4 * Copyright (c) 1998-2010 The DragonFly Project. All rights reserved.
6 * This code is derived from software contributed to The DragonFly Project
7 * by Matthew Dillon <dillon@backplane.com>
9 * Redistribution and use in source and binary forms, with or without
10 * modification, are permitted provided that the following conditions
13 * 1. Redistributions of source code must retain the above copyright
14 * notice, this list of conditions and the following disclaimer.
15 * 2. Redistributions in binary form must reproduce the above copyright
16 * notice, this list of conditions and the following disclaimer in
17 * the documentation and/or other materials provided with the
19 * 3. Neither the name of The DragonFly Project nor the names of its
20 * contributors may be used to endorse or promote products derived
21 * from this software without specific, prior written permission.
23 * THIS SOFTWARE IS PROVIDED BY THE COPYRIGHT HOLDERS AND CONTRIBUTORS
24 * ``AS IS'' AND ANY EXPRESS OR IMPLIED WARRANTIES, INCLUDING, BUT NOT
25 * LIMITED TO, THE IMPLIED WARRANTIES OF MERCHANTABILITY AND FITNESS
26 * FOR A PARTICULAR PURPOSE ARE DISCLAIMED. IN NO EVENT SHALL THE
27 * COPYRIGHT HOLDERS OR CONTRIBUTORS BE LIABLE FOR ANY DIRECT, INDIRECT,
28 * INCIDENTAL, SPECIAL, EXEMPLARY OR CONSEQUENTIAL DAMAGES (INCLUDING,
29 * BUT NOT LIMITED TO, PROCUREMENT OF SUBSTITUTE GOODS OR SERVICES;
30 * LOSS OF USE, DATA, OR PROFITS; OR BUSINESS INTERRUPTION) HOWEVER CAUSED
31 * AND ON ANY THEORY OF LIABILITY, WHETHER IN CONTRACT, STRICT LIABILITY,
32 * OR TORT (INCLUDING NEGLIGENCE OR OTHERWISE) ARISING IN ANY WAY OUT
33 * OF THE USE OF THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF
36 * Copyright (c) 1994 John S. Dyson
37 * Copyright (c) 1990 University of Utah.
38 * Copyright (c) 1991, 1993
39 * The Regents of the University of California. All rights reserved.
41 * This code is derived from software contributed to Berkeley by
42 * the Systems Programming Group of the University of Utah Computer
45 * Redistribution and use in source and binary forms, with or without
46 * modification, are permitted provided that the following conditions
48 * 1. Redistributions of source code must retain the above copyright
49 * notice, this list of conditions and the following disclaimer.
50 * 2. Redistributions in binary form must reproduce the above copyright
51 * notice, this list of conditions and the following disclaimer in the
52 * documentation and/or other materials provided with the distribution.
53 * 3. Neither the name of the University nor the names of its contributors
54 * may be used to endorse or promote products derived from this software
55 * without specific prior written permission.
57 * THIS SOFTWARE IS PROVIDED BY THE REGENTS AND CONTRIBUTORS ``AS IS'' AND
58 * ANY EXPRESS OR IMPLIED WARRANTIES, INCLUDING, BUT NOT LIMITED TO, THE
59 * IMPLIED WARRANTIES OF MERCHANTABILITY AND FITNESS FOR A PARTICULAR PURPOSE
60 * ARE DISCLAIMED. IN NO EVENT SHALL THE REGENTS OR CONTRIBUTORS BE LIABLE
61 * FOR ANY DIRECT, INDIRECT, INCIDENTAL, SPECIAL, EXEMPLARY, OR CONSEQUENTIAL
62 * DAMAGES (INCLUDING, BUT NOT LIMITED TO, PROCUREMENT OF SUBSTITUTE GOODS
63 * OR SERVICES; LOSS OF USE, DATA, OR PROFITS; OR BUSINESS INTERRUPTION)
64 * HOWEVER CAUSED AND ON ANY THEORY OF LIABILITY, WHETHER IN CONTRACT, STRICT
65 * LIABILITY, OR TORT (INCLUDING NEGLIGENCE OR OTHERWISE) ARISING IN ANY WAY
66 * OUT OF THE USE OF THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF
72 * Radix Bitmap 'blists'.
74 * - The new swapper uses the new radix bitmap code. This should scale
75 * to arbitrarily small or arbitrarily large swap spaces and an almost
76 * arbitrary degree of fragmentation.
80 * - on the fly reallocation of swap during putpages. The new system
81 * does not try to keep previously allocated swap blocks for dirty
84 * - on the fly deallocation of swap
86 * - No more garbage collection required. Unnecessarily allocated swap
87 * blocks only exist for dirty vm_page_t's now and these are already
88 * cycled (in a high-load system) by the pager. We also do on-the-fly
89 * removal of invalidated swap blocks when a page is destroyed
92 * from: Utah $Hdr: swap_pager.c 1.4 91/04/30$
93 * @(#)swap_pager.c 8.9 (Berkeley) 3/21/94
94 * $FreeBSD: src/sys/vm/swap_pager.c,v 1.130.2.12 2002/08/31 21:15:55 dillon Exp $
97 #include <sys/param.h>
98 #include <sys/systm.h>
100 #include <sys/kernel.h>
101 #include <sys/proc.h>
103 #include <sys/vnode.h>
104 #include <sys/malloc.h>
105 #include <sys/vmmeter.h>
106 #include <sys/sysctl.h>
107 #include <sys/blist.h>
108 #include <sys/lock.h>
109 #include <sys/thread2.h>
111 #include "opt_swap.h"
113 #include <vm/vm_object.h>
114 #include <vm/vm_page.h>
115 #include <vm/vm_pager.h>
116 #include <vm/vm_pageout.h>
117 #include <vm/swap_pager.h>
118 #include <vm/vm_extern.h>
119 #include <vm/vm_zone.h>
120 #include <vm/vnode_pager.h>
122 #include <sys/buf2.h>
123 #include <vm/vm_page2.h>
125 #ifndef MAX_PAGEOUT_CLUSTER
126 #define MAX_PAGEOUT_CLUSTER SWB_NPAGES
129 #define SWM_FREE 0x02 /* free, period */
130 #define SWM_POP 0x04 /* pop out */
132 #define SWBIO_READ 0x01
133 #define SWBIO_WRITE 0x02
134 #define SWBIO_SYNC 0x04
140 vm_pindex_t endi; /* inclusive */
143 struct swswapoffinfo {
150 * vm_swap_size is in page-sized chunks now. It was DEV_BSIZE'd chunks
154 int swap_pager_full; /* swap space exhaustion (task killing) */
155 int swap_fail_ticks; /* when we became exhausted */
156 int swap_pager_almost_full; /* swap space exhaustion (w/ hysteresis)*/
157 int vm_swap_cache_use;
158 int vm_swap_anon_use;
159 static int vm_report_swap_allocs;
161 static int nsw_rcount; /* free read buffers */
162 static int nsw_wcount_sync; /* limit write buffers / synchronous */
163 static int nsw_wcount_async; /* limit write buffers / asynchronous */
164 static int nsw_wcount_async_max;/* assigned maximum */
165 static int nsw_cluster_max; /* maximum VOP I/O allowed */
167 struct blist *swapblist;
168 static int swap_async_max = 4; /* maximum in-progress async I/O's */
169 static int swap_burst_read = 0; /* allow burst reading */
170 static swblk_t swapiterator; /* linearize allocations */
172 static struct spinlock swapbp_spin = SPINLOCK_INITIALIZER(&swapbp_spin, "swapbp_spin");
175 extern struct vnode *swapdev_vp;
176 extern struct swdevt *swdevt;
179 #define BLK2DEVIDX(blk) (nswdev > 1 ? blk / dmmax % nswdev : 0)
181 SYSCTL_INT(_vm, OID_AUTO, swap_async_max,
182 CTLFLAG_RW, &swap_async_max, 0, "Maximum running async swap ops");
183 SYSCTL_INT(_vm, OID_AUTO, swap_burst_read,
184 CTLFLAG_RW, &swap_burst_read, 0, "Allow burst reads for pageins");
186 SYSCTL_INT(_vm, OID_AUTO, swap_cache_use,
187 CTLFLAG_RD, &vm_swap_cache_use, 0, "");
188 SYSCTL_INT(_vm, OID_AUTO, swap_anon_use,
189 CTLFLAG_RD, &vm_swap_anon_use, 0, "");
190 SYSCTL_INT(_vm, OID_AUTO, swap_size,
191 CTLFLAG_RD, &vm_swap_size, 0, "");
192 SYSCTL_INT(_vm, OID_AUTO, report_swap_allocs,
193 CTLFLAG_RW, &vm_report_swap_allocs, 0, "");
198 * Red-Black tree for swblock entries
200 * The caller must hold vm_token
202 RB_GENERATE2(swblock_rb_tree, swblock, swb_entry, rb_swblock_compare,
203 vm_pindex_t, swb_index);
206 rb_swblock_compare(struct swblock *swb1, struct swblock *swb2)
208 if (swb1->swb_index < swb2->swb_index)
210 if (swb1->swb_index > swb2->swb_index)
217 rb_swblock_scancmp(struct swblock *swb, void *data)
219 struct swfreeinfo *info = data;
221 if (swb->swb_index < info->basei)
223 if (swb->swb_index > info->endi)
230 rb_swblock_condcmp(struct swblock *swb, void *data)
232 struct swfreeinfo *info = data;
234 if (swb->swb_index < info->basei)
240 * pagerops for OBJT_SWAP - "swap pager". Some ops are also global procedure
241 * calls hooked from other parts of the VM system and do not appear here.
242 * (see vm/swap_pager.h).
245 static void swap_pager_dealloc (vm_object_t object);
246 static int swap_pager_getpage (vm_object_t, vm_page_t *, int);
247 static void swap_chain_iodone(struct bio *biox);
249 struct pagerops swappagerops = {
250 swap_pager_dealloc, /* deallocate an OBJT_SWAP object */
251 swap_pager_getpage, /* pagein */
252 swap_pager_putpages, /* pageout */
253 swap_pager_haspage /* get backing store status for page */
257 * dmmax is in page-sized chunks with the new swap system. It was
258 * dev-bsized chunks in the old. dmmax is always a power of 2.
260 * swap_*() routines are externally accessible. swp_*() routines are
265 static int dmmax_mask;
266 int nswap_lowat = 128; /* in pages, swap_pager_almost_full warn */
267 int nswap_hiwat = 512; /* in pages, swap_pager_almost_full warn */
269 static __inline void swp_sizecheck (void);
270 static void swp_pager_async_iodone (struct bio *bio);
273 * Swap bitmap functions
276 static __inline void swp_pager_freeswapspace(vm_object_t object,
277 swblk_t blk, int npages);
278 static __inline swblk_t swp_pager_getswapspace(vm_object_t object, int npages);
284 static void swp_pager_meta_convert(vm_object_t);
285 static void swp_pager_meta_build(vm_object_t, vm_pindex_t, swblk_t);
286 static void swp_pager_meta_free(vm_object_t, vm_pindex_t, vm_pindex_t);
287 static void swp_pager_meta_free_all(vm_object_t);
288 static swblk_t swp_pager_meta_ctl(vm_object_t, vm_pindex_t, int);
291 * SWP_SIZECHECK() - update swap_pager_full indication
293 * update the swap_pager_almost_full indication and warn when we are
294 * about to run out of swap space, using lowat/hiwat hysteresis.
296 * Clear swap_pager_full ( task killing ) indication when lowat is met.
298 * No restrictions on call
299 * This routine may not block.
305 if (vm_swap_size < nswap_lowat) {
306 if (swap_pager_almost_full == 0) {
307 kprintf("swap_pager: out of swap space\n");
308 swap_pager_almost_full = 1;
309 swap_fail_ticks = ticks;
313 if (vm_swap_size > nswap_hiwat)
314 swap_pager_almost_full = 0;
319 * SWAP_PAGER_INIT() - initialize the swap pager!
321 * Expected to be started from system init. NOTE: This code is run
322 * before much else so be careful what you depend on. Most of the VM
323 * system has yet to be initialized at this point.
325 * Called from the low level boot code only.
328 swap_pager_init(void *arg __unused)
331 * Device Stripe, in PAGE_SIZE'd blocks
333 dmmax = SWB_NPAGES * 2;
334 dmmax_mask = ~(dmmax - 1);
336 SYSINIT(vm_mem, SI_BOOT1_VM, SI_ORDER_THIRD, swap_pager_init, NULL);
339 * SWAP_PAGER_SWAP_INIT() - swap pager initialization from pageout process
341 * Expected to be started from pageout process once, prior to entering
344 * Called from the low level boot code only.
347 swap_pager_swap_init(void)
352 * Number of in-transit swap bp operations. Don't
353 * exhaust the pbufs completely. Make sure we
354 * initialize workable values (0 will work for hysteresis
355 * but it isn't very efficient).
357 * The nsw_cluster_max is constrained by the number of pages an XIO
358 * holds, i.e., (MAXPHYS/PAGE_SIZE) and our locally defined
359 * MAX_PAGEOUT_CLUSTER. Also be aware that swap ops are
360 * constrained by the swap device interleave stripe size.
362 * Currently we hardwire nsw_wcount_async to 4. This limit is
363 * designed to prevent other I/O from having high latencies due to
364 * our pageout I/O. The value 4 works well for one or two active swap
365 * devices but is probably a little low if you have more. Even so,
366 * a higher value would probably generate only a limited improvement
367 * with three or four active swap devices since the system does not
368 * typically have to pageout at extreme bandwidths. We will want
369 * at least 2 per swap devices, and 4 is a pretty good value if you
370 * have one NFS swap device due to the command/ack latency over NFS.
371 * So it all works out pretty well.
374 nsw_cluster_max = min((MAXPHYS/PAGE_SIZE), MAX_PAGEOUT_CLUSTER);
376 nsw_rcount = (nswbuf_kva + 1) / 2;
377 nsw_wcount_sync = (nswbuf_kva + 3) / 4;
378 nsw_wcount_async = 4;
379 nsw_wcount_async_max = nsw_wcount_async;
382 * The zone is dynamically allocated so generally size it to
383 * maxswzone (32MB to 512MB of KVM). Set a minimum size based
384 * on physical memory of around 8x (each swblock can hold 16 pages).
386 * With the advent of SSDs (vs HDs) the practical (swap:memory) ratio
387 * has increased dramatically.
389 n = vmstats.v_page_count / 2;
390 if (maxswzone && n < maxswzone / sizeof(struct swblock))
391 n = maxswzone / sizeof(struct swblock);
397 sizeof(struct swblock),
401 if (swap_zone != NULL)
404 * if the allocation failed, try a zone two thirds the
405 * size of the previous attempt.
410 if (swap_zone == NULL)
411 panic("swap_pager_swap_init: swap_zone == NULL");
413 kprintf("Swap zone entries reduced from %d to %d.\n", n2, n);
417 * SWAP_PAGER_ALLOC() - allocate a new OBJT_SWAP VM object and instantiate
418 * its metadata structures.
420 * This routine is called from the mmap and fork code to create a new
421 * OBJT_SWAP object. We do this by creating an OBJT_DEFAULT object
422 * and then converting it with swp_pager_meta_convert().
424 * We only support unnamed objects.
429 swap_pager_alloc(void *handle, off_t size, vm_prot_t prot, off_t offset)
433 KKASSERT(handle == NULL);
434 object = vm_object_allocate_hold(OBJT_DEFAULT,
435 OFF_TO_IDX(offset + PAGE_MASK + size));
436 swp_pager_meta_convert(object);
437 vm_object_drop(object);
443 * SWAP_PAGER_DEALLOC() - remove swap metadata from object
445 * The swap backing for the object is destroyed. The code is
446 * designed such that we can reinstantiate it later, but this
447 * routine is typically called only when the entire object is
448 * about to be destroyed.
450 * The object must be locked or unreferenceable.
451 * No other requirements.
454 swap_pager_dealloc(vm_object_t object)
456 vm_object_hold(object);
457 vm_object_pip_wait(object, "swpdea");
460 * Free all remaining metadata. We only bother to free it from
461 * the swap meta data. We do not attempt to free swapblk's still
462 * associated with vm_page_t's for this object. We do not care
463 * if paging is still in progress on some objects.
465 swp_pager_meta_free_all(object);
466 vm_object_drop(object);
469 /************************************************************************
470 * SWAP PAGER BITMAP ROUTINES *
471 ************************************************************************/
474 * SWP_PAGER_GETSWAPSPACE() - allocate raw swap space
476 * Allocate swap for the requested number of pages. The starting
477 * swap block number (a page index) is returned or SWAPBLK_NONE
478 * if the allocation failed.
480 * Also has the side effect of advising that somebody made a mistake
481 * when they configured swap and didn't configure enough.
483 * The caller must hold the object.
484 * This routine may not block.
486 static __inline swblk_t
487 swp_pager_getswapspace(vm_object_t object, int npages)
491 lwkt_gettoken(&vm_token);
492 blk = blist_allocat(swapblist, npages, swapiterator);
493 if (blk == SWAPBLK_NONE)
494 blk = blist_allocat(swapblist, npages, 0);
495 if (blk == SWAPBLK_NONE) {
496 if (swap_pager_full != 2) {
497 if (vm_swap_size == 0)
498 kprintf("Warning: The system would like to "
499 "page to swap but no swap space "
502 kprintf("swap_pager_getswapspace: "
503 "unable to allocate=%d pages\n",
506 if (swap_pager_almost_full == 0)
507 swap_fail_ticks = ticks;
508 swap_pager_almost_full = 1;
511 /* swapiterator = blk; disable for now, doesn't work well */
512 swapacctspace(blk, -npages);
513 if (object->type == OBJT_SWAP)
514 vm_swap_anon_use += npages;
516 vm_swap_cache_use += npages;
519 lwkt_reltoken(&vm_token);
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.
538 swp_pager_freeswapspace(vm_object_t object, swblk_t blk, int npages)
540 struct swdevt *sp = &swdevt[BLK2DEVIDX(blk)];
542 lwkt_gettoken(&vm_token);
543 sp->sw_nused -= npages;
544 if (object->type == OBJT_SWAP)
545 vm_swap_anon_use -= npages;
547 vm_swap_cache_use -= npages;
549 if (sp->sw_flags & SW_CLOSING) {
550 lwkt_reltoken(&vm_token);
554 blist_free(swapblist, blk, npages);
555 vm_swap_size += npages;
557 lwkt_reltoken(&vm_token);
561 * SWAP_PAGER_FREESPACE() - frees swap blocks associated with a page
562 * range within an object.
564 * This is a globally accessible routine.
566 * This routine removes swapblk assignments from swap metadata.
568 * The external callers of this routine typically have already destroyed
569 * or renamed vm_page_t's associated with this range in the object so
575 swap_pager_freespace(vm_object_t object, vm_pindex_t start, vm_pindex_t size)
577 vm_object_hold(object);
578 swp_pager_meta_free(object, start, size);
579 vm_object_drop(object);
586 swap_pager_freespace_all(vm_object_t object)
588 vm_object_hold(object);
589 swp_pager_meta_free_all(object);
590 vm_object_drop(object);
594 * This function conditionally frees swap cache swap starting at
595 * (*basei) in the object. (count) swap blocks will be nominally freed.
596 * The actual number of blocks freed can be more or less than the
599 * This function nominally returns the number of blocks freed. However,
600 * the actual number of blocks freed may be less then the returned value.
601 * If the function is unable to exhaust the object or if it is able to
602 * free (approximately) the requested number of blocks it returns
605 * If we exhaust the object we will return a value n <= count.
607 * The caller must hold the object.
609 * WARNING! If count == 0 then -1 can be returned as a degenerate case,
610 * callers should always pass a count value > 0.
612 static int swap_pager_condfree_callback(struct swblock *swap, void *data);
615 swap_pager_condfree(vm_object_t object, vm_pindex_t *basei, int count)
617 struct swfreeinfo info;
621 ASSERT_LWKT_TOKEN_HELD(vm_object_token(object));
623 info.object = object;
624 info.basei = *basei; /* skip up to this page index */
625 info.begi = count; /* max swap pages to destroy */
626 info.endi = count * 8; /* max swblocks to scan */
628 swblock_rb_tree_RB_SCAN(&object->swblock_root, rb_swblock_condcmp,
629 swap_pager_condfree_callback, &info);
633 * Take the higher difference swblocks vs pages
635 n = count - (int)info.begi;
636 t = count * 8 - (int)info.endi;
645 * The idea is to free whole meta-block to avoid fragmenting
646 * the swap space or disk I/O. We only do this if NO VM pages
649 * We do not have to deal with clearing PG_SWAPPED in related VM
650 * pages because there are no related VM pages.
652 * The caller must hold the object.
655 swap_pager_condfree_callback(struct swblock *swap, void *data)
657 struct swfreeinfo *info = data;
658 vm_object_t object = info->object;
661 for (i = 0; i < SWAP_META_PAGES; ++i) {
662 if (vm_page_lookup(object, swap->swb_index + i))
665 info->basei = swap->swb_index + SWAP_META_PAGES;
666 if (i == SWAP_META_PAGES) {
667 info->begi -= swap->swb_count;
668 swap_pager_freespace(object, swap->swb_index, SWAP_META_PAGES);
671 if ((int)info->begi < 0 || (int)info->endi < 0)
678 * Called by vm_page_alloc() when a new VM page is inserted
679 * into a VM object. Checks whether swap has been assigned to
680 * the page and sets PG_SWAPPED as necessary.
685 swap_pager_page_inserted(vm_page_t m)
687 if (m->object->swblock_count) {
688 vm_object_hold(m->object);
689 if (swp_pager_meta_ctl(m->object, m->pindex, 0) != SWAPBLK_NONE)
690 vm_page_flag_set(m, PG_SWAPPED);
691 vm_object_drop(m->object);
696 * SWAP_PAGER_RESERVE() - reserve swap blocks in object
698 * Assigns swap blocks to the specified range within the object. The
699 * swap blocks are not zerod. Any previous swap assignment is destroyed.
701 * Returns 0 on success, -1 on failure.
703 * The caller is responsible for avoiding races in the specified range.
704 * No other requirements.
707 swap_pager_reserve(vm_object_t object, vm_pindex_t start, vm_size_t size)
710 swblk_t blk = SWAPBLK_NONE;
711 vm_pindex_t beg = start; /* save start index */
713 vm_object_hold(object);
718 while ((blk = swp_pager_getswapspace(object, n)) ==
723 swp_pager_meta_free(object, beg,
725 vm_object_drop(object);
730 swp_pager_meta_build(object, start, blk);
736 swp_pager_meta_free(object, start, n);
737 vm_object_drop(object);
742 * SWAP_PAGER_COPY() - copy blocks from source pager to destination pager
743 * and destroy the source.
745 * Copy any valid swapblks from the source to the destination. In
746 * cases where both the source and destination have a valid swapblk,
747 * we keep the destination's.
749 * This routine is allowed to block. It may block allocating metadata
750 * indirectly through swp_pager_meta_build() or if paging is still in
751 * progress on the source.
753 * XXX vm_page_collapse() kinda expects us not to block because we
754 * supposedly do not need to allocate memory, but for the moment we
755 * *may* have to get a little memory from the zone allocator, but
756 * it is taken from the interrupt memory. We should be ok.
758 * The source object contains no vm_page_t's (which is just as well)
759 * The source object is of type OBJT_SWAP.
761 * The source and destination objects must be held by the caller.
764 swap_pager_copy(vm_object_t srcobject, vm_object_t dstobject,
765 vm_pindex_t base_index, int destroysource)
769 ASSERT_LWKT_TOKEN_HELD(vm_object_token(srcobject));
770 ASSERT_LWKT_TOKEN_HELD(vm_object_token(dstobject));
773 * transfer source to destination.
775 for (i = 0; i < dstobject->size; ++i) {
779 * Locate (without changing) the swapblk on the destination,
780 * unless it is invalid in which case free it silently, or
781 * if the destination is a resident page, in which case the
782 * source is thrown away.
784 dstaddr = swp_pager_meta_ctl(dstobject, i, 0);
786 if (dstaddr == SWAPBLK_NONE) {
788 * Destination has no swapblk and is not resident,
793 srcaddr = swp_pager_meta_ctl(srcobject,
794 base_index + i, SWM_POP);
796 if (srcaddr != SWAPBLK_NONE)
797 swp_pager_meta_build(dstobject, i, srcaddr);
800 * Destination has valid swapblk or it is represented
801 * by a resident page. We destroy the sourceblock.
803 swp_pager_meta_ctl(srcobject, base_index + i, SWM_FREE);
808 * Free left over swap blocks in source.
810 * We have to revert the type to OBJT_DEFAULT so we do not accidently
811 * double-remove the object from the swap queues.
815 * Reverting the type is not necessary, the caller is going
816 * to destroy srcobject directly, but I'm doing it here
817 * for consistency since we've removed the object from its
820 swp_pager_meta_free_all(srcobject);
821 if (srcobject->type == OBJT_SWAP)
822 srcobject->type = OBJT_DEFAULT;
827 * SWAP_PAGER_HASPAGE() - determine if we have good backing store for
828 * the requested page.
830 * We determine whether good backing store exists for the requested
831 * page and return TRUE if it does, FALSE if it doesn't.
833 * If TRUE, we also try to determine how much valid, contiguous backing
834 * store exists before and after the requested page within a reasonable
835 * distance. We do not try to restrict it to the swap device stripe
836 * (that is handled in getpages/putpages). It probably isn't worth
842 swap_pager_haspage(vm_object_t object, vm_pindex_t pindex)
847 * do we have good backing store at the requested index ?
849 vm_object_hold(object);
850 blk0 = swp_pager_meta_ctl(object, pindex, 0);
852 if (blk0 == SWAPBLK_NONE) {
853 vm_object_drop(object);
856 vm_object_drop(object);
861 * SWAP_PAGER_PAGE_UNSWAPPED() - remove swap backing store related to page
863 * This removes any associated swap backing store, whether valid or
864 * not, from the page. This operates on any VM object, not just OBJT_SWAP
867 * This routine is typically called when a page is made dirty, at
868 * which point any associated swap can be freed. MADV_FREE also
869 * calls us in a special-case situation
871 * NOTE!!! If the page is clean and the swap was valid, the caller
872 * should make the page dirty before calling this routine. This routine
873 * does NOT change the m->dirty status of the page. Also: MADV_FREE
876 * The page must be busied or soft-busied.
877 * The caller can hold the object to avoid blocking, else we might block.
878 * No other requirements.
881 swap_pager_unswapped(vm_page_t m)
883 if (m->flags & PG_SWAPPED) {
884 vm_object_hold(m->object);
885 KKASSERT(m->flags & PG_SWAPPED);
886 swp_pager_meta_ctl(m->object, m->pindex, SWM_FREE);
887 vm_page_flag_clear(m, PG_SWAPPED);
888 vm_object_drop(m->object);
893 * SWAP_PAGER_STRATEGY() - read, write, free blocks
895 * This implements a VM OBJECT strategy function using swap backing store.
896 * This can operate on any VM OBJECT type, not necessarily just OBJT_SWAP
899 * This is intended to be a cacheless interface (i.e. caching occurs at
900 * higher levels), and is also used as a swap-based SSD cache for vnode
901 * and device objects.
903 * All I/O goes directly to and from the swap device.
905 * We currently attempt to run I/O synchronously or asynchronously as
906 * the caller requests. This isn't perfect because we loose error
907 * sequencing when we run multiple ops in parallel to satisfy a request.
908 * But this is swap, so we let it all hang out.
913 swap_pager_strategy(vm_object_t object, struct bio *bio)
915 struct buf *bp = bio->bio_buf;
918 vm_pindex_t biox_blkno = 0;
924 struct bio_track *track;
929 * tracking for swapdev vnode I/Os
931 if (bp->b_cmd == BUF_CMD_READ)
932 track = &swapdev_vp->v_track_read;
934 track = &swapdev_vp->v_track_write;
937 if (bp->b_bcount & PAGE_MASK) {
938 bp->b_error = EINVAL;
939 bp->b_flags |= B_ERROR | B_INVAL;
941 kprintf("swap_pager_strategy: bp %p offset %lld size %d, "
942 "not page bounded\n",
943 bp, (long long)bio->bio_offset, (int)bp->b_bcount);
948 * Clear error indication, initialize page index, count, data pointer.
951 bp->b_flags &= ~B_ERROR;
952 bp->b_resid = bp->b_bcount;
954 start = (vm_pindex_t)(bio->bio_offset >> PAGE_SHIFT);
955 count = howmany(bp->b_bcount, PAGE_SIZE);
959 * Deal with BUF_CMD_FREEBLKS
961 if (bp->b_cmd == BUF_CMD_FREEBLKS) {
963 * FREE PAGE(s) - destroy underlying swap that is no longer
966 vm_object_hold(object);
967 swp_pager_meta_free(object, start, count);
968 vm_object_drop(object);
975 * We need to be able to create a new cluster of I/O's. We cannot
976 * use the caller fields of the passed bio so push a new one.
978 * Because nbio is just a placeholder for the cluster links,
979 * we can biodone() the original bio instead of nbio to make
980 * things a bit more efficient.
982 nbio = push_bio(bio);
983 nbio->bio_offset = bio->bio_offset;
984 nbio->bio_caller_info1.cluster_head = NULL;
985 nbio->bio_caller_info2.cluster_tail = NULL;
991 * Execute read or write
993 vm_object_hold(object);
999 * Obtain block. If block not found and writing, allocate a
1000 * new block and build it into the object.
1002 blk = swp_pager_meta_ctl(object, start, 0);
1003 if ((blk == SWAPBLK_NONE) && bp->b_cmd != BUF_CMD_READ) {
1004 blk = swp_pager_getswapspace(object, 1);
1005 if (blk == SWAPBLK_NONE) {
1006 bp->b_error = ENOMEM;
1007 bp->b_flags |= B_ERROR;
1010 swp_pager_meta_build(object, start, blk);
1014 * Do we have to flush our current collection? Yes if:
1016 * - no swap block at this index
1017 * - swap block is not contiguous
1018 * - we cross a physical disk boundry in the
1022 biox && (biox_blkno + btoc(bufx->b_bcount) != blk ||
1023 ((biox_blkno ^ blk) & dmmax_mask)
1026 if (bp->b_cmd == BUF_CMD_READ) {
1027 ++mycpu->gd_cnt.v_swapin;
1028 mycpu->gd_cnt.v_swappgsin += btoc(bufx->b_bcount);
1030 ++mycpu->gd_cnt.v_swapout;
1031 mycpu->gd_cnt.v_swappgsout += btoc(bufx->b_bcount);
1032 bufx->b_dirtyend = bufx->b_bcount;
1036 * Finished with this buf.
1038 KKASSERT(bufx->b_bcount != 0);
1039 if (bufx->b_cmd != BUF_CMD_READ)
1040 bufx->b_dirtyend = bufx->b_bcount;
1046 * Add new swapblk to biox, instantiating biox if necessary.
1047 * Zero-fill reads are able to take a shortcut.
1049 if (blk == SWAPBLK_NONE) {
1051 * We can only get here if we are reading. Since
1052 * we are at splvm() we can safely modify b_resid,
1053 * even if chain ops are in progress.
1055 bzero(data, PAGE_SIZE);
1056 bp->b_resid -= PAGE_SIZE;
1059 /* XXX chain count > 4, wait to <= 4 */
1061 bufx = getpbuf(NULL);
1062 biox = &bufx->b_bio1;
1063 cluster_append(nbio, bufx);
1064 bufx->b_flags |= (bp->b_flags & B_ORDERED);
1065 bufx->b_cmd = bp->b_cmd;
1066 biox->bio_done = swap_chain_iodone;
1067 biox->bio_offset = (off_t)blk << PAGE_SHIFT;
1068 biox->bio_caller_info1.cluster_parent = nbio;
1071 bufx->b_data = data;
1073 bufx->b_bcount += PAGE_SIZE;
1080 vm_object_drop(object);
1083 * Flush out last buffer
1086 if (bufx->b_cmd == BUF_CMD_READ) {
1087 ++mycpu->gd_cnt.v_swapin;
1088 mycpu->gd_cnt.v_swappgsin += btoc(bufx->b_bcount);
1090 ++mycpu->gd_cnt.v_swapout;
1091 mycpu->gd_cnt.v_swappgsout += btoc(bufx->b_bcount);
1092 bufx->b_dirtyend = bufx->b_bcount;
1094 KKASSERT(bufx->b_bcount);
1095 if (bufx->b_cmd != BUF_CMD_READ)
1096 bufx->b_dirtyend = bufx->b_bcount;
1097 /* biox, bufx = NULL */
1101 * Now initiate all the I/O. Be careful looping on our chain as
1102 * I/O's may complete while we are still initiating them.
1104 * If the request is a 100% sparse read no bios will be present
1105 * and we just biodone() the buffer.
1107 nbio->bio_caller_info2.cluster_tail = NULL;
1108 bufx = nbio->bio_caller_info1.cluster_head;
1112 biox = &bufx->b_bio1;
1114 bufx = bufx->b_cluster_next;
1115 vn_strategy(swapdev_vp, biox);
1122 * Completion of the cluster will also call biodone_chain(nbio).
1123 * We never call biodone(nbio) so we don't have to worry about
1124 * setting up a bio_done callback. It's handled in the sub-IO.
1135 swap_chain_iodone(struct bio *biox)
1138 struct buf *bufx; /* chained sub-buffer */
1139 struct bio *nbio; /* parent nbio with chain glue */
1140 struct buf *bp; /* original bp associated with nbio */
1143 bufx = biox->bio_buf;
1144 nbio = biox->bio_caller_info1.cluster_parent;
1148 * Update the original buffer
1150 KKASSERT(bp != NULL);
1151 if (bufx->b_flags & B_ERROR) {
1152 atomic_set_int(&bufx->b_flags, B_ERROR);
1153 bp->b_error = bufx->b_error; /* race ok */
1154 } else if (bufx->b_resid != 0) {
1155 atomic_set_int(&bufx->b_flags, B_ERROR);
1156 bp->b_error = EINVAL; /* race ok */
1158 atomic_subtract_int(&bp->b_resid, bufx->b_bcount);
1162 * Remove us from the chain.
1164 spin_lock(&swapbp_spin);
1165 nextp = &nbio->bio_caller_info1.cluster_head;
1166 while (*nextp != bufx) {
1167 KKASSERT(*nextp != NULL);
1168 nextp = &(*nextp)->b_cluster_next;
1170 *nextp = bufx->b_cluster_next;
1171 chain_empty = (nbio->bio_caller_info1.cluster_head == NULL);
1172 spin_unlock(&swapbp_spin);
1175 * Clean up bufx. If the chain is now empty we finish out
1176 * the parent. Note that we may be racing other completions
1177 * so we must use the chain_empty status from above.
1180 if (bp->b_resid != 0 && !(bp->b_flags & B_ERROR)) {
1181 atomic_set_int(&bp->b_flags, B_ERROR);
1182 bp->b_error = EINVAL;
1184 biodone_chain(nbio);
1186 relpbuf(bufx, NULL);
1190 * SWAP_PAGER_GETPAGES() - bring page in from swap
1192 * The requested page may have to be brought in from swap. Calculate the
1193 * swap block and bring in additional pages if possible. All pages must
1194 * have contiguous swap block assignments and reside in the same object.
1196 * The caller has a single vm_object_pip_add() reference prior to
1197 * calling us and we should return with the same.
1199 * The caller has BUSY'd the page. We should return with (*mpp) left busy,
1200 * and any additinal pages unbusied.
1202 * If the caller encounters a PG_RAM page it will pass it to us even though
1203 * it may be valid and dirty. We cannot overwrite the page in this case!
1204 * The case is used to allow us to issue pure read-aheads.
1206 * NOTE! XXX This code does not entirely pipeline yet due to the fact that
1207 * the PG_RAM page is validated at the same time as mreq. What we
1208 * really need to do is issue a separate read-ahead pbuf.
1213 swap_pager_getpage(vm_object_t object, vm_page_t *mpp, int seqaccess)
1226 vm_page_t marray[XIO_INTERNAL_PAGES];
1230 vm_object_hold(object);
1231 if (mreq->object != object) {
1232 panic("swap_pager_getpages: object mismatch %p/%p",
1239 * We don't want to overwrite a fully valid page as it might be
1240 * dirty. This case can occur when e.g. vm_fault hits a perfectly
1241 * valid page with PG_RAM set.
1243 * In this case we see if the next page is a suitable page-in
1244 * candidate and if it is we issue read-ahead. PG_RAM will be
1245 * set on the last page of the read-ahead to continue the pipeline.
1247 if (mreq->valid == VM_PAGE_BITS_ALL) {
1248 if (swap_burst_read == 0 || mreq->pindex + 1 >= object->size) {
1249 vm_object_drop(object);
1250 return(VM_PAGER_OK);
1252 blk = swp_pager_meta_ctl(object, mreq->pindex + 1, 0);
1253 if (blk == SWAPBLK_NONE) {
1254 vm_object_drop(object);
1255 return(VM_PAGER_OK);
1257 m = vm_page_lookup_busy_try(object, mreq->pindex + 1,
1260 vm_object_drop(object);
1261 return(VM_PAGER_OK);
1262 } else if (m == NULL) {
1264 * Use VM_ALLOC_QUICK to avoid blocking on cache
1267 m = vm_page_alloc(object, mreq->pindex + 1,
1270 vm_object_drop(object);
1271 return(VM_PAGER_OK);
1276 vm_object_drop(object);
1277 return(VM_PAGER_OK);
1279 vm_page_unqueue_nowakeup(m);
1289 * Try to block-read contiguous pages from swap if sequential,
1290 * otherwise just read one page. Contiguous pages from swap must
1291 * reside within a single device stripe because the I/O cannot be
1292 * broken up across multiple stripes.
1294 * Note that blk and iblk can be SWAPBLK_NONE but the loop is
1295 * set up such that the case(s) are handled implicitly.
1297 blk = swp_pager_meta_ctl(mreq->object, mreq->pindex, 0);
1300 for (i = 1; swap_burst_read &&
1301 i < XIO_INTERNAL_PAGES &&
1302 mreq->pindex + i < object->size; ++i) {
1305 iblk = swp_pager_meta_ctl(object, mreq->pindex + i, 0);
1306 if (iblk != blk + i)
1308 if ((blk ^ iblk) & dmmax_mask)
1310 m = vm_page_lookup_busy_try(object, mreq->pindex + i,
1314 } else if (m == NULL) {
1316 * Use VM_ALLOC_QUICK to avoid blocking on cache
1319 m = vm_page_alloc(object, mreq->pindex + i,
1328 vm_page_unqueue_nowakeup(m);
1334 vm_page_flag_set(marray[i - 1], PG_RAM);
1337 * If mreq is the requested page and we have nothing to do return
1338 * VM_PAGER_FAIL. If raonly is set mreq is just another read-ahead
1339 * page and must be cleaned up.
1341 if (blk == SWAPBLK_NONE) {
1344 vnode_pager_freepage(mreq);
1345 vm_object_drop(object);
1346 return(VM_PAGER_OK);
1348 vm_object_drop(object);
1349 return(VM_PAGER_FAIL);
1354 * map our page(s) into kva for input
1356 bp = getpbuf_kva(&nsw_rcount);
1358 kva = (vm_offset_t) bp->b_kvabase;
1359 bcopy(marray, bp->b_xio.xio_pages, i * sizeof(vm_page_t));
1360 pmap_qenter(kva, bp->b_xio.xio_pages, i);
1362 bp->b_data = (caddr_t)kva;
1363 bp->b_bcount = PAGE_SIZE * i;
1364 bp->b_xio.xio_npages = i;
1365 bio->bio_done = swp_pager_async_iodone;
1366 bio->bio_offset = (off_t)blk << PAGE_SHIFT;
1367 bio->bio_caller_info1.index = SWBIO_READ;
1370 * Set index. If raonly set the index beyond the array so all
1371 * the pages are treated the same, otherwise the original mreq is
1375 bio->bio_driver_info = (void *)(intptr_t)i;
1377 bio->bio_driver_info = (void *)(intptr_t)0;
1379 for (j = 0; j < i; ++j)
1380 vm_page_flag_set(bp->b_xio.xio_pages[j], PG_SWAPINPROG);
1382 mycpu->gd_cnt.v_swapin++;
1383 mycpu->gd_cnt.v_swappgsin += bp->b_xio.xio_npages;
1386 * We still hold the lock on mreq, and our automatic completion routine
1387 * does not remove it.
1389 vm_object_pip_add(object, bp->b_xio.xio_npages);
1392 * perform the I/O. NOTE!!! bp cannot be considered valid after
1393 * this point because we automatically release it on completion.
1394 * Instead, we look at the one page we are interested in which we
1395 * still hold a lock on even through the I/O completion.
1397 * The other pages in our m[] array are also released on completion,
1398 * so we cannot assume they are valid anymore either.
1400 bp->b_cmd = BUF_CMD_READ;
1402 vn_strategy(swapdev_vp, bio);
1405 * Wait for the page we want to complete. PG_SWAPINPROG is always
1406 * cleared on completion. If an I/O error occurs, SWAPBLK_NONE
1407 * is set in the meta-data.
1409 * If this is a read-ahead only we return immediately without
1413 vm_object_drop(object);
1414 return(VM_PAGER_OK);
1418 * Read-ahead includes originally requested page case.
1421 flags = mreq->flags;
1423 if ((flags & PG_SWAPINPROG) == 0)
1425 tsleep_interlock(mreq, 0);
1426 if (!atomic_cmpset_int(&mreq->flags, flags,
1427 flags | PG_WANTED | PG_REFERENCED)) {
1430 mycpu->gd_cnt.v_intrans++;
1431 if (tsleep(mreq, PINTERLOCKED, "swread", hz*20)) {
1433 "swap_pager: indefinite wait buffer: "
1434 " offset: %lld, size: %ld\n",
1435 (long long)bio->bio_offset,
1442 * mreq is left bussied after completion, but all the other pages
1443 * are freed. If we had an unrecoverable read error the page will
1446 vm_object_drop(object);
1447 if (mreq->valid != VM_PAGE_BITS_ALL)
1448 return(VM_PAGER_ERROR);
1450 return(VM_PAGER_OK);
1453 * A final note: in a low swap situation, we cannot deallocate swap
1454 * and mark a page dirty here because the caller is likely to mark
1455 * the page clean when we return, causing the page to possibly revert
1456 * to all-zero's later.
1461 * swap_pager_putpages:
1463 * Assign swap (if necessary) and initiate I/O on the specified pages.
1465 * We support both OBJT_DEFAULT and OBJT_SWAP objects. DEFAULT objects
1466 * are automatically converted to SWAP objects.
1468 * In a low memory situation we may block in vn_strategy(), but the new
1469 * vm_page reservation system coupled with properly written VFS devices
1470 * should ensure that no low-memory deadlock occurs. This is an area
1473 * The parent has N vm_object_pip_add() references prior to
1474 * calling us and will remove references for rtvals[] that are
1475 * not set to VM_PAGER_PEND. We need to remove the rest on I/O
1478 * The parent has soft-busy'd the pages it passes us and will unbusy
1479 * those whos rtvals[] entry is not set to VM_PAGER_PEND on return.
1480 * We need to unbusy the rest on I/O completion.
1485 swap_pager_putpages(vm_object_t object, vm_page_t *m, int count,
1486 int sync, int *rtvals)
1491 vm_object_hold(object);
1493 if (count && m[0]->object != object) {
1494 panic("swap_pager_getpages: object mismatch %p/%p",
1503 * Turn object into OBJT_SWAP
1504 * check for bogus sysops
1505 * force sync if not pageout process
1507 if (object->type == OBJT_DEFAULT) {
1508 if (object->type == OBJT_DEFAULT)
1509 swp_pager_meta_convert(object);
1512 if (curthread != pagethread)
1518 * Update nsw parameters from swap_async_max sysctl values.
1519 * Do not let the sysop crash the machine with bogus numbers.
1521 if (swap_async_max != nsw_wcount_async_max) {
1527 if ((n = swap_async_max) > nswbuf_kva / 2)
1534 * Adjust difference ( if possible ). If the current async
1535 * count is too low, we may not be able to make the adjustment
1538 * vm_token needed for nsw_wcount sleep interlock
1540 lwkt_gettoken(&vm_token);
1541 n -= nsw_wcount_async_max;
1542 if (nsw_wcount_async + n >= 0) {
1543 nsw_wcount_async_max += n;
1544 pbuf_adjcount(&nsw_wcount_async, n);
1546 lwkt_reltoken(&vm_token);
1552 * Assign swap blocks and issue I/O. We reallocate swap on the fly.
1553 * The page is left dirty until the pageout operation completes
1557 for (i = 0; i < count; i += n) {
1564 * Maximum I/O size is limited by a number of factors.
1567 n = min(BLIST_MAX_ALLOC, count - i);
1568 n = min(n, nsw_cluster_max);
1570 lwkt_gettoken(&vm_token);
1573 * Get biggest block of swap we can. If we fail, fall
1574 * back and try to allocate a smaller block. Don't go
1575 * overboard trying to allocate space if it would overly
1579 (blk = swp_pager_getswapspace(object, n)) == SWAPBLK_NONE &&
1584 if (blk == SWAPBLK_NONE) {
1585 for (j = 0; j < n; ++j)
1586 rtvals[i+j] = VM_PAGER_FAIL;
1587 lwkt_reltoken(&vm_token);
1590 if (vm_report_swap_allocs > 0) {
1591 kprintf("swap_alloc %08jx,%d\n", (intmax_t)blk, n);
1592 --vm_report_swap_allocs;
1596 * The I/O we are constructing cannot cross a physical
1597 * disk boundry in the swap stripe. Note: we are still
1600 if ((blk ^ (blk + n)) & dmmax_mask) {
1601 j = ((blk + dmmax) & dmmax_mask) - blk;
1602 swp_pager_freeswapspace(object, blk + j, n - j);
1607 * All I/O parameters have been satisfied, build the I/O
1608 * request and assign the swap space.
1611 bp = getpbuf_kva(&nsw_wcount_sync);
1613 bp = getpbuf_kva(&nsw_wcount_async);
1616 lwkt_reltoken(&vm_token);
1618 pmap_qenter((vm_offset_t)bp->b_data, &m[i], n);
1620 bp->b_bcount = PAGE_SIZE * n;
1621 bio->bio_offset = (off_t)blk << PAGE_SHIFT;
1623 for (j = 0; j < n; ++j) {
1624 vm_page_t mreq = m[i+j];
1626 swp_pager_meta_build(mreq->object, mreq->pindex,
1628 if (object->type == OBJT_SWAP)
1629 vm_page_dirty(mreq);
1630 rtvals[i+j] = VM_PAGER_OK;
1632 vm_page_flag_set(mreq, PG_SWAPINPROG);
1633 bp->b_xio.xio_pages[j] = mreq;
1635 bp->b_xio.xio_npages = n;
1637 mycpu->gd_cnt.v_swapout++;
1638 mycpu->gd_cnt.v_swappgsout += bp->b_xio.xio_npages;
1640 bp->b_dirtyoff = 0; /* req'd for NFS */
1641 bp->b_dirtyend = bp->b_bcount; /* req'd for NFS */
1642 bp->b_cmd = BUF_CMD_WRITE;
1643 bio->bio_caller_info1.index = SWBIO_WRITE;
1648 if (sync == FALSE) {
1649 bio->bio_done = swp_pager_async_iodone;
1651 vn_strategy(swapdev_vp, bio);
1653 for (j = 0; j < n; ++j)
1654 rtvals[i+j] = VM_PAGER_PEND;
1659 * Issue synchrnously.
1661 * Wait for the sync I/O to complete, then update rtvals.
1662 * We just set the rtvals[] to VM_PAGER_PEND so we can call
1663 * our async completion routine at the end, thus avoiding a
1666 bio->bio_caller_info1.index |= SWBIO_SYNC;
1667 bio->bio_done = biodone_sync;
1668 bio->bio_flags |= BIO_SYNC;
1669 vn_strategy(swapdev_vp, bio);
1670 biowait(bio, "swwrt");
1672 for (j = 0; j < n; ++j)
1673 rtvals[i+j] = VM_PAGER_PEND;
1676 * Now that we are through with the bp, we can call the
1677 * normal async completion, which frees everything up.
1679 swp_pager_async_iodone(bio);
1681 vm_object_drop(object);
1687 * Recalculate the low and high-water marks.
1690 swap_pager_newswap(void)
1693 nswap_lowat = vm_swap_max * 4 / 100; /* 4% left */
1694 nswap_hiwat = vm_swap_max * 6 / 100; /* 6% left */
1695 kprintf("swap low/high-water marks set to %d/%d\n",
1696 nswap_lowat, nswap_hiwat);
1705 * swp_pager_async_iodone:
1707 * Completion routine for asynchronous reads and writes from/to swap.
1708 * Also called manually by synchronous code to finish up a bp.
1710 * For READ operations, the pages are PG_BUSY'd. For WRITE operations,
1711 * the pages are vm_page_t->busy'd. For READ operations, we PG_BUSY
1712 * unbusy all pages except the 'main' request page. For WRITE
1713 * operations, we vm_page_t->busy'd unbusy all pages ( we can do this
1714 * because we marked them all VM_PAGER_PEND on return from putpages ).
1716 * This routine may not block.
1721 swp_pager_async_iodone(struct bio *bio)
1723 struct buf *bp = bio->bio_buf;
1724 vm_object_t object = NULL;
1731 if (bp->b_flags & B_ERROR) {
1733 "swap_pager: I/O error - %s failed; offset %lld,"
1734 "size %ld, error %d\n",
1735 ((bio->bio_caller_info1.index & SWBIO_READ) ?
1736 "pagein" : "pageout"),
1737 (long long)bio->bio_offset,
1744 * set object, raise to splvm().
1746 if (bp->b_xio.xio_npages)
1747 object = bp->b_xio.xio_pages[0]->object;
1750 * remove the mapping for kernel virtual
1752 pmap_qremove((vm_offset_t)bp->b_data, bp->b_xio.xio_npages);
1755 * cleanup pages. If an error occurs writing to swap, we are in
1756 * very serious trouble. If it happens to be a disk error, though,
1757 * we may be able to recover by reassigning the swap later on. So
1758 * in this case we remove the m->swapblk assignment for the page
1759 * but do not free it in the rlist. The errornous block(s) are thus
1760 * never reallocated as swap. Redirty the page and continue.
1762 for (i = 0; i < bp->b_xio.xio_npages; ++i) {
1763 vm_page_t m = bp->b_xio.xio_pages[i];
1765 if (bp->b_flags & B_ERROR) {
1767 * If an error occurs I'd love to throw the swapblk
1768 * away without freeing it back to swapspace, so it
1769 * can never be used again. But I can't from an
1773 if (bio->bio_caller_info1.index & SWBIO_READ) {
1775 * When reading, reqpage needs to stay
1776 * locked for the parent, but all other
1777 * pages can be freed. We still want to
1778 * wakeup the parent waiting on the page,
1779 * though. ( also: pg_reqpage can be -1 and
1780 * not match anything ).
1782 * We have to wake specifically requested pages
1783 * up too because we cleared PG_SWAPINPROG and
1784 * someone may be waiting for that.
1786 * NOTE: for reads, m->dirty will probably
1787 * be overridden by the original caller of
1788 * getpages so don't play cute tricks here.
1790 * NOTE: We can't actually free the page from
1791 * here, because this is an interrupt. It
1792 * is not legal to mess with object->memq
1793 * from an interrupt. Deactivate the page
1798 vm_page_flag_clear(m, PG_ZERO);
1799 vm_page_flag_clear(m, PG_SWAPINPROG);
1802 * bio_driver_info holds the requested page
1805 if (i != (int)(intptr_t)bio->bio_driver_info) {
1806 vm_page_deactivate(m);
1812 * If i == bp->b_pager.pg_reqpage, do not wake
1813 * the page up. The caller needs to.
1817 * If a write error occurs remove the swap
1818 * assignment (note that PG_SWAPPED may or
1819 * may not be set depending on prior activity).
1821 * Re-dirty OBJT_SWAP pages as there is no
1822 * other backing store, we can't throw the
1825 * Non-OBJT_SWAP pages (aka swapcache) must
1826 * not be dirtied since they may not have
1827 * been dirty in the first place, and they
1828 * do have backing store (the vnode).
1830 vm_page_busy_wait(m, FALSE, "swadpg");
1831 swp_pager_meta_ctl(m->object, m->pindex,
1833 vm_page_flag_clear(m, PG_SWAPPED);
1834 if (m->object->type == OBJT_SWAP) {
1836 vm_page_activate(m);
1838 vm_page_flag_clear(m, PG_SWAPINPROG);
1839 vm_page_io_finish(m);
1842 } else if (bio->bio_caller_info1.index & SWBIO_READ) {
1844 * NOTE: for reads, m->dirty will probably be
1845 * overridden by the original caller of getpages so
1846 * we cannot set them in order to free the underlying
1847 * swap in a low-swap situation. I don't think we'd
1848 * want to do that anyway, but it was an optimization
1849 * that existed in the old swapper for a time before
1850 * it got ripped out due to precisely this problem.
1852 * clear PG_ZERO in page.
1854 * If not the requested page then deactivate it.
1856 * Note that the requested page, reqpage, is left
1857 * busied, but we still have to wake it up. The
1858 * other pages are released (unbusied) by
1859 * vm_page_wakeup(). We do not set reqpage's
1860 * valid bits here, it is up to the caller.
1864 * NOTE: can't call pmap_clear_modify(m) from an
1865 * interrupt thread, the pmap code may have to map
1866 * non-kernel pmaps and currently asserts the case.
1868 /*pmap_clear_modify(m);*/
1869 m->valid = VM_PAGE_BITS_ALL;
1871 vm_page_flag_clear(m, PG_ZERO | PG_SWAPINPROG);
1872 vm_page_flag_set(m, PG_SWAPPED);
1875 * We have to wake specifically requested pages
1876 * up too because we cleared PG_SWAPINPROG and
1877 * could be waiting for it in getpages. However,
1878 * be sure to not unbusy getpages specifically
1879 * requested page - getpages expects it to be
1882 * bio_driver_info holds the requested page
1884 if (i != (int)(intptr_t)bio->bio_driver_info) {
1885 vm_page_deactivate(m);
1892 * Mark the page clean but do not mess with the
1893 * pmap-layer's modified state. That state should
1894 * also be clear since the caller protected the
1895 * page VM_PROT_READ, but allow the case.
1897 * We are in an interrupt, avoid pmap operations.
1899 * If we have a severe page deficit, deactivate the
1900 * page. Do not try to cache it (which would also
1901 * involve a pmap op), because the page might still
1904 * When using the swap to cache clean vnode pages
1905 * we do not mess with the page dirty bits.
1907 vm_page_busy_wait(m, FALSE, "swadpg");
1908 if (m->object->type == OBJT_SWAP)
1910 vm_page_flag_clear(m, PG_SWAPINPROG);
1911 vm_page_flag_set(m, PG_SWAPPED);
1912 if (vm_page_count_severe())
1913 vm_page_deactivate(m);
1915 if (!vm_page_count_severe() || !vm_page_try_to_cache(m))
1916 vm_page_protect(m, VM_PROT_READ);
1918 vm_page_io_finish(m);
1924 * adjust pip. NOTE: the original parent may still have its own
1925 * pip refs on the object.
1929 vm_object_pip_wakeup_n(object, bp->b_xio.xio_npages);
1932 * Release the physical I/O buffer.
1934 * NOTE: Due to synchronous operations in the write case b_cmd may
1935 * already be set to BUF_CMD_DONE and BIO_SYNC may have already
1938 * Use vm_token to interlock nsw_rcount/wcount wakeup?
1940 lwkt_gettoken(&vm_token);
1941 if (bio->bio_caller_info1.index & SWBIO_READ)
1942 nswptr = &nsw_rcount;
1943 else if (bio->bio_caller_info1.index & SWBIO_SYNC)
1944 nswptr = &nsw_wcount_sync;
1946 nswptr = &nsw_wcount_async;
1947 bp->b_cmd = BUF_CMD_DONE;
1948 relpbuf(bp, nswptr);
1949 lwkt_reltoken(&vm_token);
1953 * Fault-in a potentially swapped page and remove the swap reference.
1954 * (used by swapoff code)
1956 * object must be held.
1958 static __inline void
1959 swp_pager_fault_page(vm_object_t object, int *sharedp, vm_pindex_t pindex)
1965 ASSERT_LWKT_TOKEN_HELD(vm_object_token(object));
1967 if (object->type == OBJT_VNODE) {
1969 * Any swap related to a vnode is due to swapcache. We must
1970 * vget() the vnode in case it is not active (otherwise
1971 * vref() will panic). Calling vm_object_page_remove() will
1972 * ensure that any swap ref is removed interlocked with the
1973 * page. clean_only is set to TRUE so we don't throw away
1976 vp = object->handle;
1977 error = vget(vp, LK_SHARED | LK_RETRY | LK_CANRECURSE);
1979 vm_object_page_remove(object, pindex, pindex + 1, TRUE);
1984 * Otherwise it is a normal OBJT_SWAP object and we can
1985 * fault the page in and remove the swap.
1987 m = vm_fault_object_page(object, IDX_TO_OFF(pindex),
1989 VM_FAULT_DIRTY | VM_FAULT_UNSWAP,
1997 * This removes all swap blocks related to a particular device. We have
1998 * to be careful of ripups during the scan.
2000 static int swp_pager_swapoff_callback(struct swblock *swap, void *data);
2003 swap_pager_swapoff(int devidx)
2005 struct swswapoffinfo info;
2006 struct vm_object marker;
2010 bzero(&marker, sizeof(marker));
2011 marker.type = OBJT_MARKER;
2013 for (n = 0; n < VMOBJ_HSIZE; ++n) {
2014 lwkt_gettoken(&vmobj_tokens[n]);
2015 TAILQ_INSERT_HEAD(&vm_object_lists[n], &marker, object_list);
2017 while ((object = TAILQ_NEXT(&marker, object_list)) != NULL) {
2018 if (object->type == OBJT_MARKER)
2020 if (object->type != OBJT_SWAP &&
2021 object->type != OBJT_VNODE)
2023 vm_object_hold(object);
2024 if (object->type != OBJT_SWAP &&
2025 object->type != OBJT_VNODE) {
2026 vm_object_drop(object);
2029 info.object = object;
2031 info.devidx = devidx;
2032 swblock_rb_tree_RB_SCAN(&object->swblock_root,
2033 NULL, swp_pager_swapoff_callback,
2035 vm_object_drop(object);
2037 if (object == TAILQ_NEXT(&marker, object_list)) {
2038 TAILQ_REMOVE(&vm_object_lists[n],
2039 &marker, object_list);
2040 TAILQ_INSERT_AFTER(&vm_object_lists[n], object,
2041 &marker, object_list);
2044 TAILQ_REMOVE(&vm_object_lists[n], &marker, object_list);
2045 lwkt_reltoken(&vmobj_tokens[n]);
2049 * If we fail to locate all swblocks we just fail gracefully and
2050 * do not bother to restore paging on the swap device. If the
2051 * user wants to retry the user can retry.
2053 if (swdevt[devidx].sw_nused)
2061 swp_pager_swapoff_callback(struct swblock *swap, void *data)
2063 struct swswapoffinfo *info = data;
2064 vm_object_t object = info->object;
2069 index = swap->swb_index;
2070 for (i = 0; i < SWAP_META_PAGES; ++i) {
2072 * Make sure we don't race a dying object. This will
2073 * kill the scan of the object's swap blocks entirely.
2075 if (object->flags & OBJ_DEAD)
2079 * Fault the page, which can obviously block. If the swap
2080 * structure disappears break out.
2082 v = swap->swb_pages[i];
2083 if (v != SWAPBLK_NONE && BLK2DEVIDX(v) == info->devidx) {
2084 swp_pager_fault_page(object, &info->shared,
2085 swap->swb_index + i);
2086 /* swap ptr might go away */
2087 if (RB_LOOKUP(swblock_rb_tree,
2088 &object->swblock_root, index) != swap) {
2096 /************************************************************************
2098 ************************************************************************
2100 * These routines manipulate the swap metadata stored in the
2101 * OBJT_SWAP object. All swp_*() routines must be called at
2102 * splvm() because swap can be freed up by the low level vm_page
2103 * code which might be called from interrupts beyond what splbio() covers.
2105 * Swap metadata is implemented with a global hash and not directly
2106 * linked into the object. Instead the object simply contains
2107 * appropriate tracking counters.
2111 * Lookup the swblock containing the specified swap block index.
2113 * The caller must hold the object.
2117 swp_pager_lookup(vm_object_t object, vm_pindex_t index)
2119 ASSERT_LWKT_TOKEN_HELD(vm_object_token(object));
2120 index &= ~(vm_pindex_t)SWAP_META_MASK;
2121 return (RB_LOOKUP(swblock_rb_tree, &object->swblock_root, index));
2125 * Remove a swblock from the RB tree.
2127 * The caller must hold the object.
2131 swp_pager_remove(vm_object_t object, struct swblock *swap)
2133 ASSERT_LWKT_TOKEN_HELD(vm_object_token(object));
2134 RB_REMOVE(swblock_rb_tree, &object->swblock_root, swap);
2138 * Convert default object to swap object if necessary
2140 * The caller must hold the object.
2143 swp_pager_meta_convert(vm_object_t object)
2145 if (object->type == OBJT_DEFAULT) {
2146 object->type = OBJT_SWAP;
2147 KKASSERT(object->swblock_count == 0);
2152 * SWP_PAGER_META_BUILD() - add swap block to swap meta data for object
2154 * We first convert the object to a swap object if it is a default
2155 * object. Vnode objects do not need to be converted.
2157 * The specified swapblk is added to the object's swap metadata. If
2158 * the swapblk is not valid, it is freed instead. Any previously
2159 * assigned swapblk is freed.
2161 * The caller must hold the object.
2164 swp_pager_meta_build(vm_object_t object, vm_pindex_t index, swblk_t swapblk)
2166 struct swblock *swap;
2167 struct swblock *oswap;
2170 KKASSERT(swapblk != SWAPBLK_NONE);
2171 ASSERT_LWKT_TOKEN_HELD(vm_object_token(object));
2174 * Convert object if necessary
2176 if (object->type == OBJT_DEFAULT)
2177 swp_pager_meta_convert(object);
2180 * Locate swblock. If not found create, but if we aren't adding
2181 * anything just return. If we run out of space in the map we wait
2182 * and, since the hash table may have changed, retry.
2185 swap = swp_pager_lookup(object, index);
2190 swap = zalloc(swap_zone);
2195 swap->swb_index = index & ~(vm_pindex_t)SWAP_META_MASK;
2196 swap->swb_count = 0;
2198 ++object->swblock_count;
2200 for (i = 0; i < SWAP_META_PAGES; ++i)
2201 swap->swb_pages[i] = SWAPBLK_NONE;
2202 oswap = RB_INSERT(swblock_rb_tree, &object->swblock_root, swap);
2203 KKASSERT(oswap == NULL);
2207 * Delete prior contents of metadata.
2209 * NOTE: Decrement swb_count after the freeing operation (which
2210 * might block) to prevent racing destruction of the swblock.
2212 index &= SWAP_META_MASK;
2214 while ((v = swap->swb_pages[index]) != SWAPBLK_NONE) {
2215 swap->swb_pages[index] = SWAPBLK_NONE;
2217 swp_pager_freeswapspace(object, v, 1);
2219 --mycpu->gd_vmtotal.t_vm;
2223 * Enter block into metadata
2225 swap->swb_pages[index] = swapblk;
2226 if (swapblk != SWAPBLK_NONE) {
2228 ++mycpu->gd_vmtotal.t_vm;
2233 * SWP_PAGER_META_FREE() - free a range of blocks in the object's swap metadata
2235 * The requested range of blocks is freed, with any associated swap
2236 * returned to the swap bitmap.
2238 * This routine will free swap metadata structures as they are cleaned
2239 * out. This routine does *NOT* operate on swap metadata associated
2240 * with resident pages.
2242 * The caller must hold the object.
2244 static int swp_pager_meta_free_callback(struct swblock *swb, void *data);
2247 swp_pager_meta_free(vm_object_t object, vm_pindex_t index, vm_pindex_t count)
2249 struct swfreeinfo info;
2251 ASSERT_LWKT_TOKEN_HELD(vm_object_token(object));
2256 if (object->swblock_count == 0) {
2257 KKASSERT(RB_EMPTY(&object->swblock_root));
2264 * Setup for RB tree scan. Note that the pindex range can be huge
2265 * due to the 64 bit page index space so we cannot safely iterate.
2267 info.object = object;
2268 info.basei = index & ~(vm_pindex_t)SWAP_META_MASK;
2270 info.endi = index + count - 1;
2271 swblock_rb_tree_RB_SCAN(&object->swblock_root, rb_swblock_scancmp,
2272 swp_pager_meta_free_callback, &info);
2276 * The caller must hold the object.
2280 swp_pager_meta_free_callback(struct swblock *swap, void *data)
2282 struct swfreeinfo *info = data;
2283 vm_object_t object = info->object;
2288 * Figure out the range within the swblock. The wider scan may
2289 * return edge-case swap blocks when the start and/or end points
2290 * are in the middle of a block.
2292 if (swap->swb_index < info->begi)
2293 index = (int)info->begi & SWAP_META_MASK;
2297 if (swap->swb_index + SWAP_META_PAGES > info->endi)
2298 eindex = (int)info->endi & SWAP_META_MASK;
2300 eindex = SWAP_META_MASK;
2303 * Scan and free the blocks. The loop terminates early
2304 * if (swap) runs out of blocks and could be freed.
2306 * NOTE: Decrement swb_count after swp_pager_freeswapspace()
2307 * to deal with a zfree race.
2309 while (index <= eindex) {
2310 swblk_t v = swap->swb_pages[index];
2312 if (v != SWAPBLK_NONE) {
2313 swap->swb_pages[index] = SWAPBLK_NONE;
2315 swp_pager_freeswapspace(object, v, 1);
2316 --mycpu->gd_vmtotal.t_vm;
2317 if (--swap->swb_count == 0) {
2318 swp_pager_remove(object, swap);
2319 zfree(swap_zone, swap);
2320 --object->swblock_count;
2327 /* swap may be invalid here due to zfree above */
2334 * SWP_PAGER_META_FREE_ALL() - destroy all swap metadata associated with object
2336 * This routine locates and destroys all swap metadata associated with
2339 * NOTE: Decrement swb_count after the freeing operation (which
2340 * might block) to prevent racing destruction of the swblock.
2342 * The caller must hold the object.
2345 swp_pager_meta_free_all(vm_object_t object)
2347 struct swblock *swap;
2350 ASSERT_LWKT_TOKEN_HELD(vm_object_token(object));
2352 while ((swap = RB_ROOT(&object->swblock_root)) != NULL) {
2353 swp_pager_remove(object, swap);
2354 for (i = 0; i < SWAP_META_PAGES; ++i) {
2355 swblk_t v = swap->swb_pages[i];
2356 if (v != SWAPBLK_NONE) {
2358 swp_pager_freeswapspace(object, v, 1);
2360 --mycpu->gd_vmtotal.t_vm;
2363 if (swap->swb_count != 0)
2364 panic("swap_pager_meta_free_all: swb_count != 0");
2365 zfree(swap_zone, swap);
2366 --object->swblock_count;
2369 KKASSERT(object->swblock_count == 0);
2373 * SWP_PAGER_METACTL() - misc control of swap and vm_page_t meta data.
2375 * This routine is capable of looking up, popping, or freeing
2376 * swapblk assignments in the swap meta data or in the vm_page_t.
2377 * The routine typically returns the swapblk being looked-up, or popped,
2378 * or SWAPBLK_NONE if the block was freed, or SWAPBLK_NONE if the block
2379 * was invalid. This routine will automatically free any invalid
2380 * meta-data swapblks.
2382 * It is not possible to store invalid swapblks in the swap meta data
2383 * (other then a literal 'SWAPBLK_NONE'), so we don't bother checking.
2385 * When acting on a busy resident page and paging is in progress, we
2386 * have to wait until paging is complete but otherwise can act on the
2389 * SWM_FREE remove and free swap block from metadata
2390 * SWM_POP remove from meta data but do not free.. pop it out
2392 * The caller must hold the object.
2395 swp_pager_meta_ctl(vm_object_t object, vm_pindex_t index, int flags)
2397 struct swblock *swap;
2400 if (object->swblock_count == 0)
2401 return(SWAPBLK_NONE);
2404 swap = swp_pager_lookup(object, index);
2407 index &= SWAP_META_MASK;
2408 r1 = swap->swb_pages[index];
2410 if (r1 != SWAPBLK_NONE) {
2411 if (flags & (SWM_FREE|SWM_POP)) {
2412 swap->swb_pages[index] = SWAPBLK_NONE;
2413 --mycpu->gd_vmtotal.t_vm;
2414 if (--swap->swb_count == 0) {
2415 swp_pager_remove(object, swap);
2416 zfree(swap_zone, swap);
2417 --object->swblock_count;
2420 /* swap ptr may be invalid */
2421 if (flags & SWM_FREE) {
2422 swp_pager_freeswapspace(object, r1, 1);
2426 /* swap ptr may be invalid */