xref: /freebsd/sys/vm/swap_pager.c (revision 6af83ee0d2941d18880b6aaa2b4facd1d30c6106)
1 /*-
2  * Copyright (c) 1998 Matthew Dillon,
3  * Copyright (c) 1994 John S. Dyson
4  * Copyright (c) 1990 University of Utah.
5  * Copyright (c) 1982, 1986, 1989, 1993
6  *	The Regents of the University of California.  All rights reserved.
7  *
8  * This code is derived from software contributed to Berkeley by
9  * the Systems Programming Group of the University of Utah Computer
10  * Science Department.
11  *
12  * Redistribution and use in source and binary forms, with or without
13  * modification, are permitted provided that the following conditions
14  * are met:
15  * 1. Redistributions of source code must retain the above copyright
16  *    notice, this list of conditions and the following disclaimer.
17  * 2. Redistributions in binary form must reproduce the above copyright
18  *    notice, this list of conditions and the following disclaimer in the
19  *    documentation and/or other materials provided with the distribution.
20  * 3. All advertising materials mentioning features or use of this software
21  *    must display the following acknowledgement:
22  *	This product includes software developed by the University of
23  *	California, Berkeley and its contributors.
24  * 4. Neither the name of the University nor the names of its contributors
25  *    may be used to endorse or promote products derived from this software
26  *    without specific prior written permission.
27  *
28  * THIS SOFTWARE IS PROVIDED BY THE REGENTS AND CONTRIBUTORS ``AS IS'' AND
29  * ANY EXPRESS OR IMPLIED WARRANTIES, INCLUDING, BUT NOT LIMITED TO, THE
30  * IMPLIED WARRANTIES OF MERCHANTABILITY AND FITNESS FOR A PARTICULAR PURPOSE
31  * ARE DISCLAIMED.  IN NO EVENT SHALL THE REGENTS OR CONTRIBUTORS BE LIABLE
32  * FOR ANY DIRECT, INDIRECT, INCIDENTAL, SPECIAL, EXEMPLARY, OR CONSEQUENTIAL
33  * DAMAGES (INCLUDING, BUT NOT LIMITED TO, PROCUREMENT OF SUBSTITUTE GOODS
34  * OR SERVICES; LOSS OF USE, DATA, OR PROFITS; OR BUSINESS INTERRUPTION)
35  * HOWEVER CAUSED AND ON ANY THEORY OF LIABILITY, WHETHER IN CONTRACT, STRICT
36  * LIABILITY, OR TORT (INCLUDING NEGLIGENCE OR OTHERWISE) ARISING IN ANY WAY
37  * OUT OF THE USE OF THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF
38  * SUCH DAMAGE.
39  *
40  *				New Swap System
41  *				Matthew Dillon
42  *
43  * Radix Bitmap 'blists'.
44  *
45  *	- The new swapper uses the new radix bitmap code.  This should scale
46  *	  to arbitrarily small or arbitrarily large swap spaces and an almost
47  *	  arbitrary degree of fragmentation.
48  *
49  * Features:
50  *
51  *	- on the fly reallocation of swap during putpages.  The new system
52  *	  does not try to keep previously allocated swap blocks for dirty
53  *	  pages.
54  *
55  *	- on the fly deallocation of swap
56  *
57  *	- No more garbage collection required.  Unnecessarily allocated swap
58  *	  blocks only exist for dirty vm_page_t's now and these are already
59  *	  cycled (in a high-load system) by the pager.  We also do on-the-fly
60  *	  removal of invalidated swap blocks when a page is destroyed
61  *	  or renamed.
62  *
63  * from: Utah $Hdr: swap_pager.c 1.4 91/04/30$
64  *
65  *	@(#)swap_pager.c	8.9 (Berkeley) 3/21/94
66  *	@(#)vm_swap.c	8.5 (Berkeley) 2/17/94
67  */
68 
69 #include <sys/cdefs.h>
70 __FBSDID("$FreeBSD$");
71 
72 #include "opt_mac.h"
73 #include "opt_swap.h"
74 #include "opt_vm.h"
75 
76 #include <sys/param.h>
77 #include <sys/systm.h>
78 #include <sys/conf.h>
79 #include <sys/kernel.h>
80 #include <sys/proc.h>
81 #include <sys/bio.h>
82 #include <sys/buf.h>
83 #include <sys/disk.h>
84 #include <sys/fcntl.h>
85 #include <sys/mount.h>
86 #include <sys/namei.h>
87 #include <sys/vnode.h>
88 #include <sys/mac.h>
89 #include <sys/malloc.h>
90 #include <sys/sysctl.h>
91 #include <sys/sysproto.h>
92 #include <sys/blist.h>
93 #include <sys/lock.h>
94 #include <sys/sx.h>
95 #include <sys/vmmeter.h>
96 
97 #include <vm/vm.h>
98 #include <vm/pmap.h>
99 #include <vm/vm_map.h>
100 #include <vm/vm_kern.h>
101 #include <vm/vm_object.h>
102 #include <vm/vm_page.h>
103 #include <vm/vm_pager.h>
104 #include <vm/vm_pageout.h>
105 #include <vm/vm_param.h>
106 #include <vm/swap_pager.h>
107 #include <vm/vm_extern.h>
108 #include <vm/uma.h>
109 
110 #include <geom/geom.h>
111 
112 /*
113  * SWB_NPAGES must be a power of 2.  It may be set to 1, 2, 4, 8, or 16
114  * pages per allocation.  We recommend you stick with the default of 8.
115  * The 16-page limit is due to the radix code (kern/subr_blist.c).
116  */
117 #ifndef MAX_PAGEOUT_CLUSTER
118 #define MAX_PAGEOUT_CLUSTER 16
119 #endif
120 
121 #if !defined(SWB_NPAGES)
122 #define SWB_NPAGES	MAX_PAGEOUT_CLUSTER
123 #endif
124 
125 /*
126  * Piecemeal swap metadata structure.  Swap is stored in a radix tree.
127  *
128  * If SWB_NPAGES is 8 and sizeof(char *) == sizeof(daddr_t), our radix
129  * is basically 8.  Assuming PAGE_SIZE == 4096, one tree level represents
130  * 32K worth of data, two levels represent 256K, three levels represent
131  * 2 MBytes.   This is acceptable.
132  *
133  * Overall memory utilization is about the same as the old swap structure.
134  */
135 #define SWCORRECT(n) (sizeof(void *) * (n) / sizeof(daddr_t))
136 #define SWAP_META_PAGES		(SWB_NPAGES * 2)
137 #define SWAP_META_MASK		(SWAP_META_PAGES - 1)
138 
139 typedef	int32_t	swblk_t;	/*
140 				 * swap offset.  This is the type used to
141 				 * address the "virtual swap device" and
142 				 * therefore the maximum swap space is
143 				 * 2^32 pages.
144 				 */
145 
146 struct swdevt;
147 typedef void sw_strategy_t(struct buf *bp, struct swdevt *sw);
148 typedef void sw_close_t(struct thread *td, struct swdevt *sw);
149 
150 /*
151  * Swap device table
152  */
153 struct swdevt {
154 	int	sw_flags;
155 	int	sw_nblks;
156 	int     sw_used;
157 	dev_t	sw_dev;
158 	struct vnode *sw_vp;
159 	void	*sw_id;
160 	swblk_t	sw_first;
161 	swblk_t	sw_end;
162 	struct blist *sw_blist;
163 	TAILQ_ENTRY(swdevt)	sw_list;
164 	sw_strategy_t		*sw_strategy;
165 	sw_close_t		*sw_close;
166 };
167 
168 #define	SW_CLOSING	0x04
169 
170 struct swblock {
171 	struct swblock	*swb_hnext;
172 	vm_object_t	swb_object;
173 	vm_pindex_t	swb_index;
174 	int		swb_count;
175 	daddr_t		swb_pages[SWAP_META_PAGES];
176 };
177 
178 static struct mtx sw_dev_mtx;
179 static TAILQ_HEAD(, swdevt) swtailq = TAILQ_HEAD_INITIALIZER(swtailq);
180 static struct swdevt *swdevhd;	/* Allocate from here next */
181 static int nswapdev;		/* Number of swap devices */
182 int swap_pager_avail;
183 static int swdev_syscall_active = 0; /* serialize swap(on|off) */
184 
185 static void swapdev_strategy(struct buf *, struct swdevt *sw);
186 
187 #define SWM_FREE	0x02	/* free, period			*/
188 #define SWM_POP		0x04	/* pop out			*/
189 
190 int swap_pager_full = 2;	/* swap space exhaustion (task killing) */
191 static int swap_pager_almost_full = 1; /* swap space exhaustion (w/hysteresis)*/
192 static int nsw_rcount;		/* free read buffers			*/
193 static int nsw_wcount_sync;	/* limit write buffers / synchronous	*/
194 static int nsw_wcount_async;	/* limit write buffers / asynchronous	*/
195 static int nsw_wcount_async_max;/* assigned maximum			*/
196 static int nsw_cluster_max;	/* maximum VOP I/O allowed		*/
197 
198 static struct swblock **swhash;
199 static int swhash_mask;
200 static struct mtx swhash_mtx;
201 
202 static int swap_async_max = 4;	/* maximum in-progress async I/O's	*/
203 static struct sx sw_alloc_sx;
204 
205 
206 SYSCTL_INT(_vm, OID_AUTO, swap_async_max,
207         CTLFLAG_RW, &swap_async_max, 0, "Maximum running async swap ops");
208 
209 /*
210  * "named" and "unnamed" anon region objects.  Try to reduce the overhead
211  * of searching a named list by hashing it just a little.
212  */
213 
214 #define NOBJLISTS		8
215 
216 #define NOBJLIST(handle)	\
217 	(&swap_pager_object_list[((int)(intptr_t)handle >> 4) & (NOBJLISTS-1)])
218 
219 static struct mtx sw_alloc_mtx;	/* protect list manipulation */
220 static struct pagerlst	swap_pager_object_list[NOBJLISTS];
221 static uma_zone_t	swap_zone;
222 static struct vm_object	swap_zone_obj;
223 
224 /*
225  * pagerops for OBJT_SWAP - "swap pager".  Some ops are also global procedure
226  * calls hooked from other parts of the VM system and do not appear here.
227  * (see vm/swap_pager.h).
228  */
229 static vm_object_t
230 		swap_pager_alloc(void *handle, vm_ooffset_t size,
231 				      vm_prot_t prot, vm_ooffset_t offset);
232 static void	swap_pager_dealloc(vm_object_t object);
233 static int	swap_pager_getpages(vm_object_t, vm_page_t *, int, int);
234 static void	swap_pager_putpages(vm_object_t, vm_page_t *, int, boolean_t, int *);
235 static boolean_t
236 		swap_pager_haspage(vm_object_t object, vm_pindex_t pindex, int *before, int *after);
237 static void	swap_pager_init(void);
238 static void	swap_pager_unswapped(vm_page_t);
239 static void	swap_pager_swapoff(struct swdevt *sp);
240 
241 struct pagerops swappagerops = {
242 	.pgo_init =	swap_pager_init,	/* early system initialization of pager	*/
243 	.pgo_alloc =	swap_pager_alloc,	/* allocate an OBJT_SWAP object		*/
244 	.pgo_dealloc =	swap_pager_dealloc,	/* deallocate an OBJT_SWAP object	*/
245 	.pgo_getpages =	swap_pager_getpages,	/* pagein				*/
246 	.pgo_putpages =	swap_pager_putpages,	/* pageout				*/
247 	.pgo_haspage =	swap_pager_haspage,	/* get backing store status for page	*/
248 	.pgo_pageunswapped = swap_pager_unswapped,	/* remove swap related to page		*/
249 };
250 
251 /*
252  * dmmax is in page-sized chunks with the new swap system.  It was
253  * dev-bsized chunks in the old.  dmmax is always a power of 2.
254  *
255  * swap_*() routines are externally accessible.  swp_*() routines are
256  * internal.
257  */
258 static int dmmax;
259 static int nswap_lowat = 128;	/* in pages, swap_pager_almost_full warn */
260 static int nswap_hiwat = 512;	/* in pages, swap_pager_almost_full warn */
261 
262 SYSCTL_INT(_vm, OID_AUTO, dmmax,
263 	CTLFLAG_RD, &dmmax, 0, "Maximum size of a swap block");
264 
265 static void	swp_sizecheck(void);
266 static void	swp_pager_async_iodone(struct buf *bp);
267 static int	swapongeom(struct thread *, struct vnode *);
268 static int	swaponvp(struct thread *, struct vnode *, u_long);
269 
270 /*
271  * Swap bitmap functions
272  */
273 static void	swp_pager_freeswapspace(daddr_t blk, int npages);
274 static daddr_t	swp_pager_getswapspace(int npages);
275 
276 /*
277  * Metadata functions
278  */
279 static struct swblock **swp_pager_hash(vm_object_t object, vm_pindex_t index);
280 static void swp_pager_meta_build(vm_object_t, vm_pindex_t, daddr_t);
281 static void swp_pager_meta_free(vm_object_t, vm_pindex_t, daddr_t);
282 static void swp_pager_meta_free_all(vm_object_t);
283 static daddr_t swp_pager_meta_ctl(vm_object_t, vm_pindex_t, int);
284 
285 /*
286  * SWP_SIZECHECK() -	update swap_pager_full indication
287  *
288  *	update the swap_pager_almost_full indication and warn when we are
289  *	about to run out of swap space, using lowat/hiwat hysteresis.
290  *
291  *	Clear swap_pager_full ( task killing ) indication when lowat is met.
292  *
293  *	No restrictions on call
294  *	This routine may not block.
295  *	This routine must be called at splvm()
296  */
297 static void
298 swp_sizecheck(void)
299 {
300 
301 	if (swap_pager_avail < nswap_lowat) {
302 		if (swap_pager_almost_full == 0) {
303 			printf("swap_pager: out of swap space\n");
304 			swap_pager_almost_full = 1;
305 		}
306 	} else {
307 		swap_pager_full = 0;
308 		if (swap_pager_avail > nswap_hiwat)
309 			swap_pager_almost_full = 0;
310 	}
311 }
312 
313 /*
314  * SWP_PAGER_HASH() -	hash swap meta data
315  *
316  *	This is an helper function which hashes the swapblk given
317  *	the object and page index.  It returns a pointer to a pointer
318  *	to the object, or a pointer to a NULL pointer if it could not
319  *	find a swapblk.
320  *
321  *	This routine must be called at splvm().
322  */
323 static struct swblock **
324 swp_pager_hash(vm_object_t object, vm_pindex_t index)
325 {
326 	struct swblock **pswap;
327 	struct swblock *swap;
328 
329 	index &= ~(vm_pindex_t)SWAP_META_MASK;
330 	pswap = &swhash[(index ^ (int)(intptr_t)object) & swhash_mask];
331 	while ((swap = *pswap) != NULL) {
332 		if (swap->swb_object == object &&
333 		    swap->swb_index == index
334 		) {
335 			break;
336 		}
337 		pswap = &swap->swb_hnext;
338 	}
339 	return (pswap);
340 }
341 
342 /*
343  * SWAP_PAGER_INIT() -	initialize the swap pager!
344  *
345  *	Expected to be started from system init.  NOTE:  This code is run
346  *	before much else so be careful what you depend on.  Most of the VM
347  *	system has yet to be initialized at this point.
348  */
349 static void
350 swap_pager_init(void)
351 {
352 	/*
353 	 * Initialize object lists
354 	 */
355 	int i;
356 
357 	for (i = 0; i < NOBJLISTS; ++i)
358 		TAILQ_INIT(&swap_pager_object_list[i]);
359 	mtx_init(&sw_alloc_mtx, "swap_pager list", NULL, MTX_DEF);
360 	mtx_init(&sw_dev_mtx, "swapdev", NULL, MTX_DEF);
361 
362 	/*
363 	 * Device Stripe, in PAGE_SIZE'd blocks
364 	 */
365 	dmmax = SWB_NPAGES * 2;
366 }
367 
368 /*
369  * SWAP_PAGER_SWAP_INIT() - swap pager initialization from pageout process
370  *
371  *	Expected to be started from pageout process once, prior to entering
372  *	its main loop.
373  */
374 void
375 swap_pager_swap_init(void)
376 {
377 	int n, n2;
378 
379 	/*
380 	 * Number of in-transit swap bp operations.  Don't
381 	 * exhaust the pbufs completely.  Make sure we
382 	 * initialize workable values (0 will work for hysteresis
383 	 * but it isn't very efficient).
384 	 *
385 	 * The nsw_cluster_max is constrained by the bp->b_pages[]
386 	 * array (MAXPHYS/PAGE_SIZE) and our locally defined
387 	 * MAX_PAGEOUT_CLUSTER.   Also be aware that swap ops are
388 	 * constrained by the swap device interleave stripe size.
389 	 *
390 	 * Currently we hardwire nsw_wcount_async to 4.  This limit is
391 	 * designed to prevent other I/O from having high latencies due to
392 	 * our pageout I/O.  The value 4 works well for one or two active swap
393 	 * devices but is probably a little low if you have more.  Even so,
394 	 * a higher value would probably generate only a limited improvement
395 	 * with three or four active swap devices since the system does not
396 	 * typically have to pageout at extreme bandwidths.   We will want
397 	 * at least 2 per swap devices, and 4 is a pretty good value if you
398 	 * have one NFS swap device due to the command/ack latency over NFS.
399 	 * So it all works out pretty well.
400 	 */
401 	nsw_cluster_max = min((MAXPHYS/PAGE_SIZE), MAX_PAGEOUT_CLUSTER);
402 
403 	mtx_lock(&pbuf_mtx);
404 	nsw_rcount = (nswbuf + 1) / 2;
405 	nsw_wcount_sync = (nswbuf + 3) / 4;
406 	nsw_wcount_async = 4;
407 	nsw_wcount_async_max = nsw_wcount_async;
408 	mtx_unlock(&pbuf_mtx);
409 
410 	/*
411 	 * Initialize our zone.  Right now I'm just guessing on the number
412 	 * we need based on the number of pages in the system.  Each swblock
413 	 * can hold 16 pages, so this is probably overkill.  This reservation
414 	 * is typically limited to around 32MB by default.
415 	 */
416 	n = cnt.v_page_count / 2;
417 	if (maxswzone && n > maxswzone / sizeof(struct swblock))
418 		n = maxswzone / sizeof(struct swblock);
419 	n2 = n;
420 	swap_zone = uma_zcreate("SWAPMETA", sizeof(struct swblock), NULL, NULL,
421 	    NULL, NULL, UMA_ALIGN_PTR, UMA_ZONE_NOFREE | UMA_ZONE_VM);
422 	do {
423 		if (uma_zone_set_obj(swap_zone, &swap_zone_obj, n))
424 			break;
425 		/*
426 		 * if the allocation failed, try a zone two thirds the
427 		 * size of the previous attempt.
428 		 */
429 		n -= ((n + 2) / 3);
430 	} while (n > 0);
431 	if (swap_zone == NULL)
432 		panic("failed to create swap_zone.");
433 	if (n2 != n)
434 		printf("Swap zone entries reduced from %d to %d.\n", n2, n);
435 	n2 = n;
436 
437 	/*
438 	 * Initialize our meta-data hash table.  The swapper does not need to
439 	 * be quite as efficient as the VM system, so we do not use an
440 	 * oversized hash table.
441 	 *
442 	 * 	n: 		size of hash table, must be power of 2
443 	 *	swhash_mask:	hash table index mask
444 	 */
445 	for (n = 1; n < n2 / 8; n *= 2)
446 		;
447 	swhash = malloc(sizeof(struct swblock *) * n, M_VMPGDATA, M_WAITOK | M_ZERO);
448 	swhash_mask = n - 1;
449 	mtx_init(&swhash_mtx, "swap_pager swhash", NULL, MTX_DEF);
450 }
451 
452 /*
453  * SWAP_PAGER_ALLOC() -	allocate a new OBJT_SWAP VM object and instantiate
454  *			its metadata structures.
455  *
456  *	This routine is called from the mmap and fork code to create a new
457  *	OBJT_SWAP object.  We do this by creating an OBJT_DEFAULT object
458  *	and then converting it with swp_pager_meta_build().
459  *
460  *	This routine may block in vm_object_allocate() and create a named
461  *	object lookup race, so we must interlock.   We must also run at
462  *	splvm() for the object lookup to handle races with interrupts, but
463  *	we do not have to maintain splvm() in between the lookup and the
464  *	add because (I believe) it is not possible to attempt to create
465  *	a new swap object w/handle when a default object with that handle
466  *	already exists.
467  *
468  * MPSAFE
469  */
470 static vm_object_t
471 swap_pager_alloc(void *handle, vm_ooffset_t size, vm_prot_t prot,
472 		 vm_ooffset_t offset)
473 {
474 	vm_object_t object;
475 	vm_pindex_t pindex;
476 
477 	pindex = OFF_TO_IDX(offset + PAGE_MASK + size);
478 
479 	if (handle) {
480 		mtx_lock(&Giant);
481 		/*
482 		 * Reference existing named region or allocate new one.  There
483 		 * should not be a race here against swp_pager_meta_build()
484 		 * as called from vm_page_remove() in regards to the lookup
485 		 * of the handle.
486 		 */
487 		sx_xlock(&sw_alloc_sx);
488 		object = vm_pager_object_lookup(NOBJLIST(handle), handle);
489 
490 		if (object != NULL) {
491 			vm_object_reference(object);
492 		} else {
493 			object = vm_object_allocate(OBJT_DEFAULT, pindex);
494 			object->handle = handle;
495 
496 			VM_OBJECT_LOCK(object);
497 			swp_pager_meta_build(object, 0, SWAPBLK_NONE);
498 			VM_OBJECT_UNLOCK(object);
499 		}
500 		sx_xunlock(&sw_alloc_sx);
501 		mtx_unlock(&Giant);
502 	} else {
503 		object = vm_object_allocate(OBJT_DEFAULT, pindex);
504 
505 		VM_OBJECT_LOCK(object);
506 		swp_pager_meta_build(object, 0, SWAPBLK_NONE);
507 		VM_OBJECT_UNLOCK(object);
508 	}
509 	return (object);
510 }
511 
512 /*
513  * SWAP_PAGER_DEALLOC() -	remove swap metadata from object
514  *
515  *	The swap backing for the object is destroyed.  The code is
516  *	designed such that we can reinstantiate it later, but this
517  *	routine is typically called only when the entire object is
518  *	about to be destroyed.
519  *
520  *	This routine may block, but no longer does.
521  *
522  *	The object must be locked or unreferenceable.
523  */
524 static void
525 swap_pager_dealloc(vm_object_t object)
526 {
527 	int s;
528 
529 	/*
530 	 * Remove from list right away so lookups will fail if we block for
531 	 * pageout completion.
532 	 */
533 	if (object->handle != NULL) {
534 		mtx_lock(&sw_alloc_mtx);
535 		TAILQ_REMOVE(NOBJLIST(object->handle), object, pager_object_list);
536 		mtx_unlock(&sw_alloc_mtx);
537 	}
538 
539 	VM_OBJECT_LOCK_ASSERT(object, MA_OWNED);
540 	vm_object_pip_wait(object, "swpdea");
541 
542 	/*
543 	 * Free all remaining metadata.  We only bother to free it from
544 	 * the swap meta data.  We do not attempt to free swapblk's still
545 	 * associated with vm_page_t's for this object.  We do not care
546 	 * if paging is still in progress on some objects.
547 	 */
548 	s = splvm();
549 	swp_pager_meta_free_all(object);
550 	splx(s);
551 }
552 
553 /************************************************************************
554  *			SWAP PAGER BITMAP ROUTINES			*
555  ************************************************************************/
556 
557 /*
558  * SWP_PAGER_GETSWAPSPACE() -	allocate raw swap space
559  *
560  *	Allocate swap for the requested number of pages.  The starting
561  *	swap block number (a page index) is returned or SWAPBLK_NONE
562  *	if the allocation failed.
563  *
564  *	Also has the side effect of advising that somebody made a mistake
565  *	when they configured swap and didn't configure enough.
566  *
567  *	Must be called at splvm() to avoid races with bitmap frees from
568  *	vm_page_remove() aka swap_pager_page_removed().
569  *
570  *	This routine may not block
571  *	This routine must be called at splvm().
572  *
573  *	We allocate in round-robin fashion from the configured devices.
574  */
575 static daddr_t
576 swp_pager_getswapspace(int npages)
577 {
578 	daddr_t blk;
579 	struct swdevt *sp;
580 	int i;
581 
582 	blk = SWAPBLK_NONE;
583 	mtx_lock(&sw_dev_mtx);
584 	sp = swdevhd;
585 	for (i = 0; i < nswapdev; i++) {
586 		if (sp == NULL)
587 			sp = TAILQ_FIRST(&swtailq);
588 		if (!(sp->sw_flags & SW_CLOSING)) {
589 			blk = blist_alloc(sp->sw_blist, npages);
590 			if (blk != SWAPBLK_NONE) {
591 				blk += sp->sw_first;
592 				sp->sw_used += npages;
593 				swap_pager_avail -= npages;
594 				swp_sizecheck();
595 				swdevhd = TAILQ_NEXT(sp, sw_list);
596 				goto done;
597 			}
598 		}
599 		sp = TAILQ_NEXT(sp, sw_list);
600 	}
601 	if (swap_pager_full != 2) {
602 		printf("swap_pager_getswapspace(%d): failed\n", npages);
603 		swap_pager_full = 2;
604 		swap_pager_almost_full = 1;
605 	}
606 	swdevhd = NULL;
607 done:
608 	mtx_unlock(&sw_dev_mtx);
609 	return (blk);
610 }
611 
612 static int
613 swp_pager_isondev(daddr_t blk, struct swdevt *sp)
614 {
615 
616 	return (blk >= sp->sw_first && blk < sp->sw_end);
617 }
618 
619 static void
620 swp_pager_strategy(struct buf *bp)
621 {
622 	struct swdevt *sp;
623 
624 	mtx_lock(&sw_dev_mtx);
625 	TAILQ_FOREACH(sp, &swtailq, sw_list) {
626 		if (bp->b_blkno >= sp->sw_first && bp->b_blkno < sp->sw_end) {
627 			mtx_unlock(&sw_dev_mtx);
628 			sp->sw_strategy(bp, sp);
629 			return;
630 		}
631 	}
632 	panic("Swapdev not found");
633 }
634 
635 
636 /*
637  * SWP_PAGER_FREESWAPSPACE() -	free raw swap space
638  *
639  *	This routine returns the specified swap blocks back to the bitmap.
640  *
641  *	Note:  This routine may not block (it could in the old swap code),
642  *	and through the use of the new blist routines it does not block.
643  *
644  *	We must be called at splvm() to avoid races with bitmap frees from
645  *	vm_page_remove() aka swap_pager_page_removed().
646  *
647  *	This routine may not block
648  *	This routine must be called at splvm().
649  */
650 static void
651 swp_pager_freeswapspace(daddr_t blk, int npages)
652 {
653 	struct swdevt *sp;
654 
655 	mtx_lock(&sw_dev_mtx);
656 	TAILQ_FOREACH(sp, &swtailq, sw_list) {
657 		if (blk >= sp->sw_first && blk < sp->sw_end) {
658 			sp->sw_used -= npages;
659 			/*
660 			 * If we are attempting to stop swapping on
661 			 * this device, we don't want to mark any
662 			 * blocks free lest they be reused.
663 			 */
664 			if ((sp->sw_flags & SW_CLOSING) == 0) {
665 				blist_free(sp->sw_blist, blk - sp->sw_first,
666 				    npages);
667 				swap_pager_avail += npages;
668 				swp_sizecheck();
669 			}
670 			mtx_unlock(&sw_dev_mtx);
671 			return;
672 		}
673 	}
674 	panic("Swapdev not found");
675 }
676 
677 /*
678  * SWAP_PAGER_FREESPACE() -	frees swap blocks associated with a page
679  *				range within an object.
680  *
681  *	This is a globally accessible routine.
682  *
683  *	This routine removes swapblk assignments from swap metadata.
684  *
685  *	The external callers of this routine typically have already destroyed
686  *	or renamed vm_page_t's associated with this range in the object so
687  *	we should be ok.
688  *
689  *	This routine may be called at any spl.  We up our spl to splvm temporarily
690  *	in order to perform the metadata removal.
691  */
692 void
693 swap_pager_freespace(vm_object_t object, vm_pindex_t start, vm_size_t size)
694 {
695 	int s = splvm();
696 
697 	VM_OBJECT_LOCK_ASSERT(object, MA_OWNED);
698 	swp_pager_meta_free(object, start, size);
699 	splx(s);
700 }
701 
702 /*
703  * SWAP_PAGER_RESERVE() - reserve swap blocks in object
704  *
705  *	Assigns swap blocks to the specified range within the object.  The
706  *	swap blocks are not zerod.  Any previous swap assignment is destroyed.
707  *
708  *	Returns 0 on success, -1 on failure.
709  */
710 int
711 swap_pager_reserve(vm_object_t object, vm_pindex_t start, vm_size_t size)
712 {
713 	int s;
714 	int n = 0;
715 	daddr_t blk = SWAPBLK_NONE;
716 	vm_pindex_t beg = start;	/* save start index */
717 
718 	s = splvm();
719 	VM_OBJECT_LOCK(object);
720 	while (size) {
721 		if (n == 0) {
722 			n = BLIST_MAX_ALLOC;
723 			while ((blk = swp_pager_getswapspace(n)) == SWAPBLK_NONE) {
724 				n >>= 1;
725 				if (n == 0) {
726 					swp_pager_meta_free(object, beg, start - beg);
727 					VM_OBJECT_UNLOCK(object);
728 					splx(s);
729 					return (-1);
730 				}
731 			}
732 		}
733 		swp_pager_meta_build(object, start, blk);
734 		--size;
735 		++start;
736 		++blk;
737 		--n;
738 	}
739 	swp_pager_meta_free(object, start, n);
740 	VM_OBJECT_UNLOCK(object);
741 	splx(s);
742 	return (0);
743 }
744 
745 /*
746  * SWAP_PAGER_COPY() -  copy blocks from source pager to destination pager
747  *			and destroy the source.
748  *
749  *	Copy any valid swapblks from the source to the destination.  In
750  *	cases where both the source and destination have a valid swapblk,
751  *	we keep the destination's.
752  *
753  *	This routine is allowed to block.  It may block allocating metadata
754  *	indirectly through swp_pager_meta_build() or if paging is still in
755  *	progress on the source.
756  *
757  *	This routine can be called at any spl
758  *
759  *	XXX vm_page_collapse() kinda expects us not to block because we
760  *	supposedly do not need to allocate memory, but for the moment we
761  *	*may* have to get a little memory from the zone allocator, but
762  *	it is taken from the interrupt memory.  We should be ok.
763  *
764  *	The source object contains no vm_page_t's (which is just as well)
765  *
766  *	The source object is of type OBJT_SWAP.
767  *
768  *	The source and destination objects must be locked or
769  *	inaccessible (XXX are they ?)
770  */
771 void
772 swap_pager_copy(vm_object_t srcobject, vm_object_t dstobject,
773     vm_pindex_t offset, int destroysource)
774 {
775 	vm_pindex_t i;
776 	int s;
777 
778 	VM_OBJECT_LOCK_ASSERT(srcobject, MA_OWNED);
779 	VM_OBJECT_LOCK_ASSERT(dstobject, MA_OWNED);
780 
781 	s = splvm();
782 	/*
783 	 * If destroysource is set, we remove the source object from the
784 	 * swap_pager internal queue now.
785 	 */
786 	if (destroysource) {
787 		if (srcobject->handle != NULL) {
788 			mtx_lock(&sw_alloc_mtx);
789 			TAILQ_REMOVE(
790 			    NOBJLIST(srcobject->handle),
791 			    srcobject,
792 			    pager_object_list
793 			);
794 			mtx_unlock(&sw_alloc_mtx);
795 		}
796 	}
797 
798 	/*
799 	 * transfer source to destination.
800 	 */
801 	for (i = 0; i < dstobject->size; ++i) {
802 		daddr_t dstaddr;
803 
804 		/*
805 		 * Locate (without changing) the swapblk on the destination,
806 		 * unless it is invalid in which case free it silently, or
807 		 * if the destination is a resident page, in which case the
808 		 * source is thrown away.
809 		 */
810 		dstaddr = swp_pager_meta_ctl(dstobject, i, 0);
811 
812 		if (dstaddr == SWAPBLK_NONE) {
813 			/*
814 			 * Destination has no swapblk and is not resident,
815 			 * copy source.
816 			 */
817 			daddr_t srcaddr;
818 
819 			srcaddr = swp_pager_meta_ctl(
820 			    srcobject,
821 			    i + offset,
822 			    SWM_POP
823 			);
824 
825 			if (srcaddr != SWAPBLK_NONE) {
826 				/*
827 				 * swp_pager_meta_build() can sleep.
828 				 */
829 				vm_object_pip_add(srcobject, 1);
830 				VM_OBJECT_UNLOCK(srcobject);
831 				vm_object_pip_add(dstobject, 1);
832 				swp_pager_meta_build(dstobject, i, srcaddr);
833 				vm_object_pip_wakeup(dstobject);
834 				VM_OBJECT_LOCK(srcobject);
835 				vm_object_pip_wakeup(srcobject);
836 			}
837 		} else {
838 			/*
839 			 * Destination has valid swapblk or it is represented
840 			 * by a resident page.  We destroy the sourceblock.
841 			 */
842 
843 			swp_pager_meta_ctl(srcobject, i + offset, SWM_FREE);
844 		}
845 	}
846 
847 	/*
848 	 * Free left over swap blocks in source.
849 	 *
850 	 * We have to revert the type to OBJT_DEFAULT so we do not accidently
851 	 * double-remove the object from the swap queues.
852 	 */
853 	if (destroysource) {
854 		swp_pager_meta_free_all(srcobject);
855 		/*
856 		 * Reverting the type is not necessary, the caller is going
857 		 * to destroy srcobject directly, but I'm doing it here
858 		 * for consistency since we've removed the object from its
859 		 * queues.
860 		 */
861 		srcobject->type = OBJT_DEFAULT;
862 	}
863 	splx(s);
864 }
865 
866 /*
867  * SWAP_PAGER_HASPAGE() -	determine if we have good backing store for
868  *				the requested page.
869  *
870  *	We determine whether good backing store exists for the requested
871  *	page and return TRUE if it does, FALSE if it doesn't.
872  *
873  *	If TRUE, we also try to determine how much valid, contiguous backing
874  *	store exists before and after the requested page within a reasonable
875  *	distance.  We do not try to restrict it to the swap device stripe
876  *	(that is handled in getpages/putpages).  It probably isn't worth
877  *	doing here.
878  */
879 static boolean_t
880 swap_pager_haspage(vm_object_t object, vm_pindex_t pindex, int *before, int *after)
881 {
882 	daddr_t blk0;
883 	int s;
884 
885 	VM_OBJECT_LOCK_ASSERT(object, MA_OWNED);
886 	/*
887 	 * do we have good backing store at the requested index ?
888 	 */
889 	s = splvm();
890 	blk0 = swp_pager_meta_ctl(object, pindex, 0);
891 
892 	if (blk0 == SWAPBLK_NONE) {
893 		splx(s);
894 		if (before)
895 			*before = 0;
896 		if (after)
897 			*after = 0;
898 		return (FALSE);
899 	}
900 
901 	/*
902 	 * find backwards-looking contiguous good backing store
903 	 */
904 	if (before != NULL) {
905 		int i;
906 
907 		for (i = 1; i < (SWB_NPAGES/2); ++i) {
908 			daddr_t blk;
909 
910 			if (i > pindex)
911 				break;
912 			blk = swp_pager_meta_ctl(object, pindex - i, 0);
913 			if (blk != blk0 - i)
914 				break;
915 		}
916 		*before = (i - 1);
917 	}
918 
919 	/*
920 	 * find forward-looking contiguous good backing store
921 	 */
922 	if (after != NULL) {
923 		int i;
924 
925 		for (i = 1; i < (SWB_NPAGES/2); ++i) {
926 			daddr_t blk;
927 
928 			blk = swp_pager_meta_ctl(object, pindex + i, 0);
929 			if (blk != blk0 + i)
930 				break;
931 		}
932 		*after = (i - 1);
933 	}
934 	splx(s);
935 	return (TRUE);
936 }
937 
938 /*
939  * SWAP_PAGER_PAGE_UNSWAPPED() - remove swap backing store related to page
940  *
941  *	This removes any associated swap backing store, whether valid or
942  *	not, from the page.
943  *
944  *	This routine is typically called when a page is made dirty, at
945  *	which point any associated swap can be freed.  MADV_FREE also
946  *	calls us in a special-case situation
947  *
948  *	NOTE!!!  If the page is clean and the swap was valid, the caller
949  *	should make the page dirty before calling this routine.  This routine
950  *	does NOT change the m->dirty status of the page.  Also: MADV_FREE
951  *	depends on it.
952  *
953  *	This routine may not block
954  *	This routine must be called at splvm()
955  */
956 static void
957 swap_pager_unswapped(vm_page_t m)
958 {
959 
960 	VM_OBJECT_LOCK_ASSERT(m->object, MA_OWNED);
961 	swp_pager_meta_ctl(m->object, m->pindex, SWM_FREE);
962 }
963 
964 /*
965  * SWAP_PAGER_GETPAGES() - bring pages in from swap
966  *
967  *	Attempt to retrieve (m, count) pages from backing store, but make
968  *	sure we retrieve at least m[reqpage].  We try to load in as large
969  *	a chunk surrounding m[reqpage] as is contiguous in swap and which
970  *	belongs to the same object.
971  *
972  *	The code is designed for asynchronous operation and
973  *	immediate-notification of 'reqpage' but tends not to be
974  *	used that way.  Please do not optimize-out this algorithmic
975  *	feature, I intend to improve on it in the future.
976  *
977  *	The parent has a single vm_object_pip_add() reference prior to
978  *	calling us and we should return with the same.
979  *
980  *	The parent has BUSY'd the pages.  We should return with 'm'
981  *	left busy, but the others adjusted.
982  */
983 static int
984 swap_pager_getpages(vm_object_t object, vm_page_t *m, int count, int reqpage)
985 {
986 	struct buf *bp;
987 	vm_page_t mreq;
988 	int s;
989 	int i;
990 	int j;
991 	daddr_t blk;
992 
993 	mreq = m[reqpage];
994 
995 	KASSERT(mreq->object == object,
996 	    ("swap_pager_getpages: object mismatch %p/%p",
997 	    object, mreq->object));
998 
999 	/*
1000 	 * Calculate range to retrieve.  The pages have already been assigned
1001 	 * their swapblks.  We require a *contiguous* range but we know it to
1002 	 * not span devices.   If we do not supply it, bad things
1003 	 * happen.  Note that blk, iblk & jblk can be SWAPBLK_NONE, but the
1004 	 * loops are set up such that the case(s) are handled implicitly.
1005 	 *
1006 	 * The swp_*() calls must be made at splvm().  vm_page_free() does
1007 	 * not need to be, but it will go a little faster if it is.
1008 	 */
1009 	s = splvm();
1010 	blk = swp_pager_meta_ctl(mreq->object, mreq->pindex, 0);
1011 
1012 	for (i = reqpage - 1; i >= 0; --i) {
1013 		daddr_t iblk;
1014 
1015 		iblk = swp_pager_meta_ctl(m[i]->object, m[i]->pindex, 0);
1016 		if (blk != iblk + (reqpage - i))
1017 			break;
1018 	}
1019 	++i;
1020 
1021 	for (j = reqpage + 1; j < count; ++j) {
1022 		daddr_t jblk;
1023 
1024 		jblk = swp_pager_meta_ctl(m[j]->object, m[j]->pindex, 0);
1025 		if (blk != jblk - (j - reqpage))
1026 			break;
1027 	}
1028 
1029 	/*
1030 	 * free pages outside our collection range.   Note: we never free
1031 	 * mreq, it must remain busy throughout.
1032 	 */
1033 	vm_page_lock_queues();
1034 	{
1035 		int k;
1036 
1037 		for (k = 0; k < i; ++k)
1038 			vm_page_free(m[k]);
1039 		for (k = j; k < count; ++k)
1040 			vm_page_free(m[k]);
1041 	}
1042 	vm_page_unlock_queues();
1043 	splx(s);
1044 
1045 
1046 	/*
1047 	 * Return VM_PAGER_FAIL if we have nothing to do.  Return mreq
1048 	 * still busy, but the others unbusied.
1049 	 */
1050 	if (blk == SWAPBLK_NONE)
1051 		return (VM_PAGER_FAIL);
1052 
1053 	/*
1054 	 * Getpbuf() can sleep.
1055 	 */
1056 	VM_OBJECT_UNLOCK(object);
1057 	/*
1058 	 * Get a swap buffer header to perform the IO
1059 	 */
1060 	bp = getpbuf(&nsw_rcount);
1061 	bp->b_flags |= B_PAGING;
1062 
1063 	/*
1064 	 * map our page(s) into kva for input
1065 	 */
1066 	pmap_qenter((vm_offset_t)bp->b_data, m + i, j - i);
1067 
1068 	bp->b_iocmd = BIO_READ;
1069 	bp->b_iodone = swp_pager_async_iodone;
1070 	bp->b_rcred = crhold(thread0.td_ucred);
1071 	bp->b_wcred = crhold(thread0.td_ucred);
1072 	bp->b_blkno = blk - (reqpage - i);
1073 	bp->b_bcount = PAGE_SIZE * (j - i);
1074 	bp->b_bufsize = PAGE_SIZE * (j - i);
1075 	bp->b_pager.pg_reqpage = reqpage - i;
1076 
1077 	VM_OBJECT_LOCK(object);
1078 	vm_page_lock_queues();
1079 	{
1080 		int k;
1081 
1082 		for (k = i; k < j; ++k) {
1083 			bp->b_pages[k - i] = m[k];
1084 			vm_page_flag_set(m[k], PG_SWAPINPROG);
1085 		}
1086 	}
1087 	vm_page_unlock_queues();
1088 	VM_OBJECT_UNLOCK(object);
1089 	bp->b_npages = j - i;
1090 
1091 	cnt.v_swapin++;
1092 	cnt.v_swappgsin += bp->b_npages;
1093 
1094 	/*
1095 	 * We still hold the lock on mreq, and our automatic completion routine
1096 	 * does not remove it.
1097 	 */
1098 	VM_OBJECT_LOCK(mreq->object);
1099 	vm_object_pip_add(mreq->object, bp->b_npages);
1100 	VM_OBJECT_UNLOCK(mreq->object);
1101 
1102 	/*
1103 	 * perform the I/O.  NOTE!!!  bp cannot be considered valid after
1104 	 * this point because we automatically release it on completion.
1105 	 * Instead, we look at the one page we are interested in which we
1106 	 * still hold a lock on even through the I/O completion.
1107 	 *
1108 	 * The other pages in our m[] array are also released on completion,
1109 	 * so we cannot assume they are valid anymore either.
1110 	 *
1111 	 * NOTE: b_blkno is destroyed by the call to swapdev_strategy
1112 	 */
1113 	BUF_KERNPROC(bp);
1114 	swp_pager_strategy(bp);
1115 
1116 	/*
1117 	 * wait for the page we want to complete.  PG_SWAPINPROG is always
1118 	 * cleared on completion.  If an I/O error occurs, SWAPBLK_NONE
1119 	 * is set in the meta-data.
1120 	 */
1121 	s = splvm();
1122 	vm_page_lock_queues();
1123 	while ((mreq->flags & PG_SWAPINPROG) != 0) {
1124 		vm_page_flag_set(mreq, PG_WANTED | PG_REFERENCED);
1125 		cnt.v_intrans++;
1126 		if (msleep(mreq, &vm_page_queue_mtx, PSWP, "swread", hz*20)) {
1127 			printf(
1128 "swap_pager: indefinite wait buffer: bufobj: %p, blkno: %jd, size: %ld\n",
1129 			    bp->b_bufobj, (intmax_t)bp->b_blkno, bp->b_bcount);
1130 		}
1131 	}
1132 	vm_page_unlock_queues();
1133 	splx(s);
1134 
1135 	VM_OBJECT_LOCK(mreq->object);
1136 	/*
1137 	 * mreq is left busied after completion, but all the other pages
1138 	 * are freed.  If we had an unrecoverable read error the page will
1139 	 * not be valid.
1140 	 */
1141 	if (mreq->valid != VM_PAGE_BITS_ALL) {
1142 		return (VM_PAGER_ERROR);
1143 	} else {
1144 		return (VM_PAGER_OK);
1145 	}
1146 
1147 	/*
1148 	 * A final note: in a low swap situation, we cannot deallocate swap
1149 	 * and mark a page dirty here because the caller is likely to mark
1150 	 * the page clean when we return, causing the page to possibly revert
1151 	 * to all-zero's later.
1152 	 */
1153 }
1154 
1155 /*
1156  *	swap_pager_putpages:
1157  *
1158  *	Assign swap (if necessary) and initiate I/O on the specified pages.
1159  *
1160  *	We support both OBJT_DEFAULT and OBJT_SWAP objects.  DEFAULT objects
1161  *	are automatically converted to SWAP objects.
1162  *
1163  *	In a low memory situation we may block in VOP_STRATEGY(), but the new
1164  *	vm_page reservation system coupled with properly written VFS devices
1165  *	should ensure that no low-memory deadlock occurs.  This is an area
1166  *	which needs work.
1167  *
1168  *	The parent has N vm_object_pip_add() references prior to
1169  *	calling us and will remove references for rtvals[] that are
1170  *	not set to VM_PAGER_PEND.  We need to remove the rest on I/O
1171  *	completion.
1172  *
1173  *	The parent has soft-busy'd the pages it passes us and will unbusy
1174  *	those whos rtvals[] entry is not set to VM_PAGER_PEND on return.
1175  *	We need to unbusy the rest on I/O completion.
1176  */
1177 void
1178 swap_pager_putpages(vm_object_t object, vm_page_t *m, int count,
1179     boolean_t sync, int *rtvals)
1180 {
1181 	int i;
1182 	int n = 0;
1183 
1184 	GIANT_REQUIRED;
1185 	if (count && m[0]->object != object) {
1186 		panic("swap_pager_getpages: object mismatch %p/%p",
1187 		    object,
1188 		    m[0]->object
1189 		);
1190 	}
1191 
1192 	/*
1193 	 * Step 1
1194 	 *
1195 	 * Turn object into OBJT_SWAP
1196 	 * check for bogus sysops
1197 	 * force sync if not pageout process
1198 	 */
1199 	if (object->type != OBJT_SWAP)
1200 		swp_pager_meta_build(object, 0, SWAPBLK_NONE);
1201 	VM_OBJECT_UNLOCK(object);
1202 
1203 	if (curproc != pageproc)
1204 		sync = TRUE;
1205 
1206 	/*
1207 	 * Step 2
1208 	 *
1209 	 * Update nsw parameters from swap_async_max sysctl values.
1210 	 * Do not let the sysop crash the machine with bogus numbers.
1211 	 */
1212 	mtx_lock(&pbuf_mtx);
1213 	if (swap_async_max != nsw_wcount_async_max) {
1214 		int n;
1215 		int s;
1216 
1217 		/*
1218 		 * limit range
1219 		 */
1220 		if ((n = swap_async_max) > nswbuf / 2)
1221 			n = nswbuf / 2;
1222 		if (n < 1)
1223 			n = 1;
1224 		swap_async_max = n;
1225 
1226 		/*
1227 		 * Adjust difference ( if possible ).  If the current async
1228 		 * count is too low, we may not be able to make the adjustment
1229 		 * at this time.
1230 		 */
1231 		s = splvm();
1232 		n -= nsw_wcount_async_max;
1233 		if (nsw_wcount_async + n >= 0) {
1234 			nsw_wcount_async += n;
1235 			nsw_wcount_async_max += n;
1236 			wakeup(&nsw_wcount_async);
1237 		}
1238 		splx(s);
1239 	}
1240 	mtx_unlock(&pbuf_mtx);
1241 
1242 	/*
1243 	 * Step 3
1244 	 *
1245 	 * Assign swap blocks and issue I/O.  We reallocate swap on the fly.
1246 	 * The page is left dirty until the pageout operation completes
1247 	 * successfully.
1248 	 */
1249 	for (i = 0; i < count; i += n) {
1250 		int s;
1251 		int j;
1252 		struct buf *bp;
1253 		daddr_t blk;
1254 
1255 		/*
1256 		 * Maximum I/O size is limited by a number of factors.
1257 		 */
1258 		n = min(BLIST_MAX_ALLOC, count - i);
1259 		n = min(n, nsw_cluster_max);
1260 
1261 		s = splvm();
1262 
1263 		/*
1264 		 * Get biggest block of swap we can.  If we fail, fall
1265 		 * back and try to allocate a smaller block.  Don't go
1266 		 * overboard trying to allocate space if it would overly
1267 		 * fragment swap.
1268 		 */
1269 		while (
1270 		    (blk = swp_pager_getswapspace(n)) == SWAPBLK_NONE &&
1271 		    n > 4
1272 		) {
1273 			n >>= 1;
1274 		}
1275 		if (blk == SWAPBLK_NONE) {
1276 			for (j = 0; j < n; ++j)
1277 				rtvals[i+j] = VM_PAGER_FAIL;
1278 			splx(s);
1279 			continue;
1280 		}
1281 
1282 		/*
1283 		 * All I/O parameters have been satisfied, build the I/O
1284 		 * request and assign the swap space.
1285 		 */
1286 		if (sync == TRUE) {
1287 			bp = getpbuf(&nsw_wcount_sync);
1288 		} else {
1289 			bp = getpbuf(&nsw_wcount_async);
1290 			bp->b_flags = B_ASYNC;
1291 		}
1292 		bp->b_flags |= B_PAGING;
1293 		bp->b_iocmd = BIO_WRITE;
1294 
1295 		pmap_qenter((vm_offset_t)bp->b_data, &m[i], n);
1296 
1297 		bp->b_rcred = crhold(thread0.td_ucred);
1298 		bp->b_wcred = crhold(thread0.td_ucred);
1299 		bp->b_bcount = PAGE_SIZE * n;
1300 		bp->b_bufsize = PAGE_SIZE * n;
1301 		bp->b_blkno = blk;
1302 
1303 		VM_OBJECT_LOCK(object);
1304 		for (j = 0; j < n; ++j) {
1305 			vm_page_t mreq = m[i+j];
1306 
1307 			swp_pager_meta_build(
1308 			    mreq->object,
1309 			    mreq->pindex,
1310 			    blk + j
1311 			);
1312 			vm_page_dirty(mreq);
1313 			rtvals[i+j] = VM_PAGER_OK;
1314 
1315 			vm_page_lock_queues();
1316 			vm_page_flag_set(mreq, PG_SWAPINPROG);
1317 			vm_page_unlock_queues();
1318 			bp->b_pages[j] = mreq;
1319 		}
1320 		VM_OBJECT_UNLOCK(object);
1321 		bp->b_npages = n;
1322 		/*
1323 		 * Must set dirty range for NFS to work.
1324 		 */
1325 		bp->b_dirtyoff = 0;
1326 		bp->b_dirtyend = bp->b_bcount;
1327 
1328 		cnt.v_swapout++;
1329 		cnt.v_swappgsout += bp->b_npages;
1330 
1331 		splx(s);
1332 
1333 		/*
1334 		 * asynchronous
1335 		 *
1336 		 * NOTE: b_blkno is destroyed by the call to swapdev_strategy
1337 		 */
1338 		if (sync == FALSE) {
1339 			bp->b_iodone = swp_pager_async_iodone;
1340 			BUF_KERNPROC(bp);
1341 			swp_pager_strategy(bp);
1342 
1343 			for (j = 0; j < n; ++j)
1344 				rtvals[i+j] = VM_PAGER_PEND;
1345 			/* restart outter loop */
1346 			continue;
1347 		}
1348 
1349 		/*
1350 		 * synchronous
1351 		 *
1352 		 * NOTE: b_blkno is destroyed by the call to swapdev_strategy
1353 		 */
1354 		bp->b_iodone = bdone;
1355 		swp_pager_strategy(bp);
1356 
1357 		/*
1358 		 * Wait for the sync I/O to complete, then update rtvals.
1359 		 * We just set the rtvals[] to VM_PAGER_PEND so we can call
1360 		 * our async completion routine at the end, thus avoiding a
1361 		 * double-free.
1362 		 */
1363 		s = splbio();
1364 		bwait(bp, PVM, "swwrt");
1365 		for (j = 0; j < n; ++j)
1366 			rtvals[i+j] = VM_PAGER_PEND;
1367 		/*
1368 		 * Now that we are through with the bp, we can call the
1369 		 * normal async completion, which frees everything up.
1370 		 */
1371 		swp_pager_async_iodone(bp);
1372 		splx(s);
1373 	}
1374 	VM_OBJECT_LOCK(object);
1375 }
1376 
1377 /*
1378  *	swp_pager_async_iodone:
1379  *
1380  *	Completion routine for asynchronous reads and writes from/to swap.
1381  *	Also called manually by synchronous code to finish up a bp.
1382  *
1383  *	For READ operations, the pages are PG_BUSY'd.  For WRITE operations,
1384  *	the pages are vm_page_t->busy'd.  For READ operations, we PG_BUSY
1385  *	unbusy all pages except the 'main' request page.  For WRITE
1386  *	operations, we vm_page_t->busy'd unbusy all pages ( we can do this
1387  *	because we marked them all VM_PAGER_PEND on return from putpages ).
1388  *
1389  *	This routine may not block.
1390  *	This routine is called at splbio() or better
1391  *
1392  *	We up ourselves to splvm() as required for various vm_page related
1393  *	calls.
1394  */
1395 static void
1396 swp_pager_async_iodone(struct buf *bp)
1397 {
1398 	int s;
1399 	int i;
1400 	vm_object_t object = NULL;
1401 
1402 	bp->b_flags |= B_DONE;
1403 
1404 	/*
1405 	 * report error
1406 	 */
1407 	if (bp->b_ioflags & BIO_ERROR) {
1408 		printf(
1409 		    "swap_pager: I/O error - %s failed; blkno %ld,"
1410 			"size %ld, error %d\n",
1411 		    ((bp->b_iocmd == BIO_READ) ? "pagein" : "pageout"),
1412 		    (long)bp->b_blkno,
1413 		    (long)bp->b_bcount,
1414 		    bp->b_error
1415 		);
1416 	}
1417 
1418 	/*
1419 	 * set object, raise to splvm().
1420 	 */
1421 	s = splvm();
1422 
1423 	/*
1424 	 * remove the mapping for kernel virtual
1425 	 */
1426 	pmap_qremove((vm_offset_t)bp->b_data, bp->b_npages);
1427 
1428 	if (bp->b_npages) {
1429 		object = bp->b_pages[0]->object;
1430 		VM_OBJECT_LOCK(object);
1431 	}
1432 	vm_page_lock_queues();
1433 	/*
1434 	 * cleanup pages.  If an error occurs writing to swap, we are in
1435 	 * very serious trouble.  If it happens to be a disk error, though,
1436 	 * we may be able to recover by reassigning the swap later on.  So
1437 	 * in this case we remove the m->swapblk assignment for the page
1438 	 * but do not free it in the rlist.  The errornous block(s) are thus
1439 	 * never reallocated as swap.  Redirty the page and continue.
1440 	 */
1441 	for (i = 0; i < bp->b_npages; ++i) {
1442 		vm_page_t m = bp->b_pages[i];
1443 
1444 		vm_page_flag_clear(m, PG_SWAPINPROG);
1445 
1446 		if (bp->b_ioflags & BIO_ERROR) {
1447 			/*
1448 			 * If an error occurs I'd love to throw the swapblk
1449 			 * away without freeing it back to swapspace, so it
1450 			 * can never be used again.  But I can't from an
1451 			 * interrupt.
1452 			 */
1453 			if (bp->b_iocmd == BIO_READ) {
1454 				/*
1455 				 * When reading, reqpage needs to stay
1456 				 * locked for the parent, but all other
1457 				 * pages can be freed.  We still want to
1458 				 * wakeup the parent waiting on the page,
1459 				 * though.  ( also: pg_reqpage can be -1 and
1460 				 * not match anything ).
1461 				 *
1462 				 * We have to wake specifically requested pages
1463 				 * up too because we cleared PG_SWAPINPROG and
1464 				 * someone may be waiting for that.
1465 				 *
1466 				 * NOTE: for reads, m->dirty will probably
1467 				 * be overridden by the original caller of
1468 				 * getpages so don't play cute tricks here.
1469 				 *
1470 				 * XXX IT IS NOT LEGAL TO FREE THE PAGE HERE
1471 				 * AS THIS MESSES WITH object->memq, and it is
1472 				 * not legal to mess with object->memq from an
1473 				 * interrupt.
1474 				 */
1475 				m->valid = 0;
1476 				if (i != bp->b_pager.pg_reqpage)
1477 					vm_page_free(m);
1478 				else
1479 					vm_page_flash(m);
1480 				/*
1481 				 * If i == bp->b_pager.pg_reqpage, do not wake
1482 				 * the page up.  The caller needs to.
1483 				 */
1484 			} else {
1485 				/*
1486 				 * If a write error occurs, reactivate page
1487 				 * so it doesn't clog the inactive list,
1488 				 * then finish the I/O.
1489 				 */
1490 				vm_page_dirty(m);
1491 				vm_page_activate(m);
1492 				vm_page_io_finish(m);
1493 			}
1494 		} else if (bp->b_iocmd == BIO_READ) {
1495 			/*
1496 			 * For read success, clear dirty bits.  Nobody should
1497 			 * have this page mapped but don't take any chances,
1498 			 * make sure the pmap modify bits are also cleared.
1499 			 *
1500 			 * NOTE: for reads, m->dirty will probably be
1501 			 * overridden by the original caller of getpages so
1502 			 * we cannot set them in order to free the underlying
1503 			 * swap in a low-swap situation.  I don't think we'd
1504 			 * want to do that anyway, but it was an optimization
1505 			 * that existed in the old swapper for a time before
1506 			 * it got ripped out due to precisely this problem.
1507 			 *
1508 			 * If not the requested page then deactivate it.
1509 			 *
1510 			 * Note that the requested page, reqpage, is left
1511 			 * busied, but we still have to wake it up.  The
1512 			 * other pages are released (unbusied) by
1513 			 * vm_page_wakeup().  We do not set reqpage's
1514 			 * valid bits here, it is up to the caller.
1515 			 */
1516 			pmap_clear_modify(m);
1517 			m->valid = VM_PAGE_BITS_ALL;
1518 			vm_page_undirty(m);
1519 
1520 			/*
1521 			 * We have to wake specifically requested pages
1522 			 * up too because we cleared PG_SWAPINPROG and
1523 			 * could be waiting for it in getpages.  However,
1524 			 * be sure to not unbusy getpages specifically
1525 			 * requested page - getpages expects it to be
1526 			 * left busy.
1527 			 */
1528 			if (i != bp->b_pager.pg_reqpage) {
1529 				vm_page_deactivate(m);
1530 				vm_page_wakeup(m);
1531 			} else {
1532 				vm_page_flash(m);
1533 			}
1534 		} else {
1535 			/*
1536 			 * For write success, clear the modify and dirty
1537 			 * status, then finish the I/O ( which decrements the
1538 			 * busy count and possibly wakes waiter's up ).
1539 			 */
1540 			pmap_clear_modify(m);
1541 			vm_page_undirty(m);
1542 			vm_page_io_finish(m);
1543 			if (vm_page_count_severe())
1544 				vm_page_try_to_cache(m);
1545 		}
1546 	}
1547 	vm_page_unlock_queues();
1548 
1549 	/*
1550 	 * adjust pip.  NOTE: the original parent may still have its own
1551 	 * pip refs on the object.
1552 	 */
1553 	if (object != NULL) {
1554 		vm_object_pip_wakeupn(object, bp->b_npages);
1555 		VM_OBJECT_UNLOCK(object);
1556 	}
1557 
1558 	/*
1559 	 * release the physical I/O buffer
1560 	 */
1561 	relpbuf(
1562 	    bp,
1563 	    ((bp->b_iocmd == BIO_READ) ? &nsw_rcount :
1564 		((bp->b_flags & B_ASYNC) ?
1565 		    &nsw_wcount_async :
1566 		    &nsw_wcount_sync
1567 		)
1568 	    )
1569 	);
1570 	splx(s);
1571 }
1572 
1573 /*
1574  *	swap_pager_isswapped:
1575  *
1576  *	Return 1 if at least one page in the given object is paged
1577  *	out to the given swap device.
1578  *
1579  *	This routine may not block.
1580  */
1581 int
1582 swap_pager_isswapped(vm_object_t object, struct swdevt *sp)
1583 {
1584 	daddr_t index = 0;
1585 	int bcount;
1586 	int i;
1587 
1588 	VM_OBJECT_LOCK_ASSERT(object, MA_OWNED);
1589 	if (object->type != OBJT_SWAP)
1590 		return (0);
1591 
1592 	mtx_lock(&swhash_mtx);
1593 	for (bcount = 0; bcount < object->un_pager.swp.swp_bcount; bcount++) {
1594 		struct swblock *swap;
1595 
1596 		if ((swap = *swp_pager_hash(object, index)) != NULL) {
1597 			for (i = 0; i < SWAP_META_PAGES; ++i) {
1598 				if (swp_pager_isondev(swap->swb_pages[i], sp)) {
1599 					mtx_unlock(&swhash_mtx);
1600 					return (1);
1601 				}
1602 			}
1603 		}
1604 		index += SWAP_META_PAGES;
1605 		if (index > 0x20000000)
1606 			panic("swap_pager_isswapped: failed to locate all swap meta blocks");
1607 	}
1608 	mtx_unlock(&swhash_mtx);
1609 	return (0);
1610 }
1611 
1612 /*
1613  * SWP_PAGER_FORCE_PAGEIN() - force a swap block to be paged in
1614  *
1615  *	This routine dissociates the page at the given index within a
1616  *	swap block from its backing store, paging it in if necessary.
1617  *	If the page is paged in, it is placed in the inactive queue,
1618  *	since it had its backing store ripped out from under it.
1619  *	We also attempt to swap in all other pages in the swap block,
1620  *	we only guarantee that the one at the specified index is
1621  *	paged in.
1622  *
1623  *	XXX - The code to page the whole block in doesn't work, so we
1624  *	      revert to the one-by-one behavior for now.  Sigh.
1625  */
1626 static __inline void
1627 swp_pager_force_pagein(vm_object_t object, vm_pindex_t pindex)
1628 {
1629 	vm_page_t m;
1630 
1631 	vm_object_pip_add(object, 1);
1632 	m = vm_page_grab(object, pindex, VM_ALLOC_NORMAL|VM_ALLOC_RETRY);
1633 	if (m->valid == VM_PAGE_BITS_ALL) {
1634 		vm_object_pip_subtract(object, 1);
1635 		vm_page_lock_queues();
1636 		vm_page_activate(m);
1637 		vm_page_dirty(m);
1638 		vm_page_wakeup(m);
1639 		vm_page_unlock_queues();
1640 		vm_pager_page_unswapped(m);
1641 		return;
1642 	}
1643 
1644 	if (swap_pager_getpages(object, &m, 1, 0) != VM_PAGER_OK)
1645 		panic("swap_pager_force_pagein: read from swap failed");/*XXX*/
1646 	vm_object_pip_subtract(object, 1);
1647 	vm_page_lock_queues();
1648 	vm_page_dirty(m);
1649 	vm_page_dontneed(m);
1650 	vm_page_wakeup(m);
1651 	vm_page_unlock_queues();
1652 	vm_pager_page_unswapped(m);
1653 }
1654 
1655 /*
1656  *	swap_pager_swapoff:
1657  *
1658  *	Page in all of the pages that have been paged out to the
1659  *	given device.  The corresponding blocks in the bitmap must be
1660  *	marked as allocated and the device must be flagged SW_CLOSING.
1661  *	There may be no processes swapped out to the device.
1662  *
1663  *	This routine may block.
1664  */
1665 static void
1666 swap_pager_swapoff(struct swdevt *sp)
1667 {
1668 	struct swblock *swap;
1669 	int i, j, retries;
1670 
1671 	GIANT_REQUIRED;
1672 
1673 	retries = 0;
1674 full_rescan:
1675 	mtx_lock(&swhash_mtx);
1676 	for (i = 0; i <= swhash_mask; i++) { /* '<=' is correct here */
1677 restart:
1678 		for (swap = swhash[i]; swap != NULL; swap = swap->swb_hnext) {
1679 			vm_object_t object = swap->swb_object;
1680 			vm_pindex_t pindex = swap->swb_index;
1681                         for (j = 0; j < SWAP_META_PAGES; ++j) {
1682                                 if (swp_pager_isondev(swap->swb_pages[j], sp)) {
1683 					/* avoid deadlock */
1684 					if (!VM_OBJECT_TRYLOCK(object)) {
1685 						break;
1686 					} else {
1687 						mtx_unlock(&swhash_mtx);
1688 						swp_pager_force_pagein(object,
1689 						    pindex + j);
1690 						VM_OBJECT_UNLOCK(object);
1691 						mtx_lock(&swhash_mtx);
1692 						goto restart;
1693 					}
1694 				}
1695                         }
1696 		}
1697 	}
1698 	mtx_unlock(&swhash_mtx);
1699 	if (sp->sw_used) {
1700 		int dummy;
1701 		/*
1702 		 * Objects may be locked or paging to the device being
1703 		 * removed, so we will miss their pages and need to
1704 		 * make another pass.  We have marked this device as
1705 		 * SW_CLOSING, so the activity should finish soon.
1706 		 */
1707 		retries++;
1708 		if (retries > 100) {
1709 			panic("swapoff: failed to locate %d swap blocks",
1710 			    sp->sw_used);
1711 		}
1712 		tsleep(&dummy, PVM, "swpoff", hz / 20);
1713 		goto full_rescan;
1714 	}
1715 }
1716 
1717 /************************************************************************
1718  *				SWAP META DATA 				*
1719  ************************************************************************
1720  *
1721  *	These routines manipulate the swap metadata stored in the
1722  *	OBJT_SWAP object.  All swp_*() routines must be called at
1723  *	splvm() because swap can be freed up by the low level vm_page
1724  *	code which might be called from interrupts beyond what splbio() covers.
1725  *
1726  *	Swap metadata is implemented with a global hash and not directly
1727  *	linked into the object.  Instead the object simply contains
1728  *	appropriate tracking counters.
1729  */
1730 
1731 /*
1732  * SWP_PAGER_META_BUILD() -	add swap block to swap meta data for object
1733  *
1734  *	We first convert the object to a swap object if it is a default
1735  *	object.
1736  *
1737  *	The specified swapblk is added to the object's swap metadata.  If
1738  *	the swapblk is not valid, it is freed instead.  Any previously
1739  *	assigned swapblk is freed.
1740  *
1741  *	This routine must be called at splvm(), except when used to convert
1742  *	an OBJT_DEFAULT object into an OBJT_SWAP object.
1743  */
1744 static void
1745 swp_pager_meta_build(vm_object_t object, vm_pindex_t pindex, daddr_t swapblk)
1746 {
1747 	struct swblock *swap;
1748 	struct swblock **pswap;
1749 	int idx;
1750 
1751 	VM_OBJECT_LOCK_ASSERT(object, MA_OWNED);
1752 	/*
1753 	 * Convert default object to swap object if necessary
1754 	 */
1755 	if (object->type != OBJT_SWAP) {
1756 		object->type = OBJT_SWAP;
1757 		object->un_pager.swp.swp_bcount = 0;
1758 
1759 		if (object->handle != NULL) {
1760 			mtx_lock(&sw_alloc_mtx);
1761 			TAILQ_INSERT_TAIL(
1762 			    NOBJLIST(object->handle),
1763 			    object,
1764 			    pager_object_list
1765 			);
1766 			mtx_unlock(&sw_alloc_mtx);
1767 		}
1768 	}
1769 
1770 	/*
1771 	 * Locate hash entry.  If not found create, but if we aren't adding
1772 	 * anything just return.  If we run out of space in the map we wait
1773 	 * and, since the hash table may have changed, retry.
1774 	 */
1775 retry:
1776 	mtx_lock(&swhash_mtx);
1777 	pswap = swp_pager_hash(object, pindex);
1778 
1779 	if ((swap = *pswap) == NULL) {
1780 		int i;
1781 
1782 		if (swapblk == SWAPBLK_NONE)
1783 			goto done;
1784 
1785 		swap = *pswap = uma_zalloc(swap_zone, M_NOWAIT);
1786 		if (swap == NULL) {
1787 			mtx_unlock(&swhash_mtx);
1788 			VM_OBJECT_UNLOCK(object);
1789 			VM_WAIT;
1790 			VM_OBJECT_LOCK(object);
1791 			goto retry;
1792 		}
1793 
1794 		swap->swb_hnext = NULL;
1795 		swap->swb_object = object;
1796 		swap->swb_index = pindex & ~(vm_pindex_t)SWAP_META_MASK;
1797 		swap->swb_count = 0;
1798 
1799 		++object->un_pager.swp.swp_bcount;
1800 
1801 		for (i = 0; i < SWAP_META_PAGES; ++i)
1802 			swap->swb_pages[i] = SWAPBLK_NONE;
1803 	}
1804 
1805 	/*
1806 	 * Delete prior contents of metadata
1807 	 */
1808 	idx = pindex & SWAP_META_MASK;
1809 
1810 	if (swap->swb_pages[idx] != SWAPBLK_NONE) {
1811 		swp_pager_freeswapspace(swap->swb_pages[idx], 1);
1812 		--swap->swb_count;
1813 	}
1814 
1815 	/*
1816 	 * Enter block into metadata
1817 	 */
1818 	swap->swb_pages[idx] = swapblk;
1819 	if (swapblk != SWAPBLK_NONE)
1820 		++swap->swb_count;
1821 done:
1822 	mtx_unlock(&swhash_mtx);
1823 }
1824 
1825 /*
1826  * SWP_PAGER_META_FREE() - free a range of blocks in the object's swap metadata
1827  *
1828  *	The requested range of blocks is freed, with any associated swap
1829  *	returned to the swap bitmap.
1830  *
1831  *	This routine will free swap metadata structures as they are cleaned
1832  *	out.  This routine does *NOT* operate on swap metadata associated
1833  *	with resident pages.
1834  *
1835  *	This routine must be called at splvm()
1836  */
1837 static void
1838 swp_pager_meta_free(vm_object_t object, vm_pindex_t index, daddr_t count)
1839 {
1840 
1841 	VM_OBJECT_LOCK_ASSERT(object, MA_OWNED);
1842 	if (object->type != OBJT_SWAP)
1843 		return;
1844 
1845 	while (count > 0) {
1846 		struct swblock **pswap;
1847 		struct swblock *swap;
1848 
1849 		mtx_lock(&swhash_mtx);
1850 		pswap = swp_pager_hash(object, index);
1851 
1852 		if ((swap = *pswap) != NULL) {
1853 			daddr_t v = swap->swb_pages[index & SWAP_META_MASK];
1854 
1855 			if (v != SWAPBLK_NONE) {
1856 				swp_pager_freeswapspace(v, 1);
1857 				swap->swb_pages[index & SWAP_META_MASK] =
1858 					SWAPBLK_NONE;
1859 				if (--swap->swb_count == 0) {
1860 					*pswap = swap->swb_hnext;
1861 					uma_zfree(swap_zone, swap);
1862 					--object->un_pager.swp.swp_bcount;
1863 				}
1864 			}
1865 			--count;
1866 			++index;
1867 		} else {
1868 			int n = SWAP_META_PAGES - (index & SWAP_META_MASK);
1869 			count -= n;
1870 			index += n;
1871 		}
1872 		mtx_unlock(&swhash_mtx);
1873 	}
1874 }
1875 
1876 /*
1877  * SWP_PAGER_META_FREE_ALL() - destroy all swap metadata associated with object
1878  *
1879  *	This routine locates and destroys all swap metadata associated with
1880  *	an object.
1881  *
1882  *	This routine must be called at splvm()
1883  */
1884 static void
1885 swp_pager_meta_free_all(vm_object_t object)
1886 {
1887 	daddr_t index = 0;
1888 
1889 	VM_OBJECT_LOCK_ASSERT(object, MA_OWNED);
1890 	if (object->type != OBJT_SWAP)
1891 		return;
1892 
1893 	while (object->un_pager.swp.swp_bcount) {
1894 		struct swblock **pswap;
1895 		struct swblock *swap;
1896 
1897 		mtx_lock(&swhash_mtx);
1898 		pswap = swp_pager_hash(object, index);
1899 		if ((swap = *pswap) != NULL) {
1900 			int i;
1901 
1902 			for (i = 0; i < SWAP_META_PAGES; ++i) {
1903 				daddr_t v = swap->swb_pages[i];
1904 				if (v != SWAPBLK_NONE) {
1905 					--swap->swb_count;
1906 					swp_pager_freeswapspace(v, 1);
1907 				}
1908 			}
1909 			if (swap->swb_count != 0)
1910 				panic("swap_pager_meta_free_all: swb_count != 0");
1911 			*pswap = swap->swb_hnext;
1912 			uma_zfree(swap_zone, swap);
1913 			--object->un_pager.swp.swp_bcount;
1914 		}
1915 		mtx_unlock(&swhash_mtx);
1916 		index += SWAP_META_PAGES;
1917 		if (index > 0x20000000)
1918 			panic("swp_pager_meta_free_all: failed to locate all swap meta blocks");
1919 	}
1920 }
1921 
1922 /*
1923  * SWP_PAGER_METACTL() -  misc control of swap and vm_page_t meta data.
1924  *
1925  *	This routine is capable of looking up, popping, or freeing
1926  *	swapblk assignments in the swap meta data or in the vm_page_t.
1927  *	The routine typically returns the swapblk being looked-up, or popped,
1928  *	or SWAPBLK_NONE if the block was freed, or SWAPBLK_NONE if the block
1929  *	was invalid.  This routine will automatically free any invalid
1930  *	meta-data swapblks.
1931  *
1932  *	It is not possible to store invalid swapblks in the swap meta data
1933  *	(other then a literal 'SWAPBLK_NONE'), so we don't bother checking.
1934  *
1935  *	When acting on a busy resident page and paging is in progress, we
1936  *	have to wait until paging is complete but otherwise can act on the
1937  *	busy page.
1938  *
1939  *	This routine must be called at splvm().
1940  *
1941  *	SWM_FREE	remove and free swap block from metadata
1942  *	SWM_POP		remove from meta data but do not free.. pop it out
1943  */
1944 static daddr_t
1945 swp_pager_meta_ctl(vm_object_t object, vm_pindex_t pindex, int flags)
1946 {
1947 	struct swblock **pswap;
1948 	struct swblock *swap;
1949 	daddr_t r1;
1950 	int idx;
1951 
1952 	VM_OBJECT_LOCK_ASSERT(object, MA_OWNED);
1953 	/*
1954 	 * The meta data only exists of the object is OBJT_SWAP
1955 	 * and even then might not be allocated yet.
1956 	 */
1957 	if (object->type != OBJT_SWAP)
1958 		return (SWAPBLK_NONE);
1959 
1960 	r1 = SWAPBLK_NONE;
1961 	mtx_lock(&swhash_mtx);
1962 	pswap = swp_pager_hash(object, pindex);
1963 
1964 	if ((swap = *pswap) != NULL) {
1965 		idx = pindex & SWAP_META_MASK;
1966 		r1 = swap->swb_pages[idx];
1967 
1968 		if (r1 != SWAPBLK_NONE) {
1969 			if (flags & SWM_FREE) {
1970 				swp_pager_freeswapspace(r1, 1);
1971 				r1 = SWAPBLK_NONE;
1972 			}
1973 			if (flags & (SWM_FREE|SWM_POP)) {
1974 				swap->swb_pages[idx] = SWAPBLK_NONE;
1975 				if (--swap->swb_count == 0) {
1976 					*pswap = swap->swb_hnext;
1977 					uma_zfree(swap_zone, swap);
1978 					--object->un_pager.swp.swp_bcount;
1979 				}
1980 			}
1981 		}
1982 	}
1983 	mtx_unlock(&swhash_mtx);
1984 	return (r1);
1985 }
1986 
1987 /*
1988  * System call swapon(name) enables swapping on device name,
1989  * which must be in the swdevsw.  Return EBUSY
1990  * if already swapping on this device.
1991  */
1992 #ifndef _SYS_SYSPROTO_H_
1993 struct swapon_args {
1994 	char *name;
1995 };
1996 #endif
1997 
1998 /*
1999  * MPSAFE
2000  */
2001 /* ARGSUSED */
2002 int
2003 swapon(struct thread *td, struct swapon_args *uap)
2004 {
2005 	struct vattr attr;
2006 	struct vnode *vp;
2007 	struct nameidata nd;
2008 	int error;
2009 
2010 	mtx_lock(&Giant);
2011 	error = suser(td);
2012 	if (error)
2013 		goto done2;
2014 
2015 	while (swdev_syscall_active)
2016 	    tsleep(&swdev_syscall_active, PUSER - 1, "swpon", 0);
2017 	swdev_syscall_active = 1;
2018 
2019 	/*
2020 	 * Swap metadata may not fit in the KVM if we have physical
2021 	 * memory of >1GB.
2022 	 */
2023 	if (swap_zone == NULL) {
2024 		error = ENOMEM;
2025 		goto done;
2026 	}
2027 
2028 	NDINIT(&nd, LOOKUP, FOLLOW, UIO_USERSPACE, uap->name, td);
2029 	error = namei(&nd);
2030 	if (error)
2031 		goto done;
2032 
2033 	NDFREE(&nd, NDF_ONLY_PNBUF);
2034 	vp = nd.ni_vp;
2035 
2036 	if (vn_isdisk(vp, &error)) {
2037 		error = swapongeom(td, vp);
2038 	} else if (vp->v_type == VREG &&
2039 	    (vp->v_mount->mnt_vfc->vfc_flags & VFCF_NETWORK) != 0 &&
2040 	    (error = VOP_GETATTR(vp, &attr, td->td_ucred, td)) == 0) {
2041 		/*
2042 		 * Allow direct swapping to NFS regular files in the same
2043 		 * way that nfs_mountroot() sets up diskless swapping.
2044 		 */
2045 		error = swaponvp(td, vp, attr.va_size / DEV_BSIZE);
2046 	}
2047 
2048 	if (error)
2049 		vrele(vp);
2050 done:
2051 	swdev_syscall_active = 0;
2052 	wakeup_one(&swdev_syscall_active);
2053 done2:
2054 	mtx_unlock(&Giant);
2055 	return (error);
2056 }
2057 
2058 static void
2059 swaponsomething(struct vnode *vp, void *id, u_long nblks, sw_strategy_t *strategy, sw_close_t *close, dev_t dev)
2060 {
2061 	struct swdevt *sp, *tsp;
2062 	swblk_t dvbase;
2063 	u_long mblocks;
2064 
2065 	/*
2066 	 * If we go beyond this, we get overflows in the radix
2067 	 * tree bitmap code.
2068 	 */
2069 	mblocks = 0x40000000 / BLIST_META_RADIX;
2070 	if (nblks > mblocks) {
2071 		printf("WARNING: reducing size to maximum of %lu blocks per swap unit\n",
2072 			mblocks);
2073 		nblks = mblocks;
2074 	}
2075 	/*
2076 	 * nblks is in DEV_BSIZE'd chunks, convert to PAGE_SIZE'd chunks.
2077 	 * First chop nblks off to page-align it, then convert.
2078 	 *
2079 	 * sw->sw_nblks is in page-sized chunks now too.
2080 	 */
2081 	nblks &= ~(ctodb(1) - 1);
2082 	nblks = dbtoc(nblks);
2083 
2084 	sp = malloc(sizeof *sp, M_VMPGDATA, M_WAITOK | M_ZERO);
2085 	sp->sw_vp = vp;
2086 	sp->sw_id = id;
2087 	sp->sw_dev = dev;
2088 	sp->sw_flags = 0;
2089 	sp->sw_nblks = nblks;
2090 	sp->sw_used = 0;
2091 	sp->sw_strategy = strategy;
2092 	sp->sw_close = close;
2093 
2094 	sp->sw_blist = blist_create(nblks);
2095 	/*
2096 	 * Do not free the first two block in order to avoid overwriting
2097 	 * any bsd label at the front of the partition
2098 	 */
2099 	blist_free(sp->sw_blist, 2, nblks - 2);
2100 
2101 	dvbase = 0;
2102 	mtx_lock(&sw_dev_mtx);
2103 	TAILQ_FOREACH(tsp, &swtailq, sw_list) {
2104 		if (tsp->sw_end >= dvbase) {
2105 			/*
2106 			 * We put one uncovered page between the devices
2107 			 * in order to definitively prevent any cross-device
2108 			 * I/O requests
2109 			 */
2110 			dvbase = tsp->sw_end + 1;
2111 		}
2112 	}
2113 	sp->sw_first = dvbase;
2114 	sp->sw_end = dvbase + nblks;
2115 	TAILQ_INSERT_TAIL(&swtailq, sp, sw_list);
2116 	nswapdev++;
2117 	swap_pager_avail += nblks;
2118 	swp_sizecheck();
2119 	mtx_unlock(&sw_dev_mtx);
2120 }
2121 
2122 /*
2123  * SYSCALL: swapoff(devname)
2124  *
2125  * Disable swapping on the given device.
2126  *
2127  * XXX: Badly designed system call: it should use a device index
2128  * rather than filename as specification.  We keep sw_vp around
2129  * only to make this work.
2130  */
2131 #ifndef _SYS_SYSPROTO_H_
2132 struct swapoff_args {
2133 	char *name;
2134 };
2135 #endif
2136 
2137 /*
2138  * MPSAFE
2139  */
2140 /* ARGSUSED */
2141 int
2142 swapoff(struct thread *td, struct swapoff_args *uap)
2143 {
2144 	struct vnode *vp;
2145 	struct nameidata nd;
2146 	struct swdevt *sp;
2147 	u_long nblks, dvbase;
2148 	int error;
2149 
2150 	mtx_lock(&Giant);
2151 
2152 	error = suser(td);
2153 	if (error)
2154 		goto done2;
2155 
2156 	while (swdev_syscall_active)
2157 	    tsleep(&swdev_syscall_active, PUSER - 1, "swpoff", 0);
2158 	swdev_syscall_active = 1;
2159 
2160 	NDINIT(&nd, LOOKUP, FOLLOW, UIO_USERSPACE, uap->name, td);
2161 	error = namei(&nd);
2162 	if (error)
2163 		goto done;
2164 	NDFREE(&nd, NDF_ONLY_PNBUF);
2165 	vp = nd.ni_vp;
2166 
2167 	mtx_lock(&sw_dev_mtx);
2168 	TAILQ_FOREACH(sp, &swtailq, sw_list) {
2169 		if (sp->sw_vp == vp)
2170 			goto found;
2171 	}
2172 	mtx_unlock(&sw_dev_mtx);
2173 	error = EINVAL;
2174 	goto done;
2175 found:
2176 	mtx_unlock(&sw_dev_mtx);
2177 #ifdef MAC
2178 	(void) vn_lock(vp, LK_EXCLUSIVE | LK_RETRY, td);
2179 	error = mac_check_system_swapoff(td->td_ucred, vp);
2180 	(void) VOP_UNLOCK(vp, 0, td);
2181 	if (error != 0)
2182 		goto done;
2183 #endif
2184 
2185 	nblks = sp->sw_nblks;
2186 
2187 	/*
2188 	 * We can turn off this swap device safely only if the
2189 	 * available virtual memory in the system will fit the amount
2190 	 * of data we will have to page back in, plus an epsilon so
2191 	 * the system doesn't become critically low on swap space.
2192 	 */
2193 	if (cnt.v_free_count + cnt.v_cache_count + swap_pager_avail <
2194 	    nblks + nswap_lowat) {
2195 		error = ENOMEM;
2196 		goto done;
2197 	}
2198 
2199 	/*
2200 	 * Prevent further allocations on this device.
2201 	 */
2202 	mtx_lock(&sw_dev_mtx);
2203 	sp->sw_flags |= SW_CLOSING;
2204 	for (dvbase = 0; dvbase < sp->sw_end; dvbase += dmmax) {
2205 		swap_pager_avail -= blist_fill(sp->sw_blist,
2206 		     dvbase, dmmax);
2207 	}
2208 	mtx_unlock(&sw_dev_mtx);
2209 
2210 	/*
2211 	 * Page in the contents of the device and close it.
2212 	 */
2213 	swap_pager_swapoff(sp);
2214 
2215 	sp->sw_close(td, sp);
2216 	sp->sw_id = NULL;
2217 	mtx_lock(&sw_dev_mtx);
2218 	TAILQ_REMOVE(&swtailq, sp, sw_list);
2219 	nswapdev--;
2220 	if (nswapdev == 0) {
2221 		swap_pager_full = 2;
2222 		swap_pager_almost_full = 1;
2223 	}
2224 	if (swdevhd == sp)
2225 		swdevhd = NULL;
2226 	mtx_unlock(&sw_dev_mtx);
2227 	blist_destroy(sp->sw_blist);
2228 	free(sp, M_VMPGDATA);
2229 
2230 done:
2231 	swdev_syscall_active = 0;
2232 	wakeup_one(&swdev_syscall_active);
2233 done2:
2234 	mtx_unlock(&Giant);
2235 	return (error);
2236 }
2237 
2238 void
2239 swap_pager_status(int *total, int *used)
2240 {
2241 	struct swdevt *sp;
2242 
2243 	*total = 0;
2244 	*used = 0;
2245 	mtx_lock(&sw_dev_mtx);
2246 	TAILQ_FOREACH(sp, &swtailq, sw_list) {
2247 		*total += sp->sw_nblks;
2248 		*used += sp->sw_used;
2249 	}
2250 	mtx_unlock(&sw_dev_mtx);
2251 }
2252 
2253 static int
2254 sysctl_vm_swap_info(SYSCTL_HANDLER_ARGS)
2255 {
2256 	int	*name = (int *)arg1;
2257 	int	error, n;
2258 	struct xswdev xs;
2259 	struct swdevt *sp;
2260 
2261 	if (arg2 != 1) /* name length */
2262 		return (EINVAL);
2263 
2264 	n = 0;
2265 	mtx_lock(&sw_dev_mtx);
2266 	TAILQ_FOREACH(sp, &swtailq, sw_list) {
2267 		if (n == *name) {
2268 			mtx_unlock(&sw_dev_mtx);
2269 			xs.xsw_version = XSWDEV_VERSION;
2270 			xs.xsw_dev = sp->sw_dev;
2271 			xs.xsw_flags = sp->sw_flags;
2272 			xs.xsw_nblks = sp->sw_nblks;
2273 			xs.xsw_used = sp->sw_used;
2274 
2275 			error = SYSCTL_OUT(req, &xs, sizeof(xs));
2276 			return (error);
2277 		}
2278 		n++;
2279 	}
2280 	mtx_unlock(&sw_dev_mtx);
2281 	return (ENOENT);
2282 }
2283 
2284 SYSCTL_INT(_vm, OID_AUTO, nswapdev, CTLFLAG_RD, &nswapdev, 0,
2285     "Number of swap devices");
2286 SYSCTL_NODE(_vm, OID_AUTO, swap_info, CTLFLAG_RD, sysctl_vm_swap_info,
2287     "Swap statistics by device");
2288 
2289 /*
2290  * vmspace_swap_count() - count the approximate swap useage in pages for a
2291  *			  vmspace.
2292  *
2293  *	The map must be locked.
2294  *
2295  *	Swap useage is determined by taking the proportional swap used by
2296  *	VM objects backing the VM map.  To make up for fractional losses,
2297  *	if the VM object has any swap use at all the associated map entries
2298  *	count for at least 1 swap page.
2299  */
2300 int
2301 vmspace_swap_count(struct vmspace *vmspace)
2302 {
2303 	vm_map_t map = &vmspace->vm_map;
2304 	vm_map_entry_t cur;
2305 	int count = 0;
2306 
2307 	for (cur = map->header.next; cur != &map->header; cur = cur->next) {
2308 		vm_object_t object;
2309 
2310 		if ((cur->eflags & MAP_ENTRY_IS_SUB_MAP) == 0 &&
2311 		    (object = cur->object.vm_object) != NULL) {
2312 			VM_OBJECT_LOCK(object);
2313 			if (object->type == OBJT_SWAP &&
2314 			    object->un_pager.swp.swp_bcount != 0) {
2315 				int n = (cur->end - cur->start) / PAGE_SIZE;
2316 
2317 				count += object->un_pager.swp.swp_bcount *
2318 				    SWAP_META_PAGES * n / object->size + 1;
2319 			}
2320 			VM_OBJECT_UNLOCK(object);
2321 		}
2322 	}
2323 	return (count);
2324 }
2325 
2326 /*
2327  * GEOM backend
2328  *
2329  * Swapping onto disk devices.
2330  *
2331  */
2332 
2333 static g_orphan_t swapgeom_orphan;
2334 
2335 static struct g_class g_swap_class = {
2336 	.name = "SWAP",
2337 	.version = G_VERSION,
2338 	.orphan = swapgeom_orphan,
2339 };
2340 
2341 DECLARE_GEOM_CLASS(g_swap_class, g_class);
2342 
2343 
2344 static void
2345 swapgeom_done(struct bio *bp2)
2346 {
2347 	struct buf *bp;
2348 
2349 	bp = bp2->bio_caller2;
2350 	bp->b_ioflags = bp2->bio_flags;
2351 	if (bp2->bio_error)
2352 		bp->b_ioflags |= BIO_ERROR;
2353 	bp->b_resid = bp->b_bcount - bp2->bio_completed;
2354 	bp->b_error = bp2->bio_error;
2355 	bufdone(bp);
2356 	g_destroy_bio(bp2);
2357 }
2358 
2359 static void
2360 swapgeom_strategy(struct buf *bp, struct swdevt *sp)
2361 {
2362 	struct bio *bio;
2363 	struct g_consumer *cp;
2364 
2365 	cp = sp->sw_id;
2366 	if (cp == NULL) {
2367 		bp->b_error = ENXIO;
2368 		bp->b_ioflags |= BIO_ERROR;
2369 		bufdone(bp);
2370 		return;
2371 	}
2372 	bio = g_alloc_bio();
2373 #if 0
2374 	/*
2375 	 * XXX: We shouldn't really sleep here when we run out of buffers
2376 	 * XXX: but the alternative is worse right now.
2377 	 */
2378 	if (bio == NULL) {
2379 		bp->b_error = ENOMEM;
2380 		bp->b_ioflags |= BIO_ERROR;
2381 		bufdone(bp);
2382 		return;
2383 	}
2384 #endif
2385 	bio->bio_caller2 = bp;
2386 	bio->bio_cmd = bp->b_iocmd;
2387 	bio->bio_data = bp->b_data;
2388 	bio->bio_offset = (bp->b_blkno - sp->sw_first) * PAGE_SIZE;
2389 	bio->bio_length = bp->b_bcount;
2390 	bio->bio_done = swapgeom_done;
2391 	g_io_request(bio, cp);
2392 	return;
2393 }
2394 
2395 static void
2396 swapgeom_orphan(struct g_consumer *cp)
2397 {
2398 	struct swdevt *sp;
2399 
2400 	mtx_lock(&sw_dev_mtx);
2401 	TAILQ_FOREACH(sp, &swtailq, sw_list)
2402 		if (sp->sw_id == cp)
2403 			sp->sw_id = NULL;
2404 	mtx_unlock(&sw_dev_mtx);
2405 }
2406 
2407 static void
2408 swapgeom_close_ev(void *arg, int flags)
2409 {
2410 	struct g_consumer *cp;
2411 
2412 	cp = arg;
2413 	g_access(cp, -1, -1, 0);
2414 	g_detach(cp);
2415 	g_destroy_consumer(cp);
2416 }
2417 
2418 static void
2419 swapgeom_close(struct thread *td, struct swdevt *sw)
2420 {
2421 
2422 	/* XXX: direct call when Giant untangled */
2423 	g_waitfor_event(swapgeom_close_ev, sw->sw_id, M_WAITOK, NULL);
2424 }
2425 
2426 
2427 struct swh0h0 {
2428 	struct cdev *dev;
2429 	struct vnode *vp;
2430 	int	error;
2431 };
2432 
2433 static void
2434 swapongeom_ev(void *arg, int flags)
2435 {
2436 	struct swh0h0 *swh;
2437 	struct g_provider *pp;
2438 	struct g_consumer *cp;
2439 	static struct g_geom *gp;
2440 	struct swdevt *sp;
2441 	u_long nblks;
2442 	int error;
2443 
2444 	swh = arg;
2445 	swh->error = 0;
2446 	pp = g_dev_getprovider(swh->dev);
2447 	if (pp == NULL) {
2448 		swh->error = ENODEV;
2449 		return;
2450 	}
2451 	mtx_lock(&sw_dev_mtx);
2452 	TAILQ_FOREACH(sp, &swtailq, sw_list) {
2453 		cp = sp->sw_id;
2454 		if (cp != NULL && cp->provider == pp) {
2455 			mtx_unlock(&sw_dev_mtx);
2456 			swh->error = EBUSY;
2457 			return;
2458 		}
2459 	}
2460 	mtx_unlock(&sw_dev_mtx);
2461 	if (gp == NULL)
2462 		gp = g_new_geomf(&g_swap_class, "swap", NULL);
2463 	cp = g_new_consumer(gp);
2464 	g_attach(cp, pp);
2465 	/*
2466 	 * XXX: Everytime you think you can improve the margin for
2467 	 * footshooting, somebody depends on the ability to do so:
2468 	 * savecore(8) wants to write to our swapdev so we cannot
2469 	 * set an exclusive count :-(
2470 	 */
2471 	error = g_access(cp, 1, 1, 0);
2472 	if (error) {
2473 		g_detach(cp);
2474 		g_destroy_consumer(cp);
2475 		swh->error = error;
2476 		return;
2477 	}
2478 	nblks = pp->mediasize / DEV_BSIZE;
2479 	swaponsomething(swh->vp, cp, nblks, swapgeom_strategy,
2480 	    swapgeom_close, dev2udev(swh->dev));
2481 	swh->error = 0;
2482 	return;
2483 }
2484 
2485 static int
2486 swapongeom(struct thread *td, struct vnode *vp)
2487 {
2488 	int error;
2489 	struct swh0h0 swh;
2490 
2491 	vn_lock(vp, LK_EXCLUSIVE | LK_RETRY, td);
2492 
2493 	swh.dev = vp->v_rdev;
2494 	swh.vp = vp;
2495 	swh.error = 0;
2496 	/* XXX: direct call when Giant untangled */
2497 	error = g_waitfor_event(swapongeom_ev, &swh, M_WAITOK, NULL);
2498 	if (!error)
2499 		error = swh.error;
2500 	VOP_UNLOCK(vp, 0, td);
2501 	return (error);
2502 }
2503 
2504 /*
2505  * VNODE backend
2506  *
2507  * This is used mainly for network filesystem (read: probably only tested
2508  * with NFS) swapfiles.
2509  *
2510  */
2511 
2512 static void
2513 swapdev_strategy(struct buf *bp, struct swdevt *sp)
2514 {
2515 	int s;
2516 	struct vnode *vp2;
2517 
2518 	bp->b_blkno = ctodb(bp->b_blkno - sp->sw_first);
2519 
2520 	vp2 = sp->sw_id;
2521 	vhold(vp2);
2522 	s = splvm();
2523 	if (bp->b_iocmd == BIO_WRITE) {
2524 		if (bp->b_bufobj) /* XXX: should always be true /phk */
2525 			bufobj_wdrop(bp->b_bufobj);
2526 		bufobj_wref(&vp2->v_bufobj);
2527 	}
2528 	bp->b_vp = vp2;
2529 	splx(s);
2530 	bp->b_iooffset = dbtob(bp->b_blkno);
2531 	bstrategy(bp);
2532 	return;
2533 }
2534 
2535 static void
2536 swapdev_close(struct thread *td, struct swdevt *sp)
2537 {
2538 
2539 	VOP_CLOSE(sp->sw_vp, FREAD | FWRITE, td->td_ucred, td);
2540 	vrele(sp->sw_vp);
2541 }
2542 
2543 
2544 static int
2545 swaponvp(struct thread *td, struct vnode *vp, u_long nblks)
2546 {
2547 	struct swdevt *sp;
2548 	int error;
2549 
2550 	if (nblks == 0)
2551 		return (ENXIO);
2552 	mtx_lock(&sw_dev_mtx);
2553 	TAILQ_FOREACH(sp, &swtailq, sw_list) {
2554 		if (sp->sw_id == vp) {
2555 			mtx_unlock(&sw_dev_mtx);
2556 			return (EBUSY);
2557 		}
2558 	}
2559 	mtx_unlock(&sw_dev_mtx);
2560 
2561 	(void) vn_lock(vp, LK_EXCLUSIVE | LK_RETRY, td);
2562 #ifdef MAC
2563 	error = mac_check_system_swapon(td->td_ucred, vp);
2564 	if (error == 0)
2565 #endif
2566 		error = VOP_OPEN(vp, FREAD | FWRITE, td->td_ucred, td, -1);
2567 	(void) VOP_UNLOCK(vp, 0, td);
2568 	if (error)
2569 		return (error);
2570 
2571 	swaponsomething(vp, vp, nblks, swapdev_strategy, swapdev_close,
2572 	    NODEV);
2573 	return (0);
2574 }
2575