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("$FreeBSD$");
77 #include <sys/param.h>
78 #include <sys/systm.h>
80 #include <sys/kernel.h>
86 #include <sys/disklabel.h>
87 #include <sys/fcntl.h>
88 #include <sys/mount.h>
89 #include <sys/namei.h>
90 #include <sys/vnode.h>
91 #include <sys/malloc.h>
92 #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/blist.h>
101 #include <sys/lock.h>
103 #include <sys/vmmeter.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 int nsw_rcount; /* free read buffers */
329 static int nsw_wcount_sync; /* limit write buffers / synchronous */
330 static int nsw_wcount_async; /* limit write buffers / asynchronous */
331 static int nsw_wcount_async_max;/* assigned maximum */
332 static int nsw_cluster_max; /* maximum VOP I/O allowed */
334 static int sysctl_swap_async_max(SYSCTL_HANDLER_ARGS);
335 SYSCTL_PROC(_vm, OID_AUTO, swap_async_max, CTLTYPE_INT | CTLFLAG_RW |
336 CTLFLAG_MPSAFE, NULL, 0, sysctl_swap_async_max, "I",
337 "Maximum running async swap ops");
338 static int sysctl_swap_fragmentation(SYSCTL_HANDLER_ARGS);
339 SYSCTL_PROC(_vm, OID_AUTO, swap_fragmentation, CTLTYPE_STRING | CTLFLAG_RD |
340 CTLFLAG_MPSAFE, NULL, 0, sysctl_swap_fragmentation, "A",
341 "Swap Fragmentation Info");
343 static struct sx sw_alloc_sx;
346 * "named" and "unnamed" anon region objects. Try to reduce the overhead
347 * of searching a named list by hashing it just a little.
352 #define NOBJLIST(handle) \
353 (&swap_pager_object_list[((int)(intptr_t)handle >> 4) & (NOBJLISTS-1)])
355 static struct pagerlst swap_pager_object_list[NOBJLISTS];
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);
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);
544 nsw_rcount = (nswbuf + 1) / 2;
545 nsw_wcount_sync = (nswbuf + 3) / 4;
546 nsw_wcount_async = 4;
547 nsw_wcount_async_max = nsw_wcount_async;
548 mtx_unlock(&pbuf_mtx);
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 the requested number of pages. The starting
713 * swap block number (a page index) is returned or SWAPBLK_NONE
714 * if the allocation failed.
716 * Also has the side effect of advising that somebody made a mistake
717 * when they configured swap and didn't configure enough.
719 * This routine may not sleep.
721 * We allocate in round-robin fashion from the configured devices.
724 swp_pager_getswapspace(int npages)
731 mtx_lock(&sw_dev_mtx);
733 for (i = 0; i < nswapdev; i++) {
735 sp = TAILQ_FIRST(&swtailq);
736 if (!(sp->sw_flags & SW_CLOSING)) {
737 blk = blist_alloc(sp->sw_blist, npages);
738 if (blk != SWAPBLK_NONE) {
740 sp->sw_used += npages;
741 swap_pager_avail -= npages;
743 swdevhd = TAILQ_NEXT(sp, sw_list);
747 sp = TAILQ_NEXT(sp, sw_list);
749 if (swap_pager_full != 2) {
750 printf("swap_pager_getswapspace(%d): failed\n", npages);
752 swap_pager_almost_full = 1;
756 mtx_unlock(&sw_dev_mtx);
761 swp_pager_isondev(daddr_t blk, struct swdevt *sp)
764 return (blk >= sp->sw_first && blk < sp->sw_end);
768 swp_pager_strategy(struct buf *bp)
772 mtx_lock(&sw_dev_mtx);
773 TAILQ_FOREACH(sp, &swtailq, sw_list) {
774 if (bp->b_blkno >= sp->sw_first && bp->b_blkno < sp->sw_end) {
775 mtx_unlock(&sw_dev_mtx);
776 if ((sp->sw_flags & SW_UNMAPPED) != 0 &&
777 unmapped_buf_allowed) {
778 bp->b_data = unmapped_buf;
781 pmap_qenter((vm_offset_t)bp->b_data,
782 &bp->b_pages[0], bp->b_bcount / PAGE_SIZE);
784 sp->sw_strategy(bp, sp);
788 panic("Swapdev not found");
793 * SWP_PAGER_FREESWAPSPACE() - free raw swap space
795 * This routine returns the specified swap blocks back to the bitmap.
797 * This routine may not sleep.
800 swp_pager_freeswapspace(daddr_t blk, daddr_t npages)
806 mtx_lock(&sw_dev_mtx);
807 TAILQ_FOREACH(sp, &swtailq, sw_list) {
808 if (blk >= sp->sw_first && blk < sp->sw_end) {
809 sp->sw_used -= npages;
811 * If we are attempting to stop swapping on
812 * this device, we don't want to mark any
813 * blocks free lest they be reused.
815 if ((sp->sw_flags & SW_CLOSING) == 0) {
816 blist_free(sp->sw_blist, blk - sp->sw_first,
818 swap_pager_avail += npages;
821 mtx_unlock(&sw_dev_mtx);
825 panic("Swapdev not found");
829 * SYSCTL_SWAP_FRAGMENTATION() - produce raw swap space stats
832 sysctl_swap_fragmentation(SYSCTL_HANDLER_ARGS)
839 error = sysctl_wire_old_buffer(req, 0);
842 sbuf_new_for_sysctl(&sbuf, NULL, 128, req);
843 mtx_lock(&sw_dev_mtx);
844 TAILQ_FOREACH(sp, &swtailq, sw_list) {
845 if (vn_isdisk(sp->sw_vp, NULL))
846 devname = devtoname(sp->sw_vp->v_rdev);
849 sbuf_printf(&sbuf, "\nFree space on device %s:\n", devname);
850 blist_stats(sp->sw_blist, &sbuf);
852 mtx_unlock(&sw_dev_mtx);
853 error = sbuf_finish(&sbuf);
859 * SWAP_PAGER_FREESPACE() - frees swap blocks associated with a page
860 * range within an object.
862 * This is a globally accessible routine.
864 * This routine removes swapblk assignments from swap metadata.
866 * The external callers of this routine typically have already destroyed
867 * or renamed vm_page_t's associated with this range in the object so
870 * The object must be locked.
873 swap_pager_freespace(vm_object_t object, vm_pindex_t start, vm_size_t size)
876 swp_pager_meta_free(object, start, size);
880 * SWAP_PAGER_RESERVE() - reserve swap blocks in object
882 * Assigns swap blocks to the specified range within the object. The
883 * swap blocks are not zeroed. Any previous swap assignment is destroyed.
885 * Returns 0 on success, -1 on failure.
888 swap_pager_reserve(vm_object_t object, vm_pindex_t start, vm_size_t size)
891 daddr_t blk = SWAPBLK_NONE;
892 vm_pindex_t beg = start; /* save start index */
893 daddr_t addr, n_free, s_free;
895 swp_pager_init_freerange(&s_free, &n_free);
896 VM_OBJECT_WLOCK(object);
900 while ((blk = swp_pager_getswapspace(n)) == SWAPBLK_NONE) {
903 swp_pager_meta_free(object, beg, start - beg);
904 VM_OBJECT_WUNLOCK(object);
909 addr = swp_pager_meta_build(object, start, blk);
910 if (addr != SWAPBLK_NONE)
911 swp_pager_update_freerange(&s_free, &n_free, addr);
917 swp_pager_freeswapspace(s_free, n_free);
918 swp_pager_meta_free(object, start, n);
919 VM_OBJECT_WUNLOCK(object);
924 * SWAP_PAGER_COPY() - copy blocks from source pager to destination pager
925 * and destroy the source.
927 * Copy any valid swapblks from the source to the destination. In
928 * cases where both the source and destination have a valid swapblk,
929 * we keep the destination's.
931 * This routine is allowed to sleep. It may sleep allocating metadata
932 * indirectly through swp_pager_meta_build() or if paging is still in
933 * progress on the source.
935 * The source object contains no vm_page_t's (which is just as well)
937 * The source object is of type OBJT_SWAP.
939 * The source and destination objects must be locked.
940 * Both object locks may temporarily be released.
943 swap_pager_copy(vm_object_t srcobject, vm_object_t dstobject,
944 vm_pindex_t offset, int destroysource)
947 daddr_t dstaddr, n_free, s_free, srcaddr;
949 VM_OBJECT_ASSERT_WLOCKED(srcobject);
950 VM_OBJECT_ASSERT_WLOCKED(dstobject);
953 * If destroysource is set, we remove the source object from the
954 * swap_pager internal queue now.
956 if (destroysource && srcobject->handle != NULL) {
957 vm_object_pip_add(srcobject, 1);
958 VM_OBJECT_WUNLOCK(srcobject);
959 vm_object_pip_add(dstobject, 1);
960 VM_OBJECT_WUNLOCK(dstobject);
961 sx_xlock(&sw_alloc_sx);
962 TAILQ_REMOVE(NOBJLIST(srcobject->handle), srcobject,
964 sx_xunlock(&sw_alloc_sx);
965 VM_OBJECT_WLOCK(dstobject);
966 vm_object_pip_wakeup(dstobject);
967 VM_OBJECT_WLOCK(srcobject);
968 vm_object_pip_wakeup(srcobject);
972 * Transfer source to destination.
974 swp_pager_init_freerange(&s_free, &n_free);
975 for (i = 0; i < dstobject->size; ++i) {
976 srcaddr = swp_pager_meta_ctl(srcobject, i + offset, SWM_POP);
977 if (srcaddr == SWAPBLK_NONE)
979 dstaddr = swp_pager_meta_ctl(dstobject, i, 0);
980 if (dstaddr != SWAPBLK_NONE) {
982 * Destination has valid swapblk or it is represented
983 * by a resident page. We destroy the source block.
985 swp_pager_update_freerange(&s_free, &n_free, srcaddr);
990 * Destination has no swapblk and is not resident,
993 * swp_pager_meta_build() can sleep.
995 vm_object_pip_add(srcobject, 1);
996 VM_OBJECT_WUNLOCK(srcobject);
997 vm_object_pip_add(dstobject, 1);
998 dstaddr = swp_pager_meta_build(dstobject, i, srcaddr);
999 KASSERT(dstaddr == SWAPBLK_NONE,
1000 ("Unexpected destination swapblk"));
1001 vm_object_pip_wakeup(dstobject);
1002 VM_OBJECT_WLOCK(srcobject);
1003 vm_object_pip_wakeup(srcobject);
1005 swp_pager_freeswapspace(s_free, n_free);
1008 * Free left over swap blocks in source.
1010 * We have to revert the type to OBJT_DEFAULT so we do not accidentally
1011 * double-remove the object from the swap queues.
1013 if (destroysource) {
1014 swp_pager_meta_free_all(srcobject);
1016 * Reverting the type is not necessary, the caller is going
1017 * to destroy srcobject directly, but I'm doing it here
1018 * for consistency since we've removed the object from its
1021 srcobject->type = OBJT_DEFAULT;
1026 * SWAP_PAGER_HASPAGE() - determine if we have good backing store for
1027 * the requested page.
1029 * We determine whether good backing store exists for the requested
1030 * page and return TRUE if it does, FALSE if it doesn't.
1032 * If TRUE, we also try to determine how much valid, contiguous backing
1033 * store exists before and after the requested page.
1036 swap_pager_haspage(vm_object_t object, vm_pindex_t pindex, int *before,
1042 VM_OBJECT_ASSERT_LOCKED(object);
1045 * do we have good backing store at the requested index ?
1047 blk0 = swp_pager_meta_ctl(object, pindex, 0);
1048 if (blk0 == SWAPBLK_NONE) {
1057 * find backwards-looking contiguous good backing store
1059 if (before != NULL) {
1060 for (i = 1; i < SWB_NPAGES; i++) {
1063 blk = swp_pager_meta_ctl(object, pindex - i, 0);
1064 if (blk != blk0 - i)
1071 * find forward-looking contiguous good backing store
1073 if (after != NULL) {
1074 for (i = 1; i < SWB_NPAGES; i++) {
1075 blk = swp_pager_meta_ctl(object, pindex + i, 0);
1076 if (blk != blk0 + i)
1085 * SWAP_PAGER_PAGE_UNSWAPPED() - remove swap backing store related to page
1087 * This removes any associated swap backing store, whether valid or
1088 * not, from the page.
1090 * This routine is typically called when a page is made dirty, at
1091 * which point any associated swap can be freed. MADV_FREE also
1092 * calls us in a special-case situation
1094 * NOTE!!! If the page is clean and the swap was valid, the caller
1095 * should make the page dirty before calling this routine. This routine
1096 * does NOT change the m->dirty status of the page. Also: MADV_FREE
1099 * This routine may not sleep.
1101 * The object containing the page must be locked.
1104 swap_pager_unswapped(vm_page_t m)
1108 srcaddr = swp_pager_meta_ctl(m->object, m->pindex, SWM_POP);
1109 if (srcaddr != SWAPBLK_NONE)
1110 swp_pager_freeswapspace(srcaddr, 1);
1114 * swap_pager_getpages() - bring pages in from swap
1116 * Attempt to page in the pages in array "ma" of length "count". The
1117 * caller may optionally specify that additional pages preceding and
1118 * succeeding the specified range be paged in. The number of such pages
1119 * is returned in the "rbehind" and "rahead" parameters, and they will
1120 * be in the inactive queue upon return.
1122 * The pages in "ma" must be busied and will remain busied upon return.
1125 swap_pager_getpages(vm_object_t object, vm_page_t *ma, int count, int *rbehind,
1129 vm_page_t bm, mpred, msucc, p;
1132 int i, maxahead, maxbehind, reqcount;
1137 * Determine the final number of read-behind pages and
1138 * allocate them BEFORE releasing the object lock. Otherwise,
1139 * there can be a problematic race with vm_object_split().
1140 * Specifically, vm_object_split() might first transfer pages
1141 * that precede ma[0] in the current object to a new object,
1142 * and then this function incorrectly recreates those pages as
1143 * read-behind pages in the current object.
1145 if (!swap_pager_haspage(object, ma[0]->pindex, &maxbehind, &maxahead))
1146 return (VM_PAGER_FAIL);
1149 * Clip the readahead and readbehind ranges to exclude resident pages.
1151 if (rahead != NULL) {
1152 KASSERT(reqcount - 1 <= maxahead,
1153 ("page count %d extends beyond swap block", reqcount));
1154 *rahead = imin(*rahead, maxahead - (reqcount - 1));
1155 pindex = ma[reqcount - 1]->pindex;
1156 msucc = TAILQ_NEXT(ma[reqcount - 1], listq);
1157 if (msucc != NULL && msucc->pindex - pindex - 1 < *rahead)
1158 *rahead = msucc->pindex - pindex - 1;
1160 if (rbehind != NULL) {
1161 *rbehind = imin(*rbehind, maxbehind);
1162 pindex = ma[0]->pindex;
1163 mpred = TAILQ_PREV(ma[0], pglist, listq);
1164 if (mpred != NULL && pindex - mpred->pindex - 1 < *rbehind)
1165 *rbehind = pindex - mpred->pindex - 1;
1169 for (i = 0; i < count; i++)
1170 ma[i]->oflags |= VPO_SWAPINPROG;
1173 * Allocate readahead and readbehind pages.
1175 if (rbehind != NULL) {
1176 for (i = 1; i <= *rbehind; i++) {
1177 p = vm_page_alloc(object, ma[0]->pindex - i,
1181 p->oflags |= VPO_SWAPINPROG;
1186 if (rahead != NULL) {
1187 for (i = 0; i < *rahead; i++) {
1188 p = vm_page_alloc(object,
1189 ma[reqcount - 1]->pindex + i + 1, VM_ALLOC_NORMAL);
1192 p->oflags |= VPO_SWAPINPROG;
1196 if (rbehind != NULL)
1201 vm_object_pip_add(object, count);
1203 pindex = bm->pindex;
1204 blk = swp_pager_meta_ctl(object, pindex, 0);
1205 KASSERT(blk != SWAPBLK_NONE,
1206 ("no swap blocking containing %p(%jx)", object, (uintmax_t)pindex));
1208 VM_OBJECT_WUNLOCK(object);
1209 bp = getpbuf(&nsw_rcount);
1210 /* Pages cannot leave the object while busy. */
1211 for (i = 0, p = bm; i < count; i++, p = TAILQ_NEXT(p, listq)) {
1212 MPASS(p->pindex == bm->pindex + i);
1216 bp->b_flags |= B_PAGING;
1217 bp->b_iocmd = BIO_READ;
1218 bp->b_iodone = swp_pager_async_iodone;
1219 bp->b_rcred = crhold(thread0.td_ucred);
1220 bp->b_wcred = crhold(thread0.td_ucred);
1222 bp->b_bcount = PAGE_SIZE * count;
1223 bp->b_bufsize = PAGE_SIZE * count;
1224 bp->b_npages = count;
1225 bp->b_pgbefore = rbehind != NULL ? *rbehind : 0;
1226 bp->b_pgafter = rahead != NULL ? *rahead : 0;
1228 VM_CNT_INC(v_swapin);
1229 VM_CNT_ADD(v_swappgsin, count);
1232 * perform the I/O. NOTE!!! bp cannot be considered valid after
1233 * this point because we automatically release it on completion.
1234 * Instead, we look at the one page we are interested in which we
1235 * still hold a lock on even through the I/O completion.
1237 * The other pages in our ma[] array are also released on completion,
1238 * so we cannot assume they are valid anymore either.
1240 * NOTE: b_blkno is destroyed by the call to swapdev_strategy
1243 swp_pager_strategy(bp);
1246 * Wait for the pages we want to complete. VPO_SWAPINPROG is always
1247 * cleared on completion. If an I/O error occurs, SWAPBLK_NONE
1248 * is set in the metadata for each page in the request.
1250 VM_OBJECT_WLOCK(object);
1251 while ((ma[0]->oflags & VPO_SWAPINPROG) != 0) {
1252 ma[0]->oflags |= VPO_SWAPSLEEP;
1253 VM_CNT_INC(v_intrans);
1254 if (VM_OBJECT_SLEEP(object, &object->paging_in_progress, PSWP,
1255 "swread", hz * 20)) {
1257 "swap_pager: indefinite wait buffer: bufobj: %p, blkno: %jd, size: %ld\n",
1258 bp->b_bufobj, (intmax_t)bp->b_blkno, bp->b_bcount);
1263 * If we had an unrecoverable read error pages will not be valid.
1265 for (i = 0; i < reqcount; i++)
1266 if (ma[i]->valid != VM_PAGE_BITS_ALL)
1267 return (VM_PAGER_ERROR);
1269 return (VM_PAGER_OK);
1272 * A final note: in a low swap situation, we cannot deallocate swap
1273 * and mark a page dirty here because the caller is likely to mark
1274 * the page clean when we return, causing the page to possibly revert
1275 * to all-zero's later.
1280 * swap_pager_getpages_async():
1282 * Right now this is emulation of asynchronous operation on top of
1283 * swap_pager_getpages().
1286 swap_pager_getpages_async(vm_object_t object, vm_page_t *ma, int count,
1287 int *rbehind, int *rahead, pgo_getpages_iodone_t iodone, void *arg)
1291 r = swap_pager_getpages(object, ma, count, rbehind, rahead);
1292 VM_OBJECT_WUNLOCK(object);
1297 case VM_PAGER_ERROR:
1304 panic("unhandled swap_pager_getpages() error %d", r);
1306 (iodone)(arg, ma, count, error);
1307 VM_OBJECT_WLOCK(object);
1313 * swap_pager_putpages:
1315 * Assign swap (if necessary) and initiate I/O on the specified pages.
1317 * We support both OBJT_DEFAULT and OBJT_SWAP objects. DEFAULT objects
1318 * are automatically converted to SWAP objects.
1320 * In a low memory situation we may block in VOP_STRATEGY(), but the new
1321 * vm_page reservation system coupled with properly written VFS devices
1322 * should ensure that no low-memory deadlock occurs. This is an area
1325 * The parent has N vm_object_pip_add() references prior to
1326 * calling us and will remove references for rtvals[] that are
1327 * not set to VM_PAGER_PEND. We need to remove the rest on I/O
1330 * The parent has soft-busy'd the pages it passes us and will unbusy
1331 * those whos rtvals[] entry is not set to VM_PAGER_PEND on return.
1332 * We need to unbusy the rest on I/O completion.
1335 swap_pager_putpages(vm_object_t object, vm_page_t *ma, int count,
1336 int flags, int *rtvals)
1340 daddr_t addr, n_free, s_free;
1342 swp_pager_init_freerange(&s_free, &n_free);
1343 if (count && ma[0]->object != object) {
1344 panic("swap_pager_putpages: object mismatch %p/%p",
1353 * Turn object into OBJT_SWAP
1354 * check for bogus sysops
1355 * force sync if not pageout process
1357 if (object->type != OBJT_SWAP) {
1358 addr = swp_pager_meta_build(object, 0, SWAPBLK_NONE);
1359 KASSERT(addr == SWAPBLK_NONE,
1360 ("unexpected object swap block"));
1362 VM_OBJECT_WUNLOCK(object);
1365 if (curproc != pageproc)
1368 sync = (flags & VM_PAGER_PUT_SYNC) != 0;
1373 * Assign swap blocks and issue I/O. We reallocate swap on the fly.
1374 * The page is left dirty until the pageout operation completes
1377 for (i = 0; i < count; i += n) {
1383 * Maximum I/O size is limited by a number of factors.
1385 n = min(BLIST_MAX_ALLOC, count - i);
1386 n = min(n, nsw_cluster_max);
1389 * Get biggest block of swap we can. If we fail, fall
1390 * back and try to allocate a smaller block. Don't go
1391 * overboard trying to allocate space if it would overly
1395 (blk = swp_pager_getswapspace(n)) == SWAPBLK_NONE &&
1400 if (blk == SWAPBLK_NONE) {
1401 for (j = 0; j < n; ++j)
1402 rtvals[i+j] = VM_PAGER_FAIL;
1407 * All I/O parameters have been satisfied, build the I/O
1408 * request and assign the swap space.
1411 bp = getpbuf(&nsw_wcount_sync);
1413 bp = getpbuf(&nsw_wcount_async);
1414 bp->b_flags = B_ASYNC;
1416 bp->b_flags |= B_PAGING;
1417 bp->b_iocmd = BIO_WRITE;
1419 bp->b_rcred = crhold(thread0.td_ucred);
1420 bp->b_wcred = crhold(thread0.td_ucred);
1421 bp->b_bcount = PAGE_SIZE * n;
1422 bp->b_bufsize = PAGE_SIZE * n;
1425 VM_OBJECT_WLOCK(object);
1426 for (j = 0; j < n; ++j) {
1427 vm_page_t mreq = ma[i+j];
1429 addr = swp_pager_meta_build(mreq->object, mreq->pindex,
1431 if (addr != SWAPBLK_NONE)
1432 swp_pager_update_freerange(&s_free, &n_free,
1434 MPASS(mreq->dirty == VM_PAGE_BITS_ALL);
1435 mreq->oflags |= VPO_SWAPINPROG;
1436 bp->b_pages[j] = mreq;
1438 VM_OBJECT_WUNLOCK(object);
1441 * Must set dirty range for NFS to work.
1444 bp->b_dirtyend = bp->b_bcount;
1446 VM_CNT_INC(v_swapout);
1447 VM_CNT_ADD(v_swappgsout, bp->b_npages);
1450 * We unconditionally set rtvals[] to VM_PAGER_PEND so that we
1451 * can call the async completion routine at the end of a
1452 * synchronous I/O operation. Otherwise, our caller would
1453 * perform duplicate unbusy and wakeup operations on the page
1454 * and object, respectively.
1456 for (j = 0; j < n; j++)
1457 rtvals[i + j] = VM_PAGER_PEND;
1462 * NOTE: b_blkno is destroyed by the call to swapdev_strategy
1464 if (sync == FALSE) {
1465 bp->b_iodone = swp_pager_async_iodone;
1467 swp_pager_strategy(bp);
1474 * NOTE: b_blkno is destroyed by the call to swapdev_strategy
1476 bp->b_iodone = bdone;
1477 swp_pager_strategy(bp);
1480 * Wait for the sync I/O to complete.
1482 bwait(bp, PVM, "swwrt");
1485 * Now that we are through with the bp, we can call the
1486 * normal async completion, which frees everything up.
1488 swp_pager_async_iodone(bp);
1490 VM_OBJECT_WLOCK(object);
1491 swp_pager_freeswapspace(s_free, n_free);
1495 * swp_pager_async_iodone:
1497 * Completion routine for asynchronous reads and writes from/to swap.
1498 * Also called manually by synchronous code to finish up a bp.
1500 * This routine may not sleep.
1503 swp_pager_async_iodone(struct buf *bp)
1506 vm_object_t object = NULL;
1511 if (bp->b_ioflags & BIO_ERROR) {
1513 "swap_pager: I/O error - %s failed; blkno %ld,"
1514 "size %ld, error %d\n",
1515 ((bp->b_iocmd == BIO_READ) ? "pagein" : "pageout"),
1523 * remove the mapping for kernel virtual
1526 pmap_qremove((vm_offset_t)bp->b_data, bp->b_npages);
1528 bp->b_data = bp->b_kvabase;
1531 object = bp->b_pages[0]->object;
1532 VM_OBJECT_WLOCK(object);
1536 * cleanup pages. If an error occurs writing to swap, we are in
1537 * very serious trouble. If it happens to be a disk error, though,
1538 * we may be able to recover by reassigning the swap later on. So
1539 * in this case we remove the m->swapblk assignment for the page
1540 * but do not free it in the rlist. The errornous block(s) are thus
1541 * never reallocated as swap. Redirty the page and continue.
1543 for (i = 0; i < bp->b_npages; ++i) {
1544 vm_page_t m = bp->b_pages[i];
1546 m->oflags &= ~VPO_SWAPINPROG;
1547 if (m->oflags & VPO_SWAPSLEEP) {
1548 m->oflags &= ~VPO_SWAPSLEEP;
1549 wakeup(&object->paging_in_progress);
1552 if (bp->b_ioflags & BIO_ERROR) {
1554 * If an error occurs I'd love to throw the swapblk
1555 * away without freeing it back to swapspace, so it
1556 * can never be used again. But I can't from an
1559 if (bp->b_iocmd == BIO_READ) {
1561 * NOTE: for reads, m->dirty will probably
1562 * be overridden by the original caller of
1563 * getpages so don't play cute tricks here.
1568 * If a write error occurs, reactivate page
1569 * so it doesn't clog the inactive list,
1570 * then finish the I/O.
1572 MPASS(m->dirty == VM_PAGE_BITS_ALL);
1574 vm_page_activate(m);
1578 } else if (bp->b_iocmd == BIO_READ) {
1580 * NOTE: for reads, m->dirty will probably be
1581 * overridden by the original caller of getpages so
1582 * we cannot set them in order to free the underlying
1583 * swap in a low-swap situation. I don't think we'd
1584 * want to do that anyway, but it was an optimization
1585 * that existed in the old swapper for a time before
1586 * it got ripped out due to precisely this problem.
1588 KASSERT(!pmap_page_is_mapped(m),
1589 ("swp_pager_async_iodone: page %p is mapped", m));
1590 KASSERT(m->dirty == 0,
1591 ("swp_pager_async_iodone: page %p is dirty", m));
1593 m->valid = VM_PAGE_BITS_ALL;
1594 if (i < bp->b_pgbefore ||
1595 i >= bp->b_npages - bp->b_pgafter)
1596 vm_page_readahead_finish(m);
1599 * For write success, clear the dirty
1600 * status, then finish the I/O ( which decrements the
1601 * busy count and possibly wakes waiter's up ).
1602 * A page is only written to swap after a period of
1603 * inactivity. Therefore, we do not expect it to be
1606 KASSERT(!pmap_page_is_write_mapped(m),
1607 ("swp_pager_async_iodone: page %p is not write"
1611 vm_page_deactivate_noreuse(m);
1618 * adjust pip. NOTE: the original parent may still have its own
1619 * pip refs on the object.
1621 if (object != NULL) {
1622 vm_object_pip_wakeupn(object, bp->b_npages);
1623 VM_OBJECT_WUNLOCK(object);
1627 * swapdev_strategy() manually sets b_vp and b_bufobj before calling
1628 * bstrategy(). Set them back to NULL now we're done with it, or we'll
1629 * trigger a KASSERT in relpbuf().
1633 bp->b_bufobj = NULL;
1636 * release the physical I/O buffer
1640 ((bp->b_iocmd == BIO_READ) ? &nsw_rcount :
1641 ((bp->b_flags & B_ASYNC) ?
1650 swap_pager_nswapdev(void)
1657 * SWP_PAGER_FORCE_PAGEIN() - force a swap block to be paged in
1659 * This routine dissociates the page at the given index within an object
1660 * from its backing store, paging it in if it does not reside in memory.
1661 * If the page is paged in, it is marked dirty and placed in the laundry
1662 * queue. The page is marked dirty because it no longer has backing
1663 * store. It is placed in the laundry queue because it has not been
1664 * accessed recently. Otherwise, it would already reside in memory.
1666 * We also attempt to swap in all other pages in the swap block.
1667 * However, we only guarantee that the one at the specified index is
1670 * XXX - The code to page the whole block in doesn't work, so we
1671 * revert to the one-by-one behavior for now. Sigh.
1674 swp_pager_force_pagein(vm_object_t object, vm_pindex_t pindex)
1678 vm_object_pip_add(object, 1);
1679 m = vm_page_grab(object, pindex, VM_ALLOC_NORMAL);
1680 if (m->valid == VM_PAGE_BITS_ALL) {
1681 vm_object_pip_wakeup(object);
1685 if (!vm_page_wired(m) && m->queue == PQ_NONE)
1686 panic("page %p is neither wired nor queued", m);
1690 vm_pager_page_unswapped(m);
1694 if (swap_pager_getpages(object, &m, 1, NULL, NULL) != VM_PAGER_OK)
1695 panic("swap_pager_force_pagein: read from swap failed");/*XXX*/
1696 vm_object_pip_wakeup(object);
1702 vm_pager_page_unswapped(m);
1706 * swap_pager_swapoff:
1708 * Page in all of the pages that have been paged out to the
1709 * given device. The corresponding blocks in the bitmap must be
1710 * marked as allocated and the device must be flagged SW_CLOSING.
1711 * There may be no processes swapped out to the device.
1713 * This routine may block.
1716 swap_pager_swapoff(struct swdevt *sp)
1723 sx_assert(&swdev_syscall_lock, SA_XLOCKED);
1727 mtx_lock(&vm_object_list_mtx);
1728 TAILQ_FOREACH(object, &vm_object_list, object_list) {
1729 if (object->type != OBJT_SWAP)
1731 mtx_unlock(&vm_object_list_mtx);
1732 /* Depends on type-stability. */
1733 VM_OBJECT_WLOCK(object);
1736 * Dead objects are eventually terminated on their own.
1738 if ((object->flags & OBJ_DEAD) != 0)
1742 * Sync with fences placed after pctrie
1743 * initialization. We must not access pctrie below
1744 * unless we checked that our object is swap and not
1747 atomic_thread_fence_acq();
1748 if (object->type != OBJT_SWAP)
1751 for (pi = 0; (sb = SWAP_PCTRIE_LOOKUP_GE(
1752 &object->un_pager.swp.swp_blks, pi)) != NULL; ) {
1753 pi = sb->p + SWAP_META_PAGES;
1754 for (i = 0; i < SWAP_META_PAGES; i++) {
1755 if (sb->d[i] == SWAPBLK_NONE)
1757 if (swp_pager_isondev(sb->d[i], sp))
1758 swp_pager_force_pagein(object,
1763 VM_OBJECT_WUNLOCK(object);
1764 mtx_lock(&vm_object_list_mtx);
1766 mtx_unlock(&vm_object_list_mtx);
1770 * Objects may be locked or paging to the device being
1771 * removed, so we will miss their pages and need to
1772 * make another pass. We have marked this device as
1773 * SW_CLOSING, so the activity should finish soon.
1776 if (retries > 100) {
1777 panic("swapoff: failed to locate %d swap blocks",
1780 pause("swpoff", hz / 20);
1783 EVENTHANDLER_INVOKE(swapoff, sp);
1786 /************************************************************************
1788 ************************************************************************
1790 * These routines manipulate the swap metadata stored in the
1793 * Swap metadata is implemented with a global hash and not directly
1794 * linked into the object. Instead the object simply contains
1795 * appropriate tracking counters.
1799 * SWP_PAGER_SWBLK_EMPTY() - is a range of blocks free?
1802 swp_pager_swblk_empty(struct swblk *sb, int start, int limit)
1806 MPASS(0 <= start && start <= limit && limit <= SWAP_META_PAGES);
1807 for (i = start; i < limit; i++) {
1808 if (sb->d[i] != SWAPBLK_NONE)
1815 * SWP_PAGER_META_BUILD() - add swap block to swap meta data for object
1817 * We first convert the object to a swap object if it is a default
1820 * The specified swapblk is added to the object's swap metadata. If
1821 * the swapblk is not valid, it is freed instead. Any previously
1822 * assigned swapblk is returned.
1825 swp_pager_meta_build(vm_object_t object, vm_pindex_t pindex, daddr_t swapblk)
1827 static volatile int swblk_zone_exhausted, swpctrie_zone_exhausted;
1828 struct swblk *sb, *sb1;
1829 vm_pindex_t modpi, rdpi;
1830 daddr_t prev_swapblk;
1833 VM_OBJECT_ASSERT_WLOCKED(object);
1836 * Convert default object to swap object if necessary
1838 if (object->type != OBJT_SWAP) {
1839 pctrie_init(&object->un_pager.swp.swp_blks);
1842 * Ensure that swap_pager_swapoff()'s iteration over
1843 * object_list does not see a garbage pctrie.
1845 atomic_thread_fence_rel();
1847 object->type = OBJT_SWAP;
1848 KASSERT(object->handle == NULL, ("default pager with handle"));
1851 rdpi = rounddown(pindex, SWAP_META_PAGES);
1852 sb = SWAP_PCTRIE_LOOKUP(&object->un_pager.swp.swp_blks, rdpi);
1854 if (swapblk == SWAPBLK_NONE)
1855 return (SWAPBLK_NONE);
1857 sb = uma_zalloc(swblk_zone, M_NOWAIT | (curproc ==
1858 pageproc ? M_USE_RESERVE : 0));
1861 for (i = 0; i < SWAP_META_PAGES; i++)
1862 sb->d[i] = SWAPBLK_NONE;
1863 if (atomic_cmpset_int(&swblk_zone_exhausted,
1865 printf("swblk zone ok\n");
1868 VM_OBJECT_WUNLOCK(object);
1869 if (uma_zone_exhausted(swblk_zone)) {
1870 if (atomic_cmpset_int(&swblk_zone_exhausted,
1872 printf("swap blk zone exhausted, "
1873 "increase kern.maxswzone\n");
1874 vm_pageout_oom(VM_OOM_SWAPZ);
1875 pause("swzonxb", 10);
1877 uma_zwait(swblk_zone);
1878 VM_OBJECT_WLOCK(object);
1879 sb = SWAP_PCTRIE_LOOKUP(&object->un_pager.swp.swp_blks,
1883 * Somebody swapped out a nearby page,
1884 * allocating swblk at the rdpi index,
1885 * while we dropped the object lock.
1890 error = SWAP_PCTRIE_INSERT(
1891 &object->un_pager.swp.swp_blks, sb);
1893 if (atomic_cmpset_int(&swpctrie_zone_exhausted,
1895 printf("swpctrie zone ok\n");
1898 VM_OBJECT_WUNLOCK(object);
1899 if (uma_zone_exhausted(swpctrie_zone)) {
1900 if (atomic_cmpset_int(&swpctrie_zone_exhausted,
1902 printf("swap pctrie zone exhausted, "
1903 "increase kern.maxswzone\n");
1904 vm_pageout_oom(VM_OOM_SWAPZ);
1905 pause("swzonxp", 10);
1907 uma_zwait(swpctrie_zone);
1908 VM_OBJECT_WLOCK(object);
1909 sb1 = SWAP_PCTRIE_LOOKUP(&object->un_pager.swp.swp_blks,
1912 uma_zfree(swblk_zone, sb);
1919 MPASS(sb->p == rdpi);
1921 modpi = pindex % SWAP_META_PAGES;
1922 /* Return prior contents of metadata. */
1923 prev_swapblk = sb->d[modpi];
1924 /* Enter block into metadata. */
1925 sb->d[modpi] = swapblk;
1928 * Free the swblk if we end up with the empty page run.
1930 if (swapblk == SWAPBLK_NONE &&
1931 swp_pager_swblk_empty(sb, 0, SWAP_META_PAGES)) {
1932 SWAP_PCTRIE_REMOVE(&object->un_pager.swp.swp_blks, rdpi);
1933 uma_zfree(swblk_zone, sb);
1935 return (prev_swapblk);
1939 * SWP_PAGER_META_FREE() - free a range of blocks in the object's swap metadata
1941 * The requested range of blocks is freed, with any associated swap
1942 * returned to the swap bitmap.
1944 * This routine will free swap metadata structures as they are cleaned
1945 * out. This routine does *NOT* operate on swap metadata associated
1946 * with resident pages.
1949 swp_pager_meta_free(vm_object_t object, vm_pindex_t pindex, vm_pindex_t count)
1952 daddr_t n_free, s_free;
1954 int i, limit, start;
1956 VM_OBJECT_ASSERT_WLOCKED(object);
1957 if (object->type != OBJT_SWAP || count == 0)
1960 swp_pager_init_freerange(&s_free, &n_free);
1961 last = pindex + count;
1963 sb = SWAP_PCTRIE_LOOKUP_GE(&object->un_pager.swp.swp_blks,
1964 rounddown(pindex, SWAP_META_PAGES));
1965 if (sb == NULL || sb->p >= last)
1967 start = pindex > sb->p ? pindex - sb->p : 0;
1968 limit = last - sb->p < SWAP_META_PAGES ? last - sb->p :
1970 for (i = start; i < limit; i++) {
1971 if (sb->d[i] == SWAPBLK_NONE)
1973 swp_pager_update_freerange(&s_free, &n_free, sb->d[i]);
1974 sb->d[i] = SWAPBLK_NONE;
1976 pindex = sb->p + SWAP_META_PAGES;
1977 if (swp_pager_swblk_empty(sb, 0, start) &&
1978 swp_pager_swblk_empty(sb, limit, SWAP_META_PAGES)) {
1979 SWAP_PCTRIE_REMOVE(&object->un_pager.swp.swp_blks,
1981 uma_zfree(swblk_zone, sb);
1984 swp_pager_freeswapspace(s_free, n_free);
1988 * SWP_PAGER_META_FREE_ALL() - destroy all swap metadata associated with object
1990 * This routine locates and destroys all swap metadata associated with
1994 swp_pager_meta_free_all(vm_object_t object)
1997 daddr_t n_free, s_free;
2001 VM_OBJECT_ASSERT_WLOCKED(object);
2002 if (object->type != OBJT_SWAP)
2005 swp_pager_init_freerange(&s_free, &n_free);
2006 for (pindex = 0; (sb = SWAP_PCTRIE_LOOKUP_GE(
2007 &object->un_pager.swp.swp_blks, pindex)) != NULL;) {
2008 pindex = sb->p + SWAP_META_PAGES;
2009 for (i = 0; i < SWAP_META_PAGES; i++) {
2010 if (sb->d[i] == SWAPBLK_NONE)
2012 swp_pager_update_freerange(&s_free, &n_free, sb->d[i]);
2014 SWAP_PCTRIE_REMOVE(&object->un_pager.swp.swp_blks, sb->p);
2015 uma_zfree(swblk_zone, sb);
2017 swp_pager_freeswapspace(s_free, n_free);
2021 * SWP_PAGER_METACTL() - misc control of swap meta data.
2023 * This routine is capable of looking up, or removing swapblk
2024 * assignments in the swap meta data. It returns the swapblk being
2025 * looked-up, popped, or SWAPBLK_NONE if the block was invalid.
2027 * When acting on a busy resident page and paging is in progress, we
2028 * have to wait until paging is complete but otherwise can act on the
2031 * SWM_POP remove from meta data but do not free it
2034 swp_pager_meta_ctl(vm_object_t object, vm_pindex_t pindex, int flags)
2039 if ((flags & SWM_POP) != 0)
2040 VM_OBJECT_ASSERT_WLOCKED(object);
2042 VM_OBJECT_ASSERT_LOCKED(object);
2045 * The meta data only exists if the object is OBJT_SWAP
2046 * and even then might not be allocated yet.
2048 if (object->type != OBJT_SWAP)
2049 return (SWAPBLK_NONE);
2051 sb = SWAP_PCTRIE_LOOKUP(&object->un_pager.swp.swp_blks,
2052 rounddown(pindex, SWAP_META_PAGES));
2054 return (SWAPBLK_NONE);
2055 r1 = sb->d[pindex % SWAP_META_PAGES];
2056 if (r1 == SWAPBLK_NONE)
2057 return (SWAPBLK_NONE);
2058 if ((flags & SWM_POP) != 0) {
2059 sb->d[pindex % SWAP_META_PAGES] = SWAPBLK_NONE;
2060 if (swp_pager_swblk_empty(sb, 0, SWAP_META_PAGES)) {
2061 SWAP_PCTRIE_REMOVE(&object->un_pager.swp.swp_blks,
2062 rounddown(pindex, SWAP_META_PAGES));
2063 uma_zfree(swblk_zone, sb);
2070 * Returns the least page index which is greater than or equal to the
2071 * parameter pindex and for which there is a swap block allocated.
2072 * Returns object's size if the object's type is not swap or if there
2073 * are no allocated swap blocks for the object after the requested
2077 swap_pager_find_least(vm_object_t object, vm_pindex_t pindex)
2082 VM_OBJECT_ASSERT_LOCKED(object);
2083 if (object->type != OBJT_SWAP)
2084 return (object->size);
2086 sb = SWAP_PCTRIE_LOOKUP_GE(&object->un_pager.swp.swp_blks,
2087 rounddown(pindex, SWAP_META_PAGES));
2089 return (object->size);
2090 if (sb->p < pindex) {
2091 for (i = pindex % SWAP_META_PAGES; i < SWAP_META_PAGES; i++) {
2092 if (sb->d[i] != SWAPBLK_NONE)
2095 sb = SWAP_PCTRIE_LOOKUP_GE(&object->un_pager.swp.swp_blks,
2096 roundup(pindex, SWAP_META_PAGES));
2098 return (object->size);
2100 for (i = 0; i < SWAP_META_PAGES; i++) {
2101 if (sb->d[i] != SWAPBLK_NONE)
2106 * We get here if a swblk is present in the trie but it
2107 * doesn't map any blocks.
2110 return (object->size);
2114 * System call swapon(name) enables swapping on device name,
2115 * which must be in the swdevsw. Return EBUSY
2116 * if already swapping on this device.
2118 #ifndef _SYS_SYSPROTO_H_
2119 struct swapon_args {
2129 sys_swapon(struct thread *td, struct swapon_args *uap)
2133 struct nameidata nd;
2136 error = priv_check(td, PRIV_SWAPON);
2140 sx_xlock(&swdev_syscall_lock);
2143 * Swap metadata may not fit in the KVM if we have physical
2146 if (swblk_zone == NULL) {
2151 NDINIT(&nd, LOOKUP, ISOPEN | FOLLOW | AUDITVNODE1, UIO_USERSPACE,
2157 NDFREE(&nd, NDF_ONLY_PNBUF);
2160 if (vn_isdisk(vp, &error)) {
2161 error = swapongeom(vp);
2162 } else if (vp->v_type == VREG &&
2163 (vp->v_mount->mnt_vfc->vfc_flags & VFCF_NETWORK) != 0 &&
2164 (error = VOP_GETATTR(vp, &attr, td->td_ucred)) == 0) {
2166 * Allow direct swapping to NFS regular files in the same
2167 * way that nfs_mountroot() sets up diskless swapping.
2169 error = swaponvp(td, vp, attr.va_size / DEV_BSIZE);
2175 sx_xunlock(&swdev_syscall_lock);
2180 * Check that the total amount of swap currently configured does not
2181 * exceed half the theoretical maximum. If it does, print a warning
2185 swapon_check_swzone(void)
2187 unsigned long maxpages, npages;
2189 npages = swap_total;
2190 /* absolute maximum we can handle assuming 100% efficiency */
2191 maxpages = uma_zone_get_max(swblk_zone) * SWAP_META_PAGES;
2193 /* recommend using no more than half that amount */
2194 if (npages > maxpages / 2) {
2195 printf("warning: total configured swap (%lu pages) "
2196 "exceeds maximum recommended amount (%lu pages).\n",
2197 npages, maxpages / 2);
2198 printf("warning: increase kern.maxswzone "
2199 "or reduce amount of swap.\n");
2204 swaponsomething(struct vnode *vp, void *id, u_long nblks,
2205 sw_strategy_t *strategy, sw_close_t *close, dev_t dev, int flags)
2207 struct swdevt *sp, *tsp;
2212 * nblks is in DEV_BSIZE'd chunks, convert to PAGE_SIZE'd chunks.
2213 * First chop nblks off to page-align it, then convert.
2215 * sw->sw_nblks is in page-sized chunks now too.
2217 nblks &= ~(ctodb(1) - 1);
2218 nblks = dbtoc(nblks);
2221 * If we go beyond this, we get overflows in the radix
2224 mblocks = 0x40000000 / BLIST_META_RADIX;
2225 if (nblks > mblocks) {
2227 "WARNING: reducing swap size to maximum of %luMB per unit\n",
2228 mblocks / 1024 / 1024 * PAGE_SIZE);
2232 sp = malloc(sizeof *sp, M_VMPGDATA, M_WAITOK | M_ZERO);
2236 sp->sw_nblks = nblks;
2238 sp->sw_strategy = strategy;
2239 sp->sw_close = close;
2240 sp->sw_flags = flags;
2242 sp->sw_blist = blist_create(nblks, M_WAITOK);
2244 * Do not free the first blocks in order to avoid overwriting
2245 * any bsd label at the front of the partition
2247 blist_free(sp->sw_blist, howmany(BBSIZE, PAGE_SIZE),
2248 nblks - howmany(BBSIZE, PAGE_SIZE));
2251 mtx_lock(&sw_dev_mtx);
2252 TAILQ_FOREACH(tsp, &swtailq, sw_list) {
2253 if (tsp->sw_end >= dvbase) {
2255 * We put one uncovered page between the devices
2256 * in order to definitively prevent any cross-device
2259 dvbase = tsp->sw_end + 1;
2262 sp->sw_first = dvbase;
2263 sp->sw_end = dvbase + nblks;
2264 TAILQ_INSERT_TAIL(&swtailq, sp, sw_list);
2266 swap_pager_avail += nblks - howmany(BBSIZE, PAGE_SIZE);
2267 swap_total += nblks;
2268 swapon_check_swzone();
2270 mtx_unlock(&sw_dev_mtx);
2271 EVENTHANDLER_INVOKE(swapon, sp);
2275 * SYSCALL: swapoff(devname)
2277 * Disable swapping on the given device.
2279 * XXX: Badly designed system call: it should use a device index
2280 * rather than filename as specification. We keep sw_vp around
2281 * only to make this work.
2283 #ifndef _SYS_SYSPROTO_H_
2284 struct swapoff_args {
2294 sys_swapoff(struct thread *td, struct swapoff_args *uap)
2297 struct nameidata nd;
2301 error = priv_check(td, PRIV_SWAPOFF);
2305 sx_xlock(&swdev_syscall_lock);
2307 NDINIT(&nd, LOOKUP, FOLLOW | AUDITVNODE1, UIO_USERSPACE, uap->name,
2312 NDFREE(&nd, NDF_ONLY_PNBUF);
2315 mtx_lock(&sw_dev_mtx);
2316 TAILQ_FOREACH(sp, &swtailq, sw_list) {
2317 if (sp->sw_vp == vp)
2320 mtx_unlock(&sw_dev_mtx);
2325 error = swapoff_one(sp, td->td_ucred);
2327 sx_xunlock(&swdev_syscall_lock);
2332 swapoff_one(struct swdevt *sp, struct ucred *cred)
2339 sx_assert(&swdev_syscall_lock, SA_XLOCKED);
2341 (void) vn_lock(sp->sw_vp, LK_EXCLUSIVE | LK_RETRY);
2342 error = mac_system_check_swapoff(cred, sp->sw_vp);
2343 (void) VOP_UNLOCK(sp->sw_vp, 0);
2347 nblks = sp->sw_nblks;
2350 * We can turn off this swap device safely only if the
2351 * available virtual memory in the system will fit the amount
2352 * of data we will have to page back in, plus an epsilon so
2353 * the system doesn't become critically low on swap space.
2355 if (vm_free_count() + swap_pager_avail < nblks + nswap_lowat)
2359 * Prevent further allocations on this device.
2361 mtx_lock(&sw_dev_mtx);
2362 sp->sw_flags |= SW_CLOSING;
2363 swap_pager_avail -= blist_fill(sp->sw_blist, 0, nblks);
2364 swap_total -= nblks;
2365 mtx_unlock(&sw_dev_mtx);
2368 * Page in the contents of the device and close it.
2370 swap_pager_swapoff(sp);
2372 sp->sw_close(curthread, sp);
2373 mtx_lock(&sw_dev_mtx);
2375 TAILQ_REMOVE(&swtailq, sp, sw_list);
2377 if (nswapdev == 0) {
2378 swap_pager_full = 2;
2379 swap_pager_almost_full = 1;
2383 mtx_unlock(&sw_dev_mtx);
2384 blist_destroy(sp->sw_blist);
2385 free(sp, M_VMPGDATA);
2392 struct swdevt *sp, *spt;
2393 const char *devname;
2396 sx_xlock(&swdev_syscall_lock);
2398 mtx_lock(&sw_dev_mtx);
2399 TAILQ_FOREACH_SAFE(sp, &swtailq, sw_list, spt) {
2400 mtx_unlock(&sw_dev_mtx);
2401 if (vn_isdisk(sp->sw_vp, NULL))
2402 devname = devtoname(sp->sw_vp->v_rdev);
2405 error = swapoff_one(sp, thread0.td_ucred);
2407 printf("Cannot remove swap device %s (error=%d), "
2408 "skipping.\n", devname, error);
2409 } else if (bootverbose) {
2410 printf("Swap device %s removed.\n", devname);
2412 mtx_lock(&sw_dev_mtx);
2414 mtx_unlock(&sw_dev_mtx);
2416 sx_xunlock(&swdev_syscall_lock);
2420 swap_pager_status(int *total, int *used)
2426 mtx_lock(&sw_dev_mtx);
2427 TAILQ_FOREACH(sp, &swtailq, sw_list) {
2428 *total += sp->sw_nblks;
2429 *used += sp->sw_used;
2431 mtx_unlock(&sw_dev_mtx);
2435 swap_dev_info(int name, struct xswdev *xs, char *devname, size_t len)
2438 const char *tmp_devname;
2443 mtx_lock(&sw_dev_mtx);
2444 TAILQ_FOREACH(sp, &swtailq, sw_list) {
2449 xs->xsw_version = XSWDEV_VERSION;
2450 xs->xsw_dev = sp->sw_dev;
2451 xs->xsw_flags = sp->sw_flags;
2452 xs->xsw_nblks = sp->sw_nblks;
2453 xs->xsw_used = sp->sw_used;
2454 if (devname != NULL) {
2455 if (vn_isdisk(sp->sw_vp, NULL))
2456 tmp_devname = devtoname(sp->sw_vp->v_rdev);
2458 tmp_devname = "[file]";
2459 strncpy(devname, tmp_devname, len);
2464 mtx_unlock(&sw_dev_mtx);
2468 #if defined(COMPAT_FREEBSD11)
2469 #define XSWDEV_VERSION_11 1
2479 #if defined(__amd64__) && defined(COMPAT_FREEBSD32)
2482 u_int xsw_dev1, xsw_dev2;
2490 sysctl_vm_swap_info(SYSCTL_HANDLER_ARGS)
2493 #if defined(__amd64__) && defined(COMPAT_FREEBSD32)
2494 struct xswdev32 xs32;
2496 #if defined(COMPAT_FREEBSD11)
2497 struct xswdev11 xs11;
2501 if (arg2 != 1) /* name length */
2503 error = swap_dev_info(*(int *)arg1, &xs, NULL, 0);
2506 #if defined(__amd64__) && defined(COMPAT_FREEBSD32)
2507 if (req->oldlen == sizeof(xs32)) {
2508 xs32.xsw_version = XSWDEV_VERSION;
2509 xs32.xsw_dev1 = xs.xsw_dev;
2510 xs32.xsw_dev2 = xs.xsw_dev >> 32;
2511 xs32.xsw_flags = xs.xsw_flags;
2512 xs32.xsw_nblks = xs.xsw_nblks;
2513 xs32.xsw_used = xs.xsw_used;
2514 error = SYSCTL_OUT(req, &xs32, sizeof(xs32));
2518 #if defined(COMPAT_FREEBSD11)
2519 if (req->oldlen == sizeof(xs11)) {
2520 xs11.xsw_version = XSWDEV_VERSION_11;
2521 xs11.xsw_dev = xs.xsw_dev; /* truncation */
2522 xs11.xsw_flags = xs.xsw_flags;
2523 xs11.xsw_nblks = xs.xsw_nblks;
2524 xs11.xsw_used = xs.xsw_used;
2525 error = SYSCTL_OUT(req, &xs11, sizeof(xs11));
2529 error = SYSCTL_OUT(req, &xs, sizeof(xs));
2533 SYSCTL_INT(_vm, OID_AUTO, nswapdev, CTLFLAG_RD, &nswapdev, 0,
2534 "Number of swap devices");
2535 SYSCTL_NODE(_vm, OID_AUTO, swap_info, CTLFLAG_RD | CTLFLAG_MPSAFE,
2536 sysctl_vm_swap_info,
2537 "Swap statistics by device");
2540 * Count the approximate swap usage in pages for a vmspace. The
2541 * shadowed or not yet copied on write swap blocks are not accounted.
2542 * The map must be locked.
2545 vmspace_swap_count(struct vmspace *vmspace)
2555 map = &vmspace->vm_map;
2558 for (cur = map->header.next; cur != &map->header; cur = cur->next) {
2559 if ((cur->eflags & MAP_ENTRY_IS_SUB_MAP) != 0)
2561 object = cur->object.vm_object;
2562 if (object == NULL || object->type != OBJT_SWAP)
2564 VM_OBJECT_RLOCK(object);
2565 if (object->type != OBJT_SWAP)
2567 pi = OFF_TO_IDX(cur->offset);
2568 e = pi + OFF_TO_IDX(cur->end - cur->start);
2569 for (;; pi = sb->p + SWAP_META_PAGES) {
2570 sb = SWAP_PCTRIE_LOOKUP_GE(
2571 &object->un_pager.swp.swp_blks, pi);
2572 if (sb == NULL || sb->p >= e)
2574 for (i = 0; i < SWAP_META_PAGES; i++) {
2575 if (sb->p + i < e &&
2576 sb->d[i] != SWAPBLK_NONE)
2581 VM_OBJECT_RUNLOCK(object);
2589 * Swapping onto disk devices.
2593 static g_orphan_t swapgeom_orphan;
2595 static struct g_class g_swap_class = {
2597 .version = G_VERSION,
2598 .orphan = swapgeom_orphan,
2601 DECLARE_GEOM_CLASS(g_swap_class, g_class);
2605 swapgeom_close_ev(void *arg, int flags)
2607 struct g_consumer *cp;
2610 g_access(cp, -1, -1, 0);
2612 g_destroy_consumer(cp);
2616 * Add a reference to the g_consumer for an inflight transaction.
2619 swapgeom_acquire(struct g_consumer *cp)
2622 mtx_assert(&sw_dev_mtx, MA_OWNED);
2627 * Remove a reference from the g_consumer. Post a close event if all
2628 * references go away, since the function might be called from the
2632 swapgeom_release(struct g_consumer *cp, struct swdevt *sp)
2635 mtx_assert(&sw_dev_mtx, MA_OWNED);
2637 if (cp->index == 0) {
2638 if (g_post_event(swapgeom_close_ev, cp, M_NOWAIT, NULL) == 0)
2644 swapgeom_done(struct bio *bp2)
2648 struct g_consumer *cp;
2650 bp = bp2->bio_caller2;
2652 bp->b_ioflags = bp2->bio_flags;
2654 bp->b_ioflags |= BIO_ERROR;
2655 bp->b_resid = bp->b_bcount - bp2->bio_completed;
2656 bp->b_error = bp2->bio_error;
2658 sp = bp2->bio_caller1;
2659 mtx_lock(&sw_dev_mtx);
2660 swapgeom_release(cp, sp);
2661 mtx_unlock(&sw_dev_mtx);
2666 swapgeom_strategy(struct buf *bp, struct swdevt *sp)
2669 struct g_consumer *cp;
2671 mtx_lock(&sw_dev_mtx);
2674 mtx_unlock(&sw_dev_mtx);
2675 bp->b_error = ENXIO;
2676 bp->b_ioflags |= BIO_ERROR;
2680 swapgeom_acquire(cp);
2681 mtx_unlock(&sw_dev_mtx);
2682 if (bp->b_iocmd == BIO_WRITE)
2685 bio = g_alloc_bio();
2687 mtx_lock(&sw_dev_mtx);
2688 swapgeom_release(cp, sp);
2689 mtx_unlock(&sw_dev_mtx);
2690 bp->b_error = ENOMEM;
2691 bp->b_ioflags |= BIO_ERROR;
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(&pbuf_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, &pbuf_mtx, PSWP,
2929 mtx_unlock(&pbuf_mtx);