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 static vm_ooffset_t swap_reserved;
160 SYSCTL_QUAD(_vm, OID_AUTO, swap_reserved, CTLFLAG_RD, &swap_reserved, 0, "");
161 static int overcommit = 0;
162 SYSCTL_INT(_vm, OID_AUTO, overcommit, CTLFLAG_RW, &overcommit, 0, "");
164 /* bits from overcommit */
165 #define SWAP_RESERVE_FORCE_ON (1 << 0)
166 #define SWAP_RESERVE_RLIMIT_ON (1 << 1)
167 #define SWAP_RESERVE_ALLOW_NONWIRED (1 << 2)
170 swap_reserve(vm_ooffset_t incr)
173 return (swap_reserve_by_uid(incr, curthread->td_ucred->cr_ruidinfo));
177 swap_reserve_by_uid(vm_ooffset_t incr, struct uidinfo *uip)
182 static struct timeval lastfail;
184 if (incr & PAGE_MASK)
185 panic("swap_reserve: & PAGE_MASK");
188 mtx_lock(&sw_dev_mtx);
189 r = swap_reserved + incr;
190 if (overcommit & SWAP_RESERVE_ALLOW_NONWIRED) {
191 s = cnt.v_page_count - cnt.v_free_reserved - cnt.v_wire_count;
196 if ((overcommit & SWAP_RESERVE_FORCE_ON) == 0 || r <= s ||
197 (error = priv_check(curthread, PRIV_VM_SWAP_NOQUOTA)) == 0) {
201 mtx_unlock(&sw_dev_mtx);
205 UIDINFO_VMSIZE_LOCK(uip);
206 if ((overcommit & SWAP_RESERVE_RLIMIT_ON) != 0 &&
207 uip->ui_vmsize + incr > lim_cur(curproc, RLIMIT_SWAP) &&
208 priv_check(curthread, PRIV_VM_SWAP_NORLIMIT))
211 uip->ui_vmsize += incr;
212 UIDINFO_VMSIZE_UNLOCK(uip);
213 PROC_UNLOCK(curproc);
215 mtx_lock(&sw_dev_mtx);
216 swap_reserved -= incr;
217 mtx_unlock(&sw_dev_mtx);
220 if (!res && ppsratecheck(&lastfail, &curfail, 1)) {
221 printf("uid %d, pid %d: swap reservation for %jd bytes failed\n",
222 curproc->p_pid, uip->ui_uid, incr);
229 swap_reserve_force(vm_ooffset_t incr)
233 mtx_lock(&sw_dev_mtx);
234 swap_reserved += incr;
235 mtx_unlock(&sw_dev_mtx);
237 uip = curthread->td_ucred->cr_ruidinfo;
239 UIDINFO_VMSIZE_LOCK(uip);
240 uip->ui_vmsize += incr;
241 UIDINFO_VMSIZE_UNLOCK(uip);
242 PROC_UNLOCK(curproc);
246 swap_release(vm_ooffset_t decr)
251 uip = curthread->td_ucred->cr_ruidinfo;
252 swap_release_by_uid(decr, uip);
253 PROC_UNLOCK(curproc);
257 swap_release_by_uid(vm_ooffset_t decr, struct uidinfo *uip)
260 if (decr & PAGE_MASK)
261 panic("swap_release: & PAGE_MASK");
263 mtx_lock(&sw_dev_mtx);
264 if (swap_reserved < decr)
265 panic("swap_reserved < decr");
266 swap_reserved -= decr;
267 mtx_unlock(&sw_dev_mtx);
269 UIDINFO_VMSIZE_LOCK(uip);
270 if (uip->ui_vmsize < decr)
271 printf("negative vmsize for uid = %d\n", uip->ui_uid);
272 uip->ui_vmsize -= decr;
273 UIDINFO_VMSIZE_UNLOCK(uip);
276 static void swapdev_strategy(struct buf *, struct swdevt *sw);
278 #define SWM_FREE 0x02 /* free, period */
279 #define SWM_POP 0x04 /* pop out */
281 int swap_pager_full = 2; /* swap space exhaustion (task killing) */
282 static int swap_pager_almost_full = 1; /* swap space exhaustion (w/hysteresis)*/
283 static int nsw_rcount; /* free read buffers */
284 static int nsw_wcount_sync; /* limit write buffers / synchronous */
285 static int nsw_wcount_async; /* limit write buffers / asynchronous */
286 static int nsw_wcount_async_max;/* assigned maximum */
287 static int nsw_cluster_max; /* maximum VOP I/O allowed */
289 static struct swblock **swhash;
290 static int swhash_mask;
291 static struct mtx swhash_mtx;
293 static int swap_async_max = 4; /* maximum in-progress async I/O's */
294 static struct sx sw_alloc_sx;
297 SYSCTL_INT(_vm, OID_AUTO, swap_async_max,
298 CTLFLAG_RW, &swap_async_max, 0, "Maximum running async swap ops");
301 * "named" and "unnamed" anon region objects. Try to reduce the overhead
302 * of searching a named list by hashing it just a little.
307 #define NOBJLIST(handle) \
308 (&swap_pager_object_list[((int)(intptr_t)handle >> 4) & (NOBJLISTS-1)])
310 static struct mtx sw_alloc_mtx; /* protect list manipulation */
311 static struct pagerlst swap_pager_object_list[NOBJLISTS];
312 static uma_zone_t swap_zone;
313 static struct vm_object swap_zone_obj;
316 * pagerops for OBJT_SWAP - "swap pager". Some ops are also global procedure
317 * calls hooked from other parts of the VM system and do not appear here.
318 * (see vm/swap_pager.h).
321 swap_pager_alloc(void *handle, vm_ooffset_t size,
322 vm_prot_t prot, vm_ooffset_t offset, struct ucred *);
323 static void swap_pager_dealloc(vm_object_t object);
324 static int swap_pager_getpages(vm_object_t, vm_page_t *, int, int);
325 static void swap_pager_putpages(vm_object_t, vm_page_t *, int, boolean_t, int *);
327 swap_pager_haspage(vm_object_t object, vm_pindex_t pindex, int *before, int *after);
328 static void swap_pager_init(void);
329 static void swap_pager_unswapped(vm_page_t);
330 static void swap_pager_swapoff(struct swdevt *sp);
332 struct pagerops swappagerops = {
333 .pgo_init = swap_pager_init, /* early system initialization of pager */
334 .pgo_alloc = swap_pager_alloc, /* allocate an OBJT_SWAP object */
335 .pgo_dealloc = swap_pager_dealloc, /* deallocate an OBJT_SWAP object */
336 .pgo_getpages = swap_pager_getpages, /* pagein */
337 .pgo_putpages = swap_pager_putpages, /* pageout */
338 .pgo_haspage = swap_pager_haspage, /* get backing store status for page */
339 .pgo_pageunswapped = swap_pager_unswapped, /* remove swap related to page */
343 * dmmax is in page-sized chunks with the new swap system. It was
344 * dev-bsized chunks in the old. dmmax is always a power of 2.
346 * swap_*() routines are externally accessible. swp_*() routines are
350 static int nswap_lowat = 128; /* in pages, swap_pager_almost_full warn */
351 static int nswap_hiwat = 512; /* in pages, swap_pager_almost_full warn */
353 SYSCTL_INT(_vm, OID_AUTO, dmmax,
354 CTLFLAG_RD, &dmmax, 0, "Maximum size of a swap block");
356 static void swp_sizecheck(void);
357 static void swp_pager_async_iodone(struct buf *bp);
358 static int swapongeom(struct thread *, struct vnode *);
359 static int swaponvp(struct thread *, struct vnode *, u_long);
360 static int swapoff_one(struct swdevt *sp, struct ucred *cred);
363 * Swap bitmap functions
365 static void swp_pager_freeswapspace(daddr_t blk, int npages);
366 static daddr_t swp_pager_getswapspace(int npages);
371 static struct swblock **swp_pager_hash(vm_object_t object, vm_pindex_t index);
372 static void swp_pager_meta_build(vm_object_t, vm_pindex_t, daddr_t);
373 static void swp_pager_meta_free(vm_object_t, vm_pindex_t, daddr_t);
374 static void swp_pager_meta_free_all(vm_object_t);
375 static daddr_t swp_pager_meta_ctl(vm_object_t, vm_pindex_t, int);
378 swp_pager_free_nrpage(vm_page_t m)
381 if (m->wire_count == 0)
386 * SWP_SIZECHECK() - update swap_pager_full indication
388 * update the swap_pager_almost_full indication and warn when we are
389 * about to run out of swap space, using lowat/hiwat hysteresis.
391 * Clear swap_pager_full ( task killing ) indication when lowat is met.
393 * No restrictions on call
394 * This routine may not block.
395 * This routine must be called at splvm()
401 if (swap_pager_avail < nswap_lowat) {
402 if (swap_pager_almost_full == 0) {
403 printf("swap_pager: out of swap space\n");
404 swap_pager_almost_full = 1;
408 if (swap_pager_avail > nswap_hiwat)
409 swap_pager_almost_full = 0;
414 * SWP_PAGER_HASH() - hash swap meta data
416 * This is an helper function which hashes the swapblk given
417 * the object and page index. It returns a pointer to a pointer
418 * to the object, or a pointer to a NULL pointer if it could not
421 * This routine must be called at splvm().
423 static struct swblock **
424 swp_pager_hash(vm_object_t object, vm_pindex_t index)
426 struct swblock **pswap;
427 struct swblock *swap;
429 index &= ~(vm_pindex_t)SWAP_META_MASK;
430 pswap = &swhash[(index ^ (int)(intptr_t)object) & swhash_mask];
431 while ((swap = *pswap) != NULL) {
432 if (swap->swb_object == object &&
433 swap->swb_index == index
437 pswap = &swap->swb_hnext;
443 * SWAP_PAGER_INIT() - initialize the swap pager!
445 * Expected to be started from system init. NOTE: This code is run
446 * before much else so be careful what you depend on. Most of the VM
447 * system has yet to be initialized at this point.
450 swap_pager_init(void)
453 * Initialize object lists
457 for (i = 0; i < NOBJLISTS; ++i)
458 TAILQ_INIT(&swap_pager_object_list[i]);
459 mtx_init(&sw_alloc_mtx, "swap_pager list", NULL, MTX_DEF);
460 mtx_init(&sw_dev_mtx, "swapdev", NULL, MTX_DEF);
463 * Device Stripe, in PAGE_SIZE'd blocks
465 dmmax = SWB_NPAGES * 2;
469 * SWAP_PAGER_SWAP_INIT() - swap pager initialization from pageout process
471 * Expected to be started from pageout process once, prior to entering
475 swap_pager_swap_init(void)
480 * Number of in-transit swap bp operations. Don't
481 * exhaust the pbufs completely. Make sure we
482 * initialize workable values (0 will work for hysteresis
483 * but it isn't very efficient).
485 * The nsw_cluster_max is constrained by the bp->b_pages[]
486 * array (MAXPHYS/PAGE_SIZE) and our locally defined
487 * MAX_PAGEOUT_CLUSTER. Also be aware that swap ops are
488 * constrained by the swap device interleave stripe size.
490 * Currently we hardwire nsw_wcount_async to 4. This limit is
491 * designed to prevent other I/O from having high latencies due to
492 * our pageout I/O. The value 4 works well for one or two active swap
493 * devices but is probably a little low if you have more. Even so,
494 * a higher value would probably generate only a limited improvement
495 * with three or four active swap devices since the system does not
496 * typically have to pageout at extreme bandwidths. We will want
497 * at least 2 per swap devices, and 4 is a pretty good value if you
498 * have one NFS swap device due to the command/ack latency over NFS.
499 * So it all works out pretty well.
501 nsw_cluster_max = min((MAXPHYS/PAGE_SIZE), MAX_PAGEOUT_CLUSTER);
504 nsw_rcount = (nswbuf + 1) / 2;
505 nsw_wcount_sync = (nswbuf + 3) / 4;
506 nsw_wcount_async = 4;
507 nsw_wcount_async_max = nsw_wcount_async;
508 mtx_unlock(&pbuf_mtx);
511 * Initialize our zone. Right now I'm just guessing on the number
512 * we need based on the number of pages in the system. Each swblock
513 * can hold 16 pages, so this is probably overkill. This reservation
514 * is typically limited to around 32MB by default.
516 n = cnt.v_page_count / 2;
517 if (maxswzone && n > maxswzone / sizeof(struct swblock))
518 n = maxswzone / sizeof(struct swblock);
520 swap_zone = uma_zcreate("SWAPMETA", sizeof(struct swblock), NULL, NULL,
521 NULL, NULL, UMA_ALIGN_PTR, UMA_ZONE_NOFREE | UMA_ZONE_VM);
522 if (swap_zone == NULL)
523 panic("failed to create swap_zone.");
525 if (uma_zone_set_obj(swap_zone, &swap_zone_obj, n))
528 * if the allocation failed, try a zone two thirds the
529 * size of the previous attempt.
534 printf("Swap zone entries reduced from %d to %d.\n", n2, n);
538 * Initialize our meta-data hash table. The swapper does not need to
539 * be quite as efficient as the VM system, so we do not use an
540 * oversized hash table.
542 * n: size of hash table, must be power of 2
543 * swhash_mask: hash table index mask
545 for (n = 1; n < n2 / 8; n *= 2)
547 swhash = malloc(sizeof(struct swblock *) * n, M_VMPGDATA, M_WAITOK | M_ZERO);
549 mtx_init(&swhash_mtx, "swap_pager swhash", NULL, MTX_DEF);
553 * SWAP_PAGER_ALLOC() - allocate a new OBJT_SWAP VM object and instantiate
554 * its metadata structures.
556 * This routine is called from the mmap and fork code to create a new
557 * OBJT_SWAP object. We do this by creating an OBJT_DEFAULT object
558 * and then converting it with swp_pager_meta_build().
560 * This routine may block in vm_object_allocate() and create a named
561 * object lookup race, so we must interlock. We must also run at
562 * splvm() for the object lookup to handle races with interrupts, but
563 * we do not have to maintain splvm() in between the lookup and the
564 * add because (I believe) it is not possible to attempt to create
565 * a new swap object w/handle when a default object with that handle
571 swap_pager_alloc(void *handle, vm_ooffset_t size, vm_prot_t prot,
572 vm_ooffset_t offset, struct ucred *cred)
579 pindex = OFF_TO_IDX(offset + PAGE_MASK + size);
583 * Reference existing named region or allocate new one. There
584 * should not be a race here against swp_pager_meta_build()
585 * as called from vm_page_remove() in regards to the lookup
588 sx_xlock(&sw_alloc_sx);
589 object = vm_pager_object_lookup(NOBJLIST(handle), handle);
590 if (object == NULL) {
592 uip = cred->cr_ruidinfo;
593 if (!swap_reserve_by_uid(size, uip)) {
594 sx_xunlock(&sw_alloc_sx);
600 object = vm_object_allocate(OBJT_DEFAULT, pindex);
601 VM_OBJECT_LOCK(object);
602 object->handle = handle;
605 object->charge = size;
607 swp_pager_meta_build(object, 0, SWAPBLK_NONE);
608 VM_OBJECT_UNLOCK(object);
610 sx_xunlock(&sw_alloc_sx);
614 uip = cred->cr_ruidinfo;
615 if (!swap_reserve_by_uid(size, uip))
619 object = vm_object_allocate(OBJT_DEFAULT, pindex);
620 VM_OBJECT_LOCK(object);
623 object->charge = size;
625 swp_pager_meta_build(object, 0, SWAPBLK_NONE);
626 VM_OBJECT_UNLOCK(object);
632 * SWAP_PAGER_DEALLOC() - remove swap metadata from object
634 * The swap backing for the object is destroyed. The code is
635 * designed such that we can reinstantiate it later, but this
636 * routine is typically called only when the entire object is
637 * about to be destroyed.
639 * This routine may block, but no longer does.
641 * The object must be locked or unreferenceable.
644 swap_pager_dealloc(vm_object_t object)
648 * Remove from list right away so lookups will fail if we block for
649 * pageout completion.
651 if (object->handle != NULL) {
652 mtx_lock(&sw_alloc_mtx);
653 TAILQ_REMOVE(NOBJLIST(object->handle), object, pager_object_list);
654 mtx_unlock(&sw_alloc_mtx);
657 VM_OBJECT_LOCK_ASSERT(object, MA_OWNED);
658 vm_object_pip_wait(object, "swpdea");
661 * Free all remaining metadata. We only bother to free it from
662 * the swap meta data. We do not attempt to free swapblk's still
663 * associated with vm_page_t's for this object. We do not care
664 * if paging is still in progress on some objects.
666 swp_pager_meta_free_all(object);
669 /************************************************************************
670 * SWAP PAGER BITMAP ROUTINES *
671 ************************************************************************/
674 * SWP_PAGER_GETSWAPSPACE() - allocate raw swap space
676 * Allocate swap for the requested number of pages. The starting
677 * swap block number (a page index) is returned or SWAPBLK_NONE
678 * if the allocation failed.
680 * Also has the side effect of advising that somebody made a mistake
681 * when they configured swap and didn't configure enough.
683 * Must be called at splvm() to avoid races with bitmap frees from
684 * vm_page_remove() aka swap_pager_page_removed().
686 * This routine may not block
687 * This routine must be called at splvm().
689 * We allocate in round-robin fashion from the configured devices.
692 swp_pager_getswapspace(int npages)
699 mtx_lock(&sw_dev_mtx);
701 for (i = 0; i < nswapdev; i++) {
703 sp = TAILQ_FIRST(&swtailq);
704 if (!(sp->sw_flags & SW_CLOSING)) {
705 blk = blist_alloc(sp->sw_blist, npages);
706 if (blk != SWAPBLK_NONE) {
708 sp->sw_used += npages;
709 swap_pager_avail -= npages;
711 swdevhd = TAILQ_NEXT(sp, sw_list);
715 sp = TAILQ_NEXT(sp, sw_list);
717 if (swap_pager_full != 2) {
718 printf("swap_pager_getswapspace(%d): failed\n", npages);
720 swap_pager_almost_full = 1;
724 mtx_unlock(&sw_dev_mtx);
729 swp_pager_isondev(daddr_t blk, struct swdevt *sp)
732 return (blk >= sp->sw_first && blk < sp->sw_end);
736 swp_pager_strategy(struct buf *bp)
740 mtx_lock(&sw_dev_mtx);
741 TAILQ_FOREACH(sp, &swtailq, sw_list) {
742 if (bp->b_blkno >= sp->sw_first && bp->b_blkno < sp->sw_end) {
743 mtx_unlock(&sw_dev_mtx);
744 sp->sw_strategy(bp, sp);
748 panic("Swapdev not found");
753 * SWP_PAGER_FREESWAPSPACE() - free raw swap space
755 * This routine returns the specified swap blocks back to the bitmap.
757 * Note: This routine may not block (it could in the old swap code),
758 * and through the use of the new blist routines it does not block.
760 * We must be called at splvm() to avoid races with bitmap frees from
761 * vm_page_remove() aka swap_pager_page_removed().
763 * This routine may not block
764 * This routine must be called at splvm().
767 swp_pager_freeswapspace(daddr_t blk, int npages)
771 mtx_lock(&sw_dev_mtx);
772 TAILQ_FOREACH(sp, &swtailq, sw_list) {
773 if (blk >= sp->sw_first && blk < sp->sw_end) {
774 sp->sw_used -= npages;
776 * If we are attempting to stop swapping on
777 * this device, we don't want to mark any
778 * blocks free lest they be reused.
780 if ((sp->sw_flags & SW_CLOSING) == 0) {
781 blist_free(sp->sw_blist, blk - sp->sw_first,
783 swap_pager_avail += npages;
786 mtx_unlock(&sw_dev_mtx);
790 panic("Swapdev not found");
794 * SWAP_PAGER_FREESPACE() - frees swap blocks associated with a page
795 * range within an object.
797 * This is a globally accessible routine.
799 * This routine removes swapblk assignments from swap metadata.
801 * The external callers of this routine typically have already destroyed
802 * or renamed vm_page_t's associated with this range in the object so
805 * This routine may be called at any spl. We up our spl to splvm temporarily
806 * in order to perform the metadata removal.
809 swap_pager_freespace(vm_object_t object, vm_pindex_t start, vm_size_t size)
812 VM_OBJECT_LOCK_ASSERT(object, MA_OWNED);
813 swp_pager_meta_free(object, start, size);
817 * SWAP_PAGER_RESERVE() - reserve swap blocks in object
819 * Assigns swap blocks to the specified range within the object. The
820 * swap blocks are not zerod. Any previous swap assignment is destroyed.
822 * Returns 0 on success, -1 on failure.
825 swap_pager_reserve(vm_object_t object, vm_pindex_t start, vm_size_t size)
828 daddr_t blk = SWAPBLK_NONE;
829 vm_pindex_t beg = start; /* save start index */
831 VM_OBJECT_LOCK(object);
835 while ((blk = swp_pager_getswapspace(n)) == SWAPBLK_NONE) {
838 swp_pager_meta_free(object, beg, start - beg);
839 VM_OBJECT_UNLOCK(object);
844 swp_pager_meta_build(object, start, blk);
850 swp_pager_meta_free(object, start, n);
851 VM_OBJECT_UNLOCK(object);
856 * SWAP_PAGER_COPY() - copy blocks from source pager to destination pager
857 * and destroy the source.
859 * Copy any valid swapblks from the source to the destination. In
860 * cases where both the source and destination have a valid swapblk,
861 * we keep the destination's.
863 * This routine is allowed to block. It may block allocating metadata
864 * indirectly through swp_pager_meta_build() or if paging is still in
865 * progress on the source.
867 * This routine can be called at any spl
869 * XXX vm_page_collapse() kinda expects us not to block because we
870 * supposedly do not need to allocate memory, but for the moment we
871 * *may* have to get a little memory from the zone allocator, but
872 * it is taken from the interrupt memory. We should be ok.
874 * The source object contains no vm_page_t's (which is just as well)
876 * The source object is of type OBJT_SWAP.
878 * The source and destination objects must be locked or
879 * inaccessible (XXX are they ?)
882 swap_pager_copy(vm_object_t srcobject, vm_object_t dstobject,
883 vm_pindex_t offset, int destroysource)
887 VM_OBJECT_LOCK_ASSERT(srcobject, MA_OWNED);
888 VM_OBJECT_LOCK_ASSERT(dstobject, MA_OWNED);
891 * If destroysource is set, we remove the source object from the
892 * swap_pager internal queue now.
895 if (srcobject->handle != NULL) {
896 mtx_lock(&sw_alloc_mtx);
898 NOBJLIST(srcobject->handle),
902 mtx_unlock(&sw_alloc_mtx);
907 * transfer source to destination.
909 for (i = 0; i < dstobject->size; ++i) {
913 * Locate (without changing) the swapblk on the destination,
914 * unless it is invalid in which case free it silently, or
915 * if the destination is a resident page, in which case the
916 * source is thrown away.
918 dstaddr = swp_pager_meta_ctl(dstobject, i, 0);
920 if (dstaddr == SWAPBLK_NONE) {
922 * Destination has no swapblk and is not resident,
927 srcaddr = swp_pager_meta_ctl(
933 if (srcaddr != SWAPBLK_NONE) {
935 * swp_pager_meta_build() can sleep.
937 vm_object_pip_add(srcobject, 1);
938 VM_OBJECT_UNLOCK(srcobject);
939 vm_object_pip_add(dstobject, 1);
940 swp_pager_meta_build(dstobject, i, srcaddr);
941 vm_object_pip_wakeup(dstobject);
942 VM_OBJECT_LOCK(srcobject);
943 vm_object_pip_wakeup(srcobject);
947 * Destination has valid swapblk or it is represented
948 * by a resident page. We destroy the sourceblock.
951 swp_pager_meta_ctl(srcobject, i + offset, SWM_FREE);
956 * Free left over swap blocks in source.
958 * We have to revert the type to OBJT_DEFAULT so we do not accidently
959 * double-remove the object from the swap queues.
962 swp_pager_meta_free_all(srcobject);
964 * Reverting the type is not necessary, the caller is going
965 * to destroy srcobject directly, but I'm doing it here
966 * for consistency since we've removed the object from its
969 srcobject->type = OBJT_DEFAULT;
974 * SWAP_PAGER_HASPAGE() - determine if we have good backing store for
975 * the requested page.
977 * We determine whether good backing store exists for the requested
978 * page and return TRUE if it does, FALSE if it doesn't.
980 * If TRUE, we also try to determine how much valid, contiguous backing
981 * store exists before and after the requested page within a reasonable
982 * distance. We do not try to restrict it to the swap device stripe
983 * (that is handled in getpages/putpages). It probably isn't worth
987 swap_pager_haspage(vm_object_t object, vm_pindex_t pindex, int *before, int *after)
991 VM_OBJECT_LOCK_ASSERT(object, MA_OWNED);
993 * do we have good backing store at the requested index ?
995 blk0 = swp_pager_meta_ctl(object, pindex, 0);
997 if (blk0 == SWAPBLK_NONE) {
1006 * find backwards-looking contiguous good backing store
1008 if (before != NULL) {
1011 for (i = 1; i < (SWB_NPAGES/2); ++i) {
1016 blk = swp_pager_meta_ctl(object, pindex - i, 0);
1017 if (blk != blk0 - i)
1024 * find forward-looking contiguous good backing store
1026 if (after != NULL) {
1029 for (i = 1; i < (SWB_NPAGES/2); ++i) {
1032 blk = swp_pager_meta_ctl(object, pindex + i, 0);
1033 if (blk != blk0 + i)
1042 * SWAP_PAGER_PAGE_UNSWAPPED() - remove swap backing store related to page
1044 * This removes any associated swap backing store, whether valid or
1045 * not, from the page.
1047 * This routine is typically called when a page is made dirty, at
1048 * which point any associated swap can be freed. MADV_FREE also
1049 * calls us in a special-case situation
1051 * NOTE!!! If the page is clean and the swap was valid, the caller
1052 * should make the page dirty before calling this routine. This routine
1053 * does NOT change the m->dirty status of the page. Also: MADV_FREE
1056 * This routine may not block
1057 * This routine must be called at splvm()
1060 swap_pager_unswapped(vm_page_t m)
1063 VM_OBJECT_LOCK_ASSERT(m->object, MA_OWNED);
1064 swp_pager_meta_ctl(m->object, m->pindex, SWM_FREE);
1068 * SWAP_PAGER_GETPAGES() - bring pages in from swap
1070 * Attempt to retrieve (m, count) pages from backing store, but make
1071 * sure we retrieve at least m[reqpage]. We try to load in as large
1072 * a chunk surrounding m[reqpage] as is contiguous in swap and which
1073 * belongs to the same object.
1075 * The code is designed for asynchronous operation and
1076 * immediate-notification of 'reqpage' but tends not to be
1077 * used that way. Please do not optimize-out this algorithmic
1078 * feature, I intend to improve on it in the future.
1080 * The parent has a single vm_object_pip_add() reference prior to
1081 * calling us and we should return with the same.
1083 * The parent has BUSY'd the pages. We should return with 'm'
1084 * left busy, but the others adjusted.
1087 swap_pager_getpages(vm_object_t object, vm_page_t *m, int count, int reqpage)
1097 KASSERT(mreq->object == object,
1098 ("swap_pager_getpages: object mismatch %p/%p",
1099 object, mreq->object));
1102 * Calculate range to retrieve. The pages have already been assigned
1103 * their swapblks. We require a *contiguous* range but we know it to
1104 * not span devices. If we do not supply it, bad things
1105 * happen. Note that blk, iblk & jblk can be SWAPBLK_NONE, but the
1106 * loops are set up such that the case(s) are handled implicitly.
1108 * The swp_*() calls must be made with the object locked.
1110 blk = swp_pager_meta_ctl(mreq->object, mreq->pindex, 0);
1112 for (i = reqpage - 1; i >= 0; --i) {
1115 iblk = swp_pager_meta_ctl(m[i]->object, m[i]->pindex, 0);
1116 if (blk != iblk + (reqpage - i))
1121 for (j = reqpage + 1; j < count; ++j) {
1124 jblk = swp_pager_meta_ctl(m[j]->object, m[j]->pindex, 0);
1125 if (blk != jblk - (j - reqpage))
1130 * free pages outside our collection range. Note: we never free
1131 * mreq, it must remain busy throughout.
1133 if (0 < i || j < count) {
1136 vm_page_lock_queues();
1137 for (k = 0; k < i; ++k)
1138 swp_pager_free_nrpage(m[k]);
1139 for (k = j; k < count; ++k)
1140 swp_pager_free_nrpage(m[k]);
1141 vm_page_unlock_queues();
1145 * Return VM_PAGER_FAIL if we have nothing to do. Return mreq
1146 * still busy, but the others unbusied.
1148 if (blk == SWAPBLK_NONE)
1149 return (VM_PAGER_FAIL);
1152 * Getpbuf() can sleep.
1154 VM_OBJECT_UNLOCK(object);
1156 * Get a swap buffer header to perform the IO
1158 bp = getpbuf(&nsw_rcount);
1159 bp->b_flags |= B_PAGING;
1162 * map our page(s) into kva for input
1164 pmap_qenter((vm_offset_t)bp->b_data, m + i, j - i);
1166 bp->b_iocmd = BIO_READ;
1167 bp->b_iodone = swp_pager_async_iodone;
1168 bp->b_rcred = crhold(thread0.td_ucred);
1169 bp->b_wcred = crhold(thread0.td_ucred);
1170 bp->b_blkno = blk - (reqpage - i);
1171 bp->b_bcount = PAGE_SIZE * (j - i);
1172 bp->b_bufsize = PAGE_SIZE * (j - i);
1173 bp->b_pager.pg_reqpage = reqpage - i;
1175 VM_OBJECT_LOCK(object);
1179 for (k = i; k < j; ++k) {
1180 bp->b_pages[k - i] = m[k];
1181 m[k]->oflags |= VPO_SWAPINPROG;
1184 bp->b_npages = j - i;
1186 PCPU_INC(cnt.v_swapin);
1187 PCPU_ADD(cnt.v_swappgsin, bp->b_npages);
1190 * We still hold the lock on mreq, and our automatic completion routine
1191 * does not remove it.
1193 vm_object_pip_add(object, bp->b_npages);
1194 VM_OBJECT_UNLOCK(object);
1197 * perform the I/O. NOTE!!! bp cannot be considered valid after
1198 * this point because we automatically release it on completion.
1199 * Instead, we look at the one page we are interested in which we
1200 * still hold a lock on even through the I/O completion.
1202 * The other pages in our m[] array are also released on completion,
1203 * so we cannot assume they are valid anymore either.
1205 * NOTE: b_blkno is destroyed by the call to swapdev_strategy
1208 swp_pager_strategy(bp);
1211 * wait for the page we want to complete. VPO_SWAPINPROG is always
1212 * cleared on completion. If an I/O error occurs, SWAPBLK_NONE
1213 * is set in the meta-data.
1215 VM_OBJECT_LOCK(object);
1216 while ((mreq->oflags & VPO_SWAPINPROG) != 0) {
1217 mreq->oflags |= VPO_WANTED;
1218 PCPU_INC(cnt.v_intrans);
1219 if (msleep(mreq, VM_OBJECT_MTX(object), PSWP, "swread", hz*20)) {
1221 "swap_pager: indefinite wait buffer: bufobj: %p, blkno: %jd, size: %ld\n",
1222 bp->b_bufobj, (intmax_t)bp->b_blkno, bp->b_bcount);
1227 * mreq is left busied after completion, but all the other pages
1228 * are freed. If we had an unrecoverable read error the page will
1231 if (mreq->valid != VM_PAGE_BITS_ALL) {
1232 return (VM_PAGER_ERROR);
1234 return (VM_PAGER_OK);
1238 * A final note: in a low swap situation, we cannot deallocate swap
1239 * and mark a page dirty here because the caller is likely to mark
1240 * the page clean when we return, causing the page to possibly revert
1241 * to all-zero's later.
1246 * swap_pager_putpages:
1248 * Assign swap (if necessary) and initiate I/O on the specified pages.
1250 * We support both OBJT_DEFAULT and OBJT_SWAP objects. DEFAULT objects
1251 * are automatically converted to SWAP objects.
1253 * In a low memory situation we may block in VOP_STRATEGY(), but the new
1254 * vm_page reservation system coupled with properly written VFS devices
1255 * should ensure that no low-memory deadlock occurs. This is an area
1258 * The parent has N vm_object_pip_add() references prior to
1259 * calling us and will remove references for rtvals[] that are
1260 * not set to VM_PAGER_PEND. We need to remove the rest on I/O
1263 * The parent has soft-busy'd the pages it passes us and will unbusy
1264 * those whos rtvals[] entry is not set to VM_PAGER_PEND on return.
1265 * We need to unbusy the rest on I/O completion.
1268 swap_pager_putpages(vm_object_t object, vm_page_t *m, int count,
1269 boolean_t sync, int *rtvals)
1274 if (count && m[0]->object != object) {
1275 panic("swap_pager_putpages: object mismatch %p/%p",
1284 * Turn object into OBJT_SWAP
1285 * check for bogus sysops
1286 * force sync if not pageout process
1288 if (object->type != OBJT_SWAP)
1289 swp_pager_meta_build(object, 0, SWAPBLK_NONE);
1290 VM_OBJECT_UNLOCK(object);
1292 if (curproc != pageproc)
1298 * Update nsw parameters from swap_async_max sysctl values.
1299 * Do not let the sysop crash the machine with bogus numbers.
1301 mtx_lock(&pbuf_mtx);
1302 if (swap_async_max != nsw_wcount_async_max) {
1308 if ((n = swap_async_max) > nswbuf / 2)
1315 * Adjust difference ( if possible ). If the current async
1316 * count is too low, we may not be able to make the adjustment
1319 n -= nsw_wcount_async_max;
1320 if (nsw_wcount_async + n >= 0) {
1321 nsw_wcount_async += n;
1322 nsw_wcount_async_max += n;
1323 wakeup(&nsw_wcount_async);
1326 mtx_unlock(&pbuf_mtx);
1331 * Assign swap blocks and issue I/O. We reallocate swap on the fly.
1332 * The page is left dirty until the pageout operation completes
1335 for (i = 0; i < count; i += n) {
1341 * Maximum I/O size is limited by a number of factors.
1343 n = min(BLIST_MAX_ALLOC, count - i);
1344 n = min(n, nsw_cluster_max);
1347 * Get biggest block of swap we can. If we fail, fall
1348 * back and try to allocate a smaller block. Don't go
1349 * overboard trying to allocate space if it would overly
1353 (blk = swp_pager_getswapspace(n)) == SWAPBLK_NONE &&
1358 if (blk == SWAPBLK_NONE) {
1359 for (j = 0; j < n; ++j)
1360 rtvals[i+j] = VM_PAGER_FAIL;
1365 * All I/O parameters have been satisfied, build the I/O
1366 * request and assign the swap space.
1369 bp = getpbuf(&nsw_wcount_sync);
1371 bp = getpbuf(&nsw_wcount_async);
1372 bp->b_flags = B_ASYNC;
1374 bp->b_flags |= B_PAGING;
1375 bp->b_iocmd = BIO_WRITE;
1377 pmap_qenter((vm_offset_t)bp->b_data, &m[i], n);
1379 bp->b_rcred = crhold(thread0.td_ucred);
1380 bp->b_wcred = crhold(thread0.td_ucred);
1381 bp->b_bcount = PAGE_SIZE * n;
1382 bp->b_bufsize = PAGE_SIZE * n;
1385 VM_OBJECT_LOCK(object);
1386 for (j = 0; j < n; ++j) {
1387 vm_page_t mreq = m[i+j];
1389 swp_pager_meta_build(
1394 vm_page_dirty(mreq);
1395 rtvals[i+j] = VM_PAGER_OK;
1397 mreq->oflags |= VPO_SWAPINPROG;
1398 bp->b_pages[j] = mreq;
1400 VM_OBJECT_UNLOCK(object);
1403 * Must set dirty range for NFS to work.
1406 bp->b_dirtyend = bp->b_bcount;
1408 PCPU_INC(cnt.v_swapout);
1409 PCPU_ADD(cnt.v_swappgsout, bp->b_npages);
1414 * NOTE: b_blkno is destroyed by the call to swapdev_strategy
1416 if (sync == FALSE) {
1417 bp->b_iodone = swp_pager_async_iodone;
1419 swp_pager_strategy(bp);
1421 for (j = 0; j < n; ++j)
1422 rtvals[i+j] = VM_PAGER_PEND;
1423 /* restart outter loop */
1430 * NOTE: b_blkno is destroyed by the call to swapdev_strategy
1432 bp->b_iodone = bdone;
1433 swp_pager_strategy(bp);
1436 * Wait for the sync I/O to complete, then update rtvals.
1437 * We just set the rtvals[] to VM_PAGER_PEND so we can call
1438 * our async completion routine at the end, thus avoiding a
1441 bwait(bp, PVM, "swwrt");
1442 for (j = 0; j < n; ++j)
1443 rtvals[i+j] = VM_PAGER_PEND;
1445 * Now that we are through with the bp, we can call the
1446 * normal async completion, which frees everything up.
1448 swp_pager_async_iodone(bp);
1450 VM_OBJECT_LOCK(object);
1454 * swp_pager_async_iodone:
1456 * Completion routine for asynchronous reads and writes from/to swap.
1457 * Also called manually by synchronous code to finish up a bp.
1459 * For READ operations, the pages are PG_BUSY'd. For WRITE operations,
1460 * the pages are vm_page_t->busy'd. For READ operations, we PG_BUSY
1461 * unbusy all pages except the 'main' request page. For WRITE
1462 * operations, we vm_page_t->busy'd unbusy all pages ( we can do this
1463 * because we marked them all VM_PAGER_PEND on return from putpages ).
1465 * This routine may not block.
1468 swp_pager_async_iodone(struct buf *bp)
1471 vm_object_t object = NULL;
1476 if (bp->b_ioflags & BIO_ERROR) {
1478 "swap_pager: I/O error - %s failed; blkno %ld,"
1479 "size %ld, error %d\n",
1480 ((bp->b_iocmd == BIO_READ) ? "pagein" : "pageout"),
1488 * remove the mapping for kernel virtual
1490 pmap_qremove((vm_offset_t)bp->b_data, bp->b_npages);
1493 object = bp->b_pages[0]->object;
1494 VM_OBJECT_LOCK(object);
1496 vm_page_lock_queues();
1498 * cleanup pages. If an error occurs writing to swap, we are in
1499 * very serious trouble. If it happens to be a disk error, though,
1500 * we may be able to recover by reassigning the swap later on. So
1501 * in this case we remove the m->swapblk assignment for the page
1502 * but do not free it in the rlist. The errornous block(s) are thus
1503 * never reallocated as swap. Redirty the page and continue.
1505 for (i = 0; i < bp->b_npages; ++i) {
1506 vm_page_t m = bp->b_pages[i];
1508 m->oflags &= ~VPO_SWAPINPROG;
1510 if (bp->b_ioflags & BIO_ERROR) {
1512 * If an error occurs I'd love to throw the swapblk
1513 * away without freeing it back to swapspace, so it
1514 * can never be used again. But I can't from an
1517 if (bp->b_iocmd == BIO_READ) {
1519 * When reading, reqpage needs to stay
1520 * locked for the parent, but all other
1521 * pages can be freed. We still want to
1522 * wakeup the parent waiting on the page,
1523 * though. ( also: pg_reqpage can be -1 and
1524 * not match anything ).
1526 * We have to wake specifically requested pages
1527 * up too because we cleared VPO_SWAPINPROG and
1528 * someone may be waiting for that.
1530 * NOTE: for reads, m->dirty will probably
1531 * be overridden by the original caller of
1532 * getpages so don't play cute tricks here.
1535 if (i != bp->b_pager.pg_reqpage)
1536 swp_pager_free_nrpage(m);
1540 * If i == bp->b_pager.pg_reqpage, do not wake
1541 * the page up. The caller needs to.
1545 * If a write error occurs, reactivate page
1546 * so it doesn't clog the inactive list,
1547 * then finish the I/O.
1550 vm_page_activate(m);
1551 vm_page_io_finish(m);
1553 } else if (bp->b_iocmd == BIO_READ) {
1555 * NOTE: for reads, m->dirty will probably be
1556 * overridden by the original caller of getpages so
1557 * we cannot set them in order to free the underlying
1558 * swap in a low-swap situation. I don't think we'd
1559 * want to do that anyway, but it was an optimization
1560 * that existed in the old swapper for a time before
1561 * it got ripped out due to precisely this problem.
1563 * If not the requested page then deactivate it.
1565 * Note that the requested page, reqpage, is left
1566 * busied, but we still have to wake it up. The
1567 * other pages are released (unbusied) by
1570 KASSERT(!pmap_page_is_mapped(m),
1571 ("swp_pager_async_iodone: page %p is mapped", m));
1572 m->valid = VM_PAGE_BITS_ALL;
1573 KASSERT(m->dirty == 0,
1574 ("swp_pager_async_iodone: page %p is dirty", m));
1577 * We have to wake specifically requested pages
1578 * up too because we cleared VPO_SWAPINPROG and
1579 * could be waiting for it in getpages. However,
1580 * be sure to not unbusy getpages specifically
1581 * requested page - getpages expects it to be
1584 if (i != bp->b_pager.pg_reqpage) {
1585 vm_page_deactivate(m);
1592 * For write success, clear the dirty
1593 * status, then finish the I/O ( which decrements the
1594 * busy count and possibly wakes waiter's up ).
1596 KASSERT((m->flags & PG_WRITEABLE) == 0,
1597 ("swp_pager_async_iodone: page %p is not write"
1600 vm_page_io_finish(m);
1601 if (vm_page_count_severe())
1602 vm_page_try_to_cache(m);
1605 vm_page_unlock_queues();
1608 * adjust pip. NOTE: the original parent may still have its own
1609 * pip refs on the object.
1611 if (object != NULL) {
1612 vm_object_pip_wakeupn(object, bp->b_npages);
1613 VM_OBJECT_UNLOCK(object);
1617 * swapdev_strategy() manually sets b_vp and b_bufobj before calling
1618 * bstrategy(). Set them back to NULL now we're done with it, or we'll
1619 * trigger a KASSERT in relpbuf().
1623 bp->b_bufobj = NULL;
1626 * release the physical I/O buffer
1630 ((bp->b_iocmd == BIO_READ) ? &nsw_rcount :
1631 ((bp->b_flags & B_ASYNC) ?
1640 * swap_pager_isswapped:
1642 * Return 1 if at least one page in the given object is paged
1643 * out to the given swap device.
1645 * This routine may not block.
1648 swap_pager_isswapped(vm_object_t object, struct swdevt *sp)
1654 VM_OBJECT_LOCK_ASSERT(object, MA_OWNED);
1655 if (object->type != OBJT_SWAP)
1658 mtx_lock(&swhash_mtx);
1659 for (bcount = 0; bcount < object->un_pager.swp.swp_bcount; bcount++) {
1660 struct swblock *swap;
1662 if ((swap = *swp_pager_hash(object, index)) != NULL) {
1663 for (i = 0; i < SWAP_META_PAGES; ++i) {
1664 if (swp_pager_isondev(swap->swb_pages[i], sp)) {
1665 mtx_unlock(&swhash_mtx);
1670 index += SWAP_META_PAGES;
1671 if (index > 0x20000000)
1672 panic("swap_pager_isswapped: failed to locate all swap meta blocks");
1674 mtx_unlock(&swhash_mtx);
1679 * SWP_PAGER_FORCE_PAGEIN() - force a swap block to be paged in
1681 * This routine dissociates the page at the given index within a
1682 * swap block from its backing store, paging it in if necessary.
1683 * If the page is paged in, it is placed in the inactive queue,
1684 * since it had its backing store ripped out from under it.
1685 * We also attempt to swap in all other pages in the swap block,
1686 * we only guarantee that the one at the specified index is
1689 * XXX - The code to page the whole block in doesn't work, so we
1690 * revert to the one-by-one behavior for now. Sigh.
1693 swp_pager_force_pagein(vm_object_t object, vm_pindex_t pindex)
1697 vm_object_pip_add(object, 1);
1698 m = vm_page_grab(object, pindex, VM_ALLOC_NORMAL|VM_ALLOC_RETRY);
1699 if (m->valid == VM_PAGE_BITS_ALL) {
1700 vm_object_pip_subtract(object, 1);
1701 vm_page_lock_queues();
1702 vm_page_activate(m);
1704 vm_page_unlock_queues();
1706 vm_pager_page_unswapped(m);
1710 if (swap_pager_getpages(object, &m, 1, 0) != VM_PAGER_OK)
1711 panic("swap_pager_force_pagein: read from swap failed");/*XXX*/
1712 vm_object_pip_subtract(object, 1);
1713 vm_page_lock_queues();
1715 vm_page_dontneed(m);
1716 vm_page_unlock_queues();
1718 vm_pager_page_unswapped(m);
1722 * swap_pager_swapoff:
1724 * Page in all of the pages that have been paged out to the
1725 * given device. The corresponding blocks in the bitmap must be
1726 * marked as allocated and the device must be flagged SW_CLOSING.
1727 * There may be no processes swapped out to the device.
1729 * This routine may block.
1732 swap_pager_swapoff(struct swdevt *sp)
1734 struct swblock *swap;
1741 mtx_lock(&swhash_mtx);
1742 for (i = 0; i <= swhash_mask; i++) { /* '<=' is correct here */
1744 for (swap = swhash[i]; swap != NULL; swap = swap->swb_hnext) {
1745 vm_object_t object = swap->swb_object;
1746 vm_pindex_t pindex = swap->swb_index;
1747 for (j = 0; j < SWAP_META_PAGES; ++j) {
1748 if (swp_pager_isondev(swap->swb_pages[j], sp)) {
1749 /* avoid deadlock */
1750 if (!VM_OBJECT_TRYLOCK(object)) {
1753 mtx_unlock(&swhash_mtx);
1754 swp_pager_force_pagein(object,
1756 VM_OBJECT_UNLOCK(object);
1757 mtx_lock(&swhash_mtx);
1764 mtx_unlock(&swhash_mtx);
1767 * Objects may be locked or paging to the device being
1768 * removed, so we will miss their pages and need to
1769 * make another pass. We have marked this device as
1770 * SW_CLOSING, so the activity should finish soon.
1773 if (retries > 100) {
1774 panic("swapoff: failed to locate %d swap blocks",
1777 pause("swpoff", hz / 20);
1782 /************************************************************************
1784 ************************************************************************
1786 * These routines manipulate the swap metadata stored in the
1787 * OBJT_SWAP object. All swp_*() routines must be called at
1788 * splvm() because swap can be freed up by the low level vm_page
1789 * code which might be called from interrupts beyond what splbio() covers.
1791 * Swap metadata is implemented with a global hash and not directly
1792 * linked into the object. Instead the object simply contains
1793 * appropriate tracking counters.
1797 * SWP_PAGER_META_BUILD() - add swap block to swap meta data for object
1799 * We first convert the object to a swap object if it is a default
1802 * The specified swapblk is added to the object's swap metadata. If
1803 * the swapblk is not valid, it is freed instead. Any previously
1804 * assigned swapblk is freed.
1806 * This routine must be called at splvm(), except when used to convert
1807 * an OBJT_DEFAULT object into an OBJT_SWAP object.
1810 swp_pager_meta_build(vm_object_t object, vm_pindex_t pindex, daddr_t swapblk)
1812 struct swblock *swap;
1813 struct swblock **pswap;
1816 VM_OBJECT_LOCK_ASSERT(object, MA_OWNED);
1818 * Convert default object to swap object if necessary
1820 if (object->type != OBJT_SWAP) {
1821 object->type = OBJT_SWAP;
1822 object->un_pager.swp.swp_bcount = 0;
1824 if (object->handle != NULL) {
1825 mtx_lock(&sw_alloc_mtx);
1827 NOBJLIST(object->handle),
1831 mtx_unlock(&sw_alloc_mtx);
1836 * Locate hash entry. If not found create, but if we aren't adding
1837 * anything just return. If we run out of space in the map we wait
1838 * and, since the hash table may have changed, retry.
1841 mtx_lock(&swhash_mtx);
1842 pswap = swp_pager_hash(object, pindex);
1844 if ((swap = *pswap) == NULL) {
1847 if (swapblk == SWAPBLK_NONE)
1850 swap = *pswap = uma_zalloc(swap_zone, M_NOWAIT);
1852 mtx_unlock(&swhash_mtx);
1853 VM_OBJECT_UNLOCK(object);
1854 if (uma_zone_exhausted(swap_zone)) {
1855 printf("swap zone exhausted, increase kern.maxswzone\n");
1856 vm_pageout_oom(VM_OOM_SWAPZ);
1857 pause("swzonex", 10);
1860 VM_OBJECT_LOCK(object);
1864 swap->swb_hnext = NULL;
1865 swap->swb_object = object;
1866 swap->swb_index = pindex & ~(vm_pindex_t)SWAP_META_MASK;
1867 swap->swb_count = 0;
1869 ++object->un_pager.swp.swp_bcount;
1871 for (i = 0; i < SWAP_META_PAGES; ++i)
1872 swap->swb_pages[i] = SWAPBLK_NONE;
1876 * Delete prior contents of metadata
1878 idx = pindex & SWAP_META_MASK;
1880 if (swap->swb_pages[idx] != SWAPBLK_NONE) {
1881 swp_pager_freeswapspace(swap->swb_pages[idx], 1);
1886 * Enter block into metadata
1888 swap->swb_pages[idx] = swapblk;
1889 if (swapblk != SWAPBLK_NONE)
1892 mtx_unlock(&swhash_mtx);
1896 * SWP_PAGER_META_FREE() - free a range of blocks in the object's swap metadata
1898 * The requested range of blocks is freed, with any associated swap
1899 * returned to the swap bitmap.
1901 * This routine will free swap metadata structures as they are cleaned
1902 * out. This routine does *NOT* operate on swap metadata associated
1903 * with resident pages.
1905 * This routine must be called at splvm()
1908 swp_pager_meta_free(vm_object_t object, vm_pindex_t index, daddr_t count)
1911 VM_OBJECT_LOCK_ASSERT(object, MA_OWNED);
1912 if (object->type != OBJT_SWAP)
1916 struct swblock **pswap;
1917 struct swblock *swap;
1919 mtx_lock(&swhash_mtx);
1920 pswap = swp_pager_hash(object, index);
1922 if ((swap = *pswap) != NULL) {
1923 daddr_t v = swap->swb_pages[index & SWAP_META_MASK];
1925 if (v != SWAPBLK_NONE) {
1926 swp_pager_freeswapspace(v, 1);
1927 swap->swb_pages[index & SWAP_META_MASK] =
1929 if (--swap->swb_count == 0) {
1930 *pswap = swap->swb_hnext;
1931 uma_zfree(swap_zone, swap);
1932 --object->un_pager.swp.swp_bcount;
1938 int n = SWAP_META_PAGES - (index & SWAP_META_MASK);
1942 mtx_unlock(&swhash_mtx);
1947 * SWP_PAGER_META_FREE_ALL() - destroy all swap metadata associated with object
1949 * This routine locates and destroys all swap metadata associated with
1952 * This routine must be called at splvm()
1955 swp_pager_meta_free_all(vm_object_t object)
1959 VM_OBJECT_LOCK_ASSERT(object, MA_OWNED);
1960 if (object->type != OBJT_SWAP)
1963 while (object->un_pager.swp.swp_bcount) {
1964 struct swblock **pswap;
1965 struct swblock *swap;
1967 mtx_lock(&swhash_mtx);
1968 pswap = swp_pager_hash(object, index);
1969 if ((swap = *pswap) != NULL) {
1972 for (i = 0; i < SWAP_META_PAGES; ++i) {
1973 daddr_t v = swap->swb_pages[i];
1974 if (v != SWAPBLK_NONE) {
1976 swp_pager_freeswapspace(v, 1);
1979 if (swap->swb_count != 0)
1980 panic("swap_pager_meta_free_all: swb_count != 0");
1981 *pswap = swap->swb_hnext;
1982 uma_zfree(swap_zone, swap);
1983 --object->un_pager.swp.swp_bcount;
1985 mtx_unlock(&swhash_mtx);
1986 index += SWAP_META_PAGES;
1987 if (index > 0x20000000)
1988 panic("swp_pager_meta_free_all: failed to locate all swap meta blocks");
1993 * SWP_PAGER_METACTL() - misc control of swap and vm_page_t meta data.
1995 * This routine is capable of looking up, popping, or freeing
1996 * swapblk assignments in the swap meta data or in the vm_page_t.
1997 * The routine typically returns the swapblk being looked-up, or popped,
1998 * or SWAPBLK_NONE if the block was freed, or SWAPBLK_NONE if the block
1999 * was invalid. This routine will automatically free any invalid
2000 * meta-data swapblks.
2002 * It is not possible to store invalid swapblks in the swap meta data
2003 * (other then a literal 'SWAPBLK_NONE'), so we don't bother checking.
2005 * When acting on a busy resident page and paging is in progress, we
2006 * have to wait until paging is complete but otherwise can act on the
2009 * This routine must be called at splvm().
2011 * SWM_FREE remove and free swap block from metadata
2012 * SWM_POP remove from meta data but do not free.. pop it out
2015 swp_pager_meta_ctl(vm_object_t object, vm_pindex_t pindex, int flags)
2017 struct swblock **pswap;
2018 struct swblock *swap;
2022 VM_OBJECT_LOCK_ASSERT(object, MA_OWNED);
2024 * The meta data only exists of the object is OBJT_SWAP
2025 * and even then might not be allocated yet.
2027 if (object->type != OBJT_SWAP)
2028 return (SWAPBLK_NONE);
2031 mtx_lock(&swhash_mtx);
2032 pswap = swp_pager_hash(object, pindex);
2034 if ((swap = *pswap) != NULL) {
2035 idx = pindex & SWAP_META_MASK;
2036 r1 = swap->swb_pages[idx];
2038 if (r1 != SWAPBLK_NONE) {
2039 if (flags & SWM_FREE) {
2040 swp_pager_freeswapspace(r1, 1);
2043 if (flags & (SWM_FREE|SWM_POP)) {
2044 swap->swb_pages[idx] = SWAPBLK_NONE;
2045 if (--swap->swb_count == 0) {
2046 *pswap = swap->swb_hnext;
2047 uma_zfree(swap_zone, swap);
2048 --object->un_pager.swp.swp_bcount;
2053 mtx_unlock(&swhash_mtx);
2058 * System call swapon(name) enables swapping on device name,
2059 * which must be in the swdevsw. Return EBUSY
2060 * if already swapping on this device.
2062 #ifndef _SYS_SYSPROTO_H_
2063 struct swapon_args {
2073 swapon(struct thread *td, struct swapon_args *uap)
2077 struct nameidata nd;
2080 error = priv_check(td, PRIV_SWAPON);
2085 while (swdev_syscall_active)
2086 tsleep(&swdev_syscall_active, PUSER - 1, "swpon", 0);
2087 swdev_syscall_active = 1;
2090 * Swap metadata may not fit in the KVM if we have physical
2093 if (swap_zone == NULL) {
2098 NDINIT(&nd, LOOKUP, ISOPEN | FOLLOW | AUDITVNODE1, UIO_USERSPACE,
2104 NDFREE(&nd, NDF_ONLY_PNBUF);
2107 if (vn_isdisk(vp, &error)) {
2108 error = swapongeom(td, vp);
2109 } else if (vp->v_type == VREG &&
2110 (vp->v_mount->mnt_vfc->vfc_flags & VFCF_NETWORK) != 0 &&
2111 (error = VOP_GETATTR(vp, &attr, td->td_ucred)) == 0) {
2113 * Allow direct swapping to NFS regular files in the same
2114 * way that nfs_mountroot() sets up diskless swapping.
2116 error = swaponvp(td, vp, attr.va_size / DEV_BSIZE);
2122 swdev_syscall_active = 0;
2123 wakeup_one(&swdev_syscall_active);
2129 swaponsomething(struct vnode *vp, void *id, u_long nblks, sw_strategy_t *strategy, sw_close_t *close, dev_t dev)
2131 struct swdevt *sp, *tsp;
2136 * If we go beyond this, we get overflows in the radix
2139 mblocks = 0x40000000 / BLIST_META_RADIX;
2140 if (nblks > mblocks) {
2141 printf("WARNING: reducing size to maximum of %lu blocks per swap unit\n",
2146 * nblks is in DEV_BSIZE'd chunks, convert to PAGE_SIZE'd chunks.
2147 * First chop nblks off to page-align it, then convert.
2149 * sw->sw_nblks is in page-sized chunks now too.
2151 nblks &= ~(ctodb(1) - 1);
2152 nblks = dbtoc(nblks);
2154 sp = malloc(sizeof *sp, M_VMPGDATA, M_WAITOK | M_ZERO);
2159 sp->sw_nblks = nblks;
2161 sp->sw_strategy = strategy;
2162 sp->sw_close = close;
2164 sp->sw_blist = blist_create(nblks, M_WAITOK);
2166 * Do not free the first two block in order to avoid overwriting
2167 * any bsd label at the front of the partition
2169 blist_free(sp->sw_blist, 2, nblks - 2);
2172 mtx_lock(&sw_dev_mtx);
2173 TAILQ_FOREACH(tsp, &swtailq, sw_list) {
2174 if (tsp->sw_end >= dvbase) {
2176 * We put one uncovered page between the devices
2177 * in order to definitively prevent any cross-device
2180 dvbase = tsp->sw_end + 1;
2183 sp->sw_first = dvbase;
2184 sp->sw_end = dvbase + nblks;
2185 TAILQ_INSERT_TAIL(&swtailq, sp, sw_list);
2187 swap_pager_avail += nblks;
2188 swap_total += (vm_ooffset_t)nblks * PAGE_SIZE;
2190 mtx_unlock(&sw_dev_mtx);
2194 * SYSCALL: swapoff(devname)
2196 * Disable swapping on the given device.
2198 * XXX: Badly designed system call: it should use a device index
2199 * rather than filename as specification. We keep sw_vp around
2200 * only to make this work.
2202 #ifndef _SYS_SYSPROTO_H_
2203 struct swapoff_args {
2213 swapoff(struct thread *td, struct swapoff_args *uap)
2216 struct nameidata nd;
2220 error = priv_check(td, PRIV_SWAPOFF);
2225 while (swdev_syscall_active)
2226 tsleep(&swdev_syscall_active, PUSER - 1, "swpoff", 0);
2227 swdev_syscall_active = 1;
2229 NDINIT(&nd, LOOKUP, FOLLOW | AUDITVNODE1, UIO_USERSPACE, uap->name,
2234 NDFREE(&nd, NDF_ONLY_PNBUF);
2237 mtx_lock(&sw_dev_mtx);
2238 TAILQ_FOREACH(sp, &swtailq, sw_list) {
2239 if (sp->sw_vp == vp)
2242 mtx_unlock(&sw_dev_mtx);
2247 error = swapoff_one(sp, td->td_ucred);
2249 swdev_syscall_active = 0;
2250 wakeup_one(&swdev_syscall_active);
2256 swapoff_one(struct swdevt *sp, struct ucred *cred)
2258 u_long nblks, dvbase;
2263 mtx_assert(&Giant, MA_OWNED);
2265 (void) vn_lock(sp->sw_vp, LK_EXCLUSIVE | LK_RETRY);
2266 error = mac_system_check_swapoff(cred, sp->sw_vp);
2267 (void) VOP_UNLOCK(sp->sw_vp, 0);
2271 nblks = sp->sw_nblks;
2274 * We can turn off this swap device safely only if the
2275 * available virtual memory in the system will fit the amount
2276 * of data we will have to page back in, plus an epsilon so
2277 * the system doesn't become critically low on swap space.
2279 if (cnt.v_free_count + cnt.v_cache_count + swap_pager_avail <
2280 nblks + nswap_lowat) {
2285 * Prevent further allocations on this device.
2287 mtx_lock(&sw_dev_mtx);
2288 sp->sw_flags |= SW_CLOSING;
2289 for (dvbase = 0; dvbase < sp->sw_end; dvbase += dmmax) {
2290 swap_pager_avail -= blist_fill(sp->sw_blist,
2293 swap_total -= (vm_ooffset_t)nblks * PAGE_SIZE;
2294 mtx_unlock(&sw_dev_mtx);
2297 * Page in the contents of the device and close it.
2299 swap_pager_swapoff(sp);
2301 sp->sw_close(curthread, sp);
2303 mtx_lock(&sw_dev_mtx);
2304 TAILQ_REMOVE(&swtailq, sp, sw_list);
2306 if (nswapdev == 0) {
2307 swap_pager_full = 2;
2308 swap_pager_almost_full = 1;
2312 mtx_unlock(&sw_dev_mtx);
2313 blist_destroy(sp->sw_blist);
2314 free(sp, M_VMPGDATA);
2321 struct swdevt *sp, *spt;
2322 const char *devname;
2326 while (swdev_syscall_active)
2327 tsleep(&swdev_syscall_active, PUSER - 1, "swpoff", 0);
2328 swdev_syscall_active = 1;
2330 mtx_lock(&sw_dev_mtx);
2331 TAILQ_FOREACH_SAFE(sp, &swtailq, sw_list, spt) {
2332 mtx_unlock(&sw_dev_mtx);
2333 if (vn_isdisk(sp->sw_vp, NULL))
2334 devname = sp->sw_vp->v_rdev->si_name;
2337 error = swapoff_one(sp, thread0.td_ucred);
2339 printf("Cannot remove swap device %s (error=%d), "
2340 "skipping.\n", devname, error);
2341 } else if (bootverbose) {
2342 printf("Swap device %s removed.\n", devname);
2344 mtx_lock(&sw_dev_mtx);
2346 mtx_unlock(&sw_dev_mtx);
2348 swdev_syscall_active = 0;
2349 wakeup_one(&swdev_syscall_active);
2354 swap_pager_status(int *total, int *used)
2360 mtx_lock(&sw_dev_mtx);
2361 TAILQ_FOREACH(sp, &swtailq, sw_list) {
2362 *total += sp->sw_nblks;
2363 *used += sp->sw_used;
2365 mtx_unlock(&sw_dev_mtx);
2369 sysctl_vm_swap_info(SYSCTL_HANDLER_ARGS)
2371 int *name = (int *)arg1;
2376 if (arg2 != 1) /* name length */
2380 mtx_lock(&sw_dev_mtx);
2381 TAILQ_FOREACH(sp, &swtailq, sw_list) {
2383 mtx_unlock(&sw_dev_mtx);
2384 xs.xsw_version = XSWDEV_VERSION;
2385 xs.xsw_dev = sp->sw_dev;
2386 xs.xsw_flags = sp->sw_flags;
2387 xs.xsw_nblks = sp->sw_nblks;
2388 xs.xsw_used = sp->sw_used;
2390 error = SYSCTL_OUT(req, &xs, sizeof(xs));
2395 mtx_unlock(&sw_dev_mtx);
2399 SYSCTL_INT(_vm, OID_AUTO, nswapdev, CTLFLAG_RD, &nswapdev, 0,
2400 "Number of swap devices");
2401 SYSCTL_NODE(_vm, OID_AUTO, swap_info, CTLFLAG_RD, sysctl_vm_swap_info,
2402 "Swap statistics by device");
2405 * vmspace_swap_count() - count the approximate swap usage in pages for a
2408 * The map must be locked.
2410 * Swap usage is determined by taking the proportional swap used by
2411 * VM objects backing the VM map. To make up for fractional losses,
2412 * if the VM object has any swap use at all the associated map entries
2413 * count for at least 1 swap page.
2416 vmspace_swap_count(struct vmspace *vmspace)
2418 vm_map_t map = &vmspace->vm_map;
2422 for (cur = map->header.next; cur != &map->header; cur = cur->next) {
2425 if ((cur->eflags & MAP_ENTRY_IS_SUB_MAP) == 0 &&
2426 (object = cur->object.vm_object) != NULL) {
2427 VM_OBJECT_LOCK(object);
2428 if (object->type == OBJT_SWAP &&
2429 object->un_pager.swp.swp_bcount != 0) {
2430 int n = (cur->end - cur->start) / PAGE_SIZE;
2432 count += object->un_pager.swp.swp_bcount *
2433 SWAP_META_PAGES * n / object->size + 1;
2435 VM_OBJECT_UNLOCK(object);
2444 * Swapping onto disk devices.
2448 static g_orphan_t swapgeom_orphan;
2450 static struct g_class g_swap_class = {
2452 .version = G_VERSION,
2453 .orphan = swapgeom_orphan,
2456 DECLARE_GEOM_CLASS(g_swap_class, g_class);
2460 swapgeom_done(struct bio *bp2)
2464 bp = bp2->bio_caller2;
2465 bp->b_ioflags = bp2->bio_flags;
2467 bp->b_ioflags |= BIO_ERROR;
2468 bp->b_resid = bp->b_bcount - bp2->bio_completed;
2469 bp->b_error = bp2->bio_error;
2475 swapgeom_strategy(struct buf *bp, struct swdevt *sp)
2478 struct g_consumer *cp;
2482 bp->b_error = ENXIO;
2483 bp->b_ioflags |= BIO_ERROR;
2487 if (bp->b_iocmd == BIO_WRITE)
2490 bio = g_alloc_bio();
2492 bp->b_error = ENOMEM;
2493 bp->b_ioflags |= BIO_ERROR;
2498 bio->bio_caller2 = bp;
2499 bio->bio_cmd = bp->b_iocmd;
2500 bio->bio_data = bp->b_data;
2501 bio->bio_offset = (bp->b_blkno - sp->sw_first) * PAGE_SIZE;
2502 bio->bio_length = bp->b_bcount;
2503 bio->bio_done = swapgeom_done;
2504 g_io_request(bio, cp);
2509 swapgeom_orphan(struct g_consumer *cp)
2513 mtx_lock(&sw_dev_mtx);
2514 TAILQ_FOREACH(sp, &swtailq, sw_list)
2515 if (sp->sw_id == cp)
2517 mtx_unlock(&sw_dev_mtx);
2521 swapgeom_close_ev(void *arg, int flags)
2523 struct g_consumer *cp;
2526 g_access(cp, -1, -1, 0);
2528 g_destroy_consumer(cp);
2532 swapgeom_close(struct thread *td, struct swdevt *sw)
2535 /* XXX: direct call when Giant untangled */
2536 g_waitfor_event(swapgeom_close_ev, sw->sw_id, M_WAITOK, NULL);
2547 swapongeom_ev(void *arg, int flags)
2550 struct g_provider *pp;
2551 struct g_consumer *cp;
2552 static struct g_geom *gp;
2559 pp = g_dev_getprovider(swh->dev);
2561 swh->error = ENODEV;
2564 mtx_lock(&sw_dev_mtx);
2565 TAILQ_FOREACH(sp, &swtailq, sw_list) {
2567 if (cp != NULL && cp->provider == pp) {
2568 mtx_unlock(&sw_dev_mtx);
2573 mtx_unlock(&sw_dev_mtx);
2575 gp = g_new_geomf(&g_swap_class, "swap", NULL);
2576 cp = g_new_consumer(gp);
2579 * XXX: Everytime you think you can improve the margin for
2580 * footshooting, somebody depends on the ability to do so:
2581 * savecore(8) wants to write to our swapdev so we cannot
2582 * set an exclusive count :-(
2584 error = g_access(cp, 1, 1, 0);
2587 g_destroy_consumer(cp);
2591 nblks = pp->mediasize / DEV_BSIZE;
2592 swaponsomething(swh->vp, cp, nblks, swapgeom_strategy,
2593 swapgeom_close, dev2udev(swh->dev));
2599 swapongeom(struct thread *td, struct vnode *vp)
2604 vn_lock(vp, LK_EXCLUSIVE | LK_RETRY);
2606 swh.dev = vp->v_rdev;
2609 /* XXX: direct call when Giant untangled */
2610 error = g_waitfor_event(swapongeom_ev, &swh, M_WAITOK, NULL);
2620 * This is used mainly for network filesystem (read: probably only tested
2621 * with NFS) swapfiles.
2626 swapdev_strategy(struct buf *bp, struct swdevt *sp)
2630 bp->b_blkno = ctodb(bp->b_blkno - sp->sw_first);
2634 if (bp->b_iocmd == BIO_WRITE) {
2636 bufobj_wdrop(bp->b_bufobj);
2637 bufobj_wref(&vp2->v_bufobj);
2639 if (bp->b_bufobj != &vp2->v_bufobj)
2640 bp->b_bufobj = &vp2->v_bufobj;
2642 bp->b_iooffset = dbtob(bp->b_blkno);
2648 swapdev_close(struct thread *td, struct swdevt *sp)
2651 VOP_CLOSE(sp->sw_vp, FREAD | FWRITE, td->td_ucred, td);
2657 swaponvp(struct thread *td, struct vnode *vp, u_long nblks)
2664 mtx_lock(&sw_dev_mtx);
2665 TAILQ_FOREACH(sp, &swtailq, sw_list) {
2666 if (sp->sw_id == vp) {
2667 mtx_unlock(&sw_dev_mtx);
2671 mtx_unlock(&sw_dev_mtx);
2673 (void) vn_lock(vp, LK_EXCLUSIVE | LK_RETRY);
2675 error = mac_system_check_swapon(td->td_ucred, vp);
2678 error = VOP_OPEN(vp, FREAD | FWRITE, td->td_ucred, td, NULL);
2679 (void) VOP_UNLOCK(vp, 0);
2683 swaponsomething(vp, vp, nblks, swapdev_strategy, swapdev_close,