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("$FreeBSD$");
75 #include <sys/param.h>
76 #include <sys/systm.h>
78 #include <sys/kernel.h>
84 #include <sys/fcntl.h>
85 #include <sys/mount.h>
86 #include <sys/namei.h>
87 #include <sys/vnode.h>
88 #include <sys/malloc.h>
89 #include <sys/resource.h>
90 #include <sys/resourcevar.h>
91 #include <sys/sysctl.h>
92 #include <sys/sysproto.h>
93 #include <sys/blist.h>
96 #include <sys/vmmeter.h>
98 #include <security/mac/mac_framework.h>
102 #include <vm/vm_map.h>
103 #include <vm/vm_kern.h>
104 #include <vm/vm_object.h>
105 #include <vm/vm_page.h>
106 #include <vm/vm_pager.h>
107 #include <vm/vm_pageout.h>
108 #include <vm/vm_param.h>
109 #include <vm/swap_pager.h>
110 #include <vm/vm_extern.h>
113 #include <geom/geom.h>
116 * SWB_NPAGES must be a power of 2. It may be set to 1, 2, 4, 8, or 16
117 * pages per allocation. We recommend you stick with the default of 8.
118 * The 16-page limit is due to the radix code (kern/subr_blist.c).
120 #ifndef MAX_PAGEOUT_CLUSTER
121 #define MAX_PAGEOUT_CLUSTER 16
124 #if !defined(SWB_NPAGES)
125 #define SWB_NPAGES MAX_PAGEOUT_CLUSTER
129 * Piecemeal swap metadata structure. Swap is stored in a radix tree.
131 * If SWB_NPAGES is 8 and sizeof(char *) == sizeof(daddr_t), our radix
132 * is basically 8. Assuming PAGE_SIZE == 4096, one tree level represents
133 * 32K worth of data, two levels represent 256K, three levels represent
134 * 2 MBytes. This is acceptable.
136 * Overall memory utilization is about the same as the old swap structure.
138 #define SWCORRECT(n) (sizeof(void *) * (n) / sizeof(daddr_t))
139 #define SWAP_META_PAGES (SWB_NPAGES * 2)
140 #define SWAP_META_MASK (SWAP_META_PAGES - 1)
143 struct swblock *swb_hnext;
144 vm_object_t swb_object;
145 vm_pindex_t swb_index;
147 daddr_t swb_pages[SWAP_META_PAGES];
150 static struct mtx sw_dev_mtx;
151 static TAILQ_HEAD(, swdevt) swtailq = TAILQ_HEAD_INITIALIZER(swtailq);
152 static struct swdevt *swdevhd; /* Allocate from here next */
153 static int nswapdev; /* Number of swap devices */
154 int swap_pager_avail;
155 static int swdev_syscall_active = 0; /* serialize swap(on|off) */
157 static vm_ooffset_t swap_total;
158 SYSCTL_QUAD(_vm, OID_AUTO, swap_total, CTLFLAG_RD, &swap_total, 0,
159 "Total amount of available swap storage.");
160 static vm_ooffset_t swap_reserved;
161 SYSCTL_QUAD(_vm, OID_AUTO, swap_reserved, CTLFLAG_RD, &swap_reserved, 0,
162 "Amount of swap storage needed to back all allocated anonymous memory.");
163 static int overcommit = 0;
164 SYSCTL_INT(_vm, OID_AUTO, overcommit, CTLFLAG_RW, &overcommit, 0,
165 "Configure virtual memory overcommit behavior. See tuning(7) "
168 /* bits from overcommit */
169 #define SWAP_RESERVE_FORCE_ON (1 << 0)
170 #define SWAP_RESERVE_RLIMIT_ON (1 << 1)
171 #define SWAP_RESERVE_ALLOW_NONWIRED (1 << 2)
174 swap_reserve(vm_ooffset_t incr)
177 return (swap_reserve_by_uid(incr, curthread->td_ucred->cr_ruidinfo));
181 swap_reserve_by_uid(vm_ooffset_t incr, struct uidinfo *uip)
186 static struct timeval lastfail;
188 if (incr & PAGE_MASK)
189 panic("swap_reserve: & PAGE_MASK");
192 mtx_lock(&sw_dev_mtx);
193 r = swap_reserved + incr;
194 if (overcommit & SWAP_RESERVE_ALLOW_NONWIRED) {
195 s = cnt.v_page_count - cnt.v_free_reserved - cnt.v_wire_count;
200 if ((overcommit & SWAP_RESERVE_FORCE_ON) == 0 || r <= s ||
201 (error = priv_check(curthread, PRIV_VM_SWAP_NOQUOTA)) == 0) {
205 mtx_unlock(&sw_dev_mtx);
209 UIDINFO_VMSIZE_LOCK(uip);
210 if ((overcommit & SWAP_RESERVE_RLIMIT_ON) != 0 &&
211 uip->ui_vmsize + incr > lim_cur(curproc, RLIMIT_SWAP) &&
212 priv_check(curthread, PRIV_VM_SWAP_NORLIMIT))
215 uip->ui_vmsize += incr;
216 UIDINFO_VMSIZE_UNLOCK(uip);
217 PROC_UNLOCK(curproc);
219 mtx_lock(&sw_dev_mtx);
220 swap_reserved -= incr;
221 mtx_unlock(&sw_dev_mtx);
224 if (!res && ppsratecheck(&lastfail, &curfail, 1)) {
225 printf("uid %d, pid %d: swap reservation for %jd bytes failed\n",
226 curproc->p_pid, uip->ui_uid, incr);
233 swap_reserve_force(vm_ooffset_t incr)
237 mtx_lock(&sw_dev_mtx);
238 swap_reserved += incr;
239 mtx_unlock(&sw_dev_mtx);
241 uip = curthread->td_ucred->cr_ruidinfo;
243 UIDINFO_VMSIZE_LOCK(uip);
244 uip->ui_vmsize += incr;
245 UIDINFO_VMSIZE_UNLOCK(uip);
246 PROC_UNLOCK(curproc);
250 swap_release(vm_ooffset_t decr)
255 uip = curthread->td_ucred->cr_ruidinfo;
256 swap_release_by_uid(decr, uip);
257 PROC_UNLOCK(curproc);
261 swap_release_by_uid(vm_ooffset_t decr, struct uidinfo *uip)
264 if (decr & PAGE_MASK)
265 panic("swap_release: & PAGE_MASK");
267 mtx_lock(&sw_dev_mtx);
268 if (swap_reserved < decr)
269 panic("swap_reserved < decr");
270 swap_reserved -= decr;
271 mtx_unlock(&sw_dev_mtx);
273 UIDINFO_VMSIZE_LOCK(uip);
274 if (uip->ui_vmsize < decr)
275 printf("negative vmsize for uid = %d\n", uip->ui_uid);
276 uip->ui_vmsize -= decr;
277 UIDINFO_VMSIZE_UNLOCK(uip);
280 static void swapdev_strategy(struct buf *, struct swdevt *sw);
282 #define SWM_FREE 0x02 /* free, period */
283 #define SWM_POP 0x04 /* pop out */
285 int swap_pager_full = 2; /* swap space exhaustion (task killing) */
286 static int swap_pager_almost_full = 1; /* swap space exhaustion (w/hysteresis)*/
287 static int nsw_rcount; /* free read buffers */
288 static int nsw_wcount_sync; /* limit write buffers / synchronous */
289 static int nsw_wcount_async; /* limit write buffers / asynchronous */
290 static int nsw_wcount_async_max;/* assigned maximum */
291 static int nsw_cluster_max; /* maximum VOP I/O allowed */
293 static struct swblock **swhash;
294 static int swhash_mask;
295 static struct mtx swhash_mtx;
297 static int swap_async_max = 4; /* maximum in-progress async I/O's */
298 static struct sx sw_alloc_sx;
301 SYSCTL_INT(_vm, OID_AUTO, swap_async_max,
302 CTLFLAG_RW, &swap_async_max, 0, "Maximum running async swap ops");
305 * "named" and "unnamed" anon region objects. Try to reduce the overhead
306 * of searching a named list by hashing it just a little.
311 #define NOBJLIST(handle) \
312 (&swap_pager_object_list[((int)(intptr_t)handle >> 4) & (NOBJLISTS-1)])
314 static struct mtx sw_alloc_mtx; /* protect list manipulation */
315 static struct pagerlst swap_pager_object_list[NOBJLISTS];
316 static uma_zone_t swap_zone;
317 static struct vm_object swap_zone_obj;
320 * pagerops for OBJT_SWAP - "swap pager". Some ops are also global procedure
321 * calls hooked from other parts of the VM system and do not appear here.
322 * (see vm/swap_pager.h).
325 swap_pager_alloc(void *handle, vm_ooffset_t size,
326 vm_prot_t prot, vm_ooffset_t offset, struct ucred *);
327 static void swap_pager_dealloc(vm_object_t object);
328 static int swap_pager_getpages(vm_object_t, vm_page_t *, int, int);
329 static void swap_pager_putpages(vm_object_t, vm_page_t *, int, boolean_t, int *);
331 swap_pager_haspage(vm_object_t object, vm_pindex_t pindex, int *before, int *after);
332 static void swap_pager_init(void);
333 static void swap_pager_unswapped(vm_page_t);
334 static void swap_pager_swapoff(struct swdevt *sp);
336 struct pagerops swappagerops = {
337 .pgo_init = swap_pager_init, /* early system initialization of pager */
338 .pgo_alloc = swap_pager_alloc, /* allocate an OBJT_SWAP object */
339 .pgo_dealloc = swap_pager_dealloc, /* deallocate an OBJT_SWAP object */
340 .pgo_getpages = swap_pager_getpages, /* pagein */
341 .pgo_putpages = swap_pager_putpages, /* pageout */
342 .pgo_haspage = swap_pager_haspage, /* get backing store status for page */
343 .pgo_pageunswapped = swap_pager_unswapped, /* remove swap related to page */
347 * dmmax is in page-sized chunks with the new swap system. It was
348 * dev-bsized chunks in the old. dmmax is always a power of 2.
350 * swap_*() routines are externally accessible. swp_*() routines are
354 static int nswap_lowat = 128; /* in pages, swap_pager_almost_full warn */
355 static int nswap_hiwat = 512; /* in pages, swap_pager_almost_full warn */
357 SYSCTL_INT(_vm, OID_AUTO, dmmax,
358 CTLFLAG_RD, &dmmax, 0, "Maximum size of a swap block");
360 static void swp_sizecheck(void);
361 static void swp_pager_async_iodone(struct buf *bp);
362 static int swapongeom(struct thread *, struct vnode *);
363 static int swaponvp(struct thread *, struct vnode *, u_long);
364 static int swapoff_one(struct swdevt *sp, struct ucred *cred);
367 * Swap bitmap functions
369 static void swp_pager_freeswapspace(daddr_t blk, int npages);
370 static daddr_t swp_pager_getswapspace(int npages);
375 static struct swblock **swp_pager_hash(vm_object_t object, vm_pindex_t index);
376 static void swp_pager_meta_build(vm_object_t, vm_pindex_t, daddr_t);
377 static void swp_pager_meta_free(vm_object_t, vm_pindex_t, daddr_t);
378 static void swp_pager_meta_free_all(vm_object_t);
379 static daddr_t swp_pager_meta_ctl(vm_object_t, vm_pindex_t, int);
382 swp_pager_free_nrpage(vm_page_t m)
385 if (m->wire_count == 0)
390 * SWP_SIZECHECK() - update swap_pager_full indication
392 * update the swap_pager_almost_full indication and warn when we are
393 * about to run out of swap space, using lowat/hiwat hysteresis.
395 * Clear swap_pager_full ( task killing ) indication when lowat is met.
397 * No restrictions on call
398 * This routine may not block.
399 * This routine must be called at splvm()
405 if (swap_pager_avail < nswap_lowat) {
406 if (swap_pager_almost_full == 0) {
407 printf("swap_pager: out of swap space\n");
408 swap_pager_almost_full = 1;
412 if (swap_pager_avail > nswap_hiwat)
413 swap_pager_almost_full = 0;
418 * SWP_PAGER_HASH() - hash swap meta data
420 * This is an helper function which hashes the swapblk given
421 * the object and page index. It returns a pointer to a pointer
422 * to the object, or a pointer to a NULL pointer if it could not
425 * This routine must be called at splvm().
427 static struct swblock **
428 swp_pager_hash(vm_object_t object, vm_pindex_t index)
430 struct swblock **pswap;
431 struct swblock *swap;
433 index &= ~(vm_pindex_t)SWAP_META_MASK;
434 pswap = &swhash[(index ^ (int)(intptr_t)object) & swhash_mask];
435 while ((swap = *pswap) != NULL) {
436 if (swap->swb_object == object &&
437 swap->swb_index == index
441 pswap = &swap->swb_hnext;
447 * SWAP_PAGER_INIT() - initialize the swap pager!
449 * Expected to be started from system init. NOTE: This code is run
450 * before much else so be careful what you depend on. Most of the VM
451 * system has yet to be initialized at this point.
454 swap_pager_init(void)
457 * Initialize object lists
461 for (i = 0; i < NOBJLISTS; ++i)
462 TAILQ_INIT(&swap_pager_object_list[i]);
463 mtx_init(&sw_alloc_mtx, "swap_pager list", NULL, MTX_DEF);
464 mtx_init(&sw_dev_mtx, "swapdev", NULL, MTX_DEF);
467 * Device Stripe, in PAGE_SIZE'd blocks
469 dmmax = SWB_NPAGES * 2;
473 * SWAP_PAGER_SWAP_INIT() - swap pager initialization from pageout process
475 * Expected to be started from pageout process once, prior to entering
479 swap_pager_swap_init(void)
484 * Number of in-transit swap bp operations. Don't
485 * exhaust the pbufs completely. Make sure we
486 * initialize workable values (0 will work for hysteresis
487 * but it isn't very efficient).
489 * The nsw_cluster_max is constrained by the bp->b_pages[]
490 * array (MAXPHYS/PAGE_SIZE) and our locally defined
491 * MAX_PAGEOUT_CLUSTER. Also be aware that swap ops are
492 * constrained by the swap device interleave stripe size.
494 * Currently we hardwire nsw_wcount_async to 4. This limit is
495 * designed to prevent other I/O from having high latencies due to
496 * our pageout I/O. The value 4 works well for one or two active swap
497 * devices but is probably a little low if you have more. Even so,
498 * a higher value would probably generate only a limited improvement
499 * with three or four active swap devices since the system does not
500 * typically have to pageout at extreme bandwidths. We will want
501 * at least 2 per swap devices, and 4 is a pretty good value if you
502 * have one NFS swap device due to the command/ack latency over NFS.
503 * So it all works out pretty well.
505 nsw_cluster_max = min((MAXPHYS/PAGE_SIZE), MAX_PAGEOUT_CLUSTER);
508 nsw_rcount = (nswbuf + 1) / 2;
509 nsw_wcount_sync = (nswbuf + 3) / 4;
510 nsw_wcount_async = 4;
511 nsw_wcount_async_max = nsw_wcount_async;
512 mtx_unlock(&pbuf_mtx);
515 * Initialize our zone. Right now I'm just guessing on the number
516 * we need based on the number of pages in the system. Each swblock
517 * can hold 16 pages, so this is probably overkill. This reservation
518 * is typically limited to around 32MB by default.
520 n = cnt.v_page_count / 2;
521 if (maxswzone && n > maxswzone / sizeof(struct swblock))
522 n = maxswzone / sizeof(struct swblock);
524 swap_zone = uma_zcreate("SWAPMETA", sizeof(struct swblock), NULL, NULL,
525 NULL, NULL, UMA_ALIGN_PTR, UMA_ZONE_NOFREE | UMA_ZONE_VM);
526 if (swap_zone == NULL)
527 panic("failed to create swap_zone.");
529 if (uma_zone_set_obj(swap_zone, &swap_zone_obj, n))
532 * if the allocation failed, try a zone two thirds the
533 * size of the previous attempt.
538 printf("Swap zone entries reduced from %d to %d.\n", n2, n);
542 * Initialize our meta-data hash table. The swapper does not need to
543 * be quite as efficient as the VM system, so we do not use an
544 * oversized hash table.
546 * n: size of hash table, must be power of 2
547 * swhash_mask: hash table index mask
549 for (n = 1; n < n2 / 8; n *= 2)
551 swhash = malloc(sizeof(struct swblock *) * n, M_VMPGDATA, M_WAITOK | M_ZERO);
553 mtx_init(&swhash_mtx, "swap_pager swhash", NULL, MTX_DEF);
557 * SWAP_PAGER_ALLOC() - allocate a new OBJT_SWAP VM object and instantiate
558 * its metadata structures.
560 * This routine is called from the mmap and fork code to create a new
561 * OBJT_SWAP object. We do this by creating an OBJT_DEFAULT object
562 * and then converting it with swp_pager_meta_build().
564 * This routine may block in vm_object_allocate() and create a named
565 * object lookup race, so we must interlock. We must also run at
566 * splvm() for the object lookup to handle races with interrupts, but
567 * we do not have to maintain splvm() in between the lookup and the
568 * add because (I believe) it is not possible to attempt to create
569 * a new swap object w/handle when a default object with that handle
575 swap_pager_alloc(void *handle, vm_ooffset_t size, vm_prot_t prot,
576 vm_ooffset_t offset, struct ucred *cred)
583 pindex = OFF_TO_IDX(offset + PAGE_MASK + size);
587 * Reference existing named region or allocate new one. There
588 * should not be a race here against swp_pager_meta_build()
589 * as called from vm_page_remove() in regards to the lookup
592 sx_xlock(&sw_alloc_sx);
593 object = vm_pager_object_lookup(NOBJLIST(handle), handle);
594 if (object == NULL) {
596 uip = cred->cr_ruidinfo;
597 if (!swap_reserve_by_uid(size, uip)) {
598 sx_xunlock(&sw_alloc_sx);
604 object = vm_object_allocate(OBJT_DEFAULT, pindex);
605 VM_OBJECT_LOCK(object);
606 object->handle = handle;
609 object->charge = size;
611 swp_pager_meta_build(object, 0, SWAPBLK_NONE);
612 VM_OBJECT_UNLOCK(object);
614 sx_xunlock(&sw_alloc_sx);
618 uip = cred->cr_ruidinfo;
619 if (!swap_reserve_by_uid(size, uip))
623 object = vm_object_allocate(OBJT_DEFAULT, pindex);
624 VM_OBJECT_LOCK(object);
627 object->charge = size;
629 swp_pager_meta_build(object, 0, SWAPBLK_NONE);
630 VM_OBJECT_UNLOCK(object);
636 * SWAP_PAGER_DEALLOC() - remove swap metadata from object
638 * The swap backing for the object is destroyed. The code is
639 * designed such that we can reinstantiate it later, but this
640 * routine is typically called only when the entire object is
641 * about to be destroyed.
643 * This routine may block, but no longer does.
645 * The object must be locked or unreferenceable.
648 swap_pager_dealloc(vm_object_t object)
652 * Remove from list right away so lookups will fail if we block for
653 * pageout completion.
655 if (object->handle != NULL) {
656 mtx_lock(&sw_alloc_mtx);
657 TAILQ_REMOVE(NOBJLIST(object->handle), object, pager_object_list);
658 mtx_unlock(&sw_alloc_mtx);
661 VM_OBJECT_LOCK_ASSERT(object, MA_OWNED);
662 vm_object_pip_wait(object, "swpdea");
665 * Free all remaining metadata. We only bother to free it from
666 * the swap meta data. We do not attempt to free swapblk's still
667 * associated with vm_page_t's for this object. We do not care
668 * if paging is still in progress on some objects.
670 swp_pager_meta_free_all(object);
673 /************************************************************************
674 * SWAP PAGER BITMAP ROUTINES *
675 ************************************************************************/
678 * SWP_PAGER_GETSWAPSPACE() - allocate raw swap space
680 * Allocate swap for the requested number of pages. The starting
681 * swap block number (a page index) is returned or SWAPBLK_NONE
682 * if the allocation failed.
684 * Also has the side effect of advising that somebody made a mistake
685 * when they configured swap and didn't configure enough.
687 * Must be called at splvm() to avoid races with bitmap frees from
688 * vm_page_remove() aka swap_pager_page_removed().
690 * This routine may not block
691 * This routine must be called at splvm().
693 * We allocate in round-robin fashion from the configured devices.
696 swp_pager_getswapspace(int npages)
703 mtx_lock(&sw_dev_mtx);
705 for (i = 0; i < nswapdev; i++) {
707 sp = TAILQ_FIRST(&swtailq);
708 if (!(sp->sw_flags & SW_CLOSING)) {
709 blk = blist_alloc(sp->sw_blist, npages);
710 if (blk != SWAPBLK_NONE) {
712 sp->sw_used += npages;
713 swap_pager_avail -= npages;
715 swdevhd = TAILQ_NEXT(sp, sw_list);
719 sp = TAILQ_NEXT(sp, sw_list);
721 if (swap_pager_full != 2) {
722 printf("swap_pager_getswapspace(%d): failed\n", npages);
724 swap_pager_almost_full = 1;
728 mtx_unlock(&sw_dev_mtx);
733 swp_pager_isondev(daddr_t blk, struct swdevt *sp)
736 return (blk >= sp->sw_first && blk < sp->sw_end);
740 swp_pager_strategy(struct buf *bp)
744 mtx_lock(&sw_dev_mtx);
745 TAILQ_FOREACH(sp, &swtailq, sw_list) {
746 if (bp->b_blkno >= sp->sw_first && bp->b_blkno < sp->sw_end) {
747 mtx_unlock(&sw_dev_mtx);
748 sp->sw_strategy(bp, sp);
752 panic("Swapdev not found");
757 * SWP_PAGER_FREESWAPSPACE() - free raw swap space
759 * This routine returns the specified swap blocks back to the bitmap.
761 * Note: This routine may not block (it could in the old swap code),
762 * and through the use of the new blist routines it does not block.
764 * We must be called at splvm() to avoid races with bitmap frees from
765 * vm_page_remove() aka swap_pager_page_removed().
767 * This routine may not block
768 * This routine must be called at splvm().
771 swp_pager_freeswapspace(daddr_t blk, int npages)
775 mtx_lock(&sw_dev_mtx);
776 TAILQ_FOREACH(sp, &swtailq, sw_list) {
777 if (blk >= sp->sw_first && blk < sp->sw_end) {
778 sp->sw_used -= npages;
780 * If we are attempting to stop swapping on
781 * this device, we don't want to mark any
782 * blocks free lest they be reused.
784 if ((sp->sw_flags & SW_CLOSING) == 0) {
785 blist_free(sp->sw_blist, blk - sp->sw_first,
787 swap_pager_avail += npages;
790 mtx_unlock(&sw_dev_mtx);
794 panic("Swapdev not found");
798 * SWAP_PAGER_FREESPACE() - frees swap blocks associated with a page
799 * range within an object.
801 * This is a globally accessible routine.
803 * This routine removes swapblk assignments from swap metadata.
805 * The external callers of this routine typically have already destroyed
806 * or renamed vm_page_t's associated with this range in the object so
809 * This routine may be called at any spl. We up our spl to splvm temporarily
810 * in order to perform the metadata removal.
813 swap_pager_freespace(vm_object_t object, vm_pindex_t start, vm_size_t size)
816 VM_OBJECT_LOCK_ASSERT(object, MA_OWNED);
817 swp_pager_meta_free(object, start, size);
821 * SWAP_PAGER_RESERVE() - reserve swap blocks in object
823 * Assigns swap blocks to the specified range within the object. The
824 * swap blocks are not zerod. Any previous swap assignment is destroyed.
826 * Returns 0 on success, -1 on failure.
829 swap_pager_reserve(vm_object_t object, vm_pindex_t start, vm_size_t size)
832 daddr_t blk = SWAPBLK_NONE;
833 vm_pindex_t beg = start; /* save start index */
835 VM_OBJECT_LOCK(object);
839 while ((blk = swp_pager_getswapspace(n)) == SWAPBLK_NONE) {
842 swp_pager_meta_free(object, beg, start - beg);
843 VM_OBJECT_UNLOCK(object);
848 swp_pager_meta_build(object, start, blk);
854 swp_pager_meta_free(object, start, n);
855 VM_OBJECT_UNLOCK(object);
860 * SWAP_PAGER_COPY() - copy blocks from source pager to destination pager
861 * and destroy the source.
863 * Copy any valid swapblks from the source to the destination. In
864 * cases where both the source and destination have a valid swapblk,
865 * we keep the destination's.
867 * This routine is allowed to block. It may block allocating metadata
868 * indirectly through swp_pager_meta_build() or if paging is still in
869 * progress on the source.
871 * This routine can be called at any spl
873 * XXX vm_page_collapse() kinda expects us not to block because we
874 * supposedly do not need to allocate memory, but for the moment we
875 * *may* have to get a little memory from the zone allocator, but
876 * it is taken from the interrupt memory. We should be ok.
878 * The source object contains no vm_page_t's (which is just as well)
880 * The source object is of type OBJT_SWAP.
882 * The source and destination objects must be locked or
883 * inaccessible (XXX are they ?)
886 swap_pager_copy(vm_object_t srcobject, vm_object_t dstobject,
887 vm_pindex_t offset, int destroysource)
891 VM_OBJECT_LOCK_ASSERT(srcobject, MA_OWNED);
892 VM_OBJECT_LOCK_ASSERT(dstobject, MA_OWNED);
895 * If destroysource is set, we remove the source object from the
896 * swap_pager internal queue now.
899 if (srcobject->handle != NULL) {
900 mtx_lock(&sw_alloc_mtx);
902 NOBJLIST(srcobject->handle),
906 mtx_unlock(&sw_alloc_mtx);
911 * transfer source to destination.
913 for (i = 0; i < dstobject->size; ++i) {
917 * Locate (without changing) the swapblk on the destination,
918 * unless it is invalid in which case free it silently, or
919 * if the destination is a resident page, in which case the
920 * source is thrown away.
922 dstaddr = swp_pager_meta_ctl(dstobject, i, 0);
924 if (dstaddr == SWAPBLK_NONE) {
926 * Destination has no swapblk and is not resident,
931 srcaddr = swp_pager_meta_ctl(
937 if (srcaddr != SWAPBLK_NONE) {
939 * swp_pager_meta_build() can sleep.
941 vm_object_pip_add(srcobject, 1);
942 VM_OBJECT_UNLOCK(srcobject);
943 vm_object_pip_add(dstobject, 1);
944 swp_pager_meta_build(dstobject, i, srcaddr);
945 vm_object_pip_wakeup(dstobject);
946 VM_OBJECT_LOCK(srcobject);
947 vm_object_pip_wakeup(srcobject);
951 * Destination has valid swapblk or it is represented
952 * by a resident page. We destroy the sourceblock.
955 swp_pager_meta_ctl(srcobject, i + offset, SWM_FREE);
960 * Free left over swap blocks in source.
962 * We have to revert the type to OBJT_DEFAULT so we do not accidently
963 * double-remove the object from the swap queues.
966 swp_pager_meta_free_all(srcobject);
968 * Reverting the type is not necessary, the caller is going
969 * to destroy srcobject directly, but I'm doing it here
970 * for consistency since we've removed the object from its
973 srcobject->type = OBJT_DEFAULT;
978 * SWAP_PAGER_HASPAGE() - determine if we have good backing store for
979 * the requested page.
981 * We determine whether good backing store exists for the requested
982 * page and return TRUE if it does, FALSE if it doesn't.
984 * If TRUE, we also try to determine how much valid, contiguous backing
985 * store exists before and after the requested page within a reasonable
986 * distance. We do not try to restrict it to the swap device stripe
987 * (that is handled in getpages/putpages). It probably isn't worth
991 swap_pager_haspage(vm_object_t object, vm_pindex_t pindex, int *before, int *after)
995 VM_OBJECT_LOCK_ASSERT(object, MA_OWNED);
997 * do we have good backing store at the requested index ?
999 blk0 = swp_pager_meta_ctl(object, pindex, 0);
1001 if (blk0 == SWAPBLK_NONE) {
1010 * find backwards-looking contiguous good backing store
1012 if (before != NULL) {
1015 for (i = 1; i < (SWB_NPAGES/2); ++i) {
1020 blk = swp_pager_meta_ctl(object, pindex - i, 0);
1021 if (blk != blk0 - i)
1028 * find forward-looking contiguous good backing store
1030 if (after != NULL) {
1033 for (i = 1; i < (SWB_NPAGES/2); ++i) {
1036 blk = swp_pager_meta_ctl(object, pindex + i, 0);
1037 if (blk != blk0 + i)
1046 * SWAP_PAGER_PAGE_UNSWAPPED() - remove swap backing store related to page
1048 * This removes any associated swap backing store, whether valid or
1049 * not, from the page.
1051 * This routine is typically called when a page is made dirty, at
1052 * which point any associated swap can be freed. MADV_FREE also
1053 * calls us in a special-case situation
1055 * NOTE!!! If the page is clean and the swap was valid, the caller
1056 * should make the page dirty before calling this routine. This routine
1057 * does NOT change the m->dirty status of the page. Also: MADV_FREE
1060 * This routine may not block
1061 * This routine must be called at splvm()
1064 swap_pager_unswapped(vm_page_t m)
1067 VM_OBJECT_LOCK_ASSERT(m->object, MA_OWNED);
1068 swp_pager_meta_ctl(m->object, m->pindex, SWM_FREE);
1072 * SWAP_PAGER_GETPAGES() - bring pages in from swap
1074 * Attempt to retrieve (m, count) pages from backing store, but make
1075 * sure we retrieve at least m[reqpage]. We try to load in as large
1076 * a chunk surrounding m[reqpage] as is contiguous in swap and which
1077 * belongs to the same object.
1079 * The code is designed for asynchronous operation and
1080 * immediate-notification of 'reqpage' but tends not to be
1081 * used that way. Please do not optimize-out this algorithmic
1082 * feature, I intend to improve on it in the future.
1084 * The parent has a single vm_object_pip_add() reference prior to
1085 * calling us and we should return with the same.
1087 * The parent has BUSY'd the pages. We should return with 'm'
1088 * left busy, but the others adjusted.
1091 swap_pager_getpages(vm_object_t object, vm_page_t *m, int count, int reqpage)
1101 KASSERT(mreq->object == object,
1102 ("swap_pager_getpages: object mismatch %p/%p",
1103 object, mreq->object));
1106 * Calculate range to retrieve. The pages have already been assigned
1107 * their swapblks. We require a *contiguous* range but we know it to
1108 * not span devices. If we do not supply it, bad things
1109 * happen. Note that blk, iblk & jblk can be SWAPBLK_NONE, but the
1110 * loops are set up such that the case(s) are handled implicitly.
1112 * The swp_*() calls must be made with the object locked.
1114 blk = swp_pager_meta_ctl(mreq->object, mreq->pindex, 0);
1116 for (i = reqpage - 1; i >= 0; --i) {
1119 iblk = swp_pager_meta_ctl(m[i]->object, m[i]->pindex, 0);
1120 if (blk != iblk + (reqpage - i))
1125 for (j = reqpage + 1; j < count; ++j) {
1128 jblk = swp_pager_meta_ctl(m[j]->object, m[j]->pindex, 0);
1129 if (blk != jblk - (j - reqpage))
1134 * free pages outside our collection range. Note: we never free
1135 * mreq, it must remain busy throughout.
1137 if (0 < i || j < count) {
1141 for (k = 0; k < i; ++k) {
1143 vm_page_lock_queues();
1144 swp_pager_free_nrpage(m[k]);
1145 vm_page_unlock_queues();
1146 vm_page_unlock(m[k]);
1148 for (k = j; k < count; ++k) {
1150 vm_page_lock_queues();
1151 swp_pager_free_nrpage(m[k]);
1152 vm_page_unlock_queues();
1153 vm_page_unlock(m[k]);
1158 * Return VM_PAGER_FAIL if we have nothing to do. Return mreq
1159 * still busy, but the others unbusied.
1161 if (blk == SWAPBLK_NONE)
1162 return (VM_PAGER_FAIL);
1165 * Getpbuf() can sleep.
1167 VM_OBJECT_UNLOCK(object);
1169 * Get a swap buffer header to perform the IO
1171 bp = getpbuf(&nsw_rcount);
1172 bp->b_flags |= B_PAGING;
1175 * map our page(s) into kva for input
1177 pmap_qenter((vm_offset_t)bp->b_data, m + i, j - i);
1179 bp->b_iocmd = BIO_READ;
1180 bp->b_iodone = swp_pager_async_iodone;
1181 bp->b_rcred = crhold(thread0.td_ucred);
1182 bp->b_wcred = crhold(thread0.td_ucred);
1183 bp->b_blkno = blk - (reqpage - i);
1184 bp->b_bcount = PAGE_SIZE * (j - i);
1185 bp->b_bufsize = PAGE_SIZE * (j - i);
1186 bp->b_pager.pg_reqpage = reqpage - i;
1188 VM_OBJECT_LOCK(object);
1192 for (k = i; k < j; ++k) {
1193 bp->b_pages[k - i] = m[k];
1194 m[k]->oflags |= VPO_SWAPINPROG;
1197 bp->b_npages = j - i;
1199 PCPU_INC(cnt.v_swapin);
1200 PCPU_ADD(cnt.v_swappgsin, bp->b_npages);
1203 * We still hold the lock on mreq, and our automatic completion routine
1204 * does not remove it.
1206 vm_object_pip_add(object, bp->b_npages);
1207 VM_OBJECT_UNLOCK(object);
1210 * perform the I/O. NOTE!!! bp cannot be considered valid after
1211 * this point because we automatically release it on completion.
1212 * Instead, we look at the one page we are interested in which we
1213 * still hold a lock on even through the I/O completion.
1215 * The other pages in our m[] array are also released on completion,
1216 * so we cannot assume they are valid anymore either.
1218 * NOTE: b_blkno is destroyed by the call to swapdev_strategy
1221 swp_pager_strategy(bp);
1224 * wait for the page we want to complete. VPO_SWAPINPROG is always
1225 * cleared on completion. If an I/O error occurs, SWAPBLK_NONE
1226 * is set in the meta-data.
1228 VM_OBJECT_LOCK(object);
1229 while ((mreq->oflags & VPO_SWAPINPROG) != 0) {
1230 mreq->oflags |= VPO_WANTED;
1231 PCPU_INC(cnt.v_intrans);
1232 if (msleep(mreq, VM_OBJECT_MTX(object), PSWP, "swread", hz*20)) {
1234 "swap_pager: indefinite wait buffer: bufobj: %p, blkno: %jd, size: %ld\n",
1235 bp->b_bufobj, (intmax_t)bp->b_blkno, bp->b_bcount);
1240 * mreq is left busied after completion, but all the other pages
1241 * are freed. If we had an unrecoverable read error the page will
1244 if (mreq->valid != VM_PAGE_BITS_ALL) {
1245 return (VM_PAGER_ERROR);
1247 return (VM_PAGER_OK);
1251 * A final note: in a low swap situation, we cannot deallocate swap
1252 * and mark a page dirty here because the caller is likely to mark
1253 * the page clean when we return, causing the page to possibly revert
1254 * to all-zero's later.
1259 * swap_pager_putpages:
1261 * Assign swap (if necessary) and initiate I/O on the specified pages.
1263 * We support both OBJT_DEFAULT and OBJT_SWAP objects. DEFAULT objects
1264 * are automatically converted to SWAP objects.
1266 * In a low memory situation we may block in VOP_STRATEGY(), but the new
1267 * vm_page reservation system coupled with properly written VFS devices
1268 * should ensure that no low-memory deadlock occurs. This is an area
1271 * The parent has N vm_object_pip_add() references prior to
1272 * calling us and will remove references for rtvals[] that are
1273 * not set to VM_PAGER_PEND. We need to remove the rest on I/O
1276 * The parent has soft-busy'd the pages it passes us and will unbusy
1277 * those whos rtvals[] entry is not set to VM_PAGER_PEND on return.
1278 * We need to unbusy the rest on I/O completion.
1281 swap_pager_putpages(vm_object_t object, vm_page_t *m, int count,
1282 boolean_t sync, int *rtvals)
1287 if (count && m[0]->object != object) {
1288 panic("swap_pager_putpages: object mismatch %p/%p",
1297 * Turn object into OBJT_SWAP
1298 * check for bogus sysops
1299 * force sync if not pageout process
1301 if (object->type != OBJT_SWAP)
1302 swp_pager_meta_build(object, 0, SWAPBLK_NONE);
1303 VM_OBJECT_UNLOCK(object);
1305 if (curproc != pageproc)
1311 * Update nsw parameters from swap_async_max sysctl values.
1312 * Do not let the sysop crash the machine with bogus numbers.
1314 mtx_lock(&pbuf_mtx);
1315 if (swap_async_max != nsw_wcount_async_max) {
1321 if ((n = swap_async_max) > nswbuf / 2)
1328 * Adjust difference ( if possible ). If the current async
1329 * count is too low, we may not be able to make the adjustment
1332 n -= nsw_wcount_async_max;
1333 if (nsw_wcount_async + n >= 0) {
1334 nsw_wcount_async += n;
1335 nsw_wcount_async_max += n;
1336 wakeup(&nsw_wcount_async);
1339 mtx_unlock(&pbuf_mtx);
1344 * Assign swap blocks and issue I/O. We reallocate swap on the fly.
1345 * The page is left dirty until the pageout operation completes
1348 for (i = 0; i < count; i += n) {
1354 * Maximum I/O size is limited by a number of factors.
1356 n = min(BLIST_MAX_ALLOC, count - i);
1357 n = min(n, nsw_cluster_max);
1360 * Get biggest block of swap we can. If we fail, fall
1361 * back and try to allocate a smaller block. Don't go
1362 * overboard trying to allocate space if it would overly
1366 (blk = swp_pager_getswapspace(n)) == SWAPBLK_NONE &&
1371 if (blk == SWAPBLK_NONE) {
1372 for (j = 0; j < n; ++j)
1373 rtvals[i+j] = VM_PAGER_FAIL;
1378 * All I/O parameters have been satisfied, build the I/O
1379 * request and assign the swap space.
1382 bp = getpbuf(&nsw_wcount_sync);
1384 bp = getpbuf(&nsw_wcount_async);
1385 bp->b_flags = B_ASYNC;
1387 bp->b_flags |= B_PAGING;
1388 bp->b_iocmd = BIO_WRITE;
1390 pmap_qenter((vm_offset_t)bp->b_data, &m[i], n);
1392 bp->b_rcred = crhold(thread0.td_ucred);
1393 bp->b_wcred = crhold(thread0.td_ucred);
1394 bp->b_bcount = PAGE_SIZE * n;
1395 bp->b_bufsize = PAGE_SIZE * n;
1398 VM_OBJECT_LOCK(object);
1399 for (j = 0; j < n; ++j) {
1400 vm_page_t mreq = m[i+j];
1402 swp_pager_meta_build(
1407 vm_page_dirty(mreq);
1408 rtvals[i+j] = VM_PAGER_OK;
1410 mreq->oflags |= VPO_SWAPINPROG;
1411 bp->b_pages[j] = mreq;
1413 VM_OBJECT_UNLOCK(object);
1416 * Must set dirty range for NFS to work.
1419 bp->b_dirtyend = bp->b_bcount;
1421 PCPU_INC(cnt.v_swapout);
1422 PCPU_ADD(cnt.v_swappgsout, bp->b_npages);
1427 * NOTE: b_blkno is destroyed by the call to swapdev_strategy
1429 if (sync == FALSE) {
1430 bp->b_iodone = swp_pager_async_iodone;
1432 swp_pager_strategy(bp);
1434 for (j = 0; j < n; ++j)
1435 rtvals[i+j] = VM_PAGER_PEND;
1436 /* restart outter loop */
1443 * NOTE: b_blkno is destroyed by the call to swapdev_strategy
1445 bp->b_iodone = bdone;
1446 swp_pager_strategy(bp);
1449 * Wait for the sync I/O to complete, then update rtvals.
1450 * We just set the rtvals[] to VM_PAGER_PEND so we can call
1451 * our async completion routine at the end, thus avoiding a
1454 bwait(bp, PVM, "swwrt");
1455 for (j = 0; j < n; ++j)
1456 rtvals[i+j] = VM_PAGER_PEND;
1458 * Now that we are through with the bp, we can call the
1459 * normal async completion, which frees everything up.
1461 swp_pager_async_iodone(bp);
1463 VM_OBJECT_LOCK(object);
1467 * swp_pager_async_iodone:
1469 * Completion routine for asynchronous reads and writes from/to swap.
1470 * Also called manually by synchronous code to finish up a bp.
1472 * For READ operations, the pages are PG_BUSY'd. For WRITE operations,
1473 * the pages are vm_page_t->busy'd. For READ operations, we PG_BUSY
1474 * unbusy all pages except the 'main' request page. For WRITE
1475 * operations, we vm_page_t->busy'd unbusy all pages ( we can do this
1476 * because we marked them all VM_PAGER_PEND on return from putpages ).
1478 * This routine may not block.
1481 swp_pager_async_iodone(struct buf *bp)
1484 vm_object_t object = NULL;
1489 if (bp->b_ioflags & BIO_ERROR) {
1491 "swap_pager: I/O error - %s failed; blkno %ld,"
1492 "size %ld, error %d\n",
1493 ((bp->b_iocmd == BIO_READ) ? "pagein" : "pageout"),
1501 * remove the mapping for kernel virtual
1503 pmap_qremove((vm_offset_t)bp->b_data, bp->b_npages);
1506 object = bp->b_pages[0]->object;
1507 VM_OBJECT_LOCK(object);
1511 * cleanup pages. If an error occurs writing to swap, we are in
1512 * very serious trouble. If it happens to be a disk error, though,
1513 * we may be able to recover by reassigning the swap later on. So
1514 * in this case we remove the m->swapblk assignment for the page
1515 * but do not free it in the rlist. The errornous block(s) are thus
1516 * never reallocated as swap. Redirty the page and continue.
1518 for (i = 0; i < bp->b_npages; ++i) {
1519 vm_page_t m = bp->b_pages[i];
1522 vm_page_lock_queues();
1523 m->oflags &= ~VPO_SWAPINPROG;
1525 if (bp->b_ioflags & BIO_ERROR) {
1527 * If an error occurs I'd love to throw the swapblk
1528 * away without freeing it back to swapspace, so it
1529 * can never be used again. But I can't from an
1532 if (bp->b_iocmd == BIO_READ) {
1534 * When reading, reqpage needs to stay
1535 * locked for the parent, but all other
1536 * pages can be freed. We still want to
1537 * wakeup the parent waiting on the page,
1538 * though. ( also: pg_reqpage can be -1 and
1539 * not match anything ).
1541 * We have to wake specifically requested pages
1542 * up too because we cleared VPO_SWAPINPROG and
1543 * someone may be waiting for that.
1545 * NOTE: for reads, m->dirty will probably
1546 * be overridden by the original caller of
1547 * getpages so don't play cute tricks here.
1550 if (i != bp->b_pager.pg_reqpage)
1551 swp_pager_free_nrpage(m);
1555 * If i == bp->b_pager.pg_reqpage, do not wake
1556 * the page up. The caller needs to.
1560 * If a write error occurs, reactivate page
1561 * so it doesn't clog the inactive list,
1562 * then finish the I/O.
1565 vm_page_activate(m);
1566 vm_page_io_finish(m);
1568 } else if (bp->b_iocmd == BIO_READ) {
1570 * NOTE: for reads, m->dirty will probably be
1571 * overridden by the original caller of getpages so
1572 * we cannot set them in order to free the underlying
1573 * swap in a low-swap situation. I don't think we'd
1574 * want to do that anyway, but it was an optimization
1575 * that existed in the old swapper for a time before
1576 * it got ripped out due to precisely this problem.
1578 * If not the requested page then deactivate it.
1580 * Note that the requested page, reqpage, is left
1581 * busied, but we still have to wake it up. The
1582 * other pages are released (unbusied) by
1585 KASSERT(!pmap_page_is_mapped(m),
1586 ("swp_pager_async_iodone: page %p is mapped", m));
1587 m->valid = VM_PAGE_BITS_ALL;
1588 KASSERT(m->dirty == 0,
1589 ("swp_pager_async_iodone: page %p is dirty", m));
1592 * We have to wake specifically requested pages
1593 * up too because we cleared VPO_SWAPINPROG and
1594 * could be waiting for it in getpages. However,
1595 * be sure to not unbusy getpages specifically
1596 * requested page - getpages expects it to be
1599 if (i != bp->b_pager.pg_reqpage) {
1600 vm_page_deactivate(m);
1607 * For write success, clear the dirty
1608 * status, then finish the I/O ( which decrements the
1609 * busy count and possibly wakes waiter's up ).
1611 KASSERT((m->flags & PG_WRITEABLE) == 0,
1612 ("swp_pager_async_iodone: page %p is not write"
1615 vm_page_io_finish(m);
1616 if (vm_page_count_severe())
1617 vm_page_try_to_cache(m);
1619 vm_page_unlock_queues();
1624 * adjust pip. NOTE: the original parent may still have its own
1625 * pip refs on the object.
1627 if (object != NULL) {
1628 vm_object_pip_wakeupn(object, bp->b_npages);
1629 VM_OBJECT_UNLOCK(object);
1633 * swapdev_strategy() manually sets b_vp and b_bufobj before calling
1634 * bstrategy(). Set them back to NULL now we're done with it, or we'll
1635 * trigger a KASSERT in relpbuf().
1639 bp->b_bufobj = NULL;
1642 * release the physical I/O buffer
1646 ((bp->b_iocmd == BIO_READ) ? &nsw_rcount :
1647 ((bp->b_flags & B_ASYNC) ?
1656 * swap_pager_isswapped:
1658 * Return 1 if at least one page in the given object is paged
1659 * out to the given swap device.
1661 * This routine may not block.
1664 swap_pager_isswapped(vm_object_t object, struct swdevt *sp)
1670 VM_OBJECT_LOCK_ASSERT(object, MA_OWNED);
1671 if (object->type != OBJT_SWAP)
1674 mtx_lock(&swhash_mtx);
1675 for (bcount = 0; bcount < object->un_pager.swp.swp_bcount; bcount++) {
1676 struct swblock *swap;
1678 if ((swap = *swp_pager_hash(object, index)) != NULL) {
1679 for (i = 0; i < SWAP_META_PAGES; ++i) {
1680 if (swp_pager_isondev(swap->swb_pages[i], sp)) {
1681 mtx_unlock(&swhash_mtx);
1686 index += SWAP_META_PAGES;
1687 if (index > 0x20000000)
1688 panic("swap_pager_isswapped: failed to locate all swap meta blocks");
1690 mtx_unlock(&swhash_mtx);
1695 * SWP_PAGER_FORCE_PAGEIN() - force a swap block to be paged in
1697 * This routine dissociates the page at the given index within a
1698 * swap block from its backing store, paging it in if necessary.
1699 * If the page is paged in, it is placed in the inactive queue,
1700 * since it had its backing store ripped out from under it.
1701 * We also attempt to swap in all other pages in the swap block,
1702 * we only guarantee that the one at the specified index is
1705 * XXX - The code to page the whole block in doesn't work, so we
1706 * revert to the one-by-one behavior for now. Sigh.
1709 swp_pager_force_pagein(vm_object_t object, vm_pindex_t pindex)
1713 vm_object_pip_add(object, 1);
1714 m = vm_page_grab(object, pindex, VM_ALLOC_NORMAL|VM_ALLOC_RETRY);
1715 if (m->valid == VM_PAGE_BITS_ALL) {
1716 vm_object_pip_subtract(object, 1);
1718 vm_page_lock_queues();
1719 vm_page_activate(m);
1721 vm_page_unlock_queues();
1724 vm_pager_page_unswapped(m);
1728 if (swap_pager_getpages(object, &m, 1, 0) != VM_PAGER_OK)
1729 panic("swap_pager_force_pagein: read from swap failed");/*XXX*/
1730 vm_object_pip_subtract(object, 1);
1732 vm_page_lock_queues();
1734 vm_page_dontneed(m);
1735 vm_page_unlock_queues();
1738 vm_pager_page_unswapped(m);
1742 * swap_pager_swapoff:
1744 * Page in all of the pages that have been paged out to the
1745 * given device. The corresponding blocks in the bitmap must be
1746 * marked as allocated and the device must be flagged SW_CLOSING.
1747 * There may be no processes swapped out to the device.
1749 * This routine may block.
1752 swap_pager_swapoff(struct swdevt *sp)
1754 struct swblock *swap;
1761 mtx_lock(&swhash_mtx);
1762 for (i = 0; i <= swhash_mask; i++) { /* '<=' is correct here */
1764 for (swap = swhash[i]; swap != NULL; swap = swap->swb_hnext) {
1765 vm_object_t object = swap->swb_object;
1766 vm_pindex_t pindex = swap->swb_index;
1767 for (j = 0; j < SWAP_META_PAGES; ++j) {
1768 if (swp_pager_isondev(swap->swb_pages[j], sp)) {
1769 /* avoid deadlock */
1770 if (!VM_OBJECT_TRYLOCK(object)) {
1773 mtx_unlock(&swhash_mtx);
1774 swp_pager_force_pagein(object,
1776 VM_OBJECT_UNLOCK(object);
1777 mtx_lock(&swhash_mtx);
1784 mtx_unlock(&swhash_mtx);
1787 * Objects may be locked or paging to the device being
1788 * removed, so we will miss their pages and need to
1789 * make another pass. We have marked this device as
1790 * SW_CLOSING, so the activity should finish soon.
1793 if (retries > 100) {
1794 panic("swapoff: failed to locate %d swap blocks",
1797 pause("swpoff", hz / 20);
1802 /************************************************************************
1804 ************************************************************************
1806 * These routines manipulate the swap metadata stored in the
1807 * OBJT_SWAP object. All swp_*() routines must be called at
1808 * splvm() because swap can be freed up by the low level vm_page
1809 * code which might be called from interrupts beyond what splbio() covers.
1811 * Swap metadata is implemented with a global hash and not directly
1812 * linked into the object. Instead the object simply contains
1813 * appropriate tracking counters.
1817 * SWP_PAGER_META_BUILD() - add swap block to swap meta data for object
1819 * We first convert the object to a swap object if it is a default
1822 * The specified swapblk is added to the object's swap metadata. If
1823 * the swapblk is not valid, it is freed instead. Any previously
1824 * assigned swapblk is freed.
1826 * This routine must be called at splvm(), except when used to convert
1827 * an OBJT_DEFAULT object into an OBJT_SWAP object.
1830 swp_pager_meta_build(vm_object_t object, vm_pindex_t pindex, daddr_t swapblk)
1832 struct swblock *swap;
1833 struct swblock **pswap;
1836 VM_OBJECT_LOCK_ASSERT(object, MA_OWNED);
1838 * Convert default object to swap object if necessary
1840 if (object->type != OBJT_SWAP) {
1841 object->type = OBJT_SWAP;
1842 object->un_pager.swp.swp_bcount = 0;
1844 if (object->handle != NULL) {
1845 mtx_lock(&sw_alloc_mtx);
1847 NOBJLIST(object->handle),
1851 mtx_unlock(&sw_alloc_mtx);
1856 * Locate hash entry. If not found create, but if we aren't adding
1857 * anything just return. If we run out of space in the map we wait
1858 * and, since the hash table may have changed, retry.
1861 mtx_lock(&swhash_mtx);
1862 pswap = swp_pager_hash(object, pindex);
1864 if ((swap = *pswap) == NULL) {
1867 if (swapblk == SWAPBLK_NONE)
1870 swap = *pswap = uma_zalloc(swap_zone, M_NOWAIT);
1872 mtx_unlock(&swhash_mtx);
1873 VM_OBJECT_UNLOCK(object);
1874 if (uma_zone_exhausted(swap_zone)) {
1875 printf("swap zone exhausted, increase kern.maxswzone\n");
1876 vm_pageout_oom(VM_OOM_SWAPZ);
1877 pause("swzonex", 10);
1880 VM_OBJECT_LOCK(object);
1884 swap->swb_hnext = NULL;
1885 swap->swb_object = object;
1886 swap->swb_index = pindex & ~(vm_pindex_t)SWAP_META_MASK;
1887 swap->swb_count = 0;
1889 ++object->un_pager.swp.swp_bcount;
1891 for (i = 0; i < SWAP_META_PAGES; ++i)
1892 swap->swb_pages[i] = SWAPBLK_NONE;
1896 * Delete prior contents of metadata
1898 idx = pindex & SWAP_META_MASK;
1900 if (swap->swb_pages[idx] != SWAPBLK_NONE) {
1901 swp_pager_freeswapspace(swap->swb_pages[idx], 1);
1906 * Enter block into metadata
1908 swap->swb_pages[idx] = swapblk;
1909 if (swapblk != SWAPBLK_NONE)
1912 mtx_unlock(&swhash_mtx);
1916 * SWP_PAGER_META_FREE() - free a range of blocks in the object's swap metadata
1918 * The requested range of blocks is freed, with any associated swap
1919 * returned to the swap bitmap.
1921 * This routine will free swap metadata structures as they are cleaned
1922 * out. This routine does *NOT* operate on swap metadata associated
1923 * with resident pages.
1925 * This routine must be called at splvm()
1928 swp_pager_meta_free(vm_object_t object, vm_pindex_t index, daddr_t count)
1931 VM_OBJECT_LOCK_ASSERT(object, MA_OWNED);
1932 if (object->type != OBJT_SWAP)
1936 struct swblock **pswap;
1937 struct swblock *swap;
1939 mtx_lock(&swhash_mtx);
1940 pswap = swp_pager_hash(object, index);
1942 if ((swap = *pswap) != NULL) {
1943 daddr_t v = swap->swb_pages[index & SWAP_META_MASK];
1945 if (v != SWAPBLK_NONE) {
1946 swp_pager_freeswapspace(v, 1);
1947 swap->swb_pages[index & SWAP_META_MASK] =
1949 if (--swap->swb_count == 0) {
1950 *pswap = swap->swb_hnext;
1951 uma_zfree(swap_zone, swap);
1952 --object->un_pager.swp.swp_bcount;
1958 int n = SWAP_META_PAGES - (index & SWAP_META_MASK);
1962 mtx_unlock(&swhash_mtx);
1967 * SWP_PAGER_META_FREE_ALL() - destroy all swap metadata associated with object
1969 * This routine locates and destroys all swap metadata associated with
1972 * This routine must be called at splvm()
1975 swp_pager_meta_free_all(vm_object_t object)
1979 VM_OBJECT_LOCK_ASSERT(object, MA_OWNED);
1980 if (object->type != OBJT_SWAP)
1983 while (object->un_pager.swp.swp_bcount) {
1984 struct swblock **pswap;
1985 struct swblock *swap;
1987 mtx_lock(&swhash_mtx);
1988 pswap = swp_pager_hash(object, index);
1989 if ((swap = *pswap) != NULL) {
1992 for (i = 0; i < SWAP_META_PAGES; ++i) {
1993 daddr_t v = swap->swb_pages[i];
1994 if (v != SWAPBLK_NONE) {
1996 swp_pager_freeswapspace(v, 1);
1999 if (swap->swb_count != 0)
2000 panic("swap_pager_meta_free_all: swb_count != 0");
2001 *pswap = swap->swb_hnext;
2002 uma_zfree(swap_zone, swap);
2003 --object->un_pager.swp.swp_bcount;
2005 mtx_unlock(&swhash_mtx);
2006 index += SWAP_META_PAGES;
2007 if (index > 0x20000000)
2008 panic("swp_pager_meta_free_all: failed to locate all swap meta blocks");
2013 * SWP_PAGER_METACTL() - misc control of swap and vm_page_t meta data.
2015 * This routine is capable of looking up, popping, or freeing
2016 * swapblk assignments in the swap meta data or in the vm_page_t.
2017 * The routine typically returns the swapblk being looked-up, or popped,
2018 * or SWAPBLK_NONE if the block was freed, or SWAPBLK_NONE if the block
2019 * was invalid. This routine will automatically free any invalid
2020 * meta-data swapblks.
2022 * It is not possible to store invalid swapblks in the swap meta data
2023 * (other then a literal 'SWAPBLK_NONE'), so we don't bother checking.
2025 * When acting on a busy resident page and paging is in progress, we
2026 * have to wait until paging is complete but otherwise can act on the
2029 * This routine must be called at splvm().
2031 * SWM_FREE remove and free swap block from metadata
2032 * SWM_POP remove from meta data but do not free.. pop it out
2035 swp_pager_meta_ctl(vm_object_t object, vm_pindex_t pindex, int flags)
2037 struct swblock **pswap;
2038 struct swblock *swap;
2042 VM_OBJECT_LOCK_ASSERT(object, MA_OWNED);
2044 * The meta data only exists of the object is OBJT_SWAP
2045 * and even then might not be allocated yet.
2047 if (object->type != OBJT_SWAP)
2048 return (SWAPBLK_NONE);
2051 mtx_lock(&swhash_mtx);
2052 pswap = swp_pager_hash(object, pindex);
2054 if ((swap = *pswap) != NULL) {
2055 idx = pindex & SWAP_META_MASK;
2056 r1 = swap->swb_pages[idx];
2058 if (r1 != SWAPBLK_NONE) {
2059 if (flags & SWM_FREE) {
2060 swp_pager_freeswapspace(r1, 1);
2063 if (flags & (SWM_FREE|SWM_POP)) {
2064 swap->swb_pages[idx] = SWAPBLK_NONE;
2065 if (--swap->swb_count == 0) {
2066 *pswap = swap->swb_hnext;
2067 uma_zfree(swap_zone, swap);
2068 --object->un_pager.swp.swp_bcount;
2073 mtx_unlock(&swhash_mtx);
2078 * System call swapon(name) enables swapping on device name,
2079 * which must be in the swdevsw. Return EBUSY
2080 * if already swapping on this device.
2082 #ifndef _SYS_SYSPROTO_H_
2083 struct swapon_args {
2093 swapon(struct thread *td, struct swapon_args *uap)
2097 struct nameidata nd;
2100 error = priv_check(td, PRIV_SWAPON);
2105 while (swdev_syscall_active)
2106 tsleep(&swdev_syscall_active, PUSER - 1, "swpon", 0);
2107 swdev_syscall_active = 1;
2110 * Swap metadata may not fit in the KVM if we have physical
2113 if (swap_zone == NULL) {
2118 NDINIT(&nd, LOOKUP, ISOPEN | FOLLOW | AUDITVNODE1, UIO_USERSPACE,
2124 NDFREE(&nd, NDF_ONLY_PNBUF);
2127 if (vn_isdisk(vp, &error)) {
2128 error = swapongeom(td, vp);
2129 } else if (vp->v_type == VREG &&
2130 (vp->v_mount->mnt_vfc->vfc_flags & VFCF_NETWORK) != 0 &&
2131 (error = VOP_GETATTR(vp, &attr, td->td_ucred)) == 0) {
2133 * Allow direct swapping to NFS regular files in the same
2134 * way that nfs_mountroot() sets up diskless swapping.
2136 error = swaponvp(td, vp, attr.va_size / DEV_BSIZE);
2142 swdev_syscall_active = 0;
2143 wakeup_one(&swdev_syscall_active);
2149 swaponsomething(struct vnode *vp, void *id, u_long nblks, sw_strategy_t *strategy, sw_close_t *close, dev_t dev)
2151 struct swdevt *sp, *tsp;
2156 * If we go beyond this, we get overflows in the radix
2159 mblocks = 0x40000000 / BLIST_META_RADIX;
2160 if (nblks > mblocks) {
2161 printf("WARNING: reducing size to maximum of %lu blocks per swap unit\n",
2166 * nblks is in DEV_BSIZE'd chunks, convert to PAGE_SIZE'd chunks.
2167 * First chop nblks off to page-align it, then convert.
2169 * sw->sw_nblks is in page-sized chunks now too.
2171 nblks &= ~(ctodb(1) - 1);
2172 nblks = dbtoc(nblks);
2174 sp = malloc(sizeof *sp, M_VMPGDATA, M_WAITOK | M_ZERO);
2179 sp->sw_nblks = nblks;
2181 sp->sw_strategy = strategy;
2182 sp->sw_close = close;
2184 sp->sw_blist = blist_create(nblks, M_WAITOK);
2186 * Do not free the first two block in order to avoid overwriting
2187 * any bsd label at the front of the partition
2189 blist_free(sp->sw_blist, 2, nblks - 2);
2192 mtx_lock(&sw_dev_mtx);
2193 TAILQ_FOREACH(tsp, &swtailq, sw_list) {
2194 if (tsp->sw_end >= dvbase) {
2196 * We put one uncovered page between the devices
2197 * in order to definitively prevent any cross-device
2200 dvbase = tsp->sw_end + 1;
2203 sp->sw_first = dvbase;
2204 sp->sw_end = dvbase + nblks;
2205 TAILQ_INSERT_TAIL(&swtailq, sp, sw_list);
2207 swap_pager_avail += nblks;
2208 swap_total += (vm_ooffset_t)nblks * PAGE_SIZE;
2210 mtx_unlock(&sw_dev_mtx);
2214 * SYSCALL: swapoff(devname)
2216 * Disable swapping on the given device.
2218 * XXX: Badly designed system call: it should use a device index
2219 * rather than filename as specification. We keep sw_vp around
2220 * only to make this work.
2222 #ifndef _SYS_SYSPROTO_H_
2223 struct swapoff_args {
2233 swapoff(struct thread *td, struct swapoff_args *uap)
2236 struct nameidata nd;
2240 error = priv_check(td, PRIV_SWAPOFF);
2245 while (swdev_syscall_active)
2246 tsleep(&swdev_syscall_active, PUSER - 1, "swpoff", 0);
2247 swdev_syscall_active = 1;
2249 NDINIT(&nd, LOOKUP, FOLLOW | AUDITVNODE1, UIO_USERSPACE, uap->name,
2254 NDFREE(&nd, NDF_ONLY_PNBUF);
2257 mtx_lock(&sw_dev_mtx);
2258 TAILQ_FOREACH(sp, &swtailq, sw_list) {
2259 if (sp->sw_vp == vp)
2262 mtx_unlock(&sw_dev_mtx);
2267 error = swapoff_one(sp, td->td_ucred);
2269 swdev_syscall_active = 0;
2270 wakeup_one(&swdev_syscall_active);
2276 swapoff_one(struct swdevt *sp, struct ucred *cred)
2278 u_long nblks, dvbase;
2283 mtx_assert(&Giant, MA_OWNED);
2285 (void) vn_lock(sp->sw_vp, LK_EXCLUSIVE | LK_RETRY);
2286 error = mac_system_check_swapoff(cred, sp->sw_vp);
2287 (void) VOP_UNLOCK(sp->sw_vp, 0);
2291 nblks = sp->sw_nblks;
2294 * We can turn off this swap device safely only if the
2295 * available virtual memory in the system will fit the amount
2296 * of data we will have to page back in, plus an epsilon so
2297 * the system doesn't become critically low on swap space.
2299 if (cnt.v_free_count + cnt.v_cache_count + swap_pager_avail <
2300 nblks + nswap_lowat) {
2305 * Prevent further allocations on this device.
2307 mtx_lock(&sw_dev_mtx);
2308 sp->sw_flags |= SW_CLOSING;
2309 for (dvbase = 0; dvbase < sp->sw_end; dvbase += dmmax) {
2310 swap_pager_avail -= blist_fill(sp->sw_blist,
2313 swap_total -= (vm_ooffset_t)nblks * PAGE_SIZE;
2314 mtx_unlock(&sw_dev_mtx);
2317 * Page in the contents of the device and close it.
2319 swap_pager_swapoff(sp);
2321 sp->sw_close(curthread, sp);
2323 mtx_lock(&sw_dev_mtx);
2324 TAILQ_REMOVE(&swtailq, sp, sw_list);
2326 if (nswapdev == 0) {
2327 swap_pager_full = 2;
2328 swap_pager_almost_full = 1;
2332 mtx_unlock(&sw_dev_mtx);
2333 blist_destroy(sp->sw_blist);
2334 free(sp, M_VMPGDATA);
2341 struct swdevt *sp, *spt;
2342 const char *devname;
2346 while (swdev_syscall_active)
2347 tsleep(&swdev_syscall_active, PUSER - 1, "swpoff", 0);
2348 swdev_syscall_active = 1;
2350 mtx_lock(&sw_dev_mtx);
2351 TAILQ_FOREACH_SAFE(sp, &swtailq, sw_list, spt) {
2352 mtx_unlock(&sw_dev_mtx);
2353 if (vn_isdisk(sp->sw_vp, NULL))
2354 devname = sp->sw_vp->v_rdev->si_name;
2357 error = swapoff_one(sp, thread0.td_ucred);
2359 printf("Cannot remove swap device %s (error=%d), "
2360 "skipping.\n", devname, error);
2361 } else if (bootverbose) {
2362 printf("Swap device %s removed.\n", devname);
2364 mtx_lock(&sw_dev_mtx);
2366 mtx_unlock(&sw_dev_mtx);
2368 swdev_syscall_active = 0;
2369 wakeup_one(&swdev_syscall_active);
2374 swap_pager_status(int *total, int *used)
2380 mtx_lock(&sw_dev_mtx);
2381 TAILQ_FOREACH(sp, &swtailq, sw_list) {
2382 *total += sp->sw_nblks;
2383 *used += sp->sw_used;
2385 mtx_unlock(&sw_dev_mtx);
2389 sysctl_vm_swap_info(SYSCTL_HANDLER_ARGS)
2391 int *name = (int *)arg1;
2396 if (arg2 != 1) /* name length */
2400 mtx_lock(&sw_dev_mtx);
2401 TAILQ_FOREACH(sp, &swtailq, sw_list) {
2403 mtx_unlock(&sw_dev_mtx);
2404 xs.xsw_version = XSWDEV_VERSION;
2405 xs.xsw_dev = sp->sw_dev;
2406 xs.xsw_flags = sp->sw_flags;
2407 xs.xsw_nblks = sp->sw_nblks;
2408 xs.xsw_used = sp->sw_used;
2410 error = SYSCTL_OUT(req, &xs, sizeof(xs));
2415 mtx_unlock(&sw_dev_mtx);
2419 SYSCTL_INT(_vm, OID_AUTO, nswapdev, CTLFLAG_RD, &nswapdev, 0,
2420 "Number of swap devices");
2421 SYSCTL_NODE(_vm, OID_AUTO, swap_info, CTLFLAG_RD, sysctl_vm_swap_info,
2422 "Swap statistics by device");
2425 * vmspace_swap_count() - count the approximate swap usage in pages for a
2428 * The map must be locked.
2430 * Swap usage is determined by taking the proportional swap used by
2431 * VM objects backing the VM map. To make up for fractional losses,
2432 * if the VM object has any swap use at all the associated map entries
2433 * count for at least 1 swap page.
2436 vmspace_swap_count(struct vmspace *vmspace)
2438 vm_map_t map = &vmspace->vm_map;
2442 for (cur = map->header.next; cur != &map->header; cur = cur->next) {
2445 if ((cur->eflags & MAP_ENTRY_IS_SUB_MAP) == 0 &&
2446 (object = cur->object.vm_object) != NULL) {
2447 VM_OBJECT_LOCK(object);
2448 if (object->type == OBJT_SWAP &&
2449 object->un_pager.swp.swp_bcount != 0) {
2450 int n = (cur->end - cur->start) / PAGE_SIZE;
2452 count += object->un_pager.swp.swp_bcount *
2453 SWAP_META_PAGES * n / object->size + 1;
2455 VM_OBJECT_UNLOCK(object);
2464 * Swapping onto disk devices.
2468 static g_orphan_t swapgeom_orphan;
2470 static struct g_class g_swap_class = {
2472 .version = G_VERSION,
2473 .orphan = swapgeom_orphan,
2476 DECLARE_GEOM_CLASS(g_swap_class, g_class);
2480 swapgeom_done(struct bio *bp2)
2484 bp = bp2->bio_caller2;
2485 bp->b_ioflags = bp2->bio_flags;
2487 bp->b_ioflags |= BIO_ERROR;
2488 bp->b_resid = bp->b_bcount - bp2->bio_completed;
2489 bp->b_error = bp2->bio_error;
2495 swapgeom_strategy(struct buf *bp, struct swdevt *sp)
2498 struct g_consumer *cp;
2502 bp->b_error = ENXIO;
2503 bp->b_ioflags |= BIO_ERROR;
2507 if (bp->b_iocmd == BIO_WRITE)
2510 bio = g_alloc_bio();
2512 bp->b_error = ENOMEM;
2513 bp->b_ioflags |= BIO_ERROR;
2518 bio->bio_caller2 = bp;
2519 bio->bio_cmd = bp->b_iocmd;
2520 bio->bio_data = bp->b_data;
2521 bio->bio_offset = (bp->b_blkno - sp->sw_first) * PAGE_SIZE;
2522 bio->bio_length = bp->b_bcount;
2523 bio->bio_done = swapgeom_done;
2524 g_io_request(bio, cp);
2529 swapgeom_orphan(struct g_consumer *cp)
2533 mtx_lock(&sw_dev_mtx);
2534 TAILQ_FOREACH(sp, &swtailq, sw_list)
2535 if (sp->sw_id == cp)
2537 mtx_unlock(&sw_dev_mtx);
2541 swapgeom_close_ev(void *arg, int flags)
2543 struct g_consumer *cp;
2546 g_access(cp, -1, -1, 0);
2548 g_destroy_consumer(cp);
2552 swapgeom_close(struct thread *td, struct swdevt *sw)
2555 /* XXX: direct call when Giant untangled */
2556 g_waitfor_event(swapgeom_close_ev, sw->sw_id, M_WAITOK, NULL);
2567 swapongeom_ev(void *arg, int flags)
2570 struct g_provider *pp;
2571 struct g_consumer *cp;
2572 static struct g_geom *gp;
2579 pp = g_dev_getprovider(swh->dev);
2581 swh->error = ENODEV;
2584 mtx_lock(&sw_dev_mtx);
2585 TAILQ_FOREACH(sp, &swtailq, sw_list) {
2587 if (cp != NULL && cp->provider == pp) {
2588 mtx_unlock(&sw_dev_mtx);
2593 mtx_unlock(&sw_dev_mtx);
2595 gp = g_new_geomf(&g_swap_class, "swap", NULL);
2596 cp = g_new_consumer(gp);
2599 * XXX: Everytime you think you can improve the margin for
2600 * footshooting, somebody depends on the ability to do so:
2601 * savecore(8) wants to write to our swapdev so we cannot
2602 * set an exclusive count :-(
2604 error = g_access(cp, 1, 1, 0);
2607 g_destroy_consumer(cp);
2611 nblks = pp->mediasize / DEV_BSIZE;
2612 swaponsomething(swh->vp, cp, nblks, swapgeom_strategy,
2613 swapgeom_close, dev2udev(swh->dev));
2619 swapongeom(struct thread *td, struct vnode *vp)
2624 vn_lock(vp, LK_EXCLUSIVE | LK_RETRY);
2626 swh.dev = vp->v_rdev;
2629 /* XXX: direct call when Giant untangled */
2630 error = g_waitfor_event(swapongeom_ev, &swh, M_WAITOK, NULL);
2640 * This is used mainly for network filesystem (read: probably only tested
2641 * with NFS) swapfiles.
2646 swapdev_strategy(struct buf *bp, struct swdevt *sp)
2650 bp->b_blkno = ctodb(bp->b_blkno - sp->sw_first);
2654 if (bp->b_iocmd == BIO_WRITE) {
2656 bufobj_wdrop(bp->b_bufobj);
2657 bufobj_wref(&vp2->v_bufobj);
2659 if (bp->b_bufobj != &vp2->v_bufobj)
2660 bp->b_bufobj = &vp2->v_bufobj;
2662 bp->b_iooffset = dbtob(bp->b_blkno);
2668 swapdev_close(struct thread *td, struct swdevt *sp)
2671 VOP_CLOSE(sp->sw_vp, FREAD | FWRITE, td->td_ucred, td);
2677 swaponvp(struct thread *td, struct vnode *vp, u_long nblks)
2684 mtx_lock(&sw_dev_mtx);
2685 TAILQ_FOREACH(sp, &swtailq, sw_list) {
2686 if (sp->sw_id == vp) {
2687 mtx_unlock(&sw_dev_mtx);
2691 mtx_unlock(&sw_dev_mtx);
2693 (void) vn_lock(vp, LK_EXCLUSIVE | LK_RETRY);
2695 error = mac_system_check_swapon(td->td_ucred, vp);
2698 error = VOP_OPEN(vp, FREAD | FWRITE, td->td_ucred, td, NULL);
2699 (void) VOP_UNLOCK(vp, 0);
2703 swaponsomething(vp, vp, nblks, swapdev_strategy, swapdev_close,