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 * 4. Neither the name of the University nor the names of its contributors
54 * may be used to endorse or promote products derived from this software
55 * without specific prior written permission.
57 * THIS SOFTWARE IS PROVIDED BY THE REGENTS AND CONTRIBUTORS ``AS IS'' AND
58 * ANY EXPRESS OR IMPLIED WARRANTIES, INCLUDING, BUT NOT LIMITED TO, THE
59 * IMPLIED WARRANTIES OF MERCHANTABILITY AND FITNESS FOR A PARTICULAR PURPOSE
60 * ARE DISCLAIMED. IN NO EVENT SHALL THE REGENTS OR CONTRIBUTORS BE LIABLE
61 * FOR ANY DIRECT, INDIRECT, INCIDENTAL, SPECIAL, EXEMPLARY, OR CONSEQUENTIAL
62 * DAMAGES (INCLUDING, BUT NOT LIMITED TO, PROCUREMENT OF SUBSTITUTE GOODS
63 * OR SERVICES; LOSS OF USE, DATA, OR PROFITS; OR BUSINESS INTERRUPTION)
64 * HOWEVER CAUSED AND ON ANY THEORY OF LIABILITY, WHETHER IN CONTRACT, STRICT
65 * LIABILITY, OR TORT (INCLUDING NEGLIGENCE OR OTHERWISE) ARISING IN ANY WAY
66 * OUT OF THE USE OF THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF
72 * Radix Bitmap 'blists'.
74 * - The new swapper uses the new radix bitmap code. This should scale
75 * to arbitrarily small or arbitrarily large swap spaces and an almost
76 * arbitrary degree of fragmentation.
80 * - on the fly reallocation of swap during putpages. The new system
81 * does not try to keep previously allocated swap blocks for dirty
84 * - on the fly deallocation of swap
86 * - No more garbage collection required. Unnecessarily allocated swap
87 * blocks only exist for dirty vm_page_t's now and these are already
88 * cycled (in a high-load system) by the pager. We also do on-the-fly
89 * removal of invalidated swap blocks when a page is destroyed
92 * from: Utah $Hdr: swap_pager.c 1.4 91/04/30$
93 * @(#)swap_pager.c 8.9 (Berkeley) 3/21/94
94 * $FreeBSD: src/sys/vm/swap_pager.c,v 1.130.2.12 2002/08/31 21:15:55 dillon Exp $
97 #include <sys/param.h>
98 #include <sys/systm.h>
100 #include <sys/kernel.h>
101 #include <sys/proc.h>
103 #include <sys/vnode.h>
104 #include <sys/malloc.h>
105 #include <sys/vmmeter.h>
106 #include <sys/sysctl.h>
107 #include <sys/blist.h>
108 #include <sys/lock.h>
109 #include <sys/thread2.h>
111 #include "opt_swap.h"
113 #include <vm/vm_object.h>
114 #include <vm/vm_page.h>
115 #include <vm/vm_pager.h>
116 #include <vm/vm_pageout.h>
117 #include <vm/swap_pager.h>
118 #include <vm/vm_extern.h>
119 #include <vm/vm_zone.h>
120 #include <vm/vnode_pager.h>
122 #include <sys/buf2.h>
123 #include <vm/vm_page2.h>
125 #ifndef MAX_PAGEOUT_CLUSTER
126 #define MAX_PAGEOUT_CLUSTER SWB_NPAGES
129 #define SWM_FREE 0x02 /* free, period */
130 #define SWM_POP 0x04 /* pop out */
132 #define SWBIO_READ 0x01
133 #define SWBIO_WRITE 0x02
134 #define SWBIO_SYNC 0x04
140 vm_pindex_t endi; /* inclusive */
143 struct swswapoffinfo {
149 * vm_swap_size is in page-sized chunks now. It was DEV_BSIZE'd chunks
153 int swap_pager_full; /* swap space exhaustion (task killing) */
154 int vm_swap_cache_use;
155 int vm_swap_anon_use;
157 static int swap_pager_almost_full; /* swap space exhaustion (w/ hysteresis)*/
158 static int nsw_rcount; /* free read buffers */
159 static int nsw_wcount_sync; /* limit write buffers / synchronous */
160 static int nsw_wcount_async; /* limit write buffers / asynchronous */
161 static int nsw_wcount_async_max;/* assigned maximum */
162 static int nsw_cluster_max; /* maximum VOP I/O allowed */
164 struct blist *swapblist;
165 static int swap_async_max = 4; /* maximum in-progress async I/O's */
166 static int swap_burst_read = 0; /* allow burst reading */
167 static swblk_t swapiterator; /* linearize allocations */
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 object = vm_object_allocate_hold(OBJT_DEFAULT,
427 OFF_TO_IDX(offset + PAGE_MASK + size));
428 swp_pager_meta_convert(object);
429 vm_object_drop(object);
435 * SWAP_PAGER_DEALLOC() - remove swap metadata from object
437 * The swap backing for the object is destroyed. The code is
438 * designed such that we can reinstantiate it later, but this
439 * routine is typically called only when the entire object is
440 * about to be destroyed.
442 * The object must be locked or unreferenceable.
443 * No other requirements.
446 swap_pager_dealloc(vm_object_t object)
448 vm_object_hold(object);
449 vm_object_pip_wait(object, "swpdea");
452 * Free all remaining metadata. We only bother to free it from
453 * the swap meta data. We do not attempt to free swapblk's still
454 * associated with vm_page_t's for this object. We do not care
455 * if paging is still in progress on some objects.
457 swp_pager_meta_free_all(object);
458 vm_object_drop(object);
461 /************************************************************************
462 * SWAP PAGER BITMAP ROUTINES *
463 ************************************************************************/
466 * SWP_PAGER_GETSWAPSPACE() - allocate raw swap space
468 * Allocate swap for the requested number of pages. The starting
469 * swap block number (a page index) is returned or SWAPBLK_NONE
470 * if the allocation failed.
472 * Also has the side effect of advising that somebody made a mistake
473 * when they configured swap and didn't configure enough.
475 * The caller must hold the object.
476 * This routine may not block.
478 static __inline swblk_t
479 swp_pager_getswapspace(vm_object_t object, int npages)
483 lwkt_gettoken(&vm_token);
484 blk = blist_allocat(swapblist, npages, swapiterator);
485 if (blk == SWAPBLK_NONE)
486 blk = blist_allocat(swapblist, npages, 0);
487 if (blk == SWAPBLK_NONE) {
488 if (swap_pager_full != 2) {
489 kprintf("swap_pager_getswapspace: failed alloc=%d\n",
492 swap_pager_almost_full = 1;
496 swapacctspace(blk, -npages);
497 if (object->type == OBJT_SWAP)
498 vm_swap_anon_use += npages;
500 vm_swap_cache_use += npages;
503 lwkt_reltoken(&vm_token);
508 * SWP_PAGER_FREESWAPSPACE() - free raw swap space
510 * This routine returns the specified swap blocks back to the bitmap.
512 * Note: This routine may not block (it could in the old swap code),
513 * and through the use of the new blist routines it does not block.
515 * We must be called at splvm() to avoid races with bitmap frees from
516 * vm_page_remove() aka swap_pager_page_removed().
518 * This routine may not block.
522 swp_pager_freeswapspace(vm_object_t object, swblk_t blk, int npages)
524 struct swdevt *sp = &swdevt[BLK2DEVIDX(blk)];
526 lwkt_gettoken(&vm_token);
527 sp->sw_nused -= npages;
528 if (object->type == OBJT_SWAP)
529 vm_swap_anon_use -= npages;
531 vm_swap_cache_use -= npages;
533 if (sp->sw_flags & SW_CLOSING) {
534 lwkt_reltoken(&vm_token);
538 blist_free(swapblist, blk, npages);
539 vm_swap_size += npages;
541 lwkt_reltoken(&vm_token);
545 * SWAP_PAGER_FREESPACE() - frees swap blocks associated with a page
546 * range within an object.
548 * This is a globally accessible routine.
550 * This routine removes swapblk assignments from swap metadata.
552 * The external callers of this routine typically have already destroyed
553 * or renamed vm_page_t's associated with this range in the object so
559 swap_pager_freespace(vm_object_t object, vm_pindex_t start, vm_pindex_t size)
561 vm_object_hold(object);
562 swp_pager_meta_free(object, start, size);
563 vm_object_drop(object);
570 swap_pager_freespace_all(vm_object_t object)
572 vm_object_hold(object);
573 swp_pager_meta_free_all(object);
574 vm_object_drop(object);
578 * This function conditionally frees swap cache swap starting at
579 * (*basei) in the object. (count) swap blocks will be nominally freed.
580 * The actual number of blocks freed can be more or less than the
583 * This function nominally returns the number of blocks freed. However,
584 * the actual number of blocks freed may be less then the returned value.
585 * If the function is unable to exhaust the object or if it is able to
586 * free (approximately) the requested number of blocks it returns
589 * If we exhaust the object we will return a value n <= count.
591 * The caller must hold the object.
593 * WARNING! If count == 0 then -1 can be returned as a degenerate case,
594 * callers should always pass a count value > 0.
596 static int swap_pager_condfree_callback(struct swblock *swap, void *data);
599 swap_pager_condfree(vm_object_t object, vm_pindex_t *basei, int count)
601 struct swfreeinfo info;
605 ASSERT_LWKT_TOKEN_HELD(vm_object_token(object));
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);
617 * Take the higher difference swblocks vs pages
619 n = count - (int)info.begi;
620 t = count * 8 - (int)info.endi;
629 * The idea is to free whole meta-block to avoid fragmenting
630 * the swap space or disk I/O. We only do this if NO VM pages
633 * We do not have to deal with clearing PG_SWAPPED in related VM
634 * pages because there are no related VM pages.
636 * The caller must hold the object.
639 swap_pager_condfree_callback(struct swblock *swap, void *data)
641 struct swfreeinfo *info = data;
642 vm_object_t object = info->object;
645 for (i = 0; i < SWAP_META_PAGES; ++i) {
646 if (vm_page_lookup(object, swap->swb_index + i))
649 info->basei = swap->swb_index + SWAP_META_PAGES;
650 if (i == SWAP_META_PAGES) {
651 info->begi -= swap->swb_count;
652 swap_pager_freespace(object, swap->swb_index, SWAP_META_PAGES);
655 if ((int)info->begi < 0 || (int)info->endi < 0)
662 * Called by vm_page_alloc() when a new VM page is inserted
663 * into a VM object. Checks whether swap has been assigned to
664 * the page and sets PG_SWAPPED as necessary.
669 swap_pager_page_inserted(vm_page_t m)
671 if (m->object->swblock_count) {
672 vm_object_hold(m->object);
673 if (swp_pager_meta_ctl(m->object, m->pindex, 0) != SWAPBLK_NONE)
674 vm_page_flag_set(m, PG_SWAPPED);
675 vm_object_drop(m->object);
680 * SWAP_PAGER_RESERVE() - reserve swap blocks in object
682 * Assigns swap blocks to the specified range within the object. The
683 * swap blocks are not zerod. Any previous swap assignment is destroyed.
685 * Returns 0 on success, -1 on failure.
687 * The caller is responsible for avoiding races in the specified range.
688 * No other requirements.
691 swap_pager_reserve(vm_object_t object, vm_pindex_t start, vm_size_t size)
694 swblk_t blk = SWAPBLK_NONE;
695 vm_pindex_t beg = start; /* save start index */
697 vm_object_hold(object);
702 while ((blk = swp_pager_getswapspace(object, n)) ==
707 swp_pager_meta_free(object, beg,
709 vm_object_drop(object);
714 swp_pager_meta_build(object, start, blk);
720 swp_pager_meta_free(object, start, n);
721 vm_object_drop(object);
726 * SWAP_PAGER_COPY() - copy blocks from source pager to destination pager
727 * and destroy the source.
729 * Copy any valid swapblks from the source to the destination. In
730 * cases where both the source and destination have a valid swapblk,
731 * we keep the destination's.
733 * This routine is allowed to block. It may block allocating metadata
734 * indirectly through swp_pager_meta_build() or if paging is still in
735 * progress on the source.
737 * XXX vm_page_collapse() kinda expects us not to block because we
738 * supposedly do not need to allocate memory, but for the moment we
739 * *may* have to get a little memory from the zone allocator, but
740 * it is taken from the interrupt memory. We should be ok.
742 * The source object contains no vm_page_t's (which is just as well)
743 * The source object is of type OBJT_SWAP.
745 * The source and destination objects must be held by the caller.
748 swap_pager_copy(vm_object_t srcobject, vm_object_t dstobject,
749 vm_pindex_t base_index, int destroysource)
753 ASSERT_LWKT_TOKEN_HELD(vm_object_token(srcobject));
754 ASSERT_LWKT_TOKEN_HELD(vm_object_token(dstobject));
757 * transfer source to destination.
759 for (i = 0; i < dstobject->size; ++i) {
763 * Locate (without changing) the swapblk on the destination,
764 * unless it is invalid in which case free it silently, or
765 * if the destination is a resident page, in which case the
766 * source is thrown away.
768 dstaddr = swp_pager_meta_ctl(dstobject, i, 0);
770 if (dstaddr == SWAPBLK_NONE) {
772 * Destination has no swapblk and is not resident,
777 srcaddr = swp_pager_meta_ctl(srcobject,
778 base_index + i, SWM_POP);
780 if (srcaddr != SWAPBLK_NONE)
781 swp_pager_meta_build(dstobject, i, srcaddr);
784 * Destination has valid swapblk or it is represented
785 * by a resident page. We destroy the sourceblock.
787 swp_pager_meta_ctl(srcobject, base_index + i, SWM_FREE);
792 * Free left over swap blocks in source.
794 * We have to revert the type to OBJT_DEFAULT so we do not accidently
795 * double-remove the object from the swap queues.
799 * Reverting the type is not necessary, the caller is going
800 * to destroy srcobject directly, but I'm doing it here
801 * for consistency since we've removed the object from its
804 swp_pager_meta_free_all(srcobject);
805 if (srcobject->type == OBJT_SWAP)
806 srcobject->type = OBJT_DEFAULT;
811 * SWAP_PAGER_HASPAGE() - determine if we have good backing store for
812 * the requested page.
814 * We determine whether good backing store exists for the requested
815 * page and return TRUE if it does, FALSE if it doesn't.
817 * If TRUE, we also try to determine how much valid, contiguous backing
818 * store exists before and after the requested page within a reasonable
819 * distance. We do not try to restrict it to the swap device stripe
820 * (that is handled in getpages/putpages). It probably isn't worth
826 swap_pager_haspage(vm_object_t object, vm_pindex_t pindex)
831 * do we have good backing store at the requested index ?
833 vm_object_hold(object);
834 blk0 = swp_pager_meta_ctl(object, pindex, 0);
836 if (blk0 == SWAPBLK_NONE) {
837 vm_object_drop(object);
840 vm_object_drop(object);
845 * SWAP_PAGER_PAGE_UNSWAPPED() - remove swap backing store related to page
847 * This removes any associated swap backing store, whether valid or
848 * not, from the page. This operates on any VM object, not just OBJT_SWAP
851 * This routine is typically called when a page is made dirty, at
852 * which point any associated swap can be freed. MADV_FREE also
853 * calls us in a special-case situation
855 * NOTE!!! If the page is clean and the swap was valid, the caller
856 * should make the page dirty before calling this routine. This routine
857 * does NOT change the m->dirty status of the page. Also: MADV_FREE
860 * The page must be busied or soft-busied.
861 * The caller can hold the object to avoid blocking, else we might block.
862 * No other requirements.
865 swap_pager_unswapped(vm_page_t m)
867 if (m->flags & PG_SWAPPED) {
868 vm_object_hold(m->object);
869 KKASSERT(m->flags & PG_SWAPPED);
870 swp_pager_meta_ctl(m->object, m->pindex, SWM_FREE);
871 vm_page_flag_clear(m, PG_SWAPPED);
872 vm_object_drop(m->object);
877 * SWAP_PAGER_STRATEGY() - read, write, free blocks
879 * This implements a VM OBJECT strategy function using swap backing store.
880 * This can operate on any VM OBJECT type, not necessarily just OBJT_SWAP
883 * This is intended to be a cacheless interface (i.e. caching occurs at
884 * higher levels), and is also used as a swap-based SSD cache for vnode
885 * and device objects.
887 * All I/O goes directly to and from the swap device.
889 * We currently attempt to run I/O synchronously or asynchronously as
890 * the caller requests. This isn't perfect because we loose error
891 * sequencing when we run multiple ops in parallel to satisfy a request.
892 * But this is swap, so we let it all hang out.
897 swap_pager_strategy(vm_object_t object, struct bio *bio)
899 struct buf *bp = bio->bio_buf;
902 vm_pindex_t biox_blkno = 0;
908 struct bio_track *track;
913 * tracking for swapdev vnode I/Os
915 if (bp->b_cmd == BUF_CMD_READ)
916 track = &swapdev_vp->v_track_read;
918 track = &swapdev_vp->v_track_write;
921 if (bp->b_bcount & PAGE_MASK) {
922 bp->b_error = EINVAL;
923 bp->b_flags |= B_ERROR | B_INVAL;
925 kprintf("swap_pager_strategy: bp %p offset %lld size %d, "
926 "not page bounded\n",
927 bp, (long long)bio->bio_offset, (int)bp->b_bcount);
932 * Clear error indication, initialize page index, count, data pointer.
935 bp->b_flags &= ~B_ERROR;
936 bp->b_resid = bp->b_bcount;
938 start = (vm_pindex_t)(bio->bio_offset >> PAGE_SHIFT);
939 count = howmany(bp->b_bcount, PAGE_SIZE);
943 * Deal with BUF_CMD_FREEBLKS
945 if (bp->b_cmd == BUF_CMD_FREEBLKS) {
947 * FREE PAGE(s) - destroy underlying swap that is no longer
950 vm_object_hold(object);
951 swp_pager_meta_free(object, start, count);
952 vm_object_drop(object);
959 * We need to be able to create a new cluster of I/O's. We cannot
960 * use the caller fields of the passed bio so push a new one.
962 * Because nbio is just a placeholder for the cluster links,
963 * we can biodone() the original bio instead of nbio to make
964 * things a bit more efficient.
966 nbio = push_bio(bio);
967 nbio->bio_offset = bio->bio_offset;
968 nbio->bio_caller_info1.cluster_head = NULL;
969 nbio->bio_caller_info2.cluster_tail = NULL;
975 * Execute read or write
977 vm_object_hold(object);
983 * Obtain block. If block not found and writing, allocate a
984 * new block and build it into the object.
986 blk = swp_pager_meta_ctl(object, start, 0);
987 if ((blk == SWAPBLK_NONE) && bp->b_cmd != BUF_CMD_READ) {
988 blk = swp_pager_getswapspace(object, 1);
989 if (blk == SWAPBLK_NONE) {
990 bp->b_error = ENOMEM;
991 bp->b_flags |= B_ERROR;
994 swp_pager_meta_build(object, start, blk);
998 * Do we have to flush our current collection? Yes if:
1000 * - no swap block at this index
1001 * - swap block is not contiguous
1002 * - we cross a physical disk boundry in the
1006 biox && (biox_blkno + btoc(bufx->b_bcount) != blk ||
1007 ((biox_blkno ^ blk) & dmmax_mask)
1010 if (bp->b_cmd == BUF_CMD_READ) {
1011 ++mycpu->gd_cnt.v_swapin;
1012 mycpu->gd_cnt.v_swappgsin += btoc(bufx->b_bcount);
1014 ++mycpu->gd_cnt.v_swapout;
1015 mycpu->gd_cnt.v_swappgsout += btoc(bufx->b_bcount);
1016 bufx->b_dirtyend = bufx->b_bcount;
1020 * Finished with this buf.
1022 KKASSERT(bufx->b_bcount != 0);
1023 if (bufx->b_cmd != BUF_CMD_READ)
1024 bufx->b_dirtyend = bufx->b_bcount;
1030 * Add new swapblk to biox, instantiating biox if necessary.
1031 * Zero-fill reads are able to take a shortcut.
1033 if (blk == SWAPBLK_NONE) {
1035 * We can only get here if we are reading. Since
1036 * we are at splvm() we can safely modify b_resid,
1037 * even if chain ops are in progress.
1039 bzero(data, PAGE_SIZE);
1040 bp->b_resid -= PAGE_SIZE;
1043 /* XXX chain count > 4, wait to <= 4 */
1045 bufx = getpbuf(NULL);
1046 biox = &bufx->b_bio1;
1047 cluster_append(nbio, bufx);
1048 bufx->b_flags |= (bp->b_flags & B_ORDERED);
1049 bufx->b_cmd = bp->b_cmd;
1050 biox->bio_done = swap_chain_iodone;
1051 biox->bio_offset = (off_t)blk << PAGE_SHIFT;
1052 biox->bio_caller_info1.cluster_parent = nbio;
1055 bufx->b_data = data;
1057 bufx->b_bcount += PAGE_SIZE;
1064 vm_object_drop(object);
1067 * Flush out last buffer
1070 if (bufx->b_cmd == BUF_CMD_READ) {
1071 ++mycpu->gd_cnt.v_swapin;
1072 mycpu->gd_cnt.v_swappgsin += btoc(bufx->b_bcount);
1074 ++mycpu->gd_cnt.v_swapout;
1075 mycpu->gd_cnt.v_swappgsout += btoc(bufx->b_bcount);
1076 bufx->b_dirtyend = bufx->b_bcount;
1078 KKASSERT(bufx->b_bcount);
1079 if (bufx->b_cmd != BUF_CMD_READ)
1080 bufx->b_dirtyend = bufx->b_bcount;
1081 /* biox, bufx = NULL */
1085 * Now initiate all the I/O. Be careful looping on our chain as
1086 * I/O's may complete while we are still initiating them.
1088 * If the request is a 100% sparse read no bios will be present
1089 * and we just biodone() the buffer.
1091 nbio->bio_caller_info2.cluster_tail = NULL;
1092 bufx = nbio->bio_caller_info1.cluster_head;
1096 biox = &bufx->b_bio1;
1098 bufx = bufx->b_cluster_next;
1099 vn_strategy(swapdev_vp, biox);
1106 * Completion of the cluster will also call biodone_chain(nbio).
1107 * We never call biodone(nbio) so we don't have to worry about
1108 * setting up a bio_done callback. It's handled in the sub-IO.
1119 swap_chain_iodone(struct bio *biox)
1122 struct buf *bufx; /* chained sub-buffer */
1123 struct bio *nbio; /* parent nbio with chain glue */
1124 struct buf *bp; /* original bp associated with nbio */
1127 bufx = biox->bio_buf;
1128 nbio = biox->bio_caller_info1.cluster_parent;
1132 * Update the original buffer
1134 KKASSERT(bp != NULL);
1135 if (bufx->b_flags & B_ERROR) {
1136 atomic_set_int(&bufx->b_flags, B_ERROR);
1137 bp->b_error = bufx->b_error; /* race ok */
1138 } else if (bufx->b_resid != 0) {
1139 atomic_set_int(&bufx->b_flags, B_ERROR);
1140 bp->b_error = EINVAL; /* race ok */
1142 atomic_subtract_int(&bp->b_resid, bufx->b_bcount);
1146 * Remove us from the chain.
1148 spin_lock(&bp->b_lock.lk_spinlock);
1149 nextp = &nbio->bio_caller_info1.cluster_head;
1150 while (*nextp != bufx) {
1151 KKASSERT(*nextp != NULL);
1152 nextp = &(*nextp)->b_cluster_next;
1154 *nextp = bufx->b_cluster_next;
1155 chain_empty = (nbio->bio_caller_info1.cluster_head == NULL);
1156 spin_unlock(&bp->b_lock.lk_spinlock);
1159 * Clean up bufx. If the chain is now empty we finish out
1160 * the parent. Note that we may be racing other completions
1161 * so we must use the chain_empty status from above.
1164 if (bp->b_resid != 0 && !(bp->b_flags & B_ERROR)) {
1165 atomic_set_int(&bp->b_flags, B_ERROR);
1166 bp->b_error = EINVAL;
1168 biodone_chain(nbio);
1170 relpbuf(bufx, NULL);
1174 * SWAP_PAGER_GETPAGES() - bring page in from swap
1176 * The requested page may have to be brought in from swap. Calculate the
1177 * swap block and bring in additional pages if possible. All pages must
1178 * have contiguous swap block assignments and reside in the same object.
1180 * The caller has a single vm_object_pip_add() reference prior to
1181 * calling us and we should return with the same.
1183 * The caller has BUSY'd the page. We should return with (*mpp) left busy,
1184 * and any additinal pages unbusied.
1186 * If the caller encounters a PG_RAM page it will pass it to us even though
1187 * it may be valid and dirty. We cannot overwrite the page in this case!
1188 * The case is used to allow us to issue pure read-aheads.
1190 * NOTE! XXX This code does not entirely pipeline yet due to the fact that
1191 * the PG_RAM page is validated at the same time as mreq. What we
1192 * really need to do is issue a separate read-ahead pbuf.
1197 swap_pager_getpage(vm_object_t object, vm_page_t *mpp, int seqaccess)
1210 vm_page_t marray[XIO_INTERNAL_PAGES];
1214 vm_object_hold(object);
1215 if (mreq->object != object) {
1216 panic("swap_pager_getpages: object mismatch %p/%p",
1223 * We don't want to overwrite a fully valid page as it might be
1224 * dirty. This case can occur when e.g. vm_fault hits a perfectly
1225 * valid page with PG_RAM set.
1227 * In this case we see if the next page is a suitable page-in
1228 * candidate and if it is we issue read-ahead. PG_RAM will be
1229 * set on the last page of the read-ahead to continue the pipeline.
1231 if (mreq->valid == VM_PAGE_BITS_ALL) {
1232 if (swap_burst_read == 0 || mreq->pindex + 1 >= object->size) {
1233 vm_object_drop(object);
1234 return(VM_PAGER_OK);
1236 blk = swp_pager_meta_ctl(object, mreq->pindex + 1, 0);
1237 if (blk == SWAPBLK_NONE) {
1238 vm_object_drop(object);
1239 return(VM_PAGER_OK);
1241 m = vm_page_lookup_busy_try(object, mreq->pindex + 1,
1244 vm_object_drop(object);
1245 return(VM_PAGER_OK);
1246 } else if (m == NULL) {
1248 * Use VM_ALLOC_QUICK to avoid blocking on cache
1251 m = vm_page_alloc(object, mreq->pindex + 1,
1254 vm_object_drop(object);
1255 return(VM_PAGER_OK);
1260 vm_object_drop(object);
1261 return(VM_PAGER_OK);
1263 vm_page_unqueue_nowakeup(m);
1273 * Try to block-read contiguous pages from swap if sequential,
1274 * otherwise just read one page. Contiguous pages from swap must
1275 * reside within a single device stripe because the I/O cannot be
1276 * broken up across multiple stripes.
1278 * Note that blk and iblk can be SWAPBLK_NONE but the loop is
1279 * set up such that the case(s) are handled implicitly.
1281 blk = swp_pager_meta_ctl(mreq->object, mreq->pindex, 0);
1284 for (i = 1; swap_burst_read &&
1285 i < XIO_INTERNAL_PAGES &&
1286 mreq->pindex + i < object->size; ++i) {
1289 iblk = swp_pager_meta_ctl(object, mreq->pindex + i, 0);
1290 if (iblk != blk + i)
1292 if ((blk ^ iblk) & dmmax_mask)
1294 m = vm_page_lookup_busy_try(object, mreq->pindex + i,
1298 } else if (m == NULL) {
1300 * Use VM_ALLOC_QUICK to avoid blocking on cache
1303 m = vm_page_alloc(object, mreq->pindex + i,
1312 vm_page_unqueue_nowakeup(m);
1318 vm_page_flag_set(marray[i - 1], PG_RAM);
1321 * If mreq is the requested page and we have nothing to do return
1322 * VM_PAGER_FAIL. If raonly is set mreq is just another read-ahead
1323 * page and must be cleaned up.
1325 if (blk == SWAPBLK_NONE) {
1328 vnode_pager_freepage(mreq);
1329 vm_object_drop(object);
1330 return(VM_PAGER_OK);
1332 vm_object_drop(object);
1333 return(VM_PAGER_FAIL);
1338 * map our page(s) into kva for input
1340 bp = getpbuf_kva(&nsw_rcount);
1342 kva = (vm_offset_t) bp->b_kvabase;
1343 bcopy(marray, bp->b_xio.xio_pages, i * sizeof(vm_page_t));
1344 pmap_qenter(kva, bp->b_xio.xio_pages, i);
1346 bp->b_data = (caddr_t)kva;
1347 bp->b_bcount = PAGE_SIZE * i;
1348 bp->b_xio.xio_npages = i;
1349 bio->bio_done = swp_pager_async_iodone;
1350 bio->bio_offset = (off_t)blk << PAGE_SHIFT;
1351 bio->bio_caller_info1.index = SWBIO_READ;
1354 * Set index. If raonly set the index beyond the array so all
1355 * the pages are treated the same, otherwise the original mreq is
1359 bio->bio_driver_info = (void *)(intptr_t)i;
1361 bio->bio_driver_info = (void *)(intptr_t)0;
1363 for (j = 0; j < i; ++j)
1364 vm_page_flag_set(bp->b_xio.xio_pages[j], PG_SWAPINPROG);
1366 mycpu->gd_cnt.v_swapin++;
1367 mycpu->gd_cnt.v_swappgsin += bp->b_xio.xio_npages;
1370 * We still hold the lock on mreq, and our automatic completion routine
1371 * does not remove it.
1373 vm_object_pip_add(object, bp->b_xio.xio_npages);
1376 * perform the I/O. NOTE!!! bp cannot be considered valid after
1377 * this point because we automatically release it on completion.
1378 * Instead, we look at the one page we are interested in which we
1379 * still hold a lock on even through the I/O completion.
1381 * The other pages in our m[] array are also released on completion,
1382 * so we cannot assume they are valid anymore either.
1384 bp->b_cmd = BUF_CMD_READ;
1386 vn_strategy(swapdev_vp, bio);
1389 * Wait for the page we want to complete. PG_SWAPINPROG is always
1390 * cleared on completion. If an I/O error occurs, SWAPBLK_NONE
1391 * is set in the meta-data.
1393 * If this is a read-ahead only we return immediately without
1397 vm_object_drop(object);
1398 return(VM_PAGER_OK);
1402 * Read-ahead includes originally requested page case.
1405 flags = mreq->flags;
1407 if ((flags & PG_SWAPINPROG) == 0)
1409 tsleep_interlock(mreq, 0);
1410 if (!atomic_cmpset_int(&mreq->flags, flags,
1411 flags | PG_WANTED | PG_REFERENCED)) {
1414 mycpu->gd_cnt.v_intrans++;
1415 if (tsleep(mreq, PINTERLOCKED, "swread", hz*20)) {
1417 "swap_pager: indefinite wait buffer: "
1418 " offset: %lld, size: %ld\n",
1419 (long long)bio->bio_offset,
1426 * mreq is left bussied after completion, but all the other pages
1427 * are freed. If we had an unrecoverable read error the page will
1430 vm_object_drop(object);
1431 if (mreq->valid != VM_PAGE_BITS_ALL)
1432 return(VM_PAGER_ERROR);
1434 return(VM_PAGER_OK);
1437 * A final note: in a low swap situation, we cannot deallocate swap
1438 * and mark a page dirty here because the caller is likely to mark
1439 * the page clean when we return, causing the page to possibly revert
1440 * to all-zero's later.
1445 * swap_pager_putpages:
1447 * Assign swap (if necessary) and initiate I/O on the specified pages.
1449 * We support both OBJT_DEFAULT and OBJT_SWAP objects. DEFAULT objects
1450 * are automatically converted to SWAP objects.
1452 * In a low memory situation we may block in vn_strategy(), but the new
1453 * vm_page reservation system coupled with properly written VFS devices
1454 * should ensure that no low-memory deadlock occurs. This is an area
1457 * The parent has N vm_object_pip_add() references prior to
1458 * calling us and will remove references for rtvals[] that are
1459 * not set to VM_PAGER_PEND. We need to remove the rest on I/O
1462 * The parent has soft-busy'd the pages it passes us and will unbusy
1463 * those whos rtvals[] entry is not set to VM_PAGER_PEND on return.
1464 * We need to unbusy the rest on I/O completion.
1469 swap_pager_putpages(vm_object_t object, vm_page_t *m, int count,
1470 boolean_t sync, int *rtvals)
1475 vm_object_hold(object);
1477 if (count && m[0]->object != object) {
1478 panic("swap_pager_getpages: object mismatch %p/%p",
1487 * Turn object into OBJT_SWAP
1488 * check for bogus sysops
1489 * force sync if not pageout process
1491 if (object->type == OBJT_DEFAULT) {
1492 if (object->type == OBJT_DEFAULT)
1493 swp_pager_meta_convert(object);
1496 if (curthread != pagethread)
1502 * Update nsw parameters from swap_async_max sysctl values.
1503 * Do not let the sysop crash the machine with bogus numbers.
1505 if (swap_async_max != nsw_wcount_async_max) {
1511 if ((n = swap_async_max) > nswbuf / 2)
1518 * Adjust difference ( if possible ). If the current async
1519 * count is too low, we may not be able to make the adjustment
1522 * vm_token needed for nsw_wcount sleep interlock
1524 lwkt_gettoken(&vm_token);
1525 n -= nsw_wcount_async_max;
1526 if (nsw_wcount_async + n >= 0) {
1527 nsw_wcount_async_max += n;
1528 pbuf_adjcount(&nsw_wcount_async, n);
1530 lwkt_reltoken(&vm_token);
1536 * Assign swap blocks and issue I/O. We reallocate swap on the fly.
1537 * The page is left dirty until the pageout operation completes
1541 for (i = 0; i < count; i += n) {
1548 * Maximum I/O size is limited by a number of factors.
1551 n = min(BLIST_MAX_ALLOC, count - i);
1552 n = min(n, nsw_cluster_max);
1554 lwkt_gettoken(&vm_token);
1557 * Get biggest block of swap we can. If we fail, fall
1558 * back and try to allocate a smaller block. Don't go
1559 * overboard trying to allocate space if it would overly
1563 (blk = swp_pager_getswapspace(object, n)) == SWAPBLK_NONE &&
1568 if (blk == SWAPBLK_NONE) {
1569 for (j = 0; j < n; ++j)
1570 rtvals[i+j] = VM_PAGER_FAIL;
1571 lwkt_reltoken(&vm_token);
1576 * The I/O we are constructing cannot cross a physical
1577 * disk boundry in the swap stripe. Note: we are still
1580 if ((blk ^ (blk + n)) & dmmax_mask) {
1581 j = ((blk + dmmax) & dmmax_mask) - blk;
1582 swp_pager_freeswapspace(object, blk + j, n - j);
1587 * All I/O parameters have been satisfied, build the I/O
1588 * request and assign the swap space.
1591 bp = getpbuf_kva(&nsw_wcount_sync);
1593 bp = getpbuf_kva(&nsw_wcount_async);
1596 lwkt_reltoken(&vm_token);
1598 pmap_qenter((vm_offset_t)bp->b_data, &m[i], n);
1600 bp->b_bcount = PAGE_SIZE * n;
1601 bio->bio_offset = (off_t)blk << PAGE_SHIFT;
1603 for (j = 0; j < n; ++j) {
1604 vm_page_t mreq = m[i+j];
1606 swp_pager_meta_build(mreq->object, mreq->pindex,
1608 if (object->type == OBJT_SWAP)
1609 vm_page_dirty(mreq);
1610 rtvals[i+j] = VM_PAGER_OK;
1612 vm_page_flag_set(mreq, PG_SWAPINPROG);
1613 bp->b_xio.xio_pages[j] = mreq;
1615 bp->b_xio.xio_npages = n;
1617 mycpu->gd_cnt.v_swapout++;
1618 mycpu->gd_cnt.v_swappgsout += bp->b_xio.xio_npages;
1620 bp->b_dirtyoff = 0; /* req'd for NFS */
1621 bp->b_dirtyend = bp->b_bcount; /* req'd for NFS */
1622 bp->b_cmd = BUF_CMD_WRITE;
1623 bio->bio_caller_info1.index = SWBIO_WRITE;
1628 if (sync == FALSE) {
1629 bio->bio_done = swp_pager_async_iodone;
1631 vn_strategy(swapdev_vp, bio);
1633 for (j = 0; j < n; ++j)
1634 rtvals[i+j] = VM_PAGER_PEND;
1639 * Issue synchrnously.
1641 * Wait for the sync I/O to complete, then update rtvals.
1642 * We just set the rtvals[] to VM_PAGER_PEND so we can call
1643 * our async completion routine at the end, thus avoiding a
1646 bio->bio_caller_info1.index |= SWBIO_SYNC;
1647 bio->bio_done = biodone_sync;
1648 bio->bio_flags |= BIO_SYNC;
1649 vn_strategy(swapdev_vp, bio);
1650 biowait(bio, "swwrt");
1652 for (j = 0; j < n; ++j)
1653 rtvals[i+j] = VM_PAGER_PEND;
1656 * Now that we are through with the bp, we can call the
1657 * normal async completion, which frees everything up.
1659 swp_pager_async_iodone(bio);
1661 vm_object_drop(object);
1668 swap_pager_newswap(void)
1674 * swp_pager_async_iodone:
1676 * Completion routine for asynchronous reads and writes from/to swap.
1677 * Also called manually by synchronous code to finish up a bp.
1679 * For READ operations, the pages are PG_BUSY'd. For WRITE operations,
1680 * the pages are vm_page_t->busy'd. For READ operations, we PG_BUSY
1681 * unbusy all pages except the 'main' request page. For WRITE
1682 * operations, we vm_page_t->busy'd unbusy all pages ( we can do this
1683 * because we marked them all VM_PAGER_PEND on return from putpages ).
1685 * This routine may not block.
1690 swp_pager_async_iodone(struct bio *bio)
1692 struct buf *bp = bio->bio_buf;
1693 vm_object_t object = NULL;
1700 if (bp->b_flags & B_ERROR) {
1702 "swap_pager: I/O error - %s failed; offset %lld,"
1703 "size %ld, error %d\n",
1704 ((bio->bio_caller_info1.index & SWBIO_READ) ?
1705 "pagein" : "pageout"),
1706 (long long)bio->bio_offset,
1713 * set object, raise to splvm().
1715 if (bp->b_xio.xio_npages)
1716 object = bp->b_xio.xio_pages[0]->object;
1719 * remove the mapping for kernel virtual
1721 pmap_qremove((vm_offset_t)bp->b_data, bp->b_xio.xio_npages);
1724 * cleanup pages. If an error occurs writing to swap, we are in
1725 * very serious trouble. If it happens to be a disk error, though,
1726 * we may be able to recover by reassigning the swap later on. So
1727 * in this case we remove the m->swapblk assignment for the page
1728 * but do not free it in the rlist. The errornous block(s) are thus
1729 * never reallocated as swap. Redirty the page and continue.
1731 for (i = 0; i < bp->b_xio.xio_npages; ++i) {
1732 vm_page_t m = bp->b_xio.xio_pages[i];
1734 if (bp->b_flags & B_ERROR) {
1736 * If an error occurs I'd love to throw the swapblk
1737 * away without freeing it back to swapspace, so it
1738 * can never be used again. But I can't from an
1742 if (bio->bio_caller_info1.index & SWBIO_READ) {
1744 * When reading, reqpage needs to stay
1745 * locked for the parent, but all other
1746 * pages can be freed. We still want to
1747 * wakeup the parent waiting on the page,
1748 * though. ( also: pg_reqpage can be -1 and
1749 * not match anything ).
1751 * We have to wake specifically requested pages
1752 * up too because we cleared PG_SWAPINPROG and
1753 * someone may be waiting for that.
1755 * NOTE: for reads, m->dirty will probably
1756 * be overridden by the original caller of
1757 * getpages so don't play cute tricks here.
1759 * NOTE: We can't actually free the page from
1760 * here, because this is an interrupt. It
1761 * is not legal to mess with object->memq
1762 * from an interrupt. Deactivate the page
1767 vm_page_flag_clear(m, PG_ZERO);
1768 vm_page_flag_clear(m, PG_SWAPINPROG);
1771 * bio_driver_info holds the requested page
1774 if (i != (int)(intptr_t)bio->bio_driver_info) {
1775 vm_page_deactivate(m);
1781 * If i == bp->b_pager.pg_reqpage, do not wake
1782 * the page up. The caller needs to.
1786 * If a write error occurs remove the swap
1787 * assignment (note that PG_SWAPPED may or
1788 * may not be set depending on prior activity).
1790 * Re-dirty OBJT_SWAP pages as there is no
1791 * other backing store, we can't throw the
1794 * Non-OBJT_SWAP pages (aka swapcache) must
1795 * not be dirtied since they may not have
1796 * been dirty in the first place, and they
1797 * do have backing store (the vnode).
1799 vm_page_busy_wait(m, FALSE, "swadpg");
1800 swp_pager_meta_ctl(m->object, m->pindex,
1802 vm_page_flag_clear(m, PG_SWAPPED);
1803 if (m->object->type == OBJT_SWAP) {
1805 vm_page_activate(m);
1807 vm_page_flag_clear(m, PG_SWAPINPROG);
1808 vm_page_io_finish(m);
1811 } else if (bio->bio_caller_info1.index & SWBIO_READ) {
1813 * NOTE: for reads, m->dirty will probably be
1814 * overridden by the original caller of getpages so
1815 * we cannot set them in order to free the underlying
1816 * swap in a low-swap situation. I don't think we'd
1817 * want to do that anyway, but it was an optimization
1818 * that existed in the old swapper for a time before
1819 * it got ripped out due to precisely this problem.
1821 * clear PG_ZERO in page.
1823 * If not the requested page then deactivate it.
1825 * Note that the requested page, reqpage, is left
1826 * busied, but we still have to wake it up. The
1827 * other pages are released (unbusied) by
1828 * vm_page_wakeup(). We do not set reqpage's
1829 * valid bits here, it is up to the caller.
1833 * NOTE: can't call pmap_clear_modify(m) from an
1834 * interrupt thread, the pmap code may have to map
1835 * non-kernel pmaps and currently asserts the case.
1837 /*pmap_clear_modify(m);*/
1838 m->valid = VM_PAGE_BITS_ALL;
1840 vm_page_flag_clear(m, PG_ZERO | PG_SWAPINPROG);
1841 vm_page_flag_set(m, PG_SWAPPED);
1844 * We have to wake specifically requested pages
1845 * up too because we cleared PG_SWAPINPROG and
1846 * could be waiting for it in getpages. However,
1847 * be sure to not unbusy getpages specifically
1848 * requested page - getpages expects it to be
1851 * bio_driver_info holds the requested page
1853 if (i != (int)(intptr_t)bio->bio_driver_info) {
1854 vm_page_deactivate(m);
1861 * Mark the page clean but do not mess with the
1862 * pmap-layer's modified state. That state should
1863 * also be clear since the caller protected the
1864 * page VM_PROT_READ, but allow the case.
1866 * We are in an interrupt, avoid pmap operations.
1868 * If we have a severe page deficit, deactivate the
1869 * page. Do not try to cache it (which would also
1870 * involve a pmap op), because the page might still
1873 * When using the swap to cache clean vnode pages
1874 * we do not mess with the page dirty bits.
1876 vm_page_busy_wait(m, FALSE, "swadpg");
1877 if (m->object->type == OBJT_SWAP)
1879 vm_page_flag_clear(m, PG_SWAPINPROG);
1880 vm_page_flag_set(m, PG_SWAPPED);
1881 if (vm_page_count_severe())
1882 vm_page_deactivate(m);
1884 if (!vm_page_count_severe() || !vm_page_try_to_cache(m))
1885 vm_page_protect(m, VM_PROT_READ);
1887 vm_page_io_finish(m);
1893 * adjust pip. NOTE: the original parent may still have its own
1894 * pip refs on the object.
1898 vm_object_pip_wakeup_n(object, bp->b_xio.xio_npages);
1901 * Release the physical I/O buffer.
1903 * NOTE: Due to synchronous operations in the write case b_cmd may
1904 * already be set to BUF_CMD_DONE and BIO_SYNC may have already
1907 * Use vm_token to interlock nsw_rcount/wcount wakeup?
1909 lwkt_gettoken(&vm_token);
1910 if (bio->bio_caller_info1.index & SWBIO_READ)
1911 nswptr = &nsw_rcount;
1912 else if (bio->bio_caller_info1.index & SWBIO_SYNC)
1913 nswptr = &nsw_wcount_sync;
1915 nswptr = &nsw_wcount_async;
1916 bp->b_cmd = BUF_CMD_DONE;
1917 relpbuf(bp, nswptr);
1918 lwkt_reltoken(&vm_token);
1922 * Fault-in a potentially swapped page and remove the swap reference.
1923 * (used by swapoff code)
1925 * object must be held.
1927 static __inline void
1928 swp_pager_fault_page(vm_object_t object, vm_pindex_t pindex)
1934 ASSERT_LWKT_TOKEN_HELD(vm_object_token(object));
1936 if (object->type == OBJT_VNODE) {
1938 * Any swap related to a vnode is due to swapcache. We must
1939 * vget() the vnode in case it is not active (otherwise
1940 * vref() will panic). Calling vm_object_page_remove() will
1941 * ensure that any swap ref is removed interlocked with the
1942 * page. clean_only is set to TRUE so we don't throw away
1945 vp = object->handle;
1946 error = vget(vp, LK_SHARED | LK_RETRY | LK_CANRECURSE);
1948 vm_object_page_remove(object, pindex, pindex + 1, TRUE);
1953 * Otherwise it is a normal OBJT_SWAP object and we can
1954 * fault the page in and remove the swap.
1956 m = vm_fault_object_page(object, IDX_TO_OFF(pindex),
1958 VM_FAULT_DIRTY | VM_FAULT_UNSWAP,
1966 * This removes all swap blocks related to a particular device. We have
1967 * to be careful of ripups during the scan.
1969 static int swp_pager_swapoff_callback(struct swblock *swap, void *data);
1972 swap_pager_swapoff(int devidx)
1974 struct vm_object marker;
1976 struct swswapoffinfo info;
1978 bzero(&marker, sizeof(marker));
1979 marker.type = OBJT_MARKER;
1981 lwkt_gettoken(&vmobj_token);
1982 TAILQ_INSERT_HEAD(&vm_object_list, &marker, object_list);
1984 while ((object = TAILQ_NEXT(&marker, object_list)) != NULL) {
1985 if (object->type == OBJT_MARKER)
1987 if (object->type != OBJT_SWAP && object->type != OBJT_VNODE)
1989 vm_object_hold(object);
1990 if (object->type != OBJT_SWAP && object->type != OBJT_VNODE) {
1991 vm_object_drop(object);
1994 info.object = object;
1995 info.devidx = devidx;
1996 swblock_rb_tree_RB_SCAN(&object->swblock_root,
1998 swp_pager_swapoff_callback,
2000 vm_object_drop(object);
2002 if (object == TAILQ_NEXT(&marker, object_list)) {
2003 TAILQ_REMOVE(&vm_object_list, &marker, object_list);
2004 TAILQ_INSERT_AFTER(&vm_object_list, object,
2005 &marker, object_list);
2008 TAILQ_REMOVE(&vm_object_list, &marker, object_list);
2009 lwkt_reltoken(&vmobj_token);
2012 * If we fail to locate all swblocks we just fail gracefully and
2013 * do not bother to restore paging on the swap device. If the
2014 * user wants to retry the user can retry.
2016 if (swdevt[devidx].sw_nused)
2024 swp_pager_swapoff_callback(struct swblock *swap, void *data)
2026 struct swswapoffinfo *info = data;
2027 vm_object_t object = info->object;
2032 index = swap->swb_index;
2033 for (i = 0; i < SWAP_META_PAGES; ++i) {
2035 * Make sure we don't race a dying object. This will
2036 * kill the scan of the object's swap blocks entirely.
2038 if (object->flags & OBJ_DEAD)
2042 * Fault the page, which can obviously block. If the swap
2043 * structure disappears break out.
2045 v = swap->swb_pages[i];
2046 if (v != SWAPBLK_NONE && BLK2DEVIDX(v) == info->devidx) {
2047 swp_pager_fault_page(object, swap->swb_index + i);
2048 /* swap ptr might go away */
2049 if (RB_LOOKUP(swblock_rb_tree,
2050 &object->swblock_root, index) != swap) {
2058 /************************************************************************
2060 ************************************************************************
2062 * These routines manipulate the swap metadata stored in the
2063 * OBJT_SWAP object. All swp_*() routines must be called at
2064 * splvm() because swap can be freed up by the low level vm_page
2065 * code which might be called from interrupts beyond what splbio() covers.
2067 * Swap metadata is implemented with a global hash and not directly
2068 * linked into the object. Instead the object simply contains
2069 * appropriate tracking counters.
2073 * Lookup the swblock containing the specified swap block index.
2075 * The caller must hold the object.
2079 swp_pager_lookup(vm_object_t object, vm_pindex_t index)
2081 ASSERT_LWKT_TOKEN_HELD(vm_object_token(object));
2082 index &= ~(vm_pindex_t)SWAP_META_MASK;
2083 return (RB_LOOKUP(swblock_rb_tree, &object->swblock_root, index));
2087 * Remove a swblock from the RB tree.
2089 * The caller must hold the object.
2093 swp_pager_remove(vm_object_t object, struct swblock *swap)
2095 ASSERT_LWKT_TOKEN_HELD(vm_object_token(object));
2096 RB_REMOVE(swblock_rb_tree, &object->swblock_root, swap);
2100 * Convert default object to swap object if necessary
2102 * The caller must hold the object.
2105 swp_pager_meta_convert(vm_object_t object)
2107 if (object->type == OBJT_DEFAULT) {
2108 object->type = OBJT_SWAP;
2109 KKASSERT(object->swblock_count == 0);
2114 * SWP_PAGER_META_BUILD() - add swap block to swap meta data for object
2116 * We first convert the object to a swap object if it is a default
2117 * object. Vnode objects do not need to be converted.
2119 * The specified swapblk is added to the object's swap metadata. If
2120 * the swapblk is not valid, it is freed instead. Any previously
2121 * assigned swapblk is freed.
2123 * The caller must hold the object.
2126 swp_pager_meta_build(vm_object_t object, vm_pindex_t index, swblk_t swapblk)
2128 struct swblock *swap;
2129 struct swblock *oswap;
2132 KKASSERT(swapblk != SWAPBLK_NONE);
2133 ASSERT_LWKT_TOKEN_HELD(vm_object_token(object));
2136 * Convert object if necessary
2138 if (object->type == OBJT_DEFAULT)
2139 swp_pager_meta_convert(object);
2142 * Locate swblock. If not found create, but if we aren't adding
2143 * anything just return. If we run out of space in the map we wait
2144 * and, since the hash table may have changed, retry.
2147 swap = swp_pager_lookup(object, index);
2152 swap = zalloc(swap_zone);
2157 swap->swb_index = index & ~(vm_pindex_t)SWAP_META_MASK;
2158 swap->swb_count = 0;
2160 ++object->swblock_count;
2162 for (i = 0; i < SWAP_META_PAGES; ++i)
2163 swap->swb_pages[i] = SWAPBLK_NONE;
2164 oswap = RB_INSERT(swblock_rb_tree, &object->swblock_root, swap);
2165 KKASSERT(oswap == NULL);
2169 * Delete prior contents of metadata.
2171 * NOTE: Decrement swb_count after the freeing operation (which
2172 * might block) to prevent racing destruction of the swblock.
2174 index &= SWAP_META_MASK;
2176 while ((v = swap->swb_pages[index]) != SWAPBLK_NONE) {
2177 swap->swb_pages[index] = SWAPBLK_NONE;
2179 swp_pager_freeswapspace(object, v, 1);
2184 * Enter block into metadata
2186 swap->swb_pages[index] = swapblk;
2187 if (swapblk != SWAPBLK_NONE)
2192 * SWP_PAGER_META_FREE() - free a range of blocks in the object's swap metadata
2194 * The requested range of blocks is freed, with any associated swap
2195 * returned to the swap bitmap.
2197 * This routine will free swap metadata structures as they are cleaned
2198 * out. This routine does *NOT* operate on swap metadata associated
2199 * with resident pages.
2201 * The caller must hold the object.
2203 static int swp_pager_meta_free_callback(struct swblock *swb, void *data);
2206 swp_pager_meta_free(vm_object_t object, vm_pindex_t index, vm_pindex_t count)
2208 struct swfreeinfo info;
2210 ASSERT_LWKT_TOKEN_HELD(vm_object_token(object));
2215 if (object->swblock_count == 0) {
2216 KKASSERT(RB_EMPTY(&object->swblock_root));
2223 * Setup for RB tree scan. Note that the pindex range can be huge
2224 * due to the 64 bit page index space so we cannot safely iterate.
2226 info.object = object;
2227 info.basei = index & ~(vm_pindex_t)SWAP_META_MASK;
2229 info.endi = index + count - 1;
2230 swblock_rb_tree_RB_SCAN(&object->swblock_root, rb_swblock_scancmp,
2231 swp_pager_meta_free_callback, &info);
2235 * The caller must hold the object.
2239 swp_pager_meta_free_callback(struct swblock *swap, void *data)
2241 struct swfreeinfo *info = data;
2242 vm_object_t object = info->object;
2247 * Figure out the range within the swblock. The wider scan may
2248 * return edge-case swap blocks when the start and/or end points
2249 * are in the middle of a block.
2251 if (swap->swb_index < info->begi)
2252 index = (int)info->begi & SWAP_META_MASK;
2256 if (swap->swb_index + SWAP_META_PAGES > info->endi)
2257 eindex = (int)info->endi & SWAP_META_MASK;
2259 eindex = SWAP_META_MASK;
2262 * Scan and free the blocks. The loop terminates early
2263 * if (swap) runs out of blocks and could be freed.
2265 * NOTE: Decrement swb_count after swp_pager_freeswapspace()
2266 * to deal with a zfree race.
2268 while (index <= eindex) {
2269 swblk_t v = swap->swb_pages[index];
2271 if (v != SWAPBLK_NONE) {
2272 swap->swb_pages[index] = SWAPBLK_NONE;
2274 swp_pager_freeswapspace(object, v, 1);
2275 if (--swap->swb_count == 0) {
2276 swp_pager_remove(object, swap);
2277 zfree(swap_zone, swap);
2278 --object->swblock_count;
2285 /* swap may be invalid here due to zfree above */
2292 * SWP_PAGER_META_FREE_ALL() - destroy all swap metadata associated with object
2294 * This routine locates and destroys all swap metadata associated with
2297 * NOTE: Decrement swb_count after the freeing operation (which
2298 * might block) to prevent racing destruction of the swblock.
2300 * The caller must hold the object.
2303 swp_pager_meta_free_all(vm_object_t object)
2305 struct swblock *swap;
2308 ASSERT_LWKT_TOKEN_HELD(vm_object_token(object));
2310 while ((swap = RB_ROOT(&object->swblock_root)) != NULL) {
2311 swp_pager_remove(object, swap);
2312 for (i = 0; i < SWAP_META_PAGES; ++i) {
2313 swblk_t v = swap->swb_pages[i];
2314 if (v != SWAPBLK_NONE) {
2316 swp_pager_freeswapspace(object, v, 1);
2320 if (swap->swb_count != 0)
2321 panic("swap_pager_meta_free_all: swb_count != 0");
2322 zfree(swap_zone, swap);
2323 --object->swblock_count;
2326 KKASSERT(object->swblock_count == 0);
2330 * SWP_PAGER_METACTL() - misc control of swap and vm_page_t meta data.
2332 * This routine is capable of looking up, popping, or freeing
2333 * swapblk assignments in the swap meta data or in the vm_page_t.
2334 * The routine typically returns the swapblk being looked-up, or popped,
2335 * or SWAPBLK_NONE if the block was freed, or SWAPBLK_NONE if the block
2336 * was invalid. This routine will automatically free any invalid
2337 * meta-data swapblks.
2339 * It is not possible to store invalid swapblks in the swap meta data
2340 * (other then a literal 'SWAPBLK_NONE'), so we don't bother checking.
2342 * When acting on a busy resident page and paging is in progress, we
2343 * have to wait until paging is complete but otherwise can act on the
2346 * SWM_FREE remove and free swap block from metadata
2347 * SWM_POP remove from meta data but do not free.. pop it out
2349 * The caller must hold the object.
2352 swp_pager_meta_ctl(vm_object_t object, vm_pindex_t index, int flags)
2354 struct swblock *swap;
2357 if (object->swblock_count == 0)
2358 return(SWAPBLK_NONE);
2361 swap = swp_pager_lookup(object, index);
2364 index &= SWAP_META_MASK;
2365 r1 = swap->swb_pages[index];
2367 if (r1 != SWAPBLK_NONE) {
2368 if (flags & (SWM_FREE|SWM_POP)) {
2369 swap->swb_pages[index] = SWAPBLK_NONE;
2370 if (--swap->swb_count == 0) {
2371 swp_pager_remove(object, swap);
2372 zfree(swap_zone, swap);
2373 --object->swblock_count;
2376 /* swap ptr may be invalid */
2377 if (flags & SWM_FREE) {
2378 swp_pager_freeswapspace(object, r1, 1);
2382 /* swap ptr may be invalid */