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/racct.h>
90 #include <sys/resource.h>
91 #include <sys/resourcevar.h>
92 #include <sys/sysctl.h>
93 #include <sys/sysproto.h>
94 #include <sys/blist.h>
97 #include <sys/vmmeter.h>
99 #include <security/mac/mac_framework.h>
103 #include <vm/vm_map.h>
104 #include <vm/vm_kern.h>
105 #include <vm/vm_object.h>
106 #include <vm/vm_page.h>
107 #include <vm/vm_pager.h>
108 #include <vm/vm_pageout.h>
109 #include <vm/vm_param.h>
110 #include <vm/swap_pager.h>
111 #include <vm/vm_extern.h>
114 #include <geom/geom.h>
117 * SWB_NPAGES must be a power of 2. It may be set to 1, 2, 4, 8, or 16
118 * pages per allocation. We recommend you stick with the default of 8.
119 * The 16-page limit is due to the radix code (kern/subr_blist.c).
121 #ifndef MAX_PAGEOUT_CLUSTER
122 #define MAX_PAGEOUT_CLUSTER 16
125 #if !defined(SWB_NPAGES)
126 #define SWB_NPAGES MAX_PAGEOUT_CLUSTER
130 * Piecemeal swap metadata structure. Swap is stored in a radix tree.
132 * If SWB_NPAGES is 8 and sizeof(char *) == sizeof(daddr_t), our radix
133 * is basically 8. Assuming PAGE_SIZE == 4096, one tree level represents
134 * 32K worth of data, two levels represent 256K, three levels represent
135 * 2 MBytes. This is acceptable.
137 * Overall memory utilization is about the same as the old swap structure.
139 #define SWCORRECT(n) (sizeof(void *) * (n) / sizeof(daddr_t))
140 #define SWAP_META_PAGES (SWB_NPAGES * 2)
141 #define SWAP_META_MASK (SWAP_META_PAGES - 1)
144 struct swblock *swb_hnext;
145 vm_object_t swb_object;
146 vm_pindex_t swb_index;
148 daddr_t swb_pages[SWAP_META_PAGES];
151 static MALLOC_DEFINE(M_VMPGDATA, "vm_pgdata", "swap pager private data");
152 static struct mtx sw_dev_mtx;
153 static TAILQ_HEAD(, swdevt) swtailq = TAILQ_HEAD_INITIALIZER(swtailq);
154 static struct swdevt *swdevhd; /* Allocate from here next */
155 static int nswapdev; /* Number of swap devices */
156 int swap_pager_avail;
157 static int swdev_syscall_active = 0; /* serialize swap(on|off) */
159 static vm_ooffset_t swap_total;
160 SYSCTL_QUAD(_vm, OID_AUTO, swap_total, CTLFLAG_RD, &swap_total, 0,
161 "Total amount of available swap storage.");
162 static vm_ooffset_t swap_reserved;
163 SYSCTL_QUAD(_vm, OID_AUTO, swap_reserved, CTLFLAG_RD, &swap_reserved, 0,
164 "Amount of swap storage needed to back all allocated anonymous memory.");
165 static int overcommit = 0;
166 SYSCTL_INT(_vm, OID_AUTO, overcommit, CTLFLAG_RW, &overcommit, 0,
167 "Configure virtual memory overcommit behavior. See tuning(7) "
170 /* bits from overcommit */
171 #define SWAP_RESERVE_FORCE_ON (1 << 0)
172 #define SWAP_RESERVE_RLIMIT_ON (1 << 1)
173 #define SWAP_RESERVE_ALLOW_NONWIRED (1 << 2)
176 swap_reserve(vm_ooffset_t incr)
179 return (swap_reserve_by_cred(incr, curthread->td_ucred));
183 swap_reserve_by_cred(vm_ooffset_t incr, struct ucred *cred)
188 static struct timeval lastfail;
191 uip = cred->cr_ruidinfo;
193 if (incr & PAGE_MASK)
194 panic("swap_reserve: & PAGE_MASK");
197 error = racct_add(curproc, RACCT_SWAP, incr);
198 PROC_UNLOCK(curproc);
203 mtx_lock(&sw_dev_mtx);
204 r = swap_reserved + incr;
205 if (overcommit & SWAP_RESERVE_ALLOW_NONWIRED) {
206 s = cnt.v_page_count - cnt.v_free_reserved - cnt.v_wire_count;
211 if ((overcommit & SWAP_RESERVE_FORCE_ON) == 0 || r <= s ||
212 (error = priv_check(curthread, PRIV_VM_SWAP_NOQUOTA)) == 0) {
216 mtx_unlock(&sw_dev_mtx);
220 UIDINFO_VMSIZE_LOCK(uip);
221 if ((overcommit & SWAP_RESERVE_RLIMIT_ON) != 0 &&
222 uip->ui_vmsize + incr > lim_cur(curproc, RLIMIT_SWAP) &&
223 priv_check(curthread, PRIV_VM_SWAP_NORLIMIT))
226 uip->ui_vmsize += incr;
227 UIDINFO_VMSIZE_UNLOCK(uip);
228 PROC_UNLOCK(curproc);
230 mtx_lock(&sw_dev_mtx);
231 swap_reserved -= incr;
232 mtx_unlock(&sw_dev_mtx);
235 if (!res && ppsratecheck(&lastfail, &curfail, 1)) {
236 printf("uid %d, pid %d: swap reservation for %jd bytes failed\n",
237 curproc->p_pid, uip->ui_uid, incr);
242 racct_sub(curproc, RACCT_SWAP, incr);
243 PROC_UNLOCK(curproc);
250 swap_reserve_force(vm_ooffset_t incr)
254 mtx_lock(&sw_dev_mtx);
255 swap_reserved += incr;
256 mtx_unlock(&sw_dev_mtx);
259 racct_add_force(curproc, RACCT_SWAP, incr);
260 PROC_UNLOCK(curproc);
262 uip = curthread->td_ucred->cr_ruidinfo;
264 UIDINFO_VMSIZE_LOCK(uip);
265 uip->ui_vmsize += incr;
266 UIDINFO_VMSIZE_UNLOCK(uip);
267 PROC_UNLOCK(curproc);
271 swap_release(vm_ooffset_t decr)
276 cred = curthread->td_ucred;
277 swap_release_by_cred(decr, cred);
278 PROC_UNLOCK(curproc);
282 swap_release_by_cred(vm_ooffset_t decr, struct ucred *cred)
286 uip = cred->cr_ruidinfo;
288 if (decr & PAGE_MASK)
289 panic("swap_release: & PAGE_MASK");
291 mtx_lock(&sw_dev_mtx);
292 if (swap_reserved < decr)
293 panic("swap_reserved < decr");
294 swap_reserved -= decr;
295 mtx_unlock(&sw_dev_mtx);
297 UIDINFO_VMSIZE_LOCK(uip);
298 if (uip->ui_vmsize < decr)
299 printf("negative vmsize for uid = %d\n", uip->ui_uid);
300 uip->ui_vmsize -= decr;
301 UIDINFO_VMSIZE_UNLOCK(uip);
303 racct_sub_cred(cred, RACCT_SWAP, decr);
306 static void swapdev_strategy(struct buf *, struct swdevt *sw);
308 #define SWM_FREE 0x02 /* free, period */
309 #define SWM_POP 0x04 /* pop out */
311 int swap_pager_full = 2; /* swap space exhaustion (task killing) */
312 static int swap_pager_almost_full = 1; /* swap space exhaustion (w/hysteresis)*/
313 static int nsw_rcount; /* free read buffers */
314 static int nsw_wcount_sync; /* limit write buffers / synchronous */
315 static int nsw_wcount_async; /* limit write buffers / asynchronous */
316 static int nsw_wcount_async_max;/* assigned maximum */
317 static int nsw_cluster_max; /* maximum VOP I/O allowed */
319 static struct swblock **swhash;
320 static int swhash_mask;
321 static struct mtx swhash_mtx;
323 static int swap_async_max = 4; /* maximum in-progress async I/O's */
324 static struct sx sw_alloc_sx;
327 SYSCTL_INT(_vm, OID_AUTO, swap_async_max,
328 CTLFLAG_RW, &swap_async_max, 0, "Maximum running async swap ops");
331 * "named" and "unnamed" anon region objects. Try to reduce the overhead
332 * of searching a named list by hashing it just a little.
337 #define NOBJLIST(handle) \
338 (&swap_pager_object_list[((int)(intptr_t)handle >> 4) & (NOBJLISTS-1)])
340 static struct mtx sw_alloc_mtx; /* protect list manipulation */
341 static struct pagerlst swap_pager_object_list[NOBJLISTS];
342 static uma_zone_t swap_zone;
343 static struct vm_object swap_zone_obj;
346 * pagerops for OBJT_SWAP - "swap pager". Some ops are also global procedure
347 * calls hooked from other parts of the VM system and do not appear here.
348 * (see vm/swap_pager.h).
351 swap_pager_alloc(void *handle, vm_ooffset_t size,
352 vm_prot_t prot, vm_ooffset_t offset, struct ucred *);
353 static void swap_pager_dealloc(vm_object_t object);
354 static int swap_pager_getpages(vm_object_t, vm_page_t *, int, int);
355 static void swap_pager_putpages(vm_object_t, vm_page_t *, int, boolean_t, int *);
357 swap_pager_haspage(vm_object_t object, vm_pindex_t pindex, int *before, int *after);
358 static void swap_pager_init(void);
359 static void swap_pager_unswapped(vm_page_t);
360 static void swap_pager_swapoff(struct swdevt *sp);
362 struct pagerops swappagerops = {
363 .pgo_init = swap_pager_init, /* early system initialization of pager */
364 .pgo_alloc = swap_pager_alloc, /* allocate an OBJT_SWAP object */
365 .pgo_dealloc = swap_pager_dealloc, /* deallocate an OBJT_SWAP object */
366 .pgo_getpages = swap_pager_getpages, /* pagein */
367 .pgo_putpages = swap_pager_putpages, /* pageout */
368 .pgo_haspage = swap_pager_haspage, /* get backing store status for page */
369 .pgo_pageunswapped = swap_pager_unswapped, /* remove swap related to page */
373 * dmmax is in page-sized chunks with the new swap system. It was
374 * dev-bsized chunks in the old. dmmax is always a power of 2.
376 * swap_*() routines are externally accessible. swp_*() routines are
380 static int nswap_lowat = 128; /* in pages, swap_pager_almost_full warn */
381 static int nswap_hiwat = 512; /* in pages, swap_pager_almost_full warn */
383 SYSCTL_INT(_vm, OID_AUTO, dmmax,
384 CTLFLAG_RD, &dmmax, 0, "Maximum size of a swap block");
386 static void swp_sizecheck(void);
387 static void swp_pager_async_iodone(struct buf *bp);
388 static int swapongeom(struct thread *, struct vnode *);
389 static int swaponvp(struct thread *, struct vnode *, u_long);
390 static int swapoff_one(struct swdevt *sp, struct ucred *cred);
393 * Swap bitmap functions
395 static void swp_pager_freeswapspace(daddr_t blk, int npages);
396 static daddr_t swp_pager_getswapspace(int npages);
401 static struct swblock **swp_pager_hash(vm_object_t object, vm_pindex_t index);
402 static void swp_pager_meta_build(vm_object_t, vm_pindex_t, daddr_t);
403 static void swp_pager_meta_free(vm_object_t, vm_pindex_t, daddr_t);
404 static void swp_pager_meta_free_all(vm_object_t);
405 static daddr_t swp_pager_meta_ctl(vm_object_t, vm_pindex_t, int);
408 swp_pager_free_nrpage(vm_page_t m)
412 if (m->wire_count == 0)
418 * SWP_SIZECHECK() - update swap_pager_full indication
420 * update the swap_pager_almost_full indication and warn when we are
421 * about to run out of swap space, using lowat/hiwat hysteresis.
423 * Clear swap_pager_full ( task killing ) indication when lowat is met.
425 * No restrictions on call
426 * This routine may not block.
432 if (swap_pager_avail < nswap_lowat) {
433 if (swap_pager_almost_full == 0) {
434 printf("swap_pager: out of swap space\n");
435 swap_pager_almost_full = 1;
439 if (swap_pager_avail > nswap_hiwat)
440 swap_pager_almost_full = 0;
445 * SWP_PAGER_HASH() - hash swap meta data
447 * This is an helper function which hashes the swapblk given
448 * the object and page index. It returns a pointer to a pointer
449 * to the object, or a pointer to a NULL pointer if it could not
452 static struct swblock **
453 swp_pager_hash(vm_object_t object, vm_pindex_t index)
455 struct swblock **pswap;
456 struct swblock *swap;
458 index &= ~(vm_pindex_t)SWAP_META_MASK;
459 pswap = &swhash[(index ^ (int)(intptr_t)object) & swhash_mask];
460 while ((swap = *pswap) != NULL) {
461 if (swap->swb_object == object &&
462 swap->swb_index == index
466 pswap = &swap->swb_hnext;
472 * SWAP_PAGER_INIT() - initialize the swap pager!
474 * Expected to be started from system init. NOTE: This code is run
475 * before much else so be careful what you depend on. Most of the VM
476 * system has yet to be initialized at this point.
479 swap_pager_init(void)
482 * Initialize object lists
486 for (i = 0; i < NOBJLISTS; ++i)
487 TAILQ_INIT(&swap_pager_object_list[i]);
488 mtx_init(&sw_alloc_mtx, "swap_pager list", NULL, MTX_DEF);
489 mtx_init(&sw_dev_mtx, "swapdev", NULL, MTX_DEF);
492 * Device Stripe, in PAGE_SIZE'd blocks
494 dmmax = SWB_NPAGES * 2;
498 * SWAP_PAGER_SWAP_INIT() - swap pager initialization from pageout process
500 * Expected to be started from pageout process once, prior to entering
504 swap_pager_swap_init(void)
509 * Number of in-transit swap bp operations. Don't
510 * exhaust the pbufs completely. Make sure we
511 * initialize workable values (0 will work for hysteresis
512 * but it isn't very efficient).
514 * The nsw_cluster_max is constrained by the bp->b_pages[]
515 * array (MAXPHYS/PAGE_SIZE) and our locally defined
516 * MAX_PAGEOUT_CLUSTER. Also be aware that swap ops are
517 * constrained by the swap device interleave stripe size.
519 * Currently we hardwire nsw_wcount_async to 4. This limit is
520 * designed to prevent other I/O from having high latencies due to
521 * our pageout I/O. The value 4 works well for one or two active swap
522 * devices but is probably a little low if you have more. Even so,
523 * a higher value would probably generate only a limited improvement
524 * with three or four active swap devices since the system does not
525 * typically have to pageout at extreme bandwidths. We will want
526 * at least 2 per swap devices, and 4 is a pretty good value if you
527 * have one NFS swap device due to the command/ack latency over NFS.
528 * So it all works out pretty well.
530 nsw_cluster_max = min((MAXPHYS/PAGE_SIZE), MAX_PAGEOUT_CLUSTER);
533 nsw_rcount = (nswbuf + 1) / 2;
534 nsw_wcount_sync = (nswbuf + 3) / 4;
535 nsw_wcount_async = 4;
536 nsw_wcount_async_max = nsw_wcount_async;
537 mtx_unlock(&pbuf_mtx);
540 * Initialize our zone. Right now I'm just guessing on the number
541 * we need based on the number of pages in the system. Each swblock
542 * can hold 16 pages, so this is probably overkill. This reservation
543 * is typically limited to around 32MB by default.
545 n = cnt.v_page_count / 2;
546 if (maxswzone && n > maxswzone / sizeof(struct swblock))
547 n = maxswzone / sizeof(struct swblock);
549 swap_zone = uma_zcreate("SWAPMETA", sizeof(struct swblock), NULL, NULL,
550 NULL, NULL, UMA_ALIGN_PTR, UMA_ZONE_NOFREE | UMA_ZONE_VM);
551 if (swap_zone == NULL)
552 panic("failed to create swap_zone.");
554 if (uma_zone_set_obj(swap_zone, &swap_zone_obj, n))
557 * if the allocation failed, try a zone two thirds the
558 * size of the previous attempt.
563 printf("Swap zone entries reduced from %d to %d.\n", n2, n);
567 * Initialize our meta-data hash table. The swapper does not need to
568 * be quite as efficient as the VM system, so we do not use an
569 * oversized hash table.
571 * n: size of hash table, must be power of 2
572 * swhash_mask: hash table index mask
574 for (n = 1; n < n2 / 8; n *= 2)
576 swhash = malloc(sizeof(struct swblock *) * n, M_VMPGDATA, M_WAITOK | M_ZERO);
578 mtx_init(&swhash_mtx, "swap_pager swhash", NULL, MTX_DEF);
582 * SWAP_PAGER_ALLOC() - allocate a new OBJT_SWAP VM object and instantiate
583 * its metadata structures.
585 * This routine is called from the mmap and fork code to create a new
586 * OBJT_SWAP object. We do this by creating an OBJT_DEFAULT object
587 * and then converting it with swp_pager_meta_build().
589 * This routine may block in vm_object_allocate() and create a named
590 * object lookup race, so we must interlock.
595 swap_pager_alloc(void *handle, vm_ooffset_t size, vm_prot_t prot,
596 vm_ooffset_t offset, struct ucred *cred)
601 pindex = OFF_TO_IDX(offset + PAGE_MASK + size);
605 * Reference existing named region or allocate new one. There
606 * should not be a race here against swp_pager_meta_build()
607 * as called from vm_page_remove() in regards to the lookup
610 sx_xlock(&sw_alloc_sx);
611 object = vm_pager_object_lookup(NOBJLIST(handle), handle);
612 if (object == NULL) {
614 if (!swap_reserve_by_cred(size, cred)) {
615 sx_xunlock(&sw_alloc_sx);
621 object = vm_object_allocate(OBJT_DEFAULT, pindex);
622 VM_OBJECT_LOCK(object);
623 object->handle = handle;
626 object->charge = size;
628 swp_pager_meta_build(object, 0, SWAPBLK_NONE);
629 VM_OBJECT_UNLOCK(object);
631 sx_xunlock(&sw_alloc_sx);
635 if (!swap_reserve_by_cred(size, cred))
639 object = vm_object_allocate(OBJT_DEFAULT, pindex);
640 VM_OBJECT_LOCK(object);
643 object->charge = size;
645 swp_pager_meta_build(object, 0, SWAPBLK_NONE);
646 VM_OBJECT_UNLOCK(object);
652 * SWAP_PAGER_DEALLOC() - remove swap metadata from object
654 * The swap backing for the object is destroyed. The code is
655 * designed such that we can reinstantiate it later, but this
656 * routine is typically called only when the entire object is
657 * about to be destroyed.
659 * This routine may block, but no longer does.
661 * The object must be locked or unreferenceable.
664 swap_pager_dealloc(vm_object_t object)
668 * Remove from list right away so lookups will fail if we block for
669 * pageout completion.
671 if (object->handle != NULL) {
672 mtx_lock(&sw_alloc_mtx);
673 TAILQ_REMOVE(NOBJLIST(object->handle), object, pager_object_list);
674 mtx_unlock(&sw_alloc_mtx);
677 VM_OBJECT_LOCK_ASSERT(object, MA_OWNED);
678 vm_object_pip_wait(object, "swpdea");
681 * Free all remaining metadata. We only bother to free it from
682 * the swap meta data. We do not attempt to free swapblk's still
683 * associated with vm_page_t's for this object. We do not care
684 * if paging is still in progress on some objects.
686 swp_pager_meta_free_all(object);
689 /************************************************************************
690 * SWAP PAGER BITMAP ROUTINES *
691 ************************************************************************/
694 * SWP_PAGER_GETSWAPSPACE() - allocate raw swap space
696 * Allocate swap for the requested number of pages. The starting
697 * swap block number (a page index) is returned or SWAPBLK_NONE
698 * if the allocation failed.
700 * Also has the side effect of advising that somebody made a mistake
701 * when they configured swap and didn't configure enough.
703 * This routine may not block
705 * We allocate in round-robin fashion from the configured devices.
708 swp_pager_getswapspace(int npages)
715 mtx_lock(&sw_dev_mtx);
717 for (i = 0; i < nswapdev; i++) {
719 sp = TAILQ_FIRST(&swtailq);
720 if (!(sp->sw_flags & SW_CLOSING)) {
721 blk = blist_alloc(sp->sw_blist, npages);
722 if (blk != SWAPBLK_NONE) {
724 sp->sw_used += npages;
725 swap_pager_avail -= npages;
727 swdevhd = TAILQ_NEXT(sp, sw_list);
731 sp = TAILQ_NEXT(sp, sw_list);
733 if (swap_pager_full != 2) {
734 printf("swap_pager_getswapspace(%d): failed\n", npages);
736 swap_pager_almost_full = 1;
740 mtx_unlock(&sw_dev_mtx);
745 swp_pager_isondev(daddr_t blk, struct swdevt *sp)
748 return (blk >= sp->sw_first && blk < sp->sw_end);
752 swp_pager_strategy(struct buf *bp)
756 mtx_lock(&sw_dev_mtx);
757 TAILQ_FOREACH(sp, &swtailq, sw_list) {
758 if (bp->b_blkno >= sp->sw_first && bp->b_blkno < sp->sw_end) {
759 mtx_unlock(&sw_dev_mtx);
760 sp->sw_strategy(bp, sp);
764 panic("Swapdev not found");
769 * SWP_PAGER_FREESWAPSPACE() - free raw swap space
771 * This routine returns the specified swap blocks back to the bitmap.
773 * Note: This routine may not block (it could in the old swap code),
774 * and through the use of the new blist routines it does not block.
776 * This routine may not block
779 swp_pager_freeswapspace(daddr_t blk, int npages)
783 mtx_lock(&sw_dev_mtx);
784 TAILQ_FOREACH(sp, &swtailq, sw_list) {
785 if (blk >= sp->sw_first && blk < sp->sw_end) {
786 sp->sw_used -= npages;
788 * If we are attempting to stop swapping on
789 * this device, we don't want to mark any
790 * blocks free lest they be reused.
792 if ((sp->sw_flags & SW_CLOSING) == 0) {
793 blist_free(sp->sw_blist, blk - sp->sw_first,
795 swap_pager_avail += npages;
798 mtx_unlock(&sw_dev_mtx);
802 panic("Swapdev not found");
806 * SWAP_PAGER_FREESPACE() - frees swap blocks associated with a page
807 * range within an object.
809 * This is a globally accessible routine.
811 * This routine removes swapblk assignments from swap metadata.
813 * The external callers of this routine typically have already destroyed
814 * or renamed vm_page_t's associated with this range in the object so
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 * XXX vm_page_collapse() kinda expects us not to block because we
877 * supposedly do not need to allocate memory, but for the moment we
878 * *may* have to get a little memory from the zone allocator, but
879 * it is taken from the interrupt memory. We should be ok.
881 * The source object contains no vm_page_t's (which is just as well)
883 * The source object is of type OBJT_SWAP.
885 * The source and destination objects must be locked or
886 * inaccessible (XXX are they ?)
889 swap_pager_copy(vm_object_t srcobject, vm_object_t dstobject,
890 vm_pindex_t offset, int destroysource)
894 VM_OBJECT_LOCK_ASSERT(srcobject, MA_OWNED);
895 VM_OBJECT_LOCK_ASSERT(dstobject, MA_OWNED);
898 * If destroysource is set, we remove the source object from the
899 * swap_pager internal queue now.
902 if (srcobject->handle != NULL) {
903 mtx_lock(&sw_alloc_mtx);
905 NOBJLIST(srcobject->handle),
909 mtx_unlock(&sw_alloc_mtx);
914 * transfer source to destination.
916 for (i = 0; i < dstobject->size; ++i) {
920 * Locate (without changing) the swapblk on the destination,
921 * unless it is invalid in which case free it silently, or
922 * if the destination is a resident page, in which case the
923 * source is thrown away.
925 dstaddr = swp_pager_meta_ctl(dstobject, i, 0);
927 if (dstaddr == SWAPBLK_NONE) {
929 * Destination has no swapblk and is not resident,
934 srcaddr = swp_pager_meta_ctl(
940 if (srcaddr != SWAPBLK_NONE) {
942 * swp_pager_meta_build() can sleep.
944 vm_object_pip_add(srcobject, 1);
945 VM_OBJECT_UNLOCK(srcobject);
946 vm_object_pip_add(dstobject, 1);
947 swp_pager_meta_build(dstobject, i, srcaddr);
948 vm_object_pip_wakeup(dstobject);
949 VM_OBJECT_LOCK(srcobject);
950 vm_object_pip_wakeup(srcobject);
954 * Destination has valid swapblk or it is represented
955 * by a resident page. We destroy the sourceblock.
958 swp_pager_meta_ctl(srcobject, i + offset, SWM_FREE);
963 * Free left over swap blocks in source.
965 * We have to revert the type to OBJT_DEFAULT so we do not accidently
966 * double-remove the object from the swap queues.
969 swp_pager_meta_free_all(srcobject);
971 * Reverting the type is not necessary, the caller is going
972 * to destroy srcobject directly, but I'm doing it here
973 * for consistency since we've removed the object from its
976 srcobject->type = OBJT_DEFAULT;
981 * SWAP_PAGER_HASPAGE() - determine if we have good backing store for
982 * the requested page.
984 * We determine whether good backing store exists for the requested
985 * page and return TRUE if it does, FALSE if it doesn't.
987 * If TRUE, we also try to determine how much valid, contiguous backing
988 * store exists before and after the requested page within a reasonable
989 * distance. We do not try to restrict it to the swap device stripe
990 * (that is handled in getpages/putpages). It probably isn't worth
994 swap_pager_haspage(vm_object_t object, vm_pindex_t pindex, int *before, int *after)
998 VM_OBJECT_LOCK_ASSERT(object, MA_OWNED);
1000 * do we have good backing store at the requested index ?
1002 blk0 = swp_pager_meta_ctl(object, pindex, 0);
1004 if (blk0 == SWAPBLK_NONE) {
1013 * find backwards-looking contiguous good backing store
1015 if (before != NULL) {
1018 for (i = 1; i < (SWB_NPAGES/2); ++i) {
1023 blk = swp_pager_meta_ctl(object, pindex - i, 0);
1024 if (blk != blk0 - i)
1031 * find forward-looking contiguous good backing store
1033 if (after != NULL) {
1036 for (i = 1; i < (SWB_NPAGES/2); ++i) {
1039 blk = swp_pager_meta_ctl(object, pindex + i, 0);
1040 if (blk != blk0 + i)
1049 * SWAP_PAGER_PAGE_UNSWAPPED() - remove swap backing store related to page
1051 * This removes any associated swap backing store, whether valid or
1052 * not, from the page.
1054 * This routine is typically called when a page is made dirty, at
1055 * which point any associated swap can be freed. MADV_FREE also
1056 * calls us in a special-case situation
1058 * NOTE!!! If the page is clean and the swap was valid, the caller
1059 * should make the page dirty before calling this routine. This routine
1060 * does NOT change the m->dirty status of the page. Also: MADV_FREE
1063 * This routine may not block
1066 swap_pager_unswapped(vm_page_t m)
1069 VM_OBJECT_LOCK_ASSERT(m->object, MA_OWNED);
1070 swp_pager_meta_ctl(m->object, m->pindex, SWM_FREE);
1074 * SWAP_PAGER_GETPAGES() - bring pages in from swap
1076 * Attempt to retrieve (m, count) pages from backing store, but make
1077 * sure we retrieve at least m[reqpage]. We try to load in as large
1078 * a chunk surrounding m[reqpage] as is contiguous in swap and which
1079 * belongs to the same object.
1081 * The code is designed for asynchronous operation and
1082 * immediate-notification of 'reqpage' but tends not to be
1083 * used that way. Please do not optimize-out this algorithmic
1084 * feature, I intend to improve on it in the future.
1086 * The parent has a single vm_object_pip_add() reference prior to
1087 * calling us and we should return with the same.
1089 * The parent has BUSY'd the pages. We should return with 'm'
1090 * left busy, but the others adjusted.
1093 swap_pager_getpages(vm_object_t object, vm_page_t *m, int count, int reqpage)
1103 KASSERT(mreq->object == object,
1104 ("swap_pager_getpages: object mismatch %p/%p",
1105 object, mreq->object));
1108 * Calculate range to retrieve. The pages have already been assigned
1109 * their swapblks. We require a *contiguous* range but we know it to
1110 * not span devices. If we do not supply it, bad things
1111 * happen. Note that blk, iblk & jblk can be SWAPBLK_NONE, but the
1112 * loops are set up such that the case(s) are handled implicitly.
1114 * The swp_*() calls must be made with the object locked.
1116 blk = swp_pager_meta_ctl(mreq->object, mreq->pindex, 0);
1118 for (i = reqpage - 1; i >= 0; --i) {
1121 iblk = swp_pager_meta_ctl(m[i]->object, m[i]->pindex, 0);
1122 if (blk != iblk + (reqpage - i))
1127 for (j = reqpage + 1; j < count; ++j) {
1130 jblk = swp_pager_meta_ctl(m[j]->object, m[j]->pindex, 0);
1131 if (blk != jblk - (j - reqpage))
1136 * free pages outside our collection range. Note: we never free
1137 * mreq, it must remain busy throughout.
1139 if (0 < i || j < count) {
1142 for (k = 0; k < i; ++k)
1143 swp_pager_free_nrpage(m[k]);
1144 for (k = j; k < count; ++k)
1145 swp_pager_free_nrpage(m[k]);
1149 * Return VM_PAGER_FAIL if we have nothing to do. Return mreq
1150 * still busy, but the others unbusied.
1152 if (blk == SWAPBLK_NONE)
1153 return (VM_PAGER_FAIL);
1156 * Getpbuf() can sleep.
1158 VM_OBJECT_UNLOCK(object);
1160 * Get a swap buffer header to perform the IO
1162 bp = getpbuf(&nsw_rcount);
1163 bp->b_flags |= B_PAGING;
1166 * map our page(s) into kva for input
1168 pmap_qenter((vm_offset_t)bp->b_data, m + i, j - i);
1170 bp->b_iocmd = BIO_READ;
1171 bp->b_iodone = swp_pager_async_iodone;
1172 bp->b_rcred = crhold(thread0.td_ucred);
1173 bp->b_wcred = crhold(thread0.td_ucred);
1174 bp->b_blkno = blk - (reqpage - i);
1175 bp->b_bcount = PAGE_SIZE * (j - i);
1176 bp->b_bufsize = PAGE_SIZE * (j - i);
1177 bp->b_pager.pg_reqpage = reqpage - i;
1179 VM_OBJECT_LOCK(object);
1183 for (k = i; k < j; ++k) {
1184 bp->b_pages[k - i] = m[k];
1185 m[k]->oflags |= VPO_SWAPINPROG;
1188 bp->b_npages = j - i;
1190 PCPU_INC(cnt.v_swapin);
1191 PCPU_ADD(cnt.v_swappgsin, bp->b_npages);
1194 * We still hold the lock on mreq, and our automatic completion routine
1195 * does not remove it.
1197 vm_object_pip_add(object, bp->b_npages);
1198 VM_OBJECT_UNLOCK(object);
1201 * perform the I/O. NOTE!!! bp cannot be considered valid after
1202 * this point because we automatically release it on completion.
1203 * Instead, we look at the one page we are interested in which we
1204 * still hold a lock on even through the I/O completion.
1206 * The other pages in our m[] array are also released on completion,
1207 * so we cannot assume they are valid anymore either.
1209 * NOTE: b_blkno is destroyed by the call to swapdev_strategy
1212 swp_pager_strategy(bp);
1215 * wait for the page we want to complete. VPO_SWAPINPROG is always
1216 * cleared on completion. If an I/O error occurs, SWAPBLK_NONE
1217 * is set in the meta-data.
1219 VM_OBJECT_LOCK(object);
1220 while ((mreq->oflags & VPO_SWAPINPROG) != 0) {
1221 mreq->oflags |= VPO_WANTED;
1222 PCPU_INC(cnt.v_intrans);
1223 if (msleep(mreq, VM_OBJECT_MTX(object), PSWP, "swread", hz*20)) {
1225 "swap_pager: indefinite wait buffer: bufobj: %p, blkno: %jd, size: %ld\n",
1226 bp->b_bufobj, (intmax_t)bp->b_blkno, bp->b_bcount);
1231 * mreq is left busied after completion, but all the other pages
1232 * are freed. If we had an unrecoverable read error the page will
1235 if (mreq->valid != VM_PAGE_BITS_ALL) {
1236 return (VM_PAGER_ERROR);
1238 return (VM_PAGER_OK);
1242 * A final note: in a low swap situation, we cannot deallocate swap
1243 * and mark a page dirty here because the caller is likely to mark
1244 * the page clean when we return, causing the page to possibly revert
1245 * to all-zero's later.
1250 * swap_pager_putpages:
1252 * Assign swap (if necessary) and initiate I/O on the specified pages.
1254 * We support both OBJT_DEFAULT and OBJT_SWAP objects. DEFAULT objects
1255 * are automatically converted to SWAP objects.
1257 * In a low memory situation we may block in VOP_STRATEGY(), but the new
1258 * vm_page reservation system coupled with properly written VFS devices
1259 * should ensure that no low-memory deadlock occurs. This is an area
1262 * The parent has N vm_object_pip_add() references prior to
1263 * calling us and will remove references for rtvals[] that are
1264 * not set to VM_PAGER_PEND. We need to remove the rest on I/O
1267 * The parent has soft-busy'd the pages it passes us and will unbusy
1268 * those whos rtvals[] entry is not set to VM_PAGER_PEND on return.
1269 * We need to unbusy the rest on I/O completion.
1272 swap_pager_putpages(vm_object_t object, vm_page_t *m, int count,
1273 boolean_t sync, int *rtvals)
1278 if (count && m[0]->object != object) {
1279 panic("swap_pager_putpages: object mismatch %p/%p",
1288 * Turn object into OBJT_SWAP
1289 * check for bogus sysops
1290 * force sync if not pageout process
1292 if (object->type != OBJT_SWAP)
1293 swp_pager_meta_build(object, 0, SWAPBLK_NONE);
1294 VM_OBJECT_UNLOCK(object);
1296 if (curproc != pageproc)
1302 * Update nsw parameters from swap_async_max sysctl values.
1303 * Do not let the sysop crash the machine with bogus numbers.
1305 mtx_lock(&pbuf_mtx);
1306 if (swap_async_max != nsw_wcount_async_max) {
1312 if ((n = swap_async_max) > nswbuf / 2)
1319 * Adjust difference ( if possible ). If the current async
1320 * count is too low, we may not be able to make the adjustment
1323 n -= nsw_wcount_async_max;
1324 if (nsw_wcount_async + n >= 0) {
1325 nsw_wcount_async += n;
1326 nsw_wcount_async_max += n;
1327 wakeup(&nsw_wcount_async);
1330 mtx_unlock(&pbuf_mtx);
1335 * Assign swap blocks and issue I/O. We reallocate swap on the fly.
1336 * The page is left dirty until the pageout operation completes
1339 for (i = 0; i < count; i += n) {
1345 * Maximum I/O size is limited by a number of factors.
1347 n = min(BLIST_MAX_ALLOC, count - i);
1348 n = min(n, nsw_cluster_max);
1351 * Get biggest block of swap we can. If we fail, fall
1352 * back and try to allocate a smaller block. Don't go
1353 * overboard trying to allocate space if it would overly
1357 (blk = swp_pager_getswapspace(n)) == SWAPBLK_NONE &&
1362 if (blk == SWAPBLK_NONE) {
1363 for (j = 0; j < n; ++j)
1364 rtvals[i+j] = VM_PAGER_FAIL;
1369 * All I/O parameters have been satisfied, build the I/O
1370 * request and assign the swap space.
1373 bp = getpbuf(&nsw_wcount_sync);
1375 bp = getpbuf(&nsw_wcount_async);
1376 bp->b_flags = B_ASYNC;
1378 bp->b_flags |= B_PAGING;
1379 bp->b_iocmd = BIO_WRITE;
1381 pmap_qenter((vm_offset_t)bp->b_data, &m[i], n);
1383 bp->b_rcred = crhold(thread0.td_ucred);
1384 bp->b_wcred = crhold(thread0.td_ucred);
1385 bp->b_bcount = PAGE_SIZE * n;
1386 bp->b_bufsize = PAGE_SIZE * n;
1389 VM_OBJECT_LOCK(object);
1390 for (j = 0; j < n; ++j) {
1391 vm_page_t mreq = m[i+j];
1393 swp_pager_meta_build(
1398 vm_page_dirty(mreq);
1399 rtvals[i+j] = VM_PAGER_OK;
1401 mreq->oflags |= VPO_SWAPINPROG;
1402 bp->b_pages[j] = mreq;
1404 VM_OBJECT_UNLOCK(object);
1407 * Must set dirty range for NFS to work.
1410 bp->b_dirtyend = bp->b_bcount;
1412 PCPU_INC(cnt.v_swapout);
1413 PCPU_ADD(cnt.v_swappgsout, bp->b_npages);
1418 * NOTE: b_blkno is destroyed by the call to swapdev_strategy
1420 if (sync == FALSE) {
1421 bp->b_iodone = swp_pager_async_iodone;
1423 swp_pager_strategy(bp);
1425 for (j = 0; j < n; ++j)
1426 rtvals[i+j] = VM_PAGER_PEND;
1427 /* restart outter loop */
1434 * NOTE: b_blkno is destroyed by the call to swapdev_strategy
1436 bp->b_iodone = bdone;
1437 swp_pager_strategy(bp);
1440 * Wait for the sync I/O to complete, then update rtvals.
1441 * We just set the rtvals[] to VM_PAGER_PEND so we can call
1442 * our async completion routine at the end, thus avoiding a
1445 bwait(bp, PVM, "swwrt");
1446 for (j = 0; j < n; ++j)
1447 rtvals[i+j] = VM_PAGER_PEND;
1449 * Now that we are through with the bp, we can call the
1450 * normal async completion, which frees everything up.
1452 swp_pager_async_iodone(bp);
1454 VM_OBJECT_LOCK(object);
1458 * swp_pager_async_iodone:
1460 * Completion routine for asynchronous reads and writes from/to swap.
1461 * Also called manually by synchronous code to finish up a bp.
1463 * For READ operations, the pages are VPO_BUSY'd. For WRITE operations,
1464 * the pages are vm_page_t->busy'd. For READ operations, we VPO_BUSY
1465 * unbusy all pages except the 'main' request page. For WRITE
1466 * operations, we vm_page_t->busy'd unbusy all pages ( we can do this
1467 * because we marked them all VM_PAGER_PEND on return from putpages ).
1469 * This routine may not block.
1472 swp_pager_async_iodone(struct buf *bp)
1475 vm_object_t object = NULL;
1480 if (bp->b_ioflags & BIO_ERROR) {
1482 "swap_pager: I/O error - %s failed; blkno %ld,"
1483 "size %ld, error %d\n",
1484 ((bp->b_iocmd == BIO_READ) ? "pagein" : "pageout"),
1492 * remove the mapping for kernel virtual
1494 pmap_qremove((vm_offset_t)bp->b_data, bp->b_npages);
1497 object = bp->b_pages[0]->object;
1498 VM_OBJECT_LOCK(object);
1502 * cleanup pages. If an error occurs writing to swap, we are in
1503 * very serious trouble. If it happens to be a disk error, though,
1504 * we may be able to recover by reassigning the swap later on. So
1505 * in this case we remove the m->swapblk assignment for the page
1506 * but do not free it in the rlist. The errornous block(s) are thus
1507 * never reallocated as swap. Redirty the page and continue.
1509 for (i = 0; i < bp->b_npages; ++i) {
1510 vm_page_t m = bp->b_pages[i];
1512 m->oflags &= ~VPO_SWAPINPROG;
1514 if (bp->b_ioflags & BIO_ERROR) {
1516 * If an error occurs I'd love to throw the swapblk
1517 * away without freeing it back to swapspace, so it
1518 * can never be used again. But I can't from an
1521 if (bp->b_iocmd == BIO_READ) {
1523 * When reading, reqpage needs to stay
1524 * locked for the parent, but all other
1525 * pages can be freed. We still want to
1526 * wakeup the parent waiting on the page,
1527 * though. ( also: pg_reqpage can be -1 and
1528 * not match anything ).
1530 * We have to wake specifically requested pages
1531 * up too because we cleared VPO_SWAPINPROG and
1532 * someone may be waiting for that.
1534 * NOTE: for reads, m->dirty will probably
1535 * be overridden by the original caller of
1536 * getpages so don't play cute tricks here.
1539 if (i != bp->b_pager.pg_reqpage)
1540 swp_pager_free_nrpage(m);
1544 * If i == bp->b_pager.pg_reqpage, do not wake
1545 * the page up. The caller needs to.
1549 * If a write error occurs, reactivate page
1550 * so it doesn't clog the inactive list,
1551 * then finish the I/O.
1555 vm_page_activate(m);
1557 vm_page_io_finish(m);
1559 } else if (bp->b_iocmd == BIO_READ) {
1561 * NOTE: for reads, m->dirty will probably be
1562 * overridden by the original caller of getpages so
1563 * we cannot set them in order to free the underlying
1564 * swap in a low-swap situation. I don't think we'd
1565 * want to do that anyway, but it was an optimization
1566 * that existed in the old swapper for a time before
1567 * it got ripped out due to precisely this problem.
1569 * If not the requested page then deactivate it.
1571 * Note that the requested page, reqpage, is left
1572 * busied, but we still have to wake it up. The
1573 * other pages are released (unbusied) by
1576 KASSERT(!pmap_page_is_mapped(m),
1577 ("swp_pager_async_iodone: page %p is mapped", m));
1578 m->valid = VM_PAGE_BITS_ALL;
1579 KASSERT(m->dirty == 0,
1580 ("swp_pager_async_iodone: page %p is dirty", m));
1583 * We have to wake specifically requested pages
1584 * up too because we cleared VPO_SWAPINPROG and
1585 * could be waiting for it in getpages. However,
1586 * be sure to not unbusy getpages specifically
1587 * requested page - getpages expects it to be
1590 if (i != bp->b_pager.pg_reqpage) {
1592 vm_page_deactivate(m);
1599 * For write success, clear the dirty
1600 * status, then finish the I/O ( which decrements the
1601 * busy count and possibly wakes waiter's up ).
1603 KASSERT((m->flags & PG_WRITEABLE) == 0,
1604 ("swp_pager_async_iodone: page %p is not write"
1607 vm_page_io_finish(m);
1608 if (vm_page_count_severe()) {
1610 vm_page_try_to_cache(m);
1617 * adjust pip. NOTE: the original parent may still have its own
1618 * pip refs on the object.
1620 if (object != NULL) {
1621 vm_object_pip_wakeupn(object, bp->b_npages);
1622 VM_OBJECT_UNLOCK(object);
1626 * swapdev_strategy() manually sets b_vp and b_bufobj before calling
1627 * bstrategy(). Set them back to NULL now we're done with it, or we'll
1628 * trigger a KASSERT in relpbuf().
1632 bp->b_bufobj = NULL;
1635 * release the physical I/O buffer
1639 ((bp->b_iocmd == BIO_READ) ? &nsw_rcount :
1640 ((bp->b_flags & B_ASYNC) ?
1649 * swap_pager_isswapped:
1651 * Return 1 if at least one page in the given object is paged
1652 * out to the given swap device.
1654 * This routine may not block.
1657 swap_pager_isswapped(vm_object_t object, struct swdevt *sp)
1663 VM_OBJECT_LOCK_ASSERT(object, MA_OWNED);
1664 if (object->type != OBJT_SWAP)
1667 mtx_lock(&swhash_mtx);
1668 for (bcount = 0; bcount < object->un_pager.swp.swp_bcount; bcount++) {
1669 struct swblock *swap;
1671 if ((swap = *swp_pager_hash(object, index)) != NULL) {
1672 for (i = 0; i < SWAP_META_PAGES; ++i) {
1673 if (swp_pager_isondev(swap->swb_pages[i], sp)) {
1674 mtx_unlock(&swhash_mtx);
1679 index += SWAP_META_PAGES;
1681 mtx_unlock(&swhash_mtx);
1686 * SWP_PAGER_FORCE_PAGEIN() - force a swap block to be paged in
1688 * This routine dissociates the page at the given index within a
1689 * swap block from its backing store, paging it in if necessary.
1690 * If the page is paged in, it is placed in the inactive queue,
1691 * since it had its backing store ripped out from under it.
1692 * We also attempt to swap in all other pages in the swap block,
1693 * we only guarantee that the one at the specified index is
1696 * XXX - The code to page the whole block in doesn't work, so we
1697 * revert to the one-by-one behavior for now. Sigh.
1700 swp_pager_force_pagein(vm_object_t object, vm_pindex_t pindex)
1704 vm_object_pip_add(object, 1);
1705 m = vm_page_grab(object, pindex, VM_ALLOC_NORMAL|VM_ALLOC_RETRY);
1706 if (m->valid == VM_PAGE_BITS_ALL) {
1707 vm_object_pip_subtract(object, 1);
1710 vm_page_activate(m);
1713 vm_pager_page_unswapped(m);
1717 if (swap_pager_getpages(object, &m, 1, 0) != VM_PAGER_OK)
1718 panic("swap_pager_force_pagein: read from swap failed");/*XXX*/
1719 vm_object_pip_subtract(object, 1);
1722 vm_page_deactivate(m);
1725 vm_pager_page_unswapped(m);
1729 * swap_pager_swapoff:
1731 * Page in all of the pages that have been paged out to the
1732 * given device. The corresponding blocks in the bitmap must be
1733 * marked as allocated and the device must be flagged SW_CLOSING.
1734 * There may be no processes swapped out to the device.
1736 * This routine may block.
1739 swap_pager_swapoff(struct swdevt *sp)
1741 struct swblock *swap;
1748 mtx_lock(&swhash_mtx);
1749 for (i = 0; i <= swhash_mask; i++) { /* '<=' is correct here */
1751 for (swap = swhash[i]; swap != NULL; swap = swap->swb_hnext) {
1752 vm_object_t object = swap->swb_object;
1753 vm_pindex_t pindex = swap->swb_index;
1754 for (j = 0; j < SWAP_META_PAGES; ++j) {
1755 if (swp_pager_isondev(swap->swb_pages[j], sp)) {
1756 /* avoid deadlock */
1757 if (!VM_OBJECT_TRYLOCK(object)) {
1760 mtx_unlock(&swhash_mtx);
1761 swp_pager_force_pagein(object,
1763 VM_OBJECT_UNLOCK(object);
1764 mtx_lock(&swhash_mtx);
1771 mtx_unlock(&swhash_mtx);
1774 * Objects may be locked or paging to the device being
1775 * removed, so we will miss their pages and need to
1776 * make another pass. We have marked this device as
1777 * SW_CLOSING, so the activity should finish soon.
1780 if (retries > 100) {
1781 panic("swapoff: failed to locate %d swap blocks",
1784 pause("swpoff", hz / 20);
1789 /************************************************************************
1791 ************************************************************************
1793 * These routines manipulate the swap metadata stored in the
1796 * Swap metadata is implemented with a global hash and not directly
1797 * linked into the object. Instead the object simply contains
1798 * appropriate tracking counters.
1802 * SWP_PAGER_META_BUILD() - add swap block to swap meta data for object
1804 * We first convert the object to a swap object if it is a default
1807 * The specified swapblk is added to the object's swap metadata. If
1808 * the swapblk is not valid, it is freed instead. Any previously
1809 * assigned swapblk is freed.
1812 swp_pager_meta_build(vm_object_t object, vm_pindex_t pindex, daddr_t swapblk)
1814 struct swblock *swap;
1815 struct swblock **pswap;
1818 VM_OBJECT_LOCK_ASSERT(object, MA_OWNED);
1820 * Convert default object to swap object if necessary
1822 if (object->type != OBJT_SWAP) {
1823 object->type = OBJT_SWAP;
1824 object->un_pager.swp.swp_bcount = 0;
1826 if (object->handle != NULL) {
1827 mtx_lock(&sw_alloc_mtx);
1829 NOBJLIST(object->handle),
1833 mtx_unlock(&sw_alloc_mtx);
1838 * Locate hash entry. If not found create, but if we aren't adding
1839 * anything just return. If we run out of space in the map we wait
1840 * and, since the hash table may have changed, retry.
1843 mtx_lock(&swhash_mtx);
1844 pswap = swp_pager_hash(object, pindex);
1846 if ((swap = *pswap) == NULL) {
1849 if (swapblk == SWAPBLK_NONE)
1852 swap = *pswap = uma_zalloc(swap_zone, M_NOWAIT);
1854 mtx_unlock(&swhash_mtx);
1855 VM_OBJECT_UNLOCK(object);
1856 if (uma_zone_exhausted(swap_zone)) {
1857 printf("swap zone exhausted, increase kern.maxswzone\n");
1858 vm_pageout_oom(VM_OOM_SWAPZ);
1859 pause("swzonex", 10);
1862 VM_OBJECT_LOCK(object);
1866 swap->swb_hnext = NULL;
1867 swap->swb_object = object;
1868 swap->swb_index = pindex & ~(vm_pindex_t)SWAP_META_MASK;
1869 swap->swb_count = 0;
1871 ++object->un_pager.swp.swp_bcount;
1873 for (i = 0; i < SWAP_META_PAGES; ++i)
1874 swap->swb_pages[i] = SWAPBLK_NONE;
1878 * Delete prior contents of metadata
1880 idx = pindex & SWAP_META_MASK;
1882 if (swap->swb_pages[idx] != SWAPBLK_NONE) {
1883 swp_pager_freeswapspace(swap->swb_pages[idx], 1);
1888 * Enter block into metadata
1890 swap->swb_pages[idx] = swapblk;
1891 if (swapblk != SWAPBLK_NONE)
1894 mtx_unlock(&swhash_mtx);
1898 * SWP_PAGER_META_FREE() - free a range of blocks in the object's swap metadata
1900 * The requested range of blocks is freed, with any associated swap
1901 * returned to the swap bitmap.
1903 * This routine will free swap metadata structures as they are cleaned
1904 * out. This routine does *NOT* operate on swap metadata associated
1905 * with resident pages.
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
1953 swp_pager_meta_free_all(vm_object_t object)
1957 VM_OBJECT_LOCK_ASSERT(object, MA_OWNED);
1958 if (object->type != OBJT_SWAP)
1961 while (object->un_pager.swp.swp_bcount) {
1962 struct swblock **pswap;
1963 struct swblock *swap;
1965 mtx_lock(&swhash_mtx);
1966 pswap = swp_pager_hash(object, index);
1967 if ((swap = *pswap) != NULL) {
1970 for (i = 0; i < SWAP_META_PAGES; ++i) {
1971 daddr_t v = swap->swb_pages[i];
1972 if (v != SWAPBLK_NONE) {
1974 swp_pager_freeswapspace(v, 1);
1977 if (swap->swb_count != 0)
1978 panic("swap_pager_meta_free_all: swb_count != 0");
1979 *pswap = swap->swb_hnext;
1980 uma_zfree(swap_zone, swap);
1981 --object->un_pager.swp.swp_bcount;
1983 mtx_unlock(&swhash_mtx);
1984 index += SWAP_META_PAGES;
1989 * SWP_PAGER_METACTL() - misc control of swap and vm_page_t meta data.
1991 * This routine is capable of looking up, popping, or freeing
1992 * swapblk assignments in the swap meta data or in the vm_page_t.
1993 * The routine typically returns the swapblk being looked-up, or popped,
1994 * or SWAPBLK_NONE if the block was freed, or SWAPBLK_NONE if the block
1995 * was invalid. This routine will automatically free any invalid
1996 * meta-data swapblks.
1998 * It is not possible to store invalid swapblks in the swap meta data
1999 * (other then a literal 'SWAPBLK_NONE'), so we don't bother checking.
2001 * When acting on a busy resident page and paging is in progress, we
2002 * have to wait until paging is complete but otherwise can act on the
2005 * SWM_FREE remove and free swap block from metadata
2006 * SWM_POP remove from meta data but do not free.. pop it out
2009 swp_pager_meta_ctl(vm_object_t object, vm_pindex_t pindex, int flags)
2011 struct swblock **pswap;
2012 struct swblock *swap;
2016 VM_OBJECT_LOCK_ASSERT(object, MA_OWNED);
2018 * The meta data only exists of the object is OBJT_SWAP
2019 * and even then might not be allocated yet.
2021 if (object->type != OBJT_SWAP)
2022 return (SWAPBLK_NONE);
2025 mtx_lock(&swhash_mtx);
2026 pswap = swp_pager_hash(object, pindex);
2028 if ((swap = *pswap) != NULL) {
2029 idx = pindex & SWAP_META_MASK;
2030 r1 = swap->swb_pages[idx];
2032 if (r1 != SWAPBLK_NONE) {
2033 if (flags & SWM_FREE) {
2034 swp_pager_freeswapspace(r1, 1);
2037 if (flags & (SWM_FREE|SWM_POP)) {
2038 swap->swb_pages[idx] = SWAPBLK_NONE;
2039 if (--swap->swb_count == 0) {
2040 *pswap = swap->swb_hnext;
2041 uma_zfree(swap_zone, swap);
2042 --object->un_pager.swp.swp_bcount;
2047 mtx_unlock(&swhash_mtx);
2052 * System call swapon(name) enables swapping on device name,
2053 * which must be in the swdevsw. Return EBUSY
2054 * if already swapping on this device.
2056 #ifndef _SYS_SYSPROTO_H_
2057 struct swapon_args {
2067 swapon(struct thread *td, struct swapon_args *uap)
2071 struct nameidata nd;
2074 error = priv_check(td, PRIV_SWAPON);
2079 while (swdev_syscall_active)
2080 tsleep(&swdev_syscall_active, PUSER - 1, "swpon", 0);
2081 swdev_syscall_active = 1;
2084 * Swap metadata may not fit in the KVM if we have physical
2087 if (swap_zone == NULL) {
2092 NDINIT(&nd, LOOKUP, ISOPEN | FOLLOW | AUDITVNODE1, UIO_USERSPACE,
2098 NDFREE(&nd, NDF_ONLY_PNBUF);
2101 if (vn_isdisk(vp, &error)) {
2102 error = swapongeom(td, vp);
2103 } else if (vp->v_type == VREG &&
2104 (vp->v_mount->mnt_vfc->vfc_flags & VFCF_NETWORK) != 0 &&
2105 (error = VOP_GETATTR(vp, &attr, td->td_ucred)) == 0) {
2107 * Allow direct swapping to NFS regular files in the same
2108 * way that nfs_mountroot() sets up diskless swapping.
2110 error = swaponvp(td, vp, attr.va_size / DEV_BSIZE);
2116 swdev_syscall_active = 0;
2117 wakeup_one(&swdev_syscall_active);
2123 swaponsomething(struct vnode *vp, void *id, u_long nblks, sw_strategy_t *strategy, sw_close_t *close, dev_t dev)
2125 struct swdevt *sp, *tsp;
2130 * If we go beyond this, we get overflows in the radix
2133 mblocks = 0x40000000 / BLIST_META_RADIX;
2134 if (nblks > mblocks) {
2135 printf("WARNING: reducing size to maximum of %lu blocks per swap unit\n",
2140 * nblks is in DEV_BSIZE'd chunks, convert to PAGE_SIZE'd chunks.
2141 * First chop nblks off to page-align it, then convert.
2143 * sw->sw_nblks is in page-sized chunks now too.
2145 nblks &= ~(ctodb(1) - 1);
2146 nblks = dbtoc(nblks);
2148 sp = malloc(sizeof *sp, M_VMPGDATA, M_WAITOK | M_ZERO);
2153 sp->sw_nblks = nblks;
2155 sp->sw_strategy = strategy;
2156 sp->sw_close = close;
2158 sp->sw_blist = blist_create(nblks, M_WAITOK);
2160 * Do not free the first two block in order to avoid overwriting
2161 * any bsd label at the front of the partition
2163 blist_free(sp->sw_blist, 2, nblks - 2);
2166 mtx_lock(&sw_dev_mtx);
2167 TAILQ_FOREACH(tsp, &swtailq, sw_list) {
2168 if (tsp->sw_end >= dvbase) {
2170 * We put one uncovered page between the devices
2171 * in order to definitively prevent any cross-device
2174 dvbase = tsp->sw_end + 1;
2177 sp->sw_first = dvbase;
2178 sp->sw_end = dvbase + nblks;
2179 TAILQ_INSERT_TAIL(&swtailq, sp, sw_list);
2181 swap_pager_avail += nblks;
2182 swap_total += (vm_ooffset_t)nblks * PAGE_SIZE;
2184 mtx_unlock(&sw_dev_mtx);
2188 * SYSCALL: swapoff(devname)
2190 * Disable swapping on the given device.
2192 * XXX: Badly designed system call: it should use a device index
2193 * rather than filename as specification. We keep sw_vp around
2194 * only to make this work.
2196 #ifndef _SYS_SYSPROTO_H_
2197 struct swapoff_args {
2207 swapoff(struct thread *td, struct swapoff_args *uap)
2210 struct nameidata nd;
2214 error = priv_check(td, PRIV_SWAPOFF);
2219 while (swdev_syscall_active)
2220 tsleep(&swdev_syscall_active, PUSER - 1, "swpoff", 0);
2221 swdev_syscall_active = 1;
2223 NDINIT(&nd, LOOKUP, FOLLOW | AUDITVNODE1, UIO_USERSPACE, uap->name,
2228 NDFREE(&nd, NDF_ONLY_PNBUF);
2231 mtx_lock(&sw_dev_mtx);
2232 TAILQ_FOREACH(sp, &swtailq, sw_list) {
2233 if (sp->sw_vp == vp)
2236 mtx_unlock(&sw_dev_mtx);
2241 error = swapoff_one(sp, td->td_ucred);
2243 swdev_syscall_active = 0;
2244 wakeup_one(&swdev_syscall_active);
2250 swapoff_one(struct swdevt *sp, struct ucred *cred)
2252 u_long nblks, dvbase;
2257 mtx_assert(&Giant, MA_OWNED);
2259 (void) vn_lock(sp->sw_vp, LK_EXCLUSIVE | LK_RETRY);
2260 error = mac_system_check_swapoff(cred, sp->sw_vp);
2261 (void) VOP_UNLOCK(sp->sw_vp, 0);
2265 nblks = sp->sw_nblks;
2268 * We can turn off this swap device safely only if the
2269 * available virtual memory in the system will fit the amount
2270 * of data we will have to page back in, plus an epsilon so
2271 * the system doesn't become critically low on swap space.
2273 if (cnt.v_free_count + cnt.v_cache_count + swap_pager_avail <
2274 nblks + nswap_lowat) {
2279 * Prevent further allocations on this device.
2281 mtx_lock(&sw_dev_mtx);
2282 sp->sw_flags |= SW_CLOSING;
2283 for (dvbase = 0; dvbase < sp->sw_end; dvbase += dmmax) {
2284 swap_pager_avail -= blist_fill(sp->sw_blist,
2287 swap_total -= (vm_ooffset_t)nblks * PAGE_SIZE;
2288 mtx_unlock(&sw_dev_mtx);
2291 * Page in the contents of the device and close it.
2293 swap_pager_swapoff(sp);
2295 sp->sw_close(curthread, sp);
2297 mtx_lock(&sw_dev_mtx);
2298 TAILQ_REMOVE(&swtailq, sp, sw_list);
2300 if (nswapdev == 0) {
2301 swap_pager_full = 2;
2302 swap_pager_almost_full = 1;
2306 mtx_unlock(&sw_dev_mtx);
2307 blist_destroy(sp->sw_blist);
2308 free(sp, M_VMPGDATA);
2315 struct swdevt *sp, *spt;
2316 const char *devname;
2320 while (swdev_syscall_active)
2321 tsleep(&swdev_syscall_active, PUSER - 1, "swpoff", 0);
2322 swdev_syscall_active = 1;
2324 mtx_lock(&sw_dev_mtx);
2325 TAILQ_FOREACH_SAFE(sp, &swtailq, sw_list, spt) {
2326 mtx_unlock(&sw_dev_mtx);
2327 if (vn_isdisk(sp->sw_vp, NULL))
2328 devname = sp->sw_vp->v_rdev->si_name;
2331 error = swapoff_one(sp, thread0.td_ucred);
2333 printf("Cannot remove swap device %s (error=%d), "
2334 "skipping.\n", devname, error);
2335 } else if (bootverbose) {
2336 printf("Swap device %s removed.\n", devname);
2338 mtx_lock(&sw_dev_mtx);
2340 mtx_unlock(&sw_dev_mtx);
2342 swdev_syscall_active = 0;
2343 wakeup_one(&swdev_syscall_active);
2348 swap_pager_status(int *total, int *used)
2354 mtx_lock(&sw_dev_mtx);
2355 TAILQ_FOREACH(sp, &swtailq, sw_list) {
2356 *total += sp->sw_nblks;
2357 *used += sp->sw_used;
2359 mtx_unlock(&sw_dev_mtx);
2363 sysctl_vm_swap_info(SYSCTL_HANDLER_ARGS)
2365 int *name = (int *)arg1;
2370 if (arg2 != 1) /* name length */
2374 mtx_lock(&sw_dev_mtx);
2375 TAILQ_FOREACH(sp, &swtailq, sw_list) {
2377 mtx_unlock(&sw_dev_mtx);
2378 xs.xsw_version = XSWDEV_VERSION;
2379 xs.xsw_dev = sp->sw_dev;
2380 xs.xsw_flags = sp->sw_flags;
2381 xs.xsw_nblks = sp->sw_nblks;
2382 xs.xsw_used = sp->sw_used;
2384 error = SYSCTL_OUT(req, &xs, sizeof(xs));
2389 mtx_unlock(&sw_dev_mtx);
2393 SYSCTL_INT(_vm, OID_AUTO, nswapdev, CTLFLAG_RD, &nswapdev, 0,
2394 "Number of swap devices");
2395 SYSCTL_NODE(_vm, OID_AUTO, swap_info, CTLFLAG_RD, sysctl_vm_swap_info,
2396 "Swap statistics by device");
2399 * vmspace_swap_count() - count the approximate swap usage in pages for a
2402 * The map must be locked.
2404 * Swap usage is determined by taking the proportional swap used by
2405 * VM objects backing the VM map. To make up for fractional losses,
2406 * if the VM object has any swap use at all the associated map entries
2407 * count for at least 1 swap page.
2410 vmspace_swap_count(struct vmspace *vmspace)
2417 map = &vmspace->vm_map;
2420 for (cur = map->header.next; cur != &map->header; cur = cur->next) {
2421 if ((cur->eflags & MAP_ENTRY_IS_SUB_MAP) == 0 &&
2422 (object = cur->object.vm_object) != NULL) {
2423 VM_OBJECT_LOCK(object);
2424 if (object->type == OBJT_SWAP &&
2425 object->un_pager.swp.swp_bcount != 0) {
2426 n = (cur->end - cur->start) / PAGE_SIZE;
2427 count += object->un_pager.swp.swp_bcount *
2428 SWAP_META_PAGES * n / object->size + 1;
2430 VM_OBJECT_UNLOCK(object);
2439 * Swapping onto disk devices.
2443 static g_orphan_t swapgeom_orphan;
2445 static struct g_class g_swap_class = {
2447 .version = G_VERSION,
2448 .orphan = swapgeom_orphan,
2451 DECLARE_GEOM_CLASS(g_swap_class, g_class);
2455 swapgeom_done(struct bio *bp2)
2459 bp = bp2->bio_caller2;
2460 bp->b_ioflags = bp2->bio_flags;
2462 bp->b_ioflags |= BIO_ERROR;
2463 bp->b_resid = bp->b_bcount - bp2->bio_completed;
2464 bp->b_error = bp2->bio_error;
2470 swapgeom_strategy(struct buf *bp, struct swdevt *sp)
2473 struct g_consumer *cp;
2477 bp->b_error = ENXIO;
2478 bp->b_ioflags |= BIO_ERROR;
2482 if (bp->b_iocmd == BIO_WRITE)
2485 bio = g_alloc_bio();
2487 bp->b_error = ENOMEM;
2488 bp->b_ioflags |= BIO_ERROR;
2493 bio->bio_caller2 = bp;
2494 bio->bio_cmd = bp->b_iocmd;
2495 bio->bio_data = bp->b_data;
2496 bio->bio_offset = (bp->b_blkno - sp->sw_first) * PAGE_SIZE;
2497 bio->bio_length = bp->b_bcount;
2498 bio->bio_done = swapgeom_done;
2499 g_io_request(bio, cp);
2504 swapgeom_orphan(struct g_consumer *cp)
2508 mtx_lock(&sw_dev_mtx);
2509 TAILQ_FOREACH(sp, &swtailq, sw_list)
2510 if (sp->sw_id == cp)
2512 mtx_unlock(&sw_dev_mtx);
2516 swapgeom_close_ev(void *arg, int flags)
2518 struct g_consumer *cp;
2521 g_access(cp, -1, -1, 0);
2523 g_destroy_consumer(cp);
2527 swapgeom_close(struct thread *td, struct swdevt *sw)
2530 /* XXX: direct call when Giant untangled */
2531 g_waitfor_event(swapgeom_close_ev, sw->sw_id, M_WAITOK, NULL);
2542 swapongeom_ev(void *arg, int flags)
2545 struct g_provider *pp;
2546 struct g_consumer *cp;
2547 static struct g_geom *gp;
2554 pp = g_dev_getprovider(swh->dev);
2556 swh->error = ENODEV;
2559 mtx_lock(&sw_dev_mtx);
2560 TAILQ_FOREACH(sp, &swtailq, sw_list) {
2562 if (cp != NULL && cp->provider == pp) {
2563 mtx_unlock(&sw_dev_mtx);
2568 mtx_unlock(&sw_dev_mtx);
2570 gp = g_new_geomf(&g_swap_class, "swap", NULL);
2571 cp = g_new_consumer(gp);
2574 * XXX: Everytime you think you can improve the margin for
2575 * footshooting, somebody depends on the ability to do so:
2576 * savecore(8) wants to write to our swapdev so we cannot
2577 * set an exclusive count :-(
2579 error = g_access(cp, 1, 1, 0);
2582 g_destroy_consumer(cp);
2586 nblks = pp->mediasize / DEV_BSIZE;
2587 swaponsomething(swh->vp, cp, nblks, swapgeom_strategy,
2588 swapgeom_close, dev2udev(swh->dev));
2594 swapongeom(struct thread *td, struct vnode *vp)
2599 vn_lock(vp, LK_EXCLUSIVE | LK_RETRY);
2601 swh.dev = vp->v_rdev;
2604 /* XXX: direct call when Giant untangled */
2605 error = g_waitfor_event(swapongeom_ev, &swh, M_WAITOK, NULL);
2615 * This is used mainly for network filesystem (read: probably only tested
2616 * with NFS) swapfiles.
2621 swapdev_strategy(struct buf *bp, struct swdevt *sp)
2625 bp->b_blkno = ctodb(bp->b_blkno - sp->sw_first);
2629 if (bp->b_iocmd == BIO_WRITE) {
2631 bufobj_wdrop(bp->b_bufobj);
2632 bufobj_wref(&vp2->v_bufobj);
2634 if (bp->b_bufobj != &vp2->v_bufobj)
2635 bp->b_bufobj = &vp2->v_bufobj;
2637 bp->b_iooffset = dbtob(bp->b_blkno);
2643 swapdev_close(struct thread *td, struct swdevt *sp)
2646 VOP_CLOSE(sp->sw_vp, FREAD | FWRITE, td->td_ucred, td);
2652 swaponvp(struct thread *td, struct vnode *vp, u_long nblks)
2659 mtx_lock(&sw_dev_mtx);
2660 TAILQ_FOREACH(sp, &swtailq, sw_list) {
2661 if (sp->sw_id == vp) {
2662 mtx_unlock(&sw_dev_mtx);
2666 mtx_unlock(&sw_dev_mtx);
2668 (void) vn_lock(vp, LK_EXCLUSIVE | LK_RETRY);
2670 error = mac_system_check_swapon(td->td_ucred, vp);
2673 error = VOP_OPEN(vp, FREAD | FWRITE, td->td_ucred, td, NULL);
2674 (void) VOP_UNLOCK(vp, 0);
2678 swaponsomething(vp, vp, nblks, swapdev_strategy, swapdev_close,