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. All advertising materials mentioning features or use of this software
54 * must display the following acknowledgement:
55 * This product includes software developed by the University of
56 * California, Berkeley and its contributors.
57 * 4. Neither the name of the University nor the names of its contributors
58 * may be used to endorse or promote products derived from this software
59 * without specific prior written permission.
61 * THIS SOFTWARE IS PROVIDED BY THE REGENTS AND CONTRIBUTORS ``AS IS'' AND
62 * ANY EXPRESS OR IMPLIED WARRANTIES, INCLUDING, BUT NOT LIMITED TO, THE
63 * IMPLIED WARRANTIES OF MERCHANTABILITY AND FITNESS FOR A PARTICULAR PURPOSE
64 * ARE DISCLAIMED. IN NO EVENT SHALL THE REGENTS OR CONTRIBUTORS BE LIABLE
65 * FOR ANY DIRECT, INDIRECT, INCIDENTAL, SPECIAL, EXEMPLARY, OR CONSEQUENTIAL
66 * DAMAGES (INCLUDING, BUT NOT LIMITED TO, PROCUREMENT OF SUBSTITUTE GOODS
67 * OR SERVICES; LOSS OF USE, DATA, OR PROFITS; OR BUSINESS INTERRUPTION)
68 * HOWEVER CAUSED AND ON ANY THEORY OF LIABILITY, WHETHER IN CONTRACT, STRICT
69 * LIABILITY, OR TORT (INCLUDING NEGLIGENCE OR OTHERWISE) ARISING IN ANY WAY
70 * OUT OF THE USE OF THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF
76 * Radix Bitmap 'blists'.
78 * - The new swapper uses the new radix bitmap code. This should scale
79 * to arbitrarily small or arbitrarily large swap spaces and an almost
80 * arbitrary degree of fragmentation.
84 * - on the fly reallocation of swap during putpages. The new system
85 * does not try to keep previously allocated swap blocks for dirty
88 * - on the fly deallocation of swap
90 * - No more garbage collection required. Unnecessarily allocated swap
91 * blocks only exist for dirty vm_page_t's now and these are already
92 * cycled (in a high-load system) by the pager. We also do on-the-fly
93 * removal of invalidated swap blocks when a page is destroyed
96 * from: Utah $Hdr: swap_pager.c 1.4 91/04/30$
97 * @(#)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 $
101 #include <sys/param.h>
102 #include <sys/systm.h>
103 #include <sys/conf.h>
104 #include <sys/kernel.h>
105 #include <sys/proc.h>
107 #include <sys/vnode.h>
108 #include <sys/malloc.h>
109 #include <sys/vmmeter.h>
110 #include <sys/sysctl.h>
111 #include <sys/blist.h>
112 #include <sys/lock.h>
113 #include <sys/thread2.h>
115 #ifndef MAX_PAGEOUT_CLUSTER
116 #define MAX_PAGEOUT_CLUSTER 16
119 #define SWB_NPAGES MAX_PAGEOUT_CLUSTER
121 #include "opt_swap.h"
123 #include <vm/vm_object.h>
124 #include <vm/vm_page.h>
125 #include <vm/vm_pager.h>
126 #include <vm/vm_pageout.h>
127 #include <vm/swap_pager.h>
128 #include <vm/vm_extern.h>
129 #include <vm/vm_zone.h>
130 #include <vm/vnode_pager.h>
132 #include <sys/buf2.h>
133 #include <vm/vm_page2.h>
135 #define SWM_FREE 0x02 /* free, period */
136 #define SWM_POP 0x04 /* pop out */
138 #define SWBIO_READ 0x01
139 #define SWBIO_WRITE 0x02
140 #define SWBIO_SYNC 0x04
146 vm_pindex_t endi; /* inclusive */
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 vm_swap_cache_use;
156 int vm_swap_anon_use;
158 static int swap_pager_almost_full; /* swap space exhaustion (w/ hysteresis)*/
159 static int nsw_rcount; /* free read buffers */
160 static int nsw_wcount_sync; /* limit write buffers / synchronous */
161 static int nsw_wcount_async; /* limit write buffers / asynchronous */
162 static int nsw_wcount_async_max;/* assigned maximum */
163 static int nsw_cluster_max; /* maximum VOP I/O allowed */
165 struct blist *swapblist;
166 static int swap_async_max = 4; /* maximum in-progress async I/O's */
167 static int swap_burst_read = 0; /* allow burst reading */
170 extern struct vnode *swapdev_vp;
171 extern struct swdevt *swdevt;
174 #define BLK2DEVIDX(blk) (nswdev > 1 ? blk / dmmax % nswdev : 0)
176 SYSCTL_INT(_vm, OID_AUTO, swap_async_max,
177 CTLFLAG_RW, &swap_async_max, 0, "Maximum running async swap ops");
178 SYSCTL_INT(_vm, OID_AUTO, swap_burst_read,
179 CTLFLAG_RW, &swap_burst_read, 0, "Allow burst reads for pageins");
181 SYSCTL_INT(_vm, OID_AUTO, swap_cache_use,
182 CTLFLAG_RD, &vm_swap_cache_use, 0, "");
183 SYSCTL_INT(_vm, OID_AUTO, swap_anon_use,
184 CTLFLAG_RD, &vm_swap_anon_use, 0, "");
185 SYSCTL_INT(_vm, OID_AUTO, swap_size,
186 CTLFLAG_RD, &vm_swap_size, 0, "");
191 * Red-Black tree for swblock entries
193 * The caller must hold vm_token
195 RB_GENERATE2(swblock_rb_tree, swblock, swb_entry, rb_swblock_compare,
196 vm_pindex_t, swb_index);
199 rb_swblock_compare(struct swblock *swb1, struct swblock *swb2)
201 if (swb1->swb_index < swb2->swb_index)
203 if (swb1->swb_index > swb2->swb_index)
210 rb_swblock_scancmp(struct swblock *swb, void *data)
212 struct swfreeinfo *info = data;
214 if (swb->swb_index < info->basei)
216 if (swb->swb_index > info->endi)
223 rb_swblock_condcmp(struct swblock *swb, void *data)
225 struct swfreeinfo *info = data;
227 if (swb->swb_index < info->basei)
233 * pagerops for OBJT_SWAP - "swap pager". Some ops are also global procedure
234 * calls hooked from other parts of the VM system and do not appear here.
235 * (see vm/swap_pager.h).
238 static void swap_pager_dealloc (vm_object_t object);
239 static int swap_pager_getpage (vm_object_t, vm_page_t *, int);
240 static void swap_chain_iodone(struct bio *biox);
242 struct pagerops swappagerops = {
243 swap_pager_dealloc, /* deallocate an OBJT_SWAP object */
244 swap_pager_getpage, /* pagein */
245 swap_pager_putpages, /* pageout */
246 swap_pager_haspage /* get backing store status for page */
250 * dmmax is in page-sized chunks with the new swap system. It was
251 * dev-bsized chunks in the old. dmmax is always a power of 2.
253 * swap_*() routines are externally accessible. swp_*() routines are
258 static int dmmax_mask;
259 int nswap_lowat = 128; /* in pages, swap_pager_almost_full warn */
260 int nswap_hiwat = 512; /* in pages, swap_pager_almost_full warn */
262 static __inline void swp_sizecheck (void);
263 static void swp_pager_async_iodone (struct bio *bio);
266 * Swap bitmap functions
269 static __inline void swp_pager_freeswapspace(vm_object_t object,
270 swblk_t blk, int npages);
271 static __inline swblk_t swp_pager_getswapspace(vm_object_t object, int npages);
277 static void swp_pager_meta_convert(vm_object_t);
278 static void swp_pager_meta_build(vm_object_t, vm_pindex_t, swblk_t);
279 static void swp_pager_meta_free(vm_object_t, vm_pindex_t, vm_pindex_t);
280 static void swp_pager_meta_free_all(vm_object_t);
281 static swblk_t swp_pager_meta_ctl(vm_object_t, vm_pindex_t, int);
284 * SWP_SIZECHECK() - update swap_pager_full indication
286 * update the swap_pager_almost_full indication and warn when we are
287 * about to run out of swap space, using lowat/hiwat hysteresis.
289 * Clear swap_pager_full ( task killing ) indication when lowat is met.
291 * No restrictions on call
292 * This routine may not block.
298 if (vm_swap_size < nswap_lowat) {
299 if (swap_pager_almost_full == 0) {
300 kprintf("swap_pager: out of swap space\n");
301 swap_pager_almost_full = 1;
305 if (vm_swap_size > nswap_hiwat)
306 swap_pager_almost_full = 0;
311 * SWAP_PAGER_INIT() - initialize the swap pager!
313 * Expected to be started from system init. NOTE: This code is run
314 * before much else so be careful what you depend on. Most of the VM
315 * system has yet to be initialized at this point.
317 * Called from the low level boot code only.
320 swap_pager_init(void *arg __unused)
323 * Device Stripe, in PAGE_SIZE'd blocks
325 dmmax = SWB_NPAGES * 2;
326 dmmax_mask = ~(dmmax - 1);
328 SYSINIT(vm_mem, SI_BOOT1_VM, SI_ORDER_THIRD, swap_pager_init, NULL)
331 * SWAP_PAGER_SWAP_INIT() - swap pager initialization from pageout process
333 * Expected to be started from pageout process once, prior to entering
336 * Called from the low level boot code only.
339 swap_pager_swap_init(void)
344 * Number of in-transit swap bp operations. Don't
345 * exhaust the pbufs completely. Make sure we
346 * initialize workable values (0 will work for hysteresis
347 * but it isn't very efficient).
349 * The nsw_cluster_max is constrained by the number of pages an XIO
350 * holds, i.e., (MAXPHYS/PAGE_SIZE) and our locally defined
351 * MAX_PAGEOUT_CLUSTER. Also be aware that swap ops are
352 * constrained by the swap device interleave stripe size.
354 * Currently we hardwire nsw_wcount_async to 4. This limit is
355 * designed to prevent other I/O from having high latencies due to
356 * our pageout I/O. The value 4 works well for one or two active swap
357 * devices but is probably a little low if you have more. Even so,
358 * a higher value would probably generate only a limited improvement
359 * with three or four active swap devices since the system does not
360 * typically have to pageout at extreme bandwidths. We will want
361 * at least 2 per swap devices, and 4 is a pretty good value if you
362 * have one NFS swap device due to the command/ack latency over NFS.
363 * So it all works out pretty well.
366 nsw_cluster_max = min((MAXPHYS/PAGE_SIZE), MAX_PAGEOUT_CLUSTER);
368 nsw_rcount = (nswbuf + 1) / 2;
369 nsw_wcount_sync = (nswbuf + 3) / 4;
370 nsw_wcount_async = 4;
371 nsw_wcount_async_max = nsw_wcount_async;
374 * The zone is dynamically allocated so generally size it to
375 * maxswzone (32MB to 512MB of KVM). Set a minimum size based
376 * on physical memory of around 8x (each swblock can hold 16 pages).
378 * With the advent of SSDs (vs HDs) the practical (swap:memory) ratio
379 * has increased dramatically.
381 n = vmstats.v_page_count / 2;
382 if (maxswzone && n < maxswzone / sizeof(struct swblock))
383 n = maxswzone / sizeof(struct swblock);
389 sizeof(struct swblock),
393 if (swap_zone != NULL)
396 * if the allocation failed, try a zone two thirds the
397 * size of the previous attempt.
402 if (swap_zone == NULL)
403 panic("swap_pager_swap_init: swap_zone == NULL");
405 kprintf("Swap zone entries reduced from %d to %d.\n", n2, n);
409 * SWAP_PAGER_ALLOC() - allocate a new OBJT_SWAP VM object and instantiate
410 * its metadata structures.
412 * This routine is called from the mmap and fork code to create a new
413 * OBJT_SWAP object. We do this by creating an OBJT_DEFAULT object
414 * and then converting it with swp_pager_meta_convert().
416 * We only support unnamed objects.
421 swap_pager_alloc(void *handle, off_t size, vm_prot_t prot, off_t offset)
425 KKASSERT(handle == NULL);
426 lwkt_gettoken(&vm_token);
427 object = vm_object_allocate(OBJT_DEFAULT,
428 OFF_TO_IDX(offset + PAGE_MASK + size));
429 swp_pager_meta_convert(object);
430 lwkt_reltoken(&vm_token);
436 * SWAP_PAGER_DEALLOC() - remove swap metadata from object
438 * The swap backing for the object is destroyed. The code is
439 * designed such that we can reinstantiate it later, but this
440 * routine is typically called only when the entire object is
441 * about to be destroyed.
443 * The object must be locked or unreferenceable.
444 * No other requirements.
447 swap_pager_dealloc(vm_object_t object)
449 lwkt_gettoken(&vm_token);
450 vm_object_pip_wait(object, "swpdea");
453 * Free all remaining metadata. We only bother to free it from
454 * the swap meta data. We do not attempt to free swapblk's still
455 * associated with vm_page_t's for this object. We do not care
456 * if paging is still in progress on some objects.
459 swp_pager_meta_free_all(object);
461 lwkt_reltoken(&vm_token);
464 /************************************************************************
465 * SWAP PAGER BITMAP ROUTINES *
466 ************************************************************************/
469 * SWP_PAGER_GETSWAPSPACE() - allocate raw swap space
471 * Allocate swap for the requested number of pages. The starting
472 * swap block number (a page index) is returned or SWAPBLK_NONE
473 * if the allocation failed.
475 * Also has the side effect of advising that somebody made a mistake
476 * when they configured swap and didn't configure enough.
478 * The caller must hold vm_token.
479 * This routine may not block.
481 * NOTE: vm_token must be held to avoid races with bitmap frees from
482 * vm_page_remove() via swap_pager_page_removed().
484 static __inline swblk_t
485 swp_pager_getswapspace(vm_object_t object, int npages)
489 ASSERT_LWKT_TOKEN_HELD(&vm_token);
491 if ((blk = blist_alloc(swapblist, npages)) == SWAPBLK_NONE) {
492 if (swap_pager_full != 2) {
493 kprintf("swap_pager_getswapspace: failed\n");
495 swap_pager_almost_full = 1;
498 swapacctspace(blk, -npages);
499 if (object->type == OBJT_SWAP)
500 vm_swap_anon_use += npages;
502 vm_swap_cache_use += npages;
509 * SWP_PAGER_FREESWAPSPACE() - free raw swap space
511 * This routine returns the specified swap blocks back to the bitmap.
513 * Note: This routine may not block (it could in the old swap code),
514 * and through the use of the new blist routines it does not block.
516 * We must be called at splvm() to avoid races with bitmap frees from
517 * vm_page_remove() aka swap_pager_page_removed().
519 * The caller must hold vm_token.
520 * This routine may not block.
524 swp_pager_freeswapspace(vm_object_t object, swblk_t blk, int npages)
526 struct swdevt *sp = &swdevt[BLK2DEVIDX(blk)];
528 sp->sw_nused -= npages;
529 if (object->type == OBJT_SWAP)
530 vm_swap_anon_use -= npages;
532 vm_swap_cache_use -= npages;
534 if (sp->sw_flags & SW_CLOSING)
537 blist_free(swapblist, blk, npages);
538 vm_swap_size += npages;
543 * SWAP_PAGER_FREESPACE() - frees swap blocks associated with a page
544 * range within an object.
546 * This is a globally accessible routine.
548 * This routine removes swapblk assignments from swap metadata.
550 * The external callers of this routine typically have already destroyed
551 * or renamed vm_page_t's associated with this range in the object so
557 swap_pager_freespace(vm_object_t object, vm_pindex_t start, vm_pindex_t size)
560 lwkt_gettoken(&vm_token);
561 swp_pager_meta_free(object, start, size);
562 lwkt_reltoken(&vm_token);
570 swap_pager_freespace_all(vm_object_t object)
573 lwkt_gettoken(&vm_token);
574 swp_pager_meta_free_all(object);
575 lwkt_reltoken(&vm_token);
580 * This function conditionally frees swap cache swap starting at
581 * (*basei) in the object. (count) swap blocks will be nominally freed.
582 * The actual number of blocks freed can be more or less than the
585 * This function nominally returns the number of blocks freed. However,
586 * the actual number of blocks freed may be less then the returned value.
587 * If the function is unable to exhaust the object or if it is able to
588 * free (approximately) the requested number of blocks it returns
591 * If we exhaust the object we will return a value n <= count.
593 * The caller must hold vm_token.
595 * WARNING! If count == 0 then -1 can be returned as a degenerate case,
596 * callers should always pass a count value > 0.
598 static int swap_pager_condfree_callback(struct swblock *swap, void *data);
601 swap_pager_condfree(vm_object_t object, vm_pindex_t *basei, int count)
603 struct swfreeinfo info;
605 ASSERT_LWKT_TOKEN_HELD(&vm_token);
607 info.object = object;
608 info.basei = *basei; /* skip up to this page index */
609 info.begi = count; /* max swap pages to destroy */
610 info.endi = count * 8; /* max swblocks to scan */
612 swblock_rb_tree_RB_SCAN(&object->swblock_root, rb_swblock_condcmp,
613 swap_pager_condfree_callback, &info);
615 if (info.endi < 0 && info.begi <= count)
616 info.begi = count + 1;
617 return(count - (int)info.begi);
621 * The idea is to free whole meta-block to avoid fragmenting
622 * the swap space or disk I/O. We only do this if NO VM pages
625 * We do not have to deal with clearing PG_SWAPPED in related VM
626 * pages because there are no related VM pages.
628 * The caller must hold vm_token.
631 swap_pager_condfree_callback(struct swblock *swap, void *data)
633 struct swfreeinfo *info = data;
634 vm_object_t object = info->object;
637 for (i = 0; i < SWAP_META_PAGES; ++i) {
638 if (vm_page_lookup(object, swap->swb_index + i))
641 info->basei = swap->swb_index + SWAP_META_PAGES;
642 if (i == SWAP_META_PAGES) {
643 info->begi -= swap->swb_count;
644 swap_pager_freespace(object, swap->swb_index, SWAP_META_PAGES);
647 if ((int)info->begi < 0 || (int)info->endi < 0)
653 * Called by vm_page_alloc() when a new VM page is inserted
654 * into a VM object. Checks whether swap has been assigned to
655 * the page and sets PG_SWAPPED as necessary.
660 swap_pager_page_inserted(vm_page_t m)
662 if (m->object->swblock_count) {
664 lwkt_gettoken(&vm_token);
665 if (swp_pager_meta_ctl(m->object, m->pindex, 0) != SWAPBLK_NONE)
666 vm_page_flag_set(m, PG_SWAPPED);
667 lwkt_reltoken(&vm_token);
673 * SWAP_PAGER_RESERVE() - reserve swap blocks in object
675 * Assigns swap blocks to the specified range within the object. The
676 * swap blocks are not zerod. Any previous swap assignment is destroyed.
678 * Returns 0 on success, -1 on failure.
680 * The caller is responsible for avoiding races in the specified range.
681 * No other requirements.
684 swap_pager_reserve(vm_object_t object, vm_pindex_t start, vm_size_t size)
687 swblk_t blk = SWAPBLK_NONE;
688 vm_pindex_t beg = start; /* save start index */
691 lwkt_gettoken(&vm_token);
695 while ((blk = swp_pager_getswapspace(object, n)) ==
700 swp_pager_meta_free(object, beg,
702 lwkt_reltoken(&vm_token);
708 swp_pager_meta_build(object, start, blk);
714 swp_pager_meta_free(object, start, n);
715 lwkt_reltoken(&vm_token);
721 * SWAP_PAGER_COPY() - copy blocks from source pager to destination pager
722 * and destroy the source.
724 * Copy any valid swapblks from the source to the destination. In
725 * cases where both the source and destination have a valid swapblk,
726 * we keep the destination's.
728 * This routine is allowed to block. It may block allocating metadata
729 * indirectly through swp_pager_meta_build() or if paging is still in
730 * progress on the source.
732 * This routine can be called at any spl
734 * XXX vm_page_collapse() kinda expects us not to block because we
735 * supposedly do not need to allocate memory, but for the moment we
736 * *may* have to get a little memory from the zone allocator, but
737 * it is taken from the interrupt memory. We should be ok.
739 * The source object contains no vm_page_t's (which is just as well)
741 * The source object is of type OBJT_SWAP.
743 * The source and destination objects must be locked or
744 * inaccessible (XXX are they ?)
746 * The caller must hold vm_token.
749 swap_pager_copy(vm_object_t srcobject, vm_object_t dstobject,
750 vm_pindex_t base_index, int destroysource)
754 ASSERT_LWKT_TOKEN_HELD(&vm_token);
758 * transfer source to destination.
760 for (i = 0; i < dstobject->size; ++i) {
764 * Locate (without changing) the swapblk on the destination,
765 * unless it is invalid in which case free it silently, or
766 * if the destination is a resident page, in which case the
767 * source is thrown away.
769 dstaddr = swp_pager_meta_ctl(dstobject, i, 0);
771 if (dstaddr == SWAPBLK_NONE) {
773 * Destination has no swapblk and is not resident,
778 srcaddr = swp_pager_meta_ctl(srcobject,
779 base_index + i, SWM_POP);
781 if (srcaddr != SWAPBLK_NONE)
782 swp_pager_meta_build(dstobject, i, srcaddr);
785 * Destination has valid swapblk or it is represented
786 * by a resident page. We destroy the sourceblock.
788 swp_pager_meta_ctl(srcobject, base_index + i, SWM_FREE);
793 * Free left over swap blocks in source.
795 * We have to revert the type to OBJT_DEFAULT so we do not accidently
796 * double-remove the object from the swap queues.
800 * Reverting the type is not necessary, the caller is going
801 * to destroy srcobject directly, but I'm doing it here
802 * for consistency since we've removed the object from its
805 swp_pager_meta_free_all(srcobject);
806 if (srcobject->type == OBJT_SWAP)
807 srcobject->type = OBJT_DEFAULT;
813 * SWAP_PAGER_HASPAGE() - determine if we have good backing store for
814 * the requested page.
816 * We determine whether good backing store exists for the requested
817 * page and return TRUE if it does, FALSE if it doesn't.
819 * If TRUE, we also try to determine how much valid, contiguous backing
820 * store exists before and after the requested page within a reasonable
821 * distance. We do not try to restrict it to the swap device stripe
822 * (that is handled in getpages/putpages). It probably isn't worth
828 swap_pager_haspage(vm_object_t object, vm_pindex_t pindex)
833 * do we have good backing store at the requested index ?
837 lwkt_gettoken(&vm_token);
838 blk0 = swp_pager_meta_ctl(object, pindex, 0);
840 if (blk0 == SWAPBLK_NONE) {
841 lwkt_reltoken(&vm_token);
845 lwkt_reltoken(&vm_token);
851 * SWAP_PAGER_PAGE_UNSWAPPED() - remove swap backing store related to page
853 * This removes any associated swap backing store, whether valid or
854 * not, from the page. This operates on any VM object, not just OBJT_SWAP
857 * This routine is typically called when a page is made dirty, at
858 * which point any associated swap can be freed. MADV_FREE also
859 * calls us in a special-case situation
861 * NOTE!!! If the page is clean and the swap was valid, the caller
862 * should make the page dirty before calling this routine. This routine
863 * does NOT change the m->dirty status of the page. Also: MADV_FREE
866 * The page must be busied or soft-busied.
867 * The caller must hold vm_token if the caller does not wish to block here.
868 * No other requirements.
871 swap_pager_unswapped(vm_page_t m)
873 if (m->flags & PG_SWAPPED) {
875 lwkt_gettoken(&vm_token);
876 KKASSERT(m->flags & PG_SWAPPED);
877 swp_pager_meta_ctl(m->object, m->pindex, SWM_FREE);
878 vm_page_flag_clear(m, PG_SWAPPED);
879 lwkt_reltoken(&vm_token);
885 * SWAP_PAGER_STRATEGY() - read, write, free blocks
887 * This implements a VM OBJECT strategy function using swap backing store.
888 * This can operate on any VM OBJECT type, not necessarily just OBJT_SWAP
891 * This is intended to be a cacheless interface (i.e. caching occurs at
892 * higher levels), and is also used as a swap-based SSD cache for vnode
893 * and device objects.
895 * All I/O goes directly to and from the swap device.
897 * We currently attempt to run I/O synchronously or asynchronously as
898 * the caller requests. This isn't perfect because we loose error
899 * sequencing when we run multiple ops in parallel to satisfy a request.
900 * But this is swap, so we let it all hang out.
905 swap_pager_strategy(vm_object_t object, struct bio *bio)
907 struct buf *bp = bio->bio_buf;
910 vm_pindex_t biox_blkno = 0;
915 struct bio_track *track;
918 * tracking for swapdev vnode I/Os
920 if (bp->b_cmd == BUF_CMD_READ)
921 track = &swapdev_vp->v_track_read;
923 track = &swapdev_vp->v_track_write;
925 if (bp->b_bcount & PAGE_MASK) {
926 bp->b_error = EINVAL;
927 bp->b_flags |= B_ERROR | B_INVAL;
929 kprintf("swap_pager_strategy: bp %p offset %lld size %d, "
930 "not page bounded\n",
931 bp, (long long)bio->bio_offset, (int)bp->b_bcount);
936 * Clear error indication, initialize page index, count, data pointer.
939 bp->b_flags &= ~B_ERROR;
940 bp->b_resid = bp->b_bcount;
942 start = (vm_pindex_t)(bio->bio_offset >> PAGE_SHIFT);
943 count = howmany(bp->b_bcount, PAGE_SIZE);
947 * Deal with BUF_CMD_FREEBLKS
949 if (bp->b_cmd == BUF_CMD_FREEBLKS) {
951 * FREE PAGE(s) - destroy underlying swap that is no longer
955 lwkt_gettoken(&vm_token);
956 swp_pager_meta_free(object, start, count);
957 lwkt_reltoken(&vm_token);
965 * We need to be able to create a new cluster of I/O's. We cannot
966 * use the caller fields of the passed bio so push a new one.
968 * Because nbio is just a placeholder for the cluster links,
969 * we can biodone() the original bio instead of nbio to make
970 * things a bit more efficient.
972 nbio = push_bio(bio);
973 nbio->bio_offset = bio->bio_offset;
974 nbio->bio_caller_info1.cluster_head = NULL;
975 nbio->bio_caller_info2.cluster_tail = NULL;
981 * Execute read or write
984 lwkt_gettoken(&vm_token);
989 * Obtain block. If block not found and writing, allocate a
990 * new block and build it into the object.
992 blk = swp_pager_meta_ctl(object, start, 0);
993 if ((blk == SWAPBLK_NONE) && bp->b_cmd != BUF_CMD_READ) {
994 blk = swp_pager_getswapspace(object, 1);
995 if (blk == SWAPBLK_NONE) {
996 bp->b_error = ENOMEM;
997 bp->b_flags |= B_ERROR;
1000 swp_pager_meta_build(object, start, blk);
1004 * Do we have to flush our current collection? Yes if:
1006 * - no swap block at this index
1007 * - swap block is not contiguous
1008 * - we cross a physical disk boundry in the
1012 biox && (biox_blkno + btoc(bufx->b_bcount) != blk ||
1013 ((biox_blkno ^ blk) & dmmax_mask)
1016 if (bp->b_cmd == BUF_CMD_READ) {
1017 ++mycpu->gd_cnt.v_swapin;
1018 mycpu->gd_cnt.v_swappgsin += btoc(bufx->b_bcount);
1020 ++mycpu->gd_cnt.v_swapout;
1021 mycpu->gd_cnt.v_swappgsout += btoc(bufx->b_bcount);
1022 bufx->b_dirtyend = bufx->b_bcount;
1026 * Finished with this buf.
1028 KKASSERT(bufx->b_bcount != 0);
1029 if (bufx->b_cmd != BUF_CMD_READ)
1030 bufx->b_dirtyend = bufx->b_bcount;
1036 * Add new swapblk to biox, instantiating biox if necessary.
1037 * Zero-fill reads are able to take a shortcut.
1039 if (blk == SWAPBLK_NONE) {
1041 * We can only get here if we are reading. Since
1042 * we are at splvm() we can safely modify b_resid,
1043 * even if chain ops are in progress.
1045 bzero(data, PAGE_SIZE);
1046 bp->b_resid -= PAGE_SIZE;
1049 /* XXX chain count > 4, wait to <= 4 */
1051 bufx = getpbuf(NULL);
1052 biox = &bufx->b_bio1;
1053 cluster_append(nbio, bufx);
1054 bufx->b_flags |= (bufx->b_flags & B_ORDERED);
1055 bufx->b_cmd = bp->b_cmd;
1056 biox->bio_done = swap_chain_iodone;
1057 biox->bio_offset = (off_t)blk << PAGE_SHIFT;
1058 biox->bio_caller_info1.cluster_parent = nbio;
1061 bufx->b_data = data;
1063 bufx->b_bcount += PAGE_SIZE;
1069 lwkt_reltoken(&vm_token);
1073 * Flush out last buffer
1076 if (bufx->b_cmd == BUF_CMD_READ) {
1077 ++mycpu->gd_cnt.v_swapin;
1078 mycpu->gd_cnt.v_swappgsin += btoc(bufx->b_bcount);
1080 ++mycpu->gd_cnt.v_swapout;
1081 mycpu->gd_cnt.v_swappgsout += btoc(bufx->b_bcount);
1082 bufx->b_dirtyend = bufx->b_bcount;
1084 KKASSERT(bufx->b_bcount);
1085 if (bufx->b_cmd != BUF_CMD_READ)
1086 bufx->b_dirtyend = bufx->b_bcount;
1087 /* biox, bufx = NULL */
1091 * Now initiate all the I/O. Be careful looping on our chain as
1092 * I/O's may complete while we are still initiating them.
1094 * If the request is a 100% sparse read no bios will be present
1095 * and we just biodone() the buffer.
1097 nbio->bio_caller_info2.cluster_tail = NULL;
1098 bufx = nbio->bio_caller_info1.cluster_head;
1102 biox = &bufx->b_bio1;
1104 bufx = bufx->b_cluster_next;
1105 vn_strategy(swapdev_vp, biox);
1112 * Completion of the cluster will also call biodone_chain(nbio).
1113 * We never call biodone(nbio) so we don't have to worry about
1114 * setting up a bio_done callback. It's handled in the sub-IO.
1125 swap_chain_iodone(struct bio *biox)
1128 struct buf *bufx; /* chained sub-buffer */
1129 struct bio *nbio; /* parent nbio with chain glue */
1130 struct buf *bp; /* original bp associated with nbio */
1133 bufx = biox->bio_buf;
1134 nbio = biox->bio_caller_info1.cluster_parent;
1138 * Update the original buffer
1140 KKASSERT(bp != NULL);
1141 if (bufx->b_flags & B_ERROR) {
1142 atomic_set_int(&bufx->b_flags, B_ERROR);
1143 bp->b_error = bufx->b_error; /* race ok */
1144 } else if (bufx->b_resid != 0) {
1145 atomic_set_int(&bufx->b_flags, B_ERROR);
1146 bp->b_error = EINVAL; /* race ok */
1148 atomic_subtract_int(&bp->b_resid, bufx->b_bcount);
1152 * Remove us from the chain.
1154 spin_lock(&bp->b_lock.lk_spinlock);
1155 nextp = &nbio->bio_caller_info1.cluster_head;
1156 while (*nextp != bufx) {
1157 KKASSERT(*nextp != NULL);
1158 nextp = &(*nextp)->b_cluster_next;
1160 *nextp = bufx->b_cluster_next;
1161 chain_empty = (nbio->bio_caller_info1.cluster_head == NULL);
1162 spin_unlock(&bp->b_lock.lk_spinlock);
1165 * Clean up bufx. If the chain is now empty we finish out
1166 * the parent. Note that we may be racing other completions
1167 * so we must use the chain_empty status from above.
1170 if (bp->b_resid != 0 && !(bp->b_flags & B_ERROR)) {
1171 atomic_set_int(&bp->b_flags, B_ERROR);
1172 bp->b_error = EINVAL;
1174 biodone_chain(nbio);
1176 relpbuf(bufx, NULL);
1180 * SWAP_PAGER_GETPAGES() - bring page in from swap
1182 * The requested page may have to be brought in from swap. Calculate the
1183 * swap block and bring in additional pages if possible. All pages must
1184 * have contiguous swap block assignments and reside in the same object.
1186 * The caller has a single vm_object_pip_add() reference prior to
1187 * calling us and we should return with the same.
1189 * The caller has BUSY'd the page. We should return with (*mpp) left busy,
1190 * and any additinal pages unbusied.
1192 * If the caller encounters a PG_RAM page it will pass it to us even though
1193 * it may be valid and dirty. We cannot overwrite the page in this case!
1194 * The case is used to allow us to issue pure read-aheads.
1196 * NOTE! XXX This code does not entirely pipeline yet due to the fact that
1197 * the PG_RAM page is validated at the same time as mreq. What we
1198 * really need to do is issue a separate read-ahead pbuf.
1203 swap_pager_getpage(vm_object_t object, vm_page_t *mpp, int seqaccess)
1214 vm_page_t marray[XIO_INTERNAL_PAGES];
1218 if (mreq->object != object) {
1219 panic("swap_pager_getpages: object mismatch %p/%p",
1226 * We don't want to overwrite a fully valid page as it might be
1227 * dirty. This case can occur when e.g. vm_fault hits a perfectly
1228 * valid page with PG_RAM set.
1230 * In this case we see if the next page is a suitable page-in
1231 * candidate and if it is we issue read-ahead. PG_RAM will be
1232 * set on the last page of the read-ahead to continue the pipeline.
1234 if (mreq->valid == VM_PAGE_BITS_ALL) {
1235 if (swap_burst_read == 0 || mreq->pindex + 1 >= object->size)
1236 return(VM_PAGER_OK);
1238 lwkt_gettoken(&vm_token);
1239 blk = swp_pager_meta_ctl(object, mreq->pindex + 1, 0);
1240 if (blk == SWAPBLK_NONE) {
1241 lwkt_reltoken(&vm_token);
1243 return(VM_PAGER_OK);
1245 m = vm_page_lookup(object, mreq->pindex + 1);
1247 m = vm_page_alloc(object, mreq->pindex + 1,
1250 lwkt_reltoken(&vm_token);
1252 return(VM_PAGER_OK);
1255 if ((m->flags & PG_BUSY) || m->busy || m->valid) {
1256 lwkt_reltoken(&vm_token);
1258 return(VM_PAGER_OK);
1260 vm_page_unqueue_nowakeup(m);
1265 lwkt_reltoken(&vm_token);
1272 * Try to block-read contiguous pages from swap if sequential,
1273 * otherwise just read one page. Contiguous pages from swap must
1274 * reside within a single device stripe because the I/O cannot be
1275 * broken up across multiple stripes.
1277 * Note that blk and iblk can be SWAPBLK_NONE but the loop is
1278 * set up such that the case(s) are handled implicitly.
1281 lwkt_gettoken(&vm_token);
1282 blk = swp_pager_meta_ctl(mreq->object, mreq->pindex, 0);
1285 for (i = 1; swap_burst_read &&
1286 i < XIO_INTERNAL_PAGES &&
1287 mreq->pindex + i < object->size; ++i) {
1290 iblk = swp_pager_meta_ctl(object, mreq->pindex + i, 0);
1291 if (iblk != blk + i)
1293 if ((blk ^ iblk) & dmmax_mask)
1295 m = vm_page_lookup(object, mreq->pindex + i);
1297 m = vm_page_alloc(object, mreq->pindex + i,
1302 if ((m->flags & PG_BUSY) || m->busy || m->valid)
1304 vm_page_unqueue_nowakeup(m);
1310 vm_page_flag_set(marray[i - 1], PG_RAM);
1312 lwkt_reltoken(&vm_token);
1316 * If mreq is the requested page and we have nothing to do return
1317 * VM_PAGER_FAIL. If raonly is set mreq is just another read-ahead
1318 * page and must be cleaned up.
1320 if (blk == SWAPBLK_NONE) {
1323 vnode_pager_freepage(mreq);
1324 return(VM_PAGER_OK);
1326 return(VM_PAGER_FAIL);
1331 * map our page(s) into kva for input
1333 bp = getpbuf_kva(&nsw_rcount);
1335 kva = (vm_offset_t) bp->b_kvabase;
1336 bcopy(marray, bp->b_xio.xio_pages, i * sizeof(vm_page_t));
1337 pmap_qenter(kva, bp->b_xio.xio_pages, i);
1339 bp->b_data = (caddr_t)kva;
1340 bp->b_bcount = PAGE_SIZE * i;
1341 bp->b_xio.xio_npages = i;
1342 bio->bio_done = swp_pager_async_iodone;
1343 bio->bio_offset = (off_t)blk << PAGE_SHIFT;
1344 bio->bio_caller_info1.index = SWBIO_READ;
1347 * Set index. If raonly set the index beyond the array so all
1348 * the pages are treated the same, otherwise the original mreq is
1352 bio->bio_driver_info = (void *)(intptr_t)i;
1354 bio->bio_driver_info = (void *)(intptr_t)0;
1356 for (j = 0; j < i; ++j)
1357 vm_page_flag_set(bp->b_xio.xio_pages[j], PG_SWAPINPROG);
1359 mycpu->gd_cnt.v_swapin++;
1360 mycpu->gd_cnt.v_swappgsin += bp->b_xio.xio_npages;
1363 * We still hold the lock on mreq, and our automatic completion routine
1364 * does not remove it.
1366 vm_object_pip_add(object, bp->b_xio.xio_npages);
1369 * perform the I/O. NOTE!!! bp cannot be considered valid after
1370 * this point because we automatically release it on completion.
1371 * Instead, we look at the one page we are interested in which we
1372 * still hold a lock on even through the I/O completion.
1374 * The other pages in our m[] array are also released on completion,
1375 * so we cannot assume they are valid anymore either.
1377 bp->b_cmd = BUF_CMD_READ;
1379 vn_strategy(swapdev_vp, bio);
1382 * Wait for the page we want to complete. PG_SWAPINPROG is always
1383 * cleared on completion. If an I/O error occurs, SWAPBLK_NONE
1384 * is set in the meta-data.
1386 * If this is a read-ahead only we return immediately without
1390 return(VM_PAGER_OK);
1393 * Read-ahead includes originally requested page case.
1396 lwkt_gettoken(&vm_token);
1397 while ((mreq->flags & PG_SWAPINPROG) != 0) {
1398 vm_page_flag_set(mreq, PG_WANTED | PG_REFERENCED);
1399 mycpu->gd_cnt.v_intrans++;
1400 if (tsleep(mreq, 0, "swread", hz*20)) {
1402 "swap_pager: indefinite wait buffer: "
1403 " offset: %lld, size: %ld\n",
1404 (long long)bio->bio_offset,
1409 lwkt_reltoken(&vm_token);
1413 * mreq is left bussied after completion, but all the other pages
1414 * are freed. If we had an unrecoverable read error the page will
1417 if (mreq->valid != VM_PAGE_BITS_ALL)
1418 return(VM_PAGER_ERROR);
1420 return(VM_PAGER_OK);
1423 * A final note: in a low swap situation, we cannot deallocate swap
1424 * and mark a page dirty here because the caller is likely to mark
1425 * the page clean when we return, causing the page to possibly revert
1426 * to all-zero's later.
1431 * swap_pager_putpages:
1433 * Assign swap (if necessary) and initiate I/O on the specified pages.
1435 * We support both OBJT_DEFAULT and OBJT_SWAP objects. DEFAULT objects
1436 * are automatically converted to SWAP objects.
1438 * In a low memory situation we may block in vn_strategy(), but the new
1439 * vm_page reservation system coupled with properly written VFS devices
1440 * should ensure that no low-memory deadlock occurs. This is an area
1443 * The parent has N vm_object_pip_add() references prior to
1444 * calling us and will remove references for rtvals[] that are
1445 * not set to VM_PAGER_PEND. We need to remove the rest on I/O
1448 * The parent has soft-busy'd the pages it passes us and will unbusy
1449 * those whos rtvals[] entry is not set to VM_PAGER_PEND on return.
1450 * We need to unbusy the rest on I/O completion.
1455 swap_pager_putpages(vm_object_t object, vm_page_t *m, int count,
1456 boolean_t sync, int *rtvals)
1461 if (count && m[0]->object != object) {
1462 panic("swap_pager_getpages: object mismatch %p/%p",
1471 * Turn object into OBJT_SWAP
1472 * check for bogus sysops
1473 * force sync if not pageout process
1475 if (object->type == OBJT_DEFAULT) {
1476 lwkt_gettoken(&vm_token);
1477 if (object->type == OBJT_DEFAULT)
1478 swp_pager_meta_convert(object);
1479 lwkt_reltoken(&vm_token);
1482 if (curthread != pagethread)
1488 * Update nsw parameters from swap_async_max sysctl values.
1489 * Do not let the sysop crash the machine with bogus numbers.
1492 if (swap_async_max != nsw_wcount_async_max) {
1498 if ((n = swap_async_max) > nswbuf / 2)
1505 * Adjust difference ( if possible ). If the current async
1506 * count is too low, we may not be able to make the adjustment
1510 lwkt_gettoken(&vm_token);
1511 n -= nsw_wcount_async_max;
1512 if (nsw_wcount_async + n >= 0) {
1513 nsw_wcount_async += n;
1514 nsw_wcount_async_max += n;
1515 wakeup(&nsw_wcount_async);
1517 lwkt_reltoken(&vm_token);
1524 * Assign swap blocks and issue I/O. We reallocate swap on the fly.
1525 * The page is left dirty until the pageout operation completes
1529 for (i = 0; i < count; i += n) {
1536 * Maximum I/O size is limited by a number of factors.
1539 n = min(BLIST_MAX_ALLOC, count - i);
1540 n = min(n, nsw_cluster_max);
1543 lwkt_gettoken(&vm_token);
1546 * Get biggest block of swap we can. If we fail, fall
1547 * back and try to allocate a smaller block. Don't go
1548 * overboard trying to allocate space if it would overly
1552 (blk = swp_pager_getswapspace(object, n)) == SWAPBLK_NONE &&
1557 if (blk == SWAPBLK_NONE) {
1558 for (j = 0; j < n; ++j)
1559 rtvals[i+j] = VM_PAGER_FAIL;
1560 lwkt_reltoken(&vm_token);
1566 * The I/O we are constructing cannot cross a physical
1567 * disk boundry in the swap stripe. Note: we are still
1570 if ((blk ^ (blk + n)) & dmmax_mask) {
1571 j = ((blk + dmmax) & dmmax_mask) - blk;
1572 swp_pager_freeswapspace(object, blk + j, n - j);
1577 * All I/O parameters have been satisfied, build the I/O
1578 * request and assign the swap space.
1581 bp = getpbuf_kva(&nsw_wcount_sync);
1583 bp = getpbuf_kva(&nsw_wcount_async);
1586 pmap_qenter((vm_offset_t)bp->b_data, &m[i], n);
1588 bp->b_bcount = PAGE_SIZE * n;
1589 bio->bio_offset = (off_t)blk << PAGE_SHIFT;
1591 for (j = 0; j < n; ++j) {
1592 vm_page_t mreq = m[i+j];
1594 swp_pager_meta_build(mreq->object, mreq->pindex,
1596 if (object->type == OBJT_SWAP)
1597 vm_page_dirty(mreq);
1598 rtvals[i+j] = VM_PAGER_OK;
1600 vm_page_flag_set(mreq, PG_SWAPINPROG);
1601 bp->b_xio.xio_pages[j] = mreq;
1603 bp->b_xio.xio_npages = n;
1605 mycpu->gd_cnt.v_swapout++;
1606 mycpu->gd_cnt.v_swappgsout += bp->b_xio.xio_npages;
1608 lwkt_reltoken(&vm_token);
1611 bp->b_dirtyoff = 0; /* req'd for NFS */
1612 bp->b_dirtyend = bp->b_bcount; /* req'd for NFS */
1613 bp->b_cmd = BUF_CMD_WRITE;
1614 bio->bio_caller_info1.index = SWBIO_WRITE;
1619 if (sync == FALSE) {
1620 bio->bio_done = swp_pager_async_iodone;
1622 vn_strategy(swapdev_vp, bio);
1624 for (j = 0; j < n; ++j)
1625 rtvals[i+j] = VM_PAGER_PEND;
1630 * Issue synchrnously.
1632 * Wait for the sync I/O to complete, then update rtvals.
1633 * We just set the rtvals[] to VM_PAGER_PEND so we can call
1634 * our async completion routine at the end, thus avoiding a
1637 bio->bio_caller_info1.index |= SWBIO_SYNC;
1638 bio->bio_done = biodone_sync;
1639 bio->bio_flags |= BIO_SYNC;
1640 vn_strategy(swapdev_vp, bio);
1641 biowait(bio, "swwrt");
1643 for (j = 0; j < n; ++j)
1644 rtvals[i+j] = VM_PAGER_PEND;
1647 * Now that we are through with the bp, we can call the
1648 * normal async completion, which frees everything up.
1650 swp_pager_async_iodone(bio);
1658 swap_pager_newswap(void)
1664 * swp_pager_async_iodone:
1666 * Completion routine for asynchronous reads and writes from/to swap.
1667 * Also called manually by synchronous code to finish up a bp.
1669 * For READ operations, the pages are PG_BUSY'd. For WRITE operations,
1670 * the pages are vm_page_t->busy'd. For READ operations, we PG_BUSY
1671 * unbusy all pages except the 'main' request page. For WRITE
1672 * operations, we vm_page_t->busy'd unbusy all pages ( we can do this
1673 * because we marked them all VM_PAGER_PEND on return from putpages ).
1675 * This routine may not block.
1680 swp_pager_async_iodone(struct bio *bio)
1682 struct buf *bp = bio->bio_buf;
1683 vm_object_t object = NULL;
1690 if (bp->b_flags & B_ERROR) {
1692 "swap_pager: I/O error - %s failed; offset %lld,"
1693 "size %ld, error %d\n",
1694 ((bio->bio_caller_info1.index & SWBIO_READ) ?
1695 "pagein" : "pageout"),
1696 (long long)bio->bio_offset,
1703 * set object, raise to splvm().
1705 if (bp->b_xio.xio_npages)
1706 object = bp->b_xio.xio_pages[0]->object;
1708 lwkt_gettoken(&vm_token);
1711 * remove the mapping for kernel virtual
1713 pmap_qremove((vm_offset_t)bp->b_data, bp->b_xio.xio_npages);
1716 * cleanup pages. If an error occurs writing to swap, we are in
1717 * very serious trouble. If it happens to be a disk error, though,
1718 * we may be able to recover by reassigning the swap later on. So
1719 * in this case we remove the m->swapblk assignment for the page
1720 * but do not free it in the rlist. The errornous block(s) are thus
1721 * never reallocated as swap. Redirty the page and continue.
1723 for (i = 0; i < bp->b_xio.xio_npages; ++i) {
1724 vm_page_t m = bp->b_xio.xio_pages[i];
1726 if (bp->b_flags & B_ERROR) {
1728 * If an error occurs I'd love to throw the swapblk
1729 * away without freeing it back to swapspace, so it
1730 * can never be used again. But I can't from an
1734 if (bio->bio_caller_info1.index & SWBIO_READ) {
1736 * When reading, reqpage needs to stay
1737 * locked for the parent, but all other
1738 * pages can be freed. We still want to
1739 * wakeup the parent waiting on the page,
1740 * though. ( also: pg_reqpage can be -1 and
1741 * not match anything ).
1743 * We have to wake specifically requested pages
1744 * up too because we cleared PG_SWAPINPROG and
1745 * someone may be waiting for that.
1747 * NOTE: for reads, m->dirty will probably
1748 * be overridden by the original caller of
1749 * getpages so don't play cute tricks here.
1751 * NOTE: We can't actually free the page from
1752 * here, because this is an interrupt. It
1753 * is not legal to mess with object->memq
1754 * from an interrupt. Deactivate the page
1759 vm_page_flag_clear(m, PG_ZERO);
1760 vm_page_flag_clear(m, PG_SWAPINPROG);
1763 * bio_driver_info holds the requested page
1766 if (i != (int)(intptr_t)bio->bio_driver_info) {
1767 vm_page_deactivate(m);
1773 * If i == bp->b_pager.pg_reqpage, do not wake
1774 * the page up. The caller needs to.
1778 * If a write error occurs remove the swap
1779 * assignment (note that PG_SWAPPED may or
1780 * may not be set depending on prior activity).
1782 * Re-dirty OBJT_SWAP pages as there is no
1783 * other backing store, we can't throw the
1786 * Non-OBJT_SWAP pages (aka swapcache) must
1787 * not be dirtied since they may not have
1788 * been dirty in the first place, and they
1789 * do have backing store (the vnode).
1791 swp_pager_meta_ctl(m->object, m->pindex,
1793 vm_page_flag_clear(m, PG_SWAPPED);
1794 if (m->object->type == OBJT_SWAP) {
1796 vm_page_activate(m);
1798 vm_page_flag_clear(m, PG_SWAPINPROG);
1799 vm_page_io_finish(m);
1801 } else if (bio->bio_caller_info1.index & SWBIO_READ) {
1803 * NOTE: for reads, m->dirty will probably be
1804 * overridden by the original caller of getpages so
1805 * we cannot set them in order to free the underlying
1806 * swap in a low-swap situation. I don't think we'd
1807 * want to do that anyway, but it was an optimization
1808 * that existed in the old swapper for a time before
1809 * it got ripped out due to precisely this problem.
1811 * clear PG_ZERO in page.
1813 * If not the requested page then deactivate it.
1815 * Note that the requested page, reqpage, is left
1816 * busied, but we still have to wake it up. The
1817 * other pages are released (unbusied) by
1818 * vm_page_wakeup(). We do not set reqpage's
1819 * valid bits here, it is up to the caller.
1823 * NOTE: can't call pmap_clear_modify(m) from an
1824 * interrupt thread, the pmap code may have to map
1825 * non-kernel pmaps and currently asserts the case.
1827 /*pmap_clear_modify(m);*/
1828 m->valid = VM_PAGE_BITS_ALL;
1830 vm_page_flag_clear(m, PG_ZERO | PG_SWAPINPROG);
1831 vm_page_flag_set(m, PG_SWAPPED);
1834 * We have to wake specifically requested pages
1835 * up too because we cleared PG_SWAPINPROG and
1836 * could be waiting for it in getpages. However,
1837 * be sure to not unbusy getpages specifically
1838 * requested page - getpages expects it to be
1841 * bio_driver_info holds the requested page
1843 if (i != (int)(intptr_t)bio->bio_driver_info) {
1844 vm_page_deactivate(m);
1851 * Mark the page clean but do not mess with the
1852 * pmap-layer's modified state. That state should
1853 * also be clear since the caller protected the
1854 * page VM_PROT_READ, but allow the case.
1856 * We are in an interrupt, avoid pmap operations.
1858 * If we have a severe page deficit, deactivate the
1859 * page. Do not try to cache it (which would also
1860 * involve a pmap op), because the page might still
1863 * When using the swap to cache clean vnode pages
1864 * we do not mess with the page dirty bits.
1866 if (m->object->type == OBJT_SWAP)
1868 vm_page_flag_clear(m, PG_SWAPINPROG);
1869 vm_page_flag_set(m, PG_SWAPPED);
1870 if (vm_page_count_severe())
1871 vm_page_deactivate(m);
1873 if (!vm_page_count_severe() || !vm_page_try_to_cache(m))
1874 vm_page_protect(m, VM_PROT_READ);
1876 vm_page_io_finish(m);
1881 * adjust pip. NOTE: the original parent may still have its own
1882 * pip refs on the object.
1886 vm_object_pip_wakeupn(object, bp->b_xio.xio_npages);
1889 * Release the physical I/O buffer.
1891 * NOTE: Due to synchronous operations in the write case b_cmd may
1892 * already be set to BUF_CMD_DONE and BIO_SYNC may have already
1895 if (bio->bio_caller_info1.index & SWBIO_READ)
1896 nswptr = &nsw_rcount;
1897 else if (bio->bio_caller_info1.index & SWBIO_SYNC)
1898 nswptr = &nsw_wcount_sync;
1900 nswptr = &nsw_wcount_async;
1901 bp->b_cmd = BUF_CMD_DONE;
1902 relpbuf(bp, nswptr);
1903 lwkt_reltoken(&vm_token);
1908 * Fault-in a potentially swapped page and remove the swap reference.
1910 static __inline void
1911 swp_pager_fault_page(vm_object_t object, vm_pindex_t pindex)
1917 if (object->type == OBJT_VNODE) {
1919 * Any swap related to a vnode is due to swapcache. We must
1920 * vget() the vnode in case it is not active (otherwise
1921 * vref() will panic). Calling vm_object_page_remove() will
1922 * ensure that any swap ref is removed interlocked with the
1923 * page. clean_only is set to TRUE so we don't throw away
1926 vp = object->handle;
1927 error = vget(vp, LK_SHARED | LK_RETRY | LK_CANRECURSE);
1929 vm_object_page_remove(object, pindex, pindex + 1, TRUE);
1934 * Otherwise it is a normal OBJT_SWAP object and we can
1935 * fault the page in and remove the swap.
1937 m = vm_fault_object_page(object, IDX_TO_OFF(pindex),
1939 VM_FAULT_DIRTY | VM_FAULT_UNSWAP,
1947 swap_pager_swapoff(int devidx)
1950 struct swblock *swap;
1954 lwkt_gettoken(&vm_token);
1955 lwkt_gettoken(&vmobj_token);
1957 TAILQ_FOREACH(object, &vm_object_list, object_list) {
1958 if (object->type == OBJT_SWAP || object->type == OBJT_VNODE) {
1959 RB_FOREACH(swap, swblock_rb_tree, &object->swblock_root) {
1960 for (i = 0; i < SWAP_META_PAGES; ++i) {
1961 v = swap->swb_pages[i];
1962 if (v != SWAPBLK_NONE &&
1963 BLK2DEVIDX(v) == devidx) {
1964 swp_pager_fault_page(
1966 swap->swb_index + i);
1973 lwkt_reltoken(&vmobj_token);
1974 lwkt_reltoken(&vm_token);
1977 * If we fail to locate all swblocks we just fail gracefully and
1978 * do not bother to restore paging on the swap device. If the
1979 * user wants to retry the user can retry.
1981 if (swdevt[devidx].sw_nused)
1987 /************************************************************************
1989 ************************************************************************
1991 * These routines manipulate the swap metadata stored in the
1992 * OBJT_SWAP object. All swp_*() routines must be called at
1993 * splvm() because swap can be freed up by the low level vm_page
1994 * code which might be called from interrupts beyond what splbio() covers.
1996 * Swap metadata is implemented with a global hash and not directly
1997 * linked into the object. Instead the object simply contains
1998 * appropriate tracking counters.
2002 * Lookup the swblock containing the specified swap block index.
2004 * The caller must hold vm_token.
2008 swp_pager_lookup(vm_object_t object, vm_pindex_t index)
2010 index &= ~SWAP_META_MASK;
2011 return (RB_LOOKUP(swblock_rb_tree, &object->swblock_root, index));
2015 * Remove a swblock from the RB tree.
2017 * The caller must hold vm_token.
2021 swp_pager_remove(vm_object_t object, struct swblock *swap)
2023 RB_REMOVE(swblock_rb_tree, &object->swblock_root, swap);
2027 * Convert default object to swap object if necessary
2029 * The caller must hold vm_token.
2032 swp_pager_meta_convert(vm_object_t object)
2034 if (object->type == OBJT_DEFAULT) {
2035 object->type = OBJT_SWAP;
2036 KKASSERT(object->swblock_count == 0);
2041 * SWP_PAGER_META_BUILD() - add swap block to swap meta data for object
2043 * We first convert the object to a swap object if it is a default
2044 * object. Vnode objects do not need to be converted.
2046 * The specified swapblk is added to the object's swap metadata. If
2047 * the swapblk is not valid, it is freed instead. Any previously
2048 * assigned swapblk is freed.
2050 * The caller must hold vm_token.
2053 swp_pager_meta_build(vm_object_t object, vm_pindex_t index, swblk_t swapblk)
2055 struct swblock *swap;
2056 struct swblock *oswap;
2058 KKASSERT(swapblk != SWAPBLK_NONE);
2061 * Convert object if necessary
2063 if (object->type == OBJT_DEFAULT)
2064 swp_pager_meta_convert(object);
2067 * Locate swblock. If not found create, but if we aren't adding
2068 * anything just return. If we run out of space in the map we wait
2069 * and, since the hash table may have changed, retry.
2072 swap = swp_pager_lookup(object, index);
2077 swap = zalloc(swap_zone);
2082 swap->swb_index = index & ~SWAP_META_MASK;
2083 swap->swb_count = 0;
2085 ++object->swblock_count;
2087 for (i = 0; i < SWAP_META_PAGES; ++i)
2088 swap->swb_pages[i] = SWAPBLK_NONE;
2089 oswap = RB_INSERT(swblock_rb_tree, &object->swblock_root, swap);
2090 KKASSERT(oswap == NULL);
2094 * Delete prior contents of metadata
2097 index &= SWAP_META_MASK;
2099 if (swap->swb_pages[index] != SWAPBLK_NONE) {
2100 swp_pager_freeswapspace(object, swap->swb_pages[index], 1);
2105 * Enter block into metadata
2107 swap->swb_pages[index] = swapblk;
2108 if (swapblk != SWAPBLK_NONE)
2113 * SWP_PAGER_META_FREE() - free a range of blocks in the object's swap metadata
2115 * The requested range of blocks is freed, with any associated swap
2116 * returned to the swap bitmap.
2118 * This routine will free swap metadata structures as they are cleaned
2119 * out. This routine does *NOT* operate on swap metadata associated
2120 * with resident pages.
2122 * The caller must hold vm_token.
2124 static int swp_pager_meta_free_callback(struct swblock *swb, void *data);
2127 swp_pager_meta_free(vm_object_t object, vm_pindex_t index, vm_pindex_t count)
2129 struct swfreeinfo info;
2134 if (object->swblock_count == 0) {
2135 KKASSERT(RB_EMPTY(&object->swblock_root));
2142 * Setup for RB tree scan. Note that the pindex range can be huge
2143 * due to the 64 bit page index space so we cannot safely iterate.
2145 info.object = object;
2146 info.basei = index & ~SWAP_META_MASK;
2148 info.endi = index + count - 1;
2149 swblock_rb_tree_RB_SCAN(&object->swblock_root, rb_swblock_scancmp,
2150 swp_pager_meta_free_callback, &info);
2154 * The caller must hold vm_token.
2158 swp_pager_meta_free_callback(struct swblock *swap, void *data)
2160 struct swfreeinfo *info = data;
2161 vm_object_t object = info->object;
2166 * Figure out the range within the swblock. The wider scan may
2167 * return edge-case swap blocks when the start and/or end points
2168 * are in the middle of a block.
2170 if (swap->swb_index < info->begi)
2171 index = (int)info->begi & SWAP_META_MASK;
2175 if (swap->swb_index + SWAP_META_PAGES > info->endi)
2176 eindex = (int)info->endi & SWAP_META_MASK;
2178 eindex = SWAP_META_MASK;
2181 * Scan and free the blocks. The loop terminates early
2182 * if (swap) runs out of blocks and could be freed.
2184 while (index <= eindex) {
2185 swblk_t v = swap->swb_pages[index];
2187 if (v != SWAPBLK_NONE) {
2188 swp_pager_freeswapspace(object, v, 1);
2189 swap->swb_pages[index] = SWAPBLK_NONE;
2190 if (--swap->swb_count == 0) {
2191 swp_pager_remove(object, swap);
2192 zfree(swap_zone, swap);
2193 --object->swblock_count;
2199 /* swap may be invalid here due to zfree above */
2204 * SWP_PAGER_META_FREE_ALL() - destroy all swap metadata associated with object
2206 * This routine locates and destroys all swap metadata associated with
2209 * The caller must hold vm_token.
2212 swp_pager_meta_free_all(vm_object_t object)
2214 struct swblock *swap;
2217 while ((swap = RB_ROOT(&object->swblock_root)) != NULL) {
2218 swp_pager_remove(object, swap);
2219 for (i = 0; i < SWAP_META_PAGES; ++i) {
2220 swblk_t v = swap->swb_pages[i];
2221 if (v != SWAPBLK_NONE) {
2223 swp_pager_freeswapspace(object, v, 1);
2226 if (swap->swb_count != 0)
2227 panic("swap_pager_meta_free_all: swb_count != 0");
2228 zfree(swap_zone, swap);
2229 --object->swblock_count;
2231 KKASSERT(object->swblock_count == 0);
2235 * SWP_PAGER_METACTL() - misc control of swap and vm_page_t meta data.
2237 * This routine is capable of looking up, popping, or freeing
2238 * swapblk assignments in the swap meta data or in the vm_page_t.
2239 * The routine typically returns the swapblk being looked-up, or popped,
2240 * or SWAPBLK_NONE if the block was freed, or SWAPBLK_NONE if the block
2241 * was invalid. This routine will automatically free any invalid
2242 * meta-data swapblks.
2244 * It is not possible to store invalid swapblks in the swap meta data
2245 * (other then a literal 'SWAPBLK_NONE'), so we don't bother checking.
2247 * When acting on a busy resident page and paging is in progress, we
2248 * have to wait until paging is complete but otherwise can act on the
2251 * SWM_FREE remove and free swap block from metadata
2252 * SWM_POP remove from meta data but do not free.. pop it out
2254 * The caller must hold vm_token.
2257 swp_pager_meta_ctl(vm_object_t object, vm_pindex_t index, int flags)
2259 struct swblock *swap;
2262 if (object->swblock_count == 0)
2263 return(SWAPBLK_NONE);
2266 swap = swp_pager_lookup(object, index);
2269 index &= SWAP_META_MASK;
2270 r1 = swap->swb_pages[index];
2272 if (r1 != SWAPBLK_NONE) {
2273 if (flags & SWM_FREE) {
2274 swp_pager_freeswapspace(object, r1, 1);
2277 if (flags & (SWM_FREE|SWM_POP)) {
2278 swap->swb_pages[index] = SWAPBLK_NONE;
2279 if (--swap->swb_count == 0) {
2280 swp_pager_remove(object, swap);
2281 zfree(swap_zone, swap);
2282 --object->swblock_count;