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);
379 static void swap_pager_update_writecount(vm_object_t object,
380 vm_offset_t start, vm_offset_t end);
381 static void swap_pager_release_writecount(vm_object_t object,
382 vm_offset_t start, vm_offset_t end);
384 struct pagerops swappagerops = {
385 .pgo_init = swap_pager_init, /* early system initialization of pager */
386 .pgo_alloc = swap_pager_alloc, /* allocate an OBJT_SWAP object */
387 .pgo_dealloc = swap_pager_dealloc, /* deallocate an OBJT_SWAP object */
388 .pgo_getpages = swap_pager_getpages, /* pagein */
389 .pgo_getpages_async = swap_pager_getpages_async, /* pagein (async) */
390 .pgo_putpages = swap_pager_putpages, /* pageout */
391 .pgo_haspage = swap_pager_haspage, /* get backing store status for page */
392 .pgo_pageunswapped = swap_pager_unswapped, /* remove swap related to page */
393 .pgo_update_writecount = swap_pager_update_writecount,
394 .pgo_release_writecount = swap_pager_release_writecount,
398 * swap_*() routines are externally accessible. swp_*() routines are
401 static int nswap_lowat = 128; /* in pages, swap_pager_almost_full warn */
402 static int nswap_hiwat = 512; /* in pages, swap_pager_almost_full warn */
404 SYSCTL_INT(_vm, OID_AUTO, dmmax, CTLFLAG_RD, &nsw_cluster_max, 0,
405 "Maximum size of a swap block in pages");
407 static void swp_sizecheck(void);
408 static void swp_pager_async_iodone(struct buf *bp);
409 static bool swp_pager_swblk_empty(struct swblk *sb, int start, int limit);
410 static int swapongeom(struct vnode *);
411 static int swaponvp(struct thread *, struct vnode *, u_long);
412 static int swapoff_one(struct swdevt *sp, struct ucred *cred);
415 * Swap bitmap functions
417 static void swp_pager_freeswapspace(daddr_t blk, daddr_t npages);
418 static daddr_t swp_pager_getswapspace(int *npages, int limit);
423 static daddr_t swp_pager_meta_build(vm_object_t, vm_pindex_t, daddr_t);
424 static void swp_pager_meta_free(vm_object_t, vm_pindex_t, vm_pindex_t);
425 static void swp_pager_meta_free_all(vm_object_t);
426 static daddr_t swp_pager_meta_ctl(vm_object_t, vm_pindex_t, int);
429 swp_pager_init_freerange(daddr_t *start, daddr_t *num)
432 *start = SWAPBLK_NONE;
437 swp_pager_update_freerange(daddr_t *start, daddr_t *num, daddr_t addr)
440 if (*start + *num == addr) {
443 swp_pager_freeswapspace(*start, *num);
450 swblk_trie_alloc(struct pctrie *ptree)
453 return (uma_zalloc(swpctrie_zone, M_NOWAIT | (curproc == pageproc ?
454 M_USE_RESERVE : 0)));
458 swblk_trie_free(struct pctrie *ptree, void *node)
461 uma_zfree(swpctrie_zone, node);
464 PCTRIE_DEFINE(SWAP, swblk, p, swblk_trie_alloc, swblk_trie_free);
467 * SWP_SIZECHECK() - update swap_pager_full indication
469 * update the swap_pager_almost_full indication and warn when we are
470 * about to run out of swap space, using lowat/hiwat hysteresis.
472 * Clear swap_pager_full ( task killing ) indication when lowat is met.
474 * No restrictions on call
475 * This routine may not block.
481 if (swap_pager_avail < nswap_lowat) {
482 if (swap_pager_almost_full == 0) {
483 printf("swap_pager: out of swap space\n");
484 swap_pager_almost_full = 1;
488 if (swap_pager_avail > nswap_hiwat)
489 swap_pager_almost_full = 0;
494 * SWAP_PAGER_INIT() - initialize the swap pager!
496 * Expected to be started from system init. NOTE: This code is run
497 * before much else so be careful what you depend on. Most of the VM
498 * system has yet to be initialized at this point.
501 swap_pager_init(void)
504 * Initialize object lists
508 for (i = 0; i < NOBJLISTS; ++i)
509 TAILQ_INIT(&swap_pager_object_list[i]);
510 mtx_init(&sw_dev_mtx, "swapdev", NULL, MTX_DEF);
511 sx_init(&sw_alloc_sx, "swspsx");
512 sx_init(&swdev_syscall_lock, "swsysc");
516 * SWAP_PAGER_SWAP_INIT() - swap pager initialization from pageout process
518 * Expected to be started from pageout process once, prior to entering
522 swap_pager_swap_init(void)
527 * Number of in-transit swap bp operations. Don't
528 * exhaust the pbufs completely. Make sure we
529 * initialize workable values (0 will work for hysteresis
530 * but it isn't very efficient).
532 * The nsw_cluster_max is constrained by the bp->b_pages[]
533 * array, which has MAXPHYS / PAGE_SIZE entries, and our locally
534 * defined MAX_PAGEOUT_CLUSTER. Also be aware that swap ops are
535 * constrained by the swap device interleave stripe size.
537 * Currently we hardwire nsw_wcount_async to 4. This limit is
538 * designed to prevent other I/O from having high latencies due to
539 * our pageout I/O. The value 4 works well for one or two active swap
540 * devices but is probably a little low if you have more. Even so,
541 * a higher value would probably generate only a limited improvement
542 * with three or four active swap devices since the system does not
543 * typically have to pageout at extreme bandwidths. We will want
544 * at least 2 per swap devices, and 4 is a pretty good value if you
545 * have one NFS swap device due to the command/ack latency over NFS.
546 * So it all works out pretty well.
548 nsw_cluster_max = min(MAXPHYS / PAGE_SIZE, MAX_PAGEOUT_CLUSTER);
550 nsw_wcount_async = 4;
551 nsw_wcount_async_max = nsw_wcount_async;
552 mtx_init(&swbuf_mtx, "async swbuf mutex", NULL, MTX_DEF);
554 swwbuf_zone = pbuf_zsecond_create("swwbuf", nswbuf / 4);
555 swrbuf_zone = pbuf_zsecond_create("swrbuf", nswbuf / 2);
558 * Initialize our zone, taking the user's requested size or
559 * estimating the number we need based on the number of pages
562 n = maxswzone != 0 ? maxswzone / sizeof(struct swblk) :
563 vm_cnt.v_page_count / 2;
564 swpctrie_zone = uma_zcreate("swpctrie", pctrie_node_size(), NULL, NULL,
565 pctrie_zone_init, NULL, UMA_ALIGN_PTR, UMA_ZONE_VM);
566 if (swpctrie_zone == NULL)
567 panic("failed to create swap pctrie zone.");
568 swblk_zone = uma_zcreate("swblk", sizeof(struct swblk), NULL, NULL,
569 NULL, NULL, _Alignof(struct swblk) - 1, UMA_ZONE_VM);
570 if (swblk_zone == NULL)
571 panic("failed to create swap blk zone.");
574 if (uma_zone_reserve_kva(swblk_zone, n))
577 * if the allocation failed, try a zone two thirds the
578 * size of the previous attempt.
584 * Often uma_zone_reserve_kva() cannot reserve exactly the
585 * requested size. Account for the difference when
586 * calculating swap_maxpages.
588 n = uma_zone_get_max(swblk_zone);
591 printf("Swap blk zone entries changed from %lu to %lu.\n",
593 swap_maxpages = n * SWAP_META_PAGES;
594 swzone = n * sizeof(struct swblk);
595 if (!uma_zone_reserve_kva(swpctrie_zone, n))
596 printf("Cannot reserve swap pctrie zone, "
597 "reduce kern.maxswzone.\n");
601 swap_pager_alloc_init(void *handle, struct ucred *cred, vm_ooffset_t size,
607 if (!swap_reserve_by_cred(size, cred))
613 * The un_pager.swp.swp_blks trie is initialized by
614 * vm_object_allocate() to ensure the correct order of
615 * visibility to other threads.
617 object = vm_object_allocate(OBJT_SWAP, OFF_TO_IDX(offset +
620 object->un_pager.swp.writemappings = 0;
621 object->handle = handle;
624 object->charge = size;
630 * SWAP_PAGER_ALLOC() - allocate a new OBJT_SWAP VM object and instantiate
631 * its metadata structures.
633 * This routine is called from the mmap and fork code to create a new
636 * This routine must ensure that no live duplicate is created for
637 * the named object request, which is protected against by
638 * holding the sw_alloc_sx lock in case handle != NULL.
641 swap_pager_alloc(void *handle, vm_ooffset_t size, vm_prot_t prot,
642 vm_ooffset_t offset, struct ucred *cred)
646 if (handle != NULL) {
648 * Reference existing named region or allocate new one. There
649 * should not be a race here against swp_pager_meta_build()
650 * as called from vm_page_remove() in regards to the lookup
653 sx_xlock(&sw_alloc_sx);
654 object = vm_pager_object_lookup(NOBJLIST(handle), handle);
655 if (object == NULL) {
656 object = swap_pager_alloc_init(handle, cred, size,
658 if (object != NULL) {
659 TAILQ_INSERT_TAIL(NOBJLIST(object->handle),
660 object, pager_object_list);
663 sx_xunlock(&sw_alloc_sx);
665 object = swap_pager_alloc_init(handle, cred, size, offset);
671 * SWAP_PAGER_DEALLOC() - remove swap metadata from object
673 * The swap backing for the object is destroyed. The code is
674 * designed such that we can reinstantiate it later, but this
675 * routine is typically called only when the entire object is
676 * about to be destroyed.
678 * The object must be locked.
681 swap_pager_dealloc(vm_object_t object)
684 VM_OBJECT_ASSERT_WLOCKED(object);
685 KASSERT((object->flags & OBJ_DEAD) != 0, ("dealloc of reachable obj"));
688 * Remove from list right away so lookups will fail if we block for
689 * pageout completion.
691 if (object->handle != NULL) {
692 VM_OBJECT_WUNLOCK(object);
693 sx_xlock(&sw_alloc_sx);
694 TAILQ_REMOVE(NOBJLIST(object->handle), object,
696 sx_xunlock(&sw_alloc_sx);
697 VM_OBJECT_WLOCK(object);
700 vm_object_pip_wait(object, "swpdea");
703 * Free all remaining metadata. We only bother to free it from
704 * the swap meta data. We do not attempt to free swapblk's still
705 * associated with vm_page_t's for this object. We do not care
706 * if paging is still in progress on some objects.
708 swp_pager_meta_free_all(object);
709 object->handle = NULL;
710 object->type = OBJT_DEAD;
713 /************************************************************************
714 * SWAP PAGER BITMAP ROUTINES *
715 ************************************************************************/
718 * SWP_PAGER_GETSWAPSPACE() - allocate raw swap space
720 * Allocate swap for up to the requested number of pages, and at
721 * least a minimum number of pages. The starting swap block number
722 * (a page index) is returned or SWAPBLK_NONE if the allocation
725 * Also has the side effect of advising that somebody made a mistake
726 * when they configured swap and didn't configure enough.
728 * This routine may not sleep.
730 * We allocate in round-robin fashion from the configured devices.
733 swp_pager_getswapspace(int *io_npages, int limit)
741 npages = imin(BLIST_MAX_ALLOC, mpages);
742 mtx_lock(&sw_dev_mtx);
744 while (!TAILQ_EMPTY(&swtailq)) {
746 sp = TAILQ_FIRST(&swtailq);
747 if ((sp->sw_flags & SW_CLOSING) == 0)
748 blk = blist_alloc(sp->sw_blist, &npages, mpages);
749 if (blk != SWAPBLK_NONE)
751 sp = TAILQ_NEXT(sp, sw_list);
759 if (blk != SWAPBLK_NONE) {
762 sp->sw_used += npages;
763 swap_pager_avail -= npages;
765 swdevhd = TAILQ_NEXT(sp, sw_list);
767 if (swap_pager_full != 2) {
768 printf("swp_pager_getswapspace(%d): failed\n",
771 swap_pager_almost_full = 1;
775 mtx_unlock(&sw_dev_mtx);
780 swp_pager_isondev(daddr_t blk, struct swdevt *sp)
783 return (blk >= sp->sw_first && blk < sp->sw_end);
787 swp_pager_strategy(struct buf *bp)
791 mtx_lock(&sw_dev_mtx);
792 TAILQ_FOREACH(sp, &swtailq, sw_list) {
793 if (swp_pager_isondev(bp->b_blkno, sp)) {
794 mtx_unlock(&sw_dev_mtx);
795 if ((sp->sw_flags & SW_UNMAPPED) != 0 &&
796 unmapped_buf_allowed) {
797 bp->b_data = unmapped_buf;
800 pmap_qenter((vm_offset_t)bp->b_data,
801 &bp->b_pages[0], bp->b_bcount / PAGE_SIZE);
803 sp->sw_strategy(bp, sp);
807 panic("Swapdev not found");
812 * SWP_PAGER_FREESWAPSPACE() - free raw swap space
814 * This routine returns the specified swap blocks back to the bitmap.
816 * This routine may not sleep.
819 swp_pager_freeswapspace(daddr_t blk, daddr_t npages)
825 mtx_lock(&sw_dev_mtx);
826 TAILQ_FOREACH(sp, &swtailq, sw_list) {
827 if (swp_pager_isondev(blk, sp)) {
828 sp->sw_used -= npages;
830 * If we are attempting to stop swapping on
831 * this device, we don't want to mark any
832 * blocks free lest they be reused.
834 if ((sp->sw_flags & SW_CLOSING) == 0) {
835 blist_free(sp->sw_blist, blk - sp->sw_first,
837 swap_pager_avail += npages;
840 mtx_unlock(&sw_dev_mtx);
844 panic("Swapdev not found");
848 * SYSCTL_SWAP_FRAGMENTATION() - produce raw swap space stats
851 sysctl_swap_fragmentation(SYSCTL_HANDLER_ARGS)
858 error = sysctl_wire_old_buffer(req, 0);
861 sbuf_new_for_sysctl(&sbuf, NULL, 128, req);
862 mtx_lock(&sw_dev_mtx);
863 TAILQ_FOREACH(sp, &swtailq, sw_list) {
864 if (vn_isdisk(sp->sw_vp, NULL))
865 devname = devtoname(sp->sw_vp->v_rdev);
868 sbuf_printf(&sbuf, "\nFree space on device %s:\n", devname);
869 blist_stats(sp->sw_blist, &sbuf);
871 mtx_unlock(&sw_dev_mtx);
872 error = sbuf_finish(&sbuf);
878 * SWAP_PAGER_FREESPACE() - frees swap blocks associated with a page
879 * range within an object.
881 * This is a globally accessible routine.
883 * This routine removes swapblk assignments from swap metadata.
885 * The external callers of this routine typically have already destroyed
886 * or renamed vm_page_t's associated with this range in the object so
889 * The object must be locked.
892 swap_pager_freespace(vm_object_t object, vm_pindex_t start, vm_size_t size)
895 swp_pager_meta_free(object, start, size);
899 * SWAP_PAGER_RESERVE() - reserve swap blocks in object
901 * Assigns swap blocks to the specified range within the object. The
902 * swap blocks are not zeroed. Any previous swap assignment is destroyed.
904 * Returns 0 on success, -1 on failure.
907 swap_pager_reserve(vm_object_t object, vm_pindex_t start, vm_size_t size)
909 daddr_t addr, blk, n_free, s_free;
912 swp_pager_init_freerange(&s_free, &n_free);
913 VM_OBJECT_WLOCK(object);
914 for (i = 0; i < size; i += n) {
916 blk = swp_pager_getswapspace(&n, 1);
917 if (blk == SWAPBLK_NONE) {
918 swp_pager_meta_free(object, start, i);
919 VM_OBJECT_WUNLOCK(object);
922 for (j = 0; j < n; ++j) {
923 addr = swp_pager_meta_build(object,
924 start + i + j, blk + j);
925 if (addr != SWAPBLK_NONE)
926 swp_pager_update_freerange(&s_free, &n_free,
930 swp_pager_freeswapspace(s_free, n_free);
931 VM_OBJECT_WUNLOCK(object);
936 * SWAP_PAGER_COPY() - copy blocks from source pager to destination pager
937 * and destroy the source.
939 * Copy any valid swapblks from the source to the destination. In
940 * cases where both the source and destination have a valid swapblk,
941 * we keep the destination's.
943 * This routine is allowed to sleep. It may sleep allocating metadata
944 * indirectly through swp_pager_meta_build() or if paging is still in
945 * progress on the source.
947 * The source object contains no vm_page_t's (which is just as well)
949 * The source object is of type OBJT_SWAP.
951 * The source and destination objects must be locked.
952 * Both object locks may temporarily be released.
955 swap_pager_copy(vm_object_t srcobject, vm_object_t dstobject,
956 vm_pindex_t offset, int destroysource)
959 daddr_t dstaddr, n_free, s_free, srcaddr;
961 VM_OBJECT_ASSERT_WLOCKED(srcobject);
962 VM_OBJECT_ASSERT_WLOCKED(dstobject);
965 * If destroysource is set, we remove the source object from the
966 * swap_pager internal queue now.
968 if (destroysource && srcobject->handle != NULL) {
969 vm_object_pip_add(srcobject, 1);
970 VM_OBJECT_WUNLOCK(srcobject);
971 vm_object_pip_add(dstobject, 1);
972 VM_OBJECT_WUNLOCK(dstobject);
973 sx_xlock(&sw_alloc_sx);
974 TAILQ_REMOVE(NOBJLIST(srcobject->handle), srcobject,
976 sx_xunlock(&sw_alloc_sx);
977 VM_OBJECT_WLOCK(dstobject);
978 vm_object_pip_wakeup(dstobject);
979 VM_OBJECT_WLOCK(srcobject);
980 vm_object_pip_wakeup(srcobject);
984 * Transfer source to destination.
986 swp_pager_init_freerange(&s_free, &n_free);
987 for (i = 0; i < dstobject->size; ++i) {
988 srcaddr = swp_pager_meta_ctl(srcobject, i + offset, SWM_POP);
989 if (srcaddr == SWAPBLK_NONE)
991 dstaddr = swp_pager_meta_ctl(dstobject, i, 0);
992 if (dstaddr != SWAPBLK_NONE) {
994 * Destination has valid swapblk or it is represented
995 * by a resident page. We destroy the source block.
997 swp_pager_update_freerange(&s_free, &n_free, srcaddr);
1002 * Destination has no swapblk and is not resident,
1005 * swp_pager_meta_build() can sleep.
1007 vm_object_pip_add(srcobject, 1);
1008 VM_OBJECT_WUNLOCK(srcobject);
1009 vm_object_pip_add(dstobject, 1);
1010 dstaddr = swp_pager_meta_build(dstobject, i, srcaddr);
1011 KASSERT(dstaddr == SWAPBLK_NONE,
1012 ("Unexpected destination swapblk"));
1013 vm_object_pip_wakeup(dstobject);
1014 VM_OBJECT_WLOCK(srcobject);
1015 vm_object_pip_wakeup(srcobject);
1017 swp_pager_freeswapspace(s_free, n_free);
1020 * Free left over swap blocks in source.
1022 * We have to revert the type to OBJT_DEFAULT so we do not accidentally
1023 * double-remove the object from the swap queues.
1025 if (destroysource) {
1026 swp_pager_meta_free_all(srcobject);
1028 * Reverting the type is not necessary, the caller is going
1029 * to destroy srcobject directly, but I'm doing it here
1030 * for consistency since we've removed the object from its
1033 srcobject->type = OBJT_DEFAULT;
1038 * SWAP_PAGER_HASPAGE() - determine if we have good backing store for
1039 * the requested page.
1041 * We determine whether good backing store exists for the requested
1042 * page and return TRUE if it does, FALSE if it doesn't.
1044 * If TRUE, we also try to determine how much valid, contiguous backing
1045 * store exists before and after the requested page.
1048 swap_pager_haspage(vm_object_t object, vm_pindex_t pindex, int *before,
1054 VM_OBJECT_ASSERT_LOCKED(object);
1057 * do we have good backing store at the requested index ?
1059 blk0 = swp_pager_meta_ctl(object, pindex, 0);
1060 if (blk0 == SWAPBLK_NONE) {
1069 * find backwards-looking contiguous good backing store
1071 if (before != NULL) {
1072 for (i = 1; i < SWB_NPAGES; i++) {
1075 blk = swp_pager_meta_ctl(object, pindex - i, 0);
1076 if (blk != blk0 - i)
1083 * find forward-looking contiguous good backing store
1085 if (after != NULL) {
1086 for (i = 1; i < SWB_NPAGES; i++) {
1087 blk = swp_pager_meta_ctl(object, pindex + i, 0);
1088 if (blk != blk0 + i)
1097 * SWAP_PAGER_PAGE_UNSWAPPED() - remove swap backing store related to page
1099 * This removes any associated swap backing store, whether valid or
1100 * not, from the page.
1102 * This routine is typically called when a page is made dirty, at
1103 * which point any associated swap can be freed. MADV_FREE also
1104 * calls us in a special-case situation
1106 * NOTE!!! If the page is clean and the swap was valid, the caller
1107 * should make the page dirty before calling this routine. This routine
1108 * does NOT change the m->dirty status of the page. Also: MADV_FREE
1111 * This routine may not sleep.
1113 * The object containing the page must be locked.
1116 swap_pager_unswapped(vm_page_t m)
1120 srcaddr = swp_pager_meta_ctl(m->object, m->pindex, SWM_POP);
1121 if (srcaddr != SWAPBLK_NONE)
1122 swp_pager_freeswapspace(srcaddr, 1);
1126 * swap_pager_getpages() - bring pages in from swap
1128 * Attempt to page in the pages in array "ma" of length "count". The
1129 * caller may optionally specify that additional pages preceding and
1130 * succeeding the specified range be paged in. The number of such pages
1131 * is returned in the "rbehind" and "rahead" parameters, and they will
1132 * be in the inactive queue upon return.
1134 * The pages in "ma" must be busied and will remain busied upon return.
1137 swap_pager_getpages(vm_object_t object, vm_page_t *ma, int count, int *rbehind,
1141 vm_page_t bm, mpred, msucc, p;
1144 int i, maxahead, maxbehind, reqcount;
1149 * Determine the final number of read-behind pages and
1150 * allocate them BEFORE releasing the object lock. Otherwise,
1151 * there can be a problematic race with vm_object_split().
1152 * Specifically, vm_object_split() might first transfer pages
1153 * that precede ma[0] in the current object to a new object,
1154 * and then this function incorrectly recreates those pages as
1155 * read-behind pages in the current object.
1157 if (!swap_pager_haspage(object, ma[0]->pindex, &maxbehind, &maxahead))
1158 return (VM_PAGER_FAIL);
1161 * Clip the readahead and readbehind ranges to exclude resident pages.
1163 if (rahead != NULL) {
1164 KASSERT(reqcount - 1 <= maxahead,
1165 ("page count %d extends beyond swap block", reqcount));
1166 *rahead = imin(*rahead, maxahead - (reqcount - 1));
1167 pindex = ma[reqcount - 1]->pindex;
1168 msucc = TAILQ_NEXT(ma[reqcount - 1], listq);
1169 if (msucc != NULL && msucc->pindex - pindex - 1 < *rahead)
1170 *rahead = msucc->pindex - pindex - 1;
1172 if (rbehind != NULL) {
1173 *rbehind = imin(*rbehind, maxbehind);
1174 pindex = ma[0]->pindex;
1175 mpred = TAILQ_PREV(ma[0], pglist, listq);
1176 if (mpred != NULL && pindex - mpred->pindex - 1 < *rbehind)
1177 *rbehind = pindex - mpred->pindex - 1;
1181 for (i = 0; i < count; i++)
1182 ma[i]->oflags |= VPO_SWAPINPROG;
1185 * Allocate readahead and readbehind pages.
1187 if (rbehind != NULL) {
1188 for (i = 1; i <= *rbehind; i++) {
1189 p = vm_page_alloc(object, ma[0]->pindex - i,
1193 p->oflags |= VPO_SWAPINPROG;
1198 if (rahead != NULL) {
1199 for (i = 0; i < *rahead; i++) {
1200 p = vm_page_alloc(object,
1201 ma[reqcount - 1]->pindex + i + 1, VM_ALLOC_NORMAL);
1204 p->oflags |= VPO_SWAPINPROG;
1208 if (rbehind != NULL)
1213 vm_object_pip_add(object, count);
1215 pindex = bm->pindex;
1216 blk = swp_pager_meta_ctl(object, pindex, 0);
1217 KASSERT(blk != SWAPBLK_NONE,
1218 ("no swap blocking containing %p(%jx)", object, (uintmax_t)pindex));
1220 VM_OBJECT_WUNLOCK(object);
1221 bp = uma_zalloc(swrbuf_zone, M_WAITOK);
1222 /* Pages cannot leave the object while busy. */
1223 for (i = 0, p = bm; i < count; i++, p = TAILQ_NEXT(p, listq)) {
1224 MPASS(p->pindex == bm->pindex + i);
1228 bp->b_flags |= B_PAGING;
1229 bp->b_iocmd = BIO_READ;
1230 bp->b_iodone = swp_pager_async_iodone;
1231 bp->b_rcred = crhold(thread0.td_ucred);
1232 bp->b_wcred = crhold(thread0.td_ucred);
1234 bp->b_bcount = PAGE_SIZE * count;
1235 bp->b_bufsize = PAGE_SIZE * count;
1236 bp->b_npages = count;
1237 bp->b_pgbefore = rbehind != NULL ? *rbehind : 0;
1238 bp->b_pgafter = rahead != NULL ? *rahead : 0;
1240 VM_CNT_INC(v_swapin);
1241 VM_CNT_ADD(v_swappgsin, count);
1244 * perform the I/O. NOTE!!! bp cannot be considered valid after
1245 * this point because we automatically release it on completion.
1246 * Instead, we look at the one page we are interested in which we
1247 * still hold a lock on even through the I/O completion.
1249 * The other pages in our ma[] array are also released on completion,
1250 * so we cannot assume they are valid anymore either.
1252 * NOTE: b_blkno is destroyed by the call to swapdev_strategy
1255 swp_pager_strategy(bp);
1258 * Wait for the pages we want to complete. VPO_SWAPINPROG is always
1259 * cleared on completion. If an I/O error occurs, SWAPBLK_NONE
1260 * is set in the metadata for each page in the request.
1262 VM_OBJECT_WLOCK(object);
1263 while ((ma[0]->oflags & VPO_SWAPINPROG) != 0) {
1264 ma[0]->oflags |= VPO_SWAPSLEEP;
1265 VM_CNT_INC(v_intrans);
1266 if (VM_OBJECT_SLEEP(object, &object->handle, PSWP,
1267 "swread", hz * 20)) {
1269 "swap_pager: indefinite wait buffer: bufobj: %p, blkno: %jd, size: %ld\n",
1270 bp->b_bufobj, (intmax_t)bp->b_blkno, bp->b_bcount);
1275 * If we had an unrecoverable read error pages will not be valid.
1277 for (i = 0; i < reqcount; i++)
1278 if (ma[i]->valid != VM_PAGE_BITS_ALL)
1279 return (VM_PAGER_ERROR);
1281 return (VM_PAGER_OK);
1284 * A final note: in a low swap situation, we cannot deallocate swap
1285 * and mark a page dirty here because the caller is likely to mark
1286 * the page clean when we return, causing the page to possibly revert
1287 * to all-zero's later.
1292 * swap_pager_getpages_async():
1294 * Right now this is emulation of asynchronous operation on top of
1295 * swap_pager_getpages().
1298 swap_pager_getpages_async(vm_object_t object, vm_page_t *ma, int count,
1299 int *rbehind, int *rahead, pgo_getpages_iodone_t iodone, void *arg)
1303 r = swap_pager_getpages(object, ma, count, rbehind, rahead);
1304 VM_OBJECT_WUNLOCK(object);
1309 case VM_PAGER_ERROR:
1316 panic("unhandled swap_pager_getpages() error %d", r);
1318 (iodone)(arg, ma, count, error);
1319 VM_OBJECT_WLOCK(object);
1325 * swap_pager_putpages:
1327 * Assign swap (if necessary) and initiate I/O on the specified pages.
1329 * We support both OBJT_DEFAULT and OBJT_SWAP objects. DEFAULT objects
1330 * are automatically converted to SWAP objects.
1332 * In a low memory situation we may block in VOP_STRATEGY(), but the new
1333 * vm_page reservation system coupled with properly written VFS devices
1334 * should ensure that no low-memory deadlock occurs. This is an area
1337 * The parent has N vm_object_pip_add() references prior to
1338 * calling us and will remove references for rtvals[] that are
1339 * not set to VM_PAGER_PEND. We need to remove the rest on I/O
1342 * The parent has soft-busy'd the pages it passes us and will unbusy
1343 * those whose rtvals[] entry is not set to VM_PAGER_PEND on return.
1344 * We need to unbusy the rest on I/O completion.
1347 swap_pager_putpages(vm_object_t object, vm_page_t *ma, int count,
1348 int flags, int *rtvals)
1351 daddr_t addr, blk, n_free, s_free;
1356 KASSERT(count == 0 || ma[0]->object == object,
1357 ("%s: object mismatch %p/%p",
1358 __func__, object, ma[0]->object));
1363 * Turn object into OBJT_SWAP. Force sync if not a pageout process.
1365 if (object->type != OBJT_SWAP) {
1366 addr = swp_pager_meta_build(object, 0, SWAPBLK_NONE);
1367 KASSERT(addr == SWAPBLK_NONE,
1368 ("unexpected object swap block"));
1370 VM_OBJECT_WUNLOCK(object);
1371 async = curproc == pageproc && (flags & VM_PAGER_PUT_SYNC) == 0;
1372 swp_pager_init_freerange(&s_free, &n_free);
1377 * Assign swap blocks and issue I/O. We reallocate swap on the fly.
1378 * The page is left dirty until the pageout operation completes
1381 for (i = 0; i < count; i += n) {
1382 /* Maximum I/O size is limited by maximum swap block size. */
1383 n = min(count - i, nsw_cluster_max);
1385 /* Get a block of swap of size up to size n. */
1386 blk = swp_pager_getswapspace(&n, 4);
1387 if (blk == SWAPBLK_NONE) {
1388 for (j = 0; j < n; ++j)
1389 rtvals[i + j] = VM_PAGER_FAIL;
1394 * All I/O parameters have been satisfied. Build the I/O
1395 * request and assign the swap space.
1398 mtx_lock(&swbuf_mtx);
1399 while (nsw_wcount_async == 0)
1400 msleep(&nsw_wcount_async, &swbuf_mtx, PVM,
1403 mtx_unlock(&swbuf_mtx);
1405 bp = uma_zalloc(swwbuf_zone, M_WAITOK);
1407 bp->b_flags = B_ASYNC;
1408 bp->b_flags |= B_PAGING;
1409 bp->b_iocmd = BIO_WRITE;
1411 bp->b_rcred = crhold(thread0.td_ucred);
1412 bp->b_wcred = crhold(thread0.td_ucred);
1413 bp->b_bcount = PAGE_SIZE * n;
1414 bp->b_bufsize = PAGE_SIZE * n;
1417 VM_OBJECT_WLOCK(object);
1418 for (j = 0; j < n; ++j) {
1420 addr = swp_pager_meta_build(mreq->object, mreq->pindex,
1422 if (addr != SWAPBLK_NONE)
1423 swp_pager_update_freerange(&s_free, &n_free,
1425 MPASS(mreq->dirty == VM_PAGE_BITS_ALL);
1426 mreq->oflags |= VPO_SWAPINPROG;
1427 bp->b_pages[j] = mreq;
1429 VM_OBJECT_WUNLOCK(object);
1432 * Must set dirty range for NFS to work.
1435 bp->b_dirtyend = bp->b_bcount;
1437 VM_CNT_INC(v_swapout);
1438 VM_CNT_ADD(v_swappgsout, bp->b_npages);
1441 * We unconditionally set rtvals[] to VM_PAGER_PEND so that we
1442 * can call the async completion routine at the end of a
1443 * synchronous I/O operation. Otherwise, our caller would
1444 * perform duplicate unbusy and wakeup operations on the page
1445 * and object, respectively.
1447 for (j = 0; j < n; j++)
1448 rtvals[i + j] = VM_PAGER_PEND;
1453 * NOTE: b_blkno is destroyed by the call to swapdev_strategy.
1456 bp->b_iodone = swp_pager_async_iodone;
1458 swp_pager_strategy(bp);
1465 * NOTE: b_blkno is destroyed by the call to swapdev_strategy.
1467 bp->b_iodone = bdone;
1468 swp_pager_strategy(bp);
1471 * Wait for the sync I/O to complete.
1473 bwait(bp, PVM, "swwrt");
1476 * Now that we are through with the bp, we can call the
1477 * normal async completion, which frees everything up.
1479 swp_pager_async_iodone(bp);
1481 swp_pager_freeswapspace(s_free, n_free);
1482 VM_OBJECT_WLOCK(object);
1486 * swp_pager_async_iodone:
1488 * Completion routine for asynchronous reads and writes from/to swap.
1489 * Also called manually by synchronous code to finish up a bp.
1491 * This routine may not sleep.
1494 swp_pager_async_iodone(struct buf *bp)
1497 vm_object_t object = NULL;
1500 * Report error - unless we ran out of memory, in which case
1501 * we've already logged it in swapgeom_strategy().
1503 if (bp->b_ioflags & BIO_ERROR && bp->b_error != ENOMEM) {
1505 "swap_pager: I/O error - %s failed; blkno %ld,"
1506 "size %ld, error %d\n",
1507 ((bp->b_iocmd == BIO_READ) ? "pagein" : "pageout"),
1515 * remove the mapping for kernel virtual
1518 pmap_qremove((vm_offset_t)bp->b_data, bp->b_npages);
1520 bp->b_data = bp->b_kvabase;
1523 object = bp->b_pages[0]->object;
1524 VM_OBJECT_WLOCK(object);
1528 * cleanup pages. If an error occurs writing to swap, we are in
1529 * very serious trouble. If it happens to be a disk error, though,
1530 * we may be able to recover by reassigning the swap later on. So
1531 * in this case we remove the m->swapblk assignment for the page
1532 * but do not free it in the rlist. The errornous block(s) are thus
1533 * never reallocated as swap. Redirty the page and continue.
1535 for (i = 0; i < bp->b_npages; ++i) {
1536 vm_page_t m = bp->b_pages[i];
1538 m->oflags &= ~VPO_SWAPINPROG;
1539 if (m->oflags & VPO_SWAPSLEEP) {
1540 m->oflags &= ~VPO_SWAPSLEEP;
1541 wakeup(&object->handle);
1544 if (bp->b_ioflags & BIO_ERROR) {
1546 * If an error occurs I'd love to throw the swapblk
1547 * away without freeing it back to swapspace, so it
1548 * can never be used again. But I can't from an
1551 if (bp->b_iocmd == BIO_READ) {
1553 * NOTE: for reads, m->dirty will probably
1554 * be overridden by the original caller of
1555 * getpages so don't play cute tricks here.
1560 * If a write error occurs, reactivate page
1561 * so it doesn't clog the inactive list,
1562 * then finish the I/O.
1564 MPASS(m->dirty == VM_PAGE_BITS_ALL);
1566 vm_page_activate(m);
1570 } else if (bp->b_iocmd == BIO_READ) {
1572 * NOTE: for reads, m->dirty will probably be
1573 * overridden by the original caller of getpages so
1574 * we cannot set them in order to free the underlying
1575 * swap in a low-swap situation. I don't think we'd
1576 * want to do that anyway, but it was an optimization
1577 * that existed in the old swapper for a time before
1578 * it got ripped out due to precisely this problem.
1580 KASSERT(!pmap_page_is_mapped(m),
1581 ("swp_pager_async_iodone: page %p is mapped", m));
1582 KASSERT(m->dirty == 0,
1583 ("swp_pager_async_iodone: page %p is dirty", m));
1585 m->valid = VM_PAGE_BITS_ALL;
1586 if (i < bp->b_pgbefore ||
1587 i >= bp->b_npages - bp->b_pgafter)
1588 vm_page_readahead_finish(m);
1591 * For write success, clear the dirty
1592 * status, then finish the I/O ( which decrements the
1593 * busy count and possibly wakes waiter's up ).
1594 * A page is only written to swap after a period of
1595 * inactivity. Therefore, we do not expect it to be
1598 KASSERT(!pmap_page_is_write_mapped(m),
1599 ("swp_pager_async_iodone: page %p is not write"
1603 vm_page_deactivate_noreuse(m);
1610 * adjust pip. NOTE: the original parent may still have its own
1611 * pip refs on the object.
1613 if (object != NULL) {
1614 vm_object_pip_wakeupn(object, bp->b_npages);
1615 VM_OBJECT_WUNLOCK(object);
1619 * swapdev_strategy() manually sets b_vp and b_bufobj before calling
1620 * bstrategy(). Set them back to NULL now we're done with it, or we'll
1621 * trigger a KASSERT in relpbuf().
1625 bp->b_bufobj = NULL;
1628 * release the physical I/O buffer
1630 if (bp->b_flags & B_ASYNC) {
1631 mtx_lock(&swbuf_mtx);
1632 if (++nsw_wcount_async == 1)
1633 wakeup(&nsw_wcount_async);
1634 mtx_unlock(&swbuf_mtx);
1636 uma_zfree((bp->b_iocmd == BIO_READ) ? swrbuf_zone : swwbuf_zone, bp);
1640 swap_pager_nswapdev(void)
1647 swp_pager_force_dirty(vm_page_t m)
1652 swap_pager_unswapped(m);
1656 swp_pager_force_launder(vm_page_t m)
1664 swap_pager_unswapped(m);
1668 * SWP_PAGER_FORCE_PAGEIN() - force swap blocks to be paged in
1670 * This routine dissociates pages starting at the given index within an
1671 * object from their backing store, paging them in if they do not reside
1672 * in memory. Pages that are paged in are marked dirty and placed in the
1673 * laundry queue. Pages are marked dirty because they no longer have
1674 * backing store. They are placed in the laundry queue because they have
1675 * not been accessed recently. Otherwise, they would already reside in
1679 swp_pager_force_pagein(vm_object_t object, vm_pindex_t pindex, int npages)
1681 vm_page_t ma[npages];
1684 KASSERT(npages > 0, ("%s: No pages", __func__));
1685 KASSERT(npages <= MAXPHYS / PAGE_SIZE,
1686 ("%s: Too many pages: %d", __func__, npages));
1687 vm_object_pip_add(object, npages);
1688 vm_page_grab_pages(object, pindex, VM_ALLOC_NORMAL, ma, npages);
1689 for (i = j = 0;; i++) {
1690 /* Count nonresident pages, to page-in all at once. */
1691 if (i < npages && ma[i]->valid != VM_PAGE_BITS_ALL)
1694 /* Page-in nonresident pages. Mark for laundering. */
1695 if (swap_pager_getpages(object, &ma[j], i - j, NULL,
1696 NULL) != VM_PAGER_OK)
1697 panic("%s: read from swap failed", __func__);
1699 swp_pager_force_launder(ma[j]);
1704 /* Mark dirty a resident page. */
1705 swp_pager_force_dirty(ma[j++]);
1707 vm_object_pip_wakeupn(object, npages);
1711 * swap_pager_swapoff_object:
1713 * Page in all of the pages that have been paged out for an object
1717 swap_pager_swapoff_object(struct swdevt *sp, vm_object_t object)
1720 vm_pindex_t pi, s_pindex;
1721 daddr_t blk, n_blks, s_blk;
1725 for (pi = 0; (sb = SWAP_PCTRIE_LOOKUP_GE(
1726 &object->un_pager.swp.swp_blks, pi)) != NULL; ) {
1727 for (i = 0; i < SWAP_META_PAGES; i++) {
1729 if (!swp_pager_isondev(blk, sp))
1733 * If there are no blocks/pages accumulated, start a new
1734 * accumulation here.
1737 if (blk != SWAPBLK_NONE) {
1739 s_pindex = sb->p + i;
1746 * If the accumulation can be extended without breaking
1747 * the sequence of consecutive blocks and pages that
1748 * swp_pager_force_pagein() depends on, do so.
1750 if (n_blks < MAXPHYS / PAGE_SIZE &&
1751 s_blk + n_blks == blk &&
1752 s_pindex + n_blks == sb->p + i) {
1758 * The sequence of consecutive blocks and pages cannot
1759 * be extended, so page them all in here. Then,
1760 * because doing so involves releasing and reacquiring
1761 * a lock that protects the swap block pctrie, do not
1762 * rely on the current swap block. Break this loop and
1763 * re-fetch the same pindex from the pctrie again.
1765 swp_pager_force_pagein(object, s_pindex, n_blks);
1769 if (i == SWAP_META_PAGES)
1770 pi = sb->p + SWAP_META_PAGES;
1773 swp_pager_force_pagein(object, s_pindex, n_blks);
1777 * swap_pager_swapoff:
1779 * Page in all of the pages that have been paged out to the
1780 * given device. The corresponding blocks in the bitmap must be
1781 * marked as allocated and the device must be flagged SW_CLOSING.
1782 * There may be no processes swapped out to the device.
1784 * This routine may block.
1787 swap_pager_swapoff(struct swdevt *sp)
1792 sx_assert(&swdev_syscall_lock, SA_XLOCKED);
1796 mtx_lock(&vm_object_list_mtx);
1797 TAILQ_FOREACH(object, &vm_object_list, object_list) {
1798 if (object->type != OBJT_SWAP)
1800 mtx_unlock(&vm_object_list_mtx);
1801 /* Depends on type-stability. */
1802 VM_OBJECT_WLOCK(object);
1805 * Dead objects are eventually terminated on their own.
1807 if ((object->flags & OBJ_DEAD) != 0)
1811 * Sync with fences placed after pctrie
1812 * initialization. We must not access pctrie below
1813 * unless we checked that our object is swap and not
1816 atomic_thread_fence_acq();
1817 if (object->type != OBJT_SWAP)
1820 swap_pager_swapoff_object(sp, object);
1822 VM_OBJECT_WUNLOCK(object);
1823 mtx_lock(&vm_object_list_mtx);
1825 mtx_unlock(&vm_object_list_mtx);
1829 * Objects may be locked or paging to the device being
1830 * removed, so we will miss their pages and need to
1831 * make another pass. We have marked this device as
1832 * SW_CLOSING, so the activity should finish soon.
1835 if (retries > 100) {
1836 panic("swapoff: failed to locate %d swap blocks",
1839 pause("swpoff", hz / 20);
1842 EVENTHANDLER_INVOKE(swapoff, sp);
1845 /************************************************************************
1847 ************************************************************************
1849 * These routines manipulate the swap metadata stored in the
1852 * Swap metadata is implemented with a global hash and not directly
1853 * linked into the object. Instead the object simply contains
1854 * appropriate tracking counters.
1858 * SWP_PAGER_SWBLK_EMPTY() - is a range of blocks free?
1861 swp_pager_swblk_empty(struct swblk *sb, int start, int limit)
1865 MPASS(0 <= start && start <= limit && limit <= SWAP_META_PAGES);
1866 for (i = start; i < limit; i++) {
1867 if (sb->d[i] != SWAPBLK_NONE)
1874 * SWP_PAGER_META_BUILD() - add swap block to swap meta data for object
1876 * We first convert the object to a swap object if it is a default
1879 * The specified swapblk is added to the object's swap metadata. If
1880 * the swapblk is not valid, it is freed instead. Any previously
1881 * assigned swapblk is returned.
1884 swp_pager_meta_build(vm_object_t object, vm_pindex_t pindex, daddr_t swapblk)
1886 static volatile int swblk_zone_exhausted, swpctrie_zone_exhausted;
1887 struct swblk *sb, *sb1;
1888 vm_pindex_t modpi, rdpi;
1889 daddr_t prev_swapblk;
1892 VM_OBJECT_ASSERT_WLOCKED(object);
1895 * Convert default object to swap object if necessary
1897 if (object->type != OBJT_SWAP) {
1898 pctrie_init(&object->un_pager.swp.swp_blks);
1901 * Ensure that swap_pager_swapoff()'s iteration over
1902 * object_list does not see a garbage pctrie.
1904 atomic_thread_fence_rel();
1906 object->type = OBJT_SWAP;
1907 object->un_pager.swp.writemappings = 0;
1908 KASSERT(object->handle == NULL, ("default pager with handle"));
1911 rdpi = rounddown(pindex, SWAP_META_PAGES);
1912 sb = SWAP_PCTRIE_LOOKUP(&object->un_pager.swp.swp_blks, rdpi);
1914 if (swapblk == SWAPBLK_NONE)
1915 return (SWAPBLK_NONE);
1917 sb = uma_zalloc(swblk_zone, M_NOWAIT | (curproc ==
1918 pageproc ? M_USE_RESERVE : 0));
1921 for (i = 0; i < SWAP_META_PAGES; i++)
1922 sb->d[i] = SWAPBLK_NONE;
1923 if (atomic_cmpset_int(&swblk_zone_exhausted,
1925 printf("swblk zone ok\n");
1928 VM_OBJECT_WUNLOCK(object);
1929 if (uma_zone_exhausted(swblk_zone)) {
1930 if (atomic_cmpset_int(&swblk_zone_exhausted,
1932 printf("swap blk zone exhausted, "
1933 "increase kern.maxswzone\n");
1934 vm_pageout_oom(VM_OOM_SWAPZ);
1935 pause("swzonxb", 10);
1937 uma_zwait(swblk_zone);
1938 VM_OBJECT_WLOCK(object);
1939 sb = SWAP_PCTRIE_LOOKUP(&object->un_pager.swp.swp_blks,
1943 * Somebody swapped out a nearby page,
1944 * allocating swblk at the rdpi index,
1945 * while we dropped the object lock.
1950 error = SWAP_PCTRIE_INSERT(
1951 &object->un_pager.swp.swp_blks, sb);
1953 if (atomic_cmpset_int(&swpctrie_zone_exhausted,
1955 printf("swpctrie zone ok\n");
1958 VM_OBJECT_WUNLOCK(object);
1959 if (uma_zone_exhausted(swpctrie_zone)) {
1960 if (atomic_cmpset_int(&swpctrie_zone_exhausted,
1962 printf("swap pctrie zone exhausted, "
1963 "increase kern.maxswzone\n");
1964 vm_pageout_oom(VM_OOM_SWAPZ);
1965 pause("swzonxp", 10);
1967 uma_zwait(swpctrie_zone);
1968 VM_OBJECT_WLOCK(object);
1969 sb1 = SWAP_PCTRIE_LOOKUP(&object->un_pager.swp.swp_blks,
1972 uma_zfree(swblk_zone, sb);
1979 MPASS(sb->p == rdpi);
1981 modpi = pindex % SWAP_META_PAGES;
1982 /* Return prior contents of metadata. */
1983 prev_swapblk = sb->d[modpi];
1984 /* Enter block into metadata. */
1985 sb->d[modpi] = swapblk;
1988 * Free the swblk if we end up with the empty page run.
1990 if (swapblk == SWAPBLK_NONE &&
1991 swp_pager_swblk_empty(sb, 0, SWAP_META_PAGES)) {
1992 SWAP_PCTRIE_REMOVE(&object->un_pager.swp.swp_blks, rdpi);
1993 uma_zfree(swblk_zone, sb);
1995 return (prev_swapblk);
1999 * SWP_PAGER_META_FREE() - free a range of blocks in the object's swap metadata
2001 * The requested range of blocks is freed, with any associated swap
2002 * returned to the swap bitmap.
2004 * This routine will free swap metadata structures as they are cleaned
2005 * out. This routine does *NOT* operate on swap metadata associated
2006 * with resident pages.
2009 swp_pager_meta_free(vm_object_t object, vm_pindex_t pindex, vm_pindex_t count)
2012 daddr_t n_free, s_free;
2014 int i, limit, start;
2016 VM_OBJECT_ASSERT_WLOCKED(object);
2017 if (object->type != OBJT_SWAP || count == 0)
2020 swp_pager_init_freerange(&s_free, &n_free);
2021 last = pindex + count;
2023 sb = SWAP_PCTRIE_LOOKUP_GE(&object->un_pager.swp.swp_blks,
2024 rounddown(pindex, SWAP_META_PAGES));
2025 if (sb == NULL || sb->p >= last)
2027 start = pindex > sb->p ? pindex - sb->p : 0;
2028 limit = last - sb->p < SWAP_META_PAGES ? last - sb->p :
2030 for (i = start; i < limit; i++) {
2031 if (sb->d[i] == SWAPBLK_NONE)
2033 swp_pager_update_freerange(&s_free, &n_free, sb->d[i]);
2034 sb->d[i] = SWAPBLK_NONE;
2036 pindex = sb->p + SWAP_META_PAGES;
2037 if (swp_pager_swblk_empty(sb, 0, start) &&
2038 swp_pager_swblk_empty(sb, limit, SWAP_META_PAGES)) {
2039 SWAP_PCTRIE_REMOVE(&object->un_pager.swp.swp_blks,
2041 uma_zfree(swblk_zone, sb);
2044 swp_pager_freeswapspace(s_free, n_free);
2048 * SWP_PAGER_META_FREE_ALL() - destroy all swap metadata associated with object
2050 * This routine locates and destroys all swap metadata associated with
2054 swp_pager_meta_free_all(vm_object_t object)
2057 daddr_t n_free, s_free;
2061 VM_OBJECT_ASSERT_WLOCKED(object);
2062 if (object->type != OBJT_SWAP)
2065 swp_pager_init_freerange(&s_free, &n_free);
2066 for (pindex = 0; (sb = SWAP_PCTRIE_LOOKUP_GE(
2067 &object->un_pager.swp.swp_blks, pindex)) != NULL;) {
2068 pindex = sb->p + SWAP_META_PAGES;
2069 for (i = 0; i < SWAP_META_PAGES; i++) {
2070 if (sb->d[i] == SWAPBLK_NONE)
2072 swp_pager_update_freerange(&s_free, &n_free, sb->d[i]);
2074 SWAP_PCTRIE_REMOVE(&object->un_pager.swp.swp_blks, sb->p);
2075 uma_zfree(swblk_zone, sb);
2077 swp_pager_freeswapspace(s_free, n_free);
2081 * SWP_PAGER_METACTL() - misc control of swap meta data.
2083 * This routine is capable of looking up, or removing swapblk
2084 * assignments in the swap meta data. It returns the swapblk being
2085 * looked-up, popped, or SWAPBLK_NONE if the block was invalid.
2087 * When acting on a busy resident page and paging is in progress, we
2088 * have to wait until paging is complete but otherwise can act on the
2091 * SWM_POP remove from meta data but do not free it
2094 swp_pager_meta_ctl(vm_object_t object, vm_pindex_t pindex, int flags)
2099 if ((flags & SWM_POP) != 0)
2100 VM_OBJECT_ASSERT_WLOCKED(object);
2102 VM_OBJECT_ASSERT_LOCKED(object);
2105 * The meta data only exists if the object is OBJT_SWAP
2106 * and even then might not be allocated yet.
2108 if (object->type != OBJT_SWAP)
2109 return (SWAPBLK_NONE);
2111 sb = SWAP_PCTRIE_LOOKUP(&object->un_pager.swp.swp_blks,
2112 rounddown(pindex, SWAP_META_PAGES));
2114 return (SWAPBLK_NONE);
2115 r1 = sb->d[pindex % SWAP_META_PAGES];
2116 if (r1 == SWAPBLK_NONE)
2117 return (SWAPBLK_NONE);
2118 if ((flags & SWM_POP) != 0) {
2119 sb->d[pindex % SWAP_META_PAGES] = SWAPBLK_NONE;
2120 if (swp_pager_swblk_empty(sb, 0, SWAP_META_PAGES)) {
2121 SWAP_PCTRIE_REMOVE(&object->un_pager.swp.swp_blks,
2122 rounddown(pindex, SWAP_META_PAGES));
2123 uma_zfree(swblk_zone, sb);
2130 * Returns the least page index which is greater than or equal to the
2131 * parameter pindex and for which there is a swap block allocated.
2132 * Returns object's size if the object's type is not swap or if there
2133 * are no allocated swap blocks for the object after the requested
2137 swap_pager_find_least(vm_object_t object, vm_pindex_t pindex)
2142 VM_OBJECT_ASSERT_LOCKED(object);
2143 if (object->type != OBJT_SWAP)
2144 return (object->size);
2146 sb = SWAP_PCTRIE_LOOKUP_GE(&object->un_pager.swp.swp_blks,
2147 rounddown(pindex, SWAP_META_PAGES));
2149 return (object->size);
2150 if (sb->p < pindex) {
2151 for (i = pindex % SWAP_META_PAGES; i < SWAP_META_PAGES; i++) {
2152 if (sb->d[i] != SWAPBLK_NONE)
2155 sb = SWAP_PCTRIE_LOOKUP_GE(&object->un_pager.swp.swp_blks,
2156 roundup(pindex, SWAP_META_PAGES));
2158 return (object->size);
2160 for (i = 0; i < SWAP_META_PAGES; i++) {
2161 if (sb->d[i] != SWAPBLK_NONE)
2166 * We get here if a swblk is present in the trie but it
2167 * doesn't map any blocks.
2170 return (object->size);
2174 * System call swapon(name) enables swapping on device name,
2175 * which must be in the swdevsw. Return EBUSY
2176 * if already swapping on this device.
2178 #ifndef _SYS_SYSPROTO_H_
2179 struct swapon_args {
2189 sys_swapon(struct thread *td, struct swapon_args *uap)
2193 struct nameidata nd;
2196 error = priv_check(td, PRIV_SWAPON);
2200 sx_xlock(&swdev_syscall_lock);
2203 * Swap metadata may not fit in the KVM if we have physical
2206 if (swblk_zone == NULL) {
2211 NDINIT(&nd, LOOKUP, ISOPEN | FOLLOW | AUDITVNODE1, UIO_USERSPACE,
2217 NDFREE(&nd, NDF_ONLY_PNBUF);
2220 if (vn_isdisk(vp, &error)) {
2221 error = swapongeom(vp);
2222 } else if (vp->v_type == VREG &&
2223 (vp->v_mount->mnt_vfc->vfc_flags & VFCF_NETWORK) != 0 &&
2224 (error = VOP_GETATTR(vp, &attr, td->td_ucred)) == 0) {
2226 * Allow direct swapping to NFS regular files in the same
2227 * way that nfs_mountroot() sets up diskless swapping.
2229 error = swaponvp(td, vp, attr.va_size / DEV_BSIZE);
2235 sx_xunlock(&swdev_syscall_lock);
2240 * Check that the total amount of swap currently configured does not
2241 * exceed half the theoretical maximum. If it does, print a warning
2245 swapon_check_swzone(void)
2247 unsigned long maxpages, npages;
2249 npages = swap_total;
2250 /* absolute maximum we can handle assuming 100% efficiency */
2251 maxpages = uma_zone_get_max(swblk_zone) * SWAP_META_PAGES;
2253 /* recommend using no more than half that amount */
2254 if (npages > maxpages / 2) {
2255 printf("warning: total configured swap (%lu pages) "
2256 "exceeds maximum recommended amount (%lu pages).\n",
2257 npages, maxpages / 2);
2258 printf("warning: increase kern.maxswzone "
2259 "or reduce amount of swap.\n");
2264 swaponsomething(struct vnode *vp, void *id, u_long nblks,
2265 sw_strategy_t *strategy, sw_close_t *close, dev_t dev, int flags)
2267 struct swdevt *sp, *tsp;
2272 * nblks is in DEV_BSIZE'd chunks, convert to PAGE_SIZE'd chunks.
2273 * First chop nblks off to page-align it, then convert.
2275 * sw->sw_nblks is in page-sized chunks now too.
2277 nblks &= ~(ctodb(1) - 1);
2278 nblks = dbtoc(nblks);
2281 * If we go beyond this, we get overflows in the radix
2284 mblocks = 0x40000000 / BLIST_META_RADIX;
2285 if (nblks > mblocks) {
2287 "WARNING: reducing swap size to maximum of %luMB per unit\n",
2288 mblocks / 1024 / 1024 * PAGE_SIZE);
2292 sp = malloc(sizeof *sp, M_VMPGDATA, M_WAITOK | M_ZERO);
2296 sp->sw_nblks = nblks;
2298 sp->sw_strategy = strategy;
2299 sp->sw_close = close;
2300 sp->sw_flags = flags;
2302 sp->sw_blist = blist_create(nblks, M_WAITOK);
2304 * Do not free the first blocks in order to avoid overwriting
2305 * any bsd label at the front of the partition
2307 blist_free(sp->sw_blist, howmany(BBSIZE, PAGE_SIZE),
2308 nblks - howmany(BBSIZE, PAGE_SIZE));
2311 mtx_lock(&sw_dev_mtx);
2312 TAILQ_FOREACH(tsp, &swtailq, sw_list) {
2313 if (tsp->sw_end >= dvbase) {
2315 * We put one uncovered page between the devices
2316 * in order to definitively prevent any cross-device
2319 dvbase = tsp->sw_end + 1;
2322 sp->sw_first = dvbase;
2323 sp->sw_end = dvbase + nblks;
2324 TAILQ_INSERT_TAIL(&swtailq, sp, sw_list);
2326 swap_pager_avail += nblks - howmany(BBSIZE, PAGE_SIZE);
2327 swap_total += nblks;
2328 swapon_check_swzone();
2330 mtx_unlock(&sw_dev_mtx);
2331 EVENTHANDLER_INVOKE(swapon, sp);
2335 * SYSCALL: swapoff(devname)
2337 * Disable swapping on the given device.
2339 * XXX: Badly designed system call: it should use a device index
2340 * rather than filename as specification. We keep sw_vp around
2341 * only to make this work.
2343 #ifndef _SYS_SYSPROTO_H_
2344 struct swapoff_args {
2354 sys_swapoff(struct thread *td, struct swapoff_args *uap)
2357 struct nameidata nd;
2361 error = priv_check(td, PRIV_SWAPOFF);
2365 sx_xlock(&swdev_syscall_lock);
2367 NDINIT(&nd, LOOKUP, FOLLOW | AUDITVNODE1, UIO_USERSPACE, uap->name,
2372 NDFREE(&nd, NDF_ONLY_PNBUF);
2375 mtx_lock(&sw_dev_mtx);
2376 TAILQ_FOREACH(sp, &swtailq, sw_list) {
2377 if (sp->sw_vp == vp)
2380 mtx_unlock(&sw_dev_mtx);
2385 error = swapoff_one(sp, td->td_ucred);
2387 sx_xunlock(&swdev_syscall_lock);
2392 swapoff_one(struct swdevt *sp, struct ucred *cred)
2399 sx_assert(&swdev_syscall_lock, SA_XLOCKED);
2401 (void) vn_lock(sp->sw_vp, LK_EXCLUSIVE | LK_RETRY);
2402 error = mac_system_check_swapoff(cred, sp->sw_vp);
2403 (void) VOP_UNLOCK(sp->sw_vp, 0);
2407 nblks = sp->sw_nblks;
2410 * We can turn off this swap device safely only if the
2411 * available virtual memory in the system will fit the amount
2412 * of data we will have to page back in, plus an epsilon so
2413 * the system doesn't become critically low on swap space.
2415 if (vm_free_count() + swap_pager_avail < nblks + nswap_lowat)
2419 * Prevent further allocations on this device.
2421 mtx_lock(&sw_dev_mtx);
2422 sp->sw_flags |= SW_CLOSING;
2423 swap_pager_avail -= blist_fill(sp->sw_blist, 0, nblks);
2424 swap_total -= nblks;
2425 mtx_unlock(&sw_dev_mtx);
2428 * Page in the contents of the device and close it.
2430 swap_pager_swapoff(sp);
2432 sp->sw_close(curthread, sp);
2433 mtx_lock(&sw_dev_mtx);
2435 TAILQ_REMOVE(&swtailq, sp, sw_list);
2437 if (nswapdev == 0) {
2438 swap_pager_full = 2;
2439 swap_pager_almost_full = 1;
2443 mtx_unlock(&sw_dev_mtx);
2444 blist_destroy(sp->sw_blist);
2445 free(sp, M_VMPGDATA);
2452 struct swdevt *sp, *spt;
2453 const char *devname;
2456 sx_xlock(&swdev_syscall_lock);
2458 mtx_lock(&sw_dev_mtx);
2459 TAILQ_FOREACH_SAFE(sp, &swtailq, sw_list, spt) {
2460 mtx_unlock(&sw_dev_mtx);
2461 if (vn_isdisk(sp->sw_vp, NULL))
2462 devname = devtoname(sp->sw_vp->v_rdev);
2465 error = swapoff_one(sp, thread0.td_ucred);
2467 printf("Cannot remove swap device %s (error=%d), "
2468 "skipping.\n", devname, error);
2469 } else if (bootverbose) {
2470 printf("Swap device %s removed.\n", devname);
2472 mtx_lock(&sw_dev_mtx);
2474 mtx_unlock(&sw_dev_mtx);
2476 sx_xunlock(&swdev_syscall_lock);
2480 swap_pager_status(int *total, int *used)
2486 mtx_lock(&sw_dev_mtx);
2487 TAILQ_FOREACH(sp, &swtailq, sw_list) {
2488 *total += sp->sw_nblks;
2489 *used += sp->sw_used;
2491 mtx_unlock(&sw_dev_mtx);
2495 swap_dev_info(int name, struct xswdev *xs, char *devname, size_t len)
2498 const char *tmp_devname;
2503 mtx_lock(&sw_dev_mtx);
2504 TAILQ_FOREACH(sp, &swtailq, sw_list) {
2509 xs->xsw_version = XSWDEV_VERSION;
2510 xs->xsw_dev = sp->sw_dev;
2511 xs->xsw_flags = sp->sw_flags;
2512 xs->xsw_nblks = sp->sw_nblks;
2513 xs->xsw_used = sp->sw_used;
2514 if (devname != NULL) {
2515 if (vn_isdisk(sp->sw_vp, NULL))
2516 tmp_devname = devtoname(sp->sw_vp->v_rdev);
2518 tmp_devname = "[file]";
2519 strncpy(devname, tmp_devname, len);
2524 mtx_unlock(&sw_dev_mtx);
2528 #if defined(COMPAT_FREEBSD11)
2529 #define XSWDEV_VERSION_11 1
2539 #if defined(__amd64__) && defined(COMPAT_FREEBSD32)
2542 u_int xsw_dev1, xsw_dev2;
2550 sysctl_vm_swap_info(SYSCTL_HANDLER_ARGS)
2553 #if defined(__amd64__) && defined(COMPAT_FREEBSD32)
2554 struct xswdev32 xs32;
2556 #if defined(COMPAT_FREEBSD11)
2557 struct xswdev11 xs11;
2561 if (arg2 != 1) /* name length */
2563 error = swap_dev_info(*(int *)arg1, &xs, NULL, 0);
2566 #if defined(__amd64__) && defined(COMPAT_FREEBSD32)
2567 if (req->oldlen == sizeof(xs32)) {
2568 xs32.xsw_version = XSWDEV_VERSION;
2569 xs32.xsw_dev1 = xs.xsw_dev;
2570 xs32.xsw_dev2 = xs.xsw_dev >> 32;
2571 xs32.xsw_flags = xs.xsw_flags;
2572 xs32.xsw_nblks = xs.xsw_nblks;
2573 xs32.xsw_used = xs.xsw_used;
2574 error = SYSCTL_OUT(req, &xs32, sizeof(xs32));
2578 #if defined(COMPAT_FREEBSD11)
2579 if (req->oldlen == sizeof(xs11)) {
2580 xs11.xsw_version = XSWDEV_VERSION_11;
2581 xs11.xsw_dev = xs.xsw_dev; /* truncation */
2582 xs11.xsw_flags = xs.xsw_flags;
2583 xs11.xsw_nblks = xs.xsw_nblks;
2584 xs11.xsw_used = xs.xsw_used;
2585 error = SYSCTL_OUT(req, &xs11, sizeof(xs11));
2589 error = SYSCTL_OUT(req, &xs, sizeof(xs));
2593 SYSCTL_INT(_vm, OID_AUTO, nswapdev, CTLFLAG_RD, &nswapdev, 0,
2594 "Number of swap devices");
2595 SYSCTL_NODE(_vm, OID_AUTO, swap_info, CTLFLAG_RD | CTLFLAG_MPSAFE,
2596 sysctl_vm_swap_info,
2597 "Swap statistics by device");
2600 * Count the approximate swap usage in pages for a vmspace. The
2601 * shadowed or not yet copied on write swap blocks are not accounted.
2602 * The map must be locked.
2605 vmspace_swap_count(struct vmspace *vmspace)
2615 map = &vmspace->vm_map;
2618 for (cur = map->header.next; cur != &map->header; cur = cur->next) {
2619 if ((cur->eflags & MAP_ENTRY_IS_SUB_MAP) != 0)
2621 object = cur->object.vm_object;
2622 if (object == NULL || object->type != OBJT_SWAP)
2624 VM_OBJECT_RLOCK(object);
2625 if (object->type != OBJT_SWAP)
2627 pi = OFF_TO_IDX(cur->offset);
2628 e = pi + OFF_TO_IDX(cur->end - cur->start);
2629 for (;; pi = sb->p + SWAP_META_PAGES) {
2630 sb = SWAP_PCTRIE_LOOKUP_GE(
2631 &object->un_pager.swp.swp_blks, pi);
2632 if (sb == NULL || sb->p >= e)
2634 for (i = 0; i < SWAP_META_PAGES; i++) {
2635 if (sb->p + i < e &&
2636 sb->d[i] != SWAPBLK_NONE)
2641 VM_OBJECT_RUNLOCK(object);
2649 * Swapping onto disk devices.
2653 static g_orphan_t swapgeom_orphan;
2655 static struct g_class g_swap_class = {
2657 .version = G_VERSION,
2658 .orphan = swapgeom_orphan,
2661 DECLARE_GEOM_CLASS(g_swap_class, g_class);
2665 swapgeom_close_ev(void *arg, int flags)
2667 struct g_consumer *cp;
2670 g_access(cp, -1, -1, 0);
2672 g_destroy_consumer(cp);
2676 * Add a reference to the g_consumer for an inflight transaction.
2679 swapgeom_acquire(struct g_consumer *cp)
2682 mtx_assert(&sw_dev_mtx, MA_OWNED);
2687 * Remove a reference from the g_consumer. Post a close event if all
2688 * references go away, since the function might be called from the
2692 swapgeom_release(struct g_consumer *cp, struct swdevt *sp)
2695 mtx_assert(&sw_dev_mtx, MA_OWNED);
2697 if (cp->index == 0) {
2698 if (g_post_event(swapgeom_close_ev, cp, M_NOWAIT, NULL) == 0)
2704 swapgeom_done(struct bio *bp2)
2708 struct g_consumer *cp;
2710 bp = bp2->bio_caller2;
2712 bp->b_ioflags = bp2->bio_flags;
2714 bp->b_ioflags |= BIO_ERROR;
2715 bp->b_resid = bp->b_bcount - bp2->bio_completed;
2716 bp->b_error = bp2->bio_error;
2717 bp->b_caller1 = NULL;
2719 sp = bp2->bio_caller1;
2720 mtx_lock(&sw_dev_mtx);
2721 swapgeom_release(cp, sp);
2722 mtx_unlock(&sw_dev_mtx);
2727 swapgeom_strategy(struct buf *bp, struct swdevt *sp)
2730 struct g_consumer *cp;
2732 mtx_lock(&sw_dev_mtx);
2735 mtx_unlock(&sw_dev_mtx);
2736 bp->b_error = ENXIO;
2737 bp->b_ioflags |= BIO_ERROR;
2741 swapgeom_acquire(cp);
2742 mtx_unlock(&sw_dev_mtx);
2743 if (bp->b_iocmd == BIO_WRITE)
2746 bio = g_alloc_bio();
2748 mtx_lock(&sw_dev_mtx);
2749 swapgeom_release(cp, sp);
2750 mtx_unlock(&sw_dev_mtx);
2751 bp->b_error = ENOMEM;
2752 bp->b_ioflags |= BIO_ERROR;
2753 printf("swap_pager: cannot allocate bio\n");
2758 bp->b_caller1 = bio;
2759 bio->bio_caller1 = sp;
2760 bio->bio_caller2 = bp;
2761 bio->bio_cmd = bp->b_iocmd;
2762 bio->bio_offset = (bp->b_blkno - sp->sw_first) * PAGE_SIZE;
2763 bio->bio_length = bp->b_bcount;
2764 bio->bio_done = swapgeom_done;
2765 if (!buf_mapped(bp)) {
2766 bio->bio_ma = bp->b_pages;
2767 bio->bio_data = unmapped_buf;
2768 bio->bio_ma_offset = (vm_offset_t)bp->b_offset & PAGE_MASK;
2769 bio->bio_ma_n = bp->b_npages;
2770 bio->bio_flags |= BIO_UNMAPPED;
2772 bio->bio_data = bp->b_data;
2775 g_io_request(bio, cp);
2780 swapgeom_orphan(struct g_consumer *cp)
2785 mtx_lock(&sw_dev_mtx);
2786 TAILQ_FOREACH(sp, &swtailq, sw_list) {
2787 if (sp->sw_id == cp) {
2788 sp->sw_flags |= SW_CLOSING;
2793 * Drop reference we were created with. Do directly since we're in a
2794 * special context where we don't have to queue the call to
2795 * swapgeom_close_ev().
2798 destroy = ((sp != NULL) && (cp->index == 0));
2801 mtx_unlock(&sw_dev_mtx);
2803 swapgeom_close_ev(cp, 0);
2807 swapgeom_close(struct thread *td, struct swdevt *sw)
2809 struct g_consumer *cp;
2811 mtx_lock(&sw_dev_mtx);
2814 mtx_unlock(&sw_dev_mtx);
2817 * swapgeom_close() may be called from the biodone context,
2818 * where we cannot perform topology changes. Delegate the
2819 * work to the events thread.
2822 g_waitfor_event(swapgeom_close_ev, cp, M_WAITOK, NULL);
2826 swapongeom_locked(struct cdev *dev, struct vnode *vp)
2828 struct g_provider *pp;
2829 struct g_consumer *cp;
2830 static struct g_geom *gp;
2835 pp = g_dev_getprovider(dev);
2838 mtx_lock(&sw_dev_mtx);
2839 TAILQ_FOREACH(sp, &swtailq, sw_list) {
2841 if (cp != NULL && cp->provider == pp) {
2842 mtx_unlock(&sw_dev_mtx);
2846 mtx_unlock(&sw_dev_mtx);
2848 gp = g_new_geomf(&g_swap_class, "swap");
2849 cp = g_new_consumer(gp);
2850 cp->index = 1; /* Number of active I/Os, plus one for being active. */
2851 cp->flags |= G_CF_DIRECT_SEND | G_CF_DIRECT_RECEIVE;
2854 * XXX: Every time you think you can improve the margin for
2855 * footshooting, somebody depends on the ability to do so:
2856 * savecore(8) wants to write to our swapdev so we cannot
2857 * set an exclusive count :-(
2859 error = g_access(cp, 1, 1, 0);
2862 g_destroy_consumer(cp);
2865 nblks = pp->mediasize / DEV_BSIZE;
2866 swaponsomething(vp, cp, nblks, swapgeom_strategy,
2867 swapgeom_close, dev2udev(dev),
2868 (pp->flags & G_PF_ACCEPT_UNMAPPED) != 0 ? SW_UNMAPPED : 0);
2873 swapongeom(struct vnode *vp)
2877 vn_lock(vp, LK_EXCLUSIVE | LK_RETRY);
2878 if (vp->v_type != VCHR || (vp->v_iflag & VI_DOOMED) != 0) {
2882 error = swapongeom_locked(vp->v_rdev, vp);
2883 g_topology_unlock();
2892 * This is used mainly for network filesystem (read: probably only tested
2893 * with NFS) swapfiles.
2898 swapdev_strategy(struct buf *bp, struct swdevt *sp)
2902 bp->b_blkno = ctodb(bp->b_blkno - sp->sw_first);
2906 if (bp->b_iocmd == BIO_WRITE) {
2908 bufobj_wdrop(bp->b_bufobj);
2909 bufobj_wref(&vp2->v_bufobj);
2911 if (bp->b_bufobj != &vp2->v_bufobj)
2912 bp->b_bufobj = &vp2->v_bufobj;
2914 bp->b_iooffset = dbtob(bp->b_blkno);
2920 swapdev_close(struct thread *td, struct swdevt *sp)
2923 VOP_CLOSE(sp->sw_vp, FREAD | FWRITE, td->td_ucred, td);
2929 swaponvp(struct thread *td, struct vnode *vp, u_long nblks)
2936 mtx_lock(&sw_dev_mtx);
2937 TAILQ_FOREACH(sp, &swtailq, sw_list) {
2938 if (sp->sw_id == vp) {
2939 mtx_unlock(&sw_dev_mtx);
2943 mtx_unlock(&sw_dev_mtx);
2945 (void) vn_lock(vp, LK_EXCLUSIVE | LK_RETRY);
2947 error = mac_system_check_swapon(td->td_ucred, vp);
2950 error = VOP_OPEN(vp, FREAD | FWRITE, td->td_ucred, td, NULL);
2951 (void) VOP_UNLOCK(vp, 0);
2955 swaponsomething(vp, vp, nblks, swapdev_strategy, swapdev_close,
2961 sysctl_swap_async_max(SYSCTL_HANDLER_ARGS)
2965 new = nsw_wcount_async_max;
2966 error = sysctl_handle_int(oidp, &new, 0, req);
2967 if (error != 0 || req->newptr == NULL)
2970 if (new > nswbuf / 2 || new < 1)
2973 mtx_lock(&swbuf_mtx);
2974 while (nsw_wcount_async_max != new) {
2976 * Adjust difference. If the current async count is too low,
2977 * we will need to sqeeze our update slowly in. Sleep with a
2978 * higher priority than getpbuf() to finish faster.
2980 n = new - nsw_wcount_async_max;
2981 if (nsw_wcount_async + n >= 0) {
2982 nsw_wcount_async += n;
2983 nsw_wcount_async_max += n;
2984 wakeup(&nsw_wcount_async);
2986 nsw_wcount_async_max -= nsw_wcount_async;
2987 nsw_wcount_async = 0;
2988 msleep(&nsw_wcount_async, &swbuf_mtx, PSWP,
2992 mtx_unlock(&swbuf_mtx);
2998 swap_pager_update_writecount(vm_object_t object, vm_offset_t start,
3002 VM_OBJECT_WLOCK(object);
3003 KASSERT((object->flags & OBJ_NOSPLIT) != 0,
3004 ("Splittable object with writecount"));
3005 object->un_pager.swp.writemappings += (vm_ooffset_t)end - start;
3006 VM_OBJECT_WUNLOCK(object);
3010 swap_pager_release_writecount(vm_object_t object, vm_offset_t start,
3014 VM_OBJECT_WLOCK(object);
3015 KASSERT((object->flags & OBJ_NOSPLIT) != 0,
3016 ("Splittable object with writecount"));
3017 object->un_pager.swp.writemappings -= (vm_ooffset_t)end - start;
3018 VM_OBJECT_WUNLOCK(object);