2 * Copyright (c) 1998 Matthew Dillon,
3 * Copyright (c) 1994 John S. Dyson
4 * Copyright (c) 1990 University of Utah.
5 * Copyright (c) 1991, 1993
6 * The Regents of the University of California. All rights reserved.
8 * This code is derived from software contributed to Berkeley by
9 * the Systems Programming Group of the University of Utah Computer
12 * Redistribution and use in source and binary forms, with or without
13 * modification, are permitted provided that the following conditions
15 * 1. Redistributions of source code must retain the above copyright
16 * notice, this list of conditions and the following disclaimer.
17 * 2. Redistributions in binary form must reproduce the above copyright
18 * notice, this list of conditions and the following disclaimer in the
19 * documentation and/or other materials provided with the distribution.
20 * 3. All advertising materials mentioning features or use of this software
21 * must display the following acknowledgement:
22 * This product includes software developed by the University of
23 * California, Berkeley and its contributors.
24 * 4. Neither the name of the University nor the names of its contributors
25 * may be used to endorse or promote products derived from this software
26 * without specific prior written permission.
28 * THIS SOFTWARE IS PROVIDED BY THE REGENTS AND CONTRIBUTORS ``AS IS'' AND
29 * ANY EXPRESS OR IMPLIED WARRANTIES, INCLUDING, BUT NOT LIMITED TO, THE
30 * IMPLIED WARRANTIES OF MERCHANTABILITY AND FITNESS FOR A PARTICULAR PURPOSE
31 * ARE DISCLAIMED. IN NO EVENT SHALL THE REGENTS OR CONTRIBUTORS BE LIABLE
32 * FOR ANY DIRECT, INDIRECT, INCIDENTAL, SPECIAL, EXEMPLARY, OR CONSEQUENTIAL
33 * DAMAGES (INCLUDING, BUT NOT LIMITED TO, PROCUREMENT OF SUBSTITUTE GOODS
34 * OR SERVICES; LOSS OF USE, DATA, OR PROFITS; OR BUSINESS INTERRUPTION)
35 * HOWEVER CAUSED AND ON ANY THEORY OF LIABILITY, WHETHER IN CONTRACT, STRICT
36 * LIABILITY, OR TORT (INCLUDING NEGLIGENCE OR OTHERWISE) ARISING IN ANY WAY
37 * OUT OF THE USE OF THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF
43 * Radix Bitmap 'blists'.
45 * - The new swapper uses the new radix bitmap code. This should scale
46 * to arbitrarily small or arbitrarily large swap spaces and an almost
47 * arbitrary degree of fragmentation.
51 * - on the fly reallocation of swap during putpages. The new system
52 * does not try to keep previously allocated swap blocks for dirty
55 * - on the fly deallocation of swap
57 * - No more garbage collection required. Unnecessarily allocated swap
58 * blocks only exist for dirty vm_page_t's now and these are already
59 * cycled (in a high-load system) by the pager. We also do on-the-fly
60 * removal of invalidated swap blocks when a page is destroyed
63 * from: Utah $Hdr: swap_pager.c 1.4 91/04/30$
65 * @(#)swap_pager.c 8.9 (Berkeley) 3/21/94
67 * $FreeBSD: src/sys/vm/swap_pager.c,v 1.130.2.12 2002/08/31 21:15:55 dillon Exp $
70 #include <sys/param.h>
71 #include <sys/systm.h>
73 #include <sys/kernel.h>
76 #include <sys/vnode.h>
77 #include <sys/malloc.h>
78 #include <sys/vmmeter.h>
79 #include <sys/sysctl.h>
80 #include <sys/blist.h>
82 #include <sys/vmmeter.h>
84 #ifndef MAX_PAGEOUT_CLUSTER
85 #define MAX_PAGEOUT_CLUSTER 16
88 #define SWB_NPAGES MAX_PAGEOUT_CLUSTER
92 #include <vm/vm_object.h>
93 #include <vm/vm_page.h>
94 #include <vm/vm_pager.h>
95 #include <vm/vm_pageout.h>
96 #include <vm/swap_pager.h>
97 #include <vm/vm_extern.h>
98 #include <vm/vm_zone.h>
100 #define SWM_FREE 0x02 /* free, period */
101 #define SWM_POP 0x04 /* pop out */
104 * vm_swap_size is in page-sized chunks now. It was DEV_BSIZE'd chunks
108 extern int vm_swap_size; /* number of free swap blocks, in pages */
110 int swap_pager_full; /* swap space exhaustion (task killing) */
111 static int swap_pager_almost_full; /* swap space exhaustion (w/ hysteresis)*/
112 static int nsw_rcount; /* free read buffers */
113 static int nsw_wcount_sync; /* limit write buffers / synchronous */
114 static int nsw_wcount_async; /* limit write buffers / asynchronous */
115 static int nsw_wcount_async_max;/* assigned maximum */
116 static int nsw_cluster_max; /* maximum VOP I/O allowed */
117 static int sw_alloc_interlock; /* swap pager allocation interlock */
119 struct blist *swapblist;
120 static struct swblock **swhash;
121 static int swhash_mask;
122 static int swap_async_max = 4; /* maximum in-progress async I/O's */
124 extern struct vnode *swapdev_vp; /* from vm_swap.c */
126 SYSCTL_INT(_vm, OID_AUTO, swap_async_max,
127 CTLFLAG_RW, &swap_async_max, 0, "Maximum running async swap ops");
130 * "named" and "unnamed" anon region objects. Try to reduce the overhead
131 * of searching a named list by hashing it just a little.
136 #define NOBJLIST(handle) \
137 (&swap_pager_object_list[((int)(intptr_t)handle >> 4) & (NOBJLISTS-1)])
139 static struct pagerlst swap_pager_object_list[NOBJLISTS];
140 struct pagerlst swap_pager_un_object_list;
144 * pagerops for OBJT_SWAP - "swap pager". Some ops are also global procedure
145 * calls hooked from other parts of the VM system and do not appear here.
146 * (see vm/swap_pager.h).
150 swap_pager_alloc __P((void *handle, vm_ooffset_t size,
151 vm_prot_t prot, vm_ooffset_t offset));
152 static void swap_pager_dealloc __P((vm_object_t object));
153 static int swap_pager_getpages __P((vm_object_t, vm_page_t *, int, int));
154 static void swap_pager_init __P((void));
155 static void swap_pager_unswapped __P((vm_page_t));
156 static void swap_pager_strategy __P((vm_object_t, struct buf *));
158 struct pagerops swappagerops = {
159 swap_pager_init, /* early system initialization of pager */
160 swap_pager_alloc, /* allocate an OBJT_SWAP object */
161 swap_pager_dealloc, /* deallocate an OBJT_SWAP object */
162 swap_pager_getpages, /* pagein */
163 swap_pager_putpages, /* pageout */
164 swap_pager_haspage, /* get backing store status for page */
165 swap_pager_unswapped, /* remove swap related to page */
166 swap_pager_strategy /* pager strategy call */
170 * dmmax is in page-sized chunks with the new swap system. It was
171 * dev-bsized chunks in the old. dmmax is always a power of 2.
173 * swap_*() routines are externally accessible. swp_*() routines are
178 static int dmmax_mask;
179 int nswap_lowat = 128; /* in pages, swap_pager_almost_full warn */
180 int nswap_hiwat = 512; /* in pages, swap_pager_almost_full warn */
182 static __inline void swp_sizecheck __P((void));
183 static void swp_pager_sync_iodone __P((struct buf *bp));
184 static void swp_pager_async_iodone __P((struct buf *bp));
187 * Swap bitmap functions
190 static __inline void swp_pager_freeswapspace __P((daddr_t blk, int npages));
191 static __inline daddr_t swp_pager_getswapspace __P((int npages));
197 static void swp_pager_meta_build __P((vm_object_t, vm_pindex_t, daddr_t));
198 static void swp_pager_meta_free __P((vm_object_t, vm_pindex_t, daddr_t));
199 static void swp_pager_meta_free_all __P((vm_object_t));
200 static daddr_t swp_pager_meta_ctl __P((vm_object_t, vm_pindex_t, int));
203 * SWP_SIZECHECK() - update swap_pager_full indication
205 * update the swap_pager_almost_full indication and warn when we are
206 * about to run out of swap space, using lowat/hiwat hysteresis.
208 * Clear swap_pager_full ( task killing ) indication when lowat is met.
210 * No restrictions on call
211 * This routine may not block.
212 * This routine must be called at splvm()
218 if (vm_swap_size < nswap_lowat) {
219 if (swap_pager_almost_full == 0) {
220 printf("swap_pager: out of swap space\n");
221 swap_pager_almost_full = 1;
225 if (vm_swap_size > nswap_hiwat)
226 swap_pager_almost_full = 0;
231 * SWAP_PAGER_INIT() - initialize the swap pager!
233 * Expected to be started from system init. NOTE: This code is run
234 * before much else so be careful what you depend on. Most of the VM
235 * system has yet to be initialized at this point.
242 * Initialize object lists
246 for (i = 0; i < NOBJLISTS; ++i)
247 TAILQ_INIT(&swap_pager_object_list[i]);
248 TAILQ_INIT(&swap_pager_un_object_list);
251 * Device Stripe, in PAGE_SIZE'd blocks
254 dmmax = SWB_NPAGES * 2;
255 dmmax_mask = ~(dmmax - 1);
259 * SWAP_PAGER_SWAP_INIT() - swap pager initialization from pageout process
261 * Expected to be started from pageout process once, prior to entering
266 swap_pager_swap_init()
271 * Number of in-transit swap bp operations. Don't
272 * exhaust the pbufs completely. Make sure we
273 * initialize workable values (0 will work for hysteresis
274 * but it isn't very efficient).
276 * The nsw_cluster_max is constrained by the bp->b_pages[]
277 * array (MAXPHYS/PAGE_SIZE) and our locally defined
278 * MAX_PAGEOUT_CLUSTER. Also be aware that swap ops are
279 * constrained by the swap device interleave stripe size.
281 * Currently we hardwire nsw_wcount_async to 4. This limit is
282 * designed to prevent other I/O from having high latencies due to
283 * our pageout I/O. The value 4 works well for one or two active swap
284 * devices but is probably a little low if you have more. Even so,
285 * a higher value would probably generate only a limited improvement
286 * with three or four active swap devices since the system does not
287 * typically have to pageout at extreme bandwidths. We will want
288 * at least 2 per swap devices, and 4 is a pretty good value if you
289 * have one NFS swap device due to the command/ack latency over NFS.
290 * So it all works out pretty well.
293 nsw_cluster_max = min((MAXPHYS/PAGE_SIZE), MAX_PAGEOUT_CLUSTER);
295 nsw_rcount = (nswbuf + 1) / 2;
296 nsw_wcount_sync = (nswbuf + 3) / 4;
297 nsw_wcount_async = 4;
298 nsw_wcount_async_max = nsw_wcount_async;
301 * Initialize our zone. Right now I'm just guessing on the number
302 * we need based on the number of pages in the system. Each swblock
303 * can hold 16 pages, so this is probably overkill. This reservation
304 * is typically limited to around 32MB by default.
306 n = cnt.v_page_count / 2;
307 if (maxswzone && n > maxswzone / sizeof(struct swblock))
308 n = maxswzone / sizeof(struct swblock);
314 sizeof(struct swblock),
318 if (swap_zone != NULL)
321 * if the allocation failed, try a zone two thirds the
322 * size of the previous attempt.
327 if (swap_zone == NULL)
328 panic("swap_pager_swap_init: swap_zone == NULL");
330 printf("Swap zone entries reduced from %d to %d.\n", n2, n);
334 * Initialize our meta-data hash table. The swapper does not need to
335 * be quite as efficient as the VM system, so we do not use an
336 * oversized hash table.
338 * n: size of hash table, must be power of 2
339 * swhash_mask: hash table index mask
342 for (n = 1; n < n2 / 8; n *= 2)
345 swhash = malloc(sizeof(struct swblock *) * n, M_VMPGDATA, M_WAITOK);
346 bzero(swhash, sizeof(struct swblock *) * n);
352 * SWAP_PAGER_ALLOC() - allocate a new OBJT_SWAP VM object and instantiate
353 * its metadata structures.
355 * This routine is called from the mmap and fork code to create a new
356 * OBJT_SWAP object. We do this by creating an OBJT_DEFAULT object
357 * and then converting it with swp_pager_meta_build().
359 * This routine may block in vm_object_allocate() and create a named
360 * object lookup race, so we must interlock. We must also run at
361 * splvm() for the object lookup to handle races with interrupts, but
362 * we do not have to maintain splvm() in between the lookup and the
363 * add because (I believe) it is not possible to attempt to create
364 * a new swap object w/handle when a default object with that handle
369 swap_pager_alloc(void *handle, vm_ooffset_t size, vm_prot_t prot,
376 * Reference existing named region or allocate new one. There
377 * should not be a race here against swp_pager_meta_build()
378 * as called from vm_page_remove() in regards to the lookup
382 while (sw_alloc_interlock) {
383 sw_alloc_interlock = -1;
384 tsleep(&sw_alloc_interlock, PVM, "swpalc", 0);
386 sw_alloc_interlock = 1;
388 object = vm_pager_object_lookup(NOBJLIST(handle), handle);
390 if (object != NULL) {
391 vm_object_reference(object);
393 object = vm_object_allocate(OBJT_DEFAULT,
394 OFF_TO_IDX(offset + PAGE_MASK + size));
395 object->handle = handle;
397 swp_pager_meta_build(object, 0, SWAPBLK_NONE);
400 if (sw_alloc_interlock < 0)
401 wakeup(&sw_alloc_interlock);
403 sw_alloc_interlock = 0;
405 object = vm_object_allocate(OBJT_DEFAULT,
406 OFF_TO_IDX(offset + PAGE_MASK + size));
408 swp_pager_meta_build(object, 0, SWAPBLK_NONE);
415 * SWAP_PAGER_DEALLOC() - remove swap metadata from object
417 * The swap backing for the object is destroyed. The code is
418 * designed such that we can reinstantiate it later, but this
419 * routine is typically called only when the entire object is
420 * about to be destroyed.
422 * This routine may block, but no longer does.
424 * The object must be locked or unreferenceable.
428 swap_pager_dealloc(object)
434 * Remove from list right away so lookups will fail if we block for
435 * pageout completion.
438 if (object->handle == NULL) {
439 TAILQ_REMOVE(&swap_pager_un_object_list, object, pager_object_list);
441 TAILQ_REMOVE(NOBJLIST(object->handle), object, pager_object_list);
444 vm_object_pip_wait(object, "swpdea");
447 * Free all remaining metadata. We only bother to free it from
448 * the swap meta data. We do not attempt to free swapblk's still
449 * associated with vm_page_t's for this object. We do not care
450 * if paging is still in progress on some objects.
453 swp_pager_meta_free_all(object);
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 * Must be called at splvm() to avoid races with bitmap frees from
472 * vm_page_remove() aka swap_pager_page_removed().
474 * This routine may not block
475 * This routine must be called at splvm().
478 static __inline daddr_t
479 swp_pager_getswapspace(npages)
484 if ((blk = blist_alloc(swapblist, npages)) == SWAPBLK_NONE) {
485 if (swap_pager_full != 2) {
486 printf("swap_pager_getswapspace: failed\n");
488 swap_pager_almost_full = 1;
491 vm_swap_size -= npages;
498 * SWP_PAGER_FREESWAPSPACE() - free raw swap space
500 * This routine returns the specified swap blocks back to the bitmap.
502 * Note: This routine may not block (it could in the old swap code),
503 * and through the use of the new blist routines it does not block.
505 * We must be called at splvm() to avoid races with bitmap frees from
506 * vm_page_remove() aka swap_pager_page_removed().
508 * This routine may not block
509 * This routine must be called at splvm().
513 swp_pager_freeswapspace(blk, npages)
517 blist_free(swapblist, blk, npages);
518 vm_swap_size += npages;
523 * SWAP_PAGER_FREESPACE() - frees swap blocks associated with a page
524 * range within an object.
526 * This is a globally accessible routine.
528 * This routine removes swapblk assignments from swap metadata.
530 * The external callers of this routine typically have already destroyed
531 * or renamed vm_page_t's associated with this range in the object so
534 * This routine may be called at any spl. We up our spl to splvm temporarily
535 * in order to perform the metadata removal.
539 swap_pager_freespace(object, start, size)
545 swp_pager_meta_free(object, start, size);
550 * SWAP_PAGER_RESERVE() - reserve swap blocks in object
552 * Assigns swap blocks to the specified range within the object. The
553 * swap blocks are not zerod. Any previous swap assignment is destroyed.
555 * Returns 0 on success, -1 on failure.
559 swap_pager_reserve(vm_object_t object, vm_pindex_t start, vm_size_t size)
563 daddr_t blk = SWAPBLK_NONE;
564 vm_pindex_t beg = start; /* save start index */
570 while ((blk = swp_pager_getswapspace(n)) == SWAPBLK_NONE) {
573 swp_pager_meta_free(object, beg, start - beg);
579 swp_pager_meta_build(object, start, blk);
585 swp_pager_meta_free(object, start, n);
591 * SWAP_PAGER_COPY() - copy blocks from source pager to destination pager
592 * and destroy the source.
594 * Copy any valid swapblks from the source to the destination. In
595 * cases where both the source and destination have a valid swapblk,
596 * we keep the destination's.
598 * This routine is allowed to block. It may block allocating metadata
599 * indirectly through swp_pager_meta_build() or if paging is still in
600 * progress on the source.
602 * This routine can be called at any spl
604 * XXX vm_page_collapse() kinda expects us not to block because we
605 * supposedly do not need to allocate memory, but for the moment we
606 * *may* have to get a little memory from the zone allocator, but
607 * it is taken from the interrupt memory. We should be ok.
609 * The source object contains no vm_page_t's (which is just as well)
611 * The source object is of type OBJT_SWAP.
613 * The source and destination objects must be locked or
614 * inaccessible (XXX are they ?)
618 swap_pager_copy(srcobject, dstobject, offset, destroysource)
619 vm_object_t srcobject;
620 vm_object_t dstobject;
630 * If destroysource is set, we remove the source object from the
631 * swap_pager internal queue now.
635 if (srcobject->handle == NULL) {
637 &swap_pager_un_object_list,
643 NOBJLIST(srcobject->handle),
651 * transfer source to destination.
654 for (i = 0; i < dstobject->size; ++i) {
658 * Locate (without changing) the swapblk on the destination,
659 * unless it is invalid in which case free it silently, or
660 * if the destination is a resident page, in which case the
661 * source is thrown away.
664 dstaddr = swp_pager_meta_ctl(dstobject, i, 0);
666 if (dstaddr == SWAPBLK_NONE) {
668 * Destination has no swapblk and is not resident,
673 srcaddr = swp_pager_meta_ctl(
679 if (srcaddr != SWAPBLK_NONE)
680 swp_pager_meta_build(dstobject, i, srcaddr);
683 * Destination has valid swapblk or it is represented
684 * by a resident page. We destroy the sourceblock.
687 swp_pager_meta_ctl(srcobject, i + offset, SWM_FREE);
692 * Free left over swap blocks in source.
694 * We have to revert the type to OBJT_DEFAULT so we do not accidently
695 * double-remove the object from the swap queues.
699 swp_pager_meta_free_all(srcobject);
701 * Reverting the type is not necessary, the caller is going
702 * to destroy srcobject directly, but I'm doing it here
703 * for consistency since we've removed the object from its
706 srcobject->type = OBJT_DEFAULT;
712 * SWAP_PAGER_HASPAGE() - determine if we have good backing store for
713 * the requested page.
715 * We determine whether good backing store exists for the requested
716 * page and return TRUE if it does, FALSE if it doesn't.
718 * If TRUE, we also try to determine how much valid, contiguous backing
719 * store exists before and after the requested page within a reasonable
720 * distance. We do not try to restrict it to the swap device stripe
721 * (that is handled in getpages/putpages). It probably isn't worth
726 swap_pager_haspage(object, pindex, before, after)
736 * do we have good backing store at the requested index ?
740 blk0 = swp_pager_meta_ctl(object, pindex, 0);
742 if (blk0 == SWAPBLK_NONE) {
752 * find backwards-looking contiguous good backing store
755 if (before != NULL) {
758 for (i = 1; i < (SWB_NPAGES/2); ++i) {
763 blk = swp_pager_meta_ctl(object, pindex - i, 0);
771 * find forward-looking contiguous good backing store
777 for (i = 1; i < (SWB_NPAGES/2); ++i) {
780 blk = swp_pager_meta_ctl(object, pindex + i, 0);
791 * SWAP_PAGER_PAGE_UNSWAPPED() - remove swap backing store related to page
793 * This removes any associated swap backing store, whether valid or
794 * not, from the page.
796 * This routine is typically called when a page is made dirty, at
797 * which point any associated swap can be freed. MADV_FREE also
798 * calls us in a special-case situation
800 * NOTE!!! If the page is clean and the swap was valid, the caller
801 * should make the page dirty before calling this routine. This routine
802 * does NOT change the m->dirty status of the page. Also: MADV_FREE
805 * This routine may not block
806 * This routine must be called at splvm()
810 swap_pager_unswapped(m)
813 swp_pager_meta_ctl(m->object, m->pindex, SWM_FREE);
817 * SWAP_PAGER_STRATEGY() - read, write, free blocks
819 * This implements the vm_pager_strategy() interface to swap and allows
820 * other parts of the system to directly access swap as backing store
821 * through vm_objects of type OBJT_SWAP. This is intended to be a
822 * cacheless interface ( i.e. caching occurs at higher levels ).
823 * Therefore we do not maintain any resident pages. All I/O goes
824 * directly to and from the swap device.
826 * Note that b_blkno is scaled for PAGE_SIZE
828 * We currently attempt to run I/O synchronously or asynchronously as
829 * the caller requests. This isn't perfect because we loose error
830 * sequencing when we run multiple ops in parallel to satisfy a request.
831 * But this is swap, so we let it all hang out.
835 swap_pager_strategy(vm_object_t object, struct buf *bp)
841 struct buf *nbp = NULL;
843 if (bp->b_bcount & PAGE_MASK) {
844 bp->b_error = EINVAL;
845 bp->b_flags |= B_ERROR | B_INVAL;
847 printf("swap_pager_strategy: bp %p b_vp %p blk %d size %d, not page bounded\n", bp, bp->b_vp, (int)bp->b_pblkno, (int)bp->b_bcount);
852 * Clear error indication, initialize page index, count, data pointer.
856 bp->b_flags &= ~B_ERROR;
857 bp->b_resid = bp->b_bcount;
859 start = bp->b_pblkno;
860 count = howmany(bp->b_bcount, PAGE_SIZE);
866 * Deal with B_FREEBUF
869 if (bp->b_flags & B_FREEBUF) {
871 * FREE PAGE(s) - destroy underlying swap that is no longer
874 swp_pager_meta_free(object, start, count);
882 * Execute read or write
889 * Obtain block. If block not found and writing, allocate a
890 * new block and build it into the object.
893 blk = swp_pager_meta_ctl(object, start, 0);
894 if ((blk == SWAPBLK_NONE) && (bp->b_flags & B_READ) == 0) {
895 blk = swp_pager_getswapspace(1);
896 if (blk == SWAPBLK_NONE) {
897 bp->b_error = ENOMEM;
898 bp->b_flags |= B_ERROR;
901 swp_pager_meta_build(object, start, blk);
905 * Do we have to flush our current collection? Yes if:
907 * - no swap block at this index
908 * - swap block is not contiguous
909 * - we cross a physical disk boundry in the
914 nbp && (nbp->b_blkno + btoc(nbp->b_bcount) != blk ||
915 ((nbp->b_blkno ^ blk) & dmmax_mask)
919 if (bp->b_flags & B_READ) {
921 cnt.v_swappgsin += btoc(nbp->b_bcount);
924 cnt.v_swappgsout += btoc(nbp->b_bcount);
925 nbp->b_dirtyend = nbp->b_bcount;
933 * Add new swapblk to nbp, instantiating nbp if necessary.
934 * Zero-fill reads are able to take a shortcut.
937 if (blk == SWAPBLK_NONE) {
939 * We can only get here if we are reading. Since
940 * we are at splvm() we can safely modify b_resid,
941 * even if chain ops are in progress.
943 bzero(data, PAGE_SIZE);
944 bp->b_resid -= PAGE_SIZE;
947 nbp = getchainbuf(bp, swapdev_vp, (bp->b_flags & B_READ) | B_ASYNC);
952 nbp->b_bcount += PAGE_SIZE;
960 * Flush out last buffer
966 if ((bp->b_flags & B_ASYNC) == 0)
967 nbp->b_flags &= ~B_ASYNC;
968 if (nbp->b_flags & B_READ) {
970 cnt.v_swappgsin += btoc(nbp->b_bcount);
973 cnt.v_swappgsout += btoc(nbp->b_bcount);
974 nbp->b_dirtyend = nbp->b_bcount;
981 * Wait for completion.
984 if (bp->b_flags & B_ASYNC) {
987 waitchainbuf(bp, 0, 1);
992 * SWAP_PAGER_GETPAGES() - bring pages in from swap
994 * Attempt to retrieve (m, count) pages from backing store, but make
995 * sure we retrieve at least m[reqpage]. We try to load in as large
996 * a chunk surrounding m[reqpage] as is contiguous in swap and which
997 * belongs to the same object.
999 * The code is designed for asynchronous operation and
1000 * immediate-notification of 'reqpage' but tends not to be
1001 * used that way. Please do not optimize-out this algorithmic
1002 * feature, I intend to improve on it in the future.
1004 * The parent has a single vm_object_pip_add() reference prior to
1005 * calling us and we should return with the same.
1007 * The parent has BUSY'd the pages. We should return with 'm'
1008 * left busy, but the others adjusted.
1012 swap_pager_getpages(object, m, count, reqpage)
1024 vm_pindex_t lastpindex;
1028 if (mreq->object != object) {
1029 panic("swap_pager_getpages: object mismatch %p/%p",
1035 * Calculate range to retrieve. The pages have already been assigned
1036 * their swapblks. We require a *contiguous* range that falls entirely
1037 * within a single device stripe. If we do not supply it, bad things
1038 * happen. Note that blk, iblk & jblk can be SWAPBLK_NONE, but the
1039 * loops are set up such that the case(s) are handled implicitly.
1041 * The swp_*() calls must be made at splvm(). vm_page_free() does
1042 * not need to be, but it will go a little faster if it is.
1046 blk = swp_pager_meta_ctl(mreq->object, mreq->pindex, 0);
1048 for (i = reqpage - 1; i >= 0; --i) {
1051 iblk = swp_pager_meta_ctl(m[i]->object, m[i]->pindex, 0);
1052 if (blk != iblk + (reqpage - i))
1054 if ((blk ^ iblk) & dmmax_mask)
1059 for (j = reqpage + 1; j < count; ++j) {
1062 jblk = swp_pager_meta_ctl(m[j]->object, m[j]->pindex, 0);
1063 if (blk != jblk - (j - reqpage))
1065 if ((blk ^ jblk) & dmmax_mask)
1070 * free pages outside our collection range. Note: we never free
1071 * mreq, it must remain busy throughout.
1077 for (k = 0; k < i; ++k)
1079 for (k = j; k < count; ++k)
1086 * Return VM_PAGER_FAIL if we have nothing to do. Return mreq
1087 * still busy, but the others unbusied.
1090 if (blk == SWAPBLK_NONE)
1091 return(VM_PAGER_FAIL);
1094 * Get a swap buffer header to perform the IO
1097 bp = getpbuf(&nsw_rcount);
1098 kva = (vm_offset_t) bp->b_data;
1101 * map our page(s) into kva for input
1103 * NOTE: B_PAGING is set by pbgetvp()
1106 pmap_qenter(kva, m + i, j - i);
1108 bp->b_flags = B_READ | B_CALL;
1109 bp->b_iodone = swp_pager_async_iodone;
1110 bp->b_rcred = bp->b_wcred = proc0.p_ucred;
1111 bp->b_data = (caddr_t) kva;
1112 crhold(bp->b_rcred);
1113 crhold(bp->b_wcred);
1114 bp->b_blkno = blk - (reqpage - i);
1115 bp->b_bcount = PAGE_SIZE * (j - i);
1116 bp->b_bufsize = PAGE_SIZE * (j - i);
1117 bp->b_pager.pg_reqpage = reqpage - i;
1122 for (k = i; k < j; ++k) {
1123 bp->b_pages[k - i] = m[k];
1124 vm_page_flag_set(m[k], PG_SWAPINPROG);
1127 bp->b_npages = j - i;
1129 pbgetvp(swapdev_vp, bp);
1132 cnt.v_swappgsin += bp->b_npages;
1135 * We still hold the lock on mreq, and our automatic completion routine
1136 * does not remove it.
1139 vm_object_pip_add(mreq->object, bp->b_npages);
1140 lastpindex = m[j-1]->pindex;
1143 * perform the I/O. NOTE!!! bp cannot be considered valid after
1144 * this point because we automatically release it on completion.
1145 * Instead, we look at the one page we are interested in which we
1146 * still hold a lock on even through the I/O completion.
1148 * The other pages in our m[] array are also released on completion,
1149 * so we cannot assume they are valid anymore either.
1151 * NOTE: b_blkno is destroyed by the call to VOP_STRATEGY
1155 VOP_STRATEGY(bp->b_vp, bp);
1158 * wait for the page we want to complete. PG_SWAPINPROG is always
1159 * cleared on completion. If an I/O error occurs, SWAPBLK_NONE
1160 * is set in the meta-data.
1165 while ((mreq->flags & PG_SWAPINPROG) != 0) {
1166 vm_page_flag_set(mreq, PG_WANTED | PG_REFERENCED);
1168 if (tsleep(mreq, PSWP, "swread", hz*20)) {
1170 "swap_pager: indefinite wait buffer: device:"
1171 " %s, blkno: %ld, size: %ld\n",
1172 devtoname(bp->b_dev), (long)bp->b_blkno,
1181 * mreq is left bussied after completion, but all the other pages
1182 * are freed. If we had an unrecoverable read error the page will
1186 if (mreq->valid != VM_PAGE_BITS_ALL) {
1187 return(VM_PAGER_ERROR);
1189 return(VM_PAGER_OK);
1193 * A final note: in a low swap situation, we cannot deallocate swap
1194 * and mark a page dirty here because the caller is likely to mark
1195 * the page clean when we return, causing the page to possibly revert
1196 * to all-zero's later.
1201 * swap_pager_putpages:
1203 * Assign swap (if necessary) and initiate I/O on the specified pages.
1205 * We support both OBJT_DEFAULT and OBJT_SWAP objects. DEFAULT objects
1206 * are automatically converted to SWAP objects.
1208 * In a low memory situation we may block in VOP_STRATEGY(), but the new
1209 * vm_page reservation system coupled with properly written VFS devices
1210 * should ensure that no low-memory deadlock occurs. This is an area
1213 * The parent has N vm_object_pip_add() references prior to
1214 * calling us and will remove references for rtvals[] that are
1215 * not set to VM_PAGER_PEND. We need to remove the rest on I/O
1218 * The parent has soft-busy'd the pages it passes us and will unbusy
1219 * those whos rtvals[] entry is not set to VM_PAGER_PEND on return.
1220 * We need to unbusy the rest on I/O completion.
1224 swap_pager_putpages(object, m, count, sync, rtvals)
1234 if (count && m[0]->object != object) {
1235 panic("swap_pager_getpages: object mismatch %p/%p",
1243 * Turn object into OBJT_SWAP
1244 * check for bogus sysops
1245 * force sync if not pageout process
1248 if (object->type != OBJT_SWAP)
1249 swp_pager_meta_build(object, 0, SWAPBLK_NONE);
1251 if (curproc != pageproc)
1257 * Update nsw parameters from swap_async_max sysctl values.
1258 * Do not let the sysop crash the machine with bogus numbers.
1261 if (swap_async_max != nsw_wcount_async_max) {
1268 if ((n = swap_async_max) > nswbuf / 2)
1275 * Adjust difference ( if possible ). If the current async
1276 * count is too low, we may not be able to make the adjustment
1280 n -= nsw_wcount_async_max;
1281 if (nsw_wcount_async + n >= 0) {
1282 nsw_wcount_async += n;
1283 nsw_wcount_async_max += n;
1284 wakeup(&nsw_wcount_async);
1292 * Assign swap blocks and issue I/O. We reallocate swap on the fly.
1293 * The page is left dirty until the pageout operation completes
1297 for (i = 0; i < count; i += n) {
1304 * Maximum I/O size is limited by a number of factors.
1307 n = min(BLIST_MAX_ALLOC, count - i);
1308 n = min(n, nsw_cluster_max);
1313 * Get biggest block of swap we can. If we fail, fall
1314 * back and try to allocate a smaller block. Don't go
1315 * overboard trying to allocate space if it would overly
1319 (blk = swp_pager_getswapspace(n)) == SWAPBLK_NONE &&
1324 if (blk == SWAPBLK_NONE) {
1325 for (j = 0; j < n; ++j)
1326 rtvals[i+j] = VM_PAGER_FAIL;
1332 * The I/O we are constructing cannot cross a physical
1333 * disk boundry in the swap stripe. Note: we are still
1336 if ((blk ^ (blk + n)) & dmmax_mask) {
1337 j = ((blk + dmmax) & dmmax_mask) - blk;
1338 swp_pager_freeswapspace(blk + j, n - j);
1343 * All I/O parameters have been satisfied, build the I/O
1344 * request and assign the swap space.
1346 * NOTE: B_PAGING is set by pbgetvp()
1350 bp = getpbuf(&nsw_wcount_sync);
1351 bp->b_flags = B_CALL;
1353 bp = getpbuf(&nsw_wcount_async);
1354 bp->b_flags = B_CALL | B_ASYNC;
1356 bp->b_spc = NULL; /* not used, but NULL-out anyway */
1358 pmap_qenter((vm_offset_t)bp->b_data, &m[i], n);
1360 bp->b_rcred = bp->b_wcred = proc0.p_ucred;
1361 bp->b_bcount = PAGE_SIZE * n;
1362 bp->b_bufsize = PAGE_SIZE * n;
1365 crhold(bp->b_rcred);
1366 crhold(bp->b_wcred);
1368 pbgetvp(swapdev_vp, bp);
1370 for (j = 0; j < n; ++j) {
1371 vm_page_t mreq = m[i+j];
1373 swp_pager_meta_build(
1378 vm_page_dirty(mreq);
1379 rtvals[i+j] = VM_PAGER_OK;
1381 vm_page_flag_set(mreq, PG_SWAPINPROG);
1382 bp->b_pages[j] = mreq;
1386 * Must set dirty range for NFS to work.
1389 bp->b_dirtyend = bp->b_bcount;
1392 cnt.v_swappgsout += bp->b_npages;
1393 swapdev_vp->v_numoutput++;
1400 * NOTE: b_blkno is destroyed by the call to VOP_STRATEGY
1403 if (sync == FALSE) {
1404 bp->b_iodone = swp_pager_async_iodone;
1406 VOP_STRATEGY(bp->b_vp, bp);
1408 for (j = 0; j < n; ++j)
1409 rtvals[i+j] = VM_PAGER_PEND;
1416 * NOTE: b_blkno is destroyed by the call to VOP_STRATEGY
1419 bp->b_iodone = swp_pager_sync_iodone;
1420 VOP_STRATEGY(bp->b_vp, bp);
1423 * Wait for the sync I/O to complete, then update rtvals.
1424 * We just set the rtvals[] to VM_PAGER_PEND so we can call
1425 * our async completion routine at the end, thus avoiding a
1430 while ((bp->b_flags & B_DONE) == 0) {
1431 tsleep(bp, PVM, "swwrt", 0);
1434 for (j = 0; j < n; ++j)
1435 rtvals[i+j] = VM_PAGER_PEND;
1438 * Now that we are through with the bp, we can call the
1439 * normal async completion, which frees everything up.
1442 swp_pager_async_iodone(bp);
1449 * swap_pager_sync_iodone:
1451 * Completion routine for synchronous reads and writes from/to swap.
1452 * We just mark the bp is complete and wake up anyone waiting on it.
1454 * This routine may not block. This routine is called at splbio() or better.
1458 swp_pager_sync_iodone(bp)
1461 bp->b_flags |= B_DONE;
1462 bp->b_flags &= ~B_ASYNC;
1467 * swp_pager_async_iodone:
1469 * Completion routine for asynchronous reads and writes from/to swap.
1470 * Also called manually by synchronous code to finish up a bp.
1472 * For READ operations, the pages are PG_BUSY'd. For WRITE operations,
1473 * the pages are vm_page_t->busy'd. For READ operations, we PG_BUSY
1474 * unbusy all pages except the 'main' request page. For WRITE
1475 * operations, we vm_page_t->busy'd unbusy all pages ( we can do this
1476 * because we marked them all VM_PAGER_PEND on return from putpages ).
1478 * This routine may not block.
1479 * This routine is called at splbio() or better
1481 * We up ourselves to splvm() as required for various vm_page related
1486 swp_pager_async_iodone(bp)
1487 register struct buf *bp;
1491 vm_object_t object = NULL;
1493 bp->b_flags |= B_DONE;
1499 if (bp->b_flags & B_ERROR) {
1501 "swap_pager: I/O error - %s failed; blkno %ld,"
1502 "size %ld, error %d\n",
1503 ((bp->b_flags & B_READ) ? "pagein" : "pageout"),
1511 * set object, raise to splvm().
1515 object = bp->b_pages[0]->object;
1519 * remove the mapping for kernel virtual
1522 pmap_qremove((vm_offset_t)bp->b_data, bp->b_npages);
1525 * cleanup pages. If an error occurs writing to swap, we are in
1526 * very serious trouble. If it happens to be a disk error, though,
1527 * we may be able to recover by reassigning the swap later on. So
1528 * in this case we remove the m->swapblk assignment for the page
1529 * but do not free it in the rlist. The errornous block(s) are thus
1530 * never reallocated as swap. Redirty the page and continue.
1533 for (i = 0; i < bp->b_npages; ++i) {
1534 vm_page_t m = bp->b_pages[i];
1536 vm_page_flag_clear(m, PG_SWAPINPROG);
1538 if (bp->b_flags & B_ERROR) {
1540 * If an error occurs I'd love to throw the swapblk
1541 * away without freeing it back to swapspace, so it
1542 * can never be used again. But I can't from an
1546 if (bp->b_flags & B_READ) {
1548 * When reading, reqpage needs to stay
1549 * locked for the parent, but all other
1550 * pages can be freed. We still want to
1551 * wakeup the parent waiting on the page,
1552 * though. ( also: pg_reqpage can be -1 and
1553 * not match anything ).
1555 * We have to wake specifically requested pages
1556 * up too because we cleared PG_SWAPINPROG and
1557 * someone may be waiting for that.
1559 * NOTE: for reads, m->dirty will probably
1560 * be overridden by the original caller of
1561 * getpages so don't play cute tricks here.
1563 * XXX IT IS NOT LEGAL TO FREE THE PAGE HERE
1564 * AS THIS MESSES WITH object->memq, and it is
1565 * not legal to mess with object->memq from an
1570 vm_page_flag_clear(m, PG_ZERO);
1572 if (i != bp->b_pager.pg_reqpage)
1577 * If i == bp->b_pager.pg_reqpage, do not wake
1578 * the page up. The caller needs to.
1582 * If a write error occurs, reactivate page
1583 * so it doesn't clog the inactive list,
1584 * then finish the I/O.
1587 vm_page_activate(m);
1588 vm_page_io_finish(m);
1590 } else if (bp->b_flags & B_READ) {
1592 * For read success, clear dirty bits. Nobody should
1593 * have this page mapped but don't take any chances,
1594 * make sure the pmap modify bits are also cleared.
1596 * NOTE: for reads, m->dirty will probably be
1597 * overridden by the original caller of getpages so
1598 * we cannot set them in order to free the underlying
1599 * swap in a low-swap situation. I don't think we'd
1600 * want to do that anyway, but it was an optimization
1601 * that existed in the old swapper for a time before
1602 * it got ripped out due to precisely this problem.
1604 * clear PG_ZERO in page.
1606 * If not the requested page then deactivate it.
1608 * Note that the requested page, reqpage, is left
1609 * busied, but we still have to wake it up. The
1610 * other pages are released (unbusied) by
1611 * vm_page_wakeup(). We do not set reqpage's
1612 * valid bits here, it is up to the caller.
1615 pmap_clear_modify(m);
1616 m->valid = VM_PAGE_BITS_ALL;
1618 vm_page_flag_clear(m, PG_ZERO);
1621 * We have to wake specifically requested pages
1622 * up too because we cleared PG_SWAPINPROG and
1623 * could be waiting for it in getpages. However,
1624 * be sure to not unbusy getpages specifically
1625 * requested page - getpages expects it to be
1628 if (i != bp->b_pager.pg_reqpage) {
1629 vm_page_deactivate(m);
1636 * For write success, clear the modify and dirty
1637 * status, then finish the I/O ( which decrements the
1638 * busy count and possibly wakes waiter's up ).
1640 pmap_clear_modify(m);
1642 vm_page_io_finish(m);
1643 if (!vm_page_count_severe() || !vm_page_try_to_cache(m))
1644 vm_page_protect(m, VM_PROT_READ);
1649 * adjust pip. NOTE: the original parent may still have its own
1650 * pip refs on the object.
1654 vm_object_pip_wakeupn(object, bp->b_npages);
1657 * release the physical I/O buffer
1662 ((bp->b_flags & B_READ) ? &nsw_rcount :
1663 ((bp->b_flags & B_ASYNC) ?
1672 /************************************************************************
1674 ************************************************************************
1676 * These routines manipulate the swap metadata stored in the
1677 * OBJT_SWAP object. All swp_*() routines must be called at
1678 * splvm() because swap can be freed up by the low level vm_page
1679 * code which might be called from interrupts beyond what splbio() covers.
1681 * Swap metadata is implemented with a global hash and not directly
1682 * linked into the object. Instead the object simply contains
1683 * appropriate tracking counters.
1687 * SWP_PAGER_HASH() - hash swap meta data
1689 * This is an inline helper function which hashes the swapblk given
1690 * the object and page index. It returns a pointer to a pointer
1691 * to the object, or a pointer to a NULL pointer if it could not
1694 * This routine must be called at splvm().
1697 static __inline struct swblock **
1698 swp_pager_hash(vm_object_t object, vm_pindex_t index)
1700 struct swblock **pswap;
1701 struct swblock *swap;
1703 index &= ~SWAP_META_MASK;
1704 pswap = &swhash[(index ^ (int)(intptr_t)object) & swhash_mask];
1706 while ((swap = *pswap) != NULL) {
1707 if (swap->swb_object == object &&
1708 swap->swb_index == index
1712 pswap = &swap->swb_hnext;
1718 * SWP_PAGER_META_BUILD() - add swap block to swap meta data for object
1720 * We first convert the object to a swap object if it is a default
1723 * The specified swapblk is added to the object's swap metadata. If
1724 * the swapblk is not valid, it is freed instead. Any previously
1725 * assigned swapblk is freed.
1727 * This routine must be called at splvm(), except when used to convert
1728 * an OBJT_DEFAULT object into an OBJT_SWAP object.
1733 swp_pager_meta_build(
1738 struct swblock *swap;
1739 struct swblock **pswap;
1742 * Convert default object to swap object if necessary
1745 if (object->type != OBJT_SWAP) {
1746 object->type = OBJT_SWAP;
1747 object->un_pager.swp.swp_bcount = 0;
1749 if (object->handle != NULL) {
1751 NOBJLIST(object->handle),
1757 &swap_pager_un_object_list,
1765 * Locate hash entry. If not found create, but if we aren't adding
1766 * anything just return. If we run out of space in the map we wait
1767 * and, since the hash table may have changed, retry.
1771 pswap = swp_pager_hash(object, index);
1773 if ((swap = *pswap) == NULL) {
1776 if (swapblk == SWAPBLK_NONE)
1779 swap = *pswap = zalloc(swap_zone);
1784 swap->swb_hnext = NULL;
1785 swap->swb_object = object;
1786 swap->swb_index = index & ~SWAP_META_MASK;
1787 swap->swb_count = 0;
1789 ++object->un_pager.swp.swp_bcount;
1791 for (i = 0; i < SWAP_META_PAGES; ++i)
1792 swap->swb_pages[i] = SWAPBLK_NONE;
1796 * Delete prior contents of metadata
1799 index &= SWAP_META_MASK;
1801 if (swap->swb_pages[index] != SWAPBLK_NONE) {
1802 swp_pager_freeswapspace(swap->swb_pages[index], 1);
1807 * Enter block into metadata
1810 swap->swb_pages[index] = swapblk;
1811 if (swapblk != SWAPBLK_NONE)
1816 * SWP_PAGER_META_FREE() - free a range of blocks in the object's swap metadata
1818 * The requested range of blocks is freed, with any associated swap
1819 * returned to the swap bitmap.
1821 * This routine will free swap metadata structures as they are cleaned
1822 * out. This routine does *NOT* operate on swap metadata associated
1823 * with resident pages.
1825 * This routine must be called at splvm()
1829 swp_pager_meta_free(vm_object_t object, vm_pindex_t index, daddr_t count)
1831 if (object->type != OBJT_SWAP)
1835 struct swblock **pswap;
1836 struct swblock *swap;
1838 pswap = swp_pager_hash(object, index);
1840 if ((swap = *pswap) != NULL) {
1841 daddr_t v = swap->swb_pages[index & SWAP_META_MASK];
1843 if (v != SWAPBLK_NONE) {
1844 swp_pager_freeswapspace(v, 1);
1845 swap->swb_pages[index & SWAP_META_MASK] =
1847 if (--swap->swb_count == 0) {
1848 *pswap = swap->swb_hnext;
1849 zfree(swap_zone, swap);
1850 --object->un_pager.swp.swp_bcount;
1856 int n = SWAP_META_PAGES - (index & SWAP_META_MASK);
1864 * SWP_PAGER_META_FREE_ALL() - destroy all swap metadata associated with object
1866 * This routine locates and destroys all swap metadata associated with
1869 * This routine must be called at splvm()
1873 swp_pager_meta_free_all(vm_object_t object)
1877 if (object->type != OBJT_SWAP)
1880 while (object->un_pager.swp.swp_bcount) {
1881 struct swblock **pswap;
1882 struct swblock *swap;
1884 pswap = swp_pager_hash(object, index);
1885 if ((swap = *pswap) != NULL) {
1888 for (i = 0; i < SWAP_META_PAGES; ++i) {
1889 daddr_t v = swap->swb_pages[i];
1890 if (v != SWAPBLK_NONE) {
1892 swp_pager_freeswapspace(v, 1);
1895 if (swap->swb_count != 0)
1896 panic("swap_pager_meta_free_all: swb_count != 0");
1897 *pswap = swap->swb_hnext;
1898 zfree(swap_zone, swap);
1899 --object->un_pager.swp.swp_bcount;
1901 index += SWAP_META_PAGES;
1902 if (index > 0x20000000)
1903 panic("swp_pager_meta_free_all: failed to locate all swap meta blocks");
1908 * SWP_PAGER_METACTL() - misc control of swap and vm_page_t meta data.
1910 * This routine is capable of looking up, popping, or freeing
1911 * swapblk assignments in the swap meta data or in the vm_page_t.
1912 * The routine typically returns the swapblk being looked-up, or popped,
1913 * or SWAPBLK_NONE if the block was freed, or SWAPBLK_NONE if the block
1914 * was invalid. This routine will automatically free any invalid
1915 * meta-data swapblks.
1917 * It is not possible to store invalid swapblks in the swap meta data
1918 * (other then a literal 'SWAPBLK_NONE'), so we don't bother checking.
1920 * When acting on a busy resident page and paging is in progress, we
1921 * have to wait until paging is complete but otherwise can act on the
1924 * This routine must be called at splvm().
1926 * SWM_FREE remove and free swap block from metadata
1927 * SWM_POP remove from meta data but do not free.. pop it out
1936 struct swblock **pswap;
1937 struct swblock *swap;
1941 * The meta data only exists of the object is OBJT_SWAP
1942 * and even then might not be allocated yet.
1945 if (object->type != OBJT_SWAP)
1946 return(SWAPBLK_NONE);
1949 pswap = swp_pager_hash(object, index);
1951 if ((swap = *pswap) != NULL) {
1952 index &= SWAP_META_MASK;
1953 r1 = swap->swb_pages[index];
1955 if (r1 != SWAPBLK_NONE) {
1956 if (flags & SWM_FREE) {
1957 swp_pager_freeswapspace(r1, 1);
1960 if (flags & (SWM_FREE|SWM_POP)) {
1961 swap->swb_pages[index] = SWAPBLK_NONE;
1962 if (--swap->swb_count == 0) {
1963 *pswap = swap->swb_hnext;
1964 zfree(swap_zone, swap);
1965 --object->un_pager.swp.swp_bcount;