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>
78 #include <sys/blist.h>
82 #include <sys/eventhandler.h>
83 #include <sys/fcntl.h>
85 #include <sys/kernel.h>
86 #include <sys/mount.h>
87 #include <sys/namei.h>
88 #include <sys/malloc.h>
89 #include <sys/pctrie.h>
92 #include <sys/racct.h>
93 #include <sys/resource.h>
94 #include <sys/resourcevar.h>
95 #include <sys/rwlock.h>
97 #include <sys/sysctl.h>
98 #include <sys/sysproto.h>
99 #include <sys/systm.h>
101 #include <sys/vmmeter.h>
102 #include <sys/vnode.h>
104 #include <security/mac/mac_framework.h>
108 #include <vm/vm_map.h>
109 #include <vm/vm_kern.h>
110 #include <vm/vm_object.h>
111 #include <vm/vm_page.h>
112 #include <vm/vm_pager.h>
113 #include <vm/vm_pageout.h>
114 #include <vm/vm_param.h>
115 #include <vm/swap_pager.h>
116 #include <vm/vm_extern.h>
119 #include <geom/geom.h>
122 * MAX_PAGEOUT_CLUSTER must be a power of 2 between 1 and 64.
123 * The 64-page limit is due to the radix code (kern/subr_blist.c).
125 #ifndef MAX_PAGEOUT_CLUSTER
126 #define MAX_PAGEOUT_CLUSTER 32
129 #if !defined(SWB_NPAGES)
130 #define SWB_NPAGES MAX_PAGEOUT_CLUSTER
133 #define SWAP_META_PAGES PCTRIE_COUNT
136 * A swblk structure maps each page index within a
137 * SWAP_META_PAGES-aligned and sized range to the address of an
138 * on-disk swap block (or SWAPBLK_NONE). The collection of these
139 * mappings for an entire vm object is implemented as a pc-trie.
143 daddr_t d[SWAP_META_PAGES];
146 static MALLOC_DEFINE(M_VMPGDATA, "vm_pgdata", "swap pager private data");
147 static struct mtx sw_dev_mtx;
148 static TAILQ_HEAD(, swdevt) swtailq = TAILQ_HEAD_INITIALIZER(swtailq);
149 static struct swdevt *swdevhd; /* Allocate from here next */
150 static int nswapdev; /* Number of swap devices */
151 int swap_pager_avail;
152 static struct sx swdev_syscall_lock; /* serialize swap(on|off) */
154 static u_long swap_reserved;
155 static u_long swap_total;
156 static int sysctl_page_shift(SYSCTL_HANDLER_ARGS);
157 SYSCTL_PROC(_vm, OID_AUTO, swap_reserved, CTLTYPE_U64 | CTLFLAG_RD | CTLFLAG_MPSAFE,
158 &swap_reserved, 0, sysctl_page_shift, "A",
159 "Amount of swap storage needed to back all allocated anonymous memory.");
160 SYSCTL_PROC(_vm, OID_AUTO, swap_total, CTLTYPE_U64 | CTLFLAG_RD | CTLFLAG_MPSAFE,
161 &swap_total, 0, sysctl_page_shift, "A",
162 "Total amount of available swap storage.");
164 static int overcommit = 0;
165 SYSCTL_INT(_vm, VM_OVERCOMMIT, overcommit, CTLFLAG_RW, &overcommit, 0,
166 "Configure virtual memory overcommit behavior. See tuning(7) "
168 static unsigned long swzone;
169 SYSCTL_ULONG(_vm, OID_AUTO, swzone, CTLFLAG_RD, &swzone, 0,
170 "Actual size of swap metadata zone");
171 static unsigned long swap_maxpages;
172 SYSCTL_ULONG(_vm, OID_AUTO, swap_maxpages, CTLFLAG_RD, &swap_maxpages, 0,
173 "Maximum amount of swap supported");
175 /* bits from overcommit */
176 #define SWAP_RESERVE_FORCE_ON (1 << 0)
177 #define SWAP_RESERVE_RLIMIT_ON (1 << 1)
178 #define SWAP_RESERVE_ALLOW_NONWIRED (1 << 2)
181 sysctl_page_shift(SYSCTL_HANDLER_ARGS)
184 u_long value = *(u_long *)arg1;
186 newval = ((uint64_t)value) << PAGE_SHIFT;
187 return (sysctl_handle_64(oidp, &newval, 0, req));
191 swap_reserve(vm_ooffset_t incr)
194 return (swap_reserve_by_cred(incr, curthread->td_ucred));
198 swap_reserve_by_cred(vm_ooffset_t incr, struct ucred *cred)
200 u_long r, s, prev, pincr;
203 static struct timeval lastfail;
206 uip = cred->cr_ruidinfo;
208 KASSERT((incr & PAGE_MASK) == 0, ("%s: incr: %ju & PAGE_MASK", __func__,
214 error = racct_add(curproc, RACCT_SWAP, incr);
215 PROC_UNLOCK(curproc);
223 prev = atomic_fetchadd_long(&swap_reserved, pincr);
225 if (overcommit & SWAP_RESERVE_ALLOW_NONWIRED) {
226 s = vm_cnt.v_page_count - vm_cnt.v_free_reserved -
231 if ((overcommit & SWAP_RESERVE_FORCE_ON) == 0 || r <= s ||
232 (error = priv_check(curthread, PRIV_VM_SWAP_NOQUOTA)) == 0) {
235 prev = atomic_fetchadd_long(&swap_reserved, -pincr);
237 panic("swap_reserved < incr on overcommit fail");
240 prev = atomic_fetchadd_long(&uip->ui_vmsize, pincr);
241 if ((overcommit & SWAP_RESERVE_RLIMIT_ON) != 0 &&
242 prev + pincr > lim_cur(curthread, RLIMIT_SWAP) &&
243 priv_check(curthread, PRIV_VM_SWAP_NORLIMIT)) {
245 prev = atomic_fetchadd_long(&uip->ui_vmsize, -pincr);
247 panic("uip->ui_vmsize < incr on overcommit fail");
250 if (!res && ppsratecheck(&lastfail, &curfail, 1)) {
251 printf("uid %d, pid %d: swap reservation for %jd bytes failed\n",
252 uip->ui_uid, curproc->p_pid, incr);
256 if (racct_enable && !res) {
258 racct_sub(curproc, RACCT_SWAP, incr);
259 PROC_UNLOCK(curproc);
267 swap_reserve_force(vm_ooffset_t incr)
272 KASSERT((incr & PAGE_MASK) == 0, ("%s: incr: %ju & PAGE_MASK", __func__,
278 racct_add_force(curproc, RACCT_SWAP, incr);
281 atomic_add_long(&swap_reserved, pincr);
282 uip = curproc->p_ucred->cr_ruidinfo;
283 atomic_add_long(&uip->ui_vmsize, pincr);
284 PROC_UNLOCK(curproc);
288 swap_release(vm_ooffset_t decr)
293 cred = curproc->p_ucred;
294 swap_release_by_cred(decr, cred);
295 PROC_UNLOCK(curproc);
299 swap_release_by_cred(vm_ooffset_t decr, struct ucred *cred)
304 uip = cred->cr_ruidinfo;
306 KASSERT((decr & PAGE_MASK) == 0, ("%s: decr: %ju & PAGE_MASK", __func__,
310 prev = atomic_fetchadd_long(&swap_reserved, -pdecr);
312 panic("swap_reserved < decr");
314 prev = atomic_fetchadd_long(&uip->ui_vmsize, -pdecr);
316 printf("negative vmsize for uid = %d\n", uip->ui_uid);
319 racct_sub_cred(cred, RACCT_SWAP, decr);
323 #define SWM_POP 0x01 /* pop out */
325 static int swap_pager_full = 2; /* swap space exhaustion (task killing) */
326 static int swap_pager_almost_full = 1; /* swap space exhaustion (w/hysteresis)*/
327 static struct mtx swbuf_mtx; /* to sync nsw_wcount_async */
328 static int nsw_wcount_async; /* limit async write buffers */
329 static int nsw_wcount_async_max;/* assigned maximum */
330 static int nsw_cluster_max; /* maximum VOP I/O allowed */
332 static int sysctl_swap_async_max(SYSCTL_HANDLER_ARGS);
333 SYSCTL_PROC(_vm, OID_AUTO, swap_async_max, CTLTYPE_INT | CTLFLAG_RW |
334 CTLFLAG_MPSAFE, NULL, 0, sysctl_swap_async_max, "I",
335 "Maximum running async swap ops");
336 static int sysctl_swap_fragmentation(SYSCTL_HANDLER_ARGS);
337 SYSCTL_PROC(_vm, OID_AUTO, swap_fragmentation, CTLTYPE_STRING | CTLFLAG_RD |
338 CTLFLAG_MPSAFE, NULL, 0, sysctl_swap_fragmentation, "A",
339 "Swap Fragmentation Info");
341 static struct sx sw_alloc_sx;
344 * "named" and "unnamed" anon region objects. Try to reduce the overhead
345 * of searching a named list by hashing it just a little.
350 #define NOBJLIST(handle) \
351 (&swap_pager_object_list[((int)(intptr_t)handle >> 4) & (NOBJLISTS-1)])
353 static struct pagerlst swap_pager_object_list[NOBJLISTS];
354 static uma_zone_t swwbuf_zone;
355 static uma_zone_t swrbuf_zone;
356 static uma_zone_t swblk_zone;
357 static uma_zone_t swpctrie_zone;
360 * pagerops for OBJT_SWAP - "swap pager". Some ops are also global procedure
361 * calls hooked from other parts of the VM system and do not appear here.
362 * (see vm/swap_pager.h).
365 swap_pager_alloc(void *handle, vm_ooffset_t size,
366 vm_prot_t prot, vm_ooffset_t offset, struct ucred *);
367 static void swap_pager_dealloc(vm_object_t object);
368 static int swap_pager_getpages(vm_object_t, vm_page_t *, int, int *,
370 static int swap_pager_getpages_async(vm_object_t, vm_page_t *, int, int *,
371 int *, pgo_getpages_iodone_t, void *);
372 static void swap_pager_putpages(vm_object_t, vm_page_t *, int, boolean_t, int *);
374 swap_pager_haspage(vm_object_t object, vm_pindex_t pindex, int *before, int *after);
375 static void swap_pager_init(void);
376 static void swap_pager_unswapped(vm_page_t);
377 static void swap_pager_swapoff(struct swdevt *sp);
379 struct pagerops swappagerops = {
380 .pgo_init = swap_pager_init, /* early system initialization of pager */
381 .pgo_alloc = swap_pager_alloc, /* allocate an OBJT_SWAP object */
382 .pgo_dealloc = swap_pager_dealloc, /* deallocate an OBJT_SWAP object */
383 .pgo_getpages = swap_pager_getpages, /* pagein */
384 .pgo_getpages_async = swap_pager_getpages_async, /* pagein (async) */
385 .pgo_putpages = swap_pager_putpages, /* pageout */
386 .pgo_haspage = swap_pager_haspage, /* get backing store status for page */
387 .pgo_pageunswapped = swap_pager_unswapped, /* remove swap related to page */
391 * swap_*() routines are externally accessible. swp_*() routines are
394 static int nswap_lowat = 128; /* in pages, swap_pager_almost_full warn */
395 static int nswap_hiwat = 512; /* in pages, swap_pager_almost_full warn */
397 SYSCTL_INT(_vm, OID_AUTO, dmmax, CTLFLAG_RD, &nsw_cluster_max, 0,
398 "Maximum size of a swap block in pages");
400 static void swp_sizecheck(void);
401 static void swp_pager_async_iodone(struct buf *bp);
402 static bool swp_pager_swblk_empty(struct swblk *sb, int start, int limit);
403 static int swapongeom(struct vnode *);
404 static int swaponvp(struct thread *, struct vnode *, u_long);
405 static int swapoff_one(struct swdevt *sp, struct ucred *cred);
408 * Swap bitmap functions
410 static void swp_pager_freeswapspace(daddr_t blk, daddr_t npages);
411 static daddr_t swp_pager_getswapspace(int *npages, int limit);
416 static daddr_t swp_pager_meta_build(vm_object_t, vm_pindex_t, daddr_t);
417 static void swp_pager_meta_free(vm_object_t, vm_pindex_t, vm_pindex_t);
418 static void swp_pager_meta_free_all(vm_object_t);
419 static daddr_t swp_pager_meta_ctl(vm_object_t, vm_pindex_t, int);
422 swp_pager_init_freerange(daddr_t *start, daddr_t *num)
425 *start = SWAPBLK_NONE;
430 swp_pager_update_freerange(daddr_t *start, daddr_t *num, daddr_t addr)
433 if (*start + *num == addr) {
436 swp_pager_freeswapspace(*start, *num);
443 swblk_trie_alloc(struct pctrie *ptree)
446 return (uma_zalloc(swpctrie_zone, M_NOWAIT | (curproc == pageproc ?
447 M_USE_RESERVE : 0)));
451 swblk_trie_free(struct pctrie *ptree, void *node)
454 uma_zfree(swpctrie_zone, node);
457 PCTRIE_DEFINE(SWAP, swblk, p, swblk_trie_alloc, swblk_trie_free);
460 * SWP_SIZECHECK() - update swap_pager_full indication
462 * update the swap_pager_almost_full indication and warn when we are
463 * about to run out of swap space, using lowat/hiwat hysteresis.
465 * Clear swap_pager_full ( task killing ) indication when lowat is met.
467 * No restrictions on call
468 * This routine may not block.
474 if (swap_pager_avail < nswap_lowat) {
475 if (swap_pager_almost_full == 0) {
476 printf("swap_pager: out of swap space\n");
477 swap_pager_almost_full = 1;
481 if (swap_pager_avail > nswap_hiwat)
482 swap_pager_almost_full = 0;
487 * SWAP_PAGER_INIT() - initialize the swap pager!
489 * Expected to be started from system init. NOTE: This code is run
490 * before much else so be careful what you depend on. Most of the VM
491 * system has yet to be initialized at this point.
494 swap_pager_init(void)
497 * Initialize object lists
501 for (i = 0; i < NOBJLISTS; ++i)
502 TAILQ_INIT(&swap_pager_object_list[i]);
503 mtx_init(&sw_dev_mtx, "swapdev", NULL, MTX_DEF);
504 sx_init(&sw_alloc_sx, "swspsx");
505 sx_init(&swdev_syscall_lock, "swsysc");
509 * SWAP_PAGER_SWAP_INIT() - swap pager initialization from pageout process
511 * Expected to be started from pageout process once, prior to entering
515 swap_pager_swap_init(void)
520 * Number of in-transit swap bp operations. Don't
521 * exhaust the pbufs completely. Make sure we
522 * initialize workable values (0 will work for hysteresis
523 * but it isn't very efficient).
525 * The nsw_cluster_max is constrained by the bp->b_pages[]
526 * array, which has MAXPHYS / PAGE_SIZE entries, and our locally
527 * defined MAX_PAGEOUT_CLUSTER. Also be aware that swap ops are
528 * constrained by the swap device interleave stripe size.
530 * Currently we hardwire nsw_wcount_async to 4. This limit is
531 * designed to prevent other I/O from having high latencies due to
532 * our pageout I/O. The value 4 works well for one or two active swap
533 * devices but is probably a little low if you have more. Even so,
534 * a higher value would probably generate only a limited improvement
535 * with three or four active swap devices since the system does not
536 * typically have to pageout at extreme bandwidths. We will want
537 * at least 2 per swap devices, and 4 is a pretty good value if you
538 * have one NFS swap device due to the command/ack latency over NFS.
539 * So it all works out pretty well.
541 nsw_cluster_max = min(MAXPHYS / PAGE_SIZE, MAX_PAGEOUT_CLUSTER);
543 nsw_wcount_async = 4;
544 nsw_wcount_async_max = nsw_wcount_async;
545 mtx_init(&swbuf_mtx, "async swbuf mutex", NULL, MTX_DEF);
547 swwbuf_zone = pbuf_zsecond_create("swwbuf", nswbuf / 4);
548 swrbuf_zone = pbuf_zsecond_create("swrbuf", nswbuf / 2);
551 * Initialize our zone, taking the user's requested size or
552 * estimating the number we need based on the number of pages
555 n = maxswzone != 0 ? maxswzone / sizeof(struct swblk) :
556 vm_cnt.v_page_count / 2;
557 swpctrie_zone = uma_zcreate("swpctrie", pctrie_node_size(), NULL, NULL,
558 pctrie_zone_init, NULL, UMA_ALIGN_PTR, UMA_ZONE_VM);
559 if (swpctrie_zone == NULL)
560 panic("failed to create swap pctrie zone.");
561 swblk_zone = uma_zcreate("swblk", sizeof(struct swblk), NULL, NULL,
562 NULL, NULL, _Alignof(struct swblk) - 1, UMA_ZONE_VM);
563 if (swblk_zone == NULL)
564 panic("failed to create swap blk zone.");
567 if (uma_zone_reserve_kva(swblk_zone, n))
570 * if the allocation failed, try a zone two thirds the
571 * size of the previous attempt.
577 * Often uma_zone_reserve_kva() cannot reserve exactly the
578 * requested size. Account for the difference when
579 * calculating swap_maxpages.
581 n = uma_zone_get_max(swblk_zone);
584 printf("Swap blk zone entries changed from %lu to %lu.\n",
586 swap_maxpages = n * SWAP_META_PAGES;
587 swzone = n * sizeof(struct swblk);
588 if (!uma_zone_reserve_kva(swpctrie_zone, n))
589 printf("Cannot reserve swap pctrie zone, "
590 "reduce kern.maxswzone.\n");
594 swap_pager_alloc_init(void *handle, struct ucred *cred, vm_ooffset_t size,
600 if (!swap_reserve_by_cred(size, cred))
606 * The un_pager.swp.swp_blks trie is initialized by
607 * vm_object_allocate() to ensure the correct order of
608 * visibility to other threads.
610 object = vm_object_allocate(OBJT_SWAP, OFF_TO_IDX(offset +
613 object->handle = handle;
616 object->charge = size;
622 * SWAP_PAGER_ALLOC() - allocate a new OBJT_SWAP VM object and instantiate
623 * its metadata structures.
625 * This routine is called from the mmap and fork code to create a new
628 * This routine must ensure that no live duplicate is created for
629 * the named object request, which is protected against by
630 * holding the sw_alloc_sx lock in case handle != NULL.
633 swap_pager_alloc(void *handle, vm_ooffset_t size, vm_prot_t prot,
634 vm_ooffset_t offset, struct ucred *cred)
638 if (handle != NULL) {
640 * Reference existing named region or allocate new one. There
641 * should not be a race here against swp_pager_meta_build()
642 * as called from vm_page_remove() in regards to the lookup
645 sx_xlock(&sw_alloc_sx);
646 object = vm_pager_object_lookup(NOBJLIST(handle), handle);
647 if (object == NULL) {
648 object = swap_pager_alloc_init(handle, cred, size,
650 if (object != NULL) {
651 TAILQ_INSERT_TAIL(NOBJLIST(object->handle),
652 object, pager_object_list);
655 sx_xunlock(&sw_alloc_sx);
657 object = swap_pager_alloc_init(handle, cred, size, offset);
663 * SWAP_PAGER_DEALLOC() - remove swap metadata from object
665 * The swap backing for the object is destroyed. The code is
666 * designed such that we can reinstantiate it later, but this
667 * routine is typically called only when the entire object is
668 * about to be destroyed.
670 * The object must be locked.
673 swap_pager_dealloc(vm_object_t object)
676 VM_OBJECT_ASSERT_WLOCKED(object);
677 KASSERT((object->flags & OBJ_DEAD) != 0, ("dealloc of reachable obj"));
680 * Remove from list right away so lookups will fail if we block for
681 * pageout completion.
683 if (object->handle != NULL) {
684 VM_OBJECT_WUNLOCK(object);
685 sx_xlock(&sw_alloc_sx);
686 TAILQ_REMOVE(NOBJLIST(object->handle), object,
688 sx_xunlock(&sw_alloc_sx);
689 VM_OBJECT_WLOCK(object);
692 vm_object_pip_wait(object, "swpdea");
695 * Free all remaining metadata. We only bother to free it from
696 * the swap meta data. We do not attempt to free swapblk's still
697 * associated with vm_page_t's for this object. We do not care
698 * if paging is still in progress on some objects.
700 swp_pager_meta_free_all(object);
701 object->handle = NULL;
702 object->type = OBJT_DEAD;
705 /************************************************************************
706 * SWAP PAGER BITMAP ROUTINES *
707 ************************************************************************/
710 * SWP_PAGER_GETSWAPSPACE() - allocate raw swap space
712 * Allocate swap for up to the requested number of pages, and at
713 * least a minimum number of pages. The starting swap block number
714 * (a page index) is returned or SWAPBLK_NONE if the allocation
717 * Also has the side effect of advising that somebody made a mistake
718 * when they configured swap and didn't configure enough.
720 * This routine may not sleep.
722 * We allocate in round-robin fashion from the configured devices.
725 swp_pager_getswapspace(int *io_npages, int limit)
733 npages = imin(BLIST_MAX_ALLOC, mpages);
734 mtx_lock(&sw_dev_mtx);
736 while (!TAILQ_EMPTY(&swtailq)) {
738 sp = TAILQ_FIRST(&swtailq);
739 if ((sp->sw_flags & SW_CLOSING) == 0)
740 blk = blist_alloc(sp->sw_blist, &npages, mpages);
741 if (blk != SWAPBLK_NONE)
743 sp = TAILQ_NEXT(sp, sw_list);
751 if (blk != SWAPBLK_NONE) {
754 sp->sw_used += npages;
755 swap_pager_avail -= npages;
757 swdevhd = TAILQ_NEXT(sp, sw_list);
759 if (swap_pager_full != 2) {
760 printf("swp_pager_getswapspace(%d): failed\n",
763 swap_pager_almost_full = 1;
767 mtx_unlock(&sw_dev_mtx);
772 swp_pager_isondev(daddr_t blk, struct swdevt *sp)
775 return (blk >= sp->sw_first && blk < sp->sw_end);
779 swp_pager_strategy(struct buf *bp)
783 mtx_lock(&sw_dev_mtx);
784 TAILQ_FOREACH(sp, &swtailq, sw_list) {
785 if (swp_pager_isondev(bp->b_blkno, sp)) {
786 mtx_unlock(&sw_dev_mtx);
787 if ((sp->sw_flags & SW_UNMAPPED) != 0 &&
788 unmapped_buf_allowed) {
789 bp->b_data = unmapped_buf;
792 pmap_qenter((vm_offset_t)bp->b_data,
793 &bp->b_pages[0], bp->b_bcount / PAGE_SIZE);
795 sp->sw_strategy(bp, sp);
799 panic("Swapdev not found");
804 * SWP_PAGER_FREESWAPSPACE() - free raw swap space
806 * This routine returns the specified swap blocks back to the bitmap.
808 * This routine may not sleep.
811 swp_pager_freeswapspace(daddr_t blk, daddr_t npages)
817 mtx_lock(&sw_dev_mtx);
818 TAILQ_FOREACH(sp, &swtailq, sw_list) {
819 if (swp_pager_isondev(blk, sp)) {
820 sp->sw_used -= npages;
822 * If we are attempting to stop swapping on
823 * this device, we don't want to mark any
824 * blocks free lest they be reused.
826 if ((sp->sw_flags & SW_CLOSING) == 0) {
827 blist_free(sp->sw_blist, blk - sp->sw_first,
829 swap_pager_avail += npages;
832 mtx_unlock(&sw_dev_mtx);
836 panic("Swapdev not found");
840 * SYSCTL_SWAP_FRAGMENTATION() - produce raw swap space stats
843 sysctl_swap_fragmentation(SYSCTL_HANDLER_ARGS)
850 error = sysctl_wire_old_buffer(req, 0);
853 sbuf_new_for_sysctl(&sbuf, NULL, 128, req);
854 mtx_lock(&sw_dev_mtx);
855 TAILQ_FOREACH(sp, &swtailq, sw_list) {
856 if (vn_isdisk(sp->sw_vp, NULL))
857 devname = devtoname(sp->sw_vp->v_rdev);
860 sbuf_printf(&sbuf, "\nFree space on device %s:\n", devname);
861 blist_stats(sp->sw_blist, &sbuf);
863 mtx_unlock(&sw_dev_mtx);
864 error = sbuf_finish(&sbuf);
870 * SWAP_PAGER_FREESPACE() - frees swap blocks associated with a page
871 * range within an object.
873 * This is a globally accessible routine.
875 * This routine removes swapblk assignments from swap metadata.
877 * The external callers of this routine typically have already destroyed
878 * or renamed vm_page_t's associated with this range in the object so
881 * The object must be locked.
884 swap_pager_freespace(vm_object_t object, vm_pindex_t start, vm_size_t size)
887 swp_pager_meta_free(object, start, size);
891 * SWAP_PAGER_RESERVE() - reserve swap blocks in object
893 * Assigns swap blocks to the specified range within the object. The
894 * swap blocks are not zeroed. Any previous swap assignment is destroyed.
896 * Returns 0 on success, -1 on failure.
899 swap_pager_reserve(vm_object_t object, vm_pindex_t start, vm_size_t size)
901 daddr_t addr, blk, n_free, s_free;
904 swp_pager_init_freerange(&s_free, &n_free);
905 VM_OBJECT_WLOCK(object);
906 for (i = 0; i < size; i += n) {
908 blk = swp_pager_getswapspace(&n, 1);
909 if (blk == SWAPBLK_NONE) {
910 swp_pager_meta_free(object, start, i);
911 VM_OBJECT_WUNLOCK(object);
914 for (j = 0; j < n; ++j) {
915 addr = swp_pager_meta_build(object,
916 start + i + j, blk + j);
917 if (addr != SWAPBLK_NONE)
918 swp_pager_update_freerange(&s_free, &n_free,
922 swp_pager_freeswapspace(s_free, n_free);
923 VM_OBJECT_WUNLOCK(object);
928 * SWAP_PAGER_COPY() - copy blocks from source pager to destination pager
929 * and destroy the source.
931 * Copy any valid swapblks from the source to the destination. In
932 * cases where both the source and destination have a valid swapblk,
933 * we keep the destination's.
935 * This routine is allowed to sleep. It may sleep allocating metadata
936 * indirectly through swp_pager_meta_build() or if paging is still in
937 * progress on the source.
939 * The source object contains no vm_page_t's (which is just as well)
941 * The source object is of type OBJT_SWAP.
943 * The source and destination objects must be locked.
944 * Both object locks may temporarily be released.
947 swap_pager_copy(vm_object_t srcobject, vm_object_t dstobject,
948 vm_pindex_t offset, int destroysource)
951 daddr_t dstaddr, n_free, s_free, srcaddr;
953 VM_OBJECT_ASSERT_WLOCKED(srcobject);
954 VM_OBJECT_ASSERT_WLOCKED(dstobject);
957 * If destroysource is set, we remove the source object from the
958 * swap_pager internal queue now.
960 if (destroysource && srcobject->handle != NULL) {
961 vm_object_pip_add(srcobject, 1);
962 VM_OBJECT_WUNLOCK(srcobject);
963 vm_object_pip_add(dstobject, 1);
964 VM_OBJECT_WUNLOCK(dstobject);
965 sx_xlock(&sw_alloc_sx);
966 TAILQ_REMOVE(NOBJLIST(srcobject->handle), srcobject,
968 sx_xunlock(&sw_alloc_sx);
969 VM_OBJECT_WLOCK(dstobject);
970 vm_object_pip_wakeup(dstobject);
971 VM_OBJECT_WLOCK(srcobject);
972 vm_object_pip_wakeup(srcobject);
976 * Transfer source to destination.
978 swp_pager_init_freerange(&s_free, &n_free);
979 for (i = 0; i < dstobject->size; ++i) {
980 srcaddr = swp_pager_meta_ctl(srcobject, i + offset, SWM_POP);
981 if (srcaddr == SWAPBLK_NONE)
983 dstaddr = swp_pager_meta_ctl(dstobject, i, 0);
984 if (dstaddr != SWAPBLK_NONE) {
986 * Destination has valid swapblk or it is represented
987 * by a resident page. We destroy the source block.
989 swp_pager_update_freerange(&s_free, &n_free, srcaddr);
994 * Destination has no swapblk and is not resident,
997 * swp_pager_meta_build() can sleep.
999 vm_object_pip_add(srcobject, 1);
1000 VM_OBJECT_WUNLOCK(srcobject);
1001 vm_object_pip_add(dstobject, 1);
1002 dstaddr = swp_pager_meta_build(dstobject, i, srcaddr);
1003 KASSERT(dstaddr == SWAPBLK_NONE,
1004 ("Unexpected destination swapblk"));
1005 vm_object_pip_wakeup(dstobject);
1006 VM_OBJECT_WLOCK(srcobject);
1007 vm_object_pip_wakeup(srcobject);
1009 swp_pager_freeswapspace(s_free, n_free);
1012 * Free left over swap blocks in source.
1014 * We have to revert the type to OBJT_DEFAULT so we do not accidentally
1015 * double-remove the object from the swap queues.
1017 if (destroysource) {
1018 swp_pager_meta_free_all(srcobject);
1020 * Reverting the type is not necessary, the caller is going
1021 * to destroy srcobject directly, but I'm doing it here
1022 * for consistency since we've removed the object from its
1025 srcobject->type = OBJT_DEFAULT;
1030 * SWAP_PAGER_HASPAGE() - determine if we have good backing store for
1031 * the requested page.
1033 * We determine whether good backing store exists for the requested
1034 * page and return TRUE if it does, FALSE if it doesn't.
1036 * If TRUE, we also try to determine how much valid, contiguous backing
1037 * store exists before and after the requested page.
1040 swap_pager_haspage(vm_object_t object, vm_pindex_t pindex, int *before,
1046 VM_OBJECT_ASSERT_LOCKED(object);
1049 * do we have good backing store at the requested index ?
1051 blk0 = swp_pager_meta_ctl(object, pindex, 0);
1052 if (blk0 == SWAPBLK_NONE) {
1061 * find backwards-looking contiguous good backing store
1063 if (before != NULL) {
1064 for (i = 1; i < SWB_NPAGES; i++) {
1067 blk = swp_pager_meta_ctl(object, pindex - i, 0);
1068 if (blk != blk0 - i)
1075 * find forward-looking contiguous good backing store
1077 if (after != NULL) {
1078 for (i = 1; i < SWB_NPAGES; i++) {
1079 blk = swp_pager_meta_ctl(object, pindex + i, 0);
1080 if (blk != blk0 + i)
1089 * SWAP_PAGER_PAGE_UNSWAPPED() - remove swap backing store related to page
1091 * This removes any associated swap backing store, whether valid or
1092 * not, from the page.
1094 * This routine is typically called when a page is made dirty, at
1095 * which point any associated swap can be freed. MADV_FREE also
1096 * calls us in a special-case situation
1098 * NOTE!!! If the page is clean and the swap was valid, the caller
1099 * should make the page dirty before calling this routine. This routine
1100 * does NOT change the m->dirty status of the page. Also: MADV_FREE
1103 * This routine may not sleep.
1105 * The object containing the page must be locked.
1108 swap_pager_unswapped(vm_page_t m)
1112 srcaddr = swp_pager_meta_ctl(m->object, m->pindex, SWM_POP);
1113 if (srcaddr != SWAPBLK_NONE)
1114 swp_pager_freeswapspace(srcaddr, 1);
1118 * swap_pager_getpages() - bring pages in from swap
1120 * Attempt to page in the pages in array "ma" of length "count". The
1121 * caller may optionally specify that additional pages preceding and
1122 * succeeding the specified range be paged in. The number of such pages
1123 * is returned in the "rbehind" and "rahead" parameters, and they will
1124 * be in the inactive queue upon return.
1126 * The pages in "ma" must be busied and will remain busied upon return.
1129 swap_pager_getpages(vm_object_t object, vm_page_t *ma, int count, int *rbehind,
1133 vm_page_t bm, mpred, msucc, p;
1136 int i, maxahead, maxbehind, reqcount;
1141 * Determine the final number of read-behind pages and
1142 * allocate them BEFORE releasing the object lock. Otherwise,
1143 * there can be a problematic race with vm_object_split().
1144 * Specifically, vm_object_split() might first transfer pages
1145 * that precede ma[0] in the current object to a new object,
1146 * and then this function incorrectly recreates those pages as
1147 * read-behind pages in the current object.
1149 if (!swap_pager_haspage(object, ma[0]->pindex, &maxbehind, &maxahead))
1150 return (VM_PAGER_FAIL);
1153 * Clip the readahead and readbehind ranges to exclude resident pages.
1155 if (rahead != NULL) {
1156 KASSERT(reqcount - 1 <= maxahead,
1157 ("page count %d extends beyond swap block", reqcount));
1158 *rahead = imin(*rahead, maxahead - (reqcount - 1));
1159 pindex = ma[reqcount - 1]->pindex;
1160 msucc = TAILQ_NEXT(ma[reqcount - 1], listq);
1161 if (msucc != NULL && msucc->pindex - pindex - 1 < *rahead)
1162 *rahead = msucc->pindex - pindex - 1;
1164 if (rbehind != NULL) {
1165 *rbehind = imin(*rbehind, maxbehind);
1166 pindex = ma[0]->pindex;
1167 mpred = TAILQ_PREV(ma[0], pglist, listq);
1168 if (mpred != NULL && pindex - mpred->pindex - 1 < *rbehind)
1169 *rbehind = pindex - mpred->pindex - 1;
1173 for (i = 0; i < count; i++)
1174 ma[i]->oflags |= VPO_SWAPINPROG;
1177 * Allocate readahead and readbehind pages.
1179 if (rbehind != NULL) {
1180 for (i = 1; i <= *rbehind; i++) {
1181 p = vm_page_alloc(object, ma[0]->pindex - i,
1185 p->oflags |= VPO_SWAPINPROG;
1190 if (rahead != NULL) {
1191 for (i = 0; i < *rahead; i++) {
1192 p = vm_page_alloc(object,
1193 ma[reqcount - 1]->pindex + i + 1, VM_ALLOC_NORMAL);
1196 p->oflags |= VPO_SWAPINPROG;
1200 if (rbehind != NULL)
1205 vm_object_pip_add(object, count);
1207 pindex = bm->pindex;
1208 blk = swp_pager_meta_ctl(object, pindex, 0);
1209 KASSERT(blk != SWAPBLK_NONE,
1210 ("no swap blocking containing %p(%jx)", object, (uintmax_t)pindex));
1212 VM_OBJECT_WUNLOCK(object);
1213 bp = uma_zalloc(swrbuf_zone, M_WAITOK);
1214 /* Pages cannot leave the object while busy. */
1215 for (i = 0, p = bm; i < count; i++, p = TAILQ_NEXT(p, listq)) {
1216 MPASS(p->pindex == bm->pindex + i);
1220 bp->b_flags |= B_PAGING;
1221 bp->b_iocmd = BIO_READ;
1222 bp->b_iodone = swp_pager_async_iodone;
1223 bp->b_rcred = crhold(thread0.td_ucred);
1224 bp->b_wcred = crhold(thread0.td_ucred);
1226 bp->b_bcount = PAGE_SIZE * count;
1227 bp->b_bufsize = PAGE_SIZE * count;
1228 bp->b_npages = count;
1229 bp->b_pgbefore = rbehind != NULL ? *rbehind : 0;
1230 bp->b_pgafter = rahead != NULL ? *rahead : 0;
1232 VM_CNT_INC(v_swapin);
1233 VM_CNT_ADD(v_swappgsin, count);
1236 * perform the I/O. NOTE!!! bp cannot be considered valid after
1237 * this point because we automatically release it on completion.
1238 * Instead, we look at the one page we are interested in which we
1239 * still hold a lock on even through the I/O completion.
1241 * The other pages in our ma[] array are also released on completion,
1242 * so we cannot assume they are valid anymore either.
1244 * NOTE: b_blkno is destroyed by the call to swapdev_strategy
1247 swp_pager_strategy(bp);
1250 * Wait for the pages we want to complete. VPO_SWAPINPROG is always
1251 * cleared on completion. If an I/O error occurs, SWAPBLK_NONE
1252 * is set in the metadata for each page in the request.
1254 VM_OBJECT_WLOCK(object);
1255 while ((ma[0]->oflags & VPO_SWAPINPROG) != 0) {
1256 ma[0]->oflags |= VPO_SWAPSLEEP;
1257 VM_CNT_INC(v_intrans);
1258 if (VM_OBJECT_SLEEP(object, &object->paging_in_progress, PSWP,
1259 "swread", hz * 20)) {
1261 "swap_pager: indefinite wait buffer: bufobj: %p, blkno: %jd, size: %ld\n",
1262 bp->b_bufobj, (intmax_t)bp->b_blkno, bp->b_bcount);
1267 * If we had an unrecoverable read error pages will not be valid.
1269 for (i = 0; i < reqcount; i++)
1270 if (ma[i]->valid != VM_PAGE_BITS_ALL)
1271 return (VM_PAGER_ERROR);
1273 return (VM_PAGER_OK);
1276 * A final note: in a low swap situation, we cannot deallocate swap
1277 * and mark a page dirty here because the caller is likely to mark
1278 * the page clean when we return, causing the page to possibly revert
1279 * to all-zero's later.
1284 * swap_pager_getpages_async():
1286 * Right now this is emulation of asynchronous operation on top of
1287 * swap_pager_getpages().
1290 swap_pager_getpages_async(vm_object_t object, vm_page_t *ma, int count,
1291 int *rbehind, int *rahead, pgo_getpages_iodone_t iodone, void *arg)
1295 r = swap_pager_getpages(object, ma, count, rbehind, rahead);
1296 VM_OBJECT_WUNLOCK(object);
1301 case VM_PAGER_ERROR:
1308 panic("unhandled swap_pager_getpages() error %d", r);
1310 (iodone)(arg, ma, count, error);
1311 VM_OBJECT_WLOCK(object);
1317 * swap_pager_putpages:
1319 * Assign swap (if necessary) and initiate I/O on the specified pages.
1321 * We support both OBJT_DEFAULT and OBJT_SWAP objects. DEFAULT objects
1322 * are automatically converted to SWAP objects.
1324 * In a low memory situation we may block in VOP_STRATEGY(), but the new
1325 * vm_page reservation system coupled with properly written VFS devices
1326 * should ensure that no low-memory deadlock occurs. This is an area
1329 * The parent has N vm_object_pip_add() references prior to
1330 * calling us and will remove references for rtvals[] that are
1331 * not set to VM_PAGER_PEND. We need to remove the rest on I/O
1334 * The parent has soft-busy'd the pages it passes us and will unbusy
1335 * those whose rtvals[] entry is not set to VM_PAGER_PEND on return.
1336 * We need to unbusy the rest on I/O completion.
1339 swap_pager_putpages(vm_object_t object, vm_page_t *ma, int count,
1340 int flags, int *rtvals)
1343 daddr_t addr, blk, n_free, s_free;
1348 KASSERT(count == 0 || ma[0]->object == object,
1349 ("%s: object mismatch %p/%p",
1350 __func__, object, ma[0]->object));
1355 * Turn object into OBJT_SWAP. Force sync if not a pageout process.
1357 if (object->type != OBJT_SWAP) {
1358 addr = swp_pager_meta_build(object, 0, SWAPBLK_NONE);
1359 KASSERT(addr == SWAPBLK_NONE,
1360 ("unexpected object swap block"));
1362 VM_OBJECT_WUNLOCK(object);
1363 async = curproc == pageproc && (flags & VM_PAGER_PUT_SYNC) == 0;
1364 swp_pager_init_freerange(&s_free, &n_free);
1369 * Assign swap blocks and issue I/O. We reallocate swap on the fly.
1370 * The page is left dirty until the pageout operation completes
1373 for (i = 0; i < count; i += n) {
1374 /* Maximum I/O size is limited by maximum swap block size. */
1375 n = min(count - i, nsw_cluster_max);
1377 /* Get a block of swap of size up to size n. */
1378 blk = swp_pager_getswapspace(&n, 4);
1379 if (blk == SWAPBLK_NONE) {
1380 for (j = 0; j < n; ++j)
1381 rtvals[i + j] = VM_PAGER_FAIL;
1386 * All I/O parameters have been satisfied. Build the I/O
1387 * request and assign the swap space.
1390 mtx_lock(&swbuf_mtx);
1391 while (nsw_wcount_async == 0)
1392 msleep(&nsw_wcount_async, &swbuf_mtx, PVM,
1395 mtx_unlock(&swbuf_mtx);
1397 bp = uma_zalloc(swwbuf_zone, M_WAITOK);
1399 bp->b_flags = B_ASYNC;
1400 bp->b_flags |= B_PAGING;
1401 bp->b_iocmd = BIO_WRITE;
1403 bp->b_rcred = crhold(thread0.td_ucred);
1404 bp->b_wcred = crhold(thread0.td_ucred);
1405 bp->b_bcount = PAGE_SIZE * n;
1406 bp->b_bufsize = PAGE_SIZE * n;
1409 VM_OBJECT_WLOCK(object);
1410 for (j = 0; j < n; ++j) {
1412 addr = swp_pager_meta_build(mreq->object, mreq->pindex,
1414 if (addr != SWAPBLK_NONE)
1415 swp_pager_update_freerange(&s_free, &n_free,
1417 MPASS(mreq->dirty == VM_PAGE_BITS_ALL);
1418 mreq->oflags |= VPO_SWAPINPROG;
1419 bp->b_pages[j] = mreq;
1421 VM_OBJECT_WUNLOCK(object);
1424 * Must set dirty range for NFS to work.
1427 bp->b_dirtyend = bp->b_bcount;
1429 VM_CNT_INC(v_swapout);
1430 VM_CNT_ADD(v_swappgsout, bp->b_npages);
1433 * We unconditionally set rtvals[] to VM_PAGER_PEND so that we
1434 * can call the async completion routine at the end of a
1435 * synchronous I/O operation. Otherwise, our caller would
1436 * perform duplicate unbusy and wakeup operations on the page
1437 * and object, respectively.
1439 for (j = 0; j < n; j++)
1440 rtvals[i + j] = VM_PAGER_PEND;
1445 * NOTE: b_blkno is destroyed by the call to swapdev_strategy.
1448 bp->b_iodone = swp_pager_async_iodone;
1450 swp_pager_strategy(bp);
1457 * NOTE: b_blkno is destroyed by the call to swapdev_strategy.
1459 bp->b_iodone = bdone;
1460 swp_pager_strategy(bp);
1463 * Wait for the sync I/O to complete.
1465 bwait(bp, PVM, "swwrt");
1468 * Now that we are through with the bp, we can call the
1469 * normal async completion, which frees everything up.
1471 swp_pager_async_iodone(bp);
1473 swp_pager_freeswapspace(s_free, n_free);
1474 VM_OBJECT_WLOCK(object);
1478 * swp_pager_async_iodone:
1480 * Completion routine for asynchronous reads and writes from/to swap.
1481 * Also called manually by synchronous code to finish up a bp.
1483 * This routine may not sleep.
1486 swp_pager_async_iodone(struct buf *bp)
1489 vm_object_t object = NULL;
1492 * Report error - unless we ran out of memory, in which case
1493 * we've already logged it in swapgeom_strategy().
1495 if (bp->b_ioflags & BIO_ERROR && bp->b_error != ENOMEM) {
1497 "swap_pager: I/O error - %s failed; blkno %ld,"
1498 "size %ld, error %d\n",
1499 ((bp->b_iocmd == BIO_READ) ? "pagein" : "pageout"),
1507 * remove the mapping for kernel virtual
1510 pmap_qremove((vm_offset_t)bp->b_data, bp->b_npages);
1512 bp->b_data = bp->b_kvabase;
1515 object = bp->b_pages[0]->object;
1516 VM_OBJECT_WLOCK(object);
1520 * cleanup pages. If an error occurs writing to swap, we are in
1521 * very serious trouble. If it happens to be a disk error, though,
1522 * we may be able to recover by reassigning the swap later on. So
1523 * in this case we remove the m->swapblk assignment for the page
1524 * but do not free it in the rlist. The errornous block(s) are thus
1525 * never reallocated as swap. Redirty the page and continue.
1527 for (i = 0; i < bp->b_npages; ++i) {
1528 vm_page_t m = bp->b_pages[i];
1530 m->oflags &= ~VPO_SWAPINPROG;
1531 if (m->oflags & VPO_SWAPSLEEP) {
1532 m->oflags &= ~VPO_SWAPSLEEP;
1533 wakeup(&object->paging_in_progress);
1536 if (bp->b_ioflags & BIO_ERROR) {
1538 * If an error occurs I'd love to throw the swapblk
1539 * away without freeing it back to swapspace, so it
1540 * can never be used again. But I can't from an
1543 if (bp->b_iocmd == BIO_READ) {
1545 * NOTE: for reads, m->dirty will probably
1546 * be overridden by the original caller of
1547 * getpages so don't play cute tricks here.
1552 * If a write error occurs, reactivate page
1553 * so it doesn't clog the inactive list,
1554 * then finish the I/O.
1556 MPASS(m->dirty == VM_PAGE_BITS_ALL);
1558 vm_page_activate(m);
1562 } else if (bp->b_iocmd == BIO_READ) {
1564 * NOTE: for reads, m->dirty will probably be
1565 * overridden by the original caller of getpages so
1566 * we cannot set them in order to free the underlying
1567 * swap in a low-swap situation. I don't think we'd
1568 * want to do that anyway, but it was an optimization
1569 * that existed in the old swapper for a time before
1570 * it got ripped out due to precisely this problem.
1572 KASSERT(!pmap_page_is_mapped(m),
1573 ("swp_pager_async_iodone: page %p is mapped", m));
1574 KASSERT(m->dirty == 0,
1575 ("swp_pager_async_iodone: page %p is dirty", m));
1577 m->valid = VM_PAGE_BITS_ALL;
1578 if (i < bp->b_pgbefore ||
1579 i >= bp->b_npages - bp->b_pgafter)
1580 vm_page_readahead_finish(m);
1583 * For write success, clear the dirty
1584 * status, then finish the I/O ( which decrements the
1585 * busy count and possibly wakes waiter's up ).
1586 * A page is only written to swap after a period of
1587 * inactivity. Therefore, we do not expect it to be
1590 KASSERT(!pmap_page_is_write_mapped(m),
1591 ("swp_pager_async_iodone: page %p is not write"
1595 vm_page_deactivate_noreuse(m);
1602 * adjust pip. NOTE: the original parent may still have its own
1603 * pip refs on the object.
1605 if (object != NULL) {
1606 vm_object_pip_wakeupn(object, bp->b_npages);
1607 VM_OBJECT_WUNLOCK(object);
1611 * swapdev_strategy() manually sets b_vp and b_bufobj before calling
1612 * bstrategy(). Set them back to NULL now we're done with it, or we'll
1613 * trigger a KASSERT in relpbuf().
1617 bp->b_bufobj = NULL;
1620 * release the physical I/O buffer
1622 if (bp->b_flags & B_ASYNC) {
1623 mtx_lock(&swbuf_mtx);
1624 if (++nsw_wcount_async == 1)
1625 wakeup(&nsw_wcount_async);
1626 mtx_unlock(&swbuf_mtx);
1628 uma_zfree((bp->b_iocmd == BIO_READ) ? swrbuf_zone : swwbuf_zone, bp);
1632 swap_pager_nswapdev(void)
1639 swp_pager_force_dirty(vm_page_t m)
1645 if (!vm_page_wired(m) && m->queue == PQ_NONE)
1646 panic("page %p is neither wired nor queued", m);
1650 swap_pager_unswapped(m);
1654 swp_pager_force_launder(vm_page_t m)
1662 swap_pager_unswapped(m);
1666 * SWP_PAGER_FORCE_PAGEIN() - force swap blocks to be paged in
1668 * This routine dissociates pages starting at the given index within an
1669 * object from their backing store, paging them in if they do not reside
1670 * in memory. Pages that are paged in are marked dirty and placed in the
1671 * laundry queue. Pages are marked dirty because they no longer have
1672 * backing store. They are placed in the laundry queue because they have
1673 * not been accessed recently. Otherwise, they would already reside in
1677 swp_pager_force_pagein(vm_object_t object, vm_pindex_t pindex, int npages)
1679 vm_page_t ma[npages];
1682 KASSERT(npages > 0, ("%s: No pages", __func__));
1683 KASSERT(npages <= MAXPHYS / PAGE_SIZE,
1684 ("%s: Too many pages: %d", __func__, npages));
1685 vm_object_pip_add(object, npages);
1686 vm_page_grab_pages(object, pindex, VM_ALLOC_NORMAL, ma, npages);
1687 for (i = j = 0;; i++) {
1688 /* Count nonresident pages, to page-in all at once. */
1689 if (i < npages && ma[i]->valid != VM_PAGE_BITS_ALL)
1692 /* Page-in nonresident pages. Mark for laundering. */
1693 if (swap_pager_getpages(object, &ma[j], i - j, NULL,
1694 NULL) != VM_PAGER_OK)
1695 panic("%s: read from swap failed", __func__);
1697 swp_pager_force_launder(ma[j]);
1702 /* Mark dirty a resident page. */
1703 swp_pager_force_dirty(ma[j++]);
1705 vm_object_pip_wakeupn(object, npages);
1709 * swap_pager_swapoff_object:
1711 * Page in all of the pages that have been paged out for an object
1715 swap_pager_swapoff_object(struct swdevt *sp, vm_object_t object)
1718 vm_pindex_t pi, s_pindex;
1719 daddr_t blk, n_blks, s_blk;
1723 for (pi = 0; (sb = SWAP_PCTRIE_LOOKUP_GE(
1724 &object->un_pager.swp.swp_blks, pi)) != NULL; ) {
1725 for (i = 0; i < SWAP_META_PAGES; i++) {
1727 if (!swp_pager_isondev(blk, sp))
1731 * If there are no blocks/pages accumulated, start a new
1732 * accumulation here.
1735 if (blk != SWAPBLK_NONE) {
1737 s_pindex = sb->p + i;
1744 * If the accumulation can be extended without breaking
1745 * the sequence of consecutive blocks and pages that
1746 * swp_pager_force_pagein() depends on, do so.
1748 if (n_blks < MAXPHYS / PAGE_SIZE &&
1749 s_blk + n_blks == blk &&
1750 s_pindex + n_blks == sb->p + i) {
1756 * The sequence of consecutive blocks and pages cannot
1757 * be extended, so page them all in here. Then,
1758 * because doing so involves releasing and reacquiring
1759 * a lock that protects the swap block pctrie, do not
1760 * rely on the current swap block. Break this loop and
1761 * re-fetch the same pindex from the pctrie again.
1763 swp_pager_force_pagein(object, s_pindex, n_blks);
1767 if (i == SWAP_META_PAGES)
1768 pi = sb->p + SWAP_META_PAGES;
1771 swp_pager_force_pagein(object, s_pindex, n_blks);
1775 * swap_pager_swapoff:
1777 * Page in all of the pages that have been paged out to the
1778 * given device. The corresponding blocks in the bitmap must be
1779 * marked as allocated and the device must be flagged SW_CLOSING.
1780 * There may be no processes swapped out to the device.
1782 * This routine may block.
1785 swap_pager_swapoff(struct swdevt *sp)
1790 sx_assert(&swdev_syscall_lock, SA_XLOCKED);
1794 mtx_lock(&vm_object_list_mtx);
1795 TAILQ_FOREACH(object, &vm_object_list, object_list) {
1796 if (object->type != OBJT_SWAP)
1798 mtx_unlock(&vm_object_list_mtx);
1799 /* Depends on type-stability. */
1800 VM_OBJECT_WLOCK(object);
1803 * Dead objects are eventually terminated on their own.
1805 if ((object->flags & OBJ_DEAD) != 0)
1809 * Sync with fences placed after pctrie
1810 * initialization. We must not access pctrie below
1811 * unless we checked that our object is swap and not
1814 atomic_thread_fence_acq();
1815 if (object->type != OBJT_SWAP)
1818 swap_pager_swapoff_object(sp, object);
1820 VM_OBJECT_WUNLOCK(object);
1821 mtx_lock(&vm_object_list_mtx);
1823 mtx_unlock(&vm_object_list_mtx);
1827 * Objects may be locked or paging to the device being
1828 * removed, so we will miss their pages and need to
1829 * make another pass. We have marked this device as
1830 * SW_CLOSING, so the activity should finish soon.
1833 if (retries > 100) {
1834 panic("swapoff: failed to locate %d swap blocks",
1837 pause("swpoff", hz / 20);
1840 EVENTHANDLER_INVOKE(swapoff, sp);
1843 /************************************************************************
1845 ************************************************************************
1847 * These routines manipulate the swap metadata stored in the
1850 * Swap metadata is implemented with a global hash and not directly
1851 * linked into the object. Instead the object simply contains
1852 * appropriate tracking counters.
1856 * SWP_PAGER_SWBLK_EMPTY() - is a range of blocks free?
1859 swp_pager_swblk_empty(struct swblk *sb, int start, int limit)
1863 MPASS(0 <= start && start <= limit && limit <= SWAP_META_PAGES);
1864 for (i = start; i < limit; i++) {
1865 if (sb->d[i] != SWAPBLK_NONE)
1872 * SWP_PAGER_META_BUILD() - add swap block to swap meta data for object
1874 * We first convert the object to a swap object if it is a default
1877 * The specified swapblk is added to the object's swap metadata. If
1878 * the swapblk is not valid, it is freed instead. Any previously
1879 * assigned swapblk is returned.
1882 swp_pager_meta_build(vm_object_t object, vm_pindex_t pindex, daddr_t swapblk)
1884 static volatile int swblk_zone_exhausted, swpctrie_zone_exhausted;
1885 struct swblk *sb, *sb1;
1886 vm_pindex_t modpi, rdpi;
1887 daddr_t prev_swapblk;
1890 VM_OBJECT_ASSERT_WLOCKED(object);
1893 * Convert default object to swap object if necessary
1895 if (object->type != OBJT_SWAP) {
1896 pctrie_init(&object->un_pager.swp.swp_blks);
1899 * Ensure that swap_pager_swapoff()'s iteration over
1900 * object_list does not see a garbage pctrie.
1902 atomic_thread_fence_rel();
1904 object->type = OBJT_SWAP;
1905 KASSERT(object->handle == NULL, ("default pager with handle"));
1908 rdpi = rounddown(pindex, SWAP_META_PAGES);
1909 sb = SWAP_PCTRIE_LOOKUP(&object->un_pager.swp.swp_blks, rdpi);
1911 if (swapblk == SWAPBLK_NONE)
1912 return (SWAPBLK_NONE);
1914 sb = uma_zalloc(swblk_zone, M_NOWAIT | (curproc ==
1915 pageproc ? M_USE_RESERVE : 0));
1918 for (i = 0; i < SWAP_META_PAGES; i++)
1919 sb->d[i] = SWAPBLK_NONE;
1920 if (atomic_cmpset_int(&swblk_zone_exhausted,
1922 printf("swblk zone ok\n");
1925 VM_OBJECT_WUNLOCK(object);
1926 if (uma_zone_exhausted(swblk_zone)) {
1927 if (atomic_cmpset_int(&swblk_zone_exhausted,
1929 printf("swap blk zone exhausted, "
1930 "increase kern.maxswzone\n");
1931 vm_pageout_oom(VM_OOM_SWAPZ);
1932 pause("swzonxb", 10);
1934 uma_zwait(swblk_zone);
1935 VM_OBJECT_WLOCK(object);
1936 sb = SWAP_PCTRIE_LOOKUP(&object->un_pager.swp.swp_blks,
1940 * Somebody swapped out a nearby page,
1941 * allocating swblk at the rdpi index,
1942 * while we dropped the object lock.
1947 error = SWAP_PCTRIE_INSERT(
1948 &object->un_pager.swp.swp_blks, sb);
1950 if (atomic_cmpset_int(&swpctrie_zone_exhausted,
1952 printf("swpctrie zone ok\n");
1955 VM_OBJECT_WUNLOCK(object);
1956 if (uma_zone_exhausted(swpctrie_zone)) {
1957 if (atomic_cmpset_int(&swpctrie_zone_exhausted,
1959 printf("swap pctrie zone exhausted, "
1960 "increase kern.maxswzone\n");
1961 vm_pageout_oom(VM_OOM_SWAPZ);
1962 pause("swzonxp", 10);
1964 uma_zwait(swpctrie_zone);
1965 VM_OBJECT_WLOCK(object);
1966 sb1 = SWAP_PCTRIE_LOOKUP(&object->un_pager.swp.swp_blks,
1969 uma_zfree(swblk_zone, sb);
1976 MPASS(sb->p == rdpi);
1978 modpi = pindex % SWAP_META_PAGES;
1979 /* Return prior contents of metadata. */
1980 prev_swapblk = sb->d[modpi];
1981 /* Enter block into metadata. */
1982 sb->d[modpi] = swapblk;
1985 * Free the swblk if we end up with the empty page run.
1987 if (swapblk == SWAPBLK_NONE &&
1988 swp_pager_swblk_empty(sb, 0, SWAP_META_PAGES)) {
1989 SWAP_PCTRIE_REMOVE(&object->un_pager.swp.swp_blks, rdpi);
1990 uma_zfree(swblk_zone, sb);
1992 return (prev_swapblk);
1996 * SWP_PAGER_META_FREE() - free a range of blocks in the object's swap metadata
1998 * The requested range of blocks is freed, with any associated swap
1999 * returned to the swap bitmap.
2001 * This routine will free swap metadata structures as they are cleaned
2002 * out. This routine does *NOT* operate on swap metadata associated
2003 * with resident pages.
2006 swp_pager_meta_free(vm_object_t object, vm_pindex_t pindex, vm_pindex_t count)
2009 daddr_t n_free, s_free;
2011 int i, limit, start;
2013 VM_OBJECT_ASSERT_WLOCKED(object);
2014 if (object->type != OBJT_SWAP || count == 0)
2017 swp_pager_init_freerange(&s_free, &n_free);
2018 last = pindex + count;
2020 sb = SWAP_PCTRIE_LOOKUP_GE(&object->un_pager.swp.swp_blks,
2021 rounddown(pindex, SWAP_META_PAGES));
2022 if (sb == NULL || sb->p >= last)
2024 start = pindex > sb->p ? pindex - sb->p : 0;
2025 limit = last - sb->p < SWAP_META_PAGES ? last - sb->p :
2027 for (i = start; i < limit; i++) {
2028 if (sb->d[i] == SWAPBLK_NONE)
2030 swp_pager_update_freerange(&s_free, &n_free, sb->d[i]);
2031 sb->d[i] = SWAPBLK_NONE;
2033 pindex = sb->p + SWAP_META_PAGES;
2034 if (swp_pager_swblk_empty(sb, 0, start) &&
2035 swp_pager_swblk_empty(sb, limit, SWAP_META_PAGES)) {
2036 SWAP_PCTRIE_REMOVE(&object->un_pager.swp.swp_blks,
2038 uma_zfree(swblk_zone, sb);
2041 swp_pager_freeswapspace(s_free, n_free);
2045 * SWP_PAGER_META_FREE_ALL() - destroy all swap metadata associated with object
2047 * This routine locates and destroys all swap metadata associated with
2051 swp_pager_meta_free_all(vm_object_t object)
2054 daddr_t n_free, s_free;
2058 VM_OBJECT_ASSERT_WLOCKED(object);
2059 if (object->type != OBJT_SWAP)
2062 swp_pager_init_freerange(&s_free, &n_free);
2063 for (pindex = 0; (sb = SWAP_PCTRIE_LOOKUP_GE(
2064 &object->un_pager.swp.swp_blks, pindex)) != NULL;) {
2065 pindex = sb->p + SWAP_META_PAGES;
2066 for (i = 0; i < SWAP_META_PAGES; i++) {
2067 if (sb->d[i] == SWAPBLK_NONE)
2069 swp_pager_update_freerange(&s_free, &n_free, sb->d[i]);
2071 SWAP_PCTRIE_REMOVE(&object->un_pager.swp.swp_blks, sb->p);
2072 uma_zfree(swblk_zone, sb);
2074 swp_pager_freeswapspace(s_free, n_free);
2078 * SWP_PAGER_METACTL() - misc control of swap meta data.
2080 * This routine is capable of looking up, or removing swapblk
2081 * assignments in the swap meta data. It returns the swapblk being
2082 * looked-up, popped, or SWAPBLK_NONE if the block was invalid.
2084 * When acting on a busy resident page and paging is in progress, we
2085 * have to wait until paging is complete but otherwise can act on the
2088 * SWM_POP remove from meta data but do not free it
2091 swp_pager_meta_ctl(vm_object_t object, vm_pindex_t pindex, int flags)
2096 if ((flags & SWM_POP) != 0)
2097 VM_OBJECT_ASSERT_WLOCKED(object);
2099 VM_OBJECT_ASSERT_LOCKED(object);
2102 * The meta data only exists if the object is OBJT_SWAP
2103 * and even then might not be allocated yet.
2105 if (object->type != OBJT_SWAP)
2106 return (SWAPBLK_NONE);
2108 sb = SWAP_PCTRIE_LOOKUP(&object->un_pager.swp.swp_blks,
2109 rounddown(pindex, SWAP_META_PAGES));
2111 return (SWAPBLK_NONE);
2112 r1 = sb->d[pindex % SWAP_META_PAGES];
2113 if (r1 == SWAPBLK_NONE)
2114 return (SWAPBLK_NONE);
2115 if ((flags & SWM_POP) != 0) {
2116 sb->d[pindex % SWAP_META_PAGES] = SWAPBLK_NONE;
2117 if (swp_pager_swblk_empty(sb, 0, SWAP_META_PAGES)) {
2118 SWAP_PCTRIE_REMOVE(&object->un_pager.swp.swp_blks,
2119 rounddown(pindex, SWAP_META_PAGES));
2120 uma_zfree(swblk_zone, sb);
2127 * Returns the least page index which is greater than or equal to the
2128 * parameter pindex and for which there is a swap block allocated.
2129 * Returns object's size if the object's type is not swap or if there
2130 * are no allocated swap blocks for the object after the requested
2134 swap_pager_find_least(vm_object_t object, vm_pindex_t pindex)
2139 VM_OBJECT_ASSERT_LOCKED(object);
2140 if (object->type != OBJT_SWAP)
2141 return (object->size);
2143 sb = SWAP_PCTRIE_LOOKUP_GE(&object->un_pager.swp.swp_blks,
2144 rounddown(pindex, SWAP_META_PAGES));
2146 return (object->size);
2147 if (sb->p < pindex) {
2148 for (i = pindex % SWAP_META_PAGES; i < SWAP_META_PAGES; i++) {
2149 if (sb->d[i] != SWAPBLK_NONE)
2152 sb = SWAP_PCTRIE_LOOKUP_GE(&object->un_pager.swp.swp_blks,
2153 roundup(pindex, SWAP_META_PAGES));
2155 return (object->size);
2157 for (i = 0; i < SWAP_META_PAGES; i++) {
2158 if (sb->d[i] != SWAPBLK_NONE)
2163 * We get here if a swblk is present in the trie but it
2164 * doesn't map any blocks.
2167 return (object->size);
2171 * System call swapon(name) enables swapping on device name,
2172 * which must be in the swdevsw. Return EBUSY
2173 * if already swapping on this device.
2175 #ifndef _SYS_SYSPROTO_H_
2176 struct swapon_args {
2186 sys_swapon(struct thread *td, struct swapon_args *uap)
2190 struct nameidata nd;
2193 error = priv_check(td, PRIV_SWAPON);
2197 sx_xlock(&swdev_syscall_lock);
2200 * Swap metadata may not fit in the KVM if we have physical
2203 if (swblk_zone == NULL) {
2208 NDINIT(&nd, LOOKUP, ISOPEN | FOLLOW | AUDITVNODE1, UIO_USERSPACE,
2214 NDFREE(&nd, NDF_ONLY_PNBUF);
2217 if (vn_isdisk(vp, &error)) {
2218 error = swapongeom(vp);
2219 } else if (vp->v_type == VREG &&
2220 (vp->v_mount->mnt_vfc->vfc_flags & VFCF_NETWORK) != 0 &&
2221 (error = VOP_GETATTR(vp, &attr, td->td_ucred)) == 0) {
2223 * Allow direct swapping to NFS regular files in the same
2224 * way that nfs_mountroot() sets up diskless swapping.
2226 error = swaponvp(td, vp, attr.va_size / DEV_BSIZE);
2232 sx_xunlock(&swdev_syscall_lock);
2237 * Check that the total amount of swap currently configured does not
2238 * exceed half the theoretical maximum. If it does, print a warning
2242 swapon_check_swzone(void)
2244 unsigned long maxpages, npages;
2246 npages = swap_total;
2247 /* absolute maximum we can handle assuming 100% efficiency */
2248 maxpages = uma_zone_get_max(swblk_zone) * SWAP_META_PAGES;
2250 /* recommend using no more than half that amount */
2251 if (npages > maxpages / 2) {
2252 printf("warning: total configured swap (%lu pages) "
2253 "exceeds maximum recommended amount (%lu pages).\n",
2254 npages, maxpages / 2);
2255 printf("warning: increase kern.maxswzone "
2256 "or reduce amount of swap.\n");
2261 swaponsomething(struct vnode *vp, void *id, u_long nblks,
2262 sw_strategy_t *strategy, sw_close_t *close, dev_t dev, int flags)
2264 struct swdevt *sp, *tsp;
2269 * nblks is in DEV_BSIZE'd chunks, convert to PAGE_SIZE'd chunks.
2270 * First chop nblks off to page-align it, then convert.
2272 * sw->sw_nblks is in page-sized chunks now too.
2274 nblks &= ~(ctodb(1) - 1);
2275 nblks = dbtoc(nblks);
2278 * If we go beyond this, we get overflows in the radix
2281 mblocks = 0x40000000 / BLIST_META_RADIX;
2282 if (nblks > mblocks) {
2284 "WARNING: reducing swap size to maximum of %luMB per unit\n",
2285 mblocks / 1024 / 1024 * PAGE_SIZE);
2289 sp = malloc(sizeof *sp, M_VMPGDATA, M_WAITOK | M_ZERO);
2293 sp->sw_nblks = nblks;
2295 sp->sw_strategy = strategy;
2296 sp->sw_close = close;
2297 sp->sw_flags = flags;
2299 sp->sw_blist = blist_create(nblks, M_WAITOK);
2301 * Do not free the first two block in order to avoid overwriting
2302 * any bsd label at the front of the partition
2304 blist_free(sp->sw_blist, 2, nblks - 2);
2307 mtx_lock(&sw_dev_mtx);
2308 TAILQ_FOREACH(tsp, &swtailq, sw_list) {
2309 if (tsp->sw_end >= dvbase) {
2311 * We put one uncovered page between the devices
2312 * in order to definitively prevent any cross-device
2315 dvbase = tsp->sw_end + 1;
2318 sp->sw_first = dvbase;
2319 sp->sw_end = dvbase + nblks;
2320 TAILQ_INSERT_TAIL(&swtailq, sp, sw_list);
2322 swap_pager_avail += nblks - 2;
2323 swap_total += nblks;
2324 swapon_check_swzone();
2326 mtx_unlock(&sw_dev_mtx);
2327 EVENTHANDLER_INVOKE(swapon, sp);
2331 * SYSCALL: swapoff(devname)
2333 * Disable swapping on the given device.
2335 * XXX: Badly designed system call: it should use a device index
2336 * rather than filename as specification. We keep sw_vp around
2337 * only to make this work.
2339 #ifndef _SYS_SYSPROTO_H_
2340 struct swapoff_args {
2350 sys_swapoff(struct thread *td, struct swapoff_args *uap)
2353 struct nameidata nd;
2357 error = priv_check(td, PRIV_SWAPOFF);
2361 sx_xlock(&swdev_syscall_lock);
2363 NDINIT(&nd, LOOKUP, FOLLOW | AUDITVNODE1, UIO_USERSPACE, uap->name,
2368 NDFREE(&nd, NDF_ONLY_PNBUF);
2371 mtx_lock(&sw_dev_mtx);
2372 TAILQ_FOREACH(sp, &swtailq, sw_list) {
2373 if (sp->sw_vp == vp)
2376 mtx_unlock(&sw_dev_mtx);
2381 error = swapoff_one(sp, td->td_ucred);
2383 sx_xunlock(&swdev_syscall_lock);
2388 swapoff_one(struct swdevt *sp, struct ucred *cred)
2395 sx_assert(&swdev_syscall_lock, SA_XLOCKED);
2397 (void) vn_lock(sp->sw_vp, LK_EXCLUSIVE | LK_RETRY);
2398 error = mac_system_check_swapoff(cred, sp->sw_vp);
2399 (void) VOP_UNLOCK(sp->sw_vp, 0);
2403 nblks = sp->sw_nblks;
2406 * We can turn off this swap device safely only if the
2407 * available virtual memory in the system will fit the amount
2408 * of data we will have to page back in, plus an epsilon so
2409 * the system doesn't become critically low on swap space.
2411 if (vm_free_count() + swap_pager_avail < nblks + nswap_lowat)
2415 * Prevent further allocations on this device.
2417 mtx_lock(&sw_dev_mtx);
2418 sp->sw_flags |= SW_CLOSING;
2419 swap_pager_avail -= blist_fill(sp->sw_blist, 0, nblks);
2420 swap_total -= nblks;
2421 mtx_unlock(&sw_dev_mtx);
2424 * Page in the contents of the device and close it.
2426 swap_pager_swapoff(sp);
2428 sp->sw_close(curthread, sp);
2429 mtx_lock(&sw_dev_mtx);
2431 TAILQ_REMOVE(&swtailq, sp, sw_list);
2433 if (nswapdev == 0) {
2434 swap_pager_full = 2;
2435 swap_pager_almost_full = 1;
2439 mtx_unlock(&sw_dev_mtx);
2440 blist_destroy(sp->sw_blist);
2441 free(sp, M_VMPGDATA);
2448 struct swdevt *sp, *spt;
2449 const char *devname;
2452 sx_xlock(&swdev_syscall_lock);
2454 mtx_lock(&sw_dev_mtx);
2455 TAILQ_FOREACH_SAFE(sp, &swtailq, sw_list, spt) {
2456 mtx_unlock(&sw_dev_mtx);
2457 if (vn_isdisk(sp->sw_vp, NULL))
2458 devname = devtoname(sp->sw_vp->v_rdev);
2461 error = swapoff_one(sp, thread0.td_ucred);
2463 printf("Cannot remove swap device %s (error=%d), "
2464 "skipping.\n", devname, error);
2465 } else if (bootverbose) {
2466 printf("Swap device %s removed.\n", devname);
2468 mtx_lock(&sw_dev_mtx);
2470 mtx_unlock(&sw_dev_mtx);
2472 sx_xunlock(&swdev_syscall_lock);
2476 swap_pager_status(int *total, int *used)
2482 mtx_lock(&sw_dev_mtx);
2483 TAILQ_FOREACH(sp, &swtailq, sw_list) {
2484 *total += sp->sw_nblks;
2485 *used += sp->sw_used;
2487 mtx_unlock(&sw_dev_mtx);
2491 swap_dev_info(int name, struct xswdev *xs, char *devname, size_t len)
2494 const char *tmp_devname;
2499 mtx_lock(&sw_dev_mtx);
2500 TAILQ_FOREACH(sp, &swtailq, sw_list) {
2505 xs->xsw_version = XSWDEV_VERSION;
2506 xs->xsw_dev = sp->sw_dev;
2507 xs->xsw_flags = sp->sw_flags;
2508 xs->xsw_nblks = sp->sw_nblks;
2509 xs->xsw_used = sp->sw_used;
2510 if (devname != NULL) {
2511 if (vn_isdisk(sp->sw_vp, NULL))
2512 tmp_devname = devtoname(sp->sw_vp->v_rdev);
2514 tmp_devname = "[file]";
2515 strncpy(devname, tmp_devname, len);
2520 mtx_unlock(&sw_dev_mtx);
2524 #if defined(COMPAT_FREEBSD11)
2525 #define XSWDEV_VERSION_11 1
2535 #if defined(__amd64__) && defined(COMPAT_FREEBSD32)
2538 u_int xsw_dev1, xsw_dev2;
2546 sysctl_vm_swap_info(SYSCTL_HANDLER_ARGS)
2549 #if defined(__amd64__) && defined(COMPAT_FREEBSD32)
2550 struct xswdev32 xs32;
2552 #if defined(COMPAT_FREEBSD11)
2553 struct xswdev11 xs11;
2557 if (arg2 != 1) /* name length */
2559 error = swap_dev_info(*(int *)arg1, &xs, NULL, 0);
2562 #if defined(__amd64__) && defined(COMPAT_FREEBSD32)
2563 if (req->oldlen == sizeof(xs32)) {
2564 xs32.xsw_version = XSWDEV_VERSION;
2565 xs32.xsw_dev1 = xs.xsw_dev;
2566 xs32.xsw_dev2 = xs.xsw_dev >> 32;
2567 xs32.xsw_flags = xs.xsw_flags;
2568 xs32.xsw_nblks = xs.xsw_nblks;
2569 xs32.xsw_used = xs.xsw_used;
2570 error = SYSCTL_OUT(req, &xs32, sizeof(xs32));
2574 #if defined(COMPAT_FREEBSD11)
2575 if (req->oldlen == sizeof(xs11)) {
2576 xs11.xsw_version = XSWDEV_VERSION_11;
2577 xs11.xsw_dev = xs.xsw_dev; /* truncation */
2578 xs11.xsw_flags = xs.xsw_flags;
2579 xs11.xsw_nblks = xs.xsw_nblks;
2580 xs11.xsw_used = xs.xsw_used;
2581 error = SYSCTL_OUT(req, &xs11, sizeof(xs11));
2585 error = SYSCTL_OUT(req, &xs, sizeof(xs));
2589 SYSCTL_INT(_vm, OID_AUTO, nswapdev, CTLFLAG_RD, &nswapdev, 0,
2590 "Number of swap devices");
2591 SYSCTL_NODE(_vm, OID_AUTO, swap_info, CTLFLAG_RD | CTLFLAG_MPSAFE,
2592 sysctl_vm_swap_info,
2593 "Swap statistics by device");
2596 * Count the approximate swap usage in pages for a vmspace. The
2597 * shadowed or not yet copied on write swap blocks are not accounted.
2598 * The map must be locked.
2601 vmspace_swap_count(struct vmspace *vmspace)
2611 map = &vmspace->vm_map;
2614 for (cur = map->header.next; cur != &map->header; cur = cur->next) {
2615 if ((cur->eflags & MAP_ENTRY_IS_SUB_MAP) != 0)
2617 object = cur->object.vm_object;
2618 if (object == NULL || object->type != OBJT_SWAP)
2620 VM_OBJECT_RLOCK(object);
2621 if (object->type != OBJT_SWAP)
2623 pi = OFF_TO_IDX(cur->offset);
2624 e = pi + OFF_TO_IDX(cur->end - cur->start);
2625 for (;; pi = sb->p + SWAP_META_PAGES) {
2626 sb = SWAP_PCTRIE_LOOKUP_GE(
2627 &object->un_pager.swp.swp_blks, pi);
2628 if (sb == NULL || sb->p >= e)
2630 for (i = 0; i < SWAP_META_PAGES; i++) {
2631 if (sb->p + i < e &&
2632 sb->d[i] != SWAPBLK_NONE)
2637 VM_OBJECT_RUNLOCK(object);
2645 * Swapping onto disk devices.
2649 static g_orphan_t swapgeom_orphan;
2651 static struct g_class g_swap_class = {
2653 .version = G_VERSION,
2654 .orphan = swapgeom_orphan,
2657 DECLARE_GEOM_CLASS(g_swap_class, g_class);
2661 swapgeom_close_ev(void *arg, int flags)
2663 struct g_consumer *cp;
2666 g_access(cp, -1, -1, 0);
2668 g_destroy_consumer(cp);
2672 * Add a reference to the g_consumer for an inflight transaction.
2675 swapgeom_acquire(struct g_consumer *cp)
2678 mtx_assert(&sw_dev_mtx, MA_OWNED);
2683 * Remove a reference from the g_consumer. Post a close event if all
2684 * references go away, since the function might be called from the
2688 swapgeom_release(struct g_consumer *cp, struct swdevt *sp)
2691 mtx_assert(&sw_dev_mtx, MA_OWNED);
2693 if (cp->index == 0) {
2694 if (g_post_event(swapgeom_close_ev, cp, M_NOWAIT, NULL) == 0)
2700 swapgeom_done(struct bio *bp2)
2704 struct g_consumer *cp;
2706 bp = bp2->bio_caller2;
2708 bp->b_ioflags = bp2->bio_flags;
2710 bp->b_ioflags |= BIO_ERROR;
2711 bp->b_resid = bp->b_bcount - bp2->bio_completed;
2712 bp->b_error = bp2->bio_error;
2713 bp->b_caller1 = NULL;
2715 sp = bp2->bio_caller1;
2716 mtx_lock(&sw_dev_mtx);
2717 swapgeom_release(cp, sp);
2718 mtx_unlock(&sw_dev_mtx);
2723 swapgeom_strategy(struct buf *bp, struct swdevt *sp)
2726 struct g_consumer *cp;
2728 mtx_lock(&sw_dev_mtx);
2731 mtx_unlock(&sw_dev_mtx);
2732 bp->b_error = ENXIO;
2733 bp->b_ioflags |= BIO_ERROR;
2737 swapgeom_acquire(cp);
2738 mtx_unlock(&sw_dev_mtx);
2739 if (bp->b_iocmd == BIO_WRITE)
2742 bio = g_alloc_bio();
2744 mtx_lock(&sw_dev_mtx);
2745 swapgeom_release(cp, sp);
2746 mtx_unlock(&sw_dev_mtx);
2747 bp->b_error = ENOMEM;
2748 bp->b_ioflags |= BIO_ERROR;
2749 printf("swap_pager: cannot allocate bio\n");
2754 bp->b_caller1 = bio;
2755 bio->bio_caller1 = sp;
2756 bio->bio_caller2 = bp;
2757 bio->bio_cmd = bp->b_iocmd;
2758 bio->bio_offset = (bp->b_blkno - sp->sw_first) * PAGE_SIZE;
2759 bio->bio_length = bp->b_bcount;
2760 bio->bio_done = swapgeom_done;
2761 if (!buf_mapped(bp)) {
2762 bio->bio_ma = bp->b_pages;
2763 bio->bio_data = unmapped_buf;
2764 bio->bio_ma_offset = (vm_offset_t)bp->b_offset & PAGE_MASK;
2765 bio->bio_ma_n = bp->b_npages;
2766 bio->bio_flags |= BIO_UNMAPPED;
2768 bio->bio_data = bp->b_data;
2771 g_io_request(bio, cp);
2776 swapgeom_orphan(struct g_consumer *cp)
2781 mtx_lock(&sw_dev_mtx);
2782 TAILQ_FOREACH(sp, &swtailq, sw_list) {
2783 if (sp->sw_id == cp) {
2784 sp->sw_flags |= SW_CLOSING;
2789 * Drop reference we were created with. Do directly since we're in a
2790 * special context where we don't have to queue the call to
2791 * swapgeom_close_ev().
2794 destroy = ((sp != NULL) && (cp->index == 0));
2797 mtx_unlock(&sw_dev_mtx);
2799 swapgeom_close_ev(cp, 0);
2803 swapgeom_close(struct thread *td, struct swdevt *sw)
2805 struct g_consumer *cp;
2807 mtx_lock(&sw_dev_mtx);
2810 mtx_unlock(&sw_dev_mtx);
2813 * swapgeom_close() may be called from the biodone context,
2814 * where we cannot perform topology changes. Delegate the
2815 * work to the events thread.
2818 g_waitfor_event(swapgeom_close_ev, cp, M_WAITOK, NULL);
2822 swapongeom_locked(struct cdev *dev, struct vnode *vp)
2824 struct g_provider *pp;
2825 struct g_consumer *cp;
2826 static struct g_geom *gp;
2831 pp = g_dev_getprovider(dev);
2834 mtx_lock(&sw_dev_mtx);
2835 TAILQ_FOREACH(sp, &swtailq, sw_list) {
2837 if (cp != NULL && cp->provider == pp) {
2838 mtx_unlock(&sw_dev_mtx);
2842 mtx_unlock(&sw_dev_mtx);
2844 gp = g_new_geomf(&g_swap_class, "swap");
2845 cp = g_new_consumer(gp);
2846 cp->index = 1; /* Number of active I/Os, plus one for being active. */
2847 cp->flags |= G_CF_DIRECT_SEND | G_CF_DIRECT_RECEIVE;
2850 * XXX: Every time you think you can improve the margin for
2851 * footshooting, somebody depends on the ability to do so:
2852 * savecore(8) wants to write to our swapdev so we cannot
2853 * set an exclusive count :-(
2855 error = g_access(cp, 1, 1, 0);
2858 g_destroy_consumer(cp);
2861 nblks = pp->mediasize / DEV_BSIZE;
2862 swaponsomething(vp, cp, nblks, swapgeom_strategy,
2863 swapgeom_close, dev2udev(dev),
2864 (pp->flags & G_PF_ACCEPT_UNMAPPED) != 0 ? SW_UNMAPPED : 0);
2869 swapongeom(struct vnode *vp)
2873 vn_lock(vp, LK_EXCLUSIVE | LK_RETRY);
2874 if (vp->v_type != VCHR || (vp->v_iflag & VI_DOOMED) != 0) {
2878 error = swapongeom_locked(vp->v_rdev, vp);
2879 g_topology_unlock();
2888 * This is used mainly for network filesystem (read: probably only tested
2889 * with NFS) swapfiles.
2894 swapdev_strategy(struct buf *bp, struct swdevt *sp)
2898 bp->b_blkno = ctodb(bp->b_blkno - sp->sw_first);
2902 if (bp->b_iocmd == BIO_WRITE) {
2904 bufobj_wdrop(bp->b_bufobj);
2905 bufobj_wref(&vp2->v_bufobj);
2907 if (bp->b_bufobj != &vp2->v_bufobj)
2908 bp->b_bufobj = &vp2->v_bufobj;
2910 bp->b_iooffset = dbtob(bp->b_blkno);
2916 swapdev_close(struct thread *td, struct swdevt *sp)
2919 VOP_CLOSE(sp->sw_vp, FREAD | FWRITE, td->td_ucred, td);
2925 swaponvp(struct thread *td, struct vnode *vp, u_long nblks)
2932 mtx_lock(&sw_dev_mtx);
2933 TAILQ_FOREACH(sp, &swtailq, sw_list) {
2934 if (sp->sw_id == vp) {
2935 mtx_unlock(&sw_dev_mtx);
2939 mtx_unlock(&sw_dev_mtx);
2941 (void) vn_lock(vp, LK_EXCLUSIVE | LK_RETRY);
2943 error = mac_system_check_swapon(td->td_ucred, vp);
2946 error = VOP_OPEN(vp, FREAD | FWRITE, td->td_ucred, td, NULL);
2947 (void) VOP_UNLOCK(vp, 0);
2951 swaponsomething(vp, vp, nblks, swapdev_strategy, swapdev_close,
2957 sysctl_swap_async_max(SYSCTL_HANDLER_ARGS)
2961 new = nsw_wcount_async_max;
2962 error = sysctl_handle_int(oidp, &new, 0, req);
2963 if (error != 0 || req->newptr == NULL)
2966 if (new > nswbuf / 2 || new < 1)
2969 mtx_lock(&swbuf_mtx);
2970 while (nsw_wcount_async_max != new) {
2972 * Adjust difference. If the current async count is too low,
2973 * we will need to sqeeze our update slowly in. Sleep with a
2974 * higher priority than getpbuf() to finish faster.
2976 n = new - nsw_wcount_async_max;
2977 if (nsw_wcount_async + n >= 0) {
2978 nsw_wcount_async += n;
2979 nsw_wcount_async_max += n;
2980 wakeup(&nsw_wcount_async);
2982 nsw_wcount_async_max -= nsw_wcount_async;
2983 nsw_wcount_async = 0;
2984 msleep(&nsw_wcount_async, &swbuf_mtx, PSWP,
2988 mtx_unlock(&swbuf_mtx);