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 * SWB_NPAGES must be a power of 2. It may be set to 1, 2, 4, 8, 16
119 * or 32 pages per allocation.
120 * The 32-page limit is due to the radix code (kern/subr_blist.c).
122 #ifndef MAX_PAGEOUT_CLUSTER
123 #define MAX_PAGEOUT_CLUSTER 16
126 #if !defined(SWB_NPAGES)
127 #define SWB_NPAGES MAX_PAGEOUT_CLUSTER
131 * The swblock structure maps an object and a small, fixed-size range
132 * of page indices to disk addresses within a swap area.
133 * The collection of these mappings is implemented as a hash table.
134 * Unused disk addresses within a swap area are allocated and managed
137 #define SWCORRECT(n) (sizeof(void *) * (n) / sizeof(daddr_t))
138 #define SWAP_META_PAGES (SWB_NPAGES * 2)
139 #define SWAP_META_MASK (SWAP_META_PAGES - 1)
142 struct swblock *swb_hnext;
143 vm_object_t swb_object;
144 vm_pindex_t swb_index;
146 daddr_t swb_pages[SWAP_META_PAGES];
149 static MALLOC_DEFINE(M_VMPGDATA, "vm_pgdata", "swap pager private data");
150 static struct mtx sw_dev_mtx;
151 static TAILQ_HEAD(, swdevt) swtailq = TAILQ_HEAD_INITIALIZER(swtailq);
152 static struct swdevt *swdevhd; /* Allocate from here next */
153 static int nswapdev; /* Number of swap devices */
154 int swap_pager_avail;
155 static int swdev_syscall_active = 0; /* serialize swap(on|off) */
157 static vm_ooffset_t swap_total;
158 SYSCTL_QUAD(_vm, OID_AUTO, swap_total, CTLFLAG_RD, &swap_total, 0,
159 "Total amount of available swap storage.");
160 static vm_ooffset_t swap_reserved;
161 SYSCTL_QUAD(_vm, OID_AUTO, swap_reserved, CTLFLAG_RD, &swap_reserved, 0,
162 "Amount of swap storage needed to back all allocated anonymous memory.");
163 static int overcommit = 0;
164 SYSCTL_INT(_vm, OID_AUTO, overcommit, CTLFLAG_RW, &overcommit, 0,
165 "Configure virtual memory overcommit behavior. See tuning(7) "
167 static unsigned long swzone;
168 SYSCTL_ULONG(_vm, OID_AUTO, swzone, CTLFLAG_RD, &swzone, 0,
169 "Actual size of swap metadata zone");
170 static unsigned long swap_maxpages;
171 SYSCTL_ULONG(_vm, OID_AUTO, swap_maxpages, CTLFLAG_RD, &swap_maxpages, 0,
172 "Maximum amount of swap supported");
174 /* bits from overcommit */
175 #define SWAP_RESERVE_FORCE_ON (1 << 0)
176 #define SWAP_RESERVE_RLIMIT_ON (1 << 1)
177 #define SWAP_RESERVE_ALLOW_NONWIRED (1 << 2)
180 swap_reserve(vm_ooffset_t incr)
183 return (swap_reserve_by_cred(incr, curthread->td_ucred));
187 swap_reserve_by_cred(vm_ooffset_t incr, struct ucred *cred)
192 static struct timeval lastfail;
195 uip = cred->cr_ruidinfo;
197 if (incr & PAGE_MASK)
198 panic("swap_reserve: & PAGE_MASK");
202 error = racct_add(curproc, RACCT_SWAP, incr);
203 PROC_UNLOCK(curproc);
209 mtx_lock(&sw_dev_mtx);
210 r = swap_reserved + incr;
211 if (overcommit & SWAP_RESERVE_ALLOW_NONWIRED) {
212 s = vm_cnt.v_page_count - vm_cnt.v_free_reserved - vm_cnt.v_wire_count;
217 if ((overcommit & SWAP_RESERVE_FORCE_ON) == 0 || r <= s ||
218 (error = priv_check(curthread, PRIV_VM_SWAP_NOQUOTA)) == 0) {
222 mtx_unlock(&sw_dev_mtx);
226 UIDINFO_VMSIZE_LOCK(uip);
227 if ((overcommit & SWAP_RESERVE_RLIMIT_ON) != 0 &&
228 uip->ui_vmsize + incr > lim_cur(curproc, RLIMIT_SWAP) &&
229 priv_check(curthread, PRIV_VM_SWAP_NORLIMIT))
232 uip->ui_vmsize += incr;
233 UIDINFO_VMSIZE_UNLOCK(uip);
234 PROC_UNLOCK(curproc);
236 mtx_lock(&sw_dev_mtx);
237 swap_reserved -= incr;
238 mtx_unlock(&sw_dev_mtx);
241 if (!res && ppsratecheck(&lastfail, &curfail, 1)) {
242 printf("uid %d, pid %d: swap reservation for %jd bytes failed\n",
243 uip->ui_uid, curproc->p_pid, incr);
249 racct_sub(curproc, RACCT_SWAP, incr);
250 PROC_UNLOCK(curproc);
258 swap_reserve_force(vm_ooffset_t incr)
262 mtx_lock(&sw_dev_mtx);
263 swap_reserved += incr;
264 mtx_unlock(&sw_dev_mtx);
268 racct_add_force(curproc, RACCT_SWAP, incr);
269 PROC_UNLOCK(curproc);
272 uip = curthread->td_ucred->cr_ruidinfo;
274 UIDINFO_VMSIZE_LOCK(uip);
275 uip->ui_vmsize += incr;
276 UIDINFO_VMSIZE_UNLOCK(uip);
277 PROC_UNLOCK(curproc);
281 swap_release(vm_ooffset_t decr)
286 cred = curthread->td_ucred;
287 swap_release_by_cred(decr, cred);
288 PROC_UNLOCK(curproc);
292 swap_release_by_cred(vm_ooffset_t decr, struct ucred *cred)
296 uip = cred->cr_ruidinfo;
298 if (decr & PAGE_MASK)
299 panic("swap_release: & PAGE_MASK");
301 mtx_lock(&sw_dev_mtx);
302 if (swap_reserved < decr)
303 panic("swap_reserved < decr");
304 swap_reserved -= decr;
305 mtx_unlock(&sw_dev_mtx);
307 UIDINFO_VMSIZE_LOCK(uip);
308 if (uip->ui_vmsize < decr)
309 printf("negative vmsize for uid = %d\n", uip->ui_uid);
310 uip->ui_vmsize -= decr;
311 UIDINFO_VMSIZE_UNLOCK(uip);
313 racct_sub_cred(cred, RACCT_SWAP, decr);
316 static void swapdev_strategy(struct buf *, struct swdevt *sw);
318 #define SWM_FREE 0x02 /* free, period */
319 #define SWM_POP 0x04 /* pop out */
321 int swap_pager_full = 2; /* swap space exhaustion (task killing) */
322 static int swap_pager_almost_full = 1; /* swap space exhaustion (w/hysteresis)*/
323 static int nsw_rcount; /* free read buffers */
324 static int nsw_wcount_sync; /* limit write buffers / synchronous */
325 static int nsw_wcount_async; /* limit write buffers / asynchronous */
326 static int nsw_wcount_async_max;/* assigned maximum */
327 static int nsw_cluster_max; /* maximum VOP I/O allowed */
329 static struct swblock **swhash;
330 static int swhash_mask;
331 static struct mtx swhash_mtx;
333 static int swap_async_max = 4; /* maximum in-progress async I/O's */
334 static struct sx sw_alloc_sx;
337 SYSCTL_INT(_vm, OID_AUTO, swap_async_max,
338 CTLFLAG_RW, &swap_async_max, 0, "Maximum running async swap ops");
341 * "named" and "unnamed" anon region objects. Try to reduce the overhead
342 * of searching a named list by hashing it just a little.
347 #define NOBJLIST(handle) \
348 (&swap_pager_object_list[((int)(intptr_t)handle >> 4) & (NOBJLISTS-1)])
350 static struct mtx sw_alloc_mtx; /* protect list manipulation */
351 static struct pagerlst swap_pager_object_list[NOBJLISTS];
352 static uma_zone_t swap_zone;
355 * pagerops for OBJT_SWAP - "swap pager". Some ops are also global procedure
356 * calls hooked from other parts of the VM system and do not appear here.
357 * (see vm/swap_pager.h).
360 swap_pager_alloc(void *handle, vm_ooffset_t size,
361 vm_prot_t prot, vm_ooffset_t offset, struct ucred *);
362 static void swap_pager_dealloc(vm_object_t object);
363 static int swap_pager_getpages(vm_object_t, vm_page_t *, int, int);
364 static int swap_pager_getpages_async(vm_object_t, vm_page_t *, int, int,
365 pgo_getpages_iodone_t, void *);
366 static void swap_pager_putpages(vm_object_t, vm_page_t *, int, boolean_t, int *);
368 swap_pager_haspage(vm_object_t object, vm_pindex_t pindex, int *before, int *after);
369 static void swap_pager_init(void);
370 static void swap_pager_unswapped(vm_page_t);
371 static void swap_pager_swapoff(struct swdevt *sp);
373 struct pagerops swappagerops = {
374 .pgo_init = swap_pager_init, /* early system initialization of pager */
375 .pgo_alloc = swap_pager_alloc, /* allocate an OBJT_SWAP object */
376 .pgo_dealloc = swap_pager_dealloc, /* deallocate an OBJT_SWAP object */
377 .pgo_getpages = swap_pager_getpages, /* pagein */
378 .pgo_getpages_async = swap_pager_getpages_async, /* pagein (async) */
379 .pgo_putpages = swap_pager_putpages, /* pageout */
380 .pgo_haspage = swap_pager_haspage, /* get backing store status for page */
381 .pgo_pageunswapped = swap_pager_unswapped, /* remove swap related to page */
385 * dmmax is in page-sized chunks with the new swap system. It was
386 * dev-bsized chunks in the old. dmmax is always a power of 2.
388 * swap_*() routines are externally accessible. swp_*() routines are
392 static int nswap_lowat = 128; /* in pages, swap_pager_almost_full warn */
393 static int nswap_hiwat = 512; /* in pages, swap_pager_almost_full warn */
395 SYSCTL_INT(_vm, OID_AUTO, dmmax,
396 CTLFLAG_RD, &dmmax, 0, "Maximum size of a swap block");
398 static void swp_sizecheck(void);
399 static void swp_pager_async_iodone(struct buf *bp);
400 static int swapongeom(struct thread *, struct vnode *);
401 static int swaponvp(struct thread *, struct vnode *, u_long);
402 static int swapoff_one(struct swdevt *sp, struct ucred *cred);
405 * Swap bitmap functions
407 static void swp_pager_freeswapspace(daddr_t blk, int npages);
408 static daddr_t swp_pager_getswapspace(int npages);
413 static struct swblock **swp_pager_hash(vm_object_t object, vm_pindex_t index);
414 static void swp_pager_meta_build(vm_object_t, vm_pindex_t, daddr_t);
415 static void swp_pager_meta_free(vm_object_t, vm_pindex_t, daddr_t);
416 static void swp_pager_meta_free_all(vm_object_t);
417 static daddr_t swp_pager_meta_ctl(vm_object_t, vm_pindex_t, int);
420 swp_pager_free_nrpage(vm_page_t m)
424 if (m->wire_count == 0)
430 * SWP_SIZECHECK() - update swap_pager_full indication
432 * update the swap_pager_almost_full indication and warn when we are
433 * about to run out of swap space, using lowat/hiwat hysteresis.
435 * Clear swap_pager_full ( task killing ) indication when lowat is met.
437 * No restrictions on call
438 * This routine may not block.
444 if (swap_pager_avail < nswap_lowat) {
445 if (swap_pager_almost_full == 0) {
446 printf("swap_pager: out of swap space\n");
447 swap_pager_almost_full = 1;
451 if (swap_pager_avail > nswap_hiwat)
452 swap_pager_almost_full = 0;
457 * SWP_PAGER_HASH() - hash swap meta data
459 * This is an helper function which hashes the swapblk given
460 * the object and page index. It returns a pointer to a pointer
461 * to the object, or a pointer to a NULL pointer if it could not
464 static struct swblock **
465 swp_pager_hash(vm_object_t object, vm_pindex_t index)
467 struct swblock **pswap;
468 struct swblock *swap;
470 index &= ~(vm_pindex_t)SWAP_META_MASK;
471 pswap = &swhash[(index ^ (int)(intptr_t)object) & swhash_mask];
472 while ((swap = *pswap) != NULL) {
473 if (swap->swb_object == object &&
474 swap->swb_index == index
478 pswap = &swap->swb_hnext;
484 * SWAP_PAGER_INIT() - initialize the swap pager!
486 * Expected to be started from system init. NOTE: This code is run
487 * before much else so be careful what you depend on. Most of the VM
488 * system has yet to be initialized at this point.
491 swap_pager_init(void)
494 * Initialize object lists
498 for (i = 0; i < NOBJLISTS; ++i)
499 TAILQ_INIT(&swap_pager_object_list[i]);
500 mtx_init(&sw_alloc_mtx, "swap_pager list", NULL, MTX_DEF);
501 mtx_init(&sw_dev_mtx, "swapdev", NULL, MTX_DEF);
504 * Device Stripe, in PAGE_SIZE'd blocks
506 dmmax = SWB_NPAGES * 2;
510 * SWAP_PAGER_SWAP_INIT() - swap pager initialization from pageout process
512 * Expected to be started from pageout process once, prior to entering
516 swap_pager_swap_init(void)
521 * Number of in-transit swap bp operations. Don't
522 * exhaust the pbufs completely. Make sure we
523 * initialize workable values (0 will work for hysteresis
524 * but it isn't very efficient).
526 * The nsw_cluster_max is constrained by the bp->b_pages[]
527 * array (MAXPHYS/PAGE_SIZE) and our locally defined
528 * MAX_PAGEOUT_CLUSTER. Also be aware that swap ops are
529 * constrained by the swap device interleave stripe size.
531 * Currently we hardwire nsw_wcount_async to 4. This limit is
532 * designed to prevent other I/O from having high latencies due to
533 * our pageout I/O. The value 4 works well for one or two active swap
534 * devices but is probably a little low if you have more. Even so,
535 * a higher value would probably generate only a limited improvement
536 * with three or four active swap devices since the system does not
537 * typically have to pageout at extreme bandwidths. We will want
538 * at least 2 per swap devices, and 4 is a pretty good value if you
539 * have one NFS swap device due to the command/ack latency over NFS.
540 * So it all works out pretty well.
542 nsw_cluster_max = min((MAXPHYS/PAGE_SIZE), MAX_PAGEOUT_CLUSTER);
545 nsw_rcount = (nswbuf + 1) / 2;
546 nsw_wcount_sync = (nswbuf + 3) / 4;
547 nsw_wcount_async = 4;
548 nsw_wcount_async_max = nsw_wcount_async;
549 mtx_unlock(&pbuf_mtx);
552 * Initialize our zone. Right now I'm just guessing on the number
553 * we need based on the number of pages in the system. Each swblock
554 * can hold 32 pages, so this is probably overkill. This reservation
555 * is typically limited to around 32MB by default.
557 n = vm_cnt.v_page_count / 2;
558 if (maxswzone && n > maxswzone / sizeof(struct swblock))
559 n = maxswzone / sizeof(struct swblock);
561 swap_zone = uma_zcreate("SWAPMETA", sizeof(struct swblock), NULL, NULL,
562 NULL, NULL, UMA_ALIGN_PTR, UMA_ZONE_NOFREE | UMA_ZONE_VM);
563 if (swap_zone == NULL)
564 panic("failed to create swap_zone.");
566 if (uma_zone_reserve_kva(swap_zone, n))
569 * if the allocation failed, try a zone two thirds the
570 * size of the previous attempt.
575 printf("Swap zone entries reduced from %lu to %lu.\n", n2, n);
576 swap_maxpages = n * SWAP_META_PAGES;
577 swzone = n * sizeof(struct swblock);
581 * Initialize our meta-data hash table. The swapper does not need to
582 * be quite as efficient as the VM system, so we do not use an
583 * oversized hash table.
585 * n: size of hash table, must be power of 2
586 * swhash_mask: hash table index mask
588 for (n = 1; n < n2 / 8; n *= 2)
590 swhash = malloc(sizeof(struct swblock *) * n, M_VMPGDATA, M_WAITOK | M_ZERO);
592 mtx_init(&swhash_mtx, "swap_pager swhash", NULL, MTX_DEF);
596 * SWAP_PAGER_ALLOC() - allocate a new OBJT_SWAP VM object and instantiate
597 * its metadata structures.
599 * This routine is called from the mmap and fork code to create a new
600 * OBJT_SWAP object. We do this by creating an OBJT_DEFAULT object
601 * and then converting it with swp_pager_meta_build().
603 * This routine may block in vm_object_allocate() and create a named
604 * object lookup race, so we must interlock.
609 swap_pager_alloc(void *handle, vm_ooffset_t size, vm_prot_t prot,
610 vm_ooffset_t offset, struct ucred *cred)
615 pindex = OFF_TO_IDX(offset + PAGE_MASK + size);
619 * Reference existing named region or allocate new one. There
620 * should not be a race here against swp_pager_meta_build()
621 * as called from vm_page_remove() in regards to the lookup
624 sx_xlock(&sw_alloc_sx);
625 object = vm_pager_object_lookup(NOBJLIST(handle), handle);
626 if (object == NULL) {
628 if (!swap_reserve_by_cred(size, cred)) {
629 sx_xunlock(&sw_alloc_sx);
635 object = vm_object_allocate(OBJT_DEFAULT, pindex);
636 VM_OBJECT_WLOCK(object);
637 object->handle = handle;
640 object->charge = size;
642 swp_pager_meta_build(object, 0, SWAPBLK_NONE);
643 VM_OBJECT_WUNLOCK(object);
645 sx_xunlock(&sw_alloc_sx);
649 if (!swap_reserve_by_cred(size, cred))
653 object = vm_object_allocate(OBJT_DEFAULT, pindex);
654 VM_OBJECT_WLOCK(object);
657 object->charge = size;
659 swp_pager_meta_build(object, 0, SWAPBLK_NONE);
660 VM_OBJECT_WUNLOCK(object);
666 * SWAP_PAGER_DEALLOC() - remove swap metadata from object
668 * The swap backing for the object is destroyed. The code is
669 * designed such that we can reinstantiate it later, but this
670 * routine is typically called only when the entire object is
671 * about to be destroyed.
673 * The object must be locked.
676 swap_pager_dealloc(vm_object_t object)
680 * Remove from list right away so lookups will fail if we block for
681 * pageout completion.
683 if (object->handle != NULL) {
684 mtx_lock(&sw_alloc_mtx);
685 TAILQ_REMOVE(NOBJLIST(object->handle), object, pager_object_list);
686 mtx_unlock(&sw_alloc_mtx);
689 VM_OBJECT_ASSERT_WLOCKED(object);
690 vm_object_pip_wait(object, "swpdea");
693 * Free all remaining metadata. We only bother to free it from
694 * the swap meta data. We do not attempt to free swapblk's still
695 * associated with vm_page_t's for this object. We do not care
696 * if paging is still in progress on some objects.
698 swp_pager_meta_free_all(object);
701 /************************************************************************
702 * SWAP PAGER BITMAP ROUTINES *
703 ************************************************************************/
706 * SWP_PAGER_GETSWAPSPACE() - allocate raw swap space
708 * Allocate swap for the requested number of pages. The starting
709 * swap block number (a page index) is returned or SWAPBLK_NONE
710 * if the allocation failed.
712 * Also has the side effect of advising that somebody made a mistake
713 * when they configured swap and didn't configure enough.
715 * This routine may not sleep.
717 * We allocate in round-robin fashion from the configured devices.
720 swp_pager_getswapspace(int npages)
727 mtx_lock(&sw_dev_mtx);
729 for (i = 0; i < nswapdev; i++) {
731 sp = TAILQ_FIRST(&swtailq);
732 if (!(sp->sw_flags & SW_CLOSING)) {
733 blk = blist_alloc(sp->sw_blist, npages);
734 if (blk != SWAPBLK_NONE) {
736 sp->sw_used += npages;
737 swap_pager_avail -= npages;
739 swdevhd = TAILQ_NEXT(sp, sw_list);
743 sp = TAILQ_NEXT(sp, sw_list);
745 if (swap_pager_full != 2) {
746 printf("swap_pager_getswapspace(%d): failed\n", npages);
748 swap_pager_almost_full = 1;
752 mtx_unlock(&sw_dev_mtx);
757 swp_pager_isondev(daddr_t blk, struct swdevt *sp)
760 return (blk >= sp->sw_first && blk < sp->sw_end);
764 swp_pager_strategy(struct buf *bp)
768 mtx_lock(&sw_dev_mtx);
769 TAILQ_FOREACH(sp, &swtailq, sw_list) {
770 if (bp->b_blkno >= sp->sw_first && bp->b_blkno < sp->sw_end) {
771 mtx_unlock(&sw_dev_mtx);
772 if ((sp->sw_flags & SW_UNMAPPED) != 0 &&
773 unmapped_buf_allowed) {
774 bp->b_kvaalloc = bp->b_data;
775 bp->b_data = unmapped_buf;
776 bp->b_kvabase = unmapped_buf;
778 bp->b_flags |= B_UNMAPPED;
780 pmap_qenter((vm_offset_t)bp->b_data,
781 &bp->b_pages[0], bp->b_bcount / PAGE_SIZE);
783 sp->sw_strategy(bp, sp);
787 panic("Swapdev not found");
792 * SWP_PAGER_FREESWAPSPACE() - free raw swap space
794 * This routine returns the specified swap blocks back to the bitmap.
796 * This routine may not sleep.
799 swp_pager_freeswapspace(daddr_t blk, int npages)
803 mtx_lock(&sw_dev_mtx);
804 TAILQ_FOREACH(sp, &swtailq, sw_list) {
805 if (blk >= sp->sw_first && blk < sp->sw_end) {
806 sp->sw_used -= npages;
808 * If we are attempting to stop swapping on
809 * this device, we don't want to mark any
810 * blocks free lest they be reused.
812 if ((sp->sw_flags & SW_CLOSING) == 0) {
813 blist_free(sp->sw_blist, blk - sp->sw_first,
815 swap_pager_avail += npages;
818 mtx_unlock(&sw_dev_mtx);
822 panic("Swapdev not found");
826 * SWAP_PAGER_FREESPACE() - frees swap blocks associated with a page
827 * range within an object.
829 * This is a globally accessible routine.
831 * This routine removes swapblk assignments from swap metadata.
833 * The external callers of this routine typically have already destroyed
834 * or renamed vm_page_t's associated with this range in the object so
837 * The object must be locked.
840 swap_pager_freespace(vm_object_t object, vm_pindex_t start, vm_size_t size)
843 swp_pager_meta_free(object, start, size);
847 * SWAP_PAGER_RESERVE() - reserve swap blocks in object
849 * Assigns swap blocks to the specified range within the object. The
850 * swap blocks are not zeroed. Any previous swap assignment is destroyed.
852 * Returns 0 on success, -1 on failure.
855 swap_pager_reserve(vm_object_t object, vm_pindex_t start, vm_size_t size)
858 daddr_t blk = SWAPBLK_NONE;
859 vm_pindex_t beg = start; /* save start index */
861 VM_OBJECT_WLOCK(object);
865 while ((blk = swp_pager_getswapspace(n)) == SWAPBLK_NONE) {
868 swp_pager_meta_free(object, beg, start - beg);
869 VM_OBJECT_WUNLOCK(object);
874 swp_pager_meta_build(object, start, blk);
880 swp_pager_meta_free(object, start, n);
881 VM_OBJECT_WUNLOCK(object);
886 * SWAP_PAGER_COPY() - copy blocks from source pager to destination pager
887 * and destroy the source.
889 * Copy any valid swapblks from the source to the destination. In
890 * cases where both the source and destination have a valid swapblk,
891 * we keep the destination's.
893 * This routine is allowed to sleep. It may sleep allocating metadata
894 * indirectly through swp_pager_meta_build() or if paging is still in
895 * progress on the source.
897 * The source object contains no vm_page_t's (which is just as well)
899 * The source object is of type OBJT_SWAP.
901 * The source and destination objects must be locked.
902 * Both object locks may temporarily be released.
905 swap_pager_copy(vm_object_t srcobject, vm_object_t dstobject,
906 vm_pindex_t offset, int destroysource)
910 VM_OBJECT_ASSERT_WLOCKED(srcobject);
911 VM_OBJECT_ASSERT_WLOCKED(dstobject);
914 * If destroysource is set, we remove the source object from the
915 * swap_pager internal queue now.
918 if (srcobject->handle != NULL) {
919 mtx_lock(&sw_alloc_mtx);
921 NOBJLIST(srcobject->handle),
925 mtx_unlock(&sw_alloc_mtx);
930 * transfer source to destination.
932 for (i = 0; i < dstobject->size; ++i) {
936 * Locate (without changing) the swapblk on the destination,
937 * unless it is invalid in which case free it silently, or
938 * if the destination is a resident page, in which case the
939 * source is thrown away.
941 dstaddr = swp_pager_meta_ctl(dstobject, i, 0);
943 if (dstaddr == SWAPBLK_NONE) {
945 * Destination has no swapblk and is not resident,
950 srcaddr = swp_pager_meta_ctl(
956 if (srcaddr != SWAPBLK_NONE) {
958 * swp_pager_meta_build() can sleep.
960 vm_object_pip_add(srcobject, 1);
961 VM_OBJECT_WUNLOCK(srcobject);
962 vm_object_pip_add(dstobject, 1);
963 swp_pager_meta_build(dstobject, i, srcaddr);
964 vm_object_pip_wakeup(dstobject);
965 VM_OBJECT_WLOCK(srcobject);
966 vm_object_pip_wakeup(srcobject);
970 * Destination has valid swapblk or it is represented
971 * by a resident page. We destroy the sourceblock.
974 swp_pager_meta_ctl(srcobject, i + offset, SWM_FREE);
979 * Free left over swap blocks in source.
981 * We have to revert the type to OBJT_DEFAULT so we do not accidently
982 * double-remove the object from the swap queues.
985 swp_pager_meta_free_all(srcobject);
987 * Reverting the type is not necessary, the caller is going
988 * to destroy srcobject directly, but I'm doing it here
989 * for consistency since we've removed the object from its
992 srcobject->type = OBJT_DEFAULT;
997 * SWAP_PAGER_HASPAGE() - determine if we have good backing store for
998 * the requested page.
1000 * We determine whether good backing store exists for the requested
1001 * page and return TRUE if it does, FALSE if it doesn't.
1003 * If TRUE, we also try to determine how much valid, contiguous backing
1004 * store exists before and after the requested page within a reasonable
1005 * distance. We do not try to restrict it to the swap device stripe
1006 * (that is handled in getpages/putpages). It probably isn't worth
1010 swap_pager_haspage(vm_object_t object, vm_pindex_t pindex, int *before, int *after)
1014 VM_OBJECT_ASSERT_LOCKED(object);
1016 * do we have good backing store at the requested index ?
1018 blk0 = swp_pager_meta_ctl(object, pindex, 0);
1020 if (blk0 == SWAPBLK_NONE) {
1029 * find backwards-looking contiguous good backing store
1031 if (before != NULL) {
1034 for (i = 1; i < (SWB_NPAGES/2); ++i) {
1039 blk = swp_pager_meta_ctl(object, pindex - i, 0);
1040 if (blk != blk0 - i)
1047 * find forward-looking contiguous good backing store
1049 if (after != NULL) {
1052 for (i = 1; i < (SWB_NPAGES/2); ++i) {
1055 blk = swp_pager_meta_ctl(object, pindex + i, 0);
1056 if (blk != blk0 + i)
1065 * SWAP_PAGER_PAGE_UNSWAPPED() - remove swap backing store related to page
1067 * This removes any associated swap backing store, whether valid or
1068 * not, from the page.
1070 * This routine is typically called when a page is made dirty, at
1071 * which point any associated swap can be freed. MADV_FREE also
1072 * calls us in a special-case situation
1074 * NOTE!!! If the page is clean and the swap was valid, the caller
1075 * should make the page dirty before calling this routine. This routine
1076 * does NOT change the m->dirty status of the page. Also: MADV_FREE
1079 * This routine may not sleep.
1081 * The object containing the page must be locked.
1084 swap_pager_unswapped(vm_page_t m)
1087 swp_pager_meta_ctl(m->object, m->pindex, SWM_FREE);
1091 * SWAP_PAGER_GETPAGES() - bring pages in from swap
1093 * Attempt to retrieve (m, count) pages from backing store, but make
1094 * sure we retrieve at least m[reqpage]. We try to load in as large
1095 * a chunk surrounding m[reqpage] as is contiguous in swap and which
1096 * belongs to the same object.
1098 * The code is designed for asynchronous operation and
1099 * immediate-notification of 'reqpage' but tends not to be
1100 * used that way. Please do not optimize-out this algorithmic
1101 * feature, I intend to improve on it in the future.
1103 * The parent has a single vm_object_pip_add() reference prior to
1104 * calling us and we should return with the same.
1106 * The parent has BUSY'd the pages. We should return with 'm'
1107 * left busy, but the others adjusted.
1110 swap_pager_getpages(vm_object_t object, vm_page_t *m, int count, int reqpage)
1120 KASSERT(mreq->object == object,
1121 ("swap_pager_getpages: object mismatch %p/%p",
1122 object, mreq->object));
1125 * Calculate range to retrieve. The pages have already been assigned
1126 * their swapblks. We require a *contiguous* range but we know it to
1127 * not span devices. If we do not supply it, bad things
1128 * happen. Note that blk, iblk & jblk can be SWAPBLK_NONE, but the
1129 * loops are set up such that the case(s) are handled implicitly.
1131 * The swp_*() calls must be made with the object locked.
1133 blk = swp_pager_meta_ctl(mreq->object, mreq->pindex, 0);
1135 for (i = reqpage - 1; i >= 0; --i) {
1138 iblk = swp_pager_meta_ctl(m[i]->object, m[i]->pindex, 0);
1139 if (blk != iblk + (reqpage - i))
1144 for (j = reqpage + 1; j < count; ++j) {
1147 jblk = swp_pager_meta_ctl(m[j]->object, m[j]->pindex, 0);
1148 if (blk != jblk - (j - reqpage))
1153 * free pages outside our collection range. Note: we never free
1154 * mreq, it must remain busy throughout.
1156 if (0 < i || j < count) {
1159 for (k = 0; k < i; ++k)
1160 swp_pager_free_nrpage(m[k]);
1161 for (k = j; k < count; ++k)
1162 swp_pager_free_nrpage(m[k]);
1166 * Return VM_PAGER_FAIL if we have nothing to do. Return mreq
1167 * still busy, but the others unbusied.
1169 if (blk == SWAPBLK_NONE)
1170 return (VM_PAGER_FAIL);
1173 * Getpbuf() can sleep.
1175 VM_OBJECT_WUNLOCK(object);
1177 * Get a swap buffer header to perform the IO
1179 bp = getpbuf(&nsw_rcount);
1180 bp->b_flags |= B_PAGING;
1182 bp->b_iocmd = BIO_READ;
1183 bp->b_iodone = swp_pager_async_iodone;
1184 bp->b_rcred = crhold(thread0.td_ucred);
1185 bp->b_wcred = crhold(thread0.td_ucred);
1186 bp->b_blkno = blk - (reqpage - i);
1187 bp->b_bcount = PAGE_SIZE * (j - i);
1188 bp->b_bufsize = PAGE_SIZE * (j - i);
1189 bp->b_pager.pg_reqpage = reqpage - i;
1191 VM_OBJECT_WLOCK(object);
1195 for (k = i; k < j; ++k) {
1196 bp->b_pages[k - i] = m[k];
1197 m[k]->oflags |= VPO_SWAPINPROG;
1200 bp->b_npages = j - i;
1202 PCPU_INC(cnt.v_swapin);
1203 PCPU_ADD(cnt.v_swappgsin, bp->b_npages);
1206 * We still hold the lock on mreq, and our automatic completion routine
1207 * does not remove it.
1209 vm_object_pip_add(object, bp->b_npages);
1210 VM_OBJECT_WUNLOCK(object);
1213 * perform the I/O. NOTE!!! bp cannot be considered valid after
1214 * this point because we automatically release it on completion.
1215 * Instead, we look at the one page we are interested in which we
1216 * still hold a lock on even through the I/O completion.
1218 * The other pages in our m[] array are also released on completion,
1219 * so we cannot assume they are valid anymore either.
1221 * NOTE: b_blkno is destroyed by the call to swapdev_strategy
1224 swp_pager_strategy(bp);
1227 * wait for the page we want to complete. VPO_SWAPINPROG is always
1228 * cleared on completion. If an I/O error occurs, SWAPBLK_NONE
1229 * is set in the meta-data.
1231 VM_OBJECT_WLOCK(object);
1232 while ((mreq->oflags & VPO_SWAPINPROG) != 0) {
1233 mreq->oflags |= VPO_SWAPSLEEP;
1234 PCPU_INC(cnt.v_intrans);
1235 if (VM_OBJECT_SLEEP(object, &object->paging_in_progress, PSWP,
1236 "swread", hz * 20)) {
1238 "swap_pager: indefinite wait buffer: bufobj: %p, blkno: %jd, size: %ld\n",
1239 bp->b_bufobj, (intmax_t)bp->b_blkno, bp->b_bcount);
1244 * mreq is left busied after completion, but all the other pages
1245 * are freed. If we had an unrecoverable read error the page will
1248 if (mreq->valid != VM_PAGE_BITS_ALL) {
1249 return (VM_PAGER_ERROR);
1251 return (VM_PAGER_OK);
1255 * A final note: in a low swap situation, we cannot deallocate swap
1256 * and mark a page dirty here because the caller is likely to mark
1257 * the page clean when we return, causing the page to possibly revert
1258 * to all-zero's later.
1263 * swap_pager_getpages_async():
1265 * Right now this is emulation of asynchronous operation on top of
1266 * swap_pager_getpages().
1269 swap_pager_getpages_async(vm_object_t object, vm_page_t *m, int count,
1270 int reqpage, pgo_getpages_iodone_t iodone, void *arg)
1274 r = swap_pager_getpages(object, m, count, reqpage);
1275 VM_OBJECT_WUNLOCK(object);
1280 case VM_PAGER_ERROR:
1287 panic("unhandled swap_pager_getpages() error %d", r);
1289 (iodone)(arg, m, count, error);
1290 VM_OBJECT_WLOCK(object);
1296 * swap_pager_putpages:
1298 * Assign swap (if necessary) and initiate I/O on the specified pages.
1300 * We support both OBJT_DEFAULT and OBJT_SWAP objects. DEFAULT objects
1301 * are automatically converted to SWAP objects.
1303 * In a low memory situation we may block in VOP_STRATEGY(), but the new
1304 * vm_page reservation system coupled with properly written VFS devices
1305 * should ensure that no low-memory deadlock occurs. This is an area
1308 * The parent has N vm_object_pip_add() references prior to
1309 * calling us and will remove references for rtvals[] that are
1310 * not set to VM_PAGER_PEND. We need to remove the rest on I/O
1313 * The parent has soft-busy'd the pages it passes us and will unbusy
1314 * those whos rtvals[] entry is not set to VM_PAGER_PEND on return.
1315 * We need to unbusy the rest on I/O completion.
1318 swap_pager_putpages(vm_object_t object, vm_page_t *m, int count,
1319 int flags, int *rtvals)
1324 if (count && m[0]->object != object) {
1325 panic("swap_pager_putpages: object mismatch %p/%p",
1334 * Turn object into OBJT_SWAP
1335 * check for bogus sysops
1336 * force sync if not pageout process
1338 if (object->type != OBJT_SWAP)
1339 swp_pager_meta_build(object, 0, SWAPBLK_NONE);
1340 VM_OBJECT_WUNLOCK(object);
1343 if (curproc != pageproc)
1346 sync = (flags & VM_PAGER_PUT_SYNC) != 0;
1351 * Update nsw parameters from swap_async_max sysctl values.
1352 * Do not let the sysop crash the machine with bogus numbers.
1354 mtx_lock(&pbuf_mtx);
1355 if (swap_async_max != nsw_wcount_async_max) {
1361 if ((n = swap_async_max) > nswbuf / 2)
1368 * Adjust difference ( if possible ). If the current async
1369 * count is too low, we may not be able to make the adjustment
1372 n -= nsw_wcount_async_max;
1373 if (nsw_wcount_async + n >= 0) {
1374 nsw_wcount_async += n;
1375 nsw_wcount_async_max += n;
1376 wakeup(&nsw_wcount_async);
1379 mtx_unlock(&pbuf_mtx);
1384 * Assign swap blocks and issue I/O. We reallocate swap on the fly.
1385 * The page is left dirty until the pageout operation completes
1388 for (i = 0; i < count; i += n) {
1394 * Maximum I/O size is limited by a number of factors.
1396 n = min(BLIST_MAX_ALLOC, count - i);
1397 n = min(n, nsw_cluster_max);
1400 * Get biggest block of swap we can. If we fail, fall
1401 * back and try to allocate a smaller block. Don't go
1402 * overboard trying to allocate space if it would overly
1406 (blk = swp_pager_getswapspace(n)) == SWAPBLK_NONE &&
1411 if (blk == SWAPBLK_NONE) {
1412 for (j = 0; j < n; ++j)
1413 rtvals[i+j] = VM_PAGER_FAIL;
1418 * All I/O parameters have been satisfied, build the I/O
1419 * request and assign the swap space.
1422 bp = getpbuf(&nsw_wcount_sync);
1424 bp = getpbuf(&nsw_wcount_async);
1425 bp->b_flags = B_ASYNC;
1427 bp->b_flags |= B_PAGING;
1428 bp->b_iocmd = BIO_WRITE;
1430 bp->b_rcred = crhold(thread0.td_ucred);
1431 bp->b_wcred = crhold(thread0.td_ucred);
1432 bp->b_bcount = PAGE_SIZE * n;
1433 bp->b_bufsize = PAGE_SIZE * n;
1436 VM_OBJECT_WLOCK(object);
1437 for (j = 0; j < n; ++j) {
1438 vm_page_t mreq = m[i+j];
1440 swp_pager_meta_build(
1445 vm_page_dirty(mreq);
1446 rtvals[i+j] = VM_PAGER_OK;
1448 mreq->oflags |= VPO_SWAPINPROG;
1449 bp->b_pages[j] = mreq;
1451 VM_OBJECT_WUNLOCK(object);
1454 * Must set dirty range for NFS to work.
1457 bp->b_dirtyend = bp->b_bcount;
1459 PCPU_INC(cnt.v_swapout);
1460 PCPU_ADD(cnt.v_swappgsout, bp->b_npages);
1465 * NOTE: b_blkno is destroyed by the call to swapdev_strategy
1467 if (sync == FALSE) {
1468 bp->b_iodone = swp_pager_async_iodone;
1470 swp_pager_strategy(bp);
1472 for (j = 0; j < n; ++j)
1473 rtvals[i+j] = VM_PAGER_PEND;
1474 /* restart outter loop */
1481 * NOTE: b_blkno is destroyed by the call to swapdev_strategy
1483 bp->b_iodone = bdone;
1484 swp_pager_strategy(bp);
1487 * Wait for the sync I/O to complete, then update rtvals.
1488 * We just set the rtvals[] to VM_PAGER_PEND so we can call
1489 * our async completion routine at the end, thus avoiding a
1492 bwait(bp, PVM, "swwrt");
1493 for (j = 0; j < n; ++j)
1494 rtvals[i+j] = VM_PAGER_PEND;
1496 * Now that we are through with the bp, we can call the
1497 * normal async completion, which frees everything up.
1499 swp_pager_async_iodone(bp);
1501 VM_OBJECT_WLOCK(object);
1505 * swp_pager_async_iodone:
1507 * Completion routine for asynchronous reads and writes from/to swap.
1508 * Also called manually by synchronous code to finish up a bp.
1510 * This routine may not sleep.
1513 swp_pager_async_iodone(struct buf *bp)
1516 vm_object_t object = NULL;
1521 if (bp->b_ioflags & BIO_ERROR) {
1523 "swap_pager: I/O error - %s failed; blkno %ld,"
1524 "size %ld, error %d\n",
1525 ((bp->b_iocmd == BIO_READ) ? "pagein" : "pageout"),
1533 * remove the mapping for kernel virtual
1535 if ((bp->b_flags & B_UNMAPPED) != 0) {
1536 bp->b_data = bp->b_kvaalloc;
1537 bp->b_kvabase = bp->b_kvaalloc;
1538 bp->b_flags &= ~B_UNMAPPED;
1540 pmap_qremove((vm_offset_t)bp->b_data, bp->b_npages);
1543 object = bp->b_pages[0]->object;
1544 VM_OBJECT_WLOCK(object);
1548 * cleanup pages. If an error occurs writing to swap, we are in
1549 * very serious trouble. If it happens to be a disk error, though,
1550 * we may be able to recover by reassigning the swap later on. So
1551 * in this case we remove the m->swapblk assignment for the page
1552 * but do not free it in the rlist. The errornous block(s) are thus
1553 * never reallocated as swap. Redirty the page and continue.
1555 for (i = 0; i < bp->b_npages; ++i) {
1556 vm_page_t m = bp->b_pages[i];
1558 m->oflags &= ~VPO_SWAPINPROG;
1559 if (m->oflags & VPO_SWAPSLEEP) {
1560 m->oflags &= ~VPO_SWAPSLEEP;
1561 wakeup(&object->paging_in_progress);
1564 if (bp->b_ioflags & BIO_ERROR) {
1566 * If an error occurs I'd love to throw the swapblk
1567 * away without freeing it back to swapspace, so it
1568 * can never be used again. But I can't from an
1571 if (bp->b_iocmd == BIO_READ) {
1573 * When reading, reqpage needs to stay
1574 * locked for the parent, but all other
1575 * pages can be freed. We still want to
1576 * wakeup the parent waiting on the page,
1577 * though. ( also: pg_reqpage can be -1 and
1578 * not match anything ).
1580 * We have to wake specifically requested pages
1581 * up too because we cleared VPO_SWAPINPROG and
1582 * someone may be waiting for that.
1584 * NOTE: for reads, m->dirty will probably
1585 * be overridden by the original caller of
1586 * getpages so don't play cute tricks here.
1589 if (i != bp->b_pager.pg_reqpage)
1590 swp_pager_free_nrpage(m);
1597 * If i == bp->b_pager.pg_reqpage, do not wake
1598 * the page up. The caller needs to.
1602 * If a write error occurs, reactivate page
1603 * so it doesn't clog the inactive list,
1604 * then finish the I/O.
1608 vm_page_activate(m);
1612 } else if (bp->b_iocmd == BIO_READ) {
1614 * NOTE: for reads, m->dirty will probably be
1615 * overridden by the original caller of getpages so
1616 * we cannot set them in order to free the underlying
1617 * swap in a low-swap situation. I don't think we'd
1618 * want to do that anyway, but it was an optimization
1619 * that existed in the old swapper for a time before
1620 * it got ripped out due to precisely this problem.
1622 * If not the requested page then deactivate it.
1624 * Note that the requested page, reqpage, is left
1625 * busied, but we still have to wake it up. The
1626 * other pages are released (unbusied) by
1627 * vm_page_xunbusy().
1629 KASSERT(!pmap_page_is_mapped(m),
1630 ("swp_pager_async_iodone: page %p is mapped", m));
1631 m->valid = VM_PAGE_BITS_ALL;
1632 KASSERT(m->dirty == 0,
1633 ("swp_pager_async_iodone: page %p is dirty", m));
1636 * We have to wake specifically requested pages
1637 * up too because we cleared VPO_SWAPINPROG and
1638 * could be waiting for it in getpages. However,
1639 * be sure to not unbusy getpages specifically
1640 * requested page - getpages expects it to be
1643 if (i != bp->b_pager.pg_reqpage) {
1645 vm_page_deactivate(m);
1655 * For write success, clear the dirty
1656 * status, then finish the I/O ( which decrements the
1657 * busy count and possibly wakes waiter's up ).
1659 KASSERT(!pmap_page_is_write_mapped(m),
1660 ("swp_pager_async_iodone: page %p is not write"
1664 if (vm_page_count_severe()) {
1666 vm_page_try_to_cache(m);
1673 * adjust pip. NOTE: the original parent may still have its own
1674 * pip refs on the object.
1676 if (object != NULL) {
1677 vm_object_pip_wakeupn(object, bp->b_npages);
1678 VM_OBJECT_WUNLOCK(object);
1682 * swapdev_strategy() manually sets b_vp and b_bufobj before calling
1683 * bstrategy(). Set them back to NULL now we're done with it, or we'll
1684 * trigger a KASSERT in relpbuf().
1688 bp->b_bufobj = NULL;
1691 * release the physical I/O buffer
1695 ((bp->b_iocmd == BIO_READ) ? &nsw_rcount :
1696 ((bp->b_flags & B_ASYNC) ?
1705 * swap_pager_isswapped:
1707 * Return 1 if at least one page in the given object is paged
1708 * out to the given swap device.
1710 * This routine may not sleep.
1713 swap_pager_isswapped(vm_object_t object, struct swdevt *sp)
1719 VM_OBJECT_ASSERT_WLOCKED(object);
1720 if (object->type != OBJT_SWAP)
1723 mtx_lock(&swhash_mtx);
1724 for (bcount = 0; bcount < object->un_pager.swp.swp_bcount; bcount++) {
1725 struct swblock *swap;
1727 if ((swap = *swp_pager_hash(object, index)) != NULL) {
1728 for (i = 0; i < SWAP_META_PAGES; ++i) {
1729 if (swp_pager_isondev(swap->swb_pages[i], sp)) {
1730 mtx_unlock(&swhash_mtx);
1735 index += SWAP_META_PAGES;
1737 mtx_unlock(&swhash_mtx);
1742 * SWP_PAGER_FORCE_PAGEIN() - force a swap block to be paged in
1744 * This routine dissociates the page at the given index within a
1745 * swap block from its backing store, paging it in if necessary.
1746 * If the page is paged in, it is placed in the inactive queue,
1747 * since it had its backing store ripped out from under it.
1748 * We also attempt to swap in all other pages in the swap block,
1749 * we only guarantee that the one at the specified index is
1752 * XXX - The code to page the whole block in doesn't work, so we
1753 * revert to the one-by-one behavior for now. Sigh.
1756 swp_pager_force_pagein(vm_object_t object, vm_pindex_t pindex)
1760 vm_object_pip_add(object, 1);
1761 m = vm_page_grab(object, pindex, VM_ALLOC_NORMAL);
1762 if (m->valid == VM_PAGE_BITS_ALL) {
1763 vm_object_pip_wakeup(object);
1766 vm_page_activate(m);
1769 vm_pager_page_unswapped(m);
1773 if (swap_pager_getpages(object, &m, 1, 0) != VM_PAGER_OK)
1774 panic("swap_pager_force_pagein: read from swap failed");/*XXX*/
1775 vm_object_pip_wakeup(object);
1778 vm_page_deactivate(m);
1781 vm_pager_page_unswapped(m);
1785 * swap_pager_swapoff:
1787 * Page in all of the pages that have been paged out to the
1788 * given device. The corresponding blocks in the bitmap must be
1789 * marked as allocated and the device must be flagged SW_CLOSING.
1790 * There may be no processes swapped out to the device.
1792 * This routine may block.
1795 swap_pager_swapoff(struct swdevt *sp)
1797 struct swblock *swap;
1804 mtx_lock(&swhash_mtx);
1805 for (i = 0; i <= swhash_mask; i++) { /* '<=' is correct here */
1807 for (swap = swhash[i]; swap != NULL; swap = swap->swb_hnext) {
1808 vm_object_t object = swap->swb_object;
1809 vm_pindex_t pindex = swap->swb_index;
1810 for (j = 0; j < SWAP_META_PAGES; ++j) {
1811 if (swp_pager_isondev(swap->swb_pages[j], sp)) {
1812 /* avoid deadlock */
1813 if (!VM_OBJECT_TRYWLOCK(object)) {
1816 mtx_unlock(&swhash_mtx);
1817 swp_pager_force_pagein(object,
1819 VM_OBJECT_WUNLOCK(object);
1820 mtx_lock(&swhash_mtx);
1827 mtx_unlock(&swhash_mtx);
1830 * Objects may be locked or paging to the device being
1831 * removed, so we will miss their pages and need to
1832 * make another pass. We have marked this device as
1833 * SW_CLOSING, so the activity should finish soon.
1836 if (retries > 100) {
1837 panic("swapoff: failed to locate %d swap blocks",
1840 pause("swpoff", hz / 20);
1845 /************************************************************************
1847 ************************************************************************
1849 * These routines manipulate the swap metadata stored in the
1852 * Swap metadata is implemented with a global hash and not directly
1853 * linked into the object. Instead the object simply contains
1854 * appropriate tracking counters.
1858 * SWP_PAGER_META_BUILD() - add swap block to swap meta data for object
1860 * We first convert the object to a swap object if it is a default
1863 * The specified swapblk is added to the object's swap metadata. If
1864 * the swapblk is not valid, it is freed instead. Any previously
1865 * assigned swapblk is freed.
1868 swp_pager_meta_build(vm_object_t object, vm_pindex_t pindex, daddr_t swapblk)
1870 static volatile int exhausted;
1871 struct swblock *swap;
1872 struct swblock **pswap;
1875 VM_OBJECT_ASSERT_WLOCKED(object);
1877 * Convert default object to swap object if necessary
1879 if (object->type != OBJT_SWAP) {
1880 object->type = OBJT_SWAP;
1881 object->un_pager.swp.swp_bcount = 0;
1883 if (object->handle != NULL) {
1884 mtx_lock(&sw_alloc_mtx);
1886 NOBJLIST(object->handle),
1890 mtx_unlock(&sw_alloc_mtx);
1895 * Locate hash entry. If not found create, but if we aren't adding
1896 * anything just return. If we run out of space in the map we wait
1897 * and, since the hash table may have changed, retry.
1900 mtx_lock(&swhash_mtx);
1901 pswap = swp_pager_hash(object, pindex);
1903 if ((swap = *pswap) == NULL) {
1906 if (swapblk == SWAPBLK_NONE)
1909 swap = *pswap = uma_zalloc(swap_zone, M_NOWAIT |
1910 (curproc == pageproc ? M_USE_RESERVE : 0));
1912 mtx_unlock(&swhash_mtx);
1913 VM_OBJECT_WUNLOCK(object);
1914 if (uma_zone_exhausted(swap_zone)) {
1915 if (atomic_cmpset_int(&exhausted, 0, 1))
1916 printf("swap zone exhausted, "
1917 "increase kern.maxswzone\n");
1918 vm_pageout_oom(VM_OOM_SWAPZ);
1919 pause("swzonex", 10);
1922 VM_OBJECT_WLOCK(object);
1926 if (atomic_cmpset_int(&exhausted, 1, 0))
1927 printf("swap zone ok\n");
1929 swap->swb_hnext = NULL;
1930 swap->swb_object = object;
1931 swap->swb_index = pindex & ~(vm_pindex_t)SWAP_META_MASK;
1932 swap->swb_count = 0;
1934 ++object->un_pager.swp.swp_bcount;
1936 for (i = 0; i < SWAP_META_PAGES; ++i)
1937 swap->swb_pages[i] = SWAPBLK_NONE;
1941 * Delete prior contents of metadata
1943 idx = pindex & SWAP_META_MASK;
1945 if (swap->swb_pages[idx] != SWAPBLK_NONE) {
1946 swp_pager_freeswapspace(swap->swb_pages[idx], 1);
1951 * Enter block into metadata
1953 swap->swb_pages[idx] = swapblk;
1954 if (swapblk != SWAPBLK_NONE)
1957 mtx_unlock(&swhash_mtx);
1961 * SWP_PAGER_META_FREE() - free a range of blocks in the object's swap metadata
1963 * The requested range of blocks is freed, with any associated swap
1964 * returned to the swap bitmap.
1966 * This routine will free swap metadata structures as they are cleaned
1967 * out. This routine does *NOT* operate on swap metadata associated
1968 * with resident pages.
1971 swp_pager_meta_free(vm_object_t object, vm_pindex_t index, daddr_t count)
1974 VM_OBJECT_ASSERT_LOCKED(object);
1975 if (object->type != OBJT_SWAP)
1979 struct swblock **pswap;
1980 struct swblock *swap;
1982 mtx_lock(&swhash_mtx);
1983 pswap = swp_pager_hash(object, index);
1985 if ((swap = *pswap) != NULL) {
1986 daddr_t v = swap->swb_pages[index & SWAP_META_MASK];
1988 if (v != SWAPBLK_NONE) {
1989 swp_pager_freeswapspace(v, 1);
1990 swap->swb_pages[index & SWAP_META_MASK] =
1992 if (--swap->swb_count == 0) {
1993 *pswap = swap->swb_hnext;
1994 uma_zfree(swap_zone, swap);
1995 --object->un_pager.swp.swp_bcount;
2001 int n = SWAP_META_PAGES - (index & SWAP_META_MASK);
2005 mtx_unlock(&swhash_mtx);
2010 * SWP_PAGER_META_FREE_ALL() - destroy all swap metadata associated with object
2012 * This routine locates and destroys all swap metadata associated with
2016 swp_pager_meta_free_all(vm_object_t object)
2020 VM_OBJECT_ASSERT_WLOCKED(object);
2021 if (object->type != OBJT_SWAP)
2024 while (object->un_pager.swp.swp_bcount) {
2025 struct swblock **pswap;
2026 struct swblock *swap;
2028 mtx_lock(&swhash_mtx);
2029 pswap = swp_pager_hash(object, index);
2030 if ((swap = *pswap) != NULL) {
2033 for (i = 0; i < SWAP_META_PAGES; ++i) {
2034 daddr_t v = swap->swb_pages[i];
2035 if (v != SWAPBLK_NONE) {
2037 swp_pager_freeswapspace(v, 1);
2040 if (swap->swb_count != 0)
2041 panic("swap_pager_meta_free_all: swb_count != 0");
2042 *pswap = swap->swb_hnext;
2043 uma_zfree(swap_zone, swap);
2044 --object->un_pager.swp.swp_bcount;
2046 mtx_unlock(&swhash_mtx);
2047 index += SWAP_META_PAGES;
2052 * SWP_PAGER_METACTL() - misc control of swap and vm_page_t meta data.
2054 * This routine is capable of looking up, popping, or freeing
2055 * swapblk assignments in the swap meta data or in the vm_page_t.
2056 * The routine typically returns the swapblk being looked-up, or popped,
2057 * or SWAPBLK_NONE if the block was freed, or SWAPBLK_NONE if the block
2058 * was invalid. This routine will automatically free any invalid
2059 * meta-data swapblks.
2061 * It is not possible to store invalid swapblks in the swap meta data
2062 * (other then a literal 'SWAPBLK_NONE'), so we don't bother checking.
2064 * When acting on a busy resident page and paging is in progress, we
2065 * have to wait until paging is complete but otherwise can act on the
2068 * SWM_FREE remove and free swap block from metadata
2069 * SWM_POP remove from meta data but do not free.. pop it out
2072 swp_pager_meta_ctl(vm_object_t object, vm_pindex_t pindex, int flags)
2074 struct swblock **pswap;
2075 struct swblock *swap;
2079 VM_OBJECT_ASSERT_LOCKED(object);
2081 * The meta data only exists of the object is OBJT_SWAP
2082 * and even then might not be allocated yet.
2084 if (object->type != OBJT_SWAP)
2085 return (SWAPBLK_NONE);
2088 mtx_lock(&swhash_mtx);
2089 pswap = swp_pager_hash(object, pindex);
2091 if ((swap = *pswap) != NULL) {
2092 idx = pindex & SWAP_META_MASK;
2093 r1 = swap->swb_pages[idx];
2095 if (r1 != SWAPBLK_NONE) {
2096 if (flags & SWM_FREE) {
2097 swp_pager_freeswapspace(r1, 1);
2100 if (flags & (SWM_FREE|SWM_POP)) {
2101 swap->swb_pages[idx] = SWAPBLK_NONE;
2102 if (--swap->swb_count == 0) {
2103 *pswap = swap->swb_hnext;
2104 uma_zfree(swap_zone, swap);
2105 --object->un_pager.swp.swp_bcount;
2110 mtx_unlock(&swhash_mtx);
2115 * System call swapon(name) enables swapping on device name,
2116 * which must be in the swdevsw. Return EBUSY
2117 * if already swapping on this device.
2119 #ifndef _SYS_SYSPROTO_H_
2120 struct swapon_args {
2130 sys_swapon(struct thread *td, struct swapon_args *uap)
2134 struct nameidata nd;
2137 error = priv_check(td, PRIV_SWAPON);
2142 while (swdev_syscall_active)
2143 tsleep(&swdev_syscall_active, PUSER - 1, "swpon", 0);
2144 swdev_syscall_active = 1;
2147 * Swap metadata may not fit in the KVM if we have physical
2150 if (swap_zone == NULL) {
2155 NDINIT(&nd, LOOKUP, ISOPEN | FOLLOW | AUDITVNODE1, UIO_USERSPACE,
2161 NDFREE(&nd, NDF_ONLY_PNBUF);
2164 if (vn_isdisk(vp, &error)) {
2165 error = swapongeom(td, vp);
2166 } else if (vp->v_type == VREG &&
2167 (vp->v_mount->mnt_vfc->vfc_flags & VFCF_NETWORK) != 0 &&
2168 (error = VOP_GETATTR(vp, &attr, td->td_ucred)) == 0) {
2170 * Allow direct swapping to NFS regular files in the same
2171 * way that nfs_mountroot() sets up diskless swapping.
2173 error = swaponvp(td, vp, attr.va_size / DEV_BSIZE);
2179 swdev_syscall_active = 0;
2180 wakeup_one(&swdev_syscall_active);
2186 * Check that the total amount of swap currently configured does not
2187 * exceed half the theoretical maximum. If it does, print a warning
2188 * message and return -1; otherwise, return 0.
2191 swapon_check_swzone(unsigned long npages)
2193 unsigned long maxpages;
2195 /* absolute maximum we can handle assuming 100% efficiency */
2196 maxpages = uma_zone_get_max(swap_zone) * SWAP_META_PAGES;
2198 /* recommend using no more than half that amount */
2199 if (npages > maxpages / 2) {
2200 printf("warning: total configured swap (%lu pages) "
2201 "exceeds maximum recommended amount (%lu pages).\n",
2202 npages, maxpages / 2);
2203 printf("warning: increase kern.maxswzone "
2204 "or reduce amount of swap.\n");
2211 swaponsomething(struct vnode *vp, void *id, u_long nblks,
2212 sw_strategy_t *strategy, sw_close_t *close, dev_t dev, int flags)
2214 struct swdevt *sp, *tsp;
2219 * nblks is in DEV_BSIZE'd chunks, convert to PAGE_SIZE'd chunks.
2220 * First chop nblks off to page-align it, then convert.
2222 * sw->sw_nblks is in page-sized chunks now too.
2224 nblks &= ~(ctodb(1) - 1);
2225 nblks = dbtoc(nblks);
2228 * If we go beyond this, we get overflows in the radix
2231 mblocks = 0x40000000 / BLIST_META_RADIX;
2232 if (nblks > mblocks) {
2234 "WARNING: reducing swap size to maximum of %luMB per unit\n",
2235 mblocks / 1024 / 1024 * PAGE_SIZE);
2239 sp = malloc(sizeof *sp, M_VMPGDATA, M_WAITOK | M_ZERO);
2244 sp->sw_nblks = nblks;
2246 sp->sw_strategy = strategy;
2247 sp->sw_close = close;
2248 sp->sw_flags = flags;
2250 sp->sw_blist = blist_create(nblks, M_WAITOK);
2252 * Do not free the first two block in order to avoid overwriting
2253 * any bsd label at the front of the partition
2255 blist_free(sp->sw_blist, 2, nblks - 2);
2258 mtx_lock(&sw_dev_mtx);
2259 TAILQ_FOREACH(tsp, &swtailq, sw_list) {
2260 if (tsp->sw_end >= dvbase) {
2262 * We put one uncovered page between the devices
2263 * in order to definitively prevent any cross-device
2266 dvbase = tsp->sw_end + 1;
2269 sp->sw_first = dvbase;
2270 sp->sw_end = dvbase + nblks;
2271 TAILQ_INSERT_TAIL(&swtailq, sp, sw_list);
2273 swap_pager_avail += nblks;
2274 swap_total += (vm_ooffset_t)nblks * PAGE_SIZE;
2275 swapon_check_swzone(swap_total / PAGE_SIZE);
2277 mtx_unlock(&sw_dev_mtx);
2281 * SYSCALL: swapoff(devname)
2283 * Disable swapping on the given device.
2285 * XXX: Badly designed system call: it should use a device index
2286 * rather than filename as specification. We keep sw_vp around
2287 * only to make this work.
2289 #ifndef _SYS_SYSPROTO_H_
2290 struct swapoff_args {
2300 sys_swapoff(struct thread *td, struct swapoff_args *uap)
2303 struct nameidata nd;
2307 error = priv_check(td, PRIV_SWAPOFF);
2312 while (swdev_syscall_active)
2313 tsleep(&swdev_syscall_active, PUSER - 1, "swpoff", 0);
2314 swdev_syscall_active = 1;
2316 NDINIT(&nd, LOOKUP, FOLLOW | AUDITVNODE1, UIO_USERSPACE, uap->name,
2321 NDFREE(&nd, NDF_ONLY_PNBUF);
2324 mtx_lock(&sw_dev_mtx);
2325 TAILQ_FOREACH(sp, &swtailq, sw_list) {
2326 if (sp->sw_vp == vp)
2329 mtx_unlock(&sw_dev_mtx);
2334 error = swapoff_one(sp, td->td_ucred);
2336 swdev_syscall_active = 0;
2337 wakeup_one(&swdev_syscall_active);
2343 swapoff_one(struct swdevt *sp, struct ucred *cred)
2345 u_long nblks, dvbase;
2350 mtx_assert(&Giant, MA_OWNED);
2352 (void) vn_lock(sp->sw_vp, LK_EXCLUSIVE | LK_RETRY);
2353 error = mac_system_check_swapoff(cred, sp->sw_vp);
2354 (void) VOP_UNLOCK(sp->sw_vp, 0);
2358 nblks = sp->sw_nblks;
2361 * We can turn off this swap device safely only if the
2362 * available virtual memory in the system will fit the amount
2363 * of data we will have to page back in, plus an epsilon so
2364 * the system doesn't become critically low on swap space.
2366 if (vm_cnt.v_free_count + vm_cnt.v_cache_count + swap_pager_avail <
2367 nblks + nswap_lowat) {
2372 * Prevent further allocations on this device.
2374 mtx_lock(&sw_dev_mtx);
2375 sp->sw_flags |= SW_CLOSING;
2376 for (dvbase = 0; dvbase < sp->sw_end; dvbase += dmmax) {
2377 swap_pager_avail -= blist_fill(sp->sw_blist,
2380 swap_total -= (vm_ooffset_t)nblks * PAGE_SIZE;
2381 mtx_unlock(&sw_dev_mtx);
2384 * Page in the contents of the device and close it.
2386 swap_pager_swapoff(sp);
2388 sp->sw_close(curthread, sp);
2390 mtx_lock(&sw_dev_mtx);
2391 TAILQ_REMOVE(&swtailq, sp, sw_list);
2393 if (nswapdev == 0) {
2394 swap_pager_full = 2;
2395 swap_pager_almost_full = 1;
2399 mtx_unlock(&sw_dev_mtx);
2400 blist_destroy(sp->sw_blist);
2401 free(sp, M_VMPGDATA);
2408 struct swdevt *sp, *spt;
2409 const char *devname;
2413 while (swdev_syscall_active)
2414 tsleep(&swdev_syscall_active, PUSER - 1, "swpoff", 0);
2415 swdev_syscall_active = 1;
2417 mtx_lock(&sw_dev_mtx);
2418 TAILQ_FOREACH_SAFE(sp, &swtailq, sw_list, spt) {
2419 mtx_unlock(&sw_dev_mtx);
2420 if (vn_isdisk(sp->sw_vp, NULL))
2421 devname = devtoname(sp->sw_vp->v_rdev);
2424 error = swapoff_one(sp, thread0.td_ucred);
2426 printf("Cannot remove swap device %s (error=%d), "
2427 "skipping.\n", devname, error);
2428 } else if (bootverbose) {
2429 printf("Swap device %s removed.\n", devname);
2431 mtx_lock(&sw_dev_mtx);
2433 mtx_unlock(&sw_dev_mtx);
2435 swdev_syscall_active = 0;
2436 wakeup_one(&swdev_syscall_active);
2441 swap_pager_status(int *total, int *used)
2447 mtx_lock(&sw_dev_mtx);
2448 TAILQ_FOREACH(sp, &swtailq, sw_list) {
2449 *total += sp->sw_nblks;
2450 *used += sp->sw_used;
2452 mtx_unlock(&sw_dev_mtx);
2456 swap_dev_info(int name, struct xswdev *xs, char *devname, size_t len)
2459 const char *tmp_devname;
2464 mtx_lock(&sw_dev_mtx);
2465 TAILQ_FOREACH(sp, &swtailq, sw_list) {
2470 xs->xsw_version = XSWDEV_VERSION;
2471 xs->xsw_dev = sp->sw_dev;
2472 xs->xsw_flags = sp->sw_flags;
2473 xs->xsw_nblks = sp->sw_nblks;
2474 xs->xsw_used = sp->sw_used;
2475 if (devname != NULL) {
2476 if (vn_isdisk(sp->sw_vp, NULL))
2477 tmp_devname = devtoname(sp->sw_vp->v_rdev);
2479 tmp_devname = "[file]";
2480 strncpy(devname, tmp_devname, len);
2485 mtx_unlock(&sw_dev_mtx);
2490 sysctl_vm_swap_info(SYSCTL_HANDLER_ARGS)
2495 if (arg2 != 1) /* name length */
2497 error = swap_dev_info(*(int *)arg1, &xs, NULL, 0);
2500 error = SYSCTL_OUT(req, &xs, sizeof(xs));
2504 SYSCTL_INT(_vm, OID_AUTO, nswapdev, CTLFLAG_RD, &nswapdev, 0,
2505 "Number of swap devices");
2506 SYSCTL_NODE(_vm, OID_AUTO, swap_info, CTLFLAG_RD, sysctl_vm_swap_info,
2507 "Swap statistics by device");
2510 * vmspace_swap_count() - count the approximate swap usage in pages for a
2513 * The map must be locked.
2515 * Swap usage is determined by taking the proportional swap used by
2516 * VM objects backing the VM map. To make up for fractional losses,
2517 * if the VM object has any swap use at all the associated map entries
2518 * count for at least 1 swap page.
2521 vmspace_swap_count(struct vmspace *vmspace)
2528 map = &vmspace->vm_map;
2531 for (cur = map->header.next; cur != &map->header; cur = cur->next) {
2532 if ((cur->eflags & MAP_ENTRY_IS_SUB_MAP) == 0 &&
2533 (object = cur->object.vm_object) != NULL) {
2534 VM_OBJECT_WLOCK(object);
2535 if (object->type == OBJT_SWAP &&
2536 object->un_pager.swp.swp_bcount != 0) {
2537 n = (cur->end - cur->start) / PAGE_SIZE;
2538 count += object->un_pager.swp.swp_bcount *
2539 SWAP_META_PAGES * n / object->size + 1;
2541 VM_OBJECT_WUNLOCK(object);
2550 * Swapping onto disk devices.
2554 static g_orphan_t swapgeom_orphan;
2556 static struct g_class g_swap_class = {
2558 .version = G_VERSION,
2559 .orphan = swapgeom_orphan,
2562 DECLARE_GEOM_CLASS(g_swap_class, g_class);
2566 swapgeom_done(struct bio *bp2)
2570 bp = bp2->bio_caller2;
2571 bp->b_ioflags = bp2->bio_flags;
2573 bp->b_ioflags |= BIO_ERROR;
2574 bp->b_resid = bp->b_bcount - bp2->bio_completed;
2575 bp->b_error = bp2->bio_error;
2581 swapgeom_strategy(struct buf *bp, struct swdevt *sp)
2584 struct g_consumer *cp;
2588 bp->b_error = ENXIO;
2589 bp->b_ioflags |= BIO_ERROR;
2593 if (bp->b_iocmd == BIO_WRITE)
2596 bio = g_alloc_bio();
2598 bp->b_error = ENOMEM;
2599 bp->b_ioflags |= BIO_ERROR;
2604 bio->bio_caller2 = bp;
2605 bio->bio_cmd = bp->b_iocmd;
2606 bio->bio_offset = (bp->b_blkno - sp->sw_first) * PAGE_SIZE;
2607 bio->bio_length = bp->b_bcount;
2608 bio->bio_done = swapgeom_done;
2609 if ((bp->b_flags & B_UNMAPPED) != 0) {
2610 bio->bio_ma = bp->b_pages;
2611 bio->bio_data = unmapped_buf;
2612 bio->bio_ma_offset = (vm_offset_t)bp->b_offset & PAGE_MASK;
2613 bio->bio_ma_n = bp->b_npages;
2614 bio->bio_flags |= BIO_UNMAPPED;
2616 bio->bio_data = bp->b_data;
2619 g_io_request(bio, cp);
2624 swapgeom_orphan(struct g_consumer *cp)
2628 mtx_lock(&sw_dev_mtx);
2629 TAILQ_FOREACH(sp, &swtailq, sw_list)
2630 if (sp->sw_id == cp)
2631 sp->sw_flags |= SW_CLOSING;
2632 mtx_unlock(&sw_dev_mtx);
2636 swapgeom_close_ev(void *arg, int flags)
2638 struct g_consumer *cp;
2641 g_access(cp, -1, -1, 0);
2643 g_destroy_consumer(cp);
2647 swapgeom_close(struct thread *td, struct swdevt *sw)
2650 /* XXX: direct call when Giant untangled */
2651 g_waitfor_event(swapgeom_close_ev, sw->sw_id, M_WAITOK, NULL);
2662 swapongeom_ev(void *arg, int flags)
2665 struct g_provider *pp;
2666 struct g_consumer *cp;
2667 static struct g_geom *gp;
2674 pp = g_dev_getprovider(swh->dev);
2676 swh->error = ENODEV;
2679 mtx_lock(&sw_dev_mtx);
2680 TAILQ_FOREACH(sp, &swtailq, sw_list) {
2682 if (cp != NULL && cp->provider == pp) {
2683 mtx_unlock(&sw_dev_mtx);
2688 mtx_unlock(&sw_dev_mtx);
2690 gp = g_new_geomf(&g_swap_class, "swap");
2691 cp = g_new_consumer(gp);
2694 * XXX: Everytime you think you can improve the margin for
2695 * footshooting, somebody depends on the ability to do so:
2696 * savecore(8) wants to write to our swapdev so we cannot
2697 * set an exclusive count :-(
2699 error = g_access(cp, 1, 1, 0);
2702 g_destroy_consumer(cp);
2706 nblks = pp->mediasize / DEV_BSIZE;
2707 swaponsomething(swh->vp, cp, nblks, swapgeom_strategy,
2708 swapgeom_close, dev2udev(swh->dev),
2709 (pp->flags & G_PF_ACCEPT_UNMAPPED) != 0 ? SW_UNMAPPED : 0);
2714 swapongeom(struct thread *td, struct vnode *vp)
2719 vn_lock(vp, LK_EXCLUSIVE | LK_RETRY);
2721 swh.dev = vp->v_rdev;
2724 /* XXX: direct call when Giant untangled */
2725 error = g_waitfor_event(swapongeom_ev, &swh, M_WAITOK, NULL);
2735 * This is used mainly for network filesystem (read: probably only tested
2736 * with NFS) swapfiles.
2741 swapdev_strategy(struct buf *bp, struct swdevt *sp)
2745 bp->b_blkno = ctodb(bp->b_blkno - sp->sw_first);
2749 if (bp->b_iocmd == BIO_WRITE) {
2751 bufobj_wdrop(bp->b_bufobj);
2752 bufobj_wref(&vp2->v_bufobj);
2754 if (bp->b_bufobj != &vp2->v_bufobj)
2755 bp->b_bufobj = &vp2->v_bufobj;
2757 bp->b_iooffset = dbtob(bp->b_blkno);
2763 swapdev_close(struct thread *td, struct swdevt *sp)
2766 VOP_CLOSE(sp->sw_vp, FREAD | FWRITE, td->td_ucred, td);
2772 swaponvp(struct thread *td, struct vnode *vp, u_long nblks)
2779 mtx_lock(&sw_dev_mtx);
2780 TAILQ_FOREACH(sp, &swtailq, sw_list) {
2781 if (sp->sw_id == vp) {
2782 mtx_unlock(&sw_dev_mtx);
2786 mtx_unlock(&sw_dev_mtx);
2788 (void) vn_lock(vp, LK_EXCLUSIVE | LK_RETRY);
2790 error = mac_system_check_swapon(td->td_ucred, vp);
2793 error = VOP_OPEN(vp, FREAD | FWRITE, td->td_ucred, td, NULL);
2794 (void) VOP_UNLOCK(vp, 0);
2798 swaponsomething(vp, vp, nblks, swapdev_strategy, swapdev_close,