2 * Copyright (c) 1998,2004 The DragonFly Project. All rights reserved.
4 * This code is derived from software contributed to The DragonFly Project
5 * by Matthew Dillon <dillon@backplane.com>
7 * Redistribution and use in source and binary forms, with or without
8 * modification, are permitted provided that the following conditions
11 * 1. Redistributions of source code must retain the above copyright
12 * notice, this list of conditions and the following disclaimer.
13 * 2. Redistributions in binary form must reproduce the above copyright
14 * notice, this list of conditions and the following disclaimer in
15 * the documentation and/or other materials provided with the
17 * 3. Neither the name of The DragonFly Project nor the names of its
18 * contributors may be used to endorse or promote products derived
19 * from this software without specific, prior written permission.
21 * THIS SOFTWARE IS PROVIDED BY THE COPYRIGHT HOLDERS AND CONTRIBUTORS
22 * ``AS IS'' AND ANY EXPRESS OR IMPLIED WARRANTIES, INCLUDING, BUT NOT
23 * LIMITED TO, THE IMPLIED WARRANTIES OF MERCHANTABILITY AND FITNESS
24 * FOR A PARTICULAR PURPOSE ARE DISCLAIMED. IN NO EVENT SHALL THE
25 * COPYRIGHT HOLDERS OR CONTRIBUTORS BE LIABLE FOR ANY DIRECT, INDIRECT,
26 * INCIDENTAL, SPECIAL, EXEMPLARY OR CONSEQUENTIAL DAMAGES (INCLUDING,
27 * BUT NOT LIMITED TO, PROCUREMENT OF SUBSTITUTE GOODS OR SERVICES;
28 * LOSS OF USE, DATA, OR PROFITS; OR BUSINESS INTERRUPTION) HOWEVER CAUSED
29 * AND ON ANY THEORY OF LIABILITY, WHETHER IN CONTRACT, STRICT LIABILITY,
30 * OR TORT (INCLUDING NEGLIGENCE OR OTHERWISE) ARISING IN ANY WAY OUT
31 * OF THE USE OF THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF
34 * Copyright (c) 1994 John S. Dyson
35 * Copyright (c) 1990 University of Utah.
36 * Copyright (c) 1991, 1993
37 * The Regents of the University of California. All rights reserved.
39 * This code is derived from software contributed to Berkeley by
40 * the Systems Programming Group of the University of Utah Computer
43 * Redistribution and use in source and binary forms, with or without
44 * modification, are permitted provided that the following conditions
46 * 1. Redistributions of source code must retain the above copyright
47 * notice, this list of conditions and the following disclaimer.
48 * 2. Redistributions in binary form must reproduce the above copyright
49 * notice, this list of conditions and the following disclaimer in the
50 * documentation and/or other materials provided with the distribution.
51 * 3. All advertising materials mentioning features or use of this software
52 * must display the following acknowledgement:
53 * This product includes software developed by the University of
54 * California, Berkeley and its contributors.
55 * 4. Neither the name of the University nor the names of its contributors
56 * may be used to endorse or promote products derived from this software
57 * without specific prior written permission.
59 * THIS SOFTWARE IS PROVIDED BY THE REGENTS AND CONTRIBUTORS ``AS IS'' AND
60 * ANY EXPRESS OR IMPLIED WARRANTIES, INCLUDING, BUT NOT LIMITED TO, THE
61 * IMPLIED WARRANTIES OF MERCHANTABILITY AND FITNESS FOR A PARTICULAR PURPOSE
62 * ARE DISCLAIMED. IN NO EVENT SHALL THE REGENTS OR CONTRIBUTORS BE LIABLE
63 * FOR ANY DIRECT, INDIRECT, INCIDENTAL, SPECIAL, EXEMPLARY, OR CONSEQUENTIAL
64 * DAMAGES (INCLUDING, BUT NOT LIMITED TO, PROCUREMENT OF SUBSTITUTE GOODS
65 * OR SERVICES; LOSS OF USE, DATA, OR PROFITS; OR BUSINESS INTERRUPTION)
66 * HOWEVER CAUSED AND ON ANY THEORY OF LIABILITY, WHETHER IN CONTRACT, STRICT
67 * LIABILITY, OR TORT (INCLUDING NEGLIGENCE OR OTHERWISE) ARISING IN ANY WAY
68 * OUT OF THE USE OF THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF
74 * Radix Bitmap 'blists'.
76 * - The new swapper uses the new radix bitmap code. This should scale
77 * to arbitrarily small or arbitrarily large swap spaces and an almost
78 * arbitrary degree of fragmentation.
82 * - on the fly reallocation of swap during putpages. The new system
83 * does not try to keep previously allocated swap blocks for dirty
86 * - on the fly deallocation of swap
88 * - No more garbage collection required. Unnecessarily allocated swap
89 * blocks only exist for dirty vm_page_t's now and these are already
90 * cycled (in a high-load system) by the pager. We also do on-the-fly
91 * removal of invalidated swap blocks when a page is destroyed
94 * from: Utah $Hdr: swap_pager.c 1.4 91/04/30$
96 * @(#)swap_pager.c 8.9 (Berkeley) 3/21/94
98 * $FreeBSD: src/sys/vm/swap_pager.c,v 1.130.2.12 2002/08/31 21:15:55 dillon Exp $
99 * $DragonFly: src/sys/vm/swap_pager.c,v 1.32 2008/07/01 02:02:56 dillon Exp $
102 #include <sys/param.h>
103 #include <sys/systm.h>
104 #include <sys/conf.h>
105 #include <sys/kernel.h>
106 #include <sys/proc.h>
108 #include <sys/vnode.h>
109 #include <sys/malloc.h>
110 #include <sys/vmmeter.h>
111 #include <sys/sysctl.h>
112 #include <sys/blist.h>
113 #include <sys/lock.h>
114 #include <sys/thread2.h>
116 #ifndef MAX_PAGEOUT_CLUSTER
117 #define MAX_PAGEOUT_CLUSTER 16
120 #define SWB_NPAGES MAX_PAGEOUT_CLUSTER
122 #include "opt_swap.h"
124 #include <vm/vm_object.h>
125 #include <vm/vm_page.h>
126 #include <vm/vm_pager.h>
127 #include <vm/vm_pageout.h>
128 #include <vm/swap_pager.h>
129 #include <vm/vm_extern.h>
130 #include <vm/vm_zone.h>
131 #include <vm/vnode_pager.h>
133 #include <sys/buf2.h>
134 #include <vm/vm_page2.h>
136 #define SWM_FREE 0x02 /* free, period */
137 #define SWM_POP 0x04 /* pop out */
139 #define SWBIO_READ 0x01
140 #define SWBIO_WRITE 0x02
141 #define SWBIO_SYNC 0x04
147 vm_pindex_t endi; /* inclusive */
151 * vm_swap_size is in page-sized chunks now. It was DEV_BSIZE'd chunks
155 extern int vm_swap_size; /* number of free swap blocks, in pages */
157 int swap_pager_full; /* swap space exhaustion (task killing) */
158 int vm_swap_cache_use;
159 int vm_swap_anon_use;
161 static int swap_pager_almost_full; /* swap space exhaustion (w/ hysteresis)*/
162 static int nsw_rcount; /* free read buffers */
163 static int nsw_wcount_sync; /* limit write buffers / synchronous */
164 static int nsw_wcount_async; /* limit write buffers / asynchronous */
165 static int nsw_wcount_async_max;/* assigned maximum */
166 static int nsw_cluster_max; /* maximum VOP I/O allowed */
168 struct blist *swapblist;
169 static int swap_async_max = 4; /* maximum in-progress async I/O's */
170 static int swap_burst_read = 0; /* allow burst reading */
172 extern struct vnode *swapdev_vp; /* from vm_swap.c */
174 SYSCTL_INT(_vm, OID_AUTO, swap_async_max,
175 CTLFLAG_RW, &swap_async_max, 0, "Maximum running async swap ops");
176 SYSCTL_INT(_vm, OID_AUTO, swap_burst_read,
177 CTLFLAG_RW, &swap_burst_read, 0, "Allow burst reads for pageins");
179 SYSCTL_INT(_vm, OID_AUTO, swap_cache_use,
180 CTLFLAG_RD, &vm_swap_cache_use, 0, "");
181 SYSCTL_INT(_vm, OID_AUTO, swap_anon_use,
182 CTLFLAG_RD, &vm_swap_anon_use, 0, "");
187 * Red-Black tree for swblock entries
189 RB_GENERATE2(swblock_rb_tree, swblock, swb_entry, rb_swblock_compare,
190 vm_pindex_t, swb_index);
193 rb_swblock_compare(struct swblock *swb1, struct swblock *swb2)
195 if (swb1->swb_index < swb2->swb_index)
197 if (swb1->swb_index > swb2->swb_index)
204 rb_swblock_scancmp(struct swblock *swb, void *data)
206 struct swfreeinfo *info = data;
208 if (swb->swb_index < info->basei)
210 if (swb->swb_index > info->endi)
216 * pagerops for OBJT_SWAP - "swap pager". Some ops are also global procedure
217 * calls hooked from other parts of the VM system and do not appear here.
218 * (see vm/swap_pager.h).
222 swap_pager_alloc (void *handle, off_t size,
223 vm_prot_t prot, off_t offset);
224 static void swap_pager_dealloc (vm_object_t object);
225 static int swap_pager_getpage (vm_object_t, vm_page_t *, int);
226 static void swap_chain_iodone(struct bio *biox);
228 struct pagerops swappagerops = {
229 swap_pager_alloc, /* allocate an OBJT_SWAP object */
230 swap_pager_dealloc, /* deallocate an OBJT_SWAP object */
231 swap_pager_getpage, /* pagein */
232 swap_pager_putpages, /* pageout */
233 swap_pager_haspage /* get backing store status for page */
237 * dmmax is in page-sized chunks with the new swap system. It was
238 * dev-bsized chunks in the old. dmmax is always a power of 2.
240 * swap_*() routines are externally accessible. swp_*() routines are
245 static int dmmax_mask;
246 int nswap_lowat = 128; /* in pages, swap_pager_almost_full warn */
247 int nswap_hiwat = 512; /* in pages, swap_pager_almost_full warn */
249 static __inline void swp_sizecheck (void);
250 static void swp_pager_async_iodone (struct bio *bio);
253 * Swap bitmap functions
256 static __inline void swp_pager_freeswapspace (vm_object_t object, daddr_t blk, int npages);
257 static __inline daddr_t swp_pager_getswapspace (vm_object_t object, int npages);
263 static void swp_pager_meta_convert (vm_object_t);
264 static void swp_pager_meta_build (vm_object_t, vm_pindex_t, daddr_t);
265 static void swp_pager_meta_free (vm_object_t, vm_pindex_t, vm_pindex_t);
266 static void swp_pager_meta_free_all (vm_object_t);
267 static daddr_t swp_pager_meta_ctl (vm_object_t, vm_pindex_t, int);
270 * SWP_SIZECHECK() - update swap_pager_full indication
272 * update the swap_pager_almost_full indication and warn when we are
273 * about to run out of swap space, using lowat/hiwat hysteresis.
275 * Clear swap_pager_full ( task killing ) indication when lowat is met.
277 * No restrictions on call
278 * This routine may not block.
279 * This routine must be called at splvm()
285 if (vm_swap_size < nswap_lowat) {
286 if (swap_pager_almost_full == 0) {
287 kprintf("swap_pager: out of swap space\n");
288 swap_pager_almost_full = 1;
292 if (vm_swap_size > nswap_hiwat)
293 swap_pager_almost_full = 0;
298 * SWAP_PAGER_INIT() - initialize the swap pager!
300 * Expected to be started from system init. NOTE: This code is run
301 * before much else so be careful what you depend on. Most of the VM
302 * system has yet to be initialized at this point.
305 swap_pager_init(void *arg __unused)
308 * Device Stripe, in PAGE_SIZE'd blocks
310 dmmax = SWB_NPAGES * 2;
311 dmmax_mask = ~(dmmax - 1);
313 SYSINIT(vm_mem, SI_BOOT1_VM, SI_ORDER_THIRD, swap_pager_init, NULL)
316 * SWAP_PAGER_SWAP_INIT() - swap pager initialization from pageout process
318 * Expected to be started from pageout process once, prior to entering
323 swap_pager_swap_init(void)
328 * Number of in-transit swap bp operations. Don't
329 * exhaust the pbufs completely. Make sure we
330 * initialize workable values (0 will work for hysteresis
331 * but it isn't very efficient).
333 * The nsw_cluster_max is constrained by the number of pages an XIO
334 * holds, i.e., (MAXPHYS/PAGE_SIZE) and our locally defined
335 * MAX_PAGEOUT_CLUSTER. Also be aware that swap ops are
336 * constrained by the swap device interleave stripe size.
338 * Currently we hardwire nsw_wcount_async to 4. This limit is
339 * designed to prevent other I/O from having high latencies due to
340 * our pageout I/O. The value 4 works well for one or two active swap
341 * devices but is probably a little low if you have more. Even so,
342 * a higher value would probably generate only a limited improvement
343 * with three or four active swap devices since the system does not
344 * typically have to pageout at extreme bandwidths. We will want
345 * at least 2 per swap devices, and 4 is a pretty good value if you
346 * have one NFS swap device due to the command/ack latency over NFS.
347 * So it all works out pretty well.
350 nsw_cluster_max = min((MAXPHYS/PAGE_SIZE), MAX_PAGEOUT_CLUSTER);
352 nsw_rcount = (nswbuf + 1) / 2;
353 nsw_wcount_sync = (nswbuf + 3) / 4;
354 nsw_wcount_async = 4;
355 nsw_wcount_async_max = nsw_wcount_async;
358 * The zone is dynamically allocated so generally size it to
359 * maxswzone (32MB to 512MB of KVM). Set a minimum size based
360 * on physical memory of around 8x (each swblock can hold 16 pages).
362 * With the advent of SSDs (vs HDs) the practical (swap:memory) ratio
363 * has increased dramatically.
365 n = vmstats.v_page_count / 2;
366 if (maxswzone && n < maxswzone / sizeof(struct swblock))
367 n = maxswzone / sizeof(struct swblock);
373 sizeof(struct swblock),
377 if (swap_zone != NULL)
380 * if the allocation failed, try a zone two thirds the
381 * size of the previous attempt.
386 if (swap_zone == NULL)
387 panic("swap_pager_swap_init: swap_zone == NULL");
389 kprintf("Swap zone entries reduced from %d to %d.\n", n2, n);
393 * SWAP_PAGER_ALLOC() - allocate a new OBJT_SWAP VM object and instantiate
394 * its metadata structures.
396 * This routine is called from the mmap and fork code to create a new
397 * OBJT_SWAP object. We do this by creating an OBJT_DEFAULT object
398 * and then converting it with swp_pager_meta_convert().
400 * This routine may block in vm_object_allocate() and create a named
401 * object lookup race, so we must interlock. We must also run at
402 * splvm() for the object lookup to handle races with interrupts, but
403 * we do not have to maintain splvm() in between the lookup and the
404 * add because (I believe) it is not possible to attempt to create
405 * a new swap object w/handle when a default object with that handle
410 swap_pager_alloc(void *handle, off_t size, vm_prot_t prot, off_t offset)
414 KKASSERT(handle == NULL);
418 * Reference existing named region or allocate new one. There
419 * should not be a race here against swp_pager_meta_build()
420 * as called from vm_page_remove() in regards to the lookup
423 while (sw_alloc_interlock) {
424 sw_alloc_interlock = -1;
425 tsleep(&sw_alloc_interlock, 0, "swpalc", 0);
427 sw_alloc_interlock = 1;
429 object = vm_pager_object_lookup(NOBJLIST(handle), handle);
431 if (object != NULL) {
432 vm_object_reference(object);
434 object = vm_object_allocate(OBJT_DEFAULT,
435 OFF_TO_IDX(offset + PAGE_MASK + size));
436 object->handle = handle;
437 swp_pager_meta_convert(object);
440 if (sw_alloc_interlock < 0)
441 wakeup(&sw_alloc_interlock);
442 sw_alloc_interlock = 0;
445 object = vm_object_allocate(OBJT_DEFAULT,
446 OFF_TO_IDX(offset + PAGE_MASK + size));
447 swp_pager_meta_convert(object);
453 * SWAP_PAGER_DEALLOC() - remove swap metadata from object
455 * The swap backing for the object is destroyed. The code is
456 * designed such that we can reinstantiate it later, but this
457 * routine is typically called only when the entire object is
458 * about to be destroyed.
460 * This routine may block, but no longer does.
462 * The object must be locked or unreferenceable.
466 swap_pager_dealloc(vm_object_t object)
468 vm_object_pip_wait(object, "swpdea");
471 * Free all remaining metadata. We only bother to free it from
472 * the swap meta data. We do not attempt to free swapblk's still
473 * associated with vm_page_t's for this object. We do not care
474 * if paging is still in progress on some objects.
477 swp_pager_meta_free_all(object);
481 /************************************************************************
482 * SWAP PAGER BITMAP ROUTINES *
483 ************************************************************************/
486 * SWP_PAGER_GETSWAPSPACE() - allocate raw swap space
488 * Allocate swap for the requested number of pages. The starting
489 * swap block number (a page index) is returned or SWAPBLK_NONE
490 * if the allocation failed.
492 * Also has the side effect of advising that somebody made a mistake
493 * when they configured swap and didn't configure enough.
495 * Must be called at splvm() to avoid races with bitmap frees from
496 * vm_page_remove() aka swap_pager_page_removed().
498 * This routine may not block
499 * This routine must be called at splvm().
501 static __inline daddr_t
502 swp_pager_getswapspace(vm_object_t object, int npages)
506 if ((blk = blist_alloc(swapblist, npages)) == SWAPBLK_NONE) {
507 if (swap_pager_full != 2) {
508 kprintf("swap_pager_getswapspace: failed\n");
510 swap_pager_almost_full = 1;
513 vm_swap_size -= npages;
514 if (object->type == OBJT_SWAP)
515 vm_swap_anon_use += npages;
517 vm_swap_cache_use += npages;
524 * SWP_PAGER_FREESWAPSPACE() - free raw swap space
526 * This routine returns the specified swap blocks back to the bitmap.
528 * Note: This routine may not block (it could in the old swap code),
529 * and through the use of the new blist routines it does not block.
531 * We must be called at splvm() to avoid races with bitmap frees from
532 * vm_page_remove() aka swap_pager_page_removed().
534 * This routine may not block
535 * This routine must be called at splvm().
539 swp_pager_freeswapspace(vm_object_t object, daddr_t blk, int npages)
541 blist_free(swapblist, blk, npages);
542 vm_swap_size += npages;
543 if (object->type == OBJT_SWAP)
544 vm_swap_anon_use -= npages;
546 vm_swap_cache_use -= npages;
551 * SWAP_PAGER_FREESPACE() - frees swap blocks associated with a page
552 * range within an object.
554 * This is a globally accessible routine.
556 * This routine removes swapblk assignments from swap metadata.
558 * The external callers of this routine typically have already destroyed
559 * or renamed vm_page_t's associated with this range in the object so
562 * This routine may be called at any spl. We up our spl to splvm
563 * temporarily in order to perform the metadata removal.
566 swap_pager_freespace(vm_object_t object, vm_pindex_t start, vm_pindex_t size)
569 swp_pager_meta_free(object, start, size);
574 swap_pager_freespace_all(vm_object_t object)
577 swp_pager_meta_free_all(object);
582 * Called by vm_page_alloc() when a new VM page is inserted
583 * into a VM object. Checks whether swap has been assigned to
584 * the page and sets PG_SWAPPED as necessary.
587 swap_pager_page_inserted(vm_page_t m)
589 if (m->object->swblock_count) {
591 if (swp_pager_meta_ctl(m->object, m->pindex, 0) != SWAPBLK_NONE)
592 vm_page_flag_set(m, PG_SWAPPED);
598 * SWAP_PAGER_RESERVE() - reserve swap blocks in object
600 * Assigns swap blocks to the specified range within the object. The
601 * swap blocks are not zerod. Any previous swap assignment is destroyed.
603 * Returns 0 on success, -1 on failure.
606 swap_pager_reserve(vm_object_t object, vm_pindex_t start, vm_size_t size)
609 daddr_t blk = SWAPBLK_NONE;
610 vm_pindex_t beg = start; /* save start index */
616 while ((blk = swp_pager_getswapspace(object, n)) ==
621 swp_pager_meta_free(object, beg,
628 swp_pager_meta_build(object, start, blk);
634 swp_pager_meta_free(object, start, n);
640 * SWAP_PAGER_COPY() - copy blocks from source pager to destination pager
641 * and destroy the source.
643 * Copy any valid swapblks from the source to the destination. In
644 * cases where both the source and destination have a valid swapblk,
645 * we keep the destination's.
647 * This routine is allowed to block. It may block allocating metadata
648 * indirectly through swp_pager_meta_build() or if paging is still in
649 * progress on the source.
651 * This routine can be called at any spl
653 * XXX vm_page_collapse() kinda expects us not to block because we
654 * supposedly do not need to allocate memory, but for the moment we
655 * *may* have to get a little memory from the zone allocator, but
656 * it is taken from the interrupt memory. We should be ok.
658 * The source object contains no vm_page_t's (which is just as well)
660 * The source object is of type OBJT_SWAP.
662 * The source and destination objects must be locked or
663 * inaccessible (XXX are they ?)
667 swap_pager_copy(vm_object_t srcobject, vm_object_t dstobject,
668 vm_pindex_t base_index, int destroysource)
675 * transfer source to destination.
677 for (i = 0; i < dstobject->size; ++i) {
681 * Locate (without changing) the swapblk on the destination,
682 * unless it is invalid in which case free it silently, or
683 * if the destination is a resident page, in which case the
684 * source is thrown away.
686 dstaddr = swp_pager_meta_ctl(dstobject, i, 0);
688 if (dstaddr == SWAPBLK_NONE) {
690 * Destination has no swapblk and is not resident,
695 srcaddr = swp_pager_meta_ctl(srcobject,
696 base_index + i, SWM_POP);
698 if (srcaddr != SWAPBLK_NONE)
699 swp_pager_meta_build(dstobject, i, srcaddr);
702 * Destination has valid swapblk or it is represented
703 * by a resident page. We destroy the sourceblock.
705 swp_pager_meta_ctl(srcobject, base_index + i, SWM_FREE);
710 * Free left over swap blocks in source.
712 * We have to revert the type to OBJT_DEFAULT so we do not accidently
713 * double-remove the object from the swap queues.
717 * Reverting the type is not necessary, the caller is going
718 * to destroy srcobject directly, but I'm doing it here
719 * for consistency since we've removed the object from its
722 swp_pager_meta_free_all(srcobject);
723 if (srcobject->type == OBJT_SWAP)
724 srcobject->type = OBJT_DEFAULT;
730 * SWAP_PAGER_HASPAGE() - determine if we have good backing store for
731 * the requested page.
733 * We determine whether good backing store exists for the requested
734 * page and return TRUE if it does, FALSE if it doesn't.
736 * If TRUE, we also try to determine how much valid, contiguous backing
737 * store exists before and after the requested page within a reasonable
738 * distance. We do not try to restrict it to the swap device stripe
739 * (that is handled in getpages/putpages). It probably isn't worth
744 swap_pager_haspage(vm_object_t object, vm_pindex_t pindex)
749 * do we have good backing store at the requested index ?
753 blk0 = swp_pager_meta_ctl(object, pindex, 0);
755 if (blk0 == SWAPBLK_NONE) {
762 * find backwards-looking contiguous good backing store
764 if (before != NULL) {
767 for (i = 1; i < (SWB_NPAGES/2); ++i) {
772 blk = swp_pager_meta_ctl(object, pindex - i, 0);
780 * find forward-looking contiguous good backing store
786 for (i = 1; i < (SWB_NPAGES/2); ++i) {
789 blk = swp_pager_meta_ctl(object, pindex + i, 0);
801 * SWAP_PAGER_PAGE_UNSWAPPED() - remove swap backing store related to page
803 * This removes any associated swap backing store, whether valid or
804 * not, from the page. This operates on any VM object, not just OBJT_SWAP
807 * This routine is typically called when a page is made dirty, at
808 * which point any associated swap can be freed. MADV_FREE also
809 * calls us in a special-case situation
811 * NOTE!!! If the page is clean and the swap was valid, the caller
812 * should make the page dirty before calling this routine. This routine
813 * does NOT change the m->dirty status of the page. Also: MADV_FREE
816 * This routine may not block
817 * This routine must be called at splvm()
820 swap_pager_unswapped(vm_page_t m)
822 if (m->flags & PG_SWAPPED) {
823 swp_pager_meta_ctl(m->object, m->pindex, SWM_FREE);
824 vm_page_flag_clear(m, PG_SWAPPED);
829 * SWAP_PAGER_STRATEGY() - read, write, free blocks
831 * This implements a VM OBJECT strategy function using swap backing store.
832 * This can operate on any VM OBJECT type, not necessarily just OBJT_SWAP
835 * This is intended to be a cacheless interface (i.e. caching occurs at
836 * higher levels), and is also used as a swap-based SSD cache for vnode
837 * and device objects.
839 * All I/O goes directly to and from the swap device.
841 * We currently attempt to run I/O synchronously or asynchronously as
842 * the caller requests. This isn't perfect because we loose error
843 * sequencing when we run multiple ops in parallel to satisfy a request.
844 * But this is swap, so we let it all hang out.
847 swap_pager_strategy(vm_object_t object, struct bio *bio)
849 struct buf *bp = bio->bio_buf;
852 vm_pindex_t biox_blkno = 0;
857 struct bio_track *track;
860 * tracking for swapdev vnode I/Os
862 if (bp->b_cmd == BUF_CMD_READ)
863 track = &swapdev_vp->v_track_read;
865 track = &swapdev_vp->v_track_write;
867 if (bp->b_bcount & PAGE_MASK) {
868 bp->b_error = EINVAL;
869 bp->b_flags |= B_ERROR | B_INVAL;
871 kprintf("swap_pager_strategy: bp %p offset %lld size %d, "
872 "not page bounded\n",
873 bp, (long long)bio->bio_offset, (int)bp->b_bcount);
878 * Clear error indication, initialize page index, count, data pointer.
881 bp->b_flags &= ~B_ERROR;
882 bp->b_resid = bp->b_bcount;
884 start = (vm_pindex_t)(bio->bio_offset >> PAGE_SHIFT);
885 count = howmany(bp->b_bcount, PAGE_SIZE);
889 * Deal with BUF_CMD_FREEBLKS
891 if (bp->b_cmd == BUF_CMD_FREEBLKS) {
893 * FREE PAGE(s) - destroy underlying swap that is no longer
896 swp_pager_meta_free(object, start, count);
903 * We need to be able to create a new cluster of I/O's. We cannot
904 * use the caller fields of the passed bio so push a new one.
906 * Because nbio is just a placeholder for the cluster links,
907 * we can biodone() the original bio instead of nbio to make
908 * things a bit more efficient.
910 nbio = push_bio(bio);
911 nbio->bio_offset = bio->bio_offset;
912 nbio->bio_caller_info1.cluster_head = NULL;
913 nbio->bio_caller_info2.cluster_tail = NULL;
919 * Execute read or write
925 * Obtain block. If block not found and writing, allocate a
926 * new block and build it into the object.
928 blk = swp_pager_meta_ctl(object, start, 0);
929 if ((blk == SWAPBLK_NONE) && bp->b_cmd != BUF_CMD_READ) {
930 blk = swp_pager_getswapspace(object, 1);
931 if (blk == SWAPBLK_NONE) {
932 bp->b_error = ENOMEM;
933 bp->b_flags |= B_ERROR;
936 swp_pager_meta_build(object, start, blk);
940 * Do we have to flush our current collection? Yes if:
942 * - no swap block at this index
943 * - swap block is not contiguous
944 * - we cross a physical disk boundry in the
948 biox && (biox_blkno + btoc(bufx->b_bcount) != blk ||
949 ((biox_blkno ^ blk) & dmmax_mask)
952 if (bp->b_cmd == BUF_CMD_READ) {
953 ++mycpu->gd_cnt.v_swapin;
954 mycpu->gd_cnt.v_swappgsin += btoc(bufx->b_bcount);
956 ++mycpu->gd_cnt.v_swapout;
957 mycpu->gd_cnt.v_swappgsout += btoc(bufx->b_bcount);
958 bufx->b_dirtyend = bufx->b_bcount;
962 * Finished with this buf.
964 KKASSERT(bufx->b_bcount != 0);
965 if (bufx->b_cmd != BUF_CMD_READ)
966 bufx->b_dirtyend = bufx->b_bcount;
972 * Add new swapblk to biox, instantiating biox if necessary.
973 * Zero-fill reads are able to take a shortcut.
975 if (blk == SWAPBLK_NONE) {
977 * We can only get here if we are reading. Since
978 * we are at splvm() we can safely modify b_resid,
979 * even if chain ops are in progress.
981 bzero(data, PAGE_SIZE);
982 bp->b_resid -= PAGE_SIZE;
985 /* XXX chain count > 4, wait to <= 4 */
987 bufx = getpbuf(NULL);
988 biox = &bufx->b_bio1;
989 cluster_append(nbio, bufx);
990 bufx->b_flags |= (bufx->b_flags & B_ORDERED);
991 bufx->b_cmd = bp->b_cmd;
992 biox->bio_done = swap_chain_iodone;
993 biox->bio_offset = (off_t)blk << PAGE_SHIFT;
994 biox->bio_caller_info1.cluster_parent = nbio;
999 bufx->b_bcount += PAGE_SIZE;
1007 * Flush out last buffer
1010 if (bufx->b_cmd == BUF_CMD_READ) {
1011 ++mycpu->gd_cnt.v_swapin;
1012 mycpu->gd_cnt.v_swappgsin += btoc(bufx->b_bcount);
1014 ++mycpu->gd_cnt.v_swapout;
1015 mycpu->gd_cnt.v_swappgsout += btoc(bufx->b_bcount);
1016 bufx->b_dirtyend = bufx->b_bcount;
1018 KKASSERT(bufx->b_bcount);
1019 if (bufx->b_cmd != BUF_CMD_READ)
1020 bufx->b_dirtyend = bufx->b_bcount;
1021 /* biox, bufx = NULL */
1025 * Now initiate all the I/O. Be careful looping on our chain as
1026 * I/O's may complete while we are still initiating them.
1028 nbio->bio_caller_info2.cluster_tail = NULL;
1029 bufx = nbio->bio_caller_info1.cluster_head;
1032 biox = &bufx->b_bio1;
1034 bufx = bufx->b_cluster_next;
1035 vn_strategy(swapdev_vp, biox);
1039 * Completion of the cluster will also call biodone_chain(nbio).
1040 * We never call biodone(nbio) so we don't have to worry about
1041 * setting up a bio_done callback. It's handled in the sub-IO.
1047 swap_chain_iodone(struct bio *biox)
1050 struct buf *bufx; /* chained sub-buffer */
1051 struct bio *nbio; /* parent nbio with chain glue */
1052 struct buf *bp; /* original bp associated with nbio */
1055 bufx = biox->bio_buf;
1056 nbio = biox->bio_caller_info1.cluster_parent;
1060 * Update the original buffer
1062 KKASSERT(bp != NULL);
1063 if (bufx->b_flags & B_ERROR) {
1064 atomic_set_int(&bufx->b_flags, B_ERROR);
1065 bp->b_error = bufx->b_error;
1066 } else if (bufx->b_resid != 0) {
1067 atomic_set_int(&bufx->b_flags, B_ERROR);
1068 bp->b_error = EINVAL;
1070 atomic_subtract_int(&bp->b_resid, bufx->b_bcount);
1074 * Remove us from the chain.
1076 spin_lock_wr(&bp->b_lock.lk_spinlock);
1077 nextp = &nbio->bio_caller_info1.cluster_head;
1078 while (*nextp != bufx) {
1079 KKASSERT(*nextp != NULL);
1080 nextp = &(*nextp)->b_cluster_next;
1082 *nextp = bufx->b_cluster_next;
1083 chain_empty = (nbio->bio_caller_info1.cluster_head == NULL);
1084 spin_unlock_wr(&bp->b_lock.lk_spinlock);
1087 * Clean up bufx. If the chain is now empty we finish out
1088 * the parent. Note that we may be racing other completions
1089 * so we must use the chain_empty status from above.
1092 if (bp->b_resid != 0 && !(bp->b_flags & B_ERROR)) {
1093 atomic_set_int(&bp->b_flags, B_ERROR);
1094 bp->b_error = EINVAL;
1096 biodone_chain(nbio);
1098 relpbuf(bufx, NULL);
1102 * SWAP_PAGER_GETPAGES() - bring page in from swap
1104 * The requested page may have to be brought in from swap. Calculate the
1105 * swap block and bring in additional pages if possible. All pages must
1106 * have contiguous swap block assignments and reside in the same object.
1108 * The caller has a single vm_object_pip_add() reference prior to
1109 * calling us and we should return with the same.
1111 * The caller has BUSY'd the page. We should return with (*mpp) left busy,
1112 * and any additinal pages unbusied.
1114 * If the caller encounters a PG_RAM page it will pass it to us even though
1115 * it may be valid and dirty. We cannot overwrite the page in this case!
1116 * The case is used to allow us to issue pure read-aheads.
1118 * NOTE! XXX This code does not entirely pipeline yet due to the fact that
1119 * the PG_RAM page is validated at the same time as mreq. What we
1120 * really need to do is issue a separate read-ahead pbuf.
1123 swap_pager_getpage(vm_object_t object, vm_page_t *mpp, int seqaccess)
1134 vm_page_t marray[XIO_INTERNAL_PAGES];
1138 if (mreq->object != object) {
1139 panic("swap_pager_getpages: object mismatch %p/%p",
1146 * We don't want to overwrite a fully valid page as it might be
1147 * dirty. This case can occur when e.g. vm_fault hits a perfectly
1148 * valid page with PG_RAM set.
1150 * In this case we see if the next page is a suitable page-in
1151 * candidate and if it is we issue read-ahead. PG_RAM will be
1152 * set on the last page of the read-ahead to continue the pipeline.
1154 if (mreq->valid == VM_PAGE_BITS_ALL) {
1155 if (swap_burst_read == 0 || mreq->pindex + 1 >= object->size)
1156 return(VM_PAGER_OK);
1158 blk = swp_pager_meta_ctl(object, mreq->pindex + 1, 0);
1159 if (blk == SWAPBLK_NONE) {
1161 return(VM_PAGER_OK);
1163 m = vm_page_lookup(object, mreq->pindex + 1);
1165 m = vm_page_alloc(object, mreq->pindex + 1,
1169 return(VM_PAGER_OK);
1172 if ((m->flags & PG_BUSY) || m->busy || m->valid) {
1174 return(VM_PAGER_OK);
1176 vm_page_unqueue_nowakeup(m);
1187 * Try to block-read contiguous pages from swap if sequential,
1188 * otherwise just read one page. Contiguous pages from swap must
1189 * reside within a single device stripe because the I/O cannot be
1190 * broken up across multiple stripes.
1192 * Note that blk and iblk can be SWAPBLK_NONE but the loop is
1193 * set up such that the case(s) are handled implicitly.
1196 blk = swp_pager_meta_ctl(mreq->object, mreq->pindex, 0);
1199 for (i = 1; swap_burst_read &&
1200 i < XIO_INTERNAL_PAGES &&
1201 mreq->pindex + i < object->size; ++i) {
1204 iblk = swp_pager_meta_ctl(object, mreq->pindex + i, 0);
1205 if (iblk != blk + i)
1207 if ((blk ^ iblk) & dmmax_mask)
1209 m = vm_page_lookup(object, mreq->pindex + i);
1211 m = vm_page_alloc(object, mreq->pindex + i,
1216 if ((m->flags & PG_BUSY) || m->busy || m->valid)
1218 vm_page_unqueue_nowakeup(m);
1224 vm_page_flag_set(marray[i - 1], PG_RAM);
1229 * If mreq is the requested page and we have nothing to do return
1230 * VM_PAGER_FAIL. If raonly is set mreq is just another read-ahead
1231 * page and must be cleaned up.
1233 if (blk == SWAPBLK_NONE) {
1236 vnode_pager_freepage(mreq);
1237 return(VM_PAGER_OK);
1239 return(VM_PAGER_FAIL);
1244 * map our page(s) into kva for input
1246 bp = getpbuf(&nsw_rcount);
1248 kva = (vm_offset_t) bp->b_kvabase;
1249 bcopy(marray, bp->b_xio.xio_pages, i * sizeof(vm_page_t));
1250 pmap_qenter(kva, bp->b_xio.xio_pages, i);
1252 bp->b_data = (caddr_t)kva;
1253 bp->b_bcount = PAGE_SIZE * i;
1254 bp->b_xio.xio_npages = i;
1255 bio->bio_done = swp_pager_async_iodone;
1256 bio->bio_offset = (off_t)blk << PAGE_SHIFT;
1257 bio->bio_caller_info1.index = SWBIO_READ;
1260 * Set index. If raonly set the index beyond the array so all
1261 * the pages are treated the same, otherwise the original mreq is
1265 bio->bio_driver_info = (void *)(intptr_t)i;
1267 bio->bio_driver_info = (void *)(intptr_t)0;
1269 for (j = 0; j < i; ++j)
1270 vm_page_flag_set(bp->b_xio.xio_pages[j], PG_SWAPINPROG);
1272 mycpu->gd_cnt.v_swapin++;
1273 mycpu->gd_cnt.v_swappgsin += bp->b_xio.xio_npages;
1276 * We still hold the lock on mreq, and our automatic completion routine
1277 * does not remove it.
1279 vm_object_pip_add(object, bp->b_xio.xio_npages);
1282 * perform the I/O. NOTE!!! bp cannot be considered valid after
1283 * this point because we automatically release it on completion.
1284 * Instead, we look at the one page we are interested in which we
1285 * still hold a lock on even through the I/O completion.
1287 * The other pages in our m[] array are also released on completion,
1288 * so we cannot assume they are valid anymore either.
1290 bp->b_cmd = BUF_CMD_READ;
1292 vn_strategy(swapdev_vp, bio);
1295 * Wait for the page we want to complete. PG_SWAPINPROG is always
1296 * cleared on completion. If an I/O error occurs, SWAPBLK_NONE
1297 * is set in the meta-data.
1299 * If this is a read-ahead only we return immediately without
1303 return(VM_PAGER_OK);
1306 * Read-ahead includes originally requested page case.
1309 while ((mreq->flags & PG_SWAPINPROG) != 0) {
1310 vm_page_flag_set(mreq, PG_WANTED | PG_REFERENCED);
1311 mycpu->gd_cnt.v_intrans++;
1312 if (tsleep(mreq, 0, "swread", hz*20)) {
1314 "swap_pager: indefinite wait buffer: "
1315 " offset: %lld, size: %ld\n",
1316 (long long)bio->bio_offset,
1324 * mreq is left bussied after completion, but all the other pages
1325 * are freed. If we had an unrecoverable read error the page will
1328 if (mreq->valid != VM_PAGE_BITS_ALL)
1329 return(VM_PAGER_ERROR);
1331 return(VM_PAGER_OK);
1334 * A final note: in a low swap situation, we cannot deallocate swap
1335 * and mark a page dirty here because the caller is likely to mark
1336 * the page clean when we return, causing the page to possibly revert
1337 * to all-zero's later.
1342 * swap_pager_putpages:
1344 * Assign swap (if necessary) and initiate I/O on the specified pages.
1346 * We support both OBJT_DEFAULT and OBJT_SWAP objects. DEFAULT objects
1347 * are automatically converted to SWAP objects.
1349 * In a low memory situation we may block in vn_strategy(), but the new
1350 * vm_page reservation system coupled with properly written VFS devices
1351 * should ensure that no low-memory deadlock occurs. This is an area
1354 * The parent has N vm_object_pip_add() references prior to
1355 * calling us and will remove references for rtvals[] that are
1356 * not set to VM_PAGER_PEND. We need to remove the rest on I/O
1359 * The parent has soft-busy'd the pages it passes us and will unbusy
1360 * those whos rtvals[] entry is not set to VM_PAGER_PEND on return.
1361 * We need to unbusy the rest on I/O completion.
1364 swap_pager_putpages(vm_object_t object, vm_page_t *m, int count,
1365 boolean_t sync, int *rtvals)
1370 if (count && m[0]->object != object) {
1371 panic("swap_pager_getpages: object mismatch %p/%p",
1380 * Turn object into OBJT_SWAP
1381 * check for bogus sysops
1382 * force sync if not pageout process
1384 if (object->type == OBJT_DEFAULT)
1385 swp_pager_meta_convert(object);
1387 if (curthread != pagethread)
1393 * Update nsw parameters from swap_async_max sysctl values.
1394 * Do not let the sysop crash the machine with bogus numbers.
1397 if (swap_async_max != nsw_wcount_async_max) {
1403 if ((n = swap_async_max) > nswbuf / 2)
1410 * Adjust difference ( if possible ). If the current async
1411 * count is too low, we may not be able to make the adjustment
1415 n -= nsw_wcount_async_max;
1416 if (nsw_wcount_async + n >= 0) {
1417 nsw_wcount_async += n;
1418 nsw_wcount_async_max += n;
1419 wakeup(&nsw_wcount_async);
1427 * Assign swap blocks and issue I/O. We reallocate swap on the fly.
1428 * The page is left dirty until the pageout operation completes
1432 for (i = 0; i < count; i += n) {
1439 * Maximum I/O size is limited by a number of factors.
1442 n = min(BLIST_MAX_ALLOC, count - i);
1443 n = min(n, nsw_cluster_max);
1448 * Get biggest block of swap we can. If we fail, fall
1449 * back and try to allocate a smaller block. Don't go
1450 * overboard trying to allocate space if it would overly
1454 (blk = swp_pager_getswapspace(object, n)) == SWAPBLK_NONE &&
1459 if (blk == SWAPBLK_NONE) {
1460 for (j = 0; j < n; ++j)
1461 rtvals[i+j] = VM_PAGER_FAIL;
1467 * The I/O we are constructing cannot cross a physical
1468 * disk boundry in the swap stripe. Note: we are still
1471 if ((blk ^ (blk + n)) & dmmax_mask) {
1472 j = ((blk + dmmax) & dmmax_mask) - blk;
1473 swp_pager_freeswapspace(object, blk + j, n - j);
1478 * All I/O parameters have been satisfied, build the I/O
1479 * request and assign the swap space.
1482 bp = getpbuf(&nsw_wcount_sync);
1484 bp = getpbuf(&nsw_wcount_async);
1487 pmap_qenter((vm_offset_t)bp->b_data, &m[i], n);
1489 bp->b_bcount = PAGE_SIZE * n;
1490 bio->bio_offset = (off_t)blk << PAGE_SHIFT;
1492 for (j = 0; j < n; ++j) {
1493 vm_page_t mreq = m[i+j];
1495 swp_pager_meta_build(mreq->object, mreq->pindex,
1497 if (object->type == OBJT_SWAP)
1498 vm_page_dirty(mreq);
1499 rtvals[i+j] = VM_PAGER_OK;
1501 vm_page_flag_set(mreq, PG_SWAPINPROG);
1502 bp->b_xio.xio_pages[j] = mreq;
1504 bp->b_xio.xio_npages = n;
1506 mycpu->gd_cnt.v_swapout++;
1507 mycpu->gd_cnt.v_swappgsout += bp->b_xio.xio_npages;
1511 bp->b_dirtyoff = 0; /* req'd for NFS */
1512 bp->b_dirtyend = bp->b_bcount; /* req'd for NFS */
1513 bp->b_cmd = BUF_CMD_WRITE;
1514 bio->bio_caller_info1.index = SWBIO_WRITE;
1519 if (sync == FALSE) {
1520 bio->bio_done = swp_pager_async_iodone;
1522 vn_strategy(swapdev_vp, bio);
1524 for (j = 0; j < n; ++j)
1525 rtvals[i+j] = VM_PAGER_PEND;
1530 * Issue synchrnously.
1532 * Wait for the sync I/O to complete, then update rtvals.
1533 * We just set the rtvals[] to VM_PAGER_PEND so we can call
1534 * our async completion routine at the end, thus avoiding a
1537 bio->bio_caller_info1.index |= SWBIO_SYNC;
1538 bio->bio_done = biodone_sync;
1539 bio->bio_flags |= BIO_SYNC;
1540 vn_strategy(swapdev_vp, bio);
1541 biowait(bio, "swwrt");
1543 for (j = 0; j < n; ++j)
1544 rtvals[i+j] = VM_PAGER_PEND;
1547 * Now that we are through with the bp, we can call the
1548 * normal async completion, which frees everything up.
1550 swp_pager_async_iodone(bio);
1555 swap_pager_newswap(void)
1561 * swp_pager_async_iodone:
1563 * Completion routine for asynchronous reads and writes from/to swap.
1564 * Also called manually by synchronous code to finish up a bp.
1566 * For READ operations, the pages are PG_BUSY'd. For WRITE operations,
1567 * the pages are vm_page_t->busy'd. For READ operations, we PG_BUSY
1568 * unbusy all pages except the 'main' request page. For WRITE
1569 * operations, we vm_page_t->busy'd unbusy all pages ( we can do this
1570 * because we marked them all VM_PAGER_PEND on return from putpages ).
1572 * This routine may not block.
1575 swp_pager_async_iodone(struct bio *bio)
1577 struct buf *bp = bio->bio_buf;
1578 vm_object_t object = NULL;
1585 if (bp->b_flags & B_ERROR) {
1587 "swap_pager: I/O error - %s failed; offset %lld,"
1588 "size %ld, error %d\n",
1589 ((bio->bio_caller_info1.index & SWBIO_READ) ?
1590 "pagein" : "pageout"),
1591 (long long)bio->bio_offset,
1598 * set object, raise to splvm().
1600 if (bp->b_xio.xio_npages)
1601 object = bp->b_xio.xio_pages[0]->object;
1605 * remove the mapping for kernel virtual
1607 pmap_qremove((vm_offset_t)bp->b_data, bp->b_xio.xio_npages);
1610 * cleanup pages. If an error occurs writing to swap, we are in
1611 * very serious trouble. If it happens to be a disk error, though,
1612 * we may be able to recover by reassigning the swap later on. So
1613 * in this case we remove the m->swapblk assignment for the page
1614 * but do not free it in the rlist. The errornous block(s) are thus
1615 * never reallocated as swap. Redirty the page and continue.
1617 for (i = 0; i < bp->b_xio.xio_npages; ++i) {
1618 vm_page_t m = bp->b_xio.xio_pages[i];
1620 if (bp->b_flags & B_ERROR) {
1622 * If an error occurs I'd love to throw the swapblk
1623 * away without freeing it back to swapspace, so it
1624 * can never be used again. But I can't from an
1628 if (bio->bio_caller_info1.index & SWBIO_READ) {
1630 * When reading, reqpage needs to stay
1631 * locked for the parent, but all other
1632 * pages can be freed. We still want to
1633 * wakeup the parent waiting on the page,
1634 * though. ( also: pg_reqpage can be -1 and
1635 * not match anything ).
1637 * We have to wake specifically requested pages
1638 * up too because we cleared PG_SWAPINPROG and
1639 * someone may be waiting for that.
1641 * NOTE: for reads, m->dirty will probably
1642 * be overridden by the original caller of
1643 * getpages so don't play cute tricks here.
1645 * NOTE: We can't actually free the page from
1646 * here, because this is an interrupt. It
1647 * is not legal to mess with object->memq
1648 * from an interrupt. Deactivate the page
1653 vm_page_flag_clear(m, PG_ZERO);
1654 vm_page_flag_clear(m, PG_SWAPINPROG);
1657 * bio_driver_info holds the requested page
1660 if (i != (int)(intptr_t)bio->bio_driver_info) {
1661 vm_page_deactivate(m);
1667 * If i == bp->b_pager.pg_reqpage, do not wake
1668 * the page up. The caller needs to.
1672 * If a write error occurs, reactivate page
1673 * so it doesn't clog the inactive list,
1674 * then finish the I/O.
1676 * Only for OBJT_SWAP. When using the swap
1677 * as a cache for clean vnode-backed pages
1678 * we don't mess with the page dirty state.
1680 vm_page_flag_clear(m, PG_SWAPINPROG);
1681 if (m->object->type == OBJT_SWAP) {
1683 vm_page_activate(m);
1685 vm_page_io_finish(m);
1687 } else if (bio->bio_caller_info1.index & SWBIO_READ) {
1689 * NOTE: for reads, m->dirty will probably be
1690 * overridden by the original caller of getpages so
1691 * we cannot set them in order to free the underlying
1692 * swap in a low-swap situation. I don't think we'd
1693 * want to do that anyway, but it was an optimization
1694 * that existed in the old swapper for a time before
1695 * it got ripped out due to precisely this problem.
1697 * clear PG_ZERO in page.
1699 * If not the requested page then deactivate it.
1701 * Note that the requested page, reqpage, is left
1702 * busied, but we still have to wake it up. The
1703 * other pages are released (unbusied) by
1704 * vm_page_wakeup(). We do not set reqpage's
1705 * valid bits here, it is up to the caller.
1709 * NOTE: can't call pmap_clear_modify(m) from an
1710 * interrupt thread, the pmap code may have to map
1711 * non-kernel pmaps and currently asserts the case.
1713 /*pmap_clear_modify(m);*/
1714 m->valid = VM_PAGE_BITS_ALL;
1716 vm_page_flag_clear(m, PG_ZERO | PG_SWAPINPROG);
1717 vm_page_flag_set(m, PG_SWAPPED);
1720 * We have to wake specifically requested pages
1721 * up too because we cleared PG_SWAPINPROG and
1722 * could be waiting for it in getpages. However,
1723 * be sure to not unbusy getpages specifically
1724 * requested page - getpages expects it to be
1727 * bio_driver_info holds the requested page
1729 if (i != (int)(intptr_t)bio->bio_driver_info) {
1730 vm_page_deactivate(m);
1737 * Mark the page clean but do not mess with the
1738 * pmap-layer's modified state. That state should
1739 * also be clear since the caller protected the
1740 * page VM_PROT_READ, but allow the case.
1742 * We are in an interrupt, avoid pmap operations.
1744 * If we have a severe page deficit, deactivate the
1745 * page. Do not try to cache it (which would also
1746 * involve a pmap op), because the page might still
1749 * When using the swap to cache clean vnode pages
1750 * we do not mess with the page dirty bits.
1752 if (m->object->type == OBJT_SWAP)
1754 vm_page_flag_clear(m, PG_SWAPINPROG);
1755 vm_page_flag_set(m, PG_SWAPPED);
1756 vm_page_io_finish(m);
1757 if (vm_page_count_severe())
1758 vm_page_deactivate(m);
1760 if (!vm_page_count_severe() || !vm_page_try_to_cache(m))
1761 vm_page_protect(m, VM_PROT_READ);
1767 * adjust pip. NOTE: the original parent may still have its own
1768 * pip refs on the object.
1772 vm_object_pip_wakeupn(object, bp->b_xio.xio_npages);
1775 * Release the physical I/O buffer.
1777 * NOTE: Due to synchronous operations in the write case b_cmd may
1778 * already be set to BUF_CMD_DONE and BIO_SYNC may have already
1781 if (bio->bio_caller_info1.index & SWBIO_READ)
1782 nswptr = &nsw_rcount;
1783 else if (bio->bio_caller_info1.index & SWBIO_SYNC)
1784 nswptr = &nsw_wcount_sync;
1786 nswptr = &nsw_wcount_async;
1787 bp->b_cmd = BUF_CMD_DONE;
1788 relpbuf(bp, nswptr);
1792 /************************************************************************
1794 ************************************************************************
1796 * These routines manipulate the swap metadata stored in the
1797 * OBJT_SWAP object. All swp_*() routines must be called at
1798 * splvm() because swap can be freed up by the low level vm_page
1799 * code which might be called from interrupts beyond what splbio() covers.
1801 * Swap metadata is implemented with a global hash and not directly
1802 * linked into the object. Instead the object simply contains
1803 * appropriate tracking counters.
1807 * Lookup the swblock containing the specified swap block index.
1811 swp_pager_lookup(vm_object_t object, vm_pindex_t index)
1813 index &= ~SWAP_META_MASK;
1814 return (RB_LOOKUP(swblock_rb_tree, &object->swblock_root, index));
1818 * Remove a swblock from the RB tree.
1822 swp_pager_remove(vm_object_t object, struct swblock *swap)
1824 RB_REMOVE(swblock_rb_tree, &object->swblock_root, swap);
1828 * Convert default object to swap object if necessary
1831 swp_pager_meta_convert(vm_object_t object)
1833 if (object->type == OBJT_DEFAULT) {
1834 object->type = OBJT_SWAP;
1835 KKASSERT(object->swblock_count == 0);
1840 * SWP_PAGER_META_BUILD() - add swap block to swap meta data for object
1842 * We first convert the object to a swap object if it is a default
1843 * object. Vnode objects do not need to be converted.
1845 * The specified swapblk is added to the object's swap metadata. If
1846 * the swapblk is not valid, it is freed instead. Any previously
1847 * assigned swapblk is freed.
1850 swp_pager_meta_build(vm_object_t object, vm_pindex_t index, daddr_t swapblk)
1852 struct swblock *swap;
1853 struct swblock *oswap;
1855 KKASSERT(swapblk != SWAPBLK_NONE);
1858 * Convert object if necessary
1860 if (object->type == OBJT_DEFAULT)
1861 swp_pager_meta_convert(object);
1864 * Locate swblock. If not found create, but if we aren't adding
1865 * anything just return. If we run out of space in the map we wait
1866 * and, since the hash table may have changed, retry.
1869 swap = swp_pager_lookup(object, index);
1874 swap = zalloc(swap_zone);
1879 swap->swb_index = index & ~SWAP_META_MASK;
1880 swap->swb_count = 0;
1882 ++object->swblock_count;
1884 for (i = 0; i < SWAP_META_PAGES; ++i)
1885 swap->swb_pages[i] = SWAPBLK_NONE;
1886 oswap = RB_INSERT(swblock_rb_tree, &object->swblock_root, swap);
1887 KKASSERT(oswap == NULL);
1891 * Delete prior contents of metadata
1894 index &= SWAP_META_MASK;
1896 if (swap->swb_pages[index] != SWAPBLK_NONE) {
1897 swp_pager_freeswapspace(object, swap->swb_pages[index], 1);
1902 * Enter block into metadata
1904 swap->swb_pages[index] = swapblk;
1905 if (swapblk != SWAPBLK_NONE)
1910 * SWP_PAGER_META_FREE() - free a range of blocks in the object's swap metadata
1912 * The requested range of blocks is freed, with any associated swap
1913 * returned to the swap bitmap.
1915 * This routine will free swap metadata structures as they are cleaned
1916 * out. This routine does *NOT* operate on swap metadata associated
1917 * with resident pages.
1919 * This routine must be called at splvm()
1921 static int swp_pager_meta_free_callback(struct swblock *swb, void *data);
1924 swp_pager_meta_free(vm_object_t object, vm_pindex_t index, vm_pindex_t count)
1926 struct swfreeinfo info;
1931 if (object->swblock_count == 0) {
1932 KKASSERT(RB_EMPTY(&object->swblock_root));
1939 * Setup for RB tree scan. Note that the pindex range can be huge
1940 * due to the 64 bit page index space so we cannot safely iterate.
1942 info.object = object;
1943 info.basei = index & ~SWAP_META_MASK;
1945 info.endi = index + count - 1;
1946 swblock_rb_tree_RB_SCAN(&object->swblock_root, rb_swblock_scancmp,
1947 swp_pager_meta_free_callback, &info);
1952 swp_pager_meta_free_callback(struct swblock *swap, void *data)
1954 struct swfreeinfo *info = data;
1955 vm_object_t object = info->object;
1960 * Figure out the range within the swblock. The wider scan may
1961 * return edge-case swap blocks when the start and/or end points
1962 * are in the middle of a block.
1964 if (swap->swb_index < info->begi)
1965 index = (int)info->begi & SWAP_META_MASK;
1969 if (swap->swb_index + SWAP_META_PAGES > info->endi)
1970 eindex = (int)info->endi & SWAP_META_MASK;
1972 eindex = SWAP_META_MASK;
1975 * Scan and free the blocks. The loop terminates early
1976 * if (swap) runs out of blocks and could be freed.
1978 while (index <= eindex) {
1979 daddr_t v = swap->swb_pages[index];
1981 if (v != SWAPBLK_NONE) {
1982 swp_pager_freeswapspace(object, v, 1);
1983 swap->swb_pages[index] = SWAPBLK_NONE;
1984 if (--swap->swb_count == 0) {
1985 swp_pager_remove(object, swap);
1986 zfree(swap_zone, swap);
1987 --object->swblock_count;
1993 /* swap may be invalid here due to zfree above */
1998 * SWP_PAGER_META_FREE_ALL() - destroy all swap metadata associated with object
2000 * This routine locates and destroys all swap metadata associated with
2003 * This routine must be called at splvm()
2006 swp_pager_meta_free_all(vm_object_t object)
2008 struct swblock *swap;
2011 while ((swap = RB_ROOT(&object->swblock_root)) != NULL) {
2012 swp_pager_remove(object, swap);
2013 for (i = 0; i < SWAP_META_PAGES; ++i) {
2014 daddr_t v = swap->swb_pages[i];
2015 if (v != SWAPBLK_NONE) {
2017 swp_pager_freeswapspace(object, v, 1);
2020 if (swap->swb_count != 0)
2021 panic("swap_pager_meta_free_all: swb_count != 0");
2022 zfree(swap_zone, swap);
2023 --object->swblock_count;
2025 KKASSERT(object->swblock_count == 0);
2029 * SWP_PAGER_METACTL() - misc control of swap and vm_page_t meta data.
2031 * This routine is capable of looking up, popping, or freeing
2032 * swapblk assignments in the swap meta data or in the vm_page_t.
2033 * The routine typically returns the swapblk being looked-up, or popped,
2034 * or SWAPBLK_NONE if the block was freed, or SWAPBLK_NONE if the block
2035 * was invalid. This routine will automatically free any invalid
2036 * meta-data swapblks.
2038 * It is not possible to store invalid swapblks in the swap meta data
2039 * (other then a literal 'SWAPBLK_NONE'), so we don't bother checking.
2041 * When acting on a busy resident page and paging is in progress, we
2042 * have to wait until paging is complete but otherwise can act on the
2045 * This routine must be called at splvm().
2047 * SWM_FREE remove and free swap block from metadata
2048 * SWM_POP remove from meta data but do not free.. pop it out
2051 swp_pager_meta_ctl(vm_object_t object, vm_pindex_t index, int flags)
2053 struct swblock *swap;
2056 if (object->swblock_count == 0)
2057 return(SWAPBLK_NONE);
2060 swap = swp_pager_lookup(object, index);
2063 index &= SWAP_META_MASK;
2064 r1 = swap->swb_pages[index];
2066 if (r1 != SWAPBLK_NONE) {
2067 if (flags & SWM_FREE) {
2068 swp_pager_freeswapspace(object, r1, 1);
2071 if (flags & (SWM_FREE|SWM_POP)) {
2072 swap->swb_pages[index] = SWAPBLK_NONE;
2073 if (--swap->swb_count == 0) {
2074 swp_pager_remove(object, swap);
2075 zfree(swap_zone, swap);
2076 --object->swblock_count;