2 * Copyright (c) 1998,2004 The DragonFly Project. All rights reserved.
4 * This code is derived from software contributed to The DragonFly Project
5 * by Matthew Dillon <dillon@backplane.com>
7 * Redistribution and use in source and binary forms, with or without
8 * modification, are permitted provided that the following conditions
11 * 1. Redistributions of source code must retain the above copyright
12 * notice, this list of conditions and the following disclaimer.
13 * 2. Redistributions in binary form must reproduce the above copyright
14 * notice, this list of conditions and the following disclaimer in
15 * the documentation and/or other materials provided with the
17 * 3. Neither the name of The DragonFly Project nor the names of its
18 * contributors may be used to endorse or promote products derived
19 * from this software without specific, prior written permission.
21 * THIS SOFTWARE IS PROVIDED BY THE COPYRIGHT HOLDERS AND CONTRIBUTORS
22 * ``AS IS'' AND ANY EXPRESS OR IMPLIED WARRANTIES, INCLUDING, BUT NOT
23 * LIMITED TO, THE IMPLIED WARRANTIES OF MERCHANTABILITY AND FITNESS
24 * FOR A PARTICULAR PURPOSE ARE DISCLAIMED. IN NO EVENT SHALL THE
25 * COPYRIGHT HOLDERS OR CONTRIBUTORS BE LIABLE FOR ANY DIRECT, INDIRECT,
26 * INCIDENTAL, SPECIAL, EXEMPLARY OR CONSEQUENTIAL DAMAGES (INCLUDING,
27 * BUT NOT LIMITED TO, PROCUREMENT OF SUBSTITUTE GOODS OR SERVICES;
28 * LOSS OF USE, DATA, OR PROFITS; OR BUSINESS INTERRUPTION) HOWEVER CAUSED
29 * AND ON ANY THEORY OF LIABILITY, WHETHER IN CONTRACT, STRICT LIABILITY,
30 * OR TORT (INCLUDING NEGLIGENCE OR OTHERWISE) ARISING IN ANY WAY OUT
31 * OF THE USE OF THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF
34 * Copyright (c) 1994 John S. Dyson
35 * Copyright (c) 1990 University of Utah.
36 * Copyright (c) 1991, 1993
37 * The Regents of the University of California. All rights reserved.
39 * This code is derived from software contributed to Berkeley by
40 * the Systems Programming Group of the University of Utah Computer
43 * Redistribution and use in source and binary forms, with or without
44 * modification, are permitted provided that the following conditions
46 * 1. Redistributions of source code must retain the above copyright
47 * notice, this list of conditions and the following disclaimer.
48 * 2. Redistributions in binary form must reproduce the above copyright
49 * notice, this list of conditions and the following disclaimer in the
50 * documentation and/or other materials provided with the distribution.
51 * 3. All advertising materials mentioning features or use of this software
52 * must display the following acknowledgement:
53 * This product includes software developed by the University of
54 * California, Berkeley and its contributors.
55 * 4. Neither the name of the University nor the names of its contributors
56 * may be used to endorse or promote products derived from this software
57 * without specific prior written permission.
59 * THIS SOFTWARE IS PROVIDED BY THE REGENTS AND CONTRIBUTORS ``AS IS'' AND
60 * ANY EXPRESS OR IMPLIED WARRANTIES, INCLUDING, BUT NOT LIMITED TO, THE
61 * IMPLIED WARRANTIES OF MERCHANTABILITY AND FITNESS FOR A PARTICULAR PURPOSE
62 * ARE DISCLAIMED. IN NO EVENT SHALL THE REGENTS OR CONTRIBUTORS BE LIABLE
63 * FOR ANY DIRECT, INDIRECT, INCIDENTAL, SPECIAL, EXEMPLARY, OR CONSEQUENTIAL
64 * DAMAGES (INCLUDING, BUT NOT LIMITED TO, PROCUREMENT OF SUBSTITUTE GOODS
65 * OR SERVICES; LOSS OF USE, DATA, OR PROFITS; OR BUSINESS INTERRUPTION)
66 * HOWEVER CAUSED AND ON ANY THEORY OF LIABILITY, WHETHER IN CONTRACT, STRICT
67 * LIABILITY, OR TORT (INCLUDING NEGLIGENCE OR OTHERWISE) ARISING IN ANY WAY
68 * OUT OF THE USE OF THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF
74 * Radix Bitmap 'blists'.
76 * - The new swapper uses the new radix bitmap code. This should scale
77 * to arbitrarily small or arbitrarily large swap spaces and an almost
78 * arbitrary degree of fragmentation.
82 * - on the fly reallocation of swap during putpages. The new system
83 * does not try to keep previously allocated swap blocks for dirty
86 * - on the fly deallocation of swap
88 * - No more garbage collection required. Unnecessarily allocated swap
89 * blocks only exist for dirty vm_page_t's now and these are already
90 * cycled (in a high-load system) by the pager. We also do on-the-fly
91 * removal of invalidated swap blocks when a page is destroyed
94 * from: Utah $Hdr: swap_pager.c 1.4 91/04/30$
96 * @(#)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 $
99 * $DragonFly: src/sys/vm/swap_pager.c,v 1.32 2008/07/01 02:02:56 dillon Exp $
102 #include <sys/param.h>
103 #include <sys/systm.h>
104 #include <sys/conf.h>
105 #include <sys/kernel.h>
106 #include <sys/proc.h>
108 #include <sys/vnode.h>
109 #include <sys/malloc.h>
110 #include <sys/vmmeter.h>
111 #include <sys/sysctl.h>
112 #include <sys/blist.h>
113 #include <sys/lock.h>
114 #include <sys/thread2.h>
116 #ifndef MAX_PAGEOUT_CLUSTER
117 #define MAX_PAGEOUT_CLUSTER 16
120 #define SWB_NPAGES MAX_PAGEOUT_CLUSTER
122 #include "opt_swap.h"
124 #include <vm/vm_object.h>
125 #include <vm/vm_page.h>
126 #include <vm/vm_pager.h>
127 #include <vm/vm_pageout.h>
128 #include <vm/swap_pager.h>
129 #include <vm/vm_extern.h>
130 #include <vm/vm_zone.h>
131 #include <vm/vnode_pager.h>
133 #include <sys/buf2.h>
134 #include <vm/vm_page2.h>
136 #define SWM_FREE 0x02 /* free, period */
137 #define SWM_POP 0x04 /* pop out */
139 #define SWBIO_READ 0x01
140 #define SWBIO_WRITE 0x02
141 #define SWBIO_SYNC 0x04
147 vm_pindex_t endi; /* inclusive */
151 * vm_swap_size is in page-sized chunks now. It was DEV_BSIZE'd chunks
155 extern int vm_swap_size; /* number of free swap blocks, in pages */
157 int swap_pager_full; /* swap space exhaustion (task killing) */
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");
179 * Red-Black tree for swblock entries
181 RB_GENERATE2(swblock_rb_tree, swblock, swb_entry, rb_swblock_compare,
182 vm_pindex_t, swb_index);
185 rb_swblock_compare(struct swblock *swb1, struct swblock *swb2)
187 if (swb1->swb_index < swb2->swb_index)
189 if (swb1->swb_index > swb2->swb_index)
196 rb_swblock_scancmp(struct swblock *swb, void *data)
198 struct swfreeinfo *info = data;
200 if (swb->swb_index < info->basei)
202 if (swb->swb_index > info->endi)
208 * pagerops for OBJT_SWAP - "swap pager". Some ops are also global procedure
209 * calls hooked from other parts of the VM system and do not appear here.
210 * (see vm/swap_pager.h).
214 swap_pager_alloc (void *handle, off_t size,
215 vm_prot_t prot, off_t offset);
216 static void swap_pager_dealloc (vm_object_t object);
217 static int swap_pager_getpage (vm_object_t, vm_page_t *, int);
218 static void swap_chain_iodone(struct bio *biox);
220 struct pagerops swappagerops = {
221 swap_pager_alloc, /* allocate an OBJT_SWAP object */
222 swap_pager_dealloc, /* deallocate an OBJT_SWAP object */
223 swap_pager_getpage, /* pagein */
224 swap_pager_putpages, /* pageout */
225 swap_pager_haspage /* get backing store status for page */
229 * dmmax is in page-sized chunks with the new swap system. It was
230 * dev-bsized chunks in the old. dmmax is always a power of 2.
232 * swap_*() routines are externally accessible. swp_*() routines are
237 static int dmmax_mask;
238 int nswap_lowat = 128; /* in pages, swap_pager_almost_full warn */
239 int nswap_hiwat = 512; /* in pages, swap_pager_almost_full warn */
241 static __inline void swp_sizecheck (void);
242 static void swp_pager_async_iodone (struct bio *bio);
245 * Swap bitmap functions
248 static __inline void swp_pager_freeswapspace (daddr_t blk, int npages);
249 static __inline daddr_t swp_pager_getswapspace (int npages);
255 static void swp_pager_meta_convert (vm_object_t);
256 static void swp_pager_meta_build (vm_object_t, vm_pindex_t, daddr_t);
257 static void swp_pager_meta_free (vm_object_t, vm_pindex_t, vm_pindex_t);
258 static void swp_pager_meta_free_all (vm_object_t);
259 static daddr_t swp_pager_meta_ctl (vm_object_t, vm_pindex_t, int);
262 * SWP_SIZECHECK() - update swap_pager_full indication
264 * update the swap_pager_almost_full indication and warn when we are
265 * about to run out of swap space, using lowat/hiwat hysteresis.
267 * Clear swap_pager_full ( task killing ) indication when lowat is met.
269 * No restrictions on call
270 * This routine may not block.
271 * This routine must be called at splvm()
277 if (vm_swap_size < nswap_lowat) {
278 if (swap_pager_almost_full == 0) {
279 kprintf("swap_pager: out of swap space\n");
280 swap_pager_almost_full = 1;
284 if (vm_swap_size > nswap_hiwat)
285 swap_pager_almost_full = 0;
290 * SWAP_PAGER_INIT() - initialize the swap pager!
292 * Expected to be started from system init. NOTE: This code is run
293 * before much else so be careful what you depend on. Most of the VM
294 * system has yet to be initialized at this point.
297 swap_pager_init(void *arg __unused)
300 * Device Stripe, in PAGE_SIZE'd blocks
302 dmmax = SWB_NPAGES * 2;
303 dmmax_mask = ~(dmmax - 1);
305 SYSINIT(vm_mem, SI_BOOT1_VM, SI_ORDER_THIRD, swap_pager_init, NULL)
308 * SWAP_PAGER_SWAP_INIT() - swap pager initialization from pageout process
310 * Expected to be started from pageout process once, prior to entering
315 swap_pager_swap_init(void)
320 * Number of in-transit swap bp operations. Don't
321 * exhaust the pbufs completely. Make sure we
322 * initialize workable values (0 will work for hysteresis
323 * but it isn't very efficient).
325 * The nsw_cluster_max is constrained by the number of pages an XIO
326 * holds, i.e., (MAXPHYS/PAGE_SIZE) and our locally defined
327 * MAX_PAGEOUT_CLUSTER. Also be aware that swap ops are
328 * constrained by the swap device interleave stripe size.
330 * Currently we hardwire nsw_wcount_async to 4. This limit is
331 * designed to prevent other I/O from having high latencies due to
332 * our pageout I/O. The value 4 works well for one or two active swap
333 * devices but is probably a little low if you have more. Even so,
334 * a higher value would probably generate only a limited improvement
335 * with three or four active swap devices since the system does not
336 * typically have to pageout at extreme bandwidths. We will want
337 * at least 2 per swap devices, and 4 is a pretty good value if you
338 * have one NFS swap device due to the command/ack latency over NFS.
339 * So it all works out pretty well.
342 nsw_cluster_max = min((MAXPHYS/PAGE_SIZE), MAX_PAGEOUT_CLUSTER);
344 nsw_rcount = (nswbuf + 1) / 2;
345 nsw_wcount_sync = (nswbuf + 3) / 4;
346 nsw_wcount_async = 4;
347 nsw_wcount_async_max = nsw_wcount_async;
350 * The zone is dynamically allocated so generally size it to
351 * maxswzone (32MB to 512MB of KVM). Set a minimum size based
352 * on physical memory of around 8x (each swblock can hold 16 pages).
354 * With the advent of SSDs (vs HDs) the practical (swap:memory) ratio
355 * has increased dramatically.
357 n = vmstats.v_page_count / 2;
358 if (maxswzone && n < maxswzone / sizeof(struct swblock))
359 n = maxswzone / sizeof(struct swblock);
365 sizeof(struct swblock),
369 if (swap_zone != NULL)
372 * if the allocation failed, try a zone two thirds the
373 * size of the previous attempt.
378 if (swap_zone == NULL)
379 panic("swap_pager_swap_init: swap_zone == NULL");
381 kprintf("Swap zone entries reduced from %d to %d.\n", n2, n);
385 * SWAP_PAGER_ALLOC() - allocate a new OBJT_SWAP VM object and instantiate
386 * its metadata structures.
388 * This routine is called from the mmap and fork code to create a new
389 * OBJT_SWAP object. We do this by creating an OBJT_DEFAULT object
390 * and then converting it with swp_pager_meta_convert().
392 * This routine may block in vm_object_allocate() and create a named
393 * object lookup race, so we must interlock. We must also run at
394 * splvm() for the object lookup to handle races with interrupts, but
395 * we do not have to maintain splvm() in between the lookup and the
396 * add because (I believe) it is not possible to attempt to create
397 * a new swap object w/handle when a default object with that handle
402 swap_pager_alloc(void *handle, off_t size, vm_prot_t prot, off_t offset)
406 KKASSERT(handle == NULL);
410 * Reference existing named region or allocate new one. There
411 * should not be a race here against swp_pager_meta_build()
412 * as called from vm_page_remove() in regards to the lookup
415 while (sw_alloc_interlock) {
416 sw_alloc_interlock = -1;
417 tsleep(&sw_alloc_interlock, 0, "swpalc", 0);
419 sw_alloc_interlock = 1;
421 object = vm_pager_object_lookup(NOBJLIST(handle), handle);
423 if (object != NULL) {
424 vm_object_reference(object);
426 object = vm_object_allocate(OBJT_DEFAULT,
427 OFF_TO_IDX(offset + PAGE_MASK + size));
428 object->handle = handle;
429 swp_pager_meta_convert(object);
432 if (sw_alloc_interlock < 0)
433 wakeup(&sw_alloc_interlock);
434 sw_alloc_interlock = 0;
437 object = vm_object_allocate(OBJT_DEFAULT,
438 OFF_TO_IDX(offset + PAGE_MASK + size));
439 swp_pager_meta_convert(object);
445 * SWAP_PAGER_DEALLOC() - remove swap metadata from object
447 * The swap backing for the object is destroyed. The code is
448 * designed such that we can reinstantiate it later, but this
449 * routine is typically called only when the entire object is
450 * about to be destroyed.
452 * This routine may block, but no longer does.
454 * The object must be locked or unreferenceable.
458 swap_pager_dealloc(vm_object_t object)
460 vm_object_pip_wait(object, "swpdea");
463 * Free all remaining metadata. We only bother to free it from
464 * the swap meta data. We do not attempt to free swapblk's still
465 * associated with vm_page_t's for this object. We do not care
466 * if paging is still in progress on some objects.
469 swp_pager_meta_free_all(object);
473 /************************************************************************
474 * SWAP PAGER BITMAP ROUTINES *
475 ************************************************************************/
478 * SWP_PAGER_GETSWAPSPACE() - allocate raw swap space
480 * Allocate swap for the requested number of pages. The starting
481 * swap block number (a page index) is returned or SWAPBLK_NONE
482 * if the allocation failed.
484 * Also has the side effect of advising that somebody made a mistake
485 * when they configured swap and didn't configure enough.
487 * Must be called at splvm() to avoid races with bitmap frees from
488 * vm_page_remove() aka swap_pager_page_removed().
490 * This routine may not block
491 * This routine must be called at splvm().
494 static __inline daddr_t
495 swp_pager_getswapspace(int npages)
499 if ((blk = blist_alloc(swapblist, npages)) == SWAPBLK_NONE) {
500 if (swap_pager_full != 2) {
501 kprintf("swap_pager_getswapspace: failed\n");
503 swap_pager_almost_full = 1;
506 vm_swap_size -= npages;
513 * SWP_PAGER_FREESWAPSPACE() - free raw swap space
515 * This routine returns the specified swap blocks back to the bitmap.
517 * Note: This routine may not block (it could in the old swap code),
518 * and through the use of the new blist routines it does not block.
520 * We must be called at splvm() to avoid races with bitmap frees from
521 * vm_page_remove() aka swap_pager_page_removed().
523 * This routine may not block
524 * This routine must be called at splvm().
528 swp_pager_freeswapspace(daddr_t blk, int npages)
530 blist_free(swapblist, blk, npages);
531 vm_swap_size += npages;
536 * SWAP_PAGER_FREESPACE() - frees swap blocks associated with a page
537 * range within an object.
539 * This is a globally accessible routine.
541 * This routine removes swapblk assignments from swap metadata.
543 * The external callers of this routine typically have already destroyed
544 * or renamed vm_page_t's associated with this range in the object so
547 * This routine may be called at any spl. We up our spl to splvm
548 * temporarily in order to perform the metadata removal.
551 swap_pager_freespace(vm_object_t object, vm_pindex_t start, vm_pindex_t size)
554 swp_pager_meta_free(object, start, size);
559 swap_pager_freespace_all(vm_object_t object)
562 swp_pager_meta_free_all(object);
567 * SWAP_PAGER_RESERVE() - reserve swap blocks in object
569 * Assigns swap blocks to the specified range within the object. The
570 * swap blocks are not zerod. Any previous swap assignment is destroyed.
572 * Returns 0 on success, -1 on failure.
575 swap_pager_reserve(vm_object_t object, vm_pindex_t start, vm_size_t size)
578 daddr_t blk = SWAPBLK_NONE;
579 vm_pindex_t beg = start; /* save start index */
585 while ((blk = swp_pager_getswapspace(n)) == SWAPBLK_NONE) {
588 swp_pager_meta_free(object, beg,
595 swp_pager_meta_build(object, start, blk);
601 swp_pager_meta_free(object, start, n);
607 * SWAP_PAGER_COPY() - copy blocks from source pager to destination pager
608 * and destroy the source.
610 * Copy any valid swapblks from the source to the destination. In
611 * cases where both the source and destination have a valid swapblk,
612 * we keep the destination's.
614 * This routine is allowed to block. It may block allocating metadata
615 * indirectly through swp_pager_meta_build() or if paging is still in
616 * progress on the source.
618 * This routine can be called at any spl
620 * XXX vm_page_collapse() kinda expects us not to block because we
621 * supposedly do not need to allocate memory, but for the moment we
622 * *may* have to get a little memory from the zone allocator, but
623 * it is taken from the interrupt memory. We should be ok.
625 * The source object contains no vm_page_t's (which is just as well)
627 * The source object is of type OBJT_SWAP.
629 * The source and destination objects must be locked or
630 * inaccessible (XXX are they ?)
634 swap_pager_copy(vm_object_t srcobject, vm_object_t dstobject,
635 vm_pindex_t base_index, int destroysource)
642 * transfer source to destination.
644 for (i = 0; i < dstobject->size; ++i) {
648 * Locate (without changing) the swapblk on the destination,
649 * unless it is invalid in which case free it silently, or
650 * if the destination is a resident page, in which case the
651 * source is thrown away.
653 dstaddr = swp_pager_meta_ctl(dstobject, i, 0);
655 if (dstaddr == SWAPBLK_NONE) {
657 * Destination has no swapblk and is not resident,
662 srcaddr = swp_pager_meta_ctl(srcobject,
663 base_index + i, SWM_POP);
665 if (srcaddr != SWAPBLK_NONE)
666 swp_pager_meta_build(dstobject, i, srcaddr);
669 * Destination has valid swapblk or it is represented
670 * by a resident page. We destroy the sourceblock.
672 swp_pager_meta_ctl(srcobject, base_index + i, SWM_FREE);
677 * Free left over swap blocks in source.
679 * We have to revert the type to OBJT_DEFAULT so we do not accidently
680 * double-remove the object from the swap queues.
684 * Reverting the type is not necessary, the caller is going
685 * to destroy srcobject directly, but I'm doing it here
686 * for consistency since we've removed the object from its
689 swp_pager_meta_free_all(srcobject);
690 if (srcobject->type == OBJT_SWAP)
691 srcobject->type = OBJT_DEFAULT;
697 * SWAP_PAGER_HASPAGE() - determine if we have good backing store for
698 * the requested page.
700 * We determine whether good backing store exists for the requested
701 * page and return TRUE if it does, FALSE if it doesn't.
703 * If TRUE, we also try to determine how much valid, contiguous backing
704 * store exists before and after the requested page within a reasonable
705 * distance. We do not try to restrict it to the swap device stripe
706 * (that is handled in getpages/putpages). It probably isn't worth
711 swap_pager_haspage(vm_object_t object, vm_pindex_t pindex)
716 * do we have good backing store at the requested index ?
720 blk0 = swp_pager_meta_ctl(object, pindex, 0);
722 if (blk0 == SWAPBLK_NONE) {
729 * find backwards-looking contiguous good backing store
731 if (before != NULL) {
734 for (i = 1; i < (SWB_NPAGES/2); ++i) {
739 blk = swp_pager_meta_ctl(object, pindex - i, 0);
747 * find forward-looking contiguous good backing store
753 for (i = 1; i < (SWB_NPAGES/2); ++i) {
756 blk = swp_pager_meta_ctl(object, pindex + i, 0);
768 * SWAP_PAGER_PAGE_UNSWAPPED() - remove swap backing store related to page
770 * This removes any associated swap backing store, whether valid or
771 * not, from the page. This operates on any VM object, not just OBJT_SWAP
774 * This routine is typically called when a page is made dirty, at
775 * which point any associated swap can be freed. MADV_FREE also
776 * calls us in a special-case situation
778 * NOTE!!! If the page is clean and the swap was valid, the caller
779 * should make the page dirty before calling this routine. This routine
780 * does NOT change the m->dirty status of the page. Also: MADV_FREE
783 * This routine may not block
784 * This routine must be called at splvm()
787 swap_pager_unswapped(vm_page_t m)
789 if (m->flags & PG_SWAPPED) {
790 swp_pager_meta_ctl(m->object, m->pindex, SWM_FREE);
791 vm_page_flag_clear(m, PG_SWAPPED);
796 * SWAP_PAGER_STRATEGY() - read, write, free blocks
798 * This implements a VM OBJECT strategy function using swap backing store.
799 * This can operate on any VM OBJECT type, not necessarily just OBJT_SWAP
802 * This is intended to be a cacheless interface (i.e. caching occurs at
803 * higher levels), and is also used as a swap-based SSD cache for vnode
804 * and device objects.
806 * All I/O goes directly to and from the swap device.
808 * We currently attempt to run I/O synchronously or asynchronously as
809 * the caller requests. This isn't perfect because we loose error
810 * sequencing when we run multiple ops in parallel to satisfy a request.
811 * But this is swap, so we let it all hang out.
814 swap_pager_strategy(vm_object_t object, struct bio *bio)
816 struct buf *bp = bio->bio_buf;
819 vm_pindex_t biox_blkno = 0;
824 struct bio_track *track;
827 * tracking for swapdev vnode I/Os
829 if (bp->b_cmd == BUF_CMD_READ)
830 track = &swapdev_vp->v_track_read;
832 track = &swapdev_vp->v_track_write;
834 if (bp->b_bcount & PAGE_MASK) {
835 bp->b_error = EINVAL;
836 bp->b_flags |= B_ERROR | B_INVAL;
838 kprintf("swap_pager_strategy: bp %p offset %lld size %d, "
839 "not page bounded\n",
840 bp, (long long)bio->bio_offset, (int)bp->b_bcount);
845 * Clear error indication, initialize page index, count, data pointer.
848 bp->b_flags &= ~B_ERROR;
849 bp->b_resid = bp->b_bcount;
851 start = (vm_pindex_t)(bio->bio_offset >> PAGE_SHIFT);
852 count = howmany(bp->b_bcount, PAGE_SIZE);
856 * Deal with BUF_CMD_FREEBLKS
858 if (bp->b_cmd == BUF_CMD_FREEBLKS) {
860 * FREE PAGE(s) - destroy underlying swap that is no longer
863 swp_pager_meta_free(object, start, count);
870 * We need to be able to create a new cluster of I/O's. We cannot
871 * use the caller fields of the passed bio so push a new one.
873 * Because nbio is just a placeholder for the cluster links,
874 * we can biodone() the original bio instead of nbio to make
875 * things a bit more efficient.
877 nbio = push_bio(bio);
878 nbio->bio_offset = bio->bio_offset;
879 nbio->bio_caller_info1.cluster_head = NULL;
880 nbio->bio_caller_info2.cluster_tail = NULL;
886 * Execute read or write
892 * Obtain block. If block not found and writing, allocate a
893 * new block and build it into the object.
895 blk = swp_pager_meta_ctl(object, start, 0);
896 if ((blk == SWAPBLK_NONE) && bp->b_cmd != BUF_CMD_READ) {
897 blk = swp_pager_getswapspace(1);
898 if (blk == SWAPBLK_NONE) {
899 bp->b_error = ENOMEM;
900 bp->b_flags |= B_ERROR;
903 swp_pager_meta_build(object, start, blk);
907 * Do we have to flush our current collection? Yes if:
909 * - no swap block at this index
910 * - swap block is not contiguous
911 * - we cross a physical disk boundry in the
915 biox && (biox_blkno + btoc(bufx->b_bcount) != blk ||
916 ((biox_blkno ^ blk) & dmmax_mask)
919 if (bp->b_cmd == BUF_CMD_READ) {
920 ++mycpu->gd_cnt.v_swapin;
921 mycpu->gd_cnt.v_swappgsin += btoc(bufx->b_bcount);
923 ++mycpu->gd_cnt.v_swapout;
924 mycpu->gd_cnt.v_swappgsout += btoc(bufx->b_bcount);
925 bufx->b_dirtyend = bufx->b_bcount;
929 * Finished with this buf.
931 KKASSERT(bufx->b_bcount != 0);
932 if (bufx->b_cmd != BUF_CMD_READ)
933 bufx->b_dirtyend = bufx->b_bcount;
939 * Add new swapblk to biox, instantiating biox if necessary.
940 * Zero-fill reads are able to take a shortcut.
942 if (blk == SWAPBLK_NONE) {
944 * We can only get here if we are reading. Since
945 * we are at splvm() we can safely modify b_resid,
946 * even if chain ops are in progress.
948 bzero(data, PAGE_SIZE);
949 bp->b_resid -= PAGE_SIZE;
952 /* XXX chain count > 4, wait to <= 4 */
954 bufx = getpbuf(NULL);
955 biox = &bufx->b_bio1;
956 cluster_append(nbio, bufx);
957 bufx->b_flags |= (bufx->b_flags & B_ORDERED);
958 bufx->b_cmd = bp->b_cmd;
959 biox->bio_done = swap_chain_iodone;
960 biox->bio_offset = (off_t)blk << PAGE_SHIFT;
961 biox->bio_caller_info1.cluster_parent = nbio;
966 bufx->b_bcount += PAGE_SIZE;
974 * Flush out last buffer
977 if (bufx->b_cmd == BUF_CMD_READ) {
978 ++mycpu->gd_cnt.v_swapin;
979 mycpu->gd_cnt.v_swappgsin += btoc(bufx->b_bcount);
981 ++mycpu->gd_cnt.v_swapout;
982 mycpu->gd_cnt.v_swappgsout += btoc(bufx->b_bcount);
983 bufx->b_dirtyend = bufx->b_bcount;
985 KKASSERT(bufx->b_bcount);
986 if (bufx->b_cmd != BUF_CMD_READ)
987 bufx->b_dirtyend = bufx->b_bcount;
988 /* biox, bufx = NULL */
992 * Now initiate all the I/O. Be careful looping on our chain as
993 * I/O's may complete while we are still initiating them.
995 nbio->bio_caller_info2.cluster_tail = NULL;
996 bufx = nbio->bio_caller_info1.cluster_head;
999 biox = &bufx->b_bio1;
1001 bufx = bufx->b_cluster_next;
1002 vn_strategy(swapdev_vp, biox);
1006 * Completion of the cluster will also call biodone_chain(nbio).
1007 * We never call biodone(nbio) so we don't have to worry about
1008 * setting up a bio_done callback. It's handled in the sub-IO.
1014 swap_chain_iodone(struct bio *biox)
1017 struct buf *bufx; /* chained sub-buffer */
1018 struct bio *nbio; /* parent nbio with chain glue */
1019 struct buf *bp; /* original bp associated with nbio */
1022 bufx = biox->bio_buf;
1023 nbio = biox->bio_caller_info1.cluster_parent;
1027 * Update the original buffer
1029 KKASSERT(bp != NULL);
1030 if (bufx->b_flags & B_ERROR) {
1031 atomic_set_int(&bufx->b_flags, B_ERROR);
1032 bp->b_error = bufx->b_error;
1033 } else if (bufx->b_resid != 0) {
1034 atomic_set_int(&bufx->b_flags, B_ERROR);
1035 bp->b_error = EINVAL;
1037 atomic_subtract_int(&bp->b_resid, bufx->b_bcount);
1041 * Remove us from the chain.
1043 spin_lock_wr(&bp->b_lock.lk_spinlock);
1044 nextp = &nbio->bio_caller_info1.cluster_head;
1045 while (*nextp != bufx) {
1046 KKASSERT(*nextp != NULL);
1047 nextp = &(*nextp)->b_cluster_next;
1049 *nextp = bufx->b_cluster_next;
1050 chain_empty = (nbio->bio_caller_info1.cluster_head == NULL);
1051 spin_unlock_wr(&bp->b_lock.lk_spinlock);
1054 * Clean up bufx. If the chain is now empty we finish out
1055 * the parent. Note that we may be racing other completions
1056 * so we must use the chain_empty status from above.
1059 if (bp->b_resid != 0 && !(bp->b_flags & B_ERROR)) {
1060 atomic_set_int(&bp->b_flags, B_ERROR);
1061 bp->b_error = EINVAL;
1063 biodone_chain(nbio);
1065 relpbuf(bufx, NULL);
1069 * SWAP_PAGER_GETPAGES() - bring page in from swap
1071 * The requested page may have to be brought in from swap. Calculate the
1072 * swap block and bring in additional pages if possible. All pages must
1073 * have contiguous swap block assignments and reside in the same object.
1075 * The caller has a single vm_object_pip_add() reference prior to
1076 * calling us and we should return with the same.
1078 * The caller has BUSY'd the page. We should return with (*mpp) left busy,
1079 * and any additinal pages unbusied.
1081 * If the caller encounters a PG_RAM page it will pass it to us even though
1082 * it may be valid and dirty. We cannot overwrite the page in this case!
1083 * The case is used to allow us to issue pure read-aheads.
1085 * NOTE! XXX This code does not entirely pipeline yet due to the fact that
1086 * the PG_RAM page is validated at the same time as mreq. What we
1087 * really need to do is issue a separate read-ahead pbuf.
1090 swap_pager_getpage(vm_object_t object, vm_page_t *mpp, int seqaccess)
1101 vm_page_t marray[XIO_INTERNAL_PAGES];
1105 if (mreq->object != object) {
1106 panic("swap_pager_getpages: object mismatch %p/%p",
1113 * We don't want to overwrite a fully valid page as it might be
1114 * dirty. This case can occur when e.g. vm_fault hits a perfectly
1115 * valid page with PG_RAM set.
1117 * In this case we see if the next page is a suitable page-in
1118 * candidate and if it is we issue read-ahead. PG_RAM will be
1119 * set on the last page of the read-ahead to continue the pipeline.
1121 if (mreq->valid == VM_PAGE_BITS_ALL) {
1122 if (swap_burst_read == 0 || mreq->pindex + 1 >= object->size)
1123 return(VM_PAGER_OK);
1125 blk = swp_pager_meta_ctl(object, mreq->pindex + 1, 0);
1126 if (blk == SWAPBLK_NONE) {
1128 return(VM_PAGER_OK);
1130 m = vm_page_lookup(object, mreq->pindex + 1);
1132 m = vm_page_alloc(object, mreq->pindex + 1,
1136 return(VM_PAGER_OK);
1139 if ((m->flags & PG_BUSY) || m->busy || m->valid) {
1141 return(VM_PAGER_OK);
1143 vm_page_unqueue_nowakeup(m);
1154 * Try to block-read contiguous pages from swap if sequential,
1155 * otherwise just read one page. Contiguous pages from swap must
1156 * reside within a single device stripe because the I/O cannot be
1157 * broken up across multiple stripes.
1159 * Note that blk and iblk can be SWAPBLK_NONE but the loop is
1160 * set up such that the case(s) are handled implicitly.
1163 blk = swp_pager_meta_ctl(mreq->object, mreq->pindex, 0);
1166 for (i = 1; swap_burst_read &&
1167 i < XIO_INTERNAL_PAGES &&
1168 mreq->pindex + i < object->size; ++i) {
1171 iblk = swp_pager_meta_ctl(object, mreq->pindex + i, 0);
1172 if (iblk != blk + i)
1174 if ((blk ^ iblk) & dmmax_mask)
1176 m = vm_page_lookup(object, mreq->pindex + i);
1178 m = vm_page_alloc(object, mreq->pindex + i,
1183 if ((m->flags & PG_BUSY) || m->busy || m->valid)
1185 vm_page_unqueue_nowakeup(m);
1191 vm_page_flag_set(marray[i - 1], PG_RAM);
1196 * If mreq is the requested page and we have nothing to do return
1197 * VM_PAGER_FAIL. If raonly is set mreq is just another read-ahead
1198 * page and must be cleaned up.
1200 if (blk == SWAPBLK_NONE) {
1203 vnode_pager_freepage(mreq);
1204 return(VM_PAGER_OK);
1206 return(VM_PAGER_FAIL);
1211 * map our page(s) into kva for input
1213 bp = getpbuf(&nsw_rcount);
1215 kva = (vm_offset_t) bp->b_kvabase;
1216 bcopy(marray, bp->b_xio.xio_pages, i * sizeof(vm_page_t));
1217 pmap_qenter(kva, bp->b_xio.xio_pages, i);
1219 bp->b_data = (caddr_t)kva;
1220 bp->b_bcount = PAGE_SIZE * i;
1221 bp->b_xio.xio_npages = i;
1222 bio->bio_done = swp_pager_async_iodone;
1223 bio->bio_offset = (off_t)blk << PAGE_SHIFT;
1224 bio->bio_caller_info1.index = SWBIO_READ;
1227 * Set index. If raonly set the index beyond the array so all
1228 * the pages are treated the same, otherwise the original mreq is
1232 bio->bio_driver_info = (void *)(intptr_t)i;
1234 bio->bio_driver_info = (void *)(intptr_t)0;
1236 for (j = 0; j < i; ++j)
1237 vm_page_flag_set(bp->b_xio.xio_pages[j], PG_SWAPINPROG);
1239 mycpu->gd_cnt.v_swapin++;
1240 mycpu->gd_cnt.v_swappgsin += bp->b_xio.xio_npages;
1243 * We still hold the lock on mreq, and our automatic completion routine
1244 * does not remove it.
1246 vm_object_pip_add(object, bp->b_xio.xio_npages);
1249 * perform the I/O. NOTE!!! bp cannot be considered valid after
1250 * this point because we automatically release it on completion.
1251 * Instead, we look at the one page we are interested in which we
1252 * still hold a lock on even through the I/O completion.
1254 * The other pages in our m[] array are also released on completion,
1255 * so we cannot assume they are valid anymore either.
1257 bp->b_cmd = BUF_CMD_READ;
1259 vn_strategy(swapdev_vp, bio);
1262 * Wait for the page we want to complete. PG_SWAPINPROG is always
1263 * cleared on completion. If an I/O error occurs, SWAPBLK_NONE
1264 * is set in the meta-data.
1266 * If this is a read-ahead only we return immediately without
1270 return(VM_PAGER_OK);
1273 * Read-ahead includes originally requested page case.
1276 while ((mreq->flags & PG_SWAPINPROG) != 0) {
1277 vm_page_flag_set(mreq, PG_WANTED | PG_REFERENCED);
1278 mycpu->gd_cnt.v_intrans++;
1279 if (tsleep(mreq, 0, "swread", hz*20)) {
1281 "swap_pager: indefinite wait buffer: "
1282 " offset: %lld, size: %ld\n",
1283 (long long)bio->bio_offset,
1291 * mreq is left bussied after completion, but all the other pages
1292 * are freed. If we had an unrecoverable read error the page will
1295 if (mreq->valid != VM_PAGE_BITS_ALL)
1296 return(VM_PAGER_ERROR);
1298 return(VM_PAGER_OK);
1301 * A final note: in a low swap situation, we cannot deallocate swap
1302 * and mark a page dirty here because the caller is likely to mark
1303 * the page clean when we return, causing the page to possibly revert
1304 * to all-zero's later.
1309 * swap_pager_putpages:
1311 * Assign swap (if necessary) and initiate I/O on the specified pages.
1313 * We support both OBJT_DEFAULT and OBJT_SWAP objects. DEFAULT objects
1314 * are automatically converted to SWAP objects.
1316 * In a low memory situation we may block in vn_strategy(), but the new
1317 * vm_page reservation system coupled with properly written VFS devices
1318 * should ensure that no low-memory deadlock occurs. This is an area
1321 * The parent has N vm_object_pip_add() references prior to
1322 * calling us and will remove references for rtvals[] that are
1323 * not set to VM_PAGER_PEND. We need to remove the rest on I/O
1326 * The parent has soft-busy'd the pages it passes us and will unbusy
1327 * those whos rtvals[] entry is not set to VM_PAGER_PEND on return.
1328 * We need to unbusy the rest on I/O completion.
1331 swap_pager_putpages(vm_object_t object, vm_page_t *m, int count,
1332 boolean_t sync, int *rtvals)
1337 if (count && m[0]->object != object) {
1338 panic("swap_pager_getpages: object mismatch %p/%p",
1347 * Turn object into OBJT_SWAP
1348 * check for bogus sysops
1349 * force sync if not pageout process
1351 if (object->type == OBJT_DEFAULT)
1352 swp_pager_meta_convert(object);
1354 if (curthread != pagethread)
1360 * Update nsw parameters from swap_async_max sysctl values.
1361 * Do not let the sysop crash the machine with bogus numbers.
1364 if (swap_async_max != nsw_wcount_async_max) {
1370 if ((n = swap_async_max) > nswbuf / 2)
1377 * Adjust difference ( if possible ). If the current async
1378 * count is too low, we may not be able to make the adjustment
1382 n -= nsw_wcount_async_max;
1383 if (nsw_wcount_async + n >= 0) {
1384 nsw_wcount_async += n;
1385 nsw_wcount_async_max += n;
1386 wakeup(&nsw_wcount_async);
1394 * Assign swap blocks and issue I/O. We reallocate swap on the fly.
1395 * The page is left dirty until the pageout operation completes
1399 for (i = 0; i < count; i += n) {
1406 * Maximum I/O size is limited by a number of factors.
1409 n = min(BLIST_MAX_ALLOC, count - i);
1410 n = min(n, nsw_cluster_max);
1415 * Get biggest block of swap we can. If we fail, fall
1416 * back and try to allocate a smaller block. Don't go
1417 * overboard trying to allocate space if it would overly
1421 (blk = swp_pager_getswapspace(n)) == SWAPBLK_NONE &&
1426 if (blk == SWAPBLK_NONE) {
1427 for (j = 0; j < n; ++j)
1428 rtvals[i+j] = VM_PAGER_FAIL;
1434 * The I/O we are constructing cannot cross a physical
1435 * disk boundry in the swap stripe. Note: we are still
1438 if ((blk ^ (blk + n)) & dmmax_mask) {
1439 j = ((blk + dmmax) & dmmax_mask) - blk;
1440 swp_pager_freeswapspace(blk + j, n - j);
1445 * All I/O parameters have been satisfied, build the I/O
1446 * request and assign the swap space.
1450 bp = getpbuf(&nsw_wcount_sync);
1452 bp = getpbuf(&nsw_wcount_async);
1455 pmap_qenter((vm_offset_t)bp->b_data, &m[i], n);
1457 bp->b_bcount = PAGE_SIZE * n;
1458 bio->bio_offset = (off_t)blk << PAGE_SHIFT;
1460 for (j = 0; j < n; ++j) {
1461 vm_page_t mreq = m[i+j];
1463 swp_pager_meta_build(
1468 vm_page_dirty(mreq);
1469 rtvals[i+j] = VM_PAGER_OK;
1471 vm_page_flag_set(mreq, PG_SWAPINPROG);
1472 bp->b_xio.xio_pages[j] = mreq;
1474 bp->b_xio.xio_npages = n;
1476 mycpu->gd_cnt.v_swapout++;
1477 mycpu->gd_cnt.v_swappgsout += bp->b_xio.xio_npages;
1481 bp->b_dirtyoff = 0; /* req'd for NFS */
1482 bp->b_dirtyend = bp->b_bcount; /* req'd for NFS */
1483 bp->b_cmd = BUF_CMD_WRITE;
1484 bio->bio_caller_info1.index = SWBIO_WRITE;
1489 if (sync == FALSE) {
1490 bio->bio_done = swp_pager_async_iodone;
1492 vn_strategy(swapdev_vp, bio);
1494 for (j = 0; j < n; ++j)
1495 rtvals[i+j] = VM_PAGER_PEND;
1500 * Issue synchrnously.
1502 * Wait for the sync I/O to complete, then update rtvals.
1503 * We just set the rtvals[] to VM_PAGER_PEND so we can call
1504 * our async completion routine at the end, thus avoiding a
1507 bio->bio_caller_info1.index |= SWBIO_SYNC;
1508 bio->bio_done = biodone_sync;
1509 bio->bio_flags |= BIO_SYNC;
1510 vn_strategy(swapdev_vp, bio);
1511 biowait(bio, "swwrt");
1513 for (j = 0; j < n; ++j)
1514 rtvals[i+j] = VM_PAGER_PEND;
1517 * Now that we are through with the bp, we can call the
1518 * normal async completion, which frees everything up.
1520 swp_pager_async_iodone(bio);
1525 swap_pager_newswap(void)
1531 * swp_pager_async_iodone:
1533 * Completion routine for asynchronous reads and writes from/to swap.
1534 * Also called manually by synchronous code to finish up a bp.
1536 * For READ operations, the pages are PG_BUSY'd. For WRITE operations,
1537 * the pages are vm_page_t->busy'd. For READ operations, we PG_BUSY
1538 * unbusy all pages except the 'main' request page. For WRITE
1539 * operations, we vm_page_t->busy'd unbusy all pages ( we can do this
1540 * because we marked them all VM_PAGER_PEND on return from putpages ).
1542 * This routine may not block.
1545 swp_pager_async_iodone(struct bio *bio)
1547 struct buf *bp = bio->bio_buf;
1548 vm_object_t object = NULL;
1555 if (bp->b_flags & B_ERROR) {
1557 "swap_pager: I/O error - %s failed; offset %lld,"
1558 "size %ld, error %d\n",
1559 ((bio->bio_caller_info1.index & SWBIO_READ) ?
1560 "pagein" : "pageout"),
1561 (long long)bio->bio_offset,
1568 * set object, raise to splvm().
1570 if (bp->b_xio.xio_npages)
1571 object = bp->b_xio.xio_pages[0]->object;
1575 * remove the mapping for kernel virtual
1577 pmap_qremove((vm_offset_t)bp->b_data, bp->b_xio.xio_npages);
1580 * cleanup pages. If an error occurs writing to swap, we are in
1581 * very serious trouble. If it happens to be a disk error, though,
1582 * we may be able to recover by reassigning the swap later on. So
1583 * in this case we remove the m->swapblk assignment for the page
1584 * but do not free it in the rlist. The errornous block(s) are thus
1585 * never reallocated as swap. Redirty the page and continue.
1587 for (i = 0; i < bp->b_xio.xio_npages; ++i) {
1588 vm_page_t m = bp->b_xio.xio_pages[i];
1590 if (bp->b_flags & B_ERROR) {
1592 * If an error occurs I'd love to throw the swapblk
1593 * away without freeing it back to swapspace, so it
1594 * can never be used again. But I can't from an
1598 if (bio->bio_caller_info1.index & SWBIO_READ) {
1600 * When reading, reqpage needs to stay
1601 * locked for the parent, but all other
1602 * pages can be freed. We still want to
1603 * wakeup the parent waiting on the page,
1604 * though. ( also: pg_reqpage can be -1 and
1605 * not match anything ).
1607 * We have to wake specifically requested pages
1608 * up too because we cleared PG_SWAPINPROG and
1609 * someone may be waiting for that.
1611 * NOTE: for reads, m->dirty will probably
1612 * be overridden by the original caller of
1613 * getpages so don't play cute tricks here.
1615 * NOTE: We can't actually free the page from
1616 * here, because this is an interrupt. It
1617 * is not legal to mess with object->memq
1618 * from an interrupt. Deactivate the page
1623 vm_page_flag_clear(m, PG_ZERO);
1624 vm_page_flag_clear(m, PG_SWAPINPROG);
1627 * bio_driver_info holds the requested page
1630 if (i != (int)(intptr_t)bio->bio_driver_info) {
1631 vm_page_deactivate(m);
1637 * If i == bp->b_pager.pg_reqpage, do not wake
1638 * the page up. The caller needs to.
1642 * If a write error occurs, reactivate page
1643 * so it doesn't clog the inactive list,
1644 * then finish the I/O.
1647 vm_page_flag_clear(m, PG_SWAPINPROG);
1648 vm_page_activate(m);
1649 vm_page_io_finish(m);
1651 } else if (bio->bio_caller_info1.index & SWBIO_READ) {
1653 * NOTE: for reads, m->dirty will probably be
1654 * overridden by the original caller of getpages so
1655 * we cannot set them in order to free the underlying
1656 * swap in a low-swap situation. I don't think we'd
1657 * want to do that anyway, but it was an optimization
1658 * that existed in the old swapper for a time before
1659 * it got ripped out due to precisely this problem.
1661 * clear PG_ZERO in page.
1663 * If not the requested page then deactivate it.
1665 * Note that the requested page, reqpage, is left
1666 * busied, but we still have to wake it up. The
1667 * other pages are released (unbusied) by
1668 * vm_page_wakeup(). We do not set reqpage's
1669 * valid bits here, it is up to the caller.
1673 * NOTE: can't call pmap_clear_modify(m) from an
1674 * interrupt thread, the pmap code may have to map
1675 * non-kernel pmaps and currently asserts the case.
1677 /*pmap_clear_modify(m);*/
1678 m->valid = VM_PAGE_BITS_ALL;
1680 vm_page_flag_clear(m, PG_ZERO | PG_SWAPINPROG);
1681 vm_page_flag_set(m, PG_SWAPPED);
1684 * We have to wake specifically requested pages
1685 * up too because we cleared PG_SWAPINPROG and
1686 * could be waiting for it in getpages. However,
1687 * be sure to not unbusy getpages specifically
1688 * requested page - getpages expects it to be
1691 * bio_driver_info holds the requested page
1693 if (i != (int)(intptr_t)bio->bio_driver_info) {
1694 vm_page_deactivate(m);
1701 * Mark the page clean but do not mess with the
1702 * pmap-layer's modified state. That state should
1703 * also be clear since the caller protected the
1704 * page VM_PROT_READ, but allow the case.
1706 * We are in an interrupt, avoid pmap operations.
1708 * If we have a severe page deficit, deactivate the
1709 * page. Do not try to cache it (which would also
1710 * involve a pmap op), because the page might still
1714 vm_page_flag_clear(m, PG_SWAPINPROG);
1715 vm_page_flag_set(m, PG_SWAPPED);
1716 vm_page_io_finish(m);
1717 if (vm_page_count_severe())
1718 vm_page_deactivate(m);
1720 if (!vm_page_count_severe() || !vm_page_try_to_cache(m))
1721 vm_page_protect(m, VM_PROT_READ);
1727 * adjust pip. NOTE: the original parent may still have its own
1728 * pip refs on the object.
1732 vm_object_pip_wakeupn(object, bp->b_xio.xio_npages);
1735 * Release the physical I/O buffer.
1737 * NOTE: Due to synchronous operations in the write case b_cmd may
1738 * already be set to BUF_CMD_DONE and BIO_SYNC may have already
1741 if (bio->bio_caller_info1.index & SWBIO_READ)
1742 nswptr = &nsw_rcount;
1743 else if (bio->bio_caller_info1.index & SWBIO_SYNC)
1744 nswptr = &nsw_wcount_sync;
1746 nswptr = &nsw_wcount_async;
1747 bp->b_cmd = BUF_CMD_DONE;
1748 relpbuf(bp, nswptr);
1752 /************************************************************************
1754 ************************************************************************
1756 * These routines manipulate the swap metadata stored in the
1757 * OBJT_SWAP object. All swp_*() routines must be called at
1758 * splvm() because swap can be freed up by the low level vm_page
1759 * code which might be called from interrupts beyond what splbio() covers.
1761 * Swap metadata is implemented with a global hash and not directly
1762 * linked into the object. Instead the object simply contains
1763 * appropriate tracking counters.
1767 * Lookup the swblock containing the specified swap block index.
1771 swp_pager_lookup(vm_object_t object, vm_pindex_t index)
1773 index &= ~SWAP_META_MASK;
1774 return (RB_LOOKUP(swblock_rb_tree, &object->swblock_root, index));
1778 * Remove a swblock from the RB tree.
1782 swp_pager_remove(vm_object_t object, struct swblock *swap)
1784 RB_REMOVE(swblock_rb_tree, &object->swblock_root, swap);
1788 * Convert default object to swap object if necessary
1791 swp_pager_meta_convert(vm_object_t object)
1793 if (object->type == OBJT_DEFAULT) {
1794 object->type = OBJT_SWAP;
1795 KKASSERT(object->swblock_count == 0);
1800 * SWP_PAGER_META_BUILD() - add swap block to swap meta data for object
1802 * We first convert the object to a swap object if it is a default
1803 * object. Vnode objects do not need to be converted.
1805 * The specified swapblk is added to the object's swap metadata. If
1806 * the swapblk is not valid, it is freed instead. Any previously
1807 * assigned swapblk is freed.
1810 swp_pager_meta_build(vm_object_t object, vm_pindex_t index, daddr_t swapblk)
1812 struct swblock *swap;
1813 struct swblock *oswap;
1815 KKASSERT(swapblk != SWAPBLK_NONE);
1818 * Convert object if necessary
1820 if (object->type == OBJT_DEFAULT)
1821 swp_pager_meta_convert(object);
1824 * Locate swblock. If not found create, but if we aren't adding
1825 * anything just return. If we run out of space in the map we wait
1826 * and, since the hash table may have changed, retry.
1829 swap = swp_pager_lookup(object, index);
1834 swap = zalloc(swap_zone);
1839 swap->swb_index = index & ~SWAP_META_MASK;
1840 swap->swb_count = 0;
1842 ++object->swblock_count;
1844 for (i = 0; i < SWAP_META_PAGES; ++i)
1845 swap->swb_pages[i] = SWAPBLK_NONE;
1846 oswap = RB_INSERT(swblock_rb_tree, &object->swblock_root, swap);
1847 KKASSERT(oswap == NULL);
1851 * Delete prior contents of metadata
1854 index &= SWAP_META_MASK;
1856 if (swap->swb_pages[index] != SWAPBLK_NONE) {
1857 swp_pager_freeswapspace(swap->swb_pages[index], 1);
1862 * Enter block into metadata
1864 swap->swb_pages[index] = swapblk;
1865 if (swapblk != SWAPBLK_NONE)
1870 * SWP_PAGER_META_FREE() - free a range of blocks in the object's swap metadata
1872 * The requested range of blocks is freed, with any associated swap
1873 * returned to the swap bitmap.
1875 * This routine will free swap metadata structures as they are cleaned
1876 * out. This routine does *NOT* operate on swap metadata associated
1877 * with resident pages.
1879 * This routine must be called at splvm()
1881 static int swp_pager_meta_free_callback(struct swblock *swb, void *data);
1884 swp_pager_meta_free(vm_object_t object, vm_pindex_t index, vm_pindex_t count)
1886 struct swfreeinfo info;
1891 if (object->swblock_count == 0) {
1892 KKASSERT(RB_EMPTY(&object->swblock_root));
1899 * Setup for RB tree scan. Note that the pindex range can be huge
1900 * due to the 64 bit page index space so we cannot safely iterate.
1902 info.object = object;
1903 info.basei = index & ~SWAP_META_MASK;
1905 info.endi = index + count - 1;
1906 swblock_rb_tree_RB_SCAN(&object->swblock_root, rb_swblock_scancmp,
1907 swp_pager_meta_free_callback, &info);
1912 swp_pager_meta_free_callback(struct swblock *swap, void *data)
1914 struct swfreeinfo *info = data;
1915 vm_object_t object = info->object;
1920 * Figure out the range within the swblock. The wider scan may
1921 * return edge-case swap blocks when the start and/or end points
1922 * are in the middle of a block.
1924 if (swap->swb_index < info->begi)
1925 index = (int)info->begi & SWAP_META_MASK;
1929 if (swap->swb_index + SWAP_META_PAGES > info->endi)
1930 eindex = (int)info->endi & SWAP_META_MASK;
1932 eindex = SWAP_META_MASK;
1935 * Scan and free the blocks. The loop terminates early
1936 * if (swap) runs out of blocks and could be freed.
1938 while (index <= eindex) {
1939 daddr_t v = swap->swb_pages[index];
1941 if (v != SWAPBLK_NONE) {
1942 swp_pager_freeswapspace(v, 1);
1943 swap->swb_pages[index] = SWAPBLK_NONE;
1944 if (--swap->swb_count == 0) {
1945 swp_pager_remove(object, swap);
1946 zfree(swap_zone, swap);
1947 --object->swblock_count;
1953 /* swap may be invalid here due to zfree above */
1958 * SWP_PAGER_META_FREE_ALL() - destroy all swap metadata associated with object
1960 * This routine locates and destroys all swap metadata associated with
1963 * This routine must be called at splvm()
1966 swp_pager_meta_free_all(vm_object_t object)
1968 struct swblock *swap;
1971 while ((swap = RB_ROOT(&object->swblock_root)) != NULL) {
1972 swp_pager_remove(object, swap);
1973 for (i = 0; i < SWAP_META_PAGES; ++i) {
1974 daddr_t v = swap->swb_pages[i];
1975 if (v != SWAPBLK_NONE) {
1977 swp_pager_freeswapspace(v, 1);
1980 if (swap->swb_count != 0)
1981 panic("swap_pager_meta_free_all: swb_count != 0");
1982 zfree(swap_zone, swap);
1983 --object->swblock_count;
1985 KKASSERT(object->swblock_count == 0);
1989 * SWP_PAGER_METACTL() - misc control of swap and vm_page_t meta data.
1991 * This routine is capable of looking up, popping, or freeing
1992 * swapblk assignments in the swap meta data or in the vm_page_t.
1993 * The routine typically returns the swapblk being looked-up, or popped,
1994 * or SWAPBLK_NONE if the block was freed, or SWAPBLK_NONE if the block
1995 * was invalid. This routine will automatically free any invalid
1996 * meta-data swapblks.
1998 * It is not possible to store invalid swapblks in the swap meta data
1999 * (other then a literal 'SWAPBLK_NONE'), so we don't bother checking.
2001 * When acting on a busy resident page and paging is in progress, we
2002 * have to wait until paging is complete but otherwise can act on the
2005 * This routine must be called at splvm().
2007 * SWM_FREE remove and free swap block from metadata
2008 * SWM_POP remove from meta data but do not free.. pop it out
2011 swp_pager_meta_ctl(vm_object_t object, vm_pindex_t index, int flags)
2013 struct swblock *swap;
2016 if (object->swblock_count == 0)
2017 return(SWAPBLK_NONE);
2020 swap = swp_pager_lookup(object, index);
2023 index &= SWAP_META_MASK;
2024 r1 = swap->swb_pages[index];
2026 if (r1 != SWAPBLK_NONE) {
2027 if (flags & SWM_FREE) {
2028 swp_pager_freeswapspace(r1, 1);
2031 if (flags & (SWM_FREE|SWM_POP)) {
2032 swap->swb_pages[index] = SWAPBLK_NONE;
2033 if (--swap->swb_count == 0) {
2034 swp_pager_remove(object, swap);
2035 zfree(swap_zone, swap);
2036 --object->swblock_count;