xref: /openbsd-src/sys/uvm/uvm_aobj.c (revision 3374c67d44f9b75b98444cbf63020f777792342e)
1 /*	$OpenBSD: uvm_aobj.c,v 1.107 2022/08/29 02:58:13 jsg 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 anonymous UVM object (aobj) manages anonymous-memory.  In addition to
52  * keeping the list of resident pages, it may also keep a list of allocated
53  * swap blocks.  Depending on the size of the object, this list is either
54  * stored in an array (small objects) or in a hash table (large objects).
55  */
56 
57 /*
58  * Note: for hash tables, we break the address space of the aobj into blocks
59  * of UAO_SWHASH_CLUSTER_SIZE pages, which shall be a power of two.
60  */
61 #define	UAO_SWHASH_CLUSTER_SHIFT	4
62 #define	UAO_SWHASH_CLUSTER_SIZE		(1 << UAO_SWHASH_CLUSTER_SHIFT)
63 
64 /* Get the "tag" for this page index. */
65 #define	UAO_SWHASH_ELT_TAG(idx)		((idx) >> UAO_SWHASH_CLUSTER_SHIFT)
66 #define UAO_SWHASH_ELT_PAGESLOT_IDX(idx) \
67     ((idx) & (UAO_SWHASH_CLUSTER_SIZE - 1))
68 
69 /* Given an ELT and a page index, find the swap slot. */
70 #define	UAO_SWHASH_ELT_PAGESLOT(elt, idx) \
71     ((elt)->slots[UAO_SWHASH_ELT_PAGESLOT_IDX(idx)])
72 
73 /* Given an ELT, return its pageidx base. */
74 #define	UAO_SWHASH_ELT_PAGEIDX_BASE(elt) \
75     ((elt)->tag << UAO_SWHASH_CLUSTER_SHIFT)
76 
77 /* The hash function. */
78 #define	UAO_SWHASH_HASH(aobj, idx) \
79     (&(aobj)->u_swhash[(((idx) >> UAO_SWHASH_CLUSTER_SHIFT) \
80     & (aobj)->u_swhashmask)])
81 
82 /*
83  * The threshold which determines whether we will use an array or a
84  * hash table to store the list of allocated swap blocks.
85  */
86 #define	UAO_SWHASH_THRESHOLD		(UAO_SWHASH_CLUSTER_SIZE * 4)
87 #define	UAO_USES_SWHASH(aobj) \
88     ((aobj)->u_pages > UAO_SWHASH_THRESHOLD)
89 
90 /* The number of buckets in a hash, with an upper bound. */
91 #define	UAO_SWHASH_MAXBUCKETS		256
92 #define	UAO_SWHASH_BUCKETS(pages) \
93     (min((pages) >> UAO_SWHASH_CLUSTER_SHIFT, UAO_SWHASH_MAXBUCKETS))
94 
95 
96 /*
97  * uao_swhash_elt: when a hash table is being used, this structure defines
98  * the format of an entry in the bucket list.
99  */
100 struct uao_swhash_elt {
101 	LIST_ENTRY(uao_swhash_elt) list;	/* the hash list */
102 	voff_t tag;				/* our 'tag' */
103 	int count;				/* our number of active slots */
104 	int slots[UAO_SWHASH_CLUSTER_SIZE];	/* the slots */
105 };
106 
107 /*
108  * uao_swhash: the swap hash table structure
109  */
110 LIST_HEAD(uao_swhash, uao_swhash_elt);
111 
112 /*
113  * uao_swhash_elt_pool: pool of uao_swhash_elt structures
114  */
115 struct pool uao_swhash_elt_pool;
116 
117 /*
118  * uvm_aobj: the actual anon-backed uvm_object
119  *
120  * => the uvm_object is at the top of the structure, this allows
121  *   (struct uvm_aobj *) == (struct uvm_object *)
122  * => only one of u_swslots and u_swhash is used in any given aobj
123  */
124 struct uvm_aobj {
125 	struct uvm_object u_obj; /* has: pgops, memt, #pages, #refs */
126 	int u_pages;		 /* number of pages in entire object */
127 	int u_flags;		 /* the flags (see uvm_aobj.h) */
128 	/*
129 	 * Either an array or hashtable (array of bucket heads) of
130 	 * offset -> swapslot mappings for the aobj.
131 	 */
132 #define u_swslots	u_swap.slot_array
133 #define u_swhash	u_swap.slot_hash
134 	union swslots {
135 		int			*slot_array;
136 		struct uao_swhash	*slot_hash;
137 	} u_swap;
138 	u_long u_swhashmask;		/* mask for hashtable */
139 	LIST_ENTRY(uvm_aobj) u_list;	/* global list of aobjs */
140 };
141 
142 struct pool uvm_aobj_pool;
143 
144 static struct uao_swhash_elt	*uao_find_swhash_elt(struct uvm_aobj *, int,
145 				     boolean_t);
146 static int			 uao_find_swslot(struct uvm_object *, int);
147 static boolean_t		 uao_flush(struct uvm_object *, voff_t,
148 				     voff_t, int);
149 static void			 uao_free(struct uvm_aobj *);
150 static int			 uao_get(struct uvm_object *, voff_t,
151 				     vm_page_t *, int *, int, vm_prot_t,
152 				     int, int);
153 static boolean_t		 uao_pagein(struct uvm_aobj *, int, int);
154 static boolean_t		 uao_pagein_page(struct uvm_aobj *, int);
155 
156 void	uao_dropswap_range(struct uvm_object *, voff_t, voff_t);
157 void	uao_shrink_flush(struct uvm_object *, int, int);
158 int	uao_shrink_hash(struct uvm_object *, int);
159 int	uao_shrink_array(struct uvm_object *, int);
160 int	uao_shrink_convert(struct uvm_object *, int);
161 
162 int	uao_grow_hash(struct uvm_object *, int);
163 int	uao_grow_array(struct uvm_object *, int);
164 int	uao_grow_convert(struct uvm_object *, int);
165 
166 /*
167  * aobj_pager
168  *
169  * note that some functions (e.g. put) are handled elsewhere
170  */
171 const struct uvm_pagerops aobj_pager = {
172 	.pgo_reference = uao_reference,
173 	.pgo_detach = uao_detach,
174 	.pgo_flush = uao_flush,
175 	.pgo_get = uao_get,
176 };
177 
178 /*
179  * uao_list: global list of active aobjs, locked by uao_list_lock
180  *
181  * Lock ordering: generally the locking order is object lock, then list lock.
182  * in the case of swap off we have to iterate over the list, and thus the
183  * ordering is reversed. In that case we must use trylocking to prevent
184  * deadlock.
185  */
186 static LIST_HEAD(aobjlist, uvm_aobj) uao_list = LIST_HEAD_INITIALIZER(uao_list);
187 static struct mutex uao_list_lock = MUTEX_INITIALIZER(IPL_MPFLOOR);
188 
189 
190 /*
191  * functions
192  */
193 /*
194  * hash table/array related functions
195  */
196 /*
197  * uao_find_swhash_elt: find (or create) a hash table entry for a page
198  * offset.
199  */
200 static struct uao_swhash_elt *
201 uao_find_swhash_elt(struct uvm_aobj *aobj, int pageidx, boolean_t create)
202 {
203 	struct uao_swhash *swhash;
204 	struct uao_swhash_elt *elt;
205 	voff_t page_tag;
206 
207 	swhash = UAO_SWHASH_HASH(aobj, pageidx); /* first hash to get bucket */
208 	page_tag = UAO_SWHASH_ELT_TAG(pageidx);	/* tag to search for */
209 
210 	/*
211 	 * now search the bucket for the requested tag
212 	 */
213 	LIST_FOREACH(elt, swhash, list) {
214 		if (elt->tag == page_tag)
215 			return elt;
216 	}
217 
218 	if (!create)
219 		return NULL;
220 
221 	/*
222 	 * allocate a new entry for the bucket and init/insert it in
223 	 */
224 	elt = pool_get(&uao_swhash_elt_pool, PR_NOWAIT | PR_ZERO);
225 	/*
226 	 * XXX We cannot sleep here as the hash table might disappear
227 	 * from under our feet.  And we run the risk of deadlocking
228 	 * the pagedeamon.  In fact this code will only be called by
229 	 * the pagedaemon and allocation will only fail if we
230 	 * exhausted the pagedeamon reserve.  In that case we're
231 	 * doomed anyway, so panic.
232 	 */
233 	if (elt == NULL)
234 		panic("%s: can't allocate entry", __func__);
235 	LIST_INSERT_HEAD(swhash, elt, list);
236 	elt->tag = page_tag;
237 
238 	return elt;
239 }
240 
241 /*
242  * uao_find_swslot: find the swap slot number for an aobj/pageidx
243  */
244 static inline int
245 uao_find_swslot(struct uvm_object *uobj, int pageidx)
246 {
247 	struct uvm_aobj *aobj = (struct uvm_aobj *)uobj;
248 
249 	KASSERT(UVM_OBJ_IS_AOBJ(uobj));
250 
251 	/*
252 	 * if noswap flag is set, then we never return a slot
253 	 */
254 	if (aobj->u_flags & UAO_FLAG_NOSWAP)
255 		return 0;
256 
257 	/*
258 	 * if hashing, look in hash table.
259 	 */
260 	if (UAO_USES_SWHASH(aobj)) {
261 		struct uao_swhash_elt *elt =
262 		    uao_find_swhash_elt(aobj, pageidx, FALSE);
263 
264 		if (elt)
265 			return UAO_SWHASH_ELT_PAGESLOT(elt, pageidx);
266 		else
267 			return 0;
268 	}
269 
270 	/*
271 	 * otherwise, look in the array
272 	 */
273 	return aobj->u_swslots[pageidx];
274 }
275 
276 /*
277  * uao_set_swslot: set the swap slot for a page in an aobj.
278  *
279  * => setting a slot to zero frees the slot
280  * => object must be locked by caller
281  * => we return the old slot number, or -1 if we failed to allocate
282  *    memory to record the new slot number
283  */
284 int
285 uao_set_swslot(struct uvm_object *uobj, int pageidx, int slot)
286 {
287 	struct uvm_aobj *aobj = (struct uvm_aobj *)uobj;
288 	int oldslot;
289 
290 	KASSERT(rw_write_held(uobj->vmobjlock) || uobj->uo_refs == 0);
291 	KASSERT(UVM_OBJ_IS_AOBJ(uobj));
292 
293 	/*
294 	 * if noswap flag is set, then we can't set a slot
295 	 */
296 	if (aobj->u_flags & UAO_FLAG_NOSWAP) {
297 		if (slot == 0)
298 			return 0;		/* a clear is ok */
299 
300 		/* but a set is not */
301 		printf("uao_set_swslot: uobj = %p\n", uobj);
302 	    panic("uao_set_swslot: attempt to set a slot on a NOSWAP object");
303 	}
304 
305 	/*
306 	 * are we using a hash table?  if so, add it in the hash.
307 	 */
308 	if (UAO_USES_SWHASH(aobj)) {
309 		/*
310 		 * Avoid allocating an entry just to free it again if
311 		 * the page had not swap slot in the first place, and
312 		 * we are freeing.
313 		 */
314 		struct uao_swhash_elt *elt =
315 		    uao_find_swhash_elt(aobj, pageidx, slot ? TRUE : FALSE);
316 		if (elt == NULL) {
317 			KASSERT(slot == 0);
318 			return 0;
319 		}
320 
321 		oldslot = UAO_SWHASH_ELT_PAGESLOT(elt, pageidx);
322 		UAO_SWHASH_ELT_PAGESLOT(elt, pageidx) = slot;
323 
324 		/*
325 		 * now adjust the elt's reference counter and free it if we've
326 		 * dropped it to zero.
327 		 */
328 		if (slot) {
329 			if (oldslot == 0)
330 				elt->count++;
331 		} else {
332 			if (oldslot)
333 				elt->count--;
334 
335 			if (elt->count == 0) {
336 				LIST_REMOVE(elt, list);
337 				pool_put(&uao_swhash_elt_pool, elt);
338 			}
339 		}
340 	} else {
341 		/* we are using an array */
342 		oldslot = aobj->u_swslots[pageidx];
343 		aobj->u_swslots[pageidx] = slot;
344 	}
345 	return oldslot;
346 }
347 /*
348  * end of hash/array functions
349  */
350 
351 /*
352  * uao_free: free all resources held by an aobj, and then free the aobj
353  *
354  * => the aobj should be dead
355  */
356 static void
357 uao_free(struct uvm_aobj *aobj)
358 {
359 	struct uvm_object *uobj = &aobj->u_obj;
360 
361 	KASSERT(UVM_OBJ_IS_AOBJ(uobj));
362 	KASSERT(rw_write_held(uobj->vmobjlock));
363 	uao_dropswap_range(uobj, 0, 0);
364 	rw_exit(uobj->vmobjlock);
365 
366 	if (UAO_USES_SWHASH(aobj)) {
367 		/*
368 		 * free the hash table itself.
369 		 */
370 		hashfree(aobj->u_swhash, UAO_SWHASH_BUCKETS(aobj->u_pages), M_UVMAOBJ);
371 	} else {
372 		free(aobj->u_swslots, M_UVMAOBJ, aobj->u_pages * sizeof(int));
373 	}
374 
375 	/*
376 	 * finally free the aobj itself
377 	 */
378 	uvm_obj_destroy(uobj);
379 	pool_put(&uvm_aobj_pool, aobj);
380 }
381 
382 /*
383  * pager functions
384  */
385 
386 #ifdef TMPFS
387 /*
388  * Shrink an aobj to a given number of pages. The procedure is always the same:
389  * assess the necessity of data structure conversion (hash to array), secure
390  * resources, flush pages and drop swap slots.
391  *
392  */
393 
394 void
395 uao_shrink_flush(struct uvm_object *uobj, int startpg, int endpg)
396 {
397 	KASSERT(startpg < endpg);
398 	KASSERT(uobj->uo_refs == 1);
399 	uao_flush(uobj, (voff_t)startpg << PAGE_SHIFT,
400 	    (voff_t)endpg << PAGE_SHIFT, PGO_FREE);
401 	uao_dropswap_range(uobj, startpg, endpg);
402 }
403 
404 int
405 uao_shrink_hash(struct uvm_object *uobj, int pages)
406 {
407 	struct uvm_aobj *aobj = (struct uvm_aobj *)uobj;
408 	struct uao_swhash *new_swhash;
409 	struct uao_swhash_elt *elt;
410 	unsigned long new_hashmask;
411 	int i;
412 
413 	KASSERT(UAO_USES_SWHASH(aobj));
414 
415 	/*
416 	 * If the size of the hash table doesn't change, all we need to do is
417 	 * to adjust the page count.
418 	 */
419 	if (UAO_SWHASH_BUCKETS(aobj->u_pages) == UAO_SWHASH_BUCKETS(pages)) {
420 		uao_shrink_flush(uobj, pages, aobj->u_pages);
421 		aobj->u_pages = pages;
422 		return 0;
423 	}
424 
425 	new_swhash = hashinit(UAO_SWHASH_BUCKETS(pages), M_UVMAOBJ,
426 	    M_WAITOK | M_CANFAIL, &new_hashmask);
427 	if (new_swhash == NULL)
428 		return ENOMEM;
429 
430 	uao_shrink_flush(uobj, pages, aobj->u_pages);
431 
432 	/*
433 	 * Even though the hash table size is changing, the hash of the buckets
434 	 * we are interested in copying should not change.
435 	 */
436 	for (i = 0; i < UAO_SWHASH_BUCKETS(aobj->u_pages); i++) {
437 		while (LIST_EMPTY(&aobj->u_swhash[i]) == 0) {
438 			elt = LIST_FIRST(&aobj->u_swhash[i]);
439 			LIST_REMOVE(elt, list);
440 			LIST_INSERT_HEAD(&new_swhash[i], elt, list);
441 		}
442 	}
443 
444 	hashfree(aobj->u_swhash, UAO_SWHASH_BUCKETS(aobj->u_pages), M_UVMAOBJ);
445 
446 	aobj->u_swhash = new_swhash;
447 	aobj->u_pages = pages;
448 	aobj->u_swhashmask = new_hashmask;
449 
450 	return 0;
451 }
452 
453 int
454 uao_shrink_convert(struct uvm_object *uobj, int pages)
455 {
456 	struct uvm_aobj *aobj = (struct uvm_aobj *)uobj;
457 	struct uao_swhash_elt *elt;
458 	int i, *new_swslots;
459 
460 	new_swslots = mallocarray(pages, sizeof(int), M_UVMAOBJ,
461 	    M_WAITOK | M_CANFAIL | M_ZERO);
462 	if (new_swslots == NULL)
463 		return ENOMEM;
464 
465 	uao_shrink_flush(uobj, pages, aobj->u_pages);
466 
467 	/* Convert swap slots from hash to array.  */
468 	for (i = 0; i < pages; i++) {
469 		elt = uao_find_swhash_elt(aobj, i, FALSE);
470 		if (elt != NULL) {
471 			new_swslots[i] = UAO_SWHASH_ELT_PAGESLOT(elt, i);
472 			if (new_swslots[i] != 0)
473 				elt->count--;
474 			if (elt->count == 0) {
475 				LIST_REMOVE(elt, list);
476 				pool_put(&uao_swhash_elt_pool, elt);
477 			}
478 		}
479 	}
480 
481 	hashfree(aobj->u_swhash, UAO_SWHASH_BUCKETS(aobj->u_pages), M_UVMAOBJ);
482 
483 	aobj->u_swslots = new_swslots;
484 	aobj->u_pages = pages;
485 
486 	return 0;
487 }
488 
489 int
490 uao_shrink_array(struct uvm_object *uobj, int pages)
491 {
492 	struct uvm_aobj *aobj = (struct uvm_aobj *)uobj;
493 	int i, *new_swslots;
494 
495 	new_swslots = mallocarray(pages, sizeof(int), M_UVMAOBJ,
496 	    M_WAITOK | M_CANFAIL | M_ZERO);
497 	if (new_swslots == NULL)
498 		return ENOMEM;
499 
500 	uao_shrink_flush(uobj, pages, aobj->u_pages);
501 
502 	for (i = 0; i < pages; i++)
503 		new_swslots[i] = aobj->u_swslots[i];
504 
505 	free(aobj->u_swslots, M_UVMAOBJ, aobj->u_pages * sizeof(int));
506 
507 	aobj->u_swslots = new_swslots;
508 	aobj->u_pages = pages;
509 
510 	return 0;
511 }
512 
513 int
514 uao_shrink(struct uvm_object *uobj, int pages)
515 {
516 	struct uvm_aobj *aobj = (struct uvm_aobj *)uobj;
517 
518 	KASSERT(pages < aobj->u_pages);
519 
520 	/*
521 	 * Distinguish between three possible cases:
522 	 * 1. aobj uses hash and must be converted to array.
523 	 * 2. aobj uses array and array size needs to be adjusted.
524 	 * 3. aobj uses hash and hash size needs to be adjusted.
525 	 */
526 	if (pages > UAO_SWHASH_THRESHOLD)
527 		return uao_shrink_hash(uobj, pages);	/* case 3 */
528 	else if (aobj->u_pages > UAO_SWHASH_THRESHOLD)
529 		return uao_shrink_convert(uobj, pages);	/* case 1 */
530 	else
531 		return uao_shrink_array(uobj, pages);	/* case 2 */
532 }
533 
534 /*
535  * Grow an aobj to a given number of pages. Right now we only adjust the swap
536  * slots. We could additionally handle page allocation directly, so that they
537  * don't happen through uvm_fault(). That would allow us to use another
538  * mechanism for the swap slots other than malloc(). It is thus mandatory that
539  * the caller of these functions does not allow faults to happen in case of
540  * growth error.
541  */
542 int
543 uao_grow_array(struct uvm_object *uobj, int pages)
544 {
545 	struct uvm_aobj *aobj = (struct uvm_aobj *)uobj;
546 	int i, *new_swslots;
547 
548 	KASSERT(aobj->u_pages <= UAO_SWHASH_THRESHOLD);
549 
550 	new_swslots = mallocarray(pages, sizeof(int), M_UVMAOBJ,
551 	    M_WAITOK | M_CANFAIL | M_ZERO);
552 	if (new_swslots == NULL)
553 		return ENOMEM;
554 
555 	for (i = 0; i < aobj->u_pages; i++)
556 		new_swslots[i] = aobj->u_swslots[i];
557 
558 	free(aobj->u_swslots, M_UVMAOBJ, aobj->u_pages * sizeof(int));
559 
560 	aobj->u_swslots = new_swslots;
561 	aobj->u_pages = pages;
562 
563 	return 0;
564 }
565 
566 int
567 uao_grow_hash(struct uvm_object *uobj, int pages)
568 {
569 	struct uvm_aobj *aobj = (struct uvm_aobj *)uobj;
570 	struct uao_swhash *new_swhash;
571 	struct uao_swhash_elt *elt;
572 	unsigned long new_hashmask;
573 	int i;
574 
575 	KASSERT(pages > UAO_SWHASH_THRESHOLD);
576 
577 	/*
578 	 * If the size of the hash table doesn't change, all we need to do is
579 	 * to adjust the page count.
580 	 */
581 	if (UAO_SWHASH_BUCKETS(aobj->u_pages) == UAO_SWHASH_BUCKETS(pages)) {
582 		aobj->u_pages = pages;
583 		return 0;
584 	}
585 
586 	KASSERT(UAO_SWHASH_BUCKETS(aobj->u_pages) < UAO_SWHASH_BUCKETS(pages));
587 
588 	new_swhash = hashinit(UAO_SWHASH_BUCKETS(pages), M_UVMAOBJ,
589 	    M_WAITOK | M_CANFAIL, &new_hashmask);
590 	if (new_swhash == NULL)
591 		return ENOMEM;
592 
593 	for (i = 0; i < UAO_SWHASH_BUCKETS(aobj->u_pages); i++) {
594 		while (LIST_EMPTY(&aobj->u_swhash[i]) == 0) {
595 			elt = LIST_FIRST(&aobj->u_swhash[i]);
596 			LIST_REMOVE(elt, list);
597 			LIST_INSERT_HEAD(&new_swhash[i], elt, list);
598 		}
599 	}
600 
601 	hashfree(aobj->u_swhash, UAO_SWHASH_BUCKETS(aobj->u_pages), M_UVMAOBJ);
602 
603 	aobj->u_swhash = new_swhash;
604 	aobj->u_pages = pages;
605 	aobj->u_swhashmask = new_hashmask;
606 
607 	return 0;
608 }
609 
610 int
611 uao_grow_convert(struct uvm_object *uobj, int pages)
612 {
613 	struct uvm_aobj *aobj = (struct uvm_aobj *)uobj;
614 	struct uao_swhash *new_swhash;
615 	struct uao_swhash_elt *elt;
616 	unsigned long new_hashmask;
617 	int i, *old_swslots;
618 
619 	new_swhash = hashinit(UAO_SWHASH_BUCKETS(pages), M_UVMAOBJ,
620 	    M_WAITOK | M_CANFAIL, &new_hashmask);
621 	if (new_swhash == NULL)
622 		return ENOMEM;
623 
624 	/* Set these now, so we can use uao_find_swhash_elt(). */
625 	old_swslots = aobj->u_swslots;
626 	aobj->u_swhash = new_swhash;
627 	aobj->u_swhashmask = new_hashmask;
628 
629 	for (i = 0; i < aobj->u_pages; i++) {
630 		if (old_swslots[i] != 0) {
631 			elt = uao_find_swhash_elt(aobj, i, TRUE);
632 			elt->count++;
633 			UAO_SWHASH_ELT_PAGESLOT(elt, i) = old_swslots[i];
634 		}
635 	}
636 
637 	free(old_swslots, M_UVMAOBJ, aobj->u_pages * sizeof(int));
638 	aobj->u_pages = pages;
639 
640 	return 0;
641 }
642 
643 int
644 uao_grow(struct uvm_object *uobj, int pages)
645 {
646 	struct uvm_aobj *aobj = (struct uvm_aobj *)uobj;
647 
648 	KASSERT(pages > aobj->u_pages);
649 
650 	/*
651 	 * Distinguish between three possible cases:
652 	 * 1. aobj uses hash and hash size needs to be adjusted.
653 	 * 2. aobj uses array and array size needs to be adjusted.
654 	 * 3. aobj uses array and must be converted to hash.
655 	 */
656 	if (pages <= UAO_SWHASH_THRESHOLD)
657 		return uao_grow_array(uobj, pages);	/* case 2 */
658 	else if (aobj->u_pages > UAO_SWHASH_THRESHOLD)
659 		return uao_grow_hash(uobj, pages);	/* case 1 */
660 	else
661 		return uao_grow_convert(uobj, pages);
662 }
663 #endif /* TMPFS */
664 
665 /*
666  * uao_create: create an aobj of the given size and return its uvm_object.
667  *
668  * => for normal use, flags are zero or UAO_FLAG_CANFAIL.
669  * => for the kernel object, the flags are:
670  *	UAO_FLAG_KERNOBJ - allocate the kernel object (can only happen once)
671  *	UAO_FLAG_KERNSWAP - enable swapping of kernel object ("           ")
672  */
673 struct uvm_object *
674 uao_create(vsize_t size, int flags)
675 {
676 	static struct uvm_aobj kernel_object_store;
677 	static struct rwlock bootstrap_kernel_object_lock;
678 	static int kobj_alloced = 0;
679 	int pages = round_page(size) >> PAGE_SHIFT;
680 	struct uvm_aobj *aobj;
681 	int refs;
682 
683 	/*
684 	 * Allocate a new aobj, unless kernel object is requested.
685 	 */
686 	if (flags & UAO_FLAG_KERNOBJ) {
687 		KASSERT(!kobj_alloced);
688 		aobj = &kernel_object_store;
689 		aobj->u_pages = pages;
690 		aobj->u_flags = UAO_FLAG_NOSWAP;
691 		refs = UVM_OBJ_KERN;
692 		kobj_alloced = UAO_FLAG_KERNOBJ;
693 	} else if (flags & UAO_FLAG_KERNSWAP) {
694 		KASSERT(kobj_alloced == UAO_FLAG_KERNOBJ);
695 		aobj = &kernel_object_store;
696 		kobj_alloced = UAO_FLAG_KERNSWAP;
697 	} else {
698 		aobj = pool_get(&uvm_aobj_pool, PR_WAITOK);
699 		aobj->u_pages = pages;
700 		aobj->u_flags = 0;
701 		refs = 1;
702 	}
703 
704 	/*
705 	 * allocate hash/array if necessary
706 	 */
707  	if (flags == 0 || (flags & (UAO_FLAG_KERNSWAP | UAO_FLAG_CANFAIL))) {
708 		int mflags;
709 
710 		if (flags)
711 			mflags = M_NOWAIT;
712 		else
713 			mflags = M_WAITOK;
714 
715 		/* allocate hash table or array depending on object size */
716 		if (UAO_USES_SWHASH(aobj)) {
717 			aobj->u_swhash = hashinit(UAO_SWHASH_BUCKETS(pages),
718 			    M_UVMAOBJ, mflags, &aobj->u_swhashmask);
719 			if (aobj->u_swhash == NULL) {
720 				if (flags & UAO_FLAG_CANFAIL) {
721 					pool_put(&uvm_aobj_pool, aobj);
722 					return NULL;
723 				}
724 				panic("uao_create: hashinit swhash failed");
725 			}
726 		} else {
727 			aobj->u_swslots = mallocarray(pages, sizeof(int),
728 			    M_UVMAOBJ, mflags|M_ZERO);
729 			if (aobj->u_swslots == NULL) {
730 				if (flags & UAO_FLAG_CANFAIL) {
731 					pool_put(&uvm_aobj_pool, aobj);
732 					return NULL;
733 				}
734 				panic("uao_create: malloc swslots failed");
735 			}
736 		}
737 
738 		if (flags & UAO_FLAG_KERNSWAP) {
739 			aobj->u_flags &= ~UAO_FLAG_NOSWAP; /* clear noswap */
740 			return &aobj->u_obj;
741 			/* done! */
742 		}
743 	}
744 
745 	/*
746 	 * Initialise UVM object.
747 	 */
748 	uvm_obj_init(&aobj->u_obj, &aobj_pager, refs);
749 	if (flags & UAO_FLAG_KERNOBJ) {
750 		/* Use a temporary static lock for kernel_object. */
751 		rw_init(&bootstrap_kernel_object_lock, "kobjlk");
752 		uvm_obj_setlock(&aobj->u_obj, &bootstrap_kernel_object_lock);
753 	}
754 
755 	/*
756  	 * now that aobj is ready, add it to the global list
757  	 */
758 	mtx_enter(&uao_list_lock);
759 	LIST_INSERT_HEAD(&uao_list, aobj, u_list);
760 	mtx_leave(&uao_list_lock);
761 
762 	return &aobj->u_obj;
763 }
764 
765 
766 
767 /*
768  * uao_init: set up aobj pager subsystem
769  *
770  * => called at boot time from uvm_pager_init()
771  */
772 void
773 uao_init(void)
774 {
775 	/*
776 	 * NOTE: Pages for this pool must not come from a pageable
777 	 * kernel map!
778 	 */
779 	pool_init(&uao_swhash_elt_pool, sizeof(struct uao_swhash_elt), 0,
780 	    IPL_NONE, PR_WAITOK, "uaoeltpl", NULL);
781 	pool_init(&uvm_aobj_pool, sizeof(struct uvm_aobj), 0,
782 	    IPL_NONE, PR_WAITOK, "aobjpl", NULL);
783 }
784 
785 /*
786  * uao_reference: hold a reference to an anonymous UVM object.
787  */
788 void
789 uao_reference(struct uvm_object *uobj)
790 {
791 	/* Kernel object is persistent. */
792 	if (UVM_OBJ_IS_KERN_OBJECT(uobj))
793 		return;
794 
795 	atomic_inc_int(&uobj->uo_refs);
796 }
797 
798 
799 /*
800  * uao_detach: drop a reference to an anonymous UVM object.
801  */
802 void
803 uao_detach(struct uvm_object *uobj)
804 {
805 	struct uvm_aobj *aobj = (struct uvm_aobj *)uobj;
806 	struct vm_page *pg;
807 
808 	/*
809 	 * Detaching from kernel_object is a NOP.
810 	 */
811 	if (UVM_OBJ_IS_KERN_OBJECT(uobj))
812 		return;
813 
814 	/*
815 	 * Drop the reference.  If it was the last one, destroy the object.
816 	 */
817 	if (atomic_dec_int_nv(&uobj->uo_refs) > 0) {
818 		return;
819 	}
820 
821 	/*
822 	 * Remove the aobj from the global list.
823 	 */
824 	mtx_enter(&uao_list_lock);
825 	LIST_REMOVE(aobj, u_list);
826 	mtx_leave(&uao_list_lock);
827 
828 	/*
829 	 * Free all the pages left in the aobj.  For each page, when the
830 	 * page is no longer busy (and thus after any disk I/O that it is
831 	 * involved in is complete), release any swap resources and free
832 	 * the page itself.
833 	 */
834 	rw_enter(uobj->vmobjlock, RW_WRITE);
835 	while ((pg = RBT_ROOT(uvm_objtree, &uobj->memt)) != NULL) {
836 		pmap_page_protect(pg, PROT_NONE);
837 		if (pg->pg_flags & PG_BUSY) {
838 			uvm_pagewait(pg, uobj->vmobjlock, "uao_det");
839 			rw_enter(uobj->vmobjlock, RW_WRITE);
840 			continue;
841 		}
842 		uao_dropswap(&aobj->u_obj, pg->offset >> PAGE_SHIFT);
843 		uvm_lock_pageq();
844 		uvm_pagefree(pg);
845 		uvm_unlock_pageq();
846 	}
847 
848 	/*
849 	 * Finally, free the anonymous UVM object itself.
850 	 */
851 	uao_free(aobj);
852 }
853 
854 /*
855  * uao_flush: flush pages out of a uvm object
856  *
857  * => if PGO_CLEANIT is not set, then we will not block.
858  * => if PGO_ALLPAGE is set, then all pages in the object are valid targets
859  *	for flushing.
860  * => NOTE: we are allowed to lock the page queues, so the caller
861  *	must not be holding the lock on them [e.g. pagedaemon had
862  *	better not call us with the queues locked]
863  * => we return TRUE unless we encountered some sort of I/O error
864  *	XXXJRT currently never happens, as we never directly initiate
865  *	XXXJRT I/O
866  */
867 boolean_t
868 uao_flush(struct uvm_object *uobj, voff_t start, voff_t stop, int flags)
869 {
870 	struct uvm_aobj *aobj = (struct uvm_aobj *) uobj;
871 	struct vm_page *pg;
872 	voff_t curoff;
873 
874 	KASSERT(UVM_OBJ_IS_AOBJ(uobj));
875 	KASSERT(rw_write_held(uobj->vmobjlock));
876 
877 	if (flags & PGO_ALLPAGES) {
878 		start = 0;
879 		stop = (voff_t)aobj->u_pages << PAGE_SHIFT;
880 	} else {
881 		start = trunc_page(start);
882 		stop = round_page(stop);
883 		if (stop > ((voff_t)aobj->u_pages << PAGE_SHIFT)) {
884 			printf("uao_flush: strange, got an out of range "
885 			    "flush (fixed)\n");
886 			stop = (voff_t)aobj->u_pages << PAGE_SHIFT;
887 		}
888 	}
889 
890 	/*
891 	 * Don't need to do any work here if we're not freeing
892 	 * or deactivating pages.
893 	 */
894 	if ((flags & (PGO_DEACTIVATE|PGO_FREE)) == 0) {
895 		return TRUE;
896 	}
897 
898 	curoff = start;
899 	for (;;) {
900 		if (curoff < stop) {
901 			pg = uvm_pagelookup(uobj, curoff);
902 			curoff += PAGE_SIZE;
903 			if (pg == NULL)
904 				continue;
905 		} else {
906 			break;
907 		}
908 
909 		/* Make sure page is unbusy, else wait for it. */
910 		if (pg->pg_flags & PG_BUSY) {
911 			uvm_pagewait(pg, uobj->vmobjlock, "uaoflsh");
912 			rw_enter(uobj->vmobjlock, RW_WRITE);
913 			curoff -= PAGE_SIZE;
914 			continue;
915 		}
916 
917 		switch (flags & (PGO_CLEANIT|PGO_FREE|PGO_DEACTIVATE)) {
918 		/*
919 		 * XXX In these first 3 cases, we always just
920 		 * XXX deactivate the page.  We may want to
921 		 * XXX handle the different cases more specifically
922 		 * XXX in the future.
923 		 */
924 		case PGO_CLEANIT|PGO_FREE:
925 			/* FALLTHROUGH */
926 		case PGO_CLEANIT|PGO_DEACTIVATE:
927 			/* FALLTHROUGH */
928 		case PGO_DEACTIVATE:
929  deactivate_it:
930 			if (pg->wire_count != 0)
931 				continue;
932 
933 			uvm_lock_pageq();
934 			pmap_page_protect(pg, PROT_NONE);
935 			uvm_pagedeactivate(pg);
936 			uvm_unlock_pageq();
937 
938 			continue;
939 		case PGO_FREE:
940 			/*
941 			 * If there are multiple references to
942 			 * the object, just deactivate the page.
943 			 */
944 			if (uobj->uo_refs > 1)
945 				goto deactivate_it;
946 
947 			/* XXX skip the page if it's wired */
948 			if (pg->wire_count != 0)
949 				continue;
950 
951 			/*
952 			 * free the swap slot and the page.
953 			 */
954 			pmap_page_protect(pg, PROT_NONE);
955 
956 			/*
957 			 * freeing swapslot here is not strictly necessary.
958 			 * however, leaving it here doesn't save much
959 			 * because we need to update swap accounting anyway.
960 			 */
961 			uao_dropswap(uobj, pg->offset >> PAGE_SHIFT);
962 			uvm_lock_pageq();
963 			uvm_pagefree(pg);
964 			uvm_unlock_pageq();
965 
966 			continue;
967 		default:
968 			panic("uao_flush: weird flags");
969 		}
970 	}
971 
972 	return TRUE;
973 }
974 
975 /*
976  * uao_get: fetch me a page
977  *
978  * we have three cases:
979  * 1: page is resident     -> just return the page.
980  * 2: page is zero-fill    -> allocate a new page and zero it.
981  * 3: page is swapped out  -> fetch the page from swap.
982  *
983  * cases 1 can be handled with PGO_LOCKED, cases 2 and 3 cannot.
984  * so, if the "center" page hits case 3 (or any page, with PGO_ALLPAGES),
985  * then we will need to return VM_PAGER_UNLOCK.
986  *
987  * => flags: PGO_ALLPAGES: get all of the pages
988  *           PGO_LOCKED: fault data structures are locked
989  * => NOTE: offset is the offset of pps[0], _NOT_ pps[centeridx]
990  * => NOTE: caller must check for released pages!!
991  */
992 static int
993 uao_get(struct uvm_object *uobj, voff_t offset, struct vm_page **pps,
994     int *npagesp, int centeridx, vm_prot_t access_type, int advice, int flags)
995 {
996 	struct uvm_aobj *aobj = (struct uvm_aobj *)uobj;
997 	voff_t current_offset;
998 	vm_page_t ptmp;
999 	int lcv, gotpages, maxpages, swslot, rv, pageidx;
1000 	boolean_t done;
1001 
1002 	KASSERT(UVM_OBJ_IS_AOBJ(uobj));
1003 	KASSERT(rw_write_held(uobj->vmobjlock));
1004 
1005 	/*
1006  	 * get number of pages
1007  	 */
1008 	maxpages = *npagesp;
1009 
1010 	if (flags & PGO_LOCKED) {
1011 		/*
1012  		 * step 1a: get pages that are already resident.   only do
1013 		 * this if the data structures are locked (i.e. the first
1014 		 * time through).
1015  		 */
1016 
1017 		done = TRUE;	/* be optimistic */
1018 		gotpages = 0;	/* # of pages we got so far */
1019 
1020 		for (lcv = 0, current_offset = offset ; lcv < maxpages ;
1021 		    lcv++, current_offset += PAGE_SIZE) {
1022 			/* do we care about this page?  if not, skip it */
1023 			if (pps[lcv] == PGO_DONTCARE)
1024 				continue;
1025 
1026 			ptmp = uvm_pagelookup(uobj, current_offset);
1027 
1028 			/*
1029  			 * if page is new, attempt to allocate the page,
1030 			 * zero-fill'd.
1031  			 */
1032 			if (ptmp == NULL && uao_find_swslot(uobj,
1033 			    current_offset >> PAGE_SHIFT) == 0) {
1034 				ptmp = uvm_pagealloc(uobj, current_offset,
1035 				    NULL, UVM_PGA_ZERO);
1036 				if (ptmp) {
1037 					/* new page */
1038 					atomic_clearbits_int(&ptmp->pg_flags,
1039 					    PG_BUSY|PG_FAKE);
1040 					atomic_setbits_int(&ptmp->pg_flags,
1041 					    PQ_AOBJ);
1042 					UVM_PAGE_OWN(ptmp, NULL);
1043 				}
1044 			}
1045 
1046 			/*
1047 			 * to be useful must get a non-busy page
1048 			 */
1049 			if (ptmp == NULL ||
1050 			    (ptmp->pg_flags & PG_BUSY) != 0) {
1051 				if (lcv == centeridx ||
1052 				    (flags & PGO_ALLPAGES) != 0)
1053 					/* need to do a wait or I/O! */
1054 					done = FALSE;
1055 				continue;
1056 			}
1057 
1058 			/*
1059 			 * useful page: plug it in our result array
1060 			 */
1061 			atomic_setbits_int(&ptmp->pg_flags, PG_BUSY);
1062 			UVM_PAGE_OWN(ptmp, "uao_get1");
1063 			pps[lcv] = ptmp;
1064 			gotpages++;
1065 
1066 		}
1067 
1068 		/*
1069  		 * step 1b: now we've either done everything needed or we
1070 		 * to unlock and do some waiting or I/O.
1071  		 */
1072 		*npagesp = gotpages;
1073 		if (done)
1074 			/* bingo! */
1075 			return VM_PAGER_OK;
1076 		else
1077 			/* EEK!   Need to unlock and I/O */
1078 			return VM_PAGER_UNLOCK;
1079 	}
1080 
1081 	/*
1082  	 * step 2: get non-resident or busy pages.
1083  	 * data structures are unlocked.
1084  	 */
1085 	for (lcv = 0, current_offset = offset ; lcv < maxpages ;
1086 	    lcv++, current_offset += PAGE_SIZE) {
1087 		/*
1088 		 * - skip over pages we've already gotten or don't want
1089 		 * - skip over pages we don't _have_ to get
1090 		 */
1091 		if (pps[lcv] != NULL ||
1092 		    (lcv != centeridx && (flags & PGO_ALLPAGES) == 0))
1093 			continue;
1094 
1095 		pageidx = current_offset >> PAGE_SHIFT;
1096 
1097 		/*
1098  		 * we have yet to locate the current page (pps[lcv]).   we
1099 		 * first look for a page that is already at the current offset.
1100 		 * if we find a page, we check to see if it is busy or
1101 		 * released.  if that is the case, then we sleep on the page
1102 		 * until it is no longer busy or released and repeat the lookup.
1103 		 * if the page we found is neither busy nor released, then we
1104 		 * busy it (so we own it) and plug it into pps[lcv].   this
1105 		 * 'break's the following while loop and indicates we are
1106 		 * ready to move on to the next page in the "lcv" loop above.
1107  		 *
1108  		 * if we exit the while loop with pps[lcv] still set to NULL,
1109 		 * then it means that we allocated a new busy/fake/clean page
1110 		 * ptmp in the object and we need to do I/O to fill in the data.
1111  		 */
1112 
1113 		/* top of "pps" while loop */
1114 		while (pps[lcv] == NULL) {
1115 			/* look for a resident page */
1116 			ptmp = uvm_pagelookup(uobj, current_offset);
1117 
1118 			/* not resident?   allocate one now (if we can) */
1119 			if (ptmp == NULL) {
1120 
1121 				ptmp = uvm_pagealloc(uobj, current_offset,
1122 				    NULL, 0);
1123 
1124 				/* out of RAM? */
1125 				if (ptmp == NULL) {
1126 					rw_exit(uobj->vmobjlock);
1127 					uvm_wait("uao_getpage");
1128 					rw_enter(uobj->vmobjlock, RW_WRITE);
1129 					/* goto top of pps while loop */
1130 					continue;
1131 				}
1132 
1133 				/*
1134 				 * safe with PQ's unlocked: because we just
1135 				 * alloc'd the page
1136 				 */
1137 				atomic_setbits_int(&ptmp->pg_flags, PQ_AOBJ);
1138 
1139 				/*
1140 				 * got new page ready for I/O.  break pps while
1141 				 * loop.  pps[lcv] is still NULL.
1142 				 */
1143 				break;
1144 			}
1145 
1146 			/* page is there, see if we need to wait on it */
1147 			if ((ptmp->pg_flags & PG_BUSY) != 0) {
1148 				uvm_pagewait(ptmp, uobj->vmobjlock, "uao_get");
1149 				rw_enter(uobj->vmobjlock, RW_WRITE);
1150 				continue;	/* goto top of pps while loop */
1151 			}
1152 
1153 			/*
1154  			 * if we get here then the page is resident and
1155 			 * unbusy.  we busy it now (so we own it).
1156  			 */
1157 			/* we own it, caller must un-busy */
1158 			atomic_setbits_int(&ptmp->pg_flags, PG_BUSY);
1159 			UVM_PAGE_OWN(ptmp, "uao_get2");
1160 			pps[lcv] = ptmp;
1161 		}
1162 
1163 		/*
1164  		 * if we own the valid page at the correct offset, pps[lcv] will
1165  		 * point to it.   nothing more to do except go to the next page.
1166  		 */
1167 		if (pps[lcv])
1168 			continue;			/* next lcv */
1169 
1170 		/*
1171  		 * we have a "fake/busy/clean" page that we just allocated.
1172  		 * do the needed "i/o", either reading from swap or zeroing.
1173  		 */
1174 		swslot = uao_find_swslot(uobj, pageidx);
1175 
1176 		/* just zero the page if there's nothing in swap.  */
1177 		if (swslot == 0) {
1178 			/* page hasn't existed before, just zero it. */
1179 			uvm_pagezero(ptmp);
1180 		} else {
1181 			/*
1182 			 * page in the swapped-out page.
1183 			 * unlock object for i/o, relock when done.
1184 			 */
1185 
1186 			rw_exit(uobj->vmobjlock);
1187 			rv = uvm_swap_get(ptmp, swslot, PGO_SYNCIO);
1188 			rw_enter(uobj->vmobjlock, RW_WRITE);
1189 
1190 			/*
1191 			 * I/O done.  check for errors.
1192 			 */
1193 			if (rv != VM_PAGER_OK) {
1194 				/*
1195 				 * remove the swap slot from the aobj
1196 				 * and mark the aobj as having no real slot.
1197 				 * don't free the swap slot, thus preventing
1198 				 * it from being used again.
1199 				 */
1200 				swslot = uao_set_swslot(&aobj->u_obj, pageidx,
1201 							SWSLOT_BAD);
1202 				uvm_swap_markbad(swslot, 1);
1203 
1204 				if (ptmp->pg_flags & PG_WANTED)
1205 					wakeup(ptmp);
1206 				atomic_clearbits_int(&ptmp->pg_flags,
1207 				    PG_WANTED|PG_BUSY);
1208 				UVM_PAGE_OWN(ptmp, NULL);
1209 				uvm_lock_pageq();
1210 				uvm_pagefree(ptmp);
1211 				uvm_unlock_pageq();
1212 				rw_exit(uobj->vmobjlock);
1213 
1214 				return rv;
1215 			}
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 		atomic_clearbits_int(&ptmp->pg_flags, PG_FAKE);
1229 		pmap_clear_modify(ptmp);		/* ... and clean */
1230 		pps[lcv] = ptmp;
1231 
1232 	}	/* lcv loop */
1233 
1234 	rw_exit(uobj->vmobjlock);
1235 	return VM_PAGER_OK;
1236 }
1237 
1238 /*
1239  * uao_dropswap:  release any swap resources from this aobj page.
1240  *
1241  * => aobj must be locked or have a reference count of 0.
1242  */
1243 int
1244 uao_dropswap(struct uvm_object *uobj, int pageidx)
1245 {
1246 	int slot;
1247 
1248 	KASSERT(UVM_OBJ_IS_AOBJ(uobj));
1249 
1250 	slot = uao_set_swslot(uobj, pageidx, 0);
1251 	if (slot) {
1252 		uvm_swap_free(slot, 1);
1253 	}
1254 	return slot;
1255 }
1256 
1257 /*
1258  * page in every page in every aobj that is paged-out to a range of swslots.
1259  *
1260  * => aobj must be locked and is returned locked.
1261  * => returns TRUE if pagein was aborted due to lack of memory.
1262  */
1263 boolean_t
1264 uao_swap_off(int startslot, int endslot)
1265 {
1266 	struct uvm_aobj *aobj;
1267 
1268 	/*
1269 	 * Walk the list of all anonymous UVM objects.  Grab the first.
1270 	 */
1271 	mtx_enter(&uao_list_lock);
1272 	if ((aobj = LIST_FIRST(&uao_list)) == NULL) {
1273 		mtx_leave(&uao_list_lock);
1274 		return FALSE;
1275 	}
1276 	uao_reference(&aobj->u_obj);
1277 
1278 	do {
1279 		struct uvm_aobj *nextaobj;
1280 		boolean_t rv;
1281 
1282 		/*
1283 		 * Prefetch the next object and immediately hold a reference
1284 		 * on it, so neither the current nor the next entry could
1285 		 * disappear while we are iterating.
1286 		 */
1287 		if ((nextaobj = LIST_NEXT(aobj, u_list)) != NULL) {
1288 			uao_reference(&nextaobj->u_obj);
1289 		}
1290 		mtx_leave(&uao_list_lock);
1291 
1292 		/*
1293 		 * Page in all pages in the swap slot range.
1294 		 */
1295 		rw_enter(aobj->u_obj.vmobjlock, RW_WRITE);
1296 		rv = uao_pagein(aobj, startslot, endslot);
1297 		rw_exit(aobj->u_obj.vmobjlock);
1298 
1299 		/* Drop the reference of the current object. */
1300 		uao_detach(&aobj->u_obj);
1301 		if (rv) {
1302 			if (nextaobj) {
1303 				uao_detach(&nextaobj->u_obj);
1304 			}
1305 			return rv;
1306 		}
1307 
1308 		aobj = nextaobj;
1309 		mtx_enter(&uao_list_lock);
1310 	} while (aobj);
1311 
1312 	/*
1313 	 * done with traversal, unlock the list
1314 	 */
1315 	mtx_leave(&uao_list_lock);
1316 	return FALSE;
1317 }
1318 
1319 /*
1320  * page in any pages from aobj in the given range.
1321  *
1322  * => returns TRUE if pagein was aborted due to lack of memory.
1323  */
1324 static boolean_t
1325 uao_pagein(struct uvm_aobj *aobj, int startslot, int endslot)
1326 {
1327 	boolean_t rv;
1328 
1329 	if (UAO_USES_SWHASH(aobj)) {
1330 		struct uao_swhash_elt *elt;
1331 		int bucket;
1332 
1333 restart:
1334 		for (bucket = aobj->u_swhashmask; bucket >= 0; bucket--) {
1335 			for (elt = LIST_FIRST(&aobj->u_swhash[bucket]);
1336 			     elt != NULL;
1337 			     elt = LIST_NEXT(elt, list)) {
1338 				int i;
1339 
1340 				for (i = 0; i < UAO_SWHASH_CLUSTER_SIZE; i++) {
1341 					int slot = elt->slots[i];
1342 
1343 					/*
1344 					 * if the slot isn't in range, skip it.
1345 					 */
1346 					if (slot < startslot ||
1347 					    slot >= endslot) {
1348 						continue;
1349 					}
1350 
1351 					/*
1352 					 * process the page,
1353 					 * the start over on this object
1354 					 * since the swhash elt
1355 					 * may have been freed.
1356 					 */
1357 					rv = uao_pagein_page(aobj,
1358 					  UAO_SWHASH_ELT_PAGEIDX_BASE(elt) + i);
1359 					if (rv) {
1360 						return rv;
1361 					}
1362 					goto restart;
1363 				}
1364 			}
1365 		}
1366 	} else {
1367 		int i;
1368 
1369 		for (i = 0; i < aobj->u_pages; i++) {
1370 			int slot = aobj->u_swslots[i];
1371 
1372 			/*
1373 			 * if the slot isn't in range, skip it
1374 			 */
1375 			if (slot < startslot || slot >= endslot) {
1376 				continue;
1377 			}
1378 
1379 			/*
1380 			 * process the page.
1381 			 */
1382 			rv = uao_pagein_page(aobj, i);
1383 			if (rv) {
1384 				return rv;
1385 			}
1386 		}
1387 	}
1388 
1389 	return FALSE;
1390 }
1391 
1392 /*
1393  * uao_pagein_page: page in a single page from an anonymous UVM object.
1394  *
1395  * => Returns TRUE if pagein was aborted due to lack of memory.
1396  */
1397 static boolean_t
1398 uao_pagein_page(struct uvm_aobj *aobj, int pageidx)
1399 {
1400 	struct uvm_object *uobj = &aobj->u_obj;
1401 	struct vm_page *pg;
1402 	int rv, slot, npages;
1403 
1404 	pg = NULL;
1405 	npages = 1;
1406 
1407 	KASSERT(rw_write_held(uobj->vmobjlock));
1408 	rv = uao_get(&aobj->u_obj, (voff_t)pageidx << PAGE_SHIFT,
1409 	    &pg, &npages, 0, PROT_READ | PROT_WRITE, 0, 0);
1410 
1411 	/*
1412 	 * relock and finish up.
1413 	 */
1414 	rw_enter(uobj->vmobjlock, RW_WRITE);
1415 	switch (rv) {
1416 	case VM_PAGER_OK:
1417 		break;
1418 
1419 	case VM_PAGER_ERROR:
1420 	case VM_PAGER_REFAULT:
1421 		/*
1422 		 * nothing more to do on errors.
1423 		 * VM_PAGER_REFAULT can only mean that the anon was freed,
1424 		 * so again there's nothing to do.
1425 		 */
1426 		return FALSE;
1427 	}
1428 
1429 	/*
1430 	 * ok, we've got the page now.
1431 	 * mark it as dirty, clear its swslot and un-busy it.
1432 	 */
1433 	slot = uao_set_swslot(&aobj->u_obj, pageidx, 0);
1434 	uvm_swap_free(slot, 1);
1435 	atomic_clearbits_int(&pg->pg_flags, PG_BUSY|PG_CLEAN|PG_FAKE);
1436 	UVM_PAGE_OWN(pg, NULL);
1437 
1438 	/*
1439 	 * deactivate the page (to put it on a page queue).
1440 	 */
1441 	pmap_clear_reference(pg);
1442 	uvm_lock_pageq();
1443 	uvm_pagedeactivate(pg);
1444 	uvm_unlock_pageq();
1445 
1446 	return FALSE;
1447 }
1448 
1449 /*
1450  * uao_dropswap_range: drop swapslots in the range.
1451  *
1452  * => aobj must be locked and is returned locked.
1453  * => start is inclusive.  end is exclusive.
1454  */
1455 void
1456 uao_dropswap_range(struct uvm_object *uobj, voff_t start, voff_t end)
1457 {
1458 	struct uvm_aobj *aobj = (struct uvm_aobj *)uobj;
1459 	int swpgonlydelta = 0;
1460 
1461 	KASSERT(UVM_OBJ_IS_AOBJ(uobj));
1462 	KASSERT(rw_write_held(uobj->vmobjlock));
1463 
1464 	if (end == 0) {
1465 		end = INT64_MAX;
1466 	}
1467 
1468 	if (UAO_USES_SWHASH(aobj)) {
1469 		int i, hashbuckets = aobj->u_swhashmask + 1;
1470 		voff_t taghi;
1471 		voff_t taglo;
1472 
1473 		taglo = UAO_SWHASH_ELT_TAG(start);
1474 		taghi = UAO_SWHASH_ELT_TAG(end);
1475 
1476 		for (i = 0; i < hashbuckets; i++) {
1477 			struct uao_swhash_elt *elt, *next;
1478 
1479 			for (elt = LIST_FIRST(&aobj->u_swhash[i]);
1480 			     elt != NULL;
1481 			     elt = next) {
1482 				int startidx, endidx;
1483 				int j;
1484 
1485 				next = LIST_NEXT(elt, list);
1486 
1487 				if (elt->tag < taglo || taghi < elt->tag) {
1488 					continue;
1489 				}
1490 
1491 				if (elt->tag == taglo) {
1492 					startidx =
1493 					    UAO_SWHASH_ELT_PAGESLOT_IDX(start);
1494 				} else {
1495 					startidx = 0;
1496 				}
1497 
1498 				if (elt->tag == taghi) {
1499 					endidx =
1500 					    UAO_SWHASH_ELT_PAGESLOT_IDX(end);
1501 				} else {
1502 					endidx = UAO_SWHASH_CLUSTER_SIZE;
1503 				}
1504 
1505 				for (j = startidx; j < endidx; j++) {
1506 					int slot = elt->slots[j];
1507 
1508 					KASSERT(uvm_pagelookup(&aobj->u_obj,
1509 					    (voff_t)(UAO_SWHASH_ELT_PAGEIDX_BASE(elt)
1510 					    + j) << PAGE_SHIFT) == NULL);
1511 
1512 					if (slot > 0) {
1513 						uvm_swap_free(slot, 1);
1514 						swpgonlydelta++;
1515 						KASSERT(elt->count > 0);
1516 						elt->slots[j] = 0;
1517 						elt->count--;
1518 					}
1519 				}
1520 
1521 				if (elt->count == 0) {
1522 					LIST_REMOVE(elt, list);
1523 					pool_put(&uao_swhash_elt_pool, elt);
1524 				}
1525 			}
1526 		}
1527 	} else {
1528 		int i;
1529 
1530 		if (aobj->u_pages < end) {
1531 			end = aobj->u_pages;
1532 		}
1533 		for (i = start; i < end; i++) {
1534 			int slot = aobj->u_swslots[i];
1535 
1536 			if (slot > 0) {
1537 				uvm_swap_free(slot, 1);
1538 				swpgonlydelta++;
1539 			}
1540 		}
1541 	}
1542 
1543 	/*
1544 	 * adjust the counter of pages only in swap for all
1545 	 * the swap slots we've freed.
1546 	 */
1547 	if (swpgonlydelta > 0) {
1548 		KASSERT(uvmexp.swpgonly >= swpgonlydelta);
1549 		atomic_add_int(&uvmexp.swpgonly, -swpgonlydelta);
1550 	}
1551 }
1552