2 * Copyright (c) 1998 Matthew Dillon,
3 * Copyright (c) 1994 John S. Dyson
4 * Copyright (c) 1990 University of Utah.
5 * Copyright (c) 1982, 1986, 1989, 1993
6 * The Regents of the University of California. All rights reserved.
8 * This code is derived from software contributed to Berkeley by
9 * the Systems Programming Group of the University of Utah Computer
12 * Redistribution and use in source and binary forms, with or without
13 * modification, are permitted provided that the following conditions
15 * 1. Redistributions of source code must retain the above copyright
16 * notice, this list of conditions and the following disclaimer.
17 * 2. Redistributions in binary form must reproduce the above copyright
18 * notice, this list of conditions and the following disclaimer in the
19 * documentation and/or other materials provided with the distribution.
20 * 3. All advertising materials mentioning features or use of this software
21 * must display the following acknowledgement:
22 * This product includes software developed by the University of
23 * California, Berkeley and its contributors.
24 * 4. Neither the name of the University nor the names of its contributors
25 * may be used to endorse or promote products derived from this software
26 * without specific prior written permission.
28 * THIS SOFTWARE IS PROVIDED BY THE REGENTS AND CONTRIBUTORS ``AS IS'' AND
29 * ANY EXPRESS OR IMPLIED WARRANTIES, INCLUDING, BUT NOT LIMITED TO, THE
30 * IMPLIED WARRANTIES OF MERCHANTABILITY AND FITNESS FOR A PARTICULAR PURPOSE
31 * ARE DISCLAIMED. IN NO EVENT SHALL THE REGENTS OR CONTRIBUTORS BE LIABLE
32 * FOR ANY DIRECT, INDIRECT, INCIDENTAL, SPECIAL, EXEMPLARY, OR CONSEQUENTIAL
33 * DAMAGES (INCLUDING, BUT NOT LIMITED TO, PROCUREMENT OF SUBSTITUTE GOODS
34 * OR SERVICES; LOSS OF USE, DATA, OR PROFITS; OR BUSINESS INTERRUPTION)
35 * HOWEVER CAUSED AND ON ANY THEORY OF LIABILITY, WHETHER IN CONTRACT, STRICT
36 * LIABILITY, OR TORT (INCLUDING NEGLIGENCE OR OTHERWISE) ARISING IN ANY WAY
37 * OUT OF THE USE OF THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF
43 * Radix Bitmap 'blists'.
45 * - The new swapper uses the new radix bitmap code. This should scale
46 * to arbitrarily small or arbitrarily large swap spaces and an almost
47 * arbitrary degree of fragmentation.
51 * - on the fly reallocation of swap during putpages. The new system
52 * does not try to keep previously allocated swap blocks for dirty
55 * - on the fly deallocation of swap
57 * - No more garbage collection required. Unnecessarily allocated swap
58 * blocks only exist for dirty vm_page_t's now and these are already
59 * cycled (in a high-load system) by the pager. We also do on-the-fly
60 * removal of invalidated swap blocks when a page is destroyed
63 * from: Utah $Hdr: swap_pager.c 1.4 91/04/30$
65 * @(#)swap_pager.c 8.9 (Berkeley) 3/21/94
66 * @(#)vm_swap.c 8.5 (Berkeley) 2/17/94
69 #include <sys/cdefs.h>
70 __FBSDID("$FreeBSD$");
72 #include "opt_compat.h"
76 #include <sys/param.h>
77 #include <sys/systm.h>
79 #include <sys/kernel.h>
85 #include <sys/fcntl.h>
86 #include <sys/mount.h>
87 #include <sys/namei.h>
88 #include <sys/vnode.h>
89 #include <sys/malloc.h>
90 #include <sys/pctrie.h>
91 #include <sys/racct.h>
92 #include <sys/resource.h>
93 #include <sys/resourcevar.h>
94 #include <sys/rwlock.h>
95 #include <sys/sysctl.h>
96 #include <sys/sysproto.h>
97 #include <sys/blist.h>
100 #include <sys/vmmeter.h>
102 #include <security/mac/mac_framework.h>
106 #include <vm/vm_map.h>
107 #include <vm/vm_kern.h>
108 #include <vm/vm_object.h>
109 #include <vm/vm_page.h>
110 #include <vm/vm_pager.h>
111 #include <vm/vm_pageout.h>
112 #include <vm/vm_param.h>
113 #include <vm/swap_pager.h>
114 #include <vm/vm_extern.h>
117 #include <geom/geom.h>
120 * MAX_PAGEOUT_CLUSTER must be a power of 2 between 1 and 64.
121 * The 64-page limit is due to the radix code (kern/subr_blist.c).
123 #ifndef MAX_PAGEOUT_CLUSTER
124 #define MAX_PAGEOUT_CLUSTER 32
127 #if !defined(SWB_NPAGES)
128 #define SWB_NPAGES MAX_PAGEOUT_CLUSTER
131 #define SWAP_META_PAGES PCTRIE_COUNT
134 * A swblk structure maps each page index within a
135 * SWAP_META_PAGES-aligned and sized range to the address of an
136 * on-disk swap block (or SWAPBLK_NONE). The collection of these
137 * mappings for an entire vm object is implemented as a pc-trie.
141 daddr_t d[SWAP_META_PAGES];
144 static MALLOC_DEFINE(M_VMPGDATA, "vm_pgdata", "swap pager private data");
145 static struct mtx sw_dev_mtx;
146 static TAILQ_HEAD(, swdevt) swtailq = TAILQ_HEAD_INITIALIZER(swtailq);
147 static struct swdevt *swdevhd; /* Allocate from here next */
148 static int nswapdev; /* Number of swap devices */
149 int swap_pager_avail;
150 static struct sx swdev_syscall_lock; /* serialize swap(on|off) */
152 static vm_ooffset_t swap_total;
153 SYSCTL_QUAD(_vm, OID_AUTO, swap_total, CTLFLAG_RD, &swap_total, 0,
154 "Total amount of available swap storage.");
155 static vm_ooffset_t swap_reserved;
156 SYSCTL_QUAD(_vm, OID_AUTO, swap_reserved, CTLFLAG_RD, &swap_reserved, 0,
157 "Amount of swap storage needed to back all allocated anonymous memory.");
158 static int overcommit = 0;
159 SYSCTL_INT(_vm, OID_AUTO, overcommit, CTLFLAG_RW, &overcommit, 0,
160 "Configure virtual memory overcommit behavior. See tuning(7) "
162 static unsigned long swzone;
163 SYSCTL_ULONG(_vm, OID_AUTO, swzone, CTLFLAG_RD, &swzone, 0,
164 "Actual size of swap metadata zone");
165 static unsigned long swap_maxpages;
166 SYSCTL_ULONG(_vm, OID_AUTO, swap_maxpages, CTLFLAG_RD, &swap_maxpages, 0,
167 "Maximum amount of swap supported");
169 /* bits from overcommit */
170 #define SWAP_RESERVE_FORCE_ON (1 << 0)
171 #define SWAP_RESERVE_RLIMIT_ON (1 << 1)
172 #define SWAP_RESERVE_ALLOW_NONWIRED (1 << 2)
175 swap_reserve(vm_ooffset_t incr)
178 return (swap_reserve_by_cred(incr, curthread->td_ucred));
182 swap_reserve_by_cred(vm_ooffset_t incr, struct ucred *cred)
187 static struct timeval lastfail;
190 uip = cred->cr_ruidinfo;
192 if (incr & PAGE_MASK)
193 panic("swap_reserve: & PAGE_MASK");
198 error = racct_add(curproc, RACCT_SWAP, incr);
199 PROC_UNLOCK(curproc);
206 mtx_lock(&sw_dev_mtx);
207 r = swap_reserved + incr;
208 if (overcommit & SWAP_RESERVE_ALLOW_NONWIRED) {
209 s = vm_cnt.v_page_count - vm_cnt.v_free_reserved - vm_cnt.v_wire_count;
214 if ((overcommit & SWAP_RESERVE_FORCE_ON) == 0 || r <= s ||
215 (error = priv_check(curthread, PRIV_VM_SWAP_NOQUOTA)) == 0) {
219 mtx_unlock(&sw_dev_mtx);
222 UIDINFO_VMSIZE_LOCK(uip);
223 if ((overcommit & SWAP_RESERVE_RLIMIT_ON) != 0 &&
224 uip->ui_vmsize + incr > lim_cur(curthread, RLIMIT_SWAP) &&
225 priv_check(curthread, PRIV_VM_SWAP_NORLIMIT))
228 uip->ui_vmsize += incr;
229 UIDINFO_VMSIZE_UNLOCK(uip);
231 mtx_lock(&sw_dev_mtx);
232 swap_reserved -= incr;
233 mtx_unlock(&sw_dev_mtx);
236 if (!res && ppsratecheck(&lastfail, &curfail, 1)) {
237 printf("uid %d, pid %d: swap reservation for %jd bytes failed\n",
238 uip->ui_uid, curproc->p_pid, incr);
244 racct_sub(curproc, RACCT_SWAP, incr);
245 PROC_UNLOCK(curproc);
253 swap_reserve_force(vm_ooffset_t incr)
257 mtx_lock(&sw_dev_mtx);
258 swap_reserved += incr;
259 mtx_unlock(&sw_dev_mtx);
263 racct_add_force(curproc, RACCT_SWAP, incr);
264 PROC_UNLOCK(curproc);
267 uip = curthread->td_ucred->cr_ruidinfo;
269 UIDINFO_VMSIZE_LOCK(uip);
270 uip->ui_vmsize += incr;
271 UIDINFO_VMSIZE_UNLOCK(uip);
272 PROC_UNLOCK(curproc);
276 swap_release(vm_ooffset_t decr)
281 cred = curthread->td_ucred;
282 swap_release_by_cred(decr, cred);
283 PROC_UNLOCK(curproc);
287 swap_release_by_cred(vm_ooffset_t decr, struct ucred *cred)
291 uip = cred->cr_ruidinfo;
293 if (decr & PAGE_MASK)
294 panic("swap_release: & PAGE_MASK");
296 mtx_lock(&sw_dev_mtx);
297 if (swap_reserved < decr)
298 panic("swap_reserved < decr");
299 swap_reserved -= decr;
300 mtx_unlock(&sw_dev_mtx);
302 UIDINFO_VMSIZE_LOCK(uip);
303 if (uip->ui_vmsize < decr)
304 printf("negative vmsize for uid = %d\n", uip->ui_uid);
305 uip->ui_vmsize -= decr;
306 UIDINFO_VMSIZE_UNLOCK(uip);
308 racct_sub_cred(cred, RACCT_SWAP, decr);
311 #define SWM_FREE 0x02 /* free, period */
312 #define SWM_POP 0x04 /* pop out */
314 int swap_pager_full = 2; /* swap space exhaustion (task killing) */
315 static int swap_pager_almost_full = 1; /* swap space exhaustion (w/hysteresis)*/
316 static int nsw_rcount; /* free read buffers */
317 static int nsw_wcount_sync; /* limit write buffers / synchronous */
318 static int nsw_wcount_async; /* limit write buffers / asynchronous */
319 static int nsw_wcount_async_max;/* assigned maximum */
320 static int nsw_cluster_max; /* maximum VOP I/O allowed */
322 static int sysctl_swap_async_max(SYSCTL_HANDLER_ARGS);
323 SYSCTL_PROC(_vm, OID_AUTO, swap_async_max, CTLTYPE_INT | CTLFLAG_RW |
324 CTLFLAG_MPSAFE, NULL, 0, sysctl_swap_async_max, "I",
325 "Maximum running async swap ops");
327 static struct sx sw_alloc_sx;
330 * "named" and "unnamed" anon region objects. Try to reduce the overhead
331 * of searching a named list by hashing it just a little.
336 #define NOBJLIST(handle) \
337 (&swap_pager_object_list[((int)(intptr_t)handle >> 4) & (NOBJLISTS-1)])
339 static struct pagerlst swap_pager_object_list[NOBJLISTS];
340 static uma_zone_t swblk_zone;
341 static uma_zone_t swpctrie_zone;
344 * pagerops for OBJT_SWAP - "swap pager". Some ops are also global procedure
345 * calls hooked from other parts of the VM system and do not appear here.
346 * (see vm/swap_pager.h).
349 swap_pager_alloc(void *handle, vm_ooffset_t size,
350 vm_prot_t prot, vm_ooffset_t offset, struct ucred *);
351 static void swap_pager_dealloc(vm_object_t object);
352 static int swap_pager_getpages(vm_object_t, vm_page_t *, int, int *,
354 static int swap_pager_getpages_async(vm_object_t, vm_page_t *, int, int *,
355 int *, pgo_getpages_iodone_t, void *);
356 static void swap_pager_putpages(vm_object_t, vm_page_t *, int, boolean_t, int *);
358 swap_pager_haspage(vm_object_t object, vm_pindex_t pindex, int *before, int *after);
359 static void swap_pager_init(void);
360 static void swap_pager_unswapped(vm_page_t);
361 static void swap_pager_swapoff(struct swdevt *sp);
363 struct pagerops swappagerops = {
364 .pgo_init = swap_pager_init, /* early system initialization of pager */
365 .pgo_alloc = swap_pager_alloc, /* allocate an OBJT_SWAP object */
366 .pgo_dealloc = swap_pager_dealloc, /* deallocate an OBJT_SWAP object */
367 .pgo_getpages = swap_pager_getpages, /* pagein */
368 .pgo_getpages_async = swap_pager_getpages_async, /* pagein (async) */
369 .pgo_putpages = swap_pager_putpages, /* pageout */
370 .pgo_haspage = swap_pager_haspage, /* get backing store status for page */
371 .pgo_pageunswapped = swap_pager_unswapped, /* remove swap related to page */
375 * swap_*() routines are externally accessible. swp_*() routines are
378 static int nswap_lowat = 128; /* in pages, swap_pager_almost_full warn */
379 static int nswap_hiwat = 512; /* in pages, swap_pager_almost_full warn */
381 SYSCTL_INT(_vm, OID_AUTO, dmmax, CTLFLAG_RD, &nsw_cluster_max, 0,
382 "Maximum size of a swap block in pages");
384 static void swp_sizecheck(void);
385 static void swp_pager_async_iodone(struct buf *bp);
386 static int swapongeom(struct vnode *);
387 static int swaponvp(struct thread *, struct vnode *, u_long);
388 static int swapoff_one(struct swdevt *sp, struct ucred *cred);
391 * Swap bitmap functions
393 static void swp_pager_freeswapspace(daddr_t blk, int npages);
394 static daddr_t swp_pager_getswapspace(int npages);
399 static void swp_pager_meta_build(vm_object_t, vm_pindex_t, daddr_t);
400 static void swp_pager_meta_free(vm_object_t, vm_pindex_t, vm_pindex_t);
401 static void swp_pager_meta_free_all(vm_object_t);
402 static daddr_t swp_pager_meta_ctl(vm_object_t, vm_pindex_t, int);
405 swblk_trie_alloc(struct pctrie *ptree)
408 return (uma_zalloc(swpctrie_zone, M_NOWAIT | (curproc == pageproc ?
409 M_USE_RESERVE : 0)));
413 swblk_trie_free(struct pctrie *ptree, void *node)
416 uma_zfree(swpctrie_zone, node);
419 PCTRIE_DEFINE(SWAP, swblk, p, swblk_trie_alloc, swblk_trie_free);
422 * SWP_SIZECHECK() - update swap_pager_full indication
424 * update the swap_pager_almost_full indication and warn when we are
425 * about to run out of swap space, using lowat/hiwat hysteresis.
427 * Clear swap_pager_full ( task killing ) indication when lowat is met.
429 * No restrictions on call
430 * This routine may not block.
436 if (swap_pager_avail < nswap_lowat) {
437 if (swap_pager_almost_full == 0) {
438 printf("swap_pager: out of swap space\n");
439 swap_pager_almost_full = 1;
443 if (swap_pager_avail > nswap_hiwat)
444 swap_pager_almost_full = 0;
449 * SWAP_PAGER_INIT() - initialize the swap pager!
451 * Expected to be started from system init. NOTE: This code is run
452 * before much else so be careful what you depend on. Most of the VM
453 * system has yet to be initialized at this point.
456 swap_pager_init(void)
459 * Initialize object lists
463 for (i = 0; i < NOBJLISTS; ++i)
464 TAILQ_INIT(&swap_pager_object_list[i]);
465 mtx_init(&sw_dev_mtx, "swapdev", NULL, MTX_DEF);
466 sx_init(&sw_alloc_sx, "swspsx");
467 sx_init(&swdev_syscall_lock, "swsysc");
471 * SWAP_PAGER_SWAP_INIT() - swap pager initialization from pageout process
473 * Expected to be started from pageout process once, prior to entering
477 swap_pager_swap_init(void)
482 * Number of in-transit swap bp operations. Don't
483 * exhaust the pbufs completely. Make sure we
484 * initialize workable values (0 will work for hysteresis
485 * but it isn't very efficient).
487 * The nsw_cluster_max is constrained by the bp->b_pages[]
488 * array (MAXPHYS/PAGE_SIZE) and our locally defined
489 * MAX_PAGEOUT_CLUSTER. Also be aware that swap ops are
490 * constrained by the swap device interleave stripe size.
492 * Currently we hardwire nsw_wcount_async to 4. This limit is
493 * designed to prevent other I/O from having high latencies due to
494 * our pageout I/O. The value 4 works well for one or two active swap
495 * devices but is probably a little low if you have more. Even so,
496 * a higher value would probably generate only a limited improvement
497 * with three or four active swap devices since the system does not
498 * typically have to pageout at extreme bandwidths. We will want
499 * at least 2 per swap devices, and 4 is a pretty good value if you
500 * have one NFS swap device due to the command/ack latency over NFS.
501 * So it all works out pretty well.
503 nsw_cluster_max = min((MAXPHYS/PAGE_SIZE), MAX_PAGEOUT_CLUSTER);
506 nsw_rcount = (nswbuf + 1) / 2;
507 nsw_wcount_sync = (nswbuf + 3) / 4;
508 nsw_wcount_async = 4;
509 nsw_wcount_async_max = nsw_wcount_async;
510 mtx_unlock(&pbuf_mtx);
513 * Initialize our zone, guessing on the number we need based
514 * on the number of pages in the system.
516 n = vm_cnt.v_page_count / 2;
517 if (maxswzone && n > maxswzone / sizeof(struct swblk))
518 n = maxswzone / sizeof(struct swblk);
519 swpctrie_zone = uma_zcreate("swpctrie", pctrie_node_size(), NULL, NULL,
520 pctrie_zone_init, NULL, UMA_ALIGN_PTR,
521 UMA_ZONE_NOFREE | UMA_ZONE_VM);
522 if (swpctrie_zone == NULL)
523 panic("failed to create swap pctrie zone.");
524 swblk_zone = uma_zcreate("swblk", sizeof(struct swblk), NULL, NULL,
525 NULL, NULL, _Alignof(struct swblk) - 1,
526 UMA_ZONE_NOFREE | UMA_ZONE_VM);
527 if (swblk_zone == NULL)
528 panic("failed to create swap blk zone.");
531 if (uma_zone_reserve_kva(swblk_zone, n))
534 * if the allocation failed, try a zone two thirds the
535 * size of the previous attempt.
540 printf("Swap blk zone entries reduced from %lu to %lu.\n",
542 swap_maxpages = n * SWAP_META_PAGES;
543 swzone = n * sizeof(struct swblk);
544 if (!uma_zone_reserve_kva(swpctrie_zone, n))
545 printf("Cannot reserve swap pctrie zone, "
546 "reduce kern.maxswzone.\n");
550 swap_pager_alloc_init(void *handle, struct ucred *cred, vm_ooffset_t size,
556 if (!swap_reserve_by_cred(size, cred))
562 * The un_pager.swp.swp_blks trie is initialized by
563 * vm_object_allocate() to ensure the correct order of
564 * visibility to other threads.
566 object = vm_object_allocate(OBJT_SWAP, OFF_TO_IDX(offset +
569 object->handle = handle;
572 object->charge = size;
578 * SWAP_PAGER_ALLOC() - allocate a new OBJT_SWAP VM object and instantiate
579 * its metadata structures.
581 * This routine is called from the mmap and fork code to create a new
584 * This routine must ensure that no live duplicate is created for
585 * the named object request, which is protected against by
586 * holding the sw_alloc_sx lock in case handle != NULL.
589 swap_pager_alloc(void *handle, vm_ooffset_t size, vm_prot_t prot,
590 vm_ooffset_t offset, struct ucred *cred)
594 if (handle != NULL) {
596 * Reference existing named region or allocate new one. There
597 * should not be a race here against swp_pager_meta_build()
598 * as called from vm_page_remove() in regards to the lookup
601 sx_xlock(&sw_alloc_sx);
602 object = vm_pager_object_lookup(NOBJLIST(handle), handle);
603 if (object == NULL) {
604 object = swap_pager_alloc_init(handle, cred, size,
606 if (object != NULL) {
607 TAILQ_INSERT_TAIL(NOBJLIST(object->handle),
608 object, pager_object_list);
611 sx_xunlock(&sw_alloc_sx);
613 object = swap_pager_alloc_init(handle, cred, size, offset);
619 * SWAP_PAGER_DEALLOC() - remove swap metadata from object
621 * The swap backing for the object is destroyed. The code is
622 * designed such that we can reinstantiate it later, but this
623 * routine is typically called only when the entire object is
624 * about to be destroyed.
626 * The object must be locked.
629 swap_pager_dealloc(vm_object_t object)
632 VM_OBJECT_ASSERT_WLOCKED(object);
633 KASSERT((object->flags & OBJ_DEAD) != 0, ("dealloc of reachable obj"));
636 * Remove from list right away so lookups will fail if we block for
637 * pageout completion.
639 if (object->handle != NULL) {
640 VM_OBJECT_WUNLOCK(object);
641 sx_xlock(&sw_alloc_sx);
642 TAILQ_REMOVE(NOBJLIST(object->handle), object,
644 sx_xunlock(&sw_alloc_sx);
645 VM_OBJECT_WLOCK(object);
648 vm_object_pip_wait(object, "swpdea");
651 * Free all remaining metadata. We only bother to free it from
652 * the swap meta data. We do not attempt to free swapblk's still
653 * associated with vm_page_t's for this object. We do not care
654 * if paging is still in progress on some objects.
656 swp_pager_meta_free_all(object);
657 object->handle = NULL;
658 object->type = OBJT_DEAD;
661 /************************************************************************
662 * SWAP PAGER BITMAP ROUTINES *
663 ************************************************************************/
666 * SWP_PAGER_GETSWAPSPACE() - allocate raw swap space
668 * Allocate swap for the requested number of pages. The starting
669 * swap block number (a page index) is returned or SWAPBLK_NONE
670 * if the allocation failed.
672 * Also has the side effect of advising that somebody made a mistake
673 * when they configured swap and didn't configure enough.
675 * This routine may not sleep.
677 * We allocate in round-robin fashion from the configured devices.
680 swp_pager_getswapspace(int npages)
687 mtx_lock(&sw_dev_mtx);
689 for (i = 0; i < nswapdev; i++) {
691 sp = TAILQ_FIRST(&swtailq);
692 if (!(sp->sw_flags & SW_CLOSING)) {
693 blk = blist_alloc(sp->sw_blist, npages);
694 if (blk != SWAPBLK_NONE) {
696 sp->sw_used += npages;
697 swap_pager_avail -= npages;
699 swdevhd = TAILQ_NEXT(sp, sw_list);
703 sp = TAILQ_NEXT(sp, sw_list);
705 if (swap_pager_full != 2) {
706 printf("swap_pager_getswapspace(%d): failed\n", npages);
708 swap_pager_almost_full = 1;
712 mtx_unlock(&sw_dev_mtx);
717 swp_pager_isondev(daddr_t blk, struct swdevt *sp)
720 return (blk >= sp->sw_first && blk < sp->sw_end);
724 swp_pager_strategy(struct buf *bp)
728 mtx_lock(&sw_dev_mtx);
729 TAILQ_FOREACH(sp, &swtailq, sw_list) {
730 if (bp->b_blkno >= sp->sw_first && bp->b_blkno < sp->sw_end) {
731 mtx_unlock(&sw_dev_mtx);
732 if ((sp->sw_flags & SW_UNMAPPED) != 0 &&
733 unmapped_buf_allowed) {
734 bp->b_data = unmapped_buf;
737 pmap_qenter((vm_offset_t)bp->b_data,
738 &bp->b_pages[0], bp->b_bcount / PAGE_SIZE);
740 sp->sw_strategy(bp, sp);
744 panic("Swapdev not found");
749 * SWP_PAGER_FREESWAPSPACE() - free raw swap space
751 * This routine returns the specified swap blocks back to the bitmap.
753 * This routine may not sleep.
756 swp_pager_freeswapspace(daddr_t blk, int npages)
760 mtx_lock(&sw_dev_mtx);
761 TAILQ_FOREACH(sp, &swtailq, sw_list) {
762 if (blk >= sp->sw_first && blk < sp->sw_end) {
763 sp->sw_used -= npages;
765 * If we are attempting to stop swapping on
766 * this device, we don't want to mark any
767 * blocks free lest they be reused.
769 if ((sp->sw_flags & SW_CLOSING) == 0) {
770 blist_free(sp->sw_blist, blk - sp->sw_first,
772 swap_pager_avail += npages;
775 mtx_unlock(&sw_dev_mtx);
779 panic("Swapdev not found");
783 * SWAP_PAGER_FREESPACE() - frees swap blocks associated with a page
784 * range within an object.
786 * This is a globally accessible routine.
788 * This routine removes swapblk assignments from swap metadata.
790 * The external callers of this routine typically have already destroyed
791 * or renamed vm_page_t's associated with this range in the object so
794 * The object must be locked.
797 swap_pager_freespace(vm_object_t object, vm_pindex_t start, vm_size_t size)
800 swp_pager_meta_free(object, start, size);
804 * SWAP_PAGER_RESERVE() - reserve swap blocks in object
806 * Assigns swap blocks to the specified range within the object. The
807 * swap blocks are not zeroed. Any previous swap assignment is destroyed.
809 * Returns 0 on success, -1 on failure.
812 swap_pager_reserve(vm_object_t object, vm_pindex_t start, vm_size_t size)
815 daddr_t blk = SWAPBLK_NONE;
816 vm_pindex_t beg = start; /* save start index */
818 VM_OBJECT_WLOCK(object);
822 while ((blk = swp_pager_getswapspace(n)) == SWAPBLK_NONE) {
825 swp_pager_meta_free(object, beg, start - beg);
826 VM_OBJECT_WUNLOCK(object);
831 swp_pager_meta_build(object, start, blk);
837 swp_pager_meta_free(object, start, n);
838 VM_OBJECT_WUNLOCK(object);
843 * SWAP_PAGER_COPY() - copy blocks from source pager to destination pager
844 * and destroy the source.
846 * Copy any valid swapblks from the source to the destination. In
847 * cases where both the source and destination have a valid swapblk,
848 * we keep the destination's.
850 * This routine is allowed to sleep. It may sleep allocating metadata
851 * indirectly through swp_pager_meta_build() or if paging is still in
852 * progress on the source.
854 * The source object contains no vm_page_t's (which is just as well)
856 * The source object is of type OBJT_SWAP.
858 * The source and destination objects must be locked.
859 * Both object locks may temporarily be released.
862 swap_pager_copy(vm_object_t srcobject, vm_object_t dstobject,
863 vm_pindex_t offset, int destroysource)
867 VM_OBJECT_ASSERT_WLOCKED(srcobject);
868 VM_OBJECT_ASSERT_WLOCKED(dstobject);
871 * If destroysource is set, we remove the source object from the
872 * swap_pager internal queue now.
874 if (destroysource && srcobject->handle != NULL) {
875 vm_object_pip_add(srcobject, 1);
876 VM_OBJECT_WUNLOCK(srcobject);
877 vm_object_pip_add(dstobject, 1);
878 VM_OBJECT_WUNLOCK(dstobject);
879 sx_xlock(&sw_alloc_sx);
880 TAILQ_REMOVE(NOBJLIST(srcobject->handle), srcobject,
882 sx_xunlock(&sw_alloc_sx);
883 VM_OBJECT_WLOCK(dstobject);
884 vm_object_pip_wakeup(dstobject);
885 VM_OBJECT_WLOCK(srcobject);
886 vm_object_pip_wakeup(srcobject);
890 * transfer source to destination.
892 for (i = 0; i < dstobject->size; ++i) {
896 * Locate (without changing) the swapblk on the destination,
897 * unless it is invalid in which case free it silently, or
898 * if the destination is a resident page, in which case the
899 * source is thrown away.
901 dstaddr = swp_pager_meta_ctl(dstobject, i, 0);
903 if (dstaddr == SWAPBLK_NONE) {
905 * Destination has no swapblk and is not resident,
910 srcaddr = swp_pager_meta_ctl(
916 if (srcaddr != SWAPBLK_NONE) {
918 * swp_pager_meta_build() can sleep.
920 vm_object_pip_add(srcobject, 1);
921 VM_OBJECT_WUNLOCK(srcobject);
922 vm_object_pip_add(dstobject, 1);
923 swp_pager_meta_build(dstobject, i, srcaddr);
924 vm_object_pip_wakeup(dstobject);
925 VM_OBJECT_WLOCK(srcobject);
926 vm_object_pip_wakeup(srcobject);
930 * Destination has valid swapblk or it is represented
931 * by a resident page. We destroy the sourceblock.
934 swp_pager_meta_ctl(srcobject, i + offset, SWM_FREE);
939 * Free left over swap blocks in source.
941 * We have to revert the type to OBJT_DEFAULT so we do not accidentally
942 * double-remove the object from the swap queues.
945 swp_pager_meta_free_all(srcobject);
947 * Reverting the type is not necessary, the caller is going
948 * to destroy srcobject directly, but I'm doing it here
949 * for consistency since we've removed the object from its
952 srcobject->type = OBJT_DEFAULT;
957 * SWAP_PAGER_HASPAGE() - determine if we have good backing store for
958 * the requested page.
960 * We determine whether good backing store exists for the requested
961 * page and return TRUE if it does, FALSE if it doesn't.
963 * If TRUE, we also try to determine how much valid, contiguous backing
964 * store exists before and after the requested page.
967 swap_pager_haspage(vm_object_t object, vm_pindex_t pindex, int *before,
973 VM_OBJECT_ASSERT_LOCKED(object);
976 * do we have good backing store at the requested index ?
978 blk0 = swp_pager_meta_ctl(object, pindex, 0);
979 if (blk0 == SWAPBLK_NONE) {
988 * find backwards-looking contiguous good backing store
990 if (before != NULL) {
991 for (i = 1; i < SWB_NPAGES; i++) {
994 blk = swp_pager_meta_ctl(object, pindex - i, 0);
1002 * find forward-looking contiguous good backing store
1004 if (after != NULL) {
1005 for (i = 1; i < SWB_NPAGES; i++) {
1006 blk = swp_pager_meta_ctl(object, pindex + i, 0);
1007 if (blk != blk0 + i)
1016 * SWAP_PAGER_PAGE_UNSWAPPED() - remove swap backing store related to page
1018 * This removes any associated swap backing store, whether valid or
1019 * not, from the page.
1021 * This routine is typically called when a page is made dirty, at
1022 * which point any associated swap can be freed. MADV_FREE also
1023 * calls us in a special-case situation
1025 * NOTE!!! If the page is clean and the swap was valid, the caller
1026 * should make the page dirty before calling this routine. This routine
1027 * does NOT change the m->dirty status of the page. Also: MADV_FREE
1030 * This routine may not sleep.
1032 * The object containing the page must be locked.
1035 swap_pager_unswapped(vm_page_t m)
1038 swp_pager_meta_ctl(m->object, m->pindex, SWM_FREE);
1042 * swap_pager_getpages() - bring pages in from swap
1044 * Attempt to page in the pages in array "m" of length "count". The caller
1045 * may optionally specify that additional pages preceding and succeeding
1046 * the specified range be paged in. The number of such pages is returned
1047 * in the "rbehind" and "rahead" parameters, and they will be in the
1048 * inactive queue upon return.
1050 * The pages in "m" must be busied and will remain busied upon return.
1053 swap_pager_getpages(vm_object_t object, vm_page_t *m, int count, int *rbehind,
1057 vm_page_t mpred, msucc, p;
1060 int i, j, maxahead, maxbehind, reqcount, shift;
1064 VM_OBJECT_WUNLOCK(object);
1065 bp = getpbuf(&nsw_rcount);
1066 VM_OBJECT_WLOCK(object);
1068 if (!swap_pager_haspage(object, m[0]->pindex, &maxbehind, &maxahead)) {
1069 relpbuf(bp, &nsw_rcount);
1070 return (VM_PAGER_FAIL);
1074 * Clip the readahead and readbehind ranges to exclude resident pages.
1076 if (rahead != NULL) {
1077 KASSERT(reqcount - 1 <= maxahead,
1078 ("page count %d extends beyond swap block", reqcount));
1079 *rahead = imin(*rahead, maxahead - (reqcount - 1));
1080 pindex = m[reqcount - 1]->pindex;
1081 msucc = TAILQ_NEXT(m[reqcount - 1], listq);
1082 if (msucc != NULL && msucc->pindex - pindex - 1 < *rahead)
1083 *rahead = msucc->pindex - pindex - 1;
1085 if (rbehind != NULL) {
1086 *rbehind = imin(*rbehind, maxbehind);
1087 pindex = m[0]->pindex;
1088 mpred = TAILQ_PREV(m[0], pglist, listq);
1089 if (mpred != NULL && pindex - mpred->pindex - 1 < *rbehind)
1090 *rbehind = pindex - mpred->pindex - 1;
1094 * Allocate readahead and readbehind pages.
1096 shift = rbehind != NULL ? *rbehind : 0;
1098 for (i = 1; i <= shift; i++) {
1099 p = vm_page_alloc(object, m[0]->pindex - i,
1102 /* Shift allocated pages to the left. */
1103 for (j = 0; j < i - 1; j++)
1105 bp->b_pages[j + shift - i + 1];
1108 bp->b_pages[shift - i] = p;
1113 for (i = 0; i < reqcount; i++)
1114 bp->b_pages[i + shift] = m[i];
1115 if (rahead != NULL) {
1116 for (i = 0; i < *rahead; i++) {
1117 p = vm_page_alloc(object,
1118 m[reqcount - 1]->pindex + i + 1, VM_ALLOC_NORMAL);
1121 bp->b_pages[shift + reqcount + i] = p;
1125 if (rbehind != NULL)
1130 vm_object_pip_add(object, count);
1132 for (i = 0; i < count; i++)
1133 bp->b_pages[i]->oflags |= VPO_SWAPINPROG;
1135 pindex = bp->b_pages[0]->pindex;
1136 blk = swp_pager_meta_ctl(object, pindex, 0);
1137 KASSERT(blk != SWAPBLK_NONE,
1138 ("no swap blocking containing %p(%jx)", object, (uintmax_t)pindex));
1140 VM_OBJECT_WUNLOCK(object);
1142 bp->b_flags |= B_PAGING;
1143 bp->b_iocmd = BIO_READ;
1144 bp->b_iodone = swp_pager_async_iodone;
1145 bp->b_rcred = crhold(thread0.td_ucred);
1146 bp->b_wcred = crhold(thread0.td_ucred);
1148 bp->b_bcount = PAGE_SIZE * count;
1149 bp->b_bufsize = PAGE_SIZE * count;
1150 bp->b_npages = count;
1151 bp->b_pgbefore = rbehind != NULL ? *rbehind : 0;
1152 bp->b_pgafter = rahead != NULL ? *rahead : 0;
1154 VM_CNT_INC(v_swapin);
1155 VM_CNT_ADD(v_swappgsin, count);
1158 * perform the I/O. NOTE!!! bp cannot be considered valid after
1159 * this point because we automatically release it on completion.
1160 * Instead, we look at the one page we are interested in which we
1161 * still hold a lock on even through the I/O completion.
1163 * The other pages in our m[] array are also released on completion,
1164 * so we cannot assume they are valid anymore either.
1166 * NOTE: b_blkno is destroyed by the call to swapdev_strategy
1169 swp_pager_strategy(bp);
1172 * Wait for the pages we want to complete. VPO_SWAPINPROG is always
1173 * cleared on completion. If an I/O error occurs, SWAPBLK_NONE
1174 * is set in the metadata for each page in the request.
1176 VM_OBJECT_WLOCK(object);
1177 while ((m[0]->oflags & VPO_SWAPINPROG) != 0) {
1178 m[0]->oflags |= VPO_SWAPSLEEP;
1179 VM_CNT_INC(v_intrans);
1180 if (VM_OBJECT_SLEEP(object, &object->paging_in_progress, PSWP,
1181 "swread", hz * 20)) {
1183 "swap_pager: indefinite wait buffer: bufobj: %p, blkno: %jd, size: %ld\n",
1184 bp->b_bufobj, (intmax_t)bp->b_blkno, bp->b_bcount);
1189 * If we had an unrecoverable read error pages will not be valid.
1191 for (i = 0; i < reqcount; i++)
1192 if (m[i]->valid != VM_PAGE_BITS_ALL)
1193 return (VM_PAGER_ERROR);
1195 return (VM_PAGER_OK);
1198 * A final note: in a low swap situation, we cannot deallocate swap
1199 * and mark a page dirty here because the caller is likely to mark
1200 * the page clean when we return, causing the page to possibly revert
1201 * to all-zero's later.
1206 * swap_pager_getpages_async():
1208 * Right now this is emulation of asynchronous operation on top of
1209 * swap_pager_getpages().
1212 swap_pager_getpages_async(vm_object_t object, vm_page_t *m, int count,
1213 int *rbehind, int *rahead, pgo_getpages_iodone_t iodone, void *arg)
1217 r = swap_pager_getpages(object, m, count, rbehind, rahead);
1218 VM_OBJECT_WUNLOCK(object);
1223 case VM_PAGER_ERROR:
1230 panic("unhandled swap_pager_getpages() error %d", r);
1232 (iodone)(arg, m, count, error);
1233 VM_OBJECT_WLOCK(object);
1239 * swap_pager_putpages:
1241 * Assign swap (if necessary) and initiate I/O on the specified pages.
1243 * We support both OBJT_DEFAULT and OBJT_SWAP objects. DEFAULT objects
1244 * are automatically converted to SWAP objects.
1246 * In a low memory situation we may block in VOP_STRATEGY(), but the new
1247 * vm_page reservation system coupled with properly written VFS devices
1248 * should ensure that no low-memory deadlock occurs. This is an area
1251 * The parent has N vm_object_pip_add() references prior to
1252 * calling us and will remove references for rtvals[] that are
1253 * not set to VM_PAGER_PEND. We need to remove the rest on I/O
1256 * The parent has soft-busy'd the pages it passes us and will unbusy
1257 * those whos rtvals[] entry is not set to VM_PAGER_PEND on return.
1258 * We need to unbusy the rest on I/O completion.
1261 swap_pager_putpages(vm_object_t object, vm_page_t *m, int count,
1262 int flags, int *rtvals)
1267 if (count && m[0]->object != object) {
1268 panic("swap_pager_putpages: object mismatch %p/%p",
1277 * Turn object into OBJT_SWAP
1278 * check for bogus sysops
1279 * force sync if not pageout process
1281 if (object->type != OBJT_SWAP)
1282 swp_pager_meta_build(object, 0, SWAPBLK_NONE);
1283 VM_OBJECT_WUNLOCK(object);
1286 if (curproc != pageproc)
1289 sync = (flags & VM_PAGER_PUT_SYNC) != 0;
1294 * Assign swap blocks and issue I/O. We reallocate swap on the fly.
1295 * The page is left dirty until the pageout operation completes
1298 for (i = 0; i < count; i += n) {
1304 * Maximum I/O size is limited by a number of factors.
1306 n = min(BLIST_MAX_ALLOC, count - i);
1307 n = min(n, nsw_cluster_max);
1310 * Get biggest block of swap we can. If we fail, fall
1311 * back and try to allocate a smaller block. Don't go
1312 * overboard trying to allocate space if it would overly
1316 (blk = swp_pager_getswapspace(n)) == SWAPBLK_NONE &&
1321 if (blk == SWAPBLK_NONE) {
1322 for (j = 0; j < n; ++j)
1323 rtvals[i+j] = VM_PAGER_FAIL;
1328 * All I/O parameters have been satisfied, build the I/O
1329 * request and assign the swap space.
1332 bp = getpbuf(&nsw_wcount_sync);
1334 bp = getpbuf(&nsw_wcount_async);
1335 bp->b_flags = B_ASYNC;
1337 bp->b_flags |= B_PAGING;
1338 bp->b_iocmd = BIO_WRITE;
1340 bp->b_rcred = crhold(thread0.td_ucred);
1341 bp->b_wcred = crhold(thread0.td_ucred);
1342 bp->b_bcount = PAGE_SIZE * n;
1343 bp->b_bufsize = PAGE_SIZE * n;
1346 VM_OBJECT_WLOCK(object);
1347 for (j = 0; j < n; ++j) {
1348 vm_page_t mreq = m[i+j];
1350 swp_pager_meta_build(
1355 MPASS(mreq->dirty == VM_PAGE_BITS_ALL);
1356 mreq->oflags |= VPO_SWAPINPROG;
1357 bp->b_pages[j] = mreq;
1359 VM_OBJECT_WUNLOCK(object);
1362 * Must set dirty range for NFS to work.
1365 bp->b_dirtyend = bp->b_bcount;
1367 VM_CNT_INC(v_swapout);
1368 VM_CNT_ADD(v_swappgsout, bp->b_npages);
1371 * We unconditionally set rtvals[] to VM_PAGER_PEND so that we
1372 * can call the async completion routine at the end of a
1373 * synchronous I/O operation. Otherwise, our caller would
1374 * perform duplicate unbusy and wakeup operations on the page
1375 * and object, respectively.
1377 for (j = 0; j < n; j++)
1378 rtvals[i + j] = VM_PAGER_PEND;
1383 * NOTE: b_blkno is destroyed by the call to swapdev_strategy
1385 if (sync == FALSE) {
1386 bp->b_iodone = swp_pager_async_iodone;
1388 swp_pager_strategy(bp);
1395 * NOTE: b_blkno is destroyed by the call to swapdev_strategy
1397 bp->b_iodone = bdone;
1398 swp_pager_strategy(bp);
1401 * Wait for the sync I/O to complete.
1403 bwait(bp, PVM, "swwrt");
1406 * Now that we are through with the bp, we can call the
1407 * normal async completion, which frees everything up.
1409 swp_pager_async_iodone(bp);
1411 VM_OBJECT_WLOCK(object);
1415 * swp_pager_async_iodone:
1417 * Completion routine for asynchronous reads and writes from/to swap.
1418 * Also called manually by synchronous code to finish up a bp.
1420 * This routine may not sleep.
1423 swp_pager_async_iodone(struct buf *bp)
1426 vm_object_t object = NULL;
1431 if (bp->b_ioflags & BIO_ERROR) {
1433 "swap_pager: I/O error - %s failed; blkno %ld,"
1434 "size %ld, error %d\n",
1435 ((bp->b_iocmd == BIO_READ) ? "pagein" : "pageout"),
1443 * remove the mapping for kernel virtual
1446 pmap_qremove((vm_offset_t)bp->b_data, bp->b_npages);
1448 bp->b_data = bp->b_kvabase;
1451 object = bp->b_pages[0]->object;
1452 VM_OBJECT_WLOCK(object);
1456 * cleanup pages. If an error occurs writing to swap, we are in
1457 * very serious trouble. If it happens to be a disk error, though,
1458 * we may be able to recover by reassigning the swap later on. So
1459 * in this case we remove the m->swapblk assignment for the page
1460 * but do not free it in the rlist. The errornous block(s) are thus
1461 * never reallocated as swap. Redirty the page and continue.
1463 for (i = 0; i < bp->b_npages; ++i) {
1464 vm_page_t m = bp->b_pages[i];
1466 m->oflags &= ~VPO_SWAPINPROG;
1467 if (m->oflags & VPO_SWAPSLEEP) {
1468 m->oflags &= ~VPO_SWAPSLEEP;
1469 wakeup(&object->paging_in_progress);
1472 if (bp->b_ioflags & BIO_ERROR) {
1474 * If an error occurs I'd love to throw the swapblk
1475 * away without freeing it back to swapspace, so it
1476 * can never be used again. But I can't from an
1479 if (bp->b_iocmd == BIO_READ) {
1481 * NOTE: for reads, m->dirty will probably
1482 * be overridden by the original caller of
1483 * getpages so don't play cute tricks here.
1488 * If a write error occurs, reactivate page
1489 * so it doesn't clog the inactive list,
1490 * then finish the I/O.
1494 vm_page_activate(m);
1498 } else if (bp->b_iocmd == BIO_READ) {
1500 * NOTE: for reads, m->dirty will probably be
1501 * overridden by the original caller of getpages so
1502 * we cannot set them in order to free the underlying
1503 * swap in a low-swap situation. I don't think we'd
1504 * want to do that anyway, but it was an optimization
1505 * that existed in the old swapper for a time before
1506 * it got ripped out due to precisely this problem.
1508 KASSERT(!pmap_page_is_mapped(m),
1509 ("swp_pager_async_iodone: page %p is mapped", m));
1510 KASSERT(m->dirty == 0,
1511 ("swp_pager_async_iodone: page %p is dirty", m));
1513 m->valid = VM_PAGE_BITS_ALL;
1514 if (i < bp->b_pgbefore ||
1515 i >= bp->b_npages - bp->b_pgafter)
1516 vm_page_readahead_finish(m);
1519 * For write success, clear the dirty
1520 * status, then finish the I/O ( which decrements the
1521 * busy count and possibly wakes waiter's up ).
1522 * A page is only written to swap after a period of
1523 * inactivity. Therefore, we do not expect it to be
1526 KASSERT(!pmap_page_is_write_mapped(m),
1527 ("swp_pager_async_iodone: page %p is not write"
1531 vm_page_deactivate_noreuse(m);
1538 * adjust pip. NOTE: the original parent may still have its own
1539 * pip refs on the object.
1541 if (object != NULL) {
1542 vm_object_pip_wakeupn(object, bp->b_npages);
1543 VM_OBJECT_WUNLOCK(object);
1547 * swapdev_strategy() manually sets b_vp and b_bufobj before calling
1548 * bstrategy(). Set them back to NULL now we're done with it, or we'll
1549 * trigger a KASSERT in relpbuf().
1553 bp->b_bufobj = NULL;
1556 * release the physical I/O buffer
1560 ((bp->b_iocmd == BIO_READ) ? &nsw_rcount :
1561 ((bp->b_flags & B_ASYNC) ?
1570 swap_pager_nswapdev(void)
1577 * SWP_PAGER_FORCE_PAGEIN() - force a swap block to be paged in
1579 * This routine dissociates the page at the given index within an object
1580 * from its backing store, paging it in if it does not reside in memory.
1581 * If the page is paged in, it is marked dirty and placed in the laundry
1582 * queue. The page is marked dirty because it no longer has backing
1583 * store. It is placed in the laundry queue because it has not been
1584 * accessed recently. Otherwise, it would already reside in memory.
1586 * We also attempt to swap in all other pages in the swap block.
1587 * However, we only guarantee that the one at the specified index is
1590 * XXX - The code to page the whole block in doesn't work, so we
1591 * revert to the one-by-one behavior for now. Sigh.
1594 swp_pager_force_pagein(vm_object_t object, vm_pindex_t pindex)
1598 vm_object_pip_add(object, 1);
1599 m = vm_page_grab(object, pindex, VM_ALLOC_NORMAL);
1600 if (m->valid == VM_PAGE_BITS_ALL) {
1601 vm_object_pip_wakeup(object);
1604 vm_page_activate(m);
1607 vm_pager_page_unswapped(m);
1611 if (swap_pager_getpages(object, &m, 1, NULL, NULL) != VM_PAGER_OK)
1612 panic("swap_pager_force_pagein: read from swap failed");/*XXX*/
1613 vm_object_pip_wakeup(object);
1619 vm_pager_page_unswapped(m);
1623 * swap_pager_swapoff:
1625 * Page in all of the pages that have been paged out to the
1626 * given device. The corresponding blocks in the bitmap must be
1627 * marked as allocated and the device must be flagged SW_CLOSING.
1628 * There may be no processes swapped out to the device.
1630 * This routine may block.
1633 swap_pager_swapoff(struct swdevt *sp)
1640 sx_assert(&swdev_syscall_lock, SA_XLOCKED);
1644 mtx_lock(&vm_object_list_mtx);
1645 TAILQ_FOREACH(object, &vm_object_list, object_list) {
1646 if (object->type != OBJT_SWAP)
1648 mtx_unlock(&vm_object_list_mtx);
1649 /* Depends on type-stability. */
1650 VM_OBJECT_WLOCK(object);
1653 * Dead objects are eventually terminated on their own.
1655 if ((object->flags & OBJ_DEAD) != 0)
1659 * Sync with fences placed after pctrie
1660 * initialization. We must not access pctrie below
1661 * unless we checked that our object is swap and not
1664 atomic_thread_fence_acq();
1665 if (object->type != OBJT_SWAP)
1668 for (pi = 0; (sb = SWAP_PCTRIE_LOOKUP_GE(
1669 &object->un_pager.swp.swp_blks, pi)) != NULL; ) {
1670 pi = sb->p + SWAP_META_PAGES;
1671 for (i = 0; i < SWAP_META_PAGES; i++) {
1672 if (sb->d[i] == SWAPBLK_NONE)
1674 if (swp_pager_isondev(sb->d[i], sp))
1675 swp_pager_force_pagein(object,
1680 VM_OBJECT_WUNLOCK(object);
1681 mtx_lock(&vm_object_list_mtx);
1683 mtx_unlock(&vm_object_list_mtx);
1687 * Objects may be locked or paging to the device being
1688 * removed, so we will miss their pages and need to
1689 * make another pass. We have marked this device as
1690 * SW_CLOSING, so the activity should finish soon.
1693 if (retries > 100) {
1694 panic("swapoff: failed to locate %d swap blocks",
1697 pause("swpoff", hz / 20);
1700 EVENTHANDLER_INVOKE(swapoff, sp);
1703 /************************************************************************
1705 ************************************************************************
1707 * These routines manipulate the swap metadata stored in the
1710 * Swap metadata is implemented with a global hash and not directly
1711 * linked into the object. Instead the object simply contains
1712 * appropriate tracking counters.
1716 * SWP_PAGER_META_BUILD() - add swap block to swap meta data for object
1718 * We first convert the object to a swap object if it is a default
1721 * The specified swapblk is added to the object's swap metadata. If
1722 * the swapblk is not valid, it is freed instead. Any previously
1723 * assigned swapblk is freed.
1726 swp_pager_meta_build(vm_object_t object, vm_pindex_t pindex, daddr_t swapblk)
1728 static volatile int swblk_zone_exhausted, swpctrie_zone_exhausted;
1730 vm_pindex_t modpi, rdpi;
1733 VM_OBJECT_ASSERT_WLOCKED(object);
1736 * Convert default object to swap object if necessary
1738 if (object->type != OBJT_SWAP) {
1739 pctrie_init(&object->un_pager.swp.swp_blks);
1742 * Ensure that swap_pager_swapoff()'s iteration over
1743 * object_list does not see a garbage pctrie.
1745 atomic_thread_fence_rel();
1747 object->type = OBJT_SWAP;
1748 KASSERT(object->handle == NULL, ("default pager with handle"));
1751 rdpi = rounddown(pindex, SWAP_META_PAGES);
1752 sb = SWAP_PCTRIE_LOOKUP(&object->un_pager.swp.swp_blks, rdpi);
1754 if (swapblk == SWAPBLK_NONE)
1757 sb = uma_zalloc(swblk_zone, M_NOWAIT | (curproc ==
1758 pageproc ? M_USE_RESERVE : 0));
1761 for (i = 0; i < SWAP_META_PAGES; i++)
1762 sb->d[i] = SWAPBLK_NONE;
1763 if (atomic_cmpset_int(&swblk_zone_exhausted,
1765 printf("swblk zone ok\n");
1768 VM_OBJECT_WUNLOCK(object);
1769 if (uma_zone_exhausted(swblk_zone)) {
1770 if (atomic_cmpset_int(&swblk_zone_exhausted,
1772 printf("swap blk zone exhausted, "
1773 "increase kern.maxswzone\n");
1774 vm_pageout_oom(VM_OOM_SWAPZ);
1775 pause("swzonxb", 10);
1778 VM_OBJECT_WLOCK(object);
1781 error = SWAP_PCTRIE_INSERT(
1782 &object->un_pager.swp.swp_blks, sb);
1784 if (atomic_cmpset_int(&swpctrie_zone_exhausted,
1786 printf("swpctrie zone ok\n");
1789 VM_OBJECT_WUNLOCK(object);
1790 if (uma_zone_exhausted(swpctrie_zone)) {
1791 if (atomic_cmpset_int(&swpctrie_zone_exhausted,
1793 printf("swap pctrie zone exhausted, "
1794 "increase kern.maxswzone\n");
1795 vm_pageout_oom(VM_OOM_SWAPZ);
1796 pause("swzonxp", 10);
1799 VM_OBJECT_WLOCK(object);
1802 MPASS(sb->p == rdpi);
1804 modpi = pindex % SWAP_META_PAGES;
1805 /* Delete prior contents of metadata. */
1806 if (sb->d[modpi] != SWAPBLK_NONE)
1807 swp_pager_freeswapspace(sb->d[modpi], 1);
1808 /* Enter block into metadata. */
1809 sb->d[modpi] = swapblk;
1813 * SWP_PAGER_META_FREE() - free a range of blocks in the object's swap metadata
1815 * The requested range of blocks is freed, with any associated swap
1816 * returned to the swap bitmap.
1818 * This routine will free swap metadata structures as they are cleaned
1819 * out. This routine does *NOT* operate on swap metadata associated
1820 * with resident pages.
1823 swp_pager_meta_free(vm_object_t object, vm_pindex_t pindex, vm_pindex_t count)
1830 VM_OBJECT_ASSERT_WLOCKED(object);
1831 if (object->type != OBJT_SWAP || count == 0)
1834 last = pindex + count - 1;
1836 sb = SWAP_PCTRIE_LOOKUP_GE(&object->un_pager.swp.swp_blks,
1837 rounddown(pindex, SWAP_META_PAGES));
1838 if (sb == NULL || sb->p > last)
1841 for (i = 0; i < SWAP_META_PAGES; i++) {
1842 if (sb->d[i] == SWAPBLK_NONE)
1844 if (pindex <= sb->p + i && sb->p + i <= last) {
1845 swp_pager_freeswapspace(sb->d[i], 1);
1846 sb->d[i] = SWAPBLK_NONE;
1850 pindex = sb->p + SWAP_META_PAGES;
1852 SWAP_PCTRIE_REMOVE(&object->un_pager.swp.swp_blks,
1854 uma_zfree(swblk_zone, sb);
1860 * SWP_PAGER_META_FREE_ALL() - destroy all swap metadata associated with object
1862 * This routine locates and destroys all swap metadata associated with
1866 swp_pager_meta_free_all(vm_object_t object)
1872 VM_OBJECT_ASSERT_WLOCKED(object);
1873 if (object->type != OBJT_SWAP)
1876 for (pindex = 0; (sb = SWAP_PCTRIE_LOOKUP_GE(
1877 &object->un_pager.swp.swp_blks, pindex)) != NULL;) {
1878 pindex = sb->p + SWAP_META_PAGES;
1879 for (i = 0; i < SWAP_META_PAGES; i++) {
1880 if (sb->d[i] != SWAPBLK_NONE)
1881 swp_pager_freeswapspace(sb->d[i], 1);
1883 SWAP_PCTRIE_REMOVE(&object->un_pager.swp.swp_blks, sb->p);
1884 uma_zfree(swblk_zone, sb);
1889 * SWP_PAGER_METACTL() - misc control of swap and vm_page_t meta data.
1891 * This routine is capable of looking up, popping, or freeing
1892 * swapblk assignments in the swap meta data or in the vm_page_t.
1893 * The routine typically returns the swapblk being looked-up, or popped,
1894 * or SWAPBLK_NONE if the block was freed, or SWAPBLK_NONE if the block
1895 * was invalid. This routine will automatically free any invalid
1896 * meta-data swapblks.
1898 * When acting on a busy resident page and paging is in progress, we
1899 * have to wait until paging is complete but otherwise can act on the
1902 * SWM_FREE remove and free swap block from metadata
1903 * SWM_POP remove from meta data but do not free.. pop it out
1906 swp_pager_meta_ctl(vm_object_t object, vm_pindex_t pindex, int flags)
1912 if ((flags & (SWM_FREE | SWM_POP)) != 0)
1913 VM_OBJECT_ASSERT_WLOCKED(object);
1915 VM_OBJECT_ASSERT_LOCKED(object);
1918 * The meta data only exists if the object is OBJT_SWAP
1919 * and even then might not be allocated yet.
1921 if (object->type != OBJT_SWAP)
1922 return (SWAPBLK_NONE);
1924 sb = SWAP_PCTRIE_LOOKUP(&object->un_pager.swp.swp_blks,
1925 rounddown(pindex, SWAP_META_PAGES));
1927 return (SWAPBLK_NONE);
1928 r1 = sb->d[pindex % SWAP_META_PAGES];
1929 if (r1 == SWAPBLK_NONE)
1930 return (SWAPBLK_NONE);
1931 if ((flags & (SWM_FREE | SWM_POP)) != 0) {
1932 sb->d[pindex % SWAP_META_PAGES] = SWAPBLK_NONE;
1933 for (i = 0; i < SWAP_META_PAGES; i++) {
1934 if (sb->d[i] != SWAPBLK_NONE)
1937 if (i == SWAP_META_PAGES) {
1938 SWAP_PCTRIE_REMOVE(&object->un_pager.swp.swp_blks,
1939 rounddown(pindex, SWAP_META_PAGES));
1940 uma_zfree(swblk_zone, sb);
1943 if ((flags & SWM_FREE) != 0) {
1944 swp_pager_freeswapspace(r1, 1);
1951 * Returns the least page index which is greater than or equal to the
1952 * parameter pindex and for which there is a swap block allocated.
1953 * Returns object's size if the object's type is not swap or if there
1954 * are no allocated swap blocks for the object after the requested
1958 swap_pager_find_least(vm_object_t object, vm_pindex_t pindex)
1963 VM_OBJECT_ASSERT_LOCKED(object);
1964 if (object->type != OBJT_SWAP)
1965 return (object->size);
1967 sb = SWAP_PCTRIE_LOOKUP_GE(&object->un_pager.swp.swp_blks,
1968 rounddown(pindex, SWAP_META_PAGES));
1970 return (object->size);
1971 if (sb->p < pindex) {
1972 for (i = pindex % SWAP_META_PAGES; i < SWAP_META_PAGES; i++) {
1973 if (sb->d[i] != SWAPBLK_NONE)
1976 sb = SWAP_PCTRIE_LOOKUP_GE(&object->un_pager.swp.swp_blks,
1977 roundup(pindex, SWAP_META_PAGES));
1979 return (object->size);
1981 for (i = 0; i < SWAP_META_PAGES; i++) {
1982 if (sb->d[i] != SWAPBLK_NONE)
1987 * We get here if a swblk is present in the trie but it
1988 * doesn't map any blocks.
1991 return (object->size);
1995 * System call swapon(name) enables swapping on device name,
1996 * which must be in the swdevsw. Return EBUSY
1997 * if already swapping on this device.
1999 #ifndef _SYS_SYSPROTO_H_
2000 struct swapon_args {
2010 sys_swapon(struct thread *td, struct swapon_args *uap)
2014 struct nameidata nd;
2017 error = priv_check(td, PRIV_SWAPON);
2021 sx_xlock(&swdev_syscall_lock);
2024 * Swap metadata may not fit in the KVM if we have physical
2027 if (swblk_zone == NULL) {
2032 NDINIT(&nd, LOOKUP, ISOPEN | FOLLOW | AUDITVNODE1, UIO_USERSPACE,
2038 NDFREE(&nd, NDF_ONLY_PNBUF);
2041 if (vn_isdisk(vp, &error)) {
2042 error = swapongeom(vp);
2043 } else if (vp->v_type == VREG &&
2044 (vp->v_mount->mnt_vfc->vfc_flags & VFCF_NETWORK) != 0 &&
2045 (error = VOP_GETATTR(vp, &attr, td->td_ucred)) == 0) {
2047 * Allow direct swapping to NFS regular files in the same
2048 * way that nfs_mountroot() sets up diskless swapping.
2050 error = swaponvp(td, vp, attr.va_size / DEV_BSIZE);
2056 sx_xunlock(&swdev_syscall_lock);
2061 * Check that the total amount of swap currently configured does not
2062 * exceed half the theoretical maximum. If it does, print a warning
2063 * message and return -1; otherwise, return 0.
2066 swapon_check_swzone(unsigned long npages)
2068 unsigned long maxpages;
2070 /* absolute maximum we can handle assuming 100% efficiency */
2071 maxpages = uma_zone_get_max(swblk_zone) * SWAP_META_PAGES;
2073 /* recommend using no more than half that amount */
2074 if (npages > maxpages / 2) {
2075 printf("warning: total configured swap (%lu pages) "
2076 "exceeds maximum recommended amount (%lu pages).\n",
2077 npages, maxpages / 2);
2078 printf("warning: increase kern.maxswzone "
2079 "or reduce amount of swap.\n");
2086 swaponsomething(struct vnode *vp, void *id, u_long nblks,
2087 sw_strategy_t *strategy, sw_close_t *close, dev_t dev, int flags)
2089 struct swdevt *sp, *tsp;
2094 * nblks is in DEV_BSIZE'd chunks, convert to PAGE_SIZE'd chunks.
2095 * First chop nblks off to page-align it, then convert.
2097 * sw->sw_nblks is in page-sized chunks now too.
2099 nblks &= ~(ctodb(1) - 1);
2100 nblks = dbtoc(nblks);
2103 * If we go beyond this, we get overflows in the radix
2106 mblocks = 0x40000000 / BLIST_META_RADIX;
2107 if (nblks > mblocks) {
2109 "WARNING: reducing swap size to maximum of %luMB per unit\n",
2110 mblocks / 1024 / 1024 * PAGE_SIZE);
2114 sp = malloc(sizeof *sp, M_VMPGDATA, M_WAITOK | M_ZERO);
2119 sp->sw_nblks = nblks;
2121 sp->sw_strategy = strategy;
2122 sp->sw_close = close;
2123 sp->sw_flags = flags;
2125 sp->sw_blist = blist_create(nblks, M_WAITOK);
2127 * Do not free the first two block in order to avoid overwriting
2128 * any bsd label at the front of the partition
2130 blist_free(sp->sw_blist, 2, nblks - 2);
2133 mtx_lock(&sw_dev_mtx);
2134 TAILQ_FOREACH(tsp, &swtailq, sw_list) {
2135 if (tsp->sw_end >= dvbase) {
2137 * We put one uncovered page between the devices
2138 * in order to definitively prevent any cross-device
2141 dvbase = tsp->sw_end + 1;
2144 sp->sw_first = dvbase;
2145 sp->sw_end = dvbase + nblks;
2146 TAILQ_INSERT_TAIL(&swtailq, sp, sw_list);
2148 swap_pager_avail += nblks - 2;
2149 swap_total += (vm_ooffset_t)nblks * PAGE_SIZE;
2150 swapon_check_swzone(swap_total / PAGE_SIZE);
2152 mtx_unlock(&sw_dev_mtx);
2153 EVENTHANDLER_INVOKE(swapon, sp);
2157 * SYSCALL: swapoff(devname)
2159 * Disable swapping on the given device.
2161 * XXX: Badly designed system call: it should use a device index
2162 * rather than filename as specification. We keep sw_vp around
2163 * only to make this work.
2165 #ifndef _SYS_SYSPROTO_H_
2166 struct swapoff_args {
2176 sys_swapoff(struct thread *td, struct swapoff_args *uap)
2179 struct nameidata nd;
2183 error = priv_check(td, PRIV_SWAPOFF);
2187 sx_xlock(&swdev_syscall_lock);
2189 NDINIT(&nd, LOOKUP, FOLLOW | AUDITVNODE1, UIO_USERSPACE, uap->name,
2194 NDFREE(&nd, NDF_ONLY_PNBUF);
2197 mtx_lock(&sw_dev_mtx);
2198 TAILQ_FOREACH(sp, &swtailq, sw_list) {
2199 if (sp->sw_vp == vp)
2202 mtx_unlock(&sw_dev_mtx);
2207 error = swapoff_one(sp, td->td_ucred);
2209 sx_xunlock(&swdev_syscall_lock);
2214 swapoff_one(struct swdevt *sp, struct ucred *cred)
2221 sx_assert(&swdev_syscall_lock, SA_XLOCKED);
2223 (void) vn_lock(sp->sw_vp, LK_EXCLUSIVE | LK_RETRY);
2224 error = mac_system_check_swapoff(cred, sp->sw_vp);
2225 (void) VOP_UNLOCK(sp->sw_vp, 0);
2229 nblks = sp->sw_nblks;
2232 * We can turn off this swap device safely only if the
2233 * available virtual memory in the system will fit the amount
2234 * of data we will have to page back in, plus an epsilon so
2235 * the system doesn't become critically low on swap space.
2237 if (vm_cnt.v_free_count + swap_pager_avail < nblks + nswap_lowat)
2241 * Prevent further allocations on this device.
2243 mtx_lock(&sw_dev_mtx);
2244 sp->sw_flags |= SW_CLOSING;
2245 swap_pager_avail -= blist_fill(sp->sw_blist, 0, nblks);
2246 swap_total -= (vm_ooffset_t)nblks * PAGE_SIZE;
2247 mtx_unlock(&sw_dev_mtx);
2250 * Page in the contents of the device and close it.
2252 swap_pager_swapoff(sp);
2254 sp->sw_close(curthread, sp);
2255 mtx_lock(&sw_dev_mtx);
2257 TAILQ_REMOVE(&swtailq, sp, sw_list);
2259 if (nswapdev == 0) {
2260 swap_pager_full = 2;
2261 swap_pager_almost_full = 1;
2265 mtx_unlock(&sw_dev_mtx);
2266 blist_destroy(sp->sw_blist);
2267 free(sp, M_VMPGDATA);
2274 struct swdevt *sp, *spt;
2275 const char *devname;
2278 sx_xlock(&swdev_syscall_lock);
2280 mtx_lock(&sw_dev_mtx);
2281 TAILQ_FOREACH_SAFE(sp, &swtailq, sw_list, spt) {
2282 mtx_unlock(&sw_dev_mtx);
2283 if (vn_isdisk(sp->sw_vp, NULL))
2284 devname = devtoname(sp->sw_vp->v_rdev);
2287 error = swapoff_one(sp, thread0.td_ucred);
2289 printf("Cannot remove swap device %s (error=%d), "
2290 "skipping.\n", devname, error);
2291 } else if (bootverbose) {
2292 printf("Swap device %s removed.\n", devname);
2294 mtx_lock(&sw_dev_mtx);
2296 mtx_unlock(&sw_dev_mtx);
2298 sx_xunlock(&swdev_syscall_lock);
2302 swap_pager_status(int *total, int *used)
2308 mtx_lock(&sw_dev_mtx);
2309 TAILQ_FOREACH(sp, &swtailq, sw_list) {
2310 *total += sp->sw_nblks;
2311 *used += sp->sw_used;
2313 mtx_unlock(&sw_dev_mtx);
2317 swap_dev_info(int name, struct xswdev *xs, char *devname, size_t len)
2320 const char *tmp_devname;
2325 mtx_lock(&sw_dev_mtx);
2326 TAILQ_FOREACH(sp, &swtailq, sw_list) {
2331 xs->xsw_version = XSWDEV_VERSION;
2332 xs->xsw_dev = sp->sw_dev;
2333 xs->xsw_flags = sp->sw_flags;
2334 xs->xsw_nblks = sp->sw_nblks;
2335 xs->xsw_used = sp->sw_used;
2336 if (devname != NULL) {
2337 if (vn_isdisk(sp->sw_vp, NULL))
2338 tmp_devname = devtoname(sp->sw_vp->v_rdev);
2340 tmp_devname = "[file]";
2341 strncpy(devname, tmp_devname, len);
2346 mtx_unlock(&sw_dev_mtx);
2350 #if defined(COMPAT_FREEBSD11)
2351 #define XSWDEV_VERSION_11 1
2362 sysctl_vm_swap_info(SYSCTL_HANDLER_ARGS)
2365 #if defined(COMPAT_FREEBSD11)
2366 struct xswdev11 xs11;
2370 if (arg2 != 1) /* name length */
2372 error = swap_dev_info(*(int *)arg1, &xs, NULL, 0);
2375 #if defined(COMPAT_FREEBSD11)
2376 if (req->oldlen == sizeof(xs11)) {
2377 xs11.xsw_version = XSWDEV_VERSION_11;
2378 xs11.xsw_dev = xs.xsw_dev; /* truncation */
2379 xs11.xsw_flags = xs.xsw_flags;
2380 xs11.xsw_nblks = xs.xsw_nblks;
2381 xs11.xsw_used = xs.xsw_used;
2382 error = SYSCTL_OUT(req, &xs11, sizeof(xs11));
2385 error = SYSCTL_OUT(req, &xs, sizeof(xs));
2389 SYSCTL_INT(_vm, OID_AUTO, nswapdev, CTLFLAG_RD, &nswapdev, 0,
2390 "Number of swap devices");
2391 SYSCTL_NODE(_vm, OID_AUTO, swap_info, CTLFLAG_RD | CTLFLAG_MPSAFE,
2392 sysctl_vm_swap_info,
2393 "Swap statistics by device");
2396 * Count the approximate swap usage in pages for a vmspace. The
2397 * shadowed or not yet copied on write swap blocks are not accounted.
2398 * The map must be locked.
2401 vmspace_swap_count(struct vmspace *vmspace)
2411 map = &vmspace->vm_map;
2414 for (cur = map->header.next; cur != &map->header; cur = cur->next) {
2415 if ((cur->eflags & MAP_ENTRY_IS_SUB_MAP) != 0)
2417 object = cur->object.vm_object;
2418 if (object == NULL || object->type != OBJT_SWAP)
2420 VM_OBJECT_RLOCK(object);
2421 if (object->type != OBJT_SWAP)
2423 pi = OFF_TO_IDX(cur->offset);
2424 e = pi + OFF_TO_IDX(cur->end - cur->start);
2425 for (;; pi = sb->p + SWAP_META_PAGES) {
2426 sb = SWAP_PCTRIE_LOOKUP_GE(
2427 &object->un_pager.swp.swp_blks, pi);
2428 if (sb == NULL || sb->p >= e)
2430 for (i = 0; i < SWAP_META_PAGES; i++) {
2431 if (sb->p + i < e &&
2432 sb->d[i] != SWAPBLK_NONE)
2437 VM_OBJECT_RUNLOCK(object);
2445 * Swapping onto disk devices.
2449 static g_orphan_t swapgeom_orphan;
2451 static struct g_class g_swap_class = {
2453 .version = G_VERSION,
2454 .orphan = swapgeom_orphan,
2457 DECLARE_GEOM_CLASS(g_swap_class, g_class);
2461 swapgeom_close_ev(void *arg, int flags)
2463 struct g_consumer *cp;
2466 g_access(cp, -1, -1, 0);
2468 g_destroy_consumer(cp);
2472 * Add a reference to the g_consumer for an inflight transaction.
2475 swapgeom_acquire(struct g_consumer *cp)
2478 mtx_assert(&sw_dev_mtx, MA_OWNED);
2483 * Remove a reference from the g_consumer. Post a close event if all
2484 * references go away, since the function might be called from the
2488 swapgeom_release(struct g_consumer *cp, struct swdevt *sp)
2491 mtx_assert(&sw_dev_mtx, MA_OWNED);
2493 if (cp->index == 0) {
2494 if (g_post_event(swapgeom_close_ev, cp, M_NOWAIT, NULL) == 0)
2500 swapgeom_done(struct bio *bp2)
2504 struct g_consumer *cp;
2506 bp = bp2->bio_caller2;
2508 bp->b_ioflags = bp2->bio_flags;
2510 bp->b_ioflags |= BIO_ERROR;
2511 bp->b_resid = bp->b_bcount - bp2->bio_completed;
2512 bp->b_error = bp2->bio_error;
2514 sp = bp2->bio_caller1;
2515 mtx_lock(&sw_dev_mtx);
2516 swapgeom_release(cp, sp);
2517 mtx_unlock(&sw_dev_mtx);
2522 swapgeom_strategy(struct buf *bp, struct swdevt *sp)
2525 struct g_consumer *cp;
2527 mtx_lock(&sw_dev_mtx);
2530 mtx_unlock(&sw_dev_mtx);
2531 bp->b_error = ENXIO;
2532 bp->b_ioflags |= BIO_ERROR;
2536 swapgeom_acquire(cp);
2537 mtx_unlock(&sw_dev_mtx);
2538 if (bp->b_iocmd == BIO_WRITE)
2541 bio = g_alloc_bio();
2543 mtx_lock(&sw_dev_mtx);
2544 swapgeom_release(cp, sp);
2545 mtx_unlock(&sw_dev_mtx);
2546 bp->b_error = ENOMEM;
2547 bp->b_ioflags |= BIO_ERROR;
2552 bio->bio_caller1 = sp;
2553 bio->bio_caller2 = bp;
2554 bio->bio_cmd = bp->b_iocmd;
2555 bio->bio_offset = (bp->b_blkno - sp->sw_first) * PAGE_SIZE;
2556 bio->bio_length = bp->b_bcount;
2557 bio->bio_done = swapgeom_done;
2558 if (!buf_mapped(bp)) {
2559 bio->bio_ma = bp->b_pages;
2560 bio->bio_data = unmapped_buf;
2561 bio->bio_ma_offset = (vm_offset_t)bp->b_offset & PAGE_MASK;
2562 bio->bio_ma_n = bp->b_npages;
2563 bio->bio_flags |= BIO_UNMAPPED;
2565 bio->bio_data = bp->b_data;
2568 g_io_request(bio, cp);
2573 swapgeom_orphan(struct g_consumer *cp)
2578 mtx_lock(&sw_dev_mtx);
2579 TAILQ_FOREACH(sp, &swtailq, sw_list) {
2580 if (sp->sw_id == cp) {
2581 sp->sw_flags |= SW_CLOSING;
2586 * Drop reference we were created with. Do directly since we're in a
2587 * special context where we don't have to queue the call to
2588 * swapgeom_close_ev().
2591 destroy = ((sp != NULL) && (cp->index == 0));
2594 mtx_unlock(&sw_dev_mtx);
2596 swapgeom_close_ev(cp, 0);
2600 swapgeom_close(struct thread *td, struct swdevt *sw)
2602 struct g_consumer *cp;
2604 mtx_lock(&sw_dev_mtx);
2607 mtx_unlock(&sw_dev_mtx);
2610 * swapgeom_close() may be called from the biodone context,
2611 * where we cannot perform topology changes. Delegate the
2612 * work to the events thread.
2615 g_waitfor_event(swapgeom_close_ev, cp, M_WAITOK, NULL);
2619 swapongeom_locked(struct cdev *dev, struct vnode *vp)
2621 struct g_provider *pp;
2622 struct g_consumer *cp;
2623 static struct g_geom *gp;
2628 pp = g_dev_getprovider(dev);
2631 mtx_lock(&sw_dev_mtx);
2632 TAILQ_FOREACH(sp, &swtailq, sw_list) {
2634 if (cp != NULL && cp->provider == pp) {
2635 mtx_unlock(&sw_dev_mtx);
2639 mtx_unlock(&sw_dev_mtx);
2641 gp = g_new_geomf(&g_swap_class, "swap");
2642 cp = g_new_consumer(gp);
2643 cp->index = 1; /* Number of active I/Os, plus one for being active. */
2644 cp->flags |= G_CF_DIRECT_SEND | G_CF_DIRECT_RECEIVE;
2647 * XXX: Every time you think you can improve the margin for
2648 * footshooting, somebody depends on the ability to do so:
2649 * savecore(8) wants to write to our swapdev so we cannot
2650 * set an exclusive count :-(
2652 error = g_access(cp, 1, 1, 0);
2655 g_destroy_consumer(cp);
2658 nblks = pp->mediasize / DEV_BSIZE;
2659 swaponsomething(vp, cp, nblks, swapgeom_strategy,
2660 swapgeom_close, dev2udev(dev),
2661 (pp->flags & G_PF_ACCEPT_UNMAPPED) != 0 ? SW_UNMAPPED : 0);
2666 swapongeom(struct vnode *vp)
2670 vn_lock(vp, LK_EXCLUSIVE | LK_RETRY);
2671 if (vp->v_type != VCHR || (vp->v_iflag & VI_DOOMED) != 0) {
2675 error = swapongeom_locked(vp->v_rdev, vp);
2676 g_topology_unlock();
2685 * This is used mainly for network filesystem (read: probably only tested
2686 * with NFS) swapfiles.
2691 swapdev_strategy(struct buf *bp, struct swdevt *sp)
2695 bp->b_blkno = ctodb(bp->b_blkno - sp->sw_first);
2699 if (bp->b_iocmd == BIO_WRITE) {
2701 bufobj_wdrop(bp->b_bufobj);
2702 bufobj_wref(&vp2->v_bufobj);
2704 if (bp->b_bufobj != &vp2->v_bufobj)
2705 bp->b_bufobj = &vp2->v_bufobj;
2707 bp->b_iooffset = dbtob(bp->b_blkno);
2713 swapdev_close(struct thread *td, struct swdevt *sp)
2716 VOP_CLOSE(sp->sw_vp, FREAD | FWRITE, td->td_ucred, td);
2722 swaponvp(struct thread *td, struct vnode *vp, u_long nblks)
2729 mtx_lock(&sw_dev_mtx);
2730 TAILQ_FOREACH(sp, &swtailq, sw_list) {
2731 if (sp->sw_id == vp) {
2732 mtx_unlock(&sw_dev_mtx);
2736 mtx_unlock(&sw_dev_mtx);
2738 (void) vn_lock(vp, LK_EXCLUSIVE | LK_RETRY);
2740 error = mac_system_check_swapon(td->td_ucred, vp);
2743 error = VOP_OPEN(vp, FREAD | FWRITE, td->td_ucred, td, NULL);
2744 (void) VOP_UNLOCK(vp, 0);
2748 swaponsomething(vp, vp, nblks, swapdev_strategy, swapdev_close,
2754 sysctl_swap_async_max(SYSCTL_HANDLER_ARGS)
2758 new = nsw_wcount_async_max;
2759 error = sysctl_handle_int(oidp, &new, 0, req);
2760 if (error != 0 || req->newptr == NULL)
2763 if (new > nswbuf / 2 || new < 1)
2766 mtx_lock(&pbuf_mtx);
2767 while (nsw_wcount_async_max != new) {
2769 * Adjust difference. If the current async count is too low,
2770 * we will need to sqeeze our update slowly in. Sleep with a
2771 * higher priority than getpbuf() to finish faster.
2773 n = new - nsw_wcount_async_max;
2774 if (nsw_wcount_async + n >= 0) {
2775 nsw_wcount_async += n;
2776 nsw_wcount_async_max += n;
2777 wakeup(&nsw_wcount_async);
2779 nsw_wcount_async_max -= nsw_wcount_async;
2780 nsw_wcount_async = 0;
2781 msleep(&nsw_wcount_async, &pbuf_mtx, PSWP,
2785 mtx_unlock(&pbuf_mtx);