xref: /dflybsd-src/sys/vm/swap_pager.c (revision abf21138fc9d6b9536faf556858abf264f704f26)
1 /*
2  * (MPSAFE)
3  *
4  * Copyright (c) 1998-2010 The DragonFly Project.  All rights reserved.
5  *
6  * This code is derived from software contributed to The DragonFly Project
7  * by Matthew Dillon <dillon@backplane.com>
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  *
13  * 1. Redistributions of source code must retain the above copyright
14  *    notice, this list of conditions and the following disclaimer.
15  * 2. Redistributions in binary form must reproduce the above copyright
16  *    notice, this list of conditions and the following disclaimer in
17  *    the documentation and/or other materials provided with the
18  *    distribution.
19  * 3. Neither the name of The DragonFly Project nor the names of its
20  *    contributors may be used to endorse or promote products derived
21  *    from this software without specific, prior written permission.
22  *
23  * THIS SOFTWARE IS PROVIDED BY THE COPYRIGHT HOLDERS AND CONTRIBUTORS
24  * ``AS IS'' AND ANY EXPRESS OR IMPLIED WARRANTIES, INCLUDING, BUT NOT
25  * LIMITED TO, THE IMPLIED WARRANTIES OF MERCHANTABILITY AND FITNESS
26  * FOR A PARTICULAR PURPOSE ARE DISCLAIMED.  IN NO EVENT SHALL THE
27  * COPYRIGHT HOLDERS OR CONTRIBUTORS BE LIABLE FOR ANY DIRECT, INDIRECT,
28  * INCIDENTAL, SPECIAL, EXEMPLARY OR CONSEQUENTIAL DAMAGES (INCLUDING,
29  * BUT NOT LIMITED TO, PROCUREMENT OF SUBSTITUTE GOODS OR SERVICES;
30  * LOSS OF USE, DATA, OR PROFITS; OR BUSINESS INTERRUPTION) HOWEVER CAUSED
31  * AND ON ANY THEORY OF LIABILITY, WHETHER IN CONTRACT, STRICT LIABILITY,
32  * OR TORT (INCLUDING NEGLIGENCE OR OTHERWISE) ARISING IN ANY WAY OUT
33  * OF THE USE OF THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF
34  * SUCH DAMAGE.
35  *
36  * Copyright (c) 1994 John S. Dyson
37  * Copyright (c) 1990 University of Utah.
38  * Copyright (c) 1991, 1993
39  *	The Regents of the University of California.  All rights reserved.
40  *
41  * This code is derived from software contributed to Berkeley by
42  * the Systems Programming Group of the University of Utah Computer
43  * Science Department.
44  *
45  * Redistribution and use in source and binary forms, with or without
46  * modification, are permitted provided that the following conditions
47  * are met:
48  * 1. Redistributions of source code must retain the above copyright
49  *    notice, this list of conditions and the following disclaimer.
50  * 2. Redistributions in binary form must reproduce the above copyright
51  *    notice, this list of conditions and the following disclaimer in the
52  *    documentation and/or other materials provided with the distribution.
53  * 3. All advertising materials mentioning features or use of this software
54  *    must display the following acknowledgement:
55  *	This product includes software developed by the University of
56  *	California, Berkeley and its contributors.
57  * 4. Neither the name of the University nor the names of its contributors
58  *    may be used to endorse or promote products derived from this software
59  *    without specific prior written permission.
60  *
61  * THIS SOFTWARE IS PROVIDED BY THE REGENTS AND CONTRIBUTORS ``AS IS'' AND
62  * ANY EXPRESS OR IMPLIED WARRANTIES, INCLUDING, BUT NOT LIMITED TO, THE
63  * IMPLIED WARRANTIES OF MERCHANTABILITY AND FITNESS FOR A PARTICULAR PURPOSE
64  * ARE DISCLAIMED.  IN NO EVENT SHALL THE REGENTS OR CONTRIBUTORS BE LIABLE
65  * FOR ANY DIRECT, INDIRECT, INCIDENTAL, SPECIAL, EXEMPLARY, OR CONSEQUENTIAL
66  * DAMAGES (INCLUDING, BUT NOT LIMITED TO, PROCUREMENT OF SUBSTITUTE GOODS
67  * OR SERVICES; LOSS OF USE, DATA, OR PROFITS; OR BUSINESS INTERRUPTION)
68  * HOWEVER CAUSED AND ON ANY THEORY OF LIABILITY, WHETHER IN CONTRACT, STRICT
69  * LIABILITY, OR TORT (INCLUDING NEGLIGENCE OR OTHERWISE) ARISING IN ANY WAY
70  * OUT OF THE USE OF THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF
71  * SUCH DAMAGE.
72  *
73  *				New Swap System
74  *				Matthew Dillon
75  *
76  * Radix Bitmap 'blists'.
77  *
78  *	- The new swapper uses the new radix bitmap code.  This should scale
79  *	  to arbitrarily small or arbitrarily large swap spaces and an almost
80  *	  arbitrary degree of fragmentation.
81  *
82  * Features:
83  *
84  *	- on the fly reallocation of swap during putpages.  The new system
85  *	  does not try to keep previously allocated swap blocks for dirty
86  *	  pages.
87  *
88  *	- on the fly deallocation of swap
89  *
90  *	- No more garbage collection required.  Unnecessarily allocated swap
91  *	  blocks only exist for dirty vm_page_t's now and these are already
92  *	  cycled (in a high-load system) by the pager.  We also do on-the-fly
93  *	  removal of invalidated swap blocks when a page is destroyed
94  *	  or renamed.
95  *
96  * from: Utah $Hdr: swap_pager.c 1.4 91/04/30$
97  * @(#)swap_pager.c	8.9 (Berkeley) 3/21/94
98  * $FreeBSD: src/sys/vm/swap_pager.c,v 1.130.2.12 2002/08/31 21:15:55 dillon Exp $
99  */
100 
101 #include <sys/param.h>
102 #include <sys/systm.h>
103 #include <sys/conf.h>
104 #include <sys/kernel.h>
105 #include <sys/proc.h>
106 #include <sys/buf.h>
107 #include <sys/vnode.h>
108 #include <sys/malloc.h>
109 #include <sys/vmmeter.h>
110 #include <sys/sysctl.h>
111 #include <sys/blist.h>
112 #include <sys/lock.h>
113 #include <sys/thread2.h>
114 
115 #ifndef MAX_PAGEOUT_CLUSTER
116 #define MAX_PAGEOUT_CLUSTER 16
117 #endif
118 
119 #define SWB_NPAGES	MAX_PAGEOUT_CLUSTER
120 
121 #include "opt_swap.h"
122 #include <vm/vm.h>
123 #include <vm/vm_object.h>
124 #include <vm/vm_page.h>
125 #include <vm/vm_pager.h>
126 #include <vm/vm_pageout.h>
127 #include <vm/swap_pager.h>
128 #include <vm/vm_extern.h>
129 #include <vm/vm_zone.h>
130 #include <vm/vnode_pager.h>
131 
132 #include <sys/buf2.h>
133 #include <vm/vm_page2.h>
134 
135 #define SWM_FREE	0x02	/* free, period			*/
136 #define SWM_POP		0x04	/* pop out			*/
137 
138 #define SWBIO_READ	0x01
139 #define SWBIO_WRITE	0x02
140 #define SWBIO_SYNC	0x04
141 
142 struct swfreeinfo {
143 	vm_object_t	object;
144 	vm_pindex_t	basei;
145 	vm_pindex_t	begi;
146 	vm_pindex_t	endi;	/* inclusive */
147 };
148 
149 /*
150  * vm_swap_size is in page-sized chunks now.  It was DEV_BSIZE'd chunks
151  * in the old system.
152  */
153 
154 int swap_pager_full;		/* swap space exhaustion (task killing) */
155 int vm_swap_cache_use;
156 int vm_swap_anon_use;
157 
158 static int swap_pager_almost_full; /* swap space exhaustion (w/ hysteresis)*/
159 static int nsw_rcount;		/* free read buffers			*/
160 static int nsw_wcount_sync;	/* limit write buffers / synchronous	*/
161 static int nsw_wcount_async;	/* limit write buffers / asynchronous	*/
162 static int nsw_wcount_async_max;/* assigned maximum			*/
163 static int nsw_cluster_max;	/* maximum VOP I/O allowed		*/
164 
165 struct blist *swapblist;
166 static int swap_async_max = 4;	/* maximum in-progress async I/O's	*/
167 static int swap_burst_read = 0;	/* allow burst reading */
168 
169 /* from vm_swap.c */
170 extern struct vnode *swapdev_vp;
171 extern struct swdevt *swdevt;
172 extern int nswdev;
173 
174 #define BLK2DEVIDX(blk) (nswdev > 1 ? blk / dmmax % nswdev : 0)
175 
176 SYSCTL_INT(_vm, OID_AUTO, swap_async_max,
177         CTLFLAG_RW, &swap_async_max, 0, "Maximum running async swap ops");
178 SYSCTL_INT(_vm, OID_AUTO, swap_burst_read,
179         CTLFLAG_RW, &swap_burst_read, 0, "Allow burst reads for pageins");
180 
181 SYSCTL_INT(_vm, OID_AUTO, swap_cache_use,
182         CTLFLAG_RD, &vm_swap_cache_use, 0, "");
183 SYSCTL_INT(_vm, OID_AUTO, swap_anon_use,
184         CTLFLAG_RD, &vm_swap_anon_use, 0, "");
185 SYSCTL_INT(_vm, OID_AUTO, swap_size,
186         CTLFLAG_RD, &vm_swap_size, 0, "");
187 
188 vm_zone_t		swap_zone;
189 
190 /*
191  * Red-Black tree for swblock entries
192  *
193  * The caller must hold vm_token
194  */
195 RB_GENERATE2(swblock_rb_tree, swblock, swb_entry, rb_swblock_compare,
196 	     vm_pindex_t, swb_index);
197 
198 int
199 rb_swblock_compare(struct swblock *swb1, struct swblock *swb2)
200 {
201 	if (swb1->swb_index < swb2->swb_index)
202 		return(-1);
203 	if (swb1->swb_index > swb2->swb_index)
204 		return(1);
205 	return(0);
206 }
207 
208 static
209 int
210 rb_swblock_scancmp(struct swblock *swb, void *data)
211 {
212 	struct swfreeinfo *info = data;
213 
214 	if (swb->swb_index < info->basei)
215 		return(-1);
216 	if (swb->swb_index > info->endi)
217 		return(1);
218 	return(0);
219 }
220 
221 static
222 int
223 rb_swblock_condcmp(struct swblock *swb, void *data)
224 {
225 	struct swfreeinfo *info = data;
226 
227 	if (swb->swb_index < info->basei)
228 		return(-1);
229 	return(0);
230 }
231 
232 /*
233  * pagerops for OBJT_SWAP - "swap pager".  Some ops are also global procedure
234  * calls hooked from other parts of the VM system and do not appear here.
235  * (see vm/swap_pager.h).
236  */
237 
238 static void	swap_pager_dealloc (vm_object_t object);
239 static int	swap_pager_getpage (vm_object_t, vm_page_t *, int);
240 static void	swap_chain_iodone(struct bio *biox);
241 
242 struct pagerops swappagerops = {
243 	swap_pager_dealloc,	/* deallocate an OBJT_SWAP object	*/
244 	swap_pager_getpage,	/* pagein				*/
245 	swap_pager_putpages,	/* pageout				*/
246 	swap_pager_haspage	/* get backing store status for page	*/
247 };
248 
249 /*
250  * dmmax is in page-sized chunks with the new swap system.  It was
251  * dev-bsized chunks in the old.  dmmax is always a power of 2.
252  *
253  * swap_*() routines are externally accessible.  swp_*() routines are
254  * internal.
255  */
256 
257 int dmmax;
258 static int dmmax_mask;
259 int nswap_lowat = 128;		/* in pages, swap_pager_almost_full warn */
260 int nswap_hiwat = 512;		/* in pages, swap_pager_almost_full warn */
261 
262 static __inline void	swp_sizecheck (void);
263 static void	swp_pager_async_iodone (struct bio *bio);
264 
265 /*
266  * Swap bitmap functions
267  */
268 
269 static __inline void	swp_pager_freeswapspace(vm_object_t object,
270 						swblk_t blk, int npages);
271 static __inline swblk_t	swp_pager_getswapspace(vm_object_t object, int npages);
272 
273 /*
274  * Metadata functions
275  */
276 
277 static void swp_pager_meta_convert(vm_object_t);
278 static void swp_pager_meta_build(vm_object_t, vm_pindex_t, swblk_t);
279 static void swp_pager_meta_free(vm_object_t, vm_pindex_t, vm_pindex_t);
280 static void swp_pager_meta_free_all(vm_object_t);
281 static swblk_t swp_pager_meta_ctl(vm_object_t, vm_pindex_t, int);
282 
283 /*
284  * SWP_SIZECHECK() -	update swap_pager_full indication
285  *
286  *	update the swap_pager_almost_full indication and warn when we are
287  *	about to run out of swap space, using lowat/hiwat hysteresis.
288  *
289  *	Clear swap_pager_full ( task killing ) indication when lowat is met.
290  *
291  * No restrictions on call
292  * This routine may not block.
293  * SMP races are ok.
294  */
295 static __inline void
296 swp_sizecheck(void)
297 {
298 	if (vm_swap_size < nswap_lowat) {
299 		if (swap_pager_almost_full == 0) {
300 			kprintf("swap_pager: out of swap space\n");
301 			swap_pager_almost_full = 1;
302 		}
303 	} else {
304 		swap_pager_full = 0;
305 		if (vm_swap_size > nswap_hiwat)
306 			swap_pager_almost_full = 0;
307 	}
308 }
309 
310 /*
311  * SWAP_PAGER_INIT() -	initialize the swap pager!
312  *
313  *	Expected to be started from system init.  NOTE:  This code is run
314  *	before much else so be careful what you depend on.  Most of the VM
315  *	system has yet to be initialized at this point.
316  *
317  * Called from the low level boot code only.
318  */
319 static void
320 swap_pager_init(void *arg __unused)
321 {
322 	/*
323 	 * Device Stripe, in PAGE_SIZE'd blocks
324 	 */
325 	dmmax = SWB_NPAGES * 2;
326 	dmmax_mask = ~(dmmax - 1);
327 }
328 SYSINIT(vm_mem, SI_BOOT1_VM, SI_ORDER_THIRD, swap_pager_init, NULL)
329 
330 /*
331  * SWAP_PAGER_SWAP_INIT() - swap pager initialization from pageout process
332  *
333  *	Expected to be started from pageout process once, prior to entering
334  *	its main loop.
335  *
336  * Called from the low level boot code only.
337  */
338 void
339 swap_pager_swap_init(void)
340 {
341 	int n, n2;
342 
343 	/*
344 	 * Number of in-transit swap bp operations.  Don't
345 	 * exhaust the pbufs completely.  Make sure we
346 	 * initialize workable values (0 will work for hysteresis
347 	 * but it isn't very efficient).
348 	 *
349 	 * The nsw_cluster_max is constrained by the number of pages an XIO
350 	 * holds, i.e., (MAXPHYS/PAGE_SIZE) and our locally defined
351 	 * MAX_PAGEOUT_CLUSTER.   Also be aware that swap ops are
352 	 * constrained by the swap device interleave stripe size.
353 	 *
354 	 * Currently we hardwire nsw_wcount_async to 4.  This limit is
355 	 * designed to prevent other I/O from having high latencies due to
356 	 * our pageout I/O.  The value 4 works well for one or two active swap
357 	 * devices but is probably a little low if you have more.  Even so,
358 	 * a higher value would probably generate only a limited improvement
359 	 * with three or four active swap devices since the system does not
360 	 * typically have to pageout at extreme bandwidths.   We will want
361 	 * at least 2 per swap devices, and 4 is a pretty good value if you
362 	 * have one NFS swap device due to the command/ack latency over NFS.
363 	 * So it all works out pretty well.
364 	 */
365 
366 	nsw_cluster_max = min((MAXPHYS/PAGE_SIZE), MAX_PAGEOUT_CLUSTER);
367 
368 	nsw_rcount = (nswbuf + 1) / 2;
369 	nsw_wcount_sync = (nswbuf + 3) / 4;
370 	nsw_wcount_async = 4;
371 	nsw_wcount_async_max = nsw_wcount_async;
372 
373 	/*
374 	 * The zone is dynamically allocated so generally size it to
375 	 * maxswzone (32MB to 512MB of KVM).  Set a minimum size based
376 	 * on physical memory of around 8x (each swblock can hold 16 pages).
377 	 *
378 	 * With the advent of SSDs (vs HDs) the practical (swap:memory) ratio
379 	 * has increased dramatically.
380 	 */
381 	n = vmstats.v_page_count / 2;
382 	if (maxswzone && n < maxswzone / sizeof(struct swblock))
383 		n = maxswzone / sizeof(struct swblock);
384 	n2 = n;
385 
386 	do {
387 		swap_zone = zinit(
388 			"SWAPMETA",
389 			sizeof(struct swblock),
390 			n,
391 			ZONE_INTERRUPT,
392 			1);
393 		if (swap_zone != NULL)
394 			break;
395 		/*
396 		 * if the allocation failed, try a zone two thirds the
397 		 * size of the previous attempt.
398 		 */
399 		n -= ((n + 2) / 3);
400 	} while (n > 0);
401 
402 	if (swap_zone == NULL)
403 		panic("swap_pager_swap_init: swap_zone == NULL");
404 	if (n2 != n)
405 		kprintf("Swap zone entries reduced from %d to %d.\n", n2, n);
406 }
407 
408 /*
409  * SWAP_PAGER_ALLOC() -	allocate a new OBJT_SWAP VM object and instantiate
410  *			its metadata structures.
411  *
412  *	This routine is called from the mmap and fork code to create a new
413  *	OBJT_SWAP object.  We do this by creating an OBJT_DEFAULT object
414  *	and then converting it with swp_pager_meta_convert().
415  *
416  *	We only support unnamed objects.
417  *
418  * No restrictions.
419  */
420 vm_object_t
421 swap_pager_alloc(void *handle, off_t size, vm_prot_t prot, off_t offset)
422 {
423 	vm_object_t object;
424 
425 	KKASSERT(handle == NULL);
426 	lwkt_gettoken(&vm_token);
427 	object = vm_object_allocate(OBJT_DEFAULT,
428 				    OFF_TO_IDX(offset + PAGE_MASK + size));
429 	swp_pager_meta_convert(object);
430 	lwkt_reltoken(&vm_token);
431 
432 	return (object);
433 }
434 
435 /*
436  * SWAP_PAGER_DEALLOC() -	remove swap metadata from object
437  *
438  *	The swap backing for the object is destroyed.  The code is
439  *	designed such that we can reinstantiate it later, but this
440  *	routine is typically called only when the entire object is
441  *	about to be destroyed.
442  *
443  * The object must be locked or unreferenceable.
444  * No other requirements.
445  */
446 static void
447 swap_pager_dealloc(vm_object_t object)
448 {
449 	lwkt_gettoken(&vm_token);
450 	vm_object_pip_wait(object, "swpdea");
451 
452 	/*
453 	 * Free all remaining metadata.  We only bother to free it from
454 	 * the swap meta data.  We do not attempt to free swapblk's still
455 	 * associated with vm_page_t's for this object.  We do not care
456 	 * if paging is still in progress on some objects.
457 	 */
458 	swp_pager_meta_free_all(object);
459 	lwkt_reltoken(&vm_token);
460 }
461 
462 /************************************************************************
463  *			SWAP PAGER BITMAP ROUTINES			*
464  ************************************************************************/
465 
466 /*
467  * SWP_PAGER_GETSWAPSPACE() -	allocate raw swap space
468  *
469  *	Allocate swap for the requested number of pages.  The starting
470  *	swap block number (a page index) is returned or SWAPBLK_NONE
471  *	if the allocation failed.
472  *
473  *	Also has the side effect of advising that somebody made a mistake
474  *	when they configured swap and didn't configure enough.
475  *
476  * The caller must hold vm_token.
477  * This routine may not block.
478  *
479  * NOTE: vm_token must be held to avoid races with bitmap frees from
480  *	 vm_page_remove() via swap_pager_page_removed().
481  */
482 static __inline swblk_t
483 swp_pager_getswapspace(vm_object_t object, int npages)
484 {
485 	swblk_t blk;
486 
487 	ASSERT_LWKT_TOKEN_HELD(&vm_token);
488 
489 	if ((blk = blist_alloc(swapblist, npages)) == SWAPBLK_NONE) {
490 		if (swap_pager_full != 2) {
491 			kprintf("swap_pager_getswapspace: failed\n");
492 			swap_pager_full = 2;
493 			swap_pager_almost_full = 1;
494 		}
495 	} else {
496 		swapacctspace(blk, -npages);
497 		if (object->type == OBJT_SWAP)
498 			vm_swap_anon_use += npages;
499 		else
500 			vm_swap_cache_use += npages;
501 		swp_sizecheck();
502 	}
503 	return(blk);
504 }
505 
506 /*
507  * SWP_PAGER_FREESWAPSPACE() -	free raw swap space
508  *
509  *	This routine returns the specified swap blocks back to the bitmap.
510  *
511  *	Note:  This routine may not block (it could in the old swap code),
512  *	and through the use of the new blist routines it does not block.
513  *
514  *	We must be called at splvm() to avoid races with bitmap frees from
515  *	vm_page_remove() aka swap_pager_page_removed().
516  *
517  * The caller must hold vm_token.
518  * This routine may not block.
519  */
520 
521 static __inline void
522 swp_pager_freeswapspace(vm_object_t object, swblk_t blk, int npages)
523 {
524 	struct swdevt *sp = &swdevt[BLK2DEVIDX(blk)];
525 
526 	sp->sw_nused -= npages;
527 	if (object->type == OBJT_SWAP)
528 		vm_swap_anon_use -= npages;
529 	else
530 		vm_swap_cache_use -= npages;
531 
532 	if (sp->sw_flags & SW_CLOSING)
533 		return;
534 
535 	blist_free(swapblist, blk, npages);
536 	vm_swap_size += npages;
537 	swp_sizecheck();
538 }
539 
540 /*
541  * SWAP_PAGER_FREESPACE() -	frees swap blocks associated with a page
542  *				range within an object.
543  *
544  *	This is a globally accessible routine.
545  *
546  *	This routine removes swapblk assignments from swap metadata.
547  *
548  *	The external callers of this routine typically have already destroyed
549  *	or renamed vm_page_t's associated with this range in the object so
550  *	we should be ok.
551  *
552  * No requirements.
553  */
554 void
555 swap_pager_freespace(vm_object_t object, vm_pindex_t start, vm_pindex_t size)
556 {
557 	lwkt_gettoken(&vm_token);
558 	swp_pager_meta_free(object, start, size);
559 	lwkt_reltoken(&vm_token);
560 }
561 
562 /*
563  * No requirements.
564  */
565 void
566 swap_pager_freespace_all(vm_object_t object)
567 {
568 	lwkt_gettoken(&vm_token);
569 	swp_pager_meta_free_all(object);
570 	lwkt_reltoken(&vm_token);
571 }
572 
573 /*
574  * This function conditionally frees swap cache swap starting at
575  * (*basei) in the object.  (count) swap blocks will be nominally freed.
576  * The actual number of blocks freed can be more or less than the
577  * requested number.
578  *
579  * This function nominally returns the number of blocks freed.  However,
580  * the actual number of blocks freed may be less then the returned value.
581  * If the function is unable to exhaust the object or if it is able to
582  * free (approximately) the requested number of blocks it returns
583  * a value n > count.
584  *
585  * If we exhaust the object we will return a value n <= count.
586  *
587  * The caller must hold vm_token.
588  *
589  * WARNING!  If count == 0 then -1 can be returned as a degenerate case,
590  *	     callers should always pass a count value > 0.
591  */
592 static int swap_pager_condfree_callback(struct swblock *swap, void *data);
593 
594 int
595 swap_pager_condfree(vm_object_t object, vm_pindex_t *basei, int count)
596 {
597 	struct swfreeinfo info;
598 
599 	ASSERT_LWKT_TOKEN_HELD(&vm_token);
600 
601 	info.object = object;
602 	info.basei = *basei;	/* skip up to this page index */
603 	info.begi = count;	/* max swap pages to destroy */
604 	info.endi = count * 8;	/* max swblocks to scan */
605 
606 	swblock_rb_tree_RB_SCAN(&object->swblock_root, rb_swblock_condcmp,
607 				swap_pager_condfree_callback, &info);
608 	*basei = info.basei;
609 	if (info.endi < 0 && info.begi <= count)
610 		info.begi = count + 1;
611 	return(count - (int)info.begi);
612 }
613 
614 /*
615  * The idea is to free whole meta-block to avoid fragmenting
616  * the swap space or disk I/O.  We only do this if NO VM pages
617  * are present.
618  *
619  * We do not have to deal with clearing PG_SWAPPED in related VM
620  * pages because there are no related VM pages.
621  *
622  * The caller must hold vm_token.
623  */
624 static int
625 swap_pager_condfree_callback(struct swblock *swap, void *data)
626 {
627 	struct swfreeinfo *info = data;
628 	vm_object_t object = info->object;
629 	int i;
630 
631 	for (i = 0; i < SWAP_META_PAGES; ++i) {
632 		if (vm_page_lookup(object, swap->swb_index + i))
633 			break;
634 	}
635 	info->basei = swap->swb_index + SWAP_META_PAGES;
636 	if (i == SWAP_META_PAGES) {
637 		info->begi -= swap->swb_count;
638 		swap_pager_freespace(object, swap->swb_index, SWAP_META_PAGES);
639 	}
640 	--info->endi;
641 	if ((int)info->begi < 0 || (int)info->endi < 0)
642 		return(-1);
643 	return(0);
644 }
645 
646 /*
647  * Called by vm_page_alloc() when a new VM page is inserted
648  * into a VM object.  Checks whether swap has been assigned to
649  * the page and sets PG_SWAPPED as necessary.
650  *
651  * No requirements.
652  */
653 void
654 swap_pager_page_inserted(vm_page_t m)
655 {
656 	if (m->object->swblock_count) {
657 		lwkt_gettoken(&vm_token);
658 		if (swp_pager_meta_ctl(m->object, m->pindex, 0) != SWAPBLK_NONE)
659 			vm_page_flag_set(m, PG_SWAPPED);
660 		lwkt_reltoken(&vm_token);
661 	}
662 }
663 
664 /*
665  * SWAP_PAGER_RESERVE() - reserve swap blocks in object
666  *
667  *	Assigns swap blocks to the specified range within the object.  The
668  *	swap blocks are not zerod.  Any previous swap assignment is destroyed.
669  *
670  *	Returns 0 on success, -1 on failure.
671  *
672  * The caller is responsible for avoiding races in the specified range.
673  * No other requirements.
674  */
675 int
676 swap_pager_reserve(vm_object_t object, vm_pindex_t start, vm_size_t size)
677 {
678 	int n = 0;
679 	swblk_t blk = SWAPBLK_NONE;
680 	vm_pindex_t beg = start;	/* save start index */
681 
682 	lwkt_gettoken(&vm_token);
683 	while (size) {
684 		if (n == 0) {
685 			n = BLIST_MAX_ALLOC;
686 			while ((blk = swp_pager_getswapspace(object, n)) ==
687 			       SWAPBLK_NONE)
688 			{
689 				n >>= 1;
690 				if (n == 0) {
691 					swp_pager_meta_free(object, beg,
692 							    start - beg);
693 					lwkt_reltoken(&vm_token);
694 					return(-1);
695 				}
696 			}
697 		}
698 		swp_pager_meta_build(object, start, blk);
699 		--size;
700 		++start;
701 		++blk;
702 		--n;
703 	}
704 	swp_pager_meta_free(object, start, n);
705 	lwkt_reltoken(&vm_token);
706 	return(0);
707 }
708 
709 /*
710  * SWAP_PAGER_COPY() -  copy blocks from source pager to destination pager
711  *			and destroy the source.
712  *
713  *	Copy any valid swapblks from the source to the destination.  In
714  *	cases where both the source and destination have a valid swapblk,
715  *	we keep the destination's.
716  *
717  *	This routine is allowed to block.  It may block allocating metadata
718  *	indirectly through swp_pager_meta_build() or if paging is still in
719  *	progress on the source.
720  *
721  *	This routine can be called at any spl
722  *
723  *	XXX vm_page_collapse() kinda expects us not to block because we
724  *	supposedly do not need to allocate memory, but for the moment we
725  *	*may* have to get a little memory from the zone allocator, but
726  *	it is taken from the interrupt memory.  We should be ok.
727  *
728  *	The source object contains no vm_page_t's (which is just as well)
729  *
730  *	The source object is of type OBJT_SWAP.
731  *
732  *	The source and destination objects must be locked or
733  *	inaccessible (XXX are they ?)
734  *
735  * The caller must hold vm_token.
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 	ASSERT_LWKT_TOKEN_HELD(&vm_token);
744 
745 	/*
746 	 * transfer source to destination.
747 	 */
748 	for (i = 0; i < dstobject->size; ++i) {
749 		swblk_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 			swblk_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 }
798 
799 /*
800  * SWAP_PAGER_HASPAGE() -	determine if we have good backing store for
801  *				the requested page.
802  *
803  *	We determine whether good backing store exists for the requested
804  *	page and return TRUE if it does, FALSE if it doesn't.
805  *
806  *	If TRUE, we also try to determine how much valid, contiguous backing
807  *	store exists before and after the requested page within a reasonable
808  *	distance.  We do not try to restrict it to the swap device stripe
809  *	(that is handled in getpages/putpages).  It probably isn't worth
810  *	doing here.
811  *
812  * No requirements.
813  */
814 boolean_t
815 swap_pager_haspage(vm_object_t object, vm_pindex_t pindex)
816 {
817 	swblk_t blk0;
818 
819 	/*
820 	 * do we have good backing store at the requested index ?
821 	 */
822 
823 	lwkt_gettoken(&vm_token);
824 	blk0 = swp_pager_meta_ctl(object, pindex, 0);
825 
826 	if (blk0 == SWAPBLK_NONE) {
827 		lwkt_reltoken(&vm_token);
828 		return (FALSE);
829 	}
830 	lwkt_reltoken(&vm_token);
831 	return (TRUE);
832 }
833 
834 /*
835  * SWAP_PAGER_PAGE_UNSWAPPED() - remove swap backing store related to page
836  *
837  * This removes any associated swap backing store, whether valid or
838  * not, from the page.  This operates on any VM object, not just OBJT_SWAP
839  * objects.
840  *
841  * This routine is typically called when a page is made dirty, at
842  * which point any associated swap can be freed.  MADV_FREE also
843  * calls us in a special-case situation
844  *
845  * NOTE!!!  If the page is clean and the swap was valid, the caller
846  * should make the page dirty before calling this routine.  This routine
847  * does NOT change the m->dirty status of the page.  Also: MADV_FREE
848  * depends on it.
849  *
850  * The page must be busied or soft-busied.
851  * The caller must hold vm_token if the caller does not wish to block here.
852  * No other requirements.
853  */
854 void
855 swap_pager_unswapped(vm_page_t m)
856 {
857 	if (m->flags & PG_SWAPPED) {
858 		lwkt_gettoken(&vm_token);
859 		KKASSERT(m->flags & PG_SWAPPED);
860 		swp_pager_meta_ctl(m->object, m->pindex, SWM_FREE);
861 		vm_page_flag_clear(m, PG_SWAPPED);
862 		lwkt_reltoken(&vm_token);
863 	}
864 }
865 
866 /*
867  * SWAP_PAGER_STRATEGY() - read, write, free blocks
868  *
869  * This implements a VM OBJECT strategy function using swap backing store.
870  * This can operate on any VM OBJECT type, not necessarily just OBJT_SWAP
871  * types.
872  *
873  * This is intended to be a cacheless interface (i.e. caching occurs at
874  * higher levels), and is also used as a swap-based SSD cache for vnode
875  * and device objects.
876  *
877  * All I/O goes directly to and from the swap device.
878  *
879  * We currently attempt to run I/O synchronously or asynchronously as
880  * the caller requests.  This isn't perfect because we loose error
881  * sequencing when we run multiple ops in parallel to satisfy a request.
882  * But this is swap, so we let it all hang out.
883  *
884  * No requirements.
885  */
886 void
887 swap_pager_strategy(vm_object_t object, struct bio *bio)
888 {
889 	struct buf *bp = bio->bio_buf;
890 	struct bio *nbio;
891 	vm_pindex_t start;
892 	vm_pindex_t biox_blkno = 0;
893 	int count;
894 	char *data;
895 	struct bio *biox;
896 	struct buf *bufx;
897 	struct bio_track *track;
898 
899 	/*
900 	 * tracking for swapdev vnode I/Os
901 	 */
902 	if (bp->b_cmd == BUF_CMD_READ)
903 		track = &swapdev_vp->v_track_read;
904 	else
905 		track = &swapdev_vp->v_track_write;
906 
907 	if (bp->b_bcount & PAGE_MASK) {
908 		bp->b_error = EINVAL;
909 		bp->b_flags |= B_ERROR | B_INVAL;
910 		biodone(bio);
911 		kprintf("swap_pager_strategy: bp %p offset %lld size %d, "
912 			"not page bounded\n",
913 			bp, (long long)bio->bio_offset, (int)bp->b_bcount);
914 		return;
915 	}
916 
917 	/*
918 	 * Clear error indication, initialize page index, count, data pointer.
919 	 */
920 	bp->b_error = 0;
921 	bp->b_flags &= ~B_ERROR;
922 	bp->b_resid = bp->b_bcount;
923 
924 	start = (vm_pindex_t)(bio->bio_offset >> PAGE_SHIFT);
925 	count = howmany(bp->b_bcount, PAGE_SIZE);
926 	data = bp->b_data;
927 
928 	/*
929 	 * Deal with BUF_CMD_FREEBLKS
930 	 */
931 	if (bp->b_cmd == BUF_CMD_FREEBLKS) {
932 		/*
933 		 * FREE PAGE(s) - destroy underlying swap that is no longer
934 		 *		  needed.
935 		 */
936 		lwkt_gettoken(&vm_token);
937 		swp_pager_meta_free(object, start, count);
938 		lwkt_reltoken(&vm_token);
939 		bp->b_resid = 0;
940 		biodone(bio);
941 		return;
942 	}
943 
944 	/*
945 	 * We need to be able to create a new cluster of I/O's.  We cannot
946 	 * use the caller fields of the passed bio so push a new one.
947 	 *
948 	 * Because nbio is just a placeholder for the cluster links,
949 	 * we can biodone() the original bio instead of nbio to make
950 	 * things a bit more efficient.
951 	 */
952 	nbio = push_bio(bio);
953 	nbio->bio_offset = bio->bio_offset;
954 	nbio->bio_caller_info1.cluster_head = NULL;
955 	nbio->bio_caller_info2.cluster_tail = NULL;
956 
957 	biox = NULL;
958 	bufx = NULL;
959 
960 	/*
961 	 * Execute read or write
962 	 */
963 	lwkt_gettoken(&vm_token);
964 	while (count > 0) {
965 		swblk_t blk;
966 
967 		/*
968 		 * Obtain block.  If block not found and writing, allocate a
969 		 * new block and build it into the object.
970 		 */
971 		blk = swp_pager_meta_ctl(object, start, 0);
972 		if ((blk == SWAPBLK_NONE) && bp->b_cmd != BUF_CMD_READ) {
973 			blk = swp_pager_getswapspace(object, 1);
974 			if (blk == SWAPBLK_NONE) {
975 				bp->b_error = ENOMEM;
976 				bp->b_flags |= B_ERROR;
977 				break;
978 			}
979 			swp_pager_meta_build(object, start, blk);
980 		}
981 
982 		/*
983 		 * Do we have to flush our current collection?  Yes if:
984 		 *
985 		 *	- no swap block at this index
986 		 *	- swap block is not contiguous
987 		 *	- we cross a physical disk boundry in the
988 		 *	  stripe.
989 		 */
990 		if (
991 		    biox && (biox_blkno + btoc(bufx->b_bcount) != blk ||
992 		     ((biox_blkno ^ blk) & dmmax_mask)
993 		    )
994 		) {
995 			if (bp->b_cmd == BUF_CMD_READ) {
996 				++mycpu->gd_cnt.v_swapin;
997 				mycpu->gd_cnt.v_swappgsin += btoc(bufx->b_bcount);
998 			} else {
999 				++mycpu->gd_cnt.v_swapout;
1000 				mycpu->gd_cnt.v_swappgsout += btoc(bufx->b_bcount);
1001 				bufx->b_dirtyend = bufx->b_bcount;
1002 			}
1003 
1004 			/*
1005 			 * Finished with this buf.
1006 			 */
1007 			KKASSERT(bufx->b_bcount != 0);
1008 			if (bufx->b_cmd != BUF_CMD_READ)
1009 				bufx->b_dirtyend = bufx->b_bcount;
1010 			biox = NULL;
1011 			bufx = NULL;
1012 		}
1013 
1014 		/*
1015 		 * Add new swapblk to biox, instantiating biox if necessary.
1016 		 * Zero-fill reads are able to take a shortcut.
1017 		 */
1018 		if (blk == SWAPBLK_NONE) {
1019 			/*
1020 			 * We can only get here if we are reading.  Since
1021 			 * we are at splvm() we can safely modify b_resid,
1022 			 * even if chain ops are in progress.
1023 			 */
1024 			bzero(data, PAGE_SIZE);
1025 			bp->b_resid -= PAGE_SIZE;
1026 		} else {
1027 			if (biox == NULL) {
1028 				/* XXX chain count > 4, wait to <= 4 */
1029 
1030 				bufx = getpbuf(NULL);
1031 				biox = &bufx->b_bio1;
1032 				cluster_append(nbio, bufx);
1033 				bufx->b_flags |= (bufx->b_flags & B_ORDERED);
1034 				bufx->b_cmd = bp->b_cmd;
1035 				biox->bio_done = swap_chain_iodone;
1036 				biox->bio_offset = (off_t)blk << PAGE_SHIFT;
1037 				biox->bio_caller_info1.cluster_parent = nbio;
1038 				biox_blkno = blk;
1039 				bufx->b_bcount = 0;
1040 				bufx->b_data = data;
1041 			}
1042 			bufx->b_bcount += PAGE_SIZE;
1043 		}
1044 		--count;
1045 		++start;
1046 		data += PAGE_SIZE;
1047 	}
1048 	lwkt_reltoken(&vm_token);
1049 
1050 	/*
1051 	 *  Flush out last buffer
1052 	 */
1053 	if (biox) {
1054 		if (bufx->b_cmd == BUF_CMD_READ) {
1055 			++mycpu->gd_cnt.v_swapin;
1056 			mycpu->gd_cnt.v_swappgsin += btoc(bufx->b_bcount);
1057 		} else {
1058 			++mycpu->gd_cnt.v_swapout;
1059 			mycpu->gd_cnt.v_swappgsout += btoc(bufx->b_bcount);
1060 			bufx->b_dirtyend = bufx->b_bcount;
1061 		}
1062 		KKASSERT(bufx->b_bcount);
1063 		if (bufx->b_cmd != BUF_CMD_READ)
1064 			bufx->b_dirtyend = bufx->b_bcount;
1065 		/* biox, bufx = NULL */
1066 	}
1067 
1068 	/*
1069 	 * Now initiate all the I/O.  Be careful looping on our chain as
1070 	 * I/O's may complete while we are still initiating them.
1071 	 *
1072 	 * If the request is a 100% sparse read no bios will be present
1073 	 * and we just biodone() the buffer.
1074 	 */
1075 	nbio->bio_caller_info2.cluster_tail = NULL;
1076 	bufx = nbio->bio_caller_info1.cluster_head;
1077 
1078 	if (bufx) {
1079 		while (bufx) {
1080 			biox = &bufx->b_bio1;
1081 			BUF_KERNPROC(bufx);
1082 			bufx = bufx->b_cluster_next;
1083 			vn_strategy(swapdev_vp, biox);
1084 		}
1085 	} else {
1086 		biodone(bio);
1087 	}
1088 
1089 	/*
1090 	 * Completion of the cluster will also call biodone_chain(nbio).
1091 	 * We never call biodone(nbio) so we don't have to worry about
1092 	 * setting up a bio_done callback.  It's handled in the sub-IO.
1093 	 */
1094 	/**/
1095 }
1096 
1097 /*
1098  * biodone callback
1099  *
1100  * No requirements.
1101  */
1102 static void
1103 swap_chain_iodone(struct bio *biox)
1104 {
1105 	struct buf **nextp;
1106 	struct buf *bufx;	/* chained sub-buffer */
1107 	struct bio *nbio;	/* parent nbio with chain glue */
1108 	struct buf *bp;		/* original bp associated with nbio */
1109 	int chain_empty;
1110 
1111 	bufx = biox->bio_buf;
1112 	nbio = biox->bio_caller_info1.cluster_parent;
1113 	bp = nbio->bio_buf;
1114 
1115 	/*
1116 	 * Update the original buffer
1117 	 */
1118         KKASSERT(bp != NULL);
1119 	if (bufx->b_flags & B_ERROR) {
1120 		atomic_set_int(&bufx->b_flags, B_ERROR);
1121 		bp->b_error = bufx->b_error;	/* race ok */
1122 	} else if (bufx->b_resid != 0) {
1123 		atomic_set_int(&bufx->b_flags, B_ERROR);
1124 		bp->b_error = EINVAL;		/* race ok */
1125 	} else {
1126 		atomic_subtract_int(&bp->b_resid, bufx->b_bcount);
1127 	}
1128 
1129 	/*
1130 	 * Remove us from the chain.
1131 	 */
1132 	spin_lock(&bp->b_lock.lk_spinlock);
1133 	nextp = &nbio->bio_caller_info1.cluster_head;
1134 	while (*nextp != bufx) {
1135 		KKASSERT(*nextp != NULL);
1136 		nextp = &(*nextp)->b_cluster_next;
1137 	}
1138 	*nextp = bufx->b_cluster_next;
1139 	chain_empty = (nbio->bio_caller_info1.cluster_head == NULL);
1140 	spin_unlock(&bp->b_lock.lk_spinlock);
1141 
1142 	/*
1143 	 * Clean up bufx.  If the chain is now empty we finish out
1144 	 * the parent.  Note that we may be racing other completions
1145 	 * so we must use the chain_empty status from above.
1146 	 */
1147 	if (chain_empty) {
1148 		if (bp->b_resid != 0 && !(bp->b_flags & B_ERROR)) {
1149 			atomic_set_int(&bp->b_flags, B_ERROR);
1150 			bp->b_error = EINVAL;
1151 		}
1152 		biodone_chain(nbio);
1153         }
1154         relpbuf(bufx, NULL);
1155 }
1156 
1157 /*
1158  * SWAP_PAGER_GETPAGES() - bring page in from swap
1159  *
1160  * The requested page may have to be brought in from swap.  Calculate the
1161  * swap block and bring in additional pages if possible.  All pages must
1162  * have contiguous swap block assignments and reside in the same object.
1163  *
1164  * The caller has a single vm_object_pip_add() reference prior to
1165  * calling us and we should return with the same.
1166  *
1167  * The caller has BUSY'd the page.  We should return with (*mpp) left busy,
1168  * and any additinal pages unbusied.
1169  *
1170  * If the caller encounters a PG_RAM page it will pass it to us even though
1171  * it may be valid and dirty.  We cannot overwrite the page in this case!
1172  * The case is used to allow us to issue pure read-aheads.
1173  *
1174  * NOTE! XXX This code does not entirely pipeline yet due to the fact that
1175  *       the PG_RAM page is validated at the same time as mreq.  What we
1176  *	 really need to do is issue a separate read-ahead pbuf.
1177  *
1178  * No requirements.
1179  */
1180 static int
1181 swap_pager_getpage(vm_object_t object, vm_page_t *mpp, int seqaccess)
1182 {
1183 	struct buf *bp;
1184 	struct bio *bio;
1185 	vm_page_t mreq;
1186 	vm_page_t m;
1187 	vm_offset_t kva;
1188 	swblk_t blk;
1189 	int i;
1190 	int j;
1191 	int raonly;
1192 	vm_page_t marray[XIO_INTERNAL_PAGES];
1193 
1194 	mreq = *mpp;
1195 
1196 	if (mreq->object != object) {
1197 		panic("swap_pager_getpages: object mismatch %p/%p",
1198 		    object,
1199 		    mreq->object
1200 		);
1201 	}
1202 
1203 	/*
1204 	 * We don't want to overwrite a fully valid page as it might be
1205 	 * dirty.  This case can occur when e.g. vm_fault hits a perfectly
1206 	 * valid page with PG_RAM set.
1207 	 *
1208 	 * In this case we see if the next page is a suitable page-in
1209 	 * candidate and if it is we issue read-ahead.  PG_RAM will be
1210 	 * set on the last page of the read-ahead to continue the pipeline.
1211 	 */
1212 	if (mreq->valid == VM_PAGE_BITS_ALL) {
1213 		if (swap_burst_read == 0 || mreq->pindex + 1 >= object->size)
1214 			return(VM_PAGER_OK);
1215 		lwkt_gettoken(&vm_token);
1216 		blk = swp_pager_meta_ctl(object, mreq->pindex + 1, 0);
1217 		if (blk == SWAPBLK_NONE) {
1218 			lwkt_reltoken(&vm_token);
1219 			return(VM_PAGER_OK);
1220 		}
1221 		m = vm_page_lookup(object, mreq->pindex + 1);
1222 		if (m == NULL) {
1223 			m = vm_page_alloc(object, mreq->pindex + 1,
1224 					  VM_ALLOC_QUICK);
1225 			if (m == NULL) {
1226 				lwkt_reltoken(&vm_token);
1227 				return(VM_PAGER_OK);
1228 			}
1229 		} else {
1230 			if ((m->flags & PG_BUSY) || m->busy || m->valid) {
1231 				lwkt_reltoken(&vm_token);
1232 				return(VM_PAGER_OK);
1233 			}
1234 			vm_page_unqueue_nowakeup(m);
1235 			vm_page_busy(m);
1236 		}
1237 		mreq = m;
1238 		raonly = 1;
1239 		lwkt_reltoken(&vm_token);
1240 	} else {
1241 		raonly = 0;
1242 	}
1243 
1244 	/*
1245 	 * Try to block-read contiguous pages from swap if sequential,
1246 	 * otherwise just read one page.  Contiguous pages from swap must
1247 	 * reside within a single device stripe because the I/O cannot be
1248 	 * broken up across multiple stripes.
1249 	 *
1250 	 * Note that blk and iblk can be SWAPBLK_NONE but the loop is
1251 	 * set up such that the case(s) are handled implicitly.
1252 	 */
1253 	lwkt_gettoken(&vm_token);
1254 	blk = swp_pager_meta_ctl(mreq->object, mreq->pindex, 0);
1255 	marray[0] = mreq;
1256 
1257 	for (i = 1; swap_burst_read &&
1258 		    i < XIO_INTERNAL_PAGES &&
1259 		    mreq->pindex + i < object->size; ++i) {
1260 		swblk_t iblk;
1261 
1262 		iblk = swp_pager_meta_ctl(object, mreq->pindex + i, 0);
1263 		if (iblk != blk + i)
1264 			break;
1265 		if ((blk ^ iblk) & dmmax_mask)
1266 			break;
1267 		m = vm_page_lookup(object, mreq->pindex + i);
1268 		if (m == NULL) {
1269 			m = vm_page_alloc(object, mreq->pindex + i,
1270 					  VM_ALLOC_QUICK);
1271 			if (m == NULL)
1272 				break;
1273 		} else {
1274 			if ((m->flags & PG_BUSY) || m->busy || m->valid)
1275 				break;
1276 			vm_page_unqueue_nowakeup(m);
1277 			vm_page_busy(m);
1278 		}
1279 		marray[i] = m;
1280 	}
1281 	if (i > 1)
1282 		vm_page_flag_set(marray[i - 1], PG_RAM);
1283 
1284 	lwkt_reltoken(&vm_token);
1285 
1286 	/*
1287 	 * If mreq is the requested page and we have nothing to do return
1288 	 * VM_PAGER_FAIL.  If raonly is set mreq is just another read-ahead
1289 	 * page and must be cleaned up.
1290 	 */
1291 	if (blk == SWAPBLK_NONE) {
1292 		KKASSERT(i == 1);
1293 		if (raonly) {
1294 			vnode_pager_freepage(mreq);
1295 			return(VM_PAGER_OK);
1296 		} else {
1297 			return(VM_PAGER_FAIL);
1298 		}
1299 	}
1300 
1301 	/*
1302 	 * map our page(s) into kva for input
1303 	 */
1304 	bp = getpbuf_kva(&nsw_rcount);
1305 	bio = &bp->b_bio1;
1306 	kva = (vm_offset_t) bp->b_kvabase;
1307 	bcopy(marray, bp->b_xio.xio_pages, i * sizeof(vm_page_t));
1308 	pmap_qenter(kva, bp->b_xio.xio_pages, i);
1309 
1310 	bp->b_data = (caddr_t)kva;
1311 	bp->b_bcount = PAGE_SIZE * i;
1312 	bp->b_xio.xio_npages = i;
1313 	bio->bio_done = swp_pager_async_iodone;
1314 	bio->bio_offset = (off_t)blk << PAGE_SHIFT;
1315 	bio->bio_caller_info1.index = SWBIO_READ;
1316 
1317 	/*
1318 	 * Set index.  If raonly set the index beyond the array so all
1319 	 * the pages are treated the same, otherwise the original mreq is
1320 	 * at index 0.
1321 	 */
1322 	if (raonly)
1323 		bio->bio_driver_info = (void *)(intptr_t)i;
1324 	else
1325 		bio->bio_driver_info = (void *)(intptr_t)0;
1326 
1327 	for (j = 0; j < i; ++j)
1328 		vm_page_flag_set(bp->b_xio.xio_pages[j], PG_SWAPINPROG);
1329 
1330 	mycpu->gd_cnt.v_swapin++;
1331 	mycpu->gd_cnt.v_swappgsin += bp->b_xio.xio_npages;
1332 
1333 	/*
1334 	 * We still hold the lock on mreq, and our automatic completion routine
1335 	 * does not remove it.
1336 	 */
1337 	vm_object_pip_add(object, bp->b_xio.xio_npages);
1338 
1339 	/*
1340 	 * perform the I/O.  NOTE!!!  bp cannot be considered valid after
1341 	 * this point because we automatically release it on completion.
1342 	 * Instead, we look at the one page we are interested in which we
1343 	 * still hold a lock on even through the I/O completion.
1344 	 *
1345 	 * The other pages in our m[] array are also released on completion,
1346 	 * so we cannot assume they are valid anymore either.
1347 	 */
1348 	bp->b_cmd = BUF_CMD_READ;
1349 	BUF_KERNPROC(bp);
1350 	vn_strategy(swapdev_vp, bio);
1351 
1352 	/*
1353 	 * Wait for the page we want to complete.  PG_SWAPINPROG is always
1354 	 * cleared on completion.  If an I/O error occurs, SWAPBLK_NONE
1355 	 * is set in the meta-data.
1356 	 *
1357 	 * If this is a read-ahead only we return immediately without
1358 	 * waiting for I/O.
1359 	 */
1360 	if (raonly)
1361 		return(VM_PAGER_OK);
1362 
1363 	/*
1364 	 * Read-ahead includes originally requested page case.
1365 	 */
1366 	lwkt_gettoken(&vm_token);
1367 	while ((mreq->flags & PG_SWAPINPROG) != 0) {
1368 		vm_page_flag_set(mreq, PG_WANTED | PG_REFERENCED);
1369 		mycpu->gd_cnt.v_intrans++;
1370 		if (tsleep(mreq, 0, "swread", hz*20)) {
1371 			kprintf(
1372 			    "swap_pager: indefinite wait buffer: "
1373 				" offset: %lld, size: %ld\n",
1374 			    (long long)bio->bio_offset,
1375 			    (long)bp->b_bcount
1376 			);
1377 		}
1378 	}
1379 	lwkt_reltoken(&vm_token);
1380 
1381 	/*
1382 	 * mreq is left bussied after completion, but all the other pages
1383 	 * are freed.  If we had an unrecoverable read error the page will
1384 	 * not be valid.
1385 	 */
1386 	if (mreq->valid != VM_PAGE_BITS_ALL)
1387 		return(VM_PAGER_ERROR);
1388 	else
1389 		return(VM_PAGER_OK);
1390 
1391 	/*
1392 	 * A final note: in a low swap situation, we cannot deallocate swap
1393 	 * and mark a page dirty here because the caller is likely to mark
1394 	 * the page clean when we return, causing the page to possibly revert
1395 	 * to all-zero's later.
1396 	 */
1397 }
1398 
1399 /*
1400  *	swap_pager_putpages:
1401  *
1402  *	Assign swap (if necessary) and initiate I/O on the specified pages.
1403  *
1404  *	We support both OBJT_DEFAULT and OBJT_SWAP objects.  DEFAULT objects
1405  *	are automatically converted to SWAP objects.
1406  *
1407  *	In a low memory situation we may block in vn_strategy(), but the new
1408  *	vm_page reservation system coupled with properly written VFS devices
1409  *	should ensure that no low-memory deadlock occurs.  This is an area
1410  *	which needs work.
1411  *
1412  *	The parent has N vm_object_pip_add() references prior to
1413  *	calling us and will remove references for rtvals[] that are
1414  *	not set to VM_PAGER_PEND.  We need to remove the rest on I/O
1415  *	completion.
1416  *
1417  *	The parent has soft-busy'd the pages it passes us and will unbusy
1418  *	those whos rtvals[] entry is not set to VM_PAGER_PEND on return.
1419  *	We need to unbusy the rest on I/O completion.
1420  *
1421  * No requirements.
1422  */
1423 void
1424 swap_pager_putpages(vm_object_t object, vm_page_t *m, int count,
1425 		    boolean_t sync, int *rtvals)
1426 {
1427 	int i;
1428 	int n = 0;
1429 
1430 	if (count && m[0]->object != object) {
1431 		panic("swap_pager_getpages: object mismatch %p/%p",
1432 		    object,
1433 		    m[0]->object
1434 		);
1435 	}
1436 
1437 	/*
1438 	 * Step 1
1439 	 *
1440 	 * Turn object into OBJT_SWAP
1441 	 * check for bogus sysops
1442 	 * force sync if not pageout process
1443 	 */
1444 	if (object->type == OBJT_DEFAULT) {
1445 		lwkt_gettoken(&vm_token);
1446 		if (object->type == OBJT_DEFAULT)
1447 			swp_pager_meta_convert(object);
1448 		lwkt_reltoken(&vm_token);
1449 	}
1450 
1451 	if (curthread != pagethread)
1452 		sync = TRUE;
1453 
1454 	/*
1455 	 * Step 2
1456 	 *
1457 	 * Update nsw parameters from swap_async_max sysctl values.
1458 	 * Do not let the sysop crash the machine with bogus numbers.
1459 	 */
1460 
1461 	if (swap_async_max != nsw_wcount_async_max) {
1462 		int n;
1463 
1464 		/*
1465 		 * limit range
1466 		 */
1467 		if ((n = swap_async_max) > nswbuf / 2)
1468 			n = nswbuf / 2;
1469 		if (n < 1)
1470 			n = 1;
1471 		swap_async_max = n;
1472 
1473 		/*
1474 		 * Adjust difference ( if possible ).  If the current async
1475 		 * count is too low, we may not be able to make the adjustment
1476 		 * at this time.
1477 		 */
1478 		lwkt_gettoken(&vm_token);
1479 		n -= nsw_wcount_async_max;
1480 		if (nsw_wcount_async + n >= 0) {
1481 			nsw_wcount_async += n;
1482 			nsw_wcount_async_max += n;
1483 			wakeup(&nsw_wcount_async);
1484 		}
1485 		lwkt_reltoken(&vm_token);
1486 	}
1487 
1488 	/*
1489 	 * Step 3
1490 	 *
1491 	 * Assign swap blocks and issue I/O.  We reallocate swap on the fly.
1492 	 * The page is left dirty until the pageout operation completes
1493 	 * successfully.
1494 	 */
1495 
1496 	for (i = 0; i < count; i += n) {
1497 		struct buf *bp;
1498 		struct bio *bio;
1499 		swblk_t blk;
1500 		int j;
1501 
1502 		/*
1503 		 * Maximum I/O size is limited by a number of factors.
1504 		 */
1505 
1506 		n = min(BLIST_MAX_ALLOC, count - i);
1507 		n = min(n, nsw_cluster_max);
1508 
1509 		lwkt_gettoken(&vm_token);
1510 
1511 		/*
1512 		 * Get biggest block of swap we can.  If we fail, fall
1513 		 * back and try to allocate a smaller block.  Don't go
1514 		 * overboard trying to allocate space if it would overly
1515 		 * fragment swap.
1516 		 */
1517 		while (
1518 		    (blk = swp_pager_getswapspace(object, n)) == SWAPBLK_NONE &&
1519 		    n > 4
1520 		) {
1521 			n >>= 1;
1522 		}
1523 		if (blk == SWAPBLK_NONE) {
1524 			for (j = 0; j < n; ++j)
1525 				rtvals[i+j] = VM_PAGER_FAIL;
1526 			lwkt_reltoken(&vm_token);
1527 			continue;
1528 		}
1529 
1530 		/*
1531 		 * The I/O we are constructing cannot cross a physical
1532 		 * disk boundry in the swap stripe.  Note: we are still
1533 		 * at splvm().
1534 		 */
1535 		if ((blk ^ (blk + n)) & dmmax_mask) {
1536 			j = ((blk + dmmax) & dmmax_mask) - blk;
1537 			swp_pager_freeswapspace(object, blk + j, n - j);
1538 			n = j;
1539 		}
1540 
1541 		/*
1542 		 * All I/O parameters have been satisfied, build the I/O
1543 		 * request and assign the swap space.
1544 		 */
1545 		if (sync == TRUE)
1546 			bp = getpbuf_kva(&nsw_wcount_sync);
1547 		else
1548 			bp = getpbuf_kva(&nsw_wcount_async);
1549 		bio = &bp->b_bio1;
1550 
1551 		pmap_qenter((vm_offset_t)bp->b_data, &m[i], n);
1552 
1553 		bp->b_bcount = PAGE_SIZE * n;
1554 		bio->bio_offset = (off_t)blk << PAGE_SHIFT;
1555 
1556 		for (j = 0; j < n; ++j) {
1557 			vm_page_t mreq = m[i+j];
1558 
1559 			swp_pager_meta_build(mreq->object, mreq->pindex,
1560 					     blk + j);
1561 			if (object->type == OBJT_SWAP)
1562 				vm_page_dirty(mreq);
1563 			rtvals[i+j] = VM_PAGER_OK;
1564 
1565 			vm_page_flag_set(mreq, PG_SWAPINPROG);
1566 			bp->b_xio.xio_pages[j] = mreq;
1567 		}
1568 		bp->b_xio.xio_npages = n;
1569 
1570 		mycpu->gd_cnt.v_swapout++;
1571 		mycpu->gd_cnt.v_swappgsout += bp->b_xio.xio_npages;
1572 
1573 		lwkt_reltoken(&vm_token);
1574 
1575 		bp->b_dirtyoff = 0;		/* req'd for NFS */
1576 		bp->b_dirtyend = bp->b_bcount;	/* req'd for NFS */
1577 		bp->b_cmd = BUF_CMD_WRITE;
1578 		bio->bio_caller_info1.index = SWBIO_WRITE;
1579 
1580 		/*
1581 		 * asynchronous
1582 		 */
1583 		if (sync == FALSE) {
1584 			bio->bio_done = swp_pager_async_iodone;
1585 			BUF_KERNPROC(bp);
1586 			vn_strategy(swapdev_vp, bio);
1587 
1588 			for (j = 0; j < n; ++j)
1589 				rtvals[i+j] = VM_PAGER_PEND;
1590 			continue;
1591 		}
1592 
1593 		/*
1594 		 * Issue synchrnously.
1595 		 *
1596 		 * Wait for the sync I/O to complete, then update rtvals.
1597 		 * We just set the rtvals[] to VM_PAGER_PEND so we can call
1598 		 * our async completion routine at the end, thus avoiding a
1599 		 * double-free.
1600 		 */
1601 		bio->bio_caller_info1.index |= SWBIO_SYNC;
1602 		bio->bio_done = biodone_sync;
1603 		bio->bio_flags |= BIO_SYNC;
1604 		vn_strategy(swapdev_vp, bio);
1605 		biowait(bio, "swwrt");
1606 
1607 		for (j = 0; j < n; ++j)
1608 			rtvals[i+j] = VM_PAGER_PEND;
1609 
1610 		/*
1611 		 * Now that we are through with the bp, we can call the
1612 		 * normal async completion, which frees everything up.
1613 		 */
1614 		swp_pager_async_iodone(bio);
1615 	}
1616 }
1617 
1618 /*
1619  * No requirements.
1620  */
1621 void
1622 swap_pager_newswap(void)
1623 {
1624 	swp_sizecheck();
1625 }
1626 
1627 /*
1628  *	swp_pager_async_iodone:
1629  *
1630  *	Completion routine for asynchronous reads and writes from/to swap.
1631  *	Also called manually by synchronous code to finish up a bp.
1632  *
1633  *	For READ operations, the pages are PG_BUSY'd.  For WRITE operations,
1634  *	the pages are vm_page_t->busy'd.  For READ operations, we PG_BUSY
1635  *	unbusy all pages except the 'main' request page.  For WRITE
1636  *	operations, we vm_page_t->busy'd unbusy all pages ( we can do this
1637  *	because we marked them all VM_PAGER_PEND on return from putpages ).
1638  *
1639  *	This routine may not block.
1640  *
1641  * No requirements.
1642  */
1643 static void
1644 swp_pager_async_iodone(struct bio *bio)
1645 {
1646 	struct buf *bp = bio->bio_buf;
1647 	vm_object_t object = NULL;
1648 	int i;
1649 	int *nswptr;
1650 
1651 	/*
1652 	 * report error
1653 	 */
1654 	if (bp->b_flags & B_ERROR) {
1655 		kprintf(
1656 		    "swap_pager: I/O error - %s failed; offset %lld,"
1657 			"size %ld, error %d\n",
1658 		    ((bio->bio_caller_info1.index & SWBIO_READ) ?
1659 			"pagein" : "pageout"),
1660 		    (long long)bio->bio_offset,
1661 		    (long)bp->b_bcount,
1662 		    bp->b_error
1663 		);
1664 	}
1665 
1666 	/*
1667 	 * set object, raise to splvm().
1668 	 */
1669 	if (bp->b_xio.xio_npages)
1670 		object = bp->b_xio.xio_pages[0]->object;
1671 	lwkt_gettoken(&vm_token);
1672 
1673 	/*
1674 	 * remove the mapping for kernel virtual
1675 	 */
1676 	pmap_qremove((vm_offset_t)bp->b_data, bp->b_xio.xio_npages);
1677 
1678 	/*
1679 	 * cleanup pages.  If an error occurs writing to swap, we are in
1680 	 * very serious trouble.  If it happens to be a disk error, though,
1681 	 * we may be able to recover by reassigning the swap later on.  So
1682 	 * in this case we remove the m->swapblk assignment for the page
1683 	 * but do not free it in the rlist.  The errornous block(s) are thus
1684 	 * never reallocated as swap.  Redirty the page and continue.
1685 	 */
1686 	for (i = 0; i < bp->b_xio.xio_npages; ++i) {
1687 		vm_page_t m = bp->b_xio.xio_pages[i];
1688 
1689 		if (bp->b_flags & B_ERROR) {
1690 			/*
1691 			 * If an error occurs I'd love to throw the swapblk
1692 			 * away without freeing it back to swapspace, so it
1693 			 * can never be used again.  But I can't from an
1694 			 * interrupt.
1695 			 */
1696 
1697 			if (bio->bio_caller_info1.index & SWBIO_READ) {
1698 				/*
1699 				 * When reading, reqpage needs to stay
1700 				 * locked for the parent, but all other
1701 				 * pages can be freed.  We still want to
1702 				 * wakeup the parent waiting on the page,
1703 				 * though.  ( also: pg_reqpage can be -1 and
1704 				 * not match anything ).
1705 				 *
1706 				 * We have to wake specifically requested pages
1707 				 * up too because we cleared PG_SWAPINPROG and
1708 				 * someone may be waiting for that.
1709 				 *
1710 				 * NOTE: for reads, m->dirty will probably
1711 				 * be overridden by the original caller of
1712 				 * getpages so don't play cute tricks here.
1713 				 *
1714 				 * NOTE: We can't actually free the page from
1715 				 * here, because this is an interrupt.  It
1716 				 * is not legal to mess with object->memq
1717 				 * from an interrupt.  Deactivate the page
1718 				 * instead.
1719 				 */
1720 
1721 				m->valid = 0;
1722 				vm_page_flag_clear(m, PG_ZERO);
1723 				vm_page_flag_clear(m, PG_SWAPINPROG);
1724 
1725 				/*
1726 				 * bio_driver_info holds the requested page
1727 				 * index.
1728 				 */
1729 				if (i != (int)(intptr_t)bio->bio_driver_info) {
1730 					vm_page_deactivate(m);
1731 					vm_page_wakeup(m);
1732 				} else {
1733 					vm_page_flash(m);
1734 				}
1735 				/*
1736 				 * If i == bp->b_pager.pg_reqpage, do not wake
1737 				 * the page up.  The caller needs to.
1738 				 */
1739 			} else {
1740 				/*
1741 				 * If a write error occurs remove the swap
1742 				 * assignment (note that PG_SWAPPED may or
1743 				 * may not be set depending on prior activity).
1744 				 *
1745 				 * Re-dirty OBJT_SWAP pages as there is no
1746 				 * other backing store, we can't throw the
1747 				 * page away.
1748 				 *
1749 				 * Non-OBJT_SWAP pages (aka swapcache) must
1750 				 * not be dirtied since they may not have
1751 				 * been dirty in the first place, and they
1752 				 * do have backing store (the vnode).
1753 				 */
1754 				swp_pager_meta_ctl(m->object, m->pindex,
1755 						   SWM_FREE);
1756 				vm_page_flag_clear(m, PG_SWAPPED);
1757 				if (m->object->type == OBJT_SWAP) {
1758 					vm_page_dirty(m);
1759 					vm_page_activate(m);
1760 				}
1761 				vm_page_flag_clear(m, PG_SWAPINPROG);
1762 				vm_page_io_finish(m);
1763 			}
1764 		} else if (bio->bio_caller_info1.index & SWBIO_READ) {
1765 			/*
1766 			 * NOTE: for reads, m->dirty will probably be
1767 			 * overridden by the original caller of getpages so
1768 			 * we cannot set them in order to free the underlying
1769 			 * swap in a low-swap situation.  I don't think we'd
1770 			 * want to do that anyway, but it was an optimization
1771 			 * that existed in the old swapper for a time before
1772 			 * it got ripped out due to precisely this problem.
1773 			 *
1774 			 * clear PG_ZERO in page.
1775 			 *
1776 			 * If not the requested page then deactivate it.
1777 			 *
1778 			 * Note that the requested page, reqpage, is left
1779 			 * busied, but we still have to wake it up.  The
1780 			 * other pages are released (unbusied) by
1781 			 * vm_page_wakeup().  We do not set reqpage's
1782 			 * valid bits here, it is up to the caller.
1783 			 */
1784 
1785 			/*
1786 			 * NOTE: can't call pmap_clear_modify(m) from an
1787 			 * interrupt thread, the pmap code may have to map
1788 			 * non-kernel pmaps and currently asserts the case.
1789 			 */
1790 			/*pmap_clear_modify(m);*/
1791 			m->valid = VM_PAGE_BITS_ALL;
1792 			vm_page_undirty(m);
1793 			vm_page_flag_clear(m, PG_ZERO | PG_SWAPINPROG);
1794 			vm_page_flag_set(m, PG_SWAPPED);
1795 
1796 			/*
1797 			 * We have to wake specifically requested pages
1798 			 * up too because we cleared PG_SWAPINPROG and
1799 			 * could be waiting for it in getpages.  However,
1800 			 * be sure to not unbusy getpages specifically
1801 			 * requested page - getpages expects it to be
1802 			 * left busy.
1803 			 *
1804 			 * bio_driver_info holds the requested page
1805 			 */
1806 			if (i != (int)(intptr_t)bio->bio_driver_info) {
1807 				vm_page_deactivate(m);
1808 				vm_page_wakeup(m);
1809 			} else {
1810 				vm_page_flash(m);
1811 			}
1812 		} else {
1813 			/*
1814 			 * Mark the page clean but do not mess with the
1815 			 * pmap-layer's modified state.  That state should
1816 			 * also be clear since the caller protected the
1817 			 * page VM_PROT_READ, but allow the case.
1818 			 *
1819 			 * We are in an interrupt, avoid pmap operations.
1820 			 *
1821 			 * If we have a severe page deficit, deactivate the
1822 			 * page.  Do not try to cache it (which would also
1823 			 * involve a pmap op), because the page might still
1824 			 * be read-heavy.
1825 			 *
1826 			 * When using the swap to cache clean vnode pages
1827 			 * we do not mess with the page dirty bits.
1828 			 */
1829 			if (m->object->type == OBJT_SWAP)
1830 				vm_page_undirty(m);
1831 			vm_page_flag_clear(m, PG_SWAPINPROG);
1832 			vm_page_flag_set(m, PG_SWAPPED);
1833 			if (vm_page_count_severe())
1834 				vm_page_deactivate(m);
1835 #if 0
1836 			if (!vm_page_count_severe() || !vm_page_try_to_cache(m))
1837 				vm_page_protect(m, VM_PROT_READ);
1838 #endif
1839 			vm_page_io_finish(m);
1840 		}
1841 	}
1842 
1843 	/*
1844 	 * adjust pip.  NOTE: the original parent may still have its own
1845 	 * pip refs on the object.
1846 	 */
1847 
1848 	if (object)
1849 		vm_object_pip_wakeup_n(object, bp->b_xio.xio_npages);
1850 
1851 	/*
1852 	 * Release the physical I/O buffer.
1853 	 *
1854 	 * NOTE: Due to synchronous operations in the write case b_cmd may
1855 	 *	 already be set to BUF_CMD_DONE and BIO_SYNC may have already
1856 	 *	 been cleared.
1857 	 */
1858 	if (bio->bio_caller_info1.index & SWBIO_READ)
1859 		nswptr = &nsw_rcount;
1860 	else if (bio->bio_caller_info1.index & SWBIO_SYNC)
1861 		nswptr = &nsw_wcount_sync;
1862 	else
1863 		nswptr = &nsw_wcount_async;
1864 	bp->b_cmd = BUF_CMD_DONE;
1865 	relpbuf(bp, nswptr);
1866 	lwkt_reltoken(&vm_token);
1867 }
1868 
1869 /*
1870  * Fault-in a potentially swapped page and remove the swap reference.
1871  */
1872 static __inline void
1873 swp_pager_fault_page(vm_object_t object, vm_pindex_t pindex)
1874 {
1875 	struct vnode *vp;
1876 	vm_page_t m;
1877 	int error;
1878 
1879 	if (object->type == OBJT_VNODE) {
1880 		/*
1881 		 * Any swap related to a vnode is due to swapcache.  We must
1882 		 * vget() the vnode in case it is not active (otherwise
1883 		 * vref() will panic).  Calling vm_object_page_remove() will
1884 		 * ensure that any swap ref is removed interlocked with the
1885 		 * page.  clean_only is set to TRUE so we don't throw away
1886 		 * dirty pages.
1887 		 */
1888 		vp = object->handle;
1889 		error = vget(vp, LK_SHARED | LK_RETRY | LK_CANRECURSE);
1890 		if (error == 0) {
1891 			vm_object_page_remove(object, pindex, pindex + 1, TRUE);
1892 			vput(vp);
1893 		}
1894 	} else {
1895 		/*
1896 		 * Otherwise it is a normal OBJT_SWAP object and we can
1897 		 * fault the page in and remove the swap.
1898 		 */
1899 		m = vm_fault_object_page(object, IDX_TO_OFF(pindex),
1900 					 VM_PROT_NONE,
1901 					 VM_FAULT_DIRTY | VM_FAULT_UNSWAP,
1902 					 &error);
1903 		if (m)
1904 			vm_page_unhold(m);
1905 	}
1906 }
1907 
1908 int
1909 swap_pager_swapoff(int devidx)
1910 {
1911 	vm_object_t object;
1912 	struct swblock *swap;
1913 	swblk_t v;
1914 	int i;
1915 
1916 	lwkt_gettoken(&vm_token);
1917 	lwkt_gettoken(&vmobj_token);
1918 rescan:
1919 	TAILQ_FOREACH(object, &vm_object_list, object_list) {
1920 		if (object->type == OBJT_SWAP || object->type == OBJT_VNODE) {
1921 			RB_FOREACH(swap, swblock_rb_tree, &object->swblock_root) {
1922 				for (i = 0; i < SWAP_META_PAGES; ++i) {
1923 					v = swap->swb_pages[i];
1924 					if (v != SWAPBLK_NONE &&
1925 					    BLK2DEVIDX(v) == devidx) {
1926 						swp_pager_fault_page(
1927 						    object,
1928 						    swap->swb_index + i);
1929 						goto rescan;
1930 					}
1931 				}
1932 			}
1933 		}
1934 	}
1935 	lwkt_reltoken(&vmobj_token);
1936 	lwkt_reltoken(&vm_token);
1937 
1938 	/*
1939 	 * If we fail to locate all swblocks we just fail gracefully and
1940 	 * do not bother to restore paging on the swap device.  If the
1941 	 * user wants to retry the user can retry.
1942 	 */
1943 	if (swdevt[devidx].sw_nused)
1944 		return (1);
1945 	else
1946 		return (0);
1947 }
1948 
1949 /************************************************************************
1950  *				SWAP META DATA 				*
1951  ************************************************************************
1952  *
1953  *	These routines manipulate the swap metadata stored in the
1954  *	OBJT_SWAP object.  All swp_*() routines must be called at
1955  *	splvm() because swap can be freed up by the low level vm_page
1956  *	code which might be called from interrupts beyond what splbio() covers.
1957  *
1958  *	Swap metadata is implemented with a global hash and not directly
1959  *	linked into the object.  Instead the object simply contains
1960  *	appropriate tracking counters.
1961  */
1962 
1963 /*
1964  * Lookup the swblock containing the specified swap block index.
1965  *
1966  * The caller must hold vm_token.
1967  */
1968 static __inline
1969 struct swblock *
1970 swp_pager_lookup(vm_object_t object, vm_pindex_t index)
1971 {
1972 	index &= ~SWAP_META_MASK;
1973 	return (RB_LOOKUP(swblock_rb_tree, &object->swblock_root, index));
1974 }
1975 
1976 /*
1977  * Remove a swblock from the RB tree.
1978  *
1979  * The caller must hold vm_token.
1980  */
1981 static __inline
1982 void
1983 swp_pager_remove(vm_object_t object, struct swblock *swap)
1984 {
1985 	RB_REMOVE(swblock_rb_tree, &object->swblock_root, swap);
1986 }
1987 
1988 /*
1989  * Convert default object to swap object if necessary
1990  *
1991  * The caller must hold vm_token.
1992  */
1993 static void
1994 swp_pager_meta_convert(vm_object_t object)
1995 {
1996 	if (object->type == OBJT_DEFAULT) {
1997 		object->type = OBJT_SWAP;
1998 		KKASSERT(object->swblock_count == 0);
1999 	}
2000 }
2001 
2002 /*
2003  * SWP_PAGER_META_BUILD() -	add swap block to swap meta data for object
2004  *
2005  *	We first convert the object to a swap object if it is a default
2006  *	object.  Vnode objects do not need to be converted.
2007  *
2008  *	The specified swapblk is added to the object's swap metadata.  If
2009  *	the swapblk is not valid, it is freed instead.  Any previously
2010  *	assigned swapblk is freed.
2011  *
2012  * The caller must hold vm_token.
2013  */
2014 static void
2015 swp_pager_meta_build(vm_object_t object, vm_pindex_t index, swblk_t swapblk)
2016 {
2017 	struct swblock *swap;
2018 	struct swblock *oswap;
2019 
2020 	KKASSERT(swapblk != SWAPBLK_NONE);
2021 
2022 	/*
2023 	 * Convert object if necessary
2024 	 */
2025 	if (object->type == OBJT_DEFAULT)
2026 		swp_pager_meta_convert(object);
2027 
2028 	/*
2029 	 * Locate swblock.  If not found create, but if we aren't adding
2030 	 * anything just return.  If we run out of space in the map we wait
2031 	 * and, since the hash table may have changed, retry.
2032 	 */
2033 retry:
2034 	swap = swp_pager_lookup(object, index);
2035 
2036 	if (swap == NULL) {
2037 		int i;
2038 
2039 		swap = zalloc(swap_zone);
2040 		if (swap == NULL) {
2041 			vm_wait(0);
2042 			goto retry;
2043 		}
2044 		swap->swb_index = index & ~SWAP_META_MASK;
2045 		swap->swb_count = 0;
2046 
2047 		++object->swblock_count;
2048 
2049 		for (i = 0; i < SWAP_META_PAGES; ++i)
2050 			swap->swb_pages[i] = SWAPBLK_NONE;
2051 		oswap = RB_INSERT(swblock_rb_tree, &object->swblock_root, swap);
2052 		KKASSERT(oswap == NULL);
2053 	}
2054 
2055 	/*
2056 	 * Delete prior contents of metadata
2057 	 */
2058 
2059 	index &= SWAP_META_MASK;
2060 
2061 	if (swap->swb_pages[index] != SWAPBLK_NONE) {
2062 		swp_pager_freeswapspace(object, swap->swb_pages[index], 1);
2063 		--swap->swb_count;
2064 	}
2065 
2066 	/*
2067 	 * Enter block into metadata
2068 	 */
2069 	swap->swb_pages[index] = swapblk;
2070 	if (swapblk != SWAPBLK_NONE)
2071 		++swap->swb_count;
2072 }
2073 
2074 /*
2075  * SWP_PAGER_META_FREE() - free a range of blocks in the object's swap metadata
2076  *
2077  *	The requested range of blocks is freed, with any associated swap
2078  *	returned to the swap bitmap.
2079  *
2080  *	This routine will free swap metadata structures as they are cleaned
2081  *	out.  This routine does *NOT* operate on swap metadata associated
2082  *	with resident pages.
2083  *
2084  * The caller must hold vm_token.
2085  */
2086 static int swp_pager_meta_free_callback(struct swblock *swb, void *data);
2087 
2088 static void
2089 swp_pager_meta_free(vm_object_t object, vm_pindex_t index, vm_pindex_t count)
2090 {
2091 	struct swfreeinfo info;
2092 
2093 	/*
2094 	 * Nothing to do
2095 	 */
2096 	if (object->swblock_count == 0) {
2097 		KKASSERT(RB_EMPTY(&object->swblock_root));
2098 		return;
2099 	}
2100 	if (count == 0)
2101 		return;
2102 
2103 	/*
2104 	 * Setup for RB tree scan.  Note that the pindex range can be huge
2105 	 * due to the 64 bit page index space so we cannot safely iterate.
2106 	 */
2107 	info.object = object;
2108 	info.basei = index & ~SWAP_META_MASK;
2109 	info.begi = index;
2110 	info.endi = index + count - 1;
2111 	swblock_rb_tree_RB_SCAN(&object->swblock_root, rb_swblock_scancmp,
2112 				swp_pager_meta_free_callback, &info);
2113 }
2114 
2115 /*
2116  * The caller must hold vm_token.
2117  */
2118 static
2119 int
2120 swp_pager_meta_free_callback(struct swblock *swap, void *data)
2121 {
2122 	struct swfreeinfo *info = data;
2123 	vm_object_t object = info->object;
2124 	int index;
2125 	int eindex;
2126 
2127 	/*
2128 	 * Figure out the range within the swblock.  The wider scan may
2129 	 * return edge-case swap blocks when the start and/or end points
2130 	 * are in the middle of a block.
2131 	 */
2132 	if (swap->swb_index < info->begi)
2133 		index = (int)info->begi & SWAP_META_MASK;
2134 	else
2135 		index = 0;
2136 
2137 	if (swap->swb_index + SWAP_META_PAGES > info->endi)
2138 		eindex = (int)info->endi & SWAP_META_MASK;
2139 	else
2140 		eindex = SWAP_META_MASK;
2141 
2142 	/*
2143 	 * Scan and free the blocks.  The loop terminates early
2144 	 * if (swap) runs out of blocks and could be freed.
2145 	 */
2146 	while (index <= eindex) {
2147 		swblk_t v = swap->swb_pages[index];
2148 
2149 		if (v != SWAPBLK_NONE) {
2150 			swp_pager_freeswapspace(object, v, 1);
2151 			swap->swb_pages[index] = SWAPBLK_NONE;
2152 			if (--swap->swb_count == 0) {
2153 				swp_pager_remove(object, swap);
2154 				zfree(swap_zone, swap);
2155 				--object->swblock_count;
2156 				break;
2157 			}
2158 		}
2159 		++index;
2160 	}
2161 	/* swap may be invalid here due to zfree above */
2162 	return(0);
2163 }
2164 
2165 /*
2166  * SWP_PAGER_META_FREE_ALL() - destroy all swap metadata associated with object
2167  *
2168  *	This routine locates and destroys all swap metadata associated with
2169  *	an object.
2170  *
2171  * The caller must hold vm_token.
2172  */
2173 static void
2174 swp_pager_meta_free_all(vm_object_t object)
2175 {
2176 	struct swblock *swap;
2177 	int i;
2178 
2179 	while ((swap = RB_ROOT(&object->swblock_root)) != NULL) {
2180 		swp_pager_remove(object, swap);
2181 		for (i = 0; i < SWAP_META_PAGES; ++i) {
2182 			swblk_t v = swap->swb_pages[i];
2183 			if (v != SWAPBLK_NONE) {
2184 				--swap->swb_count;
2185 				swp_pager_freeswapspace(object, v, 1);
2186 			}
2187 		}
2188 		if (swap->swb_count != 0)
2189 			panic("swap_pager_meta_free_all: swb_count != 0");
2190 		zfree(swap_zone, swap);
2191 		--object->swblock_count;
2192 	}
2193 	KKASSERT(object->swblock_count == 0);
2194 }
2195 
2196 /*
2197  * SWP_PAGER_METACTL() -  misc control of swap and vm_page_t meta data.
2198  *
2199  *	This routine is capable of looking up, popping, or freeing
2200  *	swapblk assignments in the swap meta data or in the vm_page_t.
2201  *	The routine typically returns the swapblk being looked-up, or popped,
2202  *	or SWAPBLK_NONE if the block was freed, or SWAPBLK_NONE if the block
2203  *	was invalid.  This routine will automatically free any invalid
2204  *	meta-data swapblks.
2205  *
2206  *	It is not possible to store invalid swapblks in the swap meta data
2207  *	(other then a literal 'SWAPBLK_NONE'), so we don't bother checking.
2208  *
2209  *	When acting on a busy resident page and paging is in progress, we
2210  *	have to wait until paging is complete but otherwise can act on the
2211  *	busy page.
2212  *
2213  *	SWM_FREE	remove and free swap block from metadata
2214  *	SWM_POP		remove from meta data but do not free.. pop it out
2215  *
2216  * The caller must hold vm_token.
2217  */
2218 static swblk_t
2219 swp_pager_meta_ctl(vm_object_t object, vm_pindex_t index, int flags)
2220 {
2221 	struct swblock *swap;
2222 	swblk_t r1;
2223 
2224 	if (object->swblock_count == 0)
2225 		return(SWAPBLK_NONE);
2226 
2227 	r1 = SWAPBLK_NONE;
2228 	swap = swp_pager_lookup(object, index);
2229 
2230 	if (swap != NULL) {
2231 		index &= SWAP_META_MASK;
2232 		r1 = swap->swb_pages[index];
2233 
2234 		if (r1 != SWAPBLK_NONE) {
2235 			if (flags & SWM_FREE) {
2236 				swp_pager_freeswapspace(object, r1, 1);
2237 				r1 = SWAPBLK_NONE;
2238 			}
2239 			if (flags & (SWM_FREE|SWM_POP)) {
2240 				swap->swb_pages[index] = SWAPBLK_NONE;
2241 				if (--swap->swb_count == 0) {
2242 					swp_pager_remove(object, swap);
2243 					zfree(swap_zone, swap);
2244 					--object->swblock_count;
2245 				}
2246 			}
2247 		}
2248 	}
2249 	return(r1);
2250 }
2251