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/pctrie.h>
91 #include <sys/racct.h>
92 #include <sys/resource.h>
93 #include <sys/resourcevar.h>
94 #include <sys/rwlock.h>
96 #include <sys/sysctl.h>
97 #include <sys/sysproto.h>
98 #include <sys/blist.h>
101 #include <sys/vmmeter.h>
103 #include <security/mac/mac_framework.h>
107 #include <vm/vm_map.h>
108 #include <vm/vm_kern.h>
109 #include <vm/vm_object.h>
110 #include <vm/vm_page.h>
111 #include <vm/vm_pager.h>
112 #include <vm/vm_pageout.h>
113 #include <vm/vm_param.h>
114 #include <vm/swap_pager.h>
115 #include <vm/vm_extern.h>
118 #include <geom/geom.h>
121 * MAX_PAGEOUT_CLUSTER must be a power of 2 between 1 and 64.
122 * The 64-page limit is due to the radix code (kern/subr_blist.c).
124 #ifndef MAX_PAGEOUT_CLUSTER
125 #define MAX_PAGEOUT_CLUSTER 32
128 #if !defined(SWB_NPAGES)
129 #define SWB_NPAGES MAX_PAGEOUT_CLUSTER
132 #define SWAP_META_PAGES PCTRIE_COUNT
135 * A swblk structure maps each page index within a
136 * SWAP_META_PAGES-aligned and sized range to the address of an
137 * on-disk swap block (or SWAPBLK_NONE). The collection of these
138 * mappings for an entire vm object is implemented as a pc-trie.
142 daddr_t d[SWAP_META_PAGES];
145 static MALLOC_DEFINE(M_VMPGDATA, "vm_pgdata", "swap pager private data");
146 static struct mtx sw_dev_mtx;
147 static TAILQ_HEAD(, swdevt) swtailq = TAILQ_HEAD_INITIALIZER(swtailq);
148 static struct swdevt *swdevhd; /* Allocate from here next */
149 static int nswapdev; /* Number of swap devices */
150 int swap_pager_avail;
151 static struct sx swdev_syscall_lock; /* serialize swap(on|off) */
153 static vm_ooffset_t swap_total;
154 SYSCTL_QUAD(_vm, OID_AUTO, swap_total, CTLFLAG_RD, &swap_total, 0,
155 "Total amount of available swap storage.");
156 static vm_ooffset_t swap_reserved;
157 SYSCTL_QUAD(_vm, OID_AUTO, swap_reserved, CTLFLAG_RD, &swap_reserved, 0,
158 "Amount of swap storage needed to back all allocated anonymous memory.");
159 static int overcommit = 0;
160 SYSCTL_INT(_vm, VM_OVERCOMMIT, overcommit, CTLFLAG_RW, &overcommit, 0,
161 "Configure virtual memory overcommit behavior. See tuning(7) "
163 static unsigned long swzone;
164 SYSCTL_ULONG(_vm, OID_AUTO, swzone, CTLFLAG_RD, &swzone, 0,
165 "Actual size of swap metadata zone");
166 static unsigned long swap_maxpages;
167 SYSCTL_ULONG(_vm, OID_AUTO, swap_maxpages, CTLFLAG_RD, &swap_maxpages, 0,
168 "Maximum amount of swap supported");
170 /* bits from overcommit */
171 #define SWAP_RESERVE_FORCE_ON (1 << 0)
172 #define SWAP_RESERVE_RLIMIT_ON (1 << 1)
173 #define SWAP_RESERVE_ALLOW_NONWIRED (1 << 2)
176 swap_reserve(vm_ooffset_t incr)
179 return (swap_reserve_by_cred(incr, curthread->td_ucred));
183 swap_reserve_by_cred(vm_ooffset_t incr, struct ucred *cred)
188 static struct timeval lastfail;
191 uip = cred->cr_ruidinfo;
193 if (incr & PAGE_MASK)
194 panic("swap_reserve: & PAGE_MASK");
199 error = racct_add(curproc, RACCT_SWAP, incr);
200 PROC_UNLOCK(curproc);
207 mtx_lock(&sw_dev_mtx);
208 r = swap_reserved + incr;
209 if (overcommit & SWAP_RESERVE_ALLOW_NONWIRED) {
210 s = vm_cnt.v_page_count - vm_cnt.v_free_reserved - vm_cnt.v_wire_count;
215 if ((overcommit & SWAP_RESERVE_FORCE_ON) == 0 || r <= s ||
216 (error = priv_check(curthread, PRIV_VM_SWAP_NOQUOTA)) == 0) {
220 mtx_unlock(&sw_dev_mtx);
223 UIDINFO_VMSIZE_LOCK(uip);
224 if ((overcommit & SWAP_RESERVE_RLIMIT_ON) != 0 &&
225 uip->ui_vmsize + incr > lim_cur(curthread, RLIMIT_SWAP) &&
226 priv_check(curthread, PRIV_VM_SWAP_NORLIMIT))
229 uip->ui_vmsize += incr;
230 UIDINFO_VMSIZE_UNLOCK(uip);
232 mtx_lock(&sw_dev_mtx);
233 swap_reserved -= incr;
234 mtx_unlock(&sw_dev_mtx);
237 if (!res && ppsratecheck(&lastfail, &curfail, 1)) {
238 printf("uid %d, pid %d: swap reservation for %jd bytes failed\n",
239 uip->ui_uid, curproc->p_pid, incr);
245 racct_sub(curproc, RACCT_SWAP, incr);
246 PROC_UNLOCK(curproc);
254 swap_reserve_force(vm_ooffset_t incr)
258 mtx_lock(&sw_dev_mtx);
259 swap_reserved += incr;
260 mtx_unlock(&sw_dev_mtx);
264 racct_add_force(curproc, RACCT_SWAP, incr);
265 PROC_UNLOCK(curproc);
268 uip = curthread->td_ucred->cr_ruidinfo;
270 UIDINFO_VMSIZE_LOCK(uip);
271 uip->ui_vmsize += incr;
272 UIDINFO_VMSIZE_UNLOCK(uip);
273 PROC_UNLOCK(curproc);
277 swap_release(vm_ooffset_t decr)
282 cred = curthread->td_ucred;
283 swap_release_by_cred(decr, cred);
284 PROC_UNLOCK(curproc);
288 swap_release_by_cred(vm_ooffset_t decr, struct ucred *cred)
292 uip = cred->cr_ruidinfo;
294 if (decr & PAGE_MASK)
295 panic("swap_release: & PAGE_MASK");
297 mtx_lock(&sw_dev_mtx);
298 if (swap_reserved < decr)
299 panic("swap_reserved < decr");
300 swap_reserved -= decr;
301 mtx_unlock(&sw_dev_mtx);
303 UIDINFO_VMSIZE_LOCK(uip);
304 if (uip->ui_vmsize < decr)
305 printf("negative vmsize for uid = %d\n", uip->ui_uid);
306 uip->ui_vmsize -= decr;
307 UIDINFO_VMSIZE_UNLOCK(uip);
309 racct_sub_cred(cred, RACCT_SWAP, decr);
312 #define SWM_FREE 0x02 /* free, period */
313 #define SWM_POP 0x04 /* pop out */
315 static int swap_pager_full = 2; /* swap space exhaustion (task killing) */
316 static int swap_pager_almost_full = 1; /* swap space exhaustion (w/hysteresis)*/
317 static int nsw_rcount; /* free read buffers */
318 static int nsw_wcount_sync; /* limit write buffers / synchronous */
319 static int nsw_wcount_async; /* limit write buffers / asynchronous */
320 static int nsw_wcount_async_max;/* assigned maximum */
321 static int nsw_cluster_max; /* maximum VOP I/O allowed */
323 static int sysctl_swap_async_max(SYSCTL_HANDLER_ARGS);
324 SYSCTL_PROC(_vm, OID_AUTO, swap_async_max, CTLTYPE_INT | CTLFLAG_RW |
325 CTLFLAG_MPSAFE, NULL, 0, sysctl_swap_async_max, "I",
326 "Maximum running async swap ops");
327 static int sysctl_swap_fragmentation(SYSCTL_HANDLER_ARGS);
328 SYSCTL_PROC(_vm, OID_AUTO, swap_fragmentation, CTLTYPE_STRING | CTLFLAG_RD |
329 CTLFLAG_MPSAFE, NULL, 0, sysctl_swap_fragmentation, "A",
330 "Swap Fragmentation Info");
332 static struct sx sw_alloc_sx;
335 * "named" and "unnamed" anon region objects. Try to reduce the overhead
336 * of searching a named list by hashing it just a little.
341 #define NOBJLIST(handle) \
342 (&swap_pager_object_list[((int)(intptr_t)handle >> 4) & (NOBJLISTS-1)])
344 static struct pagerlst swap_pager_object_list[NOBJLISTS];
345 static uma_zone_t swblk_zone;
346 static uma_zone_t swpctrie_zone;
349 * pagerops for OBJT_SWAP - "swap pager". Some ops are also global procedure
350 * calls hooked from other parts of the VM system and do not appear here.
351 * (see vm/swap_pager.h).
354 swap_pager_alloc(void *handle, vm_ooffset_t size,
355 vm_prot_t prot, vm_ooffset_t offset, struct ucred *);
356 static void swap_pager_dealloc(vm_object_t object);
357 static int swap_pager_getpages(vm_object_t, vm_page_t *, int, int *,
359 static int swap_pager_getpages_async(vm_object_t, vm_page_t *, int, int *,
360 int *, pgo_getpages_iodone_t, void *);
361 static void swap_pager_putpages(vm_object_t, vm_page_t *, int, boolean_t, int *);
363 swap_pager_haspage(vm_object_t object, vm_pindex_t pindex, int *before, int *after);
364 static void swap_pager_init(void);
365 static void swap_pager_unswapped(vm_page_t);
366 static void swap_pager_swapoff(struct swdevt *sp);
368 struct pagerops swappagerops = {
369 .pgo_init = swap_pager_init, /* early system initialization of pager */
370 .pgo_alloc = swap_pager_alloc, /* allocate an OBJT_SWAP object */
371 .pgo_dealloc = swap_pager_dealloc, /* deallocate an OBJT_SWAP object */
372 .pgo_getpages = swap_pager_getpages, /* pagein */
373 .pgo_getpages_async = swap_pager_getpages_async, /* pagein (async) */
374 .pgo_putpages = swap_pager_putpages, /* pageout */
375 .pgo_haspage = swap_pager_haspage, /* get backing store status for page */
376 .pgo_pageunswapped = swap_pager_unswapped, /* remove swap related to page */
380 * swap_*() routines are externally accessible. swp_*() routines are
383 static int nswap_lowat = 128; /* in pages, swap_pager_almost_full warn */
384 static int nswap_hiwat = 512; /* in pages, swap_pager_almost_full warn */
386 SYSCTL_INT(_vm, OID_AUTO, dmmax, CTLFLAG_RD, &nsw_cluster_max, 0,
387 "Maximum size of a swap block in pages");
389 static void swp_sizecheck(void);
390 static void swp_pager_async_iodone(struct buf *bp);
391 static int swapongeom(struct vnode *);
392 static int swaponvp(struct thread *, struct vnode *, u_long);
393 static int swapoff_one(struct swdevt *sp, struct ucred *cred);
396 * Swap bitmap functions
398 static void swp_pager_freeswapspace(daddr_t blk, int npages);
399 static daddr_t swp_pager_getswapspace(int npages);
404 static void swp_pager_meta_build(vm_object_t, vm_pindex_t, daddr_t);
405 static void swp_pager_meta_free(vm_object_t, vm_pindex_t, vm_pindex_t);
406 static void swp_pager_meta_free_all(vm_object_t);
407 static daddr_t swp_pager_meta_ctl(vm_object_t, vm_pindex_t, int);
410 swblk_trie_alloc(struct pctrie *ptree)
413 return (uma_zalloc(swpctrie_zone, M_NOWAIT | (curproc == pageproc ?
414 M_USE_RESERVE : 0)));
418 swblk_trie_free(struct pctrie *ptree, void *node)
421 uma_zfree(swpctrie_zone, node);
424 PCTRIE_DEFINE(SWAP, swblk, p, swblk_trie_alloc, swblk_trie_free);
427 * SWP_SIZECHECK() - update swap_pager_full indication
429 * update the swap_pager_almost_full indication and warn when we are
430 * about to run out of swap space, using lowat/hiwat hysteresis.
432 * Clear swap_pager_full ( task killing ) indication when lowat is met.
434 * No restrictions on call
435 * This routine may not block.
441 if (swap_pager_avail < nswap_lowat) {
442 if (swap_pager_almost_full == 0) {
443 printf("swap_pager: out of swap space\n");
444 swap_pager_almost_full = 1;
448 if (swap_pager_avail > nswap_hiwat)
449 swap_pager_almost_full = 0;
454 * SWAP_PAGER_INIT() - initialize the swap pager!
456 * Expected to be started from system init. NOTE: This code is run
457 * before much else so be careful what you depend on. Most of the VM
458 * system has yet to be initialized at this point.
461 swap_pager_init(void)
464 * Initialize object lists
468 for (i = 0; i < NOBJLISTS; ++i)
469 TAILQ_INIT(&swap_pager_object_list[i]);
470 mtx_init(&sw_dev_mtx, "swapdev", NULL, MTX_DEF);
471 sx_init(&sw_alloc_sx, "swspsx");
472 sx_init(&swdev_syscall_lock, "swsysc");
476 * SWAP_PAGER_SWAP_INIT() - swap pager initialization from pageout process
478 * Expected to be started from pageout process once, prior to entering
482 swap_pager_swap_init(void)
487 * Number of in-transit swap bp operations. Don't
488 * exhaust the pbufs completely. Make sure we
489 * initialize workable values (0 will work for hysteresis
490 * but it isn't very efficient).
492 * The nsw_cluster_max is constrained by the bp->b_pages[]
493 * array (MAXPHYS/PAGE_SIZE) and our locally defined
494 * MAX_PAGEOUT_CLUSTER. Also be aware that swap ops are
495 * constrained by the swap device interleave stripe size.
497 * Currently we hardwire nsw_wcount_async to 4. This limit is
498 * designed to prevent other I/O from having high latencies due to
499 * our pageout I/O. The value 4 works well for one or two active swap
500 * devices but is probably a little low if you have more. Even so,
501 * a higher value would probably generate only a limited improvement
502 * with three or four active swap devices since the system does not
503 * typically have to pageout at extreme bandwidths. We will want
504 * at least 2 per swap devices, and 4 is a pretty good value if you
505 * have one NFS swap device due to the command/ack latency over NFS.
506 * So it all works out pretty well.
508 nsw_cluster_max = min((MAXPHYS/PAGE_SIZE), MAX_PAGEOUT_CLUSTER);
511 nsw_rcount = (nswbuf + 1) / 2;
512 nsw_wcount_sync = (nswbuf + 3) / 4;
513 nsw_wcount_async = 4;
514 nsw_wcount_async_max = nsw_wcount_async;
515 mtx_unlock(&pbuf_mtx);
518 * Initialize our zone, guessing on the number we need based
519 * on the number of pages in the system.
521 n = vm_cnt.v_page_count / 2;
522 if (maxswzone && n > maxswzone / sizeof(struct swblk))
523 n = maxswzone / sizeof(struct swblk);
524 swpctrie_zone = uma_zcreate("swpctrie", pctrie_node_size(), NULL, NULL,
525 pctrie_zone_init, NULL, UMA_ALIGN_PTR,
526 UMA_ZONE_NOFREE | UMA_ZONE_VM);
527 if (swpctrie_zone == NULL)
528 panic("failed to create swap pctrie zone.");
529 swblk_zone = uma_zcreate("swblk", sizeof(struct swblk), NULL, NULL,
530 NULL, NULL, _Alignof(struct swblk) - 1,
531 UMA_ZONE_NOFREE | UMA_ZONE_VM);
532 if (swblk_zone == NULL)
533 panic("failed to create swap blk zone.");
536 if (uma_zone_reserve_kva(swblk_zone, n))
539 * if the allocation failed, try a zone two thirds the
540 * size of the previous attempt.
546 * Often uma_zone_reserve_kva() cannot reserve exactly the
547 * requested size. Account for the difference when
548 * calculating swap_maxpages.
550 n = uma_zone_get_max(swblk_zone);
553 printf("Swap blk zone entries reduced from %lu to %lu.\n",
555 swap_maxpages = n * SWAP_META_PAGES;
556 swzone = n * sizeof(struct swblk);
557 if (!uma_zone_reserve_kva(swpctrie_zone, n))
558 printf("Cannot reserve swap pctrie zone, "
559 "reduce kern.maxswzone.\n");
563 swap_pager_alloc_init(void *handle, struct ucred *cred, vm_ooffset_t size,
569 if (!swap_reserve_by_cred(size, cred))
575 * The un_pager.swp.swp_blks trie is initialized by
576 * vm_object_allocate() to ensure the correct order of
577 * visibility to other threads.
579 object = vm_object_allocate(OBJT_SWAP, OFF_TO_IDX(offset +
582 object->handle = handle;
585 object->charge = size;
591 * SWAP_PAGER_ALLOC() - allocate a new OBJT_SWAP VM object and instantiate
592 * its metadata structures.
594 * This routine is called from the mmap and fork code to create a new
597 * This routine must ensure that no live duplicate is created for
598 * the named object request, which is protected against by
599 * holding the sw_alloc_sx lock in case handle != NULL.
602 swap_pager_alloc(void *handle, vm_ooffset_t size, vm_prot_t prot,
603 vm_ooffset_t offset, struct ucred *cred)
607 if (handle != NULL) {
609 * Reference existing named region or allocate new one. There
610 * should not be a race here against swp_pager_meta_build()
611 * as called from vm_page_remove() in regards to the lookup
614 sx_xlock(&sw_alloc_sx);
615 object = vm_pager_object_lookup(NOBJLIST(handle), handle);
616 if (object == NULL) {
617 object = swap_pager_alloc_init(handle, cred, size,
619 if (object != NULL) {
620 TAILQ_INSERT_TAIL(NOBJLIST(object->handle),
621 object, pager_object_list);
624 sx_xunlock(&sw_alloc_sx);
626 object = swap_pager_alloc_init(handle, cred, size, offset);
632 * SWAP_PAGER_DEALLOC() - remove swap metadata from object
634 * The swap backing for the object is destroyed. The code is
635 * designed such that we can reinstantiate it later, but this
636 * routine is typically called only when the entire object is
637 * about to be destroyed.
639 * The object must be locked.
642 swap_pager_dealloc(vm_object_t object)
645 VM_OBJECT_ASSERT_WLOCKED(object);
646 KASSERT((object->flags & OBJ_DEAD) != 0, ("dealloc of reachable obj"));
649 * Remove from list right away so lookups will fail if we block for
650 * pageout completion.
652 if (object->handle != NULL) {
653 VM_OBJECT_WUNLOCK(object);
654 sx_xlock(&sw_alloc_sx);
655 TAILQ_REMOVE(NOBJLIST(object->handle), object,
657 sx_xunlock(&sw_alloc_sx);
658 VM_OBJECT_WLOCK(object);
661 vm_object_pip_wait(object, "swpdea");
664 * Free all remaining metadata. We only bother to free it from
665 * the swap meta data. We do not attempt to free swapblk's still
666 * associated with vm_page_t's for this object. We do not care
667 * if paging is still in progress on some objects.
669 swp_pager_meta_free_all(object);
670 object->handle = NULL;
671 object->type = OBJT_DEAD;
674 /************************************************************************
675 * SWAP PAGER BITMAP ROUTINES *
676 ************************************************************************/
679 * SWP_PAGER_GETSWAPSPACE() - allocate raw swap space
681 * Allocate swap for the requested number of pages. The starting
682 * swap block number (a page index) is returned or SWAPBLK_NONE
683 * if the allocation failed.
685 * Also has the side effect of advising that somebody made a mistake
686 * when they configured swap and didn't configure enough.
688 * This routine may not sleep.
690 * We allocate in round-robin fashion from the configured devices.
693 swp_pager_getswapspace(int npages)
700 mtx_lock(&sw_dev_mtx);
702 for (i = 0; i < nswapdev; i++) {
704 sp = TAILQ_FIRST(&swtailq);
705 if (!(sp->sw_flags & SW_CLOSING)) {
706 blk = blist_alloc(sp->sw_blist, npages);
707 if (blk != SWAPBLK_NONE) {
709 sp->sw_used += npages;
710 swap_pager_avail -= npages;
712 swdevhd = TAILQ_NEXT(sp, sw_list);
716 sp = TAILQ_NEXT(sp, sw_list);
718 if (swap_pager_full != 2) {
719 printf("swap_pager_getswapspace(%d): failed\n", npages);
721 swap_pager_almost_full = 1;
725 mtx_unlock(&sw_dev_mtx);
730 swp_pager_isondev(daddr_t blk, struct swdevt *sp)
733 return (blk >= sp->sw_first && blk < sp->sw_end);
737 swp_pager_strategy(struct buf *bp)
741 mtx_lock(&sw_dev_mtx);
742 TAILQ_FOREACH(sp, &swtailq, sw_list) {
743 if (bp->b_blkno >= sp->sw_first && bp->b_blkno < sp->sw_end) {
744 mtx_unlock(&sw_dev_mtx);
745 if ((sp->sw_flags & SW_UNMAPPED) != 0 &&
746 unmapped_buf_allowed) {
747 bp->b_data = unmapped_buf;
750 pmap_qenter((vm_offset_t)bp->b_data,
751 &bp->b_pages[0], bp->b_bcount / PAGE_SIZE);
753 sp->sw_strategy(bp, sp);
757 panic("Swapdev not found");
762 * SWP_PAGER_FREESWAPSPACE() - free raw swap space
764 * This routine returns the specified swap blocks back to the bitmap.
766 * This routine may not sleep.
769 swp_pager_freeswapspace(daddr_t blk, int npages)
773 mtx_lock(&sw_dev_mtx);
774 TAILQ_FOREACH(sp, &swtailq, sw_list) {
775 if (blk >= sp->sw_first && blk < sp->sw_end) {
776 sp->sw_used -= npages;
778 * If we are attempting to stop swapping on
779 * this device, we don't want to mark any
780 * blocks free lest they be reused.
782 if ((sp->sw_flags & SW_CLOSING) == 0) {
783 blist_free(sp->sw_blist, blk - sp->sw_first,
785 swap_pager_avail += npages;
788 mtx_unlock(&sw_dev_mtx);
792 panic("Swapdev not found");
796 * SYSCTL_SWAP_FRAGMENTATION() - produce raw swap space stats
799 sysctl_swap_fragmentation(SYSCTL_HANDLER_ARGS)
806 error = sysctl_wire_old_buffer(req, 0);
809 sbuf_new_for_sysctl(&sbuf, NULL, 128, req);
810 mtx_lock(&sw_dev_mtx);
811 TAILQ_FOREACH(sp, &swtailq, sw_list) {
812 if (vn_isdisk(sp->sw_vp, NULL))
813 devname = devtoname(sp->sw_vp->v_rdev);
816 sbuf_printf(&sbuf, "\nFree space on device %s:\n", devname);
817 blist_stats(sp->sw_blist, &sbuf);
819 mtx_unlock(&sw_dev_mtx);
820 error = sbuf_finish(&sbuf);
826 * SWAP_PAGER_FREESPACE() - frees swap blocks associated with a page
827 * range within an object.
829 * This is a globally accessible routine.
831 * This routine removes swapblk assignments from swap metadata.
833 * The external callers of this routine typically have already destroyed
834 * or renamed vm_page_t's associated with this range in the object so
837 * The object must be locked.
840 swap_pager_freespace(vm_object_t object, vm_pindex_t start, vm_size_t size)
843 swp_pager_meta_free(object, start, size);
847 * SWAP_PAGER_RESERVE() - reserve swap blocks in object
849 * Assigns swap blocks to the specified range within the object. The
850 * swap blocks are not zeroed. Any previous swap assignment is destroyed.
852 * Returns 0 on success, -1 on failure.
855 swap_pager_reserve(vm_object_t object, vm_pindex_t start, vm_size_t size)
858 daddr_t blk = SWAPBLK_NONE;
859 vm_pindex_t beg = start; /* save start index */
861 VM_OBJECT_WLOCK(object);
865 while ((blk = swp_pager_getswapspace(n)) == SWAPBLK_NONE) {
868 swp_pager_meta_free(object, beg, start - beg);
869 VM_OBJECT_WUNLOCK(object);
874 swp_pager_meta_build(object, start, blk);
880 swp_pager_meta_free(object, start, n);
881 VM_OBJECT_WUNLOCK(object);
886 * SWAP_PAGER_COPY() - copy blocks from source pager to destination pager
887 * and destroy the source.
889 * Copy any valid swapblks from the source to the destination. In
890 * cases where both the source and destination have a valid swapblk,
891 * we keep the destination's.
893 * This routine is allowed to sleep. It may sleep allocating metadata
894 * indirectly through swp_pager_meta_build() or if paging is still in
895 * progress on the source.
897 * The source object contains no vm_page_t's (which is just as well)
899 * The source object is of type OBJT_SWAP.
901 * The source and destination objects must be locked.
902 * Both object locks may temporarily be released.
905 swap_pager_copy(vm_object_t srcobject, vm_object_t dstobject,
906 vm_pindex_t offset, int destroysource)
910 VM_OBJECT_ASSERT_WLOCKED(srcobject);
911 VM_OBJECT_ASSERT_WLOCKED(dstobject);
914 * If destroysource is set, we remove the source object from the
915 * swap_pager internal queue now.
917 if (destroysource && srcobject->handle != NULL) {
918 vm_object_pip_add(srcobject, 1);
919 VM_OBJECT_WUNLOCK(srcobject);
920 vm_object_pip_add(dstobject, 1);
921 VM_OBJECT_WUNLOCK(dstobject);
922 sx_xlock(&sw_alloc_sx);
923 TAILQ_REMOVE(NOBJLIST(srcobject->handle), srcobject,
925 sx_xunlock(&sw_alloc_sx);
926 VM_OBJECT_WLOCK(dstobject);
927 vm_object_pip_wakeup(dstobject);
928 VM_OBJECT_WLOCK(srcobject);
929 vm_object_pip_wakeup(srcobject);
933 * transfer source to destination.
935 for (i = 0; i < dstobject->size; ++i) {
939 * Locate (without changing) the swapblk on the destination,
940 * unless it is invalid in which case free it silently, or
941 * if the destination is a resident page, in which case the
942 * source is thrown away.
944 dstaddr = swp_pager_meta_ctl(dstobject, i, 0);
946 if (dstaddr == SWAPBLK_NONE) {
948 * Destination has no swapblk and is not resident,
953 srcaddr = swp_pager_meta_ctl(
959 if (srcaddr != SWAPBLK_NONE) {
961 * swp_pager_meta_build() can sleep.
963 vm_object_pip_add(srcobject, 1);
964 VM_OBJECT_WUNLOCK(srcobject);
965 vm_object_pip_add(dstobject, 1);
966 swp_pager_meta_build(dstobject, i, srcaddr);
967 vm_object_pip_wakeup(dstobject);
968 VM_OBJECT_WLOCK(srcobject);
969 vm_object_pip_wakeup(srcobject);
973 * Destination has valid swapblk or it is represented
974 * by a resident page. We destroy the sourceblock.
977 swp_pager_meta_ctl(srcobject, i + offset, SWM_FREE);
982 * Free left over swap blocks in source.
984 * We have to revert the type to OBJT_DEFAULT so we do not accidentally
985 * double-remove the object from the swap queues.
988 swp_pager_meta_free_all(srcobject);
990 * Reverting the type is not necessary, the caller is going
991 * to destroy srcobject directly, but I'm doing it here
992 * for consistency since we've removed the object from its
995 srcobject->type = OBJT_DEFAULT;
1000 * SWAP_PAGER_HASPAGE() - determine if we have good backing store for
1001 * the requested page.
1003 * We determine whether good backing store exists for the requested
1004 * page and return TRUE if it does, FALSE if it doesn't.
1006 * If TRUE, we also try to determine how much valid, contiguous backing
1007 * store exists before and after the requested page.
1010 swap_pager_haspage(vm_object_t object, vm_pindex_t pindex, int *before,
1016 VM_OBJECT_ASSERT_LOCKED(object);
1019 * do we have good backing store at the requested index ?
1021 blk0 = swp_pager_meta_ctl(object, pindex, 0);
1022 if (blk0 == SWAPBLK_NONE) {
1031 * find backwards-looking contiguous good backing store
1033 if (before != NULL) {
1034 for (i = 1; i < SWB_NPAGES; i++) {
1037 blk = swp_pager_meta_ctl(object, pindex - i, 0);
1038 if (blk != blk0 - i)
1045 * find forward-looking contiguous good backing store
1047 if (after != NULL) {
1048 for (i = 1; i < SWB_NPAGES; i++) {
1049 blk = swp_pager_meta_ctl(object, pindex + i, 0);
1050 if (blk != blk0 + i)
1059 * SWAP_PAGER_PAGE_UNSWAPPED() - remove swap backing store related to page
1061 * This removes any associated swap backing store, whether valid or
1062 * not, from the page.
1064 * This routine is typically called when a page is made dirty, at
1065 * which point any associated swap can be freed. MADV_FREE also
1066 * calls us in a special-case situation
1068 * NOTE!!! If the page is clean and the swap was valid, the caller
1069 * should make the page dirty before calling this routine. This routine
1070 * does NOT change the m->dirty status of the page. Also: MADV_FREE
1073 * This routine may not sleep.
1075 * The object containing the page must be locked.
1078 swap_pager_unswapped(vm_page_t m)
1081 swp_pager_meta_ctl(m->object, m->pindex, SWM_FREE);
1085 * swap_pager_getpages() - bring pages in from swap
1087 * Attempt to page in the pages in array "m" of length "count". The caller
1088 * may optionally specify that additional pages preceding and succeeding
1089 * the specified range be paged in. The number of such pages is returned
1090 * in the "rbehind" and "rahead" parameters, and they will be in the
1091 * inactive queue upon return.
1093 * The pages in "m" must be busied and will remain busied upon return.
1096 swap_pager_getpages(vm_object_t object, vm_page_t *m, int count, int *rbehind,
1100 vm_page_t mpred, msucc, p;
1103 int i, j, maxahead, maxbehind, reqcount, shift;
1107 VM_OBJECT_WUNLOCK(object);
1108 bp = getpbuf(&nsw_rcount);
1109 VM_OBJECT_WLOCK(object);
1111 if (!swap_pager_haspage(object, m[0]->pindex, &maxbehind, &maxahead)) {
1112 relpbuf(bp, &nsw_rcount);
1113 return (VM_PAGER_FAIL);
1117 * Clip the readahead and readbehind ranges to exclude resident pages.
1119 if (rahead != NULL) {
1120 KASSERT(reqcount - 1 <= maxahead,
1121 ("page count %d extends beyond swap block", reqcount));
1122 *rahead = imin(*rahead, maxahead - (reqcount - 1));
1123 pindex = m[reqcount - 1]->pindex;
1124 msucc = TAILQ_NEXT(m[reqcount - 1], listq);
1125 if (msucc != NULL && msucc->pindex - pindex - 1 < *rahead)
1126 *rahead = msucc->pindex - pindex - 1;
1128 if (rbehind != NULL) {
1129 *rbehind = imin(*rbehind, maxbehind);
1130 pindex = m[0]->pindex;
1131 mpred = TAILQ_PREV(m[0], pglist, listq);
1132 if (mpred != NULL && pindex - mpred->pindex - 1 < *rbehind)
1133 *rbehind = pindex - mpred->pindex - 1;
1137 * Allocate readahead and readbehind pages.
1139 shift = rbehind != NULL ? *rbehind : 0;
1141 for (i = 1; i <= shift; i++) {
1142 p = vm_page_alloc(object, m[0]->pindex - i,
1145 /* Shift allocated pages to the left. */
1146 for (j = 0; j < i - 1; j++)
1148 bp->b_pages[j + shift - i + 1];
1151 bp->b_pages[shift - i] = p;
1156 for (i = 0; i < reqcount; i++)
1157 bp->b_pages[i + shift] = m[i];
1158 if (rahead != NULL) {
1159 for (i = 0; i < *rahead; i++) {
1160 p = vm_page_alloc(object,
1161 m[reqcount - 1]->pindex + i + 1, VM_ALLOC_NORMAL);
1164 bp->b_pages[shift + reqcount + i] = p;
1168 if (rbehind != NULL)
1173 vm_object_pip_add(object, count);
1175 for (i = 0; i < count; i++)
1176 bp->b_pages[i]->oflags |= VPO_SWAPINPROG;
1178 pindex = bp->b_pages[0]->pindex;
1179 blk = swp_pager_meta_ctl(object, pindex, 0);
1180 KASSERT(blk != SWAPBLK_NONE,
1181 ("no swap blocking containing %p(%jx)", object, (uintmax_t)pindex));
1183 VM_OBJECT_WUNLOCK(object);
1185 bp->b_flags |= B_PAGING;
1186 bp->b_iocmd = BIO_READ;
1187 bp->b_iodone = swp_pager_async_iodone;
1188 bp->b_rcred = crhold(thread0.td_ucred);
1189 bp->b_wcred = crhold(thread0.td_ucred);
1191 bp->b_bcount = PAGE_SIZE * count;
1192 bp->b_bufsize = PAGE_SIZE * count;
1193 bp->b_npages = count;
1194 bp->b_pgbefore = rbehind != NULL ? *rbehind : 0;
1195 bp->b_pgafter = rahead != NULL ? *rahead : 0;
1197 VM_CNT_INC(v_swapin);
1198 VM_CNT_ADD(v_swappgsin, count);
1201 * perform the I/O. NOTE!!! bp cannot be considered valid after
1202 * this point because we automatically release it on completion.
1203 * Instead, we look at the one page we are interested in which we
1204 * still hold a lock on even through the I/O completion.
1206 * The other pages in our m[] array are also released on completion,
1207 * so we cannot assume they are valid anymore either.
1209 * NOTE: b_blkno is destroyed by the call to swapdev_strategy
1212 swp_pager_strategy(bp);
1215 * Wait for the pages we want to complete. VPO_SWAPINPROG is always
1216 * cleared on completion. If an I/O error occurs, SWAPBLK_NONE
1217 * is set in the metadata for each page in the request.
1219 VM_OBJECT_WLOCK(object);
1220 while ((m[0]->oflags & VPO_SWAPINPROG) != 0) {
1221 m[0]->oflags |= VPO_SWAPSLEEP;
1222 VM_CNT_INC(v_intrans);
1223 if (VM_OBJECT_SLEEP(object, &object->paging_in_progress, PSWP,
1224 "swread", hz * 20)) {
1226 "swap_pager: indefinite wait buffer: bufobj: %p, blkno: %jd, size: %ld\n",
1227 bp->b_bufobj, (intmax_t)bp->b_blkno, bp->b_bcount);
1232 * If we had an unrecoverable read error pages will not be valid.
1234 for (i = 0; i < reqcount; i++)
1235 if (m[i]->valid != VM_PAGE_BITS_ALL)
1236 return (VM_PAGER_ERROR);
1238 return (VM_PAGER_OK);
1241 * A final note: in a low swap situation, we cannot deallocate swap
1242 * and mark a page dirty here because the caller is likely to mark
1243 * the page clean when we return, causing the page to possibly revert
1244 * to all-zero's later.
1249 * swap_pager_getpages_async():
1251 * Right now this is emulation of asynchronous operation on top of
1252 * swap_pager_getpages().
1255 swap_pager_getpages_async(vm_object_t object, vm_page_t *m, int count,
1256 int *rbehind, int *rahead, pgo_getpages_iodone_t iodone, void *arg)
1260 r = swap_pager_getpages(object, m, count, rbehind, rahead);
1261 VM_OBJECT_WUNLOCK(object);
1266 case VM_PAGER_ERROR:
1273 panic("unhandled swap_pager_getpages() error %d", r);
1275 (iodone)(arg, m, count, error);
1276 VM_OBJECT_WLOCK(object);
1282 * swap_pager_putpages:
1284 * Assign swap (if necessary) and initiate I/O on the specified pages.
1286 * We support both OBJT_DEFAULT and OBJT_SWAP objects. DEFAULT objects
1287 * are automatically converted to SWAP objects.
1289 * In a low memory situation we may block in VOP_STRATEGY(), but the new
1290 * vm_page reservation system coupled with properly written VFS devices
1291 * should ensure that no low-memory deadlock occurs. This is an area
1294 * The parent has N vm_object_pip_add() references prior to
1295 * calling us and will remove references for rtvals[] that are
1296 * not set to VM_PAGER_PEND. We need to remove the rest on I/O
1299 * The parent has soft-busy'd the pages it passes us and will unbusy
1300 * those whos rtvals[] entry is not set to VM_PAGER_PEND on return.
1301 * We need to unbusy the rest on I/O completion.
1304 swap_pager_putpages(vm_object_t object, vm_page_t *m, int count,
1305 int flags, int *rtvals)
1310 if (count && m[0]->object != object) {
1311 panic("swap_pager_putpages: object mismatch %p/%p",
1320 * Turn object into OBJT_SWAP
1321 * check for bogus sysops
1322 * force sync if not pageout process
1324 if (object->type != OBJT_SWAP)
1325 swp_pager_meta_build(object, 0, SWAPBLK_NONE);
1326 VM_OBJECT_WUNLOCK(object);
1329 if (curproc != pageproc)
1332 sync = (flags & VM_PAGER_PUT_SYNC) != 0;
1337 * Assign swap blocks and issue I/O. We reallocate swap on the fly.
1338 * The page is left dirty until the pageout operation completes
1341 for (i = 0; i < count; i += n) {
1347 * Maximum I/O size is limited by a number of factors.
1349 n = min(BLIST_MAX_ALLOC, count - i);
1350 n = min(n, nsw_cluster_max);
1353 * Get biggest block of swap we can. If we fail, fall
1354 * back and try to allocate a smaller block. Don't go
1355 * overboard trying to allocate space if it would overly
1359 (blk = swp_pager_getswapspace(n)) == SWAPBLK_NONE &&
1364 if (blk == SWAPBLK_NONE) {
1365 for (j = 0; j < n; ++j)
1366 rtvals[i+j] = VM_PAGER_FAIL;
1371 * All I/O parameters have been satisfied, build the I/O
1372 * request and assign the swap space.
1375 bp = getpbuf(&nsw_wcount_sync);
1377 bp = getpbuf(&nsw_wcount_async);
1378 bp->b_flags = B_ASYNC;
1380 bp->b_flags |= B_PAGING;
1381 bp->b_iocmd = BIO_WRITE;
1383 bp->b_rcred = crhold(thread0.td_ucred);
1384 bp->b_wcred = crhold(thread0.td_ucred);
1385 bp->b_bcount = PAGE_SIZE * n;
1386 bp->b_bufsize = PAGE_SIZE * n;
1389 VM_OBJECT_WLOCK(object);
1390 for (j = 0; j < n; ++j) {
1391 vm_page_t mreq = m[i+j];
1393 swp_pager_meta_build(
1398 MPASS(mreq->dirty == VM_PAGE_BITS_ALL);
1399 mreq->oflags |= VPO_SWAPINPROG;
1400 bp->b_pages[j] = mreq;
1402 VM_OBJECT_WUNLOCK(object);
1405 * Must set dirty range for NFS to work.
1408 bp->b_dirtyend = bp->b_bcount;
1410 VM_CNT_INC(v_swapout);
1411 VM_CNT_ADD(v_swappgsout, bp->b_npages);
1414 * We unconditionally set rtvals[] to VM_PAGER_PEND so that we
1415 * can call the async completion routine at the end of a
1416 * synchronous I/O operation. Otherwise, our caller would
1417 * perform duplicate unbusy and wakeup operations on the page
1418 * and object, respectively.
1420 for (j = 0; j < n; j++)
1421 rtvals[i + j] = VM_PAGER_PEND;
1426 * NOTE: b_blkno is destroyed by the call to swapdev_strategy
1428 if (sync == FALSE) {
1429 bp->b_iodone = swp_pager_async_iodone;
1431 swp_pager_strategy(bp);
1438 * NOTE: b_blkno is destroyed by the call to swapdev_strategy
1440 bp->b_iodone = bdone;
1441 swp_pager_strategy(bp);
1444 * Wait for the sync I/O to complete.
1446 bwait(bp, PVM, "swwrt");
1449 * Now that we are through with the bp, we can call the
1450 * normal async completion, which frees everything up.
1452 swp_pager_async_iodone(bp);
1454 VM_OBJECT_WLOCK(object);
1458 * swp_pager_async_iodone:
1460 * Completion routine for asynchronous reads and writes from/to swap.
1461 * Also called manually by synchronous code to finish up a bp.
1463 * This routine may not sleep.
1466 swp_pager_async_iodone(struct buf *bp)
1469 vm_object_t object = NULL;
1474 if (bp->b_ioflags & BIO_ERROR) {
1476 "swap_pager: I/O error - %s failed; blkno %ld,"
1477 "size %ld, error %d\n",
1478 ((bp->b_iocmd == BIO_READ) ? "pagein" : "pageout"),
1486 * remove the mapping for kernel virtual
1489 pmap_qremove((vm_offset_t)bp->b_data, bp->b_npages);
1491 bp->b_data = bp->b_kvabase;
1494 object = bp->b_pages[0]->object;
1495 VM_OBJECT_WLOCK(object);
1499 * cleanup pages. If an error occurs writing to swap, we are in
1500 * very serious trouble. If it happens to be a disk error, though,
1501 * we may be able to recover by reassigning the swap later on. So
1502 * in this case we remove the m->swapblk assignment for the page
1503 * but do not free it in the rlist. The errornous block(s) are thus
1504 * never reallocated as swap. Redirty the page and continue.
1506 for (i = 0; i < bp->b_npages; ++i) {
1507 vm_page_t m = bp->b_pages[i];
1509 m->oflags &= ~VPO_SWAPINPROG;
1510 if (m->oflags & VPO_SWAPSLEEP) {
1511 m->oflags &= ~VPO_SWAPSLEEP;
1512 wakeup(&object->paging_in_progress);
1515 if (bp->b_ioflags & BIO_ERROR) {
1517 * If an error occurs I'd love to throw the swapblk
1518 * away without freeing it back to swapspace, so it
1519 * can never be used again. But I can't from an
1522 if (bp->b_iocmd == BIO_READ) {
1524 * NOTE: for reads, m->dirty will probably
1525 * be overridden by the original caller of
1526 * getpages so don't play cute tricks here.
1531 * If a write error occurs, reactivate page
1532 * so it doesn't clog the inactive list,
1533 * then finish the I/O.
1535 MPASS(m->dirty == VM_PAGE_BITS_ALL);
1537 vm_page_activate(m);
1541 } else if (bp->b_iocmd == BIO_READ) {
1543 * NOTE: for reads, m->dirty will probably be
1544 * overridden by the original caller of getpages so
1545 * we cannot set them in order to free the underlying
1546 * swap in a low-swap situation. I don't think we'd
1547 * want to do that anyway, but it was an optimization
1548 * that existed in the old swapper for a time before
1549 * it got ripped out due to precisely this problem.
1551 KASSERT(!pmap_page_is_mapped(m),
1552 ("swp_pager_async_iodone: page %p is mapped", m));
1553 KASSERT(m->dirty == 0,
1554 ("swp_pager_async_iodone: page %p is dirty", m));
1556 m->valid = VM_PAGE_BITS_ALL;
1557 if (i < bp->b_pgbefore ||
1558 i >= bp->b_npages - bp->b_pgafter)
1559 vm_page_readahead_finish(m);
1562 * For write success, clear the dirty
1563 * status, then finish the I/O ( which decrements the
1564 * busy count and possibly wakes waiter's up ).
1565 * A page is only written to swap after a period of
1566 * inactivity. Therefore, we do not expect it to be
1569 KASSERT(!pmap_page_is_write_mapped(m),
1570 ("swp_pager_async_iodone: page %p is not write"
1574 vm_page_deactivate_noreuse(m);
1581 * adjust pip. NOTE: the original parent may still have its own
1582 * pip refs on the object.
1584 if (object != NULL) {
1585 vm_object_pip_wakeupn(object, bp->b_npages);
1586 VM_OBJECT_WUNLOCK(object);
1590 * swapdev_strategy() manually sets b_vp and b_bufobj before calling
1591 * bstrategy(). Set them back to NULL now we're done with it, or we'll
1592 * trigger a KASSERT in relpbuf().
1596 bp->b_bufobj = NULL;
1599 * release the physical I/O buffer
1603 ((bp->b_iocmd == BIO_READ) ? &nsw_rcount :
1604 ((bp->b_flags & B_ASYNC) ?
1613 swap_pager_nswapdev(void)
1620 * SWP_PAGER_FORCE_PAGEIN() - force a swap block to be paged in
1622 * This routine dissociates the page at the given index within an object
1623 * from its backing store, paging it in if it does not reside in memory.
1624 * If the page is paged in, it is marked dirty and placed in the laundry
1625 * queue. The page is marked dirty because it no longer has backing
1626 * store. It is placed in the laundry queue because it has not been
1627 * accessed recently. Otherwise, it would already reside in memory.
1629 * We also attempt to swap in all other pages in the swap block.
1630 * However, we only guarantee that the one at the specified index is
1633 * XXX - The code to page the whole block in doesn't work, so we
1634 * revert to the one-by-one behavior for now. Sigh.
1637 swp_pager_force_pagein(vm_object_t object, vm_pindex_t pindex)
1641 vm_object_pip_add(object, 1);
1642 m = vm_page_grab(object, pindex, VM_ALLOC_NORMAL);
1643 if (m->valid == VM_PAGE_BITS_ALL) {
1644 vm_object_pip_wakeup(object);
1648 if (m->wire_count == 0 && m->queue == PQ_NONE)
1649 panic("page %p is neither wired nor queued", m);
1653 vm_pager_page_unswapped(m);
1657 if (swap_pager_getpages(object, &m, 1, NULL, NULL) != VM_PAGER_OK)
1658 panic("swap_pager_force_pagein: read from swap failed");/*XXX*/
1659 vm_object_pip_wakeup(object);
1665 vm_pager_page_unswapped(m);
1669 * swap_pager_swapoff:
1671 * Page in all of the pages that have been paged out to the
1672 * given device. The corresponding blocks in the bitmap must be
1673 * marked as allocated and the device must be flagged SW_CLOSING.
1674 * There may be no processes swapped out to the device.
1676 * This routine may block.
1679 swap_pager_swapoff(struct swdevt *sp)
1686 sx_assert(&swdev_syscall_lock, SA_XLOCKED);
1690 mtx_lock(&vm_object_list_mtx);
1691 TAILQ_FOREACH(object, &vm_object_list, object_list) {
1692 if (object->type != OBJT_SWAP)
1694 mtx_unlock(&vm_object_list_mtx);
1695 /* Depends on type-stability. */
1696 VM_OBJECT_WLOCK(object);
1699 * Dead objects are eventually terminated on their own.
1701 if ((object->flags & OBJ_DEAD) != 0)
1705 * Sync with fences placed after pctrie
1706 * initialization. We must not access pctrie below
1707 * unless we checked that our object is swap and not
1710 atomic_thread_fence_acq();
1711 if (object->type != OBJT_SWAP)
1714 for (pi = 0; (sb = SWAP_PCTRIE_LOOKUP_GE(
1715 &object->un_pager.swp.swp_blks, pi)) != NULL; ) {
1716 pi = sb->p + SWAP_META_PAGES;
1717 for (i = 0; i < SWAP_META_PAGES; i++) {
1718 if (sb->d[i] == SWAPBLK_NONE)
1720 if (swp_pager_isondev(sb->d[i], sp))
1721 swp_pager_force_pagein(object,
1726 VM_OBJECT_WUNLOCK(object);
1727 mtx_lock(&vm_object_list_mtx);
1729 mtx_unlock(&vm_object_list_mtx);
1733 * Objects may be locked or paging to the device being
1734 * removed, so we will miss their pages and need to
1735 * make another pass. We have marked this device as
1736 * SW_CLOSING, so the activity should finish soon.
1739 if (retries > 100) {
1740 panic("swapoff: failed to locate %d swap blocks",
1743 pause("swpoff", hz / 20);
1746 EVENTHANDLER_INVOKE(swapoff, sp);
1749 /************************************************************************
1751 ************************************************************************
1753 * These routines manipulate the swap metadata stored in the
1756 * Swap metadata is implemented with a global hash and not directly
1757 * linked into the object. Instead the object simply contains
1758 * appropriate tracking counters.
1762 * SWP_PAGER_META_BUILD() - add swap block to swap meta data for object
1764 * We first convert the object to a swap object if it is a default
1767 * The specified swapblk is added to the object's swap metadata. If
1768 * the swapblk is not valid, it is freed instead. Any previously
1769 * assigned swapblk is freed.
1772 swp_pager_meta_build(vm_object_t object, vm_pindex_t pindex, daddr_t swapblk)
1774 static volatile int swblk_zone_exhausted, swpctrie_zone_exhausted;
1775 struct swblk *sb, *sb1;
1776 vm_pindex_t modpi, rdpi;
1779 VM_OBJECT_ASSERT_WLOCKED(object);
1782 * Convert default object to swap object if necessary
1784 if (object->type != OBJT_SWAP) {
1785 pctrie_init(&object->un_pager.swp.swp_blks);
1788 * Ensure that swap_pager_swapoff()'s iteration over
1789 * object_list does not see a garbage pctrie.
1791 atomic_thread_fence_rel();
1793 object->type = OBJT_SWAP;
1794 KASSERT(object->handle == NULL, ("default pager with handle"));
1797 rdpi = rounddown(pindex, SWAP_META_PAGES);
1798 sb = SWAP_PCTRIE_LOOKUP(&object->un_pager.swp.swp_blks, rdpi);
1800 if (swapblk == SWAPBLK_NONE)
1803 sb = uma_zalloc(swblk_zone, M_NOWAIT | (curproc ==
1804 pageproc ? M_USE_RESERVE : 0));
1807 for (i = 0; i < SWAP_META_PAGES; i++)
1808 sb->d[i] = SWAPBLK_NONE;
1809 if (atomic_cmpset_int(&swblk_zone_exhausted,
1811 printf("swblk zone ok\n");
1814 VM_OBJECT_WUNLOCK(object);
1815 if (uma_zone_exhausted(swblk_zone)) {
1816 if (atomic_cmpset_int(&swblk_zone_exhausted,
1818 printf("swap blk zone exhausted, "
1819 "increase kern.maxswzone\n");
1820 vm_pageout_oom(VM_OOM_SWAPZ);
1821 pause("swzonxb", 10);
1824 VM_OBJECT_WLOCK(object);
1825 sb = SWAP_PCTRIE_LOOKUP(&object->un_pager.swp.swp_blks,
1829 * Somebody swapped out a nearby page,
1830 * allocating swblk at the rdpi index,
1831 * while we dropped the object lock.
1836 error = SWAP_PCTRIE_INSERT(
1837 &object->un_pager.swp.swp_blks, sb);
1839 if (atomic_cmpset_int(&swpctrie_zone_exhausted,
1841 printf("swpctrie zone ok\n");
1844 VM_OBJECT_WUNLOCK(object);
1845 if (uma_zone_exhausted(swpctrie_zone)) {
1846 if (atomic_cmpset_int(&swpctrie_zone_exhausted,
1848 printf("swap pctrie zone exhausted, "
1849 "increase kern.maxswzone\n");
1850 vm_pageout_oom(VM_OOM_SWAPZ);
1851 pause("swzonxp", 10);
1854 VM_OBJECT_WLOCK(object);
1855 sb1 = SWAP_PCTRIE_LOOKUP(&object->un_pager.swp.swp_blks,
1858 uma_zfree(swblk_zone, sb);
1865 MPASS(sb->p == rdpi);
1867 modpi = pindex % SWAP_META_PAGES;
1868 /* Delete prior contents of metadata. */
1869 if (sb->d[modpi] != SWAPBLK_NONE)
1870 swp_pager_freeswapspace(sb->d[modpi], 1);
1871 /* Enter block into metadata. */
1872 sb->d[modpi] = swapblk;
1875 * Free the swblk if we end up with the empty page run.
1877 if (swapblk == SWAPBLK_NONE) {
1878 for (i = 0; i < SWAP_META_PAGES; i++) {
1879 if (sb->d[i] != SWAPBLK_NONE)
1882 if (i == SWAP_META_PAGES) {
1883 SWAP_PCTRIE_REMOVE(&object->un_pager.swp.swp_blks,
1885 uma_zfree(swblk_zone, sb);
1891 * SWP_PAGER_META_FREE() - free a range of blocks in the object's swap metadata
1893 * The requested range of blocks is freed, with any associated swap
1894 * returned to the swap bitmap.
1896 * This routine will free swap metadata structures as they are cleaned
1897 * out. This routine does *NOT* operate on swap metadata associated
1898 * with resident pages.
1901 swp_pager_meta_free(vm_object_t object, vm_pindex_t pindex, vm_pindex_t count)
1908 VM_OBJECT_ASSERT_WLOCKED(object);
1909 if (object->type != OBJT_SWAP || count == 0)
1912 last = pindex + count - 1;
1914 sb = SWAP_PCTRIE_LOOKUP_GE(&object->un_pager.swp.swp_blks,
1915 rounddown(pindex, SWAP_META_PAGES));
1916 if (sb == NULL || sb->p > last)
1919 for (i = 0; i < SWAP_META_PAGES; i++) {
1920 if (sb->d[i] == SWAPBLK_NONE)
1922 if (pindex <= sb->p + i && sb->p + i <= last) {
1923 swp_pager_freeswapspace(sb->d[i], 1);
1924 sb->d[i] = SWAPBLK_NONE;
1928 pindex = sb->p + SWAP_META_PAGES;
1930 SWAP_PCTRIE_REMOVE(&object->un_pager.swp.swp_blks,
1932 uma_zfree(swblk_zone, sb);
1938 * SWP_PAGER_META_FREE_ALL() - destroy all swap metadata associated with object
1940 * This routine locates and destroys all swap metadata associated with
1944 swp_pager_meta_free_all(vm_object_t object)
1950 VM_OBJECT_ASSERT_WLOCKED(object);
1951 if (object->type != OBJT_SWAP)
1954 for (pindex = 0; (sb = SWAP_PCTRIE_LOOKUP_GE(
1955 &object->un_pager.swp.swp_blks, pindex)) != NULL;) {
1956 pindex = sb->p + SWAP_META_PAGES;
1957 for (i = 0; i < SWAP_META_PAGES; i++) {
1958 if (sb->d[i] != SWAPBLK_NONE)
1959 swp_pager_freeswapspace(sb->d[i], 1);
1961 SWAP_PCTRIE_REMOVE(&object->un_pager.swp.swp_blks, sb->p);
1962 uma_zfree(swblk_zone, sb);
1967 * SWP_PAGER_METACTL() - misc control of swap and vm_page_t meta data.
1969 * This routine is capable of looking up, popping, or freeing
1970 * swapblk assignments in the swap meta data or in the vm_page_t.
1971 * The routine typically returns the swapblk being looked-up, or popped,
1972 * or SWAPBLK_NONE if the block was freed, or SWAPBLK_NONE if the block
1973 * was invalid. This routine will automatically free any invalid
1974 * meta-data swapblks.
1976 * When acting on a busy resident page and paging is in progress, we
1977 * have to wait until paging is complete but otherwise can act on the
1980 * SWM_FREE remove and free swap block from metadata
1981 * SWM_POP remove from meta data but do not free.. pop it out
1984 swp_pager_meta_ctl(vm_object_t object, vm_pindex_t pindex, int flags)
1990 if ((flags & (SWM_FREE | SWM_POP)) != 0)
1991 VM_OBJECT_ASSERT_WLOCKED(object);
1993 VM_OBJECT_ASSERT_LOCKED(object);
1996 * The meta data only exists if the object is OBJT_SWAP
1997 * and even then might not be allocated yet.
1999 if (object->type != OBJT_SWAP)
2000 return (SWAPBLK_NONE);
2002 sb = SWAP_PCTRIE_LOOKUP(&object->un_pager.swp.swp_blks,
2003 rounddown(pindex, SWAP_META_PAGES));
2005 return (SWAPBLK_NONE);
2006 r1 = sb->d[pindex % SWAP_META_PAGES];
2007 if (r1 == SWAPBLK_NONE)
2008 return (SWAPBLK_NONE);
2009 if ((flags & (SWM_FREE | SWM_POP)) != 0) {
2010 sb->d[pindex % SWAP_META_PAGES] = SWAPBLK_NONE;
2011 for (i = 0; i < SWAP_META_PAGES; i++) {
2012 if (sb->d[i] != SWAPBLK_NONE)
2015 if (i == SWAP_META_PAGES) {
2016 SWAP_PCTRIE_REMOVE(&object->un_pager.swp.swp_blks,
2017 rounddown(pindex, SWAP_META_PAGES));
2018 uma_zfree(swblk_zone, sb);
2021 if ((flags & SWM_FREE) != 0) {
2022 swp_pager_freeswapspace(r1, 1);
2029 * Returns the least page index which is greater than or equal to the
2030 * parameter pindex and for which there is a swap block allocated.
2031 * Returns object's size if the object's type is not swap or if there
2032 * are no allocated swap blocks for the object after the requested
2036 swap_pager_find_least(vm_object_t object, vm_pindex_t pindex)
2041 VM_OBJECT_ASSERT_LOCKED(object);
2042 if (object->type != OBJT_SWAP)
2043 return (object->size);
2045 sb = SWAP_PCTRIE_LOOKUP_GE(&object->un_pager.swp.swp_blks,
2046 rounddown(pindex, SWAP_META_PAGES));
2048 return (object->size);
2049 if (sb->p < pindex) {
2050 for (i = pindex % SWAP_META_PAGES; i < SWAP_META_PAGES; i++) {
2051 if (sb->d[i] != SWAPBLK_NONE)
2054 sb = SWAP_PCTRIE_LOOKUP_GE(&object->un_pager.swp.swp_blks,
2055 roundup(pindex, SWAP_META_PAGES));
2057 return (object->size);
2059 for (i = 0; i < SWAP_META_PAGES; i++) {
2060 if (sb->d[i] != SWAPBLK_NONE)
2065 * We get here if a swblk is present in the trie but it
2066 * doesn't map any blocks.
2069 return (object->size);
2073 * System call swapon(name) enables swapping on device name,
2074 * which must be in the swdevsw. Return EBUSY
2075 * if already swapping on this device.
2077 #ifndef _SYS_SYSPROTO_H_
2078 struct swapon_args {
2088 sys_swapon(struct thread *td, struct swapon_args *uap)
2092 struct nameidata nd;
2095 error = priv_check(td, PRIV_SWAPON);
2099 sx_xlock(&swdev_syscall_lock);
2102 * Swap metadata may not fit in the KVM if we have physical
2105 if (swblk_zone == NULL) {
2110 NDINIT(&nd, LOOKUP, ISOPEN | FOLLOW | AUDITVNODE1, UIO_USERSPACE,
2116 NDFREE(&nd, NDF_ONLY_PNBUF);
2119 if (vn_isdisk(vp, &error)) {
2120 error = swapongeom(vp);
2121 } else if (vp->v_type == VREG &&
2122 (vp->v_mount->mnt_vfc->vfc_flags & VFCF_NETWORK) != 0 &&
2123 (error = VOP_GETATTR(vp, &attr, td->td_ucred)) == 0) {
2125 * Allow direct swapping to NFS regular files in the same
2126 * way that nfs_mountroot() sets up diskless swapping.
2128 error = swaponvp(td, vp, attr.va_size / DEV_BSIZE);
2134 sx_xunlock(&swdev_syscall_lock);
2139 * Check that the total amount of swap currently configured does not
2140 * exceed half the theoretical maximum. If it does, print a warning
2144 swapon_check_swzone(void)
2146 unsigned long maxpages, npages;
2148 npages = swap_total / PAGE_SIZE;
2149 /* absolute maximum we can handle assuming 100% efficiency */
2150 maxpages = uma_zone_get_max(swblk_zone) * SWAP_META_PAGES;
2152 /* recommend using no more than half that amount */
2153 if (npages > maxpages / 2) {
2154 printf("warning: total configured swap (%lu pages) "
2155 "exceeds maximum recommended amount (%lu pages).\n",
2156 npages, maxpages / 2);
2157 printf("warning: increase kern.maxswzone "
2158 "or reduce amount of swap.\n");
2163 swaponsomething(struct vnode *vp, void *id, u_long nblks,
2164 sw_strategy_t *strategy, sw_close_t *close, dev_t dev, int flags)
2166 struct swdevt *sp, *tsp;
2171 * nblks is in DEV_BSIZE'd chunks, convert to PAGE_SIZE'd chunks.
2172 * First chop nblks off to page-align it, then convert.
2174 * sw->sw_nblks is in page-sized chunks now too.
2176 nblks &= ~(ctodb(1) - 1);
2177 nblks = dbtoc(nblks);
2180 * If we go beyond this, we get overflows in the radix
2183 mblocks = 0x40000000 / BLIST_META_RADIX;
2184 if (nblks > mblocks) {
2186 "WARNING: reducing swap size to maximum of %luMB per unit\n",
2187 mblocks / 1024 / 1024 * PAGE_SIZE);
2191 sp = malloc(sizeof *sp, M_VMPGDATA, M_WAITOK | M_ZERO);
2196 sp->sw_nblks = nblks;
2198 sp->sw_strategy = strategy;
2199 sp->sw_close = close;
2200 sp->sw_flags = flags;
2202 sp->sw_blist = blist_create(nblks, M_WAITOK);
2204 * Do not free the first two block in order to avoid overwriting
2205 * any bsd label at the front of the partition
2207 blist_free(sp->sw_blist, 2, nblks - 2);
2210 mtx_lock(&sw_dev_mtx);
2211 TAILQ_FOREACH(tsp, &swtailq, sw_list) {
2212 if (tsp->sw_end >= dvbase) {
2214 * We put one uncovered page between the devices
2215 * in order to definitively prevent any cross-device
2218 dvbase = tsp->sw_end + 1;
2221 sp->sw_first = dvbase;
2222 sp->sw_end = dvbase + nblks;
2223 TAILQ_INSERT_TAIL(&swtailq, sp, sw_list);
2225 swap_pager_avail += nblks - 2;
2226 swap_total += (vm_ooffset_t)nblks * PAGE_SIZE;
2227 swapon_check_swzone();
2229 mtx_unlock(&sw_dev_mtx);
2230 EVENTHANDLER_INVOKE(swapon, sp);
2234 * SYSCALL: swapoff(devname)
2236 * Disable swapping on the given device.
2238 * XXX: Badly designed system call: it should use a device index
2239 * rather than filename as specification. We keep sw_vp around
2240 * only to make this work.
2242 #ifndef _SYS_SYSPROTO_H_
2243 struct swapoff_args {
2253 sys_swapoff(struct thread *td, struct swapoff_args *uap)
2256 struct nameidata nd;
2260 error = priv_check(td, PRIV_SWAPOFF);
2264 sx_xlock(&swdev_syscall_lock);
2266 NDINIT(&nd, LOOKUP, FOLLOW | AUDITVNODE1, UIO_USERSPACE, uap->name,
2271 NDFREE(&nd, NDF_ONLY_PNBUF);
2274 mtx_lock(&sw_dev_mtx);
2275 TAILQ_FOREACH(sp, &swtailq, sw_list) {
2276 if (sp->sw_vp == vp)
2279 mtx_unlock(&sw_dev_mtx);
2284 error = swapoff_one(sp, td->td_ucred);
2286 sx_xunlock(&swdev_syscall_lock);
2291 swapoff_one(struct swdevt *sp, struct ucred *cred)
2298 sx_assert(&swdev_syscall_lock, SA_XLOCKED);
2300 (void) vn_lock(sp->sw_vp, LK_EXCLUSIVE | LK_RETRY);
2301 error = mac_system_check_swapoff(cred, sp->sw_vp);
2302 (void) VOP_UNLOCK(sp->sw_vp, 0);
2306 nblks = sp->sw_nblks;
2309 * We can turn off this swap device safely only if the
2310 * available virtual memory in the system will fit the amount
2311 * of data we will have to page back in, plus an epsilon so
2312 * the system doesn't become critically low on swap space.
2314 if (vm_cnt.v_free_count + swap_pager_avail < nblks + nswap_lowat)
2318 * Prevent further allocations on this device.
2320 mtx_lock(&sw_dev_mtx);
2321 sp->sw_flags |= SW_CLOSING;
2322 swap_pager_avail -= blist_fill(sp->sw_blist, 0, nblks);
2323 swap_total -= (vm_ooffset_t)nblks * PAGE_SIZE;
2324 mtx_unlock(&sw_dev_mtx);
2327 * Page in the contents of the device and close it.
2329 swap_pager_swapoff(sp);
2331 sp->sw_close(curthread, sp);
2332 mtx_lock(&sw_dev_mtx);
2334 TAILQ_REMOVE(&swtailq, sp, sw_list);
2336 if (nswapdev == 0) {
2337 swap_pager_full = 2;
2338 swap_pager_almost_full = 1;
2342 mtx_unlock(&sw_dev_mtx);
2343 blist_destroy(sp->sw_blist);
2344 free(sp, M_VMPGDATA);
2351 struct swdevt *sp, *spt;
2352 const char *devname;
2355 sx_xlock(&swdev_syscall_lock);
2357 mtx_lock(&sw_dev_mtx);
2358 TAILQ_FOREACH_SAFE(sp, &swtailq, sw_list, spt) {
2359 mtx_unlock(&sw_dev_mtx);
2360 if (vn_isdisk(sp->sw_vp, NULL))
2361 devname = devtoname(sp->sw_vp->v_rdev);
2364 error = swapoff_one(sp, thread0.td_ucred);
2366 printf("Cannot remove swap device %s (error=%d), "
2367 "skipping.\n", devname, error);
2368 } else if (bootverbose) {
2369 printf("Swap device %s removed.\n", devname);
2371 mtx_lock(&sw_dev_mtx);
2373 mtx_unlock(&sw_dev_mtx);
2375 sx_xunlock(&swdev_syscall_lock);
2379 swap_pager_status(int *total, int *used)
2385 mtx_lock(&sw_dev_mtx);
2386 TAILQ_FOREACH(sp, &swtailq, sw_list) {
2387 *total += sp->sw_nblks;
2388 *used += sp->sw_used;
2390 mtx_unlock(&sw_dev_mtx);
2394 swap_dev_info(int name, struct xswdev *xs, char *devname, size_t len)
2397 const char *tmp_devname;
2402 mtx_lock(&sw_dev_mtx);
2403 TAILQ_FOREACH(sp, &swtailq, sw_list) {
2408 xs->xsw_version = XSWDEV_VERSION;
2409 xs->xsw_dev = sp->sw_dev;
2410 xs->xsw_flags = sp->sw_flags;
2411 xs->xsw_nblks = sp->sw_nblks;
2412 xs->xsw_used = sp->sw_used;
2413 if (devname != NULL) {
2414 if (vn_isdisk(sp->sw_vp, NULL))
2415 tmp_devname = devtoname(sp->sw_vp->v_rdev);
2417 tmp_devname = "[file]";
2418 strncpy(devname, tmp_devname, len);
2423 mtx_unlock(&sw_dev_mtx);
2427 #if defined(COMPAT_FREEBSD11)
2428 #define XSWDEV_VERSION_11 1
2439 sysctl_vm_swap_info(SYSCTL_HANDLER_ARGS)
2442 #if defined(COMPAT_FREEBSD11)
2443 struct xswdev11 xs11;
2447 if (arg2 != 1) /* name length */
2449 error = swap_dev_info(*(int *)arg1, &xs, NULL, 0);
2452 #if defined(COMPAT_FREEBSD11)
2453 if (req->oldlen == sizeof(xs11)) {
2454 xs11.xsw_version = XSWDEV_VERSION_11;
2455 xs11.xsw_dev = xs.xsw_dev; /* truncation */
2456 xs11.xsw_flags = xs.xsw_flags;
2457 xs11.xsw_nblks = xs.xsw_nblks;
2458 xs11.xsw_used = xs.xsw_used;
2459 error = SYSCTL_OUT(req, &xs11, sizeof(xs11));
2462 error = SYSCTL_OUT(req, &xs, sizeof(xs));
2466 SYSCTL_INT(_vm, OID_AUTO, nswapdev, CTLFLAG_RD, &nswapdev, 0,
2467 "Number of swap devices");
2468 SYSCTL_NODE(_vm, OID_AUTO, swap_info, CTLFLAG_RD | CTLFLAG_MPSAFE,
2469 sysctl_vm_swap_info,
2470 "Swap statistics by device");
2473 * Count the approximate swap usage in pages for a vmspace. The
2474 * shadowed or not yet copied on write swap blocks are not accounted.
2475 * The map must be locked.
2478 vmspace_swap_count(struct vmspace *vmspace)
2488 map = &vmspace->vm_map;
2491 for (cur = map->header.next; cur != &map->header; cur = cur->next) {
2492 if ((cur->eflags & MAP_ENTRY_IS_SUB_MAP) != 0)
2494 object = cur->object.vm_object;
2495 if (object == NULL || object->type != OBJT_SWAP)
2497 VM_OBJECT_RLOCK(object);
2498 if (object->type != OBJT_SWAP)
2500 pi = OFF_TO_IDX(cur->offset);
2501 e = pi + OFF_TO_IDX(cur->end - cur->start);
2502 for (;; pi = sb->p + SWAP_META_PAGES) {
2503 sb = SWAP_PCTRIE_LOOKUP_GE(
2504 &object->un_pager.swp.swp_blks, pi);
2505 if (sb == NULL || sb->p >= e)
2507 for (i = 0; i < SWAP_META_PAGES; i++) {
2508 if (sb->p + i < e &&
2509 sb->d[i] != SWAPBLK_NONE)
2514 VM_OBJECT_RUNLOCK(object);
2522 * Swapping onto disk devices.
2526 static g_orphan_t swapgeom_orphan;
2528 static struct g_class g_swap_class = {
2530 .version = G_VERSION,
2531 .orphan = swapgeom_orphan,
2534 DECLARE_GEOM_CLASS(g_swap_class, g_class);
2538 swapgeom_close_ev(void *arg, int flags)
2540 struct g_consumer *cp;
2543 g_access(cp, -1, -1, 0);
2545 g_destroy_consumer(cp);
2549 * Add a reference to the g_consumer for an inflight transaction.
2552 swapgeom_acquire(struct g_consumer *cp)
2555 mtx_assert(&sw_dev_mtx, MA_OWNED);
2560 * Remove a reference from the g_consumer. Post a close event if all
2561 * references go away, since the function might be called from the
2565 swapgeom_release(struct g_consumer *cp, struct swdevt *sp)
2568 mtx_assert(&sw_dev_mtx, MA_OWNED);
2570 if (cp->index == 0) {
2571 if (g_post_event(swapgeom_close_ev, cp, M_NOWAIT, NULL) == 0)
2577 swapgeom_done(struct bio *bp2)
2581 struct g_consumer *cp;
2583 bp = bp2->bio_caller2;
2585 bp->b_ioflags = bp2->bio_flags;
2587 bp->b_ioflags |= BIO_ERROR;
2588 bp->b_resid = bp->b_bcount - bp2->bio_completed;
2589 bp->b_error = bp2->bio_error;
2591 sp = bp2->bio_caller1;
2592 mtx_lock(&sw_dev_mtx);
2593 swapgeom_release(cp, sp);
2594 mtx_unlock(&sw_dev_mtx);
2599 swapgeom_strategy(struct buf *bp, struct swdevt *sp)
2602 struct g_consumer *cp;
2604 mtx_lock(&sw_dev_mtx);
2607 mtx_unlock(&sw_dev_mtx);
2608 bp->b_error = ENXIO;
2609 bp->b_ioflags |= BIO_ERROR;
2613 swapgeom_acquire(cp);
2614 mtx_unlock(&sw_dev_mtx);
2615 if (bp->b_iocmd == BIO_WRITE)
2618 bio = g_alloc_bio();
2620 mtx_lock(&sw_dev_mtx);
2621 swapgeom_release(cp, sp);
2622 mtx_unlock(&sw_dev_mtx);
2623 bp->b_error = ENOMEM;
2624 bp->b_ioflags |= BIO_ERROR;
2629 bio->bio_caller1 = sp;
2630 bio->bio_caller2 = bp;
2631 bio->bio_cmd = bp->b_iocmd;
2632 bio->bio_offset = (bp->b_blkno - sp->sw_first) * PAGE_SIZE;
2633 bio->bio_length = bp->b_bcount;
2634 bio->bio_done = swapgeom_done;
2635 if (!buf_mapped(bp)) {
2636 bio->bio_ma = bp->b_pages;
2637 bio->bio_data = unmapped_buf;
2638 bio->bio_ma_offset = (vm_offset_t)bp->b_offset & PAGE_MASK;
2639 bio->bio_ma_n = bp->b_npages;
2640 bio->bio_flags |= BIO_UNMAPPED;
2642 bio->bio_data = bp->b_data;
2645 g_io_request(bio, cp);
2650 swapgeom_orphan(struct g_consumer *cp)
2655 mtx_lock(&sw_dev_mtx);
2656 TAILQ_FOREACH(sp, &swtailq, sw_list) {
2657 if (sp->sw_id == cp) {
2658 sp->sw_flags |= SW_CLOSING;
2663 * Drop reference we were created with. Do directly since we're in a
2664 * special context where we don't have to queue the call to
2665 * swapgeom_close_ev().
2668 destroy = ((sp != NULL) && (cp->index == 0));
2671 mtx_unlock(&sw_dev_mtx);
2673 swapgeom_close_ev(cp, 0);
2677 swapgeom_close(struct thread *td, struct swdevt *sw)
2679 struct g_consumer *cp;
2681 mtx_lock(&sw_dev_mtx);
2684 mtx_unlock(&sw_dev_mtx);
2687 * swapgeom_close() may be called from the biodone context,
2688 * where we cannot perform topology changes. Delegate the
2689 * work to the events thread.
2692 g_waitfor_event(swapgeom_close_ev, cp, M_WAITOK, NULL);
2696 swapongeom_locked(struct cdev *dev, struct vnode *vp)
2698 struct g_provider *pp;
2699 struct g_consumer *cp;
2700 static struct g_geom *gp;
2705 pp = g_dev_getprovider(dev);
2708 mtx_lock(&sw_dev_mtx);
2709 TAILQ_FOREACH(sp, &swtailq, sw_list) {
2711 if (cp != NULL && cp->provider == pp) {
2712 mtx_unlock(&sw_dev_mtx);
2716 mtx_unlock(&sw_dev_mtx);
2718 gp = g_new_geomf(&g_swap_class, "swap");
2719 cp = g_new_consumer(gp);
2720 cp->index = 1; /* Number of active I/Os, plus one for being active. */
2721 cp->flags |= G_CF_DIRECT_SEND | G_CF_DIRECT_RECEIVE;
2724 * XXX: Every time you think you can improve the margin for
2725 * footshooting, somebody depends on the ability to do so:
2726 * savecore(8) wants to write to our swapdev so we cannot
2727 * set an exclusive count :-(
2729 error = g_access(cp, 1, 1, 0);
2732 g_destroy_consumer(cp);
2735 nblks = pp->mediasize / DEV_BSIZE;
2736 swaponsomething(vp, cp, nblks, swapgeom_strategy,
2737 swapgeom_close, dev2udev(dev),
2738 (pp->flags & G_PF_ACCEPT_UNMAPPED) != 0 ? SW_UNMAPPED : 0);
2743 swapongeom(struct vnode *vp)
2747 vn_lock(vp, LK_EXCLUSIVE | LK_RETRY);
2748 if (vp->v_type != VCHR || (vp->v_iflag & VI_DOOMED) != 0) {
2752 error = swapongeom_locked(vp->v_rdev, vp);
2753 g_topology_unlock();
2762 * This is used mainly for network filesystem (read: probably only tested
2763 * with NFS) swapfiles.
2768 swapdev_strategy(struct buf *bp, struct swdevt *sp)
2772 bp->b_blkno = ctodb(bp->b_blkno - sp->sw_first);
2776 if (bp->b_iocmd == BIO_WRITE) {
2778 bufobj_wdrop(bp->b_bufobj);
2779 bufobj_wref(&vp2->v_bufobj);
2781 if (bp->b_bufobj != &vp2->v_bufobj)
2782 bp->b_bufobj = &vp2->v_bufobj;
2784 bp->b_iooffset = dbtob(bp->b_blkno);
2790 swapdev_close(struct thread *td, struct swdevt *sp)
2793 VOP_CLOSE(sp->sw_vp, FREAD | FWRITE, td->td_ucred, td);
2799 swaponvp(struct thread *td, struct vnode *vp, u_long nblks)
2806 mtx_lock(&sw_dev_mtx);
2807 TAILQ_FOREACH(sp, &swtailq, sw_list) {
2808 if (sp->sw_id == vp) {
2809 mtx_unlock(&sw_dev_mtx);
2813 mtx_unlock(&sw_dev_mtx);
2815 (void) vn_lock(vp, LK_EXCLUSIVE | LK_RETRY);
2817 error = mac_system_check_swapon(td->td_ucred, vp);
2820 error = VOP_OPEN(vp, FREAD | FWRITE, td->td_ucred, td, NULL);
2821 (void) VOP_UNLOCK(vp, 0);
2825 swaponsomething(vp, vp, nblks, swapdev_strategy, swapdev_close,
2831 sysctl_swap_async_max(SYSCTL_HANDLER_ARGS)
2835 new = nsw_wcount_async_max;
2836 error = sysctl_handle_int(oidp, &new, 0, req);
2837 if (error != 0 || req->newptr == NULL)
2840 if (new > nswbuf / 2 || new < 1)
2843 mtx_lock(&pbuf_mtx);
2844 while (nsw_wcount_async_max != new) {
2846 * Adjust difference. If the current async count is too low,
2847 * we will need to sqeeze our update slowly in. Sleep with a
2848 * higher priority than getpbuf() to finish faster.
2850 n = new - nsw_wcount_async_max;
2851 if (nsw_wcount_async + n >= 0) {
2852 nsw_wcount_async += n;
2853 nsw_wcount_async_max += n;
2854 wakeup(&nsw_wcount_async);
2856 nsw_wcount_async_max -= nsw_wcount_async;
2857 nsw_wcount_async = 0;
2858 msleep(&nsw_wcount_async, &pbuf_mtx, PSWP,
2862 mtx_unlock(&pbuf_mtx);