4 * Copyright (c) 1998-2010 The DragonFly Project. All rights reserved.
6 * This code is derived from software contributed to The DragonFly Project
7 * by Matthew Dillon <dillon@backplane.com>
9 * Redistribution and use in source and binary forms, with or without
10 * modification, are permitted provided that the following conditions
13 * 1. Redistributions of source code must retain the above copyright
14 * notice, this list of conditions and the following disclaimer.
15 * 2. Redistributions in binary form must reproduce the above copyright
16 * notice, this list of conditions and the following disclaimer in
17 * the documentation and/or other materials provided with the
19 * 3. Neither the name of The DragonFly Project nor the names of its
20 * contributors may be used to endorse or promote products derived
21 * from this software without specific, prior written permission.
23 * THIS SOFTWARE IS PROVIDED BY THE COPYRIGHT HOLDERS AND CONTRIBUTORS
24 * ``AS IS'' AND ANY EXPRESS OR IMPLIED WARRANTIES, INCLUDING, BUT NOT
25 * LIMITED TO, THE IMPLIED WARRANTIES OF MERCHANTABILITY AND FITNESS
26 * FOR A PARTICULAR PURPOSE ARE DISCLAIMED. IN NO EVENT SHALL THE
27 * COPYRIGHT HOLDERS OR CONTRIBUTORS BE LIABLE FOR ANY DIRECT, INDIRECT,
28 * INCIDENTAL, SPECIAL, EXEMPLARY OR CONSEQUENTIAL DAMAGES (INCLUDING,
29 * BUT NOT LIMITED TO, PROCUREMENT OF SUBSTITUTE GOODS OR SERVICES;
30 * LOSS OF USE, DATA, OR PROFITS; OR BUSINESS INTERRUPTION) HOWEVER CAUSED
31 * AND ON ANY THEORY OF LIABILITY, WHETHER IN CONTRACT, STRICT LIABILITY,
32 * OR TORT (INCLUDING NEGLIGENCE OR OTHERWISE) ARISING IN ANY WAY OUT
33 * OF THE USE OF THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF
36 * Copyright (c) 1994 John S. Dyson
37 * Copyright (c) 1990 University of Utah.
38 * Copyright (c) 1991, 1993
39 * The Regents of the University of California. All rights reserved.
41 * This code is derived from software contributed to Berkeley by
42 * the Systems Programming Group of the University of Utah Computer
45 * Redistribution and use in source and binary forms, with or without
46 * modification, are permitted provided that the following conditions
48 * 1. Redistributions of source code must retain the above copyright
49 * notice, this list of conditions and the following disclaimer.
50 * 2. Redistributions in binary form must reproduce the above copyright
51 * notice, this list of conditions and the following disclaimer in the
52 * documentation and/or other materials provided with the distribution.
53 * 3. All advertising materials mentioning features or use of this software
54 * must display the following acknowledgement:
55 * This product includes software developed by the University of
56 * California, Berkeley and its contributors.
57 * 4. Neither the name of the University nor the names of its contributors
58 * may be used to endorse or promote products derived from this software
59 * without specific prior written permission.
61 * THIS SOFTWARE IS PROVIDED BY THE REGENTS AND CONTRIBUTORS ``AS IS'' AND
62 * ANY EXPRESS OR IMPLIED WARRANTIES, INCLUDING, BUT NOT LIMITED TO, THE
63 * IMPLIED WARRANTIES OF MERCHANTABILITY AND FITNESS FOR A PARTICULAR PURPOSE
64 * ARE DISCLAIMED. IN NO EVENT SHALL THE REGENTS OR CONTRIBUTORS BE LIABLE
65 * FOR ANY DIRECT, INDIRECT, INCIDENTAL, SPECIAL, EXEMPLARY, OR CONSEQUENTIAL
66 * DAMAGES (INCLUDING, BUT NOT LIMITED TO, PROCUREMENT OF SUBSTITUTE GOODS
67 * OR SERVICES; LOSS OF USE, DATA, OR PROFITS; OR BUSINESS INTERRUPTION)
68 * HOWEVER CAUSED AND ON ANY THEORY OF LIABILITY, WHETHER IN CONTRACT, STRICT
69 * LIABILITY, OR TORT (INCLUDING NEGLIGENCE OR OTHERWISE) ARISING IN ANY WAY
70 * OUT OF THE USE OF THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF
76 * Radix Bitmap 'blists'.
78 * - The new swapper uses the new radix bitmap code. This should scale
79 * to arbitrarily small or arbitrarily large swap spaces and an almost
80 * arbitrary degree of fragmentation.
84 * - on the fly reallocation of swap during putpages. The new system
85 * does not try to keep previously allocated swap blocks for dirty
88 * - on the fly deallocation of swap
90 * - No more garbage collection required. Unnecessarily allocated swap
91 * blocks only exist for dirty vm_page_t's now and these are already
92 * cycled (in a high-load system) by the pager. We also do on-the-fly
93 * removal of invalidated swap blocks when a page is destroyed
96 * from: Utah $Hdr: swap_pager.c 1.4 91/04/30$
97 * @(#)swap_pager.c 8.9 (Berkeley) 3/21/94
98 * $FreeBSD: src/sys/vm/swap_pager.c,v 1.130.2.12 2002/08/31 21:15:55 dillon Exp $
101 #include <sys/param.h>
102 #include <sys/systm.h>
103 #include <sys/conf.h>
104 #include <sys/kernel.h>
105 #include <sys/proc.h>
107 #include <sys/vnode.h>
108 #include <sys/malloc.h>
109 #include <sys/vmmeter.h>
110 #include <sys/sysctl.h>
111 #include <sys/blist.h>
112 #include <sys/lock.h>
113 #include <sys/thread2.h>
115 #ifndef MAX_PAGEOUT_CLUSTER
116 #define MAX_PAGEOUT_CLUSTER 16
119 #define SWB_NPAGES MAX_PAGEOUT_CLUSTER
121 #include "opt_swap.h"
123 #include <vm/vm_object.h>
124 #include <vm/vm_page.h>
125 #include <vm/vm_pager.h>
126 #include <vm/vm_pageout.h>
127 #include <vm/swap_pager.h>
128 #include <vm/vm_extern.h>
129 #include <vm/vm_zone.h>
130 #include <vm/vnode_pager.h>
132 #include <sys/buf2.h>
133 #include <vm/vm_page2.h>
135 #define SWM_FREE 0x02 /* free, period */
136 #define SWM_POP 0x04 /* pop out */
138 #define SWBIO_READ 0x01
139 #define SWBIO_WRITE 0x02
140 #define SWBIO_SYNC 0x04
146 vm_pindex_t endi; /* inclusive */
150 * vm_swap_size is in page-sized chunks now. It was DEV_BSIZE'd chunks
154 int swap_pager_full; /* swap space exhaustion (task killing) */
155 int vm_swap_cache_use;
156 int vm_swap_anon_use;
158 static int swap_pager_almost_full; /* swap space exhaustion (w/ hysteresis)*/
159 static int nsw_rcount; /* free read buffers */
160 static int nsw_wcount_sync; /* limit write buffers / synchronous */
161 static int nsw_wcount_async; /* limit write buffers / asynchronous */
162 static int nsw_wcount_async_max;/* assigned maximum */
163 static int nsw_cluster_max; /* maximum VOP I/O allowed */
165 struct blist *swapblist;
166 static int swap_async_max = 4; /* maximum in-progress async I/O's */
167 static int swap_burst_read = 0; /* allow burst reading */
170 extern struct vnode *swapdev_vp;
171 extern struct swdevt *swdevt;
174 #define BLK2DEVIDX(blk) (nswdev > 1 ? blk / dmmax % nswdev : 0)
176 SYSCTL_INT(_vm, OID_AUTO, swap_async_max,
177 CTLFLAG_RW, &swap_async_max, 0, "Maximum running async swap ops");
178 SYSCTL_INT(_vm, OID_AUTO, swap_burst_read,
179 CTLFLAG_RW, &swap_burst_read, 0, "Allow burst reads for pageins");
181 SYSCTL_INT(_vm, OID_AUTO, swap_cache_use,
182 CTLFLAG_RD, &vm_swap_cache_use, 0, "");
183 SYSCTL_INT(_vm, OID_AUTO, swap_anon_use,
184 CTLFLAG_RD, &vm_swap_anon_use, 0, "");
185 SYSCTL_INT(_vm, OID_AUTO, swap_size,
186 CTLFLAG_RD, &vm_swap_size, 0, "");
191 * Red-Black tree for swblock entries
193 * The caller must hold vm_token
195 RB_GENERATE2(swblock_rb_tree, swblock, swb_entry, rb_swblock_compare,
196 vm_pindex_t, swb_index);
199 rb_swblock_compare(struct swblock *swb1, struct swblock *swb2)
201 if (swb1->swb_index < swb2->swb_index)
203 if (swb1->swb_index > swb2->swb_index)
210 rb_swblock_scancmp(struct swblock *swb, void *data)
212 struct swfreeinfo *info = data;
214 if (swb->swb_index < info->basei)
216 if (swb->swb_index > info->endi)
223 rb_swblock_condcmp(struct swblock *swb, void *data)
225 struct swfreeinfo *info = data;
227 if (swb->swb_index < info->basei)
233 * pagerops for OBJT_SWAP - "swap pager". Some ops are also global procedure
234 * calls hooked from other parts of the VM system and do not appear here.
235 * (see vm/swap_pager.h).
238 static void swap_pager_dealloc (vm_object_t object);
239 static int swap_pager_getpage (vm_object_t, vm_page_t *, int);
240 static void swap_chain_iodone(struct bio *biox);
242 struct pagerops swappagerops = {
243 swap_pager_dealloc, /* deallocate an OBJT_SWAP object */
244 swap_pager_getpage, /* pagein */
245 swap_pager_putpages, /* pageout */
246 swap_pager_haspage /* get backing store status for page */
250 * dmmax is in page-sized chunks with the new swap system. It was
251 * dev-bsized chunks in the old. dmmax is always a power of 2.
253 * swap_*() routines are externally accessible. swp_*() routines are
258 static int dmmax_mask;
259 int nswap_lowat = 128; /* in pages, swap_pager_almost_full warn */
260 int nswap_hiwat = 512; /* in pages, swap_pager_almost_full warn */
262 static __inline void swp_sizecheck (void);
263 static void swp_pager_async_iodone (struct bio *bio);
266 * Swap bitmap functions
269 static __inline void swp_pager_freeswapspace(vm_object_t object,
270 swblk_t blk, int npages);
271 static __inline swblk_t swp_pager_getswapspace(vm_object_t object, int npages);
277 static void swp_pager_meta_convert(vm_object_t);
278 static void swp_pager_meta_build(vm_object_t, vm_pindex_t, swblk_t);
279 static void swp_pager_meta_free(vm_object_t, vm_pindex_t, vm_pindex_t);
280 static void swp_pager_meta_free_all(vm_object_t);
281 static swblk_t swp_pager_meta_ctl(vm_object_t, vm_pindex_t, int);
284 * SWP_SIZECHECK() - update swap_pager_full indication
286 * update the swap_pager_almost_full indication and warn when we are
287 * about to run out of swap space, using lowat/hiwat hysteresis.
289 * Clear swap_pager_full ( task killing ) indication when lowat is met.
291 * No restrictions on call
292 * This routine may not block.
298 if (vm_swap_size < nswap_lowat) {
299 if (swap_pager_almost_full == 0) {
300 kprintf("swap_pager: out of swap space\n");
301 swap_pager_almost_full = 1;
305 if (vm_swap_size > nswap_hiwat)
306 swap_pager_almost_full = 0;
311 * SWAP_PAGER_INIT() - initialize the swap pager!
313 * Expected to be started from system init. NOTE: This code is run
314 * before much else so be careful what you depend on. Most of the VM
315 * system has yet to be initialized at this point.
317 * Called from the low level boot code only.
320 swap_pager_init(void *arg __unused)
323 * Device Stripe, in PAGE_SIZE'd blocks
325 dmmax = SWB_NPAGES * 2;
326 dmmax_mask = ~(dmmax - 1);
328 SYSINIT(vm_mem, SI_BOOT1_VM, SI_ORDER_THIRD, swap_pager_init, NULL)
331 * SWAP_PAGER_SWAP_INIT() - swap pager initialization from pageout process
333 * Expected to be started from pageout process once, prior to entering
336 * Called from the low level boot code only.
339 swap_pager_swap_init(void)
344 * Number of in-transit swap bp operations. Don't
345 * exhaust the pbufs completely. Make sure we
346 * initialize workable values (0 will work for hysteresis
347 * but it isn't very efficient).
349 * The nsw_cluster_max is constrained by the number of pages an XIO
350 * holds, i.e., (MAXPHYS/PAGE_SIZE) and our locally defined
351 * MAX_PAGEOUT_CLUSTER. Also be aware that swap ops are
352 * constrained by the swap device interleave stripe size.
354 * Currently we hardwire nsw_wcount_async to 4. This limit is
355 * designed to prevent other I/O from having high latencies due to
356 * our pageout I/O. The value 4 works well for one or two active swap
357 * devices but is probably a little low if you have more. Even so,
358 * a higher value would probably generate only a limited improvement
359 * with three or four active swap devices since the system does not
360 * typically have to pageout at extreme bandwidths. We will want
361 * at least 2 per swap devices, and 4 is a pretty good value if you
362 * have one NFS swap device due to the command/ack latency over NFS.
363 * So it all works out pretty well.
366 nsw_cluster_max = min((MAXPHYS/PAGE_SIZE), MAX_PAGEOUT_CLUSTER);
368 nsw_rcount = (nswbuf + 1) / 2;
369 nsw_wcount_sync = (nswbuf + 3) / 4;
370 nsw_wcount_async = 4;
371 nsw_wcount_async_max = nsw_wcount_async;
374 * The zone is dynamically allocated so generally size it to
375 * maxswzone (32MB to 512MB of KVM). Set a minimum size based
376 * on physical memory of around 8x (each swblock can hold 16 pages).
378 * With the advent of SSDs (vs HDs) the practical (swap:memory) ratio
379 * has increased dramatically.
381 n = vmstats.v_page_count / 2;
382 if (maxswzone && n < maxswzone / sizeof(struct swblock))
383 n = maxswzone / sizeof(struct swblock);
389 sizeof(struct swblock),
393 if (swap_zone != NULL)
396 * if the allocation failed, try a zone two thirds the
397 * size of the previous attempt.
402 if (swap_zone == NULL)
403 panic("swap_pager_swap_init: swap_zone == NULL");
405 kprintf("Swap zone entries reduced from %d to %d.\n", n2, n);
409 * SWAP_PAGER_ALLOC() - allocate a new OBJT_SWAP VM object and instantiate
410 * its metadata structures.
412 * This routine is called from the mmap and fork code to create a new
413 * OBJT_SWAP object. We do this by creating an OBJT_DEFAULT object
414 * and then converting it with swp_pager_meta_convert().
416 * We only support unnamed objects.
421 swap_pager_alloc(void *handle, off_t size, vm_prot_t prot, off_t offset)
425 KKASSERT(handle == NULL);
426 object = vm_object_allocate(OBJT_DEFAULT,
427 OFF_TO_IDX(offset + PAGE_MASK + size));
428 vm_object_hold(object);
429 swp_pager_meta_convert(object);
430 vm_object_drop(object);
436 * SWAP_PAGER_DEALLOC() - remove swap metadata from object
438 * The swap backing for the object is destroyed. The code is
439 * designed such that we can reinstantiate it later, but this
440 * routine is typically called only when the entire object is
441 * about to be destroyed.
443 * The object must be locked or unreferenceable.
444 * No other requirements.
447 swap_pager_dealloc(vm_object_t object)
449 vm_object_hold(object);
450 vm_object_pip_wait(object, "swpdea");
453 * Free all remaining metadata. We only bother to free it from
454 * the swap meta data. We do not attempt to free swapblk's still
455 * associated with vm_page_t's for this object. We do not care
456 * if paging is still in progress on some objects.
458 swp_pager_meta_free_all(object);
459 vm_object_drop(object);
462 /************************************************************************
463 * SWAP PAGER BITMAP ROUTINES *
464 ************************************************************************/
467 * SWP_PAGER_GETSWAPSPACE() - allocate raw swap space
469 * Allocate swap for the requested number of pages. The starting
470 * swap block number (a page index) is returned or SWAPBLK_NONE
471 * if the allocation failed.
473 * Also has the side effect of advising that somebody made a mistake
474 * when they configured swap and didn't configure enough.
476 * The caller must hold the object.
477 * This routine may not block.
479 static __inline swblk_t
480 swp_pager_getswapspace(vm_object_t object, int npages)
484 lwkt_gettoken(&vm_token);
485 if ((blk = blist_alloc(swapblist, npages)) == SWAPBLK_NONE) {
486 if (swap_pager_full != 2) {
487 kprintf("swap_pager_getswapspace: failed\n");
489 swap_pager_almost_full = 1;
492 swapacctspace(blk, -npages);
493 if (object->type == OBJT_SWAP)
494 vm_swap_anon_use += npages;
496 vm_swap_cache_use += npages;
499 lwkt_reltoken(&vm_token);
504 * SWP_PAGER_FREESWAPSPACE() - free raw swap space
506 * This routine returns the specified swap blocks back to the bitmap.
508 * Note: This routine may not block (it could in the old swap code),
509 * and through the use of the new blist routines it does not block.
511 * We must be called at splvm() to avoid races with bitmap frees from
512 * vm_page_remove() aka swap_pager_page_removed().
514 * This routine may not block.
518 swp_pager_freeswapspace(vm_object_t object, swblk_t blk, int npages)
520 struct swdevt *sp = &swdevt[BLK2DEVIDX(blk)];
522 lwkt_gettoken(&vm_token);
523 sp->sw_nused -= npages;
524 if (object->type == OBJT_SWAP)
525 vm_swap_anon_use -= npages;
527 vm_swap_cache_use -= npages;
529 if (sp->sw_flags & SW_CLOSING) {
530 lwkt_reltoken(&vm_token);
534 blist_free(swapblist, blk, npages);
535 vm_swap_size += npages;
537 lwkt_reltoken(&vm_token);
541 * SWAP_PAGER_FREESPACE() - frees swap blocks associated with a page
542 * range within an object.
544 * This is a globally accessible routine.
546 * This routine removes swapblk assignments from swap metadata.
548 * The external callers of this routine typically have already destroyed
549 * or renamed vm_page_t's associated with this range in the object so
555 swap_pager_freespace(vm_object_t object, vm_pindex_t start, vm_pindex_t size)
557 vm_object_hold(object);
558 swp_pager_meta_free(object, start, size);
559 vm_object_drop(object);
566 swap_pager_freespace_all(vm_object_t object)
568 vm_object_hold(object);
569 swp_pager_meta_free_all(object);
570 vm_object_drop(object);
574 * This function conditionally frees swap cache swap starting at
575 * (*basei) in the object. (count) swap blocks will be nominally freed.
576 * The actual number of blocks freed can be more or less than the
579 * This function nominally returns the number of blocks freed. However,
580 * the actual number of blocks freed may be less then the returned value.
581 * If the function is unable to exhaust the object or if it is able to
582 * free (approximately) the requested number of blocks it returns
585 * If we exhaust the object we will return a value n <= count.
587 * The caller must hold the object.
589 * WARNING! If count == 0 then -1 can be returned as a degenerate case,
590 * callers should always pass a count value > 0.
592 static int swap_pager_condfree_callback(struct swblock *swap, void *data);
595 swap_pager_condfree(vm_object_t object, vm_pindex_t *basei, int count)
597 struct swfreeinfo info;
599 ASSERT_LWKT_TOKEN_HELD(vm_object_token(object));
601 info.object = object;
602 info.basei = *basei; /* skip up to this page index */
603 info.begi = count; /* max swap pages to destroy */
604 info.endi = count * 8; /* max swblocks to scan */
606 swblock_rb_tree_RB_SCAN(&object->swblock_root, rb_swblock_condcmp,
607 swap_pager_condfree_callback, &info);
609 if (info.endi < 0 && info.begi <= count)
610 info.begi = count + 1;
611 return(count - (int)info.begi);
615 * The idea is to free whole meta-block to avoid fragmenting
616 * the swap space or disk I/O. We only do this if NO VM pages
619 * We do not have to deal with clearing PG_SWAPPED in related VM
620 * pages because there are no related VM pages.
622 * The caller must hold the object.
625 swap_pager_condfree_callback(struct swblock *swap, void *data)
627 struct swfreeinfo *info = data;
628 vm_object_t object = info->object;
631 for (i = 0; i < SWAP_META_PAGES; ++i) {
632 if (vm_page_lookup(object, swap->swb_index + i))
635 info->basei = swap->swb_index + SWAP_META_PAGES;
636 if (i == SWAP_META_PAGES) {
637 info->begi -= swap->swb_count;
638 swap_pager_freespace(object, swap->swb_index, SWAP_META_PAGES);
641 if ((int)info->begi < 0 || (int)info->endi < 0)
648 * Called by vm_page_alloc() when a new VM page is inserted
649 * into a VM object. Checks whether swap has been assigned to
650 * the page and sets PG_SWAPPED as necessary.
655 swap_pager_page_inserted(vm_page_t m)
657 if (m->object->swblock_count) {
658 vm_object_hold(m->object);
659 if (swp_pager_meta_ctl(m->object, m->pindex, 0) != SWAPBLK_NONE)
660 vm_page_flag_set(m, PG_SWAPPED);
661 vm_object_drop(m->object);
666 * SWAP_PAGER_RESERVE() - reserve swap blocks in object
668 * Assigns swap blocks to the specified range within the object. The
669 * swap blocks are not zerod. Any previous swap assignment is destroyed.
671 * Returns 0 on success, -1 on failure.
673 * The caller is responsible for avoiding races in the specified range.
674 * No other requirements.
677 swap_pager_reserve(vm_object_t object, vm_pindex_t start, vm_size_t size)
680 swblk_t blk = SWAPBLK_NONE;
681 vm_pindex_t beg = start; /* save start index */
683 vm_object_hold(object);
688 while ((blk = swp_pager_getswapspace(object, n)) ==
693 swp_pager_meta_free(object, beg,
695 vm_object_drop(object);
700 swp_pager_meta_build(object, start, blk);
706 swp_pager_meta_free(object, start, n);
707 vm_object_drop(object);
712 * SWAP_PAGER_COPY() - copy blocks from source pager to destination pager
713 * and destroy the source.
715 * Copy any valid swapblks from the source to the destination. In
716 * cases where both the source and destination have a valid swapblk,
717 * we keep the destination's.
719 * This routine is allowed to block. It may block allocating metadata
720 * indirectly through swp_pager_meta_build() or if paging is still in
721 * progress on the source.
723 * XXX vm_page_collapse() kinda expects us not to block because we
724 * supposedly do not need to allocate memory, but for the moment we
725 * *may* have to get a little memory from the zone allocator, but
726 * it is taken from the interrupt memory. We should be ok.
728 * The source object contains no vm_page_t's (which is just as well)
729 * The source object is of type OBJT_SWAP.
731 * The source and destination objects must be held by the caller.
734 swap_pager_copy(vm_object_t srcobject, vm_object_t dstobject,
735 vm_pindex_t base_index, int destroysource)
739 ASSERT_LWKT_TOKEN_HELD(vm_object_token(srcobject));
740 ASSERT_LWKT_TOKEN_HELD(vm_object_token(dstobject));
743 * transfer source to destination.
745 for (i = 0; i < dstobject->size; ++i) {
749 * Locate (without changing) the swapblk on the destination,
750 * unless it is invalid in which case free it silently, or
751 * if the destination is a resident page, in which case the
752 * source is thrown away.
754 dstaddr = swp_pager_meta_ctl(dstobject, i, 0);
756 if (dstaddr == SWAPBLK_NONE) {
758 * Destination has no swapblk and is not resident,
763 srcaddr = swp_pager_meta_ctl(srcobject,
764 base_index + i, SWM_POP);
766 if (srcaddr != SWAPBLK_NONE)
767 swp_pager_meta_build(dstobject, i, srcaddr);
770 * Destination has valid swapblk or it is represented
771 * by a resident page. We destroy the sourceblock.
773 swp_pager_meta_ctl(srcobject, base_index + i, SWM_FREE);
778 * Free left over swap blocks in source.
780 * We have to revert the type to OBJT_DEFAULT so we do not accidently
781 * double-remove the object from the swap queues.
785 * Reverting the type is not necessary, the caller is going
786 * to destroy srcobject directly, but I'm doing it here
787 * for consistency since we've removed the object from its
790 swp_pager_meta_free_all(srcobject);
791 if (srcobject->type == OBJT_SWAP)
792 srcobject->type = OBJT_DEFAULT;
797 * SWAP_PAGER_HASPAGE() - determine if we have good backing store for
798 * the requested page.
800 * We determine whether good backing store exists for the requested
801 * page and return TRUE if it does, FALSE if it doesn't.
803 * If TRUE, we also try to determine how much valid, contiguous backing
804 * store exists before and after the requested page within a reasonable
805 * distance. We do not try to restrict it to the swap device stripe
806 * (that is handled in getpages/putpages). It probably isn't worth
812 swap_pager_haspage(vm_object_t object, vm_pindex_t pindex)
817 * do we have good backing store at the requested index ?
819 vm_object_hold(object);
820 blk0 = swp_pager_meta_ctl(object, pindex, 0);
822 if (blk0 == SWAPBLK_NONE) {
823 vm_object_drop(object);
826 vm_object_drop(object);
831 * SWAP_PAGER_PAGE_UNSWAPPED() - remove swap backing store related to page
833 * This removes any associated swap backing store, whether valid or
834 * not, from the page. This operates on any VM object, not just OBJT_SWAP
837 * This routine is typically called when a page is made dirty, at
838 * which point any associated swap can be freed. MADV_FREE also
839 * calls us in a special-case situation
841 * NOTE!!! If the page is clean and the swap was valid, the caller
842 * should make the page dirty before calling this routine. This routine
843 * does NOT change the m->dirty status of the page. Also: MADV_FREE
846 * The page must be busied or soft-busied.
847 * The caller can hold the object to avoid blocking, else we might block.
848 * No other requirements.
851 swap_pager_unswapped(vm_page_t m)
853 if (m->flags & PG_SWAPPED) {
854 vm_object_hold(m->object);
855 KKASSERT(m->flags & PG_SWAPPED);
856 swp_pager_meta_ctl(m->object, m->pindex, SWM_FREE);
857 vm_page_flag_clear(m, PG_SWAPPED);
858 vm_object_drop(m->object);
863 * SWAP_PAGER_STRATEGY() - read, write, free blocks
865 * This implements a VM OBJECT strategy function using swap backing store.
866 * This can operate on any VM OBJECT type, not necessarily just OBJT_SWAP
869 * This is intended to be a cacheless interface (i.e. caching occurs at
870 * higher levels), and is also used as a swap-based SSD cache for vnode
871 * and device objects.
873 * All I/O goes directly to and from the swap device.
875 * We currently attempt to run I/O synchronously or asynchronously as
876 * the caller requests. This isn't perfect because we loose error
877 * sequencing when we run multiple ops in parallel to satisfy a request.
878 * But this is swap, so we let it all hang out.
883 swap_pager_strategy(vm_object_t object, struct bio *bio)
885 struct buf *bp = bio->bio_buf;
888 vm_pindex_t biox_blkno = 0;
893 struct bio_track *track;
896 * tracking for swapdev vnode I/Os
898 if (bp->b_cmd == BUF_CMD_READ)
899 track = &swapdev_vp->v_track_read;
901 track = &swapdev_vp->v_track_write;
903 if (bp->b_bcount & PAGE_MASK) {
904 bp->b_error = EINVAL;
905 bp->b_flags |= B_ERROR | B_INVAL;
907 kprintf("swap_pager_strategy: bp %p offset %lld size %d, "
908 "not page bounded\n",
909 bp, (long long)bio->bio_offset, (int)bp->b_bcount);
914 * Clear error indication, initialize page index, count, data pointer.
917 bp->b_flags &= ~B_ERROR;
918 bp->b_resid = bp->b_bcount;
920 start = (vm_pindex_t)(bio->bio_offset >> PAGE_SHIFT);
921 count = howmany(bp->b_bcount, PAGE_SIZE);
925 * Deal with BUF_CMD_FREEBLKS
927 if (bp->b_cmd == BUF_CMD_FREEBLKS) {
929 * FREE PAGE(s) - destroy underlying swap that is no longer
932 vm_object_hold(object);
933 swp_pager_meta_free(object, start, count);
934 vm_object_drop(object);
941 * We need to be able to create a new cluster of I/O's. We cannot
942 * use the caller fields of the passed bio so push a new one.
944 * Because nbio is just a placeholder for the cluster links,
945 * we can biodone() the original bio instead of nbio to make
946 * things a bit more efficient.
948 nbio = push_bio(bio);
949 nbio->bio_offset = bio->bio_offset;
950 nbio->bio_caller_info1.cluster_head = NULL;
951 nbio->bio_caller_info2.cluster_tail = NULL;
957 * Execute read or write
959 vm_object_hold(object);
965 * Obtain block. If block not found and writing, allocate a
966 * new block and build it into the object.
968 blk = swp_pager_meta_ctl(object, start, 0);
969 if ((blk == SWAPBLK_NONE) && bp->b_cmd != BUF_CMD_READ) {
970 blk = swp_pager_getswapspace(object, 1);
971 if (blk == SWAPBLK_NONE) {
972 bp->b_error = ENOMEM;
973 bp->b_flags |= B_ERROR;
976 swp_pager_meta_build(object, start, blk);
980 * Do we have to flush our current collection? Yes if:
982 * - no swap block at this index
983 * - swap block is not contiguous
984 * - we cross a physical disk boundry in the
988 biox && (biox_blkno + btoc(bufx->b_bcount) != blk ||
989 ((biox_blkno ^ blk) & dmmax_mask)
992 if (bp->b_cmd == BUF_CMD_READ) {
993 ++mycpu->gd_cnt.v_swapin;
994 mycpu->gd_cnt.v_swappgsin += btoc(bufx->b_bcount);
996 ++mycpu->gd_cnt.v_swapout;
997 mycpu->gd_cnt.v_swappgsout += btoc(bufx->b_bcount);
998 bufx->b_dirtyend = bufx->b_bcount;
1002 * Finished with this buf.
1004 KKASSERT(bufx->b_bcount != 0);
1005 if (bufx->b_cmd != BUF_CMD_READ)
1006 bufx->b_dirtyend = bufx->b_bcount;
1012 * Add new swapblk to biox, instantiating biox if necessary.
1013 * Zero-fill reads are able to take a shortcut.
1015 if (blk == SWAPBLK_NONE) {
1017 * We can only get here if we are reading. Since
1018 * we are at splvm() we can safely modify b_resid,
1019 * even if chain ops are in progress.
1021 bzero(data, PAGE_SIZE);
1022 bp->b_resid -= PAGE_SIZE;
1025 /* XXX chain count > 4, wait to <= 4 */
1027 bufx = getpbuf(NULL);
1028 biox = &bufx->b_bio1;
1029 cluster_append(nbio, bufx);
1030 bufx->b_flags |= (bufx->b_flags & B_ORDERED);
1031 bufx->b_cmd = bp->b_cmd;
1032 biox->bio_done = swap_chain_iodone;
1033 biox->bio_offset = (off_t)blk << PAGE_SHIFT;
1034 biox->bio_caller_info1.cluster_parent = nbio;
1037 bufx->b_data = data;
1039 bufx->b_bcount += PAGE_SIZE;
1046 vm_object_drop(object);
1049 * Flush out last buffer
1052 if (bufx->b_cmd == BUF_CMD_READ) {
1053 ++mycpu->gd_cnt.v_swapin;
1054 mycpu->gd_cnt.v_swappgsin += btoc(bufx->b_bcount);
1056 ++mycpu->gd_cnt.v_swapout;
1057 mycpu->gd_cnt.v_swappgsout += btoc(bufx->b_bcount);
1058 bufx->b_dirtyend = bufx->b_bcount;
1060 KKASSERT(bufx->b_bcount);
1061 if (bufx->b_cmd != BUF_CMD_READ)
1062 bufx->b_dirtyend = bufx->b_bcount;
1063 /* biox, bufx = NULL */
1067 * Now initiate all the I/O. Be careful looping on our chain as
1068 * I/O's may complete while we are still initiating them.
1070 * If the request is a 100% sparse read no bios will be present
1071 * and we just biodone() the buffer.
1073 nbio->bio_caller_info2.cluster_tail = NULL;
1074 bufx = nbio->bio_caller_info1.cluster_head;
1078 biox = &bufx->b_bio1;
1080 bufx = bufx->b_cluster_next;
1081 vn_strategy(swapdev_vp, biox);
1088 * Completion of the cluster will also call biodone_chain(nbio).
1089 * We never call biodone(nbio) so we don't have to worry about
1090 * setting up a bio_done callback. It's handled in the sub-IO.
1101 swap_chain_iodone(struct bio *biox)
1104 struct buf *bufx; /* chained sub-buffer */
1105 struct bio *nbio; /* parent nbio with chain glue */
1106 struct buf *bp; /* original bp associated with nbio */
1109 bufx = biox->bio_buf;
1110 nbio = biox->bio_caller_info1.cluster_parent;
1114 * Update the original buffer
1116 KKASSERT(bp != NULL);
1117 if (bufx->b_flags & B_ERROR) {
1118 atomic_set_int(&bufx->b_flags, B_ERROR);
1119 bp->b_error = bufx->b_error; /* race ok */
1120 } else if (bufx->b_resid != 0) {
1121 atomic_set_int(&bufx->b_flags, B_ERROR);
1122 bp->b_error = EINVAL; /* race ok */
1124 atomic_subtract_int(&bp->b_resid, bufx->b_bcount);
1128 * Remove us from the chain.
1130 spin_lock(&bp->b_lock.lk_spinlock);
1131 nextp = &nbio->bio_caller_info1.cluster_head;
1132 while (*nextp != bufx) {
1133 KKASSERT(*nextp != NULL);
1134 nextp = &(*nextp)->b_cluster_next;
1136 *nextp = bufx->b_cluster_next;
1137 chain_empty = (nbio->bio_caller_info1.cluster_head == NULL);
1138 spin_unlock(&bp->b_lock.lk_spinlock);
1141 * Clean up bufx. If the chain is now empty we finish out
1142 * the parent. Note that we may be racing other completions
1143 * so we must use the chain_empty status from above.
1146 if (bp->b_resid != 0 && !(bp->b_flags & B_ERROR)) {
1147 atomic_set_int(&bp->b_flags, B_ERROR);
1148 bp->b_error = EINVAL;
1150 biodone_chain(nbio);
1152 relpbuf(bufx, NULL);
1156 * SWAP_PAGER_GETPAGES() - bring page in from swap
1158 * The requested page may have to be brought in from swap. Calculate the
1159 * swap block and bring in additional pages if possible. All pages must
1160 * have contiguous swap block assignments and reside in the same object.
1162 * The caller has a single vm_object_pip_add() reference prior to
1163 * calling us and we should return with the same.
1165 * The caller has BUSY'd the page. We should return with (*mpp) left busy,
1166 * and any additinal pages unbusied.
1168 * If the caller encounters a PG_RAM page it will pass it to us even though
1169 * it may be valid and dirty. We cannot overwrite the page in this case!
1170 * The case is used to allow us to issue pure read-aheads.
1172 * NOTE! XXX This code does not entirely pipeline yet due to the fact that
1173 * the PG_RAM page is validated at the same time as mreq. What we
1174 * really need to do is issue a separate read-ahead pbuf.
1179 swap_pager_getpage(vm_object_t object, vm_page_t *mpp, int seqaccess)
1192 vm_page_t marray[XIO_INTERNAL_PAGES];
1196 vm_object_hold(object);
1197 if (mreq->object != object) {
1198 panic("swap_pager_getpages: object mismatch %p/%p",
1205 * We don't want to overwrite a fully valid page as it might be
1206 * dirty. This case can occur when e.g. vm_fault hits a perfectly
1207 * valid page with PG_RAM set.
1209 * In this case we see if the next page is a suitable page-in
1210 * candidate and if it is we issue read-ahead. PG_RAM will be
1211 * set on the last page of the read-ahead to continue the pipeline.
1213 if (mreq->valid == VM_PAGE_BITS_ALL) {
1214 if (swap_burst_read == 0 || mreq->pindex + 1 >= object->size) {
1215 vm_object_drop(object);
1216 return(VM_PAGER_OK);
1218 blk = swp_pager_meta_ctl(object, mreq->pindex + 1, 0);
1219 if (blk == SWAPBLK_NONE) {
1220 vm_object_drop(object);
1221 return(VM_PAGER_OK);
1223 m = vm_page_lookup_busy_try(object, mreq->pindex + 1,
1226 vm_object_drop(object);
1227 return(VM_PAGER_OK);
1228 } else if (m == NULL) {
1230 * Use VM_ALLOC_QUICK to avoid blocking on cache
1233 m = vm_page_alloc(object, mreq->pindex + 1,
1236 vm_object_drop(object);
1237 return(VM_PAGER_OK);
1242 vm_object_drop(object);
1243 return(VM_PAGER_OK);
1245 vm_page_unqueue_nowakeup(m);
1255 * Try to block-read contiguous pages from swap if sequential,
1256 * otherwise just read one page. Contiguous pages from swap must
1257 * reside within a single device stripe because the I/O cannot be
1258 * broken up across multiple stripes.
1260 * Note that blk and iblk can be SWAPBLK_NONE but the loop is
1261 * set up such that the case(s) are handled implicitly.
1263 blk = swp_pager_meta_ctl(mreq->object, mreq->pindex, 0);
1266 for (i = 1; swap_burst_read &&
1267 i < XIO_INTERNAL_PAGES &&
1268 mreq->pindex + i < object->size; ++i) {
1271 iblk = swp_pager_meta_ctl(object, mreq->pindex + i, 0);
1272 if (iblk != blk + i)
1274 if ((blk ^ iblk) & dmmax_mask)
1276 m = vm_page_lookup_busy_try(object, mreq->pindex + i,
1280 } else if (m == NULL) {
1282 * Use VM_ALLOC_QUICK to avoid blocking on cache
1285 m = vm_page_alloc(object, mreq->pindex + i,
1294 vm_page_unqueue_nowakeup(m);
1300 vm_page_flag_set(marray[i - 1], PG_RAM);
1303 * If mreq is the requested page and we have nothing to do return
1304 * VM_PAGER_FAIL. If raonly is set mreq is just another read-ahead
1305 * page and must be cleaned up.
1307 if (blk == SWAPBLK_NONE) {
1310 vnode_pager_freepage(mreq);
1311 vm_object_drop(object);
1312 return(VM_PAGER_OK);
1314 vm_object_drop(object);
1315 return(VM_PAGER_FAIL);
1320 * map our page(s) into kva for input
1322 bp = getpbuf_kva(&nsw_rcount);
1324 kva = (vm_offset_t) bp->b_kvabase;
1325 bcopy(marray, bp->b_xio.xio_pages, i * sizeof(vm_page_t));
1326 pmap_qenter(kva, bp->b_xio.xio_pages, i);
1328 bp->b_data = (caddr_t)kva;
1329 bp->b_bcount = PAGE_SIZE * i;
1330 bp->b_xio.xio_npages = i;
1331 bio->bio_done = swp_pager_async_iodone;
1332 bio->bio_offset = (off_t)blk << PAGE_SHIFT;
1333 bio->bio_caller_info1.index = SWBIO_READ;
1336 * Set index. If raonly set the index beyond the array so all
1337 * the pages are treated the same, otherwise the original mreq is
1341 bio->bio_driver_info = (void *)(intptr_t)i;
1343 bio->bio_driver_info = (void *)(intptr_t)0;
1345 for (j = 0; j < i; ++j)
1346 vm_page_flag_set(bp->b_xio.xio_pages[j], PG_SWAPINPROG);
1348 mycpu->gd_cnt.v_swapin++;
1349 mycpu->gd_cnt.v_swappgsin += bp->b_xio.xio_npages;
1352 * We still hold the lock on mreq, and our automatic completion routine
1353 * does not remove it.
1355 vm_object_pip_add(object, bp->b_xio.xio_npages);
1358 * perform the I/O. NOTE!!! bp cannot be considered valid after
1359 * this point because we automatically release it on completion.
1360 * Instead, we look at the one page we are interested in which we
1361 * still hold a lock on even through the I/O completion.
1363 * The other pages in our m[] array are also released on completion,
1364 * so we cannot assume they are valid anymore either.
1366 bp->b_cmd = BUF_CMD_READ;
1368 vn_strategy(swapdev_vp, bio);
1371 * Wait for the page we want to complete. PG_SWAPINPROG is always
1372 * cleared on completion. If an I/O error occurs, SWAPBLK_NONE
1373 * is set in the meta-data.
1375 * If this is a read-ahead only we return immediately without
1379 vm_object_drop(object);
1380 return(VM_PAGER_OK);
1384 * Read-ahead includes originally requested page case.
1387 flags = mreq->flags;
1389 if ((flags & PG_SWAPINPROG) == 0)
1391 tsleep_interlock(mreq, 0);
1392 if (!atomic_cmpset_int(&mreq->flags, flags,
1393 flags | PG_WANTED | PG_REFERENCED)) {
1396 mycpu->gd_cnt.v_intrans++;
1397 if (tsleep(mreq, PINTERLOCKED, "swread", hz*20)) {
1399 "swap_pager: indefinite wait buffer: "
1400 " offset: %lld, size: %ld\n",
1401 (long long)bio->bio_offset,
1408 * mreq is left bussied after completion, but all the other pages
1409 * are freed. If we had an unrecoverable read error the page will
1412 vm_object_drop(object);
1413 if (mreq->valid != VM_PAGE_BITS_ALL)
1414 return(VM_PAGER_ERROR);
1416 return(VM_PAGER_OK);
1419 * A final note: in a low swap situation, we cannot deallocate swap
1420 * and mark a page dirty here because the caller is likely to mark
1421 * the page clean when we return, causing the page to possibly revert
1422 * to all-zero's later.
1427 * swap_pager_putpages:
1429 * Assign swap (if necessary) and initiate I/O on the specified pages.
1431 * We support both OBJT_DEFAULT and OBJT_SWAP objects. DEFAULT objects
1432 * are automatically converted to SWAP objects.
1434 * In a low memory situation we may block in vn_strategy(), but the new
1435 * vm_page reservation system coupled with properly written VFS devices
1436 * should ensure that no low-memory deadlock occurs. This is an area
1439 * The parent has N vm_object_pip_add() references prior to
1440 * calling us and will remove references for rtvals[] that are
1441 * not set to VM_PAGER_PEND. We need to remove the rest on I/O
1444 * The parent has soft-busy'd the pages it passes us and will unbusy
1445 * those whos rtvals[] entry is not set to VM_PAGER_PEND on return.
1446 * We need to unbusy the rest on I/O completion.
1451 swap_pager_putpages(vm_object_t object, vm_page_t *m, int count,
1452 boolean_t sync, int *rtvals)
1457 vm_object_hold(object);
1459 if (count && m[0]->object != object) {
1460 panic("swap_pager_getpages: object mismatch %p/%p",
1469 * Turn object into OBJT_SWAP
1470 * check for bogus sysops
1471 * force sync if not pageout process
1473 if (object->type == OBJT_DEFAULT) {
1474 if (object->type == OBJT_DEFAULT)
1475 swp_pager_meta_convert(object);
1478 if (curthread != pagethread)
1484 * Update nsw parameters from swap_async_max sysctl values.
1485 * Do not let the sysop crash the machine with bogus numbers.
1488 if (swap_async_max != nsw_wcount_async_max) {
1494 if ((n = swap_async_max) > nswbuf / 2)
1501 * Adjust difference ( if possible ). If the current async
1502 * count is too low, we may not be able to make the adjustment
1505 * vm_token needed for nsw_wcount sleep interlock
1507 lwkt_gettoken(&vm_token);
1508 n -= nsw_wcount_async_max;
1509 if (nsw_wcount_async + n >= 0) {
1510 nsw_wcount_async += n;
1511 nsw_wcount_async_max += n;
1512 wakeup(&nsw_wcount_async);
1514 lwkt_reltoken(&vm_token);
1520 * Assign swap blocks and issue I/O. We reallocate swap on the fly.
1521 * The page is left dirty until the pageout operation completes
1525 for (i = 0; i < count; i += n) {
1532 * Maximum I/O size is limited by a number of factors.
1535 n = min(BLIST_MAX_ALLOC, count - i);
1536 n = min(n, nsw_cluster_max);
1538 lwkt_gettoken(&vm_token);
1541 * Get biggest block of swap we can. If we fail, fall
1542 * back and try to allocate a smaller block. Don't go
1543 * overboard trying to allocate space if it would overly
1547 (blk = swp_pager_getswapspace(object, n)) == SWAPBLK_NONE &&
1552 if (blk == SWAPBLK_NONE) {
1553 for (j = 0; j < n; ++j)
1554 rtvals[i+j] = VM_PAGER_FAIL;
1555 lwkt_reltoken(&vm_token);
1560 * The I/O we are constructing cannot cross a physical
1561 * disk boundry in the swap stripe. Note: we are still
1564 if ((blk ^ (blk + n)) & dmmax_mask) {
1565 j = ((blk + dmmax) & dmmax_mask) - blk;
1566 swp_pager_freeswapspace(object, blk + j, n - j);
1571 * All I/O parameters have been satisfied, build the I/O
1572 * request and assign the swap space.
1575 bp = getpbuf_kva(&nsw_wcount_sync);
1577 bp = getpbuf_kva(&nsw_wcount_async);
1580 lwkt_reltoken(&vm_token);
1582 pmap_qenter((vm_offset_t)bp->b_data, &m[i], n);
1584 bp->b_bcount = PAGE_SIZE * n;
1585 bio->bio_offset = (off_t)blk << PAGE_SHIFT;
1587 for (j = 0; j < n; ++j) {
1588 vm_page_t mreq = m[i+j];
1590 swp_pager_meta_build(mreq->object, mreq->pindex,
1592 if (object->type == OBJT_SWAP)
1593 vm_page_dirty(mreq);
1594 rtvals[i+j] = VM_PAGER_OK;
1596 vm_page_flag_set(mreq, PG_SWAPINPROG);
1597 bp->b_xio.xio_pages[j] = mreq;
1599 bp->b_xio.xio_npages = n;
1601 mycpu->gd_cnt.v_swapout++;
1602 mycpu->gd_cnt.v_swappgsout += bp->b_xio.xio_npages;
1604 bp->b_dirtyoff = 0; /* req'd for NFS */
1605 bp->b_dirtyend = bp->b_bcount; /* req'd for NFS */
1606 bp->b_cmd = BUF_CMD_WRITE;
1607 bio->bio_caller_info1.index = SWBIO_WRITE;
1612 if (sync == FALSE) {
1613 bio->bio_done = swp_pager_async_iodone;
1615 vn_strategy(swapdev_vp, bio);
1617 for (j = 0; j < n; ++j)
1618 rtvals[i+j] = VM_PAGER_PEND;
1623 * Issue synchrnously.
1625 * Wait for the sync I/O to complete, then update rtvals.
1626 * We just set the rtvals[] to VM_PAGER_PEND so we can call
1627 * our async completion routine at the end, thus avoiding a
1630 bio->bio_caller_info1.index |= SWBIO_SYNC;
1631 bio->bio_done = biodone_sync;
1632 bio->bio_flags |= BIO_SYNC;
1633 vn_strategy(swapdev_vp, bio);
1634 biowait(bio, "swwrt");
1636 for (j = 0; j < n; ++j)
1637 rtvals[i+j] = VM_PAGER_PEND;
1640 * Now that we are through with the bp, we can call the
1641 * normal async completion, which frees everything up.
1643 swp_pager_async_iodone(bio);
1645 vm_object_drop(object);
1652 swap_pager_newswap(void)
1658 * swp_pager_async_iodone:
1660 * Completion routine for asynchronous reads and writes from/to swap.
1661 * Also called manually by synchronous code to finish up a bp.
1663 * For READ operations, the pages are PG_BUSY'd. For WRITE operations,
1664 * the pages are vm_page_t->busy'd. For READ operations, we PG_BUSY
1665 * unbusy all pages except the 'main' request page. For WRITE
1666 * operations, we vm_page_t->busy'd unbusy all pages ( we can do this
1667 * because we marked them all VM_PAGER_PEND on return from putpages ).
1669 * This routine may not block.
1674 swp_pager_async_iodone(struct bio *bio)
1676 struct buf *bp = bio->bio_buf;
1677 vm_object_t object = NULL;
1684 if (bp->b_flags & B_ERROR) {
1686 "swap_pager: I/O error - %s failed; offset %lld,"
1687 "size %ld, error %d\n",
1688 ((bio->bio_caller_info1.index & SWBIO_READ) ?
1689 "pagein" : "pageout"),
1690 (long long)bio->bio_offset,
1697 * set object, raise to splvm().
1699 if (bp->b_xio.xio_npages)
1700 object = bp->b_xio.xio_pages[0]->object;
1703 * remove the mapping for kernel virtual
1705 pmap_qremove((vm_offset_t)bp->b_data, bp->b_xio.xio_npages);
1708 * cleanup pages. If an error occurs writing to swap, we are in
1709 * very serious trouble. If it happens to be a disk error, though,
1710 * we may be able to recover by reassigning the swap later on. So
1711 * in this case we remove the m->swapblk assignment for the page
1712 * but do not free it in the rlist. The errornous block(s) are thus
1713 * never reallocated as swap. Redirty the page and continue.
1715 for (i = 0; i < bp->b_xio.xio_npages; ++i) {
1716 vm_page_t m = bp->b_xio.xio_pages[i];
1718 if (bp->b_flags & B_ERROR) {
1720 * If an error occurs I'd love to throw the swapblk
1721 * away without freeing it back to swapspace, so it
1722 * can never be used again. But I can't from an
1726 if (bio->bio_caller_info1.index & SWBIO_READ) {
1728 * When reading, reqpage needs to stay
1729 * locked for the parent, but all other
1730 * pages can be freed. We still want to
1731 * wakeup the parent waiting on the page,
1732 * though. ( also: pg_reqpage can be -1 and
1733 * not match anything ).
1735 * We have to wake specifically requested pages
1736 * up too because we cleared PG_SWAPINPROG and
1737 * someone may be waiting for that.
1739 * NOTE: for reads, m->dirty will probably
1740 * be overridden by the original caller of
1741 * getpages so don't play cute tricks here.
1743 * NOTE: We can't actually free the page from
1744 * here, because this is an interrupt. It
1745 * is not legal to mess with object->memq
1746 * from an interrupt. Deactivate the page
1751 vm_page_flag_clear(m, PG_ZERO);
1752 vm_page_flag_clear(m, PG_SWAPINPROG);
1755 * bio_driver_info holds the requested page
1758 if (i != (int)(intptr_t)bio->bio_driver_info) {
1759 vm_page_deactivate(m);
1765 * If i == bp->b_pager.pg_reqpage, do not wake
1766 * the page up. The caller needs to.
1770 * If a write error occurs remove the swap
1771 * assignment (note that PG_SWAPPED may or
1772 * may not be set depending on prior activity).
1774 * Re-dirty OBJT_SWAP pages as there is no
1775 * other backing store, we can't throw the
1778 * Non-OBJT_SWAP pages (aka swapcache) must
1779 * not be dirtied since they may not have
1780 * been dirty in the first place, and they
1781 * do have backing store (the vnode).
1783 vm_page_busy_wait(m, FALSE, "swadpg");
1784 swp_pager_meta_ctl(m->object, m->pindex,
1786 vm_page_flag_clear(m, PG_SWAPPED);
1787 if (m->object->type == OBJT_SWAP) {
1789 vm_page_activate(m);
1791 vm_page_flag_clear(m, PG_SWAPINPROG);
1792 vm_page_io_finish(m);
1795 } else if (bio->bio_caller_info1.index & SWBIO_READ) {
1797 * NOTE: for reads, m->dirty will probably be
1798 * overridden by the original caller of getpages so
1799 * we cannot set them in order to free the underlying
1800 * swap in a low-swap situation. I don't think we'd
1801 * want to do that anyway, but it was an optimization
1802 * that existed in the old swapper for a time before
1803 * it got ripped out due to precisely this problem.
1805 * clear PG_ZERO in page.
1807 * If not the requested page then deactivate it.
1809 * Note that the requested page, reqpage, is left
1810 * busied, but we still have to wake it up. The
1811 * other pages are released (unbusied) by
1812 * vm_page_wakeup(). We do not set reqpage's
1813 * valid bits here, it is up to the caller.
1817 * NOTE: can't call pmap_clear_modify(m) from an
1818 * interrupt thread, the pmap code may have to map
1819 * non-kernel pmaps and currently asserts the case.
1821 /*pmap_clear_modify(m);*/
1822 m->valid = VM_PAGE_BITS_ALL;
1824 vm_page_flag_clear(m, PG_ZERO | PG_SWAPINPROG);
1825 vm_page_flag_set(m, PG_SWAPPED);
1828 * We have to wake specifically requested pages
1829 * up too because we cleared PG_SWAPINPROG and
1830 * could be waiting for it in getpages. However,
1831 * be sure to not unbusy getpages specifically
1832 * requested page - getpages expects it to be
1835 * bio_driver_info holds the requested page
1837 if (i != (int)(intptr_t)bio->bio_driver_info) {
1838 vm_page_deactivate(m);
1845 * Mark the page clean but do not mess with the
1846 * pmap-layer's modified state. That state should
1847 * also be clear since the caller protected the
1848 * page VM_PROT_READ, but allow the case.
1850 * We are in an interrupt, avoid pmap operations.
1852 * If we have a severe page deficit, deactivate the
1853 * page. Do not try to cache it (which would also
1854 * involve a pmap op), because the page might still
1857 * When using the swap to cache clean vnode pages
1858 * we do not mess with the page dirty bits.
1860 vm_page_busy_wait(m, FALSE, "swadpg");
1861 if (m->object->type == OBJT_SWAP)
1863 vm_page_flag_clear(m, PG_SWAPINPROG);
1864 vm_page_flag_set(m, PG_SWAPPED);
1865 if (vm_page_count_severe())
1866 vm_page_deactivate(m);
1868 if (!vm_page_count_severe() || !vm_page_try_to_cache(m))
1869 vm_page_protect(m, VM_PROT_READ);
1871 vm_page_io_finish(m);
1877 * adjust pip. NOTE: the original parent may still have its own
1878 * pip refs on the object.
1882 vm_object_pip_wakeup_n(object, bp->b_xio.xio_npages);
1885 * Release the physical I/O buffer.
1887 * NOTE: Due to synchronous operations in the write case b_cmd may
1888 * already be set to BUF_CMD_DONE and BIO_SYNC may have already
1891 * Use vm_token to interlock nsw_rcount/wcount wakeup?
1893 lwkt_gettoken(&vm_token);
1894 if (bio->bio_caller_info1.index & SWBIO_READ)
1895 nswptr = &nsw_rcount;
1896 else if (bio->bio_caller_info1.index & SWBIO_SYNC)
1897 nswptr = &nsw_wcount_sync;
1899 nswptr = &nsw_wcount_async;
1900 bp->b_cmd = BUF_CMD_DONE;
1901 relpbuf(bp, nswptr);
1902 lwkt_reltoken(&vm_token);
1906 * Fault-in a potentially swapped page and remove the swap reference.
1908 * object must be held.
1910 static __inline void
1911 swp_pager_fault_page(vm_object_t object, vm_pindex_t pindex)
1917 ASSERT_LWKT_TOKEN_HELD(vm_object_token(object));
1919 if (object->type == OBJT_VNODE) {
1921 * Any swap related to a vnode is due to swapcache. We must
1922 * vget() the vnode in case it is not active (otherwise
1923 * vref() will panic). Calling vm_object_page_remove() will
1924 * ensure that any swap ref is removed interlocked with the
1925 * page. clean_only is set to TRUE so we don't throw away
1928 vp = object->handle;
1929 error = vget(vp, LK_SHARED | LK_RETRY | LK_CANRECURSE);
1931 vm_object_page_remove(object, pindex, pindex + 1, TRUE);
1936 * Otherwise it is a normal OBJT_SWAP object and we can
1937 * fault the page in and remove the swap.
1939 m = vm_fault_object_page(object, IDX_TO_OFF(pindex),
1941 VM_FAULT_DIRTY | VM_FAULT_UNSWAP,
1949 swap_pager_swapoff(int devidx)
1952 struct swblock *swap;
1956 lwkt_gettoken(&vmobj_token);
1958 TAILQ_FOREACH(object, &vm_object_list, object_list) {
1959 if (object->type != OBJT_SWAP && object->type != OBJT_VNODE)
1961 vm_object_hold(object);
1962 if (object->type == OBJT_SWAP || object->type == OBJT_VNODE) {
1964 swblock_rb_tree, &object->swblock_root) {
1965 for (i = 0; i < SWAP_META_PAGES; ++i) {
1966 v = swap->swb_pages[i];
1967 if (v != SWAPBLK_NONE &&
1968 BLK2DEVIDX(v) == devidx) {
1969 swp_pager_fault_page(
1971 swap->swb_index + i);
1972 vm_object_drop(object);
1978 vm_object_drop(object);
1980 lwkt_reltoken(&vmobj_token);
1983 * If we fail to locate all swblocks we just fail gracefully and
1984 * do not bother to restore paging on the swap device. If the
1985 * user wants to retry the user can retry.
1987 if (swdevt[devidx].sw_nused)
1993 /************************************************************************
1995 ************************************************************************
1997 * These routines manipulate the swap metadata stored in the
1998 * OBJT_SWAP object. All swp_*() routines must be called at
1999 * splvm() because swap can be freed up by the low level vm_page
2000 * code which might be called from interrupts beyond what splbio() covers.
2002 * Swap metadata is implemented with a global hash and not directly
2003 * linked into the object. Instead the object simply contains
2004 * appropriate tracking counters.
2008 * Lookup the swblock containing the specified swap block index.
2010 * The caller must hold the object.
2014 swp_pager_lookup(vm_object_t object, vm_pindex_t index)
2016 ASSERT_LWKT_TOKEN_HELD(vm_object_token(object));
2017 index &= ~SWAP_META_MASK;
2018 return (RB_LOOKUP(swblock_rb_tree, &object->swblock_root, index));
2022 * Remove a swblock from the RB tree.
2024 * The caller must hold the object.
2028 swp_pager_remove(vm_object_t object, struct swblock *swap)
2030 ASSERT_LWKT_TOKEN_HELD(vm_object_token(object));
2031 RB_REMOVE(swblock_rb_tree, &object->swblock_root, swap);
2035 * Convert default object to swap object if necessary
2037 * The caller must hold the object.
2040 swp_pager_meta_convert(vm_object_t object)
2042 if (object->type == OBJT_DEFAULT) {
2043 object->type = OBJT_SWAP;
2044 KKASSERT(object->swblock_count == 0);
2049 * SWP_PAGER_META_BUILD() - add swap block to swap meta data for object
2051 * We first convert the object to a swap object if it is a default
2052 * object. Vnode objects do not need to be converted.
2054 * The specified swapblk is added to the object's swap metadata. If
2055 * the swapblk is not valid, it is freed instead. Any previously
2056 * assigned swapblk is freed.
2058 * The caller must hold the object.
2061 swp_pager_meta_build(vm_object_t object, vm_pindex_t index, swblk_t swapblk)
2063 struct swblock *swap;
2064 struct swblock *oswap;
2066 KKASSERT(swapblk != SWAPBLK_NONE);
2067 ASSERT_LWKT_TOKEN_HELD(vm_object_token(object));
2070 * Convert object if necessary
2072 if (object->type == OBJT_DEFAULT)
2073 swp_pager_meta_convert(object);
2076 * Locate swblock. If not found create, but if we aren't adding
2077 * anything just return. If we run out of space in the map we wait
2078 * and, since the hash table may have changed, retry.
2081 swap = swp_pager_lookup(object, index);
2086 swap = zalloc(swap_zone);
2091 swap->swb_index = index & ~SWAP_META_MASK;
2092 swap->swb_count = 0;
2094 ++object->swblock_count;
2096 for (i = 0; i < SWAP_META_PAGES; ++i)
2097 swap->swb_pages[i] = SWAPBLK_NONE;
2098 oswap = RB_INSERT(swblock_rb_tree, &object->swblock_root, swap);
2099 KKASSERT(oswap == NULL);
2103 * Delete prior contents of metadata
2106 index &= SWAP_META_MASK;
2108 if (swap->swb_pages[index] != SWAPBLK_NONE) {
2109 swp_pager_freeswapspace(object, swap->swb_pages[index], 1);
2114 * Enter block into metadata
2116 swap->swb_pages[index] = swapblk;
2117 if (swapblk != SWAPBLK_NONE)
2122 * SWP_PAGER_META_FREE() - free a range of blocks in the object's swap metadata
2124 * The requested range of blocks is freed, with any associated swap
2125 * returned to the swap bitmap.
2127 * This routine will free swap metadata structures as they are cleaned
2128 * out. This routine does *NOT* operate on swap metadata associated
2129 * with resident pages.
2131 * The caller must hold the object.
2133 static int swp_pager_meta_free_callback(struct swblock *swb, void *data);
2136 swp_pager_meta_free(vm_object_t object, vm_pindex_t index, vm_pindex_t count)
2138 struct swfreeinfo info;
2140 ASSERT_LWKT_TOKEN_HELD(vm_object_token(object));
2145 if (object->swblock_count == 0) {
2146 KKASSERT(RB_EMPTY(&object->swblock_root));
2153 * Setup for RB tree scan. Note that the pindex range can be huge
2154 * due to the 64 bit page index space so we cannot safely iterate.
2156 info.object = object;
2157 info.basei = index & ~SWAP_META_MASK;
2159 info.endi = index + count - 1;
2160 swblock_rb_tree_RB_SCAN(&object->swblock_root, rb_swblock_scancmp,
2161 swp_pager_meta_free_callback, &info);
2165 * The caller must hold the object.
2169 swp_pager_meta_free_callback(struct swblock *swap, void *data)
2171 struct swfreeinfo *info = data;
2172 vm_object_t object = info->object;
2177 * Figure out the range within the swblock. The wider scan may
2178 * return edge-case swap blocks when the start and/or end points
2179 * are in the middle of a block.
2181 if (swap->swb_index < info->begi)
2182 index = (int)info->begi & SWAP_META_MASK;
2186 if (swap->swb_index + SWAP_META_PAGES > info->endi)
2187 eindex = (int)info->endi & SWAP_META_MASK;
2189 eindex = SWAP_META_MASK;
2192 * Scan and free the blocks. The loop terminates early
2193 * if (swap) runs out of blocks and could be freed.
2195 while (index <= eindex) {
2196 swblk_t v = swap->swb_pages[index];
2198 if (v != SWAPBLK_NONE) {
2199 swap->swb_pages[index] = SWAPBLK_NONE;
2200 if (--swap->swb_count == 0) {
2201 swp_pager_remove(object, swap);
2202 zfree(swap_zone, swap);
2203 --object->swblock_count;
2206 swp_pager_freeswapspace(object, v, 1); /* can block */
2210 /* swap may be invalid here due to zfree above */
2215 * SWP_PAGER_META_FREE_ALL() - destroy all swap metadata associated with object
2217 * This routine locates and destroys all swap metadata associated with
2220 * The caller must hold the object.
2223 swp_pager_meta_free_all(vm_object_t object)
2225 struct swblock *swap;
2228 ASSERT_LWKT_TOKEN_HELD(vm_object_token(object));
2230 while ((swap = RB_ROOT(&object->swblock_root)) != NULL) {
2231 swp_pager_remove(object, swap);
2232 for (i = 0; i < SWAP_META_PAGES; ++i) {
2233 swblk_t v = swap->swb_pages[i];
2234 if (v != SWAPBLK_NONE) {
2236 swp_pager_freeswapspace(object, v, 1);
2239 if (swap->swb_count != 0)
2240 panic("swap_pager_meta_free_all: swb_count != 0");
2241 zfree(swap_zone, swap);
2242 --object->swblock_count;
2244 KKASSERT(object->swblock_count == 0);
2248 * SWP_PAGER_METACTL() - misc control of swap and vm_page_t meta data.
2250 * This routine is capable of looking up, popping, or freeing
2251 * swapblk assignments in the swap meta data or in the vm_page_t.
2252 * The routine typically returns the swapblk being looked-up, or popped,
2253 * or SWAPBLK_NONE if the block was freed, or SWAPBLK_NONE if the block
2254 * was invalid. This routine will automatically free any invalid
2255 * meta-data swapblks.
2257 * It is not possible to store invalid swapblks in the swap meta data
2258 * (other then a literal 'SWAPBLK_NONE'), so we don't bother checking.
2260 * When acting on a busy resident page and paging is in progress, we
2261 * have to wait until paging is complete but otherwise can act on the
2264 * SWM_FREE remove and free swap block from metadata
2265 * SWM_POP remove from meta data but do not free.. pop it out
2267 * The caller must hold the object.
2270 swp_pager_meta_ctl(vm_object_t object, vm_pindex_t index, int flags)
2272 struct swblock *swap;
2275 if (object->swblock_count == 0)
2276 return(SWAPBLK_NONE);
2279 swap = swp_pager_lookup(object, index);
2282 index &= SWAP_META_MASK;
2283 r1 = swap->swb_pages[index];
2285 if (r1 != SWAPBLK_NONE) {
2286 if (flags & (SWM_FREE|SWM_POP)) {
2287 swap->swb_pages[index] = SWAPBLK_NONE;
2288 if (--swap->swb_count == 0) {
2289 swp_pager_remove(object, swap);
2290 zfree(swap_zone, swap);
2291 --object->swblock_count;
2294 if (flags & SWM_FREE) {
2295 swp_pager_freeswapspace(object, r1, 1);