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/rwlock.h>
93 #include <sys/sysctl.h>
94 #include <sys/sysproto.h>
95 #include <sys/blist.h>
98 #include <sys/vmmeter.h>
100 #include <security/mac/mac_framework.h>
104 #include <vm/vm_map.h>
105 #include <vm/vm_kern.h>
106 #include <vm/vm_object.h>
107 #include <vm/vm_page.h>
108 #include <vm/vm_pager.h>
109 #include <vm/vm_pageout.h>
110 #include <vm/vm_param.h>
111 #include <vm/swap_pager.h>
112 #include <vm/vm_extern.h>
115 #include <geom/geom.h>
118 * MAX_PAGEOUT_CLUSTER must be a power of 2 between 1 and 64.
119 * The 64-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) "
166 static unsigned long swzone;
167 SYSCTL_ULONG(_vm, OID_AUTO, swzone, CTLFLAG_RD, &swzone, 0,
168 "Actual size of swap metadata zone");
169 static unsigned long swap_maxpages;
170 SYSCTL_ULONG(_vm, OID_AUTO, swap_maxpages, CTLFLAG_RD, &swap_maxpages, 0,
171 "Maximum amount of swap supported");
173 /* bits from overcommit */
174 #define SWAP_RESERVE_FORCE_ON (1 << 0)
175 #define SWAP_RESERVE_RLIMIT_ON (1 << 1)
176 #define SWAP_RESERVE_ALLOW_NONWIRED (1 << 2)
179 swap_reserve(vm_ooffset_t incr)
182 return (swap_reserve_by_cred(incr, curthread->td_ucred));
186 swap_reserve_by_cred(vm_ooffset_t incr, struct ucred *cred)
191 static struct timeval lastfail;
194 uip = cred->cr_ruidinfo;
196 if (incr & PAGE_MASK)
197 panic("swap_reserve: & PAGE_MASK");
202 error = racct_add(curproc, RACCT_SWAP, incr);
203 PROC_UNLOCK(curproc);
210 mtx_lock(&sw_dev_mtx);
211 r = swap_reserved + incr;
212 if (overcommit & SWAP_RESERVE_ALLOW_NONWIRED) {
213 s = cnt.v_page_count - cnt.v_free_reserved - cnt.v_wire_count;
218 if ((overcommit & SWAP_RESERVE_FORCE_ON) == 0 || r <= s ||
219 (error = priv_check(curthread, PRIV_VM_SWAP_NOQUOTA)) == 0) {
223 mtx_unlock(&sw_dev_mtx);
227 UIDINFO_VMSIZE_LOCK(uip);
228 if ((overcommit & SWAP_RESERVE_RLIMIT_ON) != 0 &&
229 uip->ui_vmsize + incr > lim_cur(curproc, RLIMIT_SWAP) &&
230 priv_check(curthread, PRIV_VM_SWAP_NORLIMIT))
233 uip->ui_vmsize += incr;
234 UIDINFO_VMSIZE_UNLOCK(uip);
235 PROC_UNLOCK(curproc);
237 mtx_lock(&sw_dev_mtx);
238 swap_reserved -= incr;
239 mtx_unlock(&sw_dev_mtx);
242 if (!res && ppsratecheck(&lastfail, &curfail, 1)) {
243 printf("uid %d, pid %d: swap reservation for %jd bytes failed\n",
244 uip->ui_uid, curproc->p_pid, incr);
250 racct_sub(curproc, RACCT_SWAP, incr);
251 PROC_UNLOCK(curproc);
259 swap_reserve_force(vm_ooffset_t incr)
263 mtx_lock(&sw_dev_mtx);
264 swap_reserved += incr;
265 mtx_unlock(&sw_dev_mtx);
269 racct_add_force(curproc, RACCT_SWAP, incr);
270 PROC_UNLOCK(curproc);
273 uip = curthread->td_ucred->cr_ruidinfo;
275 UIDINFO_VMSIZE_LOCK(uip);
276 uip->ui_vmsize += incr;
277 UIDINFO_VMSIZE_UNLOCK(uip);
278 PROC_UNLOCK(curproc);
282 swap_release(vm_ooffset_t decr)
287 cred = curthread->td_ucred;
288 swap_release_by_cred(decr, cred);
289 PROC_UNLOCK(curproc);
293 swap_release_by_cred(vm_ooffset_t decr, struct ucred *cred)
297 uip = cred->cr_ruidinfo;
299 if (decr & PAGE_MASK)
300 panic("swap_release: & PAGE_MASK");
302 mtx_lock(&sw_dev_mtx);
303 if (swap_reserved < decr)
304 panic("swap_reserved < decr");
305 swap_reserved -= decr;
306 mtx_unlock(&sw_dev_mtx);
308 UIDINFO_VMSIZE_LOCK(uip);
309 if (uip->ui_vmsize < decr)
310 printf("negative vmsize for uid = %d\n", uip->ui_uid);
311 uip->ui_vmsize -= decr;
312 UIDINFO_VMSIZE_UNLOCK(uip);
314 racct_sub_cred(cred, RACCT_SWAP, decr);
317 static void swapdev_strategy(struct buf *, struct swdevt *sw);
319 #define SWM_FREE 0x02 /* free, period */
320 #define SWM_POP 0x04 /* pop out */
322 int swap_pager_full = 2; /* swap space exhaustion (task killing) */
323 static int swap_pager_almost_full = 1; /* swap space exhaustion (w/hysteresis)*/
324 static int nsw_rcount; /* free read buffers */
325 static int nsw_wcount_sync; /* limit write buffers / synchronous */
326 static int nsw_wcount_async; /* limit write buffers / asynchronous */
327 static int nsw_wcount_async_max;/* assigned maximum */
328 static int nsw_cluster_max; /* maximum VOP I/O allowed */
330 static struct swblock **swhash;
331 static int swhash_mask;
332 static struct mtx swhash_mtx;
334 static int swap_async_max = 4; /* maximum in-progress async I/O's */
335 static struct sx sw_alloc_sx;
338 SYSCTL_INT(_vm, OID_AUTO, swap_async_max,
339 CTLFLAG_RW, &swap_async_max, 0, "Maximum running async swap ops");
342 * "named" and "unnamed" anon region objects. Try to reduce the overhead
343 * of searching a named list by hashing it just a little.
348 #define NOBJLIST(handle) \
349 (&swap_pager_object_list[((int)(intptr_t)handle >> 4) & (NOBJLISTS-1)])
351 static struct mtx sw_alloc_mtx; /* protect list manipulation */
352 static struct pagerlst swap_pager_object_list[NOBJLISTS];
353 static uma_zone_t swap_zone;
356 * pagerops for OBJT_SWAP - "swap pager". Some ops are also global procedure
357 * calls hooked from other parts of the VM system and do not appear here.
358 * (see vm/swap_pager.h).
361 swap_pager_alloc(void *handle, vm_ooffset_t size,
362 vm_prot_t prot, vm_ooffset_t offset, struct ucred *);
363 static void swap_pager_dealloc(vm_object_t object);
364 static int swap_pager_getpages(vm_object_t, vm_page_t *, int, int);
365 static void swap_pager_putpages(vm_object_t, vm_page_t *, int, boolean_t, int *);
367 swap_pager_haspage(vm_object_t object, vm_pindex_t pindex, int *before, int *after);
368 static void swap_pager_init(void);
369 static void swap_pager_unswapped(vm_page_t);
370 static void swap_pager_swapoff(struct swdevt *sp);
372 struct pagerops swappagerops = {
373 .pgo_init = swap_pager_init, /* early system initialization of pager */
374 .pgo_alloc = swap_pager_alloc, /* allocate an OBJT_SWAP object */
375 .pgo_dealloc = swap_pager_dealloc, /* deallocate an OBJT_SWAP object */
376 .pgo_getpages = swap_pager_getpages, /* pagein */
377 .pgo_putpages = swap_pager_putpages, /* pageout */
378 .pgo_haspage = swap_pager_haspage, /* get backing store status for page */
379 .pgo_pageunswapped = swap_pager_unswapped, /* remove swap related to page */
383 * swap_*() routines are externally accessible. swp_*() routines are
386 static int nswap_lowat = 128; /* in pages, swap_pager_almost_full warn */
387 static int nswap_hiwat = 512; /* in pages, swap_pager_almost_full warn */
389 SYSCTL_INT(_vm, OID_AUTO, dmmax, CTLFLAG_RD, &nsw_cluster_max, 0,
390 "Maximum size of a swap block in pages");
392 static void swp_sizecheck(void);
393 static void swp_pager_async_iodone(struct buf *bp);
394 static int swapongeom(struct thread *, struct vnode *);
395 static int swaponvp(struct thread *, struct vnode *, u_long);
396 static int swapoff_one(struct swdevt *sp, struct ucred *cred);
399 * Swap bitmap functions
401 static void swp_pager_freeswapspace(daddr_t blk, int npages);
402 static daddr_t swp_pager_getswapspace(int npages);
407 static struct swblock **swp_pager_hash(vm_object_t object, vm_pindex_t index);
408 static void swp_pager_meta_build(vm_object_t, vm_pindex_t, daddr_t);
409 static void swp_pager_meta_free(vm_object_t, vm_pindex_t, daddr_t);
410 static void swp_pager_meta_free_all(vm_object_t);
411 static daddr_t swp_pager_meta_ctl(vm_object_t, vm_pindex_t, int);
414 swp_pager_free_nrpage(vm_page_t m)
418 if (m->wire_count == 0)
424 * SWP_SIZECHECK() - update swap_pager_full indication
426 * update the swap_pager_almost_full indication and warn when we are
427 * about to run out of swap space, using lowat/hiwat hysteresis.
429 * Clear swap_pager_full ( task killing ) indication when lowat is met.
431 * No restrictions on call
432 * This routine may not block.
438 if (swap_pager_avail < nswap_lowat) {
439 if (swap_pager_almost_full == 0) {
440 printf("swap_pager: out of swap space\n");
441 swap_pager_almost_full = 1;
445 if (swap_pager_avail > nswap_hiwat)
446 swap_pager_almost_full = 0;
451 * SWP_PAGER_HASH() - hash swap meta data
453 * This is an helper function which hashes the swapblk given
454 * the object and page index. It returns a pointer to a pointer
455 * to the object, or a pointer to a NULL pointer if it could not
458 static struct swblock **
459 swp_pager_hash(vm_object_t object, vm_pindex_t index)
461 struct swblock **pswap;
462 struct swblock *swap;
464 index &= ~(vm_pindex_t)SWAP_META_MASK;
465 pswap = &swhash[(index ^ (int)(intptr_t)object) & swhash_mask];
466 while ((swap = *pswap) != NULL) {
467 if (swap->swb_object == object &&
468 swap->swb_index == index
472 pswap = &swap->swb_hnext;
478 * SWAP_PAGER_INIT() - initialize the swap pager!
480 * Expected to be started from system init. NOTE: This code is run
481 * before much else so be careful what you depend on. Most of the VM
482 * system has yet to be initialized at this point.
485 swap_pager_init(void)
488 * Initialize object lists
492 for (i = 0; i < NOBJLISTS; ++i)
493 TAILQ_INIT(&swap_pager_object_list[i]);
494 mtx_init(&sw_alloc_mtx, "swap_pager list", NULL, MTX_DEF);
495 mtx_init(&sw_dev_mtx, "swapdev", NULL, MTX_DEF);
496 sx_init(&sw_alloc_sx, "swspsx");
500 * SWAP_PAGER_SWAP_INIT() - swap pager initialization from pageout process
502 * Expected to be started from pageout process once, prior to entering
506 swap_pager_swap_init(void)
511 * Number of in-transit swap bp operations. Don't
512 * exhaust the pbufs completely. Make sure we
513 * initialize workable values (0 will work for hysteresis
514 * but it isn't very efficient).
516 * The nsw_cluster_max is constrained by the bp->b_pages[]
517 * array (MAXPHYS/PAGE_SIZE) and our locally defined
518 * MAX_PAGEOUT_CLUSTER. Also be aware that swap ops are
519 * constrained by the swap device interleave stripe size.
521 * Currently we hardwire nsw_wcount_async to 4. This limit is
522 * designed to prevent other I/O from having high latencies due to
523 * our pageout I/O. The value 4 works well for one or two active swap
524 * devices but is probably a little low if you have more. Even so,
525 * a higher value would probably generate only a limited improvement
526 * with three or four active swap devices since the system does not
527 * typically have to pageout at extreme bandwidths. We will want
528 * at least 2 per swap devices, and 4 is a pretty good value if you
529 * have one NFS swap device due to the command/ack latency over NFS.
530 * So it all works out pretty well.
532 nsw_cluster_max = min((MAXPHYS/PAGE_SIZE), MAX_PAGEOUT_CLUSTER);
535 nsw_rcount = (nswbuf + 1) / 2;
536 nsw_wcount_sync = (nswbuf + 3) / 4;
537 nsw_wcount_async = 4;
538 nsw_wcount_async_max = nsw_wcount_async;
539 mtx_unlock(&pbuf_mtx);
542 * Initialize our zone. Right now I'm just guessing on the number
543 * we need based on the number of pages in the system. Each swblock
544 * can hold 32 pages, so this is probably overkill. This reservation
545 * is typically limited to around 32MB by default.
547 n = cnt.v_page_count / 2;
548 if (maxswzone && n > maxswzone / sizeof(struct swblock))
549 n = maxswzone / sizeof(struct swblock);
551 swap_zone = uma_zcreate("SWAPMETA", sizeof(struct swblock), NULL, NULL,
552 NULL, NULL, UMA_ALIGN_PTR, UMA_ZONE_NOFREE | UMA_ZONE_VM);
553 if (swap_zone == NULL)
554 panic("failed to create swap_zone.");
556 if (uma_zone_reserve_kva(swap_zone, n))
559 * if the allocation failed, try a zone two thirds the
560 * size of the previous attempt.
565 printf("Swap zone entries reduced from %lu to %lu.\n", n2, n);
566 swap_maxpages = n * SWAP_META_PAGES;
567 swzone = n * sizeof(struct swblock);
571 * Initialize our meta-data hash table. The swapper does not need to
572 * be quite as efficient as the VM system, so we do not use an
573 * oversized hash table.
575 * n: size of hash table, must be power of 2
576 * swhash_mask: hash table index mask
578 for (n = 1; n < n2 / 8; n *= 2)
580 swhash = malloc(sizeof(struct swblock *) * n, M_VMPGDATA, M_WAITOK | M_ZERO);
582 mtx_init(&swhash_mtx, "swap_pager swhash", NULL, MTX_DEF);
586 * SWAP_PAGER_ALLOC() - allocate a new OBJT_SWAP VM object and instantiate
587 * its metadata structures.
589 * This routine is called from the mmap and fork code to create a new
590 * OBJT_SWAP object. We do this by creating an OBJT_DEFAULT object
591 * and then converting it with swp_pager_meta_build().
593 * This routine may block in vm_object_allocate() and create a named
594 * object lookup race, so we must interlock.
599 swap_pager_alloc(void *handle, vm_ooffset_t size, vm_prot_t prot,
600 vm_ooffset_t offset, struct ucred *cred)
605 pindex = OFF_TO_IDX(offset + PAGE_MASK + size);
609 * Reference existing named region or allocate new one. There
610 * should not be a race here against swp_pager_meta_build()
611 * as called from vm_page_remove() in regards to the lookup
614 sx_xlock(&sw_alloc_sx);
615 object = vm_pager_object_lookup(NOBJLIST(handle), handle);
616 if (object == NULL) {
618 if (!swap_reserve_by_cred(size, cred)) {
619 sx_xunlock(&sw_alloc_sx);
625 object = vm_object_allocate(OBJT_DEFAULT, pindex);
626 VM_OBJECT_WLOCK(object);
627 object->handle = handle;
630 object->charge = size;
632 swp_pager_meta_build(object, 0, SWAPBLK_NONE);
633 VM_OBJECT_WUNLOCK(object);
635 sx_xunlock(&sw_alloc_sx);
639 if (!swap_reserve_by_cred(size, cred))
643 object = vm_object_allocate(OBJT_DEFAULT, pindex);
644 VM_OBJECT_WLOCK(object);
647 object->charge = size;
649 swp_pager_meta_build(object, 0, SWAPBLK_NONE);
650 VM_OBJECT_WUNLOCK(object);
656 * SWAP_PAGER_DEALLOC() - remove swap metadata from object
658 * The swap backing for the object is destroyed. The code is
659 * designed such that we can reinstantiate it later, but this
660 * routine is typically called only when the entire object is
661 * about to be destroyed.
663 * The object must be locked.
666 swap_pager_dealloc(vm_object_t object)
670 * Remove from list right away so lookups will fail if we block for
671 * pageout completion.
673 if (object->handle != NULL) {
674 mtx_lock(&sw_alloc_mtx);
675 TAILQ_REMOVE(NOBJLIST(object->handle), object, pager_object_list);
676 mtx_unlock(&sw_alloc_mtx);
679 VM_OBJECT_ASSERT_WLOCKED(object);
680 vm_object_pip_wait(object, "swpdea");
683 * Free all remaining metadata. We only bother to free it from
684 * the swap meta data. We do not attempt to free swapblk's still
685 * associated with vm_page_t's for this object. We do not care
686 * if paging is still in progress on some objects.
688 swp_pager_meta_free_all(object);
689 object->handle = NULL;
690 object->type = OBJT_DEAD;
693 /************************************************************************
694 * SWAP PAGER BITMAP ROUTINES *
695 ************************************************************************/
698 * SWP_PAGER_GETSWAPSPACE() - allocate raw swap space
700 * Allocate swap for the requested number of pages. The starting
701 * swap block number (a page index) is returned or SWAPBLK_NONE
702 * if the allocation failed.
704 * Also has the side effect of advising that somebody made a mistake
705 * when they configured swap and didn't configure enough.
707 * This routine may not sleep.
709 * We allocate in round-robin fashion from the configured devices.
712 swp_pager_getswapspace(int npages)
719 mtx_lock(&sw_dev_mtx);
721 for (i = 0; i < nswapdev; i++) {
723 sp = TAILQ_FIRST(&swtailq);
724 if (!(sp->sw_flags & SW_CLOSING)) {
725 blk = blist_alloc(sp->sw_blist, npages);
726 if (blk != SWAPBLK_NONE) {
728 sp->sw_used += npages;
729 swap_pager_avail -= npages;
731 swdevhd = TAILQ_NEXT(sp, sw_list);
735 sp = TAILQ_NEXT(sp, sw_list);
737 if (swap_pager_full != 2) {
738 printf("swap_pager_getswapspace(%d): failed\n", npages);
740 swap_pager_almost_full = 1;
744 mtx_unlock(&sw_dev_mtx);
749 swp_pager_isondev(daddr_t blk, struct swdevt *sp)
752 return (blk >= sp->sw_first && blk < sp->sw_end);
756 swp_pager_strategy(struct buf *bp)
760 mtx_lock(&sw_dev_mtx);
761 TAILQ_FOREACH(sp, &swtailq, sw_list) {
762 if (bp->b_blkno >= sp->sw_first && bp->b_blkno < sp->sw_end) {
763 mtx_unlock(&sw_dev_mtx);
764 if ((sp->sw_flags & SW_UNMAPPED) != 0 &&
765 unmapped_buf_allowed) {
766 bp->b_kvaalloc = bp->b_data;
767 bp->b_data = unmapped_buf;
768 bp->b_kvabase = unmapped_buf;
770 bp->b_flags |= B_UNMAPPED;
772 pmap_qenter((vm_offset_t)bp->b_data,
773 &bp->b_pages[0], bp->b_bcount / PAGE_SIZE);
775 sp->sw_strategy(bp, sp);
779 panic("Swapdev not found");
784 * SWP_PAGER_FREESWAPSPACE() - free raw swap space
786 * This routine returns the specified swap blocks back to the bitmap.
788 * This routine may not sleep.
791 swp_pager_freeswapspace(daddr_t blk, int npages)
795 mtx_lock(&sw_dev_mtx);
796 TAILQ_FOREACH(sp, &swtailq, sw_list) {
797 if (blk >= sp->sw_first && blk < sp->sw_end) {
798 sp->sw_used -= npages;
800 * If we are attempting to stop swapping on
801 * this device, we don't want to mark any
802 * blocks free lest they be reused.
804 if ((sp->sw_flags & SW_CLOSING) == 0) {
805 blist_free(sp->sw_blist, blk - sp->sw_first,
807 swap_pager_avail += npages;
810 mtx_unlock(&sw_dev_mtx);
814 panic("Swapdev not found");
818 * SWAP_PAGER_FREESPACE() - frees swap blocks associated with a page
819 * range within an object.
821 * This is a globally accessible routine.
823 * This routine removes swapblk assignments from swap metadata.
825 * The external callers of this routine typically have already destroyed
826 * or renamed vm_page_t's associated with this range in the object so
829 * The object must be locked.
832 swap_pager_freespace(vm_object_t object, vm_pindex_t start, vm_size_t size)
835 swp_pager_meta_free(object, start, size);
839 * SWAP_PAGER_RESERVE() - reserve swap blocks in object
841 * Assigns swap blocks to the specified range within the object. The
842 * swap blocks are not zeroed. Any previous swap assignment is destroyed.
844 * Returns 0 on success, -1 on failure.
847 swap_pager_reserve(vm_object_t object, vm_pindex_t start, vm_size_t size)
850 daddr_t blk = SWAPBLK_NONE;
851 vm_pindex_t beg = start; /* save start index */
853 VM_OBJECT_WLOCK(object);
857 while ((blk = swp_pager_getswapspace(n)) == SWAPBLK_NONE) {
860 swp_pager_meta_free(object, beg, start - beg);
861 VM_OBJECT_WUNLOCK(object);
866 swp_pager_meta_build(object, start, blk);
872 swp_pager_meta_free(object, start, n);
873 VM_OBJECT_WUNLOCK(object);
878 * SWAP_PAGER_COPY() - copy blocks from source pager to destination pager
879 * and destroy the source.
881 * Copy any valid swapblks from the source to the destination. In
882 * cases where both the source and destination have a valid swapblk,
883 * we keep the destination's.
885 * This routine is allowed to sleep. It may sleep allocating metadata
886 * indirectly through swp_pager_meta_build() or if paging is still in
887 * progress on the source.
889 * The source object contains no vm_page_t's (which is just as well)
891 * The source object is of type OBJT_SWAP.
893 * The source and destination objects must be locked.
894 * Both object locks may temporarily be released.
897 swap_pager_copy(vm_object_t srcobject, vm_object_t dstobject,
898 vm_pindex_t offset, int destroysource)
902 VM_OBJECT_ASSERT_WLOCKED(srcobject);
903 VM_OBJECT_ASSERT_WLOCKED(dstobject);
906 * If destroysource is set, we remove the source object from the
907 * swap_pager internal queue now.
910 if (srcobject->handle != NULL) {
911 mtx_lock(&sw_alloc_mtx);
913 NOBJLIST(srcobject->handle),
917 mtx_unlock(&sw_alloc_mtx);
922 * transfer source to destination.
924 for (i = 0; i < dstobject->size; ++i) {
928 * Locate (without changing) the swapblk on the destination,
929 * unless it is invalid in which case free it silently, or
930 * if the destination is a resident page, in which case the
931 * source is thrown away.
933 dstaddr = swp_pager_meta_ctl(dstobject, i, 0);
935 if (dstaddr == SWAPBLK_NONE) {
937 * Destination has no swapblk and is not resident,
942 srcaddr = swp_pager_meta_ctl(
948 if (srcaddr != SWAPBLK_NONE) {
950 * swp_pager_meta_build() can sleep.
952 vm_object_pip_add(srcobject, 1);
953 VM_OBJECT_WUNLOCK(srcobject);
954 vm_object_pip_add(dstobject, 1);
955 swp_pager_meta_build(dstobject, i, srcaddr);
956 vm_object_pip_wakeup(dstobject);
957 VM_OBJECT_WLOCK(srcobject);
958 vm_object_pip_wakeup(srcobject);
962 * Destination has valid swapblk or it is represented
963 * by a resident page. We destroy the sourceblock.
966 swp_pager_meta_ctl(srcobject, i + offset, SWM_FREE);
971 * Free left over swap blocks in source.
973 * We have to revert the type to OBJT_DEFAULT so we do not accidently
974 * double-remove the object from the swap queues.
977 swp_pager_meta_free_all(srcobject);
979 * Reverting the type is not necessary, the caller is going
980 * to destroy srcobject directly, but I'm doing it here
981 * for consistency since we've removed the object from its
984 srcobject->type = OBJT_DEFAULT;
989 * SWAP_PAGER_HASPAGE() - determine if we have good backing store for
990 * the requested page.
992 * We determine whether good backing store exists for the requested
993 * page and return TRUE if it does, FALSE if it doesn't.
995 * If TRUE, we also try to determine how much valid, contiguous backing
996 * store exists before and after the requested page within a reasonable
997 * distance. We do not try to restrict it to the swap device stripe
998 * (that is handled in getpages/putpages). It probably isn't worth
1002 swap_pager_haspage(vm_object_t object, vm_pindex_t pindex, int *before, int *after)
1006 VM_OBJECT_ASSERT_LOCKED(object);
1008 * do we have good backing store at the requested index ?
1010 blk0 = swp_pager_meta_ctl(object, pindex, 0);
1012 if (blk0 == SWAPBLK_NONE) {
1021 * find backwards-looking contiguous good backing store
1023 if (before != NULL) {
1026 for (i = 1; i < (SWB_NPAGES/2); ++i) {
1031 blk = swp_pager_meta_ctl(object, pindex - i, 0);
1032 if (blk != blk0 - i)
1039 * find forward-looking contiguous good backing store
1041 if (after != NULL) {
1044 for (i = 1; i < (SWB_NPAGES/2); ++i) {
1047 blk = swp_pager_meta_ctl(object, pindex + i, 0);
1048 if (blk != blk0 + i)
1057 * SWAP_PAGER_PAGE_UNSWAPPED() - remove swap backing store related to page
1059 * This removes any associated swap backing store, whether valid or
1060 * not, from the page.
1062 * This routine is typically called when a page is made dirty, at
1063 * which point any associated swap can be freed. MADV_FREE also
1064 * calls us in a special-case situation
1066 * NOTE!!! If the page is clean and the swap was valid, the caller
1067 * should make the page dirty before calling this routine. This routine
1068 * does NOT change the m->dirty status of the page. Also: MADV_FREE
1071 * This routine may not sleep.
1073 * The object containing the page must be locked.
1076 swap_pager_unswapped(vm_page_t m)
1079 swp_pager_meta_ctl(m->object, m->pindex, SWM_FREE);
1083 * SWAP_PAGER_GETPAGES() - bring pages in from swap
1085 * Attempt to retrieve (m, count) pages from backing store, but make
1086 * sure we retrieve at least m[reqpage]. We try to load in as large
1087 * a chunk surrounding m[reqpage] as is contiguous in swap and which
1088 * belongs to the same object.
1090 * The code is designed for asynchronous operation and
1091 * immediate-notification of 'reqpage' but tends not to be
1092 * used that way. Please do not optimize-out this algorithmic
1093 * feature, I intend to improve on it in the future.
1095 * The parent has a single vm_object_pip_add() reference prior to
1096 * calling us and we should return with the same.
1098 * The parent has BUSY'd the pages. We should return with 'm'
1099 * left busy, but the others adjusted.
1102 swap_pager_getpages(vm_object_t object, vm_page_t *m, int count, int reqpage)
1112 KASSERT(mreq->object == object,
1113 ("swap_pager_getpages: object mismatch %p/%p",
1114 object, mreq->object));
1117 * Calculate range to retrieve. The pages have already been assigned
1118 * their swapblks. We require a *contiguous* range but we know it to
1119 * not span devices. If we do not supply it, bad things
1120 * happen. Note that blk, iblk & jblk can be SWAPBLK_NONE, but the
1121 * loops are set up such that the case(s) are handled implicitly.
1123 * The swp_*() calls must be made with the object locked.
1125 blk = swp_pager_meta_ctl(mreq->object, mreq->pindex, 0);
1127 for (i = reqpage - 1; i >= 0; --i) {
1130 iblk = swp_pager_meta_ctl(m[i]->object, m[i]->pindex, 0);
1131 if (blk != iblk + (reqpage - i))
1136 for (j = reqpage + 1; j < count; ++j) {
1139 jblk = swp_pager_meta_ctl(m[j]->object, m[j]->pindex, 0);
1140 if (blk != jblk - (j - reqpage))
1145 * free pages outside our collection range. Note: we never free
1146 * mreq, it must remain busy throughout.
1148 if (0 < i || j < count) {
1151 for (k = 0; k < i; ++k)
1152 swp_pager_free_nrpage(m[k]);
1153 for (k = j; k < count; ++k)
1154 swp_pager_free_nrpage(m[k]);
1158 * Return VM_PAGER_FAIL if we have nothing to do. Return mreq
1159 * still busy, but the others unbusied.
1161 if (blk == SWAPBLK_NONE)
1162 return (VM_PAGER_FAIL);
1165 * Getpbuf() can sleep.
1167 VM_OBJECT_WUNLOCK(object);
1169 * Get a swap buffer header to perform the IO
1171 bp = getpbuf(&nsw_rcount);
1172 bp->b_flags |= B_PAGING;
1174 bp->b_iocmd = BIO_READ;
1175 bp->b_iodone = swp_pager_async_iodone;
1176 bp->b_rcred = crhold(thread0.td_ucred);
1177 bp->b_wcred = crhold(thread0.td_ucred);
1178 bp->b_blkno = blk - (reqpage - i);
1179 bp->b_bcount = PAGE_SIZE * (j - i);
1180 bp->b_bufsize = PAGE_SIZE * (j - i);
1181 bp->b_pager.pg_reqpage = reqpage - i;
1183 VM_OBJECT_WLOCK(object);
1187 for (k = i; k < j; ++k) {
1188 bp->b_pages[k - i] = m[k];
1189 m[k]->oflags |= VPO_SWAPINPROG;
1192 bp->b_npages = j - i;
1194 PCPU_INC(cnt.v_swapin);
1195 PCPU_ADD(cnt.v_swappgsin, bp->b_npages);
1198 * We still hold the lock on mreq, and our automatic completion routine
1199 * does not remove it.
1201 vm_object_pip_add(object, bp->b_npages);
1202 VM_OBJECT_WUNLOCK(object);
1205 * perform the I/O. NOTE!!! bp cannot be considered valid after
1206 * this point because we automatically release it on completion.
1207 * Instead, we look at the one page we are interested in which we
1208 * still hold a lock on even through the I/O completion.
1210 * The other pages in our m[] array are also released on completion,
1211 * so we cannot assume they are valid anymore either.
1213 * NOTE: b_blkno is destroyed by the call to swapdev_strategy
1216 swp_pager_strategy(bp);
1219 * wait for the page we want to complete. VPO_SWAPINPROG is always
1220 * cleared on completion. If an I/O error occurs, SWAPBLK_NONE
1221 * is set in the meta-data.
1223 VM_OBJECT_WLOCK(object);
1224 while ((mreq->oflags & VPO_SWAPINPROG) != 0) {
1225 mreq->oflags |= VPO_SWAPSLEEP;
1226 PCPU_INC(cnt.v_intrans);
1227 if (VM_OBJECT_SLEEP(object, &object->paging_in_progress, PSWP,
1228 "swread", hz * 20)) {
1230 "swap_pager: indefinite wait buffer: bufobj: %p, blkno: %jd, size: %ld\n",
1231 bp->b_bufobj, (intmax_t)bp->b_blkno, bp->b_bcount);
1236 * mreq is left busied after completion, but all the other pages
1237 * are freed. If we had an unrecoverable read error the page will
1240 if (mreq->valid != VM_PAGE_BITS_ALL) {
1241 return (VM_PAGER_ERROR);
1243 return (VM_PAGER_OK);
1247 * A final note: in a low swap situation, we cannot deallocate swap
1248 * and mark a page dirty here because the caller is likely to mark
1249 * the page clean when we return, causing the page to possibly revert
1250 * to all-zero's later.
1255 * swap_pager_putpages:
1257 * Assign swap (if necessary) and initiate I/O on the specified pages.
1259 * We support both OBJT_DEFAULT and OBJT_SWAP objects. DEFAULT objects
1260 * are automatically converted to SWAP objects.
1262 * In a low memory situation we may block in VOP_STRATEGY(), but the new
1263 * vm_page reservation system coupled with properly written VFS devices
1264 * should ensure that no low-memory deadlock occurs. This is an area
1267 * The parent has N vm_object_pip_add() references prior to
1268 * calling us and will remove references for rtvals[] that are
1269 * not set to VM_PAGER_PEND. We need to remove the rest on I/O
1272 * The parent has soft-busy'd the pages it passes us and will unbusy
1273 * those whos rtvals[] entry is not set to VM_PAGER_PEND on return.
1274 * We need to unbusy the rest on I/O completion.
1277 swap_pager_putpages(vm_object_t object, vm_page_t *m, int count,
1278 int flags, int *rtvals)
1283 if (count && m[0]->object != object) {
1284 panic("swap_pager_putpages: object mismatch %p/%p",
1293 * Turn object into OBJT_SWAP
1294 * check for bogus sysops
1295 * force sync if not pageout process
1297 if (object->type != OBJT_SWAP)
1298 swp_pager_meta_build(object, 0, SWAPBLK_NONE);
1299 VM_OBJECT_WUNLOCK(object);
1302 if (curproc != pageproc)
1305 sync = (flags & VM_PAGER_PUT_SYNC) != 0;
1310 * Update nsw parameters from swap_async_max sysctl values.
1311 * Do not let the sysop crash the machine with bogus numbers.
1313 mtx_lock(&pbuf_mtx);
1314 if (swap_async_max != nsw_wcount_async_max) {
1320 if ((n = swap_async_max) > nswbuf / 2)
1327 * Adjust difference ( if possible ). If the current async
1328 * count is too low, we may not be able to make the adjustment
1331 n -= nsw_wcount_async_max;
1332 if (nsw_wcount_async + n >= 0) {
1333 nsw_wcount_async += n;
1334 nsw_wcount_async_max += n;
1335 wakeup(&nsw_wcount_async);
1338 mtx_unlock(&pbuf_mtx);
1343 * Assign swap blocks and issue I/O. We reallocate swap on the fly.
1344 * The page is left dirty until the pageout operation completes
1347 for (i = 0; i < count; i += n) {
1353 * Maximum I/O size is limited by a number of factors.
1355 n = min(BLIST_MAX_ALLOC, count - i);
1356 n = min(n, nsw_cluster_max);
1359 * Get biggest block of swap we can. If we fail, fall
1360 * back and try to allocate a smaller block. Don't go
1361 * overboard trying to allocate space if it would overly
1365 (blk = swp_pager_getswapspace(n)) == SWAPBLK_NONE &&
1370 if (blk == SWAPBLK_NONE) {
1371 for (j = 0; j < n; ++j)
1372 rtvals[i+j] = VM_PAGER_FAIL;
1377 * All I/O parameters have been satisfied, build the I/O
1378 * request and assign the swap space.
1381 bp = getpbuf(&nsw_wcount_sync);
1383 bp = getpbuf(&nsw_wcount_async);
1384 bp->b_flags = B_ASYNC;
1386 bp->b_flags |= B_PAGING;
1387 bp->b_iocmd = BIO_WRITE;
1389 bp->b_rcred = crhold(thread0.td_ucred);
1390 bp->b_wcred = crhold(thread0.td_ucred);
1391 bp->b_bcount = PAGE_SIZE * n;
1392 bp->b_bufsize = PAGE_SIZE * n;
1395 VM_OBJECT_WLOCK(object);
1396 for (j = 0; j < n; ++j) {
1397 vm_page_t mreq = m[i+j];
1399 swp_pager_meta_build(
1404 vm_page_dirty(mreq);
1405 rtvals[i+j] = VM_PAGER_OK;
1407 mreq->oflags |= VPO_SWAPINPROG;
1408 bp->b_pages[j] = mreq;
1410 VM_OBJECT_WUNLOCK(object);
1413 * Must set dirty range for NFS to work.
1416 bp->b_dirtyend = bp->b_bcount;
1418 PCPU_INC(cnt.v_swapout);
1419 PCPU_ADD(cnt.v_swappgsout, bp->b_npages);
1424 * NOTE: b_blkno is destroyed by the call to swapdev_strategy
1426 if (sync == FALSE) {
1427 bp->b_iodone = swp_pager_async_iodone;
1429 swp_pager_strategy(bp);
1431 for (j = 0; j < n; ++j)
1432 rtvals[i+j] = VM_PAGER_PEND;
1433 /* restart outter loop */
1440 * NOTE: b_blkno is destroyed by the call to swapdev_strategy
1442 bp->b_iodone = bdone;
1443 swp_pager_strategy(bp);
1446 * Wait for the sync I/O to complete, then update rtvals.
1447 * We just set the rtvals[] to VM_PAGER_PEND so we can call
1448 * our async completion routine at the end, thus avoiding a
1451 bwait(bp, PVM, "swwrt");
1452 for (j = 0; j < n; ++j)
1453 rtvals[i+j] = VM_PAGER_PEND;
1455 * Now that we are through with the bp, we can call the
1456 * normal async completion, which frees everything up.
1458 swp_pager_async_iodone(bp);
1460 VM_OBJECT_WLOCK(object);
1464 * swp_pager_async_iodone:
1466 * Completion routine for asynchronous reads and writes from/to swap.
1467 * Also called manually by synchronous code to finish up a bp.
1469 * This routine may not sleep.
1472 swp_pager_async_iodone(struct buf *bp)
1475 vm_object_t object = NULL;
1480 if (bp->b_ioflags & BIO_ERROR) {
1482 "swap_pager: I/O error - %s failed; blkno %ld,"
1483 "size %ld, error %d\n",
1484 ((bp->b_iocmd == BIO_READ) ? "pagein" : "pageout"),
1492 * remove the mapping for kernel virtual
1494 if ((bp->b_flags & B_UNMAPPED) != 0) {
1495 bp->b_data = bp->b_kvaalloc;
1496 bp->b_kvabase = bp->b_kvaalloc;
1497 bp->b_flags &= ~B_UNMAPPED;
1499 pmap_qremove((vm_offset_t)bp->b_data, bp->b_npages);
1502 object = bp->b_pages[0]->object;
1503 VM_OBJECT_WLOCK(object);
1507 * cleanup pages. If an error occurs writing to swap, we are in
1508 * very serious trouble. If it happens to be a disk error, though,
1509 * we may be able to recover by reassigning the swap later on. So
1510 * in this case we remove the m->swapblk assignment for the page
1511 * but do not free it in the rlist. The errornous block(s) are thus
1512 * never reallocated as swap. Redirty the page and continue.
1514 for (i = 0; i < bp->b_npages; ++i) {
1515 vm_page_t m = bp->b_pages[i];
1517 m->oflags &= ~VPO_SWAPINPROG;
1518 if (m->oflags & VPO_SWAPSLEEP) {
1519 m->oflags &= ~VPO_SWAPSLEEP;
1520 wakeup(&object->paging_in_progress);
1523 if (bp->b_ioflags & BIO_ERROR) {
1525 * If an error occurs I'd love to throw the swapblk
1526 * away without freeing it back to swapspace, so it
1527 * can never be used again. But I can't from an
1530 if (bp->b_iocmd == BIO_READ) {
1532 * When reading, reqpage needs to stay
1533 * locked for the parent, but all other
1534 * pages can be freed. We still want to
1535 * wakeup the parent waiting on the page,
1536 * though. ( also: pg_reqpage can be -1 and
1537 * not match anything ).
1539 * We have to wake specifically requested pages
1540 * up too because we cleared VPO_SWAPINPROG and
1541 * someone may be waiting for that.
1543 * NOTE: for reads, m->dirty will probably
1544 * be overridden by the original caller of
1545 * getpages so don't play cute tricks here.
1548 if (i != bp->b_pager.pg_reqpage)
1549 swp_pager_free_nrpage(m);
1556 * If i == bp->b_pager.pg_reqpage, do not wake
1557 * the page up. The caller needs to.
1561 * If a write error occurs, reactivate page
1562 * so it doesn't clog the inactive list,
1563 * then finish the I/O.
1567 vm_page_activate(m);
1571 } else if (bp->b_iocmd == BIO_READ) {
1573 * NOTE: for reads, m->dirty will probably be
1574 * overridden by the original caller of getpages so
1575 * we cannot set them in order to free the underlying
1576 * swap in a low-swap situation. I don't think we'd
1577 * want to do that anyway, but it was an optimization
1578 * that existed in the old swapper for a time before
1579 * it got ripped out due to precisely this problem.
1581 * If not the requested page then deactivate it.
1583 * Note that the requested page, reqpage, is left
1584 * busied, but we still have to wake it up. The
1585 * other pages are released (unbusied) by
1586 * vm_page_xunbusy().
1588 KASSERT(!pmap_page_is_mapped(m),
1589 ("swp_pager_async_iodone: page %p is mapped", m));
1590 m->valid = VM_PAGE_BITS_ALL;
1591 KASSERT(m->dirty == 0,
1592 ("swp_pager_async_iodone: page %p is dirty", m));
1595 * We have to wake specifically requested pages
1596 * up too because we cleared VPO_SWAPINPROG and
1597 * could be waiting for it in getpages. However,
1598 * be sure to not unbusy getpages specifically
1599 * requested page - getpages expects it to be
1602 if (i != bp->b_pager.pg_reqpage) {
1604 vm_page_deactivate(m);
1614 * For write success, clear the dirty
1615 * status, then finish the I/O ( which decrements the
1616 * busy count and possibly wakes waiter's up ).
1618 KASSERT(!pmap_page_is_write_mapped(m),
1619 ("swp_pager_async_iodone: page %p is not write"
1623 if (vm_page_count_severe()) {
1625 vm_page_try_to_cache(m);
1632 * adjust pip. NOTE: the original parent may still have its own
1633 * pip refs on the object.
1635 if (object != NULL) {
1636 vm_object_pip_wakeupn(object, bp->b_npages);
1637 VM_OBJECT_WUNLOCK(object);
1641 * swapdev_strategy() manually sets b_vp and b_bufobj before calling
1642 * bstrategy(). Set them back to NULL now we're done with it, or we'll
1643 * trigger a KASSERT in relpbuf().
1647 bp->b_bufobj = NULL;
1650 * release the physical I/O buffer
1654 ((bp->b_iocmd == BIO_READ) ? &nsw_rcount :
1655 ((bp->b_flags & B_ASYNC) ?
1664 * swap_pager_isswapped:
1666 * Return 1 if at least one page in the given object is paged
1667 * out to the given swap device.
1669 * This routine may not sleep.
1672 swap_pager_isswapped(vm_object_t object, struct swdevt *sp)
1678 VM_OBJECT_ASSERT_WLOCKED(object);
1679 if (object->type != OBJT_SWAP)
1682 mtx_lock(&swhash_mtx);
1683 for (bcount = 0; bcount < object->un_pager.swp.swp_bcount; bcount++) {
1684 struct swblock *swap;
1686 if ((swap = *swp_pager_hash(object, index)) != NULL) {
1687 for (i = 0; i < SWAP_META_PAGES; ++i) {
1688 if (swp_pager_isondev(swap->swb_pages[i], sp)) {
1689 mtx_unlock(&swhash_mtx);
1694 index += SWAP_META_PAGES;
1696 mtx_unlock(&swhash_mtx);
1701 * SWP_PAGER_FORCE_PAGEIN() - force a swap block to be paged in
1703 * This routine dissociates the page at the given index within a
1704 * swap block from its backing store, paging it in if necessary.
1705 * If the page is paged in, it is placed in the inactive queue,
1706 * since it had its backing store ripped out from under it.
1707 * We also attempt to swap in all other pages in the swap block,
1708 * we only guarantee that the one at the specified index is
1711 * XXX - The code to page the whole block in doesn't work, so we
1712 * revert to the one-by-one behavior for now. Sigh.
1715 swp_pager_force_pagein(vm_object_t object, vm_pindex_t pindex)
1719 vm_object_pip_add(object, 1);
1720 m = vm_page_grab(object, pindex, VM_ALLOC_NORMAL);
1721 if (m->valid == VM_PAGE_BITS_ALL) {
1722 vm_object_pip_subtract(object, 1);
1725 vm_page_activate(m);
1728 vm_pager_page_unswapped(m);
1732 if (swap_pager_getpages(object, &m, 1, 0) != VM_PAGER_OK)
1733 panic("swap_pager_force_pagein: read from swap failed");/*XXX*/
1734 vm_object_pip_subtract(object, 1);
1737 vm_page_deactivate(m);
1740 vm_pager_page_unswapped(m);
1744 * swap_pager_swapoff:
1746 * Page in all of the pages that have been paged out to the
1747 * given device. The corresponding blocks in the bitmap must be
1748 * marked as allocated and the device must be flagged SW_CLOSING.
1749 * There may be no processes swapped out to the device.
1751 * This routine may block.
1754 swap_pager_swapoff(struct swdevt *sp)
1756 struct swblock *swap;
1757 vm_object_t locked_obj, object;
1766 mtx_lock(&swhash_mtx);
1767 for (i = 0; i <= swhash_mask; i++) { /* '<=' is correct here */
1769 for (swap = swhash[i]; swap != NULL; swap = swap->swb_hnext) {
1770 object = swap->swb_object;
1771 pindex = swap->swb_index;
1772 for (j = 0; j < SWAP_META_PAGES; ++j) {
1773 if (!swp_pager_isondev(swap->swb_pages[j], sp))
1775 if (locked_obj != object) {
1776 if (locked_obj != NULL)
1777 VM_OBJECT_WUNLOCK(locked_obj);
1778 locked_obj = object;
1779 if (!VM_OBJECT_TRYWLOCK(object)) {
1780 mtx_unlock(&swhash_mtx);
1781 /* Depends on type-stability. */
1782 VM_OBJECT_WLOCK(object);
1783 mtx_lock(&swhash_mtx);
1787 MPASS(locked_obj == object);
1788 mtx_unlock(&swhash_mtx);
1789 swp_pager_force_pagein(object, pindex + j);
1790 mtx_lock(&swhash_mtx);
1795 mtx_unlock(&swhash_mtx);
1796 if (locked_obj != NULL) {
1797 VM_OBJECT_WUNLOCK(locked_obj);
1802 * Objects may be locked or paging to the device being
1803 * removed, so we will miss their pages and need to
1804 * make another pass. We have marked this device as
1805 * SW_CLOSING, so the activity should finish soon.
1808 if (retries > 100) {
1809 panic("swapoff: failed to locate %d swap blocks",
1812 pause("swpoff", hz / 20);
1817 /************************************************************************
1819 ************************************************************************
1821 * These routines manipulate the swap metadata stored in the
1824 * Swap metadata is implemented with a global hash and not directly
1825 * linked into the object. Instead the object simply contains
1826 * appropriate tracking counters.
1830 * SWP_PAGER_META_BUILD() - add swap block to swap meta data for object
1832 * We first convert the object to a swap object if it is a default
1835 * The specified swapblk is added to the object's swap metadata. If
1836 * the swapblk is not valid, it is freed instead. Any previously
1837 * assigned swapblk is freed.
1840 swp_pager_meta_build(vm_object_t object, vm_pindex_t pindex, daddr_t swapblk)
1842 static volatile int exhausted;
1843 struct swblock *swap;
1844 struct swblock **pswap;
1847 VM_OBJECT_ASSERT_WLOCKED(object);
1849 * Convert default object to swap object if necessary
1851 if (object->type != OBJT_SWAP) {
1852 object->type = OBJT_SWAP;
1853 object->un_pager.swp.swp_bcount = 0;
1855 if (object->handle != NULL) {
1856 mtx_lock(&sw_alloc_mtx);
1858 NOBJLIST(object->handle),
1862 mtx_unlock(&sw_alloc_mtx);
1867 * Locate hash entry. If not found create, but if we aren't adding
1868 * anything just return. If we run out of space in the map we wait
1869 * and, since the hash table may have changed, retry.
1872 mtx_lock(&swhash_mtx);
1873 pswap = swp_pager_hash(object, pindex);
1875 if ((swap = *pswap) == NULL) {
1878 if (swapblk == SWAPBLK_NONE)
1881 swap = *pswap = uma_zalloc(swap_zone, M_NOWAIT |
1882 (curproc == pageproc ? M_USE_RESERVE : 0));
1884 mtx_unlock(&swhash_mtx);
1885 VM_OBJECT_WUNLOCK(object);
1886 if (uma_zone_exhausted(swap_zone)) {
1887 if (atomic_cmpset_int(&exhausted, 0, 1))
1888 printf("swap zone exhausted, "
1889 "increase kern.maxswzone\n");
1890 vm_pageout_oom(VM_OOM_SWAPZ);
1891 pause("swzonex", 10);
1894 VM_OBJECT_WLOCK(object);
1898 if (atomic_cmpset_int(&exhausted, 1, 0))
1899 printf("swap zone ok\n");
1901 swap->swb_hnext = NULL;
1902 swap->swb_object = object;
1903 swap->swb_index = pindex & ~(vm_pindex_t)SWAP_META_MASK;
1904 swap->swb_count = 0;
1906 ++object->un_pager.swp.swp_bcount;
1908 for (i = 0; i < SWAP_META_PAGES; ++i)
1909 swap->swb_pages[i] = SWAPBLK_NONE;
1913 * Delete prior contents of metadata
1915 idx = pindex & SWAP_META_MASK;
1917 if (swap->swb_pages[idx] != SWAPBLK_NONE) {
1918 swp_pager_freeswapspace(swap->swb_pages[idx], 1);
1923 * Enter block into metadata
1925 swap->swb_pages[idx] = swapblk;
1926 if (swapblk != SWAPBLK_NONE)
1929 mtx_unlock(&swhash_mtx);
1933 * SWP_PAGER_META_FREE() - free a range of blocks in the object's swap metadata
1935 * The requested range of blocks is freed, with any associated swap
1936 * returned to the swap bitmap.
1938 * This routine will free swap metadata structures as they are cleaned
1939 * out. This routine does *NOT* operate on swap metadata associated
1940 * with resident pages.
1943 swp_pager_meta_free(vm_object_t object, vm_pindex_t index, daddr_t count)
1946 VM_OBJECT_ASSERT_LOCKED(object);
1947 if (object->type != OBJT_SWAP)
1951 struct swblock **pswap;
1952 struct swblock *swap;
1954 mtx_lock(&swhash_mtx);
1955 pswap = swp_pager_hash(object, index);
1957 if ((swap = *pswap) != NULL) {
1958 daddr_t v = swap->swb_pages[index & SWAP_META_MASK];
1960 if (v != SWAPBLK_NONE) {
1961 swp_pager_freeswapspace(v, 1);
1962 swap->swb_pages[index & SWAP_META_MASK] =
1964 if (--swap->swb_count == 0) {
1965 *pswap = swap->swb_hnext;
1966 uma_zfree(swap_zone, swap);
1967 --object->un_pager.swp.swp_bcount;
1973 int n = SWAP_META_PAGES - (index & SWAP_META_MASK);
1977 mtx_unlock(&swhash_mtx);
1982 * SWP_PAGER_META_FREE_ALL() - destroy all swap metadata associated with object
1984 * This routine locates and destroys all swap metadata associated with
1988 swp_pager_meta_free_all(vm_object_t object)
1990 struct swblock **pswap, *swap;
1995 VM_OBJECT_ASSERT_WLOCKED(object);
1996 if (object->type != OBJT_SWAP)
2000 while (object->un_pager.swp.swp_bcount != 0) {
2001 mtx_lock(&swhash_mtx);
2002 pswap = swp_pager_hash(object, index);
2003 if ((swap = *pswap) != NULL) {
2004 for (i = 0; i < SWAP_META_PAGES; ++i) {
2005 v = swap->swb_pages[i];
2006 if (v != SWAPBLK_NONE) {
2008 swp_pager_freeswapspace(v, 1);
2011 if (swap->swb_count != 0)
2013 "swap_pager_meta_free_all: swb_count != 0");
2014 *pswap = swap->swb_hnext;
2015 uma_zfree(swap_zone, swap);
2016 --object->un_pager.swp.swp_bcount;
2018 mtx_unlock(&swhash_mtx);
2019 index += SWAP_META_PAGES;
2024 * SWP_PAGER_METACTL() - misc control of swap and vm_page_t meta data.
2026 * This routine is capable of looking up, popping, or freeing
2027 * swapblk assignments in the swap meta data or in the vm_page_t.
2028 * The routine typically returns the swapblk being looked-up, or popped,
2029 * or SWAPBLK_NONE if the block was freed, or SWAPBLK_NONE if the block
2030 * was invalid. This routine will automatically free any invalid
2031 * meta-data swapblks.
2033 * It is not possible to store invalid swapblks in the swap meta data
2034 * (other then a literal 'SWAPBLK_NONE'), so we don't bother checking.
2036 * When acting on a busy resident page and paging is in progress, we
2037 * have to wait until paging is complete but otherwise can act on the
2040 * SWM_FREE remove and free swap block from metadata
2041 * SWM_POP remove from meta data but do not free.. pop it out
2044 swp_pager_meta_ctl(vm_object_t object, vm_pindex_t pindex, int flags)
2046 struct swblock **pswap;
2047 struct swblock *swap;
2051 VM_OBJECT_ASSERT_LOCKED(object);
2053 * The meta data only exists of the object is OBJT_SWAP
2054 * and even then might not be allocated yet.
2056 if (object->type != OBJT_SWAP)
2057 return (SWAPBLK_NONE);
2060 mtx_lock(&swhash_mtx);
2061 pswap = swp_pager_hash(object, pindex);
2063 if ((swap = *pswap) != NULL) {
2064 idx = pindex & SWAP_META_MASK;
2065 r1 = swap->swb_pages[idx];
2067 if (r1 != SWAPBLK_NONE) {
2068 if (flags & SWM_FREE) {
2069 swp_pager_freeswapspace(r1, 1);
2072 if (flags & (SWM_FREE|SWM_POP)) {
2073 swap->swb_pages[idx] = SWAPBLK_NONE;
2074 if (--swap->swb_count == 0) {
2075 *pswap = swap->swb_hnext;
2076 uma_zfree(swap_zone, swap);
2077 --object->un_pager.swp.swp_bcount;
2082 mtx_unlock(&swhash_mtx);
2087 * System call swapon(name) enables swapping on device name,
2088 * which must be in the swdevsw. Return EBUSY
2089 * if already swapping on this device.
2091 #ifndef _SYS_SYSPROTO_H_
2092 struct swapon_args {
2102 sys_swapon(struct thread *td, struct swapon_args *uap)
2106 struct nameidata nd;
2109 error = priv_check(td, PRIV_SWAPON);
2114 while (swdev_syscall_active)
2115 tsleep(&swdev_syscall_active, PUSER - 1, "swpon", 0);
2116 swdev_syscall_active = 1;
2119 * Swap metadata may not fit in the KVM if we have physical
2122 if (swap_zone == NULL) {
2127 NDINIT(&nd, LOOKUP, ISOPEN | FOLLOW | AUDITVNODE1, UIO_USERSPACE,
2133 NDFREE(&nd, NDF_ONLY_PNBUF);
2136 if (vn_isdisk(vp, &error)) {
2137 error = swapongeom(td, vp);
2138 } else if (vp->v_type == VREG &&
2139 (vp->v_mount->mnt_vfc->vfc_flags & VFCF_NETWORK) != 0 &&
2140 (error = VOP_GETATTR(vp, &attr, td->td_ucred)) == 0) {
2142 * Allow direct swapping to NFS regular files in the same
2143 * way that nfs_mountroot() sets up diskless swapping.
2145 error = swaponvp(td, vp, attr.va_size / DEV_BSIZE);
2151 swdev_syscall_active = 0;
2152 wakeup_one(&swdev_syscall_active);
2158 * Check that the total amount of swap currently configured does not
2159 * exceed half the theoretical maximum. If it does, print a warning
2160 * message and return -1; otherwise, return 0.
2163 swapon_check_swzone(unsigned long npages)
2165 unsigned long maxpages;
2167 /* absolute maximum we can handle assuming 100% efficiency */
2168 maxpages = uma_zone_get_max(swap_zone) * SWAP_META_PAGES;
2170 /* recommend using no more than half that amount */
2171 if (npages > maxpages / 2) {
2172 printf("warning: total configured swap (%lu pages) "
2173 "exceeds maximum recommended amount (%lu pages).\n",
2174 npages, maxpages / 2);
2175 printf("warning: increase kern.maxswzone "
2176 "or reduce amount of swap.\n");
2183 swaponsomething(struct vnode *vp, void *id, u_long nblks,
2184 sw_strategy_t *strategy, sw_close_t *close, dev_t dev, int flags)
2186 struct swdevt *sp, *tsp;
2191 * nblks is in DEV_BSIZE'd chunks, convert to PAGE_SIZE'd chunks.
2192 * First chop nblks off to page-align it, then convert.
2194 * sw->sw_nblks is in page-sized chunks now too.
2196 nblks &= ~(ctodb(1) - 1);
2197 nblks = dbtoc(nblks);
2200 * If we go beyond this, we get overflows in the radix
2203 mblocks = 0x40000000 / BLIST_META_RADIX;
2204 if (nblks > mblocks) {
2206 "WARNING: reducing swap size to maximum of %luMB per unit\n",
2207 mblocks / 1024 / 1024 * PAGE_SIZE);
2211 sp = malloc(sizeof *sp, M_VMPGDATA, M_WAITOK | M_ZERO);
2216 sp->sw_nblks = nblks;
2218 sp->sw_strategy = strategy;
2219 sp->sw_close = close;
2220 sp->sw_flags = flags;
2222 sp->sw_blist = blist_create(nblks, M_WAITOK);
2224 * Do not free the first two block in order to avoid overwriting
2225 * any bsd label at the front of the partition
2227 blist_free(sp->sw_blist, 2, nblks - 2);
2230 mtx_lock(&sw_dev_mtx);
2231 TAILQ_FOREACH(tsp, &swtailq, sw_list) {
2232 if (tsp->sw_end >= dvbase) {
2234 * We put one uncovered page between the devices
2235 * in order to definitively prevent any cross-device
2238 dvbase = tsp->sw_end + 1;
2241 sp->sw_first = dvbase;
2242 sp->sw_end = dvbase + nblks;
2243 TAILQ_INSERT_TAIL(&swtailq, sp, sw_list);
2245 swap_pager_avail += nblks - 2;
2246 swap_total += (vm_ooffset_t)nblks * PAGE_SIZE;
2247 swapon_check_swzone(swap_total / PAGE_SIZE);
2249 mtx_unlock(&sw_dev_mtx);
2253 * SYSCALL: swapoff(devname)
2255 * Disable swapping on the given device.
2257 * XXX: Badly designed system call: it should use a device index
2258 * rather than filename as specification. We keep sw_vp around
2259 * only to make this work.
2261 #ifndef _SYS_SYSPROTO_H_
2262 struct swapoff_args {
2272 sys_swapoff(struct thread *td, struct swapoff_args *uap)
2275 struct nameidata nd;
2279 error = priv_check(td, PRIV_SWAPOFF);
2284 while (swdev_syscall_active)
2285 tsleep(&swdev_syscall_active, PUSER - 1, "swpoff", 0);
2286 swdev_syscall_active = 1;
2288 NDINIT(&nd, LOOKUP, FOLLOW | AUDITVNODE1, UIO_USERSPACE, uap->name,
2293 NDFREE(&nd, NDF_ONLY_PNBUF);
2296 mtx_lock(&sw_dev_mtx);
2297 TAILQ_FOREACH(sp, &swtailq, sw_list) {
2298 if (sp->sw_vp == vp)
2301 mtx_unlock(&sw_dev_mtx);
2306 error = swapoff_one(sp, td->td_ucred);
2308 swdev_syscall_active = 0;
2309 wakeup_one(&swdev_syscall_active);
2315 swapoff_one(struct swdevt *sp, struct ucred *cred)
2322 mtx_assert(&Giant, MA_OWNED);
2324 (void) vn_lock(sp->sw_vp, LK_EXCLUSIVE | LK_RETRY);
2325 error = mac_system_check_swapoff(cred, sp->sw_vp);
2326 (void) VOP_UNLOCK(sp->sw_vp, 0);
2330 nblks = sp->sw_nblks;
2333 * We can turn off this swap device safely only if the
2334 * available virtual memory in the system will fit the amount
2335 * of data we will have to page back in, plus an epsilon so
2336 * the system doesn't become critically low on swap space.
2338 if (cnt.v_free_count + cnt.v_cache_count + swap_pager_avail <
2339 nblks + nswap_lowat) {
2344 * Prevent further allocations on this device.
2346 mtx_lock(&sw_dev_mtx);
2347 sp->sw_flags |= SW_CLOSING;
2348 swap_pager_avail -= blist_fill(sp->sw_blist, 0, nblks);
2349 swap_total -= (vm_ooffset_t)nblks * PAGE_SIZE;
2350 mtx_unlock(&sw_dev_mtx);
2353 * Page in the contents of the device and close it.
2355 swap_pager_swapoff(sp);
2357 sp->sw_close(curthread, sp);
2358 mtx_lock(&sw_dev_mtx);
2360 TAILQ_REMOVE(&swtailq, sp, sw_list);
2362 if (nswapdev == 0) {
2363 swap_pager_full = 2;
2364 swap_pager_almost_full = 1;
2368 mtx_unlock(&sw_dev_mtx);
2369 blist_destroy(sp->sw_blist);
2370 free(sp, M_VMPGDATA);
2377 struct swdevt *sp, *spt;
2378 const char *devname;
2382 while (swdev_syscall_active)
2383 tsleep(&swdev_syscall_active, PUSER - 1, "swpoff", 0);
2384 swdev_syscall_active = 1;
2386 mtx_lock(&sw_dev_mtx);
2387 TAILQ_FOREACH_SAFE(sp, &swtailq, sw_list, spt) {
2388 mtx_unlock(&sw_dev_mtx);
2389 if (vn_isdisk(sp->sw_vp, NULL))
2390 devname = devtoname(sp->sw_vp->v_rdev);
2393 error = swapoff_one(sp, thread0.td_ucred);
2395 printf("Cannot remove swap device %s (error=%d), "
2396 "skipping.\n", devname, error);
2397 } else if (bootverbose) {
2398 printf("Swap device %s removed.\n", devname);
2400 mtx_lock(&sw_dev_mtx);
2402 mtx_unlock(&sw_dev_mtx);
2404 swdev_syscall_active = 0;
2405 wakeup_one(&swdev_syscall_active);
2410 swap_pager_status(int *total, int *used)
2416 mtx_lock(&sw_dev_mtx);
2417 TAILQ_FOREACH(sp, &swtailq, sw_list) {
2418 *total += sp->sw_nblks;
2419 *used += sp->sw_used;
2421 mtx_unlock(&sw_dev_mtx);
2425 swap_dev_info(int name, struct xswdev *xs, char *devname, size_t len)
2428 const char *tmp_devname;
2433 mtx_lock(&sw_dev_mtx);
2434 TAILQ_FOREACH(sp, &swtailq, sw_list) {
2439 xs->xsw_version = XSWDEV_VERSION;
2440 xs->xsw_dev = sp->sw_dev;
2441 xs->xsw_flags = sp->sw_flags;
2442 xs->xsw_nblks = sp->sw_nblks;
2443 xs->xsw_used = sp->sw_used;
2444 if (devname != NULL) {
2445 if (vn_isdisk(sp->sw_vp, NULL))
2446 tmp_devname = devtoname(sp->sw_vp->v_rdev);
2448 tmp_devname = "[file]";
2449 strncpy(devname, tmp_devname, len);
2454 mtx_unlock(&sw_dev_mtx);
2459 sysctl_vm_swap_info(SYSCTL_HANDLER_ARGS)
2464 if (arg2 != 1) /* name length */
2466 error = swap_dev_info(*(int *)arg1, &xs, NULL, 0);
2469 error = SYSCTL_OUT(req, &xs, sizeof(xs));
2473 SYSCTL_INT(_vm, OID_AUTO, nswapdev, CTLFLAG_RD, &nswapdev, 0,
2474 "Number of swap devices");
2475 SYSCTL_NODE(_vm, OID_AUTO, swap_info, CTLFLAG_RD, sysctl_vm_swap_info,
2476 "Swap statistics by device");
2479 * vmspace_swap_count() - count the approximate swap usage in pages for a
2482 * The map must be locked.
2484 * Swap usage is determined by taking the proportional swap used by
2485 * VM objects backing the VM map. To make up for fractional losses,
2486 * if the VM object has any swap use at all the associated map entries
2487 * count for at least 1 swap page.
2490 vmspace_swap_count(struct vmspace *vmspace)
2497 map = &vmspace->vm_map;
2500 for (cur = map->header.next; cur != &map->header; cur = cur->next) {
2501 if ((cur->eflags & MAP_ENTRY_IS_SUB_MAP) == 0 &&
2502 (object = cur->object.vm_object) != NULL) {
2503 VM_OBJECT_WLOCK(object);
2504 if (object->type == OBJT_SWAP &&
2505 object->un_pager.swp.swp_bcount != 0) {
2506 n = (cur->end - cur->start) / PAGE_SIZE;
2507 count += object->un_pager.swp.swp_bcount *
2508 SWAP_META_PAGES * n / object->size + 1;
2510 VM_OBJECT_WUNLOCK(object);
2519 * Swapping onto disk devices.
2523 static g_orphan_t swapgeom_orphan;
2525 static struct g_class g_swap_class = {
2527 .version = G_VERSION,
2528 .orphan = swapgeom_orphan,
2531 DECLARE_GEOM_CLASS(g_swap_class, g_class);
2535 swapgeom_close_ev(void *arg, int flags)
2537 struct g_consumer *cp;
2540 g_access(cp, -1, -1, 0);
2542 g_destroy_consumer(cp);
2546 * Add a reference to the g_consumer for an inflight transaction.
2549 swapgeom_acquire(struct g_consumer *cp)
2552 mtx_assert(&sw_dev_mtx, MA_OWNED);
2557 * Remove a reference from the g_consumer. Post a close event if
2558 * all referneces go away.
2561 swapgeom_release(struct g_consumer *cp, struct swdevt *sp)
2564 mtx_assert(&sw_dev_mtx, MA_OWNED);
2566 if (cp->index == 0) {
2567 if (g_post_event(swapgeom_close_ev, cp, M_NOWAIT, NULL) == 0)
2573 swapgeom_done(struct bio *bp2)
2577 struct g_consumer *cp;
2579 bp = bp2->bio_caller2;
2581 bp->b_ioflags = bp2->bio_flags;
2583 bp->b_ioflags |= BIO_ERROR;
2584 bp->b_resid = bp->b_bcount - bp2->bio_completed;
2585 bp->b_error = bp2->bio_error;
2587 sp = bp2->bio_caller1;
2588 mtx_lock(&sw_dev_mtx);
2589 swapgeom_release(cp, sp);
2590 mtx_unlock(&sw_dev_mtx);
2595 swapgeom_strategy(struct buf *bp, struct swdevt *sp)
2598 struct g_consumer *cp;
2600 mtx_lock(&sw_dev_mtx);
2603 mtx_unlock(&sw_dev_mtx);
2604 bp->b_error = ENXIO;
2605 bp->b_ioflags |= BIO_ERROR;
2609 swapgeom_acquire(cp);
2610 mtx_unlock(&sw_dev_mtx);
2611 if (bp->b_iocmd == BIO_WRITE)
2614 bio = g_alloc_bio();
2616 mtx_lock(&sw_dev_mtx);
2617 swapgeom_release(cp, sp);
2618 mtx_unlock(&sw_dev_mtx);
2619 bp->b_error = ENOMEM;
2620 bp->b_ioflags |= BIO_ERROR;
2625 bio->bio_caller1 = sp;
2626 bio->bio_caller2 = bp;
2627 bio->bio_cmd = bp->b_iocmd;
2628 bio->bio_offset = (bp->b_blkno - sp->sw_first) * PAGE_SIZE;
2629 bio->bio_length = bp->b_bcount;
2630 bio->bio_done = swapgeom_done;
2631 if ((bp->b_flags & B_UNMAPPED) != 0) {
2632 bio->bio_ma = bp->b_pages;
2633 bio->bio_data = unmapped_buf;
2634 bio->bio_ma_offset = (vm_offset_t)bp->b_offset & PAGE_MASK;
2635 bio->bio_ma_n = bp->b_npages;
2636 bio->bio_flags |= BIO_UNMAPPED;
2638 bio->bio_data = bp->b_data;
2641 g_io_request(bio, cp);
2646 swapgeom_orphan(struct g_consumer *cp)
2651 mtx_lock(&sw_dev_mtx);
2652 TAILQ_FOREACH(sp, &swtailq, sw_list) {
2653 if (sp->sw_id == cp) {
2654 sp->sw_flags |= SW_CLOSING;
2659 * Drop reference we were created with. Do directly since we're in a
2660 * special context where we don't have to queue the call to
2661 * swapgeom_close_ev().
2664 destroy = ((sp != NULL) && (cp->index == 0));
2667 mtx_unlock(&sw_dev_mtx);
2669 swapgeom_close_ev(cp, 0);
2673 swapgeom_close(struct thread *td, struct swdevt *sw)
2675 struct g_consumer *cp;
2677 mtx_lock(&sw_dev_mtx);
2680 mtx_unlock(&sw_dev_mtx);
2681 /* XXX: direct call when Giant untangled */
2683 g_waitfor_event(swapgeom_close_ev, cp, M_WAITOK, NULL);
2694 swapongeom_ev(void *arg, int flags)
2697 struct g_provider *pp;
2698 struct g_consumer *cp;
2699 static struct g_geom *gp;
2706 pp = g_dev_getprovider(swh->dev);
2708 swh->error = ENODEV;
2711 mtx_lock(&sw_dev_mtx);
2712 TAILQ_FOREACH(sp, &swtailq, sw_list) {
2714 if (cp != NULL && cp->provider == pp) {
2715 mtx_unlock(&sw_dev_mtx);
2720 mtx_unlock(&sw_dev_mtx);
2722 gp = g_new_geomf(&g_swap_class, "swap");
2723 cp = g_new_consumer(gp);
2724 cp->index = 1; /* Number of active I/Os, plus one for being active. */
2725 cp->flags |= G_CF_DIRECT_SEND | G_CF_DIRECT_RECEIVE;
2728 * XXX: Everytime you think you can improve the margin for
2729 * footshooting, somebody depends on the ability to do so:
2730 * savecore(8) wants to write to our swapdev so we cannot
2731 * set an exclusive count :-(
2733 error = g_access(cp, 1, 1, 0);
2736 g_destroy_consumer(cp);
2740 nblks = pp->mediasize / DEV_BSIZE;
2741 swaponsomething(swh->vp, cp, nblks, swapgeom_strategy,
2742 swapgeom_close, dev2udev(swh->dev),
2743 (pp->flags & G_PF_ACCEPT_UNMAPPED) != 0 ? SW_UNMAPPED : 0);
2748 swapongeom(struct thread *td, struct vnode *vp)
2753 vn_lock(vp, LK_EXCLUSIVE | LK_RETRY);
2755 swh.dev = vp->v_rdev;
2758 /* XXX: direct call when Giant untangled */
2759 error = g_waitfor_event(swapongeom_ev, &swh, M_WAITOK, NULL);
2769 * This is used mainly for network filesystem (read: probably only tested
2770 * with NFS) swapfiles.
2775 swapdev_strategy(struct buf *bp, struct swdevt *sp)
2779 bp->b_blkno = ctodb(bp->b_blkno - sp->sw_first);
2783 if (bp->b_iocmd == BIO_WRITE) {
2785 bufobj_wdrop(bp->b_bufobj);
2786 bufobj_wref(&vp2->v_bufobj);
2788 if (bp->b_bufobj != &vp2->v_bufobj)
2789 bp->b_bufobj = &vp2->v_bufobj;
2791 bp->b_iooffset = dbtob(bp->b_blkno);
2797 swapdev_close(struct thread *td, struct swdevt *sp)
2800 VOP_CLOSE(sp->sw_vp, FREAD | FWRITE, td->td_ucred, td);
2806 swaponvp(struct thread *td, struct vnode *vp, u_long nblks)
2813 mtx_lock(&sw_dev_mtx);
2814 TAILQ_FOREACH(sp, &swtailq, sw_list) {
2815 if (sp->sw_id == vp) {
2816 mtx_unlock(&sw_dev_mtx);
2820 mtx_unlock(&sw_dev_mtx);
2822 (void) vn_lock(vp, LK_EXCLUSIVE | LK_RETRY);
2824 error = mac_system_check_swapon(td->td_ucred, vp);
2827 error = VOP_OPEN(vp, FREAD | FWRITE, td->td_ucred, td, NULL);
2828 (void) VOP_UNLOCK(vp, 0);
2832 swaponsomething(vp, vp, nblks, swapdev_strategy, swapdev_close,