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, OID_AUTO, 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.
545 printf("Swap blk zone entries reduced from %lu to %lu.\n",
547 swap_maxpages = n * SWAP_META_PAGES;
548 swzone = n * sizeof(struct swblk);
549 if (!uma_zone_reserve_kva(swpctrie_zone, n))
550 printf("Cannot reserve swap pctrie zone, "
551 "reduce kern.maxswzone.\n");
555 swap_pager_alloc_init(void *handle, struct ucred *cred, vm_ooffset_t size,
561 if (!swap_reserve_by_cred(size, cred))
567 * The un_pager.swp.swp_blks trie is initialized by
568 * vm_object_allocate() to ensure the correct order of
569 * visibility to other threads.
571 object = vm_object_allocate(OBJT_SWAP, OFF_TO_IDX(offset +
574 object->handle = handle;
577 object->charge = size;
583 * SWAP_PAGER_ALLOC() - allocate a new OBJT_SWAP VM object and instantiate
584 * its metadata structures.
586 * This routine is called from the mmap and fork code to create a new
589 * This routine must ensure that no live duplicate is created for
590 * the named object request, which is protected against by
591 * holding the sw_alloc_sx lock in case handle != NULL.
594 swap_pager_alloc(void *handle, vm_ooffset_t size, vm_prot_t prot,
595 vm_ooffset_t offset, struct ucred *cred)
599 if (handle != NULL) {
601 * Reference existing named region or allocate new one. There
602 * should not be a race here against swp_pager_meta_build()
603 * as called from vm_page_remove() in regards to the lookup
606 sx_xlock(&sw_alloc_sx);
607 object = vm_pager_object_lookup(NOBJLIST(handle), handle);
608 if (object == NULL) {
609 object = swap_pager_alloc_init(handle, cred, size,
611 if (object != NULL) {
612 TAILQ_INSERT_TAIL(NOBJLIST(object->handle),
613 object, pager_object_list);
616 sx_xunlock(&sw_alloc_sx);
618 object = swap_pager_alloc_init(handle, cred, size, offset);
624 * SWAP_PAGER_DEALLOC() - remove swap metadata from object
626 * The swap backing for the object is destroyed. The code is
627 * designed such that we can reinstantiate it later, but this
628 * routine is typically called only when the entire object is
629 * about to be destroyed.
631 * The object must be locked.
634 swap_pager_dealloc(vm_object_t object)
637 VM_OBJECT_ASSERT_WLOCKED(object);
638 KASSERT((object->flags & OBJ_DEAD) != 0, ("dealloc of reachable obj"));
641 * Remove from list right away so lookups will fail if we block for
642 * pageout completion.
644 if (object->handle != NULL) {
645 VM_OBJECT_WUNLOCK(object);
646 sx_xlock(&sw_alloc_sx);
647 TAILQ_REMOVE(NOBJLIST(object->handle), object,
649 sx_xunlock(&sw_alloc_sx);
650 VM_OBJECT_WLOCK(object);
653 vm_object_pip_wait(object, "swpdea");
656 * Free all remaining metadata. We only bother to free it from
657 * the swap meta data. We do not attempt to free swapblk's still
658 * associated with vm_page_t's for this object. We do not care
659 * if paging is still in progress on some objects.
661 swp_pager_meta_free_all(object);
662 object->handle = NULL;
663 object->type = OBJT_DEAD;
666 /************************************************************************
667 * SWAP PAGER BITMAP ROUTINES *
668 ************************************************************************/
671 * SWP_PAGER_GETSWAPSPACE() - allocate raw swap space
673 * Allocate swap for the requested number of pages. The starting
674 * swap block number (a page index) is returned or SWAPBLK_NONE
675 * if the allocation failed.
677 * Also has the side effect of advising that somebody made a mistake
678 * when they configured swap and didn't configure enough.
680 * This routine may not sleep.
682 * We allocate in round-robin fashion from the configured devices.
685 swp_pager_getswapspace(int npages)
692 mtx_lock(&sw_dev_mtx);
694 for (i = 0; i < nswapdev; i++) {
696 sp = TAILQ_FIRST(&swtailq);
697 if (!(sp->sw_flags & SW_CLOSING)) {
698 blk = blist_alloc(sp->sw_blist, npages);
699 if (blk != SWAPBLK_NONE) {
701 sp->sw_used += npages;
702 swap_pager_avail -= npages;
704 swdevhd = TAILQ_NEXT(sp, sw_list);
708 sp = TAILQ_NEXT(sp, sw_list);
710 if (swap_pager_full != 2) {
711 printf("swap_pager_getswapspace(%d): failed\n", npages);
713 swap_pager_almost_full = 1;
717 mtx_unlock(&sw_dev_mtx);
722 swp_pager_isondev(daddr_t blk, struct swdevt *sp)
725 return (blk >= sp->sw_first && blk < sp->sw_end);
729 swp_pager_strategy(struct buf *bp)
733 mtx_lock(&sw_dev_mtx);
734 TAILQ_FOREACH(sp, &swtailq, sw_list) {
735 if (bp->b_blkno >= sp->sw_first && bp->b_blkno < sp->sw_end) {
736 mtx_unlock(&sw_dev_mtx);
737 if ((sp->sw_flags & SW_UNMAPPED) != 0 &&
738 unmapped_buf_allowed) {
739 bp->b_data = unmapped_buf;
742 pmap_qenter((vm_offset_t)bp->b_data,
743 &bp->b_pages[0], bp->b_bcount / PAGE_SIZE);
745 sp->sw_strategy(bp, sp);
749 panic("Swapdev not found");
754 * SWP_PAGER_FREESWAPSPACE() - free raw swap space
756 * This routine returns the specified swap blocks back to the bitmap.
758 * This routine may not sleep.
761 swp_pager_freeswapspace(daddr_t blk, int npages)
765 mtx_lock(&sw_dev_mtx);
766 TAILQ_FOREACH(sp, &swtailq, sw_list) {
767 if (blk >= sp->sw_first && blk < sp->sw_end) {
768 sp->sw_used -= npages;
770 * If we are attempting to stop swapping on
771 * this device, we don't want to mark any
772 * blocks free lest they be reused.
774 if ((sp->sw_flags & SW_CLOSING) == 0) {
775 blist_free(sp->sw_blist, blk - sp->sw_first,
777 swap_pager_avail += npages;
780 mtx_unlock(&sw_dev_mtx);
784 panic("Swapdev not found");
788 * SYSCTL_SWAP_FRAGMENTATION() - produce raw swap space stats
791 sysctl_swap_fragmentation(SYSCTL_HANDLER_ARGS)
798 error = sysctl_wire_old_buffer(req, 0);
801 sbuf_new_for_sysctl(&sbuf, NULL, 128, req);
802 mtx_lock(&sw_dev_mtx);
803 TAILQ_FOREACH(sp, &swtailq, sw_list) {
804 if (vn_isdisk(sp->sw_vp, NULL))
805 devname = devtoname(sp->sw_vp->v_rdev);
808 sbuf_printf(&sbuf, "\nFree space on device %s:\n", devname);
809 blist_stats(sp->sw_blist, &sbuf);
811 mtx_unlock(&sw_dev_mtx);
812 error = sbuf_finish(&sbuf);
818 * SWAP_PAGER_FREESPACE() - frees swap blocks associated with a page
819 * range within an object.
821 * This is a globally accessible routine.
823 * This routine removes swapblk assignments from swap metadata.
825 * The external callers of this routine typically have already destroyed
826 * or renamed vm_page_t's associated with this range in the object so
829 * The object must be locked.
832 swap_pager_freespace(vm_object_t object, vm_pindex_t start, vm_size_t size)
835 swp_pager_meta_free(object, start, size);
839 * SWAP_PAGER_RESERVE() - reserve swap blocks in object
841 * Assigns swap blocks to the specified range within the object. The
842 * swap blocks are not zeroed. Any previous swap assignment is destroyed.
844 * Returns 0 on success, -1 on failure.
847 swap_pager_reserve(vm_object_t object, vm_pindex_t start, vm_size_t size)
850 daddr_t blk = SWAPBLK_NONE;
851 vm_pindex_t beg = start; /* save start index */
853 VM_OBJECT_WLOCK(object);
857 while ((blk = swp_pager_getswapspace(n)) == SWAPBLK_NONE) {
860 swp_pager_meta_free(object, beg, start - beg);
861 VM_OBJECT_WUNLOCK(object);
866 swp_pager_meta_build(object, start, blk);
872 swp_pager_meta_free(object, start, n);
873 VM_OBJECT_WUNLOCK(object);
878 * SWAP_PAGER_COPY() - copy blocks from source pager to destination pager
879 * and destroy the source.
881 * Copy any valid swapblks from the source to the destination. In
882 * cases where both the source and destination have a valid swapblk,
883 * we keep the destination's.
885 * This routine is allowed to sleep. It may sleep allocating metadata
886 * indirectly through swp_pager_meta_build() or if paging is still in
887 * progress on the source.
889 * The source object contains no vm_page_t's (which is just as well)
891 * The source object is of type OBJT_SWAP.
893 * The source and destination objects must be locked.
894 * Both object locks may temporarily be released.
897 swap_pager_copy(vm_object_t srcobject, vm_object_t dstobject,
898 vm_pindex_t offset, int destroysource)
902 VM_OBJECT_ASSERT_WLOCKED(srcobject);
903 VM_OBJECT_ASSERT_WLOCKED(dstobject);
906 * If destroysource is set, we remove the source object from the
907 * swap_pager internal queue now.
909 if (destroysource && srcobject->handle != NULL) {
910 vm_object_pip_add(srcobject, 1);
911 VM_OBJECT_WUNLOCK(srcobject);
912 vm_object_pip_add(dstobject, 1);
913 VM_OBJECT_WUNLOCK(dstobject);
914 sx_xlock(&sw_alloc_sx);
915 TAILQ_REMOVE(NOBJLIST(srcobject->handle), srcobject,
917 sx_xunlock(&sw_alloc_sx);
918 VM_OBJECT_WLOCK(dstobject);
919 vm_object_pip_wakeup(dstobject);
920 VM_OBJECT_WLOCK(srcobject);
921 vm_object_pip_wakeup(srcobject);
925 * transfer source to destination.
927 for (i = 0; i < dstobject->size; ++i) {
931 * Locate (without changing) the swapblk on the destination,
932 * unless it is invalid in which case free it silently, or
933 * if the destination is a resident page, in which case the
934 * source is thrown away.
936 dstaddr = swp_pager_meta_ctl(dstobject, i, 0);
938 if (dstaddr == SWAPBLK_NONE) {
940 * Destination has no swapblk and is not resident,
945 srcaddr = swp_pager_meta_ctl(
951 if (srcaddr != SWAPBLK_NONE) {
953 * swp_pager_meta_build() can sleep.
955 vm_object_pip_add(srcobject, 1);
956 VM_OBJECT_WUNLOCK(srcobject);
957 vm_object_pip_add(dstobject, 1);
958 swp_pager_meta_build(dstobject, i, srcaddr);
959 vm_object_pip_wakeup(dstobject);
960 VM_OBJECT_WLOCK(srcobject);
961 vm_object_pip_wakeup(srcobject);
965 * Destination has valid swapblk or it is represented
966 * by a resident page. We destroy the sourceblock.
969 swp_pager_meta_ctl(srcobject, i + offset, SWM_FREE);
974 * Free left over swap blocks in source.
976 * We have to revert the type to OBJT_DEFAULT so we do not accidentally
977 * double-remove the object from the swap queues.
980 swp_pager_meta_free_all(srcobject);
982 * Reverting the type is not necessary, the caller is going
983 * to destroy srcobject directly, but I'm doing it here
984 * for consistency since we've removed the object from its
987 srcobject->type = OBJT_DEFAULT;
992 * SWAP_PAGER_HASPAGE() - determine if we have good backing store for
993 * the requested page.
995 * We determine whether good backing store exists for the requested
996 * page and return TRUE if it does, FALSE if it doesn't.
998 * If TRUE, we also try to determine how much valid, contiguous backing
999 * store exists before and after the requested page.
1002 swap_pager_haspage(vm_object_t object, vm_pindex_t pindex, int *before,
1008 VM_OBJECT_ASSERT_LOCKED(object);
1011 * do we have good backing store at the requested index ?
1013 blk0 = swp_pager_meta_ctl(object, pindex, 0);
1014 if (blk0 == SWAPBLK_NONE) {
1023 * find backwards-looking contiguous good backing store
1025 if (before != NULL) {
1026 for (i = 1; i < SWB_NPAGES; i++) {
1029 blk = swp_pager_meta_ctl(object, pindex - i, 0);
1030 if (blk != blk0 - i)
1037 * find forward-looking contiguous good backing store
1039 if (after != NULL) {
1040 for (i = 1; i < SWB_NPAGES; i++) {
1041 blk = swp_pager_meta_ctl(object, pindex + i, 0);
1042 if (blk != blk0 + i)
1051 * SWAP_PAGER_PAGE_UNSWAPPED() - remove swap backing store related to page
1053 * This removes any associated swap backing store, whether valid or
1054 * not, from the page.
1056 * This routine is typically called when a page is made dirty, at
1057 * which point any associated swap can be freed. MADV_FREE also
1058 * calls us in a special-case situation
1060 * NOTE!!! If the page is clean and the swap was valid, the caller
1061 * should make the page dirty before calling this routine. This routine
1062 * does NOT change the m->dirty status of the page. Also: MADV_FREE
1065 * This routine may not sleep.
1067 * The object containing the page must be locked.
1070 swap_pager_unswapped(vm_page_t m)
1073 swp_pager_meta_ctl(m->object, m->pindex, SWM_FREE);
1077 * swap_pager_getpages() - bring pages in from swap
1079 * Attempt to page in the pages in array "m" of length "count". The caller
1080 * may optionally specify that additional pages preceding and succeeding
1081 * the specified range be paged in. The number of such pages is returned
1082 * in the "rbehind" and "rahead" parameters, and they will be in the
1083 * inactive queue upon return.
1085 * The pages in "m" must be busied and will remain busied upon return.
1088 swap_pager_getpages(vm_object_t object, vm_page_t *m, int count, int *rbehind,
1092 vm_page_t mpred, msucc, p;
1095 int i, j, maxahead, maxbehind, reqcount, shift;
1099 VM_OBJECT_WUNLOCK(object);
1100 bp = getpbuf(&nsw_rcount);
1101 VM_OBJECT_WLOCK(object);
1103 if (!swap_pager_haspage(object, m[0]->pindex, &maxbehind, &maxahead)) {
1104 relpbuf(bp, &nsw_rcount);
1105 return (VM_PAGER_FAIL);
1109 * Clip the readahead and readbehind ranges to exclude resident pages.
1111 if (rahead != NULL) {
1112 KASSERT(reqcount - 1 <= maxahead,
1113 ("page count %d extends beyond swap block", reqcount));
1114 *rahead = imin(*rahead, maxahead - (reqcount - 1));
1115 pindex = m[reqcount - 1]->pindex;
1116 msucc = TAILQ_NEXT(m[reqcount - 1], listq);
1117 if (msucc != NULL && msucc->pindex - pindex - 1 < *rahead)
1118 *rahead = msucc->pindex - pindex - 1;
1120 if (rbehind != NULL) {
1121 *rbehind = imin(*rbehind, maxbehind);
1122 pindex = m[0]->pindex;
1123 mpred = TAILQ_PREV(m[0], pglist, listq);
1124 if (mpred != NULL && pindex - mpred->pindex - 1 < *rbehind)
1125 *rbehind = pindex - mpred->pindex - 1;
1129 * Allocate readahead and readbehind pages.
1131 shift = rbehind != NULL ? *rbehind : 0;
1133 for (i = 1; i <= shift; i++) {
1134 p = vm_page_alloc(object, m[0]->pindex - i,
1137 /* Shift allocated pages to the left. */
1138 for (j = 0; j < i - 1; j++)
1140 bp->b_pages[j + shift - i + 1];
1143 bp->b_pages[shift - i] = p;
1148 for (i = 0; i < reqcount; i++)
1149 bp->b_pages[i + shift] = m[i];
1150 if (rahead != NULL) {
1151 for (i = 0; i < *rahead; i++) {
1152 p = vm_page_alloc(object,
1153 m[reqcount - 1]->pindex + i + 1, VM_ALLOC_NORMAL);
1156 bp->b_pages[shift + reqcount + i] = p;
1160 if (rbehind != NULL)
1165 vm_object_pip_add(object, count);
1167 for (i = 0; i < count; i++)
1168 bp->b_pages[i]->oflags |= VPO_SWAPINPROG;
1170 pindex = bp->b_pages[0]->pindex;
1171 blk = swp_pager_meta_ctl(object, pindex, 0);
1172 KASSERT(blk != SWAPBLK_NONE,
1173 ("no swap blocking containing %p(%jx)", object, (uintmax_t)pindex));
1175 VM_OBJECT_WUNLOCK(object);
1177 bp->b_flags |= B_PAGING;
1178 bp->b_iocmd = BIO_READ;
1179 bp->b_iodone = swp_pager_async_iodone;
1180 bp->b_rcred = crhold(thread0.td_ucred);
1181 bp->b_wcred = crhold(thread0.td_ucred);
1183 bp->b_bcount = PAGE_SIZE * count;
1184 bp->b_bufsize = PAGE_SIZE * count;
1185 bp->b_npages = count;
1186 bp->b_pgbefore = rbehind != NULL ? *rbehind : 0;
1187 bp->b_pgafter = rahead != NULL ? *rahead : 0;
1189 VM_CNT_INC(v_swapin);
1190 VM_CNT_ADD(v_swappgsin, count);
1193 * perform the I/O. NOTE!!! bp cannot be considered valid after
1194 * this point because we automatically release it on completion.
1195 * Instead, we look at the one page we are interested in which we
1196 * still hold a lock on even through the I/O completion.
1198 * The other pages in our m[] array are also released on completion,
1199 * so we cannot assume they are valid anymore either.
1201 * NOTE: b_blkno is destroyed by the call to swapdev_strategy
1204 swp_pager_strategy(bp);
1207 * Wait for the pages we want to complete. VPO_SWAPINPROG is always
1208 * cleared on completion. If an I/O error occurs, SWAPBLK_NONE
1209 * is set in the metadata for each page in the request.
1211 VM_OBJECT_WLOCK(object);
1212 while ((m[0]->oflags & VPO_SWAPINPROG) != 0) {
1213 m[0]->oflags |= VPO_SWAPSLEEP;
1214 VM_CNT_INC(v_intrans);
1215 if (VM_OBJECT_SLEEP(object, &object->paging_in_progress, PSWP,
1216 "swread", hz * 20)) {
1218 "swap_pager: indefinite wait buffer: bufobj: %p, blkno: %jd, size: %ld\n",
1219 bp->b_bufobj, (intmax_t)bp->b_blkno, bp->b_bcount);
1224 * If we had an unrecoverable read error pages will not be valid.
1226 for (i = 0; i < reqcount; i++)
1227 if (m[i]->valid != VM_PAGE_BITS_ALL)
1228 return (VM_PAGER_ERROR);
1230 return (VM_PAGER_OK);
1233 * A final note: in a low swap situation, we cannot deallocate swap
1234 * and mark a page dirty here because the caller is likely to mark
1235 * the page clean when we return, causing the page to possibly revert
1236 * to all-zero's later.
1241 * swap_pager_getpages_async():
1243 * Right now this is emulation of asynchronous operation on top of
1244 * swap_pager_getpages().
1247 swap_pager_getpages_async(vm_object_t object, vm_page_t *m, int count,
1248 int *rbehind, int *rahead, pgo_getpages_iodone_t iodone, void *arg)
1252 r = swap_pager_getpages(object, m, count, rbehind, rahead);
1253 VM_OBJECT_WUNLOCK(object);
1258 case VM_PAGER_ERROR:
1265 panic("unhandled swap_pager_getpages() error %d", r);
1267 (iodone)(arg, m, count, error);
1268 VM_OBJECT_WLOCK(object);
1274 * swap_pager_putpages:
1276 * Assign swap (if necessary) and initiate I/O on the specified pages.
1278 * We support both OBJT_DEFAULT and OBJT_SWAP objects. DEFAULT objects
1279 * are automatically converted to SWAP objects.
1281 * In a low memory situation we may block in VOP_STRATEGY(), but the new
1282 * vm_page reservation system coupled with properly written VFS devices
1283 * should ensure that no low-memory deadlock occurs. This is an area
1286 * The parent has N vm_object_pip_add() references prior to
1287 * calling us and will remove references for rtvals[] that are
1288 * not set to VM_PAGER_PEND. We need to remove the rest on I/O
1291 * The parent has soft-busy'd the pages it passes us and will unbusy
1292 * those whos rtvals[] entry is not set to VM_PAGER_PEND on return.
1293 * We need to unbusy the rest on I/O completion.
1296 swap_pager_putpages(vm_object_t object, vm_page_t *m, int count,
1297 int flags, int *rtvals)
1302 if (count && m[0]->object != object) {
1303 panic("swap_pager_putpages: object mismatch %p/%p",
1312 * Turn object into OBJT_SWAP
1313 * check for bogus sysops
1314 * force sync if not pageout process
1316 if (object->type != OBJT_SWAP)
1317 swp_pager_meta_build(object, 0, SWAPBLK_NONE);
1318 VM_OBJECT_WUNLOCK(object);
1321 if (curproc != pageproc)
1324 sync = (flags & VM_PAGER_PUT_SYNC) != 0;
1329 * Assign swap blocks and issue I/O. We reallocate swap on the fly.
1330 * The page is left dirty until the pageout operation completes
1333 for (i = 0; i < count; i += n) {
1339 * Maximum I/O size is limited by a number of factors.
1341 n = min(BLIST_MAX_ALLOC, count - i);
1342 n = min(n, nsw_cluster_max);
1345 * Get biggest block of swap we can. If we fail, fall
1346 * back and try to allocate a smaller block. Don't go
1347 * overboard trying to allocate space if it would overly
1351 (blk = swp_pager_getswapspace(n)) == SWAPBLK_NONE &&
1356 if (blk == SWAPBLK_NONE) {
1357 for (j = 0; j < n; ++j)
1358 rtvals[i+j] = VM_PAGER_FAIL;
1363 * All I/O parameters have been satisfied, build the I/O
1364 * request and assign the swap space.
1367 bp = getpbuf(&nsw_wcount_sync);
1369 bp = getpbuf(&nsw_wcount_async);
1370 bp->b_flags = B_ASYNC;
1372 bp->b_flags |= B_PAGING;
1373 bp->b_iocmd = BIO_WRITE;
1375 bp->b_rcred = crhold(thread0.td_ucred);
1376 bp->b_wcred = crhold(thread0.td_ucred);
1377 bp->b_bcount = PAGE_SIZE * n;
1378 bp->b_bufsize = PAGE_SIZE * n;
1381 VM_OBJECT_WLOCK(object);
1382 for (j = 0; j < n; ++j) {
1383 vm_page_t mreq = m[i+j];
1385 swp_pager_meta_build(
1390 MPASS(mreq->dirty == VM_PAGE_BITS_ALL);
1391 mreq->oflags |= VPO_SWAPINPROG;
1392 bp->b_pages[j] = mreq;
1394 VM_OBJECT_WUNLOCK(object);
1397 * Must set dirty range for NFS to work.
1400 bp->b_dirtyend = bp->b_bcount;
1402 VM_CNT_INC(v_swapout);
1403 VM_CNT_ADD(v_swappgsout, bp->b_npages);
1406 * We unconditionally set rtvals[] to VM_PAGER_PEND so that we
1407 * can call the async completion routine at the end of a
1408 * synchronous I/O operation. Otherwise, our caller would
1409 * perform duplicate unbusy and wakeup operations on the page
1410 * and object, respectively.
1412 for (j = 0; j < n; j++)
1413 rtvals[i + j] = VM_PAGER_PEND;
1418 * NOTE: b_blkno is destroyed by the call to swapdev_strategy
1420 if (sync == FALSE) {
1421 bp->b_iodone = swp_pager_async_iodone;
1423 swp_pager_strategy(bp);
1430 * NOTE: b_blkno is destroyed by the call to swapdev_strategy
1432 bp->b_iodone = bdone;
1433 swp_pager_strategy(bp);
1436 * Wait for the sync I/O to complete.
1438 bwait(bp, PVM, "swwrt");
1441 * Now that we are through with the bp, we can call the
1442 * normal async completion, which frees everything up.
1444 swp_pager_async_iodone(bp);
1446 VM_OBJECT_WLOCK(object);
1450 * swp_pager_async_iodone:
1452 * Completion routine for asynchronous reads and writes from/to swap.
1453 * Also called manually by synchronous code to finish up a bp.
1455 * This routine may not sleep.
1458 swp_pager_async_iodone(struct buf *bp)
1461 vm_object_t object = NULL;
1466 if (bp->b_ioflags & BIO_ERROR) {
1468 "swap_pager: I/O error - %s failed; blkno %ld,"
1469 "size %ld, error %d\n",
1470 ((bp->b_iocmd == BIO_READ) ? "pagein" : "pageout"),
1478 * remove the mapping for kernel virtual
1481 pmap_qremove((vm_offset_t)bp->b_data, bp->b_npages);
1483 bp->b_data = bp->b_kvabase;
1486 object = bp->b_pages[0]->object;
1487 VM_OBJECT_WLOCK(object);
1491 * cleanup pages. If an error occurs writing to swap, we are in
1492 * very serious trouble. If it happens to be a disk error, though,
1493 * we may be able to recover by reassigning the swap later on. So
1494 * in this case we remove the m->swapblk assignment for the page
1495 * but do not free it in the rlist. The errornous block(s) are thus
1496 * never reallocated as swap. Redirty the page and continue.
1498 for (i = 0; i < bp->b_npages; ++i) {
1499 vm_page_t m = bp->b_pages[i];
1501 m->oflags &= ~VPO_SWAPINPROG;
1502 if (m->oflags & VPO_SWAPSLEEP) {
1503 m->oflags &= ~VPO_SWAPSLEEP;
1504 wakeup(&object->paging_in_progress);
1507 if (bp->b_ioflags & BIO_ERROR) {
1509 * If an error occurs I'd love to throw the swapblk
1510 * away without freeing it back to swapspace, so it
1511 * can never be used again. But I can't from an
1514 if (bp->b_iocmd == BIO_READ) {
1516 * NOTE: for reads, m->dirty will probably
1517 * be overridden by the original caller of
1518 * getpages so don't play cute tricks here.
1523 * If a write error occurs, reactivate page
1524 * so it doesn't clog the inactive list,
1525 * then finish the I/O.
1529 vm_page_activate(m);
1533 } else if (bp->b_iocmd == BIO_READ) {
1535 * NOTE: for reads, m->dirty will probably be
1536 * overridden by the original caller of getpages so
1537 * we cannot set them in order to free the underlying
1538 * swap in a low-swap situation. I don't think we'd
1539 * want to do that anyway, but it was an optimization
1540 * that existed in the old swapper for a time before
1541 * it got ripped out due to precisely this problem.
1543 KASSERT(!pmap_page_is_mapped(m),
1544 ("swp_pager_async_iodone: page %p is mapped", m));
1545 KASSERT(m->dirty == 0,
1546 ("swp_pager_async_iodone: page %p is dirty", m));
1548 m->valid = VM_PAGE_BITS_ALL;
1549 if (i < bp->b_pgbefore ||
1550 i >= bp->b_npages - bp->b_pgafter)
1551 vm_page_readahead_finish(m);
1554 * For write success, clear the dirty
1555 * status, then finish the I/O ( which decrements the
1556 * busy count and possibly wakes waiter's up ).
1557 * A page is only written to swap after a period of
1558 * inactivity. Therefore, we do not expect it to be
1561 KASSERT(!pmap_page_is_write_mapped(m),
1562 ("swp_pager_async_iodone: page %p is not write"
1566 vm_page_deactivate_noreuse(m);
1573 * adjust pip. NOTE: the original parent may still have its own
1574 * pip refs on the object.
1576 if (object != NULL) {
1577 vm_object_pip_wakeupn(object, bp->b_npages);
1578 VM_OBJECT_WUNLOCK(object);
1582 * swapdev_strategy() manually sets b_vp and b_bufobj before calling
1583 * bstrategy(). Set them back to NULL now we're done with it, or we'll
1584 * trigger a KASSERT in relpbuf().
1588 bp->b_bufobj = NULL;
1591 * release the physical I/O buffer
1595 ((bp->b_iocmd == BIO_READ) ? &nsw_rcount :
1596 ((bp->b_flags & B_ASYNC) ?
1605 swap_pager_nswapdev(void)
1612 * SWP_PAGER_FORCE_PAGEIN() - force a swap block to be paged in
1614 * This routine dissociates the page at the given index within an object
1615 * from its backing store, paging it in if it does not reside in memory.
1616 * If the page is paged in, it is marked dirty and placed in the laundry
1617 * queue. The page is marked dirty because it no longer has backing
1618 * store. It is placed in the laundry queue because it has not been
1619 * accessed recently. Otherwise, it would already reside in memory.
1621 * We also attempt to swap in all other pages in the swap block.
1622 * However, we only guarantee that the one at the specified index is
1625 * XXX - The code to page the whole block in doesn't work, so we
1626 * revert to the one-by-one behavior for now. Sigh.
1629 swp_pager_force_pagein(vm_object_t object, vm_pindex_t pindex)
1633 vm_object_pip_add(object, 1);
1634 m = vm_page_grab(object, pindex, VM_ALLOC_NORMAL);
1635 if (m->valid == VM_PAGE_BITS_ALL) {
1636 vm_object_pip_wakeup(object);
1639 vm_page_activate(m);
1642 vm_pager_page_unswapped(m);
1646 if (swap_pager_getpages(object, &m, 1, NULL, NULL) != VM_PAGER_OK)
1647 panic("swap_pager_force_pagein: read from swap failed");/*XXX*/
1648 vm_object_pip_wakeup(object);
1654 vm_pager_page_unswapped(m);
1658 * swap_pager_swapoff:
1660 * Page in all of the pages that have been paged out to the
1661 * given device. The corresponding blocks in the bitmap must be
1662 * marked as allocated and the device must be flagged SW_CLOSING.
1663 * There may be no processes swapped out to the device.
1665 * This routine may block.
1668 swap_pager_swapoff(struct swdevt *sp)
1675 sx_assert(&swdev_syscall_lock, SA_XLOCKED);
1679 mtx_lock(&vm_object_list_mtx);
1680 TAILQ_FOREACH(object, &vm_object_list, object_list) {
1681 if (object->type != OBJT_SWAP)
1683 mtx_unlock(&vm_object_list_mtx);
1684 /* Depends on type-stability. */
1685 VM_OBJECT_WLOCK(object);
1688 * Dead objects are eventually terminated on their own.
1690 if ((object->flags & OBJ_DEAD) != 0)
1694 * Sync with fences placed after pctrie
1695 * initialization. We must not access pctrie below
1696 * unless we checked that our object is swap and not
1699 atomic_thread_fence_acq();
1700 if (object->type != OBJT_SWAP)
1703 for (pi = 0; (sb = SWAP_PCTRIE_LOOKUP_GE(
1704 &object->un_pager.swp.swp_blks, pi)) != NULL; ) {
1705 pi = sb->p + SWAP_META_PAGES;
1706 for (i = 0; i < SWAP_META_PAGES; i++) {
1707 if (sb->d[i] == SWAPBLK_NONE)
1709 if (swp_pager_isondev(sb->d[i], sp))
1710 swp_pager_force_pagein(object,
1715 VM_OBJECT_WUNLOCK(object);
1716 mtx_lock(&vm_object_list_mtx);
1718 mtx_unlock(&vm_object_list_mtx);
1722 * Objects may be locked or paging to the device being
1723 * removed, so we will miss their pages and need to
1724 * make another pass. We have marked this device as
1725 * SW_CLOSING, so the activity should finish soon.
1728 if (retries > 100) {
1729 panic("swapoff: failed to locate %d swap blocks",
1732 pause("swpoff", hz / 20);
1735 EVENTHANDLER_INVOKE(swapoff, sp);
1738 /************************************************************************
1740 ************************************************************************
1742 * These routines manipulate the swap metadata stored in the
1745 * Swap metadata is implemented with a global hash and not directly
1746 * linked into the object. Instead the object simply contains
1747 * appropriate tracking counters.
1751 * SWP_PAGER_META_BUILD() - add swap block to swap meta data for object
1753 * We first convert the object to a swap object if it is a default
1756 * The specified swapblk is added to the object's swap metadata. If
1757 * the swapblk is not valid, it is freed instead. Any previously
1758 * assigned swapblk is freed.
1761 swp_pager_meta_build(vm_object_t object, vm_pindex_t pindex, daddr_t swapblk)
1763 static volatile int swblk_zone_exhausted, swpctrie_zone_exhausted;
1764 struct swblk *sb, *sb1;
1765 vm_pindex_t modpi, rdpi;
1768 VM_OBJECT_ASSERT_WLOCKED(object);
1771 * Convert default object to swap object if necessary
1773 if (object->type != OBJT_SWAP) {
1774 pctrie_init(&object->un_pager.swp.swp_blks);
1777 * Ensure that swap_pager_swapoff()'s iteration over
1778 * object_list does not see a garbage pctrie.
1780 atomic_thread_fence_rel();
1782 object->type = OBJT_SWAP;
1783 KASSERT(object->handle == NULL, ("default pager with handle"));
1786 rdpi = rounddown(pindex, SWAP_META_PAGES);
1787 sb = SWAP_PCTRIE_LOOKUP(&object->un_pager.swp.swp_blks, rdpi);
1789 if (swapblk == SWAPBLK_NONE)
1792 sb = uma_zalloc(swblk_zone, M_NOWAIT | (curproc ==
1793 pageproc ? M_USE_RESERVE : 0));
1796 for (i = 0; i < SWAP_META_PAGES; i++)
1797 sb->d[i] = SWAPBLK_NONE;
1798 if (atomic_cmpset_int(&swblk_zone_exhausted,
1800 printf("swblk zone ok\n");
1803 VM_OBJECT_WUNLOCK(object);
1804 if (uma_zone_exhausted(swblk_zone)) {
1805 if (atomic_cmpset_int(&swblk_zone_exhausted,
1807 printf("swap blk zone exhausted, "
1808 "increase kern.maxswzone\n");
1809 vm_pageout_oom(VM_OOM_SWAPZ);
1810 pause("swzonxb", 10);
1813 VM_OBJECT_WLOCK(object);
1814 sb = SWAP_PCTRIE_LOOKUP(&object->un_pager.swp.swp_blks,
1818 * Somebody swapped out a nearby page,
1819 * allocating swblk at the rdpi index,
1820 * while we dropped the object lock.
1825 error = SWAP_PCTRIE_INSERT(
1826 &object->un_pager.swp.swp_blks, sb);
1828 if (atomic_cmpset_int(&swpctrie_zone_exhausted,
1830 printf("swpctrie zone ok\n");
1833 VM_OBJECT_WUNLOCK(object);
1834 if (uma_zone_exhausted(swpctrie_zone)) {
1835 if (atomic_cmpset_int(&swpctrie_zone_exhausted,
1837 printf("swap pctrie zone exhausted, "
1838 "increase kern.maxswzone\n");
1839 vm_pageout_oom(VM_OOM_SWAPZ);
1840 pause("swzonxp", 10);
1843 VM_OBJECT_WLOCK(object);
1844 sb1 = SWAP_PCTRIE_LOOKUP(&object->un_pager.swp.swp_blks,
1847 uma_zfree(swblk_zone, sb);
1854 MPASS(sb->p == rdpi);
1856 modpi = pindex % SWAP_META_PAGES;
1857 /* Delete prior contents of metadata. */
1858 if (sb->d[modpi] != SWAPBLK_NONE)
1859 swp_pager_freeswapspace(sb->d[modpi], 1);
1860 /* Enter block into metadata. */
1861 sb->d[modpi] = swapblk;
1864 * Free the swblk if we end up with the empty page run.
1866 if (swapblk == SWAPBLK_NONE) {
1867 for (i = 0; i < SWAP_META_PAGES; i++) {
1868 if (sb->d[i] != SWAPBLK_NONE)
1871 if (i == SWAP_META_PAGES) {
1872 SWAP_PCTRIE_REMOVE(&object->un_pager.swp.swp_blks,
1874 uma_zfree(swblk_zone, sb);
1880 * SWP_PAGER_META_FREE() - free a range of blocks in the object's swap metadata
1882 * The requested range of blocks is freed, with any associated swap
1883 * returned to the swap bitmap.
1885 * This routine will free swap metadata structures as they are cleaned
1886 * out. This routine does *NOT* operate on swap metadata associated
1887 * with resident pages.
1890 swp_pager_meta_free(vm_object_t object, vm_pindex_t pindex, vm_pindex_t count)
1897 VM_OBJECT_ASSERT_WLOCKED(object);
1898 if (object->type != OBJT_SWAP || count == 0)
1901 last = pindex + count - 1;
1903 sb = SWAP_PCTRIE_LOOKUP_GE(&object->un_pager.swp.swp_blks,
1904 rounddown(pindex, SWAP_META_PAGES));
1905 if (sb == NULL || sb->p > last)
1908 for (i = 0; i < SWAP_META_PAGES; i++) {
1909 if (sb->d[i] == SWAPBLK_NONE)
1911 if (pindex <= sb->p + i && sb->p + i <= last) {
1912 swp_pager_freeswapspace(sb->d[i], 1);
1913 sb->d[i] = SWAPBLK_NONE;
1917 pindex = sb->p + SWAP_META_PAGES;
1919 SWAP_PCTRIE_REMOVE(&object->un_pager.swp.swp_blks,
1921 uma_zfree(swblk_zone, sb);
1927 * SWP_PAGER_META_FREE_ALL() - destroy all swap metadata associated with object
1929 * This routine locates and destroys all swap metadata associated with
1933 swp_pager_meta_free_all(vm_object_t object)
1939 VM_OBJECT_ASSERT_WLOCKED(object);
1940 if (object->type != OBJT_SWAP)
1943 for (pindex = 0; (sb = SWAP_PCTRIE_LOOKUP_GE(
1944 &object->un_pager.swp.swp_blks, pindex)) != NULL;) {
1945 pindex = sb->p + SWAP_META_PAGES;
1946 for (i = 0; i < SWAP_META_PAGES; i++) {
1947 if (sb->d[i] != SWAPBLK_NONE)
1948 swp_pager_freeswapspace(sb->d[i], 1);
1950 SWAP_PCTRIE_REMOVE(&object->un_pager.swp.swp_blks, sb->p);
1951 uma_zfree(swblk_zone, sb);
1956 * SWP_PAGER_METACTL() - misc control of swap and vm_page_t meta data.
1958 * This routine is capable of looking up, popping, or freeing
1959 * swapblk assignments in the swap meta data or in the vm_page_t.
1960 * The routine typically returns the swapblk being looked-up, or popped,
1961 * or SWAPBLK_NONE if the block was freed, or SWAPBLK_NONE if the block
1962 * was invalid. This routine will automatically free any invalid
1963 * meta-data swapblks.
1965 * When acting on a busy resident page and paging is in progress, we
1966 * have to wait until paging is complete but otherwise can act on the
1969 * SWM_FREE remove and free swap block from metadata
1970 * SWM_POP remove from meta data but do not free.. pop it out
1973 swp_pager_meta_ctl(vm_object_t object, vm_pindex_t pindex, int flags)
1979 if ((flags & (SWM_FREE | SWM_POP)) != 0)
1980 VM_OBJECT_ASSERT_WLOCKED(object);
1982 VM_OBJECT_ASSERT_LOCKED(object);
1985 * The meta data only exists if the object is OBJT_SWAP
1986 * and even then might not be allocated yet.
1988 if (object->type != OBJT_SWAP)
1989 return (SWAPBLK_NONE);
1991 sb = SWAP_PCTRIE_LOOKUP(&object->un_pager.swp.swp_blks,
1992 rounddown(pindex, SWAP_META_PAGES));
1994 return (SWAPBLK_NONE);
1995 r1 = sb->d[pindex % SWAP_META_PAGES];
1996 if (r1 == SWAPBLK_NONE)
1997 return (SWAPBLK_NONE);
1998 if ((flags & (SWM_FREE | SWM_POP)) != 0) {
1999 sb->d[pindex % SWAP_META_PAGES] = SWAPBLK_NONE;
2000 for (i = 0; i < SWAP_META_PAGES; i++) {
2001 if (sb->d[i] != SWAPBLK_NONE)
2004 if (i == SWAP_META_PAGES) {
2005 SWAP_PCTRIE_REMOVE(&object->un_pager.swp.swp_blks,
2006 rounddown(pindex, SWAP_META_PAGES));
2007 uma_zfree(swblk_zone, sb);
2010 if ((flags & SWM_FREE) != 0) {
2011 swp_pager_freeswapspace(r1, 1);
2018 * Returns the least page index which is greater than or equal to the
2019 * parameter pindex and for which there is a swap block allocated.
2020 * Returns object's size if the object's type is not swap or if there
2021 * are no allocated swap blocks for the object after the requested
2025 swap_pager_find_least(vm_object_t object, vm_pindex_t pindex)
2030 VM_OBJECT_ASSERT_LOCKED(object);
2031 if (object->type != OBJT_SWAP)
2032 return (object->size);
2034 sb = SWAP_PCTRIE_LOOKUP_GE(&object->un_pager.swp.swp_blks,
2035 rounddown(pindex, SWAP_META_PAGES));
2037 return (object->size);
2038 if (sb->p < pindex) {
2039 for (i = pindex % SWAP_META_PAGES; i < SWAP_META_PAGES; i++) {
2040 if (sb->d[i] != SWAPBLK_NONE)
2043 sb = SWAP_PCTRIE_LOOKUP_GE(&object->un_pager.swp.swp_blks,
2044 roundup(pindex, SWAP_META_PAGES));
2046 return (object->size);
2048 for (i = 0; i < SWAP_META_PAGES; i++) {
2049 if (sb->d[i] != SWAPBLK_NONE)
2054 * We get here if a swblk is present in the trie but it
2055 * doesn't map any blocks.
2058 return (object->size);
2062 * System call swapon(name) enables swapping on device name,
2063 * which must be in the swdevsw. Return EBUSY
2064 * if already swapping on this device.
2066 #ifndef _SYS_SYSPROTO_H_
2067 struct swapon_args {
2077 sys_swapon(struct thread *td, struct swapon_args *uap)
2081 struct nameidata nd;
2084 error = priv_check(td, PRIV_SWAPON);
2088 sx_xlock(&swdev_syscall_lock);
2091 * Swap metadata may not fit in the KVM if we have physical
2094 if (swblk_zone == NULL) {
2099 NDINIT(&nd, LOOKUP, ISOPEN | FOLLOW | AUDITVNODE1, UIO_USERSPACE,
2105 NDFREE(&nd, NDF_ONLY_PNBUF);
2108 if (vn_isdisk(vp, &error)) {
2109 error = swapongeom(vp);
2110 } else if (vp->v_type == VREG &&
2111 (vp->v_mount->mnt_vfc->vfc_flags & VFCF_NETWORK) != 0 &&
2112 (error = VOP_GETATTR(vp, &attr, td->td_ucred)) == 0) {
2114 * Allow direct swapping to NFS regular files in the same
2115 * way that nfs_mountroot() sets up diskless swapping.
2117 error = swaponvp(td, vp, attr.va_size / DEV_BSIZE);
2123 sx_xunlock(&swdev_syscall_lock);
2128 * Check that the total amount of swap currently configured does not
2129 * exceed half the theoretical maximum. If it does, print a warning
2133 swapon_check_swzone(void)
2135 unsigned long maxpages, npages;
2137 npages = swap_total / PAGE_SIZE;
2138 /* absolute maximum we can handle assuming 100% efficiency */
2139 maxpages = uma_zone_get_max(swblk_zone) * SWAP_META_PAGES;
2141 /* recommend using no more than half that amount */
2142 if (npages > maxpages / 2) {
2143 printf("warning: total configured swap (%lu pages) "
2144 "exceeds maximum recommended amount (%lu pages).\n",
2145 npages, maxpages / 2);
2146 printf("warning: increase kern.maxswzone "
2147 "or reduce amount of swap.\n");
2152 swaponsomething(struct vnode *vp, void *id, u_long nblks,
2153 sw_strategy_t *strategy, sw_close_t *close, dev_t dev, int flags)
2155 struct swdevt *sp, *tsp;
2160 * nblks is in DEV_BSIZE'd chunks, convert to PAGE_SIZE'd chunks.
2161 * First chop nblks off to page-align it, then convert.
2163 * sw->sw_nblks is in page-sized chunks now too.
2165 nblks &= ~(ctodb(1) - 1);
2166 nblks = dbtoc(nblks);
2169 * If we go beyond this, we get overflows in the radix
2172 mblocks = 0x40000000 / BLIST_META_RADIX;
2173 if (nblks > mblocks) {
2175 "WARNING: reducing swap size to maximum of %luMB per unit\n",
2176 mblocks / 1024 / 1024 * PAGE_SIZE);
2180 sp = malloc(sizeof *sp, M_VMPGDATA, M_WAITOK | M_ZERO);
2185 sp->sw_nblks = nblks;
2187 sp->sw_strategy = strategy;
2188 sp->sw_close = close;
2189 sp->sw_flags = flags;
2191 sp->sw_blist = blist_create(nblks, M_WAITOK);
2193 * Do not free the first two block in order to avoid overwriting
2194 * any bsd label at the front of the partition
2196 blist_free(sp->sw_blist, 2, nblks - 2);
2199 mtx_lock(&sw_dev_mtx);
2200 TAILQ_FOREACH(tsp, &swtailq, sw_list) {
2201 if (tsp->sw_end >= dvbase) {
2203 * We put one uncovered page between the devices
2204 * in order to definitively prevent any cross-device
2207 dvbase = tsp->sw_end + 1;
2210 sp->sw_first = dvbase;
2211 sp->sw_end = dvbase + nblks;
2212 TAILQ_INSERT_TAIL(&swtailq, sp, sw_list);
2214 swap_pager_avail += nblks - 2;
2215 swap_total += (vm_ooffset_t)nblks * PAGE_SIZE;
2216 swapon_check_swzone();
2218 mtx_unlock(&sw_dev_mtx);
2219 EVENTHANDLER_INVOKE(swapon, sp);
2223 * SYSCALL: swapoff(devname)
2225 * Disable swapping on the given device.
2227 * XXX: Badly designed system call: it should use a device index
2228 * rather than filename as specification. We keep sw_vp around
2229 * only to make this work.
2231 #ifndef _SYS_SYSPROTO_H_
2232 struct swapoff_args {
2242 sys_swapoff(struct thread *td, struct swapoff_args *uap)
2245 struct nameidata nd;
2249 error = priv_check(td, PRIV_SWAPOFF);
2253 sx_xlock(&swdev_syscall_lock);
2255 NDINIT(&nd, LOOKUP, FOLLOW | AUDITVNODE1, UIO_USERSPACE, uap->name,
2260 NDFREE(&nd, NDF_ONLY_PNBUF);
2263 mtx_lock(&sw_dev_mtx);
2264 TAILQ_FOREACH(sp, &swtailq, sw_list) {
2265 if (sp->sw_vp == vp)
2268 mtx_unlock(&sw_dev_mtx);
2273 error = swapoff_one(sp, td->td_ucred);
2275 sx_xunlock(&swdev_syscall_lock);
2280 swapoff_one(struct swdevt *sp, struct ucred *cred)
2287 sx_assert(&swdev_syscall_lock, SA_XLOCKED);
2289 (void) vn_lock(sp->sw_vp, LK_EXCLUSIVE | LK_RETRY);
2290 error = mac_system_check_swapoff(cred, sp->sw_vp);
2291 (void) VOP_UNLOCK(sp->sw_vp, 0);
2295 nblks = sp->sw_nblks;
2298 * We can turn off this swap device safely only if the
2299 * available virtual memory in the system will fit the amount
2300 * of data we will have to page back in, plus an epsilon so
2301 * the system doesn't become critically low on swap space.
2303 if (vm_cnt.v_free_count + swap_pager_avail < nblks + nswap_lowat)
2307 * Prevent further allocations on this device.
2309 mtx_lock(&sw_dev_mtx);
2310 sp->sw_flags |= SW_CLOSING;
2311 swap_pager_avail -= blist_fill(sp->sw_blist, 0, nblks);
2312 swap_total -= (vm_ooffset_t)nblks * PAGE_SIZE;
2313 mtx_unlock(&sw_dev_mtx);
2316 * Page in the contents of the device and close it.
2318 swap_pager_swapoff(sp);
2320 sp->sw_close(curthread, sp);
2321 mtx_lock(&sw_dev_mtx);
2323 TAILQ_REMOVE(&swtailq, sp, sw_list);
2325 if (nswapdev == 0) {
2326 swap_pager_full = 2;
2327 swap_pager_almost_full = 1;
2331 mtx_unlock(&sw_dev_mtx);
2332 blist_destroy(sp->sw_blist);
2333 free(sp, M_VMPGDATA);
2340 struct swdevt *sp, *spt;
2341 const char *devname;
2344 sx_xlock(&swdev_syscall_lock);
2346 mtx_lock(&sw_dev_mtx);
2347 TAILQ_FOREACH_SAFE(sp, &swtailq, sw_list, spt) {
2348 mtx_unlock(&sw_dev_mtx);
2349 if (vn_isdisk(sp->sw_vp, NULL))
2350 devname = devtoname(sp->sw_vp->v_rdev);
2353 error = swapoff_one(sp, thread0.td_ucred);
2355 printf("Cannot remove swap device %s (error=%d), "
2356 "skipping.\n", devname, error);
2357 } else if (bootverbose) {
2358 printf("Swap device %s removed.\n", devname);
2360 mtx_lock(&sw_dev_mtx);
2362 mtx_unlock(&sw_dev_mtx);
2364 sx_xunlock(&swdev_syscall_lock);
2368 swap_pager_status(int *total, int *used)
2374 mtx_lock(&sw_dev_mtx);
2375 TAILQ_FOREACH(sp, &swtailq, sw_list) {
2376 *total += sp->sw_nblks;
2377 *used += sp->sw_used;
2379 mtx_unlock(&sw_dev_mtx);
2383 swap_dev_info(int name, struct xswdev *xs, char *devname, size_t len)
2386 const char *tmp_devname;
2391 mtx_lock(&sw_dev_mtx);
2392 TAILQ_FOREACH(sp, &swtailq, sw_list) {
2397 xs->xsw_version = XSWDEV_VERSION;
2398 xs->xsw_dev = sp->sw_dev;
2399 xs->xsw_flags = sp->sw_flags;
2400 xs->xsw_nblks = sp->sw_nblks;
2401 xs->xsw_used = sp->sw_used;
2402 if (devname != NULL) {
2403 if (vn_isdisk(sp->sw_vp, NULL))
2404 tmp_devname = devtoname(sp->sw_vp->v_rdev);
2406 tmp_devname = "[file]";
2407 strncpy(devname, tmp_devname, len);
2412 mtx_unlock(&sw_dev_mtx);
2416 #if defined(COMPAT_FREEBSD11)
2417 #define XSWDEV_VERSION_11 1
2428 sysctl_vm_swap_info(SYSCTL_HANDLER_ARGS)
2431 #if defined(COMPAT_FREEBSD11)
2432 struct xswdev11 xs11;
2436 if (arg2 != 1) /* name length */
2438 error = swap_dev_info(*(int *)arg1, &xs, NULL, 0);
2441 #if defined(COMPAT_FREEBSD11)
2442 if (req->oldlen == sizeof(xs11)) {
2443 xs11.xsw_version = XSWDEV_VERSION_11;
2444 xs11.xsw_dev = xs.xsw_dev; /* truncation */
2445 xs11.xsw_flags = xs.xsw_flags;
2446 xs11.xsw_nblks = xs.xsw_nblks;
2447 xs11.xsw_used = xs.xsw_used;
2448 error = SYSCTL_OUT(req, &xs11, sizeof(xs11));
2451 error = SYSCTL_OUT(req, &xs, sizeof(xs));
2455 SYSCTL_INT(_vm, OID_AUTO, nswapdev, CTLFLAG_RD, &nswapdev, 0,
2456 "Number of swap devices");
2457 SYSCTL_NODE(_vm, OID_AUTO, swap_info, CTLFLAG_RD | CTLFLAG_MPSAFE,
2458 sysctl_vm_swap_info,
2459 "Swap statistics by device");
2462 * Count the approximate swap usage in pages for a vmspace. The
2463 * shadowed or not yet copied on write swap blocks are not accounted.
2464 * The map must be locked.
2467 vmspace_swap_count(struct vmspace *vmspace)
2477 map = &vmspace->vm_map;
2480 for (cur = map->header.next; cur != &map->header; cur = cur->next) {
2481 if ((cur->eflags & MAP_ENTRY_IS_SUB_MAP) != 0)
2483 object = cur->object.vm_object;
2484 if (object == NULL || object->type != OBJT_SWAP)
2486 VM_OBJECT_RLOCK(object);
2487 if (object->type != OBJT_SWAP)
2489 pi = OFF_TO_IDX(cur->offset);
2490 e = pi + OFF_TO_IDX(cur->end - cur->start);
2491 for (;; pi = sb->p + SWAP_META_PAGES) {
2492 sb = SWAP_PCTRIE_LOOKUP_GE(
2493 &object->un_pager.swp.swp_blks, pi);
2494 if (sb == NULL || sb->p >= e)
2496 for (i = 0; i < SWAP_META_PAGES; i++) {
2497 if (sb->p + i < e &&
2498 sb->d[i] != SWAPBLK_NONE)
2503 VM_OBJECT_RUNLOCK(object);
2511 * Swapping onto disk devices.
2515 static g_orphan_t swapgeom_orphan;
2517 static struct g_class g_swap_class = {
2519 .version = G_VERSION,
2520 .orphan = swapgeom_orphan,
2523 DECLARE_GEOM_CLASS(g_swap_class, g_class);
2527 swapgeom_close_ev(void *arg, int flags)
2529 struct g_consumer *cp;
2532 g_access(cp, -1, -1, 0);
2534 g_destroy_consumer(cp);
2538 * Add a reference to the g_consumer for an inflight transaction.
2541 swapgeom_acquire(struct g_consumer *cp)
2544 mtx_assert(&sw_dev_mtx, MA_OWNED);
2549 * Remove a reference from the g_consumer. Post a close event if all
2550 * references go away, since the function might be called from the
2554 swapgeom_release(struct g_consumer *cp, struct swdevt *sp)
2557 mtx_assert(&sw_dev_mtx, MA_OWNED);
2559 if (cp->index == 0) {
2560 if (g_post_event(swapgeom_close_ev, cp, M_NOWAIT, NULL) == 0)
2566 swapgeom_done(struct bio *bp2)
2570 struct g_consumer *cp;
2572 bp = bp2->bio_caller2;
2574 bp->b_ioflags = bp2->bio_flags;
2576 bp->b_ioflags |= BIO_ERROR;
2577 bp->b_resid = bp->b_bcount - bp2->bio_completed;
2578 bp->b_error = bp2->bio_error;
2580 sp = bp2->bio_caller1;
2581 mtx_lock(&sw_dev_mtx);
2582 swapgeom_release(cp, sp);
2583 mtx_unlock(&sw_dev_mtx);
2588 swapgeom_strategy(struct buf *bp, struct swdevt *sp)
2591 struct g_consumer *cp;
2593 mtx_lock(&sw_dev_mtx);
2596 mtx_unlock(&sw_dev_mtx);
2597 bp->b_error = ENXIO;
2598 bp->b_ioflags |= BIO_ERROR;
2602 swapgeom_acquire(cp);
2603 mtx_unlock(&sw_dev_mtx);
2604 if (bp->b_iocmd == BIO_WRITE)
2607 bio = g_alloc_bio();
2609 mtx_lock(&sw_dev_mtx);
2610 swapgeom_release(cp, sp);
2611 mtx_unlock(&sw_dev_mtx);
2612 bp->b_error = ENOMEM;
2613 bp->b_ioflags |= BIO_ERROR;
2618 bio->bio_caller1 = sp;
2619 bio->bio_caller2 = bp;
2620 bio->bio_cmd = bp->b_iocmd;
2621 bio->bio_offset = (bp->b_blkno - sp->sw_first) * PAGE_SIZE;
2622 bio->bio_length = bp->b_bcount;
2623 bio->bio_done = swapgeom_done;
2624 if (!buf_mapped(bp)) {
2625 bio->bio_ma = bp->b_pages;
2626 bio->bio_data = unmapped_buf;
2627 bio->bio_ma_offset = (vm_offset_t)bp->b_offset & PAGE_MASK;
2628 bio->bio_ma_n = bp->b_npages;
2629 bio->bio_flags |= BIO_UNMAPPED;
2631 bio->bio_data = bp->b_data;
2634 g_io_request(bio, cp);
2639 swapgeom_orphan(struct g_consumer *cp)
2644 mtx_lock(&sw_dev_mtx);
2645 TAILQ_FOREACH(sp, &swtailq, sw_list) {
2646 if (sp->sw_id == cp) {
2647 sp->sw_flags |= SW_CLOSING;
2652 * Drop reference we were created with. Do directly since we're in a
2653 * special context where we don't have to queue the call to
2654 * swapgeom_close_ev().
2657 destroy = ((sp != NULL) && (cp->index == 0));
2660 mtx_unlock(&sw_dev_mtx);
2662 swapgeom_close_ev(cp, 0);
2666 swapgeom_close(struct thread *td, struct swdevt *sw)
2668 struct g_consumer *cp;
2670 mtx_lock(&sw_dev_mtx);
2673 mtx_unlock(&sw_dev_mtx);
2676 * swapgeom_close() may be called from the biodone context,
2677 * where we cannot perform topology changes. Delegate the
2678 * work to the events thread.
2681 g_waitfor_event(swapgeom_close_ev, cp, M_WAITOK, NULL);
2685 swapongeom_locked(struct cdev *dev, struct vnode *vp)
2687 struct g_provider *pp;
2688 struct g_consumer *cp;
2689 static struct g_geom *gp;
2694 pp = g_dev_getprovider(dev);
2697 mtx_lock(&sw_dev_mtx);
2698 TAILQ_FOREACH(sp, &swtailq, sw_list) {
2700 if (cp != NULL && cp->provider == pp) {
2701 mtx_unlock(&sw_dev_mtx);
2705 mtx_unlock(&sw_dev_mtx);
2707 gp = g_new_geomf(&g_swap_class, "swap");
2708 cp = g_new_consumer(gp);
2709 cp->index = 1; /* Number of active I/Os, plus one for being active. */
2710 cp->flags |= G_CF_DIRECT_SEND | G_CF_DIRECT_RECEIVE;
2713 * XXX: Every time you think you can improve the margin for
2714 * footshooting, somebody depends on the ability to do so:
2715 * savecore(8) wants to write to our swapdev so we cannot
2716 * set an exclusive count :-(
2718 error = g_access(cp, 1, 1, 0);
2721 g_destroy_consumer(cp);
2724 nblks = pp->mediasize / DEV_BSIZE;
2725 swaponsomething(vp, cp, nblks, swapgeom_strategy,
2726 swapgeom_close, dev2udev(dev),
2727 (pp->flags & G_PF_ACCEPT_UNMAPPED) != 0 ? SW_UNMAPPED : 0);
2732 swapongeom(struct vnode *vp)
2736 vn_lock(vp, LK_EXCLUSIVE | LK_RETRY);
2737 if (vp->v_type != VCHR || (vp->v_iflag & VI_DOOMED) != 0) {
2741 error = swapongeom_locked(vp->v_rdev, vp);
2742 g_topology_unlock();
2751 * This is used mainly for network filesystem (read: probably only tested
2752 * with NFS) swapfiles.
2757 swapdev_strategy(struct buf *bp, struct swdevt *sp)
2761 bp->b_blkno = ctodb(bp->b_blkno - sp->sw_first);
2765 if (bp->b_iocmd == BIO_WRITE) {
2767 bufobj_wdrop(bp->b_bufobj);
2768 bufobj_wref(&vp2->v_bufobj);
2770 if (bp->b_bufobj != &vp2->v_bufobj)
2771 bp->b_bufobj = &vp2->v_bufobj;
2773 bp->b_iooffset = dbtob(bp->b_blkno);
2779 swapdev_close(struct thread *td, struct swdevt *sp)
2782 VOP_CLOSE(sp->sw_vp, FREAD | FWRITE, td->td_ucred, td);
2788 swaponvp(struct thread *td, struct vnode *vp, u_long nblks)
2795 mtx_lock(&sw_dev_mtx);
2796 TAILQ_FOREACH(sp, &swtailq, sw_list) {
2797 if (sp->sw_id == vp) {
2798 mtx_unlock(&sw_dev_mtx);
2802 mtx_unlock(&sw_dev_mtx);
2804 (void) vn_lock(vp, LK_EXCLUSIVE | LK_RETRY);
2806 error = mac_system_check_swapon(td->td_ucred, vp);
2809 error = VOP_OPEN(vp, FREAD | FWRITE, td->td_ucred, td, NULL);
2810 (void) VOP_UNLOCK(vp, 0);
2814 swaponsomething(vp, vp, nblks, swapdev_strategy, swapdev_close,
2820 sysctl_swap_async_max(SYSCTL_HANDLER_ARGS)
2824 new = nsw_wcount_async_max;
2825 error = sysctl_handle_int(oidp, &new, 0, req);
2826 if (error != 0 || req->newptr == NULL)
2829 if (new > nswbuf / 2 || new < 1)
2832 mtx_lock(&pbuf_mtx);
2833 while (nsw_wcount_async_max != new) {
2835 * Adjust difference. If the current async count is too low,
2836 * we will need to sqeeze our update slowly in. Sleep with a
2837 * higher priority than getpbuf() to finish faster.
2839 n = new - nsw_wcount_async_max;
2840 if (nsw_wcount_async + n >= 0) {
2841 nsw_wcount_async += n;
2842 nsw_wcount_async_max += n;
2843 wakeup(&nsw_wcount_async);
2845 nsw_wcount_async_max -= nsw_wcount_async;
2846 nsw_wcount_async = 0;
2847 msleep(&nsw_wcount_async, &pbuf_mtx, PSWP,
2851 mtx_unlock(&pbuf_mtx);