xref: /openbsd-src/sys/uvm/uvm_aobj.c (revision fb8aa7497fded39583f40e800732f9c046411717)
1 /*	$OpenBSD: uvm_aobj.c,v 1.81 2016/06/17 10:48:25 dlg Exp $	*/
2 /*	$NetBSD: uvm_aobj.c,v 1.39 2001/02/18 21:19:08 chs Exp $	*/
3 
4 /*
5  * Copyright (c) 1998 Chuck Silvers, Charles D. Cranor and
6  *                    Washington University.
7  * All rights reserved.
8  *
9  * Redistribution and use in source and binary forms, with or without
10  * modification, are permitted provided that the following conditions
11  * are met:
12  * 1. Redistributions of source code must retain the above copyright
13  *    notice, this list of conditions and the following disclaimer.
14  * 2. Redistributions in binary form must reproduce the above copyright
15  *    notice, this list of conditions and the following disclaimer in the
16  *    documentation and/or other materials provided with the distribution.
17  *
18  * THIS SOFTWARE IS PROVIDED BY THE AUTHOR ``AS IS'' AND ANY EXPRESS OR
19  * IMPLIED WARRANTIES, INCLUDING, BUT NOT LIMITED TO, THE IMPLIED WARRANTIES
20  * OF MERCHANTABILITY AND FITNESS FOR A PARTICULAR PURPOSE ARE DISCLAIMED.
21  * IN NO EVENT SHALL THE AUTHOR BE LIABLE FOR ANY DIRECT, INDIRECT,
22  * INCIDENTAL, SPECIAL, EXEMPLARY, OR CONSEQUENTIAL DAMAGES (INCLUDING, BUT
23  * NOT LIMITED TO, PROCUREMENT OF SUBSTITUTE GOODS OR SERVICES; LOSS OF USE,
24  * DATA, OR PROFITS; OR BUSINESS INTERRUPTION) HOWEVER CAUSED AND ON ANY
25  * THEORY OF LIABILITY, WHETHER IN CONTRACT, STRICT LIABILITY, OR TORT
26  * (INCLUDING NEGLIGENCE OR OTHERWISE) ARISING IN ANY WAY OUT OF THE USE OF
27  * THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF SUCH DAMAGE.
28  *
29  * from: Id: uvm_aobj.c,v 1.1.2.5 1998/02/06 05:14:38 chs Exp
30  */
31 /*
32  * uvm_aobj.c: anonymous memory uvm_object pager
33  *
34  * author: Chuck Silvers <chuq@chuq.com>
35  * started: Jan-1998
36  *
37  * - design mostly from Chuck Cranor
38  */
39 
40 #include <sys/param.h>
41 #include <sys/systm.h>
42 #include <sys/malloc.h>
43 #include <sys/kernel.h>
44 #include <sys/pool.h>
45 #include <sys/stdint.h>
46 #include <sys/atomic.h>
47 
48 #include <uvm/uvm.h>
49 
50 /*
51  * an aobj manages anonymous-memory backed uvm_objects.   in addition
52  * to keeping the list of resident pages, it also keeps a list of
53  * allocated swap blocks.  depending on the size of the aobj this list
54  * of allocated swap blocks is either stored in an array (small objects)
55  * or in a hash table (large objects).
56  */
57 
58 /*
59  * local structures
60  */
61 
62 /*
63  * for hash tables, we break the address space of the aobj into blocks
64  * of UAO_SWHASH_CLUSTER_SIZE pages.   we require the cluster size to
65  * be a power of two.
66  */
67 #define UAO_SWHASH_CLUSTER_SHIFT 4
68 #define UAO_SWHASH_CLUSTER_SIZE (1 << UAO_SWHASH_CLUSTER_SHIFT)
69 
70 /* get the "tag" for this page index */
71 #define UAO_SWHASH_ELT_TAG(PAGEIDX) \
72 	((PAGEIDX) >> UAO_SWHASH_CLUSTER_SHIFT)
73 
74 /* given an ELT and a page index, find the swap slot */
75 #define UAO_SWHASH_ELT_PAGESLOT_IDX(PAGEIDX) \
76 	((PAGEIDX) & (UAO_SWHASH_CLUSTER_SIZE - 1))
77 #define UAO_SWHASH_ELT_PAGESLOT(ELT, PAGEIDX) \
78 	((ELT)->slots[(PAGEIDX) & (UAO_SWHASH_CLUSTER_SIZE - 1)])
79 
80 /* given an ELT, return its pageidx base */
81 #define UAO_SWHASH_ELT_PAGEIDX_BASE(ELT) \
82 	((ELT)->tag << UAO_SWHASH_CLUSTER_SHIFT)
83 
84 /*
85  * the swhash hash function
86  */
87 #define UAO_SWHASH_HASH(AOBJ, PAGEIDX) \
88 	(&(AOBJ)->u_swhash[(((PAGEIDX) >> UAO_SWHASH_CLUSTER_SHIFT) \
89 			    & (AOBJ)->u_swhashmask)])
90 
91 /*
92  * the swhash threshold determines if we will use an array or a
93  * hash table to store the list of allocated swap blocks.
94  */
95 
96 #define UAO_SWHASH_THRESHOLD (UAO_SWHASH_CLUSTER_SIZE * 4)
97 
98 /*
99  * the number of buckets in a swhash, with an upper bound
100  */
101 #define UAO_SWHASH_MAXBUCKETS 256
102 #define UAO_SWHASH_BUCKETS(pages) \
103 	(min((pages) >> UAO_SWHASH_CLUSTER_SHIFT, UAO_SWHASH_MAXBUCKETS))
104 
105 
106 /*
107  * uao_swhash_elt: when a hash table is being used, this structure defines
108  * the format of an entry in the bucket list.
109  */
110 struct uao_swhash_elt {
111 	LIST_ENTRY(uao_swhash_elt) list;	/* the hash list */
112 	voff_t tag;				/* our 'tag' */
113 	int count;				/* our number of active slots */
114 	int slots[UAO_SWHASH_CLUSTER_SIZE];	/* the slots */
115 };
116 
117 /*
118  * uao_swhash: the swap hash table structure
119  */
120 LIST_HEAD(uao_swhash, uao_swhash_elt);
121 
122 /*
123  * uao_swhash_elt_pool: pool of uao_swhash_elt structures
124  */
125 struct pool uao_swhash_elt_pool;
126 
127 /*
128  * uvm_aobj: the actual anon-backed uvm_object
129  *
130  * => the uvm_object is at the top of the structure, this allows
131  *   (struct uvm_aobj *) == (struct uvm_object *)
132  * => only one of u_swslots and u_swhash is used in any given aobj
133  */
134 struct uvm_aobj {
135 	struct uvm_object u_obj; /* has: pgops, memt, #pages, #refs */
136 	int u_pages;		 /* number of pages in entire object */
137 	int u_flags;		 /* the flags (see uvm_aobj.h) */
138 	/*
139 	 * Either an array or hashtable (array of bucket heads) of
140 	 * offset -> swapslot mappings for the aobj.
141 	 */
142 #define u_swslots	u_swap.slot_array
143 #define u_swhash	u_swap.slot_hash
144 	union swslots {
145 		int			*slot_array;
146 		struct uao_swhash	*slot_hash;
147 	} u_swap;
148 	u_long u_swhashmask;		/* mask for hashtable */
149 	LIST_ENTRY(uvm_aobj) u_list;	/* global list of aobjs */
150 };
151 
152 /*
153  * uvm_aobj_pool: pool of uvm_aobj structures
154  */
155 struct pool uvm_aobj_pool;
156 
157 /*
158  * local functions
159  */
160 static struct uao_swhash_elt	*uao_find_swhash_elt(struct uvm_aobj *, int,
161 				     boolean_t);
162 static int			 uao_find_swslot(struct uvm_aobj *, int);
163 static boolean_t		 uao_flush(struct uvm_object *, voff_t,
164 				     voff_t, int);
165 static void			 uao_free(struct uvm_aobj *);
166 static int			 uao_get(struct uvm_object *, voff_t,
167 				     vm_page_t *, int *, int, vm_prot_t,
168 				     int, int);
169 static boolean_t		 uao_pagein(struct uvm_aobj *, int, int);
170 static boolean_t		 uao_pagein_page(struct uvm_aobj *, int);
171 
172 void	uao_dropswap_range(struct uvm_object *, voff_t, voff_t);
173 void	uao_shrink_flush(struct uvm_object *, int, int);
174 int	uao_shrink_hash(struct uvm_object *, int);
175 int	uao_shrink_array(struct uvm_object *, int);
176 int	uao_shrink_convert(struct uvm_object *, int);
177 
178 int	uao_grow_hash(struct uvm_object *, int);
179 int	uao_grow_array(struct uvm_object *, int);
180 int	uao_grow_convert(struct uvm_object *, int);
181 
182 /*
183  * aobj_pager
184  *
185  * note that some functions (e.g. put) are handled elsewhere
186  */
187 struct uvm_pagerops aobj_pager = {
188 	NULL,			/* init */
189 	uao_reference,		/* reference */
190 	uao_detach,		/* detach */
191 	NULL,			/* fault */
192 	uao_flush,		/* flush */
193 	uao_get,		/* get */
194 };
195 
196 /*
197  * uao_list: global list of active aobjs, locked by uao_list_lock
198  *
199  * Lock ordering: generally the locking order is object lock, then list lock.
200  * in the case of swap off we have to iterate over the list, and thus the
201  * ordering is reversed. In that case we must use trylocking to prevent
202  * deadlock.
203  */
204 static LIST_HEAD(aobjlist, uvm_aobj) uao_list = LIST_HEAD_INITIALIZER(uao_list);
205 static struct mutex uao_list_lock = MUTEX_INITIALIZER(IPL_NONE);
206 
207 
208 /*
209  * functions
210  */
211 /*
212  * hash table/array related functions
213  */
214 /*
215  * uao_find_swhash_elt: find (or create) a hash table entry for a page
216  * offset.
217  */
218 static struct uao_swhash_elt *
219 uao_find_swhash_elt(struct uvm_aobj *aobj, int pageidx, boolean_t create)
220 {
221 	struct uao_swhash *swhash;
222 	struct uao_swhash_elt *elt;
223 	voff_t page_tag;
224 
225 	swhash = UAO_SWHASH_HASH(aobj, pageidx); /* first hash to get bucket */
226 	page_tag = UAO_SWHASH_ELT_TAG(pageidx);	/* tag to search for */
227 
228 	/* now search the bucket for the requested tag */
229 	LIST_FOREACH(elt, swhash, list) {
230 		if (elt->tag == page_tag)
231 			return(elt);
232 	}
233 
234 	/* fail now if we are not allowed to create a new entry in the bucket */
235 	if (!create)
236 		return NULL;
237 
238 	/* allocate a new entry for the bucket and init/insert it in */
239 	elt = pool_get(&uao_swhash_elt_pool, PR_NOWAIT | PR_ZERO);
240 	/*
241 	 * XXX We cannot sleep here as the hash table might disappear
242 	 * from under our feet.  And we run the risk of deadlocking
243 	 * the pagedeamon.  In fact this code will only be called by
244 	 * the pagedaemon and allocation will only fail if we
245 	 * exhausted the pagedeamon reserve.  In that case we're
246 	 * doomed anyway, so panic.
247 	 */
248 	if (elt == NULL)
249 		panic("%s: can't allocate entry", __func__);
250 	LIST_INSERT_HEAD(swhash, elt, list);
251 	elt->tag = page_tag;
252 
253 	return(elt);
254 }
255 
256 /*
257  * uao_find_swslot: find the swap slot number for an aobj/pageidx
258  */
259 __inline static int
260 uao_find_swslot(struct uvm_aobj *aobj, int pageidx)
261 {
262 
263 	/* if noswap flag is set, then we never return a slot */
264 	if (aobj->u_flags & UAO_FLAG_NOSWAP)
265 		return(0);
266 
267 	/* if hashing, look in hash table.  */
268 	if (aobj->u_pages > UAO_SWHASH_THRESHOLD) {
269 		struct uao_swhash_elt *elt =
270 		    uao_find_swhash_elt(aobj, pageidx, FALSE);
271 
272 		if (elt)
273 			return(UAO_SWHASH_ELT_PAGESLOT(elt, pageidx));
274 		else
275 			return(0);
276 	}
277 
278 	/* otherwise, look in the array */
279 	return(aobj->u_swslots[pageidx]);
280 }
281 
282 /*
283  * uao_set_swslot: set the swap slot for a page in an aobj.
284  *
285  * => setting a slot to zero frees the slot
286  */
287 int
288 uao_set_swslot(struct uvm_object *uobj, int pageidx, int slot)
289 {
290 	struct uvm_aobj *aobj = (struct uvm_aobj *)uobj;
291 	int oldslot;
292 
293 	/* if noswap flag is set, then we can't set a slot */
294 	if (aobj->u_flags & UAO_FLAG_NOSWAP) {
295 		if (slot == 0)
296 			return(0);		/* a clear is ok */
297 
298 		/* but a set is not */
299 		printf("uao_set_swslot: uobj = %p\n", uobj);
300 		panic("uao_set_swslot: attempt to set a slot"
301 		    " on a NOSWAP object");
302 	}
303 
304 	/* are we using a hash table?  if so, add it in the hash.  */
305 	if (aobj->u_pages > UAO_SWHASH_THRESHOLD) {
306 		/*
307 		 * Avoid allocating an entry just to free it again if
308 		 * the page had not swap slot in the first place, and
309 		 * we are freeing.
310 		 */
311 		struct uao_swhash_elt *elt =
312 		    uao_find_swhash_elt(aobj, pageidx, slot ? TRUE : FALSE);
313 		if (elt == NULL) {
314 			KASSERT(slot == 0);
315 			return (0);
316 		}
317 
318 		oldslot = UAO_SWHASH_ELT_PAGESLOT(elt, pageidx);
319 		UAO_SWHASH_ELT_PAGESLOT(elt, pageidx) = slot;
320 
321 		/*
322 		 * now adjust the elt's reference counter and free it if we've
323 		 * dropped it to zero.
324 		 */
325 		/* an allocation? */
326 		if (slot) {
327 			if (oldslot == 0)
328 				elt->count++;
329 		} else {		/* freeing slot ... */
330 			if (oldslot)	/* to be safe */
331 				elt->count--;
332 
333 			if (elt->count == 0) {
334 				LIST_REMOVE(elt, list);
335 				pool_put(&uao_swhash_elt_pool, elt);
336 			}
337 		}
338 	} else {
339 		/* we are using an array */
340 		oldslot = aobj->u_swslots[pageidx];
341 		aobj->u_swslots[pageidx] = slot;
342 	}
343 	return (oldslot);
344 }
345 /*
346  * end of hash/array functions
347  */
348 
349 /*
350  * uao_free: free all resources held by an aobj, and then free the aobj
351  *
352  * => the aobj should be dead
353  */
354 static void
355 uao_free(struct uvm_aobj *aobj)
356 {
357 
358 	if (aobj->u_pages > UAO_SWHASH_THRESHOLD) {
359 		int i, hashbuckets = aobj->u_swhashmask + 1;
360 
361 		/*
362 		 * free the swslots from each hash bucket,
363 		 * then the hash bucket, and finally the hash table itself.
364 		 */
365 		for (i = 0; i < hashbuckets; i++) {
366 			struct uao_swhash_elt *elt, *next;
367 
368 			for (elt = LIST_FIRST(&aobj->u_swhash[i]);
369 			     elt != NULL;
370 			     elt = next) {
371 				int j;
372 
373 				for (j = 0; j < UAO_SWHASH_CLUSTER_SIZE; j++) {
374 					int slot = elt->slots[j];
375 
376 					if (slot == 0) {
377 						continue;
378 					}
379 					uvm_swap_free(slot, 1);
380 					/*
381 					 * this page is no longer
382 					 * only in swap.
383 					 */
384 					uvmexp.swpgonly--;
385 				}
386 
387 				next = LIST_NEXT(elt, list);
388 				pool_put(&uao_swhash_elt_pool, elt);
389 			}
390 		}
391 		free(aobj->u_swhash, M_UVMAOBJ, 0);
392 	} else {
393 		int i;
394 
395 		/* free the array */
396 		for (i = 0; i < aobj->u_pages; i++) {
397 			int slot = aobj->u_swslots[i];
398 
399 			if (slot) {
400 				uvm_swap_free(slot, 1);
401 				/* this page is no longer only in swap. */
402 				uvmexp.swpgonly--;
403 			}
404 		}
405 		free(aobj->u_swslots, M_UVMAOBJ, 0);
406 	}
407 
408 	/* finally free the aobj itself */
409 	pool_put(&uvm_aobj_pool, aobj);
410 }
411 
412 /*
413  * pager functions
414  */
415 
416 /*
417  * Shrink an aobj to a given number of pages. The procedure is always the same:
418  * assess the necessity of data structure conversion (hash to array), secure
419  * resources, flush pages and drop swap slots.
420  *
421  */
422 
423 void
424 uao_shrink_flush(struct uvm_object *uobj, int startpg, int endpg)
425 {
426 	KASSERT(startpg < endpg);
427 	KASSERT(uobj->uo_refs == 1);
428 	uao_flush(uobj, (voff_t)startpg << PAGE_SHIFT,
429 	    (voff_t)endpg << PAGE_SHIFT, PGO_FREE);
430 	uao_dropswap_range(uobj, startpg, endpg);
431 }
432 
433 int
434 uao_shrink_hash(struct uvm_object *uobj, int pages)
435 {
436 	struct uvm_aobj *aobj = (struct uvm_aobj *)uobj;
437 	struct uao_swhash *new_swhash;
438 	struct uao_swhash_elt *elt;
439 	unsigned long new_hashmask;
440 	int i;
441 
442 	KASSERT(aobj->u_pages > UAO_SWHASH_THRESHOLD);
443 
444 	/*
445 	 * If the size of the hash table doesn't change, all we need to do is
446 	 * to adjust the page count.
447 	 */
448 	if (UAO_SWHASH_BUCKETS(aobj->u_pages) == UAO_SWHASH_BUCKETS(pages)) {
449 		uao_shrink_flush(uobj, pages, aobj->u_pages);
450 		aobj->u_pages = pages;
451 		return 0;
452 	}
453 
454 	new_swhash = hashinit(UAO_SWHASH_BUCKETS(pages), M_UVMAOBJ,
455 	    M_WAITOK | M_CANFAIL, &new_hashmask);
456 	if (new_swhash == NULL)
457 		return ENOMEM;
458 
459 	uao_shrink_flush(uobj, pages, aobj->u_pages);
460 
461 	/*
462 	 * Even though the hash table size is changing, the hash of the buckets
463 	 * we are interested in copying should not change.
464 	 */
465 	for (i = 0; i < UAO_SWHASH_BUCKETS(aobj->u_pages); i++) {
466 		while (LIST_EMPTY(&aobj->u_swhash[i]) == 0) {
467 			elt = LIST_FIRST(&aobj->u_swhash[i]);
468 			LIST_REMOVE(elt, list);
469 			LIST_INSERT_HEAD(&new_swhash[i], elt, list);
470 		}
471 	}
472 
473 	free(aobj->u_swhash, M_UVMAOBJ, 0);
474 
475 	aobj->u_swhash = new_swhash;
476 	aobj->u_pages = pages;
477 	aobj->u_swhashmask = new_hashmask;
478 
479 	return 0;
480 }
481 
482 int
483 uao_shrink_convert(struct uvm_object *uobj, int pages)
484 {
485 	struct uvm_aobj *aobj = (struct uvm_aobj *)uobj;
486 	struct uao_swhash_elt *elt;
487 	int i, *new_swslots;
488 
489 	new_swslots = mallocarray(pages, sizeof(int), M_UVMAOBJ,
490 	    M_WAITOK | M_CANFAIL | M_ZERO);
491 	if (new_swslots == NULL)
492 		return ENOMEM;
493 
494 	uao_shrink_flush(uobj, pages, aobj->u_pages);
495 
496 	/* Convert swap slots from hash to array.  */
497 	for (i = 0; i < pages; i++) {
498 		elt = uao_find_swhash_elt(aobj, i, FALSE);
499 		if (elt != NULL) {
500 			new_swslots[i] = UAO_SWHASH_ELT_PAGESLOT(elt, i);
501 			if (new_swslots[i] != 0)
502 				elt->count--;
503 			if (elt->count == 0) {
504 				LIST_REMOVE(elt, list);
505 				pool_put(&uao_swhash_elt_pool, elt);
506 			}
507 		}
508 	}
509 
510 	free(aobj->u_swhash, M_UVMAOBJ, 0);
511 
512 	aobj->u_swslots = new_swslots;
513 	aobj->u_pages = pages;
514 
515 	return 0;
516 }
517 
518 int
519 uao_shrink_array(struct uvm_object *uobj, int pages)
520 {
521 	struct uvm_aobj *aobj = (struct uvm_aobj *)uobj;
522 	int i, *new_swslots;
523 
524 	new_swslots = mallocarray(pages, sizeof(int), M_UVMAOBJ,
525 	    M_WAITOK | M_CANFAIL | M_ZERO);
526 	if (new_swslots == NULL)
527 		return ENOMEM;
528 
529 	uao_shrink_flush(uobj, pages, aobj->u_pages);
530 
531 	for (i = 0; i < pages; i++)
532 		new_swslots[i] = aobj->u_swslots[i];
533 
534 	free(aobj->u_swslots, M_UVMAOBJ, 0);
535 
536 	aobj->u_swslots = new_swslots;
537 	aobj->u_pages = pages;
538 
539 	return 0;
540 }
541 
542 int
543 uao_shrink(struct uvm_object *uobj, int pages)
544 {
545 	struct uvm_aobj *aobj = (struct uvm_aobj *)uobj;
546 
547 	KASSERT(pages < aobj->u_pages);
548 
549 	/*
550 	 * Distinguish between three possible cases:
551 	 * 1. aobj uses hash and must be converted to array.
552 	 * 2. aobj uses array and array size needs to be adjusted.
553 	 * 3. aobj uses hash and hash size needs to be adjusted.
554 	 */
555 	if (pages > UAO_SWHASH_THRESHOLD)
556 		return uao_shrink_hash(uobj, pages);	/* case 3 */
557 	else if (aobj->u_pages > UAO_SWHASH_THRESHOLD)
558 		return uao_shrink_convert(uobj, pages);	/* case 1 */
559 	else
560 		return uao_shrink_array(uobj, pages);	/* case 2 */
561 }
562 
563 /*
564  * Grow an aobj to a given number of pages. Right now we only adjust the swap
565  * slots. We could additionally handle page allocation directly, so that they
566  * don't happen through uvm_fault(). That would allow us to use another
567  * mechanism for the swap slots other than malloc(). It is thus mandatory that
568  * the caller of these functions does not allow faults to happen in case of
569  * growth error.
570  */
571 int
572 uao_grow_array(struct uvm_object *uobj, int pages)
573 {
574 	struct uvm_aobj *aobj = (struct uvm_aobj *)uobj;
575 	int i, *new_swslots;
576 
577 	KASSERT(aobj->u_pages <= UAO_SWHASH_THRESHOLD);
578 
579 	new_swslots = mallocarray(pages, sizeof(int), M_UVMAOBJ,
580 	    M_WAITOK | M_CANFAIL | M_ZERO);
581 	if (new_swslots == NULL)
582 		return ENOMEM;
583 
584 	for (i = 0; i < aobj->u_pages; i++)
585 		new_swslots[i] = aobj->u_swslots[i];
586 
587 	free(aobj->u_swslots, M_UVMAOBJ, 0);
588 
589 	aobj->u_swslots = new_swslots;
590 	aobj->u_pages = pages;
591 
592 	return 0;
593 }
594 
595 int
596 uao_grow_hash(struct uvm_object *uobj, int pages)
597 {
598 	struct uvm_aobj *aobj = (struct uvm_aobj *)uobj;
599 	struct uao_swhash *new_swhash;
600 	struct uao_swhash_elt *elt;
601 	unsigned long new_hashmask;
602 	int i;
603 
604 	KASSERT(pages > UAO_SWHASH_THRESHOLD);
605 
606 	/*
607 	 * If the size of the hash table doesn't change, all we need to do is
608 	 * to adjust the page count.
609 	 */
610 	if (UAO_SWHASH_BUCKETS(aobj->u_pages) == UAO_SWHASH_BUCKETS(pages)) {
611 		aobj->u_pages = pages;
612 		return 0;
613 	}
614 
615 	KASSERT(UAO_SWHASH_BUCKETS(aobj->u_pages) < UAO_SWHASH_BUCKETS(pages));
616 
617 	new_swhash = hashinit(UAO_SWHASH_BUCKETS(pages), M_UVMAOBJ,
618 	    M_WAITOK | M_CANFAIL, &new_hashmask);
619 	if (new_swhash == NULL)
620 		return ENOMEM;
621 
622 	for (i = 0; i < UAO_SWHASH_BUCKETS(aobj->u_pages); i++) {
623 		while (LIST_EMPTY(&aobj->u_swhash[i]) == 0) {
624 			elt = LIST_FIRST(&aobj->u_swhash[i]);
625 			LIST_REMOVE(elt, list);
626 			LIST_INSERT_HEAD(&new_swhash[i], elt, list);
627 		}
628 	}
629 
630 	free(aobj->u_swhash, M_UVMAOBJ, 0);
631 
632 	aobj->u_swhash = new_swhash;
633 	aobj->u_pages = pages;
634 	aobj->u_swhashmask = new_hashmask;
635 
636 	return 0;
637 }
638 
639 int
640 uao_grow_convert(struct uvm_object *uobj, int pages)
641 {
642 	struct uvm_aobj *aobj = (struct uvm_aobj *)uobj;
643 	struct uao_swhash *new_swhash;
644 	struct uao_swhash_elt *elt;
645 	unsigned long new_hashmask;
646 	int i, *old_swslots;
647 
648 	new_swhash = hashinit(UAO_SWHASH_BUCKETS(pages), M_UVMAOBJ,
649 	    M_WAITOK | M_CANFAIL, &new_hashmask);
650 	if (new_swhash == NULL)
651 		return ENOMEM;
652 
653 	/* Set these now, so we can use uao_find_swhash_elt(). */
654 	old_swslots = aobj->u_swslots;
655 	aobj->u_swhash = new_swhash;
656 	aobj->u_swhashmask = new_hashmask;
657 
658 	for (i = 0; i < aobj->u_pages; i++) {
659 		if (old_swslots[i] != 0) {
660 			elt = uao_find_swhash_elt(aobj, i, TRUE);
661 			elt->count++;
662 			UAO_SWHASH_ELT_PAGESLOT(elt, i) = old_swslots[i];
663 		}
664 	}
665 
666 	free(old_swslots, M_UVMAOBJ, 0);
667 	aobj->u_pages = pages;
668 
669 	return 0;
670 }
671 
672 int
673 uao_grow(struct uvm_object *uobj, int pages)
674 {
675 	struct uvm_aobj *aobj = (struct uvm_aobj *)uobj;
676 
677 	KASSERT(pages > aobj->u_pages);
678 
679 	/*
680 	 * Distinguish between three possible cases:
681 	 * 1. aobj uses hash and hash size needs to be adjusted.
682 	 * 2. aobj uses array and array size needs to be adjusted.
683 	 * 3. aobj uses array and must be converted to hash.
684 	 */
685 	if (pages <= UAO_SWHASH_THRESHOLD)
686 		return uao_grow_array(uobj, pages);	/* case 2 */
687 	else if (aobj->u_pages > UAO_SWHASH_THRESHOLD)
688 		return uao_grow_hash(uobj, pages);	/* case 1 */
689 	else
690 		return uao_grow_convert(uobj, pages);
691 }
692 
693 /*
694  * uao_create: create an aobj of the given size and return its uvm_object.
695  *
696  * => for normal use, flags are zero or UAO_FLAG_CANFAIL.
697  * => for the kernel object, the flags are:
698  *	UAO_FLAG_KERNOBJ - allocate the kernel object (can only happen once)
699  *	UAO_FLAG_KERNSWAP - enable swapping of kernel object ("           ")
700  */
701 struct uvm_object *
702 uao_create(vsize_t size, int flags)
703 {
704 	static struct uvm_aobj kernel_object_store; /* home of kernel_object */
705 	static int kobj_alloced = 0;			/* not allocated yet */
706 	int pages = round_page(size) >> PAGE_SHIFT;
707 	int refs = UVM_OBJ_KERN;
708 	int mflags;
709 	struct uvm_aobj *aobj;
710 
711 	/* malloc a new aobj unless we are asked for the kernel object */
712 	if (flags & UAO_FLAG_KERNOBJ) {		/* want kernel object? */
713 		if (kobj_alloced)
714 			panic("uao_create: kernel object already allocated");
715 
716 		aobj = &kernel_object_store;
717 		aobj->u_pages = pages;
718 		aobj->u_flags = UAO_FLAG_NOSWAP;	/* no swap to start */
719 		/* we are special, we never die */
720 		kobj_alloced = UAO_FLAG_KERNOBJ;
721 	} else if (flags & UAO_FLAG_KERNSWAP) {
722 		aobj = &kernel_object_store;
723 		if (kobj_alloced != UAO_FLAG_KERNOBJ)
724 		    panic("uao_create: asked to enable swap on kernel object");
725 		kobj_alloced = UAO_FLAG_KERNSWAP;
726 	} else {	/* normal object */
727 		aobj = pool_get(&uvm_aobj_pool, PR_WAITOK);
728 		aobj->u_pages = pages;
729 		aobj->u_flags = 0;		/* normal object */
730 		refs = 1;			/* normal object so 1 ref */
731 	}
732 
733 	/* allocate hash/array if necessary */
734  	if (flags == 0 || (flags & (UAO_FLAG_KERNSWAP | UAO_FLAG_CANFAIL))) {
735 		if (flags)
736 			mflags = M_NOWAIT;
737 		else
738 			mflags = M_WAITOK;
739 
740 		/* allocate hash table or array depending on object size */
741 		if (aobj->u_pages > UAO_SWHASH_THRESHOLD) {
742 			aobj->u_swhash = hashinit(UAO_SWHASH_BUCKETS(pages),
743 			    M_UVMAOBJ, mflags, &aobj->u_swhashmask);
744 			if (aobj->u_swhash == NULL) {
745 				if (flags & UAO_FLAG_CANFAIL) {
746 					pool_put(&uvm_aobj_pool, aobj);
747 					return (NULL);
748 				}
749 				panic("uao_create: hashinit swhash failed");
750 			}
751 		} else {
752 			aobj->u_swslots = mallocarray(pages, sizeof(int),
753 			    M_UVMAOBJ, mflags|M_ZERO);
754 			if (aobj->u_swslots == NULL) {
755 				if (flags & UAO_FLAG_CANFAIL) {
756 					pool_put(&uvm_aobj_pool, aobj);
757 					return (NULL);
758 				}
759 				panic("uao_create: malloc swslots failed");
760 			}
761 		}
762 
763 		if (flags & UAO_FLAG_KERNSWAP) {
764 			aobj->u_flags &= ~UAO_FLAG_NOSWAP; /* clear noswap */
765 			return(&aobj->u_obj);
766 			/* done! */
767 		}
768 	}
769 
770 	uvm_objinit(&aobj->u_obj, &aobj_pager, refs);
771 
772 	/* now that aobj is ready, add it to the global list */
773 	mtx_enter(&uao_list_lock);
774 	LIST_INSERT_HEAD(&uao_list, aobj, u_list);
775 	mtx_leave(&uao_list_lock);
776 
777 	return(&aobj->u_obj);
778 }
779 
780 
781 
782 /*
783  * uao_init: set up aobj pager subsystem
784  *
785  * => called at boot time from uvm_pager_init()
786  */
787 void
788 uao_init(void)
789 {
790 	static int uao_initialized;
791 
792 	if (uao_initialized)
793 		return;
794 	uao_initialized = TRUE;
795 
796 	/*
797 	 * NOTE: Pages for this pool must not come from a pageable
798 	 * kernel map!
799 	 */
800 	pool_init(&uao_swhash_elt_pool, sizeof(struct uao_swhash_elt),
801 	    0, 0, PR_WAITOK, "uaoeltpl", NULL);
802 	pool_setipl(&uao_swhash_elt_pool, IPL_NONE);
803 
804 	pool_init(&uvm_aobj_pool, sizeof(struct uvm_aobj), 0, 0, PR_WAITOK,
805 	    "aobjpl", NULL);
806 	pool_setipl(&uvm_aobj_pool, IPL_NONE);
807 }
808 
809 /*
810  * uao_reference: add a ref to an aobj
811  */
812 void
813 uao_reference(struct uvm_object *uobj)
814 {
815 	uao_reference_locked(uobj);
816 }
817 
818 /*
819  * uao_reference_locked: add a ref to an aobj
820  */
821 void
822 uao_reference_locked(struct uvm_object *uobj)
823 {
824 
825 	/* kernel_object already has plenty of references, leave it alone. */
826 	if (UVM_OBJ_IS_KERN_OBJECT(uobj))
827 		return;
828 
829 	uobj->uo_refs++;		/* bump! */
830 }
831 
832 
833 /*
834  * uao_detach: drop a reference to an aobj
835  */
836 void
837 uao_detach(struct uvm_object *uobj)
838 {
839 	uao_detach_locked(uobj);
840 }
841 
842 
843 /*
844  * uao_detach_locked: drop a reference to an aobj
845  *
846  * => aobj may freed upon return.
847  */
848 void
849 uao_detach_locked(struct uvm_object *uobj)
850 {
851 	struct uvm_aobj *aobj = (struct uvm_aobj *)uobj;
852 	struct vm_page *pg;
853 
854 	/* detaching from kernel_object is a noop. */
855 	if (UVM_OBJ_IS_KERN_OBJECT(uobj)) {
856 		return;
857 	}
858 
859 	uobj->uo_refs--;				/* drop ref! */
860 	if (uobj->uo_refs) {				/* still more refs? */
861 		return;
862 	}
863 
864 	/* remove the aobj from the global list. */
865 	mtx_enter(&uao_list_lock);
866 	LIST_REMOVE(aobj, u_list);
867 	mtx_leave(&uao_list_lock);
868 
869 	/*
870 	 * Free all pages left in the object. If they're busy, wait
871 	 * for them to become available before we kill it.
872 	 * Release swap resources then free the page.
873  	 */
874 	uvm_lock_pageq();
875 	while((pg = RB_ROOT(&uobj->memt)) != NULL) {
876 		if (pg->pg_flags & PG_BUSY) {
877 			atomic_setbits_int(&pg->pg_flags, PG_WANTED);
878 			uvm_unlock_pageq();
879 			UVM_WAIT(pg, 0, "uao_det", 0);
880 			uvm_lock_pageq();
881 			continue;
882 		}
883 		pmap_page_protect(pg, PROT_NONE);
884 		uao_dropswap(&aobj->u_obj, pg->offset >> PAGE_SHIFT);
885 		uvm_pagefree(pg);
886 	}
887 	uvm_unlock_pageq();
888 
889 	/* finally, free the rest. */
890 	uao_free(aobj);
891 }
892 
893 /*
894  * uao_flush: "flush" pages out of a uvm object
895  *
896  * => if PGO_CLEANIT is not set, then we will not block.
897  * => if PGO_ALLPAGE is set, then all pages in the object are valid targets
898  *	for flushing.
899  * => NOTE: we are allowed to lock the page queues, so the caller
900  *	must not be holding the lock on them [e.g. pagedaemon had
901  *	better not call us with the queues locked]
902  * => we return TRUE unless we encountered some sort of I/O error
903  *	XXXJRT currently never happens, as we never directly initiate
904  *	XXXJRT I/O
905  */
906 boolean_t
907 uao_flush(struct uvm_object *uobj, voff_t start, voff_t stop, int flags)
908 {
909 	struct uvm_aobj *aobj = (struct uvm_aobj *) uobj;
910 	struct vm_page *pp;
911 	voff_t curoff;
912 
913 	if (flags & PGO_ALLPAGES) {
914 		start = 0;
915 		stop = (voff_t)aobj->u_pages << PAGE_SHIFT;
916 	} else {
917 		start = trunc_page(start);
918 		stop = round_page(stop);
919 		if (stop > ((voff_t)aobj->u_pages << PAGE_SHIFT)) {
920 			printf("uao_flush: strange, got an out of range "
921 			    "flush (fixed)\n");
922 			stop = (voff_t)aobj->u_pages << PAGE_SHIFT;
923 		}
924 	}
925 
926 	/*
927 	 * Don't need to do any work here if we're not freeing
928 	 * or deactivating pages.
929 	 */
930 	if ((flags & (PGO_DEACTIVATE|PGO_FREE)) == 0)
931 		return (TRUE);
932 
933 	curoff = start;
934 	for (;;) {
935 		if (curoff < stop) {
936 			pp = uvm_pagelookup(uobj, curoff);
937 			curoff += PAGE_SIZE;
938 			if (pp == NULL)
939 				continue;
940 		} else {
941 			break;
942 		}
943 
944 		/* Make sure page is unbusy, else wait for it. */
945 		if (pp->pg_flags & PG_BUSY) {
946 			atomic_setbits_int(&pp->pg_flags, PG_WANTED);
947 			UVM_WAIT(pp, 0, "uaoflsh", 0);
948 			curoff -= PAGE_SIZE;
949 			continue;
950 		}
951 
952 		switch (flags & (PGO_CLEANIT|PGO_FREE|PGO_DEACTIVATE)) {
953 		/*
954 		 * XXX In these first 3 cases, we always just
955 		 * XXX deactivate the page.  We may want to
956 		 * XXX handle the different cases more specifically
957 		 * XXX in the future.
958 		 */
959 		case PGO_CLEANIT|PGO_FREE:
960 			/* FALLTHROUGH */
961 		case PGO_CLEANIT|PGO_DEACTIVATE:
962 			/* FALLTHROUGH */
963 		case PGO_DEACTIVATE:
964  deactivate_it:
965 			/* skip the page if it's wired */
966 			if (pp->wire_count != 0)
967 				continue;
968 
969 			uvm_lock_pageq();
970 			/* zap all mappings for the page. */
971 			pmap_page_protect(pp, PROT_NONE);
972 
973 			/* ...and deactivate the page. */
974 			uvm_pagedeactivate(pp);
975 			uvm_unlock_pageq();
976 
977 			continue;
978 		case PGO_FREE:
979 			/*
980 			 * If there are multiple references to
981 			 * the object, just deactivate the page.
982 			 */
983 			if (uobj->uo_refs > 1)
984 				goto deactivate_it;
985 
986 			/* XXX skip the page if it's wired */
987 			if (pp->wire_count != 0)
988 				continue;
989 
990 			/* zap all mappings for the page. */
991 			pmap_page_protect(pp, PROT_NONE);
992 
993 			uao_dropswap(uobj, pp->offset >> PAGE_SHIFT);
994 			uvm_lock_pageq();
995 			uvm_pagefree(pp);
996 			uvm_unlock_pageq();
997 
998 			continue;
999 		default:
1000 			panic("uao_flush: weird flags");
1001 		}
1002 	}
1003 
1004 	return (TRUE);
1005 }
1006 
1007 /*
1008  * uao_get: fetch me a page
1009  *
1010  * we have three cases:
1011  * 1: page is resident     -> just return the page.
1012  * 2: page is zero-fill    -> allocate a new page and zero it.
1013  * 3: page is swapped out  -> fetch the page from swap.
1014  *
1015  * cases 1 and 2 can be handled with PGO_LOCKED, case 3 cannot.
1016  * so, if the "center" page hits case 3 (or any page, with PGO_ALLPAGES),
1017  * then we will need to return VM_PAGER_UNLOCK.
1018  *
1019  * => flags: PGO_ALLPAGES: get all of the pages
1020  *           PGO_LOCKED: fault data structures are locked
1021  * => NOTE: offset is the offset of pps[0], _NOT_ pps[centeridx]
1022  * => NOTE: caller must check for released pages!!
1023  */
1024 static int
1025 uao_get(struct uvm_object *uobj, voff_t offset, struct vm_page **pps,
1026     int *npagesp, int centeridx, vm_prot_t access_type, int advice, int flags)
1027 {
1028 	struct uvm_aobj *aobj = (struct uvm_aobj *)uobj;
1029 	voff_t current_offset;
1030 	vm_page_t ptmp;
1031 	int lcv, gotpages, maxpages, swslot, rv, pageidx;
1032 	boolean_t done;
1033 
1034 	/* get number of pages */
1035 	maxpages = *npagesp;
1036 
1037 	/* step 1: handled the case where fault data structures are locked. */
1038 	if (flags & PGO_LOCKED) {
1039 		/* step 1a: get pages that are already resident. */
1040 
1041 		done = TRUE;	/* be optimistic */
1042 		gotpages = 0;	/* # of pages we got so far */
1043 
1044 		for (lcv = 0, current_offset = offset ; lcv < maxpages ;
1045 		    lcv++, current_offset += PAGE_SIZE) {
1046 			/* do we care about this page?  if not, skip it */
1047 			if (pps[lcv] == PGO_DONTCARE)
1048 				continue;
1049 
1050 			ptmp = uvm_pagelookup(uobj, current_offset);
1051 
1052 			/*
1053  			 * if page is new, attempt to allocate the page,
1054 			 * zero-fill'd.
1055  			 */
1056 			if (ptmp == NULL && uao_find_swslot(aobj,
1057 			    current_offset >> PAGE_SHIFT) == 0) {
1058 				ptmp = uvm_pagealloc(uobj, current_offset,
1059 				    NULL, UVM_PGA_ZERO);
1060 				if (ptmp) {
1061 					/* new page */
1062 					atomic_clearbits_int(&ptmp->pg_flags,
1063 					    PG_BUSY|PG_FAKE);
1064 					atomic_setbits_int(&ptmp->pg_flags,
1065 					    PQ_AOBJ);
1066 					UVM_PAGE_OWN(ptmp, NULL);
1067 				}
1068 			}
1069 
1070 			/* to be useful must get a non-busy page */
1071 			if (ptmp == NULL ||
1072 			    (ptmp->pg_flags & PG_BUSY) != 0) {
1073 				if (lcv == centeridx ||
1074 				    (flags & PGO_ALLPAGES) != 0)
1075 					/* need to do a wait or I/O! */
1076 					done = FALSE;
1077 				continue;
1078 			}
1079 
1080 			/*
1081 			 * useful page: busy it and plug it in our
1082 			 * result array
1083 			 */
1084 			/* caller must un-busy this page */
1085 			atomic_setbits_int(&ptmp->pg_flags, PG_BUSY);
1086 			UVM_PAGE_OWN(ptmp, "uao_get1");
1087 			pps[lcv] = ptmp;
1088 			gotpages++;
1089 
1090 		}
1091 
1092 		/*
1093  		 * step 1b: now we've either done everything needed or we
1094 		 * to unlock and do some waiting or I/O.
1095  		 */
1096 		*npagesp = gotpages;
1097 		if (done)
1098 			/* bingo! */
1099 			return(VM_PAGER_OK);
1100 		else
1101 			/* EEK!   Need to unlock and I/O */
1102 			return(VM_PAGER_UNLOCK);
1103 	}
1104 
1105 	/*
1106  	 * step 2: get non-resident or busy pages.
1107  	 * data structures are unlocked.
1108  	 */
1109 	for (lcv = 0, current_offset = offset ; lcv < maxpages ;
1110 	    lcv++, current_offset += PAGE_SIZE) {
1111 		/*
1112 		 * - skip over pages we've already gotten or don't want
1113 		 * - skip over pages we don't _have_ to get
1114 		 */
1115 		if (pps[lcv] != NULL ||
1116 		    (lcv != centeridx && (flags & PGO_ALLPAGES) == 0))
1117 			continue;
1118 
1119 		pageidx = current_offset >> PAGE_SHIFT;
1120 
1121 		/*
1122  		 * we have yet to locate the current page (pps[lcv]).   we
1123 		 * first look for a page that is already at the current offset.
1124 		 * if we find a page, we check to see if it is busy or
1125 		 * released.  if that is the case, then we sleep on the page
1126 		 * until it is no longer busy or released and repeat the lookup.
1127 		 * if the page we found is neither busy nor released, then we
1128 		 * busy it (so we own it) and plug it into pps[lcv].   this
1129 		 * 'break's the following while loop and indicates we are
1130 		 * ready to move on to the next page in the "lcv" loop above.
1131  		 *
1132  		 * if we exit the while loop with pps[lcv] still set to NULL,
1133 		 * then it means that we allocated a new busy/fake/clean page
1134 		 * ptmp in the object and we need to do I/O to fill in the data.
1135  		 */
1136 
1137 		/* top of "pps" while loop */
1138 		while (pps[lcv] == NULL) {
1139 			/* look for a resident page */
1140 			ptmp = uvm_pagelookup(uobj, current_offset);
1141 
1142 			/* not resident?   allocate one now (if we can) */
1143 			if (ptmp == NULL) {
1144 
1145 				ptmp = uvm_pagealloc(uobj, current_offset,
1146 				    NULL, 0);
1147 
1148 				/* out of RAM? */
1149 				if (ptmp == NULL) {
1150 					uvm_wait("uao_getpage");
1151 					/* goto top of pps while loop */
1152 					continue;
1153 				}
1154 
1155 				/*
1156 				 * safe with PQ's unlocked: because we just
1157 				 * alloc'd the page
1158 				 */
1159 				atomic_setbits_int(&ptmp->pg_flags, PQ_AOBJ);
1160 
1161 				/*
1162 				 * got new page ready for I/O.  break pps while
1163 				 * loop.  pps[lcv] is still NULL.
1164 				 */
1165 				break;
1166 			}
1167 
1168 			/* page is there, see if we need to wait on it */
1169 			if ((ptmp->pg_flags & PG_BUSY) != 0) {
1170 				atomic_setbits_int(&ptmp->pg_flags, PG_WANTED);
1171 				UVM_WAIT(ptmp, FALSE, "uao_get", 0);
1172 				continue;	/* goto top of pps while loop */
1173 			}
1174 
1175 			/*
1176  			 * if we get here then the page has become resident and
1177 			 * unbusy between steps 1 and 2.  we busy it now (so we
1178 			 * own it) and set pps[lcv] (so that we exit the while
1179 			 * loop).
1180  			 */
1181 			/* we own it, caller must un-busy */
1182 			atomic_setbits_int(&ptmp->pg_flags, PG_BUSY);
1183 			UVM_PAGE_OWN(ptmp, "uao_get2");
1184 			pps[lcv] = ptmp;
1185 		}
1186 
1187 		/*
1188  		 * if we own the valid page at the correct offset, pps[lcv] will
1189  		 * point to it.   nothing more to do except go to the next page.
1190  		 */
1191 		if (pps[lcv])
1192 			continue;			/* next lcv */
1193 
1194 		/*
1195  		 * we have a "fake/busy/clean" page that we just allocated.
1196  		 * do the needed "i/o", either reading from swap or zeroing.
1197  		 */
1198 		swslot = uao_find_swslot(aobj, pageidx);
1199 
1200 		/* just zero the page if there's nothing in swap.  */
1201 		if (swslot == 0) {
1202 			/* page hasn't existed before, just zero it. */
1203 			uvm_pagezero(ptmp);
1204 		} else {
1205 			/* page in the swapped-out page. */
1206 			rv = uvm_swap_get(ptmp, swslot, PGO_SYNCIO);
1207 
1208 			/* I/O done.  check for errors. */
1209 			if (rv != VM_PAGER_OK) {
1210 				/*
1211 				 * remove the swap slot from the aobj
1212 				 * and mark the aobj as having no real slot.
1213 				 * don't free the swap slot, thus preventing
1214 				 * it from being used again.
1215 				 */
1216 				swslot = uao_set_swslot(&aobj->u_obj, pageidx,
1217 							SWSLOT_BAD);
1218 				uvm_swap_markbad(swslot, 1);
1219 
1220 				if (ptmp->pg_flags & PG_WANTED)
1221 					wakeup(ptmp);
1222 				atomic_clearbits_int(&ptmp->pg_flags,
1223 				    PG_WANTED|PG_BUSY);
1224 				UVM_PAGE_OWN(ptmp, NULL);
1225 				uvm_lock_pageq();
1226 				uvm_pagefree(ptmp);
1227 				uvm_unlock_pageq();
1228 
1229 				return (rv);
1230 			}
1231 		}
1232 
1233 		/*
1234  		 * we got the page!   clear the fake flag (indicates valid
1235 		 * data now in page) and plug into our result array.   note
1236 		 * that page is still busy.
1237  		 *
1238  		 * it is the callers job to:
1239  		 * => check if the page is released
1240  		 * => unbusy the page
1241  		 * => activate the page
1242  		 */
1243 
1244 		/* data is valid ... */
1245 		atomic_clearbits_int(&ptmp->pg_flags, PG_FAKE);
1246 		pmap_clear_modify(ptmp);		/* ... and clean */
1247 		pps[lcv] = ptmp;
1248 
1249 	}	/* lcv loop */
1250 
1251 	return(VM_PAGER_OK);
1252 }
1253 
1254 /*
1255  * uao_dropswap:  release any swap resources from this aobj page.
1256  */
1257 int
1258 uao_dropswap(struct uvm_object *uobj, int pageidx)
1259 {
1260 	int slot;
1261 
1262 	slot = uao_set_swslot(uobj, pageidx, 0);
1263 	if (slot) {
1264 		uvm_swap_free(slot, 1);
1265 	}
1266 	return (slot);
1267 }
1268 
1269 /*
1270  * page in every page in every aobj that is paged-out to a range of swslots.
1271  *
1272  * => returns TRUE if pagein was aborted due to lack of memory.
1273  */
1274 boolean_t
1275 uao_swap_off(int startslot, int endslot)
1276 {
1277 	struct uvm_aobj *aobj, *nextaobj, *prevaobj = NULL;
1278 
1279 	/* walk the list of all aobjs. */
1280 	mtx_enter(&uao_list_lock);
1281 
1282 	for (aobj = LIST_FIRST(&uao_list);
1283 	     aobj != NULL;
1284 	     aobj = nextaobj) {
1285 		boolean_t rv;
1286 
1287 		/*
1288 		 * add a ref to the aobj so it doesn't disappear
1289 		 * while we're working.
1290 		 */
1291 		uao_reference_locked(&aobj->u_obj);
1292 
1293 		/*
1294 		 * now it's safe to unlock the uao list.
1295 		 * note that lock interleaving is alright with IPL_NONE mutexes.
1296 		 */
1297 		mtx_leave(&uao_list_lock);
1298 
1299 		if (prevaobj) {
1300 			uao_detach_locked(&prevaobj->u_obj);
1301 			prevaobj = NULL;
1302 		}
1303 
1304 		/*
1305 		 * page in any pages in the swslot range.
1306 		 * if there's an error, abort and return the error.
1307 		 */
1308 		rv = uao_pagein(aobj, startslot, endslot);
1309 		if (rv) {
1310 			uao_detach_locked(&aobj->u_obj);
1311 			return rv;
1312 		}
1313 
1314 		/*
1315 		 * we're done with this aobj.
1316 		 * relock the list and drop our ref on the aobj.
1317 		 */
1318 		mtx_enter(&uao_list_lock);
1319 		nextaobj = LIST_NEXT(aobj, u_list);
1320 		/*
1321 		 * prevaobj means that we have an object that we need
1322 		 * to drop a reference for. We can't just drop it now with
1323 		 * the list locked since that could cause lock recursion in
1324 		 * the case where we reduce the refcount to 0. It will be
1325 		 * released the next time we drop the list lock.
1326 		 */
1327 		prevaobj = aobj;
1328 	}
1329 
1330 	/* done with traversal, unlock the list */
1331 	mtx_leave(&uao_list_lock);
1332 	if (prevaobj) {
1333 		uao_detach_locked(&prevaobj->u_obj);
1334 	}
1335 	return FALSE;
1336 }
1337 
1338 /*
1339  * page in any pages from aobj in the given range.
1340  *
1341  * => returns TRUE if pagein was aborted due to lack of memory.
1342  */
1343 static boolean_t
1344 uao_pagein(struct uvm_aobj *aobj, int startslot, int endslot)
1345 {
1346 	boolean_t rv;
1347 
1348 	if (aobj->u_pages > UAO_SWHASH_THRESHOLD) {
1349 		struct uao_swhash_elt *elt;
1350 		int bucket;
1351 
1352 restart:
1353 		for (bucket = aobj->u_swhashmask; bucket >= 0; bucket--) {
1354 			for (elt = LIST_FIRST(&aobj->u_swhash[bucket]);
1355 			     elt != NULL;
1356 			     elt = LIST_NEXT(elt, list)) {
1357 				int i;
1358 
1359 				for (i = 0; i < UAO_SWHASH_CLUSTER_SIZE; i++) {
1360 					int slot = elt->slots[i];
1361 
1362 					/* if slot isn't in range, skip it. */
1363 					if (slot < startslot ||
1364 					    slot >= endslot) {
1365 						continue;
1366 					}
1367 
1368 					/*
1369 					 * process the page,
1370 					 * the start over on this object
1371 					 * since the swhash elt
1372 					 * may have been freed.
1373 					 */
1374 					rv = uao_pagein_page(aobj,
1375 					  UAO_SWHASH_ELT_PAGEIDX_BASE(elt) + i);
1376 					if (rv) {
1377 						return rv;
1378 					}
1379 					goto restart;
1380 				}
1381 			}
1382 		}
1383 	} else {
1384 		int i;
1385 
1386 		for (i = 0; i < aobj->u_pages; i++) {
1387 			int slot = aobj->u_swslots[i];
1388 
1389 			/* if the slot isn't in range, skip it */
1390 			if (slot < startslot || slot >= endslot) {
1391 				continue;
1392 			}
1393 
1394 			/* process the page.  */
1395 			rv = uao_pagein_page(aobj, i);
1396 			if (rv) {
1397 				return rv;
1398 			}
1399 		}
1400 	}
1401 
1402 	return FALSE;
1403 }
1404 
1405 /*
1406  * page in a page from an aobj.  used for swap_off.
1407  * returns TRUE if pagein was aborted due to lack of memory.
1408  */
1409 static boolean_t
1410 uao_pagein_page(struct uvm_aobj *aobj, int pageidx)
1411 {
1412 	struct vm_page *pg;
1413 	int rv, slot, npages;
1414 
1415 	pg = NULL;
1416 	npages = 1;
1417 	rv = uao_get(&aobj->u_obj, (voff_t)pageidx << PAGE_SHIFT,
1418 	    &pg, &npages, 0, PROT_READ | PROT_WRITE, 0, 0);
1419 
1420 	switch (rv) {
1421 	case VM_PAGER_OK:
1422 		break;
1423 
1424 	case VM_PAGER_ERROR:
1425 	case VM_PAGER_REFAULT:
1426 		/*
1427 		 * nothing more to do on errors.
1428 		 * VM_PAGER_REFAULT can only mean that the anon was freed,
1429 		 * so again there's nothing to do.
1430 		 */
1431 		return FALSE;
1432 	}
1433 
1434 	/*
1435 	 * ok, we've got the page now.
1436 	 * mark it as dirty, clear its swslot and un-busy it.
1437 	 */
1438 	slot = uao_set_swslot(&aobj->u_obj, pageidx, 0);
1439 	uvm_swap_free(slot, 1);
1440 	atomic_clearbits_int(&pg->pg_flags, PG_BUSY|PG_CLEAN|PG_FAKE);
1441 	UVM_PAGE_OWN(pg, NULL);
1442 
1443 	/* deactivate the page (to put it on a page queue). */
1444 	pmap_clear_reference(pg);
1445 	uvm_lock_pageq();
1446 	uvm_pagedeactivate(pg);
1447 	uvm_unlock_pageq();
1448 
1449 	return FALSE;
1450 }
1451 
1452 /*
1453  * XXX pedro: Once we are comfortable enough with this function, we can adapt
1454  * uao_free() to use it.
1455  *
1456  * uao_dropswap_range: drop swapslots in the range.
1457  *
1458  * => aobj must be locked and is returned locked.
1459  * => start is inclusive.  end is exclusive.
1460  */
1461 void
1462 uao_dropswap_range(struct uvm_object *uobj, voff_t start, voff_t end)
1463 {
1464 	struct uvm_aobj *aobj = (struct uvm_aobj *)uobj;
1465 	int swpgonlydelta = 0;
1466 
1467 	/* KASSERT(mutex_owned(uobj->vmobjlock)); */
1468 
1469 	if (end == 0) {
1470 		end = INT64_MAX;
1471 	}
1472 
1473 	if (aobj->u_pages > UAO_SWHASH_THRESHOLD) {
1474 		int i, hashbuckets = aobj->u_swhashmask + 1;
1475 		voff_t taghi;
1476 		voff_t taglo;
1477 
1478 		taglo = UAO_SWHASH_ELT_TAG(start);
1479 		taghi = UAO_SWHASH_ELT_TAG(end);
1480 
1481 		for (i = 0; i < hashbuckets; i++) {
1482 			struct uao_swhash_elt *elt, *next;
1483 
1484 			for (elt = LIST_FIRST(&aobj->u_swhash[i]);
1485 			     elt != NULL;
1486 			     elt = next) {
1487 				int startidx, endidx;
1488 				int j;
1489 
1490 				next = LIST_NEXT(elt, list);
1491 
1492 				if (elt->tag < taglo || taghi < elt->tag) {
1493 					continue;
1494 				}
1495 
1496 				if (elt->tag == taglo) {
1497 					startidx =
1498 					    UAO_SWHASH_ELT_PAGESLOT_IDX(start);
1499 				} else {
1500 					startidx = 0;
1501 				}
1502 
1503 				if (elt->tag == taghi) {
1504 					endidx =
1505 					    UAO_SWHASH_ELT_PAGESLOT_IDX(end);
1506 				} else {
1507 					endidx = UAO_SWHASH_CLUSTER_SIZE;
1508 				}
1509 
1510 				for (j = startidx; j < endidx; j++) {
1511 					int slot = elt->slots[j];
1512 
1513 					KASSERT(uvm_pagelookup(&aobj->u_obj,
1514 					    (voff_t)(UAO_SWHASH_ELT_PAGEIDX_BASE(elt)
1515 					    + j) << PAGE_SHIFT) == NULL);
1516 
1517 					if (slot > 0) {
1518 						uvm_swap_free(slot, 1);
1519 						swpgonlydelta++;
1520 						KASSERT(elt->count > 0);
1521 						elt->slots[j] = 0;
1522 						elt->count--;
1523 					}
1524 				}
1525 
1526 				if (elt->count == 0) {
1527 					LIST_REMOVE(elt, list);
1528 					pool_put(&uao_swhash_elt_pool, elt);
1529 				}
1530 			}
1531 		}
1532 	} else {
1533 		int i;
1534 
1535 		if (aobj->u_pages < end) {
1536 			end = aobj->u_pages;
1537 		}
1538 		for (i = start; i < end; i++) {
1539 			int slot = aobj->u_swslots[i];
1540 
1541 			if (slot > 0) {
1542 				uvm_swap_free(slot, 1);
1543 				swpgonlydelta++;
1544 			}
1545 		}
1546 	}
1547 
1548 	/*
1549 	 * adjust the counter of pages only in swap for all
1550 	 * the swap slots we've freed.
1551 	 */
1552 	if (swpgonlydelta > 0) {
1553 		KASSERT(uvmexp.swpgonly >= swpgonlydelta);
1554 		uvmexp.swpgonly -= swpgonlydelta;
1555 	}
1556 }
1557