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 1.23 2006/04/30 18:25:37 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 */
138 #define AUTOCHAINDONE ((struct buf *)(intptr_t)-1)
141 * vm_swap_size is in page-sized chunks now. It was DEV_BSIZE'd chunks
145 extern int vm_swap_size; /* number of free swap blocks, in pages */
147 int swap_pager_full; /* swap space exhaustion (task killing) */
148 static int swap_pager_almost_full; /* swap space exhaustion (w/ hysteresis)*/
149 static int nsw_rcount; /* free read buffers */
150 static int nsw_wcount_sync; /* limit write buffers / synchronous */
151 static int nsw_wcount_async; /* limit write buffers / asynchronous */
152 static int nsw_wcount_async_max;/* assigned maximum */
153 static int nsw_cluster_max; /* maximum VOP I/O allowed */
154 static int sw_alloc_interlock; /* swap pager allocation interlock */
156 struct blist *swapblist;
157 static struct swblock **swhash;
158 static int swhash_mask;
159 static int swap_async_max = 4; /* maximum in-progress async I/O's */
161 extern struct vnode *swapdev_vp; /* from vm_swap.c */
163 SYSCTL_INT(_vm, OID_AUTO, swap_async_max,
164 CTLFLAG_RW, &swap_async_max, 0, "Maximum running async swap ops");
167 * "named" and "unnamed" anon region objects. Try to reduce the overhead
168 * of searching a named list by hashing it just a little.
173 #define NOBJLIST(handle) \
174 (&swap_pager_object_list[((int)(intptr_t)handle >> 4) & (NOBJLISTS-1)])
176 static struct pagerlst swap_pager_object_list[NOBJLISTS];
177 struct pagerlst swap_pager_un_object_list;
181 * pagerops for OBJT_SWAP - "swap pager". Some ops are also global procedure
182 * calls hooked from other parts of the VM system and do not appear here.
183 * (see vm/swap_pager.h).
187 swap_pager_alloc (void *handle, off_t size,
188 vm_prot_t prot, off_t offset);
189 static void swap_pager_dealloc (vm_object_t object);
190 static int swap_pager_getpages (vm_object_t, vm_page_t *, int, int);
191 static void swap_pager_init (void);
192 static void swap_pager_unswapped (vm_page_t);
193 static void swap_pager_strategy (vm_object_t, struct bio *);
194 static void swap_chain_iodone(struct bio *biox);
196 struct pagerops swappagerops = {
197 swap_pager_init, /* early system initialization of pager */
198 swap_pager_alloc, /* allocate an OBJT_SWAP object */
199 swap_pager_dealloc, /* deallocate an OBJT_SWAP object */
200 swap_pager_getpages, /* pagein */
201 swap_pager_putpages, /* pageout */
202 swap_pager_haspage, /* get backing store status for page */
203 swap_pager_unswapped, /* remove swap related to page */
204 swap_pager_strategy /* pager strategy call */
208 * dmmax is in page-sized chunks with the new swap system. It was
209 * dev-bsized chunks in the old. dmmax is always a power of 2.
211 * swap_*() routines are externally accessible. swp_*() routines are
216 static int dmmax_mask;
217 int nswap_lowat = 128; /* in pages, swap_pager_almost_full warn */
218 int nswap_hiwat = 512; /* in pages, swap_pager_almost_full warn */
220 static __inline void swp_sizecheck (void);
221 static void swp_pager_sync_iodone (struct bio *bio);
222 static void swp_pager_async_iodone (struct bio *bio);
225 * Swap bitmap functions
228 static __inline void swp_pager_freeswapspace (daddr_t blk, int npages);
229 static __inline daddr_t swp_pager_getswapspace (int npages);
235 static void swp_pager_meta_build (vm_object_t, vm_pindex_t, daddr_t);
236 static void swp_pager_meta_free (vm_object_t, vm_pindex_t, daddr_t);
237 static void swp_pager_meta_free_all (vm_object_t);
238 static daddr_t swp_pager_meta_ctl (vm_object_t, vm_pindex_t, int);
241 * SWP_SIZECHECK() - update swap_pager_full indication
243 * update the swap_pager_almost_full indication and warn when we are
244 * about to run out of swap space, using lowat/hiwat hysteresis.
246 * Clear swap_pager_full ( task killing ) indication when lowat is met.
248 * No restrictions on call
249 * This routine may not block.
250 * This routine must be called at splvm()
256 if (vm_swap_size < nswap_lowat) {
257 if (swap_pager_almost_full == 0) {
258 printf("swap_pager: out of swap space\n");
259 swap_pager_almost_full = 1;
263 if (vm_swap_size > nswap_hiwat)
264 swap_pager_almost_full = 0;
269 * SWAP_PAGER_INIT() - initialize the swap pager!
271 * Expected to be started from system init. NOTE: This code is run
272 * before much else so be careful what you depend on. Most of the VM
273 * system has yet to be initialized at this point.
277 swap_pager_init(void)
280 * Initialize object lists
284 for (i = 0; i < NOBJLISTS; ++i)
285 TAILQ_INIT(&swap_pager_object_list[i]);
286 TAILQ_INIT(&swap_pager_un_object_list);
289 * Device Stripe, in PAGE_SIZE'd blocks
292 dmmax = SWB_NPAGES * 2;
293 dmmax_mask = ~(dmmax - 1);
297 * SWAP_PAGER_SWAP_INIT() - swap pager initialization from pageout process
299 * Expected to be started from pageout process once, prior to entering
304 swap_pager_swap_init(void)
309 * Number of in-transit swap bp operations. Don't
310 * exhaust the pbufs completely. Make sure we
311 * initialize workable values (0 will work for hysteresis
312 * but it isn't very efficient).
314 * The nsw_cluster_max is constrained by the number of pages an XIO
315 * holds, i.e., (MAXPHYS/PAGE_SIZE) and our locally defined
316 * MAX_PAGEOUT_CLUSTER. Also be aware that swap ops are
317 * constrained by the swap device interleave stripe size.
319 * Currently we hardwire nsw_wcount_async to 4. This limit is
320 * designed to prevent other I/O from having high latencies due to
321 * our pageout I/O. The value 4 works well for one or two active swap
322 * devices but is probably a little low if you have more. Even so,
323 * a higher value would probably generate only a limited improvement
324 * with three or four active swap devices since the system does not
325 * typically have to pageout at extreme bandwidths. We will want
326 * at least 2 per swap devices, and 4 is a pretty good value if you
327 * have one NFS swap device due to the command/ack latency over NFS.
328 * So it all works out pretty well.
331 nsw_cluster_max = min((MAXPHYS/PAGE_SIZE), MAX_PAGEOUT_CLUSTER);
333 nsw_rcount = (nswbuf + 1) / 2;
334 nsw_wcount_sync = (nswbuf + 3) / 4;
335 nsw_wcount_async = 4;
336 nsw_wcount_async_max = nsw_wcount_async;
339 * Initialize our zone. Right now I'm just guessing on the number
340 * we need based on the number of pages in the system. Each swblock
341 * can hold 16 pages, so this is probably overkill. This reservation
342 * is typically limited to around 32MB by default.
344 n = vmstats.v_page_count / 2;
345 if (maxswzone && n > maxswzone / sizeof(struct swblock))
346 n = maxswzone / sizeof(struct swblock);
352 sizeof(struct swblock),
356 if (swap_zone != NULL)
359 * if the allocation failed, try a zone two thirds the
360 * size of the previous attempt.
365 if (swap_zone == NULL)
366 panic("swap_pager_swap_init: swap_zone == NULL");
368 printf("Swap zone entries reduced from %d to %d.\n", n2, n);
372 * Initialize our meta-data hash table. The swapper does not need to
373 * be quite as efficient as the VM system, so we do not use an
374 * oversized hash table.
376 * n: size of hash table, must be power of 2
377 * swhash_mask: hash table index mask
380 for (n = 1; n < n2 / 8; n *= 2)
383 swhash = malloc(sizeof(struct swblock *) * n, M_VMPGDATA, M_WAITOK);
384 bzero(swhash, sizeof(struct swblock *) * n);
390 * SWAP_PAGER_ALLOC() - allocate a new OBJT_SWAP VM object and instantiate
391 * its metadata structures.
393 * This routine is called from the mmap and fork code to create a new
394 * OBJT_SWAP object. We do this by creating an OBJT_DEFAULT object
395 * and then converting it with swp_pager_meta_build().
397 * This routine may block in vm_object_allocate() and create a named
398 * object lookup race, so we must interlock. We must also run at
399 * splvm() for the object lookup to handle races with interrupts, but
400 * we do not have to maintain splvm() in between the lookup and the
401 * add because (I believe) it is not possible to attempt to create
402 * a new swap object w/handle when a default object with that handle
407 swap_pager_alloc(void *handle, off_t size, vm_prot_t prot, off_t offset)
413 * Reference existing named region or allocate new one. There
414 * should not be a race here against swp_pager_meta_build()
415 * as called from vm_page_remove() in regards to the lookup
419 while (sw_alloc_interlock) {
420 sw_alloc_interlock = -1;
421 tsleep(&sw_alloc_interlock, 0, "swpalc", 0);
423 sw_alloc_interlock = 1;
425 object = vm_pager_object_lookup(NOBJLIST(handle), handle);
427 if (object != NULL) {
428 vm_object_reference(object);
430 object = vm_object_allocate(OBJT_DEFAULT,
431 OFF_TO_IDX(offset + PAGE_MASK + size));
432 object->handle = handle;
434 swp_pager_meta_build(object, 0, SWAPBLK_NONE);
437 if (sw_alloc_interlock < 0)
438 wakeup(&sw_alloc_interlock);
440 sw_alloc_interlock = 0;
442 object = vm_object_allocate(OBJT_DEFAULT,
443 OFF_TO_IDX(offset + PAGE_MASK + size));
445 swp_pager_meta_build(object, 0, SWAPBLK_NONE);
452 * SWAP_PAGER_DEALLOC() - remove swap metadata from object
454 * The swap backing for the object is destroyed. The code is
455 * designed such that we can reinstantiate it later, but this
456 * routine is typically called only when the entire object is
457 * about to be destroyed.
459 * This routine may block, but no longer does.
461 * The object must be locked or unreferenceable.
465 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)
590 daddr_t blk = SWAPBLK_NONE;
591 vm_pindex_t beg = start; /* save start index */
597 while ((blk = swp_pager_getswapspace(n)) == SWAPBLK_NONE) {
600 swp_pager_meta_free(object, beg, start - beg);
606 swp_pager_meta_build(object, start, blk);
612 swp_pager_meta_free(object, start, n);
618 * SWAP_PAGER_COPY() - copy blocks from source pager to destination pager
619 * and destroy the source.
621 * Copy any valid swapblks from the source to the destination. In
622 * cases where both the source and destination have a valid swapblk,
623 * we keep the destination's.
625 * This routine is allowed to block. It may block allocating metadata
626 * indirectly through swp_pager_meta_build() or if paging is still in
627 * progress on the source.
629 * This routine can be called at any spl
631 * XXX vm_page_collapse() kinda expects us not to block because we
632 * supposedly do not need to allocate memory, but for the moment we
633 * *may* have to get a little memory from the zone allocator, but
634 * it is taken from the interrupt memory. We should be ok.
636 * The source object contains no vm_page_t's (which is just as well)
638 * The source object is of type OBJT_SWAP.
640 * The source and destination objects must be locked or
641 * inaccessible (XXX are they ?)
645 swap_pager_copy(vm_object_t srcobject, vm_object_t dstobject,
646 vm_pindex_t offset, int destroysource)
653 * If destroysource is set, we remove the source object from the
654 * swap_pager internal queue now.
658 if (srcobject->handle == NULL) {
660 &swap_pager_un_object_list,
666 NOBJLIST(srcobject->handle),
674 * transfer source to destination.
677 for (i = 0; i < dstobject->size; ++i) {
681 * Locate (without changing) the swapblk on the destination,
682 * unless it is invalid in which case free it silently, or
683 * if the destination is a resident page, in which case the
684 * source is thrown away.
687 dstaddr = swp_pager_meta_ctl(dstobject, i, 0);
689 if (dstaddr == SWAPBLK_NONE) {
691 * Destination has no swapblk and is not resident,
696 srcaddr = swp_pager_meta_ctl(
702 if (srcaddr != SWAPBLK_NONE)
703 swp_pager_meta_build(dstobject, i, srcaddr);
706 * Destination has valid swapblk or it is represented
707 * by a resident page. We destroy the sourceblock.
710 swp_pager_meta_ctl(srcobject, i + offset, SWM_FREE);
715 * Free left over swap blocks in source.
717 * We have to revert the type to OBJT_DEFAULT so we do not accidently
718 * double-remove the object from the swap queues.
722 swp_pager_meta_free_all(srcobject);
724 * Reverting the type is not necessary, the caller is going
725 * to destroy srcobject directly, but I'm doing it here
726 * for consistency since we've removed the object from its
729 srcobject->type = OBJT_DEFAULT;
735 * SWAP_PAGER_HASPAGE() - determine if we have good backing store for
736 * the requested page.
738 * We determine whether good backing store exists for the requested
739 * page and return TRUE if it does, FALSE if it doesn't.
741 * If TRUE, we also try to determine how much valid, contiguous backing
742 * store exists before and after the requested page within a reasonable
743 * distance. We do not try to restrict it to the swap device stripe
744 * (that is handled in getpages/putpages). It probably isn't worth
749 swap_pager_haspage(vm_object_t object, vm_pindex_t pindex, int *before,
755 * do we have good backing store at the requested index ?
759 blk0 = swp_pager_meta_ctl(object, pindex, 0);
761 if (blk0 == SWAPBLK_NONE) {
771 * find backwards-looking contiguous good backing store
774 if (before != NULL) {
777 for (i = 1; i < (SWB_NPAGES/2); ++i) {
782 blk = swp_pager_meta_ctl(object, pindex - i, 0);
790 * find forward-looking contiguous good backing store
796 for (i = 1; i < (SWB_NPAGES/2); ++i) {
799 blk = swp_pager_meta_ctl(object, pindex + i, 0);
810 * SWAP_PAGER_PAGE_UNSWAPPED() - remove swap backing store related to page
812 * This removes any associated swap backing store, whether valid or
813 * not, from the page.
815 * This routine is typically called when a page is made dirty, at
816 * which point any associated swap can be freed. MADV_FREE also
817 * calls us in a special-case situation
819 * NOTE!!! If the page is clean and the swap was valid, the caller
820 * should make the page dirty before calling this routine. This routine
821 * does NOT change the m->dirty status of the page. Also: MADV_FREE
824 * This routine may not block
825 * This routine must be called at splvm()
829 swap_pager_unswapped(vm_page_t m)
831 swp_pager_meta_ctl(m->object, m->pindex, SWM_FREE);
835 * SWAP_PAGER_STRATEGY() - read, write, free blocks
837 * This implements the vm_pager_strategy() interface to swap and allows
838 * other parts of the system to directly access swap as backing store
839 * through vm_objects of type OBJT_SWAP. This is intended to be a
840 * cacheless interface ( i.e. caching occurs at higher levels ).
841 * Therefore we do not maintain any resident pages. All I/O goes
842 * directly to and from the swap device.
844 * We currently attempt to run I/O synchronously or asynchronously as
845 * the caller requests. This isn't perfect because we loose error
846 * sequencing when we run multiple ops in parallel to satisfy a request.
847 * But this is swap, so we let it all hang out.
851 swap_pager_strategy(vm_object_t object, struct bio *bio)
853 struct buf *bp = bio->bio_buf;
856 vm_pindex_t biox_blkno = 0;
859 struct bio *biox = NULL;
860 struct buf *bufx = NULL;
861 struct bio_track *track;
864 * tracking for swapdev vnode I/Os
866 if (bp->b_cmd == BUF_CMD_READ)
867 track = &swapdev_vp->v_track_read;
869 track = &swapdev_vp->v_track_write;
871 if (bp->b_bcount & PAGE_MASK) {
872 bp->b_error = EINVAL;
873 bp->b_flags |= B_ERROR | B_INVAL;
875 printf("swap_pager_strategy: bp %p offset %lld size %d, not page bounded\n", bp, bio->bio_offset, (int)bp->b_bcount);
880 * Clear error indication, initialize page index, count, data pointer.
883 bp->b_flags &= ~B_ERROR;
884 bp->b_resid = bp->b_bcount;
886 start = (vm_pindex_t)(bio->bio_offset >> PAGE_SHIFT);
887 count = howmany(bp->b_bcount, PAGE_SIZE);
893 * Deal with BUF_CMD_FREEBLKS
895 if (bp->b_cmd == BUF_CMD_FREEBLKS) {
897 * FREE PAGE(s) - destroy underlying swap that is no longer
900 swp_pager_meta_free(object, start, count);
908 * We need to be able to create a new cluster of I/O's. We cannot
909 * use the caller fields of the passed bio so push a new one.
911 * Because nbio is just a placeholder for the cluster links,
912 * we can biodone() the original bio instead of nbio to make
913 * things a bit more efficient.
915 nbio = push_bio(bio);
916 nbio->bio_offset = bio->bio_offset;
917 nbio->bio_caller_info1.cluster_head = NULL;
918 nbio->bio_caller_info2.cluster_tail = NULL;
921 * Execute read or write
928 * Obtain block. If block not found and writing, allocate a
929 * new block and build it into the object.
932 blk = swp_pager_meta_ctl(object, start, 0);
933 if ((blk == SWAPBLK_NONE) && bp->b_cmd != BUF_CMD_READ) {
934 blk = swp_pager_getswapspace(1);
935 if (blk == SWAPBLK_NONE) {
936 bp->b_error = ENOMEM;
937 bp->b_flags |= B_ERROR;
940 swp_pager_meta_build(object, start, blk);
944 * Do we have to flush our current collection? Yes if:
946 * - no swap block at this index
947 * - swap block is not contiguous
948 * - we cross a physical disk boundry in the
953 biox && (biox_blkno + btoc(bufx->b_bcount) != blk ||
954 ((biox_blkno ^ blk) & dmmax_mask)
958 if (bp->b_cmd == BUF_CMD_READ) {
959 ++mycpu->gd_cnt.v_swapin;
960 mycpu->gd_cnt.v_swappgsin += btoc(bufx->b_bcount);
962 ++mycpu->gd_cnt.v_swapout;
963 mycpu->gd_cnt.v_swappgsout += btoc(bufx->b_bcount);
964 bufx->b_dirtyend = bufx->b_bcount;
968 * Flush the biox to the swap device.
970 if (bufx->b_bcount) {
971 bufx->b_bufsize = bufx->b_bcount;
972 if (bufx->b_cmd != BUF_CMD_READ)
973 bufx->b_dirtyend = bufx->b_bcount;
975 vn_strategy(swapdev_vp, biox);
985 * Add new swapblk to biox, instantiating biox if necessary.
986 * Zero-fill reads are able to take a shortcut.
988 if (blk == SWAPBLK_NONE) {
990 * We can only get here if we are reading. Since
991 * we are at splvm() we can safely modify b_resid,
992 * even if chain ops are in progress.
994 bzero(data, PAGE_SIZE);
995 bp->b_resid -= PAGE_SIZE;
998 /* XXX chain count > 4, wait to <= 4 */
1000 bufx = getpbuf(NULL);
1001 biox = &bufx->b_bio1;
1002 cluster_append(nbio, bufx);
1003 bufx->b_flags |= (bufx->b_flags & B_ORDERED) |
1005 bufx->b_cmd = bp->b_cmd;
1006 biox->bio_done = swap_chain_iodone;
1007 biox->bio_offset = (off_t)blk << PAGE_SHIFT;
1008 biox->bio_caller_info1.cluster_parent = nbio;
1011 bufx->b_data = data;
1013 bufx->b_bcount += PAGE_SIZE;
1021 * Flush out last buffer
1026 if ((bp->b_flags & B_ASYNC) == 0)
1027 bufx->b_flags &= ~B_ASYNC;
1028 if (bufx->b_cmd == BUF_CMD_READ) {
1029 ++mycpu->gd_cnt.v_swapin;
1030 mycpu->gd_cnt.v_swappgsin += btoc(bufx->b_bcount);
1032 ++mycpu->gd_cnt.v_swapout;
1033 mycpu->gd_cnt.v_swappgsout += btoc(bufx->b_bcount);
1034 bufx->b_dirtyend = bufx->b_bcount;
1036 if (bufx->b_bcount) {
1037 bufx->b_bufsize = bufx->b_bcount;
1038 if (bufx->b_cmd != BUF_CMD_READ)
1039 bufx->b_dirtyend = bufx->b_bcount;
1041 vn_strategy(swapdev_vp, biox);
1045 /* biox, bufx = NULL */
1049 * Wait for completion. Now that we are no longer using
1050 * cluster_append, use the cluster_tail field to indicate
1051 * auto-completion if there are still I/O's in progress.
1053 if (bp->b_flags & B_ASYNC) {
1055 if (nbio->bio_caller_info1.cluster_head == NULL) {
1058 nbio->bio_caller_info2.cluster_tail = AUTOCHAINDONE;
1063 while (nbio->bio_caller_info1.cluster_head != NULL) {
1064 bp->b_flags |= B_WANT;
1065 tsleep(bp, 0, "bpchain", 0);
1067 if (bp->b_resid != 0 && !(bp->b_flags & B_ERROR)) {
1068 bp->b_flags |= B_ERROR;
1069 bp->b_error = EINVAL;
1077 swap_chain_iodone(struct bio *biox)
1080 struct buf *bufx; /* chained sub-buffer */
1081 struct bio *nbio; /* parent nbio with chain glue */
1082 struct buf *bp; /* original bp associated with nbio */
1084 bufx = biox->bio_buf;
1085 nbio = biox->bio_caller_info1.cluster_parent;
1089 * Update the original buffer
1091 KKASSERT(bp != NULL);
1092 if (bufx->b_flags & B_ERROR) {
1093 bp->b_flags |= B_ERROR;
1094 bp->b_error = bufx->b_error;
1095 } else if (bufx->b_resid != 0) {
1096 bp->b_flags |= B_ERROR;
1097 bp->b_error = EINVAL;
1099 bp->b_resid -= bufx->b_bcount;
1103 * Remove us from the chain. It is sufficient to clean up
1104 * cluster_head. Once the chain is operational cluster_tail
1105 * may be used to indicate AUTOCHAINDONE. Note that I/O's
1106 * can complete while the swap system is still appending new
1107 * BIOs to the chain.
1109 nextp = &nbio->bio_caller_info1.cluster_head;
1110 while (*nextp != bufx) {
1111 KKASSERT(*nextp != NULL);
1112 nextp = &(*nextp)->b_cluster_next;
1114 *nextp = bufx->b_cluster_next;
1115 if (bp->b_flags & B_WANT) {
1116 bp->b_flags &= ~B_WANT;
1121 * Clean up bufx. If this was the last buffer in the chain
1122 * and AUTOCHAINDONE was set, finish off the original I/O
1125 * nbio was just a fake BIO layer to hold the cluster links,
1126 * we can issue the biodone() on the layer above it.
1128 if (nbio->bio_caller_info1.cluster_head == NULL &&
1129 nbio->bio_caller_info2.cluster_tail == AUTOCHAINDONE
1131 nbio->bio_caller_info2.cluster_tail = NULL;
1132 if (bp->b_resid != 0 && !(bp->b_flags & B_ERROR)) {
1133 bp->b_flags |= B_ERROR;
1134 bp->b_error = EINVAL;
1136 biodone(nbio->bio_prev);
1138 bufx->b_flags &= ~B_ASYNC;
1139 relpbuf(bufx, NULL);
1143 * SWAP_PAGER_GETPAGES() - bring pages in from swap
1145 * Attempt to retrieve (m, count) pages from backing store, but make
1146 * sure we retrieve at least m[reqpage]. We try to load in as large
1147 * a chunk surrounding m[reqpage] as is contiguous in swap and which
1148 * belongs to the same object.
1150 * The code is designed for asynchronous operation and
1151 * immediate-notification of 'reqpage' but tends not to be
1152 * used that way. Please do not optimize-out this algorithmic
1153 * feature, I intend to improve on it in the future.
1155 * The parent has a single vm_object_pip_add() reference prior to
1156 * calling us and we should return with the same.
1158 * The parent has BUSY'd the pages. We should return with 'm'
1159 * left busy, but the others adjusted.
1163 swap_pager_getpages(vm_object_t object, vm_page_t *m, int count, int reqpage)
1172 vm_pindex_t lastpindex;
1176 if (mreq->object != object) {
1177 panic("swap_pager_getpages: object mismatch %p/%p",
1183 * Calculate range to retrieve. The pages have already been assigned
1184 * their swapblks. We require a *contiguous* range that falls entirely
1185 * within a single device stripe. If we do not supply it, bad things
1186 * happen. Note that blk, iblk & jblk can be SWAPBLK_NONE, but the
1187 * loops are set up such that the case(s) are handled implicitly.
1189 * The swp_*() calls must be made at splvm(). vm_page_free() does
1190 * not need to be, but it will go a little faster if it is.
1194 blk = swp_pager_meta_ctl(mreq->object, mreq->pindex, 0);
1196 for (i = reqpage - 1; i >= 0; --i) {
1199 iblk = swp_pager_meta_ctl(m[i]->object, m[i]->pindex, 0);
1200 if (blk != iblk + (reqpage - i))
1202 if ((blk ^ iblk) & dmmax_mask)
1207 for (j = reqpage + 1; j < count; ++j) {
1210 jblk = swp_pager_meta_ctl(m[j]->object, m[j]->pindex, 0);
1211 if (blk != jblk - (j - reqpage))
1213 if ((blk ^ jblk) & dmmax_mask)
1218 * free pages outside our collection range. Note: we never free
1219 * mreq, it must remain busy throughout.
1225 for (k = 0; k < i; ++k)
1227 for (k = j; k < count; ++k)
1234 * Return VM_PAGER_FAIL if we have nothing to do. Return mreq
1235 * still busy, but the others unbusied.
1238 if (blk == SWAPBLK_NONE)
1239 return(VM_PAGER_FAIL);
1242 * Get a swap buffer header to perform the IO
1245 bp = getpbuf(&nsw_rcount);
1247 kva = (vm_offset_t) bp->b_data;
1250 * map our page(s) into kva for input
1253 pmap_qenter(kva, m + i, j - i);
1255 bp->b_data = (caddr_t) kva;
1256 bp->b_bcount = PAGE_SIZE * (j - i);
1257 bp->b_bufsize = PAGE_SIZE * (j - i);
1258 bio->bio_done = swp_pager_async_iodone;
1259 bio->bio_offset = (off_t)(blk - (reqpage - i)) << PAGE_SHIFT;
1260 bio->bio_driver_info = (void *)(reqpage - i);
1265 for (k = i; k < j; ++k) {
1266 bp->b_xio.xio_pages[k - i] = m[k];
1267 vm_page_flag_set(m[k], PG_SWAPINPROG);
1270 bp->b_xio.xio_npages = j - i;
1272 mycpu->gd_cnt.v_swapin++;
1273 mycpu->gd_cnt.v_swappgsin += bp->b_xio.xio_npages;
1276 * We still hold the lock on mreq, and our automatic completion routine
1277 * does not remove it.
1280 vm_object_pip_add(mreq->object, bp->b_xio.xio_npages);
1281 lastpindex = m[j-1]->pindex;
1284 * perform the I/O. NOTE!!! bp cannot be considered valid after
1285 * this point because we automatically release it on completion.
1286 * Instead, we look at the one page we are interested in which we
1287 * still hold a lock on even through the I/O completion.
1289 * The other pages in our m[] array are also released on completion,
1290 * so we cannot assume they are valid anymore either.
1293 bp->b_cmd = BUF_CMD_READ;
1295 vn_strategy(swapdev_vp, bio);
1298 * wait for the page we want to complete. PG_SWAPINPROG is always
1299 * cleared on completion. If an I/O error occurs, SWAPBLK_NONE
1300 * is set in the meta-data.
1305 while ((mreq->flags & PG_SWAPINPROG) != 0) {
1306 vm_page_flag_set(mreq, PG_WANTED | PG_REFERENCED);
1307 mycpu->gd_cnt.v_intrans++;
1308 if (tsleep(mreq, 0, "swread", hz*20)) {
1310 "swap_pager: indefinite wait buffer: "
1311 " offset: %lld, size: %d\n",
1312 bio->bio_offset, bp->b_bcount
1320 * mreq is left bussied after completion, but all the other pages
1321 * are freed. If we had an unrecoverable read error the page will
1325 if (mreq->valid != VM_PAGE_BITS_ALL) {
1326 return(VM_PAGER_ERROR);
1328 return(VM_PAGER_OK);
1332 * A final note: in a low swap situation, we cannot deallocate swap
1333 * and mark a page dirty here because the caller is likely to mark
1334 * the page clean when we return, causing the page to possibly revert
1335 * to all-zero's later.
1340 * swap_pager_putpages:
1342 * Assign swap (if necessary) and initiate I/O on the specified pages.
1344 * We support both OBJT_DEFAULT and OBJT_SWAP objects. DEFAULT objects
1345 * are automatically converted to SWAP objects.
1347 * In a low memory situation we may block in vn_strategy(), but the new
1348 * vm_page reservation system coupled with properly written VFS devices
1349 * should ensure that no low-memory deadlock occurs. This is an area
1352 * The parent has N vm_object_pip_add() references prior to
1353 * calling us and will remove references for rtvals[] that are
1354 * not set to VM_PAGER_PEND. We need to remove the rest on I/O
1357 * The parent has soft-busy'd the pages it passes us and will unbusy
1358 * those whos rtvals[] entry is not set to VM_PAGER_PEND on return.
1359 * We need to unbusy the rest on I/O completion.
1363 swap_pager_putpages(vm_object_t object, vm_page_t *m, int count, boolean_t sync,
1369 if (count && m[0]->object != object) {
1370 panic("swap_pager_getpages: object mismatch %p/%p",
1378 * Turn object into OBJT_SWAP
1379 * check for bogus sysops
1380 * force sync if not pageout process
1383 if (object->type != OBJT_SWAP)
1384 swp_pager_meta_build(object, 0, SWAPBLK_NONE);
1386 if (curthread != pagethread)
1392 * Update nsw parameters from swap_async_max sysctl values.
1393 * Do not let the sysop crash the machine with bogus numbers.
1396 if (swap_async_max != nsw_wcount_async_max) {
1402 if ((n = swap_async_max) > nswbuf / 2)
1409 * Adjust difference ( if possible ). If the current async
1410 * count is too low, we may not be able to make the adjustment
1414 n -= nsw_wcount_async_max;
1415 if (nsw_wcount_async + n >= 0) {
1416 nsw_wcount_async += n;
1417 nsw_wcount_async_max += n;
1418 wakeup(&nsw_wcount_async);
1426 * Assign swap blocks and issue I/O. We reallocate swap on the fly.
1427 * The page is left dirty until the pageout operation completes
1431 for (i = 0; i < count; i += n) {
1438 * Maximum I/O size is limited by a number of factors.
1441 n = min(BLIST_MAX_ALLOC, count - i);
1442 n = min(n, nsw_cluster_max);
1447 * Get biggest block of swap we can. If we fail, fall
1448 * back and try to allocate a smaller block. Don't go
1449 * overboard trying to allocate space if it would overly
1453 (blk = swp_pager_getswapspace(n)) == SWAPBLK_NONE &&
1458 if (blk == SWAPBLK_NONE) {
1459 for (j = 0; j < n; ++j)
1460 rtvals[i+j] = VM_PAGER_FAIL;
1466 * The I/O we are constructing cannot cross a physical
1467 * disk boundry in the swap stripe. Note: we are still
1470 if ((blk ^ (blk + n)) & dmmax_mask) {
1471 j = ((blk + dmmax) & dmmax_mask) - blk;
1472 swp_pager_freeswapspace(blk + j, n - j);
1477 * All I/O parameters have been satisfied, build the I/O
1478 * request and assign the swap space.
1482 bp = getpbuf(&nsw_wcount_sync);
1484 bp = getpbuf(&nsw_wcount_async);
1487 pmap_qenter((vm_offset_t)bp->b_data, &m[i], n);
1489 bp->b_bcount = PAGE_SIZE * n;
1490 bp->b_bufsize = PAGE_SIZE * n;
1491 bio->bio_offset = (off_t)blk << PAGE_SHIFT;
1493 for (j = 0; j < n; ++j) {
1494 vm_page_t mreq = m[i+j];
1496 swp_pager_meta_build(
1501 vm_page_dirty(mreq);
1502 rtvals[i+j] = VM_PAGER_OK;
1504 vm_page_flag_set(mreq, PG_SWAPINPROG);
1505 bp->b_xio.xio_pages[j] = mreq;
1507 bp->b_xio.xio_npages = n;
1509 mycpu->gd_cnt.v_swapout++;
1510 mycpu->gd_cnt.v_swappgsout += bp->b_xio.xio_npages;
1514 bp->b_dirtyoff = 0; /* req'd for NFS */
1515 bp->b_dirtyend = bp->b_bcount; /* req'd for NFS */
1516 bp->b_cmd = BUF_CMD_WRITE;
1521 if (sync == FALSE) {
1522 bp->b_flags |= B_ASYNC;
1523 bio->bio_done = swp_pager_async_iodone;
1525 vn_strategy(swapdev_vp, bio);
1527 for (j = 0; j < n; ++j)
1528 rtvals[i+j] = VM_PAGER_PEND;
1536 bio->bio_done = swp_pager_sync_iodone;
1537 vn_strategy(swapdev_vp, bio);
1540 * Wait for the sync I/O to complete, then update rtvals.
1541 * We just set the rtvals[] to VM_PAGER_PEND so we can call
1542 * our async completion routine at the end, thus avoiding a
1547 while (bp->b_cmd != BUF_CMD_DONE)
1548 tsleep(bp, 0, "swwrt", 0);
1550 for (j = 0; j < n; ++j)
1551 rtvals[i+j] = VM_PAGER_PEND;
1554 * Now that we are through with the bp, we can call the
1555 * normal async completion, which frees everything up.
1558 swp_pager_async_iodone(bio);
1565 * swap_pager_sync_iodone:
1567 * Completion routine for synchronous reads and writes from/to swap.
1568 * We just mark the bp is complete and wake up anyone waiting on it.
1570 * This routine may not block. This routine is called at splbio() or better.
1574 swp_pager_sync_iodone(struct bio *bio)
1576 struct buf *bp = bio->bio_buf;
1578 bp->b_flags &= ~B_ASYNC;
1579 bp->b_cmd = BUF_CMD_DONE;
1584 * swp_pager_async_iodone:
1586 * Completion routine for asynchronous reads and writes from/to swap.
1587 * Also called manually by synchronous code to finish up a bp.
1589 * For READ operations, the pages are PG_BUSY'd. For WRITE operations,
1590 * the pages are vm_page_t->busy'd. For READ operations, we PG_BUSY
1591 * unbusy all pages except the 'main' request page. For WRITE
1592 * operations, we vm_page_t->busy'd unbusy all pages ( we can do this
1593 * because we marked them all VM_PAGER_PEND on return from putpages ).
1595 * This routine may not block.
1599 swp_pager_async_iodone(struct bio *bio)
1601 struct buf *bp = bio->bio_buf;
1602 vm_object_t object = NULL;
1609 if (bp->b_flags & B_ERROR) {
1611 "swap_pager: I/O error - %s failed; offset %lld,"
1612 "size %ld, error %d\n",
1613 ((bp->b_cmd == BUF_CMD_READ) ? "pagein" : "pageout"),
1621 * set object, raise to splvm().
1624 if (bp->b_xio.xio_npages)
1625 object = bp->b_xio.xio_pages[0]->object;
1629 * remove the mapping for kernel virtual
1632 pmap_qremove((vm_offset_t)bp->b_data, bp->b_xio.xio_npages);
1635 * cleanup pages. If an error occurs writing to swap, we are in
1636 * very serious trouble. If it happens to be a disk error, though,
1637 * we may be able to recover by reassigning the swap later on. So
1638 * in this case we remove the m->swapblk assignment for the page
1639 * but do not free it in the rlist. The errornous block(s) are thus
1640 * never reallocated as swap. Redirty the page and continue.
1643 for (i = 0; i < bp->b_xio.xio_npages; ++i) {
1644 vm_page_t m = bp->b_xio.xio_pages[i];
1646 vm_page_flag_clear(m, PG_SWAPINPROG);
1648 if (bp->b_flags & B_ERROR) {
1650 * If an error occurs I'd love to throw the swapblk
1651 * away without freeing it back to swapspace, so it
1652 * can never be used again. But I can't from an
1656 if (bp->b_cmd == BUF_CMD_READ) {
1658 * When reading, reqpage needs to stay
1659 * locked for the parent, but all other
1660 * pages can be freed. We still want to
1661 * wakeup the parent waiting on the page,
1662 * though. ( also: pg_reqpage can be -1 and
1663 * not match anything ).
1665 * We have to wake specifically requested pages
1666 * up too because we cleared PG_SWAPINPROG and
1667 * someone may be waiting for that.
1669 * NOTE: for reads, m->dirty will probably
1670 * be overridden by the original caller of
1671 * getpages so don't play cute tricks here.
1673 * XXX IT IS NOT LEGAL TO FREE THE PAGE HERE
1674 * AS THIS MESSES WITH object->memq, and it is
1675 * not legal to mess with object->memq from an
1680 vm_page_flag_clear(m, PG_ZERO);
1683 * bio_driver_info holds the requested page
1686 if (i != (int)bio->bio_driver_info)
1691 * If i == bp->b_pager.pg_reqpage, do not wake
1692 * the page up. The caller needs to.
1696 * If a write error occurs, reactivate page
1697 * so it doesn't clog the inactive list,
1698 * then finish the I/O.
1701 vm_page_activate(m);
1702 vm_page_io_finish(m);
1704 } else if (bp->b_cmd == BUF_CMD_READ) {
1706 * For read success, clear dirty bits. Nobody should
1707 * have this page mapped but don't take any chances,
1708 * make sure the pmap modify bits are also cleared.
1710 * NOTE: for reads, m->dirty will probably be
1711 * overridden by the original caller of getpages so
1712 * we cannot set them in order to free the underlying
1713 * swap in a low-swap situation. I don't think we'd
1714 * want to do that anyway, but it was an optimization
1715 * that existed in the old swapper for a time before
1716 * it got ripped out due to precisely this problem.
1718 * clear PG_ZERO in page.
1720 * If not the requested page then deactivate it.
1722 * Note that the requested page, reqpage, is left
1723 * busied, but we still have to wake it up. The
1724 * other pages are released (unbusied) by
1725 * vm_page_wakeup(). We do not set reqpage's
1726 * valid bits here, it is up to the caller.
1729 pmap_clear_modify(m);
1730 m->valid = VM_PAGE_BITS_ALL;
1732 vm_page_flag_clear(m, PG_ZERO);
1735 * We have to wake specifically requested pages
1736 * up too because we cleared PG_SWAPINPROG and
1737 * could be waiting for it in getpages. However,
1738 * be sure to not unbusy getpages specifically
1739 * requested page - getpages expects it to be
1742 * bio_driver_info holds the requested page
1744 if (i != (int)bio->bio_driver_info) {
1745 vm_page_deactivate(m);
1752 * For write success, clear the modify and dirty
1753 * status, then finish the I/O ( which decrements the
1754 * busy count and possibly wakes waiter's up ).
1756 pmap_clear_modify(m);
1758 vm_page_io_finish(m);
1759 if (!vm_page_count_severe() || !vm_page_try_to_cache(m))
1760 vm_page_protect(m, VM_PROT_READ);
1765 * adjust pip. NOTE: the original parent may still have its own
1766 * pip refs on the object.
1770 vm_object_pip_wakeupn(object, bp->b_xio.xio_npages);
1773 * release the physical I/O buffer
1775 if (bp->b_cmd == BUF_CMD_READ)
1776 nswptr = &nsw_rcount;
1777 else if (bp->b_flags & B_ASYNC)
1778 nswptr = &nsw_wcount_async;
1780 nswptr = &nsw_wcount_sync;
1781 bp->b_cmd = BUF_CMD_DONE;
1782 relpbuf(bp, nswptr);
1786 /************************************************************************
1788 ************************************************************************
1790 * These routines manipulate the swap metadata stored in the
1791 * OBJT_SWAP object. All swp_*() routines must be called at
1792 * splvm() because swap can be freed up by the low level vm_page
1793 * code which might be called from interrupts beyond what splbio() covers.
1795 * Swap metadata is implemented with a global hash and not directly
1796 * linked into the object. Instead the object simply contains
1797 * appropriate tracking counters.
1801 * SWP_PAGER_HASH() - hash swap meta data
1803 * This is an inline helper function which hashes the swapblk given
1804 * the object and page index. It returns a pointer to a pointer
1805 * to the object, or a pointer to a NULL pointer if it could not
1808 * This routine must be called at splvm().
1811 static __inline struct swblock **
1812 swp_pager_hash(vm_object_t object, vm_pindex_t index)
1814 struct swblock **pswap;
1815 struct swblock *swap;
1817 index &= ~SWAP_META_MASK;
1818 pswap = &swhash[(index ^ (int)(intptr_t)object) & swhash_mask];
1820 while ((swap = *pswap) != NULL) {
1821 if (swap->swb_object == object &&
1822 swap->swb_index == index
1826 pswap = &swap->swb_hnext;
1832 * SWP_PAGER_META_BUILD() - add swap block to swap meta data for object
1834 * We first convert the object to a swap object if it is a default
1837 * The specified swapblk is added to the object's swap metadata. If
1838 * the swapblk is not valid, it is freed instead. Any previously
1839 * assigned swapblk is freed.
1841 * This routine must be called at splvm(), except when used to convert
1842 * an OBJT_DEFAULT object into an OBJT_SWAP object.
1847 swp_pager_meta_build(
1852 struct swblock *swap;
1853 struct swblock **pswap;
1856 * Convert default object to swap object if necessary
1859 if (object->type != OBJT_SWAP) {
1860 object->type = OBJT_SWAP;
1861 object->un_pager.swp.swp_bcount = 0;
1863 if (object->handle != NULL) {
1865 NOBJLIST(object->handle),
1871 &swap_pager_un_object_list,
1879 * Locate hash entry. If not found create, but if we aren't adding
1880 * anything just return. If we run out of space in the map we wait
1881 * and, since the hash table may have changed, retry.
1885 pswap = swp_pager_hash(object, index);
1887 if ((swap = *pswap) == NULL) {
1890 if (swapblk == SWAPBLK_NONE)
1893 swap = *pswap = zalloc(swap_zone);
1898 swap->swb_hnext = NULL;
1899 swap->swb_object = object;
1900 swap->swb_index = index & ~SWAP_META_MASK;
1901 swap->swb_count = 0;
1903 ++object->un_pager.swp.swp_bcount;
1905 for (i = 0; i < SWAP_META_PAGES; ++i)
1906 swap->swb_pages[i] = SWAPBLK_NONE;
1910 * Delete prior contents of metadata
1913 index &= SWAP_META_MASK;
1915 if (swap->swb_pages[index] != SWAPBLK_NONE) {
1916 swp_pager_freeswapspace(swap->swb_pages[index], 1);
1921 * Enter block into metadata
1924 swap->swb_pages[index] = swapblk;
1925 if (swapblk != SWAPBLK_NONE)
1930 * SWP_PAGER_META_FREE() - free a range of blocks in the object's swap metadata
1932 * The requested range of blocks is freed, with any associated swap
1933 * returned to the swap bitmap.
1935 * This routine will free swap metadata structures as they are cleaned
1936 * out. This routine does *NOT* operate on swap metadata associated
1937 * with resident pages.
1939 * This routine must be called at splvm()
1943 swp_pager_meta_free(vm_object_t object, vm_pindex_t index, daddr_t count)
1945 if (object->type != OBJT_SWAP)
1949 struct swblock **pswap;
1950 struct swblock *swap;
1952 pswap = swp_pager_hash(object, index);
1954 if ((swap = *pswap) != NULL) {
1955 daddr_t v = swap->swb_pages[index & SWAP_META_MASK];
1957 if (v != SWAPBLK_NONE) {
1958 swp_pager_freeswapspace(v, 1);
1959 swap->swb_pages[index & SWAP_META_MASK] =
1961 if (--swap->swb_count == 0) {
1962 *pswap = swap->swb_hnext;
1963 zfree(swap_zone, swap);
1964 --object->un_pager.swp.swp_bcount;
1970 int n = SWAP_META_PAGES - (index & SWAP_META_MASK);
1978 * SWP_PAGER_META_FREE_ALL() - destroy all swap metadata associated with object
1980 * This routine locates and destroys all swap metadata associated with
1983 * This routine must be called at splvm()
1987 swp_pager_meta_free_all(vm_object_t object)
1991 if (object->type != OBJT_SWAP)
1994 while (object->un_pager.swp.swp_bcount) {
1995 struct swblock **pswap;
1996 struct swblock *swap;
1998 pswap = swp_pager_hash(object, index);
1999 if ((swap = *pswap) != NULL) {
2002 for (i = 0; i < SWAP_META_PAGES; ++i) {
2003 daddr_t v = swap->swb_pages[i];
2004 if (v != SWAPBLK_NONE) {
2006 swp_pager_freeswapspace(v, 1);
2009 if (swap->swb_count != 0)
2010 panic("swap_pager_meta_free_all: swb_count != 0");
2011 *pswap = swap->swb_hnext;
2012 zfree(swap_zone, swap);
2013 --object->un_pager.swp.swp_bcount;
2015 index += SWAP_META_PAGES;
2016 if (index > 0x20000000)
2017 panic("swp_pager_meta_free_all: failed to locate all swap meta blocks");
2022 * SWP_PAGER_METACTL() - misc control of swap and vm_page_t meta data.
2024 * This routine is capable of looking up, popping, or freeing
2025 * swapblk assignments in the swap meta data or in the vm_page_t.
2026 * The routine typically returns the swapblk being looked-up, or popped,
2027 * or SWAPBLK_NONE if the block was freed, or SWAPBLK_NONE if the block
2028 * was invalid. This routine will automatically free any invalid
2029 * meta-data swapblks.
2031 * It is not possible to store invalid swapblks in the swap meta data
2032 * (other then a literal 'SWAPBLK_NONE'), so we don't bother checking.
2034 * When acting on a busy resident page and paging is in progress, we
2035 * have to wait until paging is complete but otherwise can act on the
2038 * This routine must be called at splvm().
2040 * SWM_FREE remove and free swap block from metadata
2041 * SWM_POP remove from meta data but do not free.. pop it out
2050 struct swblock **pswap;
2051 struct swblock *swap;
2055 * The meta data only exists of the object is OBJT_SWAP
2056 * and even then might not be allocated yet.
2059 if (object->type != OBJT_SWAP)
2060 return(SWAPBLK_NONE);
2063 pswap = swp_pager_hash(object, index);
2065 if ((swap = *pswap) != NULL) {
2066 index &= SWAP_META_MASK;
2067 r1 = swap->swb_pages[index];
2069 if (r1 != SWAPBLK_NONE) {
2070 if (flags & SWM_FREE) {
2071 swp_pager_freeswapspace(r1, 1);
2074 if (flags & (SWM_FREE|SWM_POP)) {
2075 swap->swb_pages[index] = SWAPBLK_NONE;
2076 if (--swap->swb_count == 0) {
2077 *pswap = swap->swb_hnext;
2078 zfree(swap_zone, swap);
2079 --object->un_pager.swp.swp_bcount;