xref: /netbsd-src/sys/uvm/uvm_aobj.c (revision e77448e07be3174235c13f58032a0d6d0ab7638d)
1 /*	$NetBSD: uvm_aobj.c,v 1.100 2008/05/05 17:11:17 ad Exp $	*/
2 
3 /*
4  * Copyright (c) 1998 Chuck Silvers, Charles D. Cranor and
5  *                    Washington University.
6  * All rights reserved.
7  *
8  * Redistribution and use in source and binary forms, with or without
9  * modification, are permitted provided that the following conditions
10  * are met:
11  * 1. Redistributions of source code must retain the above copyright
12  *    notice, this list of conditions and the following disclaimer.
13  * 2. Redistributions in binary form must reproduce the above copyright
14  *    notice, this list of conditions and the following disclaimer in the
15  *    documentation and/or other materials provided with the distribution.
16  * 3. All advertising materials mentioning features or use of this software
17  *    must display the following acknowledgement:
18  *      This product includes software developed by Charles D. Cranor and
19  *      Washington University.
20  * 4. The name of the author may not be used to endorse or promote products
21  *    derived from this software without specific prior written permission.
22  *
23  * THIS SOFTWARE IS PROVIDED BY THE AUTHOR ``AS IS'' AND ANY EXPRESS OR
24  * IMPLIED WARRANTIES, INCLUDING, BUT NOT LIMITED TO, THE IMPLIED WARRANTIES
25  * OF MERCHANTABILITY AND FITNESS FOR A PARTICULAR PURPOSE ARE DISCLAIMED.
26  * IN NO EVENT SHALL THE AUTHOR BE LIABLE FOR ANY DIRECT, INDIRECT,
27  * INCIDENTAL, SPECIAL, EXEMPLARY, OR CONSEQUENTIAL DAMAGES (INCLUDING, BUT
28  * NOT LIMITED TO, PROCUREMENT OF SUBSTITUTE GOODS OR SERVICES; LOSS OF USE,
29  * DATA, OR PROFITS; OR BUSINESS INTERRUPTION) HOWEVER CAUSED AND ON ANY
30  * THEORY OF LIABILITY, WHETHER IN CONTRACT, STRICT LIABILITY, OR TORT
31  * (INCLUDING NEGLIGENCE OR OTHERWISE) ARISING IN ANY WAY OUT OF THE USE OF
32  * THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF SUCH DAMAGE.
33  *
34  * from: Id: uvm_aobj.c,v 1.1.2.5 1998/02/06 05:14:38 chs Exp
35  */
36 /*
37  * uvm_aobj.c: anonymous memory uvm_object pager
38  *
39  * author: Chuck Silvers <chuq@chuq.com>
40  * started: Jan-1998
41  *
42  * - design mostly from Chuck Cranor
43  */
44 
45 #include <sys/cdefs.h>
46 __KERNEL_RCSID(0, "$NetBSD: uvm_aobj.c,v 1.100 2008/05/05 17:11:17 ad Exp $");
47 
48 #include "opt_uvmhist.h"
49 
50 #include <sys/param.h>
51 #include <sys/systm.h>
52 #include <sys/proc.h>
53 #include <sys/malloc.h>
54 #include <sys/kernel.h>
55 #include <sys/pool.h>
56 
57 #include <uvm/uvm.h>
58 
59 /*
60  * an aobj manages anonymous-memory backed uvm_objects.   in addition
61  * to keeping the list of resident pages, it also keeps a list of
62  * allocated swap blocks.  depending on the size of the aobj this list
63  * of allocated swap blocks is either stored in an array (small objects)
64  * or in a hash table (large objects).
65  */
66 
67 /*
68  * local structures
69  */
70 
71 /*
72  * for hash tables, we break the address space of the aobj into blocks
73  * of UAO_SWHASH_CLUSTER_SIZE pages.   we require the cluster size to
74  * be a power of two.
75  */
76 
77 #define UAO_SWHASH_CLUSTER_SHIFT 4
78 #define UAO_SWHASH_CLUSTER_SIZE (1 << UAO_SWHASH_CLUSTER_SHIFT)
79 
80 /* get the "tag" for this page index */
81 #define UAO_SWHASH_ELT_TAG(PAGEIDX) \
82 	((PAGEIDX) >> UAO_SWHASH_CLUSTER_SHIFT)
83 
84 #define UAO_SWHASH_ELT_PAGESLOT_IDX(PAGEIDX) \
85 	((PAGEIDX) & (UAO_SWHASH_CLUSTER_SIZE - 1))
86 
87 /* given an ELT and a page index, find the swap slot */
88 #define UAO_SWHASH_ELT_PAGESLOT(ELT, PAGEIDX) \
89 	((ELT)->slots[UAO_SWHASH_ELT_PAGESLOT_IDX(PAGEIDX)])
90 
91 /* given an ELT, return its pageidx base */
92 #define UAO_SWHASH_ELT_PAGEIDX_BASE(ELT) \
93 	((ELT)->tag << UAO_SWHASH_CLUSTER_SHIFT)
94 
95 /*
96  * the swhash hash function
97  */
98 
99 #define UAO_SWHASH_HASH(AOBJ, PAGEIDX) \
100 	(&(AOBJ)->u_swhash[(((PAGEIDX) >> UAO_SWHASH_CLUSTER_SHIFT) \
101 			    & (AOBJ)->u_swhashmask)])
102 
103 /*
104  * the swhash threshhold determines if we will use an array or a
105  * hash table to store the list of allocated swap blocks.
106  */
107 
108 #define UAO_SWHASH_THRESHOLD (UAO_SWHASH_CLUSTER_SIZE * 4)
109 #define UAO_USES_SWHASH(AOBJ) \
110 	((AOBJ)->u_pages > UAO_SWHASH_THRESHOLD)	/* use hash? */
111 
112 /*
113  * the number of buckets in a swhash, with an upper bound
114  */
115 
116 #define UAO_SWHASH_MAXBUCKETS 256
117 #define UAO_SWHASH_BUCKETS(AOBJ) \
118 	(MIN((AOBJ)->u_pages >> UAO_SWHASH_CLUSTER_SHIFT, \
119 	     UAO_SWHASH_MAXBUCKETS))
120 
121 
122 /*
123  * uao_swhash_elt: when a hash table is being used, this structure defines
124  * the format of an entry in the bucket list.
125  */
126 
127 struct uao_swhash_elt {
128 	LIST_ENTRY(uao_swhash_elt) list;	/* the hash list */
129 	voff_t tag;				/* our 'tag' */
130 	int count;				/* our number of active slots */
131 	int slots[UAO_SWHASH_CLUSTER_SIZE];	/* the slots */
132 };
133 
134 /*
135  * uao_swhash: the swap hash table structure
136  */
137 
138 LIST_HEAD(uao_swhash, uao_swhash_elt);
139 
140 /*
141  * uao_swhash_elt_pool: pool of uao_swhash_elt structures
142  * NOTE: Pages for this pool must not come from a pageable kernel map!
143  */
144 POOL_INIT(uao_swhash_elt_pool, sizeof(struct uao_swhash_elt), 0, 0, 0,
145     "uaoeltpl", NULL, IPL_VM);
146 
147 /*
148  * uvm_aobj: the actual anon-backed uvm_object
149  *
150  * => the uvm_object is at the top of the structure, this allows
151  *   (struct uvm_aobj *) == (struct uvm_object *)
152  * => only one of u_swslots and u_swhash is used in any given aobj
153  */
154 
155 struct uvm_aobj {
156 	struct uvm_object u_obj; /* has: lock, pgops, memq, #pages, #refs */
157 	pgoff_t u_pages;	 /* number of pages in entire object */
158 	int u_flags;		 /* the flags (see uvm_aobj.h) */
159 	int *u_swslots;		 /* array of offset->swapslot mappings */
160 				 /*
161 				  * hashtable of offset->swapslot mappings
162 				  * (u_swhash is an array of bucket heads)
163 				  */
164 	struct uao_swhash *u_swhash;
165 	u_long u_swhashmask;		/* mask for hashtable */
166 	LIST_ENTRY(uvm_aobj) u_list;	/* global list of aobjs */
167 };
168 
169 /*
170  * uvm_aobj_pool: pool of uvm_aobj structures
171  */
172 POOL_INIT(uvm_aobj_pool, sizeof(struct uvm_aobj), 0, 0, 0, "aobjpl",
173     &pool_allocator_nointr, IPL_NONE);
174 
175 MALLOC_DEFINE(M_UVMAOBJ, "UVM aobj", "UVM aobj and related structures");
176 
177 /*
178  * local functions
179  */
180 
181 static void	uao_free(struct uvm_aobj *);
182 static int	uao_get(struct uvm_object *, voff_t, struct vm_page **,
183 		    int *, int, vm_prot_t, int, int);
184 static int	uao_put(struct uvm_object *, voff_t, voff_t, int);
185 
186 #if defined(VMSWAP)
187 static struct uao_swhash_elt *uao_find_swhash_elt
188     (struct uvm_aobj *, int, bool);
189 
190 static bool uao_pagein(struct uvm_aobj *, int, int);
191 static bool uao_pagein_page(struct uvm_aobj *, int);
192 static void uao_dropswap_range1(struct uvm_aobj *, voff_t, voff_t);
193 #endif /* defined(VMSWAP) */
194 
195 /*
196  * aobj_pager
197  *
198  * note that some functions (e.g. put) are handled elsewhere
199  */
200 
201 const struct uvm_pagerops aobj_pager = {
202 	.pgo_reference = uao_reference,
203 	.pgo_detach = uao_detach,
204 	.pgo_get = uao_get,
205 	.pgo_put = uao_put,
206 };
207 
208 /*
209  * uao_list: global list of active aobjs, locked by uao_list_lock
210  */
211 
212 static LIST_HEAD(aobjlist, uvm_aobj) uao_list;
213 static kmutex_t uao_list_lock;
214 
215 /*
216  * functions
217  */
218 
219 /*
220  * hash table/array related functions
221  */
222 
223 #if defined(VMSWAP)
224 
225 /*
226  * uao_hashinit: limited version of hashinit() that uses malloc(). XXX
227  */
228 static void *
229 uao_hashinit(u_int elements, int mflags, u_long *hashmask)
230 {
231 	LIST_HEAD(, generic) *elm, *emx;
232 	u_long hashsize;
233 	void *p;
234 
235 	for (hashsize = 1; hashsize < elements; hashsize <<= 1)
236 		continue;
237 	if ((p = malloc(hashsize * sizeof(*elm), M_UVMAOBJ, mflags)) == NULL)
238 		return (NULL);
239 	for (elm = p, emx = elm + hashsize; elm < emx; elm++)
240 		LIST_INIT(elm);
241 	*hashmask = hashsize - 1;
242 
243 	return (p);
244 }
245 
246 /*
247  * uao_find_swhash_elt: find (or create) a hash table entry for a page
248  * offset.
249  *
250  * => the object should be locked by the caller
251  */
252 
253 static struct uao_swhash_elt *
254 uao_find_swhash_elt(struct uvm_aobj *aobj, int pageidx, bool create)
255 {
256 	struct uao_swhash *swhash;
257 	struct uao_swhash_elt *elt;
258 	voff_t page_tag;
259 
260 	swhash = UAO_SWHASH_HASH(aobj, pageidx);
261 	page_tag = UAO_SWHASH_ELT_TAG(pageidx);
262 
263 	/*
264 	 * now search the bucket for the requested tag
265 	 */
266 
267 	LIST_FOREACH(elt, swhash, list) {
268 		if (elt->tag == page_tag) {
269 			return elt;
270 		}
271 	}
272 	if (!create) {
273 		return NULL;
274 	}
275 
276 	/*
277 	 * allocate a new entry for the bucket and init/insert it in
278 	 */
279 
280 	elt = pool_get(&uao_swhash_elt_pool, PR_NOWAIT);
281 	if (elt == NULL) {
282 		return NULL;
283 	}
284 	LIST_INSERT_HEAD(swhash, elt, list);
285 	elt->tag = page_tag;
286 	elt->count = 0;
287 	memset(elt->slots, 0, sizeof(elt->slots));
288 	return elt;
289 }
290 
291 /*
292  * uao_find_swslot: find the swap slot number for an aobj/pageidx
293  *
294  * => object must be locked by caller
295  */
296 
297 int
298 uao_find_swslot(struct uvm_object *uobj, int pageidx)
299 {
300 	struct uvm_aobj *aobj = (struct uvm_aobj *)uobj;
301 	struct uao_swhash_elt *elt;
302 
303 	/*
304 	 * if noswap flag is set, then we never return a slot
305 	 */
306 
307 	if (aobj->u_flags & UAO_FLAG_NOSWAP)
308 		return(0);
309 
310 	/*
311 	 * if hashing, look in hash table.
312 	 */
313 
314 	if (UAO_USES_SWHASH(aobj)) {
315 		elt = uao_find_swhash_elt(aobj, pageidx, false);
316 		if (elt)
317 			return(UAO_SWHASH_ELT_PAGESLOT(elt, pageidx));
318 		else
319 			return(0);
320 	}
321 
322 	/*
323 	 * otherwise, look in the array
324 	 */
325 
326 	return(aobj->u_swslots[pageidx]);
327 }
328 
329 /*
330  * uao_set_swslot: set the swap slot for a page in an aobj.
331  *
332  * => setting a slot to zero frees the slot
333  * => object must be locked by caller
334  * => we return the old slot number, or -1 if we failed to allocate
335  *    memory to record the new slot number
336  */
337 
338 int
339 uao_set_swslot(struct uvm_object *uobj, int pageidx, int slot)
340 {
341 	struct uvm_aobj *aobj = (struct uvm_aobj *)uobj;
342 	struct uao_swhash_elt *elt;
343 	int oldslot;
344 	UVMHIST_FUNC("uao_set_swslot"); UVMHIST_CALLED(pdhist);
345 	UVMHIST_LOG(pdhist, "aobj %p pageidx %d slot %d",
346 	    aobj, pageidx, slot, 0);
347 
348 	/*
349 	 * if noswap flag is set, then we can't set a non-zero slot.
350 	 */
351 
352 	if (aobj->u_flags & UAO_FLAG_NOSWAP) {
353 		if (slot == 0)
354 			return(0);
355 
356 		printf("uao_set_swslot: uobj = %p\n", uobj);
357 		panic("uao_set_swslot: NOSWAP object");
358 	}
359 
360 	/*
361 	 * are we using a hash table?  if so, add it in the hash.
362 	 */
363 
364 	if (UAO_USES_SWHASH(aobj)) {
365 
366 		/*
367 		 * Avoid allocating an entry just to free it again if
368 		 * the page had not swap slot in the first place, and
369 		 * we are freeing.
370 		 */
371 
372 		elt = uao_find_swhash_elt(aobj, pageidx, slot != 0);
373 		if (elt == NULL) {
374 			return slot ? -1 : 0;
375 		}
376 
377 		oldslot = UAO_SWHASH_ELT_PAGESLOT(elt, pageidx);
378 		UAO_SWHASH_ELT_PAGESLOT(elt, pageidx) = slot;
379 
380 		/*
381 		 * now adjust the elt's reference counter and free it if we've
382 		 * dropped it to zero.
383 		 */
384 
385 		if (slot) {
386 			if (oldslot == 0)
387 				elt->count++;
388 		} else {
389 			if (oldslot)
390 				elt->count--;
391 
392 			if (elt->count == 0) {
393 				LIST_REMOVE(elt, list);
394 				pool_put(&uao_swhash_elt_pool, elt);
395 			}
396 		}
397 	} else {
398 		/* we are using an array */
399 		oldslot = aobj->u_swslots[pageidx];
400 		aobj->u_swslots[pageidx] = slot;
401 	}
402 	return (oldslot);
403 }
404 
405 #endif /* defined(VMSWAP) */
406 
407 /*
408  * end of hash/array functions
409  */
410 
411 /*
412  * uao_free: free all resources held by an aobj, and then free the aobj
413  *
414  * => the aobj should be dead
415  */
416 
417 static void
418 uao_free(struct uvm_aobj *aobj)
419 {
420 	int swpgonlydelta = 0;
421 
422 
423 #if defined(VMSWAP)
424 	uao_dropswap_range1(aobj, 0, 0);
425 #endif /* defined(VMSWAP) */
426 
427 	mutex_exit(&aobj->u_obj.vmobjlock);
428 
429 #if defined(VMSWAP)
430 	if (UAO_USES_SWHASH(aobj)) {
431 
432 		/*
433 		 * free the hash table itself.
434 		 */
435 
436 		free(aobj->u_swhash, M_UVMAOBJ);
437 	} else {
438 
439 		/*
440 		 * free the array itsself.
441 		 */
442 
443 		free(aobj->u_swslots, M_UVMAOBJ);
444 	}
445 #endif /* defined(VMSWAP) */
446 
447 	/*
448 	 * finally free the aobj itself
449 	 */
450 
451 	UVM_OBJ_DESTROY(&aobj->u_obj);
452 	pool_put(&uvm_aobj_pool, aobj);
453 
454 	/*
455 	 * adjust the counter of pages only in swap for all
456 	 * the swap slots we've freed.
457 	 */
458 
459 	if (swpgonlydelta > 0) {
460 		mutex_enter(&uvm_swap_data_lock);
461 		KASSERT(uvmexp.swpgonly >= swpgonlydelta);
462 		uvmexp.swpgonly -= swpgonlydelta;
463 		mutex_exit(&uvm_swap_data_lock);
464 	}
465 }
466 
467 /*
468  * pager functions
469  */
470 
471 /*
472  * uao_create: create an aobj of the given size and return its uvm_object.
473  *
474  * => for normal use, flags are always zero
475  * => for the kernel object, the flags are:
476  *	UAO_FLAG_KERNOBJ - allocate the kernel object (can only happen once)
477  *	UAO_FLAG_KERNSWAP - enable swapping of kernel object ("           ")
478  */
479 
480 struct uvm_object *
481 uao_create(vsize_t size, int flags)
482 {
483 	static struct uvm_aobj kernel_object_store;
484 	static int kobj_alloced = 0;
485 	pgoff_t pages = round_page(size) >> PAGE_SHIFT;
486 	struct uvm_aobj *aobj;
487 	int refs;
488 
489 	/*
490 	 * malloc a new aobj unless we are asked for the kernel object
491 	 */
492 
493 	if (flags & UAO_FLAG_KERNOBJ) {
494 		KASSERT(!kobj_alloced);
495 		aobj = &kernel_object_store;
496 		aobj->u_pages = pages;
497 		aobj->u_flags = UAO_FLAG_NOSWAP;
498 		refs = UVM_OBJ_KERN;
499 		kobj_alloced = UAO_FLAG_KERNOBJ;
500 	} else if (flags & UAO_FLAG_KERNSWAP) {
501 		KASSERT(kobj_alloced == UAO_FLAG_KERNOBJ);
502 		aobj = &kernel_object_store;
503 		kobj_alloced = UAO_FLAG_KERNSWAP;
504 		refs = 0xdeadbeaf; /* XXX: gcc */
505 	} else {
506 		aobj = pool_get(&uvm_aobj_pool, PR_WAITOK);
507 		aobj->u_pages = pages;
508 		aobj->u_flags = 0;
509 		refs = 1;
510 	}
511 
512 	/*
513  	 * allocate hash/array if necessary
514  	 *
515  	 * note: in the KERNSWAP case no need to worry about locking since
516  	 * we are still booting we should be the only thread around.
517  	 */
518 
519 	if (flags == 0 || (flags & UAO_FLAG_KERNSWAP) != 0) {
520 #if defined(VMSWAP)
521 		int mflags = (flags & UAO_FLAG_KERNSWAP) != 0 ?
522 		    M_NOWAIT : M_WAITOK;
523 
524 		/* allocate hash table or array depending on object size */
525 		if (UAO_USES_SWHASH(aobj)) {
526 			aobj->u_swhash = uao_hashinit(UAO_SWHASH_BUCKETS(aobj),
527 			    mflags, &aobj->u_swhashmask);
528 			if (aobj->u_swhash == NULL)
529 				panic("uao_create: hashinit swhash failed");
530 		} else {
531 			aobj->u_swslots = malloc(pages * sizeof(int),
532 			    M_UVMAOBJ, mflags);
533 			if (aobj->u_swslots == NULL)
534 				panic("uao_create: malloc swslots failed");
535 			memset(aobj->u_swslots, 0, pages * sizeof(int));
536 		}
537 #endif /* defined(VMSWAP) */
538 
539 		if (flags) {
540 			aobj->u_flags &= ~UAO_FLAG_NOSWAP; /* clear noswap */
541 			return(&aobj->u_obj);
542 		}
543 	}
544 
545 	/*
546  	 * init aobj fields
547  	 */
548 
549 	UVM_OBJ_INIT(&aobj->u_obj, &aobj_pager, refs);
550 
551 	/*
552  	 * now that aobj is ready, add it to the global list
553  	 */
554 
555 	mutex_enter(&uao_list_lock);
556 	LIST_INSERT_HEAD(&uao_list, aobj, u_list);
557 	mutex_exit(&uao_list_lock);
558 	return(&aobj->u_obj);
559 }
560 
561 
562 
563 /*
564  * uao_init: set up aobj pager subsystem
565  *
566  * => called at boot time from uvm_pager_init()
567  */
568 
569 void
570 uao_init(void)
571 {
572 	static int uao_initialized;
573 
574 	if (uao_initialized)
575 		return;
576 	uao_initialized = true;
577 	LIST_INIT(&uao_list);
578 	mutex_init(&uao_list_lock, MUTEX_DEFAULT, IPL_NONE);
579 }
580 
581 /*
582  * uao_reference: add a ref to an aobj
583  *
584  * => aobj must be unlocked
585  * => just lock it and call the locked version
586  */
587 
588 void
589 uao_reference(struct uvm_object *uobj)
590 {
591 	mutex_enter(&uobj->vmobjlock);
592 	uao_reference_locked(uobj);
593 	mutex_exit(&uobj->vmobjlock);
594 }
595 
596 /*
597  * uao_reference_locked: add a ref to an aobj that is already locked
598  *
599  * => aobj must be locked
600  * this needs to be separate from the normal routine
601  * since sometimes we need to add a reference to an aobj when
602  * it's already locked.
603  */
604 
605 void
606 uao_reference_locked(struct uvm_object *uobj)
607 {
608 	UVMHIST_FUNC("uao_reference"); UVMHIST_CALLED(maphist);
609 
610 	/*
611  	 * kernel_object already has plenty of references, leave it alone.
612  	 */
613 
614 	if (UVM_OBJ_IS_KERN_OBJECT(uobj))
615 		return;
616 
617 	uobj->uo_refs++;
618 	UVMHIST_LOG(maphist, "<- done (uobj=0x%x, ref = %d)",
619 		    uobj, uobj->uo_refs,0,0);
620 }
621 
622 /*
623  * uao_detach: drop a reference to an aobj
624  *
625  * => aobj must be unlocked
626  * => just lock it and call the locked version
627  */
628 
629 void
630 uao_detach(struct uvm_object *uobj)
631 {
632 	mutex_enter(&uobj->vmobjlock);
633 	uao_detach_locked(uobj);
634 }
635 
636 /*
637  * uao_detach_locked: drop a reference to an aobj
638  *
639  * => aobj must be locked, and is unlocked (or freed) upon return.
640  * this needs to be separate from the normal routine
641  * since sometimes we need to detach from an aobj when
642  * it's already locked.
643  */
644 
645 void
646 uao_detach_locked(struct uvm_object *uobj)
647 {
648 	struct uvm_aobj *aobj = (struct uvm_aobj *)uobj;
649 	struct vm_page *pg;
650 	UVMHIST_FUNC("uao_detach"); UVMHIST_CALLED(maphist);
651 
652 	/*
653  	 * detaching from kernel_object is a noop.
654  	 */
655 
656 	if (UVM_OBJ_IS_KERN_OBJECT(uobj)) {
657 		mutex_exit(&uobj->vmobjlock);
658 		return;
659 	}
660 
661 	UVMHIST_LOG(maphist,"  (uobj=0x%x)  ref=%d", uobj,uobj->uo_refs,0,0);
662 	uobj->uo_refs--;
663 	if (uobj->uo_refs) {
664 		mutex_exit(&uobj->vmobjlock);
665 		UVMHIST_LOG(maphist, "<- done (rc>0)", 0,0,0,0);
666 		return;
667 	}
668 
669 	/*
670  	 * remove the aobj from the global list.
671  	 */
672 
673 	mutex_enter(&uao_list_lock);
674 	LIST_REMOVE(aobj, u_list);
675 	mutex_exit(&uao_list_lock);
676 
677 	/*
678  	 * free all the pages left in the aobj.  for each page,
679 	 * when the page is no longer busy (and thus after any disk i/o that
680 	 * it's involved in is complete), release any swap resources and
681 	 * free the page itself.
682  	 */
683 
684 	mutex_enter(&uvm_pageqlock);
685 	while ((pg = TAILQ_FIRST(&uobj->memq)) != NULL) {
686 		pmap_page_protect(pg, VM_PROT_NONE);
687 		if (pg->flags & PG_BUSY) {
688 			pg->flags |= PG_WANTED;
689 			mutex_exit(&uvm_pageqlock);
690 			UVM_UNLOCK_AND_WAIT(pg, &uobj->vmobjlock, false,
691 			    "uao_det", 0);
692 			mutex_enter(&uobj->vmobjlock);
693 			mutex_enter(&uvm_pageqlock);
694 			continue;
695 		}
696 		uao_dropswap(&aobj->u_obj, pg->offset >> PAGE_SHIFT);
697 		uvm_pagefree(pg);
698 	}
699 	mutex_exit(&uvm_pageqlock);
700 
701 	/*
702  	 * finally, free the aobj itself.
703  	 */
704 
705 	uao_free(aobj);
706 }
707 
708 /*
709  * uao_put: flush pages out of a uvm object
710  *
711  * => object should be locked by caller.  we may _unlock_ the object
712  *	if (and only if) we need to clean a page (PGO_CLEANIT).
713  *	XXXJRT Currently, however, we don't.  In the case of cleaning
714  *	XXXJRT a page, we simply just deactivate it.  Should probably
715  *	XXXJRT handle this better, in the future (although "flushing"
716  *	XXXJRT anonymous memory isn't terribly important).
717  * => if PGO_CLEANIT is not set, then we will neither unlock the object
718  *	or block.
719  * => if PGO_ALLPAGE is set, then all pages in the object are valid targets
720  *	for flushing.
721  * => NOTE: we rely on the fact that the object's memq is a TAILQ and
722  *	that new pages are inserted on the tail end of the list.  thus,
723  *	we can make a complete pass through the object in one go by starting
724  *	at the head and working towards the tail (new pages are put in
725  *	front of us).
726  * => NOTE: we are allowed to lock the page queues, so the caller
727  *	must not be holding the lock on them [e.g. pagedaemon had
728  *	better not call us with the queues locked]
729  * => we return 0 unless we encountered some sort of I/O error
730  *	XXXJRT currently never happens, as we never directly initiate
731  *	XXXJRT I/O
732  *
733  * note on page traversal:
734  *	we can traverse the pages in an object either by going down the
735  *	linked list in "uobj->memq", or we can go over the address range
736  *	by page doing hash table lookups for each address.  depending
737  *	on how many pages are in the object it may be cheaper to do one
738  *	or the other.  we set "by_list" to true if we are using memq.
739  *	if the cost of a hash lookup was equal to the cost of the list
740  *	traversal we could compare the number of pages in the start->stop
741  *	range to the total number of pages in the object.  however, it
742  *	seems that a hash table lookup is more expensive than the linked
743  *	list traversal, so we multiply the number of pages in the
744  *	start->stop range by a penalty which we define below.
745  */
746 
747 static int
748 uao_put(struct uvm_object *uobj, voff_t start, voff_t stop, int flags)
749 {
750 	struct uvm_aobj *aobj = (struct uvm_aobj *)uobj;
751 	struct vm_page *pg, *nextpg, curmp, endmp;
752 	bool by_list;
753 	voff_t curoff;
754 	UVMHIST_FUNC("uao_put"); UVMHIST_CALLED(maphist);
755 
756 	KASSERT(mutex_owned(&uobj->vmobjlock));
757 
758 	curoff = 0;
759 	if (flags & PGO_ALLPAGES) {
760 		start = 0;
761 		stop = aobj->u_pages << PAGE_SHIFT;
762 		by_list = true;		/* always go by the list */
763 	} else {
764 		start = trunc_page(start);
765 		if (stop == 0) {
766 			stop = aobj->u_pages << PAGE_SHIFT;
767 		} else {
768 			stop = round_page(stop);
769 		}
770 		if (stop > (aobj->u_pages << PAGE_SHIFT)) {
771 			printf("uao_flush: strange, got an out of range "
772 			    "flush (fixed)\n");
773 			stop = aobj->u_pages << PAGE_SHIFT;
774 		}
775 		by_list = (uobj->uo_npages <=
776 		    ((stop - start) >> PAGE_SHIFT) * UVM_PAGE_HASH_PENALTY);
777 	}
778 	UVMHIST_LOG(maphist,
779 	    " flush start=0x%lx, stop=0x%x, by_list=%d, flags=0x%x",
780 	    start, stop, by_list, flags);
781 
782 	/*
783 	 * Don't need to do any work here if we're not freeing
784 	 * or deactivating pages.
785 	 */
786 
787 	if ((flags & (PGO_DEACTIVATE|PGO_FREE)) == 0) {
788 		mutex_exit(&uobj->vmobjlock);
789 		return 0;
790 	}
791 
792 	/*
793 	 * Initialize the marker pages.  See the comment in
794 	 * genfs_putpages() also.
795 	 */
796 
797 	curmp.uobject = uobj;
798 	curmp.offset = (voff_t)-1;
799 	curmp.flags = PG_BUSY;
800 	endmp.uobject = uobj;
801 	endmp.offset = (voff_t)-1;
802 	endmp.flags = PG_BUSY;
803 
804 	/*
805 	 * now do it.  note: we must update nextpg in the body of loop or we
806 	 * will get stuck.  we need to use nextpg if we'll traverse the list
807 	 * because we may free "pg" before doing the next loop.
808 	 */
809 
810 	if (by_list) {
811 		TAILQ_INSERT_TAIL(&uobj->memq, &endmp, listq);
812 		nextpg = TAILQ_FIRST(&uobj->memq);
813 		uvm_lwp_hold(curlwp);
814 	} else {
815 		curoff = start;
816 		nextpg = NULL;	/* Quell compiler warning */
817 	}
818 
819 	/* locked: uobj */
820 	for (;;) {
821 		if (by_list) {
822 			pg = nextpg;
823 			if (pg == &endmp)
824 				break;
825 			nextpg = TAILQ_NEXT(pg, listq);
826 			if (pg->offset < start || pg->offset >= stop)
827 				continue;
828 		} else {
829 			if (curoff < stop) {
830 				pg = uvm_pagelookup(uobj, curoff);
831 				curoff += PAGE_SIZE;
832 			} else
833 				break;
834 			if (pg == NULL)
835 				continue;
836 		}
837 
838 		/*
839 		 * wait and try again if the page is busy.
840 		 */
841 
842 		if (pg->flags & PG_BUSY) {
843 			if (by_list) {
844 				TAILQ_INSERT_BEFORE(pg, &curmp, listq);
845 			}
846 			pg->flags |= PG_WANTED;
847 			UVM_UNLOCK_AND_WAIT(pg, &uobj->vmobjlock, 0,
848 			    "uao_put", 0);
849 			mutex_enter(&uobj->vmobjlock);
850 			if (by_list) {
851 				nextpg = TAILQ_NEXT(&curmp, listq);
852 				TAILQ_REMOVE(&uobj->memq, &curmp,
853 				    listq);
854 			} else
855 				curoff -= PAGE_SIZE;
856 			continue;
857 		}
858 
859 		switch (flags & (PGO_CLEANIT|PGO_FREE|PGO_DEACTIVATE)) {
860 
861 		/*
862 		 * XXX In these first 3 cases, we always just
863 		 * XXX deactivate the page.  We may want to
864 		 * XXX handle the different cases more specifically
865 		 * XXX in the future.
866 		 */
867 
868 		case PGO_CLEANIT|PGO_FREE:
869 		case PGO_CLEANIT|PGO_DEACTIVATE:
870 		case PGO_DEACTIVATE:
871  deactivate_it:
872 			mutex_enter(&uvm_pageqlock);
873 			/* skip the page if it's wired */
874 			if (pg->wire_count == 0) {
875 				uvm_pagedeactivate(pg);
876 			}
877 			mutex_exit(&uvm_pageqlock);
878 			break;
879 
880 		case PGO_FREE:
881 			/*
882 			 * If there are multiple references to
883 			 * the object, just deactivate the page.
884 			 */
885 
886 			if (uobj->uo_refs > 1)
887 				goto deactivate_it;
888 
889 			/*
890 			 * free the swap slot and the page.
891 			 */
892 
893 			pmap_page_protect(pg, VM_PROT_NONE);
894 
895 			/*
896 			 * freeing swapslot here is not strictly necessary.
897 			 * however, leaving it here doesn't save much
898 			 * because we need to update swap accounting anyway.
899 			 */
900 
901 			uao_dropswap(uobj, pg->offset >> PAGE_SHIFT);
902 			mutex_enter(&uvm_pageqlock);
903 			uvm_pagefree(pg);
904 			mutex_exit(&uvm_pageqlock);
905 			break;
906 
907 		default:
908 			panic("%s: impossible", __func__);
909 		}
910 	}
911 	if (by_list) {
912 		TAILQ_REMOVE(&uobj->memq, &endmp, listq);
913 		uvm_lwp_rele(curlwp);
914 	}
915 	mutex_exit(&uobj->vmobjlock);
916 	return 0;
917 }
918 
919 /*
920  * uao_get: fetch me a page
921  *
922  * we have three cases:
923  * 1: page is resident     -> just return the page.
924  * 2: page is zero-fill    -> allocate a new page and zero it.
925  * 3: page is swapped out  -> fetch the page from swap.
926  *
927  * cases 1 and 2 can be handled with PGO_LOCKED, case 3 cannot.
928  * so, if the "center" page hits case 3 (or any page, with PGO_ALLPAGES),
929  * then we will need to return EBUSY.
930  *
931  * => prefer map unlocked (not required)
932  * => object must be locked!  we will _unlock_ it before starting any I/O.
933  * => flags: PGO_ALLPAGES: get all of the pages
934  *           PGO_LOCKED: fault data structures are locked
935  * => NOTE: offset is the offset of pps[0], _NOT_ pps[centeridx]
936  * => NOTE: caller must check for released pages!!
937  */
938 
939 static int
940 uao_get(struct uvm_object *uobj, voff_t offset, struct vm_page **pps,
941     int *npagesp, int centeridx, vm_prot_t access_type, int advice, int flags)
942 {
943 #if defined(VMSWAP)
944 	struct uvm_aobj *aobj = (struct uvm_aobj *)uobj;
945 #endif /* defined(VMSWAP) */
946 	voff_t current_offset;
947 	struct vm_page *ptmp = NULL;	/* Quell compiler warning */
948 	int lcv, gotpages, maxpages, swslot, pageidx;
949 	bool done;
950 	UVMHIST_FUNC("uao_get"); UVMHIST_CALLED(pdhist);
951 
952 	UVMHIST_LOG(pdhist, "aobj=%p offset=%d, flags=%d",
953 		    (struct uvm_aobj *)uobj, offset, flags,0);
954 
955 	/*
956  	 * get number of pages
957  	 */
958 
959 	maxpages = *npagesp;
960 
961 	/*
962  	 * step 1: handled the case where fault data structures are locked.
963  	 */
964 
965 	if (flags & PGO_LOCKED) {
966 
967 		/*
968  		 * step 1a: get pages that are already resident.   only do
969 		 * this if the data structures are locked (i.e. the first
970 		 * time through).
971  		 */
972 
973 		done = true;	/* be optimistic */
974 		gotpages = 0;	/* # of pages we got so far */
975 		for (lcv = 0, current_offset = offset ; lcv < maxpages ;
976 		    lcv++, current_offset += PAGE_SIZE) {
977 			/* do we care about this page?  if not, skip it */
978 			if (pps[lcv] == PGO_DONTCARE)
979 				continue;
980 			ptmp = uvm_pagelookup(uobj, current_offset);
981 
982 			/*
983  			 * if page is new, attempt to allocate the page,
984 			 * zero-fill'd.
985  			 */
986 
987 			if (ptmp == NULL && uao_find_swslot(&aobj->u_obj,
988 			    current_offset >> PAGE_SHIFT) == 0) {
989 				ptmp = uvm_pagealloc(uobj, current_offset,
990 				    NULL, UVM_PGA_ZERO);
991 				if (ptmp) {
992 					/* new page */
993 					ptmp->flags &= ~(PG_FAKE);
994 					ptmp->pqflags |= PQ_AOBJ;
995 					goto gotpage;
996 				}
997 			}
998 
999 			/*
1000 			 * to be useful must get a non-busy page
1001 			 */
1002 
1003 			if (ptmp == NULL || (ptmp->flags & PG_BUSY) != 0) {
1004 				if (lcv == centeridx ||
1005 				    (flags & PGO_ALLPAGES) != 0)
1006 					/* need to do a wait or I/O! */
1007 					done = false;
1008 					continue;
1009 			}
1010 
1011 			/*
1012 			 * useful page: busy/lock it and plug it in our
1013 			 * result array
1014 			 */
1015 
1016 			/* caller must un-busy this page */
1017 			ptmp->flags |= PG_BUSY;
1018 			UVM_PAGE_OWN(ptmp, "uao_get1");
1019 gotpage:
1020 			pps[lcv] = ptmp;
1021 			gotpages++;
1022 		}
1023 
1024 		/*
1025  		 * step 1b: now we've either done everything needed or we
1026 		 * to unlock and do some waiting or I/O.
1027  		 */
1028 
1029 		UVMHIST_LOG(pdhist, "<- done (done=%d)", done, 0,0,0);
1030 		*npagesp = gotpages;
1031 		if (done)
1032 			return 0;
1033 		else
1034 			return EBUSY;
1035 	}
1036 
1037 	/*
1038  	 * step 2: get non-resident or busy pages.
1039  	 * object is locked.   data structures are unlocked.
1040  	 */
1041 
1042 	if ((flags & PGO_SYNCIO) == 0) {
1043 		goto done;
1044 	}
1045 
1046 	for (lcv = 0, current_offset = offset ; lcv < maxpages ;
1047 	    lcv++, current_offset += PAGE_SIZE) {
1048 
1049 		/*
1050 		 * - skip over pages we've already gotten or don't want
1051 		 * - skip over pages we don't _have_ to get
1052 		 */
1053 
1054 		if (pps[lcv] != NULL ||
1055 		    (lcv != centeridx && (flags & PGO_ALLPAGES) == 0))
1056 			continue;
1057 
1058 		pageidx = current_offset >> PAGE_SHIFT;
1059 
1060 		/*
1061  		 * we have yet to locate the current page (pps[lcv]).   we
1062 		 * first look for a page that is already at the current offset.
1063 		 * if we find a page, we check to see if it is busy or
1064 		 * released.  if that is the case, then we sleep on the page
1065 		 * until it is no longer busy or released and repeat the lookup.
1066 		 * if the page we found is neither busy nor released, then we
1067 		 * busy it (so we own it) and plug it into pps[lcv].   this
1068 		 * 'break's the following while loop and indicates we are
1069 		 * ready to move on to the next page in the "lcv" loop above.
1070  		 *
1071  		 * if we exit the while loop with pps[lcv] still set to NULL,
1072 		 * then it means that we allocated a new busy/fake/clean page
1073 		 * ptmp in the object and we need to do I/O to fill in the data.
1074  		 */
1075 
1076 		/* top of "pps" while loop */
1077 		while (pps[lcv] == NULL) {
1078 			/* look for a resident page */
1079 			ptmp = uvm_pagelookup(uobj, current_offset);
1080 
1081 			/* not resident?   allocate one now (if we can) */
1082 			if (ptmp == NULL) {
1083 
1084 				ptmp = uvm_pagealloc(uobj, current_offset,
1085 				    NULL, 0);
1086 
1087 				/* out of RAM? */
1088 				if (ptmp == NULL) {
1089 					mutex_exit(&uobj->vmobjlock);
1090 					UVMHIST_LOG(pdhist,
1091 					    "sleeping, ptmp == NULL\n",0,0,0,0);
1092 					uvm_wait("uao_getpage");
1093 					mutex_enter(&uobj->vmobjlock);
1094 					continue;
1095 				}
1096 
1097 				/*
1098 				 * safe with PQ's unlocked: because we just
1099 				 * alloc'd the page
1100 				 */
1101 
1102 				ptmp->pqflags |= PQ_AOBJ;
1103 
1104 				/*
1105 				 * got new page ready for I/O.  break pps while
1106 				 * loop.  pps[lcv] is still NULL.
1107 				 */
1108 
1109 				break;
1110 			}
1111 
1112 			/* page is there, see if we need to wait on it */
1113 			if ((ptmp->flags & PG_BUSY) != 0) {
1114 				ptmp->flags |= PG_WANTED;
1115 				UVMHIST_LOG(pdhist,
1116 				    "sleeping, ptmp->flags 0x%x\n",
1117 				    ptmp->flags,0,0,0);
1118 				UVM_UNLOCK_AND_WAIT(ptmp, &uobj->vmobjlock,
1119 				    false, "uao_get", 0);
1120 				mutex_enter(&uobj->vmobjlock);
1121 				continue;
1122 			}
1123 
1124 			/*
1125  			 * if we get here then the page has become resident and
1126 			 * unbusy between steps 1 and 2.  we busy it now (so we
1127 			 * own it) and set pps[lcv] (so that we exit the while
1128 			 * loop).
1129  			 */
1130 
1131 			/* we own it, caller must un-busy */
1132 			ptmp->flags |= PG_BUSY;
1133 			UVM_PAGE_OWN(ptmp, "uao_get2");
1134 			pps[lcv] = ptmp;
1135 		}
1136 
1137 		/*
1138  		 * if we own the valid page at the correct offset, pps[lcv] will
1139  		 * point to it.   nothing more to do except go to the next page.
1140  		 */
1141 
1142 		if (pps[lcv])
1143 			continue;			/* next lcv */
1144 
1145 		/*
1146  		 * we have a "fake/busy/clean" page that we just allocated.
1147  		 * do the needed "i/o", either reading from swap or zeroing.
1148  		 */
1149 
1150 		swslot = uao_find_swslot(&aobj->u_obj, pageidx);
1151 
1152 		/*
1153  		 * just zero the page if there's nothing in swap.
1154  		 */
1155 
1156 		if (swslot == 0) {
1157 
1158 			/*
1159 			 * page hasn't existed before, just zero it.
1160 			 */
1161 
1162 			uvm_pagezero(ptmp);
1163 		} else {
1164 #if defined(VMSWAP)
1165 			int error;
1166 
1167 			UVMHIST_LOG(pdhist, "pagein from swslot %d",
1168 			     swslot, 0,0,0);
1169 
1170 			/*
1171 			 * page in the swapped-out page.
1172 			 * unlock object for i/o, relock when done.
1173 			 */
1174 
1175 			mutex_exit(&uobj->vmobjlock);
1176 			error = uvm_swap_get(ptmp, swslot, PGO_SYNCIO);
1177 			mutex_enter(&uobj->vmobjlock);
1178 
1179 			/*
1180 			 * I/O done.  check for errors.
1181 			 */
1182 
1183 			if (error != 0) {
1184 				UVMHIST_LOG(pdhist, "<- done (error=%d)",
1185 				    error,0,0,0);
1186 				if (ptmp->flags & PG_WANTED)
1187 					wakeup(ptmp);
1188 
1189 				/*
1190 				 * remove the swap slot from the aobj
1191 				 * and mark the aobj as having no real slot.
1192 				 * don't free the swap slot, thus preventing
1193 				 * it from being used again.
1194 				 */
1195 
1196 				swslot = uao_set_swslot(&aobj->u_obj, pageidx,
1197 							SWSLOT_BAD);
1198 				if (swslot > 0) {
1199 					uvm_swap_markbad(swslot, 1);
1200 				}
1201 
1202 				mutex_enter(&uvm_pageqlock);
1203 				uvm_pagefree(ptmp);
1204 				mutex_exit(&uvm_pageqlock);
1205 				mutex_exit(&uobj->vmobjlock);
1206 				return error;
1207 			}
1208 #else /* defined(VMSWAP) */
1209 			panic("%s: pagein", __func__);
1210 #endif /* defined(VMSWAP) */
1211 		}
1212 
1213 		if ((access_type & VM_PROT_WRITE) == 0) {
1214 			ptmp->flags |= PG_CLEAN;
1215 			pmap_clear_modify(ptmp);
1216 		}
1217 
1218 		/*
1219  		 * we got the page!   clear the fake flag (indicates valid
1220 		 * data now in page) and plug into our result array.   note
1221 		 * that page is still busy.
1222  		 *
1223  		 * it is the callers job to:
1224  		 * => check if the page is released
1225  		 * => unbusy the page
1226  		 * => activate the page
1227  		 */
1228 
1229 		ptmp->flags &= ~PG_FAKE;
1230 		pps[lcv] = ptmp;
1231 	}
1232 
1233 	/*
1234  	 * finally, unlock object and return.
1235  	 */
1236 
1237 done:
1238 	mutex_exit(&uobj->vmobjlock);
1239 	UVMHIST_LOG(pdhist, "<- done (OK)",0,0,0,0);
1240 	return 0;
1241 }
1242 
1243 #if defined(VMSWAP)
1244 
1245 /*
1246  * uao_dropswap:  release any swap resources from this aobj page.
1247  *
1248  * => aobj must be locked or have a reference count of 0.
1249  */
1250 
1251 void
1252 uao_dropswap(struct uvm_object *uobj, int pageidx)
1253 {
1254 	int slot;
1255 
1256 	slot = uao_set_swslot(uobj, pageidx, 0);
1257 	if (slot) {
1258 		uvm_swap_free(slot, 1);
1259 	}
1260 }
1261 
1262 /*
1263  * page in every page in every aobj that is paged-out to a range of swslots.
1264  *
1265  * => nothing should be locked.
1266  * => returns true if pagein was aborted due to lack of memory.
1267  */
1268 
1269 bool
1270 uao_swap_off(int startslot, int endslot)
1271 {
1272 	struct uvm_aobj *aobj, *nextaobj;
1273 	bool rv;
1274 
1275 	/*
1276 	 * walk the list of all aobjs.
1277 	 */
1278 
1279 restart:
1280 	mutex_enter(&uao_list_lock);
1281 	for (aobj = LIST_FIRST(&uao_list);
1282 	     aobj != NULL;
1283 	     aobj = nextaobj) {
1284 
1285 		/*
1286 		 * try to get the object lock, start all over if we fail.
1287 		 * most of the time we'll get the aobj lock,
1288 		 * so this should be a rare case.
1289 		 */
1290 
1291 		if (!mutex_tryenter(&aobj->u_obj.vmobjlock)) {
1292 			mutex_exit(&uao_list_lock);
1293 			/* XXX Better than yielding but inadequate. */
1294 			kpause("livelock", false, 1, NULL);
1295 			goto restart;
1296 		}
1297 
1298 		/*
1299 		 * add a ref to the aobj so it doesn't disappear
1300 		 * while we're working.
1301 		 */
1302 
1303 		uao_reference_locked(&aobj->u_obj);
1304 
1305 		/*
1306 		 * now it's safe to unlock the uao list.
1307 		 */
1308 
1309 		mutex_exit(&uao_list_lock);
1310 
1311 		/*
1312 		 * page in any pages in the swslot range.
1313 		 * if there's an error, abort and return the error.
1314 		 */
1315 
1316 		rv = uao_pagein(aobj, startslot, endslot);
1317 		if (rv) {
1318 			uao_detach_locked(&aobj->u_obj);
1319 			return rv;
1320 		}
1321 
1322 		/*
1323 		 * we're done with this aobj.
1324 		 * relock the list and drop our ref on the aobj.
1325 		 */
1326 
1327 		mutex_enter(&uao_list_lock);
1328 		nextaobj = LIST_NEXT(aobj, u_list);
1329 		uao_detach_locked(&aobj->u_obj);
1330 	}
1331 
1332 	/*
1333 	 * done with traversal, unlock the list
1334 	 */
1335 	mutex_exit(&uao_list_lock);
1336 	return false;
1337 }
1338 
1339 
1340 /*
1341  * page in any pages from aobj in the given range.
1342  *
1343  * => aobj must be locked and is returned locked.
1344  * => returns true if pagein was aborted due to lack of memory.
1345  */
1346 static bool
1347 uao_pagein(struct uvm_aobj *aobj, int startslot, int endslot)
1348 {
1349 	bool rv;
1350 
1351 	if (UAO_USES_SWHASH(aobj)) {
1352 		struct uao_swhash_elt *elt;
1353 		int buck;
1354 
1355 restart:
1356 		for (buck = aobj->u_swhashmask; buck >= 0; buck--) {
1357 			for (elt = LIST_FIRST(&aobj->u_swhash[buck]);
1358 			     elt != NULL;
1359 			     elt = LIST_NEXT(elt, list)) {
1360 				int i;
1361 
1362 				for (i = 0; i < UAO_SWHASH_CLUSTER_SIZE; i++) {
1363 					int slot = elt->slots[i];
1364 
1365 					/*
1366 					 * if the slot isn't in range, skip it.
1367 					 */
1368 
1369 					if (slot < startslot ||
1370 					    slot >= endslot) {
1371 						continue;
1372 					}
1373 
1374 					/*
1375 					 * process the page,
1376 					 * the start over on this object
1377 					 * since the swhash elt
1378 					 * may have been freed.
1379 					 */
1380 
1381 					rv = uao_pagein_page(aobj,
1382 					  UAO_SWHASH_ELT_PAGEIDX_BASE(elt) + i);
1383 					if (rv) {
1384 						return rv;
1385 					}
1386 					goto restart;
1387 				}
1388 			}
1389 		}
1390 	} else {
1391 		int i;
1392 
1393 		for (i = 0; i < aobj->u_pages; i++) {
1394 			int slot = aobj->u_swslots[i];
1395 
1396 			/*
1397 			 * if the slot isn't in range, skip it
1398 			 */
1399 
1400 			if (slot < startslot || slot >= endslot) {
1401 				continue;
1402 			}
1403 
1404 			/*
1405 			 * process the page.
1406 			 */
1407 
1408 			rv = uao_pagein_page(aobj, i);
1409 			if (rv) {
1410 				return rv;
1411 			}
1412 		}
1413 	}
1414 
1415 	return false;
1416 }
1417 
1418 /*
1419  * page in a page from an aobj.  used for swap_off.
1420  * returns true if pagein was aborted due to lack of memory.
1421  *
1422  * => aobj must be locked and is returned locked.
1423  */
1424 
1425 static bool
1426 uao_pagein_page(struct uvm_aobj *aobj, int pageidx)
1427 {
1428 	struct vm_page *pg;
1429 	int rv, npages;
1430 
1431 	pg = NULL;
1432 	npages = 1;
1433 	/* locked: aobj */
1434 	rv = uao_get(&aobj->u_obj, pageidx << PAGE_SHIFT,
1435 	    &pg, &npages, 0, VM_PROT_READ|VM_PROT_WRITE, 0, PGO_SYNCIO);
1436 	/* unlocked: aobj */
1437 
1438 	/*
1439 	 * relock and finish up.
1440 	 */
1441 
1442 	mutex_enter(&aobj->u_obj.vmobjlock);
1443 	switch (rv) {
1444 	case 0:
1445 		break;
1446 
1447 	case EIO:
1448 	case ERESTART:
1449 
1450 		/*
1451 		 * nothing more to do on errors.
1452 		 * ERESTART can only mean that the anon was freed,
1453 		 * so again there's nothing to do.
1454 		 */
1455 
1456 		return false;
1457 
1458 	default:
1459 		return true;
1460 	}
1461 
1462 	/*
1463 	 * ok, we've got the page now.
1464 	 * mark it as dirty, clear its swslot and un-busy it.
1465 	 */
1466 	uao_dropswap(&aobj->u_obj, pageidx);
1467 
1468 	/*
1469 	 * make sure it's on a page queue.
1470 	 */
1471 	mutex_enter(&uvm_pageqlock);
1472 	if (pg->wire_count == 0)
1473 		uvm_pageenqueue(pg);
1474 	mutex_exit(&uvm_pageqlock);
1475 
1476 	if (pg->flags & PG_WANTED) {
1477 		wakeup(pg);
1478 	}
1479 	pg->flags &= ~(PG_WANTED|PG_BUSY|PG_CLEAN|PG_FAKE);
1480 	UVM_PAGE_OWN(pg, NULL);
1481 
1482 	return false;
1483 }
1484 
1485 /*
1486  * uao_dropswap_range: drop swapslots in the range.
1487  *
1488  * => aobj must be locked and is returned locked.
1489  * => start is inclusive.  end is exclusive.
1490  */
1491 
1492 void
1493 uao_dropswap_range(struct uvm_object *uobj, voff_t start, voff_t end)
1494 {
1495 	struct uvm_aobj *aobj = (struct uvm_aobj *)uobj;
1496 
1497 	KASSERT(mutex_owned(&uobj->vmobjlock));
1498 
1499 	uao_dropswap_range1(aobj, start, end);
1500 }
1501 
1502 static void
1503 uao_dropswap_range1(struct uvm_aobj *aobj, voff_t start, voff_t end)
1504 {
1505 	int swpgonlydelta = 0;
1506 
1507 	if (end == 0) {
1508 		end = INT64_MAX;
1509 	}
1510 
1511 	if (UAO_USES_SWHASH(aobj)) {
1512 		int i, hashbuckets = aobj->u_swhashmask + 1;
1513 		voff_t taghi;
1514 		voff_t taglo;
1515 
1516 		taglo = UAO_SWHASH_ELT_TAG(start);
1517 		taghi = UAO_SWHASH_ELT_TAG(end);
1518 
1519 		for (i = 0; i < hashbuckets; i++) {
1520 			struct uao_swhash_elt *elt, *next;
1521 
1522 			for (elt = LIST_FIRST(&aobj->u_swhash[i]);
1523 			     elt != NULL;
1524 			     elt = next) {
1525 				int startidx, endidx;
1526 				int j;
1527 
1528 				next = LIST_NEXT(elt, list);
1529 
1530 				if (elt->tag < taglo || taghi < elt->tag) {
1531 					continue;
1532 				}
1533 
1534 				if (elt->tag == taglo) {
1535 					startidx =
1536 					    UAO_SWHASH_ELT_PAGESLOT_IDX(start);
1537 				} else {
1538 					startidx = 0;
1539 				}
1540 
1541 				if (elt->tag == taghi) {
1542 					endidx =
1543 					    UAO_SWHASH_ELT_PAGESLOT_IDX(end);
1544 				} else {
1545 					endidx = UAO_SWHASH_CLUSTER_SIZE;
1546 				}
1547 
1548 				for (j = startidx; j < endidx; j++) {
1549 					int slot = elt->slots[j];
1550 
1551 					KASSERT(uvm_pagelookup(&aobj->u_obj,
1552 					    (UAO_SWHASH_ELT_PAGEIDX_BASE(elt)
1553 					    + j) << PAGE_SHIFT) == NULL);
1554 					if (slot > 0) {
1555 						uvm_swap_free(slot, 1);
1556 						swpgonlydelta++;
1557 						KASSERT(elt->count > 0);
1558 						elt->slots[j] = 0;
1559 						elt->count--;
1560 					}
1561 				}
1562 
1563 				if (elt->count == 0) {
1564 					LIST_REMOVE(elt, list);
1565 					pool_put(&uao_swhash_elt_pool, elt);
1566 				}
1567 			}
1568 		}
1569 	} else {
1570 		int i;
1571 
1572 		if (aobj->u_pages < end) {
1573 			end = aobj->u_pages;
1574 		}
1575 		for (i = start; i < end; i++) {
1576 			int slot = aobj->u_swslots[i];
1577 
1578 			if (slot > 0) {
1579 				uvm_swap_free(slot, 1);
1580 				swpgonlydelta++;
1581 			}
1582 		}
1583 	}
1584 
1585 	/*
1586 	 * adjust the counter of pages only in swap for all
1587 	 * the swap slots we've freed.
1588 	 */
1589 
1590 	if (swpgonlydelta > 0) {
1591 		mutex_enter(&uvm_swap_data_lock);
1592 		KASSERT(uvmexp.swpgonly >= swpgonlydelta);
1593 		uvmexp.swpgonly -= swpgonlydelta;
1594 		mutex_exit(&uvm_swap_data_lock);
1595 	}
1596 }
1597 
1598 #endif /* defined(VMSWAP) */
1599