xref: /freebsd/sys/vm/swap_pager.c (revision 6829dae12bb055451fa467da4589c43bd03b1e64)
1 /*-
2  * SPDX-License-Identifier: BSD-4-Clause
3  *
4  * Copyright (c) 1998 Matthew Dillon,
5  * Copyright (c) 1994 John S. Dyson
6  * Copyright (c) 1990 University of Utah.
7  * Copyright (c) 1982, 1986, 1989, 1993
8  *	The Regents of the University of California.  All rights reserved.
9  *
10  * This code is derived from software contributed to Berkeley by
11  * the Systems Programming Group of the University of Utah Computer
12  * Science Department.
13  *
14  * Redistribution and use in source and binary forms, with or without
15  * modification, are permitted provided that the following conditions
16  * are met:
17  * 1. Redistributions of source code must retain the above copyright
18  *    notice, this list of conditions and the following disclaimer.
19  * 2. Redistributions in binary form must reproduce the above copyright
20  *    notice, this list of conditions and the following disclaimer in the
21  *    documentation and/or other materials provided with the distribution.
22  * 3. All advertising materials mentioning features or use of this software
23  *    must display the following acknowledgement:
24  *	This product includes software developed by the University of
25  *	California, Berkeley and its contributors.
26  * 4. Neither the name of the University nor the names of its contributors
27  *    may be used to endorse or promote products derived from this software
28  *    without specific prior written permission.
29  *
30  * THIS SOFTWARE IS PROVIDED BY THE REGENTS AND CONTRIBUTORS ``AS IS'' AND
31  * ANY EXPRESS OR IMPLIED WARRANTIES, INCLUDING, BUT NOT LIMITED TO, THE
32  * IMPLIED WARRANTIES OF MERCHANTABILITY AND FITNESS FOR A PARTICULAR PURPOSE
33  * ARE DISCLAIMED.  IN NO EVENT SHALL THE REGENTS OR CONTRIBUTORS BE LIABLE
34  * FOR ANY DIRECT, INDIRECT, INCIDENTAL, SPECIAL, EXEMPLARY, OR CONSEQUENTIAL
35  * DAMAGES (INCLUDING, BUT NOT LIMITED TO, PROCUREMENT OF SUBSTITUTE GOODS
36  * OR SERVICES; LOSS OF USE, DATA, OR PROFITS; OR BUSINESS INTERRUPTION)
37  * HOWEVER CAUSED AND ON ANY THEORY OF LIABILITY, WHETHER IN CONTRACT, STRICT
38  * LIABILITY, OR TORT (INCLUDING NEGLIGENCE OR OTHERWISE) ARISING IN ANY WAY
39  * OUT OF THE USE OF THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF
40  * SUCH DAMAGE.
41  *
42  *				New Swap System
43  *				Matthew Dillon
44  *
45  * Radix Bitmap 'blists'.
46  *
47  *	- The new swapper uses the new radix bitmap code.  This should scale
48  *	  to arbitrarily small or arbitrarily large swap spaces and an almost
49  *	  arbitrary degree of fragmentation.
50  *
51  * Features:
52  *
53  *	- on the fly reallocation of swap during putpages.  The new system
54  *	  does not try to keep previously allocated swap blocks for dirty
55  *	  pages.
56  *
57  *	- on the fly deallocation of swap
58  *
59  *	- No more garbage collection required.  Unnecessarily allocated swap
60  *	  blocks only exist for dirty vm_page_t's now and these are already
61  *	  cycled (in a high-load system) by the pager.  We also do on-the-fly
62  *	  removal of invalidated swap blocks when a page is destroyed
63  *	  or renamed.
64  *
65  * from: Utah $Hdr: swap_pager.c 1.4 91/04/30$
66  *
67  *	@(#)swap_pager.c	8.9 (Berkeley) 3/21/94
68  *	@(#)vm_swap.c	8.5 (Berkeley) 2/17/94
69  */
70 
71 #include <sys/cdefs.h>
72 __FBSDID("$FreeBSD$");
73 
74 #include "opt_vm.h"
75 
76 #include <sys/param.h>
77 #include <sys/systm.h>
78 #include <sys/conf.h>
79 #include <sys/kernel.h>
80 #include <sys/priv.h>
81 #include <sys/proc.h>
82 #include <sys/bio.h>
83 #include <sys/buf.h>
84 #include <sys/disk.h>
85 #include <sys/fcntl.h>
86 #include <sys/mount.h>
87 #include <sys/namei.h>
88 #include <sys/vnode.h>
89 #include <sys/malloc.h>
90 #include <sys/pctrie.h>
91 #include <sys/racct.h>
92 #include <sys/resource.h>
93 #include <sys/resourcevar.h>
94 #include <sys/rwlock.h>
95 #include <sys/sbuf.h>
96 #include <sys/sysctl.h>
97 #include <sys/sysproto.h>
98 #include <sys/blist.h>
99 #include <sys/lock.h>
100 #include <sys/sx.h>
101 #include <sys/vmmeter.h>
102 
103 #include <security/mac/mac_framework.h>
104 
105 #include <vm/vm.h>
106 #include <vm/pmap.h>
107 #include <vm/vm_map.h>
108 #include <vm/vm_kern.h>
109 #include <vm/vm_object.h>
110 #include <vm/vm_page.h>
111 #include <vm/vm_pager.h>
112 #include <vm/vm_pageout.h>
113 #include <vm/vm_param.h>
114 #include <vm/swap_pager.h>
115 #include <vm/vm_extern.h>
116 #include <vm/uma.h>
117 
118 #include <geom/geom.h>
119 
120 /*
121  * MAX_PAGEOUT_CLUSTER must be a power of 2 between 1 and 64.
122  * The 64-page limit is due to the radix code (kern/subr_blist.c).
123  */
124 #ifndef MAX_PAGEOUT_CLUSTER
125 #define	MAX_PAGEOUT_CLUSTER	32
126 #endif
127 
128 #if !defined(SWB_NPAGES)
129 #define SWB_NPAGES	MAX_PAGEOUT_CLUSTER
130 #endif
131 
132 #define	SWAP_META_PAGES		PCTRIE_COUNT
133 
134 /*
135  * A swblk structure maps each page index within a
136  * SWAP_META_PAGES-aligned and sized range to the address of an
137  * on-disk swap block (or SWAPBLK_NONE). The collection of these
138  * mappings for an entire vm object is implemented as a pc-trie.
139  */
140 struct swblk {
141 	vm_pindex_t	p;
142 	daddr_t		d[SWAP_META_PAGES];
143 };
144 
145 static MALLOC_DEFINE(M_VMPGDATA, "vm_pgdata", "swap pager private data");
146 static struct mtx sw_dev_mtx;
147 static TAILQ_HEAD(, swdevt) swtailq = TAILQ_HEAD_INITIALIZER(swtailq);
148 static struct swdevt *swdevhd;	/* Allocate from here next */
149 static int nswapdev;		/* Number of swap devices */
150 int swap_pager_avail;
151 static struct sx swdev_syscall_lock;	/* serialize swap(on|off) */
152 
153 static u_long swap_reserved;
154 static u_long swap_total;
155 static int sysctl_page_shift(SYSCTL_HANDLER_ARGS);
156 SYSCTL_PROC(_vm, OID_AUTO, swap_reserved, CTLTYPE_U64 | CTLFLAG_RD | CTLFLAG_MPSAFE,
157     &swap_reserved, 0, sysctl_page_shift, "A",
158     "Amount of swap storage needed to back all allocated anonymous memory.");
159 SYSCTL_PROC(_vm, OID_AUTO, swap_total, CTLTYPE_U64 | CTLFLAG_RD | CTLFLAG_MPSAFE,
160     &swap_total, 0, sysctl_page_shift, "A",
161     "Total amount of available swap storage.");
162 
163 static int overcommit = 0;
164 SYSCTL_INT(_vm, VM_OVERCOMMIT, overcommit, CTLFLAG_RW, &overcommit, 0,
165     "Configure virtual memory overcommit behavior. See tuning(7) "
166     "for details.");
167 static unsigned long swzone;
168 SYSCTL_ULONG(_vm, OID_AUTO, swzone, CTLFLAG_RD, &swzone, 0,
169     "Actual size of swap metadata zone");
170 static unsigned long swap_maxpages;
171 SYSCTL_ULONG(_vm, OID_AUTO, swap_maxpages, CTLFLAG_RD, &swap_maxpages, 0,
172     "Maximum amount of swap supported");
173 
174 /* bits from overcommit */
175 #define	SWAP_RESERVE_FORCE_ON		(1 << 0)
176 #define	SWAP_RESERVE_RLIMIT_ON		(1 << 1)
177 #define	SWAP_RESERVE_ALLOW_NONWIRED	(1 << 2)
178 
179 static int
180 sysctl_page_shift(SYSCTL_HANDLER_ARGS)
181 {
182 	uint64_t newval;
183 	u_long value = *(u_long *)arg1;
184 
185 	newval = ((uint64_t)value) << PAGE_SHIFT;
186 	return (sysctl_handle_64(oidp, &newval, 0, req));
187 }
188 
189 int
190 swap_reserve(vm_ooffset_t incr)
191 {
192 
193 	return (swap_reserve_by_cred(incr, curthread->td_ucred));
194 }
195 
196 int
197 swap_reserve_by_cred(vm_ooffset_t incr, struct ucred *cred)
198 {
199 	u_long r, s, prev, pincr;
200 	int res, error;
201 	static int curfail;
202 	static struct timeval lastfail;
203 	struct uidinfo *uip;
204 
205 	uip = cred->cr_ruidinfo;
206 
207 	KASSERT((incr & PAGE_MASK) == 0, ("%s: incr: %ju & PAGE_MASK", __func__,
208 	    (uintmax_t)incr));
209 
210 #ifdef RACCT
211 	if (racct_enable) {
212 		PROC_LOCK(curproc);
213 		error = racct_add(curproc, RACCT_SWAP, incr);
214 		PROC_UNLOCK(curproc);
215 		if (error != 0)
216 			return (0);
217 	}
218 #endif
219 
220 	pincr = atop(incr);
221 	res = 0;
222 	prev = atomic_fetchadd_long(&swap_reserved, pincr);
223 	r = prev + pincr;
224 	if (overcommit & SWAP_RESERVE_ALLOW_NONWIRED) {
225 		s = vm_cnt.v_page_count - vm_cnt.v_free_reserved -
226 		    vm_wire_count();
227 	} else
228 		s = 0;
229 	s += swap_total;
230 	if ((overcommit & SWAP_RESERVE_FORCE_ON) == 0 || r <= s ||
231 	    (error = priv_check(curthread, PRIV_VM_SWAP_NOQUOTA)) == 0) {
232 		res = 1;
233 	} else {
234 		prev = atomic_fetchadd_long(&swap_reserved, -pincr);
235 		if (prev < pincr)
236 			panic("swap_reserved < incr on overcommit fail");
237 	}
238 	if (res) {
239 		prev = atomic_fetchadd_long(&uip->ui_vmsize, pincr);
240 		if ((overcommit & SWAP_RESERVE_RLIMIT_ON) != 0 &&
241 		    prev + pincr > lim_cur(curthread, RLIMIT_SWAP) &&
242 		    priv_check(curthread, PRIV_VM_SWAP_NORLIMIT)) {
243 			res = 0;
244 			prev = atomic_fetchadd_long(&uip->ui_vmsize, -pincr);
245 			if (prev < pincr)
246 				panic("uip->ui_vmsize < incr on overcommit fail");
247 		}
248 	}
249 	if (!res && ppsratecheck(&lastfail, &curfail, 1)) {
250 		printf("uid %d, pid %d: swap reservation for %jd bytes failed\n",
251 		    uip->ui_uid, curproc->p_pid, incr);
252 	}
253 
254 #ifdef RACCT
255 	if (racct_enable && !res) {
256 		PROC_LOCK(curproc);
257 		racct_sub(curproc, RACCT_SWAP, incr);
258 		PROC_UNLOCK(curproc);
259 	}
260 #endif
261 
262 	return (res);
263 }
264 
265 void
266 swap_reserve_force(vm_ooffset_t incr)
267 {
268 	struct uidinfo *uip;
269 	u_long pincr;
270 
271 	KASSERT((incr & PAGE_MASK) == 0, ("%s: incr: %ju & PAGE_MASK", __func__,
272 	    (uintmax_t)incr));
273 
274 	PROC_LOCK(curproc);
275 #ifdef RACCT
276 	if (racct_enable)
277 		racct_add_force(curproc, RACCT_SWAP, incr);
278 #endif
279 	pincr = atop(incr);
280 	atomic_add_long(&swap_reserved, pincr);
281 	uip = curproc->p_ucred->cr_ruidinfo;
282 	atomic_add_long(&uip->ui_vmsize, pincr);
283 	PROC_UNLOCK(curproc);
284 }
285 
286 void
287 swap_release(vm_ooffset_t decr)
288 {
289 	struct ucred *cred;
290 
291 	PROC_LOCK(curproc);
292 	cred = curproc->p_ucred;
293 	swap_release_by_cred(decr, cred);
294 	PROC_UNLOCK(curproc);
295 }
296 
297 void
298 swap_release_by_cred(vm_ooffset_t decr, struct ucred *cred)
299 {
300 	u_long prev, pdecr;
301  	struct uidinfo *uip;
302 
303 	uip = cred->cr_ruidinfo;
304 
305 	KASSERT((decr & PAGE_MASK) == 0, ("%s: decr: %ju & PAGE_MASK", __func__,
306 	    (uintmax_t)decr));
307 
308 	pdecr = atop(decr);
309 	prev = atomic_fetchadd_long(&swap_reserved, -pdecr);
310 	if (prev < pdecr)
311 		panic("swap_reserved < decr");
312 
313 	prev = atomic_fetchadd_long(&uip->ui_vmsize, -pdecr);
314 	if (prev < pdecr)
315 		printf("negative vmsize for uid = %d\n", uip->ui_uid);
316 #ifdef RACCT
317 	if (racct_enable)
318 		racct_sub_cred(cred, RACCT_SWAP, decr);
319 #endif
320 }
321 
322 #define SWM_POP		0x01	/* pop out			*/
323 
324 static int swap_pager_full = 2;	/* swap space exhaustion (task killing) */
325 static int swap_pager_almost_full = 1; /* swap space exhaustion (w/hysteresis)*/
326 static struct mtx swbuf_mtx;	/* to sync nsw_wcount_async */
327 static int nsw_wcount_async;	/* limit async write buffers */
328 static int nsw_wcount_async_max;/* assigned maximum			*/
329 static int nsw_cluster_max;	/* maximum VOP I/O allowed		*/
330 
331 static int sysctl_swap_async_max(SYSCTL_HANDLER_ARGS);
332 SYSCTL_PROC(_vm, OID_AUTO, swap_async_max, CTLTYPE_INT | CTLFLAG_RW |
333     CTLFLAG_MPSAFE, NULL, 0, sysctl_swap_async_max, "I",
334     "Maximum running async swap ops");
335 static int sysctl_swap_fragmentation(SYSCTL_HANDLER_ARGS);
336 SYSCTL_PROC(_vm, OID_AUTO, swap_fragmentation, CTLTYPE_STRING | CTLFLAG_RD |
337     CTLFLAG_MPSAFE, NULL, 0, sysctl_swap_fragmentation, "A",
338     "Swap Fragmentation Info");
339 
340 static struct sx sw_alloc_sx;
341 
342 /*
343  * "named" and "unnamed" anon region objects.  Try to reduce the overhead
344  * of searching a named list by hashing it just a little.
345  */
346 
347 #define NOBJLISTS		8
348 
349 #define NOBJLIST(handle)	\
350 	(&swap_pager_object_list[((int)(intptr_t)handle >> 4) & (NOBJLISTS-1)])
351 
352 static struct pagerlst	swap_pager_object_list[NOBJLISTS];
353 static uma_zone_t swwbuf_zone;
354 static uma_zone_t swrbuf_zone;
355 static uma_zone_t swblk_zone;
356 static uma_zone_t swpctrie_zone;
357 
358 /*
359  * pagerops for OBJT_SWAP - "swap pager".  Some ops are also global procedure
360  * calls hooked from other parts of the VM system and do not appear here.
361  * (see vm/swap_pager.h).
362  */
363 static vm_object_t
364 		swap_pager_alloc(void *handle, vm_ooffset_t size,
365 		    vm_prot_t prot, vm_ooffset_t offset, struct ucred *);
366 static void	swap_pager_dealloc(vm_object_t object);
367 static int	swap_pager_getpages(vm_object_t, vm_page_t *, int, int *,
368     int *);
369 static int	swap_pager_getpages_async(vm_object_t, vm_page_t *, int, int *,
370     int *, pgo_getpages_iodone_t, void *);
371 static void	swap_pager_putpages(vm_object_t, vm_page_t *, int, boolean_t, int *);
372 static boolean_t
373 		swap_pager_haspage(vm_object_t object, vm_pindex_t pindex, int *before, int *after);
374 static void	swap_pager_init(void);
375 static void	swap_pager_unswapped(vm_page_t);
376 static void	swap_pager_swapoff(struct swdevt *sp);
377 
378 struct pagerops swappagerops = {
379 	.pgo_init =	swap_pager_init,	/* early system initialization of pager	*/
380 	.pgo_alloc =	swap_pager_alloc,	/* allocate an OBJT_SWAP object		*/
381 	.pgo_dealloc =	swap_pager_dealloc,	/* deallocate an OBJT_SWAP object	*/
382 	.pgo_getpages =	swap_pager_getpages,	/* pagein				*/
383 	.pgo_getpages_async = swap_pager_getpages_async, /* pagein (async)		*/
384 	.pgo_putpages =	swap_pager_putpages,	/* pageout				*/
385 	.pgo_haspage =	swap_pager_haspage,	/* get backing store status for page	*/
386 	.pgo_pageunswapped = swap_pager_unswapped,	/* remove swap related to page		*/
387 };
388 
389 /*
390  * swap_*() routines are externally accessible.  swp_*() routines are
391  * internal.
392  */
393 static int nswap_lowat = 128;	/* in pages, swap_pager_almost_full warn */
394 static int nswap_hiwat = 512;	/* in pages, swap_pager_almost_full warn */
395 
396 SYSCTL_INT(_vm, OID_AUTO, dmmax, CTLFLAG_RD, &nsw_cluster_max, 0,
397     "Maximum size of a swap block in pages");
398 
399 static void	swp_sizecheck(void);
400 static void	swp_pager_async_iodone(struct buf *bp);
401 static bool	swp_pager_swblk_empty(struct swblk *sb, int start, int limit);
402 static int	swapongeom(struct vnode *);
403 static int	swaponvp(struct thread *, struct vnode *, u_long);
404 static int	swapoff_one(struct swdevt *sp, struct ucred *cred);
405 
406 /*
407  * Swap bitmap functions
408  */
409 static void	swp_pager_freeswapspace(daddr_t blk, daddr_t npages);
410 static daddr_t	swp_pager_getswapspace(int npages);
411 
412 /*
413  * Metadata functions
414  */
415 static daddr_t swp_pager_meta_build(vm_object_t, vm_pindex_t, daddr_t);
416 static void swp_pager_meta_free(vm_object_t, vm_pindex_t, vm_pindex_t);
417 static void swp_pager_meta_free_all(vm_object_t);
418 static daddr_t swp_pager_meta_ctl(vm_object_t, vm_pindex_t, int);
419 
420 static void
421 swp_pager_init_freerange(daddr_t *start, daddr_t *num)
422 {
423 
424 	*start = SWAPBLK_NONE;
425 	*num = 0;
426 }
427 
428 static void
429 swp_pager_update_freerange(daddr_t *start, daddr_t *num, daddr_t addr)
430 {
431 
432 	if (*start + *num == addr) {
433 		(*num)++;
434 	} else {
435 		swp_pager_freeswapspace(*start, *num);
436 		*start = addr;
437 		*num = 1;
438 	}
439 }
440 
441 static void *
442 swblk_trie_alloc(struct pctrie *ptree)
443 {
444 
445 	return (uma_zalloc(swpctrie_zone, M_NOWAIT | (curproc == pageproc ?
446 	    M_USE_RESERVE : 0)));
447 }
448 
449 static void
450 swblk_trie_free(struct pctrie *ptree, void *node)
451 {
452 
453 	uma_zfree(swpctrie_zone, node);
454 }
455 
456 PCTRIE_DEFINE(SWAP, swblk, p, swblk_trie_alloc, swblk_trie_free);
457 
458 /*
459  * SWP_SIZECHECK() -	update swap_pager_full indication
460  *
461  *	update the swap_pager_almost_full indication and warn when we are
462  *	about to run out of swap space, using lowat/hiwat hysteresis.
463  *
464  *	Clear swap_pager_full ( task killing ) indication when lowat is met.
465  *
466  *	No restrictions on call
467  *	This routine may not block.
468  */
469 static void
470 swp_sizecheck(void)
471 {
472 
473 	if (swap_pager_avail < nswap_lowat) {
474 		if (swap_pager_almost_full == 0) {
475 			printf("swap_pager: out of swap space\n");
476 			swap_pager_almost_full = 1;
477 		}
478 	} else {
479 		swap_pager_full = 0;
480 		if (swap_pager_avail > nswap_hiwat)
481 			swap_pager_almost_full = 0;
482 	}
483 }
484 
485 /*
486  * SWAP_PAGER_INIT() -	initialize the swap pager!
487  *
488  *	Expected to be started from system init.  NOTE:  This code is run
489  *	before much else so be careful what you depend on.  Most of the VM
490  *	system has yet to be initialized at this point.
491  */
492 static void
493 swap_pager_init(void)
494 {
495 	/*
496 	 * Initialize object lists
497 	 */
498 	int i;
499 
500 	for (i = 0; i < NOBJLISTS; ++i)
501 		TAILQ_INIT(&swap_pager_object_list[i]);
502 	mtx_init(&sw_dev_mtx, "swapdev", NULL, MTX_DEF);
503 	sx_init(&sw_alloc_sx, "swspsx");
504 	sx_init(&swdev_syscall_lock, "swsysc");
505 }
506 
507 /*
508  * SWAP_PAGER_SWAP_INIT() - swap pager initialization from pageout process
509  *
510  *	Expected to be started from pageout process once, prior to entering
511  *	its main loop.
512  */
513 void
514 swap_pager_swap_init(void)
515 {
516 	unsigned long n, n2;
517 
518 	/*
519 	 * Number of in-transit swap bp operations.  Don't
520 	 * exhaust the pbufs completely.  Make sure we
521 	 * initialize workable values (0 will work for hysteresis
522 	 * but it isn't very efficient).
523 	 *
524 	 * The nsw_cluster_max is constrained by the bp->b_pages[]
525 	 * array (MAXPHYS/PAGE_SIZE) and our locally defined
526 	 * MAX_PAGEOUT_CLUSTER.   Also be aware that swap ops are
527 	 * constrained by the swap device interleave stripe size.
528 	 *
529 	 * Currently we hardwire nsw_wcount_async to 4.  This limit is
530 	 * designed to prevent other I/O from having high latencies due to
531 	 * our pageout I/O.  The value 4 works well for one or two active swap
532 	 * devices but is probably a little low if you have more.  Even so,
533 	 * a higher value would probably generate only a limited improvement
534 	 * with three or four active swap devices since the system does not
535 	 * typically have to pageout at extreme bandwidths.   We will want
536 	 * at least 2 per swap devices, and 4 is a pretty good value if you
537 	 * have one NFS swap device due to the command/ack latency over NFS.
538 	 * So it all works out pretty well.
539 	 */
540 	nsw_cluster_max = min((MAXPHYS/PAGE_SIZE), MAX_PAGEOUT_CLUSTER);
541 
542 	nsw_wcount_async = 4;
543 	nsw_wcount_async_max = nsw_wcount_async;
544 	mtx_init(&swbuf_mtx, "async swbuf mutex", NULL, MTX_DEF);
545 
546 	swwbuf_zone = pbuf_zsecond_create("swwbuf", nswbuf / 4);
547 	swrbuf_zone = pbuf_zsecond_create("swrbuf", nswbuf / 2);
548 
549 	/*
550 	 * Initialize our zone, taking the user's requested size or
551 	 * estimating the number we need based on the number of pages
552 	 * in the system.
553 	 */
554 	n = maxswzone != 0 ? maxswzone / sizeof(struct swblk) :
555 	    vm_cnt.v_page_count / 2;
556 	swpctrie_zone = uma_zcreate("swpctrie", pctrie_node_size(), NULL, NULL,
557 	    pctrie_zone_init, NULL, UMA_ALIGN_PTR, UMA_ZONE_VM);
558 	if (swpctrie_zone == NULL)
559 		panic("failed to create swap pctrie zone.");
560 	swblk_zone = uma_zcreate("swblk", sizeof(struct swblk), NULL, NULL,
561 	    NULL, NULL, _Alignof(struct swblk) - 1, UMA_ZONE_VM);
562 	if (swblk_zone == NULL)
563 		panic("failed to create swap blk zone.");
564 	n2 = n;
565 	do {
566 		if (uma_zone_reserve_kva(swblk_zone, n))
567 			break;
568 		/*
569 		 * if the allocation failed, try a zone two thirds the
570 		 * size of the previous attempt.
571 		 */
572 		n -= ((n + 2) / 3);
573 	} while (n > 0);
574 
575 	/*
576 	 * Often uma_zone_reserve_kva() cannot reserve exactly the
577 	 * requested size.  Account for the difference when
578 	 * calculating swap_maxpages.
579 	 */
580 	n = uma_zone_get_max(swblk_zone);
581 
582 	if (n < n2)
583 		printf("Swap blk zone entries changed from %lu to %lu.\n",
584 		    n2, n);
585 	swap_maxpages = n * SWAP_META_PAGES;
586 	swzone = n * sizeof(struct swblk);
587 	if (!uma_zone_reserve_kva(swpctrie_zone, n))
588 		printf("Cannot reserve swap pctrie zone, "
589 		    "reduce kern.maxswzone.\n");
590 }
591 
592 static vm_object_t
593 swap_pager_alloc_init(void *handle, struct ucred *cred, vm_ooffset_t size,
594     vm_ooffset_t offset)
595 {
596 	vm_object_t object;
597 
598 	if (cred != NULL) {
599 		if (!swap_reserve_by_cred(size, cred))
600 			return (NULL);
601 		crhold(cred);
602 	}
603 
604 	/*
605 	 * The un_pager.swp.swp_blks trie is initialized by
606 	 * vm_object_allocate() to ensure the correct order of
607 	 * visibility to other threads.
608 	 */
609 	object = vm_object_allocate(OBJT_SWAP, OFF_TO_IDX(offset +
610 	    PAGE_MASK + size));
611 
612 	object->handle = handle;
613 	if (cred != NULL) {
614 		object->cred = cred;
615 		object->charge = size;
616 	}
617 	return (object);
618 }
619 
620 /*
621  * SWAP_PAGER_ALLOC() -	allocate a new OBJT_SWAP VM object and instantiate
622  *			its metadata structures.
623  *
624  *	This routine is called from the mmap and fork code to create a new
625  *	OBJT_SWAP object.
626  *
627  *	This routine must ensure that no live duplicate is created for
628  *	the named object request, which is protected against by
629  *	holding the sw_alloc_sx lock in case handle != NULL.
630  */
631 static vm_object_t
632 swap_pager_alloc(void *handle, vm_ooffset_t size, vm_prot_t prot,
633     vm_ooffset_t offset, struct ucred *cred)
634 {
635 	vm_object_t object;
636 
637 	if (handle != NULL) {
638 		/*
639 		 * Reference existing named region or allocate new one.  There
640 		 * should not be a race here against swp_pager_meta_build()
641 		 * as called from vm_page_remove() in regards to the lookup
642 		 * of the handle.
643 		 */
644 		sx_xlock(&sw_alloc_sx);
645 		object = vm_pager_object_lookup(NOBJLIST(handle), handle);
646 		if (object == NULL) {
647 			object = swap_pager_alloc_init(handle, cred, size,
648 			    offset);
649 			if (object != NULL) {
650 				TAILQ_INSERT_TAIL(NOBJLIST(object->handle),
651 				    object, pager_object_list);
652 			}
653 		}
654 		sx_xunlock(&sw_alloc_sx);
655 	} else {
656 		object = swap_pager_alloc_init(handle, cred, size, offset);
657 	}
658 	return (object);
659 }
660 
661 /*
662  * SWAP_PAGER_DEALLOC() -	remove swap metadata from object
663  *
664  *	The swap backing for the object is destroyed.  The code is
665  *	designed such that we can reinstantiate it later, but this
666  *	routine is typically called only when the entire object is
667  *	about to be destroyed.
668  *
669  *	The object must be locked.
670  */
671 static void
672 swap_pager_dealloc(vm_object_t object)
673 {
674 
675 	VM_OBJECT_ASSERT_WLOCKED(object);
676 	KASSERT((object->flags & OBJ_DEAD) != 0, ("dealloc of reachable obj"));
677 
678 	/*
679 	 * Remove from list right away so lookups will fail if we block for
680 	 * pageout completion.
681 	 */
682 	if (object->handle != NULL) {
683 		VM_OBJECT_WUNLOCK(object);
684 		sx_xlock(&sw_alloc_sx);
685 		TAILQ_REMOVE(NOBJLIST(object->handle), object,
686 		    pager_object_list);
687 		sx_xunlock(&sw_alloc_sx);
688 		VM_OBJECT_WLOCK(object);
689 	}
690 
691 	vm_object_pip_wait(object, "swpdea");
692 
693 	/*
694 	 * Free all remaining metadata.  We only bother to free it from
695 	 * the swap meta data.  We do not attempt to free swapblk's still
696 	 * associated with vm_page_t's for this object.  We do not care
697 	 * if paging is still in progress on some objects.
698 	 */
699 	swp_pager_meta_free_all(object);
700 	object->handle = NULL;
701 	object->type = OBJT_DEAD;
702 }
703 
704 /************************************************************************
705  *			SWAP PAGER BITMAP ROUTINES			*
706  ************************************************************************/
707 
708 /*
709  * SWP_PAGER_GETSWAPSPACE() -	allocate raw swap space
710  *
711  *	Allocate swap for the requested number of pages.  The starting
712  *	swap block number (a page index) is returned or SWAPBLK_NONE
713  *	if the allocation failed.
714  *
715  *	Also has the side effect of advising that somebody made a mistake
716  *	when they configured swap and didn't configure enough.
717  *
718  *	This routine may not sleep.
719  *
720  *	We allocate in round-robin fashion from the configured devices.
721  */
722 static daddr_t
723 swp_pager_getswapspace(int npages)
724 {
725 	daddr_t blk;
726 	struct swdevt *sp;
727 	int i;
728 
729 	blk = SWAPBLK_NONE;
730 	mtx_lock(&sw_dev_mtx);
731 	sp = swdevhd;
732 	for (i = 0; i < nswapdev; i++) {
733 		if (sp == NULL)
734 			sp = TAILQ_FIRST(&swtailq);
735 		if (!(sp->sw_flags & SW_CLOSING)) {
736 			blk = blist_alloc(sp->sw_blist, npages);
737 			if (blk != SWAPBLK_NONE) {
738 				blk += sp->sw_first;
739 				sp->sw_used += npages;
740 				swap_pager_avail -= npages;
741 				swp_sizecheck();
742 				swdevhd = TAILQ_NEXT(sp, sw_list);
743 				goto done;
744 			}
745 		}
746 		sp = TAILQ_NEXT(sp, sw_list);
747 	}
748 	if (swap_pager_full != 2) {
749 		printf("swap_pager_getswapspace(%d): failed\n", npages);
750 		swap_pager_full = 2;
751 		swap_pager_almost_full = 1;
752 	}
753 	swdevhd = NULL;
754 done:
755 	mtx_unlock(&sw_dev_mtx);
756 	return (blk);
757 }
758 
759 static bool
760 swp_pager_isondev(daddr_t blk, struct swdevt *sp)
761 {
762 
763 	return (blk >= sp->sw_first && blk < sp->sw_end);
764 }
765 
766 static void
767 swp_pager_strategy(struct buf *bp)
768 {
769 	struct swdevt *sp;
770 
771 	mtx_lock(&sw_dev_mtx);
772 	TAILQ_FOREACH(sp, &swtailq, sw_list) {
773 		if (swp_pager_isondev(bp->b_blkno, sp)) {
774 			mtx_unlock(&sw_dev_mtx);
775 			if ((sp->sw_flags & SW_UNMAPPED) != 0 &&
776 			    unmapped_buf_allowed) {
777 				bp->b_data = unmapped_buf;
778 				bp->b_offset = 0;
779 			} else {
780 				pmap_qenter((vm_offset_t)bp->b_data,
781 				    &bp->b_pages[0], bp->b_bcount / PAGE_SIZE);
782 			}
783 			sp->sw_strategy(bp, sp);
784 			return;
785 		}
786 	}
787 	panic("Swapdev not found");
788 }
789 
790 
791 /*
792  * SWP_PAGER_FREESWAPSPACE() -	free raw swap space
793  *
794  *	This routine returns the specified swap blocks back to the bitmap.
795  *
796  *	This routine may not sleep.
797  */
798 static void
799 swp_pager_freeswapspace(daddr_t blk, daddr_t npages)
800 {
801 	struct swdevt *sp;
802 
803 	if (npages == 0)
804 		return;
805 	mtx_lock(&sw_dev_mtx);
806 	TAILQ_FOREACH(sp, &swtailq, sw_list) {
807 		if (swp_pager_isondev(blk, sp)) {
808 			sp->sw_used -= npages;
809 			/*
810 			 * If we are attempting to stop swapping on
811 			 * this device, we don't want to mark any
812 			 * blocks free lest they be reused.
813 			 */
814 			if ((sp->sw_flags & SW_CLOSING) == 0) {
815 				blist_free(sp->sw_blist, blk - sp->sw_first,
816 				    npages);
817 				swap_pager_avail += npages;
818 				swp_sizecheck();
819 			}
820 			mtx_unlock(&sw_dev_mtx);
821 			return;
822 		}
823 	}
824 	panic("Swapdev not found");
825 }
826 
827 /*
828  * SYSCTL_SWAP_FRAGMENTATION() -	produce raw swap space stats
829  */
830 static int
831 sysctl_swap_fragmentation(SYSCTL_HANDLER_ARGS)
832 {
833 	struct sbuf sbuf;
834 	struct swdevt *sp;
835 	const char *devname;
836 	int error;
837 
838 	error = sysctl_wire_old_buffer(req, 0);
839 	if (error != 0)
840 		return (error);
841 	sbuf_new_for_sysctl(&sbuf, NULL, 128, req);
842 	mtx_lock(&sw_dev_mtx);
843 	TAILQ_FOREACH(sp, &swtailq, sw_list) {
844 		if (vn_isdisk(sp->sw_vp, NULL))
845 			devname = devtoname(sp->sw_vp->v_rdev);
846 		else
847 			devname = "[file]";
848 		sbuf_printf(&sbuf, "\nFree space on device %s:\n", devname);
849 		blist_stats(sp->sw_blist, &sbuf);
850 	}
851 	mtx_unlock(&sw_dev_mtx);
852 	error = sbuf_finish(&sbuf);
853 	sbuf_delete(&sbuf);
854 	return (error);
855 }
856 
857 /*
858  * SWAP_PAGER_FREESPACE() -	frees swap blocks associated with a page
859  *				range within an object.
860  *
861  *	This is a globally accessible routine.
862  *
863  *	This routine removes swapblk assignments from swap metadata.
864  *
865  *	The external callers of this routine typically have already destroyed
866  *	or renamed vm_page_t's associated with this range in the object so
867  *	we should be ok.
868  *
869  *	The object must be locked.
870  */
871 void
872 swap_pager_freespace(vm_object_t object, vm_pindex_t start, vm_size_t size)
873 {
874 
875 	swp_pager_meta_free(object, start, size);
876 }
877 
878 /*
879  * SWAP_PAGER_RESERVE() - reserve swap blocks in object
880  *
881  *	Assigns swap blocks to the specified range within the object.  The
882  *	swap blocks are not zeroed.  Any previous swap assignment is destroyed.
883  *
884  *	Returns 0 on success, -1 on failure.
885  */
886 int
887 swap_pager_reserve(vm_object_t object, vm_pindex_t start, vm_size_t size)
888 {
889 	int n = 0;
890 	daddr_t blk = SWAPBLK_NONE;
891 	vm_pindex_t beg = start;	/* save start index */
892 	daddr_t addr, n_free, s_free;
893 
894 	swp_pager_init_freerange(&s_free, &n_free);
895 	VM_OBJECT_WLOCK(object);
896 	while (size) {
897 		if (n == 0) {
898 			n = BLIST_MAX_ALLOC;
899 			while ((blk = swp_pager_getswapspace(n)) == SWAPBLK_NONE) {
900 				n >>= 1;
901 				if (n == 0) {
902 					swp_pager_meta_free(object, beg, start - beg);
903 					VM_OBJECT_WUNLOCK(object);
904 					return (-1);
905 				}
906 			}
907 		}
908 		addr = swp_pager_meta_build(object, start, blk);
909 		if (addr != SWAPBLK_NONE)
910 			swp_pager_update_freerange(&s_free, &n_free, addr);
911 		--size;
912 		++start;
913 		++blk;
914 		--n;
915 	}
916 	swp_pager_freeswapspace(s_free, n_free);
917 	swp_pager_meta_free(object, start, n);
918 	VM_OBJECT_WUNLOCK(object);
919 	return (0);
920 }
921 
922 /*
923  * SWAP_PAGER_COPY() -  copy blocks from source pager to destination pager
924  *			and destroy the source.
925  *
926  *	Copy any valid swapblks from the source to the destination.  In
927  *	cases where both the source and destination have a valid swapblk,
928  *	we keep the destination's.
929  *
930  *	This routine is allowed to sleep.  It may sleep allocating metadata
931  *	indirectly through swp_pager_meta_build() or if paging is still in
932  *	progress on the source.
933  *
934  *	The source object contains no vm_page_t's (which is just as well)
935  *
936  *	The source object is of type OBJT_SWAP.
937  *
938  *	The source and destination objects must be locked.
939  *	Both object locks may temporarily be released.
940  */
941 void
942 swap_pager_copy(vm_object_t srcobject, vm_object_t dstobject,
943     vm_pindex_t offset, int destroysource)
944 {
945 	vm_pindex_t i;
946 	daddr_t dstaddr, n_free, s_free, srcaddr;
947 
948 	VM_OBJECT_ASSERT_WLOCKED(srcobject);
949 	VM_OBJECT_ASSERT_WLOCKED(dstobject);
950 
951 	/*
952 	 * If destroysource is set, we remove the source object from the
953 	 * swap_pager internal queue now.
954 	 */
955 	if (destroysource && srcobject->handle != NULL) {
956 		vm_object_pip_add(srcobject, 1);
957 		VM_OBJECT_WUNLOCK(srcobject);
958 		vm_object_pip_add(dstobject, 1);
959 		VM_OBJECT_WUNLOCK(dstobject);
960 		sx_xlock(&sw_alloc_sx);
961 		TAILQ_REMOVE(NOBJLIST(srcobject->handle), srcobject,
962 		    pager_object_list);
963 		sx_xunlock(&sw_alloc_sx);
964 		VM_OBJECT_WLOCK(dstobject);
965 		vm_object_pip_wakeup(dstobject);
966 		VM_OBJECT_WLOCK(srcobject);
967 		vm_object_pip_wakeup(srcobject);
968 	}
969 
970 	/*
971 	 * Transfer source to destination.
972 	 */
973 	swp_pager_init_freerange(&s_free, &n_free);
974 	for (i = 0; i < dstobject->size; ++i) {
975 		srcaddr = swp_pager_meta_ctl(srcobject, i + offset, SWM_POP);
976 		if (srcaddr == SWAPBLK_NONE)
977 			continue;
978 		dstaddr = swp_pager_meta_ctl(dstobject, i, 0);
979 		if (dstaddr != SWAPBLK_NONE) {
980 			/*
981 			 * Destination has valid swapblk or it is represented
982 			 * by a resident page.  We destroy the source block.
983 			 */
984 			swp_pager_update_freerange(&s_free, &n_free, srcaddr);
985 			continue;
986 		}
987 
988 		/*
989 		 * Destination has no swapblk and is not resident,
990 		 * copy source.
991 		 *
992 		 * swp_pager_meta_build() can sleep.
993 		 */
994 		vm_object_pip_add(srcobject, 1);
995 		VM_OBJECT_WUNLOCK(srcobject);
996 		vm_object_pip_add(dstobject, 1);
997 		dstaddr = swp_pager_meta_build(dstobject, i, srcaddr);
998 		KASSERT(dstaddr == SWAPBLK_NONE,
999 		    ("Unexpected destination swapblk"));
1000 		vm_object_pip_wakeup(dstobject);
1001 		VM_OBJECT_WLOCK(srcobject);
1002 		vm_object_pip_wakeup(srcobject);
1003 	}
1004 	swp_pager_freeswapspace(s_free, n_free);
1005 
1006 	/*
1007 	 * Free left over swap blocks in source.
1008 	 *
1009 	 * We have to revert the type to OBJT_DEFAULT so we do not accidentally
1010 	 * double-remove the object from the swap queues.
1011 	 */
1012 	if (destroysource) {
1013 		swp_pager_meta_free_all(srcobject);
1014 		/*
1015 		 * Reverting the type is not necessary, the caller is going
1016 		 * to destroy srcobject directly, but I'm doing it here
1017 		 * for consistency since we've removed the object from its
1018 		 * queues.
1019 		 */
1020 		srcobject->type = OBJT_DEFAULT;
1021 	}
1022 }
1023 
1024 /*
1025  * SWAP_PAGER_HASPAGE() -	determine if we have good backing store for
1026  *				the requested page.
1027  *
1028  *	We determine whether good backing store exists for the requested
1029  *	page and return TRUE if it does, FALSE if it doesn't.
1030  *
1031  *	If TRUE, we also try to determine how much valid, contiguous backing
1032  *	store exists before and after the requested page.
1033  */
1034 static boolean_t
1035 swap_pager_haspage(vm_object_t object, vm_pindex_t pindex, int *before,
1036     int *after)
1037 {
1038 	daddr_t blk, blk0;
1039 	int i;
1040 
1041 	VM_OBJECT_ASSERT_LOCKED(object);
1042 
1043 	/*
1044 	 * do we have good backing store at the requested index ?
1045 	 */
1046 	blk0 = swp_pager_meta_ctl(object, pindex, 0);
1047 	if (blk0 == SWAPBLK_NONE) {
1048 		if (before)
1049 			*before = 0;
1050 		if (after)
1051 			*after = 0;
1052 		return (FALSE);
1053 	}
1054 
1055 	/*
1056 	 * find backwards-looking contiguous good backing store
1057 	 */
1058 	if (before != NULL) {
1059 		for (i = 1; i < SWB_NPAGES; i++) {
1060 			if (i > pindex)
1061 				break;
1062 			blk = swp_pager_meta_ctl(object, pindex - i, 0);
1063 			if (blk != blk0 - i)
1064 				break;
1065 		}
1066 		*before = i - 1;
1067 	}
1068 
1069 	/*
1070 	 * find forward-looking contiguous good backing store
1071 	 */
1072 	if (after != NULL) {
1073 		for (i = 1; i < SWB_NPAGES; i++) {
1074 			blk = swp_pager_meta_ctl(object, pindex + i, 0);
1075 			if (blk != blk0 + i)
1076 				break;
1077 		}
1078 		*after = i - 1;
1079 	}
1080 	return (TRUE);
1081 }
1082 
1083 /*
1084  * SWAP_PAGER_PAGE_UNSWAPPED() - remove swap backing store related to page
1085  *
1086  *	This removes any associated swap backing store, whether valid or
1087  *	not, from the page.
1088  *
1089  *	This routine is typically called when a page is made dirty, at
1090  *	which point any associated swap can be freed.  MADV_FREE also
1091  *	calls us in a special-case situation
1092  *
1093  *	NOTE!!!  If the page is clean and the swap was valid, the caller
1094  *	should make the page dirty before calling this routine.  This routine
1095  *	does NOT change the m->dirty status of the page.  Also: MADV_FREE
1096  *	depends on it.
1097  *
1098  *	This routine may not sleep.
1099  *
1100  *	The object containing the page must be locked.
1101  */
1102 static void
1103 swap_pager_unswapped(vm_page_t m)
1104 {
1105 	daddr_t srcaddr;
1106 
1107 	srcaddr = swp_pager_meta_ctl(m->object, m->pindex, SWM_POP);
1108 	if (srcaddr != SWAPBLK_NONE)
1109 		swp_pager_freeswapspace(srcaddr, 1);
1110 }
1111 
1112 /*
1113  * swap_pager_getpages() - bring pages in from swap
1114  *
1115  *	Attempt to page in the pages in array "ma" of length "count".  The
1116  *	caller may optionally specify that additional pages preceding and
1117  *	succeeding the specified range be paged in.  The number of such pages
1118  *	is returned in the "rbehind" and "rahead" parameters, and they will
1119  *	be in the inactive queue upon return.
1120  *
1121  *	The pages in "ma" must be busied and will remain busied upon return.
1122  */
1123 static int
1124 swap_pager_getpages(vm_object_t object, vm_page_t *ma, int count, int *rbehind,
1125     int *rahead)
1126 {
1127 	struct buf *bp;
1128 	vm_page_t bm, mpred, msucc, p;
1129 	vm_pindex_t pindex;
1130 	daddr_t blk;
1131 	int i, maxahead, maxbehind, reqcount;
1132 
1133 	reqcount = count;
1134 
1135 	/*
1136 	 * Determine the final number of read-behind pages and
1137 	 * allocate them BEFORE releasing the object lock.  Otherwise,
1138 	 * there can be a problematic race with vm_object_split().
1139 	 * Specifically, vm_object_split() might first transfer pages
1140 	 * that precede ma[0] in the current object to a new object,
1141 	 * and then this function incorrectly recreates those pages as
1142 	 * read-behind pages in the current object.
1143 	 */
1144 	if (!swap_pager_haspage(object, ma[0]->pindex, &maxbehind, &maxahead))
1145 		return (VM_PAGER_FAIL);
1146 
1147 	/*
1148 	 * Clip the readahead and readbehind ranges to exclude resident pages.
1149 	 */
1150 	if (rahead != NULL) {
1151 		KASSERT(reqcount - 1 <= maxahead,
1152 		    ("page count %d extends beyond swap block", reqcount));
1153 		*rahead = imin(*rahead, maxahead - (reqcount - 1));
1154 		pindex = ma[reqcount - 1]->pindex;
1155 		msucc = TAILQ_NEXT(ma[reqcount - 1], listq);
1156 		if (msucc != NULL && msucc->pindex - pindex - 1 < *rahead)
1157 			*rahead = msucc->pindex - pindex - 1;
1158 	}
1159 	if (rbehind != NULL) {
1160 		*rbehind = imin(*rbehind, maxbehind);
1161 		pindex = ma[0]->pindex;
1162 		mpred = TAILQ_PREV(ma[0], pglist, listq);
1163 		if (mpred != NULL && pindex - mpred->pindex - 1 < *rbehind)
1164 			*rbehind = pindex - mpred->pindex - 1;
1165 	}
1166 
1167 	bm = ma[0];
1168 	for (i = 0; i < count; i++)
1169 		ma[i]->oflags |= VPO_SWAPINPROG;
1170 
1171 	/*
1172 	 * Allocate readahead and readbehind pages.
1173 	 */
1174 	if (rbehind != NULL) {
1175 		for (i = 1; i <= *rbehind; i++) {
1176 			p = vm_page_alloc(object, ma[0]->pindex - i,
1177 			    VM_ALLOC_NORMAL);
1178 			if (p == NULL)
1179 				break;
1180 			p->oflags |= VPO_SWAPINPROG;
1181 			bm = p;
1182 		}
1183 		*rbehind = i - 1;
1184 	}
1185 	if (rahead != NULL) {
1186 		for (i = 0; i < *rahead; i++) {
1187 			p = vm_page_alloc(object,
1188 			    ma[reqcount - 1]->pindex + i + 1, VM_ALLOC_NORMAL);
1189 			if (p == NULL)
1190 				break;
1191 			p->oflags |= VPO_SWAPINPROG;
1192 		}
1193 		*rahead = i;
1194 	}
1195 	if (rbehind != NULL)
1196 		count += *rbehind;
1197 	if (rahead != NULL)
1198 		count += *rahead;
1199 
1200 	vm_object_pip_add(object, count);
1201 
1202 	pindex = bm->pindex;
1203 	blk = swp_pager_meta_ctl(object, pindex, 0);
1204 	KASSERT(blk != SWAPBLK_NONE,
1205 	    ("no swap blocking containing %p(%jx)", object, (uintmax_t)pindex));
1206 
1207 	VM_OBJECT_WUNLOCK(object);
1208 	bp = uma_zalloc(swrbuf_zone, M_WAITOK);
1209 	/* Pages cannot leave the object while busy. */
1210 	for (i = 0, p = bm; i < count; i++, p = TAILQ_NEXT(p, listq)) {
1211 		MPASS(p->pindex == bm->pindex + i);
1212 		bp->b_pages[i] = p;
1213 	}
1214 
1215 	bp->b_flags |= B_PAGING;
1216 	bp->b_iocmd = BIO_READ;
1217 	bp->b_iodone = swp_pager_async_iodone;
1218 	bp->b_rcred = crhold(thread0.td_ucred);
1219 	bp->b_wcred = crhold(thread0.td_ucred);
1220 	bp->b_blkno = blk;
1221 	bp->b_bcount = PAGE_SIZE * count;
1222 	bp->b_bufsize = PAGE_SIZE * count;
1223 	bp->b_npages = count;
1224 	bp->b_pgbefore = rbehind != NULL ? *rbehind : 0;
1225 	bp->b_pgafter = rahead != NULL ? *rahead : 0;
1226 
1227 	VM_CNT_INC(v_swapin);
1228 	VM_CNT_ADD(v_swappgsin, count);
1229 
1230 	/*
1231 	 * perform the I/O.  NOTE!!!  bp cannot be considered valid after
1232 	 * this point because we automatically release it on completion.
1233 	 * Instead, we look at the one page we are interested in which we
1234 	 * still hold a lock on even through the I/O completion.
1235 	 *
1236 	 * The other pages in our ma[] array are also released on completion,
1237 	 * so we cannot assume they are valid anymore either.
1238 	 *
1239 	 * NOTE: b_blkno is destroyed by the call to swapdev_strategy
1240 	 */
1241 	BUF_KERNPROC(bp);
1242 	swp_pager_strategy(bp);
1243 
1244 	/*
1245 	 * Wait for the pages we want to complete.  VPO_SWAPINPROG is always
1246 	 * cleared on completion.  If an I/O error occurs, SWAPBLK_NONE
1247 	 * is set in the metadata for each page in the request.
1248 	 */
1249 	VM_OBJECT_WLOCK(object);
1250 	while ((ma[0]->oflags & VPO_SWAPINPROG) != 0) {
1251 		ma[0]->oflags |= VPO_SWAPSLEEP;
1252 		VM_CNT_INC(v_intrans);
1253 		if (VM_OBJECT_SLEEP(object, &object->paging_in_progress, PSWP,
1254 		    "swread", hz * 20)) {
1255 			printf(
1256 "swap_pager: indefinite wait buffer: bufobj: %p, blkno: %jd, size: %ld\n",
1257 			    bp->b_bufobj, (intmax_t)bp->b_blkno, bp->b_bcount);
1258 		}
1259 	}
1260 
1261 	/*
1262 	 * If we had an unrecoverable read error pages will not be valid.
1263 	 */
1264 	for (i = 0; i < reqcount; i++)
1265 		if (ma[i]->valid != VM_PAGE_BITS_ALL)
1266 			return (VM_PAGER_ERROR);
1267 
1268 	return (VM_PAGER_OK);
1269 
1270 	/*
1271 	 * A final note: in a low swap situation, we cannot deallocate swap
1272 	 * and mark a page dirty here because the caller is likely to mark
1273 	 * the page clean when we return, causing the page to possibly revert
1274 	 * to all-zero's later.
1275 	 */
1276 }
1277 
1278 /*
1279  * 	swap_pager_getpages_async():
1280  *
1281  *	Right now this is emulation of asynchronous operation on top of
1282  *	swap_pager_getpages().
1283  */
1284 static int
1285 swap_pager_getpages_async(vm_object_t object, vm_page_t *ma, int count,
1286     int *rbehind, int *rahead, pgo_getpages_iodone_t iodone, void *arg)
1287 {
1288 	int r, error;
1289 
1290 	r = swap_pager_getpages(object, ma, count, rbehind, rahead);
1291 	VM_OBJECT_WUNLOCK(object);
1292 	switch (r) {
1293 	case VM_PAGER_OK:
1294 		error = 0;
1295 		break;
1296 	case VM_PAGER_ERROR:
1297 		error = EIO;
1298 		break;
1299 	case VM_PAGER_FAIL:
1300 		error = EINVAL;
1301 		break;
1302 	default:
1303 		panic("unhandled swap_pager_getpages() error %d", r);
1304 	}
1305 	(iodone)(arg, ma, count, error);
1306 	VM_OBJECT_WLOCK(object);
1307 
1308 	return (r);
1309 }
1310 
1311 /*
1312  *	swap_pager_putpages:
1313  *
1314  *	Assign swap (if necessary) and initiate I/O on the specified pages.
1315  *
1316  *	We support both OBJT_DEFAULT and OBJT_SWAP objects.  DEFAULT objects
1317  *	are automatically converted to SWAP objects.
1318  *
1319  *	In a low memory situation we may block in VOP_STRATEGY(), but the new
1320  *	vm_page reservation system coupled with properly written VFS devices
1321  *	should ensure that no low-memory deadlock occurs.  This is an area
1322  *	which needs work.
1323  *
1324  *	The parent has N vm_object_pip_add() references prior to
1325  *	calling us and will remove references for rtvals[] that are
1326  *	not set to VM_PAGER_PEND.  We need to remove the rest on I/O
1327  *	completion.
1328  *
1329  *	The parent has soft-busy'd the pages it passes us and will unbusy
1330  *	those whos rtvals[] entry is not set to VM_PAGER_PEND on return.
1331  *	We need to unbusy the rest on I/O completion.
1332  */
1333 static void
1334 swap_pager_putpages(vm_object_t object, vm_page_t *ma, int count,
1335     int flags, int *rtvals)
1336 {
1337 	int i, n;
1338 	boolean_t sync;
1339 	daddr_t addr, n_free, s_free;
1340 
1341 	swp_pager_init_freerange(&s_free, &n_free);
1342 	if (count && ma[0]->object != object) {
1343 		panic("swap_pager_putpages: object mismatch %p/%p",
1344 		    object,
1345 		    ma[0]->object
1346 		);
1347 	}
1348 
1349 	/*
1350 	 * Step 1
1351 	 *
1352 	 * Turn object into OBJT_SWAP
1353 	 * check for bogus sysops
1354 	 * force sync if not pageout process
1355 	 */
1356 	if (object->type != OBJT_SWAP) {
1357 		addr = swp_pager_meta_build(object, 0, SWAPBLK_NONE);
1358 		KASSERT(addr == SWAPBLK_NONE,
1359 		    ("unexpected object swap block"));
1360 	}
1361 	VM_OBJECT_WUNLOCK(object);
1362 
1363 	n = 0;
1364 	if (curproc != pageproc)
1365 		sync = TRUE;
1366 	else
1367 		sync = (flags & VM_PAGER_PUT_SYNC) != 0;
1368 
1369 	/*
1370 	 * Step 2
1371 	 *
1372 	 * Assign swap blocks and issue I/O.  We reallocate swap on the fly.
1373 	 * The page is left dirty until the pageout operation completes
1374 	 * successfully.
1375 	 */
1376 	for (i = 0; i < count; i += n) {
1377 		int j;
1378 		struct buf *bp;
1379 		daddr_t blk;
1380 
1381 		/*
1382 		 * Maximum I/O size is limited by a number of factors.
1383 		 */
1384 		n = min(BLIST_MAX_ALLOC, count - i);
1385 		n = min(n, nsw_cluster_max);
1386 
1387 		/*
1388 		 * Get biggest block of swap we can.  If we fail, fall
1389 		 * back and try to allocate a smaller block.  Don't go
1390 		 * overboard trying to allocate space if it would overly
1391 		 * fragment swap.
1392 		 */
1393 		while (
1394 		    (blk = swp_pager_getswapspace(n)) == SWAPBLK_NONE &&
1395 		    n > 4
1396 		) {
1397 			n >>= 1;
1398 		}
1399 		if (blk == SWAPBLK_NONE) {
1400 			for (j = 0; j < n; ++j)
1401 				rtvals[i+j] = VM_PAGER_FAIL;
1402 			continue;
1403 		}
1404 
1405 		/*
1406 		 * All I/O parameters have been satisfied, build the I/O
1407 		 * request and assign the swap space.
1408 		 */
1409 		if (sync != TRUE) {
1410 			mtx_lock(&swbuf_mtx);
1411 			while (nsw_wcount_async == 0)
1412 				msleep(&nsw_wcount_async, &swbuf_mtx, PVM,
1413 				    "swbufa", 0);
1414 			nsw_wcount_async--;
1415 			mtx_unlock(&swbuf_mtx);
1416 		}
1417 		bp = uma_zalloc(swwbuf_zone, M_WAITOK);
1418 		if (sync != TRUE)
1419 			bp->b_flags = B_ASYNC;
1420 		bp->b_flags |= B_PAGING;
1421 		bp->b_iocmd = BIO_WRITE;
1422 
1423 		bp->b_rcred = crhold(thread0.td_ucred);
1424 		bp->b_wcred = crhold(thread0.td_ucred);
1425 		bp->b_bcount = PAGE_SIZE * n;
1426 		bp->b_bufsize = PAGE_SIZE * n;
1427 		bp->b_blkno = blk;
1428 
1429 		VM_OBJECT_WLOCK(object);
1430 		for (j = 0; j < n; ++j) {
1431 			vm_page_t mreq = ma[i+j];
1432 
1433 			addr = swp_pager_meta_build(mreq->object, mreq->pindex,
1434 			    blk + j);
1435 			if (addr != SWAPBLK_NONE)
1436 				swp_pager_update_freerange(&s_free, &n_free,
1437 				    addr);
1438 			MPASS(mreq->dirty == VM_PAGE_BITS_ALL);
1439 			mreq->oflags |= VPO_SWAPINPROG;
1440 			bp->b_pages[j] = mreq;
1441 		}
1442 		VM_OBJECT_WUNLOCK(object);
1443 		bp->b_npages = n;
1444 		/*
1445 		 * Must set dirty range for NFS to work.
1446 		 */
1447 		bp->b_dirtyoff = 0;
1448 		bp->b_dirtyend = bp->b_bcount;
1449 
1450 		VM_CNT_INC(v_swapout);
1451 		VM_CNT_ADD(v_swappgsout, bp->b_npages);
1452 
1453 		/*
1454 		 * We unconditionally set rtvals[] to VM_PAGER_PEND so that we
1455 		 * can call the async completion routine at the end of a
1456 		 * synchronous I/O operation.  Otherwise, our caller would
1457 		 * perform duplicate unbusy and wakeup operations on the page
1458 		 * and object, respectively.
1459 		 */
1460 		for (j = 0; j < n; j++)
1461 			rtvals[i + j] = VM_PAGER_PEND;
1462 
1463 		/*
1464 		 * asynchronous
1465 		 *
1466 		 * NOTE: b_blkno is destroyed by the call to swapdev_strategy
1467 		 */
1468 		if (sync == FALSE) {
1469 			bp->b_iodone = swp_pager_async_iodone;
1470 			BUF_KERNPROC(bp);
1471 			swp_pager_strategy(bp);
1472 			continue;
1473 		}
1474 
1475 		/*
1476 		 * synchronous
1477 		 *
1478 		 * NOTE: b_blkno is destroyed by the call to swapdev_strategy
1479 		 */
1480 		bp->b_iodone = bdone;
1481 		swp_pager_strategy(bp);
1482 
1483 		/*
1484 		 * Wait for the sync I/O to complete.
1485 		 */
1486 		bwait(bp, PVM, "swwrt");
1487 
1488 		/*
1489 		 * Now that we are through with the bp, we can call the
1490 		 * normal async completion, which frees everything up.
1491 		 */
1492 		swp_pager_async_iodone(bp);
1493 	}
1494 	VM_OBJECT_WLOCK(object);
1495 	swp_pager_freeswapspace(s_free, n_free);
1496 }
1497 
1498 /*
1499  *	swp_pager_async_iodone:
1500  *
1501  *	Completion routine for asynchronous reads and writes from/to swap.
1502  *	Also called manually by synchronous code to finish up a bp.
1503  *
1504  *	This routine may not sleep.
1505  */
1506 static void
1507 swp_pager_async_iodone(struct buf *bp)
1508 {
1509 	int i;
1510 	vm_object_t object = NULL;
1511 
1512 	/*
1513 	 * report error
1514 	 */
1515 	if (bp->b_ioflags & BIO_ERROR) {
1516 		printf(
1517 		    "swap_pager: I/O error - %s failed; blkno %ld,"
1518 			"size %ld, error %d\n",
1519 		    ((bp->b_iocmd == BIO_READ) ? "pagein" : "pageout"),
1520 		    (long)bp->b_blkno,
1521 		    (long)bp->b_bcount,
1522 		    bp->b_error
1523 		);
1524 	}
1525 
1526 	/*
1527 	 * remove the mapping for kernel virtual
1528 	 */
1529 	if (buf_mapped(bp))
1530 		pmap_qremove((vm_offset_t)bp->b_data, bp->b_npages);
1531 	else
1532 		bp->b_data = bp->b_kvabase;
1533 
1534 	if (bp->b_npages) {
1535 		object = bp->b_pages[0]->object;
1536 		VM_OBJECT_WLOCK(object);
1537 	}
1538 
1539 	/*
1540 	 * cleanup pages.  If an error occurs writing to swap, we are in
1541 	 * very serious trouble.  If it happens to be a disk error, though,
1542 	 * we may be able to recover by reassigning the swap later on.  So
1543 	 * in this case we remove the m->swapblk assignment for the page
1544 	 * but do not free it in the rlist.  The errornous block(s) are thus
1545 	 * never reallocated as swap.  Redirty the page and continue.
1546 	 */
1547 	for (i = 0; i < bp->b_npages; ++i) {
1548 		vm_page_t m = bp->b_pages[i];
1549 
1550 		m->oflags &= ~VPO_SWAPINPROG;
1551 		if (m->oflags & VPO_SWAPSLEEP) {
1552 			m->oflags &= ~VPO_SWAPSLEEP;
1553 			wakeup(&object->paging_in_progress);
1554 		}
1555 
1556 		if (bp->b_ioflags & BIO_ERROR) {
1557 			/*
1558 			 * If an error occurs I'd love to throw the swapblk
1559 			 * away without freeing it back to swapspace, so it
1560 			 * can never be used again.  But I can't from an
1561 			 * interrupt.
1562 			 */
1563 			if (bp->b_iocmd == BIO_READ) {
1564 				/*
1565 				 * NOTE: for reads, m->dirty will probably
1566 				 * be overridden by the original caller of
1567 				 * getpages so don't play cute tricks here.
1568 				 */
1569 				m->valid = 0;
1570 			} else {
1571 				/*
1572 				 * If a write error occurs, reactivate page
1573 				 * so it doesn't clog the inactive list,
1574 				 * then finish the I/O.
1575 				 */
1576 				MPASS(m->dirty == VM_PAGE_BITS_ALL);
1577 				vm_page_lock(m);
1578 				vm_page_activate(m);
1579 				vm_page_unlock(m);
1580 				vm_page_sunbusy(m);
1581 			}
1582 		} else if (bp->b_iocmd == BIO_READ) {
1583 			/*
1584 			 * NOTE: for reads, m->dirty will probably be
1585 			 * overridden by the original caller of getpages so
1586 			 * we cannot set them in order to free the underlying
1587 			 * swap in a low-swap situation.  I don't think we'd
1588 			 * want to do that anyway, but it was an optimization
1589 			 * that existed in the old swapper for a time before
1590 			 * it got ripped out due to precisely this problem.
1591 			 */
1592 			KASSERT(!pmap_page_is_mapped(m),
1593 			    ("swp_pager_async_iodone: page %p is mapped", m));
1594 			KASSERT(m->dirty == 0,
1595 			    ("swp_pager_async_iodone: page %p is dirty", m));
1596 
1597 			m->valid = VM_PAGE_BITS_ALL;
1598 			if (i < bp->b_pgbefore ||
1599 			    i >= bp->b_npages - bp->b_pgafter)
1600 				vm_page_readahead_finish(m);
1601 		} else {
1602 			/*
1603 			 * For write success, clear the dirty
1604 			 * status, then finish the I/O ( which decrements the
1605 			 * busy count and possibly wakes waiter's up ).
1606 			 * A page is only written to swap after a period of
1607 			 * inactivity.  Therefore, we do not expect it to be
1608 			 * reused.
1609 			 */
1610 			KASSERT(!pmap_page_is_write_mapped(m),
1611 			    ("swp_pager_async_iodone: page %p is not write"
1612 			    " protected", m));
1613 			vm_page_undirty(m);
1614 			vm_page_lock(m);
1615 			vm_page_deactivate_noreuse(m);
1616 			vm_page_unlock(m);
1617 			vm_page_sunbusy(m);
1618 		}
1619 	}
1620 
1621 	/*
1622 	 * adjust pip.  NOTE: the original parent may still have its own
1623 	 * pip refs on the object.
1624 	 */
1625 	if (object != NULL) {
1626 		vm_object_pip_wakeupn(object, bp->b_npages);
1627 		VM_OBJECT_WUNLOCK(object);
1628 	}
1629 
1630 	/*
1631 	 * swapdev_strategy() manually sets b_vp and b_bufobj before calling
1632 	 * bstrategy(). Set them back to NULL now we're done with it, or we'll
1633 	 * trigger a KASSERT in relpbuf().
1634 	 */
1635 	if (bp->b_vp) {
1636 		    bp->b_vp = NULL;
1637 		    bp->b_bufobj = NULL;
1638 	}
1639 	/*
1640 	 * release the physical I/O buffer
1641 	 */
1642 	if (bp->b_flags & B_ASYNC) {
1643 		mtx_lock(&swbuf_mtx);
1644 		if (++nsw_wcount_async == 1)
1645 			wakeup(&nsw_wcount_async);
1646 		mtx_unlock(&swbuf_mtx);
1647 	}
1648 	uma_zfree((bp->b_iocmd == BIO_READ) ? swrbuf_zone : swwbuf_zone, bp);
1649 }
1650 
1651 int
1652 swap_pager_nswapdev(void)
1653 {
1654 
1655 	return (nswapdev);
1656 }
1657 
1658 /*
1659  * SWP_PAGER_FORCE_PAGEIN() - force a swap block to be paged in
1660  *
1661  *	This routine dissociates the page at the given index within an object
1662  *	from its backing store, paging it in if it does not reside in memory.
1663  *	If the page is paged in, it is marked dirty and placed in the laundry
1664  *	queue.  The page is marked dirty because it no longer has backing
1665  *	store.  It is placed in the laundry queue because it has not been
1666  *	accessed recently.  Otherwise, it would already reside in memory.
1667  *
1668  *	We also attempt to swap in all other pages in the swap block.
1669  *	However, we only guarantee that the one at the specified index is
1670  *	paged in.
1671  *
1672  *	XXX - The code to page the whole block in doesn't work, so we
1673  *	      revert to the one-by-one behavior for now.  Sigh.
1674  */
1675 static inline void
1676 swp_pager_force_pagein(vm_object_t object, vm_pindex_t pindex)
1677 {
1678 	vm_page_t m;
1679 
1680 	vm_object_pip_add(object, 1);
1681 	m = vm_page_grab(object, pindex, VM_ALLOC_NORMAL);
1682 	if (m->valid == VM_PAGE_BITS_ALL) {
1683 		vm_object_pip_wakeup(object);
1684 		vm_page_dirty(m);
1685 #ifdef INVARIANTS
1686 		vm_page_lock(m);
1687 		if (m->wire_count == 0 && m->queue == PQ_NONE)
1688 			panic("page %p is neither wired nor queued", m);
1689 		vm_page_unlock(m);
1690 #endif
1691 		vm_page_xunbusy(m);
1692 		vm_pager_page_unswapped(m);
1693 		return;
1694 	}
1695 
1696 	if (swap_pager_getpages(object, &m, 1, NULL, NULL) != VM_PAGER_OK)
1697 		panic("swap_pager_force_pagein: read from swap failed");/*XXX*/
1698 	vm_object_pip_wakeup(object);
1699 	vm_page_dirty(m);
1700 	vm_page_lock(m);
1701 	vm_page_launder(m);
1702 	vm_page_unlock(m);
1703 	vm_page_xunbusy(m);
1704 	vm_pager_page_unswapped(m);
1705 }
1706 
1707 /*
1708  *	swap_pager_swapoff:
1709  *
1710  *	Page in all of the pages that have been paged out to the
1711  *	given device.  The corresponding blocks in the bitmap must be
1712  *	marked as allocated and the device must be flagged SW_CLOSING.
1713  *	There may be no processes swapped out to the device.
1714  *
1715  *	This routine may block.
1716  */
1717 static void
1718 swap_pager_swapoff(struct swdevt *sp)
1719 {
1720 	struct swblk *sb;
1721 	vm_object_t object;
1722 	vm_pindex_t pi;
1723 	int i, retries;
1724 
1725 	sx_assert(&swdev_syscall_lock, SA_XLOCKED);
1726 
1727 	retries = 0;
1728 full_rescan:
1729 	mtx_lock(&vm_object_list_mtx);
1730 	TAILQ_FOREACH(object, &vm_object_list, object_list) {
1731 		if (object->type != OBJT_SWAP)
1732 			continue;
1733 		mtx_unlock(&vm_object_list_mtx);
1734 		/* Depends on type-stability. */
1735 		VM_OBJECT_WLOCK(object);
1736 
1737 		/*
1738 		 * Dead objects are eventually terminated on their own.
1739 		 */
1740 		if ((object->flags & OBJ_DEAD) != 0)
1741 			goto next_obj;
1742 
1743 		/*
1744 		 * Sync with fences placed after pctrie
1745 		 * initialization.  We must not access pctrie below
1746 		 * unless we checked that our object is swap and not
1747 		 * dead.
1748 		 */
1749 		atomic_thread_fence_acq();
1750 		if (object->type != OBJT_SWAP)
1751 			goto next_obj;
1752 
1753 		for (pi = 0; (sb = SWAP_PCTRIE_LOOKUP_GE(
1754 		    &object->un_pager.swp.swp_blks, pi)) != NULL; ) {
1755 			pi = sb->p + SWAP_META_PAGES;
1756 			for (i = 0; i < SWAP_META_PAGES; i++) {
1757 				if (sb->d[i] == SWAPBLK_NONE)
1758 					continue;
1759 				if (swp_pager_isondev(sb->d[i], sp))
1760 					swp_pager_force_pagein(object,
1761 					    sb->p + i);
1762 			}
1763 		}
1764 next_obj:
1765 		VM_OBJECT_WUNLOCK(object);
1766 		mtx_lock(&vm_object_list_mtx);
1767 	}
1768 	mtx_unlock(&vm_object_list_mtx);
1769 
1770 	if (sp->sw_used) {
1771 		/*
1772 		 * Objects may be locked or paging to the device being
1773 		 * removed, so we will miss their pages and need to
1774 		 * make another pass.  We have marked this device as
1775 		 * SW_CLOSING, so the activity should finish soon.
1776 		 */
1777 		retries++;
1778 		if (retries > 100) {
1779 			panic("swapoff: failed to locate %d swap blocks",
1780 			    sp->sw_used);
1781 		}
1782 		pause("swpoff", hz / 20);
1783 		goto full_rescan;
1784 	}
1785 	EVENTHANDLER_INVOKE(swapoff, sp);
1786 }
1787 
1788 /************************************************************************
1789  *				SWAP META DATA 				*
1790  ************************************************************************
1791  *
1792  *	These routines manipulate the swap metadata stored in the
1793  *	OBJT_SWAP object.
1794  *
1795  *	Swap metadata is implemented with a global hash and not directly
1796  *	linked into the object.  Instead the object simply contains
1797  *	appropriate tracking counters.
1798  */
1799 
1800 /*
1801  * SWP_PAGER_SWBLK_EMPTY() - is a range of blocks free?
1802  */
1803 static bool
1804 swp_pager_swblk_empty(struct swblk *sb, int start, int limit)
1805 {
1806 	int i;
1807 
1808 	MPASS(0 <= start && start <= limit && limit <= SWAP_META_PAGES);
1809 	for (i = start; i < limit; i++) {
1810 		if (sb->d[i] != SWAPBLK_NONE)
1811 			return (false);
1812 	}
1813 	return (true);
1814 }
1815 
1816 /*
1817  * SWP_PAGER_META_BUILD() -	add swap block to swap meta data for object
1818  *
1819  *	We first convert the object to a swap object if it is a default
1820  *	object.
1821  *
1822  *	The specified swapblk is added to the object's swap metadata.  If
1823  *	the swapblk is not valid, it is freed instead.  Any previously
1824  *	assigned swapblk is returned.
1825  */
1826 static daddr_t
1827 swp_pager_meta_build(vm_object_t object, vm_pindex_t pindex, daddr_t swapblk)
1828 {
1829 	static volatile int swblk_zone_exhausted, swpctrie_zone_exhausted;
1830 	struct swblk *sb, *sb1;
1831 	vm_pindex_t modpi, rdpi;
1832 	daddr_t prev_swapblk;
1833 	int error, i;
1834 
1835 	VM_OBJECT_ASSERT_WLOCKED(object);
1836 
1837 	/*
1838 	 * Convert default object to swap object if necessary
1839 	 */
1840 	if (object->type != OBJT_SWAP) {
1841 		pctrie_init(&object->un_pager.swp.swp_blks);
1842 
1843 		/*
1844 		 * Ensure that swap_pager_swapoff()'s iteration over
1845 		 * object_list does not see a garbage pctrie.
1846 		 */
1847 		atomic_thread_fence_rel();
1848 
1849 		object->type = OBJT_SWAP;
1850 		KASSERT(object->handle == NULL, ("default pager with handle"));
1851 	}
1852 
1853 	rdpi = rounddown(pindex, SWAP_META_PAGES);
1854 	sb = SWAP_PCTRIE_LOOKUP(&object->un_pager.swp.swp_blks, rdpi);
1855 	if (sb == NULL) {
1856 		if (swapblk == SWAPBLK_NONE)
1857 			return (SWAPBLK_NONE);
1858 		for (;;) {
1859 			sb = uma_zalloc(swblk_zone, M_NOWAIT | (curproc ==
1860 			    pageproc ? M_USE_RESERVE : 0));
1861 			if (sb != NULL) {
1862 				sb->p = rdpi;
1863 				for (i = 0; i < SWAP_META_PAGES; i++)
1864 					sb->d[i] = SWAPBLK_NONE;
1865 				if (atomic_cmpset_int(&swblk_zone_exhausted,
1866 				    1, 0))
1867 					printf("swblk zone ok\n");
1868 				break;
1869 			}
1870 			VM_OBJECT_WUNLOCK(object);
1871 			if (uma_zone_exhausted(swblk_zone)) {
1872 				if (atomic_cmpset_int(&swblk_zone_exhausted,
1873 				    0, 1))
1874 					printf("swap blk zone exhausted, "
1875 					    "increase kern.maxswzone\n");
1876 				vm_pageout_oom(VM_OOM_SWAPZ);
1877 				pause("swzonxb", 10);
1878 			} else
1879 				uma_zwait(swblk_zone);
1880 			VM_OBJECT_WLOCK(object);
1881 			sb = SWAP_PCTRIE_LOOKUP(&object->un_pager.swp.swp_blks,
1882 			    rdpi);
1883 			if (sb != NULL)
1884 				/*
1885 				 * Somebody swapped out a nearby page,
1886 				 * allocating swblk at the rdpi index,
1887 				 * while we dropped the object lock.
1888 				 */
1889 				goto allocated;
1890 		}
1891 		for (;;) {
1892 			error = SWAP_PCTRIE_INSERT(
1893 			    &object->un_pager.swp.swp_blks, sb);
1894 			if (error == 0) {
1895 				if (atomic_cmpset_int(&swpctrie_zone_exhausted,
1896 				    1, 0))
1897 					printf("swpctrie zone ok\n");
1898 				break;
1899 			}
1900 			VM_OBJECT_WUNLOCK(object);
1901 			if (uma_zone_exhausted(swpctrie_zone)) {
1902 				if (atomic_cmpset_int(&swpctrie_zone_exhausted,
1903 				    0, 1))
1904 					printf("swap pctrie zone exhausted, "
1905 					    "increase kern.maxswzone\n");
1906 				vm_pageout_oom(VM_OOM_SWAPZ);
1907 				pause("swzonxp", 10);
1908 			} else
1909 				uma_zwait(swpctrie_zone);
1910 			VM_OBJECT_WLOCK(object);
1911 			sb1 = SWAP_PCTRIE_LOOKUP(&object->un_pager.swp.swp_blks,
1912 			    rdpi);
1913 			if (sb1 != NULL) {
1914 				uma_zfree(swblk_zone, sb);
1915 				sb = sb1;
1916 				goto allocated;
1917 			}
1918 		}
1919 	}
1920 allocated:
1921 	MPASS(sb->p == rdpi);
1922 
1923 	modpi = pindex % SWAP_META_PAGES;
1924 	/* Return prior contents of metadata. */
1925 	prev_swapblk = sb->d[modpi];
1926 	/* Enter block into metadata. */
1927 	sb->d[modpi] = swapblk;
1928 
1929 	/*
1930 	 * Free the swblk if we end up with the empty page run.
1931 	 */
1932 	if (swapblk == SWAPBLK_NONE &&
1933 	    swp_pager_swblk_empty(sb, 0, SWAP_META_PAGES)) {
1934 		SWAP_PCTRIE_REMOVE(&object->un_pager.swp.swp_blks, rdpi);
1935 		uma_zfree(swblk_zone, sb);
1936 	}
1937 	return (prev_swapblk);
1938 }
1939 
1940 /*
1941  * SWP_PAGER_META_FREE() - free a range of blocks in the object's swap metadata
1942  *
1943  *	The requested range of blocks is freed, with any associated swap
1944  *	returned to the swap bitmap.
1945  *
1946  *	This routine will free swap metadata structures as they are cleaned
1947  *	out.  This routine does *NOT* operate on swap metadata associated
1948  *	with resident pages.
1949  */
1950 static void
1951 swp_pager_meta_free(vm_object_t object, vm_pindex_t pindex, vm_pindex_t count)
1952 {
1953 	struct swblk *sb;
1954 	daddr_t n_free, s_free;
1955 	vm_pindex_t last;
1956 	int i, limit, start;
1957 
1958 	VM_OBJECT_ASSERT_WLOCKED(object);
1959 	if (object->type != OBJT_SWAP || count == 0)
1960 		return;
1961 
1962 	swp_pager_init_freerange(&s_free, &n_free);
1963 	last = pindex + count;
1964 	for (;;) {
1965 		sb = SWAP_PCTRIE_LOOKUP_GE(&object->un_pager.swp.swp_blks,
1966 		    rounddown(pindex, SWAP_META_PAGES));
1967 		if (sb == NULL || sb->p >= last)
1968 			break;
1969 		start = pindex > sb->p ? pindex - sb->p : 0;
1970 		limit = last - sb->p < SWAP_META_PAGES ? last - sb->p :
1971 		    SWAP_META_PAGES;
1972 		for (i = start; i < limit; i++) {
1973 			if (sb->d[i] == SWAPBLK_NONE)
1974 				continue;
1975 			swp_pager_update_freerange(&s_free, &n_free, sb->d[i]);
1976 			sb->d[i] = SWAPBLK_NONE;
1977 		}
1978 		pindex = sb->p + SWAP_META_PAGES;
1979 		if (swp_pager_swblk_empty(sb, 0, start) &&
1980 		    swp_pager_swblk_empty(sb, limit, SWAP_META_PAGES)) {
1981 			SWAP_PCTRIE_REMOVE(&object->un_pager.swp.swp_blks,
1982 			    sb->p);
1983 			uma_zfree(swblk_zone, sb);
1984 		}
1985 	}
1986 	swp_pager_freeswapspace(s_free, n_free);
1987 }
1988 
1989 /*
1990  * SWP_PAGER_META_FREE_ALL() - destroy all swap metadata associated with object
1991  *
1992  *	This routine locates and destroys all swap metadata associated with
1993  *	an object.
1994  */
1995 static void
1996 swp_pager_meta_free_all(vm_object_t object)
1997 {
1998 	struct swblk *sb;
1999 	daddr_t n_free, s_free;
2000 	vm_pindex_t pindex;
2001 	int i;
2002 
2003 	VM_OBJECT_ASSERT_WLOCKED(object);
2004 	if (object->type != OBJT_SWAP)
2005 		return;
2006 
2007 	swp_pager_init_freerange(&s_free, &n_free);
2008 	for (pindex = 0; (sb = SWAP_PCTRIE_LOOKUP_GE(
2009 	    &object->un_pager.swp.swp_blks, pindex)) != NULL;) {
2010 		pindex = sb->p + SWAP_META_PAGES;
2011 		for (i = 0; i < SWAP_META_PAGES; i++) {
2012 			if (sb->d[i] == SWAPBLK_NONE)
2013 				continue;
2014 			swp_pager_update_freerange(&s_free, &n_free, sb->d[i]);
2015 		}
2016 		SWAP_PCTRIE_REMOVE(&object->un_pager.swp.swp_blks, sb->p);
2017 		uma_zfree(swblk_zone, sb);
2018 	}
2019 	swp_pager_freeswapspace(s_free, n_free);
2020 }
2021 
2022 /*
2023  * SWP_PAGER_METACTL() -  misc control of swap meta data.
2024  *
2025  *	This routine is capable of looking up, or removing swapblk
2026  *	assignments in the swap meta data.  It returns the swapblk being
2027  *	looked-up, popped, or SWAPBLK_NONE if the block was invalid.
2028  *
2029  *	When acting on a busy resident page and paging is in progress, we
2030  *	have to wait until paging is complete but otherwise can act on the
2031  *	busy page.
2032  *
2033  *	SWM_POP		remove from meta data but do not free it
2034  */
2035 static daddr_t
2036 swp_pager_meta_ctl(vm_object_t object, vm_pindex_t pindex, int flags)
2037 {
2038 	struct swblk *sb;
2039 	daddr_t r1;
2040 
2041 	if ((flags & SWM_POP) != 0)
2042 		VM_OBJECT_ASSERT_WLOCKED(object);
2043 	else
2044 		VM_OBJECT_ASSERT_LOCKED(object);
2045 
2046 	/*
2047 	 * The meta data only exists if the object is OBJT_SWAP
2048 	 * and even then might not be allocated yet.
2049 	 */
2050 	if (object->type != OBJT_SWAP)
2051 		return (SWAPBLK_NONE);
2052 
2053 	sb = SWAP_PCTRIE_LOOKUP(&object->un_pager.swp.swp_blks,
2054 	    rounddown(pindex, SWAP_META_PAGES));
2055 	if (sb == NULL)
2056 		return (SWAPBLK_NONE);
2057 	r1 = sb->d[pindex % SWAP_META_PAGES];
2058 	if (r1 == SWAPBLK_NONE)
2059 		return (SWAPBLK_NONE);
2060 	if ((flags & SWM_POP) != 0) {
2061 		sb->d[pindex % SWAP_META_PAGES] = SWAPBLK_NONE;
2062 		if (swp_pager_swblk_empty(sb, 0, SWAP_META_PAGES)) {
2063 			SWAP_PCTRIE_REMOVE(&object->un_pager.swp.swp_blks,
2064 			    rounddown(pindex, SWAP_META_PAGES));
2065 			uma_zfree(swblk_zone, sb);
2066 		}
2067 	}
2068 	return (r1);
2069 }
2070 
2071 /*
2072  * Returns the least page index which is greater than or equal to the
2073  * parameter pindex and for which there is a swap block allocated.
2074  * Returns object's size if the object's type is not swap or if there
2075  * are no allocated swap blocks for the object after the requested
2076  * pindex.
2077  */
2078 vm_pindex_t
2079 swap_pager_find_least(vm_object_t object, vm_pindex_t pindex)
2080 {
2081 	struct swblk *sb;
2082 	int i;
2083 
2084 	VM_OBJECT_ASSERT_LOCKED(object);
2085 	if (object->type != OBJT_SWAP)
2086 		return (object->size);
2087 
2088 	sb = SWAP_PCTRIE_LOOKUP_GE(&object->un_pager.swp.swp_blks,
2089 	    rounddown(pindex, SWAP_META_PAGES));
2090 	if (sb == NULL)
2091 		return (object->size);
2092 	if (sb->p < pindex) {
2093 		for (i = pindex % SWAP_META_PAGES; i < SWAP_META_PAGES; i++) {
2094 			if (sb->d[i] != SWAPBLK_NONE)
2095 				return (sb->p + i);
2096 		}
2097 		sb = SWAP_PCTRIE_LOOKUP_GE(&object->un_pager.swp.swp_blks,
2098 		    roundup(pindex, SWAP_META_PAGES));
2099 		if (sb == NULL)
2100 			return (object->size);
2101 	}
2102 	for (i = 0; i < SWAP_META_PAGES; i++) {
2103 		if (sb->d[i] != SWAPBLK_NONE)
2104 			return (sb->p + i);
2105 	}
2106 
2107 	/*
2108 	 * We get here if a swblk is present in the trie but it
2109 	 * doesn't map any blocks.
2110 	 */
2111 	MPASS(0);
2112 	return (object->size);
2113 }
2114 
2115 /*
2116  * System call swapon(name) enables swapping on device name,
2117  * which must be in the swdevsw.  Return EBUSY
2118  * if already swapping on this device.
2119  */
2120 #ifndef _SYS_SYSPROTO_H_
2121 struct swapon_args {
2122 	char *name;
2123 };
2124 #endif
2125 
2126 /*
2127  * MPSAFE
2128  */
2129 /* ARGSUSED */
2130 int
2131 sys_swapon(struct thread *td, struct swapon_args *uap)
2132 {
2133 	struct vattr attr;
2134 	struct vnode *vp;
2135 	struct nameidata nd;
2136 	int error;
2137 
2138 	error = priv_check(td, PRIV_SWAPON);
2139 	if (error)
2140 		return (error);
2141 
2142 	sx_xlock(&swdev_syscall_lock);
2143 
2144 	/*
2145 	 * Swap metadata may not fit in the KVM if we have physical
2146 	 * memory of >1GB.
2147 	 */
2148 	if (swblk_zone == NULL) {
2149 		error = ENOMEM;
2150 		goto done;
2151 	}
2152 
2153 	NDINIT(&nd, LOOKUP, ISOPEN | FOLLOW | AUDITVNODE1, UIO_USERSPACE,
2154 	    uap->name, td);
2155 	error = namei(&nd);
2156 	if (error)
2157 		goto done;
2158 
2159 	NDFREE(&nd, NDF_ONLY_PNBUF);
2160 	vp = nd.ni_vp;
2161 
2162 	if (vn_isdisk(vp, &error)) {
2163 		error = swapongeom(vp);
2164 	} else if (vp->v_type == VREG &&
2165 	    (vp->v_mount->mnt_vfc->vfc_flags & VFCF_NETWORK) != 0 &&
2166 	    (error = VOP_GETATTR(vp, &attr, td->td_ucred)) == 0) {
2167 		/*
2168 		 * Allow direct swapping to NFS regular files in the same
2169 		 * way that nfs_mountroot() sets up diskless swapping.
2170 		 */
2171 		error = swaponvp(td, vp, attr.va_size / DEV_BSIZE);
2172 	}
2173 
2174 	if (error)
2175 		vrele(vp);
2176 done:
2177 	sx_xunlock(&swdev_syscall_lock);
2178 	return (error);
2179 }
2180 
2181 /*
2182  * Check that the total amount of swap currently configured does not
2183  * exceed half the theoretical maximum.  If it does, print a warning
2184  * message.
2185  */
2186 static void
2187 swapon_check_swzone(void)
2188 {
2189 	unsigned long maxpages, npages;
2190 
2191 	npages = swap_total;
2192 	/* absolute maximum we can handle assuming 100% efficiency */
2193 	maxpages = uma_zone_get_max(swblk_zone) * SWAP_META_PAGES;
2194 
2195 	/* recommend using no more than half that amount */
2196 	if (npages > maxpages / 2) {
2197 		printf("warning: total configured swap (%lu pages) "
2198 		    "exceeds maximum recommended amount (%lu pages).\n",
2199 		    npages, maxpages / 2);
2200 		printf("warning: increase kern.maxswzone "
2201 		    "or reduce amount of swap.\n");
2202 	}
2203 }
2204 
2205 static void
2206 swaponsomething(struct vnode *vp, void *id, u_long nblks,
2207     sw_strategy_t *strategy, sw_close_t *close, dev_t dev, int flags)
2208 {
2209 	struct swdevt *sp, *tsp;
2210 	swblk_t dvbase;
2211 	u_long mblocks;
2212 
2213 	/*
2214 	 * nblks is in DEV_BSIZE'd chunks, convert to PAGE_SIZE'd chunks.
2215 	 * First chop nblks off to page-align it, then convert.
2216 	 *
2217 	 * sw->sw_nblks is in page-sized chunks now too.
2218 	 */
2219 	nblks &= ~(ctodb(1) - 1);
2220 	nblks = dbtoc(nblks);
2221 
2222 	/*
2223 	 * If we go beyond this, we get overflows in the radix
2224 	 * tree bitmap code.
2225 	 */
2226 	mblocks = 0x40000000 / BLIST_META_RADIX;
2227 	if (nblks > mblocks) {
2228 		printf(
2229     "WARNING: reducing swap size to maximum of %luMB per unit\n",
2230 		    mblocks / 1024 / 1024 * PAGE_SIZE);
2231 		nblks = mblocks;
2232 	}
2233 
2234 	sp = malloc(sizeof *sp, M_VMPGDATA, M_WAITOK | M_ZERO);
2235 	sp->sw_vp = vp;
2236 	sp->sw_id = id;
2237 	sp->sw_dev = dev;
2238 	sp->sw_flags = 0;
2239 	sp->sw_nblks = nblks;
2240 	sp->sw_used = 0;
2241 	sp->sw_strategy = strategy;
2242 	sp->sw_close = close;
2243 	sp->sw_flags = flags;
2244 
2245 	sp->sw_blist = blist_create(nblks, M_WAITOK);
2246 	/*
2247 	 * Do not free the first two block in order to avoid overwriting
2248 	 * any bsd label at the front of the partition
2249 	 */
2250 	blist_free(sp->sw_blist, 2, nblks - 2);
2251 
2252 	dvbase = 0;
2253 	mtx_lock(&sw_dev_mtx);
2254 	TAILQ_FOREACH(tsp, &swtailq, sw_list) {
2255 		if (tsp->sw_end >= dvbase) {
2256 			/*
2257 			 * We put one uncovered page between the devices
2258 			 * in order to definitively prevent any cross-device
2259 			 * I/O requests
2260 			 */
2261 			dvbase = tsp->sw_end + 1;
2262 		}
2263 	}
2264 	sp->sw_first = dvbase;
2265 	sp->sw_end = dvbase + nblks;
2266 	TAILQ_INSERT_TAIL(&swtailq, sp, sw_list);
2267 	nswapdev++;
2268 	swap_pager_avail += nblks - 2;
2269 	swap_total += nblks;
2270 	swapon_check_swzone();
2271 	swp_sizecheck();
2272 	mtx_unlock(&sw_dev_mtx);
2273 	EVENTHANDLER_INVOKE(swapon, sp);
2274 }
2275 
2276 /*
2277  * SYSCALL: swapoff(devname)
2278  *
2279  * Disable swapping on the given device.
2280  *
2281  * XXX: Badly designed system call: it should use a device index
2282  * rather than filename as specification.  We keep sw_vp around
2283  * only to make this work.
2284  */
2285 #ifndef _SYS_SYSPROTO_H_
2286 struct swapoff_args {
2287 	char *name;
2288 };
2289 #endif
2290 
2291 /*
2292  * MPSAFE
2293  */
2294 /* ARGSUSED */
2295 int
2296 sys_swapoff(struct thread *td, struct swapoff_args *uap)
2297 {
2298 	struct vnode *vp;
2299 	struct nameidata nd;
2300 	struct swdevt *sp;
2301 	int error;
2302 
2303 	error = priv_check(td, PRIV_SWAPOFF);
2304 	if (error)
2305 		return (error);
2306 
2307 	sx_xlock(&swdev_syscall_lock);
2308 
2309 	NDINIT(&nd, LOOKUP, FOLLOW | AUDITVNODE1, UIO_USERSPACE, uap->name,
2310 	    td);
2311 	error = namei(&nd);
2312 	if (error)
2313 		goto done;
2314 	NDFREE(&nd, NDF_ONLY_PNBUF);
2315 	vp = nd.ni_vp;
2316 
2317 	mtx_lock(&sw_dev_mtx);
2318 	TAILQ_FOREACH(sp, &swtailq, sw_list) {
2319 		if (sp->sw_vp == vp)
2320 			break;
2321 	}
2322 	mtx_unlock(&sw_dev_mtx);
2323 	if (sp == NULL) {
2324 		error = EINVAL;
2325 		goto done;
2326 	}
2327 	error = swapoff_one(sp, td->td_ucred);
2328 done:
2329 	sx_xunlock(&swdev_syscall_lock);
2330 	return (error);
2331 }
2332 
2333 static int
2334 swapoff_one(struct swdevt *sp, struct ucred *cred)
2335 {
2336 	u_long nblks;
2337 #ifdef MAC
2338 	int error;
2339 #endif
2340 
2341 	sx_assert(&swdev_syscall_lock, SA_XLOCKED);
2342 #ifdef MAC
2343 	(void) vn_lock(sp->sw_vp, LK_EXCLUSIVE | LK_RETRY);
2344 	error = mac_system_check_swapoff(cred, sp->sw_vp);
2345 	(void) VOP_UNLOCK(sp->sw_vp, 0);
2346 	if (error != 0)
2347 		return (error);
2348 #endif
2349 	nblks = sp->sw_nblks;
2350 
2351 	/*
2352 	 * We can turn off this swap device safely only if the
2353 	 * available virtual memory in the system will fit the amount
2354 	 * of data we will have to page back in, plus an epsilon so
2355 	 * the system doesn't become critically low on swap space.
2356 	 */
2357 	if (vm_free_count() + swap_pager_avail < nblks + nswap_lowat)
2358 		return (ENOMEM);
2359 
2360 	/*
2361 	 * Prevent further allocations on this device.
2362 	 */
2363 	mtx_lock(&sw_dev_mtx);
2364 	sp->sw_flags |= SW_CLOSING;
2365 	swap_pager_avail -= blist_fill(sp->sw_blist, 0, nblks);
2366 	swap_total -= nblks;
2367 	mtx_unlock(&sw_dev_mtx);
2368 
2369 	/*
2370 	 * Page in the contents of the device and close it.
2371 	 */
2372 	swap_pager_swapoff(sp);
2373 
2374 	sp->sw_close(curthread, sp);
2375 	mtx_lock(&sw_dev_mtx);
2376 	sp->sw_id = NULL;
2377 	TAILQ_REMOVE(&swtailq, sp, sw_list);
2378 	nswapdev--;
2379 	if (nswapdev == 0) {
2380 		swap_pager_full = 2;
2381 		swap_pager_almost_full = 1;
2382 	}
2383 	if (swdevhd == sp)
2384 		swdevhd = NULL;
2385 	mtx_unlock(&sw_dev_mtx);
2386 	blist_destroy(sp->sw_blist);
2387 	free(sp, M_VMPGDATA);
2388 	return (0);
2389 }
2390 
2391 void
2392 swapoff_all(void)
2393 {
2394 	struct swdevt *sp, *spt;
2395 	const char *devname;
2396 	int error;
2397 
2398 	sx_xlock(&swdev_syscall_lock);
2399 
2400 	mtx_lock(&sw_dev_mtx);
2401 	TAILQ_FOREACH_SAFE(sp, &swtailq, sw_list, spt) {
2402 		mtx_unlock(&sw_dev_mtx);
2403 		if (vn_isdisk(sp->sw_vp, NULL))
2404 			devname = devtoname(sp->sw_vp->v_rdev);
2405 		else
2406 			devname = "[file]";
2407 		error = swapoff_one(sp, thread0.td_ucred);
2408 		if (error != 0) {
2409 			printf("Cannot remove swap device %s (error=%d), "
2410 			    "skipping.\n", devname, error);
2411 		} else if (bootverbose) {
2412 			printf("Swap device %s removed.\n", devname);
2413 		}
2414 		mtx_lock(&sw_dev_mtx);
2415 	}
2416 	mtx_unlock(&sw_dev_mtx);
2417 
2418 	sx_xunlock(&swdev_syscall_lock);
2419 }
2420 
2421 void
2422 swap_pager_status(int *total, int *used)
2423 {
2424 	struct swdevt *sp;
2425 
2426 	*total = 0;
2427 	*used = 0;
2428 	mtx_lock(&sw_dev_mtx);
2429 	TAILQ_FOREACH(sp, &swtailq, sw_list) {
2430 		*total += sp->sw_nblks;
2431 		*used += sp->sw_used;
2432 	}
2433 	mtx_unlock(&sw_dev_mtx);
2434 }
2435 
2436 int
2437 swap_dev_info(int name, struct xswdev *xs, char *devname, size_t len)
2438 {
2439 	struct swdevt *sp;
2440 	const char *tmp_devname;
2441 	int error, n;
2442 
2443 	n = 0;
2444 	error = ENOENT;
2445 	mtx_lock(&sw_dev_mtx);
2446 	TAILQ_FOREACH(sp, &swtailq, sw_list) {
2447 		if (n != name) {
2448 			n++;
2449 			continue;
2450 		}
2451 		xs->xsw_version = XSWDEV_VERSION;
2452 		xs->xsw_dev = sp->sw_dev;
2453 		xs->xsw_flags = sp->sw_flags;
2454 		xs->xsw_nblks = sp->sw_nblks;
2455 		xs->xsw_used = sp->sw_used;
2456 		if (devname != NULL) {
2457 			if (vn_isdisk(sp->sw_vp, NULL))
2458 				tmp_devname = devtoname(sp->sw_vp->v_rdev);
2459 			else
2460 				tmp_devname = "[file]";
2461 			strncpy(devname, tmp_devname, len);
2462 		}
2463 		error = 0;
2464 		break;
2465 	}
2466 	mtx_unlock(&sw_dev_mtx);
2467 	return (error);
2468 }
2469 
2470 #if defined(COMPAT_FREEBSD11)
2471 #define XSWDEV_VERSION_11	1
2472 struct xswdev11 {
2473 	u_int	xsw_version;
2474 	uint32_t xsw_dev;
2475 	int	xsw_flags;
2476 	int	xsw_nblks;
2477 	int     xsw_used;
2478 };
2479 #endif
2480 
2481 #if defined(__amd64__) && defined(COMPAT_FREEBSD32)
2482 struct xswdev32 {
2483 	u_int	xsw_version;
2484 	u_int	xsw_dev1, xsw_dev2;
2485 	int	xsw_flags;
2486 	int	xsw_nblks;
2487 	int     xsw_used;
2488 };
2489 #endif
2490 
2491 static int
2492 sysctl_vm_swap_info(SYSCTL_HANDLER_ARGS)
2493 {
2494 	struct xswdev xs;
2495 #if defined(__amd64__) && defined(COMPAT_FREEBSD32)
2496 	struct xswdev32 xs32;
2497 #endif
2498 #if defined(COMPAT_FREEBSD11)
2499 	struct xswdev11 xs11;
2500 #endif
2501 	int error;
2502 
2503 	if (arg2 != 1)			/* name length */
2504 		return (EINVAL);
2505 	error = swap_dev_info(*(int *)arg1, &xs, NULL, 0);
2506 	if (error != 0)
2507 		return (error);
2508 #if defined(__amd64__) && defined(COMPAT_FREEBSD32)
2509 	if (req->oldlen == sizeof(xs32)) {
2510 		xs32.xsw_version = XSWDEV_VERSION;
2511 		xs32.xsw_dev1 = xs.xsw_dev;
2512 		xs32.xsw_dev2 = xs.xsw_dev >> 32;
2513 		xs32.xsw_flags = xs.xsw_flags;
2514 		xs32.xsw_nblks = xs.xsw_nblks;
2515 		xs32.xsw_used = xs.xsw_used;
2516 		error = SYSCTL_OUT(req, &xs32, sizeof(xs32));
2517 		return (error);
2518 	}
2519 #endif
2520 #if defined(COMPAT_FREEBSD11)
2521 	if (req->oldlen == sizeof(xs11)) {
2522 		xs11.xsw_version = XSWDEV_VERSION_11;
2523 		xs11.xsw_dev = xs.xsw_dev; /* truncation */
2524 		xs11.xsw_flags = xs.xsw_flags;
2525 		xs11.xsw_nblks = xs.xsw_nblks;
2526 		xs11.xsw_used = xs.xsw_used;
2527 		error = SYSCTL_OUT(req, &xs11, sizeof(xs11));
2528 		return (error);
2529 	}
2530 #endif
2531 	error = SYSCTL_OUT(req, &xs, sizeof(xs));
2532 	return (error);
2533 }
2534 
2535 SYSCTL_INT(_vm, OID_AUTO, nswapdev, CTLFLAG_RD, &nswapdev, 0,
2536     "Number of swap devices");
2537 SYSCTL_NODE(_vm, OID_AUTO, swap_info, CTLFLAG_RD | CTLFLAG_MPSAFE,
2538     sysctl_vm_swap_info,
2539     "Swap statistics by device");
2540 
2541 /*
2542  * Count the approximate swap usage in pages for a vmspace.  The
2543  * shadowed or not yet copied on write swap blocks are not accounted.
2544  * The map must be locked.
2545  */
2546 long
2547 vmspace_swap_count(struct vmspace *vmspace)
2548 {
2549 	vm_map_t map;
2550 	vm_map_entry_t cur;
2551 	vm_object_t object;
2552 	struct swblk *sb;
2553 	vm_pindex_t e, pi;
2554 	long count;
2555 	int i;
2556 
2557 	map = &vmspace->vm_map;
2558 	count = 0;
2559 
2560 	for (cur = map->header.next; cur != &map->header; cur = cur->next) {
2561 		if ((cur->eflags & MAP_ENTRY_IS_SUB_MAP) != 0)
2562 			continue;
2563 		object = cur->object.vm_object;
2564 		if (object == NULL || object->type != OBJT_SWAP)
2565 			continue;
2566 		VM_OBJECT_RLOCK(object);
2567 		if (object->type != OBJT_SWAP)
2568 			goto unlock;
2569 		pi = OFF_TO_IDX(cur->offset);
2570 		e = pi + OFF_TO_IDX(cur->end - cur->start);
2571 		for (;; pi = sb->p + SWAP_META_PAGES) {
2572 			sb = SWAP_PCTRIE_LOOKUP_GE(
2573 			    &object->un_pager.swp.swp_blks, pi);
2574 			if (sb == NULL || sb->p >= e)
2575 				break;
2576 			for (i = 0; i < SWAP_META_PAGES; i++) {
2577 				if (sb->p + i < e &&
2578 				    sb->d[i] != SWAPBLK_NONE)
2579 					count++;
2580 			}
2581 		}
2582 unlock:
2583 		VM_OBJECT_RUNLOCK(object);
2584 	}
2585 	return (count);
2586 }
2587 
2588 /*
2589  * GEOM backend
2590  *
2591  * Swapping onto disk devices.
2592  *
2593  */
2594 
2595 static g_orphan_t swapgeom_orphan;
2596 
2597 static struct g_class g_swap_class = {
2598 	.name = "SWAP",
2599 	.version = G_VERSION,
2600 	.orphan = swapgeom_orphan,
2601 };
2602 
2603 DECLARE_GEOM_CLASS(g_swap_class, g_class);
2604 
2605 
2606 static void
2607 swapgeom_close_ev(void *arg, int flags)
2608 {
2609 	struct g_consumer *cp;
2610 
2611 	cp = arg;
2612 	g_access(cp, -1, -1, 0);
2613 	g_detach(cp);
2614 	g_destroy_consumer(cp);
2615 }
2616 
2617 /*
2618  * Add a reference to the g_consumer for an inflight transaction.
2619  */
2620 static void
2621 swapgeom_acquire(struct g_consumer *cp)
2622 {
2623 
2624 	mtx_assert(&sw_dev_mtx, MA_OWNED);
2625 	cp->index++;
2626 }
2627 
2628 /*
2629  * Remove a reference from the g_consumer.  Post a close event if all
2630  * references go away, since the function might be called from the
2631  * biodone context.
2632  */
2633 static void
2634 swapgeom_release(struct g_consumer *cp, struct swdevt *sp)
2635 {
2636 
2637 	mtx_assert(&sw_dev_mtx, MA_OWNED);
2638 	cp->index--;
2639 	if (cp->index == 0) {
2640 		if (g_post_event(swapgeom_close_ev, cp, M_NOWAIT, NULL) == 0)
2641 			sp->sw_id = NULL;
2642 	}
2643 }
2644 
2645 static void
2646 swapgeom_done(struct bio *bp2)
2647 {
2648 	struct swdevt *sp;
2649 	struct buf *bp;
2650 	struct g_consumer *cp;
2651 
2652 	bp = bp2->bio_caller2;
2653 	cp = bp2->bio_from;
2654 	bp->b_ioflags = bp2->bio_flags;
2655 	if (bp2->bio_error)
2656 		bp->b_ioflags |= BIO_ERROR;
2657 	bp->b_resid = bp->b_bcount - bp2->bio_completed;
2658 	bp->b_error = bp2->bio_error;
2659 	bp->b_caller1 = NULL;
2660 	bufdone(bp);
2661 	sp = bp2->bio_caller1;
2662 	mtx_lock(&sw_dev_mtx);
2663 	swapgeom_release(cp, sp);
2664 	mtx_unlock(&sw_dev_mtx);
2665 	g_destroy_bio(bp2);
2666 }
2667 
2668 static void
2669 swapgeom_strategy(struct buf *bp, struct swdevt *sp)
2670 {
2671 	struct bio *bio;
2672 	struct g_consumer *cp;
2673 
2674 	mtx_lock(&sw_dev_mtx);
2675 	cp = sp->sw_id;
2676 	if (cp == NULL) {
2677 		mtx_unlock(&sw_dev_mtx);
2678 		bp->b_error = ENXIO;
2679 		bp->b_ioflags |= BIO_ERROR;
2680 		bufdone(bp);
2681 		return;
2682 	}
2683 	swapgeom_acquire(cp);
2684 	mtx_unlock(&sw_dev_mtx);
2685 	if (bp->b_iocmd == BIO_WRITE)
2686 		bio = g_new_bio();
2687 	else
2688 		bio = g_alloc_bio();
2689 	if (bio == NULL) {
2690 		mtx_lock(&sw_dev_mtx);
2691 		swapgeom_release(cp, sp);
2692 		mtx_unlock(&sw_dev_mtx);
2693 		bp->b_error = ENOMEM;
2694 		bp->b_ioflags |= BIO_ERROR;
2695 		bufdone(bp);
2696 		return;
2697 	}
2698 
2699 	bp->b_caller1 = bio;
2700 	bio->bio_caller1 = sp;
2701 	bio->bio_caller2 = bp;
2702 	bio->bio_cmd = bp->b_iocmd;
2703 	bio->bio_offset = (bp->b_blkno - sp->sw_first) * PAGE_SIZE;
2704 	bio->bio_length = bp->b_bcount;
2705 	bio->bio_done = swapgeom_done;
2706 	if (!buf_mapped(bp)) {
2707 		bio->bio_ma = bp->b_pages;
2708 		bio->bio_data = unmapped_buf;
2709 		bio->bio_ma_offset = (vm_offset_t)bp->b_offset & PAGE_MASK;
2710 		bio->bio_ma_n = bp->b_npages;
2711 		bio->bio_flags |= BIO_UNMAPPED;
2712 	} else {
2713 		bio->bio_data = bp->b_data;
2714 		bio->bio_ma = NULL;
2715 	}
2716 	g_io_request(bio, cp);
2717 	return;
2718 }
2719 
2720 static void
2721 swapgeom_orphan(struct g_consumer *cp)
2722 {
2723 	struct swdevt *sp;
2724 	int destroy;
2725 
2726 	mtx_lock(&sw_dev_mtx);
2727 	TAILQ_FOREACH(sp, &swtailq, sw_list) {
2728 		if (sp->sw_id == cp) {
2729 			sp->sw_flags |= SW_CLOSING;
2730 			break;
2731 		}
2732 	}
2733 	/*
2734 	 * Drop reference we were created with. Do directly since we're in a
2735 	 * special context where we don't have to queue the call to
2736 	 * swapgeom_close_ev().
2737 	 */
2738 	cp->index--;
2739 	destroy = ((sp != NULL) && (cp->index == 0));
2740 	if (destroy)
2741 		sp->sw_id = NULL;
2742 	mtx_unlock(&sw_dev_mtx);
2743 	if (destroy)
2744 		swapgeom_close_ev(cp, 0);
2745 }
2746 
2747 static void
2748 swapgeom_close(struct thread *td, struct swdevt *sw)
2749 {
2750 	struct g_consumer *cp;
2751 
2752 	mtx_lock(&sw_dev_mtx);
2753 	cp = sw->sw_id;
2754 	sw->sw_id = NULL;
2755 	mtx_unlock(&sw_dev_mtx);
2756 
2757 	/*
2758 	 * swapgeom_close() may be called from the biodone context,
2759 	 * where we cannot perform topology changes.  Delegate the
2760 	 * work to the events thread.
2761 	 */
2762 	if (cp != NULL)
2763 		g_waitfor_event(swapgeom_close_ev, cp, M_WAITOK, NULL);
2764 }
2765 
2766 static int
2767 swapongeom_locked(struct cdev *dev, struct vnode *vp)
2768 {
2769 	struct g_provider *pp;
2770 	struct g_consumer *cp;
2771 	static struct g_geom *gp;
2772 	struct swdevt *sp;
2773 	u_long nblks;
2774 	int error;
2775 
2776 	pp = g_dev_getprovider(dev);
2777 	if (pp == NULL)
2778 		return (ENODEV);
2779 	mtx_lock(&sw_dev_mtx);
2780 	TAILQ_FOREACH(sp, &swtailq, sw_list) {
2781 		cp = sp->sw_id;
2782 		if (cp != NULL && cp->provider == pp) {
2783 			mtx_unlock(&sw_dev_mtx);
2784 			return (EBUSY);
2785 		}
2786 	}
2787 	mtx_unlock(&sw_dev_mtx);
2788 	if (gp == NULL)
2789 		gp = g_new_geomf(&g_swap_class, "swap");
2790 	cp = g_new_consumer(gp);
2791 	cp->index = 1;	/* Number of active I/Os, plus one for being active. */
2792 	cp->flags |=  G_CF_DIRECT_SEND | G_CF_DIRECT_RECEIVE;
2793 	g_attach(cp, pp);
2794 	/*
2795 	 * XXX: Every time you think you can improve the margin for
2796 	 * footshooting, somebody depends on the ability to do so:
2797 	 * savecore(8) wants to write to our swapdev so we cannot
2798 	 * set an exclusive count :-(
2799 	 */
2800 	error = g_access(cp, 1, 1, 0);
2801 	if (error != 0) {
2802 		g_detach(cp);
2803 		g_destroy_consumer(cp);
2804 		return (error);
2805 	}
2806 	nblks = pp->mediasize / DEV_BSIZE;
2807 	swaponsomething(vp, cp, nblks, swapgeom_strategy,
2808 	    swapgeom_close, dev2udev(dev),
2809 	    (pp->flags & G_PF_ACCEPT_UNMAPPED) != 0 ? SW_UNMAPPED : 0);
2810 	return (0);
2811 }
2812 
2813 static int
2814 swapongeom(struct vnode *vp)
2815 {
2816 	int error;
2817 
2818 	vn_lock(vp, LK_EXCLUSIVE | LK_RETRY);
2819 	if (vp->v_type != VCHR || (vp->v_iflag & VI_DOOMED) != 0) {
2820 		error = ENOENT;
2821 	} else {
2822 		g_topology_lock();
2823 		error = swapongeom_locked(vp->v_rdev, vp);
2824 		g_topology_unlock();
2825 	}
2826 	VOP_UNLOCK(vp, 0);
2827 	return (error);
2828 }
2829 
2830 /*
2831  * VNODE backend
2832  *
2833  * This is used mainly for network filesystem (read: probably only tested
2834  * with NFS) swapfiles.
2835  *
2836  */
2837 
2838 static void
2839 swapdev_strategy(struct buf *bp, struct swdevt *sp)
2840 {
2841 	struct vnode *vp2;
2842 
2843 	bp->b_blkno = ctodb(bp->b_blkno - sp->sw_first);
2844 
2845 	vp2 = sp->sw_id;
2846 	vhold(vp2);
2847 	if (bp->b_iocmd == BIO_WRITE) {
2848 		if (bp->b_bufobj)
2849 			bufobj_wdrop(bp->b_bufobj);
2850 		bufobj_wref(&vp2->v_bufobj);
2851 	}
2852 	if (bp->b_bufobj != &vp2->v_bufobj)
2853 		bp->b_bufobj = &vp2->v_bufobj;
2854 	bp->b_vp = vp2;
2855 	bp->b_iooffset = dbtob(bp->b_blkno);
2856 	bstrategy(bp);
2857 	return;
2858 }
2859 
2860 static void
2861 swapdev_close(struct thread *td, struct swdevt *sp)
2862 {
2863 
2864 	VOP_CLOSE(sp->sw_vp, FREAD | FWRITE, td->td_ucred, td);
2865 	vrele(sp->sw_vp);
2866 }
2867 
2868 
2869 static int
2870 swaponvp(struct thread *td, struct vnode *vp, u_long nblks)
2871 {
2872 	struct swdevt *sp;
2873 	int error;
2874 
2875 	if (nblks == 0)
2876 		return (ENXIO);
2877 	mtx_lock(&sw_dev_mtx);
2878 	TAILQ_FOREACH(sp, &swtailq, sw_list) {
2879 		if (sp->sw_id == vp) {
2880 			mtx_unlock(&sw_dev_mtx);
2881 			return (EBUSY);
2882 		}
2883 	}
2884 	mtx_unlock(&sw_dev_mtx);
2885 
2886 	(void) vn_lock(vp, LK_EXCLUSIVE | LK_RETRY);
2887 #ifdef MAC
2888 	error = mac_system_check_swapon(td->td_ucred, vp);
2889 	if (error == 0)
2890 #endif
2891 		error = VOP_OPEN(vp, FREAD | FWRITE, td->td_ucred, td, NULL);
2892 	(void) VOP_UNLOCK(vp, 0);
2893 	if (error)
2894 		return (error);
2895 
2896 	swaponsomething(vp, vp, nblks, swapdev_strategy, swapdev_close,
2897 	    NODEV, 0);
2898 	return (0);
2899 }
2900 
2901 static int
2902 sysctl_swap_async_max(SYSCTL_HANDLER_ARGS)
2903 {
2904 	int error, new, n;
2905 
2906 	new = nsw_wcount_async_max;
2907 	error = sysctl_handle_int(oidp, &new, 0, req);
2908 	if (error != 0 || req->newptr == NULL)
2909 		return (error);
2910 
2911 	if (new > nswbuf / 2 || new < 1)
2912 		return (EINVAL);
2913 
2914 	mtx_lock(&swbuf_mtx);
2915 	while (nsw_wcount_async_max != new) {
2916 		/*
2917 		 * Adjust difference.  If the current async count is too low,
2918 		 * we will need to sqeeze our update slowly in.  Sleep with a
2919 		 * higher priority than getpbuf() to finish faster.
2920 		 */
2921 		n = new - nsw_wcount_async_max;
2922 		if (nsw_wcount_async + n >= 0) {
2923 			nsw_wcount_async += n;
2924 			nsw_wcount_async_max += n;
2925 			wakeup(&nsw_wcount_async);
2926 		} else {
2927 			nsw_wcount_async_max -= nsw_wcount_async;
2928 			nsw_wcount_async = 0;
2929 			msleep(&nsw_wcount_async, &swbuf_mtx, PSWP,
2930 			    "swpsysctl", 0);
2931 		}
2932 	}
2933 	mtx_unlock(&swbuf_mtx);
2934 
2935 	return (0);
2936 }
2937