4 * Copyright (c) 1998-2010 The DragonFly Project. All rights reserved.
6 * This code is derived from software contributed to The DragonFly Project
7 * by Matthew Dillon <dillon@backplane.com>
9 * Redistribution and use in source and binary forms, with or without
10 * modification, are permitted provided that the following conditions
13 * 1. Redistributions of source code must retain the above copyright
14 * notice, this list of conditions and the following disclaimer.
15 * 2. Redistributions in binary form must reproduce the above copyright
16 * notice, this list of conditions and the following disclaimer in
17 * the documentation and/or other materials provided with the
19 * 3. Neither the name of The DragonFly Project nor the names of its
20 * contributors may be used to endorse or promote products derived
21 * from this software without specific, prior written permission.
23 * THIS SOFTWARE IS PROVIDED BY THE COPYRIGHT HOLDERS AND CONTRIBUTORS
24 * ``AS IS'' AND ANY EXPRESS OR IMPLIED WARRANTIES, INCLUDING, BUT NOT
25 * LIMITED TO, THE IMPLIED WARRANTIES OF MERCHANTABILITY AND FITNESS
26 * FOR A PARTICULAR PURPOSE ARE DISCLAIMED. IN NO EVENT SHALL THE
27 * COPYRIGHT HOLDERS OR CONTRIBUTORS BE LIABLE FOR ANY DIRECT, INDIRECT,
28 * INCIDENTAL, SPECIAL, EXEMPLARY OR CONSEQUENTIAL DAMAGES (INCLUDING,
29 * BUT NOT LIMITED TO, PROCUREMENT OF SUBSTITUTE GOODS OR SERVICES;
30 * LOSS OF USE, DATA, OR PROFITS; OR BUSINESS INTERRUPTION) HOWEVER CAUSED
31 * AND ON ANY THEORY OF LIABILITY, WHETHER IN CONTRACT, STRICT LIABILITY,
32 * OR TORT (INCLUDING NEGLIGENCE OR OTHERWISE) ARISING IN ANY WAY OUT
33 * OF THE USE OF THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF
36 * Copyright (c) 1994 John S. Dyson
37 * Copyright (c) 1990 University of Utah.
38 * Copyright (c) 1991, 1993
39 * The Regents of the University of California. All rights reserved.
41 * This code is derived from software contributed to Berkeley by
42 * the Systems Programming Group of the University of Utah Computer
45 * Redistribution and use in source and binary forms, with or without
46 * modification, are permitted provided that the following conditions
48 * 1. Redistributions of source code must retain the above copyright
49 * notice, this list of conditions and the following disclaimer.
50 * 2. Redistributions in binary form must reproduce the above copyright
51 * notice, this list of conditions and the following disclaimer in the
52 * documentation and/or other materials provided with the distribution.
53 * 3. All advertising materials mentioning features or use of this software
54 * must display the following acknowledgement:
55 * This product includes software developed by the University of
56 * California, Berkeley and its contributors.
57 * 4. Neither the name of the University nor the names of its contributors
58 * may be used to endorse or promote products derived from this software
59 * without specific prior written permission.
61 * THIS SOFTWARE IS PROVIDED BY THE REGENTS AND CONTRIBUTORS ``AS IS'' AND
62 * ANY EXPRESS OR IMPLIED WARRANTIES, INCLUDING, BUT NOT LIMITED TO, THE
63 * IMPLIED WARRANTIES OF MERCHANTABILITY AND FITNESS FOR A PARTICULAR PURPOSE
64 * ARE DISCLAIMED. IN NO EVENT SHALL THE REGENTS OR CONTRIBUTORS BE LIABLE
65 * FOR ANY DIRECT, INDIRECT, INCIDENTAL, SPECIAL, EXEMPLARY, OR CONSEQUENTIAL
66 * DAMAGES (INCLUDING, BUT NOT LIMITED TO, PROCUREMENT OF SUBSTITUTE GOODS
67 * OR SERVICES; LOSS OF USE, DATA, OR PROFITS; OR BUSINESS INTERRUPTION)
68 * HOWEVER CAUSED AND ON ANY THEORY OF LIABILITY, WHETHER IN CONTRACT, STRICT
69 * LIABILITY, OR TORT (INCLUDING NEGLIGENCE OR OTHERWISE) ARISING IN ANY WAY
70 * OUT OF THE USE OF THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF
76 * Radix Bitmap 'blists'.
78 * - The new swapper uses the new radix bitmap code. This should scale
79 * to arbitrarily small or arbitrarily large swap spaces and an almost
80 * arbitrary degree of fragmentation.
84 * - on the fly reallocation of swap during putpages. The new system
85 * does not try to keep previously allocated swap blocks for dirty
88 * - on the fly deallocation of swap
90 * - No more garbage collection required. Unnecessarily allocated swap
91 * blocks only exist for dirty vm_page_t's now and these are already
92 * cycled (in a high-load system) by the pager. We also do on-the-fly
93 * removal of invalidated swap blocks when a page is destroyed
96 * from: Utah $Hdr: swap_pager.c 1.4 91/04/30$
97 * @(#)swap_pager.c 8.9 (Berkeley) 3/21/94
98 * $FreeBSD: src/sys/vm/swap_pager.c,v 1.130.2.12 2002/08/31 21:15:55 dillon Exp $
101 #include <sys/param.h>
102 #include <sys/systm.h>
103 #include <sys/conf.h>
104 #include <sys/kernel.h>
105 #include <sys/proc.h>
107 #include <sys/vnode.h>
108 #include <sys/malloc.h>
109 #include <sys/vmmeter.h>
110 #include <sys/sysctl.h>
111 #include <sys/blist.h>
112 #include <sys/lock.h>
113 #include <sys/thread2.h>
115 #ifndef MAX_PAGEOUT_CLUSTER
116 #define MAX_PAGEOUT_CLUSTER 16
119 #define SWB_NPAGES MAX_PAGEOUT_CLUSTER
121 #include "opt_swap.h"
123 #include <vm/vm_object.h>
124 #include <vm/vm_page.h>
125 #include <vm/vm_pager.h>
126 #include <vm/vm_pageout.h>
127 #include <vm/swap_pager.h>
128 #include <vm/vm_extern.h>
129 #include <vm/vm_zone.h>
130 #include <vm/vnode_pager.h>
132 #include <sys/buf2.h>
133 #include <vm/vm_page2.h>
135 #define SWM_FREE 0x02 /* free, period */
136 #define SWM_POP 0x04 /* pop out */
138 #define SWBIO_READ 0x01
139 #define SWBIO_WRITE 0x02
140 #define SWBIO_SYNC 0x04
146 vm_pindex_t endi; /* inclusive */
150 * vm_swap_size is in page-sized chunks now. It was DEV_BSIZE'd chunks
154 int swap_pager_full; /* swap space exhaustion (task killing) */
155 int vm_swap_cache_use;
156 int vm_swap_anon_use;
158 static int swap_pager_almost_full; /* swap space exhaustion (w/ hysteresis)*/
159 static int nsw_rcount; /* free read buffers */
160 static int nsw_wcount_sync; /* limit write buffers / synchronous */
161 static int nsw_wcount_async; /* limit write buffers / asynchronous */
162 static int nsw_wcount_async_max;/* assigned maximum */
163 static int nsw_cluster_max; /* maximum VOP I/O allowed */
165 struct blist *swapblist;
166 static int swap_async_max = 4; /* maximum in-progress async I/O's */
167 static int swap_burst_read = 0; /* allow burst reading */
169 extern struct vnode *swapdev_vp; /* from vm_swap.c */
171 SYSCTL_INT(_vm, OID_AUTO, swap_async_max,
172 CTLFLAG_RW, &swap_async_max, 0, "Maximum running async swap ops");
173 SYSCTL_INT(_vm, OID_AUTO, swap_burst_read,
174 CTLFLAG_RW, &swap_burst_read, 0, "Allow burst reads for pageins");
176 SYSCTL_INT(_vm, OID_AUTO, swap_cache_use,
177 CTLFLAG_RD, &vm_swap_cache_use, 0, "");
178 SYSCTL_INT(_vm, OID_AUTO, swap_anon_use,
179 CTLFLAG_RD, &vm_swap_anon_use, 0, "");
184 * Red-Black tree for swblock entries
186 * The caller must hold vm_token
188 RB_GENERATE2(swblock_rb_tree, swblock, swb_entry, rb_swblock_compare,
189 vm_pindex_t, swb_index);
192 rb_swblock_compare(struct swblock *swb1, struct swblock *swb2)
194 if (swb1->swb_index < swb2->swb_index)
196 if (swb1->swb_index > swb2->swb_index)
203 rb_swblock_scancmp(struct swblock *swb, void *data)
205 struct swfreeinfo *info = data;
207 if (swb->swb_index < info->basei)
209 if (swb->swb_index > info->endi)
216 rb_swblock_condcmp(struct swblock *swb, void *data)
218 struct swfreeinfo *info = data;
220 if (swb->swb_index < info->basei)
226 * pagerops for OBJT_SWAP - "swap pager". Some ops are also global procedure
227 * calls hooked from other parts of the VM system and do not appear here.
228 * (see vm/swap_pager.h).
231 static void swap_pager_dealloc (vm_object_t object);
232 static int swap_pager_getpage (vm_object_t, vm_page_t *, int);
233 static void swap_chain_iodone(struct bio *biox);
235 struct pagerops swappagerops = {
236 swap_pager_dealloc, /* deallocate an OBJT_SWAP object */
237 swap_pager_getpage, /* pagein */
238 swap_pager_putpages, /* pageout */
239 swap_pager_haspage /* get backing store status for page */
243 * dmmax is in page-sized chunks with the new swap system. It was
244 * dev-bsized chunks in the old. dmmax is always a power of 2.
246 * swap_*() routines are externally accessible. swp_*() routines are
251 static int dmmax_mask;
252 int nswap_lowat = 128; /* in pages, swap_pager_almost_full warn */
253 int nswap_hiwat = 512; /* in pages, swap_pager_almost_full warn */
255 static __inline void swp_sizecheck (void);
256 static void swp_pager_async_iodone (struct bio *bio);
259 * Swap bitmap functions
262 static __inline void swp_pager_freeswapspace(vm_object_t object,
263 swblk_t blk, int npages);
264 static __inline swblk_t swp_pager_getswapspace(vm_object_t object, int npages);
270 static void swp_pager_meta_convert(vm_object_t);
271 static void swp_pager_meta_build(vm_object_t, vm_pindex_t, swblk_t);
272 static void swp_pager_meta_free(vm_object_t, vm_pindex_t, vm_pindex_t);
273 static void swp_pager_meta_free_all(vm_object_t);
274 static swblk_t swp_pager_meta_ctl(vm_object_t, vm_pindex_t, int);
277 * SWP_SIZECHECK() - update swap_pager_full indication
279 * update the swap_pager_almost_full indication and warn when we are
280 * about to run out of swap space, using lowat/hiwat hysteresis.
282 * Clear swap_pager_full ( task killing ) indication when lowat is met.
284 * No restrictions on call
285 * This routine may not block.
291 if (vm_swap_size < nswap_lowat) {
292 if (swap_pager_almost_full == 0) {
293 kprintf("swap_pager: out of swap space\n");
294 swap_pager_almost_full = 1;
298 if (vm_swap_size > nswap_hiwat)
299 swap_pager_almost_full = 0;
304 * SWAP_PAGER_INIT() - initialize the swap pager!
306 * Expected to be started from system init. NOTE: This code is run
307 * before much else so be careful what you depend on. Most of the VM
308 * system has yet to be initialized at this point.
310 * Called from the low level boot code only.
313 swap_pager_init(void *arg __unused)
316 * Device Stripe, in PAGE_SIZE'd blocks
318 dmmax = SWB_NPAGES * 2;
319 dmmax_mask = ~(dmmax - 1);
321 SYSINIT(vm_mem, SI_BOOT1_VM, SI_ORDER_THIRD, swap_pager_init, NULL)
324 * SWAP_PAGER_SWAP_INIT() - swap pager initialization from pageout process
326 * Expected to be started from pageout process once, prior to entering
329 * Called from the low level boot code only.
332 swap_pager_swap_init(void)
337 * Number of in-transit swap bp operations. Don't
338 * exhaust the pbufs completely. Make sure we
339 * initialize workable values (0 will work for hysteresis
340 * but it isn't very efficient).
342 * The nsw_cluster_max is constrained by the number of pages an XIO
343 * holds, i.e., (MAXPHYS/PAGE_SIZE) and our locally defined
344 * MAX_PAGEOUT_CLUSTER. Also be aware that swap ops are
345 * constrained by the swap device interleave stripe size.
347 * Currently we hardwire nsw_wcount_async to 4. This limit is
348 * designed to prevent other I/O from having high latencies due to
349 * our pageout I/O. The value 4 works well for one or two active swap
350 * devices but is probably a little low if you have more. Even so,
351 * a higher value would probably generate only a limited improvement
352 * with three or four active swap devices since the system does not
353 * typically have to pageout at extreme bandwidths. We will want
354 * at least 2 per swap devices, and 4 is a pretty good value if you
355 * have one NFS swap device due to the command/ack latency over NFS.
356 * So it all works out pretty well.
359 nsw_cluster_max = min((MAXPHYS/PAGE_SIZE), MAX_PAGEOUT_CLUSTER);
361 nsw_rcount = (nswbuf + 1) / 2;
362 nsw_wcount_sync = (nswbuf + 3) / 4;
363 nsw_wcount_async = 4;
364 nsw_wcount_async_max = nsw_wcount_async;
367 * The zone is dynamically allocated so generally size it to
368 * maxswzone (32MB to 512MB of KVM). Set a minimum size based
369 * on physical memory of around 8x (each swblock can hold 16 pages).
371 * With the advent of SSDs (vs HDs) the practical (swap:memory) ratio
372 * has increased dramatically.
374 n = vmstats.v_page_count / 2;
375 if (maxswzone && n < maxswzone / sizeof(struct swblock))
376 n = maxswzone / sizeof(struct swblock);
382 sizeof(struct swblock),
386 if (swap_zone != NULL)
389 * if the allocation failed, try a zone two thirds the
390 * size of the previous attempt.
395 if (swap_zone == NULL)
396 panic("swap_pager_swap_init: swap_zone == NULL");
398 kprintf("Swap zone entries reduced from %d to %d.\n", n2, n);
402 * SWAP_PAGER_ALLOC() - allocate a new OBJT_SWAP VM object and instantiate
403 * its metadata structures.
405 * This routine is called from the mmap and fork code to create a new
406 * OBJT_SWAP object. We do this by creating an OBJT_DEFAULT object
407 * and then converting it with swp_pager_meta_convert().
409 * We only support unnamed objects.
414 swap_pager_alloc(void *handle, off_t size, vm_prot_t prot, off_t offset)
418 KKASSERT(handle == NULL);
419 lwkt_gettoken(&vm_token);
420 object = vm_object_allocate(OBJT_DEFAULT,
421 OFF_TO_IDX(offset + PAGE_MASK + size));
422 swp_pager_meta_convert(object);
423 lwkt_reltoken(&vm_token);
429 * SWAP_PAGER_DEALLOC() - remove swap metadata from object
431 * The swap backing for the object is destroyed. The code is
432 * designed such that we can reinstantiate it later, but this
433 * routine is typically called only when the entire object is
434 * about to be destroyed.
436 * The object must be locked or unreferenceable.
437 * No other requirements.
440 swap_pager_dealloc(vm_object_t object)
442 lwkt_gettoken(&vm_token);
443 vm_object_pip_wait(object, "swpdea");
446 * Free all remaining metadata. We only bother to free it from
447 * the swap meta data. We do not attempt to free swapblk's still
448 * associated with vm_page_t's for this object. We do not care
449 * if paging is still in progress on some objects.
452 swp_pager_meta_free_all(object);
454 lwkt_reltoken(&vm_token);
457 /************************************************************************
458 * SWAP PAGER BITMAP ROUTINES *
459 ************************************************************************/
462 * SWP_PAGER_GETSWAPSPACE() - allocate raw swap space
464 * Allocate swap for the requested number of pages. The starting
465 * swap block number (a page index) is returned or SWAPBLK_NONE
466 * if the allocation failed.
468 * Also has the side effect of advising that somebody made a mistake
469 * when they configured swap and didn't configure enough.
471 * The caller must hold vm_token.
472 * This routine may not block.
474 * NOTE: vm_token must be held to avoid races with bitmap frees from
475 * vm_page_remove() via swap_pager_page_removed().
477 static __inline swblk_t
478 swp_pager_getswapspace(vm_object_t object, int npages)
482 ASSERT_LWKT_TOKEN_HELD(&vm_token);
484 if ((blk = blist_alloc(swapblist, npages)) == SWAPBLK_NONE) {
485 if (swap_pager_full != 2) {
486 kprintf("swap_pager_getswapspace: failed\n");
488 swap_pager_almost_full = 1;
491 vm_swap_size -= npages;
492 if (object->type == OBJT_SWAP)
493 vm_swap_anon_use += npages;
495 vm_swap_cache_use += npages;
502 * SWP_PAGER_FREESWAPSPACE() - free raw swap space
504 * This routine returns the specified swap blocks back to the bitmap.
506 * Note: This routine may not block (it could in the old swap code),
507 * and through the use of the new blist routines it does not block.
509 * We must be called at splvm() to avoid races with bitmap frees from
510 * vm_page_remove() aka swap_pager_page_removed().
512 * The caller must hold vm_token.
513 * This routine may not block.
517 swp_pager_freeswapspace(vm_object_t object, swblk_t blk, int npages)
519 blist_free(swapblist, blk, npages);
520 vm_swap_size += npages;
521 if (object->type == OBJT_SWAP)
522 vm_swap_anon_use -= npages;
524 vm_swap_cache_use -= npages;
529 * SWAP_PAGER_FREESPACE() - frees swap blocks associated with a page
530 * range within an object.
532 * This is a globally accessible routine.
534 * This routine removes swapblk assignments from swap metadata.
536 * The external callers of this routine typically have already destroyed
537 * or renamed vm_page_t's associated with this range in the object so
543 swap_pager_freespace(vm_object_t object, vm_pindex_t start, vm_pindex_t size)
546 lwkt_gettoken(&vm_token);
547 swp_pager_meta_free(object, start, size);
548 lwkt_reltoken(&vm_token);
556 swap_pager_freespace_all(vm_object_t object)
559 lwkt_gettoken(&vm_token);
560 swp_pager_meta_free_all(object);
561 lwkt_reltoken(&vm_token);
566 * This function conditionally frees swap cache swap starting at
567 * (*basei) in the object. (count) swap blocks will be nominally freed.
568 * The actual number of blocks freed can be more or less than the
571 * This function nominally returns the number of blocks freed. However,
572 * the actual number of blocks freed may be less then the returned value.
573 * If the function is unable to exhaust the object or if it is able to
574 * free (approximately) the requested number of blocks it returns
577 * If we exhaust the object we will return a value n <= count.
579 * The caller must hold vm_token.
581 static int swap_pager_condfree_callback(struct swblock *swap, void *data);
584 swap_pager_condfree(vm_object_t object, vm_pindex_t *basei, int count)
586 struct swfreeinfo info;
588 ASSERT_LWKT_TOKEN_HELD(&vm_token);
590 info.object = object;
591 info.basei = *basei; /* skip up to this page index */
592 info.begi = count; /* max swap pages to destroy */
593 info.endi = count * 8; /* max swblocks to scan */
595 swblock_rb_tree_RB_SCAN(&object->swblock_root, rb_swblock_condcmp,
596 swap_pager_condfree_callback, &info);
598 if (info.endi < 0 && info.begi <= count)
599 info.begi = count + 1;
600 return(count - (int)info.begi);
604 * The idea is to free whole meta-block to avoid fragmenting
605 * the swap space or disk I/O. We only do this if NO VM pages
608 * We do not have to deal with clearing PG_SWAPPED in related VM
609 * pages because there are no related VM pages.
611 * The caller must hold vm_token.
614 swap_pager_condfree_callback(struct swblock *swap, void *data)
616 struct swfreeinfo *info = data;
617 vm_object_t object = info->object;
620 for (i = 0; i < SWAP_META_PAGES; ++i) {
621 if (vm_page_lookup(object, swap->swb_index + i))
624 info->basei = swap->swb_index + SWAP_META_PAGES;
625 if (i == SWAP_META_PAGES) {
626 info->begi -= swap->swb_count;
627 swap_pager_freespace(object, swap->swb_index, SWAP_META_PAGES);
630 if ((int)info->begi < 0 || (int)info->endi < 0)
636 * Called by vm_page_alloc() when a new VM page is inserted
637 * into a VM object. Checks whether swap has been assigned to
638 * the page and sets PG_SWAPPED as necessary.
643 swap_pager_page_inserted(vm_page_t m)
645 if (m->object->swblock_count) {
647 lwkt_gettoken(&vm_token);
648 if (swp_pager_meta_ctl(m->object, m->pindex, 0) != SWAPBLK_NONE)
649 vm_page_flag_set(m, PG_SWAPPED);
650 lwkt_reltoken(&vm_token);
656 * SWAP_PAGER_RESERVE() - reserve swap blocks in object
658 * Assigns swap blocks to the specified range within the object. The
659 * swap blocks are not zerod. Any previous swap assignment is destroyed.
661 * Returns 0 on success, -1 on failure.
663 * The caller is responsible for avoiding races in the specified range.
664 * No other requirements.
667 swap_pager_reserve(vm_object_t object, vm_pindex_t start, vm_size_t size)
670 swblk_t blk = SWAPBLK_NONE;
671 vm_pindex_t beg = start; /* save start index */
674 lwkt_gettoken(&vm_token);
678 while ((blk = swp_pager_getswapspace(object, n)) ==
683 swp_pager_meta_free(object, beg,
685 lwkt_reltoken(&vm_token);
691 swp_pager_meta_build(object, start, blk);
697 swp_pager_meta_free(object, start, n);
698 lwkt_reltoken(&vm_token);
704 * SWAP_PAGER_COPY() - copy blocks from source pager to destination pager
705 * and destroy the source.
707 * Copy any valid swapblks from the source to the destination. In
708 * cases where both the source and destination have a valid swapblk,
709 * we keep the destination's.
711 * This routine is allowed to block. It may block allocating metadata
712 * indirectly through swp_pager_meta_build() or if paging is still in
713 * progress on the source.
715 * This routine can be called at any spl
717 * XXX vm_page_collapse() kinda expects us not to block because we
718 * supposedly do not need to allocate memory, but for the moment we
719 * *may* have to get a little memory from the zone allocator, but
720 * it is taken from the interrupt memory. We should be ok.
722 * The source object contains no vm_page_t's (which is just as well)
724 * The source object is of type OBJT_SWAP.
726 * The source and destination objects must be locked or
727 * inaccessible (XXX are they ?)
729 * The caller must hold vm_token.
732 swap_pager_copy(vm_object_t srcobject, vm_object_t dstobject,
733 vm_pindex_t base_index, int destroysource)
737 ASSERT_LWKT_TOKEN_HELD(&vm_token);
741 * transfer source to destination.
743 for (i = 0; i < dstobject->size; ++i) {
747 * Locate (without changing) the swapblk on the destination,
748 * unless it is invalid in which case free it silently, or
749 * if the destination is a resident page, in which case the
750 * source is thrown away.
752 dstaddr = swp_pager_meta_ctl(dstobject, i, 0);
754 if (dstaddr == SWAPBLK_NONE) {
756 * Destination has no swapblk and is not resident,
761 srcaddr = swp_pager_meta_ctl(srcobject,
762 base_index + i, SWM_POP);
764 if (srcaddr != SWAPBLK_NONE)
765 swp_pager_meta_build(dstobject, i, srcaddr);
768 * Destination has valid swapblk or it is represented
769 * by a resident page. We destroy the sourceblock.
771 swp_pager_meta_ctl(srcobject, base_index + i, SWM_FREE);
776 * Free left over swap blocks in source.
778 * We have to revert the type to OBJT_DEFAULT so we do not accidently
779 * double-remove the object from the swap queues.
783 * Reverting the type is not necessary, the caller is going
784 * to destroy srcobject directly, but I'm doing it here
785 * for consistency since we've removed the object from its
788 swp_pager_meta_free_all(srcobject);
789 if (srcobject->type == OBJT_SWAP)
790 srcobject->type = OBJT_DEFAULT;
796 * SWAP_PAGER_HASPAGE() - determine if we have good backing store for
797 * the requested page.
799 * We determine whether good backing store exists for the requested
800 * page and return TRUE if it does, FALSE if it doesn't.
802 * If TRUE, we also try to determine how much valid, contiguous backing
803 * store exists before and after the requested page within a reasonable
804 * distance. We do not try to restrict it to the swap device stripe
805 * (that is handled in getpages/putpages). It probably isn't worth
811 swap_pager_haspage(vm_object_t object, vm_pindex_t pindex)
816 * do we have good backing store at the requested index ?
820 lwkt_gettoken(&vm_token);
821 blk0 = swp_pager_meta_ctl(object, pindex, 0);
823 if (blk0 == SWAPBLK_NONE) {
824 lwkt_reltoken(&vm_token);
828 lwkt_reltoken(&vm_token);
834 * SWAP_PAGER_PAGE_UNSWAPPED() - remove swap backing store related to page
836 * This removes any associated swap backing store, whether valid or
837 * not, from the page. This operates on any VM object, not just OBJT_SWAP
840 * This routine is typically called when a page is made dirty, at
841 * which point any associated swap can be freed. MADV_FREE also
842 * calls us in a special-case situation
844 * NOTE!!! If the page is clean and the swap was valid, the caller
845 * should make the page dirty before calling this routine. This routine
846 * does NOT change the m->dirty status of the page. Also: MADV_FREE
849 * The page must be busied or soft-busied.
850 * The caller must hold vm_token if the caller does not wish to block here.
851 * No other requirements.
854 swap_pager_unswapped(vm_page_t m)
856 if (m->flags & PG_SWAPPED) {
858 lwkt_gettoken(&vm_token);
859 KKASSERT(m->flags & PG_SWAPPED);
860 swp_pager_meta_ctl(m->object, m->pindex, SWM_FREE);
861 vm_page_flag_clear(m, PG_SWAPPED);
862 lwkt_reltoken(&vm_token);
868 * SWAP_PAGER_STRATEGY() - read, write, free blocks
870 * This implements a VM OBJECT strategy function using swap backing store.
871 * This can operate on any VM OBJECT type, not necessarily just OBJT_SWAP
874 * This is intended to be a cacheless interface (i.e. caching occurs at
875 * higher levels), and is also used as a swap-based SSD cache for vnode
876 * and device objects.
878 * All I/O goes directly to and from the swap device.
880 * We currently attempt to run I/O synchronously or asynchronously as
881 * the caller requests. This isn't perfect because we loose error
882 * sequencing when we run multiple ops in parallel to satisfy a request.
883 * But this is swap, so we let it all hang out.
888 swap_pager_strategy(vm_object_t object, struct bio *bio)
890 struct buf *bp = bio->bio_buf;
893 vm_pindex_t biox_blkno = 0;
898 struct bio_track *track;
901 * tracking for swapdev vnode I/Os
903 if (bp->b_cmd == BUF_CMD_READ)
904 track = &swapdev_vp->v_track_read;
906 track = &swapdev_vp->v_track_write;
908 if (bp->b_bcount & PAGE_MASK) {
909 bp->b_error = EINVAL;
910 bp->b_flags |= B_ERROR | B_INVAL;
912 kprintf("swap_pager_strategy: bp %p offset %lld size %d, "
913 "not page bounded\n",
914 bp, (long long)bio->bio_offset, (int)bp->b_bcount);
919 * Clear error indication, initialize page index, count, data pointer.
922 bp->b_flags &= ~B_ERROR;
923 bp->b_resid = bp->b_bcount;
925 start = (vm_pindex_t)(bio->bio_offset >> PAGE_SHIFT);
926 count = howmany(bp->b_bcount, PAGE_SIZE);
930 * Deal with BUF_CMD_FREEBLKS
932 if (bp->b_cmd == BUF_CMD_FREEBLKS) {
934 * FREE PAGE(s) - destroy underlying swap that is no longer
938 lwkt_gettoken(&vm_token);
939 swp_pager_meta_free(object, start, count);
940 lwkt_reltoken(&vm_token);
948 * We need to be able to create a new cluster of I/O's. We cannot
949 * use the caller fields of the passed bio so push a new one.
951 * Because nbio is just a placeholder for the cluster links,
952 * we can biodone() the original bio instead of nbio to make
953 * things a bit more efficient.
955 nbio = push_bio(bio);
956 nbio->bio_offset = bio->bio_offset;
957 nbio->bio_caller_info1.cluster_head = NULL;
958 nbio->bio_caller_info2.cluster_tail = NULL;
964 * Execute read or write
967 lwkt_gettoken(&vm_token);
972 * Obtain block. If block not found and writing, allocate a
973 * new block and build it into the object.
975 blk = swp_pager_meta_ctl(object, start, 0);
976 if ((blk == SWAPBLK_NONE) && bp->b_cmd != BUF_CMD_READ) {
977 blk = swp_pager_getswapspace(object, 1);
978 if (blk == SWAPBLK_NONE) {
979 bp->b_error = ENOMEM;
980 bp->b_flags |= B_ERROR;
983 swp_pager_meta_build(object, start, blk);
987 * Do we have to flush our current collection? Yes if:
989 * - no swap block at this index
990 * - swap block is not contiguous
991 * - we cross a physical disk boundry in the
995 biox && (biox_blkno + btoc(bufx->b_bcount) != blk ||
996 ((biox_blkno ^ blk) & dmmax_mask)
999 if (bp->b_cmd == BUF_CMD_READ) {
1000 ++mycpu->gd_cnt.v_swapin;
1001 mycpu->gd_cnt.v_swappgsin += btoc(bufx->b_bcount);
1003 ++mycpu->gd_cnt.v_swapout;
1004 mycpu->gd_cnt.v_swappgsout += btoc(bufx->b_bcount);
1005 bufx->b_dirtyend = bufx->b_bcount;
1009 * Finished with this buf.
1011 KKASSERT(bufx->b_bcount != 0);
1012 if (bufx->b_cmd != BUF_CMD_READ)
1013 bufx->b_dirtyend = bufx->b_bcount;
1019 * Add new swapblk to biox, instantiating biox if necessary.
1020 * Zero-fill reads are able to take a shortcut.
1022 if (blk == SWAPBLK_NONE) {
1024 * We can only get here if we are reading. Since
1025 * we are at splvm() we can safely modify b_resid,
1026 * even if chain ops are in progress.
1028 bzero(data, PAGE_SIZE);
1029 bp->b_resid -= PAGE_SIZE;
1032 /* XXX chain count > 4, wait to <= 4 */
1034 bufx = getpbuf(NULL);
1035 biox = &bufx->b_bio1;
1036 cluster_append(nbio, bufx);
1037 bufx->b_flags |= (bufx->b_flags & B_ORDERED);
1038 bufx->b_cmd = bp->b_cmd;
1039 biox->bio_done = swap_chain_iodone;
1040 biox->bio_offset = (off_t)blk << PAGE_SHIFT;
1041 biox->bio_caller_info1.cluster_parent = nbio;
1044 bufx->b_data = data;
1046 bufx->b_bcount += PAGE_SIZE;
1052 lwkt_reltoken(&vm_token);
1056 * Flush out last buffer
1059 if (bufx->b_cmd == BUF_CMD_READ) {
1060 ++mycpu->gd_cnt.v_swapin;
1061 mycpu->gd_cnt.v_swappgsin += btoc(bufx->b_bcount);
1063 ++mycpu->gd_cnt.v_swapout;
1064 mycpu->gd_cnt.v_swappgsout += btoc(bufx->b_bcount);
1065 bufx->b_dirtyend = bufx->b_bcount;
1067 KKASSERT(bufx->b_bcount);
1068 if (bufx->b_cmd != BUF_CMD_READ)
1069 bufx->b_dirtyend = bufx->b_bcount;
1070 /* biox, bufx = NULL */
1074 * Now initiate all the I/O. Be careful looping on our chain as
1075 * I/O's may complete while we are still initiating them.
1077 * If the request is a 100% sparse read no bios will be present
1078 * and we just biodone() the buffer.
1080 nbio->bio_caller_info2.cluster_tail = NULL;
1081 bufx = nbio->bio_caller_info1.cluster_head;
1085 biox = &bufx->b_bio1;
1087 bufx = bufx->b_cluster_next;
1088 vn_strategy(swapdev_vp, biox);
1095 * Completion of the cluster will also call biodone_chain(nbio).
1096 * We never call biodone(nbio) so we don't have to worry about
1097 * setting up a bio_done callback. It's handled in the sub-IO.
1108 swap_chain_iodone(struct bio *biox)
1111 struct buf *bufx; /* chained sub-buffer */
1112 struct bio *nbio; /* parent nbio with chain glue */
1113 struct buf *bp; /* original bp associated with nbio */
1116 bufx = biox->bio_buf;
1117 nbio = biox->bio_caller_info1.cluster_parent;
1121 * Update the original buffer
1123 KKASSERT(bp != NULL);
1124 if (bufx->b_flags & B_ERROR) {
1125 atomic_set_int(&bufx->b_flags, B_ERROR);
1126 bp->b_error = bufx->b_error;
1127 } else if (bufx->b_resid != 0) {
1128 atomic_set_int(&bufx->b_flags, B_ERROR);
1129 bp->b_error = EINVAL;
1131 atomic_subtract_int(&bp->b_resid, bufx->b_bcount);
1135 * Remove us from the chain.
1137 spin_lock_wr(&bp->b_lock.lk_spinlock);
1138 nextp = &nbio->bio_caller_info1.cluster_head;
1139 while (*nextp != bufx) {
1140 KKASSERT(*nextp != NULL);
1141 nextp = &(*nextp)->b_cluster_next;
1143 *nextp = bufx->b_cluster_next;
1144 chain_empty = (nbio->bio_caller_info1.cluster_head == NULL);
1145 spin_unlock_wr(&bp->b_lock.lk_spinlock);
1148 * Clean up bufx. If the chain is now empty we finish out
1149 * the parent. Note that we may be racing other completions
1150 * so we must use the chain_empty status from above.
1153 if (bp->b_resid != 0 && !(bp->b_flags & B_ERROR)) {
1154 atomic_set_int(&bp->b_flags, B_ERROR);
1155 bp->b_error = EINVAL;
1157 biodone_chain(nbio);
1159 relpbuf(bufx, NULL);
1163 * SWAP_PAGER_GETPAGES() - bring page in from swap
1165 * The requested page may have to be brought in from swap. Calculate the
1166 * swap block and bring in additional pages if possible. All pages must
1167 * have contiguous swap block assignments and reside in the same object.
1169 * The caller has a single vm_object_pip_add() reference prior to
1170 * calling us and we should return with the same.
1172 * The caller has BUSY'd the page. We should return with (*mpp) left busy,
1173 * and any additinal pages unbusied.
1175 * If the caller encounters a PG_RAM page it will pass it to us even though
1176 * it may be valid and dirty. We cannot overwrite the page in this case!
1177 * The case is used to allow us to issue pure read-aheads.
1179 * NOTE! XXX This code does not entirely pipeline yet due to the fact that
1180 * the PG_RAM page is validated at the same time as mreq. What we
1181 * really need to do is issue a separate read-ahead pbuf.
1186 swap_pager_getpage(vm_object_t object, vm_page_t *mpp, int seqaccess)
1197 vm_page_t marray[XIO_INTERNAL_PAGES];
1201 if (mreq->object != object) {
1202 panic("swap_pager_getpages: object mismatch %p/%p",
1209 * We don't want to overwrite a fully valid page as it might be
1210 * dirty. This case can occur when e.g. vm_fault hits a perfectly
1211 * valid page with PG_RAM set.
1213 * In this case we see if the next page is a suitable page-in
1214 * candidate and if it is we issue read-ahead. PG_RAM will be
1215 * set on the last page of the read-ahead to continue the pipeline.
1217 if (mreq->valid == VM_PAGE_BITS_ALL) {
1218 if (swap_burst_read == 0 || mreq->pindex + 1 >= object->size)
1219 return(VM_PAGER_OK);
1221 lwkt_gettoken(&vm_token);
1222 blk = swp_pager_meta_ctl(object, mreq->pindex + 1, 0);
1223 if (blk == SWAPBLK_NONE) {
1224 lwkt_reltoken(&vm_token);
1226 return(VM_PAGER_OK);
1228 m = vm_page_lookup(object, mreq->pindex + 1);
1230 m = vm_page_alloc(object, mreq->pindex + 1,
1233 lwkt_reltoken(&vm_token);
1235 return(VM_PAGER_OK);
1238 if ((m->flags & PG_BUSY) || m->busy || m->valid) {
1239 lwkt_reltoken(&vm_token);
1241 return(VM_PAGER_OK);
1243 vm_page_unqueue_nowakeup(m);
1248 lwkt_reltoken(&vm_token);
1255 * Try to block-read contiguous pages from swap if sequential,
1256 * otherwise just read one page. Contiguous pages from swap must
1257 * reside within a single device stripe because the I/O cannot be
1258 * broken up across multiple stripes.
1260 * Note that blk and iblk can be SWAPBLK_NONE but the loop is
1261 * set up such that the case(s) are handled implicitly.
1264 lwkt_gettoken(&vm_token);
1265 blk = swp_pager_meta_ctl(mreq->object, mreq->pindex, 0);
1268 for (i = 1; swap_burst_read &&
1269 i < XIO_INTERNAL_PAGES &&
1270 mreq->pindex + i < object->size; ++i) {
1273 iblk = swp_pager_meta_ctl(object, mreq->pindex + i, 0);
1274 if (iblk != blk + i)
1276 if ((blk ^ iblk) & dmmax_mask)
1278 m = vm_page_lookup(object, mreq->pindex + i);
1280 m = vm_page_alloc(object, mreq->pindex + i,
1285 if ((m->flags & PG_BUSY) || m->busy || m->valid)
1287 vm_page_unqueue_nowakeup(m);
1293 vm_page_flag_set(marray[i - 1], PG_RAM);
1295 lwkt_reltoken(&vm_token);
1299 * If mreq is the requested page and we have nothing to do return
1300 * VM_PAGER_FAIL. If raonly is set mreq is just another read-ahead
1301 * page and must be cleaned up.
1303 if (blk == SWAPBLK_NONE) {
1306 vnode_pager_freepage(mreq);
1307 return(VM_PAGER_OK);
1309 return(VM_PAGER_FAIL);
1314 * map our page(s) into kva for input
1316 bp = getpbuf(&nsw_rcount);
1318 kva = (vm_offset_t) bp->b_kvabase;
1319 bcopy(marray, bp->b_xio.xio_pages, i * sizeof(vm_page_t));
1320 pmap_qenter(kva, bp->b_xio.xio_pages, i);
1322 bp->b_data = (caddr_t)kva;
1323 bp->b_bcount = PAGE_SIZE * i;
1324 bp->b_xio.xio_npages = i;
1325 bio->bio_done = swp_pager_async_iodone;
1326 bio->bio_offset = (off_t)blk << PAGE_SHIFT;
1327 bio->bio_caller_info1.index = SWBIO_READ;
1330 * Set index. If raonly set the index beyond the array so all
1331 * the pages are treated the same, otherwise the original mreq is
1335 bio->bio_driver_info = (void *)(intptr_t)i;
1337 bio->bio_driver_info = (void *)(intptr_t)0;
1339 for (j = 0; j < i; ++j)
1340 vm_page_flag_set(bp->b_xio.xio_pages[j], PG_SWAPINPROG);
1342 mycpu->gd_cnt.v_swapin++;
1343 mycpu->gd_cnt.v_swappgsin += bp->b_xio.xio_npages;
1346 * We still hold the lock on mreq, and our automatic completion routine
1347 * does not remove it.
1349 vm_object_pip_add(object, bp->b_xio.xio_npages);
1352 * perform the I/O. NOTE!!! bp cannot be considered valid after
1353 * this point because we automatically release it on completion.
1354 * Instead, we look at the one page we are interested in which we
1355 * still hold a lock on even through the I/O completion.
1357 * The other pages in our m[] array are also released on completion,
1358 * so we cannot assume they are valid anymore either.
1360 bp->b_cmd = BUF_CMD_READ;
1362 vn_strategy(swapdev_vp, bio);
1365 * Wait for the page we want to complete. PG_SWAPINPROG is always
1366 * cleared on completion. If an I/O error occurs, SWAPBLK_NONE
1367 * is set in the meta-data.
1369 * If this is a read-ahead only we return immediately without
1373 return(VM_PAGER_OK);
1376 * Read-ahead includes originally requested page case.
1379 lwkt_gettoken(&vm_token);
1380 while ((mreq->flags & PG_SWAPINPROG) != 0) {
1381 vm_page_flag_set(mreq, PG_WANTED | PG_REFERENCED);
1382 mycpu->gd_cnt.v_intrans++;
1383 if (tsleep(mreq, 0, "swread", hz*20)) {
1385 "swap_pager: indefinite wait buffer: "
1386 " offset: %lld, size: %ld\n",
1387 (long long)bio->bio_offset,
1392 lwkt_reltoken(&vm_token);
1396 * mreq is left bussied after completion, but all the other pages
1397 * are freed. If we had an unrecoverable read error the page will
1400 if (mreq->valid != VM_PAGE_BITS_ALL)
1401 return(VM_PAGER_ERROR);
1403 return(VM_PAGER_OK);
1406 * A final note: in a low swap situation, we cannot deallocate swap
1407 * and mark a page dirty here because the caller is likely to mark
1408 * the page clean when we return, causing the page to possibly revert
1409 * to all-zero's later.
1414 * swap_pager_putpages:
1416 * Assign swap (if necessary) and initiate I/O on the specified pages.
1418 * We support both OBJT_DEFAULT and OBJT_SWAP objects. DEFAULT objects
1419 * are automatically converted to SWAP objects.
1421 * In a low memory situation we may block in vn_strategy(), but the new
1422 * vm_page reservation system coupled with properly written VFS devices
1423 * should ensure that no low-memory deadlock occurs. This is an area
1426 * The parent has N vm_object_pip_add() references prior to
1427 * calling us and will remove references for rtvals[] that are
1428 * not set to VM_PAGER_PEND. We need to remove the rest on I/O
1431 * The parent has soft-busy'd the pages it passes us and will unbusy
1432 * those whos rtvals[] entry is not set to VM_PAGER_PEND on return.
1433 * We need to unbusy the rest on I/O completion.
1438 swap_pager_putpages(vm_object_t object, vm_page_t *m, int count,
1439 boolean_t sync, int *rtvals)
1444 if (count && m[0]->object != object) {
1445 panic("swap_pager_getpages: object mismatch %p/%p",
1454 * Turn object into OBJT_SWAP
1455 * check for bogus sysops
1456 * force sync if not pageout process
1458 if (object->type == OBJT_DEFAULT) {
1459 lwkt_gettoken(&vm_token);
1460 if (object->type == OBJT_DEFAULT)
1461 swp_pager_meta_convert(object);
1462 lwkt_reltoken(&vm_token);
1465 if (curthread != pagethread)
1471 * Update nsw parameters from swap_async_max sysctl values.
1472 * Do not let the sysop crash the machine with bogus numbers.
1475 if (swap_async_max != nsw_wcount_async_max) {
1481 if ((n = swap_async_max) > nswbuf / 2)
1488 * Adjust difference ( if possible ). If the current async
1489 * count is too low, we may not be able to make the adjustment
1493 lwkt_gettoken(&vm_token);
1494 n -= nsw_wcount_async_max;
1495 if (nsw_wcount_async + n >= 0) {
1496 nsw_wcount_async += n;
1497 nsw_wcount_async_max += n;
1498 wakeup(&nsw_wcount_async);
1500 lwkt_reltoken(&vm_token);
1507 * Assign swap blocks and issue I/O. We reallocate swap on the fly.
1508 * The page is left dirty until the pageout operation completes
1512 for (i = 0; i < count; i += n) {
1519 * Maximum I/O size is limited by a number of factors.
1522 n = min(BLIST_MAX_ALLOC, count - i);
1523 n = min(n, nsw_cluster_max);
1526 lwkt_gettoken(&vm_token);
1529 * Get biggest block of swap we can. If we fail, fall
1530 * back and try to allocate a smaller block. Don't go
1531 * overboard trying to allocate space if it would overly
1535 (blk = swp_pager_getswapspace(object, n)) == SWAPBLK_NONE &&
1540 if (blk == SWAPBLK_NONE) {
1541 for (j = 0; j < n; ++j)
1542 rtvals[i+j] = VM_PAGER_FAIL;
1543 lwkt_reltoken(&vm_token);
1549 * The I/O we are constructing cannot cross a physical
1550 * disk boundry in the swap stripe. Note: we are still
1553 if ((blk ^ (blk + n)) & dmmax_mask) {
1554 j = ((blk + dmmax) & dmmax_mask) - blk;
1555 swp_pager_freeswapspace(object, blk + j, n - j);
1560 * All I/O parameters have been satisfied, build the I/O
1561 * request and assign the swap space.
1564 bp = getpbuf(&nsw_wcount_sync);
1566 bp = getpbuf(&nsw_wcount_async);
1569 pmap_qenter((vm_offset_t)bp->b_data, &m[i], n);
1571 bp->b_bcount = PAGE_SIZE * n;
1572 bio->bio_offset = (off_t)blk << PAGE_SHIFT;
1574 for (j = 0; j < n; ++j) {
1575 vm_page_t mreq = m[i+j];
1577 swp_pager_meta_build(mreq->object, mreq->pindex,
1579 if (object->type == OBJT_SWAP)
1580 vm_page_dirty(mreq);
1581 rtvals[i+j] = VM_PAGER_OK;
1583 vm_page_flag_set(mreq, PG_SWAPINPROG);
1584 bp->b_xio.xio_pages[j] = mreq;
1586 bp->b_xio.xio_npages = n;
1588 mycpu->gd_cnt.v_swapout++;
1589 mycpu->gd_cnt.v_swappgsout += bp->b_xio.xio_npages;
1591 lwkt_reltoken(&vm_token);
1594 bp->b_dirtyoff = 0; /* req'd for NFS */
1595 bp->b_dirtyend = bp->b_bcount; /* req'd for NFS */
1596 bp->b_cmd = BUF_CMD_WRITE;
1597 bio->bio_caller_info1.index = SWBIO_WRITE;
1602 if (sync == FALSE) {
1603 bio->bio_done = swp_pager_async_iodone;
1605 vn_strategy(swapdev_vp, bio);
1607 for (j = 0; j < n; ++j)
1608 rtvals[i+j] = VM_PAGER_PEND;
1613 * Issue synchrnously.
1615 * Wait for the sync I/O to complete, then update rtvals.
1616 * We just set the rtvals[] to VM_PAGER_PEND so we can call
1617 * our async completion routine at the end, thus avoiding a
1620 bio->bio_caller_info1.index |= SWBIO_SYNC;
1621 bio->bio_done = biodone_sync;
1622 bio->bio_flags |= BIO_SYNC;
1623 vn_strategy(swapdev_vp, bio);
1624 biowait(bio, "swwrt");
1626 for (j = 0; j < n; ++j)
1627 rtvals[i+j] = VM_PAGER_PEND;
1630 * Now that we are through with the bp, we can call the
1631 * normal async completion, which frees everything up.
1633 swp_pager_async_iodone(bio);
1641 swap_pager_newswap(void)
1647 * swp_pager_async_iodone:
1649 * Completion routine for asynchronous reads and writes from/to swap.
1650 * Also called manually by synchronous code to finish up a bp.
1652 * For READ operations, the pages are PG_BUSY'd. For WRITE operations,
1653 * the pages are vm_page_t->busy'd. For READ operations, we PG_BUSY
1654 * unbusy all pages except the 'main' request page. For WRITE
1655 * operations, we vm_page_t->busy'd unbusy all pages ( we can do this
1656 * because we marked them all VM_PAGER_PEND on return from putpages ).
1658 * This routine may not block.
1663 swp_pager_async_iodone(struct bio *bio)
1665 struct buf *bp = bio->bio_buf;
1666 vm_object_t object = NULL;
1673 if (bp->b_flags & B_ERROR) {
1675 "swap_pager: I/O error - %s failed; offset %lld,"
1676 "size %ld, error %d\n",
1677 ((bio->bio_caller_info1.index & SWBIO_READ) ?
1678 "pagein" : "pageout"),
1679 (long long)bio->bio_offset,
1686 * set object, raise to splvm().
1688 if (bp->b_xio.xio_npages)
1689 object = bp->b_xio.xio_pages[0]->object;
1691 lwkt_gettoken(&vm_token);
1694 * remove the mapping for kernel virtual
1696 pmap_qremove((vm_offset_t)bp->b_data, bp->b_xio.xio_npages);
1699 * cleanup pages. If an error occurs writing to swap, we are in
1700 * very serious trouble. If it happens to be a disk error, though,
1701 * we may be able to recover by reassigning the swap later on. So
1702 * in this case we remove the m->swapblk assignment for the page
1703 * but do not free it in the rlist. The errornous block(s) are thus
1704 * never reallocated as swap. Redirty the page and continue.
1706 for (i = 0; i < bp->b_xio.xio_npages; ++i) {
1707 vm_page_t m = bp->b_xio.xio_pages[i];
1709 if (bp->b_flags & B_ERROR) {
1711 * If an error occurs I'd love to throw the swapblk
1712 * away without freeing it back to swapspace, so it
1713 * can never be used again. But I can't from an
1717 if (bio->bio_caller_info1.index & SWBIO_READ) {
1719 * When reading, reqpage needs to stay
1720 * locked for the parent, but all other
1721 * pages can be freed. We still want to
1722 * wakeup the parent waiting on the page,
1723 * though. ( also: pg_reqpage can be -1 and
1724 * not match anything ).
1726 * We have to wake specifically requested pages
1727 * up too because we cleared PG_SWAPINPROG and
1728 * someone may be waiting for that.
1730 * NOTE: for reads, m->dirty will probably
1731 * be overridden by the original caller of
1732 * getpages so don't play cute tricks here.
1734 * NOTE: We can't actually free the page from
1735 * here, because this is an interrupt. It
1736 * is not legal to mess with object->memq
1737 * from an interrupt. Deactivate the page
1742 vm_page_flag_clear(m, PG_ZERO);
1743 vm_page_flag_clear(m, PG_SWAPINPROG);
1746 * bio_driver_info holds the requested page
1749 if (i != (int)(intptr_t)bio->bio_driver_info) {
1750 vm_page_deactivate(m);
1756 * If i == bp->b_pager.pg_reqpage, do not wake
1757 * the page up. The caller needs to.
1761 * If a write error occurs remove the swap
1762 * assignment (note that PG_SWAPPED may or
1763 * may not be set depending on prior activity).
1765 * Re-dirty OBJT_SWAP pages as there is no
1766 * other backing store, we can't throw the
1769 * Non-OBJT_SWAP pages (aka swapcache) must
1770 * not be dirtied since they may not have
1771 * been dirty in the first place, and they
1772 * do have backing store (the vnode).
1774 swp_pager_meta_ctl(m->object, m->pindex,
1776 vm_page_flag_clear(m, PG_SWAPPED);
1777 if (m->object->type == OBJT_SWAP) {
1779 vm_page_activate(m);
1781 vm_page_flag_clear(m, PG_SWAPINPROG);
1782 vm_page_io_finish(m);
1784 } else if (bio->bio_caller_info1.index & SWBIO_READ) {
1786 * NOTE: for reads, m->dirty will probably be
1787 * overridden by the original caller of getpages so
1788 * we cannot set them in order to free the underlying
1789 * swap in a low-swap situation. I don't think we'd
1790 * want to do that anyway, but it was an optimization
1791 * that existed in the old swapper for a time before
1792 * it got ripped out due to precisely this problem.
1794 * clear PG_ZERO in page.
1796 * If not the requested page then deactivate it.
1798 * Note that the requested page, reqpage, is left
1799 * busied, but we still have to wake it up. The
1800 * other pages are released (unbusied) by
1801 * vm_page_wakeup(). We do not set reqpage's
1802 * valid bits here, it is up to the caller.
1806 * NOTE: can't call pmap_clear_modify(m) from an
1807 * interrupt thread, the pmap code may have to map
1808 * non-kernel pmaps and currently asserts the case.
1810 /*pmap_clear_modify(m);*/
1811 m->valid = VM_PAGE_BITS_ALL;
1813 vm_page_flag_clear(m, PG_ZERO | PG_SWAPINPROG);
1814 vm_page_flag_set(m, PG_SWAPPED);
1817 * We have to wake specifically requested pages
1818 * up too because we cleared PG_SWAPINPROG and
1819 * could be waiting for it in getpages. However,
1820 * be sure to not unbusy getpages specifically
1821 * requested page - getpages expects it to be
1824 * bio_driver_info holds the requested page
1826 if (i != (int)(intptr_t)bio->bio_driver_info) {
1827 vm_page_deactivate(m);
1834 * Mark the page clean but do not mess with the
1835 * pmap-layer's modified state. That state should
1836 * also be clear since the caller protected the
1837 * page VM_PROT_READ, but allow the case.
1839 * We are in an interrupt, avoid pmap operations.
1841 * If we have a severe page deficit, deactivate the
1842 * page. Do not try to cache it (which would also
1843 * involve a pmap op), because the page might still
1846 * When using the swap to cache clean vnode pages
1847 * we do not mess with the page dirty bits.
1849 if (m->object->type == OBJT_SWAP)
1851 vm_page_flag_clear(m, PG_SWAPINPROG);
1852 vm_page_flag_set(m, PG_SWAPPED);
1853 vm_page_io_finish(m);
1854 if (vm_page_count_severe())
1855 vm_page_deactivate(m);
1857 if (!vm_page_count_severe() || !vm_page_try_to_cache(m))
1858 vm_page_protect(m, VM_PROT_READ);
1864 * adjust pip. NOTE: the original parent may still have its own
1865 * pip refs on the object.
1869 vm_object_pip_wakeupn(object, bp->b_xio.xio_npages);
1872 * Release the physical I/O buffer.
1874 * NOTE: Due to synchronous operations in the write case b_cmd may
1875 * already be set to BUF_CMD_DONE and BIO_SYNC may have already
1878 if (bio->bio_caller_info1.index & SWBIO_READ)
1879 nswptr = &nsw_rcount;
1880 else if (bio->bio_caller_info1.index & SWBIO_SYNC)
1881 nswptr = &nsw_wcount_sync;
1883 nswptr = &nsw_wcount_async;
1884 bp->b_cmd = BUF_CMD_DONE;
1885 relpbuf(bp, nswptr);
1886 lwkt_reltoken(&vm_token);
1890 /************************************************************************
1892 ************************************************************************
1894 * These routines manipulate the swap metadata stored in the
1895 * OBJT_SWAP object. All swp_*() routines must be called at
1896 * splvm() because swap can be freed up by the low level vm_page
1897 * code which might be called from interrupts beyond what splbio() covers.
1899 * Swap metadata is implemented with a global hash and not directly
1900 * linked into the object. Instead the object simply contains
1901 * appropriate tracking counters.
1905 * Lookup the swblock containing the specified swap block index.
1907 * The caller must hold vm_token.
1911 swp_pager_lookup(vm_object_t object, vm_pindex_t index)
1913 index &= ~SWAP_META_MASK;
1914 return (RB_LOOKUP(swblock_rb_tree, &object->swblock_root, index));
1918 * Remove a swblock from the RB tree.
1920 * The caller must hold vm_token.
1924 swp_pager_remove(vm_object_t object, struct swblock *swap)
1926 RB_REMOVE(swblock_rb_tree, &object->swblock_root, swap);
1930 * Convert default object to swap object if necessary
1932 * The caller must hold vm_token.
1935 swp_pager_meta_convert(vm_object_t object)
1937 if (object->type == OBJT_DEFAULT) {
1938 object->type = OBJT_SWAP;
1939 KKASSERT(object->swblock_count == 0);
1944 * SWP_PAGER_META_BUILD() - add swap block to swap meta data for object
1946 * We first convert the object to a swap object if it is a default
1947 * object. Vnode objects do not need to be converted.
1949 * The specified swapblk is added to the object's swap metadata. If
1950 * the swapblk is not valid, it is freed instead. Any previously
1951 * assigned swapblk is freed.
1953 * The caller must hold vm_token.
1956 swp_pager_meta_build(vm_object_t object, vm_pindex_t index, swblk_t swapblk)
1958 struct swblock *swap;
1959 struct swblock *oswap;
1961 KKASSERT(swapblk != SWAPBLK_NONE);
1964 * Convert object if necessary
1966 if (object->type == OBJT_DEFAULT)
1967 swp_pager_meta_convert(object);
1970 * Locate swblock. If not found create, but if we aren't adding
1971 * anything just return. If we run out of space in the map we wait
1972 * and, since the hash table may have changed, retry.
1975 swap = swp_pager_lookup(object, index);
1980 swap = zalloc(swap_zone);
1985 swap->swb_index = index & ~SWAP_META_MASK;
1986 swap->swb_count = 0;
1988 ++object->swblock_count;
1990 for (i = 0; i < SWAP_META_PAGES; ++i)
1991 swap->swb_pages[i] = SWAPBLK_NONE;
1992 oswap = RB_INSERT(swblock_rb_tree, &object->swblock_root, swap);
1993 KKASSERT(oswap == NULL);
1997 * Delete prior contents of metadata
2000 index &= SWAP_META_MASK;
2002 if (swap->swb_pages[index] != SWAPBLK_NONE) {
2003 swp_pager_freeswapspace(object, swap->swb_pages[index], 1);
2008 * Enter block into metadata
2010 swap->swb_pages[index] = swapblk;
2011 if (swapblk != SWAPBLK_NONE)
2016 * SWP_PAGER_META_FREE() - free a range of blocks in the object's swap metadata
2018 * The requested range of blocks is freed, with any associated swap
2019 * returned to the swap bitmap.
2021 * This routine will free swap metadata structures as they are cleaned
2022 * out. This routine does *NOT* operate on swap metadata associated
2023 * with resident pages.
2025 * The caller must hold vm_token.
2027 static int swp_pager_meta_free_callback(struct swblock *swb, void *data);
2030 swp_pager_meta_free(vm_object_t object, vm_pindex_t index, vm_pindex_t count)
2032 struct swfreeinfo info;
2037 if (object->swblock_count == 0) {
2038 KKASSERT(RB_EMPTY(&object->swblock_root));
2045 * Setup for RB tree scan. Note that the pindex range can be huge
2046 * due to the 64 bit page index space so we cannot safely iterate.
2048 info.object = object;
2049 info.basei = index & ~SWAP_META_MASK;
2051 info.endi = index + count - 1;
2052 swblock_rb_tree_RB_SCAN(&object->swblock_root, rb_swblock_scancmp,
2053 swp_pager_meta_free_callback, &info);
2057 * The caller must hold vm_token.
2061 swp_pager_meta_free_callback(struct swblock *swap, void *data)
2063 struct swfreeinfo *info = data;
2064 vm_object_t object = info->object;
2069 * Figure out the range within the swblock. The wider scan may
2070 * return edge-case swap blocks when the start and/or end points
2071 * are in the middle of a block.
2073 if (swap->swb_index < info->begi)
2074 index = (int)info->begi & SWAP_META_MASK;
2078 if (swap->swb_index + SWAP_META_PAGES > info->endi)
2079 eindex = (int)info->endi & SWAP_META_MASK;
2081 eindex = SWAP_META_MASK;
2084 * Scan and free the blocks. The loop terminates early
2085 * if (swap) runs out of blocks and could be freed.
2087 while (index <= eindex) {
2088 swblk_t v = swap->swb_pages[index];
2090 if (v != SWAPBLK_NONE) {
2091 swp_pager_freeswapspace(object, v, 1);
2092 swap->swb_pages[index] = SWAPBLK_NONE;
2093 if (--swap->swb_count == 0) {
2094 swp_pager_remove(object, swap);
2095 zfree(swap_zone, swap);
2096 --object->swblock_count;
2102 /* swap may be invalid here due to zfree above */
2107 * SWP_PAGER_META_FREE_ALL() - destroy all swap metadata associated with object
2109 * This routine locates and destroys all swap metadata associated with
2112 * The caller must hold vm_token.
2115 swp_pager_meta_free_all(vm_object_t object)
2117 struct swblock *swap;
2120 while ((swap = RB_ROOT(&object->swblock_root)) != NULL) {
2121 swp_pager_remove(object, swap);
2122 for (i = 0; i < SWAP_META_PAGES; ++i) {
2123 swblk_t v = swap->swb_pages[i];
2124 if (v != SWAPBLK_NONE) {
2126 swp_pager_freeswapspace(object, v, 1);
2129 if (swap->swb_count != 0)
2130 panic("swap_pager_meta_free_all: swb_count != 0");
2131 zfree(swap_zone, swap);
2132 --object->swblock_count;
2134 KKASSERT(object->swblock_count == 0);
2138 * SWP_PAGER_METACTL() - misc control of swap and vm_page_t meta data.
2140 * This routine is capable of looking up, popping, or freeing
2141 * swapblk assignments in the swap meta data or in the vm_page_t.
2142 * The routine typically returns the swapblk being looked-up, or popped,
2143 * or SWAPBLK_NONE if the block was freed, or SWAPBLK_NONE if the block
2144 * was invalid. This routine will automatically free any invalid
2145 * meta-data swapblks.
2147 * It is not possible to store invalid swapblks in the swap meta data
2148 * (other then a literal 'SWAPBLK_NONE'), so we don't bother checking.
2150 * When acting on a busy resident page and paging is in progress, we
2151 * have to wait until paging is complete but otherwise can act on the
2154 * SWM_FREE remove and free swap block from metadata
2155 * SWM_POP remove from meta data but do not free.. pop it out
2157 * The caller must hold vm_token.
2160 swp_pager_meta_ctl(vm_object_t object, vm_pindex_t index, int flags)
2162 struct swblock *swap;
2165 if (object->swblock_count == 0)
2166 return(SWAPBLK_NONE);
2169 swap = swp_pager_lookup(object, index);
2172 index &= SWAP_META_MASK;
2173 r1 = swap->swb_pages[index];
2175 if (r1 != SWAPBLK_NONE) {
2176 if (flags & SWM_FREE) {
2177 swp_pager_freeswapspace(object, r1, 1);
2180 if (flags & (SWM_FREE|SWM_POP)) {
2181 swap->swb_pages[index] = SWAPBLK_NONE;
2182 if (--swap->swb_count == 0) {
2183 swp_pager_remove(object, swap);
2184 zfree(swap_zone, swap);
2185 --object->swblock_count;