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>
131 #include <vm/vnode_pager.h>
133 #include <sys/buf2.h>
134 #include <vm/vm_page2.h>
136 #define SWM_FREE 0x02 /* free, period */
137 #define SWM_POP 0x04 /* pop out */
139 #define SWBIO_READ 0x01
140 #define SWBIO_WRITE 0x02
141 #define SWBIO_SYNC 0x04
147 vm_pindex_t endi
; /* inclusive */
151 * vm_swap_size is in page-sized chunks now. It was DEV_BSIZE'd chunks
155 int swap_pager_full
; /* swap space exhaustion (task killing) */
156 int vm_swap_cache_use
;
157 int vm_swap_anon_use
;
159 static int swap_pager_almost_full
; /* swap space exhaustion (w/ hysteresis)*/
160 static int nsw_rcount
; /* free read buffers */
161 static int nsw_wcount_sync
; /* limit write buffers / synchronous */
162 static int nsw_wcount_async
; /* limit write buffers / asynchronous */
163 static int nsw_wcount_async_max
;/* assigned maximum */
164 static int nsw_cluster_max
; /* maximum VOP I/O allowed */
166 struct blist
*swapblist
;
167 static int swap_async_max
= 4; /* maximum in-progress async I/O's */
168 static int swap_burst_read
= 0; /* allow burst reading */
170 extern struct vnode
*swapdev_vp
; /* from vm_swap.c */
172 SYSCTL_INT(_vm
, OID_AUTO
, swap_async_max
,
173 CTLFLAG_RW
, &swap_async_max
, 0, "Maximum running async swap ops");
174 SYSCTL_INT(_vm
, OID_AUTO
, swap_burst_read
,
175 CTLFLAG_RW
, &swap_burst_read
, 0, "Allow burst reads for pageins");
177 SYSCTL_INT(_vm
, OID_AUTO
, swap_cache_use
,
178 CTLFLAG_RD
, &vm_swap_cache_use
, 0, "");
179 SYSCTL_INT(_vm
, OID_AUTO
, swap_anon_use
,
180 CTLFLAG_RD
, &vm_swap_anon_use
, 0, "");
185 * Red-Black tree for swblock entries
187 RB_GENERATE2(swblock_rb_tree
, swblock
, swb_entry
, rb_swblock_compare
,
188 vm_pindex_t
, swb_index
);
191 rb_swblock_compare(struct swblock
*swb1
, struct swblock
*swb2
)
193 if (swb1
->swb_index
< swb2
->swb_index
)
195 if (swb1
->swb_index
> swb2
->swb_index
)
202 rb_swblock_scancmp(struct swblock
*swb
, void *data
)
204 struct swfreeinfo
*info
= data
;
206 if (swb
->swb_index
< info
->basei
)
208 if (swb
->swb_index
> info
->endi
)
215 rb_swblock_condcmp(struct swblock
*swb
, void *data
)
217 struct swfreeinfo
*info
= data
;
219 if (swb
->swb_index
< info
->basei
)
225 * pagerops for OBJT_SWAP - "swap pager". Some ops are also global procedure
226 * calls hooked from other parts of the VM system and do not appear here.
227 * (see vm/swap_pager.h).
231 swap_pager_alloc (void *handle
, off_t size
,
232 vm_prot_t prot
, off_t offset
);
233 static void swap_pager_dealloc (vm_object_t object
);
234 static int swap_pager_getpage (vm_object_t
, vm_page_t
*, int);
235 static void swap_chain_iodone(struct bio
*biox
);
237 struct pagerops swappagerops
= {
238 swap_pager_alloc
, /* allocate an OBJT_SWAP object */
239 swap_pager_dealloc
, /* deallocate an OBJT_SWAP object */
240 swap_pager_getpage
, /* pagein */
241 swap_pager_putpages
, /* pageout */
242 swap_pager_haspage
/* get backing store status for page */
246 * dmmax is in page-sized chunks with the new swap system. It was
247 * dev-bsized chunks in the old. dmmax is always a power of 2.
249 * swap_*() routines are externally accessible. swp_*() routines are
254 static int dmmax_mask
;
255 int nswap_lowat
= 128; /* in pages, swap_pager_almost_full warn */
256 int nswap_hiwat
= 512; /* in pages, swap_pager_almost_full warn */
258 static __inline
void swp_sizecheck (void);
259 static void swp_pager_async_iodone (struct bio
*bio
);
262 * Swap bitmap functions
265 static __inline
void swp_pager_freeswapspace (vm_object_t object
, daddr_t blk
, int npages
);
266 static __inline daddr_t
swp_pager_getswapspace (vm_object_t object
, int npages
);
272 static void swp_pager_meta_convert (vm_object_t
);
273 static void swp_pager_meta_build (vm_object_t
, vm_pindex_t
, daddr_t
);
274 static void swp_pager_meta_free (vm_object_t
, vm_pindex_t
, vm_pindex_t
);
275 static void swp_pager_meta_free_all (vm_object_t
);
276 static daddr_t
swp_pager_meta_ctl (vm_object_t
, vm_pindex_t
, int);
279 * SWP_SIZECHECK() - update swap_pager_full indication
281 * update the swap_pager_almost_full indication and warn when we are
282 * about to run out of swap space, using lowat/hiwat hysteresis.
284 * Clear swap_pager_full ( task killing ) indication when lowat is met.
286 * No restrictions on call
287 * This routine may not block.
288 * This routine must be called at splvm()
294 if (vm_swap_size
< nswap_lowat
) {
295 if (swap_pager_almost_full
== 0) {
296 kprintf("swap_pager: out of swap space\n");
297 swap_pager_almost_full
= 1;
301 if (vm_swap_size
> nswap_hiwat
)
302 swap_pager_almost_full
= 0;
307 * SWAP_PAGER_INIT() - initialize the swap pager!
309 * Expected to be started from system init. NOTE: This code is run
310 * before much else so be careful what you depend on. Most of the VM
311 * system has yet to be initialized at this point.
314 swap_pager_init(void *arg __unused
)
317 * Device Stripe, in PAGE_SIZE'd blocks
319 dmmax
= SWB_NPAGES
* 2;
320 dmmax_mask
= ~(dmmax
- 1);
322 SYSINIT(vm_mem
, SI_BOOT1_VM
, SI_ORDER_THIRD
, swap_pager_init
, NULL
)
325 * SWAP_PAGER_SWAP_INIT() - swap pager initialization from pageout process
327 * Expected to be started from pageout process once, prior to entering
332 swap_pager_swap_init(void)
337 * Number of in-transit swap bp operations. Don't
338 * exhaust the pbufs completely. Make sure we
339 * initialize workable values (0 will work for hysteresis
340 * but it isn't very efficient).
342 * The nsw_cluster_max is constrained by the number of pages an XIO
343 * holds, i.e., (MAXPHYS/PAGE_SIZE) and our locally defined
344 * MAX_PAGEOUT_CLUSTER. Also be aware that swap ops are
345 * constrained by the swap device interleave stripe size.
347 * Currently we hardwire nsw_wcount_async to 4. This limit is
348 * designed to prevent other I/O from having high latencies due to
349 * our pageout I/O. The value 4 works well for one or two active swap
350 * devices but is probably a little low if you have more. Even so,
351 * a higher value would probably generate only a limited improvement
352 * with three or four active swap devices since the system does not
353 * typically have to pageout at extreme bandwidths. We will want
354 * at least 2 per swap devices, and 4 is a pretty good value if you
355 * have one NFS swap device due to the command/ack latency over NFS.
356 * So it all works out pretty well.
359 nsw_cluster_max
= min((MAXPHYS
/PAGE_SIZE
), MAX_PAGEOUT_CLUSTER
);
361 nsw_rcount
= (nswbuf
+ 1) / 2;
362 nsw_wcount_sync
= (nswbuf
+ 3) / 4;
363 nsw_wcount_async
= 4;
364 nsw_wcount_async_max
= nsw_wcount_async
;
367 * The zone is dynamically allocated so generally size it to
368 * maxswzone (32MB to 512MB of KVM). Set a minimum size based
369 * on physical memory of around 8x (each swblock can hold 16 pages).
371 * With the advent of SSDs (vs HDs) the practical (swap:memory) ratio
372 * has increased dramatically.
374 n
= vmstats
.v_page_count
/ 2;
375 if (maxswzone
&& n
< maxswzone
/ sizeof(struct swblock
))
376 n
= maxswzone
/ sizeof(struct swblock
);
382 sizeof(struct swblock
),
386 if (swap_zone
!= NULL
)
389 * if the allocation failed, try a zone two thirds the
390 * size of the previous attempt.
395 if (swap_zone
== NULL
)
396 panic("swap_pager_swap_init: swap_zone == NULL");
398 kprintf("Swap zone entries reduced from %d to %d.\n", n2
, n
);
402 * SWAP_PAGER_ALLOC() - allocate a new OBJT_SWAP VM object and instantiate
403 * its metadata structures.
405 * This routine is called from the mmap and fork code to create a new
406 * OBJT_SWAP object. We do this by creating an OBJT_DEFAULT object
407 * and then converting it with swp_pager_meta_convert().
409 * This routine may block in vm_object_allocate() and create a named
410 * object lookup race, so we must interlock. We must also run at
411 * splvm() for the object lookup to handle races with interrupts, but
412 * we do not have to maintain splvm() in between the lookup and the
413 * add because (I believe) it is not possible to attempt to create
414 * a new swap object w/handle when a default object with that handle
419 swap_pager_alloc(void *handle
, off_t size
, vm_prot_t prot
, off_t offset
)
423 KKASSERT(handle
== NULL
);
427 * Reference existing named region or allocate new one. There
428 * should not be a race here against swp_pager_meta_build()
429 * as called from vm_page_remove() in regards to the lookup
432 while (sw_alloc_interlock
) {
433 sw_alloc_interlock
= -1;
434 tsleep(&sw_alloc_interlock
, 0, "swpalc", 0);
436 sw_alloc_interlock
= 1;
438 object
= vm_pager_object_lookup(NOBJLIST(handle
), handle
);
440 if (object
!= NULL
) {
441 vm_object_reference(object
);
443 object
= vm_object_allocate(OBJT_DEFAULT
,
444 OFF_TO_IDX(offset
+ PAGE_MASK
+ size
));
445 object
->handle
= handle
;
446 swp_pager_meta_convert(object
);
449 if (sw_alloc_interlock
< 0)
450 wakeup(&sw_alloc_interlock
);
451 sw_alloc_interlock
= 0;
454 object
= vm_object_allocate(OBJT_DEFAULT
,
455 OFF_TO_IDX(offset
+ PAGE_MASK
+ size
));
456 swp_pager_meta_convert(object
);
462 * SWAP_PAGER_DEALLOC() - remove swap metadata from object
464 * The swap backing for the object is destroyed. The code is
465 * designed such that we can reinstantiate it later, but this
466 * routine is typically called only when the entire object is
467 * about to be destroyed.
469 * This routine may block, but no longer does.
471 * The object must be locked or unreferenceable.
475 swap_pager_dealloc(vm_object_t object
)
477 vm_object_pip_wait(object
, "swpdea");
480 * Free all remaining metadata. We only bother to free it from
481 * the swap meta data. We do not attempt to free swapblk's still
482 * associated with vm_page_t's for this object. We do not care
483 * if paging is still in progress on some objects.
486 swp_pager_meta_free_all(object
);
490 /************************************************************************
491 * SWAP PAGER BITMAP ROUTINES *
492 ************************************************************************/
495 * SWP_PAGER_GETSWAPSPACE() - allocate raw swap space
497 * Allocate swap for the requested number of pages. The starting
498 * swap block number (a page index) is returned or SWAPBLK_NONE
499 * if the allocation failed.
501 * Also has the side effect of advising that somebody made a mistake
502 * when they configured swap and didn't configure enough.
504 * Must be called at splvm() to avoid races with bitmap frees from
505 * vm_page_remove() aka swap_pager_page_removed().
507 * This routine may not block
508 * This routine must be called at splvm().
510 static __inline daddr_t
511 swp_pager_getswapspace(vm_object_t object
, int npages
)
515 if ((blk
= blist_alloc(swapblist
, npages
)) == SWAPBLK_NONE
) {
516 if (swap_pager_full
!= 2) {
517 kprintf("swap_pager_getswapspace: failed\n");
519 swap_pager_almost_full
= 1;
522 vm_swap_size
-= npages
;
523 if (object
->type
== OBJT_SWAP
)
524 vm_swap_anon_use
+= npages
;
526 vm_swap_cache_use
+= npages
;
533 * SWP_PAGER_FREESWAPSPACE() - free raw swap space
535 * This routine returns the specified swap blocks back to the bitmap.
537 * Note: This routine may not block (it could in the old swap code),
538 * and through the use of the new blist routines it does not block.
540 * We must be called at splvm() to avoid races with bitmap frees from
541 * vm_page_remove() aka swap_pager_page_removed().
543 * This routine may not block
544 * This routine must be called at splvm().
548 swp_pager_freeswapspace(vm_object_t object
, daddr_t blk
, int npages
)
550 blist_free(swapblist
, blk
, npages
);
551 vm_swap_size
+= npages
;
552 if (object
->type
== OBJT_SWAP
)
553 vm_swap_anon_use
-= npages
;
555 vm_swap_cache_use
-= npages
;
560 * SWAP_PAGER_FREESPACE() - frees swap blocks associated with a page
561 * range within an object.
563 * This is a globally accessible routine.
565 * This routine removes swapblk assignments from swap metadata.
567 * The external callers of this routine typically have already destroyed
568 * or renamed vm_page_t's associated with this range in the object so
571 * This routine may be called at any spl. We up our spl to splvm
572 * temporarily in order to perform the metadata removal.
575 swap_pager_freespace(vm_object_t object
, vm_pindex_t start
, vm_pindex_t size
)
578 swp_pager_meta_free(object
, start
, size
);
583 swap_pager_freespace_all(vm_object_t object
)
586 swp_pager_meta_free_all(object
);
591 * This function conditionally frees swap cache swap starting at
592 * (*basei) in the object. (count) swap blocks will be nominally freed.
593 * The actual number of blocks freed can be more or less than the
596 * This function nominally returns the number of blocks freed. However,
597 * the actual number of blocks freed may be less then the returned value.
598 * If the function is unable to exhaust the object or if it is able to
599 * free (approximately) the requested number of blocks it returns
602 * If we exhaust the object we will return a value n <= count.
604 static int swap_pager_condfree_callback(struct swblock
*swap
, void *data
);
607 swap_pager_condfree(vm_object_t object
, vm_pindex_t
*basei
, int count
)
609 struct swfreeinfo info
;
611 info
.object
= object
;
612 info
.basei
= *basei
; /* skip up to this page index */
613 info
.begi
= count
; /* max swap pages to destroy */
614 info
.endi
= count
* 8; /* max swblocks to scan */
616 swblock_rb_tree_RB_SCAN(&object
->swblock_root
, rb_swblock_condcmp
,
617 swap_pager_condfree_callback
, &info
);
619 if (info
.endi
< 0 && info
.begi
<= count
)
620 info
.begi
= count
+ 1;
621 return(count
- (int)info
.begi
);
625 * The idea is to free whole meta-block to avoid fragmenting
626 * the swap space or disk I/O. We only do this if NO VM pages
629 * We do not have to deal with clearing PG_SWAPPED in related VM
630 * pages because there are no related VM pages.
633 swap_pager_condfree_callback(struct swblock
*swap
, void *data
)
635 struct swfreeinfo
*info
= data
;
636 vm_object_t object
= info
->object
;
639 for (i
= 0; i
< SWAP_META_PAGES
; ++i
) {
640 if (vm_page_lookup(object
, swap
->swb_index
+ i
))
643 info
->basei
= swap
->swb_index
+ SWAP_META_PAGES
;
644 if (i
== SWAP_META_PAGES
) {
645 info
->begi
-= swap
->swb_count
;
646 swap_pager_freespace(object
, swap
->swb_index
, SWAP_META_PAGES
);
649 if ((int)info
->begi
< 0 || (int)info
->endi
< 0)
655 * Called by vm_page_alloc() when a new VM page is inserted
656 * into a VM object. Checks whether swap has been assigned to
657 * the page and sets PG_SWAPPED as necessary.
660 swap_pager_page_inserted(vm_page_t m
)
662 if (m
->object
->swblock_count
) {
664 if (swp_pager_meta_ctl(m
->object
, m
->pindex
, 0) != SWAPBLK_NONE
)
665 vm_page_flag_set(m
, PG_SWAPPED
);
671 * SWAP_PAGER_RESERVE() - reserve swap blocks in object
673 * Assigns swap blocks to the specified range within the object. The
674 * swap blocks are not zerod. Any previous swap assignment is destroyed.
676 * Returns 0 on success, -1 on failure.
679 swap_pager_reserve(vm_object_t object
, vm_pindex_t start
, vm_size_t size
)
682 daddr_t blk
= SWAPBLK_NONE
;
683 vm_pindex_t beg
= start
; /* save start index */
689 while ((blk
= swp_pager_getswapspace(object
, n
)) ==
694 swp_pager_meta_free(object
, beg
,
701 swp_pager_meta_build(object
, start
, blk
);
707 swp_pager_meta_free(object
, start
, n
);
713 * SWAP_PAGER_COPY() - copy blocks from source pager to destination pager
714 * and destroy the source.
716 * Copy any valid swapblks from the source to the destination. In
717 * cases where both the source and destination have a valid swapblk,
718 * we keep the destination's.
720 * This routine is allowed to block. It may block allocating metadata
721 * indirectly through swp_pager_meta_build() or if paging is still in
722 * progress on the source.
724 * This routine can be called at any spl
726 * XXX vm_page_collapse() kinda expects us not to block because we
727 * supposedly do not need to allocate memory, but for the moment we
728 * *may* have to get a little memory from the zone allocator, but
729 * it is taken from the interrupt memory. We should be ok.
731 * The source object contains no vm_page_t's (which is just as well)
733 * The source object is of type OBJT_SWAP.
735 * The source and destination objects must be locked or
736 * inaccessible (XXX are they ?)
740 swap_pager_copy(vm_object_t srcobject
, vm_object_t dstobject
,
741 vm_pindex_t base_index
, int destroysource
)
748 * transfer source to destination.
750 for (i
= 0; i
< dstobject
->size
; ++i
) {
754 * Locate (without changing) the swapblk on the destination,
755 * unless it is invalid in which case free it silently, or
756 * if the destination is a resident page, in which case the
757 * source is thrown away.
759 dstaddr
= swp_pager_meta_ctl(dstobject
, i
, 0);
761 if (dstaddr
== SWAPBLK_NONE
) {
763 * Destination has no swapblk and is not resident,
768 srcaddr
= swp_pager_meta_ctl(srcobject
,
769 base_index
+ i
, SWM_POP
);
771 if (srcaddr
!= SWAPBLK_NONE
)
772 swp_pager_meta_build(dstobject
, i
, srcaddr
);
775 * Destination has valid swapblk or it is represented
776 * by a resident page. We destroy the sourceblock.
778 swp_pager_meta_ctl(srcobject
, base_index
+ i
, SWM_FREE
);
783 * Free left over swap blocks in source.
785 * We have to revert the type to OBJT_DEFAULT so we do not accidently
786 * double-remove the object from the swap queues.
790 * Reverting the type is not necessary, the caller is going
791 * to destroy srcobject directly, but I'm doing it here
792 * for consistency since we've removed the object from its
795 swp_pager_meta_free_all(srcobject
);
796 if (srcobject
->type
== OBJT_SWAP
)
797 srcobject
->type
= OBJT_DEFAULT
;
803 * SWAP_PAGER_HASPAGE() - determine if we have good backing store for
804 * the requested page.
806 * We determine whether good backing store exists for the requested
807 * page and return TRUE if it does, FALSE if it doesn't.
809 * If TRUE, we also try to determine how much valid, contiguous backing
810 * store exists before and after the requested page within a reasonable
811 * distance. We do not try to restrict it to the swap device stripe
812 * (that is handled in getpages/putpages). It probably isn't worth
817 swap_pager_haspage(vm_object_t object
, vm_pindex_t pindex
)
822 * do we have good backing store at the requested index ?
826 blk0
= swp_pager_meta_ctl(object
, pindex
, 0);
828 if (blk0
== SWAPBLK_NONE
) {
835 * find backwards-looking contiguous good backing store
837 if (before
!= NULL
) {
840 for (i
= 1; i
< (SWB_NPAGES
/2); ++i
) {
845 blk
= swp_pager_meta_ctl(object
, pindex
- i
, 0);
853 * find forward-looking contiguous good backing store
859 for (i
= 1; i
< (SWB_NPAGES
/2); ++i
) {
862 blk
= swp_pager_meta_ctl(object
, pindex
+ i
, 0);
874 * SWAP_PAGER_PAGE_UNSWAPPED() - remove swap backing store related to page
876 * This removes any associated swap backing store, whether valid or
877 * not, from the page. This operates on any VM object, not just OBJT_SWAP
880 * This routine is typically called when a page is made dirty, at
881 * which point any associated swap can be freed. MADV_FREE also
882 * calls us in a special-case situation
884 * NOTE!!! If the page is clean and the swap was valid, the caller
885 * should make the page dirty before calling this routine. This routine
886 * does NOT change the m->dirty status of the page. Also: MADV_FREE
889 * This routine may not block
890 * This routine must be called at splvm()
893 swap_pager_unswapped(vm_page_t m
)
895 if (m
->flags
& PG_SWAPPED
) {
896 swp_pager_meta_ctl(m
->object
, m
->pindex
, SWM_FREE
);
897 vm_page_flag_clear(m
, PG_SWAPPED
);
902 * SWAP_PAGER_STRATEGY() - read, write, free blocks
904 * This implements a VM OBJECT strategy function using swap backing store.
905 * This can operate on any VM OBJECT type, not necessarily just OBJT_SWAP
908 * This is intended to be a cacheless interface (i.e. caching occurs at
909 * higher levels), and is also used as a swap-based SSD cache for vnode
910 * and device objects.
912 * All I/O goes directly to and from the swap device.
914 * We currently attempt to run I/O synchronously or asynchronously as
915 * the caller requests. This isn't perfect because we loose error
916 * sequencing when we run multiple ops in parallel to satisfy a request.
917 * But this is swap, so we let it all hang out.
920 swap_pager_strategy(vm_object_t object
, struct bio
*bio
)
922 struct buf
*bp
= bio
->bio_buf
;
925 vm_pindex_t biox_blkno
= 0;
930 struct bio_track
*track
;
933 * tracking for swapdev vnode I/Os
935 if (bp
->b_cmd
== BUF_CMD_READ
)
936 track
= &swapdev_vp
->v_track_read
;
938 track
= &swapdev_vp
->v_track_write
;
940 if (bp
->b_bcount
& PAGE_MASK
) {
941 bp
->b_error
= EINVAL
;
942 bp
->b_flags
|= B_ERROR
| B_INVAL
;
944 kprintf("swap_pager_strategy: bp %p offset %lld size %d, "
945 "not page bounded\n",
946 bp
, (long long)bio
->bio_offset
, (int)bp
->b_bcount
);
951 * Clear error indication, initialize page index, count, data pointer.
954 bp
->b_flags
&= ~B_ERROR
;
955 bp
->b_resid
= bp
->b_bcount
;
957 start
= (vm_pindex_t
)(bio
->bio_offset
>> PAGE_SHIFT
);
958 count
= howmany(bp
->b_bcount
, PAGE_SIZE
);
962 * Deal with BUF_CMD_FREEBLKS
964 if (bp
->b_cmd
== BUF_CMD_FREEBLKS
) {
966 * FREE PAGE(s) - destroy underlying swap that is no longer
969 swp_pager_meta_free(object
, start
, count
);
976 * We need to be able to create a new cluster of I/O's. We cannot
977 * use the caller fields of the passed bio so push a new one.
979 * Because nbio is just a placeholder for the cluster links,
980 * we can biodone() the original bio instead of nbio to make
981 * things a bit more efficient.
983 nbio
= push_bio(bio
);
984 nbio
->bio_offset
= bio
->bio_offset
;
985 nbio
->bio_caller_info1
.cluster_head
= NULL
;
986 nbio
->bio_caller_info2
.cluster_tail
= NULL
;
992 * Execute read or write
998 * Obtain block. If block not found and writing, allocate a
999 * new block and build it into the object.
1001 blk
= swp_pager_meta_ctl(object
, start
, 0);
1002 if ((blk
== SWAPBLK_NONE
) && bp
->b_cmd
!= BUF_CMD_READ
) {
1003 blk
= swp_pager_getswapspace(object
, 1);
1004 if (blk
== SWAPBLK_NONE
) {
1005 bp
->b_error
= ENOMEM
;
1006 bp
->b_flags
|= B_ERROR
;
1009 swp_pager_meta_build(object
, start
, blk
);
1013 * Do we have to flush our current collection? Yes if:
1015 * - no swap block at this index
1016 * - swap block is not contiguous
1017 * - we cross a physical disk boundry in the
1021 biox
&& (biox_blkno
+ btoc(bufx
->b_bcount
) != blk
||
1022 ((biox_blkno
^ blk
) & dmmax_mask
)
1025 if (bp
->b_cmd
== BUF_CMD_READ
) {
1026 ++mycpu
->gd_cnt
.v_swapin
;
1027 mycpu
->gd_cnt
.v_swappgsin
+= btoc(bufx
->b_bcount
);
1029 ++mycpu
->gd_cnt
.v_swapout
;
1030 mycpu
->gd_cnt
.v_swappgsout
+= btoc(bufx
->b_bcount
);
1031 bufx
->b_dirtyend
= bufx
->b_bcount
;
1035 * Finished with this buf.
1037 KKASSERT(bufx
->b_bcount
!= 0);
1038 if (bufx
->b_cmd
!= BUF_CMD_READ
)
1039 bufx
->b_dirtyend
= bufx
->b_bcount
;
1045 * Add new swapblk to biox, instantiating biox if necessary.
1046 * Zero-fill reads are able to take a shortcut.
1048 if (blk
== SWAPBLK_NONE
) {
1050 * We can only get here if we are reading. Since
1051 * we are at splvm() we can safely modify b_resid,
1052 * even if chain ops are in progress.
1054 bzero(data
, PAGE_SIZE
);
1055 bp
->b_resid
-= PAGE_SIZE
;
1058 /* XXX chain count > 4, wait to <= 4 */
1060 bufx
= getpbuf(NULL
);
1061 biox
= &bufx
->b_bio1
;
1062 cluster_append(nbio
, bufx
);
1063 bufx
->b_flags
|= (bufx
->b_flags
& B_ORDERED
);
1064 bufx
->b_cmd
= bp
->b_cmd
;
1065 biox
->bio_done
= swap_chain_iodone
;
1066 biox
->bio_offset
= (off_t
)blk
<< PAGE_SHIFT
;
1067 biox
->bio_caller_info1
.cluster_parent
= nbio
;
1070 bufx
->b_data
= data
;
1072 bufx
->b_bcount
+= PAGE_SIZE
;
1080 * Flush out last buffer
1083 if (bufx
->b_cmd
== BUF_CMD_READ
) {
1084 ++mycpu
->gd_cnt
.v_swapin
;
1085 mycpu
->gd_cnt
.v_swappgsin
+= btoc(bufx
->b_bcount
);
1087 ++mycpu
->gd_cnt
.v_swapout
;
1088 mycpu
->gd_cnt
.v_swappgsout
+= btoc(bufx
->b_bcount
);
1089 bufx
->b_dirtyend
= bufx
->b_bcount
;
1091 KKASSERT(bufx
->b_bcount
);
1092 if (bufx
->b_cmd
!= BUF_CMD_READ
)
1093 bufx
->b_dirtyend
= bufx
->b_bcount
;
1094 /* biox, bufx = NULL */
1098 * Now initiate all the I/O. Be careful looping on our chain as
1099 * I/O's may complete while we are still initiating them.
1101 nbio
->bio_caller_info2
.cluster_tail
= NULL
;
1102 bufx
= nbio
->bio_caller_info1
.cluster_head
;
1105 biox
= &bufx
->b_bio1
;
1107 bufx
= bufx
->b_cluster_next
;
1108 vn_strategy(swapdev_vp
, biox
);
1112 * Completion of the cluster will also call biodone_chain(nbio).
1113 * We never call biodone(nbio) so we don't have to worry about
1114 * setting up a bio_done callback. It's handled in the sub-IO.
1120 swap_chain_iodone(struct bio
*biox
)
1123 struct buf
*bufx
; /* chained sub-buffer */
1124 struct bio
*nbio
; /* parent nbio with chain glue */
1125 struct buf
*bp
; /* original bp associated with nbio */
1128 bufx
= biox
->bio_buf
;
1129 nbio
= biox
->bio_caller_info1
.cluster_parent
;
1133 * Update the original buffer
1135 KKASSERT(bp
!= NULL
);
1136 if (bufx
->b_flags
& B_ERROR
) {
1137 atomic_set_int(&bufx
->b_flags
, B_ERROR
);
1138 bp
->b_error
= bufx
->b_error
;
1139 } else if (bufx
->b_resid
!= 0) {
1140 atomic_set_int(&bufx
->b_flags
, B_ERROR
);
1141 bp
->b_error
= EINVAL
;
1143 atomic_subtract_int(&bp
->b_resid
, bufx
->b_bcount
);
1147 * Remove us from the chain.
1149 spin_lock_wr(&bp
->b_lock
.lk_spinlock
);
1150 nextp
= &nbio
->bio_caller_info1
.cluster_head
;
1151 while (*nextp
!= bufx
) {
1152 KKASSERT(*nextp
!= NULL
);
1153 nextp
= &(*nextp
)->b_cluster_next
;
1155 *nextp
= bufx
->b_cluster_next
;
1156 chain_empty
= (nbio
->bio_caller_info1
.cluster_head
== NULL
);
1157 spin_unlock_wr(&bp
->b_lock
.lk_spinlock
);
1160 * Clean up bufx. If the chain is now empty we finish out
1161 * the parent. Note that we may be racing other completions
1162 * so we must use the chain_empty status from above.
1165 if (bp
->b_resid
!= 0 && !(bp
->b_flags
& B_ERROR
)) {
1166 atomic_set_int(&bp
->b_flags
, B_ERROR
);
1167 bp
->b_error
= EINVAL
;
1169 biodone_chain(nbio
);
1171 relpbuf(bufx
, NULL
);
1175 * SWAP_PAGER_GETPAGES() - bring page in from swap
1177 * The requested page may have to be brought in from swap. Calculate the
1178 * swap block and bring in additional pages if possible. All pages must
1179 * have contiguous swap block assignments and reside in the same object.
1181 * The caller has a single vm_object_pip_add() reference prior to
1182 * calling us and we should return with the same.
1184 * The caller has BUSY'd the page. We should return with (*mpp) left busy,
1185 * and any additinal pages unbusied.
1187 * If the caller encounters a PG_RAM page it will pass it to us even though
1188 * it may be valid and dirty. We cannot overwrite the page in this case!
1189 * The case is used to allow us to issue pure read-aheads.
1191 * NOTE! XXX This code does not entirely pipeline yet due to the fact that
1192 * the PG_RAM page is validated at the same time as mreq. What we
1193 * really need to do is issue a separate read-ahead pbuf.
1196 swap_pager_getpage(vm_object_t object
, vm_page_t
*mpp
, int seqaccess
)
1207 vm_page_t marray
[XIO_INTERNAL_PAGES
];
1211 if (mreq
->object
!= object
) {
1212 panic("swap_pager_getpages: object mismatch %p/%p",
1219 * We don't want to overwrite a fully valid page as it might be
1220 * dirty. This case can occur when e.g. vm_fault hits a perfectly
1221 * valid page with PG_RAM set.
1223 * In this case we see if the next page is a suitable page-in
1224 * candidate and if it is we issue read-ahead. PG_RAM will be
1225 * set on the last page of the read-ahead to continue the pipeline.
1227 if (mreq
->valid
== VM_PAGE_BITS_ALL
) {
1228 if (swap_burst_read
== 0 || mreq
->pindex
+ 1 >= object
->size
)
1229 return(VM_PAGER_OK
);
1231 blk
= swp_pager_meta_ctl(object
, mreq
->pindex
+ 1, 0);
1232 if (blk
== SWAPBLK_NONE
) {
1234 return(VM_PAGER_OK
);
1236 m
= vm_page_lookup(object
, mreq
->pindex
+ 1);
1238 m
= vm_page_alloc(object
, mreq
->pindex
+ 1,
1242 return(VM_PAGER_OK
);
1245 if ((m
->flags
& PG_BUSY
) || m
->busy
|| m
->valid
) {
1247 return(VM_PAGER_OK
);
1249 vm_page_unqueue_nowakeup(m
);
1260 * Try to block-read contiguous pages from swap if sequential,
1261 * otherwise just read one page. Contiguous pages from swap must
1262 * reside within a single device stripe because the I/O cannot be
1263 * broken up across multiple stripes.
1265 * Note that blk and iblk can be SWAPBLK_NONE but the loop is
1266 * set up such that the case(s) are handled implicitly.
1269 blk
= swp_pager_meta_ctl(mreq
->object
, mreq
->pindex
, 0);
1272 for (i
= 1; swap_burst_read
&&
1273 i
< XIO_INTERNAL_PAGES
&&
1274 mreq
->pindex
+ i
< object
->size
; ++i
) {
1277 iblk
= swp_pager_meta_ctl(object
, mreq
->pindex
+ i
, 0);
1278 if (iblk
!= blk
+ i
)
1280 if ((blk
^ iblk
) & dmmax_mask
)
1282 m
= vm_page_lookup(object
, mreq
->pindex
+ i
);
1284 m
= vm_page_alloc(object
, mreq
->pindex
+ i
,
1289 if ((m
->flags
& PG_BUSY
) || m
->busy
|| m
->valid
)
1291 vm_page_unqueue_nowakeup(m
);
1297 vm_page_flag_set(marray
[i
- 1], PG_RAM
);
1302 * If mreq is the requested page and we have nothing to do return
1303 * VM_PAGER_FAIL. If raonly is set mreq is just another read-ahead
1304 * page and must be cleaned up.
1306 if (blk
== SWAPBLK_NONE
) {
1309 vnode_pager_freepage(mreq
);
1310 return(VM_PAGER_OK
);
1312 return(VM_PAGER_FAIL
);
1317 * map our page(s) into kva for input
1319 bp
= getpbuf(&nsw_rcount
);
1321 kva
= (vm_offset_t
) bp
->b_kvabase
;
1322 bcopy(marray
, bp
->b_xio
.xio_pages
, i
* sizeof(vm_page_t
));
1323 pmap_qenter(kva
, bp
->b_xio
.xio_pages
, i
);
1325 bp
->b_data
= (caddr_t
)kva
;
1326 bp
->b_bcount
= PAGE_SIZE
* i
;
1327 bp
->b_xio
.xio_npages
= i
;
1328 bio
->bio_done
= swp_pager_async_iodone
;
1329 bio
->bio_offset
= (off_t
)blk
<< PAGE_SHIFT
;
1330 bio
->bio_caller_info1
.index
= SWBIO_READ
;
1333 * Set index. If raonly set the index beyond the array so all
1334 * the pages are treated the same, otherwise the original mreq is
1338 bio
->bio_driver_info
= (void *)(intptr_t)i
;
1340 bio
->bio_driver_info
= (void *)(intptr_t)0;
1342 for (j
= 0; j
< i
; ++j
)
1343 vm_page_flag_set(bp
->b_xio
.xio_pages
[j
], PG_SWAPINPROG
);
1345 mycpu
->gd_cnt
.v_swapin
++;
1346 mycpu
->gd_cnt
.v_swappgsin
+= bp
->b_xio
.xio_npages
;
1349 * We still hold the lock on mreq, and our automatic completion routine
1350 * does not remove it.
1352 vm_object_pip_add(object
, bp
->b_xio
.xio_npages
);
1355 * perform the I/O. NOTE!!! bp cannot be considered valid after
1356 * this point because we automatically release it on completion.
1357 * Instead, we look at the one page we are interested in which we
1358 * still hold a lock on even through the I/O completion.
1360 * The other pages in our m[] array are also released on completion,
1361 * so we cannot assume they are valid anymore either.
1363 bp
->b_cmd
= BUF_CMD_READ
;
1365 vn_strategy(swapdev_vp
, bio
);
1368 * Wait for the page we want to complete. PG_SWAPINPROG is always
1369 * cleared on completion. If an I/O error occurs, SWAPBLK_NONE
1370 * is set in the meta-data.
1372 * If this is a read-ahead only we return immediately without
1376 return(VM_PAGER_OK
);
1379 * Read-ahead includes originally requested page case.
1382 while ((mreq
->flags
& PG_SWAPINPROG
) != 0) {
1383 vm_page_flag_set(mreq
, PG_WANTED
| PG_REFERENCED
);
1384 mycpu
->gd_cnt
.v_intrans
++;
1385 if (tsleep(mreq
, 0, "swread", hz
*20)) {
1387 "swap_pager: indefinite wait buffer: "
1388 " offset: %lld, size: %ld\n",
1389 (long long)bio
->bio_offset
,
1397 * mreq is left bussied after completion, but all the other pages
1398 * are freed. If we had an unrecoverable read error the page will
1401 if (mreq
->valid
!= VM_PAGE_BITS_ALL
)
1402 return(VM_PAGER_ERROR
);
1404 return(VM_PAGER_OK
);
1407 * A final note: in a low swap situation, we cannot deallocate swap
1408 * and mark a page dirty here because the caller is likely to mark
1409 * the page clean when we return, causing the page to possibly revert
1410 * to all-zero's later.
1415 * swap_pager_putpages:
1417 * Assign swap (if necessary) and initiate I/O on the specified pages.
1419 * We support both OBJT_DEFAULT and OBJT_SWAP objects. DEFAULT objects
1420 * are automatically converted to SWAP objects.
1422 * In a low memory situation we may block in vn_strategy(), but the new
1423 * vm_page reservation system coupled with properly written VFS devices
1424 * should ensure that no low-memory deadlock occurs. This is an area
1427 * The parent has N vm_object_pip_add() references prior to
1428 * calling us and will remove references for rtvals[] that are
1429 * not set to VM_PAGER_PEND. We need to remove the rest on I/O
1432 * The parent has soft-busy'd the pages it passes us and will unbusy
1433 * those whos rtvals[] entry is not set to VM_PAGER_PEND on return.
1434 * We need to unbusy the rest on I/O completion.
1437 swap_pager_putpages(vm_object_t object
, vm_page_t
*m
, int count
,
1438 boolean_t sync
, int *rtvals
)
1443 if (count
&& m
[0]->object
!= object
) {
1444 panic("swap_pager_getpages: object mismatch %p/%p",
1453 * Turn object into OBJT_SWAP
1454 * check for bogus sysops
1455 * force sync if not pageout process
1457 if (object
->type
== OBJT_DEFAULT
)
1458 swp_pager_meta_convert(object
);
1460 if (curthread
!= pagethread
)
1466 * Update nsw parameters from swap_async_max sysctl values.
1467 * Do not let the sysop crash the machine with bogus numbers.
1470 if (swap_async_max
!= nsw_wcount_async_max
) {
1476 if ((n
= swap_async_max
) > nswbuf
/ 2)
1483 * Adjust difference ( if possible ). If the current async
1484 * count is too low, we may not be able to make the adjustment
1488 n
-= nsw_wcount_async_max
;
1489 if (nsw_wcount_async
+ n
>= 0) {
1490 nsw_wcount_async
+= n
;
1491 nsw_wcount_async_max
+= n
;
1492 wakeup(&nsw_wcount_async
);
1500 * Assign swap blocks and issue I/O. We reallocate swap on the fly.
1501 * The page is left dirty until the pageout operation completes
1505 for (i
= 0; i
< count
; i
+= n
) {
1512 * Maximum I/O size is limited by a number of factors.
1515 n
= min(BLIST_MAX_ALLOC
, count
- i
);
1516 n
= min(n
, nsw_cluster_max
);
1521 * Get biggest block of swap we can. If we fail, fall
1522 * back and try to allocate a smaller block. Don't go
1523 * overboard trying to allocate space if it would overly
1527 (blk
= swp_pager_getswapspace(object
, n
)) == SWAPBLK_NONE
&&
1532 if (blk
== SWAPBLK_NONE
) {
1533 for (j
= 0; j
< n
; ++j
)
1534 rtvals
[i
+j
] = VM_PAGER_FAIL
;
1540 * The I/O we are constructing cannot cross a physical
1541 * disk boundry in the swap stripe. Note: we are still
1544 if ((blk
^ (blk
+ n
)) & dmmax_mask
) {
1545 j
= ((blk
+ dmmax
) & dmmax_mask
) - blk
;
1546 swp_pager_freeswapspace(object
, blk
+ j
, n
- j
);
1551 * All I/O parameters have been satisfied, build the I/O
1552 * request and assign the swap space.
1555 bp
= getpbuf(&nsw_wcount_sync
);
1557 bp
= getpbuf(&nsw_wcount_async
);
1560 pmap_qenter((vm_offset_t
)bp
->b_data
, &m
[i
], n
);
1562 bp
->b_bcount
= PAGE_SIZE
* n
;
1563 bio
->bio_offset
= (off_t
)blk
<< PAGE_SHIFT
;
1565 for (j
= 0; j
< n
; ++j
) {
1566 vm_page_t mreq
= m
[i
+j
];
1568 swp_pager_meta_build(mreq
->object
, mreq
->pindex
,
1570 if (object
->type
== OBJT_SWAP
)
1571 vm_page_dirty(mreq
);
1572 rtvals
[i
+j
] = VM_PAGER_OK
;
1574 vm_page_flag_set(mreq
, PG_SWAPINPROG
);
1575 bp
->b_xio
.xio_pages
[j
] = mreq
;
1577 bp
->b_xio
.xio_npages
= n
;
1579 mycpu
->gd_cnt
.v_swapout
++;
1580 mycpu
->gd_cnt
.v_swappgsout
+= bp
->b_xio
.xio_npages
;
1584 bp
->b_dirtyoff
= 0; /* req'd for NFS */
1585 bp
->b_dirtyend
= bp
->b_bcount
; /* req'd for NFS */
1586 bp
->b_cmd
= BUF_CMD_WRITE
;
1587 bio
->bio_caller_info1
.index
= SWBIO_WRITE
;
1592 if (sync
== FALSE
) {
1593 bio
->bio_done
= swp_pager_async_iodone
;
1595 vn_strategy(swapdev_vp
, bio
);
1597 for (j
= 0; j
< n
; ++j
)
1598 rtvals
[i
+j
] = VM_PAGER_PEND
;
1603 * Issue synchrnously.
1605 * Wait for the sync I/O to complete, then update rtvals.
1606 * We just set the rtvals[] to VM_PAGER_PEND so we can call
1607 * our async completion routine at the end, thus avoiding a
1610 bio
->bio_caller_info1
.index
|= SWBIO_SYNC
;
1611 bio
->bio_done
= biodone_sync
;
1612 bio
->bio_flags
|= BIO_SYNC
;
1613 vn_strategy(swapdev_vp
, bio
);
1614 biowait(bio
, "swwrt");
1616 for (j
= 0; j
< n
; ++j
)
1617 rtvals
[i
+j
] = VM_PAGER_PEND
;
1620 * Now that we are through with the bp, we can call the
1621 * normal async completion, which frees everything up.
1623 swp_pager_async_iodone(bio
);
1628 swap_pager_newswap(void)
1634 * swp_pager_async_iodone:
1636 * Completion routine for asynchronous reads and writes from/to swap.
1637 * Also called manually by synchronous code to finish up a bp.
1639 * For READ operations, the pages are PG_BUSY'd. For WRITE operations,
1640 * the pages are vm_page_t->busy'd. For READ operations, we PG_BUSY
1641 * unbusy all pages except the 'main' request page. For WRITE
1642 * operations, we vm_page_t->busy'd unbusy all pages ( we can do this
1643 * because we marked them all VM_PAGER_PEND on return from putpages ).
1645 * This routine may not block.
1648 swp_pager_async_iodone(struct bio
*bio
)
1650 struct buf
*bp
= bio
->bio_buf
;
1651 vm_object_t object
= NULL
;
1658 if (bp
->b_flags
& B_ERROR
) {
1660 "swap_pager: I/O error - %s failed; offset %lld,"
1661 "size %ld, error %d\n",
1662 ((bio
->bio_caller_info1
.index
& SWBIO_READ
) ?
1663 "pagein" : "pageout"),
1664 (long long)bio
->bio_offset
,
1671 * set object, raise to splvm().
1673 if (bp
->b_xio
.xio_npages
)
1674 object
= bp
->b_xio
.xio_pages
[0]->object
;
1678 * remove the mapping for kernel virtual
1680 pmap_qremove((vm_offset_t
)bp
->b_data
, bp
->b_xio
.xio_npages
);
1683 * cleanup pages. If an error occurs writing to swap, we are in
1684 * very serious trouble. If it happens to be a disk error, though,
1685 * we may be able to recover by reassigning the swap later on. So
1686 * in this case we remove the m->swapblk assignment for the page
1687 * but do not free it in the rlist. The errornous block(s) are thus
1688 * never reallocated as swap. Redirty the page and continue.
1690 for (i
= 0; i
< bp
->b_xio
.xio_npages
; ++i
) {
1691 vm_page_t m
= bp
->b_xio
.xio_pages
[i
];
1693 if (bp
->b_flags
& B_ERROR
) {
1695 * If an error occurs I'd love to throw the swapblk
1696 * away without freeing it back to swapspace, so it
1697 * can never be used again. But I can't from an
1701 if (bio
->bio_caller_info1
.index
& SWBIO_READ
) {
1703 * When reading, reqpage needs to stay
1704 * locked for the parent, but all other
1705 * pages can be freed. We still want to
1706 * wakeup the parent waiting on the page,
1707 * though. ( also: pg_reqpage can be -1 and
1708 * not match anything ).
1710 * We have to wake specifically requested pages
1711 * up too because we cleared PG_SWAPINPROG and
1712 * someone may be waiting for that.
1714 * NOTE: for reads, m->dirty will probably
1715 * be overridden by the original caller of
1716 * getpages so don't play cute tricks here.
1718 * NOTE: We can't actually free the page from
1719 * here, because this is an interrupt. It
1720 * is not legal to mess with object->memq
1721 * from an interrupt. Deactivate the page
1726 vm_page_flag_clear(m
, PG_ZERO
);
1727 vm_page_flag_clear(m
, PG_SWAPINPROG
);
1730 * bio_driver_info holds the requested page
1733 if (i
!= (int)(intptr_t)bio
->bio_driver_info
) {
1734 vm_page_deactivate(m
);
1740 * If i == bp->b_pager.pg_reqpage, do not wake
1741 * the page up. The caller needs to.
1745 * If a write error occurs, reactivate page
1746 * so it doesn't clog the inactive list,
1747 * then finish the I/O.
1749 * Only for OBJT_SWAP. When using the swap
1750 * as a cache for clean vnode-backed pages
1751 * we don't mess with the page dirty state.
1753 vm_page_flag_clear(m
, PG_SWAPINPROG
);
1754 if (m
->object
->type
== OBJT_SWAP
) {
1756 vm_page_activate(m
);
1758 vm_page_io_finish(m
);
1760 } else if (bio
->bio_caller_info1
.index
& SWBIO_READ
) {
1762 * NOTE: for reads, m->dirty will probably be
1763 * overridden by the original caller of getpages so
1764 * we cannot set them in order to free the underlying
1765 * swap in a low-swap situation. I don't think we'd
1766 * want to do that anyway, but it was an optimization
1767 * that existed in the old swapper for a time before
1768 * it got ripped out due to precisely this problem.
1770 * clear PG_ZERO in page.
1772 * If not the requested page then deactivate it.
1774 * Note that the requested page, reqpage, is left
1775 * busied, but we still have to wake it up. The
1776 * other pages are released (unbusied) by
1777 * vm_page_wakeup(). We do not set reqpage's
1778 * valid bits here, it is up to the caller.
1782 * NOTE: can't call pmap_clear_modify(m) from an
1783 * interrupt thread, the pmap code may have to map
1784 * non-kernel pmaps and currently asserts the case.
1786 /*pmap_clear_modify(m);*/
1787 m
->valid
= VM_PAGE_BITS_ALL
;
1789 vm_page_flag_clear(m
, PG_ZERO
| PG_SWAPINPROG
);
1790 vm_page_flag_set(m
, PG_SWAPPED
);
1793 * We have to wake specifically requested pages
1794 * up too because we cleared PG_SWAPINPROG and
1795 * could be waiting for it in getpages. However,
1796 * be sure to not unbusy getpages specifically
1797 * requested page - getpages expects it to be
1800 * bio_driver_info holds the requested page
1802 if (i
!= (int)(intptr_t)bio
->bio_driver_info
) {
1803 vm_page_deactivate(m
);
1810 * Mark the page clean but do not mess with the
1811 * pmap-layer's modified state. That state should
1812 * also be clear since the caller protected the
1813 * page VM_PROT_READ, but allow the case.
1815 * We are in an interrupt, avoid pmap operations.
1817 * If we have a severe page deficit, deactivate the
1818 * page. Do not try to cache it (which would also
1819 * involve a pmap op), because the page might still
1822 * When using the swap to cache clean vnode pages
1823 * we do not mess with the page dirty bits.
1825 if (m
->object
->type
== OBJT_SWAP
)
1827 vm_page_flag_clear(m
, PG_SWAPINPROG
);
1828 vm_page_flag_set(m
, PG_SWAPPED
);
1829 vm_page_io_finish(m
);
1830 if (vm_page_count_severe())
1831 vm_page_deactivate(m
);
1833 if (!vm_page_count_severe() || !vm_page_try_to_cache(m
))
1834 vm_page_protect(m
, VM_PROT_READ
);
1840 * adjust pip. NOTE: the original parent may still have its own
1841 * pip refs on the object.
1845 vm_object_pip_wakeupn(object
, bp
->b_xio
.xio_npages
);
1848 * Release the physical I/O buffer.
1850 * NOTE: Due to synchronous operations in the write case b_cmd may
1851 * already be set to BUF_CMD_DONE and BIO_SYNC may have already
1854 if (bio
->bio_caller_info1
.index
& SWBIO_READ
)
1855 nswptr
= &nsw_rcount
;
1856 else if (bio
->bio_caller_info1
.index
& SWBIO_SYNC
)
1857 nswptr
= &nsw_wcount_sync
;
1859 nswptr
= &nsw_wcount_async
;
1860 bp
->b_cmd
= BUF_CMD_DONE
;
1861 relpbuf(bp
, nswptr
);
1865 /************************************************************************
1867 ************************************************************************
1869 * These routines manipulate the swap metadata stored in the
1870 * OBJT_SWAP object. All swp_*() routines must be called at
1871 * splvm() because swap can be freed up by the low level vm_page
1872 * code which might be called from interrupts beyond what splbio() covers.
1874 * Swap metadata is implemented with a global hash and not directly
1875 * linked into the object. Instead the object simply contains
1876 * appropriate tracking counters.
1880 * Lookup the swblock containing the specified swap block index.
1884 swp_pager_lookup(vm_object_t object
, vm_pindex_t index
)
1886 index
&= ~SWAP_META_MASK
;
1887 return (RB_LOOKUP(swblock_rb_tree
, &object
->swblock_root
, index
));
1891 * Remove a swblock from the RB tree.
1895 swp_pager_remove(vm_object_t object
, struct swblock
*swap
)
1897 RB_REMOVE(swblock_rb_tree
, &object
->swblock_root
, swap
);
1901 * Convert default object to swap object if necessary
1904 swp_pager_meta_convert(vm_object_t object
)
1906 if (object
->type
== OBJT_DEFAULT
) {
1907 object
->type
= OBJT_SWAP
;
1908 KKASSERT(object
->swblock_count
== 0);
1913 * SWP_PAGER_META_BUILD() - add swap block to swap meta data for object
1915 * We first convert the object to a swap object if it is a default
1916 * object. Vnode objects do not need to be converted.
1918 * The specified swapblk is added to the object's swap metadata. If
1919 * the swapblk is not valid, it is freed instead. Any previously
1920 * assigned swapblk is freed.
1923 swp_pager_meta_build(vm_object_t object
, vm_pindex_t index
, daddr_t swapblk
)
1925 struct swblock
*swap
;
1926 struct swblock
*oswap
;
1928 KKASSERT(swapblk
!= SWAPBLK_NONE
);
1931 * Convert object if necessary
1933 if (object
->type
== OBJT_DEFAULT
)
1934 swp_pager_meta_convert(object
);
1937 * Locate swblock. If not found create, but if we aren't adding
1938 * anything just return. If we run out of space in the map we wait
1939 * and, since the hash table may have changed, retry.
1942 swap
= swp_pager_lookup(object
, index
);
1947 swap
= zalloc(swap_zone
);
1952 swap
->swb_index
= index
& ~SWAP_META_MASK
;
1953 swap
->swb_count
= 0;
1955 ++object
->swblock_count
;
1957 for (i
= 0; i
< SWAP_META_PAGES
; ++i
)
1958 swap
->swb_pages
[i
] = SWAPBLK_NONE
;
1959 oswap
= RB_INSERT(swblock_rb_tree
, &object
->swblock_root
, swap
);
1960 KKASSERT(oswap
== NULL
);
1964 * Delete prior contents of metadata
1967 index
&= SWAP_META_MASK
;
1969 if (swap
->swb_pages
[index
] != SWAPBLK_NONE
) {
1970 swp_pager_freeswapspace(object
, swap
->swb_pages
[index
], 1);
1975 * Enter block into metadata
1977 swap
->swb_pages
[index
] = swapblk
;
1978 if (swapblk
!= SWAPBLK_NONE
)
1983 * SWP_PAGER_META_FREE() - free a range of blocks in the object's swap metadata
1985 * The requested range of blocks is freed, with any associated swap
1986 * returned to the swap bitmap.
1988 * This routine will free swap metadata structures as they are cleaned
1989 * out. This routine does *NOT* operate on swap metadata associated
1990 * with resident pages.
1992 * This routine must be called at splvm()
1994 static int swp_pager_meta_free_callback(struct swblock
*swb
, void *data
);
1997 swp_pager_meta_free(vm_object_t object
, vm_pindex_t index
, vm_pindex_t count
)
1999 struct swfreeinfo info
;
2004 if (object
->swblock_count
== 0) {
2005 KKASSERT(RB_EMPTY(&object
->swblock_root
));
2012 * Setup for RB tree scan. Note that the pindex range can be huge
2013 * due to the 64 bit page index space so we cannot safely iterate.
2015 info
.object
= object
;
2016 info
.basei
= index
& ~SWAP_META_MASK
;
2018 info
.endi
= index
+ count
- 1;
2019 swblock_rb_tree_RB_SCAN(&object
->swblock_root
, rb_swblock_scancmp
,
2020 swp_pager_meta_free_callback
, &info
);
2025 swp_pager_meta_free_callback(struct swblock
*swap
, void *data
)
2027 struct swfreeinfo
*info
= data
;
2028 vm_object_t object
= info
->object
;
2033 * Figure out the range within the swblock. The wider scan may
2034 * return edge-case swap blocks when the start and/or end points
2035 * are in the middle of a block.
2037 if (swap
->swb_index
< info
->begi
)
2038 index
= (int)info
->begi
& SWAP_META_MASK
;
2042 if (swap
->swb_index
+ SWAP_META_PAGES
> info
->endi
)
2043 eindex
= (int)info
->endi
& SWAP_META_MASK
;
2045 eindex
= SWAP_META_MASK
;
2048 * Scan and free the blocks. The loop terminates early
2049 * if (swap) runs out of blocks and could be freed.
2051 while (index
<= eindex
) {
2052 daddr_t v
= swap
->swb_pages
[index
];
2054 if (v
!= SWAPBLK_NONE
) {
2055 swp_pager_freeswapspace(object
, v
, 1);
2056 swap
->swb_pages
[index
] = SWAPBLK_NONE
;
2057 if (--swap
->swb_count
== 0) {
2058 swp_pager_remove(object
, swap
);
2059 zfree(swap_zone
, swap
);
2060 --object
->swblock_count
;
2066 /* swap may be invalid here due to zfree above */
2071 * SWP_PAGER_META_FREE_ALL() - destroy all swap metadata associated with object
2073 * This routine locates and destroys all swap metadata associated with
2076 * This routine must be called at splvm()
2079 swp_pager_meta_free_all(vm_object_t object
)
2081 struct swblock
*swap
;
2084 while ((swap
= RB_ROOT(&object
->swblock_root
)) != NULL
) {
2085 swp_pager_remove(object
, swap
);
2086 for (i
= 0; i
< SWAP_META_PAGES
; ++i
) {
2087 daddr_t v
= swap
->swb_pages
[i
];
2088 if (v
!= SWAPBLK_NONE
) {
2090 swp_pager_freeswapspace(object
, v
, 1);
2093 if (swap
->swb_count
!= 0)
2094 panic("swap_pager_meta_free_all: swb_count != 0");
2095 zfree(swap_zone
, swap
);
2096 --object
->swblock_count
;
2098 KKASSERT(object
->swblock_count
== 0);
2102 * SWP_PAGER_METACTL() - misc control of swap and vm_page_t meta data.
2104 * This routine is capable of looking up, popping, or freeing
2105 * swapblk assignments in the swap meta data or in the vm_page_t.
2106 * The routine typically returns the swapblk being looked-up, or popped,
2107 * or SWAPBLK_NONE if the block was freed, or SWAPBLK_NONE if the block
2108 * was invalid. This routine will automatically free any invalid
2109 * meta-data swapblks.
2111 * It is not possible to store invalid swapblks in the swap meta data
2112 * (other then a literal 'SWAPBLK_NONE'), so we don't bother checking.
2114 * When acting on a busy resident page and paging is in progress, we
2115 * have to wait until paging is complete but otherwise can act on the
2118 * This routine must be called at splvm().
2120 * SWM_FREE remove and free swap block from metadata
2121 * SWM_POP remove from meta data but do not free.. pop it out
2124 swp_pager_meta_ctl(vm_object_t object
, vm_pindex_t index
, int flags
)
2126 struct swblock
*swap
;
2129 if (object
->swblock_count
== 0)
2130 return(SWAPBLK_NONE
);
2133 swap
= swp_pager_lookup(object
, index
);
2136 index
&= SWAP_META_MASK
;
2137 r1
= swap
->swb_pages
[index
];
2139 if (r1
!= SWAPBLK_NONE
) {
2140 if (flags
& SWM_FREE
) {
2141 swp_pager_freeswapspace(object
, r1
, 1);
2144 if (flags
& (SWM_FREE
|SWM_POP
)) {
2145 swap
->swb_pages
[index
] = SWAPBLK_NONE
;
2146 if (--swap
->swb_count
== 0) {
2147 swp_pager_remove(object
, swap
);
2148 zfree(swap_zone
, swap
);
2149 --object
->swblock_count
;