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, 16
118 * or 32 pages per allocation.
119 * The 32-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 * The swblock structure maps an object and a small, fixed-size range
131 * of page indices to disk addresses within a swap area.
132 * The collection of these mappings is implemented as a hash table.
133 * Unused disk addresses within a swap area are allocated and managed
136 #define SWCORRECT(n) (sizeof(void *) * (n) / sizeof(daddr_t))
137 #define SWAP_META_PAGES (SWB_NPAGES * 2)
138 #define SWAP_META_MASK (SWAP_META_PAGES - 1)
141 struct swblock *swb_hnext;
142 vm_object_t swb_object;
143 vm_pindex_t swb_index;
145 daddr_t swb_pages[SWAP_META_PAGES];
148 static MALLOC_DEFINE(M_VMPGDATA, "vm_pgdata", "swap pager private data");
149 static struct mtx sw_dev_mtx;
150 static TAILQ_HEAD(, swdevt) swtailq = TAILQ_HEAD_INITIALIZER(swtailq);
151 static struct swdevt *swdevhd; /* Allocate from here next */
152 static int nswapdev; /* Number of swap devices */
153 int swap_pager_avail;
154 static int swdev_syscall_active = 0; /* serialize swap(on|off) */
156 static vm_ooffset_t swap_total;
157 SYSCTL_QUAD(_vm, OID_AUTO, swap_total, CTLFLAG_RD, &swap_total, 0,
158 "Total amount of available swap storage.");
159 static vm_ooffset_t swap_reserved;
160 SYSCTL_QUAD(_vm, OID_AUTO, swap_reserved, CTLFLAG_RD, &swap_reserved, 0,
161 "Amount of swap storage needed to back all allocated anonymous memory.");
162 static int overcommit = 0;
163 SYSCTL_INT(_vm, OID_AUTO, overcommit, CTLFLAG_RW, &overcommit, 0,
164 "Configure virtual memory overcommit behavior. See tuning(7) "
167 /* bits from overcommit */
168 #define SWAP_RESERVE_FORCE_ON (1 << 0)
169 #define SWAP_RESERVE_RLIMIT_ON (1 << 1)
170 #define SWAP_RESERVE_ALLOW_NONWIRED (1 << 2)
173 swap_reserve(vm_ooffset_t incr)
176 return (swap_reserve_by_cred(incr, curthread->td_ucred));
180 swap_reserve_by_cred(vm_ooffset_t incr, struct ucred *cred)
185 static struct timeval lastfail;
188 uip = cred->cr_ruidinfo;
190 if (incr & PAGE_MASK)
191 panic("swap_reserve: & PAGE_MASK");
195 error = racct_add(curproc, RACCT_SWAP, incr);
196 PROC_UNLOCK(curproc);
202 mtx_lock(&sw_dev_mtx);
203 r = swap_reserved + incr;
204 if (overcommit & SWAP_RESERVE_ALLOW_NONWIRED) {
205 s = cnt.v_page_count - cnt.v_free_reserved - cnt.v_wire_count;
210 if ((overcommit & SWAP_RESERVE_FORCE_ON) == 0 || r <= s ||
211 (error = priv_check(curthread, PRIV_VM_SWAP_NOQUOTA)) == 0) {
215 mtx_unlock(&sw_dev_mtx);
219 UIDINFO_VMSIZE_LOCK(uip);
220 if ((overcommit & SWAP_RESERVE_RLIMIT_ON) != 0 &&
221 uip->ui_vmsize + incr > lim_cur(curproc, RLIMIT_SWAP) &&
222 priv_check(curthread, PRIV_VM_SWAP_NORLIMIT))
225 uip->ui_vmsize += incr;
226 UIDINFO_VMSIZE_UNLOCK(uip);
227 PROC_UNLOCK(curproc);
229 mtx_lock(&sw_dev_mtx);
230 swap_reserved -= incr;
231 mtx_unlock(&sw_dev_mtx);
234 if (!res && ppsratecheck(&lastfail, &curfail, 1)) {
235 printf("uid %d, pid %d: swap reservation for %jd bytes failed\n",
236 uip->ui_uid, curproc->p_pid, incr);
242 racct_sub(curproc, RACCT_SWAP, incr);
243 PROC_UNLOCK(curproc);
251 swap_reserve_force(vm_ooffset_t incr)
255 mtx_lock(&sw_dev_mtx);
256 swap_reserved += incr;
257 mtx_unlock(&sw_dev_mtx);
261 racct_add_force(curproc, RACCT_SWAP, incr);
262 PROC_UNLOCK(curproc);
265 uip = curthread->td_ucred->cr_ruidinfo;
267 UIDINFO_VMSIZE_LOCK(uip);
268 uip->ui_vmsize += incr;
269 UIDINFO_VMSIZE_UNLOCK(uip);
270 PROC_UNLOCK(curproc);
274 swap_release(vm_ooffset_t decr)
279 cred = curthread->td_ucred;
280 swap_release_by_cred(decr, cred);
281 PROC_UNLOCK(curproc);
285 swap_release_by_cred(vm_ooffset_t decr, struct ucred *cred)
289 uip = cred->cr_ruidinfo;
291 if (decr & PAGE_MASK)
292 panic("swap_release: & PAGE_MASK");
294 mtx_lock(&sw_dev_mtx);
295 if (swap_reserved < decr)
296 panic("swap_reserved < decr");
297 swap_reserved -= decr;
298 mtx_unlock(&sw_dev_mtx);
300 UIDINFO_VMSIZE_LOCK(uip);
301 if (uip->ui_vmsize < decr)
302 printf("negative vmsize for uid = %d\n", uip->ui_uid);
303 uip->ui_vmsize -= decr;
304 UIDINFO_VMSIZE_UNLOCK(uip);
306 racct_sub_cred(cred, RACCT_SWAP, decr);
309 static void swapdev_strategy(struct buf *, struct swdevt *sw);
311 #define SWM_FREE 0x02 /* free, period */
312 #define SWM_POP 0x04 /* pop out */
314 int swap_pager_full = 2; /* swap space exhaustion (task killing) */
315 static int swap_pager_almost_full = 1; /* swap space exhaustion (w/hysteresis)*/
316 static int nsw_rcount; /* free read buffers */
317 static int nsw_wcount_sync; /* limit write buffers / synchronous */
318 static int nsw_wcount_async; /* limit write buffers / asynchronous */
319 static int nsw_wcount_async_max;/* assigned maximum */
320 static int nsw_cluster_max; /* maximum VOP I/O allowed */
322 static struct swblock **swhash;
323 static int swhash_mask;
324 static struct mtx swhash_mtx;
326 static int swap_async_max = 4; /* maximum in-progress async I/O's */
327 static struct sx sw_alloc_sx;
330 SYSCTL_INT(_vm, OID_AUTO, swap_async_max,
331 CTLFLAG_RW, &swap_async_max, 0, "Maximum running async swap ops");
334 * "named" and "unnamed" anon region objects. Try to reduce the overhead
335 * of searching a named list by hashing it just a little.
340 #define NOBJLIST(handle) \
341 (&swap_pager_object_list[((int)(intptr_t)handle >> 4) & (NOBJLISTS-1)])
343 static struct mtx sw_alloc_mtx; /* protect list manipulation */
344 static struct pagerlst swap_pager_object_list[NOBJLISTS];
345 static uma_zone_t swap_zone;
346 static struct vm_object swap_zone_obj;
349 * pagerops for OBJT_SWAP - "swap pager". Some ops are also global procedure
350 * calls hooked from other parts of the VM system and do not appear here.
351 * (see vm/swap_pager.h).
354 swap_pager_alloc(void *handle, vm_ooffset_t size,
355 vm_prot_t prot, vm_ooffset_t offset, struct ucred *);
356 static void swap_pager_dealloc(vm_object_t object);
357 static int swap_pager_getpages(vm_object_t, vm_page_t *, int, int);
358 static void swap_pager_putpages(vm_object_t, vm_page_t *, int, boolean_t, int *);
360 swap_pager_haspage(vm_object_t object, vm_pindex_t pindex, int *before, int *after);
361 static void swap_pager_init(void);
362 static void swap_pager_unswapped(vm_page_t);
363 static void swap_pager_swapoff(struct swdevt *sp);
365 struct pagerops swappagerops = {
366 .pgo_init = swap_pager_init, /* early system initialization of pager */
367 .pgo_alloc = swap_pager_alloc, /* allocate an OBJT_SWAP object */
368 .pgo_dealloc = swap_pager_dealloc, /* deallocate an OBJT_SWAP object */
369 .pgo_getpages = swap_pager_getpages, /* pagein */
370 .pgo_putpages = swap_pager_putpages, /* pageout */
371 .pgo_haspage = swap_pager_haspage, /* get backing store status for page */
372 .pgo_pageunswapped = swap_pager_unswapped, /* remove swap related to page */
376 * dmmax is in page-sized chunks with the new swap system. It was
377 * dev-bsized chunks in the old. dmmax is always a power of 2.
379 * swap_*() routines are externally accessible. swp_*() routines are
383 static int nswap_lowat = 128; /* in pages, swap_pager_almost_full warn */
384 static int nswap_hiwat = 512; /* in pages, swap_pager_almost_full warn */
386 SYSCTL_INT(_vm, OID_AUTO, dmmax,
387 CTLFLAG_RD, &dmmax, 0, "Maximum size of a swap block");
389 static void swp_sizecheck(void);
390 static void swp_pager_async_iodone(struct buf *bp);
391 static int swapongeom(struct thread *, struct vnode *);
392 static int swaponvp(struct thread *, struct vnode *, u_long);
393 static int swapoff_one(struct swdevt *sp, struct ucred *cred);
396 * Swap bitmap functions
398 static void swp_pager_freeswapspace(daddr_t blk, int npages);
399 static daddr_t swp_pager_getswapspace(int npages);
404 static struct swblock **swp_pager_hash(vm_object_t object, vm_pindex_t index);
405 static void swp_pager_meta_build(vm_object_t, vm_pindex_t, daddr_t);
406 static void swp_pager_meta_free(vm_object_t, vm_pindex_t, daddr_t);
407 static void swp_pager_meta_free_all(vm_object_t);
408 static daddr_t swp_pager_meta_ctl(vm_object_t, vm_pindex_t, int);
411 swp_pager_free_nrpage(vm_page_t m)
415 if (m->wire_count == 0)
421 * SWP_SIZECHECK() - update swap_pager_full indication
423 * update the swap_pager_almost_full indication and warn when we are
424 * about to run out of swap space, using lowat/hiwat hysteresis.
426 * Clear swap_pager_full ( task killing ) indication when lowat is met.
428 * No restrictions on call
429 * This routine may not block.
435 if (swap_pager_avail < nswap_lowat) {
436 if (swap_pager_almost_full == 0) {
437 printf("swap_pager: out of swap space\n");
438 swap_pager_almost_full = 1;
442 if (swap_pager_avail > nswap_hiwat)
443 swap_pager_almost_full = 0;
448 * SWP_PAGER_HASH() - hash swap meta data
450 * This is an helper function which hashes the swapblk given
451 * the object and page index. It returns a pointer to a pointer
452 * to the object, or a pointer to a NULL pointer if it could not
455 static struct swblock **
456 swp_pager_hash(vm_object_t object, vm_pindex_t index)
458 struct swblock **pswap;
459 struct swblock *swap;
461 index &= ~(vm_pindex_t)SWAP_META_MASK;
462 pswap = &swhash[(index ^ (int)(intptr_t)object) & swhash_mask];
463 while ((swap = *pswap) != NULL) {
464 if (swap->swb_object == object &&
465 swap->swb_index == index
469 pswap = &swap->swb_hnext;
475 * SWAP_PAGER_INIT() - initialize the swap pager!
477 * Expected to be started from system init. NOTE: This code is run
478 * before much else so be careful what you depend on. Most of the VM
479 * system has yet to be initialized at this point.
482 swap_pager_init(void)
485 * Initialize object lists
489 for (i = 0; i < NOBJLISTS; ++i)
490 TAILQ_INIT(&swap_pager_object_list[i]);
491 mtx_init(&sw_alloc_mtx, "swap_pager list", NULL, MTX_DEF);
492 mtx_init(&sw_dev_mtx, "swapdev", NULL, MTX_DEF);
495 * Device Stripe, in PAGE_SIZE'd blocks
497 dmmax = SWB_NPAGES * 2;
501 * SWAP_PAGER_SWAP_INIT() - swap pager initialization from pageout process
503 * Expected to be started from pageout process once, prior to entering
507 swap_pager_swap_init(void)
512 * Number of in-transit swap bp operations. Don't
513 * exhaust the pbufs completely. Make sure we
514 * initialize workable values (0 will work for hysteresis
515 * but it isn't very efficient).
517 * The nsw_cluster_max is constrained by the bp->b_pages[]
518 * array (MAXPHYS/PAGE_SIZE) and our locally defined
519 * MAX_PAGEOUT_CLUSTER. Also be aware that swap ops are
520 * constrained by the swap device interleave stripe size.
522 * Currently we hardwire nsw_wcount_async to 4. This limit is
523 * designed to prevent other I/O from having high latencies due to
524 * our pageout I/O. The value 4 works well for one or two active swap
525 * devices but is probably a little low if you have more. Even so,
526 * a higher value would probably generate only a limited improvement
527 * with three or four active swap devices since the system does not
528 * typically have to pageout at extreme bandwidths. We will want
529 * at least 2 per swap devices, and 4 is a pretty good value if you
530 * have one NFS swap device due to the command/ack latency over NFS.
531 * So it all works out pretty well.
533 nsw_cluster_max = min((MAXPHYS/PAGE_SIZE), MAX_PAGEOUT_CLUSTER);
536 nsw_rcount = (nswbuf + 1) / 2;
537 nsw_wcount_sync = (nswbuf + 3) / 4;
538 nsw_wcount_async = 4;
539 nsw_wcount_async_max = nsw_wcount_async;
540 mtx_unlock(&pbuf_mtx);
543 * Initialize our zone. Right now I'm just guessing on the number
544 * we need based on the number of pages in the system. Each swblock
545 * can hold 16 pages, so this is probably overkill. This reservation
546 * is typically limited to around 32MB by default.
548 n = cnt.v_page_count / 2;
549 if (maxswzone && n > maxswzone / sizeof(struct swblock))
550 n = maxswzone / sizeof(struct swblock);
552 swap_zone = uma_zcreate("SWAPMETA", sizeof(struct swblock), NULL, NULL,
553 NULL, NULL, UMA_ALIGN_PTR, UMA_ZONE_NOFREE | UMA_ZONE_VM);
554 if (swap_zone == NULL)
555 panic("failed to create swap_zone.");
557 if (uma_zone_set_obj(swap_zone, &swap_zone_obj, n))
560 * if the allocation failed, try a zone two thirds the
561 * size of the previous attempt.
566 printf("Swap zone entries reduced from %d to %d.\n", n2, n);
570 * Initialize our meta-data hash table. The swapper does not need to
571 * be quite as efficient as the VM system, so we do not use an
572 * oversized hash table.
574 * n: size of hash table, must be power of 2
575 * swhash_mask: hash table index mask
577 for (n = 1; n < n2 / 8; n *= 2)
579 swhash = malloc(sizeof(struct swblock *) * n, M_VMPGDATA, M_WAITOK | M_ZERO);
581 mtx_init(&swhash_mtx, "swap_pager swhash", NULL, MTX_DEF);
585 * SWAP_PAGER_ALLOC() - allocate a new OBJT_SWAP VM object and instantiate
586 * its metadata structures.
588 * This routine is called from the mmap and fork code to create a new
589 * OBJT_SWAP object. We do this by creating an OBJT_DEFAULT object
590 * and then converting it with swp_pager_meta_build().
592 * This routine may block in vm_object_allocate() and create a named
593 * object lookup race, so we must interlock.
598 swap_pager_alloc(void *handle, vm_ooffset_t size, vm_prot_t prot,
599 vm_ooffset_t offset, struct ucred *cred)
604 pindex = OFF_TO_IDX(offset + PAGE_MASK + size);
608 * Reference existing named region or allocate new one. There
609 * should not be a race here against swp_pager_meta_build()
610 * as called from vm_page_remove() in regards to the lookup
613 sx_xlock(&sw_alloc_sx);
614 object = vm_pager_object_lookup(NOBJLIST(handle), handle);
615 if (object == NULL) {
617 if (!swap_reserve_by_cred(size, cred)) {
618 sx_xunlock(&sw_alloc_sx);
624 object = vm_object_allocate(OBJT_DEFAULT, pindex);
625 VM_OBJECT_LOCK(object);
626 object->handle = handle;
629 object->charge = size;
631 swp_pager_meta_build(object, 0, SWAPBLK_NONE);
632 VM_OBJECT_UNLOCK(object);
634 sx_xunlock(&sw_alloc_sx);
638 if (!swap_reserve_by_cred(size, cred))
642 object = vm_object_allocate(OBJT_DEFAULT, pindex);
643 VM_OBJECT_LOCK(object);
646 object->charge = size;
648 swp_pager_meta_build(object, 0, SWAPBLK_NONE);
649 VM_OBJECT_UNLOCK(object);
655 * SWAP_PAGER_DEALLOC() - remove swap metadata from object
657 * The swap backing for the object is destroyed. The code is
658 * designed such that we can reinstantiate it later, but this
659 * routine is typically called only when the entire object is
660 * about to be destroyed.
662 * The object must be locked.
665 swap_pager_dealloc(vm_object_t object)
669 * Remove from list right away so lookups will fail if we block for
670 * pageout completion.
672 if (object->handle != NULL) {
673 mtx_lock(&sw_alloc_mtx);
674 TAILQ_REMOVE(NOBJLIST(object->handle), object, pager_object_list);
675 mtx_unlock(&sw_alloc_mtx);
678 VM_OBJECT_LOCK_ASSERT(object, MA_OWNED);
679 vm_object_pip_wait(object, "swpdea");
682 * Free all remaining metadata. We only bother to free it from
683 * the swap meta data. We do not attempt to free swapblk's still
684 * associated with vm_page_t's for this object. We do not care
685 * if paging is still in progress on some objects.
687 swp_pager_meta_free_all(object);
690 /************************************************************************
691 * SWAP PAGER BITMAP ROUTINES *
692 ************************************************************************/
695 * SWP_PAGER_GETSWAPSPACE() - allocate raw swap space
697 * Allocate swap for the requested number of pages. The starting
698 * swap block number (a page index) is returned or SWAPBLK_NONE
699 * if the allocation failed.
701 * Also has the side effect of advising that somebody made a mistake
702 * when they configured swap and didn't configure enough.
704 * This routine may not sleep.
706 * We allocate in round-robin fashion from the configured devices.
709 swp_pager_getswapspace(int npages)
716 mtx_lock(&sw_dev_mtx);
718 for (i = 0; i < nswapdev; i++) {
720 sp = TAILQ_FIRST(&swtailq);
721 if (!(sp->sw_flags & SW_CLOSING)) {
722 blk = blist_alloc(sp->sw_blist, npages);
723 if (blk != SWAPBLK_NONE) {
725 sp->sw_used += npages;
726 swap_pager_avail -= npages;
728 swdevhd = TAILQ_NEXT(sp, sw_list);
732 sp = TAILQ_NEXT(sp, sw_list);
734 if (swap_pager_full != 2) {
735 printf("swap_pager_getswapspace(%d): failed\n", npages);
737 swap_pager_almost_full = 1;
741 mtx_unlock(&sw_dev_mtx);
746 swp_pager_isondev(daddr_t blk, struct swdevt *sp)
749 return (blk >= sp->sw_first && blk < sp->sw_end);
753 swp_pager_strategy(struct buf *bp)
757 mtx_lock(&sw_dev_mtx);
758 TAILQ_FOREACH(sp, &swtailq, sw_list) {
759 if (bp->b_blkno >= sp->sw_first && bp->b_blkno < sp->sw_end) {
760 mtx_unlock(&sw_dev_mtx);
761 if ((sp->sw_flags & SW_UNMAPPED) != 0 &&
762 unmapped_buf_allowed) {
763 bp->b_kvaalloc = bp->b_data;
764 bp->b_data = unmapped_buf;
765 bp->b_kvabase = unmapped_buf;
767 bp->b_flags |= B_UNMAPPED;
769 pmap_qenter((vm_offset_t)bp->b_data,
770 &bp->b_pages[0], bp->b_bcount / PAGE_SIZE);
772 sp->sw_strategy(bp, sp);
776 panic("Swapdev not found");
781 * SWP_PAGER_FREESWAPSPACE() - free raw swap space
783 * This routine returns the specified swap blocks back to the bitmap.
785 * This routine may not sleep.
788 swp_pager_freeswapspace(daddr_t blk, int npages)
792 mtx_lock(&sw_dev_mtx);
793 TAILQ_FOREACH(sp, &swtailq, sw_list) {
794 if (blk >= sp->sw_first && blk < sp->sw_end) {
795 sp->sw_used -= npages;
797 * If we are attempting to stop swapping on
798 * this device, we don't want to mark any
799 * blocks free lest they be reused.
801 if ((sp->sw_flags & SW_CLOSING) == 0) {
802 blist_free(sp->sw_blist, blk - sp->sw_first,
804 swap_pager_avail += npages;
807 mtx_unlock(&sw_dev_mtx);
811 panic("Swapdev not found");
815 * SWAP_PAGER_FREESPACE() - frees swap blocks associated with a page
816 * range within an object.
818 * This is a globally accessible routine.
820 * This routine removes swapblk assignments from swap metadata.
822 * The external callers of this routine typically have already destroyed
823 * or renamed vm_page_t's associated with this range in the object so
827 swap_pager_freespace(vm_object_t object, vm_pindex_t start, vm_size_t size)
830 VM_OBJECT_LOCK_ASSERT(object, MA_OWNED);
831 swp_pager_meta_free(object, start, size);
835 * SWAP_PAGER_RESERVE() - reserve swap blocks in object
837 * Assigns swap blocks to the specified range within the object. The
838 * swap blocks are not zeroed. Any previous swap assignment is destroyed.
840 * Returns 0 on success, -1 on failure.
843 swap_pager_reserve(vm_object_t object, vm_pindex_t start, vm_size_t size)
846 daddr_t blk = SWAPBLK_NONE;
847 vm_pindex_t beg = start; /* save start index */
849 VM_OBJECT_LOCK(object);
853 while ((blk = swp_pager_getswapspace(n)) == SWAPBLK_NONE) {
856 swp_pager_meta_free(object, beg, start - beg);
857 VM_OBJECT_UNLOCK(object);
862 swp_pager_meta_build(object, start, blk);
868 swp_pager_meta_free(object, start, n);
869 VM_OBJECT_UNLOCK(object);
874 * SWAP_PAGER_COPY() - copy blocks from source pager to destination pager
875 * and destroy the source.
877 * Copy any valid swapblks from the source to the destination. In
878 * cases where both the source and destination have a valid swapblk,
879 * we keep the destination's.
881 * This routine is allowed to sleep. It may sleep allocating metadata
882 * indirectly through swp_pager_meta_build() or if paging is still in
883 * progress on the source.
885 * The source object contains no vm_page_t's (which is just as well)
887 * The source object is of type OBJT_SWAP.
889 * The source and destination objects must be locked.
890 * Both object locks may temporarily be released.
893 swap_pager_copy(vm_object_t srcobject, vm_object_t dstobject,
894 vm_pindex_t offset, int destroysource)
898 VM_OBJECT_LOCK_ASSERT(srcobject, MA_OWNED);
899 VM_OBJECT_LOCK_ASSERT(dstobject, MA_OWNED);
902 * If destroysource is set, we remove the source object from the
903 * swap_pager internal queue now.
906 if (srcobject->handle != NULL) {
907 mtx_lock(&sw_alloc_mtx);
909 NOBJLIST(srcobject->handle),
913 mtx_unlock(&sw_alloc_mtx);
918 * transfer source to destination.
920 for (i = 0; i < dstobject->size; ++i) {
924 * Locate (without changing) the swapblk on the destination,
925 * unless it is invalid in which case free it silently, or
926 * if the destination is a resident page, in which case the
927 * source is thrown away.
929 dstaddr = swp_pager_meta_ctl(dstobject, i, 0);
931 if (dstaddr == SWAPBLK_NONE) {
933 * Destination has no swapblk and is not resident,
938 srcaddr = swp_pager_meta_ctl(
944 if (srcaddr != SWAPBLK_NONE) {
946 * swp_pager_meta_build() can sleep.
948 vm_object_pip_add(srcobject, 1);
949 VM_OBJECT_UNLOCK(srcobject);
950 vm_object_pip_add(dstobject, 1);
951 swp_pager_meta_build(dstobject, i, srcaddr);
952 vm_object_pip_wakeup(dstobject);
953 VM_OBJECT_LOCK(srcobject);
954 vm_object_pip_wakeup(srcobject);
958 * Destination has valid swapblk or it is represented
959 * by a resident page. We destroy the sourceblock.
962 swp_pager_meta_ctl(srcobject, i + offset, SWM_FREE);
967 * Free left over swap blocks in source.
969 * We have to revert the type to OBJT_DEFAULT so we do not accidently
970 * double-remove the object from the swap queues.
973 swp_pager_meta_free_all(srcobject);
975 * Reverting the type is not necessary, the caller is going
976 * to destroy srcobject directly, but I'm doing it here
977 * for consistency since we've removed the object from its
980 srcobject->type = OBJT_DEFAULT;
985 * SWAP_PAGER_HASPAGE() - determine if we have good backing store for
986 * the requested page.
988 * We determine whether good backing store exists for the requested
989 * page and return TRUE if it does, FALSE if it doesn't.
991 * If TRUE, we also try to determine how much valid, contiguous backing
992 * store exists before and after the requested page within a reasonable
993 * distance. We do not try to restrict it to the swap device stripe
994 * (that is handled in getpages/putpages). It probably isn't worth
998 swap_pager_haspage(vm_object_t object, vm_pindex_t pindex, int *before, int *after)
1002 VM_OBJECT_LOCK_ASSERT(object, MA_OWNED);
1004 * do we have good backing store at the requested index ?
1006 blk0 = swp_pager_meta_ctl(object, pindex, 0);
1008 if (blk0 == SWAPBLK_NONE) {
1017 * find backwards-looking contiguous good backing store
1019 if (before != NULL) {
1022 for (i = 1; i < (SWB_NPAGES/2); ++i) {
1027 blk = swp_pager_meta_ctl(object, pindex - i, 0);
1028 if (blk != blk0 - i)
1035 * find forward-looking contiguous good backing store
1037 if (after != NULL) {
1040 for (i = 1; i < (SWB_NPAGES/2); ++i) {
1043 blk = swp_pager_meta_ctl(object, pindex + i, 0);
1044 if (blk != blk0 + i)
1053 * SWAP_PAGER_PAGE_UNSWAPPED() - remove swap backing store related to page
1055 * This removes any associated swap backing store, whether valid or
1056 * not, from the page.
1058 * This routine is typically called when a page is made dirty, at
1059 * which point any associated swap can be freed. MADV_FREE also
1060 * calls us in a special-case situation
1062 * NOTE!!! If the page is clean and the swap was valid, the caller
1063 * should make the page dirty before calling this routine. This routine
1064 * does NOT change the m->dirty status of the page. Also: MADV_FREE
1067 * This routine may not sleep.
1070 swap_pager_unswapped(vm_page_t m)
1073 VM_OBJECT_LOCK_ASSERT(m->object, MA_OWNED);
1074 swp_pager_meta_ctl(m->object, m->pindex, SWM_FREE);
1078 * SWAP_PAGER_GETPAGES() - bring pages in from swap
1080 * Attempt to retrieve (m, count) pages from backing store, but make
1081 * sure we retrieve at least m[reqpage]. We try to load in as large
1082 * a chunk surrounding m[reqpage] as is contiguous in swap and which
1083 * belongs to the same object.
1085 * The code is designed for asynchronous operation and
1086 * immediate-notification of 'reqpage' but tends not to be
1087 * used that way. Please do not optimize-out this algorithmic
1088 * feature, I intend to improve on it in the future.
1090 * The parent has a single vm_object_pip_add() reference prior to
1091 * calling us and we should return with the same.
1093 * The parent has BUSY'd the pages. We should return with 'm'
1094 * left busy, but the others adjusted.
1097 swap_pager_getpages(vm_object_t object, vm_page_t *m, int count, int reqpage)
1107 KASSERT(mreq->object == object,
1108 ("swap_pager_getpages: object mismatch %p/%p",
1109 object, mreq->object));
1112 * Calculate range to retrieve. The pages have already been assigned
1113 * their swapblks. We require a *contiguous* range but we know it to
1114 * not span devices. If we do not supply it, bad things
1115 * happen. Note that blk, iblk & jblk can be SWAPBLK_NONE, but the
1116 * loops are set up such that the case(s) are handled implicitly.
1118 * The swp_*() calls must be made with the object locked.
1120 blk = swp_pager_meta_ctl(mreq->object, mreq->pindex, 0);
1122 for (i = reqpage - 1; i >= 0; --i) {
1125 iblk = swp_pager_meta_ctl(m[i]->object, m[i]->pindex, 0);
1126 if (blk != iblk + (reqpage - i))
1131 for (j = reqpage + 1; j < count; ++j) {
1134 jblk = swp_pager_meta_ctl(m[j]->object, m[j]->pindex, 0);
1135 if (blk != jblk - (j - reqpage))
1140 * free pages outside our collection range. Note: we never free
1141 * mreq, it must remain busy throughout.
1143 if (0 < i || j < count) {
1146 for (k = 0; k < i; ++k)
1147 swp_pager_free_nrpage(m[k]);
1148 for (k = j; k < count; ++k)
1149 swp_pager_free_nrpage(m[k]);
1153 * Return VM_PAGER_FAIL if we have nothing to do. Return mreq
1154 * still busy, but the others unbusied.
1156 if (blk == SWAPBLK_NONE)
1157 return (VM_PAGER_FAIL);
1160 * Getpbuf() can sleep.
1162 VM_OBJECT_UNLOCK(object);
1164 * Get a swap buffer header to perform the IO
1166 bp = getpbuf(&nsw_rcount);
1167 bp->b_flags |= B_PAGING;
1169 bp->b_iocmd = BIO_READ;
1170 bp->b_iodone = swp_pager_async_iodone;
1171 bp->b_rcred = crhold(thread0.td_ucred);
1172 bp->b_wcred = crhold(thread0.td_ucred);
1173 bp->b_blkno = blk - (reqpage - i);
1174 bp->b_bcount = PAGE_SIZE * (j - i);
1175 bp->b_bufsize = PAGE_SIZE * (j - i);
1176 bp->b_pager.pg_reqpage = reqpage - i;
1178 VM_OBJECT_LOCK(object);
1182 for (k = i; k < j; ++k) {
1183 bp->b_pages[k - i] = m[k];
1184 m[k]->oflags |= VPO_SWAPINPROG;
1187 bp->b_npages = j - i;
1189 PCPU_INC(cnt.v_swapin);
1190 PCPU_ADD(cnt.v_swappgsin, bp->b_npages);
1193 * We still hold the lock on mreq, and our automatic completion routine
1194 * does not remove it.
1196 vm_object_pip_add(object, bp->b_npages);
1197 VM_OBJECT_UNLOCK(object);
1200 * perform the I/O. NOTE!!! bp cannot be considered valid after
1201 * this point because we automatically release it on completion.
1202 * Instead, we look at the one page we are interested in which we
1203 * still hold a lock on even through the I/O completion.
1205 * The other pages in our m[] array are also released on completion,
1206 * so we cannot assume they are valid anymore either.
1208 * NOTE: b_blkno is destroyed by the call to swapdev_strategy
1211 swp_pager_strategy(bp);
1214 * wait for the page we want to complete. VPO_SWAPINPROG is always
1215 * cleared on completion. If an I/O error occurs, SWAPBLK_NONE
1216 * is set in the meta-data.
1218 VM_OBJECT_LOCK(object);
1219 while ((mreq->oflags & VPO_SWAPINPROG) != 0) {
1220 mreq->oflags |= VPO_WANTED;
1221 PCPU_INC(cnt.v_intrans);
1222 if (msleep(mreq, VM_OBJECT_MTX(object), PSWP, "swread", hz*20)) {
1224 "swap_pager: indefinite wait buffer: bufobj: %p, blkno: %jd, size: %ld\n",
1225 bp->b_bufobj, (intmax_t)bp->b_blkno, bp->b_bcount);
1230 * mreq is left busied after completion, but all the other pages
1231 * are freed. If we had an unrecoverable read error the page will
1234 if (mreq->valid != VM_PAGE_BITS_ALL) {
1235 return (VM_PAGER_ERROR);
1237 return (VM_PAGER_OK);
1241 * A final note: in a low swap situation, we cannot deallocate swap
1242 * and mark a page dirty here because the caller is likely to mark
1243 * the page clean when we return, causing the page to possibly revert
1244 * to all-zero's later.
1249 * swap_pager_putpages:
1251 * Assign swap (if necessary) and initiate I/O on the specified pages.
1253 * We support both OBJT_DEFAULT and OBJT_SWAP objects. DEFAULT objects
1254 * are automatically converted to SWAP objects.
1256 * In a low memory situation we may block in VOP_STRATEGY(), but the new
1257 * vm_page reservation system coupled with properly written VFS devices
1258 * should ensure that no low-memory deadlock occurs. This is an area
1261 * The parent has N vm_object_pip_add() references prior to
1262 * calling us and will remove references for rtvals[] that are
1263 * not set to VM_PAGER_PEND. We need to remove the rest on I/O
1266 * The parent has soft-busy'd the pages it passes us and will unbusy
1267 * those whos rtvals[] entry is not set to VM_PAGER_PEND on return.
1268 * We need to unbusy the rest on I/O completion.
1271 swap_pager_putpages(vm_object_t object, vm_page_t *m, int count,
1272 boolean_t sync, int *rtvals)
1277 if (count && m[0]->object != object) {
1278 panic("swap_pager_putpages: object mismatch %p/%p",
1287 * Turn object into OBJT_SWAP
1288 * check for bogus sysops
1289 * force sync if not pageout process
1291 if (object->type != OBJT_SWAP)
1292 swp_pager_meta_build(object, 0, SWAPBLK_NONE);
1293 VM_OBJECT_UNLOCK(object);
1295 if (curproc != pageproc)
1301 * Update nsw parameters from swap_async_max sysctl values.
1302 * Do not let the sysop crash the machine with bogus numbers.
1304 mtx_lock(&pbuf_mtx);
1305 if (swap_async_max != nsw_wcount_async_max) {
1311 if ((n = swap_async_max) > nswbuf / 2)
1318 * Adjust difference ( if possible ). If the current async
1319 * count is too low, we may not be able to make the adjustment
1322 n -= nsw_wcount_async_max;
1323 if (nsw_wcount_async + n >= 0) {
1324 nsw_wcount_async += n;
1325 nsw_wcount_async_max += n;
1326 wakeup(&nsw_wcount_async);
1329 mtx_unlock(&pbuf_mtx);
1334 * Assign swap blocks and issue I/O. We reallocate swap on the fly.
1335 * The page is left dirty until the pageout operation completes
1338 for (i = 0; i < count; i += n) {
1344 * Maximum I/O size is limited by a number of factors.
1346 n = min(BLIST_MAX_ALLOC, count - i);
1347 n = min(n, nsw_cluster_max);
1350 * Get biggest block of swap we can. If we fail, fall
1351 * back and try to allocate a smaller block. Don't go
1352 * overboard trying to allocate space if it would overly
1356 (blk = swp_pager_getswapspace(n)) == SWAPBLK_NONE &&
1361 if (blk == SWAPBLK_NONE) {
1362 for (j = 0; j < n; ++j)
1363 rtvals[i+j] = VM_PAGER_FAIL;
1368 * All I/O parameters have been satisfied, build the I/O
1369 * request and assign the swap space.
1372 bp = getpbuf(&nsw_wcount_sync);
1374 bp = getpbuf(&nsw_wcount_async);
1375 bp->b_flags = B_ASYNC;
1377 bp->b_flags |= B_PAGING;
1378 bp->b_iocmd = BIO_WRITE;
1380 bp->b_rcred = crhold(thread0.td_ucred);
1381 bp->b_wcred = crhold(thread0.td_ucred);
1382 bp->b_bcount = PAGE_SIZE * n;
1383 bp->b_bufsize = PAGE_SIZE * n;
1386 VM_OBJECT_LOCK(object);
1387 for (j = 0; j < n; ++j) {
1388 vm_page_t mreq = m[i+j];
1390 swp_pager_meta_build(
1395 vm_page_dirty(mreq);
1396 rtvals[i+j] = VM_PAGER_OK;
1398 mreq->oflags |= VPO_SWAPINPROG;
1399 bp->b_pages[j] = mreq;
1401 VM_OBJECT_UNLOCK(object);
1404 * Must set dirty range for NFS to work.
1407 bp->b_dirtyend = bp->b_bcount;
1409 PCPU_INC(cnt.v_swapout);
1410 PCPU_ADD(cnt.v_swappgsout, bp->b_npages);
1415 * NOTE: b_blkno is destroyed by the call to swapdev_strategy
1417 if (sync == FALSE) {
1418 bp->b_iodone = swp_pager_async_iodone;
1420 swp_pager_strategy(bp);
1422 for (j = 0; j < n; ++j)
1423 rtvals[i+j] = VM_PAGER_PEND;
1424 /* restart outter loop */
1431 * NOTE: b_blkno is destroyed by the call to swapdev_strategy
1433 bp->b_iodone = bdone;
1434 swp_pager_strategy(bp);
1437 * Wait for the sync I/O to complete, then update rtvals.
1438 * We just set the rtvals[] to VM_PAGER_PEND so we can call
1439 * our async completion routine at the end, thus avoiding a
1442 bwait(bp, PVM, "swwrt");
1443 for (j = 0; j < n; ++j)
1444 rtvals[i+j] = VM_PAGER_PEND;
1446 * Now that we are through with the bp, we can call the
1447 * normal async completion, which frees everything up.
1449 swp_pager_async_iodone(bp);
1451 VM_OBJECT_LOCK(object);
1455 * swp_pager_async_iodone:
1457 * Completion routine for asynchronous reads and writes from/to swap.
1458 * Also called manually by synchronous code to finish up a bp.
1460 * For READ operations, the pages are VPO_BUSY'd. For WRITE operations,
1461 * the pages are vm_page_t->busy'd. For READ operations, we VPO_BUSY
1462 * unbusy all pages except the 'main' request page. For WRITE
1463 * operations, we vm_page_t->busy'd unbusy all pages ( we can do this
1464 * because we marked them all VM_PAGER_PEND on return from putpages ).
1466 * This routine may not sleep.
1469 swp_pager_async_iodone(struct buf *bp)
1472 vm_object_t object = NULL;
1477 if (bp->b_ioflags & BIO_ERROR) {
1479 "swap_pager: I/O error - %s failed; blkno %ld,"
1480 "size %ld, error %d\n",
1481 ((bp->b_iocmd == BIO_READ) ? "pagein" : "pageout"),
1489 * remove the mapping for kernel virtual
1491 if ((bp->b_flags & B_UNMAPPED) != 0) {
1492 bp->b_data = bp->b_kvaalloc;
1493 bp->b_kvabase = bp->b_kvaalloc;
1494 bp->b_flags &= ~B_UNMAPPED;
1496 pmap_qremove((vm_offset_t)bp->b_data, bp->b_npages);
1499 object = bp->b_pages[0]->object;
1500 VM_OBJECT_LOCK(object);
1504 * cleanup pages. If an error occurs writing to swap, we are in
1505 * very serious trouble. If it happens to be a disk error, though,
1506 * we may be able to recover by reassigning the swap later on. So
1507 * in this case we remove the m->swapblk assignment for the page
1508 * but do not free it in the rlist. The errornous block(s) are thus
1509 * never reallocated as swap. Redirty the page and continue.
1511 for (i = 0; i < bp->b_npages; ++i) {
1512 vm_page_t m = bp->b_pages[i];
1514 m->oflags &= ~VPO_SWAPINPROG;
1516 if (bp->b_ioflags & BIO_ERROR) {
1518 * If an error occurs I'd love to throw the swapblk
1519 * away without freeing it back to swapspace, so it
1520 * can never be used again. But I can't from an
1523 if (bp->b_iocmd == BIO_READ) {
1525 * When reading, reqpage needs to stay
1526 * locked for the parent, but all other
1527 * pages can be freed. We still want to
1528 * wakeup the parent waiting on the page,
1529 * though. ( also: pg_reqpage can be -1 and
1530 * not match anything ).
1532 * We have to wake specifically requested pages
1533 * up too because we cleared VPO_SWAPINPROG and
1534 * someone may be waiting for that.
1536 * NOTE: for reads, m->dirty will probably
1537 * be overridden by the original caller of
1538 * getpages so don't play cute tricks here.
1541 if (i != bp->b_pager.pg_reqpage)
1542 swp_pager_free_nrpage(m);
1546 * If i == bp->b_pager.pg_reqpage, do not wake
1547 * the page up. The caller needs to.
1551 * If a write error occurs, reactivate page
1552 * so it doesn't clog the inactive list,
1553 * then finish the I/O.
1557 vm_page_activate(m);
1559 vm_page_io_finish(m);
1561 } else if (bp->b_iocmd == BIO_READ) {
1563 * NOTE: for reads, m->dirty will probably be
1564 * overridden by the original caller of getpages so
1565 * we cannot set them in order to free the underlying
1566 * swap in a low-swap situation. I don't think we'd
1567 * want to do that anyway, but it was an optimization
1568 * that existed in the old swapper for a time before
1569 * it got ripped out due to precisely this problem.
1571 * If not the requested page then deactivate it.
1573 * Note that the requested page, reqpage, is left
1574 * busied, but we still have to wake it up. The
1575 * other pages are released (unbusied) by
1578 KASSERT(!pmap_page_is_mapped(m),
1579 ("swp_pager_async_iodone: page %p is mapped", m));
1580 m->valid = VM_PAGE_BITS_ALL;
1581 KASSERT(m->dirty == 0,
1582 ("swp_pager_async_iodone: page %p is dirty", m));
1585 * We have to wake specifically requested pages
1586 * up too because we cleared VPO_SWAPINPROG and
1587 * could be waiting for it in getpages. However,
1588 * be sure to not unbusy getpages specifically
1589 * requested page - getpages expects it to be
1592 if (i != bp->b_pager.pg_reqpage) {
1594 vm_page_deactivate(m);
1601 * For write success, clear the dirty
1602 * status, then finish the I/O ( which decrements the
1603 * busy count and possibly wakes waiter's up ).
1605 KASSERT(!pmap_page_is_write_mapped(m),
1606 ("swp_pager_async_iodone: page %p is not write"
1609 vm_page_io_finish(m);
1610 if (vm_page_count_severe()) {
1612 vm_page_try_to_cache(m);
1619 * adjust pip. NOTE: the original parent may still have its own
1620 * pip refs on the object.
1622 if (object != NULL) {
1623 vm_object_pip_wakeupn(object, bp->b_npages);
1624 VM_OBJECT_UNLOCK(object);
1628 * swapdev_strategy() manually sets b_vp and b_bufobj before calling
1629 * bstrategy(). Set them back to NULL now we're done with it, or we'll
1630 * trigger a KASSERT in relpbuf().
1634 bp->b_bufobj = NULL;
1637 * release the physical I/O buffer
1641 ((bp->b_iocmd == BIO_READ) ? &nsw_rcount :
1642 ((bp->b_flags & B_ASYNC) ?
1651 * swap_pager_isswapped:
1653 * Return 1 if at least one page in the given object is paged
1654 * out to the given swap device.
1656 * This routine may not sleep.
1659 swap_pager_isswapped(vm_object_t object, struct swdevt *sp)
1665 VM_OBJECT_LOCK_ASSERT(object, MA_OWNED);
1666 if (object->type != OBJT_SWAP)
1669 mtx_lock(&swhash_mtx);
1670 for (bcount = 0; bcount < object->un_pager.swp.swp_bcount; bcount++) {
1671 struct swblock *swap;
1673 if ((swap = *swp_pager_hash(object, index)) != NULL) {
1674 for (i = 0; i < SWAP_META_PAGES; ++i) {
1675 if (swp_pager_isondev(swap->swb_pages[i], sp)) {
1676 mtx_unlock(&swhash_mtx);
1681 index += SWAP_META_PAGES;
1683 mtx_unlock(&swhash_mtx);
1688 * SWP_PAGER_FORCE_PAGEIN() - force a swap block to be paged in
1690 * This routine dissociates the page at the given index within a
1691 * swap block from its backing store, paging it in if necessary.
1692 * If the page is paged in, it is placed in the inactive queue,
1693 * since it had its backing store ripped out from under it.
1694 * We also attempt to swap in all other pages in the swap block,
1695 * we only guarantee that the one at the specified index is
1698 * XXX - The code to page the whole block in doesn't work, so we
1699 * revert to the one-by-one behavior for now. Sigh.
1702 swp_pager_force_pagein(vm_object_t object, vm_pindex_t pindex)
1706 vm_object_pip_add(object, 1);
1707 m = vm_page_grab(object, pindex, VM_ALLOC_NORMAL|VM_ALLOC_RETRY);
1708 if (m->valid == VM_PAGE_BITS_ALL) {
1709 vm_object_pip_subtract(object, 1);
1712 vm_page_activate(m);
1715 vm_pager_page_unswapped(m);
1719 if (swap_pager_getpages(object, &m, 1, 0) != VM_PAGER_OK)
1720 panic("swap_pager_force_pagein: read from swap failed");/*XXX*/
1721 vm_object_pip_subtract(object, 1);
1724 vm_page_deactivate(m);
1727 vm_pager_page_unswapped(m);
1731 * swap_pager_swapoff:
1733 * Page in all of the pages that have been paged out to the
1734 * given device. The corresponding blocks in the bitmap must be
1735 * marked as allocated and the device must be flagged SW_CLOSING.
1736 * There may be no processes swapped out to the device.
1738 * This routine may block.
1741 swap_pager_swapoff(struct swdevt *sp)
1743 struct swblock *swap;
1750 mtx_lock(&swhash_mtx);
1751 for (i = 0; i <= swhash_mask; i++) { /* '<=' is correct here */
1753 for (swap = swhash[i]; swap != NULL; swap = swap->swb_hnext) {
1754 vm_object_t object = swap->swb_object;
1755 vm_pindex_t pindex = swap->swb_index;
1756 for (j = 0; j < SWAP_META_PAGES; ++j) {
1757 if (swp_pager_isondev(swap->swb_pages[j], sp)) {
1758 /* avoid deadlock */
1759 if (!VM_OBJECT_TRYLOCK(object)) {
1762 mtx_unlock(&swhash_mtx);
1763 swp_pager_force_pagein(object,
1765 VM_OBJECT_UNLOCK(object);
1766 mtx_lock(&swhash_mtx);
1773 mtx_unlock(&swhash_mtx);
1776 * Objects may be locked or paging to the device being
1777 * removed, so we will miss their pages and need to
1778 * make another pass. We have marked this device as
1779 * SW_CLOSING, so the activity should finish soon.
1782 if (retries > 100) {
1783 panic("swapoff: failed to locate %d swap blocks",
1786 pause("swpoff", hz / 20);
1791 /************************************************************************
1793 ************************************************************************
1795 * These routines manipulate the swap metadata stored in the
1798 * Swap metadata is implemented with a global hash and not directly
1799 * linked into the object. Instead the object simply contains
1800 * appropriate tracking counters.
1804 * SWP_PAGER_META_BUILD() - add swap block to swap meta data for object
1806 * We first convert the object to a swap object if it is a default
1809 * The specified swapblk is added to the object's swap metadata. If
1810 * the swapblk is not valid, it is freed instead. Any previously
1811 * assigned swapblk is freed.
1814 swp_pager_meta_build(vm_object_t object, vm_pindex_t pindex, daddr_t swapblk)
1816 static volatile int exhausted;
1817 struct swblock *swap;
1818 struct swblock **pswap;
1821 VM_OBJECT_LOCK_ASSERT(object, MA_OWNED);
1823 * Convert default object to swap object if necessary
1825 if (object->type != OBJT_SWAP) {
1826 object->type = OBJT_SWAP;
1827 object->un_pager.swp.swp_bcount = 0;
1829 if (object->handle != NULL) {
1830 mtx_lock(&sw_alloc_mtx);
1832 NOBJLIST(object->handle),
1836 mtx_unlock(&sw_alloc_mtx);
1841 * Locate hash entry. If not found create, but if we aren't adding
1842 * anything just return. If we run out of space in the map we wait
1843 * and, since the hash table may have changed, retry.
1846 mtx_lock(&swhash_mtx);
1847 pswap = swp_pager_hash(object, pindex);
1849 if ((swap = *pswap) == NULL) {
1852 if (swapblk == SWAPBLK_NONE)
1855 swap = *pswap = uma_zalloc(swap_zone, M_NOWAIT);
1857 mtx_unlock(&swhash_mtx);
1858 VM_OBJECT_UNLOCK(object);
1859 if (uma_zone_exhausted(swap_zone)) {
1860 if (atomic_cmpset_rel_int(&exhausted, 0, 1))
1861 printf("swap zone exhausted, "
1862 "increase kern.maxswzone\n");
1863 vm_pageout_oom(VM_OOM_SWAPZ);
1864 pause("swzonex", 10);
1867 VM_OBJECT_LOCK(object);
1871 if (atomic_cmpset_rel_int(&exhausted, 1, 0))
1872 printf("swap zone ok\n");
1874 swap->swb_hnext = NULL;
1875 swap->swb_object = object;
1876 swap->swb_index = pindex & ~(vm_pindex_t)SWAP_META_MASK;
1877 swap->swb_count = 0;
1879 ++object->un_pager.swp.swp_bcount;
1881 for (i = 0; i < SWAP_META_PAGES; ++i)
1882 swap->swb_pages[i] = SWAPBLK_NONE;
1886 * Delete prior contents of metadata
1888 idx = pindex & SWAP_META_MASK;
1890 if (swap->swb_pages[idx] != SWAPBLK_NONE) {
1891 swp_pager_freeswapspace(swap->swb_pages[idx], 1);
1896 * Enter block into metadata
1898 swap->swb_pages[idx] = swapblk;
1899 if (swapblk != SWAPBLK_NONE)
1902 mtx_unlock(&swhash_mtx);
1906 * SWP_PAGER_META_FREE() - free a range of blocks in the object's swap metadata
1908 * The requested range of blocks is freed, with any associated swap
1909 * returned to the swap bitmap.
1911 * This routine will free swap metadata structures as they are cleaned
1912 * out. This routine does *NOT* operate on swap metadata associated
1913 * with resident pages.
1916 swp_pager_meta_free(vm_object_t object, vm_pindex_t index, daddr_t count)
1919 VM_OBJECT_LOCK_ASSERT(object, MA_OWNED);
1920 if (object->type != OBJT_SWAP)
1924 struct swblock **pswap;
1925 struct swblock *swap;
1927 mtx_lock(&swhash_mtx);
1928 pswap = swp_pager_hash(object, index);
1930 if ((swap = *pswap) != NULL) {
1931 daddr_t v = swap->swb_pages[index & SWAP_META_MASK];
1933 if (v != SWAPBLK_NONE) {
1934 swp_pager_freeswapspace(v, 1);
1935 swap->swb_pages[index & SWAP_META_MASK] =
1937 if (--swap->swb_count == 0) {
1938 *pswap = swap->swb_hnext;
1939 uma_zfree(swap_zone, swap);
1940 --object->un_pager.swp.swp_bcount;
1946 int n = SWAP_META_PAGES - (index & SWAP_META_MASK);
1950 mtx_unlock(&swhash_mtx);
1955 * SWP_PAGER_META_FREE_ALL() - destroy all swap metadata associated with object
1957 * This routine locates and destroys all swap metadata associated with
1961 swp_pager_meta_free_all(vm_object_t object)
1965 VM_OBJECT_LOCK_ASSERT(object, MA_OWNED);
1966 if (object->type != OBJT_SWAP)
1969 while (object->un_pager.swp.swp_bcount) {
1970 struct swblock **pswap;
1971 struct swblock *swap;
1973 mtx_lock(&swhash_mtx);
1974 pswap = swp_pager_hash(object, index);
1975 if ((swap = *pswap) != NULL) {
1978 for (i = 0; i < SWAP_META_PAGES; ++i) {
1979 daddr_t v = swap->swb_pages[i];
1980 if (v != SWAPBLK_NONE) {
1982 swp_pager_freeswapspace(v, 1);
1985 if (swap->swb_count != 0)
1986 panic("swap_pager_meta_free_all: swb_count != 0");
1987 *pswap = swap->swb_hnext;
1988 uma_zfree(swap_zone, swap);
1989 --object->un_pager.swp.swp_bcount;
1991 mtx_unlock(&swhash_mtx);
1992 index += SWAP_META_PAGES;
1997 * SWP_PAGER_METACTL() - misc control of swap and vm_page_t meta data.
1999 * This routine is capable of looking up, popping, or freeing
2000 * swapblk assignments in the swap meta data or in the vm_page_t.
2001 * The routine typically returns the swapblk being looked-up, or popped,
2002 * or SWAPBLK_NONE if the block was freed, or SWAPBLK_NONE if the block
2003 * was invalid. This routine will automatically free any invalid
2004 * meta-data swapblks.
2006 * It is not possible to store invalid swapblks in the swap meta data
2007 * (other then a literal 'SWAPBLK_NONE'), so we don't bother checking.
2009 * When acting on a busy resident page and paging is in progress, we
2010 * have to wait until paging is complete but otherwise can act on the
2013 * SWM_FREE remove and free swap block from metadata
2014 * SWM_POP remove from meta data but do not free.. pop it out
2017 swp_pager_meta_ctl(vm_object_t object, vm_pindex_t pindex, int flags)
2019 struct swblock **pswap;
2020 struct swblock *swap;
2024 VM_OBJECT_LOCK_ASSERT(object, MA_OWNED);
2026 * The meta data only exists of the object is OBJT_SWAP
2027 * and even then might not be allocated yet.
2029 if (object->type != OBJT_SWAP)
2030 return (SWAPBLK_NONE);
2033 mtx_lock(&swhash_mtx);
2034 pswap = swp_pager_hash(object, pindex);
2036 if ((swap = *pswap) != NULL) {
2037 idx = pindex & SWAP_META_MASK;
2038 r1 = swap->swb_pages[idx];
2040 if (r1 != SWAPBLK_NONE) {
2041 if (flags & SWM_FREE) {
2042 swp_pager_freeswapspace(r1, 1);
2045 if (flags & (SWM_FREE|SWM_POP)) {
2046 swap->swb_pages[idx] = SWAPBLK_NONE;
2047 if (--swap->swb_count == 0) {
2048 *pswap = swap->swb_hnext;
2049 uma_zfree(swap_zone, swap);
2050 --object->un_pager.swp.swp_bcount;
2055 mtx_unlock(&swhash_mtx);
2060 * System call swapon(name) enables swapping on device name,
2061 * which must be in the swdevsw. Return EBUSY
2062 * if already swapping on this device.
2064 #ifndef _SYS_SYSPROTO_H_
2065 struct swapon_args {
2075 sys_swapon(struct thread *td, struct swapon_args *uap)
2079 struct nameidata nd;
2082 error = priv_check(td, PRIV_SWAPON);
2087 while (swdev_syscall_active)
2088 tsleep(&swdev_syscall_active, PUSER - 1, "swpon", 0);
2089 swdev_syscall_active = 1;
2092 * Swap metadata may not fit in the KVM if we have physical
2095 if (swap_zone == NULL) {
2100 NDINIT(&nd, LOOKUP, ISOPEN | FOLLOW | AUDITVNODE1, UIO_USERSPACE,
2106 NDFREE(&nd, NDF_ONLY_PNBUF);
2109 if (vn_isdisk(vp, &error)) {
2110 error = swapongeom(td, vp);
2111 } else if (vp->v_type == VREG &&
2112 (vp->v_mount->mnt_vfc->vfc_flags & VFCF_NETWORK) != 0 &&
2113 (error = VOP_GETATTR(vp, &attr, td->td_ucred)) == 0) {
2115 * Allow direct swapping to NFS regular files in the same
2116 * way that nfs_mountroot() sets up diskless swapping.
2118 error = swaponvp(td, vp, attr.va_size / DEV_BSIZE);
2124 swdev_syscall_active = 0;
2125 wakeup_one(&swdev_syscall_active);
2131 * Check that the total amount of swap currently configured does not
2132 * exceed half the theoretical maximum. If it does, print a warning
2133 * message and return -1; otherwise, return 0.
2136 swapon_check_swzone(unsigned long npages)
2138 unsigned long maxpages;
2140 /* absolute maximum we can handle assuming 100% efficiency */
2141 maxpages = uma_zone_get_max(swap_zone) * SWAP_META_PAGES;
2143 /* recommend using no more than half that amount */
2144 if (npages > maxpages / 2) {
2145 printf("warning: total configured swap (%lu pages) "
2146 "exceeds maximum recommended amount (%lu pages).\n",
2147 npages, maxpages / 2);
2148 printf("warning: increase kern.maxswzone "
2149 "or reduce amount of swap.\n");
2156 swaponsomething(struct vnode *vp, void *id, u_long nblks,
2157 sw_strategy_t *strategy, sw_close_t *close, dev_t dev, int flags)
2159 struct swdevt *sp, *tsp;
2164 * nblks is in DEV_BSIZE'd chunks, convert to PAGE_SIZE'd chunks.
2165 * First chop nblks off to page-align it, then convert.
2167 * sw->sw_nblks is in page-sized chunks now too.
2169 nblks &= ~(ctodb(1) - 1);
2170 nblks = dbtoc(nblks);
2173 * If we go beyond this, we get overflows in the radix
2176 mblocks = 0x40000000 / BLIST_META_RADIX;
2177 if (nblks > mblocks) {
2179 "WARNING: reducing swap size to maximum of %luMB per unit\n",
2180 mblocks / 1024 / 1024 * PAGE_SIZE);
2184 sp = malloc(sizeof *sp, M_VMPGDATA, M_WAITOK | M_ZERO);
2189 sp->sw_nblks = nblks;
2191 sp->sw_strategy = strategy;
2192 sp->sw_close = close;
2193 sp->sw_flags = flags;
2195 sp->sw_blist = blist_create(nblks, M_WAITOK);
2197 * Do not free the first two block in order to avoid overwriting
2198 * any bsd label at the front of the partition
2200 blist_free(sp->sw_blist, 2, nblks - 2);
2203 mtx_lock(&sw_dev_mtx);
2204 TAILQ_FOREACH(tsp, &swtailq, sw_list) {
2205 if (tsp->sw_end >= dvbase) {
2207 * We put one uncovered page between the devices
2208 * in order to definitively prevent any cross-device
2211 dvbase = tsp->sw_end + 1;
2214 sp->sw_first = dvbase;
2215 sp->sw_end = dvbase + nblks;
2216 TAILQ_INSERT_TAIL(&swtailq, sp, sw_list);
2218 swap_pager_avail += nblks;
2219 swap_total += (vm_ooffset_t)nblks * PAGE_SIZE;
2220 swapon_check_swzone(swap_total / PAGE_SIZE);
2222 mtx_unlock(&sw_dev_mtx);
2226 * SYSCALL: swapoff(devname)
2228 * Disable swapping on the given device.
2230 * XXX: Badly designed system call: it should use a device index
2231 * rather than filename as specification. We keep sw_vp around
2232 * only to make this work.
2234 #ifndef _SYS_SYSPROTO_H_
2235 struct swapoff_args {
2245 sys_swapoff(struct thread *td, struct swapoff_args *uap)
2248 struct nameidata nd;
2252 error = priv_check(td, PRIV_SWAPOFF);
2257 while (swdev_syscall_active)
2258 tsleep(&swdev_syscall_active, PUSER - 1, "swpoff", 0);
2259 swdev_syscall_active = 1;
2261 NDINIT(&nd, LOOKUP, FOLLOW | AUDITVNODE1, UIO_USERSPACE, uap->name,
2266 NDFREE(&nd, NDF_ONLY_PNBUF);
2269 mtx_lock(&sw_dev_mtx);
2270 TAILQ_FOREACH(sp, &swtailq, sw_list) {
2271 if (sp->sw_vp == vp)
2274 mtx_unlock(&sw_dev_mtx);
2279 error = swapoff_one(sp, td->td_ucred);
2281 swdev_syscall_active = 0;
2282 wakeup_one(&swdev_syscall_active);
2288 swapoff_one(struct swdevt *sp, struct ucred *cred)
2290 u_long nblks, dvbase;
2295 mtx_assert(&Giant, MA_OWNED);
2297 (void) vn_lock(sp->sw_vp, LK_EXCLUSIVE | LK_RETRY);
2298 error = mac_system_check_swapoff(cred, sp->sw_vp);
2299 (void) VOP_UNLOCK(sp->sw_vp, 0);
2303 nblks = sp->sw_nblks;
2306 * We can turn off this swap device safely only if the
2307 * available virtual memory in the system will fit the amount
2308 * of data we will have to page back in, plus an epsilon so
2309 * the system doesn't become critically low on swap space.
2311 if (cnt.v_free_count + cnt.v_cache_count + swap_pager_avail <
2312 nblks + nswap_lowat) {
2317 * Prevent further allocations on this device.
2319 mtx_lock(&sw_dev_mtx);
2320 sp->sw_flags |= SW_CLOSING;
2321 for (dvbase = 0; dvbase < sp->sw_end; dvbase += dmmax) {
2322 swap_pager_avail -= blist_fill(sp->sw_blist,
2325 swap_total -= (vm_ooffset_t)nblks * PAGE_SIZE;
2326 mtx_unlock(&sw_dev_mtx);
2329 * Page in the contents of the device and close it.
2331 swap_pager_swapoff(sp);
2333 sp->sw_close(curthread, sp);
2335 mtx_lock(&sw_dev_mtx);
2336 TAILQ_REMOVE(&swtailq, sp, sw_list);
2338 if (nswapdev == 0) {
2339 swap_pager_full = 2;
2340 swap_pager_almost_full = 1;
2344 mtx_unlock(&sw_dev_mtx);
2345 blist_destroy(sp->sw_blist);
2346 free(sp, M_VMPGDATA);
2353 struct swdevt *sp, *spt;
2354 const char *devname;
2358 while (swdev_syscall_active)
2359 tsleep(&swdev_syscall_active, PUSER - 1, "swpoff", 0);
2360 swdev_syscall_active = 1;
2362 mtx_lock(&sw_dev_mtx);
2363 TAILQ_FOREACH_SAFE(sp, &swtailq, sw_list, spt) {
2364 mtx_unlock(&sw_dev_mtx);
2365 if (vn_isdisk(sp->sw_vp, NULL))
2366 devname = devtoname(sp->sw_vp->v_rdev);
2369 error = swapoff_one(sp, thread0.td_ucred);
2371 printf("Cannot remove swap device %s (error=%d), "
2372 "skipping.\n", devname, error);
2373 } else if (bootverbose) {
2374 printf("Swap device %s removed.\n", devname);
2376 mtx_lock(&sw_dev_mtx);
2378 mtx_unlock(&sw_dev_mtx);
2380 swdev_syscall_active = 0;
2381 wakeup_one(&swdev_syscall_active);
2386 swap_pager_status(int *total, int *used)
2392 mtx_lock(&sw_dev_mtx);
2393 TAILQ_FOREACH(sp, &swtailq, sw_list) {
2394 *total += sp->sw_nblks;
2395 *used += sp->sw_used;
2397 mtx_unlock(&sw_dev_mtx);
2401 swap_dev_info(int name, struct xswdev *xs, char *devname, size_t len)
2404 const char *tmp_devname;
2409 mtx_lock(&sw_dev_mtx);
2410 TAILQ_FOREACH(sp, &swtailq, sw_list) {
2415 xs->xsw_version = XSWDEV_VERSION;
2416 xs->xsw_dev = sp->sw_dev;
2417 xs->xsw_flags = sp->sw_flags;
2418 xs->xsw_nblks = sp->sw_nblks;
2419 xs->xsw_used = sp->sw_used;
2420 if (devname != NULL) {
2421 if (vn_isdisk(sp->sw_vp, NULL))
2422 tmp_devname = devtoname(sp->sw_vp->v_rdev);
2424 tmp_devname = "[file]";
2425 strncpy(devname, tmp_devname, len);
2430 mtx_unlock(&sw_dev_mtx);
2435 sysctl_vm_swap_info(SYSCTL_HANDLER_ARGS)
2440 if (arg2 != 1) /* name length */
2442 error = swap_dev_info(*(int *)arg1, &xs, NULL, 0);
2445 error = SYSCTL_OUT(req, &xs, sizeof(xs));
2449 SYSCTL_INT(_vm, OID_AUTO, nswapdev, CTLFLAG_RD, &nswapdev, 0,
2450 "Number of swap devices");
2451 SYSCTL_NODE(_vm, OID_AUTO, swap_info, CTLFLAG_RD, sysctl_vm_swap_info,
2452 "Swap statistics by device");
2455 * vmspace_swap_count() - count the approximate swap usage in pages for a
2458 * The map must be locked.
2460 * Swap usage is determined by taking the proportional swap used by
2461 * VM objects backing the VM map. To make up for fractional losses,
2462 * if the VM object has any swap use at all the associated map entries
2463 * count for at least 1 swap page.
2466 vmspace_swap_count(struct vmspace *vmspace)
2473 map = &vmspace->vm_map;
2476 for (cur = map->header.next; cur != &map->header; cur = cur->next) {
2477 if ((cur->eflags & MAP_ENTRY_IS_SUB_MAP) == 0 &&
2478 (object = cur->object.vm_object) != NULL) {
2479 VM_OBJECT_LOCK(object);
2480 if (object->type == OBJT_SWAP &&
2481 object->un_pager.swp.swp_bcount != 0) {
2482 n = (cur->end - cur->start) / PAGE_SIZE;
2483 count += object->un_pager.swp.swp_bcount *
2484 SWAP_META_PAGES * n / object->size + 1;
2486 VM_OBJECT_UNLOCK(object);
2495 * Swapping onto disk devices.
2499 static g_orphan_t swapgeom_orphan;
2501 static struct g_class g_swap_class = {
2503 .version = G_VERSION,
2504 .orphan = swapgeom_orphan,
2507 DECLARE_GEOM_CLASS(g_swap_class, g_class);
2511 swapgeom_done(struct bio *bp2)
2515 bp = bp2->bio_caller2;
2516 bp->b_ioflags = bp2->bio_flags;
2518 bp->b_ioflags |= BIO_ERROR;
2519 bp->b_resid = bp->b_bcount - bp2->bio_completed;
2520 bp->b_error = bp2->bio_error;
2526 swapgeom_strategy(struct buf *bp, struct swdevt *sp)
2529 struct g_consumer *cp;
2533 bp->b_error = ENXIO;
2534 bp->b_ioflags |= BIO_ERROR;
2538 if (bp->b_iocmd == BIO_WRITE)
2541 bio = g_alloc_bio();
2543 bp->b_error = ENOMEM;
2544 bp->b_ioflags |= BIO_ERROR;
2549 bio->bio_caller2 = bp;
2550 bio->bio_cmd = bp->b_iocmd;
2551 bio->bio_offset = (bp->b_blkno - sp->sw_first) * PAGE_SIZE;
2552 bio->bio_length = bp->b_bcount;
2553 bio->bio_done = swapgeom_done;
2554 if ((bp->b_flags & B_UNMAPPED) != 0) {
2555 bio->bio_ma = bp->b_pages;
2556 bio->bio_data = unmapped_buf;
2557 bio->bio_ma_offset = (vm_offset_t)bp->b_offset & PAGE_MASK;
2558 bio->bio_ma_n = bp->b_npages;
2559 bio->bio_flags |= BIO_UNMAPPED;
2561 bio->bio_data = bp->b_data;
2564 g_io_request(bio, cp);
2569 swapgeom_orphan(struct g_consumer *cp)
2573 mtx_lock(&sw_dev_mtx);
2574 TAILQ_FOREACH(sp, &swtailq, sw_list)
2575 if (sp->sw_id == cp)
2576 sp->sw_flags |= SW_CLOSING;
2577 mtx_unlock(&sw_dev_mtx);
2581 swapgeom_close_ev(void *arg, int flags)
2583 struct g_consumer *cp;
2586 g_access(cp, -1, -1, 0);
2588 g_destroy_consumer(cp);
2592 swapgeom_close(struct thread *td, struct swdevt *sw)
2595 /* XXX: direct call when Giant untangled */
2596 g_waitfor_event(swapgeom_close_ev, sw->sw_id, M_WAITOK, NULL);
2607 swapongeom_ev(void *arg, int flags)
2610 struct g_provider *pp;
2611 struct g_consumer *cp;
2612 static struct g_geom *gp;
2619 pp = g_dev_getprovider(swh->dev);
2621 swh->error = ENODEV;
2624 mtx_lock(&sw_dev_mtx);
2625 TAILQ_FOREACH(sp, &swtailq, sw_list) {
2627 if (cp != NULL && cp->provider == pp) {
2628 mtx_unlock(&sw_dev_mtx);
2633 mtx_unlock(&sw_dev_mtx);
2635 gp = g_new_geomf(&g_swap_class, "swap");
2636 cp = g_new_consumer(gp);
2639 * XXX: Everytime you think you can improve the margin for
2640 * footshooting, somebody depends on the ability to do so:
2641 * savecore(8) wants to write to our swapdev so we cannot
2642 * set an exclusive count :-(
2644 error = g_access(cp, 1, 1, 0);
2647 g_destroy_consumer(cp);
2651 nblks = pp->mediasize / DEV_BSIZE;
2652 swaponsomething(swh->vp, cp, nblks, swapgeom_strategy,
2653 swapgeom_close, dev2udev(swh->dev),
2654 (pp->flags & G_PF_ACCEPT_UNMAPPED) != 0 ? SW_UNMAPPED : 0);
2659 swapongeom(struct thread *td, struct vnode *vp)
2664 vn_lock(vp, LK_EXCLUSIVE | LK_RETRY);
2666 swh.dev = vp->v_rdev;
2669 /* XXX: direct call when Giant untangled */
2670 error = g_waitfor_event(swapongeom_ev, &swh, M_WAITOK, NULL);
2680 * This is used mainly for network filesystem (read: probably only tested
2681 * with NFS) swapfiles.
2686 swapdev_strategy(struct buf *bp, struct swdevt *sp)
2690 bp->b_blkno = ctodb(bp->b_blkno - sp->sw_first);
2694 if (bp->b_iocmd == BIO_WRITE) {
2696 bufobj_wdrop(bp->b_bufobj);
2697 bufobj_wref(&vp2->v_bufobj);
2699 if (bp->b_bufobj != &vp2->v_bufobj)
2700 bp->b_bufobj = &vp2->v_bufobj;
2702 bp->b_iooffset = dbtob(bp->b_blkno);
2708 swapdev_close(struct thread *td, struct swdevt *sp)
2711 VOP_CLOSE(sp->sw_vp, FREAD | FWRITE, td->td_ucred, td);
2717 swaponvp(struct thread *td, struct vnode *vp, u_long nblks)
2724 mtx_lock(&sw_dev_mtx);
2725 TAILQ_FOREACH(sp, &swtailq, sw_list) {
2726 if (sp->sw_id == vp) {
2727 mtx_unlock(&sw_dev_mtx);
2731 mtx_unlock(&sw_dev_mtx);
2733 (void) vn_lock(vp, LK_EXCLUSIVE | LK_RETRY);
2735 error = mac_system_check_swapon(td->td_ucred, vp);
2738 error = VOP_OPEN(vp, FREAD | FWRITE, td->td_ucred, td, NULL);
2739 (void) VOP_UNLOCK(vp, 0);
2743 swaponsomething(vp, vp, nblks, swapdev_strategy, swapdev_close,