xref: /netbsd-src/sys/uvm/uvm_page.c (revision 4472dbe5e3bd91ef2540bada7a7ca7384627ff9b)
1 /*	$NetBSD: uvm_page.c,v 1.36 2000/05/29 19:25:56 thorpej 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/param.h>
74 #include <sys/systm.h>
75 #include <sys/malloc.h>
76 #include <sys/sched.h>
77 
78 #include <vm/vm.h>
79 #include <vm/vm_page.h>
80 #include <vm/vm_kern.h>
81 
82 #define UVM_PAGE                /* pull in uvm_page.h functions */
83 #include <uvm/uvm.h>
84 
85 /*
86  * global vars... XXXCDC: move to uvm. structure.
87  */
88 
89 /*
90  * physical memory config is stored in vm_physmem.
91  */
92 
93 struct vm_physseg vm_physmem[VM_PHYSSEG_MAX];	/* XXXCDC: uvm.physmem */
94 int vm_nphysseg = 0;				/* XXXCDC: uvm.nphysseg */
95 
96 /*
97  * Some supported CPUs in a given architecture don't support all
98  * of the things necessary to do idle page zero'ing efficiently.
99  * We therefore provide a way to disable it from machdep code here.
100  */
101 
102 boolean_t vm_page_zero_enable = TRUE;
103 
104 /*
105  * local variables
106  */
107 
108 /*
109  * these variables record the values returned by vm_page_bootstrap,
110  * for debugging purposes.  The implementation of uvm_pageboot_alloc
111  * and pmap_startup here also uses them internally.
112  */
113 
114 static vaddr_t      virtual_space_start;
115 static vaddr_t      virtual_space_end;
116 
117 /*
118  * we use a hash table with only one bucket during bootup.  we will
119  * later rehash (resize) the hash table once the allocator is ready.
120  * we static allocate the one bootstrap bucket below...
121  */
122 
123 static struct pglist uvm_bootbucket;
124 
125 /*
126  * local prototypes
127  */
128 
129 static void uvm_pageinsert __P((struct vm_page *));
130 
131 
132 /*
133  * inline functions
134  */
135 
136 /*
137  * uvm_pageinsert: insert a page in the object and the hash table
138  *
139  * => caller must lock object
140  * => caller must lock page queues
141  * => call should have already set pg's object and offset pointers
142  *    and bumped the version counter
143  */
144 
145 __inline static void
146 uvm_pageinsert(pg)
147 	struct vm_page *pg;
148 {
149 	struct pglist *buck;
150 	int s;
151 
152 #ifdef DIAGNOSTIC
153 	if (pg->flags & PG_TABLED)
154 		panic("uvm_pageinsert: already inserted");
155 #endif
156 
157 	buck = &uvm.page_hash[uvm_pagehash(pg->uobject,pg->offset)];
158 	s = splimp();
159 	simple_lock(&uvm.hashlock);
160 	TAILQ_INSERT_TAIL(buck, pg, hashq);	/* put in hash */
161 	simple_unlock(&uvm.hashlock);
162 	splx(s);
163 
164 	TAILQ_INSERT_TAIL(&pg->uobject->memq, pg, listq); /* put in object */
165 	pg->flags |= PG_TABLED;
166 	pg->uobject->uo_npages++;
167 
168 }
169 
170 /*
171  * uvm_page_remove: remove page from object and hash
172  *
173  * => caller must lock object
174  * => caller must lock page queues
175  */
176 
177 void __inline
178 uvm_pageremove(pg)
179 	struct vm_page *pg;
180 {
181 	struct pglist *buck;
182 	int s;
183 
184 #ifdef DIAGNOSTIC
185 	if ((pg->flags & (PG_FAULTING)) != 0)
186 		panic("uvm_pageremove: page is faulting");
187 #endif
188 
189 	if ((pg->flags & PG_TABLED) == 0)
190 		return;				/* XXX: log */
191 
192 	buck = &uvm.page_hash[uvm_pagehash(pg->uobject,pg->offset)];
193 	s = splimp();
194 	simple_lock(&uvm.hashlock);
195 	TAILQ_REMOVE(buck, pg, hashq);
196 	simple_unlock(&uvm.hashlock);
197 	splx(s);
198 
199 	/* object should be locked */
200 	TAILQ_REMOVE(&pg->uobject->memq, pg, listq);
201 
202 	pg->flags &= ~PG_TABLED;
203 	pg->uobject->uo_npages--;
204 	pg->uobject = NULL;
205 	pg->version++;
206 
207 }
208 
209 /*
210  * uvm_page_init: init the page system.   called from uvm_init().
211  *
212  * => we return the range of kernel virtual memory in kvm_startp/kvm_endp
213  */
214 
215 void
216 uvm_page_init(kvm_startp, kvm_endp)
217 	vaddr_t *kvm_startp, *kvm_endp;
218 {
219 	vsize_t freepages, pagecount, n;
220 	vm_page_t pagearray;
221 	int lcv, i;
222 	paddr_t paddr;
223 
224 
225 	/*
226 	 * step 1: init the page queues and page queue locks
227 	 */
228 	for (lcv = 0; lcv < VM_NFREELIST; lcv++) {
229 		for (i = 0; i < PGFL_NQUEUES; i++)
230 			TAILQ_INIT(&uvm.page_free[lcv].pgfl_queues[i]);
231 	}
232 	TAILQ_INIT(&uvm.page_active);
233 	TAILQ_INIT(&uvm.page_inactive_swp);
234 	TAILQ_INIT(&uvm.page_inactive_obj);
235 	simple_lock_init(&uvm.pageqlock);
236 	simple_lock_init(&uvm.fpageqlock);
237 
238 	/*
239 	 * step 2: init the <obj,offset> => <page> hash table. for now
240 	 * we just have one bucket (the bootstrap bucket).   later on we
241 	 * will allocate new buckets as we dynamically resize the hash table.
242 	 */
243 
244 	uvm.page_nhash = 1;			/* 1 bucket */
245 	uvm.page_hashmask = 0;		/* mask for hash function */
246 	uvm.page_hash = &uvm_bootbucket;	/* install bootstrap bucket */
247 	TAILQ_INIT(uvm.page_hash);		/* init hash table */
248 	simple_lock_init(&uvm.hashlock);	/* init hash table lock */
249 
250 	/*
251 	 * step 3: allocate vm_page structures.
252 	 */
253 
254 	/*
255 	 * sanity check:
256 	 * before calling this function the MD code is expected to register
257 	 * some free RAM with the uvm_page_physload() function.   our job
258 	 * now is to allocate vm_page structures for this memory.
259 	 */
260 
261 	if (vm_nphysseg == 0)
262 		panic("vm_page_bootstrap: no memory pre-allocated");
263 
264 	/*
265 	 * first calculate the number of free pages...
266 	 *
267 	 * note that we use start/end rather than avail_start/avail_end.
268 	 * this allows us to allocate extra vm_page structures in case we
269 	 * want to return some memory to the pool after booting.
270 	 */
271 
272 	freepages = 0;
273 	for (lcv = 0 ; lcv < vm_nphysseg ; lcv++)
274 		freepages += (vm_physmem[lcv].end - vm_physmem[lcv].start);
275 
276 	/*
277 	 * we now know we have (PAGE_SIZE * freepages) bytes of memory we can
278 	 * use.   for each page of memory we use we need a vm_page structure.
279 	 * thus, the total number of pages we can use is the total size of
280 	 * the memory divided by the PAGE_SIZE plus the size of the vm_page
281 	 * structure.   we add one to freepages as a fudge factor to avoid
282 	 * truncation errors (since we can only allocate in terms of whole
283 	 * pages).
284 	 */
285 
286 	pagecount = ((freepages + 1) << PAGE_SHIFT) /
287 	    (PAGE_SIZE + sizeof(struct vm_page));
288 	pagearray = (vm_page_t)uvm_pageboot_alloc(pagecount *
289 	    sizeof(struct vm_page));
290 	memset(pagearray, 0, pagecount * sizeof(struct vm_page));
291 
292 	/*
293 	 * step 4: init the vm_page structures and put them in the correct
294 	 * place...
295 	 */
296 
297 	for (lcv = 0 ; lcv < vm_nphysseg ; lcv++) {
298 
299 		n = vm_physmem[lcv].end - vm_physmem[lcv].start;
300 		if (n > pagecount) {
301 			printf("uvm_page_init: lost %ld page(s) in init\n",
302 			    (long)(n - pagecount));
303 			panic("uvm_page_init");  /* XXXCDC: shouldn't happen? */
304 			/* n = pagecount; */
305 		}
306 		/* set up page array pointers */
307 		vm_physmem[lcv].pgs = pagearray;
308 		pagearray += n;
309 		pagecount -= n;
310 		vm_physmem[lcv].lastpg = vm_physmem[lcv].pgs + (n - 1);
311 
312 		/* init and free vm_pages (we've already zeroed them) */
313 		paddr = ptoa(vm_physmem[lcv].start);
314 		for (i = 0 ; i < n ; i++, paddr += PAGE_SIZE) {
315 			vm_physmem[lcv].pgs[i].phys_addr = paddr;
316 			if (atop(paddr) >= vm_physmem[lcv].avail_start &&
317 			    atop(paddr) <= vm_physmem[lcv].avail_end) {
318 				uvmexp.npages++;
319 				/* add page to free pool */
320 				uvm_pagefree(&vm_physmem[lcv].pgs[i]);
321 			}
322 		}
323 	}
324 	/*
325 	 * step 5: pass up the values of virtual_space_start and
326 	 * virtual_space_end (obtained by uvm_pageboot_alloc) to the upper
327 	 * layers of the VM.
328 	 */
329 
330 	*kvm_startp = round_page(virtual_space_start);
331 	*kvm_endp = trunc_page(virtual_space_end);
332 
333 	/*
334 	 * step 6: init pagedaemon lock
335 	 */
336 
337 	simple_lock_init(&uvm.pagedaemon_lock);
338 
339 	/*
340 	 * step 7: init reserve thresholds
341 	 * XXXCDC - values may need adjusting
342 	 */
343 	uvmexp.reserve_pagedaemon = 1;
344 	uvmexp.reserve_kernel = 5;
345 
346 	/*
347 	 * step 8: determine if we should zero pages in the idle
348 	 * loop.
349 	 *
350 	 * XXXJRT - might consider zero'ing up to the target *now*,
351 	 *	    but that could take an awfully long time if you
352 	 *	    have a lot of memory.
353 	 */
354 	uvm.page_idle_zero = vm_page_zero_enable;
355 
356 	/*
357 	 * done!
358 	 */
359 
360 	uvm.page_init_done = TRUE;
361 }
362 
363 /*
364  * uvm_setpagesize: set the page size
365  *
366  * => sets page_shift and page_mask from uvmexp.pagesize.
367  * => XXXCDC: move global vars.
368  */
369 
370 void
371 uvm_setpagesize()
372 {
373 	if (uvmexp.pagesize == 0)
374 		uvmexp.pagesize = DEFAULT_PAGE_SIZE;
375 	uvmexp.pagemask = uvmexp.pagesize - 1;
376 	if ((uvmexp.pagemask & uvmexp.pagesize) != 0)
377 		panic("uvm_setpagesize: page size not a power of two");
378 	for (uvmexp.pageshift = 0; ; uvmexp.pageshift++)
379 		if ((1 << uvmexp.pageshift) == uvmexp.pagesize)
380 			break;
381 }
382 
383 /*
384  * uvm_pageboot_alloc: steal memory from physmem for bootstrapping
385  */
386 
387 vaddr_t
388 uvm_pageboot_alloc(size)
389 	vsize_t size;
390 {
391 #if defined(PMAP_STEAL_MEMORY)
392 	vaddr_t addr;
393 
394 	/*
395 	 * defer bootstrap allocation to MD code (it may want to allocate
396 	 * from a direct-mapped segment).  pmap_steal_memory should round
397 	 * off virtual_space_start/virtual_space_end.
398 	 */
399 
400 	addr = pmap_steal_memory(size, &virtual_space_start,
401 	    &virtual_space_end);
402 
403 	return(addr);
404 
405 #else /* !PMAP_STEAL_MEMORY */
406 
407 	static boolean_t initialized = FALSE;
408 	vaddr_t addr, vaddr;
409 	paddr_t paddr;
410 
411 	/* round to page size */
412 	size = round_page(size);
413 
414 	/*
415 	 * on first call to this function, initialize ourselves.
416 	 */
417 	if (initialized == FALSE) {
418 		pmap_virtual_space(&virtual_space_start, &virtual_space_end);
419 
420 		/* round it the way we like it */
421 		virtual_space_start = round_page(virtual_space_start);
422 		virtual_space_end = trunc_page(virtual_space_end);
423 
424 		initialized = TRUE;
425 	}
426 
427 	/*
428 	 * allocate virtual memory for this request
429 	 */
430 	if (virtual_space_start == virtual_space_end ||
431 	    (virtual_space_end - virtual_space_start) < size)
432 		panic("uvm_pageboot_alloc: out of virtual space");
433 
434 	addr = virtual_space_start;
435 
436 #ifdef PMAP_GROWKERNEL
437 	/*
438 	 * If the kernel pmap can't map the requested space,
439 	 * then allocate more resources for it.
440 	 */
441 	if (uvm_maxkaddr < (addr + size)) {
442 		uvm_maxkaddr = pmap_growkernel(addr + size);
443 		if (uvm_maxkaddr < (addr + size))
444 			panic("uvm_pageboot_alloc: pmap_growkernel() failed");
445 	}
446 #endif
447 
448 	virtual_space_start += size;
449 
450 	/*
451 	 * allocate and mapin physical pages to back new virtual pages
452 	 */
453 
454 	for (vaddr = round_page(addr) ; vaddr < addr + size ;
455 	    vaddr += PAGE_SIZE) {
456 
457 		if (!uvm_page_physget(&paddr))
458 			panic("uvm_pageboot_alloc: out of memory");
459 
460 		/*
461 		 * Note this memory is no longer managed, so using
462 		 * pmap_kenter is safe.
463 		 */
464 		pmap_kenter_pa(vaddr, paddr, VM_PROT_READ|VM_PROT_WRITE);
465 	}
466 	return(addr);
467 #endif	/* PMAP_STEAL_MEMORY */
468 }
469 
470 #if !defined(PMAP_STEAL_MEMORY)
471 /*
472  * uvm_page_physget: "steal" one page from the vm_physmem structure.
473  *
474  * => attempt to allocate it off the end of a segment in which the "avail"
475  *    values match the start/end values.   if we can't do that, then we
476  *    will advance both values (making them equal, and removing some
477  *    vm_page structures from the non-avail area).
478  * => return false if out of memory.
479  */
480 
481 /* subroutine: try to allocate from memory chunks on the specified freelist */
482 static boolean_t uvm_page_physget_freelist __P((paddr_t *, int));
483 
484 static boolean_t
485 uvm_page_physget_freelist(paddrp, freelist)
486 	paddr_t *paddrp;
487 	int freelist;
488 {
489 	int lcv, x;
490 
491 	/* pass 1: try allocating from a matching end */
492 #if (VM_PHYSSEG_STRAT == VM_PSTRAT_BIGFIRST)
493 	for (lcv = vm_nphysseg - 1 ; lcv >= 0 ; lcv--)
494 #else
495 	for (lcv = 0 ; lcv < vm_nphysseg ; lcv++)
496 #endif
497 	{
498 
499 		if (uvm.page_init_done == TRUE)
500 			panic("vm_page_physget: called _after_ bootstrap");
501 
502 		if (vm_physmem[lcv].free_list != freelist)
503 			continue;
504 
505 		/* try from front */
506 		if (vm_physmem[lcv].avail_start == vm_physmem[lcv].start &&
507 		    vm_physmem[lcv].avail_start < vm_physmem[lcv].avail_end) {
508 			*paddrp = ptoa(vm_physmem[lcv].avail_start);
509 			vm_physmem[lcv].avail_start++;
510 			vm_physmem[lcv].start++;
511 			/* nothing left?   nuke it */
512 			if (vm_physmem[lcv].avail_start ==
513 			    vm_physmem[lcv].end) {
514 				if (vm_nphysseg == 1)
515 				    panic("vm_page_physget: out of memory!");
516 				vm_nphysseg--;
517 				for (x = lcv ; x < vm_nphysseg ; x++)
518 					/* structure copy */
519 					vm_physmem[x] = vm_physmem[x+1];
520 			}
521 			return (TRUE);
522 		}
523 
524 		/* try from rear */
525 		if (vm_physmem[lcv].avail_end == vm_physmem[lcv].end &&
526 		    vm_physmem[lcv].avail_start < vm_physmem[lcv].avail_end) {
527 			*paddrp = ptoa(vm_physmem[lcv].avail_end - 1);
528 			vm_physmem[lcv].avail_end--;
529 			vm_physmem[lcv].end--;
530 			/* nothing left?   nuke it */
531 			if (vm_physmem[lcv].avail_end ==
532 			    vm_physmem[lcv].start) {
533 				if (vm_nphysseg == 1)
534 				    panic("vm_page_physget: out of memory!");
535 				vm_nphysseg--;
536 				for (x = lcv ; x < vm_nphysseg ; x++)
537 					/* structure copy */
538 					vm_physmem[x] = vm_physmem[x+1];
539 			}
540 			return (TRUE);
541 		}
542 	}
543 
544 	/* pass2: forget about matching ends, just allocate something */
545 #if (VM_PHYSSEG_STRAT == VM_PSTRAT_BIGFIRST)
546 	for (lcv = vm_nphysseg - 1 ; lcv >= 0 ; lcv--)
547 #else
548 	for (lcv = 0 ; lcv < vm_nphysseg ; lcv++)
549 #endif
550 	{
551 
552 		/* any room in this bank? */
553 		if (vm_physmem[lcv].avail_start >= vm_physmem[lcv].avail_end)
554 			continue;  /* nope */
555 
556 		*paddrp = ptoa(vm_physmem[lcv].avail_start);
557 		vm_physmem[lcv].avail_start++;
558 		/* truncate! */
559 		vm_physmem[lcv].start = vm_physmem[lcv].avail_start;
560 
561 		/* nothing left?   nuke it */
562 		if (vm_physmem[lcv].avail_start == vm_physmem[lcv].end) {
563 			if (vm_nphysseg == 1)
564 				panic("vm_page_physget: out of memory!");
565 			vm_nphysseg--;
566 			for (x = lcv ; x < vm_nphysseg ; x++)
567 				/* structure copy */
568 				vm_physmem[x] = vm_physmem[x+1];
569 		}
570 		return (TRUE);
571 	}
572 
573 	return (FALSE);        /* whoops! */
574 }
575 
576 boolean_t
577 uvm_page_physget(paddrp)
578 	paddr_t *paddrp;
579 {
580 	int i;
581 
582 	/* try in the order of freelist preference */
583 	for (i = 0; i < VM_NFREELIST; i++)
584 		if (uvm_page_physget_freelist(paddrp, i) == TRUE)
585 			return (TRUE);
586 	return (FALSE);
587 }
588 #endif /* PMAP_STEAL_MEMORY */
589 
590 /*
591  * uvm_page_physload: load physical memory into VM system
592  *
593  * => all args are PFs
594  * => all pages in start/end get vm_page structures
595  * => areas marked by avail_start/avail_end get added to the free page pool
596  * => we are limited to VM_PHYSSEG_MAX physical memory segments
597  */
598 
599 void
600 uvm_page_physload(start, end, avail_start, avail_end, free_list)
601 	paddr_t start, end, avail_start, avail_end;
602 	int free_list;
603 {
604 	int preload, lcv;
605 	psize_t npages;
606 	struct vm_page *pgs;
607 	struct vm_physseg *ps;
608 
609 	if (uvmexp.pagesize == 0)
610 		panic("vm_page_physload: page size not set!");
611 
612 	if (free_list >= VM_NFREELIST || free_list < VM_FREELIST_DEFAULT)
613 		panic("uvm_page_physload: bad free list %d\n", free_list);
614 
615 	if (start >= end)
616 		panic("uvm_page_physload: start >= end");
617 
618 	/*
619 	 * do we have room?
620 	 */
621 	if (vm_nphysseg == VM_PHYSSEG_MAX) {
622 		printf("vm_page_physload: unable to load physical memory "
623 		    "segment\n");
624 		printf("\t%d segments allocated, ignoring 0x%lx -> 0x%lx\n",
625 		    VM_PHYSSEG_MAX, start, end);
626 		return;
627 	}
628 
629 	/*
630 	 * check to see if this is a "preload" (i.e. uvm_mem_init hasn't been
631 	 * called yet, so malloc is not available).
632 	 */
633 	for (lcv = 0 ; lcv < vm_nphysseg ; lcv++) {
634 		if (vm_physmem[lcv].pgs)
635 			break;
636 	}
637 	preload = (lcv == vm_nphysseg);
638 
639 	/*
640 	 * if VM is already running, attempt to malloc() vm_page structures
641 	 */
642 	if (!preload) {
643 #if defined(VM_PHYSSEG_NOADD)
644 		panic("vm_page_physload: tried to add RAM after vm_mem_init");
645 #else
646 		/* XXXCDC: need some sort of lockout for this case */
647 		paddr_t paddr;
648 		npages = end - start;  /* # of pages */
649 		MALLOC(pgs, struct vm_page *, sizeof(struct vm_page) * npages,
650 					 M_VMPAGE, M_NOWAIT);
651 		if (pgs == NULL) {
652 			printf("vm_page_physload: can not malloc vm_page "
653 			    "structs for segment\n");
654 			printf("\tignoring 0x%lx -> 0x%lx\n", start, end);
655 			return;
656 		}
657 		/* zero data, init phys_addr and free_list, and free pages */
658 		memset(pgs, 0, sizeof(struct vm_page) * npages);
659 		for (lcv = 0, paddr = ptoa(start) ;
660 				 lcv < npages ; lcv++, paddr += PAGE_SIZE) {
661 			pgs[lcv].phys_addr = paddr;
662 			pgs[lcv].free_list = free_list;
663 			if (atop(paddr) >= avail_start &&
664 			    atop(paddr) <= avail_end)
665 				uvm_pagefree(&pgs[lcv]);
666 		}
667 		/* XXXCDC: incomplete: need to update uvmexp.free, what else? */
668 		/* XXXCDC: need hook to tell pmap to rebuild pv_list, etc... */
669 #endif
670 	} else {
671 
672 		/* gcc complains if these don't get init'd */
673 		pgs = NULL;
674 		npages = 0;
675 
676 	}
677 
678 	/*
679 	 * now insert us in the proper place in vm_physmem[]
680 	 */
681 
682 #if (VM_PHYSSEG_STRAT == VM_PSTRAT_RANDOM)
683 
684 	/* random: put it at the end (easy!) */
685 	ps = &vm_physmem[vm_nphysseg];
686 
687 #elif (VM_PHYSSEG_STRAT == VM_PSTRAT_BSEARCH)
688 
689 	{
690 		int x;
691 		/* sort by address for binary search */
692 		for (lcv = 0 ; lcv < vm_nphysseg ; lcv++)
693 			if (start < vm_physmem[lcv].start)
694 				break;
695 		ps = &vm_physmem[lcv];
696 		/* move back other entries, if necessary ... */
697 		for (x = vm_nphysseg ; x > lcv ; x--)
698 			/* structure copy */
699 			vm_physmem[x] = vm_physmem[x - 1];
700 	}
701 
702 #elif (VM_PHYSSEG_STRAT == VM_PSTRAT_BIGFIRST)
703 
704 	{
705 		int x;
706 		/* sort by largest segment first */
707 		for (lcv = 0 ; lcv < vm_nphysseg ; lcv++)
708 			if ((end - start) >
709 			    (vm_physmem[lcv].end - vm_physmem[lcv].start))
710 				break;
711 		ps = &vm_physmem[lcv];
712 		/* move back other entries, if necessary ... */
713 		for (x = vm_nphysseg ; x > lcv ; x--)
714 			/* structure copy */
715 			vm_physmem[x] = vm_physmem[x - 1];
716 	}
717 
718 #else
719 
720 	panic("vm_page_physload: unknown physseg strategy selected!");
721 
722 #endif
723 
724 	ps->start = start;
725 	ps->end = end;
726 	ps->avail_start = avail_start;
727 	ps->avail_end = avail_end;
728 	if (preload) {
729 		ps->pgs = NULL;
730 	} else {
731 		ps->pgs = pgs;
732 		ps->lastpg = pgs + npages - 1;
733 	}
734 	ps->free_list = free_list;
735 	vm_nphysseg++;
736 
737 	/*
738 	 * done!
739 	 */
740 
741 	if (!preload)
742 		uvm_page_rehash();
743 
744 	return;
745 }
746 
747 /*
748  * uvm_page_rehash: reallocate hash table based on number of free pages.
749  */
750 
751 void
752 uvm_page_rehash()
753 {
754 	int freepages, lcv, bucketcount, s, oldcount;
755 	struct pglist *newbuckets, *oldbuckets;
756 	struct vm_page *pg;
757 	size_t newsize, oldsize;
758 
759 	/*
760 	 * compute number of pages that can go in the free pool
761 	 */
762 
763 	freepages = 0;
764 	for (lcv = 0 ; lcv < vm_nphysseg ; lcv++)
765 		freepages +=
766 		    (vm_physmem[lcv].avail_end - vm_physmem[lcv].avail_start);
767 
768 	/*
769 	 * compute number of buckets needed for this number of pages
770 	 */
771 
772 	bucketcount = 1;
773 	while (bucketcount < freepages)
774 		bucketcount = bucketcount * 2;
775 
776 	/*
777 	 * compute the size of the current table and new table.
778 	 */
779 
780 	oldbuckets = uvm.page_hash;
781 	oldcount = uvm.page_nhash;
782 	oldsize = round_page(sizeof(struct pglist) * oldcount);
783 	newsize = round_page(sizeof(struct pglist) * bucketcount);
784 
785 	/*
786 	 * allocate the new buckets
787 	 */
788 
789 	newbuckets = (struct pglist *) uvm_km_alloc(kernel_map, newsize);
790 	if (newbuckets == NULL) {
791 		printf("uvm_page_physrehash: WARNING: could not grow page "
792 		    "hash table\n");
793 		return;
794 	}
795 	for (lcv = 0 ; lcv < bucketcount ; lcv++)
796 		TAILQ_INIT(&newbuckets[lcv]);
797 
798 	/*
799 	 * now replace the old buckets with the new ones and rehash everything
800 	 */
801 
802 	s = splimp();
803 	simple_lock(&uvm.hashlock);
804 	uvm.page_hash = newbuckets;
805 	uvm.page_nhash = bucketcount;
806 	uvm.page_hashmask = bucketcount - 1;  /* power of 2 */
807 
808 	/* ... and rehash */
809 	for (lcv = 0 ; lcv < oldcount ; lcv++) {
810 		while ((pg = oldbuckets[lcv].tqh_first) != NULL) {
811 			TAILQ_REMOVE(&oldbuckets[lcv], pg, hashq);
812 			TAILQ_INSERT_TAIL(
813 			  &uvm.page_hash[uvm_pagehash(pg->uobject, pg->offset)],
814 			  pg, hashq);
815 		}
816 	}
817 	simple_unlock(&uvm.hashlock);
818 	splx(s);
819 
820 	/*
821 	 * free old bucket array if is not the boot-time table
822 	 */
823 
824 	if (oldbuckets != &uvm_bootbucket)
825 		uvm_km_free(kernel_map, (vaddr_t) oldbuckets, oldsize);
826 
827 	/*
828 	 * done
829 	 */
830 	return;
831 }
832 
833 
834 #if 1 /* XXXCDC: TMP TMP TMP DEBUG DEBUG DEBUG */
835 
836 void uvm_page_physdump __P((void)); /* SHUT UP GCC */
837 
838 /* call from DDB */
839 void
840 uvm_page_physdump()
841 {
842 	int lcv;
843 
844 	printf("rehash: physical memory config [segs=%d of %d]:\n",
845 				 vm_nphysseg, VM_PHYSSEG_MAX);
846 	for (lcv = 0 ; lcv < vm_nphysseg ; lcv++)
847 		printf("0x%lx->0x%lx [0x%lx->0x%lx]\n", vm_physmem[lcv].start,
848 		    vm_physmem[lcv].end, vm_physmem[lcv].avail_start,
849 		    vm_physmem[lcv].avail_end);
850 	printf("STRATEGY = ");
851 	switch (VM_PHYSSEG_STRAT) {
852 	case VM_PSTRAT_RANDOM: printf("RANDOM\n"); break;
853 	case VM_PSTRAT_BSEARCH: printf("BSEARCH\n"); break;
854 	case VM_PSTRAT_BIGFIRST: printf("BIGFIRST\n"); break;
855 	default: printf("<<UNKNOWN>>!!!!\n");
856 	}
857 	printf("number of buckets = %d\n", uvm.page_nhash);
858 }
859 #endif
860 
861 /*
862  * uvm_pagealloc_strat: allocate vm_page from a particular free list.
863  *
864  * => return null if no pages free
865  * => wake up pagedaemon if number of free pages drops below low water mark
866  * => if obj != NULL, obj must be locked (to put in hash)
867  * => if anon != NULL, anon must be locked (to put in anon)
868  * => only one of obj or anon can be non-null
869  * => caller must activate/deactivate page if it is not wired.
870  * => free_list is ignored if strat == UVM_PGA_STRAT_NORMAL.
871  * => policy decision: it is more important to pull a page off of the
872  *	appropriate priority free list than it is to get a zero'd or
873  *	unknown contents page.  This is because we live with the
874  *	consequences of a bad free list decision for the entire
875  *	lifetime of the page, e.g. if the page comes from memory that
876  *	is slower to access.
877  */
878 
879 struct vm_page *
880 uvm_pagealloc_strat(obj, off, anon, flags, strat, free_list)
881 	struct uvm_object *obj;
882 	voff_t off;
883 	int flags;
884 	struct vm_anon *anon;
885 	int strat, free_list;
886 {
887 	int lcv, try1, try2, s, zeroit = 0;
888 	struct vm_page *pg;
889 	struct pglist *freeq;
890 	struct pgfreelist *pgfl;
891 	boolean_t use_reserve;
892 
893 #ifdef DIAGNOSTIC
894 	/* sanity check */
895 	if (obj && anon)
896 		panic("uvm_pagealloc: obj and anon != NULL");
897 #endif
898 
899 	s = uvm_lock_fpageq();		/* lock free page queue */
900 
901 	/*
902 	 * check to see if we need to generate some free pages waking
903 	 * the pagedaemon.
904 	 */
905 
906 	if (uvmexp.free < uvmexp.freemin || (uvmexp.free < uvmexp.freetarg &&
907 	    uvmexp.inactive < uvmexp.inactarg))
908 		wakeup(&uvm.pagedaemon);
909 
910 	/*
911 	 * fail if any of these conditions is true:
912 	 * [1]  there really are no free pages, or
913 	 * [2]  only kernel "reserved" pages remain and
914 	 *        the page isn't being allocated to a kernel object.
915 	 * [3]  only pagedaemon "reserved" pages remain and
916 	 *        the requestor isn't the pagedaemon.
917 	 */
918 
919 	use_reserve = (flags & UVM_PGA_USERESERVE) ||
920 		(obj && UVM_OBJ_IS_KERN_OBJECT(obj));
921 	if ((uvmexp.free <= uvmexp.reserve_kernel && !use_reserve) ||
922 	    (uvmexp.free <= uvmexp.reserve_pagedaemon &&
923 	     !(use_reserve && curproc == uvm.pagedaemon_proc)))
924 		goto fail;
925 
926 #if PGFL_NQUEUES != 2
927 #error uvm_pagealloc_strat needs to be updated
928 #endif
929 
930 	/*
931 	 * If we want a zero'd page, try the ZEROS queue first, otherwise
932 	 * we try the UNKNOWN queue first.
933 	 */
934 	if (flags & UVM_PGA_ZERO) {
935 		try1 = PGFL_ZEROS;
936 		try2 = PGFL_UNKNOWN;
937 	} else {
938 		try1 = PGFL_UNKNOWN;
939 		try2 = PGFL_ZEROS;
940 	}
941 
942  again:
943 	switch (strat) {
944 	case UVM_PGA_STRAT_NORMAL:
945 		/* Check all freelists in descending priority order. */
946 		for (lcv = 0; lcv < VM_NFREELIST; lcv++) {
947 			pgfl = &uvm.page_free[lcv];
948 			if ((pg = TAILQ_FIRST((freeq =
949 			      &pgfl->pgfl_queues[try1]))) != NULL ||
950 			    (pg = TAILQ_FIRST((freeq =
951 			      &pgfl->pgfl_queues[try2]))) != NULL)
952 				goto gotit;
953 		}
954 
955 		/* No pages free! */
956 		goto fail;
957 
958 	case UVM_PGA_STRAT_ONLY:
959 	case UVM_PGA_STRAT_FALLBACK:
960 		/* Attempt to allocate from the specified free list. */
961 #ifdef DIAGNOSTIC
962 		if (free_list >= VM_NFREELIST || free_list < 0)
963 			panic("uvm_pagealloc_strat: bad free list %d",
964 			    free_list);
965 #endif
966 		pgfl = &uvm.page_free[free_list];
967 		if ((pg = TAILQ_FIRST((freeq =
968 		      &pgfl->pgfl_queues[try1]))) != NULL ||
969 		    (pg = TAILQ_FIRST((freeq =
970 		      &pgfl->pgfl_queues[try2]))) != NULL)
971 			goto gotit;
972 
973 		/* Fall back, if possible. */
974 		if (strat == UVM_PGA_STRAT_FALLBACK) {
975 			strat = UVM_PGA_STRAT_NORMAL;
976 			goto again;
977 		}
978 
979 		/* No pages free! */
980 		goto fail;
981 
982 	default:
983 		panic("uvm_pagealloc_strat: bad strat %d", strat);
984 		/* NOTREACHED */
985 	}
986 
987  gotit:
988 	TAILQ_REMOVE(freeq, pg, pageq);
989 	uvmexp.free--;
990 
991 	/* update zero'd page count */
992 	if (pg->flags & PG_ZERO)
993 		uvmexp.zeropages--;
994 
995 	/*
996 	 * update allocation statistics and remember if we have to
997 	 * zero the page
998 	 */
999 	if (flags & UVM_PGA_ZERO) {
1000 		if (pg->flags & PG_ZERO) {
1001 			uvmexp.pga_zerohit++;
1002 			zeroit = 0;
1003 		} else {
1004 			uvmexp.pga_zeromiss++;
1005 			zeroit = 1;
1006 		}
1007 	}
1008 
1009 	uvm_unlock_fpageq(s);		/* unlock free page queue */
1010 
1011 	pg->offset = off;
1012 	pg->uobject = obj;
1013 	pg->uanon = anon;
1014 	pg->flags = PG_BUSY|PG_CLEAN|PG_FAKE;
1015 	pg->version++;
1016 	pg->wire_count = 0;
1017 	pg->loan_count = 0;
1018 	if (anon) {
1019 		anon->u.an_page = pg;
1020 		pg->pqflags = PQ_ANON;
1021 	} else {
1022 		if (obj)
1023 			uvm_pageinsert(pg);
1024 		pg->pqflags = 0;
1025 	}
1026 #if defined(UVM_PAGE_TRKOWN)
1027 	pg->owner_tag = NULL;
1028 #endif
1029 	UVM_PAGE_OWN(pg, "new alloc");
1030 
1031 	if (flags & UVM_PGA_ZERO) {
1032 		/*
1033 		 * A zero'd page is not clean.  If we got a page not already
1034 		 * zero'd, then we have to zero it ourselves.
1035 		 */
1036 		pg->flags &= ~PG_CLEAN;
1037 		if (zeroit)
1038 			pmap_zero_page(VM_PAGE_TO_PHYS(pg));
1039 	}
1040 
1041 	return(pg);
1042 
1043  fail:
1044 	uvm_unlock_fpageq(s);
1045 	return (NULL);
1046 }
1047 
1048 /*
1049  * uvm_pagerealloc: reallocate a page from one object to another
1050  *
1051  * => both objects must be locked
1052  */
1053 
1054 void
1055 uvm_pagerealloc(pg, newobj, newoff)
1056 	struct vm_page *pg;
1057 	struct uvm_object *newobj;
1058 	voff_t newoff;
1059 {
1060 	/*
1061 	 * remove it from the old object
1062 	 */
1063 
1064 	if (pg->uobject) {
1065 		uvm_pageremove(pg);
1066 	}
1067 
1068 	/*
1069 	 * put it in the new object
1070 	 */
1071 
1072 	if (newobj) {
1073 		pg->uobject = newobj;
1074 		pg->offset = newoff;
1075 		pg->version++;
1076 		uvm_pageinsert(pg);
1077 	}
1078 
1079 	return;
1080 }
1081 
1082 
1083 /*
1084  * uvm_pagefree: free page
1085  *
1086  * => erase page's identity (i.e. remove from hash/object)
1087  * => put page on free list
1088  * => caller must lock owning object (either anon or uvm_object)
1089  * => caller must lock page queues
1090  * => assumes all valid mappings of pg are gone
1091  */
1092 
1093 void uvm_pagefree(pg)
1094 
1095 struct vm_page *pg;
1096 
1097 {
1098 	int s;
1099 	int saved_loan_count = pg->loan_count;
1100 
1101 	/*
1102 	 * if the page was an object page (and thus "TABLED"), remove it
1103 	 * from the object.
1104 	 */
1105 
1106 	if (pg->flags & PG_TABLED) {
1107 
1108 		/*
1109 		 * if the object page is on loan we are going to drop ownership.
1110 		 * it is possible that an anon will take over as owner for this
1111 		 * page later on.   the anon will want a !PG_CLEAN page so that
1112 		 * it knows it needs to allocate swap if it wants to page the
1113 		 * page out.
1114 		 */
1115 
1116 		if (saved_loan_count)
1117 			pg->flags &= ~PG_CLEAN;	/* in case an anon takes over */
1118 
1119 		uvm_pageremove(pg);
1120 
1121 		/*
1122 		 * if our page was on loan, then we just lost control over it
1123 		 * (in fact, if it was loaned to an anon, the anon may have
1124 		 * already taken over ownership of the page by now and thus
1125 		 * changed the loan_count [e.g. in uvmfault_anonget()]) we just
1126 		 * return (when the last loan is dropped, then the page can be
1127 		 * freed by whatever was holding the last loan).
1128 		 */
1129 		if (saved_loan_count)
1130 			return;
1131 
1132 	} else if (saved_loan_count && (pg->pqflags & PQ_ANON)) {
1133 
1134 		/*
1135 		 * if our page is owned by an anon and is loaned out to the
1136 		 * kernel then we just want to drop ownership and return.
1137 		 * the kernel must free the page when all its loans clear ...
1138 		 * note that the kernel can't change the loan status of our
1139 		 * page as long as we are holding PQ lock.
1140 		 */
1141 		pg->pqflags &= ~PQ_ANON;
1142 		pg->uanon = NULL;
1143 		return;
1144 	}
1145 
1146 #ifdef DIAGNOSTIC
1147 	if (saved_loan_count) {
1148 		printf("uvm_pagefree: warning: freeing page with a loan "
1149 		    "count of %d\n", saved_loan_count);
1150 		panic("uvm_pagefree: loan count");
1151 	}
1152 #endif
1153 
1154 
1155 	/*
1156 	 * now remove the page from the queues
1157 	 */
1158 
1159 	if (pg->pqflags & PQ_ACTIVE) {
1160 		TAILQ_REMOVE(&uvm.page_active, pg, pageq);
1161 		pg->pqflags &= ~PQ_ACTIVE;
1162 		uvmexp.active--;
1163 	}
1164 	if (pg->pqflags & PQ_INACTIVE) {
1165 		if (pg->pqflags & PQ_SWAPBACKED)
1166 			TAILQ_REMOVE(&uvm.page_inactive_swp, pg, pageq);
1167 		else
1168 			TAILQ_REMOVE(&uvm.page_inactive_obj, pg, pageq);
1169 		pg->pqflags &= ~PQ_INACTIVE;
1170 		uvmexp.inactive--;
1171 	}
1172 
1173 	/*
1174 	 * if the page was wired, unwire it now.
1175 	 */
1176 	if (pg->wire_count) {
1177 		pg->wire_count = 0;
1178 		uvmexp.wired--;
1179 	}
1180 
1181 	/*
1182 	 * and put on free queue
1183 	 */
1184 
1185 	pg->flags &= ~PG_ZERO;
1186 
1187 	s = uvm_lock_fpageq();
1188 	TAILQ_INSERT_TAIL(&uvm.page_free[
1189 	    uvm_page_lookup_freelist(pg)].pgfl_queues[PGFL_UNKNOWN], pg, pageq);
1190 	pg->pqflags = PQ_FREE;
1191 #ifdef DEBUG
1192 	pg->uobject = (void *)0xdeadbeef;
1193 	pg->offset = 0xdeadbeef;
1194 	pg->uanon = (void *)0xdeadbeef;
1195 #endif
1196 	uvmexp.free++;
1197 
1198 	if (uvmexp.zeropages < UVM_PAGEZERO_TARGET)
1199 		uvm.page_idle_zero = vm_page_zero_enable;
1200 
1201 	uvm_unlock_fpageq(s);
1202 }
1203 
1204 #if defined(UVM_PAGE_TRKOWN)
1205 /*
1206  * uvm_page_own: set or release page ownership
1207  *
1208  * => this is a debugging function that keeps track of who sets PG_BUSY
1209  *	and where they do it.   it can be used to track down problems
1210  *	such a process setting "PG_BUSY" and never releasing it.
1211  * => page's object [if any] must be locked
1212  * => if "tag" is NULL then we are releasing page ownership
1213  */
1214 void
1215 uvm_page_own(pg, tag)
1216 	struct vm_page *pg;
1217 	char *tag;
1218 {
1219 	/* gain ownership? */
1220 	if (tag) {
1221 		if (pg->owner_tag) {
1222 			printf("uvm_page_own: page %p already owned "
1223 			    "by proc %d [%s]\n", pg,
1224 			     pg->owner, pg->owner_tag);
1225 			panic("uvm_page_own");
1226 		}
1227 		pg->owner = (curproc) ? curproc->p_pid :  (pid_t) -1;
1228 		pg->owner_tag = tag;
1229 		return;
1230 	}
1231 
1232 	/* drop ownership */
1233 	if (pg->owner_tag == NULL) {
1234 		printf("uvm_page_own: dropping ownership of an non-owned "
1235 		    "page (%p)\n", pg);
1236 		panic("uvm_page_own");
1237 	}
1238 	pg->owner_tag = NULL;
1239 	return;
1240 }
1241 #endif
1242 
1243 /*
1244  * uvm_pageidlezero: zero free pages while the system is idle.
1245  *
1246  * => we do at least one iteration per call, if we are below the target.
1247  * => we loop until we either reach the target or whichqs indicates that
1248  *	there is a process ready to run.
1249  */
1250 void
1251 uvm_pageidlezero()
1252 {
1253 	struct vm_page *pg;
1254 	struct pgfreelist *pgfl;
1255 	int free_list, s;
1256 
1257 	do {
1258 		s = uvm_lock_fpageq();
1259 
1260 		if (uvmexp.zeropages >= UVM_PAGEZERO_TARGET) {
1261 			uvm.page_idle_zero = FALSE;
1262 			uvm_unlock_fpageq(s);
1263 			return;
1264 		}
1265 
1266 		for (free_list = 0; free_list < VM_NFREELIST; free_list++) {
1267 			pgfl = &uvm.page_free[free_list];
1268 			if ((pg = TAILQ_FIRST(&pgfl->pgfl_queues[
1269 			    PGFL_UNKNOWN])) != NULL)
1270 				break;
1271 		}
1272 
1273 		if (pg == NULL) {
1274 			/*
1275 			 * No non-zero'd pages; don't bother trying again
1276 			 * until we know we have non-zero'd pages free.
1277 			 */
1278 			uvm.page_idle_zero = FALSE;
1279 			uvm_unlock_fpageq(s);
1280 			return;
1281 		}
1282 
1283 		TAILQ_REMOVE(&pgfl->pgfl_queues[PGFL_UNKNOWN], pg, pageq);
1284 		uvmexp.free--;
1285 		uvm_unlock_fpageq(s);
1286 
1287 #ifdef PMAP_PAGEIDLEZERO
1288 		PMAP_PAGEIDLEZERO(VM_PAGE_TO_PHYS(pg));
1289 #else
1290 		/*
1291 		 * XXX This will toast the cache unless the pmap_zero_page()
1292 		 * XXX implementation does uncached access.
1293 		 */
1294 		pmap_zero_page(VM_PAGE_TO_PHYS(pg));
1295 #endif
1296 		pg->flags |= PG_ZERO;
1297 
1298 		s = uvm_lock_fpageq();
1299 		TAILQ_INSERT_HEAD(&pgfl->pgfl_queues[PGFL_ZEROS], pg, pageq);
1300 		uvmexp.free++;
1301 		uvmexp.zeropages++;
1302 		uvm_unlock_fpageq(s);
1303 	} while (sched_whichqs == 0);
1304 }
1305