xref: /dflybsd-src/sys/vm/swap_pager.c (revision ee65b806ac08b188bcab21ef0f1efda2cd5bdef7)
1 /*
2  * Copyright (c) 1998,2004 The DragonFly Project.  All rights reserved.
3  *
4  * This code is derived from software contributed to The DragonFly Project
5  * by Matthew Dillon <dillon@backplane.com>
6  *
7  * Redistribution and use in source and binary forms, with or without
8  * modification, are permitted provided that the following conditions
9  * are met:
10  *
11  * 1. Redistributions of source code must retain the above copyright
12  *    notice, this list of conditions and the following disclaimer.
13  * 2. Redistributions in binary form must reproduce the above copyright
14  *    notice, this list of conditions and the following disclaimer in
15  *    the documentation and/or other materials provided with the
16  *    distribution.
17  * 3. Neither the name of The DragonFly Project nor the names of its
18  *    contributors may be used to endorse or promote products derived
19  *    from this software without specific, prior written permission.
20  *
21  * THIS SOFTWARE IS PROVIDED BY THE COPYRIGHT HOLDERS AND CONTRIBUTORS
22  * ``AS IS'' AND ANY EXPRESS OR IMPLIED WARRANTIES, INCLUDING, BUT NOT
23  * LIMITED TO, THE IMPLIED WARRANTIES OF MERCHANTABILITY AND FITNESS
24  * FOR A PARTICULAR PURPOSE ARE DISCLAIMED.  IN NO EVENT SHALL THE
25  * COPYRIGHT HOLDERS OR CONTRIBUTORS BE LIABLE FOR ANY DIRECT, INDIRECT,
26  * INCIDENTAL, SPECIAL, EXEMPLARY OR CONSEQUENTIAL DAMAGES (INCLUDING,
27  * BUT NOT LIMITED TO, PROCUREMENT OF SUBSTITUTE GOODS OR SERVICES;
28  * LOSS OF USE, DATA, OR PROFITS; OR BUSINESS INTERRUPTION) HOWEVER CAUSED
29  * AND ON ANY THEORY OF LIABILITY, WHETHER IN CONTRACT, STRICT LIABILITY,
30  * OR TORT (INCLUDING NEGLIGENCE OR OTHERWISE) ARISING IN ANY WAY OUT
31  * OF THE USE OF THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF
32  * SUCH DAMAGE.
33  *
34  * Copyright (c) 1994 John S. Dyson
35  * Copyright (c) 1990 University of Utah.
36  * Copyright (c) 1991, 1993
37  *	The Regents of the University of California.  All rights reserved.
38  *
39  * This code is derived from software contributed to Berkeley by
40  * the Systems Programming Group of the University of Utah Computer
41  * Science Department.
42  *
43  * Redistribution and use in source and binary forms, with or without
44  * modification, are permitted provided that the following conditions
45  * are met:
46  * 1. Redistributions of source code must retain the above copyright
47  *    notice, this list of conditions and the following disclaimer.
48  * 2. Redistributions in binary form must reproduce the above copyright
49  *    notice, this list of conditions and the following disclaimer in the
50  *    documentation and/or other materials provided with the distribution.
51  * 3. All advertising materials mentioning features or use of this software
52  *    must display the following acknowledgement:
53  *	This product includes software developed by the University of
54  *	California, Berkeley and its contributors.
55  * 4. Neither the name of the University nor the names of its contributors
56  *    may be used to endorse or promote products derived from this software
57  *    without specific prior written permission.
58  *
59  * THIS SOFTWARE IS PROVIDED BY THE REGENTS AND CONTRIBUTORS ``AS IS'' AND
60  * ANY EXPRESS OR IMPLIED WARRANTIES, INCLUDING, BUT NOT LIMITED TO, THE
61  * IMPLIED WARRANTIES OF MERCHANTABILITY AND FITNESS FOR A PARTICULAR PURPOSE
62  * ARE DISCLAIMED.  IN NO EVENT SHALL THE REGENTS OR CONTRIBUTORS BE LIABLE
63  * FOR ANY DIRECT, INDIRECT, INCIDENTAL, SPECIAL, EXEMPLARY, OR CONSEQUENTIAL
64  * DAMAGES (INCLUDING, BUT NOT LIMITED TO, PROCUREMENT OF SUBSTITUTE GOODS
65  * OR SERVICES; LOSS OF USE, DATA, OR PROFITS; OR BUSINESS INTERRUPTION)
66  * HOWEVER CAUSED AND ON ANY THEORY OF LIABILITY, WHETHER IN CONTRACT, STRICT
67  * LIABILITY, OR TORT (INCLUDING NEGLIGENCE OR OTHERWISE) ARISING IN ANY WAY
68  * OUT OF THE USE OF THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF
69  * SUCH DAMAGE.
70  *
71  *				New Swap System
72  *				Matthew Dillon
73  *
74  * Radix Bitmap 'blists'.
75  *
76  *	- The new swapper uses the new radix bitmap code.  This should scale
77  *	  to arbitrarily small or arbitrarily large swap spaces and an almost
78  *	  arbitrary degree of fragmentation.
79  *
80  * Features:
81  *
82  *	- on the fly reallocation of swap during putpages.  The new system
83  *	  does not try to keep previously allocated swap blocks for dirty
84  *	  pages.
85  *
86  *	- on the fly deallocation of swap
87  *
88  *	- No more garbage collection required.  Unnecessarily allocated swap
89  *	  blocks only exist for dirty vm_page_t's now and these are already
90  *	  cycled (in a high-load system) by the pager.  We also do on-the-fly
91  *	  removal of invalidated swap blocks when a page is destroyed
92  *	  or renamed.
93  *
94  * from: Utah $Hdr: swap_pager.c 1.4 91/04/30$
95  *
96  *	@(#)swap_pager.c	8.9 (Berkeley) 3/21/94
97  *
98  * $FreeBSD: src/sys/vm/swap_pager.c,v 1.130.2.12 2002/08/31 21:15:55 dillon Exp $
99  * $DragonFly: src/sys/vm/swap_pager.c,v 1.32 2008/07/01 02:02:56 dillon Exp $
100  */
101 
102 #include <sys/param.h>
103 #include <sys/systm.h>
104 #include <sys/conf.h>
105 #include <sys/kernel.h>
106 #include <sys/proc.h>
107 #include <sys/buf.h>
108 #include <sys/vnode.h>
109 #include <sys/malloc.h>
110 #include <sys/vmmeter.h>
111 #include <sys/sysctl.h>
112 #include <sys/blist.h>
113 #include <sys/lock.h>
114 #include <sys/thread2.h>
115 
116 #ifndef MAX_PAGEOUT_CLUSTER
117 #define MAX_PAGEOUT_CLUSTER 16
118 #endif
119 
120 #define SWB_NPAGES	MAX_PAGEOUT_CLUSTER
121 
122 #include "opt_swap.h"
123 #include <vm/vm.h>
124 #include <vm/vm_object.h>
125 #include <vm/vm_page.h>
126 #include <vm/vm_pager.h>
127 #include <vm/vm_pageout.h>
128 #include <vm/swap_pager.h>
129 #include <vm/vm_extern.h>
130 #include <vm/vm_zone.h>
131 #include <vm/vnode_pager.h>
132 
133 #include <sys/buf2.h>
134 #include <vm/vm_page2.h>
135 
136 #define SWM_FREE	0x02	/* free, period			*/
137 #define SWM_POP		0x04	/* pop out			*/
138 
139 #define SWBIO_READ	0x01
140 #define SWBIO_WRITE	0x02
141 #define SWBIO_SYNC	0x04
142 
143 struct swfreeinfo {
144 	vm_object_t	object;
145 	vm_pindex_t	basei;
146 	vm_pindex_t	begi;
147 	vm_pindex_t	endi;	/* inclusive */
148 };
149 
150 /*
151  * vm_swap_size is in page-sized chunks now.  It was DEV_BSIZE'd chunks
152  * in the old system.
153  */
154 
155 int swap_pager_full;		/* swap space exhaustion (task killing) */
156 int vm_swap_cache_use;
157 int vm_swap_anon_use;
158 
159 static int swap_pager_almost_full; /* swap space exhaustion (w/ hysteresis)*/
160 static int nsw_rcount;		/* free read buffers			*/
161 static int nsw_wcount_sync;	/* limit write buffers / synchronous	*/
162 static int nsw_wcount_async;	/* limit write buffers / asynchronous	*/
163 static int nsw_wcount_async_max;/* assigned maximum			*/
164 static int nsw_cluster_max;	/* maximum VOP I/O allowed		*/
165 
166 struct blist *swapblist;
167 static int swap_async_max = 4;	/* maximum in-progress async I/O's	*/
168 static int swap_burst_read = 0;	/* allow burst reading */
169 
170 extern struct vnode *swapdev_vp;	/* from vm_swap.c */
171 
172 SYSCTL_INT(_vm, OID_AUTO, swap_async_max,
173         CTLFLAG_RW, &swap_async_max, 0, "Maximum running async swap ops");
174 SYSCTL_INT(_vm, OID_AUTO, swap_burst_read,
175         CTLFLAG_RW, &swap_burst_read, 0, "Allow burst reads for pageins");
176 
177 SYSCTL_INT(_vm, OID_AUTO, swap_cache_use,
178         CTLFLAG_RD, &vm_swap_cache_use, 0, "");
179 SYSCTL_INT(_vm, OID_AUTO, swap_anon_use,
180         CTLFLAG_RD, &vm_swap_anon_use, 0, "");
181 
182 vm_zone_t		swap_zone;
183 
184 /*
185  * Red-Black tree for swblock entries
186  */
187 RB_GENERATE2(swblock_rb_tree, swblock, swb_entry, rb_swblock_compare,
188 	     vm_pindex_t, swb_index);
189 
190 int
191 rb_swblock_compare(struct swblock *swb1, struct swblock *swb2)
192 {
193 	if (swb1->swb_index < swb2->swb_index)
194 		return(-1);
195 	if (swb1->swb_index > swb2->swb_index)
196 		return(1);
197 	return(0);
198 }
199 
200 static
201 int
202 rb_swblock_scancmp(struct swblock *swb, void *data)
203 {
204 	struct swfreeinfo *info = data;
205 
206 	if (swb->swb_index < info->basei)
207 		return(-1);
208 	if (swb->swb_index > info->endi)
209 		return(1);
210 	return(0);
211 }
212 
213 static
214 int
215 rb_swblock_condcmp(struct swblock *swb, void *data)
216 {
217 	struct swfreeinfo *info = data;
218 
219 	if (swb->swb_index < info->basei)
220 		return(-1);
221 	return(0);
222 }
223 
224 /*
225  * pagerops for OBJT_SWAP - "swap pager".  Some ops are also global procedure
226  * calls hooked from other parts of the VM system and do not appear here.
227  * (see vm/swap_pager.h).
228  */
229 
230 static void	swap_pager_dealloc (vm_object_t object);
231 static int	swap_pager_getpage (vm_object_t, vm_page_t *, int);
232 static void	swap_chain_iodone(struct bio *biox);
233 
234 struct pagerops swappagerops = {
235 	swap_pager_dealloc,	/* deallocate an OBJT_SWAP object	*/
236 	swap_pager_getpage,	/* pagein				*/
237 	swap_pager_putpages,	/* pageout				*/
238 	swap_pager_haspage	/* get backing store status for page	*/
239 };
240 
241 /*
242  * dmmax is in page-sized chunks with the new swap system.  It was
243  * dev-bsized chunks in the old.  dmmax is always a power of 2.
244  *
245  * swap_*() routines are externally accessible.  swp_*() routines are
246  * internal.
247  */
248 
249 int dmmax;
250 static int dmmax_mask;
251 int nswap_lowat = 128;		/* in pages, swap_pager_almost_full warn */
252 int nswap_hiwat = 512;		/* in pages, swap_pager_almost_full warn */
253 
254 static __inline void	swp_sizecheck (void);
255 static void	swp_pager_async_iodone (struct bio *bio);
256 
257 /*
258  * Swap bitmap functions
259  */
260 
261 static __inline void	swp_pager_freeswapspace (vm_object_t object, daddr_t blk, int npages);
262 static __inline daddr_t	swp_pager_getswapspace (vm_object_t object, int npages);
263 
264 /*
265  * Metadata functions
266  */
267 
268 static void swp_pager_meta_convert (vm_object_t);
269 static void swp_pager_meta_build (vm_object_t, vm_pindex_t, daddr_t);
270 static void swp_pager_meta_free (vm_object_t, vm_pindex_t, vm_pindex_t);
271 static void swp_pager_meta_free_all (vm_object_t);
272 static daddr_t swp_pager_meta_ctl (vm_object_t, vm_pindex_t, int);
273 
274 /*
275  * SWP_SIZECHECK() -	update swap_pager_full indication
276  *
277  *	update the swap_pager_almost_full indication and warn when we are
278  *	about to run out of swap space, using lowat/hiwat hysteresis.
279  *
280  *	Clear swap_pager_full ( task killing ) indication when lowat is met.
281  *
282  *	No restrictions on call
283  *	This routine may not block.
284  *	This routine must be called at splvm()
285  */
286 
287 static __inline void
288 swp_sizecheck(void)
289 {
290 	if (vm_swap_size < nswap_lowat) {
291 		if (swap_pager_almost_full == 0) {
292 			kprintf("swap_pager: out of swap space\n");
293 			swap_pager_almost_full = 1;
294 		}
295 	} else {
296 		swap_pager_full = 0;
297 		if (vm_swap_size > nswap_hiwat)
298 			swap_pager_almost_full = 0;
299 	}
300 }
301 
302 /*
303  * SWAP_PAGER_INIT() -	initialize the swap pager!
304  *
305  *	Expected to be started from system init.  NOTE:  This code is run
306  *	before much else so be careful what you depend on.  Most of the VM
307  *	system has yet to be initialized at this point.
308  */
309 static void
310 swap_pager_init(void *arg __unused)
311 {
312 	/*
313 	 * Device Stripe, in PAGE_SIZE'd blocks
314 	 */
315 	dmmax = SWB_NPAGES * 2;
316 	dmmax_mask = ~(dmmax - 1);
317 }
318 SYSINIT(vm_mem, SI_BOOT1_VM, SI_ORDER_THIRD, swap_pager_init, NULL)
319 
320 /*
321  * SWAP_PAGER_SWAP_INIT() - swap pager initialization from pageout process
322  *
323  *	Expected to be started from pageout process once, prior to entering
324  *	its main loop.
325  */
326 
327 void
328 swap_pager_swap_init(void)
329 {
330 	int n, n2;
331 
332 	/*
333 	 * Number of in-transit swap bp operations.  Don't
334 	 * exhaust the pbufs completely.  Make sure we
335 	 * initialize workable values (0 will work for hysteresis
336 	 * but it isn't very efficient).
337 	 *
338 	 * The nsw_cluster_max is constrained by the number of pages an XIO
339 	 * holds, i.e., (MAXPHYS/PAGE_SIZE) and our locally defined
340 	 * MAX_PAGEOUT_CLUSTER.   Also be aware that swap ops are
341 	 * constrained by the swap device interleave stripe size.
342 	 *
343 	 * Currently we hardwire nsw_wcount_async to 4.  This limit is
344 	 * designed to prevent other I/O from having high latencies due to
345 	 * our pageout I/O.  The value 4 works well for one or two active swap
346 	 * devices but is probably a little low if you have more.  Even so,
347 	 * a higher value would probably generate only a limited improvement
348 	 * with three or four active swap devices since the system does not
349 	 * typically have to pageout at extreme bandwidths.   We will want
350 	 * at least 2 per swap devices, and 4 is a pretty good value if you
351 	 * have one NFS swap device due to the command/ack latency over NFS.
352 	 * So it all works out pretty well.
353 	 */
354 
355 	nsw_cluster_max = min((MAXPHYS/PAGE_SIZE), MAX_PAGEOUT_CLUSTER);
356 
357 	nsw_rcount = (nswbuf + 1) / 2;
358 	nsw_wcount_sync = (nswbuf + 3) / 4;
359 	nsw_wcount_async = 4;
360 	nsw_wcount_async_max = nsw_wcount_async;
361 
362 	/*
363 	 * The zone is dynamically allocated so generally size it to
364 	 * maxswzone (32MB to 512MB of KVM).  Set a minimum size based
365 	 * on physical memory of around 8x (each swblock can hold 16 pages).
366 	 *
367 	 * With the advent of SSDs (vs HDs) the practical (swap:memory) ratio
368 	 * has increased dramatically.
369 	 */
370 	n = vmstats.v_page_count / 2;
371 	if (maxswzone && n < maxswzone / sizeof(struct swblock))
372 		n = maxswzone / sizeof(struct swblock);
373 	n2 = n;
374 
375 	do {
376 		swap_zone = zinit(
377 			"SWAPMETA",
378 			sizeof(struct swblock),
379 			n,
380 			ZONE_INTERRUPT,
381 			1);
382 		if (swap_zone != NULL)
383 			break;
384 		/*
385 		 * if the allocation failed, try a zone two thirds the
386 		 * size of the previous attempt.
387 		 */
388 		n -= ((n + 2) / 3);
389 	} while (n > 0);
390 
391 	if (swap_zone == NULL)
392 		panic("swap_pager_swap_init: swap_zone == NULL");
393 	if (n2 != n)
394 		kprintf("Swap zone entries reduced from %d to %d.\n", n2, n);
395 }
396 
397 /*
398  * SWAP_PAGER_ALLOC() -	allocate a new OBJT_SWAP VM object and instantiate
399  *			its metadata structures.
400  *
401  *	This routine is called from the mmap and fork code to create a new
402  *	OBJT_SWAP object.  We do this by creating an OBJT_DEFAULT object
403  *	and then converting it with swp_pager_meta_convert().
404  *
405  *	This routine may block in vm_object_allocate() and create a named
406  *	object lookup race, so we must interlock.   We must also run at
407  *	splvm() for the object lookup to handle races with interrupts, but
408  *	we do not have to maintain splvm() in between the lookup and the
409  *	add because (I believe) it is not possible to attempt to create
410  *	a new swap object w/handle when a default object with that handle
411  *	already exists.
412  */
413 
414 vm_object_t
415 swap_pager_alloc(void *handle, off_t size, vm_prot_t prot, off_t offset)
416 {
417 	vm_object_t object;
418 
419 	KKASSERT(handle == NULL);
420 #if 0
421 	if (handle) {
422 		/*
423 		 * Reference existing named region or allocate new one.  There
424 		 * should not be a race here against swp_pager_meta_build()
425 		 * as called from vm_page_remove() in regards to the lookup
426 		 * of the handle.
427 		 */
428 		while (sw_alloc_interlock) {
429 			sw_alloc_interlock = -1;
430 			tsleep(&sw_alloc_interlock, 0, "swpalc", 0);
431 		}
432 		sw_alloc_interlock = 1;
433 
434 		object = vm_pager_object_lookup(NOBJLIST(handle), handle);
435 
436 		if (object != NULL) {
437 			vm_object_reference(object);
438 		} else {
439 			object = vm_object_allocate(OBJT_DEFAULT,
440 				OFF_TO_IDX(offset + PAGE_MASK + size));
441 			object->handle = handle;
442 			swp_pager_meta_convert(object);
443 		}
444 
445 		if (sw_alloc_interlock < 0)
446 			wakeup(&sw_alloc_interlock);
447 		sw_alloc_interlock = 0;
448 	} else { ... }
449 #endif
450 	object = vm_object_allocate(OBJT_DEFAULT,
451 				    OFF_TO_IDX(offset + PAGE_MASK + size));
452 	swp_pager_meta_convert(object);
453 
454 	return (object);
455 }
456 
457 /*
458  * SWAP_PAGER_DEALLOC() -	remove swap metadata from object
459  *
460  *	The swap backing for the object is destroyed.  The code is
461  *	designed such that we can reinstantiate it later, but this
462  *	routine is typically called only when the entire object is
463  *	about to be destroyed.
464  *
465  *	This routine may block, but no longer does.
466  *
467  *	The object must be locked or unreferenceable.
468  */
469 
470 static void
471 swap_pager_dealloc(vm_object_t object)
472 {
473 	vm_object_pip_wait(object, "swpdea");
474 
475 	/*
476 	 * Free all remaining metadata.  We only bother to free it from
477 	 * the swap meta data.  We do not attempt to free swapblk's still
478 	 * associated with vm_page_t's for this object.  We do not care
479 	 * if paging is still in progress on some objects.
480 	 */
481 	crit_enter();
482 	swp_pager_meta_free_all(object);
483 	crit_exit();
484 }
485 
486 /************************************************************************
487  *			SWAP PAGER BITMAP ROUTINES			*
488  ************************************************************************/
489 
490 /*
491  * SWP_PAGER_GETSWAPSPACE() -	allocate raw swap space
492  *
493  *	Allocate swap for the requested number of pages.  The starting
494  *	swap block number (a page index) is returned or SWAPBLK_NONE
495  *	if the allocation failed.
496  *
497  *	Also has the side effect of advising that somebody made a mistake
498  *	when they configured swap and didn't configure enough.
499  *
500  *	Must be called at splvm() to avoid races with bitmap frees from
501  *	vm_page_remove() aka swap_pager_page_removed().
502  *
503  *	This routine may not block
504  *	This routine must be called at splvm().
505  */
506 static __inline daddr_t
507 swp_pager_getswapspace(vm_object_t object, int npages)
508 {
509 	daddr_t blk;
510 
511 	if ((blk = blist_alloc(swapblist, npages)) == SWAPBLK_NONE) {
512 		if (swap_pager_full != 2) {
513 			kprintf("swap_pager_getswapspace: failed\n");
514 			swap_pager_full = 2;
515 			swap_pager_almost_full = 1;
516 		}
517 	} else {
518 		vm_swap_size -= npages;
519 		if (object->type == OBJT_SWAP)
520 			vm_swap_anon_use += npages;
521 		else
522 			vm_swap_cache_use += npages;
523 		swp_sizecheck();
524 	}
525 	return(blk);
526 }
527 
528 /*
529  * SWP_PAGER_FREESWAPSPACE() -	free raw swap space
530  *
531  *	This routine returns the specified swap blocks back to the bitmap.
532  *
533  *	Note:  This routine may not block (it could in the old swap code),
534  *	and through the use of the new blist routines it does not block.
535  *
536  *	We must be called at splvm() to avoid races with bitmap frees from
537  *	vm_page_remove() aka swap_pager_page_removed().
538  *
539  *	This routine may not block
540  *	This routine must be called at splvm().
541  */
542 
543 static __inline void
544 swp_pager_freeswapspace(vm_object_t object, daddr_t blk, int npages)
545 {
546 	blist_free(swapblist, blk, npages);
547 	vm_swap_size += npages;
548 	if (object->type == OBJT_SWAP)
549 		vm_swap_anon_use -= npages;
550 	else
551 		vm_swap_cache_use -= npages;
552 	swp_sizecheck();
553 }
554 
555 /*
556  * SWAP_PAGER_FREESPACE() -	frees swap blocks associated with a page
557  *				range within an object.
558  *
559  *	This is a globally accessible routine.
560  *
561  *	This routine removes swapblk assignments from swap metadata.
562  *
563  *	The external callers of this routine typically have already destroyed
564  *	or renamed vm_page_t's associated with this range in the object so
565  *	we should be ok.
566  *
567  *	This routine may be called at any spl.  We up our spl to splvm
568  *	temporarily in order to perform the metadata removal.
569  */
570 void
571 swap_pager_freespace(vm_object_t object, vm_pindex_t start, vm_pindex_t size)
572 {
573 	crit_enter();
574 	swp_pager_meta_free(object, start, size);
575 	crit_exit();
576 }
577 
578 void
579 swap_pager_freespace_all(vm_object_t object)
580 {
581 	crit_enter();
582 	swp_pager_meta_free_all(object);
583 	crit_exit();
584 }
585 
586 /*
587  * This function conditionally frees swap cache swap starting at
588  * (*basei) in the object.  (count) swap blocks will be nominally freed.
589  * The actual number of blocks freed can be more or less than the
590  * requested number.
591  *
592  * This function nominally returns the number of blocks freed.  However,
593  * the actual number of blocks freed may be less then the returned value.
594  * If the function is unable to exhaust the object or if it is able to
595  * free (approximately) the requested number of blocks it returns
596  * a value n > count.
597  *
598  * If we exhaust the object we will return a value n <= count.
599  *
600  * Must be called from a critical section.
601  */
602 static int swap_pager_condfree_callback(struct swblock *swap, void *data);
603 
604 int
605 swap_pager_condfree(vm_object_t object, vm_pindex_t *basei, int count)
606 {
607 	struct swfreeinfo info;
608 
609 	info.object = object;
610 	info.basei = *basei;	/* skip up to this page index */
611 	info.begi = count;	/* max swap pages to destroy */
612 	info.endi = count * 8;	/* max swblocks to scan */
613 
614 	swblock_rb_tree_RB_SCAN(&object->swblock_root, rb_swblock_condcmp,
615 				swap_pager_condfree_callback, &info);
616 	*basei = info.basei;
617 	if (info.endi < 0 && info.begi <= count)
618 		info.begi = count + 1;
619 	return(count - (int)info.begi);
620 }
621 
622 /*
623  * The idea is to free whole meta-block to avoid fragmenting
624  * the swap space or disk I/O.  We only do this if NO VM pages
625  * are present.
626  *
627  * We do not have to deal with clearing PG_SWAPPED in related VM
628  * pages because there are no related VM pages.
629  */
630 static int
631 swap_pager_condfree_callback(struct swblock *swap, void *data)
632 {
633 	struct swfreeinfo *info = data;
634 	vm_object_t object = info->object;
635 	int i;
636 
637 	for (i = 0; i < SWAP_META_PAGES; ++i) {
638 		if (vm_page_lookup(object, swap->swb_index + i))
639 			break;
640 	}
641 	info->basei = swap->swb_index + SWAP_META_PAGES;
642 	if (i == SWAP_META_PAGES) {
643 		info->begi -= swap->swb_count;
644 		swap_pager_freespace(object, swap->swb_index, SWAP_META_PAGES);
645 	}
646 	--info->endi;
647 	if ((int)info->begi < 0 || (int)info->endi < 0)
648 		return(-1);
649 	return(0);
650 }
651 
652 /*
653  * Called by vm_page_alloc() when a new VM page is inserted
654  * into a VM object.  Checks whether swap has been assigned to
655  * the page and sets PG_SWAPPED as necessary.
656  */
657 void
658 swap_pager_page_inserted(vm_page_t m)
659 {
660 	if (m->object->swblock_count) {
661 		crit_enter();
662 		if (swp_pager_meta_ctl(m->object, m->pindex, 0) != SWAPBLK_NONE)
663 			vm_page_flag_set(m, PG_SWAPPED);
664 		crit_exit();
665 	}
666 }
667 
668 /*
669  * SWAP_PAGER_RESERVE() - reserve swap blocks in object
670  *
671  *	Assigns swap blocks to the specified range within the object.  The
672  *	swap blocks are not zerod.  Any previous swap assignment is destroyed.
673  *
674  *	Returns 0 on success, -1 on failure.
675  */
676 int
677 swap_pager_reserve(vm_object_t object, vm_pindex_t start, vm_size_t size)
678 {
679 	int n = 0;
680 	daddr_t blk = SWAPBLK_NONE;
681 	vm_pindex_t beg = start;	/* save start index */
682 
683 	crit_enter();
684 	while (size) {
685 		if (n == 0) {
686 			n = BLIST_MAX_ALLOC;
687 			while ((blk = swp_pager_getswapspace(object, n)) ==
688 			       SWAPBLK_NONE)
689 			{
690 				n >>= 1;
691 				if (n == 0) {
692 					swp_pager_meta_free(object, beg,
693 							    start - beg);
694 					crit_exit();
695 					return(-1);
696 				}
697 			}
698 		}
699 		swp_pager_meta_build(object, start, blk);
700 		--size;
701 		++start;
702 		++blk;
703 		--n;
704 	}
705 	swp_pager_meta_free(object, start, n);
706 	crit_exit();
707 	return(0);
708 }
709 
710 /*
711  * SWAP_PAGER_COPY() -  copy blocks from source pager to destination pager
712  *			and destroy the source.
713  *
714  *	Copy any valid swapblks from the source to the destination.  In
715  *	cases where both the source and destination have a valid swapblk,
716  *	we keep the destination's.
717  *
718  *	This routine is allowed to block.  It may block allocating metadata
719  *	indirectly through swp_pager_meta_build() or if paging is still in
720  *	progress on the source.
721  *
722  *	This routine can be called at any spl
723  *
724  *	XXX vm_page_collapse() kinda expects us not to block because we
725  *	supposedly do not need to allocate memory, but for the moment we
726  *	*may* have to get a little memory from the zone allocator, but
727  *	it is taken from the interrupt memory.  We should be ok.
728  *
729  *	The source object contains no vm_page_t's (which is just as well)
730  *
731  *	The source object is of type OBJT_SWAP.
732  *
733  *	The source and destination objects must be locked or
734  *	inaccessible (XXX are they ?)
735  */
736 
737 void
738 swap_pager_copy(vm_object_t srcobject, vm_object_t dstobject,
739 		vm_pindex_t base_index, int destroysource)
740 {
741 	vm_pindex_t i;
742 
743 	crit_enter();
744 
745 	/*
746 	 * transfer source to destination.
747 	 */
748 	for (i = 0; i < dstobject->size; ++i) {
749 		daddr_t dstaddr;
750 
751 		/*
752 		 * Locate (without changing) the swapblk on the destination,
753 		 * unless it is invalid in which case free it silently, or
754 		 * if the destination is a resident page, in which case the
755 		 * source is thrown away.
756 		 */
757 		dstaddr = swp_pager_meta_ctl(dstobject, i, 0);
758 
759 		if (dstaddr == SWAPBLK_NONE) {
760 			/*
761 			 * Destination has no swapblk and is not resident,
762 			 * copy source.
763 			 */
764 			daddr_t srcaddr;
765 
766 			srcaddr = swp_pager_meta_ctl(srcobject,
767 						     base_index + i, SWM_POP);
768 
769 			if (srcaddr != SWAPBLK_NONE)
770 				swp_pager_meta_build(dstobject, i, srcaddr);
771 		} else {
772 			/*
773 			 * Destination has valid swapblk or it is represented
774 			 * by a resident page.  We destroy the sourceblock.
775 			 */
776 			swp_pager_meta_ctl(srcobject, base_index + i, SWM_FREE);
777 		}
778 	}
779 
780 	/*
781 	 * Free left over swap blocks in source.
782 	 *
783 	 * We have to revert the type to OBJT_DEFAULT so we do not accidently
784 	 * double-remove the object from the swap queues.
785 	 */
786 	if (destroysource) {
787 		/*
788 		 * Reverting the type is not necessary, the caller is going
789 		 * to destroy srcobject directly, but I'm doing it here
790 		 * for consistency since we've removed the object from its
791 		 * queues.
792 		 */
793 		swp_pager_meta_free_all(srcobject);
794 		if (srcobject->type == OBJT_SWAP)
795 			srcobject->type = OBJT_DEFAULT;
796 	}
797 	crit_exit();
798 }
799 
800 /*
801  * SWAP_PAGER_HASPAGE() -	determine if we have good backing store for
802  *				the requested page.
803  *
804  *	We determine whether good backing store exists for the requested
805  *	page and return TRUE if it does, FALSE if it doesn't.
806  *
807  *	If TRUE, we also try to determine how much valid, contiguous backing
808  *	store exists before and after the requested page within a reasonable
809  *	distance.  We do not try to restrict it to the swap device stripe
810  *	(that is handled in getpages/putpages).  It probably isn't worth
811  *	doing here.
812  */
813 
814 boolean_t
815 swap_pager_haspage(vm_object_t object, vm_pindex_t pindex)
816 {
817 	daddr_t blk0;
818 
819 	/*
820 	 * do we have good backing store at the requested index ?
821 	 */
822 
823 	crit_enter();
824 	blk0 = swp_pager_meta_ctl(object, pindex, 0);
825 
826 	if (blk0 == SWAPBLK_NONE) {
827 		crit_exit();
828 		return (FALSE);
829 	}
830 
831 #if 0
832 	/*
833 	 * find backwards-looking contiguous good backing store
834 	 */
835 	if (before != NULL) {
836 		int i;
837 
838 		for (i = 1; i < (SWB_NPAGES/2); ++i) {
839 			daddr_t blk;
840 
841 			if (i > pindex)
842 				break;
843 			blk = swp_pager_meta_ctl(object, pindex - i, 0);
844 			if (blk != blk0 - i)
845 				break;
846 		}
847 		*before = (i - 1);
848 	}
849 
850 	/*
851 	 * find forward-looking contiguous good backing store
852 	 */
853 
854 	if (after != NULL) {
855 		int i;
856 
857 		for (i = 1; i < (SWB_NPAGES/2); ++i) {
858 			daddr_t blk;
859 
860 			blk = swp_pager_meta_ctl(object, pindex + i, 0);
861 			if (blk != blk0 + i)
862 				break;
863 		}
864 		*after = (i - 1);
865 	}
866 #endif
867 	crit_exit();
868 	return (TRUE);
869 }
870 
871 /*
872  * SWAP_PAGER_PAGE_UNSWAPPED() - remove swap backing store related to page
873  *
874  * This removes any associated swap backing store, whether valid or
875  * not, from the page.  This operates on any VM object, not just OBJT_SWAP
876  * objects.
877  *
878  * This routine is typically called when a page is made dirty, at
879  * which point any associated swap can be freed.  MADV_FREE also
880  * calls us in a special-case situation
881  *
882  * NOTE!!!  If the page is clean and the swap was valid, the caller
883  * should make the page dirty before calling this routine.  This routine
884  * does NOT change the m->dirty status of the page.  Also: MADV_FREE
885  * depends on it.
886  *
887  * This routine may not block.
888  *
889  * The page must be busied or soft-busied.
890  */
891 void
892 swap_pager_unswapped(vm_page_t m)
893 {
894 	if (m->flags & PG_SWAPPED) {
895 		crit_enter();
896 		KKASSERT(m->flags & PG_SWAPPED);
897 		swp_pager_meta_ctl(m->object, m->pindex, SWM_FREE);
898 		vm_page_flag_clear(m, PG_SWAPPED);
899 		crit_exit();
900 	}
901 }
902 
903 /*
904  * SWAP_PAGER_STRATEGY() - read, write, free blocks
905  *
906  * This implements a VM OBJECT strategy function using swap backing store.
907  * This can operate on any VM OBJECT type, not necessarily just OBJT_SWAP
908  * types.
909  *
910  * This is intended to be a cacheless interface (i.e. caching occurs at
911  * higher levels), and is also used as a swap-based SSD cache for vnode
912  * and device objects.
913  *
914  * All I/O goes directly to and from the swap device.
915  *
916  * We currently attempt to run I/O synchronously or asynchronously as
917  * the caller requests.  This isn't perfect because we loose error
918  * sequencing when we run multiple ops in parallel to satisfy a request.
919  * But this is swap, so we let it all hang out.
920  */
921 void
922 swap_pager_strategy(vm_object_t object, struct bio *bio)
923 {
924 	struct buf *bp = bio->bio_buf;
925 	struct bio *nbio;
926 	vm_pindex_t start;
927 	vm_pindex_t biox_blkno = 0;
928 	int count;
929 	char *data;
930 	struct bio *biox;
931 	struct buf *bufx;
932 	struct bio_track *track;
933 
934 	/*
935 	 * tracking for swapdev vnode I/Os
936 	 */
937 	if (bp->b_cmd == BUF_CMD_READ)
938 		track = &swapdev_vp->v_track_read;
939 	else
940 		track = &swapdev_vp->v_track_write;
941 
942 	if (bp->b_bcount & PAGE_MASK) {
943 		bp->b_error = EINVAL;
944 		bp->b_flags |= B_ERROR | B_INVAL;
945 		biodone(bio);
946 		kprintf("swap_pager_strategy: bp %p offset %lld size %d, "
947 			"not page bounded\n",
948 			bp, (long long)bio->bio_offset, (int)bp->b_bcount);
949 		return;
950 	}
951 
952 	/*
953 	 * Clear error indication, initialize page index, count, data pointer.
954 	 */
955 	bp->b_error = 0;
956 	bp->b_flags &= ~B_ERROR;
957 	bp->b_resid = bp->b_bcount;
958 
959 	start = (vm_pindex_t)(bio->bio_offset >> PAGE_SHIFT);
960 	count = howmany(bp->b_bcount, PAGE_SIZE);
961 	data = bp->b_data;
962 
963 	/*
964 	 * Deal with BUF_CMD_FREEBLKS
965 	 */
966 	if (bp->b_cmd == BUF_CMD_FREEBLKS) {
967 		/*
968 		 * FREE PAGE(s) - destroy underlying swap that is no longer
969 		 *		  needed.
970 		 */
971 		crit_enter();
972 		swp_pager_meta_free(object, start, count);
973 		crit_exit();
974 		bp->b_resid = 0;
975 		biodone(bio);
976 		return;
977 	}
978 
979 	/*
980 	 * We need to be able to create a new cluster of I/O's.  We cannot
981 	 * use the caller fields of the passed bio so push a new one.
982 	 *
983 	 * Because nbio is just a placeholder for the cluster links,
984 	 * we can biodone() the original bio instead of nbio to make
985 	 * things a bit more efficient.
986 	 */
987 	nbio = push_bio(bio);
988 	nbio->bio_offset = bio->bio_offset;
989 	nbio->bio_caller_info1.cluster_head = NULL;
990 	nbio->bio_caller_info2.cluster_tail = NULL;
991 
992 	biox = NULL;
993 	bufx = NULL;
994 
995 	/*
996 	 * Execute read or write
997 	 */
998 	crit_enter();
999 	while (count > 0) {
1000 		daddr_t blk;
1001 
1002 		/*
1003 		 * Obtain block.  If block not found and writing, allocate a
1004 		 * new block and build it into the object.
1005 		 */
1006 		blk = swp_pager_meta_ctl(object, start, 0);
1007 		if ((blk == SWAPBLK_NONE) && bp->b_cmd != BUF_CMD_READ) {
1008 			blk = swp_pager_getswapspace(object, 1);
1009 			if (blk == SWAPBLK_NONE) {
1010 				bp->b_error = ENOMEM;
1011 				bp->b_flags |= B_ERROR;
1012 				break;
1013 			}
1014 			swp_pager_meta_build(object, start, blk);
1015 		}
1016 
1017 		/*
1018 		 * Do we have to flush our current collection?  Yes if:
1019 		 *
1020 		 *	- no swap block at this index
1021 		 *	- swap block is not contiguous
1022 		 *	- we cross a physical disk boundry in the
1023 		 *	  stripe.
1024 		 */
1025 		if (
1026 		    biox && (biox_blkno + btoc(bufx->b_bcount) != blk ||
1027 		     ((biox_blkno ^ blk) & dmmax_mask)
1028 		    )
1029 		) {
1030 			if (bp->b_cmd == BUF_CMD_READ) {
1031 				++mycpu->gd_cnt.v_swapin;
1032 				mycpu->gd_cnt.v_swappgsin += btoc(bufx->b_bcount);
1033 			} else {
1034 				++mycpu->gd_cnt.v_swapout;
1035 				mycpu->gd_cnt.v_swappgsout += btoc(bufx->b_bcount);
1036 				bufx->b_dirtyend = bufx->b_bcount;
1037 			}
1038 
1039 			/*
1040 			 * Finished with this buf.
1041 			 */
1042 			KKASSERT(bufx->b_bcount != 0);
1043 			if (bufx->b_cmd != BUF_CMD_READ)
1044 				bufx->b_dirtyend = bufx->b_bcount;
1045 			biox = NULL;
1046 			bufx = NULL;
1047 		}
1048 
1049 		/*
1050 		 * Add new swapblk to biox, instantiating biox if necessary.
1051 		 * Zero-fill reads are able to take a shortcut.
1052 		 */
1053 		if (blk == SWAPBLK_NONE) {
1054 			/*
1055 			 * We can only get here if we are reading.  Since
1056 			 * we are at splvm() we can safely modify b_resid,
1057 			 * even if chain ops are in progress.
1058 			 */
1059 			bzero(data, PAGE_SIZE);
1060 			bp->b_resid -= PAGE_SIZE;
1061 		} else {
1062 			if (biox == NULL) {
1063 				/* XXX chain count > 4, wait to <= 4 */
1064 
1065 				bufx = getpbuf(NULL);
1066 				biox = &bufx->b_bio1;
1067 				cluster_append(nbio, bufx);
1068 				bufx->b_flags |= (bufx->b_flags & B_ORDERED);
1069 				bufx->b_cmd = bp->b_cmd;
1070 				biox->bio_done = swap_chain_iodone;
1071 				biox->bio_offset = (off_t)blk << PAGE_SHIFT;
1072 				biox->bio_caller_info1.cluster_parent = nbio;
1073 				biox_blkno = blk;
1074 				bufx->b_bcount = 0;
1075 				bufx->b_data = data;
1076 			}
1077 			bufx->b_bcount += PAGE_SIZE;
1078 		}
1079 		--count;
1080 		++start;
1081 		data += PAGE_SIZE;
1082 	}
1083 	crit_exit();
1084 
1085 	/*
1086 	 *  Flush out last buffer
1087 	 */
1088 	if (biox) {
1089 		if (bufx->b_cmd == BUF_CMD_READ) {
1090 			++mycpu->gd_cnt.v_swapin;
1091 			mycpu->gd_cnt.v_swappgsin += btoc(bufx->b_bcount);
1092 		} else {
1093 			++mycpu->gd_cnt.v_swapout;
1094 			mycpu->gd_cnt.v_swappgsout += btoc(bufx->b_bcount);
1095 			bufx->b_dirtyend = bufx->b_bcount;
1096 		}
1097 		KKASSERT(bufx->b_bcount);
1098 		if (bufx->b_cmd != BUF_CMD_READ)
1099 			bufx->b_dirtyend = bufx->b_bcount;
1100 		/* biox, bufx = NULL */
1101 	}
1102 
1103 	/*
1104 	 * Now initiate all the I/O.  Be careful looping on our chain as
1105 	 * I/O's may complete while we are still initiating them.
1106 	 *
1107 	 * If the request is a 100% sparse read no bios will be present
1108 	 * and we just biodone() the buffer.
1109 	 */
1110 	nbio->bio_caller_info2.cluster_tail = NULL;
1111 	bufx = nbio->bio_caller_info1.cluster_head;
1112 
1113 	if (bufx) {
1114 		while (bufx) {
1115 			biox = &bufx->b_bio1;
1116 			BUF_KERNPROC(bufx);
1117 			bufx = bufx->b_cluster_next;
1118 			vn_strategy(swapdev_vp, biox);
1119 		}
1120 	} else {
1121 		biodone(bio);
1122 	}
1123 
1124 	/*
1125 	 * Completion of the cluster will also call biodone_chain(nbio).
1126 	 * We never call biodone(nbio) so we don't have to worry about
1127 	 * setting up a bio_done callback.  It's handled in the sub-IO.
1128 	 */
1129 	/**/
1130 }
1131 
1132 static void
1133 swap_chain_iodone(struct bio *biox)
1134 {
1135 	struct buf **nextp;
1136 	struct buf *bufx;	/* chained sub-buffer */
1137 	struct bio *nbio;	/* parent nbio with chain glue */
1138 	struct buf *bp;		/* original bp associated with nbio */
1139 	int chain_empty;
1140 
1141 	bufx = biox->bio_buf;
1142 	nbio = biox->bio_caller_info1.cluster_parent;
1143 	bp = nbio->bio_buf;
1144 
1145 	/*
1146 	 * Update the original buffer
1147 	 */
1148         KKASSERT(bp != NULL);
1149 	if (bufx->b_flags & B_ERROR) {
1150 		atomic_set_int(&bufx->b_flags, B_ERROR);
1151 		bp->b_error = bufx->b_error;
1152 	} else if (bufx->b_resid != 0) {
1153 		atomic_set_int(&bufx->b_flags, B_ERROR);
1154 		bp->b_error = EINVAL;
1155 	} else {
1156 		atomic_subtract_int(&bp->b_resid, bufx->b_bcount);
1157 	}
1158 
1159 	/*
1160 	 * Remove us from the chain.
1161 	 */
1162 	spin_lock_wr(&bp->b_lock.lk_spinlock);
1163 	nextp = &nbio->bio_caller_info1.cluster_head;
1164 	while (*nextp != bufx) {
1165 		KKASSERT(*nextp != NULL);
1166 		nextp = &(*nextp)->b_cluster_next;
1167 	}
1168 	*nextp = bufx->b_cluster_next;
1169 	chain_empty = (nbio->bio_caller_info1.cluster_head == NULL);
1170 	spin_unlock_wr(&bp->b_lock.lk_spinlock);
1171 
1172 	/*
1173 	 * Clean up bufx.  If the chain is now empty we finish out
1174 	 * the parent.  Note that we may be racing other completions
1175 	 * so we must use the chain_empty status from above.
1176 	 */
1177 	if (chain_empty) {
1178 		if (bp->b_resid != 0 && !(bp->b_flags & B_ERROR)) {
1179 			atomic_set_int(&bp->b_flags, B_ERROR);
1180 			bp->b_error = EINVAL;
1181 		}
1182 		biodone_chain(nbio);
1183         }
1184         relpbuf(bufx, NULL);
1185 }
1186 
1187 /*
1188  * SWAP_PAGER_GETPAGES() - bring page in from swap
1189  *
1190  * The requested page may have to be brought in from swap.  Calculate the
1191  * swap block and bring in additional pages if possible.  All pages must
1192  * have contiguous swap block assignments and reside in the same object.
1193  *
1194  * The caller has a single vm_object_pip_add() reference prior to
1195  * calling us and we should return with the same.
1196  *
1197  * The caller has BUSY'd the page.  We should return with (*mpp) left busy,
1198  * and any additinal pages unbusied.
1199  *
1200  * If the caller encounters a PG_RAM page it will pass it to us even though
1201  * it may be valid and dirty.  We cannot overwrite the page in this case!
1202  * The case is used to allow us to issue pure read-aheads.
1203  *
1204  * NOTE! XXX This code does not entirely pipeline yet due to the fact that
1205  *       the PG_RAM page is validated at the same time as mreq.  What we
1206  *	 really need to do is issue a separate read-ahead pbuf.
1207  */
1208 static int
1209 swap_pager_getpage(vm_object_t object, vm_page_t *mpp, int seqaccess)
1210 {
1211 	struct buf *bp;
1212 	struct bio *bio;
1213 	vm_page_t mreq;
1214 	vm_page_t m;
1215 	vm_offset_t kva;
1216 	daddr_t blk;
1217 	int i;
1218 	int j;
1219 	int raonly;
1220 	vm_page_t marray[XIO_INTERNAL_PAGES];
1221 
1222 	mreq = *mpp;
1223 
1224 	if (mreq->object != object) {
1225 		panic("swap_pager_getpages: object mismatch %p/%p",
1226 		    object,
1227 		    mreq->object
1228 		);
1229 	}
1230 
1231 	/*
1232 	 * We don't want to overwrite a fully valid page as it might be
1233 	 * dirty.  This case can occur when e.g. vm_fault hits a perfectly
1234 	 * valid page with PG_RAM set.
1235 	 *
1236 	 * In this case we see if the next page is a suitable page-in
1237 	 * candidate and if it is we issue read-ahead.  PG_RAM will be
1238 	 * set on the last page of the read-ahead to continue the pipeline.
1239 	 */
1240 	if (mreq->valid == VM_PAGE_BITS_ALL) {
1241 		if (swap_burst_read == 0 || mreq->pindex + 1 >= object->size)
1242 			return(VM_PAGER_OK);
1243 		crit_enter();
1244 		blk = swp_pager_meta_ctl(object, mreq->pindex + 1, 0);
1245 		if (blk == SWAPBLK_NONE) {
1246 			crit_exit();
1247 			return(VM_PAGER_OK);
1248 		}
1249 		m = vm_page_lookup(object, mreq->pindex + 1);
1250 		if (m == NULL) {
1251 			m = vm_page_alloc(object, mreq->pindex + 1,
1252 					  VM_ALLOC_QUICK);
1253 			if (m == NULL) {
1254 				crit_exit();
1255 				return(VM_PAGER_OK);
1256 			}
1257 		} else {
1258 			if ((m->flags & PG_BUSY) || m->busy || m->valid) {
1259 				crit_exit();
1260 				return(VM_PAGER_OK);
1261 			}
1262 			vm_page_unqueue_nowakeup(m);
1263 			vm_page_busy(m);
1264 		}
1265 		mreq = m;
1266 		raonly = 1;
1267 		crit_exit();
1268 	} else {
1269 		raonly = 0;
1270 	}
1271 
1272 	/*
1273 	 * Try to block-read contiguous pages from swap if sequential,
1274 	 * otherwise just read one page.  Contiguous pages from swap must
1275 	 * reside within a single device stripe because the I/O cannot be
1276 	 * broken up across multiple stripes.
1277 	 *
1278 	 * Note that blk and iblk can be SWAPBLK_NONE but the loop is
1279 	 * set up such that the case(s) are handled implicitly.
1280 	 */
1281 	crit_enter();
1282 	blk = swp_pager_meta_ctl(mreq->object, mreq->pindex, 0);
1283 	marray[0] = mreq;
1284 
1285 	for (i = 1; swap_burst_read &&
1286 		    i < XIO_INTERNAL_PAGES &&
1287 		    mreq->pindex + i < object->size; ++i) {
1288 		daddr_t iblk;
1289 
1290 		iblk = swp_pager_meta_ctl(object, mreq->pindex + i, 0);
1291 		if (iblk != blk + i)
1292 			break;
1293 		if ((blk ^ iblk) & dmmax_mask)
1294 			break;
1295 		m = vm_page_lookup(object, mreq->pindex + i);
1296 		if (m == NULL) {
1297 			m = vm_page_alloc(object, mreq->pindex + i,
1298 					  VM_ALLOC_QUICK);
1299 			if (m == NULL)
1300 				break;
1301 		} else {
1302 			if ((m->flags & PG_BUSY) || m->busy || m->valid)
1303 				break;
1304 			vm_page_unqueue_nowakeup(m);
1305 			vm_page_busy(m);
1306 		}
1307 		marray[i] = m;
1308 	}
1309 	if (i > 1)
1310 		vm_page_flag_set(marray[i - 1], PG_RAM);
1311 
1312 	crit_exit();
1313 
1314 	/*
1315 	 * If mreq is the requested page and we have nothing to do return
1316 	 * VM_PAGER_FAIL.  If raonly is set mreq is just another read-ahead
1317 	 * page and must be cleaned up.
1318 	 */
1319 	if (blk == SWAPBLK_NONE) {
1320 		KKASSERT(i == 1);
1321 		if (raonly) {
1322 			vnode_pager_freepage(mreq);
1323 			return(VM_PAGER_OK);
1324 		} else {
1325 			return(VM_PAGER_FAIL);
1326 		}
1327 	}
1328 
1329 	/*
1330 	 * map our page(s) into kva for input
1331 	 */
1332 	bp = getpbuf(&nsw_rcount);
1333 	bio = &bp->b_bio1;
1334 	kva = (vm_offset_t) bp->b_kvabase;
1335 	bcopy(marray, bp->b_xio.xio_pages, i * sizeof(vm_page_t));
1336 	pmap_qenter(kva, bp->b_xio.xio_pages, i);
1337 
1338 	bp->b_data = (caddr_t)kva;
1339 	bp->b_bcount = PAGE_SIZE * i;
1340 	bp->b_xio.xio_npages = i;
1341 	bio->bio_done = swp_pager_async_iodone;
1342 	bio->bio_offset = (off_t)blk << PAGE_SHIFT;
1343 	bio->bio_caller_info1.index = SWBIO_READ;
1344 
1345 	/*
1346 	 * Set index.  If raonly set the index beyond the array so all
1347 	 * the pages are treated the same, otherwise the original mreq is
1348 	 * at index 0.
1349 	 */
1350 	if (raonly)
1351 		bio->bio_driver_info = (void *)(intptr_t)i;
1352 	else
1353 		bio->bio_driver_info = (void *)(intptr_t)0;
1354 
1355 	for (j = 0; j < i; ++j)
1356 		vm_page_flag_set(bp->b_xio.xio_pages[j], PG_SWAPINPROG);
1357 
1358 	mycpu->gd_cnt.v_swapin++;
1359 	mycpu->gd_cnt.v_swappgsin += bp->b_xio.xio_npages;
1360 
1361 	/*
1362 	 * We still hold the lock on mreq, and our automatic completion routine
1363 	 * does not remove it.
1364 	 */
1365 	vm_object_pip_add(object, bp->b_xio.xio_npages);
1366 
1367 	/*
1368 	 * perform the I/O.  NOTE!!!  bp cannot be considered valid after
1369 	 * this point because we automatically release it on completion.
1370 	 * Instead, we look at the one page we are interested in which we
1371 	 * still hold a lock on even through the I/O completion.
1372 	 *
1373 	 * The other pages in our m[] array are also released on completion,
1374 	 * so we cannot assume they are valid anymore either.
1375 	 */
1376 	bp->b_cmd = BUF_CMD_READ;
1377 	BUF_KERNPROC(bp);
1378 	vn_strategy(swapdev_vp, bio);
1379 
1380 	/*
1381 	 * Wait for the page we want to complete.  PG_SWAPINPROG is always
1382 	 * cleared on completion.  If an I/O error occurs, SWAPBLK_NONE
1383 	 * is set in the meta-data.
1384 	 *
1385 	 * If this is a read-ahead only we return immediately without
1386 	 * waiting for I/O.
1387 	 */
1388 	if (raonly)
1389 		return(VM_PAGER_OK);
1390 
1391 	/*
1392 	 * Read-ahead includes originally requested page case.
1393 	 */
1394 	crit_enter();
1395 	while ((mreq->flags & PG_SWAPINPROG) != 0) {
1396 		vm_page_flag_set(mreq, PG_WANTED | PG_REFERENCED);
1397 		mycpu->gd_cnt.v_intrans++;
1398 		if (tsleep(mreq, 0, "swread", hz*20)) {
1399 			kprintf(
1400 			    "swap_pager: indefinite wait buffer: "
1401 				" offset: %lld, size: %ld\n",
1402 			    (long long)bio->bio_offset,
1403 			    (long)bp->b_bcount
1404 			);
1405 		}
1406 	}
1407 	crit_exit();
1408 
1409 	/*
1410 	 * mreq is left bussied after completion, but all the other pages
1411 	 * are freed.  If we had an unrecoverable read error the page will
1412 	 * not be valid.
1413 	 */
1414 	if (mreq->valid != VM_PAGE_BITS_ALL)
1415 		return(VM_PAGER_ERROR);
1416 	else
1417 		return(VM_PAGER_OK);
1418 
1419 	/*
1420 	 * A final note: in a low swap situation, we cannot deallocate swap
1421 	 * and mark a page dirty here because the caller is likely to mark
1422 	 * the page clean when we return, causing the page to possibly revert
1423 	 * to all-zero's later.
1424 	 */
1425 }
1426 
1427 /*
1428  *	swap_pager_putpages:
1429  *
1430  *	Assign swap (if necessary) and initiate I/O on the specified pages.
1431  *
1432  *	We support both OBJT_DEFAULT and OBJT_SWAP objects.  DEFAULT objects
1433  *	are automatically converted to SWAP objects.
1434  *
1435  *	In a low memory situation we may block in vn_strategy(), but the new
1436  *	vm_page reservation system coupled with properly written VFS devices
1437  *	should ensure that no low-memory deadlock occurs.  This is an area
1438  *	which needs work.
1439  *
1440  *	The parent has N vm_object_pip_add() references prior to
1441  *	calling us and will remove references for rtvals[] that are
1442  *	not set to VM_PAGER_PEND.  We need to remove the rest on I/O
1443  *	completion.
1444  *
1445  *	The parent has soft-busy'd the pages it passes us and will unbusy
1446  *	those whos rtvals[] entry is not set to VM_PAGER_PEND on return.
1447  *	We need to unbusy the rest on I/O completion.
1448  */
1449 void
1450 swap_pager_putpages(vm_object_t object, vm_page_t *m, int count,
1451 		    boolean_t sync, int *rtvals)
1452 {
1453 	int i;
1454 	int n = 0;
1455 
1456 	if (count && m[0]->object != object) {
1457 		panic("swap_pager_getpages: object mismatch %p/%p",
1458 		    object,
1459 		    m[0]->object
1460 		);
1461 	}
1462 
1463 	/*
1464 	 * Step 1
1465 	 *
1466 	 * Turn object into OBJT_SWAP
1467 	 * check for bogus sysops
1468 	 * force sync if not pageout process
1469 	 */
1470 	if (object->type == OBJT_DEFAULT)
1471 		swp_pager_meta_convert(object);
1472 
1473 	if (curthread != pagethread)
1474 		sync = TRUE;
1475 
1476 	/*
1477 	 * Step 2
1478 	 *
1479 	 * Update nsw parameters from swap_async_max sysctl values.
1480 	 * Do not let the sysop crash the machine with bogus numbers.
1481 	 */
1482 
1483 	if (swap_async_max != nsw_wcount_async_max) {
1484 		int n;
1485 
1486 		/*
1487 		 * limit range
1488 		 */
1489 		if ((n = swap_async_max) > nswbuf / 2)
1490 			n = nswbuf / 2;
1491 		if (n < 1)
1492 			n = 1;
1493 		swap_async_max = n;
1494 
1495 		/*
1496 		 * Adjust difference ( if possible ).  If the current async
1497 		 * count is too low, we may not be able to make the adjustment
1498 		 * at this time.
1499 		 */
1500 		crit_enter();
1501 		n -= nsw_wcount_async_max;
1502 		if (nsw_wcount_async + n >= 0) {
1503 			nsw_wcount_async += n;
1504 			nsw_wcount_async_max += n;
1505 			wakeup(&nsw_wcount_async);
1506 		}
1507 		crit_exit();
1508 	}
1509 
1510 	/*
1511 	 * Step 3
1512 	 *
1513 	 * Assign swap blocks and issue I/O.  We reallocate swap on the fly.
1514 	 * The page is left dirty until the pageout operation completes
1515 	 * successfully.
1516 	 */
1517 
1518 	for (i = 0; i < count; i += n) {
1519 		struct buf *bp;
1520 		struct bio *bio;
1521 		daddr_t blk;
1522 		int j;
1523 
1524 		/*
1525 		 * Maximum I/O size is limited by a number of factors.
1526 		 */
1527 
1528 		n = min(BLIST_MAX_ALLOC, count - i);
1529 		n = min(n, nsw_cluster_max);
1530 
1531 		crit_enter();
1532 
1533 		/*
1534 		 * Get biggest block of swap we can.  If we fail, fall
1535 		 * back and try to allocate a smaller block.  Don't go
1536 		 * overboard trying to allocate space if it would overly
1537 		 * fragment swap.
1538 		 */
1539 		while (
1540 		    (blk = swp_pager_getswapspace(object, n)) == SWAPBLK_NONE &&
1541 		    n > 4
1542 		) {
1543 			n >>= 1;
1544 		}
1545 		if (blk == SWAPBLK_NONE) {
1546 			for (j = 0; j < n; ++j)
1547 				rtvals[i+j] = VM_PAGER_FAIL;
1548 			crit_exit();
1549 			continue;
1550 		}
1551 
1552 		/*
1553 		 * The I/O we are constructing cannot cross a physical
1554 		 * disk boundry in the swap stripe.  Note: we are still
1555 		 * at splvm().
1556 		 */
1557 		if ((blk ^ (blk + n)) & dmmax_mask) {
1558 			j = ((blk + dmmax) & dmmax_mask) - blk;
1559 			swp_pager_freeswapspace(object, blk + j, n - j);
1560 			n = j;
1561 		}
1562 
1563 		/*
1564 		 * All I/O parameters have been satisfied, build the I/O
1565 		 * request and assign the swap space.
1566 		 */
1567 		if (sync == TRUE)
1568 			bp = getpbuf(&nsw_wcount_sync);
1569 		else
1570 			bp = getpbuf(&nsw_wcount_async);
1571 		bio = &bp->b_bio1;
1572 
1573 		pmap_qenter((vm_offset_t)bp->b_data, &m[i], n);
1574 
1575 		bp->b_bcount = PAGE_SIZE * n;
1576 		bio->bio_offset = (off_t)blk << PAGE_SHIFT;
1577 
1578 		for (j = 0; j < n; ++j) {
1579 			vm_page_t mreq = m[i+j];
1580 
1581 			swp_pager_meta_build(mreq->object, mreq->pindex,
1582 					     blk + j);
1583 			if (object->type == OBJT_SWAP)
1584 				vm_page_dirty(mreq);
1585 			rtvals[i+j] = VM_PAGER_OK;
1586 
1587 			vm_page_flag_set(mreq, PG_SWAPINPROG);
1588 			bp->b_xio.xio_pages[j] = mreq;
1589 		}
1590 		bp->b_xio.xio_npages = n;
1591 
1592 		mycpu->gd_cnt.v_swapout++;
1593 		mycpu->gd_cnt.v_swappgsout += bp->b_xio.xio_npages;
1594 
1595 		crit_exit();
1596 
1597 		bp->b_dirtyoff = 0;		/* req'd for NFS */
1598 		bp->b_dirtyend = bp->b_bcount;	/* req'd for NFS */
1599 		bp->b_cmd = BUF_CMD_WRITE;
1600 		bio->bio_caller_info1.index = SWBIO_WRITE;
1601 
1602 		/*
1603 		 * asynchronous
1604 		 */
1605 		if (sync == FALSE) {
1606 			bio->bio_done = swp_pager_async_iodone;
1607 			BUF_KERNPROC(bp);
1608 			vn_strategy(swapdev_vp, bio);
1609 
1610 			for (j = 0; j < n; ++j)
1611 				rtvals[i+j] = VM_PAGER_PEND;
1612 			continue;
1613 		}
1614 
1615 		/*
1616 		 * Issue synchrnously.
1617 		 *
1618 		 * Wait for the sync I/O to complete, then update rtvals.
1619 		 * We just set the rtvals[] to VM_PAGER_PEND so we can call
1620 		 * our async completion routine at the end, thus avoiding a
1621 		 * double-free.
1622 		 */
1623 		bio->bio_caller_info1.index |= SWBIO_SYNC;
1624 		bio->bio_done = biodone_sync;
1625 		bio->bio_flags |= BIO_SYNC;
1626 		vn_strategy(swapdev_vp, bio);
1627 		biowait(bio, "swwrt");
1628 
1629 		for (j = 0; j < n; ++j)
1630 			rtvals[i+j] = VM_PAGER_PEND;
1631 
1632 		/*
1633 		 * Now that we are through with the bp, we can call the
1634 		 * normal async completion, which frees everything up.
1635 		 */
1636 		swp_pager_async_iodone(bio);
1637 	}
1638 }
1639 
1640 void
1641 swap_pager_newswap(void)
1642 {
1643 	swp_sizecheck();
1644 }
1645 
1646 /*
1647  *	swp_pager_async_iodone:
1648  *
1649  *	Completion routine for asynchronous reads and writes from/to swap.
1650  *	Also called manually by synchronous code to finish up a bp.
1651  *
1652  *	For READ operations, the pages are PG_BUSY'd.  For WRITE operations,
1653  *	the pages are vm_page_t->busy'd.  For READ operations, we PG_BUSY
1654  *	unbusy all pages except the 'main' request page.  For WRITE
1655  *	operations, we vm_page_t->busy'd unbusy all pages ( we can do this
1656  *	because we marked them all VM_PAGER_PEND on return from putpages ).
1657  *
1658  *	This routine may not block.
1659  */
1660 static void
1661 swp_pager_async_iodone(struct bio *bio)
1662 {
1663 	struct buf *bp = bio->bio_buf;
1664 	vm_object_t object = NULL;
1665 	int i;
1666 	int *nswptr;
1667 
1668 	/*
1669 	 * report error
1670 	 */
1671 	if (bp->b_flags & B_ERROR) {
1672 		kprintf(
1673 		    "swap_pager: I/O error - %s failed; offset %lld,"
1674 			"size %ld, error %d\n",
1675 		    ((bio->bio_caller_info1.index & SWBIO_READ) ?
1676 			"pagein" : "pageout"),
1677 		    (long long)bio->bio_offset,
1678 		    (long)bp->b_bcount,
1679 		    bp->b_error
1680 		);
1681 	}
1682 
1683 	/*
1684 	 * set object, raise to splvm().
1685 	 */
1686 	if (bp->b_xio.xio_npages)
1687 		object = bp->b_xio.xio_pages[0]->object;
1688 	crit_enter();
1689 
1690 	/*
1691 	 * remove the mapping for kernel virtual
1692 	 */
1693 	pmap_qremove((vm_offset_t)bp->b_data, bp->b_xio.xio_npages);
1694 
1695 	/*
1696 	 * cleanup pages.  If an error occurs writing to swap, we are in
1697 	 * very serious trouble.  If it happens to be a disk error, though,
1698 	 * we may be able to recover by reassigning the swap later on.  So
1699 	 * in this case we remove the m->swapblk assignment for the page
1700 	 * but do not free it in the rlist.  The errornous block(s) are thus
1701 	 * never reallocated as swap.  Redirty the page and continue.
1702 	 */
1703 	for (i = 0; i < bp->b_xio.xio_npages; ++i) {
1704 		vm_page_t m = bp->b_xio.xio_pages[i];
1705 
1706 		if (bp->b_flags & B_ERROR) {
1707 			/*
1708 			 * If an error occurs I'd love to throw the swapblk
1709 			 * away without freeing it back to swapspace, so it
1710 			 * can never be used again.  But I can't from an
1711 			 * interrupt.
1712 			 */
1713 
1714 			if (bio->bio_caller_info1.index & SWBIO_READ) {
1715 				/*
1716 				 * When reading, reqpage needs to stay
1717 				 * locked for the parent, but all other
1718 				 * pages can be freed.  We still want to
1719 				 * wakeup the parent waiting on the page,
1720 				 * though.  ( also: pg_reqpage can be -1 and
1721 				 * not match anything ).
1722 				 *
1723 				 * We have to wake specifically requested pages
1724 				 * up too because we cleared PG_SWAPINPROG and
1725 				 * someone may be waiting for that.
1726 				 *
1727 				 * NOTE: for reads, m->dirty will probably
1728 				 * be overridden by the original caller of
1729 				 * getpages so don't play cute tricks here.
1730 				 *
1731 				 * NOTE: We can't actually free the page from
1732 				 * here, because this is an interrupt.  It
1733 				 * is not legal to mess with object->memq
1734 				 * from an interrupt.  Deactivate the page
1735 				 * instead.
1736 				 */
1737 
1738 				m->valid = 0;
1739 				vm_page_flag_clear(m, PG_ZERO);
1740 				vm_page_flag_clear(m, PG_SWAPINPROG);
1741 
1742 				/*
1743 				 * bio_driver_info holds the requested page
1744 				 * index.
1745 				 */
1746 				if (i != (int)(intptr_t)bio->bio_driver_info) {
1747 					vm_page_deactivate(m);
1748 					vm_page_wakeup(m);
1749 				} else {
1750 					vm_page_flash(m);
1751 				}
1752 				/*
1753 				 * If i == bp->b_pager.pg_reqpage, do not wake
1754 				 * the page up.  The caller needs to.
1755 				 */
1756 			} else {
1757 				/*
1758 				 * If a write error occurs remove the swap
1759 				 * assignment (note that PG_SWAPPED may or
1760 				 * may not be set depending on prior activity).
1761 				 *
1762 				 * Re-dirty OBJT_SWAP pages as there is no
1763 				 * other backing store, we can't throw the
1764 				 * page away.
1765 				 *
1766 				 * Non-OBJT_SWAP pages (aka swapcache) must
1767 				 * not be dirtied since they may not have
1768 				 * been dirty in the first place, and they
1769 				 * do have backing store (the vnode).
1770 				 */
1771 				swp_pager_meta_ctl(m->object, m->pindex,
1772 						   SWM_FREE);
1773 				vm_page_flag_clear(m, PG_SWAPPED);
1774 				if (m->object->type == OBJT_SWAP) {
1775 					vm_page_dirty(m);
1776 					vm_page_activate(m);
1777 				}
1778 				vm_page_flag_clear(m, PG_SWAPINPROG);
1779 				vm_page_io_finish(m);
1780 			}
1781 		} else if (bio->bio_caller_info1.index & SWBIO_READ) {
1782 			/*
1783 			 * NOTE: for reads, m->dirty will probably be
1784 			 * overridden by the original caller of getpages so
1785 			 * we cannot set them in order to free the underlying
1786 			 * swap in a low-swap situation.  I don't think we'd
1787 			 * want to do that anyway, but it was an optimization
1788 			 * that existed in the old swapper for a time before
1789 			 * it got ripped out due to precisely this problem.
1790 			 *
1791 			 * clear PG_ZERO in page.
1792 			 *
1793 			 * If not the requested page then deactivate it.
1794 			 *
1795 			 * Note that the requested page, reqpage, is left
1796 			 * busied, but we still have to wake it up.  The
1797 			 * other pages are released (unbusied) by
1798 			 * vm_page_wakeup().  We do not set reqpage's
1799 			 * valid bits here, it is up to the caller.
1800 			 */
1801 
1802 			/*
1803 			 * NOTE: can't call pmap_clear_modify(m) from an
1804 			 * interrupt thread, the pmap code may have to map
1805 			 * non-kernel pmaps and currently asserts the case.
1806 			 */
1807 			/*pmap_clear_modify(m);*/
1808 			m->valid = VM_PAGE_BITS_ALL;
1809 			vm_page_undirty(m);
1810 			vm_page_flag_clear(m, PG_ZERO | PG_SWAPINPROG);
1811 			vm_page_flag_set(m, PG_SWAPPED);
1812 
1813 			/*
1814 			 * We have to wake specifically requested pages
1815 			 * up too because we cleared PG_SWAPINPROG and
1816 			 * could be waiting for it in getpages.  However,
1817 			 * be sure to not unbusy getpages specifically
1818 			 * requested page - getpages expects it to be
1819 			 * left busy.
1820 			 *
1821 			 * bio_driver_info holds the requested page
1822 			 */
1823 			if (i != (int)(intptr_t)bio->bio_driver_info) {
1824 				vm_page_deactivate(m);
1825 				vm_page_wakeup(m);
1826 			} else {
1827 				vm_page_flash(m);
1828 			}
1829 		} else {
1830 			/*
1831 			 * Mark the page clean but do not mess with the
1832 			 * pmap-layer's modified state.  That state should
1833 			 * also be clear since the caller protected the
1834 			 * page VM_PROT_READ, but allow the case.
1835 			 *
1836 			 * We are in an interrupt, avoid pmap operations.
1837 			 *
1838 			 * If we have a severe page deficit, deactivate the
1839 			 * page.  Do not try to cache it (which would also
1840 			 * involve a pmap op), because the page might still
1841 			 * be read-heavy.
1842 			 *
1843 			 * When using the swap to cache clean vnode pages
1844 			 * we do not mess with the page dirty bits.
1845 			 */
1846 			if (m->object->type == OBJT_SWAP)
1847 				vm_page_undirty(m);
1848 			vm_page_flag_clear(m, PG_SWAPINPROG);
1849 			vm_page_flag_set(m, PG_SWAPPED);
1850 			vm_page_io_finish(m);
1851 			if (vm_page_count_severe())
1852 				vm_page_deactivate(m);
1853 #if 0
1854 			if (!vm_page_count_severe() || !vm_page_try_to_cache(m))
1855 				vm_page_protect(m, VM_PROT_READ);
1856 #endif
1857 		}
1858 	}
1859 
1860 	/*
1861 	 * adjust pip.  NOTE: the original parent may still have its own
1862 	 * pip refs on the object.
1863 	 */
1864 
1865 	if (object)
1866 		vm_object_pip_wakeupn(object, bp->b_xio.xio_npages);
1867 
1868 	/*
1869 	 * Release the physical I/O buffer.
1870 	 *
1871 	 * NOTE: Due to synchronous operations in the write case b_cmd may
1872 	 *	 already be set to BUF_CMD_DONE and BIO_SYNC may have already
1873 	 *	 been cleared.
1874 	 */
1875 	if (bio->bio_caller_info1.index & SWBIO_READ)
1876 		nswptr = &nsw_rcount;
1877 	else if (bio->bio_caller_info1.index & SWBIO_SYNC)
1878 		nswptr = &nsw_wcount_sync;
1879 	else
1880 		nswptr = &nsw_wcount_async;
1881 	bp->b_cmd = BUF_CMD_DONE;
1882 	relpbuf(bp, nswptr);
1883 	crit_exit();
1884 }
1885 
1886 /************************************************************************
1887  *				SWAP META DATA 				*
1888  ************************************************************************
1889  *
1890  *	These routines manipulate the swap metadata stored in the
1891  *	OBJT_SWAP object.  All swp_*() routines must be called at
1892  *	splvm() because swap can be freed up by the low level vm_page
1893  *	code which might be called from interrupts beyond what splbio() covers.
1894  *
1895  *	Swap metadata is implemented with a global hash and not directly
1896  *	linked into the object.  Instead the object simply contains
1897  *	appropriate tracking counters.
1898  */
1899 
1900 /*
1901  * Lookup the swblock containing the specified swap block index.
1902  */
1903 static __inline
1904 struct swblock *
1905 swp_pager_lookup(vm_object_t object, vm_pindex_t index)
1906 {
1907 	index &= ~SWAP_META_MASK;
1908 	return (RB_LOOKUP(swblock_rb_tree, &object->swblock_root, index));
1909 }
1910 
1911 /*
1912  * Remove a swblock from the RB tree.
1913  */
1914 static __inline
1915 void
1916 swp_pager_remove(vm_object_t object, struct swblock *swap)
1917 {
1918 	RB_REMOVE(swblock_rb_tree, &object->swblock_root, swap);
1919 }
1920 
1921 /*
1922  * Convert default object to swap object if necessary
1923  */
1924 static void
1925 swp_pager_meta_convert(vm_object_t object)
1926 {
1927 	if (object->type == OBJT_DEFAULT) {
1928 		object->type = OBJT_SWAP;
1929 		KKASSERT(object->swblock_count == 0);
1930 	}
1931 }
1932 
1933 /*
1934  * SWP_PAGER_META_BUILD() -	add swap block to swap meta data for object
1935  *
1936  *	We first convert the object to a swap object if it is a default
1937  *	object.  Vnode objects do not need to be converted.
1938  *
1939  *	The specified swapblk is added to the object's swap metadata.  If
1940  *	the swapblk is not valid, it is freed instead.  Any previously
1941  *	assigned swapblk is freed.
1942  */
1943 static void
1944 swp_pager_meta_build(vm_object_t object, vm_pindex_t index, daddr_t swapblk)
1945 {
1946 	struct swblock *swap;
1947 	struct swblock *oswap;
1948 
1949 	KKASSERT(swapblk != SWAPBLK_NONE);
1950 
1951 	/*
1952 	 * Convert object if necessary
1953 	 */
1954 	if (object->type == OBJT_DEFAULT)
1955 		swp_pager_meta_convert(object);
1956 
1957 	/*
1958 	 * Locate swblock.  If not found create, but if we aren't adding
1959 	 * anything just return.  If we run out of space in the map we wait
1960 	 * and, since the hash table may have changed, retry.
1961 	 */
1962 retry:
1963 	swap = swp_pager_lookup(object, index);
1964 
1965 	if (swap == NULL) {
1966 		int i;
1967 
1968 		swap = zalloc(swap_zone);
1969 		if (swap == NULL) {
1970 			vm_wait(0);
1971 			goto retry;
1972 		}
1973 		swap->swb_index = index & ~SWAP_META_MASK;
1974 		swap->swb_count = 0;
1975 
1976 		++object->swblock_count;
1977 
1978 		for (i = 0; i < SWAP_META_PAGES; ++i)
1979 			swap->swb_pages[i] = SWAPBLK_NONE;
1980 		oswap = RB_INSERT(swblock_rb_tree, &object->swblock_root, swap);
1981 		KKASSERT(oswap == NULL);
1982 	}
1983 
1984 	/*
1985 	 * Delete prior contents of metadata
1986 	 */
1987 
1988 	index &= SWAP_META_MASK;
1989 
1990 	if (swap->swb_pages[index] != SWAPBLK_NONE) {
1991 		swp_pager_freeswapspace(object, swap->swb_pages[index], 1);
1992 		--swap->swb_count;
1993 	}
1994 
1995 	/*
1996 	 * Enter block into metadata
1997 	 */
1998 	swap->swb_pages[index] = swapblk;
1999 	if (swapblk != SWAPBLK_NONE)
2000 		++swap->swb_count;
2001 }
2002 
2003 /*
2004  * SWP_PAGER_META_FREE() - free a range of blocks in the object's swap metadata
2005  *
2006  *	The requested range of blocks is freed, with any associated swap
2007  *	returned to the swap bitmap.
2008  *
2009  *	This routine will free swap metadata structures as they are cleaned
2010  *	out.  This routine does *NOT* operate on swap metadata associated
2011  *	with resident pages.
2012  *
2013  *	This routine must be called at splvm()
2014  */
2015 static int swp_pager_meta_free_callback(struct swblock *swb, void *data);
2016 
2017 static void
2018 swp_pager_meta_free(vm_object_t object, vm_pindex_t index, vm_pindex_t count)
2019 {
2020 	struct swfreeinfo info;
2021 
2022 	/*
2023 	 * Nothing to do
2024 	 */
2025 	if (object->swblock_count == 0) {
2026 		KKASSERT(RB_EMPTY(&object->swblock_root));
2027 		return;
2028 	}
2029 	if (count == 0)
2030 		return;
2031 
2032 	/*
2033 	 * Setup for RB tree scan.  Note that the pindex range can be huge
2034 	 * due to the 64 bit page index space so we cannot safely iterate.
2035 	 */
2036 	info.object = object;
2037 	info.basei = index & ~SWAP_META_MASK;
2038 	info.begi = index;
2039 	info.endi = index + count - 1;
2040 	swblock_rb_tree_RB_SCAN(&object->swblock_root, rb_swblock_scancmp,
2041 				swp_pager_meta_free_callback, &info);
2042 }
2043 
2044 static
2045 int
2046 swp_pager_meta_free_callback(struct swblock *swap, void *data)
2047 {
2048 	struct swfreeinfo *info = data;
2049 	vm_object_t object = info->object;
2050 	int index;
2051 	int eindex;
2052 
2053 	/*
2054 	 * Figure out the range within the swblock.  The wider scan may
2055 	 * return edge-case swap blocks when the start and/or end points
2056 	 * are in the middle of a block.
2057 	 */
2058 	if (swap->swb_index < info->begi)
2059 		index = (int)info->begi & SWAP_META_MASK;
2060 	else
2061 		index = 0;
2062 
2063 	if (swap->swb_index + SWAP_META_PAGES > info->endi)
2064 		eindex = (int)info->endi & SWAP_META_MASK;
2065 	else
2066 		eindex = SWAP_META_MASK;
2067 
2068 	/*
2069 	 * Scan and free the blocks.  The loop terminates early
2070 	 * if (swap) runs out of blocks and could be freed.
2071 	 */
2072 	while (index <= eindex) {
2073 		daddr_t v = swap->swb_pages[index];
2074 
2075 		if (v != SWAPBLK_NONE) {
2076 			swp_pager_freeswapspace(object, v, 1);
2077 			swap->swb_pages[index] = SWAPBLK_NONE;
2078 			if (--swap->swb_count == 0) {
2079 				swp_pager_remove(object, swap);
2080 				zfree(swap_zone, swap);
2081 				--object->swblock_count;
2082 				break;
2083 			}
2084 		}
2085 		++index;
2086 	}
2087 	/* swap may be invalid here due to zfree above */
2088 	return(0);
2089 }
2090 
2091 /*
2092  * SWP_PAGER_META_FREE_ALL() - destroy all swap metadata associated with object
2093  *
2094  *	This routine locates and destroys all swap metadata associated with
2095  *	an object.
2096  *
2097  *	This routine must be called at splvm()
2098  */
2099 static void
2100 swp_pager_meta_free_all(vm_object_t object)
2101 {
2102 	struct swblock *swap;
2103 	int i;
2104 
2105 	while ((swap = RB_ROOT(&object->swblock_root)) != NULL) {
2106 		swp_pager_remove(object, swap);
2107 		for (i = 0; i < SWAP_META_PAGES; ++i) {
2108 			daddr_t v = swap->swb_pages[i];
2109 			if (v != SWAPBLK_NONE) {
2110 				--swap->swb_count;
2111 				swp_pager_freeswapspace(object, v, 1);
2112 			}
2113 		}
2114 		if (swap->swb_count != 0)
2115 			panic("swap_pager_meta_free_all: swb_count != 0");
2116 		zfree(swap_zone, swap);
2117 		--object->swblock_count;
2118 	}
2119 	KKASSERT(object->swblock_count == 0);
2120 }
2121 
2122 /*
2123  * SWP_PAGER_METACTL() -  misc control of swap and vm_page_t meta data.
2124  *
2125  *	This routine is capable of looking up, popping, or freeing
2126  *	swapblk assignments in the swap meta data or in the vm_page_t.
2127  *	The routine typically returns the swapblk being looked-up, or popped,
2128  *	or SWAPBLK_NONE if the block was freed, or SWAPBLK_NONE if the block
2129  *	was invalid.  This routine will automatically free any invalid
2130  *	meta-data swapblks.
2131  *
2132  *	It is not possible to store invalid swapblks in the swap meta data
2133  *	(other then a literal 'SWAPBLK_NONE'), so we don't bother checking.
2134  *
2135  *	When acting on a busy resident page and paging is in progress, we
2136  *	have to wait until paging is complete but otherwise can act on the
2137  *	busy page.
2138  *
2139  *	This routine must be called at splvm().
2140  *
2141  *	SWM_FREE	remove and free swap block from metadata
2142  *	SWM_POP		remove from meta data but do not free.. pop it out
2143  */
2144 static daddr_t
2145 swp_pager_meta_ctl(vm_object_t object, vm_pindex_t index, int flags)
2146 {
2147 	struct swblock *swap;
2148 	daddr_t r1;
2149 
2150 	if (object->swblock_count == 0)
2151 		return(SWAPBLK_NONE);
2152 
2153 	r1 = SWAPBLK_NONE;
2154 	swap = swp_pager_lookup(object, index);
2155 
2156 	if (swap != NULL) {
2157 		index &= SWAP_META_MASK;
2158 		r1 = swap->swb_pages[index];
2159 
2160 		if (r1 != SWAPBLK_NONE) {
2161 			if (flags & SWM_FREE) {
2162 				swp_pager_freeswapspace(object, r1, 1);
2163 				r1 = SWAPBLK_NONE;
2164 			}
2165 			if (flags & (SWM_FREE|SWM_POP)) {
2166 				swap->swb_pages[index] = SWAPBLK_NONE;
2167 				if (--swap->swb_count == 0) {
2168 					swp_pager_remove(object, swap);
2169 					zfree(swap_zone, swap);
2170 					--object->swblock_count;
2171 				}
2172 			}
2173 		}
2174 	}
2175 	return(r1);
2176 }
2177