xref: /dflybsd-src/sys/vm/swap_pager.c (revision 475c7069e94570a897d1467613efd2b3f0212ff9)
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 #define SWBIO_TTC	0x08	/* for VM_PAGER_TRY_TO_CACHE */
136 
137 struct swfreeinfo {
138 	vm_object_t	object;
139 	vm_pindex_t	basei;
140 	vm_pindex_t	begi;
141 	vm_pindex_t	endi;	/* inclusive */
142 };
143 
144 struct swswapoffinfo {
145 	vm_object_t	object;
146 	int		devidx;
147 	int		shared;
148 };
149 
150 /*
151  * vm_swap_size is in page-sized chunks now.  It was DEV_BSIZE'd chunks
152  * in the old system.
153  */
154 
155 int swap_pager_full;		/* swap space exhaustion (task killing) */
156 int swap_fail_ticks;		/* when we became exhausted */
157 int swap_pager_almost_full;	/* swap space exhaustion (w/ hysteresis)*/
158 swblk_t vm_swap_cache_use;
159 swblk_t vm_swap_anon_use;
160 static int vm_report_swap_allocs;
161 
162 static int nsw_rcount;		/* free read buffers			*/
163 static int nsw_wcount_sync;	/* limit write buffers / synchronous	*/
164 static int nsw_wcount_async;	/* limit write buffers / asynchronous	*/
165 static int nsw_wcount_async_max;/* assigned maximum			*/
166 static int nsw_cluster_max;	/* maximum VOP I/O allowed		*/
167 
168 struct blist *swapblist;
169 static int swap_async_max = 4;	/* maximum in-progress async I/O's	*/
170 static int swap_burst_read = 0;	/* allow burst reading */
171 static swblk_t swapiterator;	/* linearize allocations */
172 int swap_user_async = 0;	/* user swap pager operation can be async */
173 
174 static struct spinlock swapbp_spin = SPINLOCK_INITIALIZER(&swapbp_spin, "swapbp_spin");
175 
176 /* from vm_swap.c */
177 extern struct vnode *swapdev_vp;
178 extern struct swdevt *swdevt;
179 extern int nswdev;
180 
181 #define BLK2DEVIDX(blk) (nswdev > 1 ? blk / SWB_DMMAX % nswdev : 0)
182 
183 SYSCTL_INT(_vm, OID_AUTO, swap_async_max,
184         CTLFLAG_RW, &swap_async_max, 0, "Maximum running async swap ops");
185 SYSCTL_INT(_vm, OID_AUTO, swap_burst_read,
186         CTLFLAG_RW, &swap_burst_read, 0, "Allow burst reads for pageins");
187 SYSCTL_INT(_vm, OID_AUTO, swap_user_async,
188         CTLFLAG_RW, &swap_user_async, 0, "Allow async uuser swap write I/O");
189 
190 #if SWBLK_BITS == 64
191 SYSCTL_LONG(_vm, OID_AUTO, swap_cache_use,
192         CTLFLAG_RD, &vm_swap_cache_use, 0, "");
193 SYSCTL_LONG(_vm, OID_AUTO, swap_anon_use,
194         CTLFLAG_RD, &vm_swap_anon_use, 0, "");
195 SYSCTL_LONG(_vm, OID_AUTO, swap_size,
196         CTLFLAG_RD, &vm_swap_size, 0, "");
197 #else
198 SYSCTL_INT(_vm, OID_AUTO, swap_cache_use,
199         CTLFLAG_RD, &vm_swap_cache_use, 0, "");
200 SYSCTL_INT(_vm, OID_AUTO, swap_anon_use,
201         CTLFLAG_RD, &vm_swap_anon_use, 0, "");
202 SYSCTL_INT(_vm, OID_AUTO, swap_size,
203         CTLFLAG_RD, &vm_swap_size, 0, "");
204 #endif
205 SYSCTL_INT(_vm, OID_AUTO, report_swap_allocs,
206         CTLFLAG_RW, &vm_report_swap_allocs, 0, "");
207 
208 vm_zone_t		swap_zone;
209 
210 /*
211  * Red-Black tree for swblock entries
212  *
213  * The caller must hold vm_token
214  */
215 RB_GENERATE2(swblock_rb_tree, swblock, swb_entry, rb_swblock_compare,
216 	     vm_pindex_t, swb_index);
217 
218 int
219 rb_swblock_compare(struct swblock *swb1, struct swblock *swb2)
220 {
221 	if (swb1->swb_index < swb2->swb_index)
222 		return(-1);
223 	if (swb1->swb_index > swb2->swb_index)
224 		return(1);
225 	return(0);
226 }
227 
228 static
229 int
230 rb_swblock_scancmp(struct swblock *swb, void *data)
231 {
232 	struct swfreeinfo *info = data;
233 
234 	if (swb->swb_index < info->basei)
235 		return(-1);
236 	if (swb->swb_index > info->endi)
237 		return(1);
238 	return(0);
239 }
240 
241 static
242 int
243 rb_swblock_condcmp(struct swblock *swb, void *data)
244 {
245 	struct swfreeinfo *info = data;
246 
247 	if (swb->swb_index < info->basei)
248 		return(-1);
249 	return(0);
250 }
251 
252 /*
253  * pagerops for OBJT_SWAP - "swap pager".  Some ops are also global procedure
254  * calls hooked from other parts of the VM system and do not appear here.
255  * (see vm/swap_pager.h).
256  */
257 
258 static void	swap_pager_dealloc (vm_object_t object);
259 static int	swap_pager_getpage (vm_object_t, vm_page_t *, int);
260 static void	swap_chain_iodone(struct bio *biox);
261 
262 struct pagerops swappagerops = {
263 	swap_pager_dealloc,	/* deallocate an OBJT_SWAP object	*/
264 	swap_pager_getpage,	/* pagein				*/
265 	swap_pager_putpages,	/* pageout				*/
266 	swap_pager_haspage	/* get backing store status for page	*/
267 };
268 
269 /*
270  * SWB_DMMAX is in page-sized chunks with the new swap system.  It was
271  * dev-bsized chunks in the old.  SWB_DMMAX is always a power of 2.
272  *
273  * swap_*() routines are externally accessible.  swp_*() routines are
274  * internal.
275  */
276 
277 int nswap_lowat = 128;		/* in pages, swap_pager_almost_full warn */
278 int nswap_hiwat = 512;		/* in pages, swap_pager_almost_full warn */
279 
280 static __inline void	swp_sizecheck (void);
281 static void	swp_pager_async_iodone (struct bio *bio);
282 
283 /*
284  * Swap bitmap functions
285  */
286 
287 static __inline void	swp_pager_freeswapspace(vm_object_t object,
288 						swblk_t blk, int npages);
289 static __inline swblk_t	swp_pager_getswapspace(vm_object_t object, int npages);
290 
291 /*
292  * Metadata functions
293  */
294 
295 static void swp_pager_meta_convert(vm_object_t);
296 static void swp_pager_meta_build(vm_object_t, vm_pindex_t, swblk_t);
297 static void swp_pager_meta_free(vm_object_t, vm_pindex_t, vm_pindex_t);
298 static void swp_pager_meta_free_all(vm_object_t);
299 static swblk_t swp_pager_meta_ctl(vm_object_t, vm_pindex_t, int);
300 
301 /*
302  * SWP_SIZECHECK() -	update swap_pager_full indication
303  *
304  *	update the swap_pager_almost_full indication and warn when we are
305  *	about to run out of swap space, using lowat/hiwat hysteresis.
306  *
307  *	Clear swap_pager_full ( task killing ) indication when lowat is met.
308  *
309  * No restrictions on call
310  * This routine may not block.
311  * SMP races are ok.
312  */
313 static __inline void
314 swp_sizecheck(void)
315 {
316 	if (vm_swap_size < nswap_lowat) {
317 		if (swap_pager_almost_full == 0) {
318 			kprintf("swap_pager: out of swap space\n");
319 			swap_pager_almost_full = 1;
320 			swap_fail_ticks = ticks;
321 		}
322 	} else {
323 		swap_pager_full = 0;
324 		if (vm_swap_size > nswap_hiwat)
325 			swap_pager_almost_full = 0;
326 	}
327 }
328 
329 /*
330  * SWAP_PAGER_INIT() -	initialize the swap pager!
331  *
332  *	Expected to be started from system init.  NOTE:  This code is run
333  *	before much else so be careful what you depend on.  Most of the VM
334  *	system has yet to be initialized at this point.
335  *
336  * Called from the low level boot code only.
337  */
338 static void
339 swap_pager_init(void *arg __unused)
340 {
341 }
342 SYSINIT(vm_mem, SI_BOOT1_VM, SI_ORDER_THIRD, swap_pager_init, NULL);
343 
344 /*
345  * SWAP_PAGER_SWAP_INIT() - swap pager initialization from pageout process
346  *
347  *	Expected to be started from pageout process once, prior to entering
348  *	its main loop.
349  *
350  * Called from the low level boot code only.
351  */
352 void
353 swap_pager_swap_init(void)
354 {
355 	int n, n2;
356 
357 	/*
358 	 * Number of in-transit swap bp operations.  Don't
359 	 * exhaust the pbufs completely.  Make sure we
360 	 * initialize workable values (0 will work for hysteresis
361 	 * but it isn't very efficient).
362 	 *
363 	 * The nsw_cluster_max is constrained by the number of pages an XIO
364 	 * holds, i.e., (MAXPHYS/PAGE_SIZE) and our locally defined
365 	 * MAX_PAGEOUT_CLUSTER.   Also be aware that swap ops are
366 	 * constrained by the swap device interleave stripe size.
367 	 *
368 	 * Currently we hardwire nsw_wcount_async to 4.  This limit is
369 	 * designed to prevent other I/O from having high latencies due to
370 	 * our pageout I/O.  The value 4 works well for one or two active swap
371 	 * devices but is probably a little low if you have more.  Even so,
372 	 * a higher value would probably generate only a limited improvement
373 	 * with three or four active swap devices since the system does not
374 	 * typically have to pageout at extreme bandwidths.   We will want
375 	 * at least 2 per swap devices, and 4 is a pretty good value if you
376 	 * have one NFS swap device due to the command/ack latency over NFS.
377 	 * So it all works out pretty well.
378 	 */
379 
380 	nsw_cluster_max = min((MAXPHYS/PAGE_SIZE), MAX_PAGEOUT_CLUSTER);
381 
382 	nsw_rcount = (nswbuf_kva + 1) / 2;
383 	nsw_wcount_sync = (nswbuf_kva + 3) / 4;
384 	nsw_wcount_async = 4;
385 	nsw_wcount_async_max = nsw_wcount_async;
386 
387 	/*
388 	 * The zone is dynamically allocated so generally size it to
389 	 * maxswzone (32MB to 256GB of KVM).  Set a minimum size based
390 	 * on physical memory of around 8x (each swblock can hold 16 pages).
391 	 *
392 	 * With the advent of SSDs (vs HDs) the practical (swap:memory) ratio
393 	 * has increased dramatically.
394 	 */
395 	n = vmstats.v_page_count / 2;
396 	if (maxswzone && n < maxswzone / sizeof(struct swblock))
397 		n = maxswzone / sizeof(struct swblock);
398 	n2 = n;
399 
400 	do {
401 		swap_zone = zinit(
402 			"SWAPMETA",
403 			sizeof(struct swblock),
404 			n,
405 			ZONE_INTERRUPT);
406 		if (swap_zone != NULL)
407 			break;
408 		/*
409 		 * if the allocation failed, try a zone two thirds the
410 		 * size of the previous attempt.
411 		 */
412 		n -= ((n + 2) / 3);
413 	} while (n > 0);
414 
415 	if (swap_zone == NULL)
416 		panic("swap_pager_swap_init: swap_zone == NULL");
417 	if (n2 != n)
418 		kprintf("Swap zone entries reduced from %d to %d.\n", n2, n);
419 }
420 
421 /*
422  * SWAP_PAGER_ALLOC() -	allocate a new OBJT_SWAP VM object and instantiate
423  *			its metadata structures.
424  *
425  *	This routine is called from the mmap and fork code to create a new
426  *	OBJT_SWAP object.  We do this by creating an OBJT_DEFAULT object
427  *	and then converting it with swp_pager_meta_convert().
428  *
429  *	We only support unnamed objects.
430  *
431  * No restrictions.
432  */
433 vm_object_t
434 swap_pager_alloc(void *handle, off_t size, vm_prot_t prot, off_t offset)
435 {
436 	vm_object_t object;
437 
438 	KKASSERT(handle == NULL);
439 	object = vm_object_allocate_hold(OBJT_DEFAULT,
440 					 OFF_TO_IDX(offset + PAGE_MASK + size));
441 	swp_pager_meta_convert(object);
442 	vm_object_drop(object);
443 
444 	return (object);
445 }
446 
447 /*
448  * SWAP_PAGER_DEALLOC() -	remove swap metadata from object
449  *
450  *	The swap backing for the object is destroyed.  The code is
451  *	designed such that we can reinstantiate it later, but this
452  *	routine is typically called only when the entire object is
453  *	about to be destroyed.
454  *
455  * The object must be locked or unreferenceable.
456  * No other requirements.
457  */
458 static void
459 swap_pager_dealloc(vm_object_t object)
460 {
461 	vm_object_hold(object);
462 	vm_object_pip_wait(object, "swpdea");
463 
464 	/*
465 	 * Free all remaining metadata.  We only bother to free it from
466 	 * the swap meta data.  We do not attempt to free swapblk's still
467 	 * associated with vm_page_t's for this object.  We do not care
468 	 * if paging is still in progress on some objects.
469 	 */
470 	swp_pager_meta_free_all(object);
471 	vm_object_drop(object);
472 }
473 
474 /************************************************************************
475  *			SWAP PAGER BITMAP ROUTINES			*
476  ************************************************************************/
477 
478 /*
479  * SWP_PAGER_GETSWAPSPACE() -	allocate raw swap space
480  *
481  *	Allocate swap for the requested number of pages.  The starting
482  *	swap block number (a page index) is returned or SWAPBLK_NONE
483  *	if the allocation failed.
484  *
485  *	Also has the side effect of advising that somebody made a mistake
486  *	when they configured swap and didn't configure enough.
487  *
488  * The caller must hold the object.
489  * This routine may not block.
490  */
491 static __inline swblk_t
492 swp_pager_getswapspace(vm_object_t object, int npages)
493 {
494 	swblk_t blk;
495 
496 	lwkt_gettoken(&vm_token);
497 	blk = blist_allocat(swapblist, npages, swapiterator);
498 	if (blk == SWAPBLK_NONE)
499 		blk = blist_allocat(swapblist, npages, 0);
500 	if (blk == SWAPBLK_NONE) {
501 		if (swap_pager_full != 2) {
502 			if (vm_swap_max == 0)
503 				kprintf("Warning: The system would like to "
504 					"page to swap but no swap space "
505 					"is configured!\n");
506 			else
507 				kprintf("swap_pager_getswapspace: "
508 					"swap full allocating %d pages\n",
509 					npages);
510 			swap_pager_full = 2;
511 			if (swap_pager_almost_full == 0)
512 				swap_fail_ticks = ticks;
513 			swap_pager_almost_full = 1;
514 		}
515 	} else {
516 		/* swapiterator = blk; disable for now, doesn't work well */
517 		swapacctspace(blk, -npages);
518 		if (object->type == OBJT_SWAP)
519 			vm_swap_anon_use += npages;
520 		else
521 			vm_swap_cache_use += npages;
522 		swp_sizecheck();
523 	}
524 	lwkt_reltoken(&vm_token);
525 	return(blk);
526 }
527 
528 /*
529  * SWP_PAGER_FREESWAPSPACE() -	free raw swap space
530  *
531  *	This routine returns the specified swap blocks back to the bitmap.
532  *
533  *	Note:  This routine may not block (it could in the old swap code),
534  *	and through the use of the new blist routines it does not block.
535  *
536  * This routine may not block.
537  */
538 
539 static __inline void
540 swp_pager_freeswapspace(vm_object_t object, swblk_t blk, int npages)
541 {
542 	struct swdevt *sp = &swdevt[BLK2DEVIDX(blk)];
543 
544 	lwkt_gettoken(&vm_token);
545 	sp->sw_nused -= npages;
546 	if (object->type == OBJT_SWAP)
547 		vm_swap_anon_use -= npages;
548 	else
549 		vm_swap_cache_use -= npages;
550 
551 	if (sp->sw_flags & SW_CLOSING) {
552 		lwkt_reltoken(&vm_token);
553 		return;
554 	}
555 
556 	blist_free(swapblist, blk, npages);
557 	vm_swap_size += npages;
558 	swp_sizecheck();
559 	lwkt_reltoken(&vm_token);
560 }
561 
562 /*
563  * SWAP_PAGER_FREESPACE() -	frees swap blocks associated with a page
564  *				range within an object.
565  *
566  *	This is a globally accessible routine.
567  *
568  *	This routine removes swapblk assignments from swap metadata.
569  *
570  *	The external callers of this routine typically have already destroyed
571  *	or renamed vm_page_t's associated with this range in the object so
572  *	we should be ok.
573  *
574  * No requirements.
575  */
576 void
577 swap_pager_freespace(vm_object_t object, vm_pindex_t start, vm_pindex_t size)
578 {
579 	vm_object_hold(object);
580 	swp_pager_meta_free(object, start, size);
581 	vm_object_drop(object);
582 }
583 
584 /*
585  * No requirements.
586  */
587 void
588 swap_pager_freespace_all(vm_object_t object)
589 {
590 	vm_object_hold(object);
591 	swp_pager_meta_free_all(object);
592 	vm_object_drop(object);
593 }
594 
595 /*
596  * This function conditionally frees swap cache swap starting at
597  * (*basei) in the object.  (count) swap blocks will be nominally freed.
598  * The actual number of blocks freed can be more or less than the
599  * requested number.
600  *
601  * This function nominally returns the number of blocks freed.  However,
602  * the actual number of blocks freed may be less then the returned value.
603  * If the function is unable to exhaust the object or if it is able to
604  * free (approximately) the requested number of blocks it returns
605  * a value n > count.
606  *
607  * If we exhaust the object we will return a value n <= count.
608  *
609  * The caller must hold the object.
610  *
611  * WARNING!  If count == 0 then -1 can be returned as a degenerate case,
612  *	     callers should always pass a count value > 0.
613  */
614 static int swap_pager_condfree_callback(struct swblock *swap, void *data);
615 
616 int
617 swap_pager_condfree(vm_object_t object, vm_pindex_t *basei, int count)
618 {
619 	struct swfreeinfo info;
620 	int n;
621 	int t;
622 
623 	ASSERT_LWKT_TOKEN_HELD(vm_object_token(object));
624 
625 	info.object = object;
626 	info.basei = *basei;	/* skip up to this page index */
627 	info.begi = count;	/* max swap pages to destroy */
628 	info.endi = count * 8;	/* max swblocks to scan */
629 
630 	swblock_rb_tree_RB_SCAN(&object->swblock_root, rb_swblock_condcmp,
631 				swap_pager_condfree_callback, &info);
632 	*basei = info.basei;
633 
634 	/*
635 	 * Take the higher difference swblocks vs pages
636 	 */
637 	n = count - (int)info.begi;
638 	t = count * 8 - (int)info.endi;
639 	if (n < t)
640 		n = t;
641 	if (n < 1)
642 		n = 1;
643 	return(n);
644 }
645 
646 /*
647  * The idea is to free whole meta-block to avoid fragmenting
648  * the swap space or disk I/O.  We only do this if NO VM pages
649  * are present.
650  *
651  * We do not have to deal with clearing PG_SWAPPED in related VM
652  * pages because there are no related VM pages.
653  *
654  * The caller must hold the object.
655  */
656 static int
657 swap_pager_condfree_callback(struct swblock *swap, void *data)
658 {
659 	struct swfreeinfo *info = data;
660 	vm_object_t object = info->object;
661 	int i;
662 
663 	for (i = 0; i < SWAP_META_PAGES; ++i) {
664 		if (vm_page_lookup(object, swap->swb_index + i))
665 			break;
666 	}
667 	info->basei = swap->swb_index + SWAP_META_PAGES;
668 	if (i == SWAP_META_PAGES) {
669 		info->begi -= swap->swb_count;
670 		swap_pager_freespace(object, swap->swb_index, SWAP_META_PAGES);
671 	}
672 	--info->endi;
673 	if ((int)info->begi < 0 || (int)info->endi < 0)
674 		return(-1);
675 	lwkt_yield();
676 	return(0);
677 }
678 
679 /*
680  * Called by vm_page_alloc() when a new VM page is inserted
681  * into a VM object.  Checks whether swap has been assigned to
682  * the page and sets PG_SWAPPED as necessary.
683  *
684  * No requirements.
685  */
686 void
687 swap_pager_page_inserted(vm_page_t m)
688 {
689 	if (m->object->swblock_count) {
690 		vm_object_hold(m->object);
691 		if (swp_pager_meta_ctl(m->object, m->pindex, 0) != SWAPBLK_NONE)
692 			vm_page_flag_set(m, PG_SWAPPED);
693 		vm_object_drop(m->object);
694 	}
695 }
696 
697 /*
698  * SWAP_PAGER_RESERVE() - reserve swap blocks in object
699  *
700  *	Assigns swap blocks to the specified range within the object.  The
701  *	swap blocks are not zerod.  Any previous swap assignment is destroyed.
702  *
703  *	Returns 0 on success, -1 on failure.
704  *
705  * The caller is responsible for avoiding races in the specified range.
706  * No other requirements.
707  */
708 int
709 swap_pager_reserve(vm_object_t object, vm_pindex_t start, vm_size_t size)
710 {
711 	int n = 0;
712 	swblk_t blk = SWAPBLK_NONE;
713 	vm_pindex_t beg = start;	/* save start index */
714 
715 	vm_object_hold(object);
716 
717 	while (size) {
718 		if (n == 0) {
719 			n = BLIST_MAX_ALLOC;
720 			while ((blk = swp_pager_getswapspace(object, n)) ==
721 			       SWAPBLK_NONE)
722 			{
723 				n >>= 1;
724 				if (n == 0) {
725 					swp_pager_meta_free(object, beg,
726 							    start - beg);
727 					vm_object_drop(object);
728 					return(-1);
729 				}
730 			}
731 		}
732 		swp_pager_meta_build(object, start, blk);
733 		--size;
734 		++start;
735 		++blk;
736 		--n;
737 	}
738 	swp_pager_meta_free(object, start, n);
739 	vm_object_drop(object);
740 	return(0);
741 }
742 
743 /*
744  * SWAP_PAGER_COPY() -  copy blocks from source pager to destination pager
745  *			and destroy the source.
746  *
747  *	Copy any valid swapblks from the source to the destination.  In
748  *	cases where both the source and destination have a valid swapblk,
749  *	we keep the destination's.
750  *
751  *	This routine is allowed to block.  It may block allocating metadata
752  *	indirectly through swp_pager_meta_build() or if paging is still in
753  *	progress on the source.
754  *
755  *	XXX vm_page_collapse() kinda expects us not to block because we
756  *	supposedly do not need to allocate memory, but for the moment we
757  *	*may* have to get a little memory from the zone allocator, but
758  *	it is taken from the interrupt memory.  We should be ok.
759  *
760  *	The source object contains no vm_page_t's (which is just as well)
761  *	The source object is of type OBJT_SWAP.
762  *
763  *	The source and destination objects must be held by the caller.
764  */
765 void
766 swap_pager_copy(vm_object_t srcobject, vm_object_t dstobject,
767 		vm_pindex_t base_index, int destroysource)
768 {
769 	vm_pindex_t i;
770 
771 	ASSERT_LWKT_TOKEN_HELD(vm_object_token(srcobject));
772 	ASSERT_LWKT_TOKEN_HELD(vm_object_token(dstobject));
773 
774 	/*
775 	 * transfer source to destination.
776 	 */
777 	for (i = 0; i < dstobject->size; ++i) {
778 		swblk_t dstaddr;
779 
780 		/*
781 		 * Locate (without changing) the swapblk on the destination,
782 		 * unless it is invalid in which case free it silently, or
783 		 * if the destination is a resident page, in which case the
784 		 * source is thrown away.
785 		 */
786 		dstaddr = swp_pager_meta_ctl(dstobject, i, 0);
787 
788 		if (dstaddr == SWAPBLK_NONE) {
789 			/*
790 			 * Destination has no swapblk and is not resident,
791 			 * copy source.
792 			 */
793 			swblk_t srcaddr;
794 
795 			srcaddr = swp_pager_meta_ctl(srcobject,
796 						     base_index + i, SWM_POP);
797 
798 			if (srcaddr != SWAPBLK_NONE)
799 				swp_pager_meta_build(dstobject, i, srcaddr);
800 		} else {
801 			/*
802 			 * Destination has valid swapblk or it is represented
803 			 * by a resident page.  We destroy the sourceblock.
804 			 */
805 			swp_pager_meta_ctl(srcobject, base_index + i, SWM_FREE);
806 		}
807 	}
808 
809 	/*
810 	 * Free left over swap blocks in source.
811 	 *
812 	 * We have to revert the type to OBJT_DEFAULT so we do not accidently
813 	 * double-remove the object from the swap queues.
814 	 */
815 	if (destroysource) {
816 		/*
817 		 * Reverting the type is not necessary, the caller is going
818 		 * to destroy srcobject directly, but I'm doing it here
819 		 * for consistency since we've removed the object from its
820 		 * queues.
821 		 */
822 		swp_pager_meta_free_all(srcobject);
823 		if (srcobject->type == OBJT_SWAP)
824 			srcobject->type = OBJT_DEFAULT;
825 	}
826 }
827 
828 /*
829  * SWAP_PAGER_HASPAGE() -	determine if we have good backing store for
830  *				the requested page.
831  *
832  *	We determine whether good backing store exists for the requested
833  *	page and return TRUE if it does, FALSE if it doesn't.
834  *
835  *	If TRUE, we also try to determine how much valid, contiguous backing
836  *	store exists before and after the requested page within a reasonable
837  *	distance.  We do not try to restrict it to the swap device stripe
838  *	(that is handled in getpages/putpages).  It probably isn't worth
839  *	doing here.
840  *
841  * No requirements.
842  */
843 boolean_t
844 swap_pager_haspage(vm_object_t object, vm_pindex_t pindex)
845 {
846 	swblk_t blk0;
847 
848 	/*
849 	 * do we have good backing store at the requested index ?
850 	 */
851 	vm_object_hold(object);
852 	blk0 = swp_pager_meta_ctl(object, pindex, 0);
853 
854 	if (blk0 == SWAPBLK_NONE) {
855 		vm_object_drop(object);
856 		return (FALSE);
857 	}
858 	vm_object_drop(object);
859 	return (TRUE);
860 }
861 
862 /*
863  * SWAP_PAGER_PAGE_UNSWAPPED() - remove swap backing store related to page
864  *
865  * This removes any associated swap backing store, whether valid or
866  * not, from the page.  This operates on any VM object, not just OBJT_SWAP
867  * objects.
868  *
869  * This routine is typically called when a page is made dirty, at
870  * which point any associated swap can be freed.  MADV_FREE also
871  * calls us in a special-case situation
872  *
873  * NOTE!!!  If the page is clean and the swap was valid, the caller
874  * should make the page dirty before calling this routine.  This routine
875  * does NOT change the m->dirty status of the page.  Also: MADV_FREE
876  * depends on it.
877  *
878  * The page must be busied or soft-busied.
879  * The caller can hold the object to avoid blocking, else we might block.
880  * No other requirements.
881  */
882 void
883 swap_pager_unswapped(vm_page_t m)
884 {
885 	if (m->flags & PG_SWAPPED) {
886 		vm_object_hold(m->object);
887 		KKASSERT(m->flags & PG_SWAPPED);
888 		swp_pager_meta_ctl(m->object, m->pindex, SWM_FREE);
889 		vm_page_flag_clear(m, PG_SWAPPED);
890 		vm_object_drop(m->object);
891 	}
892 }
893 
894 /*
895  * SWAP_PAGER_STRATEGY() - read, write, free blocks
896  *
897  * This implements a VM OBJECT strategy function using swap backing store.
898  * This can operate on any VM OBJECT type, not necessarily just OBJT_SWAP
899  * types.
900  *
901  * This is intended to be a cacheless interface (i.e. caching occurs at
902  * higher levels), and is also used as a swap-based SSD cache for vnode
903  * and device objects.
904  *
905  * All I/O goes directly to and from the swap device.
906  *
907  * We currently attempt to run I/O synchronously or asynchronously as
908  * the caller requests.  This isn't perfect because we loose error
909  * sequencing when we run multiple ops in parallel to satisfy a request.
910  * But this is swap, so we let it all hang out.
911  *
912  * No requirements.
913  */
914 void
915 swap_pager_strategy(vm_object_t object, struct bio *bio)
916 {
917 	struct buf *bp = bio->bio_buf;
918 	struct bio *nbio;
919 	vm_pindex_t start;
920 	vm_pindex_t biox_blkno = 0;
921 	int count;
922 	char *data;
923 	struct bio *biox;
924 	struct buf *bufx;
925 #if 0
926 	struct bio_track *track;
927 #endif
928 
929 #if 0
930 	/*
931 	 * tracking for swapdev vnode I/Os
932 	 */
933 	if (bp->b_cmd == BUF_CMD_READ)
934 		track = &swapdev_vp->v_track_read;
935 	else
936 		track = &swapdev_vp->v_track_write;
937 #endif
938 
939 	if (bp->b_bcount & PAGE_MASK) {
940 		bp->b_error = EINVAL;
941 		bp->b_flags |= B_ERROR | B_INVAL;
942 		biodone(bio);
943 		kprintf("swap_pager_strategy: bp %p offset %lld size %d, "
944 			"not page bounded\n",
945 			bp, (long long)bio->bio_offset, (int)bp->b_bcount);
946 		return;
947 	}
948 
949 	/*
950 	 * Clear error indication, initialize page index, count, data pointer.
951 	 */
952 	bp->b_error = 0;
953 	bp->b_flags &= ~B_ERROR;
954 	bp->b_resid = bp->b_bcount;
955 
956 	start = (vm_pindex_t)(bio->bio_offset >> PAGE_SHIFT);
957 	count = howmany(bp->b_bcount, PAGE_SIZE);
958 	data = bp->b_data;
959 
960 	/*
961 	 * Deal with BUF_CMD_FREEBLKS
962 	 */
963 	if (bp->b_cmd == BUF_CMD_FREEBLKS) {
964 		/*
965 		 * FREE PAGE(s) - destroy underlying swap that is no longer
966 		 *		  needed.
967 		 */
968 		vm_object_hold(object);
969 		swp_pager_meta_free(object, start, count);
970 		vm_object_drop(object);
971 		bp->b_resid = 0;
972 		biodone(bio);
973 		return;
974 	}
975 
976 	/*
977 	 * We need to be able to create a new cluster of I/O's.  We cannot
978 	 * use the caller fields of the passed bio so push a new one.
979 	 *
980 	 * Because nbio is just a placeholder for the cluster links,
981 	 * we can biodone() the original bio instead of nbio to make
982 	 * things a bit more efficient.
983 	 */
984 	nbio = push_bio(bio);
985 	nbio->bio_offset = bio->bio_offset;
986 	nbio->bio_caller_info1.cluster_head = NULL;
987 	nbio->bio_caller_info2.cluster_tail = NULL;
988 
989 	biox = NULL;
990 	bufx = NULL;
991 
992 	/*
993 	 * Execute read or write
994 	 */
995 	vm_object_hold(object);
996 
997 	while (count > 0) {
998 		swblk_t blk;
999 
1000 		/*
1001 		 * Obtain block.  If block not found and writing, allocate a
1002 		 * new block and build it into the object.
1003 		 */
1004 		blk = swp_pager_meta_ctl(object, start, 0);
1005 		if ((blk == SWAPBLK_NONE) && bp->b_cmd != BUF_CMD_READ) {
1006 			blk = swp_pager_getswapspace(object, 1);
1007 			if (blk == SWAPBLK_NONE) {
1008 				bp->b_error = ENOMEM;
1009 				bp->b_flags |= B_ERROR;
1010 				break;
1011 			}
1012 			swp_pager_meta_build(object, start, blk);
1013 		}
1014 
1015 		/*
1016 		 * Do we have to flush our current collection?  Yes if:
1017 		 *
1018 		 *	- no swap block at this index
1019 		 *	- swap block is not contiguous
1020 		 *	- we cross a physical disk boundry in the
1021 		 *	  stripe.
1022 		 */
1023 		if (
1024 		    biox && (biox_blkno + btoc(bufx->b_bcount) != blk ||
1025 		     ((biox_blkno ^ blk) & ~SWB_DMMASK)
1026 		    )
1027 		) {
1028 			if (bp->b_cmd == BUF_CMD_READ) {
1029 				++mycpu->gd_cnt.v_swapin;
1030 				mycpu->gd_cnt.v_swappgsin += btoc(bufx->b_bcount);
1031 			} else {
1032 				++mycpu->gd_cnt.v_swapout;
1033 				mycpu->gd_cnt.v_swappgsout += btoc(bufx->b_bcount);
1034 				bufx->b_dirtyend = bufx->b_bcount;
1035 			}
1036 
1037 			/*
1038 			 * Finished with this buf.
1039 			 */
1040 			KKASSERT(bufx->b_bcount != 0);
1041 			if (bufx->b_cmd != BUF_CMD_READ)
1042 				bufx->b_dirtyend = bufx->b_bcount;
1043 			biox = NULL;
1044 			bufx = NULL;
1045 		}
1046 
1047 		/*
1048 		 * Add new swapblk to biox, instantiating biox if necessary.
1049 		 * Zero-fill reads are able to take a shortcut.
1050 		 */
1051 		if (blk == SWAPBLK_NONE) {
1052 			/*
1053 			 * We can only get here if we are reading.
1054 			 */
1055 			bzero(data, PAGE_SIZE);
1056 			bp->b_resid -= PAGE_SIZE;
1057 		} else {
1058 			if (biox == NULL) {
1059 				/* XXX chain count > 4, wait to <= 4 */
1060 
1061 				bufx = getpbuf(NULL);
1062 				biox = &bufx->b_bio1;
1063 				cluster_append(nbio, bufx);
1064 				bufx->b_cmd = bp->b_cmd;
1065 				biox->bio_done = swap_chain_iodone;
1066 				biox->bio_offset = (off_t)blk << PAGE_SHIFT;
1067 				biox->bio_caller_info1.cluster_parent = nbio;
1068 				biox_blkno = blk;
1069 				bufx->b_bcount = 0;
1070 				bufx->b_data = data;
1071 			}
1072 			bufx->b_bcount += PAGE_SIZE;
1073 		}
1074 		--count;
1075 		++start;
1076 		data += PAGE_SIZE;
1077 	}
1078 
1079 	vm_object_drop(object);
1080 
1081 	/*
1082 	 *  Flush out last buffer
1083 	 */
1084 	if (biox) {
1085 		if (bufx->b_cmd == BUF_CMD_READ) {
1086 			++mycpu->gd_cnt.v_swapin;
1087 			mycpu->gd_cnt.v_swappgsin += btoc(bufx->b_bcount);
1088 		} else {
1089 			++mycpu->gd_cnt.v_swapout;
1090 			mycpu->gd_cnt.v_swappgsout += btoc(bufx->b_bcount);
1091 			bufx->b_dirtyend = bufx->b_bcount;
1092 		}
1093 		KKASSERT(bufx->b_bcount);
1094 		if (bufx->b_cmd != BUF_CMD_READ)
1095 			bufx->b_dirtyend = bufx->b_bcount;
1096 		/* biox, bufx = NULL */
1097 	}
1098 
1099 	/*
1100 	 * Now initiate all the I/O.  Be careful looping on our chain as
1101 	 * I/O's may complete while we are still initiating them.
1102 	 *
1103 	 * If the request is a 100% sparse read no bios will be present
1104 	 * and we just biodone() the buffer.
1105 	 */
1106 	nbio->bio_caller_info2.cluster_tail = NULL;
1107 	bufx = nbio->bio_caller_info1.cluster_head;
1108 
1109 	if (bufx) {
1110 		while (bufx) {
1111 			biox = &bufx->b_bio1;
1112 			BUF_KERNPROC(bufx);
1113 			bufx = bufx->b_cluster_next;
1114 			vn_strategy(swapdev_vp, biox);
1115 		}
1116 	} else {
1117 		biodone(bio);
1118 	}
1119 
1120 	/*
1121 	 * Completion of the cluster will also call biodone_chain(nbio).
1122 	 * We never call biodone(nbio) so we don't have to worry about
1123 	 * setting up a bio_done callback.  It's handled in the sub-IO.
1124 	 */
1125 	/**/
1126 }
1127 
1128 /*
1129  * biodone callback
1130  *
1131  * No requirements.
1132  */
1133 static void
1134 swap_chain_iodone(struct bio *biox)
1135 {
1136 	struct buf **nextp;
1137 	struct buf *bufx;	/* chained sub-buffer */
1138 	struct bio *nbio;	/* parent nbio with chain glue */
1139 	struct buf *bp;		/* original bp associated with nbio */
1140 	int chain_empty;
1141 
1142 	bufx = biox->bio_buf;
1143 	nbio = biox->bio_caller_info1.cluster_parent;
1144 	bp = nbio->bio_buf;
1145 
1146 	/*
1147 	 * Update the original buffer
1148 	 */
1149         KKASSERT(bp != NULL);
1150 	if (bufx->b_flags & B_ERROR) {
1151 		atomic_set_int(&bufx->b_flags, B_ERROR);
1152 		bp->b_error = bufx->b_error;	/* race ok */
1153 	} else if (bufx->b_resid != 0) {
1154 		atomic_set_int(&bufx->b_flags, B_ERROR);
1155 		bp->b_error = EINVAL;		/* race ok */
1156 	} else {
1157 		atomic_subtract_int(&bp->b_resid, bufx->b_bcount);
1158 	}
1159 
1160 	/*
1161 	 * Remove us from the chain.
1162 	 */
1163 	spin_lock(&swapbp_spin);
1164 	nextp = &nbio->bio_caller_info1.cluster_head;
1165 	while (*nextp != bufx) {
1166 		KKASSERT(*nextp != NULL);
1167 		nextp = &(*nextp)->b_cluster_next;
1168 	}
1169 	*nextp = bufx->b_cluster_next;
1170 	chain_empty = (nbio->bio_caller_info1.cluster_head == NULL);
1171 	spin_unlock(&swapbp_spin);
1172 
1173 	/*
1174 	 * Clean up bufx.  If the chain is now empty we finish out
1175 	 * the parent.  Note that we may be racing other completions
1176 	 * so we must use the chain_empty status from above.
1177 	 */
1178 	if (chain_empty) {
1179 		if (bp->b_resid != 0 && !(bp->b_flags & B_ERROR)) {
1180 			atomic_set_int(&bp->b_flags, B_ERROR);
1181 			bp->b_error = EINVAL;
1182 		}
1183 		biodone_chain(nbio);
1184         }
1185         relpbuf(bufx, NULL);
1186 }
1187 
1188 /*
1189  * SWAP_PAGER_GETPAGES() - bring page in from swap
1190  *
1191  * The requested page may have to be brought in from swap.  Calculate the
1192  * swap block and bring in additional pages if possible.  All pages must
1193  * have contiguous swap block assignments and reside in the same object.
1194  *
1195  * The caller has a single vm_object_pip_add() reference prior to
1196  * calling us and we should return with the same.
1197  *
1198  * The caller has BUSY'd the page.  We should return with (*mpp) left busy,
1199  * and any additinal pages unbusied.
1200  *
1201  * If the caller encounters a PG_RAM page it will pass it to us even though
1202  * it may be valid and dirty.  We cannot overwrite the page in this case!
1203  * The case is used to allow us to issue pure read-aheads.
1204  *
1205  * NOTE! XXX This code does not entirely pipeline yet due to the fact that
1206  *       the PG_RAM page is validated at the same time as mreq.  What we
1207  *	 really need to do is issue a separate read-ahead pbuf.
1208  *
1209  * No requirements.
1210  */
1211 static int
1212 swap_pager_getpage(vm_object_t object, vm_page_t *mpp, int seqaccess)
1213 {
1214 	struct buf *bp;
1215 	struct bio *bio;
1216 	vm_page_t mreq;
1217 	vm_page_t m;
1218 	vm_offset_t kva;
1219 	swblk_t blk;
1220 	int i;
1221 	int j;
1222 	int raonly;
1223 	int error;
1224 	u_int32_t flags;
1225 	vm_page_t marray[XIO_INTERNAL_PAGES];
1226 
1227 	mreq = *mpp;
1228 
1229 	vm_object_hold(object);
1230 	if (mreq->object != object) {
1231 		panic("swap_pager_getpages: object mismatch %p/%p",
1232 		    object,
1233 		    mreq->object
1234 		);
1235 	}
1236 
1237 	/*
1238 	 * We don't want to overwrite a fully valid page as it might be
1239 	 * dirty.  This case can occur when e.g. vm_fault hits a perfectly
1240 	 * valid page with PG_RAM set.
1241 	 *
1242 	 * In this case we see if the next page is a suitable page-in
1243 	 * candidate and if it is we issue read-ahead.  PG_RAM will be
1244 	 * set on the last page of the read-ahead to continue the pipeline.
1245 	 */
1246 	if (mreq->valid == VM_PAGE_BITS_ALL) {
1247 		if (swap_burst_read == 0 || mreq->pindex + 1 >= object->size) {
1248 			vm_object_drop(object);
1249 			return(VM_PAGER_OK);
1250 		}
1251 		blk = swp_pager_meta_ctl(object, mreq->pindex + 1, 0);
1252 		if (blk == SWAPBLK_NONE) {
1253 			vm_object_drop(object);
1254 			return(VM_PAGER_OK);
1255 		}
1256 		m = vm_page_lookup_busy_try(object, mreq->pindex + 1,
1257 					    TRUE, &error);
1258 		if (error) {
1259 			vm_object_drop(object);
1260 			return(VM_PAGER_OK);
1261 		} else if (m == NULL) {
1262 			/*
1263 			 * Use VM_ALLOC_QUICK to avoid blocking on cache
1264 			 * page reuse.
1265 			 */
1266 			m = vm_page_alloc(object, mreq->pindex + 1,
1267 					  VM_ALLOC_QUICK);
1268 			if (m == NULL) {
1269 				vm_object_drop(object);
1270 				return(VM_PAGER_OK);
1271 			}
1272 		} else {
1273 			if (m->valid) {
1274 				vm_page_wakeup(m);
1275 				vm_object_drop(object);
1276 				return(VM_PAGER_OK);
1277 			}
1278 			vm_page_unqueue_nowakeup(m);
1279 		}
1280 		/* page is busy */
1281 		mreq = m;
1282 		raonly = 1;
1283 	} else {
1284 		raonly = 0;
1285 	}
1286 
1287 	/*
1288 	 * Try to block-read contiguous pages from swap if sequential,
1289 	 * otherwise just read one page.  Contiguous pages from swap must
1290 	 * reside within a single device stripe because the I/O cannot be
1291 	 * broken up across multiple stripes.
1292 	 *
1293 	 * Note that blk and iblk can be SWAPBLK_NONE but the loop is
1294 	 * set up such that the case(s) are handled implicitly.
1295 	 */
1296 	blk = swp_pager_meta_ctl(mreq->object, mreq->pindex, 0);
1297 	marray[0] = mreq;
1298 
1299 	for (i = 1; i <= swap_burst_read &&
1300 		    i < XIO_INTERNAL_PAGES &&
1301 		    mreq->pindex + i < object->size; ++i) {
1302 		swblk_t iblk;
1303 
1304 		iblk = swp_pager_meta_ctl(object, mreq->pindex + i, 0);
1305 		if (iblk != blk + i)
1306 			break;
1307 		if ((blk ^ iblk) & ~SWB_DMMASK)
1308 			break;
1309 		m = vm_page_lookup_busy_try(object, mreq->pindex + i,
1310 					    TRUE, &error);
1311 		if (error) {
1312 			break;
1313 		} else if (m == NULL) {
1314 			/*
1315 			 * Use VM_ALLOC_QUICK to avoid blocking on cache
1316 			 * page reuse.
1317 			 */
1318 			m = vm_page_alloc(object, mreq->pindex + i,
1319 					  VM_ALLOC_QUICK);
1320 			if (m == NULL)
1321 				break;
1322 		} else {
1323 			if (m->valid) {
1324 				vm_page_wakeup(m);
1325 				break;
1326 			}
1327 			vm_page_unqueue_nowakeup(m);
1328 		}
1329 		/* page is busy */
1330 		marray[i] = m;
1331 	}
1332 	if (i > 1)
1333 		vm_page_flag_set(marray[i - 1], PG_RAM);
1334 
1335 	/*
1336 	 * If mreq is the requested page and we have nothing to do return
1337 	 * VM_PAGER_FAIL.  If raonly is set mreq is just another read-ahead
1338 	 * page and must be cleaned up.
1339 	 */
1340 	if (blk == SWAPBLK_NONE) {
1341 		KKASSERT(i == 1);
1342 		if (raonly) {
1343 			vnode_pager_freepage(mreq);
1344 			vm_object_drop(object);
1345 			return(VM_PAGER_OK);
1346 		} else {
1347 			vm_object_drop(object);
1348 			return(VM_PAGER_FAIL);
1349 		}
1350 	}
1351 
1352 	/*
1353 	 * map our page(s) into kva for input
1354 	 */
1355 	bp = getpbuf_kva(&nsw_rcount);
1356 	bio = &bp->b_bio1;
1357 	kva = (vm_offset_t) bp->b_kvabase;
1358 	bcopy(marray, bp->b_xio.xio_pages, i * sizeof(vm_page_t));
1359 	pmap_qenter(kva, bp->b_xio.xio_pages, i);
1360 
1361 	bp->b_data = (caddr_t)kva;
1362 	bp->b_bcount = PAGE_SIZE * i;
1363 	bp->b_xio.xio_npages = i;
1364 	bio->bio_done = swp_pager_async_iodone;
1365 	bio->bio_offset = (off_t)blk << PAGE_SHIFT;
1366 	bio->bio_caller_info1.index = SWBIO_READ;
1367 
1368 	/*
1369 	 * Set index.  If raonly set the index beyond the array so all
1370 	 * the pages are treated the same, otherwise the original mreq is
1371 	 * at index 0.
1372 	 */
1373 	if (raonly)
1374 		bio->bio_driver_info = (void *)(intptr_t)i;
1375 	else
1376 		bio->bio_driver_info = (void *)(intptr_t)0;
1377 
1378 	for (j = 0; j < i; ++j)
1379 		vm_page_flag_set(bp->b_xio.xio_pages[j], PG_SWAPINPROG);
1380 
1381 	mycpu->gd_cnt.v_swapin++;
1382 	mycpu->gd_cnt.v_swappgsin += bp->b_xio.xio_npages;
1383 
1384 	/*
1385 	 * We still hold the lock on mreq, and our automatic completion routine
1386 	 * does not remove it.
1387 	 */
1388 	vm_object_pip_add(object, bp->b_xio.xio_npages);
1389 
1390 	/*
1391 	 * perform the I/O.  NOTE!!!  bp cannot be considered valid after
1392 	 * this point because we automatically release it on completion.
1393 	 * Instead, we look at the one page we are interested in which we
1394 	 * still hold a lock on even through the I/O completion.
1395 	 *
1396 	 * The other pages in our m[] array are also released on completion,
1397 	 * so we cannot assume they are valid anymore either.
1398 	 */
1399 	bp->b_cmd = BUF_CMD_READ;
1400 	BUF_KERNPROC(bp);
1401 	vn_strategy(swapdev_vp, bio);
1402 
1403 	/*
1404 	 * Wait for the page we want to complete.  PG_SWAPINPROG is always
1405 	 * cleared on completion.  If an I/O error occurs, SWAPBLK_NONE
1406 	 * is set in the meta-data.
1407 	 *
1408 	 * If this is a read-ahead only we return immediately without
1409 	 * waiting for I/O.
1410 	 */
1411 	if (raonly) {
1412 		vm_object_drop(object);
1413 		return(VM_PAGER_OK);
1414 	}
1415 
1416 	/*
1417 	 * Read-ahead includes originally requested page case.
1418 	 */
1419 	for (;;) {
1420 		flags = mreq->flags;
1421 		cpu_ccfence();
1422 		if ((flags & PG_SWAPINPROG) == 0)
1423 			break;
1424 		tsleep_interlock(mreq, 0);
1425 		if (!atomic_cmpset_int(&mreq->flags, flags,
1426 				       flags | PG_WANTED | PG_REFERENCED)) {
1427 			continue;
1428 		}
1429 		mycpu->gd_cnt.v_intrans++;
1430 		if (tsleep(mreq, PINTERLOCKED, "swread", hz*20)) {
1431 			kprintf(
1432 			    "swap_pager: indefinite wait buffer: "
1433 				" bp %p offset: %lld, size: %ld\n",
1434 			    bp,
1435 			    (long long)bio->bio_offset,
1436 			    (long)bp->b_bcount
1437 			);
1438 		}
1439 	}
1440 
1441 	/*
1442 	 * mreq is left bussied after completion, but all the other pages
1443 	 * are freed.  If we had an unrecoverable read error the page will
1444 	 * not be valid.
1445 	 */
1446 	vm_object_drop(object);
1447 	if (mreq->valid != VM_PAGE_BITS_ALL)
1448 		return(VM_PAGER_ERROR);
1449 	else
1450 		return(VM_PAGER_OK);
1451 
1452 	/*
1453 	 * A final note: in a low swap situation, we cannot deallocate swap
1454 	 * and mark a page dirty here because the caller is likely to mark
1455 	 * the page clean when we return, causing the page to possibly revert
1456 	 * to all-zero's later.
1457 	 */
1458 }
1459 
1460 /*
1461  *	swap_pager_putpages:
1462  *
1463  *	Assign swap (if necessary) and initiate I/O on the specified pages.
1464  *
1465  *	We support both OBJT_DEFAULT and OBJT_SWAP objects.  DEFAULT objects
1466  *	are automatically converted to SWAP objects.
1467  *
1468  *	In a low memory situation we may block in vn_strategy(), but the new
1469  *	vm_page reservation system coupled with properly written VFS devices
1470  *	should ensure that no low-memory deadlock occurs.  This is an area
1471  *	which needs work.
1472  *
1473  *	The parent has N vm_object_pip_add() references prior to
1474  *	calling us and will remove references for rtvals[] that are
1475  *	not set to VM_PAGER_PEND.  We need to remove the rest on I/O
1476  *	completion.
1477  *
1478  *	The parent has soft-busy'd the pages it passes us and will unbusy
1479  *	those whos rtvals[] entry is not set to VM_PAGER_PEND on return.
1480  *	We need to unbusy the rest on I/O completion.
1481  *
1482  * No requirements.
1483  */
1484 void
1485 swap_pager_putpages(vm_object_t object, vm_page_t *m, int count,
1486 		    int flags, int *rtvals)
1487 {
1488 	int i;
1489 	int n = 0;
1490 
1491 	vm_object_hold(object);
1492 
1493 	if (count && m[0]->object != object) {
1494 		panic("swap_pager_getpages: object mismatch %p/%p",
1495 		    object,
1496 		    m[0]->object
1497 		);
1498 	}
1499 
1500 	/*
1501 	 * Step 1
1502 	 *
1503 	 * Turn object into OBJT_SWAP
1504 	 * Check for bogus sysops
1505 	 *
1506 	 * Force sync if not pageout process, we don't want any single
1507 	 * non-pageout process to be able to hog the I/O subsystem!  This
1508 	 * can be overridden by setting.
1509 	 */
1510 	if (object->type == OBJT_DEFAULT) {
1511 		if (object->type == OBJT_DEFAULT)
1512 			swp_pager_meta_convert(object);
1513 	}
1514 
1515 	/*
1516 	 * Normally we force synchronous swap I/O if this is not the
1517 	 * pageout daemon to prevent any single user process limited
1518 	 * via RLIMIT_RSS from hogging swap write bandwidth.
1519 	 */
1520 	if (curthread != pagethread && swap_user_async == 0)
1521 		flags |= VM_PAGER_PUT_SYNC;
1522 
1523 	/*
1524 	 * Step 2
1525 	 *
1526 	 * Update nsw parameters from swap_async_max sysctl values.
1527 	 * Do not let the sysop crash the machine with bogus numbers.
1528 	 */
1529 	if (swap_async_max != nsw_wcount_async_max) {
1530 		int n;
1531 
1532 		/*
1533 		 * limit range
1534 		 */
1535 		if ((n = swap_async_max) > nswbuf_kva / 2)
1536 			n = nswbuf_kva / 2;
1537 		if (n < 1)
1538 			n = 1;
1539 		swap_async_max = n;
1540 
1541 		/*
1542 		 * Adjust difference ( if possible ).  If the current async
1543 		 * count is too low, we may not be able to make the adjustment
1544 		 * at this time.
1545 		 *
1546 		 * vm_token needed for nsw_wcount sleep interlock
1547 		 */
1548 		lwkt_gettoken(&vm_token);
1549 		n -= nsw_wcount_async_max;
1550 		if (nsw_wcount_async + n >= 0) {
1551 			nsw_wcount_async_max += n;
1552 			pbuf_adjcount(&nsw_wcount_async, n);
1553 		}
1554 		lwkt_reltoken(&vm_token);
1555 	}
1556 
1557 	/*
1558 	 * Step 3
1559 	 *
1560 	 * Assign swap blocks and issue I/O.  We reallocate swap on the fly.
1561 	 * The page is left dirty until the pageout operation completes
1562 	 * successfully.
1563 	 */
1564 
1565 	for (i = 0; i < count; i += n) {
1566 		struct buf *bp;
1567 		struct bio *bio;
1568 		swblk_t blk;
1569 		int j;
1570 
1571 		/*
1572 		 * Maximum I/O size is limited by a number of factors.
1573 		 */
1574 
1575 		n = min(BLIST_MAX_ALLOC, count - i);
1576 		n = min(n, nsw_cluster_max);
1577 
1578 		lwkt_gettoken(&vm_token);
1579 
1580 		/*
1581 		 * Get biggest block of swap we can.  If we fail, fall
1582 		 * back and try to allocate a smaller block.  Don't go
1583 		 * overboard trying to allocate space if it would overly
1584 		 * fragment swap.
1585 		 */
1586 		while (
1587 		    (blk = swp_pager_getswapspace(object, n)) == SWAPBLK_NONE &&
1588 		    n > 4
1589 		) {
1590 			n >>= 1;
1591 		}
1592 		if (blk == SWAPBLK_NONE) {
1593 			for (j = 0; j < n; ++j)
1594 				rtvals[i+j] = VM_PAGER_FAIL;
1595 			lwkt_reltoken(&vm_token);
1596 			continue;
1597 		}
1598 		if (vm_report_swap_allocs > 0) {
1599 			kprintf("swap_alloc %08jx,%d\n", (intmax_t)blk, n);
1600 			--vm_report_swap_allocs;
1601 		}
1602 
1603 		/*
1604 		 * The I/O we are constructing cannot cross a physical
1605 		 * disk boundry in the swap stripe.
1606 		 */
1607 		if ((blk ^ (blk + n)) & ~SWB_DMMASK) {
1608 			j = ((blk + SWB_DMMAX) & ~SWB_DMMASK) - blk;
1609 			swp_pager_freeswapspace(object, blk + j, n - j);
1610 			n = j;
1611 		}
1612 
1613 		/*
1614 		 * All I/O parameters have been satisfied, build the I/O
1615 		 * request and assign the swap space.
1616 		 */
1617 		if ((flags & VM_PAGER_PUT_SYNC))
1618 			bp = getpbuf_kva(&nsw_wcount_sync);
1619 		else
1620 			bp = getpbuf_kva(&nsw_wcount_async);
1621 		bio = &bp->b_bio1;
1622 
1623 		lwkt_reltoken(&vm_token);
1624 
1625 		pmap_qenter((vm_offset_t)bp->b_data, &m[i], n);
1626 
1627 		bp->b_bcount = PAGE_SIZE * n;
1628 		bio->bio_offset = (off_t)blk << PAGE_SHIFT;
1629 
1630 		for (j = 0; j < n; ++j) {
1631 			vm_page_t mreq = m[i+j];
1632 
1633 			swp_pager_meta_build(mreq->object, mreq->pindex,
1634 					     blk + j);
1635 			if (object->type == OBJT_SWAP)
1636 				vm_page_dirty(mreq);
1637 			rtvals[i+j] = VM_PAGER_OK;
1638 
1639 			vm_page_flag_set(mreq, PG_SWAPINPROG);
1640 			bp->b_xio.xio_pages[j] = mreq;
1641 		}
1642 		bp->b_xio.xio_npages = n;
1643 
1644 		mycpu->gd_cnt.v_swapout++;
1645 		mycpu->gd_cnt.v_swappgsout += bp->b_xio.xio_npages;
1646 
1647 		bp->b_dirtyoff = 0;		/* req'd for NFS */
1648 		bp->b_dirtyend = bp->b_bcount;	/* req'd for NFS */
1649 		bp->b_cmd = BUF_CMD_WRITE;
1650 		bio->bio_caller_info1.index = SWBIO_WRITE;
1651 
1652 		/*
1653 		 * asynchronous
1654 		 */
1655 		if ((flags & VM_PAGER_PUT_SYNC) == 0) {
1656 			bio->bio_done = swp_pager_async_iodone;
1657 			BUF_KERNPROC(bp);
1658 			vn_strategy(swapdev_vp, bio);
1659 
1660 			for (j = 0; j < n; ++j)
1661 				rtvals[i+j] = VM_PAGER_PEND;
1662 			continue;
1663 		}
1664 
1665 		/*
1666 		 * Issue synchrnously.
1667 		 *
1668 		 * Wait for the sync I/O to complete, then update rtvals.
1669 		 * We just set the rtvals[] to VM_PAGER_PEND so we can call
1670 		 * our async completion routine at the end, thus avoiding a
1671 		 * double-free.
1672 		 */
1673 		bio->bio_caller_info1.index |= SWBIO_SYNC;
1674 		if (flags & VM_PAGER_TRY_TO_CACHE)
1675 			bio->bio_caller_info1.index |= SWBIO_TTC;
1676 		bio->bio_done = biodone_sync;
1677 		bio->bio_flags |= BIO_SYNC;
1678 		vn_strategy(swapdev_vp, bio);
1679 		biowait(bio, "swwrt");
1680 
1681 		for (j = 0; j < n; ++j)
1682 			rtvals[i+j] = VM_PAGER_PEND;
1683 
1684 		/*
1685 		 * Now that we are through with the bp, we can call the
1686 		 * normal async completion, which frees everything up.
1687 		 */
1688 		swp_pager_async_iodone(bio);
1689 	}
1690 	vm_object_drop(object);
1691 }
1692 
1693 /*
1694  * No requirements.
1695  *
1696  * Recalculate the low and high-water marks.
1697  */
1698 void
1699 swap_pager_newswap(void)
1700 {
1701 	/*
1702 	 * NOTE: vm_swap_max cannot exceed 1 billion blocks, which is the
1703 	 *	 limitation imposed by the blist code.  Remember that this
1704 	 *	 will be divided by NSWAP_MAX (4), so each swap device is
1705 	 *	 limited to around a terrabyte.
1706 	 */
1707 	if (vm_swap_max) {
1708 		nswap_lowat = (int64_t)vm_swap_max * 4 / 100;	/* 4% left */
1709 		nswap_hiwat = (int64_t)vm_swap_max * 6 / 100;	/* 6% left */
1710 		kprintf("swap low/high-water marks set to %d/%d\n",
1711 			nswap_lowat, nswap_hiwat);
1712 	} else {
1713 		nswap_lowat = 128;
1714 		nswap_hiwat = 512;
1715 	}
1716 	swp_sizecheck();
1717 }
1718 
1719 /*
1720  *	swp_pager_async_iodone:
1721  *
1722  *	Completion routine for asynchronous reads and writes from/to swap.
1723  *	Also called manually by synchronous code to finish up a bp.
1724  *
1725  *	For READ operations, the pages are PG_BUSY'd.  For WRITE operations,
1726  *	the pages are vm_page_t->busy'd.  For READ operations, we PG_BUSY
1727  *	unbusy all pages except the 'main' request page.  For WRITE
1728  *	operations, we vm_page_t->busy'd unbusy all pages ( we can do this
1729  *	because we marked them all VM_PAGER_PEND on return from putpages ).
1730  *
1731  *	This routine may not block.
1732  *
1733  * No requirements.
1734  */
1735 static void
1736 swp_pager_async_iodone(struct bio *bio)
1737 {
1738 	struct buf *bp = bio->bio_buf;
1739 	vm_object_t object = NULL;
1740 	int i;
1741 	int *nswptr;
1742 
1743 	/*
1744 	 * report error
1745 	 */
1746 	if (bp->b_flags & B_ERROR) {
1747 		kprintf(
1748 		    "swap_pager: I/O error - %s failed; offset %lld,"
1749 			"size %ld, error %d\n",
1750 		    ((bio->bio_caller_info1.index & SWBIO_READ) ?
1751 			"pagein" : "pageout"),
1752 		    (long long)bio->bio_offset,
1753 		    (long)bp->b_bcount,
1754 		    bp->b_error
1755 		);
1756 	}
1757 
1758 	/*
1759 	 * set object.
1760 	 */
1761 	if (bp->b_xio.xio_npages)
1762 		object = bp->b_xio.xio_pages[0]->object;
1763 
1764 	/*
1765 	 * remove the mapping for kernel virtual
1766 	 */
1767 	pmap_qremove((vm_offset_t)bp->b_data, bp->b_xio.xio_npages);
1768 
1769 	/*
1770 	 * cleanup pages.  If an error occurs writing to swap, we are in
1771 	 * very serious trouble.  If it happens to be a disk error, though,
1772 	 * we may be able to recover by reassigning the swap later on.  So
1773 	 * in this case we remove the m->swapblk assignment for the page
1774 	 * but do not free it in the rlist.  The errornous block(s) are thus
1775 	 * never reallocated as swap.  Redirty the page and continue.
1776 	 */
1777 	for (i = 0; i < bp->b_xio.xio_npages; ++i) {
1778 		vm_page_t m = bp->b_xio.xio_pages[i];
1779 
1780 		if (bp->b_flags & B_ERROR) {
1781 			/*
1782 			 * If an error occurs I'd love to throw the swapblk
1783 			 * away without freeing it back to swapspace, so it
1784 			 * can never be used again.  But I can't from an
1785 			 * interrupt.
1786 			 */
1787 
1788 			if (bio->bio_caller_info1.index & SWBIO_READ) {
1789 				/*
1790 				 * When reading, reqpage needs to stay
1791 				 * locked for the parent, but all other
1792 				 * pages can be freed.  We still want to
1793 				 * wakeup the parent waiting on the page,
1794 				 * though.  ( also: pg_reqpage can be -1 and
1795 				 * not match anything ).
1796 				 *
1797 				 * We have to wake specifically requested pages
1798 				 * up too because we cleared PG_SWAPINPROG and
1799 				 * someone may be waiting for that.
1800 				 *
1801 				 * NOTE: for reads, m->dirty will probably
1802 				 * be overridden by the original caller of
1803 				 * getpages so don't play cute tricks here.
1804 				 *
1805 				 * NOTE: We can't actually free the page from
1806 				 * here, because this is an interrupt.  It
1807 				 * is not legal to mess with object->memq
1808 				 * from an interrupt.  Deactivate the page
1809 				 * instead.
1810 				 */
1811 
1812 				m->valid = 0;
1813 				vm_page_flag_clear(m, PG_SWAPINPROG);
1814 
1815 				/*
1816 				 * bio_driver_info holds the requested page
1817 				 * index.
1818 				 */
1819 				if (i != (int)(intptr_t)bio->bio_driver_info) {
1820 					vm_page_deactivate(m);
1821 					vm_page_wakeup(m);
1822 				} else {
1823 					vm_page_flash(m);
1824 				}
1825 				/*
1826 				 * If i == bp->b_pager.pg_reqpage, do not wake
1827 				 * the page up.  The caller needs to.
1828 				 */
1829 			} else {
1830 				/*
1831 				 * If a write error occurs remove the swap
1832 				 * assignment (note that PG_SWAPPED may or
1833 				 * may not be set depending on prior activity).
1834 				 *
1835 				 * Re-dirty OBJT_SWAP pages as there is no
1836 				 * other backing store, we can't throw the
1837 				 * page away.
1838 				 *
1839 				 * Non-OBJT_SWAP pages (aka swapcache) must
1840 				 * not be dirtied since they may not have
1841 				 * been dirty in the first place, and they
1842 				 * do have backing store (the vnode).
1843 				 */
1844 				vm_page_busy_wait(m, FALSE, "swadpg");
1845 				swp_pager_meta_ctl(m->object, m->pindex,
1846 						   SWM_FREE);
1847 				vm_page_flag_clear(m, PG_SWAPPED);
1848 				if (m->object->type == OBJT_SWAP) {
1849 					vm_page_dirty(m);
1850 					vm_page_activate(m);
1851 				}
1852 				vm_page_flag_clear(m, PG_SWAPINPROG);
1853 				vm_page_io_finish(m);
1854 				vm_page_wakeup(m);
1855 			}
1856 		} else if (bio->bio_caller_info1.index & SWBIO_READ) {
1857 			/*
1858 			 * NOTE: for reads, m->dirty will probably be
1859 			 * overridden by the original caller of getpages so
1860 			 * we cannot set them in order to free the underlying
1861 			 * swap in a low-swap situation.  I don't think we'd
1862 			 * want to do that anyway, but it was an optimization
1863 			 * that existed in the old swapper for a time before
1864 			 * it got ripped out due to precisely this problem.
1865 			 *
1866 			 * If not the requested page then deactivate it.
1867 			 *
1868 			 * Note that the requested page, reqpage, is left
1869 			 * busied, but we still have to wake it up.  The
1870 			 * other pages are released (unbusied) by
1871 			 * vm_page_wakeup().  We do not set reqpage's
1872 			 * valid bits here, it is up to the caller.
1873 			 */
1874 
1875 			/*
1876 			 * NOTE: can't call pmap_clear_modify(m) from an
1877 			 * interrupt thread, the pmap code may have to map
1878 			 * non-kernel pmaps and currently asserts the case.
1879 			 */
1880 			/*pmap_clear_modify(m);*/
1881 			m->valid = VM_PAGE_BITS_ALL;
1882 			vm_page_undirty(m);
1883 			vm_page_flag_clear(m, PG_SWAPINPROG);
1884 			vm_page_flag_set(m, PG_SWAPPED);
1885 
1886 			/*
1887 			 * We have to wake specifically requested pages
1888 			 * up too because we cleared PG_SWAPINPROG and
1889 			 * could be waiting for it in getpages.  However,
1890 			 * be sure to not unbusy getpages specifically
1891 			 * requested page - getpages expects it to be
1892 			 * left busy.
1893 			 *
1894 			 * bio_driver_info holds the requested page
1895 			 */
1896 			if (i != (int)(intptr_t)bio->bio_driver_info) {
1897 				vm_page_deactivate(m);
1898 				vm_page_wakeup(m);
1899 			} else {
1900 				vm_page_flash(m);
1901 			}
1902 		} else {
1903 			/*
1904 			 * Mark the page clean but do not mess with the
1905 			 * pmap-layer's modified state.  That state should
1906 			 * also be clear since the caller protected the
1907 			 * page VM_PROT_READ, but allow the case.
1908 			 *
1909 			 * We are in an interrupt, avoid pmap operations.
1910 			 *
1911 			 * If we have a severe page deficit, deactivate the
1912 			 * page.  Do not try to cache it (which would also
1913 			 * involve a pmap op), because the page might still
1914 			 * be read-heavy.
1915 			 *
1916 			 * When using the swap to cache clean vnode pages
1917 			 * we do not mess with the page dirty bits.
1918 			 */
1919 			vm_page_busy_wait(m, FALSE, "swadpg");
1920 			if (m->object->type == OBJT_SWAP)
1921 				vm_page_undirty(m);
1922 			vm_page_flag_clear(m, PG_SWAPINPROG);
1923 			vm_page_flag_set(m, PG_SWAPPED);
1924 			if (vm_page_count_severe())
1925 				vm_page_deactivate(m);
1926 			vm_page_io_finish(m);
1927 			if (bio->bio_caller_info1.index & SWBIO_TTC)
1928 				vm_page_try_to_cache(m);
1929 			else
1930 				vm_page_wakeup(m);
1931 		}
1932 	}
1933 
1934 	/*
1935 	 * adjust pip.  NOTE: the original parent may still have its own
1936 	 * pip refs on the object.
1937 	 */
1938 
1939 	if (object)
1940 		vm_object_pip_wakeup_n(object, bp->b_xio.xio_npages);
1941 
1942 	/*
1943 	 * Release the physical I/O buffer.
1944 	 *
1945 	 * NOTE: Due to synchronous operations in the write case b_cmd may
1946 	 *	 already be set to BUF_CMD_DONE and BIO_SYNC may have already
1947 	 *	 been cleared.
1948 	 *
1949 	 * Use vm_token to interlock nsw_rcount/wcount wakeup?
1950 	 */
1951 	lwkt_gettoken(&vm_token);
1952 	if (bio->bio_caller_info1.index & SWBIO_READ)
1953 		nswptr = &nsw_rcount;
1954 	else if (bio->bio_caller_info1.index & SWBIO_SYNC)
1955 		nswptr = &nsw_wcount_sync;
1956 	else
1957 		nswptr = &nsw_wcount_async;
1958 	bp->b_cmd = BUF_CMD_DONE;
1959 	relpbuf(bp, nswptr);
1960 	lwkt_reltoken(&vm_token);
1961 }
1962 
1963 /*
1964  * Fault-in a potentially swapped page and remove the swap reference.
1965  * (used by swapoff code)
1966  *
1967  * object must be held.
1968  */
1969 static __inline void
1970 swp_pager_fault_page(vm_object_t object, int *sharedp, vm_pindex_t pindex)
1971 {
1972 	struct vnode *vp;
1973 	vm_page_t m;
1974 	int error;
1975 
1976 	ASSERT_LWKT_TOKEN_HELD(vm_object_token(object));
1977 
1978 	if (object->type == OBJT_VNODE) {
1979 		/*
1980 		 * Any swap related to a vnode is due to swapcache.  We must
1981 		 * vget() the vnode in case it is not active (otherwise
1982 		 * vref() will panic).  Calling vm_object_page_remove() will
1983 		 * ensure that any swap ref is removed interlocked with the
1984 		 * page.  clean_only is set to TRUE so we don't throw away
1985 		 * dirty pages.
1986 		 */
1987 		vp = object->handle;
1988 		error = vget(vp, LK_SHARED | LK_RETRY | LK_CANRECURSE);
1989 		if (error == 0) {
1990 			vm_object_page_remove(object, pindex, pindex + 1, TRUE);
1991 			vput(vp);
1992 		}
1993 	} else {
1994 		/*
1995 		 * Otherwise it is a normal OBJT_SWAP object and we can
1996 		 * fault the page in and remove the swap.
1997 		 */
1998 		m = vm_fault_object_page(object, IDX_TO_OFF(pindex),
1999 					 VM_PROT_NONE,
2000 					 VM_FAULT_DIRTY | VM_FAULT_UNSWAP,
2001 					 sharedp, &error);
2002 		if (m)
2003 			vm_page_unhold(m);
2004 	}
2005 }
2006 
2007 /*
2008  * This removes all swap blocks related to a particular device.  We have
2009  * to be careful of ripups during the scan.
2010  */
2011 static int swp_pager_swapoff_callback(struct swblock *swap, void *data);
2012 
2013 int
2014 swap_pager_swapoff(int devidx)
2015 {
2016 	struct vm_object_hash *hash;
2017 	struct swswapoffinfo info;
2018 	struct vm_object marker;
2019 	vm_object_t object;
2020 	int n;
2021 
2022 	bzero(&marker, sizeof(marker));
2023 	marker.type = OBJT_MARKER;
2024 
2025 	for (n = 0; n < VMOBJ_HSIZE; ++n) {
2026 		hash = &vm_object_hash[n];
2027 
2028 		lwkt_gettoken(&hash->token);
2029 		TAILQ_INSERT_HEAD(&hash->list, &marker, object_list);
2030 
2031 		while ((object = TAILQ_NEXT(&marker, object_list)) != NULL) {
2032 			if (object->type == OBJT_MARKER)
2033 				goto skip;
2034 			if (object->type != OBJT_SWAP &&
2035 			    object->type != OBJT_VNODE)
2036 				goto skip;
2037 			vm_object_hold(object);
2038 			if (object->type != OBJT_SWAP &&
2039 			    object->type != OBJT_VNODE) {
2040 				vm_object_drop(object);
2041 				goto skip;
2042 			}
2043 			info.object = object;
2044 			info.shared = 0;
2045 			info.devidx = devidx;
2046 			swblock_rb_tree_RB_SCAN(&object->swblock_root,
2047 					    NULL, swp_pager_swapoff_callback,
2048 					    &info);
2049 			vm_object_drop(object);
2050 skip:
2051 			if (object == TAILQ_NEXT(&marker, object_list)) {
2052 				TAILQ_REMOVE(&hash->list, &marker, object_list);
2053 				TAILQ_INSERT_AFTER(&hash->list, object,
2054 						   &marker, object_list);
2055 			}
2056 		}
2057 		TAILQ_REMOVE(&hash->list, &marker, object_list);
2058 		lwkt_reltoken(&hash->token);
2059 	}
2060 
2061 	/*
2062 	 * If we fail to locate all swblocks we just fail gracefully and
2063 	 * do not bother to restore paging on the swap device.  If the
2064 	 * user wants to retry the user can retry.
2065 	 */
2066 	if (swdevt[devidx].sw_nused)
2067 		return (1);
2068 	else
2069 		return (0);
2070 }
2071 
2072 static
2073 int
2074 swp_pager_swapoff_callback(struct swblock *swap, void *data)
2075 {
2076 	struct swswapoffinfo *info = data;
2077 	vm_object_t object = info->object;
2078 	vm_pindex_t index;
2079 	swblk_t v;
2080 	int i;
2081 
2082 	index = swap->swb_index;
2083 	for (i = 0; i < SWAP_META_PAGES; ++i) {
2084 		/*
2085 		 * Make sure we don't race a dying object.  This will
2086 		 * kill the scan of the object's swap blocks entirely.
2087 		 */
2088 		if (object->flags & OBJ_DEAD)
2089 			return(-1);
2090 
2091 		/*
2092 		 * Fault the page, which can obviously block.  If the swap
2093 		 * structure disappears break out.
2094 		 */
2095 		v = swap->swb_pages[i];
2096 		if (v != SWAPBLK_NONE && BLK2DEVIDX(v) == info->devidx) {
2097 			swp_pager_fault_page(object, &info->shared,
2098 					     swap->swb_index + i);
2099 			/* swap ptr might go away */
2100 			if (RB_LOOKUP(swblock_rb_tree,
2101 				      &object->swblock_root, index) != swap) {
2102 				break;
2103 			}
2104 		}
2105 	}
2106 	return(0);
2107 }
2108 
2109 /************************************************************************
2110  *				SWAP META DATA 				*
2111  ************************************************************************
2112  *
2113  *	These routines manipulate the swap metadata stored in the
2114  *	OBJT_SWAP object.
2115  *
2116  *	Swap metadata is implemented with a global hash and not directly
2117  *	linked into the object.  Instead the object simply contains
2118  *	appropriate tracking counters.
2119  */
2120 
2121 /*
2122  * Lookup the swblock containing the specified swap block index.
2123  *
2124  * The caller must hold the object.
2125  */
2126 static __inline
2127 struct swblock *
2128 swp_pager_lookup(vm_object_t object, vm_pindex_t index)
2129 {
2130 	ASSERT_LWKT_TOKEN_HELD(vm_object_token(object));
2131 	index &= ~(vm_pindex_t)SWAP_META_MASK;
2132 	return (RB_LOOKUP(swblock_rb_tree, &object->swblock_root, index));
2133 }
2134 
2135 /*
2136  * Remove a swblock from the RB tree.
2137  *
2138  * The caller must hold the object.
2139  */
2140 static __inline
2141 void
2142 swp_pager_remove(vm_object_t object, struct swblock *swap)
2143 {
2144 	ASSERT_LWKT_TOKEN_HELD(vm_object_token(object));
2145 	RB_REMOVE(swblock_rb_tree, &object->swblock_root, swap);
2146 }
2147 
2148 /*
2149  * Convert default object to swap object if necessary
2150  *
2151  * The caller must hold the object.
2152  */
2153 static void
2154 swp_pager_meta_convert(vm_object_t object)
2155 {
2156 	if (object->type == OBJT_DEFAULT) {
2157 		object->type = OBJT_SWAP;
2158 		KKASSERT(object->swblock_count == 0);
2159 	}
2160 }
2161 
2162 /*
2163  * SWP_PAGER_META_BUILD() -	add swap block to swap meta data for object
2164  *
2165  *	We first convert the object to a swap object if it is a default
2166  *	object.  Vnode objects do not need to be converted.
2167  *
2168  *	The specified swapblk is added to the object's swap metadata.  If
2169  *	the swapblk is not valid, it is freed instead.  Any previously
2170  *	assigned swapblk is freed.
2171  *
2172  * The caller must hold the object.
2173  */
2174 static void
2175 swp_pager_meta_build(vm_object_t object, vm_pindex_t index, swblk_t swapblk)
2176 {
2177 	struct swblock *swap;
2178 	struct swblock *oswap;
2179 	vm_pindex_t v;
2180 
2181 	KKASSERT(swapblk != SWAPBLK_NONE);
2182 	ASSERT_LWKT_TOKEN_HELD(vm_object_token(object));
2183 
2184 	/*
2185 	 * Convert object if necessary
2186 	 */
2187 	if (object->type == OBJT_DEFAULT)
2188 		swp_pager_meta_convert(object);
2189 
2190 	/*
2191 	 * Locate swblock.  If not found create, but if we aren't adding
2192 	 * anything just return.  If we run out of space in the map we wait
2193 	 * and, since the hash table may have changed, retry.
2194 	 */
2195 retry:
2196 	swap = swp_pager_lookup(object, index);
2197 
2198 	if (swap == NULL) {
2199 		int i;
2200 
2201 		swap = zalloc(swap_zone);
2202 		if (swap == NULL) {
2203 			vm_wait(0);
2204 			goto retry;
2205 		}
2206 		swap->swb_index = index & ~(vm_pindex_t)SWAP_META_MASK;
2207 		swap->swb_count = 0;
2208 
2209 		++object->swblock_count;
2210 
2211 		for (i = 0; i < SWAP_META_PAGES; ++i)
2212 			swap->swb_pages[i] = SWAPBLK_NONE;
2213 		oswap = RB_INSERT(swblock_rb_tree, &object->swblock_root, swap);
2214 		KKASSERT(oswap == NULL);
2215 	}
2216 
2217 	/*
2218 	 * Delete prior contents of metadata.
2219 	 *
2220 	 * NOTE: Decrement swb_count after the freeing operation (which
2221 	 *	 might block) to prevent racing destruction of the swblock.
2222 	 */
2223 	index &= SWAP_META_MASK;
2224 
2225 	while ((v = swap->swb_pages[index]) != SWAPBLK_NONE) {
2226 		swap->swb_pages[index] = SWAPBLK_NONE;
2227 		/* can block */
2228 		swp_pager_freeswapspace(object, v, 1);
2229 		--swap->swb_count;
2230 		--mycpu->gd_vmtotal.t_vm;
2231 	}
2232 
2233 	/*
2234 	 * Enter block into metadata
2235 	 */
2236 	swap->swb_pages[index] = swapblk;
2237 	if (swapblk != SWAPBLK_NONE) {
2238 		++swap->swb_count;
2239 		++mycpu->gd_vmtotal.t_vm;
2240 	}
2241 }
2242 
2243 /*
2244  * SWP_PAGER_META_FREE() - free a range of blocks in the object's swap metadata
2245  *
2246  *	The requested range of blocks is freed, with any associated swap
2247  *	returned to the swap bitmap.
2248  *
2249  *	This routine will free swap metadata structures as they are cleaned
2250  *	out.  This routine does *NOT* operate on swap metadata associated
2251  *	with resident pages.
2252  *
2253  * The caller must hold the object.
2254  */
2255 static int swp_pager_meta_free_callback(struct swblock *swb, void *data);
2256 
2257 static void
2258 swp_pager_meta_free(vm_object_t object, vm_pindex_t index, vm_pindex_t count)
2259 {
2260 	struct swfreeinfo info;
2261 
2262 	ASSERT_LWKT_TOKEN_HELD(vm_object_token(object));
2263 
2264 	/*
2265 	 * Nothing to do
2266 	 */
2267 	if (object->swblock_count == 0) {
2268 		KKASSERT(RB_EMPTY(&object->swblock_root));
2269 		return;
2270 	}
2271 	if (count == 0)
2272 		return;
2273 
2274 	/*
2275 	 * Setup for RB tree scan.  Note that the pindex range can be huge
2276 	 * due to the 64 bit page index space so we cannot safely iterate.
2277 	 */
2278 	info.object = object;
2279 	info.basei = index & ~(vm_pindex_t)SWAP_META_MASK;
2280 	info.begi = index;
2281 	info.endi = index + count - 1;
2282 	swblock_rb_tree_RB_SCAN(&object->swblock_root, rb_swblock_scancmp,
2283 				swp_pager_meta_free_callback, &info);
2284 }
2285 
2286 /*
2287  * The caller must hold the object.
2288  */
2289 static
2290 int
2291 swp_pager_meta_free_callback(struct swblock *swap, void *data)
2292 {
2293 	struct swfreeinfo *info = data;
2294 	vm_object_t object = info->object;
2295 	int index;
2296 	int eindex;
2297 
2298 	/*
2299 	 * Figure out the range within the swblock.  The wider scan may
2300 	 * return edge-case swap blocks when the start and/or end points
2301 	 * are in the middle of a block.
2302 	 */
2303 	if (swap->swb_index < info->begi)
2304 		index = (int)info->begi & SWAP_META_MASK;
2305 	else
2306 		index = 0;
2307 
2308 	if (swap->swb_index + SWAP_META_PAGES > info->endi)
2309 		eindex = (int)info->endi & SWAP_META_MASK;
2310 	else
2311 		eindex = SWAP_META_MASK;
2312 
2313 	/*
2314 	 * Scan and free the blocks.  The loop terminates early
2315 	 * if (swap) runs out of blocks and could be freed.
2316 	 *
2317 	 * NOTE: Decrement swb_count after swp_pager_freeswapspace()
2318 	 *	 to deal with a zfree race.
2319 	 */
2320 	while (index <= eindex) {
2321 		swblk_t v = swap->swb_pages[index];
2322 
2323 		if (v != SWAPBLK_NONE) {
2324 			swap->swb_pages[index] = SWAPBLK_NONE;
2325 			/* can block */
2326 			swp_pager_freeswapspace(object, v, 1);
2327 			--mycpu->gd_vmtotal.t_vm;
2328 			if (--swap->swb_count == 0) {
2329 				swp_pager_remove(object, swap);
2330 				zfree(swap_zone, swap);
2331 				--object->swblock_count;
2332 				break;
2333 			}
2334 		}
2335 		++index;
2336 	}
2337 
2338 	/* swap may be invalid here due to zfree above */
2339 	lwkt_yield();
2340 
2341 	return(0);
2342 }
2343 
2344 /*
2345  * SWP_PAGER_META_FREE_ALL() - destroy all swap metadata associated with object
2346  *
2347  *	This routine locates and destroys all swap metadata associated with
2348  *	an object.
2349  *
2350  * NOTE: Decrement swb_count after the freeing operation (which
2351  *	 might block) to prevent racing destruction of the swblock.
2352  *
2353  * The caller must hold the object.
2354  */
2355 static void
2356 swp_pager_meta_free_all(vm_object_t object)
2357 {
2358 	struct swblock *swap;
2359 	int i;
2360 
2361 	ASSERT_LWKT_TOKEN_HELD(vm_object_token(object));
2362 
2363 	while ((swap = RB_ROOT(&object->swblock_root)) != NULL) {
2364 		swp_pager_remove(object, swap);
2365 		for (i = 0; i < SWAP_META_PAGES; ++i) {
2366 			swblk_t v = swap->swb_pages[i];
2367 			if (v != SWAPBLK_NONE) {
2368 				/* can block */
2369 				swp_pager_freeswapspace(object, v, 1);
2370 				--swap->swb_count;
2371 				--mycpu->gd_vmtotal.t_vm;
2372 			}
2373 		}
2374 		if (swap->swb_count != 0)
2375 			panic("swap_pager_meta_free_all: swb_count != 0");
2376 		zfree(swap_zone, swap);
2377 		--object->swblock_count;
2378 		lwkt_yield();
2379 	}
2380 	KKASSERT(object->swblock_count == 0);
2381 }
2382 
2383 /*
2384  * SWP_PAGER_METACTL() -  misc control of swap and vm_page_t meta data.
2385  *
2386  *	This routine is capable of looking up, popping, or freeing
2387  *	swapblk assignments in the swap meta data or in the vm_page_t.
2388  *	The routine typically returns the swapblk being looked-up, or popped,
2389  *	or SWAPBLK_NONE if the block was freed, or SWAPBLK_NONE if the block
2390  *	was invalid.  This routine will automatically free any invalid
2391  *	meta-data swapblks.
2392  *
2393  *	It is not possible to store invalid swapblks in the swap meta data
2394  *	(other then a literal 'SWAPBLK_NONE'), so we don't bother checking.
2395  *
2396  *	When acting on a busy resident page and paging is in progress, we
2397  *	have to wait until paging is complete but otherwise can act on the
2398  *	busy page.
2399  *
2400  *	SWM_FREE	remove and free swap block from metadata
2401  *	SWM_POP		remove from meta data but do not free.. pop it out
2402  *
2403  * The caller must hold the object.
2404  */
2405 static swblk_t
2406 swp_pager_meta_ctl(vm_object_t object, vm_pindex_t index, int flags)
2407 {
2408 	struct swblock *swap;
2409 	swblk_t r1;
2410 
2411 	if (object->swblock_count == 0)
2412 		return(SWAPBLK_NONE);
2413 
2414 	r1 = SWAPBLK_NONE;
2415 	swap = swp_pager_lookup(object, index);
2416 
2417 	if (swap != NULL) {
2418 		index &= SWAP_META_MASK;
2419 		r1 = swap->swb_pages[index];
2420 
2421 		if (r1 != SWAPBLK_NONE) {
2422 			if (flags & (SWM_FREE|SWM_POP)) {
2423 				swap->swb_pages[index] = SWAPBLK_NONE;
2424 				--mycpu->gd_vmtotal.t_vm;
2425 				if (--swap->swb_count == 0) {
2426 					swp_pager_remove(object, swap);
2427 					zfree(swap_zone, swap);
2428 					--object->swblock_count;
2429 				}
2430 			}
2431 			/* swap ptr may be invalid */
2432 			if (flags & SWM_FREE) {
2433 				swp_pager_freeswapspace(object, r1, 1);
2434 				r1 = SWAPBLK_NONE;
2435 			}
2436 		}
2437 		/* swap ptr may be invalid */
2438 	}
2439 	return(r1);
2440 }
2441