2 * Copyright (c) 1998 Matthew Dillon,
3 * Copyright (c) 1994 John S. Dyson
4 * Copyright (c) 1990 University of Utah.
5 * Copyright (c) 1982, 1986, 1989, 1993
6 * The Regents of the University of California. All rights reserved.
8 * This code is derived from software contributed to Berkeley by
9 * the Systems Programming Group of the University of Utah Computer
12 * Redistribution and use in source and binary forms, with or without
13 * modification, are permitted provided that the following conditions
15 * 1. Redistributions of source code must retain the above copyright
16 * notice, this list of conditions and the following disclaimer.
17 * 2. Redistributions in binary form must reproduce the above copyright
18 * notice, this list of conditions and the following disclaimer in the
19 * documentation and/or other materials provided with the distribution.
20 * 3. All advertising materials mentioning features or use of this software
21 * must display the following acknowledgement:
22 * This product includes software developed by the University of
23 * California, Berkeley and its contributors.
24 * 4. Neither the name of the University nor the names of its contributors
25 * may be used to endorse or promote products derived from this software
26 * without specific prior written permission.
28 * THIS SOFTWARE IS PROVIDED BY THE REGENTS AND CONTRIBUTORS ``AS IS'' AND
29 * ANY EXPRESS OR IMPLIED WARRANTIES, INCLUDING, BUT NOT LIMITED TO, THE
30 * IMPLIED WARRANTIES OF MERCHANTABILITY AND FITNESS FOR A PARTICULAR PURPOSE
31 * ARE DISCLAIMED. IN NO EVENT SHALL THE REGENTS OR CONTRIBUTORS BE LIABLE
32 * FOR ANY DIRECT, INDIRECT, INCIDENTAL, SPECIAL, EXEMPLARY, OR CONSEQUENTIAL
33 * DAMAGES (INCLUDING, BUT NOT LIMITED TO, PROCUREMENT OF SUBSTITUTE GOODS
34 * OR SERVICES; LOSS OF USE, DATA, OR PROFITS; OR BUSINESS INTERRUPTION)
35 * HOWEVER CAUSED AND ON ANY THEORY OF LIABILITY, WHETHER IN CONTRACT, STRICT
36 * LIABILITY, OR TORT (INCLUDING NEGLIGENCE OR OTHERWISE) ARISING IN ANY WAY
37 * OUT OF THE USE OF THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF
43 * Radix Bitmap 'blists'.
45 * - The new swapper uses the new radix bitmap code. This should scale
46 * to arbitrarily small or arbitrarily large swap spaces and an almost
47 * arbitrary degree of fragmentation.
51 * - on the fly reallocation of swap during putpages. The new system
52 * does not try to keep previously allocated swap blocks for dirty
55 * - on the fly deallocation of swap
57 * - No more garbage collection required. Unnecessarily allocated swap
58 * blocks only exist for dirty vm_page_t's now and these are already
59 * cycled (in a high-load system) by the pager. We also do on-the-fly
60 * removal of invalidated swap blocks when a page is destroyed
63 * from: Utah $Hdr: swap_pager.c 1.4 91/04/30$
65 * @(#)swap_pager.c 8.9 (Berkeley) 3/21/94
66 * @(#)vm_swap.c 8.5 (Berkeley) 2/17/94
69 #include <sys/cdefs.h>
70 __FBSDID("$FreeBSD$");
75 #include <sys/param.h>
76 #include <sys/systm.h>
78 #include <sys/kernel.h>
84 #include <sys/fcntl.h>
85 #include <sys/mount.h>
86 #include <sys/namei.h>
87 #include <sys/vnode.h>
88 #include <sys/malloc.h>
89 #include <sys/resource.h>
90 #include <sys/resourcevar.h>
91 #include <sys/sysctl.h>
92 #include <sys/sysproto.h>
93 #include <sys/blist.h>
96 #include <sys/vmmeter.h>
98 #include <security/mac/mac_framework.h>
102 #include <vm/vm_map.h>
103 #include <vm/vm_kern.h>
104 #include <vm/vm_object.h>
105 #include <vm/vm_page.h>
106 #include <vm/vm_pager.h>
107 #include <vm/vm_pageout.h>
108 #include <vm/vm_param.h>
109 #include <vm/swap_pager.h>
110 #include <vm/vm_extern.h>
113 #include <geom/geom.h>
116 * SWB_NPAGES must be a power of 2. It may be set to 1, 2, 4, 8, or 16
117 * pages per allocation. We recommend you stick with the default of 8.
118 * The 16-page limit is due to the radix code (kern/subr_blist.c).
120 #ifndef MAX_PAGEOUT_CLUSTER
121 #define MAX_PAGEOUT_CLUSTER 16
124 #if !defined(SWB_NPAGES)
125 #define SWB_NPAGES MAX_PAGEOUT_CLUSTER
129 * Piecemeal swap metadata structure. Swap is stored in a radix tree.
131 * If SWB_NPAGES is 8 and sizeof(char *) == sizeof(daddr_t), our radix
132 * is basically 8. Assuming PAGE_SIZE == 4096, one tree level represents
133 * 32K worth of data, two levels represent 256K, three levels represent
134 * 2 MBytes. This is acceptable.
136 * Overall memory utilization is about the same as the old swap structure.
138 #define SWCORRECT(n) (sizeof(void *) * (n) / sizeof(daddr_t))
139 #define SWAP_META_PAGES (SWB_NPAGES * 2)
140 #define SWAP_META_MASK (SWAP_META_PAGES - 1)
143 struct swblock *swb_hnext;
144 vm_object_t swb_object;
145 vm_pindex_t swb_index;
147 daddr_t swb_pages[SWAP_META_PAGES];
150 static struct mtx sw_dev_mtx;
151 static TAILQ_HEAD(, swdevt) swtailq = TAILQ_HEAD_INITIALIZER(swtailq);
152 static struct swdevt *swdevhd; /* Allocate from here next */
153 static int nswapdev; /* Number of swap devices */
154 int swap_pager_avail;
155 static int swdev_syscall_active = 0; /* serialize swap(on|off) */
157 static vm_ooffset_t swap_total;
158 SYSCTL_QUAD(_vm, OID_AUTO, swap_total, CTLFLAG_RD, &swap_total, 0, "");
159 static vm_ooffset_t swap_reserved;
160 SYSCTL_QUAD(_vm, OID_AUTO, swap_reserved, CTLFLAG_RD, &swap_reserved, 0, "");
161 static int overcommit = 0;
162 SYSCTL_INT(_vm, OID_AUTO, overcommit, CTLFLAG_RW, &overcommit, 0, "");
164 /* bits from overcommit */
165 #define SWAP_RESERVE_FORCE_ON (1 << 0)
166 #define SWAP_RESERVE_RLIMIT_ON (1 << 1)
167 #define SWAP_RESERVE_ALLOW_NONWIRED (1 << 2)
170 swap_reserve(vm_ooffset_t incr)
173 return (swap_reserve_by_uid(incr, curthread->td_ucred->cr_ruidinfo));
177 swap_reserve_by_uid(vm_ooffset_t incr, struct uidinfo *uip)
179 vm_ooffset_t r, s, max;
182 static struct timeval lastfail;
184 if (incr & PAGE_MASK)
185 panic("swap_reserve: & PAGE_MASK");
188 error = priv_check(curthread, PRIV_VM_SWAP_NOQUOTA);
189 mtx_lock(&sw_dev_mtx);
190 r = swap_reserved + incr;
191 if (overcommit & SWAP_RESERVE_ALLOW_NONWIRED) {
192 s = cnt.v_page_count - cnt.v_free_reserved - cnt.v_wire_count;
197 if ((overcommit & SWAP_RESERVE_FORCE_ON) == 0 || r <= s ||
198 (error = priv_check(curthread, PRIV_VM_SWAP_NOQUOTA)) == 0) {
202 mtx_unlock(&sw_dev_mtx);
206 UIDINFO_VMSIZE_LOCK(uip);
207 error = priv_check(curthread, PRIV_VM_SWAP_NORLIMIT);
208 max = (error != 0) ? lim_cur(curproc, RLIMIT_SWAP) : 0;
209 if (max != 0 && uip->ui_vmsize + incr > max &&
210 (overcommit & SWAP_RESERVE_RLIMIT_ON) != 0)
213 uip->ui_vmsize += incr;
214 UIDINFO_VMSIZE_UNLOCK(uip);
215 PROC_UNLOCK(curproc);
217 mtx_lock(&sw_dev_mtx);
218 swap_reserved -= incr;
219 mtx_unlock(&sw_dev_mtx);
222 if (!res && ppsratecheck(&lastfail, &curfail, 1)) {
223 printf("uid %d, pid %d: swap reservation for %jd bytes failed\n",
224 curproc->p_pid, uip->ui_uid, incr);
231 swap_reserve_force(vm_ooffset_t incr)
235 mtx_lock(&sw_dev_mtx);
236 swap_reserved += incr;
237 mtx_unlock(&sw_dev_mtx);
239 uip = curthread->td_ucred->cr_ruidinfo;
241 UIDINFO_VMSIZE_LOCK(uip);
242 uip->ui_vmsize += incr;
243 UIDINFO_VMSIZE_UNLOCK(uip);
244 PROC_UNLOCK(curproc);
248 swap_release(vm_ooffset_t decr)
253 uip = curthread->td_ucred->cr_ruidinfo;
254 swap_release_by_uid(decr, uip);
255 PROC_UNLOCK(curproc);
259 swap_release_by_uid(vm_ooffset_t decr, struct uidinfo *uip)
262 if (decr & PAGE_MASK)
263 panic("swap_release: & PAGE_MASK");
265 mtx_lock(&sw_dev_mtx);
266 if (swap_reserved < decr)
267 panic("swap_reserved < decr");
268 swap_reserved -= decr;
269 mtx_unlock(&sw_dev_mtx);
271 UIDINFO_VMSIZE_LOCK(uip);
272 if (uip->ui_vmsize < decr)
273 printf("negative vmsize for uid = %d\n", uip->ui_uid);
274 uip->ui_vmsize -= decr;
275 UIDINFO_VMSIZE_UNLOCK(uip);
278 static void swapdev_strategy(struct buf *, struct swdevt *sw);
280 #define SWM_FREE 0x02 /* free, period */
281 #define SWM_POP 0x04 /* pop out */
283 int swap_pager_full = 2; /* swap space exhaustion (task killing) */
284 static int swap_pager_almost_full = 1; /* swap space exhaustion (w/hysteresis)*/
285 static int nsw_rcount; /* free read buffers */
286 static int nsw_wcount_sync; /* limit write buffers / synchronous */
287 static int nsw_wcount_async; /* limit write buffers / asynchronous */
288 static int nsw_wcount_async_max;/* assigned maximum */
289 static int nsw_cluster_max; /* maximum VOP I/O allowed */
291 static struct swblock **swhash;
292 static int swhash_mask;
293 static struct mtx swhash_mtx;
295 static int swap_async_max = 4; /* maximum in-progress async I/O's */
296 static struct sx sw_alloc_sx;
299 SYSCTL_INT(_vm, OID_AUTO, swap_async_max,
300 CTLFLAG_RW, &swap_async_max, 0, "Maximum running async swap ops");
303 * "named" and "unnamed" anon region objects. Try to reduce the overhead
304 * of searching a named list by hashing it just a little.
309 #define NOBJLIST(handle) \
310 (&swap_pager_object_list[((int)(intptr_t)handle >> 4) & (NOBJLISTS-1)])
312 static struct mtx sw_alloc_mtx; /* protect list manipulation */
313 static struct pagerlst swap_pager_object_list[NOBJLISTS];
314 static uma_zone_t swap_zone;
315 static struct vm_object swap_zone_obj;
318 * pagerops for OBJT_SWAP - "swap pager". Some ops are also global procedure
319 * calls hooked from other parts of the VM system and do not appear here.
320 * (see vm/swap_pager.h).
323 swap_pager_alloc(void *handle, vm_ooffset_t size,
324 vm_prot_t prot, vm_ooffset_t offset, struct ucred *);
325 static void swap_pager_dealloc(vm_object_t object);
326 static int swap_pager_getpages(vm_object_t, vm_page_t *, int, int);
327 static void swap_pager_putpages(vm_object_t, vm_page_t *, int, boolean_t, int *);
329 swap_pager_haspage(vm_object_t object, vm_pindex_t pindex, int *before, int *after);
330 static void swap_pager_init(void);
331 static void swap_pager_unswapped(vm_page_t);
332 static void swap_pager_swapoff(struct swdevt *sp);
334 struct pagerops swappagerops = {
335 .pgo_init = swap_pager_init, /* early system initialization of pager */
336 .pgo_alloc = swap_pager_alloc, /* allocate an OBJT_SWAP object */
337 .pgo_dealloc = swap_pager_dealloc, /* deallocate an OBJT_SWAP object */
338 .pgo_getpages = swap_pager_getpages, /* pagein */
339 .pgo_putpages = swap_pager_putpages, /* pageout */
340 .pgo_haspage = swap_pager_haspage, /* get backing store status for page */
341 .pgo_pageunswapped = swap_pager_unswapped, /* remove swap related to page */
345 * dmmax is in page-sized chunks with the new swap system. It was
346 * dev-bsized chunks in the old. dmmax is always a power of 2.
348 * swap_*() routines are externally accessible. swp_*() routines are
352 static int nswap_lowat = 128; /* in pages, swap_pager_almost_full warn */
353 static int nswap_hiwat = 512; /* in pages, swap_pager_almost_full warn */
355 SYSCTL_INT(_vm, OID_AUTO, dmmax,
356 CTLFLAG_RD, &dmmax, 0, "Maximum size of a swap block");
358 static void swp_sizecheck(void);
359 static void swp_pager_async_iodone(struct buf *bp);
360 static int swapongeom(struct thread *, struct vnode *);
361 static int swaponvp(struct thread *, struct vnode *, u_long);
362 static int swapoff_one(struct swdevt *sp, struct ucred *cred);
365 * Swap bitmap functions
367 static void swp_pager_freeswapspace(daddr_t blk, int npages);
368 static daddr_t swp_pager_getswapspace(int npages);
373 static struct swblock **swp_pager_hash(vm_object_t object, vm_pindex_t index);
374 static void swp_pager_meta_build(vm_object_t, vm_pindex_t, daddr_t);
375 static void swp_pager_meta_free(vm_object_t, vm_pindex_t, daddr_t);
376 static void swp_pager_meta_free_all(vm_object_t);
377 static daddr_t swp_pager_meta_ctl(vm_object_t, vm_pindex_t, int);
380 * SWP_SIZECHECK() - update swap_pager_full indication
382 * update the swap_pager_almost_full indication and warn when we are
383 * about to run out of swap space, using lowat/hiwat hysteresis.
385 * Clear swap_pager_full ( task killing ) indication when lowat is met.
387 * No restrictions on call
388 * This routine may not block.
389 * This routine must be called at splvm()
395 if (swap_pager_avail < nswap_lowat) {
396 if (swap_pager_almost_full == 0) {
397 printf("swap_pager: out of swap space\n");
398 swap_pager_almost_full = 1;
402 if (swap_pager_avail > nswap_hiwat)
403 swap_pager_almost_full = 0;
408 * SWP_PAGER_HASH() - hash swap meta data
410 * This is an helper function which hashes the swapblk given
411 * the object and page index. It returns a pointer to a pointer
412 * to the object, or a pointer to a NULL pointer if it could not
415 * This routine must be called at splvm().
417 static struct swblock **
418 swp_pager_hash(vm_object_t object, vm_pindex_t index)
420 struct swblock **pswap;
421 struct swblock *swap;
423 index &= ~(vm_pindex_t)SWAP_META_MASK;
424 pswap = &swhash[(index ^ (int)(intptr_t)object) & swhash_mask];
425 while ((swap = *pswap) != NULL) {
426 if (swap->swb_object == object &&
427 swap->swb_index == index
431 pswap = &swap->swb_hnext;
437 * SWAP_PAGER_INIT() - initialize the swap pager!
439 * Expected to be started from system init. NOTE: This code is run
440 * before much else so be careful what you depend on. Most of the VM
441 * system has yet to be initialized at this point.
444 swap_pager_init(void)
447 * Initialize object lists
451 for (i = 0; i < NOBJLISTS; ++i)
452 TAILQ_INIT(&swap_pager_object_list[i]);
453 mtx_init(&sw_alloc_mtx, "swap_pager list", NULL, MTX_DEF);
454 mtx_init(&sw_dev_mtx, "swapdev", NULL, MTX_DEF);
457 * Device Stripe, in PAGE_SIZE'd blocks
459 dmmax = SWB_NPAGES * 2;
463 * SWAP_PAGER_SWAP_INIT() - swap pager initialization from pageout process
465 * Expected to be started from pageout process once, prior to entering
469 swap_pager_swap_init(void)
474 * Number of in-transit swap bp operations. Don't
475 * exhaust the pbufs completely. Make sure we
476 * initialize workable values (0 will work for hysteresis
477 * but it isn't very efficient).
479 * The nsw_cluster_max is constrained by the bp->b_pages[]
480 * array (MAXPHYS/PAGE_SIZE) and our locally defined
481 * MAX_PAGEOUT_CLUSTER. Also be aware that swap ops are
482 * constrained by the swap device interleave stripe size.
484 * Currently we hardwire nsw_wcount_async to 4. This limit is
485 * designed to prevent other I/O from having high latencies due to
486 * our pageout I/O. The value 4 works well for one or two active swap
487 * devices but is probably a little low if you have more. Even so,
488 * a higher value would probably generate only a limited improvement
489 * with three or four active swap devices since the system does not
490 * typically have to pageout at extreme bandwidths. We will want
491 * at least 2 per swap devices, and 4 is a pretty good value if you
492 * have one NFS swap device due to the command/ack latency over NFS.
493 * So it all works out pretty well.
495 nsw_cluster_max = min((MAXPHYS/PAGE_SIZE), MAX_PAGEOUT_CLUSTER);
498 nsw_rcount = (nswbuf + 1) / 2;
499 nsw_wcount_sync = (nswbuf + 3) / 4;
500 nsw_wcount_async = 4;
501 nsw_wcount_async_max = nsw_wcount_async;
502 mtx_unlock(&pbuf_mtx);
505 * Initialize our zone. Right now I'm just guessing on the number
506 * we need based on the number of pages in the system. Each swblock
507 * can hold 16 pages, so this is probably overkill. This reservation
508 * is typically limited to around 32MB by default.
510 n = cnt.v_page_count / 2;
511 if (maxswzone && n > maxswzone / sizeof(struct swblock))
512 n = maxswzone / sizeof(struct swblock);
514 swap_zone = uma_zcreate("SWAPMETA", sizeof(struct swblock), NULL, NULL,
515 NULL, NULL, UMA_ALIGN_PTR, UMA_ZONE_NOFREE | UMA_ZONE_VM);
516 if (swap_zone == NULL)
517 panic("failed to create swap_zone.");
519 if (uma_zone_set_obj(swap_zone, &swap_zone_obj, n))
522 * if the allocation failed, try a zone two thirds the
523 * size of the previous attempt.
528 printf("Swap zone entries reduced from %d to %d.\n", n2, n);
532 * Initialize our meta-data hash table. The swapper does not need to
533 * be quite as efficient as the VM system, so we do not use an
534 * oversized hash table.
536 * n: size of hash table, must be power of 2
537 * swhash_mask: hash table index mask
539 for (n = 1; n < n2 / 8; n *= 2)
541 swhash = malloc(sizeof(struct swblock *) * n, M_VMPGDATA, M_WAITOK | M_ZERO);
543 mtx_init(&swhash_mtx, "swap_pager swhash", NULL, MTX_DEF);
547 * SWAP_PAGER_ALLOC() - allocate a new OBJT_SWAP VM object and instantiate
548 * its metadata structures.
550 * This routine is called from the mmap and fork code to create a new
551 * OBJT_SWAP object. We do this by creating an OBJT_DEFAULT object
552 * and then converting it with swp_pager_meta_build().
554 * This routine may block in vm_object_allocate() and create a named
555 * object lookup race, so we must interlock. We must also run at
556 * splvm() for the object lookup to handle races with interrupts, but
557 * we do not have to maintain splvm() in between the lookup and the
558 * add because (I believe) it is not possible to attempt to create
559 * a new swap object w/handle when a default object with that handle
565 swap_pager_alloc(void *handle, vm_ooffset_t size, vm_prot_t prot,
566 vm_ooffset_t offset, struct ucred *cred)
573 pindex = OFF_TO_IDX(offset + PAGE_MASK + size);
577 * Reference existing named region or allocate new one. There
578 * should not be a race here against swp_pager_meta_build()
579 * as called from vm_page_remove() in regards to the lookup
582 sx_xlock(&sw_alloc_sx);
583 object = vm_pager_object_lookup(NOBJLIST(handle), handle);
584 if (object == NULL) {
586 uip = cred->cr_ruidinfo;
587 if (!swap_reserve_by_uid(size, uip)) {
588 sx_xunlock(&sw_alloc_sx);
594 object = vm_object_allocate(OBJT_DEFAULT, pindex);
595 VM_OBJECT_LOCK(object);
596 object->handle = handle;
599 object->charge = size;
601 swp_pager_meta_build(object, 0, SWAPBLK_NONE);
602 VM_OBJECT_UNLOCK(object);
604 sx_xunlock(&sw_alloc_sx);
608 uip = cred->cr_ruidinfo;
609 if (!swap_reserve_by_uid(size, uip))
613 object = vm_object_allocate(OBJT_DEFAULT, pindex);
614 VM_OBJECT_LOCK(object);
617 object->charge = size;
619 swp_pager_meta_build(object, 0, SWAPBLK_NONE);
620 VM_OBJECT_UNLOCK(object);
626 * SWAP_PAGER_DEALLOC() - remove swap metadata from object
628 * The swap backing for the object is destroyed. The code is
629 * designed such that we can reinstantiate it later, but this
630 * routine is typically called only when the entire object is
631 * about to be destroyed.
633 * This routine may block, but no longer does.
635 * The object must be locked or unreferenceable.
638 swap_pager_dealloc(vm_object_t object)
642 * Remove from list right away so lookups will fail if we block for
643 * pageout completion.
645 if (object->handle != NULL) {
646 mtx_lock(&sw_alloc_mtx);
647 TAILQ_REMOVE(NOBJLIST(object->handle), object, pager_object_list);
648 mtx_unlock(&sw_alloc_mtx);
651 VM_OBJECT_LOCK_ASSERT(object, MA_OWNED);
652 vm_object_pip_wait(object, "swpdea");
655 * Free all remaining metadata. We only bother to free it from
656 * the swap meta data. We do not attempt to free swapblk's still
657 * associated with vm_page_t's for this object. We do not care
658 * if paging is still in progress on some objects.
660 swp_pager_meta_free_all(object);
663 /************************************************************************
664 * SWAP PAGER BITMAP ROUTINES *
665 ************************************************************************/
668 * SWP_PAGER_GETSWAPSPACE() - allocate raw swap space
670 * Allocate swap for the requested number of pages. The starting
671 * swap block number (a page index) is returned or SWAPBLK_NONE
672 * if the allocation failed.
674 * Also has the side effect of advising that somebody made a mistake
675 * when they configured swap and didn't configure enough.
677 * Must be called at splvm() to avoid races with bitmap frees from
678 * vm_page_remove() aka swap_pager_page_removed().
680 * This routine may not block
681 * This routine must be called at splvm().
683 * We allocate in round-robin fashion from the configured devices.
686 swp_pager_getswapspace(int npages)
693 mtx_lock(&sw_dev_mtx);
695 for (i = 0; i < nswapdev; i++) {
697 sp = TAILQ_FIRST(&swtailq);
698 if (!(sp->sw_flags & SW_CLOSING)) {
699 blk = blist_alloc(sp->sw_blist, npages);
700 if (blk != SWAPBLK_NONE) {
702 sp->sw_used += npages;
703 swap_pager_avail -= npages;
705 swdevhd = TAILQ_NEXT(sp, sw_list);
709 sp = TAILQ_NEXT(sp, sw_list);
711 if (swap_pager_full != 2) {
712 printf("swap_pager_getswapspace(%d): failed\n", npages);
714 swap_pager_almost_full = 1;
718 mtx_unlock(&sw_dev_mtx);
723 swp_pager_isondev(daddr_t blk, struct swdevt *sp)
726 return (blk >= sp->sw_first && blk < sp->sw_end);
730 swp_pager_strategy(struct buf *bp)
734 mtx_lock(&sw_dev_mtx);
735 TAILQ_FOREACH(sp, &swtailq, sw_list) {
736 if (bp->b_blkno >= sp->sw_first && bp->b_blkno < sp->sw_end) {
737 mtx_unlock(&sw_dev_mtx);
738 sp->sw_strategy(bp, sp);
742 panic("Swapdev not found");
747 * SWP_PAGER_FREESWAPSPACE() - free raw swap space
749 * This routine returns the specified swap blocks back to the bitmap.
751 * Note: This routine may not block (it could in the old swap code),
752 * and through the use of the new blist routines it does not block.
754 * We must be called at splvm() to avoid races with bitmap frees from
755 * vm_page_remove() aka swap_pager_page_removed().
757 * This routine may not block
758 * This routine must be called at splvm().
761 swp_pager_freeswapspace(daddr_t blk, int npages)
765 mtx_lock(&sw_dev_mtx);
766 TAILQ_FOREACH(sp, &swtailq, sw_list) {
767 if (blk >= sp->sw_first && blk < sp->sw_end) {
768 sp->sw_used -= npages;
770 * If we are attempting to stop swapping on
771 * this device, we don't want to mark any
772 * blocks free lest they be reused.
774 if ((sp->sw_flags & SW_CLOSING) == 0) {
775 blist_free(sp->sw_blist, blk - sp->sw_first,
777 swap_pager_avail += npages;
780 mtx_unlock(&sw_dev_mtx);
784 panic("Swapdev not found");
788 * SWAP_PAGER_FREESPACE() - frees swap blocks associated with a page
789 * range within an object.
791 * This is a globally accessible routine.
793 * This routine removes swapblk assignments from swap metadata.
795 * The external callers of this routine typically have already destroyed
796 * or renamed vm_page_t's associated with this range in the object so
799 * This routine may be called at any spl. We up our spl to splvm temporarily
800 * in order to perform the metadata removal.
803 swap_pager_freespace(vm_object_t object, vm_pindex_t start, vm_size_t size)
806 VM_OBJECT_LOCK_ASSERT(object, MA_OWNED);
807 swp_pager_meta_free(object, start, size);
811 * SWAP_PAGER_RESERVE() - reserve swap blocks in object
813 * Assigns swap blocks to the specified range within the object. The
814 * swap blocks are not zerod. Any previous swap assignment is destroyed.
816 * Returns 0 on success, -1 on failure.
819 swap_pager_reserve(vm_object_t object, vm_pindex_t start, vm_size_t size)
822 daddr_t blk = SWAPBLK_NONE;
823 vm_pindex_t beg = start; /* save start index */
825 VM_OBJECT_LOCK(object);
829 while ((blk = swp_pager_getswapspace(n)) == SWAPBLK_NONE) {
832 swp_pager_meta_free(object, beg, start - beg);
833 VM_OBJECT_UNLOCK(object);
838 swp_pager_meta_build(object, start, blk);
844 swp_pager_meta_free(object, start, n);
845 VM_OBJECT_UNLOCK(object);
850 * SWAP_PAGER_COPY() - copy blocks from source pager to destination pager
851 * and destroy the source.
853 * Copy any valid swapblks from the source to the destination. In
854 * cases where both the source and destination have a valid swapblk,
855 * we keep the destination's.
857 * This routine is allowed to block. It may block allocating metadata
858 * indirectly through swp_pager_meta_build() or if paging is still in
859 * progress on the source.
861 * This routine can be called at any spl
863 * XXX vm_page_collapse() kinda expects us not to block because we
864 * supposedly do not need to allocate memory, but for the moment we
865 * *may* have to get a little memory from the zone allocator, but
866 * it is taken from the interrupt memory. We should be ok.
868 * The source object contains no vm_page_t's (which is just as well)
870 * The source object is of type OBJT_SWAP.
872 * The source and destination objects must be locked or
873 * inaccessible (XXX are they ?)
876 swap_pager_copy(vm_object_t srcobject, vm_object_t dstobject,
877 vm_pindex_t offset, int destroysource)
881 VM_OBJECT_LOCK_ASSERT(srcobject, MA_OWNED);
882 VM_OBJECT_LOCK_ASSERT(dstobject, MA_OWNED);
885 * If destroysource is set, we remove the source object from the
886 * swap_pager internal queue now.
889 if (srcobject->handle != NULL) {
890 mtx_lock(&sw_alloc_mtx);
892 NOBJLIST(srcobject->handle),
896 mtx_unlock(&sw_alloc_mtx);
901 * transfer source to destination.
903 for (i = 0; i < dstobject->size; ++i) {
907 * Locate (without changing) the swapblk on the destination,
908 * unless it is invalid in which case free it silently, or
909 * if the destination is a resident page, in which case the
910 * source is thrown away.
912 dstaddr = swp_pager_meta_ctl(dstobject, i, 0);
914 if (dstaddr == SWAPBLK_NONE) {
916 * Destination has no swapblk and is not resident,
921 srcaddr = swp_pager_meta_ctl(
927 if (srcaddr != SWAPBLK_NONE) {
929 * swp_pager_meta_build() can sleep.
931 vm_object_pip_add(srcobject, 1);
932 VM_OBJECT_UNLOCK(srcobject);
933 vm_object_pip_add(dstobject, 1);
934 swp_pager_meta_build(dstobject, i, srcaddr);
935 vm_object_pip_wakeup(dstobject);
936 VM_OBJECT_LOCK(srcobject);
937 vm_object_pip_wakeup(srcobject);
941 * Destination has valid swapblk or it is represented
942 * by a resident page. We destroy the sourceblock.
945 swp_pager_meta_ctl(srcobject, i + offset, SWM_FREE);
950 * Free left over swap blocks in source.
952 * We have to revert the type to OBJT_DEFAULT so we do not accidently
953 * double-remove the object from the swap queues.
956 swp_pager_meta_free_all(srcobject);
958 * Reverting the type is not necessary, the caller is going
959 * to destroy srcobject directly, but I'm doing it here
960 * for consistency since we've removed the object from its
963 srcobject->type = OBJT_DEFAULT;
968 * SWAP_PAGER_HASPAGE() - determine if we have good backing store for
969 * the requested page.
971 * We determine whether good backing store exists for the requested
972 * page and return TRUE if it does, FALSE if it doesn't.
974 * If TRUE, we also try to determine how much valid, contiguous backing
975 * store exists before and after the requested page within a reasonable
976 * distance. We do not try to restrict it to the swap device stripe
977 * (that is handled in getpages/putpages). It probably isn't worth
981 swap_pager_haspage(vm_object_t object, vm_pindex_t pindex, int *before, int *after)
985 VM_OBJECT_LOCK_ASSERT(object, MA_OWNED);
987 * do we have good backing store at the requested index ?
989 blk0 = swp_pager_meta_ctl(object, pindex, 0);
991 if (blk0 == SWAPBLK_NONE) {
1000 * find backwards-looking contiguous good backing store
1002 if (before != NULL) {
1005 for (i = 1; i < (SWB_NPAGES/2); ++i) {
1010 blk = swp_pager_meta_ctl(object, pindex - i, 0);
1011 if (blk != blk0 - i)
1018 * find forward-looking contiguous good backing store
1020 if (after != NULL) {
1023 for (i = 1; i < (SWB_NPAGES/2); ++i) {
1026 blk = swp_pager_meta_ctl(object, pindex + i, 0);
1027 if (blk != blk0 + i)
1036 * SWAP_PAGER_PAGE_UNSWAPPED() - remove swap backing store related to page
1038 * This removes any associated swap backing store, whether valid or
1039 * not, from the page.
1041 * This routine is typically called when a page is made dirty, at
1042 * which point any associated swap can be freed. MADV_FREE also
1043 * calls us in a special-case situation
1045 * NOTE!!! If the page is clean and the swap was valid, the caller
1046 * should make the page dirty before calling this routine. This routine
1047 * does NOT change the m->dirty status of the page. Also: MADV_FREE
1050 * This routine may not block
1051 * This routine must be called at splvm()
1054 swap_pager_unswapped(vm_page_t m)
1057 VM_OBJECT_LOCK_ASSERT(m->object, MA_OWNED);
1058 swp_pager_meta_ctl(m->object, m->pindex, SWM_FREE);
1062 * SWAP_PAGER_GETPAGES() - bring pages in from swap
1064 * Attempt to retrieve (m, count) pages from backing store, but make
1065 * sure we retrieve at least m[reqpage]. We try to load in as large
1066 * a chunk surrounding m[reqpage] as is contiguous in swap and which
1067 * belongs to the same object.
1069 * The code is designed for asynchronous operation and
1070 * immediate-notification of 'reqpage' but tends not to be
1071 * used that way. Please do not optimize-out this algorithmic
1072 * feature, I intend to improve on it in the future.
1074 * The parent has a single vm_object_pip_add() reference prior to
1075 * calling us and we should return with the same.
1077 * The parent has BUSY'd the pages. We should return with 'm'
1078 * left busy, but the others adjusted.
1081 swap_pager_getpages(vm_object_t object, vm_page_t *m, int count, int reqpage)
1091 KASSERT(mreq->object == object,
1092 ("swap_pager_getpages: object mismatch %p/%p",
1093 object, mreq->object));
1096 * Calculate range to retrieve. The pages have already been assigned
1097 * their swapblks. We require a *contiguous* range but we know it to
1098 * not span devices. If we do not supply it, bad things
1099 * happen. Note that blk, iblk & jblk can be SWAPBLK_NONE, but the
1100 * loops are set up such that the case(s) are handled implicitly.
1102 * The swp_*() calls must be made at splvm(). vm_page_free() does
1103 * not need to be, but it will go a little faster if it is.
1105 blk = swp_pager_meta_ctl(mreq->object, mreq->pindex, 0);
1107 for (i = reqpage - 1; i >= 0; --i) {
1110 iblk = swp_pager_meta_ctl(m[i]->object, m[i]->pindex, 0);
1111 if (blk != iblk + (reqpage - i))
1116 for (j = reqpage + 1; j < count; ++j) {
1119 jblk = swp_pager_meta_ctl(m[j]->object, m[j]->pindex, 0);
1120 if (blk != jblk - (j - reqpage))
1125 * free pages outside our collection range. Note: we never free
1126 * mreq, it must remain busy throughout.
1128 if (0 < i || j < count) {
1131 vm_page_lock_queues();
1132 for (k = 0; k < i; ++k)
1134 for (k = j; k < count; ++k)
1136 vm_page_unlock_queues();
1140 * Return VM_PAGER_FAIL if we have nothing to do. Return mreq
1141 * still busy, but the others unbusied.
1143 if (blk == SWAPBLK_NONE)
1144 return (VM_PAGER_FAIL);
1147 * Getpbuf() can sleep.
1149 VM_OBJECT_UNLOCK(object);
1151 * Get a swap buffer header to perform the IO
1153 bp = getpbuf(&nsw_rcount);
1154 bp->b_flags |= B_PAGING;
1157 * map our page(s) into kva for input
1159 pmap_qenter((vm_offset_t)bp->b_data, m + i, j - i);
1161 bp->b_iocmd = BIO_READ;
1162 bp->b_iodone = swp_pager_async_iodone;
1163 bp->b_rcred = crhold(thread0.td_ucred);
1164 bp->b_wcred = crhold(thread0.td_ucred);
1165 bp->b_blkno = blk - (reqpage - i);
1166 bp->b_bcount = PAGE_SIZE * (j - i);
1167 bp->b_bufsize = PAGE_SIZE * (j - i);
1168 bp->b_pager.pg_reqpage = reqpage - i;
1170 VM_OBJECT_LOCK(object);
1174 for (k = i; k < j; ++k) {
1175 bp->b_pages[k - i] = m[k];
1176 m[k]->oflags |= VPO_SWAPINPROG;
1179 bp->b_npages = j - i;
1181 PCPU_INC(cnt.v_swapin);
1182 PCPU_ADD(cnt.v_swappgsin, bp->b_npages);
1185 * We still hold the lock on mreq, and our automatic completion routine
1186 * does not remove it.
1188 vm_object_pip_add(object, bp->b_npages);
1189 VM_OBJECT_UNLOCK(object);
1192 * perform the I/O. NOTE!!! bp cannot be considered valid after
1193 * this point because we automatically release it on completion.
1194 * Instead, we look at the one page we are interested in which we
1195 * still hold a lock on even through the I/O completion.
1197 * The other pages in our m[] array are also released on completion,
1198 * so we cannot assume they are valid anymore either.
1200 * NOTE: b_blkno is destroyed by the call to swapdev_strategy
1203 swp_pager_strategy(bp);
1206 * wait for the page we want to complete. VPO_SWAPINPROG is always
1207 * cleared on completion. If an I/O error occurs, SWAPBLK_NONE
1208 * is set in the meta-data.
1210 VM_OBJECT_LOCK(object);
1211 while ((mreq->oflags & VPO_SWAPINPROG) != 0) {
1212 mreq->oflags |= VPO_WANTED;
1213 vm_page_lock_queues();
1214 vm_page_flag_set(mreq, PG_REFERENCED);
1215 vm_page_unlock_queues();
1216 PCPU_INC(cnt.v_intrans);
1217 if (msleep(mreq, VM_OBJECT_MTX(object), PSWP, "swread", hz*20)) {
1219 "swap_pager: indefinite wait buffer: bufobj: %p, blkno: %jd, size: %ld\n",
1220 bp->b_bufobj, (intmax_t)bp->b_blkno, bp->b_bcount);
1225 * mreq is left busied after completion, but all the other pages
1226 * are freed. If we had an unrecoverable read error the page will
1229 if (mreq->valid != VM_PAGE_BITS_ALL) {
1230 return (VM_PAGER_ERROR);
1232 return (VM_PAGER_OK);
1236 * A final note: in a low swap situation, we cannot deallocate swap
1237 * and mark a page dirty here because the caller is likely to mark
1238 * the page clean when we return, causing the page to possibly revert
1239 * to all-zero's later.
1244 * swap_pager_putpages:
1246 * Assign swap (if necessary) and initiate I/O on the specified pages.
1248 * We support both OBJT_DEFAULT and OBJT_SWAP objects. DEFAULT objects
1249 * are automatically converted to SWAP objects.
1251 * In a low memory situation we may block in VOP_STRATEGY(), but the new
1252 * vm_page reservation system coupled with properly written VFS devices
1253 * should ensure that no low-memory deadlock occurs. This is an area
1256 * The parent has N vm_object_pip_add() references prior to
1257 * calling us and will remove references for rtvals[] that are
1258 * not set to VM_PAGER_PEND. We need to remove the rest on I/O
1261 * The parent has soft-busy'd the pages it passes us and will unbusy
1262 * those whos rtvals[] entry is not set to VM_PAGER_PEND on return.
1263 * We need to unbusy the rest on I/O completion.
1266 swap_pager_putpages(vm_object_t object, vm_page_t *m, int count,
1267 boolean_t sync, int *rtvals)
1272 if (count && m[0]->object != object) {
1273 panic("swap_pager_putpages: object mismatch %p/%p",
1282 * Turn object into OBJT_SWAP
1283 * check for bogus sysops
1284 * force sync if not pageout process
1286 if (object->type != OBJT_SWAP)
1287 swp_pager_meta_build(object, 0, SWAPBLK_NONE);
1288 VM_OBJECT_UNLOCK(object);
1290 if (curproc != pageproc)
1296 * Update nsw parameters from swap_async_max sysctl values.
1297 * Do not let the sysop crash the machine with bogus numbers.
1299 mtx_lock(&pbuf_mtx);
1300 if (swap_async_max != nsw_wcount_async_max) {
1306 if ((n = swap_async_max) > nswbuf / 2)
1313 * Adjust difference ( if possible ). If the current async
1314 * count is too low, we may not be able to make the adjustment
1317 n -= nsw_wcount_async_max;
1318 if (nsw_wcount_async + n >= 0) {
1319 nsw_wcount_async += n;
1320 nsw_wcount_async_max += n;
1321 wakeup(&nsw_wcount_async);
1324 mtx_unlock(&pbuf_mtx);
1329 * Assign swap blocks and issue I/O. We reallocate swap on the fly.
1330 * The page is left dirty until the pageout operation completes
1333 for (i = 0; i < count; i += n) {
1339 * Maximum I/O size is limited by a number of factors.
1341 n = min(BLIST_MAX_ALLOC, count - i);
1342 n = min(n, nsw_cluster_max);
1345 * Get biggest block of swap we can. If we fail, fall
1346 * back and try to allocate a smaller block. Don't go
1347 * overboard trying to allocate space if it would overly
1351 (blk = swp_pager_getswapspace(n)) == SWAPBLK_NONE &&
1356 if (blk == SWAPBLK_NONE) {
1357 for (j = 0; j < n; ++j)
1358 rtvals[i+j] = VM_PAGER_FAIL;
1363 * All I/O parameters have been satisfied, build the I/O
1364 * request and assign the swap space.
1367 bp = getpbuf(&nsw_wcount_sync);
1369 bp = getpbuf(&nsw_wcount_async);
1370 bp->b_flags = B_ASYNC;
1372 bp->b_flags |= B_PAGING;
1373 bp->b_iocmd = BIO_WRITE;
1375 pmap_qenter((vm_offset_t)bp->b_data, &m[i], n);
1377 bp->b_rcred = crhold(thread0.td_ucred);
1378 bp->b_wcred = crhold(thread0.td_ucred);
1379 bp->b_bcount = PAGE_SIZE * n;
1380 bp->b_bufsize = PAGE_SIZE * n;
1383 VM_OBJECT_LOCK(object);
1384 for (j = 0; j < n; ++j) {
1385 vm_page_t mreq = m[i+j];
1387 swp_pager_meta_build(
1392 vm_page_dirty(mreq);
1393 rtvals[i+j] = VM_PAGER_OK;
1395 mreq->oflags |= VPO_SWAPINPROG;
1396 bp->b_pages[j] = mreq;
1398 VM_OBJECT_UNLOCK(object);
1401 * Must set dirty range for NFS to work.
1404 bp->b_dirtyend = bp->b_bcount;
1406 PCPU_INC(cnt.v_swapout);
1407 PCPU_ADD(cnt.v_swappgsout, bp->b_npages);
1412 * NOTE: b_blkno is destroyed by the call to swapdev_strategy
1414 if (sync == FALSE) {
1415 bp->b_iodone = swp_pager_async_iodone;
1417 swp_pager_strategy(bp);
1419 for (j = 0; j < n; ++j)
1420 rtvals[i+j] = VM_PAGER_PEND;
1421 /* restart outter loop */
1428 * NOTE: b_blkno is destroyed by the call to swapdev_strategy
1430 bp->b_iodone = bdone;
1431 swp_pager_strategy(bp);
1434 * Wait for the sync I/O to complete, then update rtvals.
1435 * We just set the rtvals[] to VM_PAGER_PEND so we can call
1436 * our async completion routine at the end, thus avoiding a
1439 bwait(bp, PVM, "swwrt");
1440 for (j = 0; j < n; ++j)
1441 rtvals[i+j] = VM_PAGER_PEND;
1443 * Now that we are through with the bp, we can call the
1444 * normal async completion, which frees everything up.
1446 swp_pager_async_iodone(bp);
1448 VM_OBJECT_LOCK(object);
1452 * swp_pager_async_iodone:
1454 * Completion routine for asynchronous reads and writes from/to swap.
1455 * Also called manually by synchronous code to finish up a bp.
1457 * For READ operations, the pages are PG_BUSY'd. For WRITE operations,
1458 * the pages are vm_page_t->busy'd. For READ operations, we PG_BUSY
1459 * unbusy all pages except the 'main' request page. For WRITE
1460 * operations, we vm_page_t->busy'd unbusy all pages ( we can do this
1461 * because we marked them all VM_PAGER_PEND on return from putpages ).
1463 * This routine may not block.
1466 swp_pager_async_iodone(struct buf *bp)
1469 vm_object_t object = NULL;
1474 if (bp->b_ioflags & BIO_ERROR) {
1476 "swap_pager: I/O error - %s failed; blkno %ld,"
1477 "size %ld, error %d\n",
1478 ((bp->b_iocmd == BIO_READ) ? "pagein" : "pageout"),
1486 * remove the mapping for kernel virtual
1488 pmap_qremove((vm_offset_t)bp->b_data, bp->b_npages);
1491 object = bp->b_pages[0]->object;
1492 VM_OBJECT_LOCK(object);
1494 vm_page_lock_queues();
1496 * cleanup pages. If an error occurs writing to swap, we are in
1497 * very serious trouble. If it happens to be a disk error, though,
1498 * we may be able to recover by reassigning the swap later on. So
1499 * in this case we remove the m->swapblk assignment for the page
1500 * but do not free it in the rlist. The errornous block(s) are thus
1501 * never reallocated as swap. Redirty the page and continue.
1503 for (i = 0; i < bp->b_npages; ++i) {
1504 vm_page_t m = bp->b_pages[i];
1506 m->oflags &= ~VPO_SWAPINPROG;
1508 if (bp->b_ioflags & BIO_ERROR) {
1510 * If an error occurs I'd love to throw the swapblk
1511 * away without freeing it back to swapspace, so it
1512 * can never be used again. But I can't from an
1515 if (bp->b_iocmd == BIO_READ) {
1517 * When reading, reqpage needs to stay
1518 * locked for the parent, but all other
1519 * pages can be freed. We still want to
1520 * wakeup the parent waiting on the page,
1521 * though. ( also: pg_reqpage can be -1 and
1522 * not match anything ).
1524 * We have to wake specifically requested pages
1525 * up too because we cleared VPO_SWAPINPROG and
1526 * someone may be waiting for that.
1528 * NOTE: for reads, m->dirty will probably
1529 * be overridden by the original caller of
1530 * getpages so don't play cute tricks here.
1533 if (i != bp->b_pager.pg_reqpage)
1538 * If i == bp->b_pager.pg_reqpage, do not wake
1539 * the page up. The caller needs to.
1543 * If a write error occurs, reactivate page
1544 * so it doesn't clog the inactive list,
1545 * then finish the I/O.
1548 vm_page_activate(m);
1549 vm_page_io_finish(m);
1551 } else if (bp->b_iocmd == BIO_READ) {
1553 * NOTE: for reads, m->dirty will probably be
1554 * overridden by the original caller of getpages so
1555 * we cannot set them in order to free the underlying
1556 * swap in a low-swap situation. I don't think we'd
1557 * want to do that anyway, but it was an optimization
1558 * that existed in the old swapper for a time before
1559 * it got ripped out due to precisely this problem.
1561 * If not the requested page then deactivate it.
1563 * Note that the requested page, reqpage, is left
1564 * busied, but we still have to wake it up. The
1565 * other pages are released (unbusied) by
1568 KASSERT(!pmap_page_is_mapped(m),
1569 ("swp_pager_async_iodone: page %p is mapped", m));
1570 m->valid = VM_PAGE_BITS_ALL;
1571 KASSERT(m->dirty == 0,
1572 ("swp_pager_async_iodone: page %p is dirty", m));
1575 * We have to wake specifically requested pages
1576 * up too because we cleared VPO_SWAPINPROG and
1577 * could be waiting for it in getpages. However,
1578 * be sure to not unbusy getpages specifically
1579 * requested page - getpages expects it to be
1582 if (i != bp->b_pager.pg_reqpage) {
1583 vm_page_deactivate(m);
1590 * For write success, clear the dirty
1591 * status, then finish the I/O ( which decrements the
1592 * busy count and possibly wakes waiter's up ).
1594 KASSERT((m->flags & PG_WRITEABLE) == 0,
1595 ("swp_pager_async_iodone: page %p is not write"
1598 vm_page_io_finish(m);
1599 if (vm_page_count_severe())
1600 vm_page_try_to_cache(m);
1603 vm_page_unlock_queues();
1606 * adjust pip. NOTE: the original parent may still have its own
1607 * pip refs on the object.
1609 if (object != NULL) {
1610 vm_object_pip_wakeupn(object, bp->b_npages);
1611 VM_OBJECT_UNLOCK(object);
1615 * swapdev_strategy() manually sets b_vp and b_bufobj before calling
1616 * bstrategy(). Set them back to NULL now we're done with it, or we'll
1617 * trigger a KASSERT in relpbuf().
1621 bp->b_bufobj = NULL;
1624 * release the physical I/O buffer
1628 ((bp->b_iocmd == BIO_READ) ? &nsw_rcount :
1629 ((bp->b_flags & B_ASYNC) ?
1638 * swap_pager_isswapped:
1640 * Return 1 if at least one page in the given object is paged
1641 * out to the given swap device.
1643 * This routine may not block.
1646 swap_pager_isswapped(vm_object_t object, struct swdevt *sp)
1652 VM_OBJECT_LOCK_ASSERT(object, MA_OWNED);
1653 if (object->type != OBJT_SWAP)
1656 mtx_lock(&swhash_mtx);
1657 for (bcount = 0; bcount < object->un_pager.swp.swp_bcount; bcount++) {
1658 struct swblock *swap;
1660 if ((swap = *swp_pager_hash(object, index)) != NULL) {
1661 for (i = 0; i < SWAP_META_PAGES; ++i) {
1662 if (swp_pager_isondev(swap->swb_pages[i], sp)) {
1663 mtx_unlock(&swhash_mtx);
1668 index += SWAP_META_PAGES;
1669 if (index > 0x20000000)
1670 panic("swap_pager_isswapped: failed to locate all swap meta blocks");
1672 mtx_unlock(&swhash_mtx);
1677 * SWP_PAGER_FORCE_PAGEIN() - force a swap block to be paged in
1679 * This routine dissociates the page at the given index within a
1680 * swap block from its backing store, paging it in if necessary.
1681 * If the page is paged in, it is placed in the inactive queue,
1682 * since it had its backing store ripped out from under it.
1683 * We also attempt to swap in all other pages in the swap block,
1684 * we only guarantee that the one at the specified index is
1687 * XXX - The code to page the whole block in doesn't work, so we
1688 * revert to the one-by-one behavior for now. Sigh.
1691 swp_pager_force_pagein(vm_object_t object, vm_pindex_t pindex)
1695 vm_object_pip_add(object, 1);
1696 m = vm_page_grab(object, pindex, VM_ALLOC_NORMAL|VM_ALLOC_RETRY);
1697 if (m->valid == VM_PAGE_BITS_ALL) {
1698 vm_object_pip_subtract(object, 1);
1699 vm_page_lock_queues();
1700 vm_page_activate(m);
1702 vm_page_unlock_queues();
1704 vm_pager_page_unswapped(m);
1708 if (swap_pager_getpages(object, &m, 1, 0) != VM_PAGER_OK)
1709 panic("swap_pager_force_pagein: read from swap failed");/*XXX*/
1710 vm_object_pip_subtract(object, 1);
1711 vm_page_lock_queues();
1713 vm_page_dontneed(m);
1714 vm_page_unlock_queues();
1716 vm_pager_page_unswapped(m);
1720 * swap_pager_swapoff:
1722 * Page in all of the pages that have been paged out to the
1723 * given device. The corresponding blocks in the bitmap must be
1724 * marked as allocated and the device must be flagged SW_CLOSING.
1725 * There may be no processes swapped out to the device.
1727 * This routine may block.
1730 swap_pager_swapoff(struct swdevt *sp)
1732 struct swblock *swap;
1739 mtx_lock(&swhash_mtx);
1740 for (i = 0; i <= swhash_mask; i++) { /* '<=' is correct here */
1742 for (swap = swhash[i]; swap != NULL; swap = swap->swb_hnext) {
1743 vm_object_t object = swap->swb_object;
1744 vm_pindex_t pindex = swap->swb_index;
1745 for (j = 0; j < SWAP_META_PAGES; ++j) {
1746 if (swp_pager_isondev(swap->swb_pages[j], sp)) {
1747 /* avoid deadlock */
1748 if (!VM_OBJECT_TRYLOCK(object)) {
1751 mtx_unlock(&swhash_mtx);
1752 swp_pager_force_pagein(object,
1754 VM_OBJECT_UNLOCK(object);
1755 mtx_lock(&swhash_mtx);
1762 mtx_unlock(&swhash_mtx);
1765 * Objects may be locked or paging to the device being
1766 * removed, so we will miss their pages and need to
1767 * make another pass. We have marked this device as
1768 * SW_CLOSING, so the activity should finish soon.
1771 if (retries > 100) {
1772 panic("swapoff: failed to locate %d swap blocks",
1775 pause("swpoff", hz / 20);
1780 /************************************************************************
1782 ************************************************************************
1784 * These routines manipulate the swap metadata stored in the
1785 * OBJT_SWAP object. All swp_*() routines must be called at
1786 * splvm() because swap can be freed up by the low level vm_page
1787 * code which might be called from interrupts beyond what splbio() covers.
1789 * Swap metadata is implemented with a global hash and not directly
1790 * linked into the object. Instead the object simply contains
1791 * appropriate tracking counters.
1795 * SWP_PAGER_META_BUILD() - add swap block to swap meta data for object
1797 * We first convert the object to a swap object if it is a default
1800 * The specified swapblk is added to the object's swap metadata. If
1801 * the swapblk is not valid, it is freed instead. Any previously
1802 * assigned swapblk is freed.
1804 * This routine must be called at splvm(), except when used to convert
1805 * an OBJT_DEFAULT object into an OBJT_SWAP object.
1808 swp_pager_meta_build(vm_object_t object, vm_pindex_t pindex, daddr_t swapblk)
1810 struct swblock *swap;
1811 struct swblock **pswap;
1814 VM_OBJECT_LOCK_ASSERT(object, MA_OWNED);
1816 * Convert default object to swap object if necessary
1818 if (object->type != OBJT_SWAP) {
1819 object->type = OBJT_SWAP;
1820 object->un_pager.swp.swp_bcount = 0;
1822 if (object->handle != NULL) {
1823 mtx_lock(&sw_alloc_mtx);
1825 NOBJLIST(object->handle),
1829 mtx_unlock(&sw_alloc_mtx);
1834 * Locate hash entry. If not found create, but if we aren't adding
1835 * anything just return. If we run out of space in the map we wait
1836 * and, since the hash table may have changed, retry.
1839 mtx_lock(&swhash_mtx);
1840 pswap = swp_pager_hash(object, pindex);
1842 if ((swap = *pswap) == NULL) {
1845 if (swapblk == SWAPBLK_NONE)
1848 swap = *pswap = uma_zalloc(swap_zone, M_NOWAIT);
1850 mtx_unlock(&swhash_mtx);
1851 VM_OBJECT_UNLOCK(object);
1852 if (uma_zone_exhausted(swap_zone)) {
1853 printf("swap zone exhausted, increase kern.maxswzone\n");
1854 vm_pageout_oom(VM_OOM_SWAPZ);
1855 pause("swzonex", 10);
1858 VM_OBJECT_LOCK(object);
1862 swap->swb_hnext = NULL;
1863 swap->swb_object = object;
1864 swap->swb_index = pindex & ~(vm_pindex_t)SWAP_META_MASK;
1865 swap->swb_count = 0;
1867 ++object->un_pager.swp.swp_bcount;
1869 for (i = 0; i < SWAP_META_PAGES; ++i)
1870 swap->swb_pages[i] = SWAPBLK_NONE;
1874 * Delete prior contents of metadata
1876 idx = pindex & SWAP_META_MASK;
1878 if (swap->swb_pages[idx] != SWAPBLK_NONE) {
1879 swp_pager_freeswapspace(swap->swb_pages[idx], 1);
1884 * Enter block into metadata
1886 swap->swb_pages[idx] = swapblk;
1887 if (swapblk != SWAPBLK_NONE)
1890 mtx_unlock(&swhash_mtx);
1894 * SWP_PAGER_META_FREE() - free a range of blocks in the object's swap metadata
1896 * The requested range of blocks is freed, with any associated swap
1897 * returned to the swap bitmap.
1899 * This routine will free swap metadata structures as they are cleaned
1900 * out. This routine does *NOT* operate on swap metadata associated
1901 * with resident pages.
1903 * This routine must be called at splvm()
1906 swp_pager_meta_free(vm_object_t object, vm_pindex_t index, daddr_t count)
1909 VM_OBJECT_LOCK_ASSERT(object, MA_OWNED);
1910 if (object->type != OBJT_SWAP)
1914 struct swblock **pswap;
1915 struct swblock *swap;
1917 mtx_lock(&swhash_mtx);
1918 pswap = swp_pager_hash(object, index);
1920 if ((swap = *pswap) != NULL) {
1921 daddr_t v = swap->swb_pages[index & SWAP_META_MASK];
1923 if (v != SWAPBLK_NONE) {
1924 swp_pager_freeswapspace(v, 1);
1925 swap->swb_pages[index & SWAP_META_MASK] =
1927 if (--swap->swb_count == 0) {
1928 *pswap = swap->swb_hnext;
1929 uma_zfree(swap_zone, swap);
1930 --object->un_pager.swp.swp_bcount;
1936 int n = SWAP_META_PAGES - (index & SWAP_META_MASK);
1940 mtx_unlock(&swhash_mtx);
1945 * SWP_PAGER_META_FREE_ALL() - destroy all swap metadata associated with object
1947 * This routine locates and destroys all swap metadata associated with
1950 * This routine must be called at splvm()
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;
1985 if (index > 0x20000000)
1986 panic("swp_pager_meta_free_all: failed to locate all swap meta blocks");
1991 * SWP_PAGER_METACTL() - misc control of swap and vm_page_t meta data.
1993 * This routine is capable of looking up, popping, or freeing
1994 * swapblk assignments in the swap meta data or in the vm_page_t.
1995 * The routine typically returns the swapblk being looked-up, or popped,
1996 * or SWAPBLK_NONE if the block was freed, or SWAPBLK_NONE if the block
1997 * was invalid. This routine will automatically free any invalid
1998 * meta-data swapblks.
2000 * It is not possible to store invalid swapblks in the swap meta data
2001 * (other then a literal 'SWAPBLK_NONE'), so we don't bother checking.
2003 * When acting on a busy resident page and paging is in progress, we
2004 * have to wait until paging is complete but otherwise can act on the
2007 * This routine must be called at splvm().
2009 * SWM_FREE remove and free swap block from metadata
2010 * SWM_POP remove from meta data but do not free.. pop it out
2013 swp_pager_meta_ctl(vm_object_t object, vm_pindex_t pindex, int flags)
2015 struct swblock **pswap;
2016 struct swblock *swap;
2020 VM_OBJECT_LOCK_ASSERT(object, MA_OWNED);
2022 * The meta data only exists of the object is OBJT_SWAP
2023 * and even then might not be allocated yet.
2025 if (object->type != OBJT_SWAP)
2026 return (SWAPBLK_NONE);
2029 mtx_lock(&swhash_mtx);
2030 pswap = swp_pager_hash(object, pindex);
2032 if ((swap = *pswap) != NULL) {
2033 idx = pindex & SWAP_META_MASK;
2034 r1 = swap->swb_pages[idx];
2036 if (r1 != SWAPBLK_NONE) {
2037 if (flags & SWM_FREE) {
2038 swp_pager_freeswapspace(r1, 1);
2041 if (flags & (SWM_FREE|SWM_POP)) {
2042 swap->swb_pages[idx] = SWAPBLK_NONE;
2043 if (--swap->swb_count == 0) {
2044 *pswap = swap->swb_hnext;
2045 uma_zfree(swap_zone, swap);
2046 --object->un_pager.swp.swp_bcount;
2051 mtx_unlock(&swhash_mtx);
2056 * System call swapon(name) enables swapping on device name,
2057 * which must be in the swdevsw. Return EBUSY
2058 * if already swapping on this device.
2060 #ifndef _SYS_SYSPROTO_H_
2061 struct swapon_args {
2071 swapon(struct thread *td, struct swapon_args *uap)
2075 struct nameidata nd;
2078 error = priv_check(td, PRIV_SWAPON);
2083 while (swdev_syscall_active)
2084 tsleep(&swdev_syscall_active, PUSER - 1, "swpon", 0);
2085 swdev_syscall_active = 1;
2088 * Swap metadata may not fit in the KVM if we have physical
2091 if (swap_zone == NULL) {
2096 NDINIT(&nd, LOOKUP, ISOPEN | FOLLOW | AUDITVNODE1, UIO_USERSPACE,
2102 NDFREE(&nd, NDF_ONLY_PNBUF);
2105 if (vn_isdisk(vp, &error)) {
2106 error = swapongeom(td, vp);
2107 } else if (vp->v_type == VREG &&
2108 (vp->v_mount->mnt_vfc->vfc_flags & VFCF_NETWORK) != 0 &&
2109 (error = VOP_GETATTR(vp, &attr, td->td_ucred)) == 0) {
2111 * Allow direct swapping to NFS regular files in the same
2112 * way that nfs_mountroot() sets up diskless swapping.
2114 error = swaponvp(td, vp, attr.va_size / DEV_BSIZE);
2120 swdev_syscall_active = 0;
2121 wakeup_one(&swdev_syscall_active);
2127 swaponsomething(struct vnode *vp, void *id, u_long nblks, sw_strategy_t *strategy, sw_close_t *close, dev_t dev)
2129 struct swdevt *sp, *tsp;
2134 * If we go beyond this, we get overflows in the radix
2137 mblocks = 0x40000000 / BLIST_META_RADIX;
2138 if (nblks > mblocks) {
2139 printf("WARNING: reducing size to maximum of %lu blocks per swap unit\n",
2144 * nblks is in DEV_BSIZE'd chunks, convert to PAGE_SIZE'd chunks.
2145 * First chop nblks off to page-align it, then convert.
2147 * sw->sw_nblks is in page-sized chunks now too.
2149 nblks &= ~(ctodb(1) - 1);
2150 nblks = dbtoc(nblks);
2152 sp = malloc(sizeof *sp, M_VMPGDATA, M_WAITOK | M_ZERO);
2157 sp->sw_nblks = nblks;
2159 sp->sw_strategy = strategy;
2160 sp->sw_close = close;
2162 sp->sw_blist = blist_create(nblks, M_WAITOK);
2164 * Do not free the first two block in order to avoid overwriting
2165 * any bsd label at the front of the partition
2167 blist_free(sp->sw_blist, 2, nblks - 2);
2170 mtx_lock(&sw_dev_mtx);
2171 TAILQ_FOREACH(tsp, &swtailq, sw_list) {
2172 if (tsp->sw_end >= dvbase) {
2174 * We put one uncovered page between the devices
2175 * in order to definitively prevent any cross-device
2178 dvbase = tsp->sw_end + 1;
2181 sp->sw_first = dvbase;
2182 sp->sw_end = dvbase + nblks;
2183 TAILQ_INSERT_TAIL(&swtailq, sp, sw_list);
2185 swap_pager_avail += nblks;
2186 swap_total += (vm_ooffset_t)nblks * PAGE_SIZE;
2188 mtx_unlock(&sw_dev_mtx);
2192 * SYSCALL: swapoff(devname)
2194 * Disable swapping on the given device.
2196 * XXX: Badly designed system call: it should use a device index
2197 * rather than filename as specification. We keep sw_vp around
2198 * only to make this work.
2200 #ifndef _SYS_SYSPROTO_H_
2201 struct swapoff_args {
2211 swapoff(struct thread *td, struct swapoff_args *uap)
2214 struct nameidata nd;
2218 error = priv_check(td, PRIV_SWAPOFF);
2223 while (swdev_syscall_active)
2224 tsleep(&swdev_syscall_active, PUSER - 1, "swpoff", 0);
2225 swdev_syscall_active = 1;
2227 NDINIT(&nd, LOOKUP, FOLLOW | AUDITVNODE1, UIO_USERSPACE, uap->name,
2232 NDFREE(&nd, NDF_ONLY_PNBUF);
2235 mtx_lock(&sw_dev_mtx);
2236 TAILQ_FOREACH(sp, &swtailq, sw_list) {
2237 if (sp->sw_vp == vp)
2240 mtx_unlock(&sw_dev_mtx);
2245 error = swapoff_one(sp, td->td_ucred);
2247 swdev_syscall_active = 0;
2248 wakeup_one(&swdev_syscall_active);
2254 swapoff_one(struct swdevt *sp, struct ucred *cred)
2256 u_long nblks, dvbase;
2261 mtx_assert(&Giant, MA_OWNED);
2263 (void) vn_lock(sp->sw_vp, LK_EXCLUSIVE | LK_RETRY);
2264 error = mac_system_check_swapoff(cred, sp->sw_vp);
2265 (void) VOP_UNLOCK(sp->sw_vp, 0);
2269 nblks = sp->sw_nblks;
2272 * We can turn off this swap device safely only if the
2273 * available virtual memory in the system will fit the amount
2274 * of data we will have to page back in, plus an epsilon so
2275 * the system doesn't become critically low on swap space.
2277 if (cnt.v_free_count + cnt.v_cache_count + swap_pager_avail <
2278 nblks + nswap_lowat) {
2283 * Prevent further allocations on this device.
2285 mtx_lock(&sw_dev_mtx);
2286 sp->sw_flags |= SW_CLOSING;
2287 for (dvbase = 0; dvbase < sp->sw_end; dvbase += dmmax) {
2288 swap_pager_avail -= blist_fill(sp->sw_blist,
2291 swap_total -= (vm_ooffset_t)nblks * PAGE_SIZE;
2292 mtx_unlock(&sw_dev_mtx);
2295 * Page in the contents of the device and close it.
2297 swap_pager_swapoff(sp);
2299 sp->sw_close(curthread, sp);
2301 mtx_lock(&sw_dev_mtx);
2302 TAILQ_REMOVE(&swtailq, sp, sw_list);
2304 if (nswapdev == 0) {
2305 swap_pager_full = 2;
2306 swap_pager_almost_full = 1;
2310 mtx_unlock(&sw_dev_mtx);
2311 blist_destroy(sp->sw_blist);
2312 free(sp, M_VMPGDATA);
2319 struct swdevt *sp, *spt;
2320 const char *devname;
2324 while (swdev_syscall_active)
2325 tsleep(&swdev_syscall_active, PUSER - 1, "swpoff", 0);
2326 swdev_syscall_active = 1;
2328 mtx_lock(&sw_dev_mtx);
2329 TAILQ_FOREACH_SAFE(sp, &swtailq, sw_list, spt) {
2330 mtx_unlock(&sw_dev_mtx);
2331 if (vn_isdisk(sp->sw_vp, NULL))
2332 devname = sp->sw_vp->v_rdev->si_name;
2335 error = swapoff_one(sp, thread0.td_ucred);
2337 printf("Cannot remove swap device %s (error=%d), "
2338 "skipping.\n", devname, error);
2339 } else if (bootverbose) {
2340 printf("Swap device %s removed.\n", devname);
2342 mtx_lock(&sw_dev_mtx);
2344 mtx_unlock(&sw_dev_mtx);
2346 swdev_syscall_active = 0;
2347 wakeup_one(&swdev_syscall_active);
2352 swap_pager_status(int *total, int *used)
2358 mtx_lock(&sw_dev_mtx);
2359 TAILQ_FOREACH(sp, &swtailq, sw_list) {
2360 *total += sp->sw_nblks;
2361 *used += sp->sw_used;
2363 mtx_unlock(&sw_dev_mtx);
2367 sysctl_vm_swap_info(SYSCTL_HANDLER_ARGS)
2369 int *name = (int *)arg1;
2374 if (arg2 != 1) /* name length */
2378 mtx_lock(&sw_dev_mtx);
2379 TAILQ_FOREACH(sp, &swtailq, sw_list) {
2381 mtx_unlock(&sw_dev_mtx);
2382 xs.xsw_version = XSWDEV_VERSION;
2383 xs.xsw_dev = sp->sw_dev;
2384 xs.xsw_flags = sp->sw_flags;
2385 xs.xsw_nblks = sp->sw_nblks;
2386 xs.xsw_used = sp->sw_used;
2388 error = SYSCTL_OUT(req, &xs, sizeof(xs));
2393 mtx_unlock(&sw_dev_mtx);
2397 SYSCTL_INT(_vm, OID_AUTO, nswapdev, CTLFLAG_RD, &nswapdev, 0,
2398 "Number of swap devices");
2399 SYSCTL_NODE(_vm, OID_AUTO, swap_info, CTLFLAG_RD, sysctl_vm_swap_info,
2400 "Swap statistics by device");
2403 * vmspace_swap_count() - count the approximate swap usage in pages for a
2406 * The map must be locked.
2408 * Swap usage is determined by taking the proportional swap used by
2409 * VM objects backing the VM map. To make up for fractional losses,
2410 * if the VM object has any swap use at all the associated map entries
2411 * count for at least 1 swap page.
2414 vmspace_swap_count(struct vmspace *vmspace)
2416 vm_map_t map = &vmspace->vm_map;
2420 for (cur = map->header.next; cur != &map->header; cur = cur->next) {
2423 if ((cur->eflags & MAP_ENTRY_IS_SUB_MAP) == 0 &&
2424 (object = cur->object.vm_object) != NULL) {
2425 VM_OBJECT_LOCK(object);
2426 if (object->type == OBJT_SWAP &&
2427 object->un_pager.swp.swp_bcount != 0) {
2428 int n = (cur->end - cur->start) / PAGE_SIZE;
2430 count += object->un_pager.swp.swp_bcount *
2431 SWAP_META_PAGES * n / object->size + 1;
2433 VM_OBJECT_UNLOCK(object);
2442 * Swapping onto disk devices.
2446 static g_orphan_t swapgeom_orphan;
2448 static struct g_class g_swap_class = {
2450 .version = G_VERSION,
2451 .orphan = swapgeom_orphan,
2454 DECLARE_GEOM_CLASS(g_swap_class, g_class);
2458 swapgeom_done(struct bio *bp2)
2462 bp = bp2->bio_caller2;
2463 bp->b_ioflags = bp2->bio_flags;
2465 bp->b_ioflags |= BIO_ERROR;
2466 bp->b_resid = bp->b_bcount - bp2->bio_completed;
2467 bp->b_error = bp2->bio_error;
2473 swapgeom_strategy(struct buf *bp, struct swdevt *sp)
2476 struct g_consumer *cp;
2480 bp->b_error = ENXIO;
2481 bp->b_ioflags |= BIO_ERROR;
2485 if (bp->b_iocmd == BIO_WRITE)
2488 bio = g_alloc_bio();
2490 bp->b_error = ENOMEM;
2491 bp->b_ioflags |= BIO_ERROR;
2496 bio->bio_caller2 = bp;
2497 bio->bio_cmd = bp->b_iocmd;
2498 bio->bio_data = bp->b_data;
2499 bio->bio_offset = (bp->b_blkno - sp->sw_first) * PAGE_SIZE;
2500 bio->bio_length = bp->b_bcount;
2501 bio->bio_done = swapgeom_done;
2502 g_io_request(bio, cp);
2507 swapgeom_orphan(struct g_consumer *cp)
2511 mtx_lock(&sw_dev_mtx);
2512 TAILQ_FOREACH(sp, &swtailq, sw_list)
2513 if (sp->sw_id == cp)
2515 mtx_unlock(&sw_dev_mtx);
2519 swapgeom_close_ev(void *arg, int flags)
2521 struct g_consumer *cp;
2524 g_access(cp, -1, -1, 0);
2526 g_destroy_consumer(cp);
2530 swapgeom_close(struct thread *td, struct swdevt *sw)
2533 /* XXX: direct call when Giant untangled */
2534 g_waitfor_event(swapgeom_close_ev, sw->sw_id, M_WAITOK, NULL);
2545 swapongeom_ev(void *arg, int flags)
2548 struct g_provider *pp;
2549 struct g_consumer *cp;
2550 static struct g_geom *gp;
2557 pp = g_dev_getprovider(swh->dev);
2559 swh->error = ENODEV;
2562 mtx_lock(&sw_dev_mtx);
2563 TAILQ_FOREACH(sp, &swtailq, sw_list) {
2565 if (cp != NULL && cp->provider == pp) {
2566 mtx_unlock(&sw_dev_mtx);
2571 mtx_unlock(&sw_dev_mtx);
2573 gp = g_new_geomf(&g_swap_class, "swap", NULL);
2574 cp = g_new_consumer(gp);
2577 * XXX: Everytime you think you can improve the margin for
2578 * footshooting, somebody depends on the ability to do so:
2579 * savecore(8) wants to write to our swapdev so we cannot
2580 * set an exclusive count :-(
2582 error = g_access(cp, 1, 1, 0);
2585 g_destroy_consumer(cp);
2589 nblks = pp->mediasize / DEV_BSIZE;
2590 swaponsomething(swh->vp, cp, nblks, swapgeom_strategy,
2591 swapgeom_close, dev2udev(swh->dev));
2597 swapongeom(struct thread *td, struct vnode *vp)
2602 vn_lock(vp, LK_EXCLUSIVE | LK_RETRY);
2604 swh.dev = vp->v_rdev;
2607 /* XXX: direct call when Giant untangled */
2608 error = g_waitfor_event(swapongeom_ev, &swh, M_WAITOK, NULL);
2618 * This is used mainly for network filesystem (read: probably only tested
2619 * with NFS) swapfiles.
2624 swapdev_strategy(struct buf *bp, struct swdevt *sp)
2628 bp->b_blkno = ctodb(bp->b_blkno - sp->sw_first);
2632 if (bp->b_iocmd == BIO_WRITE) {
2634 bufobj_wdrop(bp->b_bufobj);
2635 bufobj_wref(&vp2->v_bufobj);
2637 if (bp->b_bufobj != &vp2->v_bufobj)
2638 bp->b_bufobj = &vp2->v_bufobj;
2640 bp->b_iooffset = dbtob(bp->b_blkno);
2646 swapdev_close(struct thread *td, struct swdevt *sp)
2649 VOP_CLOSE(sp->sw_vp, FREAD | FWRITE, td->td_ucred, td);
2655 swaponvp(struct thread *td, struct vnode *vp, u_long nblks)
2662 mtx_lock(&sw_dev_mtx);
2663 TAILQ_FOREACH(sp, &swtailq, sw_list) {
2664 if (sp->sw_id == vp) {
2665 mtx_unlock(&sw_dev_mtx);
2669 mtx_unlock(&sw_dev_mtx);
2671 (void) vn_lock(vp, LK_EXCLUSIVE | LK_RETRY);
2673 error = mac_system_check_swapon(td->td_ucred, vp);
2676 error = VOP_OPEN(vp, FREAD | FWRITE, td->td_ucred, td, NULL);
2677 (void) VOP_UNLOCK(vp, 0);
2681 swaponsomething(vp, vp, nblks, swapdev_strategy, swapdev_close,