xref: /openbsd-src/sys/uvm/uvm_aobj.c (revision 25c4e8bd056e974b28f4a0ffd39d76c190a56013)
1 /*	$OpenBSD: uvm_aobj.c,v 1.104 2022/07/11 11:33:17 mpi 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 boolean_t		 uao_flush(struct uvm_object *, voff_t,
147 				     voff_t, int);
148 static void			 uao_free(struct uvm_aobj *);
149 static int			 uao_get(struct uvm_object *, voff_t,
150 				     vm_page_t *, int *, int, vm_prot_t,
151 				     int, int);
152 static boolean_t		 uao_pagein(struct uvm_aobj *, int, int);
153 static boolean_t		 uao_pagein_page(struct uvm_aobj *, int);
154 
155 void	uao_dropswap_range(struct uvm_object *, voff_t, voff_t);
156 void	uao_shrink_flush(struct uvm_object *, int, int);
157 int	uao_shrink_hash(struct uvm_object *, int);
158 int	uao_shrink_array(struct uvm_object *, int);
159 int	uao_shrink_convert(struct uvm_object *, int);
160 
161 int	uao_grow_hash(struct uvm_object *, int);
162 int	uao_grow_array(struct uvm_object *, int);
163 int	uao_grow_convert(struct uvm_object *, int);
164 
165 /*
166  * aobj_pager
167  *
168  * note that some functions (e.g. put) are handled elsewhere
169  */
170 const struct uvm_pagerops aobj_pager = {
171 	.pgo_reference = uao_reference,
172 	.pgo_detach = uao_detach,
173 	.pgo_flush = uao_flush,
174 	.pgo_get = uao_get,
175 };
176 
177 /*
178  * uao_list: global list of active aobjs, locked by uao_list_lock
179  *
180  * Lock ordering: generally the locking order is object lock, then list lock.
181  * in the case of swap off we have to iterate over the list, and thus the
182  * ordering is reversed. In that case we must use trylocking to prevent
183  * deadlock.
184  */
185 static LIST_HEAD(aobjlist, uvm_aobj) uao_list = LIST_HEAD_INITIALIZER(uao_list);
186 static struct mutex uao_list_lock = MUTEX_INITIALIZER(IPL_MPFLOOR);
187 
188 
189 /*
190  * functions
191  */
192 /*
193  * hash table/array related functions
194  */
195 /*
196  * uao_find_swhash_elt: find (or create) a hash table entry for a page
197  * offset.
198  */
199 static struct uao_swhash_elt *
200 uao_find_swhash_elt(struct uvm_aobj *aobj, int pageidx, boolean_t create)
201 {
202 	struct uao_swhash *swhash;
203 	struct uao_swhash_elt *elt;
204 	voff_t page_tag;
205 
206 	swhash = UAO_SWHASH_HASH(aobj, pageidx); /* first hash to get bucket */
207 	page_tag = UAO_SWHASH_ELT_TAG(pageidx);	/* tag to search for */
208 
209 	/*
210 	 * now search the bucket for the requested tag
211 	 */
212 	LIST_FOREACH(elt, swhash, list) {
213 		if (elt->tag == page_tag)
214 			return elt;
215 	}
216 
217 	if (!create)
218 		return NULL;
219 
220 	/*
221 	 * allocate a new entry for the bucket and init/insert it in
222 	 */
223 	elt = pool_get(&uao_swhash_elt_pool, PR_NOWAIT | PR_ZERO);
224 	/*
225 	 * XXX We cannot sleep here as the hash table might disappear
226 	 * from under our feet.  And we run the risk of deadlocking
227 	 * the pagedeamon.  In fact this code will only be called by
228 	 * the pagedaemon and allocation will only fail if we
229 	 * exhausted the pagedeamon reserve.  In that case we're
230 	 * doomed anyway, so panic.
231 	 */
232 	if (elt == NULL)
233 		panic("%s: can't allocate entry", __func__);
234 	LIST_INSERT_HEAD(swhash, elt, list);
235 	elt->tag = page_tag;
236 
237 	return elt;
238 }
239 
240 /*
241  * uao_find_swslot: find the swap slot number for an aobj/pageidx
242  */
243 int
244 uao_find_swslot(struct uvm_object *uobj, int pageidx)
245 {
246 	struct uvm_aobj *aobj = (struct uvm_aobj *)uobj;
247 
248 	KASSERT(UVM_OBJ_IS_AOBJ(uobj));
249 
250 	/*
251 	 * if noswap flag is set, then we never return a slot
252 	 */
253 	if (aobj->u_flags & UAO_FLAG_NOSWAP)
254 		return 0;
255 
256 	/*
257 	 * if hashing, look in hash table.
258 	 */
259 	if (UAO_USES_SWHASH(aobj)) {
260 		struct uao_swhash_elt *elt =
261 		    uao_find_swhash_elt(aobj, pageidx, FALSE);
262 
263 		if (elt)
264 			return UAO_SWHASH_ELT_PAGESLOT(elt, pageidx);
265 		else
266 			return 0;
267 	}
268 
269 	/*
270 	 * otherwise, look in the array
271 	 */
272 	return aobj->u_swslots[pageidx];
273 }
274 
275 /*
276  * uao_set_swslot: set the swap slot for a page in an aobj.
277  *
278  * => setting a slot to zero frees the slot
279  * => object must be locked by caller
280  * => we return the old slot number, or -1 if we failed to allocate
281  *    memory to record the new slot number
282  */
283 int
284 uao_set_swslot(struct uvm_object *uobj, int pageidx, int slot)
285 {
286 	struct uvm_aobj *aobj = (struct uvm_aobj *)uobj;
287 	int oldslot;
288 
289 	KASSERT(rw_write_held(uobj->vmobjlock) || uobj->uo_refs == 0);
290 	KASSERT(UVM_OBJ_IS_AOBJ(uobj));
291 
292 	/*
293 	 * if noswap flag is set, then we can't set a slot
294 	 */
295 	if (aobj->u_flags & UAO_FLAG_NOSWAP) {
296 		if (slot == 0)
297 			return 0;		/* a clear is ok */
298 
299 		/* but a set is not */
300 		printf("uao_set_swslot: uobj = %p\n", uobj);
301 	    panic("uao_set_swslot: attempt to set a slot on a NOSWAP object");
302 	}
303 
304 	/*
305 	 * are we using a hash table?  if so, add it in the hash.
306 	 */
307 	if (UAO_USES_SWHASH(aobj)) {
308 		/*
309 		 * Avoid allocating an entry just to free it again if
310 		 * the page had not swap slot in the first place, and
311 		 * we are freeing.
312 		 */
313 		struct uao_swhash_elt *elt =
314 		    uao_find_swhash_elt(aobj, pageidx, slot ? TRUE : FALSE);
315 		if (elt == NULL) {
316 			KASSERT(slot == 0);
317 			return 0;
318 		}
319 
320 		oldslot = UAO_SWHASH_ELT_PAGESLOT(elt, pageidx);
321 		UAO_SWHASH_ELT_PAGESLOT(elt, pageidx) = slot;
322 
323 		/*
324 		 * now adjust the elt's reference counter and free it if we've
325 		 * dropped it to zero.
326 		 */
327 		if (slot) {
328 			if (oldslot == 0)
329 				elt->count++;
330 		} else {
331 			if (oldslot)
332 				elt->count--;
333 
334 			if (elt->count == 0) {
335 				LIST_REMOVE(elt, list);
336 				pool_put(&uao_swhash_elt_pool, elt);
337 			}
338 		}
339 	} else {
340 		/* we are using an array */
341 		oldslot = aobj->u_swslots[pageidx];
342 		aobj->u_swslots[pageidx] = slot;
343 	}
344 	return oldslot;
345 }
346 /*
347  * end of hash/array functions
348  */
349 
350 /*
351  * uao_free: free all resources held by an aobj, and then free the aobj
352  *
353  * => the aobj should be dead
354  */
355 static void
356 uao_free(struct uvm_aobj *aobj)
357 {
358 	struct uvm_object *uobj = &aobj->u_obj;
359 
360 	KASSERT(UVM_OBJ_IS_AOBJ(uobj));
361 	KASSERT(rw_write_held(uobj->vmobjlock));
362 	uao_dropswap_range(uobj, 0, 0);
363 	rw_exit(uobj->vmobjlock);
364 
365 	if (UAO_USES_SWHASH(aobj)) {
366 		/*
367 		 * free the hash table itself.
368 		 */
369 		hashfree(aobj->u_swhash, UAO_SWHASH_BUCKETS(aobj->u_pages), M_UVMAOBJ);
370 	} else {
371 		free(aobj->u_swslots, M_UVMAOBJ, aobj->u_pages * sizeof(int));
372 	}
373 
374 	/*
375 	 * finally free the aobj itself
376 	 */
377 	uvm_obj_destroy(uobj);
378 	pool_put(&uvm_aobj_pool, aobj);
379 }
380 
381 /*
382  * pager functions
383  */
384 
385 #ifdef TMPFS
386 /*
387  * Shrink an aobj to a given number of pages. The procedure is always the same:
388  * assess the necessity of data structure conversion (hash to array), secure
389  * resources, flush pages and drop swap slots.
390  *
391  */
392 
393 void
394 uao_shrink_flush(struct uvm_object *uobj, int startpg, int endpg)
395 {
396 	KASSERT(startpg < endpg);
397 	KASSERT(uobj->uo_refs == 1);
398 	uao_flush(uobj, (voff_t)startpg << PAGE_SHIFT,
399 	    (voff_t)endpg << PAGE_SHIFT, PGO_FREE);
400 	uao_dropswap_range(uobj, startpg, endpg);
401 }
402 
403 int
404 uao_shrink_hash(struct uvm_object *uobj, int pages)
405 {
406 	struct uvm_aobj *aobj = (struct uvm_aobj *)uobj;
407 	struct uao_swhash *new_swhash;
408 	struct uao_swhash_elt *elt;
409 	unsigned long new_hashmask;
410 	int i;
411 
412 	KASSERT(UAO_USES_SWHASH(aobj));
413 
414 	/*
415 	 * If the size of the hash table doesn't change, all we need to do is
416 	 * to adjust the page count.
417 	 */
418 	if (UAO_SWHASH_BUCKETS(aobj->u_pages) == UAO_SWHASH_BUCKETS(pages)) {
419 		uao_shrink_flush(uobj, pages, aobj->u_pages);
420 		aobj->u_pages = pages;
421 		return 0;
422 	}
423 
424 	new_swhash = hashinit(UAO_SWHASH_BUCKETS(pages), M_UVMAOBJ,
425 	    M_WAITOK | M_CANFAIL, &new_hashmask);
426 	if (new_swhash == NULL)
427 		return ENOMEM;
428 
429 	uao_shrink_flush(uobj, pages, aobj->u_pages);
430 
431 	/*
432 	 * Even though the hash table size is changing, the hash of the buckets
433 	 * we are interested in copying should not change.
434 	 */
435 	for (i = 0; i < UAO_SWHASH_BUCKETS(aobj->u_pages); i++) {
436 		while (LIST_EMPTY(&aobj->u_swhash[i]) == 0) {
437 			elt = LIST_FIRST(&aobj->u_swhash[i]);
438 			LIST_REMOVE(elt, list);
439 			LIST_INSERT_HEAD(&new_swhash[i], elt, list);
440 		}
441 	}
442 
443 	hashfree(aobj->u_swhash, UAO_SWHASH_BUCKETS(aobj->u_pages), M_UVMAOBJ);
444 
445 	aobj->u_swhash = new_swhash;
446 	aobj->u_pages = pages;
447 	aobj->u_swhashmask = new_hashmask;
448 
449 	return 0;
450 }
451 
452 int
453 uao_shrink_convert(struct uvm_object *uobj, int pages)
454 {
455 	struct uvm_aobj *aobj = (struct uvm_aobj *)uobj;
456 	struct uao_swhash_elt *elt;
457 	int i, *new_swslots;
458 
459 	new_swslots = mallocarray(pages, sizeof(int), M_UVMAOBJ,
460 	    M_WAITOK | M_CANFAIL | M_ZERO);
461 	if (new_swslots == NULL)
462 		return ENOMEM;
463 
464 	uao_shrink_flush(uobj, pages, aobj->u_pages);
465 
466 	/* Convert swap slots from hash to array.  */
467 	for (i = 0; i < pages; i++) {
468 		elt = uao_find_swhash_elt(aobj, i, FALSE);
469 		if (elt != NULL) {
470 			new_swslots[i] = UAO_SWHASH_ELT_PAGESLOT(elt, i);
471 			if (new_swslots[i] != 0)
472 				elt->count--;
473 			if (elt->count == 0) {
474 				LIST_REMOVE(elt, list);
475 				pool_put(&uao_swhash_elt_pool, elt);
476 			}
477 		}
478 	}
479 
480 	hashfree(aobj->u_swhash, UAO_SWHASH_BUCKETS(aobj->u_pages), M_UVMAOBJ);
481 
482 	aobj->u_swslots = new_swslots;
483 	aobj->u_pages = pages;
484 
485 	return 0;
486 }
487 
488 int
489 uao_shrink_array(struct uvm_object *uobj, int pages)
490 {
491 	struct uvm_aobj *aobj = (struct uvm_aobj *)uobj;
492 	int i, *new_swslots;
493 
494 	new_swslots = mallocarray(pages, sizeof(int), M_UVMAOBJ,
495 	    M_WAITOK | M_CANFAIL | M_ZERO);
496 	if (new_swslots == NULL)
497 		return ENOMEM;
498 
499 	uao_shrink_flush(uobj, pages, aobj->u_pages);
500 
501 	for (i = 0; i < pages; i++)
502 		new_swslots[i] = aobj->u_swslots[i];
503 
504 	free(aobj->u_swslots, M_UVMAOBJ, aobj->u_pages * sizeof(int));
505 
506 	aobj->u_swslots = new_swslots;
507 	aobj->u_pages = pages;
508 
509 	return 0;
510 }
511 
512 int
513 uao_shrink(struct uvm_object *uobj, int pages)
514 {
515 	struct uvm_aobj *aobj = (struct uvm_aobj *)uobj;
516 
517 	KASSERT(pages < aobj->u_pages);
518 
519 	/*
520 	 * Distinguish between three possible cases:
521 	 * 1. aobj uses hash and must be converted to array.
522 	 * 2. aobj uses array and array size needs to be adjusted.
523 	 * 3. aobj uses hash and hash size needs to be adjusted.
524 	 */
525 	if (pages > UAO_SWHASH_THRESHOLD)
526 		return uao_shrink_hash(uobj, pages);	/* case 3 */
527 	else if (aobj->u_pages > UAO_SWHASH_THRESHOLD)
528 		return uao_shrink_convert(uobj, pages);	/* case 1 */
529 	else
530 		return uao_shrink_array(uobj, pages);	/* case 2 */
531 }
532 
533 /*
534  * Grow an aobj to a given number of pages. Right now we only adjust the swap
535  * slots. We could additionally handle page allocation directly, so that they
536  * don't happen through uvm_fault(). That would allow us to use another
537  * mechanism for the swap slots other than malloc(). It is thus mandatory that
538  * the caller of these functions does not allow faults to happen in case of
539  * growth error.
540  */
541 int
542 uao_grow_array(struct uvm_object *uobj, int pages)
543 {
544 	struct uvm_aobj *aobj = (struct uvm_aobj *)uobj;
545 	int i, *new_swslots;
546 
547 	KASSERT(aobj->u_pages <= UAO_SWHASH_THRESHOLD);
548 
549 	new_swslots = mallocarray(pages, sizeof(int), M_UVMAOBJ,
550 	    M_WAITOK | M_CANFAIL | M_ZERO);
551 	if (new_swslots == NULL)
552 		return ENOMEM;
553 
554 	for (i = 0; i < aobj->u_pages; i++)
555 		new_swslots[i] = aobj->u_swslots[i];
556 
557 	free(aobj->u_swslots, M_UVMAOBJ, aobj->u_pages * sizeof(int));
558 
559 	aobj->u_swslots = new_swslots;
560 	aobj->u_pages = pages;
561 
562 	return 0;
563 }
564 
565 int
566 uao_grow_hash(struct uvm_object *uobj, int pages)
567 {
568 	struct uvm_aobj *aobj = (struct uvm_aobj *)uobj;
569 	struct uao_swhash *new_swhash;
570 	struct uao_swhash_elt *elt;
571 	unsigned long new_hashmask;
572 	int i;
573 
574 	KASSERT(pages > UAO_SWHASH_THRESHOLD);
575 
576 	/*
577 	 * If the size of the hash table doesn't change, all we need to do is
578 	 * to adjust the page count.
579 	 */
580 	if (UAO_SWHASH_BUCKETS(aobj->u_pages) == UAO_SWHASH_BUCKETS(pages)) {
581 		aobj->u_pages = pages;
582 		return 0;
583 	}
584 
585 	KASSERT(UAO_SWHASH_BUCKETS(aobj->u_pages) < UAO_SWHASH_BUCKETS(pages));
586 
587 	new_swhash = hashinit(UAO_SWHASH_BUCKETS(pages), M_UVMAOBJ,
588 	    M_WAITOK | M_CANFAIL, &new_hashmask);
589 	if (new_swhash == NULL)
590 		return ENOMEM;
591 
592 	for (i = 0; i < UAO_SWHASH_BUCKETS(aobj->u_pages); i++) {
593 		while (LIST_EMPTY(&aobj->u_swhash[i]) == 0) {
594 			elt = LIST_FIRST(&aobj->u_swhash[i]);
595 			LIST_REMOVE(elt, list);
596 			LIST_INSERT_HEAD(&new_swhash[i], elt, list);
597 		}
598 	}
599 
600 	hashfree(aobj->u_swhash, UAO_SWHASH_BUCKETS(aobj->u_pages), M_UVMAOBJ);
601 
602 	aobj->u_swhash = new_swhash;
603 	aobj->u_pages = pages;
604 	aobj->u_swhashmask = new_hashmask;
605 
606 	return 0;
607 }
608 
609 int
610 uao_grow_convert(struct uvm_object *uobj, int pages)
611 {
612 	struct uvm_aobj *aobj = (struct uvm_aobj *)uobj;
613 	struct uao_swhash *new_swhash;
614 	struct uao_swhash_elt *elt;
615 	unsigned long new_hashmask;
616 	int i, *old_swslots;
617 
618 	new_swhash = hashinit(UAO_SWHASH_BUCKETS(pages), M_UVMAOBJ,
619 	    M_WAITOK | M_CANFAIL, &new_hashmask);
620 	if (new_swhash == NULL)
621 		return ENOMEM;
622 
623 	/* Set these now, so we can use uao_find_swhash_elt(). */
624 	old_swslots = aobj->u_swslots;
625 	aobj->u_swhash = new_swhash;
626 	aobj->u_swhashmask = new_hashmask;
627 
628 	for (i = 0; i < aobj->u_pages; i++) {
629 		if (old_swslots[i] != 0) {
630 			elt = uao_find_swhash_elt(aobj, i, TRUE);
631 			elt->count++;
632 			UAO_SWHASH_ELT_PAGESLOT(elt, i) = old_swslots[i];
633 		}
634 	}
635 
636 	free(old_swslots, M_UVMAOBJ, aobj->u_pages * sizeof(int));
637 	aobj->u_pages = pages;
638 
639 	return 0;
640 }
641 
642 int
643 uao_grow(struct uvm_object *uobj, int pages)
644 {
645 	struct uvm_aobj *aobj = (struct uvm_aobj *)uobj;
646 
647 	KASSERT(pages > aobj->u_pages);
648 
649 	/*
650 	 * Distinguish between three possible cases:
651 	 * 1. aobj uses hash and hash size needs to be adjusted.
652 	 * 2. aobj uses array and array size needs to be adjusted.
653 	 * 3. aobj uses array and must be converted to hash.
654 	 */
655 	if (pages <= UAO_SWHASH_THRESHOLD)
656 		return uao_grow_array(uobj, pages);	/* case 2 */
657 	else if (aobj->u_pages > UAO_SWHASH_THRESHOLD)
658 		return uao_grow_hash(uobj, pages);	/* case 1 */
659 	else
660 		return uao_grow_convert(uobj, pages);
661 }
662 #endif /* TMPFS */
663 
664 /*
665  * uao_create: create an aobj of the given size and return its uvm_object.
666  *
667  * => for normal use, flags are zero or UAO_FLAG_CANFAIL.
668  * => for the kernel object, the flags are:
669  *	UAO_FLAG_KERNOBJ - allocate the kernel object (can only happen once)
670  *	UAO_FLAG_KERNSWAP - enable swapping of kernel object ("           ")
671  */
672 struct uvm_object *
673 uao_create(vsize_t size, int flags)
674 {
675 	static struct uvm_aobj kernel_object_store;
676 	static struct rwlock bootstrap_kernel_object_lock;
677 	static int kobj_alloced = 0;
678 	int pages = round_page(size) >> PAGE_SHIFT;
679 	struct uvm_aobj *aobj;
680 	int refs;
681 
682 	/*
683 	 * Allocate a new aobj, unless kernel object is requested.
684 	 */
685 	if (flags & UAO_FLAG_KERNOBJ) {
686 		KASSERT(!kobj_alloced);
687 		aobj = &kernel_object_store;
688 		aobj->u_pages = pages;
689 		aobj->u_flags = UAO_FLAG_NOSWAP;
690 		refs = UVM_OBJ_KERN;
691 		kobj_alloced = UAO_FLAG_KERNOBJ;
692 	} else if (flags & UAO_FLAG_KERNSWAP) {
693 		KASSERT(kobj_alloced == UAO_FLAG_KERNOBJ);
694 		aobj = &kernel_object_store;
695 		kobj_alloced = UAO_FLAG_KERNSWAP;
696 	} else {
697 		aobj = pool_get(&uvm_aobj_pool, PR_WAITOK);
698 		aobj->u_pages = pages;
699 		aobj->u_flags = 0;
700 		refs = 1;
701 	}
702 
703 	/*
704 	 * allocate hash/array if necessary
705 	 */
706  	if (flags == 0 || (flags & (UAO_FLAG_KERNSWAP | UAO_FLAG_CANFAIL))) {
707 		int mflags;
708 
709 		if (flags)
710 			mflags = M_NOWAIT;
711 		else
712 			mflags = M_WAITOK;
713 
714 		/* allocate hash table or array depending on object size */
715 		if (UAO_USES_SWHASH(aobj)) {
716 			aobj->u_swhash = hashinit(UAO_SWHASH_BUCKETS(pages),
717 			    M_UVMAOBJ, mflags, &aobj->u_swhashmask);
718 			if (aobj->u_swhash == NULL) {
719 				if (flags & UAO_FLAG_CANFAIL) {
720 					pool_put(&uvm_aobj_pool, aobj);
721 					return NULL;
722 				}
723 				panic("uao_create: hashinit swhash failed");
724 			}
725 		} else {
726 			aobj->u_swslots = mallocarray(pages, sizeof(int),
727 			    M_UVMAOBJ, mflags|M_ZERO);
728 			if (aobj->u_swslots == NULL) {
729 				if (flags & UAO_FLAG_CANFAIL) {
730 					pool_put(&uvm_aobj_pool, aobj);
731 					return NULL;
732 				}
733 				panic("uao_create: malloc swslots failed");
734 			}
735 		}
736 
737 		if (flags & UAO_FLAG_KERNSWAP) {
738 			aobj->u_flags &= ~UAO_FLAG_NOSWAP; /* clear noswap */
739 			return &aobj->u_obj;
740 			/* done! */
741 		}
742 	}
743 
744 	/*
745 	 * Initialise UVM object.
746 	 */
747 	uvm_obj_init(&aobj->u_obj, &aobj_pager, refs);
748 	if (flags & UAO_FLAG_KERNOBJ) {
749 		/* Use a temporary static lock for kernel_object. */
750 		rw_init(&bootstrap_kernel_object_lock, "kobjlk");
751 		uvm_obj_setlock(&aobj->u_obj, &bootstrap_kernel_object_lock);
752 	}
753 
754 	/*
755  	 * now that aobj is ready, add it to the global list
756  	 */
757 	mtx_enter(&uao_list_lock);
758 	LIST_INSERT_HEAD(&uao_list, aobj, u_list);
759 	mtx_leave(&uao_list_lock);
760 
761 	return &aobj->u_obj;
762 }
763 
764 
765 
766 /*
767  * uao_init: set up aobj pager subsystem
768  *
769  * => called at boot time from uvm_pager_init()
770  */
771 void
772 uao_init(void)
773 {
774 	/*
775 	 * NOTE: Pages for this pool must not come from a pageable
776 	 * kernel map!
777 	 */
778 	pool_init(&uao_swhash_elt_pool, sizeof(struct uao_swhash_elt), 0,
779 	    IPL_NONE, PR_WAITOK, "uaoeltpl", NULL);
780 	pool_init(&uvm_aobj_pool, sizeof(struct uvm_aobj), 0,
781 	    IPL_NONE, PR_WAITOK, "aobjpl", NULL);
782 }
783 
784 /*
785  * uao_reference: hold a reference to an anonymous UVM object.
786  */
787 void
788 uao_reference(struct uvm_object *uobj)
789 {
790 	/* Kernel object is persistent. */
791 	if (UVM_OBJ_IS_KERN_OBJECT(uobj))
792 		return;
793 
794 	atomic_inc_int(&uobj->uo_refs);
795 }
796 
797 
798 /*
799  * uao_detach: drop a reference to an anonymous UVM object.
800  */
801 void
802 uao_detach(struct uvm_object *uobj)
803 {
804 	struct uvm_aobj *aobj = (struct uvm_aobj *)uobj;
805 	struct vm_page *pg;
806 
807 	/*
808 	 * Detaching from kernel_object is a NOP.
809 	 */
810 	if (UVM_OBJ_IS_KERN_OBJECT(uobj))
811 		return;
812 
813 	/*
814 	 * Drop the reference.  If it was the last one, destroy the object.
815 	 */
816 	if (atomic_dec_int_nv(&uobj->uo_refs) > 0) {
817 		return;
818 	}
819 
820 	/*
821 	 * Remove the aobj from the global list.
822 	 */
823 	mtx_enter(&uao_list_lock);
824 	LIST_REMOVE(aobj, u_list);
825 	mtx_leave(&uao_list_lock);
826 
827 	/*
828 	 * Free all the pages left in the aobj.  For each page, when the
829 	 * page is no longer busy (and thus after any disk I/O that it is
830 	 * involved in is complete), release any swap resources and free
831 	 * the page itself.
832 	 */
833 	rw_enter(uobj->vmobjlock, RW_WRITE);
834 	while ((pg = RBT_ROOT(uvm_objtree, &uobj->memt)) != NULL) {
835 		pmap_page_protect(pg, PROT_NONE);
836 		if (pg->pg_flags & PG_BUSY) {
837 			atomic_setbits_int(&pg->pg_flags, PG_WANTED);
838 			rwsleep_nsec(pg, uobj->vmobjlock, PVM, "uao_det",
839 			    INFSLP);
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 			atomic_setbits_int(&pg->pg_flags, PG_WANTED);
912 			rwsleep_nsec(pg, uobj->vmobjlock, PVM, "uaoflsh",
913 			    INFSLP);
914 			curoff -= PAGE_SIZE;
915 			continue;
916 		}
917 
918 		switch (flags & (PGO_CLEANIT|PGO_FREE|PGO_DEACTIVATE)) {
919 		/*
920 		 * XXX In these first 3 cases, we always just
921 		 * XXX deactivate the page.  We may want to
922 		 * XXX handle the different cases more specifically
923 		 * XXX in the future.
924 		 */
925 		case PGO_CLEANIT|PGO_FREE:
926 			/* FALLTHROUGH */
927 		case PGO_CLEANIT|PGO_DEACTIVATE:
928 			/* FALLTHROUGH */
929 		case PGO_DEACTIVATE:
930  deactivate_it:
931 			if (pg->wire_count != 0)
932 				continue;
933 
934 			uvm_lock_pageq();
935 			pmap_page_protect(pg, PROT_NONE);
936 			uvm_pagedeactivate(pg);
937 			uvm_unlock_pageq();
938 
939 			continue;
940 		case PGO_FREE:
941 			/*
942 			 * If there are multiple references to
943 			 * the object, just deactivate the page.
944 			 */
945 			if (uobj->uo_refs > 1)
946 				goto deactivate_it;
947 
948 			/* XXX skip the page if it's wired */
949 			if (pg->wire_count != 0)
950 				continue;
951 
952 			/*
953 			 * free the swap slot and the page.
954 			 */
955 			pmap_page_protect(pg, PROT_NONE);
956 
957 			/*
958 			 * freeing swapslot here is not strictly necessary.
959 			 * however, leaving it here doesn't save much
960 			 * because we need to update swap accounting anyway.
961 			 */
962 			uao_dropswap(uobj, pg->offset >> PAGE_SHIFT);
963 			uvm_lock_pageq();
964 			uvm_pagefree(pg);
965 			uvm_unlock_pageq();
966 
967 			continue;
968 		default:
969 			panic("uao_flush: weird flags");
970 		}
971 	}
972 
973 	return TRUE;
974 }
975 
976 /*
977  * uao_get: fetch me a page
978  *
979  * we have three cases:
980  * 1: page is resident     -> just return the page.
981  * 2: page is zero-fill    -> allocate a new page and zero it.
982  * 3: page is swapped out  -> fetch the page from swap.
983  *
984  * cases 1 can be handled with PGO_LOCKED, cases 2 and 3 cannot.
985  * so, if the "center" page hits case 3 (or any page, with PGO_ALLPAGES),
986  * then we will need to return VM_PAGER_UNLOCK.
987  *
988  * => flags: PGO_ALLPAGES: get all of the pages
989  *           PGO_LOCKED: fault data structures are locked
990  * => NOTE: offset is the offset of pps[0], _NOT_ pps[centeridx]
991  * => NOTE: caller must check for released pages!!
992  */
993 static int
994 uao_get(struct uvm_object *uobj, voff_t offset, struct vm_page **pps,
995     int *npagesp, int centeridx, vm_prot_t access_type, int advice, int flags)
996 {
997 	struct uvm_aobj *aobj = (struct uvm_aobj *)uobj;
998 	voff_t current_offset;
999 	vm_page_t ptmp;
1000 	int lcv, gotpages, maxpages, swslot, rv, pageidx;
1001 	boolean_t done;
1002 
1003 	KASSERT(UVM_OBJ_IS_AOBJ(uobj));
1004 	KASSERT(rw_write_held(uobj->vmobjlock));
1005 
1006 	/*
1007  	 * get number of pages
1008  	 */
1009 	maxpages = *npagesp;
1010 
1011 	if (flags & PGO_LOCKED) {
1012 		/*
1013  		 * step 1a: get pages that are already resident.   only do
1014 		 * this if the data structures are locked (i.e. the first
1015 		 * time through).
1016  		 */
1017 
1018 		done = TRUE;	/* be optimistic */
1019 		gotpages = 0;	/* # of pages we got so far */
1020 
1021 		for (lcv = 0, current_offset = offset ; lcv < maxpages ;
1022 		    lcv++, current_offset += PAGE_SIZE) {
1023 			/* do we care about this page?  if not, skip it */
1024 			if (pps[lcv] == PGO_DONTCARE)
1025 				continue;
1026 
1027 			ptmp = uvm_pagelookup(uobj, current_offset);
1028 
1029 			/*
1030  			 * if page is new, attempt to allocate the page,
1031 			 * zero-fill'd.
1032  			 */
1033 			if (ptmp == NULL && uao_find_swslot(uobj,
1034 			    current_offset >> PAGE_SHIFT) == 0) {
1035 				ptmp = uvm_pagealloc(uobj, current_offset,
1036 				    NULL, UVM_PGA_ZERO);
1037 				if (ptmp) {
1038 					/* new page */
1039 					atomic_clearbits_int(&ptmp->pg_flags,
1040 					    PG_BUSY|PG_FAKE);
1041 					atomic_setbits_int(&ptmp->pg_flags,
1042 					    PQ_AOBJ);
1043 					UVM_PAGE_OWN(ptmp, NULL);
1044 				}
1045 			}
1046 
1047 			/*
1048 			 * to be useful must get a non-busy page
1049 			 */
1050 			if (ptmp == NULL ||
1051 			    (ptmp->pg_flags & PG_BUSY) != 0) {
1052 				if (lcv == centeridx ||
1053 				    (flags & PGO_ALLPAGES) != 0)
1054 					/* need to do a wait or I/O! */
1055 					done = FALSE;
1056 				continue;
1057 			}
1058 
1059 			/*
1060 			 * useful page: plug it in our result array
1061 			 */
1062 			atomic_setbits_int(&ptmp->pg_flags, PG_BUSY);
1063 			UVM_PAGE_OWN(ptmp, "uao_get1");
1064 			pps[lcv] = ptmp;
1065 			gotpages++;
1066 
1067 		}
1068 
1069 		/*
1070  		 * step 1b: now we've either done everything needed or we
1071 		 * to unlock and do some waiting or I/O.
1072  		 */
1073 		*npagesp = gotpages;
1074 		if (done)
1075 			/* bingo! */
1076 			return VM_PAGER_OK;
1077 		else
1078 			/* EEK!   Need to unlock and I/O */
1079 			return VM_PAGER_UNLOCK;
1080 	}
1081 
1082 	/*
1083  	 * step 2: get non-resident or busy pages.
1084  	 * data structures are unlocked.
1085  	 */
1086 	for (lcv = 0, current_offset = offset ; lcv < maxpages ;
1087 	    lcv++, current_offset += PAGE_SIZE) {
1088 		/*
1089 		 * - skip over pages we've already gotten or don't want
1090 		 * - skip over pages we don't _have_ to get
1091 		 */
1092 		if (pps[lcv] != NULL ||
1093 		    (lcv != centeridx && (flags & PGO_ALLPAGES) == 0))
1094 			continue;
1095 
1096 		pageidx = current_offset >> PAGE_SHIFT;
1097 
1098 		/*
1099  		 * we have yet to locate the current page (pps[lcv]).   we
1100 		 * first look for a page that is already at the current offset.
1101 		 * if we find a page, we check to see if it is busy or
1102 		 * released.  if that is the case, then we sleep on the page
1103 		 * until it is no longer busy or released and repeat the lookup.
1104 		 * if the page we found is neither busy nor released, then we
1105 		 * busy it (so we own it) and plug it into pps[lcv].   this
1106 		 * 'break's the following while loop and indicates we are
1107 		 * ready to move on to the next page in the "lcv" loop above.
1108  		 *
1109  		 * if we exit the while loop with pps[lcv] still set to NULL,
1110 		 * then it means that we allocated a new busy/fake/clean page
1111 		 * ptmp in the object and we need to do I/O to fill in the data.
1112  		 */
1113 
1114 		/* top of "pps" while loop */
1115 		while (pps[lcv] == NULL) {
1116 			/* look for a resident page */
1117 			ptmp = uvm_pagelookup(uobj, current_offset);
1118 
1119 			/* not resident?   allocate one now (if we can) */
1120 			if (ptmp == NULL) {
1121 
1122 				ptmp = uvm_pagealloc(uobj, current_offset,
1123 				    NULL, 0);
1124 
1125 				/* out of RAM? */
1126 				if (ptmp == NULL) {
1127 					rw_exit(uobj->vmobjlock);
1128 					uvm_wait("uao_getpage");
1129 					rw_enter(uobj->vmobjlock, RW_WRITE);
1130 					/* goto top of pps while loop */
1131 					continue;
1132 				}
1133 
1134 				/*
1135 				 * safe with PQ's unlocked: because we just
1136 				 * alloc'd the page
1137 				 */
1138 				atomic_setbits_int(&ptmp->pg_flags, PQ_AOBJ);
1139 
1140 				/*
1141 				 * got new page ready for I/O.  break pps while
1142 				 * loop.  pps[lcv] is still NULL.
1143 				 */
1144 				break;
1145 			}
1146 
1147 			/* page is there, see if we need to wait on it */
1148 			if ((ptmp->pg_flags & PG_BUSY) != 0) {
1149 				atomic_setbits_int(&ptmp->pg_flags, PG_WANTED);
1150 				rwsleep_nsec(ptmp, uobj->vmobjlock, PVM,
1151 				    "uao_get", INFSLP);
1152 				continue;	/* goto top of pps while loop */
1153 			}
1154 
1155 			/*
1156  			 * if we get here then the page is resident and
1157 			 * unbusy.  we busy it now (so we own it).
1158  			 */
1159 			/* we own it, caller must un-busy */
1160 			atomic_setbits_int(&ptmp->pg_flags, PG_BUSY);
1161 			UVM_PAGE_OWN(ptmp, "uao_get2");
1162 			pps[lcv] = ptmp;
1163 		}
1164 
1165 		/*
1166  		 * if we own the valid page at the correct offset, pps[lcv] will
1167  		 * point to it.   nothing more to do except go to the next page.
1168  		 */
1169 		if (pps[lcv])
1170 			continue;			/* next lcv */
1171 
1172 		/*
1173  		 * we have a "fake/busy/clean" page that we just allocated.
1174  		 * do the needed "i/o", either reading from swap or zeroing.
1175  		 */
1176 		swslot = uao_find_swslot(uobj, pageidx);
1177 
1178 		/* just zero the page if there's nothing in swap.  */
1179 		if (swslot == 0) {
1180 			/* page hasn't existed before, just zero it. */
1181 			uvm_pagezero(ptmp);
1182 		} else {
1183 			/*
1184 			 * page in the swapped-out page.
1185 			 * unlock object for i/o, relock when done.
1186 			 */
1187 
1188 			rw_exit(uobj->vmobjlock);
1189 			rv = uvm_swap_get(ptmp, swslot, PGO_SYNCIO);
1190 			rw_enter(uobj->vmobjlock, RW_WRITE);
1191 
1192 			/*
1193 			 * I/O done.  check for errors.
1194 			 */
1195 			if (rv != VM_PAGER_OK) {
1196 				/*
1197 				 * remove the swap slot from the aobj
1198 				 * and mark the aobj as having no real slot.
1199 				 * don't free the swap slot, thus preventing
1200 				 * it from being used again.
1201 				 */
1202 				swslot = uao_set_swslot(&aobj->u_obj, pageidx,
1203 							SWSLOT_BAD);
1204 				uvm_swap_markbad(swslot, 1);
1205 
1206 				if (ptmp->pg_flags & PG_WANTED)
1207 					wakeup(ptmp);
1208 				atomic_clearbits_int(&ptmp->pg_flags,
1209 				    PG_WANTED|PG_BUSY);
1210 				UVM_PAGE_OWN(ptmp, NULL);
1211 				uvm_lock_pageq();
1212 				uvm_pagefree(ptmp);
1213 				uvm_unlock_pageq();
1214 				rw_exit(uobj->vmobjlock);
1215 
1216 				return rv;
1217 			}
1218 		}
1219 
1220 		/*
1221  		 * we got the page!   clear the fake flag (indicates valid
1222 		 * data now in page) and plug into our result array.   note
1223 		 * that page is still busy.
1224  		 *
1225  		 * it is the callers job to:
1226  		 * => check if the page is released
1227  		 * => unbusy the page
1228  		 * => activate the page
1229  		 */
1230 		atomic_clearbits_int(&ptmp->pg_flags, PG_FAKE);
1231 		pmap_clear_modify(ptmp);		/* ... and clean */
1232 		pps[lcv] = ptmp;
1233 
1234 	}	/* lcv loop */
1235 
1236 	rw_exit(uobj->vmobjlock);
1237 	return VM_PAGER_OK;
1238 }
1239 
1240 /*
1241  * uao_dropswap:  release any swap resources from this aobj page.
1242  *
1243  * => aobj must be locked or have a reference count of 0.
1244  */
1245 int
1246 uao_dropswap(struct uvm_object *uobj, int pageidx)
1247 {
1248 	int slot;
1249 
1250 	KASSERT(UVM_OBJ_IS_AOBJ(uobj));
1251 
1252 	slot = uao_set_swslot(uobj, pageidx, 0);
1253 	if (slot) {
1254 		uvm_swap_free(slot, 1);
1255 	}
1256 	return slot;
1257 }
1258 
1259 /*
1260  * page in every page in every aobj that is paged-out to a range of swslots.
1261  *
1262  * => aobj must be locked and is returned locked.
1263  * => returns TRUE if pagein was aborted due to lack of memory.
1264  */
1265 boolean_t
1266 uao_swap_off(int startslot, int endslot)
1267 {
1268 	struct uvm_aobj *aobj;
1269 
1270 	/*
1271 	 * Walk the list of all anonymous UVM objects.  Grab the first.
1272 	 */
1273 	mtx_enter(&uao_list_lock);
1274 	if ((aobj = LIST_FIRST(&uao_list)) == NULL) {
1275 		mtx_leave(&uao_list_lock);
1276 		return FALSE;
1277 	}
1278 	uao_reference(&aobj->u_obj);
1279 
1280 	do {
1281 		struct uvm_aobj *nextaobj;
1282 		boolean_t rv;
1283 
1284 		/*
1285 		 * Prefetch the next object and immediately hold a reference
1286 		 * on it, so neither the current nor the next entry could
1287 		 * disappear while we are iterating.
1288 		 */
1289 		if ((nextaobj = LIST_NEXT(aobj, u_list)) != NULL) {
1290 			uao_reference(&nextaobj->u_obj);
1291 		}
1292 		mtx_leave(&uao_list_lock);
1293 
1294 		/*
1295 		 * Page in all pages in the swap slot range.
1296 		 */
1297 		rw_enter(aobj->u_obj.vmobjlock, RW_WRITE);
1298 		rv = uao_pagein(aobj, startslot, endslot);
1299 		rw_exit(aobj->u_obj.vmobjlock);
1300 
1301 		/* Drop the reference of the current object. */
1302 		uao_detach(&aobj->u_obj);
1303 		if (rv) {
1304 			if (nextaobj) {
1305 				uao_detach(&nextaobj->u_obj);
1306 			}
1307 			return rv;
1308 		}
1309 
1310 		aobj = nextaobj;
1311 		mtx_enter(&uao_list_lock);
1312 	} while (aobj);
1313 
1314 	/*
1315 	 * done with traversal, unlock the list
1316 	 */
1317 	mtx_leave(&uao_list_lock);
1318 	return FALSE;
1319 }
1320 
1321 /*
1322  * page in any pages from aobj in the given range.
1323  *
1324  * => returns TRUE if pagein was aborted due to lack of memory.
1325  */
1326 static boolean_t
1327 uao_pagein(struct uvm_aobj *aobj, int startslot, int endslot)
1328 {
1329 	boolean_t rv;
1330 
1331 	if (UAO_USES_SWHASH(aobj)) {
1332 		struct uao_swhash_elt *elt;
1333 		int bucket;
1334 
1335 restart:
1336 		for (bucket = aobj->u_swhashmask; bucket >= 0; bucket--) {
1337 			for (elt = LIST_FIRST(&aobj->u_swhash[bucket]);
1338 			     elt != NULL;
1339 			     elt = LIST_NEXT(elt, list)) {
1340 				int i;
1341 
1342 				for (i = 0; i < UAO_SWHASH_CLUSTER_SIZE; i++) {
1343 					int slot = elt->slots[i];
1344 
1345 					/*
1346 					 * if the slot isn't in range, skip it.
1347 					 */
1348 					if (slot < startslot ||
1349 					    slot >= endslot) {
1350 						continue;
1351 					}
1352 
1353 					/*
1354 					 * process the page,
1355 					 * the start over on this object
1356 					 * since the swhash elt
1357 					 * may have been freed.
1358 					 */
1359 					rv = uao_pagein_page(aobj,
1360 					  UAO_SWHASH_ELT_PAGEIDX_BASE(elt) + i);
1361 					if (rv) {
1362 						return rv;
1363 					}
1364 					goto restart;
1365 				}
1366 			}
1367 		}
1368 	} else {
1369 		int i;
1370 
1371 		for (i = 0; i < aobj->u_pages; i++) {
1372 			int slot = aobj->u_swslots[i];
1373 
1374 			/*
1375 			 * if the slot isn't in range, skip it
1376 			 */
1377 			if (slot < startslot || slot >= endslot) {
1378 				continue;
1379 			}
1380 
1381 			/*
1382 			 * process the page.
1383 			 */
1384 			rv = uao_pagein_page(aobj, i);
1385 			if (rv) {
1386 				return rv;
1387 			}
1388 		}
1389 	}
1390 
1391 	return FALSE;
1392 }
1393 
1394 /*
1395  * uao_pagein_page: page in a single page from an anonymous UVM object.
1396  *
1397  * => Returns TRUE if pagein was aborted due to lack of memory.
1398  */
1399 static boolean_t
1400 uao_pagein_page(struct uvm_aobj *aobj, int pageidx)
1401 {
1402 	struct uvm_object *uobj = &aobj->u_obj;
1403 	struct vm_page *pg;
1404 	int rv, slot, npages;
1405 
1406 	pg = NULL;
1407 	npages = 1;
1408 
1409 	KASSERT(rw_write_held(uobj->vmobjlock));
1410 	rv = uao_get(&aobj->u_obj, (voff_t)pageidx << PAGE_SHIFT,
1411 	    &pg, &npages, 0, PROT_READ | PROT_WRITE, 0, 0);
1412 
1413 	/*
1414 	 * relock and finish up.
1415 	 */
1416 	rw_enter(uobj->vmobjlock, RW_WRITE);
1417 	switch (rv) {
1418 	case VM_PAGER_OK:
1419 		break;
1420 
1421 	case VM_PAGER_ERROR:
1422 	case VM_PAGER_REFAULT:
1423 		/*
1424 		 * nothing more to do on errors.
1425 		 * VM_PAGER_REFAULT can only mean that the anon was freed,
1426 		 * so again there's nothing to do.
1427 		 */
1428 		return FALSE;
1429 	}
1430 
1431 	/*
1432 	 * ok, we've got the page now.
1433 	 * mark it as dirty, clear its swslot and un-busy it.
1434 	 */
1435 	slot = uao_set_swslot(&aobj->u_obj, pageidx, 0);
1436 	uvm_swap_free(slot, 1);
1437 	atomic_clearbits_int(&pg->pg_flags, PG_BUSY|PG_CLEAN|PG_FAKE);
1438 	UVM_PAGE_OWN(pg, NULL);
1439 
1440 	/*
1441 	 * deactivate the page (to put it on a page queue).
1442 	 */
1443 	pmap_clear_reference(pg);
1444 	uvm_lock_pageq();
1445 	uvm_pagedeactivate(pg);
1446 	uvm_unlock_pageq();
1447 
1448 	return FALSE;
1449 }
1450 
1451 /*
1452  * uao_dropswap_range: drop swapslots in the range.
1453  *
1454  * => aobj must be locked and is returned locked.
1455  * => start is inclusive.  end is exclusive.
1456  */
1457 void
1458 uao_dropswap_range(struct uvm_object *uobj, voff_t start, voff_t end)
1459 {
1460 	struct uvm_aobj *aobj = (struct uvm_aobj *)uobj;
1461 	int swpgonlydelta = 0;
1462 
1463 	KASSERT(UVM_OBJ_IS_AOBJ(uobj));
1464 	KASSERT(rw_write_held(uobj->vmobjlock));
1465 
1466 	if (end == 0) {
1467 		end = INT64_MAX;
1468 	}
1469 
1470 	if (UAO_USES_SWHASH(aobj)) {
1471 		int i, hashbuckets = aobj->u_swhashmask + 1;
1472 		voff_t taghi;
1473 		voff_t taglo;
1474 
1475 		taglo = UAO_SWHASH_ELT_TAG(start);
1476 		taghi = UAO_SWHASH_ELT_TAG(end);
1477 
1478 		for (i = 0; i < hashbuckets; i++) {
1479 			struct uao_swhash_elt *elt, *next;
1480 
1481 			for (elt = LIST_FIRST(&aobj->u_swhash[i]);
1482 			     elt != NULL;
1483 			     elt = next) {
1484 				int startidx, endidx;
1485 				int j;
1486 
1487 				next = LIST_NEXT(elt, list);
1488 
1489 				if (elt->tag < taglo || taghi < elt->tag) {
1490 					continue;
1491 				}
1492 
1493 				if (elt->tag == taglo) {
1494 					startidx =
1495 					    UAO_SWHASH_ELT_PAGESLOT_IDX(start);
1496 				} else {
1497 					startidx = 0;
1498 				}
1499 
1500 				if (elt->tag == taghi) {
1501 					endidx =
1502 					    UAO_SWHASH_ELT_PAGESLOT_IDX(end);
1503 				} else {
1504 					endidx = UAO_SWHASH_CLUSTER_SIZE;
1505 				}
1506 
1507 				for (j = startidx; j < endidx; j++) {
1508 					int slot = elt->slots[j];
1509 
1510 					KASSERT(uvm_pagelookup(&aobj->u_obj,
1511 					    (voff_t)(UAO_SWHASH_ELT_PAGEIDX_BASE(elt)
1512 					    + j) << PAGE_SHIFT) == NULL);
1513 
1514 					if (slot > 0) {
1515 						uvm_swap_free(slot, 1);
1516 						swpgonlydelta++;
1517 						KASSERT(elt->count > 0);
1518 						elt->slots[j] = 0;
1519 						elt->count--;
1520 					}
1521 				}
1522 
1523 				if (elt->count == 0) {
1524 					LIST_REMOVE(elt, list);
1525 					pool_put(&uao_swhash_elt_pool, elt);
1526 				}
1527 			}
1528 		}
1529 	} else {
1530 		int i;
1531 
1532 		if (aobj->u_pages < end) {
1533 			end = aobj->u_pages;
1534 		}
1535 		for (i = start; i < end; i++) {
1536 			int slot = aobj->u_swslots[i];
1537 
1538 			if (slot > 0) {
1539 				uvm_swap_free(slot, 1);
1540 				swpgonlydelta++;
1541 			}
1542 		}
1543 	}
1544 
1545 	/*
1546 	 * adjust the counter of pages only in swap for all
1547 	 * the swap slots we've freed.
1548 	 */
1549 	if (swpgonlydelta > 0) {
1550 		KASSERT(uvmexp.swpgonly >= swpgonlydelta);
1551 		atomic_add_int(&uvmexp.swpgonly, -swpgonlydelta);
1552 	}
1553 }
1554