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 */
149 struct swswapoffinfo {
155 * vm_swap_size is in page-sized chunks now. It was DEV_BSIZE'd chunks
159 int swap_pager_full; /* swap space exhaustion (task killing) */
160 int vm_swap_cache_use;
161 int vm_swap_anon_use;
163 static int swap_pager_almost_full; /* swap space exhaustion (w/ hysteresis)*/
164 static int nsw_rcount; /* free read buffers */
165 static int nsw_wcount_sync; /* limit write buffers / synchronous */
166 static int nsw_wcount_async; /* limit write buffers / asynchronous */
167 static int nsw_wcount_async_max;/* assigned maximum */
168 static int nsw_cluster_max; /* maximum VOP I/O allowed */
170 struct blist *swapblist;
171 static int swap_async_max = 4; /* maximum in-progress async I/O's */
172 static int swap_burst_read = 0; /* allow burst reading */
173 static swblk_t swapiterator; /* linearize allocations */
176 extern struct vnode *swapdev_vp;
177 extern struct swdevt *swdevt;
180 #define BLK2DEVIDX(blk) (nswdev > 1 ? blk / dmmax % nswdev : 0)
182 SYSCTL_INT(_vm, OID_AUTO, swap_async_max,
183 CTLFLAG_RW, &swap_async_max, 0, "Maximum running async swap ops");
184 SYSCTL_INT(_vm, OID_AUTO, swap_burst_read,
185 CTLFLAG_RW, &swap_burst_read, 0, "Allow burst reads for pageins");
187 SYSCTL_INT(_vm, OID_AUTO, swap_cache_use,
188 CTLFLAG_RD, &vm_swap_cache_use, 0, "");
189 SYSCTL_INT(_vm, OID_AUTO, swap_anon_use,
190 CTLFLAG_RD, &vm_swap_anon_use, 0, "");
191 SYSCTL_INT(_vm, OID_AUTO, swap_size,
192 CTLFLAG_RD, &vm_swap_size, 0, "");
197 * Red-Black tree for swblock entries
199 * The caller must hold vm_token
201 RB_GENERATE2(swblock_rb_tree, swblock, swb_entry, rb_swblock_compare,
202 vm_pindex_t, swb_index);
205 rb_swblock_compare(struct swblock *swb1, struct swblock *swb2)
207 if (swb1->swb_index < swb2->swb_index)
209 if (swb1->swb_index > swb2->swb_index)
216 rb_swblock_scancmp(struct swblock *swb, void *data)
218 struct swfreeinfo *info = data;
220 if (swb->swb_index < info->basei)
222 if (swb->swb_index > info->endi)
229 rb_swblock_condcmp(struct swblock *swb, void *data)
231 struct swfreeinfo *info = data;
233 if (swb->swb_index < info->basei)
239 * pagerops for OBJT_SWAP - "swap pager". Some ops are also global procedure
240 * calls hooked from other parts of the VM system and do not appear here.
241 * (see vm/swap_pager.h).
244 static void swap_pager_dealloc (vm_object_t object);
245 static int swap_pager_getpage (vm_object_t, vm_page_t *, int);
246 static void swap_chain_iodone(struct bio *biox);
248 struct pagerops swappagerops = {
249 swap_pager_dealloc, /* deallocate an OBJT_SWAP object */
250 swap_pager_getpage, /* pagein */
251 swap_pager_putpages, /* pageout */
252 swap_pager_haspage /* get backing store status for page */
256 * dmmax is in page-sized chunks with the new swap system. It was
257 * dev-bsized chunks in the old. dmmax is always a power of 2.
259 * swap_*() routines are externally accessible. swp_*() routines are
264 static int dmmax_mask;
265 int nswap_lowat = 128; /* in pages, swap_pager_almost_full warn */
266 int nswap_hiwat = 512; /* in pages, swap_pager_almost_full warn */
268 static __inline void swp_sizecheck (void);
269 static void swp_pager_async_iodone (struct bio *bio);
272 * Swap bitmap functions
275 static __inline void swp_pager_freeswapspace(vm_object_t object,
276 swblk_t blk, int npages);
277 static __inline swblk_t swp_pager_getswapspace(vm_object_t object, int npages);
283 static void swp_pager_meta_convert(vm_object_t);
284 static void swp_pager_meta_build(vm_object_t, vm_pindex_t, swblk_t);
285 static void swp_pager_meta_free(vm_object_t, vm_pindex_t, vm_pindex_t);
286 static void swp_pager_meta_free_all(vm_object_t);
287 static swblk_t swp_pager_meta_ctl(vm_object_t, vm_pindex_t, int);
290 * SWP_SIZECHECK() - update swap_pager_full indication
292 * update the swap_pager_almost_full indication and warn when we are
293 * about to run out of swap space, using lowat/hiwat hysteresis.
295 * Clear swap_pager_full ( task killing ) indication when lowat is met.
297 * No restrictions on call
298 * This routine may not block.
304 if (vm_swap_size < nswap_lowat) {
305 if (swap_pager_almost_full == 0) {
306 kprintf("swap_pager: out of swap space\n");
307 swap_pager_almost_full = 1;
311 if (vm_swap_size > nswap_hiwat)
312 swap_pager_almost_full = 0;
317 * SWAP_PAGER_INIT() - initialize the swap pager!
319 * Expected to be started from system init. NOTE: This code is run
320 * before much else so be careful what you depend on. Most of the VM
321 * system has yet to be initialized at this point.
323 * Called from the low level boot code only.
326 swap_pager_init(void *arg __unused)
329 * Device Stripe, in PAGE_SIZE'd blocks
331 dmmax = SWB_NPAGES * 2;
332 dmmax_mask = ~(dmmax - 1);
334 SYSINIT(vm_mem, SI_BOOT1_VM, SI_ORDER_THIRD, swap_pager_init, NULL)
337 * SWAP_PAGER_SWAP_INIT() - swap pager initialization from pageout process
339 * Expected to be started from pageout process once, prior to entering
342 * Called from the low level boot code only.
345 swap_pager_swap_init(void)
350 * Number of in-transit swap bp operations. Don't
351 * exhaust the pbufs completely. Make sure we
352 * initialize workable values (0 will work for hysteresis
353 * but it isn't very efficient).
355 * The nsw_cluster_max is constrained by the number of pages an XIO
356 * holds, i.e., (MAXPHYS/PAGE_SIZE) and our locally defined
357 * MAX_PAGEOUT_CLUSTER. Also be aware that swap ops are
358 * constrained by the swap device interleave stripe size.
360 * Currently we hardwire nsw_wcount_async to 4. This limit is
361 * designed to prevent other I/O from having high latencies due to
362 * our pageout I/O. The value 4 works well for one or two active swap
363 * devices but is probably a little low if you have more. Even so,
364 * a higher value would probably generate only a limited improvement
365 * with three or four active swap devices since the system does not
366 * typically have to pageout at extreme bandwidths. We will want
367 * at least 2 per swap devices, and 4 is a pretty good value if you
368 * have one NFS swap device due to the command/ack latency over NFS.
369 * So it all works out pretty well.
372 nsw_cluster_max = min((MAXPHYS/PAGE_SIZE), MAX_PAGEOUT_CLUSTER);
374 nsw_rcount = (nswbuf + 1) / 2;
375 nsw_wcount_sync = (nswbuf + 3) / 4;
376 nsw_wcount_async = 4;
377 nsw_wcount_async_max = nsw_wcount_async;
380 * The zone is dynamically allocated so generally size it to
381 * maxswzone (32MB to 512MB of KVM). Set a minimum size based
382 * on physical memory of around 8x (each swblock can hold 16 pages).
384 * With the advent of SSDs (vs HDs) the practical (swap:memory) ratio
385 * has increased dramatically.
387 n = vmstats.v_page_count / 2;
388 if (maxswzone && n < maxswzone / sizeof(struct swblock))
389 n = maxswzone / sizeof(struct swblock);
395 sizeof(struct swblock),
399 if (swap_zone != NULL)
402 * if the allocation failed, try a zone two thirds the
403 * size of the previous attempt.
408 if (swap_zone == NULL)
409 panic("swap_pager_swap_init: swap_zone == NULL");
411 kprintf("Swap zone entries reduced from %d to %d.\n", n2, n);
415 * SWAP_PAGER_ALLOC() - allocate a new OBJT_SWAP VM object and instantiate
416 * its metadata structures.
418 * This routine is called from the mmap and fork code to create a new
419 * OBJT_SWAP object. We do this by creating an OBJT_DEFAULT object
420 * and then converting it with swp_pager_meta_convert().
422 * We only support unnamed objects.
427 swap_pager_alloc(void *handle, off_t size, vm_prot_t prot, off_t offset)
431 KKASSERT(handle == NULL);
432 object = vm_object_allocate_hold(OBJT_DEFAULT,
433 OFF_TO_IDX(offset + PAGE_MASK + size));
434 swp_pager_meta_convert(object);
435 vm_object_drop(object);
441 * SWAP_PAGER_DEALLOC() - remove swap metadata from object
443 * The swap backing for the object is destroyed. The code is
444 * designed such that we can reinstantiate it later, but this
445 * routine is typically called only when the entire object is
446 * about to be destroyed.
448 * The object must be locked or unreferenceable.
449 * No other requirements.
452 swap_pager_dealloc(vm_object_t object)
454 vm_object_hold(object);
455 vm_object_pip_wait(object, "swpdea");
458 * Free all remaining metadata. We only bother to free it from
459 * the swap meta data. We do not attempt to free swapblk's still
460 * associated with vm_page_t's for this object. We do not care
461 * if paging is still in progress on some objects.
463 swp_pager_meta_free_all(object);
464 vm_object_drop(object);
467 /************************************************************************
468 * SWAP PAGER BITMAP ROUTINES *
469 ************************************************************************/
472 * SWP_PAGER_GETSWAPSPACE() - allocate raw swap space
474 * Allocate swap for the requested number of pages. The starting
475 * swap block number (a page index) is returned or SWAPBLK_NONE
476 * if the allocation failed.
478 * Also has the side effect of advising that somebody made a mistake
479 * when they configured swap and didn't configure enough.
481 * The caller must hold the object.
482 * This routine may not block.
484 static __inline swblk_t
485 swp_pager_getswapspace(vm_object_t object, int npages)
489 lwkt_gettoken(&vm_token);
490 blk = blist_allocat(swapblist, npages, swapiterator);
491 if (blk == SWAPBLK_NONE)
492 blk = blist_allocat(swapblist, npages, 0);
493 if (blk == SWAPBLK_NONE) {
494 if (swap_pager_full != 2) {
495 kprintf("swap_pager_getswapspace: failed alloc=%d\n",
498 swap_pager_almost_full = 1;
502 swapacctspace(blk, -npages);
503 if (object->type == OBJT_SWAP)
504 vm_swap_anon_use += npages;
506 vm_swap_cache_use += npages;
509 lwkt_reltoken(&vm_token);
514 * SWP_PAGER_FREESWAPSPACE() - free raw swap space
516 * This routine returns the specified swap blocks back to the bitmap.
518 * Note: This routine may not block (it could in the old swap code),
519 * and through the use of the new blist routines it does not block.
521 * We must be called at splvm() to avoid races with bitmap frees from
522 * vm_page_remove() aka swap_pager_page_removed().
524 * This routine may not block.
528 swp_pager_freeswapspace(vm_object_t object, swblk_t blk, int npages)
530 struct swdevt *sp = &swdevt[BLK2DEVIDX(blk)];
532 lwkt_gettoken(&vm_token);
533 sp->sw_nused -= npages;
534 if (object->type == OBJT_SWAP)
535 vm_swap_anon_use -= npages;
537 vm_swap_cache_use -= npages;
539 if (sp->sw_flags & SW_CLOSING) {
540 lwkt_reltoken(&vm_token);
544 blist_free(swapblist, blk, npages);
545 vm_swap_size += npages;
547 lwkt_reltoken(&vm_token);
551 * SWAP_PAGER_FREESPACE() - frees swap blocks associated with a page
552 * range within an object.
554 * This is a globally accessible routine.
556 * This routine removes swapblk assignments from swap metadata.
558 * The external callers of this routine typically have already destroyed
559 * or renamed vm_page_t's associated with this range in the object so
565 swap_pager_freespace(vm_object_t object, vm_pindex_t start, vm_pindex_t size)
567 vm_object_hold(object);
568 swp_pager_meta_free(object, start, size);
569 vm_object_drop(object);
576 swap_pager_freespace_all(vm_object_t object)
578 vm_object_hold(object);
579 swp_pager_meta_free_all(object);
580 vm_object_drop(object);
584 * This function conditionally frees swap cache swap starting at
585 * (*basei) in the object. (count) swap blocks will be nominally freed.
586 * The actual number of blocks freed can be more or less than the
589 * This function nominally returns the number of blocks freed. However,
590 * the actual number of blocks freed may be less then the returned value.
591 * If the function is unable to exhaust the object or if it is able to
592 * free (approximately) the requested number of blocks it returns
595 * If we exhaust the object we will return a value n <= count.
597 * The caller must hold the object.
599 * WARNING! If count == 0 then -1 can be returned as a degenerate case,
600 * callers should always pass a count value > 0.
602 static int swap_pager_condfree_callback(struct swblock *swap, void *data);
605 swap_pager_condfree(vm_object_t object, vm_pindex_t *basei, int count)
607 struct swfreeinfo info;
611 ASSERT_LWKT_TOKEN_HELD(vm_object_token(object));
613 info.object = object;
614 info.basei = *basei; /* skip up to this page index */
615 info.begi = count; /* max swap pages to destroy */
616 info.endi = count * 8; /* max swblocks to scan */
618 swblock_rb_tree_RB_SCAN(&object->swblock_root, rb_swblock_condcmp,
619 swap_pager_condfree_callback, &info);
623 * Take the higher difference swblocks vs pages
625 n = count - (int)info.begi;
626 t = count * 8 - (int)info.endi;
635 * The idea is to free whole meta-block to avoid fragmenting
636 * the swap space or disk I/O. We only do this if NO VM pages
639 * We do not have to deal with clearing PG_SWAPPED in related VM
640 * pages because there are no related VM pages.
642 * The caller must hold the object.
645 swap_pager_condfree_callback(struct swblock *swap, void *data)
647 struct swfreeinfo *info = data;
648 vm_object_t object = info->object;
651 for (i = 0; i < SWAP_META_PAGES; ++i) {
652 if (vm_page_lookup(object, swap->swb_index + i))
655 info->basei = swap->swb_index + SWAP_META_PAGES;
656 if (i == SWAP_META_PAGES) {
657 info->begi -= swap->swb_count;
658 swap_pager_freespace(object, swap->swb_index, SWAP_META_PAGES);
661 if ((int)info->begi < 0 || (int)info->endi < 0)
668 * Called by vm_page_alloc() when a new VM page is inserted
669 * into a VM object. Checks whether swap has been assigned to
670 * the page and sets PG_SWAPPED as necessary.
675 swap_pager_page_inserted(vm_page_t m)
677 if (m->object->swblock_count) {
678 vm_object_hold(m->object);
679 if (swp_pager_meta_ctl(m->object, m->pindex, 0) != SWAPBLK_NONE)
680 vm_page_flag_set(m, PG_SWAPPED);
681 vm_object_drop(m->object);
686 * SWAP_PAGER_RESERVE() - reserve swap blocks in object
688 * Assigns swap blocks to the specified range within the object. The
689 * swap blocks are not zerod. Any previous swap assignment is destroyed.
691 * Returns 0 on success, -1 on failure.
693 * The caller is responsible for avoiding races in the specified range.
694 * No other requirements.
697 swap_pager_reserve(vm_object_t object, vm_pindex_t start, vm_size_t size)
700 swblk_t blk = SWAPBLK_NONE;
701 vm_pindex_t beg = start; /* save start index */
703 vm_object_hold(object);
708 while ((blk = swp_pager_getswapspace(object, n)) ==
713 swp_pager_meta_free(object, beg,
715 vm_object_drop(object);
720 swp_pager_meta_build(object, start, blk);
726 swp_pager_meta_free(object, start, n);
727 vm_object_drop(object);
732 * SWAP_PAGER_COPY() - copy blocks from source pager to destination pager
733 * and destroy the source.
735 * Copy any valid swapblks from the source to the destination. In
736 * cases where both the source and destination have a valid swapblk,
737 * we keep the destination's.
739 * This routine is allowed to block. It may block allocating metadata
740 * indirectly through swp_pager_meta_build() or if paging is still in
741 * progress on the source.
743 * XXX vm_page_collapse() kinda expects us not to block because we
744 * supposedly do not need to allocate memory, but for the moment we
745 * *may* have to get a little memory from the zone allocator, but
746 * it is taken from the interrupt memory. We should be ok.
748 * The source object contains no vm_page_t's (which is just as well)
749 * The source object is of type OBJT_SWAP.
751 * The source and destination objects must be held by the caller.
754 swap_pager_copy(vm_object_t srcobject, vm_object_t dstobject,
755 vm_pindex_t base_index, int destroysource)
759 ASSERT_LWKT_TOKEN_HELD(vm_object_token(srcobject));
760 ASSERT_LWKT_TOKEN_HELD(vm_object_token(dstobject));
763 * transfer source to destination.
765 for (i = 0; i < dstobject->size; ++i) {
769 * Locate (without changing) the swapblk on the destination,
770 * unless it is invalid in which case free it silently, or
771 * if the destination is a resident page, in which case the
772 * source is thrown away.
774 dstaddr = swp_pager_meta_ctl(dstobject, i, 0);
776 if (dstaddr == SWAPBLK_NONE) {
778 * Destination has no swapblk and is not resident,
783 srcaddr = swp_pager_meta_ctl(srcobject,
784 base_index + i, SWM_POP);
786 if (srcaddr != SWAPBLK_NONE)
787 swp_pager_meta_build(dstobject, i, srcaddr);
790 * Destination has valid swapblk or it is represented
791 * by a resident page. We destroy the sourceblock.
793 swp_pager_meta_ctl(srcobject, base_index + i, SWM_FREE);
798 * Free left over swap blocks in source.
800 * We have to revert the type to OBJT_DEFAULT so we do not accidently
801 * double-remove the object from the swap queues.
805 * Reverting the type is not necessary, the caller is going
806 * to destroy srcobject directly, but I'm doing it here
807 * for consistency since we've removed the object from its
810 swp_pager_meta_free_all(srcobject);
811 if (srcobject->type == OBJT_SWAP)
812 srcobject->type = OBJT_DEFAULT;
817 * SWAP_PAGER_HASPAGE() - determine if we have good backing store for
818 * the requested page.
820 * We determine whether good backing store exists for the requested
821 * page and return TRUE if it does, FALSE if it doesn't.
823 * If TRUE, we also try to determine how much valid, contiguous backing
824 * store exists before and after the requested page within a reasonable
825 * distance. We do not try to restrict it to the swap device stripe
826 * (that is handled in getpages/putpages). It probably isn't worth
832 swap_pager_haspage(vm_object_t object, vm_pindex_t pindex)
837 * do we have good backing store at the requested index ?
839 vm_object_hold(object);
840 blk0 = swp_pager_meta_ctl(object, pindex, 0);
842 if (blk0 == SWAPBLK_NONE) {
843 vm_object_drop(object);
846 vm_object_drop(object);
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 can hold the object to avoid blocking, else we might block.
868 * No other requirements.
871 swap_pager_unswapped(vm_page_t m)
873 if (m->flags & PG_SWAPPED) {
874 vm_object_hold(m->object);
875 KKASSERT(m->flags & PG_SWAPPED);
876 swp_pager_meta_ctl(m->object, m->pindex, SWM_FREE);
877 vm_page_flag_clear(m, PG_SWAPPED);
878 vm_object_drop(m->object);
883 * SWAP_PAGER_STRATEGY() - read, write, free blocks
885 * This implements a VM OBJECT strategy function using swap backing store.
886 * This can operate on any VM OBJECT type, not necessarily just OBJT_SWAP
889 * This is intended to be a cacheless interface (i.e. caching occurs at
890 * higher levels), and is also used as a swap-based SSD cache for vnode
891 * and device objects.
893 * All I/O goes directly to and from the swap device.
895 * We currently attempt to run I/O synchronously or asynchronously as
896 * the caller requests. This isn't perfect because we loose error
897 * sequencing when we run multiple ops in parallel to satisfy a request.
898 * But this is swap, so we let it all hang out.
903 swap_pager_strategy(vm_object_t object, struct bio *bio)
905 struct buf *bp = bio->bio_buf;
908 vm_pindex_t biox_blkno = 0;
914 struct bio_track *track;
919 * tracking for swapdev vnode I/Os
921 if (bp->b_cmd == BUF_CMD_READ)
922 track = &swapdev_vp->v_track_read;
924 track = &swapdev_vp->v_track_write;
927 if (bp->b_bcount & PAGE_MASK) {
928 bp->b_error = EINVAL;
929 bp->b_flags |= B_ERROR | B_INVAL;
931 kprintf("swap_pager_strategy: bp %p offset %lld size %d, "
932 "not page bounded\n",
933 bp, (long long)bio->bio_offset, (int)bp->b_bcount);
938 * Clear error indication, initialize page index, count, data pointer.
941 bp->b_flags &= ~B_ERROR;
942 bp->b_resid = bp->b_bcount;
944 start = (vm_pindex_t)(bio->bio_offset >> PAGE_SHIFT);
945 count = howmany(bp->b_bcount, PAGE_SIZE);
949 * Deal with BUF_CMD_FREEBLKS
951 if (bp->b_cmd == BUF_CMD_FREEBLKS) {
953 * FREE PAGE(s) - destroy underlying swap that is no longer
956 vm_object_hold(object);
957 swp_pager_meta_free(object, start, count);
958 vm_object_drop(object);
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
983 vm_object_hold(object);
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 |= (bp->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;
1070 vm_object_drop(object);
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)
1216 vm_page_t marray[XIO_INTERNAL_PAGES];
1220 vm_object_hold(object);
1221 if (mreq->object != object) {
1222 panic("swap_pager_getpages: object mismatch %p/%p",
1229 * We don't want to overwrite a fully valid page as it might be
1230 * dirty. This case can occur when e.g. vm_fault hits a perfectly
1231 * valid page with PG_RAM set.
1233 * In this case we see if the next page is a suitable page-in
1234 * candidate and if it is we issue read-ahead. PG_RAM will be
1235 * set on the last page of the read-ahead to continue the pipeline.
1237 if (mreq->valid == VM_PAGE_BITS_ALL) {
1238 if (swap_burst_read == 0 || mreq->pindex + 1 >= object->size) {
1239 vm_object_drop(object);
1240 return(VM_PAGER_OK);
1242 blk = swp_pager_meta_ctl(object, mreq->pindex + 1, 0);
1243 if (blk == SWAPBLK_NONE) {
1244 vm_object_drop(object);
1245 return(VM_PAGER_OK);
1247 m = vm_page_lookup_busy_try(object, mreq->pindex + 1,
1250 vm_object_drop(object);
1251 return(VM_PAGER_OK);
1252 } else if (m == NULL) {
1254 * Use VM_ALLOC_QUICK to avoid blocking on cache
1257 m = vm_page_alloc(object, mreq->pindex + 1,
1260 vm_object_drop(object);
1261 return(VM_PAGER_OK);
1266 vm_object_drop(object);
1267 return(VM_PAGER_OK);
1269 vm_page_unqueue_nowakeup(m);
1279 * Try to block-read contiguous pages from swap if sequential,
1280 * otherwise just read one page. Contiguous pages from swap must
1281 * reside within a single device stripe because the I/O cannot be
1282 * broken up across multiple stripes.
1284 * Note that blk and iblk can be SWAPBLK_NONE but the loop is
1285 * set up such that the case(s) are handled implicitly.
1287 blk = swp_pager_meta_ctl(mreq->object, mreq->pindex, 0);
1290 for (i = 1; swap_burst_read &&
1291 i < XIO_INTERNAL_PAGES &&
1292 mreq->pindex + i < object->size; ++i) {
1295 iblk = swp_pager_meta_ctl(object, mreq->pindex + i, 0);
1296 if (iblk != blk + i)
1298 if ((blk ^ iblk) & dmmax_mask)
1300 m = vm_page_lookup_busy_try(object, mreq->pindex + i,
1304 } else if (m == NULL) {
1306 * Use VM_ALLOC_QUICK to avoid blocking on cache
1309 m = vm_page_alloc(object, mreq->pindex + i,
1318 vm_page_unqueue_nowakeup(m);
1324 vm_page_flag_set(marray[i - 1], PG_RAM);
1327 * If mreq is the requested page and we have nothing to do return
1328 * VM_PAGER_FAIL. If raonly is set mreq is just another read-ahead
1329 * page and must be cleaned up.
1331 if (blk == SWAPBLK_NONE) {
1334 vnode_pager_freepage(mreq);
1335 vm_object_drop(object);
1336 return(VM_PAGER_OK);
1338 vm_object_drop(object);
1339 return(VM_PAGER_FAIL);
1344 * map our page(s) into kva for input
1346 bp = getpbuf_kva(&nsw_rcount);
1348 kva = (vm_offset_t) bp->b_kvabase;
1349 bcopy(marray, bp->b_xio.xio_pages, i * sizeof(vm_page_t));
1350 pmap_qenter(kva, bp->b_xio.xio_pages, i);
1352 bp->b_data = (caddr_t)kva;
1353 bp->b_bcount = PAGE_SIZE * i;
1354 bp->b_xio.xio_npages = i;
1355 bio->bio_done = swp_pager_async_iodone;
1356 bio->bio_offset = (off_t)blk << PAGE_SHIFT;
1357 bio->bio_caller_info1.index = SWBIO_READ;
1360 * Set index. If raonly set the index beyond the array so all
1361 * the pages are treated the same, otherwise the original mreq is
1365 bio->bio_driver_info = (void *)(intptr_t)i;
1367 bio->bio_driver_info = (void *)(intptr_t)0;
1369 for (j = 0; j < i; ++j)
1370 vm_page_flag_set(bp->b_xio.xio_pages[j], PG_SWAPINPROG);
1372 mycpu->gd_cnt.v_swapin++;
1373 mycpu->gd_cnt.v_swappgsin += bp->b_xio.xio_npages;
1376 * We still hold the lock on mreq, and our automatic completion routine
1377 * does not remove it.
1379 vm_object_pip_add(object, bp->b_xio.xio_npages);
1382 * perform the I/O. NOTE!!! bp cannot be considered valid after
1383 * this point because we automatically release it on completion.
1384 * Instead, we look at the one page we are interested in which we
1385 * still hold a lock on even through the I/O completion.
1387 * The other pages in our m[] array are also released on completion,
1388 * so we cannot assume they are valid anymore either.
1390 bp->b_cmd = BUF_CMD_READ;
1392 vn_strategy(swapdev_vp, bio);
1395 * Wait for the page we want to complete. PG_SWAPINPROG is always
1396 * cleared on completion. If an I/O error occurs, SWAPBLK_NONE
1397 * is set in the meta-data.
1399 * If this is a read-ahead only we return immediately without
1403 vm_object_drop(object);
1404 return(VM_PAGER_OK);
1408 * Read-ahead includes originally requested page case.
1411 flags = mreq->flags;
1413 if ((flags & PG_SWAPINPROG) == 0)
1415 tsleep_interlock(mreq, 0);
1416 if (!atomic_cmpset_int(&mreq->flags, flags,
1417 flags | PG_WANTED | PG_REFERENCED)) {
1420 mycpu->gd_cnt.v_intrans++;
1421 if (tsleep(mreq, PINTERLOCKED, "swread", hz*20)) {
1423 "swap_pager: indefinite wait buffer: "
1424 " offset: %lld, size: %ld\n",
1425 (long long)bio->bio_offset,
1432 * mreq is left bussied after completion, but all the other pages
1433 * are freed. If we had an unrecoverable read error the page will
1436 vm_object_drop(object);
1437 if (mreq->valid != VM_PAGE_BITS_ALL)
1438 return(VM_PAGER_ERROR);
1440 return(VM_PAGER_OK);
1443 * A final note: in a low swap situation, we cannot deallocate swap
1444 * and mark a page dirty here because the caller is likely to mark
1445 * the page clean when we return, causing the page to possibly revert
1446 * to all-zero's later.
1451 * swap_pager_putpages:
1453 * Assign swap (if necessary) and initiate I/O on the specified pages.
1455 * We support both OBJT_DEFAULT and OBJT_SWAP objects. DEFAULT objects
1456 * are automatically converted to SWAP objects.
1458 * In a low memory situation we may block in vn_strategy(), but the new
1459 * vm_page reservation system coupled with properly written VFS devices
1460 * should ensure that no low-memory deadlock occurs. This is an area
1463 * The parent has N vm_object_pip_add() references prior to
1464 * calling us and will remove references for rtvals[] that are
1465 * not set to VM_PAGER_PEND. We need to remove the rest on I/O
1468 * The parent has soft-busy'd the pages it passes us and will unbusy
1469 * those whos rtvals[] entry is not set to VM_PAGER_PEND on return.
1470 * We need to unbusy the rest on I/O completion.
1475 swap_pager_putpages(vm_object_t object, vm_page_t *m, int count,
1476 boolean_t sync, int *rtvals)
1481 vm_object_hold(object);
1483 if (count && m[0]->object != object) {
1484 panic("swap_pager_getpages: object mismatch %p/%p",
1493 * Turn object into OBJT_SWAP
1494 * check for bogus sysops
1495 * force sync if not pageout process
1497 if (object->type == OBJT_DEFAULT) {
1498 if (object->type == OBJT_DEFAULT)
1499 swp_pager_meta_convert(object);
1502 if (curthread != pagethread)
1508 * Update nsw parameters from swap_async_max sysctl values.
1509 * Do not let the sysop crash the machine with bogus numbers.
1511 if (swap_async_max != nsw_wcount_async_max) {
1517 if ((n = swap_async_max) > nswbuf / 2)
1524 * Adjust difference ( if possible ). If the current async
1525 * count is too low, we may not be able to make the adjustment
1528 * vm_token needed for nsw_wcount sleep interlock
1530 lwkt_gettoken(&vm_token);
1531 n -= nsw_wcount_async_max;
1532 if (nsw_wcount_async + n >= 0) {
1533 nsw_wcount_async_max += n;
1534 pbuf_adjcount(&nsw_wcount_async, n);
1536 lwkt_reltoken(&vm_token);
1542 * Assign swap blocks and issue I/O. We reallocate swap on the fly.
1543 * The page is left dirty until the pageout operation completes
1547 for (i = 0; i < count; i += n) {
1554 * Maximum I/O size is limited by a number of factors.
1557 n = min(BLIST_MAX_ALLOC, count - i);
1558 n = min(n, nsw_cluster_max);
1560 lwkt_gettoken(&vm_token);
1563 * Get biggest block of swap we can. If we fail, fall
1564 * back and try to allocate a smaller block. Don't go
1565 * overboard trying to allocate space if it would overly
1569 (blk = swp_pager_getswapspace(object, n)) == SWAPBLK_NONE &&
1574 if (blk == SWAPBLK_NONE) {
1575 for (j = 0; j < n; ++j)
1576 rtvals[i+j] = VM_PAGER_FAIL;
1577 lwkt_reltoken(&vm_token);
1582 * The I/O we are constructing cannot cross a physical
1583 * disk boundry in the swap stripe. Note: we are still
1586 if ((blk ^ (blk + n)) & dmmax_mask) {
1587 j = ((blk + dmmax) & dmmax_mask) - blk;
1588 swp_pager_freeswapspace(object, blk + j, n - j);
1593 * All I/O parameters have been satisfied, build the I/O
1594 * request and assign the swap space.
1597 bp = getpbuf_kva(&nsw_wcount_sync);
1599 bp = getpbuf_kva(&nsw_wcount_async);
1602 lwkt_reltoken(&vm_token);
1604 pmap_qenter((vm_offset_t)bp->b_data, &m[i], n);
1606 bp->b_bcount = PAGE_SIZE * n;
1607 bio->bio_offset = (off_t)blk << PAGE_SHIFT;
1609 for (j = 0; j < n; ++j) {
1610 vm_page_t mreq = m[i+j];
1612 swp_pager_meta_build(mreq->object, mreq->pindex,
1614 if (object->type == OBJT_SWAP)
1615 vm_page_dirty(mreq);
1616 rtvals[i+j] = VM_PAGER_OK;
1618 vm_page_flag_set(mreq, PG_SWAPINPROG);
1619 bp->b_xio.xio_pages[j] = mreq;
1621 bp->b_xio.xio_npages = n;
1623 mycpu->gd_cnt.v_swapout++;
1624 mycpu->gd_cnt.v_swappgsout += bp->b_xio.xio_npages;
1626 bp->b_dirtyoff = 0; /* req'd for NFS */
1627 bp->b_dirtyend = bp->b_bcount; /* req'd for NFS */
1628 bp->b_cmd = BUF_CMD_WRITE;
1629 bio->bio_caller_info1.index = SWBIO_WRITE;
1634 if (sync == FALSE) {
1635 bio->bio_done = swp_pager_async_iodone;
1637 vn_strategy(swapdev_vp, bio);
1639 for (j = 0; j < n; ++j)
1640 rtvals[i+j] = VM_PAGER_PEND;
1645 * Issue synchrnously.
1647 * Wait for the sync I/O to complete, then update rtvals.
1648 * We just set the rtvals[] to VM_PAGER_PEND so we can call
1649 * our async completion routine at the end, thus avoiding a
1652 bio->bio_caller_info1.index |= SWBIO_SYNC;
1653 bio->bio_done = biodone_sync;
1654 bio->bio_flags |= BIO_SYNC;
1655 vn_strategy(swapdev_vp, bio);
1656 biowait(bio, "swwrt");
1658 for (j = 0; j < n; ++j)
1659 rtvals[i+j] = VM_PAGER_PEND;
1662 * Now that we are through with the bp, we can call the
1663 * normal async completion, which frees everything up.
1665 swp_pager_async_iodone(bio);
1667 vm_object_drop(object);
1674 swap_pager_newswap(void)
1680 * swp_pager_async_iodone:
1682 * Completion routine for asynchronous reads and writes from/to swap.
1683 * Also called manually by synchronous code to finish up a bp.
1685 * For READ operations, the pages are PG_BUSY'd. For WRITE operations,
1686 * the pages are vm_page_t->busy'd. For READ operations, we PG_BUSY
1687 * unbusy all pages except the 'main' request page. For WRITE
1688 * operations, we vm_page_t->busy'd unbusy all pages ( we can do this
1689 * because we marked them all VM_PAGER_PEND on return from putpages ).
1691 * This routine may not block.
1696 swp_pager_async_iodone(struct bio *bio)
1698 struct buf *bp = bio->bio_buf;
1699 vm_object_t object = NULL;
1706 if (bp->b_flags & B_ERROR) {
1708 "swap_pager: I/O error - %s failed; offset %lld,"
1709 "size %ld, error %d\n",
1710 ((bio->bio_caller_info1.index & SWBIO_READ) ?
1711 "pagein" : "pageout"),
1712 (long long)bio->bio_offset,
1719 * set object, raise to splvm().
1721 if (bp->b_xio.xio_npages)
1722 object = bp->b_xio.xio_pages[0]->object;
1725 * remove the mapping for kernel virtual
1727 pmap_qremove((vm_offset_t)bp->b_data, bp->b_xio.xio_npages);
1730 * cleanup pages. If an error occurs writing to swap, we are in
1731 * very serious trouble. If it happens to be a disk error, though,
1732 * we may be able to recover by reassigning the swap later on. So
1733 * in this case we remove the m->swapblk assignment for the page
1734 * but do not free it in the rlist. The errornous block(s) are thus
1735 * never reallocated as swap. Redirty the page and continue.
1737 for (i = 0; i < bp->b_xio.xio_npages; ++i) {
1738 vm_page_t m = bp->b_xio.xio_pages[i];
1740 if (bp->b_flags & B_ERROR) {
1742 * If an error occurs I'd love to throw the swapblk
1743 * away without freeing it back to swapspace, so it
1744 * can never be used again. But I can't from an
1748 if (bio->bio_caller_info1.index & SWBIO_READ) {
1750 * When reading, reqpage needs to stay
1751 * locked for the parent, but all other
1752 * pages can be freed. We still want to
1753 * wakeup the parent waiting on the page,
1754 * though. ( also: pg_reqpage can be -1 and
1755 * not match anything ).
1757 * We have to wake specifically requested pages
1758 * up too because we cleared PG_SWAPINPROG and
1759 * someone may be waiting for that.
1761 * NOTE: for reads, m->dirty will probably
1762 * be overridden by the original caller of
1763 * getpages so don't play cute tricks here.
1765 * NOTE: We can't actually free the page from
1766 * here, because this is an interrupt. It
1767 * is not legal to mess with object->memq
1768 * from an interrupt. Deactivate the page
1773 vm_page_flag_clear(m, PG_ZERO);
1774 vm_page_flag_clear(m, PG_SWAPINPROG);
1777 * bio_driver_info holds the requested page
1780 if (i != (int)(intptr_t)bio->bio_driver_info) {
1781 vm_page_deactivate(m);
1787 * If i == bp->b_pager.pg_reqpage, do not wake
1788 * the page up. The caller needs to.
1792 * If a write error occurs remove the swap
1793 * assignment (note that PG_SWAPPED may or
1794 * may not be set depending on prior activity).
1796 * Re-dirty OBJT_SWAP pages as there is no
1797 * other backing store, we can't throw the
1800 * Non-OBJT_SWAP pages (aka swapcache) must
1801 * not be dirtied since they may not have
1802 * been dirty in the first place, and they
1803 * do have backing store (the vnode).
1805 vm_page_busy_wait(m, FALSE, "swadpg");
1806 swp_pager_meta_ctl(m->object, m->pindex,
1808 vm_page_flag_clear(m, PG_SWAPPED);
1809 if (m->object->type == OBJT_SWAP) {
1811 vm_page_activate(m);
1813 vm_page_flag_clear(m, PG_SWAPINPROG);
1814 vm_page_io_finish(m);
1817 } else if (bio->bio_caller_info1.index & SWBIO_READ) {
1819 * NOTE: for reads, m->dirty will probably be
1820 * overridden by the original caller of getpages so
1821 * we cannot set them in order to free the underlying
1822 * swap in a low-swap situation. I don't think we'd
1823 * want to do that anyway, but it was an optimization
1824 * that existed in the old swapper for a time before
1825 * it got ripped out due to precisely this problem.
1827 * clear PG_ZERO in page.
1829 * If not the requested page then deactivate it.
1831 * Note that the requested page, reqpage, is left
1832 * busied, but we still have to wake it up. The
1833 * other pages are released (unbusied) by
1834 * vm_page_wakeup(). We do not set reqpage's
1835 * valid bits here, it is up to the caller.
1839 * NOTE: can't call pmap_clear_modify(m) from an
1840 * interrupt thread, the pmap code may have to map
1841 * non-kernel pmaps and currently asserts the case.
1843 /*pmap_clear_modify(m);*/
1844 m->valid = VM_PAGE_BITS_ALL;
1846 vm_page_flag_clear(m, PG_ZERO | PG_SWAPINPROG);
1847 vm_page_flag_set(m, PG_SWAPPED);
1850 * We have to wake specifically requested pages
1851 * up too because we cleared PG_SWAPINPROG and
1852 * could be waiting for it in getpages. However,
1853 * be sure to not unbusy getpages specifically
1854 * requested page - getpages expects it to be
1857 * bio_driver_info holds the requested page
1859 if (i != (int)(intptr_t)bio->bio_driver_info) {
1860 vm_page_deactivate(m);
1867 * Mark the page clean but do not mess with the
1868 * pmap-layer's modified state. That state should
1869 * also be clear since the caller protected the
1870 * page VM_PROT_READ, but allow the case.
1872 * We are in an interrupt, avoid pmap operations.
1874 * If we have a severe page deficit, deactivate the
1875 * page. Do not try to cache it (which would also
1876 * involve a pmap op), because the page might still
1879 * When using the swap to cache clean vnode pages
1880 * we do not mess with the page dirty bits.
1882 vm_page_busy_wait(m, FALSE, "swadpg");
1883 if (m->object->type == OBJT_SWAP)
1885 vm_page_flag_clear(m, PG_SWAPINPROG);
1886 vm_page_flag_set(m, PG_SWAPPED);
1887 if (vm_page_count_severe())
1888 vm_page_deactivate(m);
1890 if (!vm_page_count_severe() || !vm_page_try_to_cache(m))
1891 vm_page_protect(m, VM_PROT_READ);
1893 vm_page_io_finish(m);
1899 * adjust pip. NOTE: the original parent may still have its own
1900 * pip refs on the object.
1904 vm_object_pip_wakeup_n(object, bp->b_xio.xio_npages);
1907 * Release the physical I/O buffer.
1909 * NOTE: Due to synchronous operations in the write case b_cmd may
1910 * already be set to BUF_CMD_DONE and BIO_SYNC may have already
1913 * Use vm_token to interlock nsw_rcount/wcount wakeup?
1915 lwkt_gettoken(&vm_token);
1916 if (bio->bio_caller_info1.index & SWBIO_READ)
1917 nswptr = &nsw_rcount;
1918 else if (bio->bio_caller_info1.index & SWBIO_SYNC)
1919 nswptr = &nsw_wcount_sync;
1921 nswptr = &nsw_wcount_async;
1922 bp->b_cmd = BUF_CMD_DONE;
1923 relpbuf(bp, nswptr);
1924 lwkt_reltoken(&vm_token);
1928 * Fault-in a potentially swapped page and remove the swap reference.
1929 * (used by swapoff code)
1931 * object must be held.
1933 static __inline void
1934 swp_pager_fault_page(vm_object_t object, vm_pindex_t pindex)
1940 ASSERT_LWKT_TOKEN_HELD(vm_object_token(object));
1942 if (object->type == OBJT_VNODE) {
1944 * Any swap related to a vnode is due to swapcache. We must
1945 * vget() the vnode in case it is not active (otherwise
1946 * vref() will panic). Calling vm_object_page_remove() will
1947 * ensure that any swap ref is removed interlocked with the
1948 * page. clean_only is set to TRUE so we don't throw away
1951 vp = object->handle;
1952 error = vget(vp, LK_SHARED | LK_RETRY | LK_CANRECURSE);
1954 vm_object_page_remove(object, pindex, pindex + 1, TRUE);
1959 * Otherwise it is a normal OBJT_SWAP object and we can
1960 * fault the page in and remove the swap.
1962 m = vm_fault_object_page(object, IDX_TO_OFF(pindex),
1964 VM_FAULT_DIRTY | VM_FAULT_UNSWAP,
1972 * This removes all swap blocks related to a particular device. We have
1973 * to be careful of ripups during the scan.
1975 static int swp_pager_swapoff_callback(struct swblock *swap, void *data);
1978 swap_pager_swapoff(int devidx)
1980 struct vm_object marker;
1982 struct swswapoffinfo info;
1984 bzero(&marker, sizeof(marker));
1985 marker.type = OBJT_MARKER;
1987 lwkt_gettoken(&vmobj_token);
1988 TAILQ_INSERT_HEAD(&vm_object_list, &marker, object_list);
1990 while ((object = TAILQ_NEXT(&marker, object_list)) != NULL) {
1991 if (object->type == OBJT_MARKER)
1993 if (object->type != OBJT_SWAP && object->type != OBJT_VNODE)
1995 vm_object_hold(object);
1996 if (object->type != OBJT_SWAP && object->type != OBJT_VNODE) {
1997 vm_object_drop(object);
2000 info.object = object;
2001 info.devidx = devidx;
2002 swblock_rb_tree_RB_SCAN(&object->swblock_root,
2004 swp_pager_swapoff_callback,
2006 vm_object_drop(object);
2008 if (object == TAILQ_NEXT(&marker, object_list)) {
2009 TAILQ_REMOVE(&vm_object_list, &marker, object_list);
2010 TAILQ_INSERT_AFTER(&vm_object_list, object,
2011 &marker, object_list);
2014 TAILQ_REMOVE(&vm_object_list, &marker, object_list);
2015 lwkt_reltoken(&vmobj_token);
2018 * If we fail to locate all swblocks we just fail gracefully and
2019 * do not bother to restore paging on the swap device. If the
2020 * user wants to retry the user can retry.
2022 if (swdevt[devidx].sw_nused)
2030 swp_pager_swapoff_callback(struct swblock *swap, void *data)
2032 struct swswapoffinfo *info = data;
2033 vm_object_t object = info->object;
2038 index = swap->swb_index;
2039 for (i = 0; i < SWAP_META_PAGES; ++i) {
2041 * Make sure we don't race a dying object. This will
2042 * kill the scan of the object's swap blocks entirely.
2044 if (object->flags & OBJ_DEAD)
2048 * Fault the page, which can obviously block. If the swap
2049 * structure disappears break out.
2051 v = swap->swb_pages[i];
2052 if (v != SWAPBLK_NONE && BLK2DEVIDX(v) == info->devidx) {
2053 swp_pager_fault_page(object, swap->swb_index + i);
2054 /* swap ptr might go away */
2055 if (RB_LOOKUP(swblock_rb_tree,
2056 &object->swblock_root, index) != swap) {
2064 /************************************************************************
2066 ************************************************************************
2068 * These routines manipulate the swap metadata stored in the
2069 * OBJT_SWAP object. All swp_*() routines must be called at
2070 * splvm() because swap can be freed up by the low level vm_page
2071 * code which might be called from interrupts beyond what splbio() covers.
2073 * Swap metadata is implemented with a global hash and not directly
2074 * linked into the object. Instead the object simply contains
2075 * appropriate tracking counters.
2079 * Lookup the swblock containing the specified swap block index.
2081 * The caller must hold the object.
2085 swp_pager_lookup(vm_object_t object, vm_pindex_t index)
2087 ASSERT_LWKT_TOKEN_HELD(vm_object_token(object));
2088 index &= ~(vm_pindex_t)SWAP_META_MASK;
2089 return (RB_LOOKUP(swblock_rb_tree, &object->swblock_root, index));
2093 * Remove a swblock from the RB tree.
2095 * The caller must hold the object.
2099 swp_pager_remove(vm_object_t object, struct swblock *swap)
2101 ASSERT_LWKT_TOKEN_HELD(vm_object_token(object));
2102 RB_REMOVE(swblock_rb_tree, &object->swblock_root, swap);
2106 * Convert default object to swap object if necessary
2108 * The caller must hold the object.
2111 swp_pager_meta_convert(vm_object_t object)
2113 if (object->type == OBJT_DEFAULT) {
2114 object->type = OBJT_SWAP;
2115 KKASSERT(object->swblock_count == 0);
2120 * SWP_PAGER_META_BUILD() - add swap block to swap meta data for object
2122 * We first convert the object to a swap object if it is a default
2123 * object. Vnode objects do not need to be converted.
2125 * The specified swapblk is added to the object's swap metadata. If
2126 * the swapblk is not valid, it is freed instead. Any previously
2127 * assigned swapblk is freed.
2129 * The caller must hold the object.
2132 swp_pager_meta_build(vm_object_t object, vm_pindex_t index, swblk_t swapblk)
2134 struct swblock *swap;
2135 struct swblock *oswap;
2138 KKASSERT(swapblk != SWAPBLK_NONE);
2139 ASSERT_LWKT_TOKEN_HELD(vm_object_token(object));
2142 * Convert object if necessary
2144 if (object->type == OBJT_DEFAULT)
2145 swp_pager_meta_convert(object);
2148 * Locate swblock. If not found create, but if we aren't adding
2149 * anything just return. If we run out of space in the map we wait
2150 * and, since the hash table may have changed, retry.
2153 swap = swp_pager_lookup(object, index);
2158 swap = zalloc(swap_zone);
2163 swap->swb_index = index & ~(vm_pindex_t)SWAP_META_MASK;
2164 swap->swb_count = 0;
2166 ++object->swblock_count;
2168 for (i = 0; i < SWAP_META_PAGES; ++i)
2169 swap->swb_pages[i] = SWAPBLK_NONE;
2170 oswap = RB_INSERT(swblock_rb_tree, &object->swblock_root, swap);
2171 KKASSERT(oswap == NULL);
2175 * Delete prior contents of metadata.
2177 * NOTE: Decrement swb_count after the freeing operation (which
2178 * might block) to prevent racing destruction of the swblock.
2180 index &= SWAP_META_MASK;
2182 while ((v = swap->swb_pages[index]) != SWAPBLK_NONE) {
2183 swap->swb_pages[index] = SWAPBLK_NONE;
2185 swp_pager_freeswapspace(object, v, 1);
2190 * Enter block into metadata
2192 swap->swb_pages[index] = swapblk;
2193 if (swapblk != SWAPBLK_NONE)
2198 * SWP_PAGER_META_FREE() - free a range of blocks in the object's swap metadata
2200 * The requested range of blocks is freed, with any associated swap
2201 * returned to the swap bitmap.
2203 * This routine will free swap metadata structures as they are cleaned
2204 * out. This routine does *NOT* operate on swap metadata associated
2205 * with resident pages.
2207 * The caller must hold the object.
2209 static int swp_pager_meta_free_callback(struct swblock *swb, void *data);
2212 swp_pager_meta_free(vm_object_t object, vm_pindex_t index, vm_pindex_t count)
2214 struct swfreeinfo info;
2216 ASSERT_LWKT_TOKEN_HELD(vm_object_token(object));
2221 if (object->swblock_count == 0) {
2222 KKASSERT(RB_EMPTY(&object->swblock_root));
2229 * Setup for RB tree scan. Note that the pindex range can be huge
2230 * due to the 64 bit page index space so we cannot safely iterate.
2232 info.object = object;
2233 info.basei = index & ~(vm_pindex_t)SWAP_META_MASK;
2235 info.endi = index + count - 1;
2236 swblock_rb_tree_RB_SCAN(&object->swblock_root, rb_swblock_scancmp,
2237 swp_pager_meta_free_callback, &info);
2241 * The caller must hold the object.
2245 swp_pager_meta_free_callback(struct swblock *swap, void *data)
2247 struct swfreeinfo *info = data;
2248 vm_object_t object = info->object;
2253 * Figure out the range within the swblock. The wider scan may
2254 * return edge-case swap blocks when the start and/or end points
2255 * are in the middle of a block.
2257 if (swap->swb_index < info->begi)
2258 index = (int)info->begi & SWAP_META_MASK;
2262 if (swap->swb_index + SWAP_META_PAGES > info->endi)
2263 eindex = (int)info->endi & SWAP_META_MASK;
2265 eindex = SWAP_META_MASK;
2268 * Scan and free the blocks. The loop terminates early
2269 * if (swap) runs out of blocks and could be freed.
2271 * NOTE: Decrement swb_count after swp_pager_freeswapspace()
2272 * to deal with a zfree race.
2274 while (index <= eindex) {
2275 swblk_t v = swap->swb_pages[index];
2277 if (v != SWAPBLK_NONE) {
2278 swap->swb_pages[index] = SWAPBLK_NONE;
2280 swp_pager_freeswapspace(object, v, 1);
2281 if (--swap->swb_count == 0) {
2282 swp_pager_remove(object, swap);
2283 zfree(swap_zone, swap);
2284 --object->swblock_count;
2291 /* swap may be invalid here due to zfree above */
2298 * SWP_PAGER_META_FREE_ALL() - destroy all swap metadata associated with object
2300 * This routine locates and destroys all swap metadata associated with
2303 * NOTE: Decrement swb_count after the freeing operation (which
2304 * might block) to prevent racing destruction of the swblock.
2306 * The caller must hold the object.
2309 swp_pager_meta_free_all(vm_object_t object)
2311 struct swblock *swap;
2314 ASSERT_LWKT_TOKEN_HELD(vm_object_token(object));
2316 while ((swap = RB_ROOT(&object->swblock_root)) != NULL) {
2317 swp_pager_remove(object, swap);
2318 for (i = 0; i < SWAP_META_PAGES; ++i) {
2319 swblk_t v = swap->swb_pages[i];
2320 if (v != SWAPBLK_NONE) {
2322 swp_pager_freeswapspace(object, v, 1);
2326 if (swap->swb_count != 0)
2327 panic("swap_pager_meta_free_all: swb_count != 0");
2328 zfree(swap_zone, swap);
2329 --object->swblock_count;
2332 KKASSERT(object->swblock_count == 0);
2336 * SWP_PAGER_METACTL() - misc control of swap and vm_page_t meta data.
2338 * This routine is capable of looking up, popping, or freeing
2339 * swapblk assignments in the swap meta data or in the vm_page_t.
2340 * The routine typically returns the swapblk being looked-up, or popped,
2341 * or SWAPBLK_NONE if the block was freed, or SWAPBLK_NONE if the block
2342 * was invalid. This routine will automatically free any invalid
2343 * meta-data swapblks.
2345 * It is not possible to store invalid swapblks in the swap meta data
2346 * (other then a literal 'SWAPBLK_NONE'), so we don't bother checking.
2348 * When acting on a busy resident page and paging is in progress, we
2349 * have to wait until paging is complete but otherwise can act on the
2352 * SWM_FREE remove and free swap block from metadata
2353 * SWM_POP remove from meta data but do not free.. pop it out
2355 * The caller must hold the object.
2358 swp_pager_meta_ctl(vm_object_t object, vm_pindex_t index, int flags)
2360 struct swblock *swap;
2363 if (object->swblock_count == 0)
2364 return(SWAPBLK_NONE);
2367 swap = swp_pager_lookup(object, index);
2370 index &= SWAP_META_MASK;
2371 r1 = swap->swb_pages[index];
2373 if (r1 != SWAPBLK_NONE) {
2374 if (flags & (SWM_FREE|SWM_POP)) {
2375 swap->swb_pages[index] = SWAPBLK_NONE;
2376 if (--swap->swb_count == 0) {
2377 swp_pager_remove(object, swap);
2378 zfree(swap_zone, swap);
2379 --object->swblock_count;
2382 /* swap ptr may be invalid */
2383 if (flags & SWM_FREE) {
2384 swp_pager_freeswapspace(object, r1, 1);
2388 /* swap ptr may be invalid */