xref: /freebsd/sys/vm/swap_pager.c (revision d429ea332342fcb98d27a350d0c4944bf9aec3f9)
1 /*-
2  * Copyright (c) 1998 Matthew Dillon,
3  * Copyright (c) 1994 John S. Dyson
4  * Copyright (c) 1990 University of Utah.
5  * Copyright (c) 1982, 1986, 1989, 1993
6  *	The Regents of the University of California.  All rights reserved.
7  *
8  * This code is derived from software contributed to Berkeley by
9  * the Systems Programming Group of the University of Utah Computer
10  * Science Department.
11  *
12  * Redistribution and use in source and binary forms, with or without
13  * modification, are permitted provided that the following conditions
14  * are met:
15  * 1. Redistributions of source code must retain the above copyright
16  *    notice, this list of conditions and the following disclaimer.
17  * 2. Redistributions in binary form must reproduce the above copyright
18  *    notice, this list of conditions and the following disclaimer in the
19  *    documentation and/or other materials provided with the distribution.
20  * 3. All advertising materials mentioning features or use of this software
21  *    must display the following acknowledgement:
22  *	This product includes software developed by the University of
23  *	California, Berkeley and its contributors.
24  * 4. Neither the name of the University nor the names of its contributors
25  *    may be used to endorse or promote products derived from this software
26  *    without specific prior written permission.
27  *
28  * THIS SOFTWARE IS PROVIDED BY THE REGENTS AND CONTRIBUTORS ``AS IS'' AND
29  * ANY EXPRESS OR IMPLIED WARRANTIES, INCLUDING, BUT NOT LIMITED TO, THE
30  * IMPLIED WARRANTIES OF MERCHANTABILITY AND FITNESS FOR A PARTICULAR PURPOSE
31  * ARE DISCLAIMED.  IN NO EVENT SHALL THE REGENTS OR CONTRIBUTORS BE LIABLE
32  * FOR ANY DIRECT, INDIRECT, INCIDENTAL, SPECIAL, EXEMPLARY, OR CONSEQUENTIAL
33  * DAMAGES (INCLUDING, BUT NOT LIMITED TO, PROCUREMENT OF SUBSTITUTE GOODS
34  * OR SERVICES; LOSS OF USE, DATA, OR PROFITS; OR BUSINESS INTERRUPTION)
35  * HOWEVER CAUSED AND ON ANY THEORY OF LIABILITY, WHETHER IN CONTRACT, STRICT
36  * LIABILITY, OR TORT (INCLUDING NEGLIGENCE OR OTHERWISE) ARISING IN ANY WAY
37  * OUT OF THE USE OF THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF
38  * SUCH DAMAGE.
39  *
40  *				New Swap System
41  *				Matthew Dillon
42  *
43  * Radix Bitmap 'blists'.
44  *
45  *	- The new swapper uses the new radix bitmap code.  This should scale
46  *	  to arbitrarily small or arbitrarily large swap spaces and an almost
47  *	  arbitrary degree of fragmentation.
48  *
49  * Features:
50  *
51  *	- on the fly reallocation of swap during putpages.  The new system
52  *	  does not try to keep previously allocated swap blocks for dirty
53  *	  pages.
54  *
55  *	- on the fly deallocation of swap
56  *
57  *	- No more garbage collection required.  Unnecessarily allocated swap
58  *	  blocks only exist for dirty vm_page_t's now and these are already
59  *	  cycled (in a high-load system) by the pager.  We also do on-the-fly
60  *	  removal of invalidated swap blocks when a page is destroyed
61  *	  or renamed.
62  *
63  * from: Utah $Hdr: swap_pager.c 1.4 91/04/30$
64  *
65  *	@(#)swap_pager.c	8.9 (Berkeley) 3/21/94
66  *	@(#)vm_swap.c	8.5 (Berkeley) 2/17/94
67  */
68 
69 #include <sys/cdefs.h>
70 __FBSDID("$FreeBSD$");
71 
72 #include "opt_mac.h"
73 #include "opt_swap.h"
74 #include "opt_vm.h"
75 
76 #include <sys/param.h>
77 #include <sys/systm.h>
78 #include <sys/conf.h>
79 #include <sys/kernel.h>
80 #include <sys/proc.h>
81 #include <sys/bio.h>
82 #include <sys/buf.h>
83 #include <sys/disk.h>
84 #include <sys/fcntl.h>
85 #include <sys/mount.h>
86 #include <sys/namei.h>
87 #include <sys/vnode.h>
88 #include <sys/mac.h>
89 #include <sys/malloc.h>
90 #include <sys/sysctl.h>
91 #include <sys/sysproto.h>
92 #include <sys/blist.h>
93 #include <sys/lock.h>
94 #include <sys/sx.h>
95 #include <sys/vmmeter.h>
96 
97 #include <vm/vm.h>
98 #include <vm/pmap.h>
99 #include <vm/vm_map.h>
100 #include <vm/vm_kern.h>
101 #include <vm/vm_object.h>
102 #include <vm/vm_page.h>
103 #include <vm/vm_pager.h>
104 #include <vm/vm_pageout.h>
105 #include <vm/vm_param.h>
106 #include <vm/swap_pager.h>
107 #include <vm/vm_extern.h>
108 #include <vm/uma.h>
109 
110 #include <geom/geom.h>
111 
112 /*
113  * SWB_NPAGES must be a power of 2.  It may be set to 1, 2, 4, 8, or 16
114  * pages per allocation.  We recommend you stick with the default of 8.
115  * The 16-page limit is due to the radix code (kern/subr_blist.c).
116  */
117 #ifndef MAX_PAGEOUT_CLUSTER
118 #define MAX_PAGEOUT_CLUSTER 16
119 #endif
120 
121 #if !defined(SWB_NPAGES)
122 #define SWB_NPAGES	MAX_PAGEOUT_CLUSTER
123 #endif
124 
125 /*
126  * Piecemeal swap metadata structure.  Swap is stored in a radix tree.
127  *
128  * If SWB_NPAGES is 8 and sizeof(char *) == sizeof(daddr_t), our radix
129  * is basically 8.  Assuming PAGE_SIZE == 4096, one tree level represents
130  * 32K worth of data, two levels represent 256K, three levels represent
131  * 2 MBytes.   This is acceptable.
132  *
133  * Overall memory utilization is about the same as the old swap structure.
134  */
135 #define SWCORRECT(n) (sizeof(void *) * (n) / sizeof(daddr_t))
136 #define SWAP_META_PAGES		(SWB_NPAGES * 2)
137 #define SWAP_META_MASK		(SWAP_META_PAGES - 1)
138 
139 typedef	int32_t	swblk_t;	/*
140 				 * swap offset.  This is the type used to
141 				 * address the "virtual swap device" and
142 				 * therefore the maximum swap space is
143 				 * 2^32 pages.
144 				 */
145 
146 struct swdevt;
147 typedef void sw_strategy_t(struct buf *bp, struct swdevt *sw);
148 typedef void sw_close_t(struct thread *td, struct swdevt *sw);
149 
150 /*
151  * Swap device table
152  */
153 struct swdevt {
154 	int	sw_flags;
155 	int	sw_nblks;
156 	int     sw_used;
157 	dev_t	sw_dev;
158 	struct vnode *sw_vp;
159 	void	*sw_id;
160 	swblk_t	sw_first;
161 	swblk_t	sw_end;
162 	struct blist *sw_blist;
163 	TAILQ_ENTRY(swdevt)	sw_list;
164 	sw_strategy_t		*sw_strategy;
165 	sw_close_t		*sw_close;
166 };
167 
168 #define	SW_CLOSING	0x04
169 
170 struct swblock {
171 	struct swblock	*swb_hnext;
172 	vm_object_t	swb_object;
173 	vm_pindex_t	swb_index;
174 	int		swb_count;
175 	daddr_t		swb_pages[SWAP_META_PAGES];
176 };
177 
178 static struct mtx sw_dev_mtx;
179 static TAILQ_HEAD(, swdevt) swtailq = TAILQ_HEAD_INITIALIZER(swtailq);
180 static struct swdevt *swdevhd;	/* Allocate from here next */
181 static int nswapdev;		/* Number of swap devices */
182 int swap_pager_avail;
183 static int swdev_syscall_active = 0; /* serialize swap(on|off) */
184 
185 static void swapdev_strategy(struct buf *, struct swdevt *sw);
186 
187 #define SWM_FREE	0x02	/* free, period			*/
188 #define SWM_POP		0x04	/* pop out			*/
189 
190 int swap_pager_full = 2;	/* swap space exhaustion (task killing) */
191 static int swap_pager_almost_full = 1; /* swap space exhaustion (w/hysteresis)*/
192 static int nsw_rcount;		/* free read buffers			*/
193 static int nsw_wcount_sync;	/* limit write buffers / synchronous	*/
194 static int nsw_wcount_async;	/* limit write buffers / asynchronous	*/
195 static int nsw_wcount_async_max;/* assigned maximum			*/
196 static int nsw_cluster_max;	/* maximum VOP I/O allowed		*/
197 
198 static struct swblock **swhash;
199 static int swhash_mask;
200 static struct mtx swhash_mtx;
201 
202 static int swap_async_max = 4;	/* maximum in-progress async I/O's	*/
203 static struct sx sw_alloc_sx;
204 
205 
206 SYSCTL_INT(_vm, OID_AUTO, swap_async_max,
207         CTLFLAG_RW, &swap_async_max, 0, "Maximum running async swap ops");
208 
209 /*
210  * "named" and "unnamed" anon region objects.  Try to reduce the overhead
211  * of searching a named list by hashing it just a little.
212  */
213 
214 #define NOBJLISTS		8
215 
216 #define NOBJLIST(handle)	\
217 	(&swap_pager_object_list[((int)(intptr_t)handle >> 4) & (NOBJLISTS-1)])
218 
219 static struct mtx sw_alloc_mtx;	/* protect list manipulation */
220 static struct pagerlst	swap_pager_object_list[NOBJLISTS];
221 static uma_zone_t	swap_zone;
222 static struct vm_object	swap_zone_obj;
223 
224 /*
225  * pagerops for OBJT_SWAP - "swap pager".  Some ops are also global procedure
226  * calls hooked from other parts of the VM system and do not appear here.
227  * (see vm/swap_pager.h).
228  */
229 static vm_object_t
230 		swap_pager_alloc(void *handle, vm_ooffset_t size,
231 				      vm_prot_t prot, vm_ooffset_t offset);
232 static void	swap_pager_dealloc(vm_object_t object);
233 static int	swap_pager_getpages(vm_object_t, vm_page_t *, int, int);
234 static void	swap_pager_putpages(vm_object_t, vm_page_t *, int, boolean_t, int *);
235 static boolean_t
236 		swap_pager_haspage(vm_object_t object, vm_pindex_t pindex, int *before, int *after);
237 static void	swap_pager_init(void);
238 static void	swap_pager_unswapped(vm_page_t);
239 static void	swap_pager_swapoff(struct swdevt *sp);
240 
241 struct pagerops swappagerops = {
242 	.pgo_init =	swap_pager_init,	/* early system initialization of pager	*/
243 	.pgo_alloc =	swap_pager_alloc,	/* allocate an OBJT_SWAP object		*/
244 	.pgo_dealloc =	swap_pager_dealloc,	/* deallocate an OBJT_SWAP object	*/
245 	.pgo_getpages =	swap_pager_getpages,	/* pagein				*/
246 	.pgo_putpages =	swap_pager_putpages,	/* pageout				*/
247 	.pgo_haspage =	swap_pager_haspage,	/* get backing store status for page	*/
248 	.pgo_pageunswapped = swap_pager_unswapped,	/* remove swap related to page		*/
249 };
250 
251 /*
252  * dmmax is in page-sized chunks with the new swap system.  It was
253  * dev-bsized chunks in the old.  dmmax is always a power of 2.
254  *
255  * swap_*() routines are externally accessible.  swp_*() routines are
256  * internal.
257  */
258 static int dmmax;
259 static int nswap_lowat = 128;	/* in pages, swap_pager_almost_full warn */
260 static int nswap_hiwat = 512;	/* in pages, swap_pager_almost_full warn */
261 
262 SYSCTL_INT(_vm, OID_AUTO, dmmax,
263 	CTLFLAG_RD, &dmmax, 0, "Maximum size of a swap block");
264 
265 static void	swp_sizecheck(void);
266 static void	swp_pager_async_iodone(struct buf *bp);
267 static int	swapongeom(struct thread *, struct vnode *);
268 static int	swaponvp(struct thread *, struct vnode *, u_long);
269 
270 /*
271  * Swap bitmap functions
272  */
273 static void	swp_pager_freeswapspace(daddr_t blk, int npages);
274 static daddr_t	swp_pager_getswapspace(int npages);
275 
276 /*
277  * Metadata functions
278  */
279 static struct swblock **swp_pager_hash(vm_object_t object, vm_pindex_t index);
280 static void swp_pager_meta_build(vm_object_t, vm_pindex_t, daddr_t);
281 static void swp_pager_meta_free(vm_object_t, vm_pindex_t, daddr_t);
282 static void swp_pager_meta_free_all(vm_object_t);
283 static daddr_t swp_pager_meta_ctl(vm_object_t, vm_pindex_t, int);
284 
285 /*
286  * SWP_SIZECHECK() -	update swap_pager_full indication
287  *
288  *	update the swap_pager_almost_full indication and warn when we are
289  *	about to run out of swap space, using lowat/hiwat hysteresis.
290  *
291  *	Clear swap_pager_full ( task killing ) indication when lowat is met.
292  *
293  *	No restrictions on call
294  *	This routine may not block.
295  *	This routine must be called at splvm()
296  */
297 static void
298 swp_sizecheck(void)
299 {
300 
301 	if (swap_pager_avail < nswap_lowat) {
302 		if (swap_pager_almost_full == 0) {
303 			printf("swap_pager: out of swap space\n");
304 			swap_pager_almost_full = 1;
305 		}
306 	} else {
307 		swap_pager_full = 0;
308 		if (swap_pager_avail > nswap_hiwat)
309 			swap_pager_almost_full = 0;
310 	}
311 }
312 
313 /*
314  * SWP_PAGER_HASH() -	hash swap meta data
315  *
316  *	This is an helper function which hashes the swapblk given
317  *	the object and page index.  It returns a pointer to a pointer
318  *	to the object, or a pointer to a NULL pointer if it could not
319  *	find a swapblk.
320  *
321  *	This routine must be called at splvm().
322  */
323 static struct swblock **
324 swp_pager_hash(vm_object_t object, vm_pindex_t index)
325 {
326 	struct swblock **pswap;
327 	struct swblock *swap;
328 
329 	index &= ~(vm_pindex_t)SWAP_META_MASK;
330 	pswap = &swhash[(index ^ (int)(intptr_t)object) & swhash_mask];
331 	while ((swap = *pswap) != NULL) {
332 		if (swap->swb_object == object &&
333 		    swap->swb_index == index
334 		) {
335 			break;
336 		}
337 		pswap = &swap->swb_hnext;
338 	}
339 	return (pswap);
340 }
341 
342 /*
343  * SWAP_PAGER_INIT() -	initialize the swap pager!
344  *
345  *	Expected to be started from system init.  NOTE:  This code is run
346  *	before much else so be careful what you depend on.  Most of the VM
347  *	system has yet to be initialized at this point.
348  */
349 static void
350 swap_pager_init(void)
351 {
352 	/*
353 	 * Initialize object lists
354 	 */
355 	int i;
356 
357 	for (i = 0; i < NOBJLISTS; ++i)
358 		TAILQ_INIT(&swap_pager_object_list[i]);
359 	mtx_init(&sw_alloc_mtx, "swap_pager list", NULL, MTX_DEF);
360 	mtx_init(&sw_dev_mtx, "swapdev", NULL, MTX_DEF);
361 
362 	/*
363 	 * Device Stripe, in PAGE_SIZE'd blocks
364 	 */
365 	dmmax = SWB_NPAGES * 2;
366 }
367 
368 /*
369  * SWAP_PAGER_SWAP_INIT() - swap pager initialization from pageout process
370  *
371  *	Expected to be started from pageout process once, prior to entering
372  *	its main loop.
373  */
374 void
375 swap_pager_swap_init(void)
376 {
377 	int n, n2;
378 
379 	/*
380 	 * Number of in-transit swap bp operations.  Don't
381 	 * exhaust the pbufs completely.  Make sure we
382 	 * initialize workable values (0 will work for hysteresis
383 	 * but it isn't very efficient).
384 	 *
385 	 * The nsw_cluster_max is constrained by the bp->b_pages[]
386 	 * array (MAXPHYS/PAGE_SIZE) and our locally defined
387 	 * MAX_PAGEOUT_CLUSTER.   Also be aware that swap ops are
388 	 * constrained by the swap device interleave stripe size.
389 	 *
390 	 * Currently we hardwire nsw_wcount_async to 4.  This limit is
391 	 * designed to prevent other I/O from having high latencies due to
392 	 * our pageout I/O.  The value 4 works well for one or two active swap
393 	 * devices but is probably a little low if you have more.  Even so,
394 	 * a higher value would probably generate only a limited improvement
395 	 * with three or four active swap devices since the system does not
396 	 * typically have to pageout at extreme bandwidths.   We will want
397 	 * at least 2 per swap devices, and 4 is a pretty good value if you
398 	 * have one NFS swap device due to the command/ack latency over NFS.
399 	 * So it all works out pretty well.
400 	 */
401 	nsw_cluster_max = min((MAXPHYS/PAGE_SIZE), MAX_PAGEOUT_CLUSTER);
402 
403 	mtx_lock(&pbuf_mtx);
404 	nsw_rcount = (nswbuf + 1) / 2;
405 	nsw_wcount_sync = (nswbuf + 3) / 4;
406 	nsw_wcount_async = 4;
407 	nsw_wcount_async_max = nsw_wcount_async;
408 	mtx_unlock(&pbuf_mtx);
409 
410 	/*
411 	 * Initialize our zone.  Right now I'm just guessing on the number
412 	 * we need based on the number of pages in the system.  Each swblock
413 	 * can hold 16 pages, so this is probably overkill.  This reservation
414 	 * is typically limited to around 32MB by default.
415 	 */
416 	n = cnt.v_page_count / 2;
417 	if (maxswzone && n > maxswzone / sizeof(struct swblock))
418 		n = maxswzone / sizeof(struct swblock);
419 	n2 = n;
420 	swap_zone = uma_zcreate("SWAPMETA", sizeof(struct swblock), NULL, NULL,
421 	    NULL, NULL, UMA_ALIGN_PTR, UMA_ZONE_NOFREE | UMA_ZONE_VM);
422 	if (swap_zone == NULL)
423 		panic("failed to create swap_zone.");
424 	do {
425 		if (uma_zone_set_obj(swap_zone, &swap_zone_obj, n))
426 			break;
427 		/*
428 		 * if the allocation failed, try a zone two thirds the
429 		 * size of the previous attempt.
430 		 */
431 		n -= ((n + 2) / 3);
432 	} while (n > 0);
433 	if (n2 != n)
434 		printf("Swap zone entries reduced from %d to %d.\n", n2, n);
435 	n2 = n;
436 
437 	/*
438 	 * Initialize our meta-data hash table.  The swapper does not need to
439 	 * be quite as efficient as the VM system, so we do not use an
440 	 * oversized hash table.
441 	 *
442 	 * 	n: 		size of hash table, must be power of 2
443 	 *	swhash_mask:	hash table index mask
444 	 */
445 	for (n = 1; n < n2 / 8; n *= 2)
446 		;
447 	swhash = malloc(sizeof(struct swblock *) * n, M_VMPGDATA, M_WAITOK | M_ZERO);
448 	swhash_mask = n - 1;
449 	mtx_init(&swhash_mtx, "swap_pager swhash", NULL, MTX_DEF);
450 }
451 
452 /*
453  * SWAP_PAGER_ALLOC() -	allocate a new OBJT_SWAP VM object and instantiate
454  *			its metadata structures.
455  *
456  *	This routine is called from the mmap and fork code to create a new
457  *	OBJT_SWAP object.  We do this by creating an OBJT_DEFAULT object
458  *	and then converting it with swp_pager_meta_build().
459  *
460  *	This routine may block in vm_object_allocate() and create a named
461  *	object lookup race, so we must interlock.   We must also run at
462  *	splvm() for the object lookup to handle races with interrupts, but
463  *	we do not have to maintain splvm() in between the lookup and the
464  *	add because (I believe) it is not possible to attempt to create
465  *	a new swap object w/handle when a default object with that handle
466  *	already exists.
467  *
468  * MPSAFE
469  */
470 static vm_object_t
471 swap_pager_alloc(void *handle, vm_ooffset_t size, vm_prot_t prot,
472 		 vm_ooffset_t offset)
473 {
474 	vm_object_t object;
475 	vm_pindex_t pindex;
476 
477 	pindex = OFF_TO_IDX(offset + PAGE_MASK + size);
478 
479 	if (handle) {
480 		mtx_lock(&Giant);
481 		/*
482 		 * Reference existing named region or allocate new one.  There
483 		 * should not be a race here against swp_pager_meta_build()
484 		 * as called from vm_page_remove() in regards to the lookup
485 		 * of the handle.
486 		 */
487 		sx_xlock(&sw_alloc_sx);
488 		object = vm_pager_object_lookup(NOBJLIST(handle), handle);
489 
490 		if (object != NULL) {
491 			vm_object_reference(object);
492 		} else {
493 			object = vm_object_allocate(OBJT_DEFAULT, pindex);
494 			object->handle = handle;
495 
496 			VM_OBJECT_LOCK(object);
497 			swp_pager_meta_build(object, 0, SWAPBLK_NONE);
498 			VM_OBJECT_UNLOCK(object);
499 		}
500 		sx_xunlock(&sw_alloc_sx);
501 		mtx_unlock(&Giant);
502 	} else {
503 		object = vm_object_allocate(OBJT_DEFAULT, pindex);
504 
505 		VM_OBJECT_LOCK(object);
506 		swp_pager_meta_build(object, 0, SWAPBLK_NONE);
507 		VM_OBJECT_UNLOCK(object);
508 	}
509 	return (object);
510 }
511 
512 /*
513  * SWAP_PAGER_DEALLOC() -	remove swap metadata from object
514  *
515  *	The swap backing for the object is destroyed.  The code is
516  *	designed such that we can reinstantiate it later, but this
517  *	routine is typically called only when the entire object is
518  *	about to be destroyed.
519  *
520  *	This routine may block, but no longer does.
521  *
522  *	The object must be locked or unreferenceable.
523  */
524 static void
525 swap_pager_dealloc(vm_object_t object)
526 {
527 
528 	/*
529 	 * Remove from list right away so lookups will fail if we block for
530 	 * pageout completion.
531 	 */
532 	if (object->handle != NULL) {
533 		mtx_lock(&sw_alloc_mtx);
534 		TAILQ_REMOVE(NOBJLIST(object->handle), object, pager_object_list);
535 		mtx_unlock(&sw_alloc_mtx);
536 	}
537 
538 	VM_OBJECT_LOCK_ASSERT(object, MA_OWNED);
539 	vm_object_pip_wait(object, "swpdea");
540 
541 	/*
542 	 * Free all remaining metadata.  We only bother to free it from
543 	 * the swap meta data.  We do not attempt to free swapblk's still
544 	 * associated with vm_page_t's for this object.  We do not care
545 	 * if paging is still in progress on some objects.
546 	 */
547 	swp_pager_meta_free_all(object);
548 }
549 
550 /************************************************************************
551  *			SWAP PAGER BITMAP ROUTINES			*
552  ************************************************************************/
553 
554 /*
555  * SWP_PAGER_GETSWAPSPACE() -	allocate raw swap space
556  *
557  *	Allocate swap for the requested number of pages.  The starting
558  *	swap block number (a page index) is returned or SWAPBLK_NONE
559  *	if the allocation failed.
560  *
561  *	Also has the side effect of advising that somebody made a mistake
562  *	when they configured swap and didn't configure enough.
563  *
564  *	Must be called at splvm() to avoid races with bitmap frees from
565  *	vm_page_remove() aka swap_pager_page_removed().
566  *
567  *	This routine may not block
568  *	This routine must be called at splvm().
569  *
570  *	We allocate in round-robin fashion from the configured devices.
571  */
572 static daddr_t
573 swp_pager_getswapspace(int npages)
574 {
575 	daddr_t blk;
576 	struct swdevt *sp;
577 	int i;
578 
579 	blk = SWAPBLK_NONE;
580 	mtx_lock(&sw_dev_mtx);
581 	sp = swdevhd;
582 	for (i = 0; i < nswapdev; i++) {
583 		if (sp == NULL)
584 			sp = TAILQ_FIRST(&swtailq);
585 		if (!(sp->sw_flags & SW_CLOSING)) {
586 			blk = blist_alloc(sp->sw_blist, npages);
587 			if (blk != SWAPBLK_NONE) {
588 				blk += sp->sw_first;
589 				sp->sw_used += npages;
590 				swap_pager_avail -= npages;
591 				swp_sizecheck();
592 				swdevhd = TAILQ_NEXT(sp, sw_list);
593 				goto done;
594 			}
595 		}
596 		sp = TAILQ_NEXT(sp, sw_list);
597 	}
598 	if (swap_pager_full != 2) {
599 		printf("swap_pager_getswapspace(%d): failed\n", npages);
600 		swap_pager_full = 2;
601 		swap_pager_almost_full = 1;
602 	}
603 	swdevhd = NULL;
604 done:
605 	mtx_unlock(&sw_dev_mtx);
606 	return (blk);
607 }
608 
609 static int
610 swp_pager_isondev(daddr_t blk, struct swdevt *sp)
611 {
612 
613 	return (blk >= sp->sw_first && blk < sp->sw_end);
614 }
615 
616 static void
617 swp_pager_strategy(struct buf *bp)
618 {
619 	struct swdevt *sp;
620 
621 	mtx_lock(&sw_dev_mtx);
622 	TAILQ_FOREACH(sp, &swtailq, sw_list) {
623 		if (bp->b_blkno >= sp->sw_first && bp->b_blkno < sp->sw_end) {
624 			mtx_unlock(&sw_dev_mtx);
625 			sp->sw_strategy(bp, sp);
626 			return;
627 		}
628 	}
629 	panic("Swapdev not found");
630 }
631 
632 
633 /*
634  * SWP_PAGER_FREESWAPSPACE() -	free raw swap space
635  *
636  *	This routine returns the specified swap blocks back to the bitmap.
637  *
638  *	Note:  This routine may not block (it could in the old swap code),
639  *	and through the use of the new blist routines it does not block.
640  *
641  *	We must be called at splvm() to avoid races with bitmap frees from
642  *	vm_page_remove() aka swap_pager_page_removed().
643  *
644  *	This routine may not block
645  *	This routine must be called at splvm().
646  */
647 static void
648 swp_pager_freeswapspace(daddr_t blk, int npages)
649 {
650 	struct swdevt *sp;
651 
652 	mtx_lock(&sw_dev_mtx);
653 	TAILQ_FOREACH(sp, &swtailq, sw_list) {
654 		if (blk >= sp->sw_first && blk < sp->sw_end) {
655 			sp->sw_used -= npages;
656 			/*
657 			 * If we are attempting to stop swapping on
658 			 * this device, we don't want to mark any
659 			 * blocks free lest they be reused.
660 			 */
661 			if ((sp->sw_flags & SW_CLOSING) == 0) {
662 				blist_free(sp->sw_blist, blk - sp->sw_first,
663 				    npages);
664 				swap_pager_avail += npages;
665 				swp_sizecheck();
666 			}
667 			mtx_unlock(&sw_dev_mtx);
668 			return;
669 		}
670 	}
671 	panic("Swapdev not found");
672 }
673 
674 /*
675  * SWAP_PAGER_FREESPACE() -	frees swap blocks associated with a page
676  *				range within an object.
677  *
678  *	This is a globally accessible routine.
679  *
680  *	This routine removes swapblk assignments from swap metadata.
681  *
682  *	The external callers of this routine typically have already destroyed
683  *	or renamed vm_page_t's associated with this range in the object so
684  *	we should be ok.
685  *
686  *	This routine may be called at any spl.  We up our spl to splvm temporarily
687  *	in order to perform the metadata removal.
688  */
689 void
690 swap_pager_freespace(vm_object_t object, vm_pindex_t start, vm_size_t size)
691 {
692 
693 	VM_OBJECT_LOCK_ASSERT(object, MA_OWNED);
694 	swp_pager_meta_free(object, start, size);
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 int
706 swap_pager_reserve(vm_object_t object, vm_pindex_t start, vm_size_t size)
707 {
708 	int n = 0;
709 	daddr_t blk = SWAPBLK_NONE;
710 	vm_pindex_t beg = start;	/* save start index */
711 
712 	VM_OBJECT_LOCK(object);
713 	while (size) {
714 		if (n == 0) {
715 			n = BLIST_MAX_ALLOC;
716 			while ((blk = swp_pager_getswapspace(n)) == SWAPBLK_NONE) {
717 				n >>= 1;
718 				if (n == 0) {
719 					swp_pager_meta_free(object, beg, start - beg);
720 					VM_OBJECT_UNLOCK(object);
721 					return (-1);
722 				}
723 			}
724 		}
725 		swp_pager_meta_build(object, start, blk);
726 		--size;
727 		++start;
728 		++blk;
729 		--n;
730 	}
731 	swp_pager_meta_free(object, start, n);
732 	VM_OBJECT_UNLOCK(object);
733 	return (0);
734 }
735 
736 /*
737  * SWAP_PAGER_COPY() -  copy blocks from source pager to destination pager
738  *			and destroy the source.
739  *
740  *	Copy any valid swapblks from the source to the destination.  In
741  *	cases where both the source and destination have a valid swapblk,
742  *	we keep the destination's.
743  *
744  *	This routine is allowed to block.  It may block allocating metadata
745  *	indirectly through swp_pager_meta_build() or if paging is still in
746  *	progress on the source.
747  *
748  *	This routine can be called at any spl
749  *
750  *	XXX vm_page_collapse() kinda expects us not to block because we
751  *	supposedly do not need to allocate memory, but for the moment we
752  *	*may* have to get a little memory from the zone allocator, but
753  *	it is taken from the interrupt memory.  We should be ok.
754  *
755  *	The source object contains no vm_page_t's (which is just as well)
756  *
757  *	The source object is of type OBJT_SWAP.
758  *
759  *	The source and destination objects must be locked or
760  *	inaccessible (XXX are they ?)
761  */
762 void
763 swap_pager_copy(vm_object_t srcobject, vm_object_t dstobject,
764     vm_pindex_t offset, int destroysource)
765 {
766 	vm_pindex_t i;
767 
768 	VM_OBJECT_LOCK_ASSERT(srcobject, MA_OWNED);
769 	VM_OBJECT_LOCK_ASSERT(dstobject, MA_OWNED);
770 
771 	/*
772 	 * If destroysource is set, we remove the source object from the
773 	 * swap_pager internal queue now.
774 	 */
775 	if (destroysource) {
776 		if (srcobject->handle != NULL) {
777 			mtx_lock(&sw_alloc_mtx);
778 			TAILQ_REMOVE(
779 			    NOBJLIST(srcobject->handle),
780 			    srcobject,
781 			    pager_object_list
782 			);
783 			mtx_unlock(&sw_alloc_mtx);
784 		}
785 	}
786 
787 	/*
788 	 * transfer source to destination.
789 	 */
790 	for (i = 0; i < dstobject->size; ++i) {
791 		daddr_t dstaddr;
792 
793 		/*
794 		 * Locate (without changing) the swapblk on the destination,
795 		 * unless it is invalid in which case free it silently, or
796 		 * if the destination is a resident page, in which case the
797 		 * source is thrown away.
798 		 */
799 		dstaddr = swp_pager_meta_ctl(dstobject, i, 0);
800 
801 		if (dstaddr == SWAPBLK_NONE) {
802 			/*
803 			 * Destination has no swapblk and is not resident,
804 			 * copy source.
805 			 */
806 			daddr_t srcaddr;
807 
808 			srcaddr = swp_pager_meta_ctl(
809 			    srcobject,
810 			    i + offset,
811 			    SWM_POP
812 			);
813 
814 			if (srcaddr != SWAPBLK_NONE) {
815 				/*
816 				 * swp_pager_meta_build() can sleep.
817 				 */
818 				vm_object_pip_add(srcobject, 1);
819 				VM_OBJECT_UNLOCK(srcobject);
820 				vm_object_pip_add(dstobject, 1);
821 				swp_pager_meta_build(dstobject, i, srcaddr);
822 				vm_object_pip_wakeup(dstobject);
823 				VM_OBJECT_LOCK(srcobject);
824 				vm_object_pip_wakeup(srcobject);
825 			}
826 		} else {
827 			/*
828 			 * Destination has valid swapblk or it is represented
829 			 * by a resident page.  We destroy the sourceblock.
830 			 */
831 
832 			swp_pager_meta_ctl(srcobject, i + offset, SWM_FREE);
833 		}
834 	}
835 
836 	/*
837 	 * Free left over swap blocks in source.
838 	 *
839 	 * We have to revert the type to OBJT_DEFAULT so we do not accidently
840 	 * double-remove the object from the swap queues.
841 	 */
842 	if (destroysource) {
843 		swp_pager_meta_free_all(srcobject);
844 		/*
845 		 * Reverting the type is not necessary, the caller is going
846 		 * to destroy srcobject directly, but I'm doing it here
847 		 * for consistency since we've removed the object from its
848 		 * queues.
849 		 */
850 		srcobject->type = OBJT_DEFAULT;
851 	}
852 }
853 
854 /*
855  * SWAP_PAGER_HASPAGE() -	determine if we have good backing store for
856  *				the requested page.
857  *
858  *	We determine whether good backing store exists for the requested
859  *	page and return TRUE if it does, FALSE if it doesn't.
860  *
861  *	If TRUE, we also try to determine how much valid, contiguous backing
862  *	store exists before and after the requested page within a reasonable
863  *	distance.  We do not try to restrict it to the swap device stripe
864  *	(that is handled in getpages/putpages).  It probably isn't worth
865  *	doing here.
866  */
867 static boolean_t
868 swap_pager_haspage(vm_object_t object, vm_pindex_t pindex, int *before, int *after)
869 {
870 	daddr_t blk0;
871 
872 	VM_OBJECT_LOCK_ASSERT(object, MA_OWNED);
873 	/*
874 	 * do we have good backing store at the requested index ?
875 	 */
876 	blk0 = swp_pager_meta_ctl(object, pindex, 0);
877 
878 	if (blk0 == SWAPBLK_NONE) {
879 		if (before)
880 			*before = 0;
881 		if (after)
882 			*after = 0;
883 		return (FALSE);
884 	}
885 
886 	/*
887 	 * find backwards-looking contiguous good backing store
888 	 */
889 	if (before != NULL) {
890 		int i;
891 
892 		for (i = 1; i < (SWB_NPAGES/2); ++i) {
893 			daddr_t blk;
894 
895 			if (i > pindex)
896 				break;
897 			blk = swp_pager_meta_ctl(object, pindex - i, 0);
898 			if (blk != blk0 - i)
899 				break;
900 		}
901 		*before = (i - 1);
902 	}
903 
904 	/*
905 	 * find forward-looking contiguous good backing store
906 	 */
907 	if (after != NULL) {
908 		int i;
909 
910 		for (i = 1; i < (SWB_NPAGES/2); ++i) {
911 			daddr_t blk;
912 
913 			blk = swp_pager_meta_ctl(object, pindex + i, 0);
914 			if (blk != blk0 + i)
915 				break;
916 		}
917 		*after = (i - 1);
918 	}
919 	return (TRUE);
920 }
921 
922 /*
923  * SWAP_PAGER_PAGE_UNSWAPPED() - remove swap backing store related to page
924  *
925  *	This removes any associated swap backing store, whether valid or
926  *	not, from the page.
927  *
928  *	This routine is typically called when a page is made dirty, at
929  *	which point any associated swap can be freed.  MADV_FREE also
930  *	calls us in a special-case situation
931  *
932  *	NOTE!!!  If the page is clean and the swap was valid, the caller
933  *	should make the page dirty before calling this routine.  This routine
934  *	does NOT change the m->dirty status of the page.  Also: MADV_FREE
935  *	depends on it.
936  *
937  *	This routine may not block
938  *	This routine must be called at splvm()
939  */
940 static void
941 swap_pager_unswapped(vm_page_t m)
942 {
943 
944 	VM_OBJECT_LOCK_ASSERT(m->object, MA_OWNED);
945 	swp_pager_meta_ctl(m->object, m->pindex, SWM_FREE);
946 }
947 
948 /*
949  * SWAP_PAGER_GETPAGES() - bring pages in from swap
950  *
951  *	Attempt to retrieve (m, count) pages from backing store, but make
952  *	sure we retrieve at least m[reqpage].  We try to load in as large
953  *	a chunk surrounding m[reqpage] as is contiguous in swap and which
954  *	belongs to the same object.
955  *
956  *	The code is designed for asynchronous operation and
957  *	immediate-notification of 'reqpage' but tends not to be
958  *	used that way.  Please do not optimize-out this algorithmic
959  *	feature, I intend to improve on it in the future.
960  *
961  *	The parent has a single vm_object_pip_add() reference prior to
962  *	calling us and we should return with the same.
963  *
964  *	The parent has BUSY'd the pages.  We should return with 'm'
965  *	left busy, but the others adjusted.
966  */
967 static int
968 swap_pager_getpages(vm_object_t object, vm_page_t *m, int count, int reqpage)
969 {
970 	struct buf *bp;
971 	vm_page_t mreq;
972 	int i;
973 	int j;
974 	daddr_t blk;
975 
976 	mreq = m[reqpage];
977 
978 	KASSERT(mreq->object == object,
979 	    ("swap_pager_getpages: object mismatch %p/%p",
980 	    object, mreq->object));
981 
982 	/*
983 	 * Calculate range to retrieve.  The pages have already been assigned
984 	 * their swapblks.  We require a *contiguous* range but we know it to
985 	 * not span devices.   If we do not supply it, bad things
986 	 * happen.  Note that blk, iblk & jblk can be SWAPBLK_NONE, but the
987 	 * loops are set up such that the case(s) are handled implicitly.
988 	 *
989 	 * The swp_*() calls must be made at splvm().  vm_page_free() does
990 	 * not need to be, but it will go a little faster if it is.
991 	 */
992 	blk = swp_pager_meta_ctl(mreq->object, mreq->pindex, 0);
993 
994 	for (i = reqpage - 1; i >= 0; --i) {
995 		daddr_t iblk;
996 
997 		iblk = swp_pager_meta_ctl(m[i]->object, m[i]->pindex, 0);
998 		if (blk != iblk + (reqpage - i))
999 			break;
1000 	}
1001 	++i;
1002 
1003 	for (j = reqpage + 1; j < count; ++j) {
1004 		daddr_t jblk;
1005 
1006 		jblk = swp_pager_meta_ctl(m[j]->object, m[j]->pindex, 0);
1007 		if (blk != jblk - (j - reqpage))
1008 			break;
1009 	}
1010 
1011 	/*
1012 	 * free pages outside our collection range.   Note: we never free
1013 	 * mreq, it must remain busy throughout.
1014 	 */
1015 	if (0 < i || j < count) {
1016 		int k;
1017 
1018 		vm_page_lock_queues();
1019 		for (k = 0; k < i; ++k)
1020 			vm_page_free(m[k]);
1021 		for (k = j; k < count; ++k)
1022 			vm_page_free(m[k]);
1023 		vm_page_unlock_queues();
1024 	}
1025 
1026 	/*
1027 	 * Return VM_PAGER_FAIL if we have nothing to do.  Return mreq
1028 	 * still busy, but the others unbusied.
1029 	 */
1030 	if (blk == SWAPBLK_NONE)
1031 		return (VM_PAGER_FAIL);
1032 
1033 	/*
1034 	 * Getpbuf() can sleep.
1035 	 */
1036 	VM_OBJECT_UNLOCK(object);
1037 	/*
1038 	 * Get a swap buffer header to perform the IO
1039 	 */
1040 	bp = getpbuf(&nsw_rcount);
1041 	bp->b_flags |= B_PAGING;
1042 
1043 	/*
1044 	 * map our page(s) into kva for input
1045 	 */
1046 	pmap_qenter((vm_offset_t)bp->b_data, m + i, j - i);
1047 
1048 	bp->b_iocmd = BIO_READ;
1049 	bp->b_iodone = swp_pager_async_iodone;
1050 	bp->b_rcred = crhold(thread0.td_ucred);
1051 	bp->b_wcred = crhold(thread0.td_ucred);
1052 	bp->b_blkno = blk - (reqpage - i);
1053 	bp->b_bcount = PAGE_SIZE * (j - i);
1054 	bp->b_bufsize = PAGE_SIZE * (j - i);
1055 	bp->b_pager.pg_reqpage = reqpage - i;
1056 
1057 	VM_OBJECT_LOCK(object);
1058 	vm_page_lock_queues();
1059 	{
1060 		int k;
1061 
1062 		for (k = i; k < j; ++k) {
1063 			bp->b_pages[k - i] = m[k];
1064 			vm_page_flag_set(m[k], PG_SWAPINPROG);
1065 		}
1066 	}
1067 	vm_page_unlock_queues();
1068 	bp->b_npages = j - i;
1069 
1070 	cnt.v_swapin++;
1071 	cnt.v_swappgsin += bp->b_npages;
1072 
1073 	/*
1074 	 * We still hold the lock on mreq, and our automatic completion routine
1075 	 * does not remove it.
1076 	 */
1077 	vm_object_pip_add(object, bp->b_npages);
1078 	VM_OBJECT_UNLOCK(object);
1079 
1080 	/*
1081 	 * perform the I/O.  NOTE!!!  bp cannot be considered valid after
1082 	 * this point because we automatically release it on completion.
1083 	 * Instead, we look at the one page we are interested in which we
1084 	 * still hold a lock on even through the I/O completion.
1085 	 *
1086 	 * The other pages in our m[] array are also released on completion,
1087 	 * so we cannot assume they are valid anymore either.
1088 	 *
1089 	 * NOTE: b_blkno is destroyed by the call to swapdev_strategy
1090 	 */
1091 	BUF_KERNPROC(bp);
1092 	swp_pager_strategy(bp);
1093 
1094 	/*
1095 	 * wait for the page we want to complete.  PG_SWAPINPROG is always
1096 	 * cleared on completion.  If an I/O error occurs, SWAPBLK_NONE
1097 	 * is set in the meta-data.
1098 	 */
1099 	vm_page_lock_queues();
1100 	while ((mreq->flags & PG_SWAPINPROG) != 0) {
1101 		vm_page_flag_set(mreq, PG_WANTED | PG_REFERENCED);
1102 		cnt.v_intrans++;
1103 		if (msleep(mreq, &vm_page_queue_mtx, PSWP, "swread", hz*20)) {
1104 			printf(
1105 "swap_pager: indefinite wait buffer: bufobj: %p, blkno: %jd, size: %ld\n",
1106 			    bp->b_bufobj, (intmax_t)bp->b_blkno, bp->b_bcount);
1107 		}
1108 	}
1109 	vm_page_unlock_queues();
1110 
1111 	VM_OBJECT_LOCK(object);
1112 	/*
1113 	 * mreq is left busied after completion, but all the other pages
1114 	 * are freed.  If we had an unrecoverable read error the page will
1115 	 * not be valid.
1116 	 */
1117 	if (mreq->valid != VM_PAGE_BITS_ALL) {
1118 		return (VM_PAGER_ERROR);
1119 	} else {
1120 		return (VM_PAGER_OK);
1121 	}
1122 
1123 	/*
1124 	 * A final note: in a low swap situation, we cannot deallocate swap
1125 	 * and mark a page dirty here because the caller is likely to mark
1126 	 * the page clean when we return, causing the page to possibly revert
1127 	 * to all-zero's later.
1128 	 */
1129 }
1130 
1131 /*
1132  *	swap_pager_putpages:
1133  *
1134  *	Assign swap (if necessary) and initiate I/O on the specified pages.
1135  *
1136  *	We support both OBJT_DEFAULT and OBJT_SWAP objects.  DEFAULT objects
1137  *	are automatically converted to SWAP objects.
1138  *
1139  *	In a low memory situation we may block in VOP_STRATEGY(), but the new
1140  *	vm_page reservation system coupled with properly written VFS devices
1141  *	should ensure that no low-memory deadlock occurs.  This is an area
1142  *	which needs work.
1143  *
1144  *	The parent has N vm_object_pip_add() references prior to
1145  *	calling us and will remove references for rtvals[] that are
1146  *	not set to VM_PAGER_PEND.  We need to remove the rest on I/O
1147  *	completion.
1148  *
1149  *	The parent has soft-busy'd the pages it passes us and will unbusy
1150  *	those whos rtvals[] entry is not set to VM_PAGER_PEND on return.
1151  *	We need to unbusy the rest on I/O completion.
1152  */
1153 void
1154 swap_pager_putpages(vm_object_t object, vm_page_t *m, int count,
1155     boolean_t sync, int *rtvals)
1156 {
1157 	int i;
1158 	int n = 0;
1159 
1160 	GIANT_REQUIRED;
1161 	if (count && m[0]->object != object) {
1162 		panic("swap_pager_getpages: object mismatch %p/%p",
1163 		    object,
1164 		    m[0]->object
1165 		);
1166 	}
1167 
1168 	/*
1169 	 * Step 1
1170 	 *
1171 	 * Turn object into OBJT_SWAP
1172 	 * check for bogus sysops
1173 	 * force sync if not pageout process
1174 	 */
1175 	if (object->type != OBJT_SWAP)
1176 		swp_pager_meta_build(object, 0, SWAPBLK_NONE);
1177 	VM_OBJECT_UNLOCK(object);
1178 
1179 	if (curproc != pageproc)
1180 		sync = TRUE;
1181 
1182 	/*
1183 	 * Step 2
1184 	 *
1185 	 * Update nsw parameters from swap_async_max sysctl values.
1186 	 * Do not let the sysop crash the machine with bogus numbers.
1187 	 */
1188 	mtx_lock(&pbuf_mtx);
1189 	if (swap_async_max != nsw_wcount_async_max) {
1190 		int n;
1191 
1192 		/*
1193 		 * limit range
1194 		 */
1195 		if ((n = swap_async_max) > nswbuf / 2)
1196 			n = nswbuf / 2;
1197 		if (n < 1)
1198 			n = 1;
1199 		swap_async_max = n;
1200 
1201 		/*
1202 		 * Adjust difference ( if possible ).  If the current async
1203 		 * count is too low, we may not be able to make the adjustment
1204 		 * at this time.
1205 		 */
1206 		n -= nsw_wcount_async_max;
1207 		if (nsw_wcount_async + n >= 0) {
1208 			nsw_wcount_async += n;
1209 			nsw_wcount_async_max += n;
1210 			wakeup(&nsw_wcount_async);
1211 		}
1212 	}
1213 	mtx_unlock(&pbuf_mtx);
1214 
1215 	/*
1216 	 * Step 3
1217 	 *
1218 	 * Assign swap blocks and issue I/O.  We reallocate swap on the fly.
1219 	 * The page is left dirty until the pageout operation completes
1220 	 * successfully.
1221 	 */
1222 	for (i = 0; i < count; i += n) {
1223 		int j;
1224 		struct buf *bp;
1225 		daddr_t blk;
1226 
1227 		/*
1228 		 * Maximum I/O size is limited by a number of factors.
1229 		 */
1230 		n = min(BLIST_MAX_ALLOC, count - i);
1231 		n = min(n, nsw_cluster_max);
1232 
1233 		/*
1234 		 * Get biggest block of swap we can.  If we fail, fall
1235 		 * back and try to allocate a smaller block.  Don't go
1236 		 * overboard trying to allocate space if it would overly
1237 		 * fragment swap.
1238 		 */
1239 		while (
1240 		    (blk = swp_pager_getswapspace(n)) == SWAPBLK_NONE &&
1241 		    n > 4
1242 		) {
1243 			n >>= 1;
1244 		}
1245 		if (blk == SWAPBLK_NONE) {
1246 			for (j = 0; j < n; ++j)
1247 				rtvals[i+j] = VM_PAGER_FAIL;
1248 			continue;
1249 		}
1250 
1251 		/*
1252 		 * All I/O parameters have been satisfied, build the I/O
1253 		 * request and assign the swap space.
1254 		 */
1255 		if (sync == TRUE) {
1256 			bp = getpbuf(&nsw_wcount_sync);
1257 		} else {
1258 			bp = getpbuf(&nsw_wcount_async);
1259 			bp->b_flags = B_ASYNC;
1260 		}
1261 		bp->b_flags |= B_PAGING;
1262 		bp->b_iocmd = BIO_WRITE;
1263 
1264 		pmap_qenter((vm_offset_t)bp->b_data, &m[i], n);
1265 
1266 		bp->b_rcred = crhold(thread0.td_ucred);
1267 		bp->b_wcred = crhold(thread0.td_ucred);
1268 		bp->b_bcount = PAGE_SIZE * n;
1269 		bp->b_bufsize = PAGE_SIZE * n;
1270 		bp->b_blkno = blk;
1271 
1272 		VM_OBJECT_LOCK(object);
1273 		for (j = 0; j < n; ++j) {
1274 			vm_page_t mreq = m[i+j];
1275 
1276 			swp_pager_meta_build(
1277 			    mreq->object,
1278 			    mreq->pindex,
1279 			    blk + j
1280 			);
1281 			vm_page_dirty(mreq);
1282 			rtvals[i+j] = VM_PAGER_OK;
1283 
1284 			vm_page_lock_queues();
1285 			vm_page_flag_set(mreq, PG_SWAPINPROG);
1286 			vm_page_unlock_queues();
1287 			bp->b_pages[j] = mreq;
1288 		}
1289 		VM_OBJECT_UNLOCK(object);
1290 		bp->b_npages = n;
1291 		/*
1292 		 * Must set dirty range for NFS to work.
1293 		 */
1294 		bp->b_dirtyoff = 0;
1295 		bp->b_dirtyend = bp->b_bcount;
1296 
1297 		cnt.v_swapout++;
1298 		cnt.v_swappgsout += bp->b_npages;
1299 
1300 		/*
1301 		 * asynchronous
1302 		 *
1303 		 * NOTE: b_blkno is destroyed by the call to swapdev_strategy
1304 		 */
1305 		if (sync == FALSE) {
1306 			bp->b_iodone = swp_pager_async_iodone;
1307 			BUF_KERNPROC(bp);
1308 			swp_pager_strategy(bp);
1309 
1310 			for (j = 0; j < n; ++j)
1311 				rtvals[i+j] = VM_PAGER_PEND;
1312 			/* restart outter loop */
1313 			continue;
1314 		}
1315 
1316 		/*
1317 		 * synchronous
1318 		 *
1319 		 * NOTE: b_blkno is destroyed by the call to swapdev_strategy
1320 		 */
1321 		bp->b_iodone = bdone;
1322 		swp_pager_strategy(bp);
1323 
1324 		/*
1325 		 * Wait for the sync I/O to complete, then update rtvals.
1326 		 * We just set the rtvals[] to VM_PAGER_PEND so we can call
1327 		 * our async completion routine at the end, thus avoiding a
1328 		 * double-free.
1329 		 */
1330 		bwait(bp, PVM, "swwrt");
1331 		for (j = 0; j < n; ++j)
1332 			rtvals[i+j] = VM_PAGER_PEND;
1333 		/*
1334 		 * Now that we are through with the bp, we can call the
1335 		 * normal async completion, which frees everything up.
1336 		 */
1337 		swp_pager_async_iodone(bp);
1338 	}
1339 	VM_OBJECT_LOCK(object);
1340 }
1341 
1342 /*
1343  *	swp_pager_async_iodone:
1344  *
1345  *	Completion routine for asynchronous reads and writes from/to swap.
1346  *	Also called manually by synchronous code to finish up a bp.
1347  *
1348  *	For READ operations, the pages are PG_BUSY'd.  For WRITE operations,
1349  *	the pages are vm_page_t->busy'd.  For READ operations, we PG_BUSY
1350  *	unbusy all pages except the 'main' request page.  For WRITE
1351  *	operations, we vm_page_t->busy'd unbusy all pages ( we can do this
1352  *	because we marked them all VM_PAGER_PEND on return from putpages ).
1353  *
1354  *	This routine may not block.
1355  *	This routine is called at splbio() or better
1356  *
1357  *	We up ourselves to splvm() as required for various vm_page related
1358  *	calls.
1359  */
1360 static void
1361 swp_pager_async_iodone(struct buf *bp)
1362 {
1363 	int i;
1364 	vm_object_t object = NULL;
1365 
1366 	bp->b_flags |= B_DONE;
1367 
1368 	/*
1369 	 * report error
1370 	 */
1371 	if (bp->b_ioflags & BIO_ERROR) {
1372 		printf(
1373 		    "swap_pager: I/O error - %s failed; blkno %ld,"
1374 			"size %ld, error %d\n",
1375 		    ((bp->b_iocmd == BIO_READ) ? "pagein" : "pageout"),
1376 		    (long)bp->b_blkno,
1377 		    (long)bp->b_bcount,
1378 		    bp->b_error
1379 		);
1380 	}
1381 
1382 	/*
1383 	 * remove the mapping for kernel virtual
1384 	 */
1385 	pmap_qremove((vm_offset_t)bp->b_data, bp->b_npages);
1386 
1387 	if (bp->b_npages) {
1388 		object = bp->b_pages[0]->object;
1389 		VM_OBJECT_LOCK(object);
1390 	}
1391 	vm_page_lock_queues();
1392 	/*
1393 	 * cleanup pages.  If an error occurs writing to swap, we are in
1394 	 * very serious trouble.  If it happens to be a disk error, though,
1395 	 * we may be able to recover by reassigning the swap later on.  So
1396 	 * in this case we remove the m->swapblk assignment for the page
1397 	 * but do not free it in the rlist.  The errornous block(s) are thus
1398 	 * never reallocated as swap.  Redirty the page and continue.
1399 	 */
1400 	for (i = 0; i < bp->b_npages; ++i) {
1401 		vm_page_t m = bp->b_pages[i];
1402 
1403 		vm_page_flag_clear(m, PG_SWAPINPROG);
1404 
1405 		if (bp->b_ioflags & BIO_ERROR) {
1406 			/*
1407 			 * If an error occurs I'd love to throw the swapblk
1408 			 * away without freeing it back to swapspace, so it
1409 			 * can never be used again.  But I can't from an
1410 			 * interrupt.
1411 			 */
1412 			if (bp->b_iocmd == BIO_READ) {
1413 				/*
1414 				 * When reading, reqpage needs to stay
1415 				 * locked for the parent, but all other
1416 				 * pages can be freed.  We still want to
1417 				 * wakeup the parent waiting on the page,
1418 				 * though.  ( also: pg_reqpage can be -1 and
1419 				 * not match anything ).
1420 				 *
1421 				 * We have to wake specifically requested pages
1422 				 * up too because we cleared PG_SWAPINPROG and
1423 				 * someone may be waiting for that.
1424 				 *
1425 				 * NOTE: for reads, m->dirty will probably
1426 				 * be overridden by the original caller of
1427 				 * getpages so don't play cute tricks here.
1428 				 *
1429 				 * XXX IT IS NOT LEGAL TO FREE THE PAGE HERE
1430 				 * AS THIS MESSES WITH object->memq, and it is
1431 				 * not legal to mess with object->memq from an
1432 				 * interrupt.
1433 				 */
1434 				m->valid = 0;
1435 				if (i != bp->b_pager.pg_reqpage)
1436 					vm_page_free(m);
1437 				else
1438 					vm_page_flash(m);
1439 				/*
1440 				 * If i == bp->b_pager.pg_reqpage, do not wake
1441 				 * the page up.  The caller needs to.
1442 				 */
1443 			} else {
1444 				/*
1445 				 * If a write error occurs, reactivate page
1446 				 * so it doesn't clog the inactive list,
1447 				 * then finish the I/O.
1448 				 */
1449 				vm_page_dirty(m);
1450 				vm_page_activate(m);
1451 				vm_page_io_finish(m);
1452 			}
1453 		} else if (bp->b_iocmd == BIO_READ) {
1454 			/*
1455 			 * For read success, clear dirty bits.  Nobody should
1456 			 * have this page mapped but don't take any chances,
1457 			 * make sure the pmap modify bits are also cleared.
1458 			 *
1459 			 * NOTE: for reads, m->dirty will probably be
1460 			 * overridden by the original caller of getpages so
1461 			 * we cannot set them in order to free the underlying
1462 			 * swap in a low-swap situation.  I don't think we'd
1463 			 * want to do that anyway, but it was an optimization
1464 			 * that existed in the old swapper for a time before
1465 			 * it got ripped out due to precisely this problem.
1466 			 *
1467 			 * If not the requested page then deactivate it.
1468 			 *
1469 			 * Note that the requested page, reqpage, is left
1470 			 * busied, but we still have to wake it up.  The
1471 			 * other pages are released (unbusied) by
1472 			 * vm_page_wakeup().  We do not set reqpage's
1473 			 * valid bits here, it is up to the caller.
1474 			 */
1475 			pmap_clear_modify(m);
1476 			m->valid = VM_PAGE_BITS_ALL;
1477 			vm_page_undirty(m);
1478 
1479 			/*
1480 			 * We have to wake specifically requested pages
1481 			 * up too because we cleared PG_SWAPINPROG and
1482 			 * could be waiting for it in getpages.  However,
1483 			 * be sure to not unbusy getpages specifically
1484 			 * requested page - getpages expects it to be
1485 			 * left busy.
1486 			 */
1487 			if (i != bp->b_pager.pg_reqpage) {
1488 				vm_page_deactivate(m);
1489 				vm_page_wakeup(m);
1490 			} else {
1491 				vm_page_flash(m);
1492 			}
1493 		} else {
1494 			/*
1495 			 * For write success, clear the modify and dirty
1496 			 * status, then finish the I/O ( which decrements the
1497 			 * busy count and possibly wakes waiter's up ).
1498 			 */
1499 			pmap_clear_modify(m);
1500 			vm_page_undirty(m);
1501 			vm_page_io_finish(m);
1502 			if (vm_page_count_severe())
1503 				vm_page_try_to_cache(m);
1504 		}
1505 	}
1506 	vm_page_unlock_queues();
1507 
1508 	/*
1509 	 * adjust pip.  NOTE: the original parent may still have its own
1510 	 * pip refs on the object.
1511 	 */
1512 	if (object != NULL) {
1513 		vm_object_pip_wakeupn(object, bp->b_npages);
1514 		VM_OBJECT_UNLOCK(object);
1515 	}
1516 
1517 	/*
1518 	 * release the physical I/O buffer
1519 	 */
1520 	relpbuf(
1521 	    bp,
1522 	    ((bp->b_iocmd == BIO_READ) ? &nsw_rcount :
1523 		((bp->b_flags & B_ASYNC) ?
1524 		    &nsw_wcount_async :
1525 		    &nsw_wcount_sync
1526 		)
1527 	    )
1528 	);
1529 }
1530 
1531 /*
1532  *	swap_pager_isswapped:
1533  *
1534  *	Return 1 if at least one page in the given object is paged
1535  *	out to the given swap device.
1536  *
1537  *	This routine may not block.
1538  */
1539 int
1540 swap_pager_isswapped(vm_object_t object, struct swdevt *sp)
1541 {
1542 	daddr_t index = 0;
1543 	int bcount;
1544 	int i;
1545 
1546 	VM_OBJECT_LOCK_ASSERT(object, MA_OWNED);
1547 	if (object->type != OBJT_SWAP)
1548 		return (0);
1549 
1550 	mtx_lock(&swhash_mtx);
1551 	for (bcount = 0; bcount < object->un_pager.swp.swp_bcount; bcount++) {
1552 		struct swblock *swap;
1553 
1554 		if ((swap = *swp_pager_hash(object, index)) != NULL) {
1555 			for (i = 0; i < SWAP_META_PAGES; ++i) {
1556 				if (swp_pager_isondev(swap->swb_pages[i], sp)) {
1557 					mtx_unlock(&swhash_mtx);
1558 					return (1);
1559 				}
1560 			}
1561 		}
1562 		index += SWAP_META_PAGES;
1563 		if (index > 0x20000000)
1564 			panic("swap_pager_isswapped: failed to locate all swap meta blocks");
1565 	}
1566 	mtx_unlock(&swhash_mtx);
1567 	return (0);
1568 }
1569 
1570 /*
1571  * SWP_PAGER_FORCE_PAGEIN() - force a swap block to be paged in
1572  *
1573  *	This routine dissociates the page at the given index within a
1574  *	swap block from its backing store, paging it in if necessary.
1575  *	If the page is paged in, it is placed in the inactive queue,
1576  *	since it had its backing store ripped out from under it.
1577  *	We also attempt to swap in all other pages in the swap block,
1578  *	we only guarantee that the one at the specified index is
1579  *	paged in.
1580  *
1581  *	XXX - The code to page the whole block in doesn't work, so we
1582  *	      revert to the one-by-one behavior for now.  Sigh.
1583  */
1584 static __inline void
1585 swp_pager_force_pagein(vm_object_t object, vm_pindex_t pindex)
1586 {
1587 	vm_page_t m;
1588 
1589 	vm_object_pip_add(object, 1);
1590 	m = vm_page_grab(object, pindex, VM_ALLOC_NORMAL|VM_ALLOC_RETRY);
1591 	if (m->valid == VM_PAGE_BITS_ALL) {
1592 		vm_object_pip_subtract(object, 1);
1593 		vm_page_lock_queues();
1594 		vm_page_activate(m);
1595 		vm_page_dirty(m);
1596 		vm_page_wakeup(m);
1597 		vm_page_unlock_queues();
1598 		vm_pager_page_unswapped(m);
1599 		return;
1600 	}
1601 
1602 	if (swap_pager_getpages(object, &m, 1, 0) != VM_PAGER_OK)
1603 		panic("swap_pager_force_pagein: read from swap failed");/*XXX*/
1604 	vm_object_pip_subtract(object, 1);
1605 	vm_page_lock_queues();
1606 	vm_page_dirty(m);
1607 	vm_page_dontneed(m);
1608 	vm_page_wakeup(m);
1609 	vm_page_unlock_queues();
1610 	vm_pager_page_unswapped(m);
1611 }
1612 
1613 /*
1614  *	swap_pager_swapoff:
1615  *
1616  *	Page in all of the pages that have been paged out to the
1617  *	given device.  The corresponding blocks in the bitmap must be
1618  *	marked as allocated and the device must be flagged SW_CLOSING.
1619  *	There may be no processes swapped out to the device.
1620  *
1621  *	This routine may block.
1622  */
1623 static void
1624 swap_pager_swapoff(struct swdevt *sp)
1625 {
1626 	struct swblock *swap;
1627 	int i, j, retries;
1628 
1629 	GIANT_REQUIRED;
1630 
1631 	retries = 0;
1632 full_rescan:
1633 	mtx_lock(&swhash_mtx);
1634 	for (i = 0; i <= swhash_mask; i++) { /* '<=' is correct here */
1635 restart:
1636 		for (swap = swhash[i]; swap != NULL; swap = swap->swb_hnext) {
1637 			vm_object_t object = swap->swb_object;
1638 			vm_pindex_t pindex = swap->swb_index;
1639                         for (j = 0; j < SWAP_META_PAGES; ++j) {
1640                                 if (swp_pager_isondev(swap->swb_pages[j], sp)) {
1641 					/* avoid deadlock */
1642 					if (!VM_OBJECT_TRYLOCK(object)) {
1643 						break;
1644 					} else {
1645 						mtx_unlock(&swhash_mtx);
1646 						swp_pager_force_pagein(object,
1647 						    pindex + j);
1648 						VM_OBJECT_UNLOCK(object);
1649 						mtx_lock(&swhash_mtx);
1650 						goto restart;
1651 					}
1652 				}
1653                         }
1654 		}
1655 	}
1656 	mtx_unlock(&swhash_mtx);
1657 	if (sp->sw_used) {
1658 		int dummy;
1659 		/*
1660 		 * Objects may be locked or paging to the device being
1661 		 * removed, so we will miss their pages and need to
1662 		 * make another pass.  We have marked this device as
1663 		 * SW_CLOSING, so the activity should finish soon.
1664 		 */
1665 		retries++;
1666 		if (retries > 100) {
1667 			panic("swapoff: failed to locate %d swap blocks",
1668 			    sp->sw_used);
1669 		}
1670 		tsleep(&dummy, PVM, "swpoff", hz / 20);
1671 		goto full_rescan;
1672 	}
1673 }
1674 
1675 /************************************************************************
1676  *				SWAP META DATA 				*
1677  ************************************************************************
1678  *
1679  *	These routines manipulate the swap metadata stored in the
1680  *	OBJT_SWAP object.  All swp_*() routines must be called at
1681  *	splvm() because swap can be freed up by the low level vm_page
1682  *	code which might be called from interrupts beyond what splbio() covers.
1683  *
1684  *	Swap metadata is implemented with a global hash and not directly
1685  *	linked into the object.  Instead the object simply contains
1686  *	appropriate tracking counters.
1687  */
1688 
1689 /*
1690  * SWP_PAGER_META_BUILD() -	add swap block to swap meta data for object
1691  *
1692  *	We first convert the object to a swap object if it is a default
1693  *	object.
1694  *
1695  *	The specified swapblk is added to the object's swap metadata.  If
1696  *	the swapblk is not valid, it is freed instead.  Any previously
1697  *	assigned swapblk is freed.
1698  *
1699  *	This routine must be called at splvm(), except when used to convert
1700  *	an OBJT_DEFAULT object into an OBJT_SWAP object.
1701  */
1702 static void
1703 swp_pager_meta_build(vm_object_t object, vm_pindex_t pindex, daddr_t swapblk)
1704 {
1705 	struct swblock *swap;
1706 	struct swblock **pswap;
1707 	int idx;
1708 
1709 	VM_OBJECT_LOCK_ASSERT(object, MA_OWNED);
1710 	/*
1711 	 * Convert default object to swap object if necessary
1712 	 */
1713 	if (object->type != OBJT_SWAP) {
1714 		object->type = OBJT_SWAP;
1715 		object->un_pager.swp.swp_bcount = 0;
1716 
1717 		if (object->handle != NULL) {
1718 			mtx_lock(&sw_alloc_mtx);
1719 			TAILQ_INSERT_TAIL(
1720 			    NOBJLIST(object->handle),
1721 			    object,
1722 			    pager_object_list
1723 			);
1724 			mtx_unlock(&sw_alloc_mtx);
1725 		}
1726 	}
1727 
1728 	/*
1729 	 * Locate hash entry.  If not found create, but if we aren't adding
1730 	 * anything just return.  If we run out of space in the map we wait
1731 	 * and, since the hash table may have changed, retry.
1732 	 */
1733 retry:
1734 	mtx_lock(&swhash_mtx);
1735 	pswap = swp_pager_hash(object, pindex);
1736 
1737 	if ((swap = *pswap) == NULL) {
1738 		int i;
1739 
1740 		if (swapblk == SWAPBLK_NONE)
1741 			goto done;
1742 
1743 		swap = *pswap = uma_zalloc(swap_zone, M_NOWAIT);
1744 		if (swap == NULL) {
1745 			mtx_unlock(&swhash_mtx);
1746 			VM_OBJECT_UNLOCK(object);
1747 			VM_WAIT;
1748 			VM_OBJECT_LOCK(object);
1749 			goto retry;
1750 		}
1751 
1752 		swap->swb_hnext = NULL;
1753 		swap->swb_object = object;
1754 		swap->swb_index = pindex & ~(vm_pindex_t)SWAP_META_MASK;
1755 		swap->swb_count = 0;
1756 
1757 		++object->un_pager.swp.swp_bcount;
1758 
1759 		for (i = 0; i < SWAP_META_PAGES; ++i)
1760 			swap->swb_pages[i] = SWAPBLK_NONE;
1761 	}
1762 
1763 	/*
1764 	 * Delete prior contents of metadata
1765 	 */
1766 	idx = pindex & SWAP_META_MASK;
1767 
1768 	if (swap->swb_pages[idx] != SWAPBLK_NONE) {
1769 		swp_pager_freeswapspace(swap->swb_pages[idx], 1);
1770 		--swap->swb_count;
1771 	}
1772 
1773 	/*
1774 	 * Enter block into metadata
1775 	 */
1776 	swap->swb_pages[idx] = swapblk;
1777 	if (swapblk != SWAPBLK_NONE)
1778 		++swap->swb_count;
1779 done:
1780 	mtx_unlock(&swhash_mtx);
1781 }
1782 
1783 /*
1784  * SWP_PAGER_META_FREE() - free a range of blocks in the object's swap metadata
1785  *
1786  *	The requested range of blocks is freed, with any associated swap
1787  *	returned to the swap bitmap.
1788  *
1789  *	This routine will free swap metadata structures as they are cleaned
1790  *	out.  This routine does *NOT* operate on swap metadata associated
1791  *	with resident pages.
1792  *
1793  *	This routine must be called at splvm()
1794  */
1795 static void
1796 swp_pager_meta_free(vm_object_t object, vm_pindex_t index, daddr_t count)
1797 {
1798 
1799 	VM_OBJECT_LOCK_ASSERT(object, MA_OWNED);
1800 	if (object->type != OBJT_SWAP)
1801 		return;
1802 
1803 	while (count > 0) {
1804 		struct swblock **pswap;
1805 		struct swblock *swap;
1806 
1807 		mtx_lock(&swhash_mtx);
1808 		pswap = swp_pager_hash(object, index);
1809 
1810 		if ((swap = *pswap) != NULL) {
1811 			daddr_t v = swap->swb_pages[index & SWAP_META_MASK];
1812 
1813 			if (v != SWAPBLK_NONE) {
1814 				swp_pager_freeswapspace(v, 1);
1815 				swap->swb_pages[index & SWAP_META_MASK] =
1816 					SWAPBLK_NONE;
1817 				if (--swap->swb_count == 0) {
1818 					*pswap = swap->swb_hnext;
1819 					uma_zfree(swap_zone, swap);
1820 					--object->un_pager.swp.swp_bcount;
1821 				}
1822 			}
1823 			--count;
1824 			++index;
1825 		} else {
1826 			int n = SWAP_META_PAGES - (index & SWAP_META_MASK);
1827 			count -= n;
1828 			index += n;
1829 		}
1830 		mtx_unlock(&swhash_mtx);
1831 	}
1832 }
1833 
1834 /*
1835  * SWP_PAGER_META_FREE_ALL() - destroy all swap metadata associated with object
1836  *
1837  *	This routine locates and destroys all swap metadata associated with
1838  *	an object.
1839  *
1840  *	This routine must be called at splvm()
1841  */
1842 static void
1843 swp_pager_meta_free_all(vm_object_t object)
1844 {
1845 	daddr_t index = 0;
1846 
1847 	VM_OBJECT_LOCK_ASSERT(object, MA_OWNED);
1848 	if (object->type != OBJT_SWAP)
1849 		return;
1850 
1851 	while (object->un_pager.swp.swp_bcount) {
1852 		struct swblock **pswap;
1853 		struct swblock *swap;
1854 
1855 		mtx_lock(&swhash_mtx);
1856 		pswap = swp_pager_hash(object, index);
1857 		if ((swap = *pswap) != NULL) {
1858 			int i;
1859 
1860 			for (i = 0; i < SWAP_META_PAGES; ++i) {
1861 				daddr_t v = swap->swb_pages[i];
1862 				if (v != SWAPBLK_NONE) {
1863 					--swap->swb_count;
1864 					swp_pager_freeswapspace(v, 1);
1865 				}
1866 			}
1867 			if (swap->swb_count != 0)
1868 				panic("swap_pager_meta_free_all: swb_count != 0");
1869 			*pswap = swap->swb_hnext;
1870 			uma_zfree(swap_zone, swap);
1871 			--object->un_pager.swp.swp_bcount;
1872 		}
1873 		mtx_unlock(&swhash_mtx);
1874 		index += SWAP_META_PAGES;
1875 		if (index > 0x20000000)
1876 			panic("swp_pager_meta_free_all: failed to locate all swap meta blocks");
1877 	}
1878 }
1879 
1880 /*
1881  * SWP_PAGER_METACTL() -  misc control of swap and vm_page_t meta data.
1882  *
1883  *	This routine is capable of looking up, popping, or freeing
1884  *	swapblk assignments in the swap meta data or in the vm_page_t.
1885  *	The routine typically returns the swapblk being looked-up, or popped,
1886  *	or SWAPBLK_NONE if the block was freed, or SWAPBLK_NONE if the block
1887  *	was invalid.  This routine will automatically free any invalid
1888  *	meta-data swapblks.
1889  *
1890  *	It is not possible to store invalid swapblks in the swap meta data
1891  *	(other then a literal 'SWAPBLK_NONE'), so we don't bother checking.
1892  *
1893  *	When acting on a busy resident page and paging is in progress, we
1894  *	have to wait until paging is complete but otherwise can act on the
1895  *	busy page.
1896  *
1897  *	This routine must be called at splvm().
1898  *
1899  *	SWM_FREE	remove and free swap block from metadata
1900  *	SWM_POP		remove from meta data but do not free.. pop it out
1901  */
1902 static daddr_t
1903 swp_pager_meta_ctl(vm_object_t object, vm_pindex_t pindex, int flags)
1904 {
1905 	struct swblock **pswap;
1906 	struct swblock *swap;
1907 	daddr_t r1;
1908 	int idx;
1909 
1910 	VM_OBJECT_LOCK_ASSERT(object, MA_OWNED);
1911 	/*
1912 	 * The meta data only exists of the object is OBJT_SWAP
1913 	 * and even then might not be allocated yet.
1914 	 */
1915 	if (object->type != OBJT_SWAP)
1916 		return (SWAPBLK_NONE);
1917 
1918 	r1 = SWAPBLK_NONE;
1919 	mtx_lock(&swhash_mtx);
1920 	pswap = swp_pager_hash(object, pindex);
1921 
1922 	if ((swap = *pswap) != NULL) {
1923 		idx = pindex & SWAP_META_MASK;
1924 		r1 = swap->swb_pages[idx];
1925 
1926 		if (r1 != SWAPBLK_NONE) {
1927 			if (flags & SWM_FREE) {
1928 				swp_pager_freeswapspace(r1, 1);
1929 				r1 = SWAPBLK_NONE;
1930 			}
1931 			if (flags & (SWM_FREE|SWM_POP)) {
1932 				swap->swb_pages[idx] = SWAPBLK_NONE;
1933 				if (--swap->swb_count == 0) {
1934 					*pswap = swap->swb_hnext;
1935 					uma_zfree(swap_zone, swap);
1936 					--object->un_pager.swp.swp_bcount;
1937 				}
1938 			}
1939 		}
1940 	}
1941 	mtx_unlock(&swhash_mtx);
1942 	return (r1);
1943 }
1944 
1945 /*
1946  * System call swapon(name) enables swapping on device name,
1947  * which must be in the swdevsw.  Return EBUSY
1948  * if already swapping on this device.
1949  */
1950 #ifndef _SYS_SYSPROTO_H_
1951 struct swapon_args {
1952 	char *name;
1953 };
1954 #endif
1955 
1956 /*
1957  * MPSAFE
1958  */
1959 /* ARGSUSED */
1960 int
1961 swapon(struct thread *td, struct swapon_args *uap)
1962 {
1963 	struct vattr attr;
1964 	struct vnode *vp;
1965 	struct nameidata nd;
1966 	int error;
1967 
1968 	mtx_lock(&Giant);
1969 	error = suser(td);
1970 	if (error)
1971 		goto done2;
1972 
1973 	while (swdev_syscall_active)
1974 	    tsleep(&swdev_syscall_active, PUSER - 1, "swpon", 0);
1975 	swdev_syscall_active = 1;
1976 
1977 	/*
1978 	 * Swap metadata may not fit in the KVM if we have physical
1979 	 * memory of >1GB.
1980 	 */
1981 	if (swap_zone == NULL) {
1982 		error = ENOMEM;
1983 		goto done;
1984 	}
1985 
1986 	NDINIT(&nd, LOOKUP, ISOPEN | FOLLOW, UIO_USERSPACE, uap->name, td);
1987 	error = namei(&nd);
1988 	if (error)
1989 		goto done;
1990 
1991 	NDFREE(&nd, NDF_ONLY_PNBUF);
1992 	vp = nd.ni_vp;
1993 
1994 	if (vn_isdisk(vp, &error)) {
1995 		error = swapongeom(td, vp);
1996 	} else if (vp->v_type == VREG &&
1997 	    (vp->v_mount->mnt_vfc->vfc_flags & VFCF_NETWORK) != 0 &&
1998 	    (error = VOP_GETATTR(vp, &attr, td->td_ucred, td)) == 0) {
1999 		/*
2000 		 * Allow direct swapping to NFS regular files in the same
2001 		 * way that nfs_mountroot() sets up diskless swapping.
2002 		 */
2003 		error = swaponvp(td, vp, attr.va_size / DEV_BSIZE);
2004 	}
2005 
2006 	if (error)
2007 		vrele(vp);
2008 done:
2009 	swdev_syscall_active = 0;
2010 	wakeup_one(&swdev_syscall_active);
2011 done2:
2012 	mtx_unlock(&Giant);
2013 	return (error);
2014 }
2015 
2016 static void
2017 swaponsomething(struct vnode *vp, void *id, u_long nblks, sw_strategy_t *strategy, sw_close_t *close, dev_t dev)
2018 {
2019 	struct swdevt *sp, *tsp;
2020 	swblk_t dvbase;
2021 	u_long mblocks;
2022 
2023 	/*
2024 	 * If we go beyond this, we get overflows in the radix
2025 	 * tree bitmap code.
2026 	 */
2027 	mblocks = 0x40000000 / BLIST_META_RADIX;
2028 	if (nblks > mblocks) {
2029 		printf("WARNING: reducing size to maximum of %lu blocks per swap unit\n",
2030 			mblocks);
2031 		nblks = mblocks;
2032 	}
2033 	/*
2034 	 * nblks is in DEV_BSIZE'd chunks, convert to PAGE_SIZE'd chunks.
2035 	 * First chop nblks off to page-align it, then convert.
2036 	 *
2037 	 * sw->sw_nblks is in page-sized chunks now too.
2038 	 */
2039 	nblks &= ~(ctodb(1) - 1);
2040 	nblks = dbtoc(nblks);
2041 
2042 	sp = malloc(sizeof *sp, M_VMPGDATA, M_WAITOK | M_ZERO);
2043 	sp->sw_vp = vp;
2044 	sp->sw_id = id;
2045 	sp->sw_dev = dev;
2046 	sp->sw_flags = 0;
2047 	sp->sw_nblks = nblks;
2048 	sp->sw_used = 0;
2049 	sp->sw_strategy = strategy;
2050 	sp->sw_close = close;
2051 
2052 	sp->sw_blist = blist_create(nblks);
2053 	/*
2054 	 * Do not free the first two block in order to avoid overwriting
2055 	 * any bsd label at the front of the partition
2056 	 */
2057 	blist_free(sp->sw_blist, 2, nblks - 2);
2058 
2059 	dvbase = 0;
2060 	mtx_lock(&sw_dev_mtx);
2061 	TAILQ_FOREACH(tsp, &swtailq, sw_list) {
2062 		if (tsp->sw_end >= dvbase) {
2063 			/*
2064 			 * We put one uncovered page between the devices
2065 			 * in order to definitively prevent any cross-device
2066 			 * I/O requests
2067 			 */
2068 			dvbase = tsp->sw_end + 1;
2069 		}
2070 	}
2071 	sp->sw_first = dvbase;
2072 	sp->sw_end = dvbase + nblks;
2073 	TAILQ_INSERT_TAIL(&swtailq, sp, sw_list);
2074 	nswapdev++;
2075 	swap_pager_avail += nblks;
2076 	swp_sizecheck();
2077 	mtx_unlock(&sw_dev_mtx);
2078 }
2079 
2080 /*
2081  * SYSCALL: swapoff(devname)
2082  *
2083  * Disable swapping on the given device.
2084  *
2085  * XXX: Badly designed system call: it should use a device index
2086  * rather than filename as specification.  We keep sw_vp around
2087  * only to make this work.
2088  */
2089 #ifndef _SYS_SYSPROTO_H_
2090 struct swapoff_args {
2091 	char *name;
2092 };
2093 #endif
2094 
2095 /*
2096  * MPSAFE
2097  */
2098 /* ARGSUSED */
2099 int
2100 swapoff(struct thread *td, struct swapoff_args *uap)
2101 {
2102 	struct vnode *vp;
2103 	struct nameidata nd;
2104 	struct swdevt *sp;
2105 	u_long nblks, dvbase;
2106 	int error;
2107 
2108 	mtx_lock(&Giant);
2109 
2110 	error = suser(td);
2111 	if (error)
2112 		goto done2;
2113 
2114 	while (swdev_syscall_active)
2115 	    tsleep(&swdev_syscall_active, PUSER - 1, "swpoff", 0);
2116 	swdev_syscall_active = 1;
2117 
2118 	NDINIT(&nd, LOOKUP, FOLLOW, UIO_USERSPACE, uap->name, td);
2119 	error = namei(&nd);
2120 	if (error)
2121 		goto done;
2122 	NDFREE(&nd, NDF_ONLY_PNBUF);
2123 	vp = nd.ni_vp;
2124 
2125 	mtx_lock(&sw_dev_mtx);
2126 	TAILQ_FOREACH(sp, &swtailq, sw_list) {
2127 		if (sp->sw_vp == vp)
2128 			goto found;
2129 	}
2130 	mtx_unlock(&sw_dev_mtx);
2131 	error = EINVAL;
2132 	goto done;
2133 found:
2134 	mtx_unlock(&sw_dev_mtx);
2135 #ifdef MAC
2136 	(void) vn_lock(vp, LK_EXCLUSIVE | LK_RETRY, td);
2137 	error = mac_check_system_swapoff(td->td_ucred, vp);
2138 	(void) VOP_UNLOCK(vp, 0, td);
2139 	if (error != 0)
2140 		goto done;
2141 #endif
2142 
2143 	nblks = sp->sw_nblks;
2144 
2145 	/*
2146 	 * We can turn off this swap device safely only if the
2147 	 * available virtual memory in the system will fit the amount
2148 	 * of data we will have to page back in, plus an epsilon so
2149 	 * the system doesn't become critically low on swap space.
2150 	 */
2151 	if (cnt.v_free_count + cnt.v_cache_count + swap_pager_avail <
2152 	    nblks + nswap_lowat) {
2153 		error = ENOMEM;
2154 		goto done;
2155 	}
2156 
2157 	/*
2158 	 * Prevent further allocations on this device.
2159 	 */
2160 	mtx_lock(&sw_dev_mtx);
2161 	sp->sw_flags |= SW_CLOSING;
2162 	for (dvbase = 0; dvbase < sp->sw_end; dvbase += dmmax) {
2163 		swap_pager_avail -= blist_fill(sp->sw_blist,
2164 		     dvbase, dmmax);
2165 	}
2166 	mtx_unlock(&sw_dev_mtx);
2167 
2168 	/*
2169 	 * Page in the contents of the device and close it.
2170 	 */
2171 	swap_pager_swapoff(sp);
2172 
2173 	sp->sw_close(td, sp);
2174 	sp->sw_id = NULL;
2175 	mtx_lock(&sw_dev_mtx);
2176 	TAILQ_REMOVE(&swtailq, sp, sw_list);
2177 	nswapdev--;
2178 	if (nswapdev == 0) {
2179 		swap_pager_full = 2;
2180 		swap_pager_almost_full = 1;
2181 	}
2182 	if (swdevhd == sp)
2183 		swdevhd = NULL;
2184 	mtx_unlock(&sw_dev_mtx);
2185 	blist_destroy(sp->sw_blist);
2186 	free(sp, M_VMPGDATA);
2187 
2188 done:
2189 	swdev_syscall_active = 0;
2190 	wakeup_one(&swdev_syscall_active);
2191 done2:
2192 	mtx_unlock(&Giant);
2193 	return (error);
2194 }
2195 
2196 void
2197 swap_pager_status(int *total, int *used)
2198 {
2199 	struct swdevt *sp;
2200 
2201 	*total = 0;
2202 	*used = 0;
2203 	mtx_lock(&sw_dev_mtx);
2204 	TAILQ_FOREACH(sp, &swtailq, sw_list) {
2205 		*total += sp->sw_nblks;
2206 		*used += sp->sw_used;
2207 	}
2208 	mtx_unlock(&sw_dev_mtx);
2209 }
2210 
2211 static int
2212 sysctl_vm_swap_info(SYSCTL_HANDLER_ARGS)
2213 {
2214 	int	*name = (int *)arg1;
2215 	int	error, n;
2216 	struct xswdev xs;
2217 	struct swdevt *sp;
2218 
2219 	if (arg2 != 1) /* name length */
2220 		return (EINVAL);
2221 
2222 	n = 0;
2223 	mtx_lock(&sw_dev_mtx);
2224 	TAILQ_FOREACH(sp, &swtailq, sw_list) {
2225 		if (n == *name) {
2226 			mtx_unlock(&sw_dev_mtx);
2227 			xs.xsw_version = XSWDEV_VERSION;
2228 			xs.xsw_dev = sp->sw_dev;
2229 			xs.xsw_flags = sp->sw_flags;
2230 			xs.xsw_nblks = sp->sw_nblks;
2231 			xs.xsw_used = sp->sw_used;
2232 
2233 			error = SYSCTL_OUT(req, &xs, sizeof(xs));
2234 			return (error);
2235 		}
2236 		n++;
2237 	}
2238 	mtx_unlock(&sw_dev_mtx);
2239 	return (ENOENT);
2240 }
2241 
2242 SYSCTL_INT(_vm, OID_AUTO, nswapdev, CTLFLAG_RD, &nswapdev, 0,
2243     "Number of swap devices");
2244 SYSCTL_NODE(_vm, OID_AUTO, swap_info, CTLFLAG_RD, sysctl_vm_swap_info,
2245     "Swap statistics by device");
2246 
2247 /*
2248  * vmspace_swap_count() - count the approximate swap useage in pages for a
2249  *			  vmspace.
2250  *
2251  *	The map must be locked.
2252  *
2253  *	Swap useage is determined by taking the proportional swap used by
2254  *	VM objects backing the VM map.  To make up for fractional losses,
2255  *	if the VM object has any swap use at all the associated map entries
2256  *	count for at least 1 swap page.
2257  */
2258 int
2259 vmspace_swap_count(struct vmspace *vmspace)
2260 {
2261 	vm_map_t map = &vmspace->vm_map;
2262 	vm_map_entry_t cur;
2263 	int count = 0;
2264 
2265 	for (cur = map->header.next; cur != &map->header; cur = cur->next) {
2266 		vm_object_t object;
2267 
2268 		if ((cur->eflags & MAP_ENTRY_IS_SUB_MAP) == 0 &&
2269 		    (object = cur->object.vm_object) != NULL) {
2270 			VM_OBJECT_LOCK(object);
2271 			if (object->type == OBJT_SWAP &&
2272 			    object->un_pager.swp.swp_bcount != 0) {
2273 				int n = (cur->end - cur->start) / PAGE_SIZE;
2274 
2275 				count += object->un_pager.swp.swp_bcount *
2276 				    SWAP_META_PAGES * n / object->size + 1;
2277 			}
2278 			VM_OBJECT_UNLOCK(object);
2279 		}
2280 	}
2281 	return (count);
2282 }
2283 
2284 /*
2285  * GEOM backend
2286  *
2287  * Swapping onto disk devices.
2288  *
2289  */
2290 
2291 static g_orphan_t swapgeom_orphan;
2292 
2293 static struct g_class g_swap_class = {
2294 	.name = "SWAP",
2295 	.version = G_VERSION,
2296 	.orphan = swapgeom_orphan,
2297 };
2298 
2299 DECLARE_GEOM_CLASS(g_swap_class, g_class);
2300 
2301 
2302 static void
2303 swapgeom_done(struct bio *bp2)
2304 {
2305 	struct buf *bp;
2306 
2307 	bp = bp2->bio_caller2;
2308 	bp->b_ioflags = bp2->bio_flags;
2309 	if (bp2->bio_error)
2310 		bp->b_ioflags |= BIO_ERROR;
2311 	bp->b_resid = bp->b_bcount - bp2->bio_completed;
2312 	bp->b_error = bp2->bio_error;
2313 	bufdone(bp);
2314 	g_destroy_bio(bp2);
2315 }
2316 
2317 static void
2318 swapgeom_strategy(struct buf *bp, struct swdevt *sp)
2319 {
2320 	struct bio *bio;
2321 	struct g_consumer *cp;
2322 
2323 	cp = sp->sw_id;
2324 	if (cp == NULL) {
2325 		bp->b_error = ENXIO;
2326 		bp->b_ioflags |= BIO_ERROR;
2327 		bufdone(bp);
2328 		return;
2329 	}
2330 	bio = g_alloc_bio();
2331 #if 0
2332 	/*
2333 	 * XXX: We shouldn't really sleep here when we run out of buffers
2334 	 * XXX: but the alternative is worse right now.
2335 	 */
2336 	if (bio == NULL) {
2337 		bp->b_error = ENOMEM;
2338 		bp->b_ioflags |= BIO_ERROR;
2339 		bufdone(bp);
2340 		return;
2341 	}
2342 #endif
2343 	bio->bio_caller2 = bp;
2344 	bio->bio_cmd = bp->b_iocmd;
2345 	bio->bio_data = bp->b_data;
2346 	bio->bio_offset = (bp->b_blkno - sp->sw_first) * PAGE_SIZE;
2347 	bio->bio_length = bp->b_bcount;
2348 	bio->bio_done = swapgeom_done;
2349 	g_io_request(bio, cp);
2350 	return;
2351 }
2352 
2353 static void
2354 swapgeom_orphan(struct g_consumer *cp)
2355 {
2356 	struct swdevt *sp;
2357 
2358 	mtx_lock(&sw_dev_mtx);
2359 	TAILQ_FOREACH(sp, &swtailq, sw_list)
2360 		if (sp->sw_id == cp)
2361 			sp->sw_id = NULL;
2362 	mtx_unlock(&sw_dev_mtx);
2363 }
2364 
2365 static void
2366 swapgeom_close_ev(void *arg, int flags)
2367 {
2368 	struct g_consumer *cp;
2369 
2370 	cp = arg;
2371 	g_access(cp, -1, -1, 0);
2372 	g_detach(cp);
2373 	g_destroy_consumer(cp);
2374 }
2375 
2376 static void
2377 swapgeom_close(struct thread *td, struct swdevt *sw)
2378 {
2379 
2380 	/* XXX: direct call when Giant untangled */
2381 	g_waitfor_event(swapgeom_close_ev, sw->sw_id, M_WAITOK, NULL);
2382 }
2383 
2384 
2385 struct swh0h0 {
2386 	struct cdev *dev;
2387 	struct vnode *vp;
2388 	int	error;
2389 };
2390 
2391 static void
2392 swapongeom_ev(void *arg, int flags)
2393 {
2394 	struct swh0h0 *swh;
2395 	struct g_provider *pp;
2396 	struct g_consumer *cp;
2397 	static struct g_geom *gp;
2398 	struct swdevt *sp;
2399 	u_long nblks;
2400 	int error;
2401 
2402 	swh = arg;
2403 	swh->error = 0;
2404 	pp = g_dev_getprovider(swh->dev);
2405 	if (pp == NULL) {
2406 		swh->error = ENODEV;
2407 		return;
2408 	}
2409 	mtx_lock(&sw_dev_mtx);
2410 	TAILQ_FOREACH(sp, &swtailq, sw_list) {
2411 		cp = sp->sw_id;
2412 		if (cp != NULL && cp->provider == pp) {
2413 			mtx_unlock(&sw_dev_mtx);
2414 			swh->error = EBUSY;
2415 			return;
2416 		}
2417 	}
2418 	mtx_unlock(&sw_dev_mtx);
2419 	if (gp == NULL)
2420 		gp = g_new_geomf(&g_swap_class, "swap", NULL);
2421 	cp = g_new_consumer(gp);
2422 	g_attach(cp, pp);
2423 	/*
2424 	 * XXX: Everytime you think you can improve the margin for
2425 	 * footshooting, somebody depends on the ability to do so:
2426 	 * savecore(8) wants to write to our swapdev so we cannot
2427 	 * set an exclusive count :-(
2428 	 */
2429 	error = g_access(cp, 1, 1, 0);
2430 	if (error) {
2431 		g_detach(cp);
2432 		g_destroy_consumer(cp);
2433 		swh->error = error;
2434 		return;
2435 	}
2436 	nblks = pp->mediasize / DEV_BSIZE;
2437 	swaponsomething(swh->vp, cp, nblks, swapgeom_strategy,
2438 	    swapgeom_close, dev2udev(swh->dev));
2439 	swh->error = 0;
2440 	return;
2441 }
2442 
2443 static int
2444 swapongeom(struct thread *td, struct vnode *vp)
2445 {
2446 	int error;
2447 	struct swh0h0 swh;
2448 
2449 	vn_lock(vp, LK_EXCLUSIVE | LK_RETRY, td);
2450 
2451 	swh.dev = vp->v_rdev;
2452 	swh.vp = vp;
2453 	swh.error = 0;
2454 	/* XXX: direct call when Giant untangled */
2455 	error = g_waitfor_event(swapongeom_ev, &swh, M_WAITOK, NULL);
2456 	if (!error)
2457 		error = swh.error;
2458 	VOP_UNLOCK(vp, 0, td);
2459 	return (error);
2460 }
2461 
2462 /*
2463  * VNODE backend
2464  *
2465  * This is used mainly for network filesystem (read: probably only tested
2466  * with NFS) swapfiles.
2467  *
2468  */
2469 
2470 static void
2471 swapdev_strategy(struct buf *bp, struct swdevt *sp)
2472 {
2473 	struct vnode *vp2;
2474 
2475 	bp->b_blkno = ctodb(bp->b_blkno - sp->sw_first);
2476 
2477 	vp2 = sp->sw_id;
2478 	vhold(vp2);
2479 	if (bp->b_iocmd == BIO_WRITE) {
2480 		if (bp->b_bufobj) /* XXX: should always be true /phk */
2481 			bufobj_wdrop(bp->b_bufobj);
2482 		bufobj_wref(&vp2->v_bufobj);
2483 	}
2484 	bp->b_vp = vp2;
2485 	bp->b_iooffset = dbtob(bp->b_blkno);
2486 	bstrategy(bp);
2487 	return;
2488 }
2489 
2490 static void
2491 swapdev_close(struct thread *td, struct swdevt *sp)
2492 {
2493 
2494 	VOP_CLOSE(sp->sw_vp, FREAD | FWRITE, td->td_ucred, td);
2495 	vrele(sp->sw_vp);
2496 }
2497 
2498 
2499 static int
2500 swaponvp(struct thread *td, struct vnode *vp, u_long nblks)
2501 {
2502 	struct swdevt *sp;
2503 	int error;
2504 
2505 	if (nblks == 0)
2506 		return (ENXIO);
2507 	mtx_lock(&sw_dev_mtx);
2508 	TAILQ_FOREACH(sp, &swtailq, sw_list) {
2509 		if (sp->sw_id == vp) {
2510 			mtx_unlock(&sw_dev_mtx);
2511 			return (EBUSY);
2512 		}
2513 	}
2514 	mtx_unlock(&sw_dev_mtx);
2515 
2516 	(void) vn_lock(vp, LK_EXCLUSIVE | LK_RETRY, td);
2517 #ifdef MAC
2518 	error = mac_check_system_swapon(td->td_ucred, vp);
2519 	if (error == 0)
2520 #endif
2521 		error = VOP_OPEN(vp, FREAD | FWRITE, td->td_ucred, td, -1);
2522 	(void) VOP_UNLOCK(vp, 0, td);
2523 	if (error)
2524 		return (error);
2525 
2526 	swaponsomething(vp, vp, nblks, swapdev_strategy, swapdev_close,
2527 	    NODEV);
2528 	return (0);
2529 }
2530