2 * Copyright (c) 1998,2004 The DragonFly Project. All rights reserved.
4 * This code is derived from software contributed to The DragonFly Project
5 * by Matthew Dillon <dillon@backplane.com>
7 * Redistribution and use in source and binary forms, with or without
8 * modification, are permitted provided that the following conditions
11 * 1. Redistributions of source code must retain the above copyright
12 * notice, this list of conditions and the following disclaimer.
13 * 2. Redistributions in binary form must reproduce the above copyright
14 * notice, this list of conditions and the following disclaimer in
15 * the documentation and/or other materials provided with the
17 * 3. Neither the name of The DragonFly Project nor the names of its
18 * contributors may be used to endorse or promote products derived
19 * from this software without specific, prior written permission.
21 * THIS SOFTWARE IS PROVIDED BY THE COPYRIGHT HOLDERS AND CONTRIBUTORS
22 * ``AS IS'' AND ANY EXPRESS OR IMPLIED WARRANTIES, INCLUDING, BUT NOT
23 * LIMITED TO, THE IMPLIED WARRANTIES OF MERCHANTABILITY AND FITNESS
24 * FOR A PARTICULAR PURPOSE ARE DISCLAIMED. IN NO EVENT SHALL THE
25 * COPYRIGHT HOLDERS OR CONTRIBUTORS BE LIABLE FOR ANY DIRECT, INDIRECT,
26 * INCIDENTAL, SPECIAL, EXEMPLARY OR CONSEQUENTIAL DAMAGES (INCLUDING,
27 * BUT NOT LIMITED TO, PROCUREMENT OF SUBSTITUTE GOODS OR SERVICES;
28 * LOSS OF USE, DATA, OR PROFITS; OR BUSINESS INTERRUPTION) HOWEVER CAUSED
29 * AND ON ANY THEORY OF LIABILITY, WHETHER IN CONTRACT, STRICT LIABILITY,
30 * OR TORT (INCLUDING NEGLIGENCE OR OTHERWISE) ARISING IN ANY WAY OUT
31 * OF THE USE OF THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF
34 * Copyright (c) 1994 John S. Dyson
35 * Copyright (c) 1990 University of Utah.
36 * Copyright (c) 1991, 1993
37 * The Regents of the University of California. All rights reserved.
39 * This code is derived from software contributed to Berkeley by
40 * the Systems Programming Group of the University of Utah Computer
43 * Redistribution and use in source and binary forms, with or without
44 * modification, are permitted provided that the following conditions
46 * 1. Redistributions of source code must retain the above copyright
47 * notice, this list of conditions and the following disclaimer.
48 * 2. Redistributions in binary form must reproduce the above copyright
49 * notice, this list of conditions and the following disclaimer in the
50 * documentation and/or other materials provided with the distribution.
51 * 3. All advertising materials mentioning features or use of this software
52 * must display the following acknowledgement:
53 * This product includes software developed by the University of
54 * California, Berkeley and its contributors.
55 * 4. Neither the name of the University nor the names of its contributors
56 * may be used to endorse or promote products derived from this software
57 * without specific prior written permission.
59 * THIS SOFTWARE IS PROVIDED BY THE REGENTS AND CONTRIBUTORS ``AS IS'' AND
60 * ANY EXPRESS OR IMPLIED WARRANTIES, INCLUDING, BUT NOT LIMITED TO, THE
61 * IMPLIED WARRANTIES OF MERCHANTABILITY AND FITNESS FOR A PARTICULAR PURPOSE
62 * ARE DISCLAIMED. IN NO EVENT SHALL THE REGENTS OR CONTRIBUTORS BE LIABLE
63 * FOR ANY DIRECT, INDIRECT, INCIDENTAL, SPECIAL, EXEMPLARY, OR CONSEQUENTIAL
64 * DAMAGES (INCLUDING, BUT NOT LIMITED TO, PROCUREMENT OF SUBSTITUTE GOODS
65 * OR SERVICES; LOSS OF USE, DATA, OR PROFITS; OR BUSINESS INTERRUPTION)
66 * HOWEVER CAUSED AND ON ANY THEORY OF LIABILITY, WHETHER IN CONTRACT, STRICT
67 * LIABILITY, OR TORT (INCLUDING NEGLIGENCE OR OTHERWISE) ARISING IN ANY WAY
68 * OUT OF THE USE OF THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF
74 * Radix Bitmap 'blists'.
76 * - The new swapper uses the new radix bitmap code. This should scale
77 * to arbitrarily small or arbitrarily large swap spaces and an almost
78 * arbitrary degree of fragmentation.
82 * - on the fly reallocation of swap during putpages. The new system
83 * does not try to keep previously allocated swap blocks for dirty
86 * - on the fly deallocation of swap
88 * - No more garbage collection required. Unnecessarily allocated swap
89 * blocks only exist for dirty vm_page_t's now and these are already
90 * cycled (in a high-load system) by the pager. We also do on-the-fly
91 * removal of invalidated swap blocks when a page is destroyed
94 * from: Utah $Hdr: swap_pager.c 1.4 91/04/30$
96 * @(#)swap_pager.c 8.9 (Berkeley) 3/21/94
98 * $FreeBSD: src/sys/vm/swap_pager.c,v 1.130.2.12 2002/08/31 21:15:55 dillon Exp $
99 * $DragonFly: src/sys/vm/swap_pager.c,v 1.14 2004/07/16 05:52:14 dillon Exp $
102 #include <sys/param.h>
103 #include <sys/systm.h>
104 #include <sys/conf.h>
105 #include <sys/kernel.h>
106 #include <sys/proc.h>
108 #include <sys/vnode.h>
109 #include <sys/malloc.h>
110 #include <sys/vmmeter.h>
111 #include <sys/sysctl.h>
112 #include <sys/blist.h>
113 #include <sys/lock.h>
114 #include <sys/vmmeter.h>
116 #ifndef MAX_PAGEOUT_CLUSTER
117 #define MAX_PAGEOUT_CLUSTER 16
120 #define SWB_NPAGES MAX_PAGEOUT_CLUSTER
122 #include "opt_swap.h"
124 #include <vm/vm_object.h>
125 #include <vm/vm_page.h>
126 #include <vm/vm_pager.h>
127 #include <vm/vm_pageout.h>
128 #include <vm/swap_pager.h>
129 #include <vm/vm_extern.h>
130 #include <vm/vm_zone.h>
132 #include <sys/buf2.h>
133 #include <vm/vm_page2.h>
135 #define SWM_FREE 0x02 /* free, period */
136 #define SWM_POP 0x04 /* pop out */
139 * vm_swap_size is in page-sized chunks now. It was DEV_BSIZE'd chunks
143 extern int vm_swap_size; /* number of free swap blocks, in pages */
145 int swap_pager_full; /* swap space exhaustion (task killing) */
146 static int swap_pager_almost_full; /* swap space exhaustion (w/ hysteresis)*/
147 static int nsw_rcount; /* free read buffers */
148 static int nsw_wcount_sync; /* limit write buffers / synchronous */
149 static int nsw_wcount_async; /* limit write buffers / asynchronous */
150 static int nsw_wcount_async_max;/* assigned maximum */
151 static int nsw_cluster_max; /* maximum VOP I/O allowed */
152 static int sw_alloc_interlock; /* swap pager allocation interlock */
154 struct blist *swapblist;
155 static struct swblock **swhash;
156 static int swhash_mask;
157 static int swap_async_max = 4; /* maximum in-progress async I/O's */
159 extern struct vnode *swapdev_vp; /* from vm_swap.c */
161 SYSCTL_INT(_vm, OID_AUTO, swap_async_max,
162 CTLFLAG_RW, &swap_async_max, 0, "Maximum running async swap ops");
165 * "named" and "unnamed" anon region objects. Try to reduce the overhead
166 * of searching a named list by hashing it just a little.
171 #define NOBJLIST(handle) \
172 (&swap_pager_object_list[((int)(intptr_t)handle >> 4) & (NOBJLISTS-1)])
174 static struct pagerlst swap_pager_object_list[NOBJLISTS];
175 struct pagerlst swap_pager_un_object_list;
179 * pagerops for OBJT_SWAP - "swap pager". Some ops are also global procedure
180 * calls hooked from other parts of the VM system and do not appear here.
181 * (see vm/swap_pager.h).
185 swap_pager_alloc (void *handle, vm_ooffset_t size,
186 vm_prot_t prot, vm_ooffset_t offset);
187 static void swap_pager_dealloc (vm_object_t object);
188 static int swap_pager_getpages (vm_object_t, vm_page_t *, int, int);
189 static void swap_pager_init (void);
190 static void swap_pager_unswapped (vm_page_t);
191 static void swap_pager_strategy (vm_object_t, struct buf *);
193 struct pagerops swappagerops = {
194 swap_pager_init, /* early system initialization of pager */
195 swap_pager_alloc, /* allocate an OBJT_SWAP object */
196 swap_pager_dealloc, /* deallocate an OBJT_SWAP object */
197 swap_pager_getpages, /* pagein */
198 swap_pager_putpages, /* pageout */
199 swap_pager_haspage, /* get backing store status for page */
200 swap_pager_unswapped, /* remove swap related to page */
201 swap_pager_strategy /* pager strategy call */
205 * dmmax is in page-sized chunks with the new swap system. It was
206 * dev-bsized chunks in the old. dmmax is always a power of 2.
208 * swap_*() routines are externally accessible. swp_*() routines are
213 static int dmmax_mask;
214 int nswap_lowat = 128; /* in pages, swap_pager_almost_full warn */
215 int nswap_hiwat = 512; /* in pages, swap_pager_almost_full warn */
217 static __inline void swp_sizecheck (void);
218 static void swp_pager_sync_iodone (struct buf *bp);
219 static void swp_pager_async_iodone (struct buf *bp);
222 * Swap bitmap functions
225 static __inline void swp_pager_freeswapspace (daddr_t blk, int npages);
226 static __inline daddr_t swp_pager_getswapspace (int npages);
232 static void swp_pager_meta_build (vm_object_t, vm_pindex_t, daddr_t);
233 static void swp_pager_meta_free (vm_object_t, vm_pindex_t, daddr_t);
234 static void swp_pager_meta_free_all (vm_object_t);
235 static daddr_t swp_pager_meta_ctl (vm_object_t, vm_pindex_t, int);
238 * SWP_SIZECHECK() - update swap_pager_full indication
240 * update the swap_pager_almost_full indication and warn when we are
241 * about to run out of swap space, using lowat/hiwat hysteresis.
243 * Clear swap_pager_full ( task killing ) indication when lowat is met.
245 * No restrictions on call
246 * This routine may not block.
247 * This routine must be called at splvm()
253 if (vm_swap_size < nswap_lowat) {
254 if (swap_pager_almost_full == 0) {
255 printf("swap_pager: out of swap space\n");
256 swap_pager_almost_full = 1;
260 if (vm_swap_size > nswap_hiwat)
261 swap_pager_almost_full = 0;
266 * SWAP_PAGER_INIT() - initialize the swap pager!
268 * Expected to be started from system init. NOTE: This code is run
269 * before much else so be careful what you depend on. Most of the VM
270 * system has yet to be initialized at this point.
274 swap_pager_init(void)
277 * Initialize object lists
281 for (i = 0; i < NOBJLISTS; ++i)
282 TAILQ_INIT(&swap_pager_object_list[i]);
283 TAILQ_INIT(&swap_pager_un_object_list);
286 * Device Stripe, in PAGE_SIZE'd blocks
289 dmmax = SWB_NPAGES * 2;
290 dmmax_mask = ~(dmmax - 1);
294 * SWAP_PAGER_SWAP_INIT() - swap pager initialization from pageout process
296 * Expected to be started from pageout process once, prior to entering
301 swap_pager_swap_init(void)
306 * Number of in-transit swap bp operations. Don't
307 * exhaust the pbufs completely. Make sure we
308 * initialize workable values (0 will work for hysteresis
309 * but it isn't very efficient).
311 * The nsw_cluster_max is constrained by the number of pages an XIO
312 * holds, i.e., (MAXPHYS/PAGE_SIZE) and our locally defined
313 * MAX_PAGEOUT_CLUSTER. Also be aware that swap ops are
314 * constrained by the swap device interleave stripe size.
316 * Currently we hardwire nsw_wcount_async to 4. This limit is
317 * designed to prevent other I/O from having high latencies due to
318 * our pageout I/O. The value 4 works well for one or two active swap
319 * devices but is probably a little low if you have more. Even so,
320 * a higher value would probably generate only a limited improvement
321 * with three or four active swap devices since the system does not
322 * typically have to pageout at extreme bandwidths. We will want
323 * at least 2 per swap devices, and 4 is a pretty good value if you
324 * have one NFS swap device due to the command/ack latency over NFS.
325 * So it all works out pretty well.
328 nsw_cluster_max = min((MAXPHYS/PAGE_SIZE), MAX_PAGEOUT_CLUSTER);
330 nsw_rcount = (nswbuf + 1) / 2;
331 nsw_wcount_sync = (nswbuf + 3) / 4;
332 nsw_wcount_async = 4;
333 nsw_wcount_async_max = nsw_wcount_async;
336 * Initialize our zone. Right now I'm just guessing on the number
337 * we need based on the number of pages in the system. Each swblock
338 * can hold 16 pages, so this is probably overkill. This reservation
339 * is typically limited to around 32MB by default.
341 n = vmstats.v_page_count / 2;
342 if (maxswzone && n > maxswzone / sizeof(struct swblock))
343 n = maxswzone / sizeof(struct swblock);
349 sizeof(struct swblock),
353 if (swap_zone != NULL)
356 * if the allocation failed, try a zone two thirds the
357 * size of the previous attempt.
362 if (swap_zone == NULL)
363 panic("swap_pager_swap_init: swap_zone == NULL");
365 printf("Swap zone entries reduced from %d to %d.\n", n2, n);
369 * Initialize our meta-data hash table. The swapper does not need to
370 * be quite as efficient as the VM system, so we do not use an
371 * oversized hash table.
373 * n: size of hash table, must be power of 2
374 * swhash_mask: hash table index mask
377 for (n = 1; n < n2 / 8; n *= 2)
380 swhash = malloc(sizeof(struct swblock *) * n, M_VMPGDATA, M_WAITOK);
381 bzero(swhash, sizeof(struct swblock *) * n);
387 * SWAP_PAGER_ALLOC() - allocate a new OBJT_SWAP VM object and instantiate
388 * its metadata structures.
390 * This routine is called from the mmap and fork code to create a new
391 * OBJT_SWAP object. We do this by creating an OBJT_DEFAULT object
392 * and then converting it with swp_pager_meta_build().
394 * This routine may block in vm_object_allocate() and create a named
395 * object lookup race, so we must interlock. We must also run at
396 * splvm() for the object lookup to handle races with interrupts, but
397 * we do not have to maintain splvm() in between the lookup and the
398 * add because (I believe) it is not possible to attempt to create
399 * a new swap object w/handle when a default object with that handle
404 swap_pager_alloc(void *handle, vm_ooffset_t size, vm_prot_t prot,
411 * Reference existing named region or allocate new one. There
412 * should not be a race here against swp_pager_meta_build()
413 * as called from vm_page_remove() in regards to the lookup
417 while (sw_alloc_interlock) {
418 sw_alloc_interlock = -1;
419 tsleep(&sw_alloc_interlock, 0, "swpalc", 0);
421 sw_alloc_interlock = 1;
423 object = vm_pager_object_lookup(NOBJLIST(handle), handle);
425 if (object != NULL) {
426 vm_object_reference(object);
428 object = vm_object_allocate(OBJT_DEFAULT,
429 OFF_TO_IDX(offset + PAGE_MASK + size));
430 object->handle = handle;
432 swp_pager_meta_build(object, 0, SWAPBLK_NONE);
435 if (sw_alloc_interlock < 0)
436 wakeup(&sw_alloc_interlock);
438 sw_alloc_interlock = 0;
440 object = vm_object_allocate(OBJT_DEFAULT,
441 OFF_TO_IDX(offset + PAGE_MASK + size));
443 swp_pager_meta_build(object, 0, SWAPBLK_NONE);
450 * SWAP_PAGER_DEALLOC() - remove swap metadata from object
452 * The swap backing for the object is destroyed. The code is
453 * designed such that we can reinstantiate it later, but this
454 * routine is typically called only when the entire object is
455 * about to be destroyed.
457 * This routine may block, but no longer does.
459 * The object must be locked or unreferenceable.
463 swap_pager_dealloc(vm_object_t object)
468 * Remove from list right away so lookups will fail if we block for
469 * pageout completion.
472 if (object->handle == NULL) {
473 TAILQ_REMOVE(&swap_pager_un_object_list, object, pager_object_list);
475 TAILQ_REMOVE(NOBJLIST(object->handle), object, pager_object_list);
478 vm_object_pip_wait(object, "swpdea");
481 * Free all remaining metadata. We only bother to free it from
482 * the swap meta data. We do not attempt to free swapblk's still
483 * associated with vm_page_t's for this object. We do not care
484 * if paging is still in progress on some objects.
487 swp_pager_meta_free_all(object);
491 /************************************************************************
492 * SWAP PAGER BITMAP ROUTINES *
493 ************************************************************************/
496 * SWP_PAGER_GETSWAPSPACE() - allocate raw swap space
498 * Allocate swap for the requested number of pages. The starting
499 * swap block number (a page index) is returned or SWAPBLK_NONE
500 * if the allocation failed.
502 * Also has the side effect of advising that somebody made a mistake
503 * when they configured swap and didn't configure enough.
505 * Must be called at splvm() to avoid races with bitmap frees from
506 * vm_page_remove() aka swap_pager_page_removed().
508 * This routine may not block
509 * This routine must be called at splvm().
512 static __inline daddr_t
513 swp_pager_getswapspace(int npages)
517 if ((blk = blist_alloc(swapblist, npages)) == SWAPBLK_NONE) {
518 if (swap_pager_full != 2) {
519 printf("swap_pager_getswapspace: failed\n");
521 swap_pager_almost_full = 1;
524 vm_swap_size -= npages;
531 * SWP_PAGER_FREESWAPSPACE() - free raw swap space
533 * This routine returns the specified swap blocks back to the bitmap.
535 * Note: This routine may not block (it could in the old swap code),
536 * and through the use of the new blist routines it does not block.
538 * We must be called at splvm() to avoid races with bitmap frees from
539 * vm_page_remove() aka swap_pager_page_removed().
541 * This routine may not block
542 * This routine must be called at splvm().
546 swp_pager_freeswapspace(daddr_t blk, int npages)
548 blist_free(swapblist, blk, npages);
549 vm_swap_size += npages;
554 * SWAP_PAGER_FREESPACE() - frees swap blocks associated with a page
555 * range within an object.
557 * This is a globally accessible routine.
559 * This routine removes swapblk assignments from swap metadata.
561 * The external callers of this routine typically have already destroyed
562 * or renamed vm_page_t's associated with this range in the object so
565 * This routine may be called at any spl. We up our spl to splvm temporarily
566 * in order to perform the metadata removal.
570 swap_pager_freespace(vm_object_t object, vm_pindex_t start, vm_size_t size)
573 swp_pager_meta_free(object, start, size);
578 * SWAP_PAGER_RESERVE() - reserve swap blocks in object
580 * Assigns swap blocks to the specified range within the object. The
581 * swap blocks are not zerod. Any previous swap assignment is destroyed.
583 * Returns 0 on success, -1 on failure.
587 swap_pager_reserve(vm_object_t object, vm_pindex_t start, vm_size_t size)
591 daddr_t blk = SWAPBLK_NONE;
592 vm_pindex_t beg = start; /* save start index */
598 while ((blk = swp_pager_getswapspace(n)) == SWAPBLK_NONE) {
601 swp_pager_meta_free(object, beg, start - beg);
607 swp_pager_meta_build(object, start, blk);
613 swp_pager_meta_free(object, start, n);
619 * SWAP_PAGER_COPY() - copy blocks from source pager to destination pager
620 * and destroy the source.
622 * Copy any valid swapblks from the source to the destination. In
623 * cases where both the source and destination have a valid swapblk,
624 * we keep the destination's.
626 * This routine is allowed to block. It may block allocating metadata
627 * indirectly through swp_pager_meta_build() or if paging is still in
628 * progress on the source.
630 * This routine can be called at any spl
632 * XXX vm_page_collapse() kinda expects us not to block because we
633 * supposedly do not need to allocate memory, but for the moment we
634 * *may* have to get a little memory from the zone allocator, but
635 * it is taken from the interrupt memory. We should be ok.
637 * The source object contains no vm_page_t's (which is just as well)
639 * The source object is of type OBJT_SWAP.
641 * The source and destination objects must be locked or
642 * inaccessible (XXX are they ?)
646 swap_pager_copy(vm_object_t srcobject, vm_object_t dstobject,
647 vm_pindex_t offset, int destroysource)
655 * If destroysource is set, we remove the source object from the
656 * swap_pager internal queue now.
660 if (srcobject->handle == NULL) {
662 &swap_pager_un_object_list,
668 NOBJLIST(srcobject->handle),
676 * transfer source to destination.
679 for (i = 0; i < dstobject->size; ++i) {
683 * Locate (without changing) the swapblk on the destination,
684 * unless it is invalid in which case free it silently, or
685 * if the destination is a resident page, in which case the
686 * source is thrown away.
689 dstaddr = swp_pager_meta_ctl(dstobject, i, 0);
691 if (dstaddr == SWAPBLK_NONE) {
693 * Destination has no swapblk and is not resident,
698 srcaddr = swp_pager_meta_ctl(
704 if (srcaddr != SWAPBLK_NONE)
705 swp_pager_meta_build(dstobject, i, srcaddr);
708 * Destination has valid swapblk or it is represented
709 * by a resident page. We destroy the sourceblock.
712 swp_pager_meta_ctl(srcobject, i + offset, SWM_FREE);
717 * Free left over swap blocks in source.
719 * We have to revert the type to OBJT_DEFAULT so we do not accidently
720 * double-remove the object from the swap queues.
724 swp_pager_meta_free_all(srcobject);
726 * Reverting the type is not necessary, the caller is going
727 * to destroy srcobject directly, but I'm doing it here
728 * for consistency since we've removed the object from its
731 srcobject->type = OBJT_DEFAULT;
737 * SWAP_PAGER_HASPAGE() - determine if we have good backing store for
738 * the requested page.
740 * We determine whether good backing store exists for the requested
741 * page and return TRUE if it does, FALSE if it doesn't.
743 * If TRUE, we also try to determine how much valid, contiguous backing
744 * store exists before and after the requested page within a reasonable
745 * distance. We do not try to restrict it to the swap device stripe
746 * (that is handled in getpages/putpages). It probably isn't worth
751 swap_pager_haspage(vm_object_t object, vm_pindex_t pindex, int *before,
758 * do we have good backing store at the requested index ?
762 blk0 = swp_pager_meta_ctl(object, pindex, 0);
764 if (blk0 == SWAPBLK_NONE) {
774 * find backwards-looking contiguous good backing store
777 if (before != NULL) {
780 for (i = 1; i < (SWB_NPAGES/2); ++i) {
785 blk = swp_pager_meta_ctl(object, pindex - i, 0);
793 * find forward-looking contiguous good backing store
799 for (i = 1; i < (SWB_NPAGES/2); ++i) {
802 blk = swp_pager_meta_ctl(object, pindex + i, 0);
813 * SWAP_PAGER_PAGE_UNSWAPPED() - remove swap backing store related to page
815 * This removes any associated swap backing store, whether valid or
816 * not, from the page.
818 * This routine is typically called when a page is made dirty, at
819 * which point any associated swap can be freed. MADV_FREE also
820 * calls us in a special-case situation
822 * NOTE!!! If the page is clean and the swap was valid, the caller
823 * should make the page dirty before calling this routine. This routine
824 * does NOT change the m->dirty status of the page. Also: MADV_FREE
827 * This routine may not block
828 * This routine must be called at splvm()
832 swap_pager_unswapped(vm_page_t m)
834 swp_pager_meta_ctl(m->object, m->pindex, SWM_FREE);
838 * SWAP_PAGER_STRATEGY() - read, write, free blocks
840 * This implements the vm_pager_strategy() interface to swap and allows
841 * other parts of the system to directly access swap as backing store
842 * through vm_objects of type OBJT_SWAP. This is intended to be a
843 * cacheless interface ( i.e. caching occurs at higher levels ).
844 * Therefore we do not maintain any resident pages. All I/O goes
845 * directly to and from the swap device.
847 * Note that b_blkno is scaled for PAGE_SIZE
849 * We currently attempt to run I/O synchronously or asynchronously as
850 * the caller requests. This isn't perfect because we loose error
851 * sequencing when we run multiple ops in parallel to satisfy a request.
852 * But this is swap, so we let it all hang out.
856 swap_pager_strategy(vm_object_t object, struct buf *bp)
862 struct buf *nbp = NULL;
864 if (bp->b_bcount & PAGE_MASK) {
865 bp->b_error = EINVAL;
866 bp->b_flags |= B_ERROR | B_INVAL;
868 printf("swap_pager_strategy: bp %p b_vp %p blk %d size %d, not page bounded\n", bp, bp->b_vp, (int)bp->b_pblkno, (int)bp->b_bcount);
873 * Clear error indication, initialize page index, count, data pointer.
877 bp->b_flags &= ~B_ERROR;
878 bp->b_resid = bp->b_bcount;
880 start = bp->b_pblkno;
881 count = howmany(bp->b_bcount, PAGE_SIZE);
887 * Deal with B_FREEBUF
890 if (bp->b_flags & B_FREEBUF) {
892 * FREE PAGE(s) - destroy underlying swap that is no longer
895 swp_pager_meta_free(object, start, count);
903 * Execute read or write
910 * Obtain block. If block not found and writing, allocate a
911 * new block and build it into the object.
914 blk = swp_pager_meta_ctl(object, start, 0);
915 if ((blk == SWAPBLK_NONE) && (bp->b_flags & B_READ) == 0) {
916 blk = swp_pager_getswapspace(1);
917 if (blk == SWAPBLK_NONE) {
918 bp->b_error = ENOMEM;
919 bp->b_flags |= B_ERROR;
922 swp_pager_meta_build(object, start, blk);
926 * Do we have to flush our current collection? Yes if:
928 * - no swap block at this index
929 * - swap block is not contiguous
930 * - we cross a physical disk boundry in the
935 nbp && (nbp->b_blkno + btoc(nbp->b_bcount) != blk ||
936 ((nbp->b_blkno ^ blk) & dmmax_mask)
940 if (bp->b_flags & B_READ) {
941 ++mycpu->gd_cnt.v_swapin;
942 mycpu->gd_cnt.v_swappgsin += btoc(nbp->b_bcount);
944 ++mycpu->gd_cnt.v_swapout;
945 mycpu->gd_cnt.v_swappgsout += btoc(nbp->b_bcount);
946 nbp->b_dirtyend = nbp->b_bcount;
954 * Add new swapblk to nbp, instantiating nbp if necessary.
955 * Zero-fill reads are able to take a shortcut.
958 if (blk == SWAPBLK_NONE) {
960 * We can only get here if we are reading. Since
961 * we are at splvm() we can safely modify b_resid,
962 * even if chain ops are in progress.
964 bzero(data, PAGE_SIZE);
965 bp->b_resid -= PAGE_SIZE;
968 nbp = getchainbuf(bp, swapdev_vp, (bp->b_flags & B_READ) | B_ASYNC);
973 nbp->b_bcount += PAGE_SIZE;
981 * Flush out last buffer
987 if ((bp->b_flags & B_ASYNC) == 0)
988 nbp->b_flags &= ~B_ASYNC;
989 if (nbp->b_flags & B_READ) {
990 ++mycpu->gd_cnt.v_swapin;
991 mycpu->gd_cnt.v_swappgsin += btoc(nbp->b_bcount);
993 ++mycpu->gd_cnt.v_swapout;
994 mycpu->gd_cnt.v_swappgsout += btoc(nbp->b_bcount);
995 nbp->b_dirtyend = nbp->b_bcount;
1002 * Wait for completion.
1005 if (bp->b_flags & B_ASYNC) {
1008 waitchainbuf(bp, 0, 1);
1013 * SWAP_PAGER_GETPAGES() - bring pages in from swap
1015 * Attempt to retrieve (m, count) pages from backing store, but make
1016 * sure we retrieve at least m[reqpage]. We try to load in as large
1017 * a chunk surrounding m[reqpage] as is contiguous in swap and which
1018 * belongs to the same object.
1020 * The code is designed for asynchronous operation and
1021 * immediate-notification of 'reqpage' but tends not to be
1022 * used that way. Please do not optimize-out this algorithmic
1023 * feature, I intend to improve on it in the future.
1025 * The parent has a single vm_object_pip_add() reference prior to
1026 * calling us and we should return with the same.
1028 * The parent has BUSY'd the pages. We should return with 'm'
1029 * left busy, but the others adjusted.
1033 swap_pager_getpages(vm_object_t object, vm_page_t *m, int count, int reqpage)
1042 vm_pindex_t lastpindex;
1046 if (mreq->object != object) {
1047 panic("swap_pager_getpages: object mismatch %p/%p",
1053 * Calculate range to retrieve. The pages have already been assigned
1054 * their swapblks. We require a *contiguous* range that falls entirely
1055 * within a single device stripe. If we do not supply it, bad things
1056 * happen. Note that blk, iblk & jblk can be SWAPBLK_NONE, but the
1057 * loops are set up such that the case(s) are handled implicitly.
1059 * The swp_*() calls must be made at splvm(). vm_page_free() does
1060 * not need to be, but it will go a little faster if it is.
1064 blk = swp_pager_meta_ctl(mreq->object, mreq->pindex, 0);
1066 for (i = reqpage - 1; i >= 0; --i) {
1069 iblk = swp_pager_meta_ctl(m[i]->object, m[i]->pindex, 0);
1070 if (blk != iblk + (reqpage - i))
1072 if ((blk ^ iblk) & dmmax_mask)
1077 for (j = reqpage + 1; j < count; ++j) {
1080 jblk = swp_pager_meta_ctl(m[j]->object, m[j]->pindex, 0);
1081 if (blk != jblk - (j - reqpage))
1083 if ((blk ^ jblk) & dmmax_mask)
1088 * free pages outside our collection range. Note: we never free
1089 * mreq, it must remain busy throughout.
1095 for (k = 0; k < i; ++k)
1097 for (k = j; k < count; ++k)
1104 * Return VM_PAGER_FAIL if we have nothing to do. Return mreq
1105 * still busy, but the others unbusied.
1108 if (blk == SWAPBLK_NONE)
1109 return(VM_PAGER_FAIL);
1112 * Get a swap buffer header to perform the IO
1115 bp = getpbuf(&nsw_rcount);
1116 kva = (vm_offset_t) bp->b_data;
1119 * map our page(s) into kva for input
1121 * NOTE: B_PAGING is set by pbgetvp()
1124 pmap_qenter(kva, m + i, j - i);
1126 bp->b_flags = B_READ | B_CALL;
1127 bp->b_iodone = swp_pager_async_iodone;
1128 bp->b_data = (caddr_t) kva;
1129 bp->b_blkno = blk - (reqpage - i);
1130 bp->b_bcount = PAGE_SIZE * (j - i);
1131 bp->b_bufsize = PAGE_SIZE * (j - i);
1132 bp->b_pager.pg_reqpage = reqpage - i;
1137 for (k = i; k < j; ++k) {
1138 bp->b_xio.xio_pages[k - i] = m[k];
1139 vm_page_flag_set(m[k], PG_SWAPINPROG);
1142 bp->b_xio.xio_npages = j - i;
1144 pbgetvp(swapdev_vp, bp);
1146 mycpu->gd_cnt.v_swapin++;
1147 mycpu->gd_cnt.v_swappgsin += bp->b_xio.xio_npages;
1150 * We still hold the lock on mreq, and our automatic completion routine
1151 * does not remove it.
1154 vm_object_pip_add(mreq->object, bp->b_xio.xio_npages);
1155 lastpindex = m[j-1]->pindex;
1158 * perform the I/O. NOTE!!! bp cannot be considered valid after
1159 * this point because we automatically release it on completion.
1160 * Instead, we look at the one page we are interested in which we
1161 * still hold a lock on even through the I/O completion.
1163 * The other pages in our m[] array are also released on completion,
1164 * so we cannot assume they are valid anymore either.
1166 * NOTE: b_blkno is destroyed by the call to VOP_STRATEGY
1170 VOP_STRATEGY(bp->b_vp, bp);
1173 * wait for the page we want to complete. PG_SWAPINPROG is always
1174 * cleared on completion. If an I/O error occurs, SWAPBLK_NONE
1175 * is set in the meta-data.
1180 while ((mreq->flags & PG_SWAPINPROG) != 0) {
1181 vm_page_flag_set(mreq, PG_WANTED | PG_REFERENCED);
1182 mycpu->gd_cnt.v_intrans++;
1183 if (tsleep(mreq, 0, "swread", hz*20)) {
1185 "swap_pager: indefinite wait buffer: device:"
1186 " %s, blkno: %ld, size: %ld\n",
1187 devtoname(bp->b_dev), (long)bp->b_blkno,
1196 * mreq is left bussied after completion, but all the other pages
1197 * are freed. If we had an unrecoverable read error the page will
1201 if (mreq->valid != VM_PAGE_BITS_ALL) {
1202 return(VM_PAGER_ERROR);
1204 return(VM_PAGER_OK);
1208 * A final note: in a low swap situation, we cannot deallocate swap
1209 * and mark a page dirty here because the caller is likely to mark
1210 * the page clean when we return, causing the page to possibly revert
1211 * to all-zero's later.
1216 * swap_pager_putpages:
1218 * Assign swap (if necessary) and initiate I/O on the specified pages.
1220 * We support both OBJT_DEFAULT and OBJT_SWAP objects. DEFAULT objects
1221 * are automatically converted to SWAP objects.
1223 * In a low memory situation we may block in VOP_STRATEGY(), but the new
1224 * vm_page reservation system coupled with properly written VFS devices
1225 * should ensure that no low-memory deadlock occurs. This is an area
1228 * The parent has N vm_object_pip_add() references prior to
1229 * calling us and will remove references for rtvals[] that are
1230 * not set to VM_PAGER_PEND. We need to remove the rest on I/O
1233 * The parent has soft-busy'd the pages it passes us and will unbusy
1234 * those whos rtvals[] entry is not set to VM_PAGER_PEND on return.
1235 * We need to unbusy the rest on I/O completion.
1239 swap_pager_putpages(vm_object_t object, vm_page_t *m, int count, boolean_t sync,
1245 if (count && m[0]->object != object) {
1246 panic("swap_pager_getpages: object mismatch %p/%p",
1254 * Turn object into OBJT_SWAP
1255 * check for bogus sysops
1256 * force sync if not pageout process
1259 if (object->type != OBJT_SWAP)
1260 swp_pager_meta_build(object, 0, SWAPBLK_NONE);
1262 if (curthread != pagethread)
1268 * Update nsw parameters from swap_async_max sysctl values.
1269 * Do not let the sysop crash the machine with bogus numbers.
1272 if (swap_async_max != nsw_wcount_async_max) {
1279 if ((n = swap_async_max) > nswbuf / 2)
1286 * Adjust difference ( if possible ). If the current async
1287 * count is too low, we may not be able to make the adjustment
1291 n -= nsw_wcount_async_max;
1292 if (nsw_wcount_async + n >= 0) {
1293 nsw_wcount_async += n;
1294 nsw_wcount_async_max += n;
1295 wakeup(&nsw_wcount_async);
1303 * Assign swap blocks and issue I/O. We reallocate swap on the fly.
1304 * The page is left dirty until the pageout operation completes
1308 for (i = 0; i < count; i += n) {
1315 * Maximum I/O size is limited by a number of factors.
1318 n = min(BLIST_MAX_ALLOC, count - i);
1319 n = min(n, nsw_cluster_max);
1324 * Get biggest block of swap we can. If we fail, fall
1325 * back and try to allocate a smaller block. Don't go
1326 * overboard trying to allocate space if it would overly
1330 (blk = swp_pager_getswapspace(n)) == SWAPBLK_NONE &&
1335 if (blk == SWAPBLK_NONE) {
1336 for (j = 0; j < n; ++j)
1337 rtvals[i+j] = VM_PAGER_FAIL;
1343 * The I/O we are constructing cannot cross a physical
1344 * disk boundry in the swap stripe. Note: we are still
1347 if ((blk ^ (blk + n)) & dmmax_mask) {
1348 j = ((blk + dmmax) & dmmax_mask) - blk;
1349 swp_pager_freeswapspace(blk + j, n - j);
1354 * All I/O parameters have been satisfied, build the I/O
1355 * request and assign the swap space.
1357 * NOTE: B_PAGING is set by pbgetvp()
1361 bp = getpbuf(&nsw_wcount_sync);
1362 bp->b_flags = B_CALL;
1364 bp = getpbuf(&nsw_wcount_async);
1365 bp->b_flags = B_CALL | B_ASYNC;
1367 bp->b_spc = NULL; /* not used, but NULL-out anyway */
1369 pmap_qenter((vm_offset_t)bp->b_data, &m[i], n);
1371 bp->b_bcount = PAGE_SIZE * n;
1372 bp->b_bufsize = PAGE_SIZE * n;
1375 pbgetvp(swapdev_vp, bp);
1377 for (j = 0; j < n; ++j) {
1378 vm_page_t mreq = m[i+j];
1380 swp_pager_meta_build(
1385 vm_page_dirty(mreq);
1386 rtvals[i+j] = VM_PAGER_OK;
1388 vm_page_flag_set(mreq, PG_SWAPINPROG);
1389 bp->b_xio.xio_pages[j] = mreq;
1391 bp->b_xio.xio_npages = n;
1393 * Must set dirty range for NFS to work.
1396 bp->b_dirtyend = bp->b_bcount;
1398 mycpu->gd_cnt.v_swapout++;
1399 mycpu->gd_cnt.v_swappgsout += bp->b_xio.xio_npages;
1400 swapdev_vp->v_numoutput++;
1407 * NOTE: b_blkno is destroyed by the call to VOP_STRATEGY
1410 if (sync == FALSE) {
1411 bp->b_iodone = swp_pager_async_iodone;
1413 VOP_STRATEGY(bp->b_vp, bp);
1415 for (j = 0; j < n; ++j)
1416 rtvals[i+j] = VM_PAGER_PEND;
1423 * NOTE: b_blkno is destroyed by the call to VOP_STRATEGY
1426 bp->b_iodone = swp_pager_sync_iodone;
1427 VOP_STRATEGY(bp->b_vp, bp);
1430 * Wait for the sync I/O to complete, then update rtvals.
1431 * We just set the rtvals[] to VM_PAGER_PEND so we can call
1432 * our async completion routine at the end, thus avoiding a
1437 while ((bp->b_flags & B_DONE) == 0) {
1438 tsleep(bp, 0, "swwrt", 0);
1441 for (j = 0; j < n; ++j)
1442 rtvals[i+j] = VM_PAGER_PEND;
1445 * Now that we are through with the bp, we can call the
1446 * normal async completion, which frees everything up.
1449 swp_pager_async_iodone(bp);
1456 * swap_pager_sync_iodone:
1458 * Completion routine for synchronous reads and writes from/to swap.
1459 * We just mark the bp is complete and wake up anyone waiting on it.
1461 * This routine may not block. This routine is called at splbio() or better.
1465 swp_pager_sync_iodone(struct buf *bp)
1467 bp->b_flags |= B_DONE;
1468 bp->b_flags &= ~B_ASYNC;
1473 * swp_pager_async_iodone:
1475 * Completion routine for asynchronous reads and writes from/to swap.
1476 * Also called manually by synchronous code to finish up a bp.
1478 * For READ operations, the pages are PG_BUSY'd. For WRITE operations,
1479 * the pages are vm_page_t->busy'd. For READ operations, we PG_BUSY
1480 * unbusy all pages except the 'main' request page. For WRITE
1481 * operations, we vm_page_t->busy'd unbusy all pages ( we can do this
1482 * because we marked them all VM_PAGER_PEND on return from putpages ).
1484 * This routine may not block.
1485 * This routine is called at splbio() or better
1487 * We up ourselves to splvm() as required for various vm_page related
1492 swp_pager_async_iodone(struct buf *bp)
1496 vm_object_t object = NULL;
1498 bp->b_flags |= B_DONE;
1504 if (bp->b_flags & B_ERROR) {
1506 "swap_pager: I/O error - %s failed; blkno %ld,"
1507 "size %ld, error %d\n",
1508 ((bp->b_flags & B_READ) ? "pagein" : "pageout"),
1516 * set object, raise to splvm().
1519 if (bp->b_xio.xio_npages)
1520 object = bp->b_xio.xio_pages[0]->object;
1524 * remove the mapping for kernel virtual
1527 pmap_qremove((vm_offset_t)bp->b_data, bp->b_xio.xio_npages);
1530 * cleanup pages. If an error occurs writing to swap, we are in
1531 * very serious trouble. If it happens to be a disk error, though,
1532 * we may be able to recover by reassigning the swap later on. So
1533 * in this case we remove the m->swapblk assignment for the page
1534 * but do not free it in the rlist. The errornous block(s) are thus
1535 * never reallocated as swap. Redirty the page and continue.
1538 for (i = 0; i < bp->b_xio.xio_npages; ++i) {
1539 vm_page_t m = bp->b_xio.xio_pages[i];
1541 vm_page_flag_clear(m, PG_SWAPINPROG);
1543 if (bp->b_flags & B_ERROR) {
1545 * If an error occurs I'd love to throw the swapblk
1546 * away without freeing it back to swapspace, so it
1547 * can never be used again. But I can't from an
1551 if (bp->b_flags & B_READ) {
1553 * When reading, reqpage needs to stay
1554 * locked for the parent, but all other
1555 * pages can be freed. We still want to
1556 * wakeup the parent waiting on the page,
1557 * though. ( also: pg_reqpage can be -1 and
1558 * not match anything ).
1560 * We have to wake specifically requested pages
1561 * up too because we cleared PG_SWAPINPROG and
1562 * someone may be waiting for that.
1564 * NOTE: for reads, m->dirty will probably
1565 * be overridden by the original caller of
1566 * getpages so don't play cute tricks here.
1568 * XXX IT IS NOT LEGAL TO FREE THE PAGE HERE
1569 * AS THIS MESSES WITH object->memq, and it is
1570 * not legal to mess with object->memq from an
1575 vm_page_flag_clear(m, PG_ZERO);
1577 if (i != bp->b_pager.pg_reqpage)
1582 * If i == bp->b_pager.pg_reqpage, do not wake
1583 * the page up. The caller needs to.
1587 * If a write error occurs, reactivate page
1588 * so it doesn't clog the inactive list,
1589 * then finish the I/O.
1592 vm_page_activate(m);
1593 vm_page_io_finish(m);
1595 } else if (bp->b_flags & B_READ) {
1597 * For read success, clear dirty bits. Nobody should
1598 * have this page mapped but don't take any chances,
1599 * make sure the pmap modify bits are also cleared.
1601 * NOTE: for reads, m->dirty will probably be
1602 * overridden by the original caller of getpages so
1603 * we cannot set them in order to free the underlying
1604 * swap in a low-swap situation. I don't think we'd
1605 * want to do that anyway, but it was an optimization
1606 * that existed in the old swapper for a time before
1607 * it got ripped out due to precisely this problem.
1609 * clear PG_ZERO in page.
1611 * If not the requested page then deactivate it.
1613 * Note that the requested page, reqpage, is left
1614 * busied, but we still have to wake it up. The
1615 * other pages are released (unbusied) by
1616 * vm_page_wakeup(). We do not set reqpage's
1617 * valid bits here, it is up to the caller.
1620 pmap_clear_modify(m);
1621 m->valid = VM_PAGE_BITS_ALL;
1623 vm_page_flag_clear(m, PG_ZERO);
1626 * We have to wake specifically requested pages
1627 * up too because we cleared PG_SWAPINPROG and
1628 * could be waiting for it in getpages. However,
1629 * be sure to not unbusy getpages specifically
1630 * requested page - getpages expects it to be
1633 if (i != bp->b_pager.pg_reqpage) {
1634 vm_page_deactivate(m);
1641 * For write success, clear the modify and dirty
1642 * status, then finish the I/O ( which decrements the
1643 * busy count and possibly wakes waiter's up ).
1645 pmap_clear_modify(m);
1647 vm_page_io_finish(m);
1648 if (!vm_page_count_severe() || !vm_page_try_to_cache(m))
1649 vm_page_protect(m, VM_PROT_READ);
1654 * adjust pip. NOTE: the original parent may still have its own
1655 * pip refs on the object.
1659 vm_object_pip_wakeupn(object, bp->b_xio.xio_npages);
1662 * release the physical I/O buffer
1667 ((bp->b_flags & B_READ) ? &nsw_rcount :
1668 ((bp->b_flags & B_ASYNC) ?
1677 /************************************************************************
1679 ************************************************************************
1681 * These routines manipulate the swap metadata stored in the
1682 * OBJT_SWAP object. All swp_*() routines must be called at
1683 * splvm() because swap can be freed up by the low level vm_page
1684 * code which might be called from interrupts beyond what splbio() covers.
1686 * Swap metadata is implemented with a global hash and not directly
1687 * linked into the object. Instead the object simply contains
1688 * appropriate tracking counters.
1692 * SWP_PAGER_HASH() - hash swap meta data
1694 * This is an inline helper function which hashes the swapblk given
1695 * the object and page index. It returns a pointer to a pointer
1696 * to the object, or a pointer to a NULL pointer if it could not
1699 * This routine must be called at splvm().
1702 static __inline struct swblock **
1703 swp_pager_hash(vm_object_t object, vm_pindex_t index)
1705 struct swblock **pswap;
1706 struct swblock *swap;
1708 index &= ~SWAP_META_MASK;
1709 pswap = &swhash[(index ^ (int)(intptr_t)object) & swhash_mask];
1711 while ((swap = *pswap) != NULL) {
1712 if (swap->swb_object == object &&
1713 swap->swb_index == index
1717 pswap = &swap->swb_hnext;
1723 * SWP_PAGER_META_BUILD() - add swap block to swap meta data for object
1725 * We first convert the object to a swap object if it is a default
1728 * The specified swapblk is added to the object's swap metadata. If
1729 * the swapblk is not valid, it is freed instead. Any previously
1730 * assigned swapblk is freed.
1732 * This routine must be called at splvm(), except when used to convert
1733 * an OBJT_DEFAULT object into an OBJT_SWAP object.
1738 swp_pager_meta_build(
1743 struct swblock *swap;
1744 struct swblock **pswap;
1747 * Convert default object to swap object if necessary
1750 if (object->type != OBJT_SWAP) {
1751 object->type = OBJT_SWAP;
1752 object->un_pager.swp.swp_bcount = 0;
1754 if (object->handle != NULL) {
1756 NOBJLIST(object->handle),
1762 &swap_pager_un_object_list,
1770 * Locate hash entry. If not found create, but if we aren't adding
1771 * anything just return. If we run out of space in the map we wait
1772 * and, since the hash table may have changed, retry.
1776 pswap = swp_pager_hash(object, index);
1778 if ((swap = *pswap) == NULL) {
1781 if (swapblk == SWAPBLK_NONE)
1784 swap = *pswap = zalloc(swap_zone);
1789 swap->swb_hnext = NULL;
1790 swap->swb_object = object;
1791 swap->swb_index = index & ~SWAP_META_MASK;
1792 swap->swb_count = 0;
1794 ++object->un_pager.swp.swp_bcount;
1796 for (i = 0; i < SWAP_META_PAGES; ++i)
1797 swap->swb_pages[i] = SWAPBLK_NONE;
1801 * Delete prior contents of metadata
1804 index &= SWAP_META_MASK;
1806 if (swap->swb_pages[index] != SWAPBLK_NONE) {
1807 swp_pager_freeswapspace(swap->swb_pages[index], 1);
1812 * Enter block into metadata
1815 swap->swb_pages[index] = swapblk;
1816 if (swapblk != SWAPBLK_NONE)
1821 * SWP_PAGER_META_FREE() - free a range of blocks in the object's swap metadata
1823 * The requested range of blocks is freed, with any associated swap
1824 * returned to the swap bitmap.
1826 * This routine will free swap metadata structures as they are cleaned
1827 * out. This routine does *NOT* operate on swap metadata associated
1828 * with resident pages.
1830 * This routine must be called at splvm()
1834 swp_pager_meta_free(vm_object_t object, vm_pindex_t index, daddr_t count)
1836 if (object->type != OBJT_SWAP)
1840 struct swblock **pswap;
1841 struct swblock *swap;
1843 pswap = swp_pager_hash(object, index);
1845 if ((swap = *pswap) != NULL) {
1846 daddr_t v = swap->swb_pages[index & SWAP_META_MASK];
1848 if (v != SWAPBLK_NONE) {
1849 swp_pager_freeswapspace(v, 1);
1850 swap->swb_pages[index & SWAP_META_MASK] =
1852 if (--swap->swb_count == 0) {
1853 *pswap = swap->swb_hnext;
1854 zfree(swap_zone, swap);
1855 --object->un_pager.swp.swp_bcount;
1861 int n = SWAP_META_PAGES - (index & SWAP_META_MASK);
1869 * SWP_PAGER_META_FREE_ALL() - destroy all swap metadata associated with object
1871 * This routine locates and destroys all swap metadata associated with
1874 * This routine must be called at splvm()
1878 swp_pager_meta_free_all(vm_object_t object)
1882 if (object->type != OBJT_SWAP)
1885 while (object->un_pager.swp.swp_bcount) {
1886 struct swblock **pswap;
1887 struct swblock *swap;
1889 pswap = swp_pager_hash(object, index);
1890 if ((swap = *pswap) != NULL) {
1893 for (i = 0; i < SWAP_META_PAGES; ++i) {
1894 daddr_t v = swap->swb_pages[i];
1895 if (v != SWAPBLK_NONE) {
1897 swp_pager_freeswapspace(v, 1);
1900 if (swap->swb_count != 0)
1901 panic("swap_pager_meta_free_all: swb_count != 0");
1902 *pswap = swap->swb_hnext;
1903 zfree(swap_zone, swap);
1904 --object->un_pager.swp.swp_bcount;
1906 index += SWAP_META_PAGES;
1907 if (index > 0x20000000)
1908 panic("swp_pager_meta_free_all: failed to locate all swap meta blocks");
1913 * SWP_PAGER_METACTL() - misc control of swap and vm_page_t meta data.
1915 * This routine is capable of looking up, popping, or freeing
1916 * swapblk assignments in the swap meta data or in the vm_page_t.
1917 * The routine typically returns the swapblk being looked-up, or popped,
1918 * or SWAPBLK_NONE if the block was freed, or SWAPBLK_NONE if the block
1919 * was invalid. This routine will automatically free any invalid
1920 * meta-data swapblks.
1922 * It is not possible to store invalid swapblks in the swap meta data
1923 * (other then a literal 'SWAPBLK_NONE'), so we don't bother checking.
1925 * When acting on a busy resident page and paging is in progress, we
1926 * have to wait until paging is complete but otherwise can act on the
1929 * This routine must be called at splvm().
1931 * SWM_FREE remove and free swap block from metadata
1932 * SWM_POP remove from meta data but do not free.. pop it out
1941 struct swblock **pswap;
1942 struct swblock *swap;
1946 * The meta data only exists of the object is OBJT_SWAP
1947 * and even then might not be allocated yet.
1950 if (object->type != OBJT_SWAP)
1951 return(SWAPBLK_NONE);
1954 pswap = swp_pager_hash(object, index);
1956 if ((swap = *pswap) != NULL) {
1957 index &= SWAP_META_MASK;
1958 r1 = swap->swb_pages[index];
1960 if (r1 != SWAPBLK_NONE) {
1961 if (flags & SWM_FREE) {
1962 swp_pager_freeswapspace(r1, 1);
1965 if (flags & (SWM_FREE|SWM_POP)) {
1966 swap->swb_pages[index] = SWAPBLK_NONE;
1967 if (--swap->swb_count == 0) {
1968 *pswap = swap->swb_hnext;
1969 zfree(swap_zone, swap);
1970 --object->un_pager.swp.swp_bcount;