xref: /netbsd-src/sys/uvm/uvm_page.c (revision a30f264f2a5f410ffefcc55600a9238b3a0c935c)
1 /*	$NetBSD: uvm_page.c,v 1.109 2005/12/24 20:45:10 perry 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 1.109 2005/12/24 20:45:10 perry 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_C              /* pull in uvm_page_i.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 MALLOC_DEFINE(M_VMPAGE, "VM page", "VM page");
143 
144 #ifdef DEBUG
145 vaddr_t uvm_zerocheckkva;
146 #endif /* DEBUG */
147 
148 /*
149  * local prototypes
150  */
151 
152 static void uvm_pageinsert(struct vm_page *);
153 static void uvm_pageinsert_after(struct vm_page *, struct vm_page *);
154 static void uvm_pageremove(struct vm_page *);
155 
156 /*
157  * inline functions
158  */
159 
160 /*
161  * uvm_pageinsert: insert a page in the object and the hash table
162  * uvm_pageinsert_after: insert a page into the specified place in listq
163  *
164  * => caller must lock object
165  * => caller must lock page queues
166  * => call should have already set pg's object and offset pointers
167  *    and bumped the version counter
168  */
169 
170 inline static void
171 uvm_pageinsert_after(struct vm_page *pg, struct vm_page *where)
172 {
173 	struct pglist *buck;
174 	struct uvm_object *uobj = pg->uobject;
175 
176 	KASSERT((pg->flags & PG_TABLED) == 0);
177 	KASSERT(where == NULL || (where->flags & PG_TABLED));
178 	KASSERT(where == NULL || (where->uobject == uobj));
179 	buck = &uvm.page_hash[uvm_pagehash(uobj, pg->offset)];
180 	simple_lock(&uvm.hashlock);
181 	TAILQ_INSERT_TAIL(buck, pg, hashq);
182 	simple_unlock(&uvm.hashlock);
183 
184 	if (UVM_OBJ_IS_VNODE(uobj)) {
185 		if (uobj->uo_npages == 0) {
186 			struct vnode *vp = (struct vnode *)uobj;
187 
188 			vholdl(vp);
189 		}
190 		if (UVM_OBJ_IS_VTEXT(uobj)) {
191 			uvmexp.execpages++;
192 		} else {
193 			uvmexp.filepages++;
194 		}
195 	} else if (UVM_OBJ_IS_AOBJ(uobj)) {
196 		uvmexp.anonpages++;
197 	}
198 
199 	if (where)
200 		TAILQ_INSERT_AFTER(&uobj->memq, where, pg, listq);
201 	else
202 		TAILQ_INSERT_TAIL(&uobj->memq, pg, listq);
203 	pg->flags |= PG_TABLED;
204 	uobj->uo_npages++;
205 }
206 
207 inline static void
208 uvm_pageinsert(struct vm_page *pg)
209 {
210 
211 	uvm_pageinsert_after(pg, NULL);
212 }
213 
214 /*
215  * uvm_page_remove: remove page from object and hash
216  *
217  * => caller must lock object
218  * => caller must lock page queues
219  */
220 
221 static inline void
222 uvm_pageremove(struct vm_page *pg)
223 {
224 	struct pglist *buck;
225 	struct uvm_object *uobj = pg->uobject;
226 
227 	KASSERT(pg->flags & PG_TABLED);
228 	buck = &uvm.page_hash[uvm_pagehash(uobj, pg->offset)];
229 	simple_lock(&uvm.hashlock);
230 	TAILQ_REMOVE(buck, pg, hashq);
231 	simple_unlock(&uvm.hashlock);
232 
233 	if (UVM_OBJ_IS_VNODE(uobj)) {
234 		if (uobj->uo_npages == 1) {
235 			struct vnode *vp = (struct vnode *)uobj;
236 
237 			holdrelel(vp);
238 		}
239 		if (UVM_OBJ_IS_VTEXT(uobj)) {
240 			uvmexp.execpages--;
241 		} else {
242 			uvmexp.filepages--;
243 		}
244 	} else if (UVM_OBJ_IS_AOBJ(uobj)) {
245 		uvmexp.anonpages--;
246 	}
247 
248 	/* object should be locked */
249 	uobj->uo_npages--;
250 	TAILQ_REMOVE(&uobj->memq, pg, listq);
251 	pg->flags &= ~PG_TABLED;
252 	pg->uobject = NULL;
253 }
254 
255 static void
256 uvm_page_init_buckets(struct pgfreelist *pgfl)
257 {
258 	int color, i;
259 
260 	for (color = 0; color < uvmexp.ncolors; color++) {
261 		for (i = 0; i < PGFL_NQUEUES; i++) {
262 			TAILQ_INIT(&pgfl->pgfl_buckets[color].pgfl_queues[i]);
263 		}
264 	}
265 }
266 
267 /*
268  * uvm_page_init: init the page system.   called from uvm_init().
269  *
270  * => we return the range of kernel virtual memory in kvm_startp/kvm_endp
271  */
272 
273 void
274 uvm_page_init(vaddr_t *kvm_startp, vaddr_t *kvm_endp)
275 {
276 	vsize_t freepages, pagecount, bucketcount, n;
277 	struct pgflbucket *bucketarray;
278 	struct vm_page *pagearray;
279 	int lcv;
280 	u_int i;
281 	paddr_t paddr;
282 
283 	/*
284 	 * init the page queues and page queue locks, except the free
285 	 * list; we allocate that later (with the initial vm_page
286 	 * structures).
287 	 */
288 
289 	TAILQ_INIT(&uvm.page_active);
290 	TAILQ_INIT(&uvm.page_inactive);
291 	simple_lock_init(&uvm.pageqlock);
292 	simple_lock_init(&uvm.fpageqlock);
293 
294 	/*
295 	 * init the <obj,offset> => <page> hash table.  for now
296 	 * we just have one bucket (the bootstrap bucket).  later on we
297 	 * will allocate new buckets as we dynamically resize the hash table.
298 	 */
299 
300 	uvm.page_nhash = 1;			/* 1 bucket */
301 	uvm.page_hashmask = 0;			/* mask for hash function */
302 	uvm.page_hash = &uvm_bootbucket;	/* install bootstrap bucket */
303 	TAILQ_INIT(uvm.page_hash);		/* init hash table */
304 	simple_lock_init(&uvm.hashlock);	/* init hash table lock */
305 
306 	/*
307 	 * allocate vm_page structures.
308 	 */
309 
310 	/*
311 	 * sanity check:
312 	 * before calling this function the MD code is expected to register
313 	 * some free RAM with the uvm_page_physload() function.   our job
314 	 * now is to allocate vm_page structures for this memory.
315 	 */
316 
317 	if (vm_nphysseg == 0)
318 		panic("uvm_page_bootstrap: no memory pre-allocated");
319 
320 	/*
321 	 * first calculate the number of free pages...
322 	 *
323 	 * note that we use start/end rather than avail_start/avail_end.
324 	 * this allows us to allocate extra vm_page structures in case we
325 	 * want to return some memory to the pool after booting.
326 	 */
327 
328 	freepages = 0;
329 	for (lcv = 0 ; lcv < vm_nphysseg ; lcv++)
330 		freepages += (vm_physmem[lcv].end - vm_physmem[lcv].start);
331 
332 	/*
333 	 * Let MD code initialize the number of colors, or default
334 	 * to 1 color if MD code doesn't care.
335 	 */
336 	if (uvmexp.ncolors == 0)
337 		uvmexp.ncolors = 1;
338 	uvmexp.colormask = uvmexp.ncolors - 1;
339 
340 	/*
341 	 * we now know we have (PAGE_SIZE * freepages) bytes of memory we can
342 	 * use.   for each page of memory we use we need a vm_page structure.
343 	 * thus, the total number of pages we can use is the total size of
344 	 * the memory divided by the PAGE_SIZE plus the size of the vm_page
345 	 * structure.   we add one to freepages as a fudge factor to avoid
346 	 * truncation errors (since we can only allocate in terms of whole
347 	 * pages).
348 	 */
349 
350 	bucketcount = uvmexp.ncolors * VM_NFREELIST;
351 	pagecount = ((freepages + 1) << PAGE_SHIFT) /
352 	    (PAGE_SIZE + sizeof(struct vm_page));
353 
354 	bucketarray = (void *)uvm_pageboot_alloc((bucketcount *
355 	    sizeof(struct pgflbucket)) + (pagecount *
356 	    sizeof(struct vm_page)));
357 	pagearray = (struct vm_page *)(bucketarray + bucketcount);
358 
359 	for (lcv = 0; lcv < VM_NFREELIST; lcv++) {
360 		uvm.page_free[lcv].pgfl_buckets =
361 		    (bucketarray + (lcv * uvmexp.ncolors));
362 		uvm_page_init_buckets(&uvm.page_free[lcv]);
363 	}
364 	memset(pagearray, 0, pagecount * sizeof(struct vm_page));
365 
366 	/*
367 	 * init the vm_page structures and put them in the correct place.
368 	 */
369 
370 	for (lcv = 0 ; lcv < vm_nphysseg ; lcv++) {
371 		n = vm_physmem[lcv].end - vm_physmem[lcv].start;
372 
373 		/* set up page array pointers */
374 		vm_physmem[lcv].pgs = pagearray;
375 		pagearray += n;
376 		pagecount -= n;
377 		vm_physmem[lcv].lastpg = vm_physmem[lcv].pgs + (n - 1);
378 
379 		/* init and free vm_pages (we've already zeroed them) */
380 		paddr = ptoa(vm_physmem[lcv].start);
381 		for (i = 0 ; i < n ; i++, paddr += PAGE_SIZE) {
382 			vm_physmem[lcv].pgs[i].phys_addr = paddr;
383 #ifdef __HAVE_VM_PAGE_MD
384 			VM_MDPAGE_INIT(&vm_physmem[lcv].pgs[i]);
385 #endif
386 			if (atop(paddr) >= vm_physmem[lcv].avail_start &&
387 			    atop(paddr) <= vm_physmem[lcv].avail_end) {
388 				uvmexp.npages++;
389 				/* add page to free pool */
390 				uvm_pagefree(&vm_physmem[lcv].pgs[i]);
391 			}
392 		}
393 	}
394 
395 	/*
396 	 * pass up the values of virtual_space_start and
397 	 * virtual_space_end (obtained by uvm_pageboot_alloc) to the upper
398 	 * layers of the VM.
399 	 */
400 
401 	*kvm_startp = round_page(virtual_space_start);
402 	*kvm_endp = trunc_page(virtual_space_end);
403 #ifdef DEBUG
404 	/*
405 	 * steal kva for uvm_pagezerocheck().
406 	 */
407 	uvm_zerocheckkva = *kvm_startp;
408 	*kvm_startp += PAGE_SIZE;
409 #endif /* DEBUG */
410 
411 	/*
412 	 * init locks for kernel threads
413 	 */
414 
415 	simple_lock_init(&uvm.pagedaemon_lock);
416 	simple_lock_init(&uvm.aiodoned_lock);
417 
418 	/*
419 	 * init various thresholds.
420 	 */
421 
422 	uvmexp.reserve_pagedaemon = 1;
423 	uvmexp.reserve_kernel = 5;
424 	uvmexp.anonminpct = 10;
425 	uvmexp.fileminpct = 10;
426 	uvmexp.execminpct = 5;
427 	uvmexp.anonmaxpct = 80;
428 	uvmexp.filemaxpct = 50;
429 	uvmexp.execmaxpct = 30;
430 	uvmexp.anonmin = uvmexp.anonminpct * 256 / 100;
431 	uvmexp.filemin = uvmexp.fileminpct * 256 / 100;
432 	uvmexp.execmin = uvmexp.execminpct * 256 / 100;
433 	uvmexp.anonmax = uvmexp.anonmaxpct * 256 / 100;
434 	uvmexp.filemax = uvmexp.filemaxpct * 256 / 100;
435 	uvmexp.execmax = uvmexp.execmaxpct * 256 / 100;
436 	uvm_pctparam_set(&uvmexp.inactivepct, 33);
437 
438 	/*
439 	 * determine if we should zero pages in the idle loop.
440 	 */
441 
442 	uvm.page_idle_zero = vm_page_zero_enable;
443 
444 	/*
445 	 * done!
446 	 */
447 
448 	uvm.page_init_done = TRUE;
449 }
450 
451 /*
452  * uvm_setpagesize: set the page size
453  *
454  * => sets page_shift and page_mask from uvmexp.pagesize.
455  */
456 
457 void
458 uvm_setpagesize(void)
459 {
460 
461 	/*
462 	 * If uvmexp.pagesize is 0 at this point, we expect PAGE_SIZE
463 	 * to be a constant (indicated by being a non-zero value).
464 	 */
465 	if (uvmexp.pagesize == 0) {
466 		if (PAGE_SIZE == 0)
467 			panic("uvm_setpagesize: uvmexp.pagesize not set");
468 		uvmexp.pagesize = PAGE_SIZE;
469 	}
470 	uvmexp.pagemask = uvmexp.pagesize - 1;
471 	if ((uvmexp.pagemask & uvmexp.pagesize) != 0)
472 		panic("uvm_setpagesize: page size not a power of two");
473 	for (uvmexp.pageshift = 0; ; uvmexp.pageshift++)
474 		if ((1 << uvmexp.pageshift) == uvmexp.pagesize)
475 			break;
476 }
477 
478 /*
479  * uvm_pageboot_alloc: steal memory from physmem for bootstrapping
480  */
481 
482 vaddr_t
483 uvm_pageboot_alloc(vsize_t size)
484 {
485 	static boolean_t initialized = FALSE;
486 	vaddr_t addr;
487 #if !defined(PMAP_STEAL_MEMORY)
488 	vaddr_t vaddr;
489 	paddr_t paddr;
490 #endif
491 
492 	/*
493 	 * on first call to this function, initialize ourselves.
494 	 */
495 	if (initialized == FALSE) {
496 		pmap_virtual_space(&virtual_space_start, &virtual_space_end);
497 
498 		/* round it the way we like it */
499 		virtual_space_start = round_page(virtual_space_start);
500 		virtual_space_end = trunc_page(virtual_space_end);
501 
502 		initialized = TRUE;
503 	}
504 
505 	/* round to page size */
506 	size = round_page(size);
507 
508 #if defined(PMAP_STEAL_MEMORY)
509 
510 	/*
511 	 * defer bootstrap allocation to MD code (it may want to allocate
512 	 * from a direct-mapped segment).  pmap_steal_memory should adjust
513 	 * virtual_space_start/virtual_space_end if necessary.
514 	 */
515 
516 	addr = pmap_steal_memory(size, &virtual_space_start,
517 	    &virtual_space_end);
518 
519 	return(addr);
520 
521 #else /* !PMAP_STEAL_MEMORY */
522 
523 	/*
524 	 * allocate virtual memory for this request
525 	 */
526 	if (virtual_space_start == virtual_space_end ||
527 	    (virtual_space_end - virtual_space_start) < size)
528 		panic("uvm_pageboot_alloc: out of virtual space");
529 
530 	addr = virtual_space_start;
531 
532 #ifdef PMAP_GROWKERNEL
533 	/*
534 	 * If the kernel pmap can't map the requested space,
535 	 * then allocate more resources for it.
536 	 */
537 	if (uvm_maxkaddr < (addr + size)) {
538 		uvm_maxkaddr = pmap_growkernel(addr + size);
539 		if (uvm_maxkaddr < (addr + size))
540 			panic("uvm_pageboot_alloc: pmap_growkernel() failed");
541 	}
542 #endif
543 
544 	virtual_space_start += size;
545 
546 	/*
547 	 * allocate and mapin physical pages to back new virtual pages
548 	 */
549 
550 	for (vaddr = round_page(addr) ; vaddr < addr + size ;
551 	    vaddr += PAGE_SIZE) {
552 
553 		if (!uvm_page_physget(&paddr))
554 			panic("uvm_pageboot_alloc: out of memory");
555 
556 		/*
557 		 * Note this memory is no longer managed, so using
558 		 * pmap_kenter is safe.
559 		 */
560 		pmap_kenter_pa(vaddr, paddr, VM_PROT_READ|VM_PROT_WRITE);
561 	}
562 	pmap_update(pmap_kernel());
563 	return(addr);
564 #endif	/* PMAP_STEAL_MEMORY */
565 }
566 
567 #if !defined(PMAP_STEAL_MEMORY)
568 /*
569  * uvm_page_physget: "steal" one page from the vm_physmem structure.
570  *
571  * => attempt to allocate it off the end of a segment in which the "avail"
572  *    values match the start/end values.   if we can't do that, then we
573  *    will advance both values (making them equal, and removing some
574  *    vm_page structures from the non-avail area).
575  * => return false if out of memory.
576  */
577 
578 /* subroutine: try to allocate from memory chunks on the specified freelist */
579 static boolean_t uvm_page_physget_freelist(paddr_t *, int);
580 
581 static boolean_t
582 uvm_page_physget_freelist(paddr_t *paddrp, int freelist)
583 {
584 	int lcv, x;
585 
586 	/* pass 1: try allocating from a matching end */
587 #if (VM_PHYSSEG_STRAT == VM_PSTRAT_BIGFIRST)
588 	for (lcv = vm_nphysseg - 1 ; lcv >= 0 ; lcv--)
589 #else
590 	for (lcv = 0 ; lcv < vm_nphysseg ; lcv++)
591 #endif
592 	{
593 
594 		if (uvm.page_init_done == TRUE)
595 			panic("uvm_page_physget: called _after_ bootstrap");
596 
597 		if (vm_physmem[lcv].free_list != freelist)
598 			continue;
599 
600 		/* try from front */
601 		if (vm_physmem[lcv].avail_start == vm_physmem[lcv].start &&
602 		    vm_physmem[lcv].avail_start < vm_physmem[lcv].avail_end) {
603 			*paddrp = ptoa(vm_physmem[lcv].avail_start);
604 			vm_physmem[lcv].avail_start++;
605 			vm_physmem[lcv].start++;
606 			/* nothing left?   nuke it */
607 			if (vm_physmem[lcv].avail_start ==
608 			    vm_physmem[lcv].end) {
609 				if (vm_nphysseg == 1)
610 				    panic("uvm_page_physget: out of memory!");
611 				vm_nphysseg--;
612 				for (x = lcv ; x < vm_nphysseg ; x++)
613 					/* structure copy */
614 					vm_physmem[x] = vm_physmem[x+1];
615 			}
616 			return (TRUE);
617 		}
618 
619 		/* try from rear */
620 		if (vm_physmem[lcv].avail_end == vm_physmem[lcv].end &&
621 		    vm_physmem[lcv].avail_start < vm_physmem[lcv].avail_end) {
622 			*paddrp = ptoa(vm_physmem[lcv].avail_end - 1);
623 			vm_physmem[lcv].avail_end--;
624 			vm_physmem[lcv].end--;
625 			/* nothing left?   nuke it */
626 			if (vm_physmem[lcv].avail_end ==
627 			    vm_physmem[lcv].start) {
628 				if (vm_nphysseg == 1)
629 				    panic("uvm_page_physget: out of memory!");
630 				vm_nphysseg--;
631 				for (x = lcv ; x < vm_nphysseg ; x++)
632 					/* structure copy */
633 					vm_physmem[x] = vm_physmem[x+1];
634 			}
635 			return (TRUE);
636 		}
637 	}
638 
639 	/* pass2: forget about matching ends, just allocate something */
640 #if (VM_PHYSSEG_STRAT == VM_PSTRAT_BIGFIRST)
641 	for (lcv = vm_nphysseg - 1 ; lcv >= 0 ; lcv--)
642 #else
643 	for (lcv = 0 ; lcv < vm_nphysseg ; lcv++)
644 #endif
645 	{
646 
647 		/* any room in this bank? */
648 		if (vm_physmem[lcv].avail_start >= vm_physmem[lcv].avail_end)
649 			continue;  /* nope */
650 
651 		*paddrp = ptoa(vm_physmem[lcv].avail_start);
652 		vm_physmem[lcv].avail_start++;
653 		/* truncate! */
654 		vm_physmem[lcv].start = vm_physmem[lcv].avail_start;
655 
656 		/* nothing left?   nuke it */
657 		if (vm_physmem[lcv].avail_start == vm_physmem[lcv].end) {
658 			if (vm_nphysseg == 1)
659 				panic("uvm_page_physget: out of memory!");
660 			vm_nphysseg--;
661 			for (x = lcv ; x < vm_nphysseg ; x++)
662 				/* structure copy */
663 				vm_physmem[x] = vm_physmem[x+1];
664 		}
665 		return (TRUE);
666 	}
667 
668 	return (FALSE);        /* whoops! */
669 }
670 
671 boolean_t
672 uvm_page_physget(paddr_t *paddrp)
673 {
674 	int i;
675 
676 	/* try in the order of freelist preference */
677 	for (i = 0; i < VM_NFREELIST; i++)
678 		if (uvm_page_physget_freelist(paddrp, i) == TRUE)
679 			return (TRUE);
680 	return (FALSE);
681 }
682 #endif /* PMAP_STEAL_MEMORY */
683 
684 /*
685  * uvm_page_physload: load physical memory into VM system
686  *
687  * => all args are PFs
688  * => all pages in start/end get vm_page structures
689  * => areas marked by avail_start/avail_end get added to the free page pool
690  * => we are limited to VM_PHYSSEG_MAX physical memory segments
691  */
692 
693 void
694 uvm_page_physload(paddr_t start, paddr_t end, paddr_t avail_start,
695     paddr_t avail_end, int free_list)
696 {
697 	int preload, lcv;
698 	psize_t npages;
699 	struct vm_page *pgs;
700 	struct vm_physseg *ps;
701 
702 	if (uvmexp.pagesize == 0)
703 		panic("uvm_page_physload: page size not set!");
704 	if (free_list >= VM_NFREELIST || free_list < VM_FREELIST_DEFAULT)
705 		panic("uvm_page_physload: bad free list %d", free_list);
706 	if (start >= end)
707 		panic("uvm_page_physload: start >= end");
708 
709 	/*
710 	 * do we have room?
711 	 */
712 
713 	if (vm_nphysseg == VM_PHYSSEG_MAX) {
714 		printf("uvm_page_physload: unable to load physical memory "
715 		    "segment\n");
716 		printf("\t%d segments allocated, ignoring 0x%llx -> 0x%llx\n",
717 		    VM_PHYSSEG_MAX, (long long)start, (long long)end);
718 		printf("\tincrease VM_PHYSSEG_MAX\n");
719 		return;
720 	}
721 
722 	/*
723 	 * check to see if this is a "preload" (i.e. uvm_mem_init hasn't been
724 	 * called yet, so malloc is not available).
725 	 */
726 
727 	for (lcv = 0 ; lcv < vm_nphysseg ; lcv++) {
728 		if (vm_physmem[lcv].pgs)
729 			break;
730 	}
731 	preload = (lcv == vm_nphysseg);
732 
733 	/*
734 	 * if VM is already running, attempt to malloc() vm_page structures
735 	 */
736 
737 	if (!preload) {
738 #if defined(VM_PHYSSEG_NOADD)
739 		panic("uvm_page_physload: tried to add RAM after vm_mem_init");
740 #else
741 		/* XXXCDC: need some sort of lockout for this case */
742 		paddr_t paddr;
743 		npages = end - start;  /* # of pages */
744 		pgs = malloc(sizeof(struct vm_page) * npages,
745 		    M_VMPAGE, M_NOWAIT);
746 		if (pgs == NULL) {
747 			printf("uvm_page_physload: can not malloc vm_page "
748 			    "structs for segment\n");
749 			printf("\tignoring 0x%lx -> 0x%lx\n", start, end);
750 			return;
751 		}
752 		/* zero data, init phys_addr and free_list, and free pages */
753 		memset(pgs, 0, sizeof(struct vm_page) * npages);
754 		for (lcv = 0, paddr = ptoa(start) ;
755 				 lcv < npages ; lcv++, paddr += PAGE_SIZE) {
756 			pgs[lcv].phys_addr = paddr;
757 			pgs[lcv].free_list = free_list;
758 			if (atop(paddr) >= avail_start &&
759 			    atop(paddr) <= avail_end)
760 				uvm_pagefree(&pgs[lcv]);
761 		}
762 		/* XXXCDC: incomplete: need to update uvmexp.free, what else? */
763 		/* XXXCDC: need hook to tell pmap to rebuild pv_list, etc... */
764 #endif
765 	} else {
766 		pgs = NULL;
767 		npages = 0;
768 	}
769 
770 	/*
771 	 * now insert us in the proper place in vm_physmem[]
772 	 */
773 
774 #if (VM_PHYSSEG_STRAT == VM_PSTRAT_RANDOM)
775 	/* random: put it at the end (easy!) */
776 	ps = &vm_physmem[vm_nphysseg];
777 #elif (VM_PHYSSEG_STRAT == VM_PSTRAT_BSEARCH)
778 	{
779 		int x;
780 		/* sort by address for binary search */
781 		for (lcv = 0 ; lcv < vm_nphysseg ; lcv++)
782 			if (start < vm_physmem[lcv].start)
783 				break;
784 		ps = &vm_physmem[lcv];
785 		/* move back other entries, if necessary ... */
786 		for (x = vm_nphysseg ; x > lcv ; x--)
787 			/* structure copy */
788 			vm_physmem[x] = vm_physmem[x - 1];
789 	}
790 #elif (VM_PHYSSEG_STRAT == VM_PSTRAT_BIGFIRST)
791 	{
792 		int x;
793 		/* sort by largest segment first */
794 		for (lcv = 0 ; lcv < vm_nphysseg ; lcv++)
795 			if ((end - start) >
796 			    (vm_physmem[lcv].end - vm_physmem[lcv].start))
797 				break;
798 		ps = &vm_physmem[lcv];
799 		/* move back other entries, if necessary ... */
800 		for (x = vm_nphysseg ; x > lcv ; x--)
801 			/* structure copy */
802 			vm_physmem[x] = vm_physmem[x - 1];
803 	}
804 #else
805 	panic("uvm_page_physload: unknown physseg strategy selected!");
806 #endif
807 
808 	ps->start = start;
809 	ps->end = end;
810 	ps->avail_start = avail_start;
811 	ps->avail_end = avail_end;
812 	if (preload) {
813 		ps->pgs = NULL;
814 	} else {
815 		ps->pgs = pgs;
816 		ps->lastpg = pgs + npages - 1;
817 	}
818 	ps->free_list = free_list;
819 	vm_nphysseg++;
820 
821 	if (!preload)
822 		uvm_page_rehash();
823 }
824 
825 /*
826  * uvm_page_rehash: reallocate hash table based on number of free pages.
827  */
828 
829 void
830 uvm_page_rehash(void)
831 {
832 	int freepages, lcv, bucketcount, oldcount;
833 	struct pglist *newbuckets, *oldbuckets;
834 	struct vm_page *pg;
835 	size_t newsize, oldsize;
836 
837 	/*
838 	 * compute number of pages that can go in the free pool
839 	 */
840 
841 	freepages = 0;
842 	for (lcv = 0 ; lcv < vm_nphysseg ; lcv++)
843 		freepages +=
844 		    (vm_physmem[lcv].avail_end - vm_physmem[lcv].avail_start);
845 
846 	/*
847 	 * compute number of buckets needed for this number of pages
848 	 */
849 
850 	bucketcount = 1;
851 	while (bucketcount < freepages)
852 		bucketcount = bucketcount * 2;
853 
854 	/*
855 	 * compute the size of the current table and new table.
856 	 */
857 
858 	oldbuckets = uvm.page_hash;
859 	oldcount = uvm.page_nhash;
860 	oldsize = round_page(sizeof(struct pglist) * oldcount);
861 	newsize = round_page(sizeof(struct pglist) * bucketcount);
862 
863 	/*
864 	 * allocate the new buckets
865 	 */
866 
867 	newbuckets = (struct pglist *) uvm_km_alloc(kernel_map, newsize,
868 	    0, UVM_KMF_WIRED);
869 	if (newbuckets == NULL) {
870 		printf("uvm_page_physrehash: WARNING: could not grow page "
871 		    "hash table\n");
872 		return;
873 	}
874 	for (lcv = 0 ; lcv < bucketcount ; lcv++)
875 		TAILQ_INIT(&newbuckets[lcv]);
876 
877 	/*
878 	 * now replace the old buckets with the new ones and rehash everything
879 	 */
880 
881 	simple_lock(&uvm.hashlock);
882 	uvm.page_hash = newbuckets;
883 	uvm.page_nhash = bucketcount;
884 	uvm.page_hashmask = bucketcount - 1;  /* power of 2 */
885 
886 	/* ... and rehash */
887 	for (lcv = 0 ; lcv < oldcount ; lcv++) {
888 		while ((pg = oldbuckets[lcv].tqh_first) != NULL) {
889 			TAILQ_REMOVE(&oldbuckets[lcv], pg, hashq);
890 			TAILQ_INSERT_TAIL(
891 			  &uvm.page_hash[uvm_pagehash(pg->uobject, pg->offset)],
892 			  pg, hashq);
893 		}
894 	}
895 	simple_unlock(&uvm.hashlock);
896 
897 	/*
898 	 * free old bucket array if is not the boot-time table
899 	 */
900 
901 	if (oldbuckets != &uvm_bootbucket)
902 		uvm_km_free(kernel_map, (vaddr_t) oldbuckets, oldsize,
903 		    UVM_KMF_WIRED);
904 }
905 
906 /*
907  * uvm_page_recolor: Recolor the pages if the new bucket count is
908  * larger than the old one.
909  */
910 
911 void
912 uvm_page_recolor(int newncolors)
913 {
914 	struct pgflbucket *bucketarray, *oldbucketarray;
915 	struct pgfreelist pgfl;
916 	struct vm_page *pg;
917 	vsize_t bucketcount;
918 	int s, lcv, color, i, ocolors;
919 
920 	if (newncolors <= uvmexp.ncolors)
921 		return;
922 
923 	if (uvm.page_init_done == FALSE) {
924 		uvmexp.ncolors = newncolors;
925 		return;
926 	}
927 
928 	bucketcount = newncolors * VM_NFREELIST;
929 	bucketarray = malloc(bucketcount * sizeof(struct pgflbucket),
930 	    M_VMPAGE, M_NOWAIT);
931 	if (bucketarray == NULL) {
932 		printf("WARNING: unable to allocate %ld page color buckets\n",
933 		    (long) bucketcount);
934 		return;
935 	}
936 
937 	s = uvm_lock_fpageq();
938 
939 	/* Make sure we should still do this. */
940 	if (newncolors <= uvmexp.ncolors) {
941 		uvm_unlock_fpageq(s);
942 		free(bucketarray, M_VMPAGE);
943 		return;
944 	}
945 
946 	oldbucketarray = uvm.page_free[0].pgfl_buckets;
947 	ocolors = uvmexp.ncolors;
948 
949 	uvmexp.ncolors = newncolors;
950 	uvmexp.colormask = uvmexp.ncolors - 1;
951 
952 	for (lcv = 0; lcv < VM_NFREELIST; lcv++) {
953 		pgfl.pgfl_buckets = (bucketarray + (lcv * newncolors));
954 		uvm_page_init_buckets(&pgfl);
955 		for (color = 0; color < ocolors; color++) {
956 			for (i = 0; i < PGFL_NQUEUES; i++) {
957 				while ((pg = TAILQ_FIRST(&uvm.page_free[
958 				    lcv].pgfl_buckets[color].pgfl_queues[i]))
959 				    != NULL) {
960 					TAILQ_REMOVE(&uvm.page_free[
961 					    lcv].pgfl_buckets[
962 					    color].pgfl_queues[i], pg, pageq);
963 					TAILQ_INSERT_TAIL(&pgfl.pgfl_buckets[
964 					    VM_PGCOLOR_BUCKET(pg)].pgfl_queues[
965 					    i], pg, pageq);
966 				}
967 			}
968 		}
969 		uvm.page_free[lcv].pgfl_buckets = pgfl.pgfl_buckets;
970 	}
971 
972 	if (have_recolored_pages) {
973 		uvm_unlock_fpageq(s);
974 		free(oldbucketarray, M_VMPAGE);
975 		return;
976 	}
977 
978 	have_recolored_pages = TRUE;
979 	uvm_unlock_fpageq(s);
980 }
981 
982 /*
983  * uvm_pagealloc_pgfl: helper routine for uvm_pagealloc_strat
984  */
985 
986 static inline struct vm_page *
987 uvm_pagealloc_pgfl(struct pgfreelist *pgfl, int try1, int try2,
988     int *trycolorp)
989 {
990 	struct pglist *freeq;
991 	struct vm_page *pg;
992 	int color, trycolor = *trycolorp;
993 
994 	color = trycolor;
995 	do {
996 		if ((pg = TAILQ_FIRST((freeq =
997 		    &pgfl->pgfl_buckets[color].pgfl_queues[try1]))) != NULL)
998 			goto gotit;
999 		if ((pg = TAILQ_FIRST((freeq =
1000 		    &pgfl->pgfl_buckets[color].pgfl_queues[try2]))) != NULL)
1001 			goto gotit;
1002 		color = (color + 1) & uvmexp.colormask;
1003 	} while (color != trycolor);
1004 
1005 	return (NULL);
1006 
1007  gotit:
1008 	TAILQ_REMOVE(freeq, pg, pageq);
1009 	uvmexp.free--;
1010 
1011 	/* update zero'd page count */
1012 	if (pg->flags & PG_ZERO)
1013 		uvmexp.zeropages--;
1014 
1015 	if (color == trycolor)
1016 		uvmexp.colorhit++;
1017 	else {
1018 		uvmexp.colormiss++;
1019 		*trycolorp = color;
1020 	}
1021 
1022 	return (pg);
1023 }
1024 
1025 /*
1026  * uvm_pagealloc_strat: allocate vm_page from a particular free list.
1027  *
1028  * => return null if no pages free
1029  * => wake up pagedaemon if number of free pages drops below low water mark
1030  * => if obj != NULL, obj must be locked (to put in hash)
1031  * => if anon != NULL, anon must be locked (to put in anon)
1032  * => only one of obj or anon can be non-null
1033  * => caller must activate/deactivate page if it is not wired.
1034  * => free_list is ignored if strat == UVM_PGA_STRAT_NORMAL.
1035  * => policy decision: it is more important to pull a page off of the
1036  *	appropriate priority free list than it is to get a zero'd or
1037  *	unknown contents page.  This is because we live with the
1038  *	consequences of a bad free list decision for the entire
1039  *	lifetime of the page, e.g. if the page comes from memory that
1040  *	is slower to access.
1041  */
1042 
1043 struct vm_page *
1044 uvm_pagealloc_strat(struct uvm_object *obj, voff_t off, struct vm_anon *anon,
1045     int flags, int strat, int free_list)
1046 {
1047 	int lcv, try1, try2, s, zeroit = 0, color;
1048 	struct vm_page *pg;
1049 	boolean_t use_reserve;
1050 
1051 	KASSERT(obj == NULL || anon == NULL);
1052 	KASSERT(off == trunc_page(off));
1053 	LOCK_ASSERT(obj == NULL || simple_lock_held(&obj->vmobjlock));
1054 	LOCK_ASSERT(anon == NULL || simple_lock_held(&anon->an_lock));
1055 
1056 	s = uvm_lock_fpageq();
1057 
1058 	/*
1059 	 * This implements a global round-robin page coloring
1060 	 * algorithm.
1061 	 *
1062 	 * XXXJRT: Should we make the `nextcolor' per-CPU?
1063 	 * XXXJRT: What about virtually-indexed caches?
1064 	 */
1065 
1066 	color = uvm.page_free_nextcolor;
1067 
1068 	/*
1069 	 * check to see if we need to generate some free pages waking
1070 	 * the pagedaemon.
1071 	 */
1072 
1073 	UVM_KICK_PDAEMON();
1074 
1075 	/*
1076 	 * fail if any of these conditions is true:
1077 	 * [1]  there really are no free pages, or
1078 	 * [2]  only kernel "reserved" pages remain and
1079 	 *        the page isn't being allocated to a kernel object.
1080 	 * [3]  only pagedaemon "reserved" pages remain and
1081 	 *        the requestor isn't the pagedaemon.
1082 	 */
1083 
1084 	use_reserve = (flags & UVM_PGA_USERESERVE) ||
1085 		(obj && UVM_OBJ_IS_KERN_OBJECT(obj));
1086 	if ((uvmexp.free <= uvmexp.reserve_kernel && !use_reserve) ||
1087 	    (uvmexp.free <= uvmexp.reserve_pagedaemon &&
1088 	     !(use_reserve && curproc == uvm.pagedaemon_proc)))
1089 		goto fail;
1090 
1091 #if PGFL_NQUEUES != 2
1092 #error uvm_pagealloc_strat needs to be updated
1093 #endif
1094 
1095 	/*
1096 	 * If we want a zero'd page, try the ZEROS queue first, otherwise
1097 	 * we try the UNKNOWN queue first.
1098 	 */
1099 	if (flags & UVM_PGA_ZERO) {
1100 		try1 = PGFL_ZEROS;
1101 		try2 = PGFL_UNKNOWN;
1102 	} else {
1103 		try1 = PGFL_UNKNOWN;
1104 		try2 = PGFL_ZEROS;
1105 	}
1106 
1107  again:
1108 	switch (strat) {
1109 	case UVM_PGA_STRAT_NORMAL:
1110 		/* Check all freelists in descending priority order. */
1111 		for (lcv = 0; lcv < VM_NFREELIST; lcv++) {
1112 			pg = uvm_pagealloc_pgfl(&uvm.page_free[lcv],
1113 			    try1, try2, &color);
1114 			if (pg != NULL)
1115 				goto gotit;
1116 		}
1117 
1118 		/* No pages free! */
1119 		goto fail;
1120 
1121 	case UVM_PGA_STRAT_ONLY:
1122 	case UVM_PGA_STRAT_FALLBACK:
1123 		/* Attempt to allocate from the specified free list. */
1124 		KASSERT(free_list >= 0 && free_list < VM_NFREELIST);
1125 		pg = uvm_pagealloc_pgfl(&uvm.page_free[free_list],
1126 		    try1, try2, &color);
1127 		if (pg != NULL)
1128 			goto gotit;
1129 
1130 		/* Fall back, if possible. */
1131 		if (strat == UVM_PGA_STRAT_FALLBACK) {
1132 			strat = UVM_PGA_STRAT_NORMAL;
1133 			goto again;
1134 		}
1135 
1136 		/* No pages free! */
1137 		goto fail;
1138 
1139 	default:
1140 		panic("uvm_pagealloc_strat: bad strat %d", strat);
1141 		/* NOTREACHED */
1142 	}
1143 
1144  gotit:
1145 	/*
1146 	 * We now know which color we actually allocated from; set
1147 	 * the next color accordingly.
1148 	 */
1149 
1150 	uvm.page_free_nextcolor = (color + 1) & uvmexp.colormask;
1151 
1152 	/*
1153 	 * update allocation statistics and remember if we have to
1154 	 * zero the page
1155 	 */
1156 
1157 	if (flags & UVM_PGA_ZERO) {
1158 		if (pg->flags & PG_ZERO) {
1159 			uvmexp.pga_zerohit++;
1160 			zeroit = 0;
1161 		} else {
1162 			uvmexp.pga_zeromiss++;
1163 			zeroit = 1;
1164 		}
1165 	}
1166 	uvm_unlock_fpageq(s);
1167 
1168 	pg->offset = off;
1169 	pg->uobject = obj;
1170 	pg->uanon = anon;
1171 	pg->flags = PG_BUSY|PG_CLEAN|PG_FAKE;
1172 	if (anon) {
1173 		anon->an_page = pg;
1174 		pg->pqflags = PQ_ANON;
1175 		uvmexp.anonpages++;
1176 	} else {
1177 		if (obj) {
1178 			uvm_pageinsert(pg);
1179 		}
1180 		pg->pqflags = 0;
1181 	}
1182 #if defined(UVM_PAGE_TRKOWN)
1183 	pg->owner_tag = NULL;
1184 #endif
1185 	UVM_PAGE_OWN(pg, "new alloc");
1186 
1187 	if (flags & UVM_PGA_ZERO) {
1188 		/*
1189 		 * A zero'd page is not clean.  If we got a page not already
1190 		 * zero'd, then we have to zero it ourselves.
1191 		 */
1192 		pg->flags &= ~PG_CLEAN;
1193 		if (zeroit)
1194 			pmap_zero_page(VM_PAGE_TO_PHYS(pg));
1195 	}
1196 
1197 	return(pg);
1198 
1199  fail:
1200 	uvm_unlock_fpageq(s);
1201 	return (NULL);
1202 }
1203 
1204 /*
1205  * uvm_pagereplace: replace a page with another
1206  *
1207  * => object must be locked
1208  */
1209 
1210 void
1211 uvm_pagereplace(struct vm_page *oldpg, struct vm_page *newpg)
1212 {
1213 
1214 	KASSERT((oldpg->flags & PG_TABLED) != 0);
1215 	KASSERT(oldpg->uobject != NULL);
1216 	KASSERT((newpg->flags & PG_TABLED) == 0);
1217 	KASSERT(newpg->uobject == NULL);
1218 	LOCK_ASSERT(simple_lock_held(&oldpg->uobject->vmobjlock));
1219 
1220 	newpg->uobject = oldpg->uobject;
1221 	newpg->offset = oldpg->offset;
1222 
1223 	uvm_pageinsert_after(newpg, oldpg);
1224 	uvm_pageremove(oldpg);
1225 }
1226 
1227 /*
1228  * uvm_pagerealloc: reallocate a page from one object to another
1229  *
1230  * => both objects must be locked
1231  */
1232 
1233 void
1234 uvm_pagerealloc(struct vm_page *pg, struct uvm_object *newobj, voff_t newoff)
1235 {
1236 	/*
1237 	 * remove it from the old object
1238 	 */
1239 
1240 	if (pg->uobject) {
1241 		uvm_pageremove(pg);
1242 	}
1243 
1244 	/*
1245 	 * put it in the new object
1246 	 */
1247 
1248 	if (newobj) {
1249 		pg->uobject = newobj;
1250 		pg->offset = newoff;
1251 		uvm_pageinsert(pg);
1252 	}
1253 }
1254 
1255 #ifdef DEBUG
1256 /*
1257  * check if page is zero-filled
1258  *
1259  *  - called with free page queue lock held.
1260  */
1261 void
1262 uvm_pagezerocheck(struct vm_page *pg)
1263 {
1264 	int *p, *ep;
1265 
1266 	KASSERT(uvm_zerocheckkva != 0);
1267 	LOCK_ASSERT(simple_lock_held(&uvm.fpageqlock));
1268 
1269 	/*
1270 	 * XXX assuming pmap_kenter_pa and pmap_kremove never call
1271 	 * uvm page allocator.
1272 	 *
1273 	 * it might be better to have "CPU-local temporary map" pmap interface.
1274 	 */
1275 	pmap_kenter_pa(uvm_zerocheckkva, VM_PAGE_TO_PHYS(pg), VM_PROT_READ);
1276 	p = (int *)uvm_zerocheckkva;
1277 	ep = (int *)((char *)p + PAGE_SIZE);
1278 	pmap_update(pmap_kernel());
1279 	while (p < ep) {
1280 		if (*p != 0)
1281 			panic("PG_ZERO page isn't zero-filled");
1282 		p++;
1283 	}
1284 	pmap_kremove(uvm_zerocheckkva, PAGE_SIZE);
1285 }
1286 #endif /* DEBUG */
1287 
1288 /*
1289  * uvm_pagefree: free page
1290  *
1291  * => erase page's identity (i.e. remove from hash/object)
1292  * => put page on free list
1293  * => caller must lock owning object (either anon or uvm_object)
1294  * => caller must lock page queues
1295  * => assumes all valid mappings of pg are gone
1296  */
1297 
1298 void
1299 uvm_pagefree(struct vm_page *pg)
1300 {
1301 	int s;
1302 	struct pglist *pgfl;
1303 	boolean_t iszero;
1304 
1305 	KASSERT((pg->flags & PG_PAGEOUT) == 0);
1306 	LOCK_ASSERT(simple_lock_held(&uvm.pageqlock) ||
1307 		    (pg->pqflags & (PQ_ACTIVE|PQ_INACTIVE)) == 0);
1308 	LOCK_ASSERT(pg->uobject == NULL ||
1309 		    simple_lock_held(&pg->uobject->vmobjlock));
1310 	LOCK_ASSERT(pg->uobject != NULL || pg->uanon == NULL ||
1311 		    simple_lock_held(&pg->uanon->an_lock));
1312 
1313 #ifdef DEBUG
1314 	if (pg->uobject == (void *)0xdeadbeef &&
1315 	    pg->uanon == (void *)0xdeadbeef) {
1316 		panic("uvm_pagefree: freeing free page %p", pg);
1317 	}
1318 #endif /* DEBUG */
1319 
1320 	/*
1321 	 * if the page is loaned, resolve the loan instead of freeing.
1322 	 */
1323 
1324 	if (pg->loan_count) {
1325 		KASSERT(pg->wire_count == 0);
1326 
1327 		/*
1328 		 * if the page is owned by an anon then we just want to
1329 		 * drop anon ownership.  the kernel will free the page when
1330 		 * it is done with it.  if the page is owned by an object,
1331 		 * remove it from the object and mark it dirty for the benefit
1332 		 * of possible anon owners.
1333 		 *
1334 		 * regardless of previous ownership, wakeup any waiters,
1335 		 * unbusy the page, and we're done.
1336 		 */
1337 
1338 		if (pg->uobject != NULL) {
1339 			uvm_pageremove(pg);
1340 			pg->flags &= ~PG_CLEAN;
1341 		} else if (pg->uanon != NULL) {
1342 			if ((pg->pqflags & PQ_ANON) == 0) {
1343 				pg->loan_count--;
1344 			} else {
1345 				pg->pqflags &= ~PQ_ANON;
1346 				uvmexp.anonpages--;
1347 			}
1348 			pg->uanon->an_page = NULL;
1349 			pg->uanon = NULL;
1350 		}
1351 		if (pg->flags & PG_WANTED) {
1352 			wakeup(pg);
1353 		}
1354 		pg->flags &= ~(PG_WANTED|PG_BUSY|PG_RELEASED|PG_PAGER1);
1355 #ifdef UVM_PAGE_TRKOWN
1356 		pg->owner_tag = NULL;
1357 #endif
1358 		if (pg->loan_count) {
1359 			uvm_pagedequeue(pg);
1360 			return;
1361 		}
1362 	}
1363 
1364 	/*
1365 	 * remove page from its object or anon.
1366 	 */
1367 
1368 	if (pg->uobject != NULL) {
1369 		uvm_pageremove(pg);
1370 	} else if (pg->uanon != NULL) {
1371 		pg->uanon->an_page = NULL;
1372 		uvmexp.anonpages--;
1373 	}
1374 
1375 	/*
1376 	 * now remove the page from the queues.
1377 	 */
1378 
1379 	uvm_pagedequeue(pg);
1380 
1381 	/*
1382 	 * if the page was wired, unwire it now.
1383 	 */
1384 
1385 	if (pg->wire_count) {
1386 		pg->wire_count = 0;
1387 		uvmexp.wired--;
1388 	}
1389 
1390 	/*
1391 	 * and put on free queue
1392 	 */
1393 
1394 	iszero = (pg->flags & PG_ZERO);
1395 	pgfl = &uvm.page_free[uvm_page_lookup_freelist(pg)].
1396 	    pgfl_buckets[VM_PGCOLOR_BUCKET(pg)].
1397 	    pgfl_queues[iszero ? PGFL_ZEROS : PGFL_UNKNOWN];
1398 
1399 	pg->pqflags = PQ_FREE;
1400 #ifdef DEBUG
1401 	pg->uobject = (void *)0xdeadbeef;
1402 	pg->offset = 0xdeadbeef;
1403 	pg->uanon = (void *)0xdeadbeef;
1404 #endif
1405 
1406 	s = uvm_lock_fpageq();
1407 
1408 #ifdef DEBUG
1409 	if (iszero)
1410 		uvm_pagezerocheck(pg);
1411 #endif /* DEBUG */
1412 
1413 	TAILQ_INSERT_HEAD(pgfl, pg, pageq);
1414 	uvmexp.free++;
1415 	if (iszero)
1416 		uvmexp.zeropages++;
1417 
1418 	if (uvmexp.zeropages < UVM_PAGEZERO_TARGET)
1419 		uvm.page_idle_zero = vm_page_zero_enable;
1420 
1421 	uvm_unlock_fpageq(s);
1422 }
1423 
1424 /*
1425  * uvm_page_unbusy: unbusy an array of pages.
1426  *
1427  * => pages must either all belong to the same object, or all belong to anons.
1428  * => if pages are object-owned, object must be locked.
1429  * => if pages are anon-owned, anons must be locked.
1430  * => caller must lock page queues if pages may be released.
1431  * => caller must make sure that anon-owned pages are not PG_RELEASED.
1432  */
1433 
1434 void
1435 uvm_page_unbusy(struct vm_page **pgs, int npgs)
1436 {
1437 	struct vm_page *pg;
1438 	int i;
1439 	UVMHIST_FUNC("uvm_page_unbusy"); UVMHIST_CALLED(ubchist);
1440 
1441 	for (i = 0; i < npgs; i++) {
1442 		pg = pgs[i];
1443 		if (pg == NULL || pg == PGO_DONTCARE) {
1444 			continue;
1445 		}
1446 
1447 		LOCK_ASSERT(pg->uobject == NULL ||
1448 		    simple_lock_held(&pg->uobject->vmobjlock));
1449 		LOCK_ASSERT(pg->uobject != NULL ||
1450 		    (pg->uanon != NULL &&
1451 		    simple_lock_held(&pg->uanon->an_lock)));
1452 
1453 		KASSERT(pg->flags & PG_BUSY);
1454 		KASSERT((pg->flags & PG_PAGEOUT) == 0);
1455 		if (pg->flags & PG_WANTED) {
1456 			wakeup(pg);
1457 		}
1458 		if (pg->flags & PG_RELEASED) {
1459 			UVMHIST_LOG(ubchist, "releasing pg %p", pg,0,0,0);
1460 			KASSERT(pg->uobject != NULL ||
1461 			    (pg->uanon != NULL && pg->uanon->an_ref > 0));
1462 			pg->flags &= ~PG_RELEASED;
1463 			uvm_pagefree(pg);
1464 		} else {
1465 			UVMHIST_LOG(ubchist, "unbusying pg %p", pg,0,0,0);
1466 			pg->flags &= ~(PG_WANTED|PG_BUSY);
1467 			UVM_PAGE_OWN(pg, NULL);
1468 		}
1469 	}
1470 }
1471 
1472 #if defined(UVM_PAGE_TRKOWN)
1473 /*
1474  * uvm_page_own: set or release page ownership
1475  *
1476  * => this is a debugging function that keeps track of who sets PG_BUSY
1477  *	and where they do it.   it can be used to track down problems
1478  *	such a process setting "PG_BUSY" and never releasing it.
1479  * => page's object [if any] must be locked
1480  * => if "tag" is NULL then we are releasing page ownership
1481  */
1482 void
1483 uvm_page_own(struct vm_page *pg, const char *tag)
1484 {
1485 	KASSERT((pg->flags & (PG_PAGEOUT|PG_RELEASED)) == 0);
1486 
1487 	/* gain ownership? */
1488 	if (tag) {
1489 		if (pg->owner_tag) {
1490 			printf("uvm_page_own: page %p already owned "
1491 			    "by proc %d [%s]\n", pg,
1492 			    pg->owner, pg->owner_tag);
1493 			panic("uvm_page_own");
1494 		}
1495 		pg->owner = (curproc) ? curproc->p_pid :  (pid_t) -1;
1496 		pg->owner_tag = tag;
1497 		return;
1498 	}
1499 
1500 	/* drop ownership */
1501 	if (pg->owner_tag == NULL) {
1502 		printf("uvm_page_own: dropping ownership of an non-owned "
1503 		    "page (%p)\n", pg);
1504 		panic("uvm_page_own");
1505 	}
1506 	KASSERT((pg->pqflags & (PQ_ACTIVE|PQ_INACTIVE)) ||
1507 	    (pg->uanon == NULL && pg->uobject == NULL) ||
1508 	    pg->uobject == uvm.kernel_object ||
1509 	    pg->wire_count > 0 ||
1510 	    (pg->loan_count == 1 && pg->uanon == NULL) ||
1511 	    pg->loan_count > 1);
1512 	pg->owner_tag = NULL;
1513 }
1514 #endif
1515 
1516 /*
1517  * uvm_pageidlezero: zero free pages while the system is idle.
1518  *
1519  * => try to complete one color bucket at a time, to reduce our impact
1520  *	on the CPU cache.
1521  * => we loop until we either reach the target or whichqs indicates that
1522  *	there is a process ready to run.
1523  */
1524 void
1525 uvm_pageidlezero(void)
1526 {
1527 	struct vm_page *pg;
1528 	struct pgfreelist *pgfl;
1529 	int free_list, s, firstbucket;
1530 	static int nextbucket;
1531 
1532 	KERNEL_LOCK(LK_EXCLUSIVE | LK_CANRECURSE);
1533 	s = uvm_lock_fpageq();
1534 	firstbucket = nextbucket;
1535 	do {
1536 		if (sched_whichqs != 0)
1537 			goto quit;
1538 		if (uvmexp.zeropages >= UVM_PAGEZERO_TARGET) {
1539 			uvm.page_idle_zero = FALSE;
1540 			goto quit;
1541 		}
1542 		for (free_list = 0; free_list < VM_NFREELIST; free_list++) {
1543 			pgfl = &uvm.page_free[free_list];
1544 			while ((pg = TAILQ_FIRST(&pgfl->pgfl_buckets[
1545 			    nextbucket].pgfl_queues[PGFL_UNKNOWN])) != NULL) {
1546 				if (sched_whichqs != 0)
1547 					goto quit;
1548 
1549 				TAILQ_REMOVE(&pgfl->pgfl_buckets[
1550 				    nextbucket].pgfl_queues[PGFL_UNKNOWN],
1551 				    pg, pageq);
1552 				uvmexp.free--;
1553 				uvm_unlock_fpageq(s);
1554 				KERNEL_UNLOCK();
1555 #ifdef PMAP_PAGEIDLEZERO
1556 				if (!PMAP_PAGEIDLEZERO(VM_PAGE_TO_PHYS(pg))) {
1557 
1558 					/*
1559 					 * The machine-dependent code detected
1560 					 * some reason for us to abort zeroing
1561 					 * pages, probably because there is a
1562 					 * process now ready to run.
1563 					 */
1564 
1565 					KERNEL_LOCK(
1566 					    LK_EXCLUSIVE | LK_CANRECURSE);
1567 					s = uvm_lock_fpageq();
1568 					TAILQ_INSERT_HEAD(&pgfl->pgfl_buckets[
1569 					    nextbucket].pgfl_queues[
1570 					    PGFL_UNKNOWN], pg, pageq);
1571 					uvmexp.free++;
1572 					uvmexp.zeroaborts++;
1573 					goto quit;
1574 				}
1575 #else
1576 				pmap_zero_page(VM_PAGE_TO_PHYS(pg));
1577 #endif /* PMAP_PAGEIDLEZERO */
1578 				pg->flags |= PG_ZERO;
1579 
1580 				KERNEL_LOCK(LK_EXCLUSIVE | LK_CANRECURSE);
1581 				s = uvm_lock_fpageq();
1582 				TAILQ_INSERT_HEAD(&pgfl->pgfl_buckets[
1583 				    nextbucket].pgfl_queues[PGFL_ZEROS],
1584 				    pg, pageq);
1585 				uvmexp.free++;
1586 				uvmexp.zeropages++;
1587 			}
1588 		}
1589 		nextbucket = (nextbucket + 1) & uvmexp.colormask;
1590 	} while (nextbucket != firstbucket);
1591 quit:
1592 	uvm_unlock_fpageq(s);
1593 	KERNEL_UNLOCK();
1594 }
1595