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 1.20 2006/03/27 01:54:18 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/thread2.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, off_t size,
186 vm_prot_t prot, off_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 bio *);
192 static void swap_chain_iodone(struct bio *biox);
194 struct pagerops swappagerops = {
195 swap_pager_init, /* early system initialization of pager */
196 swap_pager_alloc, /* allocate an OBJT_SWAP object */
197 swap_pager_dealloc, /* deallocate an OBJT_SWAP object */
198 swap_pager_getpages, /* pagein */
199 swap_pager_putpages, /* pageout */
200 swap_pager_haspage, /* get backing store status for page */
201 swap_pager_unswapped, /* remove swap related to page */
202 swap_pager_strategy /* pager strategy call */
206 * dmmax is in page-sized chunks with the new swap system. It was
207 * dev-bsized chunks in the old. dmmax is always a power of 2.
209 * swap_*() routines are externally accessible. swp_*() routines are
214 static int dmmax_mask;
215 int nswap_lowat = 128; /* in pages, swap_pager_almost_full warn */
216 int nswap_hiwat = 512; /* in pages, swap_pager_almost_full warn */
218 static __inline void swp_sizecheck (void);
219 static void swp_pager_sync_iodone (struct bio *bio);
220 static void swp_pager_async_iodone (struct bio *bio);
223 * Swap bitmap functions
226 static __inline void swp_pager_freeswapspace (daddr_t blk, int npages);
227 static __inline daddr_t swp_pager_getswapspace (int npages);
233 static void swp_pager_meta_build (vm_object_t, vm_pindex_t, daddr_t);
234 static void swp_pager_meta_free (vm_object_t, vm_pindex_t, daddr_t);
235 static void swp_pager_meta_free_all (vm_object_t);
236 static daddr_t swp_pager_meta_ctl (vm_object_t, vm_pindex_t, int);
239 * SWP_SIZECHECK() - update swap_pager_full indication
241 * update the swap_pager_almost_full indication and warn when we are
242 * about to run out of swap space, using lowat/hiwat hysteresis.
244 * Clear swap_pager_full ( task killing ) indication when lowat is met.
246 * No restrictions on call
247 * This routine may not block.
248 * This routine must be called at splvm()
254 if (vm_swap_size < nswap_lowat) {
255 if (swap_pager_almost_full == 0) {
256 printf("swap_pager: out of swap space\n");
257 swap_pager_almost_full = 1;
261 if (vm_swap_size > nswap_hiwat)
262 swap_pager_almost_full = 0;
267 * SWAP_PAGER_INIT() - initialize the swap pager!
269 * Expected to be started from system init. NOTE: This code is run
270 * before much else so be careful what you depend on. Most of the VM
271 * system has yet to be initialized at this point.
275 swap_pager_init(void)
278 * Initialize object lists
282 for (i = 0; i < NOBJLISTS; ++i)
283 TAILQ_INIT(&swap_pager_object_list[i]);
284 TAILQ_INIT(&swap_pager_un_object_list);
287 * Device Stripe, in PAGE_SIZE'd blocks
290 dmmax = SWB_NPAGES * 2;
291 dmmax_mask = ~(dmmax - 1);
295 * SWAP_PAGER_SWAP_INIT() - swap pager initialization from pageout process
297 * Expected to be started from pageout process once, prior to entering
302 swap_pager_swap_init(void)
307 * Number of in-transit swap bp operations. Don't
308 * exhaust the pbufs completely. Make sure we
309 * initialize workable values (0 will work for hysteresis
310 * but it isn't very efficient).
312 * The nsw_cluster_max is constrained by the number of pages an XIO
313 * holds, i.e., (MAXPHYS/PAGE_SIZE) and our locally defined
314 * MAX_PAGEOUT_CLUSTER. Also be aware that swap ops are
315 * constrained by the swap device interleave stripe size.
317 * Currently we hardwire nsw_wcount_async to 4. This limit is
318 * designed to prevent other I/O from having high latencies due to
319 * our pageout I/O. The value 4 works well for one or two active swap
320 * devices but is probably a little low if you have more. Even so,
321 * a higher value would probably generate only a limited improvement
322 * with three or four active swap devices since the system does not
323 * typically have to pageout at extreme bandwidths. We will want
324 * at least 2 per swap devices, and 4 is a pretty good value if you
325 * have one NFS swap device due to the command/ack latency over NFS.
326 * So it all works out pretty well.
329 nsw_cluster_max = min((MAXPHYS/PAGE_SIZE), MAX_PAGEOUT_CLUSTER);
331 nsw_rcount = (nswbuf + 1) / 2;
332 nsw_wcount_sync = (nswbuf + 3) / 4;
333 nsw_wcount_async = 4;
334 nsw_wcount_async_max = nsw_wcount_async;
337 * Initialize our zone. Right now I'm just guessing on the number
338 * we need based on the number of pages in the system. Each swblock
339 * can hold 16 pages, so this is probably overkill. This reservation
340 * is typically limited to around 32MB by default.
342 n = vmstats.v_page_count / 2;
343 if (maxswzone && n > maxswzone / sizeof(struct swblock))
344 n = maxswzone / sizeof(struct swblock);
350 sizeof(struct swblock),
354 if (swap_zone != NULL)
357 * if the allocation failed, try a zone two thirds the
358 * size of the previous attempt.
363 if (swap_zone == NULL)
364 panic("swap_pager_swap_init: swap_zone == NULL");
366 printf("Swap zone entries reduced from %d to %d.\n", n2, n);
370 * Initialize our meta-data hash table. The swapper does not need to
371 * be quite as efficient as the VM system, so we do not use an
372 * oversized hash table.
374 * n: size of hash table, must be power of 2
375 * swhash_mask: hash table index mask
378 for (n = 1; n < n2 / 8; n *= 2)
381 swhash = malloc(sizeof(struct swblock *) * n, M_VMPGDATA, M_WAITOK);
382 bzero(swhash, sizeof(struct swblock *) * n);
388 * SWAP_PAGER_ALLOC() - allocate a new OBJT_SWAP VM object and instantiate
389 * its metadata structures.
391 * This routine is called from the mmap and fork code to create a new
392 * OBJT_SWAP object. We do this by creating an OBJT_DEFAULT object
393 * and then converting it with swp_pager_meta_build().
395 * This routine may block in vm_object_allocate() and create a named
396 * object lookup race, so we must interlock. We must also run at
397 * splvm() for the object lookup to handle races with interrupts, but
398 * we do not have to maintain splvm() in between the lookup and the
399 * add because (I believe) it is not possible to attempt to create
400 * a new swap object w/handle when a default object with that handle
405 swap_pager_alloc(void *handle, off_t size, vm_prot_t prot, off_t offset)
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)
466 * Remove from list right away so lookups will fail if we block for
467 * pageout completion.
470 if (object->handle == NULL) {
471 TAILQ_REMOVE(&swap_pager_un_object_list, object, pager_object_list);
473 TAILQ_REMOVE(NOBJLIST(object->handle), object, pager_object_list);
476 vm_object_pip_wait(object, "swpdea");
479 * Free all remaining metadata. We only bother to free it from
480 * the swap meta data. We do not attempt to free swapblk's still
481 * associated with vm_page_t's for this object. We do not care
482 * if paging is still in progress on some objects.
485 swp_pager_meta_free_all(object);
489 /************************************************************************
490 * SWAP PAGER BITMAP ROUTINES *
491 ************************************************************************/
494 * SWP_PAGER_GETSWAPSPACE() - allocate raw swap space
496 * Allocate swap for the requested number of pages. The starting
497 * swap block number (a page index) is returned or SWAPBLK_NONE
498 * if the allocation failed.
500 * Also has the side effect of advising that somebody made a mistake
501 * when they configured swap and didn't configure enough.
503 * Must be called at splvm() to avoid races with bitmap frees from
504 * vm_page_remove() aka swap_pager_page_removed().
506 * This routine may not block
507 * This routine must be called at splvm().
510 static __inline daddr_t
511 swp_pager_getswapspace(int npages)
515 if ((blk = blist_alloc(swapblist, npages)) == SWAPBLK_NONE) {
516 if (swap_pager_full != 2) {
517 printf("swap_pager_getswapspace: failed\n");
519 swap_pager_almost_full = 1;
522 vm_swap_size -= npages;
529 * SWP_PAGER_FREESWAPSPACE() - free raw swap space
531 * This routine returns the specified swap blocks back to the bitmap.
533 * Note: This routine may not block (it could in the old swap code),
534 * and through the use of the new blist routines it does not block.
536 * We must be called at splvm() to avoid races with bitmap frees from
537 * vm_page_remove() aka swap_pager_page_removed().
539 * This routine may not block
540 * This routine must be called at splvm().
544 swp_pager_freeswapspace(daddr_t blk, int npages)
546 blist_free(swapblist, blk, npages);
547 vm_swap_size += npages;
552 * SWAP_PAGER_FREESPACE() - frees swap blocks associated with a page
553 * range within an object.
555 * This is a globally accessible routine.
557 * This routine removes swapblk assignments from swap metadata.
559 * The external callers of this routine typically have already destroyed
560 * or renamed vm_page_t's associated with this range in the object so
563 * This routine may be called at any spl. We up our spl to splvm temporarily
564 * in order to perform the metadata removal.
568 swap_pager_freespace(vm_object_t object, vm_pindex_t start, vm_size_t size)
571 swp_pager_meta_free(object, start, size);
576 * SWAP_PAGER_RESERVE() - reserve swap blocks in object
578 * Assigns swap blocks to the specified range within the object. The
579 * swap blocks are not zerod. Any previous swap assignment is destroyed.
581 * Returns 0 on success, -1 on failure.
585 swap_pager_reserve(vm_object_t object, vm_pindex_t start, vm_size_t size)
588 daddr_t blk = SWAPBLK_NONE;
589 vm_pindex_t beg = start; /* save start index */
595 while ((blk = swp_pager_getswapspace(n)) == SWAPBLK_NONE) {
598 swp_pager_meta_free(object, beg, start - beg);
604 swp_pager_meta_build(object, start, blk);
610 swp_pager_meta_free(object, start, n);
616 * SWAP_PAGER_COPY() - copy blocks from source pager to destination pager
617 * and destroy the source.
619 * Copy any valid swapblks from the source to the destination. In
620 * cases where both the source and destination have a valid swapblk,
621 * we keep the destination's.
623 * This routine is allowed to block. It may block allocating metadata
624 * indirectly through swp_pager_meta_build() or if paging is still in
625 * progress on the source.
627 * This routine can be called at any spl
629 * XXX vm_page_collapse() kinda expects us not to block because we
630 * supposedly do not need to allocate memory, but for the moment we
631 * *may* have to get a little memory from the zone allocator, but
632 * it is taken from the interrupt memory. We should be ok.
634 * The source object contains no vm_page_t's (which is just as well)
636 * The source object is of type OBJT_SWAP.
638 * The source and destination objects must be locked or
639 * inaccessible (XXX are they ?)
643 swap_pager_copy(vm_object_t srcobject, vm_object_t dstobject,
644 vm_pindex_t offset, int destroysource)
651 * If destroysource is set, we remove the source object from the
652 * swap_pager internal queue now.
656 if (srcobject->handle == NULL) {
658 &swap_pager_un_object_list,
664 NOBJLIST(srcobject->handle),
672 * transfer source to destination.
675 for (i = 0; i < dstobject->size; ++i) {
679 * Locate (without changing) the swapblk on the destination,
680 * unless it is invalid in which case free it silently, or
681 * if the destination is a resident page, in which case the
682 * source is thrown away.
685 dstaddr = swp_pager_meta_ctl(dstobject, i, 0);
687 if (dstaddr == SWAPBLK_NONE) {
689 * Destination has no swapblk and is not resident,
694 srcaddr = swp_pager_meta_ctl(
700 if (srcaddr != SWAPBLK_NONE)
701 swp_pager_meta_build(dstobject, i, srcaddr);
704 * Destination has valid swapblk or it is represented
705 * by a resident page. We destroy the sourceblock.
708 swp_pager_meta_ctl(srcobject, i + offset, SWM_FREE);
713 * Free left over swap blocks in source.
715 * We have to revert the type to OBJT_DEFAULT so we do not accidently
716 * double-remove the object from the swap queues.
720 swp_pager_meta_free_all(srcobject);
722 * Reverting the type is not necessary, the caller is going
723 * to destroy srcobject directly, but I'm doing it here
724 * for consistency since we've removed the object from its
727 srcobject->type = OBJT_DEFAULT;
733 * SWAP_PAGER_HASPAGE() - determine if we have good backing store for
734 * the requested page.
736 * We determine whether good backing store exists for the requested
737 * page and return TRUE if it does, FALSE if it doesn't.
739 * If TRUE, we also try to determine how much valid, contiguous backing
740 * store exists before and after the requested page within a reasonable
741 * distance. We do not try to restrict it to the swap device stripe
742 * (that is handled in getpages/putpages). It probably isn't worth
747 swap_pager_haspage(vm_object_t object, vm_pindex_t pindex, int *before,
753 * do we have good backing store at the requested index ?
757 blk0 = swp_pager_meta_ctl(object, pindex, 0);
759 if (blk0 == SWAPBLK_NONE) {
769 * find backwards-looking contiguous good backing store
772 if (before != NULL) {
775 for (i = 1; i < (SWB_NPAGES/2); ++i) {
780 blk = swp_pager_meta_ctl(object, pindex - i, 0);
788 * find forward-looking contiguous good backing store
794 for (i = 1; i < (SWB_NPAGES/2); ++i) {
797 blk = swp_pager_meta_ctl(object, pindex + i, 0);
808 * SWAP_PAGER_PAGE_UNSWAPPED() - remove swap backing store related to page
810 * This removes any associated swap backing store, whether valid or
811 * not, from the page.
813 * This routine is typically called when a page is made dirty, at
814 * which point any associated swap can be freed. MADV_FREE also
815 * calls us in a special-case situation
817 * NOTE!!! If the page is clean and the swap was valid, the caller
818 * should make the page dirty before calling this routine. This routine
819 * does NOT change the m->dirty status of the page. Also: MADV_FREE
822 * This routine may not block
823 * This routine must be called at splvm()
827 swap_pager_unswapped(vm_page_t m)
829 swp_pager_meta_ctl(m->object, m->pindex, SWM_FREE);
833 * SWAP_PAGER_STRATEGY() - read, write, free blocks
835 * This implements the vm_pager_strategy() interface to swap and allows
836 * other parts of the system to directly access swap as backing store
837 * through vm_objects of type OBJT_SWAP. This is intended to be a
838 * cacheless interface ( i.e. caching occurs at higher levels ).
839 * Therefore we do not maintain any resident pages. All I/O goes
840 * directly to and from the swap device.
842 * We currently attempt to run I/O synchronously or asynchronously as
843 * the caller requests. This isn't perfect because we loose error
844 * sequencing when we run multiple ops in parallel to satisfy a request.
845 * But this is swap, so we let it all hang out.
849 swap_pager_strategy(vm_object_t object, struct bio *bio)
851 struct buf *bp = bio->bio_buf;
854 vm_pindex_t biox_blkno = 0;
857 struct bio *biox = NULL;
858 struct buf *bufx = NULL;
859 struct bio_track *track;
862 * tracking for swapdev vnode I/Os
864 if (bp->b_flags & B_READ)
865 track = &swapdev_vp->v_track_read;
867 track = &swapdev_vp->v_track_write;
869 if (bp->b_bcount & PAGE_MASK) {
870 bp->b_error = EINVAL;
871 bp->b_flags |= B_ERROR | B_INVAL;
873 printf("swap_pager_strategy: bp %p b_vp %p offset %lld size %d, not page bounded\n", bp, bp->b_vp, bio->bio_offset, (int)bp->b_bcount);
878 * Clear error indication, initialize page index, count, data pointer.
881 bp->b_flags &= ~B_ERROR;
882 bp->b_resid = bp->b_bcount;
884 start = (vm_pindex_t)(bio->bio_offset >> PAGE_SHIFT);
885 count = howmany(bp->b_bcount, PAGE_SIZE);
891 * Deal with B_FREEBUF
893 if (bp->b_flags & B_FREEBUF) {
895 * FREE PAGE(s) - destroy underlying swap that is no longer
898 swp_pager_meta_free(object, start, count);
906 * We need to be able to create a new cluster of I/O's. We cannot
907 * use the caller fields of the passed bio so push a new one.
909 * Because nbio is just a placeholder for the cluster links,
910 * we can biodone() the original bio instead of nbio to make
911 * things a bit more efficient.
913 nbio = push_bio(bio);
914 nbio->bio_offset = bio->bio_offset;
915 nbio->bio_caller_info1.cluster_head = NULL;
916 nbio->bio_caller_info2.cluster_tail = NULL;
919 * Execute read or write
926 * Obtain block. If block not found and writing, allocate a
927 * new block and build it into the object.
930 blk = swp_pager_meta_ctl(object, start, 0);
931 if ((blk == SWAPBLK_NONE) && (bp->b_flags & B_READ) == 0) {
932 blk = swp_pager_getswapspace(1);
933 if (blk == SWAPBLK_NONE) {
934 bp->b_error = ENOMEM;
935 bp->b_flags |= B_ERROR;
938 swp_pager_meta_build(object, start, blk);
942 * Do we have to flush our current collection? Yes if:
944 * - no swap block at this index
945 * - swap block is not contiguous
946 * - we cross a physical disk boundry in the
951 biox && (biox_blkno + btoc(bufx->b_bcount) != blk ||
952 ((biox_blkno ^ blk) & dmmax_mask)
956 if (bp->b_flags & B_READ) {
957 ++mycpu->gd_cnt.v_swapin;
958 mycpu->gd_cnt.v_swappgsin += btoc(bufx->b_bcount);
960 ++mycpu->gd_cnt.v_swapout;
961 mycpu->gd_cnt.v_swappgsout += btoc(bufx->b_bcount);
962 bufx->b_dirtyend = bufx->b_bcount;
966 * Flush the biox to the swap device.
968 if (bufx->b_bcount) {
969 bufx->b_bufsize = bufx->b_bcount;
970 if ((bufx->b_flags & B_READ) == 0)
971 bufx->b_dirtyend = bufx->b_bcount;
973 vn_strategy(bufx->b_vp, biox);
983 * Add new swapblk to biox, instantiating biox if necessary.
984 * Zero-fill reads are able to take a shortcut.
986 if (blk == SWAPBLK_NONE) {
988 * We can only get here if we are reading. Since
989 * we are at splvm() we can safely modify b_resid,
990 * even if chain ops are in progress.
992 bzero(data, PAGE_SIZE);
993 bp->b_resid -= PAGE_SIZE;
996 /* XXX chain count > 4, wait to <= 4 */
998 bufx = getpbuf(NULL);
999 biox = &bufx->b_bio1;
1000 cluster_append(nbio, bufx);
1001 bufx->b_flags = (bufx->b_flags & B_ORDERED) |
1002 (bp->b_flags & B_READ) |
1004 pbgetvp(swapdev_vp, bufx);
1005 biox->bio_done = swap_chain_iodone;
1006 biox->bio_offset = (off_t)blk << PAGE_SHIFT;
1007 biox->bio_caller_info1.cluster_parent = nbio;
1010 bufx->b_data = data;
1012 bufx->b_bcount += PAGE_SIZE;
1020 * Flush out last buffer
1025 if ((bp->b_flags & B_ASYNC) == 0)
1026 bufx->b_flags &= ~B_ASYNC;
1027 if (bufx->b_flags & B_READ) {
1028 ++mycpu->gd_cnt.v_swapin;
1029 mycpu->gd_cnt.v_swappgsin += btoc(bufx->b_bcount);
1031 ++mycpu->gd_cnt.v_swapout;
1032 mycpu->gd_cnt.v_swappgsout += btoc(bufx->b_bcount);
1033 bufx->b_dirtyend = bufx->b_bcount;
1035 if (bufx->b_bcount) {
1036 bufx->b_bufsize = bufx->b_bcount;
1037 if ((bufx->b_flags & B_READ) == 0)
1038 bufx->b_dirtyend = bufx->b_bcount;
1040 vn_strategy(bufx->b_vp, biox);
1044 /* biox, bufx = NULL */
1048 * Wait for completion.
1050 if (bp->b_flags & B_ASYNC) {
1052 if (nbio->bio_caller_info1.cluster_head == NULL) {
1055 bp->b_xflags |= BX_AUTOCHAINDONE;
1060 while (nbio->bio_caller_info1.cluster_head != NULL) {
1061 bp->b_flags |= B_WANT;
1062 tsleep(bp, 0, "bpchain", 0);
1064 if (bp->b_resid != 0 && !(bp->b_flags & B_ERROR)) {
1065 bp->b_flags |= B_ERROR;
1066 bp->b_error = EINVAL;
1074 swap_chain_iodone(struct bio *biox)
1077 struct buf *bufx; /* chained sub-buffer */
1078 struct bio *nbio; /* parent nbio with chain glue */
1079 struct buf *bp; /* original bp associated with nbio */
1081 bufx = biox->bio_buf;
1082 nbio = biox->bio_caller_info1.cluster_parent;
1086 * Update the original buffer
1088 KKASSERT(bp != NULL);
1089 if (bufx->b_flags & B_ERROR) {
1090 bp->b_flags |= B_ERROR;
1091 bp->b_error = bufx->b_error;
1092 } else if (bufx->b_resid != 0) {
1093 bp->b_flags |= B_ERROR;
1094 bp->b_error = EINVAL;
1096 bp->b_resid -= bufx->b_bcount;
1100 * Remove us from the chain. It is sufficient to clean up
1101 * cluster_head. We do not have to clean up cluster_tail.
1103 nextp = &nbio->bio_caller_info1.cluster_head;
1104 while (*nextp != bufx) {
1105 KKASSERT(*nextp != NULL);
1106 nextp = &(*nextp)->b_cluster_next;
1108 *nextp = bufx->b_cluster_next;
1109 if (bp->b_flags & B_WANT) {
1110 bp->b_flags &= ~B_WANT;
1115 * Clean up bufx. If this was the last buffer in the chain
1116 * and BX_AUTOCHAINDONE was set, finish off the original I/O
1119 * nbio was just a fake BIO layer to hold the cluster links,
1120 * we can issue the biodone() on the layer above it.
1122 if (nbio->bio_caller_info1.cluster_head == NULL &&
1123 (bp->b_xflags & BX_AUTOCHAINDONE)) {
1124 bp->b_xflags &= ~BX_AUTOCHAINDONE;
1125 if (bp->b_resid != 0 && !(bp->b_flags & B_ERROR)) {
1126 bp->b_flags |= B_ERROR;
1127 bp->b_error = EINVAL;
1129 biodone(nbio->bio_prev);
1131 bufx->b_flags |= B_DONE;
1132 bufx->b_flags &= ~B_ASYNC;
1133 relpbuf(bufx, NULL);
1137 * SWAP_PAGER_GETPAGES() - bring pages in from swap
1139 * Attempt to retrieve (m, count) pages from backing store, but make
1140 * sure we retrieve at least m[reqpage]. We try to load in as large
1141 * a chunk surrounding m[reqpage] as is contiguous in swap and which
1142 * belongs to the same object.
1144 * The code is designed for asynchronous operation and
1145 * immediate-notification of 'reqpage' but tends not to be
1146 * used that way. Please do not optimize-out this algorithmic
1147 * feature, I intend to improve on it in the future.
1149 * The parent has a single vm_object_pip_add() reference prior to
1150 * calling us and we should return with the same.
1152 * The parent has BUSY'd the pages. We should return with 'm'
1153 * left busy, but the others adjusted.
1157 swap_pager_getpages(vm_object_t object, vm_page_t *m, int count, int reqpage)
1166 vm_pindex_t lastpindex;
1170 if (mreq->object != object) {
1171 panic("swap_pager_getpages: object mismatch %p/%p",
1177 * Calculate range to retrieve. The pages have already been assigned
1178 * their swapblks. We require a *contiguous* range that falls entirely
1179 * within a single device stripe. If we do not supply it, bad things
1180 * happen. Note that blk, iblk & jblk can be SWAPBLK_NONE, but the
1181 * loops are set up such that the case(s) are handled implicitly.
1183 * The swp_*() calls must be made at splvm(). vm_page_free() does
1184 * not need to be, but it will go a little faster if it is.
1188 blk = swp_pager_meta_ctl(mreq->object, mreq->pindex, 0);
1190 for (i = reqpage - 1; i >= 0; --i) {
1193 iblk = swp_pager_meta_ctl(m[i]->object, m[i]->pindex, 0);
1194 if (blk != iblk + (reqpage - i))
1196 if ((blk ^ iblk) & dmmax_mask)
1201 for (j = reqpage + 1; j < count; ++j) {
1204 jblk = swp_pager_meta_ctl(m[j]->object, m[j]->pindex, 0);
1205 if (blk != jblk - (j - reqpage))
1207 if ((blk ^ jblk) & dmmax_mask)
1212 * free pages outside our collection range. Note: we never free
1213 * mreq, it must remain busy throughout.
1219 for (k = 0; k < i; ++k)
1221 for (k = j; k < count; ++k)
1228 * Return VM_PAGER_FAIL if we have nothing to do. Return mreq
1229 * still busy, but the others unbusied.
1232 if (blk == SWAPBLK_NONE)
1233 return(VM_PAGER_FAIL);
1236 * Get a swap buffer header to perform the IO
1239 bp = getpbuf(&nsw_rcount);
1241 kva = (vm_offset_t) bp->b_data;
1244 * map our page(s) into kva for input
1246 * NOTE: B_PAGING is set by pbgetvp()
1249 pmap_qenter(kva, m + i, j - i);
1251 bp->b_flags = B_READ;
1252 bp->b_data = (caddr_t) kva;
1253 bp->b_bcount = PAGE_SIZE * (j - i);
1254 bp->b_bufsize = PAGE_SIZE * (j - i);
1255 bio->bio_done = swp_pager_async_iodone;
1256 bio->bio_offset = (off_t)(blk - (reqpage - i)) << PAGE_SHIFT;
1257 bio->bio_driver_info = (void *)(reqpage - i);
1262 for (k = i; k < j; ++k) {
1263 bp->b_xio.xio_pages[k - i] = m[k];
1264 vm_page_flag_set(m[k], PG_SWAPINPROG);
1267 bp->b_xio.xio_npages = j - i;
1269 pbgetvp(swapdev_vp, bp);
1271 mycpu->gd_cnt.v_swapin++;
1272 mycpu->gd_cnt.v_swappgsin += bp->b_xio.xio_npages;
1275 * We still hold the lock on mreq, and our automatic completion routine
1276 * does not remove it.
1279 vm_object_pip_add(mreq->object, bp->b_xio.xio_npages);
1280 lastpindex = m[j-1]->pindex;
1283 * perform the I/O. NOTE!!! bp cannot be considered valid after
1284 * this point because we automatically release it on completion.
1285 * Instead, we look at the one page we are interested in which we
1286 * still hold a lock on even through the I/O completion.
1288 * The other pages in our m[] array are also released on completion,
1289 * so we cannot assume they are valid anymore either.
1293 vn_strategy(swapdev_vp, bio);
1296 * wait for the page we want to complete. PG_SWAPINPROG is always
1297 * cleared on completion. If an I/O error occurs, SWAPBLK_NONE
1298 * is set in the meta-data.
1303 while ((mreq->flags & PG_SWAPINPROG) != 0) {
1304 vm_page_flag_set(mreq, PG_WANTED | PG_REFERENCED);
1305 mycpu->gd_cnt.v_intrans++;
1306 if (tsleep(mreq, 0, "swread", hz*20)) {
1308 "swap_pager: indefinite wait buffer: "
1309 " offset: %lld, size: %d\n",
1310 bio->bio_offset, bp->b_bcount
1318 * mreq is left bussied after completion, but all the other pages
1319 * are freed. If we had an unrecoverable read error the page will
1323 if (mreq->valid != VM_PAGE_BITS_ALL) {
1324 return(VM_PAGER_ERROR);
1326 return(VM_PAGER_OK);
1330 * A final note: in a low swap situation, we cannot deallocate swap
1331 * and mark a page dirty here because the caller is likely to mark
1332 * the page clean when we return, causing the page to possibly revert
1333 * to all-zero's later.
1338 * swap_pager_putpages:
1340 * Assign swap (if necessary) and initiate I/O on the specified pages.
1342 * We support both OBJT_DEFAULT and OBJT_SWAP objects. DEFAULT objects
1343 * are automatically converted to SWAP objects.
1345 * In a low memory situation we may block in vn_strategy(), but the new
1346 * vm_page reservation system coupled with properly written VFS devices
1347 * should ensure that no low-memory deadlock occurs. This is an area
1350 * The parent has N vm_object_pip_add() references prior to
1351 * calling us and will remove references for rtvals[] that are
1352 * not set to VM_PAGER_PEND. We need to remove the rest on I/O
1355 * The parent has soft-busy'd the pages it passes us and will unbusy
1356 * those whos rtvals[] entry is not set to VM_PAGER_PEND on return.
1357 * We need to unbusy the rest on I/O completion.
1361 swap_pager_putpages(vm_object_t object, vm_page_t *m, int count, boolean_t sync,
1367 if (count && m[0]->object != object) {
1368 panic("swap_pager_getpages: object mismatch %p/%p",
1376 * Turn object into OBJT_SWAP
1377 * check for bogus sysops
1378 * force sync if not pageout process
1381 if (object->type != OBJT_SWAP)
1382 swp_pager_meta_build(object, 0, SWAPBLK_NONE);
1384 if (curthread != pagethread)
1390 * Update nsw parameters from swap_async_max sysctl values.
1391 * Do not let the sysop crash the machine with bogus numbers.
1394 if (swap_async_max != nsw_wcount_async_max) {
1400 if ((n = swap_async_max) > nswbuf / 2)
1407 * Adjust difference ( if possible ). If the current async
1408 * count is too low, we may not be able to make the adjustment
1412 n -= nsw_wcount_async_max;
1413 if (nsw_wcount_async + n >= 0) {
1414 nsw_wcount_async += n;
1415 nsw_wcount_async_max += n;
1416 wakeup(&nsw_wcount_async);
1424 * Assign swap blocks and issue I/O. We reallocate swap on the fly.
1425 * The page is left dirty until the pageout operation completes
1429 for (i = 0; i < count; i += n) {
1436 * Maximum I/O size is limited by a number of factors.
1439 n = min(BLIST_MAX_ALLOC, count - i);
1440 n = min(n, nsw_cluster_max);
1445 * Get biggest block of swap we can. If we fail, fall
1446 * back and try to allocate a smaller block. Don't go
1447 * overboard trying to allocate space if it would overly
1451 (blk = swp_pager_getswapspace(n)) == SWAPBLK_NONE &&
1456 if (blk == SWAPBLK_NONE) {
1457 for (j = 0; j < n; ++j)
1458 rtvals[i+j] = VM_PAGER_FAIL;
1464 * The I/O we are constructing cannot cross a physical
1465 * disk boundry in the swap stripe. Note: we are still
1468 if ((blk ^ (blk + n)) & dmmax_mask) {
1469 j = ((blk + dmmax) & dmmax_mask) - blk;
1470 swp_pager_freeswapspace(blk + j, n - j);
1475 * All I/O parameters have been satisfied, build the I/O
1476 * request and assign the swap space.
1478 * NOTE: B_PAGING is set by pbgetvp()
1482 bp = getpbuf(&nsw_wcount_sync);
1484 bp = getpbuf(&nsw_wcount_async);
1485 bp->b_flags = B_ASYNC;
1489 pmap_qenter((vm_offset_t)bp->b_data, &m[i], n);
1491 bp->b_bcount = PAGE_SIZE * n;
1492 bp->b_bufsize = PAGE_SIZE * n;
1493 bio->bio_offset = (off_t)blk << PAGE_SHIFT;
1495 pbgetvp(swapdev_vp, bp);
1497 for (j = 0; j < n; ++j) {
1498 vm_page_t mreq = m[i+j];
1500 swp_pager_meta_build(
1505 vm_page_dirty(mreq);
1506 rtvals[i+j] = VM_PAGER_OK;
1508 vm_page_flag_set(mreq, PG_SWAPINPROG);
1509 bp->b_xio.xio_pages[j] = mreq;
1511 bp->b_xio.xio_npages = n;
1513 * Must set dirty range for NFS to work.
1516 bp->b_dirtyend = bp->b_bcount;
1518 mycpu->gd_cnt.v_swapout++;
1519 mycpu->gd_cnt.v_swappgsout += bp->b_xio.xio_npages;
1527 if (sync == FALSE) {
1528 bio->bio_done = swp_pager_async_iodone;
1530 vn_strategy(swapdev_vp, bio);
1532 for (j = 0; j < n; ++j)
1533 rtvals[i+j] = VM_PAGER_PEND;
1541 bio->bio_done = swp_pager_sync_iodone;
1542 vn_strategy(swapdev_vp, bio);
1545 * Wait for the sync I/O to complete, then update rtvals.
1546 * We just set the rtvals[] to VM_PAGER_PEND so we can call
1547 * our async completion routine at the end, thus avoiding a
1552 while ((bp->b_flags & B_DONE) == 0) {
1553 tsleep(bp, 0, "swwrt", 0);
1556 for (j = 0; j < n; ++j)
1557 rtvals[i+j] = VM_PAGER_PEND;
1560 * Now that we are through with the bp, we can call the
1561 * normal async completion, which frees everything up.
1564 swp_pager_async_iodone(bio);
1571 * swap_pager_sync_iodone:
1573 * Completion routine for synchronous reads and writes from/to swap.
1574 * We just mark the bp is complete and wake up anyone waiting on it.
1576 * This routine may not block. This routine is called at splbio() or better.
1580 swp_pager_sync_iodone(struct bio *bio)
1582 struct buf *bp = bio->bio_buf;
1584 bp->b_flags |= B_DONE;
1585 bp->b_flags &= ~B_ASYNC;
1590 * swp_pager_async_iodone:
1592 * Completion routine for asynchronous reads and writes from/to swap.
1593 * Also called manually by synchronous code to finish up a bp.
1595 * For READ operations, the pages are PG_BUSY'd. For WRITE operations,
1596 * the pages are vm_page_t->busy'd. For READ operations, we PG_BUSY
1597 * unbusy all pages except the 'main' request page. For WRITE
1598 * operations, we vm_page_t->busy'd unbusy all pages ( we can do this
1599 * because we marked them all VM_PAGER_PEND on return from putpages ).
1601 * This routine may not block.
1605 swp_pager_async_iodone(struct bio *bio)
1607 struct buf *bp = bio->bio_buf;
1608 vm_object_t object = NULL;
1611 bp->b_flags |= B_DONE;
1617 if (bp->b_flags & B_ERROR) {
1619 "swap_pager: I/O error - %s failed; offset %lld,"
1620 "size %ld, error %d\n",
1621 ((bp->b_flags & B_READ) ? "pagein" : "pageout"),
1629 * set object, raise to splvm().
1632 if (bp->b_xio.xio_npages)
1633 object = bp->b_xio.xio_pages[0]->object;
1637 * remove the mapping for kernel virtual
1640 pmap_qremove((vm_offset_t)bp->b_data, bp->b_xio.xio_npages);
1643 * cleanup pages. If an error occurs writing to swap, we are in
1644 * very serious trouble. If it happens to be a disk error, though,
1645 * we may be able to recover by reassigning the swap later on. So
1646 * in this case we remove the m->swapblk assignment for the page
1647 * but do not free it in the rlist. The errornous block(s) are thus
1648 * never reallocated as swap. Redirty the page and continue.
1651 for (i = 0; i < bp->b_xio.xio_npages; ++i) {
1652 vm_page_t m = bp->b_xio.xio_pages[i];
1654 vm_page_flag_clear(m, PG_SWAPINPROG);
1656 if (bp->b_flags & B_ERROR) {
1658 * If an error occurs I'd love to throw the swapblk
1659 * away without freeing it back to swapspace, so it
1660 * can never be used again. But I can't from an
1664 if (bp->b_flags & B_READ) {
1666 * When reading, reqpage needs to stay
1667 * locked for the parent, but all other
1668 * pages can be freed. We still want to
1669 * wakeup the parent waiting on the page,
1670 * though. ( also: pg_reqpage can be -1 and
1671 * not match anything ).
1673 * We have to wake specifically requested pages
1674 * up too because we cleared PG_SWAPINPROG and
1675 * someone may be waiting for that.
1677 * NOTE: for reads, m->dirty will probably
1678 * be overridden by the original caller of
1679 * getpages so don't play cute tricks here.
1681 * XXX IT IS NOT LEGAL TO FREE THE PAGE HERE
1682 * AS THIS MESSES WITH object->memq, and it is
1683 * not legal to mess with object->memq from an
1688 vm_page_flag_clear(m, PG_ZERO);
1691 * bio_driver_info holds the requested page
1694 if (i != (int)bio->bio_driver_info)
1699 * If i == bp->b_pager.pg_reqpage, do not wake
1700 * the page up. The caller needs to.
1704 * If a write error occurs, reactivate page
1705 * so it doesn't clog the inactive list,
1706 * then finish the I/O.
1709 vm_page_activate(m);
1710 vm_page_io_finish(m);
1712 } else if (bp->b_flags & B_READ) {
1714 * For read success, clear dirty bits. Nobody should
1715 * have this page mapped but don't take any chances,
1716 * make sure the pmap modify bits are also cleared.
1718 * NOTE: for reads, m->dirty will probably be
1719 * overridden by the original caller of getpages so
1720 * we cannot set them in order to free the underlying
1721 * swap in a low-swap situation. I don't think we'd
1722 * want to do that anyway, but it was an optimization
1723 * that existed in the old swapper for a time before
1724 * it got ripped out due to precisely this problem.
1726 * clear PG_ZERO in page.
1728 * If not the requested page then deactivate it.
1730 * Note that the requested page, reqpage, is left
1731 * busied, but we still have to wake it up. The
1732 * other pages are released (unbusied) by
1733 * vm_page_wakeup(). We do not set reqpage's
1734 * valid bits here, it is up to the caller.
1737 pmap_clear_modify(m);
1738 m->valid = VM_PAGE_BITS_ALL;
1740 vm_page_flag_clear(m, PG_ZERO);
1743 * We have to wake specifically requested pages
1744 * up too because we cleared PG_SWAPINPROG and
1745 * could be waiting for it in getpages. However,
1746 * be sure to not unbusy getpages specifically
1747 * requested page - getpages expects it to be
1750 * bio_driver_info holds the requested page
1752 if (i != (int)bio->bio_driver_info) {
1753 vm_page_deactivate(m);
1760 * For write success, clear the modify and dirty
1761 * status, then finish the I/O ( which decrements the
1762 * busy count and possibly wakes waiter's up ).
1764 pmap_clear_modify(m);
1766 vm_page_io_finish(m);
1767 if (!vm_page_count_severe() || !vm_page_try_to_cache(m))
1768 vm_page_protect(m, VM_PROT_READ);
1773 * adjust pip. NOTE: the original parent may still have its own
1774 * pip refs on the object.
1778 vm_object_pip_wakeupn(object, bp->b_xio.xio_npages);
1781 * release the physical I/O buffer
1786 ((bp->b_flags & B_READ) ? &nsw_rcount :
1787 ((bp->b_flags & B_ASYNC) ?
1796 /************************************************************************
1798 ************************************************************************
1800 * These routines manipulate the swap metadata stored in the
1801 * OBJT_SWAP object. All swp_*() routines must be called at
1802 * splvm() because swap can be freed up by the low level vm_page
1803 * code which might be called from interrupts beyond what splbio() covers.
1805 * Swap metadata is implemented with a global hash and not directly
1806 * linked into the object. Instead the object simply contains
1807 * appropriate tracking counters.
1811 * SWP_PAGER_HASH() - hash swap meta data
1813 * This is an inline helper function which hashes the swapblk given
1814 * the object and page index. It returns a pointer to a pointer
1815 * to the object, or a pointer to a NULL pointer if it could not
1818 * This routine must be called at splvm().
1821 static __inline struct swblock **
1822 swp_pager_hash(vm_object_t object, vm_pindex_t index)
1824 struct swblock **pswap;
1825 struct swblock *swap;
1827 index &= ~SWAP_META_MASK;
1828 pswap = &swhash[(index ^ (int)(intptr_t)object) & swhash_mask];
1830 while ((swap = *pswap) != NULL) {
1831 if (swap->swb_object == object &&
1832 swap->swb_index == index
1836 pswap = &swap->swb_hnext;
1842 * SWP_PAGER_META_BUILD() - add swap block to swap meta data for object
1844 * We first convert the object to a swap object if it is a default
1847 * The specified swapblk is added to the object's swap metadata. If
1848 * the swapblk is not valid, it is freed instead. Any previously
1849 * assigned swapblk is freed.
1851 * This routine must be called at splvm(), except when used to convert
1852 * an OBJT_DEFAULT object into an OBJT_SWAP object.
1857 swp_pager_meta_build(
1862 struct swblock *swap;
1863 struct swblock **pswap;
1866 * Convert default object to swap object if necessary
1869 if (object->type != OBJT_SWAP) {
1870 object->type = OBJT_SWAP;
1871 object->un_pager.swp.swp_bcount = 0;
1873 if (object->handle != NULL) {
1875 NOBJLIST(object->handle),
1881 &swap_pager_un_object_list,
1889 * Locate hash entry. If not found create, but if we aren't adding
1890 * anything just return. If we run out of space in the map we wait
1891 * and, since the hash table may have changed, retry.
1895 pswap = swp_pager_hash(object, index);
1897 if ((swap = *pswap) == NULL) {
1900 if (swapblk == SWAPBLK_NONE)
1903 swap = *pswap = zalloc(swap_zone);
1908 swap->swb_hnext = NULL;
1909 swap->swb_object = object;
1910 swap->swb_index = index & ~SWAP_META_MASK;
1911 swap->swb_count = 0;
1913 ++object->un_pager.swp.swp_bcount;
1915 for (i = 0; i < SWAP_META_PAGES; ++i)
1916 swap->swb_pages[i] = SWAPBLK_NONE;
1920 * Delete prior contents of metadata
1923 index &= SWAP_META_MASK;
1925 if (swap->swb_pages[index] != SWAPBLK_NONE) {
1926 swp_pager_freeswapspace(swap->swb_pages[index], 1);
1931 * Enter block into metadata
1934 swap->swb_pages[index] = swapblk;
1935 if (swapblk != SWAPBLK_NONE)
1940 * SWP_PAGER_META_FREE() - free a range of blocks in the object's swap metadata
1942 * The requested range of blocks is freed, with any associated swap
1943 * returned to the swap bitmap.
1945 * This routine will free swap metadata structures as they are cleaned
1946 * out. This routine does *NOT* operate on swap metadata associated
1947 * with resident pages.
1949 * This routine must be called at splvm()
1953 swp_pager_meta_free(vm_object_t object, vm_pindex_t index, daddr_t count)
1955 if (object->type != OBJT_SWAP)
1959 struct swblock **pswap;
1960 struct swblock *swap;
1962 pswap = swp_pager_hash(object, index);
1964 if ((swap = *pswap) != NULL) {
1965 daddr_t v = swap->swb_pages[index & SWAP_META_MASK];
1967 if (v != SWAPBLK_NONE) {
1968 swp_pager_freeswapspace(v, 1);
1969 swap->swb_pages[index & SWAP_META_MASK] =
1971 if (--swap->swb_count == 0) {
1972 *pswap = swap->swb_hnext;
1973 zfree(swap_zone, swap);
1974 --object->un_pager.swp.swp_bcount;
1980 int n = SWAP_META_PAGES - (index & SWAP_META_MASK);
1988 * SWP_PAGER_META_FREE_ALL() - destroy all swap metadata associated with object
1990 * This routine locates and destroys all swap metadata associated with
1993 * This routine must be called at splvm()
1997 swp_pager_meta_free_all(vm_object_t object)
2001 if (object->type != OBJT_SWAP)
2004 while (object->un_pager.swp.swp_bcount) {
2005 struct swblock **pswap;
2006 struct swblock *swap;
2008 pswap = swp_pager_hash(object, index);
2009 if ((swap = *pswap) != NULL) {
2012 for (i = 0; i < SWAP_META_PAGES; ++i) {
2013 daddr_t v = swap->swb_pages[i];
2014 if (v != SWAPBLK_NONE) {
2016 swp_pager_freeswapspace(v, 1);
2019 if (swap->swb_count != 0)
2020 panic("swap_pager_meta_free_all: swb_count != 0");
2021 *pswap = swap->swb_hnext;
2022 zfree(swap_zone, swap);
2023 --object->un_pager.swp.swp_bcount;
2025 index += SWAP_META_PAGES;
2026 if (index > 0x20000000)
2027 panic("swp_pager_meta_free_all: failed to locate all swap meta blocks");
2032 * SWP_PAGER_METACTL() - misc control of swap and vm_page_t meta data.
2034 * This routine is capable of looking up, popping, or freeing
2035 * swapblk assignments in the swap meta data or in the vm_page_t.
2036 * The routine typically returns the swapblk being looked-up, or popped,
2037 * or SWAPBLK_NONE if the block was freed, or SWAPBLK_NONE if the block
2038 * was invalid. This routine will automatically free any invalid
2039 * meta-data swapblks.
2041 * It is not possible to store invalid swapblks in the swap meta data
2042 * (other then a literal 'SWAPBLK_NONE'), so we don't bother checking.
2044 * When acting on a busy resident page and paging is in progress, we
2045 * have to wait until paging is complete but otherwise can act on the
2048 * This routine must be called at splvm().
2050 * SWM_FREE remove and free swap block from metadata
2051 * SWM_POP remove from meta data but do not free.. pop it out
2060 struct swblock **pswap;
2061 struct swblock *swap;
2065 * The meta data only exists of the object is OBJT_SWAP
2066 * and even then might not be allocated yet.
2069 if (object->type != OBJT_SWAP)
2070 return(SWAPBLK_NONE);
2073 pswap = swp_pager_hash(object, index);
2075 if ((swap = *pswap) != NULL) {
2076 index &= SWAP_META_MASK;
2077 r1 = swap->swb_pages[index];
2079 if (r1 != SWAPBLK_NONE) {
2080 if (flags & SWM_FREE) {
2081 swp_pager_freeswapspace(r1, 1);
2084 if (flags & (SWM_FREE|SWM_POP)) {
2085 swap->swb_pages[index] = SWAPBLK_NONE;
2086 if (--swap->swb_count == 0) {
2087 *pswap = swap->swb_hnext;
2088 zfree(swap_zone, swap);
2089 --object->un_pager.swp.swp_bcount;