xref: /freebsd/sys/vm/swap_pager.c (revision 656f49f8e2b0656824a5f10aeb760a00fdd3753f)
1 /*-
2  * Copyright (c) 1998 Matthew Dillon,
3  * Copyright (c) 1994 John S. Dyson
4  * Copyright (c) 1990 University of Utah.
5  * Copyright (c) 1982, 1986, 1989, 1993
6  *	The Regents of the University of California.  All rights reserved.
7  *
8  * This code is derived from software contributed to Berkeley by
9  * the Systems Programming Group of the University of Utah Computer
10  * Science Department.
11  *
12  * Redistribution and use in source and binary forms, with or without
13  * modification, are permitted provided that the following conditions
14  * are met:
15  * 1. Redistributions of source code must retain the above copyright
16  *    notice, this list of conditions and the following disclaimer.
17  * 2. Redistributions in binary form must reproduce the above copyright
18  *    notice, this list of conditions and the following disclaimer in the
19  *    documentation and/or other materials provided with the distribution.
20  * 3. All advertising materials mentioning features or use of this software
21  *    must display the following acknowledgement:
22  *	This product includes software developed by the University of
23  *	California, Berkeley and its contributors.
24  * 4. Neither the name of the University nor the names of its contributors
25  *    may be used to endorse or promote products derived from this software
26  *    without specific prior written permission.
27  *
28  * THIS SOFTWARE IS PROVIDED BY THE REGENTS AND CONTRIBUTORS ``AS IS'' AND
29  * ANY EXPRESS OR IMPLIED WARRANTIES, INCLUDING, BUT NOT LIMITED TO, THE
30  * IMPLIED WARRANTIES OF MERCHANTABILITY AND FITNESS FOR A PARTICULAR PURPOSE
31  * ARE DISCLAIMED.  IN NO EVENT SHALL THE REGENTS OR CONTRIBUTORS BE LIABLE
32  * FOR ANY DIRECT, INDIRECT, INCIDENTAL, SPECIAL, EXEMPLARY, OR CONSEQUENTIAL
33  * DAMAGES (INCLUDING, BUT NOT LIMITED TO, PROCUREMENT OF SUBSTITUTE GOODS
34  * OR SERVICES; LOSS OF USE, DATA, OR PROFITS; OR BUSINESS INTERRUPTION)
35  * HOWEVER CAUSED AND ON ANY THEORY OF LIABILITY, WHETHER IN CONTRACT, STRICT
36  * LIABILITY, OR TORT (INCLUDING NEGLIGENCE OR OTHERWISE) ARISING IN ANY WAY
37  * OUT OF THE USE OF THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF
38  * SUCH DAMAGE.
39  *
40  *				New Swap System
41  *				Matthew Dillon
42  *
43  * Radix Bitmap 'blists'.
44  *
45  *	- The new swapper uses the new radix bitmap code.  This should scale
46  *	  to arbitrarily small or arbitrarily large swap spaces and an almost
47  *	  arbitrary degree of fragmentation.
48  *
49  * Features:
50  *
51  *	- on the fly reallocation of swap during putpages.  The new system
52  *	  does not try to keep previously allocated swap blocks for dirty
53  *	  pages.
54  *
55  *	- on the fly deallocation of swap
56  *
57  *	- No more garbage collection required.  Unnecessarily allocated swap
58  *	  blocks only exist for dirty vm_page_t's now and these are already
59  *	  cycled (in a high-load system) by the pager.  We also do on-the-fly
60  *	  removal of invalidated swap blocks when a page is destroyed
61  *	  or renamed.
62  *
63  * from: Utah $Hdr: swap_pager.c 1.4 91/04/30$
64  *
65  *	@(#)swap_pager.c	8.9 (Berkeley) 3/21/94
66  *	@(#)vm_swap.c	8.5 (Berkeley) 2/17/94
67  */
68 
69 #include <sys/cdefs.h>
70 __FBSDID("$FreeBSD$");
71 
72 #include "opt_swap.h"
73 #include "opt_vm.h"
74 
75 #include <sys/param.h>
76 #include <sys/systm.h>
77 #include <sys/conf.h>
78 #include <sys/kernel.h>
79 #include <sys/priv.h>
80 #include <sys/proc.h>
81 #include <sys/bio.h>
82 #include <sys/buf.h>
83 #include <sys/disk.h>
84 #include <sys/fcntl.h>
85 #include <sys/mount.h>
86 #include <sys/namei.h>
87 #include <sys/vnode.h>
88 #include <sys/malloc.h>
89 #include <sys/racct.h>
90 #include <sys/resource.h>
91 #include <sys/resourcevar.h>
92 #include <sys/rwlock.h>
93 #include <sys/sysctl.h>
94 #include <sys/sysproto.h>
95 #include <sys/blist.h>
96 #include <sys/lock.h>
97 #include <sys/sx.h>
98 #include <sys/vmmeter.h>
99 
100 #include <security/mac/mac_framework.h>
101 
102 #include <vm/vm.h>
103 #include <vm/pmap.h>
104 #include <vm/vm_map.h>
105 #include <vm/vm_kern.h>
106 #include <vm/vm_object.h>
107 #include <vm/vm_page.h>
108 #include <vm/vm_pager.h>
109 #include <vm/vm_pageout.h>
110 #include <vm/vm_param.h>
111 #include <vm/swap_pager.h>
112 #include <vm/vm_extern.h>
113 #include <vm/uma.h>
114 
115 #include <geom/geom.h>
116 
117 /*
118  * SWB_NPAGES must be a power of 2.  It may be set to 1, 2, 4, 8, 16
119  * or 32 pages per allocation.
120  * The 32-page limit is due to the radix code (kern/subr_blist.c).
121  */
122 #ifndef MAX_PAGEOUT_CLUSTER
123 #define MAX_PAGEOUT_CLUSTER 16
124 #endif
125 
126 #if !defined(SWB_NPAGES)
127 #define SWB_NPAGES	MAX_PAGEOUT_CLUSTER
128 #endif
129 
130 /*
131  * The swblock structure maps an object and a small, fixed-size range
132  * of page indices to disk addresses within a swap area.
133  * The collection of these mappings is implemented as a hash table.
134  * Unused disk addresses within a swap area are allocated and managed
135  * using a blist.
136  */
137 #define SWCORRECT(n) (sizeof(void *) * (n) / sizeof(daddr_t))
138 #define SWAP_META_PAGES		(SWB_NPAGES * 2)
139 #define SWAP_META_MASK		(SWAP_META_PAGES - 1)
140 
141 struct swblock {
142 	struct swblock	*swb_hnext;
143 	vm_object_t	swb_object;
144 	vm_pindex_t	swb_index;
145 	int		swb_count;
146 	daddr_t		swb_pages[SWAP_META_PAGES];
147 };
148 
149 static MALLOC_DEFINE(M_VMPGDATA, "vm_pgdata", "swap pager private data");
150 static struct mtx sw_dev_mtx;
151 static TAILQ_HEAD(, swdevt) swtailq = TAILQ_HEAD_INITIALIZER(swtailq);
152 static struct swdevt *swdevhd;	/* Allocate from here next */
153 static int nswapdev;		/* Number of swap devices */
154 int swap_pager_avail;
155 static int swdev_syscall_active = 0; /* serialize swap(on|off) */
156 
157 static vm_ooffset_t swap_total;
158 SYSCTL_QUAD(_vm, OID_AUTO, swap_total, CTLFLAG_RD, &swap_total, 0,
159     "Total amount of available swap storage.");
160 static vm_ooffset_t swap_reserved;
161 SYSCTL_QUAD(_vm, OID_AUTO, swap_reserved, CTLFLAG_RD, &swap_reserved, 0,
162     "Amount of swap storage needed to back all allocated anonymous memory.");
163 static int overcommit = 0;
164 SYSCTL_INT(_vm, OID_AUTO, 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 int
180 swap_reserve(vm_ooffset_t incr)
181 {
182 
183 	return (swap_reserve_by_cred(incr, curthread->td_ucred));
184 }
185 
186 int
187 swap_reserve_by_cred(vm_ooffset_t incr, struct ucred *cred)
188 {
189 	vm_ooffset_t r, s;
190 	int res, error;
191 	static int curfail;
192 	static struct timeval lastfail;
193 	struct uidinfo *uip;
194 
195 	uip = cred->cr_ruidinfo;
196 
197 	if (incr & PAGE_MASK)
198 		panic("swap_reserve: & PAGE_MASK");
199 
200 #ifdef RACCT
201 	if (racct_enable) {
202 		PROC_LOCK(curproc);
203 		error = racct_add(curproc, RACCT_SWAP, incr);
204 		PROC_UNLOCK(curproc);
205 		if (error != 0)
206 			return (0);
207 	}
208 #endif
209 
210 	res = 0;
211 	mtx_lock(&sw_dev_mtx);
212 	r = swap_reserved + incr;
213 	if (overcommit & SWAP_RESERVE_ALLOW_NONWIRED) {
214 		s = vm_cnt.v_page_count - vm_cnt.v_free_reserved - vm_cnt.v_wire_count;
215 		s *= PAGE_SIZE;
216 	} else
217 		s = 0;
218 	s += swap_total;
219 	if ((overcommit & SWAP_RESERVE_FORCE_ON) == 0 || r <= s ||
220 	    (error = priv_check(curthread, PRIV_VM_SWAP_NOQUOTA)) == 0) {
221 		res = 1;
222 		swap_reserved = r;
223 	}
224 	mtx_unlock(&sw_dev_mtx);
225 
226 	if (res) {
227 		UIDINFO_VMSIZE_LOCK(uip);
228 		if ((overcommit & SWAP_RESERVE_RLIMIT_ON) != 0 &&
229 		    uip->ui_vmsize + incr > lim_cur(curthread, RLIMIT_SWAP) &&
230 		    priv_check(curthread, PRIV_VM_SWAP_NORLIMIT))
231 			res = 0;
232 		else
233 			uip->ui_vmsize += incr;
234 		UIDINFO_VMSIZE_UNLOCK(uip);
235 		if (!res) {
236 			mtx_lock(&sw_dev_mtx);
237 			swap_reserved -= incr;
238 			mtx_unlock(&sw_dev_mtx);
239 		}
240 	}
241 	if (!res && ppsratecheck(&lastfail, &curfail, 1)) {
242 		printf("uid %d, pid %d: swap reservation for %jd bytes failed\n",
243 		    uip->ui_uid, curproc->p_pid, incr);
244 	}
245 
246 #ifdef RACCT
247 	if (!res) {
248 		PROC_LOCK(curproc);
249 		racct_sub(curproc, RACCT_SWAP, incr);
250 		PROC_UNLOCK(curproc);
251 	}
252 #endif
253 
254 	return (res);
255 }
256 
257 void
258 swap_reserve_force(vm_ooffset_t incr)
259 {
260 	struct uidinfo *uip;
261 
262 	mtx_lock(&sw_dev_mtx);
263 	swap_reserved += incr;
264 	mtx_unlock(&sw_dev_mtx);
265 
266 #ifdef RACCT
267 	PROC_LOCK(curproc);
268 	racct_add_force(curproc, RACCT_SWAP, incr);
269 	PROC_UNLOCK(curproc);
270 #endif
271 
272 	uip = curthread->td_ucred->cr_ruidinfo;
273 	PROC_LOCK(curproc);
274 	UIDINFO_VMSIZE_LOCK(uip);
275 	uip->ui_vmsize += incr;
276 	UIDINFO_VMSIZE_UNLOCK(uip);
277 	PROC_UNLOCK(curproc);
278 }
279 
280 void
281 swap_release(vm_ooffset_t decr)
282 {
283 	struct ucred *cred;
284 
285 	PROC_LOCK(curproc);
286 	cred = curthread->td_ucred;
287 	swap_release_by_cred(decr, cred);
288 	PROC_UNLOCK(curproc);
289 }
290 
291 void
292 swap_release_by_cred(vm_ooffset_t decr, struct ucred *cred)
293 {
294  	struct uidinfo *uip;
295 
296 	uip = cred->cr_ruidinfo;
297 
298 	if (decr & PAGE_MASK)
299 		panic("swap_release: & PAGE_MASK");
300 
301 	mtx_lock(&sw_dev_mtx);
302 	if (swap_reserved < decr)
303 		panic("swap_reserved < decr");
304 	swap_reserved -= decr;
305 	mtx_unlock(&sw_dev_mtx);
306 
307 	UIDINFO_VMSIZE_LOCK(uip);
308 	if (uip->ui_vmsize < decr)
309 		printf("negative vmsize for uid = %d\n", uip->ui_uid);
310 	uip->ui_vmsize -= decr;
311 	UIDINFO_VMSIZE_UNLOCK(uip);
312 
313 	racct_sub_cred(cred, RACCT_SWAP, decr);
314 }
315 
316 #define SWM_FREE	0x02	/* free, period			*/
317 #define SWM_POP		0x04	/* pop out			*/
318 
319 int swap_pager_full = 2;	/* swap space exhaustion (task killing) */
320 static int swap_pager_almost_full = 1; /* swap space exhaustion (w/hysteresis)*/
321 static int nsw_rcount;		/* free read buffers			*/
322 static int nsw_wcount_sync;	/* limit write buffers / synchronous	*/
323 static int nsw_wcount_async;	/* limit write buffers / asynchronous	*/
324 static int nsw_wcount_async_max;/* assigned maximum			*/
325 static int nsw_cluster_max;	/* maximum VOP I/O allowed		*/
326 
327 static int sysctl_swap_async_max(SYSCTL_HANDLER_ARGS);
328 SYSCTL_PROC(_vm, OID_AUTO, swap_async_max, CTLTYPE_INT | CTLFLAG_RW,
329     NULL, 0, sysctl_swap_async_max, "I", "Maximum running async swap ops");
330 
331 static struct swblock **swhash;
332 static int swhash_mask;
333 static struct mtx swhash_mtx;
334 
335 static struct sx sw_alloc_sx;
336 
337 /*
338  * "named" and "unnamed" anon region objects.  Try to reduce the overhead
339  * of searching a named list by hashing it just a little.
340  */
341 
342 #define NOBJLISTS		8
343 
344 #define NOBJLIST(handle)	\
345 	(&swap_pager_object_list[((int)(intptr_t)handle >> 4) & (NOBJLISTS-1)])
346 
347 static struct mtx sw_alloc_mtx;	/* protect list manipulation */
348 static struct pagerlst	swap_pager_object_list[NOBJLISTS];
349 static uma_zone_t	swap_zone;
350 
351 /*
352  * pagerops for OBJT_SWAP - "swap pager".  Some ops are also global procedure
353  * calls hooked from other parts of the VM system and do not appear here.
354  * (see vm/swap_pager.h).
355  */
356 static vm_object_t
357 		swap_pager_alloc(void *handle, vm_ooffset_t size,
358 		    vm_prot_t prot, vm_ooffset_t offset, struct ucred *);
359 static void	swap_pager_dealloc(vm_object_t object);
360 static int	swap_pager_getpages(vm_object_t, vm_page_t *, int, int);
361 static int	swap_pager_getpages_async(vm_object_t, vm_page_t *, int, int,
362     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  * dmmax is in page-sized chunks with the new swap system.  It was
383  * dev-bsized chunks in the old.  dmmax is always a power of 2.
384  *
385  * swap_*() routines are externally accessible.  swp_*() routines are
386  * internal.
387  */
388 static int dmmax;
389 static int nswap_lowat = 128;	/* in pages, swap_pager_almost_full warn */
390 static int nswap_hiwat = 512;	/* in pages, swap_pager_almost_full warn */
391 
392 SYSCTL_INT(_vm, OID_AUTO, dmmax,
393 	CTLFLAG_RD, &dmmax, 0, "Maximum size of a swap block");
394 
395 static void	swp_sizecheck(void);
396 static void	swp_pager_async_iodone(struct buf *bp);
397 static int	swapongeom(struct thread *, struct vnode *);
398 static int	swaponvp(struct thread *, struct vnode *, u_long);
399 static int	swapoff_one(struct swdevt *sp, struct ucred *cred);
400 
401 /*
402  * Swap bitmap functions
403  */
404 static void	swp_pager_freeswapspace(daddr_t blk, int npages);
405 static daddr_t	swp_pager_getswapspace(int npages);
406 
407 /*
408  * Metadata functions
409  */
410 static struct swblock **swp_pager_hash(vm_object_t object, vm_pindex_t index);
411 static void swp_pager_meta_build(vm_object_t, vm_pindex_t, daddr_t);
412 static void swp_pager_meta_free(vm_object_t, vm_pindex_t, daddr_t);
413 static void swp_pager_meta_free_all(vm_object_t);
414 static daddr_t swp_pager_meta_ctl(vm_object_t, vm_pindex_t, int);
415 
416 static void
417 swp_pager_free_nrpage(vm_page_t m)
418 {
419 
420 	vm_page_lock(m);
421 	if (m->wire_count == 0)
422 		vm_page_free(m);
423 	vm_page_unlock(m);
424 }
425 
426 /*
427  * SWP_SIZECHECK() -	update swap_pager_full indication
428  *
429  *	update the swap_pager_almost_full indication and warn when we are
430  *	about to run out of swap space, using lowat/hiwat hysteresis.
431  *
432  *	Clear swap_pager_full ( task killing ) indication when lowat is met.
433  *
434  *	No restrictions on call
435  *	This routine may not block.
436  */
437 static void
438 swp_sizecheck(void)
439 {
440 
441 	if (swap_pager_avail < nswap_lowat) {
442 		if (swap_pager_almost_full == 0) {
443 			printf("swap_pager: out of swap space\n");
444 			swap_pager_almost_full = 1;
445 		}
446 	} else {
447 		swap_pager_full = 0;
448 		if (swap_pager_avail > nswap_hiwat)
449 			swap_pager_almost_full = 0;
450 	}
451 }
452 
453 /*
454  * SWP_PAGER_HASH() -	hash swap meta data
455  *
456  *	This is an helper function which hashes the swapblk given
457  *	the object and page index.  It returns a pointer to a pointer
458  *	to the object, or a pointer to a NULL pointer if it could not
459  *	find a swapblk.
460  */
461 static struct swblock **
462 swp_pager_hash(vm_object_t object, vm_pindex_t index)
463 {
464 	struct swblock **pswap;
465 	struct swblock *swap;
466 
467 	index &= ~(vm_pindex_t)SWAP_META_MASK;
468 	pswap = &swhash[(index ^ (int)(intptr_t)object) & swhash_mask];
469 	while ((swap = *pswap) != NULL) {
470 		if (swap->swb_object == object &&
471 		    swap->swb_index == index
472 		) {
473 			break;
474 		}
475 		pswap = &swap->swb_hnext;
476 	}
477 	return (pswap);
478 }
479 
480 /*
481  * SWAP_PAGER_INIT() -	initialize the swap pager!
482  *
483  *	Expected to be started from system init.  NOTE:  This code is run
484  *	before much else so be careful what you depend on.  Most of the VM
485  *	system has yet to be initialized at this point.
486  */
487 static void
488 swap_pager_init(void)
489 {
490 	/*
491 	 * Initialize object lists
492 	 */
493 	int i;
494 
495 	for (i = 0; i < NOBJLISTS; ++i)
496 		TAILQ_INIT(&swap_pager_object_list[i]);
497 	mtx_init(&sw_alloc_mtx, "swap_pager list", NULL, MTX_DEF);
498 	mtx_init(&sw_dev_mtx, "swapdev", NULL, MTX_DEF);
499 
500 	/*
501 	 * Device Stripe, in PAGE_SIZE'd blocks
502 	 */
503 	dmmax = SWB_NPAGES * 2;
504 }
505 
506 /*
507  * SWAP_PAGER_SWAP_INIT() - swap pager initialization from pageout process
508  *
509  *	Expected to be started from pageout process once, prior to entering
510  *	its main loop.
511  */
512 void
513 swap_pager_swap_init(void)
514 {
515 	unsigned long n, n2;
516 
517 	/*
518 	 * Number of in-transit swap bp operations.  Don't
519 	 * exhaust the pbufs completely.  Make sure we
520 	 * initialize workable values (0 will work for hysteresis
521 	 * but it isn't very efficient).
522 	 *
523 	 * The nsw_cluster_max is constrained by the bp->b_pages[]
524 	 * array (MAXPHYS/PAGE_SIZE) and our locally defined
525 	 * MAX_PAGEOUT_CLUSTER.   Also be aware that swap ops are
526 	 * constrained by the swap device interleave stripe size.
527 	 *
528 	 * Currently we hardwire nsw_wcount_async to 4.  This limit is
529 	 * designed to prevent other I/O from having high latencies due to
530 	 * our pageout I/O.  The value 4 works well for one or two active swap
531 	 * devices but is probably a little low if you have more.  Even so,
532 	 * a higher value would probably generate only a limited improvement
533 	 * with three or four active swap devices since the system does not
534 	 * typically have to pageout at extreme bandwidths.   We will want
535 	 * at least 2 per swap devices, and 4 is a pretty good value if you
536 	 * have one NFS swap device due to the command/ack latency over NFS.
537 	 * So it all works out pretty well.
538 	 */
539 	nsw_cluster_max = min((MAXPHYS/PAGE_SIZE), MAX_PAGEOUT_CLUSTER);
540 
541 	mtx_lock(&pbuf_mtx);
542 	nsw_rcount = (nswbuf + 1) / 2;
543 	nsw_wcount_sync = (nswbuf + 3) / 4;
544 	nsw_wcount_async = 4;
545 	nsw_wcount_async_max = nsw_wcount_async;
546 	mtx_unlock(&pbuf_mtx);
547 
548 	/*
549 	 * Initialize our zone.  Right now I'm just guessing on the number
550 	 * we need based on the number of pages in the system.  Each swblock
551 	 * can hold 32 pages, so this is probably overkill.  This reservation
552 	 * is typically limited to around 32MB by default.
553 	 */
554 	n = vm_cnt.v_page_count / 2;
555 	if (maxswzone && n > maxswzone / sizeof(struct swblock))
556 		n = maxswzone / sizeof(struct swblock);
557 	n2 = n;
558 	swap_zone = uma_zcreate("SWAPMETA", sizeof(struct swblock), NULL, NULL,
559 	    NULL, NULL, UMA_ALIGN_PTR, UMA_ZONE_NOFREE | UMA_ZONE_VM);
560 	if (swap_zone == NULL)
561 		panic("failed to create swap_zone.");
562 	do {
563 		if (uma_zone_reserve_kva(swap_zone, n))
564 			break;
565 		/*
566 		 * if the allocation failed, try a zone two thirds the
567 		 * size of the previous attempt.
568 		 */
569 		n -= ((n + 2) / 3);
570 	} while (n > 0);
571 	if (n2 != n)
572 		printf("Swap zone entries reduced from %lu to %lu.\n", n2, n);
573 	swap_maxpages = n * SWAP_META_PAGES;
574 	swzone = n * sizeof(struct swblock);
575 	n2 = n;
576 
577 	/*
578 	 * Initialize our meta-data hash table.  The swapper does not need to
579 	 * be quite as efficient as the VM system, so we do not use an
580 	 * oversized hash table.
581 	 *
582 	 * 	n: 		size of hash table, must be power of 2
583 	 *	swhash_mask:	hash table index mask
584 	 */
585 	for (n = 1; n < n2 / 8; n *= 2)
586 		;
587 	swhash = malloc(sizeof(struct swblock *) * n, M_VMPGDATA, M_WAITOK | M_ZERO);
588 	swhash_mask = n - 1;
589 	mtx_init(&swhash_mtx, "swap_pager swhash", NULL, MTX_DEF);
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.  We do this by creating an OBJT_DEFAULT object
598  *	and then converting it with swp_pager_meta_build().
599  *
600  *	This routine may block in vm_object_allocate() and create a named
601  *	object lookup race, so we must interlock.
602  *
603  * MPSAFE
604  */
605 static vm_object_t
606 swap_pager_alloc(void *handle, vm_ooffset_t size, vm_prot_t prot,
607     vm_ooffset_t offset, struct ucred *cred)
608 {
609 	vm_object_t object;
610 	vm_pindex_t pindex;
611 
612 	pindex = OFF_TO_IDX(offset + PAGE_MASK + size);
613 	if (handle) {
614 		mtx_lock(&Giant);
615 		/*
616 		 * Reference existing named region or allocate new one.  There
617 		 * should not be a race here against swp_pager_meta_build()
618 		 * as called from vm_page_remove() in regards to the lookup
619 		 * of the handle.
620 		 */
621 		sx_xlock(&sw_alloc_sx);
622 		object = vm_pager_object_lookup(NOBJLIST(handle), handle);
623 		if (object == NULL) {
624 			if (cred != NULL) {
625 				if (!swap_reserve_by_cred(size, cred)) {
626 					sx_xunlock(&sw_alloc_sx);
627 					mtx_unlock(&Giant);
628 					return (NULL);
629 				}
630 				crhold(cred);
631 			}
632 			object = vm_object_allocate(OBJT_DEFAULT, pindex);
633 			VM_OBJECT_WLOCK(object);
634 			object->handle = handle;
635 			if (cred != NULL) {
636 				object->cred = cred;
637 				object->charge = size;
638 			}
639 			swp_pager_meta_build(object, 0, SWAPBLK_NONE);
640 			VM_OBJECT_WUNLOCK(object);
641 		}
642 		sx_xunlock(&sw_alloc_sx);
643 		mtx_unlock(&Giant);
644 	} else {
645 		if (cred != NULL) {
646 			if (!swap_reserve_by_cred(size, cred))
647 				return (NULL);
648 			crhold(cred);
649 		}
650 		object = vm_object_allocate(OBJT_DEFAULT, pindex);
651 		VM_OBJECT_WLOCK(object);
652 		if (cred != NULL) {
653 			object->cred = cred;
654 			object->charge = size;
655 		}
656 		swp_pager_meta_build(object, 0, SWAPBLK_NONE);
657 		VM_OBJECT_WUNLOCK(object);
658 	}
659 	return (object);
660 }
661 
662 /*
663  * SWAP_PAGER_DEALLOC() -	remove swap metadata from object
664  *
665  *	The swap backing for the object is destroyed.  The code is
666  *	designed such that we can reinstantiate it later, but this
667  *	routine is typically called only when the entire object is
668  *	about to be destroyed.
669  *
670  *	The object must be locked.
671  */
672 static void
673 swap_pager_dealloc(vm_object_t object)
674 {
675 
676 	/*
677 	 * Remove from list right away so lookups will fail if we block for
678 	 * pageout completion.
679 	 */
680 	if (object->handle != NULL) {
681 		mtx_lock(&sw_alloc_mtx);
682 		TAILQ_REMOVE(NOBJLIST(object->handle), object, pager_object_list);
683 		mtx_unlock(&sw_alloc_mtx);
684 	}
685 
686 	VM_OBJECT_ASSERT_WLOCKED(object);
687 	vm_object_pip_wait(object, "swpdea");
688 
689 	/*
690 	 * Free all remaining metadata.  We only bother to free it from
691 	 * the swap meta data.  We do not attempt to free swapblk's still
692 	 * associated with vm_page_t's for this object.  We do not care
693 	 * if paging is still in progress on some objects.
694 	 */
695 	swp_pager_meta_free_all(object);
696 	object->handle = NULL;
697 	object->type = OBJT_DEAD;
698 }
699 
700 /************************************************************************
701  *			SWAP PAGER BITMAP ROUTINES			*
702  ************************************************************************/
703 
704 /*
705  * SWP_PAGER_GETSWAPSPACE() -	allocate raw swap space
706  *
707  *	Allocate swap for the requested number of pages.  The starting
708  *	swap block number (a page index) is returned or SWAPBLK_NONE
709  *	if the allocation failed.
710  *
711  *	Also has the side effect of advising that somebody made a mistake
712  *	when they configured swap and didn't configure enough.
713  *
714  *	This routine may not sleep.
715  *
716  *	We allocate in round-robin fashion from the configured devices.
717  */
718 static daddr_t
719 swp_pager_getswapspace(int npages)
720 {
721 	daddr_t blk;
722 	struct swdevt *sp;
723 	int i;
724 
725 	blk = SWAPBLK_NONE;
726 	mtx_lock(&sw_dev_mtx);
727 	sp = swdevhd;
728 	for (i = 0; i < nswapdev; i++) {
729 		if (sp == NULL)
730 			sp = TAILQ_FIRST(&swtailq);
731 		if (!(sp->sw_flags & SW_CLOSING)) {
732 			blk = blist_alloc(sp->sw_blist, npages);
733 			if (blk != SWAPBLK_NONE) {
734 				blk += sp->sw_first;
735 				sp->sw_used += npages;
736 				swap_pager_avail -= npages;
737 				swp_sizecheck();
738 				swdevhd = TAILQ_NEXT(sp, sw_list);
739 				goto done;
740 			}
741 		}
742 		sp = TAILQ_NEXT(sp, sw_list);
743 	}
744 	if (swap_pager_full != 2) {
745 		printf("swap_pager_getswapspace(%d): failed\n", npages);
746 		swap_pager_full = 2;
747 		swap_pager_almost_full = 1;
748 	}
749 	swdevhd = NULL;
750 done:
751 	mtx_unlock(&sw_dev_mtx);
752 	return (blk);
753 }
754 
755 static int
756 swp_pager_isondev(daddr_t blk, struct swdevt *sp)
757 {
758 
759 	return (blk >= sp->sw_first && blk < sp->sw_end);
760 }
761 
762 static void
763 swp_pager_strategy(struct buf *bp)
764 {
765 	struct swdevt *sp;
766 
767 	mtx_lock(&sw_dev_mtx);
768 	TAILQ_FOREACH(sp, &swtailq, sw_list) {
769 		if (bp->b_blkno >= sp->sw_first && bp->b_blkno < sp->sw_end) {
770 			mtx_unlock(&sw_dev_mtx);
771 			if ((sp->sw_flags & SW_UNMAPPED) != 0 &&
772 			    unmapped_buf_allowed) {
773 				bp->b_data = unmapped_buf;
774 				bp->b_offset = 0;
775 			} else {
776 				pmap_qenter((vm_offset_t)bp->b_data,
777 				    &bp->b_pages[0], bp->b_bcount / PAGE_SIZE);
778 			}
779 			sp->sw_strategy(bp, sp);
780 			return;
781 		}
782 	}
783 	panic("Swapdev not found");
784 }
785 
786 
787 /*
788  * SWP_PAGER_FREESWAPSPACE() -	free raw swap space
789  *
790  *	This routine returns the specified swap blocks back to the bitmap.
791  *
792  *	This routine may not sleep.
793  */
794 static void
795 swp_pager_freeswapspace(daddr_t blk, int npages)
796 {
797 	struct swdevt *sp;
798 
799 	mtx_lock(&sw_dev_mtx);
800 	TAILQ_FOREACH(sp, &swtailq, sw_list) {
801 		if (blk >= sp->sw_first && blk < sp->sw_end) {
802 			sp->sw_used -= npages;
803 			/*
804 			 * If we are attempting to stop swapping on
805 			 * this device, we don't want to mark any
806 			 * blocks free lest they be reused.
807 			 */
808 			if ((sp->sw_flags & SW_CLOSING) == 0) {
809 				blist_free(sp->sw_blist, blk - sp->sw_first,
810 				    npages);
811 				swap_pager_avail += npages;
812 				swp_sizecheck();
813 			}
814 			mtx_unlock(&sw_dev_mtx);
815 			return;
816 		}
817 	}
818 	panic("Swapdev not found");
819 }
820 
821 /*
822  * SWAP_PAGER_FREESPACE() -	frees swap blocks associated with a page
823  *				range within an object.
824  *
825  *	This is a globally accessible routine.
826  *
827  *	This routine removes swapblk assignments from swap metadata.
828  *
829  *	The external callers of this routine typically have already destroyed
830  *	or renamed vm_page_t's associated with this range in the object so
831  *	we should be ok.
832  *
833  *	The object must be locked.
834  */
835 void
836 swap_pager_freespace(vm_object_t object, vm_pindex_t start, vm_size_t size)
837 {
838 
839 	swp_pager_meta_free(object, start, size);
840 }
841 
842 /*
843  * SWAP_PAGER_RESERVE() - reserve swap blocks in object
844  *
845  *	Assigns swap blocks to the specified range within the object.  The
846  *	swap blocks are not zeroed.  Any previous swap assignment is destroyed.
847  *
848  *	Returns 0 on success, -1 on failure.
849  */
850 int
851 swap_pager_reserve(vm_object_t object, vm_pindex_t start, vm_size_t size)
852 {
853 	int n = 0;
854 	daddr_t blk = SWAPBLK_NONE;
855 	vm_pindex_t beg = start;	/* save start index */
856 
857 	VM_OBJECT_WLOCK(object);
858 	while (size) {
859 		if (n == 0) {
860 			n = BLIST_MAX_ALLOC;
861 			while ((blk = swp_pager_getswapspace(n)) == SWAPBLK_NONE) {
862 				n >>= 1;
863 				if (n == 0) {
864 					swp_pager_meta_free(object, beg, start - beg);
865 					VM_OBJECT_WUNLOCK(object);
866 					return (-1);
867 				}
868 			}
869 		}
870 		swp_pager_meta_build(object, start, blk);
871 		--size;
872 		++start;
873 		++blk;
874 		--n;
875 	}
876 	swp_pager_meta_free(object, start, n);
877 	VM_OBJECT_WUNLOCK(object);
878 	return (0);
879 }
880 
881 /*
882  * SWAP_PAGER_COPY() -  copy blocks from source pager to destination pager
883  *			and destroy the source.
884  *
885  *	Copy any valid swapblks from the source to the destination.  In
886  *	cases where both the source and destination have a valid swapblk,
887  *	we keep the destination's.
888  *
889  *	This routine is allowed to sleep.  It may sleep allocating metadata
890  *	indirectly through swp_pager_meta_build() or if paging is still in
891  *	progress on the source.
892  *
893  *	The source object contains no vm_page_t's (which is just as well)
894  *
895  *	The source object is of type OBJT_SWAP.
896  *
897  *	The source and destination objects must be locked.
898  *	Both object locks may temporarily be released.
899  */
900 void
901 swap_pager_copy(vm_object_t srcobject, vm_object_t dstobject,
902     vm_pindex_t offset, int destroysource)
903 {
904 	vm_pindex_t i;
905 
906 	VM_OBJECT_ASSERT_WLOCKED(srcobject);
907 	VM_OBJECT_ASSERT_WLOCKED(dstobject);
908 
909 	/*
910 	 * If destroysource is set, we remove the source object from the
911 	 * swap_pager internal queue now.
912 	 */
913 	if (destroysource) {
914 		if (srcobject->handle != NULL) {
915 			mtx_lock(&sw_alloc_mtx);
916 			TAILQ_REMOVE(
917 			    NOBJLIST(srcobject->handle),
918 			    srcobject,
919 			    pager_object_list
920 			);
921 			mtx_unlock(&sw_alloc_mtx);
922 		}
923 	}
924 
925 	/*
926 	 * transfer source to destination.
927 	 */
928 	for (i = 0; i < dstobject->size; ++i) {
929 		daddr_t dstaddr;
930 
931 		/*
932 		 * Locate (without changing) the swapblk on the destination,
933 		 * unless it is invalid in which case free it silently, or
934 		 * if the destination is a resident page, in which case the
935 		 * source is thrown away.
936 		 */
937 		dstaddr = swp_pager_meta_ctl(dstobject, i, 0);
938 
939 		if (dstaddr == SWAPBLK_NONE) {
940 			/*
941 			 * Destination has no swapblk and is not resident,
942 			 * copy source.
943 			 */
944 			daddr_t srcaddr;
945 
946 			srcaddr = swp_pager_meta_ctl(
947 			    srcobject,
948 			    i + offset,
949 			    SWM_POP
950 			);
951 
952 			if (srcaddr != SWAPBLK_NONE) {
953 				/*
954 				 * swp_pager_meta_build() can sleep.
955 				 */
956 				vm_object_pip_add(srcobject, 1);
957 				VM_OBJECT_WUNLOCK(srcobject);
958 				vm_object_pip_add(dstobject, 1);
959 				swp_pager_meta_build(dstobject, i, srcaddr);
960 				vm_object_pip_wakeup(dstobject);
961 				VM_OBJECT_WLOCK(srcobject);
962 				vm_object_pip_wakeup(srcobject);
963 			}
964 		} else {
965 			/*
966 			 * Destination has valid swapblk or it is represented
967 			 * by a resident page.  We destroy the sourceblock.
968 			 */
969 
970 			swp_pager_meta_ctl(srcobject, i + offset, SWM_FREE);
971 		}
972 	}
973 
974 	/*
975 	 * Free left over swap blocks in source.
976 	 *
977 	 * We have to revert the type to OBJT_DEFAULT so we do not accidently
978 	 * double-remove the object from the swap queues.
979 	 */
980 	if (destroysource) {
981 		swp_pager_meta_free_all(srcobject);
982 		/*
983 		 * Reverting the type is not necessary, the caller is going
984 		 * to destroy srcobject directly, but I'm doing it here
985 		 * for consistency since we've removed the object from its
986 		 * queues.
987 		 */
988 		srcobject->type = OBJT_DEFAULT;
989 	}
990 }
991 
992 /*
993  * SWAP_PAGER_HASPAGE() -	determine if we have good backing store for
994  *				the requested page.
995  *
996  *	We determine whether good backing store exists for the requested
997  *	page and return TRUE if it does, FALSE if it doesn't.
998  *
999  *	If TRUE, we also try to determine how much valid, contiguous backing
1000  *	store exists before and after the requested page within a reasonable
1001  *	distance.  We do not try to restrict it to the swap device stripe
1002  *	(that is handled in getpages/putpages).  It probably isn't worth
1003  *	doing here.
1004  */
1005 static boolean_t
1006 swap_pager_haspage(vm_object_t object, vm_pindex_t pindex, int *before, int *after)
1007 {
1008 	daddr_t blk0;
1009 
1010 	VM_OBJECT_ASSERT_LOCKED(object);
1011 	/*
1012 	 * do we have good backing store at the requested index ?
1013 	 */
1014 	blk0 = swp_pager_meta_ctl(object, pindex, 0);
1015 
1016 	if (blk0 == SWAPBLK_NONE) {
1017 		if (before)
1018 			*before = 0;
1019 		if (after)
1020 			*after = 0;
1021 		return (FALSE);
1022 	}
1023 
1024 	/*
1025 	 * find backwards-looking contiguous good backing store
1026 	 */
1027 	if (before != NULL) {
1028 		int i;
1029 
1030 		for (i = 1; i < (SWB_NPAGES/2); ++i) {
1031 			daddr_t blk;
1032 
1033 			if (i > pindex)
1034 				break;
1035 			blk = swp_pager_meta_ctl(object, pindex - i, 0);
1036 			if (blk != blk0 - i)
1037 				break;
1038 		}
1039 		*before = (i - 1);
1040 	}
1041 
1042 	/*
1043 	 * find forward-looking contiguous good backing store
1044 	 */
1045 	if (after != NULL) {
1046 		int i;
1047 
1048 		for (i = 1; i < (SWB_NPAGES/2); ++i) {
1049 			daddr_t blk;
1050 
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 retrieve (m, count) pages from backing store, but make
1090  *	sure we retrieve at least m[reqpage].  We try to load in as large
1091  *	a chunk surrounding m[reqpage] as is contiguous in swap and which
1092  *	belongs to the same object.
1093  *
1094  *	The code is designed for asynchronous operation and
1095  *	immediate-notification of 'reqpage' but tends not to be
1096  *	used that way.  Please do not optimize-out this algorithmic
1097  *	feature, I intend to improve on it in the future.
1098  *
1099  *	The parent has a single vm_object_pip_add() reference prior to
1100  *	calling us and we should return with the same.
1101  *
1102  *	The parent has BUSY'd the pages.  We should return with 'm'
1103  *	left busy, but the others adjusted.
1104  */
1105 static int
1106 swap_pager_getpages(vm_object_t object, vm_page_t *m, int count, int reqpage)
1107 {
1108 	struct buf *bp;
1109 	vm_page_t mreq;
1110 	int i;
1111 	int j;
1112 	daddr_t blk;
1113 
1114 	mreq = m[reqpage];
1115 
1116 	/*
1117 	 * Calculate range to retrieve.  The pages have already been assigned
1118 	 * their swapblks.  We require a *contiguous* range but we know it to
1119 	 * not span devices.   If we do not supply it, bad things
1120 	 * happen.  Note that blk, iblk & jblk can be SWAPBLK_NONE, but the
1121 	 * loops are set up such that the case(s) are handled implicitly.
1122 	 *
1123 	 * The swp_*() calls must be made with the object locked.
1124 	 */
1125 	blk = swp_pager_meta_ctl(mreq->object, mreq->pindex, 0);
1126 
1127 	for (i = reqpage - 1; i >= 0; --i) {
1128 		daddr_t iblk;
1129 
1130 		iblk = swp_pager_meta_ctl(m[i]->object, m[i]->pindex, 0);
1131 		if (blk != iblk + (reqpage - i))
1132 			break;
1133 	}
1134 	++i;
1135 
1136 	for (j = reqpage + 1; j < count; ++j) {
1137 		daddr_t jblk;
1138 
1139 		jblk = swp_pager_meta_ctl(m[j]->object, m[j]->pindex, 0);
1140 		if (blk != jblk - (j - reqpage))
1141 			break;
1142 	}
1143 
1144 	/*
1145 	 * free pages outside our collection range.   Note: we never free
1146 	 * mreq, it must remain busy throughout.
1147 	 */
1148 	if (0 < i || j < count) {
1149 		int k;
1150 
1151 		for (k = 0; k < i; ++k)
1152 			swp_pager_free_nrpage(m[k]);
1153 		for (k = j; k < count; ++k)
1154 			swp_pager_free_nrpage(m[k]);
1155 	}
1156 
1157 	/*
1158 	 * Return VM_PAGER_FAIL if we have nothing to do.  Return mreq
1159 	 * still busy, but the others unbusied.
1160 	 */
1161 	if (blk == SWAPBLK_NONE)
1162 		return (VM_PAGER_FAIL);
1163 
1164 	/*
1165 	 * Getpbuf() can sleep.
1166 	 */
1167 	VM_OBJECT_WUNLOCK(object);
1168 	/*
1169 	 * Get a swap buffer header to perform the IO
1170 	 */
1171 	bp = getpbuf(&nsw_rcount);
1172 	bp->b_flags |= B_PAGING;
1173 
1174 	bp->b_iocmd = BIO_READ;
1175 	bp->b_iodone = swp_pager_async_iodone;
1176 	bp->b_rcred = crhold(thread0.td_ucred);
1177 	bp->b_wcred = crhold(thread0.td_ucred);
1178 	bp->b_blkno = blk - (reqpage - i);
1179 	bp->b_bcount = PAGE_SIZE * (j - i);
1180 	bp->b_bufsize = PAGE_SIZE * (j - i);
1181 	bp->b_pager.pg_reqpage = reqpage - i;
1182 
1183 	VM_OBJECT_WLOCK(object);
1184 	{
1185 		int k;
1186 
1187 		for (k = i; k < j; ++k) {
1188 			bp->b_pages[k - i] = m[k];
1189 			m[k]->oflags |= VPO_SWAPINPROG;
1190 		}
1191 	}
1192 	bp->b_npages = j - i;
1193 
1194 	PCPU_INC(cnt.v_swapin);
1195 	PCPU_ADD(cnt.v_swappgsin, bp->b_npages);
1196 
1197 	/*
1198 	 * We still hold the lock on mreq, and our automatic completion routine
1199 	 * does not remove it.
1200 	 */
1201 	vm_object_pip_add(object, bp->b_npages);
1202 	VM_OBJECT_WUNLOCK(object);
1203 
1204 	/*
1205 	 * perform the I/O.  NOTE!!!  bp cannot be considered valid after
1206 	 * this point because we automatically release it on completion.
1207 	 * Instead, we look at the one page we are interested in which we
1208 	 * still hold a lock on even through the I/O completion.
1209 	 *
1210 	 * The other pages in our m[] array are also released on completion,
1211 	 * so we cannot assume they are valid anymore either.
1212 	 *
1213 	 * NOTE: b_blkno is destroyed by the call to swapdev_strategy
1214 	 */
1215 	BUF_KERNPROC(bp);
1216 	swp_pager_strategy(bp);
1217 
1218 	/*
1219 	 * wait for the page we want to complete.  VPO_SWAPINPROG is always
1220 	 * cleared on completion.  If an I/O error occurs, SWAPBLK_NONE
1221 	 * is set in the meta-data.
1222 	 */
1223 	VM_OBJECT_WLOCK(object);
1224 	while ((mreq->oflags & VPO_SWAPINPROG) != 0) {
1225 		mreq->oflags |= VPO_SWAPSLEEP;
1226 		PCPU_INC(cnt.v_intrans);
1227 		if (VM_OBJECT_SLEEP(object, &object->paging_in_progress, PSWP,
1228 		    "swread", hz * 20)) {
1229 			printf(
1230 "swap_pager: indefinite wait buffer: bufobj: %p, blkno: %jd, size: %ld\n",
1231 			    bp->b_bufobj, (intmax_t)bp->b_blkno, bp->b_bcount);
1232 		}
1233 	}
1234 
1235 	/*
1236 	 * mreq is left busied after completion, but all the other pages
1237 	 * are freed.  If we had an unrecoverable read error the page will
1238 	 * not be valid.
1239 	 */
1240 	if (mreq->valid != VM_PAGE_BITS_ALL) {
1241 		return (VM_PAGER_ERROR);
1242 	} else {
1243 		return (VM_PAGER_OK);
1244 	}
1245 
1246 	/*
1247 	 * A final note: in a low swap situation, we cannot deallocate swap
1248 	 * and mark a page dirty here because the caller is likely to mark
1249 	 * the page clean when we return, causing the page to possibly revert
1250 	 * to all-zero's later.
1251 	 */
1252 }
1253 
1254 /*
1255  * 	swap_pager_getpages_async():
1256  *
1257  *	Right now this is emulation of asynchronous operation on top of
1258  *	swap_pager_getpages().
1259  */
1260 static int
1261 swap_pager_getpages_async(vm_object_t object, vm_page_t *m, int count,
1262     int reqpage, pgo_getpages_iodone_t iodone, void *arg)
1263 {
1264 	int r, error;
1265 
1266 	r = swap_pager_getpages(object, m, count, reqpage);
1267 	VM_OBJECT_WUNLOCK(object);
1268 	switch (r) {
1269 	case VM_PAGER_OK:
1270 		error = 0;
1271 		break;
1272 	case VM_PAGER_ERROR:
1273 		error = EIO;
1274 		break;
1275 	case VM_PAGER_FAIL:
1276 		error = EINVAL;
1277 		break;
1278 	default:
1279 		panic("unhandled swap_pager_getpages() error %d", r);
1280 	}
1281 	(iodone)(arg, m, count, error);
1282 	VM_OBJECT_WLOCK(object);
1283 
1284 	return (r);
1285 }
1286 
1287 /*
1288  *	swap_pager_putpages:
1289  *
1290  *	Assign swap (if necessary) and initiate I/O on the specified pages.
1291  *
1292  *	We support both OBJT_DEFAULT and OBJT_SWAP objects.  DEFAULT objects
1293  *	are automatically converted to SWAP objects.
1294  *
1295  *	In a low memory situation we may block in VOP_STRATEGY(), but the new
1296  *	vm_page reservation system coupled with properly written VFS devices
1297  *	should ensure that no low-memory deadlock occurs.  This is an area
1298  *	which needs work.
1299  *
1300  *	The parent has N vm_object_pip_add() references prior to
1301  *	calling us and will remove references for rtvals[] that are
1302  *	not set to VM_PAGER_PEND.  We need to remove the rest on I/O
1303  *	completion.
1304  *
1305  *	The parent has soft-busy'd the pages it passes us and will unbusy
1306  *	those whos rtvals[] entry is not set to VM_PAGER_PEND on return.
1307  *	We need to unbusy the rest on I/O completion.
1308  */
1309 static void
1310 swap_pager_putpages(vm_object_t object, vm_page_t *m, int count,
1311     int flags, int *rtvals)
1312 {
1313 	int i, n;
1314 	boolean_t sync;
1315 
1316 	if (count && m[0]->object != object) {
1317 		panic("swap_pager_putpages: object mismatch %p/%p",
1318 		    object,
1319 		    m[0]->object
1320 		);
1321 	}
1322 
1323 	/*
1324 	 * Step 1
1325 	 *
1326 	 * Turn object into OBJT_SWAP
1327 	 * check for bogus sysops
1328 	 * force sync if not pageout process
1329 	 */
1330 	if (object->type != OBJT_SWAP)
1331 		swp_pager_meta_build(object, 0, SWAPBLK_NONE);
1332 	VM_OBJECT_WUNLOCK(object);
1333 
1334 	n = 0;
1335 	if (curproc != pageproc)
1336 		sync = TRUE;
1337 	else
1338 		sync = (flags & VM_PAGER_PUT_SYNC) != 0;
1339 
1340 	/*
1341 	 * Step 2
1342 	 *
1343 	 * Assign swap blocks and issue I/O.  We reallocate swap on the fly.
1344 	 * The page is left dirty until the pageout operation completes
1345 	 * successfully.
1346 	 */
1347 	for (i = 0; i < count; i += n) {
1348 		int j;
1349 		struct buf *bp;
1350 		daddr_t blk;
1351 
1352 		/*
1353 		 * Maximum I/O size is limited by a number of factors.
1354 		 */
1355 		n = min(BLIST_MAX_ALLOC, count - i);
1356 		n = min(n, nsw_cluster_max);
1357 
1358 		/*
1359 		 * Get biggest block of swap we can.  If we fail, fall
1360 		 * back and try to allocate a smaller block.  Don't go
1361 		 * overboard trying to allocate space if it would overly
1362 		 * fragment swap.
1363 		 */
1364 		while (
1365 		    (blk = swp_pager_getswapspace(n)) == SWAPBLK_NONE &&
1366 		    n > 4
1367 		) {
1368 			n >>= 1;
1369 		}
1370 		if (blk == SWAPBLK_NONE) {
1371 			for (j = 0; j < n; ++j)
1372 				rtvals[i+j] = VM_PAGER_FAIL;
1373 			continue;
1374 		}
1375 
1376 		/*
1377 		 * All I/O parameters have been satisfied, build the I/O
1378 		 * request and assign the swap space.
1379 		 */
1380 		if (sync == TRUE) {
1381 			bp = getpbuf(&nsw_wcount_sync);
1382 		} else {
1383 			bp = getpbuf(&nsw_wcount_async);
1384 			bp->b_flags = B_ASYNC;
1385 		}
1386 		bp->b_flags |= B_PAGING;
1387 		bp->b_iocmd = BIO_WRITE;
1388 
1389 		bp->b_rcred = crhold(thread0.td_ucred);
1390 		bp->b_wcred = crhold(thread0.td_ucred);
1391 		bp->b_bcount = PAGE_SIZE * n;
1392 		bp->b_bufsize = PAGE_SIZE * n;
1393 		bp->b_blkno = blk;
1394 
1395 		VM_OBJECT_WLOCK(object);
1396 		for (j = 0; j < n; ++j) {
1397 			vm_page_t mreq = m[i+j];
1398 
1399 			swp_pager_meta_build(
1400 			    mreq->object,
1401 			    mreq->pindex,
1402 			    blk + j
1403 			);
1404 			vm_page_dirty(mreq);
1405 			mreq->oflags |= VPO_SWAPINPROG;
1406 			bp->b_pages[j] = mreq;
1407 		}
1408 		VM_OBJECT_WUNLOCK(object);
1409 		bp->b_npages = n;
1410 		/*
1411 		 * Must set dirty range for NFS to work.
1412 		 */
1413 		bp->b_dirtyoff = 0;
1414 		bp->b_dirtyend = bp->b_bcount;
1415 
1416 		PCPU_INC(cnt.v_swapout);
1417 		PCPU_ADD(cnt.v_swappgsout, bp->b_npages);
1418 
1419 		/*
1420 		 * We unconditionally set rtvals[] to VM_PAGER_PEND so that we
1421 		 * can call the async completion routine at the end of a
1422 		 * synchronous I/O operation.  Otherwise, our caller would
1423 		 * perform duplicate unbusy and wakeup operations on the page
1424 		 * and object, respectively.
1425 		 */
1426 		for (j = 0; j < n; j++)
1427 			rtvals[i + j] = VM_PAGER_PEND;
1428 
1429 		/*
1430 		 * asynchronous
1431 		 *
1432 		 * NOTE: b_blkno is destroyed by the call to swapdev_strategy
1433 		 */
1434 		if (sync == FALSE) {
1435 			bp->b_iodone = swp_pager_async_iodone;
1436 			BUF_KERNPROC(bp);
1437 			swp_pager_strategy(bp);
1438 			continue;
1439 		}
1440 
1441 		/*
1442 		 * synchronous
1443 		 *
1444 		 * NOTE: b_blkno is destroyed by the call to swapdev_strategy
1445 		 */
1446 		bp->b_iodone = bdone;
1447 		swp_pager_strategy(bp);
1448 
1449 		/*
1450 		 * Wait for the sync I/O to complete.
1451 		 */
1452 		bwait(bp, PVM, "swwrt");
1453 
1454 		/*
1455 		 * Now that we are through with the bp, we can call the
1456 		 * normal async completion, which frees everything up.
1457 		 */
1458 		swp_pager_async_iodone(bp);
1459 	}
1460 	VM_OBJECT_WLOCK(object);
1461 }
1462 
1463 /*
1464  *	swp_pager_async_iodone:
1465  *
1466  *	Completion routine for asynchronous reads and writes from/to swap.
1467  *	Also called manually by synchronous code to finish up a bp.
1468  *
1469  *	This routine may not sleep.
1470  */
1471 static void
1472 swp_pager_async_iodone(struct buf *bp)
1473 {
1474 	int i;
1475 	vm_object_t object = NULL;
1476 
1477 	/*
1478 	 * report error
1479 	 */
1480 	if (bp->b_ioflags & BIO_ERROR) {
1481 		printf(
1482 		    "swap_pager: I/O error - %s failed; blkno %ld,"
1483 			"size %ld, error %d\n",
1484 		    ((bp->b_iocmd == BIO_READ) ? "pagein" : "pageout"),
1485 		    (long)bp->b_blkno,
1486 		    (long)bp->b_bcount,
1487 		    bp->b_error
1488 		);
1489 	}
1490 
1491 	/*
1492 	 * remove the mapping for kernel virtual
1493 	 */
1494 	if (buf_mapped(bp))
1495 		pmap_qremove((vm_offset_t)bp->b_data, bp->b_npages);
1496 	else
1497 		bp->b_data = bp->b_kvabase;
1498 
1499 	if (bp->b_npages) {
1500 		object = bp->b_pages[0]->object;
1501 		VM_OBJECT_WLOCK(object);
1502 	}
1503 
1504 	/*
1505 	 * cleanup pages.  If an error occurs writing to swap, we are in
1506 	 * very serious trouble.  If it happens to be a disk error, though,
1507 	 * we may be able to recover by reassigning the swap later on.  So
1508 	 * in this case we remove the m->swapblk assignment for the page
1509 	 * but do not free it in the rlist.  The errornous block(s) are thus
1510 	 * never reallocated as swap.  Redirty the page and continue.
1511 	 */
1512 	for (i = 0; i < bp->b_npages; ++i) {
1513 		vm_page_t m = bp->b_pages[i];
1514 
1515 		m->oflags &= ~VPO_SWAPINPROG;
1516 		if (m->oflags & VPO_SWAPSLEEP) {
1517 			m->oflags &= ~VPO_SWAPSLEEP;
1518 			wakeup(&object->paging_in_progress);
1519 		}
1520 
1521 		if (bp->b_ioflags & BIO_ERROR) {
1522 			/*
1523 			 * If an error occurs I'd love to throw the swapblk
1524 			 * away without freeing it back to swapspace, so it
1525 			 * can never be used again.  But I can't from an
1526 			 * interrupt.
1527 			 */
1528 			if (bp->b_iocmd == BIO_READ) {
1529 				/*
1530 				 * When reading, reqpage needs to stay
1531 				 * locked for the parent, but all other
1532 				 * pages can be freed.  We still want to
1533 				 * wakeup the parent waiting on the page,
1534 				 * though.  ( also: pg_reqpage can be -1 and
1535 				 * not match anything ).
1536 				 *
1537 				 * We have to wake specifically requested pages
1538 				 * up too because we cleared VPO_SWAPINPROG and
1539 				 * someone may be waiting for that.
1540 				 *
1541 				 * NOTE: for reads, m->dirty will probably
1542 				 * be overridden by the original caller of
1543 				 * getpages so don't play cute tricks here.
1544 				 */
1545 				m->valid = 0;
1546 				if (i != bp->b_pager.pg_reqpage)
1547 					swp_pager_free_nrpage(m);
1548 				else {
1549 					vm_page_lock(m);
1550 					vm_page_flash(m);
1551 					vm_page_unlock(m);
1552 				}
1553 				/*
1554 				 * If i == bp->b_pager.pg_reqpage, do not wake
1555 				 * the page up.  The caller needs to.
1556 				 */
1557 			} else {
1558 				/*
1559 				 * If a write error occurs, reactivate page
1560 				 * so it doesn't clog the inactive list,
1561 				 * then finish the I/O.
1562 				 */
1563 				vm_page_dirty(m);
1564 				vm_page_lock(m);
1565 				vm_page_activate(m);
1566 				vm_page_unlock(m);
1567 				vm_page_sunbusy(m);
1568 			}
1569 		} else if (bp->b_iocmd == BIO_READ) {
1570 			/*
1571 			 * NOTE: for reads, m->dirty will probably be
1572 			 * overridden by the original caller of getpages so
1573 			 * we cannot set them in order to free the underlying
1574 			 * swap in a low-swap situation.  I don't think we'd
1575 			 * want to do that anyway, but it was an optimization
1576 			 * that existed in the old swapper for a time before
1577 			 * it got ripped out due to precisely this problem.
1578 			 *
1579 			 * If not the requested page then deactivate it.
1580 			 *
1581 			 * Note that the requested page, reqpage, is left
1582 			 * busied, but we still have to wake it up.  The
1583 			 * other pages are released (unbusied) by
1584 			 * vm_page_xunbusy().
1585 			 */
1586 			KASSERT(!pmap_page_is_mapped(m),
1587 			    ("swp_pager_async_iodone: page %p is mapped", m));
1588 			m->valid = VM_PAGE_BITS_ALL;
1589 			KASSERT(m->dirty == 0,
1590 			    ("swp_pager_async_iodone: page %p is dirty", m));
1591 
1592 			/*
1593 			 * We have to wake specifically requested pages
1594 			 * up too because we cleared VPO_SWAPINPROG and
1595 			 * could be waiting for it in getpages.  However,
1596 			 * be sure to not unbusy getpages specifically
1597 			 * requested page - getpages expects it to be
1598 			 * left busy.
1599 			 */
1600 			if (i != bp->b_pager.pg_reqpage) {
1601 				vm_page_lock(m);
1602 				vm_page_deactivate(m);
1603 				vm_page_unlock(m);
1604 				vm_page_xunbusy(m);
1605 			} else {
1606 				vm_page_lock(m);
1607 				vm_page_flash(m);
1608 				vm_page_unlock(m);
1609 			}
1610 		} else {
1611 			/*
1612 			 * For write success, clear the dirty
1613 			 * status, then finish the I/O ( which decrements the
1614 			 * busy count and possibly wakes waiter's up ).
1615 			 */
1616 			KASSERT(!pmap_page_is_write_mapped(m),
1617 			    ("swp_pager_async_iodone: page %p is not write"
1618 			    " protected", m));
1619 			vm_page_undirty(m);
1620 			vm_page_sunbusy(m);
1621 			if (vm_page_count_severe()) {
1622 				vm_page_lock(m);
1623 				vm_page_try_to_cache(m);
1624 				vm_page_unlock(m);
1625 			}
1626 		}
1627 	}
1628 
1629 	/*
1630 	 * adjust pip.  NOTE: the original parent may still have its own
1631 	 * pip refs on the object.
1632 	 */
1633 	if (object != NULL) {
1634 		vm_object_pip_wakeupn(object, bp->b_npages);
1635 		VM_OBJECT_WUNLOCK(object);
1636 	}
1637 
1638 	/*
1639 	 * swapdev_strategy() manually sets b_vp and b_bufobj before calling
1640 	 * bstrategy(). Set them back to NULL now we're done with it, or we'll
1641 	 * trigger a KASSERT in relpbuf().
1642 	 */
1643 	if (bp->b_vp) {
1644 		    bp->b_vp = NULL;
1645 		    bp->b_bufobj = NULL;
1646 	}
1647 	/*
1648 	 * release the physical I/O buffer
1649 	 */
1650 	relpbuf(
1651 	    bp,
1652 	    ((bp->b_iocmd == BIO_READ) ? &nsw_rcount :
1653 		((bp->b_flags & B_ASYNC) ?
1654 		    &nsw_wcount_async :
1655 		    &nsw_wcount_sync
1656 		)
1657 	    )
1658 	);
1659 }
1660 
1661 /*
1662  *	swap_pager_isswapped:
1663  *
1664  *	Return 1 if at least one page in the given object is paged
1665  *	out to the given swap device.
1666  *
1667  *	This routine may not sleep.
1668  */
1669 int
1670 swap_pager_isswapped(vm_object_t object, struct swdevt *sp)
1671 {
1672 	daddr_t index = 0;
1673 	int bcount;
1674 	int i;
1675 
1676 	VM_OBJECT_ASSERT_WLOCKED(object);
1677 	if (object->type != OBJT_SWAP)
1678 		return (0);
1679 
1680 	mtx_lock(&swhash_mtx);
1681 	for (bcount = 0; bcount < object->un_pager.swp.swp_bcount; bcount++) {
1682 		struct swblock *swap;
1683 
1684 		if ((swap = *swp_pager_hash(object, index)) != NULL) {
1685 			for (i = 0; i < SWAP_META_PAGES; ++i) {
1686 				if (swp_pager_isondev(swap->swb_pages[i], sp)) {
1687 					mtx_unlock(&swhash_mtx);
1688 					return (1);
1689 				}
1690 			}
1691 		}
1692 		index += SWAP_META_PAGES;
1693 	}
1694 	mtx_unlock(&swhash_mtx);
1695 	return (0);
1696 }
1697 
1698 /*
1699  * SWP_PAGER_FORCE_PAGEIN() - force a swap block to be paged in
1700  *
1701  *	This routine dissociates the page at the given index within a
1702  *	swap block from its backing store, paging it in if necessary.
1703  *	If the page is paged in, it is placed in the inactive queue,
1704  *	since it had its backing store ripped out from under it.
1705  *	We also attempt to swap in all other pages in the swap block,
1706  *	we only guarantee that the one at the specified index is
1707  *	paged in.
1708  *
1709  *	XXX - The code to page the whole block in doesn't work, so we
1710  *	      revert to the one-by-one behavior for now.  Sigh.
1711  */
1712 static inline void
1713 swp_pager_force_pagein(vm_object_t object, vm_pindex_t pindex)
1714 {
1715 	vm_page_t m;
1716 
1717 	vm_object_pip_add(object, 1);
1718 	m = vm_page_grab(object, pindex, VM_ALLOC_NORMAL);
1719 	if (m->valid == VM_PAGE_BITS_ALL) {
1720 		vm_object_pip_wakeup(object);
1721 		vm_page_dirty(m);
1722 		vm_page_lock(m);
1723 		vm_page_activate(m);
1724 		vm_page_unlock(m);
1725 		vm_page_xunbusy(m);
1726 		vm_pager_page_unswapped(m);
1727 		return;
1728 	}
1729 
1730 	if (swap_pager_getpages(object, &m, 1, 0) != VM_PAGER_OK)
1731 		panic("swap_pager_force_pagein: read from swap failed");/*XXX*/
1732 	vm_object_pip_wakeup(object);
1733 	vm_page_dirty(m);
1734 	vm_page_lock(m);
1735 	vm_page_deactivate(m);
1736 	vm_page_unlock(m);
1737 	vm_page_xunbusy(m);
1738 	vm_pager_page_unswapped(m);
1739 }
1740 
1741 /*
1742  *	swap_pager_swapoff:
1743  *
1744  *	Page in all of the pages that have been paged out to the
1745  *	given device.  The corresponding blocks in the bitmap must be
1746  *	marked as allocated and the device must be flagged SW_CLOSING.
1747  *	There may be no processes swapped out to the device.
1748  *
1749  *	This routine may block.
1750  */
1751 static void
1752 swap_pager_swapoff(struct swdevt *sp)
1753 {
1754 	struct swblock *swap;
1755 	int i, j, retries;
1756 
1757 	GIANT_REQUIRED;
1758 
1759 	retries = 0;
1760 full_rescan:
1761 	mtx_lock(&swhash_mtx);
1762 	for (i = 0; i <= swhash_mask; i++) { /* '<=' is correct here */
1763 restart:
1764 		for (swap = swhash[i]; swap != NULL; swap = swap->swb_hnext) {
1765 			vm_object_t object = swap->swb_object;
1766 			vm_pindex_t pindex = swap->swb_index;
1767 			for (j = 0; j < SWAP_META_PAGES; ++j) {
1768 				if (swp_pager_isondev(swap->swb_pages[j], sp)) {
1769 					/* avoid deadlock */
1770 					if (!VM_OBJECT_TRYWLOCK(object)) {
1771 						break;
1772 					} else {
1773 						mtx_unlock(&swhash_mtx);
1774 						swp_pager_force_pagein(object,
1775 						    pindex + j);
1776 						VM_OBJECT_WUNLOCK(object);
1777 						mtx_lock(&swhash_mtx);
1778 						goto restart;
1779 					}
1780 				}
1781 			}
1782 		}
1783 	}
1784 	mtx_unlock(&swhash_mtx);
1785 	if (sp->sw_used) {
1786 		/*
1787 		 * Objects may be locked or paging to the device being
1788 		 * removed, so we will miss their pages and need to
1789 		 * make another pass.  We have marked this device as
1790 		 * SW_CLOSING, so the activity should finish soon.
1791 		 */
1792 		retries++;
1793 		if (retries > 100) {
1794 			panic("swapoff: failed to locate %d swap blocks",
1795 			    sp->sw_used);
1796 		}
1797 		pause("swpoff", hz / 20);
1798 		goto full_rescan;
1799 	}
1800 }
1801 
1802 /************************************************************************
1803  *				SWAP META DATA 				*
1804  ************************************************************************
1805  *
1806  *	These routines manipulate the swap metadata stored in the
1807  *	OBJT_SWAP object.
1808  *
1809  *	Swap metadata is implemented with a global hash and not directly
1810  *	linked into the object.  Instead the object simply contains
1811  *	appropriate tracking counters.
1812  */
1813 
1814 /*
1815  * SWP_PAGER_META_BUILD() -	add swap block to swap meta data for object
1816  *
1817  *	We first convert the object to a swap object if it is a default
1818  *	object.
1819  *
1820  *	The specified swapblk is added to the object's swap metadata.  If
1821  *	the swapblk is not valid, it is freed instead.  Any previously
1822  *	assigned swapblk is freed.
1823  */
1824 static void
1825 swp_pager_meta_build(vm_object_t object, vm_pindex_t pindex, daddr_t swapblk)
1826 {
1827 	static volatile int exhausted;
1828 	struct swblock *swap;
1829 	struct swblock **pswap;
1830 	int idx;
1831 
1832 	VM_OBJECT_ASSERT_WLOCKED(object);
1833 	/*
1834 	 * Convert default object to swap object if necessary
1835 	 */
1836 	if (object->type != OBJT_SWAP) {
1837 		object->type = OBJT_SWAP;
1838 		object->un_pager.swp.swp_bcount = 0;
1839 
1840 		if (object->handle != NULL) {
1841 			mtx_lock(&sw_alloc_mtx);
1842 			TAILQ_INSERT_TAIL(
1843 			    NOBJLIST(object->handle),
1844 			    object,
1845 			    pager_object_list
1846 			);
1847 			mtx_unlock(&sw_alloc_mtx);
1848 		}
1849 	}
1850 
1851 	/*
1852 	 * Locate hash entry.  If not found create, but if we aren't adding
1853 	 * anything just return.  If we run out of space in the map we wait
1854 	 * and, since the hash table may have changed, retry.
1855 	 */
1856 retry:
1857 	mtx_lock(&swhash_mtx);
1858 	pswap = swp_pager_hash(object, pindex);
1859 
1860 	if ((swap = *pswap) == NULL) {
1861 		int i;
1862 
1863 		if (swapblk == SWAPBLK_NONE)
1864 			goto done;
1865 
1866 		swap = *pswap = uma_zalloc(swap_zone, M_NOWAIT |
1867 		    (curproc == pageproc ? M_USE_RESERVE : 0));
1868 		if (swap == NULL) {
1869 			mtx_unlock(&swhash_mtx);
1870 			VM_OBJECT_WUNLOCK(object);
1871 			if (uma_zone_exhausted(swap_zone)) {
1872 				if (atomic_cmpset_int(&exhausted, 0, 1))
1873 					printf("swap zone exhausted, "
1874 					    "increase kern.maxswzone\n");
1875 				vm_pageout_oom(VM_OOM_SWAPZ);
1876 				pause("swzonex", 10);
1877 			} else
1878 				VM_WAIT;
1879 			VM_OBJECT_WLOCK(object);
1880 			goto retry;
1881 		}
1882 
1883 		if (atomic_cmpset_int(&exhausted, 1, 0))
1884 			printf("swap zone ok\n");
1885 
1886 		swap->swb_hnext = NULL;
1887 		swap->swb_object = object;
1888 		swap->swb_index = pindex & ~(vm_pindex_t)SWAP_META_MASK;
1889 		swap->swb_count = 0;
1890 
1891 		++object->un_pager.swp.swp_bcount;
1892 
1893 		for (i = 0; i < SWAP_META_PAGES; ++i)
1894 			swap->swb_pages[i] = SWAPBLK_NONE;
1895 	}
1896 
1897 	/*
1898 	 * Delete prior contents of metadata
1899 	 */
1900 	idx = pindex & SWAP_META_MASK;
1901 
1902 	if (swap->swb_pages[idx] != SWAPBLK_NONE) {
1903 		swp_pager_freeswapspace(swap->swb_pages[idx], 1);
1904 		--swap->swb_count;
1905 	}
1906 
1907 	/*
1908 	 * Enter block into metadata
1909 	 */
1910 	swap->swb_pages[idx] = swapblk;
1911 	if (swapblk != SWAPBLK_NONE)
1912 		++swap->swb_count;
1913 done:
1914 	mtx_unlock(&swhash_mtx);
1915 }
1916 
1917 /*
1918  * SWP_PAGER_META_FREE() - free a range of blocks in the object's swap metadata
1919  *
1920  *	The requested range of blocks is freed, with any associated swap
1921  *	returned to the swap bitmap.
1922  *
1923  *	This routine will free swap metadata structures as they are cleaned
1924  *	out.  This routine does *NOT* operate on swap metadata associated
1925  *	with resident pages.
1926  */
1927 static void
1928 swp_pager_meta_free(vm_object_t object, vm_pindex_t index, daddr_t count)
1929 {
1930 
1931 	VM_OBJECT_ASSERT_LOCKED(object);
1932 	if (object->type != OBJT_SWAP)
1933 		return;
1934 
1935 	while (count > 0) {
1936 		struct swblock **pswap;
1937 		struct swblock *swap;
1938 
1939 		mtx_lock(&swhash_mtx);
1940 		pswap = swp_pager_hash(object, index);
1941 
1942 		if ((swap = *pswap) != NULL) {
1943 			daddr_t v = swap->swb_pages[index & SWAP_META_MASK];
1944 
1945 			if (v != SWAPBLK_NONE) {
1946 				swp_pager_freeswapspace(v, 1);
1947 				swap->swb_pages[index & SWAP_META_MASK] =
1948 					SWAPBLK_NONE;
1949 				if (--swap->swb_count == 0) {
1950 					*pswap = swap->swb_hnext;
1951 					uma_zfree(swap_zone, swap);
1952 					--object->un_pager.swp.swp_bcount;
1953 				}
1954 			}
1955 			--count;
1956 			++index;
1957 		} else {
1958 			int n = SWAP_META_PAGES - (index & SWAP_META_MASK);
1959 			count -= n;
1960 			index += n;
1961 		}
1962 		mtx_unlock(&swhash_mtx);
1963 	}
1964 }
1965 
1966 /*
1967  * SWP_PAGER_META_FREE_ALL() - destroy all swap metadata associated with object
1968  *
1969  *	This routine locates and destroys all swap metadata associated with
1970  *	an object.
1971  */
1972 static void
1973 swp_pager_meta_free_all(vm_object_t object)
1974 {
1975 	daddr_t index = 0;
1976 
1977 	VM_OBJECT_ASSERT_WLOCKED(object);
1978 	if (object->type != OBJT_SWAP)
1979 		return;
1980 
1981 	while (object->un_pager.swp.swp_bcount) {
1982 		struct swblock **pswap;
1983 		struct swblock *swap;
1984 
1985 		mtx_lock(&swhash_mtx);
1986 		pswap = swp_pager_hash(object, index);
1987 		if ((swap = *pswap) != NULL) {
1988 			int i;
1989 
1990 			for (i = 0; i < SWAP_META_PAGES; ++i) {
1991 				daddr_t v = swap->swb_pages[i];
1992 				if (v != SWAPBLK_NONE) {
1993 					--swap->swb_count;
1994 					swp_pager_freeswapspace(v, 1);
1995 				}
1996 			}
1997 			if (swap->swb_count != 0)
1998 				panic("swap_pager_meta_free_all: swb_count != 0");
1999 			*pswap = swap->swb_hnext;
2000 			uma_zfree(swap_zone, swap);
2001 			--object->un_pager.swp.swp_bcount;
2002 		}
2003 		mtx_unlock(&swhash_mtx);
2004 		index += SWAP_META_PAGES;
2005 	}
2006 }
2007 
2008 /*
2009  * SWP_PAGER_METACTL() -  misc control of swap and vm_page_t meta data.
2010  *
2011  *	This routine is capable of looking up, popping, or freeing
2012  *	swapblk assignments in the swap meta data or in the vm_page_t.
2013  *	The routine typically returns the swapblk being looked-up, or popped,
2014  *	or SWAPBLK_NONE if the block was freed, or SWAPBLK_NONE if the block
2015  *	was invalid.  This routine will automatically free any invalid
2016  *	meta-data swapblks.
2017  *
2018  *	It is not possible to store invalid swapblks in the swap meta data
2019  *	(other then a literal 'SWAPBLK_NONE'), so we don't bother checking.
2020  *
2021  *	When acting on a busy resident page and paging is in progress, we
2022  *	have to wait until paging is complete but otherwise can act on the
2023  *	busy page.
2024  *
2025  *	SWM_FREE	remove and free swap block from metadata
2026  *	SWM_POP		remove from meta data but do not free.. pop it out
2027  */
2028 static daddr_t
2029 swp_pager_meta_ctl(vm_object_t object, vm_pindex_t pindex, int flags)
2030 {
2031 	struct swblock **pswap;
2032 	struct swblock *swap;
2033 	daddr_t r1;
2034 	int idx;
2035 
2036 	VM_OBJECT_ASSERT_LOCKED(object);
2037 	/*
2038 	 * The meta data only exists of the object is OBJT_SWAP
2039 	 * and even then might not be allocated yet.
2040 	 */
2041 	if (object->type != OBJT_SWAP)
2042 		return (SWAPBLK_NONE);
2043 
2044 	r1 = SWAPBLK_NONE;
2045 	mtx_lock(&swhash_mtx);
2046 	pswap = swp_pager_hash(object, pindex);
2047 
2048 	if ((swap = *pswap) != NULL) {
2049 		idx = pindex & SWAP_META_MASK;
2050 		r1 = swap->swb_pages[idx];
2051 
2052 		if (r1 != SWAPBLK_NONE) {
2053 			if (flags & SWM_FREE) {
2054 				swp_pager_freeswapspace(r1, 1);
2055 				r1 = SWAPBLK_NONE;
2056 			}
2057 			if (flags & (SWM_FREE|SWM_POP)) {
2058 				swap->swb_pages[idx] = SWAPBLK_NONE;
2059 				if (--swap->swb_count == 0) {
2060 					*pswap = swap->swb_hnext;
2061 					uma_zfree(swap_zone, swap);
2062 					--object->un_pager.swp.swp_bcount;
2063 				}
2064 			}
2065 		}
2066 	}
2067 	mtx_unlock(&swhash_mtx);
2068 	return (r1);
2069 }
2070 
2071 /*
2072  * System call swapon(name) enables swapping on device name,
2073  * which must be in the swdevsw.  Return EBUSY
2074  * if already swapping on this device.
2075  */
2076 #ifndef _SYS_SYSPROTO_H_
2077 struct swapon_args {
2078 	char *name;
2079 };
2080 #endif
2081 
2082 /*
2083  * MPSAFE
2084  */
2085 /* ARGSUSED */
2086 int
2087 sys_swapon(struct thread *td, struct swapon_args *uap)
2088 {
2089 	struct vattr attr;
2090 	struct vnode *vp;
2091 	struct nameidata nd;
2092 	int error;
2093 
2094 	error = priv_check(td, PRIV_SWAPON);
2095 	if (error)
2096 		return (error);
2097 
2098 	mtx_lock(&Giant);
2099 	while (swdev_syscall_active)
2100 	    tsleep(&swdev_syscall_active, PUSER - 1, "swpon", 0);
2101 	swdev_syscall_active = 1;
2102 
2103 	/*
2104 	 * Swap metadata may not fit in the KVM if we have physical
2105 	 * memory of >1GB.
2106 	 */
2107 	if (swap_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(td, 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 	swdev_syscall_active = 0;
2137 	wakeup_one(&swdev_syscall_active);
2138 	mtx_unlock(&Giant);
2139 	return (error);
2140 }
2141 
2142 /*
2143  * Check that the total amount of swap currently configured does not
2144  * exceed half the theoretical maximum.  If it does, print a warning
2145  * message and return -1; otherwise, return 0.
2146  */
2147 static int
2148 swapon_check_swzone(unsigned long npages)
2149 {
2150 	unsigned long maxpages;
2151 
2152 	/* absolute maximum we can handle assuming 100% efficiency */
2153 	maxpages = uma_zone_get_max(swap_zone) * SWAP_META_PAGES;
2154 
2155 	/* recommend using no more than half that amount */
2156 	if (npages > maxpages / 2) {
2157 		printf("warning: total configured swap (%lu pages) "
2158 		    "exceeds maximum recommended amount (%lu pages).\n",
2159 		    npages, maxpages / 2);
2160 		printf("warning: increase kern.maxswzone "
2161 		    "or reduce amount of swap.\n");
2162 		return (-1);
2163 	}
2164 	return (0);
2165 }
2166 
2167 static void
2168 swaponsomething(struct vnode *vp, void *id, u_long nblks,
2169     sw_strategy_t *strategy, sw_close_t *close, dev_t dev, int flags)
2170 {
2171 	struct swdevt *sp, *tsp;
2172 	swblk_t dvbase;
2173 	u_long mblocks;
2174 
2175 	/*
2176 	 * nblks is in DEV_BSIZE'd chunks, convert to PAGE_SIZE'd chunks.
2177 	 * First chop nblks off to page-align it, then convert.
2178 	 *
2179 	 * sw->sw_nblks is in page-sized chunks now too.
2180 	 */
2181 	nblks &= ~(ctodb(1) - 1);
2182 	nblks = dbtoc(nblks);
2183 
2184 	/*
2185 	 * If we go beyond this, we get overflows in the radix
2186 	 * tree bitmap code.
2187 	 */
2188 	mblocks = 0x40000000 / BLIST_META_RADIX;
2189 	if (nblks > mblocks) {
2190 		printf(
2191     "WARNING: reducing swap size to maximum of %luMB per unit\n",
2192 		    mblocks / 1024 / 1024 * PAGE_SIZE);
2193 		nblks = mblocks;
2194 	}
2195 
2196 	sp = malloc(sizeof *sp, M_VMPGDATA, M_WAITOK | M_ZERO);
2197 	sp->sw_vp = vp;
2198 	sp->sw_id = id;
2199 	sp->sw_dev = dev;
2200 	sp->sw_flags = 0;
2201 	sp->sw_nblks = nblks;
2202 	sp->sw_used = 0;
2203 	sp->sw_strategy = strategy;
2204 	sp->sw_close = close;
2205 	sp->sw_flags = flags;
2206 
2207 	sp->sw_blist = blist_create(nblks, M_WAITOK);
2208 	/*
2209 	 * Do not free the first two block in order to avoid overwriting
2210 	 * any bsd label at the front of the partition
2211 	 */
2212 	blist_free(sp->sw_blist, 2, nblks - 2);
2213 
2214 	dvbase = 0;
2215 	mtx_lock(&sw_dev_mtx);
2216 	TAILQ_FOREACH(tsp, &swtailq, sw_list) {
2217 		if (tsp->sw_end >= dvbase) {
2218 			/*
2219 			 * We put one uncovered page between the devices
2220 			 * in order to definitively prevent any cross-device
2221 			 * I/O requests
2222 			 */
2223 			dvbase = tsp->sw_end + 1;
2224 		}
2225 	}
2226 	sp->sw_first = dvbase;
2227 	sp->sw_end = dvbase + nblks;
2228 	TAILQ_INSERT_TAIL(&swtailq, sp, sw_list);
2229 	nswapdev++;
2230 	swap_pager_avail += nblks;
2231 	swap_total += (vm_ooffset_t)nblks * PAGE_SIZE;
2232 	swapon_check_swzone(swap_total / PAGE_SIZE);
2233 	swp_sizecheck();
2234 	mtx_unlock(&sw_dev_mtx);
2235 }
2236 
2237 /*
2238  * SYSCALL: swapoff(devname)
2239  *
2240  * Disable swapping on the given device.
2241  *
2242  * XXX: Badly designed system call: it should use a device index
2243  * rather than filename as specification.  We keep sw_vp around
2244  * only to make this work.
2245  */
2246 #ifndef _SYS_SYSPROTO_H_
2247 struct swapoff_args {
2248 	char *name;
2249 };
2250 #endif
2251 
2252 /*
2253  * MPSAFE
2254  */
2255 /* ARGSUSED */
2256 int
2257 sys_swapoff(struct thread *td, struct swapoff_args *uap)
2258 {
2259 	struct vnode *vp;
2260 	struct nameidata nd;
2261 	struct swdevt *sp;
2262 	int error;
2263 
2264 	error = priv_check(td, PRIV_SWAPOFF);
2265 	if (error)
2266 		return (error);
2267 
2268 	mtx_lock(&Giant);
2269 	while (swdev_syscall_active)
2270 	    tsleep(&swdev_syscall_active, PUSER - 1, "swpoff", 0);
2271 	swdev_syscall_active = 1;
2272 
2273 	NDINIT(&nd, LOOKUP, FOLLOW | AUDITVNODE1, UIO_USERSPACE, uap->name,
2274 	    td);
2275 	error = namei(&nd);
2276 	if (error)
2277 		goto done;
2278 	NDFREE(&nd, NDF_ONLY_PNBUF);
2279 	vp = nd.ni_vp;
2280 
2281 	mtx_lock(&sw_dev_mtx);
2282 	TAILQ_FOREACH(sp, &swtailq, sw_list) {
2283 		if (sp->sw_vp == vp)
2284 			break;
2285 	}
2286 	mtx_unlock(&sw_dev_mtx);
2287 	if (sp == NULL) {
2288 		error = EINVAL;
2289 		goto done;
2290 	}
2291 	error = swapoff_one(sp, td->td_ucred);
2292 done:
2293 	swdev_syscall_active = 0;
2294 	wakeup_one(&swdev_syscall_active);
2295 	mtx_unlock(&Giant);
2296 	return (error);
2297 }
2298 
2299 static int
2300 swapoff_one(struct swdevt *sp, struct ucred *cred)
2301 {
2302 	u_long nblks, dvbase;
2303 #ifdef MAC
2304 	int error;
2305 #endif
2306 
2307 	mtx_assert(&Giant, MA_OWNED);
2308 #ifdef MAC
2309 	(void) vn_lock(sp->sw_vp, LK_EXCLUSIVE | LK_RETRY);
2310 	error = mac_system_check_swapoff(cred, sp->sw_vp);
2311 	(void) VOP_UNLOCK(sp->sw_vp, 0);
2312 	if (error != 0)
2313 		return (error);
2314 #endif
2315 	nblks = sp->sw_nblks;
2316 
2317 	/*
2318 	 * We can turn off this swap device safely only if the
2319 	 * available virtual memory in the system will fit the amount
2320 	 * of data we will have to page back in, plus an epsilon so
2321 	 * the system doesn't become critically low on swap space.
2322 	 */
2323 	if (vm_cnt.v_free_count + vm_cnt.v_cache_count + swap_pager_avail <
2324 	    nblks + nswap_lowat) {
2325 		return (ENOMEM);
2326 	}
2327 
2328 	/*
2329 	 * Prevent further allocations on this device.
2330 	 */
2331 	mtx_lock(&sw_dev_mtx);
2332 	sp->sw_flags |= SW_CLOSING;
2333 	for (dvbase = 0; dvbase < sp->sw_end; dvbase += dmmax) {
2334 		swap_pager_avail -= blist_fill(sp->sw_blist,
2335 		     dvbase, dmmax);
2336 	}
2337 	swap_total -= (vm_ooffset_t)nblks * PAGE_SIZE;
2338 	mtx_unlock(&sw_dev_mtx);
2339 
2340 	/*
2341 	 * Page in the contents of the device and close it.
2342 	 */
2343 	swap_pager_swapoff(sp);
2344 
2345 	sp->sw_close(curthread, sp);
2346 	mtx_lock(&sw_dev_mtx);
2347 	sp->sw_id = NULL;
2348 	TAILQ_REMOVE(&swtailq, sp, sw_list);
2349 	nswapdev--;
2350 	if (nswapdev == 0) {
2351 		swap_pager_full = 2;
2352 		swap_pager_almost_full = 1;
2353 	}
2354 	if (swdevhd == sp)
2355 		swdevhd = NULL;
2356 	mtx_unlock(&sw_dev_mtx);
2357 	blist_destroy(sp->sw_blist);
2358 	free(sp, M_VMPGDATA);
2359 	return (0);
2360 }
2361 
2362 void
2363 swapoff_all(void)
2364 {
2365 	struct swdevt *sp, *spt;
2366 	const char *devname;
2367 	int error;
2368 
2369 	mtx_lock(&Giant);
2370 	while (swdev_syscall_active)
2371 		tsleep(&swdev_syscall_active, PUSER - 1, "swpoff", 0);
2372 	swdev_syscall_active = 1;
2373 
2374 	mtx_lock(&sw_dev_mtx);
2375 	TAILQ_FOREACH_SAFE(sp, &swtailq, sw_list, spt) {
2376 		mtx_unlock(&sw_dev_mtx);
2377 		if (vn_isdisk(sp->sw_vp, NULL))
2378 			devname = devtoname(sp->sw_vp->v_rdev);
2379 		else
2380 			devname = "[file]";
2381 		error = swapoff_one(sp, thread0.td_ucred);
2382 		if (error != 0) {
2383 			printf("Cannot remove swap device %s (error=%d), "
2384 			    "skipping.\n", devname, error);
2385 		} else if (bootverbose) {
2386 			printf("Swap device %s removed.\n", devname);
2387 		}
2388 		mtx_lock(&sw_dev_mtx);
2389 	}
2390 	mtx_unlock(&sw_dev_mtx);
2391 
2392 	swdev_syscall_active = 0;
2393 	wakeup_one(&swdev_syscall_active);
2394 	mtx_unlock(&Giant);
2395 }
2396 
2397 void
2398 swap_pager_status(int *total, int *used)
2399 {
2400 	struct swdevt *sp;
2401 
2402 	*total = 0;
2403 	*used = 0;
2404 	mtx_lock(&sw_dev_mtx);
2405 	TAILQ_FOREACH(sp, &swtailq, sw_list) {
2406 		*total += sp->sw_nblks;
2407 		*used += sp->sw_used;
2408 	}
2409 	mtx_unlock(&sw_dev_mtx);
2410 }
2411 
2412 int
2413 swap_dev_info(int name, struct xswdev *xs, char *devname, size_t len)
2414 {
2415 	struct swdevt *sp;
2416 	const char *tmp_devname;
2417 	int error, n;
2418 
2419 	n = 0;
2420 	error = ENOENT;
2421 	mtx_lock(&sw_dev_mtx);
2422 	TAILQ_FOREACH(sp, &swtailq, sw_list) {
2423 		if (n != name) {
2424 			n++;
2425 			continue;
2426 		}
2427 		xs->xsw_version = XSWDEV_VERSION;
2428 		xs->xsw_dev = sp->sw_dev;
2429 		xs->xsw_flags = sp->sw_flags;
2430 		xs->xsw_nblks = sp->sw_nblks;
2431 		xs->xsw_used = sp->sw_used;
2432 		if (devname != NULL) {
2433 			if (vn_isdisk(sp->sw_vp, NULL))
2434 				tmp_devname = devtoname(sp->sw_vp->v_rdev);
2435 			else
2436 				tmp_devname = "[file]";
2437 			strncpy(devname, tmp_devname, len);
2438 		}
2439 		error = 0;
2440 		break;
2441 	}
2442 	mtx_unlock(&sw_dev_mtx);
2443 	return (error);
2444 }
2445 
2446 static int
2447 sysctl_vm_swap_info(SYSCTL_HANDLER_ARGS)
2448 {
2449 	struct xswdev xs;
2450 	int error;
2451 
2452 	if (arg2 != 1)			/* name length */
2453 		return (EINVAL);
2454 	error = swap_dev_info(*(int *)arg1, &xs, NULL, 0);
2455 	if (error != 0)
2456 		return (error);
2457 	error = SYSCTL_OUT(req, &xs, sizeof(xs));
2458 	return (error);
2459 }
2460 
2461 SYSCTL_INT(_vm, OID_AUTO, nswapdev, CTLFLAG_RD, &nswapdev, 0,
2462     "Number of swap devices");
2463 SYSCTL_NODE(_vm, OID_AUTO, swap_info, CTLFLAG_RD, sysctl_vm_swap_info,
2464     "Swap statistics by device");
2465 
2466 /*
2467  * vmspace_swap_count() - count the approximate swap usage in pages for a
2468  *			  vmspace.
2469  *
2470  *	The map must be locked.
2471  *
2472  *	Swap usage is determined by taking the proportional swap used by
2473  *	VM objects backing the VM map.  To make up for fractional losses,
2474  *	if the VM object has any swap use at all the associated map entries
2475  *	count for at least 1 swap page.
2476  */
2477 long
2478 vmspace_swap_count(struct vmspace *vmspace)
2479 {
2480 	vm_map_t map;
2481 	vm_map_entry_t cur;
2482 	vm_object_t object;
2483 	long count, n;
2484 
2485 	map = &vmspace->vm_map;
2486 	count = 0;
2487 
2488 	for (cur = map->header.next; cur != &map->header; cur = cur->next) {
2489 		if ((cur->eflags & MAP_ENTRY_IS_SUB_MAP) == 0 &&
2490 		    (object = cur->object.vm_object) != NULL) {
2491 			VM_OBJECT_WLOCK(object);
2492 			if (object->type == OBJT_SWAP &&
2493 			    object->un_pager.swp.swp_bcount != 0) {
2494 				n = (cur->end - cur->start) / PAGE_SIZE;
2495 				count += object->un_pager.swp.swp_bcount *
2496 				    SWAP_META_PAGES * n / object->size + 1;
2497 			}
2498 			VM_OBJECT_WUNLOCK(object);
2499 		}
2500 	}
2501 	return (count);
2502 }
2503 
2504 /*
2505  * GEOM backend
2506  *
2507  * Swapping onto disk devices.
2508  *
2509  */
2510 
2511 static g_orphan_t swapgeom_orphan;
2512 
2513 static struct g_class g_swap_class = {
2514 	.name = "SWAP",
2515 	.version = G_VERSION,
2516 	.orphan = swapgeom_orphan,
2517 };
2518 
2519 DECLARE_GEOM_CLASS(g_swap_class, g_class);
2520 
2521 
2522 static void
2523 swapgeom_close_ev(void *arg, int flags)
2524 {
2525 	struct g_consumer *cp;
2526 
2527 	cp = arg;
2528 	g_access(cp, -1, -1, 0);
2529 	g_detach(cp);
2530 	g_destroy_consumer(cp);
2531 }
2532 
2533 /*
2534  * Add a reference to the g_consumer for an inflight transaction.
2535  */
2536 static void
2537 swapgeom_acquire(struct g_consumer *cp)
2538 {
2539 
2540 	mtx_assert(&sw_dev_mtx, MA_OWNED);
2541 	cp->index++;
2542 }
2543 
2544 /*
2545  * Remove a reference from the g_consumer. Post a close event if
2546  * all referneces go away.
2547  */
2548 static void
2549 swapgeom_release(struct g_consumer *cp, struct swdevt *sp)
2550 {
2551 
2552 	mtx_assert(&sw_dev_mtx, MA_OWNED);
2553 	cp->index--;
2554 	if (cp->index == 0) {
2555 		if (g_post_event(swapgeom_close_ev, cp, M_NOWAIT, NULL) == 0)
2556 			sp->sw_id = NULL;
2557 	}
2558 }
2559 
2560 static void
2561 swapgeom_done(struct bio *bp2)
2562 {
2563 	struct swdevt *sp;
2564 	struct buf *bp;
2565 	struct g_consumer *cp;
2566 
2567 	bp = bp2->bio_caller2;
2568 	cp = bp2->bio_from;
2569 	bp->b_ioflags = bp2->bio_flags;
2570 	if (bp2->bio_error)
2571 		bp->b_ioflags |= BIO_ERROR;
2572 	bp->b_resid = bp->b_bcount - bp2->bio_completed;
2573 	bp->b_error = bp2->bio_error;
2574 	bufdone(bp);
2575 	sp = bp2->bio_caller1;
2576 	mtx_lock(&sw_dev_mtx);
2577 	swapgeom_release(cp, sp);
2578 	mtx_unlock(&sw_dev_mtx);
2579 	g_destroy_bio(bp2);
2580 }
2581 
2582 static void
2583 swapgeom_strategy(struct buf *bp, struct swdevt *sp)
2584 {
2585 	struct bio *bio;
2586 	struct g_consumer *cp;
2587 
2588 	mtx_lock(&sw_dev_mtx);
2589 	cp = sp->sw_id;
2590 	if (cp == NULL) {
2591 		mtx_unlock(&sw_dev_mtx);
2592 		bp->b_error = ENXIO;
2593 		bp->b_ioflags |= BIO_ERROR;
2594 		bufdone(bp);
2595 		return;
2596 	}
2597 	swapgeom_acquire(cp);
2598 	mtx_unlock(&sw_dev_mtx);
2599 	if (bp->b_iocmd == BIO_WRITE)
2600 		bio = g_new_bio();
2601 	else
2602 		bio = g_alloc_bio();
2603 	if (bio == NULL) {
2604 		mtx_lock(&sw_dev_mtx);
2605 		swapgeom_release(cp, sp);
2606 		mtx_unlock(&sw_dev_mtx);
2607 		bp->b_error = ENOMEM;
2608 		bp->b_ioflags |= BIO_ERROR;
2609 		bufdone(bp);
2610 		return;
2611 	}
2612 
2613 	bio->bio_caller1 = sp;
2614 	bio->bio_caller2 = bp;
2615 	bio->bio_cmd = bp->b_iocmd;
2616 	bio->bio_offset = (bp->b_blkno - sp->sw_first) * PAGE_SIZE;
2617 	bio->bio_length = bp->b_bcount;
2618 	bio->bio_done = swapgeom_done;
2619 	if (!buf_mapped(bp)) {
2620 		bio->bio_ma = bp->b_pages;
2621 		bio->bio_data = unmapped_buf;
2622 		bio->bio_ma_offset = (vm_offset_t)bp->b_offset & PAGE_MASK;
2623 		bio->bio_ma_n = bp->b_npages;
2624 		bio->bio_flags |= BIO_UNMAPPED;
2625 	} else {
2626 		bio->bio_data = bp->b_data;
2627 		bio->bio_ma = NULL;
2628 	}
2629 	g_io_request(bio, cp);
2630 	return;
2631 }
2632 
2633 static void
2634 swapgeom_orphan(struct g_consumer *cp)
2635 {
2636 	struct swdevt *sp;
2637 	int destroy;
2638 
2639 	mtx_lock(&sw_dev_mtx);
2640 	TAILQ_FOREACH(sp, &swtailq, sw_list) {
2641 		if (sp->sw_id == cp) {
2642 			sp->sw_flags |= SW_CLOSING;
2643 			break;
2644 		}
2645 	}
2646 	/*
2647 	 * Drop reference we were created with. Do directly since we're in a
2648 	 * special context where we don't have to queue the call to
2649 	 * swapgeom_close_ev().
2650 	 */
2651 	cp->index--;
2652 	destroy = ((sp != NULL) && (cp->index == 0));
2653 	if (destroy)
2654 		sp->sw_id = NULL;
2655 	mtx_unlock(&sw_dev_mtx);
2656 	if (destroy)
2657 		swapgeom_close_ev(cp, 0);
2658 }
2659 
2660 static void
2661 swapgeom_close(struct thread *td, struct swdevt *sw)
2662 {
2663 	struct g_consumer *cp;
2664 
2665 	mtx_lock(&sw_dev_mtx);
2666 	cp = sw->sw_id;
2667 	sw->sw_id = NULL;
2668 	mtx_unlock(&sw_dev_mtx);
2669 	/* XXX: direct call when Giant untangled */
2670 	if (cp != NULL)
2671 		g_waitfor_event(swapgeom_close_ev, cp, M_WAITOK, NULL);
2672 }
2673 
2674 
2675 struct swh0h0 {
2676 	struct cdev *dev;
2677 	struct vnode *vp;
2678 	int	error;
2679 };
2680 
2681 static void
2682 swapongeom_ev(void *arg, int flags)
2683 {
2684 	struct swh0h0 *swh;
2685 	struct g_provider *pp;
2686 	struct g_consumer *cp;
2687 	static struct g_geom *gp;
2688 	struct swdevt *sp;
2689 	u_long nblks;
2690 	int error;
2691 
2692 	swh = arg;
2693 	swh->error = 0;
2694 	pp = g_dev_getprovider(swh->dev);
2695 	if (pp == NULL) {
2696 		swh->error = ENODEV;
2697 		return;
2698 	}
2699 	mtx_lock(&sw_dev_mtx);
2700 	TAILQ_FOREACH(sp, &swtailq, sw_list) {
2701 		cp = sp->sw_id;
2702 		if (cp != NULL && cp->provider == pp) {
2703 			mtx_unlock(&sw_dev_mtx);
2704 			swh->error = EBUSY;
2705 			return;
2706 		}
2707 	}
2708 	mtx_unlock(&sw_dev_mtx);
2709 	if (gp == NULL)
2710 		gp = g_new_geomf(&g_swap_class, "swap");
2711 	cp = g_new_consumer(gp);
2712 	cp->index = 1;		/* Number of active I/Os, plus one for being active. */
2713 	cp->flags |=  G_CF_DIRECT_SEND | G_CF_DIRECT_RECEIVE;
2714 	g_attach(cp, pp);
2715 	/*
2716 	 * XXX: Everytime you think you can improve the margin for
2717 	 * footshooting, somebody depends on the ability to do so:
2718 	 * savecore(8) wants to write to our swapdev so we cannot
2719 	 * set an exclusive count :-(
2720 	 */
2721 	error = g_access(cp, 1, 1, 0);
2722 	if (error) {
2723 		g_detach(cp);
2724 		g_destroy_consumer(cp);
2725 		swh->error = error;
2726 		return;
2727 	}
2728 	nblks = pp->mediasize / DEV_BSIZE;
2729 	swaponsomething(swh->vp, cp, nblks, swapgeom_strategy,
2730 	    swapgeom_close, dev2udev(swh->dev),
2731 	    (pp->flags & G_PF_ACCEPT_UNMAPPED) != 0 ? SW_UNMAPPED : 0);
2732 	swh->error = 0;
2733 }
2734 
2735 static int
2736 swapongeom(struct thread *td, struct vnode *vp)
2737 {
2738 	int error;
2739 	struct swh0h0 swh;
2740 
2741 	vn_lock(vp, LK_EXCLUSIVE | LK_RETRY);
2742 
2743 	swh.dev = vp->v_rdev;
2744 	swh.vp = vp;
2745 	swh.error = 0;
2746 	/* XXX: direct call when Giant untangled */
2747 	error = g_waitfor_event(swapongeom_ev, &swh, M_WAITOK, NULL);
2748 	if (!error)
2749 		error = swh.error;
2750 	VOP_UNLOCK(vp, 0);
2751 	return (error);
2752 }
2753 
2754 /*
2755  * VNODE backend
2756  *
2757  * This is used mainly for network filesystem (read: probably only tested
2758  * with NFS) swapfiles.
2759  *
2760  */
2761 
2762 static void
2763 swapdev_strategy(struct buf *bp, struct swdevt *sp)
2764 {
2765 	struct vnode *vp2;
2766 
2767 	bp->b_blkno = ctodb(bp->b_blkno - sp->sw_first);
2768 
2769 	vp2 = sp->sw_id;
2770 	vhold(vp2);
2771 	if (bp->b_iocmd == BIO_WRITE) {
2772 		if (bp->b_bufobj)
2773 			bufobj_wdrop(bp->b_bufobj);
2774 		bufobj_wref(&vp2->v_bufobj);
2775 	}
2776 	if (bp->b_bufobj != &vp2->v_bufobj)
2777 		bp->b_bufobj = &vp2->v_bufobj;
2778 	bp->b_vp = vp2;
2779 	bp->b_iooffset = dbtob(bp->b_blkno);
2780 	bstrategy(bp);
2781 	return;
2782 }
2783 
2784 static void
2785 swapdev_close(struct thread *td, struct swdevt *sp)
2786 {
2787 
2788 	VOP_CLOSE(sp->sw_vp, FREAD | FWRITE, td->td_ucred, td);
2789 	vrele(sp->sw_vp);
2790 }
2791 
2792 
2793 static int
2794 swaponvp(struct thread *td, struct vnode *vp, u_long nblks)
2795 {
2796 	struct swdevt *sp;
2797 	int error;
2798 
2799 	if (nblks == 0)
2800 		return (ENXIO);
2801 	mtx_lock(&sw_dev_mtx);
2802 	TAILQ_FOREACH(sp, &swtailq, sw_list) {
2803 		if (sp->sw_id == vp) {
2804 			mtx_unlock(&sw_dev_mtx);
2805 			return (EBUSY);
2806 		}
2807 	}
2808 	mtx_unlock(&sw_dev_mtx);
2809 
2810 	(void) vn_lock(vp, LK_EXCLUSIVE | LK_RETRY);
2811 #ifdef MAC
2812 	error = mac_system_check_swapon(td->td_ucred, vp);
2813 	if (error == 0)
2814 #endif
2815 		error = VOP_OPEN(vp, FREAD | FWRITE, td->td_ucred, td, NULL);
2816 	(void) VOP_UNLOCK(vp, 0);
2817 	if (error)
2818 		return (error);
2819 
2820 	swaponsomething(vp, vp, nblks, swapdev_strategy, swapdev_close,
2821 	    NODEV, 0);
2822 	return (0);
2823 }
2824 
2825 static int
2826 sysctl_swap_async_max(SYSCTL_HANDLER_ARGS)
2827 {
2828 	int error, new, n;
2829 
2830 	new = nsw_wcount_async_max;
2831 	error = sysctl_handle_int(oidp, &new, 0, req);
2832 	if (error != 0 || req->newptr == NULL)
2833 		return (error);
2834 
2835 	if (new > nswbuf / 2 || new < 1)
2836 		return (EINVAL);
2837 
2838 	mtx_lock(&pbuf_mtx);
2839 	while (nsw_wcount_async_max != new) {
2840 		/*
2841 		 * Adjust difference.  If the current async count is too low,
2842 		 * we will need to sqeeze our update slowly in.  Sleep with a
2843 		 * higher priority than getpbuf() to finish faster.
2844 		 */
2845 		n = new - nsw_wcount_async_max;
2846 		if (nsw_wcount_async + n >= 0) {
2847 			nsw_wcount_async += n;
2848 			nsw_wcount_async_max += n;
2849 			wakeup(&nsw_wcount_async);
2850 		} else {
2851 			nsw_wcount_async_max -= nsw_wcount_async;
2852 			nsw_wcount_async = 0;
2853 			msleep(&nsw_wcount_async, &pbuf_mtx, PSWP,
2854 			    "swpsysctl", 0);
2855 		}
2856 	}
2857 	mtx_unlock(&pbuf_mtx);
2858 
2859 	return (0);
2860 }
2861