4 * Copyright (c) 1998-2010 The DragonFly Project. All rights reserved.
6 * This code is derived from software contributed to The DragonFly Project
7 * by Matthew Dillon <dillon@backplane.com>
9 * Redistribution and use in source and binary forms, with or without
10 * modification, are permitted provided that the following conditions
13 * 1. Redistributions of source code must retain the above copyright
14 * notice, this list of conditions and the following disclaimer.
15 * 2. Redistributions in binary form must reproduce the above copyright
16 * notice, this list of conditions and the following disclaimer in
17 * the documentation and/or other materials provided with the
19 * 3. Neither the name of The DragonFly Project nor the names of its
20 * contributors may be used to endorse or promote products derived
21 * from this software without specific, prior written permission.
23 * THIS SOFTWARE IS PROVIDED BY THE COPYRIGHT HOLDERS AND CONTRIBUTORS
24 * ``AS IS'' AND ANY EXPRESS OR IMPLIED WARRANTIES, INCLUDING, BUT NOT
25 * LIMITED TO, THE IMPLIED WARRANTIES OF MERCHANTABILITY AND FITNESS
26 * FOR A PARTICULAR PURPOSE ARE DISCLAIMED. IN NO EVENT SHALL THE
27 * COPYRIGHT HOLDERS OR CONTRIBUTORS BE LIABLE FOR ANY DIRECT, INDIRECT,
28 * INCIDENTAL, SPECIAL, EXEMPLARY OR CONSEQUENTIAL DAMAGES (INCLUDING,
29 * BUT NOT LIMITED TO, PROCUREMENT OF SUBSTITUTE GOODS OR SERVICES;
30 * LOSS OF USE, DATA, OR PROFITS; OR BUSINESS INTERRUPTION) HOWEVER CAUSED
31 * AND ON ANY THEORY OF LIABILITY, WHETHER IN CONTRACT, STRICT LIABILITY,
32 * OR TORT (INCLUDING NEGLIGENCE OR OTHERWISE) ARISING IN ANY WAY OUT
33 * OF THE USE OF THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF
36 * Copyright (c) 1994 John S. Dyson
37 * Copyright (c) 1990 University of Utah.
38 * Copyright (c) 1991, 1993
39 * The Regents of the University of California. All rights reserved.
41 * This code is derived from software contributed to Berkeley by
42 * the Systems Programming Group of the University of Utah Computer
45 * Redistribution and use in source and binary forms, with or without
46 * modification, are permitted provided that the following conditions
48 * 1. Redistributions of source code must retain the above copyright
49 * notice, this list of conditions and the following disclaimer.
50 * 2. Redistributions in binary form must reproduce the above copyright
51 * notice, this list of conditions and the following disclaimer in the
52 * documentation and/or other materials provided with the distribution.
53 * 3. Neither the name of the University nor the names of its contributors
54 * may be used to endorse or promote products derived from this software
55 * without specific prior written permission.
57 * THIS SOFTWARE IS PROVIDED BY THE REGENTS AND CONTRIBUTORS ``AS IS'' AND
58 * ANY EXPRESS OR IMPLIED WARRANTIES, INCLUDING, BUT NOT LIMITED TO, THE
59 * IMPLIED WARRANTIES OF MERCHANTABILITY AND FITNESS FOR A PARTICULAR PURPOSE
60 * ARE DISCLAIMED. IN NO EVENT SHALL THE REGENTS OR CONTRIBUTORS BE LIABLE
61 * FOR ANY DIRECT, INDIRECT, INCIDENTAL, SPECIAL, EXEMPLARY, OR CONSEQUENTIAL
62 * DAMAGES (INCLUDING, BUT NOT LIMITED TO, PROCUREMENT OF SUBSTITUTE GOODS
63 * OR SERVICES; LOSS OF USE, DATA, OR PROFITS; OR BUSINESS INTERRUPTION)
64 * HOWEVER CAUSED AND ON ANY THEORY OF LIABILITY, WHETHER IN CONTRACT, STRICT
65 * LIABILITY, OR TORT (INCLUDING NEGLIGENCE OR OTHERWISE) ARISING IN ANY WAY
66 * OUT OF THE USE OF THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF
72 * Radix Bitmap 'blists'.
74 * - The new swapper uses the new radix bitmap code. This should scale
75 * to arbitrarily small or arbitrarily large swap spaces and an almost
76 * arbitrary degree of fragmentation.
80 * - on the fly reallocation of swap during putpages. The new system
81 * does not try to keep previously allocated swap blocks for dirty
84 * - on the fly deallocation of swap
86 * - No more garbage collection required. Unnecessarily allocated swap
87 * blocks only exist for dirty vm_page_t's now and these are already
88 * cycled (in a high-load system) by the pager. We also do on-the-fly
89 * removal of invalidated swap blocks when a page is destroyed
92 * from: Utah $Hdr: swap_pager.c 1.4 91/04/30$
93 * @(#)swap_pager.c 8.9 (Berkeley) 3/21/94
94 * $FreeBSD: src/sys/vm/swap_pager.c,v 1.130.2.12 2002/08/31 21:15:55 dillon Exp $
97 #include <sys/param.h>
98 #include <sys/systm.h>
100 #include <sys/kernel.h>
101 #include <sys/proc.h>
103 #include <sys/vnode.h>
104 #include <sys/malloc.h>
105 #include <sys/vmmeter.h>
106 #include <sys/sysctl.h>
107 #include <sys/blist.h>
108 #include <sys/lock.h>
109 #include <sys/thread2.h>
111 #include "opt_swap.h"
113 #include <vm/vm_object.h>
114 #include <vm/vm_page.h>
115 #include <vm/vm_pager.h>
116 #include <vm/vm_pageout.h>
117 #include <vm/swap_pager.h>
118 #include <vm/vm_extern.h>
119 #include <vm/vm_zone.h>
120 #include <vm/vnode_pager.h>
122 #include <sys/buf2.h>
123 #include <vm/vm_page2.h>
125 #ifndef MAX_PAGEOUT_CLUSTER
126 #define MAX_PAGEOUT_CLUSTER SWB_NPAGES
129 #define SWM_FREE 0x02 /* free, period */
130 #define SWM_POP 0x04 /* pop out */
132 #define SWBIO_READ 0x01
133 #define SWBIO_WRITE 0x02
134 #define SWBIO_SYNC 0x04
135 #define SWBIO_TTC 0x08 /* for VM_PAGER_TRY_TO_CACHE */
141 vm_pindex_t endi
; /* inclusive */
144 struct swswapoffinfo
{
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 swap_fail_ticks
; /* when we became exhausted */
157 int swap_pager_almost_full
; /* swap space exhaustion (w/ hysteresis)*/
158 swblk_t vm_swap_cache_use
;
159 swblk_t vm_swap_anon_use
;
160 static int vm_report_swap_allocs
;
162 static int nsw_rcount
; /* free read buffers */
163 static int nsw_wcount_sync
; /* limit write buffers / synchronous */
164 static int nsw_wcount_async
; /* limit write buffers / asynchronous */
165 static int nsw_wcount_async_max
;/* assigned maximum */
166 static int nsw_cluster_max
; /* maximum VOP I/O allowed */
168 struct blist
*swapblist
;
169 static int swap_async_max
= 4; /* maximum in-progress async I/O's */
170 static int swap_burst_read
= 0; /* allow burst reading */
171 static swblk_t swapiterator
; /* linearize allocations */
172 int swap_user_async
= 0; /* user swap pager operation can be async */
174 static struct spinlock swapbp_spin
= SPINLOCK_INITIALIZER(&swapbp_spin
, "swapbp_spin");
177 extern struct vnode
*swapdev_vp
;
178 extern struct swdevt
*swdevt
;
181 #define BLK2DEVIDX(blk) (nswdev > 1 ? blk / dmmax % nswdev : 0)
183 SYSCTL_INT(_vm
, OID_AUTO
, swap_async_max
,
184 CTLFLAG_RW
, &swap_async_max
, 0, "Maximum running async swap ops");
185 SYSCTL_INT(_vm
, OID_AUTO
, swap_burst_read
,
186 CTLFLAG_RW
, &swap_burst_read
, 0, "Allow burst reads for pageins");
187 SYSCTL_INT(_vm
, OID_AUTO
, swap_user_async
,
188 CTLFLAG_RW
, &swap_user_async
, 0, "Allow async uuser swap write I/O");
191 SYSCTL_LONG(_vm
, OID_AUTO
, swap_cache_use
,
192 CTLFLAG_RD
, &vm_swap_cache_use
, 0, "");
193 SYSCTL_LONG(_vm
, OID_AUTO
, swap_anon_use
,
194 CTLFLAG_RD
, &vm_swap_anon_use
, 0, "");
195 SYSCTL_LONG(_vm
, OID_AUTO
, swap_size
,
196 CTLFLAG_RD
, &vm_swap_size
, 0, "");
198 SYSCTL_INT(_vm
, OID_AUTO
, swap_cache_use
,
199 CTLFLAG_RD
, &vm_swap_cache_use
, 0, "");
200 SYSCTL_INT(_vm
, OID_AUTO
, swap_anon_use
,
201 CTLFLAG_RD
, &vm_swap_anon_use
, 0, "");
202 SYSCTL_INT(_vm
, OID_AUTO
, swap_size
,
203 CTLFLAG_RD
, &vm_swap_size
, 0, "");
205 SYSCTL_INT(_vm
, OID_AUTO
, report_swap_allocs
,
206 CTLFLAG_RW
, &vm_report_swap_allocs
, 0, "");
211 * Red-Black tree for swblock entries
213 * The caller must hold vm_token
215 RB_GENERATE2(swblock_rb_tree
, swblock
, swb_entry
, rb_swblock_compare
,
216 vm_pindex_t
, swb_index
);
219 rb_swblock_compare(struct swblock
*swb1
, struct swblock
*swb2
)
221 if (swb1
->swb_index
< swb2
->swb_index
)
223 if (swb1
->swb_index
> swb2
->swb_index
)
230 rb_swblock_scancmp(struct swblock
*swb
, void *data
)
232 struct swfreeinfo
*info
= data
;
234 if (swb
->swb_index
< info
->basei
)
236 if (swb
->swb_index
> info
->endi
)
243 rb_swblock_condcmp(struct swblock
*swb
, void *data
)
245 struct swfreeinfo
*info
= data
;
247 if (swb
->swb_index
< info
->basei
)
253 * pagerops for OBJT_SWAP - "swap pager". Some ops are also global procedure
254 * calls hooked from other parts of the VM system and do not appear here.
255 * (see vm/swap_pager.h).
258 static void swap_pager_dealloc (vm_object_t object
);
259 static int swap_pager_getpage (vm_object_t
, vm_page_t
*, int);
260 static void swap_chain_iodone(struct bio
*biox
);
262 struct pagerops swappagerops
= {
263 swap_pager_dealloc
, /* deallocate an OBJT_SWAP object */
264 swap_pager_getpage
, /* pagein */
265 swap_pager_putpages
, /* pageout */
266 swap_pager_haspage
/* get backing store status for page */
270 * dmmax is in page-sized chunks with the new swap system. It was
271 * dev-bsized chunks in the old. dmmax is always a power of 2.
273 * swap_*() routines are externally accessible. swp_*() routines are
278 static int dmmax_mask
;
279 int nswap_lowat
= 128; /* in pages, swap_pager_almost_full warn */
280 int nswap_hiwat
= 512; /* in pages, swap_pager_almost_full warn */
282 static __inline
void swp_sizecheck (void);
283 static void swp_pager_async_iodone (struct bio
*bio
);
286 * Swap bitmap functions
289 static __inline
void swp_pager_freeswapspace(vm_object_t object
,
290 swblk_t blk
, int npages
);
291 static __inline swblk_t
swp_pager_getswapspace(vm_object_t object
, int npages
);
297 static void swp_pager_meta_convert(vm_object_t
);
298 static void swp_pager_meta_build(vm_object_t
, vm_pindex_t
, swblk_t
);
299 static void swp_pager_meta_free(vm_object_t
, vm_pindex_t
, vm_pindex_t
);
300 static void swp_pager_meta_free_all(vm_object_t
);
301 static swblk_t
swp_pager_meta_ctl(vm_object_t
, vm_pindex_t
, int);
304 * SWP_SIZECHECK() - update swap_pager_full indication
306 * update the swap_pager_almost_full indication and warn when we are
307 * about to run out of swap space, using lowat/hiwat hysteresis.
309 * Clear swap_pager_full ( task killing ) indication when lowat is met.
311 * No restrictions on call
312 * This routine may not block.
318 if (vm_swap_size
< nswap_lowat
) {
319 if (swap_pager_almost_full
== 0) {
320 kprintf("swap_pager: out of swap space\n");
321 swap_pager_almost_full
= 1;
322 swap_fail_ticks
= ticks
;
326 if (vm_swap_size
> nswap_hiwat
)
327 swap_pager_almost_full
= 0;
332 * SWAP_PAGER_INIT() - initialize the swap pager!
334 * Expected to be started from system init. NOTE: This code is run
335 * before much else so be careful what you depend on. Most of the VM
336 * system has yet to be initialized at this point.
338 * Called from the low level boot code only.
341 swap_pager_init(void *arg __unused
)
344 * Device Stripe, in PAGE_SIZE'd blocks
346 dmmax
= SWB_NPAGES
* 2;
347 dmmax_mask
= ~(dmmax
- 1);
349 SYSINIT(vm_mem
, SI_BOOT1_VM
, SI_ORDER_THIRD
, swap_pager_init
, NULL
);
352 * SWAP_PAGER_SWAP_INIT() - swap pager initialization from pageout process
354 * Expected to be started from pageout process once, prior to entering
357 * Called from the low level boot code only.
360 swap_pager_swap_init(void)
365 * Number of in-transit swap bp operations. Don't
366 * exhaust the pbufs completely. Make sure we
367 * initialize workable values (0 will work for hysteresis
368 * but it isn't very efficient).
370 * The nsw_cluster_max is constrained by the number of pages an XIO
371 * holds, i.e., (MAXPHYS/PAGE_SIZE) and our locally defined
372 * MAX_PAGEOUT_CLUSTER. Also be aware that swap ops are
373 * constrained by the swap device interleave stripe size.
375 * Currently we hardwire nsw_wcount_async to 4. This limit is
376 * designed to prevent other I/O from having high latencies due to
377 * our pageout I/O. The value 4 works well for one or two active swap
378 * devices but is probably a little low if you have more. Even so,
379 * a higher value would probably generate only a limited improvement
380 * with three or four active swap devices since the system does not
381 * typically have to pageout at extreme bandwidths. We will want
382 * at least 2 per swap devices, and 4 is a pretty good value if you
383 * have one NFS swap device due to the command/ack latency over NFS.
384 * So it all works out pretty well.
387 nsw_cluster_max
= min((MAXPHYS
/PAGE_SIZE
), MAX_PAGEOUT_CLUSTER
);
389 nsw_rcount
= (nswbuf_kva
+ 1) / 2;
390 nsw_wcount_sync
= (nswbuf_kva
+ 3) / 4;
391 nsw_wcount_async
= 4;
392 nsw_wcount_async_max
= nsw_wcount_async
;
395 * The zone is dynamically allocated so generally size it to
396 * maxswzone (32MB to 256GB of KVM). Set a minimum size based
397 * on physical memory of around 8x (each swblock can hold 16 pages).
399 * With the advent of SSDs (vs HDs) the practical (swap:memory) ratio
400 * has increased dramatically.
402 n
= vmstats
.v_page_count
/ 2;
403 if (maxswzone
&& n
< maxswzone
/ sizeof(struct swblock
))
404 n
= maxswzone
/ sizeof(struct swblock
);
410 sizeof(struct swblock
),
413 if (swap_zone
!= NULL
)
416 * if the allocation failed, try a zone two thirds the
417 * size of the previous attempt.
422 if (swap_zone
== NULL
)
423 panic("swap_pager_swap_init: swap_zone == NULL");
425 kprintf("Swap zone entries reduced from %d to %d.\n", n2
, n
);
429 * SWAP_PAGER_ALLOC() - allocate a new OBJT_SWAP VM object and instantiate
430 * its metadata structures.
432 * This routine is called from the mmap and fork code to create a new
433 * OBJT_SWAP object. We do this by creating an OBJT_DEFAULT object
434 * and then converting it with swp_pager_meta_convert().
436 * We only support unnamed objects.
441 swap_pager_alloc(void *handle
, off_t size
, vm_prot_t prot
, off_t offset
)
445 KKASSERT(handle
== NULL
);
446 object
= vm_object_allocate_hold(OBJT_DEFAULT
,
447 OFF_TO_IDX(offset
+ PAGE_MASK
+ size
));
448 swp_pager_meta_convert(object
);
449 vm_object_drop(object
);
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 * The object must be locked or unreferenceable.
463 * No other requirements.
466 swap_pager_dealloc(vm_object_t object
)
468 vm_object_hold(object
);
469 vm_object_pip_wait(object
, "swpdea");
472 * Free all remaining metadata. We only bother to free it from
473 * the swap meta data. We do not attempt to free swapblk's still
474 * associated with vm_page_t's for this object. We do not care
475 * if paging is still in progress on some objects.
477 swp_pager_meta_free_all(object
);
478 vm_object_drop(object
);
481 /************************************************************************
482 * SWAP PAGER BITMAP ROUTINES *
483 ************************************************************************/
486 * SWP_PAGER_GETSWAPSPACE() - allocate raw swap space
488 * Allocate swap for the requested number of pages. The starting
489 * swap block number (a page index) is returned or SWAPBLK_NONE
490 * if the allocation failed.
492 * Also has the side effect of advising that somebody made a mistake
493 * when they configured swap and didn't configure enough.
495 * The caller must hold the object.
496 * This routine may not block.
498 static __inline swblk_t
499 swp_pager_getswapspace(vm_object_t object
, int npages
)
503 lwkt_gettoken(&vm_token
);
504 blk
= blist_allocat(swapblist
, npages
, swapiterator
);
505 if (blk
== SWAPBLK_NONE
)
506 blk
= blist_allocat(swapblist
, npages
, 0);
507 if (blk
== SWAPBLK_NONE
) {
508 if (swap_pager_full
!= 2) {
509 if (vm_swap_max
== 0)
510 kprintf("Warning: The system would like to "
511 "page to swap but no swap space "
514 kprintf("swap_pager_getswapspace: "
515 "swap full allocating %d pages\n",
518 if (swap_pager_almost_full
== 0)
519 swap_fail_ticks
= ticks
;
520 swap_pager_almost_full
= 1;
523 /* swapiterator = blk; disable for now, doesn't work well */
524 swapacctspace(blk
, -npages
);
525 if (object
->type
== OBJT_SWAP
)
526 vm_swap_anon_use
+= npages
;
528 vm_swap_cache_use
+= npages
;
531 lwkt_reltoken(&vm_token
);
536 * SWP_PAGER_FREESWAPSPACE() - free raw swap space
538 * This routine returns the specified swap blocks back to the bitmap.
540 * Note: This routine may not block (it could in the old swap code),
541 * and through the use of the new blist routines it does not block.
543 * We must be called at splvm() to avoid races with bitmap frees from
544 * vm_page_remove() aka swap_pager_page_removed().
546 * This routine may not block.
550 swp_pager_freeswapspace(vm_object_t object
, swblk_t blk
, int npages
)
552 struct swdevt
*sp
= &swdevt
[BLK2DEVIDX(blk
)];
554 lwkt_gettoken(&vm_token
);
555 sp
->sw_nused
-= npages
;
556 if (object
->type
== OBJT_SWAP
)
557 vm_swap_anon_use
-= npages
;
559 vm_swap_cache_use
-= npages
;
561 if (sp
->sw_flags
& SW_CLOSING
) {
562 lwkt_reltoken(&vm_token
);
566 blist_free(swapblist
, blk
, npages
);
567 vm_swap_size
+= npages
;
569 lwkt_reltoken(&vm_token
);
573 * SWAP_PAGER_FREESPACE() - frees swap blocks associated with a page
574 * range within an object.
576 * This is a globally accessible routine.
578 * This routine removes swapblk assignments from swap metadata.
580 * The external callers of this routine typically have already destroyed
581 * or renamed vm_page_t's associated with this range in the object so
587 swap_pager_freespace(vm_object_t object
, vm_pindex_t start
, vm_pindex_t size
)
589 vm_object_hold(object
);
590 swp_pager_meta_free(object
, start
, size
);
591 vm_object_drop(object
);
598 swap_pager_freespace_all(vm_object_t object
)
600 vm_object_hold(object
);
601 swp_pager_meta_free_all(object
);
602 vm_object_drop(object
);
606 * This function conditionally frees swap cache swap starting at
607 * (*basei) in the object. (count) swap blocks will be nominally freed.
608 * The actual number of blocks freed can be more or less than the
611 * This function nominally returns the number of blocks freed. However,
612 * the actual number of blocks freed may be less then the returned value.
613 * If the function is unable to exhaust the object or if it is able to
614 * free (approximately) the requested number of blocks it returns
617 * If we exhaust the object we will return a value n <= count.
619 * The caller must hold the object.
621 * WARNING! If count == 0 then -1 can be returned as a degenerate case,
622 * callers should always pass a count value > 0.
624 static int swap_pager_condfree_callback(struct swblock
*swap
, void *data
);
627 swap_pager_condfree(vm_object_t object
, vm_pindex_t
*basei
, int count
)
629 struct swfreeinfo info
;
633 ASSERT_LWKT_TOKEN_HELD(vm_object_token(object
));
635 info
.object
= object
;
636 info
.basei
= *basei
; /* skip up to this page index */
637 info
.begi
= count
; /* max swap pages to destroy */
638 info
.endi
= count
* 8; /* max swblocks to scan */
640 swblock_rb_tree_RB_SCAN(&object
->swblock_root
, rb_swblock_condcmp
,
641 swap_pager_condfree_callback
, &info
);
645 * Take the higher difference swblocks vs pages
647 n
= count
- (int)info
.begi
;
648 t
= count
* 8 - (int)info
.endi
;
657 * The idea is to free whole meta-block to avoid fragmenting
658 * the swap space or disk I/O. We only do this if NO VM pages
661 * We do not have to deal with clearing PG_SWAPPED in related VM
662 * pages because there are no related VM pages.
664 * The caller must hold the object.
667 swap_pager_condfree_callback(struct swblock
*swap
, void *data
)
669 struct swfreeinfo
*info
= data
;
670 vm_object_t object
= info
->object
;
673 for (i
= 0; i
< SWAP_META_PAGES
; ++i
) {
674 if (vm_page_lookup(object
, swap
->swb_index
+ i
))
677 info
->basei
= swap
->swb_index
+ SWAP_META_PAGES
;
678 if (i
== SWAP_META_PAGES
) {
679 info
->begi
-= swap
->swb_count
;
680 swap_pager_freespace(object
, swap
->swb_index
, SWAP_META_PAGES
);
683 if ((int)info
->begi
< 0 || (int)info
->endi
< 0)
690 * Called by vm_page_alloc() when a new VM page is inserted
691 * into a VM object. Checks whether swap has been assigned to
692 * the page and sets PG_SWAPPED as necessary.
697 swap_pager_page_inserted(vm_page_t m
)
699 if (m
->object
->swblock_count
) {
700 vm_object_hold(m
->object
);
701 if (swp_pager_meta_ctl(m
->object
, m
->pindex
, 0) != SWAPBLK_NONE
)
702 vm_page_flag_set(m
, PG_SWAPPED
);
703 vm_object_drop(m
->object
);
708 * SWAP_PAGER_RESERVE() - reserve swap blocks in object
710 * Assigns swap blocks to the specified range within the object. The
711 * swap blocks are not zerod. Any previous swap assignment is destroyed.
713 * Returns 0 on success, -1 on failure.
715 * The caller is responsible for avoiding races in the specified range.
716 * No other requirements.
719 swap_pager_reserve(vm_object_t object
, vm_pindex_t start
, vm_size_t size
)
722 swblk_t blk
= SWAPBLK_NONE
;
723 vm_pindex_t beg
= start
; /* save start index */
725 vm_object_hold(object
);
730 while ((blk
= swp_pager_getswapspace(object
, n
)) ==
735 swp_pager_meta_free(object
, beg
,
737 vm_object_drop(object
);
742 swp_pager_meta_build(object
, start
, blk
);
748 swp_pager_meta_free(object
, start
, n
);
749 vm_object_drop(object
);
754 * SWAP_PAGER_COPY() - copy blocks from source pager to destination pager
755 * and destroy the source.
757 * Copy any valid swapblks from the source to the destination. In
758 * cases where both the source and destination have a valid swapblk,
759 * we keep the destination's.
761 * This routine is allowed to block. It may block allocating metadata
762 * indirectly through swp_pager_meta_build() or if paging is still in
763 * progress on the source.
765 * XXX vm_page_collapse() kinda expects us not to block because we
766 * supposedly do not need to allocate memory, but for the moment we
767 * *may* have to get a little memory from the zone allocator, but
768 * it is taken from the interrupt memory. We should be ok.
770 * The source object contains no vm_page_t's (which is just as well)
771 * The source object is of type OBJT_SWAP.
773 * The source and destination objects must be held by the caller.
776 swap_pager_copy(vm_object_t srcobject
, vm_object_t dstobject
,
777 vm_pindex_t base_index
, int destroysource
)
781 ASSERT_LWKT_TOKEN_HELD(vm_object_token(srcobject
));
782 ASSERT_LWKT_TOKEN_HELD(vm_object_token(dstobject
));
785 * transfer source to destination.
787 for (i
= 0; i
< dstobject
->size
; ++i
) {
791 * Locate (without changing) the swapblk on the destination,
792 * unless it is invalid in which case free it silently, or
793 * if the destination is a resident page, in which case the
794 * source is thrown away.
796 dstaddr
= swp_pager_meta_ctl(dstobject
, i
, 0);
798 if (dstaddr
== SWAPBLK_NONE
) {
800 * Destination has no swapblk and is not resident,
805 srcaddr
= swp_pager_meta_ctl(srcobject
,
806 base_index
+ i
, SWM_POP
);
808 if (srcaddr
!= SWAPBLK_NONE
)
809 swp_pager_meta_build(dstobject
, i
, srcaddr
);
812 * Destination has valid swapblk or it is represented
813 * by a resident page. We destroy the sourceblock.
815 swp_pager_meta_ctl(srcobject
, base_index
+ i
, SWM_FREE
);
820 * Free left over swap blocks in source.
822 * We have to revert the type to OBJT_DEFAULT so we do not accidently
823 * double-remove the object from the swap queues.
827 * Reverting the type is not necessary, the caller is going
828 * to destroy srcobject directly, but I'm doing it here
829 * for consistency since we've removed the object from its
832 swp_pager_meta_free_all(srcobject
);
833 if (srcobject
->type
== OBJT_SWAP
)
834 srcobject
->type
= OBJT_DEFAULT
;
839 * SWAP_PAGER_HASPAGE() - determine if we have good backing store for
840 * the requested page.
842 * We determine whether good backing store exists for the requested
843 * page and return TRUE if it does, FALSE if it doesn't.
845 * If TRUE, we also try to determine how much valid, contiguous backing
846 * store exists before and after the requested page within a reasonable
847 * distance. We do not try to restrict it to the swap device stripe
848 * (that is handled in getpages/putpages). It probably isn't worth
854 swap_pager_haspage(vm_object_t object
, vm_pindex_t pindex
)
859 * do we have good backing store at the requested index ?
861 vm_object_hold(object
);
862 blk0
= swp_pager_meta_ctl(object
, pindex
, 0);
864 if (blk0
== SWAPBLK_NONE
) {
865 vm_object_drop(object
);
868 vm_object_drop(object
);
873 * SWAP_PAGER_PAGE_UNSWAPPED() - remove swap backing store related to page
875 * This removes any associated swap backing store, whether valid or
876 * not, from the page. This operates on any VM object, not just OBJT_SWAP
879 * This routine is typically called when a page is made dirty, at
880 * which point any associated swap can be freed. MADV_FREE also
881 * calls us in a special-case situation
883 * NOTE!!! If the page is clean and the swap was valid, the caller
884 * should make the page dirty before calling this routine. This routine
885 * does NOT change the m->dirty status of the page. Also: MADV_FREE
888 * The page must be busied or soft-busied.
889 * The caller can hold the object to avoid blocking, else we might block.
890 * No other requirements.
893 swap_pager_unswapped(vm_page_t m
)
895 if (m
->flags
& PG_SWAPPED
) {
896 vm_object_hold(m
->object
);
897 KKASSERT(m
->flags
& PG_SWAPPED
);
898 swp_pager_meta_ctl(m
->object
, m
->pindex
, SWM_FREE
);
899 vm_page_flag_clear(m
, PG_SWAPPED
);
900 vm_object_drop(m
->object
);
905 * SWAP_PAGER_STRATEGY() - read, write, free blocks
907 * This implements a VM OBJECT strategy function using swap backing store.
908 * This can operate on any VM OBJECT type, not necessarily just OBJT_SWAP
911 * This is intended to be a cacheless interface (i.e. caching occurs at
912 * higher levels), and is also used as a swap-based SSD cache for vnode
913 * and device objects.
915 * All I/O goes directly to and from the swap device.
917 * We currently attempt to run I/O synchronously or asynchronously as
918 * the caller requests. This isn't perfect because we loose error
919 * sequencing when we run multiple ops in parallel to satisfy a request.
920 * But this is swap, so we let it all hang out.
925 swap_pager_strategy(vm_object_t object
, struct bio
*bio
)
927 struct buf
*bp
= bio
->bio_buf
;
930 vm_pindex_t biox_blkno
= 0;
936 struct bio_track
*track
;
941 * tracking for swapdev vnode I/Os
943 if (bp
->b_cmd
== BUF_CMD_READ
)
944 track
= &swapdev_vp
->v_track_read
;
946 track
= &swapdev_vp
->v_track_write
;
949 if (bp
->b_bcount
& PAGE_MASK
) {
950 bp
->b_error
= EINVAL
;
951 bp
->b_flags
|= B_ERROR
| B_INVAL
;
953 kprintf("swap_pager_strategy: bp %p offset %lld size %d, "
954 "not page bounded\n",
955 bp
, (long long)bio
->bio_offset
, (int)bp
->b_bcount
);
960 * Clear error indication, initialize page index, count, data pointer.
963 bp
->b_flags
&= ~B_ERROR
;
964 bp
->b_resid
= bp
->b_bcount
;
966 start
= (vm_pindex_t
)(bio
->bio_offset
>> PAGE_SHIFT
);
967 count
= howmany(bp
->b_bcount
, PAGE_SIZE
);
971 * Deal with BUF_CMD_FREEBLKS
973 if (bp
->b_cmd
== BUF_CMD_FREEBLKS
) {
975 * FREE PAGE(s) - destroy underlying swap that is no longer
978 vm_object_hold(object
);
979 swp_pager_meta_free(object
, start
, count
);
980 vm_object_drop(object
);
987 * We need to be able to create a new cluster of I/O's. We cannot
988 * use the caller fields of the passed bio so push a new one.
990 * Because nbio is just a placeholder for the cluster links,
991 * we can biodone() the original bio instead of nbio to make
992 * things a bit more efficient.
994 nbio
= push_bio(bio
);
995 nbio
->bio_offset
= bio
->bio_offset
;
996 nbio
->bio_caller_info1
.cluster_head
= NULL
;
997 nbio
->bio_caller_info2
.cluster_tail
= NULL
;
1003 * Execute read or write
1005 vm_object_hold(object
);
1011 * Obtain block. If block not found and writing, allocate a
1012 * new block and build it into the object.
1014 blk
= swp_pager_meta_ctl(object
, start
, 0);
1015 if ((blk
== SWAPBLK_NONE
) && bp
->b_cmd
!= BUF_CMD_READ
) {
1016 blk
= swp_pager_getswapspace(object
, 1);
1017 if (blk
== SWAPBLK_NONE
) {
1018 bp
->b_error
= ENOMEM
;
1019 bp
->b_flags
|= B_ERROR
;
1022 swp_pager_meta_build(object
, start
, blk
);
1026 * Do we have to flush our current collection? Yes if:
1028 * - no swap block at this index
1029 * - swap block is not contiguous
1030 * - we cross a physical disk boundry in the
1034 biox
&& (biox_blkno
+ btoc(bufx
->b_bcount
) != blk
||
1035 ((biox_blkno
^ blk
) & dmmax_mask
)
1038 if (bp
->b_cmd
== BUF_CMD_READ
) {
1039 ++mycpu
->gd_cnt
.v_swapin
;
1040 mycpu
->gd_cnt
.v_swappgsin
+= btoc(bufx
->b_bcount
);
1042 ++mycpu
->gd_cnt
.v_swapout
;
1043 mycpu
->gd_cnt
.v_swappgsout
+= btoc(bufx
->b_bcount
);
1044 bufx
->b_dirtyend
= bufx
->b_bcount
;
1048 * Finished with this buf.
1050 KKASSERT(bufx
->b_bcount
!= 0);
1051 if (bufx
->b_cmd
!= BUF_CMD_READ
)
1052 bufx
->b_dirtyend
= bufx
->b_bcount
;
1058 * Add new swapblk to biox, instantiating biox if necessary.
1059 * Zero-fill reads are able to take a shortcut.
1061 if (blk
== SWAPBLK_NONE
) {
1063 * We can only get here if we are reading. Since
1064 * we are at splvm() we can safely modify b_resid,
1065 * even if chain ops are in progress.
1067 bzero(data
, PAGE_SIZE
);
1068 bp
->b_resid
-= PAGE_SIZE
;
1071 /* XXX chain count > 4, wait to <= 4 */
1073 bufx
= getpbuf(NULL
);
1074 biox
= &bufx
->b_bio1
;
1075 cluster_append(nbio
, bufx
);
1076 bufx
->b_flags
|= (bp
->b_flags
& B_ORDERED
);
1077 bufx
->b_cmd
= bp
->b_cmd
;
1078 biox
->bio_done
= swap_chain_iodone
;
1079 biox
->bio_offset
= (off_t
)blk
<< PAGE_SHIFT
;
1080 biox
->bio_caller_info1
.cluster_parent
= nbio
;
1083 bufx
->b_data
= data
;
1085 bufx
->b_bcount
+= PAGE_SIZE
;
1092 vm_object_drop(object
);
1095 * Flush out last buffer
1098 if (bufx
->b_cmd
== BUF_CMD_READ
) {
1099 ++mycpu
->gd_cnt
.v_swapin
;
1100 mycpu
->gd_cnt
.v_swappgsin
+= btoc(bufx
->b_bcount
);
1102 ++mycpu
->gd_cnt
.v_swapout
;
1103 mycpu
->gd_cnt
.v_swappgsout
+= btoc(bufx
->b_bcount
);
1104 bufx
->b_dirtyend
= bufx
->b_bcount
;
1106 KKASSERT(bufx
->b_bcount
);
1107 if (bufx
->b_cmd
!= BUF_CMD_READ
)
1108 bufx
->b_dirtyend
= bufx
->b_bcount
;
1109 /* biox, bufx = NULL */
1113 * Now initiate all the I/O. Be careful looping on our chain as
1114 * I/O's may complete while we are still initiating them.
1116 * If the request is a 100% sparse read no bios will be present
1117 * and we just biodone() the buffer.
1119 nbio
->bio_caller_info2
.cluster_tail
= NULL
;
1120 bufx
= nbio
->bio_caller_info1
.cluster_head
;
1124 biox
= &bufx
->b_bio1
;
1126 bufx
= bufx
->b_cluster_next
;
1127 vn_strategy(swapdev_vp
, biox
);
1134 * Completion of the cluster will also call biodone_chain(nbio).
1135 * We never call biodone(nbio) so we don't have to worry about
1136 * setting up a bio_done callback. It's handled in the sub-IO.
1147 swap_chain_iodone(struct bio
*biox
)
1150 struct buf
*bufx
; /* chained sub-buffer */
1151 struct bio
*nbio
; /* parent nbio with chain glue */
1152 struct buf
*bp
; /* original bp associated with nbio */
1155 bufx
= biox
->bio_buf
;
1156 nbio
= biox
->bio_caller_info1
.cluster_parent
;
1160 * Update the original buffer
1162 KKASSERT(bp
!= NULL
);
1163 if (bufx
->b_flags
& B_ERROR
) {
1164 atomic_set_int(&bufx
->b_flags
, B_ERROR
);
1165 bp
->b_error
= bufx
->b_error
; /* race ok */
1166 } else if (bufx
->b_resid
!= 0) {
1167 atomic_set_int(&bufx
->b_flags
, B_ERROR
);
1168 bp
->b_error
= EINVAL
; /* race ok */
1170 atomic_subtract_int(&bp
->b_resid
, bufx
->b_bcount
);
1174 * Remove us from the chain.
1176 spin_lock(&swapbp_spin
);
1177 nextp
= &nbio
->bio_caller_info1
.cluster_head
;
1178 while (*nextp
!= bufx
) {
1179 KKASSERT(*nextp
!= NULL
);
1180 nextp
= &(*nextp
)->b_cluster_next
;
1182 *nextp
= bufx
->b_cluster_next
;
1183 chain_empty
= (nbio
->bio_caller_info1
.cluster_head
== NULL
);
1184 spin_unlock(&swapbp_spin
);
1187 * Clean up bufx. If the chain is now empty we finish out
1188 * the parent. Note that we may be racing other completions
1189 * so we must use the chain_empty status from above.
1192 if (bp
->b_resid
!= 0 && !(bp
->b_flags
& B_ERROR
)) {
1193 atomic_set_int(&bp
->b_flags
, B_ERROR
);
1194 bp
->b_error
= EINVAL
;
1196 biodone_chain(nbio
);
1198 relpbuf(bufx
, NULL
);
1202 * SWAP_PAGER_GETPAGES() - bring page in from swap
1204 * The requested page may have to be brought in from swap. Calculate the
1205 * swap block and bring in additional pages if possible. All pages must
1206 * have contiguous swap block assignments and reside in the same object.
1208 * The caller has a single vm_object_pip_add() reference prior to
1209 * calling us and we should return with the same.
1211 * The caller has BUSY'd the page. We should return with (*mpp) left busy,
1212 * and any additinal pages unbusied.
1214 * If the caller encounters a PG_RAM page it will pass it to us even though
1215 * it may be valid and dirty. We cannot overwrite the page in this case!
1216 * The case is used to allow us to issue pure read-aheads.
1218 * NOTE! XXX This code does not entirely pipeline yet due to the fact that
1219 * the PG_RAM page is validated at the same time as mreq. What we
1220 * really need to do is issue a separate read-ahead pbuf.
1225 swap_pager_getpage(vm_object_t object
, vm_page_t
*mpp
, int seqaccess
)
1238 vm_page_t marray
[XIO_INTERNAL_PAGES
];
1242 vm_object_hold(object
);
1243 if (mreq
->object
!= object
) {
1244 panic("swap_pager_getpages: object mismatch %p/%p",
1251 * We don't want to overwrite a fully valid page as it might be
1252 * dirty. This case can occur when e.g. vm_fault hits a perfectly
1253 * valid page with PG_RAM set.
1255 * In this case we see if the next page is a suitable page-in
1256 * candidate and if it is we issue read-ahead. PG_RAM will be
1257 * set on the last page of the read-ahead to continue the pipeline.
1259 if (mreq
->valid
== VM_PAGE_BITS_ALL
) {
1260 if (swap_burst_read
== 0 || mreq
->pindex
+ 1 >= object
->size
) {
1261 vm_object_drop(object
);
1262 return(VM_PAGER_OK
);
1264 blk
= swp_pager_meta_ctl(object
, mreq
->pindex
+ 1, 0);
1265 if (blk
== SWAPBLK_NONE
) {
1266 vm_object_drop(object
);
1267 return(VM_PAGER_OK
);
1269 m
= vm_page_lookup_busy_try(object
, mreq
->pindex
+ 1,
1272 vm_object_drop(object
);
1273 return(VM_PAGER_OK
);
1274 } else if (m
== NULL
) {
1276 * Use VM_ALLOC_QUICK to avoid blocking on cache
1279 m
= vm_page_alloc(object
, mreq
->pindex
+ 1,
1282 vm_object_drop(object
);
1283 return(VM_PAGER_OK
);
1288 vm_object_drop(object
);
1289 return(VM_PAGER_OK
);
1291 vm_page_unqueue_nowakeup(m
);
1301 * Try to block-read contiguous pages from swap if sequential,
1302 * otherwise just read one page. Contiguous pages from swap must
1303 * reside within a single device stripe because the I/O cannot be
1304 * broken up across multiple stripes.
1306 * Note that blk and iblk can be SWAPBLK_NONE but the loop is
1307 * set up such that the case(s) are handled implicitly.
1309 blk
= swp_pager_meta_ctl(mreq
->object
, mreq
->pindex
, 0);
1312 for (i
= 1; i
<= swap_burst_read
&&
1313 i
< XIO_INTERNAL_PAGES
&&
1314 mreq
->pindex
+ i
< object
->size
; ++i
) {
1317 iblk
= swp_pager_meta_ctl(object
, mreq
->pindex
+ i
, 0);
1318 if (iblk
!= blk
+ i
)
1320 if ((blk
^ iblk
) & dmmax_mask
)
1322 m
= vm_page_lookup_busy_try(object
, mreq
->pindex
+ i
,
1326 } else if (m
== NULL
) {
1328 * Use VM_ALLOC_QUICK to avoid blocking on cache
1331 m
= vm_page_alloc(object
, mreq
->pindex
+ i
,
1340 vm_page_unqueue_nowakeup(m
);
1346 vm_page_flag_set(marray
[i
- 1], PG_RAM
);
1349 * If mreq is the requested page and we have nothing to do return
1350 * VM_PAGER_FAIL. If raonly is set mreq is just another read-ahead
1351 * page and must be cleaned up.
1353 if (blk
== SWAPBLK_NONE
) {
1356 vnode_pager_freepage(mreq
);
1357 vm_object_drop(object
);
1358 return(VM_PAGER_OK
);
1360 vm_object_drop(object
);
1361 return(VM_PAGER_FAIL
);
1366 * map our page(s) into kva for input
1368 bp
= getpbuf_kva(&nsw_rcount
);
1370 kva
= (vm_offset_t
) bp
->b_kvabase
;
1371 bcopy(marray
, bp
->b_xio
.xio_pages
, i
* sizeof(vm_page_t
));
1372 pmap_qenter(kva
, bp
->b_xio
.xio_pages
, i
);
1374 bp
->b_data
= (caddr_t
)kva
;
1375 bp
->b_bcount
= PAGE_SIZE
* i
;
1376 bp
->b_xio
.xio_npages
= i
;
1377 bio
->bio_done
= swp_pager_async_iodone
;
1378 bio
->bio_offset
= (off_t
)blk
<< PAGE_SHIFT
;
1379 bio
->bio_caller_info1
.index
= SWBIO_READ
;
1382 * Set index. If raonly set the index beyond the array so all
1383 * the pages are treated the same, otherwise the original mreq is
1387 bio
->bio_driver_info
= (void *)(intptr_t)i
;
1389 bio
->bio_driver_info
= (void *)(intptr_t)0;
1391 for (j
= 0; j
< i
; ++j
)
1392 vm_page_flag_set(bp
->b_xio
.xio_pages
[j
], PG_SWAPINPROG
);
1394 mycpu
->gd_cnt
.v_swapin
++;
1395 mycpu
->gd_cnt
.v_swappgsin
+= bp
->b_xio
.xio_npages
;
1398 * We still hold the lock on mreq, and our automatic completion routine
1399 * does not remove it.
1401 vm_object_pip_add(object
, bp
->b_xio
.xio_npages
);
1404 * perform the I/O. NOTE!!! bp cannot be considered valid after
1405 * this point because we automatically release it on completion.
1406 * Instead, we look at the one page we are interested in which we
1407 * still hold a lock on even through the I/O completion.
1409 * The other pages in our m[] array are also released on completion,
1410 * so we cannot assume they are valid anymore either.
1412 bp
->b_cmd
= BUF_CMD_READ
;
1414 vn_strategy(swapdev_vp
, bio
);
1417 * Wait for the page we want to complete. PG_SWAPINPROG is always
1418 * cleared on completion. If an I/O error occurs, SWAPBLK_NONE
1419 * is set in the meta-data.
1421 * If this is a read-ahead only we return immediately without
1425 vm_object_drop(object
);
1426 return(VM_PAGER_OK
);
1430 * Read-ahead includes originally requested page case.
1433 flags
= mreq
->flags
;
1435 if ((flags
& PG_SWAPINPROG
) == 0)
1437 tsleep_interlock(mreq
, 0);
1438 if (!atomic_cmpset_int(&mreq
->flags
, flags
,
1439 flags
| PG_WANTED
| PG_REFERENCED
)) {
1442 mycpu
->gd_cnt
.v_intrans
++;
1443 if (tsleep(mreq
, PINTERLOCKED
, "swread", hz
*20)) {
1445 "swap_pager: indefinite wait buffer: "
1446 " bp %p offset: %lld, size: %ld\n",
1448 (long long)bio
->bio_offset
,
1455 * mreq is left bussied after completion, but all the other pages
1456 * are freed. If we had an unrecoverable read error the page will
1459 vm_object_drop(object
);
1460 if (mreq
->valid
!= VM_PAGE_BITS_ALL
)
1461 return(VM_PAGER_ERROR
);
1463 return(VM_PAGER_OK
);
1466 * A final note: in a low swap situation, we cannot deallocate swap
1467 * and mark a page dirty here because the caller is likely to mark
1468 * the page clean when we return, causing the page to possibly revert
1469 * to all-zero's later.
1474 * swap_pager_putpages:
1476 * Assign swap (if necessary) and initiate I/O on the specified pages.
1478 * We support both OBJT_DEFAULT and OBJT_SWAP objects. DEFAULT objects
1479 * are automatically converted to SWAP objects.
1481 * In a low memory situation we may block in vn_strategy(), but the new
1482 * vm_page reservation system coupled with properly written VFS devices
1483 * should ensure that no low-memory deadlock occurs. This is an area
1486 * The parent has N vm_object_pip_add() references prior to
1487 * calling us and will remove references for rtvals[] that are
1488 * not set to VM_PAGER_PEND. We need to remove the rest on I/O
1491 * The parent has soft-busy'd the pages it passes us and will unbusy
1492 * those whos rtvals[] entry is not set to VM_PAGER_PEND on return.
1493 * We need to unbusy the rest on I/O completion.
1498 swap_pager_putpages(vm_object_t object
, vm_page_t
*m
, int count
,
1499 int flags
, int *rtvals
)
1504 vm_object_hold(object
);
1506 if (count
&& m
[0]->object
!= object
) {
1507 panic("swap_pager_getpages: object mismatch %p/%p",
1516 * Turn object into OBJT_SWAP
1517 * Check for bogus sysops
1519 * Force sync if not pageout process, we don't want any single
1520 * non-pageout process to be able to hog the I/O subsystem! This
1521 * can be overridden by setting.
1523 if (object
->type
== OBJT_DEFAULT
) {
1524 if (object
->type
== OBJT_DEFAULT
)
1525 swp_pager_meta_convert(object
);
1529 * Normally we force synchronous swap I/O if this is not the
1530 * pageout daemon to prevent any single user process limited
1531 * via RLIMIT_RSS from hogging swap write bandwidth.
1533 if (curthread
!= pagethread
&& swap_user_async
== 0)
1534 flags
|= VM_PAGER_PUT_SYNC
;
1539 * Update nsw parameters from swap_async_max sysctl values.
1540 * Do not let the sysop crash the machine with bogus numbers.
1542 if (swap_async_max
!= nsw_wcount_async_max
) {
1548 if ((n
= swap_async_max
) > nswbuf_kva
/ 2)
1555 * Adjust difference ( if possible ). If the current async
1556 * count is too low, we may not be able to make the adjustment
1559 * vm_token needed for nsw_wcount sleep interlock
1561 lwkt_gettoken(&vm_token
);
1562 n
-= nsw_wcount_async_max
;
1563 if (nsw_wcount_async
+ n
>= 0) {
1564 nsw_wcount_async_max
+= n
;
1565 pbuf_adjcount(&nsw_wcount_async
, n
);
1567 lwkt_reltoken(&vm_token
);
1573 * Assign swap blocks and issue I/O. We reallocate swap on the fly.
1574 * The page is left dirty until the pageout operation completes
1578 for (i
= 0; i
< count
; i
+= n
) {
1585 * Maximum I/O size is limited by a number of factors.
1588 n
= min(BLIST_MAX_ALLOC
, count
- i
);
1589 n
= min(n
, nsw_cluster_max
);
1591 lwkt_gettoken(&vm_token
);
1594 * Get biggest block of swap we can. If we fail, fall
1595 * back and try to allocate a smaller block. Don't go
1596 * overboard trying to allocate space if it would overly
1600 (blk
= swp_pager_getswapspace(object
, n
)) == SWAPBLK_NONE
&&
1605 if (blk
== SWAPBLK_NONE
) {
1606 for (j
= 0; j
< n
; ++j
)
1607 rtvals
[i
+j
] = VM_PAGER_FAIL
;
1608 lwkt_reltoken(&vm_token
);
1611 if (vm_report_swap_allocs
> 0) {
1612 kprintf("swap_alloc %08jx,%d\n", (intmax_t)blk
, n
);
1613 --vm_report_swap_allocs
;
1617 * The I/O we are constructing cannot cross a physical
1618 * disk boundry in the swap stripe. Note: we are still
1621 if ((blk
^ (blk
+ n
)) & dmmax_mask
) {
1622 j
= ((blk
+ dmmax
) & dmmax_mask
) - blk
;
1623 swp_pager_freeswapspace(object
, blk
+ j
, n
- j
);
1628 * All I/O parameters have been satisfied, build the I/O
1629 * request and assign the swap space.
1631 if ((flags
& VM_PAGER_PUT_SYNC
))
1632 bp
= getpbuf_kva(&nsw_wcount_sync
);
1634 bp
= getpbuf_kva(&nsw_wcount_async
);
1637 lwkt_reltoken(&vm_token
);
1639 pmap_qenter((vm_offset_t
)bp
->b_data
, &m
[i
], n
);
1641 bp
->b_bcount
= PAGE_SIZE
* n
;
1642 bio
->bio_offset
= (off_t
)blk
<< PAGE_SHIFT
;
1644 for (j
= 0; j
< n
; ++j
) {
1645 vm_page_t mreq
= m
[i
+j
];
1647 swp_pager_meta_build(mreq
->object
, mreq
->pindex
,
1649 if (object
->type
== OBJT_SWAP
)
1650 vm_page_dirty(mreq
);
1651 rtvals
[i
+j
] = VM_PAGER_OK
;
1653 vm_page_flag_set(mreq
, PG_SWAPINPROG
);
1654 bp
->b_xio
.xio_pages
[j
] = mreq
;
1656 bp
->b_xio
.xio_npages
= n
;
1658 mycpu
->gd_cnt
.v_swapout
++;
1659 mycpu
->gd_cnt
.v_swappgsout
+= bp
->b_xio
.xio_npages
;
1661 bp
->b_dirtyoff
= 0; /* req'd for NFS */
1662 bp
->b_dirtyend
= bp
->b_bcount
; /* req'd for NFS */
1663 bp
->b_cmd
= BUF_CMD_WRITE
;
1664 bio
->bio_caller_info1
.index
= SWBIO_WRITE
;
1669 if ((flags
& VM_PAGER_PUT_SYNC
) == 0) {
1670 bio
->bio_done
= swp_pager_async_iodone
;
1672 vn_strategy(swapdev_vp
, bio
);
1674 for (j
= 0; j
< n
; ++j
)
1675 rtvals
[i
+j
] = VM_PAGER_PEND
;
1680 * Issue synchrnously.
1682 * Wait for the sync I/O to complete, then update rtvals.
1683 * We just set the rtvals[] to VM_PAGER_PEND so we can call
1684 * our async completion routine at the end, thus avoiding a
1687 bio
->bio_caller_info1
.index
|= SWBIO_SYNC
;
1688 if (flags
& VM_PAGER_TRY_TO_CACHE
)
1689 bio
->bio_caller_info1
.index
|= SWBIO_TTC
;
1690 bio
->bio_done
= biodone_sync
;
1691 bio
->bio_flags
|= BIO_SYNC
;
1692 vn_strategy(swapdev_vp
, bio
);
1693 biowait(bio
, "swwrt");
1695 for (j
= 0; j
< n
; ++j
)
1696 rtvals
[i
+j
] = VM_PAGER_PEND
;
1699 * Now that we are through with the bp, we can call the
1700 * normal async completion, which frees everything up.
1702 swp_pager_async_iodone(bio
);
1704 vm_object_drop(object
);
1710 * Recalculate the low and high-water marks.
1713 swap_pager_newswap(void)
1716 * NOTE: vm_swap_max cannot exceed 1 billion blocks, which is the
1717 * limitation imposed by the blist code. Remember that this
1718 * will be divided by NSWAP_MAX (4), so each swap device is
1719 * limited to around a terrabyte.
1722 nswap_lowat
= (int64_t)vm_swap_max
* 4 / 100; /* 4% left */
1723 nswap_hiwat
= (int64_t)vm_swap_max
* 6 / 100; /* 6% left */
1724 kprintf("swap low/high-water marks set to %d/%d\n",
1725 nswap_lowat
, nswap_hiwat
);
1734 * swp_pager_async_iodone:
1736 * Completion routine for asynchronous reads and writes from/to swap.
1737 * Also called manually by synchronous code to finish up a bp.
1739 * For READ operations, the pages are PG_BUSY'd. For WRITE operations,
1740 * the pages are vm_page_t->busy'd. For READ operations, we PG_BUSY
1741 * unbusy all pages except the 'main' request page. For WRITE
1742 * operations, we vm_page_t->busy'd unbusy all pages ( we can do this
1743 * because we marked them all VM_PAGER_PEND on return from putpages ).
1745 * This routine may not block.
1750 swp_pager_async_iodone(struct bio
*bio
)
1752 struct buf
*bp
= bio
->bio_buf
;
1753 vm_object_t object
= NULL
;
1760 if (bp
->b_flags
& B_ERROR
) {
1762 "swap_pager: I/O error - %s failed; offset %lld,"
1763 "size %ld, error %d\n",
1764 ((bio
->bio_caller_info1
.index
& SWBIO_READ
) ?
1765 "pagein" : "pageout"),
1766 (long long)bio
->bio_offset
,
1773 * set object, raise to splvm().
1775 if (bp
->b_xio
.xio_npages
)
1776 object
= bp
->b_xio
.xio_pages
[0]->object
;
1779 * remove the mapping for kernel virtual
1781 pmap_qremove((vm_offset_t
)bp
->b_data
, bp
->b_xio
.xio_npages
);
1784 * cleanup pages. If an error occurs writing to swap, we are in
1785 * very serious trouble. If it happens to be a disk error, though,
1786 * we may be able to recover by reassigning the swap later on. So
1787 * in this case we remove the m->swapblk assignment for the page
1788 * but do not free it in the rlist. The errornous block(s) are thus
1789 * never reallocated as swap. Redirty the page and continue.
1791 for (i
= 0; i
< bp
->b_xio
.xio_npages
; ++i
) {
1792 vm_page_t m
= bp
->b_xio
.xio_pages
[i
];
1794 if (bp
->b_flags
& B_ERROR
) {
1796 * If an error occurs I'd love to throw the swapblk
1797 * away without freeing it back to swapspace, so it
1798 * can never be used again. But I can't from an
1802 if (bio
->bio_caller_info1
.index
& SWBIO_READ
) {
1804 * When reading, reqpage needs to stay
1805 * locked for the parent, but all other
1806 * pages can be freed. We still want to
1807 * wakeup the parent waiting on the page,
1808 * though. ( also: pg_reqpage can be -1 and
1809 * not match anything ).
1811 * We have to wake specifically requested pages
1812 * up too because we cleared PG_SWAPINPROG and
1813 * someone may be waiting for that.
1815 * NOTE: for reads, m->dirty will probably
1816 * be overridden by the original caller of
1817 * getpages so don't play cute tricks here.
1819 * NOTE: We can't actually free the page from
1820 * here, because this is an interrupt. It
1821 * is not legal to mess with object->memq
1822 * from an interrupt. Deactivate the page
1827 vm_page_flag_clear(m
, PG_SWAPINPROG
);
1830 * bio_driver_info holds the requested page
1833 if (i
!= (int)(intptr_t)bio
->bio_driver_info
) {
1834 vm_page_deactivate(m
);
1840 * If i == bp->b_pager.pg_reqpage, do not wake
1841 * the page up. The caller needs to.
1845 * If a write error occurs remove the swap
1846 * assignment (note that PG_SWAPPED may or
1847 * may not be set depending on prior activity).
1849 * Re-dirty OBJT_SWAP pages as there is no
1850 * other backing store, we can't throw the
1853 * Non-OBJT_SWAP pages (aka swapcache) must
1854 * not be dirtied since they may not have
1855 * been dirty in the first place, and they
1856 * do have backing store (the vnode).
1858 vm_page_busy_wait(m
, FALSE
, "swadpg");
1859 swp_pager_meta_ctl(m
->object
, m
->pindex
,
1861 vm_page_flag_clear(m
, PG_SWAPPED
);
1862 if (m
->object
->type
== OBJT_SWAP
) {
1864 vm_page_activate(m
);
1866 vm_page_flag_clear(m
, PG_SWAPINPROG
);
1867 vm_page_io_finish(m
);
1870 } else if (bio
->bio_caller_info1
.index
& SWBIO_READ
) {
1872 * NOTE: for reads, m->dirty will probably be
1873 * overridden by the original caller of getpages so
1874 * we cannot set them in order to free the underlying
1875 * swap in a low-swap situation. I don't think we'd
1876 * want to do that anyway, but it was an optimization
1877 * that existed in the old swapper for a time before
1878 * it got ripped out due to precisely this problem.
1880 * If not the requested page then deactivate it.
1882 * Note that the requested page, reqpage, is left
1883 * busied, but we still have to wake it up. The
1884 * other pages are released (unbusied) by
1885 * vm_page_wakeup(). We do not set reqpage's
1886 * valid bits here, it is up to the caller.
1890 * NOTE: can't call pmap_clear_modify(m) from an
1891 * interrupt thread, the pmap code may have to map
1892 * non-kernel pmaps and currently asserts the case.
1894 /*pmap_clear_modify(m);*/
1895 m
->valid
= VM_PAGE_BITS_ALL
;
1897 vm_page_flag_clear(m
, PG_SWAPINPROG
);
1898 vm_page_flag_set(m
, PG_SWAPPED
);
1901 * We have to wake specifically requested pages
1902 * up too because we cleared PG_SWAPINPROG and
1903 * could be waiting for it in getpages. However,
1904 * be sure to not unbusy getpages specifically
1905 * requested page - getpages expects it to be
1908 * bio_driver_info holds the requested page
1910 if (i
!= (int)(intptr_t)bio
->bio_driver_info
) {
1911 vm_page_deactivate(m
);
1918 * Mark the page clean but do not mess with the
1919 * pmap-layer's modified state. That state should
1920 * also be clear since the caller protected the
1921 * page VM_PROT_READ, but allow the case.
1923 * We are in an interrupt, avoid pmap operations.
1925 * If we have a severe page deficit, deactivate the
1926 * page. Do not try to cache it (which would also
1927 * involve a pmap op), because the page might still
1930 * When using the swap to cache clean vnode pages
1931 * we do not mess with the page dirty bits.
1933 vm_page_busy_wait(m
, FALSE
, "swadpg");
1934 if (m
->object
->type
== OBJT_SWAP
)
1936 vm_page_flag_clear(m
, PG_SWAPINPROG
);
1937 vm_page_flag_set(m
, PG_SWAPPED
);
1938 if (vm_page_count_severe())
1939 vm_page_deactivate(m
);
1940 vm_page_io_finish(m
);
1942 if (bio
->bio_caller_info1
.index
& SWBIO_TTC
)
1943 vm_page_try_to_cache(m
);
1948 * adjust pip. NOTE: the original parent may still have its own
1949 * pip refs on the object.
1953 vm_object_pip_wakeup_n(object
, bp
->b_xio
.xio_npages
);
1956 * Release the physical I/O buffer.
1958 * NOTE: Due to synchronous operations in the write case b_cmd may
1959 * already be set to BUF_CMD_DONE and BIO_SYNC may have already
1962 * Use vm_token to interlock nsw_rcount/wcount wakeup?
1964 lwkt_gettoken(&vm_token
);
1965 if (bio
->bio_caller_info1
.index
& SWBIO_READ
)
1966 nswptr
= &nsw_rcount
;
1967 else if (bio
->bio_caller_info1
.index
& SWBIO_SYNC
)
1968 nswptr
= &nsw_wcount_sync
;
1970 nswptr
= &nsw_wcount_async
;
1971 bp
->b_cmd
= BUF_CMD_DONE
;
1972 relpbuf(bp
, nswptr
);
1973 lwkt_reltoken(&vm_token
);
1977 * Fault-in a potentially swapped page and remove the swap reference.
1978 * (used by swapoff code)
1980 * object must be held.
1982 static __inline
void
1983 swp_pager_fault_page(vm_object_t object
, int *sharedp
, vm_pindex_t pindex
)
1989 ASSERT_LWKT_TOKEN_HELD(vm_object_token(object
));
1991 if (object
->type
== OBJT_VNODE
) {
1993 * Any swap related to a vnode is due to swapcache. We must
1994 * vget() the vnode in case it is not active (otherwise
1995 * vref() will panic). Calling vm_object_page_remove() will
1996 * ensure that any swap ref is removed interlocked with the
1997 * page. clean_only is set to TRUE so we don't throw away
2000 vp
= object
->handle
;
2001 error
= vget(vp
, LK_SHARED
| LK_RETRY
| LK_CANRECURSE
);
2003 vm_object_page_remove(object
, pindex
, pindex
+ 1, TRUE
);
2008 * Otherwise it is a normal OBJT_SWAP object and we can
2009 * fault the page in and remove the swap.
2011 m
= vm_fault_object_page(object
, IDX_TO_OFF(pindex
),
2013 VM_FAULT_DIRTY
| VM_FAULT_UNSWAP
,
2021 * This removes all swap blocks related to a particular device. We have
2022 * to be careful of ripups during the scan.
2024 static int swp_pager_swapoff_callback(struct swblock
*swap
, void *data
);
2027 swap_pager_swapoff(int devidx
)
2029 struct swswapoffinfo info
;
2030 struct vm_object marker
;
2034 bzero(&marker
, sizeof(marker
));
2035 marker
.type
= OBJT_MARKER
;
2037 for (n
= 0; n
< VMOBJ_HSIZE
; ++n
) {
2038 lwkt_gettoken(&vmobj_tokens
[n
]);
2039 TAILQ_INSERT_HEAD(&vm_object_lists
[n
], &marker
, object_list
);
2041 while ((object
= TAILQ_NEXT(&marker
, object_list
)) != NULL
) {
2042 if (object
->type
== OBJT_MARKER
)
2044 if (object
->type
!= OBJT_SWAP
&&
2045 object
->type
!= OBJT_VNODE
)
2047 vm_object_hold(object
);
2048 if (object
->type
!= OBJT_SWAP
&&
2049 object
->type
!= OBJT_VNODE
) {
2050 vm_object_drop(object
);
2053 info
.object
= object
;
2055 info
.devidx
= devidx
;
2056 swblock_rb_tree_RB_SCAN(&object
->swblock_root
,
2057 NULL
, swp_pager_swapoff_callback
,
2059 vm_object_drop(object
);
2061 if (object
== TAILQ_NEXT(&marker
, object_list
)) {
2062 TAILQ_REMOVE(&vm_object_lists
[n
],
2063 &marker
, object_list
);
2064 TAILQ_INSERT_AFTER(&vm_object_lists
[n
], object
,
2065 &marker
, object_list
);
2068 TAILQ_REMOVE(&vm_object_lists
[n
], &marker
, object_list
);
2069 lwkt_reltoken(&vmobj_tokens
[n
]);
2073 * If we fail to locate all swblocks we just fail gracefully and
2074 * do not bother to restore paging on the swap device. If the
2075 * user wants to retry the user can retry.
2077 if (swdevt
[devidx
].sw_nused
)
2085 swp_pager_swapoff_callback(struct swblock
*swap
, void *data
)
2087 struct swswapoffinfo
*info
= data
;
2088 vm_object_t object
= info
->object
;
2093 index
= swap
->swb_index
;
2094 for (i
= 0; i
< SWAP_META_PAGES
; ++i
) {
2096 * Make sure we don't race a dying object. This will
2097 * kill the scan of the object's swap blocks entirely.
2099 if (object
->flags
& OBJ_DEAD
)
2103 * Fault the page, which can obviously block. If the swap
2104 * structure disappears break out.
2106 v
= swap
->swb_pages
[i
];
2107 if (v
!= SWAPBLK_NONE
&& BLK2DEVIDX(v
) == info
->devidx
) {
2108 swp_pager_fault_page(object
, &info
->shared
,
2109 swap
->swb_index
+ i
);
2110 /* swap ptr might go away */
2111 if (RB_LOOKUP(swblock_rb_tree
,
2112 &object
->swblock_root
, index
) != swap
) {
2120 /************************************************************************
2122 ************************************************************************
2124 * These routines manipulate the swap metadata stored in the
2125 * OBJT_SWAP object. All swp_*() routines must be called at
2126 * splvm() because swap can be freed up by the low level vm_page
2127 * code which might be called from interrupts beyond what splbio() covers.
2129 * Swap metadata is implemented with a global hash and not directly
2130 * linked into the object. Instead the object simply contains
2131 * appropriate tracking counters.
2135 * Lookup the swblock containing the specified swap block index.
2137 * The caller must hold the object.
2141 swp_pager_lookup(vm_object_t object
, vm_pindex_t index
)
2143 ASSERT_LWKT_TOKEN_HELD(vm_object_token(object
));
2144 index
&= ~(vm_pindex_t
)SWAP_META_MASK
;
2145 return (RB_LOOKUP(swblock_rb_tree
, &object
->swblock_root
, index
));
2149 * Remove a swblock from the RB tree.
2151 * The caller must hold the object.
2155 swp_pager_remove(vm_object_t object
, struct swblock
*swap
)
2157 ASSERT_LWKT_TOKEN_HELD(vm_object_token(object
));
2158 RB_REMOVE(swblock_rb_tree
, &object
->swblock_root
, swap
);
2162 * Convert default object to swap object if necessary
2164 * The caller must hold the object.
2167 swp_pager_meta_convert(vm_object_t object
)
2169 if (object
->type
== OBJT_DEFAULT
) {
2170 object
->type
= OBJT_SWAP
;
2171 KKASSERT(object
->swblock_count
== 0);
2176 * SWP_PAGER_META_BUILD() - add swap block to swap meta data for object
2178 * We first convert the object to a swap object if it is a default
2179 * object. Vnode objects do not need to be converted.
2181 * The specified swapblk is added to the object's swap metadata. If
2182 * the swapblk is not valid, it is freed instead. Any previously
2183 * assigned swapblk is freed.
2185 * The caller must hold the object.
2188 swp_pager_meta_build(vm_object_t object
, vm_pindex_t index
, swblk_t swapblk
)
2190 struct swblock
*swap
;
2191 struct swblock
*oswap
;
2194 KKASSERT(swapblk
!= SWAPBLK_NONE
);
2195 ASSERT_LWKT_TOKEN_HELD(vm_object_token(object
));
2198 * Convert object if necessary
2200 if (object
->type
== OBJT_DEFAULT
)
2201 swp_pager_meta_convert(object
);
2204 * Locate swblock. If not found create, but if we aren't adding
2205 * anything just return. If we run out of space in the map we wait
2206 * and, since the hash table may have changed, retry.
2209 swap
= swp_pager_lookup(object
, index
);
2214 swap
= zalloc(swap_zone
);
2219 swap
->swb_index
= index
& ~(vm_pindex_t
)SWAP_META_MASK
;
2220 swap
->swb_count
= 0;
2222 ++object
->swblock_count
;
2224 for (i
= 0; i
< SWAP_META_PAGES
; ++i
)
2225 swap
->swb_pages
[i
] = SWAPBLK_NONE
;
2226 oswap
= RB_INSERT(swblock_rb_tree
, &object
->swblock_root
, swap
);
2227 KKASSERT(oswap
== NULL
);
2231 * Delete prior contents of metadata.
2233 * NOTE: Decrement swb_count after the freeing operation (which
2234 * might block) to prevent racing destruction of the swblock.
2236 index
&= SWAP_META_MASK
;
2238 while ((v
= swap
->swb_pages
[index
]) != SWAPBLK_NONE
) {
2239 swap
->swb_pages
[index
] = SWAPBLK_NONE
;
2241 swp_pager_freeswapspace(object
, v
, 1);
2243 --mycpu
->gd_vmtotal
.t_vm
;
2247 * Enter block into metadata
2249 swap
->swb_pages
[index
] = swapblk
;
2250 if (swapblk
!= SWAPBLK_NONE
) {
2252 ++mycpu
->gd_vmtotal
.t_vm
;
2257 * SWP_PAGER_META_FREE() - free a range of blocks in the object's swap metadata
2259 * The requested range of blocks is freed, with any associated swap
2260 * returned to the swap bitmap.
2262 * This routine will free swap metadata structures as they are cleaned
2263 * out. This routine does *NOT* operate on swap metadata associated
2264 * with resident pages.
2266 * The caller must hold the object.
2268 static int swp_pager_meta_free_callback(struct swblock
*swb
, void *data
);
2271 swp_pager_meta_free(vm_object_t object
, vm_pindex_t index
, vm_pindex_t count
)
2273 struct swfreeinfo info
;
2275 ASSERT_LWKT_TOKEN_HELD(vm_object_token(object
));
2280 if (object
->swblock_count
== 0) {
2281 KKASSERT(RB_EMPTY(&object
->swblock_root
));
2288 * Setup for RB tree scan. Note that the pindex range can be huge
2289 * due to the 64 bit page index space so we cannot safely iterate.
2291 info
.object
= object
;
2292 info
.basei
= index
& ~(vm_pindex_t
)SWAP_META_MASK
;
2294 info
.endi
= index
+ count
- 1;
2295 swblock_rb_tree_RB_SCAN(&object
->swblock_root
, rb_swblock_scancmp
,
2296 swp_pager_meta_free_callback
, &info
);
2300 * The caller must hold the object.
2304 swp_pager_meta_free_callback(struct swblock
*swap
, void *data
)
2306 struct swfreeinfo
*info
= data
;
2307 vm_object_t object
= info
->object
;
2312 * Figure out the range within the swblock. The wider scan may
2313 * return edge-case swap blocks when the start and/or end points
2314 * are in the middle of a block.
2316 if (swap
->swb_index
< info
->begi
)
2317 index
= (int)info
->begi
& SWAP_META_MASK
;
2321 if (swap
->swb_index
+ SWAP_META_PAGES
> info
->endi
)
2322 eindex
= (int)info
->endi
& SWAP_META_MASK
;
2324 eindex
= SWAP_META_MASK
;
2327 * Scan and free the blocks. The loop terminates early
2328 * if (swap) runs out of blocks and could be freed.
2330 * NOTE: Decrement swb_count after swp_pager_freeswapspace()
2331 * to deal with a zfree race.
2333 while (index
<= eindex
) {
2334 swblk_t v
= swap
->swb_pages
[index
];
2336 if (v
!= SWAPBLK_NONE
) {
2337 swap
->swb_pages
[index
] = SWAPBLK_NONE
;
2339 swp_pager_freeswapspace(object
, v
, 1);
2340 --mycpu
->gd_vmtotal
.t_vm
;
2341 if (--swap
->swb_count
== 0) {
2342 swp_pager_remove(object
, swap
);
2343 zfree(swap_zone
, swap
);
2344 --object
->swblock_count
;
2351 /* swap may be invalid here due to zfree above */
2358 * SWP_PAGER_META_FREE_ALL() - destroy all swap metadata associated with object
2360 * This routine locates and destroys all swap metadata associated with
2363 * NOTE: Decrement swb_count after the freeing operation (which
2364 * might block) to prevent racing destruction of the swblock.
2366 * The caller must hold the object.
2369 swp_pager_meta_free_all(vm_object_t object
)
2371 struct swblock
*swap
;
2374 ASSERT_LWKT_TOKEN_HELD(vm_object_token(object
));
2376 while ((swap
= RB_ROOT(&object
->swblock_root
)) != NULL
) {
2377 swp_pager_remove(object
, swap
);
2378 for (i
= 0; i
< SWAP_META_PAGES
; ++i
) {
2379 swblk_t v
= swap
->swb_pages
[i
];
2380 if (v
!= SWAPBLK_NONE
) {
2382 swp_pager_freeswapspace(object
, v
, 1);
2384 --mycpu
->gd_vmtotal
.t_vm
;
2387 if (swap
->swb_count
!= 0)
2388 panic("swap_pager_meta_free_all: swb_count != 0");
2389 zfree(swap_zone
, swap
);
2390 --object
->swblock_count
;
2393 KKASSERT(object
->swblock_count
== 0);
2397 * SWP_PAGER_METACTL() - misc control of swap and vm_page_t meta data.
2399 * This routine is capable of looking up, popping, or freeing
2400 * swapblk assignments in the swap meta data or in the vm_page_t.
2401 * The routine typically returns the swapblk being looked-up, or popped,
2402 * or SWAPBLK_NONE if the block was freed, or SWAPBLK_NONE if the block
2403 * was invalid. This routine will automatically free any invalid
2404 * meta-data swapblks.
2406 * It is not possible to store invalid swapblks in the swap meta data
2407 * (other then a literal 'SWAPBLK_NONE'), so we don't bother checking.
2409 * When acting on a busy resident page and paging is in progress, we
2410 * have to wait until paging is complete but otherwise can act on the
2413 * SWM_FREE remove and free swap block from metadata
2414 * SWM_POP remove from meta data but do not free.. pop it out
2416 * The caller must hold the object.
2419 swp_pager_meta_ctl(vm_object_t object
, vm_pindex_t index
, int flags
)
2421 struct swblock
*swap
;
2424 if (object
->swblock_count
== 0)
2425 return(SWAPBLK_NONE
);
2428 swap
= swp_pager_lookup(object
, index
);
2431 index
&= SWAP_META_MASK
;
2432 r1
= swap
->swb_pages
[index
];
2434 if (r1
!= SWAPBLK_NONE
) {
2435 if (flags
& (SWM_FREE
|SWM_POP
)) {
2436 swap
->swb_pages
[index
] = SWAPBLK_NONE
;
2437 --mycpu
->gd_vmtotal
.t_vm
;
2438 if (--swap
->swb_count
== 0) {
2439 swp_pager_remove(object
, swap
);
2440 zfree(swap_zone
, swap
);
2441 --object
->swblock_count
;
2444 /* swap ptr may be invalid */
2445 if (flags
& SWM_FREE
) {
2446 swp_pager_freeswapspace(object
, r1
, 1);
2450 /* swap ptr may be invalid */