xref: /freebsd/sys/vm/swap_pager.c (revision ddc0daea20280c3a06a910b72b14ffe3f624df71)
1 /*-
2  * SPDX-License-Identifier: BSD-4-Clause
3  *
4  * Copyright (c) 1998 Matthew Dillon,
5  * Copyright (c) 1994 John S. Dyson
6  * Copyright (c) 1990 University of Utah.
7  * Copyright (c) 1982, 1986, 1989, 1993
8  *	The Regents of the University of California.  All rights reserved.
9  *
10  * This code is derived from software contributed to Berkeley by
11  * the Systems Programming Group of the University of Utah Computer
12  * Science Department.
13  *
14  * Redistribution and use in source and binary forms, with or without
15  * modification, are permitted provided that the following conditions
16  * are met:
17  * 1. Redistributions of source code must retain the above copyright
18  *    notice, this list of conditions and the following disclaimer.
19  * 2. Redistributions in binary form must reproduce the above copyright
20  *    notice, this list of conditions and the following disclaimer in the
21  *    documentation and/or other materials provided with the distribution.
22  * 3. All advertising materials mentioning features or use of this software
23  *    must display the following acknowledgement:
24  *	This product includes software developed by the University of
25  *	California, Berkeley and its contributors.
26  * 4. Neither the name of the University nor the names of its contributors
27  *    may be used to endorse or promote products derived from this software
28  *    without specific prior written permission.
29  *
30  * THIS SOFTWARE IS PROVIDED BY THE REGENTS AND CONTRIBUTORS ``AS IS'' AND
31  * ANY EXPRESS OR IMPLIED WARRANTIES, INCLUDING, BUT NOT LIMITED TO, THE
32  * IMPLIED WARRANTIES OF MERCHANTABILITY AND FITNESS FOR A PARTICULAR PURPOSE
33  * ARE DISCLAIMED.  IN NO EVENT SHALL THE REGENTS OR CONTRIBUTORS BE LIABLE
34  * FOR ANY DIRECT, INDIRECT, INCIDENTAL, SPECIAL, EXEMPLARY, OR CONSEQUENTIAL
35  * DAMAGES (INCLUDING, BUT NOT LIMITED TO, PROCUREMENT OF SUBSTITUTE GOODS
36  * OR SERVICES; LOSS OF USE, DATA, OR PROFITS; OR BUSINESS INTERRUPTION)
37  * HOWEVER CAUSED AND ON ANY THEORY OF LIABILITY, WHETHER IN CONTRACT, STRICT
38  * LIABILITY, OR TORT (INCLUDING NEGLIGENCE OR OTHERWISE) ARISING IN ANY WAY
39  * OUT OF THE USE OF THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF
40  * SUCH DAMAGE.
41  *
42  *				New Swap System
43  *				Matthew Dillon
44  *
45  * Radix Bitmap 'blists'.
46  *
47  *	- The new swapper uses the new radix bitmap code.  This should scale
48  *	  to arbitrarily small or arbitrarily large swap spaces and an almost
49  *	  arbitrary degree of fragmentation.
50  *
51  * Features:
52  *
53  *	- on the fly reallocation of swap during putpages.  The new system
54  *	  does not try to keep previously allocated swap blocks for dirty
55  *	  pages.
56  *
57  *	- on the fly deallocation of swap
58  *
59  *	- No more garbage collection required.  Unnecessarily allocated swap
60  *	  blocks only exist for dirty vm_page_t's now and these are already
61  *	  cycled (in a high-load system) by the pager.  We also do on-the-fly
62  *	  removal of invalidated swap blocks when a page is destroyed
63  *	  or renamed.
64  *
65  * from: Utah $Hdr: swap_pager.c 1.4 91/04/30$
66  *
67  *	@(#)swap_pager.c	8.9 (Berkeley) 3/21/94
68  *	@(#)vm_swap.c	8.5 (Berkeley) 2/17/94
69  */
70 
71 #include <sys/cdefs.h>
72 __FBSDID("$FreeBSD$");
73 
74 #include "opt_vm.h"
75 
76 #include <sys/param.h>
77 #include <sys/bio.h>
78 #include <sys/blist.h>
79 #include <sys/buf.h>
80 #include <sys/conf.h>
81 #include <sys/disk.h>
82 #include <sys/disklabel.h>
83 #include <sys/eventhandler.h>
84 #include <sys/fcntl.h>
85 #include <sys/lock.h>
86 #include <sys/kernel.h>
87 #include <sys/mount.h>
88 #include <sys/namei.h>
89 #include <sys/malloc.h>
90 #include <sys/pctrie.h>
91 #include <sys/priv.h>
92 #include <sys/proc.h>
93 #include <sys/racct.h>
94 #include <sys/resource.h>
95 #include <sys/resourcevar.h>
96 #include <sys/rwlock.h>
97 #include <sys/sbuf.h>
98 #include <sys/sysctl.h>
99 #include <sys/sysproto.h>
100 #include <sys/systm.h>
101 #include <sys/sx.h>
102 #include <sys/vmmeter.h>
103 #include <sys/vnode.h>
104 
105 #include <security/mac/mac_framework.h>
106 
107 #include <vm/vm.h>
108 #include <vm/pmap.h>
109 #include <vm/vm_map.h>
110 #include <vm/vm_kern.h>
111 #include <vm/vm_object.h>
112 #include <vm/vm_page.h>
113 #include <vm/vm_pager.h>
114 #include <vm/vm_pageout.h>
115 #include <vm/vm_param.h>
116 #include <vm/swap_pager.h>
117 #include <vm/vm_extern.h>
118 #include <vm/uma.h>
119 
120 #include <geom/geom.h>
121 
122 /*
123  * MAX_PAGEOUT_CLUSTER must be a power of 2 between 1 and 64.
124  * The 64-page limit is due to the radix code (kern/subr_blist.c).
125  */
126 #ifndef MAX_PAGEOUT_CLUSTER
127 #define	MAX_PAGEOUT_CLUSTER	32
128 #endif
129 
130 #if !defined(SWB_NPAGES)
131 #define SWB_NPAGES	MAX_PAGEOUT_CLUSTER
132 #endif
133 
134 #define	SWAP_META_PAGES		PCTRIE_COUNT
135 
136 /*
137  * A swblk structure maps each page index within a
138  * SWAP_META_PAGES-aligned and sized range to the address of an
139  * on-disk swap block (or SWAPBLK_NONE). The collection of these
140  * mappings for an entire vm object is implemented as a pc-trie.
141  */
142 struct swblk {
143 	vm_pindex_t	p;
144 	daddr_t		d[SWAP_META_PAGES];
145 };
146 
147 static MALLOC_DEFINE(M_VMPGDATA, "vm_pgdata", "swap pager private data");
148 static struct mtx sw_dev_mtx;
149 static TAILQ_HEAD(, swdevt) swtailq = TAILQ_HEAD_INITIALIZER(swtailq);
150 static struct swdevt *swdevhd;	/* Allocate from here next */
151 static int nswapdev;		/* Number of swap devices */
152 int swap_pager_avail;
153 static struct sx swdev_syscall_lock;	/* serialize swap(on|off) */
154 
155 static u_long swap_reserved;
156 static u_long swap_total;
157 static int sysctl_page_shift(SYSCTL_HANDLER_ARGS);
158 SYSCTL_PROC(_vm, OID_AUTO, swap_reserved, CTLTYPE_U64 | CTLFLAG_RD | CTLFLAG_MPSAFE,
159     &swap_reserved, 0, sysctl_page_shift, "A",
160     "Amount of swap storage needed to back all allocated anonymous memory.");
161 SYSCTL_PROC(_vm, OID_AUTO, swap_total, CTLTYPE_U64 | CTLFLAG_RD | CTLFLAG_MPSAFE,
162     &swap_total, 0, sysctl_page_shift, "A",
163     "Total amount of available swap storage.");
164 
165 static int overcommit = 0;
166 SYSCTL_INT(_vm, VM_OVERCOMMIT, overcommit, CTLFLAG_RW, &overcommit, 0,
167     "Configure virtual memory overcommit behavior. See tuning(7) "
168     "for details.");
169 static unsigned long swzone;
170 SYSCTL_ULONG(_vm, OID_AUTO, swzone, CTLFLAG_RD, &swzone, 0,
171     "Actual size of swap metadata zone");
172 static unsigned long swap_maxpages;
173 SYSCTL_ULONG(_vm, OID_AUTO, swap_maxpages, CTLFLAG_RD, &swap_maxpages, 0,
174     "Maximum amount of swap supported");
175 
176 /* bits from overcommit */
177 #define	SWAP_RESERVE_FORCE_ON		(1 << 0)
178 #define	SWAP_RESERVE_RLIMIT_ON		(1 << 1)
179 #define	SWAP_RESERVE_ALLOW_NONWIRED	(1 << 2)
180 
181 static int
182 sysctl_page_shift(SYSCTL_HANDLER_ARGS)
183 {
184 	uint64_t newval;
185 	u_long value = *(u_long *)arg1;
186 
187 	newval = ((uint64_t)value) << PAGE_SHIFT;
188 	return (sysctl_handle_64(oidp, &newval, 0, req));
189 }
190 
191 int
192 swap_reserve(vm_ooffset_t incr)
193 {
194 
195 	return (swap_reserve_by_cred(incr, curthread->td_ucred));
196 }
197 
198 int
199 swap_reserve_by_cred(vm_ooffset_t incr, struct ucred *cred)
200 {
201 	u_long r, s, prev, pincr;
202 	int res, error;
203 	static int curfail;
204 	static struct timeval lastfail;
205 	struct uidinfo *uip;
206 
207 	uip = cred->cr_ruidinfo;
208 
209 	KASSERT((incr & PAGE_MASK) == 0, ("%s: incr: %ju & PAGE_MASK", __func__,
210 	    (uintmax_t)incr));
211 
212 #ifdef RACCT
213 	if (racct_enable) {
214 		PROC_LOCK(curproc);
215 		error = racct_add(curproc, RACCT_SWAP, incr);
216 		PROC_UNLOCK(curproc);
217 		if (error != 0)
218 			return (0);
219 	}
220 #endif
221 
222 	pincr = atop(incr);
223 	res = 0;
224 	prev = atomic_fetchadd_long(&swap_reserved, pincr);
225 	r = prev + pincr;
226 	if (overcommit & SWAP_RESERVE_ALLOW_NONWIRED) {
227 		s = vm_cnt.v_page_count - vm_cnt.v_free_reserved -
228 		    vm_wire_count();
229 	} else
230 		s = 0;
231 	s += swap_total;
232 	if ((overcommit & SWAP_RESERVE_FORCE_ON) == 0 || r <= s ||
233 	    (error = priv_check(curthread, PRIV_VM_SWAP_NOQUOTA)) == 0) {
234 		res = 1;
235 	} else {
236 		prev = atomic_fetchadd_long(&swap_reserved, -pincr);
237 		if (prev < pincr)
238 			panic("swap_reserved < incr on overcommit fail");
239 	}
240 	if (res) {
241 		prev = atomic_fetchadd_long(&uip->ui_vmsize, pincr);
242 		if ((overcommit & SWAP_RESERVE_RLIMIT_ON) != 0 &&
243 		    prev + pincr > lim_cur(curthread, RLIMIT_SWAP) &&
244 		    priv_check(curthread, PRIV_VM_SWAP_NORLIMIT)) {
245 			res = 0;
246 			prev = atomic_fetchadd_long(&uip->ui_vmsize, -pincr);
247 			if (prev < pincr)
248 				panic("uip->ui_vmsize < incr on overcommit fail");
249 		}
250 	}
251 	if (!res && ppsratecheck(&lastfail, &curfail, 1)) {
252 		printf("uid %d, pid %d: swap reservation for %jd bytes failed\n",
253 		    uip->ui_uid, curproc->p_pid, incr);
254 	}
255 
256 #ifdef RACCT
257 	if (racct_enable && !res) {
258 		PROC_LOCK(curproc);
259 		racct_sub(curproc, RACCT_SWAP, incr);
260 		PROC_UNLOCK(curproc);
261 	}
262 #endif
263 
264 	return (res);
265 }
266 
267 void
268 swap_reserve_force(vm_ooffset_t incr)
269 {
270 	struct uidinfo *uip;
271 	u_long pincr;
272 
273 	KASSERT((incr & PAGE_MASK) == 0, ("%s: incr: %ju & PAGE_MASK", __func__,
274 	    (uintmax_t)incr));
275 
276 	PROC_LOCK(curproc);
277 #ifdef RACCT
278 	if (racct_enable)
279 		racct_add_force(curproc, RACCT_SWAP, incr);
280 #endif
281 	pincr = atop(incr);
282 	atomic_add_long(&swap_reserved, pincr);
283 	uip = curproc->p_ucred->cr_ruidinfo;
284 	atomic_add_long(&uip->ui_vmsize, pincr);
285 	PROC_UNLOCK(curproc);
286 }
287 
288 void
289 swap_release(vm_ooffset_t decr)
290 {
291 	struct ucred *cred;
292 
293 	PROC_LOCK(curproc);
294 	cred = curproc->p_ucred;
295 	swap_release_by_cred(decr, cred);
296 	PROC_UNLOCK(curproc);
297 }
298 
299 void
300 swap_release_by_cred(vm_ooffset_t decr, struct ucred *cred)
301 {
302 	u_long prev, pdecr;
303  	struct uidinfo *uip;
304 
305 	uip = cred->cr_ruidinfo;
306 
307 	KASSERT((decr & PAGE_MASK) == 0, ("%s: decr: %ju & PAGE_MASK", __func__,
308 	    (uintmax_t)decr));
309 
310 	pdecr = atop(decr);
311 	prev = atomic_fetchadd_long(&swap_reserved, -pdecr);
312 	if (prev < pdecr)
313 		panic("swap_reserved < decr");
314 
315 	prev = atomic_fetchadd_long(&uip->ui_vmsize, -pdecr);
316 	if (prev < pdecr)
317 		printf("negative vmsize for uid = %d\n", uip->ui_uid);
318 #ifdef RACCT
319 	if (racct_enable)
320 		racct_sub_cred(cred, RACCT_SWAP, decr);
321 #endif
322 }
323 
324 #define SWM_POP		0x01	/* pop out			*/
325 
326 static int swap_pager_full = 2;	/* swap space exhaustion (task killing) */
327 static int swap_pager_almost_full = 1; /* swap space exhaustion (w/hysteresis)*/
328 static struct mtx swbuf_mtx;	/* to sync nsw_wcount_async */
329 static int nsw_wcount_async;	/* limit async write buffers */
330 static int nsw_wcount_async_max;/* assigned maximum			*/
331 static int nsw_cluster_max;	/* maximum VOP I/O allowed		*/
332 
333 static int sysctl_swap_async_max(SYSCTL_HANDLER_ARGS);
334 SYSCTL_PROC(_vm, OID_AUTO, swap_async_max, CTLTYPE_INT | CTLFLAG_RW |
335     CTLFLAG_MPSAFE, NULL, 0, sysctl_swap_async_max, "I",
336     "Maximum running async swap ops");
337 static int sysctl_swap_fragmentation(SYSCTL_HANDLER_ARGS);
338 SYSCTL_PROC(_vm, OID_AUTO, swap_fragmentation, CTLTYPE_STRING | CTLFLAG_RD |
339     CTLFLAG_MPSAFE, NULL, 0, sysctl_swap_fragmentation, "A",
340     "Swap Fragmentation Info");
341 
342 static struct sx sw_alloc_sx;
343 
344 /*
345  * "named" and "unnamed" anon region objects.  Try to reduce the overhead
346  * of searching a named list by hashing it just a little.
347  */
348 
349 #define NOBJLISTS		8
350 
351 #define NOBJLIST(handle)	\
352 	(&swap_pager_object_list[((int)(intptr_t)handle >> 4) & (NOBJLISTS-1)])
353 
354 static struct pagerlst	swap_pager_object_list[NOBJLISTS];
355 static uma_zone_t swwbuf_zone;
356 static uma_zone_t swrbuf_zone;
357 static uma_zone_t swblk_zone;
358 static uma_zone_t swpctrie_zone;
359 
360 /*
361  * pagerops for OBJT_SWAP - "swap pager".  Some ops are also global procedure
362  * calls hooked from other parts of the VM system and do not appear here.
363  * (see vm/swap_pager.h).
364  */
365 static vm_object_t
366 		swap_pager_alloc(void *handle, vm_ooffset_t size,
367 		    vm_prot_t prot, vm_ooffset_t offset, struct ucred *);
368 static void	swap_pager_dealloc(vm_object_t object);
369 static int	swap_pager_getpages(vm_object_t, vm_page_t *, int, int *,
370     int *);
371 static int	swap_pager_getpages_async(vm_object_t, vm_page_t *, int, int *,
372     int *, pgo_getpages_iodone_t, void *);
373 static void	swap_pager_putpages(vm_object_t, vm_page_t *, int, boolean_t, int *);
374 static boolean_t
375 		swap_pager_haspage(vm_object_t object, vm_pindex_t pindex, int *before, int *after);
376 static void	swap_pager_init(void);
377 static void	swap_pager_unswapped(vm_page_t);
378 static void	swap_pager_swapoff(struct swdevt *sp);
379 static void	swap_pager_update_writecount(vm_object_t object,
380     vm_offset_t start, vm_offset_t end);
381 static void	swap_pager_release_writecount(vm_object_t object,
382     vm_offset_t start, vm_offset_t end);
383 
384 struct pagerops swappagerops = {
385 	.pgo_init =	swap_pager_init,	/* early system initialization of pager	*/
386 	.pgo_alloc =	swap_pager_alloc,	/* allocate an OBJT_SWAP object		*/
387 	.pgo_dealloc =	swap_pager_dealloc,	/* deallocate an OBJT_SWAP object	*/
388 	.pgo_getpages =	swap_pager_getpages,	/* pagein				*/
389 	.pgo_getpages_async = swap_pager_getpages_async, /* pagein (async)		*/
390 	.pgo_putpages =	swap_pager_putpages,	/* pageout				*/
391 	.pgo_haspage =	swap_pager_haspage,	/* get backing store status for page	*/
392 	.pgo_pageunswapped = swap_pager_unswapped,	/* remove swap related to page		*/
393 	.pgo_update_writecount = swap_pager_update_writecount,
394 	.pgo_release_writecount = swap_pager_release_writecount,
395 };
396 
397 /*
398  * swap_*() routines are externally accessible.  swp_*() routines are
399  * internal.
400  */
401 static int nswap_lowat = 128;	/* in pages, swap_pager_almost_full warn */
402 static int nswap_hiwat = 512;	/* in pages, swap_pager_almost_full warn */
403 
404 SYSCTL_INT(_vm, OID_AUTO, dmmax, CTLFLAG_RD, &nsw_cluster_max, 0,
405     "Maximum size of a swap block in pages");
406 
407 static void	swp_sizecheck(void);
408 static void	swp_pager_async_iodone(struct buf *bp);
409 static bool	swp_pager_swblk_empty(struct swblk *sb, int start, int limit);
410 static int	swapongeom(struct vnode *);
411 static int	swaponvp(struct thread *, struct vnode *, u_long);
412 static int	swapoff_one(struct swdevt *sp, struct ucred *cred);
413 
414 /*
415  * Swap bitmap functions
416  */
417 static void	swp_pager_freeswapspace(daddr_t blk, daddr_t npages);
418 static daddr_t	swp_pager_getswapspace(int *npages, int limit);
419 
420 /*
421  * Metadata functions
422  */
423 static daddr_t swp_pager_meta_build(vm_object_t, vm_pindex_t, daddr_t);
424 static void swp_pager_meta_free(vm_object_t, vm_pindex_t, vm_pindex_t);
425 static void swp_pager_meta_free_all(vm_object_t);
426 static daddr_t swp_pager_meta_ctl(vm_object_t, vm_pindex_t, int);
427 
428 static void
429 swp_pager_init_freerange(daddr_t *start, daddr_t *num)
430 {
431 
432 	*start = SWAPBLK_NONE;
433 	*num = 0;
434 }
435 
436 static void
437 swp_pager_update_freerange(daddr_t *start, daddr_t *num, daddr_t addr)
438 {
439 
440 	if (*start + *num == addr) {
441 		(*num)++;
442 	} else {
443 		swp_pager_freeswapspace(*start, *num);
444 		*start = addr;
445 		*num = 1;
446 	}
447 }
448 
449 static void *
450 swblk_trie_alloc(struct pctrie *ptree)
451 {
452 
453 	return (uma_zalloc(swpctrie_zone, M_NOWAIT | (curproc == pageproc ?
454 	    M_USE_RESERVE : 0)));
455 }
456 
457 static void
458 swblk_trie_free(struct pctrie *ptree, void *node)
459 {
460 
461 	uma_zfree(swpctrie_zone, node);
462 }
463 
464 PCTRIE_DEFINE(SWAP, swblk, p, swblk_trie_alloc, swblk_trie_free);
465 
466 /*
467  * SWP_SIZECHECK() -	update swap_pager_full indication
468  *
469  *	update the swap_pager_almost_full indication and warn when we are
470  *	about to run out of swap space, using lowat/hiwat hysteresis.
471  *
472  *	Clear swap_pager_full ( task killing ) indication when lowat is met.
473  *
474  *	No restrictions on call
475  *	This routine may not block.
476  */
477 static void
478 swp_sizecheck(void)
479 {
480 
481 	if (swap_pager_avail < nswap_lowat) {
482 		if (swap_pager_almost_full == 0) {
483 			printf("swap_pager: out of swap space\n");
484 			swap_pager_almost_full = 1;
485 		}
486 	} else {
487 		swap_pager_full = 0;
488 		if (swap_pager_avail > nswap_hiwat)
489 			swap_pager_almost_full = 0;
490 	}
491 }
492 
493 /*
494  * SWAP_PAGER_INIT() -	initialize the swap pager!
495  *
496  *	Expected to be started from system init.  NOTE:  This code is run
497  *	before much else so be careful what you depend on.  Most of the VM
498  *	system has yet to be initialized at this point.
499  */
500 static void
501 swap_pager_init(void)
502 {
503 	/*
504 	 * Initialize object lists
505 	 */
506 	int i;
507 
508 	for (i = 0; i < NOBJLISTS; ++i)
509 		TAILQ_INIT(&swap_pager_object_list[i]);
510 	mtx_init(&sw_dev_mtx, "swapdev", NULL, MTX_DEF);
511 	sx_init(&sw_alloc_sx, "swspsx");
512 	sx_init(&swdev_syscall_lock, "swsysc");
513 }
514 
515 /*
516  * SWAP_PAGER_SWAP_INIT() - swap pager initialization from pageout process
517  *
518  *	Expected to be started from pageout process once, prior to entering
519  *	its main loop.
520  */
521 void
522 swap_pager_swap_init(void)
523 {
524 	unsigned long n, n2;
525 
526 	/*
527 	 * Number of in-transit swap bp operations.  Don't
528 	 * exhaust the pbufs completely.  Make sure we
529 	 * initialize workable values (0 will work for hysteresis
530 	 * but it isn't very efficient).
531 	 *
532 	 * The nsw_cluster_max is constrained by the bp->b_pages[]
533 	 * array, which has MAXPHYS / PAGE_SIZE entries, and our locally
534 	 * defined MAX_PAGEOUT_CLUSTER.   Also be aware that swap ops are
535 	 * constrained by the swap device interleave stripe size.
536 	 *
537 	 * Currently we hardwire nsw_wcount_async to 4.  This limit is
538 	 * designed to prevent other I/O from having high latencies due to
539 	 * our pageout I/O.  The value 4 works well for one or two active swap
540 	 * devices but is probably a little low if you have more.  Even so,
541 	 * a higher value would probably generate only a limited improvement
542 	 * with three or four active swap devices since the system does not
543 	 * typically have to pageout at extreme bandwidths.   We will want
544 	 * at least 2 per swap devices, and 4 is a pretty good value if you
545 	 * have one NFS swap device due to the command/ack latency over NFS.
546 	 * So it all works out pretty well.
547 	 */
548 	nsw_cluster_max = min(MAXPHYS / PAGE_SIZE, MAX_PAGEOUT_CLUSTER);
549 
550 	nsw_wcount_async = 4;
551 	nsw_wcount_async_max = nsw_wcount_async;
552 	mtx_init(&swbuf_mtx, "async swbuf mutex", NULL, MTX_DEF);
553 
554 	swwbuf_zone = pbuf_zsecond_create("swwbuf", nswbuf / 4);
555 	swrbuf_zone = pbuf_zsecond_create("swrbuf", nswbuf / 2);
556 
557 	/*
558 	 * Initialize our zone, taking the user's requested size or
559 	 * estimating the number we need based on the number of pages
560 	 * in the system.
561 	 */
562 	n = maxswzone != 0 ? maxswzone / sizeof(struct swblk) :
563 	    vm_cnt.v_page_count / 2;
564 	swpctrie_zone = uma_zcreate("swpctrie", pctrie_node_size(), NULL, NULL,
565 	    pctrie_zone_init, NULL, UMA_ALIGN_PTR, UMA_ZONE_VM);
566 	if (swpctrie_zone == NULL)
567 		panic("failed to create swap pctrie zone.");
568 	swblk_zone = uma_zcreate("swblk", sizeof(struct swblk), NULL, NULL,
569 	    NULL, NULL, _Alignof(struct swblk) - 1, UMA_ZONE_VM);
570 	if (swblk_zone == NULL)
571 		panic("failed to create swap blk zone.");
572 	n2 = n;
573 	do {
574 		if (uma_zone_reserve_kva(swblk_zone, n))
575 			break;
576 		/*
577 		 * if the allocation failed, try a zone two thirds the
578 		 * size of the previous attempt.
579 		 */
580 		n -= ((n + 2) / 3);
581 	} while (n > 0);
582 
583 	/*
584 	 * Often uma_zone_reserve_kva() cannot reserve exactly the
585 	 * requested size.  Account for the difference when
586 	 * calculating swap_maxpages.
587 	 */
588 	n = uma_zone_get_max(swblk_zone);
589 
590 	if (n < n2)
591 		printf("Swap blk zone entries changed from %lu to %lu.\n",
592 		    n2, n);
593 	swap_maxpages = n * SWAP_META_PAGES;
594 	swzone = n * sizeof(struct swblk);
595 	if (!uma_zone_reserve_kva(swpctrie_zone, n))
596 		printf("Cannot reserve swap pctrie zone, "
597 		    "reduce kern.maxswzone.\n");
598 }
599 
600 static vm_object_t
601 swap_pager_alloc_init(void *handle, struct ucred *cred, vm_ooffset_t size,
602     vm_ooffset_t offset)
603 {
604 	vm_object_t object;
605 
606 	if (cred != NULL) {
607 		if (!swap_reserve_by_cred(size, cred))
608 			return (NULL);
609 		crhold(cred);
610 	}
611 
612 	/*
613 	 * The un_pager.swp.swp_blks trie is initialized by
614 	 * vm_object_allocate() to ensure the correct order of
615 	 * visibility to other threads.
616 	 */
617 	object = vm_object_allocate(OBJT_SWAP, OFF_TO_IDX(offset +
618 	    PAGE_MASK + size));
619 
620 	object->un_pager.swp.writemappings = 0;
621 	object->handle = handle;
622 	if (cred != NULL) {
623 		object->cred = cred;
624 		object->charge = size;
625 	}
626 	return (object);
627 }
628 
629 /*
630  * SWAP_PAGER_ALLOC() -	allocate a new OBJT_SWAP VM object and instantiate
631  *			its metadata structures.
632  *
633  *	This routine is called from the mmap and fork code to create a new
634  *	OBJT_SWAP object.
635  *
636  *	This routine must ensure that no live duplicate is created for
637  *	the named object request, which is protected against by
638  *	holding the sw_alloc_sx lock in case handle != NULL.
639  */
640 static vm_object_t
641 swap_pager_alloc(void *handle, vm_ooffset_t size, vm_prot_t prot,
642     vm_ooffset_t offset, struct ucred *cred)
643 {
644 	vm_object_t object;
645 
646 	if (handle != NULL) {
647 		/*
648 		 * Reference existing named region or allocate new one.  There
649 		 * should not be a race here against swp_pager_meta_build()
650 		 * as called from vm_page_remove() in regards to the lookup
651 		 * of the handle.
652 		 */
653 		sx_xlock(&sw_alloc_sx);
654 		object = vm_pager_object_lookup(NOBJLIST(handle), handle);
655 		if (object == NULL) {
656 			object = swap_pager_alloc_init(handle, cred, size,
657 			    offset);
658 			if (object != NULL) {
659 				TAILQ_INSERT_TAIL(NOBJLIST(object->handle),
660 				    object, pager_object_list);
661 			}
662 		}
663 		sx_xunlock(&sw_alloc_sx);
664 	} else {
665 		object = swap_pager_alloc_init(handle, cred, size, offset);
666 	}
667 	return (object);
668 }
669 
670 /*
671  * SWAP_PAGER_DEALLOC() -	remove swap metadata from object
672  *
673  *	The swap backing for the object is destroyed.  The code is
674  *	designed such that we can reinstantiate it later, but this
675  *	routine is typically called only when the entire object is
676  *	about to be destroyed.
677  *
678  *	The object must be locked.
679  */
680 static void
681 swap_pager_dealloc(vm_object_t object)
682 {
683 
684 	VM_OBJECT_ASSERT_WLOCKED(object);
685 	KASSERT((object->flags & OBJ_DEAD) != 0, ("dealloc of reachable obj"));
686 
687 	/*
688 	 * Remove from list right away so lookups will fail if we block for
689 	 * pageout completion.
690 	 */
691 	if (object->handle != NULL) {
692 		VM_OBJECT_WUNLOCK(object);
693 		sx_xlock(&sw_alloc_sx);
694 		TAILQ_REMOVE(NOBJLIST(object->handle), object,
695 		    pager_object_list);
696 		sx_xunlock(&sw_alloc_sx);
697 		VM_OBJECT_WLOCK(object);
698 	}
699 
700 	vm_object_pip_wait(object, "swpdea");
701 
702 	/*
703 	 * Free all remaining metadata.  We only bother to free it from
704 	 * the swap meta data.  We do not attempt to free swapblk's still
705 	 * associated with vm_page_t's for this object.  We do not care
706 	 * if paging is still in progress on some objects.
707 	 */
708 	swp_pager_meta_free_all(object);
709 	object->handle = NULL;
710 	object->type = OBJT_DEAD;
711 }
712 
713 /************************************************************************
714  *			SWAP PAGER BITMAP ROUTINES			*
715  ************************************************************************/
716 
717 /*
718  * SWP_PAGER_GETSWAPSPACE() -	allocate raw swap space
719  *
720  *	Allocate swap for up to the requested number of pages, and at
721  *	least a minimum number of pages.  The starting swap block number
722  *	(a page index) is returned or SWAPBLK_NONE if the allocation
723  *	failed.
724  *
725  *	Also has the side effect of advising that somebody made a mistake
726  *	when they configured swap and didn't configure enough.
727  *
728  *	This routine may not sleep.
729  *
730  *	We allocate in round-robin fashion from the configured devices.
731  */
732 static daddr_t
733 swp_pager_getswapspace(int *io_npages, int limit)
734 {
735 	daddr_t blk;
736 	struct swdevt *sp;
737 	int mpages, npages;
738 
739 	blk = SWAPBLK_NONE;
740 	mpages = *io_npages;
741 	npages = imin(BLIST_MAX_ALLOC, mpages);
742 	mtx_lock(&sw_dev_mtx);
743 	sp = swdevhd;
744 	while (!TAILQ_EMPTY(&swtailq)) {
745 		if (sp == NULL)
746 			sp = TAILQ_FIRST(&swtailq);
747 		if ((sp->sw_flags & SW_CLOSING) == 0)
748 			blk = blist_alloc(sp->sw_blist, &npages, mpages);
749 		if (blk != SWAPBLK_NONE)
750 			break;
751 		sp = TAILQ_NEXT(sp, sw_list);
752 		if (swdevhd == sp) {
753 			if (npages <= limit)
754 				break;
755 			mpages = npages - 1;
756 			npages >>= 1;
757 		}
758 	}
759 	if (blk != SWAPBLK_NONE) {
760 		*io_npages = npages;
761 		blk += sp->sw_first;
762 		sp->sw_used += npages;
763 		swap_pager_avail -= npages;
764 		swp_sizecheck();
765 		swdevhd = TAILQ_NEXT(sp, sw_list);
766 	} else {
767 		if (swap_pager_full != 2) {
768 			printf("swp_pager_getswapspace(%d): failed\n",
769 			    *io_npages);
770 			swap_pager_full = 2;
771 			swap_pager_almost_full = 1;
772 		}
773 		swdevhd = NULL;
774 	}
775 	mtx_unlock(&sw_dev_mtx);
776 	return (blk);
777 }
778 
779 static bool
780 swp_pager_isondev(daddr_t blk, struct swdevt *sp)
781 {
782 
783 	return (blk >= sp->sw_first && blk < sp->sw_end);
784 }
785 
786 static void
787 swp_pager_strategy(struct buf *bp)
788 {
789 	struct swdevt *sp;
790 
791 	mtx_lock(&sw_dev_mtx);
792 	TAILQ_FOREACH(sp, &swtailq, sw_list) {
793 		if (swp_pager_isondev(bp->b_blkno, sp)) {
794 			mtx_unlock(&sw_dev_mtx);
795 			if ((sp->sw_flags & SW_UNMAPPED) != 0 &&
796 			    unmapped_buf_allowed) {
797 				bp->b_data = unmapped_buf;
798 				bp->b_offset = 0;
799 			} else {
800 				pmap_qenter((vm_offset_t)bp->b_data,
801 				    &bp->b_pages[0], bp->b_bcount / PAGE_SIZE);
802 			}
803 			sp->sw_strategy(bp, sp);
804 			return;
805 		}
806 	}
807 	panic("Swapdev not found");
808 }
809 
810 
811 /*
812  * SWP_PAGER_FREESWAPSPACE() -	free raw swap space
813  *
814  *	This routine returns the specified swap blocks back to the bitmap.
815  *
816  *	This routine may not sleep.
817  */
818 static void
819 swp_pager_freeswapspace(daddr_t blk, daddr_t npages)
820 {
821 	struct swdevt *sp;
822 
823 	if (npages == 0)
824 		return;
825 	mtx_lock(&sw_dev_mtx);
826 	TAILQ_FOREACH(sp, &swtailq, sw_list) {
827 		if (swp_pager_isondev(blk, sp)) {
828 			sp->sw_used -= npages;
829 			/*
830 			 * If we are attempting to stop swapping on
831 			 * this device, we don't want to mark any
832 			 * blocks free lest they be reused.
833 			 */
834 			if ((sp->sw_flags & SW_CLOSING) == 0) {
835 				blist_free(sp->sw_blist, blk - sp->sw_first,
836 				    npages);
837 				swap_pager_avail += npages;
838 				swp_sizecheck();
839 			}
840 			mtx_unlock(&sw_dev_mtx);
841 			return;
842 		}
843 	}
844 	panic("Swapdev not found");
845 }
846 
847 /*
848  * SYSCTL_SWAP_FRAGMENTATION() -	produce raw swap space stats
849  */
850 static int
851 sysctl_swap_fragmentation(SYSCTL_HANDLER_ARGS)
852 {
853 	struct sbuf sbuf;
854 	struct swdevt *sp;
855 	const char *devname;
856 	int error;
857 
858 	error = sysctl_wire_old_buffer(req, 0);
859 	if (error != 0)
860 		return (error);
861 	sbuf_new_for_sysctl(&sbuf, NULL, 128, req);
862 	mtx_lock(&sw_dev_mtx);
863 	TAILQ_FOREACH(sp, &swtailq, sw_list) {
864 		if (vn_isdisk(sp->sw_vp, NULL))
865 			devname = devtoname(sp->sw_vp->v_rdev);
866 		else
867 			devname = "[file]";
868 		sbuf_printf(&sbuf, "\nFree space on device %s:\n", devname);
869 		blist_stats(sp->sw_blist, &sbuf);
870 	}
871 	mtx_unlock(&sw_dev_mtx);
872 	error = sbuf_finish(&sbuf);
873 	sbuf_delete(&sbuf);
874 	return (error);
875 }
876 
877 /*
878  * SWAP_PAGER_FREESPACE() -	frees swap blocks associated with a page
879  *				range within an object.
880  *
881  *	This is a globally accessible routine.
882  *
883  *	This routine removes swapblk assignments from swap metadata.
884  *
885  *	The external callers of this routine typically have already destroyed
886  *	or renamed vm_page_t's associated with this range in the object so
887  *	we should be ok.
888  *
889  *	The object must be locked.
890  */
891 void
892 swap_pager_freespace(vm_object_t object, vm_pindex_t start, vm_size_t size)
893 {
894 
895 	swp_pager_meta_free(object, start, size);
896 }
897 
898 /*
899  * SWAP_PAGER_RESERVE() - reserve swap blocks in object
900  *
901  *	Assigns swap blocks to the specified range within the object.  The
902  *	swap blocks are not zeroed.  Any previous swap assignment is destroyed.
903  *
904  *	Returns 0 on success, -1 on failure.
905  */
906 int
907 swap_pager_reserve(vm_object_t object, vm_pindex_t start, vm_size_t size)
908 {
909 	daddr_t addr, blk, n_free, s_free;
910 	int i, j, n;
911 
912 	swp_pager_init_freerange(&s_free, &n_free);
913 	VM_OBJECT_WLOCK(object);
914 	for (i = 0; i < size; i += n) {
915 		n = size - i;
916 		blk = swp_pager_getswapspace(&n, 1);
917 		if (blk == SWAPBLK_NONE) {
918 			swp_pager_meta_free(object, start, i);
919 			VM_OBJECT_WUNLOCK(object);
920 			return (-1);
921 		}
922 		for (j = 0; j < n; ++j) {
923 			addr = swp_pager_meta_build(object,
924 			    start + i + j, blk + j);
925 			if (addr != SWAPBLK_NONE)
926 				swp_pager_update_freerange(&s_free, &n_free,
927 				    addr);
928 		}
929 	}
930 	swp_pager_freeswapspace(s_free, n_free);
931 	VM_OBJECT_WUNLOCK(object);
932 	return (0);
933 }
934 
935 /*
936  * SWAP_PAGER_COPY() -  copy blocks from source pager to destination pager
937  *			and destroy the source.
938  *
939  *	Copy any valid swapblks from the source to the destination.  In
940  *	cases where both the source and destination have a valid swapblk,
941  *	we keep the destination's.
942  *
943  *	This routine is allowed to sleep.  It may sleep allocating metadata
944  *	indirectly through swp_pager_meta_build() or if paging is still in
945  *	progress on the source.
946  *
947  *	The source object contains no vm_page_t's (which is just as well)
948  *
949  *	The source object is of type OBJT_SWAP.
950  *
951  *	The source and destination objects must be locked.
952  *	Both object locks may temporarily be released.
953  */
954 void
955 swap_pager_copy(vm_object_t srcobject, vm_object_t dstobject,
956     vm_pindex_t offset, int destroysource)
957 {
958 	vm_pindex_t i;
959 	daddr_t dstaddr, n_free, s_free, srcaddr;
960 
961 	VM_OBJECT_ASSERT_WLOCKED(srcobject);
962 	VM_OBJECT_ASSERT_WLOCKED(dstobject);
963 
964 	/*
965 	 * If destroysource is set, we remove the source object from the
966 	 * swap_pager internal queue now.
967 	 */
968 	if (destroysource && srcobject->handle != NULL) {
969 		vm_object_pip_add(srcobject, 1);
970 		VM_OBJECT_WUNLOCK(srcobject);
971 		vm_object_pip_add(dstobject, 1);
972 		VM_OBJECT_WUNLOCK(dstobject);
973 		sx_xlock(&sw_alloc_sx);
974 		TAILQ_REMOVE(NOBJLIST(srcobject->handle), srcobject,
975 		    pager_object_list);
976 		sx_xunlock(&sw_alloc_sx);
977 		VM_OBJECT_WLOCK(dstobject);
978 		vm_object_pip_wakeup(dstobject);
979 		VM_OBJECT_WLOCK(srcobject);
980 		vm_object_pip_wakeup(srcobject);
981 	}
982 
983 	/*
984 	 * Transfer source to destination.
985 	 */
986 	swp_pager_init_freerange(&s_free, &n_free);
987 	for (i = 0; i < dstobject->size; ++i) {
988 		srcaddr = swp_pager_meta_ctl(srcobject, i + offset, SWM_POP);
989 		if (srcaddr == SWAPBLK_NONE)
990 			continue;
991 		dstaddr = swp_pager_meta_ctl(dstobject, i, 0);
992 		if (dstaddr != SWAPBLK_NONE) {
993 			/*
994 			 * Destination has valid swapblk or it is represented
995 			 * by a resident page.  We destroy the source block.
996 			 */
997 			swp_pager_update_freerange(&s_free, &n_free, srcaddr);
998 			continue;
999 		}
1000 
1001 		/*
1002 		 * Destination has no swapblk and is not resident,
1003 		 * copy source.
1004 		 *
1005 		 * swp_pager_meta_build() can sleep.
1006 		 */
1007 		vm_object_pip_add(srcobject, 1);
1008 		VM_OBJECT_WUNLOCK(srcobject);
1009 		vm_object_pip_add(dstobject, 1);
1010 		dstaddr = swp_pager_meta_build(dstobject, i, srcaddr);
1011 		KASSERT(dstaddr == SWAPBLK_NONE,
1012 		    ("Unexpected destination swapblk"));
1013 		vm_object_pip_wakeup(dstobject);
1014 		VM_OBJECT_WLOCK(srcobject);
1015 		vm_object_pip_wakeup(srcobject);
1016 	}
1017 	swp_pager_freeswapspace(s_free, n_free);
1018 
1019 	/*
1020 	 * Free left over swap blocks in source.
1021 	 *
1022 	 * We have to revert the type to OBJT_DEFAULT so we do not accidentally
1023 	 * double-remove the object from the swap queues.
1024 	 */
1025 	if (destroysource) {
1026 		swp_pager_meta_free_all(srcobject);
1027 		/*
1028 		 * Reverting the type is not necessary, the caller is going
1029 		 * to destroy srcobject directly, but I'm doing it here
1030 		 * for consistency since we've removed the object from its
1031 		 * queues.
1032 		 */
1033 		srcobject->type = OBJT_DEFAULT;
1034 	}
1035 }
1036 
1037 /*
1038  * SWAP_PAGER_HASPAGE() -	determine if we have good backing store for
1039  *				the requested page.
1040  *
1041  *	We determine whether good backing store exists for the requested
1042  *	page and return TRUE if it does, FALSE if it doesn't.
1043  *
1044  *	If TRUE, we also try to determine how much valid, contiguous backing
1045  *	store exists before and after the requested page.
1046  */
1047 static boolean_t
1048 swap_pager_haspage(vm_object_t object, vm_pindex_t pindex, int *before,
1049     int *after)
1050 {
1051 	daddr_t blk, blk0;
1052 	int i;
1053 
1054 	VM_OBJECT_ASSERT_LOCKED(object);
1055 
1056 	/*
1057 	 * do we have good backing store at the requested index ?
1058 	 */
1059 	blk0 = swp_pager_meta_ctl(object, pindex, 0);
1060 	if (blk0 == SWAPBLK_NONE) {
1061 		if (before)
1062 			*before = 0;
1063 		if (after)
1064 			*after = 0;
1065 		return (FALSE);
1066 	}
1067 
1068 	/*
1069 	 * find backwards-looking contiguous good backing store
1070 	 */
1071 	if (before != NULL) {
1072 		for (i = 1; i < SWB_NPAGES; i++) {
1073 			if (i > pindex)
1074 				break;
1075 			blk = swp_pager_meta_ctl(object, pindex - i, 0);
1076 			if (blk != blk0 - i)
1077 				break;
1078 		}
1079 		*before = i - 1;
1080 	}
1081 
1082 	/*
1083 	 * find forward-looking contiguous good backing store
1084 	 */
1085 	if (after != NULL) {
1086 		for (i = 1; i < SWB_NPAGES; i++) {
1087 			blk = swp_pager_meta_ctl(object, pindex + i, 0);
1088 			if (blk != blk0 + i)
1089 				break;
1090 		}
1091 		*after = i - 1;
1092 	}
1093 	return (TRUE);
1094 }
1095 
1096 /*
1097  * SWAP_PAGER_PAGE_UNSWAPPED() - remove swap backing store related to page
1098  *
1099  *	This removes any associated swap backing store, whether valid or
1100  *	not, from the page.
1101  *
1102  *	This routine is typically called when a page is made dirty, at
1103  *	which point any associated swap can be freed.  MADV_FREE also
1104  *	calls us in a special-case situation
1105  *
1106  *	NOTE!!!  If the page is clean and the swap was valid, the caller
1107  *	should make the page dirty before calling this routine.  This routine
1108  *	does NOT change the m->dirty status of the page.  Also: MADV_FREE
1109  *	depends on it.
1110  *
1111  *	This routine may not sleep.
1112  *
1113  *	The object containing the page must be locked.
1114  */
1115 static void
1116 swap_pager_unswapped(vm_page_t m)
1117 {
1118 	daddr_t srcaddr;
1119 
1120 	srcaddr = swp_pager_meta_ctl(m->object, m->pindex, SWM_POP);
1121 	if (srcaddr != SWAPBLK_NONE)
1122 		swp_pager_freeswapspace(srcaddr, 1);
1123 }
1124 
1125 /*
1126  * swap_pager_getpages() - bring pages in from swap
1127  *
1128  *	Attempt to page in the pages in array "ma" of length "count".  The
1129  *	caller may optionally specify that additional pages preceding and
1130  *	succeeding the specified range be paged in.  The number of such pages
1131  *	is returned in the "rbehind" and "rahead" parameters, and they will
1132  *	be in the inactive queue upon return.
1133  *
1134  *	The pages in "ma" must be busied and will remain busied upon return.
1135  */
1136 static int
1137 swap_pager_getpages(vm_object_t object, vm_page_t *ma, int count, int *rbehind,
1138     int *rahead)
1139 {
1140 	struct buf *bp;
1141 	vm_page_t bm, mpred, msucc, p;
1142 	vm_pindex_t pindex;
1143 	daddr_t blk;
1144 	int i, maxahead, maxbehind, reqcount;
1145 
1146 	reqcount = count;
1147 
1148 	/*
1149 	 * Determine the final number of read-behind pages and
1150 	 * allocate them BEFORE releasing the object lock.  Otherwise,
1151 	 * there can be a problematic race with vm_object_split().
1152 	 * Specifically, vm_object_split() might first transfer pages
1153 	 * that precede ma[0] in the current object to a new object,
1154 	 * and then this function incorrectly recreates those pages as
1155 	 * read-behind pages in the current object.
1156 	 */
1157 	if (!swap_pager_haspage(object, ma[0]->pindex, &maxbehind, &maxahead))
1158 		return (VM_PAGER_FAIL);
1159 
1160 	/*
1161 	 * Clip the readahead and readbehind ranges to exclude resident pages.
1162 	 */
1163 	if (rahead != NULL) {
1164 		KASSERT(reqcount - 1 <= maxahead,
1165 		    ("page count %d extends beyond swap block", reqcount));
1166 		*rahead = imin(*rahead, maxahead - (reqcount - 1));
1167 		pindex = ma[reqcount - 1]->pindex;
1168 		msucc = TAILQ_NEXT(ma[reqcount - 1], listq);
1169 		if (msucc != NULL && msucc->pindex - pindex - 1 < *rahead)
1170 			*rahead = msucc->pindex - pindex - 1;
1171 	}
1172 	if (rbehind != NULL) {
1173 		*rbehind = imin(*rbehind, maxbehind);
1174 		pindex = ma[0]->pindex;
1175 		mpred = TAILQ_PREV(ma[0], pglist, listq);
1176 		if (mpred != NULL && pindex - mpred->pindex - 1 < *rbehind)
1177 			*rbehind = pindex - mpred->pindex - 1;
1178 	}
1179 
1180 	bm = ma[0];
1181 	for (i = 0; i < count; i++)
1182 		ma[i]->oflags |= VPO_SWAPINPROG;
1183 
1184 	/*
1185 	 * Allocate readahead and readbehind pages.
1186 	 */
1187 	if (rbehind != NULL) {
1188 		for (i = 1; i <= *rbehind; i++) {
1189 			p = vm_page_alloc(object, ma[0]->pindex - i,
1190 			    VM_ALLOC_NORMAL);
1191 			if (p == NULL)
1192 				break;
1193 			p->oflags |= VPO_SWAPINPROG;
1194 			bm = p;
1195 		}
1196 		*rbehind = i - 1;
1197 	}
1198 	if (rahead != NULL) {
1199 		for (i = 0; i < *rahead; i++) {
1200 			p = vm_page_alloc(object,
1201 			    ma[reqcount - 1]->pindex + i + 1, VM_ALLOC_NORMAL);
1202 			if (p == NULL)
1203 				break;
1204 			p->oflags |= VPO_SWAPINPROG;
1205 		}
1206 		*rahead = i;
1207 	}
1208 	if (rbehind != NULL)
1209 		count += *rbehind;
1210 	if (rahead != NULL)
1211 		count += *rahead;
1212 
1213 	vm_object_pip_add(object, count);
1214 
1215 	pindex = bm->pindex;
1216 	blk = swp_pager_meta_ctl(object, pindex, 0);
1217 	KASSERT(blk != SWAPBLK_NONE,
1218 	    ("no swap blocking containing %p(%jx)", object, (uintmax_t)pindex));
1219 
1220 	VM_OBJECT_WUNLOCK(object);
1221 	bp = uma_zalloc(swrbuf_zone, M_WAITOK);
1222 	/* Pages cannot leave the object while busy. */
1223 	for (i = 0, p = bm; i < count; i++, p = TAILQ_NEXT(p, listq)) {
1224 		MPASS(p->pindex == bm->pindex + i);
1225 		bp->b_pages[i] = p;
1226 	}
1227 
1228 	bp->b_flags |= B_PAGING;
1229 	bp->b_iocmd = BIO_READ;
1230 	bp->b_iodone = swp_pager_async_iodone;
1231 	bp->b_rcred = crhold(thread0.td_ucred);
1232 	bp->b_wcred = crhold(thread0.td_ucred);
1233 	bp->b_blkno = blk;
1234 	bp->b_bcount = PAGE_SIZE * count;
1235 	bp->b_bufsize = PAGE_SIZE * count;
1236 	bp->b_npages = count;
1237 	bp->b_pgbefore = rbehind != NULL ? *rbehind : 0;
1238 	bp->b_pgafter = rahead != NULL ? *rahead : 0;
1239 
1240 	VM_CNT_INC(v_swapin);
1241 	VM_CNT_ADD(v_swappgsin, count);
1242 
1243 	/*
1244 	 * perform the I/O.  NOTE!!!  bp cannot be considered valid after
1245 	 * this point because we automatically release it on completion.
1246 	 * Instead, we look at the one page we are interested in which we
1247 	 * still hold a lock on even through the I/O completion.
1248 	 *
1249 	 * The other pages in our ma[] array are also released on completion,
1250 	 * so we cannot assume they are valid anymore either.
1251 	 *
1252 	 * NOTE: b_blkno is destroyed by the call to swapdev_strategy
1253 	 */
1254 	BUF_KERNPROC(bp);
1255 	swp_pager_strategy(bp);
1256 
1257 	/*
1258 	 * Wait for the pages we want to complete.  VPO_SWAPINPROG is always
1259 	 * cleared on completion.  If an I/O error occurs, SWAPBLK_NONE
1260 	 * is set in the metadata for each page in the request.
1261 	 */
1262 	VM_OBJECT_WLOCK(object);
1263 	while ((ma[0]->oflags & VPO_SWAPINPROG) != 0) {
1264 		ma[0]->oflags |= VPO_SWAPSLEEP;
1265 		VM_CNT_INC(v_intrans);
1266 		if (VM_OBJECT_SLEEP(object, &object->handle, PSWP,
1267 		    "swread", hz * 20)) {
1268 			printf(
1269 "swap_pager: indefinite wait buffer: bufobj: %p, blkno: %jd, size: %ld\n",
1270 			    bp->b_bufobj, (intmax_t)bp->b_blkno, bp->b_bcount);
1271 		}
1272 	}
1273 
1274 	/*
1275 	 * If we had an unrecoverable read error pages will not be valid.
1276 	 */
1277 	for (i = 0; i < reqcount; i++)
1278 		if (ma[i]->valid != VM_PAGE_BITS_ALL)
1279 			return (VM_PAGER_ERROR);
1280 
1281 	return (VM_PAGER_OK);
1282 
1283 	/*
1284 	 * A final note: in a low swap situation, we cannot deallocate swap
1285 	 * and mark a page dirty here because the caller is likely to mark
1286 	 * the page clean when we return, causing the page to possibly revert
1287 	 * to all-zero's later.
1288 	 */
1289 }
1290 
1291 /*
1292  * 	swap_pager_getpages_async():
1293  *
1294  *	Right now this is emulation of asynchronous operation on top of
1295  *	swap_pager_getpages().
1296  */
1297 static int
1298 swap_pager_getpages_async(vm_object_t object, vm_page_t *ma, int count,
1299     int *rbehind, int *rahead, pgo_getpages_iodone_t iodone, void *arg)
1300 {
1301 	int r, error;
1302 
1303 	r = swap_pager_getpages(object, ma, count, rbehind, rahead);
1304 	VM_OBJECT_WUNLOCK(object);
1305 	switch (r) {
1306 	case VM_PAGER_OK:
1307 		error = 0;
1308 		break;
1309 	case VM_PAGER_ERROR:
1310 		error = EIO;
1311 		break;
1312 	case VM_PAGER_FAIL:
1313 		error = EINVAL;
1314 		break;
1315 	default:
1316 		panic("unhandled swap_pager_getpages() error %d", r);
1317 	}
1318 	(iodone)(arg, ma, count, error);
1319 	VM_OBJECT_WLOCK(object);
1320 
1321 	return (r);
1322 }
1323 
1324 /*
1325  *	swap_pager_putpages:
1326  *
1327  *	Assign swap (if necessary) and initiate I/O on the specified pages.
1328  *
1329  *	We support both OBJT_DEFAULT and OBJT_SWAP objects.  DEFAULT objects
1330  *	are automatically converted to SWAP objects.
1331  *
1332  *	In a low memory situation we may block in VOP_STRATEGY(), but the new
1333  *	vm_page reservation system coupled with properly written VFS devices
1334  *	should ensure that no low-memory deadlock occurs.  This is an area
1335  *	which needs work.
1336  *
1337  *	The parent has N vm_object_pip_add() references prior to
1338  *	calling us and will remove references for rtvals[] that are
1339  *	not set to VM_PAGER_PEND.  We need to remove the rest on I/O
1340  *	completion.
1341  *
1342  *	The parent has soft-busy'd the pages it passes us and will unbusy
1343  *	those whose rtvals[] entry is not set to VM_PAGER_PEND on return.
1344  *	We need to unbusy the rest on I/O completion.
1345  */
1346 static void
1347 swap_pager_putpages(vm_object_t object, vm_page_t *ma, int count,
1348     int flags, int *rtvals)
1349 {
1350 	struct buf *bp;
1351 	daddr_t addr, blk, n_free, s_free;
1352 	vm_page_t mreq;
1353 	int i, j, n;
1354 	bool async;
1355 
1356 	KASSERT(count == 0 || ma[0]->object == object,
1357 	    ("%s: object mismatch %p/%p",
1358 	    __func__, object, ma[0]->object));
1359 
1360 	/*
1361 	 * Step 1
1362 	 *
1363 	 * Turn object into OBJT_SWAP.  Force sync if not a pageout process.
1364 	 */
1365 	if (object->type != OBJT_SWAP) {
1366 		addr = swp_pager_meta_build(object, 0, SWAPBLK_NONE);
1367 		KASSERT(addr == SWAPBLK_NONE,
1368 		    ("unexpected object swap block"));
1369 	}
1370 	VM_OBJECT_WUNLOCK(object);
1371 	async = curproc == pageproc && (flags & VM_PAGER_PUT_SYNC) == 0;
1372 	swp_pager_init_freerange(&s_free, &n_free);
1373 
1374 	/*
1375 	 * Step 2
1376 	 *
1377 	 * Assign swap blocks and issue I/O.  We reallocate swap on the fly.
1378 	 * The page is left dirty until the pageout operation completes
1379 	 * successfully.
1380 	 */
1381 	for (i = 0; i < count; i += n) {
1382 		/* Maximum I/O size is limited by maximum swap block size. */
1383 		n = min(count - i, nsw_cluster_max);
1384 
1385 		/* Get a block of swap of size up to size n. */
1386 		blk = swp_pager_getswapspace(&n, 4);
1387 		if (blk == SWAPBLK_NONE) {
1388 			for (j = 0; j < n; ++j)
1389 				rtvals[i + j] = VM_PAGER_FAIL;
1390 			continue;
1391 		}
1392 
1393 		/*
1394 		 * All I/O parameters have been satisfied.  Build the I/O
1395 		 * request and assign the swap space.
1396 		 */
1397 		if (async) {
1398 			mtx_lock(&swbuf_mtx);
1399 			while (nsw_wcount_async == 0)
1400 				msleep(&nsw_wcount_async, &swbuf_mtx, PVM,
1401 				    "swbufa", 0);
1402 			nsw_wcount_async--;
1403 			mtx_unlock(&swbuf_mtx);
1404 		}
1405 		bp = uma_zalloc(swwbuf_zone, M_WAITOK);
1406 		if (async)
1407 			bp->b_flags = B_ASYNC;
1408 		bp->b_flags |= B_PAGING;
1409 		bp->b_iocmd = BIO_WRITE;
1410 
1411 		bp->b_rcred = crhold(thread0.td_ucred);
1412 		bp->b_wcred = crhold(thread0.td_ucred);
1413 		bp->b_bcount = PAGE_SIZE * n;
1414 		bp->b_bufsize = PAGE_SIZE * n;
1415 		bp->b_blkno = blk;
1416 
1417 		VM_OBJECT_WLOCK(object);
1418 		for (j = 0; j < n; ++j) {
1419 			mreq = ma[i + j];
1420 			addr = swp_pager_meta_build(mreq->object, mreq->pindex,
1421 			    blk + j);
1422 			if (addr != SWAPBLK_NONE)
1423 				swp_pager_update_freerange(&s_free, &n_free,
1424 				    addr);
1425 			MPASS(mreq->dirty == VM_PAGE_BITS_ALL);
1426 			mreq->oflags |= VPO_SWAPINPROG;
1427 			bp->b_pages[j] = mreq;
1428 		}
1429 		VM_OBJECT_WUNLOCK(object);
1430 		bp->b_npages = n;
1431 		/*
1432 		 * Must set dirty range for NFS to work.
1433 		 */
1434 		bp->b_dirtyoff = 0;
1435 		bp->b_dirtyend = bp->b_bcount;
1436 
1437 		VM_CNT_INC(v_swapout);
1438 		VM_CNT_ADD(v_swappgsout, bp->b_npages);
1439 
1440 		/*
1441 		 * We unconditionally set rtvals[] to VM_PAGER_PEND so that we
1442 		 * can call the async completion routine at the end of a
1443 		 * synchronous I/O operation.  Otherwise, our caller would
1444 		 * perform duplicate unbusy and wakeup operations on the page
1445 		 * and object, respectively.
1446 		 */
1447 		for (j = 0; j < n; j++)
1448 			rtvals[i + j] = VM_PAGER_PEND;
1449 
1450 		/*
1451 		 * asynchronous
1452 		 *
1453 		 * NOTE: b_blkno is destroyed by the call to swapdev_strategy.
1454 		 */
1455 		if (async) {
1456 			bp->b_iodone = swp_pager_async_iodone;
1457 			BUF_KERNPROC(bp);
1458 			swp_pager_strategy(bp);
1459 			continue;
1460 		}
1461 
1462 		/*
1463 		 * synchronous
1464 		 *
1465 		 * NOTE: b_blkno is destroyed by the call to swapdev_strategy.
1466 		 */
1467 		bp->b_iodone = bdone;
1468 		swp_pager_strategy(bp);
1469 
1470 		/*
1471 		 * Wait for the sync I/O to complete.
1472 		 */
1473 		bwait(bp, PVM, "swwrt");
1474 
1475 		/*
1476 		 * Now that we are through with the bp, we can call the
1477 		 * normal async completion, which frees everything up.
1478 		 */
1479 		swp_pager_async_iodone(bp);
1480 	}
1481 	swp_pager_freeswapspace(s_free, n_free);
1482 	VM_OBJECT_WLOCK(object);
1483 }
1484 
1485 /*
1486  *	swp_pager_async_iodone:
1487  *
1488  *	Completion routine for asynchronous reads and writes from/to swap.
1489  *	Also called manually by synchronous code to finish up a bp.
1490  *
1491  *	This routine may not sleep.
1492  */
1493 static void
1494 swp_pager_async_iodone(struct buf *bp)
1495 {
1496 	int i;
1497 	vm_object_t object = NULL;
1498 
1499 	/*
1500 	 * Report error - unless we ran out of memory, in which case
1501 	 * we've already logged it in swapgeom_strategy().
1502 	 */
1503 	if (bp->b_ioflags & BIO_ERROR && bp->b_error != ENOMEM) {
1504 		printf(
1505 		    "swap_pager: I/O error - %s failed; blkno %ld,"
1506 			"size %ld, error %d\n",
1507 		    ((bp->b_iocmd == BIO_READ) ? "pagein" : "pageout"),
1508 		    (long)bp->b_blkno,
1509 		    (long)bp->b_bcount,
1510 		    bp->b_error
1511 		);
1512 	}
1513 
1514 	/*
1515 	 * remove the mapping for kernel virtual
1516 	 */
1517 	if (buf_mapped(bp))
1518 		pmap_qremove((vm_offset_t)bp->b_data, bp->b_npages);
1519 	else
1520 		bp->b_data = bp->b_kvabase;
1521 
1522 	if (bp->b_npages) {
1523 		object = bp->b_pages[0]->object;
1524 		VM_OBJECT_WLOCK(object);
1525 	}
1526 
1527 	/*
1528 	 * cleanup pages.  If an error occurs writing to swap, we are in
1529 	 * very serious trouble.  If it happens to be a disk error, though,
1530 	 * we may be able to recover by reassigning the swap later on.  So
1531 	 * in this case we remove the m->swapblk assignment for the page
1532 	 * but do not free it in the rlist.  The errornous block(s) are thus
1533 	 * never reallocated as swap.  Redirty the page and continue.
1534 	 */
1535 	for (i = 0; i < bp->b_npages; ++i) {
1536 		vm_page_t m = bp->b_pages[i];
1537 
1538 		m->oflags &= ~VPO_SWAPINPROG;
1539 		if (m->oflags & VPO_SWAPSLEEP) {
1540 			m->oflags &= ~VPO_SWAPSLEEP;
1541 			wakeup(&object->handle);
1542 		}
1543 
1544 		if (bp->b_ioflags & BIO_ERROR) {
1545 			/*
1546 			 * If an error occurs I'd love to throw the swapblk
1547 			 * away without freeing it back to swapspace, so it
1548 			 * can never be used again.  But I can't from an
1549 			 * interrupt.
1550 			 */
1551 			if (bp->b_iocmd == BIO_READ) {
1552 				/*
1553 				 * NOTE: for reads, m->dirty will probably
1554 				 * be overridden by the original caller of
1555 				 * getpages so don't play cute tricks here.
1556 				 */
1557 				m->valid = 0;
1558 			} else {
1559 				/*
1560 				 * If a write error occurs, reactivate page
1561 				 * so it doesn't clog the inactive list,
1562 				 * then finish the I/O.
1563 				 */
1564 				MPASS(m->dirty == VM_PAGE_BITS_ALL);
1565 				vm_page_lock(m);
1566 				vm_page_activate(m);
1567 				vm_page_unlock(m);
1568 				vm_page_sunbusy(m);
1569 			}
1570 		} else if (bp->b_iocmd == BIO_READ) {
1571 			/*
1572 			 * NOTE: for reads, m->dirty will probably be
1573 			 * overridden by the original caller of getpages so
1574 			 * we cannot set them in order to free the underlying
1575 			 * swap in a low-swap situation.  I don't think we'd
1576 			 * want to do that anyway, but it was an optimization
1577 			 * that existed in the old swapper for a time before
1578 			 * it got ripped out due to precisely this problem.
1579 			 */
1580 			KASSERT(!pmap_page_is_mapped(m),
1581 			    ("swp_pager_async_iodone: page %p is mapped", m));
1582 			KASSERT(m->dirty == 0,
1583 			    ("swp_pager_async_iodone: page %p is dirty", m));
1584 
1585 			m->valid = VM_PAGE_BITS_ALL;
1586 			if (i < bp->b_pgbefore ||
1587 			    i >= bp->b_npages - bp->b_pgafter)
1588 				vm_page_readahead_finish(m);
1589 		} else {
1590 			/*
1591 			 * For write success, clear the dirty
1592 			 * status, then finish the I/O ( which decrements the
1593 			 * busy count and possibly wakes waiter's up ).
1594 			 * A page is only written to swap after a period of
1595 			 * inactivity.  Therefore, we do not expect it to be
1596 			 * reused.
1597 			 */
1598 			KASSERT(!pmap_page_is_write_mapped(m),
1599 			    ("swp_pager_async_iodone: page %p is not write"
1600 			    " protected", m));
1601 			vm_page_undirty(m);
1602 			vm_page_lock(m);
1603 			vm_page_deactivate_noreuse(m);
1604 			vm_page_unlock(m);
1605 			vm_page_sunbusy(m);
1606 		}
1607 	}
1608 
1609 	/*
1610 	 * adjust pip.  NOTE: the original parent may still have its own
1611 	 * pip refs on the object.
1612 	 */
1613 	if (object != NULL) {
1614 		vm_object_pip_wakeupn(object, bp->b_npages);
1615 		VM_OBJECT_WUNLOCK(object);
1616 	}
1617 
1618 	/*
1619 	 * swapdev_strategy() manually sets b_vp and b_bufobj before calling
1620 	 * bstrategy(). Set them back to NULL now we're done with it, or we'll
1621 	 * trigger a KASSERT in relpbuf().
1622 	 */
1623 	if (bp->b_vp) {
1624 		    bp->b_vp = NULL;
1625 		    bp->b_bufobj = NULL;
1626 	}
1627 	/*
1628 	 * release the physical I/O buffer
1629 	 */
1630 	if (bp->b_flags & B_ASYNC) {
1631 		mtx_lock(&swbuf_mtx);
1632 		if (++nsw_wcount_async == 1)
1633 			wakeup(&nsw_wcount_async);
1634 		mtx_unlock(&swbuf_mtx);
1635 	}
1636 	uma_zfree((bp->b_iocmd == BIO_READ) ? swrbuf_zone : swwbuf_zone, bp);
1637 }
1638 
1639 int
1640 swap_pager_nswapdev(void)
1641 {
1642 
1643 	return (nswapdev);
1644 }
1645 
1646 static void
1647 swp_pager_force_dirty(vm_page_t m)
1648 {
1649 
1650 	vm_page_dirty(m);
1651 #ifdef INVARIANTS
1652 	vm_page_lock(m);
1653 	if (!vm_page_wired(m) && m->queue == PQ_NONE)
1654 		panic("page %p is neither wired nor queued", m);
1655 	vm_page_unlock(m);
1656 #endif
1657 	vm_page_xunbusy(m);
1658 	swap_pager_unswapped(m);
1659 }
1660 
1661 static void
1662 swp_pager_force_launder(vm_page_t m)
1663 {
1664 
1665 	vm_page_dirty(m);
1666 	vm_page_lock(m);
1667 	vm_page_launder(m);
1668 	vm_page_unlock(m);
1669 	vm_page_xunbusy(m);
1670 	swap_pager_unswapped(m);
1671 }
1672 
1673 /*
1674  * SWP_PAGER_FORCE_PAGEIN() - force swap blocks to be paged in
1675  *
1676  *	This routine dissociates pages starting at the given index within an
1677  *	object from their backing store, paging them in if they do not reside
1678  *	in memory.  Pages that are paged in are marked dirty and placed in the
1679  *	laundry queue.  Pages are marked dirty because they no longer have
1680  *	backing store.  They are placed in the laundry queue because they have
1681  *	not been accessed recently.  Otherwise, they would already reside in
1682  *	memory.
1683  */
1684 static void
1685 swp_pager_force_pagein(vm_object_t object, vm_pindex_t pindex, int npages)
1686 {
1687 	vm_page_t ma[npages];
1688 	int i, j;
1689 
1690 	KASSERT(npages > 0, ("%s: No pages", __func__));
1691 	KASSERT(npages <= MAXPHYS / PAGE_SIZE,
1692 	    ("%s: Too many pages: %d", __func__, npages));
1693 	vm_object_pip_add(object, npages);
1694 	vm_page_grab_pages(object, pindex, VM_ALLOC_NORMAL, ma, npages);
1695 	for (i = j = 0;; i++) {
1696 		/* Count nonresident pages, to page-in all at once. */
1697 		if (i < npages && ma[i]->valid != VM_PAGE_BITS_ALL)
1698 			continue;
1699 		if (j < i) {
1700 			/* Page-in nonresident pages. Mark for laundering. */
1701 			if (swap_pager_getpages(object, &ma[j], i - j, NULL,
1702 			    NULL) != VM_PAGER_OK)
1703 				panic("%s: read from swap failed", __func__);
1704 			do {
1705 				swp_pager_force_launder(ma[j]);
1706 			} while (++j < i);
1707 		}
1708 		if (i == npages)
1709 			break;
1710 		/* Mark dirty a resident page. */
1711 		swp_pager_force_dirty(ma[j++]);
1712 	}
1713 	vm_object_pip_wakeupn(object, npages);
1714 }
1715 
1716 /*
1717  *	swap_pager_swapoff_object:
1718  *
1719  *	Page in all of the pages that have been paged out for an object
1720  *	to a swap device.
1721  */
1722 static void
1723 swap_pager_swapoff_object(struct swdevt *sp, vm_object_t object)
1724 {
1725 	struct swblk *sb;
1726 	vm_pindex_t pi, s_pindex;
1727 	daddr_t blk, n_blks, s_blk;
1728 	int i;
1729 
1730 	n_blks = 0;
1731 	for (pi = 0; (sb = SWAP_PCTRIE_LOOKUP_GE(
1732 	    &object->un_pager.swp.swp_blks, pi)) != NULL; ) {
1733 		for (i = 0; i < SWAP_META_PAGES; i++) {
1734 			blk = sb->d[i];
1735 			if (!swp_pager_isondev(blk, sp))
1736 				blk = SWAPBLK_NONE;
1737 
1738 			/*
1739 			 * If there are no blocks/pages accumulated, start a new
1740 			 * accumulation here.
1741 			 */
1742 			if (n_blks == 0) {
1743 				if (blk != SWAPBLK_NONE) {
1744 					s_blk = blk;
1745 					s_pindex = sb->p + i;
1746 					n_blks = 1;
1747 				}
1748 				continue;
1749 			}
1750 
1751 			/*
1752 			 * If the accumulation can be extended without breaking
1753 			 * the sequence of consecutive blocks and pages that
1754 			 * swp_pager_force_pagein() depends on, do so.
1755 			 */
1756 			if (n_blks < MAXPHYS / PAGE_SIZE &&
1757 			    s_blk + n_blks == blk &&
1758 			    s_pindex + n_blks == sb->p + i) {
1759 				++n_blks;
1760 				continue;
1761 			}
1762 
1763 			/*
1764 			 * The sequence of consecutive blocks and pages cannot
1765 			 * be extended, so page them all in here.  Then,
1766 			 * because doing so involves releasing and reacquiring
1767 			 * a lock that protects the swap block pctrie, do not
1768 			 * rely on the current swap block.  Break this loop and
1769 			 * re-fetch the same pindex from the pctrie again.
1770 			 */
1771 			swp_pager_force_pagein(object, s_pindex, n_blks);
1772 			n_blks = 0;
1773 			break;
1774 		}
1775 		if (i == SWAP_META_PAGES)
1776 			pi = sb->p + SWAP_META_PAGES;
1777 	}
1778 	if (n_blks > 0)
1779 		swp_pager_force_pagein(object, s_pindex, n_blks);
1780 }
1781 
1782 /*
1783  *	swap_pager_swapoff:
1784  *
1785  *	Page in all of the pages that have been paged out to the
1786  *	given device.  The corresponding blocks in the bitmap must be
1787  *	marked as allocated and the device must be flagged SW_CLOSING.
1788  *	There may be no processes swapped out to the device.
1789  *
1790  *	This routine may block.
1791  */
1792 static void
1793 swap_pager_swapoff(struct swdevt *sp)
1794 {
1795 	vm_object_t object;
1796 	int retries;
1797 
1798 	sx_assert(&swdev_syscall_lock, SA_XLOCKED);
1799 
1800 	retries = 0;
1801 full_rescan:
1802 	mtx_lock(&vm_object_list_mtx);
1803 	TAILQ_FOREACH(object, &vm_object_list, object_list) {
1804 		if (object->type != OBJT_SWAP)
1805 			continue;
1806 		mtx_unlock(&vm_object_list_mtx);
1807 		/* Depends on type-stability. */
1808 		VM_OBJECT_WLOCK(object);
1809 
1810 		/*
1811 		 * Dead objects are eventually terminated on their own.
1812 		 */
1813 		if ((object->flags & OBJ_DEAD) != 0)
1814 			goto next_obj;
1815 
1816 		/*
1817 		 * Sync with fences placed after pctrie
1818 		 * initialization.  We must not access pctrie below
1819 		 * unless we checked that our object is swap and not
1820 		 * dead.
1821 		 */
1822 		atomic_thread_fence_acq();
1823 		if (object->type != OBJT_SWAP)
1824 			goto next_obj;
1825 
1826 		swap_pager_swapoff_object(sp, object);
1827 next_obj:
1828 		VM_OBJECT_WUNLOCK(object);
1829 		mtx_lock(&vm_object_list_mtx);
1830 	}
1831 	mtx_unlock(&vm_object_list_mtx);
1832 
1833 	if (sp->sw_used) {
1834 		/*
1835 		 * Objects may be locked or paging to the device being
1836 		 * removed, so we will miss their pages and need to
1837 		 * make another pass.  We have marked this device as
1838 		 * SW_CLOSING, so the activity should finish soon.
1839 		 */
1840 		retries++;
1841 		if (retries > 100) {
1842 			panic("swapoff: failed to locate %d swap blocks",
1843 			    sp->sw_used);
1844 		}
1845 		pause("swpoff", hz / 20);
1846 		goto full_rescan;
1847 	}
1848 	EVENTHANDLER_INVOKE(swapoff, sp);
1849 }
1850 
1851 /************************************************************************
1852  *				SWAP META DATA 				*
1853  ************************************************************************
1854  *
1855  *	These routines manipulate the swap metadata stored in the
1856  *	OBJT_SWAP object.
1857  *
1858  *	Swap metadata is implemented with a global hash and not directly
1859  *	linked into the object.  Instead the object simply contains
1860  *	appropriate tracking counters.
1861  */
1862 
1863 /*
1864  * SWP_PAGER_SWBLK_EMPTY() - is a range of blocks free?
1865  */
1866 static bool
1867 swp_pager_swblk_empty(struct swblk *sb, int start, int limit)
1868 {
1869 	int i;
1870 
1871 	MPASS(0 <= start && start <= limit && limit <= SWAP_META_PAGES);
1872 	for (i = start; i < limit; i++) {
1873 		if (sb->d[i] != SWAPBLK_NONE)
1874 			return (false);
1875 	}
1876 	return (true);
1877 }
1878 
1879 /*
1880  * SWP_PAGER_META_BUILD() -	add swap block to swap meta data for object
1881  *
1882  *	We first convert the object to a swap object if it is a default
1883  *	object.
1884  *
1885  *	The specified swapblk is added to the object's swap metadata.  If
1886  *	the swapblk is not valid, it is freed instead.  Any previously
1887  *	assigned swapblk is returned.
1888  */
1889 static daddr_t
1890 swp_pager_meta_build(vm_object_t object, vm_pindex_t pindex, daddr_t swapblk)
1891 {
1892 	static volatile int swblk_zone_exhausted, swpctrie_zone_exhausted;
1893 	struct swblk *sb, *sb1;
1894 	vm_pindex_t modpi, rdpi;
1895 	daddr_t prev_swapblk;
1896 	int error, i;
1897 
1898 	VM_OBJECT_ASSERT_WLOCKED(object);
1899 
1900 	/*
1901 	 * Convert default object to swap object if necessary
1902 	 */
1903 	if (object->type != OBJT_SWAP) {
1904 		pctrie_init(&object->un_pager.swp.swp_blks);
1905 
1906 		/*
1907 		 * Ensure that swap_pager_swapoff()'s iteration over
1908 		 * object_list does not see a garbage pctrie.
1909 		 */
1910 		atomic_thread_fence_rel();
1911 
1912 		object->type = OBJT_SWAP;
1913 		object->un_pager.swp.writemappings = 0;
1914 		KASSERT(object->handle == NULL, ("default pager with handle"));
1915 	}
1916 
1917 	rdpi = rounddown(pindex, SWAP_META_PAGES);
1918 	sb = SWAP_PCTRIE_LOOKUP(&object->un_pager.swp.swp_blks, rdpi);
1919 	if (sb == NULL) {
1920 		if (swapblk == SWAPBLK_NONE)
1921 			return (SWAPBLK_NONE);
1922 		for (;;) {
1923 			sb = uma_zalloc(swblk_zone, M_NOWAIT | (curproc ==
1924 			    pageproc ? M_USE_RESERVE : 0));
1925 			if (sb != NULL) {
1926 				sb->p = rdpi;
1927 				for (i = 0; i < SWAP_META_PAGES; i++)
1928 					sb->d[i] = SWAPBLK_NONE;
1929 				if (atomic_cmpset_int(&swblk_zone_exhausted,
1930 				    1, 0))
1931 					printf("swblk zone ok\n");
1932 				break;
1933 			}
1934 			VM_OBJECT_WUNLOCK(object);
1935 			if (uma_zone_exhausted(swblk_zone)) {
1936 				if (atomic_cmpset_int(&swblk_zone_exhausted,
1937 				    0, 1))
1938 					printf("swap blk zone exhausted, "
1939 					    "increase kern.maxswzone\n");
1940 				vm_pageout_oom(VM_OOM_SWAPZ);
1941 				pause("swzonxb", 10);
1942 			} else
1943 				uma_zwait(swblk_zone);
1944 			VM_OBJECT_WLOCK(object);
1945 			sb = SWAP_PCTRIE_LOOKUP(&object->un_pager.swp.swp_blks,
1946 			    rdpi);
1947 			if (sb != NULL)
1948 				/*
1949 				 * Somebody swapped out a nearby page,
1950 				 * allocating swblk at the rdpi index,
1951 				 * while we dropped the object lock.
1952 				 */
1953 				goto allocated;
1954 		}
1955 		for (;;) {
1956 			error = SWAP_PCTRIE_INSERT(
1957 			    &object->un_pager.swp.swp_blks, sb);
1958 			if (error == 0) {
1959 				if (atomic_cmpset_int(&swpctrie_zone_exhausted,
1960 				    1, 0))
1961 					printf("swpctrie zone ok\n");
1962 				break;
1963 			}
1964 			VM_OBJECT_WUNLOCK(object);
1965 			if (uma_zone_exhausted(swpctrie_zone)) {
1966 				if (atomic_cmpset_int(&swpctrie_zone_exhausted,
1967 				    0, 1))
1968 					printf("swap pctrie zone exhausted, "
1969 					    "increase kern.maxswzone\n");
1970 				vm_pageout_oom(VM_OOM_SWAPZ);
1971 				pause("swzonxp", 10);
1972 			} else
1973 				uma_zwait(swpctrie_zone);
1974 			VM_OBJECT_WLOCK(object);
1975 			sb1 = SWAP_PCTRIE_LOOKUP(&object->un_pager.swp.swp_blks,
1976 			    rdpi);
1977 			if (sb1 != NULL) {
1978 				uma_zfree(swblk_zone, sb);
1979 				sb = sb1;
1980 				goto allocated;
1981 			}
1982 		}
1983 	}
1984 allocated:
1985 	MPASS(sb->p == rdpi);
1986 
1987 	modpi = pindex % SWAP_META_PAGES;
1988 	/* Return prior contents of metadata. */
1989 	prev_swapblk = sb->d[modpi];
1990 	/* Enter block into metadata. */
1991 	sb->d[modpi] = swapblk;
1992 
1993 	/*
1994 	 * Free the swblk if we end up with the empty page run.
1995 	 */
1996 	if (swapblk == SWAPBLK_NONE &&
1997 	    swp_pager_swblk_empty(sb, 0, SWAP_META_PAGES)) {
1998 		SWAP_PCTRIE_REMOVE(&object->un_pager.swp.swp_blks, rdpi);
1999 		uma_zfree(swblk_zone, sb);
2000 	}
2001 	return (prev_swapblk);
2002 }
2003 
2004 /*
2005  * SWP_PAGER_META_FREE() - free a range of blocks in the object's swap metadata
2006  *
2007  *	The requested range of blocks is freed, with any associated swap
2008  *	returned to the swap bitmap.
2009  *
2010  *	This routine will free swap metadata structures as they are cleaned
2011  *	out.  This routine does *NOT* operate on swap metadata associated
2012  *	with resident pages.
2013  */
2014 static void
2015 swp_pager_meta_free(vm_object_t object, vm_pindex_t pindex, vm_pindex_t count)
2016 {
2017 	struct swblk *sb;
2018 	daddr_t n_free, s_free;
2019 	vm_pindex_t last;
2020 	int i, limit, start;
2021 
2022 	VM_OBJECT_ASSERT_WLOCKED(object);
2023 	if (object->type != OBJT_SWAP || count == 0)
2024 		return;
2025 
2026 	swp_pager_init_freerange(&s_free, &n_free);
2027 	last = pindex + count;
2028 	for (;;) {
2029 		sb = SWAP_PCTRIE_LOOKUP_GE(&object->un_pager.swp.swp_blks,
2030 		    rounddown(pindex, SWAP_META_PAGES));
2031 		if (sb == NULL || sb->p >= last)
2032 			break;
2033 		start = pindex > sb->p ? pindex - sb->p : 0;
2034 		limit = last - sb->p < SWAP_META_PAGES ? last - sb->p :
2035 		    SWAP_META_PAGES;
2036 		for (i = start; i < limit; i++) {
2037 			if (sb->d[i] == SWAPBLK_NONE)
2038 				continue;
2039 			swp_pager_update_freerange(&s_free, &n_free, sb->d[i]);
2040 			sb->d[i] = SWAPBLK_NONE;
2041 		}
2042 		pindex = sb->p + SWAP_META_PAGES;
2043 		if (swp_pager_swblk_empty(sb, 0, start) &&
2044 		    swp_pager_swblk_empty(sb, limit, SWAP_META_PAGES)) {
2045 			SWAP_PCTRIE_REMOVE(&object->un_pager.swp.swp_blks,
2046 			    sb->p);
2047 			uma_zfree(swblk_zone, sb);
2048 		}
2049 	}
2050 	swp_pager_freeswapspace(s_free, n_free);
2051 }
2052 
2053 /*
2054  * SWP_PAGER_META_FREE_ALL() - destroy all swap metadata associated with object
2055  *
2056  *	This routine locates and destroys all swap metadata associated with
2057  *	an object.
2058  */
2059 static void
2060 swp_pager_meta_free_all(vm_object_t object)
2061 {
2062 	struct swblk *sb;
2063 	daddr_t n_free, s_free;
2064 	vm_pindex_t pindex;
2065 	int i;
2066 
2067 	VM_OBJECT_ASSERT_WLOCKED(object);
2068 	if (object->type != OBJT_SWAP)
2069 		return;
2070 
2071 	swp_pager_init_freerange(&s_free, &n_free);
2072 	for (pindex = 0; (sb = SWAP_PCTRIE_LOOKUP_GE(
2073 	    &object->un_pager.swp.swp_blks, pindex)) != NULL;) {
2074 		pindex = sb->p + SWAP_META_PAGES;
2075 		for (i = 0; i < SWAP_META_PAGES; i++) {
2076 			if (sb->d[i] == SWAPBLK_NONE)
2077 				continue;
2078 			swp_pager_update_freerange(&s_free, &n_free, sb->d[i]);
2079 		}
2080 		SWAP_PCTRIE_REMOVE(&object->un_pager.swp.swp_blks, sb->p);
2081 		uma_zfree(swblk_zone, sb);
2082 	}
2083 	swp_pager_freeswapspace(s_free, n_free);
2084 }
2085 
2086 /*
2087  * SWP_PAGER_METACTL() -  misc control of swap meta data.
2088  *
2089  *	This routine is capable of looking up, or removing swapblk
2090  *	assignments in the swap meta data.  It returns the swapblk being
2091  *	looked-up, popped, or SWAPBLK_NONE if the block was invalid.
2092  *
2093  *	When acting on a busy resident page and paging is in progress, we
2094  *	have to wait until paging is complete but otherwise can act on the
2095  *	busy page.
2096  *
2097  *	SWM_POP		remove from meta data but do not free it
2098  */
2099 static daddr_t
2100 swp_pager_meta_ctl(vm_object_t object, vm_pindex_t pindex, int flags)
2101 {
2102 	struct swblk *sb;
2103 	daddr_t r1;
2104 
2105 	if ((flags & SWM_POP) != 0)
2106 		VM_OBJECT_ASSERT_WLOCKED(object);
2107 	else
2108 		VM_OBJECT_ASSERT_LOCKED(object);
2109 
2110 	/*
2111 	 * The meta data only exists if the object is OBJT_SWAP
2112 	 * and even then might not be allocated yet.
2113 	 */
2114 	if (object->type != OBJT_SWAP)
2115 		return (SWAPBLK_NONE);
2116 
2117 	sb = SWAP_PCTRIE_LOOKUP(&object->un_pager.swp.swp_blks,
2118 	    rounddown(pindex, SWAP_META_PAGES));
2119 	if (sb == NULL)
2120 		return (SWAPBLK_NONE);
2121 	r1 = sb->d[pindex % SWAP_META_PAGES];
2122 	if (r1 == SWAPBLK_NONE)
2123 		return (SWAPBLK_NONE);
2124 	if ((flags & SWM_POP) != 0) {
2125 		sb->d[pindex % SWAP_META_PAGES] = SWAPBLK_NONE;
2126 		if (swp_pager_swblk_empty(sb, 0, SWAP_META_PAGES)) {
2127 			SWAP_PCTRIE_REMOVE(&object->un_pager.swp.swp_blks,
2128 			    rounddown(pindex, SWAP_META_PAGES));
2129 			uma_zfree(swblk_zone, sb);
2130 		}
2131 	}
2132 	return (r1);
2133 }
2134 
2135 /*
2136  * Returns the least page index which is greater than or equal to the
2137  * parameter pindex and for which there is a swap block allocated.
2138  * Returns object's size if the object's type is not swap or if there
2139  * are no allocated swap blocks for the object after the requested
2140  * pindex.
2141  */
2142 vm_pindex_t
2143 swap_pager_find_least(vm_object_t object, vm_pindex_t pindex)
2144 {
2145 	struct swblk *sb;
2146 	int i;
2147 
2148 	VM_OBJECT_ASSERT_LOCKED(object);
2149 	if (object->type != OBJT_SWAP)
2150 		return (object->size);
2151 
2152 	sb = SWAP_PCTRIE_LOOKUP_GE(&object->un_pager.swp.swp_blks,
2153 	    rounddown(pindex, SWAP_META_PAGES));
2154 	if (sb == NULL)
2155 		return (object->size);
2156 	if (sb->p < pindex) {
2157 		for (i = pindex % SWAP_META_PAGES; i < SWAP_META_PAGES; i++) {
2158 			if (sb->d[i] != SWAPBLK_NONE)
2159 				return (sb->p + i);
2160 		}
2161 		sb = SWAP_PCTRIE_LOOKUP_GE(&object->un_pager.swp.swp_blks,
2162 		    roundup(pindex, SWAP_META_PAGES));
2163 		if (sb == NULL)
2164 			return (object->size);
2165 	}
2166 	for (i = 0; i < SWAP_META_PAGES; i++) {
2167 		if (sb->d[i] != SWAPBLK_NONE)
2168 			return (sb->p + i);
2169 	}
2170 
2171 	/*
2172 	 * We get here if a swblk is present in the trie but it
2173 	 * doesn't map any blocks.
2174 	 */
2175 	MPASS(0);
2176 	return (object->size);
2177 }
2178 
2179 /*
2180  * System call swapon(name) enables swapping on device name,
2181  * which must be in the swdevsw.  Return EBUSY
2182  * if already swapping on this device.
2183  */
2184 #ifndef _SYS_SYSPROTO_H_
2185 struct swapon_args {
2186 	char *name;
2187 };
2188 #endif
2189 
2190 /*
2191  * MPSAFE
2192  */
2193 /* ARGSUSED */
2194 int
2195 sys_swapon(struct thread *td, struct swapon_args *uap)
2196 {
2197 	struct vattr attr;
2198 	struct vnode *vp;
2199 	struct nameidata nd;
2200 	int error;
2201 
2202 	error = priv_check(td, PRIV_SWAPON);
2203 	if (error)
2204 		return (error);
2205 
2206 	sx_xlock(&swdev_syscall_lock);
2207 
2208 	/*
2209 	 * Swap metadata may not fit in the KVM if we have physical
2210 	 * memory of >1GB.
2211 	 */
2212 	if (swblk_zone == NULL) {
2213 		error = ENOMEM;
2214 		goto done;
2215 	}
2216 
2217 	NDINIT(&nd, LOOKUP, ISOPEN | FOLLOW | AUDITVNODE1, UIO_USERSPACE,
2218 	    uap->name, td);
2219 	error = namei(&nd);
2220 	if (error)
2221 		goto done;
2222 
2223 	NDFREE(&nd, NDF_ONLY_PNBUF);
2224 	vp = nd.ni_vp;
2225 
2226 	if (vn_isdisk(vp, &error)) {
2227 		error = swapongeom(vp);
2228 	} else if (vp->v_type == VREG &&
2229 	    (vp->v_mount->mnt_vfc->vfc_flags & VFCF_NETWORK) != 0 &&
2230 	    (error = VOP_GETATTR(vp, &attr, td->td_ucred)) == 0) {
2231 		/*
2232 		 * Allow direct swapping to NFS regular files in the same
2233 		 * way that nfs_mountroot() sets up diskless swapping.
2234 		 */
2235 		error = swaponvp(td, vp, attr.va_size / DEV_BSIZE);
2236 	}
2237 
2238 	if (error)
2239 		vrele(vp);
2240 done:
2241 	sx_xunlock(&swdev_syscall_lock);
2242 	return (error);
2243 }
2244 
2245 /*
2246  * Check that the total amount of swap currently configured does not
2247  * exceed half the theoretical maximum.  If it does, print a warning
2248  * message.
2249  */
2250 static void
2251 swapon_check_swzone(void)
2252 {
2253 	unsigned long maxpages, npages;
2254 
2255 	npages = swap_total;
2256 	/* absolute maximum we can handle assuming 100% efficiency */
2257 	maxpages = uma_zone_get_max(swblk_zone) * SWAP_META_PAGES;
2258 
2259 	/* recommend using no more than half that amount */
2260 	if (npages > maxpages / 2) {
2261 		printf("warning: total configured swap (%lu pages) "
2262 		    "exceeds maximum recommended amount (%lu pages).\n",
2263 		    npages, maxpages / 2);
2264 		printf("warning: increase kern.maxswzone "
2265 		    "or reduce amount of swap.\n");
2266 	}
2267 }
2268 
2269 static void
2270 swaponsomething(struct vnode *vp, void *id, u_long nblks,
2271     sw_strategy_t *strategy, sw_close_t *close, dev_t dev, int flags)
2272 {
2273 	struct swdevt *sp, *tsp;
2274 	swblk_t dvbase;
2275 	u_long mblocks;
2276 
2277 	/*
2278 	 * nblks is in DEV_BSIZE'd chunks, convert to PAGE_SIZE'd chunks.
2279 	 * First chop nblks off to page-align it, then convert.
2280 	 *
2281 	 * sw->sw_nblks is in page-sized chunks now too.
2282 	 */
2283 	nblks &= ~(ctodb(1) - 1);
2284 	nblks = dbtoc(nblks);
2285 
2286 	/*
2287 	 * If we go beyond this, we get overflows in the radix
2288 	 * tree bitmap code.
2289 	 */
2290 	mblocks = 0x40000000 / BLIST_META_RADIX;
2291 	if (nblks > mblocks) {
2292 		printf(
2293     "WARNING: reducing swap size to maximum of %luMB per unit\n",
2294 		    mblocks / 1024 / 1024 * PAGE_SIZE);
2295 		nblks = mblocks;
2296 	}
2297 
2298 	sp = malloc(sizeof *sp, M_VMPGDATA, M_WAITOK | M_ZERO);
2299 	sp->sw_vp = vp;
2300 	sp->sw_id = id;
2301 	sp->sw_dev = dev;
2302 	sp->sw_nblks = nblks;
2303 	sp->sw_used = 0;
2304 	sp->sw_strategy = strategy;
2305 	sp->sw_close = close;
2306 	sp->sw_flags = flags;
2307 
2308 	sp->sw_blist = blist_create(nblks, M_WAITOK);
2309 	/*
2310 	 * Do not free the first blocks in order to avoid overwriting
2311 	 * any bsd label at the front of the partition
2312 	 */
2313 	blist_free(sp->sw_blist, howmany(BBSIZE, PAGE_SIZE),
2314 	    nblks - howmany(BBSIZE, PAGE_SIZE));
2315 
2316 	dvbase = 0;
2317 	mtx_lock(&sw_dev_mtx);
2318 	TAILQ_FOREACH(tsp, &swtailq, sw_list) {
2319 		if (tsp->sw_end >= dvbase) {
2320 			/*
2321 			 * We put one uncovered page between the devices
2322 			 * in order to definitively prevent any cross-device
2323 			 * I/O requests
2324 			 */
2325 			dvbase = tsp->sw_end + 1;
2326 		}
2327 	}
2328 	sp->sw_first = dvbase;
2329 	sp->sw_end = dvbase + nblks;
2330 	TAILQ_INSERT_TAIL(&swtailq, sp, sw_list);
2331 	nswapdev++;
2332 	swap_pager_avail += nblks - howmany(BBSIZE, PAGE_SIZE);
2333 	swap_total += nblks;
2334 	swapon_check_swzone();
2335 	swp_sizecheck();
2336 	mtx_unlock(&sw_dev_mtx);
2337 	EVENTHANDLER_INVOKE(swapon, sp);
2338 }
2339 
2340 /*
2341  * SYSCALL: swapoff(devname)
2342  *
2343  * Disable swapping on the given device.
2344  *
2345  * XXX: Badly designed system call: it should use a device index
2346  * rather than filename as specification.  We keep sw_vp around
2347  * only to make this work.
2348  */
2349 #ifndef _SYS_SYSPROTO_H_
2350 struct swapoff_args {
2351 	char *name;
2352 };
2353 #endif
2354 
2355 /*
2356  * MPSAFE
2357  */
2358 /* ARGSUSED */
2359 int
2360 sys_swapoff(struct thread *td, struct swapoff_args *uap)
2361 {
2362 	struct vnode *vp;
2363 	struct nameidata nd;
2364 	struct swdevt *sp;
2365 	int error;
2366 
2367 	error = priv_check(td, PRIV_SWAPOFF);
2368 	if (error)
2369 		return (error);
2370 
2371 	sx_xlock(&swdev_syscall_lock);
2372 
2373 	NDINIT(&nd, LOOKUP, FOLLOW | AUDITVNODE1, UIO_USERSPACE, uap->name,
2374 	    td);
2375 	error = namei(&nd);
2376 	if (error)
2377 		goto done;
2378 	NDFREE(&nd, NDF_ONLY_PNBUF);
2379 	vp = nd.ni_vp;
2380 
2381 	mtx_lock(&sw_dev_mtx);
2382 	TAILQ_FOREACH(sp, &swtailq, sw_list) {
2383 		if (sp->sw_vp == vp)
2384 			break;
2385 	}
2386 	mtx_unlock(&sw_dev_mtx);
2387 	if (sp == NULL) {
2388 		error = EINVAL;
2389 		goto done;
2390 	}
2391 	error = swapoff_one(sp, td->td_ucred);
2392 done:
2393 	sx_xunlock(&swdev_syscall_lock);
2394 	return (error);
2395 }
2396 
2397 static int
2398 swapoff_one(struct swdevt *sp, struct ucred *cred)
2399 {
2400 	u_long nblks;
2401 #ifdef MAC
2402 	int error;
2403 #endif
2404 
2405 	sx_assert(&swdev_syscall_lock, SA_XLOCKED);
2406 #ifdef MAC
2407 	(void) vn_lock(sp->sw_vp, LK_EXCLUSIVE | LK_RETRY);
2408 	error = mac_system_check_swapoff(cred, sp->sw_vp);
2409 	(void) VOP_UNLOCK(sp->sw_vp, 0);
2410 	if (error != 0)
2411 		return (error);
2412 #endif
2413 	nblks = sp->sw_nblks;
2414 
2415 	/*
2416 	 * We can turn off this swap device safely only if the
2417 	 * available virtual memory in the system will fit the amount
2418 	 * of data we will have to page back in, plus an epsilon so
2419 	 * the system doesn't become critically low on swap space.
2420 	 */
2421 	if (vm_free_count() + swap_pager_avail < nblks + nswap_lowat)
2422 		return (ENOMEM);
2423 
2424 	/*
2425 	 * Prevent further allocations on this device.
2426 	 */
2427 	mtx_lock(&sw_dev_mtx);
2428 	sp->sw_flags |= SW_CLOSING;
2429 	swap_pager_avail -= blist_fill(sp->sw_blist, 0, nblks);
2430 	swap_total -= nblks;
2431 	mtx_unlock(&sw_dev_mtx);
2432 
2433 	/*
2434 	 * Page in the contents of the device and close it.
2435 	 */
2436 	swap_pager_swapoff(sp);
2437 
2438 	sp->sw_close(curthread, sp);
2439 	mtx_lock(&sw_dev_mtx);
2440 	sp->sw_id = NULL;
2441 	TAILQ_REMOVE(&swtailq, sp, sw_list);
2442 	nswapdev--;
2443 	if (nswapdev == 0) {
2444 		swap_pager_full = 2;
2445 		swap_pager_almost_full = 1;
2446 	}
2447 	if (swdevhd == sp)
2448 		swdevhd = NULL;
2449 	mtx_unlock(&sw_dev_mtx);
2450 	blist_destroy(sp->sw_blist);
2451 	free(sp, M_VMPGDATA);
2452 	return (0);
2453 }
2454 
2455 void
2456 swapoff_all(void)
2457 {
2458 	struct swdevt *sp, *spt;
2459 	const char *devname;
2460 	int error;
2461 
2462 	sx_xlock(&swdev_syscall_lock);
2463 
2464 	mtx_lock(&sw_dev_mtx);
2465 	TAILQ_FOREACH_SAFE(sp, &swtailq, sw_list, spt) {
2466 		mtx_unlock(&sw_dev_mtx);
2467 		if (vn_isdisk(sp->sw_vp, NULL))
2468 			devname = devtoname(sp->sw_vp->v_rdev);
2469 		else
2470 			devname = "[file]";
2471 		error = swapoff_one(sp, thread0.td_ucred);
2472 		if (error != 0) {
2473 			printf("Cannot remove swap device %s (error=%d), "
2474 			    "skipping.\n", devname, error);
2475 		} else if (bootverbose) {
2476 			printf("Swap device %s removed.\n", devname);
2477 		}
2478 		mtx_lock(&sw_dev_mtx);
2479 	}
2480 	mtx_unlock(&sw_dev_mtx);
2481 
2482 	sx_xunlock(&swdev_syscall_lock);
2483 }
2484 
2485 void
2486 swap_pager_status(int *total, int *used)
2487 {
2488 	struct swdevt *sp;
2489 
2490 	*total = 0;
2491 	*used = 0;
2492 	mtx_lock(&sw_dev_mtx);
2493 	TAILQ_FOREACH(sp, &swtailq, sw_list) {
2494 		*total += sp->sw_nblks;
2495 		*used += sp->sw_used;
2496 	}
2497 	mtx_unlock(&sw_dev_mtx);
2498 }
2499 
2500 int
2501 swap_dev_info(int name, struct xswdev *xs, char *devname, size_t len)
2502 {
2503 	struct swdevt *sp;
2504 	const char *tmp_devname;
2505 	int error, n;
2506 
2507 	n = 0;
2508 	error = ENOENT;
2509 	mtx_lock(&sw_dev_mtx);
2510 	TAILQ_FOREACH(sp, &swtailq, sw_list) {
2511 		if (n != name) {
2512 			n++;
2513 			continue;
2514 		}
2515 		xs->xsw_version = XSWDEV_VERSION;
2516 		xs->xsw_dev = sp->sw_dev;
2517 		xs->xsw_flags = sp->sw_flags;
2518 		xs->xsw_nblks = sp->sw_nblks;
2519 		xs->xsw_used = sp->sw_used;
2520 		if (devname != NULL) {
2521 			if (vn_isdisk(sp->sw_vp, NULL))
2522 				tmp_devname = devtoname(sp->sw_vp->v_rdev);
2523 			else
2524 				tmp_devname = "[file]";
2525 			strncpy(devname, tmp_devname, len);
2526 		}
2527 		error = 0;
2528 		break;
2529 	}
2530 	mtx_unlock(&sw_dev_mtx);
2531 	return (error);
2532 }
2533 
2534 #if defined(COMPAT_FREEBSD11)
2535 #define XSWDEV_VERSION_11	1
2536 struct xswdev11 {
2537 	u_int	xsw_version;
2538 	uint32_t xsw_dev;
2539 	int	xsw_flags;
2540 	int	xsw_nblks;
2541 	int     xsw_used;
2542 };
2543 #endif
2544 
2545 #if defined(__amd64__) && defined(COMPAT_FREEBSD32)
2546 struct xswdev32 {
2547 	u_int	xsw_version;
2548 	u_int	xsw_dev1, xsw_dev2;
2549 	int	xsw_flags;
2550 	int	xsw_nblks;
2551 	int     xsw_used;
2552 };
2553 #endif
2554 
2555 static int
2556 sysctl_vm_swap_info(SYSCTL_HANDLER_ARGS)
2557 {
2558 	struct xswdev xs;
2559 #if defined(__amd64__) && defined(COMPAT_FREEBSD32)
2560 	struct xswdev32 xs32;
2561 #endif
2562 #if defined(COMPAT_FREEBSD11)
2563 	struct xswdev11 xs11;
2564 #endif
2565 	int error;
2566 
2567 	if (arg2 != 1)			/* name length */
2568 		return (EINVAL);
2569 	error = swap_dev_info(*(int *)arg1, &xs, NULL, 0);
2570 	if (error != 0)
2571 		return (error);
2572 #if defined(__amd64__) && defined(COMPAT_FREEBSD32)
2573 	if (req->oldlen == sizeof(xs32)) {
2574 		xs32.xsw_version = XSWDEV_VERSION;
2575 		xs32.xsw_dev1 = xs.xsw_dev;
2576 		xs32.xsw_dev2 = xs.xsw_dev >> 32;
2577 		xs32.xsw_flags = xs.xsw_flags;
2578 		xs32.xsw_nblks = xs.xsw_nblks;
2579 		xs32.xsw_used = xs.xsw_used;
2580 		error = SYSCTL_OUT(req, &xs32, sizeof(xs32));
2581 		return (error);
2582 	}
2583 #endif
2584 #if defined(COMPAT_FREEBSD11)
2585 	if (req->oldlen == sizeof(xs11)) {
2586 		xs11.xsw_version = XSWDEV_VERSION_11;
2587 		xs11.xsw_dev = xs.xsw_dev; /* truncation */
2588 		xs11.xsw_flags = xs.xsw_flags;
2589 		xs11.xsw_nblks = xs.xsw_nblks;
2590 		xs11.xsw_used = xs.xsw_used;
2591 		error = SYSCTL_OUT(req, &xs11, sizeof(xs11));
2592 		return (error);
2593 	}
2594 #endif
2595 	error = SYSCTL_OUT(req, &xs, sizeof(xs));
2596 	return (error);
2597 }
2598 
2599 SYSCTL_INT(_vm, OID_AUTO, nswapdev, CTLFLAG_RD, &nswapdev, 0,
2600     "Number of swap devices");
2601 SYSCTL_NODE(_vm, OID_AUTO, swap_info, CTLFLAG_RD | CTLFLAG_MPSAFE,
2602     sysctl_vm_swap_info,
2603     "Swap statistics by device");
2604 
2605 /*
2606  * Count the approximate swap usage in pages for a vmspace.  The
2607  * shadowed or not yet copied on write swap blocks are not accounted.
2608  * The map must be locked.
2609  */
2610 long
2611 vmspace_swap_count(struct vmspace *vmspace)
2612 {
2613 	vm_map_t map;
2614 	vm_map_entry_t cur;
2615 	vm_object_t object;
2616 	struct swblk *sb;
2617 	vm_pindex_t e, pi;
2618 	long count;
2619 	int i;
2620 
2621 	map = &vmspace->vm_map;
2622 	count = 0;
2623 
2624 	for (cur = map->header.next; cur != &map->header; cur = cur->next) {
2625 		if ((cur->eflags & MAP_ENTRY_IS_SUB_MAP) != 0)
2626 			continue;
2627 		object = cur->object.vm_object;
2628 		if (object == NULL || object->type != OBJT_SWAP)
2629 			continue;
2630 		VM_OBJECT_RLOCK(object);
2631 		if (object->type != OBJT_SWAP)
2632 			goto unlock;
2633 		pi = OFF_TO_IDX(cur->offset);
2634 		e = pi + OFF_TO_IDX(cur->end - cur->start);
2635 		for (;; pi = sb->p + SWAP_META_PAGES) {
2636 			sb = SWAP_PCTRIE_LOOKUP_GE(
2637 			    &object->un_pager.swp.swp_blks, pi);
2638 			if (sb == NULL || sb->p >= e)
2639 				break;
2640 			for (i = 0; i < SWAP_META_PAGES; i++) {
2641 				if (sb->p + i < e &&
2642 				    sb->d[i] != SWAPBLK_NONE)
2643 					count++;
2644 			}
2645 		}
2646 unlock:
2647 		VM_OBJECT_RUNLOCK(object);
2648 	}
2649 	return (count);
2650 }
2651 
2652 /*
2653  * GEOM backend
2654  *
2655  * Swapping onto disk devices.
2656  *
2657  */
2658 
2659 static g_orphan_t swapgeom_orphan;
2660 
2661 static struct g_class g_swap_class = {
2662 	.name = "SWAP",
2663 	.version = G_VERSION,
2664 	.orphan = swapgeom_orphan,
2665 };
2666 
2667 DECLARE_GEOM_CLASS(g_swap_class, g_class);
2668 
2669 
2670 static void
2671 swapgeom_close_ev(void *arg, int flags)
2672 {
2673 	struct g_consumer *cp;
2674 
2675 	cp = arg;
2676 	g_access(cp, -1, -1, 0);
2677 	g_detach(cp);
2678 	g_destroy_consumer(cp);
2679 }
2680 
2681 /*
2682  * Add a reference to the g_consumer for an inflight transaction.
2683  */
2684 static void
2685 swapgeom_acquire(struct g_consumer *cp)
2686 {
2687 
2688 	mtx_assert(&sw_dev_mtx, MA_OWNED);
2689 	cp->index++;
2690 }
2691 
2692 /*
2693  * Remove a reference from the g_consumer.  Post a close event if all
2694  * references go away, since the function might be called from the
2695  * biodone context.
2696  */
2697 static void
2698 swapgeom_release(struct g_consumer *cp, struct swdevt *sp)
2699 {
2700 
2701 	mtx_assert(&sw_dev_mtx, MA_OWNED);
2702 	cp->index--;
2703 	if (cp->index == 0) {
2704 		if (g_post_event(swapgeom_close_ev, cp, M_NOWAIT, NULL) == 0)
2705 			sp->sw_id = NULL;
2706 	}
2707 }
2708 
2709 static void
2710 swapgeom_done(struct bio *bp2)
2711 {
2712 	struct swdevt *sp;
2713 	struct buf *bp;
2714 	struct g_consumer *cp;
2715 
2716 	bp = bp2->bio_caller2;
2717 	cp = bp2->bio_from;
2718 	bp->b_ioflags = bp2->bio_flags;
2719 	if (bp2->bio_error)
2720 		bp->b_ioflags |= BIO_ERROR;
2721 	bp->b_resid = bp->b_bcount - bp2->bio_completed;
2722 	bp->b_error = bp2->bio_error;
2723 	bp->b_caller1 = NULL;
2724 	bufdone(bp);
2725 	sp = bp2->bio_caller1;
2726 	mtx_lock(&sw_dev_mtx);
2727 	swapgeom_release(cp, sp);
2728 	mtx_unlock(&sw_dev_mtx);
2729 	g_destroy_bio(bp2);
2730 }
2731 
2732 static void
2733 swapgeom_strategy(struct buf *bp, struct swdevt *sp)
2734 {
2735 	struct bio *bio;
2736 	struct g_consumer *cp;
2737 
2738 	mtx_lock(&sw_dev_mtx);
2739 	cp = sp->sw_id;
2740 	if (cp == NULL) {
2741 		mtx_unlock(&sw_dev_mtx);
2742 		bp->b_error = ENXIO;
2743 		bp->b_ioflags |= BIO_ERROR;
2744 		bufdone(bp);
2745 		return;
2746 	}
2747 	swapgeom_acquire(cp);
2748 	mtx_unlock(&sw_dev_mtx);
2749 	if (bp->b_iocmd == BIO_WRITE)
2750 		bio = g_new_bio();
2751 	else
2752 		bio = g_alloc_bio();
2753 	if (bio == NULL) {
2754 		mtx_lock(&sw_dev_mtx);
2755 		swapgeom_release(cp, sp);
2756 		mtx_unlock(&sw_dev_mtx);
2757 		bp->b_error = ENOMEM;
2758 		bp->b_ioflags |= BIO_ERROR;
2759 		printf("swap_pager: cannot allocate bio\n");
2760 		bufdone(bp);
2761 		return;
2762 	}
2763 
2764 	bp->b_caller1 = bio;
2765 	bio->bio_caller1 = sp;
2766 	bio->bio_caller2 = bp;
2767 	bio->bio_cmd = bp->b_iocmd;
2768 	bio->bio_offset = (bp->b_blkno - sp->sw_first) * PAGE_SIZE;
2769 	bio->bio_length = bp->b_bcount;
2770 	bio->bio_done = swapgeom_done;
2771 	if (!buf_mapped(bp)) {
2772 		bio->bio_ma = bp->b_pages;
2773 		bio->bio_data = unmapped_buf;
2774 		bio->bio_ma_offset = (vm_offset_t)bp->b_offset & PAGE_MASK;
2775 		bio->bio_ma_n = bp->b_npages;
2776 		bio->bio_flags |= BIO_UNMAPPED;
2777 	} else {
2778 		bio->bio_data = bp->b_data;
2779 		bio->bio_ma = NULL;
2780 	}
2781 	g_io_request(bio, cp);
2782 	return;
2783 }
2784 
2785 static void
2786 swapgeom_orphan(struct g_consumer *cp)
2787 {
2788 	struct swdevt *sp;
2789 	int destroy;
2790 
2791 	mtx_lock(&sw_dev_mtx);
2792 	TAILQ_FOREACH(sp, &swtailq, sw_list) {
2793 		if (sp->sw_id == cp) {
2794 			sp->sw_flags |= SW_CLOSING;
2795 			break;
2796 		}
2797 	}
2798 	/*
2799 	 * Drop reference we were created with. Do directly since we're in a
2800 	 * special context where we don't have to queue the call to
2801 	 * swapgeom_close_ev().
2802 	 */
2803 	cp->index--;
2804 	destroy = ((sp != NULL) && (cp->index == 0));
2805 	if (destroy)
2806 		sp->sw_id = NULL;
2807 	mtx_unlock(&sw_dev_mtx);
2808 	if (destroy)
2809 		swapgeom_close_ev(cp, 0);
2810 }
2811 
2812 static void
2813 swapgeom_close(struct thread *td, struct swdevt *sw)
2814 {
2815 	struct g_consumer *cp;
2816 
2817 	mtx_lock(&sw_dev_mtx);
2818 	cp = sw->sw_id;
2819 	sw->sw_id = NULL;
2820 	mtx_unlock(&sw_dev_mtx);
2821 
2822 	/*
2823 	 * swapgeom_close() may be called from the biodone context,
2824 	 * where we cannot perform topology changes.  Delegate the
2825 	 * work to the events thread.
2826 	 */
2827 	if (cp != NULL)
2828 		g_waitfor_event(swapgeom_close_ev, cp, M_WAITOK, NULL);
2829 }
2830 
2831 static int
2832 swapongeom_locked(struct cdev *dev, struct vnode *vp)
2833 {
2834 	struct g_provider *pp;
2835 	struct g_consumer *cp;
2836 	static struct g_geom *gp;
2837 	struct swdevt *sp;
2838 	u_long nblks;
2839 	int error;
2840 
2841 	pp = g_dev_getprovider(dev);
2842 	if (pp == NULL)
2843 		return (ENODEV);
2844 	mtx_lock(&sw_dev_mtx);
2845 	TAILQ_FOREACH(sp, &swtailq, sw_list) {
2846 		cp = sp->sw_id;
2847 		if (cp != NULL && cp->provider == pp) {
2848 			mtx_unlock(&sw_dev_mtx);
2849 			return (EBUSY);
2850 		}
2851 	}
2852 	mtx_unlock(&sw_dev_mtx);
2853 	if (gp == NULL)
2854 		gp = g_new_geomf(&g_swap_class, "swap");
2855 	cp = g_new_consumer(gp);
2856 	cp->index = 1;	/* Number of active I/Os, plus one for being active. */
2857 	cp->flags |=  G_CF_DIRECT_SEND | G_CF_DIRECT_RECEIVE;
2858 	g_attach(cp, pp);
2859 	/*
2860 	 * XXX: Every time you think you can improve the margin for
2861 	 * footshooting, somebody depends on the ability to do so:
2862 	 * savecore(8) wants to write to our swapdev so we cannot
2863 	 * set an exclusive count :-(
2864 	 */
2865 	error = g_access(cp, 1, 1, 0);
2866 	if (error != 0) {
2867 		g_detach(cp);
2868 		g_destroy_consumer(cp);
2869 		return (error);
2870 	}
2871 	nblks = pp->mediasize / DEV_BSIZE;
2872 	swaponsomething(vp, cp, nblks, swapgeom_strategy,
2873 	    swapgeom_close, dev2udev(dev),
2874 	    (pp->flags & G_PF_ACCEPT_UNMAPPED) != 0 ? SW_UNMAPPED : 0);
2875 	return (0);
2876 }
2877 
2878 static int
2879 swapongeom(struct vnode *vp)
2880 {
2881 	int error;
2882 
2883 	vn_lock(vp, LK_EXCLUSIVE | LK_RETRY);
2884 	if (vp->v_type != VCHR || (vp->v_iflag & VI_DOOMED) != 0) {
2885 		error = ENOENT;
2886 	} else {
2887 		g_topology_lock();
2888 		error = swapongeom_locked(vp->v_rdev, vp);
2889 		g_topology_unlock();
2890 	}
2891 	VOP_UNLOCK(vp, 0);
2892 	return (error);
2893 }
2894 
2895 /*
2896  * VNODE backend
2897  *
2898  * This is used mainly for network filesystem (read: probably only tested
2899  * with NFS) swapfiles.
2900  *
2901  */
2902 
2903 static void
2904 swapdev_strategy(struct buf *bp, struct swdevt *sp)
2905 {
2906 	struct vnode *vp2;
2907 
2908 	bp->b_blkno = ctodb(bp->b_blkno - sp->sw_first);
2909 
2910 	vp2 = sp->sw_id;
2911 	vhold(vp2);
2912 	if (bp->b_iocmd == BIO_WRITE) {
2913 		if (bp->b_bufobj)
2914 			bufobj_wdrop(bp->b_bufobj);
2915 		bufobj_wref(&vp2->v_bufobj);
2916 	}
2917 	if (bp->b_bufobj != &vp2->v_bufobj)
2918 		bp->b_bufobj = &vp2->v_bufobj;
2919 	bp->b_vp = vp2;
2920 	bp->b_iooffset = dbtob(bp->b_blkno);
2921 	bstrategy(bp);
2922 	return;
2923 }
2924 
2925 static void
2926 swapdev_close(struct thread *td, struct swdevt *sp)
2927 {
2928 
2929 	VOP_CLOSE(sp->sw_vp, FREAD | FWRITE, td->td_ucred, td);
2930 	vrele(sp->sw_vp);
2931 }
2932 
2933 
2934 static int
2935 swaponvp(struct thread *td, struct vnode *vp, u_long nblks)
2936 {
2937 	struct swdevt *sp;
2938 	int error;
2939 
2940 	if (nblks == 0)
2941 		return (ENXIO);
2942 	mtx_lock(&sw_dev_mtx);
2943 	TAILQ_FOREACH(sp, &swtailq, sw_list) {
2944 		if (sp->sw_id == vp) {
2945 			mtx_unlock(&sw_dev_mtx);
2946 			return (EBUSY);
2947 		}
2948 	}
2949 	mtx_unlock(&sw_dev_mtx);
2950 
2951 	(void) vn_lock(vp, LK_EXCLUSIVE | LK_RETRY);
2952 #ifdef MAC
2953 	error = mac_system_check_swapon(td->td_ucred, vp);
2954 	if (error == 0)
2955 #endif
2956 		error = VOP_OPEN(vp, FREAD | FWRITE, td->td_ucred, td, NULL);
2957 	(void) VOP_UNLOCK(vp, 0);
2958 	if (error)
2959 		return (error);
2960 
2961 	swaponsomething(vp, vp, nblks, swapdev_strategy, swapdev_close,
2962 	    NODEV, 0);
2963 	return (0);
2964 }
2965 
2966 static int
2967 sysctl_swap_async_max(SYSCTL_HANDLER_ARGS)
2968 {
2969 	int error, new, n;
2970 
2971 	new = nsw_wcount_async_max;
2972 	error = sysctl_handle_int(oidp, &new, 0, req);
2973 	if (error != 0 || req->newptr == NULL)
2974 		return (error);
2975 
2976 	if (new > nswbuf / 2 || new < 1)
2977 		return (EINVAL);
2978 
2979 	mtx_lock(&swbuf_mtx);
2980 	while (nsw_wcount_async_max != new) {
2981 		/*
2982 		 * Adjust difference.  If the current async count is too low,
2983 		 * we will need to sqeeze our update slowly in.  Sleep with a
2984 		 * higher priority than getpbuf() to finish faster.
2985 		 */
2986 		n = new - nsw_wcount_async_max;
2987 		if (nsw_wcount_async + n >= 0) {
2988 			nsw_wcount_async += n;
2989 			nsw_wcount_async_max += n;
2990 			wakeup(&nsw_wcount_async);
2991 		} else {
2992 			nsw_wcount_async_max -= nsw_wcount_async;
2993 			nsw_wcount_async = 0;
2994 			msleep(&nsw_wcount_async, &swbuf_mtx, PSWP,
2995 			    "swpsysctl", 0);
2996 		}
2997 	}
2998 	mtx_unlock(&swbuf_mtx);
2999 
3000 	return (0);
3001 }
3002 
3003 static void
3004 swap_pager_update_writecount(vm_object_t object, vm_offset_t start,
3005     vm_offset_t end)
3006 {
3007 
3008 	VM_OBJECT_WLOCK(object);
3009 	KASSERT((object->flags & OBJ_NOSPLIT) != 0,
3010 	    ("Splittable object with writecount"));
3011 	object->un_pager.swp.writemappings += (vm_ooffset_t)end - start;
3012 	VM_OBJECT_WUNLOCK(object);
3013 }
3014 
3015 static void
3016 swap_pager_release_writecount(vm_object_t object, vm_offset_t start,
3017     vm_offset_t end)
3018 {
3019 
3020 	VM_OBJECT_WLOCK(object);
3021 	KASSERT((object->flags & OBJ_NOSPLIT) != 0,
3022 	    ("Splittable object with writecount"));
3023 	object->un_pager.swp.writemappings -= (vm_ooffset_t)end - start;
3024 	VM_OBJECT_WUNLOCK(object);
3025 }
3026