xref: /dflybsd-src/sys/vm/swap_pager.c (revision cf6a53ca558fa4bbc637ee3949e2436254bcf4c2)
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. Neither the name of the University nor the names of its contributors
54  *    may be used to endorse or promote products derived from this software
55  *    without specific prior written permission.
56  *
57  * THIS SOFTWARE IS PROVIDED BY THE REGENTS AND CONTRIBUTORS ``AS IS'' AND
58  * ANY EXPRESS OR IMPLIED WARRANTIES, INCLUDING, BUT NOT LIMITED TO, THE
59  * IMPLIED WARRANTIES OF MERCHANTABILITY AND FITNESS FOR A PARTICULAR PURPOSE
60  * ARE DISCLAIMED.  IN NO EVENT SHALL THE REGENTS OR CONTRIBUTORS BE LIABLE
61  * FOR ANY DIRECT, INDIRECT, INCIDENTAL, SPECIAL, EXEMPLARY, OR CONSEQUENTIAL
62  * DAMAGES (INCLUDING, BUT NOT LIMITED TO, PROCUREMENT OF SUBSTITUTE GOODS
63  * OR SERVICES; LOSS OF USE, DATA, OR PROFITS; OR BUSINESS INTERRUPTION)
64  * HOWEVER CAUSED AND ON ANY THEORY OF LIABILITY, WHETHER IN CONTRACT, STRICT
65  * LIABILITY, OR TORT (INCLUDING NEGLIGENCE OR OTHERWISE) ARISING IN ANY WAY
66  * OUT OF THE USE OF THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF
67  * SUCH DAMAGE.
68  *
69  *				New Swap System
70  *				Matthew Dillon
71  *
72  * Radix Bitmap 'blists'.
73  *
74  *	- The new swapper uses the new radix bitmap code.  This should scale
75  *	  to arbitrarily small or arbitrarily large swap spaces and an almost
76  *	  arbitrary degree of fragmentation.
77  *
78  * Features:
79  *
80  *	- on the fly reallocation of swap during putpages.  The new system
81  *	  does not try to keep previously allocated swap blocks for dirty
82  *	  pages.
83  *
84  *	- on the fly deallocation of swap
85  *
86  *	- No more garbage collection required.  Unnecessarily allocated swap
87  *	  blocks only exist for dirty vm_page_t's now and these are already
88  *	  cycled (in a high-load system) by the pager.  We also do on-the-fly
89  *	  removal of invalidated swap blocks when a page is destroyed
90  *	  or renamed.
91  *
92  * from: Utah $Hdr: swap_pager.c 1.4 91/04/30$
93  * @(#)swap_pager.c	8.9 (Berkeley) 3/21/94
94  * $FreeBSD: src/sys/vm/swap_pager.c,v 1.130.2.12 2002/08/31 21:15:55 dillon Exp $
95  */
96 
97 #include <sys/param.h>
98 #include <sys/systm.h>
99 #include <sys/conf.h>
100 #include <sys/kernel.h>
101 #include <sys/proc.h>
102 #include <sys/buf.h>
103 #include <sys/vnode.h>
104 #include <sys/malloc.h>
105 #include <sys/vmmeter.h>
106 #include <sys/sysctl.h>
107 #include <sys/blist.h>
108 #include <sys/lock.h>
109 #include <sys/thread2.h>
110 
111 #include "opt_swap.h"
112 #include <vm/vm.h>
113 #include <vm/vm_object.h>
114 #include <vm/vm_page.h>
115 #include <vm/vm_pager.h>
116 #include <vm/vm_pageout.h>
117 #include <vm/swap_pager.h>
118 #include <vm/vm_extern.h>
119 #include <vm/vm_zone.h>
120 #include <vm/vnode_pager.h>
121 
122 #include <sys/buf2.h>
123 #include <vm/vm_page2.h>
124 
125 #ifndef MAX_PAGEOUT_CLUSTER
126 #define MAX_PAGEOUT_CLUSTER	SWB_NPAGES
127 #endif
128 
129 #define SWM_FREE	0x02	/* free, period			*/
130 #define SWM_POP		0x04	/* pop out			*/
131 
132 #define SWBIO_READ	0x01
133 #define SWBIO_WRITE	0x02
134 #define SWBIO_SYNC	0x04
135 
136 struct swfreeinfo {
137 	vm_object_t	object;
138 	vm_pindex_t	basei;
139 	vm_pindex_t	begi;
140 	vm_pindex_t	endi;	/* inclusive */
141 };
142 
143 struct swswapoffinfo {
144 	vm_object_t	object;
145 	int		devidx;
146 	int		shared;
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 static int vm_report_swap_allocs;
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 static swblk_t swapiterator;	/* linearize allocations */
170 
171 /* from vm_swap.c */
172 extern struct vnode *swapdev_vp;
173 extern struct swdevt *swdevt;
174 extern int nswdev;
175 
176 #define BLK2DEVIDX(blk) (nswdev > 1 ? blk / dmmax % nswdev : 0)
177 
178 SYSCTL_INT(_vm, OID_AUTO, swap_async_max,
179         CTLFLAG_RW, &swap_async_max, 0, "Maximum running async swap ops");
180 SYSCTL_INT(_vm, OID_AUTO, swap_burst_read,
181         CTLFLAG_RW, &swap_burst_read, 0, "Allow burst reads for pageins");
182 
183 SYSCTL_INT(_vm, OID_AUTO, swap_cache_use,
184         CTLFLAG_RD, &vm_swap_cache_use, 0, "");
185 SYSCTL_INT(_vm, OID_AUTO, swap_anon_use,
186         CTLFLAG_RD, &vm_swap_anon_use, 0, "");
187 SYSCTL_INT(_vm, OID_AUTO, swap_size,
188         CTLFLAG_RD, &vm_swap_size, 0, "");
189 SYSCTL_INT(_vm, OID_AUTO, report_swap_allocs,
190         CTLFLAG_RW, &vm_report_swap_allocs, 0, "");
191 
192 vm_zone_t		swap_zone;
193 
194 /*
195  * Red-Black tree for swblock entries
196  *
197  * The caller must hold vm_token
198  */
199 RB_GENERATE2(swblock_rb_tree, swblock, swb_entry, rb_swblock_compare,
200 	     vm_pindex_t, swb_index);
201 
202 int
203 rb_swblock_compare(struct swblock *swb1, struct swblock *swb2)
204 {
205 	if (swb1->swb_index < swb2->swb_index)
206 		return(-1);
207 	if (swb1->swb_index > swb2->swb_index)
208 		return(1);
209 	return(0);
210 }
211 
212 static
213 int
214 rb_swblock_scancmp(struct swblock *swb, void *data)
215 {
216 	struct swfreeinfo *info = data;
217 
218 	if (swb->swb_index < info->basei)
219 		return(-1);
220 	if (swb->swb_index > info->endi)
221 		return(1);
222 	return(0);
223 }
224 
225 static
226 int
227 rb_swblock_condcmp(struct swblock *swb, void *data)
228 {
229 	struct swfreeinfo *info = data;
230 
231 	if (swb->swb_index < info->basei)
232 		return(-1);
233 	return(0);
234 }
235 
236 /*
237  * pagerops for OBJT_SWAP - "swap pager".  Some ops are also global procedure
238  * calls hooked from other parts of the VM system and do not appear here.
239  * (see vm/swap_pager.h).
240  */
241 
242 static void	swap_pager_dealloc (vm_object_t object);
243 static int	swap_pager_getpage (vm_object_t, vm_page_t *, int);
244 static void	swap_chain_iodone(struct bio *biox);
245 
246 struct pagerops swappagerops = {
247 	swap_pager_dealloc,	/* deallocate an OBJT_SWAP object	*/
248 	swap_pager_getpage,	/* pagein				*/
249 	swap_pager_putpages,	/* pageout				*/
250 	swap_pager_haspage	/* get backing store status for page	*/
251 };
252 
253 /*
254  * dmmax is in page-sized chunks with the new swap system.  It was
255  * dev-bsized chunks in the old.  dmmax is always a power of 2.
256  *
257  * swap_*() routines are externally accessible.  swp_*() routines are
258  * internal.
259  */
260 
261 int dmmax;
262 static int dmmax_mask;
263 int nswap_lowat = 128;		/* in pages, swap_pager_almost_full warn */
264 int nswap_hiwat = 512;		/* in pages, swap_pager_almost_full warn */
265 
266 static __inline void	swp_sizecheck (void);
267 static void	swp_pager_async_iodone (struct bio *bio);
268 
269 /*
270  * Swap bitmap functions
271  */
272 
273 static __inline void	swp_pager_freeswapspace(vm_object_t object,
274 						swblk_t blk, int npages);
275 static __inline swblk_t	swp_pager_getswapspace(vm_object_t object, int npages);
276 
277 /*
278  * Metadata functions
279  */
280 
281 static void swp_pager_meta_convert(vm_object_t);
282 static void swp_pager_meta_build(vm_object_t, vm_pindex_t, swblk_t);
283 static void swp_pager_meta_free(vm_object_t, vm_pindex_t, vm_pindex_t);
284 static void swp_pager_meta_free_all(vm_object_t);
285 static swblk_t swp_pager_meta_ctl(vm_object_t, vm_pindex_t, int);
286 
287 /*
288  * SWP_SIZECHECK() -	update swap_pager_full indication
289  *
290  *	update the swap_pager_almost_full indication and warn when we are
291  *	about to run out of swap space, using lowat/hiwat hysteresis.
292  *
293  *	Clear swap_pager_full ( task killing ) indication when lowat is met.
294  *
295  * No restrictions on call
296  * This routine may not block.
297  * SMP races are ok.
298  */
299 static __inline void
300 swp_sizecheck(void)
301 {
302 	if (vm_swap_size < nswap_lowat) {
303 		if (swap_pager_almost_full == 0) {
304 			kprintf("swap_pager: out of swap space\n");
305 			swap_pager_almost_full = 1;
306 		}
307 	} else {
308 		swap_pager_full = 0;
309 		if (vm_swap_size > nswap_hiwat)
310 			swap_pager_almost_full = 0;
311 	}
312 }
313 
314 /*
315  * SWAP_PAGER_INIT() -	initialize the swap pager!
316  *
317  *	Expected to be started from system init.  NOTE:  This code is run
318  *	before much else so be careful what you depend on.  Most of the VM
319  *	system has yet to be initialized at this point.
320  *
321  * Called from the low level boot code only.
322  */
323 static void
324 swap_pager_init(void *arg __unused)
325 {
326 	/*
327 	 * Device Stripe, in PAGE_SIZE'd blocks
328 	 */
329 	dmmax = SWB_NPAGES * 2;
330 	dmmax_mask = ~(dmmax - 1);
331 }
332 SYSINIT(vm_mem, SI_BOOT1_VM, SI_ORDER_THIRD, swap_pager_init, NULL)
333 
334 /*
335  * SWAP_PAGER_SWAP_INIT() - swap pager initialization from pageout process
336  *
337  *	Expected to be started from pageout process once, prior to entering
338  *	its main loop.
339  *
340  * Called from the low level boot code only.
341  */
342 void
343 swap_pager_swap_init(void)
344 {
345 	int n, n2;
346 
347 	/*
348 	 * Number of in-transit swap bp operations.  Don't
349 	 * exhaust the pbufs completely.  Make sure we
350 	 * initialize workable values (0 will work for hysteresis
351 	 * but it isn't very efficient).
352 	 *
353 	 * The nsw_cluster_max is constrained by the number of pages an XIO
354 	 * holds, i.e., (MAXPHYS/PAGE_SIZE) and our locally defined
355 	 * MAX_PAGEOUT_CLUSTER.   Also be aware that swap ops are
356 	 * constrained by the swap device interleave stripe size.
357 	 *
358 	 * Currently we hardwire nsw_wcount_async to 4.  This limit is
359 	 * designed to prevent other I/O from having high latencies due to
360 	 * our pageout I/O.  The value 4 works well for one or two active swap
361 	 * devices but is probably a little low if you have more.  Even so,
362 	 * a higher value would probably generate only a limited improvement
363 	 * with three or four active swap devices since the system does not
364 	 * typically have to pageout at extreme bandwidths.   We will want
365 	 * at least 2 per swap devices, and 4 is a pretty good value if you
366 	 * have one NFS swap device due to the command/ack latency over NFS.
367 	 * So it all works out pretty well.
368 	 */
369 
370 	nsw_cluster_max = min((MAXPHYS/PAGE_SIZE), MAX_PAGEOUT_CLUSTER);
371 
372 	nsw_rcount = (nswbuf + 1) / 2;
373 	nsw_wcount_sync = (nswbuf + 3) / 4;
374 	nsw_wcount_async = 4;
375 	nsw_wcount_async_max = nsw_wcount_async;
376 
377 	/*
378 	 * The zone is dynamically allocated so generally size it to
379 	 * maxswzone (32MB to 512MB of KVM).  Set a minimum size based
380 	 * on physical memory of around 8x (each swblock can hold 16 pages).
381 	 *
382 	 * With the advent of SSDs (vs HDs) the practical (swap:memory) ratio
383 	 * has increased dramatically.
384 	 */
385 	n = vmstats.v_page_count / 2;
386 	if (maxswzone && n < maxswzone / sizeof(struct swblock))
387 		n = maxswzone / sizeof(struct swblock);
388 	n2 = n;
389 
390 	do {
391 		swap_zone = zinit(
392 			"SWAPMETA",
393 			sizeof(struct swblock),
394 			n,
395 			ZONE_INTERRUPT,
396 			1);
397 		if (swap_zone != NULL)
398 			break;
399 		/*
400 		 * if the allocation failed, try a zone two thirds the
401 		 * size of the previous attempt.
402 		 */
403 		n -= ((n + 2) / 3);
404 	} while (n > 0);
405 
406 	if (swap_zone == NULL)
407 		panic("swap_pager_swap_init: swap_zone == NULL");
408 	if (n2 != n)
409 		kprintf("Swap zone entries reduced from %d to %d.\n", n2, n);
410 }
411 
412 /*
413  * SWAP_PAGER_ALLOC() -	allocate a new OBJT_SWAP VM object and instantiate
414  *			its metadata structures.
415  *
416  *	This routine is called from the mmap and fork code to create a new
417  *	OBJT_SWAP object.  We do this by creating an OBJT_DEFAULT object
418  *	and then converting it with swp_pager_meta_convert().
419  *
420  *	We only support unnamed objects.
421  *
422  * No restrictions.
423  */
424 vm_object_t
425 swap_pager_alloc(void *handle, off_t size, vm_prot_t prot, off_t offset)
426 {
427 	vm_object_t object;
428 
429 	KKASSERT(handle == NULL);
430 	object = vm_object_allocate_hold(OBJT_DEFAULT,
431 					 OFF_TO_IDX(offset + PAGE_MASK + size));
432 	swp_pager_meta_convert(object);
433 	vm_object_drop(object);
434 
435 	return (object);
436 }
437 
438 /*
439  * SWAP_PAGER_DEALLOC() -	remove swap metadata from object
440  *
441  *	The swap backing for the object is destroyed.  The code is
442  *	designed such that we can reinstantiate it later, but this
443  *	routine is typically called only when the entire object is
444  *	about to be destroyed.
445  *
446  * The object must be locked or unreferenceable.
447  * No other requirements.
448  */
449 static void
450 swap_pager_dealloc(vm_object_t object)
451 {
452 	vm_object_hold(object);
453 	vm_object_pip_wait(object, "swpdea");
454 
455 	/*
456 	 * Free all remaining metadata.  We only bother to free it from
457 	 * the swap meta data.  We do not attempt to free swapblk's still
458 	 * associated with vm_page_t's for this object.  We do not care
459 	 * if paging is still in progress on some objects.
460 	 */
461 	swp_pager_meta_free_all(object);
462 	vm_object_drop(object);
463 }
464 
465 /************************************************************************
466  *			SWAP PAGER BITMAP ROUTINES			*
467  ************************************************************************/
468 
469 /*
470  * SWP_PAGER_GETSWAPSPACE() -	allocate raw swap space
471  *
472  *	Allocate swap for the requested number of pages.  The starting
473  *	swap block number (a page index) is returned or SWAPBLK_NONE
474  *	if the allocation failed.
475  *
476  *	Also has the side effect of advising that somebody made a mistake
477  *	when they configured swap and didn't configure enough.
478  *
479  * The caller must hold the object.
480  * This routine may not block.
481  */
482 static __inline swblk_t
483 swp_pager_getswapspace(vm_object_t object, int npages)
484 {
485 	swblk_t blk;
486 
487 	lwkt_gettoken(&vm_token);
488 	blk = blist_allocat(swapblist, npages, swapiterator);
489 	if (blk == SWAPBLK_NONE)
490 		blk = blist_allocat(swapblist, npages, 0);
491 	if (blk == SWAPBLK_NONE) {
492 		if (swap_pager_full != 2) {
493 			kprintf("swap_pager_getswapspace: failed alloc=%d\n",
494 				npages);
495 			swap_pager_full = 2;
496 			swap_pager_almost_full = 1;
497 		}
498 	} else {
499 		/* swapiterator = blk; disable for now, doesn't work well */
500 		swapacctspace(blk, -npages);
501 		if (object->type == OBJT_SWAP)
502 			vm_swap_anon_use += npages;
503 		else
504 			vm_swap_cache_use += npages;
505 		swp_sizecheck();
506 	}
507 	lwkt_reltoken(&vm_token);
508 	return(blk);
509 }
510 
511 /*
512  * SWP_PAGER_FREESWAPSPACE() -	free raw swap space
513  *
514  *	This routine returns the specified swap blocks back to the bitmap.
515  *
516  *	Note:  This routine may not block (it could in the old swap code),
517  *	and through the use of the new blist routines it does not block.
518  *
519  *	We must be called at splvm() to avoid races with bitmap frees from
520  *	vm_page_remove() aka swap_pager_page_removed().
521  *
522  * This routine may not block.
523  */
524 
525 static __inline void
526 swp_pager_freeswapspace(vm_object_t object, swblk_t blk, int npages)
527 {
528 	struct swdevt *sp = &swdevt[BLK2DEVIDX(blk)];
529 
530 	lwkt_gettoken(&vm_token);
531 	sp->sw_nused -= npages;
532 	if (object->type == OBJT_SWAP)
533 		vm_swap_anon_use -= npages;
534 	else
535 		vm_swap_cache_use -= npages;
536 
537 	if (sp->sw_flags & SW_CLOSING) {
538 		lwkt_reltoken(&vm_token);
539 		return;
540 	}
541 
542 	blist_free(swapblist, blk, npages);
543 	vm_swap_size += npages;
544 	swp_sizecheck();
545 	lwkt_reltoken(&vm_token);
546 }
547 
548 /*
549  * SWAP_PAGER_FREESPACE() -	frees swap blocks associated with a page
550  *				range within an object.
551  *
552  *	This is a globally accessible routine.
553  *
554  *	This routine removes swapblk assignments from swap metadata.
555  *
556  *	The external callers of this routine typically have already destroyed
557  *	or renamed vm_page_t's associated with this range in the object so
558  *	we should be ok.
559  *
560  * No requirements.
561  */
562 void
563 swap_pager_freespace(vm_object_t object, vm_pindex_t start, vm_pindex_t size)
564 {
565 	vm_object_hold(object);
566 	swp_pager_meta_free(object, start, size);
567 	vm_object_drop(object);
568 }
569 
570 /*
571  * No requirements.
572  */
573 void
574 swap_pager_freespace_all(vm_object_t object)
575 {
576 	vm_object_hold(object);
577 	swp_pager_meta_free_all(object);
578 	vm_object_drop(object);
579 }
580 
581 /*
582  * This function conditionally frees swap cache swap starting at
583  * (*basei) in the object.  (count) swap blocks will be nominally freed.
584  * The actual number of blocks freed can be more or less than the
585  * requested number.
586  *
587  * This function nominally returns the number of blocks freed.  However,
588  * the actual number of blocks freed may be less then the returned value.
589  * If the function is unable to exhaust the object or if it is able to
590  * free (approximately) the requested number of blocks it returns
591  * a value n > count.
592  *
593  * If we exhaust the object we will return a value n <= count.
594  *
595  * The caller must hold the object.
596  *
597  * WARNING!  If count == 0 then -1 can be returned as a degenerate case,
598  *	     callers should always pass a count value > 0.
599  */
600 static int swap_pager_condfree_callback(struct swblock *swap, void *data);
601 
602 int
603 swap_pager_condfree(vm_object_t object, vm_pindex_t *basei, int count)
604 {
605 	struct swfreeinfo info;
606 	int n;
607 	int t;
608 
609 	ASSERT_LWKT_TOKEN_HELD(vm_object_token(object));
610 
611 	info.object = object;
612 	info.basei = *basei;	/* skip up to this page index */
613 	info.begi = count;	/* max swap pages to destroy */
614 	info.endi = count * 8;	/* max swblocks to scan */
615 
616 	swblock_rb_tree_RB_SCAN(&object->swblock_root, rb_swblock_condcmp,
617 				swap_pager_condfree_callback, &info);
618 	*basei = info.basei;
619 
620 	/*
621 	 * Take the higher difference swblocks vs pages
622 	 */
623 	n = count - (int)info.begi;
624 	t = count * 8 - (int)info.endi;
625 	if (n < t)
626 		n = t;
627 	if (n < 1)
628 		n = 1;
629 	return(n);
630 }
631 
632 /*
633  * The idea is to free whole meta-block to avoid fragmenting
634  * the swap space or disk I/O.  We only do this if NO VM pages
635  * are present.
636  *
637  * We do not have to deal with clearing PG_SWAPPED in related VM
638  * pages because there are no related VM pages.
639  *
640  * The caller must hold the object.
641  */
642 static int
643 swap_pager_condfree_callback(struct swblock *swap, void *data)
644 {
645 	struct swfreeinfo *info = data;
646 	vm_object_t object = info->object;
647 	int i;
648 
649 	for (i = 0; i < SWAP_META_PAGES; ++i) {
650 		if (vm_page_lookup(object, swap->swb_index + i))
651 			break;
652 	}
653 	info->basei = swap->swb_index + SWAP_META_PAGES;
654 	if (i == SWAP_META_PAGES) {
655 		info->begi -= swap->swb_count;
656 		swap_pager_freespace(object, swap->swb_index, SWAP_META_PAGES);
657 	}
658 	--info->endi;
659 	if ((int)info->begi < 0 || (int)info->endi < 0)
660 		return(-1);
661 	lwkt_yield();
662 	return(0);
663 }
664 
665 /*
666  * Called by vm_page_alloc() when a new VM page is inserted
667  * into a VM object.  Checks whether swap has been assigned to
668  * the page and sets PG_SWAPPED as necessary.
669  *
670  * No requirements.
671  */
672 void
673 swap_pager_page_inserted(vm_page_t m)
674 {
675 	if (m->object->swblock_count) {
676 		vm_object_hold(m->object);
677 		if (swp_pager_meta_ctl(m->object, m->pindex, 0) != SWAPBLK_NONE)
678 			vm_page_flag_set(m, PG_SWAPPED);
679 		vm_object_drop(m->object);
680 	}
681 }
682 
683 /*
684  * SWAP_PAGER_RESERVE() - reserve swap blocks in object
685  *
686  *	Assigns swap blocks to the specified range within the object.  The
687  *	swap blocks are not zerod.  Any previous swap assignment is destroyed.
688  *
689  *	Returns 0 on success, -1 on failure.
690  *
691  * The caller is responsible for avoiding races in the specified range.
692  * No other requirements.
693  */
694 int
695 swap_pager_reserve(vm_object_t object, vm_pindex_t start, vm_size_t size)
696 {
697 	int n = 0;
698 	swblk_t blk = SWAPBLK_NONE;
699 	vm_pindex_t beg = start;	/* save start index */
700 
701 	vm_object_hold(object);
702 
703 	while (size) {
704 		if (n == 0) {
705 			n = BLIST_MAX_ALLOC;
706 			while ((blk = swp_pager_getswapspace(object, n)) ==
707 			       SWAPBLK_NONE)
708 			{
709 				n >>= 1;
710 				if (n == 0) {
711 					swp_pager_meta_free(object, beg,
712 							    start - beg);
713 					vm_object_drop(object);
714 					return(-1);
715 				}
716 			}
717 		}
718 		swp_pager_meta_build(object, start, blk);
719 		--size;
720 		++start;
721 		++blk;
722 		--n;
723 	}
724 	swp_pager_meta_free(object, start, n);
725 	vm_object_drop(object);
726 	return(0);
727 }
728 
729 /*
730  * SWAP_PAGER_COPY() -  copy blocks from source pager to destination pager
731  *			and destroy the source.
732  *
733  *	Copy any valid swapblks from the source to the destination.  In
734  *	cases where both the source and destination have a valid swapblk,
735  *	we keep the destination's.
736  *
737  *	This routine is allowed to block.  It may block allocating metadata
738  *	indirectly through swp_pager_meta_build() or if paging is still in
739  *	progress on the source.
740  *
741  *	XXX vm_page_collapse() kinda expects us not to block because we
742  *	supposedly do not need to allocate memory, but for the moment we
743  *	*may* have to get a little memory from the zone allocator, but
744  *	it is taken from the interrupt memory.  We should be ok.
745  *
746  *	The source object contains no vm_page_t's (which is just as well)
747  *	The source object is of type OBJT_SWAP.
748  *
749  *	The source and destination objects must be held by the caller.
750  */
751 void
752 swap_pager_copy(vm_object_t srcobject, vm_object_t dstobject,
753 		vm_pindex_t base_index, int destroysource)
754 {
755 	vm_pindex_t i;
756 
757 	ASSERT_LWKT_TOKEN_HELD(vm_object_token(srcobject));
758 	ASSERT_LWKT_TOKEN_HELD(vm_object_token(dstobject));
759 
760 	/*
761 	 * transfer source to destination.
762 	 */
763 	for (i = 0; i < dstobject->size; ++i) {
764 		swblk_t dstaddr;
765 
766 		/*
767 		 * Locate (without changing) the swapblk on the destination,
768 		 * unless it is invalid in which case free it silently, or
769 		 * if the destination is a resident page, in which case the
770 		 * source is thrown away.
771 		 */
772 		dstaddr = swp_pager_meta_ctl(dstobject, i, 0);
773 
774 		if (dstaddr == SWAPBLK_NONE) {
775 			/*
776 			 * Destination has no swapblk and is not resident,
777 			 * copy source.
778 			 */
779 			swblk_t srcaddr;
780 
781 			srcaddr = swp_pager_meta_ctl(srcobject,
782 						     base_index + i, SWM_POP);
783 
784 			if (srcaddr != SWAPBLK_NONE)
785 				swp_pager_meta_build(dstobject, i, srcaddr);
786 		} else {
787 			/*
788 			 * Destination has valid swapblk or it is represented
789 			 * by a resident page.  We destroy the sourceblock.
790 			 */
791 			swp_pager_meta_ctl(srcobject, base_index + i, SWM_FREE);
792 		}
793 	}
794 
795 	/*
796 	 * Free left over swap blocks in source.
797 	 *
798 	 * We have to revert the type to OBJT_DEFAULT so we do not accidently
799 	 * double-remove the object from the swap queues.
800 	 */
801 	if (destroysource) {
802 		/*
803 		 * Reverting the type is not necessary, the caller is going
804 		 * to destroy srcobject directly, but I'm doing it here
805 		 * for consistency since we've removed the object from its
806 		 * queues.
807 		 */
808 		swp_pager_meta_free_all(srcobject);
809 		if (srcobject->type == OBJT_SWAP)
810 			srcobject->type = OBJT_DEFAULT;
811 	}
812 }
813 
814 /*
815  * SWAP_PAGER_HASPAGE() -	determine if we have good backing store for
816  *				the requested page.
817  *
818  *	We determine whether good backing store exists for the requested
819  *	page and return TRUE if it does, FALSE if it doesn't.
820  *
821  *	If TRUE, we also try to determine how much valid, contiguous backing
822  *	store exists before and after the requested page within a reasonable
823  *	distance.  We do not try to restrict it to the swap device stripe
824  *	(that is handled in getpages/putpages).  It probably isn't worth
825  *	doing here.
826  *
827  * No requirements.
828  */
829 boolean_t
830 swap_pager_haspage(vm_object_t object, vm_pindex_t pindex)
831 {
832 	swblk_t blk0;
833 
834 	/*
835 	 * do we have good backing store at the requested index ?
836 	 */
837 	vm_object_hold(object);
838 	blk0 = swp_pager_meta_ctl(object, pindex, 0);
839 
840 	if (blk0 == SWAPBLK_NONE) {
841 		vm_object_drop(object);
842 		return (FALSE);
843 	}
844 	vm_object_drop(object);
845 	return (TRUE);
846 }
847 
848 /*
849  * SWAP_PAGER_PAGE_UNSWAPPED() - remove swap backing store related to page
850  *
851  * This removes any associated swap backing store, whether valid or
852  * not, from the page.  This operates on any VM object, not just OBJT_SWAP
853  * objects.
854  *
855  * This routine is typically called when a page is made dirty, at
856  * which point any associated swap can be freed.  MADV_FREE also
857  * calls us in a special-case situation
858  *
859  * NOTE!!!  If the page is clean and the swap was valid, the caller
860  * should make the page dirty before calling this routine.  This routine
861  * does NOT change the m->dirty status of the page.  Also: MADV_FREE
862  * depends on it.
863  *
864  * The page must be busied or soft-busied.
865  * The caller can hold the object to avoid blocking, else we might block.
866  * No other requirements.
867  */
868 void
869 swap_pager_unswapped(vm_page_t m)
870 {
871 	if (m->flags & PG_SWAPPED) {
872 		vm_object_hold(m->object);
873 		KKASSERT(m->flags & PG_SWAPPED);
874 		swp_pager_meta_ctl(m->object, m->pindex, SWM_FREE);
875 		vm_page_flag_clear(m, PG_SWAPPED);
876 		vm_object_drop(m->object);
877 	}
878 }
879 
880 /*
881  * SWAP_PAGER_STRATEGY() - read, write, free blocks
882  *
883  * This implements a VM OBJECT strategy function using swap backing store.
884  * This can operate on any VM OBJECT type, not necessarily just OBJT_SWAP
885  * types.
886  *
887  * This is intended to be a cacheless interface (i.e. caching occurs at
888  * higher levels), and is also used as a swap-based SSD cache for vnode
889  * and device objects.
890  *
891  * All I/O goes directly to and from the swap device.
892  *
893  * We currently attempt to run I/O synchronously or asynchronously as
894  * the caller requests.  This isn't perfect because we loose error
895  * sequencing when we run multiple ops in parallel to satisfy a request.
896  * But this is swap, so we let it all hang out.
897  *
898  * No requirements.
899  */
900 void
901 swap_pager_strategy(vm_object_t object, struct bio *bio)
902 {
903 	struct buf *bp = bio->bio_buf;
904 	struct bio *nbio;
905 	vm_pindex_t start;
906 	vm_pindex_t biox_blkno = 0;
907 	int count;
908 	char *data;
909 	struct bio *biox;
910 	struct buf *bufx;
911 #if 0
912 	struct bio_track *track;
913 #endif
914 
915 #if 0
916 	/*
917 	 * tracking for swapdev vnode I/Os
918 	 */
919 	if (bp->b_cmd == BUF_CMD_READ)
920 		track = &swapdev_vp->v_track_read;
921 	else
922 		track = &swapdev_vp->v_track_write;
923 #endif
924 
925 	if (bp->b_bcount & PAGE_MASK) {
926 		bp->b_error = EINVAL;
927 		bp->b_flags |= B_ERROR | B_INVAL;
928 		biodone(bio);
929 		kprintf("swap_pager_strategy: bp %p offset %lld size %d, "
930 			"not page bounded\n",
931 			bp, (long long)bio->bio_offset, (int)bp->b_bcount);
932 		return;
933 	}
934 
935 	/*
936 	 * Clear error indication, initialize page index, count, data pointer.
937 	 */
938 	bp->b_error = 0;
939 	bp->b_flags &= ~B_ERROR;
940 	bp->b_resid = bp->b_bcount;
941 
942 	start = (vm_pindex_t)(bio->bio_offset >> PAGE_SHIFT);
943 	count = howmany(bp->b_bcount, PAGE_SIZE);
944 	data = bp->b_data;
945 
946 	/*
947 	 * Deal with BUF_CMD_FREEBLKS
948 	 */
949 	if (bp->b_cmd == BUF_CMD_FREEBLKS) {
950 		/*
951 		 * FREE PAGE(s) - destroy underlying swap that is no longer
952 		 *		  needed.
953 		 */
954 		vm_object_hold(object);
955 		swp_pager_meta_free(object, start, count);
956 		vm_object_drop(object);
957 		bp->b_resid = 0;
958 		biodone(bio);
959 		return;
960 	}
961 
962 	/*
963 	 * We need to be able to create a new cluster of I/O's.  We cannot
964 	 * use the caller fields of the passed bio so push a new one.
965 	 *
966 	 * Because nbio is just a placeholder for the cluster links,
967 	 * we can biodone() the original bio instead of nbio to make
968 	 * things a bit more efficient.
969 	 */
970 	nbio = push_bio(bio);
971 	nbio->bio_offset = bio->bio_offset;
972 	nbio->bio_caller_info1.cluster_head = NULL;
973 	nbio->bio_caller_info2.cluster_tail = NULL;
974 
975 	biox = NULL;
976 	bufx = NULL;
977 
978 	/*
979 	 * Execute read or write
980 	 */
981 	vm_object_hold(object);
982 
983 	while (count > 0) {
984 		swblk_t blk;
985 
986 		/*
987 		 * Obtain block.  If block not found and writing, allocate a
988 		 * new block and build it into the object.
989 		 */
990 		blk = swp_pager_meta_ctl(object, start, 0);
991 		if ((blk == SWAPBLK_NONE) && bp->b_cmd != BUF_CMD_READ) {
992 			blk = swp_pager_getswapspace(object, 1);
993 			if (blk == SWAPBLK_NONE) {
994 				bp->b_error = ENOMEM;
995 				bp->b_flags |= B_ERROR;
996 				break;
997 			}
998 			swp_pager_meta_build(object, start, blk);
999 		}
1000 
1001 		/*
1002 		 * Do we have to flush our current collection?  Yes if:
1003 		 *
1004 		 *	- no swap block at this index
1005 		 *	- swap block is not contiguous
1006 		 *	- we cross a physical disk boundry in the
1007 		 *	  stripe.
1008 		 */
1009 		if (
1010 		    biox && (biox_blkno + btoc(bufx->b_bcount) != blk ||
1011 		     ((biox_blkno ^ blk) & dmmax_mask)
1012 		    )
1013 		) {
1014 			if (bp->b_cmd == BUF_CMD_READ) {
1015 				++mycpu->gd_cnt.v_swapin;
1016 				mycpu->gd_cnt.v_swappgsin += btoc(bufx->b_bcount);
1017 			} else {
1018 				++mycpu->gd_cnt.v_swapout;
1019 				mycpu->gd_cnt.v_swappgsout += btoc(bufx->b_bcount);
1020 				bufx->b_dirtyend = bufx->b_bcount;
1021 			}
1022 
1023 			/*
1024 			 * Finished with this buf.
1025 			 */
1026 			KKASSERT(bufx->b_bcount != 0);
1027 			if (bufx->b_cmd != BUF_CMD_READ)
1028 				bufx->b_dirtyend = bufx->b_bcount;
1029 			biox = NULL;
1030 			bufx = NULL;
1031 		}
1032 
1033 		/*
1034 		 * Add new swapblk to biox, instantiating biox if necessary.
1035 		 * Zero-fill reads are able to take a shortcut.
1036 		 */
1037 		if (blk == SWAPBLK_NONE) {
1038 			/*
1039 			 * We can only get here if we are reading.  Since
1040 			 * we are at splvm() we can safely modify b_resid,
1041 			 * even if chain ops are in progress.
1042 			 */
1043 			bzero(data, PAGE_SIZE);
1044 			bp->b_resid -= PAGE_SIZE;
1045 		} else {
1046 			if (biox == NULL) {
1047 				/* XXX chain count > 4, wait to <= 4 */
1048 
1049 				bufx = getpbuf(NULL);
1050 				biox = &bufx->b_bio1;
1051 				cluster_append(nbio, bufx);
1052 				bufx->b_flags |= (bp->b_flags & B_ORDERED);
1053 				bufx->b_cmd = bp->b_cmd;
1054 				biox->bio_done = swap_chain_iodone;
1055 				biox->bio_offset = (off_t)blk << PAGE_SHIFT;
1056 				biox->bio_caller_info1.cluster_parent = nbio;
1057 				biox_blkno = blk;
1058 				bufx->b_bcount = 0;
1059 				bufx->b_data = data;
1060 			}
1061 			bufx->b_bcount += PAGE_SIZE;
1062 		}
1063 		--count;
1064 		++start;
1065 		data += PAGE_SIZE;
1066 	}
1067 
1068 	vm_object_drop(object);
1069 
1070 	/*
1071 	 *  Flush out last buffer
1072 	 */
1073 	if (biox) {
1074 		if (bufx->b_cmd == BUF_CMD_READ) {
1075 			++mycpu->gd_cnt.v_swapin;
1076 			mycpu->gd_cnt.v_swappgsin += btoc(bufx->b_bcount);
1077 		} else {
1078 			++mycpu->gd_cnt.v_swapout;
1079 			mycpu->gd_cnt.v_swappgsout += btoc(bufx->b_bcount);
1080 			bufx->b_dirtyend = bufx->b_bcount;
1081 		}
1082 		KKASSERT(bufx->b_bcount);
1083 		if (bufx->b_cmd != BUF_CMD_READ)
1084 			bufx->b_dirtyend = bufx->b_bcount;
1085 		/* biox, bufx = NULL */
1086 	}
1087 
1088 	/*
1089 	 * Now initiate all the I/O.  Be careful looping on our chain as
1090 	 * I/O's may complete while we are still initiating them.
1091 	 *
1092 	 * If the request is a 100% sparse read no bios will be present
1093 	 * and we just biodone() the buffer.
1094 	 */
1095 	nbio->bio_caller_info2.cluster_tail = NULL;
1096 	bufx = nbio->bio_caller_info1.cluster_head;
1097 
1098 	if (bufx) {
1099 		while (bufx) {
1100 			biox = &bufx->b_bio1;
1101 			BUF_KERNPROC(bufx);
1102 			bufx = bufx->b_cluster_next;
1103 			vn_strategy(swapdev_vp, biox);
1104 		}
1105 	} else {
1106 		biodone(bio);
1107 	}
1108 
1109 	/*
1110 	 * Completion of the cluster will also call biodone_chain(nbio).
1111 	 * We never call biodone(nbio) so we don't have to worry about
1112 	 * setting up a bio_done callback.  It's handled in the sub-IO.
1113 	 */
1114 	/**/
1115 }
1116 
1117 /*
1118  * biodone callback
1119  *
1120  * No requirements.
1121  */
1122 static void
1123 swap_chain_iodone(struct bio *biox)
1124 {
1125 	struct buf **nextp;
1126 	struct buf *bufx;	/* chained sub-buffer */
1127 	struct bio *nbio;	/* parent nbio with chain glue */
1128 	struct buf *bp;		/* original bp associated with nbio */
1129 	int chain_empty;
1130 
1131 	bufx = biox->bio_buf;
1132 	nbio = biox->bio_caller_info1.cluster_parent;
1133 	bp = nbio->bio_buf;
1134 
1135 	/*
1136 	 * Update the original buffer
1137 	 */
1138         KKASSERT(bp != NULL);
1139 	if (bufx->b_flags & B_ERROR) {
1140 		atomic_set_int(&bufx->b_flags, B_ERROR);
1141 		bp->b_error = bufx->b_error;	/* race ok */
1142 	} else if (bufx->b_resid != 0) {
1143 		atomic_set_int(&bufx->b_flags, B_ERROR);
1144 		bp->b_error = EINVAL;		/* race ok */
1145 	} else {
1146 		atomic_subtract_int(&bp->b_resid, bufx->b_bcount);
1147 	}
1148 
1149 	/*
1150 	 * Remove us from the chain.
1151 	 */
1152 	spin_lock(&bp->b_lock.lk_spinlock);
1153 	nextp = &nbio->bio_caller_info1.cluster_head;
1154 	while (*nextp != bufx) {
1155 		KKASSERT(*nextp != NULL);
1156 		nextp = &(*nextp)->b_cluster_next;
1157 	}
1158 	*nextp = bufx->b_cluster_next;
1159 	chain_empty = (nbio->bio_caller_info1.cluster_head == NULL);
1160 	spin_unlock(&bp->b_lock.lk_spinlock);
1161 
1162 	/*
1163 	 * Clean up bufx.  If the chain is now empty we finish out
1164 	 * the parent.  Note that we may be racing other completions
1165 	 * so we must use the chain_empty status from above.
1166 	 */
1167 	if (chain_empty) {
1168 		if (bp->b_resid != 0 && !(bp->b_flags & B_ERROR)) {
1169 			atomic_set_int(&bp->b_flags, B_ERROR);
1170 			bp->b_error = EINVAL;
1171 		}
1172 		biodone_chain(nbio);
1173         }
1174         relpbuf(bufx, NULL);
1175 }
1176 
1177 /*
1178  * SWAP_PAGER_GETPAGES() - bring page in from swap
1179  *
1180  * The requested page may have to be brought in from swap.  Calculate the
1181  * swap block and bring in additional pages if possible.  All pages must
1182  * have contiguous swap block assignments and reside in the same object.
1183  *
1184  * The caller has a single vm_object_pip_add() reference prior to
1185  * calling us and we should return with the same.
1186  *
1187  * The caller has BUSY'd the page.  We should return with (*mpp) left busy,
1188  * and any additinal pages unbusied.
1189  *
1190  * If the caller encounters a PG_RAM page it will pass it to us even though
1191  * it may be valid and dirty.  We cannot overwrite the page in this case!
1192  * The case is used to allow us to issue pure read-aheads.
1193  *
1194  * NOTE! XXX This code does not entirely pipeline yet due to the fact that
1195  *       the PG_RAM page is validated at the same time as mreq.  What we
1196  *	 really need to do is issue a separate read-ahead pbuf.
1197  *
1198  * No requirements.
1199  */
1200 static int
1201 swap_pager_getpage(vm_object_t object, vm_page_t *mpp, int seqaccess)
1202 {
1203 	struct buf *bp;
1204 	struct bio *bio;
1205 	vm_page_t mreq;
1206 	vm_page_t m;
1207 	vm_offset_t kva;
1208 	swblk_t blk;
1209 	int i;
1210 	int j;
1211 	int raonly;
1212 	int error;
1213 	u_int32_t flags;
1214 	vm_page_t marray[XIO_INTERNAL_PAGES];
1215 
1216 	mreq = *mpp;
1217 
1218 	vm_object_hold(object);
1219 	if (mreq->object != object) {
1220 		panic("swap_pager_getpages: object mismatch %p/%p",
1221 		    object,
1222 		    mreq->object
1223 		);
1224 	}
1225 
1226 	/*
1227 	 * We don't want to overwrite a fully valid page as it might be
1228 	 * dirty.  This case can occur when e.g. vm_fault hits a perfectly
1229 	 * valid page with PG_RAM set.
1230 	 *
1231 	 * In this case we see if the next page is a suitable page-in
1232 	 * candidate and if it is we issue read-ahead.  PG_RAM will be
1233 	 * set on the last page of the read-ahead to continue the pipeline.
1234 	 */
1235 	if (mreq->valid == VM_PAGE_BITS_ALL) {
1236 		if (swap_burst_read == 0 || mreq->pindex + 1 >= object->size) {
1237 			vm_object_drop(object);
1238 			return(VM_PAGER_OK);
1239 		}
1240 		blk = swp_pager_meta_ctl(object, mreq->pindex + 1, 0);
1241 		if (blk == SWAPBLK_NONE) {
1242 			vm_object_drop(object);
1243 			return(VM_PAGER_OK);
1244 		}
1245 		m = vm_page_lookup_busy_try(object, mreq->pindex + 1,
1246 					    TRUE, &error);
1247 		if (error) {
1248 			vm_object_drop(object);
1249 			return(VM_PAGER_OK);
1250 		} else if (m == NULL) {
1251 			/*
1252 			 * Use VM_ALLOC_QUICK to avoid blocking on cache
1253 			 * page reuse.
1254 			 */
1255 			m = vm_page_alloc(object, mreq->pindex + 1,
1256 					  VM_ALLOC_QUICK);
1257 			if (m == NULL) {
1258 				vm_object_drop(object);
1259 				return(VM_PAGER_OK);
1260 			}
1261 		} else {
1262 			if (m->valid) {
1263 				vm_page_wakeup(m);
1264 				vm_object_drop(object);
1265 				return(VM_PAGER_OK);
1266 			}
1267 			vm_page_unqueue_nowakeup(m);
1268 		}
1269 		/* page is busy */
1270 		mreq = m;
1271 		raonly = 1;
1272 	} else {
1273 		raonly = 0;
1274 	}
1275 
1276 	/*
1277 	 * Try to block-read contiguous pages from swap if sequential,
1278 	 * otherwise just read one page.  Contiguous pages from swap must
1279 	 * reside within a single device stripe because the I/O cannot be
1280 	 * broken up across multiple stripes.
1281 	 *
1282 	 * Note that blk and iblk can be SWAPBLK_NONE but the loop is
1283 	 * set up such that the case(s) are handled implicitly.
1284 	 */
1285 	blk = swp_pager_meta_ctl(mreq->object, mreq->pindex, 0);
1286 	marray[0] = mreq;
1287 
1288 	for (i = 1; swap_burst_read &&
1289 		    i < XIO_INTERNAL_PAGES &&
1290 		    mreq->pindex + i < object->size; ++i) {
1291 		swblk_t iblk;
1292 
1293 		iblk = swp_pager_meta_ctl(object, mreq->pindex + i, 0);
1294 		if (iblk != blk + i)
1295 			break;
1296 		if ((blk ^ iblk) & dmmax_mask)
1297 			break;
1298 		m = vm_page_lookup_busy_try(object, mreq->pindex + i,
1299 					    TRUE, &error);
1300 		if (error) {
1301 			break;
1302 		} else if (m == NULL) {
1303 			/*
1304 			 * Use VM_ALLOC_QUICK to avoid blocking on cache
1305 			 * page reuse.
1306 			 */
1307 			m = vm_page_alloc(object, mreq->pindex + i,
1308 					  VM_ALLOC_QUICK);
1309 			if (m == NULL)
1310 				break;
1311 		} else {
1312 			if (m->valid) {
1313 				vm_page_wakeup(m);
1314 				break;
1315 			}
1316 			vm_page_unqueue_nowakeup(m);
1317 		}
1318 		/* page is busy */
1319 		marray[i] = m;
1320 	}
1321 	if (i > 1)
1322 		vm_page_flag_set(marray[i - 1], PG_RAM);
1323 
1324 	/*
1325 	 * If mreq is the requested page and we have nothing to do return
1326 	 * VM_PAGER_FAIL.  If raonly is set mreq is just another read-ahead
1327 	 * page and must be cleaned up.
1328 	 */
1329 	if (blk == SWAPBLK_NONE) {
1330 		KKASSERT(i == 1);
1331 		if (raonly) {
1332 			vnode_pager_freepage(mreq);
1333 			vm_object_drop(object);
1334 			return(VM_PAGER_OK);
1335 		} else {
1336 			vm_object_drop(object);
1337 			return(VM_PAGER_FAIL);
1338 		}
1339 	}
1340 
1341 	/*
1342 	 * map our page(s) into kva for input
1343 	 */
1344 	bp = getpbuf_kva(&nsw_rcount);
1345 	bio = &bp->b_bio1;
1346 	kva = (vm_offset_t) bp->b_kvabase;
1347 	bcopy(marray, bp->b_xio.xio_pages, i * sizeof(vm_page_t));
1348 	pmap_qenter(kva, bp->b_xio.xio_pages, i);
1349 
1350 	bp->b_data = (caddr_t)kva;
1351 	bp->b_bcount = PAGE_SIZE * i;
1352 	bp->b_xio.xio_npages = i;
1353 	bio->bio_done = swp_pager_async_iodone;
1354 	bio->bio_offset = (off_t)blk << PAGE_SHIFT;
1355 	bio->bio_caller_info1.index = SWBIO_READ;
1356 
1357 	/*
1358 	 * Set index.  If raonly set the index beyond the array so all
1359 	 * the pages are treated the same, otherwise the original mreq is
1360 	 * at index 0.
1361 	 */
1362 	if (raonly)
1363 		bio->bio_driver_info = (void *)(intptr_t)i;
1364 	else
1365 		bio->bio_driver_info = (void *)(intptr_t)0;
1366 
1367 	for (j = 0; j < i; ++j)
1368 		vm_page_flag_set(bp->b_xio.xio_pages[j], PG_SWAPINPROG);
1369 
1370 	mycpu->gd_cnt.v_swapin++;
1371 	mycpu->gd_cnt.v_swappgsin += bp->b_xio.xio_npages;
1372 
1373 	/*
1374 	 * We still hold the lock on mreq, and our automatic completion routine
1375 	 * does not remove it.
1376 	 */
1377 	vm_object_pip_add(object, bp->b_xio.xio_npages);
1378 
1379 	/*
1380 	 * perform the I/O.  NOTE!!!  bp cannot be considered valid after
1381 	 * this point because we automatically release it on completion.
1382 	 * Instead, we look at the one page we are interested in which we
1383 	 * still hold a lock on even through the I/O completion.
1384 	 *
1385 	 * The other pages in our m[] array are also released on completion,
1386 	 * so we cannot assume they are valid anymore either.
1387 	 */
1388 	bp->b_cmd = BUF_CMD_READ;
1389 	BUF_KERNPROC(bp);
1390 	vn_strategy(swapdev_vp, bio);
1391 
1392 	/*
1393 	 * Wait for the page we want to complete.  PG_SWAPINPROG is always
1394 	 * cleared on completion.  If an I/O error occurs, SWAPBLK_NONE
1395 	 * is set in the meta-data.
1396 	 *
1397 	 * If this is a read-ahead only we return immediately without
1398 	 * waiting for I/O.
1399 	 */
1400 	if (raonly) {
1401 		vm_object_drop(object);
1402 		return(VM_PAGER_OK);
1403 	}
1404 
1405 	/*
1406 	 * Read-ahead includes originally requested page case.
1407 	 */
1408 	for (;;) {
1409 		flags = mreq->flags;
1410 		cpu_ccfence();
1411 		if ((flags & PG_SWAPINPROG) == 0)
1412 			break;
1413 		tsleep_interlock(mreq, 0);
1414 		if (!atomic_cmpset_int(&mreq->flags, flags,
1415 				       flags | PG_WANTED | PG_REFERENCED)) {
1416 			continue;
1417 		}
1418 		mycpu->gd_cnt.v_intrans++;
1419 		if (tsleep(mreq, PINTERLOCKED, "swread", hz*20)) {
1420 			kprintf(
1421 			    "swap_pager: indefinite wait buffer: "
1422 				" offset: %lld, size: %ld\n",
1423 			    (long long)bio->bio_offset,
1424 			    (long)bp->b_bcount
1425 			);
1426 		}
1427 	}
1428 
1429 	/*
1430 	 * mreq is left bussied after completion, but all the other pages
1431 	 * are freed.  If we had an unrecoverable read error the page will
1432 	 * not be valid.
1433 	 */
1434 	vm_object_drop(object);
1435 	if (mreq->valid != VM_PAGE_BITS_ALL)
1436 		return(VM_PAGER_ERROR);
1437 	else
1438 		return(VM_PAGER_OK);
1439 
1440 	/*
1441 	 * A final note: in a low swap situation, we cannot deallocate swap
1442 	 * and mark a page dirty here because the caller is likely to mark
1443 	 * the page clean when we return, causing the page to possibly revert
1444 	 * to all-zero's later.
1445 	 */
1446 }
1447 
1448 /*
1449  *	swap_pager_putpages:
1450  *
1451  *	Assign swap (if necessary) and initiate I/O on the specified pages.
1452  *
1453  *	We support both OBJT_DEFAULT and OBJT_SWAP objects.  DEFAULT objects
1454  *	are automatically converted to SWAP objects.
1455  *
1456  *	In a low memory situation we may block in vn_strategy(), but the new
1457  *	vm_page reservation system coupled with properly written VFS devices
1458  *	should ensure that no low-memory deadlock occurs.  This is an area
1459  *	which needs work.
1460  *
1461  *	The parent has N vm_object_pip_add() references prior to
1462  *	calling us and will remove references for rtvals[] that are
1463  *	not set to VM_PAGER_PEND.  We need to remove the rest on I/O
1464  *	completion.
1465  *
1466  *	The parent has soft-busy'd the pages it passes us and will unbusy
1467  *	those whos rtvals[] entry is not set to VM_PAGER_PEND on return.
1468  *	We need to unbusy the rest on I/O completion.
1469  *
1470  * No requirements.
1471  */
1472 void
1473 swap_pager_putpages(vm_object_t object, vm_page_t *m, int count,
1474 		    boolean_t sync, int *rtvals)
1475 {
1476 	int i;
1477 	int n = 0;
1478 
1479 	vm_object_hold(object);
1480 
1481 	if (count && m[0]->object != object) {
1482 		panic("swap_pager_getpages: object mismatch %p/%p",
1483 		    object,
1484 		    m[0]->object
1485 		);
1486 	}
1487 
1488 	/*
1489 	 * Step 1
1490 	 *
1491 	 * Turn object into OBJT_SWAP
1492 	 * check for bogus sysops
1493 	 * force sync if not pageout process
1494 	 */
1495 	if (object->type == OBJT_DEFAULT) {
1496 		if (object->type == OBJT_DEFAULT)
1497 			swp_pager_meta_convert(object);
1498 	}
1499 
1500 	if (curthread != pagethread)
1501 		sync = TRUE;
1502 
1503 	/*
1504 	 * Step 2
1505 	 *
1506 	 * Update nsw parameters from swap_async_max sysctl values.
1507 	 * Do not let the sysop crash the machine with bogus numbers.
1508 	 */
1509 	if (swap_async_max != nsw_wcount_async_max) {
1510 		int n;
1511 
1512 		/*
1513 		 * limit range
1514 		 */
1515 		if ((n = swap_async_max) > nswbuf / 2)
1516 			n = nswbuf / 2;
1517 		if (n < 1)
1518 			n = 1;
1519 		swap_async_max = n;
1520 
1521 		/*
1522 		 * Adjust difference ( if possible ).  If the current async
1523 		 * count is too low, we may not be able to make the adjustment
1524 		 * at this time.
1525 		 *
1526 		 * vm_token needed for nsw_wcount sleep interlock
1527 		 */
1528 		lwkt_gettoken(&vm_token);
1529 		n -= nsw_wcount_async_max;
1530 		if (nsw_wcount_async + n >= 0) {
1531 			nsw_wcount_async_max += n;
1532 			pbuf_adjcount(&nsw_wcount_async, n);
1533 		}
1534 		lwkt_reltoken(&vm_token);
1535 	}
1536 
1537 	/*
1538 	 * Step 3
1539 	 *
1540 	 * Assign swap blocks and issue I/O.  We reallocate swap on the fly.
1541 	 * The page is left dirty until the pageout operation completes
1542 	 * successfully.
1543 	 */
1544 
1545 	for (i = 0; i < count; i += n) {
1546 		struct buf *bp;
1547 		struct bio *bio;
1548 		swblk_t blk;
1549 		int j;
1550 
1551 		/*
1552 		 * Maximum I/O size is limited by a number of factors.
1553 		 */
1554 
1555 		n = min(BLIST_MAX_ALLOC, count - i);
1556 		n = min(n, nsw_cluster_max);
1557 
1558 		lwkt_gettoken(&vm_token);
1559 
1560 		/*
1561 		 * Get biggest block of swap we can.  If we fail, fall
1562 		 * back and try to allocate a smaller block.  Don't go
1563 		 * overboard trying to allocate space if it would overly
1564 		 * fragment swap.
1565 		 */
1566 		while (
1567 		    (blk = swp_pager_getswapspace(object, n)) == SWAPBLK_NONE &&
1568 		    n > 4
1569 		) {
1570 			n >>= 1;
1571 		}
1572 		if (blk == SWAPBLK_NONE) {
1573 			for (j = 0; j < n; ++j)
1574 				rtvals[i+j] = VM_PAGER_FAIL;
1575 			lwkt_reltoken(&vm_token);
1576 			continue;
1577 		}
1578 		if (vm_report_swap_allocs > 0) {
1579 			kprintf("swap_alloc %08jx,%d\n", (intmax_t)blk, n);
1580 			--vm_report_swap_allocs;
1581 		}
1582 
1583 		/*
1584 		 * The I/O we are constructing cannot cross a physical
1585 		 * disk boundry in the swap stripe.  Note: we are still
1586 		 * at splvm().
1587 		 */
1588 		if ((blk ^ (blk + n)) & dmmax_mask) {
1589 			j = ((blk + dmmax) & dmmax_mask) - blk;
1590 			swp_pager_freeswapspace(object, blk + j, n - j);
1591 			n = j;
1592 		}
1593 
1594 		/*
1595 		 * All I/O parameters have been satisfied, build the I/O
1596 		 * request and assign the swap space.
1597 		 */
1598 		if (sync == TRUE)
1599 			bp = getpbuf_kva(&nsw_wcount_sync);
1600 		else
1601 			bp = getpbuf_kva(&nsw_wcount_async);
1602 		bio = &bp->b_bio1;
1603 
1604 		lwkt_reltoken(&vm_token);
1605 
1606 		pmap_qenter((vm_offset_t)bp->b_data, &m[i], n);
1607 
1608 		bp->b_bcount = PAGE_SIZE * n;
1609 		bio->bio_offset = (off_t)blk << PAGE_SHIFT;
1610 
1611 		for (j = 0; j < n; ++j) {
1612 			vm_page_t mreq = m[i+j];
1613 
1614 			swp_pager_meta_build(mreq->object, mreq->pindex,
1615 					     blk + j);
1616 			if (object->type == OBJT_SWAP)
1617 				vm_page_dirty(mreq);
1618 			rtvals[i+j] = VM_PAGER_OK;
1619 
1620 			vm_page_flag_set(mreq, PG_SWAPINPROG);
1621 			bp->b_xio.xio_pages[j] = mreq;
1622 		}
1623 		bp->b_xio.xio_npages = n;
1624 
1625 		mycpu->gd_cnt.v_swapout++;
1626 		mycpu->gd_cnt.v_swappgsout += bp->b_xio.xio_npages;
1627 
1628 		bp->b_dirtyoff = 0;		/* req'd for NFS */
1629 		bp->b_dirtyend = bp->b_bcount;	/* req'd for NFS */
1630 		bp->b_cmd = BUF_CMD_WRITE;
1631 		bio->bio_caller_info1.index = SWBIO_WRITE;
1632 
1633 		/*
1634 		 * asynchronous
1635 		 */
1636 		if (sync == FALSE) {
1637 			bio->bio_done = swp_pager_async_iodone;
1638 			BUF_KERNPROC(bp);
1639 			vn_strategy(swapdev_vp, bio);
1640 
1641 			for (j = 0; j < n; ++j)
1642 				rtvals[i+j] = VM_PAGER_PEND;
1643 			continue;
1644 		}
1645 
1646 		/*
1647 		 * Issue synchrnously.
1648 		 *
1649 		 * Wait for the sync I/O to complete, then update rtvals.
1650 		 * We just set the rtvals[] to VM_PAGER_PEND so we can call
1651 		 * our async completion routine at the end, thus avoiding a
1652 		 * double-free.
1653 		 */
1654 		bio->bio_caller_info1.index |= SWBIO_SYNC;
1655 		bio->bio_done = biodone_sync;
1656 		bio->bio_flags |= BIO_SYNC;
1657 		vn_strategy(swapdev_vp, bio);
1658 		biowait(bio, "swwrt");
1659 
1660 		for (j = 0; j < n; ++j)
1661 			rtvals[i+j] = VM_PAGER_PEND;
1662 
1663 		/*
1664 		 * Now that we are through with the bp, we can call the
1665 		 * normal async completion, which frees everything up.
1666 		 */
1667 		swp_pager_async_iodone(bio);
1668 	}
1669 	vm_object_drop(object);
1670 }
1671 
1672 /*
1673  * No requirements.
1674  */
1675 void
1676 swap_pager_newswap(void)
1677 {
1678 	swp_sizecheck();
1679 }
1680 
1681 /*
1682  *	swp_pager_async_iodone:
1683  *
1684  *	Completion routine for asynchronous reads and writes from/to swap.
1685  *	Also called manually by synchronous code to finish up a bp.
1686  *
1687  *	For READ operations, the pages are PG_BUSY'd.  For WRITE operations,
1688  *	the pages are vm_page_t->busy'd.  For READ operations, we PG_BUSY
1689  *	unbusy all pages except the 'main' request page.  For WRITE
1690  *	operations, we vm_page_t->busy'd unbusy all pages ( we can do this
1691  *	because we marked them all VM_PAGER_PEND on return from putpages ).
1692  *
1693  *	This routine may not block.
1694  *
1695  * No requirements.
1696  */
1697 static void
1698 swp_pager_async_iodone(struct bio *bio)
1699 {
1700 	struct buf *bp = bio->bio_buf;
1701 	vm_object_t object = NULL;
1702 	int i;
1703 	int *nswptr;
1704 
1705 	/*
1706 	 * report error
1707 	 */
1708 	if (bp->b_flags & B_ERROR) {
1709 		kprintf(
1710 		    "swap_pager: I/O error - %s failed; offset %lld,"
1711 			"size %ld, error %d\n",
1712 		    ((bio->bio_caller_info1.index & SWBIO_READ) ?
1713 			"pagein" : "pageout"),
1714 		    (long long)bio->bio_offset,
1715 		    (long)bp->b_bcount,
1716 		    bp->b_error
1717 		);
1718 	}
1719 
1720 	/*
1721 	 * set object, raise to splvm().
1722 	 */
1723 	if (bp->b_xio.xio_npages)
1724 		object = bp->b_xio.xio_pages[0]->object;
1725 
1726 	/*
1727 	 * remove the mapping for kernel virtual
1728 	 */
1729 	pmap_qremove((vm_offset_t)bp->b_data, bp->b_xio.xio_npages);
1730 
1731 	/*
1732 	 * cleanup pages.  If an error occurs writing to swap, we are in
1733 	 * very serious trouble.  If it happens to be a disk error, though,
1734 	 * we may be able to recover by reassigning the swap later on.  So
1735 	 * in this case we remove the m->swapblk assignment for the page
1736 	 * but do not free it in the rlist.  The errornous block(s) are thus
1737 	 * never reallocated as swap.  Redirty the page and continue.
1738 	 */
1739 	for (i = 0; i < bp->b_xio.xio_npages; ++i) {
1740 		vm_page_t m = bp->b_xio.xio_pages[i];
1741 
1742 		if (bp->b_flags & B_ERROR) {
1743 			/*
1744 			 * If an error occurs I'd love to throw the swapblk
1745 			 * away without freeing it back to swapspace, so it
1746 			 * can never be used again.  But I can't from an
1747 			 * interrupt.
1748 			 */
1749 
1750 			if (bio->bio_caller_info1.index & SWBIO_READ) {
1751 				/*
1752 				 * When reading, reqpage needs to stay
1753 				 * locked for the parent, but all other
1754 				 * pages can be freed.  We still want to
1755 				 * wakeup the parent waiting on the page,
1756 				 * though.  ( also: pg_reqpage can be -1 and
1757 				 * not match anything ).
1758 				 *
1759 				 * We have to wake specifically requested pages
1760 				 * up too because we cleared PG_SWAPINPROG and
1761 				 * someone may be waiting for that.
1762 				 *
1763 				 * NOTE: for reads, m->dirty will probably
1764 				 * be overridden by the original caller of
1765 				 * getpages so don't play cute tricks here.
1766 				 *
1767 				 * NOTE: We can't actually free the page from
1768 				 * here, because this is an interrupt.  It
1769 				 * is not legal to mess with object->memq
1770 				 * from an interrupt.  Deactivate the page
1771 				 * instead.
1772 				 */
1773 
1774 				m->valid = 0;
1775 				vm_page_flag_clear(m, PG_ZERO);
1776 				vm_page_flag_clear(m, PG_SWAPINPROG);
1777 
1778 				/*
1779 				 * bio_driver_info holds the requested page
1780 				 * index.
1781 				 */
1782 				if (i != (int)(intptr_t)bio->bio_driver_info) {
1783 					vm_page_deactivate(m);
1784 					vm_page_wakeup(m);
1785 				} else {
1786 					vm_page_flash(m);
1787 				}
1788 				/*
1789 				 * If i == bp->b_pager.pg_reqpage, do not wake
1790 				 * the page up.  The caller needs to.
1791 				 */
1792 			} else {
1793 				/*
1794 				 * If a write error occurs remove the swap
1795 				 * assignment (note that PG_SWAPPED may or
1796 				 * may not be set depending on prior activity).
1797 				 *
1798 				 * Re-dirty OBJT_SWAP pages as there is no
1799 				 * other backing store, we can't throw the
1800 				 * page away.
1801 				 *
1802 				 * Non-OBJT_SWAP pages (aka swapcache) must
1803 				 * not be dirtied since they may not have
1804 				 * been dirty in the first place, and they
1805 				 * do have backing store (the vnode).
1806 				 */
1807 				vm_page_busy_wait(m, FALSE, "swadpg");
1808 				swp_pager_meta_ctl(m->object, m->pindex,
1809 						   SWM_FREE);
1810 				vm_page_flag_clear(m, PG_SWAPPED);
1811 				if (m->object->type == OBJT_SWAP) {
1812 					vm_page_dirty(m);
1813 					vm_page_activate(m);
1814 				}
1815 				vm_page_flag_clear(m, PG_SWAPINPROG);
1816 				vm_page_io_finish(m);
1817 				vm_page_wakeup(m);
1818 			}
1819 		} else if (bio->bio_caller_info1.index & SWBIO_READ) {
1820 			/*
1821 			 * NOTE: for reads, m->dirty will probably be
1822 			 * overridden by the original caller of getpages so
1823 			 * we cannot set them in order to free the underlying
1824 			 * swap in a low-swap situation.  I don't think we'd
1825 			 * want to do that anyway, but it was an optimization
1826 			 * that existed in the old swapper for a time before
1827 			 * it got ripped out due to precisely this problem.
1828 			 *
1829 			 * clear PG_ZERO in page.
1830 			 *
1831 			 * If not the requested page then deactivate it.
1832 			 *
1833 			 * Note that the requested page, reqpage, is left
1834 			 * busied, but we still have to wake it up.  The
1835 			 * other pages are released (unbusied) by
1836 			 * vm_page_wakeup().  We do not set reqpage's
1837 			 * valid bits here, it is up to the caller.
1838 			 */
1839 
1840 			/*
1841 			 * NOTE: can't call pmap_clear_modify(m) from an
1842 			 * interrupt thread, the pmap code may have to map
1843 			 * non-kernel pmaps and currently asserts the case.
1844 			 */
1845 			/*pmap_clear_modify(m);*/
1846 			m->valid = VM_PAGE_BITS_ALL;
1847 			vm_page_undirty(m);
1848 			vm_page_flag_clear(m, PG_ZERO | PG_SWAPINPROG);
1849 			vm_page_flag_set(m, PG_SWAPPED);
1850 
1851 			/*
1852 			 * We have to wake specifically requested pages
1853 			 * up too because we cleared PG_SWAPINPROG and
1854 			 * could be waiting for it in getpages.  However,
1855 			 * be sure to not unbusy getpages specifically
1856 			 * requested page - getpages expects it to be
1857 			 * left busy.
1858 			 *
1859 			 * bio_driver_info holds the requested page
1860 			 */
1861 			if (i != (int)(intptr_t)bio->bio_driver_info) {
1862 				vm_page_deactivate(m);
1863 				vm_page_wakeup(m);
1864 			} else {
1865 				vm_page_flash(m);
1866 			}
1867 		} else {
1868 			/*
1869 			 * Mark the page clean but do not mess with the
1870 			 * pmap-layer's modified state.  That state should
1871 			 * also be clear since the caller protected the
1872 			 * page VM_PROT_READ, but allow the case.
1873 			 *
1874 			 * We are in an interrupt, avoid pmap operations.
1875 			 *
1876 			 * If we have a severe page deficit, deactivate the
1877 			 * page.  Do not try to cache it (which would also
1878 			 * involve a pmap op), because the page might still
1879 			 * be read-heavy.
1880 			 *
1881 			 * When using the swap to cache clean vnode pages
1882 			 * we do not mess with the page dirty bits.
1883 			 */
1884 			vm_page_busy_wait(m, FALSE, "swadpg");
1885 			if (m->object->type == OBJT_SWAP)
1886 				vm_page_undirty(m);
1887 			vm_page_flag_clear(m, PG_SWAPINPROG);
1888 			vm_page_flag_set(m, PG_SWAPPED);
1889 			if (vm_page_count_severe())
1890 				vm_page_deactivate(m);
1891 #if 0
1892 			if (!vm_page_count_severe() || !vm_page_try_to_cache(m))
1893 				vm_page_protect(m, VM_PROT_READ);
1894 #endif
1895 			vm_page_io_finish(m);
1896 			vm_page_wakeup(m);
1897 		}
1898 	}
1899 
1900 	/*
1901 	 * adjust pip.  NOTE: the original parent may still have its own
1902 	 * pip refs on the object.
1903 	 */
1904 
1905 	if (object)
1906 		vm_object_pip_wakeup_n(object, bp->b_xio.xio_npages);
1907 
1908 	/*
1909 	 * Release the physical I/O buffer.
1910 	 *
1911 	 * NOTE: Due to synchronous operations in the write case b_cmd may
1912 	 *	 already be set to BUF_CMD_DONE and BIO_SYNC may have already
1913 	 *	 been cleared.
1914 	 *
1915 	 * Use vm_token to interlock nsw_rcount/wcount wakeup?
1916 	 */
1917 	lwkt_gettoken(&vm_token);
1918 	if (bio->bio_caller_info1.index & SWBIO_READ)
1919 		nswptr = &nsw_rcount;
1920 	else if (bio->bio_caller_info1.index & SWBIO_SYNC)
1921 		nswptr = &nsw_wcount_sync;
1922 	else
1923 		nswptr = &nsw_wcount_async;
1924 	bp->b_cmd = BUF_CMD_DONE;
1925 	relpbuf(bp, nswptr);
1926 	lwkt_reltoken(&vm_token);
1927 }
1928 
1929 /*
1930  * Fault-in a potentially swapped page and remove the swap reference.
1931  * (used by swapoff code)
1932  *
1933  * object must be held.
1934  */
1935 static __inline void
1936 swp_pager_fault_page(vm_object_t object, int *sharedp, vm_pindex_t pindex)
1937 {
1938 	struct vnode *vp;
1939 	vm_page_t m;
1940 	int error;
1941 
1942 	ASSERT_LWKT_TOKEN_HELD(vm_object_token(object));
1943 
1944 	if (object->type == OBJT_VNODE) {
1945 		/*
1946 		 * Any swap related to a vnode is due to swapcache.  We must
1947 		 * vget() the vnode in case it is not active (otherwise
1948 		 * vref() will panic).  Calling vm_object_page_remove() will
1949 		 * ensure that any swap ref is removed interlocked with the
1950 		 * page.  clean_only is set to TRUE so we don't throw away
1951 		 * dirty pages.
1952 		 */
1953 		vp = object->handle;
1954 		error = vget(vp, LK_SHARED | LK_RETRY | LK_CANRECURSE);
1955 		if (error == 0) {
1956 			vm_object_page_remove(object, pindex, pindex + 1, TRUE);
1957 			vput(vp);
1958 		}
1959 	} else {
1960 		/*
1961 		 * Otherwise it is a normal OBJT_SWAP object and we can
1962 		 * fault the page in and remove the swap.
1963 		 */
1964 		m = vm_fault_object_page(object, IDX_TO_OFF(pindex),
1965 					 VM_PROT_NONE,
1966 					 VM_FAULT_DIRTY | VM_FAULT_UNSWAP,
1967 					 sharedp, &error);
1968 		if (m)
1969 			vm_page_unhold(m);
1970 	}
1971 }
1972 
1973 /*
1974  * This removes all swap blocks related to a particular device.  We have
1975  * to be careful of ripups during the scan.
1976  */
1977 static int swp_pager_swapoff_callback(struct swblock *swap, void *data);
1978 
1979 int
1980 swap_pager_swapoff(int devidx)
1981 {
1982 	struct vm_object marker;
1983 	vm_object_t object;
1984 	struct swswapoffinfo info;
1985 
1986 	bzero(&marker, sizeof(marker));
1987 	marker.type = OBJT_MARKER;
1988 
1989 	lwkt_gettoken(&vmobj_token);
1990 	TAILQ_INSERT_HEAD(&vm_object_list, &marker, object_list);
1991 
1992 	while ((object = TAILQ_NEXT(&marker, object_list)) != NULL) {
1993 		if (object->type == OBJT_MARKER)
1994 			goto skip;
1995 		if (object->type != OBJT_SWAP && object->type != OBJT_VNODE)
1996 			goto skip;
1997 		vm_object_hold(object);
1998 		if (object->type != OBJT_SWAP && object->type != OBJT_VNODE) {
1999 			vm_object_drop(object);
2000 			goto skip;
2001 		}
2002 		info.object = object;
2003 		info.shared = 0;
2004 		info.devidx = devidx;
2005 		swblock_rb_tree_RB_SCAN(&object->swblock_root,
2006 					NULL,
2007 					swp_pager_swapoff_callback,
2008 					&info);
2009 		vm_object_drop(object);
2010 skip:
2011 		if (object == TAILQ_NEXT(&marker, object_list)) {
2012 			TAILQ_REMOVE(&vm_object_list, &marker, object_list);
2013 			TAILQ_INSERT_AFTER(&vm_object_list, object,
2014 					   &marker, object_list);
2015 		}
2016 	}
2017 	TAILQ_REMOVE(&vm_object_list, &marker, object_list);
2018 	lwkt_reltoken(&vmobj_token);
2019 
2020 	/*
2021 	 * If we fail to locate all swblocks we just fail gracefully and
2022 	 * do not bother to restore paging on the swap device.  If the
2023 	 * user wants to retry the user can retry.
2024 	 */
2025 	if (swdevt[devidx].sw_nused)
2026 		return (1);
2027 	else
2028 		return (0);
2029 }
2030 
2031 static
2032 int
2033 swp_pager_swapoff_callback(struct swblock *swap, void *data)
2034 {
2035 	struct swswapoffinfo *info = data;
2036 	vm_object_t object = info->object;
2037 	vm_pindex_t index;
2038 	swblk_t v;
2039 	int i;
2040 
2041 	index = swap->swb_index;
2042 	for (i = 0; i < SWAP_META_PAGES; ++i) {
2043 		/*
2044 		 * Make sure we don't race a dying object.  This will
2045 		 * kill the scan of the object's swap blocks entirely.
2046 		 */
2047 		if (object->flags & OBJ_DEAD)
2048 			return(-1);
2049 
2050 		/*
2051 		 * Fault the page, which can obviously block.  If the swap
2052 		 * structure disappears break out.
2053 		 */
2054 		v = swap->swb_pages[i];
2055 		if (v != SWAPBLK_NONE && BLK2DEVIDX(v) == info->devidx) {
2056 			swp_pager_fault_page(object, &info->shared,
2057 					     swap->swb_index + i);
2058 			/* swap ptr might go away */
2059 			if (RB_LOOKUP(swblock_rb_tree,
2060 				      &object->swblock_root, index) != swap) {
2061 				break;
2062 			}
2063 		}
2064 	}
2065 	return(0);
2066 }
2067 
2068 /************************************************************************
2069  *				SWAP META DATA 				*
2070  ************************************************************************
2071  *
2072  *	These routines manipulate the swap metadata stored in the
2073  *	OBJT_SWAP object.  All swp_*() routines must be called at
2074  *	splvm() because swap can be freed up by the low level vm_page
2075  *	code which might be called from interrupts beyond what splbio() covers.
2076  *
2077  *	Swap metadata is implemented with a global hash and not directly
2078  *	linked into the object.  Instead the object simply contains
2079  *	appropriate tracking counters.
2080  */
2081 
2082 /*
2083  * Lookup the swblock containing the specified swap block index.
2084  *
2085  * The caller must hold the object.
2086  */
2087 static __inline
2088 struct swblock *
2089 swp_pager_lookup(vm_object_t object, vm_pindex_t index)
2090 {
2091 	ASSERT_LWKT_TOKEN_HELD(vm_object_token(object));
2092 	index &= ~(vm_pindex_t)SWAP_META_MASK;
2093 	return (RB_LOOKUP(swblock_rb_tree, &object->swblock_root, index));
2094 }
2095 
2096 /*
2097  * Remove a swblock from the RB tree.
2098  *
2099  * The caller must hold the object.
2100  */
2101 static __inline
2102 void
2103 swp_pager_remove(vm_object_t object, struct swblock *swap)
2104 {
2105 	ASSERT_LWKT_TOKEN_HELD(vm_object_token(object));
2106 	RB_REMOVE(swblock_rb_tree, &object->swblock_root, swap);
2107 }
2108 
2109 /*
2110  * Convert default object to swap object if necessary
2111  *
2112  * The caller must hold the object.
2113  */
2114 static void
2115 swp_pager_meta_convert(vm_object_t object)
2116 {
2117 	if (object->type == OBJT_DEFAULT) {
2118 		object->type = OBJT_SWAP;
2119 		KKASSERT(object->swblock_count == 0);
2120 	}
2121 }
2122 
2123 /*
2124  * SWP_PAGER_META_BUILD() -	add swap block to swap meta data for object
2125  *
2126  *	We first convert the object to a swap object if it is a default
2127  *	object.  Vnode objects do not need to be converted.
2128  *
2129  *	The specified swapblk is added to the object's swap metadata.  If
2130  *	the swapblk is not valid, it is freed instead.  Any previously
2131  *	assigned swapblk is freed.
2132  *
2133  * The caller must hold the object.
2134  */
2135 static void
2136 swp_pager_meta_build(vm_object_t object, vm_pindex_t index, swblk_t swapblk)
2137 {
2138 	struct swblock *swap;
2139 	struct swblock *oswap;
2140 	vm_pindex_t v;
2141 
2142 	KKASSERT(swapblk != SWAPBLK_NONE);
2143 	ASSERT_LWKT_TOKEN_HELD(vm_object_token(object));
2144 
2145 	/*
2146 	 * Convert object if necessary
2147 	 */
2148 	if (object->type == OBJT_DEFAULT)
2149 		swp_pager_meta_convert(object);
2150 
2151 	/*
2152 	 * Locate swblock.  If not found create, but if we aren't adding
2153 	 * anything just return.  If we run out of space in the map we wait
2154 	 * and, since the hash table may have changed, retry.
2155 	 */
2156 retry:
2157 	swap = swp_pager_lookup(object, index);
2158 
2159 	if (swap == NULL) {
2160 		int i;
2161 
2162 		swap = zalloc(swap_zone);
2163 		if (swap == NULL) {
2164 			vm_wait(0);
2165 			goto retry;
2166 		}
2167 		swap->swb_index = index & ~(vm_pindex_t)SWAP_META_MASK;
2168 		swap->swb_count = 0;
2169 
2170 		++object->swblock_count;
2171 
2172 		for (i = 0; i < SWAP_META_PAGES; ++i)
2173 			swap->swb_pages[i] = SWAPBLK_NONE;
2174 		oswap = RB_INSERT(swblock_rb_tree, &object->swblock_root, swap);
2175 		KKASSERT(oswap == NULL);
2176 	}
2177 
2178 	/*
2179 	 * Delete prior contents of metadata.
2180 	 *
2181 	 * NOTE: Decrement swb_count after the freeing operation (which
2182 	 *	 might block) to prevent racing destruction of the swblock.
2183 	 */
2184 	index &= SWAP_META_MASK;
2185 
2186 	while ((v = swap->swb_pages[index]) != SWAPBLK_NONE) {
2187 		swap->swb_pages[index] = SWAPBLK_NONE;
2188 		/* can block */
2189 		swp_pager_freeswapspace(object, v, 1);
2190 		--swap->swb_count;
2191 		--mycpu->gd_vmtotal.t_vm;
2192 	}
2193 
2194 	/*
2195 	 * Enter block into metadata
2196 	 */
2197 	swap->swb_pages[index] = swapblk;
2198 	if (swapblk != SWAPBLK_NONE) {
2199 		++swap->swb_count;
2200 		++mycpu->gd_vmtotal.t_vm;
2201 	}
2202 }
2203 
2204 /*
2205  * SWP_PAGER_META_FREE() - free a range of blocks in the object's swap metadata
2206  *
2207  *	The requested range of blocks is freed, with any associated swap
2208  *	returned to the swap bitmap.
2209  *
2210  *	This routine will free swap metadata structures as they are cleaned
2211  *	out.  This routine does *NOT* operate on swap metadata associated
2212  *	with resident pages.
2213  *
2214  * The caller must hold the object.
2215  */
2216 static int swp_pager_meta_free_callback(struct swblock *swb, void *data);
2217 
2218 static void
2219 swp_pager_meta_free(vm_object_t object, vm_pindex_t index, vm_pindex_t count)
2220 {
2221 	struct swfreeinfo info;
2222 
2223 	ASSERT_LWKT_TOKEN_HELD(vm_object_token(object));
2224 
2225 	/*
2226 	 * Nothing to do
2227 	 */
2228 	if (object->swblock_count == 0) {
2229 		KKASSERT(RB_EMPTY(&object->swblock_root));
2230 		return;
2231 	}
2232 	if (count == 0)
2233 		return;
2234 
2235 	/*
2236 	 * Setup for RB tree scan.  Note that the pindex range can be huge
2237 	 * due to the 64 bit page index space so we cannot safely iterate.
2238 	 */
2239 	info.object = object;
2240 	info.basei = index & ~(vm_pindex_t)SWAP_META_MASK;
2241 	info.begi = index;
2242 	info.endi = index + count - 1;
2243 	swblock_rb_tree_RB_SCAN(&object->swblock_root, rb_swblock_scancmp,
2244 				swp_pager_meta_free_callback, &info);
2245 }
2246 
2247 /*
2248  * The caller must hold the object.
2249  */
2250 static
2251 int
2252 swp_pager_meta_free_callback(struct swblock *swap, void *data)
2253 {
2254 	struct swfreeinfo *info = data;
2255 	vm_object_t object = info->object;
2256 	int index;
2257 	int eindex;
2258 
2259 	/*
2260 	 * Figure out the range within the swblock.  The wider scan may
2261 	 * return edge-case swap blocks when the start and/or end points
2262 	 * are in the middle of a block.
2263 	 */
2264 	if (swap->swb_index < info->begi)
2265 		index = (int)info->begi & SWAP_META_MASK;
2266 	else
2267 		index = 0;
2268 
2269 	if (swap->swb_index + SWAP_META_PAGES > info->endi)
2270 		eindex = (int)info->endi & SWAP_META_MASK;
2271 	else
2272 		eindex = SWAP_META_MASK;
2273 
2274 	/*
2275 	 * Scan and free the blocks.  The loop terminates early
2276 	 * if (swap) runs out of blocks and could be freed.
2277 	 *
2278 	 * NOTE: Decrement swb_count after swp_pager_freeswapspace()
2279 	 *	 to deal with a zfree race.
2280 	 */
2281 	while (index <= eindex) {
2282 		swblk_t v = swap->swb_pages[index];
2283 
2284 		if (v != SWAPBLK_NONE) {
2285 			swap->swb_pages[index] = SWAPBLK_NONE;
2286 			/* can block */
2287 			swp_pager_freeswapspace(object, v, 1);
2288 			--mycpu->gd_vmtotal.t_vm;
2289 			if (--swap->swb_count == 0) {
2290 				swp_pager_remove(object, swap);
2291 				zfree(swap_zone, swap);
2292 				--object->swblock_count;
2293 				break;
2294 			}
2295 		}
2296 		++index;
2297 	}
2298 
2299 	/* swap may be invalid here due to zfree above */
2300 	lwkt_yield();
2301 
2302 	return(0);
2303 }
2304 
2305 /*
2306  * SWP_PAGER_META_FREE_ALL() - destroy all swap metadata associated with object
2307  *
2308  *	This routine locates and destroys all swap metadata associated with
2309  *	an object.
2310  *
2311  * NOTE: Decrement swb_count after the freeing operation (which
2312  *	 might block) to prevent racing destruction of the swblock.
2313  *
2314  * The caller must hold the object.
2315  */
2316 static void
2317 swp_pager_meta_free_all(vm_object_t object)
2318 {
2319 	struct swblock *swap;
2320 	int i;
2321 
2322 	ASSERT_LWKT_TOKEN_HELD(vm_object_token(object));
2323 
2324 	while ((swap = RB_ROOT(&object->swblock_root)) != NULL) {
2325 		swp_pager_remove(object, swap);
2326 		for (i = 0; i < SWAP_META_PAGES; ++i) {
2327 			swblk_t v = swap->swb_pages[i];
2328 			if (v != SWAPBLK_NONE) {
2329 				/* can block */
2330 				swp_pager_freeswapspace(object, v, 1);
2331 				--swap->swb_count;
2332 				--mycpu->gd_vmtotal.t_vm;
2333 			}
2334 		}
2335 		if (swap->swb_count != 0)
2336 			panic("swap_pager_meta_free_all: swb_count != 0");
2337 		zfree(swap_zone, swap);
2338 		--object->swblock_count;
2339 		lwkt_yield();
2340 	}
2341 	KKASSERT(object->swblock_count == 0);
2342 }
2343 
2344 /*
2345  * SWP_PAGER_METACTL() -  misc control of swap and vm_page_t meta data.
2346  *
2347  *	This routine is capable of looking up, popping, or freeing
2348  *	swapblk assignments in the swap meta data or in the vm_page_t.
2349  *	The routine typically returns the swapblk being looked-up, or popped,
2350  *	or SWAPBLK_NONE if the block was freed, or SWAPBLK_NONE if the block
2351  *	was invalid.  This routine will automatically free any invalid
2352  *	meta-data swapblks.
2353  *
2354  *	It is not possible to store invalid swapblks in the swap meta data
2355  *	(other then a literal 'SWAPBLK_NONE'), so we don't bother checking.
2356  *
2357  *	When acting on a busy resident page and paging is in progress, we
2358  *	have to wait until paging is complete but otherwise can act on the
2359  *	busy page.
2360  *
2361  *	SWM_FREE	remove and free swap block from metadata
2362  *	SWM_POP		remove from meta data but do not free.. pop it out
2363  *
2364  * The caller must hold the object.
2365  */
2366 static swblk_t
2367 swp_pager_meta_ctl(vm_object_t object, vm_pindex_t index, int flags)
2368 {
2369 	struct swblock *swap;
2370 	swblk_t r1;
2371 
2372 	if (object->swblock_count == 0)
2373 		return(SWAPBLK_NONE);
2374 
2375 	r1 = SWAPBLK_NONE;
2376 	swap = swp_pager_lookup(object, index);
2377 
2378 	if (swap != NULL) {
2379 		index &= SWAP_META_MASK;
2380 		r1 = swap->swb_pages[index];
2381 
2382 		if (r1 != SWAPBLK_NONE) {
2383 			if (flags & (SWM_FREE|SWM_POP)) {
2384 				swap->swb_pages[index] = SWAPBLK_NONE;
2385 				--mycpu->gd_vmtotal.t_vm;
2386 				if (--swap->swb_count == 0) {
2387 					swp_pager_remove(object, swap);
2388 					zfree(swap_zone, swap);
2389 					--object->swblock_count;
2390 				}
2391 			}
2392 			/* swap ptr may be invalid */
2393 			if (flags & SWM_FREE) {
2394 				swp_pager_freeswapspace(object, r1, 1);
2395 				r1 = SWAPBLK_NONE;
2396 			}
2397 		}
2398 		/* swap ptr may be invalid */
2399 	}
2400 	return(r1);
2401 }
2402