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. Neither the name of the University nor the names of its contributors
54 * may be used to endorse or promote products derived from this software
55 * without specific prior written permission.
57 * THIS SOFTWARE IS PROVIDED BY THE REGENTS AND CONTRIBUTORS ``AS IS'' AND
58 * ANY EXPRESS OR IMPLIED WARRANTIES, INCLUDING, BUT NOT LIMITED TO, THE
59 * IMPLIED WARRANTIES OF MERCHANTABILITY AND FITNESS FOR A PARTICULAR PURPOSE
60 * ARE DISCLAIMED. IN NO EVENT SHALL THE REGENTS OR CONTRIBUTORS BE LIABLE
61 * FOR ANY DIRECT, INDIRECT, INCIDENTAL, SPECIAL, EXEMPLARY, OR CONSEQUENTIAL
62 * DAMAGES (INCLUDING, BUT NOT LIMITED TO, PROCUREMENT OF SUBSTITUTE GOODS
63 * OR SERVICES; LOSS OF USE, DATA, OR PROFITS; OR BUSINESS INTERRUPTION)
64 * HOWEVER CAUSED AND ON ANY THEORY OF LIABILITY, WHETHER IN CONTRACT, STRICT
65 * LIABILITY, OR TORT (INCLUDING NEGLIGENCE OR OTHERWISE) ARISING IN ANY WAY
66 * OUT OF THE USE OF THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF
72 * Radix Bitmap 'blists'.
74 * - The new swapper uses the new radix bitmap code. This should scale
75 * to arbitrarily small or arbitrarily large swap spaces and an almost
76 * arbitrary degree of fragmentation.
80 * - on the fly reallocation of swap during putpages. The new system
81 * does not try to keep previously allocated swap blocks for dirty
84 * - on the fly deallocation of swap
86 * - No more garbage collection required. Unnecessarily allocated swap
87 * blocks only exist for dirty vm_page_t's now and these are already
88 * cycled (in a high-load system) by the pager. We also do on-the-fly
89 * removal of invalidated swap blocks when a page is destroyed
92 * from: Utah $Hdr: swap_pager.c 1.4 91/04/30$
93 * @(#)swap_pager.c 8.9 (Berkeley) 3/21/94
94 * $FreeBSD: src/sys/vm/swap_pager.c,v 1.130.2.12 2002/08/31 21:15:55 dillon Exp $
97 #include <sys/param.h>
98 #include <sys/systm.h>
100 #include <sys/kernel.h>
101 #include <sys/proc.h>
103 #include <sys/vnode.h>
104 #include <sys/malloc.h>
105 #include <sys/vmmeter.h>
106 #include <sys/sysctl.h>
107 #include <sys/blist.h>
108 #include <sys/lock.h>
109 #include <sys/thread2.h>
111 #include "opt_swap.h"
113 #include <vm/vm_object.h>
114 #include <vm/vm_page.h>
115 #include <vm/vm_pager.h>
116 #include <vm/vm_pageout.h>
117 #include <vm/swap_pager.h>
118 #include <vm/vm_extern.h>
119 #include <vm/vm_zone.h>
120 #include <vm/vnode_pager.h>
122 #include <sys/buf2.h>
123 #include <vm/vm_page2.h>
125 #ifndef MAX_PAGEOUT_CLUSTER
126 #define MAX_PAGEOUT_CLUSTER SWB_NPAGES
129 #define SWM_FREE 0x02 /* free, period */
130 #define SWM_POP 0x04 /* pop out */
132 #define SWBIO_READ 0x01
133 #define SWBIO_WRITE 0x02
134 #define SWBIO_SYNC 0x04
140 vm_pindex_t endi; /* inclusive */
143 struct swswapoffinfo {
149 * vm_swap_size is in page-sized chunks now. It was DEV_BSIZE'd chunks
153 int swap_pager_full; /* swap space exhaustion (task killing) */
154 int vm_swap_cache_use;
155 int vm_swap_anon_use;
156 static int vm_report_swap_allocs;
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 */
168 static swblk_t swapiterator; /* linearize allocations */
171 extern struct vnode *swapdev_vp;
172 extern struct swdevt *swdevt;
175 #define BLK2DEVIDX(blk) (nswdev > 1 ? blk / dmmax % nswdev : 0)
177 SYSCTL_INT(_vm, OID_AUTO, swap_async_max,
178 CTLFLAG_RW, &swap_async_max, 0, "Maximum running async swap ops");
179 SYSCTL_INT(_vm, OID_AUTO, swap_burst_read,
180 CTLFLAG_RW, &swap_burst_read, 0, "Allow burst reads for pageins");
182 SYSCTL_INT(_vm, OID_AUTO, swap_cache_use,
183 CTLFLAG_RD, &vm_swap_cache_use, 0, "");
184 SYSCTL_INT(_vm, OID_AUTO, swap_anon_use,
185 CTLFLAG_RD, &vm_swap_anon_use, 0, "");
186 SYSCTL_INT(_vm, OID_AUTO, swap_size,
187 CTLFLAG_RD, &vm_swap_size, 0, "");
188 SYSCTL_INT(_vm, OID_AUTO, report_swap_allocs,
189 CTLFLAG_RW, &vm_report_swap_allocs, 0, "");
194 * Red-Black tree for swblock entries
196 * The caller must hold vm_token
198 RB_GENERATE2(swblock_rb_tree, swblock, swb_entry, rb_swblock_compare,
199 vm_pindex_t, swb_index);
202 rb_swblock_compare(struct swblock *swb1, struct swblock *swb2)
204 if (swb1->swb_index < swb2->swb_index)
206 if (swb1->swb_index > swb2->swb_index)
213 rb_swblock_scancmp(struct swblock *swb, void *data)
215 struct swfreeinfo *info = data;
217 if (swb->swb_index < info->basei)
219 if (swb->swb_index > info->endi)
226 rb_swblock_condcmp(struct swblock *swb, void *data)
228 struct swfreeinfo *info = data;
230 if (swb->swb_index < info->basei)
236 * pagerops for OBJT_SWAP - "swap pager". Some ops are also global procedure
237 * calls hooked from other parts of the VM system and do not appear here.
238 * (see vm/swap_pager.h).
241 static void swap_pager_dealloc (vm_object_t object);
242 static int swap_pager_getpage (vm_object_t, vm_page_t *, int);
243 static void swap_chain_iodone(struct bio *biox);
245 struct pagerops swappagerops = {
246 swap_pager_dealloc, /* deallocate an OBJT_SWAP object */
247 swap_pager_getpage, /* pagein */
248 swap_pager_putpages, /* pageout */
249 swap_pager_haspage /* get backing store status for page */
253 * dmmax is in page-sized chunks with the new swap system. It was
254 * dev-bsized chunks in the old. dmmax is always a power of 2.
256 * swap_*() routines are externally accessible. swp_*() routines are
261 static int dmmax_mask;
262 int nswap_lowat = 128; /* in pages, swap_pager_almost_full warn */
263 int nswap_hiwat = 512; /* in pages, swap_pager_almost_full warn */
265 static __inline void swp_sizecheck (void);
266 static void swp_pager_async_iodone (struct bio *bio);
269 * Swap bitmap functions
272 static __inline void swp_pager_freeswapspace(vm_object_t object,
273 swblk_t blk, int npages);
274 static __inline swblk_t swp_pager_getswapspace(vm_object_t object, int npages);
280 static void swp_pager_meta_convert(vm_object_t);
281 static void swp_pager_meta_build(vm_object_t, vm_pindex_t, swblk_t);
282 static void swp_pager_meta_free(vm_object_t, vm_pindex_t, vm_pindex_t);
283 static void swp_pager_meta_free_all(vm_object_t);
284 static swblk_t swp_pager_meta_ctl(vm_object_t, vm_pindex_t, int);
287 * SWP_SIZECHECK() - update swap_pager_full indication
289 * update the swap_pager_almost_full indication and warn when we are
290 * about to run out of swap space, using lowat/hiwat hysteresis.
292 * Clear swap_pager_full ( task killing ) indication when lowat is met.
294 * No restrictions on call
295 * This routine may not block.
301 if (vm_swap_size < nswap_lowat) {
302 if (swap_pager_almost_full == 0) {
303 kprintf("swap_pager: out of swap space\n");
304 swap_pager_almost_full = 1;
308 if (vm_swap_size > nswap_hiwat)
309 swap_pager_almost_full = 0;
314 * SWAP_PAGER_INIT() - initialize the swap pager!
316 * Expected to be started from system init. NOTE: This code is run
317 * before much else so be careful what you depend on. Most of the VM
318 * system has yet to be initialized at this point.
320 * Called from the low level boot code only.
323 swap_pager_init(void *arg __unused)
326 * Device Stripe, in PAGE_SIZE'd blocks
328 dmmax = SWB_NPAGES * 2;
329 dmmax_mask = ~(dmmax - 1);
331 SYSINIT(vm_mem, SI_BOOT1_VM, SI_ORDER_THIRD, swap_pager_init, NULL)
334 * SWAP_PAGER_SWAP_INIT() - swap pager initialization from pageout process
336 * Expected to be started from pageout process once, prior to entering
339 * Called from the low level boot code only.
342 swap_pager_swap_init(void)
347 * Number of in-transit swap bp operations. Don't
348 * exhaust the pbufs completely. Make sure we
349 * initialize workable values (0 will work for hysteresis
350 * but it isn't very efficient).
352 * The nsw_cluster_max is constrained by the number of pages an XIO
353 * holds, i.e., (MAXPHYS/PAGE_SIZE) and our locally defined
354 * MAX_PAGEOUT_CLUSTER. Also be aware that swap ops are
355 * constrained by the swap device interleave stripe size.
357 * Currently we hardwire nsw_wcount_async to 4. This limit is
358 * designed to prevent other I/O from having high latencies due to
359 * our pageout I/O. The value 4 works well for one or two active swap
360 * devices but is probably a little low if you have more. Even so,
361 * a higher value would probably generate only a limited improvement
362 * with three or four active swap devices since the system does not
363 * typically have to pageout at extreme bandwidths. We will want
364 * at least 2 per swap devices, and 4 is a pretty good value if you
365 * have one NFS swap device due to the command/ack latency over NFS.
366 * So it all works out pretty well.
369 nsw_cluster_max = min((MAXPHYS/PAGE_SIZE), MAX_PAGEOUT_CLUSTER);
371 nsw_rcount = (nswbuf + 1) / 2;
372 nsw_wcount_sync = (nswbuf + 3) / 4;
373 nsw_wcount_async = 4;
374 nsw_wcount_async_max = nsw_wcount_async;
377 * The zone is dynamically allocated so generally size it to
378 * maxswzone (32MB to 512MB of KVM). Set a minimum size based
379 * on physical memory of around 8x (each swblock can hold 16 pages).
381 * With the advent of SSDs (vs HDs) the practical (swap:memory) ratio
382 * has increased dramatically.
384 n = vmstats.v_page_count / 2;
385 if (maxswzone && n < maxswzone / sizeof(struct swblock))
386 n = maxswzone / sizeof(struct swblock);
392 sizeof(struct swblock),
396 if (swap_zone != NULL)
399 * if the allocation failed, try a zone two thirds the
400 * size of the previous attempt.
405 if (swap_zone == NULL)
406 panic("swap_pager_swap_init: swap_zone == NULL");
408 kprintf("Swap zone entries reduced from %d to %d.\n", n2, n);
412 * SWAP_PAGER_ALLOC() - allocate a new OBJT_SWAP VM object and instantiate
413 * its metadata structures.
415 * This routine is called from the mmap and fork code to create a new
416 * OBJT_SWAP object. We do this by creating an OBJT_DEFAULT object
417 * and then converting it with swp_pager_meta_convert().
419 * We only support unnamed objects.
424 swap_pager_alloc(void *handle, off_t size, vm_prot_t prot, off_t offset)
428 KKASSERT(handle == NULL);
429 object = vm_object_allocate_hold(OBJT_DEFAULT,
430 OFF_TO_IDX(offset + PAGE_MASK + size));
431 swp_pager_meta_convert(object);
432 vm_object_drop(object);
438 * SWAP_PAGER_DEALLOC() - remove swap metadata from object
440 * The swap backing for the object is destroyed. The code is
441 * designed such that we can reinstantiate it later, but this
442 * routine is typically called only when the entire object is
443 * about to be destroyed.
445 * The object must be locked or unreferenceable.
446 * No other requirements.
449 swap_pager_dealloc(vm_object_t object)
451 vm_object_hold(object);
452 vm_object_pip_wait(object, "swpdea");
455 * Free all remaining metadata. We only bother to free it from
456 * the swap meta data. We do not attempt to free swapblk's still
457 * associated with vm_page_t's for this object. We do not care
458 * if paging is still in progress on some objects.
460 swp_pager_meta_free_all(object);
461 vm_object_drop(object);
464 /************************************************************************
465 * SWAP PAGER BITMAP ROUTINES *
466 ************************************************************************/
469 * SWP_PAGER_GETSWAPSPACE() - allocate raw swap space
471 * Allocate swap for the requested number of pages. The starting
472 * swap block number (a page index) is returned or SWAPBLK_NONE
473 * if the allocation failed.
475 * Also has the side effect of advising that somebody made a mistake
476 * when they configured swap and didn't configure enough.
478 * The caller must hold the object.
479 * This routine may not block.
481 static __inline swblk_t
482 swp_pager_getswapspace(vm_object_t object, int npages)
486 lwkt_gettoken(&vm_token);
487 blk = blist_allocat(swapblist, npages, swapiterator);
488 if (blk == SWAPBLK_NONE)
489 blk = blist_allocat(swapblist, npages, 0);
490 if (blk == SWAPBLK_NONE) {
491 if (swap_pager_full != 2) {
492 kprintf("swap_pager_getswapspace: failed alloc=%d\n",
495 swap_pager_almost_full = 1;
498 /* swapiterator = blk; disable for now, doesn't work well */
499 swapacctspace(blk, -npages);
500 if (object->type == OBJT_SWAP)
501 vm_swap_anon_use += npages;
503 vm_swap_cache_use += npages;
506 lwkt_reltoken(&vm_token);
511 * SWP_PAGER_FREESWAPSPACE() - free raw swap space
513 * This routine returns the specified swap blocks back to the bitmap.
515 * Note: This routine may not block (it could in the old swap code),
516 * and through the use of the new blist routines it does not block.
518 * We must be called at splvm() to avoid races with bitmap frees from
519 * vm_page_remove() aka swap_pager_page_removed().
521 * This routine may not block.
525 swp_pager_freeswapspace(vm_object_t object, swblk_t blk, int npages)
527 struct swdevt *sp = &swdevt[BLK2DEVIDX(blk)];
529 lwkt_gettoken(&vm_token);
530 sp->sw_nused -= npages;
531 if (object->type == OBJT_SWAP)
532 vm_swap_anon_use -= npages;
534 vm_swap_cache_use -= npages;
536 if (sp->sw_flags & SW_CLOSING) {
537 lwkt_reltoken(&vm_token);
541 blist_free(swapblist, blk, npages);
542 vm_swap_size += npages;
544 lwkt_reltoken(&vm_token);
548 * SWAP_PAGER_FREESPACE() - frees swap blocks associated with a page
549 * range within an object.
551 * This is a globally accessible routine.
553 * This routine removes swapblk assignments from swap metadata.
555 * The external callers of this routine typically have already destroyed
556 * or renamed vm_page_t's associated with this range in the object so
562 swap_pager_freespace(vm_object_t object, vm_pindex_t start, vm_pindex_t size)
564 vm_object_hold(object);
565 swp_pager_meta_free(object, start, size);
566 vm_object_drop(object);
573 swap_pager_freespace_all(vm_object_t object)
575 vm_object_hold(object);
576 swp_pager_meta_free_all(object);
577 vm_object_drop(object);
581 * This function conditionally frees swap cache swap starting at
582 * (*basei) in the object. (count) swap blocks will be nominally freed.
583 * The actual number of blocks freed can be more or less than the
586 * This function nominally returns the number of blocks freed. However,
587 * the actual number of blocks freed may be less then the returned value.
588 * If the function is unable to exhaust the object or if it is able to
589 * free (approximately) the requested number of blocks it returns
592 * If we exhaust the object we will return a value n <= count.
594 * The caller must hold the object.
596 * WARNING! If count == 0 then -1 can be returned as a degenerate case,
597 * callers should always pass a count value > 0.
599 static int swap_pager_condfree_callback(struct swblock *swap, void *data);
602 swap_pager_condfree(vm_object_t object, vm_pindex_t *basei, int count)
604 struct swfreeinfo info;
608 ASSERT_LWKT_TOKEN_HELD(vm_object_token(object));
610 info.object = object;
611 info.basei = *basei; /* skip up to this page index */
612 info.begi = count; /* max swap pages to destroy */
613 info.endi = count * 8; /* max swblocks to scan */
615 swblock_rb_tree_RB_SCAN(&object->swblock_root, rb_swblock_condcmp,
616 swap_pager_condfree_callback, &info);
620 * Take the higher difference swblocks vs pages
622 n = count - (int)info.begi;
623 t = count * 8 - (int)info.endi;
632 * The idea is to free whole meta-block to avoid fragmenting
633 * the swap space or disk I/O. We only do this if NO VM pages
636 * We do not have to deal with clearing PG_SWAPPED in related VM
637 * pages because there are no related VM pages.
639 * The caller must hold the object.
642 swap_pager_condfree_callback(struct swblock *swap, void *data)
644 struct swfreeinfo *info = data;
645 vm_object_t object = info->object;
648 for (i = 0; i < SWAP_META_PAGES; ++i) {
649 if (vm_page_lookup(object, swap->swb_index + i))
652 info->basei = swap->swb_index + SWAP_META_PAGES;
653 if (i == SWAP_META_PAGES) {
654 info->begi -= swap->swb_count;
655 swap_pager_freespace(object, swap->swb_index, SWAP_META_PAGES);
658 if ((int)info->begi < 0 || (int)info->endi < 0)
665 * Called by vm_page_alloc() when a new VM page is inserted
666 * into a VM object. Checks whether swap has been assigned to
667 * the page and sets PG_SWAPPED as necessary.
672 swap_pager_page_inserted(vm_page_t m)
674 if (m->object->swblock_count) {
675 vm_object_hold(m->object);
676 if (swp_pager_meta_ctl(m->object, m->pindex, 0) != SWAPBLK_NONE)
677 vm_page_flag_set(m, PG_SWAPPED);
678 vm_object_drop(m->object);
683 * SWAP_PAGER_RESERVE() - reserve swap blocks in object
685 * Assigns swap blocks to the specified range within the object. The
686 * swap blocks are not zerod. Any previous swap assignment is destroyed.
688 * Returns 0 on success, -1 on failure.
690 * The caller is responsible for avoiding races in the specified range.
691 * No other requirements.
694 swap_pager_reserve(vm_object_t object, vm_pindex_t start, vm_size_t size)
697 swblk_t blk = SWAPBLK_NONE;
698 vm_pindex_t beg = start; /* save start index */
700 vm_object_hold(object);
705 while ((blk = swp_pager_getswapspace(object, n)) ==
710 swp_pager_meta_free(object, beg,
712 vm_object_drop(object);
717 swp_pager_meta_build(object, start, blk);
723 swp_pager_meta_free(object, start, n);
724 vm_object_drop(object);
729 * SWAP_PAGER_COPY() - copy blocks from source pager to destination pager
730 * and destroy the source.
732 * Copy any valid swapblks from the source to the destination. In
733 * cases where both the source and destination have a valid swapblk,
734 * we keep the destination's.
736 * This routine is allowed to block. It may block allocating metadata
737 * indirectly through swp_pager_meta_build() or if paging is still in
738 * progress on the source.
740 * XXX vm_page_collapse() kinda expects us not to block because we
741 * supposedly do not need to allocate memory, but for the moment we
742 * *may* have to get a little memory from the zone allocator, but
743 * it is taken from the interrupt memory. We should be ok.
745 * The source object contains no vm_page_t's (which is just as well)
746 * The source object is of type OBJT_SWAP.
748 * The source and destination objects must be held by the caller.
751 swap_pager_copy(vm_object_t srcobject, vm_object_t dstobject,
752 vm_pindex_t base_index, int destroysource)
756 ASSERT_LWKT_TOKEN_HELD(vm_object_token(srcobject));
757 ASSERT_LWKT_TOKEN_HELD(vm_object_token(dstobject));
760 * transfer source to destination.
762 for (i = 0; i < dstobject->size; ++i) {
766 * Locate (without changing) the swapblk on the destination,
767 * unless it is invalid in which case free it silently, or
768 * if the destination is a resident page, in which case the
769 * source is thrown away.
771 dstaddr = swp_pager_meta_ctl(dstobject, i, 0);
773 if (dstaddr == SWAPBLK_NONE) {
775 * Destination has no swapblk and is not resident,
780 srcaddr = swp_pager_meta_ctl(srcobject,
781 base_index + i, SWM_POP);
783 if (srcaddr != SWAPBLK_NONE)
784 swp_pager_meta_build(dstobject, i, srcaddr);
787 * Destination has valid swapblk or it is represented
788 * by a resident page. We destroy the sourceblock.
790 swp_pager_meta_ctl(srcobject, base_index + i, SWM_FREE);
795 * Free left over swap blocks in source.
797 * We have to revert the type to OBJT_DEFAULT so we do not accidently
798 * double-remove the object from the swap queues.
802 * Reverting the type is not necessary, the caller is going
803 * to destroy srcobject directly, but I'm doing it here
804 * for consistency since we've removed the object from its
807 swp_pager_meta_free_all(srcobject);
808 if (srcobject->type == OBJT_SWAP)
809 srcobject->type = OBJT_DEFAULT;
814 * SWAP_PAGER_HASPAGE() - determine if we have good backing store for
815 * the requested page.
817 * We determine whether good backing store exists for the requested
818 * page and return TRUE if it does, FALSE if it doesn't.
820 * If TRUE, we also try to determine how much valid, contiguous backing
821 * store exists before and after the requested page within a reasonable
822 * distance. We do not try to restrict it to the swap device stripe
823 * (that is handled in getpages/putpages). It probably isn't worth
829 swap_pager_haspage(vm_object_t object, vm_pindex_t pindex)
834 * do we have good backing store at the requested index ?
836 vm_object_hold(object);
837 blk0 = swp_pager_meta_ctl(object, pindex, 0);
839 if (blk0 == SWAPBLK_NONE) {
840 vm_object_drop(object);
843 vm_object_drop(object);
848 * SWAP_PAGER_PAGE_UNSWAPPED() - remove swap backing store related to page
850 * This removes any associated swap backing store, whether valid or
851 * not, from the page. This operates on any VM object, not just OBJT_SWAP
854 * This routine is typically called when a page is made dirty, at
855 * which point any associated swap can be freed. MADV_FREE also
856 * calls us in a special-case situation
858 * NOTE!!! If the page is clean and the swap was valid, the caller
859 * should make the page dirty before calling this routine. This routine
860 * does NOT change the m->dirty status of the page. Also: MADV_FREE
863 * The page must be busied or soft-busied.
864 * The caller can hold the object to avoid blocking, else we might block.
865 * No other requirements.
868 swap_pager_unswapped(vm_page_t m)
870 if (m->flags & PG_SWAPPED) {
871 vm_object_hold(m->object);
872 KKASSERT(m->flags & PG_SWAPPED);
873 swp_pager_meta_ctl(m->object, m->pindex, SWM_FREE);
874 vm_page_flag_clear(m, PG_SWAPPED);
875 vm_object_drop(m->object);
880 * SWAP_PAGER_STRATEGY() - read, write, free blocks
882 * This implements a VM OBJECT strategy function using swap backing store.
883 * This can operate on any VM OBJECT type, not necessarily just OBJT_SWAP
886 * This is intended to be a cacheless interface (i.e. caching occurs at
887 * higher levels), and is also used as a swap-based SSD cache for vnode
888 * and device objects.
890 * All I/O goes directly to and from the swap device.
892 * We currently attempt to run I/O synchronously or asynchronously as
893 * the caller requests. This isn't perfect because we loose error
894 * sequencing when we run multiple ops in parallel to satisfy a request.
895 * But this is swap, so we let it all hang out.
900 swap_pager_strategy(vm_object_t object, struct bio *bio)
902 struct buf *bp = bio->bio_buf;
905 vm_pindex_t biox_blkno = 0;
911 struct bio_track *track;
916 * tracking for swapdev vnode I/Os
918 if (bp->b_cmd == BUF_CMD_READ)
919 track = &swapdev_vp->v_track_read;
921 track = &swapdev_vp->v_track_write;
924 if (bp->b_bcount & PAGE_MASK) {
925 bp->b_error = EINVAL;
926 bp->b_flags |= B_ERROR | B_INVAL;
928 kprintf("swap_pager_strategy: bp %p offset %lld size %d, "
929 "not page bounded\n",
930 bp, (long long)bio->bio_offset, (int)bp->b_bcount);
935 * Clear error indication, initialize page index, count, data pointer.
938 bp->b_flags &= ~B_ERROR;
939 bp->b_resid = bp->b_bcount;
941 start = (vm_pindex_t)(bio->bio_offset >> PAGE_SHIFT);
942 count = howmany(bp->b_bcount, PAGE_SIZE);
946 * Deal with BUF_CMD_FREEBLKS
948 if (bp->b_cmd == BUF_CMD_FREEBLKS) {
950 * FREE PAGE(s) - destroy underlying swap that is no longer
953 vm_object_hold(object);
954 swp_pager_meta_free(object, start, count);
955 vm_object_drop(object);
962 * We need to be able to create a new cluster of I/O's. We cannot
963 * use the caller fields of the passed bio so push a new one.
965 * Because nbio is just a placeholder for the cluster links,
966 * we can biodone() the original bio instead of nbio to make
967 * things a bit more efficient.
969 nbio = push_bio(bio);
970 nbio->bio_offset = bio->bio_offset;
971 nbio->bio_caller_info1.cluster_head = NULL;
972 nbio->bio_caller_info2.cluster_tail = NULL;
978 * Execute read or write
980 vm_object_hold(object);
986 * Obtain block. If block not found and writing, allocate a
987 * new block and build it into the object.
989 blk = swp_pager_meta_ctl(object, start, 0);
990 if ((blk == SWAPBLK_NONE) && bp->b_cmd != BUF_CMD_READ) {
991 blk = swp_pager_getswapspace(object, 1);
992 if (blk == SWAPBLK_NONE) {
993 bp->b_error = ENOMEM;
994 bp->b_flags |= B_ERROR;
997 swp_pager_meta_build(object, start, blk);
1001 * Do we have to flush our current collection? Yes if:
1003 * - no swap block at this index
1004 * - swap block is not contiguous
1005 * - we cross a physical disk boundry in the
1009 biox && (biox_blkno + btoc(bufx->b_bcount) != blk ||
1010 ((biox_blkno ^ blk) & dmmax_mask)
1013 if (bp->b_cmd == BUF_CMD_READ) {
1014 ++mycpu->gd_cnt.v_swapin;
1015 mycpu->gd_cnt.v_swappgsin += btoc(bufx->b_bcount);
1017 ++mycpu->gd_cnt.v_swapout;
1018 mycpu->gd_cnt.v_swappgsout += btoc(bufx->b_bcount);
1019 bufx->b_dirtyend = bufx->b_bcount;
1023 * Finished with this buf.
1025 KKASSERT(bufx->b_bcount != 0);
1026 if (bufx->b_cmd != BUF_CMD_READ)
1027 bufx->b_dirtyend = bufx->b_bcount;
1033 * Add new swapblk to biox, instantiating biox if necessary.
1034 * Zero-fill reads are able to take a shortcut.
1036 if (blk == SWAPBLK_NONE) {
1038 * We can only get here if we are reading. Since
1039 * we are at splvm() we can safely modify b_resid,
1040 * even if chain ops are in progress.
1042 bzero(data, PAGE_SIZE);
1043 bp->b_resid -= PAGE_SIZE;
1046 /* XXX chain count > 4, wait to <= 4 */
1048 bufx = getpbuf(NULL);
1049 biox = &bufx->b_bio1;
1050 cluster_append(nbio, bufx);
1051 bufx->b_flags |= (bp->b_flags & B_ORDERED);
1052 bufx->b_cmd = bp->b_cmd;
1053 biox->bio_done = swap_chain_iodone;
1054 biox->bio_offset = (off_t)blk << PAGE_SHIFT;
1055 biox->bio_caller_info1.cluster_parent = nbio;
1058 bufx->b_data = data;
1060 bufx->b_bcount += PAGE_SIZE;
1067 vm_object_drop(object);
1070 * Flush out last buffer
1073 if (bufx->b_cmd == BUF_CMD_READ) {
1074 ++mycpu->gd_cnt.v_swapin;
1075 mycpu->gd_cnt.v_swappgsin += btoc(bufx->b_bcount);
1077 ++mycpu->gd_cnt.v_swapout;
1078 mycpu->gd_cnt.v_swappgsout += btoc(bufx->b_bcount);
1079 bufx->b_dirtyend = bufx->b_bcount;
1081 KKASSERT(bufx->b_bcount);
1082 if (bufx->b_cmd != BUF_CMD_READ)
1083 bufx->b_dirtyend = bufx->b_bcount;
1084 /* biox, bufx = NULL */
1088 * Now initiate all the I/O. Be careful looping on our chain as
1089 * I/O's may complete while we are still initiating them.
1091 * If the request is a 100% sparse read no bios will be present
1092 * and we just biodone() the buffer.
1094 nbio->bio_caller_info2.cluster_tail = NULL;
1095 bufx = nbio->bio_caller_info1.cluster_head;
1099 biox = &bufx->b_bio1;
1101 bufx = bufx->b_cluster_next;
1102 vn_strategy(swapdev_vp, biox);
1109 * Completion of the cluster will also call biodone_chain(nbio).
1110 * We never call biodone(nbio) so we don't have to worry about
1111 * setting up a bio_done callback. It's handled in the sub-IO.
1122 swap_chain_iodone(struct bio *biox)
1125 struct buf *bufx; /* chained sub-buffer */
1126 struct bio *nbio; /* parent nbio with chain glue */
1127 struct buf *bp; /* original bp associated with nbio */
1130 bufx = biox->bio_buf;
1131 nbio = biox->bio_caller_info1.cluster_parent;
1135 * Update the original buffer
1137 KKASSERT(bp != NULL);
1138 if (bufx->b_flags & B_ERROR) {
1139 atomic_set_int(&bufx->b_flags, B_ERROR);
1140 bp->b_error = bufx->b_error; /* race ok */
1141 } else if (bufx->b_resid != 0) {
1142 atomic_set_int(&bufx->b_flags, B_ERROR);
1143 bp->b_error = EINVAL; /* race ok */
1145 atomic_subtract_int(&bp->b_resid, bufx->b_bcount);
1149 * Remove us from the chain.
1151 spin_lock(&bp->b_lock.lk_spinlock);
1152 nextp = &nbio->bio_caller_info1.cluster_head;
1153 while (*nextp != bufx) {
1154 KKASSERT(*nextp != NULL);
1155 nextp = &(*nextp)->b_cluster_next;
1157 *nextp = bufx->b_cluster_next;
1158 chain_empty = (nbio->bio_caller_info1.cluster_head == NULL);
1159 spin_unlock(&bp->b_lock.lk_spinlock);
1162 * Clean up bufx. If the chain is now empty we finish out
1163 * the parent. Note that we may be racing other completions
1164 * so we must use the chain_empty status from above.
1167 if (bp->b_resid != 0 && !(bp->b_flags & B_ERROR)) {
1168 atomic_set_int(&bp->b_flags, B_ERROR);
1169 bp->b_error = EINVAL;
1171 biodone_chain(nbio);
1173 relpbuf(bufx, NULL);
1177 * SWAP_PAGER_GETPAGES() - bring page in from swap
1179 * The requested page may have to be brought in from swap. Calculate the
1180 * swap block and bring in additional pages if possible. All pages must
1181 * have contiguous swap block assignments and reside in the same object.
1183 * The caller has a single vm_object_pip_add() reference prior to
1184 * calling us and we should return with the same.
1186 * The caller has BUSY'd the page. We should return with (*mpp) left busy,
1187 * and any additinal pages unbusied.
1189 * If the caller encounters a PG_RAM page it will pass it to us even though
1190 * it may be valid and dirty. We cannot overwrite the page in this case!
1191 * The case is used to allow us to issue pure read-aheads.
1193 * NOTE! XXX This code does not entirely pipeline yet due to the fact that
1194 * the PG_RAM page is validated at the same time as mreq. What we
1195 * really need to do is issue a separate read-ahead pbuf.
1200 swap_pager_getpage(vm_object_t object, vm_page_t *mpp, int seqaccess)
1213 vm_page_t marray[XIO_INTERNAL_PAGES];
1217 vm_object_hold(object);
1218 if (mreq->object != object) {
1219 panic("swap_pager_getpages: object mismatch %p/%p",
1226 * We don't want to overwrite a fully valid page as it might be
1227 * dirty. This case can occur when e.g. vm_fault hits a perfectly
1228 * valid page with PG_RAM set.
1230 * In this case we see if the next page is a suitable page-in
1231 * candidate and if it is we issue read-ahead. PG_RAM will be
1232 * set on the last page of the read-ahead to continue the pipeline.
1234 if (mreq->valid == VM_PAGE_BITS_ALL) {
1235 if (swap_burst_read == 0 || mreq->pindex + 1 >= object->size) {
1236 vm_object_drop(object);
1237 return(VM_PAGER_OK);
1239 blk = swp_pager_meta_ctl(object, mreq->pindex + 1, 0);
1240 if (blk == SWAPBLK_NONE) {
1241 vm_object_drop(object);
1242 return(VM_PAGER_OK);
1244 m = vm_page_lookup_busy_try(object, mreq->pindex + 1,
1247 vm_object_drop(object);
1248 return(VM_PAGER_OK);
1249 } else if (m == NULL) {
1251 * Use VM_ALLOC_QUICK to avoid blocking on cache
1254 m = vm_page_alloc(object, mreq->pindex + 1,
1257 vm_object_drop(object);
1258 return(VM_PAGER_OK);
1263 vm_object_drop(object);
1264 return(VM_PAGER_OK);
1266 vm_page_unqueue_nowakeup(m);
1276 * Try to block-read contiguous pages from swap if sequential,
1277 * otherwise just read one page. Contiguous pages from swap must
1278 * reside within a single device stripe because the I/O cannot be
1279 * broken up across multiple stripes.
1281 * Note that blk and iblk can be SWAPBLK_NONE but the loop is
1282 * set up such that the case(s) are handled implicitly.
1284 blk = swp_pager_meta_ctl(mreq->object, mreq->pindex, 0);
1287 for (i = 1; swap_burst_read &&
1288 i < XIO_INTERNAL_PAGES &&
1289 mreq->pindex + i < object->size; ++i) {
1292 iblk = swp_pager_meta_ctl(object, mreq->pindex + i, 0);
1293 if (iblk != blk + i)
1295 if ((blk ^ iblk) & dmmax_mask)
1297 m = vm_page_lookup_busy_try(object, mreq->pindex + i,
1301 } else if (m == NULL) {
1303 * Use VM_ALLOC_QUICK to avoid blocking on cache
1306 m = vm_page_alloc(object, mreq->pindex + i,
1315 vm_page_unqueue_nowakeup(m);
1321 vm_page_flag_set(marray[i - 1], PG_RAM);
1324 * If mreq is the requested page and we have nothing to do return
1325 * VM_PAGER_FAIL. If raonly is set mreq is just another read-ahead
1326 * page and must be cleaned up.
1328 if (blk == SWAPBLK_NONE) {
1331 vnode_pager_freepage(mreq);
1332 vm_object_drop(object);
1333 return(VM_PAGER_OK);
1335 vm_object_drop(object);
1336 return(VM_PAGER_FAIL);
1341 * map our page(s) into kva for input
1343 bp = getpbuf_kva(&nsw_rcount);
1345 kva = (vm_offset_t) bp->b_kvabase;
1346 bcopy(marray, bp->b_xio.xio_pages, i * sizeof(vm_page_t));
1347 pmap_qenter(kva, bp->b_xio.xio_pages, i);
1349 bp->b_data = (caddr_t)kva;
1350 bp->b_bcount = PAGE_SIZE * i;
1351 bp->b_xio.xio_npages = i;
1352 bio->bio_done = swp_pager_async_iodone;
1353 bio->bio_offset = (off_t)blk << PAGE_SHIFT;
1354 bio->bio_caller_info1.index = SWBIO_READ;
1357 * Set index. If raonly set the index beyond the array so all
1358 * the pages are treated the same, otherwise the original mreq is
1362 bio->bio_driver_info = (void *)(intptr_t)i;
1364 bio->bio_driver_info = (void *)(intptr_t)0;
1366 for (j = 0; j < i; ++j)
1367 vm_page_flag_set(bp->b_xio.xio_pages[j], PG_SWAPINPROG);
1369 mycpu->gd_cnt.v_swapin++;
1370 mycpu->gd_cnt.v_swappgsin += bp->b_xio.xio_npages;
1373 * We still hold the lock on mreq, and our automatic completion routine
1374 * does not remove it.
1376 vm_object_pip_add(object, bp->b_xio.xio_npages);
1379 * perform the I/O. NOTE!!! bp cannot be considered valid after
1380 * this point because we automatically release it on completion.
1381 * Instead, we look at the one page we are interested in which we
1382 * still hold a lock on even through the I/O completion.
1384 * The other pages in our m[] array are also released on completion,
1385 * so we cannot assume they are valid anymore either.
1387 bp->b_cmd = BUF_CMD_READ;
1389 vn_strategy(swapdev_vp, bio);
1392 * Wait for the page we want to complete. PG_SWAPINPROG is always
1393 * cleared on completion. If an I/O error occurs, SWAPBLK_NONE
1394 * is set in the meta-data.
1396 * If this is a read-ahead only we return immediately without
1400 vm_object_drop(object);
1401 return(VM_PAGER_OK);
1405 * Read-ahead includes originally requested page case.
1408 flags = mreq->flags;
1410 if ((flags & PG_SWAPINPROG) == 0)
1412 tsleep_interlock(mreq, 0);
1413 if (!atomic_cmpset_int(&mreq->flags, flags,
1414 flags | PG_WANTED | PG_REFERENCED)) {
1417 mycpu->gd_cnt.v_intrans++;
1418 if (tsleep(mreq, PINTERLOCKED, "swread", hz*20)) {
1420 "swap_pager: indefinite wait buffer: "
1421 " offset: %lld, size: %ld\n",
1422 (long long)bio->bio_offset,
1429 * mreq is left bussied after completion, but all the other pages
1430 * are freed. If we had an unrecoverable read error the page will
1433 vm_object_drop(object);
1434 if (mreq->valid != VM_PAGE_BITS_ALL)
1435 return(VM_PAGER_ERROR);
1437 return(VM_PAGER_OK);
1440 * A final note: in a low swap situation, we cannot deallocate swap
1441 * and mark a page dirty here because the caller is likely to mark
1442 * the page clean when we return, causing the page to possibly revert
1443 * to all-zero's later.
1448 * swap_pager_putpages:
1450 * Assign swap (if necessary) and initiate I/O on the specified pages.
1452 * We support both OBJT_DEFAULT and OBJT_SWAP objects. DEFAULT objects
1453 * are automatically converted to SWAP objects.
1455 * In a low memory situation we may block in vn_strategy(), but the new
1456 * vm_page reservation system coupled with properly written VFS devices
1457 * should ensure that no low-memory deadlock occurs. This is an area
1460 * The parent has N vm_object_pip_add() references prior to
1461 * calling us and will remove references for rtvals[] that are
1462 * not set to VM_PAGER_PEND. We need to remove the rest on I/O
1465 * The parent has soft-busy'd the pages it passes us and will unbusy
1466 * those whos rtvals[] entry is not set to VM_PAGER_PEND on return.
1467 * We need to unbusy the rest on I/O completion.
1472 swap_pager_putpages(vm_object_t object, vm_page_t *m, int count,
1473 boolean_t sync, int *rtvals)
1478 vm_object_hold(object);
1480 if (count && m[0]->object != object) {
1481 panic("swap_pager_getpages: object mismatch %p/%p",
1490 * Turn object into OBJT_SWAP
1491 * check for bogus sysops
1492 * force sync if not pageout process
1494 if (object->type == OBJT_DEFAULT) {
1495 if (object->type == OBJT_DEFAULT)
1496 swp_pager_meta_convert(object);
1499 if (curthread != pagethread)
1505 * Update nsw parameters from swap_async_max sysctl values.
1506 * Do not let the sysop crash the machine with bogus numbers.
1508 if (swap_async_max != nsw_wcount_async_max) {
1514 if ((n = swap_async_max) > nswbuf / 2)
1521 * Adjust difference ( if possible ). If the current async
1522 * count is too low, we may not be able to make the adjustment
1525 * vm_token needed for nsw_wcount sleep interlock
1527 lwkt_gettoken(&vm_token);
1528 n -= nsw_wcount_async_max;
1529 if (nsw_wcount_async + n >= 0) {
1530 nsw_wcount_async_max += n;
1531 pbuf_adjcount(&nsw_wcount_async, n);
1533 lwkt_reltoken(&vm_token);
1539 * Assign swap blocks and issue I/O. We reallocate swap on the fly.
1540 * The page is left dirty until the pageout operation completes
1544 for (i = 0; i < count; i += n) {
1551 * Maximum I/O size is limited by a number of factors.
1554 n = min(BLIST_MAX_ALLOC, count - i);
1555 n = min(n, nsw_cluster_max);
1557 lwkt_gettoken(&vm_token);
1560 * Get biggest block of swap we can. If we fail, fall
1561 * back and try to allocate a smaller block. Don't go
1562 * overboard trying to allocate space if it would overly
1566 (blk = swp_pager_getswapspace(object, n)) == SWAPBLK_NONE &&
1571 if (blk == SWAPBLK_NONE) {
1572 for (j = 0; j < n; ++j)
1573 rtvals[i+j] = VM_PAGER_FAIL;
1574 lwkt_reltoken(&vm_token);
1577 if (vm_report_swap_allocs > 0) {
1578 kprintf("swap_alloc %08jx,%d\n", (intmax_t)blk, n);
1579 --vm_report_swap_allocs;
1583 * The I/O we are constructing cannot cross a physical
1584 * disk boundry in the swap stripe. Note: we are still
1587 if ((blk ^ (blk + n)) & dmmax_mask) {
1588 j = ((blk + dmmax) & dmmax_mask) - blk;
1589 swp_pager_freeswapspace(object, blk + j, n - j);
1594 * All I/O parameters have been satisfied, build the I/O
1595 * request and assign the swap space.
1598 bp = getpbuf_kva(&nsw_wcount_sync);
1600 bp = getpbuf_kva(&nsw_wcount_async);
1603 lwkt_reltoken(&vm_token);
1605 pmap_qenter((vm_offset_t)bp->b_data, &m[i], n);
1607 bp->b_bcount = PAGE_SIZE * n;
1608 bio->bio_offset = (off_t)blk << PAGE_SHIFT;
1610 for (j = 0; j < n; ++j) {
1611 vm_page_t mreq = m[i+j];
1613 swp_pager_meta_build(mreq->object, mreq->pindex,
1615 if (object->type == OBJT_SWAP)
1616 vm_page_dirty(mreq);
1617 rtvals[i+j] = VM_PAGER_OK;
1619 vm_page_flag_set(mreq, PG_SWAPINPROG);
1620 bp->b_xio.xio_pages[j] = mreq;
1622 bp->b_xio.xio_npages = n;
1624 mycpu->gd_cnt.v_swapout++;
1625 mycpu->gd_cnt.v_swappgsout += bp->b_xio.xio_npages;
1627 bp->b_dirtyoff = 0; /* req'd for NFS */
1628 bp->b_dirtyend = bp->b_bcount; /* req'd for NFS */
1629 bp->b_cmd = BUF_CMD_WRITE;
1630 bio->bio_caller_info1.index = SWBIO_WRITE;
1635 if (sync == FALSE) {
1636 bio->bio_done = swp_pager_async_iodone;
1638 vn_strategy(swapdev_vp, bio);
1640 for (j = 0; j < n; ++j)
1641 rtvals[i+j] = VM_PAGER_PEND;
1646 * Issue synchrnously.
1648 * Wait for the sync I/O to complete, then update rtvals.
1649 * We just set the rtvals[] to VM_PAGER_PEND so we can call
1650 * our async completion routine at the end, thus avoiding a
1653 bio->bio_caller_info1.index |= SWBIO_SYNC;
1654 bio->bio_done = biodone_sync;
1655 bio->bio_flags |= BIO_SYNC;
1656 vn_strategy(swapdev_vp, bio);
1657 biowait(bio, "swwrt");
1659 for (j = 0; j < n; ++j)
1660 rtvals[i+j] = VM_PAGER_PEND;
1663 * Now that we are through with the bp, we can call the
1664 * normal async completion, which frees everything up.
1666 swp_pager_async_iodone(bio);
1668 vm_object_drop(object);
1675 swap_pager_newswap(void)
1681 * swp_pager_async_iodone:
1683 * Completion routine for asynchronous reads and writes from/to swap.
1684 * Also called manually by synchronous code to finish up a bp.
1686 * For READ operations, the pages are PG_BUSY'd. For WRITE operations,
1687 * the pages are vm_page_t->busy'd. For READ operations, we PG_BUSY
1688 * unbusy all pages except the 'main' request page. For WRITE
1689 * operations, we vm_page_t->busy'd unbusy all pages ( we can do this
1690 * because we marked them all VM_PAGER_PEND on return from putpages ).
1692 * This routine may not block.
1697 swp_pager_async_iodone(struct bio *bio)
1699 struct buf *bp = bio->bio_buf;
1700 vm_object_t object = NULL;
1707 if (bp->b_flags & B_ERROR) {
1709 "swap_pager: I/O error - %s failed; offset %lld,"
1710 "size %ld, error %d\n",
1711 ((bio->bio_caller_info1.index & SWBIO_READ) ?
1712 "pagein" : "pageout"),
1713 (long long)bio->bio_offset,
1720 * set object, raise to splvm().
1722 if (bp->b_xio.xio_npages)
1723 object = bp->b_xio.xio_pages[0]->object;
1726 * remove the mapping for kernel virtual
1728 pmap_qremove((vm_offset_t)bp->b_data, bp->b_xio.xio_npages);
1731 * cleanup pages. If an error occurs writing to swap, we are in
1732 * very serious trouble. If it happens to be a disk error, though,
1733 * we may be able to recover by reassigning the swap later on. So
1734 * in this case we remove the m->swapblk assignment for the page
1735 * but do not free it in the rlist. The errornous block(s) are thus
1736 * never reallocated as swap. Redirty the page and continue.
1738 for (i = 0; i < bp->b_xio.xio_npages; ++i) {
1739 vm_page_t m = bp->b_xio.xio_pages[i];
1741 if (bp->b_flags & B_ERROR) {
1743 * If an error occurs I'd love to throw the swapblk
1744 * away without freeing it back to swapspace, so it
1745 * can never be used again. But I can't from an
1749 if (bio->bio_caller_info1.index & SWBIO_READ) {
1751 * When reading, reqpage needs to stay
1752 * locked for the parent, but all other
1753 * pages can be freed. We still want to
1754 * wakeup the parent waiting on the page,
1755 * though. ( also: pg_reqpage can be -1 and
1756 * not match anything ).
1758 * We have to wake specifically requested pages
1759 * up too because we cleared PG_SWAPINPROG and
1760 * someone may be waiting for that.
1762 * NOTE: for reads, m->dirty will probably
1763 * be overridden by the original caller of
1764 * getpages so don't play cute tricks here.
1766 * NOTE: We can't actually free the page from
1767 * here, because this is an interrupt. It
1768 * is not legal to mess with object->memq
1769 * from an interrupt. Deactivate the page
1774 vm_page_flag_clear(m, PG_ZERO);
1775 vm_page_flag_clear(m, PG_SWAPINPROG);
1778 * bio_driver_info holds the requested page
1781 if (i != (int)(intptr_t)bio->bio_driver_info) {
1782 vm_page_deactivate(m);
1788 * If i == bp->b_pager.pg_reqpage, do not wake
1789 * the page up. The caller needs to.
1793 * If a write error occurs remove the swap
1794 * assignment (note that PG_SWAPPED may or
1795 * may not be set depending on prior activity).
1797 * Re-dirty OBJT_SWAP pages as there is no
1798 * other backing store, we can't throw the
1801 * Non-OBJT_SWAP pages (aka swapcache) must
1802 * not be dirtied since they may not have
1803 * been dirty in the first place, and they
1804 * do have backing store (the vnode).
1806 vm_page_busy_wait(m, FALSE, "swadpg");
1807 swp_pager_meta_ctl(m->object, m->pindex,
1809 vm_page_flag_clear(m, PG_SWAPPED);
1810 if (m->object->type == OBJT_SWAP) {
1812 vm_page_activate(m);
1814 vm_page_flag_clear(m, PG_SWAPINPROG);
1815 vm_page_io_finish(m);
1818 } else if (bio->bio_caller_info1.index & SWBIO_READ) {
1820 * NOTE: for reads, m->dirty will probably be
1821 * overridden by the original caller of getpages so
1822 * we cannot set them in order to free the underlying
1823 * swap in a low-swap situation. I don't think we'd
1824 * want to do that anyway, but it was an optimization
1825 * that existed in the old swapper for a time before
1826 * it got ripped out due to precisely this problem.
1828 * clear PG_ZERO in page.
1830 * If not the requested page then deactivate it.
1832 * Note that the requested page, reqpage, is left
1833 * busied, but we still have to wake it up. The
1834 * other pages are released (unbusied) by
1835 * vm_page_wakeup(). We do not set reqpage's
1836 * valid bits here, it is up to the caller.
1840 * NOTE: can't call pmap_clear_modify(m) from an
1841 * interrupt thread, the pmap code may have to map
1842 * non-kernel pmaps and currently asserts the case.
1844 /*pmap_clear_modify(m);*/
1845 m->valid = VM_PAGE_BITS_ALL;
1847 vm_page_flag_clear(m, PG_ZERO | PG_SWAPINPROG);
1848 vm_page_flag_set(m, PG_SWAPPED);
1851 * We have to wake specifically requested pages
1852 * up too because we cleared PG_SWAPINPROG and
1853 * could be waiting for it in getpages. However,
1854 * be sure to not unbusy getpages specifically
1855 * requested page - getpages expects it to be
1858 * bio_driver_info holds the requested page
1860 if (i != (int)(intptr_t)bio->bio_driver_info) {
1861 vm_page_deactivate(m);
1868 * Mark the page clean but do not mess with the
1869 * pmap-layer's modified state. That state should
1870 * also be clear since the caller protected the
1871 * page VM_PROT_READ, but allow the case.
1873 * We are in an interrupt, avoid pmap operations.
1875 * If we have a severe page deficit, deactivate the
1876 * page. Do not try to cache it (which would also
1877 * involve a pmap op), because the page might still
1880 * When using the swap to cache clean vnode pages
1881 * we do not mess with the page dirty bits.
1883 vm_page_busy_wait(m, FALSE, "swadpg");
1884 if (m->object->type == OBJT_SWAP)
1886 vm_page_flag_clear(m, PG_SWAPINPROG);
1887 vm_page_flag_set(m, PG_SWAPPED);
1888 if (vm_page_count_severe())
1889 vm_page_deactivate(m);
1891 if (!vm_page_count_severe() || !vm_page_try_to_cache(m))
1892 vm_page_protect(m, VM_PROT_READ);
1894 vm_page_io_finish(m);
1900 * adjust pip. NOTE: the original parent may still have its own
1901 * pip refs on the object.
1905 vm_object_pip_wakeup_n(object, bp->b_xio.xio_npages);
1908 * Release the physical I/O buffer.
1910 * NOTE: Due to synchronous operations in the write case b_cmd may
1911 * already be set to BUF_CMD_DONE and BIO_SYNC may have already
1914 * Use vm_token to interlock nsw_rcount/wcount wakeup?
1916 lwkt_gettoken(&vm_token);
1917 if (bio->bio_caller_info1.index & SWBIO_READ)
1918 nswptr = &nsw_rcount;
1919 else if (bio->bio_caller_info1.index & SWBIO_SYNC)
1920 nswptr = &nsw_wcount_sync;
1922 nswptr = &nsw_wcount_async;
1923 bp->b_cmd = BUF_CMD_DONE;
1924 relpbuf(bp, nswptr);
1925 lwkt_reltoken(&vm_token);
1929 * Fault-in a potentially swapped page and remove the swap reference.
1930 * (used by swapoff code)
1932 * object must be held.
1934 static __inline void
1935 swp_pager_fault_page(vm_object_t object, vm_pindex_t pindex)
1941 ASSERT_LWKT_TOKEN_HELD(vm_object_token(object));
1943 if (object->type == OBJT_VNODE) {
1945 * Any swap related to a vnode is due to swapcache. We must
1946 * vget() the vnode in case it is not active (otherwise
1947 * vref() will panic). Calling vm_object_page_remove() will
1948 * ensure that any swap ref is removed interlocked with the
1949 * page. clean_only is set to TRUE so we don't throw away
1952 vp = object->handle;
1953 error = vget(vp, LK_SHARED | LK_RETRY | LK_CANRECURSE);
1955 vm_object_page_remove(object, pindex, pindex + 1, TRUE);
1960 * Otherwise it is a normal OBJT_SWAP object and we can
1961 * fault the page in and remove the swap.
1963 m = vm_fault_object_page(object, IDX_TO_OFF(pindex),
1965 VM_FAULT_DIRTY | VM_FAULT_UNSWAP,
1973 * This removes all swap blocks related to a particular device. We have
1974 * to be careful of ripups during the scan.
1976 static int swp_pager_swapoff_callback(struct swblock *swap, void *data);
1979 swap_pager_swapoff(int devidx)
1981 struct vm_object marker;
1983 struct swswapoffinfo info;
1985 bzero(&marker, sizeof(marker));
1986 marker.type = OBJT_MARKER;
1988 lwkt_gettoken(&vmobj_token);
1989 TAILQ_INSERT_HEAD(&vm_object_list, &marker, object_list);
1991 while ((object = TAILQ_NEXT(&marker, object_list)) != NULL) {
1992 if (object->type == OBJT_MARKER)
1994 if (object->type != OBJT_SWAP && object->type != OBJT_VNODE)
1996 vm_object_hold(object);
1997 if (object->type != OBJT_SWAP && object->type != OBJT_VNODE) {
1998 vm_object_drop(object);
2001 info.object = object;
2002 info.devidx = devidx;
2003 swblock_rb_tree_RB_SCAN(&object->swblock_root,
2005 swp_pager_swapoff_callback,
2007 vm_object_drop(object);
2009 if (object == TAILQ_NEXT(&marker, object_list)) {
2010 TAILQ_REMOVE(&vm_object_list, &marker, object_list);
2011 TAILQ_INSERT_AFTER(&vm_object_list, object,
2012 &marker, object_list);
2015 TAILQ_REMOVE(&vm_object_list, &marker, object_list);
2016 lwkt_reltoken(&vmobj_token);
2019 * If we fail to locate all swblocks we just fail gracefully and
2020 * do not bother to restore paging on the swap device. If the
2021 * user wants to retry the user can retry.
2023 if (swdevt[devidx].sw_nused)
2031 swp_pager_swapoff_callback(struct swblock *swap, void *data)
2033 struct swswapoffinfo *info = data;
2034 vm_object_t object = info->object;
2039 index = swap->swb_index;
2040 for (i = 0; i < SWAP_META_PAGES; ++i) {
2042 * Make sure we don't race a dying object. This will
2043 * kill the scan of the object's swap blocks entirely.
2045 if (object->flags & OBJ_DEAD)
2049 * Fault the page, which can obviously block. If the swap
2050 * structure disappears break out.
2052 v = swap->swb_pages[i];
2053 if (v != SWAPBLK_NONE && BLK2DEVIDX(v) == info->devidx) {
2054 swp_pager_fault_page(object, swap->swb_index + i);
2055 /* swap ptr might go away */
2056 if (RB_LOOKUP(swblock_rb_tree,
2057 &object->swblock_root, index) != swap) {
2065 /************************************************************************
2067 ************************************************************************
2069 * These routines manipulate the swap metadata stored in the
2070 * OBJT_SWAP object. All swp_*() routines must be called at
2071 * splvm() because swap can be freed up by the low level vm_page
2072 * code which might be called from interrupts beyond what splbio() covers.
2074 * Swap metadata is implemented with a global hash and not directly
2075 * linked into the object. Instead the object simply contains
2076 * appropriate tracking counters.
2080 * Lookup the swblock containing the specified swap block index.
2082 * The caller must hold the object.
2086 swp_pager_lookup(vm_object_t object, vm_pindex_t index)
2088 ASSERT_LWKT_TOKEN_HELD(vm_object_token(object));
2089 index &= ~(vm_pindex_t)SWAP_META_MASK;
2090 return (RB_LOOKUP(swblock_rb_tree, &object->swblock_root, index));
2094 * Remove a swblock from the RB tree.
2096 * The caller must hold the object.
2100 swp_pager_remove(vm_object_t object, struct swblock *swap)
2102 ASSERT_LWKT_TOKEN_HELD(vm_object_token(object));
2103 RB_REMOVE(swblock_rb_tree, &object->swblock_root, swap);
2107 * Convert default object to swap object if necessary
2109 * The caller must hold the object.
2112 swp_pager_meta_convert(vm_object_t object)
2114 if (object->type == OBJT_DEFAULT) {
2115 object->type = OBJT_SWAP;
2116 KKASSERT(object->swblock_count == 0);
2121 * SWP_PAGER_META_BUILD() - add swap block to swap meta data for object
2123 * We first convert the object to a swap object if it is a default
2124 * object. Vnode objects do not need to be converted.
2126 * The specified swapblk is added to the object's swap metadata. If
2127 * the swapblk is not valid, it is freed instead. Any previously
2128 * assigned swapblk is freed.
2130 * The caller must hold the object.
2133 swp_pager_meta_build(vm_object_t object, vm_pindex_t index, swblk_t swapblk)
2135 struct swblock *swap;
2136 struct swblock *oswap;
2139 KKASSERT(swapblk != SWAPBLK_NONE);
2140 ASSERT_LWKT_TOKEN_HELD(vm_object_token(object));
2143 * Convert object if necessary
2145 if (object->type == OBJT_DEFAULT)
2146 swp_pager_meta_convert(object);
2149 * Locate swblock. If not found create, but if we aren't adding
2150 * anything just return. If we run out of space in the map we wait
2151 * and, since the hash table may have changed, retry.
2154 swap = swp_pager_lookup(object, index);
2159 swap = zalloc(swap_zone);
2164 swap->swb_index = index & ~(vm_pindex_t)SWAP_META_MASK;
2165 swap->swb_count = 0;
2167 ++object->swblock_count;
2169 for (i = 0; i < SWAP_META_PAGES; ++i)
2170 swap->swb_pages[i] = SWAPBLK_NONE;
2171 oswap = RB_INSERT(swblock_rb_tree, &object->swblock_root, swap);
2172 KKASSERT(oswap == NULL);
2176 * Delete prior contents of metadata.
2178 * NOTE: Decrement swb_count after the freeing operation (which
2179 * might block) to prevent racing destruction of the swblock.
2181 index &= SWAP_META_MASK;
2183 while ((v = swap->swb_pages[index]) != SWAPBLK_NONE) {
2184 swap->swb_pages[index] = SWAPBLK_NONE;
2186 swp_pager_freeswapspace(object, v, 1);
2188 --mycpu->gd_vmtotal.t_vm;
2192 * Enter block into metadata
2194 swap->swb_pages[index] = swapblk;
2195 if (swapblk != SWAPBLK_NONE) {
2197 ++mycpu->gd_vmtotal.t_vm;
2202 * SWP_PAGER_META_FREE() - free a range of blocks in the object's swap metadata
2204 * The requested range of blocks is freed, with any associated swap
2205 * returned to the swap bitmap.
2207 * This routine will free swap metadata structures as they are cleaned
2208 * out. This routine does *NOT* operate on swap metadata associated
2209 * with resident pages.
2211 * The caller must hold the object.
2213 static int swp_pager_meta_free_callback(struct swblock *swb, void *data);
2216 swp_pager_meta_free(vm_object_t object, vm_pindex_t index, vm_pindex_t count)
2218 struct swfreeinfo info;
2220 ASSERT_LWKT_TOKEN_HELD(vm_object_token(object));
2225 if (object->swblock_count == 0) {
2226 KKASSERT(RB_EMPTY(&object->swblock_root));
2233 * Setup for RB tree scan. Note that the pindex range can be huge
2234 * due to the 64 bit page index space so we cannot safely iterate.
2236 info.object = object;
2237 info.basei = index & ~(vm_pindex_t)SWAP_META_MASK;
2239 info.endi = index + count - 1;
2240 swblock_rb_tree_RB_SCAN(&object->swblock_root, rb_swblock_scancmp,
2241 swp_pager_meta_free_callback, &info);
2245 * The caller must hold the object.
2249 swp_pager_meta_free_callback(struct swblock *swap, void *data)
2251 struct swfreeinfo *info = data;
2252 vm_object_t object = info->object;
2257 * Figure out the range within the swblock. The wider scan may
2258 * return edge-case swap blocks when the start and/or end points
2259 * are in the middle of a block.
2261 if (swap->swb_index < info->begi)
2262 index = (int)info->begi & SWAP_META_MASK;
2266 if (swap->swb_index + SWAP_META_PAGES > info->endi)
2267 eindex = (int)info->endi & SWAP_META_MASK;
2269 eindex = SWAP_META_MASK;
2272 * Scan and free the blocks. The loop terminates early
2273 * if (swap) runs out of blocks and could be freed.
2275 * NOTE: Decrement swb_count after swp_pager_freeswapspace()
2276 * to deal with a zfree race.
2278 while (index <= eindex) {
2279 swblk_t v = swap->swb_pages[index];
2281 if (v != SWAPBLK_NONE) {
2282 swap->swb_pages[index] = SWAPBLK_NONE;
2284 swp_pager_freeswapspace(object, v, 1);
2285 --mycpu->gd_vmtotal.t_vm;
2286 if (--swap->swb_count == 0) {
2287 swp_pager_remove(object, swap);
2288 zfree(swap_zone, swap);
2289 --object->swblock_count;
2296 /* swap may be invalid here due to zfree above */
2303 * SWP_PAGER_META_FREE_ALL() - destroy all swap metadata associated with object
2305 * This routine locates and destroys all swap metadata associated with
2308 * NOTE: Decrement swb_count after the freeing operation (which
2309 * might block) to prevent racing destruction of the swblock.
2311 * The caller must hold the object.
2314 swp_pager_meta_free_all(vm_object_t object)
2316 struct swblock *swap;
2319 ASSERT_LWKT_TOKEN_HELD(vm_object_token(object));
2321 while ((swap = RB_ROOT(&object->swblock_root)) != NULL) {
2322 swp_pager_remove(object, swap);
2323 for (i = 0; i < SWAP_META_PAGES; ++i) {
2324 swblk_t v = swap->swb_pages[i];
2325 if (v != SWAPBLK_NONE) {
2327 swp_pager_freeswapspace(object, v, 1);
2329 --mycpu->gd_vmtotal.t_vm;
2332 if (swap->swb_count != 0)
2333 panic("swap_pager_meta_free_all: swb_count != 0");
2334 zfree(swap_zone, swap);
2335 --object->swblock_count;
2338 KKASSERT(object->swblock_count == 0);
2342 * SWP_PAGER_METACTL() - misc control of swap and vm_page_t meta data.
2344 * This routine is capable of looking up, popping, or freeing
2345 * swapblk assignments in the swap meta data or in the vm_page_t.
2346 * The routine typically returns the swapblk being looked-up, or popped,
2347 * or SWAPBLK_NONE if the block was freed, or SWAPBLK_NONE if the block
2348 * was invalid. This routine will automatically free any invalid
2349 * meta-data swapblks.
2351 * It is not possible to store invalid swapblks in the swap meta data
2352 * (other then a literal 'SWAPBLK_NONE'), so we don't bother checking.
2354 * When acting on a busy resident page and paging is in progress, we
2355 * have to wait until paging is complete but otherwise can act on the
2358 * SWM_FREE remove and free swap block from metadata
2359 * SWM_POP remove from meta data but do not free.. pop it out
2361 * The caller must hold the object.
2364 swp_pager_meta_ctl(vm_object_t object, vm_pindex_t index, int flags)
2366 struct swblock *swap;
2369 if (object->swblock_count == 0)
2370 return(SWAPBLK_NONE);
2373 swap = swp_pager_lookup(object, index);
2376 index &= SWAP_META_MASK;
2377 r1 = swap->swb_pages[index];
2379 if (r1 != SWAPBLK_NONE) {
2380 if (flags & (SWM_FREE|SWM_POP)) {
2381 swap->swb_pages[index] = SWAPBLK_NONE;
2382 --mycpu->gd_vmtotal.t_vm;
2383 if (--swap->swb_count == 0) {
2384 swp_pager_remove(object, swap);
2385 zfree(swap_zone, swap);
2386 --object->swblock_count;
2389 /* swap ptr may be invalid */
2390 if (flags & SWM_FREE) {
2391 swp_pager_freeswapspace(object, r1, 1);
2395 /* swap ptr may be invalid */