xref: /freebsd/sys/vm/swap_pager.c (revision 396c556d77189a5c474d35cec6f44a762e310b7d)
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_compat.h"
75 #include "opt_swap.h"
76 #include "opt_vm.h"
77 
78 #include <sys/param.h>
79 #include <sys/systm.h>
80 #include <sys/conf.h>
81 #include <sys/kernel.h>
82 #include <sys/priv.h>
83 #include <sys/proc.h>
84 #include <sys/bio.h>
85 #include <sys/buf.h>
86 #include <sys/disk.h>
87 #include <sys/fcntl.h>
88 #include <sys/mount.h>
89 #include <sys/namei.h>
90 #include <sys/vnode.h>
91 #include <sys/malloc.h>
92 #include <sys/pctrie.h>
93 #include <sys/racct.h>
94 #include <sys/resource.h>
95 #include <sys/resourcevar.h>
96 #include <sys/rwlock.h>
97 #include <sys/sbuf.h>
98 #include <sys/sysctl.h>
99 #include <sys/sysproto.h>
100 #include <sys/blist.h>
101 #include <sys/lock.h>
102 #include <sys/sx.h>
103 #include <sys/vmmeter.h>
104 
105 #include <security/mac/mac_framework.h>
106 
107 #include <vm/vm.h>
108 #include <vm/pmap.h>
109 #include <vm/vm_map.h>
110 #include <vm/vm_kern.h>
111 #include <vm/vm_object.h>
112 #include <vm/vm_page.h>
113 #include <vm/vm_pager.h>
114 #include <vm/vm_pageout.h>
115 #include <vm/vm_param.h>
116 #include <vm/swap_pager.h>
117 #include <vm/vm_extern.h>
118 #include <vm/uma.h>
119 
120 #include <geom/geom.h>
121 
122 /*
123  * MAX_PAGEOUT_CLUSTER must be a power of 2 between 1 and 64.
124  * The 64-page limit is due to the radix code (kern/subr_blist.c).
125  */
126 #ifndef MAX_PAGEOUT_CLUSTER
127 #define	MAX_PAGEOUT_CLUSTER	32
128 #endif
129 
130 #if !defined(SWB_NPAGES)
131 #define SWB_NPAGES	MAX_PAGEOUT_CLUSTER
132 #endif
133 
134 #define	SWAP_META_PAGES		PCTRIE_COUNT
135 
136 /*
137  * A swblk structure maps each page index within a
138  * SWAP_META_PAGES-aligned and sized range to the address of an
139  * on-disk swap block (or SWAPBLK_NONE). The collection of these
140  * mappings for an entire vm object is implemented as a pc-trie.
141  */
142 struct swblk {
143 	vm_pindex_t	p;
144 	daddr_t		d[SWAP_META_PAGES];
145 };
146 
147 static MALLOC_DEFINE(M_VMPGDATA, "vm_pgdata", "swap pager private data");
148 static struct mtx sw_dev_mtx;
149 static TAILQ_HEAD(, swdevt) swtailq = TAILQ_HEAD_INITIALIZER(swtailq);
150 static struct swdevt *swdevhd;	/* Allocate from here next */
151 static int nswapdev;		/* Number of swap devices */
152 int swap_pager_avail;
153 static struct sx swdev_syscall_lock;	/* serialize swap(on|off) */
154 
155 static vm_ooffset_t swap_total;
156 SYSCTL_QUAD(_vm, OID_AUTO, swap_total, CTLFLAG_RD, &swap_total, 0,
157     "Total amount of available swap storage.");
158 static vm_ooffset_t swap_reserved;
159 SYSCTL_QUAD(_vm, OID_AUTO, swap_reserved, CTLFLAG_RD, &swap_reserved, 0,
160     "Amount of swap storage needed to back all allocated anonymous memory.");
161 static int overcommit = 0;
162 SYSCTL_INT(_vm, VM_OVERCOMMIT, overcommit, CTLFLAG_RW, &overcommit, 0,
163     "Configure virtual memory overcommit behavior. See tuning(7) "
164     "for details.");
165 static unsigned long swzone;
166 SYSCTL_ULONG(_vm, OID_AUTO, swzone, CTLFLAG_RD, &swzone, 0,
167     "Actual size of swap metadata zone");
168 static unsigned long swap_maxpages;
169 SYSCTL_ULONG(_vm, OID_AUTO, swap_maxpages, CTLFLAG_RD, &swap_maxpages, 0,
170     "Maximum amount of swap supported");
171 
172 /* bits from overcommit */
173 #define	SWAP_RESERVE_FORCE_ON		(1 << 0)
174 #define	SWAP_RESERVE_RLIMIT_ON		(1 << 1)
175 #define	SWAP_RESERVE_ALLOW_NONWIRED	(1 << 2)
176 
177 int
178 swap_reserve(vm_ooffset_t incr)
179 {
180 
181 	return (swap_reserve_by_cred(incr, curthread->td_ucred));
182 }
183 
184 int
185 swap_reserve_by_cred(vm_ooffset_t incr, struct ucred *cred)
186 {
187 	vm_ooffset_t r, s;
188 	int res, error;
189 	static int curfail;
190 	static struct timeval lastfail;
191 	struct uidinfo *uip;
192 
193 	uip = cred->cr_ruidinfo;
194 
195 	if (incr & PAGE_MASK)
196 		panic("swap_reserve: & PAGE_MASK");
197 
198 #ifdef RACCT
199 	if (racct_enable) {
200 		PROC_LOCK(curproc);
201 		error = racct_add(curproc, RACCT_SWAP, incr);
202 		PROC_UNLOCK(curproc);
203 		if (error != 0)
204 			return (0);
205 	}
206 #endif
207 
208 	res = 0;
209 	mtx_lock(&sw_dev_mtx);
210 	r = swap_reserved + incr;
211 	if (overcommit & SWAP_RESERVE_ALLOW_NONWIRED) {
212 		s = vm_cnt.v_page_count - vm_cnt.v_free_reserved - vm_cnt.v_wire_count;
213 		s *= PAGE_SIZE;
214 	} else
215 		s = 0;
216 	s += swap_total;
217 	if ((overcommit & SWAP_RESERVE_FORCE_ON) == 0 || r <= s ||
218 	    (error = priv_check(curthread, PRIV_VM_SWAP_NOQUOTA)) == 0) {
219 		res = 1;
220 		swap_reserved = r;
221 	}
222 	mtx_unlock(&sw_dev_mtx);
223 
224 	if (res) {
225 		UIDINFO_VMSIZE_LOCK(uip);
226 		if ((overcommit & SWAP_RESERVE_RLIMIT_ON) != 0 &&
227 		    uip->ui_vmsize + incr > lim_cur(curthread, RLIMIT_SWAP) &&
228 		    priv_check(curthread, PRIV_VM_SWAP_NORLIMIT))
229 			res = 0;
230 		else
231 			uip->ui_vmsize += incr;
232 		UIDINFO_VMSIZE_UNLOCK(uip);
233 		if (!res) {
234 			mtx_lock(&sw_dev_mtx);
235 			swap_reserved -= incr;
236 			mtx_unlock(&sw_dev_mtx);
237 		}
238 	}
239 	if (!res && ppsratecheck(&lastfail, &curfail, 1)) {
240 		printf("uid %d, pid %d: swap reservation for %jd bytes failed\n",
241 		    uip->ui_uid, curproc->p_pid, incr);
242 	}
243 
244 #ifdef RACCT
245 	if (!res) {
246 		PROC_LOCK(curproc);
247 		racct_sub(curproc, RACCT_SWAP, incr);
248 		PROC_UNLOCK(curproc);
249 	}
250 #endif
251 
252 	return (res);
253 }
254 
255 void
256 swap_reserve_force(vm_ooffset_t incr)
257 {
258 	struct uidinfo *uip;
259 
260 	mtx_lock(&sw_dev_mtx);
261 	swap_reserved += incr;
262 	mtx_unlock(&sw_dev_mtx);
263 
264 #ifdef RACCT
265 	PROC_LOCK(curproc);
266 	racct_add_force(curproc, RACCT_SWAP, incr);
267 	PROC_UNLOCK(curproc);
268 #endif
269 
270 	uip = curthread->td_ucred->cr_ruidinfo;
271 	PROC_LOCK(curproc);
272 	UIDINFO_VMSIZE_LOCK(uip);
273 	uip->ui_vmsize += incr;
274 	UIDINFO_VMSIZE_UNLOCK(uip);
275 	PROC_UNLOCK(curproc);
276 }
277 
278 void
279 swap_release(vm_ooffset_t decr)
280 {
281 	struct ucred *cred;
282 
283 	PROC_LOCK(curproc);
284 	cred = curthread->td_ucred;
285 	swap_release_by_cred(decr, cred);
286 	PROC_UNLOCK(curproc);
287 }
288 
289 void
290 swap_release_by_cred(vm_ooffset_t decr, struct ucred *cred)
291 {
292  	struct uidinfo *uip;
293 
294 	uip = cred->cr_ruidinfo;
295 
296 	if (decr & PAGE_MASK)
297 		panic("swap_release: & PAGE_MASK");
298 
299 	mtx_lock(&sw_dev_mtx);
300 	if (swap_reserved < decr)
301 		panic("swap_reserved < decr");
302 	swap_reserved -= decr;
303 	mtx_unlock(&sw_dev_mtx);
304 
305 	UIDINFO_VMSIZE_LOCK(uip);
306 	if (uip->ui_vmsize < decr)
307 		printf("negative vmsize for uid = %d\n", uip->ui_uid);
308 	uip->ui_vmsize -= decr;
309 	UIDINFO_VMSIZE_UNLOCK(uip);
310 
311 	racct_sub_cred(cred, RACCT_SWAP, decr);
312 }
313 
314 #define SWM_FREE	0x02	/* free, period			*/
315 #define SWM_POP		0x04	/* pop out			*/
316 
317 static int swap_pager_full = 2;	/* swap space exhaustion (task killing) */
318 static int swap_pager_almost_full = 1; /* swap space exhaustion (w/hysteresis)*/
319 static int nsw_rcount;		/* free read buffers			*/
320 static int nsw_wcount_sync;	/* limit write buffers / synchronous	*/
321 static int nsw_wcount_async;	/* limit write buffers / asynchronous	*/
322 static int nsw_wcount_async_max;/* assigned maximum			*/
323 static int nsw_cluster_max;	/* maximum VOP I/O allowed		*/
324 
325 static int sysctl_swap_async_max(SYSCTL_HANDLER_ARGS);
326 SYSCTL_PROC(_vm, OID_AUTO, swap_async_max, CTLTYPE_INT | CTLFLAG_RW |
327     CTLFLAG_MPSAFE, NULL, 0, sysctl_swap_async_max, "I",
328     "Maximum running async swap ops");
329 static int sysctl_swap_fragmentation(SYSCTL_HANDLER_ARGS);
330 SYSCTL_PROC(_vm, OID_AUTO, swap_fragmentation, CTLTYPE_STRING | CTLFLAG_RD |
331     CTLFLAG_MPSAFE, NULL, 0, sysctl_swap_fragmentation, "A",
332     "Swap Fragmentation Info");
333 
334 static struct sx sw_alloc_sx;
335 
336 /*
337  * "named" and "unnamed" anon region objects.  Try to reduce the overhead
338  * of searching a named list by hashing it just a little.
339  */
340 
341 #define NOBJLISTS		8
342 
343 #define NOBJLIST(handle)	\
344 	(&swap_pager_object_list[((int)(intptr_t)handle >> 4) & (NOBJLISTS-1)])
345 
346 static struct pagerlst	swap_pager_object_list[NOBJLISTS];
347 static uma_zone_t swblk_zone;
348 static uma_zone_t swpctrie_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 void swp_pager_meta_build(vm_object_t, vm_pindex_t, daddr_t);
407 static void swp_pager_meta_free(vm_object_t, vm_pindex_t, vm_pindex_t);
408 static void swp_pager_meta_free_all(vm_object_t);
409 static daddr_t swp_pager_meta_ctl(vm_object_t, vm_pindex_t, int);
410 
411 static void *
412 swblk_trie_alloc(struct pctrie *ptree)
413 {
414 
415 	return (uma_zalloc(swpctrie_zone, M_NOWAIT | (curproc == pageproc ?
416 	    M_USE_RESERVE : 0)));
417 }
418 
419 static void
420 swblk_trie_free(struct pctrie *ptree, void *node)
421 {
422 
423 	uma_zfree(swpctrie_zone, node);
424 }
425 
426 PCTRIE_DEFINE(SWAP, swblk, p, swblk_trie_alloc, swblk_trie_free);
427 
428 /*
429  * SWP_SIZECHECK() -	update swap_pager_full indication
430  *
431  *	update the swap_pager_almost_full indication and warn when we are
432  *	about to run out of swap space, using lowat/hiwat hysteresis.
433  *
434  *	Clear swap_pager_full ( task killing ) indication when lowat is met.
435  *
436  *	No restrictions on call
437  *	This routine may not block.
438  */
439 static void
440 swp_sizecheck(void)
441 {
442 
443 	if (swap_pager_avail < nswap_lowat) {
444 		if (swap_pager_almost_full == 0) {
445 			printf("swap_pager: out of swap space\n");
446 			swap_pager_almost_full = 1;
447 		}
448 	} else {
449 		swap_pager_full = 0;
450 		if (swap_pager_avail > nswap_hiwat)
451 			swap_pager_almost_full = 0;
452 	}
453 }
454 
455 /*
456  * SWAP_PAGER_INIT() -	initialize the swap pager!
457  *
458  *	Expected to be started from system init.  NOTE:  This code is run
459  *	before much else so be careful what you depend on.  Most of the VM
460  *	system has yet to be initialized at this point.
461  */
462 static void
463 swap_pager_init(void)
464 {
465 	/*
466 	 * Initialize object lists
467 	 */
468 	int i;
469 
470 	for (i = 0; i < NOBJLISTS; ++i)
471 		TAILQ_INIT(&swap_pager_object_list[i]);
472 	mtx_init(&sw_dev_mtx, "swapdev", NULL, MTX_DEF);
473 	sx_init(&sw_alloc_sx, "swspsx");
474 	sx_init(&swdev_syscall_lock, "swsysc");
475 }
476 
477 /*
478  * SWAP_PAGER_SWAP_INIT() - swap pager initialization from pageout process
479  *
480  *	Expected to be started from pageout process once, prior to entering
481  *	its main loop.
482  */
483 void
484 swap_pager_swap_init(void)
485 {
486 	unsigned long n, n2;
487 
488 	/*
489 	 * Number of in-transit swap bp operations.  Don't
490 	 * exhaust the pbufs completely.  Make sure we
491 	 * initialize workable values (0 will work for hysteresis
492 	 * but it isn't very efficient).
493 	 *
494 	 * The nsw_cluster_max is constrained by the bp->b_pages[]
495 	 * array (MAXPHYS/PAGE_SIZE) and our locally defined
496 	 * MAX_PAGEOUT_CLUSTER.   Also be aware that swap ops are
497 	 * constrained by the swap device interleave stripe size.
498 	 *
499 	 * Currently we hardwire nsw_wcount_async to 4.  This limit is
500 	 * designed to prevent other I/O from having high latencies due to
501 	 * our pageout I/O.  The value 4 works well for one or two active swap
502 	 * devices but is probably a little low if you have more.  Even so,
503 	 * a higher value would probably generate only a limited improvement
504 	 * with three or four active swap devices since the system does not
505 	 * typically have to pageout at extreme bandwidths.   We will want
506 	 * at least 2 per swap devices, and 4 is a pretty good value if you
507 	 * have one NFS swap device due to the command/ack latency over NFS.
508 	 * So it all works out pretty well.
509 	 */
510 	nsw_cluster_max = min((MAXPHYS/PAGE_SIZE), MAX_PAGEOUT_CLUSTER);
511 
512 	mtx_lock(&pbuf_mtx);
513 	nsw_rcount = (nswbuf + 1) / 2;
514 	nsw_wcount_sync = (nswbuf + 3) / 4;
515 	nsw_wcount_async = 4;
516 	nsw_wcount_async_max = nsw_wcount_async;
517 	mtx_unlock(&pbuf_mtx);
518 
519 	/*
520 	 * Initialize our zone, guessing on the number we need based
521 	 * on the number of pages in the system.
522 	 */
523 	n = vm_cnt.v_page_count / 2;
524 	if (maxswzone && n > maxswzone / sizeof(struct swblk))
525 		n = maxswzone / sizeof(struct swblk);
526 	swpctrie_zone = uma_zcreate("swpctrie", pctrie_node_size(), NULL, NULL,
527 	    pctrie_zone_init, NULL, UMA_ALIGN_PTR,
528 	    UMA_ZONE_NOFREE | UMA_ZONE_VM);
529 	if (swpctrie_zone == NULL)
530 		panic("failed to create swap pctrie zone.");
531 	swblk_zone = uma_zcreate("swblk", sizeof(struct swblk), NULL, NULL,
532 	    NULL, NULL, _Alignof(struct swblk) - 1,
533 	    UMA_ZONE_NOFREE | UMA_ZONE_VM);
534 	if (swblk_zone == NULL)
535 		panic("failed to create swap blk zone.");
536 	n2 = n;
537 	do {
538 		if (uma_zone_reserve_kva(swblk_zone, n))
539 			break;
540 		/*
541 		 * if the allocation failed, try a zone two thirds the
542 		 * size of the previous attempt.
543 		 */
544 		n -= ((n + 2) / 3);
545 	} while (n > 0);
546 
547 	/*
548 	 * Often uma_zone_reserve_kva() cannot reserve exactly the
549 	 * requested size.  Account for the difference when
550 	 * calculating swap_maxpages.
551 	 */
552 	n = uma_zone_get_max(swblk_zone);
553 
554 	if (n < n2)
555 		printf("Swap blk zone entries reduced from %lu to %lu.\n",
556 		    n2, n);
557 	swap_maxpages = n * SWAP_META_PAGES;
558 	swzone = n * sizeof(struct swblk);
559 	if (!uma_zone_reserve_kva(swpctrie_zone, n))
560 		printf("Cannot reserve swap pctrie zone, "
561 		    "reduce kern.maxswzone.\n");
562 }
563 
564 static vm_object_t
565 swap_pager_alloc_init(void *handle, struct ucred *cred, vm_ooffset_t size,
566     vm_ooffset_t offset)
567 {
568 	vm_object_t object;
569 
570 	if (cred != NULL) {
571 		if (!swap_reserve_by_cred(size, cred))
572 			return (NULL);
573 		crhold(cred);
574 	}
575 
576 	/*
577 	 * The un_pager.swp.swp_blks trie is initialized by
578 	 * vm_object_allocate() to ensure the correct order of
579 	 * visibility to other threads.
580 	 */
581 	object = vm_object_allocate(OBJT_SWAP, OFF_TO_IDX(offset +
582 	    PAGE_MASK + size));
583 
584 	object->handle = handle;
585 	if (cred != NULL) {
586 		object->cred = cred;
587 		object->charge = size;
588 	}
589 	return (object);
590 }
591 
592 /*
593  * SWAP_PAGER_ALLOC() -	allocate a new OBJT_SWAP VM object and instantiate
594  *			its metadata structures.
595  *
596  *	This routine is called from the mmap and fork code to create a new
597  *	OBJT_SWAP object.
598  *
599  *	This routine must ensure that no live duplicate is created for
600  *	the named object request, which is protected against by
601  *	holding the sw_alloc_sx lock in case handle != NULL.
602  */
603 static vm_object_t
604 swap_pager_alloc(void *handle, vm_ooffset_t size, vm_prot_t prot,
605     vm_ooffset_t offset, struct ucred *cred)
606 {
607 	vm_object_t object;
608 
609 	if (handle != NULL) {
610 		/*
611 		 * Reference existing named region or allocate new one.  There
612 		 * should not be a race here against swp_pager_meta_build()
613 		 * as called from vm_page_remove() in regards to the lookup
614 		 * of the handle.
615 		 */
616 		sx_xlock(&sw_alloc_sx);
617 		object = vm_pager_object_lookup(NOBJLIST(handle), handle);
618 		if (object == NULL) {
619 			object = swap_pager_alloc_init(handle, cred, size,
620 			    offset);
621 			if (object != NULL) {
622 				TAILQ_INSERT_TAIL(NOBJLIST(object->handle),
623 				    object, pager_object_list);
624 			}
625 		}
626 		sx_xunlock(&sw_alloc_sx);
627 	} else {
628 		object = swap_pager_alloc_init(handle, cred, size, offset);
629 	}
630 	return (object);
631 }
632 
633 /*
634  * SWAP_PAGER_DEALLOC() -	remove swap metadata from object
635  *
636  *	The swap backing for the object is destroyed.  The code is
637  *	designed such that we can reinstantiate it later, but this
638  *	routine is typically called only when the entire object is
639  *	about to be destroyed.
640  *
641  *	The object must be locked.
642  */
643 static void
644 swap_pager_dealloc(vm_object_t object)
645 {
646 
647 	VM_OBJECT_ASSERT_WLOCKED(object);
648 	KASSERT((object->flags & OBJ_DEAD) != 0, ("dealloc of reachable obj"));
649 
650 	/*
651 	 * Remove from list right away so lookups will fail if we block for
652 	 * pageout completion.
653 	 */
654 	if (object->handle != NULL) {
655 		VM_OBJECT_WUNLOCK(object);
656 		sx_xlock(&sw_alloc_sx);
657 		TAILQ_REMOVE(NOBJLIST(object->handle), object,
658 		    pager_object_list);
659 		sx_xunlock(&sw_alloc_sx);
660 		VM_OBJECT_WLOCK(object);
661 	}
662 
663 	vm_object_pip_wait(object, "swpdea");
664 
665 	/*
666 	 * Free all remaining metadata.  We only bother to free it from
667 	 * the swap meta data.  We do not attempt to free swapblk's still
668 	 * associated with vm_page_t's for this object.  We do not care
669 	 * if paging is still in progress on some objects.
670 	 */
671 	swp_pager_meta_free_all(object);
672 	object->handle = NULL;
673 	object->type = OBJT_DEAD;
674 }
675 
676 /************************************************************************
677  *			SWAP PAGER BITMAP ROUTINES			*
678  ************************************************************************/
679 
680 /*
681  * SWP_PAGER_GETSWAPSPACE() -	allocate raw swap space
682  *
683  *	Allocate swap for the requested number of pages.  The starting
684  *	swap block number (a page index) is returned or SWAPBLK_NONE
685  *	if the allocation failed.
686  *
687  *	Also has the side effect of advising that somebody made a mistake
688  *	when they configured swap and didn't configure enough.
689  *
690  *	This routine may not sleep.
691  *
692  *	We allocate in round-robin fashion from the configured devices.
693  */
694 static daddr_t
695 swp_pager_getswapspace(int npages)
696 {
697 	daddr_t blk;
698 	struct swdevt *sp;
699 	int i;
700 
701 	blk = SWAPBLK_NONE;
702 	mtx_lock(&sw_dev_mtx);
703 	sp = swdevhd;
704 	for (i = 0; i < nswapdev; i++) {
705 		if (sp == NULL)
706 			sp = TAILQ_FIRST(&swtailq);
707 		if (!(sp->sw_flags & SW_CLOSING)) {
708 			blk = blist_alloc(sp->sw_blist, npages);
709 			if (blk != SWAPBLK_NONE) {
710 				blk += sp->sw_first;
711 				sp->sw_used += npages;
712 				swap_pager_avail -= npages;
713 				swp_sizecheck();
714 				swdevhd = TAILQ_NEXT(sp, sw_list);
715 				goto done;
716 			}
717 		}
718 		sp = TAILQ_NEXT(sp, sw_list);
719 	}
720 	if (swap_pager_full != 2) {
721 		printf("swap_pager_getswapspace(%d): failed\n", npages);
722 		swap_pager_full = 2;
723 		swap_pager_almost_full = 1;
724 	}
725 	swdevhd = NULL;
726 done:
727 	mtx_unlock(&sw_dev_mtx);
728 	return (blk);
729 }
730 
731 static int
732 swp_pager_isondev(daddr_t blk, struct swdevt *sp)
733 {
734 
735 	return (blk >= sp->sw_first && blk < sp->sw_end);
736 }
737 
738 static void
739 swp_pager_strategy(struct buf *bp)
740 {
741 	struct swdevt *sp;
742 
743 	mtx_lock(&sw_dev_mtx);
744 	TAILQ_FOREACH(sp, &swtailq, sw_list) {
745 		if (bp->b_blkno >= sp->sw_first && bp->b_blkno < sp->sw_end) {
746 			mtx_unlock(&sw_dev_mtx);
747 			if ((sp->sw_flags & SW_UNMAPPED) != 0 &&
748 			    unmapped_buf_allowed) {
749 				bp->b_data = unmapped_buf;
750 				bp->b_offset = 0;
751 			} else {
752 				pmap_qenter((vm_offset_t)bp->b_data,
753 				    &bp->b_pages[0], bp->b_bcount / PAGE_SIZE);
754 			}
755 			sp->sw_strategy(bp, sp);
756 			return;
757 		}
758 	}
759 	panic("Swapdev not found");
760 }
761 
762 
763 /*
764  * SWP_PAGER_FREESWAPSPACE() -	free raw swap space
765  *
766  *	This routine returns the specified swap blocks back to the bitmap.
767  *
768  *	This routine may not sleep.
769  */
770 static void
771 swp_pager_freeswapspace(daddr_t blk, int npages)
772 {
773 	struct swdevt *sp;
774 
775 	mtx_lock(&sw_dev_mtx);
776 	TAILQ_FOREACH(sp, &swtailq, sw_list) {
777 		if (blk >= sp->sw_first && blk < sp->sw_end) {
778 			sp->sw_used -= npages;
779 			/*
780 			 * If we are attempting to stop swapping on
781 			 * this device, we don't want to mark any
782 			 * blocks free lest they be reused.
783 			 */
784 			if ((sp->sw_flags & SW_CLOSING) == 0) {
785 				blist_free(sp->sw_blist, blk - sp->sw_first,
786 				    npages);
787 				swap_pager_avail += npages;
788 				swp_sizecheck();
789 			}
790 			mtx_unlock(&sw_dev_mtx);
791 			return;
792 		}
793 	}
794 	panic("Swapdev not found");
795 }
796 
797 /*
798  * SYSCTL_SWAP_FRAGMENTATION() -	produce raw swap space stats
799  */
800 static int
801 sysctl_swap_fragmentation(SYSCTL_HANDLER_ARGS)
802 {
803 	struct sbuf sbuf;
804 	struct swdevt *sp;
805 	const char *devname;
806 	int error;
807 
808 	error = sysctl_wire_old_buffer(req, 0);
809 	if (error != 0)
810 		return (error);
811 	sbuf_new_for_sysctl(&sbuf, NULL, 128, req);
812 	mtx_lock(&sw_dev_mtx);
813 	TAILQ_FOREACH(sp, &swtailq, sw_list) {
814 		if (vn_isdisk(sp->sw_vp, NULL))
815 			devname = devtoname(sp->sw_vp->v_rdev);
816 		else
817 			devname = "[file]";
818 		sbuf_printf(&sbuf, "\nFree space on device %s:\n", devname);
819 		blist_stats(sp->sw_blist, &sbuf);
820 	}
821 	mtx_unlock(&sw_dev_mtx);
822 	error = sbuf_finish(&sbuf);
823 	sbuf_delete(&sbuf);
824 	return (error);
825 }
826 
827 /*
828  * SWAP_PAGER_FREESPACE() -	frees swap blocks associated with a page
829  *				range within an object.
830  *
831  *	This is a globally accessible routine.
832  *
833  *	This routine removes swapblk assignments from swap metadata.
834  *
835  *	The external callers of this routine typically have already destroyed
836  *	or renamed vm_page_t's associated with this range in the object so
837  *	we should be ok.
838  *
839  *	The object must be locked.
840  */
841 void
842 swap_pager_freespace(vm_object_t object, vm_pindex_t start, vm_size_t size)
843 {
844 
845 	swp_pager_meta_free(object, start, size);
846 }
847 
848 /*
849  * SWAP_PAGER_RESERVE() - reserve swap blocks in object
850  *
851  *	Assigns swap blocks to the specified range within the object.  The
852  *	swap blocks are not zeroed.  Any previous swap assignment is destroyed.
853  *
854  *	Returns 0 on success, -1 on failure.
855  */
856 int
857 swap_pager_reserve(vm_object_t object, vm_pindex_t start, vm_size_t size)
858 {
859 	int n = 0;
860 	daddr_t blk = SWAPBLK_NONE;
861 	vm_pindex_t beg = start;	/* save start index */
862 
863 	VM_OBJECT_WLOCK(object);
864 	while (size) {
865 		if (n == 0) {
866 			n = BLIST_MAX_ALLOC;
867 			while ((blk = swp_pager_getswapspace(n)) == SWAPBLK_NONE) {
868 				n >>= 1;
869 				if (n == 0) {
870 					swp_pager_meta_free(object, beg, start - beg);
871 					VM_OBJECT_WUNLOCK(object);
872 					return (-1);
873 				}
874 			}
875 		}
876 		swp_pager_meta_build(object, start, blk);
877 		--size;
878 		++start;
879 		++blk;
880 		--n;
881 	}
882 	swp_pager_meta_free(object, start, n);
883 	VM_OBJECT_WUNLOCK(object);
884 	return (0);
885 }
886 
887 /*
888  * SWAP_PAGER_COPY() -  copy blocks from source pager to destination pager
889  *			and destroy the source.
890  *
891  *	Copy any valid swapblks from the source to the destination.  In
892  *	cases where both the source and destination have a valid swapblk,
893  *	we keep the destination's.
894  *
895  *	This routine is allowed to sleep.  It may sleep allocating metadata
896  *	indirectly through swp_pager_meta_build() or if paging is still in
897  *	progress on the source.
898  *
899  *	The source object contains no vm_page_t's (which is just as well)
900  *
901  *	The source object is of type OBJT_SWAP.
902  *
903  *	The source and destination objects must be locked.
904  *	Both object locks may temporarily be released.
905  */
906 void
907 swap_pager_copy(vm_object_t srcobject, vm_object_t dstobject,
908     vm_pindex_t offset, int destroysource)
909 {
910 	vm_pindex_t i;
911 
912 	VM_OBJECT_ASSERT_WLOCKED(srcobject);
913 	VM_OBJECT_ASSERT_WLOCKED(dstobject);
914 
915 	/*
916 	 * If destroysource is set, we remove the source object from the
917 	 * swap_pager internal queue now.
918 	 */
919 	if (destroysource && srcobject->handle != NULL) {
920 		vm_object_pip_add(srcobject, 1);
921 		VM_OBJECT_WUNLOCK(srcobject);
922 		vm_object_pip_add(dstobject, 1);
923 		VM_OBJECT_WUNLOCK(dstobject);
924 		sx_xlock(&sw_alloc_sx);
925 		TAILQ_REMOVE(NOBJLIST(srcobject->handle), srcobject,
926 		    pager_object_list);
927 		sx_xunlock(&sw_alloc_sx);
928 		VM_OBJECT_WLOCK(dstobject);
929 		vm_object_pip_wakeup(dstobject);
930 		VM_OBJECT_WLOCK(srcobject);
931 		vm_object_pip_wakeup(srcobject);
932 	}
933 
934 	/*
935 	 * transfer source to destination.
936 	 */
937 	for (i = 0; i < dstobject->size; ++i) {
938 		daddr_t dstaddr;
939 
940 		/*
941 		 * Locate (without changing) the swapblk on the destination,
942 		 * unless it is invalid in which case free it silently, or
943 		 * if the destination is a resident page, in which case the
944 		 * source is thrown away.
945 		 */
946 		dstaddr = swp_pager_meta_ctl(dstobject, i, 0);
947 
948 		if (dstaddr == SWAPBLK_NONE) {
949 			/*
950 			 * Destination has no swapblk and is not resident,
951 			 * copy source.
952 			 */
953 			daddr_t srcaddr;
954 
955 			srcaddr = swp_pager_meta_ctl(
956 			    srcobject,
957 			    i + offset,
958 			    SWM_POP
959 			);
960 
961 			if (srcaddr != SWAPBLK_NONE) {
962 				/*
963 				 * swp_pager_meta_build() can sleep.
964 				 */
965 				vm_object_pip_add(srcobject, 1);
966 				VM_OBJECT_WUNLOCK(srcobject);
967 				vm_object_pip_add(dstobject, 1);
968 				swp_pager_meta_build(dstobject, i, srcaddr);
969 				vm_object_pip_wakeup(dstobject);
970 				VM_OBJECT_WLOCK(srcobject);
971 				vm_object_pip_wakeup(srcobject);
972 			}
973 		} else {
974 			/*
975 			 * Destination has valid swapblk or it is represented
976 			 * by a resident page.  We destroy the sourceblock.
977 			 */
978 
979 			swp_pager_meta_ctl(srcobject, i + offset, SWM_FREE);
980 		}
981 	}
982 
983 	/*
984 	 * Free left over swap blocks in source.
985 	 *
986 	 * We have to revert the type to OBJT_DEFAULT so we do not accidentally
987 	 * double-remove the object from the swap queues.
988 	 */
989 	if (destroysource) {
990 		swp_pager_meta_free_all(srcobject);
991 		/*
992 		 * Reverting the type is not necessary, the caller is going
993 		 * to destroy srcobject directly, but I'm doing it here
994 		 * for consistency since we've removed the object from its
995 		 * queues.
996 		 */
997 		srcobject->type = OBJT_DEFAULT;
998 	}
999 }
1000 
1001 /*
1002  * SWAP_PAGER_HASPAGE() -	determine if we have good backing store for
1003  *				the requested page.
1004  *
1005  *	We determine whether good backing store exists for the requested
1006  *	page and return TRUE if it does, FALSE if it doesn't.
1007  *
1008  *	If TRUE, we also try to determine how much valid, contiguous backing
1009  *	store exists before and after the requested page.
1010  */
1011 static boolean_t
1012 swap_pager_haspage(vm_object_t object, vm_pindex_t pindex, int *before,
1013     int *after)
1014 {
1015 	daddr_t blk, blk0;
1016 	int i;
1017 
1018 	VM_OBJECT_ASSERT_LOCKED(object);
1019 
1020 	/*
1021 	 * do we have good backing store at the requested index ?
1022 	 */
1023 	blk0 = swp_pager_meta_ctl(object, pindex, 0);
1024 	if (blk0 == SWAPBLK_NONE) {
1025 		if (before)
1026 			*before = 0;
1027 		if (after)
1028 			*after = 0;
1029 		return (FALSE);
1030 	}
1031 
1032 	/*
1033 	 * find backwards-looking contiguous good backing store
1034 	 */
1035 	if (before != NULL) {
1036 		for (i = 1; i < SWB_NPAGES; i++) {
1037 			if (i > pindex)
1038 				break;
1039 			blk = swp_pager_meta_ctl(object, pindex - i, 0);
1040 			if (blk != blk0 - i)
1041 				break;
1042 		}
1043 		*before = i - 1;
1044 	}
1045 
1046 	/*
1047 	 * find forward-looking contiguous good backing store
1048 	 */
1049 	if (after != NULL) {
1050 		for (i = 1; i < SWB_NPAGES; i++) {
1051 			blk = swp_pager_meta_ctl(object, pindex + i, 0);
1052 			if (blk != blk0 + i)
1053 				break;
1054 		}
1055 		*after = i - 1;
1056 	}
1057 	return (TRUE);
1058 }
1059 
1060 /*
1061  * SWAP_PAGER_PAGE_UNSWAPPED() - remove swap backing store related to page
1062  *
1063  *	This removes any associated swap backing store, whether valid or
1064  *	not, from the page.
1065  *
1066  *	This routine is typically called when a page is made dirty, at
1067  *	which point any associated swap can be freed.  MADV_FREE also
1068  *	calls us in a special-case situation
1069  *
1070  *	NOTE!!!  If the page is clean and the swap was valid, the caller
1071  *	should make the page dirty before calling this routine.  This routine
1072  *	does NOT change the m->dirty status of the page.  Also: MADV_FREE
1073  *	depends on it.
1074  *
1075  *	This routine may not sleep.
1076  *
1077  *	The object containing the page must be locked.
1078  */
1079 static void
1080 swap_pager_unswapped(vm_page_t m)
1081 {
1082 
1083 	swp_pager_meta_ctl(m->object, m->pindex, SWM_FREE);
1084 }
1085 
1086 /*
1087  * swap_pager_getpages() - bring pages in from swap
1088  *
1089  *	Attempt to page in the pages in array "m" of length "count".  The caller
1090  *	may optionally specify that additional pages preceding and succeeding
1091  *	the specified range be paged in.  The number of such pages is returned
1092  *	in the "rbehind" and "rahead" parameters, and they will be in the
1093  *	inactive queue upon return.
1094  *
1095  *	The pages in "m" must be busied and will remain busied upon return.
1096  */
1097 static int
1098 swap_pager_getpages(vm_object_t object, vm_page_t *m, int count, int *rbehind,
1099     int *rahead)
1100 {
1101 	struct buf *bp;
1102 	vm_page_t mpred, msucc, p;
1103 	vm_pindex_t pindex;
1104 	daddr_t blk;
1105 	int i, j, maxahead, maxbehind, reqcount, shift;
1106 
1107 	reqcount = count;
1108 
1109 	VM_OBJECT_WUNLOCK(object);
1110 	bp = getpbuf(&nsw_rcount);
1111 	VM_OBJECT_WLOCK(object);
1112 
1113 	if (!swap_pager_haspage(object, m[0]->pindex, &maxbehind, &maxahead)) {
1114 		relpbuf(bp, &nsw_rcount);
1115 		return (VM_PAGER_FAIL);
1116 	}
1117 
1118 	/*
1119 	 * Clip the readahead and readbehind ranges to exclude resident pages.
1120 	 */
1121 	if (rahead != NULL) {
1122 		KASSERT(reqcount - 1 <= maxahead,
1123 		    ("page count %d extends beyond swap block", reqcount));
1124 		*rahead = imin(*rahead, maxahead - (reqcount - 1));
1125 		pindex = m[reqcount - 1]->pindex;
1126 		msucc = TAILQ_NEXT(m[reqcount - 1], listq);
1127 		if (msucc != NULL && msucc->pindex - pindex - 1 < *rahead)
1128 			*rahead = msucc->pindex - pindex - 1;
1129 	}
1130 	if (rbehind != NULL) {
1131 		*rbehind = imin(*rbehind, maxbehind);
1132 		pindex = m[0]->pindex;
1133 		mpred = TAILQ_PREV(m[0], pglist, listq);
1134 		if (mpred != NULL && pindex - mpred->pindex - 1 < *rbehind)
1135 			*rbehind = pindex - mpred->pindex - 1;
1136 	}
1137 
1138 	/*
1139 	 * Allocate readahead and readbehind pages.
1140 	 */
1141 	shift = rbehind != NULL ? *rbehind : 0;
1142 	if (shift != 0) {
1143 		for (i = 1; i <= shift; i++) {
1144 			p = vm_page_alloc(object, m[0]->pindex - i,
1145 			    VM_ALLOC_NORMAL);
1146 			if (p == NULL) {
1147 				/* Shift allocated pages to the left. */
1148 				for (j = 0; j < i - 1; j++)
1149 					bp->b_pages[j] =
1150 					    bp->b_pages[j + shift - i + 1];
1151 				break;
1152 			}
1153 			bp->b_pages[shift - i] = p;
1154 		}
1155 		shift = i - 1;
1156 		*rbehind = shift;
1157 	}
1158 	for (i = 0; i < reqcount; i++)
1159 		bp->b_pages[i + shift] = m[i];
1160 	if (rahead != NULL) {
1161 		for (i = 0; i < *rahead; i++) {
1162 			p = vm_page_alloc(object,
1163 			    m[reqcount - 1]->pindex + i + 1, VM_ALLOC_NORMAL);
1164 			if (p == NULL)
1165 				break;
1166 			bp->b_pages[shift + reqcount + i] = p;
1167 		}
1168 		*rahead = i;
1169 	}
1170 	if (rbehind != NULL)
1171 		count += *rbehind;
1172 	if (rahead != NULL)
1173 		count += *rahead;
1174 
1175 	vm_object_pip_add(object, count);
1176 
1177 	for (i = 0; i < count; i++)
1178 		bp->b_pages[i]->oflags |= VPO_SWAPINPROG;
1179 
1180 	pindex = bp->b_pages[0]->pindex;
1181 	blk = swp_pager_meta_ctl(object, pindex, 0);
1182 	KASSERT(blk != SWAPBLK_NONE,
1183 	    ("no swap blocking containing %p(%jx)", object, (uintmax_t)pindex));
1184 
1185 	VM_OBJECT_WUNLOCK(object);
1186 
1187 	bp->b_flags |= B_PAGING;
1188 	bp->b_iocmd = BIO_READ;
1189 	bp->b_iodone = swp_pager_async_iodone;
1190 	bp->b_rcred = crhold(thread0.td_ucred);
1191 	bp->b_wcred = crhold(thread0.td_ucred);
1192 	bp->b_blkno = blk;
1193 	bp->b_bcount = PAGE_SIZE * count;
1194 	bp->b_bufsize = PAGE_SIZE * count;
1195 	bp->b_npages = count;
1196 	bp->b_pgbefore = rbehind != NULL ? *rbehind : 0;
1197 	bp->b_pgafter = rahead != NULL ? *rahead : 0;
1198 
1199 	VM_CNT_INC(v_swapin);
1200 	VM_CNT_ADD(v_swappgsin, count);
1201 
1202 	/*
1203 	 * perform the I/O.  NOTE!!!  bp cannot be considered valid after
1204 	 * this point because we automatically release it on completion.
1205 	 * Instead, we look at the one page we are interested in which we
1206 	 * still hold a lock on even through the I/O completion.
1207 	 *
1208 	 * The other pages in our m[] array are also released on completion,
1209 	 * so we cannot assume they are valid anymore either.
1210 	 *
1211 	 * NOTE: b_blkno is destroyed by the call to swapdev_strategy
1212 	 */
1213 	BUF_KERNPROC(bp);
1214 	swp_pager_strategy(bp);
1215 
1216 	/*
1217 	 * Wait for the pages we want to complete.  VPO_SWAPINPROG is always
1218 	 * cleared on completion.  If an I/O error occurs, SWAPBLK_NONE
1219 	 * is set in the metadata for each page in the request.
1220 	 */
1221 	VM_OBJECT_WLOCK(object);
1222 	while ((m[0]->oflags & VPO_SWAPINPROG) != 0) {
1223 		m[0]->oflags |= VPO_SWAPSLEEP;
1224 		VM_CNT_INC(v_intrans);
1225 		if (VM_OBJECT_SLEEP(object, &object->paging_in_progress, PSWP,
1226 		    "swread", hz * 20)) {
1227 			printf(
1228 "swap_pager: indefinite wait buffer: bufobj: %p, blkno: %jd, size: %ld\n",
1229 			    bp->b_bufobj, (intmax_t)bp->b_blkno, bp->b_bcount);
1230 		}
1231 	}
1232 
1233 	/*
1234 	 * If we had an unrecoverable read error pages will not be valid.
1235 	 */
1236 	for (i = 0; i < reqcount; i++)
1237 		if (m[i]->valid != VM_PAGE_BITS_ALL)
1238 			return (VM_PAGER_ERROR);
1239 
1240 	return (VM_PAGER_OK);
1241 
1242 	/*
1243 	 * A final note: in a low swap situation, we cannot deallocate swap
1244 	 * and mark a page dirty here because the caller is likely to mark
1245 	 * the page clean when we return, causing the page to possibly revert
1246 	 * to all-zero's later.
1247 	 */
1248 }
1249 
1250 /*
1251  * 	swap_pager_getpages_async():
1252  *
1253  *	Right now this is emulation of asynchronous operation on top of
1254  *	swap_pager_getpages().
1255  */
1256 static int
1257 swap_pager_getpages_async(vm_object_t object, vm_page_t *m, int count,
1258     int *rbehind, int *rahead, pgo_getpages_iodone_t iodone, void *arg)
1259 {
1260 	int r, error;
1261 
1262 	r = swap_pager_getpages(object, m, count, rbehind, rahead);
1263 	VM_OBJECT_WUNLOCK(object);
1264 	switch (r) {
1265 	case VM_PAGER_OK:
1266 		error = 0;
1267 		break;
1268 	case VM_PAGER_ERROR:
1269 		error = EIO;
1270 		break;
1271 	case VM_PAGER_FAIL:
1272 		error = EINVAL;
1273 		break;
1274 	default:
1275 		panic("unhandled swap_pager_getpages() error %d", r);
1276 	}
1277 	(iodone)(arg, m, count, error);
1278 	VM_OBJECT_WLOCK(object);
1279 
1280 	return (r);
1281 }
1282 
1283 /*
1284  *	swap_pager_putpages:
1285  *
1286  *	Assign swap (if necessary) and initiate I/O on the specified pages.
1287  *
1288  *	We support both OBJT_DEFAULT and OBJT_SWAP objects.  DEFAULT objects
1289  *	are automatically converted to SWAP objects.
1290  *
1291  *	In a low memory situation we may block in VOP_STRATEGY(), but the new
1292  *	vm_page reservation system coupled with properly written VFS devices
1293  *	should ensure that no low-memory deadlock occurs.  This is an area
1294  *	which needs work.
1295  *
1296  *	The parent has N vm_object_pip_add() references prior to
1297  *	calling us and will remove references for rtvals[] that are
1298  *	not set to VM_PAGER_PEND.  We need to remove the rest on I/O
1299  *	completion.
1300  *
1301  *	The parent has soft-busy'd the pages it passes us and will unbusy
1302  *	those whos rtvals[] entry is not set to VM_PAGER_PEND on return.
1303  *	We need to unbusy the rest on I/O completion.
1304  */
1305 static void
1306 swap_pager_putpages(vm_object_t object, vm_page_t *m, int count,
1307     int flags, int *rtvals)
1308 {
1309 	int i, n;
1310 	boolean_t sync;
1311 
1312 	if (count && m[0]->object != object) {
1313 		panic("swap_pager_putpages: object mismatch %p/%p",
1314 		    object,
1315 		    m[0]->object
1316 		);
1317 	}
1318 
1319 	/*
1320 	 * Step 1
1321 	 *
1322 	 * Turn object into OBJT_SWAP
1323 	 * check for bogus sysops
1324 	 * force sync if not pageout process
1325 	 */
1326 	if (object->type != OBJT_SWAP)
1327 		swp_pager_meta_build(object, 0, SWAPBLK_NONE);
1328 	VM_OBJECT_WUNLOCK(object);
1329 
1330 	n = 0;
1331 	if (curproc != pageproc)
1332 		sync = TRUE;
1333 	else
1334 		sync = (flags & VM_PAGER_PUT_SYNC) != 0;
1335 
1336 	/*
1337 	 * Step 2
1338 	 *
1339 	 * Assign swap blocks and issue I/O.  We reallocate swap on the fly.
1340 	 * The page is left dirty until the pageout operation completes
1341 	 * successfully.
1342 	 */
1343 	for (i = 0; i < count; i += n) {
1344 		int j;
1345 		struct buf *bp;
1346 		daddr_t blk;
1347 
1348 		/*
1349 		 * Maximum I/O size is limited by a number of factors.
1350 		 */
1351 		n = min(BLIST_MAX_ALLOC, count - i);
1352 		n = min(n, nsw_cluster_max);
1353 
1354 		/*
1355 		 * Get biggest block of swap we can.  If we fail, fall
1356 		 * back and try to allocate a smaller block.  Don't go
1357 		 * overboard trying to allocate space if it would overly
1358 		 * fragment swap.
1359 		 */
1360 		while (
1361 		    (blk = swp_pager_getswapspace(n)) == SWAPBLK_NONE &&
1362 		    n > 4
1363 		) {
1364 			n >>= 1;
1365 		}
1366 		if (blk == SWAPBLK_NONE) {
1367 			for (j = 0; j < n; ++j)
1368 				rtvals[i+j] = VM_PAGER_FAIL;
1369 			continue;
1370 		}
1371 
1372 		/*
1373 		 * All I/O parameters have been satisfied, build the I/O
1374 		 * request and assign the swap space.
1375 		 */
1376 		if (sync == TRUE) {
1377 			bp = getpbuf(&nsw_wcount_sync);
1378 		} else {
1379 			bp = getpbuf(&nsw_wcount_async);
1380 			bp->b_flags = B_ASYNC;
1381 		}
1382 		bp->b_flags |= B_PAGING;
1383 		bp->b_iocmd = BIO_WRITE;
1384 
1385 		bp->b_rcred = crhold(thread0.td_ucred);
1386 		bp->b_wcred = crhold(thread0.td_ucred);
1387 		bp->b_bcount = PAGE_SIZE * n;
1388 		bp->b_bufsize = PAGE_SIZE * n;
1389 		bp->b_blkno = blk;
1390 
1391 		VM_OBJECT_WLOCK(object);
1392 		for (j = 0; j < n; ++j) {
1393 			vm_page_t mreq = m[i+j];
1394 
1395 			swp_pager_meta_build(
1396 			    mreq->object,
1397 			    mreq->pindex,
1398 			    blk + j
1399 			);
1400 			MPASS(mreq->dirty == VM_PAGE_BITS_ALL);
1401 			mreq->oflags |= VPO_SWAPINPROG;
1402 			bp->b_pages[j] = mreq;
1403 		}
1404 		VM_OBJECT_WUNLOCK(object);
1405 		bp->b_npages = n;
1406 		/*
1407 		 * Must set dirty range for NFS to work.
1408 		 */
1409 		bp->b_dirtyoff = 0;
1410 		bp->b_dirtyend = bp->b_bcount;
1411 
1412 		VM_CNT_INC(v_swapout);
1413 		VM_CNT_ADD(v_swappgsout, bp->b_npages);
1414 
1415 		/*
1416 		 * We unconditionally set rtvals[] to VM_PAGER_PEND so that we
1417 		 * can call the async completion routine at the end of a
1418 		 * synchronous I/O operation.  Otherwise, our caller would
1419 		 * perform duplicate unbusy and wakeup operations on the page
1420 		 * and object, respectively.
1421 		 */
1422 		for (j = 0; j < n; j++)
1423 			rtvals[i + j] = VM_PAGER_PEND;
1424 
1425 		/*
1426 		 * asynchronous
1427 		 *
1428 		 * NOTE: b_blkno is destroyed by the call to swapdev_strategy
1429 		 */
1430 		if (sync == FALSE) {
1431 			bp->b_iodone = swp_pager_async_iodone;
1432 			BUF_KERNPROC(bp);
1433 			swp_pager_strategy(bp);
1434 			continue;
1435 		}
1436 
1437 		/*
1438 		 * synchronous
1439 		 *
1440 		 * NOTE: b_blkno is destroyed by the call to swapdev_strategy
1441 		 */
1442 		bp->b_iodone = bdone;
1443 		swp_pager_strategy(bp);
1444 
1445 		/*
1446 		 * Wait for the sync I/O to complete.
1447 		 */
1448 		bwait(bp, PVM, "swwrt");
1449 
1450 		/*
1451 		 * Now that we are through with the bp, we can call the
1452 		 * normal async completion, which frees everything up.
1453 		 */
1454 		swp_pager_async_iodone(bp);
1455 	}
1456 	VM_OBJECT_WLOCK(object);
1457 }
1458 
1459 /*
1460  *	swp_pager_async_iodone:
1461  *
1462  *	Completion routine for asynchronous reads and writes from/to swap.
1463  *	Also called manually by synchronous code to finish up a bp.
1464  *
1465  *	This routine may not sleep.
1466  */
1467 static void
1468 swp_pager_async_iodone(struct buf *bp)
1469 {
1470 	int i;
1471 	vm_object_t object = NULL;
1472 
1473 	/*
1474 	 * report error
1475 	 */
1476 	if (bp->b_ioflags & BIO_ERROR) {
1477 		printf(
1478 		    "swap_pager: I/O error - %s failed; blkno %ld,"
1479 			"size %ld, error %d\n",
1480 		    ((bp->b_iocmd == BIO_READ) ? "pagein" : "pageout"),
1481 		    (long)bp->b_blkno,
1482 		    (long)bp->b_bcount,
1483 		    bp->b_error
1484 		);
1485 	}
1486 
1487 	/*
1488 	 * remove the mapping for kernel virtual
1489 	 */
1490 	if (buf_mapped(bp))
1491 		pmap_qremove((vm_offset_t)bp->b_data, bp->b_npages);
1492 	else
1493 		bp->b_data = bp->b_kvabase;
1494 
1495 	if (bp->b_npages) {
1496 		object = bp->b_pages[0]->object;
1497 		VM_OBJECT_WLOCK(object);
1498 	}
1499 
1500 	/*
1501 	 * cleanup pages.  If an error occurs writing to swap, we are in
1502 	 * very serious trouble.  If it happens to be a disk error, though,
1503 	 * we may be able to recover by reassigning the swap later on.  So
1504 	 * in this case we remove the m->swapblk assignment for the page
1505 	 * but do not free it in the rlist.  The errornous block(s) are thus
1506 	 * never reallocated as swap.  Redirty the page and continue.
1507 	 */
1508 	for (i = 0; i < bp->b_npages; ++i) {
1509 		vm_page_t m = bp->b_pages[i];
1510 
1511 		m->oflags &= ~VPO_SWAPINPROG;
1512 		if (m->oflags & VPO_SWAPSLEEP) {
1513 			m->oflags &= ~VPO_SWAPSLEEP;
1514 			wakeup(&object->paging_in_progress);
1515 		}
1516 
1517 		if (bp->b_ioflags & BIO_ERROR) {
1518 			/*
1519 			 * If an error occurs I'd love to throw the swapblk
1520 			 * away without freeing it back to swapspace, so it
1521 			 * can never be used again.  But I can't from an
1522 			 * interrupt.
1523 			 */
1524 			if (bp->b_iocmd == BIO_READ) {
1525 				/*
1526 				 * NOTE: for reads, m->dirty will probably
1527 				 * be overridden by the original caller of
1528 				 * getpages so don't play cute tricks here.
1529 				 */
1530 				m->valid = 0;
1531 			} else {
1532 				/*
1533 				 * If a write error occurs, reactivate page
1534 				 * so it doesn't clog the inactive list,
1535 				 * then finish the I/O.
1536 				 */
1537 				MPASS(m->dirty == VM_PAGE_BITS_ALL);
1538 				vm_page_lock(m);
1539 				vm_page_activate(m);
1540 				vm_page_unlock(m);
1541 				vm_page_sunbusy(m);
1542 			}
1543 		} else if (bp->b_iocmd == BIO_READ) {
1544 			/*
1545 			 * NOTE: for reads, m->dirty will probably be
1546 			 * overridden by the original caller of getpages so
1547 			 * we cannot set them in order to free the underlying
1548 			 * swap in a low-swap situation.  I don't think we'd
1549 			 * want to do that anyway, but it was an optimization
1550 			 * that existed in the old swapper for a time before
1551 			 * it got ripped out due to precisely this problem.
1552 			 */
1553 			KASSERT(!pmap_page_is_mapped(m),
1554 			    ("swp_pager_async_iodone: page %p is mapped", m));
1555 			KASSERT(m->dirty == 0,
1556 			    ("swp_pager_async_iodone: page %p is dirty", m));
1557 
1558 			m->valid = VM_PAGE_BITS_ALL;
1559 			if (i < bp->b_pgbefore ||
1560 			    i >= bp->b_npages - bp->b_pgafter)
1561 				vm_page_readahead_finish(m);
1562 		} else {
1563 			/*
1564 			 * For write success, clear the dirty
1565 			 * status, then finish the I/O ( which decrements the
1566 			 * busy count and possibly wakes waiter's up ).
1567 			 * A page is only written to swap after a period of
1568 			 * inactivity.  Therefore, we do not expect it to be
1569 			 * reused.
1570 			 */
1571 			KASSERT(!pmap_page_is_write_mapped(m),
1572 			    ("swp_pager_async_iodone: page %p is not write"
1573 			    " protected", m));
1574 			vm_page_undirty(m);
1575 			vm_page_lock(m);
1576 			vm_page_deactivate_noreuse(m);
1577 			vm_page_unlock(m);
1578 			vm_page_sunbusy(m);
1579 		}
1580 	}
1581 
1582 	/*
1583 	 * adjust pip.  NOTE: the original parent may still have its own
1584 	 * pip refs on the object.
1585 	 */
1586 	if (object != NULL) {
1587 		vm_object_pip_wakeupn(object, bp->b_npages);
1588 		VM_OBJECT_WUNLOCK(object);
1589 	}
1590 
1591 	/*
1592 	 * swapdev_strategy() manually sets b_vp and b_bufobj before calling
1593 	 * bstrategy(). Set them back to NULL now we're done with it, or we'll
1594 	 * trigger a KASSERT in relpbuf().
1595 	 */
1596 	if (bp->b_vp) {
1597 		    bp->b_vp = NULL;
1598 		    bp->b_bufobj = NULL;
1599 	}
1600 	/*
1601 	 * release the physical I/O buffer
1602 	 */
1603 	relpbuf(
1604 	    bp,
1605 	    ((bp->b_iocmd == BIO_READ) ? &nsw_rcount :
1606 		((bp->b_flags & B_ASYNC) ?
1607 		    &nsw_wcount_async :
1608 		    &nsw_wcount_sync
1609 		)
1610 	    )
1611 	);
1612 }
1613 
1614 int
1615 swap_pager_nswapdev(void)
1616 {
1617 
1618 	return (nswapdev);
1619 }
1620 
1621 /*
1622  * SWP_PAGER_FORCE_PAGEIN() - force a swap block to be paged in
1623  *
1624  *	This routine dissociates the page at the given index within an object
1625  *	from its backing store, paging it in if it does not reside in memory.
1626  *	If the page is paged in, it is marked dirty and placed in the laundry
1627  *	queue.  The page is marked dirty because it no longer has backing
1628  *	store.  It is placed in the laundry queue because it has not been
1629  *	accessed recently.  Otherwise, it would already reside in memory.
1630  *
1631  *	We also attempt to swap in all other pages in the swap block.
1632  *	However, we only guarantee that the one at the specified index is
1633  *	paged in.
1634  *
1635  *	XXX - The code to page the whole block in doesn't work, so we
1636  *	      revert to the one-by-one behavior for now.  Sigh.
1637  */
1638 static inline void
1639 swp_pager_force_pagein(vm_object_t object, vm_pindex_t pindex)
1640 {
1641 	vm_page_t m;
1642 
1643 	vm_object_pip_add(object, 1);
1644 	m = vm_page_grab(object, pindex, VM_ALLOC_NORMAL);
1645 	if (m->valid == VM_PAGE_BITS_ALL) {
1646 		vm_object_pip_wakeup(object);
1647 		vm_page_dirty(m);
1648 #ifdef INVARIANTS
1649 		vm_page_lock(m);
1650 		if (m->wire_count == 0 && m->queue == PQ_NONE)
1651 			panic("page %p is neither wired nor queued", m);
1652 		vm_page_unlock(m);
1653 #endif
1654 		vm_page_xunbusy(m);
1655 		vm_pager_page_unswapped(m);
1656 		return;
1657 	}
1658 
1659 	if (swap_pager_getpages(object, &m, 1, NULL, NULL) != VM_PAGER_OK)
1660 		panic("swap_pager_force_pagein: read from swap failed");/*XXX*/
1661 	vm_object_pip_wakeup(object);
1662 	vm_page_dirty(m);
1663 	vm_page_lock(m);
1664 	vm_page_launder(m);
1665 	vm_page_unlock(m);
1666 	vm_page_xunbusy(m);
1667 	vm_pager_page_unswapped(m);
1668 }
1669 
1670 /*
1671  *	swap_pager_swapoff:
1672  *
1673  *	Page in all of the pages that have been paged out to the
1674  *	given device.  The corresponding blocks in the bitmap must be
1675  *	marked as allocated and the device must be flagged SW_CLOSING.
1676  *	There may be no processes swapped out to the device.
1677  *
1678  *	This routine may block.
1679  */
1680 static void
1681 swap_pager_swapoff(struct swdevt *sp)
1682 {
1683 	struct swblk *sb;
1684 	vm_object_t object;
1685 	vm_pindex_t pi;
1686 	int i, retries;
1687 
1688 	sx_assert(&swdev_syscall_lock, SA_XLOCKED);
1689 
1690 	retries = 0;
1691 full_rescan:
1692 	mtx_lock(&vm_object_list_mtx);
1693 	TAILQ_FOREACH(object, &vm_object_list, object_list) {
1694 		if (object->type != OBJT_SWAP)
1695 			continue;
1696 		mtx_unlock(&vm_object_list_mtx);
1697 		/* Depends on type-stability. */
1698 		VM_OBJECT_WLOCK(object);
1699 
1700 		/*
1701 		 * Dead objects are eventually terminated on their own.
1702 		 */
1703 		if ((object->flags & OBJ_DEAD) != 0)
1704 			goto next_obj;
1705 
1706 		/*
1707 		 * Sync with fences placed after pctrie
1708 		 * initialization.  We must not access pctrie below
1709 		 * unless we checked that our object is swap and not
1710 		 * dead.
1711 		 */
1712 		atomic_thread_fence_acq();
1713 		if (object->type != OBJT_SWAP)
1714 			goto next_obj;
1715 
1716 		for (pi = 0; (sb = SWAP_PCTRIE_LOOKUP_GE(
1717 		    &object->un_pager.swp.swp_blks, pi)) != NULL; ) {
1718 			pi = sb->p + SWAP_META_PAGES;
1719 			for (i = 0; i < SWAP_META_PAGES; i++) {
1720 				if (sb->d[i] == SWAPBLK_NONE)
1721 					continue;
1722 				if (swp_pager_isondev(sb->d[i], sp))
1723 					swp_pager_force_pagein(object,
1724 					    sb->p + i);
1725 			}
1726 		}
1727 next_obj:
1728 		VM_OBJECT_WUNLOCK(object);
1729 		mtx_lock(&vm_object_list_mtx);
1730 	}
1731 	mtx_unlock(&vm_object_list_mtx);
1732 
1733 	if (sp->sw_used) {
1734 		/*
1735 		 * Objects may be locked or paging to the device being
1736 		 * removed, so we will miss their pages and need to
1737 		 * make another pass.  We have marked this device as
1738 		 * SW_CLOSING, so the activity should finish soon.
1739 		 */
1740 		retries++;
1741 		if (retries > 100) {
1742 			panic("swapoff: failed to locate %d swap blocks",
1743 			    sp->sw_used);
1744 		}
1745 		pause("swpoff", hz / 20);
1746 		goto full_rescan;
1747 	}
1748 	EVENTHANDLER_INVOKE(swapoff, sp);
1749 }
1750 
1751 /************************************************************************
1752  *				SWAP META DATA 				*
1753  ************************************************************************
1754  *
1755  *	These routines manipulate the swap metadata stored in the
1756  *	OBJT_SWAP object.
1757  *
1758  *	Swap metadata is implemented with a global hash and not directly
1759  *	linked into the object.  Instead the object simply contains
1760  *	appropriate tracking counters.
1761  */
1762 
1763 /*
1764  * SWP_PAGER_META_BUILD() -	add swap block to swap meta data for object
1765  *
1766  *	We first convert the object to a swap object if it is a default
1767  *	object.
1768  *
1769  *	The specified swapblk is added to the object's swap metadata.  If
1770  *	the swapblk is not valid, it is freed instead.  Any previously
1771  *	assigned swapblk is freed.
1772  */
1773 static void
1774 swp_pager_meta_build(vm_object_t object, vm_pindex_t pindex, daddr_t swapblk)
1775 {
1776 	static volatile int swblk_zone_exhausted, swpctrie_zone_exhausted;
1777 	struct swblk *sb, *sb1;
1778 	vm_pindex_t modpi, rdpi;
1779 	int error, i;
1780 
1781 	VM_OBJECT_ASSERT_WLOCKED(object);
1782 
1783 	/*
1784 	 * Convert default object to swap object if necessary
1785 	 */
1786 	if (object->type != OBJT_SWAP) {
1787 		pctrie_init(&object->un_pager.swp.swp_blks);
1788 
1789 		/*
1790 		 * Ensure that swap_pager_swapoff()'s iteration over
1791 		 * object_list does not see a garbage pctrie.
1792 		 */
1793 		atomic_thread_fence_rel();
1794 
1795 		object->type = OBJT_SWAP;
1796 		KASSERT(object->handle == NULL, ("default pager with handle"));
1797 	}
1798 
1799 	rdpi = rounddown(pindex, SWAP_META_PAGES);
1800 	sb = SWAP_PCTRIE_LOOKUP(&object->un_pager.swp.swp_blks, rdpi);
1801 	if (sb == NULL) {
1802 		if (swapblk == SWAPBLK_NONE)
1803 			return;
1804 		for (;;) {
1805 			sb = uma_zalloc(swblk_zone, M_NOWAIT | (curproc ==
1806 			    pageproc ? M_USE_RESERVE : 0));
1807 			if (sb != NULL) {
1808 				sb->p = rdpi;
1809 				for (i = 0; i < SWAP_META_PAGES; i++)
1810 					sb->d[i] = SWAPBLK_NONE;
1811 				if (atomic_cmpset_int(&swblk_zone_exhausted,
1812 				    1, 0))
1813 					printf("swblk zone ok\n");
1814 				break;
1815 			}
1816 			VM_OBJECT_WUNLOCK(object);
1817 			if (uma_zone_exhausted(swblk_zone)) {
1818 				if (atomic_cmpset_int(&swblk_zone_exhausted,
1819 				    0, 1))
1820 					printf("swap blk zone exhausted, "
1821 					    "increase kern.maxswzone\n");
1822 				vm_pageout_oom(VM_OOM_SWAPZ);
1823 				pause("swzonxb", 10);
1824 			} else
1825 				uma_zwait(swblk_zone);
1826 			VM_OBJECT_WLOCK(object);
1827 			sb = SWAP_PCTRIE_LOOKUP(&object->un_pager.swp.swp_blks,
1828 			    rdpi);
1829 			if (sb != NULL)
1830 				/*
1831 				 * Somebody swapped out a nearby page,
1832 				 * allocating swblk at the rdpi index,
1833 				 * while we dropped the object lock.
1834 				 */
1835 				goto allocated;
1836 		}
1837 		for (;;) {
1838 			error = SWAP_PCTRIE_INSERT(
1839 			    &object->un_pager.swp.swp_blks, sb);
1840 			if (error == 0) {
1841 				if (atomic_cmpset_int(&swpctrie_zone_exhausted,
1842 				    1, 0))
1843 					printf("swpctrie zone ok\n");
1844 				break;
1845 			}
1846 			VM_OBJECT_WUNLOCK(object);
1847 			if (uma_zone_exhausted(swpctrie_zone)) {
1848 				if (atomic_cmpset_int(&swpctrie_zone_exhausted,
1849 				    0, 1))
1850 					printf("swap pctrie zone exhausted, "
1851 					    "increase kern.maxswzone\n");
1852 				vm_pageout_oom(VM_OOM_SWAPZ);
1853 				pause("swzonxp", 10);
1854 			} else
1855 				uma_zwait(swpctrie_zone);
1856 			VM_OBJECT_WLOCK(object);
1857 			sb1 = SWAP_PCTRIE_LOOKUP(&object->un_pager.swp.swp_blks,
1858 			    rdpi);
1859 			if (sb1 != NULL) {
1860 				uma_zfree(swblk_zone, sb);
1861 				sb = sb1;
1862 				goto allocated;
1863 			}
1864 		}
1865 	}
1866 allocated:
1867 	MPASS(sb->p == rdpi);
1868 
1869 	modpi = pindex % SWAP_META_PAGES;
1870 	/* Delete prior contents of metadata. */
1871 	if (sb->d[modpi] != SWAPBLK_NONE)
1872 		swp_pager_freeswapspace(sb->d[modpi], 1);
1873 	/* Enter block into metadata. */
1874 	sb->d[modpi] = swapblk;
1875 
1876 	/*
1877 	 * Free the swblk if we end up with the empty page run.
1878 	 */
1879 	if (swapblk == SWAPBLK_NONE) {
1880 		for (i = 0; i < SWAP_META_PAGES; i++) {
1881 			if (sb->d[i] != SWAPBLK_NONE)
1882 				break;
1883 		}
1884 		if (i == SWAP_META_PAGES) {
1885 			SWAP_PCTRIE_REMOVE(&object->un_pager.swp.swp_blks,
1886 			    rdpi);
1887 			uma_zfree(swblk_zone, sb);
1888 		}
1889 	}
1890 }
1891 
1892 /*
1893  * SWP_PAGER_META_FREE() - free a range of blocks in the object's swap metadata
1894  *
1895  *	The requested range of blocks is freed, with any associated swap
1896  *	returned to the swap bitmap.
1897  *
1898  *	This routine will free swap metadata structures as they are cleaned
1899  *	out.  This routine does *NOT* operate on swap metadata associated
1900  *	with resident pages.
1901  */
1902 static void
1903 swp_pager_meta_free(vm_object_t object, vm_pindex_t pindex, vm_pindex_t count)
1904 {
1905 	struct swblk *sb;
1906 	vm_pindex_t last;
1907 	int i;
1908 	bool empty;
1909 
1910 	VM_OBJECT_ASSERT_WLOCKED(object);
1911 	if (object->type != OBJT_SWAP || count == 0)
1912 		return;
1913 
1914 	last = pindex + count - 1;
1915 	for (;;) {
1916 		sb = SWAP_PCTRIE_LOOKUP_GE(&object->un_pager.swp.swp_blks,
1917 		    rounddown(pindex, SWAP_META_PAGES));
1918 		if (sb == NULL || sb->p > last)
1919 			break;
1920 		empty = true;
1921 		for (i = 0; i < SWAP_META_PAGES; i++) {
1922 			if (sb->d[i] == SWAPBLK_NONE)
1923 				continue;
1924 			if (pindex <= sb->p + i && sb->p + i <= last) {
1925 				swp_pager_freeswapspace(sb->d[i], 1);
1926 				sb->d[i] = SWAPBLK_NONE;
1927 			} else
1928 				empty = false;
1929 		}
1930 		pindex = sb->p + SWAP_META_PAGES;
1931 		if (empty) {
1932 			SWAP_PCTRIE_REMOVE(&object->un_pager.swp.swp_blks,
1933 			    sb->p);
1934 			uma_zfree(swblk_zone, sb);
1935 		}
1936 	}
1937 }
1938 
1939 /*
1940  * SWP_PAGER_META_FREE_ALL() - destroy all swap metadata associated with object
1941  *
1942  *	This routine locates and destroys all swap metadata associated with
1943  *	an object.
1944  */
1945 static void
1946 swp_pager_meta_free_all(vm_object_t object)
1947 {
1948 	struct swblk *sb;
1949 	vm_pindex_t pindex;
1950 	int i;
1951 
1952 	VM_OBJECT_ASSERT_WLOCKED(object);
1953 	if (object->type != OBJT_SWAP)
1954 		return;
1955 
1956 	for (pindex = 0; (sb = SWAP_PCTRIE_LOOKUP_GE(
1957 	    &object->un_pager.swp.swp_blks, pindex)) != NULL;) {
1958 		pindex = sb->p + SWAP_META_PAGES;
1959 		for (i = 0; i < SWAP_META_PAGES; i++) {
1960 			if (sb->d[i] != SWAPBLK_NONE)
1961 				swp_pager_freeswapspace(sb->d[i], 1);
1962 		}
1963 		SWAP_PCTRIE_REMOVE(&object->un_pager.swp.swp_blks, sb->p);
1964 		uma_zfree(swblk_zone, sb);
1965 	}
1966 }
1967 
1968 /*
1969  * SWP_PAGER_METACTL() -  misc control of swap and vm_page_t meta data.
1970  *
1971  *	This routine is capable of looking up, popping, or freeing
1972  *	swapblk assignments in the swap meta data or in the vm_page_t.
1973  *	The routine typically returns the swapblk being looked-up, or popped,
1974  *	or SWAPBLK_NONE if the block was freed, or SWAPBLK_NONE if the block
1975  *	was invalid.  This routine will automatically free any invalid
1976  *	meta-data swapblks.
1977  *
1978  *	When acting on a busy resident page and paging is in progress, we
1979  *	have to wait until paging is complete but otherwise can act on the
1980  *	busy page.
1981  *
1982  *	SWM_FREE	remove and free swap block from metadata
1983  *	SWM_POP		remove from meta data but do not free.. pop it out
1984  */
1985 static daddr_t
1986 swp_pager_meta_ctl(vm_object_t object, vm_pindex_t pindex, int flags)
1987 {
1988 	struct swblk *sb;
1989 	daddr_t r1;
1990 	int i;
1991 
1992 	if ((flags & (SWM_FREE | SWM_POP)) != 0)
1993 		VM_OBJECT_ASSERT_WLOCKED(object);
1994 	else
1995 		VM_OBJECT_ASSERT_LOCKED(object);
1996 
1997 	/*
1998 	 * The meta data only exists if the object is OBJT_SWAP
1999 	 * and even then might not be allocated yet.
2000 	 */
2001 	if (object->type != OBJT_SWAP)
2002 		return (SWAPBLK_NONE);
2003 
2004 	sb = SWAP_PCTRIE_LOOKUP(&object->un_pager.swp.swp_blks,
2005 	    rounddown(pindex, SWAP_META_PAGES));
2006 	if (sb == NULL)
2007 		return (SWAPBLK_NONE);
2008 	r1 = sb->d[pindex % SWAP_META_PAGES];
2009 	if (r1 == SWAPBLK_NONE)
2010 		return (SWAPBLK_NONE);
2011 	if ((flags & (SWM_FREE | SWM_POP)) != 0) {
2012 		sb->d[pindex % SWAP_META_PAGES] = SWAPBLK_NONE;
2013 		for (i = 0; i < SWAP_META_PAGES; i++) {
2014 			if (sb->d[i] != SWAPBLK_NONE)
2015 				break;
2016 		}
2017 		if (i == SWAP_META_PAGES) {
2018 			SWAP_PCTRIE_REMOVE(&object->un_pager.swp.swp_blks,
2019 			    rounddown(pindex, SWAP_META_PAGES));
2020 			uma_zfree(swblk_zone, sb);
2021 		}
2022 	}
2023 	if ((flags & SWM_FREE) != 0) {
2024 		swp_pager_freeswapspace(r1, 1);
2025 		r1 = SWAPBLK_NONE;
2026 	}
2027 	return (r1);
2028 }
2029 
2030 /*
2031  * Returns the least page index which is greater than or equal to the
2032  * parameter pindex and for which there is a swap block allocated.
2033  * Returns object's size if the object's type is not swap or if there
2034  * are no allocated swap blocks for the object after the requested
2035  * pindex.
2036  */
2037 vm_pindex_t
2038 swap_pager_find_least(vm_object_t object, vm_pindex_t pindex)
2039 {
2040 	struct swblk *sb;
2041 	int i;
2042 
2043 	VM_OBJECT_ASSERT_LOCKED(object);
2044 	if (object->type != OBJT_SWAP)
2045 		return (object->size);
2046 
2047 	sb = SWAP_PCTRIE_LOOKUP_GE(&object->un_pager.swp.swp_blks,
2048 	    rounddown(pindex, SWAP_META_PAGES));
2049 	if (sb == NULL)
2050 		return (object->size);
2051 	if (sb->p < pindex) {
2052 		for (i = pindex % SWAP_META_PAGES; i < SWAP_META_PAGES; i++) {
2053 			if (sb->d[i] != SWAPBLK_NONE)
2054 				return (sb->p + i);
2055 		}
2056 		sb = SWAP_PCTRIE_LOOKUP_GE(&object->un_pager.swp.swp_blks,
2057 		    roundup(pindex, SWAP_META_PAGES));
2058 		if (sb == NULL)
2059 			return (object->size);
2060 	}
2061 	for (i = 0; i < SWAP_META_PAGES; i++) {
2062 		if (sb->d[i] != SWAPBLK_NONE)
2063 			return (sb->p + i);
2064 	}
2065 
2066 	/*
2067 	 * We get here if a swblk is present in the trie but it
2068 	 * doesn't map any blocks.
2069 	 */
2070 	MPASS(0);
2071 	return (object->size);
2072 }
2073 
2074 /*
2075  * System call swapon(name) enables swapping on device name,
2076  * which must be in the swdevsw.  Return EBUSY
2077  * if already swapping on this device.
2078  */
2079 #ifndef _SYS_SYSPROTO_H_
2080 struct swapon_args {
2081 	char *name;
2082 };
2083 #endif
2084 
2085 /*
2086  * MPSAFE
2087  */
2088 /* ARGSUSED */
2089 int
2090 sys_swapon(struct thread *td, struct swapon_args *uap)
2091 {
2092 	struct vattr attr;
2093 	struct vnode *vp;
2094 	struct nameidata nd;
2095 	int error;
2096 
2097 	error = priv_check(td, PRIV_SWAPON);
2098 	if (error)
2099 		return (error);
2100 
2101 	sx_xlock(&swdev_syscall_lock);
2102 
2103 	/*
2104 	 * Swap metadata may not fit in the KVM if we have physical
2105 	 * memory of >1GB.
2106 	 */
2107 	if (swblk_zone == NULL) {
2108 		error = ENOMEM;
2109 		goto done;
2110 	}
2111 
2112 	NDINIT(&nd, LOOKUP, ISOPEN | FOLLOW | AUDITVNODE1, UIO_USERSPACE,
2113 	    uap->name, td);
2114 	error = namei(&nd);
2115 	if (error)
2116 		goto done;
2117 
2118 	NDFREE(&nd, NDF_ONLY_PNBUF);
2119 	vp = nd.ni_vp;
2120 
2121 	if (vn_isdisk(vp, &error)) {
2122 		error = swapongeom(vp);
2123 	} else if (vp->v_type == VREG &&
2124 	    (vp->v_mount->mnt_vfc->vfc_flags & VFCF_NETWORK) != 0 &&
2125 	    (error = VOP_GETATTR(vp, &attr, td->td_ucred)) == 0) {
2126 		/*
2127 		 * Allow direct swapping to NFS regular files in the same
2128 		 * way that nfs_mountroot() sets up diskless swapping.
2129 		 */
2130 		error = swaponvp(td, vp, attr.va_size / DEV_BSIZE);
2131 	}
2132 
2133 	if (error)
2134 		vrele(vp);
2135 done:
2136 	sx_xunlock(&swdev_syscall_lock);
2137 	return (error);
2138 }
2139 
2140 /*
2141  * Check that the total amount of swap currently configured does not
2142  * exceed half the theoretical maximum.  If it does, print a warning
2143  * message.
2144  */
2145 static void
2146 swapon_check_swzone(void)
2147 {
2148 	unsigned long maxpages, npages;
2149 
2150 	npages = swap_total / PAGE_SIZE;
2151 	/* absolute maximum we can handle assuming 100% efficiency */
2152 	maxpages = uma_zone_get_max(swblk_zone) * SWAP_META_PAGES;
2153 
2154 	/* recommend using no more than half that amount */
2155 	if (npages > maxpages / 2) {
2156 		printf("warning: total configured swap (%lu pages) "
2157 		    "exceeds maximum recommended amount (%lu pages).\n",
2158 		    npages, maxpages / 2);
2159 		printf("warning: increase kern.maxswzone "
2160 		    "or reduce amount of swap.\n");
2161 	}
2162 }
2163 
2164 static void
2165 swaponsomething(struct vnode *vp, void *id, u_long nblks,
2166     sw_strategy_t *strategy, sw_close_t *close, dev_t dev, int flags)
2167 {
2168 	struct swdevt *sp, *tsp;
2169 	swblk_t dvbase;
2170 	u_long mblocks;
2171 
2172 	/*
2173 	 * nblks is in DEV_BSIZE'd chunks, convert to PAGE_SIZE'd chunks.
2174 	 * First chop nblks off to page-align it, then convert.
2175 	 *
2176 	 * sw->sw_nblks is in page-sized chunks now too.
2177 	 */
2178 	nblks &= ~(ctodb(1) - 1);
2179 	nblks = dbtoc(nblks);
2180 
2181 	/*
2182 	 * If we go beyond this, we get overflows in the radix
2183 	 * tree bitmap code.
2184 	 */
2185 	mblocks = 0x40000000 / BLIST_META_RADIX;
2186 	if (nblks > mblocks) {
2187 		printf(
2188     "WARNING: reducing swap size to maximum of %luMB per unit\n",
2189 		    mblocks / 1024 / 1024 * PAGE_SIZE);
2190 		nblks = mblocks;
2191 	}
2192 
2193 	sp = malloc(sizeof *sp, M_VMPGDATA, M_WAITOK | M_ZERO);
2194 	sp->sw_vp = vp;
2195 	sp->sw_id = id;
2196 	sp->sw_dev = dev;
2197 	sp->sw_flags = 0;
2198 	sp->sw_nblks = nblks;
2199 	sp->sw_used = 0;
2200 	sp->sw_strategy = strategy;
2201 	sp->sw_close = close;
2202 	sp->sw_flags = flags;
2203 
2204 	sp->sw_blist = blist_create(nblks, M_WAITOK);
2205 	/*
2206 	 * Do not free the first two block in order to avoid overwriting
2207 	 * any bsd label at the front of the partition
2208 	 */
2209 	blist_free(sp->sw_blist, 2, nblks - 2);
2210 
2211 	dvbase = 0;
2212 	mtx_lock(&sw_dev_mtx);
2213 	TAILQ_FOREACH(tsp, &swtailq, sw_list) {
2214 		if (tsp->sw_end >= dvbase) {
2215 			/*
2216 			 * We put one uncovered page between the devices
2217 			 * in order to definitively prevent any cross-device
2218 			 * I/O requests
2219 			 */
2220 			dvbase = tsp->sw_end + 1;
2221 		}
2222 	}
2223 	sp->sw_first = dvbase;
2224 	sp->sw_end = dvbase + nblks;
2225 	TAILQ_INSERT_TAIL(&swtailq, sp, sw_list);
2226 	nswapdev++;
2227 	swap_pager_avail += nblks - 2;
2228 	swap_total += (vm_ooffset_t)nblks * PAGE_SIZE;
2229 	swapon_check_swzone();
2230 	swp_sizecheck();
2231 	mtx_unlock(&sw_dev_mtx);
2232 	EVENTHANDLER_INVOKE(swapon, sp);
2233 }
2234 
2235 /*
2236  * SYSCALL: swapoff(devname)
2237  *
2238  * Disable swapping on the given device.
2239  *
2240  * XXX: Badly designed system call: it should use a device index
2241  * rather than filename as specification.  We keep sw_vp around
2242  * only to make this work.
2243  */
2244 #ifndef _SYS_SYSPROTO_H_
2245 struct swapoff_args {
2246 	char *name;
2247 };
2248 #endif
2249 
2250 /*
2251  * MPSAFE
2252  */
2253 /* ARGSUSED */
2254 int
2255 sys_swapoff(struct thread *td, struct swapoff_args *uap)
2256 {
2257 	struct vnode *vp;
2258 	struct nameidata nd;
2259 	struct swdevt *sp;
2260 	int error;
2261 
2262 	error = priv_check(td, PRIV_SWAPOFF);
2263 	if (error)
2264 		return (error);
2265 
2266 	sx_xlock(&swdev_syscall_lock);
2267 
2268 	NDINIT(&nd, LOOKUP, FOLLOW | AUDITVNODE1, UIO_USERSPACE, uap->name,
2269 	    td);
2270 	error = namei(&nd);
2271 	if (error)
2272 		goto done;
2273 	NDFREE(&nd, NDF_ONLY_PNBUF);
2274 	vp = nd.ni_vp;
2275 
2276 	mtx_lock(&sw_dev_mtx);
2277 	TAILQ_FOREACH(sp, &swtailq, sw_list) {
2278 		if (sp->sw_vp == vp)
2279 			break;
2280 	}
2281 	mtx_unlock(&sw_dev_mtx);
2282 	if (sp == NULL) {
2283 		error = EINVAL;
2284 		goto done;
2285 	}
2286 	error = swapoff_one(sp, td->td_ucred);
2287 done:
2288 	sx_xunlock(&swdev_syscall_lock);
2289 	return (error);
2290 }
2291 
2292 static int
2293 swapoff_one(struct swdevt *sp, struct ucred *cred)
2294 {
2295 	u_long nblks;
2296 #ifdef MAC
2297 	int error;
2298 #endif
2299 
2300 	sx_assert(&swdev_syscall_lock, SA_XLOCKED);
2301 #ifdef MAC
2302 	(void) vn_lock(sp->sw_vp, LK_EXCLUSIVE | LK_RETRY);
2303 	error = mac_system_check_swapoff(cred, sp->sw_vp);
2304 	(void) VOP_UNLOCK(sp->sw_vp, 0);
2305 	if (error != 0)
2306 		return (error);
2307 #endif
2308 	nblks = sp->sw_nblks;
2309 
2310 	/*
2311 	 * We can turn off this swap device safely only if the
2312 	 * available virtual memory in the system will fit the amount
2313 	 * of data we will have to page back in, plus an epsilon so
2314 	 * the system doesn't become critically low on swap space.
2315 	 */
2316 	if (vm_cnt.v_free_count + swap_pager_avail < nblks + nswap_lowat)
2317 		return (ENOMEM);
2318 
2319 	/*
2320 	 * Prevent further allocations on this device.
2321 	 */
2322 	mtx_lock(&sw_dev_mtx);
2323 	sp->sw_flags |= SW_CLOSING;
2324 	swap_pager_avail -= blist_fill(sp->sw_blist, 0, nblks);
2325 	swap_total -= (vm_ooffset_t)nblks * PAGE_SIZE;
2326 	mtx_unlock(&sw_dev_mtx);
2327 
2328 	/*
2329 	 * Page in the contents of the device and close it.
2330 	 */
2331 	swap_pager_swapoff(sp);
2332 
2333 	sp->sw_close(curthread, sp);
2334 	mtx_lock(&sw_dev_mtx);
2335 	sp->sw_id = NULL;
2336 	TAILQ_REMOVE(&swtailq, sp, sw_list);
2337 	nswapdev--;
2338 	if (nswapdev == 0) {
2339 		swap_pager_full = 2;
2340 		swap_pager_almost_full = 1;
2341 	}
2342 	if (swdevhd == sp)
2343 		swdevhd = NULL;
2344 	mtx_unlock(&sw_dev_mtx);
2345 	blist_destroy(sp->sw_blist);
2346 	free(sp, M_VMPGDATA);
2347 	return (0);
2348 }
2349 
2350 void
2351 swapoff_all(void)
2352 {
2353 	struct swdevt *sp, *spt;
2354 	const char *devname;
2355 	int error;
2356 
2357 	sx_xlock(&swdev_syscall_lock);
2358 
2359 	mtx_lock(&sw_dev_mtx);
2360 	TAILQ_FOREACH_SAFE(sp, &swtailq, sw_list, spt) {
2361 		mtx_unlock(&sw_dev_mtx);
2362 		if (vn_isdisk(sp->sw_vp, NULL))
2363 			devname = devtoname(sp->sw_vp->v_rdev);
2364 		else
2365 			devname = "[file]";
2366 		error = swapoff_one(sp, thread0.td_ucred);
2367 		if (error != 0) {
2368 			printf("Cannot remove swap device %s (error=%d), "
2369 			    "skipping.\n", devname, error);
2370 		} else if (bootverbose) {
2371 			printf("Swap device %s removed.\n", devname);
2372 		}
2373 		mtx_lock(&sw_dev_mtx);
2374 	}
2375 	mtx_unlock(&sw_dev_mtx);
2376 
2377 	sx_xunlock(&swdev_syscall_lock);
2378 }
2379 
2380 void
2381 swap_pager_status(int *total, int *used)
2382 {
2383 	struct swdevt *sp;
2384 
2385 	*total = 0;
2386 	*used = 0;
2387 	mtx_lock(&sw_dev_mtx);
2388 	TAILQ_FOREACH(sp, &swtailq, sw_list) {
2389 		*total += sp->sw_nblks;
2390 		*used += sp->sw_used;
2391 	}
2392 	mtx_unlock(&sw_dev_mtx);
2393 }
2394 
2395 int
2396 swap_dev_info(int name, struct xswdev *xs, char *devname, size_t len)
2397 {
2398 	struct swdevt *sp;
2399 	const char *tmp_devname;
2400 	int error, n;
2401 
2402 	n = 0;
2403 	error = ENOENT;
2404 	mtx_lock(&sw_dev_mtx);
2405 	TAILQ_FOREACH(sp, &swtailq, sw_list) {
2406 		if (n != name) {
2407 			n++;
2408 			continue;
2409 		}
2410 		xs->xsw_version = XSWDEV_VERSION;
2411 		xs->xsw_dev = sp->sw_dev;
2412 		xs->xsw_flags = sp->sw_flags;
2413 		xs->xsw_nblks = sp->sw_nblks;
2414 		xs->xsw_used = sp->sw_used;
2415 		if (devname != NULL) {
2416 			if (vn_isdisk(sp->sw_vp, NULL))
2417 				tmp_devname = devtoname(sp->sw_vp->v_rdev);
2418 			else
2419 				tmp_devname = "[file]";
2420 			strncpy(devname, tmp_devname, len);
2421 		}
2422 		error = 0;
2423 		break;
2424 	}
2425 	mtx_unlock(&sw_dev_mtx);
2426 	return (error);
2427 }
2428 
2429 #if defined(COMPAT_FREEBSD11)
2430 #define XSWDEV_VERSION_11	1
2431 struct xswdev11 {
2432 	u_int	xsw_version;
2433 	uint32_t xsw_dev;
2434 	int	xsw_flags;
2435 	int	xsw_nblks;
2436 	int     xsw_used;
2437 };
2438 #endif
2439 
2440 static int
2441 sysctl_vm_swap_info(SYSCTL_HANDLER_ARGS)
2442 {
2443 	struct xswdev xs;
2444 #if defined(COMPAT_FREEBSD11)
2445 	struct xswdev11 xs11;
2446 #endif
2447 	int error;
2448 
2449 	if (arg2 != 1)			/* name length */
2450 		return (EINVAL);
2451 	error = swap_dev_info(*(int *)arg1, &xs, NULL, 0);
2452 	if (error != 0)
2453 		return (error);
2454 #if defined(COMPAT_FREEBSD11)
2455 	if (req->oldlen == sizeof(xs11)) {
2456 		xs11.xsw_version = XSWDEV_VERSION_11;
2457 		xs11.xsw_dev = xs.xsw_dev; /* truncation */
2458 		xs11.xsw_flags = xs.xsw_flags;
2459 		xs11.xsw_nblks = xs.xsw_nblks;
2460 		xs11.xsw_used = xs.xsw_used;
2461 		error = SYSCTL_OUT(req, &xs11, sizeof(xs11));
2462 	} else
2463 #endif
2464 		error = SYSCTL_OUT(req, &xs, sizeof(xs));
2465 	return (error);
2466 }
2467 
2468 SYSCTL_INT(_vm, OID_AUTO, nswapdev, CTLFLAG_RD, &nswapdev, 0,
2469     "Number of swap devices");
2470 SYSCTL_NODE(_vm, OID_AUTO, swap_info, CTLFLAG_RD | CTLFLAG_MPSAFE,
2471     sysctl_vm_swap_info,
2472     "Swap statistics by device");
2473 
2474 /*
2475  * Count the approximate swap usage in pages for a vmspace.  The
2476  * shadowed or not yet copied on write swap blocks are not accounted.
2477  * The map must be locked.
2478  */
2479 long
2480 vmspace_swap_count(struct vmspace *vmspace)
2481 {
2482 	vm_map_t map;
2483 	vm_map_entry_t cur;
2484 	vm_object_t object;
2485 	struct swblk *sb;
2486 	vm_pindex_t e, pi;
2487 	long count;
2488 	int i;
2489 
2490 	map = &vmspace->vm_map;
2491 	count = 0;
2492 
2493 	for (cur = map->header.next; cur != &map->header; cur = cur->next) {
2494 		if ((cur->eflags & MAP_ENTRY_IS_SUB_MAP) != 0)
2495 			continue;
2496 		object = cur->object.vm_object;
2497 		if (object == NULL || object->type != OBJT_SWAP)
2498 			continue;
2499 		VM_OBJECT_RLOCK(object);
2500 		if (object->type != OBJT_SWAP)
2501 			goto unlock;
2502 		pi = OFF_TO_IDX(cur->offset);
2503 		e = pi + OFF_TO_IDX(cur->end - cur->start);
2504 		for (;; pi = sb->p + SWAP_META_PAGES) {
2505 			sb = SWAP_PCTRIE_LOOKUP_GE(
2506 			    &object->un_pager.swp.swp_blks, pi);
2507 			if (sb == NULL || sb->p >= e)
2508 				break;
2509 			for (i = 0; i < SWAP_META_PAGES; i++) {
2510 				if (sb->p + i < e &&
2511 				    sb->d[i] != SWAPBLK_NONE)
2512 					count++;
2513 			}
2514 		}
2515 unlock:
2516 		VM_OBJECT_RUNLOCK(object);
2517 	}
2518 	return (count);
2519 }
2520 
2521 /*
2522  * GEOM backend
2523  *
2524  * Swapping onto disk devices.
2525  *
2526  */
2527 
2528 static g_orphan_t swapgeom_orphan;
2529 
2530 static struct g_class g_swap_class = {
2531 	.name = "SWAP",
2532 	.version = G_VERSION,
2533 	.orphan = swapgeom_orphan,
2534 };
2535 
2536 DECLARE_GEOM_CLASS(g_swap_class, g_class);
2537 
2538 
2539 static void
2540 swapgeom_close_ev(void *arg, int flags)
2541 {
2542 	struct g_consumer *cp;
2543 
2544 	cp = arg;
2545 	g_access(cp, -1, -1, 0);
2546 	g_detach(cp);
2547 	g_destroy_consumer(cp);
2548 }
2549 
2550 /*
2551  * Add a reference to the g_consumer for an inflight transaction.
2552  */
2553 static void
2554 swapgeom_acquire(struct g_consumer *cp)
2555 {
2556 
2557 	mtx_assert(&sw_dev_mtx, MA_OWNED);
2558 	cp->index++;
2559 }
2560 
2561 /*
2562  * Remove a reference from the g_consumer.  Post a close event if all
2563  * references go away, since the function might be called from the
2564  * biodone context.
2565  */
2566 static void
2567 swapgeom_release(struct g_consumer *cp, struct swdevt *sp)
2568 {
2569 
2570 	mtx_assert(&sw_dev_mtx, MA_OWNED);
2571 	cp->index--;
2572 	if (cp->index == 0) {
2573 		if (g_post_event(swapgeom_close_ev, cp, M_NOWAIT, NULL) == 0)
2574 			sp->sw_id = NULL;
2575 	}
2576 }
2577 
2578 static void
2579 swapgeom_done(struct bio *bp2)
2580 {
2581 	struct swdevt *sp;
2582 	struct buf *bp;
2583 	struct g_consumer *cp;
2584 
2585 	bp = bp2->bio_caller2;
2586 	cp = bp2->bio_from;
2587 	bp->b_ioflags = bp2->bio_flags;
2588 	if (bp2->bio_error)
2589 		bp->b_ioflags |= BIO_ERROR;
2590 	bp->b_resid = bp->b_bcount - bp2->bio_completed;
2591 	bp->b_error = bp2->bio_error;
2592 	bufdone(bp);
2593 	sp = bp2->bio_caller1;
2594 	mtx_lock(&sw_dev_mtx);
2595 	swapgeom_release(cp, sp);
2596 	mtx_unlock(&sw_dev_mtx);
2597 	g_destroy_bio(bp2);
2598 }
2599 
2600 static void
2601 swapgeom_strategy(struct buf *bp, struct swdevt *sp)
2602 {
2603 	struct bio *bio;
2604 	struct g_consumer *cp;
2605 
2606 	mtx_lock(&sw_dev_mtx);
2607 	cp = sp->sw_id;
2608 	if (cp == NULL) {
2609 		mtx_unlock(&sw_dev_mtx);
2610 		bp->b_error = ENXIO;
2611 		bp->b_ioflags |= BIO_ERROR;
2612 		bufdone(bp);
2613 		return;
2614 	}
2615 	swapgeom_acquire(cp);
2616 	mtx_unlock(&sw_dev_mtx);
2617 	if (bp->b_iocmd == BIO_WRITE)
2618 		bio = g_new_bio();
2619 	else
2620 		bio = g_alloc_bio();
2621 	if (bio == NULL) {
2622 		mtx_lock(&sw_dev_mtx);
2623 		swapgeom_release(cp, sp);
2624 		mtx_unlock(&sw_dev_mtx);
2625 		bp->b_error = ENOMEM;
2626 		bp->b_ioflags |= BIO_ERROR;
2627 		bufdone(bp);
2628 		return;
2629 	}
2630 
2631 	bio->bio_caller1 = sp;
2632 	bio->bio_caller2 = bp;
2633 	bio->bio_cmd = bp->b_iocmd;
2634 	bio->bio_offset = (bp->b_blkno - sp->sw_first) * PAGE_SIZE;
2635 	bio->bio_length = bp->b_bcount;
2636 	bio->bio_done = swapgeom_done;
2637 	if (!buf_mapped(bp)) {
2638 		bio->bio_ma = bp->b_pages;
2639 		bio->bio_data = unmapped_buf;
2640 		bio->bio_ma_offset = (vm_offset_t)bp->b_offset & PAGE_MASK;
2641 		bio->bio_ma_n = bp->b_npages;
2642 		bio->bio_flags |= BIO_UNMAPPED;
2643 	} else {
2644 		bio->bio_data = bp->b_data;
2645 		bio->bio_ma = NULL;
2646 	}
2647 	g_io_request(bio, cp);
2648 	return;
2649 }
2650 
2651 static void
2652 swapgeom_orphan(struct g_consumer *cp)
2653 {
2654 	struct swdevt *sp;
2655 	int destroy;
2656 
2657 	mtx_lock(&sw_dev_mtx);
2658 	TAILQ_FOREACH(sp, &swtailq, sw_list) {
2659 		if (sp->sw_id == cp) {
2660 			sp->sw_flags |= SW_CLOSING;
2661 			break;
2662 		}
2663 	}
2664 	/*
2665 	 * Drop reference we were created with. Do directly since we're in a
2666 	 * special context where we don't have to queue the call to
2667 	 * swapgeom_close_ev().
2668 	 */
2669 	cp->index--;
2670 	destroy = ((sp != NULL) && (cp->index == 0));
2671 	if (destroy)
2672 		sp->sw_id = NULL;
2673 	mtx_unlock(&sw_dev_mtx);
2674 	if (destroy)
2675 		swapgeom_close_ev(cp, 0);
2676 }
2677 
2678 static void
2679 swapgeom_close(struct thread *td, struct swdevt *sw)
2680 {
2681 	struct g_consumer *cp;
2682 
2683 	mtx_lock(&sw_dev_mtx);
2684 	cp = sw->sw_id;
2685 	sw->sw_id = NULL;
2686 	mtx_unlock(&sw_dev_mtx);
2687 
2688 	/*
2689 	 * swapgeom_close() may be called from the biodone context,
2690 	 * where we cannot perform topology changes.  Delegate the
2691 	 * work to the events thread.
2692 	 */
2693 	if (cp != NULL)
2694 		g_waitfor_event(swapgeom_close_ev, cp, M_WAITOK, NULL);
2695 }
2696 
2697 static int
2698 swapongeom_locked(struct cdev *dev, struct vnode *vp)
2699 {
2700 	struct g_provider *pp;
2701 	struct g_consumer *cp;
2702 	static struct g_geom *gp;
2703 	struct swdevt *sp;
2704 	u_long nblks;
2705 	int error;
2706 
2707 	pp = g_dev_getprovider(dev);
2708 	if (pp == NULL)
2709 		return (ENODEV);
2710 	mtx_lock(&sw_dev_mtx);
2711 	TAILQ_FOREACH(sp, &swtailq, sw_list) {
2712 		cp = sp->sw_id;
2713 		if (cp != NULL && cp->provider == pp) {
2714 			mtx_unlock(&sw_dev_mtx);
2715 			return (EBUSY);
2716 		}
2717 	}
2718 	mtx_unlock(&sw_dev_mtx);
2719 	if (gp == NULL)
2720 		gp = g_new_geomf(&g_swap_class, "swap");
2721 	cp = g_new_consumer(gp);
2722 	cp->index = 1;	/* Number of active I/Os, plus one for being active. */
2723 	cp->flags |=  G_CF_DIRECT_SEND | G_CF_DIRECT_RECEIVE;
2724 	g_attach(cp, pp);
2725 	/*
2726 	 * XXX: Every time you think you can improve the margin for
2727 	 * footshooting, somebody depends on the ability to do so:
2728 	 * savecore(8) wants to write to our swapdev so we cannot
2729 	 * set an exclusive count :-(
2730 	 */
2731 	error = g_access(cp, 1, 1, 0);
2732 	if (error != 0) {
2733 		g_detach(cp);
2734 		g_destroy_consumer(cp);
2735 		return (error);
2736 	}
2737 	nblks = pp->mediasize / DEV_BSIZE;
2738 	swaponsomething(vp, cp, nblks, swapgeom_strategy,
2739 	    swapgeom_close, dev2udev(dev),
2740 	    (pp->flags & G_PF_ACCEPT_UNMAPPED) != 0 ? SW_UNMAPPED : 0);
2741 	return (0);
2742 }
2743 
2744 static int
2745 swapongeom(struct vnode *vp)
2746 {
2747 	int error;
2748 
2749 	vn_lock(vp, LK_EXCLUSIVE | LK_RETRY);
2750 	if (vp->v_type != VCHR || (vp->v_iflag & VI_DOOMED) != 0) {
2751 		error = ENOENT;
2752 	} else {
2753 		g_topology_lock();
2754 		error = swapongeom_locked(vp->v_rdev, vp);
2755 		g_topology_unlock();
2756 	}
2757 	VOP_UNLOCK(vp, 0);
2758 	return (error);
2759 }
2760 
2761 /*
2762  * VNODE backend
2763  *
2764  * This is used mainly for network filesystem (read: probably only tested
2765  * with NFS) swapfiles.
2766  *
2767  */
2768 
2769 static void
2770 swapdev_strategy(struct buf *bp, struct swdevt *sp)
2771 {
2772 	struct vnode *vp2;
2773 
2774 	bp->b_blkno = ctodb(bp->b_blkno - sp->sw_first);
2775 
2776 	vp2 = sp->sw_id;
2777 	vhold(vp2);
2778 	if (bp->b_iocmd == BIO_WRITE) {
2779 		if (bp->b_bufobj)
2780 			bufobj_wdrop(bp->b_bufobj);
2781 		bufobj_wref(&vp2->v_bufobj);
2782 	}
2783 	if (bp->b_bufobj != &vp2->v_bufobj)
2784 		bp->b_bufobj = &vp2->v_bufobj;
2785 	bp->b_vp = vp2;
2786 	bp->b_iooffset = dbtob(bp->b_blkno);
2787 	bstrategy(bp);
2788 	return;
2789 }
2790 
2791 static void
2792 swapdev_close(struct thread *td, struct swdevt *sp)
2793 {
2794 
2795 	VOP_CLOSE(sp->sw_vp, FREAD | FWRITE, td->td_ucred, td);
2796 	vrele(sp->sw_vp);
2797 }
2798 
2799 
2800 static int
2801 swaponvp(struct thread *td, struct vnode *vp, u_long nblks)
2802 {
2803 	struct swdevt *sp;
2804 	int error;
2805 
2806 	if (nblks == 0)
2807 		return (ENXIO);
2808 	mtx_lock(&sw_dev_mtx);
2809 	TAILQ_FOREACH(sp, &swtailq, sw_list) {
2810 		if (sp->sw_id == vp) {
2811 			mtx_unlock(&sw_dev_mtx);
2812 			return (EBUSY);
2813 		}
2814 	}
2815 	mtx_unlock(&sw_dev_mtx);
2816 
2817 	(void) vn_lock(vp, LK_EXCLUSIVE | LK_RETRY);
2818 #ifdef MAC
2819 	error = mac_system_check_swapon(td->td_ucred, vp);
2820 	if (error == 0)
2821 #endif
2822 		error = VOP_OPEN(vp, FREAD | FWRITE, td->td_ucred, td, NULL);
2823 	(void) VOP_UNLOCK(vp, 0);
2824 	if (error)
2825 		return (error);
2826 
2827 	swaponsomething(vp, vp, nblks, swapdev_strategy, swapdev_close,
2828 	    NODEV, 0);
2829 	return (0);
2830 }
2831 
2832 static int
2833 sysctl_swap_async_max(SYSCTL_HANDLER_ARGS)
2834 {
2835 	int error, new, n;
2836 
2837 	new = nsw_wcount_async_max;
2838 	error = sysctl_handle_int(oidp, &new, 0, req);
2839 	if (error != 0 || req->newptr == NULL)
2840 		return (error);
2841 
2842 	if (new > nswbuf / 2 || new < 1)
2843 		return (EINVAL);
2844 
2845 	mtx_lock(&pbuf_mtx);
2846 	while (nsw_wcount_async_max != new) {
2847 		/*
2848 		 * Adjust difference.  If the current async count is too low,
2849 		 * we will need to sqeeze our update slowly in.  Sleep with a
2850 		 * higher priority than getpbuf() to finish faster.
2851 		 */
2852 		n = new - nsw_wcount_async_max;
2853 		if (nsw_wcount_async + n >= 0) {
2854 			nsw_wcount_async += n;
2855 			nsw_wcount_async_max += n;
2856 			wakeup(&nsw_wcount_async);
2857 		} else {
2858 			nsw_wcount_async_max -= nsw_wcount_async;
2859 			nsw_wcount_async = 0;
2860 			msleep(&nsw_wcount_async, &pbuf_mtx, PSWP,
2861 			    "swpsysctl", 0);
2862 		}
2863 	}
2864 	mtx_unlock(&pbuf_mtx);
2865 
2866 	return (0);
2867 }
2868