2 * SPDX-License-Identifier: BSD-4-Clause
4 * Copyright (c) 1998 Matthew Dillon,
5 * Copyright (c) 1994 John S. Dyson
6 * Copyright (c) 1990 University of Utah.
7 * Copyright (c) 1982, 1986, 1989, 1993
8 * The Regents of the University of California. All rights reserved.
10 * This code is derived from software contributed to Berkeley by
11 * the Systems Programming Group of the University of Utah Computer
14 * Redistribution and use in source and binary forms, with or without
15 * modification, are permitted provided that the following conditions
17 * 1. Redistributions of source code must retain the above copyright
18 * notice, this list of conditions and the following disclaimer.
19 * 2. Redistributions in binary form must reproduce the above copyright
20 * notice, this list of conditions and the following disclaimer in the
21 * documentation and/or other materials provided with the distribution.
22 * 3. All advertising materials mentioning features or use of this software
23 * must display the following acknowledgement:
24 * This product includes software developed by the University of
25 * California, Berkeley and its contributors.
26 * 4. Neither the name of the University nor the names of its contributors
27 * may be used to endorse or promote products derived from this software
28 * without specific prior written permission.
30 * THIS SOFTWARE IS PROVIDED BY THE REGENTS AND CONTRIBUTORS ``AS IS'' AND
31 * ANY EXPRESS OR IMPLIED WARRANTIES, INCLUDING, BUT NOT LIMITED TO, THE
32 * IMPLIED WARRANTIES OF MERCHANTABILITY AND FITNESS FOR A PARTICULAR PURPOSE
33 * ARE DISCLAIMED. IN NO EVENT SHALL THE REGENTS OR CONTRIBUTORS BE LIABLE
34 * FOR ANY DIRECT, INDIRECT, INCIDENTAL, SPECIAL, EXEMPLARY, OR CONSEQUENTIAL
35 * DAMAGES (INCLUDING, BUT NOT LIMITED TO, PROCUREMENT OF SUBSTITUTE GOODS
36 * OR SERVICES; LOSS OF USE, DATA, OR PROFITS; OR BUSINESS INTERRUPTION)
37 * HOWEVER CAUSED AND ON ANY THEORY OF LIABILITY, WHETHER IN CONTRACT, STRICT
38 * LIABILITY, OR TORT (INCLUDING NEGLIGENCE OR OTHERWISE) ARISING IN ANY WAY
39 * OUT OF THE USE OF THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF
45 * Radix Bitmap 'blists'.
47 * - The new swapper uses the new radix bitmap code. This should scale
48 * to arbitrarily small or arbitrarily large swap spaces and an almost
49 * arbitrary degree of fragmentation.
53 * - on the fly reallocation of swap during putpages. The new system
54 * does not try to keep previously allocated swap blocks for dirty
57 * - on the fly deallocation of swap
59 * - No more garbage collection required. Unnecessarily allocated swap
60 * blocks only exist for dirty vm_page_t's now and these are already
61 * cycled (in a high-load system) by the pager. We also do on-the-fly
62 * removal of invalidated swap blocks when a page is destroyed
65 * from: Utah $Hdr: swap_pager.c 1.4 91/04/30$
67 * @(#)swap_pager.c 8.9 (Berkeley) 3/21/94
68 * @(#)vm_swap.c 8.5 (Berkeley) 2/17/94
71 #include <sys/cdefs.h>
72 __FBSDID("$FreeBSD$");
76 #include <sys/param.h>
77 #include <sys/systm.h>
79 #include <sys/kernel.h>
85 #include <sys/fcntl.h>
86 #include <sys/mount.h>
87 #include <sys/namei.h>
88 #include <sys/vnode.h>
89 #include <sys/malloc.h>
90 #include <sys/pctrie.h>
91 #include <sys/racct.h>
92 #include <sys/resource.h>
93 #include <sys/resourcevar.h>
94 #include <sys/rwlock.h>
96 #include <sys/sysctl.h>
97 #include <sys/sysproto.h>
98 #include <sys/blist.h>
101 #include <sys/vmmeter.h>
103 #include <security/mac/mac_framework.h>
107 #include <vm/vm_map.h>
108 #include <vm/vm_kern.h>
109 #include <vm/vm_object.h>
110 #include <vm/vm_page.h>
111 #include <vm/vm_pager.h>
112 #include <vm/vm_pageout.h>
113 #include <vm/vm_param.h>
114 #include <vm/swap_pager.h>
115 #include <vm/vm_extern.h>
118 #include <geom/geom.h>
121 * MAX_PAGEOUT_CLUSTER must be a power of 2 between 1 and 64.
122 * The 64-page limit is due to the radix code (kern/subr_blist.c).
124 #ifndef MAX_PAGEOUT_CLUSTER
125 #define MAX_PAGEOUT_CLUSTER 32
128 #if !defined(SWB_NPAGES)
129 #define SWB_NPAGES MAX_PAGEOUT_CLUSTER
132 #define SWAP_META_PAGES PCTRIE_COUNT
135 * A swblk structure maps each page index within a
136 * SWAP_META_PAGES-aligned and sized range to the address of an
137 * on-disk swap block (or SWAPBLK_NONE). The collection of these
138 * mappings for an entire vm object is implemented as a pc-trie.
142 daddr_t d[SWAP_META_PAGES];
145 static MALLOC_DEFINE(M_VMPGDATA, "vm_pgdata", "swap pager private data");
146 static struct mtx sw_dev_mtx;
147 static TAILQ_HEAD(, swdevt) swtailq = TAILQ_HEAD_INITIALIZER(swtailq);
148 static struct swdevt *swdevhd; /* Allocate from here next */
149 static int nswapdev; /* Number of swap devices */
150 int swap_pager_avail;
151 static struct sx swdev_syscall_lock; /* serialize swap(on|off) */
153 static u_long swap_reserved;
154 static u_long swap_total;
155 static int sysctl_page_shift(SYSCTL_HANDLER_ARGS);
156 SYSCTL_PROC(_vm, OID_AUTO, swap_reserved, CTLTYPE_U64 | CTLFLAG_RD | CTLFLAG_MPSAFE,
157 &swap_reserved, 0, sysctl_page_shift, "A",
158 "Amount of swap storage needed to back all allocated anonymous memory.");
159 SYSCTL_PROC(_vm, OID_AUTO, swap_total, CTLTYPE_U64 | CTLFLAG_RD | CTLFLAG_MPSAFE,
160 &swap_total, 0, sysctl_page_shift, "A",
161 "Total amount of available swap storage.");
163 static int overcommit = 0;
164 SYSCTL_INT(_vm, VM_OVERCOMMIT, overcommit, CTLFLAG_RW, &overcommit, 0,
165 "Configure virtual memory overcommit behavior. See tuning(7) "
167 static unsigned long swzone;
168 SYSCTL_ULONG(_vm, OID_AUTO, swzone, CTLFLAG_RD, &swzone, 0,
169 "Actual size of swap metadata zone");
170 static unsigned long swap_maxpages;
171 SYSCTL_ULONG(_vm, OID_AUTO, swap_maxpages, CTLFLAG_RD, &swap_maxpages, 0,
172 "Maximum amount of swap supported");
174 /* bits from overcommit */
175 #define SWAP_RESERVE_FORCE_ON (1 << 0)
176 #define SWAP_RESERVE_RLIMIT_ON (1 << 1)
177 #define SWAP_RESERVE_ALLOW_NONWIRED (1 << 2)
180 sysctl_page_shift(SYSCTL_HANDLER_ARGS)
183 u_long value = *(u_long *)arg1;
185 newval = ((uint64_t)value) << PAGE_SHIFT;
186 return (sysctl_handle_64(oidp, &newval, 0, req));
190 swap_reserve(vm_ooffset_t incr)
193 return (swap_reserve_by_cred(incr, curthread->td_ucred));
197 swap_reserve_by_cred(vm_ooffset_t incr, struct ucred *cred)
199 u_long r, s, prev, pincr;
202 static struct timeval lastfail;
205 uip = cred->cr_ruidinfo;
207 KASSERT((incr & PAGE_MASK) == 0, ("%s: incr: %ju & PAGE_MASK", __func__,
213 error = racct_add(curproc, RACCT_SWAP, incr);
214 PROC_UNLOCK(curproc);
222 prev = atomic_fetchadd_long(&swap_reserved, pincr);
224 if (overcommit & SWAP_RESERVE_ALLOW_NONWIRED) {
225 s = vm_cnt.v_page_count - vm_cnt.v_free_reserved -
230 if ((overcommit & SWAP_RESERVE_FORCE_ON) == 0 || r <= s ||
231 (error = priv_check(curthread, PRIV_VM_SWAP_NOQUOTA)) == 0) {
234 prev = atomic_fetchadd_long(&swap_reserved, -pincr);
236 panic("swap_reserved < incr on overcommit fail");
239 prev = atomic_fetchadd_long(&uip->ui_vmsize, pincr);
240 if ((overcommit & SWAP_RESERVE_RLIMIT_ON) != 0 &&
241 prev + pincr > lim_cur(curthread, RLIMIT_SWAP) &&
242 priv_check(curthread, PRIV_VM_SWAP_NORLIMIT)) {
244 prev = atomic_fetchadd_long(&uip->ui_vmsize, -pincr);
246 panic("uip->ui_vmsize < incr on overcommit fail");
249 if (!res && ppsratecheck(&lastfail, &curfail, 1)) {
250 printf("uid %d, pid %d: swap reservation for %jd bytes failed\n",
251 uip->ui_uid, curproc->p_pid, incr);
255 if (racct_enable && !res) {
257 racct_sub(curproc, RACCT_SWAP, incr);
258 PROC_UNLOCK(curproc);
266 swap_reserve_force(vm_ooffset_t incr)
271 KASSERT((incr & PAGE_MASK) == 0, ("%s: incr: %ju & PAGE_MASK", __func__,
277 racct_add_force(curproc, RACCT_SWAP, incr);
280 atomic_add_long(&swap_reserved, pincr);
281 uip = curproc->p_ucred->cr_ruidinfo;
282 atomic_add_long(&uip->ui_vmsize, pincr);
283 PROC_UNLOCK(curproc);
287 swap_release(vm_ooffset_t decr)
292 cred = curproc->p_ucred;
293 swap_release_by_cred(decr, cred);
294 PROC_UNLOCK(curproc);
298 swap_release_by_cred(vm_ooffset_t decr, struct ucred *cred)
303 uip = cred->cr_ruidinfo;
305 KASSERT((decr & PAGE_MASK) == 0, ("%s: decr: %ju & PAGE_MASK", __func__,
309 prev = atomic_fetchadd_long(&swap_reserved, -pdecr);
311 panic("swap_reserved < decr");
313 prev = atomic_fetchadd_long(&uip->ui_vmsize, -pdecr);
315 printf("negative vmsize for uid = %d\n", uip->ui_uid);
318 racct_sub_cred(cred, RACCT_SWAP, decr);
322 #define SWM_POP 0x01 /* pop out */
324 static int swap_pager_full = 2; /* swap space exhaustion (task killing) */
325 static int swap_pager_almost_full = 1; /* swap space exhaustion (w/hysteresis)*/
326 static struct mtx swbuf_mtx; /* to sync nsw_wcount_async */
327 static int nsw_wcount_async; /* limit async write buffers */
328 static int nsw_wcount_async_max;/* assigned maximum */
329 static int nsw_cluster_max; /* maximum VOP I/O allowed */
331 static int sysctl_swap_async_max(SYSCTL_HANDLER_ARGS);
332 SYSCTL_PROC(_vm, OID_AUTO, swap_async_max, CTLTYPE_INT | CTLFLAG_RW |
333 CTLFLAG_MPSAFE, NULL, 0, sysctl_swap_async_max, "I",
334 "Maximum running async swap ops");
335 static int sysctl_swap_fragmentation(SYSCTL_HANDLER_ARGS);
336 SYSCTL_PROC(_vm, OID_AUTO, swap_fragmentation, CTLTYPE_STRING | CTLFLAG_RD |
337 CTLFLAG_MPSAFE, NULL, 0, sysctl_swap_fragmentation, "A",
338 "Swap Fragmentation Info");
340 static struct sx sw_alloc_sx;
343 * "named" and "unnamed" anon region objects. Try to reduce the overhead
344 * of searching a named list by hashing it just a little.
349 #define NOBJLIST(handle) \
350 (&swap_pager_object_list[((int)(intptr_t)handle >> 4) & (NOBJLISTS-1)])
352 static struct pagerlst swap_pager_object_list[NOBJLISTS];
353 static uma_zone_t swwbuf_zone;
354 static uma_zone_t swrbuf_zone;
355 static uma_zone_t swblk_zone;
356 static uma_zone_t swpctrie_zone;
359 * pagerops for OBJT_SWAP - "swap pager". Some ops are also global procedure
360 * calls hooked from other parts of the VM system and do not appear here.
361 * (see vm/swap_pager.h).
364 swap_pager_alloc(void *handle, vm_ooffset_t size,
365 vm_prot_t prot, vm_ooffset_t offset, struct ucred *);
366 static void swap_pager_dealloc(vm_object_t object);
367 static int swap_pager_getpages(vm_object_t, vm_page_t *, int, int *,
369 static int swap_pager_getpages_async(vm_object_t, vm_page_t *, int, int *,
370 int *, pgo_getpages_iodone_t, void *);
371 static void swap_pager_putpages(vm_object_t, vm_page_t *, int, boolean_t, int *);
373 swap_pager_haspage(vm_object_t object, vm_pindex_t pindex, int *before, int *after);
374 static void swap_pager_init(void);
375 static void swap_pager_unswapped(vm_page_t);
376 static void swap_pager_swapoff(struct swdevt *sp);
378 struct pagerops swappagerops = {
379 .pgo_init = swap_pager_init, /* early system initialization of pager */
380 .pgo_alloc = swap_pager_alloc, /* allocate an OBJT_SWAP object */
381 .pgo_dealloc = swap_pager_dealloc, /* deallocate an OBJT_SWAP object */
382 .pgo_getpages = swap_pager_getpages, /* pagein */
383 .pgo_getpages_async = swap_pager_getpages_async, /* pagein (async) */
384 .pgo_putpages = swap_pager_putpages, /* pageout */
385 .pgo_haspage = swap_pager_haspage, /* get backing store status for page */
386 .pgo_pageunswapped = swap_pager_unswapped, /* remove swap related to page */
390 * swap_*() routines are externally accessible. swp_*() routines are
393 static int nswap_lowat = 128; /* in pages, swap_pager_almost_full warn */
394 static int nswap_hiwat = 512; /* in pages, swap_pager_almost_full warn */
396 SYSCTL_INT(_vm, OID_AUTO, dmmax, CTLFLAG_RD, &nsw_cluster_max, 0,
397 "Maximum size of a swap block in pages");
399 static void swp_sizecheck(void);
400 static void swp_pager_async_iodone(struct buf *bp);
401 static bool swp_pager_swblk_empty(struct swblk *sb, int start, int limit);
402 static int swapongeom(struct vnode *);
403 static int swaponvp(struct thread *, struct vnode *, u_long);
404 static int swapoff_one(struct swdevt *sp, struct ucred *cred);
407 * Swap bitmap functions
409 static void swp_pager_freeswapspace(daddr_t blk, daddr_t npages);
410 static daddr_t swp_pager_getswapspace(int *npages, int limit);
415 static daddr_t swp_pager_meta_build(vm_object_t, vm_pindex_t, daddr_t);
416 static void swp_pager_meta_free(vm_object_t, vm_pindex_t, vm_pindex_t);
417 static void swp_pager_meta_free_all(vm_object_t);
418 static daddr_t swp_pager_meta_ctl(vm_object_t, vm_pindex_t, int);
421 swp_pager_init_freerange(daddr_t *start, daddr_t *num)
424 *start = SWAPBLK_NONE;
429 swp_pager_update_freerange(daddr_t *start, daddr_t *num, daddr_t addr)
432 if (*start + *num == addr) {
435 swp_pager_freeswapspace(*start, *num);
442 swblk_trie_alloc(struct pctrie *ptree)
445 return (uma_zalloc(swpctrie_zone, M_NOWAIT | (curproc == pageproc ?
446 M_USE_RESERVE : 0)));
450 swblk_trie_free(struct pctrie *ptree, void *node)
453 uma_zfree(swpctrie_zone, node);
456 PCTRIE_DEFINE(SWAP, swblk, p, swblk_trie_alloc, swblk_trie_free);
459 * SWP_SIZECHECK() - update swap_pager_full indication
461 * update the swap_pager_almost_full indication and warn when we are
462 * about to run out of swap space, using lowat/hiwat hysteresis.
464 * Clear swap_pager_full ( task killing ) indication when lowat is met.
466 * No restrictions on call
467 * This routine may not block.
473 if (swap_pager_avail < nswap_lowat) {
474 if (swap_pager_almost_full == 0) {
475 printf("swap_pager: out of swap space\n");
476 swap_pager_almost_full = 1;
480 if (swap_pager_avail > nswap_hiwat)
481 swap_pager_almost_full = 0;
486 * SWAP_PAGER_INIT() - initialize the swap pager!
488 * Expected to be started from system init. NOTE: This code is run
489 * before much else so be careful what you depend on. Most of the VM
490 * system has yet to be initialized at this point.
493 swap_pager_init(void)
496 * Initialize object lists
500 for (i = 0; i < NOBJLISTS; ++i)
501 TAILQ_INIT(&swap_pager_object_list[i]);
502 mtx_init(&sw_dev_mtx, "swapdev", NULL, MTX_DEF);
503 sx_init(&sw_alloc_sx, "swspsx");
504 sx_init(&swdev_syscall_lock, "swsysc");
508 * SWAP_PAGER_SWAP_INIT() - swap pager initialization from pageout process
510 * Expected to be started from pageout process once, prior to entering
514 swap_pager_swap_init(void)
519 * Number of in-transit swap bp operations. Don't
520 * exhaust the pbufs completely. Make sure we
521 * initialize workable values (0 will work for hysteresis
522 * but it isn't very efficient).
524 * The nsw_cluster_max is constrained by the bp->b_pages[]
525 * array (MAXPHYS/PAGE_SIZE) and our locally defined
526 * MAX_PAGEOUT_CLUSTER. Also be aware that swap ops are
527 * constrained by the swap device interleave stripe size.
529 * Currently we hardwire nsw_wcount_async to 4. This limit is
530 * designed to prevent other I/O from having high latencies due to
531 * our pageout I/O. The value 4 works well for one or two active swap
532 * devices but is probably a little low if you have more. Even so,
533 * a higher value would probably generate only a limited improvement
534 * with three or four active swap devices since the system does not
535 * typically have to pageout at extreme bandwidths. We will want
536 * at least 2 per swap devices, and 4 is a pretty good value if you
537 * have one NFS swap device due to the command/ack latency over NFS.
538 * So it all works out pretty well.
540 nsw_cluster_max = min((MAXPHYS/PAGE_SIZE), MAX_PAGEOUT_CLUSTER);
542 nsw_wcount_async = 4;
543 nsw_wcount_async_max = nsw_wcount_async;
544 mtx_init(&swbuf_mtx, "async swbuf mutex", NULL, MTX_DEF);
546 swwbuf_zone = pbuf_zsecond_create("swwbuf", nswbuf / 4);
547 swrbuf_zone = pbuf_zsecond_create("swrbuf", nswbuf / 2);
550 * Initialize our zone, taking the user's requested size or
551 * estimating the number we need based on the number of pages
554 n = maxswzone != 0 ? maxswzone / sizeof(struct swblk) :
555 vm_cnt.v_page_count / 2;
556 swpctrie_zone = uma_zcreate("swpctrie", pctrie_node_size(), NULL, NULL,
557 pctrie_zone_init, NULL, UMA_ALIGN_PTR, UMA_ZONE_VM);
558 if (swpctrie_zone == NULL)
559 panic("failed to create swap pctrie zone.");
560 swblk_zone = uma_zcreate("swblk", sizeof(struct swblk), NULL, NULL,
561 NULL, NULL, _Alignof(struct swblk) - 1, UMA_ZONE_VM);
562 if (swblk_zone == NULL)
563 panic("failed to create swap blk zone.");
566 if (uma_zone_reserve_kva(swblk_zone, n))
569 * if the allocation failed, try a zone two thirds the
570 * size of the previous attempt.
576 * Often uma_zone_reserve_kva() cannot reserve exactly the
577 * requested size. Account for the difference when
578 * calculating swap_maxpages.
580 n = uma_zone_get_max(swblk_zone);
583 printf("Swap blk zone entries changed from %lu to %lu.\n",
585 swap_maxpages = n * SWAP_META_PAGES;
586 swzone = n * sizeof(struct swblk);
587 if (!uma_zone_reserve_kva(swpctrie_zone, n))
588 printf("Cannot reserve swap pctrie zone, "
589 "reduce kern.maxswzone.\n");
593 swap_pager_alloc_init(void *handle, struct ucred *cred, vm_ooffset_t size,
599 if (!swap_reserve_by_cred(size, cred))
605 * The un_pager.swp.swp_blks trie is initialized by
606 * vm_object_allocate() to ensure the correct order of
607 * visibility to other threads.
609 object = vm_object_allocate(OBJT_SWAP, OFF_TO_IDX(offset +
612 object->handle = handle;
615 object->charge = size;
621 * SWAP_PAGER_ALLOC() - allocate a new OBJT_SWAP VM object and instantiate
622 * its metadata structures.
624 * This routine is called from the mmap and fork code to create a new
627 * This routine must ensure that no live duplicate is created for
628 * the named object request, which is protected against by
629 * holding the sw_alloc_sx lock in case handle != NULL.
632 swap_pager_alloc(void *handle, vm_ooffset_t size, vm_prot_t prot,
633 vm_ooffset_t offset, struct ucred *cred)
637 if (handle != NULL) {
639 * Reference existing named region or allocate new one. There
640 * should not be a race here against swp_pager_meta_build()
641 * as called from vm_page_remove() in regards to the lookup
644 sx_xlock(&sw_alloc_sx);
645 object = vm_pager_object_lookup(NOBJLIST(handle), handle);
646 if (object == NULL) {
647 object = swap_pager_alloc_init(handle, cred, size,
649 if (object != NULL) {
650 TAILQ_INSERT_TAIL(NOBJLIST(object->handle),
651 object, pager_object_list);
654 sx_xunlock(&sw_alloc_sx);
656 object = swap_pager_alloc_init(handle, cred, size, offset);
662 * SWAP_PAGER_DEALLOC() - remove swap metadata from object
664 * The swap backing for the object is destroyed. The code is
665 * designed such that we can reinstantiate it later, but this
666 * routine is typically called only when the entire object is
667 * about to be destroyed.
669 * The object must be locked.
672 swap_pager_dealloc(vm_object_t object)
675 VM_OBJECT_ASSERT_WLOCKED(object);
676 KASSERT((object->flags & OBJ_DEAD) != 0, ("dealloc of reachable obj"));
679 * Remove from list right away so lookups will fail if we block for
680 * pageout completion.
682 if (object->handle != NULL) {
683 VM_OBJECT_WUNLOCK(object);
684 sx_xlock(&sw_alloc_sx);
685 TAILQ_REMOVE(NOBJLIST(object->handle), object,
687 sx_xunlock(&sw_alloc_sx);
688 VM_OBJECT_WLOCK(object);
691 vm_object_pip_wait(object, "swpdea");
694 * Free all remaining metadata. We only bother to free it from
695 * the swap meta data. We do not attempt to free swapblk's still
696 * associated with vm_page_t's for this object. We do not care
697 * if paging is still in progress on some objects.
699 swp_pager_meta_free_all(object);
700 object->handle = NULL;
701 object->type = OBJT_DEAD;
704 /************************************************************************
705 * SWAP PAGER BITMAP ROUTINES *
706 ************************************************************************/
709 * SWP_PAGER_GETSWAPSPACE() - allocate raw swap space
711 * Allocate swap for up to the requested number of pages, and at
712 * least a minimum number of pages. The starting swap block number
713 * (a page index) is returned or SWAPBLK_NONE if the allocation
716 * Also has the side effect of advising that somebody made a mistake
717 * when they configured swap and didn't configure enough.
719 * This routine may not sleep.
721 * We allocate in round-robin fashion from the configured devices.
724 swp_pager_getswapspace(int *io_npages, int limit)
732 mtx_lock(&sw_dev_mtx);
734 while (!TAILQ_EMPTY(&swtailq)) {
736 sp = TAILQ_FIRST(&swtailq);
737 if ((sp->sw_flags & SW_CLOSING) == 0)
738 blk = blist_alloc(sp->sw_blist, npages);
739 if (blk != SWAPBLK_NONE)
741 sp = TAILQ_NEXT(sp, sw_list);
748 if (blk != SWAPBLK_NONE) {
751 sp->sw_used += npages;
752 swap_pager_avail -= npages;
754 swdevhd = TAILQ_NEXT(sp, sw_list);
756 if (swap_pager_full != 2) {
757 printf("swp_pager_getswapspace(%d): failed\n",
760 swap_pager_almost_full = 1;
764 mtx_unlock(&sw_dev_mtx);
769 swp_pager_isondev(daddr_t blk, struct swdevt *sp)
772 return (blk >= sp->sw_first && blk < sp->sw_end);
776 swp_pager_strategy(struct buf *bp)
780 mtx_lock(&sw_dev_mtx);
781 TAILQ_FOREACH(sp, &swtailq, sw_list) {
782 if (swp_pager_isondev(bp->b_blkno, sp)) {
783 mtx_unlock(&sw_dev_mtx);
784 if ((sp->sw_flags & SW_UNMAPPED) != 0 &&
785 unmapped_buf_allowed) {
786 bp->b_data = unmapped_buf;
789 pmap_qenter((vm_offset_t)bp->b_data,
790 &bp->b_pages[0], bp->b_bcount / PAGE_SIZE);
792 sp->sw_strategy(bp, sp);
796 panic("Swapdev not found");
801 * SWP_PAGER_FREESWAPSPACE() - free raw swap space
803 * This routine returns the specified swap blocks back to the bitmap.
805 * This routine may not sleep.
808 swp_pager_freeswapspace(daddr_t blk, daddr_t npages)
814 mtx_lock(&sw_dev_mtx);
815 TAILQ_FOREACH(sp, &swtailq, sw_list) {
816 if (swp_pager_isondev(blk, sp)) {
817 sp->sw_used -= npages;
819 * If we are attempting to stop swapping on
820 * this device, we don't want to mark any
821 * blocks free lest they be reused.
823 if ((sp->sw_flags & SW_CLOSING) == 0) {
824 blist_free(sp->sw_blist, blk - sp->sw_first,
826 swap_pager_avail += npages;
829 mtx_unlock(&sw_dev_mtx);
833 panic("Swapdev not found");
837 * SYSCTL_SWAP_FRAGMENTATION() - produce raw swap space stats
840 sysctl_swap_fragmentation(SYSCTL_HANDLER_ARGS)
847 error = sysctl_wire_old_buffer(req, 0);
850 sbuf_new_for_sysctl(&sbuf, NULL, 128, req);
851 mtx_lock(&sw_dev_mtx);
852 TAILQ_FOREACH(sp, &swtailq, sw_list) {
853 if (vn_isdisk(sp->sw_vp, NULL))
854 devname = devtoname(sp->sw_vp->v_rdev);
857 sbuf_printf(&sbuf, "\nFree space on device %s:\n", devname);
858 blist_stats(sp->sw_blist, &sbuf);
860 mtx_unlock(&sw_dev_mtx);
861 error = sbuf_finish(&sbuf);
867 * SWAP_PAGER_FREESPACE() - frees swap blocks associated with a page
868 * range within an object.
870 * This is a globally accessible routine.
872 * This routine removes swapblk assignments from swap metadata.
874 * The external callers of this routine typically have already destroyed
875 * or renamed vm_page_t's associated with this range in the object so
878 * The object must be locked.
881 swap_pager_freespace(vm_object_t object, vm_pindex_t start, vm_size_t size)
884 swp_pager_meta_free(object, start, size);
888 * SWAP_PAGER_RESERVE() - reserve swap blocks in object
890 * Assigns swap blocks to the specified range within the object. The
891 * swap blocks are not zeroed. Any previous swap assignment is destroyed.
893 * Returns 0 on success, -1 on failure.
896 swap_pager_reserve(vm_object_t object, vm_pindex_t start, vm_size_t size)
898 daddr_t addr, blk, n_free, s_free;
901 swp_pager_init_freerange(&s_free, &n_free);
902 VM_OBJECT_WLOCK(object);
903 for (i = 0; i < size; i += n) {
904 n = min(BLIST_MAX_ALLOC, size - i);
905 blk = swp_pager_getswapspace(&n, 1);
906 if (blk == SWAPBLK_NONE) {
907 swp_pager_meta_free(object, start, i);
908 VM_OBJECT_WUNLOCK(object);
911 for (j = 0; j < n; ++j) {
912 addr = swp_pager_meta_build(object,
913 start + i + j, blk + j);
914 if (addr != SWAPBLK_NONE)
915 swp_pager_update_freerange(&s_free, &n_free,
919 swp_pager_freeswapspace(s_free, n_free);
920 VM_OBJECT_WUNLOCK(object);
925 * SWAP_PAGER_COPY() - copy blocks from source pager to destination pager
926 * and destroy the source.
928 * Copy any valid swapblks from the source to the destination. In
929 * cases where both the source and destination have a valid swapblk,
930 * we keep the destination's.
932 * This routine is allowed to sleep. It may sleep allocating metadata
933 * indirectly through swp_pager_meta_build() or if paging is still in
934 * progress on the source.
936 * The source object contains no vm_page_t's (which is just as well)
938 * The source object is of type OBJT_SWAP.
940 * The source and destination objects must be locked.
941 * Both object locks may temporarily be released.
944 swap_pager_copy(vm_object_t srcobject, vm_object_t dstobject,
945 vm_pindex_t offset, int destroysource)
948 daddr_t dstaddr, n_free, s_free, srcaddr;
950 VM_OBJECT_ASSERT_WLOCKED(srcobject);
951 VM_OBJECT_ASSERT_WLOCKED(dstobject);
954 * If destroysource is set, we remove the source object from the
955 * swap_pager internal queue now.
957 if (destroysource && srcobject->handle != NULL) {
958 vm_object_pip_add(srcobject, 1);
959 VM_OBJECT_WUNLOCK(srcobject);
960 vm_object_pip_add(dstobject, 1);
961 VM_OBJECT_WUNLOCK(dstobject);
962 sx_xlock(&sw_alloc_sx);
963 TAILQ_REMOVE(NOBJLIST(srcobject->handle), srcobject,
965 sx_xunlock(&sw_alloc_sx);
966 VM_OBJECT_WLOCK(dstobject);
967 vm_object_pip_wakeup(dstobject);
968 VM_OBJECT_WLOCK(srcobject);
969 vm_object_pip_wakeup(srcobject);
973 * Transfer source to destination.
975 swp_pager_init_freerange(&s_free, &n_free);
976 for (i = 0; i < dstobject->size; ++i) {
977 srcaddr = swp_pager_meta_ctl(srcobject, i + offset, SWM_POP);
978 if (srcaddr == SWAPBLK_NONE)
980 dstaddr = swp_pager_meta_ctl(dstobject, i, 0);
981 if (dstaddr != SWAPBLK_NONE) {
983 * Destination has valid swapblk or it is represented
984 * by a resident page. We destroy the source block.
986 swp_pager_update_freerange(&s_free, &n_free, srcaddr);
991 * Destination has no swapblk and is not resident,
994 * swp_pager_meta_build() can sleep.
996 vm_object_pip_add(srcobject, 1);
997 VM_OBJECT_WUNLOCK(srcobject);
998 vm_object_pip_add(dstobject, 1);
999 dstaddr = swp_pager_meta_build(dstobject, i, srcaddr);
1000 KASSERT(dstaddr == SWAPBLK_NONE,
1001 ("Unexpected destination swapblk"));
1002 vm_object_pip_wakeup(dstobject);
1003 VM_OBJECT_WLOCK(srcobject);
1004 vm_object_pip_wakeup(srcobject);
1006 swp_pager_freeswapspace(s_free, n_free);
1009 * Free left over swap blocks in source.
1011 * We have to revert the type to OBJT_DEFAULT so we do not accidentally
1012 * double-remove the object from the swap queues.
1014 if (destroysource) {
1015 swp_pager_meta_free_all(srcobject);
1017 * Reverting the type is not necessary, the caller is going
1018 * to destroy srcobject directly, but I'm doing it here
1019 * for consistency since we've removed the object from its
1022 srcobject->type = OBJT_DEFAULT;
1027 * SWAP_PAGER_HASPAGE() - determine if we have good backing store for
1028 * the requested page.
1030 * We determine whether good backing store exists for the requested
1031 * page and return TRUE if it does, FALSE if it doesn't.
1033 * If TRUE, we also try to determine how much valid, contiguous backing
1034 * store exists before and after the requested page.
1037 swap_pager_haspage(vm_object_t object, vm_pindex_t pindex, int *before,
1043 VM_OBJECT_ASSERT_LOCKED(object);
1046 * do we have good backing store at the requested index ?
1048 blk0 = swp_pager_meta_ctl(object, pindex, 0);
1049 if (blk0 == SWAPBLK_NONE) {
1058 * find backwards-looking contiguous good backing store
1060 if (before != NULL) {
1061 for (i = 1; i < SWB_NPAGES; i++) {
1064 blk = swp_pager_meta_ctl(object, pindex - i, 0);
1065 if (blk != blk0 - i)
1072 * find forward-looking contiguous good backing store
1074 if (after != NULL) {
1075 for (i = 1; i < SWB_NPAGES; i++) {
1076 blk = swp_pager_meta_ctl(object, pindex + i, 0);
1077 if (blk != blk0 + i)
1086 * SWAP_PAGER_PAGE_UNSWAPPED() - remove swap backing store related to page
1088 * This removes any associated swap backing store, whether valid or
1089 * not, from the page.
1091 * This routine is typically called when a page is made dirty, at
1092 * which point any associated swap can be freed. MADV_FREE also
1093 * calls us in a special-case situation
1095 * NOTE!!! If the page is clean and the swap was valid, the caller
1096 * should make the page dirty before calling this routine. This routine
1097 * does NOT change the m->dirty status of the page. Also: MADV_FREE
1100 * This routine may not sleep.
1102 * The object containing the page must be locked.
1105 swap_pager_unswapped(vm_page_t m)
1109 srcaddr = swp_pager_meta_ctl(m->object, m->pindex, SWM_POP);
1110 if (srcaddr != SWAPBLK_NONE)
1111 swp_pager_freeswapspace(srcaddr, 1);
1115 * swap_pager_getpages() - bring pages in from swap
1117 * Attempt to page in the pages in array "ma" of length "count". The
1118 * caller may optionally specify that additional pages preceding and
1119 * succeeding the specified range be paged in. The number of such pages
1120 * is returned in the "rbehind" and "rahead" parameters, and they will
1121 * be in the inactive queue upon return.
1123 * The pages in "ma" must be busied and will remain busied upon return.
1126 swap_pager_getpages(vm_object_t object, vm_page_t *ma, int count, int *rbehind,
1130 vm_page_t bm, mpred, msucc, p;
1133 int i, maxahead, maxbehind, reqcount;
1138 * Determine the final number of read-behind pages and
1139 * allocate them BEFORE releasing the object lock. Otherwise,
1140 * there can be a problematic race with vm_object_split().
1141 * Specifically, vm_object_split() might first transfer pages
1142 * that precede ma[0] in the current object to a new object,
1143 * and then this function incorrectly recreates those pages as
1144 * read-behind pages in the current object.
1146 if (!swap_pager_haspage(object, ma[0]->pindex, &maxbehind, &maxahead))
1147 return (VM_PAGER_FAIL);
1150 * Clip the readahead and readbehind ranges to exclude resident pages.
1152 if (rahead != NULL) {
1153 KASSERT(reqcount - 1 <= maxahead,
1154 ("page count %d extends beyond swap block", reqcount));
1155 *rahead = imin(*rahead, maxahead - (reqcount - 1));
1156 pindex = ma[reqcount - 1]->pindex;
1157 msucc = TAILQ_NEXT(ma[reqcount - 1], listq);
1158 if (msucc != NULL && msucc->pindex - pindex - 1 < *rahead)
1159 *rahead = msucc->pindex - pindex - 1;
1161 if (rbehind != NULL) {
1162 *rbehind = imin(*rbehind, maxbehind);
1163 pindex = ma[0]->pindex;
1164 mpred = TAILQ_PREV(ma[0], pglist, listq);
1165 if (mpred != NULL && pindex - mpred->pindex - 1 < *rbehind)
1166 *rbehind = pindex - mpred->pindex - 1;
1170 for (i = 0; i < count; i++)
1171 ma[i]->oflags |= VPO_SWAPINPROG;
1174 * Allocate readahead and readbehind pages.
1176 if (rbehind != NULL) {
1177 for (i = 1; i <= *rbehind; i++) {
1178 p = vm_page_alloc(object, ma[0]->pindex - i,
1182 p->oflags |= VPO_SWAPINPROG;
1187 if (rahead != NULL) {
1188 for (i = 0; i < *rahead; i++) {
1189 p = vm_page_alloc(object,
1190 ma[reqcount - 1]->pindex + i + 1, VM_ALLOC_NORMAL);
1193 p->oflags |= VPO_SWAPINPROG;
1197 if (rbehind != NULL)
1202 vm_object_pip_add(object, count);
1204 pindex = bm->pindex;
1205 blk = swp_pager_meta_ctl(object, pindex, 0);
1206 KASSERT(blk != SWAPBLK_NONE,
1207 ("no swap blocking containing %p(%jx)", object, (uintmax_t)pindex));
1209 VM_OBJECT_WUNLOCK(object);
1210 bp = uma_zalloc(swrbuf_zone, M_WAITOK);
1211 /* Pages cannot leave the object while busy. */
1212 for (i = 0, p = bm; i < count; i++, p = TAILQ_NEXT(p, listq)) {
1213 MPASS(p->pindex == bm->pindex + i);
1217 bp->b_flags |= B_PAGING;
1218 bp->b_iocmd = BIO_READ;
1219 bp->b_iodone = swp_pager_async_iodone;
1220 bp->b_rcred = crhold(thread0.td_ucred);
1221 bp->b_wcred = crhold(thread0.td_ucred);
1223 bp->b_bcount = PAGE_SIZE * count;
1224 bp->b_bufsize = PAGE_SIZE * count;
1225 bp->b_npages = count;
1226 bp->b_pgbefore = rbehind != NULL ? *rbehind : 0;
1227 bp->b_pgafter = rahead != NULL ? *rahead : 0;
1229 VM_CNT_INC(v_swapin);
1230 VM_CNT_ADD(v_swappgsin, count);
1233 * perform the I/O. NOTE!!! bp cannot be considered valid after
1234 * this point because we automatically release it on completion.
1235 * Instead, we look at the one page we are interested in which we
1236 * still hold a lock on even through the I/O completion.
1238 * The other pages in our ma[] array are also released on completion,
1239 * so we cannot assume they are valid anymore either.
1241 * NOTE: b_blkno is destroyed by the call to swapdev_strategy
1244 swp_pager_strategy(bp);
1247 * Wait for the pages we want to complete. VPO_SWAPINPROG is always
1248 * cleared on completion. If an I/O error occurs, SWAPBLK_NONE
1249 * is set in the metadata for each page in the request.
1251 VM_OBJECT_WLOCK(object);
1252 while ((ma[0]->oflags & VPO_SWAPINPROG) != 0) {
1253 ma[0]->oflags |= VPO_SWAPSLEEP;
1254 VM_CNT_INC(v_intrans);
1255 if (VM_OBJECT_SLEEP(object, &object->paging_in_progress, PSWP,
1256 "swread", hz * 20)) {
1258 "swap_pager: indefinite wait buffer: bufobj: %p, blkno: %jd, size: %ld\n",
1259 bp->b_bufobj, (intmax_t)bp->b_blkno, bp->b_bcount);
1264 * If we had an unrecoverable read error pages will not be valid.
1266 for (i = 0; i < reqcount; i++)
1267 if (ma[i]->valid != VM_PAGE_BITS_ALL)
1268 return (VM_PAGER_ERROR);
1270 return (VM_PAGER_OK);
1273 * A final note: in a low swap situation, we cannot deallocate swap
1274 * and mark a page dirty here because the caller is likely to mark
1275 * the page clean when we return, causing the page to possibly revert
1276 * to all-zero's later.
1281 * swap_pager_getpages_async():
1283 * Right now this is emulation of asynchronous operation on top of
1284 * swap_pager_getpages().
1287 swap_pager_getpages_async(vm_object_t object, vm_page_t *ma, int count,
1288 int *rbehind, int *rahead, pgo_getpages_iodone_t iodone, void *arg)
1292 r = swap_pager_getpages(object, ma, count, rbehind, rahead);
1293 VM_OBJECT_WUNLOCK(object);
1298 case VM_PAGER_ERROR:
1305 panic("unhandled swap_pager_getpages() error %d", r);
1307 (iodone)(arg, ma, count, error);
1308 VM_OBJECT_WLOCK(object);
1314 * swap_pager_putpages:
1316 * Assign swap (if necessary) and initiate I/O on the specified pages.
1318 * We support both OBJT_DEFAULT and OBJT_SWAP objects. DEFAULT objects
1319 * are automatically converted to SWAP objects.
1321 * In a low memory situation we may block in VOP_STRATEGY(), but the new
1322 * vm_page reservation system coupled with properly written VFS devices
1323 * should ensure that no low-memory deadlock occurs. This is an area
1326 * The parent has N vm_object_pip_add() references prior to
1327 * calling us and will remove references for rtvals[] that are
1328 * not set to VM_PAGER_PEND. We need to remove the rest on I/O
1331 * The parent has soft-busy'd the pages it passes us and will unbusy
1332 * those whos rtvals[] entry is not set to VM_PAGER_PEND on return.
1333 * We need to unbusy the rest on I/O completion.
1336 swap_pager_putpages(vm_object_t object, vm_page_t *ma, int count,
1337 int flags, int *rtvals)
1341 daddr_t addr, n_free, s_free;
1343 swp_pager_init_freerange(&s_free, &n_free);
1344 if (count && ma[0]->object != object) {
1345 panic("swap_pager_putpages: object mismatch %p/%p",
1354 * Turn object into OBJT_SWAP
1355 * check for bogus sysops
1356 * force sync if not pageout process
1358 if (object->type != OBJT_SWAP) {
1359 addr = swp_pager_meta_build(object, 0, SWAPBLK_NONE);
1360 KASSERT(addr == SWAPBLK_NONE,
1361 ("unexpected object swap block"));
1363 VM_OBJECT_WUNLOCK(object);
1366 if (curproc != pageproc)
1369 sync = (flags & VM_PAGER_PUT_SYNC) != 0;
1374 * Assign swap blocks and issue I/O. We reallocate swap on the fly.
1375 * The page is left dirty until the pageout operation completes
1378 for (i = 0; i < count; i += n) {
1384 * Maximum I/O size is limited by a number of factors.
1386 n = min(BLIST_MAX_ALLOC, count - i);
1387 n = min(n, nsw_cluster_max);
1389 /* Get a block of swap of size up to size n. */
1390 blk = swp_pager_getswapspace(&n, 4);
1391 if (blk == SWAPBLK_NONE) {
1392 for (j = 0; j < n; ++j)
1393 rtvals[i+j] = VM_PAGER_FAIL;
1398 * All I/O parameters have been satisfied, build the I/O
1399 * request and assign the swap space.
1402 mtx_lock(&swbuf_mtx);
1403 while (nsw_wcount_async == 0)
1404 msleep(&nsw_wcount_async, &swbuf_mtx, PVM,
1407 mtx_unlock(&swbuf_mtx);
1409 bp = uma_zalloc(swwbuf_zone, M_WAITOK);
1411 bp->b_flags = B_ASYNC;
1412 bp->b_flags |= B_PAGING;
1413 bp->b_iocmd = BIO_WRITE;
1415 bp->b_rcred = crhold(thread0.td_ucred);
1416 bp->b_wcred = crhold(thread0.td_ucred);
1417 bp->b_bcount = PAGE_SIZE * n;
1418 bp->b_bufsize = PAGE_SIZE * n;
1421 VM_OBJECT_WLOCK(object);
1422 for (j = 0; j < n; ++j) {
1423 vm_page_t mreq = ma[i+j];
1425 addr = swp_pager_meta_build(mreq->object, mreq->pindex,
1427 if (addr != SWAPBLK_NONE)
1428 swp_pager_update_freerange(&s_free, &n_free,
1430 MPASS(mreq->dirty == VM_PAGE_BITS_ALL);
1431 mreq->oflags |= VPO_SWAPINPROG;
1432 bp->b_pages[j] = mreq;
1434 VM_OBJECT_WUNLOCK(object);
1437 * Must set dirty range for NFS to work.
1440 bp->b_dirtyend = bp->b_bcount;
1442 VM_CNT_INC(v_swapout);
1443 VM_CNT_ADD(v_swappgsout, bp->b_npages);
1446 * We unconditionally set rtvals[] to VM_PAGER_PEND so that we
1447 * can call the async completion routine at the end of a
1448 * synchronous I/O operation. Otherwise, our caller would
1449 * perform duplicate unbusy and wakeup operations on the page
1450 * and object, respectively.
1452 for (j = 0; j < n; j++)
1453 rtvals[i + j] = VM_PAGER_PEND;
1458 * NOTE: b_blkno is destroyed by the call to swapdev_strategy
1460 if (sync == FALSE) {
1461 bp->b_iodone = swp_pager_async_iodone;
1463 swp_pager_strategy(bp);
1470 * NOTE: b_blkno is destroyed by the call to swapdev_strategy
1472 bp->b_iodone = bdone;
1473 swp_pager_strategy(bp);
1476 * Wait for the sync I/O to complete.
1478 bwait(bp, PVM, "swwrt");
1481 * Now that we are through with the bp, we can call the
1482 * normal async completion, which frees everything up.
1484 swp_pager_async_iodone(bp);
1486 VM_OBJECT_WLOCK(object);
1487 swp_pager_freeswapspace(s_free, n_free);
1491 * swp_pager_async_iodone:
1493 * Completion routine for asynchronous reads and writes from/to swap.
1494 * Also called manually by synchronous code to finish up a bp.
1496 * This routine may not sleep.
1499 swp_pager_async_iodone(struct buf *bp)
1502 vm_object_t object = NULL;
1505 * Report error - unless we ran out of memory, in which case
1506 * we've already logged it in swapgeom_strategy().
1508 if (bp->b_ioflags & BIO_ERROR && bp->b_error != ENOMEM) {
1510 "swap_pager: I/O error - %s failed; blkno %ld,"
1511 "size %ld, error %d\n",
1512 ((bp->b_iocmd == BIO_READ) ? "pagein" : "pageout"),
1520 * remove the mapping for kernel virtual
1523 pmap_qremove((vm_offset_t)bp->b_data, bp->b_npages);
1525 bp->b_data = bp->b_kvabase;
1528 object = bp->b_pages[0]->object;
1529 VM_OBJECT_WLOCK(object);
1533 * cleanup pages. If an error occurs writing to swap, we are in
1534 * very serious trouble. If it happens to be a disk error, though,
1535 * we may be able to recover by reassigning the swap later on. So
1536 * in this case we remove the m->swapblk assignment for the page
1537 * but do not free it in the rlist. The errornous block(s) are thus
1538 * never reallocated as swap. Redirty the page and continue.
1540 for (i = 0; i < bp->b_npages; ++i) {
1541 vm_page_t m = bp->b_pages[i];
1543 m->oflags &= ~VPO_SWAPINPROG;
1544 if (m->oflags & VPO_SWAPSLEEP) {
1545 m->oflags &= ~VPO_SWAPSLEEP;
1546 wakeup(&object->paging_in_progress);
1549 if (bp->b_ioflags & BIO_ERROR) {
1551 * If an error occurs I'd love to throw the swapblk
1552 * away without freeing it back to swapspace, so it
1553 * can never be used again. But I can't from an
1556 if (bp->b_iocmd == BIO_READ) {
1558 * NOTE: for reads, m->dirty will probably
1559 * be overridden by the original caller of
1560 * getpages so don't play cute tricks here.
1565 * If a write error occurs, reactivate page
1566 * so it doesn't clog the inactive list,
1567 * then finish the I/O.
1569 MPASS(m->dirty == VM_PAGE_BITS_ALL);
1571 vm_page_activate(m);
1575 } else if (bp->b_iocmd == BIO_READ) {
1577 * NOTE: for reads, m->dirty will probably be
1578 * overridden by the original caller of getpages so
1579 * we cannot set them in order to free the underlying
1580 * swap in a low-swap situation. I don't think we'd
1581 * want to do that anyway, but it was an optimization
1582 * that existed in the old swapper for a time before
1583 * it got ripped out due to precisely this problem.
1585 KASSERT(!pmap_page_is_mapped(m),
1586 ("swp_pager_async_iodone: page %p is mapped", m));
1587 KASSERT(m->dirty == 0,
1588 ("swp_pager_async_iodone: page %p is dirty", m));
1590 m->valid = VM_PAGE_BITS_ALL;
1591 if (i < bp->b_pgbefore ||
1592 i >= bp->b_npages - bp->b_pgafter)
1593 vm_page_readahead_finish(m);
1596 * For write success, clear the dirty
1597 * status, then finish the I/O ( which decrements the
1598 * busy count and possibly wakes waiter's up ).
1599 * A page is only written to swap after a period of
1600 * inactivity. Therefore, we do not expect it to be
1603 KASSERT(!pmap_page_is_write_mapped(m),
1604 ("swp_pager_async_iodone: page %p is not write"
1608 vm_page_deactivate_noreuse(m);
1615 * adjust pip. NOTE: the original parent may still have its own
1616 * pip refs on the object.
1618 if (object != NULL) {
1619 vm_object_pip_wakeupn(object, bp->b_npages);
1620 VM_OBJECT_WUNLOCK(object);
1624 * swapdev_strategy() manually sets b_vp and b_bufobj before calling
1625 * bstrategy(). Set them back to NULL now we're done with it, or we'll
1626 * trigger a KASSERT in relpbuf().
1630 bp->b_bufobj = NULL;
1633 * release the physical I/O buffer
1635 if (bp->b_flags & B_ASYNC) {
1636 mtx_lock(&swbuf_mtx);
1637 if (++nsw_wcount_async == 1)
1638 wakeup(&nsw_wcount_async);
1639 mtx_unlock(&swbuf_mtx);
1641 uma_zfree((bp->b_iocmd == BIO_READ) ? swrbuf_zone : swwbuf_zone, bp);
1645 swap_pager_nswapdev(void)
1652 * SWP_PAGER_FORCE_PAGEIN() - force a swap block to be paged in
1654 * This routine dissociates the page at the given index within an object
1655 * from its backing store, paging it in if it does not reside in memory.
1656 * If the page is paged in, it is marked dirty and placed in the laundry
1657 * queue. The page is marked dirty because it no longer has backing
1658 * store. It is placed in the laundry queue because it has not been
1659 * accessed recently. Otherwise, it would already reside in memory.
1661 * We also attempt to swap in all other pages in the swap block.
1662 * However, we only guarantee that the one at the specified index is
1665 * XXX - The code to page the whole block in doesn't work, so we
1666 * revert to the one-by-one behavior for now. Sigh.
1669 swp_pager_force_pagein(vm_object_t object, vm_pindex_t pindex)
1673 vm_object_pip_add(object, 1);
1674 m = vm_page_grab(object, pindex, VM_ALLOC_NORMAL);
1675 if (m->valid == VM_PAGE_BITS_ALL) {
1676 vm_object_pip_wakeup(object);
1680 if (m->wire_count == 0 && m->queue == PQ_NONE)
1681 panic("page %p is neither wired nor queued", m);
1685 vm_pager_page_unswapped(m);
1689 if (swap_pager_getpages(object, &m, 1, NULL, NULL) != VM_PAGER_OK)
1690 panic("swap_pager_force_pagein: read from swap failed");/*XXX*/
1691 vm_object_pip_wakeup(object);
1697 vm_pager_page_unswapped(m);
1701 * swap_pager_swapoff:
1703 * Page in all of the pages that have been paged out to the
1704 * given device. The corresponding blocks in the bitmap must be
1705 * marked as allocated and the device must be flagged SW_CLOSING.
1706 * There may be no processes swapped out to the device.
1708 * This routine may block.
1711 swap_pager_swapoff(struct swdevt *sp)
1718 sx_assert(&swdev_syscall_lock, SA_XLOCKED);
1722 mtx_lock(&vm_object_list_mtx);
1723 TAILQ_FOREACH(object, &vm_object_list, object_list) {
1724 if (object->type != OBJT_SWAP)
1726 mtx_unlock(&vm_object_list_mtx);
1727 /* Depends on type-stability. */
1728 VM_OBJECT_WLOCK(object);
1731 * Dead objects are eventually terminated on their own.
1733 if ((object->flags & OBJ_DEAD) != 0)
1737 * Sync with fences placed after pctrie
1738 * initialization. We must not access pctrie below
1739 * unless we checked that our object is swap and not
1742 atomic_thread_fence_acq();
1743 if (object->type != OBJT_SWAP)
1746 for (pi = 0; (sb = SWAP_PCTRIE_LOOKUP_GE(
1747 &object->un_pager.swp.swp_blks, pi)) != NULL; ) {
1748 pi = sb->p + SWAP_META_PAGES;
1749 for (i = 0; i < SWAP_META_PAGES; i++) {
1750 if (sb->d[i] == SWAPBLK_NONE)
1752 if (swp_pager_isondev(sb->d[i], sp))
1753 swp_pager_force_pagein(object,
1758 VM_OBJECT_WUNLOCK(object);
1759 mtx_lock(&vm_object_list_mtx);
1761 mtx_unlock(&vm_object_list_mtx);
1765 * Objects may be locked or paging to the device being
1766 * removed, so we will miss their pages and need to
1767 * make another pass. We have marked this device as
1768 * SW_CLOSING, so the activity should finish soon.
1771 if (retries > 100) {
1772 panic("swapoff: failed to locate %d swap blocks",
1775 pause("swpoff", hz / 20);
1778 EVENTHANDLER_INVOKE(swapoff, sp);
1781 /************************************************************************
1783 ************************************************************************
1785 * These routines manipulate the swap metadata stored in the
1788 * Swap metadata is implemented with a global hash and not directly
1789 * linked into the object. Instead the object simply contains
1790 * appropriate tracking counters.
1794 * SWP_PAGER_SWBLK_EMPTY() - is a range of blocks free?
1797 swp_pager_swblk_empty(struct swblk *sb, int start, int limit)
1801 MPASS(0 <= start && start <= limit && limit <= SWAP_META_PAGES);
1802 for (i = start; i < limit; i++) {
1803 if (sb->d[i] != SWAPBLK_NONE)
1810 * SWP_PAGER_META_BUILD() - add swap block to swap meta data for object
1812 * We first convert the object to a swap object if it is a default
1815 * The specified swapblk is added to the object's swap metadata. If
1816 * the swapblk is not valid, it is freed instead. Any previously
1817 * assigned swapblk is returned.
1820 swp_pager_meta_build(vm_object_t object, vm_pindex_t pindex, daddr_t swapblk)
1822 static volatile int swblk_zone_exhausted, swpctrie_zone_exhausted;
1823 struct swblk *sb, *sb1;
1824 vm_pindex_t modpi, rdpi;
1825 daddr_t prev_swapblk;
1828 VM_OBJECT_ASSERT_WLOCKED(object);
1831 * Convert default object to swap object if necessary
1833 if (object->type != OBJT_SWAP) {
1834 pctrie_init(&object->un_pager.swp.swp_blks);
1837 * Ensure that swap_pager_swapoff()'s iteration over
1838 * object_list does not see a garbage pctrie.
1840 atomic_thread_fence_rel();
1842 object->type = OBJT_SWAP;
1843 KASSERT(object->handle == NULL, ("default pager with handle"));
1846 rdpi = rounddown(pindex, SWAP_META_PAGES);
1847 sb = SWAP_PCTRIE_LOOKUP(&object->un_pager.swp.swp_blks, rdpi);
1849 if (swapblk == SWAPBLK_NONE)
1850 return (SWAPBLK_NONE);
1852 sb = uma_zalloc(swblk_zone, M_NOWAIT | (curproc ==
1853 pageproc ? M_USE_RESERVE : 0));
1856 for (i = 0; i < SWAP_META_PAGES; i++)
1857 sb->d[i] = SWAPBLK_NONE;
1858 if (atomic_cmpset_int(&swblk_zone_exhausted,
1860 printf("swblk zone ok\n");
1863 VM_OBJECT_WUNLOCK(object);
1864 if (uma_zone_exhausted(swblk_zone)) {
1865 if (atomic_cmpset_int(&swblk_zone_exhausted,
1867 printf("swap blk zone exhausted, "
1868 "increase kern.maxswzone\n");
1869 vm_pageout_oom(VM_OOM_SWAPZ);
1870 pause("swzonxb", 10);
1872 uma_zwait(swblk_zone);
1873 VM_OBJECT_WLOCK(object);
1874 sb = SWAP_PCTRIE_LOOKUP(&object->un_pager.swp.swp_blks,
1878 * Somebody swapped out a nearby page,
1879 * allocating swblk at the rdpi index,
1880 * while we dropped the object lock.
1885 error = SWAP_PCTRIE_INSERT(
1886 &object->un_pager.swp.swp_blks, sb);
1888 if (atomic_cmpset_int(&swpctrie_zone_exhausted,
1890 printf("swpctrie zone ok\n");
1893 VM_OBJECT_WUNLOCK(object);
1894 if (uma_zone_exhausted(swpctrie_zone)) {
1895 if (atomic_cmpset_int(&swpctrie_zone_exhausted,
1897 printf("swap pctrie zone exhausted, "
1898 "increase kern.maxswzone\n");
1899 vm_pageout_oom(VM_OOM_SWAPZ);
1900 pause("swzonxp", 10);
1902 uma_zwait(swpctrie_zone);
1903 VM_OBJECT_WLOCK(object);
1904 sb1 = SWAP_PCTRIE_LOOKUP(&object->un_pager.swp.swp_blks,
1907 uma_zfree(swblk_zone, sb);
1914 MPASS(sb->p == rdpi);
1916 modpi = pindex % SWAP_META_PAGES;
1917 /* Return prior contents of metadata. */
1918 prev_swapblk = sb->d[modpi];
1919 /* Enter block into metadata. */
1920 sb->d[modpi] = swapblk;
1923 * Free the swblk if we end up with the empty page run.
1925 if (swapblk == SWAPBLK_NONE &&
1926 swp_pager_swblk_empty(sb, 0, SWAP_META_PAGES)) {
1927 SWAP_PCTRIE_REMOVE(&object->un_pager.swp.swp_blks, rdpi);
1928 uma_zfree(swblk_zone, sb);
1930 return (prev_swapblk);
1934 * SWP_PAGER_META_FREE() - free a range of blocks in the object's swap metadata
1936 * The requested range of blocks is freed, with any associated swap
1937 * returned to the swap bitmap.
1939 * This routine will free swap metadata structures as they are cleaned
1940 * out. This routine does *NOT* operate on swap metadata associated
1941 * with resident pages.
1944 swp_pager_meta_free(vm_object_t object, vm_pindex_t pindex, vm_pindex_t count)
1947 daddr_t n_free, s_free;
1949 int i, limit, start;
1951 VM_OBJECT_ASSERT_WLOCKED(object);
1952 if (object->type != OBJT_SWAP || count == 0)
1955 swp_pager_init_freerange(&s_free, &n_free);
1956 last = pindex + count;
1958 sb = SWAP_PCTRIE_LOOKUP_GE(&object->un_pager.swp.swp_blks,
1959 rounddown(pindex, SWAP_META_PAGES));
1960 if (sb == NULL || sb->p >= last)
1962 start = pindex > sb->p ? pindex - sb->p : 0;
1963 limit = last - sb->p < SWAP_META_PAGES ? last - sb->p :
1965 for (i = start; i < limit; i++) {
1966 if (sb->d[i] == SWAPBLK_NONE)
1968 swp_pager_update_freerange(&s_free, &n_free, sb->d[i]);
1969 sb->d[i] = SWAPBLK_NONE;
1971 pindex = sb->p + SWAP_META_PAGES;
1972 if (swp_pager_swblk_empty(sb, 0, start) &&
1973 swp_pager_swblk_empty(sb, limit, SWAP_META_PAGES)) {
1974 SWAP_PCTRIE_REMOVE(&object->un_pager.swp.swp_blks,
1976 uma_zfree(swblk_zone, sb);
1979 swp_pager_freeswapspace(s_free, n_free);
1983 * SWP_PAGER_META_FREE_ALL() - destroy all swap metadata associated with object
1985 * This routine locates and destroys all swap metadata associated with
1989 swp_pager_meta_free_all(vm_object_t object)
1992 daddr_t n_free, s_free;
1996 VM_OBJECT_ASSERT_WLOCKED(object);
1997 if (object->type != OBJT_SWAP)
2000 swp_pager_init_freerange(&s_free, &n_free);
2001 for (pindex = 0; (sb = SWAP_PCTRIE_LOOKUP_GE(
2002 &object->un_pager.swp.swp_blks, pindex)) != NULL;) {
2003 pindex = sb->p + SWAP_META_PAGES;
2004 for (i = 0; i < SWAP_META_PAGES; i++) {
2005 if (sb->d[i] == SWAPBLK_NONE)
2007 swp_pager_update_freerange(&s_free, &n_free, sb->d[i]);
2009 SWAP_PCTRIE_REMOVE(&object->un_pager.swp.swp_blks, sb->p);
2010 uma_zfree(swblk_zone, sb);
2012 swp_pager_freeswapspace(s_free, n_free);
2016 * SWP_PAGER_METACTL() - misc control of swap meta data.
2018 * This routine is capable of looking up, or removing swapblk
2019 * assignments in the swap meta data. It returns the swapblk being
2020 * looked-up, popped, or SWAPBLK_NONE if the block was invalid.
2022 * When acting on a busy resident page and paging is in progress, we
2023 * have to wait until paging is complete but otherwise can act on the
2026 * SWM_POP remove from meta data but do not free it
2029 swp_pager_meta_ctl(vm_object_t object, vm_pindex_t pindex, int flags)
2034 if ((flags & SWM_POP) != 0)
2035 VM_OBJECT_ASSERT_WLOCKED(object);
2037 VM_OBJECT_ASSERT_LOCKED(object);
2040 * The meta data only exists if the object is OBJT_SWAP
2041 * and even then might not be allocated yet.
2043 if (object->type != OBJT_SWAP)
2044 return (SWAPBLK_NONE);
2046 sb = SWAP_PCTRIE_LOOKUP(&object->un_pager.swp.swp_blks,
2047 rounddown(pindex, SWAP_META_PAGES));
2049 return (SWAPBLK_NONE);
2050 r1 = sb->d[pindex % SWAP_META_PAGES];
2051 if (r1 == SWAPBLK_NONE)
2052 return (SWAPBLK_NONE);
2053 if ((flags & SWM_POP) != 0) {
2054 sb->d[pindex % SWAP_META_PAGES] = SWAPBLK_NONE;
2055 if (swp_pager_swblk_empty(sb, 0, SWAP_META_PAGES)) {
2056 SWAP_PCTRIE_REMOVE(&object->un_pager.swp.swp_blks,
2057 rounddown(pindex, SWAP_META_PAGES));
2058 uma_zfree(swblk_zone, sb);
2065 * Returns the least page index which is greater than or equal to the
2066 * parameter pindex and for which there is a swap block allocated.
2067 * Returns object's size if the object's type is not swap or if there
2068 * are no allocated swap blocks for the object after the requested
2072 swap_pager_find_least(vm_object_t object, vm_pindex_t pindex)
2077 VM_OBJECT_ASSERT_LOCKED(object);
2078 if (object->type != OBJT_SWAP)
2079 return (object->size);
2081 sb = SWAP_PCTRIE_LOOKUP_GE(&object->un_pager.swp.swp_blks,
2082 rounddown(pindex, SWAP_META_PAGES));
2084 return (object->size);
2085 if (sb->p < pindex) {
2086 for (i = pindex % SWAP_META_PAGES; i < SWAP_META_PAGES; i++) {
2087 if (sb->d[i] != SWAPBLK_NONE)
2090 sb = SWAP_PCTRIE_LOOKUP_GE(&object->un_pager.swp.swp_blks,
2091 roundup(pindex, SWAP_META_PAGES));
2093 return (object->size);
2095 for (i = 0; i < SWAP_META_PAGES; i++) {
2096 if (sb->d[i] != SWAPBLK_NONE)
2101 * We get here if a swblk is present in the trie but it
2102 * doesn't map any blocks.
2105 return (object->size);
2109 * System call swapon(name) enables swapping on device name,
2110 * which must be in the swdevsw. Return EBUSY
2111 * if already swapping on this device.
2113 #ifndef _SYS_SYSPROTO_H_
2114 struct swapon_args {
2124 sys_swapon(struct thread *td, struct swapon_args *uap)
2128 struct nameidata nd;
2131 error = priv_check(td, PRIV_SWAPON);
2135 sx_xlock(&swdev_syscall_lock);
2138 * Swap metadata may not fit in the KVM if we have physical
2141 if (swblk_zone == NULL) {
2146 NDINIT(&nd, LOOKUP, ISOPEN | FOLLOW | AUDITVNODE1, UIO_USERSPACE,
2152 NDFREE(&nd, NDF_ONLY_PNBUF);
2155 if (vn_isdisk(vp, &error)) {
2156 error = swapongeom(vp);
2157 } else if (vp->v_type == VREG &&
2158 (vp->v_mount->mnt_vfc->vfc_flags & VFCF_NETWORK) != 0 &&
2159 (error = VOP_GETATTR(vp, &attr, td->td_ucred)) == 0) {
2161 * Allow direct swapping to NFS regular files in the same
2162 * way that nfs_mountroot() sets up diskless swapping.
2164 error = swaponvp(td, vp, attr.va_size / DEV_BSIZE);
2170 sx_xunlock(&swdev_syscall_lock);
2175 * Check that the total amount of swap currently configured does not
2176 * exceed half the theoretical maximum. If it does, print a warning
2180 swapon_check_swzone(void)
2182 unsigned long maxpages, npages;
2184 npages = swap_total;
2185 /* absolute maximum we can handle assuming 100% efficiency */
2186 maxpages = uma_zone_get_max(swblk_zone) * SWAP_META_PAGES;
2188 /* recommend using no more than half that amount */
2189 if (npages > maxpages / 2) {
2190 printf("warning: total configured swap (%lu pages) "
2191 "exceeds maximum recommended amount (%lu pages).\n",
2192 npages, maxpages / 2);
2193 printf("warning: increase kern.maxswzone "
2194 "or reduce amount of swap.\n");
2199 swaponsomething(struct vnode *vp, void *id, u_long nblks,
2200 sw_strategy_t *strategy, sw_close_t *close, dev_t dev, int flags)
2202 struct swdevt *sp, *tsp;
2207 * nblks is in DEV_BSIZE'd chunks, convert to PAGE_SIZE'd chunks.
2208 * First chop nblks off to page-align it, then convert.
2210 * sw->sw_nblks is in page-sized chunks now too.
2212 nblks &= ~(ctodb(1) - 1);
2213 nblks = dbtoc(nblks);
2216 * If we go beyond this, we get overflows in the radix
2219 mblocks = 0x40000000 / BLIST_META_RADIX;
2220 if (nblks > mblocks) {
2222 "WARNING: reducing swap size to maximum of %luMB per unit\n",
2223 mblocks / 1024 / 1024 * PAGE_SIZE);
2227 sp = malloc(sizeof *sp, M_VMPGDATA, M_WAITOK | M_ZERO);
2232 sp->sw_nblks = nblks;
2234 sp->sw_strategy = strategy;
2235 sp->sw_close = close;
2236 sp->sw_flags = flags;
2238 sp->sw_blist = blist_create(nblks, M_WAITOK);
2240 * Do not free the first two block in order to avoid overwriting
2241 * any bsd label at the front of the partition
2243 blist_free(sp->sw_blist, 2, nblks - 2);
2246 mtx_lock(&sw_dev_mtx);
2247 TAILQ_FOREACH(tsp, &swtailq, sw_list) {
2248 if (tsp->sw_end >= dvbase) {
2250 * We put one uncovered page between the devices
2251 * in order to definitively prevent any cross-device
2254 dvbase = tsp->sw_end + 1;
2257 sp->sw_first = dvbase;
2258 sp->sw_end = dvbase + nblks;
2259 TAILQ_INSERT_TAIL(&swtailq, sp, sw_list);
2261 swap_pager_avail += nblks - 2;
2262 swap_total += nblks;
2263 swapon_check_swzone();
2265 mtx_unlock(&sw_dev_mtx);
2266 EVENTHANDLER_INVOKE(swapon, sp);
2270 * SYSCALL: swapoff(devname)
2272 * Disable swapping on the given device.
2274 * XXX: Badly designed system call: it should use a device index
2275 * rather than filename as specification. We keep sw_vp around
2276 * only to make this work.
2278 #ifndef _SYS_SYSPROTO_H_
2279 struct swapoff_args {
2289 sys_swapoff(struct thread *td, struct swapoff_args *uap)
2292 struct nameidata nd;
2296 error = priv_check(td, PRIV_SWAPOFF);
2300 sx_xlock(&swdev_syscall_lock);
2302 NDINIT(&nd, LOOKUP, FOLLOW | AUDITVNODE1, UIO_USERSPACE, uap->name,
2307 NDFREE(&nd, NDF_ONLY_PNBUF);
2310 mtx_lock(&sw_dev_mtx);
2311 TAILQ_FOREACH(sp, &swtailq, sw_list) {
2312 if (sp->sw_vp == vp)
2315 mtx_unlock(&sw_dev_mtx);
2320 error = swapoff_one(sp, td->td_ucred);
2322 sx_xunlock(&swdev_syscall_lock);
2327 swapoff_one(struct swdevt *sp, struct ucred *cred)
2334 sx_assert(&swdev_syscall_lock, SA_XLOCKED);
2336 (void) vn_lock(sp->sw_vp, LK_EXCLUSIVE | LK_RETRY);
2337 error = mac_system_check_swapoff(cred, sp->sw_vp);
2338 (void) VOP_UNLOCK(sp->sw_vp, 0);
2342 nblks = sp->sw_nblks;
2345 * We can turn off this swap device safely only if the
2346 * available virtual memory in the system will fit the amount
2347 * of data we will have to page back in, plus an epsilon so
2348 * the system doesn't become critically low on swap space.
2350 if (vm_free_count() + swap_pager_avail < nblks + nswap_lowat)
2354 * Prevent further allocations on this device.
2356 mtx_lock(&sw_dev_mtx);
2357 sp->sw_flags |= SW_CLOSING;
2358 swap_pager_avail -= blist_fill(sp->sw_blist, 0, nblks);
2359 swap_total -= nblks;
2360 mtx_unlock(&sw_dev_mtx);
2363 * Page in the contents of the device and close it.
2365 swap_pager_swapoff(sp);
2367 sp->sw_close(curthread, sp);
2368 mtx_lock(&sw_dev_mtx);
2370 TAILQ_REMOVE(&swtailq, sp, sw_list);
2372 if (nswapdev == 0) {
2373 swap_pager_full = 2;
2374 swap_pager_almost_full = 1;
2378 mtx_unlock(&sw_dev_mtx);
2379 blist_destroy(sp->sw_blist);
2380 free(sp, M_VMPGDATA);
2387 struct swdevt *sp, *spt;
2388 const char *devname;
2391 sx_xlock(&swdev_syscall_lock);
2393 mtx_lock(&sw_dev_mtx);
2394 TAILQ_FOREACH_SAFE(sp, &swtailq, sw_list, spt) {
2395 mtx_unlock(&sw_dev_mtx);
2396 if (vn_isdisk(sp->sw_vp, NULL))
2397 devname = devtoname(sp->sw_vp->v_rdev);
2400 error = swapoff_one(sp, thread0.td_ucred);
2402 printf("Cannot remove swap device %s (error=%d), "
2403 "skipping.\n", devname, error);
2404 } else if (bootverbose) {
2405 printf("Swap device %s removed.\n", devname);
2407 mtx_lock(&sw_dev_mtx);
2409 mtx_unlock(&sw_dev_mtx);
2411 sx_xunlock(&swdev_syscall_lock);
2415 swap_pager_status(int *total, int *used)
2421 mtx_lock(&sw_dev_mtx);
2422 TAILQ_FOREACH(sp, &swtailq, sw_list) {
2423 *total += sp->sw_nblks;
2424 *used += sp->sw_used;
2426 mtx_unlock(&sw_dev_mtx);
2430 swap_dev_info(int name, struct xswdev *xs, char *devname, size_t len)
2433 const char *tmp_devname;
2438 mtx_lock(&sw_dev_mtx);
2439 TAILQ_FOREACH(sp, &swtailq, sw_list) {
2444 xs->xsw_version = XSWDEV_VERSION;
2445 xs->xsw_dev = sp->sw_dev;
2446 xs->xsw_flags = sp->sw_flags;
2447 xs->xsw_nblks = sp->sw_nblks;
2448 xs->xsw_used = sp->sw_used;
2449 if (devname != NULL) {
2450 if (vn_isdisk(sp->sw_vp, NULL))
2451 tmp_devname = devtoname(sp->sw_vp->v_rdev);
2453 tmp_devname = "[file]";
2454 strncpy(devname, tmp_devname, len);
2459 mtx_unlock(&sw_dev_mtx);
2463 #if defined(COMPAT_FREEBSD11)
2464 #define XSWDEV_VERSION_11 1
2474 #if defined(__amd64__) && defined(COMPAT_FREEBSD32)
2477 u_int xsw_dev1, xsw_dev2;
2485 sysctl_vm_swap_info(SYSCTL_HANDLER_ARGS)
2488 #if defined(__amd64__) && defined(COMPAT_FREEBSD32)
2489 struct xswdev32 xs32;
2491 #if defined(COMPAT_FREEBSD11)
2492 struct xswdev11 xs11;
2496 if (arg2 != 1) /* name length */
2498 error = swap_dev_info(*(int *)arg1, &xs, NULL, 0);
2501 #if defined(__amd64__) && defined(COMPAT_FREEBSD32)
2502 if (req->oldlen == sizeof(xs32)) {
2503 xs32.xsw_version = XSWDEV_VERSION;
2504 xs32.xsw_dev1 = xs.xsw_dev;
2505 xs32.xsw_dev2 = xs.xsw_dev >> 32;
2506 xs32.xsw_flags = xs.xsw_flags;
2507 xs32.xsw_nblks = xs.xsw_nblks;
2508 xs32.xsw_used = xs.xsw_used;
2509 error = SYSCTL_OUT(req, &xs32, sizeof(xs32));
2513 #if defined(COMPAT_FREEBSD11)
2514 if (req->oldlen == sizeof(xs11)) {
2515 xs11.xsw_version = XSWDEV_VERSION_11;
2516 xs11.xsw_dev = xs.xsw_dev; /* truncation */
2517 xs11.xsw_flags = xs.xsw_flags;
2518 xs11.xsw_nblks = xs.xsw_nblks;
2519 xs11.xsw_used = xs.xsw_used;
2520 error = SYSCTL_OUT(req, &xs11, sizeof(xs11));
2524 error = SYSCTL_OUT(req, &xs, sizeof(xs));
2528 SYSCTL_INT(_vm, OID_AUTO, nswapdev, CTLFLAG_RD, &nswapdev, 0,
2529 "Number of swap devices");
2530 SYSCTL_NODE(_vm, OID_AUTO, swap_info, CTLFLAG_RD | CTLFLAG_MPSAFE,
2531 sysctl_vm_swap_info,
2532 "Swap statistics by device");
2535 * Count the approximate swap usage in pages for a vmspace. The
2536 * shadowed or not yet copied on write swap blocks are not accounted.
2537 * The map must be locked.
2540 vmspace_swap_count(struct vmspace *vmspace)
2550 map = &vmspace->vm_map;
2553 for (cur = map->header.next; cur != &map->header; cur = cur->next) {
2554 if ((cur->eflags & MAP_ENTRY_IS_SUB_MAP) != 0)
2556 object = cur->object.vm_object;
2557 if (object == NULL || object->type != OBJT_SWAP)
2559 VM_OBJECT_RLOCK(object);
2560 if (object->type != OBJT_SWAP)
2562 pi = OFF_TO_IDX(cur->offset);
2563 e = pi + OFF_TO_IDX(cur->end - cur->start);
2564 for (;; pi = sb->p + SWAP_META_PAGES) {
2565 sb = SWAP_PCTRIE_LOOKUP_GE(
2566 &object->un_pager.swp.swp_blks, pi);
2567 if (sb == NULL || sb->p >= e)
2569 for (i = 0; i < SWAP_META_PAGES; i++) {
2570 if (sb->p + i < e &&
2571 sb->d[i] != SWAPBLK_NONE)
2576 VM_OBJECT_RUNLOCK(object);
2584 * Swapping onto disk devices.
2588 static g_orphan_t swapgeom_orphan;
2590 static struct g_class g_swap_class = {
2592 .version = G_VERSION,
2593 .orphan = swapgeom_orphan,
2596 DECLARE_GEOM_CLASS(g_swap_class, g_class);
2600 swapgeom_close_ev(void *arg, int flags)
2602 struct g_consumer *cp;
2605 g_access(cp, -1, -1, 0);
2607 g_destroy_consumer(cp);
2611 * Add a reference to the g_consumer for an inflight transaction.
2614 swapgeom_acquire(struct g_consumer *cp)
2617 mtx_assert(&sw_dev_mtx, MA_OWNED);
2622 * Remove a reference from the g_consumer. Post a close event if all
2623 * references go away, since the function might be called from the
2627 swapgeom_release(struct g_consumer *cp, struct swdevt *sp)
2630 mtx_assert(&sw_dev_mtx, MA_OWNED);
2632 if (cp->index == 0) {
2633 if (g_post_event(swapgeom_close_ev, cp, M_NOWAIT, NULL) == 0)
2639 swapgeom_done(struct bio *bp2)
2643 struct g_consumer *cp;
2645 bp = bp2->bio_caller2;
2647 bp->b_ioflags = bp2->bio_flags;
2649 bp->b_ioflags |= BIO_ERROR;
2650 bp->b_resid = bp->b_bcount - bp2->bio_completed;
2651 bp->b_error = bp2->bio_error;
2652 bp->b_caller1 = NULL;
2654 sp = bp2->bio_caller1;
2655 mtx_lock(&sw_dev_mtx);
2656 swapgeom_release(cp, sp);
2657 mtx_unlock(&sw_dev_mtx);
2662 swapgeom_strategy(struct buf *bp, struct swdevt *sp)
2665 struct g_consumer *cp;
2667 mtx_lock(&sw_dev_mtx);
2670 mtx_unlock(&sw_dev_mtx);
2671 bp->b_error = ENXIO;
2672 bp->b_ioflags |= BIO_ERROR;
2676 swapgeom_acquire(cp);
2677 mtx_unlock(&sw_dev_mtx);
2678 if (bp->b_iocmd == BIO_WRITE)
2681 bio = g_alloc_bio();
2683 mtx_lock(&sw_dev_mtx);
2684 swapgeom_release(cp, sp);
2685 mtx_unlock(&sw_dev_mtx);
2686 bp->b_error = ENOMEM;
2687 bp->b_ioflags |= BIO_ERROR;
2688 printf("swap_pager: cannot allocate bio\n");
2693 bp->b_caller1 = bio;
2694 bio->bio_caller1 = sp;
2695 bio->bio_caller2 = bp;
2696 bio->bio_cmd = bp->b_iocmd;
2697 bio->bio_offset = (bp->b_blkno - sp->sw_first) * PAGE_SIZE;
2698 bio->bio_length = bp->b_bcount;
2699 bio->bio_done = swapgeom_done;
2700 if (!buf_mapped(bp)) {
2701 bio->bio_ma = bp->b_pages;
2702 bio->bio_data = unmapped_buf;
2703 bio->bio_ma_offset = (vm_offset_t)bp->b_offset & PAGE_MASK;
2704 bio->bio_ma_n = bp->b_npages;
2705 bio->bio_flags |= BIO_UNMAPPED;
2707 bio->bio_data = bp->b_data;
2710 g_io_request(bio, cp);
2715 swapgeom_orphan(struct g_consumer *cp)
2720 mtx_lock(&sw_dev_mtx);
2721 TAILQ_FOREACH(sp, &swtailq, sw_list) {
2722 if (sp->sw_id == cp) {
2723 sp->sw_flags |= SW_CLOSING;
2728 * Drop reference we were created with. Do directly since we're in a
2729 * special context where we don't have to queue the call to
2730 * swapgeom_close_ev().
2733 destroy = ((sp != NULL) && (cp->index == 0));
2736 mtx_unlock(&sw_dev_mtx);
2738 swapgeom_close_ev(cp, 0);
2742 swapgeom_close(struct thread *td, struct swdevt *sw)
2744 struct g_consumer *cp;
2746 mtx_lock(&sw_dev_mtx);
2749 mtx_unlock(&sw_dev_mtx);
2752 * swapgeom_close() may be called from the biodone context,
2753 * where we cannot perform topology changes. Delegate the
2754 * work to the events thread.
2757 g_waitfor_event(swapgeom_close_ev, cp, M_WAITOK, NULL);
2761 swapongeom_locked(struct cdev *dev, struct vnode *vp)
2763 struct g_provider *pp;
2764 struct g_consumer *cp;
2765 static struct g_geom *gp;
2770 pp = g_dev_getprovider(dev);
2773 mtx_lock(&sw_dev_mtx);
2774 TAILQ_FOREACH(sp, &swtailq, sw_list) {
2776 if (cp != NULL && cp->provider == pp) {
2777 mtx_unlock(&sw_dev_mtx);
2781 mtx_unlock(&sw_dev_mtx);
2783 gp = g_new_geomf(&g_swap_class, "swap");
2784 cp = g_new_consumer(gp);
2785 cp->index = 1; /* Number of active I/Os, plus one for being active. */
2786 cp->flags |= G_CF_DIRECT_SEND | G_CF_DIRECT_RECEIVE;
2789 * XXX: Every time you think you can improve the margin for
2790 * footshooting, somebody depends on the ability to do so:
2791 * savecore(8) wants to write to our swapdev so we cannot
2792 * set an exclusive count :-(
2794 error = g_access(cp, 1, 1, 0);
2797 g_destroy_consumer(cp);
2800 nblks = pp->mediasize / DEV_BSIZE;
2801 swaponsomething(vp, cp, nblks, swapgeom_strategy,
2802 swapgeom_close, dev2udev(dev),
2803 (pp->flags & G_PF_ACCEPT_UNMAPPED) != 0 ? SW_UNMAPPED : 0);
2808 swapongeom(struct vnode *vp)
2812 vn_lock(vp, LK_EXCLUSIVE | LK_RETRY);
2813 if (vp->v_type != VCHR || (vp->v_iflag & VI_DOOMED) != 0) {
2817 error = swapongeom_locked(vp->v_rdev, vp);
2818 g_topology_unlock();
2827 * This is used mainly for network filesystem (read: probably only tested
2828 * with NFS) swapfiles.
2833 swapdev_strategy(struct buf *bp, struct swdevt *sp)
2837 bp->b_blkno = ctodb(bp->b_blkno - sp->sw_first);
2841 if (bp->b_iocmd == BIO_WRITE) {
2843 bufobj_wdrop(bp->b_bufobj);
2844 bufobj_wref(&vp2->v_bufobj);
2846 if (bp->b_bufobj != &vp2->v_bufobj)
2847 bp->b_bufobj = &vp2->v_bufobj;
2849 bp->b_iooffset = dbtob(bp->b_blkno);
2855 swapdev_close(struct thread *td, struct swdevt *sp)
2858 VOP_CLOSE(sp->sw_vp, FREAD | FWRITE, td->td_ucred, td);
2864 swaponvp(struct thread *td, struct vnode *vp, u_long nblks)
2871 mtx_lock(&sw_dev_mtx);
2872 TAILQ_FOREACH(sp, &swtailq, sw_list) {
2873 if (sp->sw_id == vp) {
2874 mtx_unlock(&sw_dev_mtx);
2878 mtx_unlock(&sw_dev_mtx);
2880 (void) vn_lock(vp, LK_EXCLUSIVE | LK_RETRY);
2882 error = mac_system_check_swapon(td->td_ucred, vp);
2885 error = VOP_OPEN(vp, FREAD | FWRITE, td->td_ucred, td, NULL);
2886 (void) VOP_UNLOCK(vp, 0);
2890 swaponsomething(vp, vp, nblks, swapdev_strategy, swapdev_close,
2896 sysctl_swap_async_max(SYSCTL_HANDLER_ARGS)
2900 new = nsw_wcount_async_max;
2901 error = sysctl_handle_int(oidp, &new, 0, req);
2902 if (error != 0 || req->newptr == NULL)
2905 if (new > nswbuf / 2 || new < 1)
2908 mtx_lock(&swbuf_mtx);
2909 while (nsw_wcount_async_max != new) {
2911 * Adjust difference. If the current async count is too low,
2912 * we will need to sqeeze our update slowly in. Sleep with a
2913 * higher priority than getpbuf() to finish faster.
2915 n = new - nsw_wcount_async_max;
2916 if (nsw_wcount_async + n >= 0) {
2917 nsw_wcount_async += n;
2918 nsw_wcount_async_max += n;
2919 wakeup(&nsw_wcount_async);
2921 nsw_wcount_async_max -= nsw_wcount_async;
2922 nsw_wcount_async = 0;
2923 msleep(&nsw_wcount_async, &swbuf_mtx, PSWP,
2927 mtx_unlock(&swbuf_mtx);