xref: /freebsd/sys/vm/swap_pager.c (revision 4731124cace5e7a0224e29784617d2856e5c59ab)
1 /*-
2  * SPDX-License-Identifier: BSD-4-Clause
3  *
4  * Copyright (c) 1998 Matthew Dillon,
5  * Copyright (c) 1994 John S. Dyson
6  * Copyright (c) 1990 University of Utah.
7  * Copyright (c) 1982, 1986, 1989, 1993
8  *	The Regents of the University of California.  All rights reserved.
9  *
10  * This code is derived from software contributed to Berkeley by
11  * the Systems Programming Group of the University of Utah Computer
12  * Science Department.
13  *
14  * Redistribution and use in source and binary forms, with or without
15  * modification, are permitted provided that the following conditions
16  * are met:
17  * 1. Redistributions of source code must retain the above copyright
18  *    notice, this list of conditions and the following disclaimer.
19  * 2. Redistributions in binary form must reproduce the above copyright
20  *    notice, this list of conditions and the following disclaimer in the
21  *    documentation and/or other materials provided with the distribution.
22  * 3. All advertising materials mentioning features or use of this software
23  *    must display the following acknowledgement:
24  *	This product includes software developed by the University of
25  *	California, Berkeley and its contributors.
26  * 4. Neither the name of the University nor the names of its contributors
27  *    may be used to endorse or promote products derived from this software
28  *    without specific prior written permission.
29  *
30  * THIS SOFTWARE IS PROVIDED BY THE REGENTS AND CONTRIBUTORS ``AS IS'' AND
31  * ANY EXPRESS OR IMPLIED WARRANTIES, INCLUDING, BUT NOT LIMITED TO, THE
32  * IMPLIED WARRANTIES OF MERCHANTABILITY AND FITNESS FOR A PARTICULAR PURPOSE
33  * ARE DISCLAIMED.  IN NO EVENT SHALL THE REGENTS OR CONTRIBUTORS BE LIABLE
34  * FOR ANY DIRECT, INDIRECT, INCIDENTAL, SPECIAL, EXEMPLARY, OR CONSEQUENTIAL
35  * DAMAGES (INCLUDING, BUT NOT LIMITED TO, PROCUREMENT OF SUBSTITUTE GOODS
36  * OR SERVICES; LOSS OF USE, DATA, OR PROFITS; OR BUSINESS INTERRUPTION)
37  * HOWEVER CAUSED AND ON ANY THEORY OF LIABILITY, WHETHER IN CONTRACT, STRICT
38  * LIABILITY, OR TORT (INCLUDING NEGLIGENCE OR OTHERWISE) ARISING IN ANY WAY
39  * OUT OF THE USE OF THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF
40  * SUCH DAMAGE.
41  *
42  *				New Swap System
43  *				Matthew Dillon
44  *
45  * Radix Bitmap 'blists'.
46  *
47  *	- The new swapper uses the new radix bitmap code.  This should scale
48  *	  to arbitrarily small or arbitrarily large swap spaces and an almost
49  *	  arbitrary degree of fragmentation.
50  *
51  * Features:
52  *
53  *	- on the fly reallocation of swap during putpages.  The new system
54  *	  does not try to keep previously allocated swap blocks for dirty
55  *	  pages.
56  *
57  *	- on the fly deallocation of swap
58  *
59  *	- No more garbage collection required.  Unnecessarily allocated swap
60  *	  blocks only exist for dirty vm_page_t's now and these are already
61  *	  cycled (in a high-load system) by the pager.  We also do on-the-fly
62  *	  removal of invalidated swap blocks when a page is destroyed
63  *	  or renamed.
64  *
65  * from: Utah $Hdr: swap_pager.c 1.4 91/04/30$
66  *
67  *	@(#)swap_pager.c	8.9 (Berkeley) 3/21/94
68  *	@(#)vm_swap.c	8.5 (Berkeley) 2/17/94
69  */
70 
71 #include <sys/cdefs.h>
72 __FBSDID("$FreeBSD$");
73 
74 #include "opt_vm.h"
75 
76 #include <sys/param.h>
77 #include <sys/bio.h>
78 #include <sys/blist.h>
79 #include <sys/buf.h>
80 #include <sys/conf.h>
81 #include <sys/disk.h>
82 #include <sys/disklabel.h>
83 #include <sys/eventhandler.h>
84 #include <sys/fcntl.h>
85 #include <sys/limits.h>
86 #include <sys/lock.h>
87 #include <sys/kernel.h>
88 #include <sys/mount.h>
89 #include <sys/namei.h>
90 #include <sys/malloc.h>
91 #include <sys/pctrie.h>
92 #include <sys/priv.h>
93 #include <sys/proc.h>
94 #include <sys/racct.h>
95 #include <sys/resource.h>
96 #include <sys/resourcevar.h>
97 #include <sys/rwlock.h>
98 #include <sys/sbuf.h>
99 #include <sys/sysctl.h>
100 #include <sys/sysproto.h>
101 #include <sys/systm.h>
102 #include <sys/sx.h>
103 #include <sys/unistd.h>
104 #include <sys/user.h>
105 #include <sys/vmmeter.h>
106 #include <sys/vnode.h>
107 
108 #include <security/mac/mac_framework.h>
109 
110 #include <vm/vm.h>
111 #include <vm/pmap.h>
112 #include <vm/vm_map.h>
113 #include <vm/vm_kern.h>
114 #include <vm/vm_object.h>
115 #include <vm/vm_page.h>
116 #include <vm/vm_pager.h>
117 #include <vm/vm_pageout.h>
118 #include <vm/vm_param.h>
119 #include <vm/swap_pager.h>
120 #include <vm/vm_extern.h>
121 #include <vm/uma.h>
122 
123 #include <geom/geom.h>
124 
125 /*
126  * MAX_PAGEOUT_CLUSTER must be a power of 2 between 1 and 64.
127  * The 64-page limit is due to the radix code (kern/subr_blist.c).
128  */
129 #ifndef MAX_PAGEOUT_CLUSTER
130 #define	MAX_PAGEOUT_CLUSTER	32
131 #endif
132 
133 #if !defined(SWB_NPAGES)
134 #define SWB_NPAGES	MAX_PAGEOUT_CLUSTER
135 #endif
136 
137 #define	SWAP_META_PAGES		PCTRIE_COUNT
138 
139 /*
140  * A swblk structure maps each page index within a
141  * SWAP_META_PAGES-aligned and sized range to the address of an
142  * on-disk swap block (or SWAPBLK_NONE). The collection of these
143  * mappings for an entire vm object is implemented as a pc-trie.
144  */
145 struct swblk {
146 	vm_pindex_t	p;
147 	daddr_t		d[SWAP_META_PAGES];
148 };
149 
150 static MALLOC_DEFINE(M_VMPGDATA, "vm_pgdata", "swap pager private data");
151 static struct mtx sw_dev_mtx;
152 static TAILQ_HEAD(, swdevt) swtailq = TAILQ_HEAD_INITIALIZER(swtailq);
153 static struct swdevt *swdevhd;	/* Allocate from here next */
154 static int nswapdev;		/* Number of swap devices */
155 int swap_pager_avail;
156 static struct sx swdev_syscall_lock;	/* serialize swap(on|off) */
157 
158 static __exclusive_cache_line u_long swap_reserved;
159 static u_long swap_total;
160 static int sysctl_page_shift(SYSCTL_HANDLER_ARGS);
161 
162 static SYSCTL_NODE(_vm_stats, OID_AUTO, swap, CTLFLAG_RD | CTLFLAG_MPSAFE, 0,
163     "VM swap stats");
164 
165 SYSCTL_PROC(_vm, OID_AUTO, swap_reserved, CTLTYPE_U64 | CTLFLAG_RD | CTLFLAG_MPSAFE,
166     &swap_reserved, 0, sysctl_page_shift, "A",
167     "Amount of swap storage needed to back all allocated anonymous memory.");
168 SYSCTL_PROC(_vm, OID_AUTO, swap_total, CTLTYPE_U64 | CTLFLAG_RD | CTLFLAG_MPSAFE,
169     &swap_total, 0, sysctl_page_shift, "A",
170     "Total amount of available swap storage.");
171 
172 static int overcommit = 0;
173 SYSCTL_INT(_vm, VM_OVERCOMMIT, overcommit, CTLFLAG_RW, &overcommit, 0,
174     "Configure virtual memory overcommit behavior. See tuning(7) "
175     "for details.");
176 static unsigned long swzone;
177 SYSCTL_ULONG(_vm, OID_AUTO, swzone, CTLFLAG_RD, &swzone, 0,
178     "Actual size of swap metadata zone");
179 static unsigned long swap_maxpages;
180 SYSCTL_ULONG(_vm, OID_AUTO, swap_maxpages, CTLFLAG_RD, &swap_maxpages, 0,
181     "Maximum amount of swap supported");
182 
183 static COUNTER_U64_DEFINE_EARLY(swap_free_deferred);
184 SYSCTL_COUNTER_U64(_vm_stats_swap, OID_AUTO, free_deferred,
185     CTLFLAG_RD, &swap_free_deferred,
186     "Number of pages that deferred freeing swap space");
187 
188 static COUNTER_U64_DEFINE_EARLY(swap_free_completed);
189 SYSCTL_COUNTER_U64(_vm_stats_swap, OID_AUTO, free_completed,
190     CTLFLAG_RD, &swap_free_completed,
191     "Number of deferred frees completed");
192 
193 /* bits from overcommit */
194 #define	SWAP_RESERVE_FORCE_ON		(1 << 0)
195 #define	SWAP_RESERVE_RLIMIT_ON		(1 << 1)
196 #define	SWAP_RESERVE_ALLOW_NONWIRED	(1 << 2)
197 
198 static int
199 sysctl_page_shift(SYSCTL_HANDLER_ARGS)
200 {
201 	uint64_t newval;
202 	u_long value = *(u_long *)arg1;
203 
204 	newval = ((uint64_t)value) << PAGE_SHIFT;
205 	return (sysctl_handle_64(oidp, &newval, 0, req));
206 }
207 
208 static bool
209 swap_reserve_by_cred_rlimit(u_long pincr, struct ucred *cred, int oc)
210 {
211 	struct uidinfo *uip;
212 	u_long prev;
213 
214 	uip = cred->cr_ruidinfo;
215 
216 	prev = atomic_fetchadd_long(&uip->ui_vmsize, pincr);
217 	if ((oc & SWAP_RESERVE_RLIMIT_ON) != 0 &&
218 	    prev + pincr > lim_cur(curthread, RLIMIT_SWAP) &&
219 	    priv_check(curthread, PRIV_VM_SWAP_NORLIMIT) != 0) {
220 		prev = atomic_fetchadd_long(&uip->ui_vmsize, -pincr);
221 		KASSERT(prev >= pincr, ("negative vmsize for uid = %d\n", uip->ui_uid));
222 		return (false);
223 	}
224 	return (true);
225 }
226 
227 static void
228 swap_release_by_cred_rlimit(u_long pdecr, struct ucred *cred)
229 {
230 	struct uidinfo *uip;
231 #ifdef INVARIANTS
232 	u_long prev;
233 #endif
234 
235 	uip = cred->cr_ruidinfo;
236 
237 #ifdef INVARIANTS
238 	prev = atomic_fetchadd_long(&uip->ui_vmsize, -pdecr);
239 	KASSERT(prev >= pdecr, ("negative vmsize for uid = %d\n", uip->ui_uid));
240 #else
241 	atomic_subtract_long(&uip->ui_vmsize, pdecr);
242 #endif
243 }
244 
245 static void
246 swap_reserve_force_rlimit(u_long pincr, struct ucred *cred)
247 {
248 	struct uidinfo *uip;
249 
250 	uip = cred->cr_ruidinfo;
251 	atomic_add_long(&uip->ui_vmsize, pincr);
252 }
253 
254 bool
255 swap_reserve(vm_ooffset_t incr)
256 {
257 
258 	return (swap_reserve_by_cred(incr, curthread->td_ucred));
259 }
260 
261 bool
262 swap_reserve_by_cred(vm_ooffset_t incr, struct ucred *cred)
263 {
264 	u_long r, s, prev, pincr;
265 #ifdef RACCT
266 	int error;
267 #endif
268 	int oc;
269 	static int curfail;
270 	static struct timeval lastfail;
271 
272 	KASSERT((incr & PAGE_MASK) == 0, ("%s: incr: %ju & PAGE_MASK", __func__,
273 	    (uintmax_t)incr));
274 
275 #ifdef RACCT
276 	if (RACCT_ENABLED()) {
277 		PROC_LOCK(curproc);
278 		error = racct_add(curproc, RACCT_SWAP, incr);
279 		PROC_UNLOCK(curproc);
280 		if (error != 0)
281 			return (false);
282 	}
283 #endif
284 
285 	pincr = atop(incr);
286 	prev = atomic_fetchadd_long(&swap_reserved, pincr);
287 	r = prev + pincr;
288 	s = swap_total;
289 	oc = atomic_load_int(&overcommit);
290 	if (r > s && (oc & SWAP_RESERVE_ALLOW_NONWIRED) != 0) {
291 		s += vm_cnt.v_page_count - vm_cnt.v_free_reserved -
292 		    vm_wire_count();
293 	}
294 	if ((oc & SWAP_RESERVE_FORCE_ON) != 0 && r > s &&
295 	    priv_check(curthread, PRIV_VM_SWAP_NOQUOTA) != 0) {
296 		prev = atomic_fetchadd_long(&swap_reserved, -pincr);
297 		KASSERT(prev >= pincr, ("swap_reserved < incr on overcommit fail"));
298 		goto out_error;
299 	}
300 
301 	if (!swap_reserve_by_cred_rlimit(pincr, cred, oc)) {
302 		prev = atomic_fetchadd_long(&swap_reserved, -pincr);
303 		KASSERT(prev >= pincr, ("swap_reserved < incr on overcommit fail"));
304 		goto out_error;
305 	}
306 
307 	return (true);
308 
309 out_error:
310 	if (ppsratecheck(&lastfail, &curfail, 1)) {
311 		printf("uid %d, pid %d: swap reservation for %jd bytes failed\n",
312 		    cred->cr_ruidinfo->ui_uid, curproc->p_pid, incr);
313 	}
314 #ifdef RACCT
315 	if (RACCT_ENABLED()) {
316 		PROC_LOCK(curproc);
317 		racct_sub(curproc, RACCT_SWAP, incr);
318 		PROC_UNLOCK(curproc);
319 	}
320 #endif
321 
322 	return (false);
323 }
324 
325 void
326 swap_reserve_force(vm_ooffset_t incr)
327 {
328 	u_long pincr;
329 
330 	KASSERT((incr & PAGE_MASK) == 0, ("%s: incr: %ju & PAGE_MASK", __func__,
331 	    (uintmax_t)incr));
332 
333 #ifdef RACCT
334 	if (RACCT_ENABLED()) {
335 		PROC_LOCK(curproc);
336 		racct_add_force(curproc, RACCT_SWAP, incr);
337 		PROC_UNLOCK(curproc);
338 	}
339 #endif
340 	pincr = atop(incr);
341 	atomic_add_long(&swap_reserved, pincr);
342 	swap_reserve_force_rlimit(pincr, curthread->td_ucred);
343 }
344 
345 void
346 swap_release(vm_ooffset_t decr)
347 {
348 	struct ucred *cred;
349 
350 	PROC_LOCK(curproc);
351 	cred = curproc->p_ucred;
352 	swap_release_by_cred(decr, cred);
353 	PROC_UNLOCK(curproc);
354 }
355 
356 void
357 swap_release_by_cred(vm_ooffset_t decr, struct ucred *cred)
358 {
359 	u_long pdecr;
360 #ifdef INVARIANTS
361 	u_long prev;
362 #endif
363 
364 	KASSERT((decr & PAGE_MASK) == 0, ("%s: decr: %ju & PAGE_MASK", __func__,
365 	    (uintmax_t)decr));
366 
367 	pdecr = atop(decr);
368 #ifdef INVARIANTS
369 	prev = atomic_fetchadd_long(&swap_reserved, -pdecr);
370 	KASSERT(prev >= pdecr, ("swap_reserved < decr"));
371 #else
372 	atomic_subtract_long(&swap_reserved, pdecr);
373 #endif
374 
375 	swap_release_by_cred_rlimit(pdecr, cred);
376 #ifdef RACCT
377 	if (racct_enable)
378 		racct_sub_cred(cred, RACCT_SWAP, decr);
379 #endif
380 }
381 
382 static int swap_pager_full = 2;	/* swap space exhaustion (task killing) */
383 static int swap_pager_almost_full = 1; /* swap space exhaustion (w/hysteresis)*/
384 static struct mtx swbuf_mtx;	/* to sync nsw_wcount_async */
385 static int nsw_wcount_async;	/* limit async write buffers */
386 static int nsw_wcount_async_max;/* assigned maximum			*/
387 int nsw_cluster_max; 		/* maximum VOP I/O allowed		*/
388 
389 static int sysctl_swap_async_max(SYSCTL_HANDLER_ARGS);
390 SYSCTL_PROC(_vm, OID_AUTO, swap_async_max, CTLTYPE_INT | CTLFLAG_RW |
391     CTLFLAG_MPSAFE, NULL, 0, sysctl_swap_async_max, "I",
392     "Maximum running async swap ops");
393 static int sysctl_swap_fragmentation(SYSCTL_HANDLER_ARGS);
394 SYSCTL_PROC(_vm, OID_AUTO, swap_fragmentation, CTLTYPE_STRING | CTLFLAG_RD |
395     CTLFLAG_MPSAFE, NULL, 0, sysctl_swap_fragmentation, "A",
396     "Swap Fragmentation Info");
397 
398 static struct sx sw_alloc_sx;
399 
400 /*
401  * "named" and "unnamed" anon region objects.  Try to reduce the overhead
402  * of searching a named list by hashing it just a little.
403  */
404 
405 #define NOBJLISTS		8
406 
407 #define NOBJLIST(handle)	\
408 	(&swap_pager_object_list[((int)(intptr_t)handle >> 4) & (NOBJLISTS-1)])
409 
410 static struct pagerlst	swap_pager_object_list[NOBJLISTS];
411 static uma_zone_t swwbuf_zone;
412 static uma_zone_t swrbuf_zone;
413 static uma_zone_t swblk_zone;
414 static uma_zone_t swpctrie_zone;
415 
416 /*
417  * pagerops for OBJT_SWAP - "swap pager".  Some ops are also global procedure
418  * calls hooked from other parts of the VM system and do not appear here.
419  * (see vm/swap_pager.h).
420  */
421 static vm_object_t
422 		swap_pager_alloc(void *handle, vm_ooffset_t size,
423 		    vm_prot_t prot, vm_ooffset_t offset, struct ucred *);
424 static void	swap_pager_dealloc(vm_object_t object);
425 static int	swap_pager_getpages(vm_object_t, vm_page_t *, int, int *,
426     int *);
427 static int	swap_pager_getpages_async(vm_object_t, vm_page_t *, int, int *,
428     int *, pgo_getpages_iodone_t, void *);
429 static void	swap_pager_putpages(vm_object_t, vm_page_t *, int, boolean_t, int *);
430 static boolean_t
431 		swap_pager_haspage(vm_object_t object, vm_pindex_t pindex, int *before, int *after);
432 static void	swap_pager_init(void);
433 static void	swap_pager_unswapped(vm_page_t);
434 static void	swap_pager_swapoff(struct swdevt *sp);
435 static void	swap_pager_update_writecount(vm_object_t object,
436     vm_offset_t start, vm_offset_t end);
437 static void	swap_pager_release_writecount(vm_object_t object,
438     vm_offset_t start, vm_offset_t end);
439 static void	swap_pager_freespace(vm_object_t object, vm_pindex_t start,
440     vm_size_t size);
441 
442 const struct pagerops swappagerops = {
443 	.pgo_kvme_type = KVME_TYPE_SWAP,
444 	.pgo_init =	swap_pager_init,	/* early system initialization of pager	*/
445 	.pgo_alloc =	swap_pager_alloc,	/* allocate an OBJT_SWAP object */
446 	.pgo_dealloc =	swap_pager_dealloc,	/* deallocate an OBJT_SWAP object */
447 	.pgo_getpages =	swap_pager_getpages,	/* pagein */
448 	.pgo_getpages_async = swap_pager_getpages_async, /* pagein (async) */
449 	.pgo_putpages =	swap_pager_putpages,	/* pageout */
450 	.pgo_haspage =	swap_pager_haspage,	/* get backing store status for page */
451 	.pgo_pageunswapped = swap_pager_unswapped, /* remove swap related to page */
452 	.pgo_update_writecount = swap_pager_update_writecount,
453 	.pgo_release_writecount = swap_pager_release_writecount,
454 	.pgo_freespace = swap_pager_freespace,
455 };
456 
457 /*
458  * swap_*() routines are externally accessible.  swp_*() routines are
459  * internal.
460  */
461 static int nswap_lowat = 128;	/* in pages, swap_pager_almost_full warn */
462 static int nswap_hiwat = 512;	/* in pages, swap_pager_almost_full warn */
463 
464 SYSCTL_INT(_vm, OID_AUTO, dmmax, CTLFLAG_RD, &nsw_cluster_max, 0,
465     "Maximum size of a swap block in pages");
466 
467 static void	swp_sizecheck(void);
468 static void	swp_pager_async_iodone(struct buf *bp);
469 static bool	swp_pager_swblk_empty(struct swblk *sb, int start, int limit);
470 static void	swp_pager_free_empty_swblk(vm_object_t, struct swblk *sb);
471 static int	swapongeom(struct vnode *);
472 static int	swaponvp(struct thread *, struct vnode *, u_long);
473 static int	swapoff_one(struct swdevt *sp, struct ucred *cred,
474 		    u_int flags);
475 
476 /*
477  * Swap bitmap functions
478  */
479 static void	swp_pager_freeswapspace(daddr_t blk, daddr_t npages);
480 static daddr_t	swp_pager_getswapspace(int *npages);
481 
482 /*
483  * Metadata functions
484  */
485 static daddr_t swp_pager_meta_build(vm_object_t, vm_pindex_t, daddr_t);
486 static void swp_pager_meta_free(vm_object_t, vm_pindex_t, vm_pindex_t);
487 static void swp_pager_meta_transfer(vm_object_t src, vm_object_t dst,
488     vm_pindex_t pindex, vm_pindex_t count);
489 static void swp_pager_meta_free_all(vm_object_t);
490 static daddr_t swp_pager_meta_lookup(vm_object_t, vm_pindex_t);
491 
492 static void
493 swp_pager_init_freerange(daddr_t *start, daddr_t *num)
494 {
495 
496 	*start = SWAPBLK_NONE;
497 	*num = 0;
498 }
499 
500 static void
501 swp_pager_update_freerange(daddr_t *start, daddr_t *num, daddr_t addr)
502 {
503 
504 	if (*start + *num == addr) {
505 		(*num)++;
506 	} else {
507 		swp_pager_freeswapspace(*start, *num);
508 		*start = addr;
509 		*num = 1;
510 	}
511 }
512 
513 static void *
514 swblk_trie_alloc(struct pctrie *ptree)
515 {
516 
517 	return (uma_zalloc(swpctrie_zone, M_NOWAIT | (curproc == pageproc ?
518 	    M_USE_RESERVE : 0)));
519 }
520 
521 static void
522 swblk_trie_free(struct pctrie *ptree, void *node)
523 {
524 
525 	uma_zfree(swpctrie_zone, node);
526 }
527 
528 PCTRIE_DEFINE(SWAP, swblk, p, swblk_trie_alloc, swblk_trie_free);
529 
530 /*
531  * SWP_SIZECHECK() -	update swap_pager_full indication
532  *
533  *	update the swap_pager_almost_full indication and warn when we are
534  *	about to run out of swap space, using lowat/hiwat hysteresis.
535  *
536  *	Clear swap_pager_full ( task killing ) indication when lowat is met.
537  *
538  *	No restrictions on call
539  *	This routine may not block.
540  */
541 static void
542 swp_sizecheck(void)
543 {
544 
545 	if (swap_pager_avail < nswap_lowat) {
546 		if (swap_pager_almost_full == 0) {
547 			printf("swap_pager: out of swap space\n");
548 			swap_pager_almost_full = 1;
549 		}
550 	} else {
551 		swap_pager_full = 0;
552 		if (swap_pager_avail > nswap_hiwat)
553 			swap_pager_almost_full = 0;
554 	}
555 }
556 
557 /*
558  * SWAP_PAGER_INIT() -	initialize the swap pager!
559  *
560  *	Expected to be started from system init.  NOTE:  This code is run
561  *	before much else so be careful what you depend on.  Most of the VM
562  *	system has yet to be initialized at this point.
563  */
564 static void
565 swap_pager_init(void)
566 {
567 	/*
568 	 * Initialize object lists
569 	 */
570 	int i;
571 
572 	for (i = 0; i < NOBJLISTS; ++i)
573 		TAILQ_INIT(&swap_pager_object_list[i]);
574 	mtx_init(&sw_dev_mtx, "swapdev", NULL, MTX_DEF);
575 	sx_init(&sw_alloc_sx, "swspsx");
576 	sx_init(&swdev_syscall_lock, "swsysc");
577 
578 	/*
579 	 * The nsw_cluster_max is constrained by the bp->b_pages[]
580 	 * array, which has maxphys / PAGE_SIZE entries, and our locally
581 	 * defined MAX_PAGEOUT_CLUSTER.   Also be aware that swap ops are
582 	 * constrained by the swap device interleave stripe size.
583 	 *
584 	 * Initialized early so that GEOM_ELI can see it.
585 	 */
586 	nsw_cluster_max = min(maxphys / PAGE_SIZE, MAX_PAGEOUT_CLUSTER);
587 }
588 
589 /*
590  * SWAP_PAGER_SWAP_INIT() - swap pager initialization from pageout process
591  *
592  *	Expected to be started from pageout process once, prior to entering
593  *	its main loop.
594  */
595 void
596 swap_pager_swap_init(void)
597 {
598 	unsigned long n, n2;
599 
600 	/*
601 	 * Number of in-transit swap bp operations.  Don't
602 	 * exhaust the pbufs completely.  Make sure we
603 	 * initialize workable values (0 will work for hysteresis
604 	 * but it isn't very efficient).
605 	 *
606 	 * Currently we hardwire nsw_wcount_async to 4.  This limit is
607 	 * designed to prevent other I/O from having high latencies due to
608 	 * our pageout I/O.  The value 4 works well for one or two active swap
609 	 * devices but is probably a little low if you have more.  Even so,
610 	 * a higher value would probably generate only a limited improvement
611 	 * with three or four active swap devices since the system does not
612 	 * typically have to pageout at extreme bandwidths.   We will want
613 	 * at least 2 per swap devices, and 4 is a pretty good value if you
614 	 * have one NFS swap device due to the command/ack latency over NFS.
615 	 * So it all works out pretty well.
616 	 *
617 	 * nsw_cluster_max is initialized in swap_pager_init().
618 	 */
619 
620 	nsw_wcount_async = 4;
621 	nsw_wcount_async_max = nsw_wcount_async;
622 	mtx_init(&swbuf_mtx, "async swbuf mutex", NULL, MTX_DEF);
623 
624 	swwbuf_zone = pbuf_zsecond_create("swwbuf", nswbuf / 4);
625 	swrbuf_zone = pbuf_zsecond_create("swrbuf", nswbuf / 2);
626 
627 	/*
628 	 * Initialize our zone, taking the user's requested size or
629 	 * estimating the number we need based on the number of pages
630 	 * in the system.
631 	 */
632 	n = maxswzone != 0 ? maxswzone / sizeof(struct swblk) :
633 	    vm_cnt.v_page_count / 2;
634 	swpctrie_zone = uma_zcreate("swpctrie", pctrie_node_size(), NULL, NULL,
635 	    pctrie_zone_init, NULL, UMA_ALIGN_PTR, 0);
636 	swblk_zone = uma_zcreate("swblk", sizeof(struct swblk), NULL, NULL,
637 	    NULL, NULL, _Alignof(struct swblk) - 1, 0);
638 	n2 = n;
639 	do {
640 		if (uma_zone_reserve_kva(swblk_zone, n))
641 			break;
642 		/*
643 		 * if the allocation failed, try a zone two thirds the
644 		 * size of the previous attempt.
645 		 */
646 		n -= ((n + 2) / 3);
647 	} while (n > 0);
648 
649 	/*
650 	 * Often uma_zone_reserve_kva() cannot reserve exactly the
651 	 * requested size.  Account for the difference when
652 	 * calculating swap_maxpages.
653 	 */
654 	n = uma_zone_get_max(swblk_zone);
655 
656 	if (n < n2)
657 		printf("Swap blk zone entries changed from %lu to %lu.\n",
658 		    n2, n);
659 	/* absolute maximum we can handle assuming 100% efficiency */
660 	swap_maxpages = n * SWAP_META_PAGES;
661 	swzone = n * sizeof(struct swblk);
662 	if (!uma_zone_reserve_kva(swpctrie_zone, n))
663 		printf("Cannot reserve swap pctrie zone, "
664 		    "reduce kern.maxswzone.\n");
665 }
666 
667 bool
668 swap_pager_init_object(vm_object_t object, void *handle, struct ucred *cred,
669     vm_ooffset_t size, vm_ooffset_t offset)
670 {
671 	if (cred != NULL) {
672 		if (!swap_reserve_by_cred(size, cred))
673 			return (false);
674 		crhold(cred);
675 	}
676 
677 	object->un_pager.swp.writemappings = 0;
678 	object->handle = handle;
679 	if (cred != NULL) {
680 		object->cred = cred;
681 		object->charge = size;
682 	}
683 	return (true);
684 }
685 
686 static vm_object_t
687 swap_pager_alloc_init(objtype_t otype, void *handle, struct ucred *cred,
688     vm_ooffset_t size, vm_ooffset_t offset)
689 {
690 	vm_object_t object;
691 
692 	/*
693 	 * The un_pager.swp.swp_blks trie is initialized by
694 	 * vm_object_allocate() to ensure the correct order of
695 	 * visibility to other threads.
696 	 */
697 	object = vm_object_allocate(otype, OFF_TO_IDX(offset +
698 	    PAGE_MASK + size));
699 
700 	if (!swap_pager_init_object(object, handle, cred, size, offset)) {
701 		vm_object_deallocate(object);
702 		return (NULL);
703 	}
704 	return (object);
705 }
706 
707 /*
708  * SWAP_PAGER_ALLOC() -	allocate a new OBJT_SWAP VM object and instantiate
709  *			its metadata structures.
710  *
711  *	This routine is called from the mmap and fork code to create a new
712  *	OBJT_SWAP object.
713  *
714  *	This routine must ensure that no live duplicate is created for
715  *	the named object request, which is protected against by
716  *	holding the sw_alloc_sx lock in case handle != NULL.
717  */
718 static vm_object_t
719 swap_pager_alloc(void *handle, vm_ooffset_t size, vm_prot_t prot,
720     vm_ooffset_t offset, struct ucred *cred)
721 {
722 	vm_object_t object;
723 
724 	if (handle != NULL) {
725 		/*
726 		 * Reference existing named region or allocate new one.  There
727 		 * should not be a race here against swp_pager_meta_build()
728 		 * as called from vm_page_remove() in regards to the lookup
729 		 * of the handle.
730 		 */
731 		sx_xlock(&sw_alloc_sx);
732 		object = vm_pager_object_lookup(NOBJLIST(handle), handle);
733 		if (object == NULL) {
734 			object = swap_pager_alloc_init(OBJT_SWAP, handle, cred,
735 			    size, offset);
736 			if (object != NULL) {
737 				TAILQ_INSERT_TAIL(NOBJLIST(object->handle),
738 				    object, pager_object_list);
739 			}
740 		}
741 		sx_xunlock(&sw_alloc_sx);
742 	} else {
743 		object = swap_pager_alloc_init(OBJT_SWAP, handle, cred,
744 		    size, offset);
745 	}
746 	return (object);
747 }
748 
749 /*
750  * SWAP_PAGER_DEALLOC() -	remove swap metadata from object
751  *
752  *	The swap backing for the object is destroyed.  The code is
753  *	designed such that we can reinstantiate it later, but this
754  *	routine is typically called only when the entire object is
755  *	about to be destroyed.
756  *
757  *	The object must be locked.
758  */
759 static void
760 swap_pager_dealloc(vm_object_t object)
761 {
762 
763 	VM_OBJECT_ASSERT_WLOCKED(object);
764 	KASSERT((object->flags & OBJ_DEAD) != 0, ("dealloc of reachable obj"));
765 
766 	/*
767 	 * Remove from list right away so lookups will fail if we block for
768 	 * pageout completion.
769 	 */
770 	if ((object->flags & OBJ_ANON) == 0 && object->handle != NULL) {
771 		VM_OBJECT_WUNLOCK(object);
772 		sx_xlock(&sw_alloc_sx);
773 		TAILQ_REMOVE(NOBJLIST(object->handle), object,
774 		    pager_object_list);
775 		sx_xunlock(&sw_alloc_sx);
776 		VM_OBJECT_WLOCK(object);
777 	}
778 
779 	vm_object_pip_wait(object, "swpdea");
780 
781 	/*
782 	 * Free all remaining metadata.  We only bother to free it from
783 	 * the swap meta data.  We do not attempt to free swapblk's still
784 	 * associated with vm_page_t's for this object.  We do not care
785 	 * if paging is still in progress on some objects.
786 	 */
787 	swp_pager_meta_free_all(object);
788 	object->handle = NULL;
789 	object->type = OBJT_DEAD;
790 	vm_object_clear_flag(object, OBJ_SWAP);
791 }
792 
793 /************************************************************************
794  *			SWAP PAGER BITMAP ROUTINES			*
795  ************************************************************************/
796 
797 /*
798  * SWP_PAGER_GETSWAPSPACE() -	allocate raw swap space
799  *
800  *	Allocate swap for up to the requested number of pages.  The
801  *	starting swap block number (a page index) is returned or
802  *	SWAPBLK_NONE if the allocation failed.
803  *
804  *	Also has the side effect of advising that somebody made a mistake
805  *	when they configured swap and didn't configure enough.
806  *
807  *	This routine may not sleep.
808  *
809  *	We allocate in round-robin fashion from the configured devices.
810  */
811 static daddr_t
812 swp_pager_getswapspace(int *io_npages)
813 {
814 	daddr_t blk;
815 	struct swdevt *sp;
816 	int mpages, npages;
817 
818 	KASSERT(*io_npages >= 1,
819 	    ("%s: npages not positive", __func__));
820 	blk = SWAPBLK_NONE;
821 	mpages = *io_npages;
822 	npages = imin(BLIST_MAX_ALLOC, mpages);
823 	mtx_lock(&sw_dev_mtx);
824 	sp = swdevhd;
825 	while (!TAILQ_EMPTY(&swtailq)) {
826 		if (sp == NULL)
827 			sp = TAILQ_FIRST(&swtailq);
828 		if ((sp->sw_flags & SW_CLOSING) == 0)
829 			blk = blist_alloc(sp->sw_blist, &npages, mpages);
830 		if (blk != SWAPBLK_NONE)
831 			break;
832 		sp = TAILQ_NEXT(sp, sw_list);
833 		if (swdevhd == sp) {
834 			if (npages == 1)
835 				break;
836 			mpages = npages - 1;
837 			npages >>= 1;
838 		}
839 	}
840 	if (blk != SWAPBLK_NONE) {
841 		*io_npages = npages;
842 		blk += sp->sw_first;
843 		sp->sw_used += npages;
844 		swap_pager_avail -= npages;
845 		swp_sizecheck();
846 		swdevhd = TAILQ_NEXT(sp, sw_list);
847 	} else {
848 		if (swap_pager_full != 2) {
849 			printf("swp_pager_getswapspace(%d): failed\n",
850 			    *io_npages);
851 			swap_pager_full = 2;
852 			swap_pager_almost_full = 1;
853 		}
854 		swdevhd = NULL;
855 	}
856 	mtx_unlock(&sw_dev_mtx);
857 	return (blk);
858 }
859 
860 static bool
861 swp_pager_isondev(daddr_t blk, struct swdevt *sp)
862 {
863 
864 	return (blk >= sp->sw_first && blk < sp->sw_end);
865 }
866 
867 static void
868 swp_pager_strategy(struct buf *bp)
869 {
870 	struct swdevt *sp;
871 
872 	mtx_lock(&sw_dev_mtx);
873 	TAILQ_FOREACH(sp, &swtailq, sw_list) {
874 		if (swp_pager_isondev(bp->b_blkno, sp)) {
875 			mtx_unlock(&sw_dev_mtx);
876 			if ((sp->sw_flags & SW_UNMAPPED) != 0 &&
877 			    unmapped_buf_allowed) {
878 				bp->b_data = unmapped_buf;
879 				bp->b_offset = 0;
880 			} else {
881 				pmap_qenter((vm_offset_t)bp->b_data,
882 				    &bp->b_pages[0], bp->b_bcount / PAGE_SIZE);
883 			}
884 			sp->sw_strategy(bp, sp);
885 			return;
886 		}
887 	}
888 	panic("Swapdev not found");
889 }
890 
891 /*
892  * SWP_PAGER_FREESWAPSPACE() -	free raw swap space
893  *
894  *	This routine returns the specified swap blocks back to the bitmap.
895  *
896  *	This routine may not sleep.
897  */
898 static void
899 swp_pager_freeswapspace(daddr_t blk, daddr_t npages)
900 {
901 	struct swdevt *sp;
902 
903 	if (npages == 0)
904 		return;
905 	mtx_lock(&sw_dev_mtx);
906 	TAILQ_FOREACH(sp, &swtailq, sw_list) {
907 		if (swp_pager_isondev(blk, sp)) {
908 			sp->sw_used -= npages;
909 			/*
910 			 * If we are attempting to stop swapping on
911 			 * this device, we don't want to mark any
912 			 * blocks free lest they be reused.
913 			 */
914 			if ((sp->sw_flags & SW_CLOSING) == 0) {
915 				blist_free(sp->sw_blist, blk - sp->sw_first,
916 				    npages);
917 				swap_pager_avail += npages;
918 				swp_sizecheck();
919 			}
920 			mtx_unlock(&sw_dev_mtx);
921 			return;
922 		}
923 	}
924 	panic("Swapdev not found");
925 }
926 
927 /*
928  * SYSCTL_SWAP_FRAGMENTATION() -	produce raw swap space stats
929  */
930 static int
931 sysctl_swap_fragmentation(SYSCTL_HANDLER_ARGS)
932 {
933 	struct sbuf sbuf;
934 	struct swdevt *sp;
935 	const char *devname;
936 	int error;
937 
938 	error = sysctl_wire_old_buffer(req, 0);
939 	if (error != 0)
940 		return (error);
941 	sbuf_new_for_sysctl(&sbuf, NULL, 128, req);
942 	mtx_lock(&sw_dev_mtx);
943 	TAILQ_FOREACH(sp, &swtailq, sw_list) {
944 		if (vn_isdisk(sp->sw_vp))
945 			devname = devtoname(sp->sw_vp->v_rdev);
946 		else
947 			devname = "[file]";
948 		sbuf_printf(&sbuf, "\nFree space on device %s:\n", devname);
949 		blist_stats(sp->sw_blist, &sbuf);
950 	}
951 	mtx_unlock(&sw_dev_mtx);
952 	error = sbuf_finish(&sbuf);
953 	sbuf_delete(&sbuf);
954 	return (error);
955 }
956 
957 /*
958  * SWAP_PAGER_FREESPACE() -	frees swap blocks associated with a page
959  *				range within an object.
960  *
961  *	This routine removes swapblk assignments from swap metadata.
962  *
963  *	The external callers of this routine typically have already destroyed
964  *	or renamed vm_page_t's associated with this range in the object so
965  *	we should be ok.
966  *
967  *	The object must be locked.
968  */
969 static void
970 swap_pager_freespace(vm_object_t object, vm_pindex_t start, vm_size_t size)
971 {
972 
973 	swp_pager_meta_free(object, start, size);
974 }
975 
976 /*
977  * SWAP_PAGER_RESERVE() - reserve swap blocks in object
978  *
979  *	Assigns swap blocks to the specified range within the object.  The
980  *	swap blocks are not zeroed.  Any previous swap assignment is destroyed.
981  *
982  *	Returns 0 on success, -1 on failure.
983  */
984 int
985 swap_pager_reserve(vm_object_t object, vm_pindex_t start, vm_pindex_t size)
986 {
987 	daddr_t addr, blk, n_free, s_free;
988 	vm_pindex_t i, j;
989 	int n;
990 
991 	swp_pager_init_freerange(&s_free, &n_free);
992 	VM_OBJECT_WLOCK(object);
993 	for (i = 0; i < size; i += n) {
994 		n = MIN(size - i, INT_MAX);
995 		blk = swp_pager_getswapspace(&n);
996 		if (blk == SWAPBLK_NONE) {
997 			swp_pager_meta_free(object, start, i);
998 			VM_OBJECT_WUNLOCK(object);
999 			return (-1);
1000 		}
1001 		for (j = 0; j < n; ++j) {
1002 			addr = swp_pager_meta_build(object,
1003 			    start + i + j, blk + j);
1004 			if (addr != SWAPBLK_NONE)
1005 				swp_pager_update_freerange(&s_free, &n_free,
1006 				    addr);
1007 		}
1008 	}
1009 	swp_pager_freeswapspace(s_free, n_free);
1010 	VM_OBJECT_WUNLOCK(object);
1011 	return (0);
1012 }
1013 
1014 static bool
1015 swp_pager_xfer_source(vm_object_t srcobject, vm_object_t dstobject,
1016     vm_pindex_t pindex, daddr_t addr)
1017 {
1018 	daddr_t dstaddr;
1019 
1020 	KASSERT((srcobject->flags & OBJ_SWAP) != 0,
1021 	    ("%s: Srcobject not swappable", __func__));
1022 	if ((dstobject->flags & OBJ_SWAP) != 0 &&
1023 	    swp_pager_meta_lookup(dstobject, pindex) != SWAPBLK_NONE) {
1024 		/* Caller should destroy the source block. */
1025 		return (false);
1026 	}
1027 
1028 	/*
1029 	 * Destination has no swapblk and is not resident, transfer source.
1030 	 * swp_pager_meta_build() can sleep.
1031 	 */
1032 	VM_OBJECT_WUNLOCK(srcobject);
1033 	dstaddr = swp_pager_meta_build(dstobject, pindex, addr);
1034 	KASSERT(dstaddr == SWAPBLK_NONE,
1035 	    ("Unexpected destination swapblk"));
1036 	VM_OBJECT_WLOCK(srcobject);
1037 
1038 	return (true);
1039 }
1040 
1041 /*
1042  * SWAP_PAGER_COPY() -  copy blocks from source pager to destination pager
1043  *			and destroy the source.
1044  *
1045  *	Copy any valid swapblks from the source to the destination.  In
1046  *	cases where both the source and destination have a valid swapblk,
1047  *	we keep the destination's.
1048  *
1049  *	This routine is allowed to sleep.  It may sleep allocating metadata
1050  *	indirectly through swp_pager_meta_build().
1051  *
1052  *	The source object contains no vm_page_t's (which is just as well)
1053  *
1054  *	The source object is of type OBJT_SWAP.
1055  *
1056  *	The source and destination objects must be locked.
1057  *	Both object locks may temporarily be released.
1058  */
1059 void
1060 swap_pager_copy(vm_object_t srcobject, vm_object_t dstobject,
1061     vm_pindex_t offset, int destroysource)
1062 {
1063 
1064 	VM_OBJECT_ASSERT_WLOCKED(srcobject);
1065 	VM_OBJECT_ASSERT_WLOCKED(dstobject);
1066 
1067 	/*
1068 	 * If destroysource is set, we remove the source object from the
1069 	 * swap_pager internal queue now.
1070 	 */
1071 	if (destroysource && (srcobject->flags & OBJ_ANON) == 0 &&
1072 	    srcobject->handle != NULL) {
1073 		VM_OBJECT_WUNLOCK(srcobject);
1074 		VM_OBJECT_WUNLOCK(dstobject);
1075 		sx_xlock(&sw_alloc_sx);
1076 		TAILQ_REMOVE(NOBJLIST(srcobject->handle), srcobject,
1077 		    pager_object_list);
1078 		sx_xunlock(&sw_alloc_sx);
1079 		VM_OBJECT_WLOCK(dstobject);
1080 		VM_OBJECT_WLOCK(srcobject);
1081 	}
1082 
1083 	/*
1084 	 * Transfer source to destination.
1085 	 */
1086 	swp_pager_meta_transfer(srcobject, dstobject, offset, dstobject->size);
1087 
1088 	/*
1089 	 * Free left over swap blocks in source.
1090 	 *
1091 	 * We have to revert the type to OBJT_DEFAULT so we do not accidentally
1092 	 * double-remove the object from the swap queues.
1093 	 */
1094 	if (destroysource) {
1095 		swp_pager_meta_free_all(srcobject);
1096 		/*
1097 		 * Reverting the type is not necessary, the caller is going
1098 		 * to destroy srcobject directly, but I'm doing it here
1099 		 * for consistency since we've removed the object from its
1100 		 * queues.
1101 		 */
1102 		srcobject->type = OBJT_DEFAULT;
1103 		vm_object_clear_flag(srcobject, OBJ_SWAP);
1104 	}
1105 }
1106 
1107 /*
1108  * SWAP_PAGER_HASPAGE() -	determine if we have good backing store for
1109  *				the requested page.
1110  *
1111  *	We determine whether good backing store exists for the requested
1112  *	page and return TRUE if it does, FALSE if it doesn't.
1113  *
1114  *	If TRUE, we also try to determine how much valid, contiguous backing
1115  *	store exists before and after the requested page.
1116  */
1117 static boolean_t
1118 swap_pager_haspage(vm_object_t object, vm_pindex_t pindex, int *before,
1119     int *after)
1120 {
1121 	daddr_t blk, blk0;
1122 	int i;
1123 
1124 	VM_OBJECT_ASSERT_LOCKED(object);
1125 	KASSERT((object->flags & OBJ_SWAP) != 0,
1126 	    ("%s: object not swappable", __func__));
1127 
1128 	/*
1129 	 * do we have good backing store at the requested index ?
1130 	 */
1131 	blk0 = swp_pager_meta_lookup(object, pindex);
1132 	if (blk0 == SWAPBLK_NONE) {
1133 		if (before)
1134 			*before = 0;
1135 		if (after)
1136 			*after = 0;
1137 		return (FALSE);
1138 	}
1139 
1140 	/*
1141 	 * find backwards-looking contiguous good backing store
1142 	 */
1143 	if (before != NULL) {
1144 		for (i = 1; i < SWB_NPAGES; i++) {
1145 			if (i > pindex)
1146 				break;
1147 			blk = swp_pager_meta_lookup(object, pindex - i);
1148 			if (blk != blk0 - i)
1149 				break;
1150 		}
1151 		*before = i - 1;
1152 	}
1153 
1154 	/*
1155 	 * find forward-looking contiguous good backing store
1156 	 */
1157 	if (after != NULL) {
1158 		for (i = 1; i < SWB_NPAGES; i++) {
1159 			blk = swp_pager_meta_lookup(object, pindex + i);
1160 			if (blk != blk0 + i)
1161 				break;
1162 		}
1163 		*after = i - 1;
1164 	}
1165 	return (TRUE);
1166 }
1167 
1168 /*
1169  * SWAP_PAGER_PAGE_UNSWAPPED() - remove swap backing store related to page
1170  *
1171  *	This removes any associated swap backing store, whether valid or
1172  *	not, from the page.
1173  *
1174  *	This routine is typically called when a page is made dirty, at
1175  *	which point any associated swap can be freed.  MADV_FREE also
1176  *	calls us in a special-case situation
1177  *
1178  *	NOTE!!!  If the page is clean and the swap was valid, the caller
1179  *	should make the page dirty before calling this routine.  This routine
1180  *	does NOT change the m->dirty status of the page.  Also: MADV_FREE
1181  *	depends on it.
1182  *
1183  *	This routine may not sleep.
1184  *
1185  *	The object containing the page may be locked.
1186  */
1187 static void
1188 swap_pager_unswapped(vm_page_t m)
1189 {
1190 	struct swblk *sb;
1191 	vm_object_t obj;
1192 
1193 	/*
1194 	 * Handle enqueing deferred frees first.  If we do not have the
1195 	 * object lock we wait for the page daemon to clear the space.
1196 	 */
1197 	obj = m->object;
1198 	if (!VM_OBJECT_WOWNED(obj)) {
1199 		VM_PAGE_OBJECT_BUSY_ASSERT(m);
1200 		/*
1201 		 * The caller is responsible for synchronization but we
1202 		 * will harmlessly handle races.  This is typically provided
1203 		 * by only calling unswapped() when a page transitions from
1204 		 * clean to dirty.
1205 		 */
1206 		if ((m->a.flags & (PGA_SWAP_SPACE | PGA_SWAP_FREE)) ==
1207 		    PGA_SWAP_SPACE) {
1208 			vm_page_aflag_set(m, PGA_SWAP_FREE);
1209 			counter_u64_add(swap_free_deferred, 1);
1210 		}
1211 		return;
1212 	}
1213 	if ((m->a.flags & PGA_SWAP_FREE) != 0)
1214 		counter_u64_add(swap_free_completed, 1);
1215 	vm_page_aflag_clear(m, PGA_SWAP_FREE | PGA_SWAP_SPACE);
1216 
1217 	/*
1218 	 * The meta data only exists if the object is OBJT_SWAP
1219 	 * and even then might not be allocated yet.
1220 	 */
1221 	KASSERT((m->object->flags & OBJ_SWAP) != 0,
1222 	    ("Free object not swappable"));
1223 
1224 	sb = SWAP_PCTRIE_LOOKUP(&m->object->un_pager.swp.swp_blks,
1225 	    rounddown(m->pindex, SWAP_META_PAGES));
1226 	if (sb == NULL)
1227 		return;
1228 	if (sb->d[m->pindex % SWAP_META_PAGES] == SWAPBLK_NONE)
1229 		return;
1230 	swp_pager_freeswapspace(sb->d[m->pindex % SWAP_META_PAGES], 1);
1231 	sb->d[m->pindex % SWAP_META_PAGES] = SWAPBLK_NONE;
1232 	swp_pager_free_empty_swblk(m->object, sb);
1233 }
1234 
1235 /*
1236  * swap_pager_getpages() - bring pages in from swap
1237  *
1238  *	Attempt to page in the pages in array "ma" of length "count".  The
1239  *	caller may optionally specify that additional pages preceding and
1240  *	succeeding the specified range be paged in.  The number of such pages
1241  *	is returned in the "rbehind" and "rahead" parameters, and they will
1242  *	be in the inactive queue upon return.
1243  *
1244  *	The pages in "ma" must be busied and will remain busied upon return.
1245  */
1246 static int
1247 swap_pager_getpages_locked(vm_object_t object, vm_page_t *ma, int count,
1248     int *rbehind, int *rahead)
1249 {
1250 	struct buf *bp;
1251 	vm_page_t bm, mpred, msucc, p;
1252 	vm_pindex_t pindex;
1253 	daddr_t blk;
1254 	int i, maxahead, maxbehind, reqcount;
1255 
1256 	VM_OBJECT_ASSERT_WLOCKED(object);
1257 	reqcount = count;
1258 
1259 	KASSERT((object->flags & OBJ_SWAP) != 0,
1260 	    ("%s: object not swappable", __func__));
1261 	if (!swap_pager_haspage(object, ma[0]->pindex, &maxbehind, &maxahead)) {
1262 		VM_OBJECT_WUNLOCK(object);
1263 		return (VM_PAGER_FAIL);
1264 	}
1265 
1266 	KASSERT(reqcount - 1 <= maxahead,
1267 	    ("page count %d extends beyond swap block", reqcount));
1268 
1269 	/*
1270 	 * Do not transfer any pages other than those that are xbusied
1271 	 * when running during a split or collapse operation.  This
1272 	 * prevents clustering from re-creating pages which are being
1273 	 * moved into another object.
1274 	 */
1275 	if ((object->flags & (OBJ_SPLIT | OBJ_DEAD)) != 0) {
1276 		maxahead = reqcount - 1;
1277 		maxbehind = 0;
1278 	}
1279 
1280 	/*
1281 	 * Clip the readahead and readbehind ranges to exclude resident pages.
1282 	 */
1283 	if (rahead != NULL) {
1284 		*rahead = imin(*rahead, maxahead - (reqcount - 1));
1285 		pindex = ma[reqcount - 1]->pindex;
1286 		msucc = TAILQ_NEXT(ma[reqcount - 1], listq);
1287 		if (msucc != NULL && msucc->pindex - pindex - 1 < *rahead)
1288 			*rahead = msucc->pindex - pindex - 1;
1289 	}
1290 	if (rbehind != NULL) {
1291 		*rbehind = imin(*rbehind, maxbehind);
1292 		pindex = ma[0]->pindex;
1293 		mpred = TAILQ_PREV(ma[0], pglist, listq);
1294 		if (mpred != NULL && pindex - mpred->pindex - 1 < *rbehind)
1295 			*rbehind = pindex - mpred->pindex - 1;
1296 	}
1297 
1298 	bm = ma[0];
1299 	for (i = 0; i < count; i++)
1300 		ma[i]->oflags |= VPO_SWAPINPROG;
1301 
1302 	/*
1303 	 * Allocate readahead and readbehind pages.
1304 	 */
1305 	if (rbehind != NULL) {
1306 		for (i = 1; i <= *rbehind; i++) {
1307 			p = vm_page_alloc(object, ma[0]->pindex - i,
1308 			    VM_ALLOC_NORMAL);
1309 			if (p == NULL)
1310 				break;
1311 			p->oflags |= VPO_SWAPINPROG;
1312 			bm = p;
1313 		}
1314 		*rbehind = i - 1;
1315 	}
1316 	if (rahead != NULL) {
1317 		for (i = 0; i < *rahead; i++) {
1318 			p = vm_page_alloc(object,
1319 			    ma[reqcount - 1]->pindex + i + 1, VM_ALLOC_NORMAL);
1320 			if (p == NULL)
1321 				break;
1322 			p->oflags |= VPO_SWAPINPROG;
1323 		}
1324 		*rahead = i;
1325 	}
1326 	if (rbehind != NULL)
1327 		count += *rbehind;
1328 	if (rahead != NULL)
1329 		count += *rahead;
1330 
1331 	vm_object_pip_add(object, count);
1332 
1333 	pindex = bm->pindex;
1334 	blk = swp_pager_meta_lookup(object, pindex);
1335 	KASSERT(blk != SWAPBLK_NONE,
1336 	    ("no swap blocking containing %p(%jx)", object, (uintmax_t)pindex));
1337 
1338 	VM_OBJECT_WUNLOCK(object);
1339 	bp = uma_zalloc(swrbuf_zone, M_WAITOK);
1340 	MPASS((bp->b_flags & B_MAXPHYS) != 0);
1341 	/* Pages cannot leave the object while busy. */
1342 	for (i = 0, p = bm; i < count; i++, p = TAILQ_NEXT(p, listq)) {
1343 		MPASS(p->pindex == bm->pindex + i);
1344 		bp->b_pages[i] = p;
1345 	}
1346 
1347 	bp->b_flags |= B_PAGING;
1348 	bp->b_iocmd = BIO_READ;
1349 	bp->b_iodone = swp_pager_async_iodone;
1350 	bp->b_rcred = crhold(thread0.td_ucred);
1351 	bp->b_wcred = crhold(thread0.td_ucred);
1352 	bp->b_blkno = blk;
1353 	bp->b_bcount = PAGE_SIZE * count;
1354 	bp->b_bufsize = PAGE_SIZE * count;
1355 	bp->b_npages = count;
1356 	bp->b_pgbefore = rbehind != NULL ? *rbehind : 0;
1357 	bp->b_pgafter = rahead != NULL ? *rahead : 0;
1358 
1359 	VM_CNT_INC(v_swapin);
1360 	VM_CNT_ADD(v_swappgsin, count);
1361 
1362 	/*
1363 	 * perform the I/O.  NOTE!!!  bp cannot be considered valid after
1364 	 * this point because we automatically release it on completion.
1365 	 * Instead, we look at the one page we are interested in which we
1366 	 * still hold a lock on even through the I/O completion.
1367 	 *
1368 	 * The other pages in our ma[] array are also released on completion,
1369 	 * so we cannot assume they are valid anymore either.
1370 	 *
1371 	 * NOTE: b_blkno is destroyed by the call to swapdev_strategy
1372 	 */
1373 	BUF_KERNPROC(bp);
1374 	swp_pager_strategy(bp);
1375 
1376 	/*
1377 	 * Wait for the pages we want to complete.  VPO_SWAPINPROG is always
1378 	 * cleared on completion.  If an I/O error occurs, SWAPBLK_NONE
1379 	 * is set in the metadata for each page in the request.
1380 	 */
1381 	VM_OBJECT_WLOCK(object);
1382 	/* This could be implemented more efficiently with aflags */
1383 	while ((ma[0]->oflags & VPO_SWAPINPROG) != 0) {
1384 		ma[0]->oflags |= VPO_SWAPSLEEP;
1385 		VM_CNT_INC(v_intrans);
1386 		if (VM_OBJECT_SLEEP(object, &object->handle, PSWP,
1387 		    "swread", hz * 20)) {
1388 			printf(
1389 "swap_pager: indefinite wait buffer: bufobj: %p, blkno: %jd, size: %ld\n",
1390 			    bp->b_bufobj, (intmax_t)bp->b_blkno, bp->b_bcount);
1391 		}
1392 	}
1393 	VM_OBJECT_WUNLOCK(object);
1394 
1395 	/*
1396 	 * If we had an unrecoverable read error pages will not be valid.
1397 	 */
1398 	for (i = 0; i < reqcount; i++)
1399 		if (ma[i]->valid != VM_PAGE_BITS_ALL)
1400 			return (VM_PAGER_ERROR);
1401 
1402 	return (VM_PAGER_OK);
1403 
1404 	/*
1405 	 * A final note: in a low swap situation, we cannot deallocate swap
1406 	 * and mark a page dirty here because the caller is likely to mark
1407 	 * the page clean when we return, causing the page to possibly revert
1408 	 * to all-zero's later.
1409 	 */
1410 }
1411 
1412 static int
1413 swap_pager_getpages(vm_object_t object, vm_page_t *ma, int count,
1414     int *rbehind, int *rahead)
1415 {
1416 
1417 	VM_OBJECT_WLOCK(object);
1418 	return (swap_pager_getpages_locked(object, ma, count, rbehind, rahead));
1419 }
1420 
1421 /*
1422  * 	swap_pager_getpages_async():
1423  *
1424  *	Right now this is emulation of asynchronous operation on top of
1425  *	swap_pager_getpages().
1426  */
1427 static int
1428 swap_pager_getpages_async(vm_object_t object, vm_page_t *ma, int count,
1429     int *rbehind, int *rahead, pgo_getpages_iodone_t iodone, void *arg)
1430 {
1431 	int r, error;
1432 
1433 	r = swap_pager_getpages(object, ma, count, rbehind, rahead);
1434 	switch (r) {
1435 	case VM_PAGER_OK:
1436 		error = 0;
1437 		break;
1438 	case VM_PAGER_ERROR:
1439 		error = EIO;
1440 		break;
1441 	case VM_PAGER_FAIL:
1442 		error = EINVAL;
1443 		break;
1444 	default:
1445 		panic("unhandled swap_pager_getpages() error %d", r);
1446 	}
1447 	(iodone)(arg, ma, count, error);
1448 
1449 	return (r);
1450 }
1451 
1452 /*
1453  *	swap_pager_putpages:
1454  *
1455  *	Assign swap (if necessary) and initiate I/O on the specified pages.
1456  *
1457  *	We support both OBJT_DEFAULT and OBJT_SWAP objects.  DEFAULT objects
1458  *	are automatically converted to SWAP objects.
1459  *
1460  *	In a low memory situation we may block in VOP_STRATEGY(), but the new
1461  *	vm_page reservation system coupled with properly written VFS devices
1462  *	should ensure that no low-memory deadlock occurs.  This is an area
1463  *	which needs work.
1464  *
1465  *	The parent has N vm_object_pip_add() references prior to
1466  *	calling us and will remove references for rtvals[] that are
1467  *	not set to VM_PAGER_PEND.  We need to remove the rest on I/O
1468  *	completion.
1469  *
1470  *	The parent has soft-busy'd the pages it passes us and will unbusy
1471  *	those whose rtvals[] entry is not set to VM_PAGER_PEND on return.
1472  *	We need to unbusy the rest on I/O completion.
1473  */
1474 static void
1475 swap_pager_putpages(vm_object_t object, vm_page_t *ma, int count,
1476     int flags, int *rtvals)
1477 {
1478 	struct buf *bp;
1479 	daddr_t addr, blk, n_free, s_free;
1480 	vm_page_t mreq;
1481 	int i, j, n;
1482 	bool async;
1483 
1484 	KASSERT(count == 0 || ma[0]->object == object,
1485 	    ("%s: object mismatch %p/%p",
1486 	    __func__, object, ma[0]->object));
1487 
1488 	/*
1489 	 * Step 1
1490 	 *
1491 	 * Turn object into OBJT_SWAP.  Force sync if not a pageout process.
1492 	 */
1493 	if ((object->flags & OBJ_SWAP) == 0) {
1494 		addr = swp_pager_meta_build(object, 0, SWAPBLK_NONE);
1495 		KASSERT(addr == SWAPBLK_NONE,
1496 		    ("unexpected object swap block"));
1497 	}
1498 	VM_OBJECT_WUNLOCK(object);
1499 	async = curproc == pageproc && (flags & VM_PAGER_PUT_SYNC) == 0;
1500 	swp_pager_init_freerange(&s_free, &n_free);
1501 
1502 	/*
1503 	 * Step 2
1504 	 *
1505 	 * Assign swap blocks and issue I/O.  We reallocate swap on the fly.
1506 	 * The page is left dirty until the pageout operation completes
1507 	 * successfully.
1508 	 */
1509 	for (i = 0; i < count; i += n) {
1510 		/* Maximum I/O size is limited by maximum swap block size. */
1511 		n = min(count - i, nsw_cluster_max);
1512 
1513 		if (async) {
1514 			mtx_lock(&swbuf_mtx);
1515 			while (nsw_wcount_async == 0)
1516 				msleep(&nsw_wcount_async, &swbuf_mtx, PVM,
1517 				    "swbufa", 0);
1518 			nsw_wcount_async--;
1519 			mtx_unlock(&swbuf_mtx);
1520 		}
1521 
1522 		/* Get a block of swap of size up to size n. */
1523 		VM_OBJECT_WLOCK(object);
1524 		blk = swp_pager_getswapspace(&n);
1525 		if (blk == SWAPBLK_NONE) {
1526 			VM_OBJECT_WUNLOCK(object);
1527 			mtx_lock(&swbuf_mtx);
1528 			if (++nsw_wcount_async == 1)
1529 				wakeup(&nsw_wcount_async);
1530 			mtx_unlock(&swbuf_mtx);
1531 			for (j = 0; j < n; ++j)
1532 				rtvals[i + j] = VM_PAGER_FAIL;
1533 			continue;
1534 		}
1535 		for (j = 0; j < n; ++j) {
1536 			mreq = ma[i + j];
1537 			vm_page_aflag_clear(mreq, PGA_SWAP_FREE);
1538 			addr = swp_pager_meta_build(mreq->object, mreq->pindex,
1539 			    blk + j);
1540 			if (addr != SWAPBLK_NONE)
1541 				swp_pager_update_freerange(&s_free, &n_free,
1542 				    addr);
1543 			MPASS(mreq->dirty == VM_PAGE_BITS_ALL);
1544 			mreq->oflags |= VPO_SWAPINPROG;
1545 		}
1546 		VM_OBJECT_WUNLOCK(object);
1547 
1548 		bp = uma_zalloc(swwbuf_zone, M_WAITOK);
1549 		MPASS((bp->b_flags & B_MAXPHYS) != 0);
1550 		if (async)
1551 			bp->b_flags |= B_ASYNC;
1552 		bp->b_flags |= B_PAGING;
1553 		bp->b_iocmd = BIO_WRITE;
1554 
1555 		bp->b_rcred = crhold(thread0.td_ucred);
1556 		bp->b_wcred = crhold(thread0.td_ucred);
1557 		bp->b_bcount = PAGE_SIZE * n;
1558 		bp->b_bufsize = PAGE_SIZE * n;
1559 		bp->b_blkno = blk;
1560 		for (j = 0; j < n; j++)
1561 			bp->b_pages[j] = ma[i + j];
1562 		bp->b_npages = n;
1563 
1564 		/*
1565 		 * Must set dirty range for NFS to work.
1566 		 */
1567 		bp->b_dirtyoff = 0;
1568 		bp->b_dirtyend = bp->b_bcount;
1569 
1570 		VM_CNT_INC(v_swapout);
1571 		VM_CNT_ADD(v_swappgsout, bp->b_npages);
1572 
1573 		/*
1574 		 * We unconditionally set rtvals[] to VM_PAGER_PEND so that we
1575 		 * can call the async completion routine at the end of a
1576 		 * synchronous I/O operation.  Otherwise, our caller would
1577 		 * perform duplicate unbusy and wakeup operations on the page
1578 		 * and object, respectively.
1579 		 */
1580 		for (j = 0; j < n; j++)
1581 			rtvals[i + j] = VM_PAGER_PEND;
1582 
1583 		/*
1584 		 * asynchronous
1585 		 *
1586 		 * NOTE: b_blkno is destroyed by the call to swapdev_strategy.
1587 		 */
1588 		if (async) {
1589 			bp->b_iodone = swp_pager_async_iodone;
1590 			BUF_KERNPROC(bp);
1591 			swp_pager_strategy(bp);
1592 			continue;
1593 		}
1594 
1595 		/*
1596 		 * synchronous
1597 		 *
1598 		 * NOTE: b_blkno is destroyed by the call to swapdev_strategy.
1599 		 */
1600 		bp->b_iodone = bdone;
1601 		swp_pager_strategy(bp);
1602 
1603 		/*
1604 		 * Wait for the sync I/O to complete.
1605 		 */
1606 		bwait(bp, PVM, "swwrt");
1607 
1608 		/*
1609 		 * Now that we are through with the bp, we can call the
1610 		 * normal async completion, which frees everything up.
1611 		 */
1612 		swp_pager_async_iodone(bp);
1613 	}
1614 	swp_pager_freeswapspace(s_free, n_free);
1615 	VM_OBJECT_WLOCK(object);
1616 }
1617 
1618 /*
1619  *	swp_pager_async_iodone:
1620  *
1621  *	Completion routine for asynchronous reads and writes from/to swap.
1622  *	Also called manually by synchronous code to finish up a bp.
1623  *
1624  *	This routine may not sleep.
1625  */
1626 static void
1627 swp_pager_async_iodone(struct buf *bp)
1628 {
1629 	int i;
1630 	vm_object_t object = NULL;
1631 
1632 	/*
1633 	 * Report error - unless we ran out of memory, in which case
1634 	 * we've already logged it in swapgeom_strategy().
1635 	 */
1636 	if (bp->b_ioflags & BIO_ERROR && bp->b_error != ENOMEM) {
1637 		printf(
1638 		    "swap_pager: I/O error - %s failed; blkno %ld,"
1639 			"size %ld, error %d\n",
1640 		    ((bp->b_iocmd == BIO_READ) ? "pagein" : "pageout"),
1641 		    (long)bp->b_blkno,
1642 		    (long)bp->b_bcount,
1643 		    bp->b_error
1644 		);
1645 	}
1646 
1647 	/*
1648 	 * remove the mapping for kernel virtual
1649 	 */
1650 	if (buf_mapped(bp))
1651 		pmap_qremove((vm_offset_t)bp->b_data, bp->b_npages);
1652 	else
1653 		bp->b_data = bp->b_kvabase;
1654 
1655 	if (bp->b_npages) {
1656 		object = bp->b_pages[0]->object;
1657 		VM_OBJECT_WLOCK(object);
1658 	}
1659 
1660 	/*
1661 	 * cleanup pages.  If an error occurs writing to swap, we are in
1662 	 * very serious trouble.  If it happens to be a disk error, though,
1663 	 * we may be able to recover by reassigning the swap later on.  So
1664 	 * in this case we remove the m->swapblk assignment for the page
1665 	 * but do not free it in the rlist.  The errornous block(s) are thus
1666 	 * never reallocated as swap.  Redirty the page and continue.
1667 	 */
1668 	for (i = 0; i < bp->b_npages; ++i) {
1669 		vm_page_t m = bp->b_pages[i];
1670 
1671 		m->oflags &= ~VPO_SWAPINPROG;
1672 		if (m->oflags & VPO_SWAPSLEEP) {
1673 			m->oflags &= ~VPO_SWAPSLEEP;
1674 			wakeup(&object->handle);
1675 		}
1676 
1677 		/* We always have space after I/O, successful or not. */
1678 		vm_page_aflag_set(m, PGA_SWAP_SPACE);
1679 
1680 		if (bp->b_ioflags & BIO_ERROR) {
1681 			/*
1682 			 * If an error occurs I'd love to throw the swapblk
1683 			 * away without freeing it back to swapspace, so it
1684 			 * can never be used again.  But I can't from an
1685 			 * interrupt.
1686 			 */
1687 			if (bp->b_iocmd == BIO_READ) {
1688 				/*
1689 				 * NOTE: for reads, m->dirty will probably
1690 				 * be overridden by the original caller of
1691 				 * getpages so don't play cute tricks here.
1692 				 */
1693 				vm_page_invalid(m);
1694 			} else {
1695 				/*
1696 				 * If a write error occurs, reactivate page
1697 				 * so it doesn't clog the inactive list,
1698 				 * then finish the I/O.
1699 				 */
1700 				MPASS(m->dirty == VM_PAGE_BITS_ALL);
1701 
1702 				/* PQ_UNSWAPPABLE? */
1703 				vm_page_activate(m);
1704 				vm_page_sunbusy(m);
1705 			}
1706 		} else if (bp->b_iocmd == BIO_READ) {
1707 			/*
1708 			 * NOTE: for reads, m->dirty will probably be
1709 			 * overridden by the original caller of getpages so
1710 			 * we cannot set them in order to free the underlying
1711 			 * swap in a low-swap situation.  I don't think we'd
1712 			 * want to do that anyway, but it was an optimization
1713 			 * that existed in the old swapper for a time before
1714 			 * it got ripped out due to precisely this problem.
1715 			 */
1716 			KASSERT(!pmap_page_is_mapped(m),
1717 			    ("swp_pager_async_iodone: page %p is mapped", m));
1718 			KASSERT(m->dirty == 0,
1719 			    ("swp_pager_async_iodone: page %p is dirty", m));
1720 
1721 			vm_page_valid(m);
1722 			if (i < bp->b_pgbefore ||
1723 			    i >= bp->b_npages - bp->b_pgafter)
1724 				vm_page_readahead_finish(m);
1725 		} else {
1726 			/*
1727 			 * For write success, clear the dirty
1728 			 * status, then finish the I/O ( which decrements the
1729 			 * busy count and possibly wakes waiter's up ).
1730 			 * A page is only written to swap after a period of
1731 			 * inactivity.  Therefore, we do not expect it to be
1732 			 * reused.
1733 			 */
1734 			KASSERT(!pmap_page_is_write_mapped(m),
1735 			    ("swp_pager_async_iodone: page %p is not write"
1736 			    " protected", m));
1737 			vm_page_undirty(m);
1738 			vm_page_deactivate_noreuse(m);
1739 			vm_page_sunbusy(m);
1740 		}
1741 	}
1742 
1743 	/*
1744 	 * adjust pip.  NOTE: the original parent may still have its own
1745 	 * pip refs on the object.
1746 	 */
1747 	if (object != NULL) {
1748 		vm_object_pip_wakeupn(object, bp->b_npages);
1749 		VM_OBJECT_WUNLOCK(object);
1750 	}
1751 
1752 	/*
1753 	 * swapdev_strategy() manually sets b_vp and b_bufobj before calling
1754 	 * bstrategy(). Set them back to NULL now we're done with it, or we'll
1755 	 * trigger a KASSERT in relpbuf().
1756 	 */
1757 	if (bp->b_vp) {
1758 		    bp->b_vp = NULL;
1759 		    bp->b_bufobj = NULL;
1760 	}
1761 	/*
1762 	 * release the physical I/O buffer
1763 	 */
1764 	if (bp->b_flags & B_ASYNC) {
1765 		mtx_lock(&swbuf_mtx);
1766 		if (++nsw_wcount_async == 1)
1767 			wakeup(&nsw_wcount_async);
1768 		mtx_unlock(&swbuf_mtx);
1769 	}
1770 	uma_zfree((bp->b_iocmd == BIO_READ) ? swrbuf_zone : swwbuf_zone, bp);
1771 }
1772 
1773 int
1774 swap_pager_nswapdev(void)
1775 {
1776 
1777 	return (nswapdev);
1778 }
1779 
1780 static void
1781 swp_pager_force_dirty(vm_page_t m)
1782 {
1783 
1784 	vm_page_dirty(m);
1785 	swap_pager_unswapped(m);
1786 	vm_page_launder(m);
1787 }
1788 
1789 u_long
1790 swap_pager_swapped_pages(vm_object_t object)
1791 {
1792 	struct swblk *sb;
1793 	vm_pindex_t pi;
1794 	u_long res;
1795 	int i;
1796 
1797 	VM_OBJECT_ASSERT_LOCKED(object);
1798 	if ((object->flags & OBJ_SWAP) == 0)
1799 		return (0);
1800 
1801 	for (res = 0, pi = 0; (sb = SWAP_PCTRIE_LOOKUP_GE(
1802 	    &object->un_pager.swp.swp_blks, pi)) != NULL;
1803 	    pi = sb->p + SWAP_META_PAGES) {
1804 		for (i = 0; i < SWAP_META_PAGES; i++) {
1805 			if (sb->d[i] != SWAPBLK_NONE)
1806 				res++;
1807 		}
1808 	}
1809 	return (res);
1810 }
1811 
1812 /*
1813  *	swap_pager_swapoff_object:
1814  *
1815  *	Page in all of the pages that have been paged out for an object
1816  *	to a swap device.
1817  */
1818 static void
1819 swap_pager_swapoff_object(struct swdevt *sp, vm_object_t object)
1820 {
1821 	struct swblk *sb;
1822 	vm_page_t m;
1823 	vm_pindex_t pi;
1824 	daddr_t blk;
1825 	int i, nv, rahead, rv;
1826 
1827 	KASSERT((object->flags & OBJ_SWAP) != 0,
1828 	    ("%s: Object not swappable", __func__));
1829 
1830 	for (pi = 0; (sb = SWAP_PCTRIE_LOOKUP_GE(
1831 	    &object->un_pager.swp.swp_blks, pi)) != NULL; ) {
1832 		if ((object->flags & OBJ_DEAD) != 0) {
1833 			/*
1834 			 * Make sure that pending writes finish before
1835 			 * returning.
1836 			 */
1837 			vm_object_pip_wait(object, "swpoff");
1838 			swp_pager_meta_free_all(object);
1839 			break;
1840 		}
1841 		for (i = 0; i < SWAP_META_PAGES; i++) {
1842 			/*
1843 			 * Count the number of contiguous valid blocks.
1844 			 */
1845 			for (nv = 0; nv < SWAP_META_PAGES - i; nv++) {
1846 				blk = sb->d[i + nv];
1847 				if (!swp_pager_isondev(blk, sp) ||
1848 				    blk == SWAPBLK_NONE)
1849 					break;
1850 			}
1851 			if (nv == 0)
1852 				continue;
1853 
1854 			/*
1855 			 * Look for a page corresponding to the first
1856 			 * valid block and ensure that any pending paging
1857 			 * operations on it are complete.  If the page is valid,
1858 			 * mark it dirty and free the swap block.  Try to batch
1859 			 * this operation since it may cause sp to be freed,
1860 			 * meaning that we must restart the scan.  Avoid busying
1861 			 * valid pages since we may block forever on kernel
1862 			 * stack pages.
1863 			 */
1864 			m = vm_page_lookup(object, sb->p + i);
1865 			if (m == NULL) {
1866 				m = vm_page_alloc(object, sb->p + i,
1867 				    VM_ALLOC_NORMAL | VM_ALLOC_WAITFAIL);
1868 				if (m == NULL)
1869 					break;
1870 			} else {
1871 				if ((m->oflags & VPO_SWAPINPROG) != 0) {
1872 					m->oflags |= VPO_SWAPSLEEP;
1873 					VM_OBJECT_SLEEP(object, &object->handle,
1874 					    PSWP, "swpoff", 0);
1875 					break;
1876 				}
1877 				if (vm_page_all_valid(m)) {
1878 					do {
1879 						swp_pager_force_dirty(m);
1880 					} while (--nv > 0 &&
1881 					    (m = vm_page_next(m)) != NULL &&
1882 					    vm_page_all_valid(m) &&
1883 					    (m->oflags & VPO_SWAPINPROG) == 0);
1884 					break;
1885 				}
1886 				if (!vm_page_busy_acquire(m, VM_ALLOC_WAITFAIL))
1887 					break;
1888 			}
1889 
1890 			vm_object_pip_add(object, 1);
1891 			rahead = SWAP_META_PAGES;
1892 			rv = swap_pager_getpages_locked(object, &m, 1, NULL,
1893 			    &rahead);
1894 			if (rv != VM_PAGER_OK)
1895 				panic("%s: read from swap failed: %d",
1896 				    __func__, rv);
1897 			vm_object_pip_wakeupn(object, 1);
1898 			VM_OBJECT_WLOCK(object);
1899 			vm_page_xunbusy(m);
1900 
1901 			/*
1902 			 * The object lock was dropped so we must restart the
1903 			 * scan of this swap block.  Pages paged in during this
1904 			 * iteration will be marked dirty in a future iteration.
1905 			 */
1906 			break;
1907 		}
1908 		if (i == SWAP_META_PAGES)
1909 			pi = sb->p + SWAP_META_PAGES;
1910 	}
1911 }
1912 
1913 /*
1914  *	swap_pager_swapoff:
1915  *
1916  *	Page in all of the pages that have been paged out to the
1917  *	given device.  The corresponding blocks in the bitmap must be
1918  *	marked as allocated and the device must be flagged SW_CLOSING.
1919  *	There may be no processes swapped out to the device.
1920  *
1921  *	This routine may block.
1922  */
1923 static void
1924 swap_pager_swapoff(struct swdevt *sp)
1925 {
1926 	vm_object_t object;
1927 	int retries;
1928 
1929 	sx_assert(&swdev_syscall_lock, SA_XLOCKED);
1930 
1931 	retries = 0;
1932 full_rescan:
1933 	mtx_lock(&vm_object_list_mtx);
1934 	TAILQ_FOREACH(object, &vm_object_list, object_list) {
1935 		if ((object->flags & OBJ_SWAP) == 0)
1936 			continue;
1937 		mtx_unlock(&vm_object_list_mtx);
1938 		/* Depends on type-stability. */
1939 		VM_OBJECT_WLOCK(object);
1940 
1941 		/*
1942 		 * Dead objects are eventually terminated on their own.
1943 		 */
1944 		if ((object->flags & OBJ_DEAD) != 0)
1945 			goto next_obj;
1946 
1947 		/*
1948 		 * Sync with fences placed after pctrie
1949 		 * initialization.  We must not access pctrie below
1950 		 * unless we checked that our object is swap and not
1951 		 * dead.
1952 		 */
1953 		atomic_thread_fence_acq();
1954 		if ((object->flags & OBJ_SWAP) == 0)
1955 			goto next_obj;
1956 
1957 		swap_pager_swapoff_object(sp, object);
1958 next_obj:
1959 		VM_OBJECT_WUNLOCK(object);
1960 		mtx_lock(&vm_object_list_mtx);
1961 	}
1962 	mtx_unlock(&vm_object_list_mtx);
1963 
1964 	if (sp->sw_used) {
1965 		/*
1966 		 * Objects may be locked or paging to the device being
1967 		 * removed, so we will miss their pages and need to
1968 		 * make another pass.  We have marked this device as
1969 		 * SW_CLOSING, so the activity should finish soon.
1970 		 */
1971 		retries++;
1972 		if (retries > 100) {
1973 			panic("swapoff: failed to locate %d swap blocks",
1974 			    sp->sw_used);
1975 		}
1976 		pause("swpoff", hz / 20);
1977 		goto full_rescan;
1978 	}
1979 	EVENTHANDLER_INVOKE(swapoff, sp);
1980 }
1981 
1982 /************************************************************************
1983  *				SWAP META DATA 				*
1984  ************************************************************************
1985  *
1986  *	These routines manipulate the swap metadata stored in the
1987  *	OBJT_SWAP object.
1988  *
1989  *	Swap metadata is implemented with a global hash and not directly
1990  *	linked into the object.  Instead the object simply contains
1991  *	appropriate tracking counters.
1992  */
1993 
1994 /*
1995  * SWP_PAGER_SWBLK_EMPTY() - is a range of blocks free?
1996  */
1997 static bool
1998 swp_pager_swblk_empty(struct swblk *sb, int start, int limit)
1999 {
2000 	int i;
2001 
2002 	MPASS(0 <= start && start <= limit && limit <= SWAP_META_PAGES);
2003 	for (i = start; i < limit; i++) {
2004 		if (sb->d[i] != SWAPBLK_NONE)
2005 			return (false);
2006 	}
2007 	return (true);
2008 }
2009 
2010 /*
2011  * SWP_PAGER_FREE_EMPTY_SWBLK() - frees if a block is free
2012  *
2013  *  Nothing is done if the block is still in use.
2014  */
2015 static void
2016 swp_pager_free_empty_swblk(vm_object_t object, struct swblk *sb)
2017 {
2018 
2019 	if (swp_pager_swblk_empty(sb, 0, SWAP_META_PAGES)) {
2020 		SWAP_PCTRIE_REMOVE(&object->un_pager.swp.swp_blks, sb->p);
2021 		uma_zfree(swblk_zone, sb);
2022 	}
2023 }
2024 
2025 /*
2026  * SWP_PAGER_META_BUILD() -	add swap block to swap meta data for object
2027  *
2028  *	We first convert the object to a swap object if it is a default
2029  *	object.
2030  *
2031  *	The specified swapblk is added to the object's swap metadata.  If
2032  *	the swapblk is not valid, it is freed instead.  Any previously
2033  *	assigned swapblk is returned.
2034  */
2035 static daddr_t
2036 swp_pager_meta_build(vm_object_t object, vm_pindex_t pindex, daddr_t swapblk)
2037 {
2038 	static volatile int swblk_zone_exhausted, swpctrie_zone_exhausted;
2039 	struct swblk *sb, *sb1;
2040 	vm_pindex_t modpi, rdpi;
2041 	daddr_t prev_swapblk;
2042 	int error, i;
2043 
2044 	VM_OBJECT_ASSERT_WLOCKED(object);
2045 
2046 	/*
2047 	 * Convert default object to swap object if necessary
2048 	 */
2049 	if ((object->flags & OBJ_SWAP) == 0) {
2050 		pctrie_init(&object->un_pager.swp.swp_blks);
2051 
2052 		/*
2053 		 * Ensure that swap_pager_swapoff()'s iteration over
2054 		 * object_list does not see a garbage pctrie.
2055 		 */
2056 		atomic_thread_fence_rel();
2057 
2058 		object->type = OBJT_SWAP;
2059 		vm_object_set_flag(object, OBJ_SWAP);
2060 		object->un_pager.swp.writemappings = 0;
2061 		KASSERT((object->flags & OBJ_ANON) != 0 ||
2062 		    object->handle == NULL,
2063 		    ("default pager %p with handle %p",
2064 		    object, object->handle));
2065 	}
2066 
2067 	rdpi = rounddown(pindex, SWAP_META_PAGES);
2068 	sb = SWAP_PCTRIE_LOOKUP(&object->un_pager.swp.swp_blks, rdpi);
2069 	if (sb == NULL) {
2070 		if (swapblk == SWAPBLK_NONE)
2071 			return (SWAPBLK_NONE);
2072 		for (;;) {
2073 			sb = uma_zalloc(swblk_zone, M_NOWAIT | (curproc ==
2074 			    pageproc ? M_USE_RESERVE : 0));
2075 			if (sb != NULL) {
2076 				sb->p = rdpi;
2077 				for (i = 0; i < SWAP_META_PAGES; i++)
2078 					sb->d[i] = SWAPBLK_NONE;
2079 				if (atomic_cmpset_int(&swblk_zone_exhausted,
2080 				    1, 0))
2081 					printf("swblk zone ok\n");
2082 				break;
2083 			}
2084 			VM_OBJECT_WUNLOCK(object);
2085 			if (uma_zone_exhausted(swblk_zone)) {
2086 				if (atomic_cmpset_int(&swblk_zone_exhausted,
2087 				    0, 1))
2088 					printf("swap blk zone exhausted, "
2089 					    "increase kern.maxswzone\n");
2090 				vm_pageout_oom(VM_OOM_SWAPZ);
2091 				pause("swzonxb", 10);
2092 			} else
2093 				uma_zwait(swblk_zone);
2094 			VM_OBJECT_WLOCK(object);
2095 			sb = SWAP_PCTRIE_LOOKUP(&object->un_pager.swp.swp_blks,
2096 			    rdpi);
2097 			if (sb != NULL)
2098 				/*
2099 				 * Somebody swapped out a nearby page,
2100 				 * allocating swblk at the rdpi index,
2101 				 * while we dropped the object lock.
2102 				 */
2103 				goto allocated;
2104 		}
2105 		for (;;) {
2106 			error = SWAP_PCTRIE_INSERT(
2107 			    &object->un_pager.swp.swp_blks, sb);
2108 			if (error == 0) {
2109 				if (atomic_cmpset_int(&swpctrie_zone_exhausted,
2110 				    1, 0))
2111 					printf("swpctrie zone ok\n");
2112 				break;
2113 			}
2114 			VM_OBJECT_WUNLOCK(object);
2115 			if (uma_zone_exhausted(swpctrie_zone)) {
2116 				if (atomic_cmpset_int(&swpctrie_zone_exhausted,
2117 				    0, 1))
2118 					printf("swap pctrie zone exhausted, "
2119 					    "increase kern.maxswzone\n");
2120 				vm_pageout_oom(VM_OOM_SWAPZ);
2121 				pause("swzonxp", 10);
2122 			} else
2123 				uma_zwait(swpctrie_zone);
2124 			VM_OBJECT_WLOCK(object);
2125 			sb1 = SWAP_PCTRIE_LOOKUP(&object->un_pager.swp.swp_blks,
2126 			    rdpi);
2127 			if (sb1 != NULL) {
2128 				uma_zfree(swblk_zone, sb);
2129 				sb = sb1;
2130 				goto allocated;
2131 			}
2132 		}
2133 	}
2134 allocated:
2135 	MPASS(sb->p == rdpi);
2136 
2137 	modpi = pindex % SWAP_META_PAGES;
2138 	/* Return prior contents of metadata. */
2139 	prev_swapblk = sb->d[modpi];
2140 	/* Enter block into metadata. */
2141 	sb->d[modpi] = swapblk;
2142 
2143 	/*
2144 	 * Free the swblk if we end up with the empty page run.
2145 	 */
2146 	if (swapblk == SWAPBLK_NONE)
2147 		swp_pager_free_empty_swblk(object, sb);
2148 	return (prev_swapblk);
2149 }
2150 
2151 /*
2152  * SWP_PAGER_META_TRANSFER() - free a range of blocks in the srcobject's swap
2153  * metadata, or transfer it into dstobject.
2154  *
2155  *	This routine will free swap metadata structures as they are cleaned
2156  *	out.
2157  */
2158 static void
2159 swp_pager_meta_transfer(vm_object_t srcobject, vm_object_t dstobject,
2160     vm_pindex_t pindex, vm_pindex_t count)
2161 {
2162 	struct swblk *sb;
2163 	daddr_t n_free, s_free;
2164 	vm_pindex_t offset, last;
2165 	int i, limit, start;
2166 
2167 	VM_OBJECT_ASSERT_WLOCKED(srcobject);
2168 	if ((srcobject->flags & OBJ_SWAP) == 0 || count == 0)
2169 		return;
2170 
2171 	swp_pager_init_freerange(&s_free, &n_free);
2172 	offset = pindex;
2173 	last = pindex + count;
2174 	for (;;) {
2175 		sb = SWAP_PCTRIE_LOOKUP_GE(&srcobject->un_pager.swp.swp_blks,
2176 		    rounddown(pindex, SWAP_META_PAGES));
2177 		if (sb == NULL || sb->p >= last)
2178 			break;
2179 		start = pindex > sb->p ? pindex - sb->p : 0;
2180 		limit = last - sb->p < SWAP_META_PAGES ? last - sb->p :
2181 		    SWAP_META_PAGES;
2182 		for (i = start; i < limit; i++) {
2183 			if (sb->d[i] == SWAPBLK_NONE)
2184 				continue;
2185 			if (dstobject == NULL ||
2186 			    !swp_pager_xfer_source(srcobject, dstobject,
2187 			    sb->p + i - offset, sb->d[i])) {
2188 				swp_pager_update_freerange(&s_free, &n_free,
2189 				    sb->d[i]);
2190 			}
2191 			sb->d[i] = SWAPBLK_NONE;
2192 		}
2193 		pindex = sb->p + SWAP_META_PAGES;
2194 		if (swp_pager_swblk_empty(sb, 0, start) &&
2195 		    swp_pager_swblk_empty(sb, limit, SWAP_META_PAGES)) {
2196 			SWAP_PCTRIE_REMOVE(&srcobject->un_pager.swp.swp_blks,
2197 			    sb->p);
2198 			uma_zfree(swblk_zone, sb);
2199 		}
2200 	}
2201 	swp_pager_freeswapspace(s_free, n_free);
2202 }
2203 
2204 /*
2205  * SWP_PAGER_META_FREE() - free a range of blocks in the object's swap metadata
2206  *
2207  *	The requested range of blocks is freed, with any associated swap
2208  *	returned to the swap bitmap.
2209  *
2210  *	This routine will free swap metadata structures as they are cleaned
2211  *	out.  This routine does *NOT* operate on swap metadata associated
2212  *	with resident pages.
2213  */
2214 static void
2215 swp_pager_meta_free(vm_object_t object, vm_pindex_t pindex, vm_pindex_t count)
2216 {
2217 	swp_pager_meta_transfer(object, NULL, pindex, count);
2218 }
2219 
2220 /*
2221  * SWP_PAGER_META_FREE_ALL() - destroy all swap metadata associated with object
2222  *
2223  *	This routine locates and destroys all swap metadata associated with
2224  *	an object.
2225  */
2226 static void
2227 swp_pager_meta_free_all(vm_object_t object)
2228 {
2229 	struct swblk *sb;
2230 	daddr_t n_free, s_free;
2231 	vm_pindex_t pindex;
2232 	int i;
2233 
2234 	VM_OBJECT_ASSERT_WLOCKED(object);
2235 	if ((object->flags & OBJ_SWAP) == 0)
2236 		return;
2237 
2238 	swp_pager_init_freerange(&s_free, &n_free);
2239 	for (pindex = 0; (sb = SWAP_PCTRIE_LOOKUP_GE(
2240 	    &object->un_pager.swp.swp_blks, pindex)) != NULL;) {
2241 		pindex = sb->p + SWAP_META_PAGES;
2242 		for (i = 0; i < SWAP_META_PAGES; i++) {
2243 			if (sb->d[i] == SWAPBLK_NONE)
2244 				continue;
2245 			swp_pager_update_freerange(&s_free, &n_free, sb->d[i]);
2246 		}
2247 		SWAP_PCTRIE_REMOVE(&object->un_pager.swp.swp_blks, sb->p);
2248 		uma_zfree(swblk_zone, sb);
2249 	}
2250 	swp_pager_freeswapspace(s_free, n_free);
2251 }
2252 
2253 /*
2254  * SWP_PAGER_METACTL() -  misc control of swap meta data.
2255  *
2256  *	This routine is capable of looking up, or removing swapblk
2257  *	assignments in the swap meta data.  It returns the swapblk being
2258  *	looked-up, popped, or SWAPBLK_NONE if the block was invalid.
2259  *
2260  *	When acting on a busy resident page and paging is in progress, we
2261  *	have to wait until paging is complete but otherwise can act on the
2262  *	busy page.
2263  */
2264 static daddr_t
2265 swp_pager_meta_lookup(vm_object_t object, vm_pindex_t pindex)
2266 {
2267 	struct swblk *sb;
2268 
2269 	VM_OBJECT_ASSERT_LOCKED(object);
2270 
2271 	/*
2272 	 * The meta data only exists if the object is OBJT_SWAP
2273 	 * and even then might not be allocated yet.
2274 	 */
2275 	KASSERT((object->flags & OBJ_SWAP) != 0,
2276 	    ("Lookup object not swappable"));
2277 
2278 	sb = SWAP_PCTRIE_LOOKUP(&object->un_pager.swp.swp_blks,
2279 	    rounddown(pindex, SWAP_META_PAGES));
2280 	if (sb == NULL)
2281 		return (SWAPBLK_NONE);
2282 	return (sb->d[pindex % SWAP_META_PAGES]);
2283 }
2284 
2285 /*
2286  * Returns the least page index which is greater than or equal to the
2287  * parameter pindex and for which there is a swap block allocated.
2288  * Returns object's size if the object's type is not swap or if there
2289  * are no allocated swap blocks for the object after the requested
2290  * pindex.
2291  */
2292 vm_pindex_t
2293 swap_pager_find_least(vm_object_t object, vm_pindex_t pindex)
2294 {
2295 	struct swblk *sb;
2296 	int i;
2297 
2298 	VM_OBJECT_ASSERT_LOCKED(object);
2299 	if ((object->flags & OBJ_SWAP) == 0)
2300 		return (object->size);
2301 
2302 	sb = SWAP_PCTRIE_LOOKUP_GE(&object->un_pager.swp.swp_blks,
2303 	    rounddown(pindex, SWAP_META_PAGES));
2304 	if (sb == NULL)
2305 		return (object->size);
2306 	if (sb->p < pindex) {
2307 		for (i = pindex % SWAP_META_PAGES; i < SWAP_META_PAGES; i++) {
2308 			if (sb->d[i] != SWAPBLK_NONE)
2309 				return (sb->p + i);
2310 		}
2311 		sb = SWAP_PCTRIE_LOOKUP_GE(&object->un_pager.swp.swp_blks,
2312 		    roundup(pindex, SWAP_META_PAGES));
2313 		if (sb == NULL)
2314 			return (object->size);
2315 	}
2316 	for (i = 0; i < SWAP_META_PAGES; i++) {
2317 		if (sb->d[i] != SWAPBLK_NONE)
2318 			return (sb->p + i);
2319 	}
2320 
2321 	/*
2322 	 * We get here if a swblk is present in the trie but it
2323 	 * doesn't map any blocks.
2324 	 */
2325 	MPASS(0);
2326 	return (object->size);
2327 }
2328 
2329 /*
2330  * System call swapon(name) enables swapping on device name,
2331  * which must be in the swdevsw.  Return EBUSY
2332  * if already swapping on this device.
2333  */
2334 #ifndef _SYS_SYSPROTO_H_
2335 struct swapon_args {
2336 	char *name;
2337 };
2338 #endif
2339 
2340 int
2341 sys_swapon(struct thread *td, struct swapon_args *uap)
2342 {
2343 	struct vattr attr;
2344 	struct vnode *vp;
2345 	struct nameidata nd;
2346 	int error;
2347 
2348 	error = priv_check(td, PRIV_SWAPON);
2349 	if (error)
2350 		return (error);
2351 
2352 	sx_xlock(&swdev_syscall_lock);
2353 
2354 	/*
2355 	 * Swap metadata may not fit in the KVM if we have physical
2356 	 * memory of >1GB.
2357 	 */
2358 	if (swblk_zone == NULL) {
2359 		error = ENOMEM;
2360 		goto done;
2361 	}
2362 
2363 	NDINIT(&nd, LOOKUP, ISOPEN | FOLLOW | LOCKLEAF | AUDITVNODE1,
2364 	    UIO_USERSPACE, uap->name);
2365 	error = namei(&nd);
2366 	if (error)
2367 		goto done;
2368 
2369 	NDFREE(&nd, NDF_ONLY_PNBUF);
2370 	vp = nd.ni_vp;
2371 
2372 	if (vn_isdisk_error(vp, &error)) {
2373 		error = swapongeom(vp);
2374 	} else if (vp->v_type == VREG &&
2375 	    (vp->v_mount->mnt_vfc->vfc_flags & VFCF_NETWORK) != 0 &&
2376 	    (error = VOP_GETATTR(vp, &attr, td->td_ucred)) == 0) {
2377 		/*
2378 		 * Allow direct swapping to NFS regular files in the same
2379 		 * way that nfs_mountroot() sets up diskless swapping.
2380 		 */
2381 		error = swaponvp(td, vp, attr.va_size / DEV_BSIZE);
2382 	}
2383 
2384 	if (error != 0)
2385 		vput(vp);
2386 	else
2387 		VOP_UNLOCK(vp);
2388 done:
2389 	sx_xunlock(&swdev_syscall_lock);
2390 	return (error);
2391 }
2392 
2393 /*
2394  * Check that the total amount of swap currently configured does not
2395  * exceed half the theoretical maximum.  If it does, print a warning
2396  * message.
2397  */
2398 static void
2399 swapon_check_swzone(void)
2400 {
2401 
2402 	/* recommend using no more than half that amount */
2403 	if (swap_total > swap_maxpages / 2) {
2404 		printf("warning: total configured swap (%lu pages) "
2405 		    "exceeds maximum recommended amount (%lu pages).\n",
2406 		    swap_total, swap_maxpages / 2);
2407 		printf("warning: increase kern.maxswzone "
2408 		    "or reduce amount of swap.\n");
2409 	}
2410 }
2411 
2412 static void
2413 swaponsomething(struct vnode *vp, void *id, u_long nblks,
2414     sw_strategy_t *strategy, sw_close_t *close, dev_t dev, int flags)
2415 {
2416 	struct swdevt *sp, *tsp;
2417 	daddr_t dvbase;
2418 
2419 	/*
2420 	 * nblks is in DEV_BSIZE'd chunks, convert to PAGE_SIZE'd chunks.
2421 	 * First chop nblks off to page-align it, then convert.
2422 	 *
2423 	 * sw->sw_nblks is in page-sized chunks now too.
2424 	 */
2425 	nblks &= ~(ctodb(1) - 1);
2426 	nblks = dbtoc(nblks);
2427 
2428 	sp = malloc(sizeof *sp, M_VMPGDATA, M_WAITOK | M_ZERO);
2429 	sp->sw_blist = blist_create(nblks, M_WAITOK);
2430 	sp->sw_vp = vp;
2431 	sp->sw_id = id;
2432 	sp->sw_dev = dev;
2433 	sp->sw_nblks = nblks;
2434 	sp->sw_used = 0;
2435 	sp->sw_strategy = strategy;
2436 	sp->sw_close = close;
2437 	sp->sw_flags = flags;
2438 
2439 	/*
2440 	 * Do not free the first blocks in order to avoid overwriting
2441 	 * any bsd label at the front of the partition
2442 	 */
2443 	blist_free(sp->sw_blist, howmany(BBSIZE, PAGE_SIZE),
2444 	    nblks - howmany(BBSIZE, PAGE_SIZE));
2445 
2446 	dvbase = 0;
2447 	mtx_lock(&sw_dev_mtx);
2448 	TAILQ_FOREACH(tsp, &swtailq, sw_list) {
2449 		if (tsp->sw_end >= dvbase) {
2450 			/*
2451 			 * We put one uncovered page between the devices
2452 			 * in order to definitively prevent any cross-device
2453 			 * I/O requests
2454 			 */
2455 			dvbase = tsp->sw_end + 1;
2456 		}
2457 	}
2458 	sp->sw_first = dvbase;
2459 	sp->sw_end = dvbase + nblks;
2460 	TAILQ_INSERT_TAIL(&swtailq, sp, sw_list);
2461 	nswapdev++;
2462 	swap_pager_avail += nblks - howmany(BBSIZE, PAGE_SIZE);
2463 	swap_total += nblks;
2464 	swapon_check_swzone();
2465 	swp_sizecheck();
2466 	mtx_unlock(&sw_dev_mtx);
2467 	EVENTHANDLER_INVOKE(swapon, sp);
2468 }
2469 
2470 /*
2471  * SYSCALL: swapoff(devname)
2472  *
2473  * Disable swapping on the given device.
2474  *
2475  * XXX: Badly designed system call: it should use a device index
2476  * rather than filename as specification.  We keep sw_vp around
2477  * only to make this work.
2478  */
2479 static int
2480 kern_swapoff(struct thread *td, const char *name, enum uio_seg name_seg,
2481     u_int flags)
2482 {
2483 	struct vnode *vp;
2484 	struct nameidata nd;
2485 	struct swdevt *sp;
2486 	int error;
2487 
2488 	error = priv_check(td, PRIV_SWAPOFF);
2489 	if (error != 0)
2490 		return (error);
2491 	if ((flags & ~(SWAPOFF_FORCE)) != 0)
2492 		return (EINVAL);
2493 
2494 	sx_xlock(&swdev_syscall_lock);
2495 
2496 	NDINIT(&nd, LOOKUP, FOLLOW | AUDITVNODE1, name_seg, name);
2497 	error = namei(&nd);
2498 	if (error)
2499 		goto done;
2500 	NDFREE(&nd, NDF_ONLY_PNBUF);
2501 	vp = nd.ni_vp;
2502 
2503 	mtx_lock(&sw_dev_mtx);
2504 	TAILQ_FOREACH(sp, &swtailq, sw_list) {
2505 		if (sp->sw_vp == vp)
2506 			break;
2507 	}
2508 	mtx_unlock(&sw_dev_mtx);
2509 	if (sp == NULL) {
2510 		error = EINVAL;
2511 		goto done;
2512 	}
2513 	error = swapoff_one(sp, td->td_ucred, flags);
2514 done:
2515 	sx_xunlock(&swdev_syscall_lock);
2516 	return (error);
2517 }
2518 
2519 
2520 #ifdef COMPAT_FREEBSD13
2521 int
2522 freebsd13_swapoff(struct thread *td, struct freebsd13_swapoff_args *uap)
2523 {
2524 	return (kern_swapoff(td, uap->name, UIO_USERSPACE, 0));
2525 }
2526 #endif
2527 
2528 int
2529 sys_swapoff(struct thread *td, struct swapoff_args *uap)
2530 {
2531 	return (kern_swapoff(td, uap->name, UIO_USERSPACE, uap->flags));
2532 }
2533 
2534 static int
2535 swapoff_one(struct swdevt *sp, struct ucred *cred, u_int flags)
2536 {
2537 	u_long nblks;
2538 #ifdef MAC
2539 	int error;
2540 #endif
2541 
2542 	sx_assert(&swdev_syscall_lock, SA_XLOCKED);
2543 #ifdef MAC
2544 	(void) vn_lock(sp->sw_vp, LK_EXCLUSIVE | LK_RETRY);
2545 	error = mac_system_check_swapoff(cred, sp->sw_vp);
2546 	(void) VOP_UNLOCK(sp->sw_vp);
2547 	if (error != 0)
2548 		return (error);
2549 #endif
2550 	nblks = sp->sw_nblks;
2551 
2552 	/*
2553 	 * We can turn off this swap device safely only if the
2554 	 * available virtual memory in the system will fit the amount
2555 	 * of data we will have to page back in, plus an epsilon so
2556 	 * the system doesn't become critically low on swap space.
2557 	 * The vm_free_count() part does not account e.g. for clean
2558 	 * pages that can be immediately reclaimed without paging, so
2559 	 * this is a very rough estimation.
2560 	 *
2561 	 * On the other hand, not turning swap off on swapoff_all()
2562 	 * means that we can lose swap data when filesystems go away,
2563 	 * which is arguably worse.
2564 	 */
2565 	if ((flags & SWAPOFF_FORCE) == 0 &&
2566 	    vm_free_count() + swap_pager_avail < nblks + nswap_lowat)
2567 		return (ENOMEM);
2568 
2569 	/*
2570 	 * Prevent further allocations on this device.
2571 	 */
2572 	mtx_lock(&sw_dev_mtx);
2573 	sp->sw_flags |= SW_CLOSING;
2574 	swap_pager_avail -= blist_fill(sp->sw_blist, 0, nblks);
2575 	swap_total -= nblks;
2576 	mtx_unlock(&sw_dev_mtx);
2577 
2578 	/*
2579 	 * Page in the contents of the device and close it.
2580 	 */
2581 	swap_pager_swapoff(sp);
2582 
2583 	sp->sw_close(curthread, sp);
2584 	mtx_lock(&sw_dev_mtx);
2585 	sp->sw_id = NULL;
2586 	TAILQ_REMOVE(&swtailq, sp, sw_list);
2587 	nswapdev--;
2588 	if (nswapdev == 0) {
2589 		swap_pager_full = 2;
2590 		swap_pager_almost_full = 1;
2591 	}
2592 	if (swdevhd == sp)
2593 		swdevhd = NULL;
2594 	mtx_unlock(&sw_dev_mtx);
2595 	blist_destroy(sp->sw_blist);
2596 	free(sp, M_VMPGDATA);
2597 	return (0);
2598 }
2599 
2600 void
2601 swapoff_all(void)
2602 {
2603 	struct swdevt *sp, *spt;
2604 	const char *devname;
2605 	int error;
2606 
2607 	sx_xlock(&swdev_syscall_lock);
2608 
2609 	mtx_lock(&sw_dev_mtx);
2610 	TAILQ_FOREACH_SAFE(sp, &swtailq, sw_list, spt) {
2611 		mtx_unlock(&sw_dev_mtx);
2612 		if (vn_isdisk(sp->sw_vp))
2613 			devname = devtoname(sp->sw_vp->v_rdev);
2614 		else
2615 			devname = "[file]";
2616 		error = swapoff_one(sp, thread0.td_ucred, SWAPOFF_FORCE);
2617 		if (error != 0) {
2618 			printf("Cannot remove swap device %s (error=%d), "
2619 			    "skipping.\n", devname, error);
2620 		} else if (bootverbose) {
2621 			printf("Swap device %s removed.\n", devname);
2622 		}
2623 		mtx_lock(&sw_dev_mtx);
2624 	}
2625 	mtx_unlock(&sw_dev_mtx);
2626 
2627 	sx_xunlock(&swdev_syscall_lock);
2628 }
2629 
2630 void
2631 swap_pager_status(int *total, int *used)
2632 {
2633 
2634 	*total = swap_total;
2635 	*used = swap_total - swap_pager_avail -
2636 	    nswapdev * howmany(BBSIZE, PAGE_SIZE);
2637 }
2638 
2639 int
2640 swap_dev_info(int name, struct xswdev *xs, char *devname, size_t len)
2641 {
2642 	struct swdevt *sp;
2643 	const char *tmp_devname;
2644 	int error, n;
2645 
2646 	n = 0;
2647 	error = ENOENT;
2648 	mtx_lock(&sw_dev_mtx);
2649 	TAILQ_FOREACH(sp, &swtailq, sw_list) {
2650 		if (n != name) {
2651 			n++;
2652 			continue;
2653 		}
2654 		xs->xsw_version = XSWDEV_VERSION;
2655 		xs->xsw_dev = sp->sw_dev;
2656 		xs->xsw_flags = sp->sw_flags;
2657 		xs->xsw_nblks = sp->sw_nblks;
2658 		xs->xsw_used = sp->sw_used;
2659 		if (devname != NULL) {
2660 			if (vn_isdisk(sp->sw_vp))
2661 				tmp_devname = devtoname(sp->sw_vp->v_rdev);
2662 			else
2663 				tmp_devname = "[file]";
2664 			strncpy(devname, tmp_devname, len);
2665 		}
2666 		error = 0;
2667 		break;
2668 	}
2669 	mtx_unlock(&sw_dev_mtx);
2670 	return (error);
2671 }
2672 
2673 #if defined(COMPAT_FREEBSD11)
2674 #define XSWDEV_VERSION_11	1
2675 struct xswdev11 {
2676 	u_int	xsw_version;
2677 	uint32_t xsw_dev;
2678 	int	xsw_flags;
2679 	int	xsw_nblks;
2680 	int     xsw_used;
2681 };
2682 #endif
2683 
2684 #if defined(__amd64__) && defined(COMPAT_FREEBSD32)
2685 struct xswdev32 {
2686 	u_int	xsw_version;
2687 	u_int	xsw_dev1, xsw_dev2;
2688 	int	xsw_flags;
2689 	int	xsw_nblks;
2690 	int     xsw_used;
2691 };
2692 #endif
2693 
2694 static int
2695 sysctl_vm_swap_info(SYSCTL_HANDLER_ARGS)
2696 {
2697 	struct xswdev xs;
2698 #if defined(__amd64__) && defined(COMPAT_FREEBSD32)
2699 	struct xswdev32 xs32;
2700 #endif
2701 #if defined(COMPAT_FREEBSD11)
2702 	struct xswdev11 xs11;
2703 #endif
2704 	int error;
2705 
2706 	if (arg2 != 1)			/* name length */
2707 		return (EINVAL);
2708 
2709 	memset(&xs, 0, sizeof(xs));
2710 	error = swap_dev_info(*(int *)arg1, &xs, NULL, 0);
2711 	if (error != 0)
2712 		return (error);
2713 #if defined(__amd64__) && defined(COMPAT_FREEBSD32)
2714 	if (req->oldlen == sizeof(xs32)) {
2715 		memset(&xs32, 0, sizeof(xs32));
2716 		xs32.xsw_version = XSWDEV_VERSION;
2717 		xs32.xsw_dev1 = xs.xsw_dev;
2718 		xs32.xsw_dev2 = xs.xsw_dev >> 32;
2719 		xs32.xsw_flags = xs.xsw_flags;
2720 		xs32.xsw_nblks = xs.xsw_nblks;
2721 		xs32.xsw_used = xs.xsw_used;
2722 		error = SYSCTL_OUT(req, &xs32, sizeof(xs32));
2723 		return (error);
2724 	}
2725 #endif
2726 #if defined(COMPAT_FREEBSD11)
2727 	if (req->oldlen == sizeof(xs11)) {
2728 		memset(&xs11, 0, sizeof(xs11));
2729 		xs11.xsw_version = XSWDEV_VERSION_11;
2730 		xs11.xsw_dev = xs.xsw_dev; /* truncation */
2731 		xs11.xsw_flags = xs.xsw_flags;
2732 		xs11.xsw_nblks = xs.xsw_nblks;
2733 		xs11.xsw_used = xs.xsw_used;
2734 		error = SYSCTL_OUT(req, &xs11, sizeof(xs11));
2735 		return (error);
2736 	}
2737 #endif
2738 	error = SYSCTL_OUT(req, &xs, sizeof(xs));
2739 	return (error);
2740 }
2741 
2742 SYSCTL_INT(_vm, OID_AUTO, nswapdev, CTLFLAG_RD, &nswapdev, 0,
2743     "Number of swap devices");
2744 SYSCTL_NODE(_vm, OID_AUTO, swap_info, CTLFLAG_RD | CTLFLAG_MPSAFE,
2745     sysctl_vm_swap_info,
2746     "Swap statistics by device");
2747 
2748 /*
2749  * Count the approximate swap usage in pages for a vmspace.  The
2750  * shadowed or not yet copied on write swap blocks are not accounted.
2751  * The map must be locked.
2752  */
2753 long
2754 vmspace_swap_count(struct vmspace *vmspace)
2755 {
2756 	vm_map_t map;
2757 	vm_map_entry_t cur;
2758 	vm_object_t object;
2759 	struct swblk *sb;
2760 	vm_pindex_t e, pi;
2761 	long count;
2762 	int i;
2763 
2764 	map = &vmspace->vm_map;
2765 	count = 0;
2766 
2767 	VM_MAP_ENTRY_FOREACH(cur, map) {
2768 		if ((cur->eflags & MAP_ENTRY_IS_SUB_MAP) != 0)
2769 			continue;
2770 		object = cur->object.vm_object;
2771 		if (object == NULL || (object->flags & OBJ_SWAP) == 0)
2772 			continue;
2773 		VM_OBJECT_RLOCK(object);
2774 		if ((object->flags & OBJ_SWAP) == 0)
2775 			goto unlock;
2776 		pi = OFF_TO_IDX(cur->offset);
2777 		e = pi + OFF_TO_IDX(cur->end - cur->start);
2778 		for (;; pi = sb->p + SWAP_META_PAGES) {
2779 			sb = SWAP_PCTRIE_LOOKUP_GE(
2780 			    &object->un_pager.swp.swp_blks, pi);
2781 			if (sb == NULL || sb->p >= e)
2782 				break;
2783 			for (i = 0; i < SWAP_META_PAGES; i++) {
2784 				if (sb->p + i < e &&
2785 				    sb->d[i] != SWAPBLK_NONE)
2786 					count++;
2787 			}
2788 		}
2789 unlock:
2790 		VM_OBJECT_RUNLOCK(object);
2791 	}
2792 	return (count);
2793 }
2794 
2795 /*
2796  * GEOM backend
2797  *
2798  * Swapping onto disk devices.
2799  *
2800  */
2801 
2802 static g_orphan_t swapgeom_orphan;
2803 
2804 static struct g_class g_swap_class = {
2805 	.name = "SWAP",
2806 	.version = G_VERSION,
2807 	.orphan = swapgeom_orphan,
2808 };
2809 
2810 DECLARE_GEOM_CLASS(g_swap_class, g_class);
2811 
2812 static void
2813 swapgeom_close_ev(void *arg, int flags)
2814 {
2815 	struct g_consumer *cp;
2816 
2817 	cp = arg;
2818 	g_access(cp, -1, -1, 0);
2819 	g_detach(cp);
2820 	g_destroy_consumer(cp);
2821 }
2822 
2823 /*
2824  * Add a reference to the g_consumer for an inflight transaction.
2825  */
2826 static void
2827 swapgeom_acquire(struct g_consumer *cp)
2828 {
2829 
2830 	mtx_assert(&sw_dev_mtx, MA_OWNED);
2831 	cp->index++;
2832 }
2833 
2834 /*
2835  * Remove a reference from the g_consumer.  Post a close event if all
2836  * references go away, since the function might be called from the
2837  * biodone context.
2838  */
2839 static void
2840 swapgeom_release(struct g_consumer *cp, struct swdevt *sp)
2841 {
2842 
2843 	mtx_assert(&sw_dev_mtx, MA_OWNED);
2844 	cp->index--;
2845 	if (cp->index == 0) {
2846 		if (g_post_event(swapgeom_close_ev, cp, M_NOWAIT, NULL) == 0)
2847 			sp->sw_id = NULL;
2848 	}
2849 }
2850 
2851 static void
2852 swapgeom_done(struct bio *bp2)
2853 {
2854 	struct swdevt *sp;
2855 	struct buf *bp;
2856 	struct g_consumer *cp;
2857 
2858 	bp = bp2->bio_caller2;
2859 	cp = bp2->bio_from;
2860 	bp->b_ioflags = bp2->bio_flags;
2861 	if (bp2->bio_error)
2862 		bp->b_ioflags |= BIO_ERROR;
2863 	bp->b_resid = bp->b_bcount - bp2->bio_completed;
2864 	bp->b_error = bp2->bio_error;
2865 	bp->b_caller1 = NULL;
2866 	bufdone(bp);
2867 	sp = bp2->bio_caller1;
2868 	mtx_lock(&sw_dev_mtx);
2869 	swapgeom_release(cp, sp);
2870 	mtx_unlock(&sw_dev_mtx);
2871 	g_destroy_bio(bp2);
2872 }
2873 
2874 static void
2875 swapgeom_strategy(struct buf *bp, struct swdevt *sp)
2876 {
2877 	struct bio *bio;
2878 	struct g_consumer *cp;
2879 
2880 	mtx_lock(&sw_dev_mtx);
2881 	cp = sp->sw_id;
2882 	if (cp == NULL) {
2883 		mtx_unlock(&sw_dev_mtx);
2884 		bp->b_error = ENXIO;
2885 		bp->b_ioflags |= BIO_ERROR;
2886 		bufdone(bp);
2887 		return;
2888 	}
2889 	swapgeom_acquire(cp);
2890 	mtx_unlock(&sw_dev_mtx);
2891 	if (bp->b_iocmd == BIO_WRITE)
2892 		bio = g_new_bio();
2893 	else
2894 		bio = g_alloc_bio();
2895 	if (bio == NULL) {
2896 		mtx_lock(&sw_dev_mtx);
2897 		swapgeom_release(cp, sp);
2898 		mtx_unlock(&sw_dev_mtx);
2899 		bp->b_error = ENOMEM;
2900 		bp->b_ioflags |= BIO_ERROR;
2901 		printf("swap_pager: cannot allocate bio\n");
2902 		bufdone(bp);
2903 		return;
2904 	}
2905 
2906 	bp->b_caller1 = bio;
2907 	bio->bio_caller1 = sp;
2908 	bio->bio_caller2 = bp;
2909 	bio->bio_cmd = bp->b_iocmd;
2910 	bio->bio_offset = (bp->b_blkno - sp->sw_first) * PAGE_SIZE;
2911 	bio->bio_length = bp->b_bcount;
2912 	bio->bio_done = swapgeom_done;
2913 	bio->bio_flags |= BIO_SWAP;
2914 	if (!buf_mapped(bp)) {
2915 		bio->bio_ma = bp->b_pages;
2916 		bio->bio_data = unmapped_buf;
2917 		bio->bio_ma_offset = (vm_offset_t)bp->b_offset & PAGE_MASK;
2918 		bio->bio_ma_n = bp->b_npages;
2919 		bio->bio_flags |= BIO_UNMAPPED;
2920 	} else {
2921 		bio->bio_data = bp->b_data;
2922 		bio->bio_ma = NULL;
2923 	}
2924 	g_io_request(bio, cp);
2925 	return;
2926 }
2927 
2928 static void
2929 swapgeom_orphan(struct g_consumer *cp)
2930 {
2931 	struct swdevt *sp;
2932 	int destroy;
2933 
2934 	mtx_lock(&sw_dev_mtx);
2935 	TAILQ_FOREACH(sp, &swtailq, sw_list) {
2936 		if (sp->sw_id == cp) {
2937 			sp->sw_flags |= SW_CLOSING;
2938 			break;
2939 		}
2940 	}
2941 	/*
2942 	 * Drop reference we were created with. Do directly since we're in a
2943 	 * special context where we don't have to queue the call to
2944 	 * swapgeom_close_ev().
2945 	 */
2946 	cp->index--;
2947 	destroy = ((sp != NULL) && (cp->index == 0));
2948 	if (destroy)
2949 		sp->sw_id = NULL;
2950 	mtx_unlock(&sw_dev_mtx);
2951 	if (destroy)
2952 		swapgeom_close_ev(cp, 0);
2953 }
2954 
2955 static void
2956 swapgeom_close(struct thread *td, struct swdevt *sw)
2957 {
2958 	struct g_consumer *cp;
2959 
2960 	mtx_lock(&sw_dev_mtx);
2961 	cp = sw->sw_id;
2962 	sw->sw_id = NULL;
2963 	mtx_unlock(&sw_dev_mtx);
2964 
2965 	/*
2966 	 * swapgeom_close() may be called from the biodone context,
2967 	 * where we cannot perform topology changes.  Delegate the
2968 	 * work to the events thread.
2969 	 */
2970 	if (cp != NULL)
2971 		g_waitfor_event(swapgeom_close_ev, cp, M_WAITOK, NULL);
2972 }
2973 
2974 static int
2975 swapongeom_locked(struct cdev *dev, struct vnode *vp)
2976 {
2977 	struct g_provider *pp;
2978 	struct g_consumer *cp;
2979 	static struct g_geom *gp;
2980 	struct swdevt *sp;
2981 	u_long nblks;
2982 	int error;
2983 
2984 	pp = g_dev_getprovider(dev);
2985 	if (pp == NULL)
2986 		return (ENODEV);
2987 	mtx_lock(&sw_dev_mtx);
2988 	TAILQ_FOREACH(sp, &swtailq, sw_list) {
2989 		cp = sp->sw_id;
2990 		if (cp != NULL && cp->provider == pp) {
2991 			mtx_unlock(&sw_dev_mtx);
2992 			return (EBUSY);
2993 		}
2994 	}
2995 	mtx_unlock(&sw_dev_mtx);
2996 	if (gp == NULL)
2997 		gp = g_new_geomf(&g_swap_class, "swap");
2998 	cp = g_new_consumer(gp);
2999 	cp->index = 1;	/* Number of active I/Os, plus one for being active. */
3000 	cp->flags |=  G_CF_DIRECT_SEND | G_CF_DIRECT_RECEIVE;
3001 	g_attach(cp, pp);
3002 	/*
3003 	 * XXX: Every time you think you can improve the margin for
3004 	 * footshooting, somebody depends on the ability to do so:
3005 	 * savecore(8) wants to write to our swapdev so we cannot
3006 	 * set an exclusive count :-(
3007 	 */
3008 	error = g_access(cp, 1, 1, 0);
3009 	if (error != 0) {
3010 		g_detach(cp);
3011 		g_destroy_consumer(cp);
3012 		return (error);
3013 	}
3014 	nblks = pp->mediasize / DEV_BSIZE;
3015 	swaponsomething(vp, cp, nblks, swapgeom_strategy,
3016 	    swapgeom_close, dev2udev(dev),
3017 	    (pp->flags & G_PF_ACCEPT_UNMAPPED) != 0 ? SW_UNMAPPED : 0);
3018 	return (0);
3019 }
3020 
3021 static int
3022 swapongeom(struct vnode *vp)
3023 {
3024 	int error;
3025 
3026 	ASSERT_VOP_ELOCKED(vp, "swapongeom");
3027 	if (vp->v_type != VCHR || VN_IS_DOOMED(vp)) {
3028 		error = ENOENT;
3029 	} else {
3030 		g_topology_lock();
3031 		error = swapongeom_locked(vp->v_rdev, vp);
3032 		g_topology_unlock();
3033 	}
3034 	return (error);
3035 }
3036 
3037 /*
3038  * VNODE backend
3039  *
3040  * This is used mainly for network filesystem (read: probably only tested
3041  * with NFS) swapfiles.
3042  *
3043  */
3044 
3045 static void
3046 swapdev_strategy(struct buf *bp, struct swdevt *sp)
3047 {
3048 	struct vnode *vp2;
3049 
3050 	bp->b_blkno = ctodb(bp->b_blkno - sp->sw_first);
3051 
3052 	vp2 = sp->sw_id;
3053 	vhold(vp2);
3054 	if (bp->b_iocmd == BIO_WRITE) {
3055 		vn_lock(vp2, LK_EXCLUSIVE | LK_RETRY);
3056 		if (bp->b_bufobj)
3057 			bufobj_wdrop(bp->b_bufobj);
3058 		bufobj_wref(&vp2->v_bufobj);
3059 	} else {
3060 		vn_lock(vp2, LK_SHARED | LK_RETRY);
3061 	}
3062 	if (bp->b_bufobj != &vp2->v_bufobj)
3063 		bp->b_bufobj = &vp2->v_bufobj;
3064 	bp->b_vp = vp2;
3065 	bp->b_iooffset = dbtob(bp->b_blkno);
3066 	bstrategy(bp);
3067 	VOP_UNLOCK(vp2);
3068 }
3069 
3070 static void
3071 swapdev_close(struct thread *td, struct swdevt *sp)
3072 {
3073 	struct vnode *vp;
3074 
3075 	vp = sp->sw_vp;
3076 	vn_lock(vp, LK_EXCLUSIVE | LK_RETRY);
3077 	VOP_CLOSE(vp, FREAD | FWRITE, td->td_ucred, td);
3078 	vput(vp);
3079 }
3080 
3081 static int
3082 swaponvp(struct thread *td, struct vnode *vp, u_long nblks)
3083 {
3084 	struct swdevt *sp;
3085 	int error;
3086 
3087 	ASSERT_VOP_ELOCKED(vp, "swaponvp");
3088 	if (nblks == 0)
3089 		return (ENXIO);
3090 	mtx_lock(&sw_dev_mtx);
3091 	TAILQ_FOREACH(sp, &swtailq, sw_list) {
3092 		if (sp->sw_id == vp) {
3093 			mtx_unlock(&sw_dev_mtx);
3094 			return (EBUSY);
3095 		}
3096 	}
3097 	mtx_unlock(&sw_dev_mtx);
3098 
3099 #ifdef MAC
3100 	error = mac_system_check_swapon(td->td_ucred, vp);
3101 	if (error == 0)
3102 #endif
3103 		error = VOP_OPEN(vp, FREAD | FWRITE, td->td_ucred, td, NULL);
3104 	if (error != 0)
3105 		return (error);
3106 
3107 	swaponsomething(vp, vp, nblks, swapdev_strategy, swapdev_close,
3108 	    NODEV, 0);
3109 	return (0);
3110 }
3111 
3112 static int
3113 sysctl_swap_async_max(SYSCTL_HANDLER_ARGS)
3114 {
3115 	int error, new, n;
3116 
3117 	new = nsw_wcount_async_max;
3118 	error = sysctl_handle_int(oidp, &new, 0, req);
3119 	if (error != 0 || req->newptr == NULL)
3120 		return (error);
3121 
3122 	if (new > nswbuf / 2 || new < 1)
3123 		return (EINVAL);
3124 
3125 	mtx_lock(&swbuf_mtx);
3126 	while (nsw_wcount_async_max != new) {
3127 		/*
3128 		 * Adjust difference.  If the current async count is too low,
3129 		 * we will need to sqeeze our update slowly in.  Sleep with a
3130 		 * higher priority than getpbuf() to finish faster.
3131 		 */
3132 		n = new - nsw_wcount_async_max;
3133 		if (nsw_wcount_async + n >= 0) {
3134 			nsw_wcount_async += n;
3135 			nsw_wcount_async_max += n;
3136 			wakeup(&nsw_wcount_async);
3137 		} else {
3138 			nsw_wcount_async_max -= nsw_wcount_async;
3139 			nsw_wcount_async = 0;
3140 			msleep(&nsw_wcount_async, &swbuf_mtx, PSWP,
3141 			    "swpsysctl", 0);
3142 		}
3143 	}
3144 	mtx_unlock(&swbuf_mtx);
3145 
3146 	return (0);
3147 }
3148 
3149 static void
3150 swap_pager_update_writecount(vm_object_t object, vm_offset_t start,
3151     vm_offset_t end)
3152 {
3153 
3154 	VM_OBJECT_WLOCK(object);
3155 	KASSERT((object->flags & OBJ_ANON) == 0,
3156 	    ("Splittable object with writecount"));
3157 	object->un_pager.swp.writemappings += (vm_ooffset_t)end - start;
3158 	VM_OBJECT_WUNLOCK(object);
3159 }
3160 
3161 static void
3162 swap_pager_release_writecount(vm_object_t object, vm_offset_t start,
3163     vm_offset_t end)
3164 {
3165 
3166 	VM_OBJECT_WLOCK(object);
3167 	KASSERT((object->flags & OBJ_ANON) == 0,
3168 	    ("Splittable object with writecount"));
3169 	object->un_pager.swp.writemappings -= (vm_ooffset_t)end - start;
3170 	VM_OBJECT_WUNLOCK(object);
3171 }
3172