xref: /freebsd/sys/vm/swap_pager.c (revision 50896984cd36c1724905e41819f5236399a4972d)
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_compat.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/priv.h>
81 #include <sys/proc.h>
82 #include <sys/bio.h>
83 #include <sys/buf.h>
84 #include <sys/disk.h>
85 #include <sys/fcntl.h>
86 #include <sys/mount.h>
87 #include <sys/namei.h>
88 #include <sys/vnode.h>
89 #include <sys/malloc.h>
90 #include <sys/pctrie.h>
91 #include <sys/racct.h>
92 #include <sys/resource.h>
93 #include <sys/resourcevar.h>
94 #include <sys/rwlock.h>
95 #include <sys/sbuf.h>
96 #include <sys/sysctl.h>
97 #include <sys/sysproto.h>
98 #include <sys/blist.h>
99 #include <sys/lock.h>
100 #include <sys/sx.h>
101 #include <sys/vmmeter.h>
102 
103 #include <security/mac/mac_framework.h>
104 
105 #include <vm/vm.h>
106 #include <vm/pmap.h>
107 #include <vm/vm_map.h>
108 #include <vm/vm_kern.h>
109 #include <vm/vm_object.h>
110 #include <vm/vm_page.h>
111 #include <vm/vm_pager.h>
112 #include <vm/vm_pageout.h>
113 #include <vm/vm_param.h>
114 #include <vm/swap_pager.h>
115 #include <vm/vm_extern.h>
116 #include <vm/uma.h>
117 
118 #include <geom/geom.h>
119 
120 /*
121  * MAX_PAGEOUT_CLUSTER must be a power of 2 between 1 and 64.
122  * The 64-page limit is due to the radix code (kern/subr_blist.c).
123  */
124 #ifndef MAX_PAGEOUT_CLUSTER
125 #define	MAX_PAGEOUT_CLUSTER	32
126 #endif
127 
128 #if !defined(SWB_NPAGES)
129 #define SWB_NPAGES	MAX_PAGEOUT_CLUSTER
130 #endif
131 
132 #define	SWAP_META_PAGES		PCTRIE_COUNT
133 
134 /*
135  * A swblk structure maps each page index within a
136  * SWAP_META_PAGES-aligned and sized range to the address of an
137  * on-disk swap block (or SWAPBLK_NONE). The collection of these
138  * mappings for an entire vm object is implemented as a pc-trie.
139  */
140 struct swblk {
141 	vm_pindex_t	p;
142 	daddr_t		d[SWAP_META_PAGES];
143 };
144 
145 static MALLOC_DEFINE(M_VMPGDATA, "vm_pgdata", "swap pager private data");
146 static struct mtx sw_dev_mtx;
147 static TAILQ_HEAD(, swdevt) swtailq = TAILQ_HEAD_INITIALIZER(swtailq);
148 static struct swdevt *swdevhd;	/* Allocate from here next */
149 static int nswapdev;		/* Number of swap devices */
150 int swap_pager_avail;
151 static struct sx swdev_syscall_lock;	/* serialize swap(on|off) */
152 
153 static vm_ooffset_t swap_total;
154 SYSCTL_QUAD(_vm, OID_AUTO, swap_total, CTLFLAG_RD, &swap_total, 0,
155     "Total amount of available swap storage.");
156 static vm_ooffset_t swap_reserved;
157 SYSCTL_QUAD(_vm, OID_AUTO, swap_reserved, CTLFLAG_RD, &swap_reserved, 0,
158     "Amount of swap storage needed to back all allocated anonymous memory.");
159 static int overcommit = 0;
160 SYSCTL_INT(_vm, OID_AUTO, overcommit, CTLFLAG_RW, &overcommit, 0,
161     "Configure virtual memory overcommit behavior. See tuning(7) "
162     "for details.");
163 static unsigned long swzone;
164 SYSCTL_ULONG(_vm, OID_AUTO, swzone, CTLFLAG_RD, &swzone, 0,
165     "Actual size of swap metadata zone");
166 static unsigned long swap_maxpages;
167 SYSCTL_ULONG(_vm, OID_AUTO, swap_maxpages, CTLFLAG_RD, &swap_maxpages, 0,
168     "Maximum amount of swap supported");
169 
170 /* bits from overcommit */
171 #define	SWAP_RESERVE_FORCE_ON		(1 << 0)
172 #define	SWAP_RESERVE_RLIMIT_ON		(1 << 1)
173 #define	SWAP_RESERVE_ALLOW_NONWIRED	(1 << 2)
174 
175 int
176 swap_reserve(vm_ooffset_t incr)
177 {
178 
179 	return (swap_reserve_by_cred(incr, curthread->td_ucred));
180 }
181 
182 int
183 swap_reserve_by_cred(vm_ooffset_t incr, struct ucred *cred)
184 {
185 	vm_ooffset_t r, s;
186 	int res, error;
187 	static int curfail;
188 	static struct timeval lastfail;
189 	struct uidinfo *uip;
190 
191 	uip = cred->cr_ruidinfo;
192 
193 	if (incr & PAGE_MASK)
194 		panic("swap_reserve: & PAGE_MASK");
195 
196 #ifdef RACCT
197 	if (racct_enable) {
198 		PROC_LOCK(curproc);
199 		error = racct_add(curproc, RACCT_SWAP, incr);
200 		PROC_UNLOCK(curproc);
201 		if (error != 0)
202 			return (0);
203 	}
204 #endif
205 
206 	res = 0;
207 	mtx_lock(&sw_dev_mtx);
208 	r = swap_reserved + incr;
209 	if (overcommit & SWAP_RESERVE_ALLOW_NONWIRED) {
210 		s = vm_cnt.v_page_count - vm_cnt.v_free_reserved - vm_cnt.v_wire_count;
211 		s *= PAGE_SIZE;
212 	} else
213 		s = 0;
214 	s += swap_total;
215 	if ((overcommit & SWAP_RESERVE_FORCE_ON) == 0 || r <= s ||
216 	    (error = priv_check(curthread, PRIV_VM_SWAP_NOQUOTA)) == 0) {
217 		res = 1;
218 		swap_reserved = r;
219 	}
220 	mtx_unlock(&sw_dev_mtx);
221 
222 	if (res) {
223 		UIDINFO_VMSIZE_LOCK(uip);
224 		if ((overcommit & SWAP_RESERVE_RLIMIT_ON) != 0 &&
225 		    uip->ui_vmsize + incr > lim_cur(curthread, RLIMIT_SWAP) &&
226 		    priv_check(curthread, PRIV_VM_SWAP_NORLIMIT))
227 			res = 0;
228 		else
229 			uip->ui_vmsize += incr;
230 		UIDINFO_VMSIZE_UNLOCK(uip);
231 		if (!res) {
232 			mtx_lock(&sw_dev_mtx);
233 			swap_reserved -= incr;
234 			mtx_unlock(&sw_dev_mtx);
235 		}
236 	}
237 	if (!res && ppsratecheck(&lastfail, &curfail, 1)) {
238 		printf("uid %d, pid %d: swap reservation for %jd bytes failed\n",
239 		    uip->ui_uid, curproc->p_pid, incr);
240 	}
241 
242 #ifdef RACCT
243 	if (!res) {
244 		PROC_LOCK(curproc);
245 		racct_sub(curproc, RACCT_SWAP, incr);
246 		PROC_UNLOCK(curproc);
247 	}
248 #endif
249 
250 	return (res);
251 }
252 
253 void
254 swap_reserve_force(vm_ooffset_t incr)
255 {
256 	struct uidinfo *uip;
257 
258 	mtx_lock(&sw_dev_mtx);
259 	swap_reserved += incr;
260 	mtx_unlock(&sw_dev_mtx);
261 
262 #ifdef RACCT
263 	PROC_LOCK(curproc);
264 	racct_add_force(curproc, RACCT_SWAP, incr);
265 	PROC_UNLOCK(curproc);
266 #endif
267 
268 	uip = curthread->td_ucred->cr_ruidinfo;
269 	PROC_LOCK(curproc);
270 	UIDINFO_VMSIZE_LOCK(uip);
271 	uip->ui_vmsize += incr;
272 	UIDINFO_VMSIZE_UNLOCK(uip);
273 	PROC_UNLOCK(curproc);
274 }
275 
276 void
277 swap_release(vm_ooffset_t decr)
278 {
279 	struct ucred *cred;
280 
281 	PROC_LOCK(curproc);
282 	cred = curthread->td_ucred;
283 	swap_release_by_cred(decr, cred);
284 	PROC_UNLOCK(curproc);
285 }
286 
287 void
288 swap_release_by_cred(vm_ooffset_t decr, struct ucred *cred)
289 {
290  	struct uidinfo *uip;
291 
292 	uip = cred->cr_ruidinfo;
293 
294 	if (decr & PAGE_MASK)
295 		panic("swap_release: & PAGE_MASK");
296 
297 	mtx_lock(&sw_dev_mtx);
298 	if (swap_reserved < decr)
299 		panic("swap_reserved < decr");
300 	swap_reserved -= decr;
301 	mtx_unlock(&sw_dev_mtx);
302 
303 	UIDINFO_VMSIZE_LOCK(uip);
304 	if (uip->ui_vmsize < decr)
305 		printf("negative vmsize for uid = %d\n", uip->ui_uid);
306 	uip->ui_vmsize -= decr;
307 	UIDINFO_VMSIZE_UNLOCK(uip);
308 
309 	racct_sub_cred(cred, RACCT_SWAP, decr);
310 }
311 
312 #define SWM_FREE	0x02	/* free, period			*/
313 #define SWM_POP		0x04	/* pop out			*/
314 
315 static int swap_pager_full = 2;	/* swap space exhaustion (task killing) */
316 static int swap_pager_almost_full = 1; /* swap space exhaustion (w/hysteresis)*/
317 static int nsw_rcount;		/* free read buffers			*/
318 static int nsw_wcount_sync;	/* limit write buffers / synchronous	*/
319 static int nsw_wcount_async;	/* limit write buffers / asynchronous	*/
320 static int nsw_wcount_async_max;/* assigned maximum			*/
321 static int nsw_cluster_max;	/* maximum VOP I/O allowed		*/
322 
323 static int sysctl_swap_async_max(SYSCTL_HANDLER_ARGS);
324 SYSCTL_PROC(_vm, OID_AUTO, swap_async_max, CTLTYPE_INT | CTLFLAG_RW |
325     CTLFLAG_MPSAFE, NULL, 0, sysctl_swap_async_max, "I",
326     "Maximum running async swap ops");
327 static int sysctl_swap_fragmentation(SYSCTL_HANDLER_ARGS);
328 SYSCTL_PROC(_vm, OID_AUTO, swap_fragmentation, CTLTYPE_STRING | CTLFLAG_RD |
329     CTLFLAG_MPSAFE, NULL, 0, sysctl_swap_fragmentation, "A",
330     "Swap Fragmentation Info");
331 
332 static struct sx sw_alloc_sx;
333 
334 /*
335  * "named" and "unnamed" anon region objects.  Try to reduce the overhead
336  * of searching a named list by hashing it just a little.
337  */
338 
339 #define NOBJLISTS		8
340 
341 #define NOBJLIST(handle)	\
342 	(&swap_pager_object_list[((int)(intptr_t)handle >> 4) & (NOBJLISTS-1)])
343 
344 static struct pagerlst	swap_pager_object_list[NOBJLISTS];
345 static uma_zone_t swblk_zone;
346 static uma_zone_t swpctrie_zone;
347 
348 /*
349  * pagerops for OBJT_SWAP - "swap pager".  Some ops are also global procedure
350  * calls hooked from other parts of the VM system and do not appear here.
351  * (see vm/swap_pager.h).
352  */
353 static vm_object_t
354 		swap_pager_alloc(void *handle, vm_ooffset_t size,
355 		    vm_prot_t prot, vm_ooffset_t offset, struct ucred *);
356 static void	swap_pager_dealloc(vm_object_t object);
357 static int	swap_pager_getpages(vm_object_t, vm_page_t *, int, int *,
358     int *);
359 static int	swap_pager_getpages_async(vm_object_t, vm_page_t *, int, int *,
360     int *, pgo_getpages_iodone_t, void *);
361 static void	swap_pager_putpages(vm_object_t, vm_page_t *, int, boolean_t, int *);
362 static boolean_t
363 		swap_pager_haspage(vm_object_t object, vm_pindex_t pindex, int *before, int *after);
364 static void	swap_pager_init(void);
365 static void	swap_pager_unswapped(vm_page_t);
366 static void	swap_pager_swapoff(struct swdevt *sp);
367 
368 struct pagerops swappagerops = {
369 	.pgo_init =	swap_pager_init,	/* early system initialization of pager	*/
370 	.pgo_alloc =	swap_pager_alloc,	/* allocate an OBJT_SWAP object		*/
371 	.pgo_dealloc =	swap_pager_dealloc,	/* deallocate an OBJT_SWAP object	*/
372 	.pgo_getpages =	swap_pager_getpages,	/* pagein				*/
373 	.pgo_getpages_async = swap_pager_getpages_async, /* pagein (async)		*/
374 	.pgo_putpages =	swap_pager_putpages,	/* pageout				*/
375 	.pgo_haspage =	swap_pager_haspage,	/* get backing store status for page	*/
376 	.pgo_pageunswapped = swap_pager_unswapped,	/* remove swap related to page		*/
377 };
378 
379 /*
380  * swap_*() routines are externally accessible.  swp_*() routines are
381  * internal.
382  */
383 static int nswap_lowat = 128;	/* in pages, swap_pager_almost_full warn */
384 static int nswap_hiwat = 512;	/* in pages, swap_pager_almost_full warn */
385 
386 SYSCTL_INT(_vm, OID_AUTO, dmmax, CTLFLAG_RD, &nsw_cluster_max, 0,
387     "Maximum size of a swap block in pages");
388 
389 static void	swp_sizecheck(void);
390 static void	swp_pager_async_iodone(struct buf *bp);
391 static int	swapongeom(struct vnode *);
392 static int	swaponvp(struct thread *, struct vnode *, u_long);
393 static int	swapoff_one(struct swdevt *sp, struct ucred *cred);
394 
395 /*
396  * Swap bitmap functions
397  */
398 static void	swp_pager_freeswapspace(daddr_t blk, int npages);
399 static daddr_t	swp_pager_getswapspace(int npages);
400 
401 /*
402  * Metadata functions
403  */
404 static void swp_pager_meta_build(vm_object_t, vm_pindex_t, daddr_t);
405 static void swp_pager_meta_free(vm_object_t, vm_pindex_t, vm_pindex_t);
406 static void swp_pager_meta_free_all(vm_object_t);
407 static daddr_t swp_pager_meta_ctl(vm_object_t, vm_pindex_t, int);
408 
409 static void *
410 swblk_trie_alloc(struct pctrie *ptree)
411 {
412 
413 	return (uma_zalloc(swpctrie_zone, M_NOWAIT | (curproc == pageproc ?
414 	    M_USE_RESERVE : 0)));
415 }
416 
417 static void
418 swblk_trie_free(struct pctrie *ptree, void *node)
419 {
420 
421 	uma_zfree(swpctrie_zone, node);
422 }
423 
424 PCTRIE_DEFINE(SWAP, swblk, p, swblk_trie_alloc, swblk_trie_free);
425 
426 /*
427  * SWP_SIZECHECK() -	update swap_pager_full indication
428  *
429  *	update the swap_pager_almost_full indication and warn when we are
430  *	about to run out of swap space, using lowat/hiwat hysteresis.
431  *
432  *	Clear swap_pager_full ( task killing ) indication when lowat is met.
433  *
434  *	No restrictions on call
435  *	This routine may not block.
436  */
437 static void
438 swp_sizecheck(void)
439 {
440 
441 	if (swap_pager_avail < nswap_lowat) {
442 		if (swap_pager_almost_full == 0) {
443 			printf("swap_pager: out of swap space\n");
444 			swap_pager_almost_full = 1;
445 		}
446 	} else {
447 		swap_pager_full = 0;
448 		if (swap_pager_avail > nswap_hiwat)
449 			swap_pager_almost_full = 0;
450 	}
451 }
452 
453 /*
454  * SWAP_PAGER_INIT() -	initialize the swap pager!
455  *
456  *	Expected to be started from system init.  NOTE:  This code is run
457  *	before much else so be careful what you depend on.  Most of the VM
458  *	system has yet to be initialized at this point.
459  */
460 static void
461 swap_pager_init(void)
462 {
463 	/*
464 	 * Initialize object lists
465 	 */
466 	int i;
467 
468 	for (i = 0; i < NOBJLISTS; ++i)
469 		TAILQ_INIT(&swap_pager_object_list[i]);
470 	mtx_init(&sw_dev_mtx, "swapdev", NULL, MTX_DEF);
471 	sx_init(&sw_alloc_sx, "swspsx");
472 	sx_init(&swdev_syscall_lock, "swsysc");
473 }
474 
475 /*
476  * SWAP_PAGER_SWAP_INIT() - swap pager initialization from pageout process
477  *
478  *	Expected to be started from pageout process once, prior to entering
479  *	its main loop.
480  */
481 void
482 swap_pager_swap_init(void)
483 {
484 	unsigned long n, n2;
485 
486 	/*
487 	 * Number of in-transit swap bp operations.  Don't
488 	 * exhaust the pbufs completely.  Make sure we
489 	 * initialize workable values (0 will work for hysteresis
490 	 * but it isn't very efficient).
491 	 *
492 	 * The nsw_cluster_max is constrained by the bp->b_pages[]
493 	 * array (MAXPHYS/PAGE_SIZE) and our locally defined
494 	 * MAX_PAGEOUT_CLUSTER.   Also be aware that swap ops are
495 	 * constrained by the swap device interleave stripe size.
496 	 *
497 	 * Currently we hardwire nsw_wcount_async to 4.  This limit is
498 	 * designed to prevent other I/O from having high latencies due to
499 	 * our pageout I/O.  The value 4 works well for one or two active swap
500 	 * devices but is probably a little low if you have more.  Even so,
501 	 * a higher value would probably generate only a limited improvement
502 	 * with three or four active swap devices since the system does not
503 	 * typically have to pageout at extreme bandwidths.   We will want
504 	 * at least 2 per swap devices, and 4 is a pretty good value if you
505 	 * have one NFS swap device due to the command/ack latency over NFS.
506 	 * So it all works out pretty well.
507 	 */
508 	nsw_cluster_max = min((MAXPHYS/PAGE_SIZE), MAX_PAGEOUT_CLUSTER);
509 
510 	mtx_lock(&pbuf_mtx);
511 	nsw_rcount = (nswbuf + 1) / 2;
512 	nsw_wcount_sync = (nswbuf + 3) / 4;
513 	nsw_wcount_async = 4;
514 	nsw_wcount_async_max = nsw_wcount_async;
515 	mtx_unlock(&pbuf_mtx);
516 
517 	/*
518 	 * Initialize our zone, guessing on the number we need based
519 	 * on the number of pages in the system.
520 	 */
521 	n = vm_cnt.v_page_count / 2;
522 	if (maxswzone && n > maxswzone / sizeof(struct swblk))
523 		n = maxswzone / sizeof(struct swblk);
524 	swpctrie_zone = uma_zcreate("swpctrie", pctrie_node_size(), NULL, NULL,
525 	    pctrie_zone_init, NULL, UMA_ALIGN_PTR,
526 	    UMA_ZONE_NOFREE | UMA_ZONE_VM);
527 	if (swpctrie_zone == NULL)
528 		panic("failed to create swap pctrie zone.");
529 	swblk_zone = uma_zcreate("swblk", sizeof(struct swblk), NULL, NULL,
530 	    NULL, NULL, _Alignof(struct swblk) - 1,
531 	    UMA_ZONE_NOFREE | UMA_ZONE_VM);
532 	if (swblk_zone == NULL)
533 		panic("failed to create swap blk zone.");
534 	n2 = n;
535 	do {
536 		if (uma_zone_reserve_kva(swblk_zone, n))
537 			break;
538 		/*
539 		 * if the allocation failed, try a zone two thirds the
540 		 * size of the previous attempt.
541 		 */
542 		n -= ((n + 2) / 3);
543 	} while (n > 0);
544 	if (n2 != n)
545 		printf("Swap blk zone entries reduced from %lu to %lu.\n",
546 		    n2, n);
547 	swap_maxpages = n * SWAP_META_PAGES;
548 	swzone = n * sizeof(struct swblk);
549 	if (!uma_zone_reserve_kva(swpctrie_zone, n))
550 		printf("Cannot reserve swap pctrie zone, "
551 		    "reduce kern.maxswzone.\n");
552 }
553 
554 static vm_object_t
555 swap_pager_alloc_init(void *handle, struct ucred *cred, vm_ooffset_t size,
556     vm_ooffset_t offset)
557 {
558 	vm_object_t object;
559 
560 	if (cred != NULL) {
561 		if (!swap_reserve_by_cred(size, cred))
562 			return (NULL);
563 		crhold(cred);
564 	}
565 
566 	/*
567 	 * The un_pager.swp.swp_blks trie is initialized by
568 	 * vm_object_allocate() to ensure the correct order of
569 	 * visibility to other threads.
570 	 */
571 	object = vm_object_allocate(OBJT_SWAP, OFF_TO_IDX(offset +
572 	    PAGE_MASK + size));
573 
574 	object->handle = handle;
575 	if (cred != NULL) {
576 		object->cred = cred;
577 		object->charge = size;
578 	}
579 	return (object);
580 }
581 
582 /*
583  * SWAP_PAGER_ALLOC() -	allocate a new OBJT_SWAP VM object and instantiate
584  *			its metadata structures.
585  *
586  *	This routine is called from the mmap and fork code to create a new
587  *	OBJT_SWAP object.
588  *
589  *	This routine must ensure that no live duplicate is created for
590  *	the named object request, which is protected against by
591  *	holding the sw_alloc_sx lock in case handle != NULL.
592  */
593 static vm_object_t
594 swap_pager_alloc(void *handle, vm_ooffset_t size, vm_prot_t prot,
595     vm_ooffset_t offset, struct ucred *cred)
596 {
597 	vm_object_t object;
598 
599 	if (handle != NULL) {
600 		/*
601 		 * Reference existing named region or allocate new one.  There
602 		 * should not be a race here against swp_pager_meta_build()
603 		 * as called from vm_page_remove() in regards to the lookup
604 		 * of the handle.
605 		 */
606 		sx_xlock(&sw_alloc_sx);
607 		object = vm_pager_object_lookup(NOBJLIST(handle), handle);
608 		if (object == NULL) {
609 			object = swap_pager_alloc_init(handle, cred, size,
610 			    offset);
611 			if (object != NULL) {
612 				TAILQ_INSERT_TAIL(NOBJLIST(object->handle),
613 				    object, pager_object_list);
614 			}
615 		}
616 		sx_xunlock(&sw_alloc_sx);
617 	} else {
618 		object = swap_pager_alloc_init(handle, cred, size, offset);
619 	}
620 	return (object);
621 }
622 
623 /*
624  * SWAP_PAGER_DEALLOC() -	remove swap metadata from object
625  *
626  *	The swap backing for the object is destroyed.  The code is
627  *	designed such that we can reinstantiate it later, but this
628  *	routine is typically called only when the entire object is
629  *	about to be destroyed.
630  *
631  *	The object must be locked.
632  */
633 static void
634 swap_pager_dealloc(vm_object_t object)
635 {
636 
637 	VM_OBJECT_ASSERT_WLOCKED(object);
638 	KASSERT((object->flags & OBJ_DEAD) != 0, ("dealloc of reachable obj"));
639 
640 	/*
641 	 * Remove from list right away so lookups will fail if we block for
642 	 * pageout completion.
643 	 */
644 	if (object->handle != NULL) {
645 		VM_OBJECT_WUNLOCK(object);
646 		sx_xlock(&sw_alloc_sx);
647 		TAILQ_REMOVE(NOBJLIST(object->handle), object,
648 		    pager_object_list);
649 		sx_xunlock(&sw_alloc_sx);
650 		VM_OBJECT_WLOCK(object);
651 	}
652 
653 	vm_object_pip_wait(object, "swpdea");
654 
655 	/*
656 	 * Free all remaining metadata.  We only bother to free it from
657 	 * the swap meta data.  We do not attempt to free swapblk's still
658 	 * associated with vm_page_t's for this object.  We do not care
659 	 * if paging is still in progress on some objects.
660 	 */
661 	swp_pager_meta_free_all(object);
662 	object->handle = NULL;
663 	object->type = OBJT_DEAD;
664 }
665 
666 /************************************************************************
667  *			SWAP PAGER BITMAP ROUTINES			*
668  ************************************************************************/
669 
670 /*
671  * SWP_PAGER_GETSWAPSPACE() -	allocate raw swap space
672  *
673  *	Allocate swap for the requested number of pages.  The starting
674  *	swap block number (a page index) is returned or SWAPBLK_NONE
675  *	if the allocation failed.
676  *
677  *	Also has the side effect of advising that somebody made a mistake
678  *	when they configured swap and didn't configure enough.
679  *
680  *	This routine may not sleep.
681  *
682  *	We allocate in round-robin fashion from the configured devices.
683  */
684 static daddr_t
685 swp_pager_getswapspace(int npages)
686 {
687 	daddr_t blk;
688 	struct swdevt *sp;
689 	int i;
690 
691 	blk = SWAPBLK_NONE;
692 	mtx_lock(&sw_dev_mtx);
693 	sp = swdevhd;
694 	for (i = 0; i < nswapdev; i++) {
695 		if (sp == NULL)
696 			sp = TAILQ_FIRST(&swtailq);
697 		if (!(sp->sw_flags & SW_CLOSING)) {
698 			blk = blist_alloc(sp->sw_blist, npages);
699 			if (blk != SWAPBLK_NONE) {
700 				blk += sp->sw_first;
701 				sp->sw_used += npages;
702 				swap_pager_avail -= npages;
703 				swp_sizecheck();
704 				swdevhd = TAILQ_NEXT(sp, sw_list);
705 				goto done;
706 			}
707 		}
708 		sp = TAILQ_NEXT(sp, sw_list);
709 	}
710 	if (swap_pager_full != 2) {
711 		printf("swap_pager_getswapspace(%d): failed\n", npages);
712 		swap_pager_full = 2;
713 		swap_pager_almost_full = 1;
714 	}
715 	swdevhd = NULL;
716 done:
717 	mtx_unlock(&sw_dev_mtx);
718 	return (blk);
719 }
720 
721 static int
722 swp_pager_isondev(daddr_t blk, struct swdevt *sp)
723 {
724 
725 	return (blk >= sp->sw_first && blk < sp->sw_end);
726 }
727 
728 static void
729 swp_pager_strategy(struct buf *bp)
730 {
731 	struct swdevt *sp;
732 
733 	mtx_lock(&sw_dev_mtx);
734 	TAILQ_FOREACH(sp, &swtailq, sw_list) {
735 		if (bp->b_blkno >= sp->sw_first && bp->b_blkno < sp->sw_end) {
736 			mtx_unlock(&sw_dev_mtx);
737 			if ((sp->sw_flags & SW_UNMAPPED) != 0 &&
738 			    unmapped_buf_allowed) {
739 				bp->b_data = unmapped_buf;
740 				bp->b_offset = 0;
741 			} else {
742 				pmap_qenter((vm_offset_t)bp->b_data,
743 				    &bp->b_pages[0], bp->b_bcount / PAGE_SIZE);
744 			}
745 			sp->sw_strategy(bp, sp);
746 			return;
747 		}
748 	}
749 	panic("Swapdev not found");
750 }
751 
752 
753 /*
754  * SWP_PAGER_FREESWAPSPACE() -	free raw swap space
755  *
756  *	This routine returns the specified swap blocks back to the bitmap.
757  *
758  *	This routine may not sleep.
759  */
760 static void
761 swp_pager_freeswapspace(daddr_t blk, int npages)
762 {
763 	struct swdevt *sp;
764 
765 	mtx_lock(&sw_dev_mtx);
766 	TAILQ_FOREACH(sp, &swtailq, sw_list) {
767 		if (blk >= sp->sw_first && blk < sp->sw_end) {
768 			sp->sw_used -= npages;
769 			/*
770 			 * If we are attempting to stop swapping on
771 			 * this device, we don't want to mark any
772 			 * blocks free lest they be reused.
773 			 */
774 			if ((sp->sw_flags & SW_CLOSING) == 0) {
775 				blist_free(sp->sw_blist, blk - sp->sw_first,
776 				    npages);
777 				swap_pager_avail += npages;
778 				swp_sizecheck();
779 			}
780 			mtx_unlock(&sw_dev_mtx);
781 			return;
782 		}
783 	}
784 	panic("Swapdev not found");
785 }
786 
787 /*
788  * SYSCTL_SWAP_FRAGMENTATION() -	produce raw swap space stats
789  */
790 static int
791 sysctl_swap_fragmentation(SYSCTL_HANDLER_ARGS)
792 {
793 	struct sbuf sbuf;
794 	struct swdevt *sp;
795 	const char *devname;
796 	int error;
797 
798 	error = sysctl_wire_old_buffer(req, 0);
799 	if (error != 0)
800 		return (error);
801 	sbuf_new_for_sysctl(&sbuf, NULL, 128, req);
802 	mtx_lock(&sw_dev_mtx);
803 	TAILQ_FOREACH(sp, &swtailq, sw_list) {
804 		if (vn_isdisk(sp->sw_vp, NULL))
805 			devname = devtoname(sp->sw_vp->v_rdev);
806 		else
807 			devname = "[file]";
808 		sbuf_printf(&sbuf, "\nFree space on device %s:\n", devname);
809 		blist_stats(sp->sw_blist, &sbuf);
810 	}
811 	mtx_unlock(&sw_dev_mtx);
812 	error = sbuf_finish(&sbuf);
813 	sbuf_delete(&sbuf);
814 	return (error);
815 }
816 
817 /*
818  * SWAP_PAGER_FREESPACE() -	frees swap blocks associated with a page
819  *				range within an object.
820  *
821  *	This is a globally accessible routine.
822  *
823  *	This routine removes swapblk assignments from swap metadata.
824  *
825  *	The external callers of this routine typically have already destroyed
826  *	or renamed vm_page_t's associated with this range in the object so
827  *	we should be ok.
828  *
829  *	The object must be locked.
830  */
831 void
832 swap_pager_freespace(vm_object_t object, vm_pindex_t start, vm_size_t size)
833 {
834 
835 	swp_pager_meta_free(object, start, size);
836 }
837 
838 /*
839  * SWAP_PAGER_RESERVE() - reserve swap blocks in object
840  *
841  *	Assigns swap blocks to the specified range within the object.  The
842  *	swap blocks are not zeroed.  Any previous swap assignment is destroyed.
843  *
844  *	Returns 0 on success, -1 on failure.
845  */
846 int
847 swap_pager_reserve(vm_object_t object, vm_pindex_t start, vm_size_t size)
848 {
849 	int n = 0;
850 	daddr_t blk = SWAPBLK_NONE;
851 	vm_pindex_t beg = start;	/* save start index */
852 
853 	VM_OBJECT_WLOCK(object);
854 	while (size) {
855 		if (n == 0) {
856 			n = BLIST_MAX_ALLOC;
857 			while ((blk = swp_pager_getswapspace(n)) == SWAPBLK_NONE) {
858 				n >>= 1;
859 				if (n == 0) {
860 					swp_pager_meta_free(object, beg, start - beg);
861 					VM_OBJECT_WUNLOCK(object);
862 					return (-1);
863 				}
864 			}
865 		}
866 		swp_pager_meta_build(object, start, blk);
867 		--size;
868 		++start;
869 		++blk;
870 		--n;
871 	}
872 	swp_pager_meta_free(object, start, n);
873 	VM_OBJECT_WUNLOCK(object);
874 	return (0);
875 }
876 
877 /*
878  * SWAP_PAGER_COPY() -  copy blocks from source pager to destination pager
879  *			and destroy the source.
880  *
881  *	Copy any valid swapblks from the source to the destination.  In
882  *	cases where both the source and destination have a valid swapblk,
883  *	we keep the destination's.
884  *
885  *	This routine is allowed to sleep.  It may sleep allocating metadata
886  *	indirectly through swp_pager_meta_build() or if paging is still in
887  *	progress on the source.
888  *
889  *	The source object contains no vm_page_t's (which is just as well)
890  *
891  *	The source object is of type OBJT_SWAP.
892  *
893  *	The source and destination objects must be locked.
894  *	Both object locks may temporarily be released.
895  */
896 void
897 swap_pager_copy(vm_object_t srcobject, vm_object_t dstobject,
898     vm_pindex_t offset, int destroysource)
899 {
900 	vm_pindex_t i;
901 
902 	VM_OBJECT_ASSERT_WLOCKED(srcobject);
903 	VM_OBJECT_ASSERT_WLOCKED(dstobject);
904 
905 	/*
906 	 * If destroysource is set, we remove the source object from the
907 	 * swap_pager internal queue now.
908 	 */
909 	if (destroysource && srcobject->handle != NULL) {
910 		vm_object_pip_add(srcobject, 1);
911 		VM_OBJECT_WUNLOCK(srcobject);
912 		vm_object_pip_add(dstobject, 1);
913 		VM_OBJECT_WUNLOCK(dstobject);
914 		sx_xlock(&sw_alloc_sx);
915 		TAILQ_REMOVE(NOBJLIST(srcobject->handle), srcobject,
916 		    pager_object_list);
917 		sx_xunlock(&sw_alloc_sx);
918 		VM_OBJECT_WLOCK(dstobject);
919 		vm_object_pip_wakeup(dstobject);
920 		VM_OBJECT_WLOCK(srcobject);
921 		vm_object_pip_wakeup(srcobject);
922 	}
923 
924 	/*
925 	 * transfer source to destination.
926 	 */
927 	for (i = 0; i < dstobject->size; ++i) {
928 		daddr_t dstaddr;
929 
930 		/*
931 		 * Locate (without changing) the swapblk on the destination,
932 		 * unless it is invalid in which case free it silently, or
933 		 * if the destination is a resident page, in which case the
934 		 * source is thrown away.
935 		 */
936 		dstaddr = swp_pager_meta_ctl(dstobject, i, 0);
937 
938 		if (dstaddr == SWAPBLK_NONE) {
939 			/*
940 			 * Destination has no swapblk and is not resident,
941 			 * copy source.
942 			 */
943 			daddr_t srcaddr;
944 
945 			srcaddr = swp_pager_meta_ctl(
946 			    srcobject,
947 			    i + offset,
948 			    SWM_POP
949 			);
950 
951 			if (srcaddr != SWAPBLK_NONE) {
952 				/*
953 				 * swp_pager_meta_build() can sleep.
954 				 */
955 				vm_object_pip_add(srcobject, 1);
956 				VM_OBJECT_WUNLOCK(srcobject);
957 				vm_object_pip_add(dstobject, 1);
958 				swp_pager_meta_build(dstobject, i, srcaddr);
959 				vm_object_pip_wakeup(dstobject);
960 				VM_OBJECT_WLOCK(srcobject);
961 				vm_object_pip_wakeup(srcobject);
962 			}
963 		} else {
964 			/*
965 			 * Destination has valid swapblk or it is represented
966 			 * by a resident page.  We destroy the sourceblock.
967 			 */
968 
969 			swp_pager_meta_ctl(srcobject, i + offset, SWM_FREE);
970 		}
971 	}
972 
973 	/*
974 	 * Free left over swap blocks in source.
975 	 *
976 	 * We have to revert the type to OBJT_DEFAULT so we do not accidentally
977 	 * double-remove the object from the swap queues.
978 	 */
979 	if (destroysource) {
980 		swp_pager_meta_free_all(srcobject);
981 		/*
982 		 * Reverting the type is not necessary, the caller is going
983 		 * to destroy srcobject directly, but I'm doing it here
984 		 * for consistency since we've removed the object from its
985 		 * queues.
986 		 */
987 		srcobject->type = OBJT_DEFAULT;
988 	}
989 }
990 
991 /*
992  * SWAP_PAGER_HASPAGE() -	determine if we have good backing store for
993  *				the requested page.
994  *
995  *	We determine whether good backing store exists for the requested
996  *	page and return TRUE if it does, FALSE if it doesn't.
997  *
998  *	If TRUE, we also try to determine how much valid, contiguous backing
999  *	store exists before and after the requested page.
1000  */
1001 static boolean_t
1002 swap_pager_haspage(vm_object_t object, vm_pindex_t pindex, int *before,
1003     int *after)
1004 {
1005 	daddr_t blk, blk0;
1006 	int i;
1007 
1008 	VM_OBJECT_ASSERT_LOCKED(object);
1009 
1010 	/*
1011 	 * do we have good backing store at the requested index ?
1012 	 */
1013 	blk0 = swp_pager_meta_ctl(object, pindex, 0);
1014 	if (blk0 == SWAPBLK_NONE) {
1015 		if (before)
1016 			*before = 0;
1017 		if (after)
1018 			*after = 0;
1019 		return (FALSE);
1020 	}
1021 
1022 	/*
1023 	 * find backwards-looking contiguous good backing store
1024 	 */
1025 	if (before != NULL) {
1026 		for (i = 1; i < SWB_NPAGES; i++) {
1027 			if (i > pindex)
1028 				break;
1029 			blk = swp_pager_meta_ctl(object, pindex - i, 0);
1030 			if (blk != blk0 - i)
1031 				break;
1032 		}
1033 		*before = i - 1;
1034 	}
1035 
1036 	/*
1037 	 * find forward-looking contiguous good backing store
1038 	 */
1039 	if (after != NULL) {
1040 		for (i = 1; i < SWB_NPAGES; i++) {
1041 			blk = swp_pager_meta_ctl(object, pindex + i, 0);
1042 			if (blk != blk0 + i)
1043 				break;
1044 		}
1045 		*after = i - 1;
1046 	}
1047 	return (TRUE);
1048 }
1049 
1050 /*
1051  * SWAP_PAGER_PAGE_UNSWAPPED() - remove swap backing store related to page
1052  *
1053  *	This removes any associated swap backing store, whether valid or
1054  *	not, from the page.
1055  *
1056  *	This routine is typically called when a page is made dirty, at
1057  *	which point any associated swap can be freed.  MADV_FREE also
1058  *	calls us in a special-case situation
1059  *
1060  *	NOTE!!!  If the page is clean and the swap was valid, the caller
1061  *	should make the page dirty before calling this routine.  This routine
1062  *	does NOT change the m->dirty status of the page.  Also: MADV_FREE
1063  *	depends on it.
1064  *
1065  *	This routine may not sleep.
1066  *
1067  *	The object containing the page must be locked.
1068  */
1069 static void
1070 swap_pager_unswapped(vm_page_t m)
1071 {
1072 
1073 	swp_pager_meta_ctl(m->object, m->pindex, SWM_FREE);
1074 }
1075 
1076 /*
1077  * swap_pager_getpages() - bring pages in from swap
1078  *
1079  *	Attempt to page in the pages in array "m" of length "count".  The caller
1080  *	may optionally specify that additional pages preceding and succeeding
1081  *	the specified range be paged in.  The number of such pages is returned
1082  *	in the "rbehind" and "rahead" parameters, and they will be in the
1083  *	inactive queue upon return.
1084  *
1085  *	The pages in "m" must be busied and will remain busied upon return.
1086  */
1087 static int
1088 swap_pager_getpages(vm_object_t object, vm_page_t *m, int count, int *rbehind,
1089     int *rahead)
1090 {
1091 	struct buf *bp;
1092 	vm_page_t mpred, msucc, p;
1093 	vm_pindex_t pindex;
1094 	daddr_t blk;
1095 	int i, j, maxahead, maxbehind, reqcount, shift;
1096 
1097 	reqcount = count;
1098 
1099 	VM_OBJECT_WUNLOCK(object);
1100 	bp = getpbuf(&nsw_rcount);
1101 	VM_OBJECT_WLOCK(object);
1102 
1103 	if (!swap_pager_haspage(object, m[0]->pindex, &maxbehind, &maxahead)) {
1104 		relpbuf(bp, &nsw_rcount);
1105 		return (VM_PAGER_FAIL);
1106 	}
1107 
1108 	/*
1109 	 * Clip the readahead and readbehind ranges to exclude resident pages.
1110 	 */
1111 	if (rahead != NULL) {
1112 		KASSERT(reqcount - 1 <= maxahead,
1113 		    ("page count %d extends beyond swap block", reqcount));
1114 		*rahead = imin(*rahead, maxahead - (reqcount - 1));
1115 		pindex = m[reqcount - 1]->pindex;
1116 		msucc = TAILQ_NEXT(m[reqcount - 1], listq);
1117 		if (msucc != NULL && msucc->pindex - pindex - 1 < *rahead)
1118 			*rahead = msucc->pindex - pindex - 1;
1119 	}
1120 	if (rbehind != NULL) {
1121 		*rbehind = imin(*rbehind, maxbehind);
1122 		pindex = m[0]->pindex;
1123 		mpred = TAILQ_PREV(m[0], pglist, listq);
1124 		if (mpred != NULL && pindex - mpred->pindex - 1 < *rbehind)
1125 			*rbehind = pindex - mpred->pindex - 1;
1126 	}
1127 
1128 	/*
1129 	 * Allocate readahead and readbehind pages.
1130 	 */
1131 	shift = rbehind != NULL ? *rbehind : 0;
1132 	if (shift != 0) {
1133 		for (i = 1; i <= shift; i++) {
1134 			p = vm_page_alloc(object, m[0]->pindex - i,
1135 			    VM_ALLOC_NORMAL);
1136 			if (p == NULL) {
1137 				/* Shift allocated pages to the left. */
1138 				for (j = 0; j < i - 1; j++)
1139 					bp->b_pages[j] =
1140 					    bp->b_pages[j + shift - i + 1];
1141 				break;
1142 			}
1143 			bp->b_pages[shift - i] = p;
1144 		}
1145 		shift = i - 1;
1146 		*rbehind = shift;
1147 	}
1148 	for (i = 0; i < reqcount; i++)
1149 		bp->b_pages[i + shift] = m[i];
1150 	if (rahead != NULL) {
1151 		for (i = 0; i < *rahead; i++) {
1152 			p = vm_page_alloc(object,
1153 			    m[reqcount - 1]->pindex + i + 1, VM_ALLOC_NORMAL);
1154 			if (p == NULL)
1155 				break;
1156 			bp->b_pages[shift + reqcount + i] = p;
1157 		}
1158 		*rahead = i;
1159 	}
1160 	if (rbehind != NULL)
1161 		count += *rbehind;
1162 	if (rahead != NULL)
1163 		count += *rahead;
1164 
1165 	vm_object_pip_add(object, count);
1166 
1167 	for (i = 0; i < count; i++)
1168 		bp->b_pages[i]->oflags |= VPO_SWAPINPROG;
1169 
1170 	pindex = bp->b_pages[0]->pindex;
1171 	blk = swp_pager_meta_ctl(object, pindex, 0);
1172 	KASSERT(blk != SWAPBLK_NONE,
1173 	    ("no swap blocking containing %p(%jx)", object, (uintmax_t)pindex));
1174 
1175 	VM_OBJECT_WUNLOCK(object);
1176 
1177 	bp->b_flags |= B_PAGING;
1178 	bp->b_iocmd = BIO_READ;
1179 	bp->b_iodone = swp_pager_async_iodone;
1180 	bp->b_rcred = crhold(thread0.td_ucred);
1181 	bp->b_wcred = crhold(thread0.td_ucred);
1182 	bp->b_blkno = blk;
1183 	bp->b_bcount = PAGE_SIZE * count;
1184 	bp->b_bufsize = PAGE_SIZE * count;
1185 	bp->b_npages = count;
1186 	bp->b_pgbefore = rbehind != NULL ? *rbehind : 0;
1187 	bp->b_pgafter = rahead != NULL ? *rahead : 0;
1188 
1189 	VM_CNT_INC(v_swapin);
1190 	VM_CNT_ADD(v_swappgsin, count);
1191 
1192 	/*
1193 	 * perform the I/O.  NOTE!!!  bp cannot be considered valid after
1194 	 * this point because we automatically release it on completion.
1195 	 * Instead, we look at the one page we are interested in which we
1196 	 * still hold a lock on even through the I/O completion.
1197 	 *
1198 	 * The other pages in our m[] array are also released on completion,
1199 	 * so we cannot assume they are valid anymore either.
1200 	 *
1201 	 * NOTE: b_blkno is destroyed by the call to swapdev_strategy
1202 	 */
1203 	BUF_KERNPROC(bp);
1204 	swp_pager_strategy(bp);
1205 
1206 	/*
1207 	 * Wait for the pages we want to complete.  VPO_SWAPINPROG is always
1208 	 * cleared on completion.  If an I/O error occurs, SWAPBLK_NONE
1209 	 * is set in the metadata for each page in the request.
1210 	 */
1211 	VM_OBJECT_WLOCK(object);
1212 	while ((m[0]->oflags & VPO_SWAPINPROG) != 0) {
1213 		m[0]->oflags |= VPO_SWAPSLEEP;
1214 		VM_CNT_INC(v_intrans);
1215 		if (VM_OBJECT_SLEEP(object, &object->paging_in_progress, PSWP,
1216 		    "swread", hz * 20)) {
1217 			printf(
1218 "swap_pager: indefinite wait buffer: bufobj: %p, blkno: %jd, size: %ld\n",
1219 			    bp->b_bufobj, (intmax_t)bp->b_blkno, bp->b_bcount);
1220 		}
1221 	}
1222 
1223 	/*
1224 	 * If we had an unrecoverable read error pages will not be valid.
1225 	 */
1226 	for (i = 0; i < reqcount; i++)
1227 		if (m[i]->valid != VM_PAGE_BITS_ALL)
1228 			return (VM_PAGER_ERROR);
1229 
1230 	return (VM_PAGER_OK);
1231 
1232 	/*
1233 	 * A final note: in a low swap situation, we cannot deallocate swap
1234 	 * and mark a page dirty here because the caller is likely to mark
1235 	 * the page clean when we return, causing the page to possibly revert
1236 	 * to all-zero's later.
1237 	 */
1238 }
1239 
1240 /*
1241  * 	swap_pager_getpages_async():
1242  *
1243  *	Right now this is emulation of asynchronous operation on top of
1244  *	swap_pager_getpages().
1245  */
1246 static int
1247 swap_pager_getpages_async(vm_object_t object, vm_page_t *m, int count,
1248     int *rbehind, int *rahead, pgo_getpages_iodone_t iodone, void *arg)
1249 {
1250 	int r, error;
1251 
1252 	r = swap_pager_getpages(object, m, count, rbehind, rahead);
1253 	VM_OBJECT_WUNLOCK(object);
1254 	switch (r) {
1255 	case VM_PAGER_OK:
1256 		error = 0;
1257 		break;
1258 	case VM_PAGER_ERROR:
1259 		error = EIO;
1260 		break;
1261 	case VM_PAGER_FAIL:
1262 		error = EINVAL;
1263 		break;
1264 	default:
1265 		panic("unhandled swap_pager_getpages() error %d", r);
1266 	}
1267 	(iodone)(arg, m, count, error);
1268 	VM_OBJECT_WLOCK(object);
1269 
1270 	return (r);
1271 }
1272 
1273 /*
1274  *	swap_pager_putpages:
1275  *
1276  *	Assign swap (if necessary) and initiate I/O on the specified pages.
1277  *
1278  *	We support both OBJT_DEFAULT and OBJT_SWAP objects.  DEFAULT objects
1279  *	are automatically converted to SWAP objects.
1280  *
1281  *	In a low memory situation we may block in VOP_STRATEGY(), but the new
1282  *	vm_page reservation system coupled with properly written VFS devices
1283  *	should ensure that no low-memory deadlock occurs.  This is an area
1284  *	which needs work.
1285  *
1286  *	The parent has N vm_object_pip_add() references prior to
1287  *	calling us and will remove references for rtvals[] that are
1288  *	not set to VM_PAGER_PEND.  We need to remove the rest on I/O
1289  *	completion.
1290  *
1291  *	The parent has soft-busy'd the pages it passes us and will unbusy
1292  *	those whos rtvals[] entry is not set to VM_PAGER_PEND on return.
1293  *	We need to unbusy the rest on I/O completion.
1294  */
1295 static void
1296 swap_pager_putpages(vm_object_t object, vm_page_t *m, int count,
1297     int flags, int *rtvals)
1298 {
1299 	int i, n;
1300 	boolean_t sync;
1301 
1302 	if (count && m[0]->object != object) {
1303 		panic("swap_pager_putpages: object mismatch %p/%p",
1304 		    object,
1305 		    m[0]->object
1306 		);
1307 	}
1308 
1309 	/*
1310 	 * Step 1
1311 	 *
1312 	 * Turn object into OBJT_SWAP
1313 	 * check for bogus sysops
1314 	 * force sync if not pageout process
1315 	 */
1316 	if (object->type != OBJT_SWAP)
1317 		swp_pager_meta_build(object, 0, SWAPBLK_NONE);
1318 	VM_OBJECT_WUNLOCK(object);
1319 
1320 	n = 0;
1321 	if (curproc != pageproc)
1322 		sync = TRUE;
1323 	else
1324 		sync = (flags & VM_PAGER_PUT_SYNC) != 0;
1325 
1326 	/*
1327 	 * Step 2
1328 	 *
1329 	 * Assign swap blocks and issue I/O.  We reallocate swap on the fly.
1330 	 * The page is left dirty until the pageout operation completes
1331 	 * successfully.
1332 	 */
1333 	for (i = 0; i < count; i += n) {
1334 		int j;
1335 		struct buf *bp;
1336 		daddr_t blk;
1337 
1338 		/*
1339 		 * Maximum I/O size is limited by a number of factors.
1340 		 */
1341 		n = min(BLIST_MAX_ALLOC, count - i);
1342 		n = min(n, nsw_cluster_max);
1343 
1344 		/*
1345 		 * Get biggest block of swap we can.  If we fail, fall
1346 		 * back and try to allocate a smaller block.  Don't go
1347 		 * overboard trying to allocate space if it would overly
1348 		 * fragment swap.
1349 		 */
1350 		while (
1351 		    (blk = swp_pager_getswapspace(n)) == SWAPBLK_NONE &&
1352 		    n > 4
1353 		) {
1354 			n >>= 1;
1355 		}
1356 		if (blk == SWAPBLK_NONE) {
1357 			for (j = 0; j < n; ++j)
1358 				rtvals[i+j] = VM_PAGER_FAIL;
1359 			continue;
1360 		}
1361 
1362 		/*
1363 		 * All I/O parameters have been satisfied, build the I/O
1364 		 * request and assign the swap space.
1365 		 */
1366 		if (sync == TRUE) {
1367 			bp = getpbuf(&nsw_wcount_sync);
1368 		} else {
1369 			bp = getpbuf(&nsw_wcount_async);
1370 			bp->b_flags = B_ASYNC;
1371 		}
1372 		bp->b_flags |= B_PAGING;
1373 		bp->b_iocmd = BIO_WRITE;
1374 
1375 		bp->b_rcred = crhold(thread0.td_ucred);
1376 		bp->b_wcred = crhold(thread0.td_ucred);
1377 		bp->b_bcount = PAGE_SIZE * n;
1378 		bp->b_bufsize = PAGE_SIZE * n;
1379 		bp->b_blkno = blk;
1380 
1381 		VM_OBJECT_WLOCK(object);
1382 		for (j = 0; j < n; ++j) {
1383 			vm_page_t mreq = m[i+j];
1384 
1385 			swp_pager_meta_build(
1386 			    mreq->object,
1387 			    mreq->pindex,
1388 			    blk + j
1389 			);
1390 			MPASS(mreq->dirty == VM_PAGE_BITS_ALL);
1391 			mreq->oflags |= VPO_SWAPINPROG;
1392 			bp->b_pages[j] = mreq;
1393 		}
1394 		VM_OBJECT_WUNLOCK(object);
1395 		bp->b_npages = n;
1396 		/*
1397 		 * Must set dirty range for NFS to work.
1398 		 */
1399 		bp->b_dirtyoff = 0;
1400 		bp->b_dirtyend = bp->b_bcount;
1401 
1402 		VM_CNT_INC(v_swapout);
1403 		VM_CNT_ADD(v_swappgsout, bp->b_npages);
1404 
1405 		/*
1406 		 * We unconditionally set rtvals[] to VM_PAGER_PEND so that we
1407 		 * can call the async completion routine at the end of a
1408 		 * synchronous I/O operation.  Otherwise, our caller would
1409 		 * perform duplicate unbusy and wakeup operations on the page
1410 		 * and object, respectively.
1411 		 */
1412 		for (j = 0; j < n; j++)
1413 			rtvals[i + j] = VM_PAGER_PEND;
1414 
1415 		/*
1416 		 * asynchronous
1417 		 *
1418 		 * NOTE: b_blkno is destroyed by the call to swapdev_strategy
1419 		 */
1420 		if (sync == FALSE) {
1421 			bp->b_iodone = swp_pager_async_iodone;
1422 			BUF_KERNPROC(bp);
1423 			swp_pager_strategy(bp);
1424 			continue;
1425 		}
1426 
1427 		/*
1428 		 * synchronous
1429 		 *
1430 		 * NOTE: b_blkno is destroyed by the call to swapdev_strategy
1431 		 */
1432 		bp->b_iodone = bdone;
1433 		swp_pager_strategy(bp);
1434 
1435 		/*
1436 		 * Wait for the sync I/O to complete.
1437 		 */
1438 		bwait(bp, PVM, "swwrt");
1439 
1440 		/*
1441 		 * Now that we are through with the bp, we can call the
1442 		 * normal async completion, which frees everything up.
1443 		 */
1444 		swp_pager_async_iodone(bp);
1445 	}
1446 	VM_OBJECT_WLOCK(object);
1447 }
1448 
1449 /*
1450  *	swp_pager_async_iodone:
1451  *
1452  *	Completion routine for asynchronous reads and writes from/to swap.
1453  *	Also called manually by synchronous code to finish up a bp.
1454  *
1455  *	This routine may not sleep.
1456  */
1457 static void
1458 swp_pager_async_iodone(struct buf *bp)
1459 {
1460 	int i;
1461 	vm_object_t object = NULL;
1462 
1463 	/*
1464 	 * report error
1465 	 */
1466 	if (bp->b_ioflags & BIO_ERROR) {
1467 		printf(
1468 		    "swap_pager: I/O error - %s failed; blkno %ld,"
1469 			"size %ld, error %d\n",
1470 		    ((bp->b_iocmd == BIO_READ) ? "pagein" : "pageout"),
1471 		    (long)bp->b_blkno,
1472 		    (long)bp->b_bcount,
1473 		    bp->b_error
1474 		);
1475 	}
1476 
1477 	/*
1478 	 * remove the mapping for kernel virtual
1479 	 */
1480 	if (buf_mapped(bp))
1481 		pmap_qremove((vm_offset_t)bp->b_data, bp->b_npages);
1482 	else
1483 		bp->b_data = bp->b_kvabase;
1484 
1485 	if (bp->b_npages) {
1486 		object = bp->b_pages[0]->object;
1487 		VM_OBJECT_WLOCK(object);
1488 	}
1489 
1490 	/*
1491 	 * cleanup pages.  If an error occurs writing to swap, we are in
1492 	 * very serious trouble.  If it happens to be a disk error, though,
1493 	 * we may be able to recover by reassigning the swap later on.  So
1494 	 * in this case we remove the m->swapblk assignment for the page
1495 	 * but do not free it in the rlist.  The errornous block(s) are thus
1496 	 * never reallocated as swap.  Redirty the page and continue.
1497 	 */
1498 	for (i = 0; i < bp->b_npages; ++i) {
1499 		vm_page_t m = bp->b_pages[i];
1500 
1501 		m->oflags &= ~VPO_SWAPINPROG;
1502 		if (m->oflags & VPO_SWAPSLEEP) {
1503 			m->oflags &= ~VPO_SWAPSLEEP;
1504 			wakeup(&object->paging_in_progress);
1505 		}
1506 
1507 		if (bp->b_ioflags & BIO_ERROR) {
1508 			/*
1509 			 * If an error occurs I'd love to throw the swapblk
1510 			 * away without freeing it back to swapspace, so it
1511 			 * can never be used again.  But I can't from an
1512 			 * interrupt.
1513 			 */
1514 			if (bp->b_iocmd == BIO_READ) {
1515 				/*
1516 				 * NOTE: for reads, m->dirty will probably
1517 				 * be overridden by the original caller of
1518 				 * getpages so don't play cute tricks here.
1519 				 */
1520 				m->valid = 0;
1521 			} else {
1522 				/*
1523 				 * If a write error occurs, reactivate page
1524 				 * so it doesn't clog the inactive list,
1525 				 * then finish the I/O.
1526 				 */
1527 				MPASS(m->dirty == VM_PAGE_BITS_ALL);
1528 				vm_page_lock(m);
1529 				vm_page_activate(m);
1530 				vm_page_unlock(m);
1531 				vm_page_sunbusy(m);
1532 			}
1533 		} else if (bp->b_iocmd == BIO_READ) {
1534 			/*
1535 			 * NOTE: for reads, m->dirty will probably be
1536 			 * overridden by the original caller of getpages so
1537 			 * we cannot set them in order to free the underlying
1538 			 * swap in a low-swap situation.  I don't think we'd
1539 			 * want to do that anyway, but it was an optimization
1540 			 * that existed in the old swapper for a time before
1541 			 * it got ripped out due to precisely this problem.
1542 			 */
1543 			KASSERT(!pmap_page_is_mapped(m),
1544 			    ("swp_pager_async_iodone: page %p is mapped", m));
1545 			KASSERT(m->dirty == 0,
1546 			    ("swp_pager_async_iodone: page %p is dirty", m));
1547 
1548 			m->valid = VM_PAGE_BITS_ALL;
1549 			if (i < bp->b_pgbefore ||
1550 			    i >= bp->b_npages - bp->b_pgafter)
1551 				vm_page_readahead_finish(m);
1552 		} else {
1553 			/*
1554 			 * For write success, clear the dirty
1555 			 * status, then finish the I/O ( which decrements the
1556 			 * busy count and possibly wakes waiter's up ).
1557 			 * A page is only written to swap after a period of
1558 			 * inactivity.  Therefore, we do not expect it to be
1559 			 * reused.
1560 			 */
1561 			KASSERT(!pmap_page_is_write_mapped(m),
1562 			    ("swp_pager_async_iodone: page %p is not write"
1563 			    " protected", m));
1564 			vm_page_undirty(m);
1565 			vm_page_lock(m);
1566 			vm_page_deactivate_noreuse(m);
1567 			vm_page_unlock(m);
1568 			vm_page_sunbusy(m);
1569 		}
1570 	}
1571 
1572 	/*
1573 	 * adjust pip.  NOTE: the original parent may still have its own
1574 	 * pip refs on the object.
1575 	 */
1576 	if (object != NULL) {
1577 		vm_object_pip_wakeupn(object, bp->b_npages);
1578 		VM_OBJECT_WUNLOCK(object);
1579 	}
1580 
1581 	/*
1582 	 * swapdev_strategy() manually sets b_vp and b_bufobj before calling
1583 	 * bstrategy(). Set them back to NULL now we're done with it, or we'll
1584 	 * trigger a KASSERT in relpbuf().
1585 	 */
1586 	if (bp->b_vp) {
1587 		    bp->b_vp = NULL;
1588 		    bp->b_bufobj = NULL;
1589 	}
1590 	/*
1591 	 * release the physical I/O buffer
1592 	 */
1593 	relpbuf(
1594 	    bp,
1595 	    ((bp->b_iocmd == BIO_READ) ? &nsw_rcount :
1596 		((bp->b_flags & B_ASYNC) ?
1597 		    &nsw_wcount_async :
1598 		    &nsw_wcount_sync
1599 		)
1600 	    )
1601 	);
1602 }
1603 
1604 int
1605 swap_pager_nswapdev(void)
1606 {
1607 
1608 	return (nswapdev);
1609 }
1610 
1611 /*
1612  * SWP_PAGER_FORCE_PAGEIN() - force a swap block to be paged in
1613  *
1614  *	This routine dissociates the page at the given index within an object
1615  *	from its backing store, paging it in if it does not reside in memory.
1616  *	If the page is paged in, it is marked dirty and placed in the laundry
1617  *	queue.  The page is marked dirty because it no longer has backing
1618  *	store.  It is placed in the laundry queue because it has not been
1619  *	accessed recently.  Otherwise, it would already reside in memory.
1620  *
1621  *	We also attempt to swap in all other pages in the swap block.
1622  *	However, we only guarantee that the one at the specified index is
1623  *	paged in.
1624  *
1625  *	XXX - The code to page the whole block in doesn't work, so we
1626  *	      revert to the one-by-one behavior for now.  Sigh.
1627  */
1628 static inline void
1629 swp_pager_force_pagein(vm_object_t object, vm_pindex_t pindex)
1630 {
1631 	vm_page_t m;
1632 
1633 	vm_object_pip_add(object, 1);
1634 	m = vm_page_grab(object, pindex, VM_ALLOC_NORMAL);
1635 	if (m->valid == VM_PAGE_BITS_ALL) {
1636 		vm_object_pip_wakeup(object);
1637 		vm_page_dirty(m);
1638 #ifdef INVARIANTS
1639 		vm_page_lock(m);
1640 		if (m->wire_count == 0 && m->queue == PQ_NONE)
1641 			panic("page %p is neither wired nor queued", m);
1642 		vm_page_unlock(m);
1643 #endif
1644 		vm_page_xunbusy(m);
1645 		vm_pager_page_unswapped(m);
1646 		return;
1647 	}
1648 
1649 	if (swap_pager_getpages(object, &m, 1, NULL, NULL) != VM_PAGER_OK)
1650 		panic("swap_pager_force_pagein: read from swap failed");/*XXX*/
1651 	vm_object_pip_wakeup(object);
1652 	vm_page_dirty(m);
1653 	vm_page_lock(m);
1654 	vm_page_launder(m);
1655 	vm_page_unlock(m);
1656 	vm_page_xunbusy(m);
1657 	vm_pager_page_unswapped(m);
1658 }
1659 
1660 /*
1661  *	swap_pager_swapoff:
1662  *
1663  *	Page in all of the pages that have been paged out to the
1664  *	given device.  The corresponding blocks in the bitmap must be
1665  *	marked as allocated and the device must be flagged SW_CLOSING.
1666  *	There may be no processes swapped out to the device.
1667  *
1668  *	This routine may block.
1669  */
1670 static void
1671 swap_pager_swapoff(struct swdevt *sp)
1672 {
1673 	struct swblk *sb;
1674 	vm_object_t object;
1675 	vm_pindex_t pi;
1676 	int i, retries;
1677 
1678 	sx_assert(&swdev_syscall_lock, SA_XLOCKED);
1679 
1680 	retries = 0;
1681 full_rescan:
1682 	mtx_lock(&vm_object_list_mtx);
1683 	TAILQ_FOREACH(object, &vm_object_list, object_list) {
1684 		if (object->type != OBJT_SWAP)
1685 			continue;
1686 		mtx_unlock(&vm_object_list_mtx);
1687 		/* Depends on type-stability. */
1688 		VM_OBJECT_WLOCK(object);
1689 
1690 		/*
1691 		 * Dead objects are eventually terminated on their own.
1692 		 */
1693 		if ((object->flags & OBJ_DEAD) != 0)
1694 			goto next_obj;
1695 
1696 		/*
1697 		 * Sync with fences placed after pctrie
1698 		 * initialization.  We must not access pctrie below
1699 		 * unless we checked that our object is swap and not
1700 		 * dead.
1701 		 */
1702 		atomic_thread_fence_acq();
1703 		if (object->type != OBJT_SWAP)
1704 			goto next_obj;
1705 
1706 		for (pi = 0; (sb = SWAP_PCTRIE_LOOKUP_GE(
1707 		    &object->un_pager.swp.swp_blks, pi)) != NULL; ) {
1708 			pi = sb->p + SWAP_META_PAGES;
1709 			for (i = 0; i < SWAP_META_PAGES; i++) {
1710 				if (sb->d[i] == SWAPBLK_NONE)
1711 					continue;
1712 				if (swp_pager_isondev(sb->d[i], sp))
1713 					swp_pager_force_pagein(object,
1714 					    sb->p + i);
1715 			}
1716 		}
1717 next_obj:
1718 		VM_OBJECT_WUNLOCK(object);
1719 		mtx_lock(&vm_object_list_mtx);
1720 	}
1721 	mtx_unlock(&vm_object_list_mtx);
1722 
1723 	if (sp->sw_used) {
1724 		/*
1725 		 * Objects may be locked or paging to the device being
1726 		 * removed, so we will miss their pages and need to
1727 		 * make another pass.  We have marked this device as
1728 		 * SW_CLOSING, so the activity should finish soon.
1729 		 */
1730 		retries++;
1731 		if (retries > 100) {
1732 			panic("swapoff: failed to locate %d swap blocks",
1733 			    sp->sw_used);
1734 		}
1735 		pause("swpoff", hz / 20);
1736 		goto full_rescan;
1737 	}
1738 	EVENTHANDLER_INVOKE(swapoff, sp);
1739 }
1740 
1741 /************************************************************************
1742  *				SWAP META DATA 				*
1743  ************************************************************************
1744  *
1745  *	These routines manipulate the swap metadata stored in the
1746  *	OBJT_SWAP object.
1747  *
1748  *	Swap metadata is implemented with a global hash and not directly
1749  *	linked into the object.  Instead the object simply contains
1750  *	appropriate tracking counters.
1751  */
1752 
1753 /*
1754  * SWP_PAGER_META_BUILD() -	add swap block to swap meta data for object
1755  *
1756  *	We first convert the object to a swap object if it is a default
1757  *	object.
1758  *
1759  *	The specified swapblk is added to the object's swap metadata.  If
1760  *	the swapblk is not valid, it is freed instead.  Any previously
1761  *	assigned swapblk is freed.
1762  */
1763 static void
1764 swp_pager_meta_build(vm_object_t object, vm_pindex_t pindex, daddr_t swapblk)
1765 {
1766 	static volatile int swblk_zone_exhausted, swpctrie_zone_exhausted;
1767 	struct swblk *sb, *sb1;
1768 	vm_pindex_t modpi, rdpi;
1769 	int error, i;
1770 
1771 	VM_OBJECT_ASSERT_WLOCKED(object);
1772 
1773 	/*
1774 	 * Convert default object to swap object if necessary
1775 	 */
1776 	if (object->type != OBJT_SWAP) {
1777 		pctrie_init(&object->un_pager.swp.swp_blks);
1778 
1779 		/*
1780 		 * Ensure that swap_pager_swapoff()'s iteration over
1781 		 * object_list does not see a garbage pctrie.
1782 		 */
1783 		atomic_thread_fence_rel();
1784 
1785 		object->type = OBJT_SWAP;
1786 		KASSERT(object->handle == NULL, ("default pager with handle"));
1787 	}
1788 
1789 	rdpi = rounddown(pindex, SWAP_META_PAGES);
1790 	sb = SWAP_PCTRIE_LOOKUP(&object->un_pager.swp.swp_blks, rdpi);
1791 	if (sb == NULL) {
1792 		if (swapblk == SWAPBLK_NONE)
1793 			return;
1794 		for (;;) {
1795 			sb = uma_zalloc(swblk_zone, M_NOWAIT | (curproc ==
1796 			    pageproc ? M_USE_RESERVE : 0));
1797 			if (sb != NULL) {
1798 				sb->p = rdpi;
1799 				for (i = 0; i < SWAP_META_PAGES; i++)
1800 					sb->d[i] = SWAPBLK_NONE;
1801 				if (atomic_cmpset_int(&swblk_zone_exhausted,
1802 				    1, 0))
1803 					printf("swblk zone ok\n");
1804 				break;
1805 			}
1806 			VM_OBJECT_WUNLOCK(object);
1807 			if (uma_zone_exhausted(swblk_zone)) {
1808 				if (atomic_cmpset_int(&swblk_zone_exhausted,
1809 				    0, 1))
1810 					printf("swap blk zone exhausted, "
1811 					    "increase kern.maxswzone\n");
1812 				vm_pageout_oom(VM_OOM_SWAPZ);
1813 				pause("swzonxb", 10);
1814 			} else
1815 				VM_WAIT;
1816 			VM_OBJECT_WLOCK(object);
1817 			sb = SWAP_PCTRIE_LOOKUP(&object->un_pager.swp.swp_blks,
1818 			    rdpi);
1819 			if (sb != NULL)
1820 				/*
1821 				 * Somebody swapped out a nearby page,
1822 				 * allocating swblk at the rdpi index,
1823 				 * while we dropped the object lock.
1824 				 */
1825 				goto allocated;
1826 		}
1827 		for (;;) {
1828 			error = SWAP_PCTRIE_INSERT(
1829 			    &object->un_pager.swp.swp_blks, sb);
1830 			if (error == 0) {
1831 				if (atomic_cmpset_int(&swpctrie_zone_exhausted,
1832 				    1, 0))
1833 					printf("swpctrie zone ok\n");
1834 				break;
1835 			}
1836 			VM_OBJECT_WUNLOCK(object);
1837 			if (uma_zone_exhausted(swpctrie_zone)) {
1838 				if (atomic_cmpset_int(&swpctrie_zone_exhausted,
1839 				    0, 1))
1840 					printf("swap pctrie zone exhausted, "
1841 					    "increase kern.maxswzone\n");
1842 				vm_pageout_oom(VM_OOM_SWAPZ);
1843 				pause("swzonxp", 10);
1844 			} else
1845 				VM_WAIT;
1846 			VM_OBJECT_WLOCK(object);
1847 			sb1 = SWAP_PCTRIE_LOOKUP(&object->un_pager.swp.swp_blks,
1848 			    rdpi);
1849 			if (sb1 != NULL) {
1850 				uma_zfree(swblk_zone, sb);
1851 				sb = sb1;
1852 				goto allocated;
1853 			}
1854 		}
1855 	}
1856 allocated:
1857 	MPASS(sb->p == rdpi);
1858 
1859 	modpi = pindex % SWAP_META_PAGES;
1860 	/* Delete prior contents of metadata. */
1861 	if (sb->d[modpi] != SWAPBLK_NONE)
1862 		swp_pager_freeswapspace(sb->d[modpi], 1);
1863 	/* Enter block into metadata. */
1864 	sb->d[modpi] = swapblk;
1865 
1866 	/*
1867 	 * Free the swblk if we end up with the empty page run.
1868 	 */
1869 	if (swapblk == SWAPBLK_NONE) {
1870 		for (i = 0; i < SWAP_META_PAGES; i++) {
1871 			if (sb->d[i] != SWAPBLK_NONE)
1872 				break;
1873 		}
1874 		if (i == SWAP_META_PAGES) {
1875 			SWAP_PCTRIE_REMOVE(&object->un_pager.swp.swp_blks,
1876 			    rdpi);
1877 			uma_zfree(swblk_zone, sb);
1878 		}
1879 	}
1880 }
1881 
1882 /*
1883  * SWP_PAGER_META_FREE() - free a range of blocks in the object's swap metadata
1884  *
1885  *	The requested range of blocks is freed, with any associated swap
1886  *	returned to the swap bitmap.
1887  *
1888  *	This routine will free swap metadata structures as they are cleaned
1889  *	out.  This routine does *NOT* operate on swap metadata associated
1890  *	with resident pages.
1891  */
1892 static void
1893 swp_pager_meta_free(vm_object_t object, vm_pindex_t pindex, vm_pindex_t count)
1894 {
1895 	struct swblk *sb;
1896 	vm_pindex_t last;
1897 	int i;
1898 	bool empty;
1899 
1900 	VM_OBJECT_ASSERT_WLOCKED(object);
1901 	if (object->type != OBJT_SWAP || count == 0)
1902 		return;
1903 
1904 	last = pindex + count - 1;
1905 	for (;;) {
1906 		sb = SWAP_PCTRIE_LOOKUP_GE(&object->un_pager.swp.swp_blks,
1907 		    rounddown(pindex, SWAP_META_PAGES));
1908 		if (sb == NULL || sb->p > last)
1909 			break;
1910 		empty = true;
1911 		for (i = 0; i < SWAP_META_PAGES; i++) {
1912 			if (sb->d[i] == SWAPBLK_NONE)
1913 				continue;
1914 			if (pindex <= sb->p + i && sb->p + i <= last) {
1915 				swp_pager_freeswapspace(sb->d[i], 1);
1916 				sb->d[i] = SWAPBLK_NONE;
1917 			} else
1918 				empty = false;
1919 		}
1920 		pindex = sb->p + SWAP_META_PAGES;
1921 		if (empty) {
1922 			SWAP_PCTRIE_REMOVE(&object->un_pager.swp.swp_blks,
1923 			    sb->p);
1924 			uma_zfree(swblk_zone, sb);
1925 		}
1926 	}
1927 }
1928 
1929 /*
1930  * SWP_PAGER_META_FREE_ALL() - destroy all swap metadata associated with object
1931  *
1932  *	This routine locates and destroys all swap metadata associated with
1933  *	an object.
1934  */
1935 static void
1936 swp_pager_meta_free_all(vm_object_t object)
1937 {
1938 	struct swblk *sb;
1939 	vm_pindex_t pindex;
1940 	int i;
1941 
1942 	VM_OBJECT_ASSERT_WLOCKED(object);
1943 	if (object->type != OBJT_SWAP)
1944 		return;
1945 
1946 	for (pindex = 0; (sb = SWAP_PCTRIE_LOOKUP_GE(
1947 	    &object->un_pager.swp.swp_blks, pindex)) != NULL;) {
1948 		pindex = sb->p + SWAP_META_PAGES;
1949 		for (i = 0; i < SWAP_META_PAGES; i++) {
1950 			if (sb->d[i] != SWAPBLK_NONE)
1951 				swp_pager_freeswapspace(sb->d[i], 1);
1952 		}
1953 		SWAP_PCTRIE_REMOVE(&object->un_pager.swp.swp_blks, sb->p);
1954 		uma_zfree(swblk_zone, sb);
1955 	}
1956 }
1957 
1958 /*
1959  * SWP_PAGER_METACTL() -  misc control of swap and vm_page_t meta data.
1960  *
1961  *	This routine is capable of looking up, popping, or freeing
1962  *	swapblk assignments in the swap meta data or in the vm_page_t.
1963  *	The routine typically returns the swapblk being looked-up, or popped,
1964  *	or SWAPBLK_NONE if the block was freed, or SWAPBLK_NONE if the block
1965  *	was invalid.  This routine will automatically free any invalid
1966  *	meta-data swapblks.
1967  *
1968  *	When acting on a busy resident page and paging is in progress, we
1969  *	have to wait until paging is complete but otherwise can act on the
1970  *	busy page.
1971  *
1972  *	SWM_FREE	remove and free swap block from metadata
1973  *	SWM_POP		remove from meta data but do not free.. pop it out
1974  */
1975 static daddr_t
1976 swp_pager_meta_ctl(vm_object_t object, vm_pindex_t pindex, int flags)
1977 {
1978 	struct swblk *sb;
1979 	daddr_t r1;
1980 	int i;
1981 
1982 	if ((flags & (SWM_FREE | SWM_POP)) != 0)
1983 		VM_OBJECT_ASSERT_WLOCKED(object);
1984 	else
1985 		VM_OBJECT_ASSERT_LOCKED(object);
1986 
1987 	/*
1988 	 * The meta data only exists if the object is OBJT_SWAP
1989 	 * and even then might not be allocated yet.
1990 	 */
1991 	if (object->type != OBJT_SWAP)
1992 		return (SWAPBLK_NONE);
1993 
1994 	sb = SWAP_PCTRIE_LOOKUP(&object->un_pager.swp.swp_blks,
1995 	    rounddown(pindex, SWAP_META_PAGES));
1996 	if (sb == NULL)
1997 		return (SWAPBLK_NONE);
1998 	r1 = sb->d[pindex % SWAP_META_PAGES];
1999 	if (r1 == SWAPBLK_NONE)
2000 		return (SWAPBLK_NONE);
2001 	if ((flags & (SWM_FREE | SWM_POP)) != 0) {
2002 		sb->d[pindex % SWAP_META_PAGES] = SWAPBLK_NONE;
2003 		for (i = 0; i < SWAP_META_PAGES; i++) {
2004 			if (sb->d[i] != SWAPBLK_NONE)
2005 				break;
2006 		}
2007 		if (i == SWAP_META_PAGES) {
2008 			SWAP_PCTRIE_REMOVE(&object->un_pager.swp.swp_blks,
2009 			    rounddown(pindex, SWAP_META_PAGES));
2010 			uma_zfree(swblk_zone, sb);
2011 		}
2012 	}
2013 	if ((flags & SWM_FREE) != 0) {
2014 		swp_pager_freeswapspace(r1, 1);
2015 		r1 = SWAPBLK_NONE;
2016 	}
2017 	return (r1);
2018 }
2019 
2020 /*
2021  * Returns the least page index which is greater than or equal to the
2022  * parameter pindex and for which there is a swap block allocated.
2023  * Returns object's size if the object's type is not swap or if there
2024  * are no allocated swap blocks for the object after the requested
2025  * pindex.
2026  */
2027 vm_pindex_t
2028 swap_pager_find_least(vm_object_t object, vm_pindex_t pindex)
2029 {
2030 	struct swblk *sb;
2031 	int i;
2032 
2033 	VM_OBJECT_ASSERT_LOCKED(object);
2034 	if (object->type != OBJT_SWAP)
2035 		return (object->size);
2036 
2037 	sb = SWAP_PCTRIE_LOOKUP_GE(&object->un_pager.swp.swp_blks,
2038 	    rounddown(pindex, SWAP_META_PAGES));
2039 	if (sb == NULL)
2040 		return (object->size);
2041 	if (sb->p < pindex) {
2042 		for (i = pindex % SWAP_META_PAGES; i < SWAP_META_PAGES; i++) {
2043 			if (sb->d[i] != SWAPBLK_NONE)
2044 				return (sb->p + i);
2045 		}
2046 		sb = SWAP_PCTRIE_LOOKUP_GE(&object->un_pager.swp.swp_blks,
2047 		    roundup(pindex, SWAP_META_PAGES));
2048 		if (sb == NULL)
2049 			return (object->size);
2050 	}
2051 	for (i = 0; i < SWAP_META_PAGES; i++) {
2052 		if (sb->d[i] != SWAPBLK_NONE)
2053 			return (sb->p + i);
2054 	}
2055 
2056 	/*
2057 	 * We get here if a swblk is present in the trie but it
2058 	 * doesn't map any blocks.
2059 	 */
2060 	MPASS(0);
2061 	return (object->size);
2062 }
2063 
2064 /*
2065  * System call swapon(name) enables swapping on device name,
2066  * which must be in the swdevsw.  Return EBUSY
2067  * if already swapping on this device.
2068  */
2069 #ifndef _SYS_SYSPROTO_H_
2070 struct swapon_args {
2071 	char *name;
2072 };
2073 #endif
2074 
2075 /*
2076  * MPSAFE
2077  */
2078 /* ARGSUSED */
2079 int
2080 sys_swapon(struct thread *td, struct swapon_args *uap)
2081 {
2082 	struct vattr attr;
2083 	struct vnode *vp;
2084 	struct nameidata nd;
2085 	int error;
2086 
2087 	error = priv_check(td, PRIV_SWAPON);
2088 	if (error)
2089 		return (error);
2090 
2091 	sx_xlock(&swdev_syscall_lock);
2092 
2093 	/*
2094 	 * Swap metadata may not fit in the KVM if we have physical
2095 	 * memory of >1GB.
2096 	 */
2097 	if (swblk_zone == NULL) {
2098 		error = ENOMEM;
2099 		goto done;
2100 	}
2101 
2102 	NDINIT(&nd, LOOKUP, ISOPEN | FOLLOW | AUDITVNODE1, UIO_USERSPACE,
2103 	    uap->name, td);
2104 	error = namei(&nd);
2105 	if (error)
2106 		goto done;
2107 
2108 	NDFREE(&nd, NDF_ONLY_PNBUF);
2109 	vp = nd.ni_vp;
2110 
2111 	if (vn_isdisk(vp, &error)) {
2112 		error = swapongeom(vp);
2113 	} else if (vp->v_type == VREG &&
2114 	    (vp->v_mount->mnt_vfc->vfc_flags & VFCF_NETWORK) != 0 &&
2115 	    (error = VOP_GETATTR(vp, &attr, td->td_ucred)) == 0) {
2116 		/*
2117 		 * Allow direct swapping to NFS regular files in the same
2118 		 * way that nfs_mountroot() sets up diskless swapping.
2119 		 */
2120 		error = swaponvp(td, vp, attr.va_size / DEV_BSIZE);
2121 	}
2122 
2123 	if (error)
2124 		vrele(vp);
2125 done:
2126 	sx_xunlock(&swdev_syscall_lock);
2127 	return (error);
2128 }
2129 
2130 /*
2131  * Check that the total amount of swap currently configured does not
2132  * exceed half the theoretical maximum.  If it does, print a warning
2133  * message.
2134  */
2135 static void
2136 swapon_check_swzone(void)
2137 {
2138 	unsigned long maxpages, npages;
2139 
2140 	npages = swap_total / PAGE_SIZE;
2141 	/* absolute maximum we can handle assuming 100% efficiency */
2142 	maxpages = uma_zone_get_max(swblk_zone) * SWAP_META_PAGES;
2143 
2144 	/* recommend using no more than half that amount */
2145 	if (npages > maxpages / 2) {
2146 		printf("warning: total configured swap (%lu pages) "
2147 		    "exceeds maximum recommended amount (%lu pages).\n",
2148 		    npages, maxpages / 2);
2149 		printf("warning: increase kern.maxswzone "
2150 		    "or reduce amount of swap.\n");
2151 	}
2152 }
2153 
2154 static void
2155 swaponsomething(struct vnode *vp, void *id, u_long nblks,
2156     sw_strategy_t *strategy, sw_close_t *close, dev_t dev, int flags)
2157 {
2158 	struct swdevt *sp, *tsp;
2159 	swblk_t dvbase;
2160 	u_long mblocks;
2161 
2162 	/*
2163 	 * nblks is in DEV_BSIZE'd chunks, convert to PAGE_SIZE'd chunks.
2164 	 * First chop nblks off to page-align it, then convert.
2165 	 *
2166 	 * sw->sw_nblks is in page-sized chunks now too.
2167 	 */
2168 	nblks &= ~(ctodb(1) - 1);
2169 	nblks = dbtoc(nblks);
2170 
2171 	/*
2172 	 * If we go beyond this, we get overflows in the radix
2173 	 * tree bitmap code.
2174 	 */
2175 	mblocks = 0x40000000 / BLIST_META_RADIX;
2176 	if (nblks > mblocks) {
2177 		printf(
2178     "WARNING: reducing swap size to maximum of %luMB per unit\n",
2179 		    mblocks / 1024 / 1024 * PAGE_SIZE);
2180 		nblks = mblocks;
2181 	}
2182 
2183 	sp = malloc(sizeof *sp, M_VMPGDATA, M_WAITOK | M_ZERO);
2184 	sp->sw_vp = vp;
2185 	sp->sw_id = id;
2186 	sp->sw_dev = dev;
2187 	sp->sw_flags = 0;
2188 	sp->sw_nblks = nblks;
2189 	sp->sw_used = 0;
2190 	sp->sw_strategy = strategy;
2191 	sp->sw_close = close;
2192 	sp->sw_flags = flags;
2193 
2194 	sp->sw_blist = blist_create(nblks, M_WAITOK);
2195 	/*
2196 	 * Do not free the first two block in order to avoid overwriting
2197 	 * any bsd label at the front of the partition
2198 	 */
2199 	blist_free(sp->sw_blist, 2, nblks - 2);
2200 
2201 	dvbase = 0;
2202 	mtx_lock(&sw_dev_mtx);
2203 	TAILQ_FOREACH(tsp, &swtailq, sw_list) {
2204 		if (tsp->sw_end >= dvbase) {
2205 			/*
2206 			 * We put one uncovered page between the devices
2207 			 * in order to definitively prevent any cross-device
2208 			 * I/O requests
2209 			 */
2210 			dvbase = tsp->sw_end + 1;
2211 		}
2212 	}
2213 	sp->sw_first = dvbase;
2214 	sp->sw_end = dvbase + nblks;
2215 	TAILQ_INSERT_TAIL(&swtailq, sp, sw_list);
2216 	nswapdev++;
2217 	swap_pager_avail += nblks - 2;
2218 	swap_total += (vm_ooffset_t)nblks * PAGE_SIZE;
2219 	swapon_check_swzone();
2220 	swp_sizecheck();
2221 	mtx_unlock(&sw_dev_mtx);
2222 	EVENTHANDLER_INVOKE(swapon, sp);
2223 }
2224 
2225 /*
2226  * SYSCALL: swapoff(devname)
2227  *
2228  * Disable swapping on the given device.
2229  *
2230  * XXX: Badly designed system call: it should use a device index
2231  * rather than filename as specification.  We keep sw_vp around
2232  * only to make this work.
2233  */
2234 #ifndef _SYS_SYSPROTO_H_
2235 struct swapoff_args {
2236 	char *name;
2237 };
2238 #endif
2239 
2240 /*
2241  * MPSAFE
2242  */
2243 /* ARGSUSED */
2244 int
2245 sys_swapoff(struct thread *td, struct swapoff_args *uap)
2246 {
2247 	struct vnode *vp;
2248 	struct nameidata nd;
2249 	struct swdevt *sp;
2250 	int error;
2251 
2252 	error = priv_check(td, PRIV_SWAPOFF);
2253 	if (error)
2254 		return (error);
2255 
2256 	sx_xlock(&swdev_syscall_lock);
2257 
2258 	NDINIT(&nd, LOOKUP, FOLLOW | AUDITVNODE1, UIO_USERSPACE, uap->name,
2259 	    td);
2260 	error = namei(&nd);
2261 	if (error)
2262 		goto done;
2263 	NDFREE(&nd, NDF_ONLY_PNBUF);
2264 	vp = nd.ni_vp;
2265 
2266 	mtx_lock(&sw_dev_mtx);
2267 	TAILQ_FOREACH(sp, &swtailq, sw_list) {
2268 		if (sp->sw_vp == vp)
2269 			break;
2270 	}
2271 	mtx_unlock(&sw_dev_mtx);
2272 	if (sp == NULL) {
2273 		error = EINVAL;
2274 		goto done;
2275 	}
2276 	error = swapoff_one(sp, td->td_ucred);
2277 done:
2278 	sx_xunlock(&swdev_syscall_lock);
2279 	return (error);
2280 }
2281 
2282 static int
2283 swapoff_one(struct swdevt *sp, struct ucred *cred)
2284 {
2285 	u_long nblks;
2286 #ifdef MAC
2287 	int error;
2288 #endif
2289 
2290 	sx_assert(&swdev_syscall_lock, SA_XLOCKED);
2291 #ifdef MAC
2292 	(void) vn_lock(sp->sw_vp, LK_EXCLUSIVE | LK_RETRY);
2293 	error = mac_system_check_swapoff(cred, sp->sw_vp);
2294 	(void) VOP_UNLOCK(sp->sw_vp, 0);
2295 	if (error != 0)
2296 		return (error);
2297 #endif
2298 	nblks = sp->sw_nblks;
2299 
2300 	/*
2301 	 * We can turn off this swap device safely only if the
2302 	 * available virtual memory in the system will fit the amount
2303 	 * of data we will have to page back in, plus an epsilon so
2304 	 * the system doesn't become critically low on swap space.
2305 	 */
2306 	if (vm_cnt.v_free_count + swap_pager_avail < nblks + nswap_lowat)
2307 		return (ENOMEM);
2308 
2309 	/*
2310 	 * Prevent further allocations on this device.
2311 	 */
2312 	mtx_lock(&sw_dev_mtx);
2313 	sp->sw_flags |= SW_CLOSING;
2314 	swap_pager_avail -= blist_fill(sp->sw_blist, 0, nblks);
2315 	swap_total -= (vm_ooffset_t)nblks * PAGE_SIZE;
2316 	mtx_unlock(&sw_dev_mtx);
2317 
2318 	/*
2319 	 * Page in the contents of the device and close it.
2320 	 */
2321 	swap_pager_swapoff(sp);
2322 
2323 	sp->sw_close(curthread, sp);
2324 	mtx_lock(&sw_dev_mtx);
2325 	sp->sw_id = NULL;
2326 	TAILQ_REMOVE(&swtailq, sp, sw_list);
2327 	nswapdev--;
2328 	if (nswapdev == 0) {
2329 		swap_pager_full = 2;
2330 		swap_pager_almost_full = 1;
2331 	}
2332 	if (swdevhd == sp)
2333 		swdevhd = NULL;
2334 	mtx_unlock(&sw_dev_mtx);
2335 	blist_destroy(sp->sw_blist);
2336 	free(sp, M_VMPGDATA);
2337 	return (0);
2338 }
2339 
2340 void
2341 swapoff_all(void)
2342 {
2343 	struct swdevt *sp, *spt;
2344 	const char *devname;
2345 	int error;
2346 
2347 	sx_xlock(&swdev_syscall_lock);
2348 
2349 	mtx_lock(&sw_dev_mtx);
2350 	TAILQ_FOREACH_SAFE(sp, &swtailq, sw_list, spt) {
2351 		mtx_unlock(&sw_dev_mtx);
2352 		if (vn_isdisk(sp->sw_vp, NULL))
2353 			devname = devtoname(sp->sw_vp->v_rdev);
2354 		else
2355 			devname = "[file]";
2356 		error = swapoff_one(sp, thread0.td_ucred);
2357 		if (error != 0) {
2358 			printf("Cannot remove swap device %s (error=%d), "
2359 			    "skipping.\n", devname, error);
2360 		} else if (bootverbose) {
2361 			printf("Swap device %s removed.\n", devname);
2362 		}
2363 		mtx_lock(&sw_dev_mtx);
2364 	}
2365 	mtx_unlock(&sw_dev_mtx);
2366 
2367 	sx_xunlock(&swdev_syscall_lock);
2368 }
2369 
2370 void
2371 swap_pager_status(int *total, int *used)
2372 {
2373 	struct swdevt *sp;
2374 
2375 	*total = 0;
2376 	*used = 0;
2377 	mtx_lock(&sw_dev_mtx);
2378 	TAILQ_FOREACH(sp, &swtailq, sw_list) {
2379 		*total += sp->sw_nblks;
2380 		*used += sp->sw_used;
2381 	}
2382 	mtx_unlock(&sw_dev_mtx);
2383 }
2384 
2385 int
2386 swap_dev_info(int name, struct xswdev *xs, char *devname, size_t len)
2387 {
2388 	struct swdevt *sp;
2389 	const char *tmp_devname;
2390 	int error, n;
2391 
2392 	n = 0;
2393 	error = ENOENT;
2394 	mtx_lock(&sw_dev_mtx);
2395 	TAILQ_FOREACH(sp, &swtailq, sw_list) {
2396 		if (n != name) {
2397 			n++;
2398 			continue;
2399 		}
2400 		xs->xsw_version = XSWDEV_VERSION;
2401 		xs->xsw_dev = sp->sw_dev;
2402 		xs->xsw_flags = sp->sw_flags;
2403 		xs->xsw_nblks = sp->sw_nblks;
2404 		xs->xsw_used = sp->sw_used;
2405 		if (devname != NULL) {
2406 			if (vn_isdisk(sp->sw_vp, NULL))
2407 				tmp_devname = devtoname(sp->sw_vp->v_rdev);
2408 			else
2409 				tmp_devname = "[file]";
2410 			strncpy(devname, tmp_devname, len);
2411 		}
2412 		error = 0;
2413 		break;
2414 	}
2415 	mtx_unlock(&sw_dev_mtx);
2416 	return (error);
2417 }
2418 
2419 #if defined(COMPAT_FREEBSD11)
2420 #define XSWDEV_VERSION_11	1
2421 struct xswdev11 {
2422 	u_int	xsw_version;
2423 	uint32_t xsw_dev;
2424 	int	xsw_flags;
2425 	int	xsw_nblks;
2426 	int     xsw_used;
2427 };
2428 #endif
2429 
2430 static int
2431 sysctl_vm_swap_info(SYSCTL_HANDLER_ARGS)
2432 {
2433 	struct xswdev xs;
2434 #if defined(COMPAT_FREEBSD11)
2435 	struct xswdev11 xs11;
2436 #endif
2437 	int error;
2438 
2439 	if (arg2 != 1)			/* name length */
2440 		return (EINVAL);
2441 	error = swap_dev_info(*(int *)arg1, &xs, NULL, 0);
2442 	if (error != 0)
2443 		return (error);
2444 #if defined(COMPAT_FREEBSD11)
2445 	if (req->oldlen == sizeof(xs11)) {
2446 		xs11.xsw_version = XSWDEV_VERSION_11;
2447 		xs11.xsw_dev = xs.xsw_dev; /* truncation */
2448 		xs11.xsw_flags = xs.xsw_flags;
2449 		xs11.xsw_nblks = xs.xsw_nblks;
2450 		xs11.xsw_used = xs.xsw_used;
2451 		error = SYSCTL_OUT(req, &xs11, sizeof(xs11));
2452 	} else
2453 #endif
2454 		error = SYSCTL_OUT(req, &xs, sizeof(xs));
2455 	return (error);
2456 }
2457 
2458 SYSCTL_INT(_vm, OID_AUTO, nswapdev, CTLFLAG_RD, &nswapdev, 0,
2459     "Number of swap devices");
2460 SYSCTL_NODE(_vm, OID_AUTO, swap_info, CTLFLAG_RD | CTLFLAG_MPSAFE,
2461     sysctl_vm_swap_info,
2462     "Swap statistics by device");
2463 
2464 /*
2465  * Count the approximate swap usage in pages for a vmspace.  The
2466  * shadowed or not yet copied on write swap blocks are not accounted.
2467  * The map must be locked.
2468  */
2469 long
2470 vmspace_swap_count(struct vmspace *vmspace)
2471 {
2472 	vm_map_t map;
2473 	vm_map_entry_t cur;
2474 	vm_object_t object;
2475 	struct swblk *sb;
2476 	vm_pindex_t e, pi;
2477 	long count;
2478 	int i;
2479 
2480 	map = &vmspace->vm_map;
2481 	count = 0;
2482 
2483 	for (cur = map->header.next; cur != &map->header; cur = cur->next) {
2484 		if ((cur->eflags & MAP_ENTRY_IS_SUB_MAP) != 0)
2485 			continue;
2486 		object = cur->object.vm_object;
2487 		if (object == NULL || object->type != OBJT_SWAP)
2488 			continue;
2489 		VM_OBJECT_RLOCK(object);
2490 		if (object->type != OBJT_SWAP)
2491 			goto unlock;
2492 		pi = OFF_TO_IDX(cur->offset);
2493 		e = pi + OFF_TO_IDX(cur->end - cur->start);
2494 		for (;; pi = sb->p + SWAP_META_PAGES) {
2495 			sb = SWAP_PCTRIE_LOOKUP_GE(
2496 			    &object->un_pager.swp.swp_blks, pi);
2497 			if (sb == NULL || sb->p >= e)
2498 				break;
2499 			for (i = 0; i < SWAP_META_PAGES; i++) {
2500 				if (sb->p + i < e &&
2501 				    sb->d[i] != SWAPBLK_NONE)
2502 					count++;
2503 			}
2504 		}
2505 unlock:
2506 		VM_OBJECT_RUNLOCK(object);
2507 	}
2508 	return (count);
2509 }
2510 
2511 /*
2512  * GEOM backend
2513  *
2514  * Swapping onto disk devices.
2515  *
2516  */
2517 
2518 static g_orphan_t swapgeom_orphan;
2519 
2520 static struct g_class g_swap_class = {
2521 	.name = "SWAP",
2522 	.version = G_VERSION,
2523 	.orphan = swapgeom_orphan,
2524 };
2525 
2526 DECLARE_GEOM_CLASS(g_swap_class, g_class);
2527 
2528 
2529 static void
2530 swapgeom_close_ev(void *arg, int flags)
2531 {
2532 	struct g_consumer *cp;
2533 
2534 	cp = arg;
2535 	g_access(cp, -1, -1, 0);
2536 	g_detach(cp);
2537 	g_destroy_consumer(cp);
2538 }
2539 
2540 /*
2541  * Add a reference to the g_consumer for an inflight transaction.
2542  */
2543 static void
2544 swapgeom_acquire(struct g_consumer *cp)
2545 {
2546 
2547 	mtx_assert(&sw_dev_mtx, MA_OWNED);
2548 	cp->index++;
2549 }
2550 
2551 /*
2552  * Remove a reference from the g_consumer.  Post a close event if all
2553  * references go away, since the function might be called from the
2554  * biodone context.
2555  */
2556 static void
2557 swapgeom_release(struct g_consumer *cp, struct swdevt *sp)
2558 {
2559 
2560 	mtx_assert(&sw_dev_mtx, MA_OWNED);
2561 	cp->index--;
2562 	if (cp->index == 0) {
2563 		if (g_post_event(swapgeom_close_ev, cp, M_NOWAIT, NULL) == 0)
2564 			sp->sw_id = NULL;
2565 	}
2566 }
2567 
2568 static void
2569 swapgeom_done(struct bio *bp2)
2570 {
2571 	struct swdevt *sp;
2572 	struct buf *bp;
2573 	struct g_consumer *cp;
2574 
2575 	bp = bp2->bio_caller2;
2576 	cp = bp2->bio_from;
2577 	bp->b_ioflags = bp2->bio_flags;
2578 	if (bp2->bio_error)
2579 		bp->b_ioflags |= BIO_ERROR;
2580 	bp->b_resid = bp->b_bcount - bp2->bio_completed;
2581 	bp->b_error = bp2->bio_error;
2582 	bufdone(bp);
2583 	sp = bp2->bio_caller1;
2584 	mtx_lock(&sw_dev_mtx);
2585 	swapgeom_release(cp, sp);
2586 	mtx_unlock(&sw_dev_mtx);
2587 	g_destroy_bio(bp2);
2588 }
2589 
2590 static void
2591 swapgeom_strategy(struct buf *bp, struct swdevt *sp)
2592 {
2593 	struct bio *bio;
2594 	struct g_consumer *cp;
2595 
2596 	mtx_lock(&sw_dev_mtx);
2597 	cp = sp->sw_id;
2598 	if (cp == NULL) {
2599 		mtx_unlock(&sw_dev_mtx);
2600 		bp->b_error = ENXIO;
2601 		bp->b_ioflags |= BIO_ERROR;
2602 		bufdone(bp);
2603 		return;
2604 	}
2605 	swapgeom_acquire(cp);
2606 	mtx_unlock(&sw_dev_mtx);
2607 	if (bp->b_iocmd == BIO_WRITE)
2608 		bio = g_new_bio();
2609 	else
2610 		bio = g_alloc_bio();
2611 	if (bio == NULL) {
2612 		mtx_lock(&sw_dev_mtx);
2613 		swapgeom_release(cp, sp);
2614 		mtx_unlock(&sw_dev_mtx);
2615 		bp->b_error = ENOMEM;
2616 		bp->b_ioflags |= BIO_ERROR;
2617 		bufdone(bp);
2618 		return;
2619 	}
2620 
2621 	bio->bio_caller1 = sp;
2622 	bio->bio_caller2 = bp;
2623 	bio->bio_cmd = bp->b_iocmd;
2624 	bio->bio_offset = (bp->b_blkno - sp->sw_first) * PAGE_SIZE;
2625 	bio->bio_length = bp->b_bcount;
2626 	bio->bio_done = swapgeom_done;
2627 	if (!buf_mapped(bp)) {
2628 		bio->bio_ma = bp->b_pages;
2629 		bio->bio_data = unmapped_buf;
2630 		bio->bio_ma_offset = (vm_offset_t)bp->b_offset & PAGE_MASK;
2631 		bio->bio_ma_n = bp->b_npages;
2632 		bio->bio_flags |= BIO_UNMAPPED;
2633 	} else {
2634 		bio->bio_data = bp->b_data;
2635 		bio->bio_ma = NULL;
2636 	}
2637 	g_io_request(bio, cp);
2638 	return;
2639 }
2640 
2641 static void
2642 swapgeom_orphan(struct g_consumer *cp)
2643 {
2644 	struct swdevt *sp;
2645 	int destroy;
2646 
2647 	mtx_lock(&sw_dev_mtx);
2648 	TAILQ_FOREACH(sp, &swtailq, sw_list) {
2649 		if (sp->sw_id == cp) {
2650 			sp->sw_flags |= SW_CLOSING;
2651 			break;
2652 		}
2653 	}
2654 	/*
2655 	 * Drop reference we were created with. Do directly since we're in a
2656 	 * special context where we don't have to queue the call to
2657 	 * swapgeom_close_ev().
2658 	 */
2659 	cp->index--;
2660 	destroy = ((sp != NULL) && (cp->index == 0));
2661 	if (destroy)
2662 		sp->sw_id = NULL;
2663 	mtx_unlock(&sw_dev_mtx);
2664 	if (destroy)
2665 		swapgeom_close_ev(cp, 0);
2666 }
2667 
2668 static void
2669 swapgeom_close(struct thread *td, struct swdevt *sw)
2670 {
2671 	struct g_consumer *cp;
2672 
2673 	mtx_lock(&sw_dev_mtx);
2674 	cp = sw->sw_id;
2675 	sw->sw_id = NULL;
2676 	mtx_unlock(&sw_dev_mtx);
2677 
2678 	/*
2679 	 * swapgeom_close() may be called from the biodone context,
2680 	 * where we cannot perform topology changes.  Delegate the
2681 	 * work to the events thread.
2682 	 */
2683 	if (cp != NULL)
2684 		g_waitfor_event(swapgeom_close_ev, cp, M_WAITOK, NULL);
2685 }
2686 
2687 static int
2688 swapongeom_locked(struct cdev *dev, struct vnode *vp)
2689 {
2690 	struct g_provider *pp;
2691 	struct g_consumer *cp;
2692 	static struct g_geom *gp;
2693 	struct swdevt *sp;
2694 	u_long nblks;
2695 	int error;
2696 
2697 	pp = g_dev_getprovider(dev);
2698 	if (pp == NULL)
2699 		return (ENODEV);
2700 	mtx_lock(&sw_dev_mtx);
2701 	TAILQ_FOREACH(sp, &swtailq, sw_list) {
2702 		cp = sp->sw_id;
2703 		if (cp != NULL && cp->provider == pp) {
2704 			mtx_unlock(&sw_dev_mtx);
2705 			return (EBUSY);
2706 		}
2707 	}
2708 	mtx_unlock(&sw_dev_mtx);
2709 	if (gp == NULL)
2710 		gp = g_new_geomf(&g_swap_class, "swap");
2711 	cp = g_new_consumer(gp);
2712 	cp->index = 1;	/* Number of active I/Os, plus one for being active. */
2713 	cp->flags |=  G_CF_DIRECT_SEND | G_CF_DIRECT_RECEIVE;
2714 	g_attach(cp, pp);
2715 	/*
2716 	 * XXX: Every time you think you can improve the margin for
2717 	 * footshooting, somebody depends on the ability to do so:
2718 	 * savecore(8) wants to write to our swapdev so we cannot
2719 	 * set an exclusive count :-(
2720 	 */
2721 	error = g_access(cp, 1, 1, 0);
2722 	if (error != 0) {
2723 		g_detach(cp);
2724 		g_destroy_consumer(cp);
2725 		return (error);
2726 	}
2727 	nblks = pp->mediasize / DEV_BSIZE;
2728 	swaponsomething(vp, cp, nblks, swapgeom_strategy,
2729 	    swapgeom_close, dev2udev(dev),
2730 	    (pp->flags & G_PF_ACCEPT_UNMAPPED) != 0 ? SW_UNMAPPED : 0);
2731 	return (0);
2732 }
2733 
2734 static int
2735 swapongeom(struct vnode *vp)
2736 {
2737 	int error;
2738 
2739 	vn_lock(vp, LK_EXCLUSIVE | LK_RETRY);
2740 	if (vp->v_type != VCHR || (vp->v_iflag & VI_DOOMED) != 0) {
2741 		error = ENOENT;
2742 	} else {
2743 		g_topology_lock();
2744 		error = swapongeom_locked(vp->v_rdev, vp);
2745 		g_topology_unlock();
2746 	}
2747 	VOP_UNLOCK(vp, 0);
2748 	return (error);
2749 }
2750 
2751 /*
2752  * VNODE backend
2753  *
2754  * This is used mainly for network filesystem (read: probably only tested
2755  * with NFS) swapfiles.
2756  *
2757  */
2758 
2759 static void
2760 swapdev_strategy(struct buf *bp, struct swdevt *sp)
2761 {
2762 	struct vnode *vp2;
2763 
2764 	bp->b_blkno = ctodb(bp->b_blkno - sp->sw_first);
2765 
2766 	vp2 = sp->sw_id;
2767 	vhold(vp2);
2768 	if (bp->b_iocmd == BIO_WRITE) {
2769 		if (bp->b_bufobj)
2770 			bufobj_wdrop(bp->b_bufobj);
2771 		bufobj_wref(&vp2->v_bufobj);
2772 	}
2773 	if (bp->b_bufobj != &vp2->v_bufobj)
2774 		bp->b_bufobj = &vp2->v_bufobj;
2775 	bp->b_vp = vp2;
2776 	bp->b_iooffset = dbtob(bp->b_blkno);
2777 	bstrategy(bp);
2778 	return;
2779 }
2780 
2781 static void
2782 swapdev_close(struct thread *td, struct swdevt *sp)
2783 {
2784 
2785 	VOP_CLOSE(sp->sw_vp, FREAD | FWRITE, td->td_ucred, td);
2786 	vrele(sp->sw_vp);
2787 }
2788 
2789 
2790 static int
2791 swaponvp(struct thread *td, struct vnode *vp, u_long nblks)
2792 {
2793 	struct swdevt *sp;
2794 	int error;
2795 
2796 	if (nblks == 0)
2797 		return (ENXIO);
2798 	mtx_lock(&sw_dev_mtx);
2799 	TAILQ_FOREACH(sp, &swtailq, sw_list) {
2800 		if (sp->sw_id == vp) {
2801 			mtx_unlock(&sw_dev_mtx);
2802 			return (EBUSY);
2803 		}
2804 	}
2805 	mtx_unlock(&sw_dev_mtx);
2806 
2807 	(void) vn_lock(vp, LK_EXCLUSIVE | LK_RETRY);
2808 #ifdef MAC
2809 	error = mac_system_check_swapon(td->td_ucred, vp);
2810 	if (error == 0)
2811 #endif
2812 		error = VOP_OPEN(vp, FREAD | FWRITE, td->td_ucred, td, NULL);
2813 	(void) VOP_UNLOCK(vp, 0);
2814 	if (error)
2815 		return (error);
2816 
2817 	swaponsomething(vp, vp, nblks, swapdev_strategy, swapdev_close,
2818 	    NODEV, 0);
2819 	return (0);
2820 }
2821 
2822 static int
2823 sysctl_swap_async_max(SYSCTL_HANDLER_ARGS)
2824 {
2825 	int error, new, n;
2826 
2827 	new = nsw_wcount_async_max;
2828 	error = sysctl_handle_int(oidp, &new, 0, req);
2829 	if (error != 0 || req->newptr == NULL)
2830 		return (error);
2831 
2832 	if (new > nswbuf / 2 || new < 1)
2833 		return (EINVAL);
2834 
2835 	mtx_lock(&pbuf_mtx);
2836 	while (nsw_wcount_async_max != new) {
2837 		/*
2838 		 * Adjust difference.  If the current async count is too low,
2839 		 * we will need to sqeeze our update slowly in.  Sleep with a
2840 		 * higher priority than getpbuf() to finish faster.
2841 		 */
2842 		n = new - nsw_wcount_async_max;
2843 		if (nsw_wcount_async + n >= 0) {
2844 			nsw_wcount_async += n;
2845 			nsw_wcount_async_max += n;
2846 			wakeup(&nsw_wcount_async);
2847 		} else {
2848 			nsw_wcount_async_max -= nsw_wcount_async;
2849 			nsw_wcount_async = 0;
2850 			msleep(&nsw_wcount_async, &pbuf_mtx, PSWP,
2851 			    "swpsysctl", 0);
2852 		}
2853 	}
2854 	mtx_unlock(&pbuf_mtx);
2855 
2856 	return (0);
2857 }
2858