2 * Copyright (c) 1998,2004 The DragonFly Project. All rights reserved.
4 * This code is derived from software contributed to The DragonFly Project
5 * by Matthew Dillon <dillon@backplane.com>
7 * Redistribution and use in source and binary forms, with or without
8 * modification, are permitted provided that the following conditions
11 * 1. Redistributions of source code must retain the above copyright
12 * notice, this list of conditions and the following disclaimer.
13 * 2. Redistributions in binary form must reproduce the above copyright
14 * notice, this list of conditions and the following disclaimer in
15 * the documentation and/or other materials provided with the
17 * 3. Neither the name of The DragonFly Project nor the names of its
18 * contributors may be used to endorse or promote products derived
19 * from this software without specific, prior written permission.
21 * THIS SOFTWARE IS PROVIDED BY THE COPYRIGHT HOLDERS AND CONTRIBUTORS
22 * ``AS IS'' AND ANY EXPRESS OR IMPLIED WARRANTIES, INCLUDING, BUT NOT
23 * LIMITED TO, THE IMPLIED WARRANTIES OF MERCHANTABILITY AND FITNESS
24 * FOR A PARTICULAR PURPOSE ARE DISCLAIMED. IN NO EVENT SHALL THE
25 * COPYRIGHT HOLDERS OR CONTRIBUTORS BE LIABLE FOR ANY DIRECT, INDIRECT,
26 * INCIDENTAL, SPECIAL, EXEMPLARY OR CONSEQUENTIAL DAMAGES (INCLUDING,
27 * BUT NOT LIMITED TO, PROCUREMENT OF SUBSTITUTE GOODS OR SERVICES;
28 * LOSS OF USE, DATA, OR PROFITS; OR BUSINESS INTERRUPTION) HOWEVER CAUSED
29 * AND ON ANY THEORY OF LIABILITY, WHETHER IN CONTRACT, STRICT LIABILITY,
30 * OR TORT (INCLUDING NEGLIGENCE OR OTHERWISE) ARISING IN ANY WAY OUT
31 * OF THE USE OF THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF
34 * Copyright (c) 1994 John S. Dyson
35 * Copyright (c) 1990 University of Utah.
36 * Copyright (c) 1991, 1993
37 * The Regents of the University of California. All rights reserved.
39 * This code is derived from software contributed to Berkeley by
40 * the Systems Programming Group of the University of Utah Computer
43 * Redistribution and use in source and binary forms, with or without
44 * modification, are permitted provided that the following conditions
46 * 1. Redistributions of source code must retain the above copyright
47 * notice, this list of conditions and the following disclaimer.
48 * 2. Redistributions in binary form must reproduce the above copyright
49 * notice, this list of conditions and the following disclaimer in the
50 * documentation and/or other materials provided with the distribution.
51 * 3. All advertising materials mentioning features or use of this software
52 * must display the following acknowledgement:
53 * This product includes software developed by the University of
54 * California, Berkeley and its contributors.
55 * 4. Neither the name of the University nor the names of its contributors
56 * may be used to endorse or promote products derived from this software
57 * without specific prior written permission.
59 * THIS SOFTWARE IS PROVIDED BY THE REGENTS AND CONTRIBUTORS ``AS IS'' AND
60 * ANY EXPRESS OR IMPLIED WARRANTIES, INCLUDING, BUT NOT LIMITED TO, THE
61 * IMPLIED WARRANTIES OF MERCHANTABILITY AND FITNESS FOR A PARTICULAR PURPOSE
62 * ARE DISCLAIMED. IN NO EVENT SHALL THE REGENTS OR CONTRIBUTORS BE LIABLE
63 * FOR ANY DIRECT, INDIRECT, INCIDENTAL, SPECIAL, EXEMPLARY, OR CONSEQUENTIAL
64 * DAMAGES (INCLUDING, BUT NOT LIMITED TO, PROCUREMENT OF SUBSTITUTE GOODS
65 * OR SERVICES; LOSS OF USE, DATA, OR PROFITS; OR BUSINESS INTERRUPTION)
66 * HOWEVER CAUSED AND ON ANY THEORY OF LIABILITY, WHETHER IN CONTRACT, STRICT
67 * LIABILITY, OR TORT (INCLUDING NEGLIGENCE OR OTHERWISE) ARISING IN ANY WAY
68 * OUT OF THE USE OF THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF
74 * Radix Bitmap 'blists'.
76 * - The new swapper uses the new radix bitmap code. This should scale
77 * to arbitrarily small or arbitrarily large swap spaces and an almost
78 * arbitrary degree of fragmentation.
82 * - on the fly reallocation of swap during putpages. The new system
83 * does not try to keep previously allocated swap blocks for dirty
86 * - on the fly deallocation of swap
88 * - No more garbage collection required. Unnecessarily allocated swap
89 * blocks only exist for dirty vm_page_t's now and these are already
90 * cycled (in a high-load system) by the pager. We also do on-the-fly
91 * removal of invalidated swap blocks when a page is destroyed
94 * from: Utah $Hdr: swap_pager.c 1.4 91/04/30$
96 * @(#)swap_pager.c 8.9 (Berkeley) 3/21/94
98 * $FreeBSD: src/sys/vm/swap_pager.c,v 1.130.2.12 2002/08/31 21:15:55 dillon Exp $
99 * $DragonFly: src/sys/vm/swap_pager.c,v 1.32 2008/07/01 02:02:56 dillon Exp $
102 #include <sys/param.h>
103 #include <sys/systm.h>
104 #include <sys/conf.h>
105 #include <sys/kernel.h>
106 #include <sys/proc.h>
108 #include <sys/vnode.h>
109 #include <sys/malloc.h>
110 #include <sys/vmmeter.h>
111 #include <sys/sysctl.h>
112 #include <sys/blist.h>
113 #include <sys/lock.h>
114 #include <sys/thread2.h>
116 #ifndef MAX_PAGEOUT_CLUSTER
117 #define MAX_PAGEOUT_CLUSTER 16
120 #define SWB_NPAGES MAX_PAGEOUT_CLUSTER
122 #include "opt_swap.h"
124 #include <vm/vm_object.h>
125 #include <vm/vm_page.h>
126 #include <vm/vm_pager.h>
127 #include <vm/vm_pageout.h>
128 #include <vm/swap_pager.h>
129 #include <vm/vm_extern.h>
130 #include <vm/vm_zone.h>
131 #include <vm/vnode_pager.h>
133 #include <sys/buf2.h>
134 #include <vm/vm_page2.h>
136 #define SWM_FREE 0x02 /* free, period */
137 #define SWM_POP 0x04 /* pop out */
139 #define SWBIO_READ 0x01
140 #define SWBIO_WRITE 0x02
141 #define SWBIO_SYNC 0x04
144 * vm_swap_size is in page-sized chunks now. It was DEV_BSIZE'd chunks
148 extern int vm_swap_size; /* number of free swap blocks, in pages */
150 int swap_pager_full; /* swap space exhaustion (task killing) */
151 static int swap_pager_almost_full; /* swap space exhaustion (w/ hysteresis)*/
152 static int nsw_rcount; /* free read buffers */
153 static int nsw_wcount_sync; /* limit write buffers / synchronous */
154 static int nsw_wcount_async; /* limit write buffers / asynchronous */
155 static int nsw_wcount_async_max;/* assigned maximum */
156 static int nsw_cluster_max; /* maximum VOP I/O allowed */
158 struct blist *swapblist;
159 static int swap_async_max = 4; /* maximum in-progress async I/O's */
160 static int swap_burst_read = 0; /* allow burst reading */
162 extern struct vnode *swapdev_vp; /* from vm_swap.c */
164 SYSCTL_INT(_vm, OID_AUTO, swap_async_max,
165 CTLFLAG_RW, &swap_async_max, 0, "Maximum running async swap ops");
166 SYSCTL_INT(_vm, OID_AUTO, swap_burst_read,
167 CTLFLAG_RW, &swap_burst_read, 0, "Allow burst reads for pageins");
172 * Red-Black tree for swblock entries
174 RB_GENERATE2(swblock_rb_tree, swblock, swb_entry, rb_swblock_compare,
175 vm_pindex_t, swb_index);
178 rb_swblock_compare(struct swblock *swb1, struct swblock *swb2)
180 if (swb1->swb_index < swb2->swb_index)
182 if (swb1->swb_index > swb2->swb_index)
188 * pagerops for OBJT_SWAP - "swap pager". Some ops are also global procedure
189 * calls hooked from other parts of the VM system and do not appear here.
190 * (see vm/swap_pager.h).
194 swap_pager_alloc (void *handle, off_t size,
195 vm_prot_t prot, off_t offset);
196 static void swap_pager_dealloc (vm_object_t object);
197 static int swap_pager_getpage (vm_object_t, vm_page_t *, int);
198 static void swap_chain_iodone(struct bio *biox);
200 struct pagerops swappagerops = {
201 swap_pager_alloc, /* allocate an OBJT_SWAP object */
202 swap_pager_dealloc, /* deallocate an OBJT_SWAP object */
203 swap_pager_getpage, /* pagein */
204 swap_pager_putpages, /* pageout */
205 swap_pager_haspage /* get backing store status for page */
209 * dmmax is in page-sized chunks with the new swap system. It was
210 * dev-bsized chunks in the old. dmmax is always a power of 2.
212 * swap_*() routines are externally accessible. swp_*() routines are
217 static int dmmax_mask;
218 int nswap_lowat = 128; /* in pages, swap_pager_almost_full warn */
219 int nswap_hiwat = 512; /* in pages, swap_pager_almost_full warn */
221 static __inline void swp_sizecheck (void);
222 static void swp_pager_async_iodone (struct bio *bio);
225 * Swap bitmap functions
228 static __inline void swp_pager_freeswapspace (daddr_t blk, int npages);
229 static __inline daddr_t swp_pager_getswapspace (int npages);
235 static void swp_pager_meta_convert (vm_object_t);
236 static void swp_pager_meta_build (vm_object_t, vm_pindex_t, daddr_t);
237 static void swp_pager_meta_free (vm_object_t, vm_pindex_t, daddr_t);
238 static void swp_pager_meta_free_all (vm_object_t);
239 static daddr_t swp_pager_meta_ctl (vm_object_t, vm_pindex_t, int);
242 * SWP_SIZECHECK() - update swap_pager_full indication
244 * update the swap_pager_almost_full indication and warn when we are
245 * about to run out of swap space, using lowat/hiwat hysteresis.
247 * Clear swap_pager_full ( task killing ) indication when lowat is met.
249 * No restrictions on call
250 * This routine may not block.
251 * This routine must be called at splvm()
257 if (vm_swap_size < nswap_lowat) {
258 if (swap_pager_almost_full == 0) {
259 kprintf("swap_pager: out of swap space\n");
260 swap_pager_almost_full = 1;
264 if (vm_swap_size > nswap_hiwat)
265 swap_pager_almost_full = 0;
270 * SWAP_PAGER_INIT() - initialize the swap pager!
272 * Expected to be started from system init. NOTE: This code is run
273 * before much else so be careful what you depend on. Most of the VM
274 * system has yet to be initialized at this point.
277 swap_pager_init(void *arg __unused)
280 * Device Stripe, in PAGE_SIZE'd blocks
282 dmmax = SWB_NPAGES * 2;
283 dmmax_mask = ~(dmmax - 1);
285 SYSINIT(vm_mem, SI_BOOT1_VM, SI_ORDER_THIRD, swap_pager_init, NULL)
288 * SWAP_PAGER_SWAP_INIT() - swap pager initialization from pageout process
290 * Expected to be started from pageout process once, prior to entering
295 swap_pager_swap_init(void)
300 * Number of in-transit swap bp operations. Don't
301 * exhaust the pbufs completely. Make sure we
302 * initialize workable values (0 will work for hysteresis
303 * but it isn't very efficient).
305 * The nsw_cluster_max is constrained by the number of pages an XIO
306 * holds, i.e., (MAXPHYS/PAGE_SIZE) and our locally defined
307 * MAX_PAGEOUT_CLUSTER. Also be aware that swap ops are
308 * constrained by the swap device interleave stripe size.
310 * Currently we hardwire nsw_wcount_async to 4. This limit is
311 * designed to prevent other I/O from having high latencies due to
312 * our pageout I/O. The value 4 works well for one or two active swap
313 * devices but is probably a little low if you have more. Even so,
314 * a higher value would probably generate only a limited improvement
315 * with three or four active swap devices since the system does not
316 * typically have to pageout at extreme bandwidths. We will want
317 * at least 2 per swap devices, and 4 is a pretty good value if you
318 * have one NFS swap device due to the command/ack latency over NFS.
319 * So it all works out pretty well.
322 nsw_cluster_max = min((MAXPHYS/PAGE_SIZE), MAX_PAGEOUT_CLUSTER);
324 nsw_rcount = (nswbuf + 1) / 2;
325 nsw_wcount_sync = (nswbuf + 3) / 4;
326 nsw_wcount_async = 4;
327 nsw_wcount_async_max = nsw_wcount_async;
330 * The zone is dynamically allocated so generally size it to
331 * maxswzone (32MB to 512MB of KVM). Set a minimum size based
332 * on physical memory of around 8x (each swblock can hold 16 pages).
334 * With the advent of SSDs (vs HDs) the practical (swap:memory) ratio
335 * has increased dramatically.
337 n = vmstats.v_page_count / 2;
338 if (maxswzone && n < maxswzone / sizeof(struct swblock))
339 n = maxswzone / sizeof(struct swblock);
345 sizeof(struct swblock),
349 if (swap_zone != NULL)
352 * if the allocation failed, try a zone two thirds the
353 * size of the previous attempt.
358 if (swap_zone == NULL)
359 panic("swap_pager_swap_init: swap_zone == NULL");
361 kprintf("Swap zone entries reduced from %d to %d.\n", n2, n);
365 * SWAP_PAGER_ALLOC() - allocate a new OBJT_SWAP VM object and instantiate
366 * its metadata structures.
368 * This routine is called from the mmap and fork code to create a new
369 * OBJT_SWAP object. We do this by creating an OBJT_DEFAULT object
370 * and then converting it with swp_pager_meta_convert().
372 * This routine may block in vm_object_allocate() and create a named
373 * object lookup race, so we must interlock. We must also run at
374 * splvm() for the object lookup to handle races with interrupts, but
375 * we do not have to maintain splvm() in between the lookup and the
376 * add because (I believe) it is not possible to attempt to create
377 * a new swap object w/handle when a default object with that handle
382 swap_pager_alloc(void *handle, off_t size, vm_prot_t prot, off_t offset)
386 KKASSERT(handle == NULL);
390 * Reference existing named region or allocate new one. There
391 * should not be a race here against swp_pager_meta_build()
392 * as called from vm_page_remove() in regards to the lookup
395 while (sw_alloc_interlock) {
396 sw_alloc_interlock = -1;
397 tsleep(&sw_alloc_interlock, 0, "swpalc", 0);
399 sw_alloc_interlock = 1;
401 object = vm_pager_object_lookup(NOBJLIST(handle), handle);
403 if (object != NULL) {
404 vm_object_reference(object);
406 object = vm_object_allocate(OBJT_DEFAULT,
407 OFF_TO_IDX(offset + PAGE_MASK + size));
408 object->handle = handle;
409 swp_pager_meta_convert(object);
412 if (sw_alloc_interlock < 0)
413 wakeup(&sw_alloc_interlock);
414 sw_alloc_interlock = 0;
417 object = vm_object_allocate(OBJT_DEFAULT,
418 OFF_TO_IDX(offset + PAGE_MASK + size));
419 swp_pager_meta_convert(object);
425 * SWAP_PAGER_DEALLOC() - remove swap metadata from object
427 * The swap backing for the object is destroyed. The code is
428 * designed such that we can reinstantiate it later, but this
429 * routine is typically called only when the entire object is
430 * about to be destroyed.
432 * This routine may block, but no longer does.
434 * The object must be locked or unreferenceable.
438 swap_pager_dealloc(vm_object_t object)
440 vm_object_pip_wait(object, "swpdea");
443 * Free all remaining metadata. We only bother to free it from
444 * the swap meta data. We do not attempt to free swapblk's still
445 * associated with vm_page_t's for this object. We do not care
446 * if paging is still in progress on some objects.
449 swp_pager_meta_free_all(object);
453 /************************************************************************
454 * SWAP PAGER BITMAP ROUTINES *
455 ************************************************************************/
458 * SWP_PAGER_GETSWAPSPACE() - allocate raw swap space
460 * Allocate swap for the requested number of pages. The starting
461 * swap block number (a page index) is returned or SWAPBLK_NONE
462 * if the allocation failed.
464 * Also has the side effect of advising that somebody made a mistake
465 * when they configured swap and didn't configure enough.
467 * Must be called at splvm() to avoid races with bitmap frees from
468 * vm_page_remove() aka swap_pager_page_removed().
470 * This routine may not block
471 * This routine must be called at splvm().
474 static __inline daddr_t
475 swp_pager_getswapspace(int npages)
479 if ((blk = blist_alloc(swapblist, npages)) == SWAPBLK_NONE) {
480 if (swap_pager_full != 2) {
481 kprintf("swap_pager_getswapspace: failed\n");
483 swap_pager_almost_full = 1;
486 vm_swap_size -= npages;
493 * SWP_PAGER_FREESWAPSPACE() - free raw swap space
495 * This routine returns the specified swap blocks back to the bitmap.
497 * Note: This routine may not block (it could in the old swap code),
498 * and through the use of the new blist routines it does not block.
500 * We must be called at splvm() to avoid races with bitmap frees from
501 * vm_page_remove() aka swap_pager_page_removed().
503 * This routine may not block
504 * This routine must be called at splvm().
508 swp_pager_freeswapspace(daddr_t blk, int npages)
510 blist_free(swapblist, blk, npages);
511 vm_swap_size += npages;
516 * SWAP_PAGER_FREESPACE() - frees swap blocks associated with a page
517 * range within an object.
519 * This is a globally accessible routine.
521 * This routine removes swapblk assignments from swap metadata.
523 * The external callers of this routine typically have already destroyed
524 * or renamed vm_page_t's associated with this range in the object so
527 * This routine may be called at any spl. We up our spl to splvm
528 * temporarily in order to perform the metadata removal.
531 swap_pager_freespace(vm_object_t object, vm_pindex_t start, vm_size_t size)
534 swp_pager_meta_free(object, start, size);
539 * SWAP_PAGER_RESERVE() - reserve swap blocks in object
541 * Assigns swap blocks to the specified range within the object. The
542 * swap blocks are not zerod. Any previous swap assignment is destroyed.
544 * Returns 0 on success, -1 on failure.
547 swap_pager_reserve(vm_object_t object, vm_pindex_t start, vm_size_t size)
550 daddr_t blk = SWAPBLK_NONE;
551 vm_pindex_t beg = start; /* save start index */
557 while ((blk = swp_pager_getswapspace(n)) == SWAPBLK_NONE) {
560 swp_pager_meta_free(object, beg, start - beg);
566 swp_pager_meta_build(object, start, blk);
572 swp_pager_meta_free(object, start, n);
578 * SWAP_PAGER_COPY() - copy blocks from source pager to destination pager
579 * and destroy the source.
581 * Copy any valid swapblks from the source to the destination. In
582 * cases where both the source and destination have a valid swapblk,
583 * we keep the destination's.
585 * This routine is allowed to block. It may block allocating metadata
586 * indirectly through swp_pager_meta_build() or if paging is still in
587 * progress on the source.
589 * This routine can be called at any spl
591 * XXX vm_page_collapse() kinda expects us not to block because we
592 * supposedly do not need to allocate memory, but for the moment we
593 * *may* have to get a little memory from the zone allocator, but
594 * it is taken from the interrupt memory. We should be ok.
596 * The source object contains no vm_page_t's (which is just as well)
598 * The source object is of type OBJT_SWAP.
600 * The source and destination objects must be locked or
601 * inaccessible (XXX are they ?)
605 swap_pager_copy(vm_object_t srcobject, vm_object_t dstobject,
606 vm_pindex_t offset, int destroysource)
613 * transfer source to destination.
616 for (i = 0; i < dstobject->size; ++i) {
620 * Locate (without changing) the swapblk on the destination,
621 * unless it is invalid in which case free it silently, or
622 * if the destination is a resident page, in which case the
623 * source is thrown away.
626 dstaddr = swp_pager_meta_ctl(dstobject, i, 0);
628 if (dstaddr == SWAPBLK_NONE) {
630 * Destination has no swapblk and is not resident,
635 srcaddr = swp_pager_meta_ctl(
641 if (srcaddr != SWAPBLK_NONE)
642 swp_pager_meta_build(dstobject, i, srcaddr);
645 * Destination has valid swapblk or it is represented
646 * by a resident page. We destroy the sourceblock.
649 swp_pager_meta_ctl(srcobject, i + offset, SWM_FREE);
654 * Free left over swap blocks in source.
656 * We have to revert the type to OBJT_DEFAULT so we do not accidently
657 * double-remove the object from the swap queues.
662 * Reverting the type is not necessary, the caller is going
663 * to destroy srcobject directly, but I'm doing it here
664 * for consistency since we've removed the object from its
667 swp_pager_meta_free_all(srcobject);
668 srcobject->type = OBJT_DEFAULT;
674 * SWAP_PAGER_HASPAGE() - determine if we have good backing store for
675 * the requested page.
677 * We determine whether good backing store exists for the requested
678 * page and return TRUE if it does, FALSE if it doesn't.
680 * If TRUE, we also try to determine how much valid, contiguous backing
681 * store exists before and after the requested page within a reasonable
682 * distance. We do not try to restrict it to the swap device stripe
683 * (that is handled in getpages/putpages). It probably isn't worth
688 swap_pager_haspage(vm_object_t object, vm_pindex_t pindex)
693 * do we have good backing store at the requested index ?
697 blk0 = swp_pager_meta_ctl(object, pindex, 0);
699 if (blk0 == SWAPBLK_NONE) {
706 * find backwards-looking contiguous good backing store
708 if (before != NULL) {
711 for (i = 1; i < (SWB_NPAGES/2); ++i) {
716 blk = swp_pager_meta_ctl(object, pindex - i, 0);
724 * find forward-looking contiguous good backing store
730 for (i = 1; i < (SWB_NPAGES/2); ++i) {
733 blk = swp_pager_meta_ctl(object, pindex + i, 0);
745 * SWAP_PAGER_PAGE_UNSWAPPED() - remove swap backing store related to page
747 * This removes any associated swap backing store, whether valid or
748 * not, from the page. This operates on any VM object, not just OBJT_SWAP
751 * This routine is typically called when a page is made dirty, at
752 * which point any associated swap can be freed. MADV_FREE also
753 * calls us in a special-case situation
755 * NOTE!!! If the page is clean and the swap was valid, the caller
756 * should make the page dirty before calling this routine. This routine
757 * does NOT change the m->dirty status of the page. Also: MADV_FREE
760 * This routine may not block
761 * This routine must be called at splvm()
764 swap_pager_unswapped(vm_page_t m)
766 swp_pager_meta_ctl(m->object, m->pindex, SWM_FREE);
770 * SWAP_PAGER_STRATEGY() - read, write, free blocks
772 * This implements a VM OBJECT strategy function using swap backing store.
773 * This can operate on any VM OBJECT type, not necessarily just OBJT_SWAP
776 * This is intended to be a cacheless interface (i.e. caching occurs at
777 * higher levels), and is also used as a swap-based SSD cache for vnode
778 * and device objects.
780 * All I/O goes directly to and from the swap device.
782 * We currently attempt to run I/O synchronously or asynchronously as
783 * the caller requests. This isn't perfect because we loose error
784 * sequencing when we run multiple ops in parallel to satisfy a request.
785 * But this is swap, so we let it all hang out.
788 swap_pager_strategy(vm_object_t object, struct bio *bio)
790 struct buf *bp = bio->bio_buf;
793 vm_pindex_t biox_blkno = 0;
798 struct bio_track *track;
801 * tracking for swapdev vnode I/Os
803 if (bp->b_cmd == BUF_CMD_READ)
804 track = &swapdev_vp->v_track_read;
806 track = &swapdev_vp->v_track_write;
808 if (bp->b_bcount & PAGE_MASK) {
809 bp->b_error = EINVAL;
810 bp->b_flags |= B_ERROR | B_INVAL;
812 kprintf("swap_pager_strategy: bp %p offset %lld size %d, "
813 "not page bounded\n",
814 bp, (long long)bio->bio_offset, (int)bp->b_bcount);
819 * Clear error indication, initialize page index, count, data pointer.
822 bp->b_flags &= ~B_ERROR;
823 bp->b_resid = bp->b_bcount;
825 start = (vm_pindex_t)(bio->bio_offset >> PAGE_SHIFT);
826 count = howmany(bp->b_bcount, PAGE_SIZE);
830 * Deal with BUF_CMD_FREEBLKS
832 if (bp->b_cmd == BUF_CMD_FREEBLKS) {
834 * FREE PAGE(s) - destroy underlying swap that is no longer
837 swp_pager_meta_free(object, start, count);
844 * We need to be able to create a new cluster of I/O's. We cannot
845 * use the caller fields of the passed bio so push a new one.
847 * Because nbio is just a placeholder for the cluster links,
848 * we can biodone() the original bio instead of nbio to make
849 * things a bit more efficient.
851 nbio = push_bio(bio);
852 nbio->bio_offset = bio->bio_offset;
853 nbio->bio_caller_info1.cluster_head = NULL;
854 nbio->bio_caller_info2.cluster_tail = NULL;
860 * Execute read or write
866 * Obtain block. If block not found and writing, allocate a
867 * new block and build it into the object.
869 blk = swp_pager_meta_ctl(object, start, 0);
870 if ((blk == SWAPBLK_NONE) && bp->b_cmd != BUF_CMD_READ) {
871 blk = swp_pager_getswapspace(1);
872 if (blk == SWAPBLK_NONE) {
873 bp->b_error = ENOMEM;
874 bp->b_flags |= B_ERROR;
877 swp_pager_meta_build(object, start, blk);
881 * Do we have to flush our current collection? Yes if:
883 * - no swap block at this index
884 * - swap block is not contiguous
885 * - we cross a physical disk boundry in the
889 biox && (biox_blkno + btoc(bufx->b_bcount) != blk ||
890 ((biox_blkno ^ blk) & dmmax_mask)
893 if (bp->b_cmd == BUF_CMD_READ) {
894 ++mycpu->gd_cnt.v_swapin;
895 mycpu->gd_cnt.v_swappgsin += btoc(bufx->b_bcount);
897 ++mycpu->gd_cnt.v_swapout;
898 mycpu->gd_cnt.v_swappgsout += btoc(bufx->b_bcount);
899 bufx->b_dirtyend = bufx->b_bcount;
903 * Finished with this buf.
905 KKASSERT(bufx->b_bcount != 0);
906 if (bufx->b_cmd != BUF_CMD_READ)
907 bufx->b_dirtyend = bufx->b_bcount;
913 * Add new swapblk to biox, instantiating biox if necessary.
914 * Zero-fill reads are able to take a shortcut.
916 if (blk == SWAPBLK_NONE) {
918 * We can only get here if we are reading. Since
919 * we are at splvm() we can safely modify b_resid,
920 * even if chain ops are in progress.
922 bzero(data, PAGE_SIZE);
923 bp->b_resid -= PAGE_SIZE;
926 /* XXX chain count > 4, wait to <= 4 */
928 bufx = getpbuf(NULL);
929 biox = &bufx->b_bio1;
930 cluster_append(nbio, bufx);
931 bufx->b_flags |= (bufx->b_flags & B_ORDERED);
932 bufx->b_cmd = bp->b_cmd;
933 biox->bio_done = swap_chain_iodone;
934 biox->bio_offset = (off_t)blk << PAGE_SHIFT;
935 biox->bio_caller_info1.cluster_parent = nbio;
940 bufx->b_bcount += PAGE_SIZE;
948 * Flush out last buffer
951 if (bufx->b_cmd == BUF_CMD_READ) {
952 ++mycpu->gd_cnt.v_swapin;
953 mycpu->gd_cnt.v_swappgsin += btoc(bufx->b_bcount);
955 ++mycpu->gd_cnt.v_swapout;
956 mycpu->gd_cnt.v_swappgsout += btoc(bufx->b_bcount);
957 bufx->b_dirtyend = bufx->b_bcount;
959 KKASSERT(bufx->b_bcount);
960 if (bufx->b_cmd != BUF_CMD_READ)
961 bufx->b_dirtyend = bufx->b_bcount;
962 /* biox, bufx = NULL */
966 * Now initiate all the I/O. Be careful looping on our chain as
967 * I/O's may complete while we are still initiating them.
969 nbio->bio_caller_info2.cluster_tail = NULL;
970 bufx = nbio->bio_caller_info1.cluster_head;
973 biox = &bufx->b_bio1;
975 bufx = bufx->b_cluster_next;
976 vn_strategy(swapdev_vp, biox);
980 * Completion of the cluster will also call biodone_chain(nbio).
981 * We never call biodone(nbio) so we don't have to worry about
982 * setting up a bio_done callback. It's handled in the sub-IO.
988 swap_chain_iodone(struct bio *biox)
991 struct buf *bufx; /* chained sub-buffer */
992 struct bio *nbio; /* parent nbio with chain glue */
993 struct buf *bp; /* original bp associated with nbio */
996 bufx = biox->bio_buf;
997 nbio = biox->bio_caller_info1.cluster_parent;
1001 * Update the original buffer
1003 KKASSERT(bp != NULL);
1004 if (bufx->b_flags & B_ERROR) {
1005 atomic_set_int(&bufx->b_flags, B_ERROR);
1006 bp->b_error = bufx->b_error;
1007 } else if (bufx->b_resid != 0) {
1008 atomic_set_int(&bufx->b_flags, B_ERROR);
1009 bp->b_error = EINVAL;
1011 atomic_subtract_int(&bp->b_resid, bufx->b_bcount);
1015 * Remove us from the chain.
1017 spin_lock_wr(&bp->b_lock.lk_spinlock);
1018 nextp = &nbio->bio_caller_info1.cluster_head;
1019 while (*nextp != bufx) {
1020 KKASSERT(*nextp != NULL);
1021 nextp = &(*nextp)->b_cluster_next;
1023 *nextp = bufx->b_cluster_next;
1024 chain_empty = (nbio->bio_caller_info1.cluster_head == NULL);
1025 spin_unlock_wr(&bp->b_lock.lk_spinlock);
1028 * Clean up bufx. If the chain is now empty we finish out
1029 * the parent. Note that we may be racing other completions
1030 * so we must use the chain_empty status from above.
1033 if (bp->b_resid != 0 && !(bp->b_flags & B_ERROR)) {
1034 atomic_set_int(&bp->b_flags, B_ERROR);
1035 bp->b_error = EINVAL;
1037 biodone_chain(nbio);
1039 relpbuf(bufx, NULL);
1043 * SWAP_PAGER_GETPAGES() - bring page in from swap
1045 * The requested page may have to be brought in from swap. Calculate the
1046 * swap block and bring in additional pages if possible. All pages must
1047 * have contiguous swap block assignments and reside in the same object.
1049 * The caller has a single vm_object_pip_add() reference prior to
1050 * calling us and we should return with the same.
1052 * The caller has BUSY'd the page. We should return with (*mpp) left busy,
1053 * and any additinal pages unbusied.
1055 * If the caller encounters a PG_RAM page it will pass it to us even though
1056 * it may be valid and dirty. We cannot overwrite the page in this case!
1057 * The case is used to allow us to issue pure read-aheads.
1059 * NOTE! XXX This code does not entirely pipeline yet due to the fact that
1060 * the PG_RAM page is validated at the same time as mreq. What we
1061 * really need to do is issue a separate read-ahead pbuf.
1064 swap_pager_getpage(vm_object_t object, vm_page_t *mpp, int seqaccess)
1075 vm_page_t marray[XIO_INTERNAL_PAGES];
1079 if (mreq->object != object) {
1080 panic("swap_pager_getpages: object mismatch %p/%p",
1087 * We don't want to overwrite a fully valid page as it might be
1088 * dirty. This case can occur when e.g. vm_fault hits a perfectly
1089 * valid page with PG_RAM set.
1091 * In this case we see if the next page is a suitable page-in
1092 * candidate and if it is we issue read-ahead. PG_RAM will be
1093 * set on the last page of the read-ahead to continue the pipeline.
1095 if (mreq->valid == VM_PAGE_BITS_ALL) {
1096 if (swap_burst_read == 0 || mreq->pindex + 1 >= object->size)
1097 return(VM_PAGER_OK);
1099 blk = swp_pager_meta_ctl(object, mreq->pindex + 1, 0);
1100 if (blk == SWAPBLK_NONE) {
1102 return(VM_PAGER_OK);
1104 m = vm_page_lookup(object, mreq->pindex + 1);
1106 m = vm_page_alloc(object, mreq->pindex + 1,
1110 return(VM_PAGER_OK);
1113 if ((m->flags & PG_BUSY) || m->busy || m->valid) {
1115 return(VM_PAGER_OK);
1117 vm_page_unqueue_nowakeup(m);
1128 * Try to block-read contiguous pages from swap if sequential,
1129 * otherwise just read one page. Contiguous pages from swap must
1130 * reside within a single device stripe because the I/O cannot be
1131 * broken up across multiple stripes.
1133 * Note that blk and iblk can be SWAPBLK_NONE but the loop is
1134 * set up such that the case(s) are handled implicitly.
1137 blk = swp_pager_meta_ctl(mreq->object, mreq->pindex, 0);
1140 for (i = 1; swap_burst_read &&
1141 i < XIO_INTERNAL_PAGES &&
1142 mreq->pindex + i < object->size; ++i) {
1145 iblk = swp_pager_meta_ctl(object, mreq->pindex + i, 0);
1146 if (iblk != blk + i)
1148 if ((blk ^ iblk) & dmmax_mask)
1150 m = vm_page_lookup(object, mreq->pindex + i);
1152 m = vm_page_alloc(object, mreq->pindex + i,
1157 if ((m->flags & PG_BUSY) || m->busy || m->valid)
1159 vm_page_unqueue_nowakeup(m);
1165 vm_page_flag_set(marray[i - 1], PG_RAM);
1170 * If mreq is the requested page and we have nothing to do return
1171 * VM_PAGER_FAIL. If raonly is set mreq is just another read-ahead
1172 * page and must be cleaned up.
1174 if (blk == SWAPBLK_NONE) {
1177 vnode_pager_freepage(mreq);
1178 return(VM_PAGER_OK);
1180 return(VM_PAGER_FAIL);
1185 * map our page(s) into kva for input
1187 bp = getpbuf(&nsw_rcount);
1189 kva = (vm_offset_t) bp->b_kvabase;
1190 bcopy(marray, bp->b_xio.xio_pages, i * sizeof(vm_page_t));
1191 pmap_qenter(kva, bp->b_xio.xio_pages, i);
1193 bp->b_data = (caddr_t)kva;
1194 bp->b_bcount = PAGE_SIZE * i;
1195 bp->b_xio.xio_npages = i;
1196 bio->bio_done = swp_pager_async_iodone;
1197 bio->bio_offset = (off_t)blk << PAGE_SHIFT;
1198 bio->bio_caller_info1.index = SWBIO_READ;
1201 * Set index. If raonly set the index beyond the array so all
1202 * the pages are treated the same, otherwise the original mreq is
1206 bio->bio_driver_info = (void *)(intptr_t)i;
1208 bio->bio_driver_info = (void *)(intptr_t)0;
1210 for (j = 0; j < i; ++j)
1211 vm_page_flag_set(bp->b_xio.xio_pages[j], PG_SWAPINPROG);
1213 mycpu->gd_cnt.v_swapin++;
1214 mycpu->gd_cnt.v_swappgsin += bp->b_xio.xio_npages;
1217 * We still hold the lock on mreq, and our automatic completion routine
1218 * does not remove it.
1220 vm_object_pip_add(object, bp->b_xio.xio_npages);
1223 * perform the I/O. NOTE!!! bp cannot be considered valid after
1224 * this point because we automatically release it on completion.
1225 * Instead, we look at the one page we are interested in which we
1226 * still hold a lock on even through the I/O completion.
1228 * The other pages in our m[] array are also released on completion,
1229 * so we cannot assume they are valid anymore either.
1231 bp->b_cmd = BUF_CMD_READ;
1233 vn_strategy(swapdev_vp, bio);
1236 * Wait for the page we want to complete. PG_SWAPINPROG is always
1237 * cleared on completion. If an I/O error occurs, SWAPBLK_NONE
1238 * is set in the meta-data.
1240 * If this is a read-ahead only we return immediately without
1244 return(VM_PAGER_OK);
1247 * Read-ahead includes originally requested page case.
1250 while ((mreq->flags & PG_SWAPINPROG) != 0) {
1251 vm_page_flag_set(mreq, PG_WANTED | PG_REFERENCED);
1252 mycpu->gd_cnt.v_intrans++;
1253 if (tsleep(mreq, 0, "swread", hz*20)) {
1255 "swap_pager: indefinite wait buffer: "
1256 " offset: %lld, size: %ld\n",
1257 (long long)bio->bio_offset,
1265 * mreq is left bussied after completion, but all the other pages
1266 * are freed. If we had an unrecoverable read error the page will
1269 if (mreq->valid != VM_PAGE_BITS_ALL)
1270 return(VM_PAGER_ERROR);
1272 return(VM_PAGER_OK);
1275 * A final note: in a low swap situation, we cannot deallocate swap
1276 * and mark a page dirty here because the caller is likely to mark
1277 * the page clean when we return, causing the page to possibly revert
1278 * to all-zero's later.
1283 * swap_pager_putpages:
1285 * Assign swap (if necessary) and initiate I/O on the specified pages.
1287 * We support both OBJT_DEFAULT and OBJT_SWAP objects. DEFAULT objects
1288 * are automatically converted to SWAP objects.
1290 * In a low memory situation we may block in vn_strategy(), but the new
1291 * vm_page reservation system coupled with properly written VFS devices
1292 * should ensure that no low-memory deadlock occurs. This is an area
1295 * The parent has N vm_object_pip_add() references prior to
1296 * calling us and will remove references for rtvals[] that are
1297 * not set to VM_PAGER_PEND. We need to remove the rest on I/O
1300 * The parent has soft-busy'd the pages it passes us and will unbusy
1301 * those whos rtvals[] entry is not set to VM_PAGER_PEND on return.
1302 * We need to unbusy the rest on I/O completion.
1305 swap_pager_putpages(vm_object_t object, vm_page_t *m, int count,
1306 boolean_t sync, int *rtvals)
1311 if (count && m[0]->object != object) {
1312 panic("swap_pager_getpages: object mismatch %p/%p",
1321 * Turn object into OBJT_SWAP
1322 * check for bogus sysops
1323 * force sync if not pageout process
1325 if (object->type == OBJT_DEFAULT)
1326 swp_pager_meta_convert(object);
1328 if (curthread != pagethread)
1334 * Update nsw parameters from swap_async_max sysctl values.
1335 * Do not let the sysop crash the machine with bogus numbers.
1338 if (swap_async_max != nsw_wcount_async_max) {
1344 if ((n = swap_async_max) > nswbuf / 2)
1351 * Adjust difference ( if possible ). If the current async
1352 * count is too low, we may not be able to make the adjustment
1356 n -= nsw_wcount_async_max;
1357 if (nsw_wcount_async + n >= 0) {
1358 nsw_wcount_async += n;
1359 nsw_wcount_async_max += n;
1360 wakeup(&nsw_wcount_async);
1368 * Assign swap blocks and issue I/O. We reallocate swap on the fly.
1369 * The page is left dirty until the pageout operation completes
1373 for (i = 0; i < count; i += n) {
1380 * Maximum I/O size is limited by a number of factors.
1383 n = min(BLIST_MAX_ALLOC, count - i);
1384 n = min(n, nsw_cluster_max);
1389 * Get biggest block of swap we can. If we fail, fall
1390 * back and try to allocate a smaller block. Don't go
1391 * overboard trying to allocate space if it would overly
1395 (blk = swp_pager_getswapspace(n)) == SWAPBLK_NONE &&
1400 if (blk == SWAPBLK_NONE) {
1401 for (j = 0; j < n; ++j)
1402 rtvals[i+j] = VM_PAGER_FAIL;
1408 * The I/O we are constructing cannot cross a physical
1409 * disk boundry in the swap stripe. Note: we are still
1412 if ((blk ^ (blk + n)) & dmmax_mask) {
1413 j = ((blk + dmmax) & dmmax_mask) - blk;
1414 swp_pager_freeswapspace(blk + j, n - j);
1419 * All I/O parameters have been satisfied, build the I/O
1420 * request and assign the swap space.
1424 bp = getpbuf(&nsw_wcount_sync);
1426 bp = getpbuf(&nsw_wcount_async);
1429 pmap_qenter((vm_offset_t)bp->b_data, &m[i], n);
1431 bp->b_bcount = PAGE_SIZE * n;
1432 bio->bio_offset = (off_t)blk << PAGE_SHIFT;
1434 for (j = 0; j < n; ++j) {
1435 vm_page_t mreq = m[i+j];
1437 swp_pager_meta_build(
1442 vm_page_dirty(mreq);
1443 rtvals[i+j] = VM_PAGER_OK;
1445 vm_page_flag_set(mreq, PG_SWAPINPROG);
1446 bp->b_xio.xio_pages[j] = mreq;
1448 bp->b_xio.xio_npages = n;
1450 mycpu->gd_cnt.v_swapout++;
1451 mycpu->gd_cnt.v_swappgsout += bp->b_xio.xio_npages;
1455 bp->b_dirtyoff = 0; /* req'd for NFS */
1456 bp->b_dirtyend = bp->b_bcount; /* req'd for NFS */
1457 bp->b_cmd = BUF_CMD_WRITE;
1458 bio->bio_caller_info1.index = SWBIO_WRITE;
1463 if (sync == FALSE) {
1464 bio->bio_done = swp_pager_async_iodone;
1466 vn_strategy(swapdev_vp, bio);
1468 for (j = 0; j < n; ++j)
1469 rtvals[i+j] = VM_PAGER_PEND;
1474 * Issue synchrnously.
1476 * Wait for the sync I/O to complete, then update rtvals.
1477 * We just set the rtvals[] to VM_PAGER_PEND so we can call
1478 * our async completion routine at the end, thus avoiding a
1481 bio->bio_caller_info1.index |= SWBIO_SYNC;
1482 bio->bio_done = biodone_sync;
1483 bio->bio_flags |= BIO_SYNC;
1484 vn_strategy(swapdev_vp, bio);
1485 biowait(bio, "swwrt");
1487 for (j = 0; j < n; ++j)
1488 rtvals[i+j] = VM_PAGER_PEND;
1491 * Now that we are through with the bp, we can call the
1492 * normal async completion, which frees everything up.
1494 swp_pager_async_iodone(bio);
1499 swap_pager_newswap(void)
1505 * swp_pager_async_iodone:
1507 * Completion routine for asynchronous reads and writes from/to swap.
1508 * Also called manually by synchronous code to finish up a bp.
1510 * For READ operations, the pages are PG_BUSY'd. For WRITE operations,
1511 * the pages are vm_page_t->busy'd. For READ operations, we PG_BUSY
1512 * unbusy all pages except the 'main' request page. For WRITE
1513 * operations, we vm_page_t->busy'd unbusy all pages ( we can do this
1514 * because we marked them all VM_PAGER_PEND on return from putpages ).
1516 * This routine may not block.
1519 swp_pager_async_iodone(struct bio *bio)
1521 struct buf *bp = bio->bio_buf;
1522 vm_object_t object = NULL;
1529 if (bp->b_flags & B_ERROR) {
1531 "swap_pager: I/O error - %s failed; offset %lld,"
1532 "size %ld, error %d\n",
1533 ((bio->bio_caller_info1.index & SWBIO_READ) ?
1534 "pagein" : "pageout"),
1535 (long long)bio->bio_offset,
1542 * set object, raise to splvm().
1544 if (bp->b_xio.xio_npages)
1545 object = bp->b_xio.xio_pages[0]->object;
1549 * remove the mapping for kernel virtual
1551 pmap_qremove((vm_offset_t)bp->b_data, bp->b_xio.xio_npages);
1554 * cleanup pages. If an error occurs writing to swap, we are in
1555 * very serious trouble. If it happens to be a disk error, though,
1556 * we may be able to recover by reassigning the swap later on. So
1557 * in this case we remove the m->swapblk assignment for the page
1558 * but do not free it in the rlist. The errornous block(s) are thus
1559 * never reallocated as swap. Redirty the page and continue.
1561 for (i = 0; i < bp->b_xio.xio_npages; ++i) {
1562 vm_page_t m = bp->b_xio.xio_pages[i];
1564 if (bp->b_flags & B_ERROR) {
1566 * If an error occurs I'd love to throw the swapblk
1567 * away without freeing it back to swapspace, so it
1568 * can never be used again. But I can't from an
1572 if (bio->bio_caller_info1.index & SWBIO_READ) {
1574 * When reading, reqpage needs to stay
1575 * locked for the parent, but all other
1576 * pages can be freed. We still want to
1577 * wakeup the parent waiting on the page,
1578 * though. ( also: pg_reqpage can be -1 and
1579 * not match anything ).
1581 * We have to wake specifically requested pages
1582 * up too because we cleared PG_SWAPINPROG and
1583 * someone may be waiting for that.
1585 * NOTE: for reads, m->dirty will probably
1586 * be overridden by the original caller of
1587 * getpages so don't play cute tricks here.
1589 * NOTE: We can't actually free the page from
1590 * here, because this is an interrupt. It
1591 * is not legal to mess with object->memq
1592 * from an interrupt. Deactivate the page
1597 vm_page_flag_clear(m, PG_ZERO);
1598 vm_page_flag_clear(m, PG_SWAPINPROG);
1601 * bio_driver_info holds the requested page
1604 if (i != (int)(intptr_t)bio->bio_driver_info) {
1605 vm_page_deactivate(m);
1611 * If i == bp->b_pager.pg_reqpage, do not wake
1612 * the page up. The caller needs to.
1616 * If a write error occurs, reactivate page
1617 * so it doesn't clog the inactive list,
1618 * then finish the I/O.
1621 vm_page_flag_clear(m, PG_SWAPINPROG);
1622 vm_page_activate(m);
1623 vm_page_io_finish(m);
1625 } else if (bio->bio_caller_info1.index & SWBIO_READ) {
1627 * NOTE: for reads, m->dirty will probably be
1628 * overridden by the original caller of getpages so
1629 * we cannot set them in order to free the underlying
1630 * swap in a low-swap situation. I don't think we'd
1631 * want to do that anyway, but it was an optimization
1632 * that existed in the old swapper for a time before
1633 * it got ripped out due to precisely this problem.
1635 * clear PG_ZERO in page.
1637 * If not the requested page then deactivate it.
1639 * Note that the requested page, reqpage, is left
1640 * busied, but we still have to wake it up. The
1641 * other pages are released (unbusied) by
1642 * vm_page_wakeup(). We do not set reqpage's
1643 * valid bits here, it is up to the caller.
1647 * NOTE: can't call pmap_clear_modify(m) from an
1648 * interrupt thread, the pmap code may have to map
1649 * non-kernel pmaps and currently asserts the case.
1651 /*pmap_clear_modify(m);*/
1652 m->valid = VM_PAGE_BITS_ALL;
1654 vm_page_flag_clear(m, PG_ZERO | PG_SWAPINPROG);
1657 * We have to wake specifically requested pages
1658 * up too because we cleared PG_SWAPINPROG and
1659 * could be waiting for it in getpages. However,
1660 * be sure to not unbusy getpages specifically
1661 * requested page - getpages expects it to be
1664 * bio_driver_info holds the requested page
1666 if (i != (int)(intptr_t)bio->bio_driver_info) {
1667 vm_page_deactivate(m);
1674 * Mark the page clean but do not mess with the
1675 * pmap-layer's modified state. That state should
1676 * also be clear since the caller protected the
1677 * page VM_PROT_READ, but allow the case.
1679 * We are in an interrupt, avoid pmap operations.
1681 * If we have a severe page deficit, deactivate the
1682 * page. Do not try to cache it (which would also
1683 * involve a pmap op), because the page might still
1687 vm_page_flag_clear(m, PG_SWAPINPROG);
1688 vm_page_io_finish(m);
1689 if (vm_page_count_severe())
1690 vm_page_deactivate(m);
1692 if (!vm_page_count_severe() || !vm_page_try_to_cache(m))
1693 vm_page_protect(m, VM_PROT_READ);
1699 * adjust pip. NOTE: the original parent may still have its own
1700 * pip refs on the object.
1704 vm_object_pip_wakeupn(object, bp->b_xio.xio_npages);
1707 * Release the physical I/O buffer.
1709 * NOTE: Due to synchronous operations in the write case b_cmd may
1710 * already be set to BUF_CMD_DONE and BIO_SYNC may have already
1713 if (bio->bio_caller_info1.index & SWBIO_READ)
1714 nswptr = &nsw_rcount;
1715 else if (bio->bio_caller_info1.index & SWBIO_SYNC)
1716 nswptr = &nsw_wcount_sync;
1718 nswptr = &nsw_wcount_async;
1719 bp->b_cmd = BUF_CMD_DONE;
1720 relpbuf(bp, nswptr);
1724 /************************************************************************
1726 ************************************************************************
1728 * These routines manipulate the swap metadata stored in the
1729 * OBJT_SWAP object. All swp_*() routines must be called at
1730 * splvm() because swap can be freed up by the low level vm_page
1731 * code which might be called from interrupts beyond what splbio() covers.
1733 * Swap metadata is implemented with a global hash and not directly
1734 * linked into the object. Instead the object simply contains
1735 * appropriate tracking counters.
1739 * Lookup the swblock containing the specified swap block index.
1743 swp_pager_lookup(vm_object_t object, vm_pindex_t index)
1745 index &= ~SWAP_META_MASK;
1746 return (RB_LOOKUP(swblock_rb_tree, &object->swblock_root, index));
1750 * Remove a swblock from the RB tree.
1754 swp_pager_remove(vm_object_t object, struct swblock *swap)
1756 RB_REMOVE(swblock_rb_tree, &object->swblock_root, swap);
1760 * Convert default object to swap object if necessary
1763 swp_pager_meta_convert(vm_object_t object)
1765 if (object->type == OBJT_DEFAULT) {
1766 object->type = OBJT_SWAP;
1767 KKASSERT(object->swblock_count == 0);
1772 * SWP_PAGER_META_BUILD() - add swap block to swap meta data for object
1774 * We first convert the object to a swap object if it is a default
1775 * object. Vnode objects do not need to be converted.
1777 * The specified swapblk is added to the object's swap metadata. If
1778 * the swapblk is not valid, it is freed instead. Any previously
1779 * assigned swapblk is freed.
1782 swp_pager_meta_build(vm_object_t object, vm_pindex_t index, daddr_t swapblk)
1784 struct swblock *swap;
1785 struct swblock *oswap;
1787 KKASSERT(swapblk != SWAPBLK_NONE);
1790 * Convert object if necessary
1792 if (object->type == OBJT_DEFAULT)
1793 swp_pager_meta_convert(object);
1796 * Locate swblock. If not found create, but if we aren't adding
1797 * anything just return. If we run out of space in the map we wait
1798 * and, since the hash table may have changed, retry.
1801 swap = swp_pager_lookup(object, index);
1806 swap = zalloc(swap_zone);
1811 swap->swb_index = index & ~SWAP_META_MASK;
1812 swap->swb_count = 0;
1814 ++object->swblock_count;
1816 for (i = 0; i < SWAP_META_PAGES; ++i)
1817 swap->swb_pages[i] = SWAPBLK_NONE;
1818 oswap = RB_INSERT(swblock_rb_tree, &object->swblock_root, swap);
1819 KKASSERT(oswap == NULL);
1823 * Delete prior contents of metadata
1826 index &= SWAP_META_MASK;
1828 if (swap->swb_pages[index] != SWAPBLK_NONE) {
1829 swp_pager_freeswapspace(swap->swb_pages[index], 1);
1834 * Enter block into metadata
1836 swap->swb_pages[index] = swapblk;
1837 if (swapblk != SWAPBLK_NONE)
1842 * SWP_PAGER_META_FREE() - free a range of blocks in the object's swap metadata
1844 * The requested range of blocks is freed, with any associated swap
1845 * returned to the swap bitmap.
1847 * This routine will free swap metadata structures as they are cleaned
1848 * out. This routine does *NOT* operate on swap metadata associated
1849 * with resident pages.
1851 * This routine must be called at splvm()
1854 swp_pager_meta_free(vm_object_t object, vm_pindex_t index, daddr_t count)
1856 struct swblock *swap;
1858 if (object->type != OBJT_SWAP && object->type != OBJT_VNODE)
1862 swap = swp_pager_lookup(object, index);
1864 daddr_t v = swap->swb_pages[index & SWAP_META_MASK];
1866 if (v != SWAPBLK_NONE) {
1867 swp_pager_freeswapspace(v, 1);
1868 swap->swb_pages[index & SWAP_META_MASK] =
1870 if (--swap->swb_count == 0) {
1871 swp_pager_remove(object, swap);
1872 zfree(swap_zone, swap);
1873 --object->swblock_count;
1879 int n = SWAP_META_PAGES - (index & SWAP_META_MASK);
1887 * SWP_PAGER_META_FREE_ALL() - destroy all swap metadata associated with object
1889 * This routine locates and destroys all swap metadata associated with
1892 * This routine must be called at splvm()
1895 swp_pager_meta_free_all(vm_object_t object)
1897 struct swblock *swap;
1900 if (object->type != OBJT_SWAP && object->type != OBJT_VNODE)
1903 while ((swap = RB_ROOT(&object->swblock_root)) != NULL) {
1904 swp_pager_remove(object, swap);
1905 for (i = 0; i < SWAP_META_PAGES; ++i) {
1906 daddr_t v = swap->swb_pages[i];
1907 if (v != SWAPBLK_NONE) {
1909 swp_pager_freeswapspace(v, 1);
1912 if (swap->swb_count != 0)
1913 panic("swap_pager_meta_free_all: swb_count != 0");
1914 zfree(swap_zone, swap);
1915 --object->swblock_count;
1917 KKASSERT(object->swblock_count == 0);
1921 * SWP_PAGER_METACTL() - misc control of swap and vm_page_t meta data.
1923 * This routine is capable of looking up, popping, or freeing
1924 * swapblk assignments in the swap meta data or in the vm_page_t.
1925 * The routine typically returns the swapblk being looked-up, or popped,
1926 * or SWAPBLK_NONE if the block was freed, or SWAPBLK_NONE if the block
1927 * was invalid. This routine will automatically free any invalid
1928 * meta-data swapblks.
1930 * It is not possible to store invalid swapblks in the swap meta data
1931 * (other then a literal 'SWAPBLK_NONE'), so we don't bother checking.
1933 * When acting on a busy resident page and paging is in progress, we
1934 * have to wait until paging is complete but otherwise can act on the
1937 * This routine must be called at splvm().
1939 * SWM_FREE remove and free swap block from metadata
1940 * SWM_POP remove from meta data but do not free.. pop it out
1943 swp_pager_meta_ctl(vm_object_t object, vm_pindex_t index, int flags)
1945 struct swblock *swap;
1949 * The meta data only exists of the object is OBJT_SWAP
1950 * and even then might not be allocated yet.
1953 if (object->type != OBJT_SWAP && object->type != OBJT_VNODE)
1954 return(SWAPBLK_NONE);
1957 swap = swp_pager_lookup(object, index);
1960 index &= SWAP_META_MASK;
1961 r1 = swap->swb_pages[index];
1963 if (r1 != SWAPBLK_NONE) {
1964 if (flags & SWM_FREE) {
1965 swp_pager_freeswapspace(r1, 1);
1968 if (flags & (SWM_FREE|SWM_POP)) {
1969 swap->swb_pages[index] = SWAPBLK_NONE;
1970 if (--swap->swb_count == 0) {
1971 swp_pager_remove(object, swap);
1972 zfree(swap_zone, swap);
1973 --object->swblock_count;