xref: /freebsd/sys/vm/swap_pager.c (revision 3416500aef140042c64bc149cb1ec6620483bc44)
1 /*-
2  * Copyright (c) 1998 Matthew Dillon,
3  * Copyright (c) 1994 John S. Dyson
4  * Copyright (c) 1990 University of Utah.
5  * Copyright (c) 1982, 1986, 1989, 1993
6  *	The Regents of the University of California.  All rights reserved.
7  *
8  * This code is derived from software contributed to Berkeley by
9  * the Systems Programming Group of the University of Utah Computer
10  * Science Department.
11  *
12  * Redistribution and use in source and binary forms, with or without
13  * modification, are permitted provided that the following conditions
14  * are met:
15  * 1. Redistributions of source code must retain the above copyright
16  *    notice, this list of conditions and the following disclaimer.
17  * 2. Redistributions in binary form must reproduce the above copyright
18  *    notice, this list of conditions and the following disclaimer in the
19  *    documentation and/or other materials provided with the distribution.
20  * 3. All advertising materials mentioning features or use of this software
21  *    must display the following acknowledgement:
22  *	This product includes software developed by the University of
23  *	California, Berkeley and its contributors.
24  * 4. Neither the name of the University nor the names of its contributors
25  *    may be used to endorse or promote products derived from this software
26  *    without specific prior written permission.
27  *
28  * THIS SOFTWARE IS PROVIDED BY THE REGENTS AND CONTRIBUTORS ``AS IS'' AND
29  * ANY EXPRESS OR IMPLIED WARRANTIES, INCLUDING, BUT NOT LIMITED TO, THE
30  * IMPLIED WARRANTIES OF MERCHANTABILITY AND FITNESS FOR A PARTICULAR PURPOSE
31  * ARE DISCLAIMED.  IN NO EVENT SHALL THE REGENTS OR CONTRIBUTORS BE LIABLE
32  * FOR ANY DIRECT, INDIRECT, INCIDENTAL, SPECIAL, EXEMPLARY, OR CONSEQUENTIAL
33  * DAMAGES (INCLUDING, BUT NOT LIMITED TO, PROCUREMENT OF SUBSTITUTE GOODS
34  * OR SERVICES; LOSS OF USE, DATA, OR PROFITS; OR BUSINESS INTERRUPTION)
35  * HOWEVER CAUSED AND ON ANY THEORY OF LIABILITY, WHETHER IN CONTRACT, STRICT
36  * LIABILITY, OR TORT (INCLUDING NEGLIGENCE OR OTHERWISE) ARISING IN ANY WAY
37  * OUT OF THE USE OF THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF
38  * SUCH DAMAGE.
39  *
40  *				New Swap System
41  *				Matthew Dillon
42  *
43  * Radix Bitmap 'blists'.
44  *
45  *	- The new swapper uses the new radix bitmap code.  This should scale
46  *	  to arbitrarily small or arbitrarily large swap spaces and an almost
47  *	  arbitrary degree of fragmentation.
48  *
49  * Features:
50  *
51  *	- on the fly reallocation of swap during putpages.  The new system
52  *	  does not try to keep previously allocated swap blocks for dirty
53  *	  pages.
54  *
55  *	- on the fly deallocation of swap
56  *
57  *	- No more garbage collection required.  Unnecessarily allocated swap
58  *	  blocks only exist for dirty vm_page_t's now and these are already
59  *	  cycled (in a high-load system) by the pager.  We also do on-the-fly
60  *	  removal of invalidated swap blocks when a page is destroyed
61  *	  or renamed.
62  *
63  * from: Utah $Hdr: swap_pager.c 1.4 91/04/30$
64  *
65  *	@(#)swap_pager.c	8.9 (Berkeley) 3/21/94
66  *	@(#)vm_swap.c	8.5 (Berkeley) 2/17/94
67  */
68 
69 #include <sys/cdefs.h>
70 __FBSDID("$FreeBSD$");
71 
72 #include "opt_compat.h"
73 #include "opt_swap.h"
74 #include "opt_vm.h"
75 
76 #include <sys/param.h>
77 #include <sys/systm.h>
78 #include <sys/conf.h>
79 #include <sys/kernel.h>
80 #include <sys/priv.h>
81 #include <sys/proc.h>
82 #include <sys/bio.h>
83 #include <sys/buf.h>
84 #include <sys/disk.h>
85 #include <sys/fcntl.h>
86 #include <sys/mount.h>
87 #include <sys/namei.h>
88 #include <sys/vnode.h>
89 #include <sys/malloc.h>
90 #include <sys/racct.h>
91 #include <sys/resource.h>
92 #include <sys/resourcevar.h>
93 #include <sys/rwlock.h>
94 #include <sys/sysctl.h>
95 #include <sys/sysproto.h>
96 #include <sys/blist.h>
97 #include <sys/lock.h>
98 #include <sys/sx.h>
99 #include <sys/vmmeter.h>
100 
101 #include <security/mac/mac_framework.h>
102 
103 #include <vm/vm.h>
104 #include <vm/pmap.h>
105 #include <vm/vm_map.h>
106 #include <vm/vm_kern.h>
107 #include <vm/vm_object.h>
108 #include <vm/vm_page.h>
109 #include <vm/vm_pager.h>
110 #include <vm/vm_pageout.h>
111 #include <vm/vm_param.h>
112 #include <vm/swap_pager.h>
113 #include <vm/vm_extern.h>
114 #include <vm/uma.h>
115 
116 #include <geom/geom.h>
117 
118 /*
119  * MAX_PAGEOUT_CLUSTER must be a power of 2 between 1 and 64.
120  * The 64-page limit is due to the radix code (kern/subr_blist.c).
121  */
122 #ifndef MAX_PAGEOUT_CLUSTER
123 #define	MAX_PAGEOUT_CLUSTER	32
124 #endif
125 
126 #if !defined(SWB_NPAGES)
127 #define SWB_NPAGES	MAX_PAGEOUT_CLUSTER
128 #endif
129 
130 /*
131  * The swblock structure maps an object and a small, fixed-size range
132  * of page indices to disk addresses within a swap area.
133  * The collection of these mappings is implemented as a hash table.
134  * Unused disk addresses within a swap area are allocated and managed
135  * using a blist.
136  */
137 #define	SWAP_META_PAGES		32
138 #define SWAP_META_MASK		(SWAP_META_PAGES - 1)
139 
140 struct swblock {
141 	struct swblock	*swb_hnext;
142 	vm_object_t	swb_object;
143 	vm_pindex_t	swb_index;
144 	int		swb_count;
145 	daddr_t		swb_pages[SWAP_META_PAGES];
146 };
147 
148 static MALLOC_DEFINE(M_VMPGDATA, "vm_pgdata", "swap pager private data");
149 static struct mtx sw_dev_mtx;
150 static TAILQ_HEAD(, swdevt) swtailq = TAILQ_HEAD_INITIALIZER(swtailq);
151 static struct swdevt *swdevhd;	/* Allocate from here next */
152 static int nswapdev;		/* Number of swap devices */
153 int swap_pager_avail;
154 static struct sx swdev_syscall_lock;	/* serialize swap(on|off) */
155 
156 static vm_ooffset_t swap_total;
157 SYSCTL_QUAD(_vm, OID_AUTO, swap_total, CTLFLAG_RD, &swap_total, 0,
158     "Total amount of available swap storage.");
159 static vm_ooffset_t swap_reserved;
160 SYSCTL_QUAD(_vm, OID_AUTO, swap_reserved, CTLFLAG_RD, &swap_reserved, 0,
161     "Amount of swap storage needed to back all allocated anonymous memory.");
162 static int overcommit = 0;
163 SYSCTL_INT(_vm, OID_AUTO, overcommit, CTLFLAG_RW, &overcommit, 0,
164     "Configure virtual memory overcommit behavior. See tuning(7) "
165     "for details.");
166 static unsigned long swzone;
167 SYSCTL_ULONG(_vm, OID_AUTO, swzone, CTLFLAG_RD, &swzone, 0,
168     "Actual size of swap metadata zone");
169 static unsigned long swap_maxpages;
170 SYSCTL_ULONG(_vm, OID_AUTO, swap_maxpages, CTLFLAG_RD, &swap_maxpages, 0,
171     "Maximum amount of swap supported");
172 
173 /* bits from overcommit */
174 #define	SWAP_RESERVE_FORCE_ON		(1 << 0)
175 #define	SWAP_RESERVE_RLIMIT_ON		(1 << 1)
176 #define	SWAP_RESERVE_ALLOW_NONWIRED	(1 << 2)
177 
178 int
179 swap_reserve(vm_ooffset_t incr)
180 {
181 
182 	return (swap_reserve_by_cred(incr, curthread->td_ucred));
183 }
184 
185 int
186 swap_reserve_by_cred(vm_ooffset_t incr, struct ucred *cred)
187 {
188 	vm_ooffset_t r, s;
189 	int res, error;
190 	static int curfail;
191 	static struct timeval lastfail;
192 	struct uidinfo *uip;
193 
194 	uip = cred->cr_ruidinfo;
195 
196 	if (incr & PAGE_MASK)
197 		panic("swap_reserve: & PAGE_MASK");
198 
199 #ifdef RACCT
200 	if (racct_enable) {
201 		PROC_LOCK(curproc);
202 		error = racct_add(curproc, RACCT_SWAP, incr);
203 		PROC_UNLOCK(curproc);
204 		if (error != 0)
205 			return (0);
206 	}
207 #endif
208 
209 	res = 0;
210 	mtx_lock(&sw_dev_mtx);
211 	r = swap_reserved + incr;
212 	if (overcommit & SWAP_RESERVE_ALLOW_NONWIRED) {
213 		s = vm_cnt.v_page_count - vm_cnt.v_free_reserved - vm_cnt.v_wire_count;
214 		s *= PAGE_SIZE;
215 	} else
216 		s = 0;
217 	s += swap_total;
218 	if ((overcommit & SWAP_RESERVE_FORCE_ON) == 0 || r <= s ||
219 	    (error = priv_check(curthread, PRIV_VM_SWAP_NOQUOTA)) == 0) {
220 		res = 1;
221 		swap_reserved = r;
222 	}
223 	mtx_unlock(&sw_dev_mtx);
224 
225 	if (res) {
226 		UIDINFO_VMSIZE_LOCK(uip);
227 		if ((overcommit & SWAP_RESERVE_RLIMIT_ON) != 0 &&
228 		    uip->ui_vmsize + incr > lim_cur(curthread, RLIMIT_SWAP) &&
229 		    priv_check(curthread, PRIV_VM_SWAP_NORLIMIT))
230 			res = 0;
231 		else
232 			uip->ui_vmsize += incr;
233 		UIDINFO_VMSIZE_UNLOCK(uip);
234 		if (!res) {
235 			mtx_lock(&sw_dev_mtx);
236 			swap_reserved -= incr;
237 			mtx_unlock(&sw_dev_mtx);
238 		}
239 	}
240 	if (!res && ppsratecheck(&lastfail, &curfail, 1)) {
241 		printf("uid %d, pid %d: swap reservation for %jd bytes failed\n",
242 		    uip->ui_uid, curproc->p_pid, incr);
243 	}
244 
245 #ifdef RACCT
246 	if (!res) {
247 		PROC_LOCK(curproc);
248 		racct_sub(curproc, RACCT_SWAP, incr);
249 		PROC_UNLOCK(curproc);
250 	}
251 #endif
252 
253 	return (res);
254 }
255 
256 void
257 swap_reserve_force(vm_ooffset_t incr)
258 {
259 	struct uidinfo *uip;
260 
261 	mtx_lock(&sw_dev_mtx);
262 	swap_reserved += incr;
263 	mtx_unlock(&sw_dev_mtx);
264 
265 #ifdef RACCT
266 	PROC_LOCK(curproc);
267 	racct_add_force(curproc, RACCT_SWAP, incr);
268 	PROC_UNLOCK(curproc);
269 #endif
270 
271 	uip = curthread->td_ucred->cr_ruidinfo;
272 	PROC_LOCK(curproc);
273 	UIDINFO_VMSIZE_LOCK(uip);
274 	uip->ui_vmsize += incr;
275 	UIDINFO_VMSIZE_UNLOCK(uip);
276 	PROC_UNLOCK(curproc);
277 }
278 
279 void
280 swap_release(vm_ooffset_t decr)
281 {
282 	struct ucred *cred;
283 
284 	PROC_LOCK(curproc);
285 	cred = curthread->td_ucred;
286 	swap_release_by_cred(decr, cred);
287 	PROC_UNLOCK(curproc);
288 }
289 
290 void
291 swap_release_by_cred(vm_ooffset_t decr, struct ucred *cred)
292 {
293  	struct uidinfo *uip;
294 
295 	uip = cred->cr_ruidinfo;
296 
297 	if (decr & PAGE_MASK)
298 		panic("swap_release: & PAGE_MASK");
299 
300 	mtx_lock(&sw_dev_mtx);
301 	if (swap_reserved < decr)
302 		panic("swap_reserved < decr");
303 	swap_reserved -= decr;
304 	mtx_unlock(&sw_dev_mtx);
305 
306 	UIDINFO_VMSIZE_LOCK(uip);
307 	if (uip->ui_vmsize < decr)
308 		printf("negative vmsize for uid = %d\n", uip->ui_uid);
309 	uip->ui_vmsize -= decr;
310 	UIDINFO_VMSIZE_UNLOCK(uip);
311 
312 	racct_sub_cred(cred, RACCT_SWAP, decr);
313 }
314 
315 #define SWM_FREE	0x02	/* free, period			*/
316 #define SWM_POP		0x04	/* pop out			*/
317 
318 int swap_pager_full = 2;	/* swap space exhaustion (task killing) */
319 static int swap_pager_almost_full = 1; /* swap space exhaustion (w/hysteresis)*/
320 static int nsw_rcount;		/* free read buffers			*/
321 static int nsw_wcount_sync;	/* limit write buffers / synchronous	*/
322 static int nsw_wcount_async;	/* limit write buffers / asynchronous	*/
323 static int nsw_wcount_async_max;/* assigned maximum			*/
324 static int nsw_cluster_max;	/* maximum VOP I/O allowed		*/
325 
326 static int sysctl_swap_async_max(SYSCTL_HANDLER_ARGS);
327 SYSCTL_PROC(_vm, OID_AUTO, swap_async_max, CTLTYPE_INT | CTLFLAG_RW |
328     CTLFLAG_MPSAFE, NULL, 0, sysctl_swap_async_max, "I",
329     "Maximum running async swap ops");
330 
331 static struct swblock **swhash;
332 static int swhash_mask;
333 static struct mtx swhash_mtx;
334 
335 static struct sx sw_alloc_sx;
336 
337 /*
338  * "named" and "unnamed" anon region objects.  Try to reduce the overhead
339  * of searching a named list by hashing it just a little.
340  */
341 
342 #define NOBJLISTS		8
343 
344 #define NOBJLIST(handle)	\
345 	(&swap_pager_object_list[((int)(intptr_t)handle >> 4) & (NOBJLISTS-1)])
346 
347 static struct pagerlst	swap_pager_object_list[NOBJLISTS];
348 static uma_zone_t	swap_zone;
349 
350 /*
351  * pagerops for OBJT_SWAP - "swap pager".  Some ops are also global procedure
352  * calls hooked from other parts of the VM system and do not appear here.
353  * (see vm/swap_pager.h).
354  */
355 static vm_object_t
356 		swap_pager_alloc(void *handle, vm_ooffset_t size,
357 		    vm_prot_t prot, vm_ooffset_t offset, struct ucred *);
358 static void	swap_pager_dealloc(vm_object_t object);
359 static int	swap_pager_getpages(vm_object_t, vm_page_t *, int, int *,
360     int *);
361 static int	swap_pager_getpages_async(vm_object_t, vm_page_t *, int, int *,
362     int *, pgo_getpages_iodone_t, void *);
363 static void	swap_pager_putpages(vm_object_t, vm_page_t *, int, boolean_t, int *);
364 static boolean_t
365 		swap_pager_haspage(vm_object_t object, vm_pindex_t pindex, int *before, int *after);
366 static void	swap_pager_init(void);
367 static void	swap_pager_unswapped(vm_page_t);
368 static void	swap_pager_swapoff(struct swdevt *sp);
369 
370 struct pagerops swappagerops = {
371 	.pgo_init =	swap_pager_init,	/* early system initialization of pager	*/
372 	.pgo_alloc =	swap_pager_alloc,	/* allocate an OBJT_SWAP object		*/
373 	.pgo_dealloc =	swap_pager_dealloc,	/* deallocate an OBJT_SWAP object	*/
374 	.pgo_getpages =	swap_pager_getpages,	/* pagein				*/
375 	.pgo_getpages_async = swap_pager_getpages_async, /* pagein (async)		*/
376 	.pgo_putpages =	swap_pager_putpages,	/* pageout				*/
377 	.pgo_haspage =	swap_pager_haspage,	/* get backing store status for page	*/
378 	.pgo_pageunswapped = swap_pager_unswapped,	/* remove swap related to page		*/
379 };
380 
381 /*
382  * swap_*() routines are externally accessible.  swp_*() routines are
383  * internal.
384  */
385 static int nswap_lowat = 128;	/* in pages, swap_pager_almost_full warn */
386 static int nswap_hiwat = 512;	/* in pages, swap_pager_almost_full warn */
387 
388 SYSCTL_INT(_vm, OID_AUTO, dmmax, CTLFLAG_RD, &nsw_cluster_max, 0,
389     "Maximum size of a swap block in pages");
390 
391 static void	swp_sizecheck(void);
392 static void	swp_pager_async_iodone(struct buf *bp);
393 static int	swapongeom(struct vnode *);
394 static int	swaponvp(struct thread *, struct vnode *, u_long);
395 static int	swapoff_one(struct swdevt *sp, struct ucred *cred);
396 
397 /*
398  * Swap bitmap functions
399  */
400 static void	swp_pager_freeswapspace(daddr_t blk, int npages);
401 static daddr_t	swp_pager_getswapspace(int npages);
402 
403 /*
404  * Metadata functions
405  */
406 static struct swblock **swp_pager_hash(vm_object_t object, vm_pindex_t index);
407 static void swp_pager_meta_build(vm_object_t, vm_pindex_t, daddr_t);
408 static void swp_pager_meta_free(vm_object_t, vm_pindex_t, vm_pindex_t);
409 static void swp_pager_meta_free_all(vm_object_t);
410 static daddr_t swp_pager_meta_ctl(vm_object_t, vm_pindex_t, int);
411 
412 /*
413  * SWP_SIZECHECK() -	update swap_pager_full indication
414  *
415  *	update the swap_pager_almost_full indication and warn when we are
416  *	about to run out of swap space, using lowat/hiwat hysteresis.
417  *
418  *	Clear swap_pager_full ( task killing ) indication when lowat is met.
419  *
420  *	No restrictions on call
421  *	This routine may not block.
422  */
423 static void
424 swp_sizecheck(void)
425 {
426 
427 	if (swap_pager_avail < nswap_lowat) {
428 		if (swap_pager_almost_full == 0) {
429 			printf("swap_pager: out of swap space\n");
430 			swap_pager_almost_full = 1;
431 		}
432 	} else {
433 		swap_pager_full = 0;
434 		if (swap_pager_avail > nswap_hiwat)
435 			swap_pager_almost_full = 0;
436 	}
437 }
438 
439 /*
440  * SWP_PAGER_HASH() -	hash swap meta data
441  *
442  *	This is an helper function which hashes the swapblk given
443  *	the object and page index.  It returns a pointer to a pointer
444  *	to the object, or a pointer to a NULL pointer if it could not
445  *	find a swapblk.
446  */
447 static struct swblock **
448 swp_pager_hash(vm_object_t object, vm_pindex_t index)
449 {
450 	struct swblock **pswap;
451 	struct swblock *swap;
452 
453 	index &= ~(vm_pindex_t)SWAP_META_MASK;
454 	pswap = &swhash[(index ^ (int)(intptr_t)object) & swhash_mask];
455 	while ((swap = *pswap) != NULL) {
456 		if (swap->swb_object == object &&
457 		    swap->swb_index == index
458 		) {
459 			break;
460 		}
461 		pswap = &swap->swb_hnext;
462 	}
463 	return (pswap);
464 }
465 
466 /*
467  * SWAP_PAGER_INIT() -	initialize the swap pager!
468  *
469  *	Expected to be started from system init.  NOTE:  This code is run
470  *	before much else so be careful what you depend on.  Most of the VM
471  *	system has yet to be initialized at this point.
472  */
473 static void
474 swap_pager_init(void)
475 {
476 	/*
477 	 * Initialize object lists
478 	 */
479 	int i;
480 
481 	for (i = 0; i < NOBJLISTS; ++i)
482 		TAILQ_INIT(&swap_pager_object_list[i]);
483 	mtx_init(&sw_dev_mtx, "swapdev", NULL, MTX_DEF);
484 	sx_init(&sw_alloc_sx, "swspsx");
485 	sx_init(&swdev_syscall_lock, "swsysc");
486 }
487 
488 /*
489  * SWAP_PAGER_SWAP_INIT() - swap pager initialization from pageout process
490  *
491  *	Expected to be started from pageout process once, prior to entering
492  *	its main loop.
493  */
494 void
495 swap_pager_swap_init(void)
496 {
497 	unsigned long n, n2;
498 
499 	/*
500 	 * Number of in-transit swap bp operations.  Don't
501 	 * exhaust the pbufs completely.  Make sure we
502 	 * initialize workable values (0 will work for hysteresis
503 	 * but it isn't very efficient).
504 	 *
505 	 * The nsw_cluster_max is constrained by the bp->b_pages[]
506 	 * array (MAXPHYS/PAGE_SIZE) and our locally defined
507 	 * MAX_PAGEOUT_CLUSTER.   Also be aware that swap ops are
508 	 * constrained by the swap device interleave stripe size.
509 	 *
510 	 * Currently we hardwire nsw_wcount_async to 4.  This limit is
511 	 * designed to prevent other I/O from having high latencies due to
512 	 * our pageout I/O.  The value 4 works well for one or two active swap
513 	 * devices but is probably a little low if you have more.  Even so,
514 	 * a higher value would probably generate only a limited improvement
515 	 * with three or four active swap devices since the system does not
516 	 * typically have to pageout at extreme bandwidths.   We will want
517 	 * at least 2 per swap devices, and 4 is a pretty good value if you
518 	 * have one NFS swap device due to the command/ack latency over NFS.
519 	 * So it all works out pretty well.
520 	 */
521 	nsw_cluster_max = min((MAXPHYS/PAGE_SIZE), MAX_PAGEOUT_CLUSTER);
522 
523 	mtx_lock(&pbuf_mtx);
524 	nsw_rcount = (nswbuf + 1) / 2;
525 	nsw_wcount_sync = (nswbuf + 3) / 4;
526 	nsw_wcount_async = 4;
527 	nsw_wcount_async_max = nsw_wcount_async;
528 	mtx_unlock(&pbuf_mtx);
529 
530 	/*
531 	 * Initialize our zone.  Right now I'm just guessing on the number
532 	 * we need based on the number of pages in the system.  Each swblock
533 	 * can hold 32 pages, so this is probably overkill.  This reservation
534 	 * is typically limited to around 32MB by default.
535 	 */
536 	n = vm_cnt.v_page_count / 2;
537 	if (maxswzone && n > maxswzone / sizeof(struct swblock))
538 		n = maxswzone / sizeof(struct swblock);
539 	n2 = n;
540 	swap_zone = uma_zcreate("SWAPMETA", sizeof(struct swblock), NULL, NULL,
541 	    NULL, NULL, UMA_ALIGN_PTR, UMA_ZONE_NOFREE | UMA_ZONE_VM);
542 	if (swap_zone == NULL)
543 		panic("failed to create swap_zone.");
544 	do {
545 		if (uma_zone_reserve_kva(swap_zone, n))
546 			break;
547 		/*
548 		 * if the allocation failed, try a zone two thirds the
549 		 * size of the previous attempt.
550 		 */
551 		n -= ((n + 2) / 3);
552 	} while (n > 0);
553 	if (n2 != n)
554 		printf("Swap zone entries reduced from %lu to %lu.\n", n2, n);
555 	swap_maxpages = n * SWAP_META_PAGES;
556 	swzone = n * sizeof(struct swblock);
557 	n2 = n;
558 
559 	/*
560 	 * Initialize our meta-data hash table.  The swapper does not need to
561 	 * be quite as efficient as the VM system, so we do not use an
562 	 * oversized hash table.
563 	 *
564 	 * 	n: 		size of hash table, must be power of 2
565 	 *	swhash_mask:	hash table index mask
566 	 */
567 	for (n = 1; n < n2 / 8; n *= 2)
568 		;
569 	swhash = malloc(sizeof(struct swblock *) * n, M_VMPGDATA, M_WAITOK | M_ZERO);
570 	swhash_mask = n - 1;
571 	mtx_init(&swhash_mtx, "swap_pager swhash", NULL, MTX_DEF);
572 }
573 
574 static vm_object_t
575 swap_pager_alloc_init(void *handle, struct ucred *cred, vm_ooffset_t size,
576     vm_ooffset_t offset)
577 {
578 	vm_object_t object;
579 
580 	if (cred != NULL) {
581 		if (!swap_reserve_by_cred(size, cred))
582 			return (NULL);
583 		crhold(cred);
584 	}
585 	object = vm_object_allocate(OBJT_SWAP, OFF_TO_IDX(offset +
586 	    PAGE_MASK + size));
587 	object->handle = handle;
588 	if (cred != NULL) {
589 		object->cred = cred;
590 		object->charge = size;
591 	}
592 	object->un_pager.swp.swp_bcount = 0;
593 	return (object);
594 }
595 
596 /*
597  * SWAP_PAGER_ALLOC() -	allocate a new OBJT_SWAP VM object and instantiate
598  *			its metadata structures.
599  *
600  *	This routine is called from the mmap and fork code to create a new
601  *	OBJT_SWAP object.
602  *
603  *	This routine must ensure that no live duplicate is created for
604  *	the named object request, which is protected against by
605  *	holding the sw_alloc_sx lock in case handle != NULL.
606  */
607 static vm_object_t
608 swap_pager_alloc(void *handle, vm_ooffset_t size, vm_prot_t prot,
609     vm_ooffset_t offset, struct ucred *cred)
610 {
611 	vm_object_t object;
612 
613 	if (handle != NULL) {
614 		/*
615 		 * Reference existing named region or allocate new one.  There
616 		 * should not be a race here against swp_pager_meta_build()
617 		 * as called from vm_page_remove() in regards to the lookup
618 		 * of the handle.
619 		 */
620 		sx_xlock(&sw_alloc_sx);
621 		object = vm_pager_object_lookup(NOBJLIST(handle), handle);
622 		if (object == NULL) {
623 			object = swap_pager_alloc_init(handle, cred, size,
624 			    offset);
625 			if (object != NULL) {
626 				TAILQ_INSERT_TAIL(NOBJLIST(object->handle),
627 				    object, pager_object_list);
628 			}
629 		}
630 		sx_xunlock(&sw_alloc_sx);
631 	} else {
632 		object = swap_pager_alloc_init(handle, cred, size, offset);
633 	}
634 	return (object);
635 }
636 
637 /*
638  * SWAP_PAGER_DEALLOC() -	remove swap metadata from object
639  *
640  *	The swap backing for the object is destroyed.  The code is
641  *	designed such that we can reinstantiate it later, but this
642  *	routine is typically called only when the entire object is
643  *	about to be destroyed.
644  *
645  *	The object must be locked.
646  */
647 static void
648 swap_pager_dealloc(vm_object_t object)
649 {
650 
651 	VM_OBJECT_ASSERT_WLOCKED(object);
652 	KASSERT((object->flags & OBJ_DEAD) != 0, ("dealloc of reachable obj"));
653 
654 	/*
655 	 * Remove from list right away so lookups will fail if we block for
656 	 * pageout completion.
657 	 */
658 	if (object->handle != NULL) {
659 		VM_OBJECT_WUNLOCK(object);
660 		sx_xlock(&sw_alloc_sx);
661 		TAILQ_REMOVE(NOBJLIST(object->handle), object,
662 		    pager_object_list);
663 		sx_xunlock(&sw_alloc_sx);
664 		VM_OBJECT_WLOCK(object);
665 	}
666 
667 	vm_object_pip_wait(object, "swpdea");
668 
669 	/*
670 	 * Free all remaining metadata.  We only bother to free it from
671 	 * the swap meta data.  We do not attempt to free swapblk's still
672 	 * associated with vm_page_t's for this object.  We do not care
673 	 * if paging is still in progress on some objects.
674 	 */
675 	swp_pager_meta_free_all(object);
676 	object->handle = NULL;
677 	object->type = OBJT_DEAD;
678 }
679 
680 /************************************************************************
681  *			SWAP PAGER BITMAP ROUTINES			*
682  ************************************************************************/
683 
684 /*
685  * SWP_PAGER_GETSWAPSPACE() -	allocate raw swap space
686  *
687  *	Allocate swap for the requested number of pages.  The starting
688  *	swap block number (a page index) is returned or SWAPBLK_NONE
689  *	if the allocation failed.
690  *
691  *	Also has the side effect of advising that somebody made a mistake
692  *	when they configured swap and didn't configure enough.
693  *
694  *	This routine may not sleep.
695  *
696  *	We allocate in round-robin fashion from the configured devices.
697  */
698 static daddr_t
699 swp_pager_getswapspace(int npages)
700 {
701 	daddr_t blk;
702 	struct swdevt *sp;
703 	int i;
704 
705 	blk = SWAPBLK_NONE;
706 	mtx_lock(&sw_dev_mtx);
707 	sp = swdevhd;
708 	for (i = 0; i < nswapdev; i++) {
709 		if (sp == NULL)
710 			sp = TAILQ_FIRST(&swtailq);
711 		if (!(sp->sw_flags & SW_CLOSING)) {
712 			blk = blist_alloc(sp->sw_blist, npages);
713 			if (blk != SWAPBLK_NONE) {
714 				blk += sp->sw_first;
715 				sp->sw_used += npages;
716 				swap_pager_avail -= npages;
717 				swp_sizecheck();
718 				swdevhd = TAILQ_NEXT(sp, sw_list);
719 				goto done;
720 			}
721 		}
722 		sp = TAILQ_NEXT(sp, sw_list);
723 	}
724 	if (swap_pager_full != 2) {
725 		printf("swap_pager_getswapspace(%d): failed\n", npages);
726 		swap_pager_full = 2;
727 		swap_pager_almost_full = 1;
728 	}
729 	swdevhd = NULL;
730 done:
731 	mtx_unlock(&sw_dev_mtx);
732 	return (blk);
733 }
734 
735 static int
736 swp_pager_isondev(daddr_t blk, struct swdevt *sp)
737 {
738 
739 	return (blk >= sp->sw_first && blk < sp->sw_end);
740 }
741 
742 static void
743 swp_pager_strategy(struct buf *bp)
744 {
745 	struct swdevt *sp;
746 
747 	mtx_lock(&sw_dev_mtx);
748 	TAILQ_FOREACH(sp, &swtailq, sw_list) {
749 		if (bp->b_blkno >= sp->sw_first && bp->b_blkno < sp->sw_end) {
750 			mtx_unlock(&sw_dev_mtx);
751 			if ((sp->sw_flags & SW_UNMAPPED) != 0 &&
752 			    unmapped_buf_allowed) {
753 				bp->b_data = unmapped_buf;
754 				bp->b_offset = 0;
755 			} else {
756 				pmap_qenter((vm_offset_t)bp->b_data,
757 				    &bp->b_pages[0], bp->b_bcount / PAGE_SIZE);
758 			}
759 			sp->sw_strategy(bp, sp);
760 			return;
761 		}
762 	}
763 	panic("Swapdev not found");
764 }
765 
766 
767 /*
768  * SWP_PAGER_FREESWAPSPACE() -	free raw swap space
769  *
770  *	This routine returns the specified swap blocks back to the bitmap.
771  *
772  *	This routine may not sleep.
773  */
774 static void
775 swp_pager_freeswapspace(daddr_t blk, int npages)
776 {
777 	struct swdevt *sp;
778 
779 	mtx_lock(&sw_dev_mtx);
780 	TAILQ_FOREACH(sp, &swtailq, sw_list) {
781 		if (blk >= sp->sw_first && blk < sp->sw_end) {
782 			sp->sw_used -= npages;
783 			/*
784 			 * If we are attempting to stop swapping on
785 			 * this device, we don't want to mark any
786 			 * blocks free lest they be reused.
787 			 */
788 			if ((sp->sw_flags & SW_CLOSING) == 0) {
789 				blist_free(sp->sw_blist, blk - sp->sw_first,
790 				    npages);
791 				swap_pager_avail += npages;
792 				swp_sizecheck();
793 			}
794 			mtx_unlock(&sw_dev_mtx);
795 			return;
796 		}
797 	}
798 	panic("Swapdev not found");
799 }
800 
801 /*
802  * SWAP_PAGER_FREESPACE() -	frees swap blocks associated with a page
803  *				range within an object.
804  *
805  *	This is a globally accessible routine.
806  *
807  *	This routine removes swapblk assignments from swap metadata.
808  *
809  *	The external callers of this routine typically have already destroyed
810  *	or renamed vm_page_t's associated with this range in the object so
811  *	we should be ok.
812  *
813  *	The object must be locked.
814  */
815 void
816 swap_pager_freespace(vm_object_t object, vm_pindex_t start, vm_size_t size)
817 {
818 
819 	swp_pager_meta_free(object, start, size);
820 }
821 
822 /*
823  * SWAP_PAGER_RESERVE() - reserve swap blocks in object
824  *
825  *	Assigns swap blocks to the specified range within the object.  The
826  *	swap blocks are not zeroed.  Any previous swap assignment is destroyed.
827  *
828  *	Returns 0 on success, -1 on failure.
829  */
830 int
831 swap_pager_reserve(vm_object_t object, vm_pindex_t start, vm_size_t size)
832 {
833 	int n = 0;
834 	daddr_t blk = SWAPBLK_NONE;
835 	vm_pindex_t beg = start;	/* save start index */
836 
837 	VM_OBJECT_WLOCK(object);
838 	while (size) {
839 		if (n == 0) {
840 			n = BLIST_MAX_ALLOC;
841 			while ((blk = swp_pager_getswapspace(n)) == SWAPBLK_NONE) {
842 				n >>= 1;
843 				if (n == 0) {
844 					swp_pager_meta_free(object, beg, start - beg);
845 					VM_OBJECT_WUNLOCK(object);
846 					return (-1);
847 				}
848 			}
849 		}
850 		swp_pager_meta_build(object, start, blk);
851 		--size;
852 		++start;
853 		++blk;
854 		--n;
855 	}
856 	swp_pager_meta_free(object, start, n);
857 	VM_OBJECT_WUNLOCK(object);
858 	return (0);
859 }
860 
861 /*
862  * SWAP_PAGER_COPY() -  copy blocks from source pager to destination pager
863  *			and destroy the source.
864  *
865  *	Copy any valid swapblks from the source to the destination.  In
866  *	cases where both the source and destination have a valid swapblk,
867  *	we keep the destination's.
868  *
869  *	This routine is allowed to sleep.  It may sleep allocating metadata
870  *	indirectly through swp_pager_meta_build() or if paging is still in
871  *	progress on the source.
872  *
873  *	The source object contains no vm_page_t's (which is just as well)
874  *
875  *	The source object is of type OBJT_SWAP.
876  *
877  *	The source and destination objects must be locked.
878  *	Both object locks may temporarily be released.
879  */
880 void
881 swap_pager_copy(vm_object_t srcobject, vm_object_t dstobject,
882     vm_pindex_t offset, int destroysource)
883 {
884 	vm_pindex_t i;
885 
886 	VM_OBJECT_ASSERT_WLOCKED(srcobject);
887 	VM_OBJECT_ASSERT_WLOCKED(dstobject);
888 
889 	/*
890 	 * If destroysource is set, we remove the source object from the
891 	 * swap_pager internal queue now.
892 	 */
893 	if (destroysource && srcobject->handle != NULL) {
894 		vm_object_pip_add(srcobject, 1);
895 		VM_OBJECT_WUNLOCK(srcobject);
896 		vm_object_pip_add(dstobject, 1);
897 		VM_OBJECT_WUNLOCK(dstobject);
898 		sx_xlock(&sw_alloc_sx);
899 		TAILQ_REMOVE(NOBJLIST(srcobject->handle), srcobject,
900 		    pager_object_list);
901 		sx_xunlock(&sw_alloc_sx);
902 		VM_OBJECT_WLOCK(dstobject);
903 		vm_object_pip_wakeup(dstobject);
904 		VM_OBJECT_WLOCK(srcobject);
905 		vm_object_pip_wakeup(srcobject);
906 	}
907 
908 	/*
909 	 * transfer source to destination.
910 	 */
911 	for (i = 0; i < dstobject->size; ++i) {
912 		daddr_t dstaddr;
913 
914 		/*
915 		 * Locate (without changing) the swapblk on the destination,
916 		 * unless it is invalid in which case free it silently, or
917 		 * if the destination is a resident page, in which case the
918 		 * source is thrown away.
919 		 */
920 		dstaddr = swp_pager_meta_ctl(dstobject, i, 0);
921 
922 		if (dstaddr == SWAPBLK_NONE) {
923 			/*
924 			 * Destination has no swapblk and is not resident,
925 			 * copy source.
926 			 */
927 			daddr_t srcaddr;
928 
929 			srcaddr = swp_pager_meta_ctl(
930 			    srcobject,
931 			    i + offset,
932 			    SWM_POP
933 			);
934 
935 			if (srcaddr != SWAPBLK_NONE) {
936 				/*
937 				 * swp_pager_meta_build() can sleep.
938 				 */
939 				vm_object_pip_add(srcobject, 1);
940 				VM_OBJECT_WUNLOCK(srcobject);
941 				vm_object_pip_add(dstobject, 1);
942 				swp_pager_meta_build(dstobject, i, srcaddr);
943 				vm_object_pip_wakeup(dstobject);
944 				VM_OBJECT_WLOCK(srcobject);
945 				vm_object_pip_wakeup(srcobject);
946 			}
947 		} else {
948 			/*
949 			 * Destination has valid swapblk or it is represented
950 			 * by a resident page.  We destroy the sourceblock.
951 			 */
952 
953 			swp_pager_meta_ctl(srcobject, i + offset, SWM_FREE);
954 		}
955 	}
956 
957 	/*
958 	 * Free left over swap blocks in source.
959 	 *
960 	 * We have to revert the type to OBJT_DEFAULT so we do not accidentally
961 	 * double-remove the object from the swap queues.
962 	 */
963 	if (destroysource) {
964 		swp_pager_meta_free_all(srcobject);
965 		/*
966 		 * Reverting the type is not necessary, the caller is going
967 		 * to destroy srcobject directly, but I'm doing it here
968 		 * for consistency since we've removed the object from its
969 		 * queues.
970 		 */
971 		srcobject->type = OBJT_DEFAULT;
972 	}
973 }
974 
975 /*
976  * SWAP_PAGER_HASPAGE() -	determine if we have good backing store for
977  *				the requested page.
978  *
979  *	We determine whether good backing store exists for the requested
980  *	page and return TRUE if it does, FALSE if it doesn't.
981  *
982  *	If TRUE, we also try to determine how much valid, contiguous backing
983  *	store exists before and after the requested page.
984  */
985 static boolean_t
986 swap_pager_haspage(vm_object_t object, vm_pindex_t pindex, int *before,
987     int *after)
988 {
989 	daddr_t blk, blk0;
990 	int i;
991 
992 	VM_OBJECT_ASSERT_LOCKED(object);
993 
994 	/*
995 	 * do we have good backing store at the requested index ?
996 	 */
997 	blk0 = swp_pager_meta_ctl(object, pindex, 0);
998 	if (blk0 == SWAPBLK_NONE) {
999 		if (before)
1000 			*before = 0;
1001 		if (after)
1002 			*after = 0;
1003 		return (FALSE);
1004 	}
1005 
1006 	/*
1007 	 * find backwards-looking contiguous good backing store
1008 	 */
1009 	if (before != NULL) {
1010 		for (i = 1; i < SWB_NPAGES; i++) {
1011 			if (i > pindex)
1012 				break;
1013 			blk = swp_pager_meta_ctl(object, pindex - i, 0);
1014 			if (blk != blk0 - i)
1015 				break;
1016 		}
1017 		*before = i - 1;
1018 	}
1019 
1020 	/*
1021 	 * find forward-looking contiguous good backing store
1022 	 */
1023 	if (after != NULL) {
1024 		for (i = 1; i < SWB_NPAGES; i++) {
1025 			blk = swp_pager_meta_ctl(object, pindex + i, 0);
1026 			if (blk != blk0 + i)
1027 				break;
1028 		}
1029 		*after = i - 1;
1030 	}
1031 	return (TRUE);
1032 }
1033 
1034 /*
1035  * SWAP_PAGER_PAGE_UNSWAPPED() - remove swap backing store related to page
1036  *
1037  *	This removes any associated swap backing store, whether valid or
1038  *	not, from the page.
1039  *
1040  *	This routine is typically called when a page is made dirty, at
1041  *	which point any associated swap can be freed.  MADV_FREE also
1042  *	calls us in a special-case situation
1043  *
1044  *	NOTE!!!  If the page is clean and the swap was valid, the caller
1045  *	should make the page dirty before calling this routine.  This routine
1046  *	does NOT change the m->dirty status of the page.  Also: MADV_FREE
1047  *	depends on it.
1048  *
1049  *	This routine may not sleep.
1050  *
1051  *	The object containing the page must be locked.
1052  */
1053 static void
1054 swap_pager_unswapped(vm_page_t m)
1055 {
1056 
1057 	swp_pager_meta_ctl(m->object, m->pindex, SWM_FREE);
1058 }
1059 
1060 /*
1061  * swap_pager_getpages() - bring pages in from swap
1062  *
1063  *	Attempt to page in the pages in array "m" of length "count".  The caller
1064  *	may optionally specify that additional pages preceding and succeeding
1065  *	the specified range be paged in.  The number of such pages is returned
1066  *	in the "rbehind" and "rahead" parameters, and they will be in the
1067  *	inactive queue upon return.
1068  *
1069  *	The pages in "m" must be busied and will remain busied upon return.
1070  */
1071 static int
1072 swap_pager_getpages(vm_object_t object, vm_page_t *m, int count, int *rbehind,
1073     int *rahead)
1074 {
1075 	struct buf *bp;
1076 	vm_page_t mpred, msucc, p;
1077 	vm_pindex_t pindex;
1078 	daddr_t blk;
1079 	int i, j, maxahead, maxbehind, reqcount, shift;
1080 
1081 	reqcount = count;
1082 
1083 	VM_OBJECT_WUNLOCK(object);
1084 	bp = getpbuf(&nsw_rcount);
1085 	VM_OBJECT_WLOCK(object);
1086 
1087 	if (!swap_pager_haspage(object, m[0]->pindex, &maxbehind, &maxahead)) {
1088 		relpbuf(bp, &nsw_rcount);
1089 		return (VM_PAGER_FAIL);
1090 	}
1091 
1092 	/*
1093 	 * Clip the readahead and readbehind ranges to exclude resident pages.
1094 	 */
1095 	if (rahead != NULL) {
1096 		KASSERT(reqcount - 1 <= maxahead,
1097 		    ("page count %d extends beyond swap block", reqcount));
1098 		*rahead = imin(*rahead, maxahead - (reqcount - 1));
1099 		pindex = m[reqcount - 1]->pindex;
1100 		msucc = TAILQ_NEXT(m[reqcount - 1], listq);
1101 		if (msucc != NULL && msucc->pindex - pindex - 1 < *rahead)
1102 			*rahead = msucc->pindex - pindex - 1;
1103 	}
1104 	if (rbehind != NULL) {
1105 		*rbehind = imin(*rbehind, maxbehind);
1106 		pindex = m[0]->pindex;
1107 		mpred = TAILQ_PREV(m[0], pglist, listq);
1108 		if (mpred != NULL && pindex - mpred->pindex - 1 < *rbehind)
1109 			*rbehind = pindex - mpred->pindex - 1;
1110 	}
1111 
1112 	/*
1113 	 * Allocate readahead and readbehind pages.
1114 	 */
1115 	shift = rbehind != NULL ? *rbehind : 0;
1116 	if (shift != 0) {
1117 		for (i = 1; i <= shift; i++) {
1118 			p = vm_page_alloc(object, m[0]->pindex - i,
1119 			    VM_ALLOC_NORMAL);
1120 			if (p == NULL) {
1121 				/* Shift allocated pages to the left. */
1122 				for (j = 0; j < i - 1; j++)
1123 					bp->b_pages[j] =
1124 					    bp->b_pages[j + shift - i + 1];
1125 				break;
1126 			}
1127 			bp->b_pages[shift - i] = p;
1128 		}
1129 		shift = i - 1;
1130 		*rbehind = shift;
1131 	}
1132 	for (i = 0; i < reqcount; i++)
1133 		bp->b_pages[i + shift] = m[i];
1134 	if (rahead != NULL) {
1135 		for (i = 0; i < *rahead; i++) {
1136 			p = vm_page_alloc(object,
1137 			    m[reqcount - 1]->pindex + i + 1, VM_ALLOC_NORMAL);
1138 			if (p == NULL)
1139 				break;
1140 			bp->b_pages[shift + reqcount + i] = p;
1141 		}
1142 		*rahead = i;
1143 	}
1144 	if (rbehind != NULL)
1145 		count += *rbehind;
1146 	if (rahead != NULL)
1147 		count += *rahead;
1148 
1149 	vm_object_pip_add(object, count);
1150 
1151 	for (i = 0; i < count; i++)
1152 		bp->b_pages[i]->oflags |= VPO_SWAPINPROG;
1153 
1154 	pindex = bp->b_pages[0]->pindex;
1155 	blk = swp_pager_meta_ctl(object, pindex, 0);
1156 	KASSERT(blk != SWAPBLK_NONE,
1157 	    ("no swap blocking containing %p(%jx)", object, (uintmax_t)pindex));
1158 
1159 	VM_OBJECT_WUNLOCK(object);
1160 
1161 	bp->b_flags |= B_PAGING;
1162 	bp->b_iocmd = BIO_READ;
1163 	bp->b_iodone = swp_pager_async_iodone;
1164 	bp->b_rcred = crhold(thread0.td_ucred);
1165 	bp->b_wcred = crhold(thread0.td_ucred);
1166 	bp->b_blkno = blk;
1167 	bp->b_bcount = PAGE_SIZE * count;
1168 	bp->b_bufsize = PAGE_SIZE * count;
1169 	bp->b_npages = count;
1170 	bp->b_pgbefore = rbehind != NULL ? *rbehind : 0;
1171 	bp->b_pgafter = rahead != NULL ? *rahead : 0;
1172 
1173 	VM_CNT_INC(v_swapin);
1174 	VM_CNT_ADD(v_swappgsin, count);
1175 
1176 	/*
1177 	 * perform the I/O.  NOTE!!!  bp cannot be considered valid after
1178 	 * this point because we automatically release it on completion.
1179 	 * Instead, we look at the one page we are interested in which we
1180 	 * still hold a lock on even through the I/O completion.
1181 	 *
1182 	 * The other pages in our m[] array are also released on completion,
1183 	 * so we cannot assume they are valid anymore either.
1184 	 *
1185 	 * NOTE: b_blkno is destroyed by the call to swapdev_strategy
1186 	 */
1187 	BUF_KERNPROC(bp);
1188 	swp_pager_strategy(bp);
1189 
1190 	/*
1191 	 * Wait for the pages we want to complete.  VPO_SWAPINPROG is always
1192 	 * cleared on completion.  If an I/O error occurs, SWAPBLK_NONE
1193 	 * is set in the metadata for each page in the request.
1194 	 */
1195 	VM_OBJECT_WLOCK(object);
1196 	while ((m[0]->oflags & VPO_SWAPINPROG) != 0) {
1197 		m[0]->oflags |= VPO_SWAPSLEEP;
1198 		VM_CNT_INC(v_intrans);
1199 		if (VM_OBJECT_SLEEP(object, &object->paging_in_progress, PSWP,
1200 		    "swread", hz * 20)) {
1201 			printf(
1202 "swap_pager: indefinite wait buffer: bufobj: %p, blkno: %jd, size: %ld\n",
1203 			    bp->b_bufobj, (intmax_t)bp->b_blkno, bp->b_bcount);
1204 		}
1205 	}
1206 
1207 	/*
1208 	 * If we had an unrecoverable read error pages will not be valid.
1209 	 */
1210 	for (i = 0; i < reqcount; i++)
1211 		if (m[i]->valid != VM_PAGE_BITS_ALL)
1212 			return (VM_PAGER_ERROR);
1213 
1214 	return (VM_PAGER_OK);
1215 
1216 	/*
1217 	 * A final note: in a low swap situation, we cannot deallocate swap
1218 	 * and mark a page dirty here because the caller is likely to mark
1219 	 * the page clean when we return, causing the page to possibly revert
1220 	 * to all-zero's later.
1221 	 */
1222 }
1223 
1224 /*
1225  * 	swap_pager_getpages_async():
1226  *
1227  *	Right now this is emulation of asynchronous operation on top of
1228  *	swap_pager_getpages().
1229  */
1230 static int
1231 swap_pager_getpages_async(vm_object_t object, vm_page_t *m, int count,
1232     int *rbehind, int *rahead, pgo_getpages_iodone_t iodone, void *arg)
1233 {
1234 	int r, error;
1235 
1236 	r = swap_pager_getpages(object, m, count, rbehind, rahead);
1237 	VM_OBJECT_WUNLOCK(object);
1238 	switch (r) {
1239 	case VM_PAGER_OK:
1240 		error = 0;
1241 		break;
1242 	case VM_PAGER_ERROR:
1243 		error = EIO;
1244 		break;
1245 	case VM_PAGER_FAIL:
1246 		error = EINVAL;
1247 		break;
1248 	default:
1249 		panic("unhandled swap_pager_getpages() error %d", r);
1250 	}
1251 	(iodone)(arg, m, count, error);
1252 	VM_OBJECT_WLOCK(object);
1253 
1254 	return (r);
1255 }
1256 
1257 /*
1258  *	swap_pager_putpages:
1259  *
1260  *	Assign swap (if necessary) and initiate I/O on the specified pages.
1261  *
1262  *	We support both OBJT_DEFAULT and OBJT_SWAP objects.  DEFAULT objects
1263  *	are automatically converted to SWAP objects.
1264  *
1265  *	In a low memory situation we may block in VOP_STRATEGY(), but the new
1266  *	vm_page reservation system coupled with properly written VFS devices
1267  *	should ensure that no low-memory deadlock occurs.  This is an area
1268  *	which needs work.
1269  *
1270  *	The parent has N vm_object_pip_add() references prior to
1271  *	calling us and will remove references for rtvals[] that are
1272  *	not set to VM_PAGER_PEND.  We need to remove the rest on I/O
1273  *	completion.
1274  *
1275  *	The parent has soft-busy'd the pages it passes us and will unbusy
1276  *	those whos rtvals[] entry is not set to VM_PAGER_PEND on return.
1277  *	We need to unbusy the rest on I/O completion.
1278  */
1279 static void
1280 swap_pager_putpages(vm_object_t object, vm_page_t *m, int count,
1281     int flags, int *rtvals)
1282 {
1283 	int i, n;
1284 	boolean_t sync;
1285 
1286 	if (count && m[0]->object != object) {
1287 		panic("swap_pager_putpages: object mismatch %p/%p",
1288 		    object,
1289 		    m[0]->object
1290 		);
1291 	}
1292 
1293 	/*
1294 	 * Step 1
1295 	 *
1296 	 * Turn object into OBJT_SWAP
1297 	 * check for bogus sysops
1298 	 * force sync if not pageout process
1299 	 */
1300 	if (object->type != OBJT_SWAP)
1301 		swp_pager_meta_build(object, 0, SWAPBLK_NONE);
1302 	VM_OBJECT_WUNLOCK(object);
1303 
1304 	n = 0;
1305 	if (curproc != pageproc)
1306 		sync = TRUE;
1307 	else
1308 		sync = (flags & VM_PAGER_PUT_SYNC) != 0;
1309 
1310 	/*
1311 	 * Step 2
1312 	 *
1313 	 * Assign swap blocks and issue I/O.  We reallocate swap on the fly.
1314 	 * The page is left dirty until the pageout operation completes
1315 	 * successfully.
1316 	 */
1317 	for (i = 0; i < count; i += n) {
1318 		int j;
1319 		struct buf *bp;
1320 		daddr_t blk;
1321 
1322 		/*
1323 		 * Maximum I/O size is limited by a number of factors.
1324 		 */
1325 		n = min(BLIST_MAX_ALLOC, count - i);
1326 		n = min(n, nsw_cluster_max);
1327 
1328 		/*
1329 		 * Get biggest block of swap we can.  If we fail, fall
1330 		 * back and try to allocate a smaller block.  Don't go
1331 		 * overboard trying to allocate space if it would overly
1332 		 * fragment swap.
1333 		 */
1334 		while (
1335 		    (blk = swp_pager_getswapspace(n)) == SWAPBLK_NONE &&
1336 		    n > 4
1337 		) {
1338 			n >>= 1;
1339 		}
1340 		if (blk == SWAPBLK_NONE) {
1341 			for (j = 0; j < n; ++j)
1342 				rtvals[i+j] = VM_PAGER_FAIL;
1343 			continue;
1344 		}
1345 
1346 		/*
1347 		 * All I/O parameters have been satisfied, build the I/O
1348 		 * request and assign the swap space.
1349 		 */
1350 		if (sync == TRUE) {
1351 			bp = getpbuf(&nsw_wcount_sync);
1352 		} else {
1353 			bp = getpbuf(&nsw_wcount_async);
1354 			bp->b_flags = B_ASYNC;
1355 		}
1356 		bp->b_flags |= B_PAGING;
1357 		bp->b_iocmd = BIO_WRITE;
1358 
1359 		bp->b_rcred = crhold(thread0.td_ucred);
1360 		bp->b_wcred = crhold(thread0.td_ucred);
1361 		bp->b_bcount = PAGE_SIZE * n;
1362 		bp->b_bufsize = PAGE_SIZE * n;
1363 		bp->b_blkno = blk;
1364 
1365 		VM_OBJECT_WLOCK(object);
1366 		for (j = 0; j < n; ++j) {
1367 			vm_page_t mreq = m[i+j];
1368 
1369 			swp_pager_meta_build(
1370 			    mreq->object,
1371 			    mreq->pindex,
1372 			    blk + j
1373 			);
1374 			MPASS(mreq->dirty == VM_PAGE_BITS_ALL);
1375 			mreq->oflags |= VPO_SWAPINPROG;
1376 			bp->b_pages[j] = mreq;
1377 		}
1378 		VM_OBJECT_WUNLOCK(object);
1379 		bp->b_npages = n;
1380 		/*
1381 		 * Must set dirty range for NFS to work.
1382 		 */
1383 		bp->b_dirtyoff = 0;
1384 		bp->b_dirtyend = bp->b_bcount;
1385 
1386 		VM_CNT_INC(v_swapout);
1387 		VM_CNT_ADD(v_swappgsout, bp->b_npages);
1388 
1389 		/*
1390 		 * We unconditionally set rtvals[] to VM_PAGER_PEND so that we
1391 		 * can call the async completion routine at the end of a
1392 		 * synchronous I/O operation.  Otherwise, our caller would
1393 		 * perform duplicate unbusy and wakeup operations on the page
1394 		 * and object, respectively.
1395 		 */
1396 		for (j = 0; j < n; j++)
1397 			rtvals[i + j] = VM_PAGER_PEND;
1398 
1399 		/*
1400 		 * asynchronous
1401 		 *
1402 		 * NOTE: b_blkno is destroyed by the call to swapdev_strategy
1403 		 */
1404 		if (sync == FALSE) {
1405 			bp->b_iodone = swp_pager_async_iodone;
1406 			BUF_KERNPROC(bp);
1407 			swp_pager_strategy(bp);
1408 			continue;
1409 		}
1410 
1411 		/*
1412 		 * synchronous
1413 		 *
1414 		 * NOTE: b_blkno is destroyed by the call to swapdev_strategy
1415 		 */
1416 		bp->b_iodone = bdone;
1417 		swp_pager_strategy(bp);
1418 
1419 		/*
1420 		 * Wait for the sync I/O to complete.
1421 		 */
1422 		bwait(bp, PVM, "swwrt");
1423 
1424 		/*
1425 		 * Now that we are through with the bp, we can call the
1426 		 * normal async completion, which frees everything up.
1427 		 */
1428 		swp_pager_async_iodone(bp);
1429 	}
1430 	VM_OBJECT_WLOCK(object);
1431 }
1432 
1433 /*
1434  *	swp_pager_async_iodone:
1435  *
1436  *	Completion routine for asynchronous reads and writes from/to swap.
1437  *	Also called manually by synchronous code to finish up a bp.
1438  *
1439  *	This routine may not sleep.
1440  */
1441 static void
1442 swp_pager_async_iodone(struct buf *bp)
1443 {
1444 	int i;
1445 	vm_object_t object = NULL;
1446 
1447 	/*
1448 	 * report error
1449 	 */
1450 	if (bp->b_ioflags & BIO_ERROR) {
1451 		printf(
1452 		    "swap_pager: I/O error - %s failed; blkno %ld,"
1453 			"size %ld, error %d\n",
1454 		    ((bp->b_iocmd == BIO_READ) ? "pagein" : "pageout"),
1455 		    (long)bp->b_blkno,
1456 		    (long)bp->b_bcount,
1457 		    bp->b_error
1458 		);
1459 	}
1460 
1461 	/*
1462 	 * remove the mapping for kernel virtual
1463 	 */
1464 	if (buf_mapped(bp))
1465 		pmap_qremove((vm_offset_t)bp->b_data, bp->b_npages);
1466 	else
1467 		bp->b_data = bp->b_kvabase;
1468 
1469 	if (bp->b_npages) {
1470 		object = bp->b_pages[0]->object;
1471 		VM_OBJECT_WLOCK(object);
1472 	}
1473 
1474 	/*
1475 	 * cleanup pages.  If an error occurs writing to swap, we are in
1476 	 * very serious trouble.  If it happens to be a disk error, though,
1477 	 * we may be able to recover by reassigning the swap later on.  So
1478 	 * in this case we remove the m->swapblk assignment for the page
1479 	 * but do not free it in the rlist.  The errornous block(s) are thus
1480 	 * never reallocated as swap.  Redirty the page and continue.
1481 	 */
1482 	for (i = 0; i < bp->b_npages; ++i) {
1483 		vm_page_t m = bp->b_pages[i];
1484 
1485 		m->oflags &= ~VPO_SWAPINPROG;
1486 		if (m->oflags & VPO_SWAPSLEEP) {
1487 			m->oflags &= ~VPO_SWAPSLEEP;
1488 			wakeup(&object->paging_in_progress);
1489 		}
1490 
1491 		if (bp->b_ioflags & BIO_ERROR) {
1492 			/*
1493 			 * If an error occurs I'd love to throw the swapblk
1494 			 * away without freeing it back to swapspace, so it
1495 			 * can never be used again.  But I can't from an
1496 			 * interrupt.
1497 			 */
1498 			if (bp->b_iocmd == BIO_READ) {
1499 				/*
1500 				 * NOTE: for reads, m->dirty will probably
1501 				 * be overridden by the original caller of
1502 				 * getpages so don't play cute tricks here.
1503 				 */
1504 				m->valid = 0;
1505 			} else {
1506 				/*
1507 				 * If a write error occurs, reactivate page
1508 				 * so it doesn't clog the inactive list,
1509 				 * then finish the I/O.
1510 				 */
1511 				vm_page_dirty(m);
1512 				vm_page_lock(m);
1513 				vm_page_activate(m);
1514 				vm_page_unlock(m);
1515 				vm_page_sunbusy(m);
1516 			}
1517 		} else if (bp->b_iocmd == BIO_READ) {
1518 			/*
1519 			 * NOTE: for reads, m->dirty will probably be
1520 			 * overridden by the original caller of getpages so
1521 			 * we cannot set them in order to free the underlying
1522 			 * swap in a low-swap situation.  I don't think we'd
1523 			 * want to do that anyway, but it was an optimization
1524 			 * that existed in the old swapper for a time before
1525 			 * it got ripped out due to precisely this problem.
1526 			 */
1527 			KASSERT(!pmap_page_is_mapped(m),
1528 			    ("swp_pager_async_iodone: page %p is mapped", m));
1529 			KASSERT(m->dirty == 0,
1530 			    ("swp_pager_async_iodone: page %p is dirty", m));
1531 
1532 			m->valid = VM_PAGE_BITS_ALL;
1533 			if (i < bp->b_pgbefore ||
1534 			    i >= bp->b_npages - bp->b_pgafter)
1535 				vm_page_readahead_finish(m);
1536 		} else {
1537 			/*
1538 			 * For write success, clear the dirty
1539 			 * status, then finish the I/O ( which decrements the
1540 			 * busy count and possibly wakes waiter's up ).
1541 			 * A page is only written to swap after a period of
1542 			 * inactivity.  Therefore, we do not expect it to be
1543 			 * reused.
1544 			 */
1545 			KASSERT(!pmap_page_is_write_mapped(m),
1546 			    ("swp_pager_async_iodone: page %p is not write"
1547 			    " protected", m));
1548 			vm_page_undirty(m);
1549 			vm_page_lock(m);
1550 			vm_page_deactivate_noreuse(m);
1551 			vm_page_unlock(m);
1552 			vm_page_sunbusy(m);
1553 		}
1554 	}
1555 
1556 	/*
1557 	 * adjust pip.  NOTE: the original parent may still have its own
1558 	 * pip refs on the object.
1559 	 */
1560 	if (object != NULL) {
1561 		vm_object_pip_wakeupn(object, bp->b_npages);
1562 		VM_OBJECT_WUNLOCK(object);
1563 	}
1564 
1565 	/*
1566 	 * swapdev_strategy() manually sets b_vp and b_bufobj before calling
1567 	 * bstrategy(). Set them back to NULL now we're done with it, or we'll
1568 	 * trigger a KASSERT in relpbuf().
1569 	 */
1570 	if (bp->b_vp) {
1571 		    bp->b_vp = NULL;
1572 		    bp->b_bufobj = NULL;
1573 	}
1574 	/*
1575 	 * release the physical I/O buffer
1576 	 */
1577 	relpbuf(
1578 	    bp,
1579 	    ((bp->b_iocmd == BIO_READ) ? &nsw_rcount :
1580 		((bp->b_flags & B_ASYNC) ?
1581 		    &nsw_wcount_async :
1582 		    &nsw_wcount_sync
1583 		)
1584 	    )
1585 	);
1586 }
1587 
1588 /*
1589  *	swap_pager_isswapped:
1590  *
1591  *	Return 1 if at least one page in the given object is paged
1592  *	out to the given swap device.
1593  *
1594  *	This routine may not sleep.
1595  */
1596 int
1597 swap_pager_isswapped(vm_object_t object, struct swdevt *sp)
1598 {
1599 	daddr_t index = 0;
1600 	int bcount;
1601 	int i;
1602 
1603 	VM_OBJECT_ASSERT_WLOCKED(object);
1604 	if (object->type != OBJT_SWAP)
1605 		return (0);
1606 
1607 	mtx_lock(&swhash_mtx);
1608 	for (bcount = 0; bcount < object->un_pager.swp.swp_bcount; bcount++) {
1609 		struct swblock *swap;
1610 
1611 		if ((swap = *swp_pager_hash(object, index)) != NULL) {
1612 			for (i = 0; i < SWAP_META_PAGES; ++i) {
1613 				if (swp_pager_isondev(swap->swb_pages[i], sp)) {
1614 					mtx_unlock(&swhash_mtx);
1615 					return (1);
1616 				}
1617 			}
1618 		}
1619 		index += SWAP_META_PAGES;
1620 	}
1621 	mtx_unlock(&swhash_mtx);
1622 	return (0);
1623 }
1624 
1625 int
1626 swap_pager_nswapdev(void)
1627 {
1628 
1629 	return (nswapdev);
1630 }
1631 
1632 /*
1633  * SWP_PAGER_FORCE_PAGEIN() - force a swap block to be paged in
1634  *
1635  *	This routine dissociates the page at the given index within an object
1636  *	from its backing store, paging it in if it does not reside in memory.
1637  *	If the page is paged in, it is marked dirty and placed in the laundry
1638  *	queue.  The page is marked dirty because it no longer has backing
1639  *	store.  It is placed in the laundry queue because it has not been
1640  *	accessed recently.  Otherwise, it would already reside in memory.
1641  *
1642  *	We also attempt to swap in all other pages in the swap block.
1643  *	However, we only guarantee that the one at the specified index is
1644  *	paged in.
1645  *
1646  *	XXX - The code to page the whole block in doesn't work, so we
1647  *	      revert to the one-by-one behavior for now.  Sigh.
1648  */
1649 static inline void
1650 swp_pager_force_pagein(vm_object_t object, vm_pindex_t pindex)
1651 {
1652 	vm_page_t m;
1653 
1654 	vm_object_pip_add(object, 1);
1655 	m = vm_page_grab(object, pindex, VM_ALLOC_NORMAL);
1656 	if (m->valid == VM_PAGE_BITS_ALL) {
1657 		vm_object_pip_wakeup(object);
1658 		vm_page_dirty(m);
1659 		vm_page_lock(m);
1660 		vm_page_activate(m);
1661 		vm_page_unlock(m);
1662 		vm_page_xunbusy(m);
1663 		vm_pager_page_unswapped(m);
1664 		return;
1665 	}
1666 
1667 	if (swap_pager_getpages(object, &m, 1, NULL, NULL) != VM_PAGER_OK)
1668 		panic("swap_pager_force_pagein: read from swap failed");/*XXX*/
1669 	vm_object_pip_wakeup(object);
1670 	vm_page_dirty(m);
1671 	vm_page_lock(m);
1672 	vm_page_launder(m);
1673 	vm_page_unlock(m);
1674 	vm_page_xunbusy(m);
1675 	vm_pager_page_unswapped(m);
1676 }
1677 
1678 /*
1679  *	swap_pager_swapoff:
1680  *
1681  *	Page in all of the pages that have been paged out to the
1682  *	given device.  The corresponding blocks in the bitmap must be
1683  *	marked as allocated and the device must be flagged SW_CLOSING.
1684  *	There may be no processes swapped out to the device.
1685  *
1686  *	This routine may block.
1687  */
1688 static void
1689 swap_pager_swapoff(struct swdevt *sp)
1690 {
1691 	struct swblock *swap;
1692 	vm_object_t locked_obj, object;
1693 	vm_pindex_t pindex;
1694 	int i, j, retries;
1695 
1696 	sx_assert(&swdev_syscall_lock, SA_XLOCKED);
1697 
1698 	retries = 0;
1699 	locked_obj = NULL;
1700 full_rescan:
1701 	mtx_lock(&swhash_mtx);
1702 	for (i = 0; i <= swhash_mask; i++) { /* '<=' is correct here */
1703 restart:
1704 		for (swap = swhash[i]; swap != NULL; swap = swap->swb_hnext) {
1705 			object = swap->swb_object;
1706 			pindex = swap->swb_index;
1707 			for (j = 0; j < SWAP_META_PAGES; ++j) {
1708 				if (!swp_pager_isondev(swap->swb_pages[j], sp))
1709 					continue;
1710 				if (locked_obj != object) {
1711 					if (locked_obj != NULL)
1712 						VM_OBJECT_WUNLOCK(locked_obj);
1713 					locked_obj = object;
1714 					if (!VM_OBJECT_TRYWLOCK(object)) {
1715 						mtx_unlock(&swhash_mtx);
1716 						/* Depends on type-stability. */
1717 						VM_OBJECT_WLOCK(object);
1718 						mtx_lock(&swhash_mtx);
1719 						goto restart;
1720 					}
1721 				}
1722 				MPASS(locked_obj == object);
1723 				mtx_unlock(&swhash_mtx);
1724 				swp_pager_force_pagein(object, pindex + j);
1725 				mtx_lock(&swhash_mtx);
1726 				goto restart;
1727 			}
1728 		}
1729 	}
1730 	mtx_unlock(&swhash_mtx);
1731 	if (locked_obj != NULL) {
1732 		VM_OBJECT_WUNLOCK(locked_obj);
1733 		locked_obj = NULL;
1734 	}
1735 	if (sp->sw_used) {
1736 		/*
1737 		 * Objects may be locked or paging to the device being
1738 		 * removed, so we will miss their pages and need to
1739 		 * make another pass.  We have marked this device as
1740 		 * SW_CLOSING, so the activity should finish soon.
1741 		 */
1742 		retries++;
1743 		if (retries > 100) {
1744 			panic("swapoff: failed to locate %d swap blocks",
1745 			    sp->sw_used);
1746 		}
1747 		pause("swpoff", hz / 20);
1748 		goto full_rescan;
1749 	}
1750 	EVENTHANDLER_INVOKE(swapoff, sp);
1751 }
1752 
1753 /************************************************************************
1754  *				SWAP META DATA 				*
1755  ************************************************************************
1756  *
1757  *	These routines manipulate the swap metadata stored in the
1758  *	OBJT_SWAP object.
1759  *
1760  *	Swap metadata is implemented with a global hash and not directly
1761  *	linked into the object.  Instead the object simply contains
1762  *	appropriate tracking counters.
1763  */
1764 
1765 /*
1766  * SWP_PAGER_META_BUILD() -	add swap block to swap meta data for object
1767  *
1768  *	We first convert the object to a swap object if it is a default
1769  *	object.
1770  *
1771  *	The specified swapblk is added to the object's swap metadata.  If
1772  *	the swapblk is not valid, it is freed instead.  Any previously
1773  *	assigned swapblk is freed.
1774  */
1775 static void
1776 swp_pager_meta_build(vm_object_t object, vm_pindex_t pindex, daddr_t swapblk)
1777 {
1778 	static volatile int exhausted;
1779 	struct swblock *swap;
1780 	struct swblock **pswap;
1781 	int idx;
1782 
1783 	VM_OBJECT_ASSERT_WLOCKED(object);
1784 	/*
1785 	 * Convert default object to swap object if necessary
1786 	 */
1787 	if (object->type != OBJT_SWAP) {
1788 		object->type = OBJT_SWAP;
1789 		object->un_pager.swp.swp_bcount = 0;
1790 		KASSERT(object->handle == NULL, ("default pager with handle"));
1791 	}
1792 
1793 	/*
1794 	 * Locate hash entry.  If not found create, but if we aren't adding
1795 	 * anything just return.  If we run out of space in the map we wait
1796 	 * and, since the hash table may have changed, retry.
1797 	 */
1798 retry:
1799 	mtx_lock(&swhash_mtx);
1800 	pswap = swp_pager_hash(object, pindex);
1801 
1802 	if ((swap = *pswap) == NULL) {
1803 		int i;
1804 
1805 		if (swapblk == SWAPBLK_NONE)
1806 			goto done;
1807 
1808 		swap = *pswap = uma_zalloc(swap_zone, M_NOWAIT |
1809 		    (curproc == pageproc ? M_USE_RESERVE : 0));
1810 		if (swap == NULL) {
1811 			mtx_unlock(&swhash_mtx);
1812 			VM_OBJECT_WUNLOCK(object);
1813 			if (uma_zone_exhausted(swap_zone)) {
1814 				if (atomic_cmpset_int(&exhausted, 0, 1))
1815 					printf("swap zone exhausted, "
1816 					    "increase kern.maxswzone\n");
1817 				vm_pageout_oom(VM_OOM_SWAPZ);
1818 				pause("swzonex", 10);
1819 			} else
1820 				VM_WAIT;
1821 			VM_OBJECT_WLOCK(object);
1822 			goto retry;
1823 		}
1824 
1825 		if (atomic_cmpset_int(&exhausted, 1, 0))
1826 			printf("swap zone ok\n");
1827 
1828 		swap->swb_hnext = NULL;
1829 		swap->swb_object = object;
1830 		swap->swb_index = pindex & ~(vm_pindex_t)SWAP_META_MASK;
1831 		swap->swb_count = 0;
1832 
1833 		++object->un_pager.swp.swp_bcount;
1834 
1835 		for (i = 0; i < SWAP_META_PAGES; ++i)
1836 			swap->swb_pages[i] = SWAPBLK_NONE;
1837 	}
1838 
1839 	/*
1840 	 * Delete prior contents of metadata
1841 	 */
1842 	idx = pindex & SWAP_META_MASK;
1843 
1844 	if (swap->swb_pages[idx] != SWAPBLK_NONE) {
1845 		swp_pager_freeswapspace(swap->swb_pages[idx], 1);
1846 		--swap->swb_count;
1847 	}
1848 
1849 	/*
1850 	 * Enter block into metadata
1851 	 */
1852 	swap->swb_pages[idx] = swapblk;
1853 	if (swapblk != SWAPBLK_NONE)
1854 		++swap->swb_count;
1855 done:
1856 	mtx_unlock(&swhash_mtx);
1857 }
1858 
1859 /*
1860  * SWP_PAGER_META_FREE() - free a range of blocks in the object's swap metadata
1861  *
1862  *	The requested range of blocks is freed, with any associated swap
1863  *	returned to the swap bitmap.
1864  *
1865  *	This routine will free swap metadata structures as they are cleaned
1866  *	out.  This routine does *NOT* operate on swap metadata associated
1867  *	with resident pages.
1868  */
1869 static void
1870 swp_pager_meta_free(vm_object_t object, vm_pindex_t index, vm_pindex_t count)
1871 {
1872 	struct swblock **pswap, *swap;
1873 	vm_pindex_t c;
1874 	daddr_t v;
1875 	int n, sidx;
1876 
1877 	VM_OBJECT_ASSERT_LOCKED(object);
1878 	if (object->type != OBJT_SWAP || count == 0)
1879 		return;
1880 
1881 	mtx_lock(&swhash_mtx);
1882 	for (c = 0; c < count;) {
1883 		pswap = swp_pager_hash(object, index);
1884 		sidx = index & SWAP_META_MASK;
1885 		n = SWAP_META_PAGES - sidx;
1886 		index += n;
1887 		if ((swap = *pswap) == NULL) {
1888 			c += n;
1889 			continue;
1890 		}
1891 		for (; c < count && sidx < SWAP_META_PAGES; ++c, ++sidx) {
1892 			if ((v = swap->swb_pages[sidx]) == SWAPBLK_NONE)
1893 				continue;
1894 			swp_pager_freeswapspace(v, 1);
1895 			swap->swb_pages[sidx] = SWAPBLK_NONE;
1896 			if (--swap->swb_count == 0) {
1897 				*pswap = swap->swb_hnext;
1898 				uma_zfree(swap_zone, swap);
1899 				--object->un_pager.swp.swp_bcount;
1900 				c += SWAP_META_PAGES - sidx;
1901 				break;
1902 			}
1903 		}
1904 	}
1905 	mtx_unlock(&swhash_mtx);
1906 }
1907 
1908 /*
1909  * SWP_PAGER_META_FREE_ALL() - destroy all swap metadata associated with object
1910  *
1911  *	This routine locates and destroys all swap metadata associated with
1912  *	an object.
1913  */
1914 static void
1915 swp_pager_meta_free_all(vm_object_t object)
1916 {
1917 	struct swblock **pswap, *swap;
1918 	vm_pindex_t index;
1919 	daddr_t v;
1920 	int i;
1921 
1922 	VM_OBJECT_ASSERT_WLOCKED(object);
1923 	if (object->type != OBJT_SWAP)
1924 		return;
1925 
1926 	index = 0;
1927 	while (object->un_pager.swp.swp_bcount != 0) {
1928 		mtx_lock(&swhash_mtx);
1929 		pswap = swp_pager_hash(object, index);
1930 		if ((swap = *pswap) != NULL) {
1931 			for (i = 0; i < SWAP_META_PAGES; ++i) {
1932 				v = swap->swb_pages[i];
1933 				if (v != SWAPBLK_NONE) {
1934 					--swap->swb_count;
1935 					swp_pager_freeswapspace(v, 1);
1936 				}
1937 			}
1938 			if (swap->swb_count != 0)
1939 				panic(
1940 				    "swap_pager_meta_free_all: swb_count != 0");
1941 			*pswap = swap->swb_hnext;
1942 			uma_zfree(swap_zone, swap);
1943 			--object->un_pager.swp.swp_bcount;
1944 		}
1945 		mtx_unlock(&swhash_mtx);
1946 		index += SWAP_META_PAGES;
1947 	}
1948 }
1949 
1950 /*
1951  * SWP_PAGER_METACTL() -  misc control of swap and vm_page_t meta data.
1952  *
1953  *	This routine is capable of looking up, popping, or freeing
1954  *	swapblk assignments in the swap meta data or in the vm_page_t.
1955  *	The routine typically returns the swapblk being looked-up, or popped,
1956  *	or SWAPBLK_NONE if the block was freed, or SWAPBLK_NONE if the block
1957  *	was invalid.  This routine will automatically free any invalid
1958  *	meta-data swapblks.
1959  *
1960  *	It is not possible to store invalid swapblks in the swap meta data
1961  *	(other then a literal 'SWAPBLK_NONE'), so we don't bother checking.
1962  *
1963  *	When acting on a busy resident page and paging is in progress, we
1964  *	have to wait until paging is complete but otherwise can act on the
1965  *	busy page.
1966  *
1967  *	SWM_FREE	remove and free swap block from metadata
1968  *	SWM_POP		remove from meta data but do not free.. pop it out
1969  */
1970 static daddr_t
1971 swp_pager_meta_ctl(vm_object_t object, vm_pindex_t pindex, int flags)
1972 {
1973 	struct swblock **pswap;
1974 	struct swblock *swap;
1975 	daddr_t r1;
1976 	int idx;
1977 
1978 	VM_OBJECT_ASSERT_LOCKED(object);
1979 	/*
1980 	 * The meta data only exists of the object is OBJT_SWAP
1981 	 * and even then might not be allocated yet.
1982 	 */
1983 	if (object->type != OBJT_SWAP)
1984 		return (SWAPBLK_NONE);
1985 
1986 	r1 = SWAPBLK_NONE;
1987 	mtx_lock(&swhash_mtx);
1988 	pswap = swp_pager_hash(object, pindex);
1989 
1990 	if ((swap = *pswap) != NULL) {
1991 		idx = pindex & SWAP_META_MASK;
1992 		r1 = swap->swb_pages[idx];
1993 
1994 		if (r1 != SWAPBLK_NONE) {
1995 			if (flags & SWM_FREE) {
1996 				swp_pager_freeswapspace(r1, 1);
1997 				r1 = SWAPBLK_NONE;
1998 			}
1999 			if (flags & (SWM_FREE|SWM_POP)) {
2000 				swap->swb_pages[idx] = SWAPBLK_NONE;
2001 				if (--swap->swb_count == 0) {
2002 					*pswap = swap->swb_hnext;
2003 					uma_zfree(swap_zone, swap);
2004 					--object->un_pager.swp.swp_bcount;
2005 				}
2006 			}
2007 		}
2008 	}
2009 	mtx_unlock(&swhash_mtx);
2010 	return (r1);
2011 }
2012 
2013 /*
2014  * Returns the least page index which is greater than or equal to the
2015  * parameter pindex and for which there is a swap block allocated.
2016  * Returns object's size if the object's type is not swap or if there
2017  * are no allocated swap blocks for the object after the requested
2018  * pindex.
2019  */
2020 vm_pindex_t
2021 swap_pager_find_least(vm_object_t object, vm_pindex_t pindex)
2022 {
2023 	struct swblock **pswap, *swap;
2024 	vm_pindex_t i, j, lim;
2025 	int idx;
2026 
2027 	VM_OBJECT_ASSERT_LOCKED(object);
2028 	if (object->type != OBJT_SWAP || object->un_pager.swp.swp_bcount == 0)
2029 		return (object->size);
2030 
2031 	mtx_lock(&swhash_mtx);
2032 	for (j = pindex; j < object->size; j = lim) {
2033 		pswap = swp_pager_hash(object, j);
2034 		lim = rounddown2(j + SWAP_META_PAGES, SWAP_META_PAGES);
2035 		if (lim > object->size)
2036 			lim = object->size;
2037 		if ((swap = *pswap) != NULL) {
2038 			for (idx = j & SWAP_META_MASK, i = j; i < lim;
2039 			    i++, idx++) {
2040 				if (swap->swb_pages[idx] != SWAPBLK_NONE)
2041 					goto found;
2042 			}
2043 		}
2044 	}
2045 	i = object->size;
2046 found:
2047 	mtx_unlock(&swhash_mtx);
2048 	return (i);
2049 }
2050 
2051 /*
2052  * System call swapon(name) enables swapping on device name,
2053  * which must be in the swdevsw.  Return EBUSY
2054  * if already swapping on this device.
2055  */
2056 #ifndef _SYS_SYSPROTO_H_
2057 struct swapon_args {
2058 	char *name;
2059 };
2060 #endif
2061 
2062 /*
2063  * MPSAFE
2064  */
2065 /* ARGSUSED */
2066 int
2067 sys_swapon(struct thread *td, struct swapon_args *uap)
2068 {
2069 	struct vattr attr;
2070 	struct vnode *vp;
2071 	struct nameidata nd;
2072 	int error;
2073 
2074 	error = priv_check(td, PRIV_SWAPON);
2075 	if (error)
2076 		return (error);
2077 
2078 	sx_xlock(&swdev_syscall_lock);
2079 
2080 	/*
2081 	 * Swap metadata may not fit in the KVM if we have physical
2082 	 * memory of >1GB.
2083 	 */
2084 	if (swap_zone == NULL) {
2085 		error = ENOMEM;
2086 		goto done;
2087 	}
2088 
2089 	NDINIT(&nd, LOOKUP, ISOPEN | FOLLOW | AUDITVNODE1, UIO_USERSPACE,
2090 	    uap->name, td);
2091 	error = namei(&nd);
2092 	if (error)
2093 		goto done;
2094 
2095 	NDFREE(&nd, NDF_ONLY_PNBUF);
2096 	vp = nd.ni_vp;
2097 
2098 	if (vn_isdisk(vp, &error)) {
2099 		error = swapongeom(vp);
2100 	} else if (vp->v_type == VREG &&
2101 	    (vp->v_mount->mnt_vfc->vfc_flags & VFCF_NETWORK) != 0 &&
2102 	    (error = VOP_GETATTR(vp, &attr, td->td_ucred)) == 0) {
2103 		/*
2104 		 * Allow direct swapping to NFS regular files in the same
2105 		 * way that nfs_mountroot() sets up diskless swapping.
2106 		 */
2107 		error = swaponvp(td, vp, attr.va_size / DEV_BSIZE);
2108 	}
2109 
2110 	if (error)
2111 		vrele(vp);
2112 done:
2113 	sx_xunlock(&swdev_syscall_lock);
2114 	return (error);
2115 }
2116 
2117 /*
2118  * Check that the total amount of swap currently configured does not
2119  * exceed half the theoretical maximum.  If it does, print a warning
2120  * message and return -1; otherwise, return 0.
2121  */
2122 static int
2123 swapon_check_swzone(unsigned long npages)
2124 {
2125 	unsigned long maxpages;
2126 
2127 	/* absolute maximum we can handle assuming 100% efficiency */
2128 	maxpages = uma_zone_get_max(swap_zone) * SWAP_META_PAGES;
2129 
2130 	/* recommend using no more than half that amount */
2131 	if (npages > maxpages / 2) {
2132 		printf("warning: total configured swap (%lu pages) "
2133 		    "exceeds maximum recommended amount (%lu pages).\n",
2134 		    npages, maxpages / 2);
2135 		printf("warning: increase kern.maxswzone "
2136 		    "or reduce amount of swap.\n");
2137 		return (-1);
2138 	}
2139 	return (0);
2140 }
2141 
2142 static void
2143 swaponsomething(struct vnode *vp, void *id, u_long nblks,
2144     sw_strategy_t *strategy, sw_close_t *close, dev_t dev, int flags)
2145 {
2146 	struct swdevt *sp, *tsp;
2147 	swblk_t dvbase;
2148 	u_long mblocks;
2149 
2150 	/*
2151 	 * nblks is in DEV_BSIZE'd chunks, convert to PAGE_SIZE'd chunks.
2152 	 * First chop nblks off to page-align it, then convert.
2153 	 *
2154 	 * sw->sw_nblks is in page-sized chunks now too.
2155 	 */
2156 	nblks &= ~(ctodb(1) - 1);
2157 	nblks = dbtoc(nblks);
2158 
2159 	/*
2160 	 * If we go beyond this, we get overflows in the radix
2161 	 * tree bitmap code.
2162 	 */
2163 	mblocks = 0x40000000 / BLIST_META_RADIX;
2164 	if (nblks > mblocks) {
2165 		printf(
2166     "WARNING: reducing swap size to maximum of %luMB per unit\n",
2167 		    mblocks / 1024 / 1024 * PAGE_SIZE);
2168 		nblks = mblocks;
2169 	}
2170 
2171 	sp = malloc(sizeof *sp, M_VMPGDATA, M_WAITOK | M_ZERO);
2172 	sp->sw_vp = vp;
2173 	sp->sw_id = id;
2174 	sp->sw_dev = dev;
2175 	sp->sw_flags = 0;
2176 	sp->sw_nblks = nblks;
2177 	sp->sw_used = 0;
2178 	sp->sw_strategy = strategy;
2179 	sp->sw_close = close;
2180 	sp->sw_flags = flags;
2181 
2182 	sp->sw_blist = blist_create(nblks, M_WAITOK);
2183 	/*
2184 	 * Do not free the first two block in order to avoid overwriting
2185 	 * any bsd label at the front of the partition
2186 	 */
2187 	blist_free(sp->sw_blist, 2, nblks - 2);
2188 
2189 	dvbase = 0;
2190 	mtx_lock(&sw_dev_mtx);
2191 	TAILQ_FOREACH(tsp, &swtailq, sw_list) {
2192 		if (tsp->sw_end >= dvbase) {
2193 			/*
2194 			 * We put one uncovered page between the devices
2195 			 * in order to definitively prevent any cross-device
2196 			 * I/O requests
2197 			 */
2198 			dvbase = tsp->sw_end + 1;
2199 		}
2200 	}
2201 	sp->sw_first = dvbase;
2202 	sp->sw_end = dvbase + nblks;
2203 	TAILQ_INSERT_TAIL(&swtailq, sp, sw_list);
2204 	nswapdev++;
2205 	swap_pager_avail += nblks - 2;
2206 	swap_total += (vm_ooffset_t)nblks * PAGE_SIZE;
2207 	swapon_check_swzone(swap_total / PAGE_SIZE);
2208 	swp_sizecheck();
2209 	mtx_unlock(&sw_dev_mtx);
2210 	EVENTHANDLER_INVOKE(swapon, sp);
2211 }
2212 
2213 /*
2214  * SYSCALL: swapoff(devname)
2215  *
2216  * Disable swapping on the given device.
2217  *
2218  * XXX: Badly designed system call: it should use a device index
2219  * rather than filename as specification.  We keep sw_vp around
2220  * only to make this work.
2221  */
2222 #ifndef _SYS_SYSPROTO_H_
2223 struct swapoff_args {
2224 	char *name;
2225 };
2226 #endif
2227 
2228 /*
2229  * MPSAFE
2230  */
2231 /* ARGSUSED */
2232 int
2233 sys_swapoff(struct thread *td, struct swapoff_args *uap)
2234 {
2235 	struct vnode *vp;
2236 	struct nameidata nd;
2237 	struct swdevt *sp;
2238 	int error;
2239 
2240 	error = priv_check(td, PRIV_SWAPOFF);
2241 	if (error)
2242 		return (error);
2243 
2244 	sx_xlock(&swdev_syscall_lock);
2245 
2246 	NDINIT(&nd, LOOKUP, FOLLOW | AUDITVNODE1, UIO_USERSPACE, uap->name,
2247 	    td);
2248 	error = namei(&nd);
2249 	if (error)
2250 		goto done;
2251 	NDFREE(&nd, NDF_ONLY_PNBUF);
2252 	vp = nd.ni_vp;
2253 
2254 	mtx_lock(&sw_dev_mtx);
2255 	TAILQ_FOREACH(sp, &swtailq, sw_list) {
2256 		if (sp->sw_vp == vp)
2257 			break;
2258 	}
2259 	mtx_unlock(&sw_dev_mtx);
2260 	if (sp == NULL) {
2261 		error = EINVAL;
2262 		goto done;
2263 	}
2264 	error = swapoff_one(sp, td->td_ucred);
2265 done:
2266 	sx_xunlock(&swdev_syscall_lock);
2267 	return (error);
2268 }
2269 
2270 static int
2271 swapoff_one(struct swdevt *sp, struct ucred *cred)
2272 {
2273 	u_long nblks;
2274 #ifdef MAC
2275 	int error;
2276 #endif
2277 
2278 	sx_assert(&swdev_syscall_lock, SA_XLOCKED);
2279 #ifdef MAC
2280 	(void) vn_lock(sp->sw_vp, LK_EXCLUSIVE | LK_RETRY);
2281 	error = mac_system_check_swapoff(cred, sp->sw_vp);
2282 	(void) VOP_UNLOCK(sp->sw_vp, 0);
2283 	if (error != 0)
2284 		return (error);
2285 #endif
2286 	nblks = sp->sw_nblks;
2287 
2288 	/*
2289 	 * We can turn off this swap device safely only if the
2290 	 * available virtual memory in the system will fit the amount
2291 	 * of data we will have to page back in, plus an epsilon so
2292 	 * the system doesn't become critically low on swap space.
2293 	 */
2294 	if (vm_cnt.v_free_count + swap_pager_avail < nblks + nswap_lowat)
2295 		return (ENOMEM);
2296 
2297 	/*
2298 	 * Prevent further allocations on this device.
2299 	 */
2300 	mtx_lock(&sw_dev_mtx);
2301 	sp->sw_flags |= SW_CLOSING;
2302 	swap_pager_avail -= blist_fill(sp->sw_blist, 0, nblks);
2303 	swap_total -= (vm_ooffset_t)nblks * PAGE_SIZE;
2304 	mtx_unlock(&sw_dev_mtx);
2305 
2306 	/*
2307 	 * Page in the contents of the device and close it.
2308 	 */
2309 	swap_pager_swapoff(sp);
2310 
2311 	sp->sw_close(curthread, sp);
2312 	mtx_lock(&sw_dev_mtx);
2313 	sp->sw_id = NULL;
2314 	TAILQ_REMOVE(&swtailq, sp, sw_list);
2315 	nswapdev--;
2316 	if (nswapdev == 0) {
2317 		swap_pager_full = 2;
2318 		swap_pager_almost_full = 1;
2319 	}
2320 	if (swdevhd == sp)
2321 		swdevhd = NULL;
2322 	mtx_unlock(&sw_dev_mtx);
2323 	blist_destroy(sp->sw_blist);
2324 	free(sp, M_VMPGDATA);
2325 	return (0);
2326 }
2327 
2328 void
2329 swapoff_all(void)
2330 {
2331 	struct swdevt *sp, *spt;
2332 	const char *devname;
2333 	int error;
2334 
2335 	sx_xlock(&swdev_syscall_lock);
2336 
2337 	mtx_lock(&sw_dev_mtx);
2338 	TAILQ_FOREACH_SAFE(sp, &swtailq, sw_list, spt) {
2339 		mtx_unlock(&sw_dev_mtx);
2340 		if (vn_isdisk(sp->sw_vp, NULL))
2341 			devname = devtoname(sp->sw_vp->v_rdev);
2342 		else
2343 			devname = "[file]";
2344 		error = swapoff_one(sp, thread0.td_ucred);
2345 		if (error != 0) {
2346 			printf("Cannot remove swap device %s (error=%d), "
2347 			    "skipping.\n", devname, error);
2348 		} else if (bootverbose) {
2349 			printf("Swap device %s removed.\n", devname);
2350 		}
2351 		mtx_lock(&sw_dev_mtx);
2352 	}
2353 	mtx_unlock(&sw_dev_mtx);
2354 
2355 	sx_xunlock(&swdev_syscall_lock);
2356 }
2357 
2358 void
2359 swap_pager_status(int *total, int *used)
2360 {
2361 	struct swdevt *sp;
2362 
2363 	*total = 0;
2364 	*used = 0;
2365 	mtx_lock(&sw_dev_mtx);
2366 	TAILQ_FOREACH(sp, &swtailq, sw_list) {
2367 		*total += sp->sw_nblks;
2368 		*used += sp->sw_used;
2369 	}
2370 	mtx_unlock(&sw_dev_mtx);
2371 }
2372 
2373 int
2374 swap_dev_info(int name, struct xswdev *xs, char *devname, size_t len)
2375 {
2376 	struct swdevt *sp;
2377 	const char *tmp_devname;
2378 	int error, n;
2379 
2380 	n = 0;
2381 	error = ENOENT;
2382 	mtx_lock(&sw_dev_mtx);
2383 	TAILQ_FOREACH(sp, &swtailq, sw_list) {
2384 		if (n != name) {
2385 			n++;
2386 			continue;
2387 		}
2388 		xs->xsw_version = XSWDEV_VERSION;
2389 		xs->xsw_dev = sp->sw_dev;
2390 		xs->xsw_flags = sp->sw_flags;
2391 		xs->xsw_nblks = sp->sw_nblks;
2392 		xs->xsw_used = sp->sw_used;
2393 		if (devname != NULL) {
2394 			if (vn_isdisk(sp->sw_vp, NULL))
2395 				tmp_devname = devtoname(sp->sw_vp->v_rdev);
2396 			else
2397 				tmp_devname = "[file]";
2398 			strncpy(devname, tmp_devname, len);
2399 		}
2400 		error = 0;
2401 		break;
2402 	}
2403 	mtx_unlock(&sw_dev_mtx);
2404 	return (error);
2405 }
2406 
2407 #if defined(COMPAT_FREEBSD11)
2408 #define XSWDEV_VERSION_11	1
2409 struct xswdev11 {
2410 	u_int	xsw_version;
2411 	uint32_t xsw_dev;
2412 	int	xsw_flags;
2413 	int	xsw_nblks;
2414 	int     xsw_used;
2415 };
2416 #endif
2417 
2418 static int
2419 sysctl_vm_swap_info(SYSCTL_HANDLER_ARGS)
2420 {
2421 	struct xswdev xs;
2422 #if defined(COMPAT_FREEBSD11)
2423 	struct xswdev11 xs11;
2424 #endif
2425 	int error;
2426 
2427 	if (arg2 != 1)			/* name length */
2428 		return (EINVAL);
2429 	error = swap_dev_info(*(int *)arg1, &xs, NULL, 0);
2430 	if (error != 0)
2431 		return (error);
2432 #if defined(COMPAT_FREEBSD11)
2433 	if (req->oldlen == sizeof(xs11)) {
2434 		xs11.xsw_version = XSWDEV_VERSION_11;
2435 		xs11.xsw_dev = xs.xsw_dev; /* truncation */
2436 		xs11.xsw_flags = xs.xsw_flags;
2437 		xs11.xsw_nblks = xs.xsw_nblks;
2438 		xs11.xsw_used = xs.xsw_used;
2439 		error = SYSCTL_OUT(req, &xs11, sizeof(xs11));
2440 	} else
2441 #endif
2442 		error = SYSCTL_OUT(req, &xs, sizeof(xs));
2443 	return (error);
2444 }
2445 
2446 SYSCTL_INT(_vm, OID_AUTO, nswapdev, CTLFLAG_RD, &nswapdev, 0,
2447     "Number of swap devices");
2448 SYSCTL_NODE(_vm, OID_AUTO, swap_info, CTLFLAG_RD | CTLFLAG_MPSAFE,
2449     sysctl_vm_swap_info,
2450     "Swap statistics by device");
2451 
2452 /*
2453  * vmspace_swap_count() - count the approximate swap usage in pages for a
2454  *			  vmspace.
2455  *
2456  *	The map must be locked.
2457  *
2458  *	Swap usage is determined by taking the proportional swap used by
2459  *	VM objects backing the VM map.  To make up for fractional losses,
2460  *	if the VM object has any swap use at all the associated map entries
2461  *	count for at least 1 swap page.
2462  */
2463 long
2464 vmspace_swap_count(struct vmspace *vmspace)
2465 {
2466 	vm_map_t map;
2467 	vm_map_entry_t cur;
2468 	vm_object_t object;
2469 	long count, n;
2470 
2471 	map = &vmspace->vm_map;
2472 	count = 0;
2473 
2474 	for (cur = map->header.next; cur != &map->header; cur = cur->next) {
2475 		if ((cur->eflags & MAP_ENTRY_IS_SUB_MAP) == 0 &&
2476 		    (object = cur->object.vm_object) != NULL) {
2477 			VM_OBJECT_WLOCK(object);
2478 			if (object->type == OBJT_SWAP &&
2479 			    object->un_pager.swp.swp_bcount != 0) {
2480 				n = (cur->end - cur->start) / PAGE_SIZE;
2481 				count += object->un_pager.swp.swp_bcount *
2482 				    SWAP_META_PAGES * n / object->size + 1;
2483 			}
2484 			VM_OBJECT_WUNLOCK(object);
2485 		}
2486 	}
2487 	return (count);
2488 }
2489 
2490 /*
2491  * GEOM backend
2492  *
2493  * Swapping onto disk devices.
2494  *
2495  */
2496 
2497 static g_orphan_t swapgeom_orphan;
2498 
2499 static struct g_class g_swap_class = {
2500 	.name = "SWAP",
2501 	.version = G_VERSION,
2502 	.orphan = swapgeom_orphan,
2503 };
2504 
2505 DECLARE_GEOM_CLASS(g_swap_class, g_class);
2506 
2507 
2508 static void
2509 swapgeom_close_ev(void *arg, int flags)
2510 {
2511 	struct g_consumer *cp;
2512 
2513 	cp = arg;
2514 	g_access(cp, -1, -1, 0);
2515 	g_detach(cp);
2516 	g_destroy_consumer(cp);
2517 }
2518 
2519 /*
2520  * Add a reference to the g_consumer for an inflight transaction.
2521  */
2522 static void
2523 swapgeom_acquire(struct g_consumer *cp)
2524 {
2525 
2526 	mtx_assert(&sw_dev_mtx, MA_OWNED);
2527 	cp->index++;
2528 }
2529 
2530 /*
2531  * Remove a reference from the g_consumer.  Post a close event if all
2532  * references go away, since the function might be called from the
2533  * biodone context.
2534  */
2535 static void
2536 swapgeom_release(struct g_consumer *cp, struct swdevt *sp)
2537 {
2538 
2539 	mtx_assert(&sw_dev_mtx, MA_OWNED);
2540 	cp->index--;
2541 	if (cp->index == 0) {
2542 		if (g_post_event(swapgeom_close_ev, cp, M_NOWAIT, NULL) == 0)
2543 			sp->sw_id = NULL;
2544 	}
2545 }
2546 
2547 static void
2548 swapgeom_done(struct bio *bp2)
2549 {
2550 	struct swdevt *sp;
2551 	struct buf *bp;
2552 	struct g_consumer *cp;
2553 
2554 	bp = bp2->bio_caller2;
2555 	cp = bp2->bio_from;
2556 	bp->b_ioflags = bp2->bio_flags;
2557 	if (bp2->bio_error)
2558 		bp->b_ioflags |= BIO_ERROR;
2559 	bp->b_resid = bp->b_bcount - bp2->bio_completed;
2560 	bp->b_error = bp2->bio_error;
2561 	bufdone(bp);
2562 	sp = bp2->bio_caller1;
2563 	mtx_lock(&sw_dev_mtx);
2564 	swapgeom_release(cp, sp);
2565 	mtx_unlock(&sw_dev_mtx);
2566 	g_destroy_bio(bp2);
2567 }
2568 
2569 static void
2570 swapgeom_strategy(struct buf *bp, struct swdevt *sp)
2571 {
2572 	struct bio *bio;
2573 	struct g_consumer *cp;
2574 
2575 	mtx_lock(&sw_dev_mtx);
2576 	cp = sp->sw_id;
2577 	if (cp == NULL) {
2578 		mtx_unlock(&sw_dev_mtx);
2579 		bp->b_error = ENXIO;
2580 		bp->b_ioflags |= BIO_ERROR;
2581 		bufdone(bp);
2582 		return;
2583 	}
2584 	swapgeom_acquire(cp);
2585 	mtx_unlock(&sw_dev_mtx);
2586 	if (bp->b_iocmd == BIO_WRITE)
2587 		bio = g_new_bio();
2588 	else
2589 		bio = g_alloc_bio();
2590 	if (bio == NULL) {
2591 		mtx_lock(&sw_dev_mtx);
2592 		swapgeom_release(cp, sp);
2593 		mtx_unlock(&sw_dev_mtx);
2594 		bp->b_error = ENOMEM;
2595 		bp->b_ioflags |= BIO_ERROR;
2596 		bufdone(bp);
2597 		return;
2598 	}
2599 
2600 	bio->bio_caller1 = sp;
2601 	bio->bio_caller2 = bp;
2602 	bio->bio_cmd = bp->b_iocmd;
2603 	bio->bio_offset = (bp->b_blkno - sp->sw_first) * PAGE_SIZE;
2604 	bio->bio_length = bp->b_bcount;
2605 	bio->bio_done = swapgeom_done;
2606 	if (!buf_mapped(bp)) {
2607 		bio->bio_ma = bp->b_pages;
2608 		bio->bio_data = unmapped_buf;
2609 		bio->bio_ma_offset = (vm_offset_t)bp->b_offset & PAGE_MASK;
2610 		bio->bio_ma_n = bp->b_npages;
2611 		bio->bio_flags |= BIO_UNMAPPED;
2612 	} else {
2613 		bio->bio_data = bp->b_data;
2614 		bio->bio_ma = NULL;
2615 	}
2616 	g_io_request(bio, cp);
2617 	return;
2618 }
2619 
2620 static void
2621 swapgeom_orphan(struct g_consumer *cp)
2622 {
2623 	struct swdevt *sp;
2624 	int destroy;
2625 
2626 	mtx_lock(&sw_dev_mtx);
2627 	TAILQ_FOREACH(sp, &swtailq, sw_list) {
2628 		if (sp->sw_id == cp) {
2629 			sp->sw_flags |= SW_CLOSING;
2630 			break;
2631 		}
2632 	}
2633 	/*
2634 	 * Drop reference we were created with. Do directly since we're in a
2635 	 * special context where we don't have to queue the call to
2636 	 * swapgeom_close_ev().
2637 	 */
2638 	cp->index--;
2639 	destroy = ((sp != NULL) && (cp->index == 0));
2640 	if (destroy)
2641 		sp->sw_id = NULL;
2642 	mtx_unlock(&sw_dev_mtx);
2643 	if (destroy)
2644 		swapgeom_close_ev(cp, 0);
2645 }
2646 
2647 static void
2648 swapgeom_close(struct thread *td, struct swdevt *sw)
2649 {
2650 	struct g_consumer *cp;
2651 
2652 	mtx_lock(&sw_dev_mtx);
2653 	cp = sw->sw_id;
2654 	sw->sw_id = NULL;
2655 	mtx_unlock(&sw_dev_mtx);
2656 
2657 	/*
2658 	 * swapgeom_close() may be called from the biodone context,
2659 	 * where we cannot perform topology changes.  Delegate the
2660 	 * work to the events thread.
2661 	 */
2662 	if (cp != NULL)
2663 		g_waitfor_event(swapgeom_close_ev, cp, M_WAITOK, NULL);
2664 }
2665 
2666 static int
2667 swapongeom_locked(struct cdev *dev, struct vnode *vp)
2668 {
2669 	struct g_provider *pp;
2670 	struct g_consumer *cp;
2671 	static struct g_geom *gp;
2672 	struct swdevt *sp;
2673 	u_long nblks;
2674 	int error;
2675 
2676 	pp = g_dev_getprovider(dev);
2677 	if (pp == NULL)
2678 		return (ENODEV);
2679 	mtx_lock(&sw_dev_mtx);
2680 	TAILQ_FOREACH(sp, &swtailq, sw_list) {
2681 		cp = sp->sw_id;
2682 		if (cp != NULL && cp->provider == pp) {
2683 			mtx_unlock(&sw_dev_mtx);
2684 			return (EBUSY);
2685 		}
2686 	}
2687 	mtx_unlock(&sw_dev_mtx);
2688 	if (gp == NULL)
2689 		gp = g_new_geomf(&g_swap_class, "swap");
2690 	cp = g_new_consumer(gp);
2691 	cp->index = 1;	/* Number of active I/Os, plus one for being active. */
2692 	cp->flags |=  G_CF_DIRECT_SEND | G_CF_DIRECT_RECEIVE;
2693 	g_attach(cp, pp);
2694 	/*
2695 	 * XXX: Every time you think you can improve the margin for
2696 	 * footshooting, somebody depends on the ability to do so:
2697 	 * savecore(8) wants to write to our swapdev so we cannot
2698 	 * set an exclusive count :-(
2699 	 */
2700 	error = g_access(cp, 1, 1, 0);
2701 	if (error != 0) {
2702 		g_detach(cp);
2703 		g_destroy_consumer(cp);
2704 		return (error);
2705 	}
2706 	nblks = pp->mediasize / DEV_BSIZE;
2707 	swaponsomething(vp, cp, nblks, swapgeom_strategy,
2708 	    swapgeom_close, dev2udev(dev),
2709 	    (pp->flags & G_PF_ACCEPT_UNMAPPED) != 0 ? SW_UNMAPPED : 0);
2710 	return (0);
2711 }
2712 
2713 static int
2714 swapongeom(struct vnode *vp)
2715 {
2716 	int error;
2717 
2718 	vn_lock(vp, LK_EXCLUSIVE | LK_RETRY);
2719 	if (vp->v_type != VCHR || (vp->v_iflag & VI_DOOMED) != 0) {
2720 		error = ENOENT;
2721 	} else {
2722 		g_topology_lock();
2723 		error = swapongeom_locked(vp->v_rdev, vp);
2724 		g_topology_unlock();
2725 	}
2726 	VOP_UNLOCK(vp, 0);
2727 	return (error);
2728 }
2729 
2730 /*
2731  * VNODE backend
2732  *
2733  * This is used mainly for network filesystem (read: probably only tested
2734  * with NFS) swapfiles.
2735  *
2736  */
2737 
2738 static void
2739 swapdev_strategy(struct buf *bp, struct swdevt *sp)
2740 {
2741 	struct vnode *vp2;
2742 
2743 	bp->b_blkno = ctodb(bp->b_blkno - sp->sw_first);
2744 
2745 	vp2 = sp->sw_id;
2746 	vhold(vp2);
2747 	if (bp->b_iocmd == BIO_WRITE) {
2748 		if (bp->b_bufobj)
2749 			bufobj_wdrop(bp->b_bufobj);
2750 		bufobj_wref(&vp2->v_bufobj);
2751 	}
2752 	if (bp->b_bufobj != &vp2->v_bufobj)
2753 		bp->b_bufobj = &vp2->v_bufobj;
2754 	bp->b_vp = vp2;
2755 	bp->b_iooffset = dbtob(bp->b_blkno);
2756 	bstrategy(bp);
2757 	return;
2758 }
2759 
2760 static void
2761 swapdev_close(struct thread *td, struct swdevt *sp)
2762 {
2763 
2764 	VOP_CLOSE(sp->sw_vp, FREAD | FWRITE, td->td_ucred, td);
2765 	vrele(sp->sw_vp);
2766 }
2767 
2768 
2769 static int
2770 swaponvp(struct thread *td, struct vnode *vp, u_long nblks)
2771 {
2772 	struct swdevt *sp;
2773 	int error;
2774 
2775 	if (nblks == 0)
2776 		return (ENXIO);
2777 	mtx_lock(&sw_dev_mtx);
2778 	TAILQ_FOREACH(sp, &swtailq, sw_list) {
2779 		if (sp->sw_id == vp) {
2780 			mtx_unlock(&sw_dev_mtx);
2781 			return (EBUSY);
2782 		}
2783 	}
2784 	mtx_unlock(&sw_dev_mtx);
2785 
2786 	(void) vn_lock(vp, LK_EXCLUSIVE | LK_RETRY);
2787 #ifdef MAC
2788 	error = mac_system_check_swapon(td->td_ucred, vp);
2789 	if (error == 0)
2790 #endif
2791 		error = VOP_OPEN(vp, FREAD | FWRITE, td->td_ucred, td, NULL);
2792 	(void) VOP_UNLOCK(vp, 0);
2793 	if (error)
2794 		return (error);
2795 
2796 	swaponsomething(vp, vp, nblks, swapdev_strategy, swapdev_close,
2797 	    NODEV, 0);
2798 	return (0);
2799 }
2800 
2801 static int
2802 sysctl_swap_async_max(SYSCTL_HANDLER_ARGS)
2803 {
2804 	int error, new, n;
2805 
2806 	new = nsw_wcount_async_max;
2807 	error = sysctl_handle_int(oidp, &new, 0, req);
2808 	if (error != 0 || req->newptr == NULL)
2809 		return (error);
2810 
2811 	if (new > nswbuf / 2 || new < 1)
2812 		return (EINVAL);
2813 
2814 	mtx_lock(&pbuf_mtx);
2815 	while (nsw_wcount_async_max != new) {
2816 		/*
2817 		 * Adjust difference.  If the current async count is too low,
2818 		 * we will need to sqeeze our update slowly in.  Sleep with a
2819 		 * higher priority than getpbuf() to finish faster.
2820 		 */
2821 		n = new - nsw_wcount_async_max;
2822 		if (nsw_wcount_async + n >= 0) {
2823 			nsw_wcount_async += n;
2824 			nsw_wcount_async_max += n;
2825 			wakeup(&nsw_wcount_async);
2826 		} else {
2827 			nsw_wcount_async_max -= nsw_wcount_async;
2828 			nsw_wcount_async = 0;
2829 			msleep(&nsw_wcount_async, &pbuf_mtx, PSWP,
2830 			    "swpsysctl", 0);
2831 		}
2832 	}
2833 	mtx_unlock(&pbuf_mtx);
2834 
2835 	return (0);
2836 }
2837