xref: /netbsd-src/sys/uvm/uvm_page.c (revision 06be8101a16cc95f40783b3cb7afd12112103a9a)
1 /*	$NetBSD: uvm_page.c,v 1.71 2001/11/10 07:37:00 lukem 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 1.71 2001/11/10 07:37:00 lukem 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.vtextpages--;
202 	} else if (UVM_OBJ_IS_VNODE(uobj)) {
203 		uvmexp.vnodepages--;
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 	 * XXXCDC - values may need adjusting
373 	 */
374 
375 	uvmexp.reserve_pagedaemon = 1;
376 	uvmexp.reserve_kernel = 5;
377 	uvmexp.anonminpct = 10;
378 	uvmexp.vnodeminpct = 10;
379 	uvmexp.vtextminpct = 5;
380 	uvmexp.anonmin = uvmexp.anonminpct * 256 / 100;
381 	uvmexp.vnodemin = uvmexp.vnodeminpct * 256 / 100;
382 	uvmexp.vtextmin = uvmexp.vtextminpct * 256 / 100;
383 
384 	/*
385 	 * determine if we should zero pages in the idle loop.
386 	 */
387 
388 	uvm.page_idle_zero = vm_page_zero_enable;
389 
390 	/*
391 	 * done!
392 	 */
393 
394 	uvm.page_init_done = TRUE;
395 }
396 
397 /*
398  * uvm_setpagesize: set the page size
399  *
400  * => sets page_shift and page_mask from uvmexp.pagesize.
401  */
402 
403 void
404 uvm_setpagesize()
405 {
406 	if (uvmexp.pagesize == 0)
407 		uvmexp.pagesize = DEFAULT_PAGE_SIZE;
408 	uvmexp.pagemask = uvmexp.pagesize - 1;
409 	if ((uvmexp.pagemask & uvmexp.pagesize) != 0)
410 		panic("uvm_setpagesize: page size not a power of two");
411 	for (uvmexp.pageshift = 0; ; uvmexp.pageshift++)
412 		if ((1 << uvmexp.pageshift) == uvmexp.pagesize)
413 			break;
414 }
415 
416 /*
417  * uvm_pageboot_alloc: steal memory from physmem for bootstrapping
418  */
419 
420 vaddr_t
421 uvm_pageboot_alloc(size)
422 	vsize_t size;
423 {
424 	static boolean_t initialized = FALSE;
425 	vaddr_t addr;
426 #if !defined(PMAP_STEAL_MEMORY)
427 	vaddr_t vaddr;
428 	paddr_t paddr;
429 #endif
430 
431 	/*
432 	 * on first call to this function, initialize ourselves.
433 	 */
434 	if (initialized == FALSE) {
435 		pmap_virtual_space(&virtual_space_start, &virtual_space_end);
436 
437 		/* round it the way we like it */
438 		virtual_space_start = round_page(virtual_space_start);
439 		virtual_space_end = trunc_page(virtual_space_end);
440 
441 		initialized = TRUE;
442 	}
443 
444 	/* round to page size */
445 	size = round_page(size);
446 
447 #if defined(PMAP_STEAL_MEMORY)
448 
449 	/*
450 	 * defer bootstrap allocation to MD code (it may want to allocate
451 	 * from a direct-mapped segment).  pmap_steal_memory should adjust
452 	 * virtual_space_start/virtual_space_end if necessary.
453 	 */
454 
455 	addr = pmap_steal_memory(size, &virtual_space_start,
456 	    &virtual_space_end);
457 
458 	return(addr);
459 
460 #else /* !PMAP_STEAL_MEMORY */
461 
462 	/*
463 	 * allocate virtual memory for this request
464 	 */
465 	if (virtual_space_start == virtual_space_end ||
466 	    (virtual_space_end - virtual_space_start) < size)
467 		panic("uvm_pageboot_alloc: out of virtual space");
468 
469 	addr = virtual_space_start;
470 
471 #ifdef PMAP_GROWKERNEL
472 	/*
473 	 * If the kernel pmap can't map the requested space,
474 	 * then allocate more resources for it.
475 	 */
476 	if (uvm_maxkaddr < (addr + size)) {
477 		uvm_maxkaddr = pmap_growkernel(addr + size);
478 		if (uvm_maxkaddr < (addr + size))
479 			panic("uvm_pageboot_alloc: pmap_growkernel() failed");
480 	}
481 #endif
482 
483 	virtual_space_start += size;
484 
485 	/*
486 	 * allocate and mapin physical pages to back new virtual pages
487 	 */
488 
489 	for (vaddr = round_page(addr) ; vaddr < addr + size ;
490 	    vaddr += PAGE_SIZE) {
491 
492 		if (!uvm_page_physget(&paddr))
493 			panic("uvm_pageboot_alloc: out of memory");
494 
495 		/*
496 		 * Note this memory is no longer managed, so using
497 		 * pmap_kenter is safe.
498 		 */
499 		pmap_kenter_pa(vaddr, paddr, VM_PROT_READ|VM_PROT_WRITE);
500 	}
501 	pmap_update(pmap_kernel());
502 	return(addr);
503 #endif	/* PMAP_STEAL_MEMORY */
504 }
505 
506 #if !defined(PMAP_STEAL_MEMORY)
507 /*
508  * uvm_page_physget: "steal" one page from the vm_physmem structure.
509  *
510  * => attempt to allocate it off the end of a segment in which the "avail"
511  *    values match the start/end values.   if we can't do that, then we
512  *    will advance both values (making them equal, and removing some
513  *    vm_page structures from the non-avail area).
514  * => return false if out of memory.
515  */
516 
517 /* subroutine: try to allocate from memory chunks on the specified freelist */
518 static boolean_t uvm_page_physget_freelist __P((paddr_t *, int));
519 
520 static boolean_t
521 uvm_page_physget_freelist(paddrp, freelist)
522 	paddr_t *paddrp;
523 	int freelist;
524 {
525 	int lcv, x;
526 
527 	/* pass 1: try allocating from a matching end */
528 #if (VM_PHYSSEG_STRAT == VM_PSTRAT_BIGFIRST)
529 	for (lcv = vm_nphysseg - 1 ; lcv >= 0 ; lcv--)
530 #else
531 	for (lcv = 0 ; lcv < vm_nphysseg ; lcv++)
532 #endif
533 	{
534 
535 		if (uvm.page_init_done == TRUE)
536 			panic("uvm_page_physget: called _after_ bootstrap");
537 
538 		if (vm_physmem[lcv].free_list != freelist)
539 			continue;
540 
541 		/* try from front */
542 		if (vm_physmem[lcv].avail_start == vm_physmem[lcv].start &&
543 		    vm_physmem[lcv].avail_start < vm_physmem[lcv].avail_end) {
544 			*paddrp = ptoa(vm_physmem[lcv].avail_start);
545 			vm_physmem[lcv].avail_start++;
546 			vm_physmem[lcv].start++;
547 			/* nothing left?   nuke it */
548 			if (vm_physmem[lcv].avail_start ==
549 			    vm_physmem[lcv].end) {
550 				if (vm_nphysseg == 1)
551 				    panic("vum_page_physget: out of memory!");
552 				vm_nphysseg--;
553 				for (x = lcv ; x < vm_nphysseg ; x++)
554 					/* structure copy */
555 					vm_physmem[x] = vm_physmem[x+1];
556 			}
557 			return (TRUE);
558 		}
559 
560 		/* try from rear */
561 		if (vm_physmem[lcv].avail_end == vm_physmem[lcv].end &&
562 		    vm_physmem[lcv].avail_start < vm_physmem[lcv].avail_end) {
563 			*paddrp = ptoa(vm_physmem[lcv].avail_end - 1);
564 			vm_physmem[lcv].avail_end--;
565 			vm_physmem[lcv].end--;
566 			/* nothing left?   nuke it */
567 			if (vm_physmem[lcv].avail_end ==
568 			    vm_physmem[lcv].start) {
569 				if (vm_nphysseg == 1)
570 				    panic("uvm_page_physget: out of memory!");
571 				vm_nphysseg--;
572 				for (x = lcv ; x < vm_nphysseg ; x++)
573 					/* structure copy */
574 					vm_physmem[x] = vm_physmem[x+1];
575 			}
576 			return (TRUE);
577 		}
578 	}
579 
580 	/* pass2: forget about matching ends, just allocate something */
581 #if (VM_PHYSSEG_STRAT == VM_PSTRAT_BIGFIRST)
582 	for (lcv = vm_nphysseg - 1 ; lcv >= 0 ; lcv--)
583 #else
584 	for (lcv = 0 ; lcv < vm_nphysseg ; lcv++)
585 #endif
586 	{
587 
588 		/* any room in this bank? */
589 		if (vm_physmem[lcv].avail_start >= vm_physmem[lcv].avail_end)
590 			continue;  /* nope */
591 
592 		*paddrp = ptoa(vm_physmem[lcv].avail_start);
593 		vm_physmem[lcv].avail_start++;
594 		/* truncate! */
595 		vm_physmem[lcv].start = vm_physmem[lcv].avail_start;
596 
597 		/* nothing left?   nuke it */
598 		if (vm_physmem[lcv].avail_start == vm_physmem[lcv].end) {
599 			if (vm_nphysseg == 1)
600 				panic("uvm_page_physget: out of memory!");
601 			vm_nphysseg--;
602 			for (x = lcv ; x < vm_nphysseg ; x++)
603 				/* structure copy */
604 				vm_physmem[x] = vm_physmem[x+1];
605 		}
606 		return (TRUE);
607 	}
608 
609 	return (FALSE);        /* whoops! */
610 }
611 
612 boolean_t
613 uvm_page_physget(paddrp)
614 	paddr_t *paddrp;
615 {
616 	int i;
617 
618 	/* try in the order of freelist preference */
619 	for (i = 0; i < VM_NFREELIST; i++)
620 		if (uvm_page_physget_freelist(paddrp, i) == TRUE)
621 			return (TRUE);
622 	return (FALSE);
623 }
624 #endif /* PMAP_STEAL_MEMORY */
625 
626 /*
627  * uvm_page_physload: load physical memory into VM system
628  *
629  * => all args are PFs
630  * => all pages in start/end get vm_page structures
631  * => areas marked by avail_start/avail_end get added to the free page pool
632  * => we are limited to VM_PHYSSEG_MAX physical memory segments
633  */
634 
635 void
636 uvm_page_physload(start, end, avail_start, avail_end, free_list)
637 	paddr_t start, end, avail_start, avail_end;
638 	int free_list;
639 {
640 	int preload, lcv;
641 	psize_t npages;
642 	struct vm_page *pgs;
643 	struct vm_physseg *ps;
644 
645 	if (uvmexp.pagesize == 0)
646 		panic("uvm_page_physload: page size not set!");
647 	if (free_list >= VM_NFREELIST || free_list < VM_FREELIST_DEFAULT)
648 		panic("uvm_page_physload: bad free list %d\n", free_list);
649 	if (start >= end)
650 		panic("uvm_page_physload: start >= end");
651 
652 	/*
653 	 * do we have room?
654 	 */
655 
656 	if (vm_nphysseg == VM_PHYSSEG_MAX) {
657 		printf("uvm_page_physload: unable to load physical memory "
658 		    "segment\n");
659 		printf("\t%d segments allocated, ignoring 0x%llx -> 0x%llx\n",
660 		    VM_PHYSSEG_MAX, (long long)start, (long long)end);
661 		printf("\tincrease VM_PHYSSEG_MAX\n");
662 		return;
663 	}
664 
665 	/*
666 	 * check to see if this is a "preload" (i.e. uvm_mem_init hasn't been
667 	 * called yet, so malloc is not available).
668 	 */
669 
670 	for (lcv = 0 ; lcv < vm_nphysseg ; lcv++) {
671 		if (vm_physmem[lcv].pgs)
672 			break;
673 	}
674 	preload = (lcv == vm_nphysseg);
675 
676 	/*
677 	 * if VM is already running, attempt to malloc() vm_page structures
678 	 */
679 
680 	if (!preload) {
681 #if defined(VM_PHYSSEG_NOADD)
682 		panic("uvm_page_physload: tried to add RAM after vm_mem_init");
683 #else
684 		/* XXXCDC: need some sort of lockout for this case */
685 		paddr_t paddr;
686 		npages = end - start;  /* # of pages */
687 		pgs = malloc(sizeof(struct vm_page) * npages,
688 		    M_VMPAGE, M_NOWAIT);
689 		if (pgs == NULL) {
690 			printf("uvm_page_physload: can not malloc vm_page "
691 			    "structs for segment\n");
692 			printf("\tignoring 0x%lx -> 0x%lx\n", start, end);
693 			return;
694 		}
695 		/* zero data, init phys_addr and free_list, and free pages */
696 		memset(pgs, 0, sizeof(struct vm_page) * npages);
697 		for (lcv = 0, paddr = ptoa(start) ;
698 				 lcv < npages ; lcv++, paddr += PAGE_SIZE) {
699 			pgs[lcv].phys_addr = paddr;
700 			pgs[lcv].free_list = free_list;
701 			if (atop(paddr) >= avail_start &&
702 			    atop(paddr) <= avail_end)
703 				uvm_pagefree(&pgs[lcv]);
704 		}
705 		/* XXXCDC: incomplete: need to update uvmexp.free, what else? */
706 		/* XXXCDC: need hook to tell pmap to rebuild pv_list, etc... */
707 #endif
708 	} else {
709 		pgs = NULL;
710 		npages = 0;
711 	}
712 
713 	/*
714 	 * now insert us in the proper place in vm_physmem[]
715 	 */
716 
717 #if (VM_PHYSSEG_STRAT == VM_PSTRAT_RANDOM)
718 	/* random: put it at the end (easy!) */
719 	ps = &vm_physmem[vm_nphysseg];
720 #elif (VM_PHYSSEG_STRAT == VM_PSTRAT_BSEARCH)
721 	{
722 		int x;
723 		/* sort by address for binary search */
724 		for (lcv = 0 ; lcv < vm_nphysseg ; lcv++)
725 			if (start < vm_physmem[lcv].start)
726 				break;
727 		ps = &vm_physmem[lcv];
728 		/* move back other entries, if necessary ... */
729 		for (x = vm_nphysseg ; x > lcv ; x--)
730 			/* structure copy */
731 			vm_physmem[x] = vm_physmem[x - 1];
732 	}
733 #elif (VM_PHYSSEG_STRAT == VM_PSTRAT_BIGFIRST)
734 	{
735 		int x;
736 		/* sort by largest segment first */
737 		for (lcv = 0 ; lcv < vm_nphysseg ; lcv++)
738 			if ((end - start) >
739 			    (vm_physmem[lcv].end - vm_physmem[lcv].start))
740 				break;
741 		ps = &vm_physmem[lcv];
742 		/* move back other entries, if necessary ... */
743 		for (x = vm_nphysseg ; x > lcv ; x--)
744 			/* structure copy */
745 			vm_physmem[x] = vm_physmem[x - 1];
746 	}
747 #else
748 	panic("uvm_page_physload: unknown physseg strategy selected!");
749 #endif
750 
751 	ps->start = start;
752 	ps->end = end;
753 	ps->avail_start = avail_start;
754 	ps->avail_end = avail_end;
755 	if (preload) {
756 		ps->pgs = NULL;
757 	} else {
758 		ps->pgs = pgs;
759 		ps->lastpg = pgs + npages - 1;
760 	}
761 	ps->free_list = free_list;
762 	vm_nphysseg++;
763 
764 	if (!preload)
765 		uvm_page_rehash();
766 }
767 
768 /*
769  * uvm_page_rehash: reallocate hash table based on number of free pages.
770  */
771 
772 void
773 uvm_page_rehash()
774 {
775 	int freepages, lcv, bucketcount, oldcount;
776 	struct pglist *newbuckets, *oldbuckets;
777 	struct vm_page *pg;
778 	size_t newsize, oldsize;
779 
780 	/*
781 	 * compute number of pages that can go in the free pool
782 	 */
783 
784 	freepages = 0;
785 	for (lcv = 0 ; lcv < vm_nphysseg ; lcv++)
786 		freepages +=
787 		    (vm_physmem[lcv].avail_end - vm_physmem[lcv].avail_start);
788 
789 	/*
790 	 * compute number of buckets needed for this number of pages
791 	 */
792 
793 	bucketcount = 1;
794 	while (bucketcount < freepages)
795 		bucketcount = bucketcount * 2;
796 
797 	/*
798 	 * compute the size of the current table and new table.
799 	 */
800 
801 	oldbuckets = uvm.page_hash;
802 	oldcount = uvm.page_nhash;
803 	oldsize = round_page(sizeof(struct pglist) * oldcount);
804 	newsize = round_page(sizeof(struct pglist) * bucketcount);
805 
806 	/*
807 	 * allocate the new buckets
808 	 */
809 
810 	newbuckets = (struct pglist *) uvm_km_alloc(kernel_map, newsize);
811 	if (newbuckets == NULL) {
812 		printf("uvm_page_physrehash: WARNING: could not grow page "
813 		    "hash table\n");
814 		return;
815 	}
816 	for (lcv = 0 ; lcv < bucketcount ; lcv++)
817 		TAILQ_INIT(&newbuckets[lcv]);
818 
819 	/*
820 	 * now replace the old buckets with the new ones and rehash everything
821 	 */
822 
823 	simple_lock(&uvm.hashlock);
824 	uvm.page_hash = newbuckets;
825 	uvm.page_nhash = bucketcount;
826 	uvm.page_hashmask = bucketcount - 1;  /* power of 2 */
827 
828 	/* ... and rehash */
829 	for (lcv = 0 ; lcv < oldcount ; lcv++) {
830 		while ((pg = oldbuckets[lcv].tqh_first) != NULL) {
831 			TAILQ_REMOVE(&oldbuckets[lcv], pg, hashq);
832 			TAILQ_INSERT_TAIL(
833 			  &uvm.page_hash[uvm_pagehash(pg->uobject, pg->offset)],
834 			  pg, hashq);
835 		}
836 	}
837 	simple_unlock(&uvm.hashlock);
838 
839 	/*
840 	 * free old bucket array if is not the boot-time table
841 	 */
842 
843 	if (oldbuckets != &uvm_bootbucket)
844 		uvm_km_free(kernel_map, (vaddr_t) oldbuckets, oldsize);
845 }
846 
847 /*
848  * uvm_page_recolor: Recolor the pages if the new bucket count is
849  * larger than the old one.
850  */
851 
852 void
853 uvm_page_recolor(int newncolors)
854 {
855 	struct pgflbucket *bucketarray, *oldbucketarray;
856 	struct pgfreelist pgfl;
857 	struct vm_page *pg;
858 	vsize_t bucketcount;
859 	int s, lcv, color, i, ocolors;
860 
861 	if (newncolors <= uvmexp.ncolors)
862 		return;
863 
864 	bucketcount = newncolors * VM_NFREELIST;
865 	bucketarray = malloc(bucketcount * sizeof(struct pgflbucket),
866 	    M_VMPAGE, M_NOWAIT);
867 	if (bucketarray == NULL) {
868 		printf("WARNING: unable to allocate %ld page color buckets\n",
869 		    (long) bucketcount);
870 		return;
871 	}
872 
873 	s = uvm_lock_fpageq();
874 
875 	/* Make sure we should still do this. */
876 	if (newncolors <= uvmexp.ncolors) {
877 		uvm_unlock_fpageq(s);
878 		free(bucketarray, M_VMPAGE);
879 		return;
880 	}
881 
882 	oldbucketarray = uvm.page_free[0].pgfl_buckets;
883 	ocolors = uvmexp.ncolors;
884 
885 	uvmexp.ncolors = newncolors;
886 	uvmexp.colormask = uvmexp.ncolors - 1;
887 
888 	for (lcv = 0; lcv < VM_NFREELIST; lcv++) {
889 		pgfl.pgfl_buckets = (bucketarray + (lcv * newncolors));
890 		uvm_page_init_buckets(&pgfl);
891 		for (color = 0; color < ocolors; color++) {
892 			for (i = 0; i < PGFL_NQUEUES; i++) {
893 				while ((pg = TAILQ_FIRST(&uvm.page_free[
894 				    lcv].pgfl_buckets[color].pgfl_queues[i]))
895 				    != NULL) {
896 					TAILQ_REMOVE(&uvm.page_free[
897 					    lcv].pgfl_buckets[
898 					    color].pgfl_queues[i], pg, pageq);
899 					TAILQ_INSERT_TAIL(&pgfl.pgfl_buckets[
900 					    VM_PGCOLOR_BUCKET(pg)].pgfl_queues[
901 					    i], pg, pageq);
902 				}
903 			}
904 		}
905 		uvm.page_free[lcv].pgfl_buckets = pgfl.pgfl_buckets;
906 	}
907 
908 	if (have_recolored_pages) {
909 		uvm_unlock_fpageq(s);
910 		free(oldbucketarray, M_VMPAGE);
911 		return;
912 	}
913 
914 	have_recolored_pages = TRUE;
915 	uvm_unlock_fpageq(s);
916 }
917 
918 /*
919  * uvm_pagealloc_pgfl: helper routine for uvm_pagealloc_strat
920  */
921 
922 static __inline struct vm_page *
923 uvm_pagealloc_pgfl(struct pgfreelist *pgfl, int try1, int try2,
924     int *trycolorp)
925 {
926 	struct pglist *freeq;
927 	struct vm_page *pg;
928 	int color, trycolor = *trycolorp;
929 
930 	color = trycolor;
931 	do {
932 		if ((pg = TAILQ_FIRST((freeq =
933 		    &pgfl->pgfl_buckets[color].pgfl_queues[try1]))) != NULL)
934 			goto gotit;
935 		if ((pg = TAILQ_FIRST((freeq =
936 		    &pgfl->pgfl_buckets[color].pgfl_queues[try2]))) != NULL)
937 			goto gotit;
938 		color = (color + 1) & uvmexp.colormask;
939 	} while (color != trycolor);
940 
941 	return (NULL);
942 
943  gotit:
944 	TAILQ_REMOVE(freeq, pg, pageq);
945 	uvmexp.free--;
946 
947 	/* update zero'd page count */
948 	if (pg->flags & PG_ZERO)
949 		uvmexp.zeropages--;
950 
951 	if (color == trycolor)
952 		uvmexp.colorhit++;
953 	else {
954 		uvmexp.colormiss++;
955 		*trycolorp = color;
956 	}
957 
958 	return (pg);
959 }
960 
961 /*
962  * uvm_pagealloc_strat: allocate vm_page from a particular free list.
963  *
964  * => return null if no pages free
965  * => wake up pagedaemon if number of free pages drops below low water mark
966  * => if obj != NULL, obj must be locked (to put in hash)
967  * => if anon != NULL, anon must be locked (to put in anon)
968  * => only one of obj or anon can be non-null
969  * => caller must activate/deactivate page if it is not wired.
970  * => free_list is ignored if strat == UVM_PGA_STRAT_NORMAL.
971  * => policy decision: it is more important to pull a page off of the
972  *	appropriate priority free list than it is to get a zero'd or
973  *	unknown contents page.  This is because we live with the
974  *	consequences of a bad free list decision for the entire
975  *	lifetime of the page, e.g. if the page comes from memory that
976  *	is slower to access.
977  */
978 
979 struct vm_page *
980 uvm_pagealloc_strat(obj, off, anon, flags, strat, free_list)
981 	struct uvm_object *obj;
982 	voff_t off;
983 	int flags;
984 	struct vm_anon *anon;
985 	int strat, free_list;
986 {
987 	int lcv, try1, try2, s, zeroit = 0, color;
988 	struct vm_page *pg;
989 	boolean_t use_reserve;
990 
991 	KASSERT(obj == NULL || anon == NULL);
992 	KASSERT(off == trunc_page(off));
993 	LOCK_ASSERT(obj == NULL || simple_lock_held(&obj->vmobjlock));
994 	LOCK_ASSERT(anon == NULL || simple_lock_held(&anon->an_lock));
995 
996 	s = uvm_lock_fpageq();
997 
998 	/*
999 	 * This implements a global round-robin page coloring
1000 	 * algorithm.
1001 	 *
1002 	 * XXXJRT: Should we make the `nextcolor' per-cpu?
1003 	 * XXXJRT: What about virtually-indexed caches?
1004 	 */
1005 
1006 	color = uvm.page_free_nextcolor;
1007 
1008 	/*
1009 	 * check to see if we need to generate some free pages waking
1010 	 * the pagedaemon.
1011 	 */
1012 
1013 	UVM_KICK_PDAEMON();
1014 
1015 	/*
1016 	 * fail if any of these conditions is true:
1017 	 * [1]  there really are no free pages, or
1018 	 * [2]  only kernel "reserved" pages remain and
1019 	 *        the page isn't being allocated to a kernel object.
1020 	 * [3]  only pagedaemon "reserved" pages remain and
1021 	 *        the requestor isn't the pagedaemon.
1022 	 */
1023 
1024 	use_reserve = (flags & UVM_PGA_USERESERVE) ||
1025 		(obj && UVM_OBJ_IS_KERN_OBJECT(obj));
1026 	if ((uvmexp.free <= uvmexp.reserve_kernel && !use_reserve) ||
1027 	    (uvmexp.free <= uvmexp.reserve_pagedaemon &&
1028 	     !(use_reserve && curproc == uvm.pagedaemon_proc)))
1029 		goto fail;
1030 
1031 #if PGFL_NQUEUES != 2
1032 #error uvm_pagealloc_strat needs to be updated
1033 #endif
1034 
1035 	/*
1036 	 * If we want a zero'd page, try the ZEROS queue first, otherwise
1037 	 * we try the UNKNOWN queue first.
1038 	 */
1039 	if (flags & UVM_PGA_ZERO) {
1040 		try1 = PGFL_ZEROS;
1041 		try2 = PGFL_UNKNOWN;
1042 	} else {
1043 		try1 = PGFL_UNKNOWN;
1044 		try2 = PGFL_ZEROS;
1045 	}
1046 
1047  again:
1048 	switch (strat) {
1049 	case UVM_PGA_STRAT_NORMAL:
1050 		/* Check all freelists in descending priority order. */
1051 		for (lcv = 0; lcv < VM_NFREELIST; lcv++) {
1052 			pg = uvm_pagealloc_pgfl(&uvm.page_free[lcv],
1053 			    try1, try2, &color);
1054 			if (pg != NULL)
1055 				goto gotit;
1056 		}
1057 
1058 		/* No pages free! */
1059 		goto fail;
1060 
1061 	case UVM_PGA_STRAT_ONLY:
1062 	case UVM_PGA_STRAT_FALLBACK:
1063 		/* Attempt to allocate from the specified free list. */
1064 		KASSERT(free_list >= 0 && free_list < VM_NFREELIST);
1065 		pg = uvm_pagealloc_pgfl(&uvm.page_free[free_list],
1066 		    try1, try2, &color);
1067 		if (pg != NULL)
1068 			goto gotit;
1069 
1070 		/* Fall back, if possible. */
1071 		if (strat == UVM_PGA_STRAT_FALLBACK) {
1072 			strat = UVM_PGA_STRAT_NORMAL;
1073 			goto again;
1074 		}
1075 
1076 		/* No pages free! */
1077 		goto fail;
1078 
1079 	default:
1080 		panic("uvm_pagealloc_strat: bad strat %d", strat);
1081 		/* NOTREACHED */
1082 	}
1083 
1084  gotit:
1085 	/*
1086 	 * We now know which color we actually allocated from; set
1087 	 * the next color accordingly.
1088 	 */
1089 
1090 	uvm.page_free_nextcolor = (color + 1) & uvmexp.colormask;
1091 
1092 	/*
1093 	 * update allocation statistics and remember if we have to
1094 	 * zero the page
1095 	 */
1096 
1097 	if (flags & UVM_PGA_ZERO) {
1098 		if (pg->flags & PG_ZERO) {
1099 			uvmexp.pga_zerohit++;
1100 			zeroit = 0;
1101 		} else {
1102 			uvmexp.pga_zeromiss++;
1103 			zeroit = 1;
1104 		}
1105 	}
1106 	uvm_unlock_fpageq(s);
1107 
1108 	pg->offset = off;
1109 	pg->uobject = obj;
1110 	pg->uanon = anon;
1111 	pg->flags = PG_BUSY|PG_CLEAN|PG_FAKE;
1112 	if (anon) {
1113 		anon->u.an_page = pg;
1114 		pg->pqflags = PQ_ANON;
1115 		uvmexp.anonpages++;
1116 	} else {
1117 		if (obj) {
1118 			uvm_pageinsert(pg);
1119 		}
1120 		pg->pqflags = 0;
1121 	}
1122 #if defined(UVM_PAGE_TRKOWN)
1123 	pg->owner_tag = NULL;
1124 #endif
1125 	UVM_PAGE_OWN(pg, "new alloc");
1126 
1127 	if (flags & UVM_PGA_ZERO) {
1128 		/*
1129 		 * A zero'd page is not clean.  If we got a page not already
1130 		 * zero'd, then we have to zero it ourselves.
1131 		 */
1132 		pg->flags &= ~PG_CLEAN;
1133 		if (zeroit)
1134 			pmap_zero_page(VM_PAGE_TO_PHYS(pg));
1135 	}
1136 
1137 	return(pg);
1138 
1139  fail:
1140 	uvm_unlock_fpageq(s);
1141 	return (NULL);
1142 }
1143 
1144 /*
1145  * uvm_pagerealloc: reallocate a page from one object to another
1146  *
1147  * => both objects must be locked
1148  */
1149 
1150 void
1151 uvm_pagerealloc(pg, newobj, newoff)
1152 	struct vm_page *pg;
1153 	struct uvm_object *newobj;
1154 	voff_t newoff;
1155 {
1156 	/*
1157 	 * remove it from the old object
1158 	 */
1159 
1160 	if (pg->uobject) {
1161 		uvm_pageremove(pg);
1162 	}
1163 
1164 	/*
1165 	 * put it in the new object
1166 	 */
1167 
1168 	if (newobj) {
1169 		pg->uobject = newobj;
1170 		pg->offset = newoff;
1171 		uvm_pageinsert(pg);
1172 	}
1173 }
1174 
1175 /*
1176  * uvm_pagefree: free page
1177  *
1178  * => erase page's identity (i.e. remove from hash/object)
1179  * => put page on free list
1180  * => caller must lock owning object (either anon or uvm_object)
1181  * => caller must lock page queues
1182  * => assumes all valid mappings of pg are gone
1183  */
1184 
1185 void
1186 uvm_pagefree(pg)
1187 	struct vm_page *pg;
1188 {
1189 	int s;
1190 
1191 	KASSERT((pg->flags & PG_PAGEOUT) == 0);
1192 	LOCK_ASSERT(simple_lock_held(&uvm.pageqlock) ||
1193 		    (pg->pqflags & (PQ_ACTIVE|PQ_INACTIVE)) == 0);
1194 	LOCK_ASSERT(pg->uobject == NULL ||
1195 		    simple_lock_held(&pg->uobject->vmobjlock));
1196 	LOCK_ASSERT(pg->uobject != NULL || pg->uanon == NULL ||
1197 		    simple_lock_held(&pg->uanon->an_lock));
1198 
1199 #ifdef DEBUG
1200 	if (pg->uobject == (void *)0xdeadbeef &&
1201 	    pg->uanon == (void *)0xdeadbeef) {
1202 		panic("uvm_pagefree: freeing free page %p\n", pg);
1203 	}
1204 #endif
1205 
1206 	/*
1207 	 * if the page is loaned, resolve the loan instead of freeing.
1208 	 */
1209 
1210 	if (pg->loan_count) {
1211 		KASSERT(pg->wire_count == 0);
1212 
1213 		/*
1214 		 * if the page is owned by an anon then we just want to
1215 		 * drop anon ownership.  the kernel will free the page when
1216 		 * it is done with it.  if the page is owned by an object,
1217 		 * remove it from the object and mark it dirty for the benefit
1218 		 * of possible anon owners.
1219 		 *
1220 		 * regardless of previous ownership, wakeup any waiters,
1221 		 * unbusy the page, and we're done.
1222 		 */
1223 
1224 		if (pg->pqflags & PQ_ANON) {
1225 			pg->pqflags &= ~PQ_ANON;
1226 			pg->uanon = NULL;
1227 		} else if (pg->flags & PG_TABLED) {
1228 			uvm_pageremove(pg);
1229 			pg->flags &= ~PG_CLEAN;
1230 		}
1231 		if (pg->flags & PG_WANTED) {
1232 			wakeup(pg);
1233 		}
1234 		pg->flags &= ~(PG_WANTED|PG_BUSY);
1235 #ifdef UVM_PAGE_TRKOWN
1236 		pg->owner_tag = NULL;
1237 #endif
1238 		return;
1239 	}
1240 
1241 	/*
1242 	 * remove page from its object or anon.
1243 	 * adjust swpgonly if the page is swap-backed.
1244 	 */
1245 
1246 	if (pg->flags & PG_TABLED) {
1247 		uvm_pageremove(pg);
1248 	} else if (pg->pqflags & PQ_ANON) {
1249 		pg->uanon->u.an_page = NULL;
1250 	}
1251 
1252 	/*
1253 	 * now remove the page from the queues.
1254 	 */
1255 
1256 	uvm_pagedequeue(pg);
1257 
1258 	/*
1259 	 * if the page was wired, unwire it now.
1260 	 */
1261 
1262 	if (pg->wire_count) {
1263 		pg->wire_count = 0;
1264 		uvmexp.wired--;
1265 	}
1266 	if (pg->pqflags & PQ_ANON) {
1267 		uvmexp.anonpages--;
1268 	}
1269 
1270 	/*
1271 	 * and put on free queue
1272 	 */
1273 
1274 	pg->flags &= ~PG_ZERO;
1275 
1276 	s = uvm_lock_fpageq();
1277 	TAILQ_INSERT_TAIL(&uvm.page_free[
1278 	    uvm_page_lookup_freelist(pg)].pgfl_buckets[
1279 	    VM_PGCOLOR_BUCKET(pg)].pgfl_queues[PGFL_UNKNOWN], pg, pageq);
1280 	pg->pqflags = PQ_FREE;
1281 #ifdef DEBUG
1282 	pg->uobject = (void *)0xdeadbeef;
1283 	pg->offset = 0xdeadbeef;
1284 	pg->uanon = (void *)0xdeadbeef;
1285 #endif
1286 	uvmexp.free++;
1287 
1288 	if (uvmexp.zeropages < UVM_PAGEZERO_TARGET)
1289 		uvm.page_idle_zero = vm_page_zero_enable;
1290 
1291 	uvm_unlock_fpageq(s);
1292 }
1293 
1294 /*
1295  * uvm_page_unbusy: unbusy an array of pages.
1296  *
1297  * => pages must either all belong to the same object, or all belong to anons.
1298  * => if pages are object-owned, object must be locked.
1299  * => if pages are anon-owned, anons must be locked.
1300  */
1301 
1302 void
1303 uvm_page_unbusy(pgs, npgs)
1304 	struct vm_page **pgs;
1305 	int npgs;
1306 {
1307 	struct vm_page *pg;
1308 	int i;
1309 	UVMHIST_FUNC("uvm_page_unbusy"); UVMHIST_CALLED(ubchist);
1310 
1311 	for (i = 0; i < npgs; i++) {
1312 		pg = pgs[i];
1313 		if (pg == NULL) {
1314 			continue;
1315 		}
1316 		if (pg->flags & PG_WANTED) {
1317 			wakeup(pg);
1318 		}
1319 		if (pg->flags & PG_RELEASED) {
1320 			UVMHIST_LOG(ubchist, "releasing pg %p", pg,0,0,0);
1321 			pg->flags &= ~PG_RELEASED;
1322 			uvm_pagefree(pg);
1323 		} else {
1324 			UVMHIST_LOG(ubchist, "unbusying pg %p", pg,0,0,0);
1325 			pg->flags &= ~(PG_WANTED|PG_BUSY);
1326 			UVM_PAGE_OWN(pg, NULL);
1327 		}
1328 	}
1329 }
1330 
1331 #if defined(UVM_PAGE_TRKOWN)
1332 /*
1333  * uvm_page_own: set or release page ownership
1334  *
1335  * => this is a debugging function that keeps track of who sets PG_BUSY
1336  *	and where they do it.   it can be used to track down problems
1337  *	such a process setting "PG_BUSY" and never releasing it.
1338  * => page's object [if any] must be locked
1339  * => if "tag" is NULL then we are releasing page ownership
1340  */
1341 void
1342 uvm_page_own(pg, tag)
1343 	struct vm_page *pg;
1344 	char *tag;
1345 {
1346 	KASSERT((pg->flags & (PG_PAGEOUT|PG_RELEASED)) == 0);
1347 
1348 	/* gain ownership? */
1349 	if (tag) {
1350 		if (pg->owner_tag) {
1351 			printf("uvm_page_own: page %p already owned "
1352 			    "by proc %d [%s]\n", pg,
1353 			     pg->owner, pg->owner_tag);
1354 			panic("uvm_page_own");
1355 		}
1356 		pg->owner = (curproc) ? curproc->p_pid :  (pid_t) -1;
1357 		pg->owner_tag = tag;
1358 		return;
1359 	}
1360 
1361 	/* drop ownership */
1362 	if (pg->owner_tag == NULL) {
1363 		printf("uvm_page_own: dropping ownership of an non-owned "
1364 		    "page (%p)\n", pg);
1365 		panic("uvm_page_own");
1366 	}
1367 	pg->owner_tag = NULL;
1368 	KASSERT((pg->pqflags & (PQ_ACTIVE|PQ_INACTIVE)) ||
1369 		(pg->uanon == NULL && pg->uobject == NULL) ||
1370 		pg->uobject == uvm.kernel_object ||
1371 		pg->wire_count > 0 ||
1372 		(pg->loan_count == 1 && pg->uanon == NULL) ||
1373 		pg->loan_count > 1);
1374 	return;
1375 }
1376 #endif
1377 
1378 /*
1379  * uvm_pageidlezero: zero free pages while the system is idle.
1380  *
1381  * => try to complete one color bucket at a time, to reduce our impact
1382  *	on the CPU cache.
1383  * => we loop until we either reach the target or whichqs indicates that
1384  *	there is a process ready to run.
1385  */
1386 void
1387 uvm_pageidlezero()
1388 {
1389 	struct vm_page *pg;
1390 	struct pgfreelist *pgfl;
1391 	int free_list, s, firstbucket;
1392 	static int nextbucket;
1393 
1394 	s = uvm_lock_fpageq();
1395 	firstbucket = nextbucket;
1396 	do {
1397 		if (sched_whichqs != 0) {
1398 			uvm_unlock_fpageq(s);
1399 			return;
1400 		}
1401 		if (uvmexp.zeropages >= UVM_PAGEZERO_TARGET) {
1402 			uvm.page_idle_zero = FALSE;
1403 			uvm_unlock_fpageq(s);
1404 			return;
1405 		}
1406 		for (free_list = 0; free_list < VM_NFREELIST; free_list++) {
1407 			pgfl = &uvm.page_free[free_list];
1408 			while ((pg = TAILQ_FIRST(&pgfl->pgfl_buckets[
1409 			    nextbucket].pgfl_queues[PGFL_UNKNOWN])) != NULL) {
1410 				if (sched_whichqs != 0) {
1411 					uvm_unlock_fpageq(s);
1412 					return;
1413 				}
1414 
1415 				TAILQ_REMOVE(&pgfl->pgfl_buckets[
1416 				    nextbucket].pgfl_queues[PGFL_UNKNOWN],
1417 				    pg, pageq);
1418 				uvmexp.free--;
1419 				uvm_unlock_fpageq(s);
1420 #ifdef PMAP_PAGEIDLEZERO
1421 				if (!PMAP_PAGEIDLEZERO(VM_PAGE_TO_PHYS(pg))) {
1422 
1423 					/*
1424 					 * The machine-dependent code detected
1425 					 * some reason for us to abort zeroing
1426 					 * pages, probably because there is a
1427 					 * process now ready to run.
1428 					 */
1429 
1430 					s = uvm_lock_fpageq();
1431 					TAILQ_INSERT_HEAD(&pgfl->pgfl_buckets[
1432 					    nextbucket].pgfl_queues[
1433 					    PGFL_UNKNOWN], pg, pageq);
1434 					uvmexp.free++;
1435 					uvmexp.zeroaborts++;
1436 					uvm_unlock_fpageq(s);
1437 					return;
1438 				}
1439 #else
1440 				pmap_zero_page(VM_PAGE_TO_PHYS(pg));
1441 #endif /* PMAP_PAGEIDLEZERO */
1442 				pg->flags |= PG_ZERO;
1443 
1444 				s = uvm_lock_fpageq();
1445 				TAILQ_INSERT_HEAD(&pgfl->pgfl_buckets[
1446 				    nextbucket].pgfl_queues[PGFL_ZEROS],
1447 				    pg, pageq);
1448 				uvmexp.free++;
1449 				uvmexp.zeropages++;
1450 			}
1451 		}
1452 		nextbucket = (nextbucket + 1) & uvmexp.colormask;
1453 	} while (nextbucket != firstbucket);
1454 	uvm_unlock_fpageq(s);
1455 }
1456