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 1.32 2008/07/01 02:02:56 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 SWBIO_READ 0x01
139 #define SWBIO_WRITE 0x02
140 #define SWBIO_SYNC 0x04
143 * vm_swap_size is in page-sized chunks now. It was DEV_BSIZE'd chunks
147 extern int vm_swap_size
; /* number of free swap blocks, in pages */
149 int swap_pager_full
; /* swap space exhaustion (task killing) */
150 static int swap_pager_almost_full
; /* swap space exhaustion (w/ hysteresis)*/
151 static int nsw_rcount
; /* free read buffers */
152 static int nsw_wcount_sync
; /* limit write buffers / synchronous */
153 static int nsw_wcount_async
; /* limit write buffers / asynchronous */
154 static int nsw_wcount_async_max
;/* assigned maximum */
155 static int nsw_cluster_max
; /* maximum VOP I/O allowed */
156 static int sw_alloc_interlock
; /* swap pager allocation interlock */
158 struct blist
*swapblist
;
159 static struct swblock
**swhash
;
160 static int swhash_mask
;
161 static int swap_async_max
= 4; /* maximum in-progress async I/O's */
163 extern struct vnode
*swapdev_vp
; /* from vm_swap.c */
165 SYSCTL_INT(_vm
, OID_AUTO
, swap_async_max
,
166 CTLFLAG_RW
, &swap_async_max
, 0, "Maximum running async swap ops");
169 * "named" and "unnamed" anon region objects. Try to reduce the overhead
170 * of searching a named list by hashing it just a little.
175 #define NOBJLIST(handle) \
176 (&swap_pager_object_list[((int)(intptr_t)handle >> 4) & (NOBJLISTS-1)])
178 static struct pagerlst swap_pager_object_list
[NOBJLISTS
];
179 struct pagerlst swap_pager_un_object_list
;
183 * pagerops for OBJT_SWAP - "swap pager". Some ops are also global procedure
184 * calls hooked from other parts of the VM system and do not appear here.
185 * (see vm/swap_pager.h).
189 swap_pager_alloc (void *handle
, off_t size
,
190 vm_prot_t prot
, off_t offset
);
191 static void swap_pager_dealloc (vm_object_t object
);
192 static int swap_pager_getpage (vm_object_t
, vm_page_t
*, int);
193 static void swap_pager_init (void);
194 static void swap_pager_unswapped (vm_page_t
);
195 static void swap_pager_strategy (vm_object_t
, struct bio
*);
196 static void swap_chain_iodone(struct bio
*biox
);
198 struct pagerops swappagerops
= {
199 swap_pager_init
, /* early system initialization of pager */
200 swap_pager_alloc
, /* allocate an OBJT_SWAP object */
201 swap_pager_dealloc
, /* deallocate an OBJT_SWAP object */
202 swap_pager_getpage
, /* pagein */
203 swap_pager_putpages
, /* pageout */
204 swap_pager_haspage
, /* get backing store status for page */
205 swap_pager_unswapped
, /* remove swap related to page */
206 swap_pager_strategy
/* pager strategy call */
210 * dmmax is in page-sized chunks with the new swap system. It was
211 * dev-bsized chunks in the old. dmmax is always a power of 2.
213 * swap_*() routines are externally accessible. swp_*() routines are
218 static int dmmax_mask
;
219 int nswap_lowat
= 128; /* in pages, swap_pager_almost_full warn */
220 int nswap_hiwat
= 512; /* in pages, swap_pager_almost_full warn */
222 static __inline
void swp_sizecheck (void);
223 static void swp_pager_async_iodone (struct bio
*bio
);
226 * Swap bitmap functions
229 static __inline
void swp_pager_freeswapspace (daddr_t blk
, int npages
);
230 static __inline daddr_t
swp_pager_getswapspace (int npages
);
236 static void swp_pager_meta_build (vm_object_t
, vm_pindex_t
, daddr_t
);
237 static void swp_pager_meta_free (vm_object_t
, vm_pindex_t
, daddr_t
);
238 static void swp_pager_meta_free_all (vm_object_t
);
239 static daddr_t
swp_pager_meta_ctl (vm_object_t
, vm_pindex_t
, int);
242 * SWP_SIZECHECK() - update swap_pager_full indication
244 * update the swap_pager_almost_full indication and warn when we are
245 * about to run out of swap space, using lowat/hiwat hysteresis.
247 * Clear swap_pager_full ( task killing ) indication when lowat is met.
249 * No restrictions on call
250 * This routine may not block.
251 * This routine must be called at splvm()
257 if (vm_swap_size
< nswap_lowat
) {
258 if (swap_pager_almost_full
== 0) {
259 kprintf("swap_pager: out of swap space\n");
260 swap_pager_almost_full
= 1;
264 if (vm_swap_size
> nswap_hiwat
)
265 swap_pager_almost_full
= 0;
270 * SWAP_PAGER_INIT() - initialize the swap pager!
272 * Expected to be started from system init. NOTE: This code is run
273 * before much else so be careful what you depend on. Most of the VM
274 * system has yet to be initialized at this point.
278 swap_pager_init(void)
281 * Initialize object lists
285 for (i
= 0; i
< NOBJLISTS
; ++i
)
286 TAILQ_INIT(&swap_pager_object_list
[i
]);
287 TAILQ_INIT(&swap_pager_un_object_list
);
290 * Device Stripe, in PAGE_SIZE'd blocks
293 dmmax
= SWB_NPAGES
* 2;
294 dmmax_mask
= ~(dmmax
- 1);
298 * SWAP_PAGER_SWAP_INIT() - swap pager initialization from pageout process
300 * Expected to be started from pageout process once, prior to entering
305 swap_pager_swap_init(void)
310 * Number of in-transit swap bp operations. Don't
311 * exhaust the pbufs completely. Make sure we
312 * initialize workable values (0 will work for hysteresis
313 * but it isn't very efficient).
315 * The nsw_cluster_max is constrained by the number of pages an XIO
316 * holds, i.e., (MAXPHYS/PAGE_SIZE) and our locally defined
317 * MAX_PAGEOUT_CLUSTER. Also be aware that swap ops are
318 * constrained by the swap device interleave stripe size.
320 * Currently we hardwire nsw_wcount_async to 4. This limit is
321 * designed to prevent other I/O from having high latencies due to
322 * our pageout I/O. The value 4 works well for one or two active swap
323 * devices but is probably a little low if you have more. Even so,
324 * a higher value would probably generate only a limited improvement
325 * with three or four active swap devices since the system does not
326 * typically have to pageout at extreme bandwidths. We will want
327 * at least 2 per swap devices, and 4 is a pretty good value if you
328 * have one NFS swap device due to the command/ack latency over NFS.
329 * So it all works out pretty well.
332 nsw_cluster_max
= min((MAXPHYS
/PAGE_SIZE
), MAX_PAGEOUT_CLUSTER
);
334 nsw_rcount
= (nswbuf
+ 1) / 2;
335 nsw_wcount_sync
= (nswbuf
+ 3) / 4;
336 nsw_wcount_async
= 4;
337 nsw_wcount_async_max
= nsw_wcount_async
;
340 * The zone is dynamically allocated so generally size it to
341 * maxswzone (32MB to 512MB of KVM). Set a minimum size based
342 * on physical memory of around 8x (each swblock can hold 16 pages).
344 * With the advent of SSDs (vs HDs) the practical (swap:memory) ratio
345 * has increased dramatically.
347 n
= vmstats
.v_page_count
/ 2;
348 if (maxswzone
&& n
< maxswzone
/ sizeof(struct swblock
))
349 n
= maxswzone
/ sizeof(struct swblock
);
355 sizeof(struct swblock
),
359 if (swap_zone
!= NULL
)
362 * if the allocation failed, try a zone two thirds the
363 * size of the previous attempt.
368 if (swap_zone
== NULL
)
369 panic("swap_pager_swap_init: swap_zone == NULL");
371 kprintf("Swap zone entries reduced from %d to %d.\n", n2
, n
);
375 * Initialize our meta-data hash table. The swapper does not need to
376 * be quite as efficient as the VM system, so we do not use an
377 * oversized hash table.
379 * n: size of hash table, must be power of 2
380 * swhash_mask: hash table index mask
383 for (n
= 1; n
< n2
/ 8; n
*= 2)
386 swhash
= kmalloc(sizeof(struct swblock
*) * n
, M_VMPGDATA
,
393 * SWAP_PAGER_ALLOC() - allocate a new OBJT_SWAP VM object and instantiate
394 * its metadata structures.
396 * This routine is called from the mmap and fork code to create a new
397 * OBJT_SWAP object. We do this by creating an OBJT_DEFAULT object
398 * and then converting it with swp_pager_meta_build().
400 * This routine may block in vm_object_allocate() and create a named
401 * object lookup race, so we must interlock. We must also run at
402 * splvm() for the object lookup to handle races with interrupts, but
403 * we do not have to maintain splvm() in between the lookup and the
404 * add because (I believe) it is not possible to attempt to create
405 * a new swap object w/handle when a default object with that handle
410 swap_pager_alloc(void *handle
, off_t size
, vm_prot_t prot
, off_t offset
)
416 * Reference existing named region or allocate new one. There
417 * should not be a race here against swp_pager_meta_build()
418 * as called from vm_page_remove() in regards to the lookup
422 while (sw_alloc_interlock
) {
423 sw_alloc_interlock
= -1;
424 tsleep(&sw_alloc_interlock
, 0, "swpalc", 0);
426 sw_alloc_interlock
= 1;
428 object
= vm_pager_object_lookup(NOBJLIST(handle
), handle
);
430 if (object
!= NULL
) {
431 vm_object_reference(object
);
433 object
= vm_object_allocate(OBJT_DEFAULT
,
434 OFF_TO_IDX(offset
+ PAGE_MASK
+ size
));
435 object
->handle
= handle
;
437 swp_pager_meta_build(object
, 0, SWAPBLK_NONE
);
440 if (sw_alloc_interlock
< 0)
441 wakeup(&sw_alloc_interlock
);
443 sw_alloc_interlock
= 0;
445 object
= vm_object_allocate(OBJT_DEFAULT
,
446 OFF_TO_IDX(offset
+ PAGE_MASK
+ size
));
448 swp_pager_meta_build(object
, 0, SWAPBLK_NONE
);
455 * SWAP_PAGER_DEALLOC() - remove swap metadata from object
457 * The swap backing for the object is destroyed. The code is
458 * designed such that we can reinstantiate it later, but this
459 * routine is typically called only when the entire object is
460 * about to be destroyed.
462 * This routine may block, but no longer does.
464 * The object must be locked or unreferenceable.
468 swap_pager_dealloc(vm_object_t object
)
471 * Remove from list right away so lookups will fail if we block for
472 * pageout completion.
475 if (object
->handle
== NULL
) {
476 TAILQ_REMOVE(&swap_pager_un_object_list
, object
, pager_object_list
);
478 TAILQ_REMOVE(NOBJLIST(object
->handle
), object
, pager_object_list
);
481 vm_object_pip_wait(object
, "swpdea");
484 * Free all remaining metadata. We only bother to free it from
485 * the swap meta data. We do not attempt to free swapblk's still
486 * associated with vm_page_t's for this object. We do not care
487 * if paging is still in progress on some objects.
490 swp_pager_meta_free_all(object
);
494 /************************************************************************
495 * SWAP PAGER BITMAP ROUTINES *
496 ************************************************************************/
499 * SWP_PAGER_GETSWAPSPACE() - allocate raw swap space
501 * Allocate swap for the requested number of pages. The starting
502 * swap block number (a page index) is returned or SWAPBLK_NONE
503 * if the allocation failed.
505 * Also has the side effect of advising that somebody made a mistake
506 * when they configured swap and didn't configure enough.
508 * Must be called at splvm() to avoid races with bitmap frees from
509 * vm_page_remove() aka swap_pager_page_removed().
511 * This routine may not block
512 * This routine must be called at splvm().
515 static __inline daddr_t
516 swp_pager_getswapspace(int npages
)
520 if ((blk
= blist_alloc(swapblist
, npages
)) == SWAPBLK_NONE
) {
521 if (swap_pager_full
!= 2) {
522 kprintf("swap_pager_getswapspace: failed\n");
524 swap_pager_almost_full
= 1;
527 vm_swap_size
-= npages
;
534 * SWP_PAGER_FREESWAPSPACE() - free raw swap space
536 * This routine returns the specified swap blocks back to the bitmap.
538 * Note: This routine may not block (it could in the old swap code),
539 * and through the use of the new blist routines it does not block.
541 * We must be called at splvm() to avoid races with bitmap frees from
542 * vm_page_remove() aka swap_pager_page_removed().
544 * This routine may not block
545 * This routine must be called at splvm().
549 swp_pager_freeswapspace(daddr_t blk
, int npages
)
551 blist_free(swapblist
, blk
, npages
);
552 vm_swap_size
+= npages
;
557 * SWAP_PAGER_FREESPACE() - frees swap blocks associated with a page
558 * range within an object.
560 * This is a globally accessible routine.
562 * This routine removes swapblk assignments from swap metadata.
564 * The external callers of this routine typically have already destroyed
565 * or renamed vm_page_t's associated with this range in the object so
568 * This routine may be called at any spl. We up our spl to splvm temporarily
569 * in order to perform the metadata removal.
573 swap_pager_freespace(vm_object_t object
, vm_pindex_t start
, vm_size_t size
)
576 swp_pager_meta_free(object
, start
, size
);
581 * SWAP_PAGER_RESERVE() - reserve swap blocks in object
583 * Assigns swap blocks to the specified range within the object. The
584 * swap blocks are not zerod. Any previous swap assignment is destroyed.
586 * Returns 0 on success, -1 on failure.
590 swap_pager_reserve(vm_object_t object
, vm_pindex_t start
, vm_size_t size
)
593 daddr_t blk
= SWAPBLK_NONE
;
594 vm_pindex_t beg
= start
; /* save start index */
600 while ((blk
= swp_pager_getswapspace(n
)) == SWAPBLK_NONE
) {
603 swp_pager_meta_free(object
, beg
, start
- beg
);
609 swp_pager_meta_build(object
, start
, blk
);
615 swp_pager_meta_free(object
, start
, n
);
621 * SWAP_PAGER_COPY() - copy blocks from source pager to destination pager
622 * and destroy the source.
624 * Copy any valid swapblks from the source to the destination. In
625 * cases where both the source and destination have a valid swapblk,
626 * we keep the destination's.
628 * This routine is allowed to block. It may block allocating metadata
629 * indirectly through swp_pager_meta_build() or if paging is still in
630 * progress on the source.
632 * This routine can be called at any spl
634 * XXX vm_page_collapse() kinda expects us not to block because we
635 * supposedly do not need to allocate memory, but for the moment we
636 * *may* have to get a little memory from the zone allocator, but
637 * it is taken from the interrupt memory. We should be ok.
639 * The source object contains no vm_page_t's (which is just as well)
641 * The source object is of type OBJT_SWAP.
643 * The source and destination objects must be locked or
644 * inaccessible (XXX are they ?)
648 swap_pager_copy(vm_object_t srcobject
, vm_object_t dstobject
,
649 vm_pindex_t offset
, int destroysource
)
656 * If destroysource is set, we remove the source object from the
657 * swap_pager internal queue now.
661 if (srcobject
->handle
== NULL
) {
663 &swap_pager_un_object_list
,
669 NOBJLIST(srcobject
->handle
),
677 * transfer source to destination.
680 for (i
= 0; i
< dstobject
->size
; ++i
) {
684 * Locate (without changing) the swapblk on the destination,
685 * unless it is invalid in which case free it silently, or
686 * if the destination is a resident page, in which case the
687 * source is thrown away.
690 dstaddr
= swp_pager_meta_ctl(dstobject
, i
, 0);
692 if (dstaddr
== SWAPBLK_NONE
) {
694 * Destination has no swapblk and is not resident,
699 srcaddr
= swp_pager_meta_ctl(
705 if (srcaddr
!= SWAPBLK_NONE
)
706 swp_pager_meta_build(dstobject
, i
, srcaddr
);
709 * Destination has valid swapblk or it is represented
710 * by a resident page. We destroy the sourceblock.
713 swp_pager_meta_ctl(srcobject
, i
+ offset
, SWM_FREE
);
718 * Free left over swap blocks in source.
720 * We have to revert the type to OBJT_DEFAULT so we do not accidently
721 * double-remove the object from the swap queues.
725 swp_pager_meta_free_all(srcobject
);
727 * Reverting the type is not necessary, the caller is going
728 * to destroy srcobject directly, but I'm doing it here
729 * for consistency since we've removed the object from its
732 srcobject
->type
= OBJT_DEFAULT
;
738 * SWAP_PAGER_HASPAGE() - determine if we have good backing store for
739 * the requested page.
741 * We determine whether good backing store exists for the requested
742 * page and return TRUE if it does, FALSE if it doesn't.
744 * If TRUE, we also try to determine how much valid, contiguous backing
745 * store exists before and after the requested page within a reasonable
746 * distance. We do not try to restrict it to the swap device stripe
747 * (that is handled in getpages/putpages). It probably isn't worth
752 swap_pager_haspage(vm_object_t object
, vm_pindex_t pindex
)
757 * do we have good backing store at the requested index ?
761 blk0
= swp_pager_meta_ctl(object
, pindex
, 0);
763 if (blk0
== SWAPBLK_NONE
) {
770 * 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);
809 * SWAP_PAGER_PAGE_UNSWAPPED() - remove swap backing store related to page
811 * This removes any associated swap backing store, whether valid or
812 * not, from the page.
814 * This routine is typically called when a page is made dirty, at
815 * which point any associated swap can be freed. MADV_FREE also
816 * calls us in a special-case situation
818 * NOTE!!! If the page is clean and the swap was valid, the caller
819 * should make the page dirty before calling this routine. This routine
820 * does NOT change the m->dirty status of the page. Also: MADV_FREE
823 * This routine may not block
824 * This routine must be called at splvm()
828 swap_pager_unswapped(vm_page_t m
)
830 swp_pager_meta_ctl(m
->object
, m
->pindex
, SWM_FREE
);
834 * SWAP_PAGER_STRATEGY() - read, write, free blocks
836 * This implements the vm_pager_strategy() interface to swap and allows
837 * other parts of the system to directly access swap as backing store
838 * through vm_objects of type OBJT_SWAP. This is intended to be a
839 * cacheless interface ( i.e. caching occurs at higher levels ).
840 * Therefore we do not maintain any resident pages. All I/O goes
841 * directly to and from the swap device.
843 * We currently attempt to run I/O synchronously or asynchronously as
844 * the caller requests. This isn't perfect because we loose error
845 * sequencing when we run multiple ops in parallel to satisfy a request.
846 * But this is swap, so we let it all hang out.
850 swap_pager_strategy(vm_object_t object
, struct bio
*bio
)
852 struct buf
*bp
= bio
->bio_buf
;
855 vm_pindex_t biox_blkno
= 0;
860 struct bio_track
*track
;
863 * tracking for swapdev vnode I/Os
865 if (bp
->b_cmd
== BUF_CMD_READ
)
866 track
= &swapdev_vp
->v_track_read
;
868 track
= &swapdev_vp
->v_track_write
;
870 if (bp
->b_bcount
& PAGE_MASK
) {
871 bp
->b_error
= EINVAL
;
872 bp
->b_flags
|= B_ERROR
| B_INVAL
;
874 kprintf("swap_pager_strategy: bp %p offset %lld size %d, "
875 "not page bounded\n",
876 bp
, (long long)bio
->bio_offset
, (int)bp
->b_bcount
);
881 * Clear error indication, initialize page index, count, data pointer.
884 bp
->b_flags
&= ~B_ERROR
;
885 bp
->b_resid
= bp
->b_bcount
;
887 start
= (vm_pindex_t
)(bio
->bio_offset
>> PAGE_SHIFT
);
888 count
= howmany(bp
->b_bcount
, PAGE_SIZE
);
892 * Deal with BUF_CMD_FREEBLKS
894 if (bp
->b_cmd
== BUF_CMD_FREEBLKS
) {
896 * FREE PAGE(s) - destroy underlying swap that is no longer
899 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
;
922 * Execute read or write
928 * Obtain block. If block not found and writing, allocate a
929 * new block and build it into the object.
931 blk
= swp_pager_meta_ctl(object
, start
, 0);
932 if ((blk
== SWAPBLK_NONE
) && bp
->b_cmd
!= BUF_CMD_READ
) {
933 blk
= swp_pager_getswapspace(1);
934 if (blk
== SWAPBLK_NONE
) {
935 bp
->b_error
= ENOMEM
;
936 bp
->b_flags
|= B_ERROR
;
939 swp_pager_meta_build(object
, start
, blk
);
943 * Do we have to flush our current collection? Yes if:
945 * - no swap block at this index
946 * - swap block is not contiguous
947 * - we cross a physical disk boundry in the
951 biox
&& (biox_blkno
+ btoc(bufx
->b_bcount
) != blk
||
952 ((biox_blkno
^ blk
) & dmmax_mask
)
955 if (bp
->b_cmd
== BUF_CMD_READ
) {
956 ++mycpu
->gd_cnt
.v_swapin
;
957 mycpu
->gd_cnt
.v_swappgsin
+= btoc(bufx
->b_bcount
);
959 ++mycpu
->gd_cnt
.v_swapout
;
960 mycpu
->gd_cnt
.v_swappgsout
+= btoc(bufx
->b_bcount
);
961 bufx
->b_dirtyend
= bufx
->b_bcount
;
965 * Finished with this buf.
967 KKASSERT(bufx
->b_bcount
!= 0);
968 if (bufx
->b_cmd
!= BUF_CMD_READ
)
969 bufx
->b_dirtyend
= bufx
->b_bcount
;
975 * Add new swapblk to biox, instantiating biox if necessary.
976 * Zero-fill reads are able to take a shortcut.
978 if (blk
== SWAPBLK_NONE
) {
980 * We can only get here if we are reading. Since
981 * we are at splvm() we can safely modify b_resid,
982 * even if chain ops are in progress.
984 bzero(data
, PAGE_SIZE
);
985 bp
->b_resid
-= PAGE_SIZE
;
988 /* XXX chain count > 4, wait to <= 4 */
990 bufx
= getpbuf(NULL
);
991 biox
= &bufx
->b_bio1
;
992 cluster_append(nbio
, bufx
);
993 bufx
->b_flags
|= (bufx
->b_flags
& B_ORDERED
);
994 bufx
->b_cmd
= bp
->b_cmd
;
995 biox
->bio_done
= swap_chain_iodone
;
996 biox
->bio_offset
= (off_t
)blk
<< PAGE_SHIFT
;
997 biox
->bio_caller_info1
.cluster_parent
= nbio
;
1000 bufx
->b_data
= data
;
1002 bufx
->b_bcount
+= PAGE_SIZE
;
1010 * Flush out last buffer
1013 if (bufx
->b_cmd
== BUF_CMD_READ
) {
1014 ++mycpu
->gd_cnt
.v_swapin
;
1015 mycpu
->gd_cnt
.v_swappgsin
+= btoc(bufx
->b_bcount
);
1017 ++mycpu
->gd_cnt
.v_swapout
;
1018 mycpu
->gd_cnt
.v_swappgsout
+= btoc(bufx
->b_bcount
);
1019 bufx
->b_dirtyend
= bufx
->b_bcount
;
1021 KKASSERT(bufx
->b_bcount
);
1022 if (bufx
->b_cmd
!= BUF_CMD_READ
)
1023 bufx
->b_dirtyend
= bufx
->b_bcount
;
1024 /* biox, bufx = NULL */
1028 * Now initiate all the I/O. Be careful looping on our chain as
1029 * I/O's may complete while we are still initiating them.
1031 nbio
->bio_caller_info2
.cluster_tail
= NULL
;
1032 bufx
= nbio
->bio_caller_info1
.cluster_head
;
1035 biox
= &bufx
->b_bio1
;
1037 bufx
= bufx
->b_cluster_next
;
1038 vn_strategy(swapdev_vp
, biox
);
1042 * Completion of the cluster will also call biodone_chain(nbio).
1043 * We never call biodone(nbio) so we don't have to worry about
1044 * setting up a bio_done callback. It's handled in the sub-IO.
1050 swap_chain_iodone(struct bio
*biox
)
1053 struct buf
*bufx
; /* chained sub-buffer */
1054 struct bio
*nbio
; /* parent nbio with chain glue */
1055 struct buf
*bp
; /* original bp associated with nbio */
1058 bufx
= biox
->bio_buf
;
1059 nbio
= biox
->bio_caller_info1
.cluster_parent
;
1063 * Update the original buffer
1065 KKASSERT(bp
!= NULL
);
1066 if (bufx
->b_flags
& B_ERROR
) {
1067 atomic_set_int(&bufx
->b_flags
, B_ERROR
);
1068 bp
->b_error
= bufx
->b_error
;
1069 } else if (bufx
->b_resid
!= 0) {
1070 atomic_set_int(&bufx
->b_flags
, B_ERROR
);
1071 bp
->b_error
= EINVAL
;
1073 atomic_subtract_int(&bp
->b_resid
, bufx
->b_bcount
);
1077 * Remove us from the chain.
1079 spin_lock_wr(&bp
->b_lock
.lk_spinlock
);
1080 nextp
= &nbio
->bio_caller_info1
.cluster_head
;
1081 while (*nextp
!= bufx
) {
1082 KKASSERT(*nextp
!= NULL
);
1083 nextp
= &(*nextp
)->b_cluster_next
;
1085 *nextp
= bufx
->b_cluster_next
;
1086 chain_empty
= (nbio
->bio_caller_info1
.cluster_head
== NULL
);
1087 spin_unlock_wr(&bp
->b_lock
.lk_spinlock
);
1090 * Clean up bufx. If the chain is now empty we finish out
1091 * the parent. Note that we may be racing other completions
1092 * so we must use the chain_empty status from above.
1095 if (bp
->b_resid
!= 0 && !(bp
->b_flags
& B_ERROR
)) {
1096 atomic_set_int(&bp
->b_flags
, B_ERROR
);
1097 bp
->b_error
= EINVAL
;
1099 biodone_chain(nbio
);
1101 relpbuf(bufx
, NULL
);
1105 * SWAP_PAGER_GETPAGES() - bring pages in from swap
1107 * Attempt to retrieve (m, count) pages from backing store, but make
1108 * sure we retrieve at least m[reqpage]. We try to load in as large
1109 * a chunk surrounding m[reqpage] as is contiguous in swap and which
1110 * belongs to the same object.
1112 * The code is designed for asynchronous operation and
1113 * immediate-notification of 'reqpage' but tends not to be
1114 * used that way. Please do not optimize-out this algorithmic
1115 * feature, I intend to improve on it in the future.
1117 * The parent has a single vm_object_pip_add() reference prior to
1118 * calling us and we should return with the same.
1120 * The parent has BUSY'd the pages. We should return with 'm'
1121 * left busy, but the others adjusted.
1125 swap_pager_getpage(vm_object_t object
, vm_page_t
*mpp
, int seqaccess
)
1135 vm_pindex_t lastpindex
;
1139 if (mreq
->object
!= object
) {
1140 panic("swap_pager_getpages: object mismatch %p/%p",
1147 * Calculate range to retrieve. The pages have already been assigned
1148 * their swapblks. We require a *contiguous* range that falls entirely
1149 * within a single device stripe. If we do not supply it, bad things
1150 * happen. Note that blk, iblk & jblk can be SWAPBLK_NONE, but the
1151 * loops are set up such that the case(s) are handled implicitly.
1153 * The swp_*() calls must be made at splvm(). vm_page_free() does
1154 * not need to be, but it will go a little faster if it is.
1157 blk
= swp_pager_meta_ctl(mreq
->object
, mreq
->pindex
, 0);
1160 for (i
= reqpage
- 1; i
>= 0; --i
) {
1163 iblk
= swp_pager_meta_ctl(m
[i
]->object
, m
[i
]->pindex
, 0);
1164 if (blk
!= iblk
+ (reqpage
- i
))
1166 if ((blk
^ iblk
) & dmmax_mask
)
1171 for (j
= reqpage
+ 1; j
< count
; ++j
) {
1174 jblk
= swp_pager_meta_ctl(m
[j
]->object
, m
[j
]->pindex
, 0);
1175 if (blk
!= jblk
- (j
- reqpage
))
1177 if ((blk
^ jblk
) & dmmax_mask
)
1182 * free pages outside our collection range. Note: we never free
1183 * mreq, it must remain busy throughout.
1189 for (k
= 0; k
< i
; ++k
)
1191 for (k
= j
; k
< count
; ++k
)
1199 * Return VM_PAGER_FAIL if we have nothing to do. Return mreq
1200 * still busy, but the others unbusied.
1203 if (blk
== SWAPBLK_NONE
)
1204 return(VM_PAGER_FAIL
);
1207 * Get a swap buffer header to perform the IO
1210 bp
= getpbuf(&nsw_rcount
);
1212 kva
= (vm_offset_t
) bp
->b_data
;
1215 * map our page(s) into kva for input
1221 pmap_qenter(kva
, mpp
+ i
, j
- i
);
1223 bp
->b_data
= (caddr_t
) kva
;
1224 bp
->b_bcount
= PAGE_SIZE
* (j
- i
);
1225 bio
->bio_done
= swp_pager_async_iodone
;
1226 bio
->bio_offset
= (off_t
)(blk
- (reqpage
- i
)) << PAGE_SHIFT
;
1227 bio
->bio_driver_info
= (void *)(intptr_t)(reqpage
- i
);
1228 bio
->bio_caller_info1
.index
= SWBIO_READ
;
1233 for (k
= i
; k
< j
; ++k
) {
1234 bp
->b_xio
.xio_pages
[k
- i
] = mpp
[k
];
1235 vm_page_flag_set(mpp
[k
], PG_SWAPINPROG
);
1238 bp
->b_xio
.xio_npages
= j
- i
;
1240 mycpu
->gd_cnt
.v_swapin
++;
1241 mycpu
->gd_cnt
.v_swappgsin
+= bp
->b_xio
.xio_npages
;
1244 * We still hold the lock on mreq, and our automatic completion routine
1245 * does not remove it.
1248 vm_object_pip_add(mreq
->object
, bp
->b_xio
.xio_npages
);
1249 lastpindex
= mpp
[j
-1]->pindex
;
1252 * perform the I/O. NOTE!!! bp cannot be considered valid after
1253 * this point because we automatically release it on completion.
1254 * Instead, we look at the one page we are interested in which we
1255 * still hold a lock on even through the I/O completion.
1257 * The other pages in our m[] array are also released on completion,
1258 * so we cannot assume they are valid anymore either.
1261 bp
->b_cmd
= BUF_CMD_READ
;
1263 vn_strategy(swapdev_vp
, bio
);
1266 * wait for the page we want to complete. PG_SWAPINPROG is always
1267 * cleared on completion. If an I/O error occurs, SWAPBLK_NONE
1268 * is set in the meta-data.
1273 while ((mreq
->flags
& PG_SWAPINPROG
) != 0) {
1274 vm_page_flag_set(mreq
, PG_WANTED
| PG_REFERENCED
);
1275 mycpu
->gd_cnt
.v_intrans
++;
1276 if (tsleep(mreq
, 0, "swread", hz
*20)) {
1278 "swap_pager: indefinite wait buffer: "
1279 " offset: %lld, size: %ld\n",
1280 (long long)bio
->bio_offset
,
1289 * mreq is left bussied after completion, but all the other pages
1290 * are freed. If we had an unrecoverable read error the page will
1294 if (mreq
->valid
!= VM_PAGE_BITS_ALL
) {
1295 return(VM_PAGER_ERROR
);
1297 return(VM_PAGER_OK
);
1301 * A final note: in a low swap situation, we cannot deallocate swap
1302 * and mark a page dirty here because the caller is likely to mark
1303 * the page clean when we return, causing the page to possibly revert
1304 * to all-zero's later.
1309 * swap_pager_putpages:
1311 * Assign swap (if necessary) and initiate I/O on the specified pages.
1313 * We support both OBJT_DEFAULT and OBJT_SWAP objects. DEFAULT objects
1314 * are automatically converted to SWAP objects.
1316 * In a low memory situation we may block in vn_strategy(), but the new
1317 * vm_page reservation system coupled with properly written VFS devices
1318 * should ensure that no low-memory deadlock occurs. This is an area
1321 * The parent has N vm_object_pip_add() references prior to
1322 * calling us and will remove references for rtvals[] that are
1323 * not set to VM_PAGER_PEND. We need to remove the rest on I/O
1326 * The parent has soft-busy'd the pages it passes us and will unbusy
1327 * those whos rtvals[] entry is not set to VM_PAGER_PEND on return.
1328 * We need to unbusy the rest on I/O completion.
1331 swap_pager_putpages(vm_object_t object
, vm_page_t
*m
, int count
,
1332 boolean_t sync
, int *rtvals
)
1337 if (count
&& m
[0]->object
!= object
) {
1338 panic("swap_pager_getpages: object mismatch %p/%p",
1347 * Turn object into OBJT_SWAP
1348 * check for bogus sysops
1349 * force sync if not pageout process
1352 if (object
->type
!= OBJT_SWAP
)
1353 swp_pager_meta_build(object
, 0, SWAPBLK_NONE
);
1355 if (curthread
!= pagethread
)
1361 * Update nsw parameters from swap_async_max sysctl values.
1362 * Do not let the sysop crash the machine with bogus numbers.
1365 if (swap_async_max
!= nsw_wcount_async_max
) {
1371 if ((n
= swap_async_max
) > nswbuf
/ 2)
1378 * Adjust difference ( if possible ). If the current async
1379 * count is too low, we may not be able to make the adjustment
1383 n
-= nsw_wcount_async_max
;
1384 if (nsw_wcount_async
+ n
>= 0) {
1385 nsw_wcount_async
+= n
;
1386 nsw_wcount_async_max
+= n
;
1387 wakeup(&nsw_wcount_async
);
1395 * Assign swap blocks and issue I/O. We reallocate swap on the fly.
1396 * The page is left dirty until the pageout operation completes
1400 for (i
= 0; i
< count
; i
+= n
) {
1407 * Maximum I/O size is limited by a number of factors.
1410 n
= min(BLIST_MAX_ALLOC
, count
- i
);
1411 n
= min(n
, nsw_cluster_max
);
1416 * Get biggest block of swap we can. If we fail, fall
1417 * back and try to allocate a smaller block. Don't go
1418 * overboard trying to allocate space if it would overly
1422 (blk
= swp_pager_getswapspace(n
)) == SWAPBLK_NONE
&&
1427 if (blk
== SWAPBLK_NONE
) {
1428 for (j
= 0; j
< n
; ++j
)
1429 rtvals
[i
+j
] = VM_PAGER_FAIL
;
1435 * The I/O we are constructing cannot cross a physical
1436 * disk boundry in the swap stripe. Note: we are still
1439 if ((blk
^ (blk
+ n
)) & dmmax_mask
) {
1440 j
= ((blk
+ dmmax
) & dmmax_mask
) - blk
;
1441 swp_pager_freeswapspace(blk
+ j
, n
- j
);
1446 * All I/O parameters have been satisfied, build the I/O
1447 * request and assign the swap space.
1451 bp
= getpbuf(&nsw_wcount_sync
);
1453 bp
= getpbuf(&nsw_wcount_async
);
1456 pmap_qenter((vm_offset_t
)bp
->b_data
, &m
[i
], n
);
1458 bp
->b_bcount
= PAGE_SIZE
* n
;
1459 bio
->bio_offset
= (off_t
)blk
<< PAGE_SHIFT
;
1461 for (j
= 0; j
< n
; ++j
) {
1462 vm_page_t mreq
= m
[i
+j
];
1464 swp_pager_meta_build(
1469 vm_page_dirty(mreq
);
1470 rtvals
[i
+j
] = VM_PAGER_OK
;
1472 vm_page_flag_set(mreq
, PG_SWAPINPROG
);
1473 bp
->b_xio
.xio_pages
[j
] = mreq
;
1475 bp
->b_xio
.xio_npages
= n
;
1477 mycpu
->gd_cnt
.v_swapout
++;
1478 mycpu
->gd_cnt
.v_swappgsout
+= bp
->b_xio
.xio_npages
;
1482 bp
->b_dirtyoff
= 0; /* req'd for NFS */
1483 bp
->b_dirtyend
= bp
->b_bcount
; /* req'd for NFS */
1484 bp
->b_cmd
= BUF_CMD_WRITE
;
1485 bio
->bio_caller_info1
.index
= SWBIO_WRITE
;
1490 if (sync
== FALSE
) {
1491 bio
->bio_done
= swp_pager_async_iodone
;
1493 vn_strategy(swapdev_vp
, bio
);
1495 for (j
= 0; j
< n
; ++j
)
1496 rtvals
[i
+j
] = VM_PAGER_PEND
;
1501 * Issue synchrnously.
1503 * Wait for the sync I/O to complete, then update rtvals.
1504 * We just set the rtvals[] to VM_PAGER_PEND so we can call
1505 * our async completion routine at the end, thus avoiding a
1508 bio
->bio_caller_info1
.index
|= SWBIO_SYNC
;
1509 bio
->bio_done
= biodone_sync
;
1510 bio
->bio_flags
|= BIO_SYNC
;
1511 vn_strategy(swapdev_vp
, bio
);
1512 biowait(bio
, "swwrt");
1514 for (j
= 0; j
< n
; ++j
)
1515 rtvals
[i
+j
] = VM_PAGER_PEND
;
1518 * Now that we are through with the bp, we can call the
1519 * normal async completion, which frees everything up.
1521 swp_pager_async_iodone(bio
);
1526 swap_pager_newswap(void)
1532 * swp_pager_async_iodone:
1534 * Completion routine for asynchronous reads and writes from/to swap.
1535 * Also called manually by synchronous code to finish up a bp.
1537 * For READ operations, the pages are PG_BUSY'd. For WRITE operations,
1538 * the pages are vm_page_t->busy'd. For READ operations, we PG_BUSY
1539 * unbusy all pages except the 'main' request page. For WRITE
1540 * operations, we vm_page_t->busy'd unbusy all pages ( we can do this
1541 * because we marked them all VM_PAGER_PEND on return from putpages ).
1543 * This routine may not block.
1546 swp_pager_async_iodone(struct bio
*bio
)
1548 struct buf
*bp
= bio
->bio_buf
;
1549 vm_object_t object
= NULL
;
1556 if (bp
->b_flags
& B_ERROR
) {
1558 "swap_pager: I/O error - %s failed; offset %lld,"
1559 "size %ld, error %d\n",
1560 ((bio
->bio_caller_info1
.index
& SWBIO_READ
) ?
1561 "pagein" : "pageout"),
1562 (long long)bio
->bio_offset
,
1569 * set object, raise to splvm().
1571 if (bp
->b_xio
.xio_npages
)
1572 object
= bp
->b_xio
.xio_pages
[0]->object
;
1576 * remove the mapping for kernel virtual
1578 pmap_qremove((vm_offset_t
)bp
->b_data
, bp
->b_xio
.xio_npages
);
1581 * cleanup pages. If an error occurs writing to swap, we are in
1582 * very serious trouble. If it happens to be a disk error, though,
1583 * we may be able to recover by reassigning the swap later on. So
1584 * in this case we remove the m->swapblk assignment for the page
1585 * but do not free it in the rlist. The errornous block(s) are thus
1586 * never reallocated as swap. Redirty the page and continue.
1588 for (i
= 0; i
< bp
->b_xio
.xio_npages
; ++i
) {
1589 vm_page_t m
= bp
->b_xio
.xio_pages
[i
];
1591 vm_page_flag_clear(m
, PG_SWAPINPROG
);
1593 if (bp
->b_flags
& B_ERROR
) {
1595 * If an error occurs I'd love to throw the swapblk
1596 * away without freeing it back to swapspace, so it
1597 * can never be used again. But I can't from an
1601 if (bio
->bio_caller_info1
.index
& SWBIO_READ
) {
1603 * When reading, reqpage needs to stay
1604 * locked for the parent, but all other
1605 * pages can be freed. We still want to
1606 * wakeup the parent waiting on the page,
1607 * though. ( also: pg_reqpage can be -1 and
1608 * not match anything ).
1610 * We have to wake specifically requested pages
1611 * up too because we cleared PG_SWAPINPROG and
1612 * someone may be waiting for that.
1614 * NOTE: for reads, m->dirty will probably
1615 * be overridden by the original caller of
1616 * getpages so don't play cute tricks here.
1618 * NOTE: We can't actually free the page from
1619 * here, because this is an interrupt. It
1620 * is not legal to mess with object->memq
1621 * from an interrupt. Deactivate the page
1626 vm_page_flag_clear(m
, PG_ZERO
);
1629 * bio_driver_info holds the requested page
1632 if (i
!= (int)(intptr_t)bio
->bio_driver_info
) {
1633 vm_page_deactivate(m
);
1639 * If i == bp->b_pager.pg_reqpage, do not wake
1640 * the page up. The caller needs to.
1644 * If a write error occurs, reactivate page
1645 * so it doesn't clog the inactive list,
1646 * then finish the I/O.
1649 vm_page_activate(m
);
1650 vm_page_io_finish(m
);
1652 } else if (bio
->bio_caller_info1
.index
& SWBIO_READ
) {
1654 * NOTE: for reads, m->dirty will probably be
1655 * overridden by the original caller of getpages so
1656 * we cannot set them in order to free the underlying
1657 * swap in a low-swap situation. I don't think we'd
1658 * want to do that anyway, but it was an optimization
1659 * that existed in the old swapper for a time before
1660 * it got ripped out due to precisely this problem.
1662 * clear PG_ZERO in page.
1664 * If not the requested page then deactivate it.
1666 * Note that the requested page, reqpage, is left
1667 * busied, but we still have to wake it up. The
1668 * other pages are released (unbusied) by
1669 * vm_page_wakeup(). We do not set reqpage's
1670 * valid bits here, it is up to the caller.
1674 * NOTE: can't call pmap_clear_modify(m) from an
1675 * interrupt thread, the pmap code may have to map
1676 * non-kernel pmaps and currently asserts the case.
1678 /*pmap_clear_modify(m);*/
1679 m
->valid
= VM_PAGE_BITS_ALL
;
1681 vm_page_flag_clear(m
, PG_ZERO
);
1684 * We have to wake specifically requested pages
1685 * up too because we cleared PG_SWAPINPROG and
1686 * could be waiting for it in getpages. However,
1687 * be sure to not unbusy getpages specifically
1688 * requested page - getpages expects it to be
1691 * bio_driver_info holds the requested page
1693 if (i
!= (int)(intptr_t)bio
->bio_driver_info
) {
1694 vm_page_deactivate(m
);
1701 * Mark the page clean but do not mess with the
1702 * pmap-layer's modified state. That state should
1703 * also be clear since the caller protected the
1704 * page VM_PROT_READ, but allow the case.
1706 * We are in an interrupt, avoid pmap operations.
1708 * If we have a severe page deficit, deactivate the
1709 * page. Do not try to cache it (which would also
1710 * involve a pmap op), because the page might still
1714 vm_page_io_finish(m
);
1715 if (vm_page_count_severe())
1716 vm_page_deactivate(m
);
1718 if (!vm_page_count_severe() || !vm_page_try_to_cache(m
))
1719 vm_page_protect(m
, VM_PROT_READ
);
1725 * adjust pip. NOTE: the original parent may still have its own
1726 * pip refs on the object.
1730 vm_object_pip_wakeupn(object
, bp
->b_xio
.xio_npages
);
1733 * Release the physical I/O buffer.
1735 * NOTE: Due to synchronous operations in the write case b_cmd may
1736 * already be set to BUF_CMD_DONE and BIO_SYNC may have already
1739 if (bio
->bio_caller_info1
.index
& SWBIO_READ
)
1740 nswptr
= &nsw_rcount
;
1741 else if (bio
->bio_caller_info1
.index
& SWBIO_SYNC
)
1742 nswptr
= &nsw_wcount_sync
;
1744 nswptr
= &nsw_wcount_async
;
1745 bp
->b_cmd
= BUF_CMD_DONE
;
1746 relpbuf(bp
, nswptr
);
1750 /************************************************************************
1752 ************************************************************************
1754 * These routines manipulate the swap metadata stored in the
1755 * OBJT_SWAP object. All swp_*() routines must be called at
1756 * splvm() because swap can be freed up by the low level vm_page
1757 * code which might be called from interrupts beyond what splbio() covers.
1759 * Swap metadata is implemented with a global hash and not directly
1760 * linked into the object. Instead the object simply contains
1761 * appropriate tracking counters.
1765 * SWP_PAGER_HASH() - hash swap meta data
1767 * This is an inline helper function which hashes the swapblk given
1768 * the object and page index. It returns a pointer to a pointer
1769 * to the object, or a pointer to a NULL pointer if it could not
1772 * This routine must be called at splvm().
1775 static __inline
struct swblock
**
1776 swp_pager_hash(vm_object_t object
, vm_pindex_t index
)
1778 struct swblock
**pswap
;
1779 struct swblock
*swap
;
1781 index
&= ~SWAP_META_MASK
;
1782 pswap
= &swhash
[(index
^ (int)(intptr_t)object
) & swhash_mask
];
1784 while ((swap
= *pswap
) != NULL
) {
1785 if (swap
->swb_object
== object
&&
1786 swap
->swb_index
== index
1790 pswap
= &swap
->swb_hnext
;
1796 * SWP_PAGER_META_BUILD() - add swap block to swap meta data for object
1798 * We first convert the object to a swap object if it is a default
1801 * The specified swapblk is added to the object's swap metadata. If
1802 * the swapblk is not valid, it is freed instead. Any previously
1803 * assigned swapblk is freed.
1805 * This routine must be called at splvm(), except when used to convert
1806 * an OBJT_DEFAULT object into an OBJT_SWAP object.
1811 swp_pager_meta_build(
1816 struct swblock
*swap
;
1817 struct swblock
**pswap
;
1820 * Convert default object to swap object if necessary
1823 if (object
->type
!= OBJT_SWAP
) {
1824 object
->type
= OBJT_SWAP
;
1825 object
->un_pager
.swp
.swp_bcount
= 0;
1827 if (object
->handle
!= NULL
) {
1829 NOBJLIST(object
->handle
),
1835 &swap_pager_un_object_list
,
1843 * Locate hash entry. If not found create, but if we aren't adding
1844 * anything just return. If we run out of space in the map we wait
1845 * and, since the hash table may have changed, retry.
1849 pswap
= swp_pager_hash(object
, index
);
1851 if ((swap
= *pswap
) == NULL
) {
1854 if (swapblk
== SWAPBLK_NONE
)
1857 swap
= *pswap
= zalloc(swap_zone
);
1862 swap
->swb_hnext
= NULL
;
1863 swap
->swb_object
= object
;
1864 swap
->swb_index
= index
& ~SWAP_META_MASK
;
1865 swap
->swb_count
= 0;
1867 ++object
->un_pager
.swp
.swp_bcount
;
1869 for (i
= 0; i
< SWAP_META_PAGES
; ++i
)
1870 swap
->swb_pages
[i
] = SWAPBLK_NONE
;
1874 * Delete prior contents of metadata
1877 index
&= SWAP_META_MASK
;
1879 if (swap
->swb_pages
[index
] != SWAPBLK_NONE
) {
1880 swp_pager_freeswapspace(swap
->swb_pages
[index
], 1);
1885 * Enter block into metadata
1888 swap
->swb_pages
[index
] = swapblk
;
1889 if (swapblk
!= SWAPBLK_NONE
)
1894 * SWP_PAGER_META_FREE() - free a range of blocks in the object's swap metadata
1896 * The requested range of blocks is freed, with any associated swap
1897 * returned to the swap bitmap.
1899 * This routine will free swap metadata structures as they are cleaned
1900 * out. This routine does *NOT* operate on swap metadata associated
1901 * with resident pages.
1903 * This routine must be called at splvm()
1907 swp_pager_meta_free(vm_object_t object
, vm_pindex_t index
, daddr_t count
)
1909 if (object
->type
!= OBJT_SWAP
)
1913 struct swblock
**pswap
;
1914 struct swblock
*swap
;
1916 pswap
= swp_pager_hash(object
, index
);
1918 if ((swap
= *pswap
) != NULL
) {
1919 daddr_t v
= swap
->swb_pages
[index
& SWAP_META_MASK
];
1921 if (v
!= SWAPBLK_NONE
) {
1922 swp_pager_freeswapspace(v
, 1);
1923 swap
->swb_pages
[index
& SWAP_META_MASK
] =
1925 if (--swap
->swb_count
== 0) {
1926 *pswap
= swap
->swb_hnext
;
1927 zfree(swap_zone
, swap
);
1928 --object
->un_pager
.swp
.swp_bcount
;
1934 int n
= SWAP_META_PAGES
- (index
& SWAP_META_MASK
);
1942 * SWP_PAGER_META_FREE_ALL() - destroy all swap metadata associated with object
1944 * This routine locates and destroys all swap metadata associated with
1947 * This routine must be called at splvm()
1951 swp_pager_meta_free_all(vm_object_t object
)
1955 if (object
->type
!= OBJT_SWAP
)
1958 while (object
->un_pager
.swp
.swp_bcount
) {
1959 struct swblock
**pswap
;
1960 struct swblock
*swap
;
1962 pswap
= swp_pager_hash(object
, index
);
1963 if ((swap
= *pswap
) != NULL
) {
1966 for (i
= 0; i
< SWAP_META_PAGES
; ++i
) {
1967 daddr_t v
= swap
->swb_pages
[i
];
1968 if (v
!= SWAPBLK_NONE
) {
1970 swp_pager_freeswapspace(v
, 1);
1973 if (swap
->swb_count
!= 0)
1974 panic("swap_pager_meta_free_all: swb_count != 0");
1975 *pswap
= swap
->swb_hnext
;
1976 zfree(swap_zone
, swap
);
1977 --object
->un_pager
.swp
.swp_bcount
;
1979 index
+= SWAP_META_PAGES
;
1980 if (index
> 0x20000000)
1981 panic("swp_pager_meta_free_all: failed to locate all swap meta blocks");
1986 * SWP_PAGER_METACTL() - misc control of swap and vm_page_t meta data.
1988 * This routine is capable of looking up, popping, or freeing
1989 * swapblk assignments in the swap meta data or in the vm_page_t.
1990 * The routine typically returns the swapblk being looked-up, or popped,
1991 * or SWAPBLK_NONE if the block was freed, or SWAPBLK_NONE if the block
1992 * was invalid. This routine will automatically free any invalid
1993 * meta-data swapblks.
1995 * It is not possible to store invalid swapblks in the swap meta data
1996 * (other then a literal 'SWAPBLK_NONE'), so we don't bother checking.
1998 * When acting on a busy resident page and paging is in progress, we
1999 * have to wait until paging is complete but otherwise can act on the
2002 * This routine must be called at splvm().
2004 * SWM_FREE remove and free swap block from metadata
2005 * SWM_POP remove from meta data but do not free.. pop it out
2014 struct swblock
**pswap
;
2015 struct swblock
*swap
;
2019 * The meta data only exists of the object is OBJT_SWAP
2020 * and even then might not be allocated yet.
2023 if (object
->type
!= OBJT_SWAP
)
2024 return(SWAPBLK_NONE
);
2027 pswap
= swp_pager_hash(object
, index
);
2029 if ((swap
= *pswap
) != NULL
) {
2030 index
&= SWAP_META_MASK
;
2031 r1
= swap
->swb_pages
[index
];
2033 if (r1
!= SWAPBLK_NONE
) {
2034 if (flags
& SWM_FREE
) {
2035 swp_pager_freeswapspace(r1
, 1);
2038 if (flags
& (SWM_FREE
|SWM_POP
)) {
2039 swap
->swb_pages
[index
] = SWAPBLK_NONE
;
2040 if (--swap
->swb_count
== 0) {
2041 *pswap
= swap
->swb_hnext
;
2042 zfree(swap_zone
, swap
);
2043 --object
->un_pager
.swp
.swp_bcount
;