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 MALLOC_DEFINE(M_VMPGDATA, "vm_pgdata", "swap pager private data");
151 static struct mtx sw_dev_mtx;
152 static TAILQ_HEAD(, swdevt) swtailq = TAILQ_HEAD_INITIALIZER(swtailq);
153 static struct swdevt *swdevhd; /* Allocate from here next */
154 static int nswapdev; /* Number of swap devices */
155 int swap_pager_avail;
156 static int swdev_syscall_active = 0; /* serialize swap(on|off) */
158 static vm_ooffset_t swap_total;
159 SYSCTL_QUAD(_vm, OID_AUTO, swap_total, CTLFLAG_RD, &swap_total, 0,
160 "Total amount of available swap storage.");
161 static vm_ooffset_t swap_reserved;
162 SYSCTL_QUAD(_vm, OID_AUTO, swap_reserved, CTLFLAG_RD, &swap_reserved, 0,
163 "Amount of swap storage needed to back all allocated anonymous memory.");
164 static int overcommit = 0;
165 SYSCTL_INT(_vm, OID_AUTO, overcommit, CTLFLAG_RW, &overcommit, 0,
166 "Configure virtual memory overcommit behavior. See tuning(7) "
169 /* bits from overcommit */
170 #define SWAP_RESERVE_FORCE_ON (1 << 0)
171 #define SWAP_RESERVE_RLIMIT_ON (1 << 1)
172 #define SWAP_RESERVE_ALLOW_NONWIRED (1 << 2)
175 swap_reserve(vm_ooffset_t incr)
178 return (swap_reserve_by_cred(incr, curthread->td_ucred));
182 swap_reserve_by_cred(vm_ooffset_t incr, struct ucred *cred)
187 static struct timeval lastfail;
190 uip = cred->cr_ruidinfo;
192 if (incr & PAGE_MASK)
193 panic("swap_reserve: & PAGE_MASK");
196 mtx_lock(&sw_dev_mtx);
197 r = swap_reserved + incr;
198 if (overcommit & SWAP_RESERVE_ALLOW_NONWIRED) {
199 s = cnt.v_page_count - cnt.v_free_reserved - cnt.v_wire_count;
204 if ((overcommit & SWAP_RESERVE_FORCE_ON) == 0 || r <= s ||
205 (error = priv_check(curthread, PRIV_VM_SWAP_NOQUOTA)) == 0) {
209 mtx_unlock(&sw_dev_mtx);
213 UIDINFO_VMSIZE_LOCK(uip);
214 if ((overcommit & SWAP_RESERVE_RLIMIT_ON) != 0 &&
215 uip->ui_vmsize + incr > lim_cur(curproc, RLIMIT_SWAP) &&
216 priv_check(curthread, PRIV_VM_SWAP_NORLIMIT))
219 uip->ui_vmsize += incr;
220 UIDINFO_VMSIZE_UNLOCK(uip);
221 PROC_UNLOCK(curproc);
223 mtx_lock(&sw_dev_mtx);
224 swap_reserved -= incr;
225 mtx_unlock(&sw_dev_mtx);
228 if (!res && ppsratecheck(&lastfail, &curfail, 1)) {
229 printf("uid %d, pid %d: swap reservation for %jd bytes failed\n",
230 curproc->p_pid, uip->ui_uid, incr);
237 swap_reserve_force(vm_ooffset_t incr)
241 mtx_lock(&sw_dev_mtx);
242 swap_reserved += incr;
243 mtx_unlock(&sw_dev_mtx);
245 uip = curthread->td_ucred->cr_ruidinfo;
247 UIDINFO_VMSIZE_LOCK(uip);
248 uip->ui_vmsize += incr;
249 UIDINFO_VMSIZE_UNLOCK(uip);
250 PROC_UNLOCK(curproc);
254 swap_release(vm_ooffset_t decr)
259 cred = curthread->td_ucred;
260 swap_release_by_cred(decr, cred);
261 PROC_UNLOCK(curproc);
265 swap_release_by_cred(vm_ooffset_t decr, struct ucred *cred)
269 uip = cred->cr_ruidinfo;
271 if (decr & PAGE_MASK)
272 panic("swap_release: & PAGE_MASK");
274 mtx_lock(&sw_dev_mtx);
275 if (swap_reserved < decr)
276 panic("swap_reserved < decr");
277 swap_reserved -= decr;
278 mtx_unlock(&sw_dev_mtx);
280 UIDINFO_VMSIZE_LOCK(uip);
281 if (uip->ui_vmsize < decr)
282 printf("negative vmsize for uid = %d\n", uip->ui_uid);
283 uip->ui_vmsize -= decr;
284 UIDINFO_VMSIZE_UNLOCK(uip);
287 static void swapdev_strategy(struct buf *, struct swdevt *sw);
289 #define SWM_FREE 0x02 /* free, period */
290 #define SWM_POP 0x04 /* pop out */
292 int swap_pager_full = 2; /* swap space exhaustion (task killing) */
293 static int swap_pager_almost_full = 1; /* swap space exhaustion (w/hysteresis)*/
294 static int nsw_rcount; /* free read buffers */
295 static int nsw_wcount_sync; /* limit write buffers / synchronous */
296 static int nsw_wcount_async; /* limit write buffers / asynchronous */
297 static int nsw_wcount_async_max;/* assigned maximum */
298 static int nsw_cluster_max; /* maximum VOP I/O allowed */
300 static struct swblock **swhash;
301 static int swhash_mask;
302 static struct mtx swhash_mtx;
304 static int swap_async_max = 4; /* maximum in-progress async I/O's */
305 static struct sx sw_alloc_sx;
308 SYSCTL_INT(_vm, OID_AUTO, swap_async_max,
309 CTLFLAG_RW, &swap_async_max, 0, "Maximum running async swap ops");
312 * "named" and "unnamed" anon region objects. Try to reduce the overhead
313 * of searching a named list by hashing it just a little.
318 #define NOBJLIST(handle) \
319 (&swap_pager_object_list[((int)(intptr_t)handle >> 4) & (NOBJLISTS-1)])
321 static struct mtx sw_alloc_mtx; /* protect list manipulation */
322 static struct pagerlst swap_pager_object_list[NOBJLISTS];
323 static uma_zone_t swap_zone;
324 static struct vm_object swap_zone_obj;
327 * pagerops for OBJT_SWAP - "swap pager". Some ops are also global procedure
328 * calls hooked from other parts of the VM system and do not appear here.
329 * (see vm/swap_pager.h).
332 swap_pager_alloc(void *handle, vm_ooffset_t size,
333 vm_prot_t prot, vm_ooffset_t offset, struct ucred *);
334 static void swap_pager_dealloc(vm_object_t object);
335 static int swap_pager_getpages(vm_object_t, vm_page_t *, int, int);
336 static void swap_pager_putpages(vm_object_t, vm_page_t *, int, boolean_t, int *);
338 swap_pager_haspage(vm_object_t object, vm_pindex_t pindex, int *before, int *after);
339 static void swap_pager_init(void);
340 static void swap_pager_unswapped(vm_page_t);
341 static void swap_pager_swapoff(struct swdevt *sp);
343 struct pagerops swappagerops = {
344 .pgo_init = swap_pager_init, /* early system initialization of pager */
345 .pgo_alloc = swap_pager_alloc, /* allocate an OBJT_SWAP object */
346 .pgo_dealloc = swap_pager_dealloc, /* deallocate an OBJT_SWAP object */
347 .pgo_getpages = swap_pager_getpages, /* pagein */
348 .pgo_putpages = swap_pager_putpages, /* pageout */
349 .pgo_haspage = swap_pager_haspage, /* get backing store status for page */
350 .pgo_pageunswapped = swap_pager_unswapped, /* remove swap related to page */
354 * dmmax is in page-sized chunks with the new swap system. It was
355 * dev-bsized chunks in the old. dmmax is always a power of 2.
357 * swap_*() routines are externally accessible. swp_*() routines are
361 static int nswap_lowat = 128; /* in pages, swap_pager_almost_full warn */
362 static int nswap_hiwat = 512; /* in pages, swap_pager_almost_full warn */
364 SYSCTL_INT(_vm, OID_AUTO, dmmax,
365 CTLFLAG_RD, &dmmax, 0, "Maximum size of a swap block");
367 static void swp_sizecheck(void);
368 static void swp_pager_async_iodone(struct buf *bp);
369 static int swapongeom(struct thread *, struct vnode *);
370 static int swaponvp(struct thread *, struct vnode *, u_long);
371 static int swapoff_one(struct swdevt *sp, struct ucred *cred);
374 * Swap bitmap functions
376 static void swp_pager_freeswapspace(daddr_t blk, int npages);
377 static daddr_t swp_pager_getswapspace(int npages);
382 static struct swblock **swp_pager_hash(vm_object_t object, vm_pindex_t index);
383 static void swp_pager_meta_build(vm_object_t, vm_pindex_t, daddr_t);
384 static void swp_pager_meta_free(vm_object_t, vm_pindex_t, daddr_t);
385 static void swp_pager_meta_free_all(vm_object_t);
386 static daddr_t swp_pager_meta_ctl(vm_object_t, vm_pindex_t, int);
389 swp_pager_free_nrpage(vm_page_t m)
393 if (m->wire_count == 0)
399 * SWP_SIZECHECK() - update swap_pager_full indication
401 * update the swap_pager_almost_full indication and warn when we are
402 * about to run out of swap space, using lowat/hiwat hysteresis.
404 * Clear swap_pager_full ( task killing ) indication when lowat is met.
406 * No restrictions on call
407 * This routine may not block.
408 * This routine must be called at splvm()
414 if (swap_pager_avail < nswap_lowat) {
415 if (swap_pager_almost_full == 0) {
416 printf("swap_pager: out of swap space\n");
417 swap_pager_almost_full = 1;
421 if (swap_pager_avail > nswap_hiwat)
422 swap_pager_almost_full = 0;
427 * SWP_PAGER_HASH() - hash swap meta data
429 * This is an helper function which hashes the swapblk given
430 * the object and page index. It returns a pointer to a pointer
431 * to the object, or a pointer to a NULL pointer if it could not
434 * This routine must be called at splvm().
436 static struct swblock **
437 swp_pager_hash(vm_object_t object, vm_pindex_t index)
439 struct swblock **pswap;
440 struct swblock *swap;
442 index &= ~(vm_pindex_t)SWAP_META_MASK;
443 pswap = &swhash[(index ^ (int)(intptr_t)object) & swhash_mask];
444 while ((swap = *pswap) != NULL) {
445 if (swap->swb_object == object &&
446 swap->swb_index == index
450 pswap = &swap->swb_hnext;
456 * SWAP_PAGER_INIT() - initialize the swap pager!
458 * Expected to be started from system init. NOTE: This code is run
459 * before much else so be careful what you depend on. Most of the VM
460 * system has yet to be initialized at this point.
463 swap_pager_init(void)
466 * Initialize object lists
470 for (i = 0; i < NOBJLISTS; ++i)
471 TAILQ_INIT(&swap_pager_object_list[i]);
472 mtx_init(&sw_alloc_mtx, "swap_pager list", NULL, MTX_DEF);
473 mtx_init(&sw_dev_mtx, "swapdev", NULL, MTX_DEF);
476 * Device Stripe, in PAGE_SIZE'd blocks
478 dmmax = SWB_NPAGES * 2;
482 * SWAP_PAGER_SWAP_INIT() - swap pager initialization from pageout process
484 * Expected to be started from pageout process once, prior to entering
488 swap_pager_swap_init(void)
493 * Number of in-transit swap bp operations. Don't
494 * exhaust the pbufs completely. Make sure we
495 * initialize workable values (0 will work for hysteresis
496 * but it isn't very efficient).
498 * The nsw_cluster_max is constrained by the bp->b_pages[]
499 * array (MAXPHYS/PAGE_SIZE) and our locally defined
500 * MAX_PAGEOUT_CLUSTER. Also be aware that swap ops are
501 * constrained by the swap device interleave stripe size.
503 * Currently we hardwire nsw_wcount_async to 4. This limit is
504 * designed to prevent other I/O from having high latencies due to
505 * our pageout I/O. The value 4 works well for one or two active swap
506 * devices but is probably a little low if you have more. Even so,
507 * a higher value would probably generate only a limited improvement
508 * with three or four active swap devices since the system does not
509 * typically have to pageout at extreme bandwidths. We will want
510 * at least 2 per swap devices, and 4 is a pretty good value if you
511 * have one NFS swap device due to the command/ack latency over NFS.
512 * So it all works out pretty well.
514 nsw_cluster_max = min((MAXPHYS/PAGE_SIZE), MAX_PAGEOUT_CLUSTER);
517 nsw_rcount = (nswbuf + 1) / 2;
518 nsw_wcount_sync = (nswbuf + 3) / 4;
519 nsw_wcount_async = 4;
520 nsw_wcount_async_max = nsw_wcount_async;
521 mtx_unlock(&pbuf_mtx);
524 * Initialize our zone. Right now I'm just guessing on the number
525 * we need based on the number of pages in the system. Each swblock
526 * can hold 16 pages, so this is probably overkill. This reservation
527 * is typically limited to around 32MB by default.
529 n = cnt.v_page_count / 2;
530 if (maxswzone && n > maxswzone / sizeof(struct swblock))
531 n = maxswzone / sizeof(struct swblock);
533 swap_zone = uma_zcreate("SWAPMETA", sizeof(struct swblock), NULL, NULL,
534 NULL, NULL, UMA_ALIGN_PTR, UMA_ZONE_NOFREE | UMA_ZONE_VM);
535 if (swap_zone == NULL)
536 panic("failed to create swap_zone.");
538 if (uma_zone_set_obj(swap_zone, &swap_zone_obj, n))
541 * if the allocation failed, try a zone two thirds the
542 * size of the previous attempt.
547 printf("Swap zone entries reduced from %d to %d.\n", n2, n);
551 * Initialize our meta-data hash table. The swapper does not need to
552 * be quite as efficient as the VM system, so we do not use an
553 * oversized hash table.
555 * n: size of hash table, must be power of 2
556 * swhash_mask: hash table index mask
558 for (n = 1; n < n2 / 8; n *= 2)
560 swhash = malloc(sizeof(struct swblock *) * n, M_VMPGDATA, M_WAITOK | M_ZERO);
562 mtx_init(&swhash_mtx, "swap_pager swhash", NULL, MTX_DEF);
566 * SWAP_PAGER_ALLOC() - allocate a new OBJT_SWAP VM object and instantiate
567 * its metadata structures.
569 * This routine is called from the mmap and fork code to create a new
570 * OBJT_SWAP object. We do this by creating an OBJT_DEFAULT object
571 * and then converting it with swp_pager_meta_build().
573 * This routine may block in vm_object_allocate() and create a named
574 * object lookup race, so we must interlock. We must also run at
575 * splvm() for the object lookup to handle races with interrupts, but
576 * we do not have to maintain splvm() in between the lookup and the
577 * add because (I believe) it is not possible to attempt to create
578 * a new swap object w/handle when a default object with that handle
584 swap_pager_alloc(void *handle, vm_ooffset_t size, vm_prot_t prot,
585 vm_ooffset_t offset, struct ucred *cred)
590 pindex = OFF_TO_IDX(offset + PAGE_MASK + size);
594 * Reference existing named region or allocate new one. There
595 * should not be a race here against swp_pager_meta_build()
596 * as called from vm_page_remove() in regards to the lookup
599 sx_xlock(&sw_alloc_sx);
600 object = vm_pager_object_lookup(NOBJLIST(handle), handle);
601 if (object == NULL) {
603 if (!swap_reserve_by_cred(size, cred)) {
604 sx_xunlock(&sw_alloc_sx);
610 object = vm_object_allocate(OBJT_DEFAULT, pindex);
611 VM_OBJECT_LOCK(object);
612 object->handle = handle;
615 object->charge = size;
617 swp_pager_meta_build(object, 0, SWAPBLK_NONE);
618 VM_OBJECT_UNLOCK(object);
620 sx_xunlock(&sw_alloc_sx);
624 if (!swap_reserve_by_cred(size, cred))
628 object = vm_object_allocate(OBJT_DEFAULT, pindex);
629 VM_OBJECT_LOCK(object);
632 object->charge = size;
634 swp_pager_meta_build(object, 0, SWAPBLK_NONE);
635 VM_OBJECT_UNLOCK(object);
641 * SWAP_PAGER_DEALLOC() - remove swap metadata from object
643 * The swap backing for the object is destroyed. The code is
644 * designed such that we can reinstantiate it later, but this
645 * routine is typically called only when the entire object is
646 * about to be destroyed.
648 * This routine may block, but no longer does.
650 * The object must be locked or unreferenceable.
653 swap_pager_dealloc(vm_object_t object)
657 * Remove from list right away so lookups will fail if we block for
658 * pageout completion.
660 if (object->handle != NULL) {
661 mtx_lock(&sw_alloc_mtx);
662 TAILQ_REMOVE(NOBJLIST(object->handle), object, pager_object_list);
663 mtx_unlock(&sw_alloc_mtx);
666 VM_OBJECT_LOCK_ASSERT(object, MA_OWNED);
667 vm_object_pip_wait(object, "swpdea");
670 * Free all remaining metadata. We only bother to free it from
671 * the swap meta data. We do not attempt to free swapblk's still
672 * associated with vm_page_t's for this object. We do not care
673 * if paging is still in progress on some objects.
675 swp_pager_meta_free_all(object);
678 /************************************************************************
679 * SWAP PAGER BITMAP ROUTINES *
680 ************************************************************************/
683 * SWP_PAGER_GETSWAPSPACE() - allocate raw swap space
685 * Allocate swap for the requested number of pages. The starting
686 * swap block number (a page index) is returned or SWAPBLK_NONE
687 * if the allocation failed.
689 * Also has the side effect of advising that somebody made a mistake
690 * when they configured swap and didn't configure enough.
692 * Must be called at splvm() to avoid races with bitmap frees from
693 * vm_page_remove() aka swap_pager_page_removed().
695 * This routine may not block
696 * This routine must be called at splvm().
698 * We allocate in round-robin fashion from the configured devices.
701 swp_pager_getswapspace(int npages)
708 mtx_lock(&sw_dev_mtx);
710 for (i = 0; i < nswapdev; i++) {
712 sp = TAILQ_FIRST(&swtailq);
713 if (!(sp->sw_flags & SW_CLOSING)) {
714 blk = blist_alloc(sp->sw_blist, npages);
715 if (blk != SWAPBLK_NONE) {
717 sp->sw_used += npages;
718 swap_pager_avail -= npages;
720 swdevhd = TAILQ_NEXT(sp, sw_list);
724 sp = TAILQ_NEXT(sp, sw_list);
726 if (swap_pager_full != 2) {
727 printf("swap_pager_getswapspace(%d): failed\n", npages);
729 swap_pager_almost_full = 1;
733 mtx_unlock(&sw_dev_mtx);
738 swp_pager_isondev(daddr_t blk, struct swdevt *sp)
741 return (blk >= sp->sw_first && blk < sp->sw_end);
745 swp_pager_strategy(struct buf *bp)
749 mtx_lock(&sw_dev_mtx);
750 TAILQ_FOREACH(sp, &swtailq, sw_list) {
751 if (bp->b_blkno >= sp->sw_first && bp->b_blkno < sp->sw_end) {
752 mtx_unlock(&sw_dev_mtx);
753 sp->sw_strategy(bp, sp);
757 panic("Swapdev not found");
762 * SWP_PAGER_FREESWAPSPACE() - free raw swap space
764 * This routine returns the specified swap blocks back to the bitmap.
766 * Note: This routine may not block (it could in the old swap code),
767 * and through the use of the new blist routines it does not block.
769 * We must be called at splvm() to avoid races with bitmap frees from
770 * vm_page_remove() aka swap_pager_page_removed().
772 * This routine may not block
773 * This routine must be called at splvm().
776 swp_pager_freeswapspace(daddr_t blk, int npages)
780 mtx_lock(&sw_dev_mtx);
781 TAILQ_FOREACH(sp, &swtailq, sw_list) {
782 if (blk >= sp->sw_first && blk < sp->sw_end) {
783 sp->sw_used -= npages;
785 * If we are attempting to stop swapping on
786 * this device, we don't want to mark any
787 * blocks free lest they be reused.
789 if ((sp->sw_flags & SW_CLOSING) == 0) {
790 blist_free(sp->sw_blist, blk - sp->sw_first,
792 swap_pager_avail += npages;
795 mtx_unlock(&sw_dev_mtx);
799 panic("Swapdev not found");
803 * SWAP_PAGER_FREESPACE() - frees swap blocks associated with a page
804 * range within an object.
806 * This is a globally accessible routine.
808 * This routine removes swapblk assignments from swap metadata.
810 * The external callers of this routine typically have already destroyed
811 * or renamed vm_page_t's associated with this range in the object so
814 * This routine may be called at any spl. We up our spl to splvm temporarily
815 * in order to perform the metadata removal.
818 swap_pager_freespace(vm_object_t object, vm_pindex_t start, vm_size_t size)
821 VM_OBJECT_LOCK_ASSERT(object, MA_OWNED);
822 swp_pager_meta_free(object, start, size);
826 * SWAP_PAGER_RESERVE() - reserve swap blocks in object
828 * Assigns swap blocks to the specified range within the object. The
829 * swap blocks are not zerod. Any previous swap assignment is destroyed.
831 * Returns 0 on success, -1 on failure.
834 swap_pager_reserve(vm_object_t object, vm_pindex_t start, vm_size_t size)
837 daddr_t blk = SWAPBLK_NONE;
838 vm_pindex_t beg = start; /* save start index */
840 VM_OBJECT_LOCK(object);
844 while ((blk = swp_pager_getswapspace(n)) == SWAPBLK_NONE) {
847 swp_pager_meta_free(object, beg, start - beg);
848 VM_OBJECT_UNLOCK(object);
853 swp_pager_meta_build(object, start, blk);
859 swp_pager_meta_free(object, start, n);
860 VM_OBJECT_UNLOCK(object);
865 * SWAP_PAGER_COPY() - copy blocks from source pager to destination pager
866 * and destroy the source.
868 * Copy any valid swapblks from the source to the destination. In
869 * cases where both the source and destination have a valid swapblk,
870 * we keep the destination's.
872 * This routine is allowed to block. It may block allocating metadata
873 * indirectly through swp_pager_meta_build() or if paging is still in
874 * progress on the source.
876 * This routine can be called at any spl
878 * XXX vm_page_collapse() kinda expects us not to block because we
879 * supposedly do not need to allocate memory, but for the moment we
880 * *may* have to get a little memory from the zone allocator, but
881 * it is taken from the interrupt memory. We should be ok.
883 * The source object contains no vm_page_t's (which is just as well)
885 * The source object is of type OBJT_SWAP.
887 * The source and destination objects must be locked or
888 * inaccessible (XXX are they ?)
891 swap_pager_copy(vm_object_t srcobject, vm_object_t dstobject,
892 vm_pindex_t offset, int destroysource)
896 VM_OBJECT_LOCK_ASSERT(srcobject, MA_OWNED);
897 VM_OBJECT_LOCK_ASSERT(dstobject, MA_OWNED);
900 * If destroysource is set, we remove the source object from the
901 * swap_pager internal queue now.
904 if (srcobject->handle != NULL) {
905 mtx_lock(&sw_alloc_mtx);
907 NOBJLIST(srcobject->handle),
911 mtx_unlock(&sw_alloc_mtx);
916 * transfer source to destination.
918 for (i = 0; i < dstobject->size; ++i) {
922 * Locate (without changing) the swapblk on the destination,
923 * unless it is invalid in which case free it silently, or
924 * if the destination is a resident page, in which case the
925 * source is thrown away.
927 dstaddr = swp_pager_meta_ctl(dstobject, i, 0);
929 if (dstaddr == SWAPBLK_NONE) {
931 * Destination has no swapblk and is not resident,
936 srcaddr = swp_pager_meta_ctl(
942 if (srcaddr != SWAPBLK_NONE) {
944 * swp_pager_meta_build() can sleep.
946 vm_object_pip_add(srcobject, 1);
947 VM_OBJECT_UNLOCK(srcobject);
948 vm_object_pip_add(dstobject, 1);
949 swp_pager_meta_build(dstobject, i, srcaddr);
950 vm_object_pip_wakeup(dstobject);
951 VM_OBJECT_LOCK(srcobject);
952 vm_object_pip_wakeup(srcobject);
956 * Destination has valid swapblk or it is represented
957 * by a resident page. We destroy the sourceblock.
960 swp_pager_meta_ctl(srcobject, i + offset, SWM_FREE);
965 * Free left over swap blocks in source.
967 * We have to revert the type to OBJT_DEFAULT so we do not accidently
968 * double-remove the object from the swap queues.
971 swp_pager_meta_free_all(srcobject);
973 * Reverting the type is not necessary, the caller is going
974 * to destroy srcobject directly, but I'm doing it here
975 * for consistency since we've removed the object from its
978 srcobject->type = OBJT_DEFAULT;
983 * SWAP_PAGER_HASPAGE() - determine if we have good backing store for
984 * the requested page.
986 * We determine whether good backing store exists for the requested
987 * page and return TRUE if it does, FALSE if it doesn't.
989 * If TRUE, we also try to determine how much valid, contiguous backing
990 * store exists before and after the requested page within a reasonable
991 * distance. We do not try to restrict it to the swap device stripe
992 * (that is handled in getpages/putpages). It probably isn't worth
996 swap_pager_haspage(vm_object_t object, vm_pindex_t pindex, int *before, int *after)
1000 VM_OBJECT_LOCK_ASSERT(object, MA_OWNED);
1002 * do we have good backing store at the requested index ?
1004 blk0 = swp_pager_meta_ctl(object, pindex, 0);
1006 if (blk0 == SWAPBLK_NONE) {
1015 * find backwards-looking contiguous good backing store
1017 if (before != NULL) {
1020 for (i = 1; i < (SWB_NPAGES/2); ++i) {
1025 blk = swp_pager_meta_ctl(object, pindex - i, 0);
1026 if (blk != blk0 - i)
1033 * find forward-looking contiguous good backing store
1035 if (after != NULL) {
1038 for (i = 1; i < (SWB_NPAGES/2); ++i) {
1041 blk = swp_pager_meta_ctl(object, pindex + i, 0);
1042 if (blk != blk0 + i)
1051 * SWAP_PAGER_PAGE_UNSWAPPED() - remove swap backing store related to page
1053 * This removes any associated swap backing store, whether valid or
1054 * not, from the page.
1056 * This routine is typically called when a page is made dirty, at
1057 * which point any associated swap can be freed. MADV_FREE also
1058 * calls us in a special-case situation
1060 * NOTE!!! If the page is clean and the swap was valid, the caller
1061 * should make the page dirty before calling this routine. This routine
1062 * does NOT change the m->dirty status of the page. Also: MADV_FREE
1065 * This routine may not block
1066 * This routine must be called at splvm()
1069 swap_pager_unswapped(vm_page_t m)
1072 VM_OBJECT_LOCK_ASSERT(m->object, MA_OWNED);
1073 swp_pager_meta_ctl(m->object, m->pindex, SWM_FREE);
1077 * SWAP_PAGER_GETPAGES() - bring pages in from swap
1079 * Attempt to retrieve (m, count) pages from backing store, but make
1080 * sure we retrieve at least m[reqpage]. We try to load in as large
1081 * a chunk surrounding m[reqpage] as is contiguous in swap and which
1082 * belongs to the same object.
1084 * The code is designed for asynchronous operation and
1085 * immediate-notification of 'reqpage' but tends not to be
1086 * used that way. Please do not optimize-out this algorithmic
1087 * feature, I intend to improve on it in the future.
1089 * The parent has a single vm_object_pip_add() reference prior to
1090 * calling us and we should return with the same.
1092 * The parent has BUSY'd the pages. We should return with 'm'
1093 * left busy, but the others adjusted.
1096 swap_pager_getpages(vm_object_t object, vm_page_t *m, int count, int reqpage)
1106 KASSERT(mreq->object == object,
1107 ("swap_pager_getpages: object mismatch %p/%p",
1108 object, mreq->object));
1111 * Calculate range to retrieve. The pages have already been assigned
1112 * their swapblks. We require a *contiguous* range but we know it to
1113 * not span devices. If we do not supply it, bad things
1114 * happen. Note that blk, iblk & jblk can be SWAPBLK_NONE, but the
1115 * loops are set up such that the case(s) are handled implicitly.
1117 * The swp_*() calls must be made with the object locked.
1119 blk = swp_pager_meta_ctl(mreq->object, mreq->pindex, 0);
1121 for (i = reqpage - 1; i >= 0; --i) {
1124 iblk = swp_pager_meta_ctl(m[i]->object, m[i]->pindex, 0);
1125 if (blk != iblk + (reqpage - i))
1130 for (j = reqpage + 1; j < count; ++j) {
1133 jblk = swp_pager_meta_ctl(m[j]->object, m[j]->pindex, 0);
1134 if (blk != jblk - (j - reqpage))
1139 * free pages outside our collection range. Note: we never free
1140 * mreq, it must remain busy throughout.
1142 if (0 < i || j < count) {
1145 for (k = 0; k < i; ++k)
1146 swp_pager_free_nrpage(m[k]);
1147 for (k = j; k < count; ++k)
1148 swp_pager_free_nrpage(m[k]);
1152 * Return VM_PAGER_FAIL if we have nothing to do. Return mreq
1153 * still busy, but the others unbusied.
1155 if (blk == SWAPBLK_NONE)
1156 return (VM_PAGER_FAIL);
1159 * Getpbuf() can sleep.
1161 VM_OBJECT_UNLOCK(object);
1163 * Get a swap buffer header to perform the IO
1165 bp = getpbuf(&nsw_rcount);
1166 bp->b_flags |= B_PAGING;
1169 * map our page(s) into kva for input
1171 pmap_qenter((vm_offset_t)bp->b_data, m + i, j - i);
1173 bp->b_iocmd = BIO_READ;
1174 bp->b_iodone = swp_pager_async_iodone;
1175 bp->b_rcred = crhold(thread0.td_ucred);
1176 bp->b_wcred = crhold(thread0.td_ucred);
1177 bp->b_blkno = blk - (reqpage - i);
1178 bp->b_bcount = PAGE_SIZE * (j - i);
1179 bp->b_bufsize = PAGE_SIZE * (j - i);
1180 bp->b_pager.pg_reqpage = reqpage - i;
1182 VM_OBJECT_LOCK(object);
1186 for (k = i; k < j; ++k) {
1187 bp->b_pages[k - i] = m[k];
1188 m[k]->oflags |= VPO_SWAPINPROG;
1191 bp->b_npages = j - i;
1193 PCPU_INC(cnt.v_swapin);
1194 PCPU_ADD(cnt.v_swappgsin, bp->b_npages);
1197 * We still hold the lock on mreq, and our automatic completion routine
1198 * does not remove it.
1200 vm_object_pip_add(object, bp->b_npages);
1201 VM_OBJECT_UNLOCK(object);
1204 * perform the I/O. NOTE!!! bp cannot be considered valid after
1205 * this point because we automatically release it on completion.
1206 * Instead, we look at the one page we are interested in which we
1207 * still hold a lock on even through the I/O completion.
1209 * The other pages in our m[] array are also released on completion,
1210 * so we cannot assume they are valid anymore either.
1212 * NOTE: b_blkno is destroyed by the call to swapdev_strategy
1215 swp_pager_strategy(bp);
1218 * wait for the page we want to complete. VPO_SWAPINPROG is always
1219 * cleared on completion. If an I/O error occurs, SWAPBLK_NONE
1220 * is set in the meta-data.
1222 VM_OBJECT_LOCK(object);
1223 while ((mreq->oflags & VPO_SWAPINPROG) != 0) {
1224 mreq->oflags |= VPO_WANTED;
1225 PCPU_INC(cnt.v_intrans);
1226 if (msleep(mreq, VM_OBJECT_MTX(object), PSWP, "swread", hz*20)) {
1228 "swap_pager: indefinite wait buffer: bufobj: %p, blkno: %jd, size: %ld\n",
1229 bp->b_bufobj, (intmax_t)bp->b_blkno, bp->b_bcount);
1234 * mreq is left busied after completion, but all the other pages
1235 * are freed. If we had an unrecoverable read error the page will
1238 if (mreq->valid != VM_PAGE_BITS_ALL) {
1239 return (VM_PAGER_ERROR);
1241 return (VM_PAGER_OK);
1245 * A final note: in a low swap situation, we cannot deallocate swap
1246 * and mark a page dirty here because the caller is likely to mark
1247 * the page clean when we return, causing the page to possibly revert
1248 * to all-zero's later.
1253 * swap_pager_putpages:
1255 * Assign swap (if necessary) and initiate I/O on the specified pages.
1257 * We support both OBJT_DEFAULT and OBJT_SWAP objects. DEFAULT objects
1258 * are automatically converted to SWAP objects.
1260 * In a low memory situation we may block in VOP_STRATEGY(), but the new
1261 * vm_page reservation system coupled with properly written VFS devices
1262 * should ensure that no low-memory deadlock occurs. This is an area
1265 * The parent has N vm_object_pip_add() references prior to
1266 * calling us and will remove references for rtvals[] that are
1267 * not set to VM_PAGER_PEND. We need to remove the rest on I/O
1270 * The parent has soft-busy'd the pages it passes us and will unbusy
1271 * those whos rtvals[] entry is not set to VM_PAGER_PEND on return.
1272 * We need to unbusy the rest on I/O completion.
1275 swap_pager_putpages(vm_object_t object, vm_page_t *m, int count,
1276 boolean_t sync, int *rtvals)
1281 if (count && m[0]->object != object) {
1282 panic("swap_pager_putpages: object mismatch %p/%p",
1291 * Turn object into OBJT_SWAP
1292 * check for bogus sysops
1293 * force sync if not pageout process
1295 if (object->type != OBJT_SWAP)
1296 swp_pager_meta_build(object, 0, SWAPBLK_NONE);
1297 VM_OBJECT_UNLOCK(object);
1299 if (curproc != pageproc)
1305 * Update nsw parameters from swap_async_max sysctl values.
1306 * Do not let the sysop crash the machine with bogus numbers.
1308 mtx_lock(&pbuf_mtx);
1309 if (swap_async_max != nsw_wcount_async_max) {
1315 if ((n = swap_async_max) > nswbuf / 2)
1322 * Adjust difference ( if possible ). If the current async
1323 * count is too low, we may not be able to make the adjustment
1326 n -= nsw_wcount_async_max;
1327 if (nsw_wcount_async + n >= 0) {
1328 nsw_wcount_async += n;
1329 nsw_wcount_async_max += n;
1330 wakeup(&nsw_wcount_async);
1333 mtx_unlock(&pbuf_mtx);
1338 * Assign swap blocks and issue I/O. We reallocate swap on the fly.
1339 * The page is left dirty until the pageout operation completes
1342 for (i = 0; i < count; i += n) {
1348 * Maximum I/O size is limited by a number of factors.
1350 n = min(BLIST_MAX_ALLOC, count - i);
1351 n = min(n, nsw_cluster_max);
1354 * Get biggest block of swap we can. If we fail, fall
1355 * back and try to allocate a smaller block. Don't go
1356 * overboard trying to allocate space if it would overly
1360 (blk = swp_pager_getswapspace(n)) == SWAPBLK_NONE &&
1365 if (blk == SWAPBLK_NONE) {
1366 for (j = 0; j < n; ++j)
1367 rtvals[i+j] = VM_PAGER_FAIL;
1372 * All I/O parameters have been satisfied, build the I/O
1373 * request and assign the swap space.
1376 bp = getpbuf(&nsw_wcount_sync);
1378 bp = getpbuf(&nsw_wcount_async);
1379 bp->b_flags = B_ASYNC;
1381 bp->b_flags |= B_PAGING;
1382 bp->b_iocmd = BIO_WRITE;
1384 pmap_qenter((vm_offset_t)bp->b_data, &m[i], n);
1386 bp->b_rcred = crhold(thread0.td_ucred);
1387 bp->b_wcred = crhold(thread0.td_ucred);
1388 bp->b_bcount = PAGE_SIZE * n;
1389 bp->b_bufsize = PAGE_SIZE * n;
1392 VM_OBJECT_LOCK(object);
1393 for (j = 0; j < n; ++j) {
1394 vm_page_t mreq = m[i+j];
1396 swp_pager_meta_build(
1401 vm_page_dirty(mreq);
1402 rtvals[i+j] = VM_PAGER_OK;
1404 mreq->oflags |= VPO_SWAPINPROG;
1405 bp->b_pages[j] = mreq;
1407 VM_OBJECT_UNLOCK(object);
1410 * Must set dirty range for NFS to work.
1413 bp->b_dirtyend = bp->b_bcount;
1415 PCPU_INC(cnt.v_swapout);
1416 PCPU_ADD(cnt.v_swappgsout, bp->b_npages);
1421 * NOTE: b_blkno is destroyed by the call to swapdev_strategy
1423 if (sync == FALSE) {
1424 bp->b_iodone = swp_pager_async_iodone;
1426 swp_pager_strategy(bp);
1428 for (j = 0; j < n; ++j)
1429 rtvals[i+j] = VM_PAGER_PEND;
1430 /* restart outter loop */
1437 * NOTE: b_blkno is destroyed by the call to swapdev_strategy
1439 bp->b_iodone = bdone;
1440 swp_pager_strategy(bp);
1443 * Wait for the sync I/O to complete, then update rtvals.
1444 * We just set the rtvals[] to VM_PAGER_PEND so we can call
1445 * our async completion routine at the end, thus avoiding a
1448 bwait(bp, PVM, "swwrt");
1449 for (j = 0; j < n; ++j)
1450 rtvals[i+j] = VM_PAGER_PEND;
1452 * Now that we are through with the bp, we can call the
1453 * normal async completion, which frees everything up.
1455 swp_pager_async_iodone(bp);
1457 VM_OBJECT_LOCK(object);
1461 * swp_pager_async_iodone:
1463 * Completion routine for asynchronous reads and writes from/to swap.
1464 * Also called manually by synchronous code to finish up a bp.
1466 * For READ operations, the pages are VPO_BUSY'd. For WRITE operations,
1467 * the pages are vm_page_t->busy'd. For READ operations, we VPO_BUSY
1468 * unbusy all pages except the 'main' request page. For WRITE
1469 * operations, we vm_page_t->busy'd unbusy all pages ( we can do this
1470 * because we marked them all VM_PAGER_PEND on return from putpages ).
1472 * This routine may not block.
1475 swp_pager_async_iodone(struct buf *bp)
1478 vm_object_t object = NULL;
1483 if (bp->b_ioflags & BIO_ERROR) {
1485 "swap_pager: I/O error - %s failed; blkno %ld,"
1486 "size %ld, error %d\n",
1487 ((bp->b_iocmd == BIO_READ) ? "pagein" : "pageout"),
1495 * remove the mapping for kernel virtual
1497 pmap_qremove((vm_offset_t)bp->b_data, bp->b_npages);
1500 object = bp->b_pages[0]->object;
1501 VM_OBJECT_LOCK(object);
1505 * cleanup pages. If an error occurs writing to swap, we are in
1506 * very serious trouble. If it happens to be a disk error, though,
1507 * we may be able to recover by reassigning the swap later on. So
1508 * in this case we remove the m->swapblk assignment for the page
1509 * but do not free it in the rlist. The errornous block(s) are thus
1510 * never reallocated as swap. Redirty the page and continue.
1512 for (i = 0; i < bp->b_npages; ++i) {
1513 vm_page_t m = bp->b_pages[i];
1515 m->oflags &= ~VPO_SWAPINPROG;
1517 if (bp->b_ioflags & BIO_ERROR) {
1519 * If an error occurs I'd love to throw the swapblk
1520 * away without freeing it back to swapspace, so it
1521 * can never be used again. But I can't from an
1524 if (bp->b_iocmd == BIO_READ) {
1526 * When reading, reqpage needs to stay
1527 * locked for the parent, but all other
1528 * pages can be freed. We still want to
1529 * wakeup the parent waiting on the page,
1530 * though. ( also: pg_reqpage can be -1 and
1531 * not match anything ).
1533 * We have to wake specifically requested pages
1534 * up too because we cleared VPO_SWAPINPROG and
1535 * someone may be waiting for that.
1537 * NOTE: for reads, m->dirty will probably
1538 * be overridden by the original caller of
1539 * getpages so don't play cute tricks here.
1542 if (i != bp->b_pager.pg_reqpage)
1543 swp_pager_free_nrpage(m);
1547 * If i == bp->b_pager.pg_reqpage, do not wake
1548 * the page up. The caller needs to.
1552 * If a write error occurs, reactivate page
1553 * so it doesn't clog the inactive list,
1554 * then finish the I/O.
1558 vm_page_activate(m);
1560 vm_page_io_finish(m);
1562 } else if (bp->b_iocmd == BIO_READ) {
1564 * NOTE: for reads, m->dirty will probably be
1565 * overridden by the original caller of getpages so
1566 * we cannot set them in order to free the underlying
1567 * swap in a low-swap situation. I don't think we'd
1568 * want to do that anyway, but it was an optimization
1569 * that existed in the old swapper for a time before
1570 * it got ripped out due to precisely this problem.
1572 * If not the requested page then deactivate it.
1574 * Note that the requested page, reqpage, is left
1575 * busied, but we still have to wake it up. The
1576 * other pages are released (unbusied) by
1579 KASSERT(!pmap_page_is_mapped(m),
1580 ("swp_pager_async_iodone: page %p is mapped", m));
1581 m->valid = VM_PAGE_BITS_ALL;
1582 KASSERT(m->dirty == 0,
1583 ("swp_pager_async_iodone: page %p is dirty", m));
1586 * We have to wake specifically requested pages
1587 * up too because we cleared VPO_SWAPINPROG and
1588 * could be waiting for it in getpages. However,
1589 * be sure to not unbusy getpages specifically
1590 * requested page - getpages expects it to be
1593 if (i != bp->b_pager.pg_reqpage) {
1595 vm_page_deactivate(m);
1602 * For write success, clear the dirty
1603 * status, then finish the I/O ( which decrements the
1604 * busy count and possibly wakes waiter's up ).
1606 KASSERT((m->flags & PG_WRITEABLE) == 0,
1607 ("swp_pager_async_iodone: page %p is not write"
1610 vm_page_io_finish(m);
1611 if (vm_page_count_severe()) {
1613 vm_page_try_to_cache(m);
1620 * adjust pip. NOTE: the original parent may still have its own
1621 * pip refs on the object.
1623 if (object != NULL) {
1624 vm_object_pip_wakeupn(object, bp->b_npages);
1625 VM_OBJECT_UNLOCK(object);
1629 * swapdev_strategy() manually sets b_vp and b_bufobj before calling
1630 * bstrategy(). Set them back to NULL now we're done with it, or we'll
1631 * trigger a KASSERT in relpbuf().
1635 bp->b_bufobj = NULL;
1638 * release the physical I/O buffer
1642 ((bp->b_iocmd == BIO_READ) ? &nsw_rcount :
1643 ((bp->b_flags & B_ASYNC) ?
1652 * swap_pager_isswapped:
1654 * Return 1 if at least one page in the given object is paged
1655 * out to the given swap device.
1657 * This routine may not block.
1660 swap_pager_isswapped(vm_object_t object, struct swdevt *sp)
1666 VM_OBJECT_LOCK_ASSERT(object, MA_OWNED);
1667 if (object->type != OBJT_SWAP)
1670 mtx_lock(&swhash_mtx);
1671 for (bcount = 0; bcount < object->un_pager.swp.swp_bcount; bcount++) {
1672 struct swblock *swap;
1674 if ((swap = *swp_pager_hash(object, index)) != NULL) {
1675 for (i = 0; i < SWAP_META_PAGES; ++i) {
1676 if (swp_pager_isondev(swap->swb_pages[i], sp)) {
1677 mtx_unlock(&swhash_mtx);
1682 index += SWAP_META_PAGES;
1684 mtx_unlock(&swhash_mtx);
1689 * SWP_PAGER_FORCE_PAGEIN() - force a swap block to be paged in
1691 * This routine dissociates the page at the given index within a
1692 * swap block from its backing store, paging it in if necessary.
1693 * If the page is paged in, it is placed in the inactive queue,
1694 * since it had its backing store ripped out from under it.
1695 * We also attempt to swap in all other pages in the swap block,
1696 * we only guarantee that the one at the specified index is
1699 * XXX - The code to page the whole block in doesn't work, so we
1700 * revert to the one-by-one behavior for now. Sigh.
1703 swp_pager_force_pagein(vm_object_t object, vm_pindex_t pindex)
1707 vm_object_pip_add(object, 1);
1708 m = vm_page_grab(object, pindex, VM_ALLOC_NORMAL|VM_ALLOC_RETRY);
1709 if (m->valid == VM_PAGE_BITS_ALL) {
1710 vm_object_pip_subtract(object, 1);
1713 vm_page_activate(m);
1716 vm_pager_page_unswapped(m);
1720 if (swap_pager_getpages(object, &m, 1, 0) != VM_PAGER_OK)
1721 panic("swap_pager_force_pagein: read from swap failed");/*XXX*/
1722 vm_object_pip_subtract(object, 1);
1725 vm_page_deactivate(m);
1728 vm_pager_page_unswapped(m);
1732 * swap_pager_swapoff:
1734 * Page in all of the pages that have been paged out to the
1735 * given device. The corresponding blocks in the bitmap must be
1736 * marked as allocated and the device must be flagged SW_CLOSING.
1737 * There may be no processes swapped out to the device.
1739 * This routine may block.
1742 swap_pager_swapoff(struct swdevt *sp)
1744 struct swblock *swap;
1751 mtx_lock(&swhash_mtx);
1752 for (i = 0; i <= swhash_mask; i++) { /* '<=' is correct here */
1754 for (swap = swhash[i]; swap != NULL; swap = swap->swb_hnext) {
1755 vm_object_t object = swap->swb_object;
1756 vm_pindex_t pindex = swap->swb_index;
1757 for (j = 0; j < SWAP_META_PAGES; ++j) {
1758 if (swp_pager_isondev(swap->swb_pages[j], sp)) {
1759 /* avoid deadlock */
1760 if (!VM_OBJECT_TRYLOCK(object)) {
1763 mtx_unlock(&swhash_mtx);
1764 swp_pager_force_pagein(object,
1766 VM_OBJECT_UNLOCK(object);
1767 mtx_lock(&swhash_mtx);
1774 mtx_unlock(&swhash_mtx);
1777 * Objects may be locked or paging to the device being
1778 * removed, so we will miss their pages and need to
1779 * make another pass. We have marked this device as
1780 * SW_CLOSING, so the activity should finish soon.
1783 if (retries > 100) {
1784 panic("swapoff: failed to locate %d swap blocks",
1787 pause("swpoff", hz / 20);
1792 /************************************************************************
1794 ************************************************************************
1796 * These routines manipulate the swap metadata stored in the
1797 * OBJT_SWAP object. All swp_*() routines must be called at
1798 * splvm() because swap can be freed up by the low level vm_page
1799 * code which might be called from interrupts beyond what splbio() covers.
1801 * Swap metadata is implemented with a global hash and not directly
1802 * linked into the object. Instead the object simply contains
1803 * appropriate tracking counters.
1807 * SWP_PAGER_META_BUILD() - add swap block to swap meta data for object
1809 * We first convert the object to a swap object if it is a default
1812 * The specified swapblk is added to the object's swap metadata. If
1813 * the swapblk is not valid, it is freed instead. Any previously
1814 * assigned swapblk is freed.
1816 * This routine must be called at splvm(), except when used to convert
1817 * an OBJT_DEFAULT object into an OBJT_SWAP object.
1820 swp_pager_meta_build(vm_object_t object, vm_pindex_t pindex, daddr_t swapblk)
1822 struct swblock *swap;
1823 struct swblock **pswap;
1826 VM_OBJECT_LOCK_ASSERT(object, MA_OWNED);
1828 * Convert default object to swap object if necessary
1830 if (object->type != OBJT_SWAP) {
1831 object->type = OBJT_SWAP;
1832 object->un_pager.swp.swp_bcount = 0;
1834 if (object->handle != NULL) {
1835 mtx_lock(&sw_alloc_mtx);
1837 NOBJLIST(object->handle),
1841 mtx_unlock(&sw_alloc_mtx);
1846 * Locate hash entry. If not found create, but if we aren't adding
1847 * anything just return. If we run out of space in the map we wait
1848 * and, since the hash table may have changed, retry.
1851 mtx_lock(&swhash_mtx);
1852 pswap = swp_pager_hash(object, pindex);
1854 if ((swap = *pswap) == NULL) {
1857 if (swapblk == SWAPBLK_NONE)
1860 swap = *pswap = uma_zalloc(swap_zone, M_NOWAIT);
1862 mtx_unlock(&swhash_mtx);
1863 VM_OBJECT_UNLOCK(object);
1864 if (uma_zone_exhausted(swap_zone)) {
1865 printf("swap zone exhausted, increase kern.maxswzone\n");
1866 vm_pageout_oom(VM_OOM_SWAPZ);
1867 pause("swzonex", 10);
1870 VM_OBJECT_LOCK(object);
1874 swap->swb_hnext = NULL;
1875 swap->swb_object = object;
1876 swap->swb_index = pindex & ~(vm_pindex_t)SWAP_META_MASK;
1877 swap->swb_count = 0;
1879 ++object->un_pager.swp.swp_bcount;
1881 for (i = 0; i < SWAP_META_PAGES; ++i)
1882 swap->swb_pages[i] = SWAPBLK_NONE;
1886 * Delete prior contents of metadata
1888 idx = pindex & SWAP_META_MASK;
1890 if (swap->swb_pages[idx] != SWAPBLK_NONE) {
1891 swp_pager_freeswapspace(swap->swb_pages[idx], 1);
1896 * Enter block into metadata
1898 swap->swb_pages[idx] = swapblk;
1899 if (swapblk != SWAPBLK_NONE)
1902 mtx_unlock(&swhash_mtx);
1906 * SWP_PAGER_META_FREE() - free a range of blocks in the object's swap metadata
1908 * The requested range of blocks is freed, with any associated swap
1909 * returned to the swap bitmap.
1911 * This routine will free swap metadata structures as they are cleaned
1912 * out. This routine does *NOT* operate on swap metadata associated
1913 * with resident pages.
1915 * This routine must be called at splvm()
1918 swp_pager_meta_free(vm_object_t object, vm_pindex_t index, daddr_t count)
1921 VM_OBJECT_LOCK_ASSERT(object, MA_OWNED);
1922 if (object->type != OBJT_SWAP)
1926 struct swblock **pswap;
1927 struct swblock *swap;
1929 mtx_lock(&swhash_mtx);
1930 pswap = swp_pager_hash(object, index);
1932 if ((swap = *pswap) != NULL) {
1933 daddr_t v = swap->swb_pages[index & SWAP_META_MASK];
1935 if (v != SWAPBLK_NONE) {
1936 swp_pager_freeswapspace(v, 1);
1937 swap->swb_pages[index & SWAP_META_MASK] =
1939 if (--swap->swb_count == 0) {
1940 *pswap = swap->swb_hnext;
1941 uma_zfree(swap_zone, swap);
1942 --object->un_pager.swp.swp_bcount;
1948 int n = SWAP_META_PAGES - (index & SWAP_META_MASK);
1952 mtx_unlock(&swhash_mtx);
1957 * SWP_PAGER_META_FREE_ALL() - destroy all swap metadata associated with object
1959 * This routine locates and destroys all swap metadata associated with
1962 * This routine must be called at splvm()
1965 swp_pager_meta_free_all(vm_object_t object)
1969 VM_OBJECT_LOCK_ASSERT(object, MA_OWNED);
1970 if (object->type != OBJT_SWAP)
1973 while (object->un_pager.swp.swp_bcount) {
1974 struct swblock **pswap;
1975 struct swblock *swap;
1977 mtx_lock(&swhash_mtx);
1978 pswap = swp_pager_hash(object, index);
1979 if ((swap = *pswap) != NULL) {
1982 for (i = 0; i < SWAP_META_PAGES; ++i) {
1983 daddr_t v = swap->swb_pages[i];
1984 if (v != SWAPBLK_NONE) {
1986 swp_pager_freeswapspace(v, 1);
1989 if (swap->swb_count != 0)
1990 panic("swap_pager_meta_free_all: swb_count != 0");
1991 *pswap = swap->swb_hnext;
1992 uma_zfree(swap_zone, swap);
1993 --object->un_pager.swp.swp_bcount;
1995 mtx_unlock(&swhash_mtx);
1996 index += SWAP_META_PAGES;
2001 * SWP_PAGER_METACTL() - misc control of swap and vm_page_t meta data.
2003 * This routine is capable of looking up, popping, or freeing
2004 * swapblk assignments in the swap meta data or in the vm_page_t.
2005 * The routine typically returns the swapblk being looked-up, or popped,
2006 * or SWAPBLK_NONE if the block was freed, or SWAPBLK_NONE if the block
2007 * was invalid. This routine will automatically free any invalid
2008 * meta-data swapblks.
2010 * It is not possible to store invalid swapblks in the swap meta data
2011 * (other then a literal 'SWAPBLK_NONE'), so we don't bother checking.
2013 * When acting on a busy resident page and paging is in progress, we
2014 * have to wait until paging is complete but otherwise can act on the
2017 * This routine must be called at splvm().
2019 * SWM_FREE remove and free swap block from metadata
2020 * SWM_POP remove from meta data but do not free.. pop it out
2023 swp_pager_meta_ctl(vm_object_t object, vm_pindex_t pindex, int flags)
2025 struct swblock **pswap;
2026 struct swblock *swap;
2030 VM_OBJECT_LOCK_ASSERT(object, MA_OWNED);
2032 * The meta data only exists of the object is OBJT_SWAP
2033 * and even then might not be allocated yet.
2035 if (object->type != OBJT_SWAP)
2036 return (SWAPBLK_NONE);
2039 mtx_lock(&swhash_mtx);
2040 pswap = swp_pager_hash(object, pindex);
2042 if ((swap = *pswap) != NULL) {
2043 idx = pindex & SWAP_META_MASK;
2044 r1 = swap->swb_pages[idx];
2046 if (r1 != SWAPBLK_NONE) {
2047 if (flags & SWM_FREE) {
2048 swp_pager_freeswapspace(r1, 1);
2051 if (flags & (SWM_FREE|SWM_POP)) {
2052 swap->swb_pages[idx] = SWAPBLK_NONE;
2053 if (--swap->swb_count == 0) {
2054 *pswap = swap->swb_hnext;
2055 uma_zfree(swap_zone, swap);
2056 --object->un_pager.swp.swp_bcount;
2061 mtx_unlock(&swhash_mtx);
2066 * System call swapon(name) enables swapping on device name,
2067 * which must be in the swdevsw. Return EBUSY
2068 * if already swapping on this device.
2070 #ifndef _SYS_SYSPROTO_H_
2071 struct swapon_args {
2081 swapon(struct thread *td, struct swapon_args *uap)
2085 struct nameidata nd;
2088 error = priv_check(td, PRIV_SWAPON);
2093 while (swdev_syscall_active)
2094 tsleep(&swdev_syscall_active, PUSER - 1, "swpon", 0);
2095 swdev_syscall_active = 1;
2098 * Swap metadata may not fit in the KVM if we have physical
2101 if (swap_zone == NULL) {
2106 NDINIT(&nd, LOOKUP, ISOPEN | FOLLOW | AUDITVNODE1, UIO_USERSPACE,
2112 NDFREE(&nd, NDF_ONLY_PNBUF);
2115 if (vn_isdisk(vp, &error)) {
2116 error = swapongeom(td, vp);
2117 } else if (vp->v_type == VREG &&
2118 (vp->v_mount->mnt_vfc->vfc_flags & VFCF_NETWORK) != 0 &&
2119 (error = VOP_GETATTR(vp, &attr, td->td_ucred)) == 0) {
2121 * Allow direct swapping to NFS regular files in the same
2122 * way that nfs_mountroot() sets up diskless swapping.
2124 error = swaponvp(td, vp, attr.va_size / DEV_BSIZE);
2130 swdev_syscall_active = 0;
2131 wakeup_one(&swdev_syscall_active);
2137 swaponsomething(struct vnode *vp, void *id, u_long nblks, sw_strategy_t *strategy, sw_close_t *close, dev_t dev)
2139 struct swdevt *sp, *tsp;
2144 * If we go beyond this, we get overflows in the radix
2147 mblocks = 0x40000000 / BLIST_META_RADIX;
2148 if (nblks > mblocks) {
2149 printf("WARNING: reducing size to maximum of %lu blocks per swap unit\n",
2154 * nblks is in DEV_BSIZE'd chunks, convert to PAGE_SIZE'd chunks.
2155 * First chop nblks off to page-align it, then convert.
2157 * sw->sw_nblks is in page-sized chunks now too.
2159 nblks &= ~(ctodb(1) - 1);
2160 nblks = dbtoc(nblks);
2162 sp = malloc(sizeof *sp, M_VMPGDATA, M_WAITOK | M_ZERO);
2167 sp->sw_nblks = nblks;
2169 sp->sw_strategy = strategy;
2170 sp->sw_close = close;
2172 sp->sw_blist = blist_create(nblks, M_WAITOK);
2174 * Do not free the first two block in order to avoid overwriting
2175 * any bsd label at the front of the partition
2177 blist_free(sp->sw_blist, 2, nblks - 2);
2180 mtx_lock(&sw_dev_mtx);
2181 TAILQ_FOREACH(tsp, &swtailq, sw_list) {
2182 if (tsp->sw_end >= dvbase) {
2184 * We put one uncovered page between the devices
2185 * in order to definitively prevent any cross-device
2188 dvbase = tsp->sw_end + 1;
2191 sp->sw_first = dvbase;
2192 sp->sw_end = dvbase + nblks;
2193 TAILQ_INSERT_TAIL(&swtailq, sp, sw_list);
2195 swap_pager_avail += nblks;
2196 swap_total += (vm_ooffset_t)nblks * PAGE_SIZE;
2198 mtx_unlock(&sw_dev_mtx);
2202 * SYSCALL: swapoff(devname)
2204 * Disable swapping on the given device.
2206 * XXX: Badly designed system call: it should use a device index
2207 * rather than filename as specification. We keep sw_vp around
2208 * only to make this work.
2210 #ifndef _SYS_SYSPROTO_H_
2211 struct swapoff_args {
2221 swapoff(struct thread *td, struct swapoff_args *uap)
2224 struct nameidata nd;
2228 error = priv_check(td, PRIV_SWAPOFF);
2233 while (swdev_syscall_active)
2234 tsleep(&swdev_syscall_active, PUSER - 1, "swpoff", 0);
2235 swdev_syscall_active = 1;
2237 NDINIT(&nd, LOOKUP, FOLLOW | AUDITVNODE1, UIO_USERSPACE, uap->name,
2242 NDFREE(&nd, NDF_ONLY_PNBUF);
2245 mtx_lock(&sw_dev_mtx);
2246 TAILQ_FOREACH(sp, &swtailq, sw_list) {
2247 if (sp->sw_vp == vp)
2250 mtx_unlock(&sw_dev_mtx);
2255 error = swapoff_one(sp, td->td_ucred);
2257 swdev_syscall_active = 0;
2258 wakeup_one(&swdev_syscall_active);
2264 swapoff_one(struct swdevt *sp, struct ucred *cred)
2266 u_long nblks, dvbase;
2271 mtx_assert(&Giant, MA_OWNED);
2273 (void) vn_lock(sp->sw_vp, LK_EXCLUSIVE | LK_RETRY);
2274 error = mac_system_check_swapoff(cred, sp->sw_vp);
2275 (void) VOP_UNLOCK(sp->sw_vp, 0);
2279 nblks = sp->sw_nblks;
2282 * We can turn off this swap device safely only if the
2283 * available virtual memory in the system will fit the amount
2284 * of data we will have to page back in, plus an epsilon so
2285 * the system doesn't become critically low on swap space.
2287 if (cnt.v_free_count + cnt.v_cache_count + swap_pager_avail <
2288 nblks + nswap_lowat) {
2293 * Prevent further allocations on this device.
2295 mtx_lock(&sw_dev_mtx);
2296 sp->sw_flags |= SW_CLOSING;
2297 for (dvbase = 0; dvbase < sp->sw_end; dvbase += dmmax) {
2298 swap_pager_avail -= blist_fill(sp->sw_blist,
2301 swap_total -= (vm_ooffset_t)nblks * PAGE_SIZE;
2302 mtx_unlock(&sw_dev_mtx);
2305 * Page in the contents of the device and close it.
2307 swap_pager_swapoff(sp);
2309 sp->sw_close(curthread, sp);
2311 mtx_lock(&sw_dev_mtx);
2312 TAILQ_REMOVE(&swtailq, sp, sw_list);
2314 if (nswapdev == 0) {
2315 swap_pager_full = 2;
2316 swap_pager_almost_full = 1;
2320 mtx_unlock(&sw_dev_mtx);
2321 blist_destroy(sp->sw_blist);
2322 free(sp, M_VMPGDATA);
2329 struct swdevt *sp, *spt;
2330 const char *devname;
2334 while (swdev_syscall_active)
2335 tsleep(&swdev_syscall_active, PUSER - 1, "swpoff", 0);
2336 swdev_syscall_active = 1;
2338 mtx_lock(&sw_dev_mtx);
2339 TAILQ_FOREACH_SAFE(sp, &swtailq, sw_list, spt) {
2340 mtx_unlock(&sw_dev_mtx);
2341 if (vn_isdisk(sp->sw_vp, NULL))
2342 devname = sp->sw_vp->v_rdev->si_name;
2345 error = swapoff_one(sp, thread0.td_ucred);
2347 printf("Cannot remove swap device %s (error=%d), "
2348 "skipping.\n", devname, error);
2349 } else if (bootverbose) {
2350 printf("Swap device %s removed.\n", devname);
2352 mtx_lock(&sw_dev_mtx);
2354 mtx_unlock(&sw_dev_mtx);
2356 swdev_syscall_active = 0;
2357 wakeup_one(&swdev_syscall_active);
2362 swap_pager_status(int *total, int *used)
2368 mtx_lock(&sw_dev_mtx);
2369 TAILQ_FOREACH(sp, &swtailq, sw_list) {
2370 *total += sp->sw_nblks;
2371 *used += sp->sw_used;
2373 mtx_unlock(&sw_dev_mtx);
2377 sysctl_vm_swap_info(SYSCTL_HANDLER_ARGS)
2379 int *name = (int *)arg1;
2384 if (arg2 != 1) /* name length */
2388 mtx_lock(&sw_dev_mtx);
2389 TAILQ_FOREACH(sp, &swtailq, sw_list) {
2391 mtx_unlock(&sw_dev_mtx);
2392 xs.xsw_version = XSWDEV_VERSION;
2393 xs.xsw_dev = sp->sw_dev;
2394 xs.xsw_flags = sp->sw_flags;
2395 xs.xsw_nblks = sp->sw_nblks;
2396 xs.xsw_used = sp->sw_used;
2398 error = SYSCTL_OUT(req, &xs, sizeof(xs));
2403 mtx_unlock(&sw_dev_mtx);
2407 SYSCTL_INT(_vm, OID_AUTO, nswapdev, CTLFLAG_RD, &nswapdev, 0,
2408 "Number of swap devices");
2409 SYSCTL_NODE(_vm, OID_AUTO, swap_info, CTLFLAG_RD, sysctl_vm_swap_info,
2410 "Swap statistics by device");
2413 * vmspace_swap_count() - count the approximate swap usage in pages for a
2416 * The map must be locked.
2418 * Swap usage is determined by taking the proportional swap used by
2419 * VM objects backing the VM map. To make up for fractional losses,
2420 * if the VM object has any swap use at all the associated map entries
2421 * count for at least 1 swap page.
2424 vmspace_swap_count(struct vmspace *vmspace)
2431 map = &vmspace->vm_map;
2434 for (cur = map->header.next; cur != &map->header; cur = cur->next) {
2435 if ((cur->eflags & MAP_ENTRY_IS_SUB_MAP) == 0 &&
2436 (object = cur->object.vm_object) != NULL) {
2437 VM_OBJECT_LOCK(object);
2438 if (object->type == OBJT_SWAP &&
2439 object->un_pager.swp.swp_bcount != 0) {
2440 n = (cur->end - cur->start) / PAGE_SIZE;
2441 count += object->un_pager.swp.swp_bcount *
2442 SWAP_META_PAGES * n / object->size + 1;
2444 VM_OBJECT_UNLOCK(object);
2453 * Swapping onto disk devices.
2457 static g_orphan_t swapgeom_orphan;
2459 static struct g_class g_swap_class = {
2461 .version = G_VERSION,
2462 .orphan = swapgeom_orphan,
2465 DECLARE_GEOM_CLASS(g_swap_class, g_class);
2469 swapgeom_done(struct bio *bp2)
2473 bp = bp2->bio_caller2;
2474 bp->b_ioflags = bp2->bio_flags;
2476 bp->b_ioflags |= BIO_ERROR;
2477 bp->b_resid = bp->b_bcount - bp2->bio_completed;
2478 bp->b_error = bp2->bio_error;
2484 swapgeom_strategy(struct buf *bp, struct swdevt *sp)
2487 struct g_consumer *cp;
2491 bp->b_error = ENXIO;
2492 bp->b_ioflags |= BIO_ERROR;
2496 if (bp->b_iocmd == BIO_WRITE)
2499 bio = g_alloc_bio();
2501 bp->b_error = ENOMEM;
2502 bp->b_ioflags |= BIO_ERROR;
2507 bio->bio_caller2 = bp;
2508 bio->bio_cmd = bp->b_iocmd;
2509 bio->bio_data = bp->b_data;
2510 bio->bio_offset = (bp->b_blkno - sp->sw_first) * PAGE_SIZE;
2511 bio->bio_length = bp->b_bcount;
2512 bio->bio_done = swapgeom_done;
2513 g_io_request(bio, cp);
2518 swapgeom_orphan(struct g_consumer *cp)
2522 mtx_lock(&sw_dev_mtx);
2523 TAILQ_FOREACH(sp, &swtailq, sw_list)
2524 if (sp->sw_id == cp)
2526 mtx_unlock(&sw_dev_mtx);
2530 swapgeom_close_ev(void *arg, int flags)
2532 struct g_consumer *cp;
2535 g_access(cp, -1, -1, 0);
2537 g_destroy_consumer(cp);
2541 swapgeom_close(struct thread *td, struct swdevt *sw)
2544 /* XXX: direct call when Giant untangled */
2545 g_waitfor_event(swapgeom_close_ev, sw->sw_id, M_WAITOK, NULL);
2556 swapongeom_ev(void *arg, int flags)
2559 struct g_provider *pp;
2560 struct g_consumer *cp;
2561 static struct g_geom *gp;
2568 pp = g_dev_getprovider(swh->dev);
2570 swh->error = ENODEV;
2573 mtx_lock(&sw_dev_mtx);
2574 TAILQ_FOREACH(sp, &swtailq, sw_list) {
2576 if (cp != NULL && cp->provider == pp) {
2577 mtx_unlock(&sw_dev_mtx);
2582 mtx_unlock(&sw_dev_mtx);
2584 gp = g_new_geomf(&g_swap_class, "swap", NULL);
2585 cp = g_new_consumer(gp);
2588 * XXX: Everytime you think you can improve the margin for
2589 * footshooting, somebody depends on the ability to do so:
2590 * savecore(8) wants to write to our swapdev so we cannot
2591 * set an exclusive count :-(
2593 error = g_access(cp, 1, 1, 0);
2596 g_destroy_consumer(cp);
2600 nblks = pp->mediasize / DEV_BSIZE;
2601 swaponsomething(swh->vp, cp, nblks, swapgeom_strategy,
2602 swapgeom_close, dev2udev(swh->dev));
2608 swapongeom(struct thread *td, struct vnode *vp)
2613 vn_lock(vp, LK_EXCLUSIVE | LK_RETRY);
2615 swh.dev = vp->v_rdev;
2618 /* XXX: direct call when Giant untangled */
2619 error = g_waitfor_event(swapongeom_ev, &swh, M_WAITOK, NULL);
2629 * This is used mainly for network filesystem (read: probably only tested
2630 * with NFS) swapfiles.
2635 swapdev_strategy(struct buf *bp, struct swdevt *sp)
2639 bp->b_blkno = ctodb(bp->b_blkno - sp->sw_first);
2643 if (bp->b_iocmd == BIO_WRITE) {
2645 bufobj_wdrop(bp->b_bufobj);
2646 bufobj_wref(&vp2->v_bufobj);
2648 if (bp->b_bufobj != &vp2->v_bufobj)
2649 bp->b_bufobj = &vp2->v_bufobj;
2651 bp->b_iooffset = dbtob(bp->b_blkno);
2657 swapdev_close(struct thread *td, struct swdevt *sp)
2660 VOP_CLOSE(sp->sw_vp, FREAD | FWRITE, td->td_ucred, td);
2666 swaponvp(struct thread *td, struct vnode *vp, u_long nblks)
2673 mtx_lock(&sw_dev_mtx);
2674 TAILQ_FOREACH(sp, &swtailq, sw_list) {
2675 if (sp->sw_id == vp) {
2676 mtx_unlock(&sw_dev_mtx);
2680 mtx_unlock(&sw_dev_mtx);
2682 (void) vn_lock(vp, LK_EXCLUSIVE | LK_RETRY);
2684 error = mac_system_check_swapon(td->td_ucred, vp);
2687 error = VOP_OPEN(vp, FREAD | FWRITE, td->td_ucred, td, NULL);
2688 (void) VOP_UNLOCK(vp, 0);
2692 swaponsomething(vp, vp, nblks, swapdev_strategy, swapdev_close,