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 u_long swap_reserved;
156 static u_long swap_total;
157 static int sysctl_page_shift(SYSCTL_HANDLER_ARGS);
158 SYSCTL_PROC(_vm, OID_AUTO, swap_reserved, CTLTYPE_U64 | CTLFLAG_RD | CTLFLAG_MPSAFE,
159 &swap_reserved, 0, sysctl_page_shift, "A",
160 "Amount of swap storage needed to back all allocated anonymous memory.");
161 SYSCTL_PROC(_vm, OID_AUTO, swap_total, CTLTYPE_U64 | CTLFLAG_RD | CTLFLAG_MPSAFE,
162 &swap_total, 0, sysctl_page_shift, "A",
163 "Total amount of available swap storage.");
165 static int overcommit = 0;
166 SYSCTL_INT(_vm, VM_OVERCOMMIT, overcommit, CTLFLAG_RW, &overcommit, 0,
167 "Configure virtual memory overcommit behavior. See tuning(7) "
169 static unsigned long swzone;
170 SYSCTL_ULONG(_vm, OID_AUTO, swzone, CTLFLAG_RD, &swzone, 0,
171 "Actual size of swap metadata zone");
172 static unsigned long swap_maxpages;
173 SYSCTL_ULONG(_vm, OID_AUTO, swap_maxpages, CTLFLAG_RD, &swap_maxpages, 0,
174 "Maximum amount of swap supported");
176 /* bits from overcommit */
177 #define SWAP_RESERVE_FORCE_ON (1 << 0)
178 #define SWAP_RESERVE_RLIMIT_ON (1 << 1)
179 #define SWAP_RESERVE_ALLOW_NONWIRED (1 << 2)
182 sysctl_page_shift(SYSCTL_HANDLER_ARGS)
185 u_long value = *(u_long *)arg1;
187 newval = ((uint64_t)value) << PAGE_SHIFT;
188 return (sysctl_handle_64(oidp, &newval, 0, req));
192 swap_reserve(vm_ooffset_t incr)
195 return (swap_reserve_by_cred(incr, curthread->td_ucred));
199 swap_reserve_by_cred(vm_ooffset_t incr, struct ucred *cred)
201 u_long r, s, prev, pincr;
204 static struct timeval lastfail;
207 uip = cred->cr_ruidinfo;
209 KASSERT((incr & PAGE_MASK) == 0, ("%s: incr: %ju & PAGE_MASK", __func__,
215 error = racct_add(curproc, RACCT_SWAP, incr);
216 PROC_UNLOCK(curproc);
224 prev = atomic_fetchadd_long(&swap_reserved, pincr);
226 if (overcommit & SWAP_RESERVE_ALLOW_NONWIRED) {
227 s = vm_cnt.v_page_count - vm_cnt.v_free_reserved -
232 if ((overcommit & SWAP_RESERVE_FORCE_ON) == 0 || r <= s ||
233 (error = priv_check(curthread, PRIV_VM_SWAP_NOQUOTA)) == 0) {
236 prev = atomic_fetchadd_long(&swap_reserved, -pincr);
238 panic("swap_reserved < incr on overcommit fail");
241 prev = atomic_fetchadd_long(&uip->ui_vmsize, pincr);
242 if ((overcommit & SWAP_RESERVE_RLIMIT_ON) != 0 &&
243 prev + pincr > lim_cur(curthread, RLIMIT_SWAP) &&
244 priv_check(curthread, PRIV_VM_SWAP_NORLIMIT)) {
246 prev = atomic_fetchadd_long(&uip->ui_vmsize, -pincr);
248 panic("uip->ui_vmsize < incr on overcommit fail");
251 if (!res && ppsratecheck(&lastfail, &curfail, 1)) {
252 printf("uid %d, pid %d: swap reservation for %jd bytes failed\n",
253 uip->ui_uid, curproc->p_pid, incr);
257 if (racct_enable && !res) {
259 racct_sub(curproc, RACCT_SWAP, incr);
260 PROC_UNLOCK(curproc);
268 swap_reserve_force(vm_ooffset_t incr)
273 KASSERT((incr & PAGE_MASK) == 0, ("%s: incr: %ju & PAGE_MASK", __func__,
279 racct_add_force(curproc, RACCT_SWAP, incr);
282 atomic_add_long(&swap_reserved, pincr);
283 uip = curproc->p_ucred->cr_ruidinfo;
284 atomic_add_long(&uip->ui_vmsize, pincr);
285 PROC_UNLOCK(curproc);
289 swap_release(vm_ooffset_t decr)
294 cred = curproc->p_ucred;
295 swap_release_by_cred(decr, cred);
296 PROC_UNLOCK(curproc);
300 swap_release_by_cred(vm_ooffset_t decr, struct ucred *cred)
305 uip = cred->cr_ruidinfo;
307 KASSERT((decr & PAGE_MASK) == 0, ("%s: decr: %ju & PAGE_MASK", __func__,
311 prev = atomic_fetchadd_long(&swap_reserved, -pdecr);
313 panic("swap_reserved < decr");
315 prev = atomic_fetchadd_long(&uip->ui_vmsize, -pdecr);
317 printf("negative vmsize for uid = %d\n", uip->ui_uid);
320 racct_sub_cred(cred, RACCT_SWAP, decr);
324 #define SWM_POP 0x01 /* pop out */
326 static int swap_pager_full = 2; /* swap space exhaustion (task killing) */
327 static int swap_pager_almost_full = 1; /* swap space exhaustion (w/hysteresis)*/
328 static struct mtx swbuf_mtx; /* to sync nsw_wcount_async */
329 static int nsw_wcount_async; /* limit async write buffers */
330 static int nsw_wcount_async_max;/* assigned maximum */
331 static int nsw_cluster_max; /* maximum VOP I/O allowed */
333 static int sysctl_swap_async_max(SYSCTL_HANDLER_ARGS);
334 SYSCTL_PROC(_vm, OID_AUTO, swap_async_max, CTLTYPE_INT | CTLFLAG_RW |
335 CTLFLAG_MPSAFE, NULL, 0, sysctl_swap_async_max, "I",
336 "Maximum running async swap ops");
337 static int sysctl_swap_fragmentation(SYSCTL_HANDLER_ARGS);
338 SYSCTL_PROC(_vm, OID_AUTO, swap_fragmentation, CTLTYPE_STRING | CTLFLAG_RD |
339 CTLFLAG_MPSAFE, NULL, 0, sysctl_swap_fragmentation, "A",
340 "Swap Fragmentation Info");
342 static struct sx sw_alloc_sx;
345 * "named" and "unnamed" anon region objects. Try to reduce the overhead
346 * of searching a named list by hashing it just a little.
351 #define NOBJLIST(handle) \
352 (&swap_pager_object_list[((int)(intptr_t)handle >> 4) & (NOBJLISTS-1)])
354 static struct pagerlst swap_pager_object_list[NOBJLISTS];
355 static uma_zone_t swwbuf_zone;
356 static uma_zone_t swrbuf_zone;
357 static uma_zone_t swblk_zone;
358 static uma_zone_t swpctrie_zone;
361 * pagerops for OBJT_SWAP - "swap pager". Some ops are also global procedure
362 * calls hooked from other parts of the VM system and do not appear here.
363 * (see vm/swap_pager.h).
366 swap_pager_alloc(void *handle, vm_ooffset_t size,
367 vm_prot_t prot, vm_ooffset_t offset, struct ucred *);
368 static void swap_pager_dealloc(vm_object_t object);
369 static int swap_pager_getpages(vm_object_t, vm_page_t *, int, int *,
371 static int swap_pager_getpages_async(vm_object_t, vm_page_t *, int, int *,
372 int *, pgo_getpages_iodone_t, void *);
373 static void swap_pager_putpages(vm_object_t, vm_page_t *, int, boolean_t, int *);
375 swap_pager_haspage(vm_object_t object, vm_pindex_t pindex, int *before, int *after);
376 static void swap_pager_init(void);
377 static void swap_pager_unswapped(vm_page_t);
378 static void swap_pager_swapoff(struct swdevt *sp);
380 struct pagerops swappagerops = {
381 .pgo_init = swap_pager_init, /* early system initialization of pager */
382 .pgo_alloc = swap_pager_alloc, /* allocate an OBJT_SWAP object */
383 .pgo_dealloc = swap_pager_dealloc, /* deallocate an OBJT_SWAP object */
384 .pgo_getpages = swap_pager_getpages, /* pagein */
385 .pgo_getpages_async = swap_pager_getpages_async, /* pagein (async) */
386 .pgo_putpages = swap_pager_putpages, /* pageout */
387 .pgo_haspage = swap_pager_haspage, /* get backing store status for page */
388 .pgo_pageunswapped = swap_pager_unswapped, /* remove swap related to page */
392 * swap_*() routines are externally accessible. swp_*() routines are
395 static int nswap_lowat = 128; /* in pages, swap_pager_almost_full warn */
396 static int nswap_hiwat = 512; /* in pages, swap_pager_almost_full warn */
398 SYSCTL_INT(_vm, OID_AUTO, dmmax, CTLFLAG_RD, &nsw_cluster_max, 0,
399 "Maximum size of a swap block in pages");
401 static void swp_sizecheck(void);
402 static void swp_pager_async_iodone(struct buf *bp);
403 static bool swp_pager_swblk_empty(struct swblk *sb, int start, int limit);
404 static int swapongeom(struct vnode *);
405 static int swaponvp(struct thread *, struct vnode *, u_long);
406 static int swapoff_one(struct swdevt *sp, struct ucred *cred);
409 * Swap bitmap functions
411 static void swp_pager_freeswapspace(daddr_t blk, daddr_t npages);
412 static daddr_t swp_pager_getswapspace(int *npages, int limit);
417 static daddr_t swp_pager_meta_build(vm_object_t, vm_pindex_t, daddr_t);
418 static void swp_pager_meta_free(vm_object_t, vm_pindex_t, vm_pindex_t);
419 static void swp_pager_meta_free_all(vm_object_t);
420 static daddr_t swp_pager_meta_ctl(vm_object_t, vm_pindex_t, int);
423 swp_pager_init_freerange(daddr_t *start, daddr_t *num)
426 *start = SWAPBLK_NONE;
431 swp_pager_update_freerange(daddr_t *start, daddr_t *num, daddr_t addr)
434 if (*start + *num == addr) {
437 swp_pager_freeswapspace(*start, *num);
444 swblk_trie_alloc(struct pctrie *ptree)
447 return (uma_zalloc(swpctrie_zone, M_NOWAIT | (curproc == pageproc ?
448 M_USE_RESERVE : 0)));
452 swblk_trie_free(struct pctrie *ptree, void *node)
455 uma_zfree(swpctrie_zone, node);
458 PCTRIE_DEFINE(SWAP, swblk, p, swblk_trie_alloc, swblk_trie_free);
461 * SWP_SIZECHECK() - update swap_pager_full indication
463 * update the swap_pager_almost_full indication and warn when we are
464 * about to run out of swap space, using lowat/hiwat hysteresis.
466 * Clear swap_pager_full ( task killing ) indication when lowat is met.
468 * No restrictions on call
469 * This routine may not block.
475 if (swap_pager_avail < nswap_lowat) {
476 if (swap_pager_almost_full == 0) {
477 printf("swap_pager: out of swap space\n");
478 swap_pager_almost_full = 1;
482 if (swap_pager_avail > nswap_hiwat)
483 swap_pager_almost_full = 0;
488 * SWAP_PAGER_INIT() - initialize the swap pager!
490 * Expected to be started from system init. NOTE: This code is run
491 * before much else so be careful what you depend on. Most of the VM
492 * system has yet to be initialized at this point.
495 swap_pager_init(void)
498 * Initialize object lists
502 for (i = 0; i < NOBJLISTS; ++i)
503 TAILQ_INIT(&swap_pager_object_list[i]);
504 mtx_init(&sw_dev_mtx, "swapdev", NULL, MTX_DEF);
505 sx_init(&sw_alloc_sx, "swspsx");
506 sx_init(&swdev_syscall_lock, "swsysc");
510 * SWAP_PAGER_SWAP_INIT() - swap pager initialization from pageout process
512 * Expected to be started from pageout process once, prior to entering
516 swap_pager_swap_init(void)
521 * Number of in-transit swap bp operations. Don't
522 * exhaust the pbufs completely. Make sure we
523 * initialize workable values (0 will work for hysteresis
524 * but it isn't very efficient).
526 * The nsw_cluster_max is constrained by the bp->b_pages[]
527 * array, which has MAXPHYS / PAGE_SIZE entries, and our locally
528 * defined MAX_PAGEOUT_CLUSTER. Also be aware that swap ops are
529 * constrained by the swap device interleave stripe size.
531 * Currently we hardwire nsw_wcount_async to 4. This limit is
532 * designed to prevent other I/O from having high latencies due to
533 * our pageout I/O. The value 4 works well for one or two active swap
534 * devices but is probably a little low if you have more. Even so,
535 * a higher value would probably generate only a limited improvement
536 * with three or four active swap devices since the system does not
537 * typically have to pageout at extreme bandwidths. We will want
538 * at least 2 per swap devices, and 4 is a pretty good value if you
539 * have one NFS swap device due to the command/ack latency over NFS.
540 * So it all works out pretty well.
542 nsw_cluster_max = min(MAXPHYS / PAGE_SIZE, MAX_PAGEOUT_CLUSTER);
544 nsw_wcount_async = 4;
545 nsw_wcount_async_max = nsw_wcount_async;
546 mtx_init(&swbuf_mtx, "async swbuf mutex", NULL, MTX_DEF);
548 swwbuf_zone = pbuf_zsecond_create("swwbuf", nswbuf / 4);
549 swrbuf_zone = pbuf_zsecond_create("swrbuf", nswbuf / 2);
552 * Initialize our zone, taking the user's requested size or
553 * estimating the number we need based on the number of pages
556 n = maxswzone != 0 ? maxswzone / sizeof(struct swblk) :
557 vm_cnt.v_page_count / 2;
558 swpctrie_zone = uma_zcreate("swpctrie", pctrie_node_size(), NULL, NULL,
559 pctrie_zone_init, NULL, UMA_ALIGN_PTR, UMA_ZONE_VM);
560 if (swpctrie_zone == NULL)
561 panic("failed to create swap pctrie zone.");
562 swblk_zone = uma_zcreate("swblk", sizeof(struct swblk), NULL, NULL,
563 NULL, NULL, _Alignof(struct swblk) - 1, UMA_ZONE_VM);
564 if (swblk_zone == NULL)
565 panic("failed to create swap blk zone.");
568 if (uma_zone_reserve_kva(swblk_zone, n))
571 * if the allocation failed, try a zone two thirds the
572 * size of the previous attempt.
578 * Often uma_zone_reserve_kva() cannot reserve exactly the
579 * requested size. Account for the difference when
580 * calculating swap_maxpages.
582 n = uma_zone_get_max(swblk_zone);
585 printf("Swap blk zone entries changed from %lu to %lu.\n",
587 swap_maxpages = n * SWAP_META_PAGES;
588 swzone = n * sizeof(struct swblk);
589 if (!uma_zone_reserve_kva(swpctrie_zone, n))
590 printf("Cannot reserve swap pctrie zone, "
591 "reduce kern.maxswzone.\n");
595 swap_pager_alloc_init(void *handle, struct ucred *cred, vm_ooffset_t size,
601 if (!swap_reserve_by_cred(size, cred))
607 * The un_pager.swp.swp_blks trie is initialized by
608 * vm_object_allocate() to ensure the correct order of
609 * visibility to other threads.
611 object = vm_object_allocate(OBJT_SWAP, OFF_TO_IDX(offset +
614 object->handle = handle;
617 object->charge = size;
623 * SWAP_PAGER_ALLOC() - allocate a new OBJT_SWAP VM object and instantiate
624 * its metadata structures.
626 * This routine is called from the mmap and fork code to create a new
629 * This routine must ensure that no live duplicate is created for
630 * the named object request, which is protected against by
631 * holding the sw_alloc_sx lock in case handle != NULL.
634 swap_pager_alloc(void *handle, vm_ooffset_t size, vm_prot_t prot,
635 vm_ooffset_t offset, struct ucred *cred)
639 if (handle != NULL) {
641 * Reference existing named region or allocate new one. There
642 * should not be a race here against swp_pager_meta_build()
643 * as called from vm_page_remove() in regards to the lookup
646 sx_xlock(&sw_alloc_sx);
647 object = vm_pager_object_lookup(NOBJLIST(handle), handle);
648 if (object == NULL) {
649 object = swap_pager_alloc_init(handle, cred, size,
651 if (object != NULL) {
652 TAILQ_INSERT_TAIL(NOBJLIST(object->handle),
653 object, pager_object_list);
656 sx_xunlock(&sw_alloc_sx);
658 object = swap_pager_alloc_init(handle, cred, size, offset);
664 * SWAP_PAGER_DEALLOC() - remove swap metadata from object
666 * The swap backing for the object is destroyed. The code is
667 * designed such that we can reinstantiate it later, but this
668 * routine is typically called only when the entire object is
669 * about to be destroyed.
671 * The object must be locked.
674 swap_pager_dealloc(vm_object_t object)
677 VM_OBJECT_ASSERT_WLOCKED(object);
678 KASSERT((object->flags & OBJ_DEAD) != 0, ("dealloc of reachable obj"));
681 * Remove from list right away so lookups will fail if we block for
682 * pageout completion.
684 if (object->handle != NULL) {
685 VM_OBJECT_WUNLOCK(object);
686 sx_xlock(&sw_alloc_sx);
687 TAILQ_REMOVE(NOBJLIST(object->handle), object,
689 sx_xunlock(&sw_alloc_sx);
690 VM_OBJECT_WLOCK(object);
693 vm_object_pip_wait(object, "swpdea");
696 * Free all remaining metadata. We only bother to free it from
697 * the swap meta data. We do not attempt to free swapblk's still
698 * associated with vm_page_t's for this object. We do not care
699 * if paging is still in progress on some objects.
701 swp_pager_meta_free_all(object);
702 object->handle = NULL;
703 object->type = OBJT_DEAD;
706 /************************************************************************
707 * SWAP PAGER BITMAP ROUTINES *
708 ************************************************************************/
711 * SWP_PAGER_GETSWAPSPACE() - allocate raw swap space
713 * Allocate swap for up to the requested number of pages, and at
714 * least a minimum number of pages. The starting swap block number
715 * (a page index) is returned or SWAPBLK_NONE if the allocation
718 * Also has the side effect of advising that somebody made a mistake
719 * when they configured swap and didn't configure enough.
721 * This routine may not sleep.
723 * We allocate in round-robin fashion from the configured devices.
726 swp_pager_getswapspace(int *io_npages, int limit)
734 npages = imin(BLIST_MAX_ALLOC, mpages);
735 mtx_lock(&sw_dev_mtx);
737 while (!TAILQ_EMPTY(&swtailq)) {
739 sp = TAILQ_FIRST(&swtailq);
740 if ((sp->sw_flags & SW_CLOSING) == 0)
741 blk = blist_alloc(sp->sw_blist, &npages, mpages);
742 if (blk != SWAPBLK_NONE)
744 sp = TAILQ_NEXT(sp, sw_list);
752 if (blk != SWAPBLK_NONE) {
755 sp->sw_used += npages;
756 swap_pager_avail -= npages;
758 swdevhd = TAILQ_NEXT(sp, sw_list);
760 if (swap_pager_full != 2) {
761 printf("swp_pager_getswapspace(%d): failed\n",
764 swap_pager_almost_full = 1;
768 mtx_unlock(&sw_dev_mtx);
773 swp_pager_isondev(daddr_t blk, struct swdevt *sp)
776 return (blk >= sp->sw_first && blk < sp->sw_end);
780 swp_pager_strategy(struct buf *bp)
784 mtx_lock(&sw_dev_mtx);
785 TAILQ_FOREACH(sp, &swtailq, sw_list) {
786 if (swp_pager_isondev(bp->b_blkno, sp)) {
787 mtx_unlock(&sw_dev_mtx);
788 if ((sp->sw_flags & SW_UNMAPPED) != 0 &&
789 unmapped_buf_allowed) {
790 bp->b_data = unmapped_buf;
793 pmap_qenter((vm_offset_t)bp->b_data,
794 &bp->b_pages[0], bp->b_bcount / PAGE_SIZE);
796 sp->sw_strategy(bp, sp);
800 panic("Swapdev not found");
805 * SWP_PAGER_FREESWAPSPACE() - free raw swap space
807 * This routine returns the specified swap blocks back to the bitmap.
809 * This routine may not sleep.
812 swp_pager_freeswapspace(daddr_t blk, daddr_t npages)
818 mtx_lock(&sw_dev_mtx);
819 TAILQ_FOREACH(sp, &swtailq, sw_list) {
820 if (swp_pager_isondev(blk, sp)) {
821 sp->sw_used -= npages;
823 * If we are attempting to stop swapping on
824 * this device, we don't want to mark any
825 * blocks free lest they be reused.
827 if ((sp->sw_flags & SW_CLOSING) == 0) {
828 blist_free(sp->sw_blist, blk - sp->sw_first,
830 swap_pager_avail += npages;
833 mtx_unlock(&sw_dev_mtx);
837 panic("Swapdev not found");
841 * SYSCTL_SWAP_FRAGMENTATION() - produce raw swap space stats
844 sysctl_swap_fragmentation(SYSCTL_HANDLER_ARGS)
851 error = sysctl_wire_old_buffer(req, 0);
854 sbuf_new_for_sysctl(&sbuf, NULL, 128, req);
855 mtx_lock(&sw_dev_mtx);
856 TAILQ_FOREACH(sp, &swtailq, sw_list) {
857 if (vn_isdisk(sp->sw_vp, NULL))
858 devname = devtoname(sp->sw_vp->v_rdev);
861 sbuf_printf(&sbuf, "\nFree space on device %s:\n", devname);
862 blist_stats(sp->sw_blist, &sbuf);
864 mtx_unlock(&sw_dev_mtx);
865 error = sbuf_finish(&sbuf);
871 * SWAP_PAGER_FREESPACE() - frees swap blocks associated with a page
872 * range within an object.
874 * This is a globally accessible routine.
876 * This routine removes swapblk assignments from swap metadata.
878 * The external callers of this routine typically have already destroyed
879 * or renamed vm_page_t's associated with this range in the object so
882 * The object must be locked.
885 swap_pager_freespace(vm_object_t object, vm_pindex_t start, vm_size_t size)
888 swp_pager_meta_free(object, start, size);
892 * SWAP_PAGER_RESERVE() - reserve swap blocks in object
894 * Assigns swap blocks to the specified range within the object. The
895 * swap blocks are not zeroed. Any previous swap assignment is destroyed.
897 * Returns 0 on success, -1 on failure.
900 swap_pager_reserve(vm_object_t object, vm_pindex_t start, vm_size_t size)
902 daddr_t addr, blk, n_free, s_free;
905 swp_pager_init_freerange(&s_free, &n_free);
906 VM_OBJECT_WLOCK(object);
907 for (i = 0; i < size; i += n) {
909 blk = swp_pager_getswapspace(&n, 1);
910 if (blk == SWAPBLK_NONE) {
911 swp_pager_meta_free(object, start, i);
912 VM_OBJECT_WUNLOCK(object);
915 for (j = 0; j < n; ++j) {
916 addr = swp_pager_meta_build(object,
917 start + i + j, blk + j);
918 if (addr != SWAPBLK_NONE)
919 swp_pager_update_freerange(&s_free, &n_free,
923 swp_pager_freeswapspace(s_free, n_free);
924 VM_OBJECT_WUNLOCK(object);
929 * SWAP_PAGER_COPY() - copy blocks from source pager to destination pager
930 * and destroy the source.
932 * Copy any valid swapblks from the source to the destination. In
933 * cases where both the source and destination have a valid swapblk,
934 * we keep the destination's.
936 * This routine is allowed to sleep. It may sleep allocating metadata
937 * indirectly through swp_pager_meta_build() or if paging is still in
938 * progress on the source.
940 * The source object contains no vm_page_t's (which is just as well)
942 * The source object is of type OBJT_SWAP.
944 * The source and destination objects must be locked.
945 * Both object locks may temporarily be released.
948 swap_pager_copy(vm_object_t srcobject, vm_object_t dstobject,
949 vm_pindex_t offset, int destroysource)
952 daddr_t dstaddr, n_free, s_free, srcaddr;
954 VM_OBJECT_ASSERT_WLOCKED(srcobject);
955 VM_OBJECT_ASSERT_WLOCKED(dstobject);
958 * If destroysource is set, we remove the source object from the
959 * swap_pager internal queue now.
961 if (destroysource && srcobject->handle != NULL) {
962 vm_object_pip_add(srcobject, 1);
963 VM_OBJECT_WUNLOCK(srcobject);
964 vm_object_pip_add(dstobject, 1);
965 VM_OBJECT_WUNLOCK(dstobject);
966 sx_xlock(&sw_alloc_sx);
967 TAILQ_REMOVE(NOBJLIST(srcobject->handle), srcobject,
969 sx_xunlock(&sw_alloc_sx);
970 VM_OBJECT_WLOCK(dstobject);
971 vm_object_pip_wakeup(dstobject);
972 VM_OBJECT_WLOCK(srcobject);
973 vm_object_pip_wakeup(srcobject);
977 * Transfer source to destination.
979 swp_pager_init_freerange(&s_free, &n_free);
980 for (i = 0; i < dstobject->size; ++i) {
981 srcaddr = swp_pager_meta_ctl(srcobject, i + offset, SWM_POP);
982 if (srcaddr == SWAPBLK_NONE)
984 dstaddr = swp_pager_meta_ctl(dstobject, i, 0);
985 if (dstaddr != SWAPBLK_NONE) {
987 * Destination has valid swapblk or it is represented
988 * by a resident page. We destroy the source block.
990 swp_pager_update_freerange(&s_free, &n_free, srcaddr);
995 * Destination has no swapblk and is not resident,
998 * swp_pager_meta_build() can sleep.
1000 vm_object_pip_add(srcobject, 1);
1001 VM_OBJECT_WUNLOCK(srcobject);
1002 vm_object_pip_add(dstobject, 1);
1003 dstaddr = swp_pager_meta_build(dstobject, i, srcaddr);
1004 KASSERT(dstaddr == SWAPBLK_NONE,
1005 ("Unexpected destination swapblk"));
1006 vm_object_pip_wakeup(dstobject);
1007 VM_OBJECT_WLOCK(srcobject);
1008 vm_object_pip_wakeup(srcobject);
1010 swp_pager_freeswapspace(s_free, n_free);
1013 * Free left over swap blocks in source.
1015 * We have to revert the type to OBJT_DEFAULT so we do not accidentally
1016 * double-remove the object from the swap queues.
1018 if (destroysource) {
1019 swp_pager_meta_free_all(srcobject);
1021 * Reverting the type is not necessary, the caller is going
1022 * to destroy srcobject directly, but I'm doing it here
1023 * for consistency since we've removed the object from its
1026 srcobject->type = OBJT_DEFAULT;
1031 * SWAP_PAGER_HASPAGE() - determine if we have good backing store for
1032 * the requested page.
1034 * We determine whether good backing store exists for the requested
1035 * page and return TRUE if it does, FALSE if it doesn't.
1037 * If TRUE, we also try to determine how much valid, contiguous backing
1038 * store exists before and after the requested page.
1041 swap_pager_haspage(vm_object_t object, vm_pindex_t pindex, int *before,
1047 VM_OBJECT_ASSERT_LOCKED(object);
1050 * do we have good backing store at the requested index ?
1052 blk0 = swp_pager_meta_ctl(object, pindex, 0);
1053 if (blk0 == SWAPBLK_NONE) {
1062 * find backwards-looking contiguous good backing store
1064 if (before != NULL) {
1065 for (i = 1; i < SWB_NPAGES; i++) {
1068 blk = swp_pager_meta_ctl(object, pindex - i, 0);
1069 if (blk != blk0 - i)
1076 * find forward-looking contiguous good backing store
1078 if (after != NULL) {
1079 for (i = 1; i < SWB_NPAGES; i++) {
1080 blk = swp_pager_meta_ctl(object, pindex + i, 0);
1081 if (blk != blk0 + i)
1090 * SWAP_PAGER_PAGE_UNSWAPPED() - remove swap backing store related to page
1092 * This removes any associated swap backing store, whether valid or
1093 * not, from the page.
1095 * This routine is typically called when a page is made dirty, at
1096 * which point any associated swap can be freed. MADV_FREE also
1097 * calls us in a special-case situation
1099 * NOTE!!! If the page is clean and the swap was valid, the caller
1100 * should make the page dirty before calling this routine. This routine
1101 * does NOT change the m->dirty status of the page. Also: MADV_FREE
1104 * This routine may not sleep.
1106 * The object containing the page must be locked.
1109 swap_pager_unswapped(vm_page_t m)
1113 srcaddr = swp_pager_meta_ctl(m->object, m->pindex, SWM_POP);
1114 if (srcaddr != SWAPBLK_NONE)
1115 swp_pager_freeswapspace(srcaddr, 1);
1119 * swap_pager_getpages() - bring pages in from swap
1121 * Attempt to page in the pages in array "ma" of length "count". The
1122 * caller may optionally specify that additional pages preceding and
1123 * succeeding the specified range be paged in. The number of such pages
1124 * is returned in the "rbehind" and "rahead" parameters, and they will
1125 * be in the inactive queue upon return.
1127 * The pages in "ma" must be busied and will remain busied upon return.
1130 swap_pager_getpages(vm_object_t object, vm_page_t *ma, int count, int *rbehind,
1134 vm_page_t bm, mpred, msucc, p;
1137 int i, maxahead, maxbehind, reqcount;
1142 * Determine the final number of read-behind pages and
1143 * allocate them BEFORE releasing the object lock. Otherwise,
1144 * there can be a problematic race with vm_object_split().
1145 * Specifically, vm_object_split() might first transfer pages
1146 * that precede ma[0] in the current object to a new object,
1147 * and then this function incorrectly recreates those pages as
1148 * read-behind pages in the current object.
1150 if (!swap_pager_haspage(object, ma[0]->pindex, &maxbehind, &maxahead))
1151 return (VM_PAGER_FAIL);
1154 * Clip the readahead and readbehind ranges to exclude resident pages.
1156 if (rahead != NULL) {
1157 KASSERT(reqcount - 1 <= maxahead,
1158 ("page count %d extends beyond swap block", reqcount));
1159 *rahead = imin(*rahead, maxahead - (reqcount - 1));
1160 pindex = ma[reqcount - 1]->pindex;
1161 msucc = TAILQ_NEXT(ma[reqcount - 1], listq);
1162 if (msucc != NULL && msucc->pindex - pindex - 1 < *rahead)
1163 *rahead = msucc->pindex - pindex - 1;
1165 if (rbehind != NULL) {
1166 *rbehind = imin(*rbehind, maxbehind);
1167 pindex = ma[0]->pindex;
1168 mpred = TAILQ_PREV(ma[0], pglist, listq);
1169 if (mpred != NULL && pindex - mpred->pindex - 1 < *rbehind)
1170 *rbehind = pindex - mpred->pindex - 1;
1174 for (i = 0; i < count; i++)
1175 ma[i]->oflags |= VPO_SWAPINPROG;
1178 * Allocate readahead and readbehind pages.
1180 if (rbehind != NULL) {
1181 for (i = 1; i <= *rbehind; i++) {
1182 p = vm_page_alloc(object, ma[0]->pindex - i,
1186 p->oflags |= VPO_SWAPINPROG;
1191 if (rahead != NULL) {
1192 for (i = 0; i < *rahead; i++) {
1193 p = vm_page_alloc(object,
1194 ma[reqcount - 1]->pindex + i + 1, VM_ALLOC_NORMAL);
1197 p->oflags |= VPO_SWAPINPROG;
1201 if (rbehind != NULL)
1206 vm_object_pip_add(object, count);
1208 pindex = bm->pindex;
1209 blk = swp_pager_meta_ctl(object, pindex, 0);
1210 KASSERT(blk != SWAPBLK_NONE,
1211 ("no swap blocking containing %p(%jx)", object, (uintmax_t)pindex));
1213 VM_OBJECT_WUNLOCK(object);
1214 bp = uma_zalloc(swrbuf_zone, M_WAITOK);
1215 /* Pages cannot leave the object while busy. */
1216 for (i = 0, p = bm; i < count; i++, p = TAILQ_NEXT(p, listq)) {
1217 MPASS(p->pindex == bm->pindex + i);
1221 bp->b_flags |= B_PAGING;
1222 bp->b_iocmd = BIO_READ;
1223 bp->b_iodone = swp_pager_async_iodone;
1224 bp->b_rcred = crhold(thread0.td_ucred);
1225 bp->b_wcred = crhold(thread0.td_ucred);
1227 bp->b_bcount = PAGE_SIZE * count;
1228 bp->b_bufsize = PAGE_SIZE * count;
1229 bp->b_npages = count;
1230 bp->b_pgbefore = rbehind != NULL ? *rbehind : 0;
1231 bp->b_pgafter = rahead != NULL ? *rahead : 0;
1233 VM_CNT_INC(v_swapin);
1234 VM_CNT_ADD(v_swappgsin, count);
1237 * perform the I/O. NOTE!!! bp cannot be considered valid after
1238 * this point because we automatically release it on completion.
1239 * Instead, we look at the one page we are interested in which we
1240 * still hold a lock on even through the I/O completion.
1242 * The other pages in our ma[] array are also released on completion,
1243 * so we cannot assume they are valid anymore either.
1245 * NOTE: b_blkno is destroyed by the call to swapdev_strategy
1248 swp_pager_strategy(bp);
1251 * Wait for the pages we want to complete. VPO_SWAPINPROG is always
1252 * cleared on completion. If an I/O error occurs, SWAPBLK_NONE
1253 * is set in the metadata for each page in the request.
1255 VM_OBJECT_WLOCK(object);
1256 while ((ma[0]->oflags & VPO_SWAPINPROG) != 0) {
1257 ma[0]->oflags |= VPO_SWAPSLEEP;
1258 VM_CNT_INC(v_intrans);
1259 if (VM_OBJECT_SLEEP(object, &object->handle, PSWP,
1260 "swread", hz * 20)) {
1262 "swap_pager: indefinite wait buffer: bufobj: %p, blkno: %jd, size: %ld\n",
1263 bp->b_bufobj, (intmax_t)bp->b_blkno, bp->b_bcount);
1268 * If we had an unrecoverable read error pages will not be valid.
1270 for (i = 0; i < reqcount; i++)
1271 if (ma[i]->valid != VM_PAGE_BITS_ALL)
1272 return (VM_PAGER_ERROR);
1274 return (VM_PAGER_OK);
1277 * A final note: in a low swap situation, we cannot deallocate swap
1278 * and mark a page dirty here because the caller is likely to mark
1279 * the page clean when we return, causing the page to possibly revert
1280 * to all-zero's later.
1285 * swap_pager_getpages_async():
1287 * Right now this is emulation of asynchronous operation on top of
1288 * swap_pager_getpages().
1291 swap_pager_getpages_async(vm_object_t object, vm_page_t *ma, int count,
1292 int *rbehind, int *rahead, pgo_getpages_iodone_t iodone, void *arg)
1296 r = swap_pager_getpages(object, ma, count, rbehind, rahead);
1297 VM_OBJECT_WUNLOCK(object);
1302 case VM_PAGER_ERROR:
1309 panic("unhandled swap_pager_getpages() error %d", r);
1311 (iodone)(arg, ma, count, error);
1312 VM_OBJECT_WLOCK(object);
1318 * swap_pager_putpages:
1320 * Assign swap (if necessary) and initiate I/O on the specified pages.
1322 * We support both OBJT_DEFAULT and OBJT_SWAP objects. DEFAULT objects
1323 * are automatically converted to SWAP objects.
1325 * In a low memory situation we may block in VOP_STRATEGY(), but the new
1326 * vm_page reservation system coupled with properly written VFS devices
1327 * should ensure that no low-memory deadlock occurs. This is an area
1330 * The parent has N vm_object_pip_add() references prior to
1331 * calling us and will remove references for rtvals[] that are
1332 * not set to VM_PAGER_PEND. We need to remove the rest on I/O
1335 * The parent has soft-busy'd the pages it passes us and will unbusy
1336 * those whose rtvals[] entry is not set to VM_PAGER_PEND on return.
1337 * We need to unbusy the rest on I/O completion.
1340 swap_pager_putpages(vm_object_t object, vm_page_t *ma, int count,
1341 int flags, int *rtvals)
1344 daddr_t addr, blk, n_free, s_free;
1349 KASSERT(count == 0 || ma[0]->object == object,
1350 ("%s: object mismatch %p/%p",
1351 __func__, object, ma[0]->object));
1356 * Turn object into OBJT_SWAP. Force sync if not a pageout process.
1358 if (object->type != OBJT_SWAP) {
1359 addr = swp_pager_meta_build(object, 0, SWAPBLK_NONE);
1360 KASSERT(addr == SWAPBLK_NONE,
1361 ("unexpected object swap block"));
1363 VM_OBJECT_WUNLOCK(object);
1364 async = curproc == pageproc && (flags & VM_PAGER_PUT_SYNC) == 0;
1365 swp_pager_init_freerange(&s_free, &n_free);
1370 * Assign swap blocks and issue I/O. We reallocate swap on the fly.
1371 * The page is left dirty until the pageout operation completes
1374 for (i = 0; i < count; i += n) {
1375 /* Maximum I/O size is limited by maximum swap block size. */
1376 n = min(count - i, nsw_cluster_max);
1378 /* Get a block of swap of size up to size n. */
1379 blk = swp_pager_getswapspace(&n, 4);
1380 if (blk == SWAPBLK_NONE) {
1381 for (j = 0; j < n; ++j)
1382 rtvals[i + j] = VM_PAGER_FAIL;
1387 * All I/O parameters have been satisfied. Build the I/O
1388 * request and assign the swap space.
1391 mtx_lock(&swbuf_mtx);
1392 while (nsw_wcount_async == 0)
1393 msleep(&nsw_wcount_async, &swbuf_mtx, PVM,
1396 mtx_unlock(&swbuf_mtx);
1398 bp = uma_zalloc(swwbuf_zone, M_WAITOK);
1400 bp->b_flags = B_ASYNC;
1401 bp->b_flags |= B_PAGING;
1402 bp->b_iocmd = BIO_WRITE;
1404 bp->b_rcred = crhold(thread0.td_ucred);
1405 bp->b_wcred = crhold(thread0.td_ucred);
1406 bp->b_bcount = PAGE_SIZE * n;
1407 bp->b_bufsize = PAGE_SIZE * n;
1410 VM_OBJECT_WLOCK(object);
1411 for (j = 0; j < n; ++j) {
1413 addr = swp_pager_meta_build(mreq->object, mreq->pindex,
1415 if (addr != SWAPBLK_NONE)
1416 swp_pager_update_freerange(&s_free, &n_free,
1418 MPASS(mreq->dirty == VM_PAGE_BITS_ALL);
1419 mreq->oflags |= VPO_SWAPINPROG;
1420 bp->b_pages[j] = mreq;
1422 VM_OBJECT_WUNLOCK(object);
1425 * Must set dirty range for NFS to work.
1428 bp->b_dirtyend = bp->b_bcount;
1430 VM_CNT_INC(v_swapout);
1431 VM_CNT_ADD(v_swappgsout, bp->b_npages);
1434 * We unconditionally set rtvals[] to VM_PAGER_PEND so that we
1435 * can call the async completion routine at the end of a
1436 * synchronous I/O operation. Otherwise, our caller would
1437 * perform duplicate unbusy and wakeup operations on the page
1438 * and object, respectively.
1440 for (j = 0; j < n; j++)
1441 rtvals[i + j] = VM_PAGER_PEND;
1446 * NOTE: b_blkno is destroyed by the call to swapdev_strategy.
1449 bp->b_iodone = swp_pager_async_iodone;
1451 swp_pager_strategy(bp);
1458 * NOTE: b_blkno is destroyed by the call to swapdev_strategy.
1460 bp->b_iodone = bdone;
1461 swp_pager_strategy(bp);
1464 * Wait for the sync I/O to complete.
1466 bwait(bp, PVM, "swwrt");
1469 * Now that we are through with the bp, we can call the
1470 * normal async completion, which frees everything up.
1472 swp_pager_async_iodone(bp);
1474 swp_pager_freeswapspace(s_free, n_free);
1475 VM_OBJECT_WLOCK(object);
1479 * swp_pager_async_iodone:
1481 * Completion routine for asynchronous reads and writes from/to swap.
1482 * Also called manually by synchronous code to finish up a bp.
1484 * This routine may not sleep.
1487 swp_pager_async_iodone(struct buf *bp)
1490 vm_object_t object = NULL;
1493 * Report error - unless we ran out of memory, in which case
1494 * we've already logged it in swapgeom_strategy().
1496 if (bp->b_ioflags & BIO_ERROR && bp->b_error != ENOMEM) {
1498 "swap_pager: I/O error - %s failed; blkno %ld,"
1499 "size %ld, error %d\n",
1500 ((bp->b_iocmd == BIO_READ) ? "pagein" : "pageout"),
1508 * remove the mapping for kernel virtual
1511 pmap_qremove((vm_offset_t)bp->b_data, bp->b_npages);
1513 bp->b_data = bp->b_kvabase;
1516 object = bp->b_pages[0]->object;
1517 VM_OBJECT_WLOCK(object);
1521 * cleanup pages. If an error occurs writing to swap, we are in
1522 * very serious trouble. If it happens to be a disk error, though,
1523 * we may be able to recover by reassigning the swap later on. So
1524 * in this case we remove the m->swapblk assignment for the page
1525 * but do not free it in the rlist. The errornous block(s) are thus
1526 * never reallocated as swap. Redirty the page and continue.
1528 for (i = 0; i < bp->b_npages; ++i) {
1529 vm_page_t m = bp->b_pages[i];
1531 m->oflags &= ~VPO_SWAPINPROG;
1532 if (m->oflags & VPO_SWAPSLEEP) {
1533 m->oflags &= ~VPO_SWAPSLEEP;
1534 wakeup(&object->handle);
1537 if (bp->b_ioflags & BIO_ERROR) {
1539 * If an error occurs I'd love to throw the swapblk
1540 * away without freeing it back to swapspace, so it
1541 * can never be used again. But I can't from an
1544 if (bp->b_iocmd == BIO_READ) {
1546 * NOTE: for reads, m->dirty will probably
1547 * be overridden by the original caller of
1548 * getpages so don't play cute tricks here.
1553 * If a write error occurs, reactivate page
1554 * so it doesn't clog the inactive list,
1555 * then finish the I/O.
1557 MPASS(m->dirty == VM_PAGE_BITS_ALL);
1559 vm_page_activate(m);
1563 } else if (bp->b_iocmd == BIO_READ) {
1565 * NOTE: for reads, m->dirty will probably be
1566 * overridden by the original caller of getpages so
1567 * we cannot set them in order to free the underlying
1568 * swap in a low-swap situation. I don't think we'd
1569 * want to do that anyway, but it was an optimization
1570 * that existed in the old swapper for a time before
1571 * it got ripped out due to precisely this problem.
1573 KASSERT(!pmap_page_is_mapped(m),
1574 ("swp_pager_async_iodone: page %p is mapped", m));
1575 KASSERT(m->dirty == 0,
1576 ("swp_pager_async_iodone: page %p is dirty", m));
1578 m->valid = VM_PAGE_BITS_ALL;
1579 if (i < bp->b_pgbefore ||
1580 i >= bp->b_npages - bp->b_pgafter)
1581 vm_page_readahead_finish(m);
1584 * For write success, clear the dirty
1585 * status, then finish the I/O ( which decrements the
1586 * busy count and possibly wakes waiter's up ).
1587 * A page is only written to swap after a period of
1588 * inactivity. Therefore, we do not expect it to be
1591 KASSERT(!pmap_page_is_write_mapped(m),
1592 ("swp_pager_async_iodone: page %p is not write"
1596 vm_page_deactivate_noreuse(m);
1603 * adjust pip. NOTE: the original parent may still have its own
1604 * pip refs on the object.
1606 if (object != NULL) {
1607 vm_object_pip_wakeupn(object, bp->b_npages);
1608 VM_OBJECT_WUNLOCK(object);
1612 * swapdev_strategy() manually sets b_vp and b_bufobj before calling
1613 * bstrategy(). Set them back to NULL now we're done with it, or we'll
1614 * trigger a KASSERT in relpbuf().
1618 bp->b_bufobj = NULL;
1621 * release the physical I/O buffer
1623 if (bp->b_flags & B_ASYNC) {
1624 mtx_lock(&swbuf_mtx);
1625 if (++nsw_wcount_async == 1)
1626 wakeup(&nsw_wcount_async);
1627 mtx_unlock(&swbuf_mtx);
1629 uma_zfree((bp->b_iocmd == BIO_READ) ? swrbuf_zone : swwbuf_zone, bp);
1633 swap_pager_nswapdev(void)
1640 swp_pager_force_dirty(vm_page_t m)
1646 if (!vm_page_wired(m) && m->queue == PQ_NONE)
1647 panic("page %p is neither wired nor queued", m);
1651 swap_pager_unswapped(m);
1655 swp_pager_force_launder(vm_page_t m)
1663 swap_pager_unswapped(m);
1667 * SWP_PAGER_FORCE_PAGEIN() - force swap blocks to be paged in
1669 * This routine dissociates pages starting at the given index within an
1670 * object from their backing store, paging them in if they do not reside
1671 * in memory. Pages that are paged in are marked dirty and placed in the
1672 * laundry queue. Pages are marked dirty because they no longer have
1673 * backing store. They are placed in the laundry queue because they have
1674 * not been accessed recently. Otherwise, they would already reside in
1678 swp_pager_force_pagein(vm_object_t object, vm_pindex_t pindex, int npages)
1680 vm_page_t ma[npages];
1683 KASSERT(npages > 0, ("%s: No pages", __func__));
1684 KASSERT(npages <= MAXPHYS / PAGE_SIZE,
1685 ("%s: Too many pages: %d", __func__, npages));
1686 vm_object_pip_add(object, npages);
1687 vm_page_grab_pages(object, pindex, VM_ALLOC_NORMAL, ma, npages);
1688 for (i = j = 0;; i++) {
1689 /* Count nonresident pages, to page-in all at once. */
1690 if (i < npages && ma[i]->valid != VM_PAGE_BITS_ALL)
1693 /* Page-in nonresident pages. Mark for laundering. */
1694 if (swap_pager_getpages(object, &ma[j], i - j, NULL,
1695 NULL) != VM_PAGER_OK)
1696 panic("%s: read from swap failed", __func__);
1698 swp_pager_force_launder(ma[j]);
1703 /* Mark dirty a resident page. */
1704 swp_pager_force_dirty(ma[j++]);
1706 vm_object_pip_wakeupn(object, npages);
1710 * swap_pager_swapoff_object:
1712 * Page in all of the pages that have been paged out for an object
1716 swap_pager_swapoff_object(struct swdevt *sp, vm_object_t object)
1719 vm_pindex_t pi, s_pindex;
1720 daddr_t blk, n_blks, s_blk;
1724 for (pi = 0; (sb = SWAP_PCTRIE_LOOKUP_GE(
1725 &object->un_pager.swp.swp_blks, pi)) != NULL; ) {
1726 for (i = 0; i < SWAP_META_PAGES; i++) {
1728 if (!swp_pager_isondev(blk, sp))
1732 * If there are no blocks/pages accumulated, start a new
1733 * accumulation here.
1736 if (blk != SWAPBLK_NONE) {
1738 s_pindex = sb->p + i;
1745 * If the accumulation can be extended without breaking
1746 * the sequence of consecutive blocks and pages that
1747 * swp_pager_force_pagein() depends on, do so.
1749 if (n_blks < MAXPHYS / PAGE_SIZE &&
1750 s_blk + n_blks == blk &&
1751 s_pindex + n_blks == sb->p + i) {
1757 * The sequence of consecutive blocks and pages cannot
1758 * be extended, so page them all in here. Then,
1759 * because doing so involves releasing and reacquiring
1760 * a lock that protects the swap block pctrie, do not
1761 * rely on the current swap block. Break this loop and
1762 * re-fetch the same pindex from the pctrie again.
1764 swp_pager_force_pagein(object, s_pindex, n_blks);
1768 if (i == SWAP_META_PAGES)
1769 pi = sb->p + SWAP_META_PAGES;
1772 swp_pager_force_pagein(object, s_pindex, n_blks);
1776 * swap_pager_swapoff:
1778 * Page in all of the pages that have been paged out to the
1779 * given device. The corresponding blocks in the bitmap must be
1780 * marked as allocated and the device must be flagged SW_CLOSING.
1781 * There may be no processes swapped out to the device.
1783 * This routine may block.
1786 swap_pager_swapoff(struct swdevt *sp)
1791 sx_assert(&swdev_syscall_lock, SA_XLOCKED);
1795 mtx_lock(&vm_object_list_mtx);
1796 TAILQ_FOREACH(object, &vm_object_list, object_list) {
1797 if (object->type != OBJT_SWAP)
1799 mtx_unlock(&vm_object_list_mtx);
1800 /* Depends on type-stability. */
1801 VM_OBJECT_WLOCK(object);
1804 * Dead objects are eventually terminated on their own.
1806 if ((object->flags & OBJ_DEAD) != 0)
1810 * Sync with fences placed after pctrie
1811 * initialization. We must not access pctrie below
1812 * unless we checked that our object is swap and not
1815 atomic_thread_fence_acq();
1816 if (object->type != OBJT_SWAP)
1819 swap_pager_swapoff_object(sp, object);
1821 VM_OBJECT_WUNLOCK(object);
1822 mtx_lock(&vm_object_list_mtx);
1824 mtx_unlock(&vm_object_list_mtx);
1828 * Objects may be locked or paging to the device being
1829 * removed, so we will miss their pages and need to
1830 * make another pass. We have marked this device as
1831 * SW_CLOSING, so the activity should finish soon.
1834 if (retries > 100) {
1835 panic("swapoff: failed to locate %d swap blocks",
1838 pause("swpoff", hz / 20);
1841 EVENTHANDLER_INVOKE(swapoff, sp);
1844 /************************************************************************
1846 ************************************************************************
1848 * These routines manipulate the swap metadata stored in the
1851 * Swap metadata is implemented with a global hash and not directly
1852 * linked into the object. Instead the object simply contains
1853 * appropriate tracking counters.
1857 * SWP_PAGER_SWBLK_EMPTY() - is a range of blocks free?
1860 swp_pager_swblk_empty(struct swblk *sb, int start, int limit)
1864 MPASS(0 <= start && start <= limit && limit <= SWAP_META_PAGES);
1865 for (i = start; i < limit; i++) {
1866 if (sb->d[i] != SWAPBLK_NONE)
1873 * SWP_PAGER_META_BUILD() - add swap block to swap meta data for object
1875 * We first convert the object to a swap object if it is a default
1878 * The specified swapblk is added to the object's swap metadata. If
1879 * the swapblk is not valid, it is freed instead. Any previously
1880 * assigned swapblk is returned.
1883 swp_pager_meta_build(vm_object_t object, vm_pindex_t pindex, daddr_t swapblk)
1885 static volatile int swblk_zone_exhausted, swpctrie_zone_exhausted;
1886 struct swblk *sb, *sb1;
1887 vm_pindex_t modpi, rdpi;
1888 daddr_t prev_swapblk;
1891 VM_OBJECT_ASSERT_WLOCKED(object);
1894 * Convert default object to swap object if necessary
1896 if (object->type != OBJT_SWAP) {
1897 pctrie_init(&object->un_pager.swp.swp_blks);
1900 * Ensure that swap_pager_swapoff()'s iteration over
1901 * object_list does not see a garbage pctrie.
1903 atomic_thread_fence_rel();
1905 object->type = OBJT_SWAP;
1906 KASSERT(object->handle == NULL, ("default pager with handle"));
1909 rdpi = rounddown(pindex, SWAP_META_PAGES);
1910 sb = SWAP_PCTRIE_LOOKUP(&object->un_pager.swp.swp_blks, rdpi);
1912 if (swapblk == SWAPBLK_NONE)
1913 return (SWAPBLK_NONE);
1915 sb = uma_zalloc(swblk_zone, M_NOWAIT | (curproc ==
1916 pageproc ? M_USE_RESERVE : 0));
1919 for (i = 0; i < SWAP_META_PAGES; i++)
1920 sb->d[i] = SWAPBLK_NONE;
1921 if (atomic_cmpset_int(&swblk_zone_exhausted,
1923 printf("swblk zone ok\n");
1926 VM_OBJECT_WUNLOCK(object);
1927 if (uma_zone_exhausted(swblk_zone)) {
1928 if (atomic_cmpset_int(&swblk_zone_exhausted,
1930 printf("swap blk zone exhausted, "
1931 "increase kern.maxswzone\n");
1932 vm_pageout_oom(VM_OOM_SWAPZ);
1933 pause("swzonxb", 10);
1935 uma_zwait(swblk_zone);
1936 VM_OBJECT_WLOCK(object);
1937 sb = SWAP_PCTRIE_LOOKUP(&object->un_pager.swp.swp_blks,
1941 * Somebody swapped out a nearby page,
1942 * allocating swblk at the rdpi index,
1943 * while we dropped the object lock.
1948 error = SWAP_PCTRIE_INSERT(
1949 &object->un_pager.swp.swp_blks, sb);
1951 if (atomic_cmpset_int(&swpctrie_zone_exhausted,
1953 printf("swpctrie zone ok\n");
1956 VM_OBJECT_WUNLOCK(object);
1957 if (uma_zone_exhausted(swpctrie_zone)) {
1958 if (atomic_cmpset_int(&swpctrie_zone_exhausted,
1960 printf("swap pctrie zone exhausted, "
1961 "increase kern.maxswzone\n");
1962 vm_pageout_oom(VM_OOM_SWAPZ);
1963 pause("swzonxp", 10);
1965 uma_zwait(swpctrie_zone);
1966 VM_OBJECT_WLOCK(object);
1967 sb1 = SWAP_PCTRIE_LOOKUP(&object->un_pager.swp.swp_blks,
1970 uma_zfree(swblk_zone, sb);
1977 MPASS(sb->p == rdpi);
1979 modpi = pindex % SWAP_META_PAGES;
1980 /* Return prior contents of metadata. */
1981 prev_swapblk = sb->d[modpi];
1982 /* Enter block into metadata. */
1983 sb->d[modpi] = swapblk;
1986 * Free the swblk if we end up with the empty page run.
1988 if (swapblk == SWAPBLK_NONE &&
1989 swp_pager_swblk_empty(sb, 0, SWAP_META_PAGES)) {
1990 SWAP_PCTRIE_REMOVE(&object->un_pager.swp.swp_blks, rdpi);
1991 uma_zfree(swblk_zone, sb);
1993 return (prev_swapblk);
1997 * SWP_PAGER_META_FREE() - free a range of blocks in the object's swap metadata
1999 * The requested range of blocks is freed, with any associated swap
2000 * returned to the swap bitmap.
2002 * This routine will free swap metadata structures as they are cleaned
2003 * out. This routine does *NOT* operate on swap metadata associated
2004 * with resident pages.
2007 swp_pager_meta_free(vm_object_t object, vm_pindex_t pindex, vm_pindex_t count)
2010 daddr_t n_free, s_free;
2012 int i, limit, start;
2014 VM_OBJECT_ASSERT_WLOCKED(object);
2015 if (object->type != OBJT_SWAP || count == 0)
2018 swp_pager_init_freerange(&s_free, &n_free);
2019 last = pindex + count;
2021 sb = SWAP_PCTRIE_LOOKUP_GE(&object->un_pager.swp.swp_blks,
2022 rounddown(pindex, SWAP_META_PAGES));
2023 if (sb == NULL || sb->p >= last)
2025 start = pindex > sb->p ? pindex - sb->p : 0;
2026 limit = last - sb->p < SWAP_META_PAGES ? last - sb->p :
2028 for (i = start; i < limit; i++) {
2029 if (sb->d[i] == SWAPBLK_NONE)
2031 swp_pager_update_freerange(&s_free, &n_free, sb->d[i]);
2032 sb->d[i] = SWAPBLK_NONE;
2034 pindex = sb->p + SWAP_META_PAGES;
2035 if (swp_pager_swblk_empty(sb, 0, start) &&
2036 swp_pager_swblk_empty(sb, limit, SWAP_META_PAGES)) {
2037 SWAP_PCTRIE_REMOVE(&object->un_pager.swp.swp_blks,
2039 uma_zfree(swblk_zone, sb);
2042 swp_pager_freeswapspace(s_free, n_free);
2046 * SWP_PAGER_META_FREE_ALL() - destroy all swap metadata associated with object
2048 * This routine locates and destroys all swap metadata associated with
2052 swp_pager_meta_free_all(vm_object_t object)
2055 daddr_t n_free, s_free;
2059 VM_OBJECT_ASSERT_WLOCKED(object);
2060 if (object->type != OBJT_SWAP)
2063 swp_pager_init_freerange(&s_free, &n_free);
2064 for (pindex = 0; (sb = SWAP_PCTRIE_LOOKUP_GE(
2065 &object->un_pager.swp.swp_blks, pindex)) != NULL;) {
2066 pindex = sb->p + SWAP_META_PAGES;
2067 for (i = 0; i < SWAP_META_PAGES; i++) {
2068 if (sb->d[i] == SWAPBLK_NONE)
2070 swp_pager_update_freerange(&s_free, &n_free, sb->d[i]);
2072 SWAP_PCTRIE_REMOVE(&object->un_pager.swp.swp_blks, sb->p);
2073 uma_zfree(swblk_zone, sb);
2075 swp_pager_freeswapspace(s_free, n_free);
2079 * SWP_PAGER_METACTL() - misc control of swap meta data.
2081 * This routine is capable of looking up, or removing swapblk
2082 * assignments in the swap meta data. It returns the swapblk being
2083 * looked-up, popped, or SWAPBLK_NONE if the block was invalid.
2085 * When acting on a busy resident page and paging is in progress, we
2086 * have to wait until paging is complete but otherwise can act on the
2089 * SWM_POP remove from meta data but do not free it
2092 swp_pager_meta_ctl(vm_object_t object, vm_pindex_t pindex, int flags)
2097 if ((flags & SWM_POP) != 0)
2098 VM_OBJECT_ASSERT_WLOCKED(object);
2100 VM_OBJECT_ASSERT_LOCKED(object);
2103 * The meta data only exists if the object is OBJT_SWAP
2104 * and even then might not be allocated yet.
2106 if (object->type != OBJT_SWAP)
2107 return (SWAPBLK_NONE);
2109 sb = SWAP_PCTRIE_LOOKUP(&object->un_pager.swp.swp_blks,
2110 rounddown(pindex, SWAP_META_PAGES));
2112 return (SWAPBLK_NONE);
2113 r1 = sb->d[pindex % SWAP_META_PAGES];
2114 if (r1 == SWAPBLK_NONE)
2115 return (SWAPBLK_NONE);
2116 if ((flags & SWM_POP) != 0) {
2117 sb->d[pindex % SWAP_META_PAGES] = SWAPBLK_NONE;
2118 if (swp_pager_swblk_empty(sb, 0, SWAP_META_PAGES)) {
2119 SWAP_PCTRIE_REMOVE(&object->un_pager.swp.swp_blks,
2120 rounddown(pindex, SWAP_META_PAGES));
2121 uma_zfree(swblk_zone, sb);
2128 * Returns the least page index which is greater than or equal to the
2129 * parameter pindex and for which there is a swap block allocated.
2130 * Returns object's size if the object's type is not swap or if there
2131 * are no allocated swap blocks for the object after the requested
2135 swap_pager_find_least(vm_object_t object, vm_pindex_t pindex)
2140 VM_OBJECT_ASSERT_LOCKED(object);
2141 if (object->type != OBJT_SWAP)
2142 return (object->size);
2144 sb = SWAP_PCTRIE_LOOKUP_GE(&object->un_pager.swp.swp_blks,
2145 rounddown(pindex, SWAP_META_PAGES));
2147 return (object->size);
2148 if (sb->p < pindex) {
2149 for (i = pindex % SWAP_META_PAGES; i < SWAP_META_PAGES; i++) {
2150 if (sb->d[i] != SWAPBLK_NONE)
2153 sb = SWAP_PCTRIE_LOOKUP_GE(&object->un_pager.swp.swp_blks,
2154 roundup(pindex, SWAP_META_PAGES));
2156 return (object->size);
2158 for (i = 0; i < SWAP_META_PAGES; i++) {
2159 if (sb->d[i] != SWAPBLK_NONE)
2164 * We get here if a swblk is present in the trie but it
2165 * doesn't map any blocks.
2168 return (object->size);
2172 * System call swapon(name) enables swapping on device name,
2173 * which must be in the swdevsw. Return EBUSY
2174 * if already swapping on this device.
2176 #ifndef _SYS_SYSPROTO_H_
2177 struct swapon_args {
2187 sys_swapon(struct thread *td, struct swapon_args *uap)
2191 struct nameidata nd;
2194 error = priv_check(td, PRIV_SWAPON);
2198 sx_xlock(&swdev_syscall_lock);
2201 * Swap metadata may not fit in the KVM if we have physical
2204 if (swblk_zone == NULL) {
2209 NDINIT(&nd, LOOKUP, ISOPEN | FOLLOW | AUDITVNODE1, UIO_USERSPACE,
2215 NDFREE(&nd, NDF_ONLY_PNBUF);
2218 if (vn_isdisk(vp, &error)) {
2219 error = swapongeom(vp);
2220 } else if (vp->v_type == VREG &&
2221 (vp->v_mount->mnt_vfc->vfc_flags & VFCF_NETWORK) != 0 &&
2222 (error = VOP_GETATTR(vp, &attr, td->td_ucred)) == 0) {
2224 * Allow direct swapping to NFS regular files in the same
2225 * way that nfs_mountroot() sets up diskless swapping.
2227 error = swaponvp(td, vp, attr.va_size / DEV_BSIZE);
2233 sx_xunlock(&swdev_syscall_lock);
2238 * Check that the total amount of swap currently configured does not
2239 * exceed half the theoretical maximum. If it does, print a warning
2243 swapon_check_swzone(void)
2245 unsigned long maxpages, npages;
2247 npages = swap_total;
2248 /* absolute maximum we can handle assuming 100% efficiency */
2249 maxpages = uma_zone_get_max(swblk_zone) * SWAP_META_PAGES;
2251 /* recommend using no more than half that amount */
2252 if (npages > maxpages / 2) {
2253 printf("warning: total configured swap (%lu pages) "
2254 "exceeds maximum recommended amount (%lu pages).\n",
2255 npages, maxpages / 2);
2256 printf("warning: increase kern.maxswzone "
2257 "or reduce amount of swap.\n");
2262 swaponsomething(struct vnode *vp, void *id, u_long nblks,
2263 sw_strategy_t *strategy, sw_close_t *close, dev_t dev, int flags)
2265 struct swdevt *sp, *tsp;
2270 * nblks is in DEV_BSIZE'd chunks, convert to PAGE_SIZE'd chunks.
2271 * First chop nblks off to page-align it, then convert.
2273 * sw->sw_nblks is in page-sized chunks now too.
2275 nblks &= ~(ctodb(1) - 1);
2276 nblks = dbtoc(nblks);
2279 * If we go beyond this, we get overflows in the radix
2282 mblocks = 0x40000000 / BLIST_META_RADIX;
2283 if (nblks > mblocks) {
2285 "WARNING: reducing swap size to maximum of %luMB per unit\n",
2286 mblocks / 1024 / 1024 * PAGE_SIZE);
2290 sp = malloc(sizeof *sp, M_VMPGDATA, M_WAITOK | M_ZERO);
2294 sp->sw_nblks = nblks;
2296 sp->sw_strategy = strategy;
2297 sp->sw_close = close;
2298 sp->sw_flags = flags;
2300 sp->sw_blist = blist_create(nblks, M_WAITOK);
2302 * Do not free the first blocks in order to avoid overwriting
2303 * any bsd label at the front of the partition
2305 blist_free(sp->sw_blist, howmany(BBSIZE, PAGE_SIZE),
2306 nblks - howmany(BBSIZE, PAGE_SIZE));
2309 mtx_lock(&sw_dev_mtx);
2310 TAILQ_FOREACH(tsp, &swtailq, sw_list) {
2311 if (tsp->sw_end >= dvbase) {
2313 * We put one uncovered page between the devices
2314 * in order to definitively prevent any cross-device
2317 dvbase = tsp->sw_end + 1;
2320 sp->sw_first = dvbase;
2321 sp->sw_end = dvbase + nblks;
2322 TAILQ_INSERT_TAIL(&swtailq, sp, sw_list);
2324 swap_pager_avail += nblks - howmany(BBSIZE, PAGE_SIZE);
2325 swap_total += nblks;
2326 swapon_check_swzone();
2328 mtx_unlock(&sw_dev_mtx);
2329 EVENTHANDLER_INVOKE(swapon, sp);
2333 * SYSCALL: swapoff(devname)
2335 * Disable swapping on the given device.
2337 * XXX: Badly designed system call: it should use a device index
2338 * rather than filename as specification. We keep sw_vp around
2339 * only to make this work.
2341 #ifndef _SYS_SYSPROTO_H_
2342 struct swapoff_args {
2352 sys_swapoff(struct thread *td, struct swapoff_args *uap)
2355 struct nameidata nd;
2359 error = priv_check(td, PRIV_SWAPOFF);
2363 sx_xlock(&swdev_syscall_lock);
2365 NDINIT(&nd, LOOKUP, FOLLOW | AUDITVNODE1, UIO_USERSPACE, uap->name,
2370 NDFREE(&nd, NDF_ONLY_PNBUF);
2373 mtx_lock(&sw_dev_mtx);
2374 TAILQ_FOREACH(sp, &swtailq, sw_list) {
2375 if (sp->sw_vp == vp)
2378 mtx_unlock(&sw_dev_mtx);
2383 error = swapoff_one(sp, td->td_ucred);
2385 sx_xunlock(&swdev_syscall_lock);
2390 swapoff_one(struct swdevt *sp, struct ucred *cred)
2397 sx_assert(&swdev_syscall_lock, SA_XLOCKED);
2399 (void) vn_lock(sp->sw_vp, LK_EXCLUSIVE | LK_RETRY);
2400 error = mac_system_check_swapoff(cred, sp->sw_vp);
2401 (void) VOP_UNLOCK(sp->sw_vp, 0);
2405 nblks = sp->sw_nblks;
2408 * We can turn off this swap device safely only if the
2409 * available virtual memory in the system will fit the amount
2410 * of data we will have to page back in, plus an epsilon so
2411 * the system doesn't become critically low on swap space.
2413 if (vm_free_count() + swap_pager_avail < nblks + nswap_lowat)
2417 * Prevent further allocations on this device.
2419 mtx_lock(&sw_dev_mtx);
2420 sp->sw_flags |= SW_CLOSING;
2421 swap_pager_avail -= blist_fill(sp->sw_blist, 0, nblks);
2422 swap_total -= nblks;
2423 mtx_unlock(&sw_dev_mtx);
2426 * Page in the contents of the device and close it.
2428 swap_pager_swapoff(sp);
2430 sp->sw_close(curthread, sp);
2431 mtx_lock(&sw_dev_mtx);
2433 TAILQ_REMOVE(&swtailq, sp, sw_list);
2435 if (nswapdev == 0) {
2436 swap_pager_full = 2;
2437 swap_pager_almost_full = 1;
2441 mtx_unlock(&sw_dev_mtx);
2442 blist_destroy(sp->sw_blist);
2443 free(sp, M_VMPGDATA);
2450 struct swdevt *sp, *spt;
2451 const char *devname;
2454 sx_xlock(&swdev_syscall_lock);
2456 mtx_lock(&sw_dev_mtx);
2457 TAILQ_FOREACH_SAFE(sp, &swtailq, sw_list, spt) {
2458 mtx_unlock(&sw_dev_mtx);
2459 if (vn_isdisk(sp->sw_vp, NULL))
2460 devname = devtoname(sp->sw_vp->v_rdev);
2463 error = swapoff_one(sp, thread0.td_ucred);
2465 printf("Cannot remove swap device %s (error=%d), "
2466 "skipping.\n", devname, error);
2467 } else if (bootverbose) {
2468 printf("Swap device %s removed.\n", devname);
2470 mtx_lock(&sw_dev_mtx);
2472 mtx_unlock(&sw_dev_mtx);
2474 sx_xunlock(&swdev_syscall_lock);
2478 swap_pager_status(int *total, int *used)
2484 mtx_lock(&sw_dev_mtx);
2485 TAILQ_FOREACH(sp, &swtailq, sw_list) {
2486 *total += sp->sw_nblks;
2487 *used += sp->sw_used;
2489 mtx_unlock(&sw_dev_mtx);
2493 swap_dev_info(int name, struct xswdev *xs, char *devname, size_t len)
2496 const char *tmp_devname;
2501 mtx_lock(&sw_dev_mtx);
2502 TAILQ_FOREACH(sp, &swtailq, sw_list) {
2507 xs->xsw_version = XSWDEV_VERSION;
2508 xs->xsw_dev = sp->sw_dev;
2509 xs->xsw_flags = sp->sw_flags;
2510 xs->xsw_nblks = sp->sw_nblks;
2511 xs->xsw_used = sp->sw_used;
2512 if (devname != NULL) {
2513 if (vn_isdisk(sp->sw_vp, NULL))
2514 tmp_devname = devtoname(sp->sw_vp->v_rdev);
2516 tmp_devname = "[file]";
2517 strncpy(devname, tmp_devname, len);
2522 mtx_unlock(&sw_dev_mtx);
2526 #if defined(COMPAT_FREEBSD11)
2527 #define XSWDEV_VERSION_11 1
2537 #if defined(__amd64__) && defined(COMPAT_FREEBSD32)
2540 u_int xsw_dev1, xsw_dev2;
2548 sysctl_vm_swap_info(SYSCTL_HANDLER_ARGS)
2551 #if defined(__amd64__) && defined(COMPAT_FREEBSD32)
2552 struct xswdev32 xs32;
2554 #if defined(COMPAT_FREEBSD11)
2555 struct xswdev11 xs11;
2559 if (arg2 != 1) /* name length */
2561 error = swap_dev_info(*(int *)arg1, &xs, NULL, 0);
2564 #if defined(__amd64__) && defined(COMPAT_FREEBSD32)
2565 if (req->oldlen == sizeof(xs32)) {
2566 xs32.xsw_version = XSWDEV_VERSION;
2567 xs32.xsw_dev1 = xs.xsw_dev;
2568 xs32.xsw_dev2 = xs.xsw_dev >> 32;
2569 xs32.xsw_flags = xs.xsw_flags;
2570 xs32.xsw_nblks = xs.xsw_nblks;
2571 xs32.xsw_used = xs.xsw_used;
2572 error = SYSCTL_OUT(req, &xs32, sizeof(xs32));
2576 #if defined(COMPAT_FREEBSD11)
2577 if (req->oldlen == sizeof(xs11)) {
2578 xs11.xsw_version = XSWDEV_VERSION_11;
2579 xs11.xsw_dev = xs.xsw_dev; /* truncation */
2580 xs11.xsw_flags = xs.xsw_flags;
2581 xs11.xsw_nblks = xs.xsw_nblks;
2582 xs11.xsw_used = xs.xsw_used;
2583 error = SYSCTL_OUT(req, &xs11, sizeof(xs11));
2587 error = SYSCTL_OUT(req, &xs, sizeof(xs));
2591 SYSCTL_INT(_vm, OID_AUTO, nswapdev, CTLFLAG_RD, &nswapdev, 0,
2592 "Number of swap devices");
2593 SYSCTL_NODE(_vm, OID_AUTO, swap_info, CTLFLAG_RD | CTLFLAG_MPSAFE,
2594 sysctl_vm_swap_info,
2595 "Swap statistics by device");
2598 * Count the approximate swap usage in pages for a vmspace. The
2599 * shadowed or not yet copied on write swap blocks are not accounted.
2600 * The map must be locked.
2603 vmspace_swap_count(struct vmspace *vmspace)
2613 map = &vmspace->vm_map;
2616 for (cur = map->header.next; cur != &map->header; cur = cur->next) {
2617 if ((cur->eflags & MAP_ENTRY_IS_SUB_MAP) != 0)
2619 object = cur->object.vm_object;
2620 if (object == NULL || object->type != OBJT_SWAP)
2622 VM_OBJECT_RLOCK(object);
2623 if (object->type != OBJT_SWAP)
2625 pi = OFF_TO_IDX(cur->offset);
2626 e = pi + OFF_TO_IDX(cur->end - cur->start);
2627 for (;; pi = sb->p + SWAP_META_PAGES) {
2628 sb = SWAP_PCTRIE_LOOKUP_GE(
2629 &object->un_pager.swp.swp_blks, pi);
2630 if (sb == NULL || sb->p >= e)
2632 for (i = 0; i < SWAP_META_PAGES; i++) {
2633 if (sb->p + i < e &&
2634 sb->d[i] != SWAPBLK_NONE)
2639 VM_OBJECT_RUNLOCK(object);
2647 * Swapping onto disk devices.
2651 static g_orphan_t swapgeom_orphan;
2653 static struct g_class g_swap_class = {
2655 .version = G_VERSION,
2656 .orphan = swapgeom_orphan,
2659 DECLARE_GEOM_CLASS(g_swap_class, g_class);
2663 swapgeom_close_ev(void *arg, int flags)
2665 struct g_consumer *cp;
2668 g_access(cp, -1, -1, 0);
2670 g_destroy_consumer(cp);
2674 * Add a reference to the g_consumer for an inflight transaction.
2677 swapgeom_acquire(struct g_consumer *cp)
2680 mtx_assert(&sw_dev_mtx, MA_OWNED);
2685 * Remove a reference from the g_consumer. Post a close event if all
2686 * references go away, since the function might be called from the
2690 swapgeom_release(struct g_consumer *cp, struct swdevt *sp)
2693 mtx_assert(&sw_dev_mtx, MA_OWNED);
2695 if (cp->index == 0) {
2696 if (g_post_event(swapgeom_close_ev, cp, M_NOWAIT, NULL) == 0)
2702 swapgeom_done(struct bio *bp2)
2706 struct g_consumer *cp;
2708 bp = bp2->bio_caller2;
2710 bp->b_ioflags = bp2->bio_flags;
2712 bp->b_ioflags |= BIO_ERROR;
2713 bp->b_resid = bp->b_bcount - bp2->bio_completed;
2714 bp->b_error = bp2->bio_error;
2715 bp->b_caller1 = NULL;
2717 sp = bp2->bio_caller1;
2718 mtx_lock(&sw_dev_mtx);
2719 swapgeom_release(cp, sp);
2720 mtx_unlock(&sw_dev_mtx);
2725 swapgeom_strategy(struct buf *bp, struct swdevt *sp)
2728 struct g_consumer *cp;
2730 mtx_lock(&sw_dev_mtx);
2733 mtx_unlock(&sw_dev_mtx);
2734 bp->b_error = ENXIO;
2735 bp->b_ioflags |= BIO_ERROR;
2739 swapgeom_acquire(cp);
2740 mtx_unlock(&sw_dev_mtx);
2741 if (bp->b_iocmd == BIO_WRITE)
2744 bio = g_alloc_bio();
2746 mtx_lock(&sw_dev_mtx);
2747 swapgeom_release(cp, sp);
2748 mtx_unlock(&sw_dev_mtx);
2749 bp->b_error = ENOMEM;
2750 bp->b_ioflags |= BIO_ERROR;
2751 printf("swap_pager: cannot allocate bio\n");
2756 bp->b_caller1 = bio;
2757 bio->bio_caller1 = sp;
2758 bio->bio_caller2 = bp;
2759 bio->bio_cmd = bp->b_iocmd;
2760 bio->bio_offset = (bp->b_blkno - sp->sw_first) * PAGE_SIZE;
2761 bio->bio_length = bp->b_bcount;
2762 bio->bio_done = swapgeom_done;
2763 if (!buf_mapped(bp)) {
2764 bio->bio_ma = bp->b_pages;
2765 bio->bio_data = unmapped_buf;
2766 bio->bio_ma_offset = (vm_offset_t)bp->b_offset & PAGE_MASK;
2767 bio->bio_ma_n = bp->b_npages;
2768 bio->bio_flags |= BIO_UNMAPPED;
2770 bio->bio_data = bp->b_data;
2773 g_io_request(bio, cp);
2778 swapgeom_orphan(struct g_consumer *cp)
2783 mtx_lock(&sw_dev_mtx);
2784 TAILQ_FOREACH(sp, &swtailq, sw_list) {
2785 if (sp->sw_id == cp) {
2786 sp->sw_flags |= SW_CLOSING;
2791 * Drop reference we were created with. Do directly since we're in a
2792 * special context where we don't have to queue the call to
2793 * swapgeom_close_ev().
2796 destroy = ((sp != NULL) && (cp->index == 0));
2799 mtx_unlock(&sw_dev_mtx);
2801 swapgeom_close_ev(cp, 0);
2805 swapgeom_close(struct thread *td, struct swdevt *sw)
2807 struct g_consumer *cp;
2809 mtx_lock(&sw_dev_mtx);
2812 mtx_unlock(&sw_dev_mtx);
2815 * swapgeom_close() may be called from the biodone context,
2816 * where we cannot perform topology changes. Delegate the
2817 * work to the events thread.
2820 g_waitfor_event(swapgeom_close_ev, cp, M_WAITOK, NULL);
2824 swapongeom_locked(struct cdev *dev, struct vnode *vp)
2826 struct g_provider *pp;
2827 struct g_consumer *cp;
2828 static struct g_geom *gp;
2833 pp = g_dev_getprovider(dev);
2836 mtx_lock(&sw_dev_mtx);
2837 TAILQ_FOREACH(sp, &swtailq, sw_list) {
2839 if (cp != NULL && cp->provider == pp) {
2840 mtx_unlock(&sw_dev_mtx);
2844 mtx_unlock(&sw_dev_mtx);
2846 gp = g_new_geomf(&g_swap_class, "swap");
2847 cp = g_new_consumer(gp);
2848 cp->index = 1; /* Number of active I/Os, plus one for being active. */
2849 cp->flags |= G_CF_DIRECT_SEND | G_CF_DIRECT_RECEIVE;
2852 * XXX: Every time you think you can improve the margin for
2853 * footshooting, somebody depends on the ability to do so:
2854 * savecore(8) wants to write to our swapdev so we cannot
2855 * set an exclusive count :-(
2857 error = g_access(cp, 1, 1, 0);
2860 g_destroy_consumer(cp);
2863 nblks = pp->mediasize / DEV_BSIZE;
2864 swaponsomething(vp, cp, nblks, swapgeom_strategy,
2865 swapgeom_close, dev2udev(dev),
2866 (pp->flags & G_PF_ACCEPT_UNMAPPED) != 0 ? SW_UNMAPPED : 0);
2871 swapongeom(struct vnode *vp)
2875 vn_lock(vp, LK_EXCLUSIVE | LK_RETRY);
2876 if (vp->v_type != VCHR || (vp->v_iflag & VI_DOOMED) != 0) {
2880 error = swapongeom_locked(vp->v_rdev, vp);
2881 g_topology_unlock();
2890 * This is used mainly for network filesystem (read: probably only tested
2891 * with NFS) swapfiles.
2896 swapdev_strategy(struct buf *bp, struct swdevt *sp)
2900 bp->b_blkno = ctodb(bp->b_blkno - sp->sw_first);
2904 if (bp->b_iocmd == BIO_WRITE) {
2906 bufobj_wdrop(bp->b_bufobj);
2907 bufobj_wref(&vp2->v_bufobj);
2909 if (bp->b_bufobj != &vp2->v_bufobj)
2910 bp->b_bufobj = &vp2->v_bufobj;
2912 bp->b_iooffset = dbtob(bp->b_blkno);
2918 swapdev_close(struct thread *td, struct swdevt *sp)
2921 VOP_CLOSE(sp->sw_vp, FREAD | FWRITE, td->td_ucred, td);
2927 swaponvp(struct thread *td, struct vnode *vp, u_long nblks)
2934 mtx_lock(&sw_dev_mtx);
2935 TAILQ_FOREACH(sp, &swtailq, sw_list) {
2936 if (sp->sw_id == vp) {
2937 mtx_unlock(&sw_dev_mtx);
2941 mtx_unlock(&sw_dev_mtx);
2943 (void) vn_lock(vp, LK_EXCLUSIVE | LK_RETRY);
2945 error = mac_system_check_swapon(td->td_ucred, vp);
2948 error = VOP_OPEN(vp, FREAD | FWRITE, td->td_ucred, td, NULL);
2949 (void) VOP_UNLOCK(vp, 0);
2953 swaponsomething(vp, vp, nblks, swapdev_strategy, swapdev_close,
2959 sysctl_swap_async_max(SYSCTL_HANDLER_ARGS)
2963 new = nsw_wcount_async_max;
2964 error = sysctl_handle_int(oidp, &new, 0, req);
2965 if (error != 0 || req->newptr == NULL)
2968 if (new > nswbuf / 2 || new < 1)
2971 mtx_lock(&swbuf_mtx);
2972 while (nsw_wcount_async_max != new) {
2974 * Adjust difference. If the current async count is too low,
2975 * we will need to sqeeze our update slowly in. Sleep with a
2976 * higher priority than getpbuf() to finish faster.
2978 n = new - nsw_wcount_async_max;
2979 if (nsw_wcount_async + n >= 0) {
2980 nsw_wcount_async += n;
2981 nsw_wcount_async_max += n;
2982 wakeup(&nsw_wcount_async);
2984 nsw_wcount_async_max -= nsw_wcount_async;
2985 nsw_wcount_async = 0;
2986 msleep(&nsw_wcount_async, &swbuf_mtx, PSWP,
2990 mtx_unlock(&swbuf_mtx);