xref: /freebsd/sys/vm/swap_pager.c (revision 2357939bc239bd5334a169b62313806178dd8f30)
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 	udev_t	sw_udev;
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 
223 /*
224  * pagerops for OBJT_SWAP - "swap pager".  Some ops are also global procedure
225  * calls hooked from other parts of the VM system and do not appear here.
226  * (see vm/swap_pager.h).
227  */
228 static vm_object_t
229 		swap_pager_alloc(void *handle, vm_ooffset_t size,
230 				      vm_prot_t prot, vm_ooffset_t offset);
231 static void	swap_pager_dealloc(vm_object_t object);
232 static int	swap_pager_getpages(vm_object_t, vm_page_t *, int, int);
233 static void	swap_pager_putpages(vm_object_t, vm_page_t *, int, boolean_t, int *);
234 static boolean_t
235 		swap_pager_haspage(vm_object_t object, vm_pindex_t pindex, int *before, int *after);
236 static void	swap_pager_init(void);
237 static void	swap_pager_unswapped(vm_page_t);
238 static void	swap_pager_swapoff(struct swdevt *sp, int *sw_used);
239 
240 struct pagerops swappagerops = {
241 	.pgo_init =	swap_pager_init,	/* early system initialization of pager	*/
242 	.pgo_alloc =	swap_pager_alloc,	/* allocate an OBJT_SWAP object		*/
243 	.pgo_dealloc =	swap_pager_dealloc,	/* deallocate an OBJT_SWAP object	*/
244 	.pgo_getpages =	swap_pager_getpages,	/* pagein				*/
245 	.pgo_putpages =	swap_pager_putpages,	/* pageout				*/
246 	.pgo_haspage =	swap_pager_haspage,	/* get backing store status for page	*/
247 	.pgo_pageunswapped = swap_pager_unswapped,	/* remove swap related to page		*/
248 };
249 
250 /*
251  * dmmax is in page-sized chunks with the new swap system.  It was
252  * dev-bsized chunks in the old.  dmmax is always a power of 2.
253  *
254  * swap_*() routines are externally accessible.  swp_*() routines are
255  * internal.
256  */
257 static int dmmax;
258 static int nswap_lowat = 128;	/* in pages, swap_pager_almost_full warn */
259 static int nswap_hiwat = 512;	/* in pages, swap_pager_almost_full warn */
260 
261 SYSCTL_INT(_vm, OID_AUTO, dmmax,
262 	CTLFLAG_RD, &dmmax, 0, "Maximum size of a swap block");
263 
264 static void	swp_sizecheck(void);
265 static void	swp_pager_async_iodone(struct buf *bp);
266 static int	swapongeom(struct thread *, struct vnode *);
267 static int	swaponvp(struct thread *, struct vnode *, u_long);
268 
269 /*
270  * Swap bitmap functions
271  */
272 static void	swp_pager_freeswapspace(daddr_t blk, int npages);
273 static daddr_t	swp_pager_getswapspace(int npages);
274 
275 /*
276  * Metadata functions
277  */
278 static struct swblock **swp_pager_hash(vm_object_t object, vm_pindex_t index);
279 static void swp_pager_meta_build(vm_object_t, vm_pindex_t, daddr_t);
280 static void swp_pager_meta_free(vm_object_t, vm_pindex_t, daddr_t);
281 static void swp_pager_meta_free_all(vm_object_t);
282 static daddr_t swp_pager_meta_ctl(vm_object_t, vm_pindex_t, int);
283 
284 /*
285  * SWP_SIZECHECK() -	update swap_pager_full indication
286  *
287  *	update the swap_pager_almost_full indication and warn when we are
288  *	about to run out of swap space, using lowat/hiwat hysteresis.
289  *
290  *	Clear swap_pager_full ( task killing ) indication when lowat is met.
291  *
292  *	No restrictions on call
293  *	This routine may not block.
294  *	This routine must be called at splvm()
295  */
296 static void
297 swp_sizecheck(void)
298 {
299 
300 	if (swap_pager_avail < nswap_lowat) {
301 		if (swap_pager_almost_full == 0) {
302 			printf("swap_pager: out of swap space\n");
303 			swap_pager_almost_full = 1;
304 		}
305 	} else {
306 		swap_pager_full = 0;
307 		if (swap_pager_avail > nswap_hiwat)
308 			swap_pager_almost_full = 0;
309 	}
310 }
311 
312 /*
313  * SWP_PAGER_HASH() -	hash swap meta data
314  *
315  *	This is an helper function which hashes the swapblk given
316  *	the object and page index.  It returns a pointer to a pointer
317  *	to the object, or a pointer to a NULL pointer if it could not
318  *	find a swapblk.
319  *
320  *	This routine must be called at splvm().
321  */
322 static struct swblock **
323 swp_pager_hash(vm_object_t object, vm_pindex_t index)
324 {
325 	struct swblock **pswap;
326 	struct swblock *swap;
327 
328 	index &= ~(vm_pindex_t)SWAP_META_MASK;
329 	pswap = &swhash[(index ^ (int)(intptr_t)object) & swhash_mask];
330 	while ((swap = *pswap) != NULL) {
331 		if (swap->swb_object == object &&
332 		    swap->swb_index == index
333 		) {
334 			break;
335 		}
336 		pswap = &swap->swb_hnext;
337 	}
338 	return (pswap);
339 }
340 
341 /*
342  * SWAP_PAGER_INIT() -	initialize the swap pager!
343  *
344  *	Expected to be started from system init.  NOTE:  This code is run
345  *	before much else so be careful what you depend on.  Most of the VM
346  *	system has yet to be initialized at this point.
347  */
348 static void
349 swap_pager_init(void)
350 {
351 	/*
352 	 * Initialize object lists
353 	 */
354 	int i;
355 
356 	for (i = 0; i < NOBJLISTS; ++i)
357 		TAILQ_INIT(&swap_pager_object_list[i]);
358 	mtx_init(&sw_alloc_mtx, "swap_pager list", NULL, MTX_DEF);
359 	mtx_init(&sw_dev_mtx, "swapdev", NULL, MTX_DEF);
360 
361 	/*
362 	 * Device Stripe, in PAGE_SIZE'd blocks
363 	 */
364 	dmmax = SWB_NPAGES * 2;
365 }
366 
367 /*
368  * SWAP_PAGER_SWAP_INIT() - swap pager initialization from pageout process
369  *
370  *	Expected to be started from pageout process once, prior to entering
371  *	its main loop.
372  */
373 void
374 swap_pager_swap_init(void)
375 {
376 	int n, n2;
377 
378 	/*
379 	 * Number of in-transit swap bp operations.  Don't
380 	 * exhaust the pbufs completely.  Make sure we
381 	 * initialize workable values (0 will work for hysteresis
382 	 * but it isn't very efficient).
383 	 *
384 	 * The nsw_cluster_max is constrained by the bp->b_pages[]
385 	 * array (MAXPHYS/PAGE_SIZE) and our locally defined
386 	 * MAX_PAGEOUT_CLUSTER.   Also be aware that swap ops are
387 	 * constrained by the swap device interleave stripe size.
388 	 *
389 	 * Currently we hardwire nsw_wcount_async to 4.  This limit is
390 	 * designed to prevent other I/O from having high latencies due to
391 	 * our pageout I/O.  The value 4 works well for one or two active swap
392 	 * devices but is probably a little low if you have more.  Even so,
393 	 * a higher value would probably generate only a limited improvement
394 	 * with three or four active swap devices since the system does not
395 	 * typically have to pageout at extreme bandwidths.   We will want
396 	 * at least 2 per swap devices, and 4 is a pretty good value if you
397 	 * have one NFS swap device due to the command/ack latency over NFS.
398 	 * So it all works out pretty well.
399 	 */
400 	nsw_cluster_max = min((MAXPHYS/PAGE_SIZE), MAX_PAGEOUT_CLUSTER);
401 
402 	mtx_lock(&pbuf_mtx);
403 	nsw_rcount = (nswbuf + 1) / 2;
404 	nsw_wcount_sync = (nswbuf + 3) / 4;
405 	nsw_wcount_async = 4;
406 	nsw_wcount_async_max = nsw_wcount_async;
407 	mtx_unlock(&pbuf_mtx);
408 
409 	/*
410 	 * Initialize our zone.  Right now I'm just guessing on the number
411 	 * we need based on the number of pages in the system.  Each swblock
412 	 * can hold 16 pages, so this is probably overkill.  This reservation
413 	 * is typically limited to around 32MB by default.
414 	 */
415 	n = cnt.v_page_count / 2;
416 	if (maxswzone && n > maxswzone / sizeof(struct swblock))
417 		n = maxswzone / sizeof(struct swblock);
418 	n2 = n;
419 	swap_zone = uma_zcreate("SWAPMETA", sizeof(struct swblock), NULL, NULL,
420 	    NULL, NULL, UMA_ALIGN_PTR, UMA_ZONE_NOFREE | UMA_ZONE_VM);
421 	do {
422 		if (uma_zone_set_obj(swap_zone, NULL, n))
423 			break;
424 		/*
425 		 * if the allocation failed, try a zone two thirds the
426 		 * size of the previous attempt.
427 		 */
428 		n -= ((n + 2) / 3);
429 	} while (n > 0);
430 	if (swap_zone == NULL)
431 		panic("failed to create swap_zone.");
432 	if (n2 != n)
433 		printf("Swap zone entries reduced from %d to %d.\n", n2, n);
434 	n2 = n;
435 
436 	/*
437 	 * Initialize our meta-data hash table.  The swapper does not need to
438 	 * be quite as efficient as the VM system, so we do not use an
439 	 * oversized hash table.
440 	 *
441 	 * 	n: 		size of hash table, must be power of 2
442 	 *	swhash_mask:	hash table index mask
443 	 */
444 	for (n = 1; n < n2 / 8; n *= 2)
445 		;
446 	swhash = malloc(sizeof(struct swblock *) * n, M_VMPGDATA, M_WAITOK | M_ZERO);
447 	swhash_mask = n - 1;
448 	mtx_init(&swhash_mtx, "swap_pager swhash", NULL, MTX_DEF);
449 }
450 
451 /*
452  * SWAP_PAGER_ALLOC() -	allocate a new OBJT_SWAP VM object and instantiate
453  *			its metadata structures.
454  *
455  *	This routine is called from the mmap and fork code to create a new
456  *	OBJT_SWAP object.  We do this by creating an OBJT_DEFAULT object
457  *	and then converting it with swp_pager_meta_build().
458  *
459  *	This routine may block in vm_object_allocate() and create a named
460  *	object lookup race, so we must interlock.   We must also run at
461  *	splvm() for the object lookup to handle races with interrupts, but
462  *	we do not have to maintain splvm() in between the lookup and the
463  *	add because (I believe) it is not possible to attempt to create
464  *	a new swap object w/handle when a default object with that handle
465  *	already exists.
466  *
467  * MPSAFE
468  */
469 static vm_object_t
470 swap_pager_alloc(void *handle, vm_ooffset_t size, vm_prot_t prot,
471 		 vm_ooffset_t offset)
472 {
473 	vm_object_t object;
474 	vm_pindex_t pindex;
475 
476 	pindex = OFF_TO_IDX(offset + PAGE_MASK + size);
477 
478 	if (handle) {
479 		mtx_lock(&Giant);
480 		/*
481 		 * Reference existing named region or allocate new one.  There
482 		 * should not be a race here against swp_pager_meta_build()
483 		 * as called from vm_page_remove() in regards to the lookup
484 		 * of the handle.
485 		 */
486 		sx_xlock(&sw_alloc_sx);
487 		object = vm_pager_object_lookup(NOBJLIST(handle), handle);
488 
489 		if (object != NULL) {
490 			vm_object_reference(object);
491 		} else {
492 			object = vm_object_allocate(OBJT_DEFAULT, pindex);
493 			object->handle = handle;
494 
495 			VM_OBJECT_LOCK(object);
496 			swp_pager_meta_build(object, 0, SWAPBLK_NONE);
497 			VM_OBJECT_UNLOCK(object);
498 		}
499 		sx_xunlock(&sw_alloc_sx);
500 		mtx_unlock(&Giant);
501 	} else {
502 		object = vm_object_allocate(OBJT_DEFAULT, pindex);
503 
504 		VM_OBJECT_LOCK(object);
505 		swp_pager_meta_build(object, 0, SWAPBLK_NONE);
506 		VM_OBJECT_UNLOCK(object);
507 	}
508 	return (object);
509 }
510 
511 /*
512  * SWAP_PAGER_DEALLOC() -	remove swap metadata from object
513  *
514  *	The swap backing for the object is destroyed.  The code is
515  *	designed such that we can reinstantiate it later, but this
516  *	routine is typically called only when the entire object is
517  *	about to be destroyed.
518  *
519  *	This routine may block, but no longer does.
520  *
521  *	The object must be locked or unreferenceable.
522  */
523 static void
524 swap_pager_dealloc(vm_object_t object)
525 {
526 	int s;
527 
528 	/*
529 	 * Remove from list right away so lookups will fail if we block for
530 	 * pageout completion.
531 	 */
532 	if (object->handle != NULL) {
533 		mtx_lock(&sw_alloc_mtx);
534 		TAILQ_REMOVE(NOBJLIST(object->handle), object, pager_object_list);
535 		mtx_unlock(&sw_alloc_mtx);
536 	}
537 
538 	VM_OBJECT_LOCK_ASSERT(object, MA_OWNED);
539 	vm_object_pip_wait(object, "swpdea");
540 
541 	/*
542 	 * Free all remaining metadata.  We only bother to free it from
543 	 * the swap meta data.  We do not attempt to free swapblk's still
544 	 * associated with vm_page_t's for this object.  We do not care
545 	 * if paging is still in progress on some objects.
546 	 */
547 	s = splvm();
548 	swp_pager_meta_free_all(object);
549 	splx(s);
550 }
551 
552 /************************************************************************
553  *			SWAP PAGER BITMAP ROUTINES			*
554  ************************************************************************/
555 
556 /*
557  * SWP_PAGER_GETSWAPSPACE() -	allocate raw swap space
558  *
559  *	Allocate swap for the requested number of pages.  The starting
560  *	swap block number (a page index) is returned or SWAPBLK_NONE
561  *	if the allocation failed.
562  *
563  *	Also has the side effect of advising that somebody made a mistake
564  *	when they configured swap and didn't configure enough.
565  *
566  *	Must be called at splvm() to avoid races with bitmap frees from
567  *	vm_page_remove() aka swap_pager_page_removed().
568  *
569  *	This routine may not block
570  *	This routine must be called at splvm().
571  *
572  *	We allocate in round-robin fashion from the configured devices.
573  */
574 static daddr_t
575 swp_pager_getswapspace(int npages)
576 {
577 	daddr_t blk;
578 	struct swdevt *sp;
579 	int i;
580 
581 	blk = SWAPBLK_NONE;
582 	mtx_lock(&sw_dev_mtx);
583 	sp = swdevhd;
584 	for (i = 0; i < nswapdev; i++) {
585 		if (sp == NULL)
586 			sp = TAILQ_FIRST(&swtailq);
587 		if (!(sp->sw_flags & SW_CLOSING)) {
588 			blk = blist_alloc(sp->sw_blist, npages);
589 			if (blk != SWAPBLK_NONE) {
590 				blk += sp->sw_first;
591 				sp->sw_used += npages;
592 				swap_pager_avail -= npages;
593 				swp_sizecheck();
594 				swdevhd = TAILQ_NEXT(sp, sw_list);
595 				goto done;
596 			}
597 		}
598 		sp = TAILQ_NEXT(sp, sw_list);
599 	}
600 	if (swap_pager_full != 2) {
601 		printf("swap_pager_getswapspace(%d): failed\n", npages);
602 		swap_pager_full = 2;
603 		swap_pager_almost_full = 1;
604 	}
605 	swdevhd = NULL;
606 done:
607 	mtx_unlock(&sw_dev_mtx);
608 	return (blk);
609 }
610 
611 static struct swdevt *
612 swp_pager_find_dev(daddr_t blk)
613 {
614 	struct swdevt *sp;
615 
616 	mtx_lock(&sw_dev_mtx);
617 	TAILQ_FOREACH(sp, &swtailq, sw_list) {
618 		if (blk >= sp->sw_first && blk < sp->sw_end) {
619 			mtx_unlock(&sw_dev_mtx);
620 			return (sp);
621 		}
622 	}
623 	panic("Swapdev not found");
624 }
625 
626 static void
627 swp_pager_strategy(struct buf *bp)
628 {
629 	struct swdevt *sp;
630 
631 	mtx_lock(&sw_dev_mtx);
632 	TAILQ_FOREACH(sp, &swtailq, sw_list) {
633 		if (bp->b_blkno >= sp->sw_first && bp->b_blkno < sp->sw_end) {
634 			mtx_unlock(&sw_dev_mtx);
635 			sp->sw_strategy(bp, sp);
636 			return;
637 		}
638 	}
639 	panic("Swapdev not found");
640 }
641 
642 
643 /*
644  * SWP_PAGER_FREESWAPSPACE() -	free raw swap space
645  *
646  *	This routine returns the specified swap blocks back to the bitmap.
647  *
648  *	Note:  This routine may not block (it could in the old swap code),
649  *	and through the use of the new blist routines it does not block.
650  *
651  *	We must be called at splvm() to avoid races with bitmap frees from
652  *	vm_page_remove() aka swap_pager_page_removed().
653  *
654  *	This routine may not block
655  *	This routine must be called at splvm().
656  */
657 static void
658 swp_pager_freeswapspace(daddr_t blk, int npages)
659 {
660 	struct swdevt *sp;
661 
662 	mtx_lock(&sw_dev_mtx);
663 	TAILQ_FOREACH(sp, &swtailq, sw_list) {
664 		if (blk >= sp->sw_first && blk < sp->sw_end) {
665 			sp->sw_used -= npages;
666 			/*
667 			 * If we are attempting to stop swapping on
668 			 * this device, we don't want to mark any
669 			 * blocks free lest they be reused.
670 			 */
671 			if ((sp->sw_flags & SW_CLOSING) == 0) {
672 				blist_free(sp->sw_blist, blk - sp->sw_first,
673 				    npages);
674 				swap_pager_avail += npages;
675 				swp_sizecheck();
676 			}
677 			mtx_unlock(&sw_dev_mtx);
678 			return;
679 		}
680 	}
681 	panic("Swapdev not found");
682 }
683 
684 /*
685  * SWAP_PAGER_FREESPACE() -	frees swap blocks associated with a page
686  *				range within an object.
687  *
688  *	This is a globally accessible routine.
689  *
690  *	This routine removes swapblk assignments from swap metadata.
691  *
692  *	The external callers of this routine typically have already destroyed
693  *	or renamed vm_page_t's associated with this range in the object so
694  *	we should be ok.
695  *
696  *	This routine may be called at any spl.  We up our spl to splvm temporarily
697  *	in order to perform the metadata removal.
698  */
699 void
700 swap_pager_freespace(vm_object_t object, vm_pindex_t start, vm_size_t size)
701 {
702 	int s = splvm();
703 
704 	VM_OBJECT_LOCK_ASSERT(object, MA_OWNED);
705 	swp_pager_meta_free(object, start, size);
706 	splx(s);
707 }
708 
709 /*
710  * SWAP_PAGER_RESERVE() - reserve swap blocks in object
711  *
712  *	Assigns swap blocks to the specified range within the object.  The
713  *	swap blocks are not zerod.  Any previous swap assignment is destroyed.
714  *
715  *	Returns 0 on success, -1 on failure.
716  */
717 int
718 swap_pager_reserve(vm_object_t object, vm_pindex_t start, vm_size_t size)
719 {
720 	int s;
721 	int n = 0;
722 	daddr_t blk = SWAPBLK_NONE;
723 	vm_pindex_t beg = start;	/* save start index */
724 
725 	s = splvm();
726 	VM_OBJECT_LOCK(object);
727 	while (size) {
728 		if (n == 0) {
729 			n = BLIST_MAX_ALLOC;
730 			while ((blk = swp_pager_getswapspace(n)) == SWAPBLK_NONE) {
731 				n >>= 1;
732 				if (n == 0) {
733 					swp_pager_meta_free(object, beg, start - beg);
734 					VM_OBJECT_UNLOCK(object);
735 					splx(s);
736 					return (-1);
737 				}
738 			}
739 		}
740 		swp_pager_meta_build(object, start, blk);
741 		--size;
742 		++start;
743 		++blk;
744 		--n;
745 	}
746 	swp_pager_meta_free(object, start, n);
747 	VM_OBJECT_UNLOCK(object);
748 	splx(s);
749 	return (0);
750 }
751 
752 /*
753  * SWAP_PAGER_COPY() -  copy blocks from source pager to destination pager
754  *			and destroy the source.
755  *
756  *	Copy any valid swapblks from the source to the destination.  In
757  *	cases where both the source and destination have a valid swapblk,
758  *	we keep the destination's.
759  *
760  *	This routine is allowed to block.  It may block allocating metadata
761  *	indirectly through swp_pager_meta_build() or if paging is still in
762  *	progress on the source.
763  *
764  *	This routine can be called at any spl
765  *
766  *	XXX vm_page_collapse() kinda expects us not to block because we
767  *	supposedly do not need to allocate memory, but for the moment we
768  *	*may* have to get a little memory from the zone allocator, but
769  *	it is taken from the interrupt memory.  We should be ok.
770  *
771  *	The source object contains no vm_page_t's (which is just as well)
772  *
773  *	The source object is of type OBJT_SWAP.
774  *
775  *	The source and destination objects must be locked or
776  *	inaccessible (XXX are they ?)
777  */
778 void
779 swap_pager_copy(vm_object_t srcobject, vm_object_t dstobject,
780     vm_pindex_t offset, int destroysource)
781 {
782 	vm_pindex_t i;
783 	int s;
784 
785 	VM_OBJECT_LOCK_ASSERT(srcobject, MA_OWNED);
786 	VM_OBJECT_LOCK_ASSERT(dstobject, MA_OWNED);
787 
788 	s = splvm();
789 	/*
790 	 * If destroysource is set, we remove the source object from the
791 	 * swap_pager internal queue now.
792 	 */
793 	if (destroysource) {
794 		if (srcobject->handle != NULL) {
795 			mtx_lock(&sw_alloc_mtx);
796 			TAILQ_REMOVE(
797 			    NOBJLIST(srcobject->handle),
798 			    srcobject,
799 			    pager_object_list
800 			);
801 			mtx_unlock(&sw_alloc_mtx);
802 		}
803 	}
804 
805 	/*
806 	 * transfer source to destination.
807 	 */
808 	for (i = 0; i < dstobject->size; ++i) {
809 		daddr_t dstaddr;
810 
811 		/*
812 		 * Locate (without changing) the swapblk on the destination,
813 		 * unless it is invalid in which case free it silently, or
814 		 * if the destination is a resident page, in which case the
815 		 * source is thrown away.
816 		 */
817 		dstaddr = swp_pager_meta_ctl(dstobject, i, 0);
818 
819 		if (dstaddr == SWAPBLK_NONE) {
820 			/*
821 			 * Destination has no swapblk and is not resident,
822 			 * copy source.
823 			 */
824 			daddr_t srcaddr;
825 
826 			srcaddr = swp_pager_meta_ctl(
827 			    srcobject,
828 			    i + offset,
829 			    SWM_POP
830 			);
831 
832 			if (srcaddr != SWAPBLK_NONE) {
833 				/*
834 				 * swp_pager_meta_build() can sleep.
835 				 */
836 				vm_object_pip_add(srcobject, 1);
837 				VM_OBJECT_UNLOCK(srcobject);
838 				vm_object_pip_add(dstobject, 1);
839 				swp_pager_meta_build(dstobject, i, srcaddr);
840 				vm_object_pip_wakeup(dstobject);
841 				VM_OBJECT_LOCK(srcobject);
842 				vm_object_pip_wakeup(srcobject);
843 			}
844 		} else {
845 			/*
846 			 * Destination has valid swapblk or it is represented
847 			 * by a resident page.  We destroy the sourceblock.
848 			 */
849 
850 			swp_pager_meta_ctl(srcobject, i + offset, SWM_FREE);
851 		}
852 	}
853 
854 	/*
855 	 * Free left over swap blocks in source.
856 	 *
857 	 * We have to revert the type to OBJT_DEFAULT so we do not accidently
858 	 * double-remove the object from the swap queues.
859 	 */
860 	if (destroysource) {
861 		swp_pager_meta_free_all(srcobject);
862 		/*
863 		 * Reverting the type is not necessary, the caller is going
864 		 * to destroy srcobject directly, but I'm doing it here
865 		 * for consistency since we've removed the object from its
866 		 * queues.
867 		 */
868 		srcobject->type = OBJT_DEFAULT;
869 	}
870 	splx(s);
871 }
872 
873 /*
874  * SWAP_PAGER_HASPAGE() -	determine if we have good backing store for
875  *				the requested page.
876  *
877  *	We determine whether good backing store exists for the requested
878  *	page and return TRUE if it does, FALSE if it doesn't.
879  *
880  *	If TRUE, we also try to determine how much valid, contiguous backing
881  *	store exists before and after the requested page within a reasonable
882  *	distance.  We do not try to restrict it to the swap device stripe
883  *	(that is handled in getpages/putpages).  It probably isn't worth
884  *	doing here.
885  */
886 static boolean_t
887 swap_pager_haspage(vm_object_t object, vm_pindex_t pindex, int *before, int *after)
888 {
889 	daddr_t blk0;
890 	int s;
891 
892 	VM_OBJECT_LOCK_ASSERT(object, MA_OWNED);
893 	/*
894 	 * do we have good backing store at the requested index ?
895 	 */
896 	s = splvm();
897 	blk0 = swp_pager_meta_ctl(object, pindex, 0);
898 
899 	if (blk0 == SWAPBLK_NONE) {
900 		splx(s);
901 		if (before)
902 			*before = 0;
903 		if (after)
904 			*after = 0;
905 		return (FALSE);
906 	}
907 
908 	/*
909 	 * find backwards-looking contiguous good backing store
910 	 */
911 	if (before != NULL) {
912 		int i;
913 
914 		for (i = 1; i < (SWB_NPAGES/2); ++i) {
915 			daddr_t blk;
916 
917 			if (i > pindex)
918 				break;
919 			blk = swp_pager_meta_ctl(object, pindex - i, 0);
920 			if (blk != blk0 - i)
921 				break;
922 		}
923 		*before = (i - 1);
924 	}
925 
926 	/*
927 	 * find forward-looking contiguous good backing store
928 	 */
929 	if (after != NULL) {
930 		int i;
931 
932 		for (i = 1; i < (SWB_NPAGES/2); ++i) {
933 			daddr_t blk;
934 
935 			blk = swp_pager_meta_ctl(object, pindex + i, 0);
936 			if (blk != blk0 + i)
937 				break;
938 		}
939 		*after = (i - 1);
940 	}
941 	splx(s);
942 	return (TRUE);
943 }
944 
945 /*
946  * SWAP_PAGER_PAGE_UNSWAPPED() - remove swap backing store related to page
947  *
948  *	This removes any associated swap backing store, whether valid or
949  *	not, from the page.
950  *
951  *	This routine is typically called when a page is made dirty, at
952  *	which point any associated swap can be freed.  MADV_FREE also
953  *	calls us in a special-case situation
954  *
955  *	NOTE!!!  If the page is clean and the swap was valid, the caller
956  *	should make the page dirty before calling this routine.  This routine
957  *	does NOT change the m->dirty status of the page.  Also: MADV_FREE
958  *	depends on it.
959  *
960  *	This routine may not block
961  *	This routine must be called at splvm()
962  */
963 static void
964 swap_pager_unswapped(vm_page_t m)
965 {
966 
967 	VM_OBJECT_LOCK_ASSERT(m->object, MA_OWNED);
968 	swp_pager_meta_ctl(m->object, m->pindex, SWM_FREE);
969 }
970 
971 /*
972  * SWAP_PAGER_GETPAGES() - bring pages in from swap
973  *
974  *	Attempt to retrieve (m, count) pages from backing store, but make
975  *	sure we retrieve at least m[reqpage].  We try to load in as large
976  *	a chunk surrounding m[reqpage] as is contiguous in swap and which
977  *	belongs to the same object.
978  *
979  *	The code is designed for asynchronous operation and
980  *	immediate-notification of 'reqpage' but tends not to be
981  *	used that way.  Please do not optimize-out this algorithmic
982  *	feature, I intend to improve on it in the future.
983  *
984  *	The parent has a single vm_object_pip_add() reference prior to
985  *	calling us and we should return with the same.
986  *
987  *	The parent has BUSY'd the pages.  We should return with 'm'
988  *	left busy, but the others adjusted.
989  */
990 static int
991 swap_pager_getpages(vm_object_t object, vm_page_t *m, int count, int reqpage)
992 {
993 	struct buf *bp;
994 	vm_page_t mreq;
995 	int s;
996 	int i;
997 	int j;
998 	daddr_t blk;
999 
1000 	mreq = m[reqpage];
1001 
1002 	KASSERT(mreq->object == object,
1003 	    ("swap_pager_getpages: object mismatch %p/%p",
1004 	    object, mreq->object));
1005 
1006 	/*
1007 	 * Calculate range to retrieve.  The pages have already been assigned
1008 	 * their swapblks.  We require a *contiguous* range but we know it to
1009 	 * not span devices.   If we do not supply it, bad things
1010 	 * happen.  Note that blk, iblk & jblk can be SWAPBLK_NONE, but the
1011 	 * loops are set up such that the case(s) are handled implicitly.
1012 	 *
1013 	 * The swp_*() calls must be made at splvm().  vm_page_free() does
1014 	 * not need to be, but it will go a little faster if it is.
1015 	 */
1016 	s = splvm();
1017 	blk = swp_pager_meta_ctl(mreq->object, mreq->pindex, 0);
1018 
1019 	for (i = reqpage - 1; i >= 0; --i) {
1020 		daddr_t iblk;
1021 
1022 		iblk = swp_pager_meta_ctl(m[i]->object, m[i]->pindex, 0);
1023 		if (blk != iblk + (reqpage - i))
1024 			break;
1025 	}
1026 	++i;
1027 
1028 	for (j = reqpage + 1; j < count; ++j) {
1029 		daddr_t jblk;
1030 
1031 		jblk = swp_pager_meta_ctl(m[j]->object, m[j]->pindex, 0);
1032 		if (blk != jblk - (j - reqpage))
1033 			break;
1034 	}
1035 
1036 	/*
1037 	 * free pages outside our collection range.   Note: we never free
1038 	 * mreq, it must remain busy throughout.
1039 	 */
1040 	vm_page_lock_queues();
1041 	{
1042 		int k;
1043 
1044 		for (k = 0; k < i; ++k)
1045 			vm_page_free(m[k]);
1046 		for (k = j; k < count; ++k)
1047 			vm_page_free(m[k]);
1048 	}
1049 	vm_page_unlock_queues();
1050 	splx(s);
1051 
1052 
1053 	/*
1054 	 * Return VM_PAGER_FAIL if we have nothing to do.  Return mreq
1055 	 * still busy, but the others unbusied.
1056 	 */
1057 	if (blk == SWAPBLK_NONE)
1058 		return (VM_PAGER_FAIL);
1059 
1060 	/*
1061 	 * Getpbuf() can sleep.
1062 	 */
1063 	VM_OBJECT_UNLOCK(object);
1064 	/*
1065 	 * Get a swap buffer header to perform the IO
1066 	 */
1067 	bp = getpbuf(&nsw_rcount);
1068 	bp->b_flags |= B_PAGING;
1069 
1070 	/*
1071 	 * map our page(s) into kva for input
1072 	 */
1073 	pmap_qenter((vm_offset_t)bp->b_data, m + i, j - i);
1074 
1075 	bp->b_iocmd = BIO_READ;
1076 	bp->b_iodone = swp_pager_async_iodone;
1077 	bp->b_rcred = crhold(thread0.td_ucred);
1078 	bp->b_wcred = crhold(thread0.td_ucred);
1079 	bp->b_blkno = blk - (reqpage - i);
1080 	bp->b_bcount = PAGE_SIZE * (j - i);
1081 	bp->b_bufsize = PAGE_SIZE * (j - i);
1082 	bp->b_pager.pg_reqpage = reqpage - i;
1083 
1084 	VM_OBJECT_LOCK(object);
1085 	vm_page_lock_queues();
1086 	{
1087 		int k;
1088 
1089 		for (k = i; k < j; ++k) {
1090 			bp->b_pages[k - i] = m[k];
1091 			vm_page_flag_set(m[k], PG_SWAPINPROG);
1092 		}
1093 	}
1094 	vm_page_unlock_queues();
1095 	VM_OBJECT_UNLOCK(object);
1096 	bp->b_npages = j - i;
1097 
1098 	cnt.v_swapin++;
1099 	cnt.v_swappgsin += bp->b_npages;
1100 
1101 	/*
1102 	 * We still hold the lock on mreq, and our automatic completion routine
1103 	 * does not remove it.
1104 	 */
1105 	VM_OBJECT_LOCK(mreq->object);
1106 	vm_object_pip_add(mreq->object, bp->b_npages);
1107 	VM_OBJECT_UNLOCK(mreq->object);
1108 
1109 	/*
1110 	 * perform the I/O.  NOTE!!!  bp cannot be considered valid after
1111 	 * this point because we automatically release it on completion.
1112 	 * Instead, we look at the one page we are interested in which we
1113 	 * still hold a lock on even through the I/O completion.
1114 	 *
1115 	 * The other pages in our m[] array are also released on completion,
1116 	 * so we cannot assume they are valid anymore either.
1117 	 *
1118 	 * NOTE: b_blkno is destroyed by the call to swapdev_strategy
1119 	 */
1120 	BUF_KERNPROC(bp);
1121 	swp_pager_strategy(bp);
1122 
1123 	/*
1124 	 * wait for the page we want to complete.  PG_SWAPINPROG is always
1125 	 * cleared on completion.  If an I/O error occurs, SWAPBLK_NONE
1126 	 * is set in the meta-data.
1127 	 */
1128 	s = splvm();
1129 	vm_page_lock_queues();
1130 	while ((mreq->flags & PG_SWAPINPROG) != 0) {
1131 		vm_page_flag_set(mreq, PG_WANTED | PG_REFERENCED);
1132 		cnt.v_intrans++;
1133 		if (msleep(mreq, &vm_page_queue_mtx, PSWP, "swread", hz*20)) {
1134 			printf(
1135 			    "swap_pager: indefinite wait buffer: device:"
1136 				" %s, blkno: %ld, size: %ld\n",
1137 			    devtoname(bp->b_dev), (long)bp->b_blkno,
1138 			    bp->b_bcount
1139 			);
1140 		}
1141 	}
1142 	vm_page_unlock_queues();
1143 	splx(s);
1144 
1145 	VM_OBJECT_LOCK(mreq->object);
1146 	/*
1147 	 * mreq is left busied after completion, but all the other pages
1148 	 * are freed.  If we had an unrecoverable read error the page will
1149 	 * not be valid.
1150 	 */
1151 	if (mreq->valid != VM_PAGE_BITS_ALL) {
1152 		return (VM_PAGER_ERROR);
1153 	} else {
1154 		return (VM_PAGER_OK);
1155 	}
1156 
1157 	/*
1158 	 * A final note: in a low swap situation, we cannot deallocate swap
1159 	 * and mark a page dirty here because the caller is likely to mark
1160 	 * the page clean when we return, causing the page to possibly revert
1161 	 * to all-zero's later.
1162 	 */
1163 }
1164 
1165 /*
1166  *	swap_pager_putpages:
1167  *
1168  *	Assign swap (if necessary) and initiate I/O on the specified pages.
1169  *
1170  *	We support both OBJT_DEFAULT and OBJT_SWAP objects.  DEFAULT objects
1171  *	are automatically converted to SWAP objects.
1172  *
1173  *	In a low memory situation we may block in VOP_STRATEGY(), but the new
1174  *	vm_page reservation system coupled with properly written VFS devices
1175  *	should ensure that no low-memory deadlock occurs.  This is an area
1176  *	which needs work.
1177  *
1178  *	The parent has N vm_object_pip_add() references prior to
1179  *	calling us and will remove references for rtvals[] that are
1180  *	not set to VM_PAGER_PEND.  We need to remove the rest on I/O
1181  *	completion.
1182  *
1183  *	The parent has soft-busy'd the pages it passes us and will unbusy
1184  *	those whos rtvals[] entry is not set to VM_PAGER_PEND on return.
1185  *	We need to unbusy the rest on I/O completion.
1186  */
1187 void
1188 swap_pager_putpages(vm_object_t object, vm_page_t *m, int count,
1189     boolean_t sync, int *rtvals)
1190 {
1191 	int i;
1192 	int n = 0;
1193 
1194 	GIANT_REQUIRED;
1195 	if (count && m[0]->object != object) {
1196 		panic("swap_pager_getpages: object mismatch %p/%p",
1197 		    object,
1198 		    m[0]->object
1199 		);
1200 	}
1201 
1202 	/*
1203 	 * Step 1
1204 	 *
1205 	 * Turn object into OBJT_SWAP
1206 	 * check for bogus sysops
1207 	 * force sync if not pageout process
1208 	 */
1209 	if (object->type != OBJT_SWAP)
1210 		swp_pager_meta_build(object, 0, SWAPBLK_NONE);
1211 	VM_OBJECT_UNLOCK(object);
1212 
1213 	if (curproc != pageproc)
1214 		sync = TRUE;
1215 
1216 	/*
1217 	 * Step 2
1218 	 *
1219 	 * Update nsw parameters from swap_async_max sysctl values.
1220 	 * Do not let the sysop crash the machine with bogus numbers.
1221 	 */
1222 	mtx_lock(&pbuf_mtx);
1223 	if (swap_async_max != nsw_wcount_async_max) {
1224 		int n;
1225 		int s;
1226 
1227 		/*
1228 		 * limit range
1229 		 */
1230 		if ((n = swap_async_max) > nswbuf / 2)
1231 			n = nswbuf / 2;
1232 		if (n < 1)
1233 			n = 1;
1234 		swap_async_max = n;
1235 
1236 		/*
1237 		 * Adjust difference ( if possible ).  If the current async
1238 		 * count is too low, we may not be able to make the adjustment
1239 		 * at this time.
1240 		 */
1241 		s = splvm();
1242 		n -= nsw_wcount_async_max;
1243 		if (nsw_wcount_async + n >= 0) {
1244 			nsw_wcount_async += n;
1245 			nsw_wcount_async_max += n;
1246 			wakeup(&nsw_wcount_async);
1247 		}
1248 		splx(s);
1249 	}
1250 	mtx_unlock(&pbuf_mtx);
1251 
1252 	/*
1253 	 * Step 3
1254 	 *
1255 	 * Assign swap blocks and issue I/O.  We reallocate swap on the fly.
1256 	 * The page is left dirty until the pageout operation completes
1257 	 * successfully.
1258 	 */
1259 	for (i = 0; i < count; i += n) {
1260 		int s;
1261 		int j;
1262 		struct buf *bp;
1263 		daddr_t blk;
1264 
1265 		/*
1266 		 * Maximum I/O size is limited by a number of factors.
1267 		 */
1268 		n = min(BLIST_MAX_ALLOC, count - i);
1269 		n = min(n, nsw_cluster_max);
1270 
1271 		s = splvm();
1272 
1273 		/*
1274 		 * Get biggest block of swap we can.  If we fail, fall
1275 		 * back and try to allocate a smaller block.  Don't go
1276 		 * overboard trying to allocate space if it would overly
1277 		 * fragment swap.
1278 		 */
1279 		while (
1280 		    (blk = swp_pager_getswapspace(n)) == SWAPBLK_NONE &&
1281 		    n > 4
1282 		) {
1283 			n >>= 1;
1284 		}
1285 		if (blk == SWAPBLK_NONE) {
1286 			for (j = 0; j < n; ++j)
1287 				rtvals[i+j] = VM_PAGER_FAIL;
1288 			splx(s);
1289 			continue;
1290 		}
1291 
1292 		/*
1293 		 * All I/O parameters have been satisfied, build the I/O
1294 		 * request and assign the swap space.
1295 		 */
1296 		if (sync == TRUE) {
1297 			bp = getpbuf(&nsw_wcount_sync);
1298 		} else {
1299 			bp = getpbuf(&nsw_wcount_async);
1300 			bp->b_flags = B_ASYNC;
1301 		}
1302 		bp->b_flags |= B_PAGING;
1303 		bp->b_iocmd = BIO_WRITE;
1304 
1305 		pmap_qenter((vm_offset_t)bp->b_data, &m[i], n);
1306 
1307 		bp->b_rcred = crhold(thread0.td_ucred);
1308 		bp->b_wcred = crhold(thread0.td_ucred);
1309 		bp->b_bcount = PAGE_SIZE * n;
1310 		bp->b_bufsize = PAGE_SIZE * n;
1311 		bp->b_blkno = blk;
1312 
1313 		VM_OBJECT_LOCK(object);
1314 		for (j = 0; j < n; ++j) {
1315 			vm_page_t mreq = m[i+j];
1316 
1317 			swp_pager_meta_build(
1318 			    mreq->object,
1319 			    mreq->pindex,
1320 			    blk + j
1321 			);
1322 			vm_page_dirty(mreq);
1323 			rtvals[i+j] = VM_PAGER_OK;
1324 
1325 			vm_page_lock_queues();
1326 			vm_page_flag_set(mreq, PG_SWAPINPROG);
1327 			vm_page_unlock_queues();
1328 			bp->b_pages[j] = mreq;
1329 		}
1330 		VM_OBJECT_UNLOCK(object);
1331 		bp->b_npages = n;
1332 		/*
1333 		 * Must set dirty range for NFS to work.
1334 		 */
1335 		bp->b_dirtyoff = 0;
1336 		bp->b_dirtyend = bp->b_bcount;
1337 
1338 		cnt.v_swapout++;
1339 		cnt.v_swappgsout += bp->b_npages;
1340 
1341 		splx(s);
1342 
1343 		/*
1344 		 * asynchronous
1345 		 *
1346 		 * NOTE: b_blkno is destroyed by the call to swapdev_strategy
1347 		 */
1348 		if (sync == FALSE) {
1349 			bp->b_iodone = swp_pager_async_iodone;
1350 			BUF_KERNPROC(bp);
1351 			swp_pager_strategy(bp);
1352 
1353 			for (j = 0; j < n; ++j)
1354 				rtvals[i+j] = VM_PAGER_PEND;
1355 			/* restart outter loop */
1356 			continue;
1357 		}
1358 
1359 		/*
1360 		 * synchronous
1361 		 *
1362 		 * NOTE: b_blkno is destroyed by the call to swapdev_strategy
1363 		 */
1364 		bp->b_iodone = bdone;
1365 		swp_pager_strategy(bp);
1366 
1367 		/*
1368 		 * Wait for the sync I/O to complete, then update rtvals.
1369 		 * We just set the rtvals[] to VM_PAGER_PEND so we can call
1370 		 * our async completion routine at the end, thus avoiding a
1371 		 * double-free.
1372 		 */
1373 		s = splbio();
1374 		bwait(bp, PVM, "swwrt");
1375 		for (j = 0; j < n; ++j)
1376 			rtvals[i+j] = VM_PAGER_PEND;
1377 		/*
1378 		 * Now that we are through with the bp, we can call the
1379 		 * normal async completion, which frees everything up.
1380 		 */
1381 		swp_pager_async_iodone(bp);
1382 		splx(s);
1383 	}
1384 	VM_OBJECT_LOCK(object);
1385 }
1386 
1387 /*
1388  *	swp_pager_async_iodone:
1389  *
1390  *	Completion routine for asynchronous reads and writes from/to swap.
1391  *	Also called manually by synchronous code to finish up a bp.
1392  *
1393  *	For READ operations, the pages are PG_BUSY'd.  For WRITE operations,
1394  *	the pages are vm_page_t->busy'd.  For READ operations, we PG_BUSY
1395  *	unbusy all pages except the 'main' request page.  For WRITE
1396  *	operations, we vm_page_t->busy'd unbusy all pages ( we can do this
1397  *	because we marked them all VM_PAGER_PEND on return from putpages ).
1398  *
1399  *	This routine may not block.
1400  *	This routine is called at splbio() or better
1401  *
1402  *	We up ourselves to splvm() as required for various vm_page related
1403  *	calls.
1404  */
1405 static void
1406 swp_pager_async_iodone(struct buf *bp)
1407 {
1408 	int s;
1409 	int i;
1410 	vm_object_t object = NULL;
1411 
1412 	bp->b_flags |= B_DONE;
1413 
1414 	/*
1415 	 * report error
1416 	 */
1417 	if (bp->b_ioflags & BIO_ERROR) {
1418 		printf(
1419 		    "swap_pager: I/O error - %s failed; blkno %ld,"
1420 			"size %ld, error %d\n",
1421 		    ((bp->b_iocmd == BIO_READ) ? "pagein" : "pageout"),
1422 		    (long)bp->b_blkno,
1423 		    (long)bp->b_bcount,
1424 		    bp->b_error
1425 		);
1426 	}
1427 
1428 	/*
1429 	 * set object, raise to splvm().
1430 	 */
1431 	s = splvm();
1432 
1433 	/*
1434 	 * remove the mapping for kernel virtual
1435 	 */
1436 	pmap_qremove((vm_offset_t)bp->b_data, bp->b_npages);
1437 
1438 	if (bp->b_npages) {
1439 		object = bp->b_pages[0]->object;
1440 		VM_OBJECT_LOCK(object);
1441 	}
1442 	vm_page_lock_queues();
1443 	/*
1444 	 * cleanup pages.  If an error occurs writing to swap, we are in
1445 	 * very serious trouble.  If it happens to be a disk error, though,
1446 	 * we may be able to recover by reassigning the swap later on.  So
1447 	 * in this case we remove the m->swapblk assignment for the page
1448 	 * but do not free it in the rlist.  The errornous block(s) are thus
1449 	 * never reallocated as swap.  Redirty the page and continue.
1450 	 */
1451 	for (i = 0; i < bp->b_npages; ++i) {
1452 		vm_page_t m = bp->b_pages[i];
1453 
1454 		vm_page_flag_clear(m, PG_SWAPINPROG);
1455 
1456 		if (bp->b_ioflags & BIO_ERROR) {
1457 			/*
1458 			 * If an error occurs I'd love to throw the swapblk
1459 			 * away without freeing it back to swapspace, so it
1460 			 * can never be used again.  But I can't from an
1461 			 * interrupt.
1462 			 */
1463 			if (bp->b_iocmd == BIO_READ) {
1464 				/*
1465 				 * When reading, reqpage needs to stay
1466 				 * locked for the parent, but all other
1467 				 * pages can be freed.  We still want to
1468 				 * wakeup the parent waiting on the page,
1469 				 * though.  ( also: pg_reqpage can be -1 and
1470 				 * not match anything ).
1471 				 *
1472 				 * We have to wake specifically requested pages
1473 				 * up too because we cleared PG_SWAPINPROG and
1474 				 * someone may be waiting for that.
1475 				 *
1476 				 * NOTE: for reads, m->dirty will probably
1477 				 * be overridden by the original caller of
1478 				 * getpages so don't play cute tricks here.
1479 				 *
1480 				 * XXX IT IS NOT LEGAL TO FREE THE PAGE HERE
1481 				 * AS THIS MESSES WITH object->memq, and it is
1482 				 * not legal to mess with object->memq from an
1483 				 * interrupt.
1484 				 */
1485 				m->valid = 0;
1486 				if (i != bp->b_pager.pg_reqpage)
1487 					vm_page_free(m);
1488 				else
1489 					vm_page_flash(m);
1490 				/*
1491 				 * If i == bp->b_pager.pg_reqpage, do not wake
1492 				 * the page up.  The caller needs to.
1493 				 */
1494 			} else {
1495 				/*
1496 				 * If a write error occurs, reactivate page
1497 				 * so it doesn't clog the inactive list,
1498 				 * then finish the I/O.
1499 				 */
1500 				vm_page_dirty(m);
1501 				vm_page_activate(m);
1502 				vm_page_io_finish(m);
1503 			}
1504 		} else if (bp->b_iocmd == BIO_READ) {
1505 			/*
1506 			 * For read success, clear dirty bits.  Nobody should
1507 			 * have this page mapped but don't take any chances,
1508 			 * make sure the pmap modify bits are also cleared.
1509 			 *
1510 			 * NOTE: for reads, m->dirty will probably be
1511 			 * overridden by the original caller of getpages so
1512 			 * we cannot set them in order to free the underlying
1513 			 * swap in a low-swap situation.  I don't think we'd
1514 			 * want to do that anyway, but it was an optimization
1515 			 * that existed in the old swapper for a time before
1516 			 * it got ripped out due to precisely this problem.
1517 			 *
1518 			 * If not the requested page then deactivate it.
1519 			 *
1520 			 * Note that the requested page, reqpage, is left
1521 			 * busied, but we still have to wake it up.  The
1522 			 * other pages are released (unbusied) by
1523 			 * vm_page_wakeup().  We do not set reqpage's
1524 			 * valid bits here, it is up to the caller.
1525 			 */
1526 			pmap_clear_modify(m);
1527 			m->valid = VM_PAGE_BITS_ALL;
1528 			vm_page_undirty(m);
1529 
1530 			/*
1531 			 * We have to wake specifically requested pages
1532 			 * up too because we cleared PG_SWAPINPROG and
1533 			 * could be waiting for it in getpages.  However,
1534 			 * be sure to not unbusy getpages specifically
1535 			 * requested page - getpages expects it to be
1536 			 * left busy.
1537 			 */
1538 			if (i != bp->b_pager.pg_reqpage) {
1539 				vm_page_deactivate(m);
1540 				vm_page_wakeup(m);
1541 			} else {
1542 				vm_page_flash(m);
1543 			}
1544 		} else {
1545 			/*
1546 			 * For write success, clear the modify and dirty
1547 			 * status, then finish the I/O ( which decrements the
1548 			 * busy count and possibly wakes waiter's up ).
1549 			 */
1550 			pmap_clear_modify(m);
1551 			vm_page_undirty(m);
1552 			vm_page_io_finish(m);
1553 			if (vm_page_count_severe())
1554 				vm_page_try_to_cache(m);
1555 		}
1556 	}
1557 	vm_page_unlock_queues();
1558 
1559 	/*
1560 	 * adjust pip.  NOTE: the original parent may still have its own
1561 	 * pip refs on the object.
1562 	 */
1563 	if (object != NULL) {
1564 		vm_object_pip_wakeupn(object, bp->b_npages);
1565 		VM_OBJECT_UNLOCK(object);
1566 	}
1567 
1568 	/*
1569 	 * release the physical I/O buffer
1570 	 */
1571 	relpbuf(
1572 	    bp,
1573 	    ((bp->b_iocmd == BIO_READ) ? &nsw_rcount :
1574 		((bp->b_flags & B_ASYNC) ?
1575 		    &nsw_wcount_async :
1576 		    &nsw_wcount_sync
1577 		)
1578 	    )
1579 	);
1580 	splx(s);
1581 }
1582 
1583 /*
1584  *	swap_pager_isswapped:
1585  *
1586  *	Return 1 if at least one page in the given object is paged
1587  *	out to the given swap device.
1588  *
1589  *	This routine may not block.
1590  */
1591 int
1592 swap_pager_isswapped(vm_object_t object, struct swdevt *sp)
1593 {
1594 	daddr_t index = 0;
1595 	int bcount;
1596 	int i;
1597 
1598 	VM_OBJECT_LOCK_ASSERT(object, MA_OWNED);
1599 	for (bcount = 0; bcount < object->un_pager.swp.swp_bcount; bcount++) {
1600 		struct swblock *swap;
1601 
1602 		mtx_lock(&swhash_mtx);
1603 		if ((swap = *swp_pager_hash(object, index)) != NULL) {
1604 			for (i = 0; i < SWAP_META_PAGES; ++i) {
1605 				daddr_t v = swap->swb_pages[i];
1606 				if (v == SWAPBLK_NONE)
1607 					continue;
1608 				if (swp_pager_find_dev(v) == sp) {
1609 					mtx_unlock(&swhash_mtx);
1610 					return 1;
1611 				}
1612 			}
1613 		}
1614 		mtx_unlock(&swhash_mtx);
1615 		index += SWAP_META_PAGES;
1616 		if (index > 0x20000000)
1617 			panic("swap_pager_isswapped: failed to locate all swap meta blocks");
1618 	}
1619 	return 0;
1620 }
1621 
1622 /*
1623  * SWP_PAGER_FORCE_PAGEIN() - force a swap block to be paged in
1624  *
1625  *	This routine dissociates the page at the given index within a
1626  *	swap block from its backing store, paging it in if necessary.
1627  *	If the page is paged in, it is placed in the inactive queue,
1628  *	since it had its backing store ripped out from under it.
1629  *	We also attempt to swap in all other pages in the swap block,
1630  *	we only guarantee that the one at the specified index is
1631  *	paged in.
1632  *
1633  *	XXX - The code to page the whole block in doesn't work, so we
1634  *	      revert to the one-by-one behavior for now.  Sigh.
1635  */
1636 static __inline void
1637 swp_pager_force_pagein(struct swblock *swap, int idx)
1638 {
1639 	vm_object_t object;
1640 	vm_page_t m;
1641 	vm_pindex_t pindex;
1642 
1643 	object = swap->swb_object;
1644 	pindex = swap->swb_index;
1645 	mtx_unlock(&swhash_mtx);
1646 
1647 	VM_OBJECT_LOCK(object);
1648 	vm_object_pip_add(object, 1);
1649 	m = vm_page_grab(object, pindex + idx, VM_ALLOC_NORMAL|VM_ALLOC_RETRY);
1650 	if (m->valid == VM_PAGE_BITS_ALL) {
1651 		vm_object_pip_subtract(object, 1);
1652 		vm_page_lock_queues();
1653 		vm_page_activate(m);
1654 		vm_page_dirty(m);
1655 		vm_page_wakeup(m);
1656 		vm_page_unlock_queues();
1657 		vm_pager_page_unswapped(m);
1658 		VM_OBJECT_UNLOCK(object);
1659 		return;
1660 	}
1661 
1662 	if (swap_pager_getpages(object, &m, 1, 0) !=
1663 	    VM_PAGER_OK)
1664 		panic("swap_pager_force_pagein: read from swap failed");/*XXX*/
1665 	vm_object_pip_subtract(object, 1);
1666 	vm_page_lock_queues();
1667 	vm_page_dirty(m);
1668 	vm_page_dontneed(m);
1669 	vm_page_wakeup(m);
1670 	vm_page_unlock_queues();
1671 	vm_pager_page_unswapped(m);
1672 	VM_OBJECT_UNLOCK(object);
1673 }
1674 
1675 
1676 /*
1677  *	swap_pager_swapoff:
1678  *
1679  *	Page in all of the pages that have been paged out to the
1680  *	given device.  The corresponding blocks in the bitmap must be
1681  *	marked as allocated and the device must be flagged SW_CLOSING.
1682  *	There may be no processes swapped out to the device.
1683  *
1684  *	The sw_used parameter points to the field in the swdev structure
1685  *	that contains a count of the number of blocks still allocated
1686  *	on the device.  If we encounter objects with a nonzero pip count
1687  *	in our scan, we use this number to determine if we're really done.
1688  *
1689  *	This routine may block.
1690  */
1691 static void
1692 swap_pager_swapoff(struct swdevt *sp, int *sw_used)
1693 {
1694 	struct swblock **pswap;
1695 	struct swblock *swap;
1696 	vm_object_t waitobj;
1697 	daddr_t v;
1698 	int i, j;
1699 
1700 	GIANT_REQUIRED;
1701 
1702 full_rescan:
1703 	waitobj = NULL;
1704 	for (i = 0; i <= swhash_mask; i++) { /* '<=' is correct here */
1705 restart:
1706 		pswap = &swhash[i];
1707 		mtx_lock(&swhash_mtx);
1708 		while ((swap = *pswap) != NULL) {
1709                         for (j = 0; j < SWAP_META_PAGES; ++j) {
1710                                 v = swap->swb_pages[j];
1711                                 if (v != SWAPBLK_NONE &&
1712 				    swp_pager_find_dev(v) == sp)
1713                                         break;
1714                         }
1715 			if (j < SWAP_META_PAGES) {
1716 				swp_pager_force_pagein(swap, j);
1717 				goto restart;
1718 			} else if (swap->swb_object->paging_in_progress) {
1719 				if (!waitobj)
1720 					waitobj = swap->swb_object;
1721 			}
1722 			pswap = &swap->swb_hnext;
1723 		}
1724 		mtx_unlock(&swhash_mtx);
1725 	}
1726 	if (waitobj && *sw_used) {
1727 	    /*
1728 	     * We wait on an arbitrary object to clock our rescans
1729 	     * to the rate of paging completion.
1730 	     */
1731 	    VM_OBJECT_LOCK(waitobj);
1732 	    vm_object_pip_wait(waitobj, "swpoff");
1733 	    VM_OBJECT_UNLOCK(waitobj);
1734 	    goto full_rescan;
1735 	}
1736 	if (*sw_used)
1737 	    panic("swapoff: failed to locate %d swap blocks", *sw_used);
1738 }
1739 
1740 /************************************************************************
1741  *				SWAP META DATA 				*
1742  ************************************************************************
1743  *
1744  *	These routines manipulate the swap metadata stored in the
1745  *	OBJT_SWAP object.  All swp_*() routines must be called at
1746  *	splvm() because swap can be freed up by the low level vm_page
1747  *	code which might be called from interrupts beyond what splbio() covers.
1748  *
1749  *	Swap metadata is implemented with a global hash and not directly
1750  *	linked into the object.  Instead the object simply contains
1751  *	appropriate tracking counters.
1752  */
1753 
1754 /*
1755  * SWP_PAGER_META_BUILD() -	add swap block to swap meta data for object
1756  *
1757  *	We first convert the object to a swap object if it is a default
1758  *	object.
1759  *
1760  *	The specified swapblk is added to the object's swap metadata.  If
1761  *	the swapblk is not valid, it is freed instead.  Any previously
1762  *	assigned swapblk is freed.
1763  *
1764  *	This routine must be called at splvm(), except when used to convert
1765  *	an OBJT_DEFAULT object into an OBJT_SWAP object.
1766  */
1767 static void
1768 swp_pager_meta_build(vm_object_t object, vm_pindex_t pindex, daddr_t swapblk)
1769 {
1770 	struct swblock *swap;
1771 	struct swblock **pswap;
1772 	int idx;
1773 
1774 	VM_OBJECT_LOCK_ASSERT(object, MA_OWNED);
1775 	/*
1776 	 * Convert default object to swap object if necessary
1777 	 */
1778 	if (object->type != OBJT_SWAP) {
1779 		object->type = OBJT_SWAP;
1780 		object->un_pager.swp.swp_bcount = 0;
1781 
1782 		if (object->handle != NULL) {
1783 			mtx_lock(&sw_alloc_mtx);
1784 			TAILQ_INSERT_TAIL(
1785 			    NOBJLIST(object->handle),
1786 			    object,
1787 			    pager_object_list
1788 			);
1789 			mtx_unlock(&sw_alloc_mtx);
1790 		}
1791 	}
1792 
1793 	/*
1794 	 * Locate hash entry.  If not found create, but if we aren't adding
1795 	 * anything just return.  If we run out of space in the map we wait
1796 	 * and, since the hash table may have changed, retry.
1797 	 */
1798 retry:
1799 	mtx_lock(&swhash_mtx);
1800 	pswap = swp_pager_hash(object, pindex);
1801 
1802 	if ((swap = *pswap) == NULL) {
1803 		int i;
1804 
1805 		if (swapblk == SWAPBLK_NONE)
1806 			goto done;
1807 
1808 		swap = *pswap = uma_zalloc(swap_zone, M_NOWAIT);
1809 		if (swap == NULL) {
1810 			mtx_unlock(&swhash_mtx);
1811 			VM_OBJECT_UNLOCK(object);
1812 			VM_WAIT;
1813 			VM_OBJECT_LOCK(object);
1814 			goto retry;
1815 		}
1816 
1817 		swap->swb_hnext = NULL;
1818 		swap->swb_object = object;
1819 		swap->swb_index = pindex & ~(vm_pindex_t)SWAP_META_MASK;
1820 		swap->swb_count = 0;
1821 
1822 		++object->un_pager.swp.swp_bcount;
1823 
1824 		for (i = 0; i < SWAP_META_PAGES; ++i)
1825 			swap->swb_pages[i] = SWAPBLK_NONE;
1826 	}
1827 
1828 	/*
1829 	 * Delete prior contents of metadata
1830 	 */
1831 	idx = pindex & SWAP_META_MASK;
1832 
1833 	if (swap->swb_pages[idx] != SWAPBLK_NONE) {
1834 		swp_pager_freeswapspace(swap->swb_pages[idx], 1);
1835 		--swap->swb_count;
1836 	}
1837 
1838 	/*
1839 	 * Enter block into metadata
1840 	 */
1841 	swap->swb_pages[idx] = swapblk;
1842 	if (swapblk != SWAPBLK_NONE)
1843 		++swap->swb_count;
1844 done:
1845 	mtx_unlock(&swhash_mtx);
1846 }
1847 
1848 /*
1849  * SWP_PAGER_META_FREE() - free a range of blocks in the object's swap metadata
1850  *
1851  *	The requested range of blocks is freed, with any associated swap
1852  *	returned to the swap bitmap.
1853  *
1854  *	This routine will free swap metadata structures as they are cleaned
1855  *	out.  This routine does *NOT* operate on swap metadata associated
1856  *	with resident pages.
1857  *
1858  *	This routine must be called at splvm()
1859  */
1860 static void
1861 swp_pager_meta_free(vm_object_t object, vm_pindex_t index, daddr_t count)
1862 {
1863 
1864 	VM_OBJECT_LOCK_ASSERT(object, MA_OWNED);
1865 	if (object->type != OBJT_SWAP)
1866 		return;
1867 
1868 	while (count > 0) {
1869 		struct swblock **pswap;
1870 		struct swblock *swap;
1871 
1872 		mtx_lock(&swhash_mtx);
1873 		pswap = swp_pager_hash(object, index);
1874 
1875 		if ((swap = *pswap) != NULL) {
1876 			daddr_t v = swap->swb_pages[index & SWAP_META_MASK];
1877 
1878 			if (v != SWAPBLK_NONE) {
1879 				swp_pager_freeswapspace(v, 1);
1880 				swap->swb_pages[index & SWAP_META_MASK] =
1881 					SWAPBLK_NONE;
1882 				if (--swap->swb_count == 0) {
1883 					*pswap = swap->swb_hnext;
1884 					uma_zfree(swap_zone, swap);
1885 					--object->un_pager.swp.swp_bcount;
1886 				}
1887 			}
1888 			--count;
1889 			++index;
1890 		} else {
1891 			int n = SWAP_META_PAGES - (index & SWAP_META_MASK);
1892 			count -= n;
1893 			index += n;
1894 		}
1895 		mtx_unlock(&swhash_mtx);
1896 	}
1897 }
1898 
1899 /*
1900  * SWP_PAGER_META_FREE_ALL() - destroy all swap metadata associated with object
1901  *
1902  *	This routine locates and destroys all swap metadata associated with
1903  *	an object.
1904  *
1905  *	This routine must be called at splvm()
1906  */
1907 static void
1908 swp_pager_meta_free_all(vm_object_t object)
1909 {
1910 	daddr_t index = 0;
1911 
1912 	VM_OBJECT_LOCK_ASSERT(object, MA_OWNED);
1913 	if (object->type != OBJT_SWAP)
1914 		return;
1915 
1916 	while (object->un_pager.swp.swp_bcount) {
1917 		struct swblock **pswap;
1918 		struct swblock *swap;
1919 
1920 		mtx_lock(&swhash_mtx);
1921 		pswap = swp_pager_hash(object, index);
1922 		if ((swap = *pswap) != NULL) {
1923 			int i;
1924 
1925 			for (i = 0; i < SWAP_META_PAGES; ++i) {
1926 				daddr_t v = swap->swb_pages[i];
1927 				if (v != SWAPBLK_NONE) {
1928 					--swap->swb_count;
1929 					swp_pager_freeswapspace(v, 1);
1930 				}
1931 			}
1932 			if (swap->swb_count != 0)
1933 				panic("swap_pager_meta_free_all: swb_count != 0");
1934 			*pswap = swap->swb_hnext;
1935 			uma_zfree(swap_zone, swap);
1936 			--object->un_pager.swp.swp_bcount;
1937 		}
1938 		mtx_unlock(&swhash_mtx);
1939 		index += SWAP_META_PAGES;
1940 		if (index > 0x20000000)
1941 			panic("swp_pager_meta_free_all: failed to locate all swap meta blocks");
1942 	}
1943 }
1944 
1945 /*
1946  * SWP_PAGER_METACTL() -  misc control of swap and vm_page_t meta data.
1947  *
1948  *	This routine is capable of looking up, popping, or freeing
1949  *	swapblk assignments in the swap meta data or in the vm_page_t.
1950  *	The routine typically returns the swapblk being looked-up, or popped,
1951  *	or SWAPBLK_NONE if the block was freed, or SWAPBLK_NONE if the block
1952  *	was invalid.  This routine will automatically free any invalid
1953  *	meta-data swapblks.
1954  *
1955  *	It is not possible to store invalid swapblks in the swap meta data
1956  *	(other then a literal 'SWAPBLK_NONE'), so we don't bother checking.
1957  *
1958  *	When acting on a busy resident page and paging is in progress, we
1959  *	have to wait until paging is complete but otherwise can act on the
1960  *	busy page.
1961  *
1962  *	This routine must be called at splvm().
1963  *
1964  *	SWM_FREE	remove and free swap block from metadata
1965  *	SWM_POP		remove from meta data but do not free.. pop it out
1966  */
1967 static daddr_t
1968 swp_pager_meta_ctl(vm_object_t object, vm_pindex_t pindex, int flags)
1969 {
1970 	struct swblock **pswap;
1971 	struct swblock *swap;
1972 	daddr_t r1;
1973 	int idx;
1974 
1975 	VM_OBJECT_LOCK_ASSERT(object, MA_OWNED);
1976 	/*
1977 	 * The meta data only exists of the object is OBJT_SWAP
1978 	 * and even then might not be allocated yet.
1979 	 */
1980 	if (object->type != OBJT_SWAP)
1981 		return (SWAPBLK_NONE);
1982 
1983 	r1 = SWAPBLK_NONE;
1984 	mtx_lock(&swhash_mtx);
1985 	pswap = swp_pager_hash(object, pindex);
1986 
1987 	if ((swap = *pswap) != NULL) {
1988 		idx = pindex & SWAP_META_MASK;
1989 		r1 = swap->swb_pages[idx];
1990 
1991 		if (r1 != SWAPBLK_NONE) {
1992 			if (flags & SWM_FREE) {
1993 				swp_pager_freeswapspace(r1, 1);
1994 				r1 = SWAPBLK_NONE;
1995 			}
1996 			if (flags & (SWM_FREE|SWM_POP)) {
1997 				swap->swb_pages[idx] = SWAPBLK_NONE;
1998 				if (--swap->swb_count == 0) {
1999 					*pswap = swap->swb_hnext;
2000 					uma_zfree(swap_zone, swap);
2001 					--object->un_pager.swp.swp_bcount;
2002 				}
2003 			}
2004 		}
2005 	}
2006 	mtx_unlock(&swhash_mtx);
2007 	return (r1);
2008 }
2009 
2010 /*
2011  * System call swapon(name) enables swapping on device name,
2012  * which must be in the swdevsw.  Return EBUSY
2013  * if already swapping on this device.
2014  */
2015 #ifndef _SYS_SYSPROTO_H_
2016 struct swapon_args {
2017 	char *name;
2018 };
2019 #endif
2020 
2021 /*
2022  * MPSAFE
2023  */
2024 /* ARGSUSED */
2025 int
2026 swapon(struct thread *td, struct swapon_args *uap)
2027 {
2028 	struct vattr attr;
2029 	struct vnode *vp;
2030 	struct nameidata nd;
2031 	int error;
2032 
2033 	mtx_lock(&Giant);
2034 	error = suser(td);
2035 	if (error)
2036 		goto done2;
2037 
2038 	while (swdev_syscall_active)
2039 	    tsleep(&swdev_syscall_active, PUSER - 1, "swpon", 0);
2040 	swdev_syscall_active = 1;
2041 
2042 	/*
2043 	 * Swap metadata may not fit in the KVM if we have physical
2044 	 * memory of >1GB.
2045 	 */
2046 	if (swap_zone == NULL) {
2047 		error = ENOMEM;
2048 		goto done;
2049 	}
2050 
2051 	NDINIT(&nd, LOOKUP, FOLLOW, UIO_USERSPACE, uap->name, td);
2052 	error = namei(&nd);
2053 	if (error)
2054 		goto done;
2055 
2056 	NDFREE(&nd, NDF_ONLY_PNBUF);
2057 	vp = nd.ni_vp;
2058 
2059 	if (vn_isdisk(vp, &error)) {
2060 		error = swapongeom(td, vp);
2061 	} else if (vp->v_type == VREG &&
2062 	    (vp->v_mount->mnt_vfc->vfc_flags & VFCF_NETWORK) != 0 &&
2063 	    (error = VOP_GETATTR(vp, &attr, td->td_ucred, td)) == 0) {
2064 		/*
2065 		 * Allow direct swapping to NFS regular files in the same
2066 		 * way that nfs_mountroot() sets up diskless swapping.
2067 		 */
2068 		error = swaponvp(td, vp, attr.va_size / DEV_BSIZE);
2069 	}
2070 
2071 	if (error)
2072 		vrele(vp);
2073 done:
2074 	swdev_syscall_active = 0;
2075 	wakeup_one(&swdev_syscall_active);
2076 done2:
2077 	mtx_unlock(&Giant);
2078 	return (error);
2079 }
2080 
2081 static void
2082 swaponsomething(struct vnode *vp, void *id, u_long nblks, sw_strategy_t *strategy, sw_close_t *close, udev_t udev)
2083 {
2084 	struct swdevt *sp, *tsp;
2085 	swblk_t dvbase;
2086 	u_long mblocks;
2087 
2088 	/*
2089 	 * If we go beyond this, we get overflows in the radix
2090 	 * tree bitmap code.
2091 	 */
2092 	mblocks = 0x40000000 / BLIST_META_RADIX;
2093 	if (nblks > mblocks) {
2094 		printf("WARNING: reducing size to maximum of %lu blocks per swap unit\n",
2095 			mblocks);
2096 		nblks = mblocks;
2097 	}
2098 	/*
2099 	 * nblks is in DEV_BSIZE'd chunks, convert to PAGE_SIZE'd chunks.
2100 	 * First chop nblks off to page-align it, then convert.
2101 	 *
2102 	 * sw->sw_nblks is in page-sized chunks now too.
2103 	 */
2104 	nblks &= ~(ctodb(1) - 1);
2105 	nblks = dbtoc(nblks);
2106 
2107 	sp = malloc(sizeof *sp, M_VMPGDATA, M_WAITOK | M_ZERO);
2108 	sp->sw_vp = vp;
2109 	sp->sw_id = id;
2110 	sp->sw_udev = udev;
2111 	sp->sw_flags = 0;
2112 	sp->sw_nblks = nblks;
2113 	sp->sw_used = 0;
2114 	sp->sw_strategy = strategy;
2115 	sp->sw_close = close;
2116 
2117 	sp->sw_blist = blist_create(nblks);
2118 	/*
2119 	 * Do not free the first two block in order to avoid overwriting
2120 	 * any bsd label at the front of the partition
2121 	 */
2122 	blist_free(sp->sw_blist, 2, nblks - 2);
2123 
2124 	dvbase = 0;
2125 	mtx_lock(&sw_dev_mtx);
2126 	TAILQ_FOREACH(tsp, &swtailq, sw_list) {
2127 		if (tsp->sw_end >= dvbase) {
2128 			/*
2129 			 * We put one uncovered page between the devices
2130 			 * in order to definitively prevent any cross-device
2131 			 * I/O requests
2132 			 */
2133 			dvbase = tsp->sw_end + 1;
2134 		}
2135 	}
2136 	sp->sw_first = dvbase;
2137 	sp->sw_end = dvbase + nblks;
2138 	TAILQ_INSERT_TAIL(&swtailq, sp, sw_list);
2139 	nswapdev++;
2140 	swap_pager_avail += nblks;
2141 	swp_sizecheck();
2142 	mtx_unlock(&sw_dev_mtx);
2143 }
2144 
2145 /*
2146  * SYSCALL: swapoff(devname)
2147  *
2148  * Disable swapping on the given device.
2149  *
2150  * XXX: Badly designed system call: it should use a device index
2151  * rather than filename as specification.  We keep sw_vp around
2152  * only to make this work.
2153  */
2154 #ifndef _SYS_SYSPROTO_H_
2155 struct swapoff_args {
2156 	char *name;
2157 };
2158 #endif
2159 
2160 /*
2161  * MPSAFE
2162  */
2163 /* ARGSUSED */
2164 int
2165 swapoff(struct thread *td, struct swapoff_args *uap)
2166 {
2167 	struct vnode *vp;
2168 	struct nameidata nd;
2169 	struct swdevt *sp;
2170 	u_long nblks, dvbase;
2171 	int error;
2172 
2173 	mtx_lock(&Giant);
2174 
2175 	error = suser(td);
2176 	if (error)
2177 		goto done2;
2178 
2179 	while (swdev_syscall_active)
2180 	    tsleep(&swdev_syscall_active, PUSER - 1, "swpoff", 0);
2181 	swdev_syscall_active = 1;
2182 
2183 	NDINIT(&nd, LOOKUP, FOLLOW, UIO_USERSPACE, uap->name, td);
2184 	error = namei(&nd);
2185 	if (error)
2186 		goto done;
2187 	NDFREE(&nd, NDF_ONLY_PNBUF);
2188 	vp = nd.ni_vp;
2189 
2190 	mtx_lock(&sw_dev_mtx);
2191 	TAILQ_FOREACH(sp, &swtailq, sw_list) {
2192 		if (sp->sw_vp == vp)
2193 			goto found;
2194 	}
2195 	mtx_unlock(&sw_dev_mtx);
2196 	error = EINVAL;
2197 	goto done;
2198 found:
2199 	mtx_unlock(&sw_dev_mtx);
2200 #ifdef MAC
2201 	(void) vn_lock(vp, LK_EXCLUSIVE | LK_RETRY, td);
2202 	error = mac_check_system_swapoff(td->td_ucred, vp);
2203 	(void) VOP_UNLOCK(vp, 0, td);
2204 	if (error != 0)
2205 		goto done;
2206 #endif
2207 
2208 	nblks = sp->sw_nblks;
2209 
2210 	/*
2211 	 * We can turn off this swap device safely only if the
2212 	 * available virtual memory in the system will fit the amount
2213 	 * of data we will have to page back in, plus an epsilon so
2214 	 * the system doesn't become critically low on swap space.
2215 	 */
2216 	if (cnt.v_free_count + cnt.v_cache_count + swap_pager_avail <
2217 	    nblks + nswap_lowat) {
2218 		error = ENOMEM;
2219 		goto done;
2220 	}
2221 
2222 	/*
2223 	 * Prevent further allocations on this device.
2224 	 */
2225 	mtx_lock(&sw_dev_mtx);
2226 	sp->sw_flags |= SW_CLOSING;
2227 	for (dvbase = 0; dvbase < sp->sw_end; dvbase += dmmax) {
2228 		swap_pager_avail -= blist_fill(sp->sw_blist,
2229 		     dvbase, dmmax);
2230 	}
2231 	mtx_unlock(&sw_dev_mtx);
2232 
2233 	/*
2234 	 * Page in the contents of the device and close it.
2235 	 */
2236 #ifndef NO_SWAPPING
2237        	vm_proc_swapin_all(sp);
2238 #endif /* !NO_SWAPPING */
2239 	swap_pager_swapoff(sp, &sp->sw_used);
2240 
2241 	sp->sw_close(td, sp);
2242 	sp->sw_id = NULL;
2243 	mtx_lock(&sw_dev_mtx);
2244 	TAILQ_REMOVE(&swtailq, sp, sw_list);
2245 	nswapdev--;
2246 	if (nswapdev == 0) {
2247 		swap_pager_full = 2;
2248 		swap_pager_almost_full = 1;
2249 	}
2250 	if (swdevhd == sp)
2251 		swdevhd = NULL;
2252 	mtx_unlock(&sw_dev_mtx);
2253 	blist_destroy(sp->sw_blist);
2254 	free(sp, M_VMPGDATA);
2255 
2256 done:
2257 	swdev_syscall_active = 0;
2258 	wakeup_one(&swdev_syscall_active);
2259 done2:
2260 	mtx_unlock(&Giant);
2261 	return (error);
2262 }
2263 
2264 void
2265 swap_pager_status(int *total, int *used)
2266 {
2267 	struct swdevt *sp;
2268 
2269 	*total = 0;
2270 	*used = 0;
2271 	mtx_lock(&sw_dev_mtx);
2272 	TAILQ_FOREACH(sp, &swtailq, sw_list) {
2273 		*total += sp->sw_nblks;
2274 		*used += sp->sw_used;
2275 	}
2276 	mtx_unlock(&sw_dev_mtx);
2277 }
2278 
2279 static int
2280 sysctl_vm_swap_info(SYSCTL_HANDLER_ARGS)
2281 {
2282 	int	*name = (int *)arg1;
2283 	int	error, n;
2284 	struct xswdev xs;
2285 	struct swdevt *sp;
2286 
2287 	if (arg2 != 1) /* name length */
2288 		return (EINVAL);
2289 
2290 	n = 0;
2291 	mtx_lock(&sw_dev_mtx);
2292 	TAILQ_FOREACH(sp, &swtailq, sw_list) {
2293 		if (n == *name) {
2294 			mtx_unlock(&sw_dev_mtx);
2295 			xs.xsw_version = XSWDEV_VERSION;
2296 			xs.xsw_dev = sp->sw_udev;
2297 			xs.xsw_flags = sp->sw_flags;
2298 			xs.xsw_nblks = sp->sw_nblks;
2299 			xs.xsw_used = sp->sw_used;
2300 
2301 			error = SYSCTL_OUT(req, &xs, sizeof(xs));
2302 			return (error);
2303 		}
2304 		n++;
2305 	}
2306 	mtx_unlock(&sw_dev_mtx);
2307 	return (ENOENT);
2308 }
2309 
2310 SYSCTL_INT(_vm, OID_AUTO, nswapdev, CTLFLAG_RD, &nswapdev, 0,
2311     "Number of swap devices");
2312 SYSCTL_NODE(_vm, OID_AUTO, swap_info, CTLFLAG_RD, sysctl_vm_swap_info,
2313     "Swap statistics by device");
2314 
2315 /*
2316  * vmspace_swap_count() - count the approximate swap useage in pages for a
2317  *			  vmspace.
2318  *
2319  *	The map must be locked.
2320  *
2321  *	Swap useage is determined by taking the proportional swap used by
2322  *	VM objects backing the VM map.  To make up for fractional losses,
2323  *	if the VM object has any swap use at all the associated map entries
2324  *	count for at least 1 swap page.
2325  */
2326 int
2327 vmspace_swap_count(struct vmspace *vmspace)
2328 {
2329 	vm_map_t map = &vmspace->vm_map;
2330 	vm_map_entry_t cur;
2331 	int count = 0;
2332 
2333 	for (cur = map->header.next; cur != &map->header; cur = cur->next) {
2334 		vm_object_t object;
2335 
2336 		if ((cur->eflags & MAP_ENTRY_IS_SUB_MAP) == 0 &&
2337 		    (object = cur->object.vm_object) != NULL) {
2338 			VM_OBJECT_LOCK(object);
2339 			if (object->type == OBJT_SWAP &&
2340 			    object->un_pager.swp.swp_bcount != 0) {
2341 				int n = (cur->end - cur->start) / PAGE_SIZE;
2342 
2343 				count += object->un_pager.swp.swp_bcount *
2344 				    SWAP_META_PAGES * n / object->size + 1;
2345 			}
2346 			VM_OBJECT_UNLOCK(object);
2347 		}
2348 	}
2349 	return (count);
2350 }
2351 
2352 /*
2353  * GEOM backend
2354  *
2355  * Swapping onto disk devices.
2356  *
2357  */
2358 
2359 static struct g_class g_swap_class = {
2360 	.name = "SWAP",
2361 };
2362 
2363 DECLARE_GEOM_CLASS(g_swap_class, g_class);
2364 
2365 
2366 static void
2367 swapgeom_done(struct bio *bp2)
2368 {
2369 	struct buf *bp;
2370 
2371 	bp = bp2->bio_caller2;
2372 	if (bp2->bio_error)
2373 		bp->b_ioflags |= BIO_ERROR;
2374 	bufdone(bp);
2375 	g_destroy_bio(bp2);
2376 }
2377 
2378 static void
2379 swapgeom_strategy(struct buf *bp, struct swdevt *sp)
2380 {
2381 	struct bio *bio;
2382 	struct g_consumer *cp;
2383 
2384 	cp = sp->sw_id;
2385 	if (cp == NULL) {
2386 		bp->b_error = ENXIO;
2387 		bp->b_ioflags |= BIO_ERROR;
2388 		bufdone(bp);
2389 		return;
2390 	}
2391 	bio = g_clone_bio(&bp->b_io);
2392 	if (bio == NULL) {
2393 		/*
2394 		 * XXX: This is better than panicing, but not much better.
2395 		 * XXX: Somehow this should be retried.  A more generic
2396 		 * XXX: implementation of ENOMEM in geom may be able to cope.
2397 		 */
2398 		bp->b_error = ENOMEM;
2399 		bp->b_ioflags |= BIO_ERROR;
2400 		bufdone(bp);
2401 		return;
2402 	}
2403 	bio->bio_caller2 = bp;
2404 	bio->bio_offset = (bp->b_blkno - sp->sw_first) * PAGE_SIZE;
2405 	bio->bio_length = bp->b_bcount;
2406 	bio->bio_done = swapgeom_done;
2407 	g_io_request(bio, cp);
2408 	return;
2409 }
2410 
2411 static void
2412 swapgeom_orphan(struct g_consumer *cp)
2413 {
2414 	struct swdevt *sp;
2415 
2416 	mtx_lock(&sw_dev_mtx);
2417 	TAILQ_FOREACH(sp, &swtailq, sw_list)
2418 		if (sp->sw_id == cp)
2419 			sp->sw_id = NULL;
2420 	mtx_unlock(&sw_dev_mtx);
2421 }
2422 
2423 static void
2424 swapgeom_close_ev(void *arg, int flags)
2425 {
2426 	struct g_consumer *cp;
2427 
2428 	cp = arg;
2429 	g_access(cp, -1, -1, 0);
2430 	g_detach(cp);
2431 	g_destroy_consumer(cp);
2432 }
2433 
2434 static void
2435 swapgeom_close(struct thread *td, struct swdevt *sw)
2436 {
2437 
2438 	/* XXX: direct call when Giant untangled */
2439 	g_waitfor_event(swapgeom_close_ev, sw->sw_id, M_WAITOK, NULL);
2440 }
2441 
2442 
2443 struct swh0h0 {
2444 	dev_t	dev;
2445 	struct vnode *vp;
2446 	int	error;
2447 };
2448 
2449 static void
2450 swapongeom_ev(void *arg, int flags)
2451 {
2452 	struct swh0h0 *swh;
2453 	struct g_provider *pp;
2454 	struct g_consumer *cp;
2455 	static struct g_geom *gp;
2456 	struct swdevt *sp;
2457 	u_long nblks;
2458 	int error;
2459 
2460 	swh = arg;
2461 	swh->error = 0;
2462 	pp = g_dev_getprovider(swh->dev);
2463 	if (pp == NULL) {
2464 		swh->error = ENODEV;
2465 		return;
2466 	}
2467 	mtx_lock(&sw_dev_mtx);
2468 	TAILQ_FOREACH(sp, &swtailq, sw_list) {
2469 		cp = sp->sw_id;
2470 		if (cp != NULL && cp->provider == pp) {
2471 			mtx_unlock(&sw_dev_mtx);
2472 			swh->error = EBUSY;
2473 			return;
2474 		}
2475 	}
2476 	mtx_unlock(&sw_dev_mtx);
2477 	if (gp == NULL) {
2478 		gp = g_new_geomf(&g_swap_class, "swap", NULL);
2479 		gp->orphan = swapgeom_orphan;
2480 	}
2481 	cp = g_new_consumer(gp);
2482 	g_attach(cp, pp);
2483 	/*
2484 	 * XXX: Everytime you think you can improve the margin for
2485 	 * footshooting, somebody depends on the ability to do so:
2486 	 * savecore(8) wants to write to our swapdev so we cannot
2487 	 * set an exclusive count :-(
2488 	 */
2489 	error = g_access(cp, 1, 1, 0);
2490 	if (error) {
2491 		g_detach(cp);
2492 		g_destroy_consumer(cp);
2493 		swh->error = error;
2494 		return;
2495 	}
2496 	nblks = pp->mediasize / DEV_BSIZE;
2497 	swaponsomething(swh->vp, cp, nblks, swapgeom_strategy,
2498 	    swapgeom_close, dev2udev(swh->dev));
2499 	swh->error = 0;
2500 	return;
2501 }
2502 
2503 static int
2504 swapongeom(struct thread *td, struct vnode *vp)
2505 {
2506 	int error;
2507 	struct swh0h0 swh;
2508 
2509 	vn_lock(vp, LK_EXCLUSIVE | LK_RETRY, td);
2510 
2511 	swh.dev = vp->v_rdev;
2512 	swh.vp = vp;
2513 	swh.error = 0;
2514 	/* XXX: direct call when Giant untangled */
2515 	error = g_waitfor_event(swapongeom_ev, &swh, M_WAITOK, NULL);
2516 	if (!error)
2517 		error = swh.error;
2518 	VOP_UNLOCK(vp, 0, td);
2519 	return (error);
2520 }
2521 
2522 /*
2523  * VNODE backend
2524  *
2525  * This is used mainly for network filesystem (read: probably only tested
2526  * with NFS) swapfiles.
2527  *
2528  */
2529 
2530 static void
2531 swapdev_strategy(struct buf *bp, struct swdevt *sp)
2532 {
2533 	int s;
2534 	struct vnode *vp, *vp2;
2535 
2536 	bp->b_dev = NODEV;
2537 	bp->b_blkno = ctodb(bp->b_blkno - sp->sw_first);
2538 
2539 	vp2 = sp->sw_id;
2540 	vhold(vp2);
2541 	s = splvm();
2542 	if (bp->b_iocmd == BIO_WRITE) {
2543 		vp = bp->b_vp;
2544 		if (vp) {
2545 			VI_LOCK(vp);
2546 			vp->v_numoutput--;
2547 			if ((vp->v_iflag & VI_BWAIT) && vp->v_numoutput <= 0) {
2548 				vp->v_iflag &= ~VI_BWAIT;
2549 				wakeup(&vp->v_numoutput);
2550 			}
2551 			VI_UNLOCK(vp);
2552 		}
2553 		VI_LOCK(vp2);
2554 		vp2->v_numoutput++;
2555 		VI_UNLOCK(vp2);
2556 	}
2557 	bp->b_vp = vp2;
2558 	splx(s);
2559 	bp->b_iooffset = dbtob(bp->b_blkno);
2560 	VOP_STRATEGY(vp2, bp);
2561 	return;
2562 }
2563 
2564 static void
2565 swapdev_close(struct thread *td, struct swdevt *sp)
2566 {
2567 
2568 	VOP_CLOSE(sp->sw_vp, FREAD | FWRITE, td->td_ucred, td);
2569 	vrele(sp->sw_vp);
2570 }
2571 
2572 
2573 static int
2574 swaponvp(struct thread *td, struct vnode *vp, u_long nblks)
2575 {
2576 	struct swdevt *sp;
2577 	int error;
2578 
2579 	if (nblks == 0)
2580 		return (ENXIO);
2581 	mtx_lock(&sw_dev_mtx);
2582 	TAILQ_FOREACH(sp, &swtailq, sw_list) {
2583 		if (sp->sw_id == vp) {
2584 			mtx_unlock(&sw_dev_mtx);
2585 			return (EBUSY);
2586 		}
2587 	}
2588 	mtx_unlock(&sw_dev_mtx);
2589 
2590 	(void) vn_lock(vp, LK_EXCLUSIVE | LK_RETRY, td);
2591 #ifdef MAC
2592 	error = mac_check_system_swapon(td->td_ucred, vp);
2593 	if (error == 0)
2594 #endif
2595 		error = VOP_OPEN(vp, FREAD | FWRITE, td->td_ucred, td, -1);
2596 	(void) VOP_UNLOCK(vp, 0, td);
2597 	if (error)
2598 		return (error);
2599 
2600 	swaponsomething(vp, vp, nblks, swapdev_strategy, swapdev_close,
2601 	    NOUDEV);
2602 	return (0);
2603 }
2604