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/disklabel.h>
83 #include <sys/eventhandler.h>
84 #include <sys/fcntl.h>
86 #include <sys/kernel.h>
87 #include <sys/mount.h>
88 #include <sys/namei.h>
89 #include <sys/malloc.h>
90 #include <sys/pctrie.h>
93 #include <sys/racct.h>
94 #include <sys/resource.h>
95 #include <sys/resourcevar.h>
96 #include <sys/rwlock.h>
98 #include <sys/sysctl.h>
99 #include <sys/sysproto.h>
100 #include <sys/systm.h>
102 #include <sys/vmmeter.h>
103 #include <sys/vnode.h>
105 #include <security/mac/mac_framework.h>
109 #include <vm/vm_map.h>
110 #include <vm/vm_kern.h>
111 #include <vm/vm_object.h>
112 #include <vm/vm_page.h>
113 #include <vm/vm_pager.h>
114 #include <vm/vm_pageout.h>
115 #include <vm/vm_param.h>
116 #include <vm/swap_pager.h>
117 #include <vm/vm_extern.h>
120 #include <geom/geom.h>
123 * MAX_PAGEOUT_CLUSTER must be a power of 2 between 1 and 64.
124 * The 64-page limit is due to the radix code (kern/subr_blist.c).
126 #ifndef MAX_PAGEOUT_CLUSTER
127 #define MAX_PAGEOUT_CLUSTER 32
130 #if !defined(SWB_NPAGES)
131 #define SWB_NPAGES MAX_PAGEOUT_CLUSTER
134 #define SWAP_META_PAGES PCTRIE_COUNT
137 * A swblk structure maps each page index within a
138 * SWAP_META_PAGES-aligned and sized range to the address of an
139 * on-disk swap block (or SWAPBLK_NONE). The collection of these
140 * mappings for an entire vm object is implemented as a pc-trie.
144 daddr_t d[SWAP_META_PAGES];
147 static MALLOC_DEFINE(M_VMPGDATA, "vm_pgdata", "swap pager private data");
148 static struct mtx sw_dev_mtx;
149 static TAILQ_HEAD(, swdevt) swtailq = TAILQ_HEAD_INITIALIZER(swtailq);
150 static struct swdevt *swdevhd; /* Allocate from here next */
151 static int nswapdev; /* Number of swap devices */
152 int swap_pager_avail;
153 static struct sx swdev_syscall_lock; /* serialize swap(on|off) */
155 static __exclusive_cache_line u_long swap_reserved;
156 static u_long swap_total;
157 static int sysctl_page_shift(SYSCTL_HANDLER_ARGS);
159 static SYSCTL_NODE(_vm_stats, OID_AUTO, swap, CTLFLAG_RD | CTLFLAG_MPSAFE, 0,
162 SYSCTL_PROC(_vm, OID_AUTO, swap_reserved, CTLTYPE_U64 | CTLFLAG_RD | CTLFLAG_MPSAFE,
163 &swap_reserved, 0, sysctl_page_shift, "A",
164 "Amount of swap storage needed to back all allocated anonymous memory.");
165 SYSCTL_PROC(_vm, OID_AUTO, swap_total, CTLTYPE_U64 | CTLFLAG_RD | CTLFLAG_MPSAFE,
166 &swap_total, 0, sysctl_page_shift, "A",
167 "Total amount of available swap storage.");
169 static int overcommit = 0;
170 SYSCTL_INT(_vm, VM_OVERCOMMIT, overcommit, CTLFLAG_RW, &overcommit, 0,
171 "Configure virtual memory overcommit behavior. See tuning(7) "
173 static unsigned long swzone;
174 SYSCTL_ULONG(_vm, OID_AUTO, swzone, CTLFLAG_RD, &swzone, 0,
175 "Actual size of swap metadata zone");
176 static unsigned long swap_maxpages;
177 SYSCTL_ULONG(_vm, OID_AUTO, swap_maxpages, CTLFLAG_RD, &swap_maxpages, 0,
178 "Maximum amount of swap supported");
180 static COUNTER_U64_DEFINE_EARLY(swap_free_deferred);
181 SYSCTL_COUNTER_U64(_vm_stats_swap, OID_AUTO, free_deferred,
182 CTLFLAG_RD, &swap_free_deferred,
183 "Number of pages that deferred freeing swap space");
185 static COUNTER_U64_DEFINE_EARLY(swap_free_completed);
186 SYSCTL_COUNTER_U64(_vm_stats_swap, OID_AUTO, free_completed,
187 CTLFLAG_RD, &swap_free_completed,
188 "Number of deferred frees completed");
190 /* bits from overcommit */
191 #define SWAP_RESERVE_FORCE_ON (1 << 0)
192 #define SWAP_RESERVE_RLIMIT_ON (1 << 1)
193 #define SWAP_RESERVE_ALLOW_NONWIRED (1 << 2)
196 sysctl_page_shift(SYSCTL_HANDLER_ARGS)
199 u_long value = *(u_long *)arg1;
201 newval = ((uint64_t)value) << PAGE_SHIFT;
202 return (sysctl_handle_64(oidp, &newval, 0, req));
206 swap_reserve_by_cred_rlimit(u_long pincr, struct ucred *cred, int oc)
211 uip = cred->cr_ruidinfo;
213 prev = atomic_fetchadd_long(&uip->ui_vmsize, pincr);
214 if ((oc & SWAP_RESERVE_RLIMIT_ON) != 0 &&
215 prev + pincr > lim_cur(curthread, RLIMIT_SWAP) &&
216 priv_check(curthread, PRIV_VM_SWAP_NORLIMIT) != 0) {
217 prev = atomic_fetchadd_long(&uip->ui_vmsize, -pincr);
218 KASSERT(prev >= pincr, ("negative vmsize for uid = %d\n", uip->ui_uid));
225 swap_release_by_cred_rlimit(u_long pdecr, struct ucred *cred)
232 uip = cred->cr_ruidinfo;
235 prev = atomic_fetchadd_long(&uip->ui_vmsize, -pdecr);
236 KASSERT(prev >= pdecr, ("negative vmsize for uid = %d\n", uip->ui_uid));
238 atomic_subtract_long(&uip->ui_vmsize, pdecr);
243 swap_reserve_force_rlimit(u_long pincr, struct ucred *cred)
247 uip = cred->cr_ruidinfo;
248 atomic_add_long(&uip->ui_vmsize, pincr);
252 swap_reserve(vm_ooffset_t incr)
255 return (swap_reserve_by_cred(incr, curthread->td_ucred));
259 swap_reserve_by_cred(vm_ooffset_t incr, struct ucred *cred)
261 u_long r, s, prev, pincr;
267 static struct timeval lastfail;
269 KASSERT((incr & PAGE_MASK) == 0, ("%s: incr: %ju & PAGE_MASK", __func__,
273 if (RACCT_ENABLED()) {
275 error = racct_add(curproc, RACCT_SWAP, incr);
276 PROC_UNLOCK(curproc);
283 prev = atomic_fetchadd_long(&swap_reserved, pincr);
286 oc = atomic_load_int(&overcommit);
287 if (r > s && (oc & SWAP_RESERVE_ALLOW_NONWIRED) != 0) {
288 s += vm_cnt.v_page_count - vm_cnt.v_free_reserved -
291 if ((oc & SWAP_RESERVE_FORCE_ON) != 0 && r > s &&
292 priv_check(curthread, PRIV_VM_SWAP_NOQUOTA) != 0) {
293 prev = atomic_fetchadd_long(&swap_reserved, -pincr);
294 KASSERT(prev >= pincr, ("swap_reserved < incr on overcommit fail"));
298 if (!swap_reserve_by_cred_rlimit(pincr, cred, oc)) {
299 prev = atomic_fetchadd_long(&swap_reserved, -pincr);
300 KASSERT(prev >= pincr, ("swap_reserved < incr on overcommit fail"));
307 if (ppsratecheck(&lastfail, &curfail, 1)) {
308 printf("uid %d, pid %d: swap reservation for %jd bytes failed\n",
309 cred->cr_ruidinfo->ui_uid, curproc->p_pid, incr);
312 if (RACCT_ENABLED()) {
314 racct_sub(curproc, RACCT_SWAP, incr);
315 PROC_UNLOCK(curproc);
323 swap_reserve_force(vm_ooffset_t incr)
327 KASSERT((incr & PAGE_MASK) == 0, ("%s: incr: %ju & PAGE_MASK", __func__,
331 if (RACCT_ENABLED()) {
333 racct_add_force(curproc, RACCT_SWAP, incr);
334 PROC_UNLOCK(curproc);
338 atomic_add_long(&swap_reserved, pincr);
339 swap_reserve_force_rlimit(pincr, curthread->td_ucred);
343 swap_release(vm_ooffset_t decr)
348 cred = curproc->p_ucred;
349 swap_release_by_cred(decr, cred);
350 PROC_UNLOCK(curproc);
354 swap_release_by_cred(vm_ooffset_t decr, struct ucred *cred)
361 KASSERT((decr & PAGE_MASK) == 0, ("%s: decr: %ju & PAGE_MASK", __func__,
366 prev = atomic_fetchadd_long(&swap_reserved, -pdecr);
367 KASSERT(prev >= pdecr, ("swap_reserved < decr"));
369 atomic_subtract_long(&swap_reserved, pdecr);
372 swap_release_by_cred_rlimit(pdecr, cred);
375 racct_sub_cred(cred, RACCT_SWAP, decr);
379 static int swap_pager_full = 2; /* swap space exhaustion (task killing) */
380 static int swap_pager_almost_full = 1; /* swap space exhaustion (w/hysteresis)*/
381 static struct mtx swbuf_mtx; /* to sync nsw_wcount_async */
382 static int nsw_wcount_async; /* limit async write buffers */
383 static int nsw_wcount_async_max;/* assigned maximum */
384 static int nsw_cluster_max; /* maximum VOP I/O allowed */
386 static int sysctl_swap_async_max(SYSCTL_HANDLER_ARGS);
387 SYSCTL_PROC(_vm, OID_AUTO, swap_async_max, CTLTYPE_INT | CTLFLAG_RW |
388 CTLFLAG_MPSAFE, NULL, 0, sysctl_swap_async_max, "I",
389 "Maximum running async swap ops");
390 static int sysctl_swap_fragmentation(SYSCTL_HANDLER_ARGS);
391 SYSCTL_PROC(_vm, OID_AUTO, swap_fragmentation, CTLTYPE_STRING | CTLFLAG_RD |
392 CTLFLAG_MPSAFE, NULL, 0, sysctl_swap_fragmentation, "A",
393 "Swap Fragmentation Info");
395 static struct sx sw_alloc_sx;
398 * "named" and "unnamed" anon region objects. Try to reduce the overhead
399 * of searching a named list by hashing it just a little.
404 #define NOBJLIST(handle) \
405 (&swap_pager_object_list[((int)(intptr_t)handle >> 4) & (NOBJLISTS-1)])
407 static struct pagerlst swap_pager_object_list[NOBJLISTS];
408 static uma_zone_t swwbuf_zone;
409 static uma_zone_t swrbuf_zone;
410 static uma_zone_t swblk_zone;
411 static uma_zone_t swpctrie_zone;
414 * pagerops for OBJT_SWAP - "swap pager". Some ops are also global procedure
415 * calls hooked from other parts of the VM system and do not appear here.
416 * (see vm/swap_pager.h).
419 swap_pager_alloc(void *handle, vm_ooffset_t size,
420 vm_prot_t prot, vm_ooffset_t offset, struct ucred *);
421 static void swap_pager_dealloc(vm_object_t object);
422 static int swap_pager_getpages(vm_object_t, vm_page_t *, int, int *,
424 static int swap_pager_getpages_async(vm_object_t, vm_page_t *, int, int *,
425 int *, pgo_getpages_iodone_t, void *);
426 static void swap_pager_putpages(vm_object_t, vm_page_t *, int, boolean_t, int *);
428 swap_pager_haspage(vm_object_t object, vm_pindex_t pindex, int *before, int *after);
429 static void swap_pager_init(void);
430 static void swap_pager_unswapped(vm_page_t);
431 static void swap_pager_swapoff(struct swdevt *sp);
432 static void swap_pager_update_writecount(vm_object_t object,
433 vm_offset_t start, vm_offset_t end);
434 static void swap_pager_release_writecount(vm_object_t object,
435 vm_offset_t start, vm_offset_t end);
437 struct pagerops swappagerops = {
438 .pgo_init = swap_pager_init, /* early system initialization of pager */
439 .pgo_alloc = swap_pager_alloc, /* allocate an OBJT_SWAP object */
440 .pgo_dealloc = swap_pager_dealloc, /* deallocate an OBJT_SWAP object */
441 .pgo_getpages = swap_pager_getpages, /* pagein */
442 .pgo_getpages_async = swap_pager_getpages_async, /* pagein (async) */
443 .pgo_putpages = swap_pager_putpages, /* pageout */
444 .pgo_haspage = swap_pager_haspage, /* get backing store status for page */
445 .pgo_pageunswapped = swap_pager_unswapped, /* remove swap related to page */
446 .pgo_update_writecount = swap_pager_update_writecount,
447 .pgo_release_writecount = swap_pager_release_writecount,
451 * swap_*() routines are externally accessible. swp_*() routines are
454 static int nswap_lowat = 128; /* in pages, swap_pager_almost_full warn */
455 static int nswap_hiwat = 512; /* in pages, swap_pager_almost_full warn */
457 SYSCTL_INT(_vm, OID_AUTO, dmmax, CTLFLAG_RD, &nsw_cluster_max, 0,
458 "Maximum size of a swap block in pages");
460 static void swp_sizecheck(void);
461 static void swp_pager_async_iodone(struct buf *bp);
462 static bool swp_pager_swblk_empty(struct swblk *sb, int start, int limit);
463 static void swp_pager_free_empty_swblk(vm_object_t, struct swblk *sb);
464 static int swapongeom(struct vnode *);
465 static int swaponvp(struct thread *, struct vnode *, u_long);
466 static int swapoff_one(struct swdevt *sp, struct ucred *cred);
469 * Swap bitmap functions
471 static void swp_pager_freeswapspace(daddr_t blk, daddr_t npages);
472 static daddr_t swp_pager_getswapspace(int *npages);
477 static daddr_t swp_pager_meta_build(vm_object_t, vm_pindex_t, daddr_t);
478 static void swp_pager_meta_free(vm_object_t, vm_pindex_t, vm_pindex_t);
479 static void swp_pager_meta_transfer(vm_object_t src, vm_object_t dst,
480 vm_pindex_t pindex, vm_pindex_t count);
481 static void swp_pager_meta_free_all(vm_object_t);
482 static daddr_t swp_pager_meta_lookup(vm_object_t, vm_pindex_t);
485 swp_pager_init_freerange(daddr_t *start, daddr_t *num)
488 *start = SWAPBLK_NONE;
493 swp_pager_update_freerange(daddr_t *start, daddr_t *num, daddr_t addr)
496 if (*start + *num == addr) {
499 swp_pager_freeswapspace(*start, *num);
506 swblk_trie_alloc(struct pctrie *ptree)
509 return (uma_zalloc(swpctrie_zone, M_NOWAIT | (curproc == pageproc ?
510 M_USE_RESERVE : 0)));
514 swblk_trie_free(struct pctrie *ptree, void *node)
517 uma_zfree(swpctrie_zone, node);
520 PCTRIE_DEFINE(SWAP, swblk, p, swblk_trie_alloc, swblk_trie_free);
523 * SWP_SIZECHECK() - update swap_pager_full indication
525 * update the swap_pager_almost_full indication and warn when we are
526 * about to run out of swap space, using lowat/hiwat hysteresis.
528 * Clear swap_pager_full ( task killing ) indication when lowat is met.
530 * No restrictions on call
531 * This routine may not block.
537 if (swap_pager_avail < nswap_lowat) {
538 if (swap_pager_almost_full == 0) {
539 printf("swap_pager: out of swap space\n");
540 swap_pager_almost_full = 1;
544 if (swap_pager_avail > nswap_hiwat)
545 swap_pager_almost_full = 0;
550 * SWAP_PAGER_INIT() - initialize the swap pager!
552 * Expected to be started from system init. NOTE: This code is run
553 * before much else so be careful what you depend on. Most of the VM
554 * system has yet to be initialized at this point.
557 swap_pager_init(void)
560 * Initialize object lists
564 for (i = 0; i < NOBJLISTS; ++i)
565 TAILQ_INIT(&swap_pager_object_list[i]);
566 mtx_init(&sw_dev_mtx, "swapdev", NULL, MTX_DEF);
567 sx_init(&sw_alloc_sx, "swspsx");
568 sx_init(&swdev_syscall_lock, "swsysc");
572 * SWAP_PAGER_SWAP_INIT() - swap pager initialization from pageout process
574 * Expected to be started from pageout process once, prior to entering
578 swap_pager_swap_init(void)
583 * Number of in-transit swap bp operations. Don't
584 * exhaust the pbufs completely. Make sure we
585 * initialize workable values (0 will work for hysteresis
586 * but it isn't very efficient).
588 * The nsw_cluster_max is constrained by the bp->b_pages[]
589 * array, which has MAXPHYS / PAGE_SIZE entries, and our locally
590 * defined MAX_PAGEOUT_CLUSTER. Also be aware that swap ops are
591 * constrained by the swap device interleave stripe size.
593 * Currently we hardwire nsw_wcount_async to 4. This limit is
594 * designed to prevent other I/O from having high latencies due to
595 * our pageout I/O. The value 4 works well for one or two active swap
596 * devices but is probably a little low if you have more. Even so,
597 * a higher value would probably generate only a limited improvement
598 * with three or four active swap devices since the system does not
599 * typically have to pageout at extreme bandwidths. We will want
600 * at least 2 per swap devices, and 4 is a pretty good value if you
601 * have one NFS swap device due to the command/ack latency over NFS.
602 * So it all works out pretty well.
604 nsw_cluster_max = min(MAXPHYS / PAGE_SIZE, MAX_PAGEOUT_CLUSTER);
606 nsw_wcount_async = 4;
607 nsw_wcount_async_max = nsw_wcount_async;
608 mtx_init(&swbuf_mtx, "async swbuf mutex", NULL, MTX_DEF);
610 swwbuf_zone = pbuf_zsecond_create("swwbuf", nswbuf / 4);
611 swrbuf_zone = pbuf_zsecond_create("swrbuf", nswbuf / 2);
614 * Initialize our zone, taking the user's requested size or
615 * estimating the number we need based on the number of pages
618 n = maxswzone != 0 ? maxswzone / sizeof(struct swblk) :
619 vm_cnt.v_page_count / 2;
620 swpctrie_zone = uma_zcreate("swpctrie", pctrie_node_size(), NULL, NULL,
621 pctrie_zone_init, NULL, UMA_ALIGN_PTR, 0);
622 if (swpctrie_zone == NULL)
623 panic("failed to create swap pctrie zone.");
624 swblk_zone = uma_zcreate("swblk", sizeof(struct swblk), NULL, NULL,
625 NULL, NULL, _Alignof(struct swblk) - 1, 0);
626 if (swblk_zone == NULL)
627 panic("failed to create swap blk zone.");
630 if (uma_zone_reserve_kva(swblk_zone, n))
633 * if the allocation failed, try a zone two thirds the
634 * size of the previous attempt.
640 * Often uma_zone_reserve_kva() cannot reserve exactly the
641 * requested size. Account for the difference when
642 * calculating swap_maxpages.
644 n = uma_zone_get_max(swblk_zone);
647 printf("Swap blk zone entries changed from %lu to %lu.\n",
649 /* absolute maximum we can handle assuming 100% efficiency */
650 swap_maxpages = n * SWAP_META_PAGES;
651 swzone = n * sizeof(struct swblk);
652 if (!uma_zone_reserve_kva(swpctrie_zone, n))
653 printf("Cannot reserve swap pctrie zone, "
654 "reduce kern.maxswzone.\n");
658 swap_pager_alloc_init(void *handle, struct ucred *cred, vm_ooffset_t size,
664 if (!swap_reserve_by_cred(size, cred))
670 * The un_pager.swp.swp_blks trie is initialized by
671 * vm_object_allocate() to ensure the correct order of
672 * visibility to other threads.
674 object = vm_object_allocate(OBJT_SWAP, OFF_TO_IDX(offset +
677 object->un_pager.swp.writemappings = 0;
678 object->handle = handle;
681 object->charge = size;
687 * SWAP_PAGER_ALLOC() - allocate a new OBJT_SWAP VM object and instantiate
688 * its metadata structures.
690 * This routine is called from the mmap and fork code to create a new
693 * This routine must ensure that no live duplicate is created for
694 * the named object request, which is protected against by
695 * holding the sw_alloc_sx lock in case handle != NULL.
698 swap_pager_alloc(void *handle, vm_ooffset_t size, vm_prot_t prot,
699 vm_ooffset_t offset, struct ucred *cred)
703 if (handle != NULL) {
705 * Reference existing named region or allocate new one. There
706 * should not be a race here against swp_pager_meta_build()
707 * as called from vm_page_remove() in regards to the lookup
710 sx_xlock(&sw_alloc_sx);
711 object = vm_pager_object_lookup(NOBJLIST(handle), handle);
712 if (object == NULL) {
713 object = swap_pager_alloc_init(handle, cred, size,
715 if (object != NULL) {
716 TAILQ_INSERT_TAIL(NOBJLIST(object->handle),
717 object, pager_object_list);
720 sx_xunlock(&sw_alloc_sx);
722 object = swap_pager_alloc_init(handle, cred, size, offset);
728 * SWAP_PAGER_DEALLOC() - remove swap metadata from object
730 * The swap backing for the object is destroyed. The code is
731 * designed such that we can reinstantiate it later, but this
732 * routine is typically called only when the entire object is
733 * about to be destroyed.
735 * The object must be locked.
738 swap_pager_dealloc(vm_object_t object)
741 VM_OBJECT_ASSERT_WLOCKED(object);
742 KASSERT((object->flags & OBJ_DEAD) != 0, ("dealloc of reachable obj"));
745 * Remove from list right away so lookups will fail if we block for
746 * pageout completion.
748 if ((object->flags & OBJ_ANON) == 0 && object->handle != NULL) {
749 VM_OBJECT_WUNLOCK(object);
750 sx_xlock(&sw_alloc_sx);
751 TAILQ_REMOVE(NOBJLIST(object->handle), object,
753 sx_xunlock(&sw_alloc_sx);
754 VM_OBJECT_WLOCK(object);
757 vm_object_pip_wait(object, "swpdea");
760 * Free all remaining metadata. We only bother to free it from
761 * the swap meta data. We do not attempt to free swapblk's still
762 * associated with vm_page_t's for this object. We do not care
763 * if paging is still in progress on some objects.
765 swp_pager_meta_free_all(object);
766 object->handle = NULL;
767 object->type = OBJT_DEAD;
770 /************************************************************************
771 * SWAP PAGER BITMAP ROUTINES *
772 ************************************************************************/
775 * SWP_PAGER_GETSWAPSPACE() - allocate raw swap space
777 * Allocate swap for up to the requested number of pages. The
778 * starting swap block number (a page index) is returned or
779 * SWAPBLK_NONE if the allocation failed.
781 * Also has the side effect of advising that somebody made a mistake
782 * when they configured swap and didn't configure enough.
784 * This routine may not sleep.
786 * We allocate in round-robin fashion from the configured devices.
789 swp_pager_getswapspace(int *io_npages)
795 KASSERT(*io_npages >= 1,
796 ("%s: npages not positive", __func__));
799 npages = imin(BLIST_MAX_ALLOC, mpages);
800 mtx_lock(&sw_dev_mtx);
802 while (!TAILQ_EMPTY(&swtailq)) {
804 sp = TAILQ_FIRST(&swtailq);
805 if ((sp->sw_flags & SW_CLOSING) == 0)
806 blk = blist_alloc(sp->sw_blist, &npages, mpages);
807 if (blk != SWAPBLK_NONE)
809 sp = TAILQ_NEXT(sp, sw_list);
817 if (blk != SWAPBLK_NONE) {
820 sp->sw_used += npages;
821 swap_pager_avail -= npages;
823 swdevhd = TAILQ_NEXT(sp, sw_list);
825 if (swap_pager_full != 2) {
826 printf("swp_pager_getswapspace(%d): failed\n",
829 swap_pager_almost_full = 1;
833 mtx_unlock(&sw_dev_mtx);
838 swp_pager_isondev(daddr_t blk, struct swdevt *sp)
841 return (blk >= sp->sw_first && blk < sp->sw_end);
845 swp_pager_strategy(struct buf *bp)
849 mtx_lock(&sw_dev_mtx);
850 TAILQ_FOREACH(sp, &swtailq, sw_list) {
851 if (swp_pager_isondev(bp->b_blkno, sp)) {
852 mtx_unlock(&sw_dev_mtx);
853 if ((sp->sw_flags & SW_UNMAPPED) != 0 &&
854 unmapped_buf_allowed) {
855 bp->b_data = unmapped_buf;
858 pmap_qenter((vm_offset_t)bp->b_data,
859 &bp->b_pages[0], bp->b_bcount / PAGE_SIZE);
861 sp->sw_strategy(bp, sp);
865 panic("Swapdev not found");
870 * SWP_PAGER_FREESWAPSPACE() - free raw swap space
872 * This routine returns the specified swap blocks back to the bitmap.
874 * This routine may not sleep.
877 swp_pager_freeswapspace(daddr_t blk, daddr_t npages)
883 mtx_lock(&sw_dev_mtx);
884 TAILQ_FOREACH(sp, &swtailq, sw_list) {
885 if (swp_pager_isondev(blk, sp)) {
886 sp->sw_used -= npages;
888 * If we are attempting to stop swapping on
889 * this device, we don't want to mark any
890 * blocks free lest they be reused.
892 if ((sp->sw_flags & SW_CLOSING) == 0) {
893 blist_free(sp->sw_blist, blk - sp->sw_first,
895 swap_pager_avail += npages;
898 mtx_unlock(&sw_dev_mtx);
902 panic("Swapdev not found");
906 * SYSCTL_SWAP_FRAGMENTATION() - produce raw swap space stats
909 sysctl_swap_fragmentation(SYSCTL_HANDLER_ARGS)
916 error = sysctl_wire_old_buffer(req, 0);
919 sbuf_new_for_sysctl(&sbuf, NULL, 128, req);
920 mtx_lock(&sw_dev_mtx);
921 TAILQ_FOREACH(sp, &swtailq, sw_list) {
922 if (vn_isdisk(sp->sw_vp, NULL))
923 devname = devtoname(sp->sw_vp->v_rdev);
926 sbuf_printf(&sbuf, "\nFree space on device %s:\n", devname);
927 blist_stats(sp->sw_blist, &sbuf);
929 mtx_unlock(&sw_dev_mtx);
930 error = sbuf_finish(&sbuf);
936 * SWAP_PAGER_FREESPACE() - frees swap blocks associated with a page
937 * range within an object.
939 * This is a globally accessible routine.
941 * This routine removes swapblk assignments from swap metadata.
943 * The external callers of this routine typically have already destroyed
944 * or renamed vm_page_t's associated with this range in the object so
947 * The object must be locked.
950 swap_pager_freespace(vm_object_t object, vm_pindex_t start, vm_size_t size)
953 swp_pager_meta_free(object, start, size);
957 * SWAP_PAGER_RESERVE() - reserve swap blocks in object
959 * Assigns swap blocks to the specified range within the object. The
960 * swap blocks are not zeroed. Any previous swap assignment is destroyed.
962 * Returns 0 on success, -1 on failure.
965 swap_pager_reserve(vm_object_t object, vm_pindex_t start, vm_size_t size)
967 daddr_t addr, blk, n_free, s_free;
970 swp_pager_init_freerange(&s_free, &n_free);
971 VM_OBJECT_WLOCK(object);
972 for (i = 0; i < size; i += n) {
974 blk = swp_pager_getswapspace(&n);
975 if (blk == SWAPBLK_NONE) {
976 swp_pager_meta_free(object, start, i);
977 VM_OBJECT_WUNLOCK(object);
980 for (j = 0; j < n; ++j) {
981 addr = swp_pager_meta_build(object,
982 start + i + j, blk + j);
983 if (addr != SWAPBLK_NONE)
984 swp_pager_update_freerange(&s_free, &n_free,
988 swp_pager_freeswapspace(s_free, n_free);
989 VM_OBJECT_WUNLOCK(object);
994 swp_pager_xfer_source(vm_object_t srcobject, vm_object_t dstobject,
995 vm_pindex_t pindex, daddr_t addr)
999 KASSERT(srcobject->type == OBJT_SWAP,
1000 ("%s: Srcobject not swappable", __func__));
1001 if (dstobject->type == OBJT_SWAP &&
1002 swp_pager_meta_lookup(dstobject, pindex) != SWAPBLK_NONE) {
1003 /* Caller should destroy the source block. */
1008 * Destination has no swapblk and is not resident, transfer source.
1009 * swp_pager_meta_build() can sleep.
1011 VM_OBJECT_WUNLOCK(srcobject);
1012 dstaddr = swp_pager_meta_build(dstobject, pindex, addr);
1013 KASSERT(dstaddr == SWAPBLK_NONE,
1014 ("Unexpected destination swapblk"));
1015 VM_OBJECT_WLOCK(srcobject);
1021 * SWAP_PAGER_COPY() - copy blocks from source pager to destination pager
1022 * and destroy the source.
1024 * Copy any valid swapblks from the source to the destination. In
1025 * cases where both the source and destination have a valid swapblk,
1026 * we keep the destination's.
1028 * This routine is allowed to sleep. It may sleep allocating metadata
1029 * indirectly through swp_pager_meta_build().
1031 * The source object contains no vm_page_t's (which is just as well)
1033 * The source object is of type OBJT_SWAP.
1035 * The source and destination objects must be locked.
1036 * Both object locks may temporarily be released.
1039 swap_pager_copy(vm_object_t srcobject, vm_object_t dstobject,
1040 vm_pindex_t offset, int destroysource)
1043 VM_OBJECT_ASSERT_WLOCKED(srcobject);
1044 VM_OBJECT_ASSERT_WLOCKED(dstobject);
1047 * If destroysource is set, we remove the source object from the
1048 * swap_pager internal queue now.
1050 if (destroysource && (srcobject->flags & OBJ_ANON) == 0 &&
1051 srcobject->handle != NULL) {
1052 VM_OBJECT_WUNLOCK(srcobject);
1053 VM_OBJECT_WUNLOCK(dstobject);
1054 sx_xlock(&sw_alloc_sx);
1055 TAILQ_REMOVE(NOBJLIST(srcobject->handle), srcobject,
1057 sx_xunlock(&sw_alloc_sx);
1058 VM_OBJECT_WLOCK(dstobject);
1059 VM_OBJECT_WLOCK(srcobject);
1063 * Transfer source to destination.
1065 swp_pager_meta_transfer(srcobject, dstobject, offset, dstobject->size);
1068 * Free left over swap blocks in source.
1070 * We have to revert the type to OBJT_DEFAULT so we do not accidentally
1071 * double-remove the object from the swap queues.
1073 if (destroysource) {
1074 swp_pager_meta_free_all(srcobject);
1076 * Reverting the type is not necessary, the caller is going
1077 * to destroy srcobject directly, but I'm doing it here
1078 * for consistency since we've removed the object from its
1081 srcobject->type = OBJT_DEFAULT;
1086 * SWAP_PAGER_HASPAGE() - determine if we have good backing store for
1087 * the requested page.
1089 * We determine whether good backing store exists for the requested
1090 * page and return TRUE if it does, FALSE if it doesn't.
1092 * If TRUE, we also try to determine how much valid, contiguous backing
1093 * store exists before and after the requested page.
1096 swap_pager_haspage(vm_object_t object, vm_pindex_t pindex, int *before,
1102 VM_OBJECT_ASSERT_LOCKED(object);
1103 KASSERT(object->type == OBJT_SWAP,
1104 ("%s: object not swappable", __func__));
1107 * do we have good backing store at the requested index ?
1109 blk0 = swp_pager_meta_lookup(object, pindex);
1110 if (blk0 == SWAPBLK_NONE) {
1119 * find backwards-looking contiguous good backing store
1121 if (before != NULL) {
1122 for (i = 1; i < SWB_NPAGES; i++) {
1125 blk = swp_pager_meta_lookup(object, pindex - i);
1126 if (blk != blk0 - i)
1133 * find forward-looking contiguous good backing store
1135 if (after != NULL) {
1136 for (i = 1; i < SWB_NPAGES; i++) {
1137 blk = swp_pager_meta_lookup(object, pindex + i);
1138 if (blk != blk0 + i)
1147 * SWAP_PAGER_PAGE_UNSWAPPED() - remove swap backing store related to page
1149 * This removes any associated swap backing store, whether valid or
1150 * not, from the page.
1152 * This routine is typically called when a page is made dirty, at
1153 * which point any associated swap can be freed. MADV_FREE also
1154 * calls us in a special-case situation
1156 * NOTE!!! If the page is clean and the swap was valid, the caller
1157 * should make the page dirty before calling this routine. This routine
1158 * does NOT change the m->dirty status of the page. Also: MADV_FREE
1161 * This routine may not sleep.
1163 * The object containing the page may be locked.
1166 swap_pager_unswapped(vm_page_t m)
1172 * Handle enqueing deferred frees first. If we do not have the
1173 * object lock we wait for the page daemon to clear the space.
1176 if (!VM_OBJECT_WOWNED(obj)) {
1177 VM_PAGE_OBJECT_BUSY_ASSERT(m);
1179 * The caller is responsible for synchronization but we
1180 * will harmlessly handle races. This is typically provided
1181 * by only calling unswapped() when a page transitions from
1184 if ((m->a.flags & (PGA_SWAP_SPACE | PGA_SWAP_FREE)) ==
1186 vm_page_aflag_set(m, PGA_SWAP_FREE);
1187 counter_u64_add(swap_free_deferred, 1);
1191 if ((m->a.flags & PGA_SWAP_FREE) != 0)
1192 counter_u64_add(swap_free_completed, 1);
1193 vm_page_aflag_clear(m, PGA_SWAP_FREE | PGA_SWAP_SPACE);
1196 * The meta data only exists if the object is OBJT_SWAP
1197 * and even then might not be allocated yet.
1199 KASSERT(m->object->type == OBJT_SWAP,
1200 ("Free object not swappable"));
1202 sb = SWAP_PCTRIE_LOOKUP(&m->object->un_pager.swp.swp_blks,
1203 rounddown(m->pindex, SWAP_META_PAGES));
1206 if (sb->d[m->pindex % SWAP_META_PAGES] == SWAPBLK_NONE)
1208 swp_pager_freeswapspace(sb->d[m->pindex % SWAP_META_PAGES], 1);
1209 sb->d[m->pindex % SWAP_META_PAGES] = SWAPBLK_NONE;
1210 swp_pager_free_empty_swblk(m->object, sb);
1214 * swap_pager_getpages() - bring pages in from swap
1216 * Attempt to page in the pages in array "ma" of length "count". The
1217 * caller may optionally specify that additional pages preceding and
1218 * succeeding the specified range be paged in. The number of such pages
1219 * is returned in the "rbehind" and "rahead" parameters, and they will
1220 * be in the inactive queue upon return.
1222 * The pages in "ma" must be busied and will remain busied upon return.
1225 swap_pager_getpages_locked(vm_object_t object, vm_page_t *ma, int count,
1226 int *rbehind, int *rahead)
1229 vm_page_t bm, mpred, msucc, p;
1232 int i, maxahead, maxbehind, reqcount;
1234 VM_OBJECT_ASSERT_WLOCKED(object);
1237 KASSERT(object->type == OBJT_SWAP,
1238 ("%s: object not swappable", __func__));
1239 if (!swap_pager_haspage(object, ma[0]->pindex, &maxbehind, &maxahead)) {
1240 VM_OBJECT_WUNLOCK(object);
1241 return (VM_PAGER_FAIL);
1244 KASSERT(reqcount - 1 <= maxahead,
1245 ("page count %d extends beyond swap block", reqcount));
1248 * Do not transfer any pages other than those that are xbusied
1249 * when running during a split or collapse operation. This
1250 * prevents clustering from re-creating pages which are being
1251 * moved into another object.
1253 if ((object->flags & (OBJ_SPLIT | OBJ_DEAD)) != 0) {
1254 maxahead = reqcount - 1;
1259 * Clip the readahead and readbehind ranges to exclude resident pages.
1261 if (rahead != NULL) {
1262 *rahead = imin(*rahead, maxahead - (reqcount - 1));
1263 pindex = ma[reqcount - 1]->pindex;
1264 msucc = TAILQ_NEXT(ma[reqcount - 1], listq);
1265 if (msucc != NULL && msucc->pindex - pindex - 1 < *rahead)
1266 *rahead = msucc->pindex - pindex - 1;
1268 if (rbehind != NULL) {
1269 *rbehind = imin(*rbehind, maxbehind);
1270 pindex = ma[0]->pindex;
1271 mpred = TAILQ_PREV(ma[0], pglist, listq);
1272 if (mpred != NULL && pindex - mpred->pindex - 1 < *rbehind)
1273 *rbehind = pindex - mpred->pindex - 1;
1277 for (i = 0; i < count; i++)
1278 ma[i]->oflags |= VPO_SWAPINPROG;
1281 * Allocate readahead and readbehind pages.
1283 if (rbehind != NULL) {
1284 for (i = 1; i <= *rbehind; i++) {
1285 p = vm_page_alloc(object, ma[0]->pindex - i,
1289 p->oflags |= VPO_SWAPINPROG;
1294 if (rahead != NULL) {
1295 for (i = 0; i < *rahead; i++) {
1296 p = vm_page_alloc(object,
1297 ma[reqcount - 1]->pindex + i + 1, VM_ALLOC_NORMAL);
1300 p->oflags |= VPO_SWAPINPROG;
1304 if (rbehind != NULL)
1309 vm_object_pip_add(object, count);
1311 pindex = bm->pindex;
1312 blk = swp_pager_meta_lookup(object, pindex);
1313 KASSERT(blk != SWAPBLK_NONE,
1314 ("no swap blocking containing %p(%jx)", object, (uintmax_t)pindex));
1316 VM_OBJECT_WUNLOCK(object);
1317 bp = uma_zalloc(swrbuf_zone, M_WAITOK);
1318 /* Pages cannot leave the object while busy. */
1319 for (i = 0, p = bm; i < count; i++, p = TAILQ_NEXT(p, listq)) {
1320 MPASS(p->pindex == bm->pindex + i);
1324 bp->b_flags |= B_PAGING;
1325 bp->b_iocmd = BIO_READ;
1326 bp->b_iodone = swp_pager_async_iodone;
1327 bp->b_rcred = crhold(thread0.td_ucred);
1328 bp->b_wcred = crhold(thread0.td_ucred);
1330 bp->b_bcount = PAGE_SIZE * count;
1331 bp->b_bufsize = PAGE_SIZE * count;
1332 bp->b_npages = count;
1333 bp->b_pgbefore = rbehind != NULL ? *rbehind : 0;
1334 bp->b_pgafter = rahead != NULL ? *rahead : 0;
1336 VM_CNT_INC(v_swapin);
1337 VM_CNT_ADD(v_swappgsin, count);
1340 * perform the I/O. NOTE!!! bp cannot be considered valid after
1341 * this point because we automatically release it on completion.
1342 * Instead, we look at the one page we are interested in which we
1343 * still hold a lock on even through the I/O completion.
1345 * The other pages in our ma[] array are also released on completion,
1346 * so we cannot assume they are valid anymore either.
1348 * NOTE: b_blkno is destroyed by the call to swapdev_strategy
1351 swp_pager_strategy(bp);
1354 * Wait for the pages we want to complete. VPO_SWAPINPROG is always
1355 * cleared on completion. If an I/O error occurs, SWAPBLK_NONE
1356 * is set in the metadata for each page in the request.
1358 VM_OBJECT_WLOCK(object);
1359 /* This could be implemented more efficiently with aflags */
1360 while ((ma[0]->oflags & VPO_SWAPINPROG) != 0) {
1361 ma[0]->oflags |= VPO_SWAPSLEEP;
1362 VM_CNT_INC(v_intrans);
1363 if (VM_OBJECT_SLEEP(object, &object->handle, PSWP,
1364 "swread", hz * 20)) {
1366 "swap_pager: indefinite wait buffer: bufobj: %p, blkno: %jd, size: %ld\n",
1367 bp->b_bufobj, (intmax_t)bp->b_blkno, bp->b_bcount);
1370 VM_OBJECT_WUNLOCK(object);
1373 * If we had an unrecoverable read error pages will not be valid.
1375 for (i = 0; i < reqcount; i++)
1376 if (ma[i]->valid != VM_PAGE_BITS_ALL)
1377 return (VM_PAGER_ERROR);
1379 return (VM_PAGER_OK);
1382 * A final note: in a low swap situation, we cannot deallocate swap
1383 * and mark a page dirty here because the caller is likely to mark
1384 * the page clean when we return, causing the page to possibly revert
1385 * to all-zero's later.
1390 swap_pager_getpages(vm_object_t object, vm_page_t *ma, int count,
1391 int *rbehind, int *rahead)
1394 VM_OBJECT_WLOCK(object);
1395 return (swap_pager_getpages_locked(object, ma, count, rbehind, rahead));
1399 * swap_pager_getpages_async():
1401 * Right now this is emulation of asynchronous operation on top of
1402 * swap_pager_getpages().
1405 swap_pager_getpages_async(vm_object_t object, vm_page_t *ma, int count,
1406 int *rbehind, int *rahead, pgo_getpages_iodone_t iodone, void *arg)
1410 r = swap_pager_getpages(object, ma, count, rbehind, rahead);
1415 case VM_PAGER_ERROR:
1422 panic("unhandled swap_pager_getpages() error %d", r);
1424 (iodone)(arg, ma, count, error);
1430 * swap_pager_putpages:
1432 * Assign swap (if necessary) and initiate I/O on the specified pages.
1434 * We support both OBJT_DEFAULT and OBJT_SWAP objects. DEFAULT objects
1435 * are automatically converted to SWAP objects.
1437 * In a low memory situation we may block in VOP_STRATEGY(), but the new
1438 * vm_page reservation system coupled with properly written VFS devices
1439 * should ensure that no low-memory deadlock occurs. This is an area
1442 * The parent has N vm_object_pip_add() references prior to
1443 * calling us and will remove references for rtvals[] that are
1444 * not set to VM_PAGER_PEND. We need to remove the rest on I/O
1447 * The parent has soft-busy'd the pages it passes us and will unbusy
1448 * those whose rtvals[] entry is not set to VM_PAGER_PEND on return.
1449 * We need to unbusy the rest on I/O completion.
1452 swap_pager_putpages(vm_object_t object, vm_page_t *ma, int count,
1453 int flags, int *rtvals)
1456 daddr_t addr, blk, n_free, s_free;
1461 KASSERT(count == 0 || ma[0]->object == object,
1462 ("%s: object mismatch %p/%p",
1463 __func__, object, ma[0]->object));
1468 * Turn object into OBJT_SWAP. Force sync if not a pageout process.
1470 if (object->type != OBJT_SWAP) {
1471 addr = swp_pager_meta_build(object, 0, SWAPBLK_NONE);
1472 KASSERT(addr == SWAPBLK_NONE,
1473 ("unexpected object swap block"));
1475 VM_OBJECT_WUNLOCK(object);
1476 async = curproc == pageproc && (flags & VM_PAGER_PUT_SYNC) == 0;
1477 swp_pager_init_freerange(&s_free, &n_free);
1482 * Assign swap blocks and issue I/O. We reallocate swap on the fly.
1483 * The page is left dirty until the pageout operation completes
1486 for (i = 0; i < count; i += n) {
1487 /* Maximum I/O size is limited by maximum swap block size. */
1488 n = min(count - i, nsw_cluster_max);
1491 mtx_lock(&swbuf_mtx);
1492 while (nsw_wcount_async == 0)
1493 msleep(&nsw_wcount_async, &swbuf_mtx, PVM,
1496 mtx_unlock(&swbuf_mtx);
1499 /* Get a block of swap of size up to size n. */
1500 VM_OBJECT_WLOCK(object);
1501 blk = swp_pager_getswapspace(&n);
1502 if (blk == SWAPBLK_NONE) {
1503 VM_OBJECT_WUNLOCK(object);
1504 mtx_lock(&swbuf_mtx);
1505 if (++nsw_wcount_async == 1)
1506 wakeup(&nsw_wcount_async);
1507 mtx_unlock(&swbuf_mtx);
1508 for (j = 0; j < n; ++j)
1509 rtvals[i + j] = VM_PAGER_FAIL;
1512 for (j = 0; j < n; ++j) {
1514 vm_page_aflag_clear(mreq, PGA_SWAP_FREE);
1515 addr = swp_pager_meta_build(mreq->object, mreq->pindex,
1517 if (addr != SWAPBLK_NONE)
1518 swp_pager_update_freerange(&s_free, &n_free,
1520 MPASS(mreq->dirty == VM_PAGE_BITS_ALL);
1521 mreq->oflags |= VPO_SWAPINPROG;
1523 VM_OBJECT_WUNLOCK(object);
1525 bp = uma_zalloc(swwbuf_zone, M_WAITOK);
1527 bp->b_flags = B_ASYNC;
1528 bp->b_flags |= B_PAGING;
1529 bp->b_iocmd = BIO_WRITE;
1531 bp->b_rcred = crhold(thread0.td_ucred);
1532 bp->b_wcred = crhold(thread0.td_ucred);
1533 bp->b_bcount = PAGE_SIZE * n;
1534 bp->b_bufsize = PAGE_SIZE * n;
1536 for (j = 0; j < n; j++)
1537 bp->b_pages[j] = ma[i + j];
1541 * Must set dirty range for NFS to work.
1544 bp->b_dirtyend = bp->b_bcount;
1546 VM_CNT_INC(v_swapout);
1547 VM_CNT_ADD(v_swappgsout, bp->b_npages);
1550 * We unconditionally set rtvals[] to VM_PAGER_PEND so that we
1551 * can call the async completion routine at the end of a
1552 * synchronous I/O operation. Otherwise, our caller would
1553 * perform duplicate unbusy and wakeup operations on the page
1554 * and object, respectively.
1556 for (j = 0; j < n; j++)
1557 rtvals[i + j] = VM_PAGER_PEND;
1562 * NOTE: b_blkno is destroyed by the call to swapdev_strategy.
1565 bp->b_iodone = swp_pager_async_iodone;
1567 swp_pager_strategy(bp);
1574 * NOTE: b_blkno is destroyed by the call to swapdev_strategy.
1576 bp->b_iodone = bdone;
1577 swp_pager_strategy(bp);
1580 * Wait for the sync I/O to complete.
1582 bwait(bp, PVM, "swwrt");
1585 * Now that we are through with the bp, we can call the
1586 * normal async completion, which frees everything up.
1588 swp_pager_async_iodone(bp);
1590 swp_pager_freeswapspace(s_free, n_free);
1591 VM_OBJECT_WLOCK(object);
1595 * swp_pager_async_iodone:
1597 * Completion routine for asynchronous reads and writes from/to swap.
1598 * Also called manually by synchronous code to finish up a bp.
1600 * This routine may not sleep.
1603 swp_pager_async_iodone(struct buf *bp)
1606 vm_object_t object = NULL;
1609 * Report error - unless we ran out of memory, in which case
1610 * we've already logged it in swapgeom_strategy().
1612 if (bp->b_ioflags & BIO_ERROR && bp->b_error != ENOMEM) {
1614 "swap_pager: I/O error - %s failed; blkno %ld,"
1615 "size %ld, error %d\n",
1616 ((bp->b_iocmd == BIO_READ) ? "pagein" : "pageout"),
1624 * remove the mapping for kernel virtual
1627 pmap_qremove((vm_offset_t)bp->b_data, bp->b_npages);
1629 bp->b_data = bp->b_kvabase;
1632 object = bp->b_pages[0]->object;
1633 VM_OBJECT_WLOCK(object);
1637 * cleanup pages. If an error occurs writing to swap, we are in
1638 * very serious trouble. If it happens to be a disk error, though,
1639 * we may be able to recover by reassigning the swap later on. So
1640 * in this case we remove the m->swapblk assignment for the page
1641 * but do not free it in the rlist. The errornous block(s) are thus
1642 * never reallocated as swap. Redirty the page and continue.
1644 for (i = 0; i < bp->b_npages; ++i) {
1645 vm_page_t m = bp->b_pages[i];
1647 m->oflags &= ~VPO_SWAPINPROG;
1648 if (m->oflags & VPO_SWAPSLEEP) {
1649 m->oflags &= ~VPO_SWAPSLEEP;
1650 wakeup(&object->handle);
1653 /* We always have space after I/O, successful or not. */
1654 vm_page_aflag_set(m, PGA_SWAP_SPACE);
1656 if (bp->b_ioflags & BIO_ERROR) {
1658 * If an error occurs I'd love to throw the swapblk
1659 * away without freeing it back to swapspace, so it
1660 * can never be used again. But I can't from an
1663 if (bp->b_iocmd == BIO_READ) {
1665 * NOTE: for reads, m->dirty will probably
1666 * be overridden by the original caller of
1667 * getpages so don't play cute tricks here.
1672 * If a write error occurs, reactivate page
1673 * so it doesn't clog the inactive list,
1674 * then finish the I/O.
1676 MPASS(m->dirty == VM_PAGE_BITS_ALL);
1678 /* PQ_UNSWAPPABLE? */
1679 vm_page_activate(m);
1682 } else if (bp->b_iocmd == BIO_READ) {
1684 * NOTE: for reads, m->dirty will probably be
1685 * overridden by the original caller of getpages so
1686 * we cannot set them in order to free the underlying
1687 * swap in a low-swap situation. I don't think we'd
1688 * want to do that anyway, but it was an optimization
1689 * that existed in the old swapper for a time before
1690 * it got ripped out due to precisely this problem.
1692 KASSERT(!pmap_page_is_mapped(m),
1693 ("swp_pager_async_iodone: page %p is mapped", m));
1694 KASSERT(m->dirty == 0,
1695 ("swp_pager_async_iodone: page %p is dirty", m));
1698 if (i < bp->b_pgbefore ||
1699 i >= bp->b_npages - bp->b_pgafter)
1700 vm_page_readahead_finish(m);
1703 * For write success, clear the dirty
1704 * status, then finish the I/O ( which decrements the
1705 * busy count and possibly wakes waiter's up ).
1706 * A page is only written to swap after a period of
1707 * inactivity. Therefore, we do not expect it to be
1710 KASSERT(!pmap_page_is_write_mapped(m),
1711 ("swp_pager_async_iodone: page %p is not write"
1714 vm_page_deactivate_noreuse(m);
1720 * adjust pip. NOTE: the original parent may still have its own
1721 * pip refs on the object.
1723 if (object != NULL) {
1724 vm_object_pip_wakeupn(object, bp->b_npages);
1725 VM_OBJECT_WUNLOCK(object);
1729 * swapdev_strategy() manually sets b_vp and b_bufobj before calling
1730 * bstrategy(). Set them back to NULL now we're done with it, or we'll
1731 * trigger a KASSERT in relpbuf().
1735 bp->b_bufobj = NULL;
1738 * release the physical I/O buffer
1740 if (bp->b_flags & B_ASYNC) {
1741 mtx_lock(&swbuf_mtx);
1742 if (++nsw_wcount_async == 1)
1743 wakeup(&nsw_wcount_async);
1744 mtx_unlock(&swbuf_mtx);
1746 uma_zfree((bp->b_iocmd == BIO_READ) ? swrbuf_zone : swwbuf_zone, bp);
1750 swap_pager_nswapdev(void)
1757 swp_pager_force_dirty(vm_page_t m)
1761 swap_pager_unswapped(m);
1766 * swap_pager_swapoff_object:
1768 * Page in all of the pages that have been paged out for an object
1772 swap_pager_swapoff_object(struct swdevt *sp, vm_object_t object)
1778 int i, nv, rahead, rv;
1780 KASSERT(object->type == OBJT_SWAP,
1781 ("%s: Object not swappable", __func__));
1783 for (pi = 0; (sb = SWAP_PCTRIE_LOOKUP_GE(
1784 &object->un_pager.swp.swp_blks, pi)) != NULL; ) {
1785 if ((object->flags & OBJ_DEAD) != 0) {
1787 * Make sure that pending writes finish before
1790 vm_object_pip_wait(object, "swpoff");
1791 swp_pager_meta_free_all(object);
1794 for (i = 0; i < SWAP_META_PAGES; i++) {
1796 * Count the number of contiguous valid blocks.
1798 for (nv = 0; nv < SWAP_META_PAGES - i; nv++) {
1799 blk = sb->d[i + nv];
1800 if (!swp_pager_isondev(blk, sp) ||
1801 blk == SWAPBLK_NONE)
1808 * Look for a page corresponding to the first
1809 * valid block and ensure that any pending paging
1810 * operations on it are complete. If the page is valid,
1811 * mark it dirty and free the swap block. Try to batch
1812 * this operation since it may cause sp to be freed,
1813 * meaning that we must restart the scan. Avoid busying
1814 * valid pages since we may block forever on kernel
1817 m = vm_page_lookup(object, sb->p + i);
1819 m = vm_page_alloc(object, sb->p + i,
1820 VM_ALLOC_NORMAL | VM_ALLOC_WAITFAIL);
1824 if ((m->oflags & VPO_SWAPINPROG) != 0) {
1825 m->oflags |= VPO_SWAPSLEEP;
1826 VM_OBJECT_SLEEP(object, &object->handle,
1830 if (vm_page_all_valid(m)) {
1832 swp_pager_force_dirty(m);
1833 } while (--nv > 0 &&
1834 (m = vm_page_next(m)) != NULL &&
1835 vm_page_all_valid(m) &&
1836 (m->oflags & VPO_SWAPINPROG) == 0);
1839 if (!vm_page_busy_acquire(m, VM_ALLOC_WAITFAIL))
1843 vm_object_pip_add(object, 1);
1844 rahead = SWAP_META_PAGES;
1845 rv = swap_pager_getpages_locked(object, &m, 1, NULL,
1847 if (rv != VM_PAGER_OK)
1848 panic("%s: read from swap failed: %d",
1850 vm_object_pip_wakeupn(object, 1);
1851 VM_OBJECT_WLOCK(object);
1855 * The object lock was dropped so we must restart the
1856 * scan of this swap block. Pages paged in during this
1857 * iteration will be marked dirty in a future iteration.
1861 if (i == SWAP_META_PAGES)
1862 pi = sb->p + SWAP_META_PAGES;
1867 * swap_pager_swapoff:
1869 * Page in all of the pages that have been paged out to the
1870 * given device. The corresponding blocks in the bitmap must be
1871 * marked as allocated and the device must be flagged SW_CLOSING.
1872 * There may be no processes swapped out to the device.
1874 * This routine may block.
1877 swap_pager_swapoff(struct swdevt *sp)
1882 sx_assert(&swdev_syscall_lock, SA_XLOCKED);
1886 mtx_lock(&vm_object_list_mtx);
1887 TAILQ_FOREACH(object, &vm_object_list, object_list) {
1888 if (object->type != OBJT_SWAP)
1890 mtx_unlock(&vm_object_list_mtx);
1891 /* Depends on type-stability. */
1892 VM_OBJECT_WLOCK(object);
1895 * Dead objects are eventually terminated on their own.
1897 if ((object->flags & OBJ_DEAD) != 0)
1901 * Sync with fences placed after pctrie
1902 * initialization. We must not access pctrie below
1903 * unless we checked that our object is swap and not
1906 atomic_thread_fence_acq();
1907 if (object->type != OBJT_SWAP)
1910 swap_pager_swapoff_object(sp, object);
1912 VM_OBJECT_WUNLOCK(object);
1913 mtx_lock(&vm_object_list_mtx);
1915 mtx_unlock(&vm_object_list_mtx);
1919 * Objects may be locked or paging to the device being
1920 * removed, so we will miss their pages and need to
1921 * make another pass. We have marked this device as
1922 * SW_CLOSING, so the activity should finish soon.
1925 if (retries > 100) {
1926 panic("swapoff: failed to locate %d swap blocks",
1929 pause("swpoff", hz / 20);
1932 EVENTHANDLER_INVOKE(swapoff, sp);
1935 /************************************************************************
1937 ************************************************************************
1939 * These routines manipulate the swap metadata stored in the
1942 * Swap metadata is implemented with a global hash and not directly
1943 * linked into the object. Instead the object simply contains
1944 * appropriate tracking counters.
1948 * SWP_PAGER_SWBLK_EMPTY() - is a range of blocks free?
1951 swp_pager_swblk_empty(struct swblk *sb, int start, int limit)
1955 MPASS(0 <= start && start <= limit && limit <= SWAP_META_PAGES);
1956 for (i = start; i < limit; i++) {
1957 if (sb->d[i] != SWAPBLK_NONE)
1964 * SWP_PAGER_FREE_EMPTY_SWBLK() - frees if a block is free
1966 * Nothing is done if the block is still in use.
1969 swp_pager_free_empty_swblk(vm_object_t object, struct swblk *sb)
1972 if (swp_pager_swblk_empty(sb, 0, SWAP_META_PAGES)) {
1973 SWAP_PCTRIE_REMOVE(&object->un_pager.swp.swp_blks, sb->p);
1974 uma_zfree(swblk_zone, sb);
1979 * SWP_PAGER_META_BUILD() - add swap block to swap meta data for object
1981 * We first convert the object to a swap object if it is a default
1984 * The specified swapblk is added to the object's swap metadata. If
1985 * the swapblk is not valid, it is freed instead. Any previously
1986 * assigned swapblk is returned.
1989 swp_pager_meta_build(vm_object_t object, vm_pindex_t pindex, daddr_t swapblk)
1991 static volatile int swblk_zone_exhausted, swpctrie_zone_exhausted;
1992 struct swblk *sb, *sb1;
1993 vm_pindex_t modpi, rdpi;
1994 daddr_t prev_swapblk;
1997 VM_OBJECT_ASSERT_WLOCKED(object);
2000 * Convert default object to swap object if necessary
2002 if (object->type != OBJT_SWAP) {
2003 pctrie_init(&object->un_pager.swp.swp_blks);
2006 * Ensure that swap_pager_swapoff()'s iteration over
2007 * object_list does not see a garbage pctrie.
2009 atomic_thread_fence_rel();
2011 object->type = OBJT_SWAP;
2012 object->un_pager.swp.writemappings = 0;
2013 KASSERT((object->flags & OBJ_ANON) != 0 ||
2014 object->handle == NULL,
2015 ("default pager %p with handle %p",
2016 object, object->handle));
2019 rdpi = rounddown(pindex, SWAP_META_PAGES);
2020 sb = SWAP_PCTRIE_LOOKUP(&object->un_pager.swp.swp_blks, rdpi);
2022 if (swapblk == SWAPBLK_NONE)
2023 return (SWAPBLK_NONE);
2025 sb = uma_zalloc(swblk_zone, M_NOWAIT | (curproc ==
2026 pageproc ? M_USE_RESERVE : 0));
2029 for (i = 0; i < SWAP_META_PAGES; i++)
2030 sb->d[i] = SWAPBLK_NONE;
2031 if (atomic_cmpset_int(&swblk_zone_exhausted,
2033 printf("swblk zone ok\n");
2036 VM_OBJECT_WUNLOCK(object);
2037 if (uma_zone_exhausted(swblk_zone)) {
2038 if (atomic_cmpset_int(&swblk_zone_exhausted,
2040 printf("swap blk zone exhausted, "
2041 "increase kern.maxswzone\n");
2042 vm_pageout_oom(VM_OOM_SWAPZ);
2043 pause("swzonxb", 10);
2045 uma_zwait(swblk_zone);
2046 VM_OBJECT_WLOCK(object);
2047 sb = SWAP_PCTRIE_LOOKUP(&object->un_pager.swp.swp_blks,
2051 * Somebody swapped out a nearby page,
2052 * allocating swblk at the rdpi index,
2053 * while we dropped the object lock.
2058 error = SWAP_PCTRIE_INSERT(
2059 &object->un_pager.swp.swp_blks, sb);
2061 if (atomic_cmpset_int(&swpctrie_zone_exhausted,
2063 printf("swpctrie zone ok\n");
2066 VM_OBJECT_WUNLOCK(object);
2067 if (uma_zone_exhausted(swpctrie_zone)) {
2068 if (atomic_cmpset_int(&swpctrie_zone_exhausted,
2070 printf("swap pctrie zone exhausted, "
2071 "increase kern.maxswzone\n");
2072 vm_pageout_oom(VM_OOM_SWAPZ);
2073 pause("swzonxp", 10);
2075 uma_zwait(swpctrie_zone);
2076 VM_OBJECT_WLOCK(object);
2077 sb1 = SWAP_PCTRIE_LOOKUP(&object->un_pager.swp.swp_blks,
2080 uma_zfree(swblk_zone, sb);
2087 MPASS(sb->p == rdpi);
2089 modpi = pindex % SWAP_META_PAGES;
2090 /* Return prior contents of metadata. */
2091 prev_swapblk = sb->d[modpi];
2092 /* Enter block into metadata. */
2093 sb->d[modpi] = swapblk;
2096 * Free the swblk if we end up with the empty page run.
2098 if (swapblk == SWAPBLK_NONE)
2099 swp_pager_free_empty_swblk(object, sb);
2100 return (prev_swapblk);
2104 * SWP_PAGER_META_TRANSFER() - free a range of blocks in the srcobject's swap
2105 * metadata, or transfer it into dstobject.
2107 * This routine will free swap metadata structures as they are cleaned
2111 swp_pager_meta_transfer(vm_object_t srcobject, vm_object_t dstobject,
2112 vm_pindex_t pindex, vm_pindex_t count)
2115 daddr_t n_free, s_free;
2116 vm_pindex_t offset, last;
2117 int i, limit, start;
2119 VM_OBJECT_ASSERT_WLOCKED(srcobject);
2120 if (srcobject->type != OBJT_SWAP || count == 0)
2123 swp_pager_init_freerange(&s_free, &n_free);
2125 last = pindex + count;
2127 sb = SWAP_PCTRIE_LOOKUP_GE(&srcobject->un_pager.swp.swp_blks,
2128 rounddown(pindex, SWAP_META_PAGES));
2129 if (sb == NULL || sb->p >= last)
2131 start = pindex > sb->p ? pindex - sb->p : 0;
2132 limit = last - sb->p < SWAP_META_PAGES ? last - sb->p :
2134 for (i = start; i < limit; i++) {
2135 if (sb->d[i] == SWAPBLK_NONE)
2137 if (dstobject == NULL ||
2138 !swp_pager_xfer_source(srcobject, dstobject,
2139 sb->p + i - offset, sb->d[i])) {
2140 swp_pager_update_freerange(&s_free, &n_free,
2143 sb->d[i] = SWAPBLK_NONE;
2145 pindex = sb->p + SWAP_META_PAGES;
2146 if (swp_pager_swblk_empty(sb, 0, start) &&
2147 swp_pager_swblk_empty(sb, limit, SWAP_META_PAGES)) {
2148 SWAP_PCTRIE_REMOVE(&srcobject->un_pager.swp.swp_blks,
2150 uma_zfree(swblk_zone, sb);
2153 swp_pager_freeswapspace(s_free, n_free);
2157 * SWP_PAGER_META_FREE() - free a range of blocks in the object's swap metadata
2159 * The requested range of blocks is freed, with any associated swap
2160 * returned to the swap bitmap.
2162 * This routine will free swap metadata structures as they are cleaned
2163 * out. This routine does *NOT* operate on swap metadata associated
2164 * with resident pages.
2167 swp_pager_meta_free(vm_object_t object, vm_pindex_t pindex, vm_pindex_t count)
2169 swp_pager_meta_transfer(object, NULL, pindex, count);
2173 * SWP_PAGER_META_FREE_ALL() - destroy all swap metadata associated with object
2175 * This routine locates and destroys all swap metadata associated with
2179 swp_pager_meta_free_all(vm_object_t object)
2182 daddr_t n_free, s_free;
2186 VM_OBJECT_ASSERT_WLOCKED(object);
2187 if (object->type != OBJT_SWAP)
2190 swp_pager_init_freerange(&s_free, &n_free);
2191 for (pindex = 0; (sb = SWAP_PCTRIE_LOOKUP_GE(
2192 &object->un_pager.swp.swp_blks, pindex)) != NULL;) {
2193 pindex = sb->p + SWAP_META_PAGES;
2194 for (i = 0; i < SWAP_META_PAGES; i++) {
2195 if (sb->d[i] == SWAPBLK_NONE)
2197 swp_pager_update_freerange(&s_free, &n_free, sb->d[i]);
2199 SWAP_PCTRIE_REMOVE(&object->un_pager.swp.swp_blks, sb->p);
2200 uma_zfree(swblk_zone, sb);
2202 swp_pager_freeswapspace(s_free, n_free);
2206 * SWP_PAGER_METACTL() - misc control of swap meta data.
2208 * This routine is capable of looking up, or removing swapblk
2209 * assignments in the swap meta data. It returns the swapblk being
2210 * looked-up, popped, or SWAPBLK_NONE if the block was invalid.
2212 * When acting on a busy resident page and paging is in progress, we
2213 * have to wait until paging is complete but otherwise can act on the
2217 swp_pager_meta_lookup(vm_object_t object, vm_pindex_t pindex)
2221 VM_OBJECT_ASSERT_LOCKED(object);
2224 * The meta data only exists if the object is OBJT_SWAP
2225 * and even then might not be allocated yet.
2227 KASSERT(object->type == OBJT_SWAP,
2228 ("Lookup object not swappable"));
2230 sb = SWAP_PCTRIE_LOOKUP(&object->un_pager.swp.swp_blks,
2231 rounddown(pindex, SWAP_META_PAGES));
2233 return (SWAPBLK_NONE);
2234 return (sb->d[pindex % SWAP_META_PAGES]);
2238 * Returns the least page index which is greater than or equal to the
2239 * parameter pindex and for which there is a swap block allocated.
2240 * Returns object's size if the object's type is not swap or if there
2241 * are no allocated swap blocks for the object after the requested
2245 swap_pager_find_least(vm_object_t object, vm_pindex_t pindex)
2250 VM_OBJECT_ASSERT_LOCKED(object);
2251 if (object->type != OBJT_SWAP)
2252 return (object->size);
2254 sb = SWAP_PCTRIE_LOOKUP_GE(&object->un_pager.swp.swp_blks,
2255 rounddown(pindex, SWAP_META_PAGES));
2257 return (object->size);
2258 if (sb->p < pindex) {
2259 for (i = pindex % SWAP_META_PAGES; i < SWAP_META_PAGES; i++) {
2260 if (sb->d[i] != SWAPBLK_NONE)
2263 sb = SWAP_PCTRIE_LOOKUP_GE(&object->un_pager.swp.swp_blks,
2264 roundup(pindex, SWAP_META_PAGES));
2266 return (object->size);
2268 for (i = 0; i < SWAP_META_PAGES; i++) {
2269 if (sb->d[i] != SWAPBLK_NONE)
2274 * We get here if a swblk is present in the trie but it
2275 * doesn't map any blocks.
2278 return (object->size);
2282 * System call swapon(name) enables swapping on device name,
2283 * which must be in the swdevsw. Return EBUSY
2284 * if already swapping on this device.
2286 #ifndef _SYS_SYSPROTO_H_
2287 struct swapon_args {
2297 sys_swapon(struct thread *td, struct swapon_args *uap)
2301 struct nameidata nd;
2304 error = priv_check(td, PRIV_SWAPON);
2308 sx_xlock(&swdev_syscall_lock);
2311 * Swap metadata may not fit in the KVM if we have physical
2314 if (swblk_zone == NULL) {
2319 NDINIT(&nd, LOOKUP, ISOPEN | FOLLOW | AUDITVNODE1, UIO_USERSPACE,
2325 NDFREE(&nd, NDF_ONLY_PNBUF);
2328 if (vn_isdisk(vp, &error)) {
2329 error = swapongeom(vp);
2330 } else if (vp->v_type == VREG &&
2331 (vp->v_mount->mnt_vfc->vfc_flags & VFCF_NETWORK) != 0 &&
2332 (error = VOP_GETATTR(vp, &attr, td->td_ucred)) == 0) {
2334 * Allow direct swapping to NFS regular files in the same
2335 * way that nfs_mountroot() sets up diskless swapping.
2337 error = swaponvp(td, vp, attr.va_size / DEV_BSIZE);
2343 sx_xunlock(&swdev_syscall_lock);
2348 * Check that the total amount of swap currently configured does not
2349 * exceed half the theoretical maximum. If it does, print a warning
2353 swapon_check_swzone(void)
2356 /* recommend using no more than half that amount */
2357 if (swap_total > swap_maxpages / 2) {
2358 printf("warning: total configured swap (%lu pages) "
2359 "exceeds maximum recommended amount (%lu pages).\n",
2360 swap_total, swap_maxpages / 2);
2361 printf("warning: increase kern.maxswzone "
2362 "or reduce amount of swap.\n");
2367 swaponsomething(struct vnode *vp, void *id, u_long nblks,
2368 sw_strategy_t *strategy, sw_close_t *close, dev_t dev, int flags)
2370 struct swdevt *sp, *tsp;
2374 * nblks is in DEV_BSIZE'd chunks, convert to PAGE_SIZE'd chunks.
2375 * First chop nblks off to page-align it, then convert.
2377 * sw->sw_nblks is in page-sized chunks now too.
2379 nblks &= ~(ctodb(1) - 1);
2380 nblks = dbtoc(nblks);
2382 sp = malloc(sizeof *sp, M_VMPGDATA, M_WAITOK | M_ZERO);
2383 sp->sw_blist = blist_create(nblks, M_WAITOK);
2387 sp->sw_nblks = nblks;
2389 sp->sw_strategy = strategy;
2390 sp->sw_close = close;
2391 sp->sw_flags = flags;
2394 * Do not free the first blocks in order to avoid overwriting
2395 * any bsd label at the front of the partition
2397 blist_free(sp->sw_blist, howmany(BBSIZE, PAGE_SIZE),
2398 nblks - howmany(BBSIZE, PAGE_SIZE));
2401 mtx_lock(&sw_dev_mtx);
2402 TAILQ_FOREACH(tsp, &swtailq, sw_list) {
2403 if (tsp->sw_end >= dvbase) {
2405 * We put one uncovered page between the devices
2406 * in order to definitively prevent any cross-device
2409 dvbase = tsp->sw_end + 1;
2412 sp->sw_first = dvbase;
2413 sp->sw_end = dvbase + nblks;
2414 TAILQ_INSERT_TAIL(&swtailq, sp, sw_list);
2416 swap_pager_avail += nblks - howmany(BBSIZE, PAGE_SIZE);
2417 swap_total += nblks;
2418 swapon_check_swzone();
2420 mtx_unlock(&sw_dev_mtx);
2421 EVENTHANDLER_INVOKE(swapon, sp);
2425 * SYSCALL: swapoff(devname)
2427 * Disable swapping on the given device.
2429 * XXX: Badly designed system call: it should use a device index
2430 * rather than filename as specification. We keep sw_vp around
2431 * only to make this work.
2433 #ifndef _SYS_SYSPROTO_H_
2434 struct swapoff_args {
2444 sys_swapoff(struct thread *td, struct swapoff_args *uap)
2447 struct nameidata nd;
2451 error = priv_check(td, PRIV_SWAPOFF);
2455 sx_xlock(&swdev_syscall_lock);
2457 NDINIT(&nd, LOOKUP, FOLLOW | AUDITVNODE1, UIO_USERSPACE, uap->name,
2462 NDFREE(&nd, NDF_ONLY_PNBUF);
2465 mtx_lock(&sw_dev_mtx);
2466 TAILQ_FOREACH(sp, &swtailq, sw_list) {
2467 if (sp->sw_vp == vp)
2470 mtx_unlock(&sw_dev_mtx);
2475 error = swapoff_one(sp, td->td_ucred);
2477 sx_xunlock(&swdev_syscall_lock);
2482 swapoff_one(struct swdevt *sp, struct ucred *cred)
2489 sx_assert(&swdev_syscall_lock, SA_XLOCKED);
2491 (void) vn_lock(sp->sw_vp, LK_EXCLUSIVE | LK_RETRY);
2492 error = mac_system_check_swapoff(cred, sp->sw_vp);
2493 (void) VOP_UNLOCK(sp->sw_vp);
2497 nblks = sp->sw_nblks;
2500 * We can turn off this swap device safely only if the
2501 * available virtual memory in the system will fit the amount
2502 * of data we will have to page back in, plus an epsilon so
2503 * the system doesn't become critically low on swap space.
2505 if (vm_free_count() + swap_pager_avail < nblks + nswap_lowat)
2509 * Prevent further allocations on this device.
2511 mtx_lock(&sw_dev_mtx);
2512 sp->sw_flags |= SW_CLOSING;
2513 swap_pager_avail -= blist_fill(sp->sw_blist, 0, nblks);
2514 swap_total -= nblks;
2515 mtx_unlock(&sw_dev_mtx);
2518 * Page in the contents of the device and close it.
2520 swap_pager_swapoff(sp);
2522 sp->sw_close(curthread, sp);
2523 mtx_lock(&sw_dev_mtx);
2525 TAILQ_REMOVE(&swtailq, sp, sw_list);
2527 if (nswapdev == 0) {
2528 swap_pager_full = 2;
2529 swap_pager_almost_full = 1;
2533 mtx_unlock(&sw_dev_mtx);
2534 blist_destroy(sp->sw_blist);
2535 free(sp, M_VMPGDATA);
2542 struct swdevt *sp, *spt;
2543 const char *devname;
2546 sx_xlock(&swdev_syscall_lock);
2548 mtx_lock(&sw_dev_mtx);
2549 TAILQ_FOREACH_SAFE(sp, &swtailq, sw_list, spt) {
2550 mtx_unlock(&sw_dev_mtx);
2551 if (vn_isdisk(sp->sw_vp, NULL))
2552 devname = devtoname(sp->sw_vp->v_rdev);
2555 error = swapoff_one(sp, thread0.td_ucred);
2557 printf("Cannot remove swap device %s (error=%d), "
2558 "skipping.\n", devname, error);
2559 } else if (bootverbose) {
2560 printf("Swap device %s removed.\n", devname);
2562 mtx_lock(&sw_dev_mtx);
2564 mtx_unlock(&sw_dev_mtx);
2566 sx_xunlock(&swdev_syscall_lock);
2570 swap_pager_status(int *total, int *used)
2573 *total = swap_total;
2574 *used = swap_total - swap_pager_avail -
2575 nswapdev * howmany(BBSIZE, PAGE_SIZE);
2579 swap_dev_info(int name, struct xswdev *xs, char *devname, size_t len)
2582 const char *tmp_devname;
2587 mtx_lock(&sw_dev_mtx);
2588 TAILQ_FOREACH(sp, &swtailq, sw_list) {
2593 xs->xsw_version = XSWDEV_VERSION;
2594 xs->xsw_dev = sp->sw_dev;
2595 xs->xsw_flags = sp->sw_flags;
2596 xs->xsw_nblks = sp->sw_nblks;
2597 xs->xsw_used = sp->sw_used;
2598 if (devname != NULL) {
2599 if (vn_isdisk(sp->sw_vp, NULL))
2600 tmp_devname = devtoname(sp->sw_vp->v_rdev);
2602 tmp_devname = "[file]";
2603 strncpy(devname, tmp_devname, len);
2608 mtx_unlock(&sw_dev_mtx);
2612 #if defined(COMPAT_FREEBSD11)
2613 #define XSWDEV_VERSION_11 1
2623 #if defined(__amd64__) && defined(COMPAT_FREEBSD32)
2626 u_int xsw_dev1, xsw_dev2;
2634 sysctl_vm_swap_info(SYSCTL_HANDLER_ARGS)
2637 #if defined(__amd64__) && defined(COMPAT_FREEBSD32)
2638 struct xswdev32 xs32;
2640 #if defined(COMPAT_FREEBSD11)
2641 struct xswdev11 xs11;
2645 if (arg2 != 1) /* name length */
2647 error = swap_dev_info(*(int *)arg1, &xs, NULL, 0);
2650 #if defined(__amd64__) && defined(COMPAT_FREEBSD32)
2651 if (req->oldlen == sizeof(xs32)) {
2652 xs32.xsw_version = XSWDEV_VERSION;
2653 xs32.xsw_dev1 = xs.xsw_dev;
2654 xs32.xsw_dev2 = xs.xsw_dev >> 32;
2655 xs32.xsw_flags = xs.xsw_flags;
2656 xs32.xsw_nblks = xs.xsw_nblks;
2657 xs32.xsw_used = xs.xsw_used;
2658 error = SYSCTL_OUT(req, &xs32, sizeof(xs32));
2662 #if defined(COMPAT_FREEBSD11)
2663 if (req->oldlen == sizeof(xs11)) {
2664 xs11.xsw_version = XSWDEV_VERSION_11;
2665 xs11.xsw_dev = xs.xsw_dev; /* truncation */
2666 xs11.xsw_flags = xs.xsw_flags;
2667 xs11.xsw_nblks = xs.xsw_nblks;
2668 xs11.xsw_used = xs.xsw_used;
2669 error = SYSCTL_OUT(req, &xs11, sizeof(xs11));
2673 error = SYSCTL_OUT(req, &xs, sizeof(xs));
2677 SYSCTL_INT(_vm, OID_AUTO, nswapdev, CTLFLAG_RD, &nswapdev, 0,
2678 "Number of swap devices");
2679 SYSCTL_NODE(_vm, OID_AUTO, swap_info, CTLFLAG_RD | CTLFLAG_MPSAFE,
2680 sysctl_vm_swap_info,
2681 "Swap statistics by device");
2684 * Count the approximate swap usage in pages for a vmspace. The
2685 * shadowed or not yet copied on write swap blocks are not accounted.
2686 * The map must be locked.
2689 vmspace_swap_count(struct vmspace *vmspace)
2699 map = &vmspace->vm_map;
2702 VM_MAP_ENTRY_FOREACH(cur, map) {
2703 if ((cur->eflags & MAP_ENTRY_IS_SUB_MAP) != 0)
2705 object = cur->object.vm_object;
2706 if (object == NULL || object->type != OBJT_SWAP)
2708 VM_OBJECT_RLOCK(object);
2709 if (object->type != OBJT_SWAP)
2711 pi = OFF_TO_IDX(cur->offset);
2712 e = pi + OFF_TO_IDX(cur->end - cur->start);
2713 for (;; pi = sb->p + SWAP_META_PAGES) {
2714 sb = SWAP_PCTRIE_LOOKUP_GE(
2715 &object->un_pager.swp.swp_blks, pi);
2716 if (sb == NULL || sb->p >= e)
2718 for (i = 0; i < SWAP_META_PAGES; i++) {
2719 if (sb->p + i < e &&
2720 sb->d[i] != SWAPBLK_NONE)
2725 VM_OBJECT_RUNLOCK(object);
2733 * Swapping onto disk devices.
2737 static g_orphan_t swapgeom_orphan;
2739 static struct g_class g_swap_class = {
2741 .version = G_VERSION,
2742 .orphan = swapgeom_orphan,
2745 DECLARE_GEOM_CLASS(g_swap_class, g_class);
2749 swapgeom_close_ev(void *arg, int flags)
2751 struct g_consumer *cp;
2754 g_access(cp, -1, -1, 0);
2756 g_destroy_consumer(cp);
2760 * Add a reference to the g_consumer for an inflight transaction.
2763 swapgeom_acquire(struct g_consumer *cp)
2766 mtx_assert(&sw_dev_mtx, MA_OWNED);
2771 * Remove a reference from the g_consumer. Post a close event if all
2772 * references go away, since the function might be called from the
2776 swapgeom_release(struct g_consumer *cp, struct swdevt *sp)
2779 mtx_assert(&sw_dev_mtx, MA_OWNED);
2781 if (cp->index == 0) {
2782 if (g_post_event(swapgeom_close_ev, cp, M_NOWAIT, NULL) == 0)
2788 swapgeom_done(struct bio *bp2)
2792 struct g_consumer *cp;
2794 bp = bp2->bio_caller2;
2796 bp->b_ioflags = bp2->bio_flags;
2798 bp->b_ioflags |= BIO_ERROR;
2799 bp->b_resid = bp->b_bcount - bp2->bio_completed;
2800 bp->b_error = bp2->bio_error;
2801 bp->b_caller1 = NULL;
2803 sp = bp2->bio_caller1;
2804 mtx_lock(&sw_dev_mtx);
2805 swapgeom_release(cp, sp);
2806 mtx_unlock(&sw_dev_mtx);
2811 swapgeom_strategy(struct buf *bp, struct swdevt *sp)
2814 struct g_consumer *cp;
2816 mtx_lock(&sw_dev_mtx);
2819 mtx_unlock(&sw_dev_mtx);
2820 bp->b_error = ENXIO;
2821 bp->b_ioflags |= BIO_ERROR;
2825 swapgeom_acquire(cp);
2826 mtx_unlock(&sw_dev_mtx);
2827 if (bp->b_iocmd == BIO_WRITE)
2830 bio = g_alloc_bio();
2832 mtx_lock(&sw_dev_mtx);
2833 swapgeom_release(cp, sp);
2834 mtx_unlock(&sw_dev_mtx);
2835 bp->b_error = ENOMEM;
2836 bp->b_ioflags |= BIO_ERROR;
2837 printf("swap_pager: cannot allocate bio\n");
2842 bp->b_caller1 = bio;
2843 bio->bio_caller1 = sp;
2844 bio->bio_caller2 = bp;
2845 bio->bio_cmd = bp->b_iocmd;
2846 bio->bio_offset = (bp->b_blkno - sp->sw_first) * PAGE_SIZE;
2847 bio->bio_length = bp->b_bcount;
2848 bio->bio_done = swapgeom_done;
2849 if (!buf_mapped(bp)) {
2850 bio->bio_ma = bp->b_pages;
2851 bio->bio_data = unmapped_buf;
2852 bio->bio_ma_offset = (vm_offset_t)bp->b_offset & PAGE_MASK;
2853 bio->bio_ma_n = bp->b_npages;
2854 bio->bio_flags |= BIO_UNMAPPED;
2856 bio->bio_data = bp->b_data;
2859 g_io_request(bio, cp);
2864 swapgeom_orphan(struct g_consumer *cp)
2869 mtx_lock(&sw_dev_mtx);
2870 TAILQ_FOREACH(sp, &swtailq, sw_list) {
2871 if (sp->sw_id == cp) {
2872 sp->sw_flags |= SW_CLOSING;
2877 * Drop reference we were created with. Do directly since we're in a
2878 * special context where we don't have to queue the call to
2879 * swapgeom_close_ev().
2882 destroy = ((sp != NULL) && (cp->index == 0));
2885 mtx_unlock(&sw_dev_mtx);
2887 swapgeom_close_ev(cp, 0);
2891 swapgeom_close(struct thread *td, struct swdevt *sw)
2893 struct g_consumer *cp;
2895 mtx_lock(&sw_dev_mtx);
2898 mtx_unlock(&sw_dev_mtx);
2901 * swapgeom_close() may be called from the biodone context,
2902 * where we cannot perform topology changes. Delegate the
2903 * work to the events thread.
2906 g_waitfor_event(swapgeom_close_ev, cp, M_WAITOK, NULL);
2910 swapongeom_locked(struct cdev *dev, struct vnode *vp)
2912 struct g_provider *pp;
2913 struct g_consumer *cp;
2914 static struct g_geom *gp;
2919 pp = g_dev_getprovider(dev);
2922 mtx_lock(&sw_dev_mtx);
2923 TAILQ_FOREACH(sp, &swtailq, sw_list) {
2925 if (cp != NULL && cp->provider == pp) {
2926 mtx_unlock(&sw_dev_mtx);
2930 mtx_unlock(&sw_dev_mtx);
2932 gp = g_new_geomf(&g_swap_class, "swap");
2933 cp = g_new_consumer(gp);
2934 cp->index = 1; /* Number of active I/Os, plus one for being active. */
2935 cp->flags |= G_CF_DIRECT_SEND | G_CF_DIRECT_RECEIVE;
2938 * XXX: Every time you think you can improve the margin for
2939 * footshooting, somebody depends on the ability to do so:
2940 * savecore(8) wants to write to our swapdev so we cannot
2941 * set an exclusive count :-(
2943 error = g_access(cp, 1, 1, 0);
2946 g_destroy_consumer(cp);
2949 nblks = pp->mediasize / DEV_BSIZE;
2950 swaponsomething(vp, cp, nblks, swapgeom_strategy,
2951 swapgeom_close, dev2udev(dev),
2952 (pp->flags & G_PF_ACCEPT_UNMAPPED) != 0 ? SW_UNMAPPED : 0);
2957 swapongeom(struct vnode *vp)
2961 vn_lock(vp, LK_EXCLUSIVE | LK_RETRY);
2962 if (vp->v_type != VCHR || VN_IS_DOOMED(vp)) {
2966 error = swapongeom_locked(vp->v_rdev, vp);
2967 g_topology_unlock();
2976 * This is used mainly for network filesystem (read: probably only tested
2977 * with NFS) swapfiles.
2982 swapdev_strategy(struct buf *bp, struct swdevt *sp)
2986 bp->b_blkno = ctodb(bp->b_blkno - sp->sw_first);
2990 if (bp->b_iocmd == BIO_WRITE) {
2992 bufobj_wdrop(bp->b_bufobj);
2993 bufobj_wref(&vp2->v_bufobj);
2995 if (bp->b_bufobj != &vp2->v_bufobj)
2996 bp->b_bufobj = &vp2->v_bufobj;
2998 bp->b_iooffset = dbtob(bp->b_blkno);
3004 swapdev_close(struct thread *td, struct swdevt *sp)
3007 VOP_CLOSE(sp->sw_vp, FREAD | FWRITE, td->td_ucred, td);
3013 swaponvp(struct thread *td, struct vnode *vp, u_long nblks)
3020 mtx_lock(&sw_dev_mtx);
3021 TAILQ_FOREACH(sp, &swtailq, sw_list) {
3022 if (sp->sw_id == vp) {
3023 mtx_unlock(&sw_dev_mtx);
3027 mtx_unlock(&sw_dev_mtx);
3029 (void) vn_lock(vp, LK_EXCLUSIVE | LK_RETRY);
3031 error = mac_system_check_swapon(td->td_ucred, vp);
3034 error = VOP_OPEN(vp, FREAD | FWRITE, td->td_ucred, td, NULL);
3035 (void) VOP_UNLOCK(vp);
3039 swaponsomething(vp, vp, nblks, swapdev_strategy, swapdev_close,
3045 sysctl_swap_async_max(SYSCTL_HANDLER_ARGS)
3049 new = nsw_wcount_async_max;
3050 error = sysctl_handle_int(oidp, &new, 0, req);
3051 if (error != 0 || req->newptr == NULL)
3054 if (new > nswbuf / 2 || new < 1)
3057 mtx_lock(&swbuf_mtx);
3058 while (nsw_wcount_async_max != new) {
3060 * Adjust difference. If the current async count is too low,
3061 * we will need to sqeeze our update slowly in. Sleep with a
3062 * higher priority than getpbuf() to finish faster.
3064 n = new - nsw_wcount_async_max;
3065 if (nsw_wcount_async + n >= 0) {
3066 nsw_wcount_async += n;
3067 nsw_wcount_async_max += n;
3068 wakeup(&nsw_wcount_async);
3070 nsw_wcount_async_max -= nsw_wcount_async;
3071 nsw_wcount_async = 0;
3072 msleep(&nsw_wcount_async, &swbuf_mtx, PSWP,
3076 mtx_unlock(&swbuf_mtx);
3082 swap_pager_update_writecount(vm_object_t object, vm_offset_t start,
3086 VM_OBJECT_WLOCK(object);
3087 KASSERT((object->flags & OBJ_ANON) == 0,
3088 ("Splittable object with writecount"));
3089 object->un_pager.swp.writemappings += (vm_ooffset_t)end - start;
3090 VM_OBJECT_WUNLOCK(object);
3094 swap_pager_release_writecount(vm_object_t object, vm_offset_t start,
3098 VM_OBJECT_WLOCK(object);
3099 KASSERT((object->flags & OBJ_ANON) == 0,
3100 ("Splittable object with writecount"));
3101 object->un_pager.swp.writemappings -= (vm_ooffset_t)end - start;
3102 VM_OBJECT_WUNLOCK(object);