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 #include "opt_swap.h"
117 #include <vm/vm_object.h>
118 #include <vm/vm_page.h>
119 #include <vm/vm_pager.h>
120 #include <vm/vm_pageout.h>
121 #include <vm/swap_pager.h>
122 #include <vm/vm_extern.h>
123 #include <vm/vm_zone.h>
124 #include <vm/vnode_pager.h>
126 #include <sys/buf2.h>
127 #include <vm/vm_page2.h>
129 #ifndef MAX_PAGEOUT_CLUSTER
130 #define MAX_PAGEOUT_CLUSTER SWB_NPAGES
133 #define SWM_FREE 0x02 /* free, period */
134 #define SWM_POP 0x04 /* pop out */
136 #define SWBIO_READ 0x01
137 #define SWBIO_WRITE 0x02
138 #define SWBIO_SYNC 0x04
144 vm_pindex_t endi; /* inclusive */
147 struct swswapoffinfo {
153 * vm_swap_size is in page-sized chunks now. It was DEV_BSIZE'd chunks
157 int swap_pager_full; /* swap space exhaustion (task killing) */
158 int vm_swap_cache_use;
159 int vm_swap_anon_use;
161 static int swap_pager_almost_full; /* swap space exhaustion (w/ hysteresis)*/
162 static int nsw_rcount; /* free read buffers */
163 static int nsw_wcount_sync; /* limit write buffers / synchronous */
164 static int nsw_wcount_async; /* limit write buffers / asynchronous */
165 static int nsw_wcount_async_max;/* assigned maximum */
166 static int nsw_cluster_max; /* maximum VOP I/O allowed */
168 struct blist *swapblist;
169 static int swap_async_max = 4; /* maximum in-progress async I/O's */
170 static int swap_burst_read = 0; /* allow burst reading */
171 static swblk_t swapiterator; /* linearize allocations */
174 extern struct vnode *swapdev_vp;
175 extern struct swdevt *swdevt;
178 #define BLK2DEVIDX(blk) (nswdev > 1 ? blk / dmmax % nswdev : 0)
180 SYSCTL_INT(_vm, OID_AUTO, swap_async_max,
181 CTLFLAG_RW, &swap_async_max, 0, "Maximum running async swap ops");
182 SYSCTL_INT(_vm, OID_AUTO, swap_burst_read,
183 CTLFLAG_RW, &swap_burst_read, 0, "Allow burst reads for pageins");
185 SYSCTL_INT(_vm, OID_AUTO, swap_cache_use,
186 CTLFLAG_RD, &vm_swap_cache_use, 0, "");
187 SYSCTL_INT(_vm, OID_AUTO, swap_anon_use,
188 CTLFLAG_RD, &vm_swap_anon_use, 0, "");
189 SYSCTL_INT(_vm, OID_AUTO, swap_size,
190 CTLFLAG_RD, &vm_swap_size, 0, "");
195 * Red-Black tree for swblock entries
197 * The caller must hold vm_token
199 RB_GENERATE2(swblock_rb_tree, swblock, swb_entry, rb_swblock_compare,
200 vm_pindex_t, swb_index);
203 rb_swblock_compare(struct swblock *swb1, struct swblock *swb2)
205 if (swb1->swb_index < swb2->swb_index)
207 if (swb1->swb_index > swb2->swb_index)
214 rb_swblock_scancmp(struct swblock *swb, void *data)
216 struct swfreeinfo *info = data;
218 if (swb->swb_index < info->basei)
220 if (swb->swb_index > info->endi)
227 rb_swblock_condcmp(struct swblock *swb, void *data)
229 struct swfreeinfo *info = data;
231 if (swb->swb_index < info->basei)
237 * pagerops for OBJT_SWAP - "swap pager". Some ops are also global procedure
238 * calls hooked from other parts of the VM system and do not appear here.
239 * (see vm/swap_pager.h).
242 static void swap_pager_dealloc (vm_object_t object);
243 static int swap_pager_getpage (vm_object_t, vm_page_t *, int);
244 static void swap_chain_iodone(struct bio *biox);
246 struct pagerops swappagerops = {
247 swap_pager_dealloc, /* deallocate an OBJT_SWAP object */
248 swap_pager_getpage, /* pagein */
249 swap_pager_putpages, /* pageout */
250 swap_pager_haspage /* get backing store status for page */
254 * dmmax is in page-sized chunks with the new swap system. It was
255 * dev-bsized chunks in the old. dmmax is always a power of 2.
257 * swap_*() routines are externally accessible. swp_*() routines are
262 static int dmmax_mask;
263 int nswap_lowat = 128; /* in pages, swap_pager_almost_full warn */
264 int nswap_hiwat = 512; /* in pages, swap_pager_almost_full warn */
266 static __inline void swp_sizecheck (void);
267 static void swp_pager_async_iodone (struct bio *bio);
270 * Swap bitmap functions
273 static __inline void swp_pager_freeswapspace(vm_object_t object,
274 swblk_t blk, int npages);
275 static __inline swblk_t swp_pager_getswapspace(vm_object_t object, int npages);
281 static void swp_pager_meta_convert(vm_object_t);
282 static void swp_pager_meta_build(vm_object_t, vm_pindex_t, swblk_t);
283 static void swp_pager_meta_free(vm_object_t, vm_pindex_t, vm_pindex_t);
284 static void swp_pager_meta_free_all(vm_object_t);
285 static swblk_t swp_pager_meta_ctl(vm_object_t, vm_pindex_t, int);
288 * SWP_SIZECHECK() - update swap_pager_full indication
290 * update the swap_pager_almost_full indication and warn when we are
291 * about to run out of swap space, using lowat/hiwat hysteresis.
293 * Clear swap_pager_full ( task killing ) indication when lowat is met.
295 * No restrictions on call
296 * This routine may not block.
302 if (vm_swap_size < nswap_lowat) {
303 if (swap_pager_almost_full == 0) {
304 kprintf("swap_pager: out of swap space\n");
305 swap_pager_almost_full = 1;
309 if (vm_swap_size > nswap_hiwat)
310 swap_pager_almost_full = 0;
315 * SWAP_PAGER_INIT() - initialize the swap pager!
317 * Expected to be started from system init. NOTE: This code is run
318 * before much else so be careful what you depend on. Most of the VM
319 * system has yet to be initialized at this point.
321 * Called from the low level boot code only.
324 swap_pager_init(void *arg __unused)
327 * Device Stripe, in PAGE_SIZE'd blocks
329 dmmax = SWB_NPAGES * 2;
330 dmmax_mask = ~(dmmax - 1);
332 SYSINIT(vm_mem, SI_BOOT1_VM, SI_ORDER_THIRD, swap_pager_init, NULL)
335 * SWAP_PAGER_SWAP_INIT() - swap pager initialization from pageout process
337 * Expected to be started from pageout process once, prior to entering
340 * Called from the low level boot code only.
343 swap_pager_swap_init(void)
348 * Number of in-transit swap bp operations. Don't
349 * exhaust the pbufs completely. Make sure we
350 * initialize workable values (0 will work for hysteresis
351 * but it isn't very efficient).
353 * The nsw_cluster_max is constrained by the number of pages an XIO
354 * holds, i.e., (MAXPHYS/PAGE_SIZE) and our locally defined
355 * MAX_PAGEOUT_CLUSTER. Also be aware that swap ops are
356 * constrained by the swap device interleave stripe size.
358 * Currently we hardwire nsw_wcount_async to 4. This limit is
359 * designed to prevent other I/O from having high latencies due to
360 * our pageout I/O. The value 4 works well for one or two active swap
361 * devices but is probably a little low if you have more. Even so,
362 * a higher value would probably generate only a limited improvement
363 * with three or four active swap devices since the system does not
364 * typically have to pageout at extreme bandwidths. We will want
365 * at least 2 per swap devices, and 4 is a pretty good value if you
366 * have one NFS swap device due to the command/ack latency over NFS.
367 * So it all works out pretty well.
370 nsw_cluster_max = min((MAXPHYS/PAGE_SIZE), MAX_PAGEOUT_CLUSTER);
372 nsw_rcount = (nswbuf + 1) / 2;
373 nsw_wcount_sync = (nswbuf + 3) / 4;
374 nsw_wcount_async = 4;
375 nsw_wcount_async_max = nsw_wcount_async;
378 * The zone is dynamically allocated so generally size it to
379 * maxswzone (32MB to 512MB of KVM). Set a minimum size based
380 * on physical memory of around 8x (each swblock can hold 16 pages).
382 * With the advent of SSDs (vs HDs) the practical (swap:memory) ratio
383 * has increased dramatically.
385 n = vmstats.v_page_count / 2;
386 if (maxswzone && n < maxswzone / sizeof(struct swblock))
387 n = maxswzone / sizeof(struct swblock);
393 sizeof(struct swblock),
397 if (swap_zone != NULL)
400 * if the allocation failed, try a zone two thirds the
401 * size of the previous attempt.
406 if (swap_zone == NULL)
407 panic("swap_pager_swap_init: swap_zone == NULL");
409 kprintf("Swap zone entries reduced from %d to %d.\n", n2, n);
413 * SWAP_PAGER_ALLOC() - allocate a new OBJT_SWAP VM object and instantiate
414 * its metadata structures.
416 * This routine is called from the mmap and fork code to create a new
417 * OBJT_SWAP object. We do this by creating an OBJT_DEFAULT object
418 * and then converting it with swp_pager_meta_convert().
420 * We only support unnamed objects.
425 swap_pager_alloc(void *handle, off_t size, vm_prot_t prot, off_t offset)
429 KKASSERT(handle == NULL);
430 object = vm_object_allocate_hold(OBJT_DEFAULT,
431 OFF_TO_IDX(offset + PAGE_MASK + size));
432 swp_pager_meta_convert(object);
433 vm_object_drop(object);
439 * SWAP_PAGER_DEALLOC() - remove swap metadata from object
441 * The swap backing for the object is destroyed. The code is
442 * designed such that we can reinstantiate it later, but this
443 * routine is typically called only when the entire object is
444 * about to be destroyed.
446 * The object must be locked or unreferenceable.
447 * No other requirements.
450 swap_pager_dealloc(vm_object_t object)
452 vm_object_hold(object);
453 vm_object_pip_wait(object, "swpdea");
456 * Free all remaining metadata. We only bother to free it from
457 * the swap meta data. We do not attempt to free swapblk's still
458 * associated with vm_page_t's for this object. We do not care
459 * if paging is still in progress on some objects.
461 swp_pager_meta_free_all(object);
462 vm_object_drop(object);
465 /************************************************************************
466 * SWAP PAGER BITMAP ROUTINES *
467 ************************************************************************/
470 * SWP_PAGER_GETSWAPSPACE() - allocate raw swap space
472 * Allocate swap for the requested number of pages. The starting
473 * swap block number (a page index) is returned or SWAPBLK_NONE
474 * if the allocation failed.
476 * Also has the side effect of advising that somebody made a mistake
477 * when they configured swap and didn't configure enough.
479 * The caller must hold the object.
480 * This routine may not block.
482 static __inline swblk_t
483 swp_pager_getswapspace(vm_object_t object, int npages)
487 lwkt_gettoken(&vm_token);
488 blk = blist_allocat(swapblist, npages, swapiterator);
489 if (blk == SWAPBLK_NONE)
490 blk = blist_allocat(swapblist, npages, 0);
491 if (blk == SWAPBLK_NONE) {
492 if (swap_pager_full != 2) {
493 kprintf("swap_pager_getswapspace: failed alloc=%d\n",
496 swap_pager_almost_full = 1;
500 swapacctspace(blk, -npages);
501 if (object->type == OBJT_SWAP)
502 vm_swap_anon_use += npages;
504 vm_swap_cache_use += npages;
507 lwkt_reltoken(&vm_token);
512 * SWP_PAGER_FREESWAPSPACE() - free raw swap space
514 * This routine returns the specified swap blocks back to the bitmap.
516 * Note: This routine may not block (it could in the old swap code),
517 * and through the use of the new blist routines it does not block.
519 * We must be called at splvm() to avoid races with bitmap frees from
520 * vm_page_remove() aka swap_pager_page_removed().
522 * This routine may not block.
526 swp_pager_freeswapspace(vm_object_t object, swblk_t blk, int npages)
528 struct swdevt *sp = &swdevt[BLK2DEVIDX(blk)];
530 lwkt_gettoken(&vm_token);
531 sp->sw_nused -= npages;
532 if (object->type == OBJT_SWAP)
533 vm_swap_anon_use -= npages;
535 vm_swap_cache_use -= npages;
537 if (sp->sw_flags & SW_CLOSING) {
538 lwkt_reltoken(&vm_token);
542 blist_free(swapblist, blk, npages);
543 vm_swap_size += npages;
545 lwkt_reltoken(&vm_token);
549 * SWAP_PAGER_FREESPACE() - frees swap blocks associated with a page
550 * range within an object.
552 * This is a globally accessible routine.
554 * This routine removes swapblk assignments from swap metadata.
556 * The external callers of this routine typically have already destroyed
557 * or renamed vm_page_t's associated with this range in the object so
563 swap_pager_freespace(vm_object_t object, vm_pindex_t start, vm_pindex_t size)
565 vm_object_hold(object);
566 swp_pager_meta_free(object, start, size);
567 vm_object_drop(object);
574 swap_pager_freespace_all(vm_object_t object)
576 vm_object_hold(object);
577 swp_pager_meta_free_all(object);
578 vm_object_drop(object);
582 * This function conditionally frees swap cache swap starting at
583 * (*basei) in the object. (count) swap blocks will be nominally freed.
584 * The actual number of blocks freed can be more or less than the
587 * This function nominally returns the number of blocks freed. However,
588 * the actual number of blocks freed may be less then the returned value.
589 * If the function is unable to exhaust the object or if it is able to
590 * free (approximately) the requested number of blocks it returns
593 * If we exhaust the object we will return a value n <= count.
595 * The caller must hold the object.
597 * WARNING! If count == 0 then -1 can be returned as a degenerate case,
598 * callers should always pass a count value > 0.
600 static int swap_pager_condfree_callback(struct swblock *swap, void *data);
603 swap_pager_condfree(vm_object_t object, vm_pindex_t *basei, int count)
605 struct swfreeinfo info;
609 ASSERT_LWKT_TOKEN_HELD(vm_object_token(object));
611 info.object = object;
612 info.basei = *basei; /* skip up to this page index */
613 info.begi = count; /* max swap pages to destroy */
614 info.endi = count * 8; /* max swblocks to scan */
616 swblock_rb_tree_RB_SCAN(&object->swblock_root, rb_swblock_condcmp,
617 swap_pager_condfree_callback, &info);
621 * Take the higher difference swblocks vs pages
623 n = count - (int)info.begi;
624 t = count * 8 - (int)info.endi;
633 * The idea is to free whole meta-block to avoid fragmenting
634 * the swap space or disk I/O. We only do this if NO VM pages
637 * We do not have to deal with clearing PG_SWAPPED in related VM
638 * pages because there are no related VM pages.
640 * The caller must hold the object.
643 swap_pager_condfree_callback(struct swblock *swap, void *data)
645 struct swfreeinfo *info = data;
646 vm_object_t object = info->object;
649 for (i = 0; i < SWAP_META_PAGES; ++i) {
650 if (vm_page_lookup(object, swap->swb_index + i))
653 info->basei = swap->swb_index + SWAP_META_PAGES;
654 if (i == SWAP_META_PAGES) {
655 info->begi -= swap->swb_count;
656 swap_pager_freespace(object, swap->swb_index, SWAP_META_PAGES);
659 if ((int)info->begi < 0 || (int)info->endi < 0)
666 * Called by vm_page_alloc() when a new VM page is inserted
667 * into a VM object. Checks whether swap has been assigned to
668 * the page and sets PG_SWAPPED as necessary.
673 swap_pager_page_inserted(vm_page_t m)
675 if (m->object->swblock_count) {
676 vm_object_hold(m->object);
677 if (swp_pager_meta_ctl(m->object, m->pindex, 0) != SWAPBLK_NONE)
678 vm_page_flag_set(m, PG_SWAPPED);
679 vm_object_drop(m->object);
684 * SWAP_PAGER_RESERVE() - reserve swap blocks in object
686 * Assigns swap blocks to the specified range within the object. The
687 * swap blocks are not zerod. Any previous swap assignment is destroyed.
689 * Returns 0 on success, -1 on failure.
691 * The caller is responsible for avoiding races in the specified range.
692 * No other requirements.
695 swap_pager_reserve(vm_object_t object, vm_pindex_t start, vm_size_t size)
698 swblk_t blk = SWAPBLK_NONE;
699 vm_pindex_t beg = start; /* save start index */
701 vm_object_hold(object);
706 while ((blk = swp_pager_getswapspace(object, n)) ==
711 swp_pager_meta_free(object, beg,
713 vm_object_drop(object);
718 swp_pager_meta_build(object, start, blk);
724 swp_pager_meta_free(object, start, n);
725 vm_object_drop(object);
730 * SWAP_PAGER_COPY() - copy blocks from source pager to destination pager
731 * and destroy the source.
733 * Copy any valid swapblks from the source to the destination. In
734 * cases where both the source and destination have a valid swapblk,
735 * we keep the destination's.
737 * This routine is allowed to block. It may block allocating metadata
738 * indirectly through swp_pager_meta_build() or if paging is still in
739 * progress on the source.
741 * XXX vm_page_collapse() kinda expects us not to block because we
742 * supposedly do not need to allocate memory, but for the moment we
743 * *may* have to get a little memory from the zone allocator, but
744 * it is taken from the interrupt memory. We should be ok.
746 * The source object contains no vm_page_t's (which is just as well)
747 * The source object is of type OBJT_SWAP.
749 * The source and destination objects must be held by the caller.
752 swap_pager_copy(vm_object_t srcobject, vm_object_t dstobject,
753 vm_pindex_t base_index, int destroysource)
757 ASSERT_LWKT_TOKEN_HELD(vm_object_token(srcobject));
758 ASSERT_LWKT_TOKEN_HELD(vm_object_token(dstobject));
761 * transfer source to destination.
763 for (i = 0; i < dstobject->size; ++i) {
767 * Locate (without changing) the swapblk on the destination,
768 * unless it is invalid in which case free it silently, or
769 * if the destination is a resident page, in which case the
770 * source is thrown away.
772 dstaddr = swp_pager_meta_ctl(dstobject, i, 0);
774 if (dstaddr == SWAPBLK_NONE) {
776 * Destination has no swapblk and is not resident,
781 srcaddr = swp_pager_meta_ctl(srcobject,
782 base_index + i, SWM_POP);
784 if (srcaddr != SWAPBLK_NONE)
785 swp_pager_meta_build(dstobject, i, srcaddr);
788 * Destination has valid swapblk or it is represented
789 * by a resident page. We destroy the sourceblock.
791 swp_pager_meta_ctl(srcobject, base_index + i, SWM_FREE);
796 * Free left over swap blocks in source.
798 * We have to revert the type to OBJT_DEFAULT so we do not accidently
799 * double-remove the object from the swap queues.
803 * Reverting the type is not necessary, the caller is going
804 * to destroy srcobject directly, but I'm doing it here
805 * for consistency since we've removed the object from its
808 swp_pager_meta_free_all(srcobject);
809 if (srcobject->type == OBJT_SWAP)
810 srcobject->type = OBJT_DEFAULT;
815 * SWAP_PAGER_HASPAGE() - determine if we have good backing store for
816 * the requested page.
818 * We determine whether good backing store exists for the requested
819 * page and return TRUE if it does, FALSE if it doesn't.
821 * If TRUE, we also try to determine how much valid, contiguous backing
822 * store exists before and after the requested page within a reasonable
823 * distance. We do not try to restrict it to the swap device stripe
824 * (that is handled in getpages/putpages). It probably isn't worth
830 swap_pager_haspage(vm_object_t object, vm_pindex_t pindex)
835 * do we have good backing store at the requested index ?
837 vm_object_hold(object);
838 blk0 = swp_pager_meta_ctl(object, pindex, 0);
840 if (blk0 == SWAPBLK_NONE) {
841 vm_object_drop(object);
844 vm_object_drop(object);
849 * SWAP_PAGER_PAGE_UNSWAPPED() - remove swap backing store related to page
851 * This removes any associated swap backing store, whether valid or
852 * not, from the page. This operates on any VM object, not just OBJT_SWAP
855 * This routine is typically called when a page is made dirty, at
856 * which point any associated swap can be freed. MADV_FREE also
857 * calls us in a special-case situation
859 * NOTE!!! If the page is clean and the swap was valid, the caller
860 * should make the page dirty before calling this routine. This routine
861 * does NOT change the m->dirty status of the page. Also: MADV_FREE
864 * The page must be busied or soft-busied.
865 * The caller can hold the object to avoid blocking, else we might block.
866 * No other requirements.
869 swap_pager_unswapped(vm_page_t m)
871 if (m->flags & PG_SWAPPED) {
872 vm_object_hold(m->object);
873 KKASSERT(m->flags & PG_SWAPPED);
874 swp_pager_meta_ctl(m->object, m->pindex, SWM_FREE);
875 vm_page_flag_clear(m, PG_SWAPPED);
876 vm_object_drop(m->object);
881 * SWAP_PAGER_STRATEGY() - read, write, free blocks
883 * This implements a VM OBJECT strategy function using swap backing store.
884 * This can operate on any VM OBJECT type, not necessarily just OBJT_SWAP
887 * This is intended to be a cacheless interface (i.e. caching occurs at
888 * higher levels), and is also used as a swap-based SSD cache for vnode
889 * and device objects.
891 * All I/O goes directly to and from the swap device.
893 * We currently attempt to run I/O synchronously or asynchronously as
894 * the caller requests. This isn't perfect because we loose error
895 * sequencing when we run multiple ops in parallel to satisfy a request.
896 * But this is swap, so we let it all hang out.
901 swap_pager_strategy(vm_object_t object, struct bio *bio)
903 struct buf *bp = bio->bio_buf;
906 vm_pindex_t biox_blkno = 0;
912 struct bio_track *track;
917 * tracking for swapdev vnode I/Os
919 if (bp->b_cmd == BUF_CMD_READ)
920 track = &swapdev_vp->v_track_read;
922 track = &swapdev_vp->v_track_write;
925 if (bp->b_bcount & PAGE_MASK) {
926 bp->b_error = EINVAL;
927 bp->b_flags |= B_ERROR | B_INVAL;
929 kprintf("swap_pager_strategy: bp %p offset %lld size %d, "
930 "not page bounded\n",
931 bp, (long long)bio->bio_offset, (int)bp->b_bcount);
936 * Clear error indication, initialize page index, count, data pointer.
939 bp->b_flags &= ~B_ERROR;
940 bp->b_resid = bp->b_bcount;
942 start = (vm_pindex_t)(bio->bio_offset >> PAGE_SHIFT);
943 count = howmany(bp->b_bcount, PAGE_SIZE);
947 * Deal with BUF_CMD_FREEBLKS
949 if (bp->b_cmd == BUF_CMD_FREEBLKS) {
951 * FREE PAGE(s) - destroy underlying swap that is no longer
954 vm_object_hold(object);
955 swp_pager_meta_free(object, start, count);
956 vm_object_drop(object);
963 * We need to be able to create a new cluster of I/O's. We cannot
964 * use the caller fields of the passed bio so push a new one.
966 * Because nbio is just a placeholder for the cluster links,
967 * we can biodone() the original bio instead of nbio to make
968 * things a bit more efficient.
970 nbio = push_bio(bio);
971 nbio->bio_offset = bio->bio_offset;
972 nbio->bio_caller_info1.cluster_head = NULL;
973 nbio->bio_caller_info2.cluster_tail = NULL;
979 * Execute read or write
981 vm_object_hold(object);
987 * Obtain block. If block not found and writing, allocate a
988 * new block and build it into the object.
990 blk = swp_pager_meta_ctl(object, start, 0);
991 if ((blk == SWAPBLK_NONE) && bp->b_cmd != BUF_CMD_READ) {
992 blk = swp_pager_getswapspace(object, 1);
993 if (blk == SWAPBLK_NONE) {
994 bp->b_error = ENOMEM;
995 bp->b_flags |= B_ERROR;
998 swp_pager_meta_build(object, start, blk);
1002 * Do we have to flush our current collection? Yes if:
1004 * - no swap block at this index
1005 * - swap block is not contiguous
1006 * - we cross a physical disk boundry in the
1010 biox && (biox_blkno + btoc(bufx->b_bcount) != blk ||
1011 ((biox_blkno ^ blk) & dmmax_mask)
1014 if (bp->b_cmd == BUF_CMD_READ) {
1015 ++mycpu->gd_cnt.v_swapin;
1016 mycpu->gd_cnt.v_swappgsin += btoc(bufx->b_bcount);
1018 ++mycpu->gd_cnt.v_swapout;
1019 mycpu->gd_cnt.v_swappgsout += btoc(bufx->b_bcount);
1020 bufx->b_dirtyend = bufx->b_bcount;
1024 * Finished with this buf.
1026 KKASSERT(bufx->b_bcount != 0);
1027 if (bufx->b_cmd != BUF_CMD_READ)
1028 bufx->b_dirtyend = bufx->b_bcount;
1034 * Add new swapblk to biox, instantiating biox if necessary.
1035 * Zero-fill reads are able to take a shortcut.
1037 if (blk == SWAPBLK_NONE) {
1039 * We can only get here if we are reading. Since
1040 * we are at splvm() we can safely modify b_resid,
1041 * even if chain ops are in progress.
1043 bzero(data, PAGE_SIZE);
1044 bp->b_resid -= PAGE_SIZE;
1047 /* XXX chain count > 4, wait to <= 4 */
1049 bufx = getpbuf(NULL);
1050 biox = &bufx->b_bio1;
1051 cluster_append(nbio, bufx);
1052 bufx->b_flags |= (bp->b_flags & B_ORDERED);
1053 bufx->b_cmd = bp->b_cmd;
1054 biox->bio_done = swap_chain_iodone;
1055 biox->bio_offset = (off_t)blk << PAGE_SHIFT;
1056 biox->bio_caller_info1.cluster_parent = nbio;
1059 bufx->b_data = data;
1061 bufx->b_bcount += PAGE_SIZE;
1068 vm_object_drop(object);
1071 * Flush out last buffer
1074 if (bufx->b_cmd == BUF_CMD_READ) {
1075 ++mycpu->gd_cnt.v_swapin;
1076 mycpu->gd_cnt.v_swappgsin += btoc(bufx->b_bcount);
1078 ++mycpu->gd_cnt.v_swapout;
1079 mycpu->gd_cnt.v_swappgsout += btoc(bufx->b_bcount);
1080 bufx->b_dirtyend = bufx->b_bcount;
1082 KKASSERT(bufx->b_bcount);
1083 if (bufx->b_cmd != BUF_CMD_READ)
1084 bufx->b_dirtyend = bufx->b_bcount;
1085 /* biox, bufx = NULL */
1089 * Now initiate all the I/O. Be careful looping on our chain as
1090 * I/O's may complete while we are still initiating them.
1092 * If the request is a 100% sparse read no bios will be present
1093 * and we just biodone() the buffer.
1095 nbio->bio_caller_info2.cluster_tail = NULL;
1096 bufx = nbio->bio_caller_info1.cluster_head;
1100 biox = &bufx->b_bio1;
1102 bufx = bufx->b_cluster_next;
1103 vn_strategy(swapdev_vp, biox);
1110 * Completion of the cluster will also call biodone_chain(nbio).
1111 * We never call biodone(nbio) so we don't have to worry about
1112 * setting up a bio_done callback. It's handled in the sub-IO.
1123 swap_chain_iodone(struct bio *biox)
1126 struct buf *bufx; /* chained sub-buffer */
1127 struct bio *nbio; /* parent nbio with chain glue */
1128 struct buf *bp; /* original bp associated with nbio */
1131 bufx = biox->bio_buf;
1132 nbio = biox->bio_caller_info1.cluster_parent;
1136 * Update the original buffer
1138 KKASSERT(bp != NULL);
1139 if (bufx->b_flags & B_ERROR) {
1140 atomic_set_int(&bufx->b_flags, B_ERROR);
1141 bp->b_error = bufx->b_error; /* race ok */
1142 } else if (bufx->b_resid != 0) {
1143 atomic_set_int(&bufx->b_flags, B_ERROR);
1144 bp->b_error = EINVAL; /* race ok */
1146 atomic_subtract_int(&bp->b_resid, bufx->b_bcount);
1150 * Remove us from the chain.
1152 spin_lock(&bp->b_lock.lk_spinlock);
1153 nextp = &nbio->bio_caller_info1.cluster_head;
1154 while (*nextp != bufx) {
1155 KKASSERT(*nextp != NULL);
1156 nextp = &(*nextp)->b_cluster_next;
1158 *nextp = bufx->b_cluster_next;
1159 chain_empty = (nbio->bio_caller_info1.cluster_head == NULL);
1160 spin_unlock(&bp->b_lock.lk_spinlock);
1163 * Clean up bufx. If the chain is now empty we finish out
1164 * the parent. Note that we may be racing other completions
1165 * so we must use the chain_empty status from above.
1168 if (bp->b_resid != 0 && !(bp->b_flags & B_ERROR)) {
1169 atomic_set_int(&bp->b_flags, B_ERROR);
1170 bp->b_error = EINVAL;
1172 biodone_chain(nbio);
1174 relpbuf(bufx, NULL);
1178 * SWAP_PAGER_GETPAGES() - bring page in from swap
1180 * The requested page may have to be brought in from swap. Calculate the
1181 * swap block and bring in additional pages if possible. All pages must
1182 * have contiguous swap block assignments and reside in the same object.
1184 * The caller has a single vm_object_pip_add() reference prior to
1185 * calling us and we should return with the same.
1187 * The caller has BUSY'd the page. We should return with (*mpp) left busy,
1188 * and any additinal pages unbusied.
1190 * If the caller encounters a PG_RAM page it will pass it to us even though
1191 * it may be valid and dirty. We cannot overwrite the page in this case!
1192 * The case is used to allow us to issue pure read-aheads.
1194 * NOTE! XXX This code does not entirely pipeline yet due to the fact that
1195 * the PG_RAM page is validated at the same time as mreq. What we
1196 * really need to do is issue a separate read-ahead pbuf.
1201 swap_pager_getpage(vm_object_t object, vm_page_t *mpp, int seqaccess)
1214 vm_page_t marray[XIO_INTERNAL_PAGES];
1218 vm_object_hold(object);
1219 if (mreq->object != object) {
1220 panic("swap_pager_getpages: object mismatch %p/%p",
1227 * We don't want to overwrite a fully valid page as it might be
1228 * dirty. This case can occur when e.g. vm_fault hits a perfectly
1229 * valid page with PG_RAM set.
1231 * In this case we see if the next page is a suitable page-in
1232 * candidate and if it is we issue read-ahead. PG_RAM will be
1233 * set on the last page of the read-ahead to continue the pipeline.
1235 if (mreq->valid == VM_PAGE_BITS_ALL) {
1236 if (swap_burst_read == 0 || mreq->pindex + 1 >= object->size) {
1237 vm_object_drop(object);
1238 return(VM_PAGER_OK);
1240 blk = swp_pager_meta_ctl(object, mreq->pindex + 1, 0);
1241 if (blk == SWAPBLK_NONE) {
1242 vm_object_drop(object);
1243 return(VM_PAGER_OK);
1245 m = vm_page_lookup_busy_try(object, mreq->pindex + 1,
1248 vm_object_drop(object);
1249 return(VM_PAGER_OK);
1250 } else if (m == NULL) {
1252 * Use VM_ALLOC_QUICK to avoid blocking on cache
1255 m = vm_page_alloc(object, mreq->pindex + 1,
1258 vm_object_drop(object);
1259 return(VM_PAGER_OK);
1264 vm_object_drop(object);
1265 return(VM_PAGER_OK);
1267 vm_page_unqueue_nowakeup(m);
1277 * Try to block-read contiguous pages from swap if sequential,
1278 * otherwise just read one page. Contiguous pages from swap must
1279 * reside within a single device stripe because the I/O cannot be
1280 * broken up across multiple stripes.
1282 * Note that blk and iblk can be SWAPBLK_NONE but the loop is
1283 * set up such that the case(s) are handled implicitly.
1285 blk = swp_pager_meta_ctl(mreq->object, mreq->pindex, 0);
1288 for (i = 1; swap_burst_read &&
1289 i < XIO_INTERNAL_PAGES &&
1290 mreq->pindex + i < object->size; ++i) {
1293 iblk = swp_pager_meta_ctl(object, mreq->pindex + i, 0);
1294 if (iblk != blk + i)
1296 if ((blk ^ iblk) & dmmax_mask)
1298 m = vm_page_lookup_busy_try(object, mreq->pindex + i,
1302 } else if (m == NULL) {
1304 * Use VM_ALLOC_QUICK to avoid blocking on cache
1307 m = vm_page_alloc(object, mreq->pindex + i,
1316 vm_page_unqueue_nowakeup(m);
1322 vm_page_flag_set(marray[i - 1], PG_RAM);
1325 * If mreq is the requested page and we have nothing to do return
1326 * VM_PAGER_FAIL. If raonly is set mreq is just another read-ahead
1327 * page and must be cleaned up.
1329 if (blk == SWAPBLK_NONE) {
1332 vnode_pager_freepage(mreq);
1333 vm_object_drop(object);
1334 return(VM_PAGER_OK);
1336 vm_object_drop(object);
1337 return(VM_PAGER_FAIL);
1342 * map our page(s) into kva for input
1344 bp = getpbuf_kva(&nsw_rcount);
1346 kva = (vm_offset_t) bp->b_kvabase;
1347 bcopy(marray, bp->b_xio.xio_pages, i * sizeof(vm_page_t));
1348 pmap_qenter(kva, bp->b_xio.xio_pages, i);
1350 bp->b_data = (caddr_t)kva;
1351 bp->b_bcount = PAGE_SIZE * i;
1352 bp->b_xio.xio_npages = i;
1353 bio->bio_done = swp_pager_async_iodone;
1354 bio->bio_offset = (off_t)blk << PAGE_SHIFT;
1355 bio->bio_caller_info1.index = SWBIO_READ;
1358 * Set index. If raonly set the index beyond the array so all
1359 * the pages are treated the same, otherwise the original mreq is
1363 bio->bio_driver_info = (void *)(intptr_t)i;
1365 bio->bio_driver_info = (void *)(intptr_t)0;
1367 for (j = 0; j < i; ++j)
1368 vm_page_flag_set(bp->b_xio.xio_pages[j], PG_SWAPINPROG);
1370 mycpu->gd_cnt.v_swapin++;
1371 mycpu->gd_cnt.v_swappgsin += bp->b_xio.xio_npages;
1374 * We still hold the lock on mreq, and our automatic completion routine
1375 * does not remove it.
1377 vm_object_pip_add(object, bp->b_xio.xio_npages);
1380 * perform the I/O. NOTE!!! bp cannot be considered valid after
1381 * this point because we automatically release it on completion.
1382 * Instead, we look at the one page we are interested in which we
1383 * still hold a lock on even through the I/O completion.
1385 * The other pages in our m[] array are also released on completion,
1386 * so we cannot assume they are valid anymore either.
1388 bp->b_cmd = BUF_CMD_READ;
1390 vn_strategy(swapdev_vp, bio);
1393 * Wait for the page we want to complete. PG_SWAPINPROG is always
1394 * cleared on completion. If an I/O error occurs, SWAPBLK_NONE
1395 * is set in the meta-data.
1397 * If this is a read-ahead only we return immediately without
1401 vm_object_drop(object);
1402 return(VM_PAGER_OK);
1406 * Read-ahead includes originally requested page case.
1409 flags = mreq->flags;
1411 if ((flags & PG_SWAPINPROG) == 0)
1413 tsleep_interlock(mreq, 0);
1414 if (!atomic_cmpset_int(&mreq->flags, flags,
1415 flags | PG_WANTED | PG_REFERENCED)) {
1418 mycpu->gd_cnt.v_intrans++;
1419 if (tsleep(mreq, PINTERLOCKED, "swread", hz*20)) {
1421 "swap_pager: indefinite wait buffer: "
1422 " offset: %lld, size: %ld\n",
1423 (long long)bio->bio_offset,
1430 * mreq is left bussied after completion, but all the other pages
1431 * are freed. If we had an unrecoverable read error the page will
1434 vm_object_drop(object);
1435 if (mreq->valid != VM_PAGE_BITS_ALL)
1436 return(VM_PAGER_ERROR);
1438 return(VM_PAGER_OK);
1441 * A final note: in a low swap situation, we cannot deallocate swap
1442 * and mark a page dirty here because the caller is likely to mark
1443 * the page clean when we return, causing the page to possibly revert
1444 * to all-zero's later.
1449 * swap_pager_putpages:
1451 * Assign swap (if necessary) and initiate I/O on the specified pages.
1453 * We support both OBJT_DEFAULT and OBJT_SWAP objects. DEFAULT objects
1454 * are automatically converted to SWAP objects.
1456 * In a low memory situation we may block in vn_strategy(), but the new
1457 * vm_page reservation system coupled with properly written VFS devices
1458 * should ensure that no low-memory deadlock occurs. This is an area
1461 * The parent has N vm_object_pip_add() references prior to
1462 * calling us and will remove references for rtvals[] that are
1463 * not set to VM_PAGER_PEND. We need to remove the rest on I/O
1466 * The parent has soft-busy'd the pages it passes us and will unbusy
1467 * those whos rtvals[] entry is not set to VM_PAGER_PEND on return.
1468 * We need to unbusy the rest on I/O completion.
1473 swap_pager_putpages(vm_object_t object, vm_page_t *m, int count,
1474 boolean_t sync, int *rtvals)
1479 vm_object_hold(object);
1481 if (count && m[0]->object != object) {
1482 panic("swap_pager_getpages: object mismatch %p/%p",
1491 * Turn object into OBJT_SWAP
1492 * check for bogus sysops
1493 * force sync if not pageout process
1495 if (object->type == OBJT_DEFAULT) {
1496 if (object->type == OBJT_DEFAULT)
1497 swp_pager_meta_convert(object);
1500 if (curthread != pagethread)
1506 * Update nsw parameters from swap_async_max sysctl values.
1507 * Do not let the sysop crash the machine with bogus numbers.
1509 if (swap_async_max != nsw_wcount_async_max) {
1515 if ((n = swap_async_max) > nswbuf / 2)
1522 * Adjust difference ( if possible ). If the current async
1523 * count is too low, we may not be able to make the adjustment
1526 * vm_token needed for nsw_wcount sleep interlock
1528 lwkt_gettoken(&vm_token);
1529 n -= nsw_wcount_async_max;
1530 if (nsw_wcount_async + n >= 0) {
1531 nsw_wcount_async_max += n;
1532 pbuf_adjcount(&nsw_wcount_async, n);
1534 lwkt_reltoken(&vm_token);
1540 * Assign swap blocks and issue I/O. We reallocate swap on the fly.
1541 * The page is left dirty until the pageout operation completes
1545 for (i = 0; i < count; i += n) {
1552 * Maximum I/O size is limited by a number of factors.
1555 n = min(BLIST_MAX_ALLOC, count - i);
1556 n = min(n, nsw_cluster_max);
1558 lwkt_gettoken(&vm_token);
1561 * Get biggest block of swap we can. If we fail, fall
1562 * back and try to allocate a smaller block. Don't go
1563 * overboard trying to allocate space if it would overly
1567 (blk = swp_pager_getswapspace(object, n)) == SWAPBLK_NONE &&
1572 if (blk == SWAPBLK_NONE) {
1573 for (j = 0; j < n; ++j)
1574 rtvals[i+j] = VM_PAGER_FAIL;
1575 lwkt_reltoken(&vm_token);
1580 * The I/O we are constructing cannot cross a physical
1581 * disk boundry in the swap stripe. Note: we are still
1584 if ((blk ^ (blk + n)) & dmmax_mask) {
1585 j = ((blk + dmmax) & dmmax_mask) - blk;
1586 swp_pager_freeswapspace(object, blk + j, n - j);
1591 * All I/O parameters have been satisfied, build the I/O
1592 * request and assign the swap space.
1595 bp = getpbuf_kva(&nsw_wcount_sync);
1597 bp = getpbuf_kva(&nsw_wcount_async);
1600 lwkt_reltoken(&vm_token);
1602 pmap_qenter((vm_offset_t)bp->b_data, &m[i], n);
1604 bp->b_bcount = PAGE_SIZE * n;
1605 bio->bio_offset = (off_t)blk << PAGE_SHIFT;
1607 for (j = 0; j < n; ++j) {
1608 vm_page_t mreq = m[i+j];
1610 swp_pager_meta_build(mreq->object, mreq->pindex,
1612 if (object->type == OBJT_SWAP)
1613 vm_page_dirty(mreq);
1614 rtvals[i+j] = VM_PAGER_OK;
1616 vm_page_flag_set(mreq, PG_SWAPINPROG);
1617 bp->b_xio.xio_pages[j] = mreq;
1619 bp->b_xio.xio_npages = n;
1621 mycpu->gd_cnt.v_swapout++;
1622 mycpu->gd_cnt.v_swappgsout += bp->b_xio.xio_npages;
1624 bp->b_dirtyoff = 0; /* req'd for NFS */
1625 bp->b_dirtyend = bp->b_bcount; /* req'd for NFS */
1626 bp->b_cmd = BUF_CMD_WRITE;
1627 bio->bio_caller_info1.index = SWBIO_WRITE;
1632 if (sync == FALSE) {
1633 bio->bio_done = swp_pager_async_iodone;
1635 vn_strategy(swapdev_vp, bio);
1637 for (j = 0; j < n; ++j)
1638 rtvals[i+j] = VM_PAGER_PEND;
1643 * Issue synchrnously.
1645 * Wait for the sync I/O to complete, then update rtvals.
1646 * We just set the rtvals[] to VM_PAGER_PEND so we can call
1647 * our async completion routine at the end, thus avoiding a
1650 bio->bio_caller_info1.index |= SWBIO_SYNC;
1651 bio->bio_done = biodone_sync;
1652 bio->bio_flags |= BIO_SYNC;
1653 vn_strategy(swapdev_vp, bio);
1654 biowait(bio, "swwrt");
1656 for (j = 0; j < n; ++j)
1657 rtvals[i+j] = VM_PAGER_PEND;
1660 * Now that we are through with the bp, we can call the
1661 * normal async completion, which frees everything up.
1663 swp_pager_async_iodone(bio);
1665 vm_object_drop(object);
1672 swap_pager_newswap(void)
1678 * swp_pager_async_iodone:
1680 * Completion routine for asynchronous reads and writes from/to swap.
1681 * Also called manually by synchronous code to finish up a bp.
1683 * For READ operations, the pages are PG_BUSY'd. For WRITE operations,
1684 * the pages are vm_page_t->busy'd. For READ operations, we PG_BUSY
1685 * unbusy all pages except the 'main' request page. For WRITE
1686 * operations, we vm_page_t->busy'd unbusy all pages ( we can do this
1687 * because we marked them all VM_PAGER_PEND on return from putpages ).
1689 * This routine may not block.
1694 swp_pager_async_iodone(struct bio *bio)
1696 struct buf *bp = bio->bio_buf;
1697 vm_object_t object = NULL;
1704 if (bp->b_flags & B_ERROR) {
1706 "swap_pager: I/O error - %s failed; offset %lld,"
1707 "size %ld, error %d\n",
1708 ((bio->bio_caller_info1.index & SWBIO_READ) ?
1709 "pagein" : "pageout"),
1710 (long long)bio->bio_offset,
1717 * set object, raise to splvm().
1719 if (bp->b_xio.xio_npages)
1720 object = bp->b_xio.xio_pages[0]->object;
1723 * remove the mapping for kernel virtual
1725 pmap_qremove((vm_offset_t)bp->b_data, bp->b_xio.xio_npages);
1728 * cleanup pages. If an error occurs writing to swap, we are in
1729 * very serious trouble. If it happens to be a disk error, though,
1730 * we may be able to recover by reassigning the swap later on. So
1731 * in this case we remove the m->swapblk assignment for the page
1732 * but do not free it in the rlist. The errornous block(s) are thus
1733 * never reallocated as swap. Redirty the page and continue.
1735 for (i = 0; i < bp->b_xio.xio_npages; ++i) {
1736 vm_page_t m = bp->b_xio.xio_pages[i];
1738 if (bp->b_flags & B_ERROR) {
1740 * If an error occurs I'd love to throw the swapblk
1741 * away without freeing it back to swapspace, so it
1742 * can never be used again. But I can't from an
1746 if (bio->bio_caller_info1.index & SWBIO_READ) {
1748 * When reading, reqpage needs to stay
1749 * locked for the parent, but all other
1750 * pages can be freed. We still want to
1751 * wakeup the parent waiting on the page,
1752 * though. ( also: pg_reqpage can be -1 and
1753 * not match anything ).
1755 * We have to wake specifically requested pages
1756 * up too because we cleared PG_SWAPINPROG and
1757 * someone may be waiting for that.
1759 * NOTE: for reads, m->dirty will probably
1760 * be overridden by the original caller of
1761 * getpages so don't play cute tricks here.
1763 * NOTE: We can't actually free the page from
1764 * here, because this is an interrupt. It
1765 * is not legal to mess with object->memq
1766 * from an interrupt. Deactivate the page
1771 vm_page_flag_clear(m, PG_ZERO);
1772 vm_page_flag_clear(m, PG_SWAPINPROG);
1775 * bio_driver_info holds the requested page
1778 if (i != (int)(intptr_t)bio->bio_driver_info) {
1779 vm_page_deactivate(m);
1785 * If i == bp->b_pager.pg_reqpage, do not wake
1786 * the page up. The caller needs to.
1790 * If a write error occurs remove the swap
1791 * assignment (note that PG_SWAPPED may or
1792 * may not be set depending on prior activity).
1794 * Re-dirty OBJT_SWAP pages as there is no
1795 * other backing store, we can't throw the
1798 * Non-OBJT_SWAP pages (aka swapcache) must
1799 * not be dirtied since they may not have
1800 * been dirty in the first place, and they
1801 * do have backing store (the vnode).
1803 vm_page_busy_wait(m, FALSE, "swadpg");
1804 swp_pager_meta_ctl(m->object, m->pindex,
1806 vm_page_flag_clear(m, PG_SWAPPED);
1807 if (m->object->type == OBJT_SWAP) {
1809 vm_page_activate(m);
1811 vm_page_flag_clear(m, PG_SWAPINPROG);
1812 vm_page_io_finish(m);
1815 } else if (bio->bio_caller_info1.index & SWBIO_READ) {
1817 * NOTE: for reads, m->dirty will probably be
1818 * overridden by the original caller of getpages so
1819 * we cannot set them in order to free the underlying
1820 * swap in a low-swap situation. I don't think we'd
1821 * want to do that anyway, but it was an optimization
1822 * that existed in the old swapper for a time before
1823 * it got ripped out due to precisely this problem.
1825 * clear PG_ZERO in page.
1827 * If not the requested page then deactivate it.
1829 * Note that the requested page, reqpage, is left
1830 * busied, but we still have to wake it up. The
1831 * other pages are released (unbusied) by
1832 * vm_page_wakeup(). We do not set reqpage's
1833 * valid bits here, it is up to the caller.
1837 * NOTE: can't call pmap_clear_modify(m) from an
1838 * interrupt thread, the pmap code may have to map
1839 * non-kernel pmaps and currently asserts the case.
1841 /*pmap_clear_modify(m);*/
1842 m->valid = VM_PAGE_BITS_ALL;
1844 vm_page_flag_clear(m, PG_ZERO | PG_SWAPINPROG);
1845 vm_page_flag_set(m, PG_SWAPPED);
1848 * We have to wake specifically requested pages
1849 * up too because we cleared PG_SWAPINPROG and
1850 * could be waiting for it in getpages. However,
1851 * be sure to not unbusy getpages specifically
1852 * requested page - getpages expects it to be
1855 * bio_driver_info holds the requested page
1857 if (i != (int)(intptr_t)bio->bio_driver_info) {
1858 vm_page_deactivate(m);
1865 * Mark the page clean but do not mess with the
1866 * pmap-layer's modified state. That state should
1867 * also be clear since the caller protected the
1868 * page VM_PROT_READ, but allow the case.
1870 * We are in an interrupt, avoid pmap operations.
1872 * If we have a severe page deficit, deactivate the
1873 * page. Do not try to cache it (which would also
1874 * involve a pmap op), because the page might still
1877 * When using the swap to cache clean vnode pages
1878 * we do not mess with the page dirty bits.
1880 vm_page_busy_wait(m, FALSE, "swadpg");
1881 if (m->object->type == OBJT_SWAP)
1883 vm_page_flag_clear(m, PG_SWAPINPROG);
1884 vm_page_flag_set(m, PG_SWAPPED);
1885 if (vm_page_count_severe())
1886 vm_page_deactivate(m);
1888 if (!vm_page_count_severe() || !vm_page_try_to_cache(m))
1889 vm_page_protect(m, VM_PROT_READ);
1891 vm_page_io_finish(m);
1897 * adjust pip. NOTE: the original parent may still have its own
1898 * pip refs on the object.
1902 vm_object_pip_wakeup_n(object, bp->b_xio.xio_npages);
1905 * Release the physical I/O buffer.
1907 * NOTE: Due to synchronous operations in the write case b_cmd may
1908 * already be set to BUF_CMD_DONE and BIO_SYNC may have already
1911 * Use vm_token to interlock nsw_rcount/wcount wakeup?
1913 lwkt_gettoken(&vm_token);
1914 if (bio->bio_caller_info1.index & SWBIO_READ)
1915 nswptr = &nsw_rcount;
1916 else if (bio->bio_caller_info1.index & SWBIO_SYNC)
1917 nswptr = &nsw_wcount_sync;
1919 nswptr = &nsw_wcount_async;
1920 bp->b_cmd = BUF_CMD_DONE;
1921 relpbuf(bp, nswptr);
1922 lwkt_reltoken(&vm_token);
1926 * Fault-in a potentially swapped page and remove the swap reference.
1927 * (used by swapoff code)
1929 * object must be held.
1931 static __inline void
1932 swp_pager_fault_page(vm_object_t object, vm_pindex_t pindex)
1938 ASSERT_LWKT_TOKEN_HELD(vm_object_token(object));
1940 if (object->type == OBJT_VNODE) {
1942 * Any swap related to a vnode is due to swapcache. We must
1943 * vget() the vnode in case it is not active (otherwise
1944 * vref() will panic). Calling vm_object_page_remove() will
1945 * ensure that any swap ref is removed interlocked with the
1946 * page. clean_only is set to TRUE so we don't throw away
1949 vp = object->handle;
1950 error = vget(vp, LK_SHARED | LK_RETRY | LK_CANRECURSE);
1952 vm_object_page_remove(object, pindex, pindex + 1, TRUE);
1957 * Otherwise it is a normal OBJT_SWAP object and we can
1958 * fault the page in and remove the swap.
1960 m = vm_fault_object_page(object, IDX_TO_OFF(pindex),
1962 VM_FAULT_DIRTY | VM_FAULT_UNSWAP,
1970 * This removes all swap blocks related to a particular device. We have
1971 * to be careful of ripups during the scan.
1973 static int swp_pager_swapoff_callback(struct swblock *swap, void *data);
1976 swap_pager_swapoff(int devidx)
1978 struct vm_object marker;
1980 struct swswapoffinfo info;
1982 bzero(&marker, sizeof(marker));
1983 marker.type = OBJT_MARKER;
1985 lwkt_gettoken(&vmobj_token);
1986 TAILQ_INSERT_HEAD(&vm_object_list, &marker, object_list);
1988 while ((object = TAILQ_NEXT(&marker, object_list)) != NULL) {
1989 if (object->type == OBJT_MARKER)
1991 if (object->type != OBJT_SWAP && object->type != OBJT_VNODE)
1993 vm_object_hold(object);
1994 if (object->type != OBJT_SWAP && object->type != OBJT_VNODE) {
1995 vm_object_drop(object);
1998 info.object = object;
1999 info.devidx = devidx;
2000 swblock_rb_tree_RB_SCAN(&object->swblock_root,
2002 swp_pager_swapoff_callback,
2004 vm_object_drop(object);
2006 if (object == TAILQ_NEXT(&marker, object_list)) {
2007 TAILQ_REMOVE(&vm_object_list, &marker, object_list);
2008 TAILQ_INSERT_AFTER(&vm_object_list, object,
2009 &marker, object_list);
2012 TAILQ_REMOVE(&vm_object_list, &marker, object_list);
2013 lwkt_reltoken(&vmobj_token);
2016 * If we fail to locate all swblocks we just fail gracefully and
2017 * do not bother to restore paging on the swap device. If the
2018 * user wants to retry the user can retry.
2020 if (swdevt[devidx].sw_nused)
2028 swp_pager_swapoff_callback(struct swblock *swap, void *data)
2030 struct swswapoffinfo *info = data;
2031 vm_object_t object = info->object;
2036 index = swap->swb_index;
2037 for (i = 0; i < SWAP_META_PAGES; ++i) {
2039 * Make sure we don't race a dying object. This will
2040 * kill the scan of the object's swap blocks entirely.
2042 if (object->flags & OBJ_DEAD)
2046 * Fault the page, which can obviously block. If the swap
2047 * structure disappears break out.
2049 v = swap->swb_pages[i];
2050 if (v != SWAPBLK_NONE && BLK2DEVIDX(v) == info->devidx) {
2051 swp_pager_fault_page(object, swap->swb_index + i);
2052 /* swap ptr might go away */
2053 if (RB_LOOKUP(swblock_rb_tree,
2054 &object->swblock_root, index) != swap) {
2062 /************************************************************************
2064 ************************************************************************
2066 * These routines manipulate the swap metadata stored in the
2067 * OBJT_SWAP object. All swp_*() routines must be called at
2068 * splvm() because swap can be freed up by the low level vm_page
2069 * code which might be called from interrupts beyond what splbio() covers.
2071 * Swap metadata is implemented with a global hash and not directly
2072 * linked into the object. Instead the object simply contains
2073 * appropriate tracking counters.
2077 * Lookup the swblock containing the specified swap block index.
2079 * The caller must hold the object.
2083 swp_pager_lookup(vm_object_t object, vm_pindex_t index)
2085 ASSERT_LWKT_TOKEN_HELD(vm_object_token(object));
2086 index &= ~(vm_pindex_t)SWAP_META_MASK;
2087 return (RB_LOOKUP(swblock_rb_tree, &object->swblock_root, index));
2091 * Remove a swblock from the RB tree.
2093 * The caller must hold the object.
2097 swp_pager_remove(vm_object_t object, struct swblock *swap)
2099 ASSERT_LWKT_TOKEN_HELD(vm_object_token(object));
2100 RB_REMOVE(swblock_rb_tree, &object->swblock_root, swap);
2104 * Convert default object to swap object if necessary
2106 * The caller must hold the object.
2109 swp_pager_meta_convert(vm_object_t object)
2111 if (object->type == OBJT_DEFAULT) {
2112 object->type = OBJT_SWAP;
2113 KKASSERT(object->swblock_count == 0);
2118 * SWP_PAGER_META_BUILD() - add swap block to swap meta data for object
2120 * We first convert the object to a swap object if it is a default
2121 * object. Vnode objects do not need to be converted.
2123 * The specified swapblk is added to the object's swap metadata. If
2124 * the swapblk is not valid, it is freed instead. Any previously
2125 * assigned swapblk is freed.
2127 * The caller must hold the object.
2130 swp_pager_meta_build(vm_object_t object, vm_pindex_t index, swblk_t swapblk)
2132 struct swblock *swap;
2133 struct swblock *oswap;
2136 KKASSERT(swapblk != SWAPBLK_NONE);
2137 ASSERT_LWKT_TOKEN_HELD(vm_object_token(object));
2140 * Convert object if necessary
2142 if (object->type == OBJT_DEFAULT)
2143 swp_pager_meta_convert(object);
2146 * Locate swblock. If not found create, but if we aren't adding
2147 * anything just return. If we run out of space in the map we wait
2148 * and, since the hash table may have changed, retry.
2151 swap = swp_pager_lookup(object, index);
2156 swap = zalloc(swap_zone);
2161 swap->swb_index = index & ~(vm_pindex_t)SWAP_META_MASK;
2162 swap->swb_count = 0;
2164 ++object->swblock_count;
2166 for (i = 0; i < SWAP_META_PAGES; ++i)
2167 swap->swb_pages[i] = SWAPBLK_NONE;
2168 oswap = RB_INSERT(swblock_rb_tree, &object->swblock_root, swap);
2169 KKASSERT(oswap == NULL);
2173 * Delete prior contents of metadata.
2175 * NOTE: Decrement swb_count after the freeing operation (which
2176 * might block) to prevent racing destruction of the swblock.
2178 index &= SWAP_META_MASK;
2180 while ((v = swap->swb_pages[index]) != SWAPBLK_NONE) {
2181 swap->swb_pages[index] = SWAPBLK_NONE;
2183 swp_pager_freeswapspace(object, v, 1);
2188 * Enter block into metadata
2190 swap->swb_pages[index] = swapblk;
2191 if (swapblk != SWAPBLK_NONE)
2196 * SWP_PAGER_META_FREE() - free a range of blocks in the object's swap metadata
2198 * The requested range of blocks is freed, with any associated swap
2199 * returned to the swap bitmap.
2201 * This routine will free swap metadata structures as they are cleaned
2202 * out. This routine does *NOT* operate on swap metadata associated
2203 * with resident pages.
2205 * The caller must hold the object.
2207 static int swp_pager_meta_free_callback(struct swblock *swb, void *data);
2210 swp_pager_meta_free(vm_object_t object, vm_pindex_t index, vm_pindex_t count)
2212 struct swfreeinfo info;
2214 ASSERT_LWKT_TOKEN_HELD(vm_object_token(object));
2219 if (object->swblock_count == 0) {
2220 KKASSERT(RB_EMPTY(&object->swblock_root));
2227 * Setup for RB tree scan. Note that the pindex range can be huge
2228 * due to the 64 bit page index space so we cannot safely iterate.
2230 info.object = object;
2231 info.basei = index & ~(vm_pindex_t)SWAP_META_MASK;
2233 info.endi = index + count - 1;
2234 swblock_rb_tree_RB_SCAN(&object->swblock_root, rb_swblock_scancmp,
2235 swp_pager_meta_free_callback, &info);
2239 * The caller must hold the object.
2243 swp_pager_meta_free_callback(struct swblock *swap, void *data)
2245 struct swfreeinfo *info = data;
2246 vm_object_t object = info->object;
2251 * Figure out the range within the swblock. The wider scan may
2252 * return edge-case swap blocks when the start and/or end points
2253 * are in the middle of a block.
2255 if (swap->swb_index < info->begi)
2256 index = (int)info->begi & SWAP_META_MASK;
2260 if (swap->swb_index + SWAP_META_PAGES > info->endi)
2261 eindex = (int)info->endi & SWAP_META_MASK;
2263 eindex = SWAP_META_MASK;
2266 * Scan and free the blocks. The loop terminates early
2267 * if (swap) runs out of blocks and could be freed.
2269 * NOTE: Decrement swb_count after swp_pager_freeswapspace()
2270 * to deal with a zfree race.
2272 while (index <= eindex) {
2273 swblk_t v = swap->swb_pages[index];
2275 if (v != SWAPBLK_NONE) {
2276 swap->swb_pages[index] = SWAPBLK_NONE;
2278 swp_pager_freeswapspace(object, v, 1);
2279 if (--swap->swb_count == 0) {
2280 swp_pager_remove(object, swap);
2281 zfree(swap_zone, swap);
2282 --object->swblock_count;
2289 /* swap may be invalid here due to zfree above */
2296 * SWP_PAGER_META_FREE_ALL() - destroy all swap metadata associated with object
2298 * This routine locates and destroys all swap metadata associated with
2301 * NOTE: Decrement swb_count after the freeing operation (which
2302 * might block) to prevent racing destruction of the swblock.
2304 * The caller must hold the object.
2307 swp_pager_meta_free_all(vm_object_t object)
2309 struct swblock *swap;
2312 ASSERT_LWKT_TOKEN_HELD(vm_object_token(object));
2314 while ((swap = RB_ROOT(&object->swblock_root)) != NULL) {
2315 swp_pager_remove(object, swap);
2316 for (i = 0; i < SWAP_META_PAGES; ++i) {
2317 swblk_t v = swap->swb_pages[i];
2318 if (v != SWAPBLK_NONE) {
2320 swp_pager_freeswapspace(object, v, 1);
2324 if (swap->swb_count != 0)
2325 panic("swap_pager_meta_free_all: swb_count != 0");
2326 zfree(swap_zone, swap);
2327 --object->swblock_count;
2330 KKASSERT(object->swblock_count == 0);
2334 * SWP_PAGER_METACTL() - misc control of swap and vm_page_t meta data.
2336 * This routine is capable of looking up, popping, or freeing
2337 * swapblk assignments in the swap meta data or in the vm_page_t.
2338 * The routine typically returns the swapblk being looked-up, or popped,
2339 * or SWAPBLK_NONE if the block was freed, or SWAPBLK_NONE if the block
2340 * was invalid. This routine will automatically free any invalid
2341 * meta-data swapblks.
2343 * It is not possible to store invalid swapblks in the swap meta data
2344 * (other then a literal 'SWAPBLK_NONE'), so we don't bother checking.
2346 * When acting on a busy resident page and paging is in progress, we
2347 * have to wait until paging is complete but otherwise can act on the
2350 * SWM_FREE remove and free swap block from metadata
2351 * SWM_POP remove from meta data but do not free.. pop it out
2353 * The caller must hold the object.
2356 swp_pager_meta_ctl(vm_object_t object, vm_pindex_t index, int flags)
2358 struct swblock *swap;
2361 if (object->swblock_count == 0)
2362 return(SWAPBLK_NONE);
2365 swap = swp_pager_lookup(object, index);
2368 index &= SWAP_META_MASK;
2369 r1 = swap->swb_pages[index];
2371 if (r1 != SWAPBLK_NONE) {
2372 if (flags & (SWM_FREE|SWM_POP)) {
2373 swap->swb_pages[index] = SWAPBLK_NONE;
2374 if (--swap->swb_count == 0) {
2375 swp_pager_remove(object, swap);
2376 zfree(swap_zone, swap);
2377 --object->swblock_count;
2380 /* swap ptr may be invalid */
2381 if (flags & SWM_FREE) {
2382 swp_pager_freeswapspace(object, r1, 1);
2386 /* swap ptr may be invalid */