2 * SPDX-License-Identifier: BSD-4-Clause
4 * Copyright (c) 1998 Matthew Dillon,
5 * Copyright (c) 1994 John S. Dyson
6 * Copyright (c) 1990 University of Utah.
7 * Copyright (c) 1982, 1986, 1989, 1993
8 * The Regents of the University of California. All rights reserved.
10 * This code is derived from software contributed to Berkeley by
11 * the Systems Programming Group of the University of Utah Computer
14 * Redistribution and use in source and binary forms, with or without
15 * modification, are permitted provided that the following conditions
17 * 1. Redistributions of source code must retain the above copyright
18 * notice, this list of conditions and the following disclaimer.
19 * 2. Redistributions in binary form must reproduce the above copyright
20 * notice, this list of conditions and the following disclaimer in the
21 * documentation and/or other materials provided with the distribution.
22 * 3. All advertising materials mentioning features or use of this software
23 * must display the following acknowledgement:
24 * This product includes software developed by the University of
25 * California, Berkeley and its contributors.
26 * 4. Neither the name of the University nor the names of its contributors
27 * may be used to endorse or promote products derived from this software
28 * without specific prior written permission.
30 * THIS SOFTWARE IS PROVIDED BY THE REGENTS AND CONTRIBUTORS ``AS IS'' AND
31 * ANY EXPRESS OR IMPLIED WARRANTIES, INCLUDING, BUT NOT LIMITED TO, THE
32 * IMPLIED WARRANTIES OF MERCHANTABILITY AND FITNESS FOR A PARTICULAR PURPOSE
33 * ARE DISCLAIMED. IN NO EVENT SHALL THE REGENTS OR CONTRIBUTORS BE LIABLE
34 * FOR ANY DIRECT, INDIRECT, INCIDENTAL, SPECIAL, EXEMPLARY, OR CONSEQUENTIAL
35 * DAMAGES (INCLUDING, BUT NOT LIMITED TO, PROCUREMENT OF SUBSTITUTE GOODS
36 * OR SERVICES; LOSS OF USE, DATA, OR PROFITS; OR BUSINESS INTERRUPTION)
37 * HOWEVER CAUSED AND ON ANY THEORY OF LIABILITY, WHETHER IN CONTRACT, STRICT
38 * LIABILITY, OR TORT (INCLUDING NEGLIGENCE OR OTHERWISE) ARISING IN ANY WAY
39 * OUT OF THE USE OF THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF
45 * Radix Bitmap 'blists'.
47 * - The new swapper uses the new radix bitmap code. This should scale
48 * to arbitrarily small or arbitrarily large swap spaces and an almost
49 * arbitrary degree of fragmentation.
53 * - on the fly reallocation of swap during putpages. The new system
54 * does not try to keep previously allocated swap blocks for dirty
57 * - on the fly deallocation of swap
59 * - No more garbage collection required. Unnecessarily allocated swap
60 * blocks only exist for dirty vm_page_t's now and these are already
61 * cycled (in a high-load system) by the pager. We also do on-the-fly
62 * removal of invalidated swap blocks when a page is destroyed
65 * from: Utah $Hdr: swap_pager.c 1.4 91/04/30$
67 * @(#)swap_pager.c 8.9 (Berkeley) 3/21/94
68 * @(#)vm_swap.c 8.5 (Berkeley) 2/17/94
71 #include <sys/cdefs.h>
72 __FBSDID("$FreeBSD$");
76 #include <sys/param.h>
78 #include <sys/blist.h>
82 #include <sys/eventhandler.h>
83 #include <sys/fcntl.h>
85 #include <sys/kernel.h>
86 #include <sys/mount.h>
87 #include <sys/namei.h>
88 #include <sys/malloc.h>
89 #include <sys/pctrie.h>
92 #include <sys/racct.h>
93 #include <sys/resource.h>
94 #include <sys/resourcevar.h>
95 #include <sys/rwlock.h>
97 #include <sys/sysctl.h>
98 #include <sys/sysproto.h>
99 #include <sys/systm.h>
101 #include <sys/vmmeter.h>
102 #include <sys/vnode.h>
104 #include <security/mac/mac_framework.h>
108 #include <vm/vm_map.h>
109 #include <vm/vm_kern.h>
110 #include <vm/vm_object.h>
111 #include <vm/vm_page.h>
112 #include <vm/vm_pager.h>
113 #include <vm/vm_pageout.h>
114 #include <vm/vm_param.h>
115 #include <vm/swap_pager.h>
116 #include <vm/vm_extern.h>
119 #include <geom/geom.h>
122 * MAX_PAGEOUT_CLUSTER must be a power of 2 between 1 and 64.
123 * The 64-page limit is due to the radix code (kern/subr_blist.c).
125 #ifndef MAX_PAGEOUT_CLUSTER
126 #define MAX_PAGEOUT_CLUSTER 32
129 #if !defined(SWB_NPAGES)
130 #define SWB_NPAGES MAX_PAGEOUT_CLUSTER
133 #define SWAP_META_PAGES PCTRIE_COUNT
136 * A swblk structure maps each page index within a
137 * SWAP_META_PAGES-aligned and sized range to the address of an
138 * on-disk swap block (or SWAPBLK_NONE). The collection of these
139 * mappings for an entire vm object is implemented as a pc-trie.
143 daddr_t d[SWAP_META_PAGES];
146 static MALLOC_DEFINE(M_VMPGDATA, "vm_pgdata", "swap pager private data");
147 static struct mtx sw_dev_mtx;
148 static TAILQ_HEAD(, swdevt) swtailq = TAILQ_HEAD_INITIALIZER(swtailq);
149 static struct swdevt *swdevhd; /* Allocate from here next */
150 static int nswapdev; /* Number of swap devices */
151 int swap_pager_avail;
152 static struct sx swdev_syscall_lock; /* serialize swap(on|off) */
154 static u_long swap_reserved;
155 static u_long swap_total;
156 static int sysctl_page_shift(SYSCTL_HANDLER_ARGS);
157 SYSCTL_PROC(_vm, OID_AUTO, swap_reserved, CTLTYPE_U64 | CTLFLAG_RD | CTLFLAG_MPSAFE,
158 &swap_reserved, 0, sysctl_page_shift, "A",
159 "Amount of swap storage needed to back all allocated anonymous memory.");
160 SYSCTL_PROC(_vm, OID_AUTO, swap_total, CTLTYPE_U64 | CTLFLAG_RD | CTLFLAG_MPSAFE,
161 &swap_total, 0, sysctl_page_shift, "A",
162 "Total amount of available swap storage.");
164 static int overcommit = 0;
165 SYSCTL_INT(_vm, VM_OVERCOMMIT, overcommit, CTLFLAG_RW, &overcommit, 0,
166 "Configure virtual memory overcommit behavior. See tuning(7) "
168 static unsigned long swzone;
169 SYSCTL_ULONG(_vm, OID_AUTO, swzone, CTLFLAG_RD, &swzone, 0,
170 "Actual size of swap metadata zone");
171 static unsigned long swap_maxpages;
172 SYSCTL_ULONG(_vm, OID_AUTO, swap_maxpages, CTLFLAG_RD, &swap_maxpages, 0,
173 "Maximum amount of swap supported");
175 /* bits from overcommit */
176 #define SWAP_RESERVE_FORCE_ON (1 << 0)
177 #define SWAP_RESERVE_RLIMIT_ON (1 << 1)
178 #define SWAP_RESERVE_ALLOW_NONWIRED (1 << 2)
181 sysctl_page_shift(SYSCTL_HANDLER_ARGS)
184 u_long value = *(u_long *)arg1;
186 newval = ((uint64_t)value) << PAGE_SHIFT;
187 return (sysctl_handle_64(oidp, &newval, 0, req));
191 swap_reserve(vm_ooffset_t incr)
194 return (swap_reserve_by_cred(incr, curthread->td_ucred));
198 swap_reserve_by_cred(vm_ooffset_t incr, struct ucred *cred)
200 u_long r, s, prev, pincr;
203 static struct timeval lastfail;
206 uip = cred->cr_ruidinfo;
208 KASSERT((incr & PAGE_MASK) == 0, ("%s: incr: %ju & PAGE_MASK", __func__,
214 error = racct_add(curproc, RACCT_SWAP, incr);
215 PROC_UNLOCK(curproc);
223 prev = atomic_fetchadd_long(&swap_reserved, pincr);
225 if (overcommit & SWAP_RESERVE_ALLOW_NONWIRED) {
226 s = vm_cnt.v_page_count - vm_cnt.v_free_reserved -
231 if ((overcommit & SWAP_RESERVE_FORCE_ON) == 0 || r <= s ||
232 (error = priv_check(curthread, PRIV_VM_SWAP_NOQUOTA)) == 0) {
235 prev = atomic_fetchadd_long(&swap_reserved, -pincr);
237 panic("swap_reserved < incr on overcommit fail");
240 prev = atomic_fetchadd_long(&uip->ui_vmsize, pincr);
241 if ((overcommit & SWAP_RESERVE_RLIMIT_ON) != 0 &&
242 prev + pincr > lim_cur(curthread, RLIMIT_SWAP) &&
243 priv_check(curthread, PRIV_VM_SWAP_NORLIMIT)) {
245 prev = atomic_fetchadd_long(&uip->ui_vmsize, -pincr);
247 panic("uip->ui_vmsize < incr on overcommit fail");
250 if (!res && ppsratecheck(&lastfail, &curfail, 1)) {
251 printf("uid %d, pid %d: swap reservation for %jd bytes failed\n",
252 uip->ui_uid, curproc->p_pid, incr);
256 if (racct_enable && !res) {
258 racct_sub(curproc, RACCT_SWAP, incr);
259 PROC_UNLOCK(curproc);
267 swap_reserve_force(vm_ooffset_t incr)
272 KASSERT((incr & PAGE_MASK) == 0, ("%s: incr: %ju & PAGE_MASK", __func__,
278 racct_add_force(curproc, RACCT_SWAP, incr);
281 atomic_add_long(&swap_reserved, pincr);
282 uip = curproc->p_ucred->cr_ruidinfo;
283 atomic_add_long(&uip->ui_vmsize, pincr);
284 PROC_UNLOCK(curproc);
288 swap_release(vm_ooffset_t decr)
293 cred = curproc->p_ucred;
294 swap_release_by_cred(decr, cred);
295 PROC_UNLOCK(curproc);
299 swap_release_by_cred(vm_ooffset_t decr, struct ucred *cred)
304 uip = cred->cr_ruidinfo;
306 KASSERT((decr & PAGE_MASK) == 0, ("%s: decr: %ju & PAGE_MASK", __func__,
310 prev = atomic_fetchadd_long(&swap_reserved, -pdecr);
312 panic("swap_reserved < decr");
314 prev = atomic_fetchadd_long(&uip->ui_vmsize, -pdecr);
316 printf("negative vmsize for uid = %d\n", uip->ui_uid);
319 racct_sub_cred(cred, RACCT_SWAP, decr);
323 #define SWM_POP 0x01 /* pop out */
325 static int swap_pager_full = 2; /* swap space exhaustion (task killing) */
326 static int swap_pager_almost_full = 1; /* swap space exhaustion (w/hysteresis)*/
327 static struct mtx swbuf_mtx; /* to sync nsw_wcount_async */
328 static int nsw_wcount_async; /* limit async write buffers */
329 static int nsw_wcount_async_max;/* assigned maximum */
330 static int nsw_cluster_max; /* maximum VOP I/O allowed */
332 static int sysctl_swap_async_max(SYSCTL_HANDLER_ARGS);
333 SYSCTL_PROC(_vm, OID_AUTO, swap_async_max, CTLTYPE_INT | CTLFLAG_RW |
334 CTLFLAG_MPSAFE, NULL, 0, sysctl_swap_async_max, "I",
335 "Maximum running async swap ops");
336 static int sysctl_swap_fragmentation(SYSCTL_HANDLER_ARGS);
337 SYSCTL_PROC(_vm, OID_AUTO, swap_fragmentation, CTLTYPE_STRING | CTLFLAG_RD |
338 CTLFLAG_MPSAFE, NULL, 0, sysctl_swap_fragmentation, "A",
339 "Swap Fragmentation Info");
341 static struct sx sw_alloc_sx;
344 * "named" and "unnamed" anon region objects. Try to reduce the overhead
345 * of searching a named list by hashing it just a little.
350 #define NOBJLIST(handle) \
351 (&swap_pager_object_list[((int)(intptr_t)handle >> 4) & (NOBJLISTS-1)])
353 static struct pagerlst swap_pager_object_list[NOBJLISTS];
354 static uma_zone_t swwbuf_zone;
355 static uma_zone_t swrbuf_zone;
356 static uma_zone_t swblk_zone;
357 static uma_zone_t swpctrie_zone;
360 * pagerops for OBJT_SWAP - "swap pager". Some ops are also global procedure
361 * calls hooked from other parts of the VM system and do not appear here.
362 * (see vm/swap_pager.h).
365 swap_pager_alloc(void *handle, vm_ooffset_t size,
366 vm_prot_t prot, vm_ooffset_t offset, struct ucred *);
367 static void swap_pager_dealloc(vm_object_t object);
368 static int swap_pager_getpages(vm_object_t, vm_page_t *, int, int *,
370 static int swap_pager_getpages_async(vm_object_t, vm_page_t *, int, int *,
371 int *, pgo_getpages_iodone_t, void *);
372 static void swap_pager_putpages(vm_object_t, vm_page_t *, int, boolean_t, int *);
374 swap_pager_haspage(vm_object_t object, vm_pindex_t pindex, int *before, int *after);
375 static void swap_pager_init(void);
376 static void swap_pager_unswapped(vm_page_t);
377 static void swap_pager_swapoff(struct swdevt *sp);
379 struct pagerops swappagerops = {
380 .pgo_init = swap_pager_init, /* early system initialization of pager */
381 .pgo_alloc = swap_pager_alloc, /* allocate an OBJT_SWAP object */
382 .pgo_dealloc = swap_pager_dealloc, /* deallocate an OBJT_SWAP object */
383 .pgo_getpages = swap_pager_getpages, /* pagein */
384 .pgo_getpages_async = swap_pager_getpages_async, /* pagein (async) */
385 .pgo_putpages = swap_pager_putpages, /* pageout */
386 .pgo_haspage = swap_pager_haspage, /* get backing store status for page */
387 .pgo_pageunswapped = swap_pager_unswapped, /* remove swap related to page */
391 * swap_*() routines are externally accessible. swp_*() routines are
394 static int nswap_lowat = 128; /* in pages, swap_pager_almost_full warn */
395 static int nswap_hiwat = 512; /* in pages, swap_pager_almost_full warn */
397 SYSCTL_INT(_vm, OID_AUTO, dmmax, CTLFLAG_RD, &nsw_cluster_max, 0,
398 "Maximum size of a swap block in pages");
400 static void swp_sizecheck(void);
401 static void swp_pager_async_iodone(struct buf *bp);
402 static bool swp_pager_swblk_empty(struct swblk *sb, int start, int limit);
403 static int swapongeom(struct vnode *);
404 static int swaponvp(struct thread *, struct vnode *, u_long);
405 static int swapoff_one(struct swdevt *sp, struct ucred *cred);
408 * Swap bitmap functions
410 static void swp_pager_freeswapspace(daddr_t blk, daddr_t npages);
411 static daddr_t swp_pager_getswapspace(int *npages, int limit);
416 static daddr_t swp_pager_meta_build(vm_object_t, vm_pindex_t, daddr_t);
417 static void swp_pager_meta_free(vm_object_t, vm_pindex_t, vm_pindex_t);
418 static void swp_pager_meta_free_all(vm_object_t);
419 static daddr_t swp_pager_meta_ctl(vm_object_t, vm_pindex_t, int);
422 swp_pager_init_freerange(daddr_t *start, daddr_t *num)
425 *start = SWAPBLK_NONE;
430 swp_pager_update_freerange(daddr_t *start, daddr_t *num, daddr_t addr)
433 if (*start + *num == addr) {
436 swp_pager_freeswapspace(*start, *num);
443 swblk_trie_alloc(struct pctrie *ptree)
446 return (uma_zalloc(swpctrie_zone, M_NOWAIT | (curproc == pageproc ?
447 M_USE_RESERVE : 0)));
451 swblk_trie_free(struct pctrie *ptree, void *node)
454 uma_zfree(swpctrie_zone, node);
457 PCTRIE_DEFINE(SWAP, swblk, p, swblk_trie_alloc, swblk_trie_free);
460 * SWP_SIZECHECK() - update swap_pager_full indication
462 * update the swap_pager_almost_full indication and warn when we are
463 * about to run out of swap space, using lowat/hiwat hysteresis.
465 * Clear swap_pager_full ( task killing ) indication when lowat is met.
467 * No restrictions on call
468 * This routine may not block.
474 if (swap_pager_avail < nswap_lowat) {
475 if (swap_pager_almost_full == 0) {
476 printf("swap_pager: out of swap space\n");
477 swap_pager_almost_full = 1;
481 if (swap_pager_avail > nswap_hiwat)
482 swap_pager_almost_full = 0;
487 * SWAP_PAGER_INIT() - initialize the swap pager!
489 * Expected to be started from system init. NOTE: This code is run
490 * before much else so be careful what you depend on. Most of the VM
491 * system has yet to be initialized at this point.
494 swap_pager_init(void)
497 * Initialize object lists
501 for (i = 0; i < NOBJLISTS; ++i)
502 TAILQ_INIT(&swap_pager_object_list[i]);
503 mtx_init(&sw_dev_mtx, "swapdev", NULL, MTX_DEF);
504 sx_init(&sw_alloc_sx, "swspsx");
505 sx_init(&swdev_syscall_lock, "swsysc");
509 * SWAP_PAGER_SWAP_INIT() - swap pager initialization from pageout process
511 * Expected to be started from pageout process once, prior to entering
515 swap_pager_swap_init(void)
520 * Number of in-transit swap bp operations. Don't
521 * exhaust the pbufs completely. Make sure we
522 * initialize workable values (0 will work for hysteresis
523 * but it isn't very efficient).
525 * The nsw_cluster_max is constrained by the bp->b_pages[]
526 * array (MAXPHYS/PAGE_SIZE) and our locally defined
527 * MAX_PAGEOUT_CLUSTER. Also be aware that swap ops are
528 * constrained by the swap device interleave stripe size.
530 * Currently we hardwire nsw_wcount_async to 4. This limit is
531 * designed to prevent other I/O from having high latencies due to
532 * our pageout I/O. The value 4 works well for one or two active swap
533 * devices but is probably a little low if you have more. Even so,
534 * a higher value would probably generate only a limited improvement
535 * with three or four active swap devices since the system does not
536 * typically have to pageout at extreme bandwidths. We will want
537 * at least 2 per swap devices, and 4 is a pretty good value if you
538 * have one NFS swap device due to the command/ack latency over NFS.
539 * So it all works out pretty well.
541 nsw_cluster_max = min((MAXPHYS/PAGE_SIZE), MAX_PAGEOUT_CLUSTER);
543 nsw_wcount_async = 4;
544 nsw_wcount_async_max = nsw_wcount_async;
545 mtx_init(&swbuf_mtx, "async swbuf mutex", NULL, MTX_DEF);
547 swwbuf_zone = pbuf_zsecond_create("swwbuf", nswbuf / 4);
548 swrbuf_zone = pbuf_zsecond_create("swrbuf", nswbuf / 2);
551 * Initialize our zone, taking the user's requested size or
552 * estimating the number we need based on the number of pages
555 n = maxswzone != 0 ? maxswzone / sizeof(struct swblk) :
556 vm_cnt.v_page_count / 2;
557 swpctrie_zone = uma_zcreate("swpctrie", pctrie_node_size(), NULL, NULL,
558 pctrie_zone_init, NULL, UMA_ALIGN_PTR, UMA_ZONE_VM);
559 if (swpctrie_zone == NULL)
560 panic("failed to create swap pctrie zone.");
561 swblk_zone = uma_zcreate("swblk", sizeof(struct swblk), NULL, NULL,
562 NULL, NULL, _Alignof(struct swblk) - 1, UMA_ZONE_VM);
563 if (swblk_zone == NULL)
564 panic("failed to create swap blk zone.");
567 if (uma_zone_reserve_kva(swblk_zone, n))
570 * if the allocation failed, try a zone two thirds the
571 * size of the previous attempt.
577 * Often uma_zone_reserve_kva() cannot reserve exactly the
578 * requested size. Account for the difference when
579 * calculating swap_maxpages.
581 n = uma_zone_get_max(swblk_zone);
584 printf("Swap blk zone entries changed from %lu to %lu.\n",
586 swap_maxpages = n * SWAP_META_PAGES;
587 swzone = n * sizeof(struct swblk);
588 if (!uma_zone_reserve_kva(swpctrie_zone, n))
589 printf("Cannot reserve swap pctrie zone, "
590 "reduce kern.maxswzone.\n");
594 swap_pager_alloc_init(void *handle, struct ucred *cred, vm_ooffset_t size,
600 if (!swap_reserve_by_cred(size, cred))
606 * The un_pager.swp.swp_blks trie is initialized by
607 * vm_object_allocate() to ensure the correct order of
608 * visibility to other threads.
610 object = vm_object_allocate(OBJT_SWAP, OFF_TO_IDX(offset +
613 object->handle = handle;
616 object->charge = size;
622 * SWAP_PAGER_ALLOC() - allocate a new OBJT_SWAP VM object and instantiate
623 * its metadata structures.
625 * This routine is called from the mmap and fork code to create a new
628 * This routine must ensure that no live duplicate is created for
629 * the named object request, which is protected against by
630 * holding the sw_alloc_sx lock in case handle != NULL.
633 swap_pager_alloc(void *handle, vm_ooffset_t size, vm_prot_t prot,
634 vm_ooffset_t offset, struct ucred *cred)
638 if (handle != NULL) {
640 * Reference existing named region or allocate new one. There
641 * should not be a race here against swp_pager_meta_build()
642 * as called from vm_page_remove() in regards to the lookup
645 sx_xlock(&sw_alloc_sx);
646 object = vm_pager_object_lookup(NOBJLIST(handle), handle);
647 if (object == NULL) {
648 object = swap_pager_alloc_init(handle, cred, size,
650 if (object != NULL) {
651 TAILQ_INSERT_TAIL(NOBJLIST(object->handle),
652 object, pager_object_list);
655 sx_xunlock(&sw_alloc_sx);
657 object = swap_pager_alloc_init(handle, cred, size, offset);
663 * SWAP_PAGER_DEALLOC() - remove swap metadata from object
665 * The swap backing for the object is destroyed. The code is
666 * designed such that we can reinstantiate it later, but this
667 * routine is typically called only when the entire object is
668 * about to be destroyed.
670 * The object must be locked.
673 swap_pager_dealloc(vm_object_t object)
676 VM_OBJECT_ASSERT_WLOCKED(object);
677 KASSERT((object->flags & OBJ_DEAD) != 0, ("dealloc of reachable obj"));
680 * Remove from list right away so lookups will fail if we block for
681 * pageout completion.
683 if (object->handle != NULL) {
684 VM_OBJECT_WUNLOCK(object);
685 sx_xlock(&sw_alloc_sx);
686 TAILQ_REMOVE(NOBJLIST(object->handle), object,
688 sx_xunlock(&sw_alloc_sx);
689 VM_OBJECT_WLOCK(object);
692 vm_object_pip_wait(object, "swpdea");
695 * Free all remaining metadata. We only bother to free it from
696 * the swap meta data. We do not attempt to free swapblk's still
697 * associated with vm_page_t's for this object. We do not care
698 * if paging is still in progress on some objects.
700 swp_pager_meta_free_all(object);
701 object->handle = NULL;
702 object->type = OBJT_DEAD;
705 /************************************************************************
706 * SWAP PAGER BITMAP ROUTINES *
707 ************************************************************************/
710 * SWP_PAGER_GETSWAPSPACE() - allocate raw swap space
712 * Allocate swap for up to the requested number of pages, and at
713 * least a minimum number of pages. The starting swap block number
714 * (a page index) is returned or SWAPBLK_NONE if the allocation
717 * Also has the side effect of advising that somebody made a mistake
718 * when they configured swap and didn't configure enough.
720 * This routine may not sleep.
722 * We allocate in round-robin fashion from the configured devices.
725 swp_pager_getswapspace(int *io_npages, int limit)
732 npages = mpages = *io_npages;
733 mtx_lock(&sw_dev_mtx);
735 while (!TAILQ_EMPTY(&swtailq)) {
737 sp = TAILQ_FIRST(&swtailq);
738 if ((sp->sw_flags & SW_CLOSING) == 0)
739 blk = blist_alloc(sp->sw_blist, &npages, mpages);
740 if (blk != SWAPBLK_NONE)
742 sp = TAILQ_NEXT(sp, sw_list);
750 if (blk != SWAPBLK_NONE) {
753 sp->sw_used += npages;
754 swap_pager_avail -= npages;
756 swdevhd = TAILQ_NEXT(sp, sw_list);
758 if (swap_pager_full != 2) {
759 printf("swp_pager_getswapspace(%d): failed\n",
762 swap_pager_almost_full = 1;
766 mtx_unlock(&sw_dev_mtx);
771 swp_pager_isondev(daddr_t blk, struct swdevt *sp)
774 return (blk >= sp->sw_first && blk < sp->sw_end);
778 swp_pager_strategy(struct buf *bp)
782 mtx_lock(&sw_dev_mtx);
783 TAILQ_FOREACH(sp, &swtailq, sw_list) {
784 if (swp_pager_isondev(bp->b_blkno, sp)) {
785 mtx_unlock(&sw_dev_mtx);
786 if ((sp->sw_flags & SW_UNMAPPED) != 0 &&
787 unmapped_buf_allowed) {
788 bp->b_data = unmapped_buf;
791 pmap_qenter((vm_offset_t)bp->b_data,
792 &bp->b_pages[0], bp->b_bcount / PAGE_SIZE);
794 sp->sw_strategy(bp, sp);
798 panic("Swapdev not found");
803 * SWP_PAGER_FREESWAPSPACE() - free raw swap space
805 * This routine returns the specified swap blocks back to the bitmap.
807 * This routine may not sleep.
810 swp_pager_freeswapspace(daddr_t blk, daddr_t npages)
816 mtx_lock(&sw_dev_mtx);
817 TAILQ_FOREACH(sp, &swtailq, sw_list) {
818 if (swp_pager_isondev(blk, sp)) {
819 sp->sw_used -= npages;
821 * If we are attempting to stop swapping on
822 * this device, we don't want to mark any
823 * blocks free lest they be reused.
825 if ((sp->sw_flags & SW_CLOSING) == 0) {
826 blist_free(sp->sw_blist, blk - sp->sw_first,
828 swap_pager_avail += npages;
831 mtx_unlock(&sw_dev_mtx);
835 panic("Swapdev not found");
839 * SYSCTL_SWAP_FRAGMENTATION() - produce raw swap space stats
842 sysctl_swap_fragmentation(SYSCTL_HANDLER_ARGS)
849 error = sysctl_wire_old_buffer(req, 0);
852 sbuf_new_for_sysctl(&sbuf, NULL, 128, req);
853 mtx_lock(&sw_dev_mtx);
854 TAILQ_FOREACH(sp, &swtailq, sw_list) {
855 if (vn_isdisk(sp->sw_vp, NULL))
856 devname = devtoname(sp->sw_vp->v_rdev);
859 sbuf_printf(&sbuf, "\nFree space on device %s:\n", devname);
860 blist_stats(sp->sw_blist, &sbuf);
862 mtx_unlock(&sw_dev_mtx);
863 error = sbuf_finish(&sbuf);
869 * SWAP_PAGER_FREESPACE() - frees swap blocks associated with a page
870 * range within an object.
872 * This is a globally accessible routine.
874 * This routine removes swapblk assignments from swap metadata.
876 * The external callers of this routine typically have already destroyed
877 * or renamed vm_page_t's associated with this range in the object so
880 * The object must be locked.
883 swap_pager_freespace(vm_object_t object, vm_pindex_t start, vm_size_t size)
886 swp_pager_meta_free(object, start, size);
890 * SWAP_PAGER_RESERVE() - reserve swap blocks in object
892 * Assigns swap blocks to the specified range within the object. The
893 * swap blocks are not zeroed. Any previous swap assignment is destroyed.
895 * Returns 0 on success, -1 on failure.
898 swap_pager_reserve(vm_object_t object, vm_pindex_t start, vm_size_t size)
900 daddr_t addr, blk, n_free, s_free;
903 swp_pager_init_freerange(&s_free, &n_free);
904 VM_OBJECT_WLOCK(object);
905 for (i = 0; i < size; i += n) {
906 n = min(BLIST_MAX_ALLOC, size - i);
907 blk = swp_pager_getswapspace(&n, 1);
908 if (blk == SWAPBLK_NONE) {
909 swp_pager_meta_free(object, start, i);
910 VM_OBJECT_WUNLOCK(object);
913 for (j = 0; j < n; ++j) {
914 addr = swp_pager_meta_build(object,
915 start + i + j, blk + j);
916 if (addr != SWAPBLK_NONE)
917 swp_pager_update_freerange(&s_free, &n_free,
921 swp_pager_freeswapspace(s_free, n_free);
922 VM_OBJECT_WUNLOCK(object);
927 * SWAP_PAGER_COPY() - copy blocks from source pager to destination pager
928 * and destroy the source.
930 * Copy any valid swapblks from the source to the destination. In
931 * cases where both the source and destination have a valid swapblk,
932 * we keep the destination's.
934 * This routine is allowed to sleep. It may sleep allocating metadata
935 * indirectly through swp_pager_meta_build() or if paging is still in
936 * progress on the source.
938 * The source object contains no vm_page_t's (which is just as well)
940 * The source object is of type OBJT_SWAP.
942 * The source and destination objects must be locked.
943 * Both object locks may temporarily be released.
946 swap_pager_copy(vm_object_t srcobject, vm_object_t dstobject,
947 vm_pindex_t offset, int destroysource)
950 daddr_t dstaddr, n_free, s_free, srcaddr;
952 VM_OBJECT_ASSERT_WLOCKED(srcobject);
953 VM_OBJECT_ASSERT_WLOCKED(dstobject);
956 * If destroysource is set, we remove the source object from the
957 * swap_pager internal queue now.
959 if (destroysource && srcobject->handle != NULL) {
960 vm_object_pip_add(srcobject, 1);
961 VM_OBJECT_WUNLOCK(srcobject);
962 vm_object_pip_add(dstobject, 1);
963 VM_OBJECT_WUNLOCK(dstobject);
964 sx_xlock(&sw_alloc_sx);
965 TAILQ_REMOVE(NOBJLIST(srcobject->handle), srcobject,
967 sx_xunlock(&sw_alloc_sx);
968 VM_OBJECT_WLOCK(dstobject);
969 vm_object_pip_wakeup(dstobject);
970 VM_OBJECT_WLOCK(srcobject);
971 vm_object_pip_wakeup(srcobject);
975 * Transfer source to destination.
977 swp_pager_init_freerange(&s_free, &n_free);
978 for (i = 0; i < dstobject->size; ++i) {
979 srcaddr = swp_pager_meta_ctl(srcobject, i + offset, SWM_POP);
980 if (srcaddr == SWAPBLK_NONE)
982 dstaddr = swp_pager_meta_ctl(dstobject, i, 0);
983 if (dstaddr != SWAPBLK_NONE) {
985 * Destination has valid swapblk or it is represented
986 * by a resident page. We destroy the source block.
988 swp_pager_update_freerange(&s_free, &n_free, srcaddr);
993 * Destination has no swapblk and is not resident,
996 * swp_pager_meta_build() can sleep.
998 vm_object_pip_add(srcobject, 1);
999 VM_OBJECT_WUNLOCK(srcobject);
1000 vm_object_pip_add(dstobject, 1);
1001 dstaddr = swp_pager_meta_build(dstobject, i, srcaddr);
1002 KASSERT(dstaddr == SWAPBLK_NONE,
1003 ("Unexpected destination swapblk"));
1004 vm_object_pip_wakeup(dstobject);
1005 VM_OBJECT_WLOCK(srcobject);
1006 vm_object_pip_wakeup(srcobject);
1008 swp_pager_freeswapspace(s_free, n_free);
1011 * Free left over swap blocks in source.
1013 * We have to revert the type to OBJT_DEFAULT so we do not accidentally
1014 * double-remove the object from the swap queues.
1016 if (destroysource) {
1017 swp_pager_meta_free_all(srcobject);
1019 * Reverting the type is not necessary, the caller is going
1020 * to destroy srcobject directly, but I'm doing it here
1021 * for consistency since we've removed the object from its
1024 srcobject->type = OBJT_DEFAULT;
1029 * SWAP_PAGER_HASPAGE() - determine if we have good backing store for
1030 * the requested page.
1032 * We determine whether good backing store exists for the requested
1033 * page and return TRUE if it does, FALSE if it doesn't.
1035 * If TRUE, we also try to determine how much valid, contiguous backing
1036 * store exists before and after the requested page.
1039 swap_pager_haspage(vm_object_t object, vm_pindex_t pindex, int *before,
1045 VM_OBJECT_ASSERT_LOCKED(object);
1048 * do we have good backing store at the requested index ?
1050 blk0 = swp_pager_meta_ctl(object, pindex, 0);
1051 if (blk0 == SWAPBLK_NONE) {
1060 * find backwards-looking contiguous good backing store
1062 if (before != NULL) {
1063 for (i = 1; i < SWB_NPAGES; i++) {
1066 blk = swp_pager_meta_ctl(object, pindex - i, 0);
1067 if (blk != blk0 - i)
1074 * find forward-looking contiguous good backing store
1076 if (after != NULL) {
1077 for (i = 1; i < SWB_NPAGES; i++) {
1078 blk = swp_pager_meta_ctl(object, pindex + i, 0);
1079 if (blk != blk0 + i)
1088 * SWAP_PAGER_PAGE_UNSWAPPED() - remove swap backing store related to page
1090 * This removes any associated swap backing store, whether valid or
1091 * not, from the page.
1093 * This routine is typically called when a page is made dirty, at
1094 * which point any associated swap can be freed. MADV_FREE also
1095 * calls us in a special-case situation
1097 * NOTE!!! If the page is clean and the swap was valid, the caller
1098 * should make the page dirty before calling this routine. This routine
1099 * does NOT change the m->dirty status of the page. Also: MADV_FREE
1102 * This routine may not sleep.
1104 * The object containing the page must be locked.
1107 swap_pager_unswapped(vm_page_t m)
1111 srcaddr = swp_pager_meta_ctl(m->object, m->pindex, SWM_POP);
1112 if (srcaddr != SWAPBLK_NONE)
1113 swp_pager_freeswapspace(srcaddr, 1);
1117 * swap_pager_getpages() - bring pages in from swap
1119 * Attempt to page in the pages in array "ma" of length "count". The
1120 * caller may optionally specify that additional pages preceding and
1121 * succeeding the specified range be paged in. The number of such pages
1122 * is returned in the "rbehind" and "rahead" parameters, and they will
1123 * be in the inactive queue upon return.
1125 * The pages in "ma" must be busied and will remain busied upon return.
1128 swap_pager_getpages(vm_object_t object, vm_page_t *ma, int count, int *rbehind,
1132 vm_page_t bm, mpred, msucc, p;
1135 int i, maxahead, maxbehind, reqcount;
1140 * Determine the final number of read-behind pages and
1141 * allocate them BEFORE releasing the object lock. Otherwise,
1142 * there can be a problematic race with vm_object_split().
1143 * Specifically, vm_object_split() might first transfer pages
1144 * that precede ma[0] in the current object to a new object,
1145 * and then this function incorrectly recreates those pages as
1146 * read-behind pages in the current object.
1148 if (!swap_pager_haspage(object, ma[0]->pindex, &maxbehind, &maxahead))
1149 return (VM_PAGER_FAIL);
1152 * Clip the readahead and readbehind ranges to exclude resident pages.
1154 if (rahead != NULL) {
1155 KASSERT(reqcount - 1 <= maxahead,
1156 ("page count %d extends beyond swap block", reqcount));
1157 *rahead = imin(*rahead, maxahead - (reqcount - 1));
1158 pindex = ma[reqcount - 1]->pindex;
1159 msucc = TAILQ_NEXT(ma[reqcount - 1], listq);
1160 if (msucc != NULL && msucc->pindex - pindex - 1 < *rahead)
1161 *rahead = msucc->pindex - pindex - 1;
1163 if (rbehind != NULL) {
1164 *rbehind = imin(*rbehind, maxbehind);
1165 pindex = ma[0]->pindex;
1166 mpred = TAILQ_PREV(ma[0], pglist, listq);
1167 if (mpred != NULL && pindex - mpred->pindex - 1 < *rbehind)
1168 *rbehind = pindex - mpred->pindex - 1;
1172 for (i = 0; i < count; i++)
1173 ma[i]->oflags |= VPO_SWAPINPROG;
1176 * Allocate readahead and readbehind pages.
1178 if (rbehind != NULL) {
1179 for (i = 1; i <= *rbehind; i++) {
1180 p = vm_page_alloc(object, ma[0]->pindex - i,
1184 p->oflags |= VPO_SWAPINPROG;
1189 if (rahead != NULL) {
1190 for (i = 0; i < *rahead; i++) {
1191 p = vm_page_alloc(object,
1192 ma[reqcount - 1]->pindex + i + 1, VM_ALLOC_NORMAL);
1195 p->oflags |= VPO_SWAPINPROG;
1199 if (rbehind != NULL)
1204 vm_object_pip_add(object, count);
1206 pindex = bm->pindex;
1207 blk = swp_pager_meta_ctl(object, pindex, 0);
1208 KASSERT(blk != SWAPBLK_NONE,
1209 ("no swap blocking containing %p(%jx)", object, (uintmax_t)pindex));
1211 VM_OBJECT_WUNLOCK(object);
1212 bp = uma_zalloc(swrbuf_zone, M_WAITOK);
1213 /* Pages cannot leave the object while busy. */
1214 for (i = 0, p = bm; i < count; i++, p = TAILQ_NEXT(p, listq)) {
1215 MPASS(p->pindex == bm->pindex + i);
1219 bp->b_flags |= B_PAGING;
1220 bp->b_iocmd = BIO_READ;
1221 bp->b_iodone = swp_pager_async_iodone;
1222 bp->b_rcred = crhold(thread0.td_ucred);
1223 bp->b_wcred = crhold(thread0.td_ucred);
1225 bp->b_bcount = PAGE_SIZE * count;
1226 bp->b_bufsize = PAGE_SIZE * count;
1227 bp->b_npages = count;
1228 bp->b_pgbefore = rbehind != NULL ? *rbehind : 0;
1229 bp->b_pgafter = rahead != NULL ? *rahead : 0;
1231 VM_CNT_INC(v_swapin);
1232 VM_CNT_ADD(v_swappgsin, count);
1235 * perform the I/O. NOTE!!! bp cannot be considered valid after
1236 * this point because we automatically release it on completion.
1237 * Instead, we look at the one page we are interested in which we
1238 * still hold a lock on even through the I/O completion.
1240 * The other pages in our ma[] array are also released on completion,
1241 * so we cannot assume they are valid anymore either.
1243 * NOTE: b_blkno is destroyed by the call to swapdev_strategy
1246 swp_pager_strategy(bp);
1249 * Wait for the pages we want to complete. VPO_SWAPINPROG is always
1250 * cleared on completion. If an I/O error occurs, SWAPBLK_NONE
1251 * is set in the metadata for each page in the request.
1253 VM_OBJECT_WLOCK(object);
1254 while ((ma[0]->oflags & VPO_SWAPINPROG) != 0) {
1255 ma[0]->oflags |= VPO_SWAPSLEEP;
1256 VM_CNT_INC(v_intrans);
1257 if (VM_OBJECT_SLEEP(object, &object->paging_in_progress, PSWP,
1258 "swread", hz * 20)) {
1260 "swap_pager: indefinite wait buffer: bufobj: %p, blkno: %jd, size: %ld\n",
1261 bp->b_bufobj, (intmax_t)bp->b_blkno, bp->b_bcount);
1266 * If we had an unrecoverable read error pages will not be valid.
1268 for (i = 0; i < reqcount; i++)
1269 if (ma[i]->valid != VM_PAGE_BITS_ALL)
1270 return (VM_PAGER_ERROR);
1272 return (VM_PAGER_OK);
1275 * A final note: in a low swap situation, we cannot deallocate swap
1276 * and mark a page dirty here because the caller is likely to mark
1277 * the page clean when we return, causing the page to possibly revert
1278 * to all-zero's later.
1283 * swap_pager_getpages_async():
1285 * Right now this is emulation of asynchronous operation on top of
1286 * swap_pager_getpages().
1289 swap_pager_getpages_async(vm_object_t object, vm_page_t *ma, int count,
1290 int *rbehind, int *rahead, pgo_getpages_iodone_t iodone, void *arg)
1294 r = swap_pager_getpages(object, ma, count, rbehind, rahead);
1295 VM_OBJECT_WUNLOCK(object);
1300 case VM_PAGER_ERROR:
1307 panic("unhandled swap_pager_getpages() error %d", r);
1309 (iodone)(arg, ma, count, error);
1310 VM_OBJECT_WLOCK(object);
1316 * swap_pager_putpages:
1318 * Assign swap (if necessary) and initiate I/O on the specified pages.
1320 * We support both OBJT_DEFAULT and OBJT_SWAP objects. DEFAULT objects
1321 * are automatically converted to SWAP objects.
1323 * In a low memory situation we may block in VOP_STRATEGY(), but the new
1324 * vm_page reservation system coupled with properly written VFS devices
1325 * should ensure that no low-memory deadlock occurs. This is an area
1328 * The parent has N vm_object_pip_add() references prior to
1329 * calling us and will remove references for rtvals[] that are
1330 * not set to VM_PAGER_PEND. We need to remove the rest on I/O
1333 * The parent has soft-busy'd the pages it passes us and will unbusy
1334 * those whos rtvals[] entry is not set to VM_PAGER_PEND on return.
1335 * We need to unbusy the rest on I/O completion.
1338 swap_pager_putpages(vm_object_t object, vm_page_t *ma, int count,
1339 int flags, int *rtvals)
1343 daddr_t addr, n_free, s_free;
1345 swp_pager_init_freerange(&s_free, &n_free);
1346 if (count && ma[0]->object != object) {
1347 panic("swap_pager_putpages: object mismatch %p/%p",
1356 * Turn object into OBJT_SWAP
1357 * check for bogus sysops
1358 * force sync if not pageout process
1360 if (object->type != OBJT_SWAP) {
1361 addr = swp_pager_meta_build(object, 0, SWAPBLK_NONE);
1362 KASSERT(addr == SWAPBLK_NONE,
1363 ("unexpected object swap block"));
1365 VM_OBJECT_WUNLOCK(object);
1368 if (curproc != pageproc)
1371 sync = (flags & VM_PAGER_PUT_SYNC) != 0;
1376 * Assign swap blocks and issue I/O. We reallocate swap on the fly.
1377 * The page is left dirty until the pageout operation completes
1380 for (i = 0; i < count; i += n) {
1386 * Maximum I/O size is limited by a number of factors.
1388 n = min(BLIST_MAX_ALLOC, count - i);
1389 n = min(n, nsw_cluster_max);
1391 /* Get a block of swap of size up to size n. */
1392 blk = swp_pager_getswapspace(&n, 4);
1393 if (blk == SWAPBLK_NONE) {
1394 for (j = 0; j < n; ++j)
1395 rtvals[i+j] = VM_PAGER_FAIL;
1400 * All I/O parameters have been satisfied, build the I/O
1401 * request and assign the swap space.
1404 mtx_lock(&swbuf_mtx);
1405 while (nsw_wcount_async == 0)
1406 msleep(&nsw_wcount_async, &swbuf_mtx, PVM,
1409 mtx_unlock(&swbuf_mtx);
1411 bp = uma_zalloc(swwbuf_zone, M_WAITOK);
1413 bp->b_flags = B_ASYNC;
1414 bp->b_flags |= B_PAGING;
1415 bp->b_iocmd = BIO_WRITE;
1417 bp->b_rcred = crhold(thread0.td_ucred);
1418 bp->b_wcred = crhold(thread0.td_ucred);
1419 bp->b_bcount = PAGE_SIZE * n;
1420 bp->b_bufsize = PAGE_SIZE * n;
1423 VM_OBJECT_WLOCK(object);
1424 for (j = 0; j < n; ++j) {
1425 vm_page_t mreq = ma[i+j];
1427 addr = swp_pager_meta_build(mreq->object, mreq->pindex,
1429 if (addr != SWAPBLK_NONE)
1430 swp_pager_update_freerange(&s_free, &n_free,
1432 MPASS(mreq->dirty == VM_PAGE_BITS_ALL);
1433 mreq->oflags |= VPO_SWAPINPROG;
1434 bp->b_pages[j] = mreq;
1436 VM_OBJECT_WUNLOCK(object);
1439 * Must set dirty range for NFS to work.
1442 bp->b_dirtyend = bp->b_bcount;
1444 VM_CNT_INC(v_swapout);
1445 VM_CNT_ADD(v_swappgsout, bp->b_npages);
1448 * We unconditionally set rtvals[] to VM_PAGER_PEND so that we
1449 * can call the async completion routine at the end of a
1450 * synchronous I/O operation. Otherwise, our caller would
1451 * perform duplicate unbusy and wakeup operations on the page
1452 * and object, respectively.
1454 for (j = 0; j < n; j++)
1455 rtvals[i + j] = VM_PAGER_PEND;
1460 * NOTE: b_blkno is destroyed by the call to swapdev_strategy
1462 if (sync == FALSE) {
1463 bp->b_iodone = swp_pager_async_iodone;
1465 swp_pager_strategy(bp);
1472 * NOTE: b_blkno is destroyed by the call to swapdev_strategy
1474 bp->b_iodone = bdone;
1475 swp_pager_strategy(bp);
1478 * Wait for the sync I/O to complete.
1480 bwait(bp, PVM, "swwrt");
1483 * Now that we are through with the bp, we can call the
1484 * normal async completion, which frees everything up.
1486 swp_pager_async_iodone(bp);
1488 VM_OBJECT_WLOCK(object);
1489 swp_pager_freeswapspace(s_free, n_free);
1493 * swp_pager_async_iodone:
1495 * Completion routine for asynchronous reads and writes from/to swap.
1496 * Also called manually by synchronous code to finish up a bp.
1498 * This routine may not sleep.
1501 swp_pager_async_iodone(struct buf *bp)
1504 vm_object_t object = NULL;
1507 * Report error - unless we ran out of memory, in which case
1508 * we've already logged it in swapgeom_strategy().
1510 if (bp->b_ioflags & BIO_ERROR && bp->b_error != ENOMEM) {
1512 "swap_pager: I/O error - %s failed; blkno %ld,"
1513 "size %ld, error %d\n",
1514 ((bp->b_iocmd == BIO_READ) ? "pagein" : "pageout"),
1522 * remove the mapping for kernel virtual
1525 pmap_qremove((vm_offset_t)bp->b_data, bp->b_npages);
1527 bp->b_data = bp->b_kvabase;
1530 object = bp->b_pages[0]->object;
1531 VM_OBJECT_WLOCK(object);
1535 * cleanup pages. If an error occurs writing to swap, we are in
1536 * very serious trouble. If it happens to be a disk error, though,
1537 * we may be able to recover by reassigning the swap later on. So
1538 * in this case we remove the m->swapblk assignment for the page
1539 * but do not free it in the rlist. The errornous block(s) are thus
1540 * never reallocated as swap. Redirty the page and continue.
1542 for (i = 0; i < bp->b_npages; ++i) {
1543 vm_page_t m = bp->b_pages[i];
1545 m->oflags &= ~VPO_SWAPINPROG;
1546 if (m->oflags & VPO_SWAPSLEEP) {
1547 m->oflags &= ~VPO_SWAPSLEEP;
1548 wakeup(&object->paging_in_progress);
1551 if (bp->b_ioflags & BIO_ERROR) {
1553 * If an error occurs I'd love to throw the swapblk
1554 * away without freeing it back to swapspace, so it
1555 * can never be used again. But I can't from an
1558 if (bp->b_iocmd == BIO_READ) {
1560 * NOTE: for reads, m->dirty will probably
1561 * be overridden by the original caller of
1562 * getpages so don't play cute tricks here.
1567 * If a write error occurs, reactivate page
1568 * so it doesn't clog the inactive list,
1569 * then finish the I/O.
1571 MPASS(m->dirty == VM_PAGE_BITS_ALL);
1573 vm_page_activate(m);
1577 } else if (bp->b_iocmd == BIO_READ) {
1579 * NOTE: for reads, m->dirty will probably be
1580 * overridden by the original caller of getpages so
1581 * we cannot set them in order to free the underlying
1582 * swap in a low-swap situation. I don't think we'd
1583 * want to do that anyway, but it was an optimization
1584 * that existed in the old swapper for a time before
1585 * it got ripped out due to precisely this problem.
1587 KASSERT(!pmap_page_is_mapped(m),
1588 ("swp_pager_async_iodone: page %p is mapped", m));
1589 KASSERT(m->dirty == 0,
1590 ("swp_pager_async_iodone: page %p is dirty", m));
1592 m->valid = VM_PAGE_BITS_ALL;
1593 if (i < bp->b_pgbefore ||
1594 i >= bp->b_npages - bp->b_pgafter)
1595 vm_page_readahead_finish(m);
1598 * For write success, clear the dirty
1599 * status, then finish the I/O ( which decrements the
1600 * busy count and possibly wakes waiter's up ).
1601 * A page is only written to swap after a period of
1602 * inactivity. Therefore, we do not expect it to be
1605 KASSERT(!pmap_page_is_write_mapped(m),
1606 ("swp_pager_async_iodone: page %p is not write"
1610 vm_page_deactivate_noreuse(m);
1617 * adjust pip. NOTE: the original parent may still have its own
1618 * pip refs on the object.
1620 if (object != NULL) {
1621 vm_object_pip_wakeupn(object, bp->b_npages);
1622 VM_OBJECT_WUNLOCK(object);
1626 * swapdev_strategy() manually sets b_vp and b_bufobj before calling
1627 * bstrategy(). Set them back to NULL now we're done with it, or we'll
1628 * trigger a KASSERT in relpbuf().
1632 bp->b_bufobj = NULL;
1635 * release the physical I/O buffer
1637 if (bp->b_flags & B_ASYNC) {
1638 mtx_lock(&swbuf_mtx);
1639 if (++nsw_wcount_async == 1)
1640 wakeup(&nsw_wcount_async);
1641 mtx_unlock(&swbuf_mtx);
1643 uma_zfree((bp->b_iocmd == BIO_READ) ? swrbuf_zone : swwbuf_zone, bp);
1647 swap_pager_nswapdev(void)
1654 * SWP_PAGER_FORCE_PAGEIN() - force a swap block to be paged in
1656 * This routine dissociates the page at the given index within an object
1657 * from its backing store, paging it in if it does not reside in memory.
1658 * If the page is paged in, it is marked dirty and placed in the laundry
1659 * queue. The page is marked dirty because it no longer has backing
1660 * store. It is placed in the laundry queue because it has not been
1661 * accessed recently. Otherwise, it would already reside in memory.
1663 * We also attempt to swap in all other pages in the swap block.
1664 * However, we only guarantee that the one at the specified index is
1667 * XXX - The code to page the whole block in doesn't work, so we
1668 * revert to the one-by-one behavior for now. Sigh.
1671 swp_pager_force_pagein(vm_object_t object, vm_pindex_t pindex)
1675 vm_object_pip_add(object, 1);
1676 m = vm_page_grab(object, pindex, VM_ALLOC_NORMAL);
1677 if (m->valid == VM_PAGE_BITS_ALL) {
1678 vm_object_pip_wakeup(object);
1682 if (m->wire_count == 0 && m->queue == PQ_NONE)
1683 panic("page %p is neither wired nor queued", m);
1687 vm_pager_page_unswapped(m);
1691 if (swap_pager_getpages(object, &m, 1, NULL, NULL) != VM_PAGER_OK)
1692 panic("swap_pager_force_pagein: read from swap failed");/*XXX*/
1693 vm_object_pip_wakeup(object);
1699 vm_pager_page_unswapped(m);
1703 * swap_pager_swapoff:
1705 * Page in all of the pages that have been paged out to the
1706 * given device. The corresponding blocks in the bitmap must be
1707 * marked as allocated and the device must be flagged SW_CLOSING.
1708 * There may be no processes swapped out to the device.
1710 * This routine may block.
1713 swap_pager_swapoff(struct swdevt *sp)
1720 sx_assert(&swdev_syscall_lock, SA_XLOCKED);
1724 mtx_lock(&vm_object_list_mtx);
1725 TAILQ_FOREACH(object, &vm_object_list, object_list) {
1726 if (object->type != OBJT_SWAP)
1728 mtx_unlock(&vm_object_list_mtx);
1729 /* Depends on type-stability. */
1730 VM_OBJECT_WLOCK(object);
1733 * Dead objects are eventually terminated on their own.
1735 if ((object->flags & OBJ_DEAD) != 0)
1739 * Sync with fences placed after pctrie
1740 * initialization. We must not access pctrie below
1741 * unless we checked that our object is swap and not
1744 atomic_thread_fence_acq();
1745 if (object->type != OBJT_SWAP)
1748 for (pi = 0; (sb = SWAP_PCTRIE_LOOKUP_GE(
1749 &object->un_pager.swp.swp_blks, pi)) != NULL; ) {
1750 pi = sb->p + SWAP_META_PAGES;
1751 for (i = 0; i < SWAP_META_PAGES; i++) {
1752 if (sb->d[i] == SWAPBLK_NONE)
1754 if (swp_pager_isondev(sb->d[i], sp))
1755 swp_pager_force_pagein(object,
1760 VM_OBJECT_WUNLOCK(object);
1761 mtx_lock(&vm_object_list_mtx);
1763 mtx_unlock(&vm_object_list_mtx);
1767 * Objects may be locked or paging to the device being
1768 * removed, so we will miss their pages and need to
1769 * make another pass. We have marked this device as
1770 * SW_CLOSING, so the activity should finish soon.
1773 if (retries > 100) {
1774 panic("swapoff: failed to locate %d swap blocks",
1777 pause("swpoff", hz / 20);
1780 EVENTHANDLER_INVOKE(swapoff, sp);
1783 /************************************************************************
1785 ************************************************************************
1787 * These routines manipulate the swap metadata stored in the
1790 * Swap metadata is implemented with a global hash and not directly
1791 * linked into the object. Instead the object simply contains
1792 * appropriate tracking counters.
1796 * SWP_PAGER_SWBLK_EMPTY() - is a range of blocks free?
1799 swp_pager_swblk_empty(struct swblk *sb, int start, int limit)
1803 MPASS(0 <= start && start <= limit && limit <= SWAP_META_PAGES);
1804 for (i = start; i < limit; i++) {
1805 if (sb->d[i] != SWAPBLK_NONE)
1812 * SWP_PAGER_META_BUILD() - add swap block to swap meta data for object
1814 * We first convert the object to a swap object if it is a default
1817 * The specified swapblk is added to the object's swap metadata. If
1818 * the swapblk is not valid, it is freed instead. Any previously
1819 * assigned swapblk is returned.
1822 swp_pager_meta_build(vm_object_t object, vm_pindex_t pindex, daddr_t swapblk)
1824 static volatile int swblk_zone_exhausted, swpctrie_zone_exhausted;
1825 struct swblk *sb, *sb1;
1826 vm_pindex_t modpi, rdpi;
1827 daddr_t prev_swapblk;
1830 VM_OBJECT_ASSERT_WLOCKED(object);
1833 * Convert default object to swap object if necessary
1835 if (object->type != OBJT_SWAP) {
1836 pctrie_init(&object->un_pager.swp.swp_blks);
1839 * Ensure that swap_pager_swapoff()'s iteration over
1840 * object_list does not see a garbage pctrie.
1842 atomic_thread_fence_rel();
1844 object->type = OBJT_SWAP;
1845 KASSERT(object->handle == NULL, ("default pager with handle"));
1848 rdpi = rounddown(pindex, SWAP_META_PAGES);
1849 sb = SWAP_PCTRIE_LOOKUP(&object->un_pager.swp.swp_blks, rdpi);
1851 if (swapblk == SWAPBLK_NONE)
1852 return (SWAPBLK_NONE);
1854 sb = uma_zalloc(swblk_zone, M_NOWAIT | (curproc ==
1855 pageproc ? M_USE_RESERVE : 0));
1858 for (i = 0; i < SWAP_META_PAGES; i++)
1859 sb->d[i] = SWAPBLK_NONE;
1860 if (atomic_cmpset_int(&swblk_zone_exhausted,
1862 printf("swblk zone ok\n");
1865 VM_OBJECT_WUNLOCK(object);
1866 if (uma_zone_exhausted(swblk_zone)) {
1867 if (atomic_cmpset_int(&swblk_zone_exhausted,
1869 printf("swap blk zone exhausted, "
1870 "increase kern.maxswzone\n");
1871 vm_pageout_oom(VM_OOM_SWAPZ);
1872 pause("swzonxb", 10);
1874 uma_zwait(swblk_zone);
1875 VM_OBJECT_WLOCK(object);
1876 sb = SWAP_PCTRIE_LOOKUP(&object->un_pager.swp.swp_blks,
1880 * Somebody swapped out a nearby page,
1881 * allocating swblk at the rdpi index,
1882 * while we dropped the object lock.
1887 error = SWAP_PCTRIE_INSERT(
1888 &object->un_pager.swp.swp_blks, sb);
1890 if (atomic_cmpset_int(&swpctrie_zone_exhausted,
1892 printf("swpctrie zone ok\n");
1895 VM_OBJECT_WUNLOCK(object);
1896 if (uma_zone_exhausted(swpctrie_zone)) {
1897 if (atomic_cmpset_int(&swpctrie_zone_exhausted,
1899 printf("swap pctrie zone exhausted, "
1900 "increase kern.maxswzone\n");
1901 vm_pageout_oom(VM_OOM_SWAPZ);
1902 pause("swzonxp", 10);
1904 uma_zwait(swpctrie_zone);
1905 VM_OBJECT_WLOCK(object);
1906 sb1 = SWAP_PCTRIE_LOOKUP(&object->un_pager.swp.swp_blks,
1909 uma_zfree(swblk_zone, sb);
1916 MPASS(sb->p == rdpi);
1918 modpi = pindex % SWAP_META_PAGES;
1919 /* Return prior contents of metadata. */
1920 prev_swapblk = sb->d[modpi];
1921 /* Enter block into metadata. */
1922 sb->d[modpi] = swapblk;
1925 * Free the swblk if we end up with the empty page run.
1927 if (swapblk == SWAPBLK_NONE &&
1928 swp_pager_swblk_empty(sb, 0, SWAP_META_PAGES)) {
1929 SWAP_PCTRIE_REMOVE(&object->un_pager.swp.swp_blks, rdpi);
1930 uma_zfree(swblk_zone, sb);
1932 return (prev_swapblk);
1936 * SWP_PAGER_META_FREE() - free a range of blocks in the object's swap metadata
1938 * The requested range of blocks is freed, with any associated swap
1939 * returned to the swap bitmap.
1941 * This routine will free swap metadata structures as they are cleaned
1942 * out. This routine does *NOT* operate on swap metadata associated
1943 * with resident pages.
1946 swp_pager_meta_free(vm_object_t object, vm_pindex_t pindex, vm_pindex_t count)
1949 daddr_t n_free, s_free;
1951 int i, limit, start;
1953 VM_OBJECT_ASSERT_WLOCKED(object);
1954 if (object->type != OBJT_SWAP || count == 0)
1957 swp_pager_init_freerange(&s_free, &n_free);
1958 last = pindex + count;
1960 sb = SWAP_PCTRIE_LOOKUP_GE(&object->un_pager.swp.swp_blks,
1961 rounddown(pindex, SWAP_META_PAGES));
1962 if (sb == NULL || sb->p >= last)
1964 start = pindex > sb->p ? pindex - sb->p : 0;
1965 limit = last - sb->p < SWAP_META_PAGES ? last - sb->p :
1967 for (i = start; i < limit; i++) {
1968 if (sb->d[i] == SWAPBLK_NONE)
1970 swp_pager_update_freerange(&s_free, &n_free, sb->d[i]);
1971 sb->d[i] = SWAPBLK_NONE;
1973 pindex = sb->p + SWAP_META_PAGES;
1974 if (swp_pager_swblk_empty(sb, 0, start) &&
1975 swp_pager_swblk_empty(sb, limit, SWAP_META_PAGES)) {
1976 SWAP_PCTRIE_REMOVE(&object->un_pager.swp.swp_blks,
1978 uma_zfree(swblk_zone, sb);
1981 swp_pager_freeswapspace(s_free, n_free);
1985 * SWP_PAGER_META_FREE_ALL() - destroy all swap metadata associated with object
1987 * This routine locates and destroys all swap metadata associated with
1991 swp_pager_meta_free_all(vm_object_t object)
1994 daddr_t n_free, s_free;
1998 VM_OBJECT_ASSERT_WLOCKED(object);
1999 if (object->type != OBJT_SWAP)
2002 swp_pager_init_freerange(&s_free, &n_free);
2003 for (pindex = 0; (sb = SWAP_PCTRIE_LOOKUP_GE(
2004 &object->un_pager.swp.swp_blks, pindex)) != NULL;) {
2005 pindex = sb->p + SWAP_META_PAGES;
2006 for (i = 0; i < SWAP_META_PAGES; i++) {
2007 if (sb->d[i] == SWAPBLK_NONE)
2009 swp_pager_update_freerange(&s_free, &n_free, sb->d[i]);
2011 SWAP_PCTRIE_REMOVE(&object->un_pager.swp.swp_blks, sb->p);
2012 uma_zfree(swblk_zone, sb);
2014 swp_pager_freeswapspace(s_free, n_free);
2018 * SWP_PAGER_METACTL() - misc control of swap meta data.
2020 * This routine is capable of looking up, or removing swapblk
2021 * assignments in the swap meta data. It returns the swapblk being
2022 * looked-up, popped, or SWAPBLK_NONE if the block was invalid.
2024 * When acting on a busy resident page and paging is in progress, we
2025 * have to wait until paging is complete but otherwise can act on the
2028 * SWM_POP remove from meta data but do not free it
2031 swp_pager_meta_ctl(vm_object_t object, vm_pindex_t pindex, int flags)
2036 if ((flags & SWM_POP) != 0)
2037 VM_OBJECT_ASSERT_WLOCKED(object);
2039 VM_OBJECT_ASSERT_LOCKED(object);
2042 * The meta data only exists if the object is OBJT_SWAP
2043 * and even then might not be allocated yet.
2045 if (object->type != OBJT_SWAP)
2046 return (SWAPBLK_NONE);
2048 sb = SWAP_PCTRIE_LOOKUP(&object->un_pager.swp.swp_blks,
2049 rounddown(pindex, SWAP_META_PAGES));
2051 return (SWAPBLK_NONE);
2052 r1 = sb->d[pindex % SWAP_META_PAGES];
2053 if (r1 == SWAPBLK_NONE)
2054 return (SWAPBLK_NONE);
2055 if ((flags & SWM_POP) != 0) {
2056 sb->d[pindex % SWAP_META_PAGES] = SWAPBLK_NONE;
2057 if (swp_pager_swblk_empty(sb, 0, SWAP_META_PAGES)) {
2058 SWAP_PCTRIE_REMOVE(&object->un_pager.swp.swp_blks,
2059 rounddown(pindex, SWAP_META_PAGES));
2060 uma_zfree(swblk_zone, sb);
2067 * Returns the least page index which is greater than or equal to the
2068 * parameter pindex and for which there is a swap block allocated.
2069 * Returns object's size if the object's type is not swap or if there
2070 * are no allocated swap blocks for the object after the requested
2074 swap_pager_find_least(vm_object_t object, vm_pindex_t pindex)
2079 VM_OBJECT_ASSERT_LOCKED(object);
2080 if (object->type != OBJT_SWAP)
2081 return (object->size);
2083 sb = SWAP_PCTRIE_LOOKUP_GE(&object->un_pager.swp.swp_blks,
2084 rounddown(pindex, SWAP_META_PAGES));
2086 return (object->size);
2087 if (sb->p < pindex) {
2088 for (i = pindex % SWAP_META_PAGES; i < SWAP_META_PAGES; i++) {
2089 if (sb->d[i] != SWAPBLK_NONE)
2092 sb = SWAP_PCTRIE_LOOKUP_GE(&object->un_pager.swp.swp_blks,
2093 roundup(pindex, SWAP_META_PAGES));
2095 return (object->size);
2097 for (i = 0; i < SWAP_META_PAGES; i++) {
2098 if (sb->d[i] != SWAPBLK_NONE)
2103 * We get here if a swblk is present in the trie but it
2104 * doesn't map any blocks.
2107 return (object->size);
2111 * System call swapon(name) enables swapping on device name,
2112 * which must be in the swdevsw. Return EBUSY
2113 * if already swapping on this device.
2115 #ifndef _SYS_SYSPROTO_H_
2116 struct swapon_args {
2126 sys_swapon(struct thread *td, struct swapon_args *uap)
2130 struct nameidata nd;
2133 error = priv_check(td, PRIV_SWAPON);
2137 sx_xlock(&swdev_syscall_lock);
2140 * Swap metadata may not fit in the KVM if we have physical
2143 if (swblk_zone == NULL) {
2148 NDINIT(&nd, LOOKUP, ISOPEN | FOLLOW | AUDITVNODE1, UIO_USERSPACE,
2154 NDFREE(&nd, NDF_ONLY_PNBUF);
2157 if (vn_isdisk(vp, &error)) {
2158 error = swapongeom(vp);
2159 } else if (vp->v_type == VREG &&
2160 (vp->v_mount->mnt_vfc->vfc_flags & VFCF_NETWORK) != 0 &&
2161 (error = VOP_GETATTR(vp, &attr, td->td_ucred)) == 0) {
2163 * Allow direct swapping to NFS regular files in the same
2164 * way that nfs_mountroot() sets up diskless swapping.
2166 error = swaponvp(td, vp, attr.va_size / DEV_BSIZE);
2172 sx_xunlock(&swdev_syscall_lock);
2177 * Check that the total amount of swap currently configured does not
2178 * exceed half the theoretical maximum. If it does, print a warning
2182 swapon_check_swzone(void)
2184 unsigned long maxpages, npages;
2186 npages = swap_total;
2187 /* absolute maximum we can handle assuming 100% efficiency */
2188 maxpages = uma_zone_get_max(swblk_zone) * SWAP_META_PAGES;
2190 /* recommend using no more than half that amount */
2191 if (npages > maxpages / 2) {
2192 printf("warning: total configured swap (%lu pages) "
2193 "exceeds maximum recommended amount (%lu pages).\n",
2194 npages, maxpages / 2);
2195 printf("warning: increase kern.maxswzone "
2196 "or reduce amount of swap.\n");
2201 swaponsomething(struct vnode *vp, void *id, u_long nblks,
2202 sw_strategy_t *strategy, sw_close_t *close, dev_t dev, int flags)
2204 struct swdevt *sp, *tsp;
2209 * nblks is in DEV_BSIZE'd chunks, convert to PAGE_SIZE'd chunks.
2210 * First chop nblks off to page-align it, then convert.
2212 * sw->sw_nblks is in page-sized chunks now too.
2214 nblks &= ~(ctodb(1) - 1);
2215 nblks = dbtoc(nblks);
2218 * If we go beyond this, we get overflows in the radix
2221 mblocks = 0x40000000 / BLIST_META_RADIX;
2222 if (nblks > mblocks) {
2224 "WARNING: reducing swap size to maximum of %luMB per unit\n",
2225 mblocks / 1024 / 1024 * PAGE_SIZE);
2229 sp = malloc(sizeof *sp, M_VMPGDATA, M_WAITOK | M_ZERO);
2234 sp->sw_nblks = nblks;
2236 sp->sw_strategy = strategy;
2237 sp->sw_close = close;
2238 sp->sw_flags = flags;
2240 sp->sw_blist = blist_create(nblks, M_WAITOK);
2242 * Do not free the first two block in order to avoid overwriting
2243 * any bsd label at the front of the partition
2245 blist_free(sp->sw_blist, 2, nblks - 2);
2248 mtx_lock(&sw_dev_mtx);
2249 TAILQ_FOREACH(tsp, &swtailq, sw_list) {
2250 if (tsp->sw_end >= dvbase) {
2252 * We put one uncovered page between the devices
2253 * in order to definitively prevent any cross-device
2256 dvbase = tsp->sw_end + 1;
2259 sp->sw_first = dvbase;
2260 sp->sw_end = dvbase + nblks;
2261 TAILQ_INSERT_TAIL(&swtailq, sp, sw_list);
2263 swap_pager_avail += nblks - 2;
2264 swap_total += nblks;
2265 swapon_check_swzone();
2267 mtx_unlock(&sw_dev_mtx);
2268 EVENTHANDLER_INVOKE(swapon, sp);
2272 * SYSCALL: swapoff(devname)
2274 * Disable swapping on the given device.
2276 * XXX: Badly designed system call: it should use a device index
2277 * rather than filename as specification. We keep sw_vp around
2278 * only to make this work.
2280 #ifndef _SYS_SYSPROTO_H_
2281 struct swapoff_args {
2291 sys_swapoff(struct thread *td, struct swapoff_args *uap)
2294 struct nameidata nd;
2298 error = priv_check(td, PRIV_SWAPOFF);
2302 sx_xlock(&swdev_syscall_lock);
2304 NDINIT(&nd, LOOKUP, FOLLOW | AUDITVNODE1, UIO_USERSPACE, uap->name,
2309 NDFREE(&nd, NDF_ONLY_PNBUF);
2312 mtx_lock(&sw_dev_mtx);
2313 TAILQ_FOREACH(sp, &swtailq, sw_list) {
2314 if (sp->sw_vp == vp)
2317 mtx_unlock(&sw_dev_mtx);
2322 error = swapoff_one(sp, td->td_ucred);
2324 sx_xunlock(&swdev_syscall_lock);
2329 swapoff_one(struct swdevt *sp, struct ucred *cred)
2336 sx_assert(&swdev_syscall_lock, SA_XLOCKED);
2338 (void) vn_lock(sp->sw_vp, LK_EXCLUSIVE | LK_RETRY);
2339 error = mac_system_check_swapoff(cred, sp->sw_vp);
2340 (void) VOP_UNLOCK(sp->sw_vp, 0);
2344 nblks = sp->sw_nblks;
2347 * We can turn off this swap device safely only if the
2348 * available virtual memory in the system will fit the amount
2349 * of data we will have to page back in, plus an epsilon so
2350 * the system doesn't become critically low on swap space.
2352 if (vm_free_count() + swap_pager_avail < nblks + nswap_lowat)
2356 * Prevent further allocations on this device.
2358 mtx_lock(&sw_dev_mtx);
2359 sp->sw_flags |= SW_CLOSING;
2360 swap_pager_avail -= blist_fill(sp->sw_blist, 0, nblks);
2361 swap_total -= nblks;
2362 mtx_unlock(&sw_dev_mtx);
2365 * Page in the contents of the device and close it.
2367 swap_pager_swapoff(sp);
2369 sp->sw_close(curthread, sp);
2370 mtx_lock(&sw_dev_mtx);
2372 TAILQ_REMOVE(&swtailq, sp, sw_list);
2374 if (nswapdev == 0) {
2375 swap_pager_full = 2;
2376 swap_pager_almost_full = 1;
2380 mtx_unlock(&sw_dev_mtx);
2381 blist_destroy(sp->sw_blist);
2382 free(sp, M_VMPGDATA);
2389 struct swdevt *sp, *spt;
2390 const char *devname;
2393 sx_xlock(&swdev_syscall_lock);
2395 mtx_lock(&sw_dev_mtx);
2396 TAILQ_FOREACH_SAFE(sp, &swtailq, sw_list, spt) {
2397 mtx_unlock(&sw_dev_mtx);
2398 if (vn_isdisk(sp->sw_vp, NULL))
2399 devname = devtoname(sp->sw_vp->v_rdev);
2402 error = swapoff_one(sp, thread0.td_ucred);
2404 printf("Cannot remove swap device %s (error=%d), "
2405 "skipping.\n", devname, error);
2406 } else if (bootverbose) {
2407 printf("Swap device %s removed.\n", devname);
2409 mtx_lock(&sw_dev_mtx);
2411 mtx_unlock(&sw_dev_mtx);
2413 sx_xunlock(&swdev_syscall_lock);
2417 swap_pager_status(int *total, int *used)
2423 mtx_lock(&sw_dev_mtx);
2424 TAILQ_FOREACH(sp, &swtailq, sw_list) {
2425 *total += sp->sw_nblks;
2426 *used += sp->sw_used;
2428 mtx_unlock(&sw_dev_mtx);
2432 swap_dev_info(int name, struct xswdev *xs, char *devname, size_t len)
2435 const char *tmp_devname;
2440 mtx_lock(&sw_dev_mtx);
2441 TAILQ_FOREACH(sp, &swtailq, sw_list) {
2446 xs->xsw_version = XSWDEV_VERSION;
2447 xs->xsw_dev = sp->sw_dev;
2448 xs->xsw_flags = sp->sw_flags;
2449 xs->xsw_nblks = sp->sw_nblks;
2450 xs->xsw_used = sp->sw_used;
2451 if (devname != NULL) {
2452 if (vn_isdisk(sp->sw_vp, NULL))
2453 tmp_devname = devtoname(sp->sw_vp->v_rdev);
2455 tmp_devname = "[file]";
2456 strncpy(devname, tmp_devname, len);
2461 mtx_unlock(&sw_dev_mtx);
2465 #if defined(COMPAT_FREEBSD11)
2466 #define XSWDEV_VERSION_11 1
2476 #if defined(__amd64__) && defined(COMPAT_FREEBSD32)
2479 u_int xsw_dev1, xsw_dev2;
2487 sysctl_vm_swap_info(SYSCTL_HANDLER_ARGS)
2490 #if defined(__amd64__) && defined(COMPAT_FREEBSD32)
2491 struct xswdev32 xs32;
2493 #if defined(COMPAT_FREEBSD11)
2494 struct xswdev11 xs11;
2498 if (arg2 != 1) /* name length */
2500 error = swap_dev_info(*(int *)arg1, &xs, NULL, 0);
2503 #if defined(__amd64__) && defined(COMPAT_FREEBSD32)
2504 if (req->oldlen == sizeof(xs32)) {
2505 xs32.xsw_version = XSWDEV_VERSION;
2506 xs32.xsw_dev1 = xs.xsw_dev;
2507 xs32.xsw_dev2 = xs.xsw_dev >> 32;
2508 xs32.xsw_flags = xs.xsw_flags;
2509 xs32.xsw_nblks = xs.xsw_nblks;
2510 xs32.xsw_used = xs.xsw_used;
2511 error = SYSCTL_OUT(req, &xs32, sizeof(xs32));
2515 #if defined(COMPAT_FREEBSD11)
2516 if (req->oldlen == sizeof(xs11)) {
2517 xs11.xsw_version = XSWDEV_VERSION_11;
2518 xs11.xsw_dev = xs.xsw_dev; /* truncation */
2519 xs11.xsw_flags = xs.xsw_flags;
2520 xs11.xsw_nblks = xs.xsw_nblks;
2521 xs11.xsw_used = xs.xsw_used;
2522 error = SYSCTL_OUT(req, &xs11, sizeof(xs11));
2526 error = SYSCTL_OUT(req, &xs, sizeof(xs));
2530 SYSCTL_INT(_vm, OID_AUTO, nswapdev, CTLFLAG_RD, &nswapdev, 0,
2531 "Number of swap devices");
2532 SYSCTL_NODE(_vm, OID_AUTO, swap_info, CTLFLAG_RD | CTLFLAG_MPSAFE,
2533 sysctl_vm_swap_info,
2534 "Swap statistics by device");
2537 * Count the approximate swap usage in pages for a vmspace. The
2538 * shadowed or not yet copied on write swap blocks are not accounted.
2539 * The map must be locked.
2542 vmspace_swap_count(struct vmspace *vmspace)
2552 map = &vmspace->vm_map;
2555 for (cur = map->header.next; cur != &map->header; cur = cur->next) {
2556 if ((cur->eflags & MAP_ENTRY_IS_SUB_MAP) != 0)
2558 object = cur->object.vm_object;
2559 if (object == NULL || object->type != OBJT_SWAP)
2561 VM_OBJECT_RLOCK(object);
2562 if (object->type != OBJT_SWAP)
2564 pi = OFF_TO_IDX(cur->offset);
2565 e = pi + OFF_TO_IDX(cur->end - cur->start);
2566 for (;; pi = sb->p + SWAP_META_PAGES) {
2567 sb = SWAP_PCTRIE_LOOKUP_GE(
2568 &object->un_pager.swp.swp_blks, pi);
2569 if (sb == NULL || sb->p >= e)
2571 for (i = 0; i < SWAP_META_PAGES; i++) {
2572 if (sb->p + i < e &&
2573 sb->d[i] != SWAPBLK_NONE)
2578 VM_OBJECT_RUNLOCK(object);
2586 * Swapping onto disk devices.
2590 static g_orphan_t swapgeom_orphan;
2592 static struct g_class g_swap_class = {
2594 .version = G_VERSION,
2595 .orphan = swapgeom_orphan,
2598 DECLARE_GEOM_CLASS(g_swap_class, g_class);
2602 swapgeom_close_ev(void *arg, int flags)
2604 struct g_consumer *cp;
2607 g_access(cp, -1, -1, 0);
2609 g_destroy_consumer(cp);
2613 * Add a reference to the g_consumer for an inflight transaction.
2616 swapgeom_acquire(struct g_consumer *cp)
2619 mtx_assert(&sw_dev_mtx, MA_OWNED);
2624 * Remove a reference from the g_consumer. Post a close event if all
2625 * references go away, since the function might be called from the
2629 swapgeom_release(struct g_consumer *cp, struct swdevt *sp)
2632 mtx_assert(&sw_dev_mtx, MA_OWNED);
2634 if (cp->index == 0) {
2635 if (g_post_event(swapgeom_close_ev, cp, M_NOWAIT, NULL) == 0)
2641 swapgeom_done(struct bio *bp2)
2645 struct g_consumer *cp;
2647 bp = bp2->bio_caller2;
2649 bp->b_ioflags = bp2->bio_flags;
2651 bp->b_ioflags |= BIO_ERROR;
2652 bp->b_resid = bp->b_bcount - bp2->bio_completed;
2653 bp->b_error = bp2->bio_error;
2654 bp->b_caller1 = NULL;
2656 sp = bp2->bio_caller1;
2657 mtx_lock(&sw_dev_mtx);
2658 swapgeom_release(cp, sp);
2659 mtx_unlock(&sw_dev_mtx);
2664 swapgeom_strategy(struct buf *bp, struct swdevt *sp)
2667 struct g_consumer *cp;
2669 mtx_lock(&sw_dev_mtx);
2672 mtx_unlock(&sw_dev_mtx);
2673 bp->b_error = ENXIO;
2674 bp->b_ioflags |= BIO_ERROR;
2678 swapgeom_acquire(cp);
2679 mtx_unlock(&sw_dev_mtx);
2680 if (bp->b_iocmd == BIO_WRITE)
2683 bio = g_alloc_bio();
2685 mtx_lock(&sw_dev_mtx);
2686 swapgeom_release(cp, sp);
2687 mtx_unlock(&sw_dev_mtx);
2688 bp->b_error = ENOMEM;
2689 bp->b_ioflags |= BIO_ERROR;
2690 printf("swap_pager: cannot allocate bio\n");
2695 bp->b_caller1 = bio;
2696 bio->bio_caller1 = sp;
2697 bio->bio_caller2 = bp;
2698 bio->bio_cmd = bp->b_iocmd;
2699 bio->bio_offset = (bp->b_blkno - sp->sw_first) * PAGE_SIZE;
2700 bio->bio_length = bp->b_bcount;
2701 bio->bio_done = swapgeom_done;
2702 if (!buf_mapped(bp)) {
2703 bio->bio_ma = bp->b_pages;
2704 bio->bio_data = unmapped_buf;
2705 bio->bio_ma_offset = (vm_offset_t)bp->b_offset & PAGE_MASK;
2706 bio->bio_ma_n = bp->b_npages;
2707 bio->bio_flags |= BIO_UNMAPPED;
2709 bio->bio_data = bp->b_data;
2712 g_io_request(bio, cp);
2717 swapgeom_orphan(struct g_consumer *cp)
2722 mtx_lock(&sw_dev_mtx);
2723 TAILQ_FOREACH(sp, &swtailq, sw_list) {
2724 if (sp->sw_id == cp) {
2725 sp->sw_flags |= SW_CLOSING;
2730 * Drop reference we were created with. Do directly since we're in a
2731 * special context where we don't have to queue the call to
2732 * swapgeom_close_ev().
2735 destroy = ((sp != NULL) && (cp->index == 0));
2738 mtx_unlock(&sw_dev_mtx);
2740 swapgeom_close_ev(cp, 0);
2744 swapgeom_close(struct thread *td, struct swdevt *sw)
2746 struct g_consumer *cp;
2748 mtx_lock(&sw_dev_mtx);
2751 mtx_unlock(&sw_dev_mtx);
2754 * swapgeom_close() may be called from the biodone context,
2755 * where we cannot perform topology changes. Delegate the
2756 * work to the events thread.
2759 g_waitfor_event(swapgeom_close_ev, cp, M_WAITOK, NULL);
2763 swapongeom_locked(struct cdev *dev, struct vnode *vp)
2765 struct g_provider *pp;
2766 struct g_consumer *cp;
2767 static struct g_geom *gp;
2772 pp = g_dev_getprovider(dev);
2775 mtx_lock(&sw_dev_mtx);
2776 TAILQ_FOREACH(sp, &swtailq, sw_list) {
2778 if (cp != NULL && cp->provider == pp) {
2779 mtx_unlock(&sw_dev_mtx);
2783 mtx_unlock(&sw_dev_mtx);
2785 gp = g_new_geomf(&g_swap_class, "swap");
2786 cp = g_new_consumer(gp);
2787 cp->index = 1; /* Number of active I/Os, plus one for being active. */
2788 cp->flags |= G_CF_DIRECT_SEND | G_CF_DIRECT_RECEIVE;
2791 * XXX: Every time you think you can improve the margin for
2792 * footshooting, somebody depends on the ability to do so:
2793 * savecore(8) wants to write to our swapdev so we cannot
2794 * set an exclusive count :-(
2796 error = g_access(cp, 1, 1, 0);
2799 g_destroy_consumer(cp);
2802 nblks = pp->mediasize / DEV_BSIZE;
2803 swaponsomething(vp, cp, nblks, swapgeom_strategy,
2804 swapgeom_close, dev2udev(dev),
2805 (pp->flags & G_PF_ACCEPT_UNMAPPED) != 0 ? SW_UNMAPPED : 0);
2810 swapongeom(struct vnode *vp)
2814 vn_lock(vp, LK_EXCLUSIVE | LK_RETRY);
2815 if (vp->v_type != VCHR || (vp->v_iflag & VI_DOOMED) != 0) {
2819 error = swapongeom_locked(vp->v_rdev, vp);
2820 g_topology_unlock();
2829 * This is used mainly for network filesystem (read: probably only tested
2830 * with NFS) swapfiles.
2835 swapdev_strategy(struct buf *bp, struct swdevt *sp)
2839 bp->b_blkno = ctodb(bp->b_blkno - sp->sw_first);
2843 if (bp->b_iocmd == BIO_WRITE) {
2845 bufobj_wdrop(bp->b_bufobj);
2846 bufobj_wref(&vp2->v_bufobj);
2848 if (bp->b_bufobj != &vp2->v_bufobj)
2849 bp->b_bufobj = &vp2->v_bufobj;
2851 bp->b_iooffset = dbtob(bp->b_blkno);
2857 swapdev_close(struct thread *td, struct swdevt *sp)
2860 VOP_CLOSE(sp->sw_vp, FREAD | FWRITE, td->td_ucred, td);
2866 swaponvp(struct thread *td, struct vnode *vp, u_long nblks)
2873 mtx_lock(&sw_dev_mtx);
2874 TAILQ_FOREACH(sp, &swtailq, sw_list) {
2875 if (sp->sw_id == vp) {
2876 mtx_unlock(&sw_dev_mtx);
2880 mtx_unlock(&sw_dev_mtx);
2882 (void) vn_lock(vp, LK_EXCLUSIVE | LK_RETRY);
2884 error = mac_system_check_swapon(td->td_ucred, vp);
2887 error = VOP_OPEN(vp, FREAD | FWRITE, td->td_ucred, td, NULL);
2888 (void) VOP_UNLOCK(vp, 0);
2892 swaponsomething(vp, vp, nblks, swapdev_strategy, swapdev_close,
2898 sysctl_swap_async_max(SYSCTL_HANDLER_ARGS)
2902 new = nsw_wcount_async_max;
2903 error = sysctl_handle_int(oidp, &new, 0, req);
2904 if (error != 0 || req->newptr == NULL)
2907 if (new > nswbuf / 2 || new < 1)
2910 mtx_lock(&swbuf_mtx);
2911 while (nsw_wcount_async_max != new) {
2913 * Adjust difference. If the current async count is too low,
2914 * we will need to sqeeze our update slowly in. Sleep with a
2915 * higher priority than getpbuf() to finish faster.
2917 n = new - nsw_wcount_async_max;
2918 if (nsw_wcount_async + n >= 0) {
2919 nsw_wcount_async += n;
2920 nsw_wcount_async_max += n;
2921 wakeup(&nsw_wcount_async);
2923 nsw_wcount_async_max -= nsw_wcount_async;
2924 nsw_wcount_async = 0;
2925 msleep(&nsw_wcount_async, &swbuf_mtx, PSWP,
2929 mtx_unlock(&swbuf_mtx);