaddr;
433 	paddr_t paddr;
434 #endif
435 
436 	/*
437 	 * on first call to this function, initialize ourselves.
438 	 */
439 	if (initialized == FALSE) {
440 		pmap_virtual_space(&virtual_space_start, &virtual_space_end);
441 
442 		/* round it the way we like it */
443 		virtual_space_start = round_page(virtual_space_start);
444 		virtual_space_end = trunc_page(virtual_space_end);
445 
446 		initialized = TRUE;
447 	}
448 
449 	/* round to page size */
450 	size = round_page(size);
451 
452 #if defined(PMAP_STEAL_MEMORY)
453 
454 	/*
455 	 * defer bootstrap allocation to MD code (it may want to allocate
456 	 * from a direct-mapped segment).  pmap_steal_memory should adjust
457 	 * virtual_space_start/virtual_space_end if necessary.
458 	 */
459 
460 	addr = pmap_steal_memory(size, &virtual_space_start,
461 	    &virtual_space_end);
462 
463 	return(addr);
464 
465 #else /* !PMAP_STEAL_MEMORY */
466 
467 	/*
468 	 * allocate virtual memory for this request
469 	 */
470 	if (virtual_space_start == virtual_space_end ||
471 	    (virtual_space_end - virtual_space_start) < size)
472 		panic("uvm_pageboot_alloc: out of virtual space");
473 
474 	addr = virtual_space_start;
475 
476 #ifdef PMAP_GROWKERNEL
477 	/*
478 	 * If the kernel pmap can't map the requested space,
479 	 * then allocate more resources for it.
480 	 */
481 	if (uvm_maxkaddr < (addr + size)) {
482 		uvm_maxkaddr = pmap_growkernel(addr + size);
483 		if (uvm_maxkaddr < (addr + size))
484 			panic("uvm_pageboot_alloc: pmap_growkernel() failed");
485 	}
486 #endif
487 
488 	virtual_space_start += size;
489 
490 	/*
491 	 * allocate and mapin physical pages to back new virtual pages
492 	 */
493 
494 	for (vaddr = round_page(addr) ; vaddr < addr + size ;
495 	    vaddr += PAGE_SIZE) {
496 
497 		if (!uvm_page_physget(&paddr))
498 			panic("uvm_pageboot_alloc: out of memory");
499 
500 		/*
501 		 * Note this memory is no longer managed, so using
502 		 * pmap_kenter is safe.
503 		 */
504 		pmap_kenter_pa(vaddr, paddr, VM_PROT_READ|VM_PROT_WRITE);
505 	}
506 	pmap_update(pmap_kernel());
507 	return(addr);
508 #endif	/* PMAP_STEAL_MEMORY */
509 }
510 
511 #if !defined(PMAP_STEAL_MEMORY)
512 /*
513  * uvm_page_physget: "steal" one page from the vm_physmem structure.
514  *
515  * => attempt to allocate it off the end of a segment in which the "avail"
516  *    values match the start/end values.   if we can't do that, then we
517  *    will advance both values (making them equal, and removing some
518  *    vm_page structures from the non-avail area).
519  * => return false if out of memory.
520  */
521 
522 /* subroutine: try to allocate from memory chunks on the specified freelist */
523 static boolean_t uvm_page_physget_freelist __P((paddr_t *, int));
524 
525 static boolean_t
526 uvm_page_physget_freelist(paddrp, freelist)
527 	paddr_t *paddrp;
528 	int freelist;
529 {
530 	int lcv, x;
531 
532 	/* pass 1: try allocating from a matching end */
533 #if (VM_PHYSSEG_STRAT == VM_PSTRAT_BIGFIRST)
534 	for (lcv = vm_nphysseg - 1 ; lcv >= 0 ; lcv--)
535 #else
536 	for (lcv = 0 ; lcv < vm_nphysseg ; lcv++)
537 #endif
538 	{
539 
540 		if (uvm.page_init_done == TRUE)
541 			panic("uvm_page_physget: called _after_ bootstrap");
542 
543 		if (vm_physmem[lcv].free_list != freelist)
544 			continue;
545 
546 		/* try from front */
547 		if (vm_physmem[lcv].avail_start == vm_physmem[lcv].start &&
548 		    vm_physmem[lcv].avail_start < vm_physmem[lcv].avail_end) {
549 			*paddrp = ptoa(vm_physmem[lcv].avail_start);
550 			vm_physmem[lcv].avail_start++;
551 			vm_physmem[lcv].start++;
552 			/* nothing left?   nuke it */
553 			if (vm_physmem[lcv].avail_start ==
554 			    vm_physmem[lcv].end) {
555 				if (vm_nphysseg == 1)
556 				    panic("vum_page_physget: out of memory!");
557 				vm_nphysseg--;
558 				for (x = lcv ; x < vm_nphysseg ; x++)
559 					/* structure copy */
560 					vm_physmem[x] = vm_physmem[x+1];
561 			}
562 			return (TRUE);
563 		}
564 
565 		/* try from rear */
566 		if (vm_physmem[lcv].avail_end == vm_physmem[lcv].end &&
567 		    vm_physmem[lcv].avail_start < vm_physmem[lcv].avail_end) {
568 			*paddrp = ptoa(vm_physmem[lcv].avail_end - 1);
569 			vm_physmem[lcv].avail_end--;
570 			vm_physmem[lcv].end--;
571 			/* nothing left?   nuke it */
572 			if (vm_physmem[lcv].avail_end ==
573 			    vm_physmem[lcv].start) {
574 				if (vm_nphysseg == 1)
575 				    panic("uvm_page_physget: out of memory!");
576 				vm_nphysseg--;
577 				for (x = lcv ; x < vm_nphysseg ; x++)
578 					/* structure copy */
579 					vm_physmem[x] = vm_physmem[x+1];
580 			}
581 			return (TRUE);
582 		}
583 	}
584 
585 	/* pass2: forget about matching ends, just allocate something */
586 #if (VM_PHYSSEG_STRAT == VM_PSTRAT_BIGFIRST)
587 	for (lcv = vm_nphysseg - 1 ; lcv >= 0 ; lcv--)
588 #else
589 	for (lcv = 0 ; lcv < vm_nphysseg ; lcv++)
590 #endif
591 	{
592 
593 		/* any room in this bank? */
594 		if (vm_physmem[lcv].avail_start >= vm_physmem[lcv].avail_end)
595 			continue;  /* nope */
596 
597 		*paddrp = ptoa(vm_physmem[lcv].avail_start);
598 		vm_physmem[lcv].avail_start++;
599 		/* truncate! */
600 		vm_physmem[lcv].start = vm_physmem[lcv].avail_start;
601 
602 		/* nothing left?   nuke it */
603 		if (vm_physmem[lcv].avail_start == vm_physmem[lcv].end) {
604 			if (vm_nphysseg == 1)
605 				panic("uvm_page_physget: out of memory!");
606 			vm_nphysseg--;
607 			for (x = lcv ; x < vm_nphysseg ; x++)
608 				/* structure copy */
609 				vm_physmem[x] = vm_physmem[x+1];
610 		}
611 		return (TRUE);
612 	}
613 
614 	return (FALSE);        /* whoops! */
615 }
616 
617 boolean_t
618 uvm_page_physget(paddrp)
619 	paddr_t *paddrp;
620 {
621 	int i;
622 
623 	/* try in the order of freelist preference */
624 	for (i = 0; i < VM_NFREELIST; i++)
625 		if (uvm_page_physget_freelist(paddrp, i) == TRUE)
626 			return (TRUE);
627 	return (FALSE);
628 }
629 #endif /* PMAP_STEAL_MEMORY */
630 
631 /*
632  * uvm_page_physload: load physical memory into VM system
633  *
634  * => all args are PFs
635  * => all pages in start/end get vm_page structures
636  * => areas marked by avail_start/avail_end get added to the free page pool
637  * => we are limited to VM_PHYSSEG_MAX physical memory segments
638  */
639 
640 void
641 uvm_page_physload(start, end, avail_start, avail_end, free_list)
642 	paddr_t start, end, avail_start, avail_end;
643 	int free_list;
644 {
645 	int preload, lcv;
646 	psize_t npages;
647 	struct vm_page *pgs;
648 	struct vm_physseg *ps;
649 
650 	if (uvmexp.pagesize == 0)
651 		panic("uvm_page_physload: page size not set!");
652 	if (free_list >= VM_NFREELIST || free_list < VM_FREELIST_DEFAULT)
653 		panic("uvm_page_physload: bad free list %d\n", free_list);
654 	if (start >= end)
655 		panic("uvm_page_physload: start >= end");
656 
657 	/*
658 	 * do we have room?
659 	 */
660 
661 	if (vm_nphysseg == VM_PHYSSEG_MAX) {
662 		printf("uvm_page_physload: unable to load physical memory "
663 		    "segment\n");
664 		printf("\t%d segments allocated, ignoring 0x%llx -> 0x%llx\n",
665 		    VM_PHYSSEG_MAX, (long long)start, (long long)end);
666 		printf("\tincrease VM_PHYSSEG_MAX\n");
667 		return;
668 	}
669 
670 	/*
671 	 * check to see if this is a "preload" (i.e. uvm_mem_init hasn't been
672 	 * called yet, so malloc is not available).
673 	 */
674 
675 	for (lcv = 0 ; lcv < vm_nphysseg ; lcv++) {
676 		if (vm_physmem[lcv].pgs)
677 			break;
678 	}
679 	preload = (lcv == vm_nphysseg);
680 
681 	/*
682 	 * if VM is already running, attempt to malloc() vm_page structures
683 	 */
684 
685 	if (!preload) {
686 #if defined(VM_PHYSSEG_NOADD)
687 		panic("uvm_page_physload: tried to add RAM after vm_mem_init");
688 #else
689 		/* XXXCDC: need some sort of lockout for this case */
690 		paddr_t paddr;
691 		npages = end - start;  /* # of pages */
692 		pgs = malloc(sizeof(struct vm_page) * npages,
693 		    M_VMPAGE, M_NOWAIT);
694 		if (pgs == NULL) {
695 			printf("uvm_page_physload: can not malloc vm_page "
696 			    "structs for segment\n");
697 			printf("\tignoring 0x%lx -> 0x%lx\n", start, end);
698 			return;
699 		}
700 		/* zero data, init phys_addr and free_list, and free pages */
701 		memset(pgs, 0, sizeof(struct vm_page) * npages);
702 		for (lcv = 0, paddr = ptoa(start) ;
703 				 lcv < npages ; lcv++, paddr += PAGE_SIZE) {
704 			pgs[lcv].phys_addr = paddr;
705 			pgs[lcv].free_list = free_list;
706 			if (atop(paddr) >= avail_start &&
707 			    atop(paddr) <= avail_end)
708 				uvm_pagefree(&pgs[lcv]);
709 		}
710 		/* XXXCDC: incomplete: need to update uvmexp.free, what else? */
711 		/* XXXCDC: need hook to tell pmap to rebuild pv_list, etc... */
712 #endif
713 	} else {
714 		pgs = NULL;
715 		npages = 0;
716 	}
717 
718 	/*
719 	 * now insert us in the proper place in vm_physmem[]
720 	 */
721 
722 #if (VM_PHYSSEG_STRAT == VM_PSTRAT_RANDOM)
723 	/* random: put it at the end (easy!) */
724 	ps = &vm_physmem[vm_nphysseg];
725 #elif (VM_PHYSSEG_STRAT == VM_PSTRAT_BSEARCH)
726 	{
727 		int x;
728 		/* sort by address for binary search */
729 		for (lcv = 0 ; lcv < vm_nphysseg ; lcv++)
730 			if (start < vm_physmem[lcv].start)
731 				break;
732 		ps = &vm_physmem[lcv];
733 		/* move back other entries, if necessary ... */
734 		for (x = vm_nphysseg ; x > lcv ; x--)
735 			/* structure copy */
736 			vm_physmem[x] = vm_physmem[x - 1];
737 	}
738 #elif (VM_PHYSSEG_STRAT == VM_PSTRAT_BIGFIRST)
739 	{
740 		int x;
741 		/* sort by largest segment first */
742 		for (lcv = 0 ; lcv < vm_nphysseg ; lcv++)
743 			if ((end - start) >
744 			    (vm_physmem[lcv].end - vm_physmem[lcv].start))
745 				break;
746 		ps = &vm_physmem[lcv];
747 		/* move back other entries, if necessary ... */
748 		for (x = vm_nphysseg ; x > lcv ; x--)
749 			/* structure copy */
750 			vm_physmem[x] = vm_physmem[x - 1];
751 	}
752 #else
753 	panic("uvm_page_physload: unknown physseg strategy selected!");
754 #endif
755 
756 	ps->start = start;
757 	ps->end = end;
758 	ps->avail_start = avail_start;
759 	ps->avail_end = avail_end;
760 	if (preload) {
761 		ps->pgs = NULL;
762 	} else {
763 		ps->pgs = pgs;
764 		ps->lastpg = pgs + npages - 1;
765 	}
766 	ps->free_list = free_list;
767 	vm_nphysseg++;
768 
769 	if (!preload)
770 		uvm_page_rehash();
771 }
772 
773 /*
774  * uvm_page_rehash: reallocate hash table based on number of free pages.
775  */
776 
777 void
778 uvm_page_rehash()
779 {
780 	int freepages, lcv, bucketcount, oldcount;
781 	struct pglist *newbuckets, *oldbuckets;
782 	struct vm_page *pg;
783 	size_t newsize, oldsize;
784 
785 	/*
786 	 * compute number of pages that can go in the free pool
787 	 */
788 
789 	freepages = 0;
790 	for (lcv = 0 ; lcv < vm_nphysseg ; lcv++)
791 		freepages +=
792 		    (vm_physmem[lcv].avail_end - vm_physmem[lcv].avail_start);
793 
794 	/*
795 	 * compute number of buckets needed for this number of pages
796 	 */
797 
798 	bucketcount = 1;
799 	while (bucketcount < freepages)
800 		bucketcount = bucketcount * 2;
801 
802 	/*
803 	 * compute the size of the current table and new table.
804 	 */
805 
806 	oldbuckets = uvm.page_hash;
807 	oldcount = uvm.page_nhash;
808 	oldsize = round_page(sizeof(struct pglist) * oldcount);
809 	newsize = round_page(sizeof(struct pglist) * bucketcount);
810 
811 	/*
812 	 * allocate the new buckets
813 	 */
814 
815 	newbuckets = (struct pglist *) uvm_km_alloc(kernel_map, newsize);
816 	if (newbuckets == NULL) {
817 		printf("uvm_page_physrehash: WARNING: could not grow page "
818 		    "hash table\n");
819 		return;
820 	}
821 	for (lcv = 0 ; lcv < bucketcount ; lcv++)
822 		TAILQ_INIT(&newbuckets[lcv]);
823 
824 	/*
825 	 * now replace the old buckets with the new ones and rehash everything
826 	 */
827 
828 	simple_lock(&uvm.hashlock);
829 	uvm.page_hash = newbuckets;
830 	uvm.page_nhash = bucketcount;
831 	uvm.page_hashmask = bucketcount - 1;  /* power of 2 */
832 
833 	/* ... and rehash */
834 	for (lcv = 0 ; lcv < oldcount ; lcv++) {
835 		while ((pg = oldbuckets[lcv].tqh_first) != NULL) {
836 			TAILQ_REMOVE(&oldbuckets[lcv], pg, hashq);
837 			TAILQ_INSERT_TAIL(
838 			  &uvm.page_hash[uvm_pagehash(pg->uobject, pg->offset)],
839 			  pg, hashq);
840 		}
841 	}
842 	simple_unlock(&uvm.hashlock);
843 
844 	/*
845 	 * free old bucket array if is not the boot-time table
846 	 */
847 
848 	if (oldbuckets != &uvm_bootbucket)
849 		uvm_km_free(kernel_map, (vaddr_t) oldbuckets, oldsize);
850 }
851 
852 /*
853  * uvm_page_recolor: Recolor the pages if the new bucket count is
854  * larger than the old one.
855  */
856 
857 void
858 uvm_page_recolor(int newncolors)
859 {
860 	struct pgflbucket *bucketarray, *oldbucketarray;
861 	struct pgfreelist pgfl;
862 	struct vm_page *pg;
863 	vsize_t bucketcount;
864 	int s, lcv, color, i, ocolors;
865 
866 	if (newncolors <= uvmexp.ncolors)
867 		return;
868 
869 	bucketcount = newncolors * VM_NFREELIST;
870 	bucketarray = malloc(bucketcount * sizeof(struct pgflbucket),
871 	    M_VMPAGE, M_NOWAIT);
872 	if (bucketarray == NULL) {
873 		printf("WARNING: unable to allocate %ld page color buckets\n",
874 		    (long) bucketcount);
875 		return;
876 	}
877 
878 	s = uvm_lock_fpageq();
879 
880 	/* Make sure we should still do this. */
881 	if (newncolors <= uvmexp.ncolors) {
882 		uvm_unlock_fpageq(s);
883 		free(bucketarray, M_VMPAGE);
884 		return;
885 	}
886 
887 	oldbucketarray = uvm.page_free[0].pgfl_buckets;
888 	ocolors = uvmexp.ncolors;
889 
890 	uvmexp.ncolors = newncolors;
891 	uvmexp.colormask = uvmexp.ncolors - 1;
892 
893 	for (lcv = 0; lcv < VM_NFREELIST; lcv++) {
894 		pgfl.pgfl_buckets = (bucketarray + (lcv * newncolors));
895 		uvm_page_init_buckets(&pgfl);
896 		for (color = 0; color < ocolors; color++) {
897 			for (i = 0; i < PGFL_NQUEUES; i++) {
898 				while ((pg = TAILQ_FIRST(&uvm.page_free[
899 				    lcv].pgfl_buckets[color].pgfl_queues[i]))
900 				    != NULL) {
901 					TAILQ_REMOVE(&uvm.page_free[
902 					    lcv].pgfl_buckets[
903 					    color].pgfl_queues[i], pg, pageq);
904 					TAILQ_INSERT_TAIL(&pgfl.pgfl_buckets[
905 					    VM_PGCOLOR_BUCKET(pg)].pgfl_queues[
906 					    i], pg, pageq);
907 				}
908 			}
909 		}
910 		uvm.page_free[lcv].pgfl_buckets = pgfl.pgfl_buckets;
911 	}
912 
913 	if (have_recolored_pages) {
914 		uvm_unlock_fpageq(s);
915 		free(oldbucketarray, M_VMPAGE);
916 		return;
917 	}
918 
919 	have_recolored_pages = TRUE;
920 	uvm_unlock_fpageq(s);
921 }
922 
923 /*
924  * uvm_pagealloc_pgfl: helper routine for uvm_pagealloc_strat
925  */
926 
927 static __inline struct vm_page *
928 uvm_pagealloc_pgfl(struct pgfreelist *pgfl, int try1, int try2,
929     int *trycolorp)
930 {
931 	struct pglist *freeq;
932 	struct vm_page *pg;
933 	int color, trycolor = *trycolorp;
934 
935 	color = trycolor;
936 	do {
937 		if ((pg = TAILQ_FIRST((freeq =
938 		    &pgfl->pgfl_buckets[color].pgfl_queues[try1]))) != NULL)
939 			goto gotit;
940 		if ((pg = TAILQ_FIRST((freeq =
941 		    &pgfl->pgfl_buckets[color].pgfl_queues[try2]))) != NULL)
942 			goto gotit;
943 		color = (color + 1) & uvmexp.colormask;
944 	} while (color != trycolor);
945 
946 	return (NULL);
947 
948  gotit:
949 	TAILQ_REMOVE(freeq, pg, pageq);
950 	uvmexp.free--;
951 
952 	/* update zero'd page count */
953 	if (pg->flags & PG_ZERO)
954 		uvmexp.zeropages--;
955 
956 	if (color == trycolor)
957 		uvmexp.colorhit++;
958 	else {
959 		uvmexp.colormiss++;
960 		*trycolorp = color;
961 	}
962 
963 	return (pg);
964 }
965 
966 /*
967  * uvm_pagealloc_strat: allocate vm_page from a particular free list.
968  *
969  * => return null if no pages free
970  * => wake up pagedaemon if number of free pages drops below low water mark
971  * => if obj != NULL, obj must be locked (to put in hash)
972  * => if anon != NULL, anon must be locked (to put in anon)
973  * => only one of obj or anon can be non-null
974  * => caller must activate/deactivate page if it is not wired.
975  * => free_list is ignored if strat == UVM_PGA_STRAT_NORMAL.
976  * => policy decision: it is more important to pull a page off of the
977  *	appropriate priority free list than it is to get a zero'd or
978  *	unknown contents page.  This is because we live with the
979  *	consequences of a bad free list decision for the entire
980  *	lifetime of the page, e.g. if the page comes from memory that
981  *	is slower to access.
982  */
983 
984 struct vm_page *
985 uvm_pagealloc_strat(obj, off, anon, flags, strat, free_list)
986 	struct uvm_object *obj;
987 	voff_t off;
988 	int flags;
989 	struct vm_anon *anon;
990 	int strat, free_list;
991 {
992 	int lcv, try1, try2, s, zeroit = 0, color;
993 	struct vm_page *pg;
994 	boolean_t use_reserve;
995 
996 	KASSERT(obj == NULL || anon == NULL);
997 	KASSERT(off == trunc_page(off));
998 	LOCK_ASSERT(obj == NULL || simple_lock_held(&obj->vmobjlock));
999 	LOCK_ASSERT(anon == NULL || simple_lock_held(&anon->an_lock));
1000 
1001 	s = uvm_lock_fpageq();
1002 
1003 	/*
1004 	 * This implements a global round-robin page coloring
1005 	 * algorithm.
1006 	 *
1007 	 * XXXJRT: Should we make the `nextcolor' per-cpu?
1008 	 * XXXJRT: What about virtually-indexed caches?
1009 	 */
1010 
1011 	color = uvm.page_free_nextcolor;
1012 
1013 	/*
1014 	 * check to see if we need to generate some free pages waking
1015 	 * the pagedaemon.
1016 	 */
1017 
1018 	UVM_KICK_PDAEMON();
1019 
1020 	/*
1021 	 * fail if any of these conditions is true:
1022 	 * [1]  there really are no free pages, or
1023 	 * [2]  only kernel "reserved" pages remain and
1024 	 *        the page isn't being allocated to a kernel object.
1025 	 * [3]  only pagedaemon "reserved" pages remain and
1026 	 *        the requestor isn't the pagedaemon.
1027 	 */
1028 
1029 	use_reserve = (flags & UVM_PGA_USERESERVE) ||
1030 		(obj && UVM_OBJ_IS_KERN_OBJECT(obj));
1031 	if ((uvmexp.free <= uvmexp.reserve_kernel && !use_reserve) ||
1032 	    (uvmexp.free <= uvmexp.reserve_pagedaemon &&
1033 	     !(use_reserve && curproc == uvm.pagedaemon_proc)))
1034 		goto fail;
1035 
1036 #if PGFL_NQUEUES != 2
1037 #error uvm_pagealloc_strat needs to be updated
1038 #endif
1039 
1040 	/*
1041 	 * If we want a zero'd page, try the ZEROS queue first, otherwise
1042 	 * we try the UNKNOWN queue first.
1043 	 */
1044 	if (flags & UVM_PGA_ZERO) {
1045 		try1 = PGFL_ZEROS;
1046 		try2 = PGFL_UNKNOWN;
1047 	} else {
1048 		try1 = PGFL_UNKNOWN;
1049 		try2 = PGFL_ZEROS;
1050 	}
1051 
1052  again:
1053 	switch (strat) {
1054 	case UVM_PGA_STRAT_NORMAL:
1055 		/* Check all freelists in descending priority order. */
1056 		for (lcv = 0; lcv < VM_NFREELIST; lcv++) {
1057 			pg = uvm_pagealloc_pgfl(&uvm.page_free[lcv],
1058 			    try1, try2, &color);
1059 			if (pg != NULL)
1060 				goto gotit;
1061 		}
1062 
1063 		/* No pages free! */
1064 		goto fail;
1065 
1066 	case UVM_PGA_STRAT_ONLY:
1067 	case UVM_PGA_STRAT_FALLBACK:
1068 		/* Attempt to allocate from the specified free list. */
1069 		KASSERT(free_list >= 0 && free_list < VM_NFREELIST);
1070 		pg = uvm_pagealloc_pgfl(&uvm.page_free[free_list],
1071 		    try1, try2, &color);
1072 		if (pg != NULL)
1073 			goto gotit;
1074 
1075 		/* Fall back, if possible. */
1076 		if (strat == UVM_PGA_STRAT_FALLBACK) {
1077 			strat = UVM_PGA_STRAT_NORMAL;
1078 			goto again;
1079 		}
1080 
1081 		/* No pages free! */
1082 		goto fail;
1083 
1084 	default:
1085 		panic("uvm_pagealloc_strat: bad strat %d", strat);
1086 		/* NOTREACHED */
1087 	}
1088 
1089  gotit:
1090 	/*
1091 	 * We now know which color we actually allocated from; set
1092 	 * the next color accordingly.
1093 	 */
1094 
1095 	uvm.page_free_nextcolor = (color + 1) & uvmexp.colormask;
1096 
1097 	/*
1098 	 * update allocation statistics and remember if we have to
1099 	 * zero the page
1100 	 */
1101 
1102 	if (flags & UVM_PGA_ZERO) {
1103 		if (pg->flags & PG_ZERO) {
1104 			uvmexp.pga_zerohit++;
1105 			zeroit = 0;
1106 		} else {
1107 			uvmexp.pga_zeromiss++;
1108 			zeroit = 1;
1109 		}
1110 	}
1111 	uvm_unlock_fpageq(s);
1112 
1113 	pg->offset = off;
1114 	pg->uobject = obj;
1115 	pg->uanon = anon;
1116 	pg->flags = PG_BUSY|PG_CLEAN|PG_FAKE;
1117 	if (anon) {
1118 		anon->u.an_page = pg;
1119 		pg->pqflags = PQ_ANON;
1120 		uvmexp.anonpages++;
1121 	} else {
1122 		if (obj) {
1123 			uvm_pageinsert(pg);
1124 		}
1125 		pg->pqflags = 0;
1126 	}
1127 #if defined(UVM_PAGE_TRKOWN)
1128 	pg->owner_tag = NULL;
1129 #endif
1130 	UVM_PAGE_OWN(pg, "new alloc");
1131 
1132 	if (flags & UVM_PGA_ZERO) {
1133 		/*
1134 		 * A zero'd page is not clean.  If we got a page not already
1135 		 * zero'd, then we have to zero it ourselves.
1136 		 */
1137 		pg->flags &= ~PG_CLEAN;
1138 		if (zeroit)
1139 			pmap_zero_page(VM_PAGE_TO_PHYS(pg));
1140 	}
1141 
1142 	return(pg);
1143 
1144  fail:
1145 	uvm_unlock_fpageq(s);
1146 	return (NULL);
1147 }
1148 
1149 /*
1150  * uvm_pagerealloc: reallocate a page from one object to another
1151  *
1152  * => both objects must be locked
1153  */
1154 
1155 void
1156 uvm_pagerealloc(pg, newobj, newoff)
1157 	struct vm_page *pg;
1158 	struct uvm_object *newobj;
1159 	voff_t newoff;
1160 {
1161 	/*
1162 	 * remove it from the old object
1163 	 */
1164 
1165 	if (pg->uobject) {
1166 		uvm_pageremove(pg);
1167 	}
1168 
1169 	/*
1170 	 * put it in the new object
1171 	 */
1172 
1173 	if (newobj) {
1174 		pg->uobject = newobj;
1175 		pg->offset = newoff;
1176 		uvm_pageinsert(pg);
1177 	}
1178 }
1179 
1180 /*
1181  * uvm_pagefree: free page
1182  *
1183  * => erase page's identity (i.e. remove from hash/object)
1184  * => put page on free list
1185  * => caller must lock owning object (either anon or uvm_object)
1186  * => caller must lock page queues
1187  * => assumes all valid mappings of pg are gone
1188  */
1189 
1190 void
1191 uvm_pagefree(pg)
1192 	struct vm_page *pg;
1193 {
1194 	int s;
1195 
1196 	KASSERT((pg->flags & PG_PAGEOUT) == 0);
1197 	LOCK_ASSERT(simple_lock_held(&uvm.pageqlock) ||
1198 		    (pg->pqflags & (PQ_ACTIVE|PQ_INACTIVE)) == 0);
1199 	LOCK_ASSERT(pg->uobject == NULL ||
1200 		    simple_lock_held(&pg->uobject->vmobjlock));
1201 	LOCK_ASSERT(pg->uobject != NULL || pg->uanon == NULL ||
1202 		    simple_lock_held(&pg->uanon->an_lock));
1203 
1204 #ifdef DEBUG
1205 	if (pg->uobject == (void *)0xdeadbeef &&
1206 	    pg->uanon == (void *)0xdeadbeef) {
1207 		panic("uvm_pagefree: freeing free page %p\n", pg);
1208 	}
1209 #endif
1210 
1211 	/*
1212 	 * if the page is loaned, resolve the loan instead of freeing.
1213 	 */
1214 
1215 	if (pg->loan_count) {
1216 		KASSERT(pg->wire_count == 0);
1217 
1218 		/*
1219 		 * if the page is owned by an anon then we just want to
1220 		 * drop anon ownership.  the kernel will free the page when
1221 		 * it is done with it.  if the page is owned by an object,
1222 		 * remove it from the object and mark it dirty for the benefit
1223 		 * of possible anon owners.
1224 		 *
1225 		 * regardless of previous ownership, wakeup any waiters,
1226 		 * unbusy the page, and we're done.
1227 		 */
1228 
1229 		if (pg->uobject != NULL) {
1230 			uvm_pageremove(pg);
1231 			pg->flags &= ~PG_CLEAN;
1232 		} else if (pg->uanon != NULL) {
1233 			if ((pg->pqflags & PQ_ANON) == 0) {
1234 				pg->loan_count--;
1235 			} else {
1236 				pg->pqflags &= ~PQ_ANON;
1237 			}
1238 			pg->uanon = NULL;
1239 		}
1240 		if (pg->flags & PG_WANTED) {
1241 			wakeup(pg);
1242 		}
1243 		pg->flags &= ~(PG_WANTED|PG_BUSY|PG_RELEASED);
1244 #ifdef UVM_PAGE_TRKOWN
1245 		pg->owner_tag = NULL;
1246 #endif
1247 		if (pg->loan_count) {
1248 			return;
1249 		}
1250 	}
1251 
1252 	/*
1253 	 * remove page from its object or anon.
1254 	 */
1255 
1256 	if (pg->uobject != NULL) {
1257 		uvm_pageremove(pg);
1258 	} else if (pg->uanon != NULL) {
1259 		pg->uanon->u.an_page = NULL;
1260 		uvmexp.anonpages--;
1261 	}
1262 
1263 	/*
1264 	 * now remove the page from the queues.
1265 	 */
1266 
1267 	uvm_pagedequeue(pg);
1268 
1269 	/*
1270 	 * if the page was wired, unwire it now.
1271 	 */
1272 
1273 	if (pg->wire_count) {
1274 		pg->wire_count = 0;
1275 		uvmexp.wired--;
1276 	}
1277 
1278 	/*
1279 	 * and put on free queue
1280 	 */
1281 
1282 	pg->flags &= ~PG_ZERO;
1283 
1284 	s = uvm_lock_fpageq();
1285 	TAILQ_INSERT_TAIL(&uvm.page_free[
1286 	    uvm_page_lookup_freelist(pg)].pgfl_buckets[
1287 	    VM_PGCOLOR_BUCKET(pg)].pgfl_queues[PGFL_UNKNOWN], pg, pageq);
1288 	pg->pqflags = PQ_FREE;
1289 #ifdef DEBUG
1290 	pg->uobject = (void *)0xdeadbeef;
1291 	pg->offset = 0xdeadbeef;
1292 	pg->uanon = (void *)0xdeadbeef;
1293 #endif
1294 	uvmexp.free++;
1295 
1296 	if (uvmexp.zeropages < UVM_PAGEZERO_TARGET)
1297 		uvm.page_idle_zero = vm_page_zero_enable;
1298 
1299 	uvm_unlock_fpageq(s);
1300 }
1301 
1302 /*
1303  * uvm_page_unbusy: unbusy an array of pages.
1304  *
1305  * => pages must either all belong to the same object, or all belong to anons.
1306  * => if pages are object-owned, object must be locked.
1307  * => if pages are anon-owned, anons must be locked.
1308  */
1309 
1310 void
1311 uvm_page_unbusy(pgs, npgs)
1312 	struct vm_page **pgs;
1313 	int npgs;
1314 {
1315 	struct vm_page *pg;
1316 	int i;
1317 	UVMHIST_FUNC("uvm_page_unbusy"); UVMHIST_CALLED(ubchist);
1318 
1319 	for (i = 0; i < npgs; i++) {
1320 		pg = pgs[i];
1321 		if (pg == NULL) {
1322 			continue;
1323 		}
1324 		if (pg->flags & PG_WANTED) {
1325 			wakeup(pg);
1326 		}
1327 		if (pg->flags & PG_RELEASED) {
1328 			UVMHIST_LOG(ubchist, "releasing pg %p", pg,0,0,0);
1329 			pg->flags &= ~PG_RELEASED;
1330 			uvm_pagefree(pg);
1331 		} else {
1332 			UVMHIST_LOG(ubchist, "unbusying pg %p", pg,0,0,0);
1333 			pg->flags &= ~(PG_WANTED|PG_BUSY);
1334 			UVM_PAGE_OWN(pg, NULL);
1335 		}
1336 	}
1337 }
1338 
1339 #if defined(UVM_PAGE_TRKOWN)
1340 /*
1341  * uvm_page_own: set or release page ownership
1342  *
1343  * => this is a debugging function that keeps track of who sets PG_BUSY
1344  *	and where they do it.   it can be used to track down problems
1345  *	such a process setting "PG_BUSY" and never releasing it.
1346  * => page's object [if any] must be locked
1347  * => if "tag" is NULL then we are releasing page ownership
1348  */
1349 void
1350 uvm_page_own(pg, tag)
1351 	struct vm_page *pg;
1352 	char *tag;
1353 {
1354 	KASSERT((pg->flags & (PG_PAGEOUT|PG_RELEASED)) == 0);
1355 
1356 	/* gain ownership? */
1357 	if (tag) {
1358 		if (pg->owner_tag) {
1359 			printf("uvm_page_own: page %p already owned "
1360 			    "by proc %d [%s]\n", pg,
1361 			     pg->owner, pg->owner_tag);
1362 			panic("uvm_page_own");
1363 		}
1364 		pg->owner = (curproc) ? curproc->p_pid :  (pid_t) -1;
1365 		pg->owner_tag = tag;
1366 		return;
1367 	}
1368 
1369 	/* drop ownership */
1370 	if (pg->owner_tag == NULL) {
1371 		printf("uvm_page_own: dropping ownership of an non-owned "
1372 		    "page (%p)\n", pg);
1373 		panic("uvm_page_own");
1374 	}
1375 	pg->owner_tag = NULL;
1376 	KASSERT((pg->pqflags & (PQ_ACTIVE|PQ_INACTIVE)) ||
1377 		(pg->uanon == NULL && pg->uobject == NULL) ||
1378 		pg->uobject == uvm.kernel_object ||
1379 		pg->wire_count > 0 ||
1380 		(pg->loan_count == 1 && pg->uanon == NULL) ||
1381 		pg->loan_count > 1);
1382 	return;
1383 }
1384 #endif
1385 
1386 /*
1387  * uvm_pageidlezero: zero free pages while the system is idle.
1388  *
1389  * => try to complete one color bucket at a time, to reduce our impact
1390  *	on the CPU cache.
1391  * => we loop until we either reach the target or whichqs indicates that
1392  *	there is a process ready to run.
1393  */
1394 void
1395 uvm_pageidlezero()
1396 {
1397 	struct vm_page *pg;
1398 	struct pgfreelist *pgfl;
1399 	int free_list, s, firstbucket;
1400 	static int nextbucket;
1401 
1402 	s = uvm_lock_fpageq();
1403 	firstbucket = nextbucket;
1404 	do {
1405 		if (sched_whichqs != 0) {
1406 			uvm_unlock_fpageq(s);
1407 			return;
1408 		}
1409 		if (uvmexp.zeropages >= UVM_PAGEZERO_TARGET) {
1410 			uvm.page_idle_zero = FALSE;
1411 			uvm_unlock_fpageq(s);
1412 			return;
1413 		}
1414 		for (free_list = 0; free_list < VM_NFREELIST; free_list++) {
1415 			pgfl = &uvm.page_free[free_list];
1416 			while ((pg = TAILQ_FIRST(&pgfl->pgfl_buckets[
1417 			    nextbucket].pgfl_queues[PGFL_UNKNOWN])) != NULL) {
1418 				if (sched_whichqs != 0) {
1419 					uvm_unlock_fpageq(s);
1420 					return;
1421 				}
1422 
1423 				TAILQ_REMOVE(&pgfl->pgfl_buckets[
1424 				    nextbucket].pgfl_queues[PGFL_UNKNOWN],
1425 				    pg, pageq);
1426 				uvmexp.free--;
1427 				uvm_unlock_fpageq(s);
1428 #ifdef PMAP_PAGEIDLEZERO
1429 				if (!PMAP_PAGEIDLEZERO(VM_PAGE_TO_PHYS(pg))) {
1430 
1431 					/*
1432 					 * The machine-dependent code detected
1433 					 * some reason for us to abort zeroing
1434 					 * pages, probably because there is a
1435 					 * process now ready to run.
1436 					 */
1437 
1438 					s = uvm_lock_fpageq();
1439 					TAILQ_INSERT_HEAD(&pgfl->pgfl_buckets[
1440 					    nextbucket].pgfl_queues[
1441 					    PGFL_UNKNOWN], pg, pageq);
1442 					uvmexp.free++;
1443 					uvmexp.zeroaborts++;
1444 					uvm_unlock_fpageq(s);
1445 					return;
1446 				}
1447 #else
1448 				pmap_zero_page(VM_PAGE_TO_PHYS(pg));
1449 #endif /* PMAP_PAGEIDLEZERO */
1450 				pg->flags |= PG_ZERO;
1451 
1452 				s = uvm_lock_fpageq();
1453 				TAILQ_INSERT_HEAD(&pgfl->pgfl_buckets[
1454 				    nextbucket].pgfl_queues[PGFL_ZEROS],
1455 				    pg, pageq);
1456 				uvmexp.free++;
1457 				uvmexp.zeropages++;
1458 			}
1459 		}
1460 		nextbucket = (nextbucket + 1) & uvmexp.colormask;
1461 	} while (nextbucket != firstbucket);
1462 	uvm_unlock_fpageq(s);
1463 }
1464