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/kcollect.h>
112 #include "opt_swap.h"
114 #include <vm/vm_object.h>
115 #include <vm/vm_page.h>
116 #include <vm/vm_pager.h>
117 #include <vm/vm_pageout.h>
118 #include <vm/swap_pager.h>
119 #include <vm/vm_extern.h>
120 #include <vm/vm_zone.h>
121 #include <vm/vnode_pager.h>
123 #include <sys/thread2.h>
124 #include <sys/buf2.h>
125 #include <vm/vm_page2.h>
127 #ifndef MAX_PAGEOUT_CLUSTER
128 #define MAX_PAGEOUT_CLUSTER SWB_NPAGES
131 #define SWM_FREE 0x02 /* free, period */
132 #define SWM_POP 0x04 /* pop out */
134 #define SWBIO_READ 0x01
135 #define SWBIO_WRITE 0x02
136 #define SWBIO_SYNC 0x04
137 #define SWBIO_TTC 0x08 /* for VM_PAGER_TRY_TO_CACHE */
143 vm_pindex_t endi
; /* inclusive */
146 struct swswapoffinfo
{
153 * vm_swap_size is in page-sized chunks now. It was DEV_BSIZE'd chunks
157 int swap_pager_full
; /* swap space exhaustion (task killing) */
158 int swap_fail_ticks
; /* when we became exhausted */
159 int swap_pager_almost_full
; /* swap space exhaustion (w/ hysteresis)*/
160 swblk_t vm_swap_cache_use
;
161 swblk_t vm_swap_anon_use
;
162 static int vm_report_swap_allocs
;
164 static int nsw_rcount
; /* free read buffers */
165 static int nsw_wcount_sync
; /* limit write buffers / synchronous */
166 static int nsw_wcount_async
; /* limit write buffers / asynchronous */
167 static int nsw_wcount_async_max
;/* assigned maximum */
168 static int nsw_cluster_max
; /* maximum VOP I/O allowed */
170 struct blist
*swapblist
;
171 static int swap_async_max
= 4; /* maximum in-progress async I/O's */
172 static int swap_burst_read
= 0; /* allow burst reading */
173 static swblk_t swapiterator
; /* linearize allocations */
174 int swap_user_async
= 0; /* user swap pager operation can be async */
176 static struct spinlock swapbp_spin
= SPINLOCK_INITIALIZER(&swapbp_spin
, "swapbp_spin");
179 extern struct vnode
*swapdev_vp
;
180 extern struct swdevt
*swdevt
;
183 #define BLK2DEVIDX(blk) (nswdev > 1 ? blk / SWB_DMMAX % nswdev : 0)
185 SYSCTL_INT(_vm
, OID_AUTO
, swap_async_max
,
186 CTLFLAG_RW
, &swap_async_max
, 0, "Maximum running async swap ops");
187 SYSCTL_INT(_vm
, OID_AUTO
, swap_burst_read
,
188 CTLFLAG_RW
, &swap_burst_read
, 0, "Allow burst reads for pageins");
189 SYSCTL_INT(_vm
, OID_AUTO
, swap_user_async
,
190 CTLFLAG_RW
, &swap_user_async
, 0, "Allow async uuser swap write I/O");
193 SYSCTL_LONG(_vm
, OID_AUTO
, swap_cache_use
,
194 CTLFLAG_RD
, &vm_swap_cache_use
, 0, "");
195 SYSCTL_LONG(_vm
, OID_AUTO
, swap_anon_use
,
196 CTLFLAG_RD
, &vm_swap_anon_use
, 0, "");
197 SYSCTL_LONG(_vm
, OID_AUTO
, swap_size
,
198 CTLFLAG_RD
, &vm_swap_size
, 0, "");
200 SYSCTL_INT(_vm
, OID_AUTO
, swap_cache_use
,
201 CTLFLAG_RD
, &vm_swap_cache_use
, 0, "");
202 SYSCTL_INT(_vm
, OID_AUTO
, swap_anon_use
,
203 CTLFLAG_RD
, &vm_swap_anon_use
, 0, "");
204 SYSCTL_INT(_vm
, OID_AUTO
, swap_size
,
205 CTLFLAG_RD
, &vm_swap_size
, 0, "");
207 SYSCTL_INT(_vm
, OID_AUTO
, report_swap_allocs
,
208 CTLFLAG_RW
, &vm_report_swap_allocs
, 0, "");
213 * Red-Black tree for swblock entries
215 * The caller must hold vm_token
217 RB_GENERATE2(swblock_rb_tree
, swblock
, swb_entry
, rb_swblock_compare
,
218 vm_pindex_t
, swb_index
);
221 rb_swblock_compare(struct swblock
*swb1
, struct swblock
*swb2
)
223 if (swb1
->swb_index
< swb2
->swb_index
)
225 if (swb1
->swb_index
> swb2
->swb_index
)
232 rb_swblock_scancmp(struct swblock
*swb
, void *data
)
234 struct swfreeinfo
*info
= data
;
236 if (swb
->swb_index
< info
->basei
)
238 if (swb
->swb_index
> info
->endi
)
245 rb_swblock_condcmp(struct swblock
*swb
, void *data
)
247 struct swfreeinfo
*info
= data
;
249 if (swb
->swb_index
< info
->basei
)
255 * pagerops for OBJT_SWAP - "swap pager". Some ops are also global procedure
256 * calls hooked from other parts of the VM system and do not appear here.
257 * (see vm/swap_pager.h).
260 static void swap_pager_dealloc (vm_object_t object
);
261 static int swap_pager_getpage (vm_object_t
, vm_page_t
*, int);
262 static void swap_chain_iodone(struct bio
*biox
);
264 struct pagerops swappagerops
= {
265 swap_pager_dealloc
, /* deallocate an OBJT_SWAP object */
266 swap_pager_getpage
, /* pagein */
267 swap_pager_putpages
, /* pageout */
268 swap_pager_haspage
/* get backing store status for page */
272 * SWB_DMMAX is in page-sized chunks with the new swap system. It was
273 * dev-bsized chunks in the old. SWB_DMMAX is always a power of 2.
275 * swap_*() routines are externally accessible. swp_*() routines are
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 * Long-term data collection on 10-second interval. Return the value
333 * for KCOLLECT_SWAPPCT and set the values for SWAPANO and SWAPCCAC.
335 * Return total swap in the scale field. This can change if swap is
336 * regularly added or removed and may cause some historical confusion
337 * in that case, but SWAPPCT will always be historically accurate.
340 collect_swap_callback(int n
)
342 uint64_t total
= vm_swap_size
;
343 uint64_t anon
= vm_swap_anon_use
;
344 uint64_t cache
= vm_swap_cache_use
;
346 if (total
== 0) /* avoid divide by zero */
348 kcollect_setvalue(KCOLLECT_SWAPANO
, anon
* PAGE_SIZE
);
349 kcollect_setvalue(KCOLLECT_SWAPCAC
, cache
* PAGE_SIZE
);
350 kcollect_setscale(KCOLLECT_SWAPANO
, total
);
351 kcollect_setscale(KCOLLECT_SWAPCAC
, total
);
352 return (((anon
+ cache
) * 10000 + (total
>> 1)) / total
);
356 * SWAP_PAGER_INIT() - initialize the swap pager!
358 * Expected to be started from system init. NOTE: This code is run
359 * before much else so be careful what you depend on. Most of the VM
360 * system has yet to be initialized at this point.
362 * Called from the low level boot code only.
365 swap_pager_init(void *arg __unused
)
367 kcollect_register(KCOLLECT_SWAPPCT
, "swapuse", collect_swap_callback
,
368 KCOLLECT_SCALE(KCOLLECT_SWAPPCT_FORMAT
, 0));
369 kcollect_register(KCOLLECT_SWAPANO
, "swapmem", NULL
,
370 KCOLLECT_SCALE(KCOLLECT_SWAPANO_FORMAT
, 0));
371 kcollect_register(KCOLLECT_SWAPCAC
, "swapcsh", NULL
,
372 KCOLLECT_SCALE(KCOLLECT_SWAPCAC_FORMAT
, 0));
374 SYSINIT(vm_mem
, SI_BOOT1_VM
, SI_ORDER_THIRD
, swap_pager_init
, NULL
);
377 * SWAP_PAGER_SWAP_INIT() - swap pager initialization from pageout process
379 * Expected to be started from pageout process once, prior to entering
382 * Called from the low level boot code only.
385 swap_pager_swap_init(void)
390 * Number of in-transit swap bp operations. Don't
391 * exhaust the pbufs completely. Make sure we
392 * initialize workable values (0 will work for hysteresis
393 * but it isn't very efficient).
395 * The nsw_cluster_max is constrained by the number of pages an XIO
396 * holds, i.e., (MAXPHYS/PAGE_SIZE) and our locally defined
397 * MAX_PAGEOUT_CLUSTER. Also be aware that swap ops are
398 * constrained by the swap device interleave stripe size.
400 * Currently we hardwire nsw_wcount_async to 4. This limit is
401 * designed to prevent other I/O from having high latencies due to
402 * our pageout I/O. The value 4 works well for one or two active swap
403 * devices but is probably a little low if you have more. Even so,
404 * a higher value would probably generate only a limited improvement
405 * with three or four active swap devices since the system does not
406 * typically have to pageout at extreme bandwidths. We will want
407 * at least 2 per swap devices, and 4 is a pretty good value if you
408 * have one NFS swap device due to the command/ack latency over NFS.
409 * So it all works out pretty well.
412 nsw_cluster_max
= min((MAXPHYS
/PAGE_SIZE
), MAX_PAGEOUT_CLUSTER
);
414 nsw_rcount
= (nswbuf_kva
+ 1) / 2;
415 nsw_wcount_sync
= (nswbuf_kva
+ 3) / 4;
416 nsw_wcount_async
= 4;
417 nsw_wcount_async_max
= nsw_wcount_async
;
420 * The zone is dynamically allocated so generally size it to
421 * maxswzone (32MB to 256GB of KVM). Set a minimum size based
422 * on physical memory of around 8x (each swblock can hold 16 pages).
424 * With the advent of SSDs (vs HDs) the practical (swap:memory) ratio
425 * has increased dramatically.
427 n
= vmstats
.v_page_count
/ 2;
428 if (maxswzone
&& n
< maxswzone
/ sizeof(struct swblock
))
429 n
= maxswzone
/ sizeof(struct swblock
);
435 sizeof(struct swblock
),
438 if (swap_zone
!= NULL
)
441 * if the allocation failed, try a zone two thirds the
442 * size of the previous attempt.
447 if (swap_zone
== NULL
)
448 panic("swap_pager_swap_init: swap_zone == NULL");
450 kprintf("Swap zone entries reduced from %d to %d.\n", n2
, n
);
454 * SWAP_PAGER_ALLOC() - allocate a new OBJT_SWAP VM object and instantiate
455 * its metadata structures.
457 * This routine is called from the mmap and fork code to create a new
458 * OBJT_SWAP object. We do this by creating an OBJT_DEFAULT object
459 * and then converting it with swp_pager_meta_convert().
461 * We only support unnamed objects.
466 swap_pager_alloc(void *handle
, off_t size
, vm_prot_t prot
, off_t offset
)
470 KKASSERT(handle
== NULL
);
471 object
= vm_object_allocate_hold(OBJT_DEFAULT
,
472 OFF_TO_IDX(offset
+ PAGE_MASK
+ size
));
473 swp_pager_meta_convert(object
);
474 vm_object_drop(object
);
480 * SWAP_PAGER_DEALLOC() - remove swap metadata from object
482 * The swap backing for the object is destroyed. The code is
483 * designed such that we can reinstantiate it later, but this
484 * routine is typically called only when the entire object is
485 * about to be destroyed.
487 * The object must be locked or unreferenceable.
488 * No other requirements.
491 swap_pager_dealloc(vm_object_t object
)
493 vm_object_hold(object
);
494 vm_object_pip_wait(object
, "swpdea");
497 * Free all remaining metadata. We only bother to free it from
498 * the swap meta data. We do not attempt to free swapblk's still
499 * associated with vm_page_t's for this object. We do not care
500 * if paging is still in progress on some objects.
502 swp_pager_meta_free_all(object
);
503 vm_object_drop(object
);
506 /************************************************************************
507 * SWAP PAGER BITMAP ROUTINES *
508 ************************************************************************/
511 * SWP_PAGER_GETSWAPSPACE() - allocate raw swap space
513 * Allocate swap for the requested number of pages. The starting
514 * swap block number (a page index) is returned or SWAPBLK_NONE
515 * if the allocation failed.
517 * Also has the side effect of advising that somebody made a mistake
518 * when they configured swap and didn't configure enough.
520 * The caller must hold the object.
521 * This routine may not block.
523 static __inline swblk_t
524 swp_pager_getswapspace(vm_object_t object
, int npages
)
528 lwkt_gettoken(&vm_token
);
529 blk
= blist_allocat(swapblist
, npages
, swapiterator
);
530 if (blk
== SWAPBLK_NONE
)
531 blk
= blist_allocat(swapblist
, npages
, 0);
532 if (blk
== SWAPBLK_NONE
) {
533 if (swap_pager_full
!= 2) {
534 if (vm_swap_max
== 0)
535 kprintf("Warning: The system would like to "
536 "page to swap but no swap space "
539 kprintf("swap_pager_getswapspace: "
540 "swap full allocating %d pages\n",
543 if (swap_pager_almost_full
== 0)
544 swap_fail_ticks
= ticks
;
545 swap_pager_almost_full
= 1;
548 /* swapiterator = blk; disable for now, doesn't work well */
549 swapacctspace(blk
, -npages
);
550 if (object
->type
== OBJT_SWAP
)
551 vm_swap_anon_use
+= npages
;
553 vm_swap_cache_use
+= npages
;
556 lwkt_reltoken(&vm_token
);
561 * SWP_PAGER_FREESWAPSPACE() - free raw swap space
563 * This routine returns the specified swap blocks back to the bitmap.
565 * Note: This routine may not block (it could in the old swap code),
566 * and through the use of the new blist routines it does not block.
568 * This routine may not block.
572 swp_pager_freeswapspace(vm_object_t object
, swblk_t blk
, int npages
)
574 struct swdevt
*sp
= &swdevt
[BLK2DEVIDX(blk
)];
576 lwkt_gettoken(&vm_token
);
577 sp
->sw_nused
-= npages
;
578 if (object
->type
== OBJT_SWAP
)
579 vm_swap_anon_use
-= npages
;
581 vm_swap_cache_use
-= npages
;
583 if (sp
->sw_flags
& SW_CLOSING
) {
584 lwkt_reltoken(&vm_token
);
588 blist_free(swapblist
, blk
, npages
);
589 vm_swap_size
+= npages
;
591 lwkt_reltoken(&vm_token
);
595 * SWAP_PAGER_FREESPACE() - frees swap blocks associated with a page
596 * range within an object.
598 * This is a globally accessible routine.
600 * This routine removes swapblk assignments from swap metadata.
602 * The external callers of this routine typically have already destroyed
603 * or renamed vm_page_t's associated with this range in the object so
609 swap_pager_freespace(vm_object_t object
, vm_pindex_t start
, vm_pindex_t size
)
611 vm_object_hold(object
);
612 swp_pager_meta_free(object
, start
, size
);
613 vm_object_drop(object
);
620 swap_pager_freespace_all(vm_object_t object
)
622 vm_object_hold(object
);
623 swp_pager_meta_free_all(object
);
624 vm_object_drop(object
);
628 * This function conditionally frees swap cache swap starting at
629 * (*basei) in the object. (count) swap blocks will be nominally freed.
630 * The actual number of blocks freed can be more or less than the
633 * This function nominally returns the number of blocks freed. However,
634 * the actual number of blocks freed may be less then the returned value.
635 * If the function is unable to exhaust the object or if it is able to
636 * free (approximately) the requested number of blocks it returns
639 * If we exhaust the object we will return a value n <= count.
641 * The caller must hold the object.
643 * WARNING! If count == 0 then -1 can be returned as a degenerate case,
644 * callers should always pass a count value > 0.
646 static int swap_pager_condfree_callback(struct swblock
*swap
, void *data
);
649 swap_pager_condfree(vm_object_t object
, vm_pindex_t
*basei
, int count
)
651 struct swfreeinfo info
;
655 ASSERT_LWKT_TOKEN_HELD(vm_object_token(object
));
657 info
.object
= object
;
658 info
.basei
= *basei
; /* skip up to this page index */
659 info
.begi
= count
; /* max swap pages to destroy */
660 info
.endi
= count
* 8; /* max swblocks to scan */
662 swblock_rb_tree_RB_SCAN(&object
->swblock_root
, rb_swblock_condcmp
,
663 swap_pager_condfree_callback
, &info
);
667 * Take the higher difference swblocks vs pages
669 n
= count
- (int)info
.begi
;
670 t
= count
* 8 - (int)info
.endi
;
679 * The idea is to free whole meta-block to avoid fragmenting
680 * the swap space or disk I/O. We only do this if NO VM pages
683 * We do not have to deal with clearing PG_SWAPPED in related VM
684 * pages because there are no related VM pages.
686 * The caller must hold the object.
689 swap_pager_condfree_callback(struct swblock
*swap
, void *data
)
691 struct swfreeinfo
*info
= data
;
692 vm_object_t object
= info
->object
;
695 for (i
= 0; i
< SWAP_META_PAGES
; ++i
) {
696 if (vm_page_lookup(object
, swap
->swb_index
+ i
))
699 info
->basei
= swap
->swb_index
+ SWAP_META_PAGES
;
700 if (i
== SWAP_META_PAGES
) {
701 info
->begi
-= swap
->swb_count
;
702 swap_pager_freespace(object
, swap
->swb_index
, SWAP_META_PAGES
);
705 if ((int)info
->begi
< 0 || (int)info
->endi
< 0)
712 * Called by vm_page_alloc() when a new VM page is inserted
713 * into a VM object. Checks whether swap has been assigned to
714 * the page and sets PG_SWAPPED as necessary.
716 * (m) must be busied by caller and remains busied on return.
719 swap_pager_page_inserted(vm_page_t m
)
721 if (m
->object
->swblock_count
) {
722 vm_object_hold(m
->object
);
723 if (swp_pager_meta_ctl(m
->object
, m
->pindex
, 0) != SWAPBLK_NONE
)
724 vm_page_flag_set(m
, PG_SWAPPED
);
725 vm_object_drop(m
->object
);
730 * SWAP_PAGER_RESERVE() - reserve swap blocks in object
732 * Assigns swap blocks to the specified range within the object. The
733 * swap blocks are not zerod. Any previous swap assignment is destroyed.
735 * Returns 0 on success, -1 on failure.
737 * The caller is responsible for avoiding races in the specified range.
738 * No other requirements.
741 swap_pager_reserve(vm_object_t object
, vm_pindex_t start
, vm_size_t size
)
744 swblk_t blk
= SWAPBLK_NONE
;
745 vm_pindex_t beg
= start
; /* save start index */
747 vm_object_hold(object
);
752 while ((blk
= swp_pager_getswapspace(object
, n
)) ==
757 swp_pager_meta_free(object
, beg
,
759 vm_object_drop(object
);
764 swp_pager_meta_build(object
, start
, blk
);
770 swp_pager_meta_free(object
, start
, n
);
771 vm_object_drop(object
);
776 * SWAP_PAGER_COPY() - copy blocks from source pager to destination pager
777 * and destroy the source.
779 * Copy any valid swapblks from the source to the destination. In
780 * cases where both the source and destination have a valid swapblk,
781 * we keep the destination's.
783 * This routine is allowed to block. It may block allocating metadata
784 * indirectly through swp_pager_meta_build() or if paging is still in
785 * progress on the source.
787 * XXX vm_page_collapse() kinda expects us not to block because we
788 * supposedly do not need to allocate memory, but for the moment we
789 * *may* have to get a little memory from the zone allocator, but
790 * it is taken from the interrupt memory. We should be ok.
792 * The source object contains no vm_page_t's (which is just as well)
793 * The source object is of type OBJT_SWAP.
795 * The source and destination objects must be held by the caller.
798 swap_pager_copy(vm_object_t srcobject
, vm_object_t dstobject
,
799 vm_pindex_t base_index
, int destroysource
)
803 ASSERT_LWKT_TOKEN_HELD(vm_object_token(srcobject
));
804 ASSERT_LWKT_TOKEN_HELD(vm_object_token(dstobject
));
807 * transfer source to destination.
809 for (i
= 0; i
< dstobject
->size
; ++i
) {
813 * Locate (without changing) the swapblk on the destination,
814 * unless it is invalid in which case free it silently, or
815 * if the destination is a resident page, in which case the
816 * source is thrown away.
818 dstaddr
= swp_pager_meta_ctl(dstobject
, i
, 0);
820 if (dstaddr
== SWAPBLK_NONE
) {
822 * Destination has no swapblk and is not resident,
827 srcaddr
= swp_pager_meta_ctl(srcobject
,
828 base_index
+ i
, SWM_POP
);
830 if (srcaddr
!= SWAPBLK_NONE
)
831 swp_pager_meta_build(dstobject
, i
, srcaddr
);
834 * Destination has valid swapblk or it is represented
835 * by a resident page. We destroy the sourceblock.
837 swp_pager_meta_ctl(srcobject
, base_index
+ i
, SWM_FREE
);
842 * Free left over swap blocks in source.
844 * We have to revert the type to OBJT_DEFAULT so we do not accidently
845 * double-remove the object from the swap queues.
849 * Reverting the type is not necessary, the caller is going
850 * to destroy srcobject directly, but I'm doing it here
851 * for consistency since we've removed the object from its
854 swp_pager_meta_free_all(srcobject
);
855 if (srcobject
->type
== OBJT_SWAP
)
856 srcobject
->type
= OBJT_DEFAULT
;
861 * SWAP_PAGER_HASPAGE() - determine if we have good backing store for
862 * the requested page.
864 * We determine whether good backing store exists for the requested
865 * page and return TRUE if it does, FALSE if it doesn't.
867 * If TRUE, we also try to determine how much valid, contiguous backing
868 * store exists before and after the requested page within a reasonable
869 * distance. We do not try to restrict it to the swap device stripe
870 * (that is handled in getpages/putpages). It probably isn't worth
876 swap_pager_haspage(vm_object_t object
, vm_pindex_t pindex
)
881 * do we have good backing store at the requested index ?
883 vm_object_hold(object
);
884 blk0
= swp_pager_meta_ctl(object
, pindex
, 0);
886 if (blk0
== SWAPBLK_NONE
) {
887 vm_object_drop(object
);
890 vm_object_drop(object
);
895 * SWAP_PAGER_PAGE_UNSWAPPED() - remove swap backing store related to page
897 * This removes any associated swap backing store, whether valid or
898 * not, from the page. This operates on any VM object, not just OBJT_SWAP
901 * This routine is typically called when a page is made dirty, at
902 * which point any associated swap can be freed. MADV_FREE also
903 * calls us in a special-case situation
905 * NOTE!!! If the page is clean and the swap was valid, the caller
906 * should make the page dirty before calling this routine.
907 * This routine does NOT change the m->dirty status of the page.
908 * Also: MADV_FREE depends on it.
910 * The page must be busied.
911 * The caller can hold the object to avoid blocking, else we might block.
912 * No other requirements.
915 swap_pager_unswapped(vm_page_t m
)
917 if (m
->flags
& PG_SWAPPED
) {
918 vm_object_hold(m
->object
);
919 KKASSERT(m
->flags
& PG_SWAPPED
);
920 swp_pager_meta_ctl(m
->object
, m
->pindex
, SWM_FREE
);
921 vm_page_flag_clear(m
, PG_SWAPPED
);
922 vm_object_drop(m
->object
);
927 * SWAP_PAGER_STRATEGY() - read, write, free blocks
929 * This implements a VM OBJECT strategy function using swap backing store.
930 * This can operate on any VM OBJECT type, not necessarily just OBJT_SWAP
933 * This is intended to be a cacheless interface (i.e. caching occurs at
934 * higher levels), and is also used as a swap-based SSD cache for vnode
935 * and device objects.
937 * All I/O goes directly to and from the swap device.
939 * We currently attempt to run I/O synchronously or asynchronously as
940 * the caller requests. This isn't perfect because we loose error
941 * sequencing when we run multiple ops in parallel to satisfy a request.
942 * But this is swap, so we let it all hang out.
947 swap_pager_strategy(vm_object_t object
, struct bio
*bio
)
949 struct buf
*bp
= bio
->bio_buf
;
952 vm_pindex_t biox_blkno
= 0;
958 struct bio_track
*track
;
963 * tracking for swapdev vnode I/Os
965 if (bp
->b_cmd
== BUF_CMD_READ
)
966 track
= &swapdev_vp
->v_track_read
;
968 track
= &swapdev_vp
->v_track_write
;
971 if (bp
->b_bcount
& PAGE_MASK
) {
972 bp
->b_error
= EINVAL
;
973 bp
->b_flags
|= B_ERROR
| B_INVAL
;
975 kprintf("swap_pager_strategy: bp %p offset %lld size %d, "
976 "not page bounded\n",
977 bp
, (long long)bio
->bio_offset
, (int)bp
->b_bcount
);
982 * Clear error indication, initialize page index, count, data pointer.
985 bp
->b_flags
&= ~B_ERROR
;
986 bp
->b_resid
= bp
->b_bcount
;
988 start
= (vm_pindex_t
)(bio
->bio_offset
>> PAGE_SHIFT
);
989 count
= howmany(bp
->b_bcount
, PAGE_SIZE
);
993 * Deal with BUF_CMD_FREEBLKS
995 if (bp
->b_cmd
== BUF_CMD_FREEBLKS
) {
997 * FREE PAGE(s) - destroy underlying swap that is no longer
1000 vm_object_hold(object
);
1001 swp_pager_meta_free(object
, start
, count
);
1002 vm_object_drop(object
);
1009 * We need to be able to create a new cluster of I/O's. We cannot
1010 * use the caller fields of the passed bio so push a new one.
1012 * Because nbio is just a placeholder for the cluster links,
1013 * we can biodone() the original bio instead of nbio to make
1014 * things a bit more efficient.
1016 nbio
= push_bio(bio
);
1017 nbio
->bio_offset
= bio
->bio_offset
;
1018 nbio
->bio_caller_info1
.cluster_head
= NULL
;
1019 nbio
->bio_caller_info2
.cluster_tail
= NULL
;
1025 * Execute read or write
1027 vm_object_hold(object
);
1033 * Obtain block. If block not found and writing, allocate a
1034 * new block and build it into the object.
1036 blk
= swp_pager_meta_ctl(object
, start
, 0);
1037 if ((blk
== SWAPBLK_NONE
) && bp
->b_cmd
!= BUF_CMD_READ
) {
1038 blk
= swp_pager_getswapspace(object
, 1);
1039 if (blk
== SWAPBLK_NONE
) {
1040 bp
->b_error
= ENOMEM
;
1041 bp
->b_flags
|= B_ERROR
;
1044 swp_pager_meta_build(object
, start
, blk
);
1048 * Do we have to flush our current collection? Yes if:
1050 * - no swap block at this index
1051 * - swap block is not contiguous
1052 * - we cross a physical disk boundry in the
1056 biox
&& (biox_blkno
+ btoc(bufx
->b_bcount
) != blk
||
1057 ((biox_blkno
^ blk
) & ~SWB_DMMASK
)
1060 if (bp
->b_cmd
== BUF_CMD_READ
) {
1061 ++mycpu
->gd_cnt
.v_swapin
;
1062 mycpu
->gd_cnt
.v_swappgsin
+= btoc(bufx
->b_bcount
);
1064 ++mycpu
->gd_cnt
.v_swapout
;
1065 mycpu
->gd_cnt
.v_swappgsout
+= btoc(bufx
->b_bcount
);
1066 bufx
->b_dirtyend
= bufx
->b_bcount
;
1070 * Finished with this buf.
1072 KKASSERT(bufx
->b_bcount
!= 0);
1073 if (bufx
->b_cmd
!= BUF_CMD_READ
)
1074 bufx
->b_dirtyend
= bufx
->b_bcount
;
1080 * Add new swapblk to biox, instantiating biox if necessary.
1081 * Zero-fill reads are able to take a shortcut.
1083 if (blk
== SWAPBLK_NONE
) {
1085 * We can only get here if we are reading.
1087 bzero(data
, PAGE_SIZE
);
1088 bp
->b_resid
-= PAGE_SIZE
;
1091 /* XXX chain count > 4, wait to <= 4 */
1093 bufx
= getpbuf(NULL
);
1094 biox
= &bufx
->b_bio1
;
1095 cluster_append(nbio
, bufx
);
1096 bufx
->b_cmd
= bp
->b_cmd
;
1097 biox
->bio_done
= swap_chain_iodone
;
1098 biox
->bio_offset
= (off_t
)blk
<< PAGE_SHIFT
;
1099 biox
->bio_caller_info1
.cluster_parent
= nbio
;
1102 bufx
->b_data
= data
;
1104 bufx
->b_bcount
+= PAGE_SIZE
;
1111 vm_object_drop(object
);
1114 * Flush out last buffer
1117 if (bufx
->b_cmd
== BUF_CMD_READ
) {
1118 ++mycpu
->gd_cnt
.v_swapin
;
1119 mycpu
->gd_cnt
.v_swappgsin
+= btoc(bufx
->b_bcount
);
1121 ++mycpu
->gd_cnt
.v_swapout
;
1122 mycpu
->gd_cnt
.v_swappgsout
+= btoc(bufx
->b_bcount
);
1123 bufx
->b_dirtyend
= bufx
->b_bcount
;
1125 KKASSERT(bufx
->b_bcount
);
1126 if (bufx
->b_cmd
!= BUF_CMD_READ
)
1127 bufx
->b_dirtyend
= bufx
->b_bcount
;
1128 /* biox, bufx = NULL */
1132 * Now initiate all the I/O. Be careful looping on our chain as
1133 * I/O's may complete while we are still initiating them.
1135 * If the request is a 100% sparse read no bios will be present
1136 * and we just biodone() the buffer.
1138 nbio
->bio_caller_info2
.cluster_tail
= NULL
;
1139 bufx
= nbio
->bio_caller_info1
.cluster_head
;
1143 biox
= &bufx
->b_bio1
;
1145 bufx
= bufx
->b_cluster_next
;
1146 vn_strategy(swapdev_vp
, biox
);
1153 * Completion of the cluster will also call biodone_chain(nbio).
1154 * We never call biodone(nbio) so we don't have to worry about
1155 * setting up a bio_done callback. It's handled in the sub-IO.
1166 swap_chain_iodone(struct bio
*biox
)
1169 struct buf
*bufx
; /* chained sub-buffer */
1170 struct bio
*nbio
; /* parent nbio with chain glue */
1171 struct buf
*bp
; /* original bp associated with nbio */
1174 bufx
= biox
->bio_buf
;
1175 nbio
= biox
->bio_caller_info1
.cluster_parent
;
1179 * Update the original buffer
1181 KKASSERT(bp
!= NULL
);
1182 if (bufx
->b_flags
& B_ERROR
) {
1183 atomic_set_int(&bufx
->b_flags
, B_ERROR
);
1184 bp
->b_error
= bufx
->b_error
; /* race ok */
1185 } else if (bufx
->b_resid
!= 0) {
1186 atomic_set_int(&bufx
->b_flags
, B_ERROR
);
1187 bp
->b_error
= EINVAL
; /* race ok */
1189 atomic_subtract_int(&bp
->b_resid
, bufx
->b_bcount
);
1193 * Remove us from the chain.
1195 spin_lock(&swapbp_spin
);
1196 nextp
= &nbio
->bio_caller_info1
.cluster_head
;
1197 while (*nextp
!= bufx
) {
1198 KKASSERT(*nextp
!= NULL
);
1199 nextp
= &(*nextp
)->b_cluster_next
;
1201 *nextp
= bufx
->b_cluster_next
;
1202 chain_empty
= (nbio
->bio_caller_info1
.cluster_head
== NULL
);
1203 spin_unlock(&swapbp_spin
);
1206 * Clean up bufx. If the chain is now empty we finish out
1207 * the parent. Note that we may be racing other completions
1208 * so we must use the chain_empty status from above.
1211 if (bp
->b_resid
!= 0 && !(bp
->b_flags
& B_ERROR
)) {
1212 atomic_set_int(&bp
->b_flags
, B_ERROR
);
1213 bp
->b_error
= EINVAL
;
1215 biodone_chain(nbio
);
1217 relpbuf(bufx
, NULL
);
1221 * SWAP_PAGER_GETPAGES() - bring page in from swap
1223 * The requested page may have to be brought in from swap. Calculate the
1224 * swap block and bring in additional pages if possible. All pages must
1225 * have contiguous swap block assignments and reside in the same object.
1227 * The caller has a single vm_object_pip_add() reference prior to
1228 * calling us and we should return with the same.
1230 * The caller has BUSY'd the page. We should return with (*mpp) left busy,
1231 * and any additinal pages unbusied.
1233 * If the caller encounters a PG_RAM page it will pass it to us even though
1234 * it may be valid and dirty. We cannot overwrite the page in this case!
1235 * The case is used to allow us to issue pure read-aheads.
1237 * NOTE! XXX This code does not entirely pipeline yet due to the fact that
1238 * the PG_RAM page is validated at the same time as mreq. What we
1239 * really need to do is issue a separate read-ahead pbuf.
1244 swap_pager_getpage(vm_object_t object
, vm_page_t
*mpp
, int seqaccess
)
1257 vm_page_t marray
[XIO_INTERNAL_PAGES
];
1261 vm_object_hold(object
);
1262 if (mreq
->object
!= object
) {
1263 panic("swap_pager_getpages: object mismatch %p/%p",
1270 * We don't want to overwrite a fully valid page as it might be
1271 * dirty. This case can occur when e.g. vm_fault hits a perfectly
1272 * valid page with PG_RAM set.
1274 * In this case we see if the next page is a suitable page-in
1275 * candidate and if it is we issue read-ahead. PG_RAM will be
1276 * set on the last page of the read-ahead to continue the pipeline.
1278 if (mreq
->valid
== VM_PAGE_BITS_ALL
) {
1279 if (swap_burst_read
== 0 || mreq
->pindex
+ 1 >= object
->size
) {
1280 vm_object_drop(object
);
1281 return(VM_PAGER_OK
);
1283 blk
= swp_pager_meta_ctl(object
, mreq
->pindex
+ 1, 0);
1284 if (blk
== SWAPBLK_NONE
) {
1285 vm_object_drop(object
);
1286 return(VM_PAGER_OK
);
1288 m
= vm_page_lookup_busy_try(object
, mreq
->pindex
+ 1,
1291 vm_object_drop(object
);
1292 return(VM_PAGER_OK
);
1293 } else if (m
== NULL
) {
1295 * Use VM_ALLOC_QUICK to avoid blocking on cache
1298 m
= vm_page_alloc(object
, mreq
->pindex
+ 1,
1301 vm_object_drop(object
);
1302 return(VM_PAGER_OK
);
1307 vm_object_drop(object
);
1308 return(VM_PAGER_OK
);
1310 vm_page_unqueue_nowakeup(m
);
1320 * Try to block-read contiguous pages from swap if sequential,
1321 * otherwise just read one page. Contiguous pages from swap must
1322 * reside within a single device stripe because the I/O cannot be
1323 * broken up across multiple stripes.
1325 * Note that blk and iblk can be SWAPBLK_NONE but the loop is
1326 * set up such that the case(s) are handled implicitly.
1328 blk
= swp_pager_meta_ctl(mreq
->object
, mreq
->pindex
, 0);
1331 for (i
= 1; i
<= swap_burst_read
&&
1332 i
< XIO_INTERNAL_PAGES
&&
1333 mreq
->pindex
+ i
< object
->size
; ++i
) {
1336 iblk
= swp_pager_meta_ctl(object
, mreq
->pindex
+ i
, 0);
1337 if (iblk
!= blk
+ i
)
1339 if ((blk
^ iblk
) & ~SWB_DMMASK
)
1341 m
= vm_page_lookup_busy_try(object
, mreq
->pindex
+ i
,
1345 } else if (m
== NULL
) {
1347 * Use VM_ALLOC_QUICK to avoid blocking on cache
1350 m
= vm_page_alloc(object
, mreq
->pindex
+ i
,
1359 vm_page_unqueue_nowakeup(m
);
1365 vm_page_flag_set(marray
[i
- 1], PG_RAM
);
1368 * If mreq is the requested page and we have nothing to do return
1369 * VM_PAGER_FAIL. If raonly is set mreq is just another read-ahead
1370 * page and must be cleaned up.
1372 if (blk
== SWAPBLK_NONE
) {
1375 vnode_pager_freepage(mreq
);
1376 vm_object_drop(object
);
1377 return(VM_PAGER_OK
);
1379 vm_object_drop(object
);
1380 return(VM_PAGER_FAIL
);
1385 * map our page(s) into kva for input
1387 bp
= getpbuf_kva(&nsw_rcount
);
1389 kva
= (vm_offset_t
) bp
->b_kvabase
;
1390 bcopy(marray
, bp
->b_xio
.xio_pages
, i
* sizeof(vm_page_t
));
1391 pmap_qenter(kva
, bp
->b_xio
.xio_pages
, i
);
1393 bp
->b_data
= (caddr_t
)kva
;
1394 bp
->b_bcount
= PAGE_SIZE
* i
;
1395 bp
->b_xio
.xio_npages
= i
;
1396 bio
->bio_done
= swp_pager_async_iodone
;
1397 bio
->bio_offset
= (off_t
)blk
<< PAGE_SHIFT
;
1398 bio
->bio_caller_info1
.index
= SWBIO_READ
;
1401 * Set index. If raonly set the index beyond the array so all
1402 * the pages are treated the same, otherwise the original mreq is
1406 bio
->bio_driver_info
= (void *)(intptr_t)i
;
1408 bio
->bio_driver_info
= (void *)(intptr_t)0;
1410 for (j
= 0; j
< i
; ++j
)
1411 vm_page_flag_set(bp
->b_xio
.xio_pages
[j
], PG_SWAPINPROG
);
1413 mycpu
->gd_cnt
.v_swapin
++;
1414 mycpu
->gd_cnt
.v_swappgsin
+= bp
->b_xio
.xio_npages
;
1417 * We still hold the lock on mreq, and our automatic completion routine
1418 * does not remove it.
1420 vm_object_pip_add(object
, bp
->b_xio
.xio_npages
);
1423 * perform the I/O. NOTE!!! bp cannot be considered valid after
1424 * this point because we automatically release it on completion.
1425 * Instead, we look at the one page we are interested in which we
1426 * still hold a lock on even through the I/O completion.
1428 * The other pages in our m[] array are also released on completion,
1429 * so we cannot assume they are valid anymore either.
1431 bp
->b_cmd
= BUF_CMD_READ
;
1433 vn_strategy(swapdev_vp
, bio
);
1436 * Wait for the page we want to complete. PG_SWAPINPROG is always
1437 * cleared on completion. If an I/O error occurs, SWAPBLK_NONE
1438 * is set in the meta-data.
1440 * If this is a read-ahead only we return immediately without
1444 vm_object_drop(object
);
1445 return(VM_PAGER_OK
);
1449 * Read-ahead includes originally requested page case.
1452 flags
= mreq
->flags
;
1454 if ((flags
& PG_SWAPINPROG
) == 0)
1456 tsleep_interlock(mreq
, 0);
1457 if (!atomic_cmpset_int(&mreq
->flags
, flags
,
1458 flags
| PG_WANTED
| PG_REFERENCED
)) {
1461 mycpu
->gd_cnt
.v_intrans
++;
1462 if (tsleep(mreq
, PINTERLOCKED
, "swread", hz
*20)) {
1464 "swap_pager: indefinite wait buffer: "
1465 " bp %p offset: %lld, size: %ld\n",
1467 (long long)bio
->bio_offset
,
1474 * Disallow speculative reads prior to the PG_SWAPINPROG test.
1479 * mreq is left busied after completion, but all the other pages
1480 * are freed. If we had an unrecoverable read error the page will
1483 vm_object_drop(object
);
1484 if (mreq
->valid
!= VM_PAGE_BITS_ALL
)
1485 return(VM_PAGER_ERROR
);
1487 return(VM_PAGER_OK
);
1490 * A final note: in a low swap situation, we cannot deallocate swap
1491 * and mark a page dirty here because the caller is likely to mark
1492 * the page clean when we return, causing the page to possibly revert
1493 * to all-zero's later.
1498 * swap_pager_putpages:
1500 * Assign swap (if necessary) and initiate I/O on the specified pages.
1502 * We support both OBJT_DEFAULT and OBJT_SWAP objects. DEFAULT objects
1503 * are automatically converted to SWAP objects.
1505 * In a low memory situation we may block in vn_strategy(), but the new
1506 * vm_page reservation system coupled with properly written VFS devices
1507 * should ensure that no low-memory deadlock occurs. This is an area
1510 * The parent has N vm_object_pip_add() references prior to
1511 * calling us and will remove references for rtvals[] that are
1512 * not set to VM_PAGER_PEND. We need to remove the rest on I/O
1515 * The parent has soft-busy'd the pages it passes us and will unbusy
1516 * those whos rtvals[] entry is not set to VM_PAGER_PEND on return.
1517 * We need to unbusy the rest on I/O completion.
1522 swap_pager_putpages(vm_object_t object
, vm_page_t
*m
, int count
,
1523 int flags
, int *rtvals
)
1528 vm_object_hold(object
);
1530 if (count
&& m
[0]->object
!= object
) {
1531 panic("swap_pager_getpages: object mismatch %p/%p",
1540 * Turn object into OBJT_SWAP
1541 * Check for bogus sysops
1543 * Force sync if not pageout process, we don't want any single
1544 * non-pageout process to be able to hog the I/O subsystem! This
1545 * can be overridden by setting.
1547 if (object
->type
== OBJT_DEFAULT
) {
1548 if (object
->type
== OBJT_DEFAULT
)
1549 swp_pager_meta_convert(object
);
1553 * Normally we force synchronous swap I/O if this is not the
1554 * pageout daemon to prevent any single user process limited
1555 * via RLIMIT_RSS from hogging swap write bandwidth.
1557 if (curthread
!= pagethread
&& swap_user_async
== 0)
1558 flags
|= VM_PAGER_PUT_SYNC
;
1563 * Update nsw parameters from swap_async_max sysctl values.
1564 * Do not let the sysop crash the machine with bogus numbers.
1566 if (swap_async_max
!= nsw_wcount_async_max
) {
1572 if ((n
= swap_async_max
) > nswbuf_kva
/ 2)
1579 * Adjust difference ( if possible ). If the current async
1580 * count is too low, we may not be able to make the adjustment
1583 * vm_token needed for nsw_wcount sleep interlock
1585 lwkt_gettoken(&vm_token
);
1586 n
-= nsw_wcount_async_max
;
1587 if (nsw_wcount_async
+ n
>= 0) {
1588 nsw_wcount_async_max
+= n
;
1589 pbuf_adjcount(&nsw_wcount_async
, n
);
1591 lwkt_reltoken(&vm_token
);
1597 * Assign swap blocks and issue I/O. We reallocate swap on the fly.
1598 * The page is left dirty until the pageout operation completes
1602 for (i
= 0; i
< count
; i
+= n
) {
1609 * Maximum I/O size is limited by a number of factors.
1612 n
= min(BLIST_MAX_ALLOC
, count
- i
);
1613 n
= min(n
, nsw_cluster_max
);
1615 lwkt_gettoken(&vm_token
);
1618 * Get biggest block of swap we can. If we fail, fall
1619 * back and try to allocate a smaller block. Don't go
1620 * overboard trying to allocate space if it would overly
1624 (blk
= swp_pager_getswapspace(object
, n
)) == SWAPBLK_NONE
&&
1629 if (blk
== SWAPBLK_NONE
) {
1630 for (j
= 0; j
< n
; ++j
)
1631 rtvals
[i
+j
] = VM_PAGER_FAIL
;
1632 lwkt_reltoken(&vm_token
);
1635 if (vm_report_swap_allocs
> 0) {
1636 kprintf("swap_alloc %08jx,%d\n", (intmax_t)blk
, n
);
1637 --vm_report_swap_allocs
;
1641 * The I/O we are constructing cannot cross a physical
1642 * disk boundry in the swap stripe.
1644 if ((blk
^ (blk
+ n
)) & ~SWB_DMMASK
) {
1645 j
= ((blk
+ SWB_DMMAX
) & ~SWB_DMMASK
) - blk
;
1646 swp_pager_freeswapspace(object
, blk
+ j
, n
- j
);
1651 * All I/O parameters have been satisfied, build the I/O
1652 * request and assign the swap space.
1654 if ((flags
& VM_PAGER_PUT_SYNC
))
1655 bp
= getpbuf_kva(&nsw_wcount_sync
);
1657 bp
= getpbuf_kva(&nsw_wcount_async
);
1660 lwkt_reltoken(&vm_token
);
1662 pmap_qenter((vm_offset_t
)bp
->b_data
, &m
[i
], n
);
1664 bp
->b_bcount
= PAGE_SIZE
* n
;
1665 bio
->bio_offset
= (off_t
)blk
<< PAGE_SHIFT
;
1667 for (j
= 0; j
< n
; ++j
) {
1668 vm_page_t mreq
= m
[i
+j
];
1670 swp_pager_meta_build(mreq
->object
, mreq
->pindex
,
1672 if (object
->type
== OBJT_SWAP
)
1673 vm_page_dirty(mreq
);
1674 rtvals
[i
+j
] = VM_PAGER_OK
;
1676 vm_page_flag_set(mreq
, PG_SWAPINPROG
);
1677 bp
->b_xio
.xio_pages
[j
] = mreq
;
1679 bp
->b_xio
.xio_npages
= n
;
1681 mycpu
->gd_cnt
.v_swapout
++;
1682 mycpu
->gd_cnt
.v_swappgsout
+= bp
->b_xio
.xio_npages
;
1684 bp
->b_dirtyoff
= 0; /* req'd for NFS */
1685 bp
->b_dirtyend
= bp
->b_bcount
; /* req'd for NFS */
1686 bp
->b_cmd
= BUF_CMD_WRITE
;
1687 bio
->bio_caller_info1
.index
= SWBIO_WRITE
;
1690 /* PMAP TESTING CODE (useful, keep it in but #if 0'd) */
1691 bio
->bio_crc
= iscsi_crc32(bp
->b_data
, bp
->b_bcount
);
1694 for (j
= 0; j
< n
; ++j
) {
1695 vm_page_t mm
= bp
->b_xio
.xio_pages
[j
];
1696 char *p
= (char *)PHYS_TO_DMAP(VM_PAGE_TO_PHYS(mm
));
1697 crc
= iscsi_crc32_ext(p
, PAGE_SIZE
, crc
);
1699 if (bio
->bio_crc
!= crc
) {
1700 kprintf("PREWRITE MISMATCH-A "
1701 "bdata=%08x dmap=%08x bdata=%08x (%d)\n",
1704 iscsi_crc32(bp
->b_data
, bp
->b_bcount
),
1706 #ifdef _KERNEL_VIRTUAL
1707 madvise(bp
->b_data
, bp
->b_bcount
, MADV_INVAL
);
1710 for (j
= 0; j
< n
; ++j
) {
1711 vm_page_t mm
= bp
->b_xio
.xio_pages
[j
];
1712 char *p
= (char *)PHYS_TO_DMAP(VM_PAGE_TO_PHYS(mm
));
1713 crc
= iscsi_crc32_ext(p
, PAGE_SIZE
, crc
);
1715 kprintf("PREWRITE MISMATCH-B "
1716 "bdata=%08x dmap=%08x\n",
1717 iscsi_crc32(bp
->b_data
, bp
->b_bcount
),
1726 if ((flags
& VM_PAGER_PUT_SYNC
) == 0) {
1727 bio
->bio_done
= swp_pager_async_iodone
;
1729 vn_strategy(swapdev_vp
, bio
);
1731 for (j
= 0; j
< n
; ++j
)
1732 rtvals
[i
+j
] = VM_PAGER_PEND
;
1737 * Issue synchrnously.
1739 * Wait for the sync I/O to complete, then update rtvals.
1740 * We just set the rtvals[] to VM_PAGER_PEND so we can call
1741 * our async completion routine at the end, thus avoiding a
1744 bio
->bio_caller_info1
.index
|= SWBIO_SYNC
;
1745 if (flags
& VM_PAGER_TRY_TO_CACHE
)
1746 bio
->bio_caller_info1
.index
|= SWBIO_TTC
;
1747 bio
->bio_done
= biodone_sync
;
1748 bio
->bio_flags
|= BIO_SYNC
;
1749 vn_strategy(swapdev_vp
, bio
);
1750 biowait(bio
, "swwrt");
1752 for (j
= 0; j
< n
; ++j
)
1753 rtvals
[i
+j
] = VM_PAGER_PEND
;
1756 * Now that we are through with the bp, we can call the
1757 * normal async completion, which frees everything up.
1759 swp_pager_async_iodone(bio
);
1761 vm_object_drop(object
);
1767 * Recalculate the low and high-water marks.
1770 swap_pager_newswap(void)
1773 * NOTE: vm_swap_max cannot exceed 1 billion blocks, which is the
1774 * limitation imposed by the blist code. Remember that this
1775 * will be divided by NSWAP_MAX (4), so each swap device is
1776 * limited to around a terrabyte.
1779 nswap_lowat
= (int64_t)vm_swap_max
* 4 / 100; /* 4% left */
1780 nswap_hiwat
= (int64_t)vm_swap_max
* 6 / 100; /* 6% left */
1781 kprintf("swap low/high-water marks set to %d/%d\n",
1782 nswap_lowat
, nswap_hiwat
);
1791 * swp_pager_async_iodone:
1793 * Completion routine for asynchronous reads and writes from/to swap.
1794 * Also called manually by synchronous code to finish up a bp.
1796 * For READ operations, the pages are PG_BUSY'd. For WRITE operations,
1797 * the pages are vm_page_t->busy'd. For READ operations, we PG_BUSY
1798 * unbusy all pages except the 'main' request page. For WRITE
1799 * operations, we vm_page_t->busy'd unbusy all pages ( we can do this
1800 * because we marked them all VM_PAGER_PEND on return from putpages ).
1802 * This routine may not block.
1807 swp_pager_async_iodone(struct bio
*bio
)
1809 struct buf
*bp
= bio
->bio_buf
;
1810 vm_object_t object
= NULL
;
1817 if (bp
->b_flags
& B_ERROR
) {
1819 "swap_pager: I/O error - %s failed; offset %lld,"
1820 "size %ld, error %d\n",
1821 ((bio
->bio_caller_info1
.index
& SWBIO_READ
) ?
1822 "pagein" : "pageout"),
1823 (long long)bio
->bio_offset
,
1832 if (bp
->b_xio
.xio_npages
)
1833 object
= bp
->b_xio
.xio_pages
[0]->object
;
1836 /* PMAP TESTING CODE (useful, keep it in but #if 0'd) */
1837 if (bio
->bio_caller_info1
.index
& SWBIO_WRITE
) {
1838 if (bio
->bio_crc
!= iscsi_crc32(bp
->b_data
, bp
->b_bcount
)) {
1839 kprintf("SWAPOUT: BADCRC %08x %08x\n",
1841 iscsi_crc32(bp
->b_data
, bp
->b_bcount
));
1842 for (i
= 0; i
< bp
->b_xio
.xio_npages
; ++i
) {
1843 vm_page_t m
= bp
->b_xio
.xio_pages
[i
];
1844 if (m
->flags
& PG_WRITEABLE
)
1846 "%d/%d %p writable\n",
1847 i
, bp
->b_xio
.xio_npages
, m
);
1854 * remove the mapping for kernel virtual
1856 pmap_qremove((vm_offset_t
)bp
->b_data
, bp
->b_xio
.xio_npages
);
1859 * cleanup pages. If an error occurs writing to swap, we are in
1860 * very serious trouble. If it happens to be a disk error, though,
1861 * we may be able to recover by reassigning the swap later on. So
1862 * in this case we remove the m->swapblk assignment for the page
1863 * but do not free it in the rlist. The errornous block(s) are thus
1864 * never reallocated as swap. Redirty the page and continue.
1866 for (i
= 0; i
< bp
->b_xio
.xio_npages
; ++i
) {
1867 vm_page_t m
= bp
->b_xio
.xio_pages
[i
];
1869 if (bp
->b_flags
& B_ERROR
) {
1871 * If an error occurs I'd love to throw the swapblk
1872 * away without freeing it back to swapspace, so it
1873 * can never be used again. But I can't from an
1877 if (bio
->bio_caller_info1
.index
& SWBIO_READ
) {
1879 * When reading, reqpage needs to stay
1880 * locked for the parent, but all other
1881 * pages can be freed. We still want to
1882 * wakeup the parent waiting on the page,
1883 * though. ( also: pg_reqpage can be -1 and
1884 * not match anything ).
1886 * We have to wake specifically requested pages
1887 * up too because we cleared PG_SWAPINPROG and
1888 * someone may be waiting for that.
1890 * NOTE: For reads, m->dirty will probably
1891 * be overridden by the original caller
1892 * of getpages so don't play cute tricks
1895 * NOTE: We can't actually free the page from
1896 * here, because this is an interrupt.
1897 * It is not legal to mess with
1898 * object->memq from an interrupt.
1899 * Deactivate the page instead.
1901 * WARNING! The instant PG_SWAPINPROG is
1902 * cleared another cpu may start
1903 * using the mreq page (it will
1904 * check m->valid immediately).
1908 vm_page_flag_clear(m
, PG_SWAPINPROG
);
1911 * bio_driver_info holds the requested page
1914 if (i
!= (int)(intptr_t)bio
->bio_driver_info
) {
1915 vm_page_deactivate(m
);
1921 * If i == bp->b_pager.pg_reqpage, do not wake
1922 * the page up. The caller needs to.
1926 * If a write error occurs remove the swap
1927 * assignment (note that PG_SWAPPED may or
1928 * may not be set depending on prior activity).
1930 * Re-dirty OBJT_SWAP pages as there is no
1931 * other backing store, we can't throw the
1934 * Non-OBJT_SWAP pages (aka swapcache) must
1935 * not be dirtied since they may not have
1936 * been dirty in the first place, and they
1937 * do have backing store (the vnode).
1939 vm_page_busy_wait(m
, FALSE
, "swadpg");
1940 vm_object_hold(m
->object
);
1941 swp_pager_meta_ctl(m
->object
, m
->pindex
,
1943 vm_page_flag_clear(m
, PG_SWAPPED
);
1944 vm_object_drop(m
->object
);
1945 if (m
->object
->type
== OBJT_SWAP
) {
1947 vm_page_activate(m
);
1949 vm_page_io_finish(m
);
1950 vm_page_flag_clear(m
, PG_SWAPINPROG
);
1953 } else if (bio
->bio_caller_info1
.index
& SWBIO_READ
) {
1955 * NOTE: for reads, m->dirty will probably be
1956 * overridden by the original caller of getpages so
1957 * we cannot set them in order to free the underlying
1958 * swap in a low-swap situation. I don't think we'd
1959 * want to do that anyway, but it was an optimization
1960 * that existed in the old swapper for a time before
1961 * it got ripped out due to precisely this problem.
1963 * If not the requested page then deactivate it.
1965 * Note that the requested page, reqpage, is left
1966 * busied, but we still have to wake it up. The
1967 * other pages are released (unbusied) by
1968 * vm_page_wakeup(). We do not set reqpage's
1969 * valid bits here, it is up to the caller.
1973 * NOTE: Can't call pmap_clear_modify(m) from an
1974 * interrupt thread, the pmap code may have to
1975 * map non-kernel pmaps and currently asserts
1978 * WARNING! The instant PG_SWAPINPROG is
1979 * cleared another cpu may start
1980 * using the mreq page (it will
1981 * check m->valid immediately).
1983 /*pmap_clear_modify(m);*/
1984 m
->valid
= VM_PAGE_BITS_ALL
;
1986 vm_page_flag_set(m
, PG_SWAPPED
);
1987 vm_page_flag_clear(m
, PG_SWAPINPROG
);
1990 * We have to wake specifically requested pages
1991 * up too because we cleared PG_SWAPINPROG and
1992 * could be waiting for it in getpages. However,
1993 * be sure to not unbusy getpages specifically
1994 * requested page - getpages expects it to be
1997 * bio_driver_info holds the requested page
1999 if (i
!= (int)(intptr_t)bio
->bio_driver_info
) {
2000 vm_page_deactivate(m
);
2007 * Mark the page clean but do not mess with the
2008 * pmap-layer's modified state. That state should
2009 * also be clear since the caller protected the
2010 * page VM_PROT_READ, but allow the case.
2012 * We are in an interrupt, avoid pmap operations.
2014 * If we have a severe page deficit, deactivate the
2015 * page. Do not try to cache it (which would also
2016 * involve a pmap op), because the page might still
2019 * When using the swap to cache clean vnode pages
2020 * we do not mess with the page dirty bits.
2022 * NOTE! Nobody is waiting for the key mreq page
2023 * on write completion.
2025 vm_page_busy_wait(m
, FALSE
, "swadpg");
2026 if (m
->object
->type
== OBJT_SWAP
)
2028 vm_page_flag_set(m
, PG_SWAPPED
);
2029 vm_page_flag_clear(m
, PG_SWAPINPROG
);
2030 if (vm_page_count_severe())
2031 vm_page_deactivate(m
);
2032 vm_page_io_finish(m
);
2033 if (bio
->bio_caller_info1
.index
& SWBIO_TTC
)
2034 vm_page_try_to_cache(m
);
2041 * adjust pip. NOTE: the original parent may still have its own
2042 * pip refs on the object.
2046 vm_object_pip_wakeup_n(object
, bp
->b_xio
.xio_npages
);
2049 * Release the physical I/O buffer.
2051 * NOTE: Due to synchronous operations in the write case b_cmd may
2052 * already be set to BUF_CMD_DONE and BIO_SYNC may have already
2055 * Use vm_token to interlock nsw_rcount/wcount wakeup?
2057 lwkt_gettoken(&vm_token
);
2058 if (bio
->bio_caller_info1
.index
& SWBIO_READ
)
2059 nswptr
= &nsw_rcount
;
2060 else if (bio
->bio_caller_info1
.index
& SWBIO_SYNC
)
2061 nswptr
= &nsw_wcount_sync
;
2063 nswptr
= &nsw_wcount_async
;
2064 bp
->b_cmd
= BUF_CMD_DONE
;
2065 relpbuf(bp
, nswptr
);
2066 lwkt_reltoken(&vm_token
);
2070 * Fault-in a potentially swapped page and remove the swap reference.
2071 * (used by swapoff code)
2073 * object must be held.
2075 static __inline
void
2076 swp_pager_fault_page(vm_object_t object
, int *sharedp
, vm_pindex_t pindex
)
2082 ASSERT_LWKT_TOKEN_HELD(vm_object_token(object
));
2084 if (object
->type
== OBJT_VNODE
) {
2086 * Any swap related to a vnode is due to swapcache. We must
2087 * vget() the vnode in case it is not active (otherwise
2088 * vref() will panic). Calling vm_object_page_remove() will
2089 * ensure that any swap ref is removed interlocked with the
2090 * page. clean_only is set to TRUE so we don't throw away
2093 vp
= object
->handle
;
2094 error
= vget(vp
, LK_SHARED
| LK_RETRY
| LK_CANRECURSE
);
2096 vm_object_page_remove(object
, pindex
, pindex
+ 1, TRUE
);
2101 * Otherwise it is a normal OBJT_SWAP object and we can
2102 * fault the page in and remove the swap.
2104 m
= vm_fault_object_page(object
, IDX_TO_OFF(pindex
),
2106 VM_FAULT_DIRTY
| VM_FAULT_UNSWAP
,
2114 * This removes all swap blocks related to a particular device. We have
2115 * to be careful of ripups during the scan.
2117 static int swp_pager_swapoff_callback(struct swblock
*swap
, void *data
);
2120 swap_pager_swapoff(int devidx
)
2122 struct vm_object_hash
*hash
;
2123 struct swswapoffinfo info
;
2124 struct vm_object marker
;
2128 bzero(&marker
, sizeof(marker
));
2129 marker
.type
= OBJT_MARKER
;
2131 for (n
= 0; n
< VMOBJ_HSIZE
; ++n
) {
2132 hash
= &vm_object_hash
[n
];
2134 lwkt_gettoken(&hash
->token
);
2135 TAILQ_INSERT_HEAD(&hash
->list
, &marker
, object_list
);
2137 while ((object
= TAILQ_NEXT(&marker
, object_list
)) != NULL
) {
2138 if (object
->type
== OBJT_MARKER
)
2140 if (object
->type
!= OBJT_SWAP
&&
2141 object
->type
!= OBJT_VNODE
)
2143 vm_object_hold(object
);
2144 if (object
->type
!= OBJT_SWAP
&&
2145 object
->type
!= OBJT_VNODE
) {
2146 vm_object_drop(object
);
2149 info
.object
= object
;
2151 info
.devidx
= devidx
;
2152 swblock_rb_tree_RB_SCAN(&object
->swblock_root
,
2153 NULL
, swp_pager_swapoff_callback
,
2155 vm_object_drop(object
);
2157 if (object
== TAILQ_NEXT(&marker
, object_list
)) {
2158 TAILQ_REMOVE(&hash
->list
, &marker
, object_list
);
2159 TAILQ_INSERT_AFTER(&hash
->list
, object
,
2160 &marker
, object_list
);
2163 TAILQ_REMOVE(&hash
->list
, &marker
, object_list
);
2164 lwkt_reltoken(&hash
->token
);
2168 * If we fail to locate all swblocks we just fail gracefully and
2169 * do not bother to restore paging on the swap device. If the
2170 * user wants to retry the user can retry.
2172 if (swdevt
[devidx
].sw_nused
)
2180 swp_pager_swapoff_callback(struct swblock
*swap
, void *data
)
2182 struct swswapoffinfo
*info
= data
;
2183 vm_object_t object
= info
->object
;
2188 index
= swap
->swb_index
;
2189 for (i
= 0; i
< SWAP_META_PAGES
; ++i
) {
2191 * Make sure we don't race a dying object. This will
2192 * kill the scan of the object's swap blocks entirely.
2194 if (object
->flags
& OBJ_DEAD
)
2198 * Fault the page, which can obviously block. If the swap
2199 * structure disappears break out.
2201 v
= swap
->swb_pages
[i
];
2202 if (v
!= SWAPBLK_NONE
&& BLK2DEVIDX(v
) == info
->devidx
) {
2203 swp_pager_fault_page(object
, &info
->shared
,
2204 swap
->swb_index
+ i
);
2205 /* swap ptr might go away */
2206 if (RB_LOOKUP(swblock_rb_tree
,
2207 &object
->swblock_root
, index
) != swap
) {
2215 /************************************************************************
2217 ************************************************************************
2219 * These routines manipulate the swap metadata stored in the
2222 * Swap metadata is implemented with a global hash and not directly
2223 * linked into the object. Instead the object simply contains
2224 * appropriate tracking counters.
2228 * Lookup the swblock containing the specified swap block index.
2230 * The caller must hold the object.
2234 swp_pager_lookup(vm_object_t object
, vm_pindex_t index
)
2236 ASSERT_LWKT_TOKEN_HELD(vm_object_token(object
));
2237 index
&= ~(vm_pindex_t
)SWAP_META_MASK
;
2238 return (RB_LOOKUP(swblock_rb_tree
, &object
->swblock_root
, index
));
2242 * Remove a swblock from the RB tree.
2244 * The caller must hold the object.
2248 swp_pager_remove(vm_object_t object
, struct swblock
*swap
)
2250 ASSERT_LWKT_TOKEN_HELD(vm_object_token(object
));
2251 RB_REMOVE(swblock_rb_tree
, &object
->swblock_root
, swap
);
2255 * Convert default object to swap object if necessary
2257 * The caller must hold the object.
2260 swp_pager_meta_convert(vm_object_t object
)
2262 if (object
->type
== OBJT_DEFAULT
) {
2263 object
->type
= OBJT_SWAP
;
2264 KKASSERT(object
->swblock_count
== 0);
2269 * SWP_PAGER_META_BUILD() - add swap block to swap meta data for object
2271 * We first convert the object to a swap object if it is a default
2272 * object. Vnode objects do not need to be converted.
2274 * The specified swapblk is added to the object's swap metadata. If
2275 * the swapblk is not valid, it is freed instead. Any previously
2276 * assigned swapblk is freed.
2278 * The caller must hold the object.
2281 swp_pager_meta_build(vm_object_t object
, vm_pindex_t index
, swblk_t swapblk
)
2283 struct swblock
*swap
;
2284 struct swblock
*oswap
;
2287 KKASSERT(swapblk
!= SWAPBLK_NONE
);
2288 ASSERT_LWKT_TOKEN_HELD(vm_object_token(object
));
2291 * Convert object if necessary
2293 if (object
->type
== OBJT_DEFAULT
)
2294 swp_pager_meta_convert(object
);
2297 * Locate swblock. If not found create, but if we aren't adding
2298 * anything just return. If we run out of space in the map we wait
2299 * and, since the hash table may have changed, retry.
2302 swap
= swp_pager_lookup(object
, index
);
2307 swap
= zalloc(swap_zone
);
2312 swap
->swb_index
= index
& ~(vm_pindex_t
)SWAP_META_MASK
;
2313 swap
->swb_count
= 0;
2315 ++object
->swblock_count
;
2317 for (i
= 0; i
< SWAP_META_PAGES
; ++i
)
2318 swap
->swb_pages
[i
] = SWAPBLK_NONE
;
2319 oswap
= RB_INSERT(swblock_rb_tree
, &object
->swblock_root
, swap
);
2320 KKASSERT(oswap
== NULL
);
2324 * Delete prior contents of metadata.
2326 * NOTE: Decrement swb_count after the freeing operation (which
2327 * might block) to prevent racing destruction of the swblock.
2329 index
&= SWAP_META_MASK
;
2331 while ((v
= swap
->swb_pages
[index
]) != SWAPBLK_NONE
) {
2332 swap
->swb_pages
[index
] = SWAPBLK_NONE
;
2334 swp_pager_freeswapspace(object
, v
, 1);
2336 --mycpu
->gd_vmtotal
.t_vm
;
2340 * Enter block into metadata
2342 swap
->swb_pages
[index
] = swapblk
;
2343 if (swapblk
!= SWAPBLK_NONE
) {
2345 ++mycpu
->gd_vmtotal
.t_vm
;
2350 * SWP_PAGER_META_FREE() - free a range of blocks in the object's swap metadata
2352 * The requested range of blocks is freed, with any associated swap
2353 * returned to the swap bitmap.
2355 * This routine will free swap metadata structures as they are cleaned
2356 * out. This routine does *NOT* operate on swap metadata associated
2357 * with resident pages.
2359 * The caller must hold the object.
2361 static int swp_pager_meta_free_callback(struct swblock
*swb
, void *data
);
2364 swp_pager_meta_free(vm_object_t object
, vm_pindex_t index
, vm_pindex_t count
)
2366 struct swfreeinfo info
;
2368 ASSERT_LWKT_TOKEN_HELD(vm_object_token(object
));
2373 if (object
->swblock_count
== 0) {
2374 KKASSERT(RB_EMPTY(&object
->swblock_root
));
2381 * Setup for RB tree scan. Note that the pindex range can be huge
2382 * due to the 64 bit page index space so we cannot safely iterate.
2384 info
.object
= object
;
2385 info
.basei
= index
& ~(vm_pindex_t
)SWAP_META_MASK
;
2387 info
.endi
= index
+ count
- 1;
2388 swblock_rb_tree_RB_SCAN(&object
->swblock_root
, rb_swblock_scancmp
,
2389 swp_pager_meta_free_callback
, &info
);
2393 * The caller must hold the object.
2397 swp_pager_meta_free_callback(struct swblock
*swap
, void *data
)
2399 struct swfreeinfo
*info
= data
;
2400 vm_object_t object
= info
->object
;
2405 * Figure out the range within the swblock. The wider scan may
2406 * return edge-case swap blocks when the start and/or end points
2407 * are in the middle of a block.
2409 if (swap
->swb_index
< info
->begi
)
2410 index
= (int)info
->begi
& SWAP_META_MASK
;
2414 if (swap
->swb_index
+ SWAP_META_PAGES
> info
->endi
)
2415 eindex
= (int)info
->endi
& SWAP_META_MASK
;
2417 eindex
= SWAP_META_MASK
;
2420 * Scan and free the blocks. The loop terminates early
2421 * if (swap) runs out of blocks and could be freed.
2423 * NOTE: Decrement swb_count after swp_pager_freeswapspace()
2424 * to deal with a zfree race.
2426 while (index
<= eindex
) {
2427 swblk_t v
= swap
->swb_pages
[index
];
2429 if (v
!= SWAPBLK_NONE
) {
2430 swap
->swb_pages
[index
] = SWAPBLK_NONE
;
2432 swp_pager_freeswapspace(object
, v
, 1);
2433 --mycpu
->gd_vmtotal
.t_vm
;
2434 if (--swap
->swb_count
== 0) {
2435 swp_pager_remove(object
, swap
);
2436 zfree(swap_zone
, swap
);
2437 --object
->swblock_count
;
2444 /* swap may be invalid here due to zfree above */
2451 * SWP_PAGER_META_FREE_ALL() - destroy all swap metadata associated with object
2453 * This routine locates and destroys all swap metadata associated with
2456 * NOTE: Decrement swb_count after the freeing operation (which
2457 * might block) to prevent racing destruction of the swblock.
2459 * The caller must hold the object.
2462 swp_pager_meta_free_all(vm_object_t object
)
2464 struct swblock
*swap
;
2467 ASSERT_LWKT_TOKEN_HELD(vm_object_token(object
));
2469 while ((swap
= RB_ROOT(&object
->swblock_root
)) != NULL
) {
2470 swp_pager_remove(object
, swap
);
2471 for (i
= 0; i
< SWAP_META_PAGES
; ++i
) {
2472 swblk_t v
= swap
->swb_pages
[i
];
2473 if (v
!= SWAPBLK_NONE
) {
2475 swp_pager_freeswapspace(object
, v
, 1);
2477 --mycpu
->gd_vmtotal
.t_vm
;
2480 if (swap
->swb_count
!= 0)
2481 panic("swap_pager_meta_free_all: swb_count != 0");
2482 zfree(swap_zone
, swap
);
2483 --object
->swblock_count
;
2486 KKASSERT(object
->swblock_count
== 0);
2490 * SWP_PAGER_METACTL() - misc control of swap and vm_page_t meta data.
2492 * This routine is capable of looking up, popping, or freeing
2493 * swapblk assignments in the swap meta data or in the vm_page_t.
2494 * The routine typically returns the swapblk being looked-up, or popped,
2495 * or SWAPBLK_NONE if the block was freed, or SWAPBLK_NONE if the block
2496 * was invalid. This routine will automatically free any invalid
2497 * meta-data swapblks.
2499 * It is not possible to store invalid swapblks in the swap meta data
2500 * (other then a literal 'SWAPBLK_NONE'), so we don't bother checking.
2502 * When acting on a busy resident page and paging is in progress, we
2503 * have to wait until paging is complete but otherwise can act on the
2506 * SWM_FREE remove and free swap block from metadata
2507 * SWM_POP remove from meta data but do not free.. pop it out
2509 * The caller must hold the object.
2512 swp_pager_meta_ctl(vm_object_t object
, vm_pindex_t index
, int flags
)
2514 struct swblock
*swap
;
2517 if (object
->swblock_count
== 0)
2518 return(SWAPBLK_NONE
);
2521 swap
= swp_pager_lookup(object
, index
);
2524 index
&= SWAP_META_MASK
;
2525 r1
= swap
->swb_pages
[index
];
2527 if (r1
!= SWAPBLK_NONE
) {
2528 if (flags
& (SWM_FREE
|SWM_POP
)) {
2529 swap
->swb_pages
[index
] = SWAPBLK_NONE
;
2530 --mycpu
->gd_vmtotal
.t_vm
;
2531 if (--swap
->swb_count
== 0) {
2532 swp_pager_remove(object
, swap
);
2533 zfree(swap_zone
, swap
);
2534 --object
->swblock_count
;
2537 /* swap ptr may be invalid */
2538 if (flags
& SWM_FREE
) {
2539 swp_pager_freeswapspace(object
, r1
, 1);
2543 /* swap ptr may be invalid */