xref: /netbsd-src/sys/uvm/uvm_page.c (revision 23c8222edbfb0f0932d88a8351d3a0cf817dfb9e)
1 /*	$NetBSD: uvm_page.c,v 1.101 2004/10/23 21:29:27 yamt 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 1.101 2004/10/23 21:29:27 yamt 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 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(pg, where)
172 	struct vm_page *pg;
173 	struct vm_page *where;
174 {
175 	struct pglist *buck;
176 	struct uvm_object *uobj = pg->uobject;
177 
178 	KASSERT((pg->flags & PG_TABLED) == 0);
179 	KASSERT(where == NULL || (where->flags & PG_TABLED));
180 	KASSERT(where == NULL || (where->uobject == uobj));
181 	buck = &uvm.page_hash[uvm_pagehash(uobj, pg->offset)];
182 	simple_lock(&uvm.hashlock);
183 	TAILQ_INSERT_TAIL(buck, pg, hashq);
184 	simple_unlock(&uvm.hashlock);
185 
186 	if (UVM_OBJ_IS_VNODE(uobj)) {
187 		if (uobj->uo_npages == 0) {
188 			struct vnode *vp = (struct vnode *)uobj;
189 
190 			vholdl(vp);
191 		}
192 		if (UVM_OBJ_IS_VTEXT(uobj)) {
193 			uvmexp.execpages++;
194 		} else {
195 			uvmexp.filepages++;
196 		}
197 	} else if (UVM_OBJ_IS_AOBJ(uobj)) {
198 		uvmexp.anonpages++;
199 	}
200 
201 	if (where)
202 		TAILQ_INSERT_AFTER(&uobj->memq, where, pg, listq);
203 	else
204 		TAILQ_INSERT_TAIL(&uobj->memq, pg, listq);
205 	pg->flags |= PG_TABLED;
206 	uobj->uo_npages++;
207 }
208 
209 __inline static void
210 uvm_pageinsert(pg)
211 	struct vm_page *pg;
212 {
213 
214 	uvm_pageinsert_after(pg, NULL);
215 }
216 
217 /*
218  * uvm_page_remove: remove page from object and hash
219  *
220  * => caller must lock object
221  * => caller must lock page queues
222  */
223 
224 static __inline void
225 uvm_pageremove(pg)
226 	struct vm_page *pg;
227 {
228 	struct pglist *buck;
229 	struct uvm_object *uobj = pg->uobject;
230 
231 	KASSERT(pg->flags & PG_TABLED);
232 	buck = &uvm.page_hash[uvm_pagehash(uobj, pg->offset)];
233 	simple_lock(&uvm.hashlock);
234 	TAILQ_REMOVE(buck, pg, hashq);
235 	simple_unlock(&uvm.hashlock);
236 
237 	if (UVM_OBJ_IS_VNODE(uobj)) {
238 		if (uobj->uo_npages == 1) {
239 			struct vnode *vp = (struct vnode *)uobj;
240 
241 			holdrelel(vp);
242 		}
243 		if (UVM_OBJ_IS_VTEXT(uobj)) {
244 			uvmexp.execpages--;
245 		} else {
246 			uvmexp.filepages--;
247 		}
248 	} else if (UVM_OBJ_IS_AOBJ(uobj)) {
249 		uvmexp.anonpages--;
250 	}
251 
252 	/* object should be locked */
253 	uobj->uo_npages--;
254 	TAILQ_REMOVE(&uobj->memq, pg, listq);
255 	pg->flags &= ~PG_TABLED;
256 	pg->uobject = NULL;
257 }
258 
259 static void
260 uvm_page_init_buckets(struct pgfreelist *pgfl)
261 {
262 	int color, i;
263 
264 	for (color = 0; color < uvmexp.ncolors; color++) {
265 		for (i = 0; i < PGFL_NQUEUES; i++) {
266 			TAILQ_INIT(&pgfl->pgfl_buckets[color].pgfl_queues[i]);
267 		}
268 	}
269 }
270 
271 /*
272  * uvm_page_init: init the page system.   called from uvm_init().
273  *
274  * => we return the range of kernel virtual memory in kvm_startp/kvm_endp
275  */
276 
277 void
278 uvm_page_init(kvm_startp, kvm_endp)
279 	vaddr_t *kvm_startp, *kvm_endp;
280 {
281 	vsize_t freepages, pagecount, bucketcount, n;
282 	struct pgflbucket *bucketarray;
283 	struct vm_page *pagearray;
284 	int lcv;
285 	u_int i;
286 	paddr_t paddr;
287 
288 	/*
289 	 * init the page queues and page queue locks, except the free
290 	 * list; we allocate that later (with the initial vm_page
291 	 * structures).
292 	 */
293 
294 	TAILQ_INIT(&uvm.page_active);
295 	TAILQ_INIT(&uvm.page_inactive);
296 	simple_lock_init(&uvm.pageqlock);
297 	simple_lock_init(&uvm.fpageqlock);
298 
299 	/*
300 	 * init the <obj,offset> => <page> hash table.  for now
301 	 * we just have one bucket (the bootstrap bucket).  later on we
302 	 * will allocate new buckets as we dynamically resize the hash table.
303 	 */
304 
305 	uvm.page_nhash = 1;			/* 1 bucket */
306 	uvm.page_hashmask = 0;			/* mask for hash function */
307 	uvm.page_hash = &uvm_bootbucket;	/* install bootstrap bucket */
308 	TAILQ_INIT(uvm.page_hash);		/* init hash table */
309 	simple_lock_init(&uvm.hashlock);	/* init hash table lock */
310 
311 	/*
312 	 * allocate vm_page structures.
313 	 */
314 
315 	/*
316 	 * sanity check:
317 	 * before calling this function the MD code is expected to register
318 	 * some free RAM with the uvm_page_physload() function.   our job
319 	 * now is to allocate vm_page structures for this memory.
320 	 */
321 
322 	if (vm_nphysseg == 0)
323 		panic("uvm_page_bootstrap: no memory pre-allocated");
324 
325 	/*
326 	 * first calculate the number of free pages...
327 	 *
328 	 * note that we use start/end rather than avail_start/avail_end.
329 	 * this allows us to allocate extra vm_page structures in case we
330 	 * want to return some memory to the pool after booting.
331 	 */
332 
333 	freepages = 0;
334 	for (lcv = 0 ; lcv < vm_nphysseg ; lcv++)
335 		freepages += (vm_physmem[lcv].end - vm_physmem[lcv].start);
336 
337 	/*
338 	 * Let MD code initialize the number of colors, or default
339 	 * to 1 color if MD code doesn't care.
340 	 */
341 	if (uvmexp.ncolors == 0)
342 		uvmexp.ncolors = 1;
343 	uvmexp.colormask = uvmexp.ncolors - 1;
344 
345 	/*
346 	 * we now know we have (PAGE_SIZE * freepages) bytes of memory we can
347 	 * use.   for each page of memory we use we need a vm_page structure.
348 	 * thus, the total number of pages we can use is the total size of
349 	 * the memory divided by the PAGE_SIZE plus the size of the vm_page
350 	 * structure.   we add one to freepages as a fudge factor to avoid
351 	 * truncation errors (since we can only allocate in terms of whole
352 	 * pages).
353 	 */
354 
355 	bucketcount = uvmexp.ncolors * VM_NFREELIST;
356 	pagecount = ((freepages + 1) << PAGE_SHIFT) /
357 	    (PAGE_SIZE + sizeof(struct vm_page));
358 
359 	bucketarray = (void *)uvm_pageboot_alloc((bucketcount *
360 	    sizeof(struct pgflbucket)) + (pagecount *
361 	    sizeof(struct vm_page)));
362 	pagearray = (struct vm_page *)(bucketarray + bucketcount);
363 
364 	for (lcv = 0; lcv < VM_NFREELIST; lcv++) {
365 		uvm.page_free[lcv].pgfl_buckets =
366 		    (bucketarray + (lcv * uvmexp.ncolors));
367 		uvm_page_init_buckets(&uvm.page_free[lcv]);
368 	}
369 	memset(pagearray, 0, pagecount * sizeof(struct vm_page));
370 
371 	/*
372 	 * init the vm_page structures and put them in the correct place.
373 	 */
374 
375 	for (lcv = 0 ; lcv < vm_nphysseg ; lcv++) {
376 		n = vm_physmem[lcv].end - vm_physmem[lcv].start;
377 
378 		/* set up page array pointers */
379 		vm_physmem[lcv].pgs = pagearray;
380 		pagearray += n;
381 		pagecount -= n;
382 		vm_physmem[lcv].lastpg = vm_physmem[lcv].pgs + (n - 1);
383 
384 		/* init and free vm_pages (we've already zeroed them) */
385 		paddr = ptoa(vm_physmem[lcv].start);
386 		for (i = 0 ; i < n ; i++, paddr += PAGE_SIZE) {
387 			vm_physmem[lcv].pgs[i].phys_addr = paddr;
388 #ifdef __HAVE_VM_PAGE_MD
389 			VM_MDPAGE_INIT(&vm_physmem[lcv].pgs[i]);
390 #endif
391 			if (atop(paddr) >= vm_physmem[lcv].avail_start &&
392 			    atop(paddr) <= vm_physmem[lcv].avail_end) {
393 				uvmexp.npages++;
394 				/* add page to free pool */
395 				uvm_pagefree(&vm_physmem[lcv].pgs[i]);
396 			}
397 		}
398 	}
399 
400 	/*
401 	 * pass up the values of virtual_space_start and
402 	 * virtual_space_end (obtained by uvm_pageboot_alloc) to the upper
403 	 * layers of the VM.
404 	 */
405 
406 	*kvm_startp = round_page(virtual_space_start);
407 	*kvm_endp = trunc_page(virtual_space_end);
408 #ifdef DEBUG
409 	/*
410 	 * steal kva for uvm_pagezerocheck().
411 	 */
412 	uvm_zerocheckkva = *kvm_startp;
413 	*kvm_startp += PAGE_SIZE;
414 #endif /* DEBUG */
415 
416 	/*
417 	 * init locks for kernel threads
418 	 */
419 
420 	simple_lock_init(&uvm.pagedaemon_lock);
421 	simple_lock_init(&uvm.aiodoned_lock);
422 
423 	/*
424 	 * init various thresholds.
425 	 */
426 
427 	uvmexp.reserve_pagedaemon = 1;
428 	uvmexp.reserve_kernel = 5;
429 	uvmexp.anonminpct = 10;
430 	uvmexp.fileminpct = 10;
431 	uvmexp.execminpct = 5;
432 	uvmexp.anonmaxpct = 80;
433 	uvmexp.filemaxpct = 50;
434 	uvmexp.execmaxpct = 30;
435 	uvmexp.anonmin = uvmexp.anonminpct * 256 / 100;
436 	uvmexp.filemin = uvmexp.fileminpct * 256 / 100;
437 	uvmexp.execmin = uvmexp.execminpct * 256 / 100;
438 	uvmexp.anonmax = uvmexp.anonmaxpct * 256 / 100;
439 	uvmexp.filemax = uvmexp.filemaxpct * 256 / 100;
440 	uvmexp.execmax = uvmexp.execmaxpct * 256 / 100;
441 
442 	/*
443 	 * determine if we should zero pages in the idle loop.
444 	 */
445 
446 	uvm.page_idle_zero = vm_page_zero_enable;
447 
448 	/*
449 	 * done!
450 	 */
451 
452 	uvm.page_init_done = TRUE;
453 }
454 
455 /*
456  * uvm_setpagesize: set the page size
457  *
458  * => sets page_shift and page_mask from uvmexp.pagesize.
459  */
460 
461 void
462 uvm_setpagesize()
463 {
464 
465 	/*
466 	 * If uvmexp.pagesize is 0 at this point, we expect PAGE_SIZE
467 	 * to be a constant (indicated by being a non-zero value).
468 	 */
469 	if (uvmexp.pagesize == 0) {
470 		if (PAGE_SIZE == 0)
471 			panic("uvm_setpagesize: uvmexp.pagesize not set");
472 		uvmexp.pagesize = PAGE_SIZE;
473 	}
474 	uvmexp.pagemask = uvmexp.pagesize - 1;
475 	if ((uvmexp.pagemask & uvmexp.pagesize) != 0)
476 		panic("uvm_setpagesize: page size not a power of two");
477 	for (uvmexp.pageshift = 0; ; uvmexp.pageshift++)
478 		if ((1 << uvmexp.pageshift) == uvmexp.pagesize)
479 			break;
480 }
481 
482 /*
483  * uvm_pageboot_alloc: steal memory from physmem for bootstrapping
484  */
485 
486 vaddr_t
487 uvm_pageboot_alloc(size)
488 	vsize_t size;
489 {
490 	static boolean_t initialized = FALSE;
491 	vaddr_t addr;
492 #if !defined(PMAP_STEAL_MEMORY)
493 	vaddr_t vaddr;
494 	paddr_t paddr;
495 #endif
496 
497 	/*
498 	 * on first call to this function, initialize ourselves.
499 	 */
500 	if (initialized == FALSE) {
501 		pmap_virtual_space(&virtual_space_start, &virtual_space_end);
502 
503 		/* round it the way we like it */
504 		virtual_space_start = round_page(virtual_space_start);
505 		virtual_space_end = trunc_page(virtual_space_end);
506 
507 		initialized = TRUE;
508 	}
509 
510 	/* round to page size */
511 	size = round_page(size);
512 
513 #if defined(PMAP_STEAL_MEMORY)
514 
515 	/*
516 	 * defer bootstrap allocation to MD code (it may want to allocate
517 	 * from a direct-mapped segment).  pmap_steal_memory should adjust
518 	 * virtual_space_start/virtual_space_end if necessary.
519 	 */
520 
521 	addr = pmap_steal_memory(size, &virtual_space_start,
522 	    &virtual_space_end);
523 
524 	return(addr);
525 
526 #else /* !PMAP_STEAL_MEMORY */
527 
528 	/*
529 	 * allocate virtual memory for this request
530 	 */
531 	if (virtual_space_start == virtual_space_end ||
532 	    (virtual_space_end - virtual_space_start) < size)
533 		panic("uvm_pageboot_alloc: out of virtual space");
534 
535 	addr = virtual_space_start;
536 
537 #ifdef PMAP_GROWKERNEL
538 	/*
539 	 * If the kernel pmap can't map the requested space,
540 	 * then allocate more resources for it.
541 	 */
542 	if (uvm_maxkaddr < (addr + size)) {
543 		uvm_maxkaddr = pmap_growkernel(addr + size);
544 		if (uvm_maxkaddr < (addr + size))
545 			panic("uvm_pageboot_alloc: pmap_growkernel() failed");
546 	}
547 #endif
548 
549 	virtual_space_start += size;
550 
551 	/*
552 	 * allocate and mapin physical pages to back new virtual pages
553 	 */
554 
555 	for (vaddr = round_page(addr) ; vaddr < addr + size ;
556 	    vaddr += PAGE_SIZE) {
557 
558 		if (!uvm_page_physget(&paddr))
559 			panic("uvm_pageboot_alloc: out of memory");
560 
561 		/*
562 		 * Note this memory is no longer managed, so using
563 		 * pmap_kenter is safe.
564 		 */
565 		pmap_kenter_pa(vaddr, paddr, VM_PROT_READ|VM_PROT_WRITE);
566 	}
567 	pmap_update(pmap_kernel());
568 	return(addr);
569 #endif	/* PMAP_STEAL_MEMORY */
570 }
571 
572 #if !defined(PMAP_STEAL_MEMORY)
573 /*
574  * uvm_page_physget: "steal" one page from the vm_physmem structure.
575  *
576  * => attempt to allocate it off the end of a segment in which the "avail"
577  *    values match the start/end values.   if we can't do that, then we
578  *    will advance both values (making them equal, and removing some
579  *    vm_page structures from the non-avail area).
580  * => return false if out of memory.
581  */
582 
583 /* subroutine: try to allocate from memory chunks on the specified freelist */
584 static boolean_t uvm_page_physget_freelist(paddr_t *, int);
585 
586 static boolean_t
587 uvm_page_physget_freelist(paddrp, freelist)
588 	paddr_t *paddrp;
589 	int freelist;
590 {
591 	int lcv, x;
592 
593 	/* pass 1: try allocating from a matching end */
594 #if (VM_PHYSSEG_STRAT == VM_PSTRAT_BIGFIRST)
595 	for (lcv = vm_nphysseg - 1 ; lcv >= 0 ; lcv--)
596 #else
597 	for (lcv = 0 ; lcv < vm_nphysseg ; lcv++)
598 #endif
599 	{
600 
601 		if (uvm.page_init_done == TRUE)
602 			panic("uvm_page_physget: called _after_ bootstrap");
603 
604 		if (vm_physmem[lcv].free_list != freelist)
605 			continue;
606 
607 		/* try from front */
608 		if (vm_physmem[lcv].avail_start == vm_physmem[lcv].start &&
609 		    vm_physmem[lcv].avail_start < vm_physmem[lcv].avail_end) {
610 			*paddrp = ptoa(vm_physmem[lcv].avail_start);
611 			vm_physmem[lcv].avail_start++;
612 			vm_physmem[lcv].start++;
613 			/* nothing left?   nuke it */
614 			if (vm_physmem[lcv].avail_start ==
615 			    vm_physmem[lcv].end) {
616 				if (vm_nphysseg == 1)
617 				    panic("uvm_page_physget: out of memory!");
618 				vm_nphysseg--;
619 				for (x = lcv ; x < vm_nphysseg ; x++)
620 					/* structure copy */
621 					vm_physmem[x] = vm_physmem[x+1];
622 			}
623 			return (TRUE);
624 		}
625 
626 		/* try from rear */
627 		if (vm_physmem[lcv].avail_end == vm_physmem[lcv].end &&
628 		    vm_physmem[lcv].avail_start < vm_physmem[lcv].avail_end) {
629 			*paddrp = ptoa(vm_physmem[lcv].avail_end - 1);
630 			vm_physmem[lcv].avail_end--;
631 			vm_physmem[lcv].end--;
632 			/* nothing left?   nuke it */
633 			if (vm_physmem[lcv].avail_end ==
634 			    vm_physmem[lcv].start) {
635 				if (vm_nphysseg == 1)
636 				    panic("uvm_page_physget: out of memory!");
637 				vm_nphysseg--;
638 				for (x = lcv ; x < vm_nphysseg ; x++)
639 					/* structure copy */
640 					vm_physmem[x] = vm_physmem[x+1];
641 			}
642 			return (TRUE);
643 		}
644 	}
645 
646 	/* pass2: forget about matching ends, just allocate something */
647 #if (VM_PHYSSEG_STRAT == VM_PSTRAT_BIGFIRST)
648 	for (lcv = vm_nphysseg - 1 ; lcv >= 0 ; lcv--)
649 #else
650 	for (lcv = 0 ; lcv < vm_nphysseg ; lcv++)
651 #endif
652 	{
653 
654 		/* any room in this bank? */
655 		if (vm_physmem[lcv].avail_start >= vm_physmem[lcv].avail_end)
656 			continue;  /* nope */
657 
658 		*paddrp = ptoa(vm_physmem[lcv].avail_start);
659 		vm_physmem[lcv].avail_start++;
660 		/* truncate! */
661 		vm_physmem[lcv].start = vm_physmem[lcv].avail_start;
662 
663 		/* nothing left?   nuke it */
664 		if (vm_physmem[lcv].avail_start == vm_physmem[lcv].end) {
665 			if (vm_nphysseg == 1)
666 				panic("uvm_page_physget: out of memory!");
667 			vm_nphysseg--;
668 			for (x = lcv ; x < vm_nphysseg ; x++)
669 				/* structure copy */
670 				vm_physmem[x] = vm_physmem[x+1];
671 		}
672 		return (TRUE);
673 	}
674 
675 	return (FALSE);        /* whoops! */
676 }
677 
678 boolean_t
679 uvm_page_physget(paddrp)
680 	paddr_t *paddrp;
681 {
682 	int i;
683 
684 	/* try in the order of freelist preference */
685 	for (i = 0; i < VM_NFREELIST; i++)
686 		if (uvm_page_physget_freelist(paddrp, i) == TRUE)
687 			return (TRUE);
688 	return (FALSE);
689 }
690 #endif /* PMAP_STEAL_MEMORY */
691 
692 /*
693  * uvm_page_physload: load physical memory into VM system
694  *
695  * => all args are PFs
696  * => all pages in start/end get vm_page structures
697  * => areas marked by avail_start/avail_end get added to the free page pool
698  * => we are limited to VM_PHYSSEG_MAX physical memory segments
699  */
700 
701 void
702 uvm_page_physload(start, end, avail_start, avail_end, free_list)
703 	paddr_t start, end, avail_start, avail_end;
704 	int free_list;
705 {
706 	int preload, lcv;
707 	psize_t npages;
708 	struct vm_page *pgs;
709 	struct vm_physseg *ps;
710 
711 	if (uvmexp.pagesize == 0)
712 		panic("uvm_page_physload: page size not set!");
713 	if (free_list >= VM_NFREELIST || free_list < VM_FREELIST_DEFAULT)
714 		panic("uvm_page_physload: bad free list %d", free_list);
715 	if (start >= end)
716 		panic("uvm_page_physload: start >= end");
717 
718 	/*
719 	 * do we have room?
720 	 */
721 
722 	if (vm_nphysseg == VM_PHYSSEG_MAX) {
723 		printf("uvm_page_physload: unable to load physical memory "
724 		    "segment\n");
725 		printf("\t%d segments allocated, ignoring 0x%llx -> 0x%llx\n",
726 		    VM_PHYSSEG_MAX, (long long)start, (long long)end);
727 		printf("\tincrease VM_PHYSSEG_MAX\n");
728 		return;
729 	}
730 
731 	/*
732 	 * check to see if this is a "preload" (i.e. uvm_mem_init hasn't been
733 	 * called yet, so malloc is not available).
734 	 */
735 
736 	for (lcv = 0 ; lcv < vm_nphysseg ; lcv++) {
737 		if (vm_physmem[lcv].pgs)
738 			break;
739 	}
740 	preload = (lcv == vm_nphysseg);
741 
742 	/*
743 	 * if VM is already running, attempt to malloc() vm_page structures
744 	 */
745 
746 	if (!preload) {
747 #if defined(VM_PHYSSEG_NOADD)
748 		panic("uvm_page_physload: tried to add RAM after vm_mem_init");
749 #else
750 		/* XXXCDC: need some sort of lockout for this case */
751 		paddr_t paddr;
752 		npages = end - start;  /* # of pages */
753 		pgs = malloc(sizeof(struct vm_page) * npages,
754 		    M_VMPAGE, M_NOWAIT);
755 		if (pgs == NULL) {
756 			printf("uvm_page_physload: can not malloc vm_page "
757 			    "structs for segment\n");
758 			printf("\tignoring 0x%lx -> 0x%lx\n", start, end);
759 			return;
760 		}
761 		/* zero data, init phys_addr and free_list, and free pages */
762 		memset(pgs, 0, sizeof(struct vm_page) * npages);
763 		for (lcv = 0, paddr = ptoa(start) ;
764 				 lcv < npages ; lcv++, paddr += PAGE_SIZE) {
765 			pgs[lcv].phys_addr = paddr;
766 			pgs[lcv].free_list = free_list;
767 			if (atop(paddr) >= avail_start &&
768 			    atop(paddr) <= avail_end)
769 				uvm_pagefree(&pgs[lcv]);
770 		}
771 		/* XXXCDC: incomplete: need to update uvmexp.free, what else? */
772 		/* XXXCDC: need hook to tell pmap to rebuild pv_list, etc... */
773 #endif
774 	} else {
775 		pgs = NULL;
776 		npages = 0;
777 	}
778 
779 	/*
780 	 * now insert us in the proper place in vm_physmem[]
781 	 */
782 
783 #if (VM_PHYSSEG_STRAT == VM_PSTRAT_RANDOM)
784 	/* random: put it at the end (easy!) */
785 	ps = &vm_physmem[vm_nphysseg];
786 #elif (VM_PHYSSEG_STRAT == VM_PSTRAT_BSEARCH)
787 	{
788 		int x;
789 		/* sort by address for binary search */
790 		for (lcv = 0 ; lcv < vm_nphysseg ; lcv++)
791 			if (start < vm_physmem[lcv].start)
792 				break;
793 		ps = &vm_physmem[lcv];
794 		/* move back other entries, if necessary ... */
795 		for (x = vm_nphysseg ; x > lcv ; x--)
796 			/* structure copy */
797 			vm_physmem[x] = vm_physmem[x - 1];
798 	}
799 #elif (VM_PHYSSEG_STRAT == VM_PSTRAT_BIGFIRST)
800 	{
801 		int x;
802 		/* sort by largest segment first */
803 		for (lcv = 0 ; lcv < vm_nphysseg ; lcv++)
804 			if ((end - start) >
805 			    (vm_physmem[lcv].end - vm_physmem[lcv].start))
806 				break;
807 		ps = &vm_physmem[lcv];
808 		/* move back other entries, if necessary ... */
809 		for (x = vm_nphysseg ; x > lcv ; x--)
810 			/* structure copy */
811 			vm_physmem[x] = vm_physmem[x - 1];
812 	}
813 #else
814 	panic("uvm_page_physload: unknown physseg strategy selected!");
815 #endif
816 
817 	ps->start = start;
818 	ps->end = end;
819 	ps->avail_start = avail_start;
820 	ps->avail_end = avail_end;
821 	if (preload) {
822 		ps->pgs = NULL;
823 	} else {
824 		ps->pgs = pgs;
825 		ps->lastpg = pgs + npages - 1;
826 	}
827 	ps->free_list = free_list;
828 	vm_nphysseg++;
829 
830 	if (!preload)
831 		uvm_page_rehash();
832 }
833 
834 /*
835  * uvm_page_rehash: reallocate hash table based on number of free pages.
836  */
837 
838 void
839 uvm_page_rehash()
840 {
841 	int freepages, lcv, bucketcount, oldcount;
842 	struct pglist *newbuckets, *oldbuckets;
843 	struct vm_page *pg;
844 	size_t newsize, oldsize;
845 
846 	/*
847 	 * compute number of pages that can go in the free pool
848 	 */
849 
850 	freepages = 0;
851 	for (lcv = 0 ; lcv < vm_nphysseg ; lcv++)
852 		freepages +=
853 		    (vm_physmem[lcv].avail_end - vm_physmem[lcv].avail_start);
854 
855 	/*
856 	 * compute number of buckets needed for this number of pages
857 	 */
858 
859 	bucketcount = 1;
860 	while (bucketcount < freepages)
861 		bucketcount = bucketcount * 2;
862 
863 	/*
864 	 * compute the size of the current table and new table.
865 	 */
866 
867 	oldbuckets = uvm.page_hash;
868 	oldcount = uvm.page_nhash;
869 	oldsize = round_page(sizeof(struct pglist) * oldcount);
870 	newsize = round_page(sizeof(struct pglist) * bucketcount);
871 
872 	/*
873 	 * allocate the new buckets
874 	 */
875 
876 	newbuckets = (struct pglist *) uvm_km_alloc(kernel_map, newsize);
877 	if (newbuckets == NULL) {
878 		printf("uvm_page_physrehash: WARNING: could not grow page "
879 		    "hash table\n");
880 		return;
881 	}
882 	for (lcv = 0 ; lcv < bucketcount ; lcv++)
883 		TAILQ_INIT(&newbuckets[lcv]);
884 
885 	/*
886 	 * now replace the old buckets with the new ones and rehash everything
887 	 */
888 
889 	simple_lock(&uvm.hashlock);
890 	uvm.page_hash = newbuckets;
891 	uvm.page_nhash = bucketcount;
892 	uvm.page_hashmask = bucketcount - 1;  /* power of 2 */
893 
894 	/* ... and rehash */
895 	for (lcv = 0 ; lcv < oldcount ; lcv++) {
896 		while ((pg = oldbuckets[lcv].tqh_first) != NULL) {
897 			TAILQ_REMOVE(&oldbuckets[lcv], pg, hashq);
898 			TAILQ_INSERT_TAIL(
899 			  &uvm.page_hash[uvm_pagehash(pg->uobject, pg->offset)],
900 			  pg, hashq);
901 		}
902 	}
903 	simple_unlock(&uvm.hashlock);
904 
905 	/*
906 	 * free old bucket array if is not the boot-time table
907 	 */
908 
909 	if (oldbuckets != &uvm_bootbucket)
910 		uvm_km_free(kernel_map, (vaddr_t) oldbuckets, oldsize);
911 }
912 
913 /*
914  * uvm_page_recolor: Recolor the pages if the new bucket count is
915  * larger than the old one.
916  */
917 
918 void
919 uvm_page_recolor(int newncolors)
920 {
921 	struct pgflbucket *bucketarray, *oldbucketarray;
922 	struct pgfreelist pgfl;
923 	struct vm_page *pg;
924 	vsize_t bucketcount;
925 	int s, lcv, color, i, ocolors;
926 
927 	if (newncolors <= uvmexp.ncolors)
928 		return;
929 
930 	if (uvm.page_init_done == FALSE) {
931 		uvmexp.ncolors = newncolors;
932 		return;
933 	}
934 
935 	bucketcount = newncolors * VM_NFREELIST;
936 	bucketarray = malloc(bucketcount * sizeof(struct pgflbucket),
937 	    M_VMPAGE, M_NOWAIT);
938 	if (bucketarray == NULL) {
939 		printf("WARNING: unable to allocate %ld page color buckets\n",
940 		    (long) bucketcount);
941 		return;
942 	}
943 
944 	s = uvm_lock_fpageq();
945 
946 	/* Make sure we should still do this. */
947 	if (newncolors <= uvmexp.ncolors) {
948 		uvm_unlock_fpageq(s);
949 		free(bucketarray, M_VMPAGE);
950 		return;
951 	}
952 
953 	oldbucketarray = uvm.page_free[0].pgfl_buckets;
954 	ocolors = uvmexp.ncolors;
955 
956 	uvmexp.ncolors = newncolors;
957 	uvmexp.colormask = uvmexp.ncolors - 1;
958 
959 	for (lcv = 0; lcv < VM_NFREELIST; lcv++) {
960 		pgfl.pgfl_buckets = (bucketarray + (lcv * newncolors));
961 		uvm_page_init_buckets(&pgfl);
962 		for (color = 0; color < ocolors; color++) {
963 			for (i = 0; i < PGFL_NQUEUES; i++) {
964 				while ((pg = TAILQ_FIRST(&uvm.page_free[
965 				    lcv].pgfl_buckets[color].pgfl_queues[i]))
966 				    != NULL) {
967 					TAILQ_REMOVE(&uvm.page_free[
968 					    lcv].pgfl_buckets[
969 					    color].pgfl_queues[i], pg, pageq);
970 					TAILQ_INSERT_TAIL(&pgfl.pgfl_buckets[
971 					    VM_PGCOLOR_BUCKET(pg)].pgfl_queues[
972 					    i], pg, pageq);
973 				}
974 			}
975 		}
976 		uvm.page_free[lcv].pgfl_buckets = pgfl.pgfl_buckets;
977 	}
978 
979 	if (have_recolored_pages) {
980 		uvm_unlock_fpageq(s);
981 		free(oldbucketarray, M_VMPAGE);
982 		return;
983 	}
984 
985 	have_recolored_pages = TRUE;
986 	uvm_unlock_fpageq(s);
987 }
988 
989 /*
990  * uvm_pagealloc_pgfl: helper routine for uvm_pagealloc_strat
991  */
992 
993 static __inline struct vm_page *
994 uvm_pagealloc_pgfl(struct pgfreelist *pgfl, int try1, int try2,
995     int *trycolorp)
996 {
997 	struct pglist *freeq;
998 	struct vm_page *pg;
999 	int color, trycolor = *trycolorp;
1000 
1001 	color = trycolor;
1002 	do {
1003 		if ((pg = TAILQ_FIRST((freeq =
1004 		    &pgfl->pgfl_buckets[color].pgfl_queues[try1]))) != NULL)
1005 			goto gotit;
1006 		if ((pg = TAILQ_FIRST((freeq =
1007 		    &pgfl->pgfl_buckets[color].pgfl_queues[try2]))) != NULL)
1008 			goto gotit;
1009 		color = (color + 1) & uvmexp.colormask;
1010 	} while (color != trycolor);
1011 
1012 	return (NULL);
1013 
1014  gotit:
1015 	TAILQ_REMOVE(freeq, pg, pageq);
1016 	uvmexp.free--;
1017 
1018 	/* update zero'd page count */
1019 	if (pg->flags & PG_ZERO)
1020 		uvmexp.zeropages--;
1021 
1022 	if (color == trycolor)
1023 		uvmexp.colorhit++;
1024 	else {
1025 		uvmexp.colormiss++;
1026 		*trycolorp = color;
1027 	}
1028 
1029 	return (pg);
1030 }
1031 
1032 /*
1033  * uvm_pagealloc_strat: allocate vm_page from a particular free list.
1034  *
1035  * => return null if no pages free
1036  * => wake up pagedaemon if number of free pages drops below low water mark
1037  * => if obj != NULL, obj must be locked (to put in hash)
1038  * => if anon != NULL, anon must be locked (to put in anon)
1039  * => only one of obj or anon can be non-null
1040  * => caller must activate/deactivate page if it is not wired.
1041  * => free_list is ignored if strat == UVM_PGA_STRAT_NORMAL.
1042  * => policy decision: it is more important to pull a page off of the
1043  *	appropriate priority free list than it is to get a zero'd or
1044  *	unknown contents page.  This is because we live with the
1045  *	consequences of a bad free list decision for the entire
1046  *	lifetime of the page, e.g. if the page comes from memory that
1047  *	is slower to access.
1048  */
1049 
1050 struct vm_page *
1051 uvm_pagealloc_strat(obj, off, anon, flags, strat, free_list)
1052 	struct uvm_object *obj;
1053 	voff_t off;
1054 	int flags;
1055 	struct vm_anon *anon;
1056 	int strat, free_list;
1057 {
1058 	int lcv, try1, try2, s, zeroit = 0, color;
1059 	struct vm_page *pg;
1060 	boolean_t use_reserve;
1061 
1062 	KASSERT(obj == NULL || anon == NULL);
1063 	KASSERT(off == trunc_page(off));
1064 	LOCK_ASSERT(obj == NULL || simple_lock_held(&obj->vmobjlock));
1065 	LOCK_ASSERT(anon == NULL || simple_lock_held(&anon->an_lock));
1066 
1067 	s = uvm_lock_fpageq();
1068 
1069 	/*
1070 	 * This implements a global round-robin page coloring
1071 	 * algorithm.
1072 	 *
1073 	 * XXXJRT: Should we make the `nextcolor' per-CPU?
1074 	 * XXXJRT: What about virtually-indexed caches?
1075 	 */
1076 
1077 	color = uvm.page_free_nextcolor;
1078 
1079 	/*
1080 	 * check to see if we need to generate some free pages waking
1081 	 * the pagedaemon.
1082 	 */
1083 
1084 	UVM_KICK_PDAEMON();
1085 
1086 	/*
1087 	 * fail if any of these conditions is true:
1088 	 * [1]  there really are no free pages, or
1089 	 * [2]  only kernel "reserved" pages remain and
1090 	 *        the page isn't being allocated to a kernel object.
1091 	 * [3]  only pagedaemon "reserved" pages remain and
1092 	 *        the requestor isn't the pagedaemon.
1093 	 */
1094 
1095 	use_reserve = (flags & UVM_PGA_USERESERVE) ||
1096 		(obj && UVM_OBJ_IS_KERN_OBJECT(obj));
1097 	if ((uvmexp.free <= uvmexp.reserve_kernel && !use_reserve) ||
1098 	    (uvmexp.free <= uvmexp.reserve_pagedaemon &&
1099 	     !(use_reserve && curproc == uvm.pagedaemon_proc)))
1100 		goto fail;
1101 
1102 #if PGFL_NQUEUES != 2
1103 #error uvm_pagealloc_strat needs to be updated
1104 #endif
1105 
1106 	/*
1107 	 * If we want a zero'd page, try the ZEROS queue first, otherwise
1108 	 * we try the UNKNOWN queue first.
1109 	 */
1110 	if (flags & UVM_PGA_ZERO) {
1111 		try1 = PGFL_ZEROS;
1112 		try2 = PGFL_UNKNOWN;
1113 	} else {
1114 		try1 = PGFL_UNKNOWN;
1115 		try2 = PGFL_ZEROS;
1116 	}
1117 
1118  again:
1119 	switch (strat) {
1120 	case UVM_PGA_STRAT_NORMAL:
1121 		/* Check all freelists in descending priority order. */
1122 		for (lcv = 0; lcv < VM_NFREELIST; lcv++) {
1123 			pg = uvm_pagealloc_pgfl(&uvm.page_free[lcv],
1124 			    try1, try2, &color);
1125 			if (pg != NULL)
1126 				goto gotit;
1127 		}
1128 
1129 		/* No pages free! */
1130 		goto fail;
1131 
1132 	case UVM_PGA_STRAT_ONLY:
1133 	case UVM_PGA_STRAT_FALLBACK:
1134 		/* Attempt to allocate from the specified free list. */
1135 		KASSERT(free_list >= 0 && free_list < VM_NFREELIST);
1136 		pg = uvm_pagealloc_pgfl(&uvm.page_free[free_list],
1137 		    try1, try2, &color);
1138 		if (pg != NULL)
1139 			goto gotit;
1140 
1141 		/* Fall back, if possible. */
1142 		if (strat == UVM_PGA_STRAT_FALLBACK) {
1143 			strat = UVM_PGA_STRAT_NORMAL;
1144 			goto again;
1145 		}
1146 
1147 		/* No pages free! */
1148 		goto fail;
1149 
1150 	default:
1151 		panic("uvm_pagealloc_strat: bad strat %d", strat);
1152 		/* NOTREACHED */
1153 	}
1154 
1155  gotit:
1156 	/*
1157 	 * We now know which color we actually allocated from; set
1158 	 * the next color accordingly.
1159 	 */
1160 
1161 	uvm.page_free_nextcolor = (color + 1) & uvmexp.colormask;
1162 
1163 	/*
1164 	 * update allocation statistics and remember if we have to
1165 	 * zero the page
1166 	 */
1167 
1168 	if (flags & UVM_PGA_ZERO) {
1169 		if (pg->flags & PG_ZERO) {
1170 			uvmexp.pga_zerohit++;
1171 			zeroit = 0;
1172 		} else {
1173 			uvmexp.pga_zeromiss++;
1174 			zeroit = 1;
1175 		}
1176 	}
1177 	uvm_unlock_fpageq(s);
1178 
1179 	pg->offset = off;
1180 	pg->uobject = obj;
1181 	pg->uanon = anon;
1182 	pg->flags = PG_BUSY|PG_CLEAN|PG_FAKE;
1183 	if (anon) {
1184 		anon->u.an_page = pg;
1185 		pg->pqflags = PQ_ANON;
1186 		uvmexp.anonpages++;
1187 	} else {
1188 		if (obj) {
1189 			uvm_pageinsert(pg);
1190 		}
1191 		pg->pqflags = 0;
1192 	}
1193 #if defined(UVM_PAGE_TRKOWN)
1194 	pg->owner_tag = NULL;
1195 #endif
1196 	UVM_PAGE_OWN(pg, "new alloc");
1197 
1198 	if (flags & UVM_PGA_ZERO) {
1199 		/*
1200 		 * A zero'd page is not clean.  If we got a page not already
1201 		 * zero'd, then we have to zero it ourselves.
1202 		 */
1203 		pg->flags &= ~PG_CLEAN;
1204 		if (zeroit)
1205 			pmap_zero_page(VM_PAGE_TO_PHYS(pg));
1206 	}
1207 
1208 	return(pg);
1209 
1210  fail:
1211 	uvm_unlock_fpageq(s);
1212 	return (NULL);
1213 }
1214 
1215 /*
1216  * uvm_pagereplace: replace a page with another
1217  *
1218  * => object must be locked
1219  */
1220 
1221 void
1222 uvm_pagereplace(oldpg, newpg)
1223 	struct vm_page *oldpg;
1224 	struct vm_page *newpg;
1225 {
1226 
1227 	KASSERT((oldpg->flags & PG_TABLED) != 0);
1228 	KASSERT(oldpg->uobject != NULL);
1229 	KASSERT((newpg->flags & PG_TABLED) == 0);
1230 	KASSERT(newpg->uobject == NULL);
1231 	LOCK_ASSERT(simple_lock_held(&oldpg->uobject->vmobjlock));
1232 
1233 	newpg->uobject = oldpg->uobject;
1234 	newpg->offset = oldpg->offset;
1235 
1236 	uvm_pageinsert_after(newpg, oldpg);
1237 	uvm_pageremove(oldpg);
1238 }
1239 
1240 /*
1241  * uvm_pagerealloc: reallocate a page from one object to another
1242  *
1243  * => both objects must be locked
1244  */
1245 
1246 void
1247 uvm_pagerealloc(pg, newobj, newoff)
1248 	struct vm_page *pg;
1249 	struct uvm_object *newobj;
1250 	voff_t newoff;
1251 {
1252 	/*
1253 	 * remove it from the old object
1254 	 */
1255 
1256 	if (pg->uobject) {
1257 		uvm_pageremove(pg);
1258 	}
1259 
1260 	/*
1261 	 * put it in the new object
1262 	 */
1263 
1264 	if (newobj) {
1265 		pg->uobject = newobj;
1266 		pg->offset = newoff;
1267 		uvm_pageinsert(pg);
1268 	}
1269 }
1270 
1271 #ifdef DEBUG
1272 /*
1273  * check if page is zero-filled
1274  *
1275  *  - called with free page queue lock held.
1276  */
1277 void
1278 uvm_pagezerocheck(struct vm_page *pg)
1279 {
1280 	int *p, *ep;
1281 
1282 	KASSERT(uvm_zerocheckkva != 0);
1283 	LOCK_ASSERT(simple_lock_held(&uvm.fpageqlock));
1284 
1285 	/*
1286 	 * XXX assuming pmap_kenter_pa and pmap_kremove never call
1287 	 * uvm page allocator.
1288 	 *
1289 	 * it might be better to have "CPU-local temporary map" pmap interface.
1290 	 */
1291 	pmap_kenter_pa(uvm_zerocheckkva, VM_PAGE_TO_PHYS(pg), VM_PROT_READ);
1292 	p = (int *)uvm_zerocheckkva;
1293 	ep = (int *)((char *)p + PAGE_SIZE);
1294 	pmap_update(pmap_kernel());
1295 	while (p < ep) {
1296 		if (*p != 0)
1297 			panic("PG_ZERO page isn't zero-filled");
1298 		p++;
1299 	}
1300 	pmap_kremove(uvm_zerocheckkva, PAGE_SIZE);
1301 }
1302 #endif /* DEBUG */
1303 
1304 /*
1305  * uvm_pagefree: free page
1306  *
1307  * => erase page's identity (i.e. remove from hash/object)
1308  * => put page on free list
1309  * => caller must lock owning object (either anon or uvm_object)
1310  * => caller must lock page queues
1311  * => assumes all valid mappings of pg are gone
1312  */
1313 
1314 void
1315 uvm_pagefree(pg)
1316 	struct vm_page *pg;
1317 {
1318 	int s;
1319 	struct pglist *pgfl;
1320 	boolean_t iszero;
1321 
1322 	KASSERT((pg->flags & PG_PAGEOUT) == 0);
1323 	LOCK_ASSERT(simple_lock_held(&uvm.pageqlock) ||
1324 		    (pg->pqflags & (PQ_ACTIVE|PQ_INACTIVE)) == 0);
1325 	LOCK_ASSERT(pg->uobject == NULL ||
1326 		    simple_lock_held(&pg->uobject->vmobjlock));
1327 	LOCK_ASSERT(pg->uobject != NULL || pg->uanon == NULL ||
1328 		    simple_lock_held(&pg->uanon->an_lock));
1329 
1330 #ifdef DEBUG
1331 	if (pg->uobject == (void *)0xdeadbeef &&
1332 	    pg->uanon == (void *)0xdeadbeef) {
1333 		panic("uvm_pagefree: freeing free page %p", pg);
1334 	}
1335 #endif /* DEBUG */
1336 
1337 	/*
1338 	 * if the page is loaned, resolve the loan instead of freeing.
1339 	 */
1340 
1341 	if (pg->loan_count) {
1342 		KASSERT(pg->wire_count == 0);
1343 
1344 		/*
1345 		 * if the page is owned by an anon then we just want to
1346 		 * drop anon ownership.  the kernel will free the page when
1347 		 * it is done with it.  if the page is owned by an object,
1348 		 * remove it from the object and mark it dirty for the benefit
1349 		 * of possible anon owners.
1350 		 *
1351 		 * regardless of previous ownership, wakeup any waiters,
1352 		 * unbusy the page, and we're done.
1353 		 */
1354 
1355 		if (pg->uobject != NULL) {
1356 			uvm_pageremove(pg);
1357 			pg->flags &= ~PG_CLEAN;
1358 		} else if (pg->uanon != NULL) {
1359 			if ((pg->pqflags & PQ_ANON) == 0) {
1360 				pg->loan_count--;
1361 			} else {
1362 				pg->pqflags &= ~PQ_ANON;
1363 				uvmexp.anonpages--;
1364 			}
1365 			pg->uanon->u.an_page = NULL;
1366 			pg->uanon = NULL;
1367 		}
1368 		if (pg->flags & PG_WANTED) {
1369 			wakeup(pg);
1370 		}
1371 		pg->flags &= ~(PG_WANTED|PG_BUSY|PG_RELEASED|PG_PAGER1);
1372 #ifdef UVM_PAGE_TRKOWN
1373 		pg->owner_tag = NULL;
1374 #endif
1375 		if (pg->loan_count) {
1376 			uvm_pagedequeue(pg);
1377 			return;
1378 		}
1379 	}
1380 
1381 	/*
1382 	 * remove page from its object or anon.
1383 	 */
1384 
1385 	if (pg->uobject != NULL) {
1386 		uvm_pageremove(pg);
1387 	} else if (pg->uanon != NULL) {
1388 		pg->uanon->u.an_page = NULL;
1389 		uvmexp.anonpages--;
1390 	}
1391 
1392 	/*
1393 	 * now remove the page from the queues.
1394 	 */
1395 
1396 	uvm_pagedequeue(pg);
1397 
1398 	/*
1399 	 * if the page was wired, unwire it now.
1400 	 */
1401 
1402 	if (pg->wire_count) {
1403 		pg->wire_count = 0;
1404 		uvmexp.wired--;
1405 	}
1406 
1407 	/*
1408 	 * and put on free queue
1409 	 */
1410 
1411 	iszero = (pg->flags & PG_ZERO);
1412 	pgfl = &uvm.page_free[uvm_page_lookup_freelist(pg)].
1413 	    pgfl_buckets[VM_PGCOLOR_BUCKET(pg)].
1414 	    pgfl_queues[iszero ? PGFL_ZEROS : PGFL_UNKNOWN];
1415 
1416 	pg->pqflags = PQ_FREE;
1417 #ifdef DEBUG
1418 	pg->uobject = (void *)0xdeadbeef;
1419 	pg->offset = 0xdeadbeef;
1420 	pg->uanon = (void *)0xdeadbeef;
1421 #endif
1422 
1423 	s = uvm_lock_fpageq();
1424 
1425 #ifdef DEBUG
1426 	if (iszero)
1427 		uvm_pagezerocheck(pg);
1428 #endif /* DEBUG */
1429 
1430 	TAILQ_INSERT_HEAD(pgfl, pg, pageq);
1431 	uvmexp.free++;
1432 	if (iszero)
1433 		uvmexp.zeropages++;
1434 
1435 	if (uvmexp.zeropages < UVM_PAGEZERO_TARGET)
1436 		uvm.page_idle_zero = vm_page_zero_enable;
1437 
1438 	uvm_unlock_fpageq(s);
1439 }
1440 
1441 /*
1442  * uvm_page_unbusy: unbusy an array of pages.
1443  *
1444  * => pages must either all belong to the same object, or all belong to anons.
1445  * => if pages are object-owned, object must be locked.
1446  * => if pages are anon-owned, anons must be locked.
1447  * => caller must lock page queues if pages may be released.
1448  * => caller must make sure that anon-owned pages are not PG_RELEASED.
1449  */
1450 
1451 void
1452 uvm_page_unbusy(pgs, npgs)
1453 	struct vm_page **pgs;
1454 	int npgs;
1455 {
1456 	struct vm_page *pg;
1457 	int i;
1458 	UVMHIST_FUNC("uvm_page_unbusy"); UVMHIST_CALLED(ubchist);
1459 
1460 	for (i = 0; i < npgs; i++) {
1461 		pg = pgs[i];
1462 		if (pg == NULL || pg == PGO_DONTCARE) {
1463 			continue;
1464 		}
1465 
1466 		LOCK_ASSERT(pg->uobject == NULL ||
1467 		    simple_lock_held(&pg->uobject->vmobjlock));
1468 		LOCK_ASSERT(pg->uobject != NULL ||
1469 		    (pg->uanon != NULL &&
1470 		    simple_lock_held(&pg->uanon->an_lock)));
1471 
1472 		KASSERT(pg->flags & PG_BUSY);
1473 		KASSERT((pg->flags & PG_PAGEOUT) == 0);
1474 		if (pg->flags & PG_WANTED) {
1475 			wakeup(pg);
1476 		}
1477 		if (pg->flags & PG_RELEASED) {
1478 			UVMHIST_LOG(ubchist, "releasing pg %p", pg,0,0,0);
1479 			KASSERT(pg->uobject != NULL ||
1480 			    (pg->uanon != NULL && pg->uanon->an_ref > 0));
1481 			pg->flags &= ~PG_RELEASED;
1482 			uvm_pagefree(pg);
1483 		} else {
1484 			UVMHIST_LOG(ubchist, "unbusying pg %p", pg,0,0,0);
1485 			pg->flags &= ~(PG_WANTED|PG_BUSY);
1486 			UVM_PAGE_OWN(pg, NULL);
1487 		}
1488 	}
1489 }
1490 
1491 #if defined(UVM_PAGE_TRKOWN)
1492 /*
1493  * uvm_page_own: set or release page ownership
1494  *
1495  * => this is a debugging function that keeps track of who sets PG_BUSY
1496  *	and where they do it.   it can be used to track down problems
1497  *	such a process setting "PG_BUSY" and never releasing it.
1498  * => page's object [if any] must be locked
1499  * => if "tag" is NULL then we are releasing page ownership
1500  */
1501 void
1502 uvm_page_own(pg, tag)
1503 	struct vm_page *pg;
1504 	char *tag;
1505 {
1506 	KASSERT((pg->flags & (PG_PAGEOUT|PG_RELEASED)) == 0);
1507 
1508 	/* gain ownership? */
1509 	if (tag) {
1510 		if (pg->owner_tag) {
1511 			printf("uvm_page_own: page %p already owned "
1512 			    "by proc %d [%s]\n", pg,
1513 			    pg->owner, pg->owner_tag);
1514 			panic("uvm_page_own");
1515 		}
1516 		pg->owner = (curproc) ? curproc->p_pid :  (pid_t) -1;
1517 		pg->owner_tag = tag;
1518 		return;
1519 	}
1520 
1521 	/* drop ownership */
1522 	if (pg->owner_tag == NULL) {
1523 		printf("uvm_page_own: dropping ownership of an non-owned "
1524 		    "page (%p)\n", pg);
1525 		panic("uvm_page_own");
1526 	}
1527 	KASSERT((pg->pqflags & (PQ_ACTIVE|PQ_INACTIVE)) ||
1528 	    (pg->uanon == NULL && pg->uobject == NULL) ||
1529 	    pg->uobject == uvm.kernel_object ||
1530 	    pg->wire_count > 0 ||
1531 	    (pg->loan_count == 1 && pg->uanon == NULL) ||
1532 	    pg->loan_count > 1);
1533 	pg->owner_tag = NULL;
1534 }
1535 #endif
1536 
1537 /*
1538  * uvm_pageidlezero: zero free pages while the system is idle.
1539  *
1540  * => try to complete one color bucket at a time, to reduce our impact
1541  *	on the CPU cache.
1542  * => we loop until we either reach the target or whichqs indicates that
1543  *	there is a process ready to run.
1544  */
1545 void
1546 uvm_pageidlezero()
1547 {
1548 	struct vm_page *pg;
1549 	struct pgfreelist *pgfl;
1550 	int free_list, s, firstbucket;
1551 	static int nextbucket;
1552 
1553 	KERNEL_LOCK(LK_EXCLUSIVE | LK_CANRECURSE);
1554 	s = uvm_lock_fpageq();
1555 	firstbucket = nextbucket;
1556 	do {
1557 		if (sched_whichqs != 0)
1558 			goto quit;
1559 		if (uvmexp.zeropages >= UVM_PAGEZERO_TARGET) {
1560 			uvm.page_idle_zero = FALSE;
1561 			goto quit;
1562 		}
1563 		for (free_list = 0; free_list < VM_NFREELIST; free_list++) {
1564 			pgfl = &uvm.page_free[free_list];
1565 			while ((pg = TAILQ_FIRST(&pgfl->pgfl_buckets[
1566 			    nextbucket].pgfl_queues[PGFL_UNKNOWN])) != NULL) {
1567 				if (sched_whichqs != 0)
1568 					goto quit;
1569 
1570 				TAILQ_REMOVE(&pgfl->pgfl_buckets[
1571 				    nextbucket].pgfl_queues[PGFL_UNKNOWN],
1572 				    pg, pageq);
1573 				uvmexp.free--;
1574 				uvm_unlock_fpageq(s);
1575 				KERNEL_UNLOCK();
1576 #ifdef PMAP_PAGEIDLEZERO
1577 				if (!PMAP_PAGEIDLEZERO(VM_PAGE_TO_PHYS(pg))) {
1578 
1579 					/*
1580 					 * The machine-dependent code detected
1581 					 * some reason for us to abort zeroing
1582 					 * pages, probably because there is a
1583 					 * process now ready to run.
1584 					 */
1585 
1586 					KERNEL_LOCK(
1587 					    LK_EXCLUSIVE | LK_CANRECURSE);
1588 					s = uvm_lock_fpageq();
1589 					TAILQ_INSERT_HEAD(&pgfl->pgfl_buckets[
1590 					    nextbucket].pgfl_queues[
1591 					    PGFL_UNKNOWN], pg, pageq);
1592 					uvmexp.free++;
1593 					uvmexp.zeroaborts++;
1594 					goto quit;
1595 				}
1596 #else
1597 				pmap_zero_page(VM_PAGE_TO_PHYS(pg));
1598 #endif /* PMAP_PAGEIDLEZERO */
1599 				pg->flags |= PG_ZERO;
1600 
1601 				KERNEL_LOCK(LK_EXCLUSIVE | LK_CANRECURSE);
1602 				s = uvm_lock_fpageq();
1603 				TAILQ_INSERT_HEAD(&pgfl->pgfl_buckets[
1604 				    nextbucket].pgfl_queues[PGFL_ZEROS],
1605 				    pg, pageq);
1606 				uvmexp.free++;
1607 				uvmexp.zeropages++;
1608 			}
1609 		}
1610 		nextbucket = (nextbucket + 1) & uvmexp.colormask;
1611 	} while (nextbucket != firstbucket);
1612 quit:
1613 	uvm_unlock_fpageq(s);
1614 	KERNEL_UNLOCK();
1615 }
1616