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("$FreeBSD$");
72 #include "opt_compat.h"
76 #include <sys/param.h>
77 #include <sys/systm.h>
79 #include <sys/kernel.h>
85 #include <sys/fcntl.h>
86 #include <sys/mount.h>
87 #include <sys/namei.h>
88 #include <sys/vnode.h>
89 #include <sys/malloc.h>
90 #include <sys/racct.h>
91 #include <sys/resource.h>
92 #include <sys/resourcevar.h>
93 #include <sys/rwlock.h>
94 #include <sys/sysctl.h>
95 #include <sys/sysproto.h>
96 #include <sys/blist.h>
99 #include <sys/vmmeter.h>
101 #include <security/mac/mac_framework.h>
105 #include <vm/vm_map.h>
106 #include <vm/vm_kern.h>
107 #include <vm/vm_object.h>
108 #include <vm/vm_page.h>
109 #include <vm/vm_pager.h>
110 #include <vm/vm_pageout.h>
111 #include <vm/vm_param.h>
112 #include <vm/swap_pager.h>
113 #include <vm/vm_extern.h>
116 #include <geom/geom.h>
119 * MAX_PAGEOUT_CLUSTER must be a power of 2 between 1 and 64.
120 * The 64-page limit is due to the radix code (kern/subr_blist.c).
122 #ifndef MAX_PAGEOUT_CLUSTER
123 #define MAX_PAGEOUT_CLUSTER 32
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 SWAP_META_PAGES 32
138 #define SWAP_META_MASK (SWAP_META_PAGES - 1)
141 struct swblock *swb_hnext;
142 vm_object_t swb_object;
143 vm_pindex_t swb_index;
145 daddr_t swb_pages[SWAP_META_PAGES];
148 static MALLOC_DEFINE(M_VMPGDATA, "vm_pgdata", "swap pager private data");
149 static struct mtx sw_dev_mtx;
150 static TAILQ_HEAD(, swdevt) swtailq = TAILQ_HEAD_INITIALIZER(swtailq);
151 static struct swdevt *swdevhd; /* Allocate from here next */
152 static int nswapdev; /* Number of swap devices */
153 int swap_pager_avail;
154 static struct sx swdev_syscall_lock; /* serialize swap(on|off) */
156 static vm_ooffset_t swap_total;
157 SYSCTL_QUAD(_vm, OID_AUTO, swap_total, CTLFLAG_RD, &swap_total, 0,
158 "Total amount of available swap storage.");
159 static vm_ooffset_t swap_reserved;
160 SYSCTL_QUAD(_vm, OID_AUTO, swap_reserved, CTLFLAG_RD, &swap_reserved, 0,
161 "Amount of swap storage needed to back all allocated anonymous memory.");
162 static int overcommit = 0;
163 SYSCTL_INT(_vm, OID_AUTO, overcommit, CTLFLAG_RW, &overcommit, 0,
164 "Configure virtual memory overcommit behavior. See tuning(7) "
166 static unsigned long swzone;
167 SYSCTL_ULONG(_vm, OID_AUTO, swzone, CTLFLAG_RD, &swzone, 0,
168 "Actual size of swap metadata zone");
169 static unsigned long swap_maxpages;
170 SYSCTL_ULONG(_vm, OID_AUTO, swap_maxpages, CTLFLAG_RD, &swap_maxpages, 0,
171 "Maximum amount of swap supported");
173 /* bits from overcommit */
174 #define SWAP_RESERVE_FORCE_ON (1 << 0)
175 #define SWAP_RESERVE_RLIMIT_ON (1 << 1)
176 #define SWAP_RESERVE_ALLOW_NONWIRED (1 << 2)
179 swap_reserve(vm_ooffset_t incr)
182 return (swap_reserve_by_cred(incr, curthread->td_ucred));
186 swap_reserve_by_cred(vm_ooffset_t incr, struct ucred *cred)
191 static struct timeval lastfail;
194 uip = cred->cr_ruidinfo;
196 if (incr & PAGE_MASK)
197 panic("swap_reserve: & PAGE_MASK");
202 error = racct_add(curproc, RACCT_SWAP, incr);
203 PROC_UNLOCK(curproc);
210 mtx_lock(&sw_dev_mtx);
211 r = swap_reserved + incr;
212 if (overcommit & SWAP_RESERVE_ALLOW_NONWIRED) {
213 s = vm_cnt.v_page_count - vm_cnt.v_free_reserved - vm_cnt.v_wire_count;
218 if ((overcommit & SWAP_RESERVE_FORCE_ON) == 0 || r <= s ||
219 (error = priv_check(curthread, PRIV_VM_SWAP_NOQUOTA)) == 0) {
223 mtx_unlock(&sw_dev_mtx);
226 UIDINFO_VMSIZE_LOCK(uip);
227 if ((overcommit & SWAP_RESERVE_RLIMIT_ON) != 0 &&
228 uip->ui_vmsize + incr > lim_cur(curthread, RLIMIT_SWAP) &&
229 priv_check(curthread, PRIV_VM_SWAP_NORLIMIT))
232 uip->ui_vmsize += incr;
233 UIDINFO_VMSIZE_UNLOCK(uip);
235 mtx_lock(&sw_dev_mtx);
236 swap_reserved -= incr;
237 mtx_unlock(&sw_dev_mtx);
240 if (!res && ppsratecheck(&lastfail, &curfail, 1)) {
241 printf("uid %d, pid %d: swap reservation for %jd bytes failed\n",
242 uip->ui_uid, curproc->p_pid, incr);
248 racct_sub(curproc, RACCT_SWAP, incr);
249 PROC_UNLOCK(curproc);
257 swap_reserve_force(vm_ooffset_t incr)
261 mtx_lock(&sw_dev_mtx);
262 swap_reserved += incr;
263 mtx_unlock(&sw_dev_mtx);
267 racct_add_force(curproc, RACCT_SWAP, incr);
268 PROC_UNLOCK(curproc);
271 uip = curthread->td_ucred->cr_ruidinfo;
273 UIDINFO_VMSIZE_LOCK(uip);
274 uip->ui_vmsize += incr;
275 UIDINFO_VMSIZE_UNLOCK(uip);
276 PROC_UNLOCK(curproc);
280 swap_release(vm_ooffset_t decr)
285 cred = curthread->td_ucred;
286 swap_release_by_cred(decr, cred);
287 PROC_UNLOCK(curproc);
291 swap_release_by_cred(vm_ooffset_t decr, struct ucred *cred)
295 uip = cred->cr_ruidinfo;
297 if (decr & PAGE_MASK)
298 panic("swap_release: & PAGE_MASK");
300 mtx_lock(&sw_dev_mtx);
301 if (swap_reserved < decr)
302 panic("swap_reserved < decr");
303 swap_reserved -= decr;
304 mtx_unlock(&sw_dev_mtx);
306 UIDINFO_VMSIZE_LOCK(uip);
307 if (uip->ui_vmsize < decr)
308 printf("negative vmsize for uid = %d\n", uip->ui_uid);
309 uip->ui_vmsize -= decr;
310 UIDINFO_VMSIZE_UNLOCK(uip);
312 racct_sub_cred(cred, RACCT_SWAP, decr);
315 #define SWM_FREE 0x02 /* free, period */
316 #define SWM_POP 0x04 /* pop out */
318 int swap_pager_full = 2; /* swap space exhaustion (task killing) */
319 static int swap_pager_almost_full = 1; /* swap space exhaustion (w/hysteresis)*/
320 static int nsw_rcount; /* free read buffers */
321 static int nsw_wcount_sync; /* limit write buffers / synchronous */
322 static int nsw_wcount_async; /* limit write buffers / asynchronous */
323 static int nsw_wcount_async_max;/* assigned maximum */
324 static int nsw_cluster_max; /* maximum VOP I/O allowed */
326 static int sysctl_swap_async_max(SYSCTL_HANDLER_ARGS);
327 SYSCTL_PROC(_vm, OID_AUTO, swap_async_max, CTLTYPE_INT | CTLFLAG_RW |
328 CTLFLAG_MPSAFE, NULL, 0, sysctl_swap_async_max, "I",
329 "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 pagerlst swap_pager_object_list[NOBJLISTS];
348 static uma_zone_t swap_zone;
351 * pagerops for OBJT_SWAP - "swap pager". Some ops are also global procedure
352 * calls hooked from other parts of the VM system and do not appear here.
353 * (see vm/swap_pager.h).
356 swap_pager_alloc(void *handle, vm_ooffset_t size,
357 vm_prot_t prot, vm_ooffset_t offset, struct ucred *);
358 static void swap_pager_dealloc(vm_object_t object);
359 static int swap_pager_getpages(vm_object_t, vm_page_t *, int, int *,
361 static int swap_pager_getpages_async(vm_object_t, vm_page_t *, int, int *,
362 int *, pgo_getpages_iodone_t, void *);
363 static void swap_pager_putpages(vm_object_t, vm_page_t *, int, boolean_t, int *);
365 swap_pager_haspage(vm_object_t object, vm_pindex_t pindex, int *before, int *after);
366 static void swap_pager_init(void);
367 static void swap_pager_unswapped(vm_page_t);
368 static void swap_pager_swapoff(struct swdevt *sp);
370 struct pagerops swappagerops = {
371 .pgo_init = swap_pager_init, /* early system initialization of pager */
372 .pgo_alloc = swap_pager_alloc, /* allocate an OBJT_SWAP object */
373 .pgo_dealloc = swap_pager_dealloc, /* deallocate an OBJT_SWAP object */
374 .pgo_getpages = swap_pager_getpages, /* pagein */
375 .pgo_getpages_async = swap_pager_getpages_async, /* pagein (async) */
376 .pgo_putpages = swap_pager_putpages, /* pageout */
377 .pgo_haspage = swap_pager_haspage, /* get backing store status for page */
378 .pgo_pageunswapped = swap_pager_unswapped, /* remove swap related to page */
382 * swap_*() routines are externally accessible. swp_*() routines are
385 static int nswap_lowat = 128; /* in pages, swap_pager_almost_full warn */
386 static int nswap_hiwat = 512; /* in pages, swap_pager_almost_full warn */
388 SYSCTL_INT(_vm, OID_AUTO, dmmax, CTLFLAG_RD, &nsw_cluster_max, 0,
389 "Maximum size of a swap block in pages");
391 static void swp_sizecheck(void);
392 static void swp_pager_async_iodone(struct buf *bp);
393 static int swapongeom(struct vnode *);
394 static int swaponvp(struct thread *, struct vnode *, u_long);
395 static int swapoff_one(struct swdevt *sp, struct ucred *cred);
398 * Swap bitmap functions
400 static void swp_pager_freeswapspace(daddr_t blk, int npages);
401 static daddr_t swp_pager_getswapspace(int npages);
406 static struct swblock **swp_pager_hash(vm_object_t object, vm_pindex_t index);
407 static void swp_pager_meta_build(vm_object_t, vm_pindex_t, daddr_t);
408 static void swp_pager_meta_free(vm_object_t, vm_pindex_t, vm_pindex_t);
409 static void swp_pager_meta_free_all(vm_object_t);
410 static daddr_t swp_pager_meta_ctl(vm_object_t, vm_pindex_t, int);
413 * SWP_SIZECHECK() - update swap_pager_full indication
415 * update the swap_pager_almost_full indication and warn when we are
416 * about to run out of swap space, using lowat/hiwat hysteresis.
418 * Clear swap_pager_full ( task killing ) indication when lowat is met.
420 * No restrictions on call
421 * This routine may not block.
427 if (swap_pager_avail < nswap_lowat) {
428 if (swap_pager_almost_full == 0) {
429 printf("swap_pager: out of swap space\n");
430 swap_pager_almost_full = 1;
434 if (swap_pager_avail > nswap_hiwat)
435 swap_pager_almost_full = 0;
440 * SWP_PAGER_HASH() - hash swap meta data
442 * This is an helper function which hashes the swapblk given
443 * the object and page index. It returns a pointer to a pointer
444 * to the object, or a pointer to a NULL pointer if it could not
447 static struct swblock **
448 swp_pager_hash(vm_object_t object, vm_pindex_t index)
450 struct swblock **pswap;
451 struct swblock *swap;
453 index &= ~(vm_pindex_t)SWAP_META_MASK;
454 pswap = &swhash[(index ^ (int)(intptr_t)object) & swhash_mask];
455 while ((swap = *pswap) != NULL) {
456 if (swap->swb_object == object &&
457 swap->swb_index == index
461 pswap = &swap->swb_hnext;
467 * SWAP_PAGER_INIT() - initialize the swap pager!
469 * Expected to be started from system init. NOTE: This code is run
470 * before much else so be careful what you depend on. Most of the VM
471 * system has yet to be initialized at this point.
474 swap_pager_init(void)
477 * Initialize object lists
481 for (i = 0; i < NOBJLISTS; ++i)
482 TAILQ_INIT(&swap_pager_object_list[i]);
483 mtx_init(&sw_dev_mtx, "swapdev", NULL, MTX_DEF);
484 sx_init(&sw_alloc_sx, "swspsx");
485 sx_init(&swdev_syscall_lock, "swsysc");
489 * SWAP_PAGER_SWAP_INIT() - swap pager initialization from pageout process
491 * Expected to be started from pageout process once, prior to entering
495 swap_pager_swap_init(void)
500 * Number of in-transit swap bp operations. Don't
501 * exhaust the pbufs completely. Make sure we
502 * initialize workable values (0 will work for hysteresis
503 * but it isn't very efficient).
505 * The nsw_cluster_max is constrained by the bp->b_pages[]
506 * array (MAXPHYS/PAGE_SIZE) and our locally defined
507 * MAX_PAGEOUT_CLUSTER. Also be aware that swap ops are
508 * constrained by the swap device interleave stripe size.
510 * Currently we hardwire nsw_wcount_async to 4. This limit is
511 * designed to prevent other I/O from having high latencies due to
512 * our pageout I/O. The value 4 works well for one or two active swap
513 * devices but is probably a little low if you have more. Even so,
514 * a higher value would probably generate only a limited improvement
515 * with three or four active swap devices since the system does not
516 * typically have to pageout at extreme bandwidths. We will want
517 * at least 2 per swap devices, and 4 is a pretty good value if you
518 * have one NFS swap device due to the command/ack latency over NFS.
519 * So it all works out pretty well.
521 nsw_cluster_max = min((MAXPHYS/PAGE_SIZE), MAX_PAGEOUT_CLUSTER);
524 nsw_rcount = (nswbuf + 1) / 2;
525 nsw_wcount_sync = (nswbuf + 3) / 4;
526 nsw_wcount_async = 4;
527 nsw_wcount_async_max = nsw_wcount_async;
528 mtx_unlock(&pbuf_mtx);
531 * Initialize our zone. Right now I'm just guessing on the number
532 * we need based on the number of pages in the system. Each swblock
533 * can hold 32 pages, so this is probably overkill. This reservation
534 * is typically limited to around 32MB by default.
536 n = vm_cnt.v_page_count / 2;
537 if (maxswzone && n > maxswzone / sizeof(struct swblock))
538 n = maxswzone / sizeof(struct swblock);
540 swap_zone = uma_zcreate("SWAPMETA", sizeof(struct swblock), NULL, NULL,
541 NULL, NULL, UMA_ALIGN_PTR, UMA_ZONE_NOFREE | UMA_ZONE_VM);
542 if (swap_zone == NULL)
543 panic("failed to create swap_zone.");
545 if (uma_zone_reserve_kva(swap_zone, n))
548 * if the allocation failed, try a zone two thirds the
549 * size of the previous attempt.
554 printf("Swap zone entries reduced from %lu to %lu.\n", n2, n);
555 swap_maxpages = n * SWAP_META_PAGES;
556 swzone = n * sizeof(struct swblock);
560 * Initialize our meta-data hash table. The swapper does not need to
561 * be quite as efficient as the VM system, so we do not use an
562 * oversized hash table.
564 * n: size of hash table, must be power of 2
565 * swhash_mask: hash table index mask
567 for (n = 1; n < n2 / 8; n *= 2)
569 swhash = malloc(sizeof(struct swblock *) * n, M_VMPGDATA, M_WAITOK | M_ZERO);
571 mtx_init(&swhash_mtx, "swap_pager swhash", NULL, MTX_DEF);
575 swap_pager_alloc_init(void *handle, struct ucred *cred, vm_ooffset_t size,
581 if (!swap_reserve_by_cred(size, cred))
585 object = vm_object_allocate(OBJT_SWAP, OFF_TO_IDX(offset +
587 object->handle = handle;
590 object->charge = size;
592 object->un_pager.swp.swp_bcount = 0;
597 * SWAP_PAGER_ALLOC() - allocate a new OBJT_SWAP VM object and instantiate
598 * its metadata structures.
600 * This routine is called from the mmap and fork code to create a new
603 * This routine must ensure that no live duplicate is created for
604 * the named object request, which is protected against by
605 * holding the sw_alloc_sx lock in case handle != NULL.
608 swap_pager_alloc(void *handle, vm_ooffset_t size, vm_prot_t prot,
609 vm_ooffset_t offset, struct ucred *cred)
613 if (handle != NULL) {
615 * Reference existing named region or allocate new one. There
616 * should not be a race here against swp_pager_meta_build()
617 * as called from vm_page_remove() in regards to the lookup
620 sx_xlock(&sw_alloc_sx);
621 object = vm_pager_object_lookup(NOBJLIST(handle), handle);
622 if (object == NULL) {
623 object = swap_pager_alloc_init(handle, cred, size,
625 if (object != NULL) {
626 TAILQ_INSERT_TAIL(NOBJLIST(object->handle),
627 object, pager_object_list);
630 sx_xunlock(&sw_alloc_sx);
632 object = swap_pager_alloc_init(handle, cred, size, offset);
638 * SWAP_PAGER_DEALLOC() - remove swap metadata from object
640 * The swap backing for the object is destroyed. The code is
641 * designed such that we can reinstantiate it later, but this
642 * routine is typically called only when the entire object is
643 * about to be destroyed.
645 * The object must be locked.
648 swap_pager_dealloc(vm_object_t object)
651 VM_OBJECT_ASSERT_WLOCKED(object);
652 KASSERT((object->flags & OBJ_DEAD) != 0, ("dealloc of reachable obj"));
655 * Remove from list right away so lookups will fail if we block for
656 * pageout completion.
658 if (object->handle != NULL) {
659 VM_OBJECT_WUNLOCK(object);
660 sx_xlock(&sw_alloc_sx);
661 TAILQ_REMOVE(NOBJLIST(object->handle), object,
663 sx_xunlock(&sw_alloc_sx);
664 VM_OBJECT_WLOCK(object);
667 vm_object_pip_wait(object, "swpdea");
670 * Free all remaining metadata. We only bother to free it from
671 * the swap meta data. We do not attempt to free swapblk's still
672 * associated with vm_page_t's for this object. We do not care
673 * if paging is still in progress on some objects.
675 swp_pager_meta_free_all(object);
676 object->handle = NULL;
677 object->type = OBJT_DEAD;
680 /************************************************************************
681 * SWAP PAGER BITMAP ROUTINES *
682 ************************************************************************/
685 * SWP_PAGER_GETSWAPSPACE() - allocate raw swap space
687 * Allocate swap for the requested number of pages. The starting
688 * swap block number (a page index) is returned or SWAPBLK_NONE
689 * if the allocation failed.
691 * Also has the side effect of advising that somebody made a mistake
692 * when they configured swap and didn't configure enough.
694 * This routine may not sleep.
696 * We allocate in round-robin fashion from the configured devices.
699 swp_pager_getswapspace(int npages)
706 mtx_lock(&sw_dev_mtx);
708 for (i = 0; i < nswapdev; i++) {
710 sp = TAILQ_FIRST(&swtailq);
711 if (!(sp->sw_flags & SW_CLOSING)) {
712 blk = blist_alloc(sp->sw_blist, npages);
713 if (blk != SWAPBLK_NONE) {
715 sp->sw_used += npages;
716 swap_pager_avail -= npages;
718 swdevhd = TAILQ_NEXT(sp, sw_list);
722 sp = TAILQ_NEXT(sp, sw_list);
724 if (swap_pager_full != 2) {
725 printf("swap_pager_getswapspace(%d): failed\n", npages);
727 swap_pager_almost_full = 1;
731 mtx_unlock(&sw_dev_mtx);
736 swp_pager_isondev(daddr_t blk, struct swdevt *sp)
739 return (blk >= sp->sw_first && blk < sp->sw_end);
743 swp_pager_strategy(struct buf *bp)
747 mtx_lock(&sw_dev_mtx);
748 TAILQ_FOREACH(sp, &swtailq, sw_list) {
749 if (bp->b_blkno >= sp->sw_first && bp->b_blkno < sp->sw_end) {
750 mtx_unlock(&sw_dev_mtx);
751 if ((sp->sw_flags & SW_UNMAPPED) != 0 &&
752 unmapped_buf_allowed) {
753 bp->b_data = unmapped_buf;
756 pmap_qenter((vm_offset_t)bp->b_data,
757 &bp->b_pages[0], bp->b_bcount / PAGE_SIZE);
759 sp->sw_strategy(bp, sp);
763 panic("Swapdev not found");
768 * SWP_PAGER_FREESWAPSPACE() - free raw swap space
770 * This routine returns the specified swap blocks back to the bitmap.
772 * This routine may not sleep.
775 swp_pager_freeswapspace(daddr_t blk, int npages)
779 mtx_lock(&sw_dev_mtx);
780 TAILQ_FOREACH(sp, &swtailq, sw_list) {
781 if (blk >= sp->sw_first && blk < sp->sw_end) {
782 sp->sw_used -= npages;
784 * If we are attempting to stop swapping on
785 * this device, we don't want to mark any
786 * blocks free lest they be reused.
788 if ((sp->sw_flags & SW_CLOSING) == 0) {
789 blist_free(sp->sw_blist, blk - sp->sw_first,
791 swap_pager_avail += npages;
794 mtx_unlock(&sw_dev_mtx);
798 panic("Swapdev not found");
802 * SWAP_PAGER_FREESPACE() - frees swap blocks associated with a page
803 * range within an object.
805 * This is a globally accessible routine.
807 * This routine removes swapblk assignments from swap metadata.
809 * The external callers of this routine typically have already destroyed
810 * or renamed vm_page_t's associated with this range in the object so
813 * The object must be locked.
816 swap_pager_freespace(vm_object_t object, vm_pindex_t start, vm_size_t size)
819 swp_pager_meta_free(object, start, size);
823 * SWAP_PAGER_RESERVE() - reserve swap blocks in object
825 * Assigns swap blocks to the specified range within the object. The
826 * swap blocks are not zeroed. Any previous swap assignment is destroyed.
828 * Returns 0 on success, -1 on failure.
831 swap_pager_reserve(vm_object_t object, vm_pindex_t start, vm_size_t size)
834 daddr_t blk = SWAPBLK_NONE;
835 vm_pindex_t beg = start; /* save start index */
837 VM_OBJECT_WLOCK(object);
841 while ((blk = swp_pager_getswapspace(n)) == SWAPBLK_NONE) {
844 swp_pager_meta_free(object, beg, start - beg);
845 VM_OBJECT_WUNLOCK(object);
850 swp_pager_meta_build(object, start, blk);
856 swp_pager_meta_free(object, start, n);
857 VM_OBJECT_WUNLOCK(object);
862 * SWAP_PAGER_COPY() - copy blocks from source pager to destination pager
863 * and destroy the source.
865 * Copy any valid swapblks from the source to the destination. In
866 * cases where both the source and destination have a valid swapblk,
867 * we keep the destination's.
869 * This routine is allowed to sleep. It may sleep allocating metadata
870 * indirectly through swp_pager_meta_build() or if paging is still in
871 * progress on the source.
873 * The source object contains no vm_page_t's (which is just as well)
875 * The source object is of type OBJT_SWAP.
877 * The source and destination objects must be locked.
878 * Both object locks may temporarily be released.
881 swap_pager_copy(vm_object_t srcobject, vm_object_t dstobject,
882 vm_pindex_t offset, int destroysource)
886 VM_OBJECT_ASSERT_WLOCKED(srcobject);
887 VM_OBJECT_ASSERT_WLOCKED(dstobject);
890 * If destroysource is set, we remove the source object from the
891 * swap_pager internal queue now.
893 if (destroysource && srcobject->handle != NULL) {
894 vm_object_pip_add(srcobject, 1);
895 VM_OBJECT_WUNLOCK(srcobject);
896 vm_object_pip_add(dstobject, 1);
897 VM_OBJECT_WUNLOCK(dstobject);
898 sx_xlock(&sw_alloc_sx);
899 TAILQ_REMOVE(NOBJLIST(srcobject->handle), srcobject,
901 sx_xunlock(&sw_alloc_sx);
902 VM_OBJECT_WLOCK(dstobject);
903 vm_object_pip_wakeup(dstobject);
904 VM_OBJECT_WLOCK(srcobject);
905 vm_object_pip_wakeup(srcobject);
909 * transfer source to destination.
911 for (i = 0; i < dstobject->size; ++i) {
915 * Locate (without changing) the swapblk on the destination,
916 * unless it is invalid in which case free it silently, or
917 * if the destination is a resident page, in which case the
918 * source is thrown away.
920 dstaddr = swp_pager_meta_ctl(dstobject, i, 0);
922 if (dstaddr == SWAPBLK_NONE) {
924 * Destination has no swapblk and is not resident,
929 srcaddr = swp_pager_meta_ctl(
935 if (srcaddr != SWAPBLK_NONE) {
937 * swp_pager_meta_build() can sleep.
939 vm_object_pip_add(srcobject, 1);
940 VM_OBJECT_WUNLOCK(srcobject);
941 vm_object_pip_add(dstobject, 1);
942 swp_pager_meta_build(dstobject, i, srcaddr);
943 vm_object_pip_wakeup(dstobject);
944 VM_OBJECT_WLOCK(srcobject);
945 vm_object_pip_wakeup(srcobject);
949 * Destination has valid swapblk or it is represented
950 * by a resident page. We destroy the sourceblock.
953 swp_pager_meta_ctl(srcobject, i + offset, SWM_FREE);
958 * Free left over swap blocks in source.
960 * We have to revert the type to OBJT_DEFAULT so we do not accidentally
961 * double-remove the object from the swap queues.
964 swp_pager_meta_free_all(srcobject);
966 * Reverting the type is not necessary, the caller is going
967 * to destroy srcobject directly, but I'm doing it here
968 * for consistency since we've removed the object from its
971 srcobject->type = OBJT_DEFAULT;
976 * SWAP_PAGER_HASPAGE() - determine if we have good backing store for
977 * the requested page.
979 * We determine whether good backing store exists for the requested
980 * page and return TRUE if it does, FALSE if it doesn't.
982 * If TRUE, we also try to determine how much valid, contiguous backing
983 * store exists before and after the requested page.
986 swap_pager_haspage(vm_object_t object, vm_pindex_t pindex, int *before,
992 VM_OBJECT_ASSERT_LOCKED(object);
995 * do we have good backing store at the requested index ?
997 blk0 = swp_pager_meta_ctl(object, pindex, 0);
998 if (blk0 == SWAPBLK_NONE) {
1007 * find backwards-looking contiguous good backing store
1009 if (before != NULL) {
1010 for (i = 1; i < SWB_NPAGES; i++) {
1013 blk = swp_pager_meta_ctl(object, pindex - i, 0);
1014 if (blk != blk0 - i)
1021 * find forward-looking contiguous good backing store
1023 if (after != NULL) {
1024 for (i = 1; i < SWB_NPAGES; i++) {
1025 blk = swp_pager_meta_ctl(object, pindex + i, 0);
1026 if (blk != blk0 + i)
1035 * SWAP_PAGER_PAGE_UNSWAPPED() - remove swap backing store related to page
1037 * This removes any associated swap backing store, whether valid or
1038 * not, from the page.
1040 * This routine is typically called when a page is made dirty, at
1041 * which point any associated swap can be freed. MADV_FREE also
1042 * calls us in a special-case situation
1044 * NOTE!!! If the page is clean and the swap was valid, the caller
1045 * should make the page dirty before calling this routine. This routine
1046 * does NOT change the m->dirty status of the page. Also: MADV_FREE
1049 * This routine may not sleep.
1051 * The object containing the page must be locked.
1054 swap_pager_unswapped(vm_page_t m)
1057 swp_pager_meta_ctl(m->object, m->pindex, SWM_FREE);
1061 * swap_pager_getpages() - bring pages in from swap
1063 * Attempt to page in the pages in array "m" of length "count". The caller
1064 * may optionally specify that additional pages preceding and succeeding
1065 * the specified range be paged in. The number of such pages is returned
1066 * in the "rbehind" and "rahead" parameters, and they will be in the
1067 * inactive queue upon return.
1069 * The pages in "m" must be busied and will remain busied upon return.
1072 swap_pager_getpages(vm_object_t object, vm_page_t *m, int count, int *rbehind,
1076 vm_page_t mpred, msucc, p;
1079 int i, j, maxahead, maxbehind, reqcount, shift;
1083 VM_OBJECT_WUNLOCK(object);
1084 bp = getpbuf(&nsw_rcount);
1085 VM_OBJECT_WLOCK(object);
1087 if (!swap_pager_haspage(object, m[0]->pindex, &maxbehind, &maxahead)) {
1088 relpbuf(bp, &nsw_rcount);
1089 return (VM_PAGER_FAIL);
1093 * Clip the readahead and readbehind ranges to exclude resident pages.
1095 if (rahead != NULL) {
1096 KASSERT(reqcount - 1 <= maxahead,
1097 ("page count %d extends beyond swap block", reqcount));
1098 *rahead = imin(*rahead, maxahead - (reqcount - 1));
1099 pindex = m[reqcount - 1]->pindex;
1100 msucc = TAILQ_NEXT(m[reqcount - 1], listq);
1101 if (msucc != NULL && msucc->pindex - pindex - 1 < *rahead)
1102 *rahead = msucc->pindex - pindex - 1;
1104 if (rbehind != NULL) {
1105 *rbehind = imin(*rbehind, maxbehind);
1106 pindex = m[0]->pindex;
1107 mpred = TAILQ_PREV(m[0], pglist, listq);
1108 if (mpred != NULL && pindex - mpred->pindex - 1 < *rbehind)
1109 *rbehind = pindex - mpred->pindex - 1;
1113 * Allocate readahead and readbehind pages.
1115 shift = rbehind != NULL ? *rbehind : 0;
1117 for (i = 1; i <= shift; i++) {
1118 p = vm_page_alloc(object, m[0]->pindex - i,
1121 /* Shift allocated pages to the left. */
1122 for (j = 0; j < i - 1; j++)
1124 bp->b_pages[j + shift - i + 1];
1127 bp->b_pages[shift - i] = p;
1132 for (i = 0; i < reqcount; i++)
1133 bp->b_pages[i + shift] = m[i];
1134 if (rahead != NULL) {
1135 for (i = 0; i < *rahead; i++) {
1136 p = vm_page_alloc(object,
1137 m[reqcount - 1]->pindex + i + 1, VM_ALLOC_NORMAL);
1140 bp->b_pages[shift + reqcount + i] = p;
1144 if (rbehind != NULL)
1149 vm_object_pip_add(object, count);
1151 for (i = 0; i < count; i++)
1152 bp->b_pages[i]->oflags |= VPO_SWAPINPROG;
1154 pindex = bp->b_pages[0]->pindex;
1155 blk = swp_pager_meta_ctl(object, pindex, 0);
1156 KASSERT(blk != SWAPBLK_NONE,
1157 ("no swap blocking containing %p(%jx)", object, (uintmax_t)pindex));
1159 VM_OBJECT_WUNLOCK(object);
1161 bp->b_flags |= B_PAGING;
1162 bp->b_iocmd = BIO_READ;
1163 bp->b_iodone = swp_pager_async_iodone;
1164 bp->b_rcred = crhold(thread0.td_ucred);
1165 bp->b_wcred = crhold(thread0.td_ucred);
1167 bp->b_bcount = PAGE_SIZE * count;
1168 bp->b_bufsize = PAGE_SIZE * count;
1169 bp->b_npages = count;
1170 bp->b_pgbefore = rbehind != NULL ? *rbehind : 0;
1171 bp->b_pgafter = rahead != NULL ? *rahead : 0;
1173 VM_CNT_INC(v_swapin);
1174 VM_CNT_ADD(v_swappgsin, count);
1177 * perform the I/O. NOTE!!! bp cannot be considered valid after
1178 * this point because we automatically release it on completion.
1179 * Instead, we look at the one page we are interested in which we
1180 * still hold a lock on even through the I/O completion.
1182 * The other pages in our m[] array are also released on completion,
1183 * so we cannot assume they are valid anymore either.
1185 * NOTE: b_blkno is destroyed by the call to swapdev_strategy
1188 swp_pager_strategy(bp);
1191 * Wait for the pages we want to complete. VPO_SWAPINPROG is always
1192 * cleared on completion. If an I/O error occurs, SWAPBLK_NONE
1193 * is set in the metadata for each page in the request.
1195 VM_OBJECT_WLOCK(object);
1196 while ((m[0]->oflags & VPO_SWAPINPROG) != 0) {
1197 m[0]->oflags |= VPO_SWAPSLEEP;
1198 VM_CNT_INC(v_intrans);
1199 if (VM_OBJECT_SLEEP(object, &object->paging_in_progress, PSWP,
1200 "swread", hz * 20)) {
1202 "swap_pager: indefinite wait buffer: bufobj: %p, blkno: %jd, size: %ld\n",
1203 bp->b_bufobj, (intmax_t)bp->b_blkno, bp->b_bcount);
1208 * If we had an unrecoverable read error pages will not be valid.
1210 for (i = 0; i < reqcount; i++)
1211 if (m[i]->valid != VM_PAGE_BITS_ALL)
1212 return (VM_PAGER_ERROR);
1214 return (VM_PAGER_OK);
1217 * A final note: in a low swap situation, we cannot deallocate swap
1218 * and mark a page dirty here because the caller is likely to mark
1219 * the page clean when we return, causing the page to possibly revert
1220 * to all-zero's later.
1225 * swap_pager_getpages_async():
1227 * Right now this is emulation of asynchronous operation on top of
1228 * swap_pager_getpages().
1231 swap_pager_getpages_async(vm_object_t object, vm_page_t *m, int count,
1232 int *rbehind, int *rahead, pgo_getpages_iodone_t iodone, void *arg)
1236 r = swap_pager_getpages(object, m, count, rbehind, rahead);
1237 VM_OBJECT_WUNLOCK(object);
1242 case VM_PAGER_ERROR:
1249 panic("unhandled swap_pager_getpages() error %d", r);
1251 (iodone)(arg, m, count, error);
1252 VM_OBJECT_WLOCK(object);
1258 * swap_pager_putpages:
1260 * Assign swap (if necessary) and initiate I/O on the specified pages.
1262 * We support both OBJT_DEFAULT and OBJT_SWAP objects. DEFAULT objects
1263 * are automatically converted to SWAP objects.
1265 * In a low memory situation we may block in VOP_STRATEGY(), but the new
1266 * vm_page reservation system coupled with properly written VFS devices
1267 * should ensure that no low-memory deadlock occurs. This is an area
1270 * The parent has N vm_object_pip_add() references prior to
1271 * calling us and will remove references for rtvals[] that are
1272 * not set to VM_PAGER_PEND. We need to remove the rest on I/O
1275 * The parent has soft-busy'd the pages it passes us and will unbusy
1276 * those whos rtvals[] entry is not set to VM_PAGER_PEND on return.
1277 * We need to unbusy the rest on I/O completion.
1280 swap_pager_putpages(vm_object_t object, vm_page_t *m, int count,
1281 int flags, int *rtvals)
1286 if (count && m[0]->object != object) {
1287 panic("swap_pager_putpages: object mismatch %p/%p",
1296 * Turn object into OBJT_SWAP
1297 * check for bogus sysops
1298 * force sync if not pageout process
1300 if (object->type != OBJT_SWAP)
1301 swp_pager_meta_build(object, 0, SWAPBLK_NONE);
1302 VM_OBJECT_WUNLOCK(object);
1305 if (curproc != pageproc)
1308 sync = (flags & VM_PAGER_PUT_SYNC) != 0;
1313 * Assign swap blocks and issue I/O. We reallocate swap on the fly.
1314 * The page is left dirty until the pageout operation completes
1317 for (i = 0; i < count; i += n) {
1323 * Maximum I/O size is limited by a number of factors.
1325 n = min(BLIST_MAX_ALLOC, count - i);
1326 n = min(n, nsw_cluster_max);
1329 * Get biggest block of swap we can. If we fail, fall
1330 * back and try to allocate a smaller block. Don't go
1331 * overboard trying to allocate space if it would overly
1335 (blk = swp_pager_getswapspace(n)) == SWAPBLK_NONE &&
1340 if (blk == SWAPBLK_NONE) {
1341 for (j = 0; j < n; ++j)
1342 rtvals[i+j] = VM_PAGER_FAIL;
1347 * All I/O parameters have been satisfied, build the I/O
1348 * request and assign the swap space.
1351 bp = getpbuf(&nsw_wcount_sync);
1353 bp = getpbuf(&nsw_wcount_async);
1354 bp->b_flags = B_ASYNC;
1356 bp->b_flags |= B_PAGING;
1357 bp->b_iocmd = BIO_WRITE;
1359 bp->b_rcred = crhold(thread0.td_ucred);
1360 bp->b_wcred = crhold(thread0.td_ucred);
1361 bp->b_bcount = PAGE_SIZE * n;
1362 bp->b_bufsize = PAGE_SIZE * n;
1365 VM_OBJECT_WLOCK(object);
1366 for (j = 0; j < n; ++j) {
1367 vm_page_t mreq = m[i+j];
1369 swp_pager_meta_build(
1374 MPASS(mreq->dirty == VM_PAGE_BITS_ALL);
1375 mreq->oflags |= VPO_SWAPINPROG;
1376 bp->b_pages[j] = mreq;
1378 VM_OBJECT_WUNLOCK(object);
1381 * Must set dirty range for NFS to work.
1384 bp->b_dirtyend = bp->b_bcount;
1386 VM_CNT_INC(v_swapout);
1387 VM_CNT_ADD(v_swappgsout, bp->b_npages);
1390 * We unconditionally set rtvals[] to VM_PAGER_PEND so that we
1391 * can call the async completion routine at the end of a
1392 * synchronous I/O operation. Otherwise, our caller would
1393 * perform duplicate unbusy and wakeup operations on the page
1394 * and object, respectively.
1396 for (j = 0; j < n; j++)
1397 rtvals[i + j] = VM_PAGER_PEND;
1402 * NOTE: b_blkno is destroyed by the call to swapdev_strategy
1404 if (sync == FALSE) {
1405 bp->b_iodone = swp_pager_async_iodone;
1407 swp_pager_strategy(bp);
1414 * NOTE: b_blkno is destroyed by the call to swapdev_strategy
1416 bp->b_iodone = bdone;
1417 swp_pager_strategy(bp);
1420 * Wait for the sync I/O to complete.
1422 bwait(bp, PVM, "swwrt");
1425 * Now that we are through with the bp, we can call the
1426 * normal async completion, which frees everything up.
1428 swp_pager_async_iodone(bp);
1430 VM_OBJECT_WLOCK(object);
1434 * swp_pager_async_iodone:
1436 * Completion routine for asynchronous reads and writes from/to swap.
1437 * Also called manually by synchronous code to finish up a bp.
1439 * This routine may not sleep.
1442 swp_pager_async_iodone(struct buf *bp)
1445 vm_object_t object = NULL;
1450 if (bp->b_ioflags & BIO_ERROR) {
1452 "swap_pager: I/O error - %s failed; blkno %ld,"
1453 "size %ld, error %d\n",
1454 ((bp->b_iocmd == BIO_READ) ? "pagein" : "pageout"),
1462 * remove the mapping for kernel virtual
1465 pmap_qremove((vm_offset_t)bp->b_data, bp->b_npages);
1467 bp->b_data = bp->b_kvabase;
1470 object = bp->b_pages[0]->object;
1471 VM_OBJECT_WLOCK(object);
1475 * cleanup pages. If an error occurs writing to swap, we are in
1476 * very serious trouble. If it happens to be a disk error, though,
1477 * we may be able to recover by reassigning the swap later on. So
1478 * in this case we remove the m->swapblk assignment for the page
1479 * but do not free it in the rlist. The errornous block(s) are thus
1480 * never reallocated as swap. Redirty the page and continue.
1482 for (i = 0; i < bp->b_npages; ++i) {
1483 vm_page_t m = bp->b_pages[i];
1485 m->oflags &= ~VPO_SWAPINPROG;
1486 if (m->oflags & VPO_SWAPSLEEP) {
1487 m->oflags &= ~VPO_SWAPSLEEP;
1488 wakeup(&object->paging_in_progress);
1491 if (bp->b_ioflags & BIO_ERROR) {
1493 * If an error occurs I'd love to throw the swapblk
1494 * away without freeing it back to swapspace, so it
1495 * can never be used again. But I can't from an
1498 if (bp->b_iocmd == BIO_READ) {
1500 * NOTE: for reads, m->dirty will probably
1501 * be overridden by the original caller of
1502 * getpages so don't play cute tricks here.
1507 * If a write error occurs, reactivate page
1508 * so it doesn't clog the inactive list,
1509 * then finish the I/O.
1513 vm_page_activate(m);
1517 } else if (bp->b_iocmd == BIO_READ) {
1519 * NOTE: for reads, m->dirty will probably be
1520 * overridden by the original caller of getpages so
1521 * we cannot set them in order to free the underlying
1522 * swap in a low-swap situation. I don't think we'd
1523 * want to do that anyway, but it was an optimization
1524 * that existed in the old swapper for a time before
1525 * it got ripped out due to precisely this problem.
1527 KASSERT(!pmap_page_is_mapped(m),
1528 ("swp_pager_async_iodone: page %p is mapped", m));
1529 KASSERT(m->dirty == 0,
1530 ("swp_pager_async_iodone: page %p is dirty", m));
1532 m->valid = VM_PAGE_BITS_ALL;
1533 if (i < bp->b_pgbefore ||
1534 i >= bp->b_npages - bp->b_pgafter)
1535 vm_page_readahead_finish(m);
1538 * For write success, clear the dirty
1539 * status, then finish the I/O ( which decrements the
1540 * busy count and possibly wakes waiter's up ).
1541 * A page is only written to swap after a period of
1542 * inactivity. Therefore, we do not expect it to be
1545 KASSERT(!pmap_page_is_write_mapped(m),
1546 ("swp_pager_async_iodone: page %p is not write"
1550 vm_page_deactivate_noreuse(m);
1557 * adjust pip. NOTE: the original parent may still have its own
1558 * pip refs on the object.
1560 if (object != NULL) {
1561 vm_object_pip_wakeupn(object, bp->b_npages);
1562 VM_OBJECT_WUNLOCK(object);
1566 * swapdev_strategy() manually sets b_vp and b_bufobj before calling
1567 * bstrategy(). Set them back to NULL now we're done with it, or we'll
1568 * trigger a KASSERT in relpbuf().
1572 bp->b_bufobj = NULL;
1575 * release the physical I/O buffer
1579 ((bp->b_iocmd == BIO_READ) ? &nsw_rcount :
1580 ((bp->b_flags & B_ASYNC) ?
1589 swap_pager_nswapdev(void)
1596 * SWP_PAGER_FORCE_PAGEIN() - force a swap block to be paged in
1598 * This routine dissociates the page at the given index within an object
1599 * from its backing store, paging it in if it does not reside in memory.
1600 * If the page is paged in, it is marked dirty and placed in the laundry
1601 * queue. The page is marked dirty because it no longer has backing
1602 * store. It is placed in the laundry queue because it has not been
1603 * accessed recently. Otherwise, it would already reside in memory.
1605 * We also attempt to swap in all other pages in the swap block.
1606 * However, we only guarantee that the one at the specified index is
1609 * XXX - The code to page the whole block in doesn't work, so we
1610 * revert to the one-by-one behavior for now. Sigh.
1613 swp_pager_force_pagein(vm_object_t object, vm_pindex_t pindex)
1617 vm_object_pip_add(object, 1);
1618 m = vm_page_grab(object, pindex, VM_ALLOC_NORMAL);
1619 if (m->valid == VM_PAGE_BITS_ALL) {
1620 vm_object_pip_wakeup(object);
1623 vm_page_activate(m);
1626 vm_pager_page_unswapped(m);
1630 if (swap_pager_getpages(object, &m, 1, NULL, NULL) != VM_PAGER_OK)
1631 panic("swap_pager_force_pagein: read from swap failed");/*XXX*/
1632 vm_object_pip_wakeup(object);
1638 vm_pager_page_unswapped(m);
1642 * swap_pager_swapoff:
1644 * Page in all of the pages that have been paged out to the
1645 * given device. The corresponding blocks in the bitmap must be
1646 * marked as allocated and the device must be flagged SW_CLOSING.
1647 * There may be no processes swapped out to the device.
1649 * This routine may block.
1652 swap_pager_swapoff(struct swdevt *sp)
1654 struct swblock *swap;
1655 vm_object_t locked_obj, object;
1659 sx_assert(&swdev_syscall_lock, SA_XLOCKED);
1664 mtx_lock(&swhash_mtx);
1665 for (i = 0; i <= swhash_mask; i++) { /* '<=' is correct here */
1667 for (swap = swhash[i]; swap != NULL; swap = swap->swb_hnext) {
1668 object = swap->swb_object;
1669 pindex = swap->swb_index;
1670 for (j = 0; j < SWAP_META_PAGES; ++j) {
1671 if (!swp_pager_isondev(swap->swb_pages[j], sp))
1673 if (locked_obj != object) {
1674 if (locked_obj != NULL)
1675 VM_OBJECT_WUNLOCK(locked_obj);
1676 locked_obj = object;
1677 if (!VM_OBJECT_TRYWLOCK(object)) {
1678 mtx_unlock(&swhash_mtx);
1679 /* Depends on type-stability. */
1680 VM_OBJECT_WLOCK(object);
1681 mtx_lock(&swhash_mtx);
1685 MPASS(locked_obj == object);
1686 mtx_unlock(&swhash_mtx);
1687 swp_pager_force_pagein(object, pindex + j);
1688 mtx_lock(&swhash_mtx);
1693 mtx_unlock(&swhash_mtx);
1694 if (locked_obj != NULL) {
1695 VM_OBJECT_WUNLOCK(locked_obj);
1700 * Objects may be locked or paging to the device being
1701 * removed, so we will miss their pages and need to
1702 * make another pass. We have marked this device as
1703 * SW_CLOSING, so the activity should finish soon.
1706 if (retries > 100) {
1707 panic("swapoff: failed to locate %d swap blocks",
1710 pause("swpoff", hz / 20);
1713 EVENTHANDLER_INVOKE(swapoff, sp);
1716 /************************************************************************
1718 ************************************************************************
1720 * These routines manipulate the swap metadata stored in the
1723 * Swap metadata is implemented with a global hash and not directly
1724 * linked into the object. Instead the object simply contains
1725 * appropriate tracking counters.
1729 * SWP_PAGER_META_BUILD() - add swap block to swap meta data for object
1731 * We first convert the object to a swap object if it is a default
1734 * The specified swapblk is added to the object's swap metadata. If
1735 * the swapblk is not valid, it is freed instead. Any previously
1736 * assigned swapblk is freed.
1739 swp_pager_meta_build(vm_object_t object, vm_pindex_t pindex, daddr_t swapblk)
1741 static volatile int exhausted;
1742 struct swblock *swap;
1743 struct swblock **pswap;
1746 VM_OBJECT_ASSERT_WLOCKED(object);
1748 * Convert default object to swap object if necessary
1750 if (object->type != OBJT_SWAP) {
1751 object->type = OBJT_SWAP;
1752 object->un_pager.swp.swp_bcount = 0;
1753 KASSERT(object->handle == NULL, ("default pager with handle"));
1757 * Locate hash entry. If not found create, but if we aren't adding
1758 * anything just return. If we run out of space in the map we wait
1759 * and, since the hash table may have changed, retry.
1762 mtx_lock(&swhash_mtx);
1763 pswap = swp_pager_hash(object, pindex);
1765 if ((swap = *pswap) == NULL) {
1768 if (swapblk == SWAPBLK_NONE)
1771 swap = *pswap = uma_zalloc(swap_zone, M_NOWAIT |
1772 (curproc == pageproc ? M_USE_RESERVE : 0));
1774 mtx_unlock(&swhash_mtx);
1775 VM_OBJECT_WUNLOCK(object);
1776 if (uma_zone_exhausted(swap_zone)) {
1777 if (atomic_cmpset_int(&exhausted, 0, 1))
1778 printf("swap zone exhausted, "
1779 "increase kern.maxswzone\n");
1780 vm_pageout_oom(VM_OOM_SWAPZ);
1781 pause("swzonex", 10);
1784 VM_OBJECT_WLOCK(object);
1788 if (atomic_cmpset_int(&exhausted, 1, 0))
1789 printf("swap zone ok\n");
1791 swap->swb_hnext = NULL;
1792 swap->swb_object = object;
1793 swap->swb_index = pindex & ~(vm_pindex_t)SWAP_META_MASK;
1794 swap->swb_count = 0;
1796 ++object->un_pager.swp.swp_bcount;
1798 for (i = 0; i < SWAP_META_PAGES; ++i)
1799 swap->swb_pages[i] = SWAPBLK_NONE;
1803 * Delete prior contents of metadata
1805 idx = pindex & SWAP_META_MASK;
1807 if (swap->swb_pages[idx] != SWAPBLK_NONE) {
1808 swp_pager_freeswapspace(swap->swb_pages[idx], 1);
1813 * Enter block into metadata
1815 swap->swb_pages[idx] = swapblk;
1816 if (swapblk != SWAPBLK_NONE)
1819 mtx_unlock(&swhash_mtx);
1823 * SWP_PAGER_META_FREE() - free a range of blocks in the object's swap metadata
1825 * The requested range of blocks is freed, with any associated swap
1826 * returned to the swap bitmap.
1828 * This routine will free swap metadata structures as they are cleaned
1829 * out. This routine does *NOT* operate on swap metadata associated
1830 * with resident pages.
1833 swp_pager_meta_free(vm_object_t object, vm_pindex_t index, vm_pindex_t count)
1835 struct swblock **pswap, *swap;
1840 VM_OBJECT_ASSERT_LOCKED(object);
1841 if (object->type != OBJT_SWAP || count == 0)
1844 mtx_lock(&swhash_mtx);
1845 for (c = 0; c < count;) {
1846 pswap = swp_pager_hash(object, index);
1847 sidx = index & SWAP_META_MASK;
1848 n = SWAP_META_PAGES - sidx;
1850 if ((swap = *pswap) == NULL) {
1854 for (; c < count && sidx < SWAP_META_PAGES; ++c, ++sidx) {
1855 if ((v = swap->swb_pages[sidx]) == SWAPBLK_NONE)
1857 swp_pager_freeswapspace(v, 1);
1858 swap->swb_pages[sidx] = SWAPBLK_NONE;
1859 if (--swap->swb_count == 0) {
1860 *pswap = swap->swb_hnext;
1861 uma_zfree(swap_zone, swap);
1862 --object->un_pager.swp.swp_bcount;
1863 c += SWAP_META_PAGES - sidx;
1868 mtx_unlock(&swhash_mtx);
1872 * SWP_PAGER_META_FREE_ALL() - destroy all swap metadata associated with object
1874 * This routine locates and destroys all swap metadata associated with
1878 swp_pager_meta_free_all(vm_object_t object)
1880 struct swblock **pswap, *swap;
1885 VM_OBJECT_ASSERT_WLOCKED(object);
1886 if (object->type != OBJT_SWAP)
1890 while (object->un_pager.swp.swp_bcount != 0) {
1891 mtx_lock(&swhash_mtx);
1892 pswap = swp_pager_hash(object, index);
1893 if ((swap = *pswap) != NULL) {
1894 for (i = 0; i < SWAP_META_PAGES; ++i) {
1895 v = swap->swb_pages[i];
1896 if (v != SWAPBLK_NONE) {
1898 swp_pager_freeswapspace(v, 1);
1901 if (swap->swb_count != 0)
1903 "swap_pager_meta_free_all: swb_count != 0");
1904 *pswap = swap->swb_hnext;
1905 uma_zfree(swap_zone, swap);
1906 --object->un_pager.swp.swp_bcount;
1908 mtx_unlock(&swhash_mtx);
1909 index += SWAP_META_PAGES;
1914 * SWP_PAGER_METACTL() - misc control of swap and vm_page_t meta data.
1916 * This routine is capable of looking up, popping, or freeing
1917 * swapblk assignments in the swap meta data or in the vm_page_t.
1918 * The routine typically returns the swapblk being looked-up, or popped,
1919 * or SWAPBLK_NONE if the block was freed, or SWAPBLK_NONE if the block
1920 * was invalid. This routine will automatically free any invalid
1921 * meta-data swapblks.
1923 * It is not possible to store invalid swapblks in the swap meta data
1924 * (other then a literal 'SWAPBLK_NONE'), so we don't bother checking.
1926 * When acting on a busy resident page and paging is in progress, we
1927 * have to wait until paging is complete but otherwise can act on the
1930 * SWM_FREE remove and free swap block from metadata
1931 * SWM_POP remove from meta data but do not free.. pop it out
1934 swp_pager_meta_ctl(vm_object_t object, vm_pindex_t pindex, int flags)
1936 struct swblock **pswap;
1937 struct swblock *swap;
1941 VM_OBJECT_ASSERT_LOCKED(object);
1943 * The meta data only exists of the object is OBJT_SWAP
1944 * and even then might not be allocated yet.
1946 if (object->type != OBJT_SWAP)
1947 return (SWAPBLK_NONE);
1950 mtx_lock(&swhash_mtx);
1951 pswap = swp_pager_hash(object, pindex);
1953 if ((swap = *pswap) != NULL) {
1954 idx = pindex & SWAP_META_MASK;
1955 r1 = swap->swb_pages[idx];
1957 if (r1 != SWAPBLK_NONE) {
1958 if (flags & SWM_FREE) {
1959 swp_pager_freeswapspace(r1, 1);
1962 if (flags & (SWM_FREE|SWM_POP)) {
1963 swap->swb_pages[idx] = SWAPBLK_NONE;
1964 if (--swap->swb_count == 0) {
1965 *pswap = swap->swb_hnext;
1966 uma_zfree(swap_zone, swap);
1967 --object->un_pager.swp.swp_bcount;
1972 mtx_unlock(&swhash_mtx);
1977 * Returns the least page index which is greater than or equal to the
1978 * parameter pindex and for which there is a swap block allocated.
1979 * Returns object's size if the object's type is not swap or if there
1980 * are no allocated swap blocks for the object after the requested
1984 swap_pager_find_least(vm_object_t object, vm_pindex_t pindex)
1986 struct swblock **pswap, *swap;
1987 vm_pindex_t i, j, lim;
1990 VM_OBJECT_ASSERT_LOCKED(object);
1991 if (object->type != OBJT_SWAP || object->un_pager.swp.swp_bcount == 0)
1992 return (object->size);
1994 mtx_lock(&swhash_mtx);
1995 for (j = pindex; j < object->size; j = lim) {
1996 pswap = swp_pager_hash(object, j);
1997 lim = rounddown2(j + SWAP_META_PAGES, SWAP_META_PAGES);
1998 if (lim > object->size)
2000 if ((swap = *pswap) != NULL) {
2001 for (idx = j & SWAP_META_MASK, i = j; i < lim;
2003 if (swap->swb_pages[idx] != SWAPBLK_NONE)
2010 mtx_unlock(&swhash_mtx);
2015 * System call swapon(name) enables swapping on device name,
2016 * which must be in the swdevsw. Return EBUSY
2017 * if already swapping on this device.
2019 #ifndef _SYS_SYSPROTO_H_
2020 struct swapon_args {
2030 sys_swapon(struct thread *td, struct swapon_args *uap)
2034 struct nameidata nd;
2037 error = priv_check(td, PRIV_SWAPON);
2041 sx_xlock(&swdev_syscall_lock);
2044 * Swap metadata may not fit in the KVM if we have physical
2047 if (swap_zone == NULL) {
2052 NDINIT(&nd, LOOKUP, ISOPEN | FOLLOW | AUDITVNODE1, UIO_USERSPACE,
2058 NDFREE(&nd, NDF_ONLY_PNBUF);
2061 if (vn_isdisk(vp, &error)) {
2062 error = swapongeom(vp);
2063 } else if (vp->v_type == VREG &&
2064 (vp->v_mount->mnt_vfc->vfc_flags & VFCF_NETWORK) != 0 &&
2065 (error = VOP_GETATTR(vp, &attr, td->td_ucred)) == 0) {
2067 * Allow direct swapping to NFS regular files in the same
2068 * way that nfs_mountroot() sets up diskless swapping.
2070 error = swaponvp(td, vp, attr.va_size / DEV_BSIZE);
2076 sx_xunlock(&swdev_syscall_lock);
2081 * Check that the total amount of swap currently configured does not
2082 * exceed half the theoretical maximum. If it does, print a warning
2083 * message and return -1; otherwise, return 0.
2086 swapon_check_swzone(unsigned long npages)
2088 unsigned long maxpages;
2090 /* absolute maximum we can handle assuming 100% efficiency */
2091 maxpages = uma_zone_get_max(swap_zone) * SWAP_META_PAGES;
2093 /* recommend using no more than half that amount */
2094 if (npages > maxpages / 2) {
2095 printf("warning: total configured swap (%lu pages) "
2096 "exceeds maximum recommended amount (%lu pages).\n",
2097 npages, maxpages / 2);
2098 printf("warning: increase kern.maxswzone "
2099 "or reduce amount of swap.\n");
2106 swaponsomething(struct vnode *vp, void *id, u_long nblks,
2107 sw_strategy_t *strategy, sw_close_t *close, dev_t dev, int flags)
2109 struct swdevt *sp, *tsp;
2114 * nblks is in DEV_BSIZE'd chunks, convert to PAGE_SIZE'd chunks.
2115 * First chop nblks off to page-align it, then convert.
2117 * sw->sw_nblks is in page-sized chunks now too.
2119 nblks &= ~(ctodb(1) - 1);
2120 nblks = dbtoc(nblks);
2123 * If we go beyond this, we get overflows in the radix
2126 mblocks = 0x40000000 / BLIST_META_RADIX;
2127 if (nblks > mblocks) {
2129 "WARNING: reducing swap size to maximum of %luMB per unit\n",
2130 mblocks / 1024 / 1024 * PAGE_SIZE);
2134 sp = malloc(sizeof *sp, M_VMPGDATA, M_WAITOK | M_ZERO);
2139 sp->sw_nblks = nblks;
2141 sp->sw_strategy = strategy;
2142 sp->sw_close = close;
2143 sp->sw_flags = flags;
2145 sp->sw_blist = blist_create(nblks, M_WAITOK);
2147 * Do not free the first two block in order to avoid overwriting
2148 * any bsd label at the front of the partition
2150 blist_free(sp->sw_blist, 2, nblks - 2);
2153 mtx_lock(&sw_dev_mtx);
2154 TAILQ_FOREACH(tsp, &swtailq, sw_list) {
2155 if (tsp->sw_end >= dvbase) {
2157 * We put one uncovered page between the devices
2158 * in order to definitively prevent any cross-device
2161 dvbase = tsp->sw_end + 1;
2164 sp->sw_first = dvbase;
2165 sp->sw_end = dvbase + nblks;
2166 TAILQ_INSERT_TAIL(&swtailq, sp, sw_list);
2168 swap_pager_avail += nblks - 2;
2169 swap_total += (vm_ooffset_t)nblks * PAGE_SIZE;
2170 swapon_check_swzone(swap_total / PAGE_SIZE);
2172 mtx_unlock(&sw_dev_mtx);
2173 EVENTHANDLER_INVOKE(swapon, sp);
2177 * SYSCALL: swapoff(devname)
2179 * Disable swapping on the given device.
2181 * XXX: Badly designed system call: it should use a device index
2182 * rather than filename as specification. We keep sw_vp around
2183 * only to make this work.
2185 #ifndef _SYS_SYSPROTO_H_
2186 struct swapoff_args {
2196 sys_swapoff(struct thread *td, struct swapoff_args *uap)
2199 struct nameidata nd;
2203 error = priv_check(td, PRIV_SWAPOFF);
2207 sx_xlock(&swdev_syscall_lock);
2209 NDINIT(&nd, LOOKUP, FOLLOW | AUDITVNODE1, UIO_USERSPACE, uap->name,
2214 NDFREE(&nd, NDF_ONLY_PNBUF);
2217 mtx_lock(&sw_dev_mtx);
2218 TAILQ_FOREACH(sp, &swtailq, sw_list) {
2219 if (sp->sw_vp == vp)
2222 mtx_unlock(&sw_dev_mtx);
2227 error = swapoff_one(sp, td->td_ucred);
2229 sx_xunlock(&swdev_syscall_lock);
2234 swapoff_one(struct swdevt *sp, struct ucred *cred)
2241 sx_assert(&swdev_syscall_lock, SA_XLOCKED);
2243 (void) vn_lock(sp->sw_vp, LK_EXCLUSIVE | LK_RETRY);
2244 error = mac_system_check_swapoff(cred, sp->sw_vp);
2245 (void) VOP_UNLOCK(sp->sw_vp, 0);
2249 nblks = sp->sw_nblks;
2252 * We can turn off this swap device safely only if the
2253 * available virtual memory in the system will fit the amount
2254 * of data we will have to page back in, plus an epsilon so
2255 * the system doesn't become critically low on swap space.
2257 if (vm_cnt.v_free_count + swap_pager_avail < nblks + nswap_lowat)
2261 * Prevent further allocations on this device.
2263 mtx_lock(&sw_dev_mtx);
2264 sp->sw_flags |= SW_CLOSING;
2265 swap_pager_avail -= blist_fill(sp->sw_blist, 0, nblks);
2266 swap_total -= (vm_ooffset_t)nblks * PAGE_SIZE;
2267 mtx_unlock(&sw_dev_mtx);
2270 * Page in the contents of the device and close it.
2272 swap_pager_swapoff(sp);
2274 sp->sw_close(curthread, sp);
2275 mtx_lock(&sw_dev_mtx);
2277 TAILQ_REMOVE(&swtailq, sp, sw_list);
2279 if (nswapdev == 0) {
2280 swap_pager_full = 2;
2281 swap_pager_almost_full = 1;
2285 mtx_unlock(&sw_dev_mtx);
2286 blist_destroy(sp->sw_blist);
2287 free(sp, M_VMPGDATA);
2294 struct swdevt *sp, *spt;
2295 const char *devname;
2298 sx_xlock(&swdev_syscall_lock);
2300 mtx_lock(&sw_dev_mtx);
2301 TAILQ_FOREACH_SAFE(sp, &swtailq, sw_list, spt) {
2302 mtx_unlock(&sw_dev_mtx);
2303 if (vn_isdisk(sp->sw_vp, NULL))
2304 devname = devtoname(sp->sw_vp->v_rdev);
2307 error = swapoff_one(sp, thread0.td_ucred);
2309 printf("Cannot remove swap device %s (error=%d), "
2310 "skipping.\n", devname, error);
2311 } else if (bootverbose) {
2312 printf("Swap device %s removed.\n", devname);
2314 mtx_lock(&sw_dev_mtx);
2316 mtx_unlock(&sw_dev_mtx);
2318 sx_xunlock(&swdev_syscall_lock);
2322 swap_pager_status(int *total, int *used)
2328 mtx_lock(&sw_dev_mtx);
2329 TAILQ_FOREACH(sp, &swtailq, sw_list) {
2330 *total += sp->sw_nblks;
2331 *used += sp->sw_used;
2333 mtx_unlock(&sw_dev_mtx);
2337 swap_dev_info(int name, struct xswdev *xs, char *devname, size_t len)
2340 const char *tmp_devname;
2345 mtx_lock(&sw_dev_mtx);
2346 TAILQ_FOREACH(sp, &swtailq, sw_list) {
2351 xs->xsw_version = XSWDEV_VERSION;
2352 xs->xsw_dev = sp->sw_dev;
2353 xs->xsw_flags = sp->sw_flags;
2354 xs->xsw_nblks = sp->sw_nblks;
2355 xs->xsw_used = sp->sw_used;
2356 if (devname != NULL) {
2357 if (vn_isdisk(sp->sw_vp, NULL))
2358 tmp_devname = devtoname(sp->sw_vp->v_rdev);
2360 tmp_devname = "[file]";
2361 strncpy(devname, tmp_devname, len);
2366 mtx_unlock(&sw_dev_mtx);
2370 #if defined(COMPAT_FREEBSD11)
2371 #define XSWDEV_VERSION_11 1
2382 sysctl_vm_swap_info(SYSCTL_HANDLER_ARGS)
2385 #if defined(COMPAT_FREEBSD11)
2386 struct xswdev11 xs11;
2390 if (arg2 != 1) /* name length */
2392 error = swap_dev_info(*(int *)arg1, &xs, NULL, 0);
2395 #if defined(COMPAT_FREEBSD11)
2396 if (req->oldlen == sizeof(xs11)) {
2397 xs11.xsw_version = XSWDEV_VERSION_11;
2398 xs11.xsw_dev = xs.xsw_dev; /* truncation */
2399 xs11.xsw_flags = xs.xsw_flags;
2400 xs11.xsw_nblks = xs.xsw_nblks;
2401 xs11.xsw_used = xs.xsw_used;
2402 error = SYSCTL_OUT(req, &xs11, sizeof(xs11));
2405 error = SYSCTL_OUT(req, &xs, sizeof(xs));
2409 SYSCTL_INT(_vm, OID_AUTO, nswapdev, CTLFLAG_RD, &nswapdev, 0,
2410 "Number of swap devices");
2411 SYSCTL_NODE(_vm, OID_AUTO, swap_info, CTLFLAG_RD | CTLFLAG_MPSAFE,
2412 sysctl_vm_swap_info,
2413 "Swap statistics by device");
2416 * vmspace_swap_count() - count the approximate swap usage in pages for a
2419 * The map must be locked.
2421 * Swap usage is determined by taking the proportional swap used by
2422 * VM objects backing the VM map. To make up for fractional losses,
2423 * if the VM object has any swap use at all the associated map entries
2424 * count for at least 1 swap page.
2427 vmspace_swap_count(struct vmspace *vmspace)
2434 map = &vmspace->vm_map;
2437 for (cur = map->header.next; cur != &map->header; cur = cur->next) {
2438 if ((cur->eflags & MAP_ENTRY_IS_SUB_MAP) == 0 &&
2439 (object = cur->object.vm_object) != NULL) {
2440 VM_OBJECT_WLOCK(object);
2441 if (object->type == OBJT_SWAP &&
2442 object->un_pager.swp.swp_bcount != 0) {
2443 n = (cur->end - cur->start) / PAGE_SIZE;
2444 count += object->un_pager.swp.swp_bcount *
2445 SWAP_META_PAGES * n / object->size + 1;
2447 VM_OBJECT_WUNLOCK(object);
2456 * Swapping onto disk devices.
2460 static g_orphan_t swapgeom_orphan;
2462 static struct g_class g_swap_class = {
2464 .version = G_VERSION,
2465 .orphan = swapgeom_orphan,
2468 DECLARE_GEOM_CLASS(g_swap_class, g_class);
2472 swapgeom_close_ev(void *arg, int flags)
2474 struct g_consumer *cp;
2477 g_access(cp, -1, -1, 0);
2479 g_destroy_consumer(cp);
2483 * Add a reference to the g_consumer for an inflight transaction.
2486 swapgeom_acquire(struct g_consumer *cp)
2489 mtx_assert(&sw_dev_mtx, MA_OWNED);
2494 * Remove a reference from the g_consumer. Post a close event if all
2495 * references go away, since the function might be called from the
2499 swapgeom_release(struct g_consumer *cp, struct swdevt *sp)
2502 mtx_assert(&sw_dev_mtx, MA_OWNED);
2504 if (cp->index == 0) {
2505 if (g_post_event(swapgeom_close_ev, cp, M_NOWAIT, NULL) == 0)
2511 swapgeom_done(struct bio *bp2)
2515 struct g_consumer *cp;
2517 bp = bp2->bio_caller2;
2519 bp->b_ioflags = bp2->bio_flags;
2521 bp->b_ioflags |= BIO_ERROR;
2522 bp->b_resid = bp->b_bcount - bp2->bio_completed;
2523 bp->b_error = bp2->bio_error;
2525 sp = bp2->bio_caller1;
2526 mtx_lock(&sw_dev_mtx);
2527 swapgeom_release(cp, sp);
2528 mtx_unlock(&sw_dev_mtx);
2533 swapgeom_strategy(struct buf *bp, struct swdevt *sp)
2536 struct g_consumer *cp;
2538 mtx_lock(&sw_dev_mtx);
2541 mtx_unlock(&sw_dev_mtx);
2542 bp->b_error = ENXIO;
2543 bp->b_ioflags |= BIO_ERROR;
2547 swapgeom_acquire(cp);
2548 mtx_unlock(&sw_dev_mtx);
2549 if (bp->b_iocmd == BIO_WRITE)
2552 bio = g_alloc_bio();
2554 mtx_lock(&sw_dev_mtx);
2555 swapgeom_release(cp, sp);
2556 mtx_unlock(&sw_dev_mtx);
2557 bp->b_error = ENOMEM;
2558 bp->b_ioflags |= BIO_ERROR;
2563 bio->bio_caller1 = sp;
2564 bio->bio_caller2 = bp;
2565 bio->bio_cmd = bp->b_iocmd;
2566 bio->bio_offset = (bp->b_blkno - sp->sw_first) * PAGE_SIZE;
2567 bio->bio_length = bp->b_bcount;
2568 bio->bio_done = swapgeom_done;
2569 if (!buf_mapped(bp)) {
2570 bio->bio_ma = bp->b_pages;
2571 bio->bio_data = unmapped_buf;
2572 bio->bio_ma_offset = (vm_offset_t)bp->b_offset & PAGE_MASK;
2573 bio->bio_ma_n = bp->b_npages;
2574 bio->bio_flags |= BIO_UNMAPPED;
2576 bio->bio_data = bp->b_data;
2579 g_io_request(bio, cp);
2584 swapgeom_orphan(struct g_consumer *cp)
2589 mtx_lock(&sw_dev_mtx);
2590 TAILQ_FOREACH(sp, &swtailq, sw_list) {
2591 if (sp->sw_id == cp) {
2592 sp->sw_flags |= SW_CLOSING;
2597 * Drop reference we were created with. Do directly since we're in a
2598 * special context where we don't have to queue the call to
2599 * swapgeom_close_ev().
2602 destroy = ((sp != NULL) && (cp->index == 0));
2605 mtx_unlock(&sw_dev_mtx);
2607 swapgeom_close_ev(cp, 0);
2611 swapgeom_close(struct thread *td, struct swdevt *sw)
2613 struct g_consumer *cp;
2615 mtx_lock(&sw_dev_mtx);
2618 mtx_unlock(&sw_dev_mtx);
2621 * swapgeom_close() may be called from the biodone context,
2622 * where we cannot perform topology changes. Delegate the
2623 * work to the events thread.
2626 g_waitfor_event(swapgeom_close_ev, cp, M_WAITOK, NULL);
2630 swapongeom_locked(struct cdev *dev, struct vnode *vp)
2632 struct g_provider *pp;
2633 struct g_consumer *cp;
2634 static struct g_geom *gp;
2639 pp = g_dev_getprovider(dev);
2642 mtx_lock(&sw_dev_mtx);
2643 TAILQ_FOREACH(sp, &swtailq, sw_list) {
2645 if (cp != NULL && cp->provider == pp) {
2646 mtx_unlock(&sw_dev_mtx);
2650 mtx_unlock(&sw_dev_mtx);
2652 gp = g_new_geomf(&g_swap_class, "swap");
2653 cp = g_new_consumer(gp);
2654 cp->index = 1; /* Number of active I/Os, plus one for being active. */
2655 cp->flags |= G_CF_DIRECT_SEND | G_CF_DIRECT_RECEIVE;
2658 * XXX: Every time you think you can improve the margin for
2659 * footshooting, somebody depends on the ability to do so:
2660 * savecore(8) wants to write to our swapdev so we cannot
2661 * set an exclusive count :-(
2663 error = g_access(cp, 1, 1, 0);
2666 g_destroy_consumer(cp);
2669 nblks = pp->mediasize / DEV_BSIZE;
2670 swaponsomething(vp, cp, nblks, swapgeom_strategy,
2671 swapgeom_close, dev2udev(dev),
2672 (pp->flags & G_PF_ACCEPT_UNMAPPED) != 0 ? SW_UNMAPPED : 0);
2677 swapongeom(struct vnode *vp)
2681 vn_lock(vp, LK_EXCLUSIVE | LK_RETRY);
2682 if (vp->v_type != VCHR || (vp->v_iflag & VI_DOOMED) != 0) {
2686 error = swapongeom_locked(vp->v_rdev, vp);
2687 g_topology_unlock();
2696 * This is used mainly for network filesystem (read: probably only tested
2697 * with NFS) swapfiles.
2702 swapdev_strategy(struct buf *bp, struct swdevt *sp)
2706 bp->b_blkno = ctodb(bp->b_blkno - sp->sw_first);
2710 if (bp->b_iocmd == BIO_WRITE) {
2712 bufobj_wdrop(bp->b_bufobj);
2713 bufobj_wref(&vp2->v_bufobj);
2715 if (bp->b_bufobj != &vp2->v_bufobj)
2716 bp->b_bufobj = &vp2->v_bufobj;
2718 bp->b_iooffset = dbtob(bp->b_blkno);
2724 swapdev_close(struct thread *td, struct swdevt *sp)
2727 VOP_CLOSE(sp->sw_vp, FREAD | FWRITE, td->td_ucred, td);
2733 swaponvp(struct thread *td, struct vnode *vp, u_long nblks)
2740 mtx_lock(&sw_dev_mtx);
2741 TAILQ_FOREACH(sp, &swtailq, sw_list) {
2742 if (sp->sw_id == vp) {
2743 mtx_unlock(&sw_dev_mtx);
2747 mtx_unlock(&sw_dev_mtx);
2749 (void) vn_lock(vp, LK_EXCLUSIVE | LK_RETRY);
2751 error = mac_system_check_swapon(td->td_ucred, vp);
2754 error = VOP_OPEN(vp, FREAD | FWRITE, td->td_ucred, td, NULL);
2755 (void) VOP_UNLOCK(vp, 0);
2759 swaponsomething(vp, vp, nblks, swapdev_strategy, swapdev_close,
2765 sysctl_swap_async_max(SYSCTL_HANDLER_ARGS)
2769 new = nsw_wcount_async_max;
2770 error = sysctl_handle_int(oidp, &new, 0, req);
2771 if (error != 0 || req->newptr == NULL)
2774 if (new > nswbuf / 2 || new < 1)
2777 mtx_lock(&pbuf_mtx);
2778 while (nsw_wcount_async_max != new) {
2780 * Adjust difference. If the current async count is too low,
2781 * we will need to sqeeze our update slowly in. Sleep with a
2782 * higher priority than getpbuf() to finish faster.
2784 n = new - nsw_wcount_async_max;
2785 if (nsw_wcount_async + n >= 0) {
2786 nsw_wcount_async += n;
2787 nsw_wcount_async_max += n;
2788 wakeup(&nsw_wcount_async);
2790 nsw_wcount_async_max -= nsw_wcount_async;
2791 nsw_wcount_async = 0;
2792 msleep(&nsw_wcount_async, &pbuf_mtx, PSWP,
2796 mtx_unlock(&pbuf_mtx);