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");
203 error = racct_add(curproc, RACCT_SWAP, incr);
204 PROC_UNLOCK(curproc);
211 mtx_lock(&sw_dev_mtx);
212 r = swap_reserved + incr;
213 if (overcommit & SWAP_RESERVE_ALLOW_NONWIRED) {
214 s = vm_cnt.v_page_count - vm_cnt.v_free_reserved - vm_cnt.v_wire_count;
219 if ((overcommit & SWAP_RESERVE_FORCE_ON) == 0 || r <= s ||
220 (error = priv_check(curthread, PRIV_VM_SWAP_NOQUOTA)) == 0) {
224 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(curthread, RLIMIT_SWAP) &&
230 priv_check(curthread, PRIV_VM_SWAP_NORLIMIT))
233 uip->ui_vmsize += incr;
234 UIDINFO_VMSIZE_UNLOCK(uip);
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 #define SWM_FREE 0x02 /* free, period */
317 #define SWM_POP 0x04 /* pop out */
319 int swap_pager_full = 2; /* swap space exhaustion (task killing) */
320 static int swap_pager_almost_full = 1; /* swap space exhaustion (w/hysteresis)*/
321 static int nsw_rcount; /* free read buffers */
322 static int nsw_wcount_sync; /* limit write buffers / synchronous */
323 static int nsw_wcount_async; /* limit write buffers / asynchronous */
324 static int nsw_wcount_async_max;/* assigned maximum */
325 static int nsw_cluster_max; /* maximum VOP I/O allowed */
327 static int sysctl_swap_async_max(SYSCTL_HANDLER_ARGS);
328 SYSCTL_PROC(_vm, OID_AUTO, swap_async_max, CTLTYPE_INT | CTLFLAG_RW,
329 NULL, 0, sysctl_swap_async_max, "I", "Maximum running async swap ops");
331 static struct swblock **swhash;
332 static int swhash_mask;
333 static struct mtx swhash_mtx;
335 static struct sx sw_alloc_sx;
338 * "named" and "unnamed" anon region objects. Try to reduce the overhead
339 * of searching a named list by hashing it just a little.
344 #define NOBJLIST(handle) \
345 (&swap_pager_object_list[((int)(intptr_t)handle >> 4) & (NOBJLISTS-1)])
347 static struct mtx sw_alloc_mtx; /* protect list manipulation */
348 static struct pagerlst swap_pager_object_list[NOBJLISTS];
349 static uma_zone_t swap_zone;
352 * pagerops for OBJT_SWAP - "swap pager". Some ops are also global procedure
353 * calls hooked from other parts of the VM system and do not appear here.
354 * (see vm/swap_pager.h).
357 swap_pager_alloc(void *handle, vm_ooffset_t size,
358 vm_prot_t prot, vm_ooffset_t offset, struct ucred *);
359 static void swap_pager_dealloc(vm_object_t object);
360 static int swap_pager_getpages(vm_object_t, vm_page_t *, int, int *,
362 static int swap_pager_getpages_async(vm_object_t, vm_page_t *, int, int *,
363 int *, pgo_getpages_iodone_t, void *);
364 static void swap_pager_putpages(vm_object_t, vm_page_t *, int, boolean_t, int *);
366 swap_pager_haspage(vm_object_t object, vm_pindex_t pindex, int *before, int *after);
367 static void swap_pager_init(void);
368 static void swap_pager_unswapped(vm_page_t);
369 static void swap_pager_swapoff(struct swdevt *sp);
371 struct pagerops swappagerops = {
372 .pgo_init = swap_pager_init, /* early system initialization of pager */
373 .pgo_alloc = swap_pager_alloc, /* allocate an OBJT_SWAP object */
374 .pgo_dealloc = swap_pager_dealloc, /* deallocate an OBJT_SWAP object */
375 .pgo_getpages = swap_pager_getpages, /* pagein */
376 .pgo_getpages_async = swap_pager_getpages_async, /* pagein (async) */
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 * dmmax is in page-sized chunks with the new swap system. It was
384 * dev-bsized chunks in the old. dmmax is always a power of 2.
386 * swap_*() routines are externally accessible. swp_*() routines are
390 static int nswap_lowat = 128; /* in pages, swap_pager_almost_full warn */
391 static int nswap_hiwat = 512; /* in pages, swap_pager_almost_full warn */
393 SYSCTL_INT(_vm, OID_AUTO, dmmax,
394 CTLFLAG_RD, &dmmax, 0, "Maximum size of a swap block");
396 static void swp_sizecheck(void);
397 static void swp_pager_async_iodone(struct buf *bp);
398 static int swapongeom(struct thread *, struct vnode *);
399 static int swaponvp(struct thread *, struct vnode *, u_long);
400 static int swapoff_one(struct swdevt *sp, struct ucred *cred);
403 * Swap bitmap functions
405 static void swp_pager_freeswapspace(daddr_t blk, int npages);
406 static daddr_t swp_pager_getswapspace(int npages);
411 static struct swblock **swp_pager_hash(vm_object_t object, vm_pindex_t index);
412 static void swp_pager_meta_build(vm_object_t, vm_pindex_t, daddr_t);
413 static void swp_pager_meta_free(vm_object_t, vm_pindex_t, daddr_t);
414 static void swp_pager_meta_free_all(vm_object_t);
415 static daddr_t swp_pager_meta_ctl(vm_object_t, vm_pindex_t, int);
418 * SWP_SIZECHECK() - update swap_pager_full indication
420 * update the swap_pager_almost_full indication and warn when we are
421 * about to run out of swap space, using lowat/hiwat hysteresis.
423 * Clear swap_pager_full ( task killing ) indication when lowat is met.
425 * No restrictions on call
426 * This routine may not block.
432 if (swap_pager_avail < nswap_lowat) {
433 if (swap_pager_almost_full == 0) {
434 printf("swap_pager: out of swap space\n");
435 swap_pager_almost_full = 1;
439 if (swap_pager_avail > nswap_hiwat)
440 swap_pager_almost_full = 0;
445 * SWP_PAGER_HASH() - hash swap meta data
447 * This is an helper function which hashes the swapblk given
448 * the object and page index. It returns a pointer to a pointer
449 * to the object, or a pointer to a NULL pointer if it could not
452 static struct swblock **
453 swp_pager_hash(vm_object_t object, vm_pindex_t index)
455 struct swblock **pswap;
456 struct swblock *swap;
458 index &= ~(vm_pindex_t)SWAP_META_MASK;
459 pswap = &swhash[(index ^ (int)(intptr_t)object) & swhash_mask];
460 while ((swap = *pswap) != NULL) {
461 if (swap->swb_object == object &&
462 swap->swb_index == index
466 pswap = &swap->swb_hnext;
472 * SWAP_PAGER_INIT() - initialize the swap pager!
474 * Expected to be started from system init. NOTE: This code is run
475 * before much else so be careful what you depend on. Most of the VM
476 * system has yet to be initialized at this point.
479 swap_pager_init(void)
482 * Initialize object lists
486 for (i = 0; i < NOBJLISTS; ++i)
487 TAILQ_INIT(&swap_pager_object_list[i]);
488 mtx_init(&sw_alloc_mtx, "swap_pager list", NULL, MTX_DEF);
489 mtx_init(&sw_dev_mtx, "swapdev", NULL, MTX_DEF);
492 * Device Stripe, in PAGE_SIZE'd blocks
494 dmmax = SWB_NPAGES * 2;
498 * SWAP_PAGER_SWAP_INIT() - swap pager initialization from pageout process
500 * Expected to be started from pageout process once, prior to entering
504 swap_pager_swap_init(void)
509 * Number of in-transit swap bp operations. Don't
510 * exhaust the pbufs completely. Make sure we
511 * initialize workable values (0 will work for hysteresis
512 * but it isn't very efficient).
514 * The nsw_cluster_max is constrained by the bp->b_pages[]
515 * array (MAXPHYS/PAGE_SIZE) and our locally defined
516 * MAX_PAGEOUT_CLUSTER. Also be aware that swap ops are
517 * constrained by the swap device interleave stripe size.
519 * Currently we hardwire nsw_wcount_async to 4. This limit is
520 * designed to prevent other I/O from having high latencies due to
521 * our pageout I/O. The value 4 works well for one or two active swap
522 * devices but is probably a little low if you have more. Even so,
523 * a higher value would probably generate only a limited improvement
524 * with three or four active swap devices since the system does not
525 * typically have to pageout at extreme bandwidths. We will want
526 * at least 2 per swap devices, and 4 is a pretty good value if you
527 * have one NFS swap device due to the command/ack latency over NFS.
528 * So it all works out pretty well.
530 nsw_cluster_max = min((MAXPHYS/PAGE_SIZE), MAX_PAGEOUT_CLUSTER);
533 nsw_rcount = (nswbuf + 1) / 2;
534 nsw_wcount_sync = (nswbuf + 3) / 4;
535 nsw_wcount_async = 4;
536 nsw_wcount_async_max = nsw_wcount_async;
537 mtx_unlock(&pbuf_mtx);
540 * Initialize our zone. Right now I'm just guessing on the number
541 * we need based on the number of pages in the system. Each swblock
542 * can hold 32 pages, so this is probably overkill. This reservation
543 * is typically limited to around 32MB by default.
545 n = vm_cnt.v_page_count / 2;
546 if (maxswzone && n > maxswzone / sizeof(struct swblock))
547 n = maxswzone / sizeof(struct swblock);
549 swap_zone = uma_zcreate("SWAPMETA", sizeof(struct swblock), NULL, NULL,
550 NULL, NULL, UMA_ALIGN_PTR, UMA_ZONE_NOFREE | UMA_ZONE_VM);
551 if (swap_zone == NULL)
552 panic("failed to create swap_zone.");
554 if (uma_zone_reserve_kva(swap_zone, n))
557 * if the allocation failed, try a zone two thirds the
558 * size of the previous attempt.
563 printf("Swap zone entries reduced from %lu to %lu.\n", n2, n);
564 swap_maxpages = n * SWAP_META_PAGES;
565 swzone = n * sizeof(struct swblock);
569 * Initialize our meta-data hash table. The swapper does not need to
570 * be quite as efficient as the VM system, so we do not use an
571 * oversized hash table.
573 * n: size of hash table, must be power of 2
574 * swhash_mask: hash table index mask
576 for (n = 1; n < n2 / 8; n *= 2)
578 swhash = malloc(sizeof(struct swblock *) * n, M_VMPGDATA, M_WAITOK | M_ZERO);
580 mtx_init(&swhash_mtx, "swap_pager swhash", NULL, MTX_DEF);
584 * SWAP_PAGER_ALLOC() - allocate a new OBJT_SWAP VM object and instantiate
585 * its metadata structures.
587 * This routine is called from the mmap and fork code to create a new
588 * OBJT_SWAP object. We do this by creating an OBJT_DEFAULT object
589 * and then converting it with swp_pager_meta_build().
591 * This routine may block in vm_object_allocate() and create a named
592 * object lookup race, so we must interlock.
597 swap_pager_alloc(void *handle, vm_ooffset_t size, vm_prot_t prot,
598 vm_ooffset_t offset, struct ucred *cred)
603 pindex = OFF_TO_IDX(offset + PAGE_MASK + size);
607 * Reference existing named region or allocate new one. There
608 * should not be a race here against swp_pager_meta_build()
609 * as called from vm_page_remove() in regards to the lookup
612 sx_xlock(&sw_alloc_sx);
613 object = vm_pager_object_lookup(NOBJLIST(handle), handle);
614 if (object == NULL) {
616 if (!swap_reserve_by_cred(size, cred)) {
617 sx_xunlock(&sw_alloc_sx);
623 object = vm_object_allocate(OBJT_DEFAULT, pindex);
624 VM_OBJECT_WLOCK(object);
625 object->handle = handle;
628 object->charge = size;
630 swp_pager_meta_build(object, 0, SWAPBLK_NONE);
631 VM_OBJECT_WUNLOCK(object);
633 sx_xunlock(&sw_alloc_sx);
637 if (!swap_reserve_by_cred(size, cred))
641 object = vm_object_allocate(OBJT_DEFAULT, pindex);
642 VM_OBJECT_WLOCK(object);
645 object->charge = size;
647 swp_pager_meta_build(object, 0, SWAPBLK_NONE);
648 VM_OBJECT_WUNLOCK(object);
654 * SWAP_PAGER_DEALLOC() - remove swap metadata from object
656 * The swap backing for the object is destroyed. The code is
657 * designed such that we can reinstantiate it later, but this
658 * routine is typically called only when the entire object is
659 * about to be destroyed.
661 * The object must be locked.
664 swap_pager_dealloc(vm_object_t object)
668 * Remove from list right away so lookups will fail if we block for
669 * pageout completion.
671 if (object->handle != NULL) {
672 mtx_lock(&sw_alloc_mtx);
673 TAILQ_REMOVE(NOBJLIST(object->handle), object, pager_object_list);
674 mtx_unlock(&sw_alloc_mtx);
677 VM_OBJECT_ASSERT_WLOCKED(object);
678 vm_object_pip_wait(object, "swpdea");
681 * Free all remaining metadata. We only bother to free it from
682 * the swap meta data. We do not attempt to free swapblk's still
683 * associated with vm_page_t's for this object. We do not care
684 * if paging is still in progress on some objects.
686 swp_pager_meta_free_all(object);
687 object->handle = NULL;
688 object->type = OBJT_DEAD;
691 /************************************************************************
692 * SWAP PAGER BITMAP ROUTINES *
693 ************************************************************************/
696 * SWP_PAGER_GETSWAPSPACE() - allocate raw swap space
698 * Allocate swap for the requested number of pages. The starting
699 * swap block number (a page index) is returned or SWAPBLK_NONE
700 * if the allocation failed.
702 * Also has the side effect of advising that somebody made a mistake
703 * when they configured swap and didn't configure enough.
705 * This routine may not sleep.
707 * We allocate in round-robin fashion from the configured devices.
710 swp_pager_getswapspace(int npages)
717 mtx_lock(&sw_dev_mtx);
719 for (i = 0; i < nswapdev; i++) {
721 sp = TAILQ_FIRST(&swtailq);
722 if (!(sp->sw_flags & SW_CLOSING)) {
723 blk = blist_alloc(sp->sw_blist, npages);
724 if (blk != SWAPBLK_NONE) {
726 sp->sw_used += npages;
727 swap_pager_avail -= npages;
729 swdevhd = TAILQ_NEXT(sp, sw_list);
733 sp = TAILQ_NEXT(sp, sw_list);
735 if (swap_pager_full != 2) {
736 printf("swap_pager_getswapspace(%d): failed\n", npages);
738 swap_pager_almost_full = 1;
742 mtx_unlock(&sw_dev_mtx);
747 swp_pager_isondev(daddr_t blk, struct swdevt *sp)
750 return (blk >= sp->sw_first && blk < sp->sw_end);
754 swp_pager_strategy(struct buf *bp)
758 mtx_lock(&sw_dev_mtx);
759 TAILQ_FOREACH(sp, &swtailq, sw_list) {
760 if (bp->b_blkno >= sp->sw_first && bp->b_blkno < sp->sw_end) {
761 mtx_unlock(&sw_dev_mtx);
762 if ((sp->sw_flags & SW_UNMAPPED) != 0 &&
763 unmapped_buf_allowed) {
764 bp->b_data = unmapped_buf;
767 pmap_qenter((vm_offset_t)bp->b_data,
768 &bp->b_pages[0], bp->b_bcount / PAGE_SIZE);
770 sp->sw_strategy(bp, sp);
774 panic("Swapdev not found");
779 * SWP_PAGER_FREESWAPSPACE() - free raw swap space
781 * This routine returns the specified swap blocks back to the bitmap.
783 * This routine may not sleep.
786 swp_pager_freeswapspace(daddr_t blk, int npages)
790 mtx_lock(&sw_dev_mtx);
791 TAILQ_FOREACH(sp, &swtailq, sw_list) {
792 if (blk >= sp->sw_first && blk < sp->sw_end) {
793 sp->sw_used -= npages;
795 * If we are attempting to stop swapping on
796 * this device, we don't want to mark any
797 * blocks free lest they be reused.
799 if ((sp->sw_flags & SW_CLOSING) == 0) {
800 blist_free(sp->sw_blist, blk - sp->sw_first,
802 swap_pager_avail += npages;
805 mtx_unlock(&sw_dev_mtx);
809 panic("Swapdev not found");
813 * SWAP_PAGER_FREESPACE() - frees swap blocks associated with a page
814 * range within an object.
816 * This is a globally accessible routine.
818 * This routine removes swapblk assignments from swap metadata.
820 * The external callers of this routine typically have already destroyed
821 * or renamed vm_page_t's associated with this range in the object so
824 * The object must be locked.
827 swap_pager_freespace(vm_object_t object, vm_pindex_t start, vm_size_t size)
830 swp_pager_meta_free(object, start, size);
834 * SWAP_PAGER_RESERVE() - reserve swap blocks in object
836 * Assigns swap blocks to the specified range within the object. The
837 * swap blocks are not zeroed. Any previous swap assignment is destroyed.
839 * Returns 0 on success, -1 on failure.
842 swap_pager_reserve(vm_object_t object, vm_pindex_t start, vm_size_t size)
845 daddr_t blk = SWAPBLK_NONE;
846 vm_pindex_t beg = start; /* save start index */
848 VM_OBJECT_WLOCK(object);
852 while ((blk = swp_pager_getswapspace(n)) == SWAPBLK_NONE) {
855 swp_pager_meta_free(object, beg, start - beg);
856 VM_OBJECT_WUNLOCK(object);
861 swp_pager_meta_build(object, start, blk);
867 swp_pager_meta_free(object, start, n);
868 VM_OBJECT_WUNLOCK(object);
873 * SWAP_PAGER_COPY() - copy blocks from source pager to destination pager
874 * and destroy the source.
876 * Copy any valid swapblks from the source to the destination. In
877 * cases where both the source and destination have a valid swapblk,
878 * we keep the destination's.
880 * This routine is allowed to sleep. It may sleep allocating metadata
881 * indirectly through swp_pager_meta_build() or if paging is still in
882 * progress on the source.
884 * The source object contains no vm_page_t's (which is just as well)
886 * The source object is of type OBJT_SWAP.
888 * The source and destination objects must be locked.
889 * Both object locks may temporarily be released.
892 swap_pager_copy(vm_object_t srcobject, vm_object_t dstobject,
893 vm_pindex_t offset, int destroysource)
897 VM_OBJECT_ASSERT_WLOCKED(srcobject);
898 VM_OBJECT_ASSERT_WLOCKED(dstobject);
901 * If destroysource is set, we remove the source object from the
902 * swap_pager internal queue now.
905 if (srcobject->handle != NULL) {
906 mtx_lock(&sw_alloc_mtx);
908 NOBJLIST(srcobject->handle),
912 mtx_unlock(&sw_alloc_mtx);
917 * transfer source to destination.
919 for (i = 0; i < dstobject->size; ++i) {
923 * Locate (without changing) the swapblk on the destination,
924 * unless it is invalid in which case free it silently, or
925 * if the destination is a resident page, in which case the
926 * source is thrown away.
928 dstaddr = swp_pager_meta_ctl(dstobject, i, 0);
930 if (dstaddr == SWAPBLK_NONE) {
932 * Destination has no swapblk and is not resident,
937 srcaddr = swp_pager_meta_ctl(
943 if (srcaddr != SWAPBLK_NONE) {
945 * swp_pager_meta_build() can sleep.
947 vm_object_pip_add(srcobject, 1);
948 VM_OBJECT_WUNLOCK(srcobject);
949 vm_object_pip_add(dstobject, 1);
950 swp_pager_meta_build(dstobject, i, srcaddr);
951 vm_object_pip_wakeup(dstobject);
952 VM_OBJECT_WLOCK(srcobject);
953 vm_object_pip_wakeup(srcobject);
957 * Destination has valid swapblk or it is represented
958 * by a resident page. We destroy the sourceblock.
961 swp_pager_meta_ctl(srcobject, i + offset, SWM_FREE);
966 * Free left over swap blocks in source.
968 * We have to revert the type to OBJT_DEFAULT so we do not accidently
969 * double-remove the object from the swap queues.
972 swp_pager_meta_free_all(srcobject);
974 * Reverting the type is not necessary, the caller is going
975 * to destroy srcobject directly, but I'm doing it here
976 * for consistency since we've removed the object from its
979 srcobject->type = OBJT_DEFAULT;
984 * SWAP_PAGER_HASPAGE() - determine if we have good backing store for
985 * the requested page.
987 * We determine whether good backing store exists for the requested
988 * page and return TRUE if it does, FALSE if it doesn't.
990 * If TRUE, we also try to determine how much valid, contiguous backing
991 * store exists before and after the requested page within a reasonable
992 * distance. We do not try to restrict it to the swap device stripe
993 * (that is handled in getpages/putpages). It probably isn't worth
997 swap_pager_haspage(vm_object_t object, vm_pindex_t pindex, int *before, int *after)
1001 VM_OBJECT_ASSERT_LOCKED(object);
1003 * do we have good backing store at the requested index ?
1005 blk0 = swp_pager_meta_ctl(object, pindex, 0);
1007 if (blk0 == SWAPBLK_NONE) {
1016 * find backwards-looking contiguous good backing store
1018 if (before != NULL) {
1021 for (i = 1; i < (SWB_NPAGES/2); ++i) {
1026 blk = swp_pager_meta_ctl(object, pindex - i, 0);
1027 if (blk != blk0 - i)
1034 * find forward-looking contiguous good backing store
1036 if (after != NULL) {
1039 for (i = 1; i < (SWB_NPAGES/2); ++i) {
1042 blk = swp_pager_meta_ctl(object, pindex + i, 0);
1043 if (blk != blk0 + i)
1052 * SWAP_PAGER_PAGE_UNSWAPPED() - remove swap backing store related to page
1054 * This removes any associated swap backing store, whether valid or
1055 * not, from the page.
1057 * This routine is typically called when a page is made dirty, at
1058 * which point any associated swap can be freed. MADV_FREE also
1059 * calls us in a special-case situation
1061 * NOTE!!! If the page is clean and the swap was valid, the caller
1062 * should make the page dirty before calling this routine. This routine
1063 * does NOT change the m->dirty status of the page. Also: MADV_FREE
1066 * This routine may not sleep.
1068 * The object containing the page must be locked.
1071 swap_pager_unswapped(vm_page_t m)
1074 swp_pager_meta_ctl(m->object, m->pindex, SWM_FREE);
1078 * SWAP_PAGER_GETPAGES() - bring pages in from swap
1080 * Attempt to retrieve (m, count) pages from backing store, but make
1081 * sure we retrieve at least m[reqpage]. We try to load in as large
1082 * a chunk surrounding m[reqpage] as is contiguous in swap and which
1083 * belongs to the same object.
1085 * The code is designed for asynchronous operation and
1086 * immediate-notification of 'reqpage' but tends not to be
1087 * used that way. Please do not optimize-out this algorithmic
1088 * feature, I intend to improve on it in the future.
1090 * The parent has a single vm_object_pip_add() reference prior to
1091 * calling us and we should return with the same.
1093 * The parent has BUSY'd the pages. We should return with 'm'
1094 * left busy, but the others adjusted.
1097 swap_pager_getpages(vm_object_t object, vm_page_t *m, int count, int *rbehind,
1104 * Calculate range to retrieve. The pages have already been assigned
1105 * their swapblks. We require a *contiguous* range but we know it to
1106 * not span devices. If we do not supply it, bad things
1107 * happen. Note that blk, iblk & jblk can be SWAPBLK_NONE, but the
1108 * loops are set up such that the case(s) are handled implicitly.
1110 * The swp_*() calls must be made with the object locked.
1112 blk = swp_pager_meta_ctl(m[0]->object, m[0]->pindex, 0);
1114 if (blk == SWAPBLK_NONE)
1115 return (VM_PAGER_FAIL);
1118 for (int i = 0; i < count; i++)
1120 swp_pager_meta_ctl(m[i]->object, m[i]->pindex, 0),
1121 ("%s: range is not contiguous", __func__));
1125 * Getpbuf() can sleep.
1127 VM_OBJECT_WUNLOCK(object);
1129 * Get a swap buffer header to perform the IO
1131 bp = getpbuf(&nsw_rcount);
1132 bp->b_flags |= B_PAGING;
1134 bp->b_iocmd = BIO_READ;
1135 bp->b_iodone = swp_pager_async_iodone;
1136 bp->b_rcred = crhold(thread0.td_ucred);
1137 bp->b_wcred = crhold(thread0.td_ucred);
1139 bp->b_bcount = PAGE_SIZE * count;
1140 bp->b_bufsize = PAGE_SIZE * count;
1141 bp->b_npages = count;
1143 VM_OBJECT_WLOCK(object);
1144 for (int i = 0; i < count; i++) {
1145 bp->b_pages[i] = m[i];
1146 m[i]->oflags |= VPO_SWAPINPROG;
1149 PCPU_INC(cnt.v_swapin);
1150 PCPU_ADD(cnt.v_swappgsin, bp->b_npages);
1153 * We still hold the lock on mreq, and our automatic completion routine
1154 * does not remove it.
1156 vm_object_pip_add(object, bp->b_npages);
1157 VM_OBJECT_WUNLOCK(object);
1160 * perform the I/O. NOTE!!! bp cannot be considered valid after
1161 * this point because we automatically release it on completion.
1162 * Instead, we look at the one page we are interested in which we
1163 * still hold a lock on even through the I/O completion.
1165 * The other pages in our m[] array are also released on completion,
1166 * so we cannot assume they are valid anymore either.
1168 * NOTE: b_blkno is destroyed by the call to swapdev_strategy
1171 swp_pager_strategy(bp);
1174 * wait for the page we want to complete. VPO_SWAPINPROG is always
1175 * cleared on completion. If an I/O error occurs, SWAPBLK_NONE
1176 * is set in the meta-data.
1178 VM_OBJECT_WLOCK(object);
1179 while ((m[0]->oflags & VPO_SWAPINPROG) != 0) {
1180 m[0]->oflags |= VPO_SWAPSLEEP;
1181 PCPU_INC(cnt.v_intrans);
1182 if (VM_OBJECT_SLEEP(object, &object->paging_in_progress, PSWP,
1183 "swread", hz * 20)) {
1185 "swap_pager: indefinite wait buffer: bufobj: %p, blkno: %jd, size: %ld\n",
1186 bp->b_bufobj, (intmax_t)bp->b_blkno, bp->b_bcount);
1191 * If we had an unrecoverable read error pages will not be valid.
1193 for (int i = 0; i < count; i++)
1194 if (m[i]->valid != VM_PAGE_BITS_ALL)
1195 return (VM_PAGER_ERROR);
1202 return (VM_PAGER_OK);
1205 * A final note: in a low swap situation, we cannot deallocate swap
1206 * and mark a page dirty here because the caller is likely to mark
1207 * the page clean when we return, causing the page to possibly revert
1208 * to all-zero's later.
1213 * swap_pager_getpages_async():
1215 * Right now this is emulation of asynchronous operation on top of
1216 * swap_pager_getpages().
1219 swap_pager_getpages_async(vm_object_t object, vm_page_t *m, int count,
1220 int *rbehind, int *rahead, pgo_getpages_iodone_t iodone, void *arg)
1224 r = swap_pager_getpages(object, m, count, rbehind, rahead);
1225 VM_OBJECT_WUNLOCK(object);
1230 case VM_PAGER_ERROR:
1237 panic("unhandled swap_pager_getpages() error %d", r);
1239 (iodone)(arg, m, count, error);
1240 VM_OBJECT_WLOCK(object);
1246 * swap_pager_putpages:
1248 * Assign swap (if necessary) and initiate I/O on the specified pages.
1250 * We support both OBJT_DEFAULT and OBJT_SWAP objects. DEFAULT objects
1251 * are automatically converted to SWAP objects.
1253 * In a low memory situation we may block in VOP_STRATEGY(), but the new
1254 * vm_page reservation system coupled with properly written VFS devices
1255 * should ensure that no low-memory deadlock occurs. This is an area
1258 * The parent has N vm_object_pip_add() references prior to
1259 * calling us and will remove references for rtvals[] that are
1260 * not set to VM_PAGER_PEND. We need to remove the rest on I/O
1263 * The parent has soft-busy'd the pages it passes us and will unbusy
1264 * those whos rtvals[] entry is not set to VM_PAGER_PEND on return.
1265 * We need to unbusy the rest on I/O completion.
1268 swap_pager_putpages(vm_object_t object, vm_page_t *m, int count,
1269 int flags, int *rtvals)
1274 if (count && m[0]->object != object) {
1275 panic("swap_pager_putpages: object mismatch %p/%p",
1284 * Turn object into OBJT_SWAP
1285 * check for bogus sysops
1286 * force sync if not pageout process
1288 if (object->type != OBJT_SWAP)
1289 swp_pager_meta_build(object, 0, SWAPBLK_NONE);
1290 VM_OBJECT_WUNLOCK(object);
1293 if (curproc != pageproc)
1296 sync = (flags & VM_PAGER_PUT_SYNC) != 0;
1301 * Assign swap blocks and issue I/O. We reallocate swap on the fly.
1302 * The page is left dirty until the pageout operation completes
1305 for (i = 0; i < count; i += n) {
1311 * Maximum I/O size is limited by a number of factors.
1313 n = min(BLIST_MAX_ALLOC, count - i);
1314 n = min(n, nsw_cluster_max);
1317 * Get biggest block of swap we can. If we fail, fall
1318 * back and try to allocate a smaller block. Don't go
1319 * overboard trying to allocate space if it would overly
1323 (blk = swp_pager_getswapspace(n)) == SWAPBLK_NONE &&
1328 if (blk == SWAPBLK_NONE) {
1329 for (j = 0; j < n; ++j)
1330 rtvals[i+j] = VM_PAGER_FAIL;
1335 * All I/O parameters have been satisfied, build the I/O
1336 * request and assign the swap space.
1339 bp = getpbuf(&nsw_wcount_sync);
1341 bp = getpbuf(&nsw_wcount_async);
1342 bp->b_flags = B_ASYNC;
1344 bp->b_flags |= B_PAGING;
1345 bp->b_iocmd = BIO_WRITE;
1347 bp->b_rcred = crhold(thread0.td_ucred);
1348 bp->b_wcred = crhold(thread0.td_ucred);
1349 bp->b_bcount = PAGE_SIZE * n;
1350 bp->b_bufsize = PAGE_SIZE * n;
1353 VM_OBJECT_WLOCK(object);
1354 for (j = 0; j < n; ++j) {
1355 vm_page_t mreq = m[i+j];
1357 swp_pager_meta_build(
1362 vm_page_dirty(mreq);
1363 mreq->oflags |= VPO_SWAPINPROG;
1364 bp->b_pages[j] = mreq;
1366 VM_OBJECT_WUNLOCK(object);
1369 * Must set dirty range for NFS to work.
1372 bp->b_dirtyend = bp->b_bcount;
1374 PCPU_INC(cnt.v_swapout);
1375 PCPU_ADD(cnt.v_swappgsout, bp->b_npages);
1378 * We unconditionally set rtvals[] to VM_PAGER_PEND so that we
1379 * can call the async completion routine at the end of a
1380 * synchronous I/O operation. Otherwise, our caller would
1381 * perform duplicate unbusy and wakeup operations on the page
1382 * and object, respectively.
1384 for (j = 0; j < n; j++)
1385 rtvals[i + j] = VM_PAGER_PEND;
1390 * NOTE: b_blkno is destroyed by the call to swapdev_strategy
1392 if (sync == FALSE) {
1393 bp->b_iodone = swp_pager_async_iodone;
1395 swp_pager_strategy(bp);
1402 * NOTE: b_blkno is destroyed by the call to swapdev_strategy
1404 bp->b_iodone = bdone;
1405 swp_pager_strategy(bp);
1408 * Wait for the sync I/O to complete.
1410 bwait(bp, PVM, "swwrt");
1413 * Now that we are through with the bp, we can call the
1414 * normal async completion, which frees everything up.
1416 swp_pager_async_iodone(bp);
1418 VM_OBJECT_WLOCK(object);
1422 * swp_pager_async_iodone:
1424 * Completion routine for asynchronous reads and writes from/to swap.
1425 * Also called manually by synchronous code to finish up a bp.
1427 * This routine may not sleep.
1430 swp_pager_async_iodone(struct buf *bp)
1433 vm_object_t object = NULL;
1438 if (bp->b_ioflags & BIO_ERROR) {
1440 "swap_pager: I/O error - %s failed; blkno %ld,"
1441 "size %ld, error %d\n",
1442 ((bp->b_iocmd == BIO_READ) ? "pagein" : "pageout"),
1450 * remove the mapping for kernel virtual
1453 pmap_qremove((vm_offset_t)bp->b_data, bp->b_npages);
1455 bp->b_data = bp->b_kvabase;
1458 object = bp->b_pages[0]->object;
1459 VM_OBJECT_WLOCK(object);
1463 * cleanup pages. If an error occurs writing to swap, we are in
1464 * very serious trouble. If it happens to be a disk error, though,
1465 * we may be able to recover by reassigning the swap later on. So
1466 * in this case we remove the m->swapblk assignment for the page
1467 * but do not free it in the rlist. The errornous block(s) are thus
1468 * never reallocated as swap. Redirty the page and continue.
1470 for (i = 0; i < bp->b_npages; ++i) {
1471 vm_page_t m = bp->b_pages[i];
1473 m->oflags &= ~VPO_SWAPINPROG;
1474 if (m->oflags & VPO_SWAPSLEEP) {
1475 m->oflags &= ~VPO_SWAPSLEEP;
1476 wakeup(&object->paging_in_progress);
1479 if (bp->b_ioflags & BIO_ERROR) {
1481 * If an error occurs I'd love to throw the swapblk
1482 * away without freeing it back to swapspace, so it
1483 * can never be used again. But I can't from an
1486 if (bp->b_iocmd == BIO_READ) {
1488 * NOTE: for reads, m->dirty will probably
1489 * be overridden by the original caller of
1490 * getpages so don't play cute tricks here.
1495 * If a write error occurs, reactivate page
1496 * so it doesn't clog the inactive list,
1497 * then finish the I/O.
1501 vm_page_activate(m);
1505 } else if (bp->b_iocmd == BIO_READ) {
1507 * NOTE: for reads, m->dirty will probably be
1508 * overridden by the original caller of getpages so
1509 * we cannot set them in order to free the underlying
1510 * swap in a low-swap situation. I don't think we'd
1511 * want to do that anyway, but it was an optimization
1512 * that existed in the old swapper for a time before
1513 * it got ripped out due to precisely this problem.
1515 KASSERT(!pmap_page_is_mapped(m),
1516 ("swp_pager_async_iodone: page %p is mapped", m));
1517 KASSERT(m->dirty == 0,
1518 ("swp_pager_async_iodone: page %p is dirty", m));
1519 m->valid = VM_PAGE_BITS_ALL;
1522 * For write success, clear the dirty
1523 * status, then finish the I/O ( which decrements the
1524 * busy count and possibly wakes waiter's up ).
1526 KASSERT(!pmap_page_is_write_mapped(m),
1527 ("swp_pager_async_iodone: page %p is not write"
1531 if (vm_page_count_severe()) {
1533 vm_page_try_to_cache(m);
1540 * adjust pip. NOTE: the original parent may still have its own
1541 * pip refs on the object.
1543 if (object != NULL) {
1544 vm_object_pip_wakeupn(object, bp->b_npages);
1545 VM_OBJECT_WUNLOCK(object);
1549 * swapdev_strategy() manually sets b_vp and b_bufobj before calling
1550 * bstrategy(). Set them back to NULL now we're done with it, or we'll
1551 * trigger a KASSERT in relpbuf().
1555 bp->b_bufobj = NULL;
1558 * release the physical I/O buffer
1562 ((bp->b_iocmd == BIO_READ) ? &nsw_rcount :
1563 ((bp->b_flags & B_ASYNC) ?
1572 * swap_pager_isswapped:
1574 * Return 1 if at least one page in the given object is paged
1575 * out to the given swap device.
1577 * This routine may not sleep.
1580 swap_pager_isswapped(vm_object_t object, struct swdevt *sp)
1586 VM_OBJECT_ASSERT_WLOCKED(object);
1587 if (object->type != OBJT_SWAP)
1590 mtx_lock(&swhash_mtx);
1591 for (bcount = 0; bcount < object->un_pager.swp.swp_bcount; bcount++) {
1592 struct swblock *swap;
1594 if ((swap = *swp_pager_hash(object, index)) != NULL) {
1595 for (i = 0; i < SWAP_META_PAGES; ++i) {
1596 if (swp_pager_isondev(swap->swb_pages[i], sp)) {
1597 mtx_unlock(&swhash_mtx);
1602 index += SWAP_META_PAGES;
1604 mtx_unlock(&swhash_mtx);
1609 * SWP_PAGER_FORCE_PAGEIN() - force a swap block to be paged in
1611 * This routine dissociates the page at the given index within a
1612 * swap block from its backing store, paging it in if necessary.
1613 * If the page is paged in, it is placed in the inactive queue,
1614 * since it had its backing store ripped out from under it.
1615 * We also attempt to swap in all other pages in the swap block,
1616 * we only guarantee that the one at the specified index is
1619 * XXX - The code to page the whole block in doesn't work, so we
1620 * revert to the one-by-one behavior for now. Sigh.
1623 swp_pager_force_pagein(vm_object_t object, vm_pindex_t pindex)
1627 vm_object_pip_add(object, 1);
1628 m = vm_page_grab(object, pindex, VM_ALLOC_NORMAL);
1629 if (m->valid == VM_PAGE_BITS_ALL) {
1630 vm_object_pip_wakeup(object);
1633 vm_page_activate(m);
1636 vm_pager_page_unswapped(m);
1640 if (swap_pager_getpages(object, &m, 1, NULL, NULL) != VM_PAGER_OK)
1641 panic("swap_pager_force_pagein: read from swap failed");/*XXX*/
1642 vm_object_pip_wakeup(object);
1645 vm_page_deactivate(m);
1648 vm_pager_page_unswapped(m);
1652 * swap_pager_swapoff:
1654 * Page in all of the pages that have been paged out to the
1655 * given device. The corresponding blocks in the bitmap must be
1656 * marked as allocated and the device must be flagged SW_CLOSING.
1657 * There may be no processes swapped out to the device.
1659 * This routine may block.
1662 swap_pager_swapoff(struct swdevt *sp)
1664 struct swblock *swap;
1671 mtx_lock(&swhash_mtx);
1672 for (i = 0; i <= swhash_mask; i++) { /* '<=' is correct here */
1674 for (swap = swhash[i]; swap != NULL; swap = swap->swb_hnext) {
1675 vm_object_t object = swap->swb_object;
1676 vm_pindex_t pindex = swap->swb_index;
1677 for (j = 0; j < SWAP_META_PAGES; ++j) {
1678 if (swp_pager_isondev(swap->swb_pages[j], sp)) {
1679 /* avoid deadlock */
1680 if (!VM_OBJECT_TRYWLOCK(object)) {
1683 mtx_unlock(&swhash_mtx);
1684 swp_pager_force_pagein(object,
1686 VM_OBJECT_WUNLOCK(object);
1687 mtx_lock(&swhash_mtx);
1694 mtx_unlock(&swhash_mtx);
1697 * Objects may be locked or paging to the device being
1698 * removed, so we will miss their pages and need to
1699 * make another pass. We have marked this device as
1700 * SW_CLOSING, so the activity should finish soon.
1703 if (retries > 100) {
1704 panic("swapoff: failed to locate %d swap blocks",
1707 pause("swpoff", hz / 20);
1712 /************************************************************************
1714 ************************************************************************
1716 * These routines manipulate the swap metadata stored in the
1719 * Swap metadata is implemented with a global hash and not directly
1720 * linked into the object. Instead the object simply contains
1721 * appropriate tracking counters.
1725 * SWP_PAGER_META_BUILD() - add swap block to swap meta data for object
1727 * We first convert the object to a swap object if it is a default
1730 * The specified swapblk is added to the object's swap metadata. If
1731 * the swapblk is not valid, it is freed instead. Any previously
1732 * assigned swapblk is freed.
1735 swp_pager_meta_build(vm_object_t object, vm_pindex_t pindex, daddr_t swapblk)
1737 static volatile int exhausted;
1738 struct swblock *swap;
1739 struct swblock **pswap;
1742 VM_OBJECT_ASSERT_WLOCKED(object);
1744 * Convert default object to swap object if necessary
1746 if (object->type != OBJT_SWAP) {
1747 object->type = OBJT_SWAP;
1748 object->un_pager.swp.swp_bcount = 0;
1750 if (object->handle != NULL) {
1751 mtx_lock(&sw_alloc_mtx);
1753 NOBJLIST(object->handle),
1757 mtx_unlock(&sw_alloc_mtx);
1762 * Locate hash entry. If not found create, but if we aren't adding
1763 * anything just return. If we run out of space in the map we wait
1764 * and, since the hash table may have changed, retry.
1767 mtx_lock(&swhash_mtx);
1768 pswap = swp_pager_hash(object, pindex);
1770 if ((swap = *pswap) == NULL) {
1773 if (swapblk == SWAPBLK_NONE)
1776 swap = *pswap = uma_zalloc(swap_zone, M_NOWAIT |
1777 (curproc == pageproc ? M_USE_RESERVE : 0));
1779 mtx_unlock(&swhash_mtx);
1780 VM_OBJECT_WUNLOCK(object);
1781 if (uma_zone_exhausted(swap_zone)) {
1782 if (atomic_cmpset_int(&exhausted, 0, 1))
1783 printf("swap zone exhausted, "
1784 "increase kern.maxswzone\n");
1785 vm_pageout_oom(VM_OOM_SWAPZ);
1786 pause("swzonex", 10);
1789 VM_OBJECT_WLOCK(object);
1793 if (atomic_cmpset_int(&exhausted, 1, 0))
1794 printf("swap zone ok\n");
1796 swap->swb_hnext = NULL;
1797 swap->swb_object = object;
1798 swap->swb_index = pindex & ~(vm_pindex_t)SWAP_META_MASK;
1799 swap->swb_count = 0;
1801 ++object->un_pager.swp.swp_bcount;
1803 for (i = 0; i < SWAP_META_PAGES; ++i)
1804 swap->swb_pages[i] = SWAPBLK_NONE;
1808 * Delete prior contents of metadata
1810 idx = pindex & SWAP_META_MASK;
1812 if (swap->swb_pages[idx] != SWAPBLK_NONE) {
1813 swp_pager_freeswapspace(swap->swb_pages[idx], 1);
1818 * Enter block into metadata
1820 swap->swb_pages[idx] = swapblk;
1821 if (swapblk != SWAPBLK_NONE)
1824 mtx_unlock(&swhash_mtx);
1828 * SWP_PAGER_META_FREE() - free a range of blocks in the object's swap metadata
1830 * The requested range of blocks is freed, with any associated swap
1831 * returned to the swap bitmap.
1833 * This routine will free swap metadata structures as they are cleaned
1834 * out. This routine does *NOT* operate on swap metadata associated
1835 * with resident pages.
1838 swp_pager_meta_free(vm_object_t object, vm_pindex_t index, daddr_t count)
1841 VM_OBJECT_ASSERT_LOCKED(object);
1842 if (object->type != OBJT_SWAP)
1846 struct swblock **pswap;
1847 struct swblock *swap;
1849 mtx_lock(&swhash_mtx);
1850 pswap = swp_pager_hash(object, index);
1852 if ((swap = *pswap) != NULL) {
1853 daddr_t v = swap->swb_pages[index & SWAP_META_MASK];
1855 if (v != SWAPBLK_NONE) {
1856 swp_pager_freeswapspace(v, 1);
1857 swap->swb_pages[index & SWAP_META_MASK] =
1859 if (--swap->swb_count == 0) {
1860 *pswap = swap->swb_hnext;
1861 uma_zfree(swap_zone, swap);
1862 --object->un_pager.swp.swp_bcount;
1868 int n = SWAP_META_PAGES - (index & SWAP_META_MASK);
1872 mtx_unlock(&swhash_mtx);
1877 * SWP_PAGER_META_FREE_ALL() - destroy all swap metadata associated with object
1879 * This routine locates and destroys all swap metadata associated with
1883 swp_pager_meta_free_all(vm_object_t object)
1887 VM_OBJECT_ASSERT_WLOCKED(object);
1888 if (object->type != OBJT_SWAP)
1891 while (object->un_pager.swp.swp_bcount) {
1892 struct swblock **pswap;
1893 struct swblock *swap;
1895 mtx_lock(&swhash_mtx);
1896 pswap = swp_pager_hash(object, index);
1897 if ((swap = *pswap) != NULL) {
1900 for (i = 0; i < SWAP_META_PAGES; ++i) {
1901 daddr_t v = swap->swb_pages[i];
1902 if (v != SWAPBLK_NONE) {
1904 swp_pager_freeswapspace(v, 1);
1907 if (swap->swb_count != 0)
1908 panic("swap_pager_meta_free_all: swb_count != 0");
1909 *pswap = swap->swb_hnext;
1910 uma_zfree(swap_zone, swap);
1911 --object->un_pager.swp.swp_bcount;
1913 mtx_unlock(&swhash_mtx);
1914 index += SWAP_META_PAGES;
1919 * SWP_PAGER_METACTL() - misc control of swap and vm_page_t meta data.
1921 * This routine is capable of looking up, popping, or freeing
1922 * swapblk assignments in the swap meta data or in the vm_page_t.
1923 * The routine typically returns the swapblk being looked-up, or popped,
1924 * or SWAPBLK_NONE if the block was freed, or SWAPBLK_NONE if the block
1925 * was invalid. This routine will automatically free any invalid
1926 * meta-data swapblks.
1928 * It is not possible to store invalid swapblks in the swap meta data
1929 * (other then a literal 'SWAPBLK_NONE'), so we don't bother checking.
1931 * When acting on a busy resident page and paging is in progress, we
1932 * have to wait until paging is complete but otherwise can act on the
1935 * SWM_FREE remove and free swap block from metadata
1936 * SWM_POP remove from meta data but do not free.. pop it out
1939 swp_pager_meta_ctl(vm_object_t object, vm_pindex_t pindex, int flags)
1941 struct swblock **pswap;
1942 struct swblock *swap;
1946 VM_OBJECT_ASSERT_LOCKED(object);
1948 * The meta data only exists of the object is OBJT_SWAP
1949 * and even then might not be allocated yet.
1951 if (object->type != OBJT_SWAP)
1952 return (SWAPBLK_NONE);
1955 mtx_lock(&swhash_mtx);
1956 pswap = swp_pager_hash(object, pindex);
1958 if ((swap = *pswap) != NULL) {
1959 idx = pindex & SWAP_META_MASK;
1960 r1 = swap->swb_pages[idx];
1962 if (r1 != SWAPBLK_NONE) {
1963 if (flags & SWM_FREE) {
1964 swp_pager_freeswapspace(r1, 1);
1967 if (flags & (SWM_FREE|SWM_POP)) {
1968 swap->swb_pages[idx] = SWAPBLK_NONE;
1969 if (--swap->swb_count == 0) {
1970 *pswap = swap->swb_hnext;
1971 uma_zfree(swap_zone, swap);
1972 --object->un_pager.swp.swp_bcount;
1977 mtx_unlock(&swhash_mtx);
1982 * System call swapon(name) enables swapping on device name,
1983 * which must be in the swdevsw. Return EBUSY
1984 * if already swapping on this device.
1986 #ifndef _SYS_SYSPROTO_H_
1987 struct swapon_args {
1997 sys_swapon(struct thread *td, struct swapon_args *uap)
2001 struct nameidata nd;
2004 error = priv_check(td, PRIV_SWAPON);
2009 while (swdev_syscall_active)
2010 tsleep(&swdev_syscall_active, PUSER - 1, "swpon", 0);
2011 swdev_syscall_active = 1;
2014 * Swap metadata may not fit in the KVM if we have physical
2017 if (swap_zone == NULL) {
2022 NDINIT(&nd, LOOKUP, ISOPEN | FOLLOW | AUDITVNODE1, UIO_USERSPACE,
2028 NDFREE(&nd, NDF_ONLY_PNBUF);
2031 if (vn_isdisk(vp, &error)) {
2032 error = swapongeom(td, vp);
2033 } else if (vp->v_type == VREG &&
2034 (vp->v_mount->mnt_vfc->vfc_flags & VFCF_NETWORK) != 0 &&
2035 (error = VOP_GETATTR(vp, &attr, td->td_ucred)) == 0) {
2037 * Allow direct swapping to NFS regular files in the same
2038 * way that nfs_mountroot() sets up diskless swapping.
2040 error = swaponvp(td, vp, attr.va_size / DEV_BSIZE);
2046 swdev_syscall_active = 0;
2047 wakeup_one(&swdev_syscall_active);
2053 * Check that the total amount of swap currently configured does not
2054 * exceed half the theoretical maximum. If it does, print a warning
2055 * message and return -1; otherwise, return 0.
2058 swapon_check_swzone(unsigned long npages)
2060 unsigned long maxpages;
2062 /* absolute maximum we can handle assuming 100% efficiency */
2063 maxpages = uma_zone_get_max(swap_zone) * SWAP_META_PAGES;
2065 /* recommend using no more than half that amount */
2066 if (npages > maxpages / 2) {
2067 printf("warning: total configured swap (%lu pages) "
2068 "exceeds maximum recommended amount (%lu pages).\n",
2069 npages, maxpages / 2);
2070 printf("warning: increase kern.maxswzone "
2071 "or reduce amount of swap.\n");
2078 swaponsomething(struct vnode *vp, void *id, u_long nblks,
2079 sw_strategy_t *strategy, sw_close_t *close, dev_t dev, int flags)
2081 struct swdevt *sp, *tsp;
2086 * nblks is in DEV_BSIZE'd chunks, convert to PAGE_SIZE'd chunks.
2087 * First chop nblks off to page-align it, then convert.
2089 * sw->sw_nblks is in page-sized chunks now too.
2091 nblks &= ~(ctodb(1) - 1);
2092 nblks = dbtoc(nblks);
2095 * If we go beyond this, we get overflows in the radix
2098 mblocks = 0x40000000 / BLIST_META_RADIX;
2099 if (nblks > mblocks) {
2101 "WARNING: reducing swap size to maximum of %luMB per unit\n",
2102 mblocks / 1024 / 1024 * PAGE_SIZE);
2106 sp = malloc(sizeof *sp, M_VMPGDATA, M_WAITOK | M_ZERO);
2111 sp->sw_nblks = nblks;
2113 sp->sw_strategy = strategy;
2114 sp->sw_close = close;
2115 sp->sw_flags = flags;
2117 sp->sw_blist = blist_create(nblks, M_WAITOK);
2119 * Do not free the first two block in order to avoid overwriting
2120 * any bsd label at the front of the partition
2122 blist_free(sp->sw_blist, 2, nblks - 2);
2125 mtx_lock(&sw_dev_mtx);
2126 TAILQ_FOREACH(tsp, &swtailq, sw_list) {
2127 if (tsp->sw_end >= dvbase) {
2129 * We put one uncovered page between the devices
2130 * in order to definitively prevent any cross-device
2133 dvbase = tsp->sw_end + 1;
2136 sp->sw_first = dvbase;
2137 sp->sw_end = dvbase + nblks;
2138 TAILQ_INSERT_TAIL(&swtailq, sp, sw_list);
2140 swap_pager_avail += nblks;
2141 swap_total += (vm_ooffset_t)nblks * PAGE_SIZE;
2142 swapon_check_swzone(swap_total / PAGE_SIZE);
2144 mtx_unlock(&sw_dev_mtx);
2148 * SYSCALL: swapoff(devname)
2150 * Disable swapping on the given device.
2152 * XXX: Badly designed system call: it should use a device index
2153 * rather than filename as specification. We keep sw_vp around
2154 * only to make this work.
2156 #ifndef _SYS_SYSPROTO_H_
2157 struct swapoff_args {
2167 sys_swapoff(struct thread *td, struct swapoff_args *uap)
2170 struct nameidata nd;
2174 error = priv_check(td, PRIV_SWAPOFF);
2179 while (swdev_syscall_active)
2180 tsleep(&swdev_syscall_active, PUSER - 1, "swpoff", 0);
2181 swdev_syscall_active = 1;
2183 NDINIT(&nd, LOOKUP, FOLLOW | AUDITVNODE1, UIO_USERSPACE, uap->name,
2188 NDFREE(&nd, NDF_ONLY_PNBUF);
2191 mtx_lock(&sw_dev_mtx);
2192 TAILQ_FOREACH(sp, &swtailq, sw_list) {
2193 if (sp->sw_vp == vp)
2196 mtx_unlock(&sw_dev_mtx);
2201 error = swapoff_one(sp, td->td_ucred);
2203 swdev_syscall_active = 0;
2204 wakeup_one(&swdev_syscall_active);
2210 swapoff_one(struct swdevt *sp, struct ucred *cred)
2212 u_long nblks, dvbase;
2217 mtx_assert(&Giant, MA_OWNED);
2219 (void) vn_lock(sp->sw_vp, LK_EXCLUSIVE | LK_RETRY);
2220 error = mac_system_check_swapoff(cred, sp->sw_vp);
2221 (void) VOP_UNLOCK(sp->sw_vp, 0);
2225 nblks = sp->sw_nblks;
2228 * We can turn off this swap device safely only if the
2229 * available virtual memory in the system will fit the amount
2230 * of data we will have to page back in, plus an epsilon so
2231 * the system doesn't become critically low on swap space.
2233 if (vm_cnt.v_free_count + vm_cnt.v_cache_count + swap_pager_avail <
2234 nblks + nswap_lowat) {
2239 * Prevent further allocations on this device.
2241 mtx_lock(&sw_dev_mtx);
2242 sp->sw_flags |= SW_CLOSING;
2243 for (dvbase = 0; dvbase < sp->sw_end; dvbase += dmmax) {
2244 swap_pager_avail -= blist_fill(sp->sw_blist,
2247 swap_total -= (vm_ooffset_t)nblks * PAGE_SIZE;
2248 mtx_unlock(&sw_dev_mtx);
2251 * Page in the contents of the device and close it.
2253 swap_pager_swapoff(sp);
2255 sp->sw_close(curthread, sp);
2256 mtx_lock(&sw_dev_mtx);
2258 TAILQ_REMOVE(&swtailq, sp, sw_list);
2260 if (nswapdev == 0) {
2261 swap_pager_full = 2;
2262 swap_pager_almost_full = 1;
2266 mtx_unlock(&sw_dev_mtx);
2267 blist_destroy(sp->sw_blist);
2268 free(sp, M_VMPGDATA);
2275 struct swdevt *sp, *spt;
2276 const char *devname;
2280 while (swdev_syscall_active)
2281 tsleep(&swdev_syscall_active, PUSER - 1, "swpoff", 0);
2282 swdev_syscall_active = 1;
2284 mtx_lock(&sw_dev_mtx);
2285 TAILQ_FOREACH_SAFE(sp, &swtailq, sw_list, spt) {
2286 mtx_unlock(&sw_dev_mtx);
2287 if (vn_isdisk(sp->sw_vp, NULL))
2288 devname = devtoname(sp->sw_vp->v_rdev);
2291 error = swapoff_one(sp, thread0.td_ucred);
2293 printf("Cannot remove swap device %s (error=%d), "
2294 "skipping.\n", devname, error);
2295 } else if (bootverbose) {
2296 printf("Swap device %s removed.\n", devname);
2298 mtx_lock(&sw_dev_mtx);
2300 mtx_unlock(&sw_dev_mtx);
2302 swdev_syscall_active = 0;
2303 wakeup_one(&swdev_syscall_active);
2308 swap_pager_status(int *total, int *used)
2314 mtx_lock(&sw_dev_mtx);
2315 TAILQ_FOREACH(sp, &swtailq, sw_list) {
2316 *total += sp->sw_nblks;
2317 *used += sp->sw_used;
2319 mtx_unlock(&sw_dev_mtx);
2323 swap_dev_info(int name, struct xswdev *xs, char *devname, size_t len)
2326 const char *tmp_devname;
2331 mtx_lock(&sw_dev_mtx);
2332 TAILQ_FOREACH(sp, &swtailq, sw_list) {
2337 xs->xsw_version = XSWDEV_VERSION;
2338 xs->xsw_dev = sp->sw_dev;
2339 xs->xsw_flags = sp->sw_flags;
2340 xs->xsw_nblks = sp->sw_nblks;
2341 xs->xsw_used = sp->sw_used;
2342 if (devname != NULL) {
2343 if (vn_isdisk(sp->sw_vp, NULL))
2344 tmp_devname = devtoname(sp->sw_vp->v_rdev);
2346 tmp_devname = "[file]";
2347 strncpy(devname, tmp_devname, len);
2352 mtx_unlock(&sw_dev_mtx);
2357 sysctl_vm_swap_info(SYSCTL_HANDLER_ARGS)
2362 if (arg2 != 1) /* name length */
2364 error = swap_dev_info(*(int *)arg1, &xs, NULL, 0);
2367 error = SYSCTL_OUT(req, &xs, sizeof(xs));
2371 SYSCTL_INT(_vm, OID_AUTO, nswapdev, CTLFLAG_RD, &nswapdev, 0,
2372 "Number of swap devices");
2373 SYSCTL_NODE(_vm, OID_AUTO, swap_info, CTLFLAG_RD, sysctl_vm_swap_info,
2374 "Swap statistics by device");
2377 * vmspace_swap_count() - count the approximate swap usage in pages for a
2380 * The map must be locked.
2382 * Swap usage is determined by taking the proportional swap used by
2383 * VM objects backing the VM map. To make up for fractional losses,
2384 * if the VM object has any swap use at all the associated map entries
2385 * count for at least 1 swap page.
2388 vmspace_swap_count(struct vmspace *vmspace)
2395 map = &vmspace->vm_map;
2398 for (cur = map->header.next; cur != &map->header; cur = cur->next) {
2399 if ((cur->eflags & MAP_ENTRY_IS_SUB_MAP) == 0 &&
2400 (object = cur->object.vm_object) != NULL) {
2401 VM_OBJECT_WLOCK(object);
2402 if (object->type == OBJT_SWAP &&
2403 object->un_pager.swp.swp_bcount != 0) {
2404 n = (cur->end - cur->start) / PAGE_SIZE;
2405 count += object->un_pager.swp.swp_bcount *
2406 SWAP_META_PAGES * n / object->size + 1;
2408 VM_OBJECT_WUNLOCK(object);
2417 * Swapping onto disk devices.
2421 static g_orphan_t swapgeom_orphan;
2423 static struct g_class g_swap_class = {
2425 .version = G_VERSION,
2426 .orphan = swapgeom_orphan,
2429 DECLARE_GEOM_CLASS(g_swap_class, g_class);
2433 swapgeom_close_ev(void *arg, int flags)
2435 struct g_consumer *cp;
2438 g_access(cp, -1, -1, 0);
2440 g_destroy_consumer(cp);
2444 * Add a reference to the g_consumer for an inflight transaction.
2447 swapgeom_acquire(struct g_consumer *cp)
2450 mtx_assert(&sw_dev_mtx, MA_OWNED);
2455 * Remove a reference from the g_consumer. Post a close event if
2456 * all referneces go away.
2459 swapgeom_release(struct g_consumer *cp, struct swdevt *sp)
2462 mtx_assert(&sw_dev_mtx, MA_OWNED);
2464 if (cp->index == 0) {
2465 if (g_post_event(swapgeom_close_ev, cp, M_NOWAIT, NULL) == 0)
2471 swapgeom_done(struct bio *bp2)
2475 struct g_consumer *cp;
2477 bp = bp2->bio_caller2;
2479 bp->b_ioflags = bp2->bio_flags;
2481 bp->b_ioflags |= BIO_ERROR;
2482 bp->b_resid = bp->b_bcount - bp2->bio_completed;
2483 bp->b_error = bp2->bio_error;
2485 sp = bp2->bio_caller1;
2486 mtx_lock(&sw_dev_mtx);
2487 swapgeom_release(cp, sp);
2488 mtx_unlock(&sw_dev_mtx);
2493 swapgeom_strategy(struct buf *bp, struct swdevt *sp)
2496 struct g_consumer *cp;
2498 mtx_lock(&sw_dev_mtx);
2501 mtx_unlock(&sw_dev_mtx);
2502 bp->b_error = ENXIO;
2503 bp->b_ioflags |= BIO_ERROR;
2507 swapgeom_acquire(cp);
2508 mtx_unlock(&sw_dev_mtx);
2509 if (bp->b_iocmd == BIO_WRITE)
2512 bio = g_alloc_bio();
2514 mtx_lock(&sw_dev_mtx);
2515 swapgeom_release(cp, sp);
2516 mtx_unlock(&sw_dev_mtx);
2517 bp->b_error = ENOMEM;
2518 bp->b_ioflags |= BIO_ERROR;
2523 bio->bio_caller1 = sp;
2524 bio->bio_caller2 = bp;
2525 bio->bio_cmd = bp->b_iocmd;
2526 bio->bio_offset = (bp->b_blkno - sp->sw_first) * PAGE_SIZE;
2527 bio->bio_length = bp->b_bcount;
2528 bio->bio_done = swapgeom_done;
2529 if (!buf_mapped(bp)) {
2530 bio->bio_ma = bp->b_pages;
2531 bio->bio_data = unmapped_buf;
2532 bio->bio_ma_offset = (vm_offset_t)bp->b_offset & PAGE_MASK;
2533 bio->bio_ma_n = bp->b_npages;
2534 bio->bio_flags |= BIO_UNMAPPED;
2536 bio->bio_data = bp->b_data;
2539 g_io_request(bio, cp);
2544 swapgeom_orphan(struct g_consumer *cp)
2549 mtx_lock(&sw_dev_mtx);
2550 TAILQ_FOREACH(sp, &swtailq, sw_list) {
2551 if (sp->sw_id == cp) {
2552 sp->sw_flags |= SW_CLOSING;
2557 * Drop reference we were created with. Do directly since we're in a
2558 * special context where we don't have to queue the call to
2559 * swapgeom_close_ev().
2562 destroy = ((sp != NULL) && (cp->index == 0));
2565 mtx_unlock(&sw_dev_mtx);
2567 swapgeom_close_ev(cp, 0);
2571 swapgeom_close(struct thread *td, struct swdevt *sw)
2573 struct g_consumer *cp;
2575 mtx_lock(&sw_dev_mtx);
2578 mtx_unlock(&sw_dev_mtx);
2579 /* XXX: direct call when Giant untangled */
2581 g_waitfor_event(swapgeom_close_ev, cp, M_WAITOK, NULL);
2592 swapongeom_ev(void *arg, int flags)
2595 struct g_provider *pp;
2596 struct g_consumer *cp;
2597 static struct g_geom *gp;
2604 pp = g_dev_getprovider(swh->dev);
2606 swh->error = ENODEV;
2609 mtx_lock(&sw_dev_mtx);
2610 TAILQ_FOREACH(sp, &swtailq, sw_list) {
2612 if (cp != NULL && cp->provider == pp) {
2613 mtx_unlock(&sw_dev_mtx);
2618 mtx_unlock(&sw_dev_mtx);
2620 gp = g_new_geomf(&g_swap_class, "swap");
2621 cp = g_new_consumer(gp);
2622 cp->index = 1; /* Number of active I/Os, plus one for being active. */
2623 cp->flags |= G_CF_DIRECT_SEND | G_CF_DIRECT_RECEIVE;
2626 * XXX: Everytime you think you can improve the margin for
2627 * footshooting, somebody depends on the ability to do so:
2628 * savecore(8) wants to write to our swapdev so we cannot
2629 * set an exclusive count :-(
2631 error = g_access(cp, 1, 1, 0);
2634 g_destroy_consumer(cp);
2638 nblks = pp->mediasize / DEV_BSIZE;
2639 swaponsomething(swh->vp, cp, nblks, swapgeom_strategy,
2640 swapgeom_close, dev2udev(swh->dev),
2641 (pp->flags & G_PF_ACCEPT_UNMAPPED) != 0 ? SW_UNMAPPED : 0);
2646 swapongeom(struct thread *td, struct vnode *vp)
2651 vn_lock(vp, LK_EXCLUSIVE | LK_RETRY);
2653 swh.dev = vp->v_rdev;
2656 /* XXX: direct call when Giant untangled */
2657 error = g_waitfor_event(swapongeom_ev, &swh, M_WAITOK, NULL);
2667 * This is used mainly for network filesystem (read: probably only tested
2668 * with NFS) swapfiles.
2673 swapdev_strategy(struct buf *bp, struct swdevt *sp)
2677 bp->b_blkno = ctodb(bp->b_blkno - sp->sw_first);
2681 if (bp->b_iocmd == BIO_WRITE) {
2683 bufobj_wdrop(bp->b_bufobj);
2684 bufobj_wref(&vp2->v_bufobj);
2686 if (bp->b_bufobj != &vp2->v_bufobj)
2687 bp->b_bufobj = &vp2->v_bufobj;
2689 bp->b_iooffset = dbtob(bp->b_blkno);
2695 swapdev_close(struct thread *td, struct swdevt *sp)
2698 VOP_CLOSE(sp->sw_vp, FREAD | FWRITE, td->td_ucred, td);
2704 swaponvp(struct thread *td, struct vnode *vp, u_long nblks)
2711 mtx_lock(&sw_dev_mtx);
2712 TAILQ_FOREACH(sp, &swtailq, sw_list) {
2713 if (sp->sw_id == vp) {
2714 mtx_unlock(&sw_dev_mtx);
2718 mtx_unlock(&sw_dev_mtx);
2720 (void) vn_lock(vp, LK_EXCLUSIVE | LK_RETRY);
2722 error = mac_system_check_swapon(td->td_ucred, vp);
2725 error = VOP_OPEN(vp, FREAD | FWRITE, td->td_ucred, td, NULL);
2726 (void) VOP_UNLOCK(vp, 0);
2730 swaponsomething(vp, vp, nblks, swapdev_strategy, swapdev_close,
2736 sysctl_swap_async_max(SYSCTL_HANDLER_ARGS)
2740 new = nsw_wcount_async_max;
2741 error = sysctl_handle_int(oidp, &new, 0, req);
2742 if (error != 0 || req->newptr == NULL)
2745 if (new > nswbuf / 2 || new < 1)
2748 mtx_lock(&pbuf_mtx);
2749 while (nsw_wcount_async_max != new) {
2751 * Adjust difference. If the current async count is too low,
2752 * we will need to sqeeze our update slowly in. Sleep with a
2753 * higher priority than getpbuf() to finish faster.
2755 n = new - nsw_wcount_async_max;
2756 if (nsw_wcount_async + n >= 0) {
2757 nsw_wcount_async += n;
2758 nsw_wcount_async_max += n;
2759 wakeup(&nsw_wcount_async);
2761 nsw_wcount_async_max -= nsw_wcount_async;
2762 nsw_wcount_async = 0;
2763 msleep(&nsw_wcount_async, &pbuf_mtx, PSWP,
2767 mtx_unlock(&pbuf_mtx);