xref: /freebsd/sys/vm/swap_pager.c (revision 59f5f100b774de8824fb2fc1a8a11a93bbc2dafd)
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 
68 #include <sys/cdefs.h>
69 #include "opt_vm.h"
70 
71 #include <sys/param.h>
72 #include <sys/bio.h>
73 #include <sys/blist.h>
74 #include <sys/buf.h>
75 #include <sys/conf.h>
76 #include <sys/disk.h>
77 #include <sys/disklabel.h>
78 #include <sys/eventhandler.h>
79 #include <sys/fcntl.h>
80 #include <sys/limits.h>
81 #include <sys/lock.h>
82 #include <sys/kernel.h>
83 #include <sys/mount.h>
84 #include <sys/namei.h>
85 #include <sys/malloc.h>
86 #include <sys/pctrie.h>
87 #include <sys/priv.h>
88 #include <sys/proc.h>
89 #include <sys/racct.h>
90 #include <sys/resource.h>
91 #include <sys/resourcevar.h>
92 #include <sys/rwlock.h>
93 #include <sys/sbuf.h>
94 #include <sys/sysctl.h>
95 #include <sys/sysproto.h>
96 #include <sys/systm.h>
97 #include <sys/sx.h>
98 #include <sys/unistd.h>
99 #include <sys/user.h>
100 #include <sys/vmmeter.h>
101 #include <sys/vnode.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 /*
146  * A page_range structure records the start address and length of a sequence of
147  * mapped page addresses.
148  */
149 struct page_range {
150 	daddr_t start;
151 	daddr_t num;
152 };
153 
154 static MALLOC_DEFINE(M_VMPGDATA, "vm_pgdata", "swap pager private data");
155 static struct mtx sw_dev_mtx;
156 static TAILQ_HEAD(, swdevt) swtailq = TAILQ_HEAD_INITIALIZER(swtailq);
157 static struct swdevt *swdevhd;	/* Allocate from here next */
158 static int nswapdev;		/* Number of swap devices */
159 int swap_pager_avail;
160 static struct sx swdev_syscall_lock;	/* serialize swap(on|off) */
161 
162 static __exclusive_cache_line u_long swap_reserved;
163 static u_long swap_total;
164 static int sysctl_page_shift(SYSCTL_HANDLER_ARGS);
165 
166 static SYSCTL_NODE(_vm_stats, OID_AUTO, swap, CTLFLAG_RD | CTLFLAG_MPSAFE, 0,
167     "VM swap stats");
168 
169 SYSCTL_PROC(_vm, OID_AUTO, swap_reserved, CTLTYPE_U64 | CTLFLAG_RD | CTLFLAG_MPSAFE,
170     &swap_reserved, 0, sysctl_page_shift, "QU",
171     "Amount of swap storage needed to back all allocated anonymous memory.");
172 SYSCTL_PROC(_vm, OID_AUTO, swap_total, CTLTYPE_U64 | CTLFLAG_RD | CTLFLAG_MPSAFE,
173     &swap_total, 0, sysctl_page_shift, "QU",
174     "Total amount of available swap storage.");
175 
176 int vm_overcommit __read_mostly = 0;
177 SYSCTL_INT(_vm, VM_OVERCOMMIT, overcommit, CTLFLAG_RW, &vm_overcommit, 0,
178     "Configure virtual memory overcommit behavior. See tuning(7) "
179     "for details.");
180 static unsigned long swzone;
181 SYSCTL_ULONG(_vm, OID_AUTO, swzone, CTLFLAG_RD, &swzone, 0,
182     "Actual size of swap metadata zone");
183 static unsigned long swap_maxpages;
184 SYSCTL_ULONG(_vm, OID_AUTO, swap_maxpages, CTLFLAG_RD, &swap_maxpages, 0,
185     "Maximum amount of swap supported");
186 
187 static COUNTER_U64_DEFINE_EARLY(swap_free_deferred);
188 SYSCTL_COUNTER_U64(_vm_stats_swap, OID_AUTO, free_deferred,
189     CTLFLAG_RD, &swap_free_deferred,
190     "Number of pages that deferred freeing swap space");
191 
192 static COUNTER_U64_DEFINE_EARLY(swap_free_completed);
193 SYSCTL_COUNTER_U64(_vm_stats_swap, OID_AUTO, free_completed,
194     CTLFLAG_RD, &swap_free_completed,
195     "Number of deferred frees completed");
196 
197 static int
198 sysctl_page_shift(SYSCTL_HANDLER_ARGS)
199 {
200 	uint64_t newval;
201 	u_long value = *(u_long *)arg1;
202 
203 	newval = ((uint64_t)value) << PAGE_SHIFT;
204 	return (sysctl_handle_64(oidp, &newval, 0, req));
205 }
206 
207 static bool
208 swap_reserve_by_cred_rlimit(u_long pincr, struct ucred *cred, int oc)
209 {
210 	struct uidinfo *uip;
211 	u_long prev;
212 
213 	uip = cred->cr_ruidinfo;
214 
215 	prev = atomic_fetchadd_long(&uip->ui_vmsize, pincr);
216 	if ((oc & SWAP_RESERVE_RLIMIT_ON) != 0 &&
217 	    prev + pincr > lim_cur(curthread, RLIMIT_SWAP) &&
218 	    priv_check(curthread, PRIV_VM_SWAP_NORLIMIT) != 0) {
219 		prev = atomic_fetchadd_long(&uip->ui_vmsize, -pincr);
220 		KASSERT(prev >= pincr,
221 		    ("negative vmsize for uid %d\n", uip->ui_uid));
222 		return (false);
223 	}
224 	return (true);
225 }
226 
227 static void
228 swap_release_by_cred_rlimit(u_long pdecr, struct ucred *cred)
229 {
230 	struct uidinfo *uip;
231 #ifdef INVARIANTS
232 	u_long prev;
233 #endif
234 
235 	uip = cred->cr_ruidinfo;
236 
237 #ifdef INVARIANTS
238 	prev = atomic_fetchadd_long(&uip->ui_vmsize, -pdecr);
239 	KASSERT(prev >= pdecr,
240 	    ("negative vmsize for uid %d\n", uip->ui_uid));
241 #else
242 	atomic_subtract_long(&uip->ui_vmsize, pdecr);
243 #endif
244 }
245 
246 static void
247 swap_reserve_force_rlimit(u_long pincr, struct ucred *cred)
248 {
249 	struct uidinfo *uip;
250 
251 	uip = cred->cr_ruidinfo;
252 	atomic_add_long(&uip->ui_vmsize, pincr);
253 }
254 
255 bool
256 swap_reserve(vm_ooffset_t incr)
257 {
258 
259 	return (swap_reserve_by_cred(incr, curthread->td_ucred));
260 }
261 
262 bool
263 swap_reserve_by_cred(vm_ooffset_t incr, struct ucred *cred)
264 {
265 	u_long r, s, prev, pincr;
266 #ifdef RACCT
267 	int error;
268 #endif
269 	int oc;
270 	static int curfail;
271 	static struct timeval lastfail;
272 
273 	KASSERT((incr & PAGE_MASK) == 0, ("%s: incr: %ju & PAGE_MASK",
274 	    __func__, (uintmax_t)incr));
275 
276 #ifdef RACCT
277 	if (RACCT_ENABLED()) {
278 		PROC_LOCK(curproc);
279 		error = racct_add(curproc, RACCT_SWAP, incr);
280 		PROC_UNLOCK(curproc);
281 		if (error != 0)
282 			return (false);
283 	}
284 #endif
285 
286 	pincr = atop(incr);
287 	prev = atomic_fetchadd_long(&swap_reserved, pincr);
288 	r = prev + pincr;
289 	s = swap_total;
290 	oc = atomic_load_int(&vm_overcommit);
291 	if (r > s && (oc & SWAP_RESERVE_ALLOW_NONWIRED) != 0) {
292 		s += vm_cnt.v_page_count - vm_cnt.v_free_reserved -
293 		    vm_wire_count();
294 	}
295 	if ((oc & SWAP_RESERVE_FORCE_ON) != 0 && r > s &&
296 	    priv_check(curthread, PRIV_VM_SWAP_NOQUOTA) != 0) {
297 		prev = atomic_fetchadd_long(&swap_reserved, -pincr);
298 		KASSERT(prev >= pincr,
299 		    ("swap_reserved < incr on overcommit fail"));
300 		goto out_error;
301 	}
302 
303 	if (!swap_reserve_by_cred_rlimit(pincr, cred, oc)) {
304 		prev = atomic_fetchadd_long(&swap_reserved, -pincr);
305 		KASSERT(prev >= pincr,
306 		    ("swap_reserved < incr on overcommit fail"));
307 		goto out_error;
308 	}
309 
310 	return (true);
311 
312 out_error:
313 	if (ppsratecheck(&lastfail, &curfail, 1)) {
314 		printf("uid %d, pid %d: swap reservation "
315 		    "for %jd bytes failed\n",
316 		    cred->cr_ruidinfo->ui_uid, curproc->p_pid, incr);
317 	}
318 #ifdef RACCT
319 	if (RACCT_ENABLED()) {
320 		PROC_LOCK(curproc);
321 		racct_sub(curproc, RACCT_SWAP, incr);
322 		PROC_UNLOCK(curproc);
323 	}
324 #endif
325 
326 	return (false);
327 }
328 
329 void
330 swap_reserve_force(vm_ooffset_t incr)
331 {
332 	u_long pincr;
333 
334 	KASSERT((incr & PAGE_MASK) == 0, ("%s: incr: %ju & PAGE_MASK",
335 	    __func__, (uintmax_t)incr));
336 
337 #ifdef RACCT
338 	if (RACCT_ENABLED()) {
339 		PROC_LOCK(curproc);
340 		racct_add_force(curproc, RACCT_SWAP, incr);
341 		PROC_UNLOCK(curproc);
342 	}
343 #endif
344 	pincr = atop(incr);
345 	atomic_add_long(&swap_reserved, pincr);
346 	swap_reserve_force_rlimit(pincr, curthread->td_ucred);
347 }
348 
349 void
350 swap_release(vm_ooffset_t decr)
351 {
352 	struct ucred *cred;
353 
354 	PROC_LOCK(curproc);
355 	cred = curproc->p_ucred;
356 	swap_release_by_cred(decr, cred);
357 	PROC_UNLOCK(curproc);
358 }
359 
360 void
361 swap_release_by_cred(vm_ooffset_t decr, struct ucred *cred)
362 {
363 	u_long pdecr;
364 #ifdef INVARIANTS
365 	u_long prev;
366 #endif
367 
368 	KASSERT((decr & PAGE_MASK) == 0, ("%s: decr: %ju & PAGE_MASK",
369 	    __func__, (uintmax_t)decr));
370 
371 	pdecr = atop(decr);
372 #ifdef INVARIANTS
373 	prev = atomic_fetchadd_long(&swap_reserved, -pdecr);
374 	KASSERT(prev >= pdecr, ("swap_reserved < decr"));
375 #else
376 	atomic_subtract_long(&swap_reserved, pdecr);
377 #endif
378 
379 	swap_release_by_cred_rlimit(pdecr, cred);
380 #ifdef RACCT
381 	if (racct_enable)
382 		racct_sub_cred(cred, RACCT_SWAP, decr);
383 #endif
384 }
385 
386 static int swap_pager_full = 2;	/* swap space exhaustion (task killing) */
387 static int swap_pager_almost_full = 1; /* swap space exhaustion (w/hysteresis)*/
388 static struct mtx swbuf_mtx;	/* to sync nsw_wcount_async */
389 static int nsw_wcount_async;	/* limit async write buffers */
390 static int nsw_wcount_async_max;/* assigned maximum			*/
391 int nsw_cluster_max; 		/* maximum VOP I/O allowed		*/
392 
393 static int sysctl_swap_async_max(SYSCTL_HANDLER_ARGS);
394 SYSCTL_PROC(_vm, OID_AUTO, swap_async_max, CTLTYPE_INT | CTLFLAG_RW |
395     CTLFLAG_MPSAFE, NULL, 0, sysctl_swap_async_max, "I",
396     "Maximum running async swap ops");
397 static int sysctl_swap_fragmentation(SYSCTL_HANDLER_ARGS);
398 SYSCTL_PROC(_vm, OID_AUTO, swap_fragmentation, CTLTYPE_STRING | CTLFLAG_RD |
399     CTLFLAG_MPSAFE, NULL, 0, sysctl_swap_fragmentation, "A",
400     "Swap Fragmentation Info");
401 
402 static struct sx sw_alloc_sx;
403 
404 /*
405  * "named" and "unnamed" anon region objects.  Try to reduce the overhead
406  * of searching a named list by hashing it just a little.
407  */
408 
409 #define NOBJLISTS		8
410 
411 #define NOBJLIST(handle)	\
412 	(&swap_pager_object_list[((int)(intptr_t)handle >> 4) & (NOBJLISTS-1)])
413 
414 static struct pagerlst	swap_pager_object_list[NOBJLISTS];
415 static uma_zone_t swwbuf_zone;
416 static uma_zone_t swrbuf_zone;
417 static uma_zone_t swblk_zone;
418 static uma_zone_t swpctrie_zone;
419 
420 /*
421  * pagerops for OBJT_SWAP - "swap pager".  Some ops are also global procedure
422  * calls hooked from other parts of the VM system and do not appear here.
423  * (see vm/swap_pager.h).
424  */
425 static vm_object_t
426 		swap_pager_alloc(void *handle, vm_ooffset_t size,
427 		    vm_prot_t prot, vm_ooffset_t offset, struct ucred *);
428 static void	swap_pager_dealloc(vm_object_t object);
429 static int	swap_pager_getpages(vm_object_t, vm_page_t *, int, int *,
430     int *);
431 static int	swap_pager_getpages_async(vm_object_t, vm_page_t *, int, int *,
432     int *, pgo_getpages_iodone_t, void *);
433 static void	swap_pager_putpages(vm_object_t, vm_page_t *, int, int, int *);
434 static boolean_t
435 		swap_pager_haspage(vm_object_t object, vm_pindex_t pindex, int *before, int *after);
436 static void	swap_pager_init(void);
437 static void	swap_pager_unswapped(vm_page_t);
438 static void	swap_pager_swapoff(struct swdevt *sp);
439 static void	swap_pager_update_writecount(vm_object_t object,
440     vm_offset_t start, vm_offset_t end);
441 static void	swap_pager_release_writecount(vm_object_t object,
442     vm_offset_t start, vm_offset_t end);
443 static void	swap_pager_freespace_pgo(vm_object_t object, vm_pindex_t start,
444     vm_size_t size);
445 
446 const struct pagerops swappagerops = {
447 	.pgo_kvme_type = KVME_TYPE_SWAP,
448 	.pgo_init =	swap_pager_init,	/* early system initialization of pager	*/
449 	.pgo_alloc =	swap_pager_alloc,	/* allocate an OBJT_SWAP object */
450 	.pgo_dealloc =	swap_pager_dealloc,	/* deallocate an OBJT_SWAP object */
451 	.pgo_getpages =	swap_pager_getpages,	/* pagein */
452 	.pgo_getpages_async = swap_pager_getpages_async, /* pagein (async) */
453 	.pgo_putpages =	swap_pager_putpages,	/* pageout */
454 	.pgo_haspage =	swap_pager_haspage,	/* get backing store status for page */
455 	.pgo_pageunswapped = swap_pager_unswapped, /* remove swap related to page */
456 	.pgo_update_writecount = swap_pager_update_writecount,
457 	.pgo_release_writecount = swap_pager_release_writecount,
458 	.pgo_freespace = swap_pager_freespace_pgo,
459 };
460 
461 /*
462  * swap_*() routines are externally accessible.  swp_*() routines are
463  * internal.
464  */
465 static int nswap_lowat = 128;	/* in pages, swap_pager_almost_full warn */
466 static int nswap_hiwat = 512;	/* in pages, swap_pager_almost_full warn */
467 
468 SYSCTL_INT(_vm, OID_AUTO, dmmax, CTLFLAG_RD, &nsw_cluster_max, 0,
469     "Maximum size of a swap block in pages");
470 
471 static void	swp_sizecheck(void);
472 static void	swp_pager_async_iodone(struct buf *bp);
473 static bool	swp_pager_swblk_empty(struct swblk *sb, int start, int limit);
474 static void	swp_pager_free_empty_swblk(vm_object_t, struct swblk *sb);
475 static int	swapongeom(struct vnode *);
476 static int	swaponvp(struct thread *, struct vnode *, u_long);
477 static int	swapoff_one(struct swdevt *sp, struct ucred *cred,
478 		    u_int flags);
479 
480 /*
481  * Swap bitmap functions
482  */
483 static void	swp_pager_freeswapspace(const struct page_range *range);
484 static daddr_t	swp_pager_getswapspace(int *npages);
485 
486 /*
487  * Metadata functions
488  */
489 static daddr_t swp_pager_meta_build(vm_object_t, vm_pindex_t, daddr_t,
490 	bool);
491 static void swp_pager_meta_free(vm_object_t, vm_pindex_t, vm_pindex_t,
492     vm_size_t *);
493 static void swp_pager_meta_transfer(vm_object_t src, vm_object_t dst,
494     vm_pindex_t pindex, vm_pindex_t count);
495 static void swp_pager_meta_free_all(vm_object_t);
496 static daddr_t swp_pager_meta_lookup(vm_object_t, vm_pindex_t);
497 
498 static void
499 swp_pager_init_freerange(struct page_range *range)
500 {
501 	range->start = SWAPBLK_NONE;
502 	range->num = 0;
503 }
504 
505 static void
506 swp_pager_update_freerange(struct page_range *range, daddr_t addr)
507 {
508 	if (range->start + range->num == addr) {
509 		range->num++;
510 	} else {
511 		swp_pager_freeswapspace(range);
512 		range->start = addr;
513 		range->num = 1;
514 	}
515 }
516 
517 static void *
518 swblk_trie_alloc(struct pctrie *ptree)
519 {
520 
521 	return (uma_zalloc(swpctrie_zone, M_NOWAIT | (curproc == pageproc ?
522 	    M_USE_RESERVE : 0)));
523 }
524 
525 static void
526 swblk_trie_free(struct pctrie *ptree, void *node)
527 {
528 
529 	uma_zfree(swpctrie_zone, node);
530 }
531 
532 static int
533 swblk_start(struct swblk *sb, vm_pindex_t pindex)
534 {
535 	return (sb == NULL || sb->p >= pindex ?
536 	    0 : pindex % SWAP_META_PAGES);
537 }
538 
539 PCTRIE_DEFINE(SWAP, swblk, p, swblk_trie_alloc, swblk_trie_free);
540 
541 static struct swblk *
542 swblk_lookup(vm_object_t object, vm_pindex_t pindex)
543 {
544 	return (SWAP_PCTRIE_LOOKUP(&object->un_pager.swp.swp_blks,
545 	    rounddown(pindex, SWAP_META_PAGES)));
546 }
547 
548 static void
549 swblk_lookup_remove(vm_object_t object, struct swblk *sb)
550 {
551 	SWAP_PCTRIE_REMOVE(&object->un_pager.swp.swp_blks, sb->p);
552 }
553 
554 static int
555 swblk_lookup_insert(vm_object_t object, struct swblk *sb)
556 {
557 	return (SWAP_PCTRIE_INSERT(&object->un_pager.swp.swp_blks, sb));
558 }
559 
560 static bool
561 swblk_is_empty(vm_object_t object)
562 {
563 	return (pctrie_is_empty(&object->un_pager.swp.swp_blks));
564 }
565 
566 static struct swblk *
567 swblk_iter(struct pctrie_iter *blks, vm_object_t object,
568     vm_pindex_t pindex)
569 {
570 	VM_OBJECT_ASSERT_LOCKED(object);
571 	MPASS((object->flags & OBJ_SWAP) != 0);
572 	pctrie_iter_init(blks, &object->un_pager.swp.swp_blks);
573 	return (SWAP_PCTRIE_ITER_LOOKUP_GE(blks,
574 	    rounddown(pindex, SWAP_META_PAGES)));
575 }
576 
577 static struct swblk *
578 swblk_iter_limit(struct pctrie_iter *blks, vm_object_t object,
579     vm_pindex_t pindex, vm_pindex_t limit)
580 {
581 	VM_OBJECT_ASSERT_LOCKED(object);
582 	MPASS((object->flags & OBJ_SWAP) != 0);
583 	pctrie_iter_limit_init(blks, &object->un_pager.swp.swp_blks, limit);
584 	return (SWAP_PCTRIE_ITER_LOOKUP_GE(blks,
585 	    rounddown(pindex, SWAP_META_PAGES)));
586 }
587 
588 static struct swblk *
589 swblk_iter_next(struct pctrie_iter *blks)
590 {
591 	return (SWAP_PCTRIE_ITER_JUMP_GE(blks, SWAP_META_PAGES));
592 }
593 
594 static void
595 swblk_iter_remove(struct pctrie_iter *blks)
596 {
597 	SWAP_PCTRIE_ITER_REMOVE(blks);
598 }
599 
600 /*
601  * SWP_SIZECHECK() -	update swap_pager_full indication
602  *
603  *	update the swap_pager_almost_full indication and warn when we are
604  *	about to run out of swap space, using lowat/hiwat hysteresis.
605  *
606  *	Clear swap_pager_full ( task killing ) indication when lowat is met.
607  *
608  *	No restrictions on call
609  *	This routine may not block.
610  */
611 static void
612 swp_sizecheck(void)
613 {
614 
615 	if (swap_pager_avail < nswap_lowat) {
616 		if (swap_pager_almost_full == 0) {
617 			printf("swap_pager: out of swap space\n");
618 			swap_pager_almost_full = 1;
619 		}
620 	} else {
621 		swap_pager_full = 0;
622 		if (swap_pager_avail > nswap_hiwat)
623 			swap_pager_almost_full = 0;
624 	}
625 }
626 
627 /*
628  * SWAP_PAGER_INIT() -	initialize the swap pager!
629  *
630  *	Expected to be started from system init.  NOTE:  This code is run
631  *	before much else so be careful what you depend on.  Most of the VM
632  *	system has yet to be initialized at this point.
633  */
634 static void
635 swap_pager_init(void)
636 {
637 	/*
638 	 * Initialize object lists
639 	 */
640 	int i;
641 
642 	for (i = 0; i < NOBJLISTS; ++i)
643 		TAILQ_INIT(&swap_pager_object_list[i]);
644 	mtx_init(&sw_dev_mtx, "swapdev", NULL, MTX_DEF);
645 	sx_init(&sw_alloc_sx, "swspsx");
646 	sx_init(&swdev_syscall_lock, "swsysc");
647 
648 	/*
649 	 * The nsw_cluster_max is constrained by the bp->b_pages[]
650 	 * array, which has maxphys / PAGE_SIZE entries, and our locally
651 	 * defined MAX_PAGEOUT_CLUSTER.   Also be aware that swap ops are
652 	 * constrained by the swap device interleave stripe size.
653 	 *
654 	 * Initialized early so that GEOM_ELI can see it.
655 	 */
656 	nsw_cluster_max = min(maxphys / PAGE_SIZE, MAX_PAGEOUT_CLUSTER);
657 }
658 
659 /*
660  * SWAP_PAGER_SWAP_INIT() - swap pager initialization from pageout process
661  *
662  *	Expected to be started from pageout process once, prior to entering
663  *	its main loop.
664  */
665 void
666 swap_pager_swap_init(void)
667 {
668 	unsigned long n, n2;
669 
670 	/*
671 	 * Number of in-transit swap bp operations.  Don't
672 	 * exhaust the pbufs completely.  Make sure we
673 	 * initialize workable values (0 will work for hysteresis
674 	 * but it isn't very efficient).
675 	 *
676 	 * Currently we hardwire nsw_wcount_async to 4.  This limit is
677 	 * designed to prevent other I/O from having high latencies due to
678 	 * our pageout I/O.  The value 4 works well for one or two active swap
679 	 * devices but is probably a little low if you have more.  Even so,
680 	 * a higher value would probably generate only a limited improvement
681 	 * with three or four active swap devices since the system does not
682 	 * typically have to pageout at extreme bandwidths.   We will want
683 	 * at least 2 per swap devices, and 4 is a pretty good value if you
684 	 * have one NFS swap device due to the command/ack latency over NFS.
685 	 * So it all works out pretty well.
686 	 *
687 	 * nsw_cluster_max is initialized in swap_pager_init().
688 	 */
689 
690 	nsw_wcount_async = 4;
691 	nsw_wcount_async_max = nsw_wcount_async;
692 	mtx_init(&swbuf_mtx, "async swbuf mutex", NULL, MTX_DEF);
693 
694 	swwbuf_zone = pbuf_zsecond_create("swwbuf", nswbuf / 4);
695 	swrbuf_zone = pbuf_zsecond_create("swrbuf", nswbuf / 2);
696 
697 	/*
698 	 * Initialize our zone, taking the user's requested size or
699 	 * estimating the number we need based on the number of pages
700 	 * in the system.
701 	 */
702 	n = maxswzone != 0 ? maxswzone / sizeof(struct swblk) :
703 	    vm_cnt.v_page_count / 2;
704 	swpctrie_zone = uma_zcreate("swpctrie", pctrie_node_size(), NULL, NULL,
705 	    pctrie_zone_init, NULL, UMA_ALIGN_PTR, 0);
706 	swblk_zone = uma_zcreate("swblk", sizeof(struct swblk), NULL, NULL,
707 	    NULL, NULL, _Alignof(struct swblk) - 1, 0);
708 	n2 = n;
709 	do {
710 		if (uma_zone_reserve_kva(swblk_zone, n))
711 			break;
712 		/*
713 		 * if the allocation failed, try a zone two thirds the
714 		 * size of the previous attempt.
715 		 */
716 		n -= ((n + 2) / 3);
717 	} while (n > 0);
718 
719 	/*
720 	 * Often uma_zone_reserve_kva() cannot reserve exactly the
721 	 * requested size.  Account for the difference when
722 	 * calculating swap_maxpages.
723 	 */
724 	n = uma_zone_get_max(swblk_zone);
725 
726 	if (n < n2)
727 		printf("Swap blk zone entries changed from %lu to %lu.\n",
728 		    n2, n);
729 	/* absolute maximum we can handle assuming 100% efficiency */
730 	swap_maxpages = n * SWAP_META_PAGES;
731 	swzone = n * sizeof(struct swblk);
732 	if (!uma_zone_reserve_kva(swpctrie_zone, n))
733 		printf("Cannot reserve swap pctrie zone, "
734 		    "reduce kern.maxswzone.\n");
735 }
736 
737 bool
738 swap_pager_init_object(vm_object_t object, void *handle, struct ucred *cred,
739     vm_ooffset_t size, vm_ooffset_t offset)
740 {
741 	if (cred != NULL) {
742 		if (!swap_reserve_by_cred(size, cred))
743 			return (false);
744 		crhold(cred);
745 	}
746 
747 	object->un_pager.swp.writemappings = 0;
748 	object->handle = handle;
749 	if (cred != NULL) {
750 		object->cred = cred;
751 		object->charge = size;
752 	}
753 	return (true);
754 }
755 
756 static vm_object_t
757 swap_pager_alloc_init(objtype_t otype, void *handle, struct ucred *cred,
758     vm_ooffset_t size, vm_ooffset_t offset)
759 {
760 	vm_object_t object;
761 
762 	/*
763 	 * The un_pager.swp.swp_blks trie is initialized by
764 	 * vm_object_allocate() to ensure the correct order of
765 	 * visibility to other threads.
766 	 */
767 	object = vm_object_allocate(otype, OFF_TO_IDX(offset +
768 	    PAGE_MASK + size));
769 
770 	if (!swap_pager_init_object(object, handle, cred, size, offset)) {
771 		vm_object_deallocate(object);
772 		return (NULL);
773 	}
774 	return (object);
775 }
776 
777 /*
778  * SWAP_PAGER_ALLOC() -	allocate a new OBJT_SWAP VM object and instantiate
779  *			its metadata structures.
780  *
781  *	This routine is called from the mmap and fork code to create a new
782  *	OBJT_SWAP object.
783  *
784  *	This routine must ensure that no live duplicate is created for
785  *	the named object request, which is protected against by
786  *	holding the sw_alloc_sx lock in case handle != NULL.
787  */
788 static vm_object_t
789 swap_pager_alloc(void *handle, vm_ooffset_t size, vm_prot_t prot,
790     vm_ooffset_t offset, struct ucred *cred)
791 {
792 	vm_object_t object;
793 
794 	if (handle != NULL) {
795 		/*
796 		 * Reference existing named region or allocate new one.  There
797 		 * should not be a race here against swp_pager_meta_build()
798 		 * as called from vm_page_remove() in regards to the lookup
799 		 * of the handle.
800 		 */
801 		sx_xlock(&sw_alloc_sx);
802 		object = vm_pager_object_lookup(NOBJLIST(handle), handle);
803 		if (object == NULL) {
804 			object = swap_pager_alloc_init(OBJT_SWAP, handle, cred,
805 			    size, offset);
806 			if (object != NULL) {
807 				TAILQ_INSERT_TAIL(NOBJLIST(object->handle),
808 				    object, pager_object_list);
809 			}
810 		}
811 		sx_xunlock(&sw_alloc_sx);
812 	} else {
813 		object = swap_pager_alloc_init(OBJT_SWAP, handle, cred,
814 		    size, offset);
815 	}
816 	return (object);
817 }
818 
819 /*
820  * SWAP_PAGER_DEALLOC() -	remove swap metadata from object
821  *
822  *	The swap backing for the object is destroyed.  The code is
823  *	designed such that we can reinstantiate it later, but this
824  *	routine is typically called only when the entire object is
825  *	about to be destroyed.
826  *
827  *	The object must be locked.
828  */
829 static void
830 swap_pager_dealloc(vm_object_t object)
831 {
832 
833 	VM_OBJECT_ASSERT_WLOCKED(object);
834 	KASSERT((object->flags & OBJ_DEAD) != 0, ("dealloc of reachable obj"));
835 
836 	/*
837 	 * Remove from list right away so lookups will fail if we block for
838 	 * pageout completion.
839 	 */
840 	if ((object->flags & OBJ_ANON) == 0 && object->handle != NULL) {
841 		VM_OBJECT_WUNLOCK(object);
842 		sx_xlock(&sw_alloc_sx);
843 		TAILQ_REMOVE(NOBJLIST(object->handle), object,
844 		    pager_object_list);
845 		sx_xunlock(&sw_alloc_sx);
846 		VM_OBJECT_WLOCK(object);
847 	}
848 
849 	vm_object_pip_wait(object, "swpdea");
850 
851 	/*
852 	 * Free all remaining metadata.  We only bother to free it from
853 	 * the swap meta data.  We do not attempt to free swapblk's still
854 	 * associated with vm_page_t's for this object.  We do not care
855 	 * if paging is still in progress on some objects.
856 	 */
857 	swp_pager_meta_free_all(object);
858 	object->handle = NULL;
859 	object->type = OBJT_DEAD;
860 
861 	/*
862 	 * Release the allocation charge.
863 	 */
864 	if (object->cred != NULL) {
865 		swap_release_by_cred(object->charge, object->cred);
866 		object->charge = 0;
867 		crfree(object->cred);
868 		object->cred = NULL;
869 	}
870 
871 	/*
872 	 * Hide the object from swap_pager_swapoff().
873 	 */
874 	vm_object_clear_flag(object, OBJ_SWAP);
875 }
876 
877 /************************************************************************
878  *			SWAP PAGER BITMAP ROUTINES			*
879  ************************************************************************/
880 
881 /*
882  * SWP_PAGER_GETSWAPSPACE() -	allocate raw swap space
883  *
884  *	Allocate swap for up to the requested number of pages.  The
885  *	starting swap block number (a page index) is returned or
886  *	SWAPBLK_NONE if the allocation failed.
887  *
888  *	Also has the side effect of advising that somebody made a mistake
889  *	when they configured swap and didn't configure enough.
890  *
891  *	This routine may not sleep.
892  *
893  *	We allocate in round-robin fashion from the configured devices.
894  */
895 static daddr_t
896 swp_pager_getswapspace(int *io_npages)
897 {
898 	daddr_t blk;
899 	struct swdevt *sp;
900 	int mpages, npages;
901 
902 	KASSERT(*io_npages >= 1,
903 	    ("%s: npages not positive", __func__));
904 	blk = SWAPBLK_NONE;
905 	mpages = *io_npages;
906 	npages = imin(BLIST_MAX_ALLOC, mpages);
907 	mtx_lock(&sw_dev_mtx);
908 	sp = swdevhd;
909 	while (!TAILQ_EMPTY(&swtailq)) {
910 		if (sp == NULL)
911 			sp = TAILQ_FIRST(&swtailq);
912 		if ((sp->sw_flags & SW_CLOSING) == 0)
913 			blk = blist_alloc(sp->sw_blist, &npages, mpages);
914 		if (blk != SWAPBLK_NONE)
915 			break;
916 		sp = TAILQ_NEXT(sp, sw_list);
917 		if (swdevhd == sp) {
918 			if (npages == 1)
919 				break;
920 			mpages = npages - 1;
921 			npages >>= 1;
922 		}
923 	}
924 	if (blk != SWAPBLK_NONE) {
925 		*io_npages = npages;
926 		blk += sp->sw_first;
927 		sp->sw_used += npages;
928 		swap_pager_avail -= npages;
929 		swp_sizecheck();
930 		swdevhd = TAILQ_NEXT(sp, sw_list);
931 	} else {
932 		if (swap_pager_full != 2) {
933 			printf("swp_pager_getswapspace(%d): failed\n",
934 			    *io_npages);
935 			swap_pager_full = 2;
936 			swap_pager_almost_full = 1;
937 		}
938 		swdevhd = NULL;
939 	}
940 	mtx_unlock(&sw_dev_mtx);
941 	return (blk);
942 }
943 
944 static bool
945 swp_pager_isondev(daddr_t blk, struct swdevt *sp)
946 {
947 
948 	return (blk >= sp->sw_first && blk < sp->sw_end);
949 }
950 
951 static void
952 swp_pager_strategy(struct buf *bp)
953 {
954 	struct swdevt *sp;
955 
956 	mtx_lock(&sw_dev_mtx);
957 	TAILQ_FOREACH(sp, &swtailq, sw_list) {
958 		if (swp_pager_isondev(bp->b_blkno, sp)) {
959 			mtx_unlock(&sw_dev_mtx);
960 			if ((sp->sw_flags & SW_UNMAPPED) != 0 &&
961 			    unmapped_buf_allowed) {
962 				bp->b_data = unmapped_buf;
963 				bp->b_offset = 0;
964 			} else {
965 				pmap_qenter((vm_offset_t)bp->b_data,
966 				    &bp->b_pages[0], bp->b_bcount / PAGE_SIZE);
967 			}
968 			sp->sw_strategy(bp, sp);
969 			return;
970 		}
971 	}
972 	panic("Swapdev not found");
973 }
974 
975 /*
976  * SWP_PAGER_FREESWAPSPACE() -	free raw swap space
977  *
978  *	This routine returns the specified swap blocks back to the bitmap.
979  *
980  *	This routine may not sleep.
981  */
982 static void
983 swp_pager_freeswapspace(const struct page_range *range)
984 {
985 	daddr_t blk, npages;
986 	struct swdevt *sp;
987 
988 	blk = range->start;
989 	npages = range->num;
990 	if (npages == 0)
991 		return;
992 	mtx_lock(&sw_dev_mtx);
993 	TAILQ_FOREACH(sp, &swtailq, sw_list) {
994 		if (swp_pager_isondev(blk, sp)) {
995 			sp->sw_used -= npages;
996 			/*
997 			 * If we are attempting to stop swapping on
998 			 * this device, we don't want to mark any
999 			 * blocks free lest they be reused.
1000 			 */
1001 			if ((sp->sw_flags & SW_CLOSING) == 0) {
1002 				blist_free(sp->sw_blist, blk - sp->sw_first,
1003 				    npages);
1004 				swap_pager_avail += npages;
1005 				swp_sizecheck();
1006 			}
1007 			mtx_unlock(&sw_dev_mtx);
1008 			return;
1009 		}
1010 	}
1011 	panic("Swapdev not found");
1012 }
1013 
1014 /*
1015  * SYSCTL_SWAP_FRAGMENTATION() -	produce raw swap space stats
1016  */
1017 static int
1018 sysctl_swap_fragmentation(SYSCTL_HANDLER_ARGS)
1019 {
1020 	struct sbuf sbuf;
1021 	struct swdevt *sp;
1022 	const char *devname;
1023 	int error;
1024 
1025 	error = sysctl_wire_old_buffer(req, 0);
1026 	if (error != 0)
1027 		return (error);
1028 	sbuf_new_for_sysctl(&sbuf, NULL, 128, req);
1029 	mtx_lock(&sw_dev_mtx);
1030 	TAILQ_FOREACH(sp, &swtailq, sw_list) {
1031 		if (vn_isdisk(sp->sw_vp))
1032 			devname = devtoname(sp->sw_vp->v_rdev);
1033 		else
1034 			devname = "[file]";
1035 		sbuf_printf(&sbuf, "\nFree space on device %s:\n", devname);
1036 		blist_stats(sp->sw_blist, &sbuf);
1037 	}
1038 	mtx_unlock(&sw_dev_mtx);
1039 	error = sbuf_finish(&sbuf);
1040 	sbuf_delete(&sbuf);
1041 	return (error);
1042 }
1043 
1044 /*
1045  * SWAP_PAGER_FREESPACE() -	frees swap blocks associated with a page
1046  *				range within an object.
1047  *
1048  *	This routine removes swapblk assignments from swap metadata.
1049  *
1050  *	The external callers of this routine typically have already destroyed
1051  *	or renamed vm_page_t's associated with this range in the object so
1052  *	we should be ok.
1053  *
1054  *	The object must be locked.
1055  */
1056 void
1057 swap_pager_freespace(vm_object_t object, vm_pindex_t start, vm_size_t size,
1058     vm_size_t *freed)
1059 {
1060 	MPASS((object->flags & OBJ_SWAP) != 0);
1061 
1062 	swp_pager_meta_free(object, start, size, freed);
1063 }
1064 
1065 static void
1066 swap_pager_freespace_pgo(vm_object_t object, vm_pindex_t start, vm_size_t size)
1067 {
1068 	MPASS((object->flags & OBJ_SWAP) != 0);
1069 
1070 	swp_pager_meta_free(object, start, size, NULL);
1071 }
1072 
1073 /*
1074  * SWAP_PAGER_RESERVE() - reserve swap blocks in object
1075  *
1076  *	Assigns swap blocks to the specified range within the object.  The
1077  *	swap blocks are not zeroed.  Any previous swap assignment is destroyed.
1078  *
1079  *	Returns 0 on success, -1 on failure.
1080  */
1081 int
1082 swap_pager_reserve(vm_object_t object, vm_pindex_t start, vm_pindex_t size)
1083 {
1084 	struct page_range range;
1085 	daddr_t addr, blk;
1086 	vm_pindex_t i, j;
1087 	int n;
1088 
1089 	swp_pager_init_freerange(&range);
1090 	VM_OBJECT_WLOCK(object);
1091 	for (i = 0; i < size; i += n) {
1092 		n = MIN(size - i, INT_MAX);
1093 		blk = swp_pager_getswapspace(&n);
1094 		if (blk == SWAPBLK_NONE) {
1095 			swp_pager_meta_free(object, start, i, NULL);
1096 			VM_OBJECT_WUNLOCK(object);
1097 			return (-1);
1098 		}
1099 		for (j = 0; j < n; ++j) {
1100 			addr = swp_pager_meta_build(object,
1101 			    start + i + j, blk + j, false);
1102 			if (addr != SWAPBLK_NONE)
1103 				swp_pager_update_freerange(&range, addr);
1104 		}
1105 	}
1106 	swp_pager_freeswapspace(&range);
1107 	VM_OBJECT_WUNLOCK(object);
1108 	return (0);
1109 }
1110 
1111 /*
1112  * SWAP_PAGER_COPY() -  copy blocks from source pager to destination pager
1113  *			and destroy the source.
1114  *
1115  *	Copy any valid swapblks from the source to the destination.  In
1116  *	cases where both the source and destination have a valid swapblk,
1117  *	we keep the destination's.
1118  *
1119  *	This routine is allowed to sleep.  It may sleep allocating metadata
1120  *	indirectly through swp_pager_meta_build().
1121  *
1122  *	The source object contains no vm_page_t's (which is just as well)
1123  *
1124  *	The source and destination objects must be locked.
1125  *	Both object locks may temporarily be released.
1126  */
1127 void
1128 swap_pager_copy(vm_object_t srcobject, vm_object_t dstobject,
1129     vm_pindex_t offset, int destroysource)
1130 {
1131 	VM_OBJECT_ASSERT_WLOCKED(srcobject);
1132 	VM_OBJECT_ASSERT_WLOCKED(dstobject);
1133 
1134 	/*
1135 	 * If destroysource is set, we remove the source object from the
1136 	 * swap_pager internal queue now.
1137 	 */
1138 	if (destroysource && (srcobject->flags & OBJ_ANON) == 0 &&
1139 	    srcobject->handle != NULL) {
1140 		VM_OBJECT_WUNLOCK(srcobject);
1141 		VM_OBJECT_WUNLOCK(dstobject);
1142 		sx_xlock(&sw_alloc_sx);
1143 		TAILQ_REMOVE(NOBJLIST(srcobject->handle), srcobject,
1144 		    pager_object_list);
1145 		sx_xunlock(&sw_alloc_sx);
1146 		VM_OBJECT_WLOCK(dstobject);
1147 		VM_OBJECT_WLOCK(srcobject);
1148 	}
1149 
1150 	/*
1151 	 * Transfer source to destination.
1152 	 */
1153 	swp_pager_meta_transfer(srcobject, dstobject, offset, dstobject->size);
1154 
1155 	/*
1156 	 * Free left over swap blocks in source.
1157 	 */
1158 	if (destroysource)
1159 		swp_pager_meta_free_all(srcobject);
1160 }
1161 
1162 /*
1163  * SWAP_PAGER_HASPAGE() -	determine if we have good backing store for
1164  *				the requested page.
1165  *
1166  *	We determine whether good backing store exists for the requested
1167  *	page and return TRUE if it does, FALSE if it doesn't.
1168  *
1169  *	If TRUE, we also try to determine how much valid, contiguous backing
1170  *	store exists before and after the requested page.
1171  */
1172 static boolean_t
1173 swap_pager_haspage(vm_object_t object, vm_pindex_t pindex, int *before,
1174     int *after)
1175 {
1176 	daddr_t blk, blk0;
1177 	int i;
1178 
1179 	VM_OBJECT_ASSERT_LOCKED(object);
1180 	KASSERT((object->flags & OBJ_SWAP) != 0,
1181 	    ("%s: object not swappable", __func__));
1182 
1183 	/*
1184 	 * do we have good backing store at the requested index ?
1185 	 */
1186 	blk0 = swp_pager_meta_lookup(object, pindex);
1187 	if (blk0 == SWAPBLK_NONE) {
1188 		if (before)
1189 			*before = 0;
1190 		if (after)
1191 			*after = 0;
1192 		return (FALSE);
1193 	}
1194 
1195 	/*
1196 	 * find backwards-looking contiguous good backing store
1197 	 */
1198 	if (before != NULL) {
1199 		for (i = 1; i < SWB_NPAGES; i++) {
1200 			if (i > pindex)
1201 				break;
1202 			blk = swp_pager_meta_lookup(object, pindex - i);
1203 			if (blk != blk0 - i)
1204 				break;
1205 		}
1206 		*before = i - 1;
1207 	}
1208 
1209 	/*
1210 	 * find forward-looking contiguous good backing store
1211 	 */
1212 	if (after != NULL) {
1213 		for (i = 1; i < SWB_NPAGES; i++) {
1214 			blk = swp_pager_meta_lookup(object, pindex + i);
1215 			if (blk != blk0 + i)
1216 				break;
1217 		}
1218 		*after = i - 1;
1219 	}
1220 	return (TRUE);
1221 }
1222 
1223 static void
1224 swap_pager_unswapped_acct(vm_page_t m)
1225 {
1226 	KASSERT((m->object->flags & OBJ_SWAP) != 0,
1227 	    ("Free object not swappable"));
1228 	if ((m->a.flags & PGA_SWAP_FREE) != 0)
1229 		counter_u64_add(swap_free_completed, 1);
1230 	vm_page_aflag_clear(m, PGA_SWAP_FREE | PGA_SWAP_SPACE);
1231 
1232 	/*
1233 	 * The meta data only exists if the object is OBJT_SWAP
1234 	 * and even then might not be allocated yet.
1235 	 */
1236 }
1237 
1238 /*
1239  * SWAP_PAGER_PAGE_UNSWAPPED() - remove swap backing store related to page
1240  *
1241  *	This removes any associated swap backing store, whether valid or
1242  *	not, from the page.
1243  *
1244  *	This routine is typically called when a page is made dirty, at
1245  *	which point any associated swap can be freed.  MADV_FREE also
1246  *	calls us in a special-case situation
1247  *
1248  *	NOTE!!!  If the page is clean and the swap was valid, the caller
1249  *	should make the page dirty before calling this routine.  This routine
1250  *	does NOT change the m->dirty status of the page.  Also: MADV_FREE
1251  *	depends on it.
1252  *
1253  *	This routine may not sleep.
1254  *
1255  *	The object containing the page may be locked.
1256  */
1257 static void
1258 swap_pager_unswapped(vm_page_t m)
1259 {
1260 	struct page_range range;
1261 	struct swblk *sb;
1262 	vm_object_t obj;
1263 
1264 	/*
1265 	 * Handle enqueing deferred frees first.  If we do not have the
1266 	 * object lock we wait for the page daemon to clear the space.
1267 	 */
1268 	obj = m->object;
1269 	if (!VM_OBJECT_WOWNED(obj)) {
1270 		VM_PAGE_OBJECT_BUSY_ASSERT(m);
1271 		/*
1272 		 * The caller is responsible for synchronization but we
1273 		 * will harmlessly handle races.  This is typically provided
1274 		 * by only calling unswapped() when a page transitions from
1275 		 * clean to dirty.
1276 		 */
1277 		if ((m->a.flags & (PGA_SWAP_SPACE | PGA_SWAP_FREE)) ==
1278 		    PGA_SWAP_SPACE) {
1279 			vm_page_aflag_set(m, PGA_SWAP_FREE);
1280 			counter_u64_add(swap_free_deferred, 1);
1281 		}
1282 		return;
1283 	}
1284 	swap_pager_unswapped_acct(m);
1285 
1286 	sb = swblk_lookup(m->object, m->pindex);
1287 	if (sb == NULL)
1288 		return;
1289 	range.start = sb->d[m->pindex % SWAP_META_PAGES];
1290 	if (range.start == SWAPBLK_NONE)
1291 		return;
1292 	range.num = 1;
1293 	swp_pager_freeswapspace(&range);
1294 	sb->d[m->pindex % SWAP_META_PAGES] = SWAPBLK_NONE;
1295 	swp_pager_free_empty_swblk(m->object, sb);
1296 }
1297 
1298 /*
1299  * swap_pager_getpages() - bring pages in from swap
1300  *
1301  *	Attempt to page in the pages in array "ma" of length "count".  The
1302  *	caller may optionally specify that additional pages preceding and
1303  *	succeeding the specified range be paged in.  The number of such pages
1304  *	is returned in the "rbehind" and "rahead" parameters, and they will
1305  *	be in the inactive queue upon return.
1306  *
1307  *	The pages in "ma" must be busied and will remain busied upon return.
1308  */
1309 static int
1310 swap_pager_getpages_locked(vm_object_t object, vm_page_t *ma, int count,
1311     int *rbehind, int *rahead)
1312 {
1313 	struct buf *bp;
1314 	vm_page_t bm, mpred, msucc, p;
1315 	vm_pindex_t pindex;
1316 	daddr_t blk;
1317 	int i, maxahead, maxbehind, reqcount;
1318 
1319 	VM_OBJECT_ASSERT_WLOCKED(object);
1320 	reqcount = count;
1321 
1322 	KASSERT((object->flags & OBJ_SWAP) != 0,
1323 	    ("%s: object not swappable", __func__));
1324 	if (!swap_pager_haspage(object, ma[0]->pindex, &maxbehind, &maxahead)) {
1325 		VM_OBJECT_WUNLOCK(object);
1326 		return (VM_PAGER_FAIL);
1327 	}
1328 
1329 	KASSERT(reqcount - 1 <= maxahead,
1330 	    ("page count %d extends beyond swap block", reqcount));
1331 
1332 	/*
1333 	 * Do not transfer any pages other than those that are xbusied
1334 	 * when running during a split or collapse operation.  This
1335 	 * prevents clustering from re-creating pages which are being
1336 	 * moved into another object.
1337 	 */
1338 	if ((object->flags & (OBJ_SPLIT | OBJ_DEAD)) != 0) {
1339 		maxahead = reqcount - 1;
1340 		maxbehind = 0;
1341 	}
1342 
1343 	/*
1344 	 * Clip the readahead and readbehind ranges to exclude resident pages.
1345 	 */
1346 	if (rahead != NULL) {
1347 		*rahead = imin(*rahead, maxahead - (reqcount - 1));
1348 		pindex = ma[reqcount - 1]->pindex;
1349 		msucc = TAILQ_NEXT(ma[reqcount - 1], listq);
1350 		if (msucc != NULL && msucc->pindex - pindex - 1 < *rahead)
1351 			*rahead = msucc->pindex - pindex - 1;
1352 	}
1353 	if (rbehind != NULL) {
1354 		*rbehind = imin(*rbehind, maxbehind);
1355 		pindex = ma[0]->pindex;
1356 		mpred = TAILQ_PREV(ma[0], pglist, listq);
1357 		if (mpred != NULL && pindex - mpred->pindex - 1 < *rbehind)
1358 			*rbehind = pindex - mpred->pindex - 1;
1359 	}
1360 
1361 	bm = ma[0];
1362 	for (i = 0; i < count; i++)
1363 		ma[i]->oflags |= VPO_SWAPINPROG;
1364 
1365 	/*
1366 	 * Allocate readahead and readbehind pages.
1367 	 */
1368 	if (rbehind != NULL) {
1369 		for (i = 1; i <= *rbehind; i++) {
1370 			p = vm_page_alloc(object, ma[0]->pindex - i,
1371 			    VM_ALLOC_NORMAL);
1372 			if (p == NULL)
1373 				break;
1374 			p->oflags |= VPO_SWAPINPROG;
1375 			bm = p;
1376 		}
1377 		*rbehind = i - 1;
1378 	}
1379 	if (rahead != NULL) {
1380 		for (i = 0; i < *rahead; i++) {
1381 			p = vm_page_alloc(object,
1382 			    ma[reqcount - 1]->pindex + i + 1, VM_ALLOC_NORMAL);
1383 			if (p == NULL)
1384 				break;
1385 			p->oflags |= VPO_SWAPINPROG;
1386 		}
1387 		*rahead = i;
1388 	}
1389 	if (rbehind != NULL)
1390 		count += *rbehind;
1391 	if (rahead != NULL)
1392 		count += *rahead;
1393 
1394 	vm_object_pip_add(object, count);
1395 
1396 	pindex = bm->pindex;
1397 	blk = swp_pager_meta_lookup(object, pindex);
1398 	KASSERT(blk != SWAPBLK_NONE,
1399 	    ("no swap blocking containing %p(%jx)", object, (uintmax_t)pindex));
1400 
1401 	VM_OBJECT_WUNLOCK(object);
1402 	bp = uma_zalloc(swrbuf_zone, M_WAITOK);
1403 	MPASS((bp->b_flags & B_MAXPHYS) != 0);
1404 	/* Pages cannot leave the object while busy. */
1405 	for (i = 0, p = bm; i < count; i++, p = TAILQ_NEXT(p, listq)) {
1406 		MPASS(p->pindex == bm->pindex + i);
1407 		bp->b_pages[i] = p;
1408 	}
1409 
1410 	bp->b_flags |= B_PAGING;
1411 	bp->b_iocmd = BIO_READ;
1412 	bp->b_iodone = swp_pager_async_iodone;
1413 	bp->b_rcred = crhold(thread0.td_ucred);
1414 	bp->b_wcred = crhold(thread0.td_ucred);
1415 	bp->b_blkno = blk;
1416 	bp->b_bcount = PAGE_SIZE * count;
1417 	bp->b_bufsize = PAGE_SIZE * count;
1418 	bp->b_npages = count;
1419 	bp->b_pgbefore = rbehind != NULL ? *rbehind : 0;
1420 	bp->b_pgafter = rahead != NULL ? *rahead : 0;
1421 
1422 	VM_CNT_INC(v_swapin);
1423 	VM_CNT_ADD(v_swappgsin, count);
1424 
1425 	/*
1426 	 * perform the I/O.  NOTE!!!  bp cannot be considered valid after
1427 	 * this point because we automatically release it on completion.
1428 	 * Instead, we look at the one page we are interested in which we
1429 	 * still hold a lock on even through the I/O completion.
1430 	 *
1431 	 * The other pages in our ma[] array are also released on completion,
1432 	 * so we cannot assume they are valid anymore either.
1433 	 *
1434 	 * NOTE: b_blkno is destroyed by the call to swapdev_strategy
1435 	 */
1436 	BUF_KERNPROC(bp);
1437 	swp_pager_strategy(bp);
1438 
1439 	/*
1440 	 * Wait for the pages we want to complete.  VPO_SWAPINPROG is always
1441 	 * cleared on completion.  If an I/O error occurs, SWAPBLK_NONE
1442 	 * is set in the metadata for each page in the request.
1443 	 */
1444 	VM_OBJECT_WLOCK(object);
1445 	/* This could be implemented more efficiently with aflags */
1446 	while ((ma[0]->oflags & VPO_SWAPINPROG) != 0) {
1447 		ma[0]->oflags |= VPO_SWAPSLEEP;
1448 		VM_CNT_INC(v_intrans);
1449 		if (VM_OBJECT_SLEEP(object, &object->handle, PSWP,
1450 		    "swread", hz * 20)) {
1451 			printf(
1452 "swap_pager: indefinite wait buffer: bufobj: %p, blkno: %jd, size: %ld\n",
1453 			    bp->b_bufobj, (intmax_t)bp->b_blkno, bp->b_bcount);
1454 		}
1455 	}
1456 	VM_OBJECT_WUNLOCK(object);
1457 
1458 	/*
1459 	 * If we had an unrecoverable read error pages will not be valid.
1460 	 */
1461 	for (i = 0; i < reqcount; i++)
1462 		if (ma[i]->valid != VM_PAGE_BITS_ALL)
1463 			return (VM_PAGER_ERROR);
1464 
1465 	return (VM_PAGER_OK);
1466 
1467 	/*
1468 	 * A final note: in a low swap situation, we cannot deallocate swap
1469 	 * and mark a page dirty here because the caller is likely to mark
1470 	 * the page clean when we return, causing the page to possibly revert
1471 	 * to all-zero's later.
1472 	 */
1473 }
1474 
1475 static int
1476 swap_pager_getpages(vm_object_t object, vm_page_t *ma, int count,
1477     int *rbehind, int *rahead)
1478 {
1479 
1480 	VM_OBJECT_WLOCK(object);
1481 	return (swap_pager_getpages_locked(object, ma, count, rbehind, rahead));
1482 }
1483 
1484 /*
1485  * 	swap_pager_getpages_async():
1486  *
1487  *	Right now this is emulation of asynchronous operation on top of
1488  *	swap_pager_getpages().
1489  */
1490 static int
1491 swap_pager_getpages_async(vm_object_t object, vm_page_t *ma, int count,
1492     int *rbehind, int *rahead, pgo_getpages_iodone_t iodone, void *arg)
1493 {
1494 	int r, error;
1495 
1496 	r = swap_pager_getpages(object, ma, count, rbehind, rahead);
1497 	switch (r) {
1498 	case VM_PAGER_OK:
1499 		error = 0;
1500 		break;
1501 	case VM_PAGER_ERROR:
1502 		error = EIO;
1503 		break;
1504 	case VM_PAGER_FAIL:
1505 		error = EINVAL;
1506 		break;
1507 	default:
1508 		panic("unhandled swap_pager_getpages() error %d", r);
1509 	}
1510 	(iodone)(arg, ma, count, error);
1511 
1512 	return (r);
1513 }
1514 
1515 /*
1516  *	swap_pager_putpages:
1517  *
1518  *	Assign swap (if necessary) and initiate I/O on the specified pages.
1519  *
1520  *	In a low memory situation we may block in VOP_STRATEGY(), but the new
1521  *	vm_page reservation system coupled with properly written VFS devices
1522  *	should ensure that no low-memory deadlock occurs.  This is an area
1523  *	which needs work.
1524  *
1525  *	The parent has N vm_object_pip_add() references prior to
1526  *	calling us and will remove references for rtvals[] that are
1527  *	not set to VM_PAGER_PEND.  We need to remove the rest on I/O
1528  *	completion.
1529  *
1530  *	The parent has soft-busy'd the pages it passes us and will unbusy
1531  *	those whose rtvals[] entry is not set to VM_PAGER_PEND on return.
1532  *	We need to unbusy the rest on I/O completion.
1533  */
1534 static void
1535 swap_pager_putpages(vm_object_t object, vm_page_t *ma, int count,
1536     int flags, int *rtvals)
1537 {
1538 	struct page_range range;
1539 	struct buf *bp;
1540 	daddr_t addr, blk;
1541 	vm_page_t mreq;
1542 	int i, j, n;
1543 	bool async;
1544 
1545 	KASSERT(count == 0 || ma[0]->object == object,
1546 	    ("%s: object mismatch %p/%p",
1547 	    __func__, object, ma[0]->object));
1548 
1549 	VM_OBJECT_WUNLOCK(object);
1550 	async = curproc == pageproc && (flags & VM_PAGER_PUT_SYNC) == 0;
1551 	swp_pager_init_freerange(&range);
1552 
1553 	/*
1554 	 * Assign swap blocks and issue I/O.  We reallocate swap on the fly.
1555 	 * The page is left dirty until the pageout operation completes
1556 	 * successfully.
1557 	 */
1558 	for (i = 0; i < count; i += n) {
1559 		/* Maximum I/O size is limited by maximum swap block size. */
1560 		n = min(count - i, nsw_cluster_max);
1561 
1562 		if (async) {
1563 			mtx_lock(&swbuf_mtx);
1564 			while (nsw_wcount_async == 0)
1565 				msleep(&nsw_wcount_async, &swbuf_mtx, PVM,
1566 				    "swbufa", 0);
1567 			nsw_wcount_async--;
1568 			mtx_unlock(&swbuf_mtx);
1569 		}
1570 
1571 		/* Get a block of swap of size up to size n. */
1572 		blk = swp_pager_getswapspace(&n);
1573 		if (blk == SWAPBLK_NONE) {
1574 			mtx_lock(&swbuf_mtx);
1575 			if (++nsw_wcount_async == 1)
1576 				wakeup(&nsw_wcount_async);
1577 			mtx_unlock(&swbuf_mtx);
1578 			for (j = 0; j < n; ++j)
1579 				rtvals[i + j] = VM_PAGER_FAIL;
1580 			continue;
1581 		}
1582 		VM_OBJECT_WLOCK(object);
1583 		for (j = 0; j < n; ++j) {
1584 			mreq = ma[i + j];
1585 			vm_page_aflag_clear(mreq, PGA_SWAP_FREE);
1586 			addr = swp_pager_meta_build(mreq->object, mreq->pindex,
1587 			    blk + j, false);
1588 			if (addr != SWAPBLK_NONE)
1589 				swp_pager_update_freerange(&range, addr);
1590 			MPASS(mreq->dirty == VM_PAGE_BITS_ALL);
1591 			mreq->oflags |= VPO_SWAPINPROG;
1592 		}
1593 		VM_OBJECT_WUNLOCK(object);
1594 
1595 		bp = uma_zalloc(swwbuf_zone, M_WAITOK);
1596 		MPASS((bp->b_flags & B_MAXPHYS) != 0);
1597 		if (async)
1598 			bp->b_flags |= B_ASYNC;
1599 		bp->b_flags |= B_PAGING;
1600 		bp->b_iocmd = BIO_WRITE;
1601 
1602 		bp->b_rcred = crhold(thread0.td_ucred);
1603 		bp->b_wcred = crhold(thread0.td_ucred);
1604 		bp->b_bcount = PAGE_SIZE * n;
1605 		bp->b_bufsize = PAGE_SIZE * n;
1606 		bp->b_blkno = blk;
1607 		for (j = 0; j < n; j++)
1608 			bp->b_pages[j] = ma[i + j];
1609 		bp->b_npages = n;
1610 
1611 		/*
1612 		 * Must set dirty range for NFS to work.
1613 		 */
1614 		bp->b_dirtyoff = 0;
1615 		bp->b_dirtyend = bp->b_bcount;
1616 
1617 		VM_CNT_INC(v_swapout);
1618 		VM_CNT_ADD(v_swappgsout, bp->b_npages);
1619 
1620 		/*
1621 		 * We unconditionally set rtvals[] to VM_PAGER_PEND so that we
1622 		 * can call the async completion routine at the end of a
1623 		 * synchronous I/O operation.  Otherwise, our caller would
1624 		 * perform duplicate unbusy and wakeup operations on the page
1625 		 * and object, respectively.
1626 		 */
1627 		for (j = 0; j < n; j++)
1628 			rtvals[i + j] = VM_PAGER_PEND;
1629 
1630 		/*
1631 		 * asynchronous
1632 		 *
1633 		 * NOTE: b_blkno is destroyed by the call to swapdev_strategy.
1634 		 */
1635 		if (async) {
1636 			bp->b_iodone = swp_pager_async_iodone;
1637 			BUF_KERNPROC(bp);
1638 			swp_pager_strategy(bp);
1639 			continue;
1640 		}
1641 
1642 		/*
1643 		 * synchronous
1644 		 *
1645 		 * NOTE: b_blkno is destroyed by the call to swapdev_strategy.
1646 		 */
1647 		bp->b_iodone = bdone;
1648 		swp_pager_strategy(bp);
1649 
1650 		/*
1651 		 * Wait for the sync I/O to complete.
1652 		 */
1653 		bwait(bp, PVM, "swwrt");
1654 
1655 		/*
1656 		 * Now that we are through with the bp, we can call the
1657 		 * normal async completion, which frees everything up.
1658 		 */
1659 		swp_pager_async_iodone(bp);
1660 	}
1661 	swp_pager_freeswapspace(&range);
1662 	VM_OBJECT_WLOCK(object);
1663 }
1664 
1665 /*
1666  *	swp_pager_async_iodone:
1667  *
1668  *	Completion routine for asynchronous reads and writes from/to swap.
1669  *	Also called manually by synchronous code to finish up a bp.
1670  *
1671  *	This routine may not sleep.
1672  */
1673 static void
1674 swp_pager_async_iodone(struct buf *bp)
1675 {
1676 	int i;
1677 	vm_object_t object = NULL;
1678 
1679 	/*
1680 	 * Report error - unless we ran out of memory, in which case
1681 	 * we've already logged it in swapgeom_strategy().
1682 	 */
1683 	if (bp->b_ioflags & BIO_ERROR && bp->b_error != ENOMEM) {
1684 		printf(
1685 		    "swap_pager: I/O error - %s failed; blkno %ld,"
1686 			"size %ld, error %d\n",
1687 		    ((bp->b_iocmd == BIO_READ) ? "pagein" : "pageout"),
1688 		    (long)bp->b_blkno,
1689 		    (long)bp->b_bcount,
1690 		    bp->b_error
1691 		);
1692 	}
1693 
1694 	/*
1695 	 * remove the mapping for kernel virtual
1696 	 */
1697 	if (buf_mapped(bp))
1698 		pmap_qremove((vm_offset_t)bp->b_data, bp->b_npages);
1699 	else
1700 		bp->b_data = bp->b_kvabase;
1701 
1702 	if (bp->b_npages) {
1703 		object = bp->b_pages[0]->object;
1704 		VM_OBJECT_WLOCK(object);
1705 	}
1706 
1707 	/*
1708 	 * cleanup pages.  If an error occurs writing to swap, we are in
1709 	 * very serious trouble.  If it happens to be a disk error, though,
1710 	 * we may be able to recover by reassigning the swap later on.  So
1711 	 * in this case we remove the m->swapblk assignment for the page
1712 	 * but do not free it in the rlist.  The errornous block(s) are thus
1713 	 * never reallocated as swap.  Redirty the page and continue.
1714 	 */
1715 	for (i = 0; i < bp->b_npages; ++i) {
1716 		vm_page_t m = bp->b_pages[i];
1717 
1718 		m->oflags &= ~VPO_SWAPINPROG;
1719 		if (m->oflags & VPO_SWAPSLEEP) {
1720 			m->oflags &= ~VPO_SWAPSLEEP;
1721 			wakeup(&object->handle);
1722 		}
1723 
1724 		/* We always have space after I/O, successful or not. */
1725 		vm_page_aflag_set(m, PGA_SWAP_SPACE);
1726 
1727 		if (bp->b_ioflags & BIO_ERROR) {
1728 			/*
1729 			 * If an error occurs I'd love to throw the swapblk
1730 			 * away without freeing it back to swapspace, so it
1731 			 * can never be used again.  But I can't from an
1732 			 * interrupt.
1733 			 */
1734 			if (bp->b_iocmd == BIO_READ) {
1735 				/*
1736 				 * NOTE: for reads, m->dirty will probably
1737 				 * be overridden by the original caller of
1738 				 * getpages so don't play cute tricks here.
1739 				 */
1740 				vm_page_invalid(m);
1741 				if (i < bp->b_pgbefore ||
1742 				    i >= bp->b_npages - bp->b_pgafter)
1743 					vm_page_free_invalid(m);
1744 			} else {
1745 				/*
1746 				 * If a write error occurs, reactivate page
1747 				 * so it doesn't clog the inactive list,
1748 				 * then finish the I/O.
1749 				 */
1750 				MPASS(m->dirty == VM_PAGE_BITS_ALL);
1751 
1752 				/* PQ_UNSWAPPABLE? */
1753 				vm_page_activate(m);
1754 				vm_page_sunbusy(m);
1755 			}
1756 		} else if (bp->b_iocmd == BIO_READ) {
1757 			/*
1758 			 * NOTE: for reads, m->dirty will probably be
1759 			 * overridden by the original caller of getpages so
1760 			 * we cannot set them in order to free the underlying
1761 			 * swap in a low-swap situation.  I don't think we'd
1762 			 * want to do that anyway, but it was an optimization
1763 			 * that existed in the old swapper for a time before
1764 			 * it got ripped out due to precisely this problem.
1765 			 */
1766 			KASSERT(!pmap_page_is_mapped(m),
1767 			    ("swp_pager_async_iodone: page %p is mapped", m));
1768 			KASSERT(m->dirty == 0,
1769 			    ("swp_pager_async_iodone: page %p is dirty", m));
1770 
1771 			vm_page_valid(m);
1772 			if (i < bp->b_pgbefore ||
1773 			    i >= bp->b_npages - bp->b_pgafter)
1774 				vm_page_readahead_finish(m);
1775 		} else {
1776 			/*
1777 			 * For write success, clear the dirty
1778 			 * status, then finish the I/O ( which decrements the
1779 			 * busy count and possibly wakes waiter's up ).
1780 			 * A page is only written to swap after a period of
1781 			 * inactivity.  Therefore, we do not expect it to be
1782 			 * reused.
1783 			 */
1784 			KASSERT(!pmap_page_is_write_mapped(m),
1785 			    ("swp_pager_async_iodone: page %p is not write"
1786 			    " protected", m));
1787 			vm_page_undirty(m);
1788 			vm_page_deactivate_noreuse(m);
1789 			vm_page_sunbusy(m);
1790 		}
1791 	}
1792 
1793 	/*
1794 	 * adjust pip.  NOTE: the original parent may still have its own
1795 	 * pip refs on the object.
1796 	 */
1797 	if (object != NULL) {
1798 		vm_object_pip_wakeupn(object, bp->b_npages);
1799 		VM_OBJECT_WUNLOCK(object);
1800 	}
1801 
1802 	/*
1803 	 * swapdev_strategy() manually sets b_vp and b_bufobj before calling
1804 	 * bstrategy(). Set them back to NULL now we're done with it, or we'll
1805 	 * trigger a KASSERT in relpbuf().
1806 	 */
1807 	if (bp->b_vp) {
1808 		    bp->b_vp = NULL;
1809 		    bp->b_bufobj = NULL;
1810 	}
1811 	/*
1812 	 * release the physical I/O buffer
1813 	 */
1814 	if (bp->b_flags & B_ASYNC) {
1815 		mtx_lock(&swbuf_mtx);
1816 		if (++nsw_wcount_async == 1)
1817 			wakeup(&nsw_wcount_async);
1818 		mtx_unlock(&swbuf_mtx);
1819 	}
1820 	uma_zfree((bp->b_iocmd == BIO_READ) ? swrbuf_zone : swwbuf_zone, bp);
1821 }
1822 
1823 int
1824 swap_pager_nswapdev(void)
1825 {
1826 
1827 	return (nswapdev);
1828 }
1829 
1830 static void
1831 swp_pager_force_dirty(struct page_range *range, vm_page_t m, daddr_t *blk)
1832 {
1833 	vm_page_dirty(m);
1834 	swap_pager_unswapped_acct(m);
1835 	swp_pager_update_freerange(range, *blk);
1836 	*blk = SWAPBLK_NONE;
1837 	vm_page_launder(m);
1838 }
1839 
1840 u_long
1841 swap_pager_swapped_pages(vm_object_t object)
1842 {
1843 	struct pctrie_iter blks;
1844 	struct swblk *sb;
1845 	u_long res;
1846 	int i;
1847 
1848 	VM_OBJECT_ASSERT_LOCKED(object);
1849 
1850 	if (swblk_is_empty(object))
1851 		return (0);
1852 
1853 	res = 0;
1854 	for (sb = swblk_iter(&blks, object, 0); sb != NULL;
1855 	    sb = swblk_iter_next(&blks)) {
1856 		for (i = 0; i < SWAP_META_PAGES; i++) {
1857 			if (sb->d[i] != SWAPBLK_NONE)
1858 				res++;
1859 		}
1860 	}
1861 	return (res);
1862 }
1863 
1864 /*
1865  *	swap_pager_swapoff_object:
1866  *
1867  *	Page in all of the pages that have been paged out for an object
1868  *	to a swap device.
1869  */
1870 static void
1871 swap_pager_swapoff_object(struct swdevt *sp, vm_object_t object)
1872 {
1873 	struct pctrie_iter blks, pages;
1874 	struct page_range range;
1875 	struct swblk *sb;
1876 	vm_page_t m;
1877 	int i, rahead, rv;
1878 	bool sb_empty;
1879 
1880 	VM_OBJECT_ASSERT_WLOCKED(object);
1881 	KASSERT((object->flags & OBJ_SWAP) != 0,
1882 	    ("%s: Object not swappable", __func__));
1883 	KASSERT((sp->sw_flags & SW_CLOSING) != 0,
1884 	    ("%s: Device not blocking further allocations", __func__));
1885 
1886 	vm_page_iter_init(&pages, object);
1887 	swp_pager_init_freerange(&range);
1888 	sb = swblk_iter(&blks, object, 0);
1889 	while (sb != NULL) {
1890 		if ((object->flags & OBJ_DEAD) != 0) {
1891 			/*
1892 			 * Make sure that pending writes finish before
1893 			 * returning.
1894 			 */
1895 			vm_object_pip_wait(object, "swpoff");
1896 			swp_pager_meta_free_all(object);
1897 			break;
1898 		}
1899 		sb_empty = true;
1900 		for (i = 0; i < SWAP_META_PAGES; i++) {
1901 			/* Skip an invalid block. */
1902 			if (sb->d[i] == SWAPBLK_NONE)
1903 				continue;
1904 			/* Skip a block not of this device. */
1905 			if (!swp_pager_isondev(sb->d[i], sp)) {
1906 				sb_empty = false;
1907 				continue;
1908 			}
1909 
1910 			/*
1911 			 * Look for a page corresponding to this block. If the
1912 			 * found page has pending operations, sleep and restart
1913 			 * the scan.
1914 			 */
1915 			m = vm_page_iter_lookup(&pages, blks.index + i);
1916 			if (m != NULL && (m->oflags & VPO_SWAPINPROG) != 0) {
1917 				m->oflags |= VPO_SWAPSLEEP;
1918 				VM_OBJECT_SLEEP(object, &object->handle, PSWP,
1919 				    "swpoff", 0);
1920 				break;
1921 			}
1922 
1923 			/*
1924 			 * If the found page is valid, mark it dirty and free
1925 			 * the swap block.
1926 			 */
1927 			if (m != NULL && vm_page_all_valid(m)) {
1928 				swp_pager_force_dirty(&range, m, &sb->d[i]);
1929 				continue;
1930 			}
1931 			/* Is there a page we can acquire or allocate? */
1932 			if (m != NULL) {
1933 				if (!vm_page_busy_acquire(m, VM_ALLOC_WAITFAIL))
1934 					break;
1935 			} else if ((m = vm_page_alloc(object, blks.index + i,
1936 			    VM_ALLOC_NORMAL | VM_ALLOC_WAITFAIL)) == NULL)
1937 				break;
1938 
1939 			/* Get the page from swap, and restart the scan. */
1940 			vm_object_pip_add(object, 1);
1941 			rahead = SWAP_META_PAGES;
1942 			rv = swap_pager_getpages_locked(object, &m, 1,
1943 			    NULL, &rahead);
1944 			if (rv != VM_PAGER_OK)
1945 				panic("%s: read from swap failed: %d",
1946 				    __func__, rv);
1947 			VM_OBJECT_WLOCK(object);
1948 			vm_object_pip_wakeupn(object, 1);
1949 			KASSERT(vm_page_all_valid(m),
1950 			    ("%s: Page %p not all valid", __func__, m));
1951 			vm_page_xunbusy(m);
1952 			break;
1953 		}
1954 		if (i < SWAP_META_PAGES) {
1955 			/*
1956 			 * With the object lock released and regained, the
1957 			 * swapblk could have been freed, so reset the pages
1958 			 * iterator and search again for the first swblk at or
1959 			 * after blks.index.
1960 			 */
1961 			pctrie_iter_reset(&pages);
1962 			sb = swblk_iter(&blks, object, blks.index);
1963 			continue;
1964 		}
1965 		if (sb_empty) {
1966 			swblk_iter_remove(&blks);
1967 			uma_zfree(swblk_zone, sb);
1968 		}
1969 
1970 		/*
1971 		 * It is safe to advance to the next block.  No allocations
1972 		 * before blk.index have happened, even with the lock released,
1973 		 * because allocations on this device are blocked.
1974 		 */
1975 		sb = swblk_iter_next(&blks);
1976 	}
1977 	swp_pager_freeswapspace(&range);
1978 }
1979 
1980 /*
1981  *	swap_pager_swapoff:
1982  *
1983  *	Page in all of the pages that have been paged out to the
1984  *	given device.  The corresponding blocks in the bitmap must be
1985  *	marked as allocated and the device must be flagged SW_CLOSING.
1986  *	There may be no processes swapped out to the device.
1987  *
1988  *	This routine may block.
1989  */
1990 static void
1991 swap_pager_swapoff(struct swdevt *sp)
1992 {
1993 	vm_object_t object;
1994 	int retries;
1995 
1996 	sx_assert(&swdev_syscall_lock, SA_XLOCKED);
1997 
1998 	retries = 0;
1999 full_rescan:
2000 	mtx_lock(&vm_object_list_mtx);
2001 	TAILQ_FOREACH(object, &vm_object_list, object_list) {
2002 		if ((object->flags & OBJ_SWAP) == 0)
2003 			continue;
2004 		mtx_unlock(&vm_object_list_mtx);
2005 		/* Depends on type-stability. */
2006 		VM_OBJECT_WLOCK(object);
2007 
2008 		/*
2009 		 * Dead objects are eventually terminated on their own.
2010 		 */
2011 		if ((object->flags & OBJ_DEAD) != 0)
2012 			goto next_obj;
2013 
2014 		/*
2015 		 * Sync with fences placed after pctrie
2016 		 * initialization.  We must not access pctrie below
2017 		 * unless we checked that our object is swap and not
2018 		 * dead.
2019 		 */
2020 		atomic_thread_fence_acq();
2021 		if ((object->flags & OBJ_SWAP) == 0)
2022 			goto next_obj;
2023 
2024 		swap_pager_swapoff_object(sp, object);
2025 next_obj:
2026 		VM_OBJECT_WUNLOCK(object);
2027 		mtx_lock(&vm_object_list_mtx);
2028 	}
2029 	mtx_unlock(&vm_object_list_mtx);
2030 
2031 	if (sp->sw_used) {
2032 		/*
2033 		 * Objects may be locked or paging to the device being
2034 		 * removed, so we will miss their pages and need to
2035 		 * make another pass.  We have marked this device as
2036 		 * SW_CLOSING, so the activity should finish soon.
2037 		 */
2038 		retries++;
2039 		if (retries > 100) {
2040 			panic("swapoff: failed to locate %d swap blocks",
2041 			    sp->sw_used);
2042 		}
2043 		pause("swpoff", hz / 20);
2044 		goto full_rescan;
2045 	}
2046 	EVENTHANDLER_INVOKE(swapoff, sp);
2047 }
2048 
2049 /************************************************************************
2050  *				SWAP META DATA 				*
2051  ************************************************************************
2052  *
2053  *	These routines manipulate the swap metadata stored in the
2054  *	OBJT_SWAP object.
2055  *
2056  *	Swap metadata is implemented with a global hash and not directly
2057  *	linked into the object.  Instead the object simply contains
2058  *	appropriate tracking counters.
2059  */
2060 
2061 /*
2062  * SWP_PAGER_SWBLK_EMPTY() - is a range of blocks free?
2063  */
2064 static bool
2065 swp_pager_swblk_empty(struct swblk *sb, int start, int limit)
2066 {
2067 	int i;
2068 
2069 	MPASS(0 <= start && start <= limit && limit <= SWAP_META_PAGES);
2070 	for (i = start; i < limit; i++) {
2071 		if (sb->d[i] != SWAPBLK_NONE)
2072 			return (false);
2073 	}
2074 	return (true);
2075 }
2076 
2077 /*
2078  * SWP_PAGER_FREE_EMPTY_SWBLK() - frees if a block is free
2079  *
2080  *  Nothing is done if the block is still in use.
2081  */
2082 static void
2083 swp_pager_free_empty_swblk(vm_object_t object, struct swblk *sb)
2084 {
2085 
2086 	if (swp_pager_swblk_empty(sb, 0, SWAP_META_PAGES)) {
2087 		swblk_lookup_remove(object, sb);
2088 		uma_zfree(swblk_zone, sb);
2089 	}
2090 }
2091 
2092 /*
2093  * SWP_PAGER_META_BUILD() -	add swap block to swap meta data for object
2094  *
2095  *	Try to add the specified swapblk to the object's swap metadata.  If
2096  *	nowait_noreplace is set, add the specified swapblk only if there is no
2097  *	previously assigned swapblk at pindex.  If the swapblk is invalid, and
2098  *	replaces a valid swapblk, empty swap metadata is freed.  If memory
2099  *	allocation fails, and nowait_noreplace is set, return the specified
2100  *	swapblk immediately to indicate failure; otherwise, wait and retry until
2101  *	memory allocation succeeds.  Return the previously assigned swapblk, if
2102  *	any.
2103  */
2104 static daddr_t
2105 swp_pager_meta_build(vm_object_t object, vm_pindex_t pindex, daddr_t swapblk,
2106     bool nowait_noreplace)
2107 {
2108 	static volatile int swblk_zone_exhausted, swpctrie_zone_exhausted;
2109 	struct swblk *sb, *sb1;
2110 	vm_pindex_t modpi, rdpi;
2111 	daddr_t prev_swapblk;
2112 	int error, i;
2113 
2114 	VM_OBJECT_ASSERT_WLOCKED(object);
2115 
2116 	rdpi = rounddown(pindex, SWAP_META_PAGES);
2117 	sb = swblk_lookup(object, rdpi);
2118 	if (sb == NULL) {
2119 		if (swapblk == SWAPBLK_NONE)
2120 			return (SWAPBLK_NONE);
2121 		for (;;) {
2122 			sb = uma_zalloc(swblk_zone, M_NOWAIT | (curproc ==
2123 			    pageproc ? M_USE_RESERVE : 0));
2124 			if (sb != NULL) {
2125 				sb->p = rdpi;
2126 				for (i = 0; i < SWAP_META_PAGES; i++)
2127 					sb->d[i] = SWAPBLK_NONE;
2128 				if (atomic_cmpset_int(&swblk_zone_exhausted,
2129 				    1, 0))
2130 					printf("swblk zone ok\n");
2131 				break;
2132 			}
2133 			if (nowait_noreplace)
2134 				return (swapblk);
2135 			VM_OBJECT_WUNLOCK(object);
2136 			if (uma_zone_exhausted(swblk_zone)) {
2137 				if (atomic_cmpset_int(&swblk_zone_exhausted,
2138 				    0, 1))
2139 					printf("swap blk zone exhausted, "
2140 					    "increase kern.maxswzone\n");
2141 				vm_pageout_oom(VM_OOM_SWAPZ);
2142 				pause("swzonxb", 10);
2143 			} else
2144 				uma_zwait(swblk_zone);
2145 			VM_OBJECT_WLOCK(object);
2146 			sb = swblk_lookup(object, rdpi);
2147 			if (sb != NULL)
2148 				/*
2149 				 * Somebody swapped out a nearby page,
2150 				 * allocating swblk at the rdpi index,
2151 				 * while we dropped the object lock.
2152 				 */
2153 				goto allocated;
2154 		}
2155 		for (;;) {
2156 			error = swblk_lookup_insert(object, sb);
2157 			if (error == 0) {
2158 				if (atomic_cmpset_int(&swpctrie_zone_exhausted,
2159 				    1, 0))
2160 					printf("swpctrie zone ok\n");
2161 				break;
2162 			}
2163 			if (nowait_noreplace) {
2164 				uma_zfree(swblk_zone, sb);
2165 				return (swapblk);
2166 			}
2167 			VM_OBJECT_WUNLOCK(object);
2168 			if (uma_zone_exhausted(swpctrie_zone)) {
2169 				if (atomic_cmpset_int(&swpctrie_zone_exhausted,
2170 				    0, 1))
2171 					printf("swap pctrie zone exhausted, "
2172 					    "increase kern.maxswzone\n");
2173 				vm_pageout_oom(VM_OOM_SWAPZ);
2174 				pause("swzonxp", 10);
2175 			} else
2176 				uma_zwait(swpctrie_zone);
2177 			VM_OBJECT_WLOCK(object);
2178 			sb1 = swblk_lookup(object, rdpi);
2179 			if (sb1 != NULL) {
2180 				uma_zfree(swblk_zone, sb);
2181 				sb = sb1;
2182 				goto allocated;
2183 			}
2184 		}
2185 	}
2186 allocated:
2187 	MPASS(sb->p == rdpi);
2188 
2189 	modpi = pindex % SWAP_META_PAGES;
2190 	/* Return prior contents of metadata. */
2191 	prev_swapblk = sb->d[modpi];
2192 	if (!nowait_noreplace || prev_swapblk == SWAPBLK_NONE) {
2193 		/* Enter block into metadata. */
2194 		sb->d[modpi] = swapblk;
2195 
2196 		/*
2197 		 * Free the swblk if we end up with the empty page run.
2198 		 */
2199 		if (swapblk == SWAPBLK_NONE)
2200 			swp_pager_free_empty_swblk(object, sb);
2201 	}
2202 	return (prev_swapblk);
2203 }
2204 
2205 /*
2206  * SWP_PAGER_META_TRANSFER() - transfer a range of blocks in the srcobject's
2207  * swap metadata into dstobject.
2208  *
2209  *	Blocks in src that correspond to holes in dst are transferred.  Blocks
2210  *	in src that correspond to blocks in dst are freed.
2211  */
2212 static void
2213 swp_pager_meta_transfer(vm_object_t srcobject, vm_object_t dstobject,
2214     vm_pindex_t pindex, vm_pindex_t count)
2215 {
2216 	struct pctrie_iter blks;
2217 	struct page_range range;
2218 	struct swblk *sb;
2219 	daddr_t blk, d[SWAP_META_PAGES];
2220 	vm_pindex_t last;
2221 	int d_mask, i, limit, start;
2222 	_Static_assert(8 * sizeof(d_mask) >= SWAP_META_PAGES,
2223 	    "d_mask not big enough");
2224 
2225 	VM_OBJECT_ASSERT_WLOCKED(srcobject);
2226 	VM_OBJECT_ASSERT_WLOCKED(dstobject);
2227 
2228 	if (count == 0 || swblk_is_empty(srcobject))
2229 		return;
2230 
2231 	swp_pager_init_freerange(&range);
2232 	d_mask = 0;
2233 	last = pindex + count;
2234 	for (sb = swblk_iter_limit(&blks, srcobject, pindex, last),
2235 	    start = swblk_start(sb, pindex);
2236 	    sb != NULL; sb = swblk_iter_next(&blks), start = 0) {
2237 		limit = MIN(last - blks.index, SWAP_META_PAGES);
2238 		for (i = start; i < limit; i++) {
2239 			if (sb->d[i] == SWAPBLK_NONE)
2240 				continue;
2241 			blk = swp_pager_meta_build(dstobject,
2242 			    blks.index + i - pindex, sb->d[i], true);
2243 			if (blk == sb->d[i]) {
2244 				/*
2245 				 * Failed memory allocation stopped transfer;
2246 				 * save this block for transfer with lock
2247 				 * released.
2248 				 */
2249 				d[i] = blk;
2250 				d_mask |= 1 << i;
2251 			} else if (blk != SWAPBLK_NONE) {
2252 				/* Dst has a block at pindex, so free block. */
2253 				swp_pager_update_freerange(&range, sb->d[i]);
2254 			}
2255 			sb->d[i] = SWAPBLK_NONE;
2256 		}
2257 		if (swp_pager_swblk_empty(sb, 0, start) &&
2258 		    swp_pager_swblk_empty(sb, limit, SWAP_META_PAGES)) {
2259 			swblk_iter_remove(&blks);
2260 			uma_zfree(swblk_zone, sb);
2261 		}
2262 		if (d_mask != 0) {
2263 			/* Finish block transfer, with the lock released. */
2264 			VM_OBJECT_WUNLOCK(srcobject);
2265 			do {
2266 				i = ffs(d_mask) - 1;
2267 				swp_pager_meta_build(dstobject,
2268 				    blks.index + i - pindex, d[i], false);
2269 				d_mask &= ~(1 << i);
2270 			} while (d_mask != 0);
2271 			VM_OBJECT_WLOCK(srcobject);
2272 
2273 			/*
2274 			 * While the lock was not held, the iterator path could
2275 			 * have become stale, so discard it.
2276 			 */
2277 			pctrie_iter_reset(&blks);
2278 		}
2279 	}
2280 	swp_pager_freeswapspace(&range);
2281 }
2282 
2283 /*
2284  * SWP_PAGER_META_FREE() - free a range of blocks in the object's swap metadata
2285  *
2286  *	Return freed swap blocks to the swap bitmap, and free emptied swblk
2287  *	metadata.  With 'freed' set, provide a count of freed blocks that were
2288  *	not associated with valid resident pages.
2289  */
2290 static void
2291 swp_pager_meta_free(vm_object_t object, vm_pindex_t pindex, vm_pindex_t count,
2292     vm_size_t *freed)
2293 {
2294 	struct pctrie_iter blks, pages;
2295 	struct page_range range;
2296 	struct swblk *sb;
2297 	vm_page_t m;
2298 	vm_pindex_t last;
2299 	vm_size_t fc;
2300 	int i, limit, start;
2301 
2302 	VM_OBJECT_ASSERT_WLOCKED(object);
2303 
2304 	fc = 0;
2305 	if (count == 0 || swblk_is_empty(object))
2306 		goto out;
2307 
2308 	swp_pager_init_freerange(&range);
2309 	vm_page_iter_init(&pages, object);
2310 	last = pindex + count;
2311 	for (sb = swblk_iter_limit(&blks, object, pindex, last),
2312 	    start = swblk_start(sb, pindex);
2313 	    sb != NULL; sb = swblk_iter_next(&blks), start = 0) {
2314 		limit = MIN(last - blks.index, SWAP_META_PAGES);
2315 		for (i = start; i < limit; i++) {
2316 			if (sb->d[i] == SWAPBLK_NONE)
2317 				continue;
2318 			swp_pager_update_freerange(&range, sb->d[i]);
2319 			if (freed != NULL) {
2320 				m = vm_page_iter_lookup(&pages, blks.index + i);
2321 				if (m == NULL || vm_page_none_valid(m))
2322 					fc++;
2323 			}
2324 			sb->d[i] = SWAPBLK_NONE;
2325 		}
2326 		if (swp_pager_swblk_empty(sb, 0, start) &&
2327 		    swp_pager_swblk_empty(sb, limit, SWAP_META_PAGES)) {
2328 			swblk_iter_remove(&blks);
2329 			uma_zfree(swblk_zone, sb);
2330 		}
2331 	}
2332 	swp_pager_freeswapspace(&range);
2333 out:
2334 	if (freed != NULL)
2335 		*freed = fc;
2336 }
2337 
2338 static void
2339 swp_pager_meta_free_block(struct swblk *sb, void *rangev)
2340 {
2341 	struct page_range *range = rangev;
2342 
2343 	for (int i = 0; i < SWAP_META_PAGES; i++) {
2344 		if (sb->d[i] != SWAPBLK_NONE)
2345 			swp_pager_update_freerange(range, sb->d[i]);
2346 	}
2347 	uma_zfree(swblk_zone, sb);
2348 }
2349 
2350 /*
2351  * SWP_PAGER_META_FREE_ALL() - destroy all swap metadata associated with object
2352  *
2353  *	This routine locates and destroys all swap metadata associated with
2354  *	an object.
2355  */
2356 static void
2357 swp_pager_meta_free_all(vm_object_t object)
2358 {
2359 	struct page_range range;
2360 
2361 	VM_OBJECT_ASSERT_WLOCKED(object);
2362 
2363 	swp_pager_init_freerange(&range);
2364 	SWAP_PCTRIE_RECLAIM_CALLBACK(&object->un_pager.swp.swp_blks,
2365 	    swp_pager_meta_free_block, &range);
2366 	swp_pager_freeswapspace(&range);
2367 }
2368 
2369 /*
2370  * SWP_PAGER_METACTL() -  misc control of swap meta data.
2371  *
2372  *	This routine is capable of looking up, or removing swapblk
2373  *	assignments in the swap meta data.  It returns the swapblk being
2374  *	looked-up, popped, or SWAPBLK_NONE if the block was invalid.
2375  *
2376  *	When acting on a busy resident page and paging is in progress, we
2377  *	have to wait until paging is complete but otherwise can act on the
2378  *	busy page.
2379  */
2380 static daddr_t
2381 swp_pager_meta_lookup(vm_object_t object, vm_pindex_t pindex)
2382 {
2383 	struct swblk *sb;
2384 
2385 	VM_OBJECT_ASSERT_LOCKED(object);
2386 
2387 	/*
2388 	 * The meta data only exists if the object is OBJT_SWAP
2389 	 * and even then might not be allocated yet.
2390 	 */
2391 	KASSERT((object->flags & OBJ_SWAP) != 0,
2392 	    ("Lookup object not swappable"));
2393 
2394 	sb = swblk_lookup(object, pindex);
2395 	if (sb == NULL)
2396 		return (SWAPBLK_NONE);
2397 	return (sb->d[pindex % SWAP_META_PAGES]);
2398 }
2399 
2400 /*
2401  * Returns the least page index which is greater than or equal to the parameter
2402  * pindex and for which there is a swap block allocated.  Returns OBJ_MAX_SIZE
2403  * if are no allocated swap blocks for the object after the requested pindex.
2404  */
2405 vm_pindex_t
2406 swap_pager_find_least(vm_object_t object, vm_pindex_t pindex)
2407 {
2408 	struct pctrie_iter blks;
2409 	struct swblk *sb;
2410 	int i;
2411 
2412 	if ((sb = swblk_iter(&blks, object, pindex)) == NULL)
2413 		return (OBJ_MAX_SIZE);
2414 	if (blks.index < pindex) {
2415 		for (i = pindex % SWAP_META_PAGES; i < SWAP_META_PAGES; i++) {
2416 			if (sb->d[i] != SWAPBLK_NONE)
2417 				return (blks.index + i);
2418 		}
2419 		if ((sb = swblk_iter_next(&blks)) == NULL)
2420 			return (OBJ_MAX_SIZE);
2421 	}
2422 	for (i = 0; i < SWAP_META_PAGES; i++) {
2423 		if (sb->d[i] != SWAPBLK_NONE)
2424 			return (blks.index + i);
2425 	}
2426 
2427 	/*
2428 	 * We get here if a swblk is present in the trie but it
2429 	 * doesn't map any blocks.
2430 	 */
2431 	MPASS(0);
2432 	return (OBJ_MAX_SIZE);
2433 }
2434 
2435 /*
2436  * System call swapon(name) enables swapping on device name,
2437  * which must be in the swdevsw.  Return EBUSY
2438  * if already swapping on this device.
2439  */
2440 #ifndef _SYS_SYSPROTO_H_
2441 struct swapon_args {
2442 	char *name;
2443 };
2444 #endif
2445 
2446 int
2447 sys_swapon(struct thread *td, struct swapon_args *uap)
2448 {
2449 	struct vattr attr;
2450 	struct vnode *vp;
2451 	struct nameidata nd;
2452 	int error;
2453 
2454 	error = priv_check(td, PRIV_SWAPON);
2455 	if (error)
2456 		return (error);
2457 
2458 	sx_xlock(&swdev_syscall_lock);
2459 
2460 	/*
2461 	 * Swap metadata may not fit in the KVM if we have physical
2462 	 * memory of >1GB.
2463 	 */
2464 	if (swblk_zone == NULL) {
2465 		error = ENOMEM;
2466 		goto done;
2467 	}
2468 
2469 	NDINIT(&nd, LOOKUP, ISOPEN | FOLLOW | LOCKLEAF | AUDITVNODE1,
2470 	    UIO_USERSPACE, uap->name);
2471 	error = namei(&nd);
2472 	if (error)
2473 		goto done;
2474 
2475 	NDFREE_PNBUF(&nd);
2476 	vp = nd.ni_vp;
2477 
2478 	if (vn_isdisk_error(vp, &error)) {
2479 		error = swapongeom(vp);
2480 	} else if (vp->v_type == VREG &&
2481 	    (vp->v_mount->mnt_vfc->vfc_flags & VFCF_NETWORK) != 0 &&
2482 	    (error = VOP_GETATTR(vp, &attr, td->td_ucred)) == 0) {
2483 		/*
2484 		 * Allow direct swapping to NFS regular files in the same
2485 		 * way that nfs_mountroot() sets up diskless swapping.
2486 		 */
2487 		error = swaponvp(td, vp, attr.va_size / DEV_BSIZE);
2488 	}
2489 
2490 	if (error != 0)
2491 		vput(vp);
2492 	else
2493 		VOP_UNLOCK(vp);
2494 done:
2495 	sx_xunlock(&swdev_syscall_lock);
2496 	return (error);
2497 }
2498 
2499 /*
2500  * Check that the total amount of swap currently configured does not
2501  * exceed half the theoretical maximum.  If it does, print a warning
2502  * message.
2503  */
2504 static void
2505 swapon_check_swzone(void)
2506 {
2507 
2508 	/* recommend using no more than half that amount */
2509 	if (swap_total > swap_maxpages / 2) {
2510 		printf("warning: total configured swap (%lu pages) "
2511 		    "exceeds maximum recommended amount (%lu pages).\n",
2512 		    swap_total, swap_maxpages / 2);
2513 		printf("warning: increase kern.maxswzone "
2514 		    "or reduce amount of swap.\n");
2515 	}
2516 }
2517 
2518 static void
2519 swaponsomething(struct vnode *vp, void *id, u_long nblks,
2520     sw_strategy_t *strategy, sw_close_t *close, dev_t dev, int flags)
2521 {
2522 	struct swdevt *sp, *tsp;
2523 	daddr_t dvbase;
2524 
2525 	/*
2526 	 * nblks is in DEV_BSIZE'd chunks, convert to PAGE_SIZE'd chunks.
2527 	 * First chop nblks off to page-align it, then convert.
2528 	 *
2529 	 * sw->sw_nblks is in page-sized chunks now too.
2530 	 */
2531 	nblks &= ~(ctodb(1) - 1);
2532 	nblks = dbtoc(nblks);
2533 
2534 	sp = malloc(sizeof *sp, M_VMPGDATA, M_WAITOK | M_ZERO);
2535 	sp->sw_blist = blist_create(nblks, M_WAITOK);
2536 	sp->sw_vp = vp;
2537 	sp->sw_id = id;
2538 	sp->sw_dev = dev;
2539 	sp->sw_nblks = nblks;
2540 	sp->sw_used = 0;
2541 	sp->sw_strategy = strategy;
2542 	sp->sw_close = close;
2543 	sp->sw_flags = flags;
2544 
2545 	/*
2546 	 * Do not free the first blocks in order to avoid overwriting
2547 	 * any bsd label at the front of the partition
2548 	 */
2549 	blist_free(sp->sw_blist, howmany(BBSIZE, PAGE_SIZE),
2550 	    nblks - howmany(BBSIZE, PAGE_SIZE));
2551 
2552 	dvbase = 0;
2553 	mtx_lock(&sw_dev_mtx);
2554 	TAILQ_FOREACH(tsp, &swtailq, sw_list) {
2555 		if (tsp->sw_end >= dvbase) {
2556 			/*
2557 			 * We put one uncovered page between the devices
2558 			 * in order to definitively prevent any cross-device
2559 			 * I/O requests
2560 			 */
2561 			dvbase = tsp->sw_end + 1;
2562 		}
2563 	}
2564 	sp->sw_first = dvbase;
2565 	sp->sw_end = dvbase + nblks;
2566 	TAILQ_INSERT_TAIL(&swtailq, sp, sw_list);
2567 	nswapdev++;
2568 	swap_pager_avail += nblks - howmany(BBSIZE, PAGE_SIZE);
2569 	swap_total += nblks;
2570 	swapon_check_swzone();
2571 	swp_sizecheck();
2572 	mtx_unlock(&sw_dev_mtx);
2573 	EVENTHANDLER_INVOKE(swapon, sp);
2574 }
2575 
2576 /*
2577  * SYSCALL: swapoff(devname)
2578  *
2579  * Disable swapping on the given device.
2580  *
2581  * XXX: Badly designed system call: it should use a device index
2582  * rather than filename as specification.  We keep sw_vp around
2583  * only to make this work.
2584  */
2585 static int
2586 kern_swapoff(struct thread *td, const char *name, enum uio_seg name_seg,
2587     u_int flags)
2588 {
2589 	struct vnode *vp;
2590 	struct nameidata nd;
2591 	struct swdevt *sp;
2592 	int error;
2593 
2594 	error = priv_check(td, PRIV_SWAPOFF);
2595 	if (error != 0)
2596 		return (error);
2597 	if ((flags & ~(SWAPOFF_FORCE)) != 0)
2598 		return (EINVAL);
2599 
2600 	sx_xlock(&swdev_syscall_lock);
2601 
2602 	NDINIT(&nd, LOOKUP, FOLLOW | AUDITVNODE1, name_seg, name);
2603 	error = namei(&nd);
2604 	if (error)
2605 		goto done;
2606 	NDFREE_PNBUF(&nd);
2607 	vp = nd.ni_vp;
2608 
2609 	mtx_lock(&sw_dev_mtx);
2610 	TAILQ_FOREACH(sp, &swtailq, sw_list) {
2611 		if (sp->sw_vp == vp)
2612 			break;
2613 	}
2614 	mtx_unlock(&sw_dev_mtx);
2615 	if (sp == NULL) {
2616 		error = EINVAL;
2617 		goto done;
2618 	}
2619 	error = swapoff_one(sp, td->td_ucred, flags);
2620 done:
2621 	sx_xunlock(&swdev_syscall_lock);
2622 	return (error);
2623 }
2624 
2625 
2626 #ifdef COMPAT_FREEBSD13
2627 int
2628 freebsd13_swapoff(struct thread *td, struct freebsd13_swapoff_args *uap)
2629 {
2630 	return (kern_swapoff(td, uap->name, UIO_USERSPACE, 0));
2631 }
2632 #endif
2633 
2634 int
2635 sys_swapoff(struct thread *td, struct swapoff_args *uap)
2636 {
2637 	return (kern_swapoff(td, uap->name, UIO_USERSPACE, uap->flags));
2638 }
2639 
2640 static int
2641 swapoff_one(struct swdevt *sp, struct ucred *cred, u_int flags)
2642 {
2643 	u_long nblks;
2644 #ifdef MAC
2645 	int error;
2646 #endif
2647 
2648 	sx_assert(&swdev_syscall_lock, SA_XLOCKED);
2649 #ifdef MAC
2650 	(void) vn_lock(sp->sw_vp, LK_EXCLUSIVE | LK_RETRY);
2651 	error = mac_system_check_swapoff(cred, sp->sw_vp);
2652 	(void) VOP_UNLOCK(sp->sw_vp);
2653 	if (error != 0)
2654 		return (error);
2655 #endif
2656 	nblks = sp->sw_nblks;
2657 
2658 	/*
2659 	 * We can turn off this swap device safely only if the
2660 	 * available virtual memory in the system will fit the amount
2661 	 * of data we will have to page back in, plus an epsilon so
2662 	 * the system doesn't become critically low on swap space.
2663 	 * The vm_free_count() part does not account e.g. for clean
2664 	 * pages that can be immediately reclaimed without paging, so
2665 	 * this is a very rough estimation.
2666 	 *
2667 	 * On the other hand, not turning swap off on swapoff_all()
2668 	 * means that we can lose swap data when filesystems go away,
2669 	 * which is arguably worse.
2670 	 */
2671 	if ((flags & SWAPOFF_FORCE) == 0 &&
2672 	    vm_free_count() + swap_pager_avail < nblks + nswap_lowat)
2673 		return (ENOMEM);
2674 
2675 	/*
2676 	 * Prevent further allocations on this device.
2677 	 */
2678 	mtx_lock(&sw_dev_mtx);
2679 	sp->sw_flags |= SW_CLOSING;
2680 	swap_pager_avail -= blist_fill(sp->sw_blist, 0, nblks);
2681 	swap_total -= nblks;
2682 	mtx_unlock(&sw_dev_mtx);
2683 
2684 	/*
2685 	 * Page in the contents of the device and close it.
2686 	 */
2687 	swap_pager_swapoff(sp);
2688 
2689 	sp->sw_close(curthread, sp);
2690 	mtx_lock(&sw_dev_mtx);
2691 	sp->sw_id = NULL;
2692 	TAILQ_REMOVE(&swtailq, sp, sw_list);
2693 	nswapdev--;
2694 	if (nswapdev == 0) {
2695 		swap_pager_full = 2;
2696 		swap_pager_almost_full = 1;
2697 	}
2698 	if (swdevhd == sp)
2699 		swdevhd = NULL;
2700 	mtx_unlock(&sw_dev_mtx);
2701 	blist_destroy(sp->sw_blist);
2702 	free(sp, M_VMPGDATA);
2703 	return (0);
2704 }
2705 
2706 void
2707 swapoff_all(void)
2708 {
2709 	struct swdevt *sp, *spt;
2710 	const char *devname;
2711 	int error;
2712 
2713 	sx_xlock(&swdev_syscall_lock);
2714 
2715 	mtx_lock(&sw_dev_mtx);
2716 	TAILQ_FOREACH_SAFE(sp, &swtailq, sw_list, spt) {
2717 		mtx_unlock(&sw_dev_mtx);
2718 		if (vn_isdisk(sp->sw_vp))
2719 			devname = devtoname(sp->sw_vp->v_rdev);
2720 		else
2721 			devname = "[file]";
2722 		error = swapoff_one(sp, thread0.td_ucred, SWAPOFF_FORCE);
2723 		if (error != 0) {
2724 			printf("Cannot remove swap device %s (error=%d), "
2725 			    "skipping.\n", devname, error);
2726 		} else if (bootverbose) {
2727 			printf("Swap device %s removed.\n", devname);
2728 		}
2729 		mtx_lock(&sw_dev_mtx);
2730 	}
2731 	mtx_unlock(&sw_dev_mtx);
2732 
2733 	sx_xunlock(&swdev_syscall_lock);
2734 }
2735 
2736 void
2737 swap_pager_status(int *total, int *used)
2738 {
2739 
2740 	*total = swap_total;
2741 	*used = swap_total - swap_pager_avail -
2742 	    nswapdev * howmany(BBSIZE, PAGE_SIZE);
2743 }
2744 
2745 int
2746 swap_dev_info(int name, struct xswdev *xs, char *devname, size_t len)
2747 {
2748 	struct swdevt *sp;
2749 	const char *tmp_devname;
2750 	int error, n;
2751 
2752 	n = 0;
2753 	error = ENOENT;
2754 	mtx_lock(&sw_dev_mtx);
2755 	TAILQ_FOREACH(sp, &swtailq, sw_list) {
2756 		if (n != name) {
2757 			n++;
2758 			continue;
2759 		}
2760 		xs->xsw_version = XSWDEV_VERSION;
2761 		xs->xsw_dev = sp->sw_dev;
2762 		xs->xsw_flags = sp->sw_flags;
2763 		xs->xsw_nblks = sp->sw_nblks;
2764 		xs->xsw_used = sp->sw_used;
2765 		if (devname != NULL) {
2766 			if (vn_isdisk(sp->sw_vp))
2767 				tmp_devname = devtoname(sp->sw_vp->v_rdev);
2768 			else
2769 				tmp_devname = "[file]";
2770 			strncpy(devname, tmp_devname, len);
2771 		}
2772 		error = 0;
2773 		break;
2774 	}
2775 	mtx_unlock(&sw_dev_mtx);
2776 	return (error);
2777 }
2778 
2779 #if defined(COMPAT_FREEBSD11)
2780 #define XSWDEV_VERSION_11	1
2781 struct xswdev11 {
2782 	u_int	xsw_version;
2783 	uint32_t xsw_dev;
2784 	int	xsw_flags;
2785 	int	xsw_nblks;
2786 	int     xsw_used;
2787 };
2788 #endif
2789 
2790 #if defined(__amd64__) && defined(COMPAT_FREEBSD32)
2791 struct xswdev32 {
2792 	u_int	xsw_version;
2793 	u_int	xsw_dev1, xsw_dev2;
2794 	int	xsw_flags;
2795 	int	xsw_nblks;
2796 	int     xsw_used;
2797 };
2798 #endif
2799 
2800 static int
2801 sysctl_vm_swap_info(SYSCTL_HANDLER_ARGS)
2802 {
2803 	struct xswdev xs;
2804 #if defined(__amd64__) && defined(COMPAT_FREEBSD32)
2805 	struct xswdev32 xs32;
2806 #endif
2807 #if defined(COMPAT_FREEBSD11)
2808 	struct xswdev11 xs11;
2809 #endif
2810 	int error;
2811 
2812 	if (arg2 != 1)			/* name length */
2813 		return (EINVAL);
2814 
2815 	memset(&xs, 0, sizeof(xs));
2816 	error = swap_dev_info(*(int *)arg1, &xs, NULL, 0);
2817 	if (error != 0)
2818 		return (error);
2819 #if defined(__amd64__) && defined(COMPAT_FREEBSD32)
2820 	if (req->oldlen == sizeof(xs32)) {
2821 		memset(&xs32, 0, sizeof(xs32));
2822 		xs32.xsw_version = XSWDEV_VERSION;
2823 		xs32.xsw_dev1 = xs.xsw_dev;
2824 		xs32.xsw_dev2 = xs.xsw_dev >> 32;
2825 		xs32.xsw_flags = xs.xsw_flags;
2826 		xs32.xsw_nblks = xs.xsw_nblks;
2827 		xs32.xsw_used = xs.xsw_used;
2828 		error = SYSCTL_OUT(req, &xs32, sizeof(xs32));
2829 		return (error);
2830 	}
2831 #endif
2832 #if defined(COMPAT_FREEBSD11)
2833 	if (req->oldlen == sizeof(xs11)) {
2834 		memset(&xs11, 0, sizeof(xs11));
2835 		xs11.xsw_version = XSWDEV_VERSION_11;
2836 		xs11.xsw_dev = xs.xsw_dev; /* truncation */
2837 		xs11.xsw_flags = xs.xsw_flags;
2838 		xs11.xsw_nblks = xs.xsw_nblks;
2839 		xs11.xsw_used = xs.xsw_used;
2840 		error = SYSCTL_OUT(req, &xs11, sizeof(xs11));
2841 		return (error);
2842 	}
2843 #endif
2844 	error = SYSCTL_OUT(req, &xs, sizeof(xs));
2845 	return (error);
2846 }
2847 
2848 SYSCTL_INT(_vm, OID_AUTO, nswapdev, CTLFLAG_RD, &nswapdev, 0,
2849     "Number of swap devices");
2850 SYSCTL_NODE(_vm, OID_AUTO, swap_info, CTLFLAG_RD | CTLFLAG_MPSAFE,
2851     sysctl_vm_swap_info,
2852     "Swap statistics by device");
2853 
2854 /*
2855  * Count the approximate swap usage in pages for a vmspace.  The
2856  * shadowed or not yet copied on write swap blocks are not accounted.
2857  * The map must be locked.
2858  */
2859 long
2860 vmspace_swap_count(struct vmspace *vmspace)
2861 {
2862 	struct pctrie_iter blks;
2863 	vm_map_t map;
2864 	vm_map_entry_t cur;
2865 	vm_object_t object;
2866 	struct swblk *sb;
2867 	vm_pindex_t e, pi;
2868 	long count;
2869 	int i, limit, start;
2870 
2871 	map = &vmspace->vm_map;
2872 	count = 0;
2873 
2874 	VM_MAP_ENTRY_FOREACH(cur, map) {
2875 		if ((cur->eflags & MAP_ENTRY_IS_SUB_MAP) != 0)
2876 			continue;
2877 		object = cur->object.vm_object;
2878 		if (object == NULL || (object->flags & OBJ_SWAP) == 0)
2879 			continue;
2880 		VM_OBJECT_RLOCK(object);
2881 		if ((object->flags & OBJ_SWAP) == 0)
2882 			goto unlock;
2883 		pi = OFF_TO_IDX(cur->offset);
2884 		e = pi + OFF_TO_IDX(cur->end - cur->start);
2885 		for (sb = swblk_iter_limit(&blks, object, pi, e),
2886 		    start = swblk_start(sb, pi);
2887 		    sb != NULL; sb = swblk_iter_next(&blks), start = 0) {
2888 			limit = MIN(e - blks.index, SWAP_META_PAGES);
2889 			for (i = start; i < limit; i++) {
2890 				if (sb->d[i] != SWAPBLK_NONE)
2891 					count++;
2892 			}
2893 		}
2894 unlock:
2895 		VM_OBJECT_RUNLOCK(object);
2896 	}
2897 	return (count);
2898 }
2899 
2900 /*
2901  * GEOM backend
2902  *
2903  * Swapping onto disk devices.
2904  *
2905  */
2906 
2907 static g_orphan_t swapgeom_orphan;
2908 
2909 static struct g_class g_swap_class = {
2910 	.name = "SWAP",
2911 	.version = G_VERSION,
2912 	.orphan = swapgeom_orphan,
2913 };
2914 
2915 DECLARE_GEOM_CLASS(g_swap_class, g_class);
2916 
2917 static void
2918 swapgeom_close_ev(void *arg, int flags)
2919 {
2920 	struct g_consumer *cp;
2921 
2922 	cp = arg;
2923 	g_access(cp, -1, -1, 0);
2924 	g_detach(cp);
2925 	g_destroy_consumer(cp);
2926 }
2927 
2928 /*
2929  * Add a reference to the g_consumer for an inflight transaction.
2930  */
2931 static void
2932 swapgeom_acquire(struct g_consumer *cp)
2933 {
2934 
2935 	mtx_assert(&sw_dev_mtx, MA_OWNED);
2936 	cp->index++;
2937 }
2938 
2939 /*
2940  * Remove a reference from the g_consumer.  Post a close event if all
2941  * references go away, since the function might be called from the
2942  * biodone context.
2943  */
2944 static void
2945 swapgeom_release(struct g_consumer *cp, struct swdevt *sp)
2946 {
2947 
2948 	mtx_assert(&sw_dev_mtx, MA_OWNED);
2949 	cp->index--;
2950 	if (cp->index == 0) {
2951 		if (g_post_event(swapgeom_close_ev, cp, M_NOWAIT, NULL) == 0)
2952 			sp->sw_id = NULL;
2953 	}
2954 }
2955 
2956 static void
2957 swapgeom_done(struct bio *bp2)
2958 {
2959 	struct swdevt *sp;
2960 	struct buf *bp;
2961 	struct g_consumer *cp;
2962 
2963 	bp = bp2->bio_caller2;
2964 	cp = bp2->bio_from;
2965 	bp->b_ioflags = bp2->bio_flags;
2966 	if (bp2->bio_error)
2967 		bp->b_ioflags |= BIO_ERROR;
2968 	bp->b_resid = bp->b_bcount - bp2->bio_completed;
2969 	bp->b_error = bp2->bio_error;
2970 	bp->b_caller1 = NULL;
2971 	bufdone(bp);
2972 	sp = bp2->bio_caller1;
2973 	mtx_lock(&sw_dev_mtx);
2974 	swapgeom_release(cp, sp);
2975 	mtx_unlock(&sw_dev_mtx);
2976 	g_destroy_bio(bp2);
2977 }
2978 
2979 static void
2980 swapgeom_strategy(struct buf *bp, struct swdevt *sp)
2981 {
2982 	struct bio *bio;
2983 	struct g_consumer *cp;
2984 
2985 	mtx_lock(&sw_dev_mtx);
2986 	cp = sp->sw_id;
2987 	if (cp == NULL) {
2988 		mtx_unlock(&sw_dev_mtx);
2989 		bp->b_error = ENXIO;
2990 		bp->b_ioflags |= BIO_ERROR;
2991 		bufdone(bp);
2992 		return;
2993 	}
2994 	swapgeom_acquire(cp);
2995 	mtx_unlock(&sw_dev_mtx);
2996 	if (bp->b_iocmd == BIO_WRITE)
2997 		bio = g_new_bio();
2998 	else
2999 		bio = g_alloc_bio();
3000 	if (bio == NULL) {
3001 		mtx_lock(&sw_dev_mtx);
3002 		swapgeom_release(cp, sp);
3003 		mtx_unlock(&sw_dev_mtx);
3004 		bp->b_error = ENOMEM;
3005 		bp->b_ioflags |= BIO_ERROR;
3006 		printf("swap_pager: cannot allocate bio\n");
3007 		bufdone(bp);
3008 		return;
3009 	}
3010 
3011 	bp->b_caller1 = bio;
3012 	bio->bio_caller1 = sp;
3013 	bio->bio_caller2 = bp;
3014 	bio->bio_cmd = bp->b_iocmd;
3015 	bio->bio_offset = (bp->b_blkno - sp->sw_first) * PAGE_SIZE;
3016 	bio->bio_length = bp->b_bcount;
3017 	bio->bio_done = swapgeom_done;
3018 	bio->bio_flags |= BIO_SWAP;
3019 	if (!buf_mapped(bp)) {
3020 		bio->bio_ma = bp->b_pages;
3021 		bio->bio_data = unmapped_buf;
3022 		bio->bio_ma_offset = (vm_offset_t)bp->b_offset & PAGE_MASK;
3023 		bio->bio_ma_n = bp->b_npages;
3024 		bio->bio_flags |= BIO_UNMAPPED;
3025 	} else {
3026 		bio->bio_data = bp->b_data;
3027 		bio->bio_ma = NULL;
3028 	}
3029 	g_io_request(bio, cp);
3030 	return;
3031 }
3032 
3033 static void
3034 swapgeom_orphan(struct g_consumer *cp)
3035 {
3036 	struct swdevt *sp;
3037 	int destroy;
3038 
3039 	mtx_lock(&sw_dev_mtx);
3040 	TAILQ_FOREACH(sp, &swtailq, sw_list) {
3041 		if (sp->sw_id == cp) {
3042 			sp->sw_flags |= SW_CLOSING;
3043 			break;
3044 		}
3045 	}
3046 	/*
3047 	 * Drop reference we were created with. Do directly since we're in a
3048 	 * special context where we don't have to queue the call to
3049 	 * swapgeom_close_ev().
3050 	 */
3051 	cp->index--;
3052 	destroy = ((sp != NULL) && (cp->index == 0));
3053 	if (destroy)
3054 		sp->sw_id = NULL;
3055 	mtx_unlock(&sw_dev_mtx);
3056 	if (destroy)
3057 		swapgeom_close_ev(cp, 0);
3058 }
3059 
3060 static void
3061 swapgeom_close(struct thread *td, struct swdevt *sw)
3062 {
3063 	struct g_consumer *cp;
3064 
3065 	mtx_lock(&sw_dev_mtx);
3066 	cp = sw->sw_id;
3067 	sw->sw_id = NULL;
3068 	mtx_unlock(&sw_dev_mtx);
3069 
3070 	/*
3071 	 * swapgeom_close() may be called from the biodone context,
3072 	 * where we cannot perform topology changes.  Delegate the
3073 	 * work to the events thread.
3074 	 */
3075 	if (cp != NULL)
3076 		g_waitfor_event(swapgeom_close_ev, cp, M_WAITOK, NULL);
3077 }
3078 
3079 static int
3080 swapongeom_locked(struct cdev *dev, struct vnode *vp)
3081 {
3082 	struct g_provider *pp;
3083 	struct g_consumer *cp;
3084 	static struct g_geom *gp;
3085 	struct swdevt *sp;
3086 	u_long nblks;
3087 	int error;
3088 
3089 	pp = g_dev_getprovider(dev);
3090 	if (pp == NULL)
3091 		return (ENODEV);
3092 	mtx_lock(&sw_dev_mtx);
3093 	TAILQ_FOREACH(sp, &swtailq, sw_list) {
3094 		cp = sp->sw_id;
3095 		if (cp != NULL && cp->provider == pp) {
3096 			mtx_unlock(&sw_dev_mtx);
3097 			return (EBUSY);
3098 		}
3099 	}
3100 	mtx_unlock(&sw_dev_mtx);
3101 	if (gp == NULL)
3102 		gp = g_new_geomf(&g_swap_class, "swap");
3103 	cp = g_new_consumer(gp);
3104 	cp->index = 1;	/* Number of active I/Os, plus one for being active. */
3105 	cp->flags |=  G_CF_DIRECT_SEND | G_CF_DIRECT_RECEIVE;
3106 	g_attach(cp, pp);
3107 	/*
3108 	 * XXX: Every time you think you can improve the margin for
3109 	 * footshooting, somebody depends on the ability to do so:
3110 	 * savecore(8) wants to write to our swapdev so we cannot
3111 	 * set an exclusive count :-(
3112 	 */
3113 	error = g_access(cp, 1, 1, 0);
3114 	if (error != 0) {
3115 		g_detach(cp);
3116 		g_destroy_consumer(cp);
3117 		return (error);
3118 	}
3119 	nblks = pp->mediasize / DEV_BSIZE;
3120 	swaponsomething(vp, cp, nblks, swapgeom_strategy,
3121 	    swapgeom_close, dev2udev(dev),
3122 	    (pp->flags & G_PF_ACCEPT_UNMAPPED) != 0 ? SW_UNMAPPED : 0);
3123 	return (0);
3124 }
3125 
3126 static int
3127 swapongeom(struct vnode *vp)
3128 {
3129 	int error;
3130 
3131 	ASSERT_VOP_ELOCKED(vp, "swapongeom");
3132 	if (vp->v_type != VCHR || VN_IS_DOOMED(vp)) {
3133 		error = ENOENT;
3134 	} else {
3135 		g_topology_lock();
3136 		error = swapongeom_locked(vp->v_rdev, vp);
3137 		g_topology_unlock();
3138 	}
3139 	return (error);
3140 }
3141 
3142 /*
3143  * VNODE backend
3144  *
3145  * This is used mainly for network filesystem (read: probably only tested
3146  * with NFS) swapfiles.
3147  *
3148  */
3149 
3150 static void
3151 swapdev_strategy(struct buf *bp, struct swdevt *sp)
3152 {
3153 	struct vnode *vp2;
3154 
3155 	bp->b_blkno = ctodb(bp->b_blkno - sp->sw_first);
3156 
3157 	vp2 = sp->sw_id;
3158 	vhold(vp2);
3159 	if (bp->b_iocmd == BIO_WRITE) {
3160 		vn_lock(vp2, LK_EXCLUSIVE | LK_RETRY);
3161 		if (bp->b_bufobj)
3162 			bufobj_wdrop(bp->b_bufobj);
3163 		bufobj_wref(&vp2->v_bufobj);
3164 	} else {
3165 		vn_lock(vp2, LK_SHARED | LK_RETRY);
3166 	}
3167 	if (bp->b_bufobj != &vp2->v_bufobj)
3168 		bp->b_bufobj = &vp2->v_bufobj;
3169 	bp->b_vp = vp2;
3170 	bp->b_iooffset = dbtob(bp->b_blkno);
3171 	bstrategy(bp);
3172 	VOP_UNLOCK(vp2);
3173 }
3174 
3175 static void
3176 swapdev_close(struct thread *td, struct swdevt *sp)
3177 {
3178 	struct vnode *vp;
3179 
3180 	vp = sp->sw_vp;
3181 	vn_lock(vp, LK_EXCLUSIVE | LK_RETRY);
3182 	VOP_CLOSE(vp, FREAD | FWRITE, td->td_ucred, td);
3183 	vput(vp);
3184 }
3185 
3186 static int
3187 swaponvp(struct thread *td, struct vnode *vp, u_long nblks)
3188 {
3189 	struct swdevt *sp;
3190 	int error;
3191 
3192 	ASSERT_VOP_ELOCKED(vp, "swaponvp");
3193 	if (nblks == 0)
3194 		return (ENXIO);
3195 	mtx_lock(&sw_dev_mtx);
3196 	TAILQ_FOREACH(sp, &swtailq, sw_list) {
3197 		if (sp->sw_id == vp) {
3198 			mtx_unlock(&sw_dev_mtx);
3199 			return (EBUSY);
3200 		}
3201 	}
3202 	mtx_unlock(&sw_dev_mtx);
3203 
3204 #ifdef MAC
3205 	error = mac_system_check_swapon(td->td_ucred, vp);
3206 	if (error == 0)
3207 #endif
3208 		error = VOP_OPEN(vp, FREAD | FWRITE, td->td_ucred, td, NULL);
3209 	if (error != 0)
3210 		return (error);
3211 
3212 	swaponsomething(vp, vp, nblks, swapdev_strategy, swapdev_close,
3213 	    NODEV, 0);
3214 	return (0);
3215 }
3216 
3217 static int
3218 sysctl_swap_async_max(SYSCTL_HANDLER_ARGS)
3219 {
3220 	int error, new, n;
3221 
3222 	new = nsw_wcount_async_max;
3223 	error = sysctl_handle_int(oidp, &new, 0, req);
3224 	if (error != 0 || req->newptr == NULL)
3225 		return (error);
3226 
3227 	if (new > nswbuf / 2 || new < 1)
3228 		return (EINVAL);
3229 
3230 	mtx_lock(&swbuf_mtx);
3231 	while (nsw_wcount_async_max != new) {
3232 		/*
3233 		 * Adjust difference.  If the current async count is too low,
3234 		 * we will need to sqeeze our update slowly in.  Sleep with a
3235 		 * higher priority than getpbuf() to finish faster.
3236 		 */
3237 		n = new - nsw_wcount_async_max;
3238 		if (nsw_wcount_async + n >= 0) {
3239 			nsw_wcount_async += n;
3240 			nsw_wcount_async_max += n;
3241 			wakeup(&nsw_wcount_async);
3242 		} else {
3243 			nsw_wcount_async_max -= nsw_wcount_async;
3244 			nsw_wcount_async = 0;
3245 			msleep(&nsw_wcount_async, &swbuf_mtx, PSWP,
3246 			    "swpsysctl", 0);
3247 		}
3248 	}
3249 	mtx_unlock(&swbuf_mtx);
3250 
3251 	return (0);
3252 }
3253 
3254 static void
3255 swap_pager_update_writecount(vm_object_t object, vm_offset_t start,
3256     vm_offset_t end)
3257 {
3258 
3259 	VM_OBJECT_WLOCK(object);
3260 	KASSERT((object->flags & OBJ_ANON) == 0,
3261 	    ("Splittable object with writecount"));
3262 	object->un_pager.swp.writemappings += (vm_ooffset_t)end - start;
3263 	VM_OBJECT_WUNLOCK(object);
3264 }
3265 
3266 static void
3267 swap_pager_release_writecount(vm_object_t object, vm_offset_t start,
3268     vm_offset_t end)
3269 {
3270 
3271 	VM_OBJECT_WLOCK(object);
3272 	KASSERT((object->flags & OBJ_ANON) == 0,
3273 	    ("Splittable object with writecount"));
3274 	KASSERT(object->un_pager.swp.writemappings >= (vm_ooffset_t)end - start,
3275 	    ("swap obj %p writecount %jx dec %jx", object,
3276 	    (uintmax_t)object->un_pager.swp.writemappings,
3277 	    (uintmax_t)((vm_ooffset_t)end - start)));
3278 	object->un_pager.swp.writemappings -= (vm_ooffset_t)end - start;
3279 	VM_OBJECT_WUNLOCK(object);
3280 }
3281