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
70 #include <sys/param.h>
71 #include <sys/systm.h>
73 #include <sys/kernel.h>
77 #include <sys/vnode.h>
78 #include <sys/malloc.h>
79 #include <sys/vmmeter.h>
80 #include <sys/sysctl.h>
81 #include <sys/blist.h>
84 #include <sys/vmmeter.h>
86 #ifndef MAX_PAGEOUT_CLUSTER
87 #define MAX_PAGEOUT_CLUSTER 16
90 #define SWB_NPAGES MAX_PAGEOUT_CLUSTER
95 #include <vm/vm_map.h>
96 #include <vm/vm_kern.h>
97 #include <vm/vm_object.h>
98 #include <vm/vm_page.h>
99 #include <vm/vm_pager.h>
100 #include <vm/vm_pageout.h>
101 #include <vm/swap_pager.h>
102 #include <vm/vm_extern.h>
105 #define SWM_FREE 0x02 /* free, period */
106 #define SWM_POP 0x04 /* pop out */
109 * vm_swap_size is in page-sized chunks now. It was DEV_BSIZE'd chunks
112 extern int vm_swap_size; /* number of free swap blocks, in pages */
114 int swap_pager_full; /* swap space exhaustion (task killing) */
115 static int swap_pager_almost_full; /* swap space exhaustion (w/ hysteresis)*/
116 static int nsw_rcount; /* free read buffers */
117 static int nsw_wcount_sync; /* limit write buffers / synchronous */
118 static int nsw_wcount_async; /* limit write buffers / asynchronous */
119 static int nsw_wcount_async_max;/* assigned maximum */
120 static int nsw_cluster_max; /* maximum VOP I/O allowed */
122 struct blist *swapblist;
123 static struct swblock **swhash;
124 static int swhash_mask;
125 static int swap_async_max = 4; /* maximum in-progress async I/O's */
126 static struct sx sw_alloc_sx;
129 extern struct vnode *swapdev_vp;
130 extern struct swdevt *swdevt;
133 SYSCTL_INT(_vm, OID_AUTO, swap_async_max,
134 CTLFLAG_RW, &swap_async_max, 0, "Maximum running async swap ops");
136 #define BLK2DEVIDX(blk) (nswdev > 1 ? blk / dmmax % nswdev : 0)
139 * "named" and "unnamed" anon region objects. Try to reduce the overhead
140 * of searching a named list by hashing it just a little.
145 #define NOBJLIST(handle) \
146 (&swap_pager_object_list[((int)(intptr_t)handle >> 4) & (NOBJLISTS-1)])
148 static struct mtx sw_alloc_mtx; /* protect list manipulation */
149 static struct pagerlst swap_pager_object_list[NOBJLISTS];
150 struct pagerlst swap_pager_un_object_list;
151 uma_zone_t swap_zone;
154 * pagerops for OBJT_SWAP - "swap pager". Some ops are also global procedure
155 * calls hooked from other parts of the VM system and do not appear here.
156 * (see vm/swap_pager.h).
159 swap_pager_alloc(void *handle, vm_ooffset_t size,
160 vm_prot_t prot, vm_ooffset_t offset);
161 static void swap_pager_dealloc(vm_object_t object);
162 static int swap_pager_getpages(vm_object_t, vm_page_t *, int, int);
163 static void swap_pager_init(void);
164 static void swap_pager_unswapped(vm_page_t);
165 static void swap_pager_strategy(vm_object_t, struct bio *);
167 struct pagerops swappagerops = {
168 swap_pager_init, /* early system initialization of pager */
169 swap_pager_alloc, /* allocate an OBJT_SWAP object */
170 swap_pager_dealloc, /* deallocate an OBJT_SWAP object */
171 swap_pager_getpages, /* pagein */
172 swap_pager_putpages, /* pageout */
173 swap_pager_haspage, /* get backing store status for page */
174 swap_pager_unswapped, /* remove swap related to page */
175 swap_pager_strategy /* pager strategy call */
178 static struct buf *getchainbuf(struct bio *bp, struct vnode *vp, int flags);
179 static void flushchainbuf(struct buf *nbp);
180 static void waitchainbuf(struct bio *bp, int count, int done);
183 * dmmax is in page-sized chunks with the new swap system. It was
184 * dev-bsized chunks in the old. dmmax is always a power of 2.
186 * swap_*() routines are externally accessible. swp_*() routines are
190 static int dmmax_mask;
191 int nswap_lowat = 128; /* in pages, swap_pager_almost_full warn */
192 int nswap_hiwat = 512; /* in pages, swap_pager_almost_full warn */
194 SYSCTL_INT(_vm, OID_AUTO, dmmax,
195 CTLFLAG_RD, &dmmax, 0, "Maximum size of a swap block");
197 static __inline void swp_sizecheck(void);
198 static void swp_pager_sync_iodone(struct buf *bp);
199 static void swp_pager_async_iodone(struct buf *bp);
202 * Swap bitmap functions
204 static __inline void swp_pager_freeswapspace(daddr_t blk, int npages);
205 static __inline daddr_t swp_pager_getswapspace(int npages);
210 static void swp_pager_meta_build(vm_object_t, vm_pindex_t, daddr_t);
211 static void swp_pager_meta_free(vm_object_t, vm_pindex_t, daddr_t);
212 static void swp_pager_meta_free_all(vm_object_t);
213 static daddr_t swp_pager_meta_ctl(vm_object_t, vm_pindex_t, int);
216 * SWP_SIZECHECK() - update swap_pager_full indication
218 * update the swap_pager_almost_full indication and warn when we are
219 * about to run out of swap space, using lowat/hiwat hysteresis.
221 * Clear swap_pager_full ( task killing ) indication when lowat is met.
223 * No restrictions on call
224 * This routine may not block.
225 * This routine must be called at splvm()
232 if (vm_swap_size < nswap_lowat) {
233 if (swap_pager_almost_full == 0) {
234 printf("swap_pager: out of swap space\n");
235 swap_pager_almost_full = 1;
239 if (vm_swap_size > nswap_hiwat)
240 swap_pager_almost_full = 0;
245 * SWAP_PAGER_INIT() - initialize the swap pager!
247 * Expected to be started from system init. NOTE: This code is run
248 * before much else so be careful what you depend on. Most of the VM
249 * system has yet to be initialized at this point.
255 * Initialize object lists
259 for (i = 0; i < NOBJLISTS; ++i)
260 TAILQ_INIT(&swap_pager_object_list[i]);
261 TAILQ_INIT(&swap_pager_un_object_list);
262 mtx_init(&sw_alloc_mtx, "swap_pager list", NULL, MTX_DEF);
265 * Device Stripe, in PAGE_SIZE'd blocks
267 dmmax = SWB_NPAGES * 2;
268 dmmax_mask = ~(dmmax - 1);
272 * SWAP_PAGER_SWAP_INIT() - swap pager initialization from pageout process
274 * Expected to be started from pageout process once, prior to entering
278 swap_pager_swap_init()
283 * Number of in-transit swap bp operations. Don't
284 * exhaust the pbufs completely. Make sure we
285 * initialize workable values (0 will work for hysteresis
286 * but it isn't very efficient).
288 * The nsw_cluster_max is constrained by the bp->b_pages[]
289 * array (MAXPHYS/PAGE_SIZE) and our locally defined
290 * MAX_PAGEOUT_CLUSTER. Also be aware that swap ops are
291 * constrained by the swap device interleave stripe size.
293 * Currently we hardwire nsw_wcount_async to 4. This limit is
294 * designed to prevent other I/O from having high latencies due to
295 * our pageout I/O. The value 4 works well for one or two active swap
296 * devices but is probably a little low if you have more. Even so,
297 * a higher value would probably generate only a limited improvement
298 * with three or four active swap devices since the system does not
299 * typically have to pageout at extreme bandwidths. We will want
300 * at least 2 per swap devices, and 4 is a pretty good value if you
301 * have one NFS swap device due to the command/ack latency over NFS.
302 * So it all works out pretty well.
304 nsw_cluster_max = min((MAXPHYS/PAGE_SIZE), MAX_PAGEOUT_CLUSTER);
307 nsw_rcount = (nswbuf + 1) / 2;
308 nsw_wcount_sync = (nswbuf + 3) / 4;
309 nsw_wcount_async = 4;
310 nsw_wcount_async_max = nsw_wcount_async;
311 mtx_unlock(&pbuf_mtx);
314 * Initialize our zone. Right now I'm just guessing on the number
315 * we need based on the number of pages in the system. Each swblock
316 * can hold 16 pages, so this is probably overkill. This reservation
317 * is typically limited to around 70MB by default.
319 n = cnt.v_page_count;
320 if (maxswzone && n > maxswzone / sizeof(struct swblock))
321 n = maxswzone / sizeof(struct swblock);
323 swap_zone = uma_zcreate("SWAPMETA", sizeof(struct swblock), NULL, NULL,
324 NULL, NULL, UMA_ALIGN_PTR, UMA_ZONE_NOFREE);
326 if (uma_zone_set_obj(swap_zone, NULL, n))
329 * if the allocation failed, try a zone two thirds the
330 * size of the previous attempt.
334 if (swap_zone == NULL)
335 panic("failed to create swap_zone.");
337 printf("Swap zone entries reduced from %d to %d.\n", n2, n);
341 * Initialize our meta-data hash table. The swapper does not need to
342 * be quite as efficient as the VM system, so we do not use an
343 * oversized hash table.
345 * n: size of hash table, must be power of 2
346 * swhash_mask: hash table index mask
348 for (n = 1; n < n2 / 8; n *= 2)
350 swhash = malloc(sizeof(struct swblock *) * n, M_VMPGDATA, M_WAITOK | M_ZERO);
355 * SWAP_PAGER_ALLOC() - allocate a new OBJT_SWAP VM object and instantiate
356 * its metadata structures.
358 * This routine is called from the mmap and fork code to create a new
359 * OBJT_SWAP object. We do this by creating an OBJT_DEFAULT object
360 * and then converting it with swp_pager_meta_build().
362 * This routine may block in vm_object_allocate() and create a named
363 * object lookup race, so we must interlock. We must also run at
364 * splvm() for the object lookup to handle races with interrupts, but
365 * we do not have to maintain splvm() in between the lookup and the
366 * add because (I believe) it is not possible to attempt to create
367 * a new swap object w/handle when a default object with that handle
373 swap_pager_alloc(void *handle, vm_ooffset_t size, vm_prot_t prot,
381 * Reference existing named region or allocate new one. There
382 * should not be a race here against swp_pager_meta_build()
383 * as called from vm_page_remove() in regards to the lookup
386 sx_xlock(&sw_alloc_sx);
387 object = vm_pager_object_lookup(NOBJLIST(handle), handle);
389 if (object != NULL) {
390 vm_object_reference(object);
392 object = vm_object_allocate(OBJT_DEFAULT,
393 OFF_TO_IDX(offset + PAGE_MASK + size));
394 object->handle = handle;
396 swp_pager_meta_build(object, 0, SWAPBLK_NONE);
398 sx_xunlock(&sw_alloc_sx);
400 object = vm_object_allocate(OBJT_DEFAULT,
401 OFF_TO_IDX(offset + PAGE_MASK + size));
403 swp_pager_meta_build(object, 0, SWAPBLK_NONE);
410 * SWAP_PAGER_DEALLOC() - remove swap metadata from object
412 * The swap backing for the object is destroyed. The code is
413 * designed such that we can reinstantiate it later, but this
414 * routine is typically called only when the entire object is
415 * about to be destroyed.
417 * This routine may block, but no longer does.
419 * The object must be locked or unreferenceable.
422 swap_pager_dealloc(object)
430 * Remove from list right away so lookups will fail if we block for
431 * pageout completion.
433 mtx_lock(&sw_alloc_mtx);
434 if (object->handle == NULL) {
435 TAILQ_REMOVE(&swap_pager_un_object_list, object, pager_object_list);
437 TAILQ_REMOVE(NOBJLIST(object->handle), object, pager_object_list);
439 mtx_unlock(&sw_alloc_mtx);
441 vm_object_pip_wait(object, "swpdea");
444 * Free all remaining metadata. We only bother to free it from
445 * the swap meta data. We do not attempt to free swapblk's still
446 * associated with vm_page_t's for this object. We do not care
447 * if paging is still in progress on some objects.
450 swp_pager_meta_free_all(object);
454 /************************************************************************
455 * SWAP PAGER BITMAP ROUTINES *
456 ************************************************************************/
459 * SWP_PAGER_GETSWAPSPACE() - allocate raw swap space
461 * Allocate swap for the requested number of pages. The starting
462 * swap block number (a page index) is returned or SWAPBLK_NONE
463 * if the allocation failed.
465 * Also has the side effect of advising that somebody made a mistake
466 * when they configured swap and didn't configure enough.
468 * Must be called at splvm() to avoid races with bitmap frees from
469 * vm_page_remove() aka swap_pager_page_removed().
471 * This routine may not block
472 * This routine must be called at splvm().
474 static __inline daddr_t
475 swp_pager_getswapspace(npages)
482 if ((blk = blist_alloc(swapblist, npages)) == SWAPBLK_NONE) {
483 if (swap_pager_full != 2) {
484 printf("swap_pager_getswapspace: failed\n");
486 swap_pager_almost_full = 1;
489 vm_swap_size -= npages;
490 /* per-swap area stats */
491 swdevt[BLK2DEVIDX(blk)].sw_used += 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().
512 swp_pager_freeswapspace(blk, npages)
518 blist_free(swapblist, blk, npages);
519 vm_swap_size += npages;
520 /* per-swap area stats */
521 swdevt[BLK2DEVIDX(blk)].sw_used -= npages;
526 * SWAP_PAGER_FREESPACE() - frees swap blocks associated with a page
527 * range within an object.
529 * This is a globally accessible routine.
531 * This routine removes swapblk assignments from swap metadata.
533 * The external callers of this routine typically have already destroyed
534 * or renamed vm_page_t's associated with this range in the object so
537 * This routine may be called at any spl. We up our spl to splvm temporarily
538 * in order to perform the metadata removal.
541 swap_pager_freespace(object, start, size)
549 swp_pager_meta_free(object, start, size);
554 * SWAP_PAGER_RESERVE() - reserve swap blocks in object
556 * Assigns swap blocks to the specified range within the object. The
557 * swap blocks are not zerod. Any previous swap assignment is destroyed.
559 * Returns 0 on success, -1 on failure.
562 swap_pager_reserve(vm_object_t object, vm_pindex_t start, vm_size_t size)
566 daddr_t blk = SWAPBLK_NONE;
567 vm_pindex_t beg = start; /* save start index */
573 while ((blk = swp_pager_getswapspace(n)) == SWAPBLK_NONE) {
576 swp_pager_meta_free(object, beg, start - beg);
582 swp_pager_meta_build(object, start, blk);
588 swp_pager_meta_free(object, start, n);
594 * SWAP_PAGER_COPY() - copy blocks from source pager to destination pager
595 * and destroy the source.
597 * Copy any valid swapblks from the source to the destination. In
598 * cases where both the source and destination have a valid swapblk,
599 * we keep the destination's.
601 * This routine is allowed to block. It may block allocating metadata
602 * indirectly through swp_pager_meta_build() or if paging is still in
603 * progress on the source.
605 * This routine can be called at any spl
607 * XXX vm_page_collapse() kinda expects us not to block because we
608 * supposedly do not need to allocate memory, but for the moment we
609 * *may* have to get a little memory from the zone allocator, but
610 * it is taken from the interrupt memory. We should be ok.
612 * The source object contains no vm_page_t's (which is just as well)
614 * The source object is of type OBJT_SWAP.
616 * The source and destination objects must be locked or
617 * inaccessible (XXX are they ?)
620 swap_pager_copy(srcobject, dstobject, offset, destroysource)
621 vm_object_t srcobject;
622 vm_object_t dstobject;
633 * If destroysource is set, we remove the source object from the
634 * swap_pager internal queue now.
637 mtx_lock(&sw_alloc_mtx);
638 if (srcobject->handle == NULL) {
640 &swap_pager_un_object_list,
646 NOBJLIST(srcobject->handle),
651 mtx_unlock(&sw_alloc_mtx);
655 * transfer source to destination.
657 for (i = 0; i < dstobject->size; ++i) {
661 * Locate (without changing) the swapblk on the destination,
662 * unless it is invalid in which case free it silently, or
663 * if the destination is a resident page, in which case the
664 * source is thrown away.
666 dstaddr = swp_pager_meta_ctl(dstobject, i, 0);
668 if (dstaddr == SWAPBLK_NONE) {
670 * Destination has no swapblk and is not resident,
675 srcaddr = swp_pager_meta_ctl(
681 if (srcaddr != SWAPBLK_NONE)
682 swp_pager_meta_build(dstobject, i, srcaddr);
685 * Destination has valid swapblk or it is represented
686 * by a resident page. We destroy the sourceblock.
689 swp_pager_meta_ctl(srcobject, i + offset, SWM_FREE);
694 * Free left over swap blocks in source.
696 * We have to revert the type to OBJT_DEFAULT so we do not accidently
697 * double-remove the object from the swap queues.
700 swp_pager_meta_free_all(srcobject);
702 * Reverting the type is not necessary, the caller is going
703 * to destroy srcobject directly, but I'm doing it here
704 * for consistency since we've removed the object from its
707 srcobject->type = OBJT_DEFAULT;
713 * SWAP_PAGER_HASPAGE() - determine if we have good backing store for
714 * the requested page.
716 * We determine whether good backing store exists for the requested
717 * page and return TRUE if it does, FALSE if it doesn't.
719 * If TRUE, we also try to determine how much valid, contiguous backing
720 * store exists before and after the requested page within a reasonable
721 * distance. We do not try to restrict it to the swap device stripe
722 * (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 ?
739 blk0 = swp_pager_meta_ctl(object, pindex, 0);
741 if (blk0 == SWAPBLK_NONE) {
751 * find backwards-looking contiguous good backing store
753 if (before != NULL) {
756 for (i = 1; i < (SWB_NPAGES/2); ++i) {
761 blk = swp_pager_meta_ctl(object, pindex - i, 0);
769 * find forward-looking contiguous good backing store
774 for (i = 1; i < (SWB_NPAGES/2); ++i) {
777 blk = swp_pager_meta_ctl(object, pindex + i, 0);
788 * SWAP_PAGER_PAGE_UNSWAPPED() - remove swap backing store related to page
790 * This removes any associated swap backing store, whether valid or
791 * not, from the page.
793 * This routine is typically called when a page is made dirty, at
794 * which point any associated swap can be freed. MADV_FREE also
795 * calls us in a special-case situation
797 * NOTE!!! If the page is clean and the swap was valid, the caller
798 * should make the page dirty before calling this routine. This routine
799 * does NOT change the m->dirty status of the page. Also: MADV_FREE
802 * This routine may not block
803 * This routine must be called at splvm()
806 swap_pager_unswapped(m)
809 swp_pager_meta_ctl(m->object, m->pindex, SWM_FREE);
813 * SWAP_PAGER_STRATEGY() - read, write, free blocks
815 * This implements the vm_pager_strategy() interface to swap and allows
816 * other parts of the system to directly access swap as backing store
817 * through vm_objects of type OBJT_SWAP. This is intended to be a
818 * cacheless interface ( i.e. caching occurs at higher levels ).
819 * Therefore we do not maintain any resident pages. All I/O goes
820 * directly to and from the swap device.
822 * Note that b_blkno is scaled for PAGE_SIZE
824 * We currently attempt to run I/O synchronously or asynchronously as
825 * the caller requests. This isn't perfect because we loose error
826 * sequencing when we run multiple ops in parallel to satisfy a request.
827 * But this is swap, so we let it all hang out.
830 swap_pager_strategy(vm_object_t object, struct bio *bp)
836 struct buf *nbp = NULL;
840 /* XXX: KASSERT instead ? */
841 if (bp->bio_bcount & PAGE_MASK) {
842 biofinish(bp, NULL, EINVAL);
843 printf("swap_pager_strategy: bp %p blk %d size %d, not page bounded\n", bp, (int)bp->bio_pblkno, (int)bp->bio_bcount);
848 * Clear error indication, initialize page index, count, data pointer.
851 bp->bio_flags &= ~BIO_ERROR;
852 bp->bio_resid = bp->bio_bcount;
853 *(u_int *) &bp->bio_driver1 = 0;
855 start = bp->bio_pblkno;
856 count = howmany(bp->bio_bcount, PAGE_SIZE);
862 * Deal with BIO_DELETE
864 if (bp->bio_cmd == BIO_DELETE) {
866 * FREE PAGE(s) - destroy underlying swap that is no longer
869 swp_pager_meta_free(object, start, count);
877 * Execute read or write
883 * Obtain block. If block not found and writing, allocate a
884 * new block and build it into the object.
887 blk = swp_pager_meta_ctl(object, start, 0);
888 if ((blk == SWAPBLK_NONE) && (bp->bio_cmd == BIO_WRITE)) {
889 blk = swp_pager_getswapspace(1);
890 if (blk == SWAPBLK_NONE) {
891 bp->bio_error = ENOMEM;
892 bp->bio_flags |= BIO_ERROR;
895 swp_pager_meta_build(object, start, blk);
899 * Do we have to flush our current collection? Yes if:
901 * - no swap block at this index
902 * - swap block is not contiguous
903 * - we cross a physical disk boundry in the
907 nbp && (nbp->b_blkno + btoc(nbp->b_bcount) != blk ||
908 ((nbp->b_blkno ^ blk) & dmmax_mask)
912 if (bp->bio_cmd == BIO_READ) {
914 cnt.v_swappgsin += btoc(nbp->b_bcount);
917 cnt.v_swappgsout += btoc(nbp->b_bcount);
918 nbp->b_dirtyend = nbp->b_bcount;
926 * Add new swapblk to nbp, instantiating nbp if necessary.
927 * Zero-fill reads are able to take a shortcut.
929 if (blk == SWAPBLK_NONE) {
931 * We can only get here if we are reading. Since
932 * we are at splvm() we can safely modify b_resid,
933 * even if chain ops are in progress.
935 bzero(data, PAGE_SIZE);
936 bp->bio_resid -= PAGE_SIZE;
939 nbp = getchainbuf(bp, swapdev_vp, B_ASYNC);
944 nbp->b_bcount += PAGE_SIZE;
952 * Flush out last buffer
957 if (nbp->b_iocmd == BIO_READ) {
959 cnt.v_swappgsin += btoc(nbp->b_bcount);
962 cnt.v_swappgsout += btoc(nbp->b_bcount);
963 nbp->b_dirtyend = nbp->b_bcount;
969 * Wait for completion.
971 waitchainbuf(bp, 0, 1);
975 * SWAP_PAGER_GETPAGES() - bring pages in from swap
977 * Attempt to retrieve (m, count) pages from backing store, but make
978 * sure we retrieve at least m[reqpage]. We try to load in as large
979 * a chunk surrounding m[reqpage] as is contiguous in swap and which
980 * belongs to the same object.
982 * The code is designed for asynchronous operation and
983 * immediate-notification of 'reqpage' but tends not to be
984 * used that way. Please do not optimize-out this algorithmic
985 * feature, I intend to improve on it in the future.
987 * The parent has a single vm_object_pip_add() reference prior to
988 * calling us and we should return with the same.
990 * The parent has BUSY'd the pages. We should return with 'm'
991 * left busy, but the others adjusted.
994 swap_pager_getpages(object, m, count, reqpage)
1006 vm_pindex_t lastpindex;
1012 if (mreq->object != object) {
1013 panic("swap_pager_getpages: object mismatch %p/%p",
1019 * Calculate range to retrieve. The pages have already been assigned
1020 * their swapblks. We require a *contiguous* range that falls entirely
1021 * within a single device stripe. If we do not supply it, bad things
1022 * happen. Note that blk, iblk & jblk can be SWAPBLK_NONE, but the
1023 * loops are set up such that the case(s) are handled implicitly.
1025 * The swp_*() calls must be made at splvm(). vm_page_free() does
1026 * not need to be, but it will go a little faster if it is.
1029 blk = swp_pager_meta_ctl(mreq->object, mreq->pindex, 0);
1031 for (i = reqpage - 1; i >= 0; --i) {
1034 iblk = swp_pager_meta_ctl(m[i]->object, m[i]->pindex, 0);
1035 if (blk != iblk + (reqpage - i))
1037 if ((blk ^ iblk) & dmmax_mask)
1042 for (j = reqpage + 1; j < count; ++j) {
1045 jblk = swp_pager_meta_ctl(m[j]->object, m[j]->pindex, 0);
1046 if (blk != jblk - (j - reqpage))
1048 if ((blk ^ jblk) & dmmax_mask)
1053 * free pages outside our collection range. Note: we never free
1054 * mreq, it must remain busy throughout.
1059 for (k = 0; k < i; ++k)
1061 for (k = j; k < count; ++k)
1068 * Return VM_PAGER_FAIL if we have nothing to do. Return mreq
1069 * still busy, but the others unbusied.
1071 if (blk == SWAPBLK_NONE)
1072 return (VM_PAGER_FAIL);
1075 * Get a swap buffer header to perform the IO
1077 bp = getpbuf(&nsw_rcount);
1078 kva = (vm_offset_t) bp->b_data;
1081 * map our page(s) into kva for input
1083 * NOTE: B_PAGING is set by pbgetvp()
1085 pmap_qenter(kva, m + i, j - i);
1087 bp->b_iocmd = BIO_READ;
1088 bp->b_iodone = swp_pager_async_iodone;
1089 bp->b_rcred = crhold(thread0.td_ucred);
1090 bp->b_wcred = crhold(thread0.td_ucred);
1091 bp->b_data = (caddr_t) kva;
1092 bp->b_blkno = blk - (reqpage - i);
1093 bp->b_bcount = PAGE_SIZE * (j - i);
1094 bp->b_bufsize = PAGE_SIZE * (j - i);
1095 bp->b_pager.pg_reqpage = reqpage - i;
1100 for (k = i; k < j; ++k) {
1101 bp->b_pages[k - i] = m[k];
1102 vm_page_flag_set(m[k], PG_SWAPINPROG);
1105 bp->b_npages = j - i;
1107 pbgetvp(swapdev_vp, bp);
1110 cnt.v_swappgsin += bp->b_npages;
1113 * We still hold the lock on mreq, and our automatic completion routine
1114 * does not remove it.
1116 vm_object_pip_add(mreq->object, bp->b_npages);
1117 lastpindex = m[j-1]->pindex;
1120 * perform the I/O. NOTE!!! bp cannot be considered valid after
1121 * this point because we automatically release it on completion.
1122 * Instead, we look at the one page we are interested in which we
1123 * still hold a lock on even through the I/O completion.
1125 * The other pages in our m[] array are also released on completion,
1126 * so we cannot assume they are valid anymore either.
1128 * NOTE: b_blkno is destroyed by the call to VOP_STRATEGY
1134 * wait for the page we want to complete. PG_SWAPINPROG is always
1135 * cleared on completion. If an I/O error occurs, SWAPBLK_NONE
1136 * is set in the meta-data.
1139 while ((mreq->flags & PG_SWAPINPROG) != 0) {
1140 vm_page_flag_set(mreq, PG_WANTED | PG_REFERENCED);
1142 if (tsleep(mreq, PSWP, "swread", hz*20)) {
1144 "swap_pager: indefinite wait buffer: device:"
1145 " %s, blkno: %ld, size: %ld\n",
1146 devtoname(bp->b_dev), (long)bp->b_blkno,
1154 * mreq is left busied after completion, but all the other pages
1155 * are freed. If we had an unrecoverable read error the page will
1158 if (mreq->valid != VM_PAGE_BITS_ALL) {
1159 return (VM_PAGER_ERROR);
1161 return (VM_PAGER_OK);
1165 * A final note: in a low swap situation, we cannot deallocate swap
1166 * and mark a page dirty here because the caller is likely to mark
1167 * the page clean when we return, causing the page to possibly revert
1168 * to all-zero's later.
1173 * swap_pager_putpages:
1175 * Assign swap (if necessary) and initiate I/O on the specified pages.
1177 * We support both OBJT_DEFAULT and OBJT_SWAP objects. DEFAULT objects
1178 * are automatically converted to SWAP objects.
1180 * In a low memory situation we may block in VOP_STRATEGY(), but the new
1181 * vm_page reservation system coupled with properly written VFS devices
1182 * should ensure that no low-memory deadlock occurs. This is an area
1185 * The parent has N vm_object_pip_add() references prior to
1186 * calling us and will remove references for rtvals[] that are
1187 * not set to VM_PAGER_PEND. We need to remove the rest on I/O
1190 * The parent has soft-busy'd the pages it passes us and will unbusy
1191 * those whos rtvals[] entry is not set to VM_PAGER_PEND on return.
1192 * We need to unbusy the rest on I/O completion.
1195 swap_pager_putpages(object, m, count, sync, rtvals)
1206 if (count && m[0]->object != object) {
1207 panic("swap_pager_getpages: object mismatch %p/%p",
1215 * Turn object into OBJT_SWAP
1216 * check for bogus sysops
1217 * force sync if not pageout process
1219 if (object->type != OBJT_SWAP)
1220 swp_pager_meta_build(object, 0, SWAPBLK_NONE);
1222 if (curproc != pageproc)
1228 * Update nsw parameters from swap_async_max sysctl values.
1229 * Do not let the sysop crash the machine with bogus numbers.
1231 mtx_lock(&pbuf_mtx);
1232 if (swap_async_max != nsw_wcount_async_max) {
1239 if ((n = swap_async_max) > nswbuf / 2)
1246 * Adjust difference ( if possible ). If the current async
1247 * count is too low, we may not be able to make the adjustment
1251 n -= nsw_wcount_async_max;
1252 if (nsw_wcount_async + n >= 0) {
1253 nsw_wcount_async += n;
1254 nsw_wcount_async_max += n;
1255 wakeup(&nsw_wcount_async);
1259 mtx_unlock(&pbuf_mtx);
1264 * Assign swap blocks and issue I/O. We reallocate swap on the fly.
1265 * The page is left dirty until the pageout operation completes
1268 for (i = 0; i < count; i += n) {
1275 * Maximum I/O size is limited by a number of factors.
1277 n = min(BLIST_MAX_ALLOC, count - i);
1278 n = min(n, nsw_cluster_max);
1283 * Get biggest block of swap we can. If we fail, fall
1284 * back and try to allocate a smaller block. Don't go
1285 * overboard trying to allocate space if it would overly
1289 (blk = swp_pager_getswapspace(n)) == SWAPBLK_NONE &&
1294 if (blk == SWAPBLK_NONE) {
1295 for (j = 0; j < n; ++j)
1296 rtvals[i+j] = VM_PAGER_FAIL;
1302 * The I/O we are constructing cannot cross a physical
1303 * disk boundry in the swap stripe. Note: we are still
1306 if ((blk ^ (blk + n)) & dmmax_mask) {
1307 j = ((blk + dmmax) & dmmax_mask) - blk;
1308 swp_pager_freeswapspace(blk + j, n - j);
1313 * All I/O parameters have been satisfied, build the I/O
1314 * request and assign the swap space.
1316 * NOTE: B_PAGING is set by pbgetvp()
1319 bp = getpbuf(&nsw_wcount_sync);
1321 bp = getpbuf(&nsw_wcount_async);
1322 bp->b_flags = B_ASYNC;
1324 bp->b_iocmd = BIO_WRITE;
1325 bp->b_spc = NULL; /* not used, but NULL-out anyway */
1327 pmap_qenter((vm_offset_t)bp->b_data, &m[i], n);
1329 bp->b_rcred = crhold(thread0.td_ucred);
1330 bp->b_wcred = crhold(thread0.td_ucred);
1331 bp->b_bcount = PAGE_SIZE * n;
1332 bp->b_bufsize = PAGE_SIZE * n;
1335 pbgetvp(swapdev_vp, bp);
1337 for (j = 0; j < n; ++j) {
1338 vm_page_t mreq = m[i+j];
1340 swp_pager_meta_build(
1345 vm_page_dirty(mreq);
1346 rtvals[i+j] = VM_PAGER_OK;
1348 vm_page_flag_set(mreq, PG_SWAPINPROG);
1349 bp->b_pages[j] = mreq;
1353 * Must set dirty range for NFS to work.
1356 bp->b_dirtyend = bp->b_bcount;
1359 cnt.v_swappgsout += bp->b_npages;
1360 swapdev_vp->v_numoutput++;
1367 * NOTE: b_blkno is destroyed by the call to VOP_STRATEGY
1369 if (sync == FALSE) {
1370 bp->b_iodone = swp_pager_async_iodone;
1374 for (j = 0; j < n; ++j)
1375 rtvals[i+j] = VM_PAGER_PEND;
1376 /* restart outter loop */
1383 * NOTE: b_blkno is destroyed by the call to VOP_STRATEGY
1385 bp->b_iodone = swp_pager_sync_iodone;
1389 * Wait for the sync I/O to complete, then update rtvals.
1390 * We just set the rtvals[] to VM_PAGER_PEND so we can call
1391 * our async completion routine at the end, thus avoiding a
1395 while ((bp->b_flags & B_DONE) == 0) {
1396 tsleep(bp, PVM, "swwrt", 0);
1398 for (j = 0; j < n; ++j)
1399 rtvals[i+j] = VM_PAGER_PEND;
1401 * Now that we are through with the bp, we can call the
1402 * normal async completion, which frees everything up.
1404 swp_pager_async_iodone(bp);
1410 * swap_pager_sync_iodone:
1412 * Completion routine for synchronous reads and writes from/to swap.
1413 * We just mark the bp is complete and wake up anyone waiting on it.
1415 * This routine may not block. This routine is called at splbio() or better.
1418 swp_pager_sync_iodone(bp)
1421 bp->b_flags |= B_DONE;
1422 bp->b_flags &= ~B_ASYNC;
1427 * swp_pager_async_iodone:
1429 * Completion routine for asynchronous reads and writes from/to swap.
1430 * Also called manually by synchronous code to finish up a bp.
1432 * For READ operations, the pages are PG_BUSY'd. For WRITE operations,
1433 * the pages are vm_page_t->busy'd. For READ operations, we PG_BUSY
1434 * unbusy all pages except the 'main' request page. For WRITE
1435 * operations, we vm_page_t->busy'd unbusy all pages ( we can do this
1436 * because we marked them all VM_PAGER_PEND on return from putpages ).
1438 * This routine may not block.
1439 * This routine is called at splbio() or better
1441 * We up ourselves to splvm() as required for various vm_page related
1445 swp_pager_async_iodone(bp)
1450 vm_object_t object = NULL;
1453 bp->b_flags |= B_DONE;
1458 if (bp->b_ioflags & BIO_ERROR) {
1460 "swap_pager: I/O error - %s failed; blkno %ld,"
1461 "size %ld, error %d\n",
1462 ((bp->b_iocmd == BIO_READ) ? "pagein" : "pageout"),
1470 * set object, raise to splvm().
1473 object = bp->b_pages[0]->object;
1477 * remove the mapping for kernel virtual
1479 pmap_qremove((vm_offset_t)bp->b_data, bp->b_npages);
1482 * cleanup pages. If an error occurs writing to swap, we are in
1483 * very serious trouble. If it happens to be a disk error, though,
1484 * we may be able to recover by reassigning the swap later on. So
1485 * in this case we remove the m->swapblk assignment for the page
1486 * but do not free it in the rlist. The errornous block(s) are thus
1487 * never reallocated as swap. Redirty the page and continue.
1489 for (i = 0; i < bp->b_npages; ++i) {
1490 vm_page_t m = bp->b_pages[i];
1492 vm_page_flag_clear(m, PG_SWAPINPROG);
1494 if (bp->b_ioflags & BIO_ERROR) {
1496 * If an error occurs I'd love to throw the swapblk
1497 * away without freeing it back to swapspace, so it
1498 * can never be used again. But I can't from an
1501 if (bp->b_iocmd == BIO_READ) {
1503 * When reading, reqpage needs to stay
1504 * locked for the parent, but all other
1505 * pages can be freed. We still want to
1506 * wakeup the parent waiting on the page,
1507 * though. ( also: pg_reqpage can be -1 and
1508 * not match anything ).
1510 * We have to wake specifically requested pages
1511 * up too because we cleared PG_SWAPINPROG and
1512 * someone may be waiting for that.
1514 * NOTE: for reads, m->dirty will probably
1515 * be overridden by the original caller of
1516 * getpages so don't play cute tricks here.
1518 * XXX IT IS NOT LEGAL TO FREE THE PAGE HERE
1519 * AS THIS MESSES WITH object->memq, and it is
1520 * not legal to mess with object->memq from an
1524 vm_page_flag_clear(m, PG_ZERO);
1525 if (i != bp->b_pager.pg_reqpage)
1530 * If i == bp->b_pager.pg_reqpage, do not wake
1531 * the page up. The caller needs to.
1535 * If a write error occurs, reactivate page
1536 * so it doesn't clog the inactive list,
1537 * then finish the I/O.
1540 vm_page_activate(m);
1541 vm_page_io_finish(m);
1543 } else if (bp->b_iocmd == BIO_READ) {
1545 * For read success, clear dirty bits. Nobody should
1546 * have this page mapped but don't take any chances,
1547 * make sure the pmap modify bits are also cleared.
1549 * NOTE: for reads, m->dirty will probably be
1550 * overridden by the original caller of getpages so
1551 * we cannot set them in order to free the underlying
1552 * swap in a low-swap situation. I don't think we'd
1553 * want to do that anyway, but it was an optimization
1554 * that existed in the old swapper for a time before
1555 * it got ripped out due to precisely this problem.
1557 * clear PG_ZERO in page.
1559 * If not the requested page then deactivate it.
1561 * Note that the requested page, reqpage, is left
1562 * busied, but we still have to wake it up. The
1563 * other pages are released (unbusied) by
1564 * vm_page_wakeup(). We do not set reqpage's
1565 * valid bits here, it is up to the caller.
1567 pmap_clear_modify(m);
1568 m->valid = VM_PAGE_BITS_ALL;
1570 vm_page_flag_clear(m, PG_ZERO);
1573 * We have to wake specifically requested pages
1574 * up too because we cleared PG_SWAPINPROG and
1575 * could be waiting for it in getpages. However,
1576 * be sure to not unbusy getpages specifically
1577 * requested page - getpages expects it to be
1580 if (i != bp->b_pager.pg_reqpage) {
1581 vm_page_deactivate(m);
1588 * For write success, clear the modify and dirty
1589 * status, then finish the I/O ( which decrements the
1590 * busy count and possibly wakes waiter's up ).
1592 pmap_clear_modify(m);
1594 vm_page_io_finish(m);
1595 if (!vm_page_count_severe() || !vm_page_try_to_cache(m))
1596 vm_page_protect(m, VM_PROT_READ);
1601 * adjust pip. NOTE: the original parent may still have its own
1602 * pip refs on the object.
1605 vm_object_pip_wakeupn(object, bp->b_npages);
1608 * release the physical I/O buffer
1612 ((bp->b_iocmd == BIO_READ) ? &nsw_rcount :
1613 ((bp->b_flags & B_ASYNC) ?
1622 /************************************************************************
1624 ************************************************************************
1626 * These routines manipulate the swap metadata stored in the
1627 * OBJT_SWAP object. All swp_*() routines must be called at
1628 * splvm() because swap can be freed up by the low level vm_page
1629 * code which might be called from interrupts beyond what splbio() covers.
1631 * Swap metadata is implemented with a global hash and not directly
1632 * linked into the object. Instead the object simply contains
1633 * appropriate tracking counters.
1637 * SWP_PAGER_HASH() - hash swap meta data
1639 * This is an inline helper function which hashes the swapblk given
1640 * the object and page index. It returns a pointer to a pointer
1641 * to the object, or a pointer to a NULL pointer if it could not
1644 * This routine must be called at splvm().
1646 static __inline struct swblock **
1647 swp_pager_hash(vm_object_t object, vm_pindex_t index)
1649 struct swblock **pswap;
1650 struct swblock *swap;
1652 index &= ~(vm_pindex_t)SWAP_META_MASK;
1653 pswap = &swhash[(index ^ (int)(intptr_t)object) & swhash_mask];
1654 while ((swap = *pswap) != NULL) {
1655 if (swap->swb_object == object &&
1656 swap->swb_index == index
1660 pswap = &swap->swb_hnext;
1666 * SWP_PAGER_META_BUILD() - add swap block to swap meta data for object
1668 * We first convert the object to a swap object if it is a default
1671 * The specified swapblk is added to the object's swap metadata. If
1672 * the swapblk is not valid, it is freed instead. Any previously
1673 * assigned swapblk is freed.
1675 * This routine must be called at splvm(), except when used to convert
1676 * an OBJT_DEFAULT object into an OBJT_SWAP object.
1679 swp_pager_meta_build(
1684 struct swblock *swap;
1685 struct swblock **pswap;
1690 * Convert default object to swap object if necessary
1692 if (object->type != OBJT_SWAP) {
1693 object->type = OBJT_SWAP;
1694 object->un_pager.swp.swp_bcount = 0;
1696 mtx_lock(&sw_alloc_mtx);
1697 if (object->handle != NULL) {
1699 NOBJLIST(object->handle),
1705 &swap_pager_un_object_list,
1710 mtx_unlock(&sw_alloc_mtx);
1714 * Locate hash entry. If not found create, but if we aren't adding
1715 * anything just return. If we run out of space in the map we wait
1716 * and, since the hash table may have changed, retry.
1719 pswap = swp_pager_hash(object, pindex);
1721 if ((swap = *pswap) == NULL) {
1724 if (swapblk == SWAPBLK_NONE)
1727 swap = *pswap = uma_zalloc(swap_zone, M_NOWAIT);
1733 swap->swb_hnext = NULL;
1734 swap->swb_object = object;
1735 swap->swb_index = pindex & ~(vm_pindex_t)SWAP_META_MASK;
1736 swap->swb_count = 0;
1738 ++object->un_pager.swp.swp_bcount;
1740 for (i = 0; i < SWAP_META_PAGES; ++i)
1741 swap->swb_pages[i] = SWAPBLK_NONE;
1745 * Delete prior contents of metadata
1747 idx = pindex & SWAP_META_MASK;
1749 if (swap->swb_pages[idx] != SWAPBLK_NONE) {
1750 swp_pager_freeswapspace(swap->swb_pages[idx], 1);
1755 * Enter block into metadata
1757 swap->swb_pages[idx] = swapblk;
1758 if (swapblk != SWAPBLK_NONE)
1763 * SWP_PAGER_META_FREE() - free a range of blocks in the object's swap metadata
1765 * The requested range of blocks is freed, with any associated swap
1766 * returned to the swap bitmap.
1768 * This routine will free swap metadata structures as they are cleaned
1769 * out. This routine does *NOT* operate on swap metadata associated
1770 * with resident pages.
1772 * This routine must be called at splvm()
1775 swp_pager_meta_free(vm_object_t object, vm_pindex_t index, daddr_t count)
1779 if (object->type != OBJT_SWAP)
1783 struct swblock **pswap;
1784 struct swblock *swap;
1786 pswap = swp_pager_hash(object, index);
1788 if ((swap = *pswap) != NULL) {
1789 daddr_t v = swap->swb_pages[index & SWAP_META_MASK];
1791 if (v != SWAPBLK_NONE) {
1792 swp_pager_freeswapspace(v, 1);
1793 swap->swb_pages[index & SWAP_META_MASK] =
1795 if (--swap->swb_count == 0) {
1796 *pswap = swap->swb_hnext;
1797 uma_zfree(swap_zone, swap);
1798 --object->un_pager.swp.swp_bcount;
1804 int n = SWAP_META_PAGES - (index & SWAP_META_MASK);
1812 * SWP_PAGER_META_FREE_ALL() - destroy all swap metadata associated with object
1814 * This routine locates and destroys all swap metadata associated with
1817 * This routine must be called at splvm()
1820 swp_pager_meta_free_all(vm_object_t object)
1826 if (object->type != OBJT_SWAP)
1829 while (object->un_pager.swp.swp_bcount) {
1830 struct swblock **pswap;
1831 struct swblock *swap;
1833 pswap = swp_pager_hash(object, index);
1834 if ((swap = *pswap) != NULL) {
1837 for (i = 0; i < SWAP_META_PAGES; ++i) {
1838 daddr_t v = swap->swb_pages[i];
1839 if (v != SWAPBLK_NONE) {
1841 swp_pager_freeswapspace(v, 1);
1844 if (swap->swb_count != 0)
1845 panic("swap_pager_meta_free_all: swb_count != 0");
1846 *pswap = swap->swb_hnext;
1847 uma_zfree(swap_zone, swap);
1848 --object->un_pager.swp.swp_bcount;
1850 index += SWAP_META_PAGES;
1851 if (index > 0x20000000)
1852 panic("swp_pager_meta_free_all: failed to locate all swap meta blocks");
1857 * SWP_PAGER_METACTL() - misc control of swap and vm_page_t meta data.
1859 * This routine is capable of looking up, popping, or freeing
1860 * swapblk assignments in the swap meta data or in the vm_page_t.
1861 * The routine typically returns the swapblk being looked-up, or popped,
1862 * or SWAPBLK_NONE if the block was freed, or SWAPBLK_NONE if the block
1863 * was invalid. This routine will automatically free any invalid
1864 * meta-data swapblks.
1866 * It is not possible to store invalid swapblks in the swap meta data
1867 * (other then a literal 'SWAPBLK_NONE'), so we don't bother checking.
1869 * When acting on a busy resident page and paging is in progress, we
1870 * have to wait until paging is complete but otherwise can act on the
1873 * This routine must be called at splvm().
1875 * SWM_FREE remove and free swap block from metadata
1876 * SWM_POP remove from meta data but do not free.. pop it out
1884 struct swblock **pswap;
1885 struct swblock *swap;
1891 * The meta data only exists of the object is OBJT_SWAP
1892 * and even then might not be allocated yet.
1894 if (object->type != OBJT_SWAP)
1895 return (SWAPBLK_NONE);
1898 pswap = swp_pager_hash(object, pindex);
1900 if ((swap = *pswap) != NULL) {
1901 idx = pindex & SWAP_META_MASK;
1902 r1 = swap->swb_pages[idx];
1904 if (r1 != SWAPBLK_NONE) {
1905 if (flags & SWM_FREE) {
1906 swp_pager_freeswapspace(r1, 1);
1909 if (flags & (SWM_FREE|SWM_POP)) {
1910 swap->swb_pages[idx] = SWAPBLK_NONE;
1911 if (--swap->swb_count == 0) {
1912 *pswap = swap->swb_hnext;
1913 uma_zfree(swap_zone, swap);
1914 --object->un_pager.swp.swp_bcount;
1922 /********************************************************
1923 * CHAINING FUNCTIONS *
1924 ********************************************************
1926 * These functions support recursion of I/O operations
1927 * on bp's, typically by chaining one or more 'child' bp's
1928 * to the parent. Synchronous, asynchronous, and semi-synchronous
1929 * chaining is possible.
1933 * vm_pager_chain_iodone:
1935 * io completion routine for child bp. Currently we fudge a bit
1936 * on dealing with b_resid. Since users of these routines may issue
1937 * multiple children simultaneously, sequencing of the error can be lost.
1940 vm_pager_chain_iodone(struct buf *nbp)
1945 bp = nbp->b_caller1;
1946 count = (u_int *)&(bp->bio_driver1);
1948 if (nbp->b_ioflags & BIO_ERROR) {
1949 bp->bio_flags |= BIO_ERROR;
1950 bp->bio_error = nbp->b_error;
1951 } else if (nbp->b_resid != 0) {
1952 bp->bio_flags |= BIO_ERROR;
1953 bp->bio_error = EINVAL;
1955 bp->bio_resid -= nbp->b_bcount;
1957 nbp->b_caller1 = NULL;
1959 if (bp->bio_flags & BIO_FLAG1) {
1960 bp->bio_flags &= ~BIO_FLAG1;
1964 nbp->b_flags |= B_DONE;
1965 nbp->b_flags &= ~B_ASYNC;
1972 * Obtain a physical buffer and chain it to its parent buffer. When
1973 * I/O completes, the parent buffer will be B_SIGNAL'd. Errors are
1974 * automatically propagated to the parent
1977 getchainbuf(struct bio *bp, struct vnode *vp, int flags)
1983 nbp = getpbuf(NULL);
1984 count = (u_int *)&(bp->bio_driver1);
1986 nbp->b_caller1 = bp;
1990 waitchainbuf(bp, 4, 0);
1992 nbp->b_iocmd = bp->bio_cmd;
1994 nbp->b_flags = flags;
1995 nbp->b_rcred = crhold(thread0.td_ucred);
1996 nbp->b_wcred = crhold(thread0.td_ucred);
1997 nbp->b_iodone = vm_pager_chain_iodone;
2005 flushchainbuf(struct buf *nbp)
2008 if (nbp->b_bcount) {
2009 nbp->b_bufsize = nbp->b_bcount;
2010 if (nbp->b_iocmd == BIO_WRITE)
2011 nbp->b_dirtyend = nbp->b_bcount;
2020 waitchainbuf(struct bio *bp, int limit, int done)
2026 count = (u_int *)&(bp->bio_driver1);
2028 while (*count > limit) {
2029 bp->bio_flags |= BIO_FLAG1;
2030 tsleep(bp, PRIBIO + 4, "bpchain", 0);
2033 if (bp->bio_resid != 0 && !(bp->bio_flags & BIO_ERROR)) {
2034 bp->bio_flags |= BIO_ERROR;
2035 bp->bio_error = EINVAL;