2 * Copyright (c) 1998 Matthew Dillon,
3 * Copyright (c) 1994 John S. Dyson
4 * Copyright (c) 1990 University of Utah.
5 * Copyright (c) 1982, 1986, 1989, 1993
6 * The Regents of the University of California. All rights reserved.
8 * This code is derived from software contributed to Berkeley by
9 * the Systems Programming Group of the University of Utah Computer
12 * Redistribution and use in source and binary forms, with or without
13 * modification, are permitted provided that the following conditions
15 * 1. Redistributions of source code must retain the above copyright
16 * notice, this list of conditions and the following disclaimer.
17 * 2. Redistributions in binary form must reproduce the above copyright
18 * notice, this list of conditions and the following disclaimer in the
19 * documentation and/or other materials provided with the distribution.
20 * 3. All advertising materials mentioning features or use of this software
21 * must display the following acknowledgement:
22 * This product includes software developed by the University of
23 * California, Berkeley and its contributors.
24 * 4. Neither the name of the University nor the names of its contributors
25 * may be used to endorse or promote products derived from this software
26 * without specific prior written permission.
28 * THIS SOFTWARE IS PROVIDED BY THE REGENTS AND CONTRIBUTORS ``AS IS'' AND
29 * ANY EXPRESS OR IMPLIED WARRANTIES, INCLUDING, BUT NOT LIMITED TO, THE
30 * IMPLIED WARRANTIES OF MERCHANTABILITY AND FITNESS FOR A PARTICULAR PURPOSE
31 * ARE DISCLAIMED. IN NO EVENT SHALL THE REGENTS OR CONTRIBUTORS BE LIABLE
32 * FOR ANY DIRECT, INDIRECT, INCIDENTAL, SPECIAL, EXEMPLARY, OR CONSEQUENTIAL
33 * DAMAGES (INCLUDING, BUT NOT LIMITED TO, PROCUREMENT OF SUBSTITUTE GOODS
34 * OR SERVICES; LOSS OF USE, DATA, OR PROFITS; OR BUSINESS INTERRUPTION)
35 * HOWEVER CAUSED AND ON ANY THEORY OF LIABILITY, WHETHER IN CONTRACT, STRICT
36 * LIABILITY, OR TORT (INCLUDING NEGLIGENCE OR OTHERWISE) ARISING IN ANY WAY
37 * OUT OF THE USE OF THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF
43 * Radix Bitmap 'blists'.
45 * - The new swapper uses the new radix bitmap code. This should scale
46 * to arbitrarily small or arbitrarily large swap spaces and an almost
47 * arbitrary degree of fragmentation.
51 * - on the fly reallocation of swap during putpages. The new system
52 * does not try to keep previously allocated swap blocks for dirty
55 * - on the fly deallocation of swap
57 * - No more garbage collection required. Unnecessarily allocated swap
58 * blocks only exist for dirty vm_page_t's now and these are already
59 * cycled (in a high-load system) by the pager. We also do on-the-fly
60 * removal of invalidated swap blocks when a page is destroyed
63 * from: Utah $Hdr: swap_pager.c 1.4 91/04/30$
65 * @(#)swap_pager.c 8.9 (Berkeley) 3/21/94
66 * @(#)vm_swap.c 8.5 (Berkeley) 2/17/94
69 #include <sys/cdefs.h>
70 __FBSDID("$FreeBSD$");
75 #include <sys/param.h>
76 #include <sys/systm.h>
78 #include <sys/kernel.h>
84 #include <sys/fcntl.h>
85 #include <sys/mount.h>
86 #include <sys/namei.h>
87 #include <sys/vnode.h>
88 #include <sys/malloc.h>
89 #include <sys/racct.h>
90 #include <sys/resource.h>
91 #include <sys/resourcevar.h>
92 #include <sys/rwlock.h>
93 #include <sys/sysctl.h>
94 #include <sys/sysproto.h>
95 #include <sys/blist.h>
98 #include <sys/vmmeter.h>
100 #include <security/mac/mac_framework.h>
104 #include <vm/vm_map.h>
105 #include <vm/vm_kern.h>
106 #include <vm/vm_object.h>
107 #include <vm/vm_page.h>
108 #include <vm/vm_pager.h>
109 #include <vm/vm_pageout.h>
110 #include <vm/vm_param.h>
111 #include <vm/swap_pager.h>
112 #include <vm/vm_extern.h>
115 #include <geom/geom.h>
118 * SWB_NPAGES must be a power of 2. It may be set to 1, 2, 4, 8, 16
119 * or 32 pages per allocation.
120 * The 32-page limit is due to the radix code (kern/subr_blist.c).
122 #ifndef MAX_PAGEOUT_CLUSTER
123 #define MAX_PAGEOUT_CLUSTER 16
126 #if !defined(SWB_NPAGES)
127 #define SWB_NPAGES MAX_PAGEOUT_CLUSTER
131 * The swblock structure maps an object and a small, fixed-size range
132 * of page indices to disk addresses within a swap area.
133 * The collection of these mappings is implemented as a hash table.
134 * Unused disk addresses within a swap area are allocated and managed
137 #define SWCORRECT(n) (sizeof(void *) * (n) / sizeof(daddr_t))
138 #define SWAP_META_PAGES (SWB_NPAGES * 2)
139 #define SWAP_META_MASK (SWAP_META_PAGES - 1)
142 struct swblock *swb_hnext;
143 vm_object_t swb_object;
144 vm_pindex_t swb_index;
146 daddr_t swb_pages[SWAP_META_PAGES];
149 static MALLOC_DEFINE(M_VMPGDATA, "vm_pgdata", "swap pager private data");
150 static struct mtx sw_dev_mtx;
151 static TAILQ_HEAD(, swdevt) swtailq = TAILQ_HEAD_INITIALIZER(swtailq);
152 static struct swdevt *swdevhd; /* Allocate from here next */
153 static int nswapdev; /* Number of swap devices */
154 int swap_pager_avail;
155 static int swdev_syscall_active = 0; /* serialize swap(on|off) */
157 static vm_ooffset_t swap_total;
158 SYSCTL_QUAD(_vm, OID_AUTO, swap_total, CTLFLAG_RD, &swap_total, 0,
159 "Total amount of available swap storage.");
160 static vm_ooffset_t swap_reserved;
161 SYSCTL_QUAD(_vm, OID_AUTO, swap_reserved, CTLFLAG_RD, &swap_reserved, 0,
162 "Amount of swap storage needed to back all allocated anonymous memory.");
163 static int overcommit = 0;
164 SYSCTL_INT(_vm, OID_AUTO, overcommit, CTLFLAG_RW, &overcommit, 0,
165 "Configure virtual memory overcommit behavior. See tuning(7) "
167 static unsigned long swzone;
168 SYSCTL_ULONG(_vm, OID_AUTO, swzone, CTLFLAG_RD, &swzone, 0,
169 "Actual size of swap metadata zone");
170 static unsigned long swap_maxpages;
171 SYSCTL_ULONG(_vm, OID_AUTO, swap_maxpages, CTLFLAG_RD, &swap_maxpages, 0,
172 "Maximum amount of swap supported");
174 /* bits from overcommit */
175 #define SWAP_RESERVE_FORCE_ON (1 << 0)
176 #define SWAP_RESERVE_RLIMIT_ON (1 << 1)
177 #define SWAP_RESERVE_ALLOW_NONWIRED (1 << 2)
180 swap_reserve(vm_ooffset_t incr)
183 return (swap_reserve_by_cred(incr, curthread->td_ucred));
187 swap_reserve_by_cred(vm_ooffset_t incr, struct ucred *cred)
192 static struct timeval lastfail;
195 uip = cred->cr_ruidinfo;
197 if (incr & PAGE_MASK)
198 panic("swap_reserve: & PAGE_MASK");
203 error = racct_add(curproc, RACCT_SWAP, incr);
204 PROC_UNLOCK(curproc);
211 mtx_lock(&sw_dev_mtx);
212 r = swap_reserved + incr;
213 if (overcommit & SWAP_RESERVE_ALLOW_NONWIRED) {
214 s = cnt.v_page_count - cnt.v_free_reserved - cnt.v_wire_count;
219 if ((overcommit & SWAP_RESERVE_FORCE_ON) == 0 || r <= s ||
220 (error = priv_check(curthread, PRIV_VM_SWAP_NOQUOTA)) == 0) {
224 mtx_unlock(&sw_dev_mtx);
228 UIDINFO_VMSIZE_LOCK(uip);
229 if ((overcommit & SWAP_RESERVE_RLIMIT_ON) != 0 &&
230 uip->ui_vmsize + incr > lim_cur(curproc, RLIMIT_SWAP) &&
231 priv_check(curthread, PRIV_VM_SWAP_NORLIMIT))
234 uip->ui_vmsize += incr;
235 UIDINFO_VMSIZE_UNLOCK(uip);
236 PROC_UNLOCK(curproc);
238 mtx_lock(&sw_dev_mtx);
239 swap_reserved -= incr;
240 mtx_unlock(&sw_dev_mtx);
243 if (!res && ppsratecheck(&lastfail, &curfail, 1)) {
244 printf("uid %d, pid %d: swap reservation for %jd bytes failed\n",
245 uip->ui_uid, curproc->p_pid, incr);
251 racct_sub(curproc, RACCT_SWAP, incr);
252 PROC_UNLOCK(curproc);
260 swap_reserve_force(vm_ooffset_t incr)
264 mtx_lock(&sw_dev_mtx);
265 swap_reserved += incr;
266 mtx_unlock(&sw_dev_mtx);
270 racct_add_force(curproc, RACCT_SWAP, incr);
271 PROC_UNLOCK(curproc);
274 uip = curthread->td_ucred->cr_ruidinfo;
276 UIDINFO_VMSIZE_LOCK(uip);
277 uip->ui_vmsize += incr;
278 UIDINFO_VMSIZE_UNLOCK(uip);
279 PROC_UNLOCK(curproc);
283 swap_release(vm_ooffset_t decr)
288 cred = curthread->td_ucred;
289 swap_release_by_cred(decr, cred);
290 PROC_UNLOCK(curproc);
294 swap_release_by_cred(vm_ooffset_t decr, struct ucred *cred)
298 uip = cred->cr_ruidinfo;
300 if (decr & PAGE_MASK)
301 panic("swap_release: & PAGE_MASK");
303 mtx_lock(&sw_dev_mtx);
304 if (swap_reserved < decr)
305 panic("swap_reserved < decr");
306 swap_reserved -= decr;
307 mtx_unlock(&sw_dev_mtx);
309 UIDINFO_VMSIZE_LOCK(uip);
310 if (uip->ui_vmsize < decr)
311 printf("negative vmsize for uid = %d\n", uip->ui_uid);
312 uip->ui_vmsize -= decr;
313 UIDINFO_VMSIZE_UNLOCK(uip);
315 racct_sub_cred(cred, RACCT_SWAP, decr);
318 static void swapdev_strategy(struct buf *, struct swdevt *sw);
320 #define SWM_FREE 0x02 /* free, period */
321 #define SWM_POP 0x04 /* pop out */
323 int swap_pager_full = 2; /* swap space exhaustion (task killing) */
324 static int swap_pager_almost_full = 1; /* swap space exhaustion (w/hysteresis)*/
325 static int nsw_rcount; /* free read buffers */
326 static int nsw_wcount_sync; /* limit write buffers / synchronous */
327 static int nsw_wcount_async; /* limit write buffers / asynchronous */
328 static int nsw_wcount_async_max;/* assigned maximum */
329 static int nsw_cluster_max; /* maximum VOP I/O allowed */
331 static struct swblock **swhash;
332 static int swhash_mask;
333 static struct mtx swhash_mtx;
335 static int swap_async_max = 4; /* maximum in-progress async I/O's */
336 static struct sx sw_alloc_sx;
339 SYSCTL_INT(_vm, OID_AUTO, swap_async_max,
340 CTLFLAG_RW, &swap_async_max, 0, "Maximum running async swap ops");
343 * "named" and "unnamed" anon region objects. Try to reduce the overhead
344 * of searching a named list by hashing it just a little.
349 #define NOBJLIST(handle) \
350 (&swap_pager_object_list[((int)(intptr_t)handle >> 4) & (NOBJLISTS-1)])
352 static struct mtx sw_alloc_mtx; /* protect list manipulation */
353 static struct pagerlst swap_pager_object_list[NOBJLISTS];
354 static uma_zone_t swap_zone;
357 * pagerops for OBJT_SWAP - "swap pager". Some ops are also global procedure
358 * calls hooked from other parts of the VM system and do not appear here.
359 * (see vm/swap_pager.h).
362 swap_pager_alloc(void *handle, vm_ooffset_t size,
363 vm_prot_t prot, vm_ooffset_t offset, struct ucred *);
364 static void swap_pager_dealloc(vm_object_t object);
365 static int swap_pager_getpages(vm_object_t, vm_page_t *, int, int);
366 static void swap_pager_putpages(vm_object_t, vm_page_t *, int, boolean_t, int *);
368 swap_pager_haspage(vm_object_t object, vm_pindex_t pindex, int *before, int *after);
369 static void swap_pager_init(void);
370 static void swap_pager_unswapped(vm_page_t);
371 static void swap_pager_swapoff(struct swdevt *sp);
373 struct pagerops swappagerops = {
374 .pgo_init = swap_pager_init, /* early system initialization of pager */
375 .pgo_alloc = swap_pager_alloc, /* allocate an OBJT_SWAP object */
376 .pgo_dealloc = swap_pager_dealloc, /* deallocate an OBJT_SWAP object */
377 .pgo_getpages = swap_pager_getpages, /* pagein */
378 .pgo_putpages = swap_pager_putpages, /* pageout */
379 .pgo_haspage = swap_pager_haspage, /* get backing store status for page */
380 .pgo_pageunswapped = swap_pager_unswapped, /* remove swap related to page */
384 * dmmax is in page-sized chunks with the new swap system. It was
385 * dev-bsized chunks in the old. dmmax is always a power of 2.
387 * swap_*() routines are externally accessible. swp_*() routines are
391 static int nswap_lowat = 128; /* in pages, swap_pager_almost_full warn */
392 static int nswap_hiwat = 512; /* in pages, swap_pager_almost_full warn */
394 SYSCTL_INT(_vm, OID_AUTO, dmmax,
395 CTLFLAG_RD, &dmmax, 0, "Maximum size of a swap block");
397 static void swp_sizecheck(void);
398 static void swp_pager_async_iodone(struct buf *bp);
399 static int swapongeom(struct thread *, struct vnode *);
400 static int swaponvp(struct thread *, struct vnode *, u_long);
401 static int swapoff_one(struct swdevt *sp, struct ucred *cred);
404 * Swap bitmap functions
406 static void swp_pager_freeswapspace(daddr_t blk, int npages);
407 static daddr_t swp_pager_getswapspace(int npages);
412 static struct swblock **swp_pager_hash(vm_object_t object, vm_pindex_t index);
413 static void swp_pager_meta_build(vm_object_t, vm_pindex_t, daddr_t);
414 static void swp_pager_meta_free(vm_object_t, vm_pindex_t, daddr_t);
415 static void swp_pager_meta_free_all(vm_object_t);
416 static daddr_t swp_pager_meta_ctl(vm_object_t, vm_pindex_t, int);
419 swp_pager_free_nrpage(vm_page_t m)
423 if (m->wire_count == 0)
429 * SWP_SIZECHECK() - update swap_pager_full indication
431 * update the swap_pager_almost_full indication and warn when we are
432 * about to run out of swap space, using lowat/hiwat hysteresis.
434 * Clear swap_pager_full ( task killing ) indication when lowat is met.
436 * No restrictions on call
437 * This routine may not block.
443 if (swap_pager_avail < nswap_lowat) {
444 if (swap_pager_almost_full == 0) {
445 printf("swap_pager: out of swap space\n");
446 swap_pager_almost_full = 1;
450 if (swap_pager_avail > nswap_hiwat)
451 swap_pager_almost_full = 0;
456 * SWP_PAGER_HASH() - hash swap meta data
458 * This is an helper function which hashes the swapblk given
459 * the object and page index. It returns a pointer to a pointer
460 * to the object, or a pointer to a NULL pointer if it could not
463 static struct swblock **
464 swp_pager_hash(vm_object_t object, vm_pindex_t index)
466 struct swblock **pswap;
467 struct swblock *swap;
469 index &= ~(vm_pindex_t)SWAP_META_MASK;
470 pswap = &swhash[(index ^ (int)(intptr_t)object) & swhash_mask];
471 while ((swap = *pswap) != NULL) {
472 if (swap->swb_object == object &&
473 swap->swb_index == index
477 pswap = &swap->swb_hnext;
483 * SWAP_PAGER_INIT() - initialize the swap pager!
485 * Expected to be started from system init. NOTE: This code is run
486 * before much else so be careful what you depend on. Most of the VM
487 * system has yet to be initialized at this point.
490 swap_pager_init(void)
493 * Initialize object lists
497 for (i = 0; i < NOBJLISTS; ++i)
498 TAILQ_INIT(&swap_pager_object_list[i]);
499 mtx_init(&sw_alloc_mtx, "swap_pager list", NULL, MTX_DEF);
500 mtx_init(&sw_dev_mtx, "swapdev", NULL, MTX_DEF);
501 sx_init(&sw_alloc_sx, "swspsx");
504 * Device Stripe, in PAGE_SIZE'd blocks
506 dmmax = SWB_NPAGES * 2;
510 * SWAP_PAGER_SWAP_INIT() - swap pager initialization from pageout process
512 * Expected to be started from pageout process once, prior to entering
516 swap_pager_swap_init(void)
521 * Number of in-transit swap bp operations. Don't
522 * exhaust the pbufs completely. Make sure we
523 * initialize workable values (0 will work for hysteresis
524 * but it isn't very efficient).
526 * The nsw_cluster_max is constrained by the bp->b_pages[]
527 * array (MAXPHYS/PAGE_SIZE) and our locally defined
528 * MAX_PAGEOUT_CLUSTER. Also be aware that swap ops are
529 * constrained by the swap device interleave stripe size.
531 * Currently we hardwire nsw_wcount_async to 4. This limit is
532 * designed to prevent other I/O from having high latencies due to
533 * our pageout I/O. The value 4 works well for one or two active swap
534 * devices but is probably a little low if you have more. Even so,
535 * a higher value would probably generate only a limited improvement
536 * with three or four active swap devices since the system does not
537 * typically have to pageout at extreme bandwidths. We will want
538 * at least 2 per swap devices, and 4 is a pretty good value if you
539 * have one NFS swap device due to the command/ack latency over NFS.
540 * So it all works out pretty well.
542 nsw_cluster_max = min((MAXPHYS/PAGE_SIZE), MAX_PAGEOUT_CLUSTER);
545 nsw_rcount = (nswbuf + 1) / 2;
546 nsw_wcount_sync = (nswbuf + 3) / 4;
547 nsw_wcount_async = 4;
548 nsw_wcount_async_max = nsw_wcount_async;
549 mtx_unlock(&pbuf_mtx);
552 * Initialize our zone. Right now I'm just guessing on the number
553 * we need based on the number of pages in the system. Each swblock
554 * can hold 32 pages, so this is probably overkill. This reservation
555 * is typically limited to around 32MB by default.
557 n = cnt.v_page_count / 2;
558 if (maxswzone && n > maxswzone / sizeof(struct swblock))
559 n = maxswzone / sizeof(struct swblock);
561 swap_zone = uma_zcreate("SWAPMETA", sizeof(struct swblock), NULL, NULL,
562 NULL, NULL, UMA_ALIGN_PTR, UMA_ZONE_NOFREE | UMA_ZONE_VM);
563 if (swap_zone == NULL)
564 panic("failed to create swap_zone.");
566 if (uma_zone_reserve_kva(swap_zone, n))
569 * if the allocation failed, try a zone two thirds the
570 * size of the previous attempt.
575 printf("Swap zone entries reduced from %lu to %lu.\n", n2, n);
576 swap_maxpages = n * SWAP_META_PAGES;
577 swzone = n * sizeof(struct swblock);
581 * Initialize our meta-data hash table. The swapper does not need to
582 * be quite as efficient as the VM system, so we do not use an
583 * oversized hash table.
585 * n: size of hash table, must be power of 2
586 * swhash_mask: hash table index mask
588 for (n = 1; n < n2 / 8; n *= 2)
590 swhash = malloc(sizeof(struct swblock *) * n, M_VMPGDATA, M_WAITOK | M_ZERO);
592 mtx_init(&swhash_mtx, "swap_pager swhash", NULL, MTX_DEF);
596 * SWAP_PAGER_ALLOC() - allocate a new OBJT_SWAP VM object and instantiate
597 * its metadata structures.
599 * This routine is called from the mmap and fork code to create a new
600 * OBJT_SWAP object. We do this by creating an OBJT_DEFAULT object
601 * and then converting it with swp_pager_meta_build().
603 * This routine may block in vm_object_allocate() and create a named
604 * object lookup race, so we must interlock.
609 swap_pager_alloc(void *handle, vm_ooffset_t size, vm_prot_t prot,
610 vm_ooffset_t offset, struct ucred *cred)
615 pindex = OFF_TO_IDX(offset + PAGE_MASK + size);
619 * Reference existing named region or allocate new one. There
620 * should not be a race here against swp_pager_meta_build()
621 * as called from vm_page_remove() in regards to the lookup
624 sx_xlock(&sw_alloc_sx);
625 object = vm_pager_object_lookup(NOBJLIST(handle), handle);
626 if (object == NULL) {
628 if (!swap_reserve_by_cred(size, cred)) {
629 sx_xunlock(&sw_alloc_sx);
635 object = vm_object_allocate(OBJT_DEFAULT, pindex);
636 VM_OBJECT_WLOCK(object);
637 object->handle = handle;
640 object->charge = size;
642 swp_pager_meta_build(object, 0, SWAPBLK_NONE);
643 VM_OBJECT_WUNLOCK(object);
645 sx_xunlock(&sw_alloc_sx);
649 if (!swap_reserve_by_cred(size, cred))
653 object = vm_object_allocate(OBJT_DEFAULT, pindex);
654 VM_OBJECT_WLOCK(object);
657 object->charge = size;
659 swp_pager_meta_build(object, 0, SWAPBLK_NONE);
660 VM_OBJECT_WUNLOCK(object);
666 * SWAP_PAGER_DEALLOC() - remove swap metadata from object
668 * The swap backing for the object is destroyed. The code is
669 * designed such that we can reinstantiate it later, but this
670 * routine is typically called only when the entire object is
671 * about to be destroyed.
673 * The object must be locked.
676 swap_pager_dealloc(vm_object_t object)
680 * Remove from list right away so lookups will fail if we block for
681 * pageout completion.
683 if (object->handle != NULL) {
684 mtx_lock(&sw_alloc_mtx);
685 TAILQ_REMOVE(NOBJLIST(object->handle), object, pager_object_list);
686 mtx_unlock(&sw_alloc_mtx);
689 VM_OBJECT_ASSERT_WLOCKED(object);
690 vm_object_pip_wait(object, "swpdea");
693 * Free all remaining metadata. We only bother to free it from
694 * the swap meta data. We do not attempt to free swapblk's still
695 * associated with vm_page_t's for this object. We do not care
696 * if paging is still in progress on some objects.
698 swp_pager_meta_free_all(object);
699 object->handle = NULL;
700 object->type = OBJT_DEAD;
703 /************************************************************************
704 * SWAP PAGER BITMAP ROUTINES *
705 ************************************************************************/
708 * SWP_PAGER_GETSWAPSPACE() - allocate raw swap space
710 * Allocate swap for the requested number of pages. The starting
711 * swap block number (a page index) is returned or SWAPBLK_NONE
712 * if the allocation failed.
714 * Also has the side effect of advising that somebody made a mistake
715 * when they configured swap and didn't configure enough.
717 * This routine may not sleep.
719 * We allocate in round-robin fashion from the configured devices.
722 swp_pager_getswapspace(int npages)
729 mtx_lock(&sw_dev_mtx);
731 for (i = 0; i < nswapdev; i++) {
733 sp = TAILQ_FIRST(&swtailq);
734 if (!(sp->sw_flags & SW_CLOSING)) {
735 blk = blist_alloc(sp->sw_blist, npages);
736 if (blk != SWAPBLK_NONE) {
738 sp->sw_used += npages;
739 swap_pager_avail -= npages;
741 swdevhd = TAILQ_NEXT(sp, sw_list);
745 sp = TAILQ_NEXT(sp, sw_list);
747 if (swap_pager_full != 2) {
748 printf("swap_pager_getswapspace(%d): failed\n", npages);
750 swap_pager_almost_full = 1;
754 mtx_unlock(&sw_dev_mtx);
759 swp_pager_isondev(daddr_t blk, struct swdevt *sp)
762 return (blk >= sp->sw_first && blk < sp->sw_end);
766 swp_pager_strategy(struct buf *bp)
770 mtx_lock(&sw_dev_mtx);
771 TAILQ_FOREACH(sp, &swtailq, sw_list) {
772 if (bp->b_blkno >= sp->sw_first && bp->b_blkno < sp->sw_end) {
773 mtx_unlock(&sw_dev_mtx);
774 if ((sp->sw_flags & SW_UNMAPPED) != 0 &&
775 unmapped_buf_allowed) {
776 bp->b_kvaalloc = bp->b_data;
777 bp->b_data = unmapped_buf;
778 bp->b_kvabase = unmapped_buf;
780 bp->b_flags |= B_UNMAPPED;
782 pmap_qenter((vm_offset_t)bp->b_data,
783 &bp->b_pages[0], bp->b_bcount / PAGE_SIZE);
785 sp->sw_strategy(bp, sp);
789 panic("Swapdev not found");
794 * SWP_PAGER_FREESWAPSPACE() - free raw swap space
796 * This routine returns the specified swap blocks back to the bitmap.
798 * This routine may not sleep.
801 swp_pager_freeswapspace(daddr_t blk, int npages)
805 mtx_lock(&sw_dev_mtx);
806 TAILQ_FOREACH(sp, &swtailq, sw_list) {
807 if (blk >= sp->sw_first && blk < sp->sw_end) {
808 sp->sw_used -= npages;
810 * If we are attempting to stop swapping on
811 * this device, we don't want to mark any
812 * blocks free lest they be reused.
814 if ((sp->sw_flags & SW_CLOSING) == 0) {
815 blist_free(sp->sw_blist, blk - sp->sw_first,
817 swap_pager_avail += npages;
820 mtx_unlock(&sw_dev_mtx);
824 panic("Swapdev not found");
828 * SWAP_PAGER_FREESPACE() - frees swap blocks associated with a page
829 * range within an object.
831 * This is a globally accessible routine.
833 * This routine removes swapblk assignments from swap metadata.
835 * The external callers of this routine typically have already destroyed
836 * or renamed vm_page_t's associated with this range in the object so
839 * The object must be locked.
842 swap_pager_freespace(vm_object_t object, vm_pindex_t start, vm_size_t size)
845 swp_pager_meta_free(object, start, size);
849 * SWAP_PAGER_RESERVE() - reserve swap blocks in object
851 * Assigns swap blocks to the specified range within the object. The
852 * swap blocks are not zeroed. Any previous swap assignment is destroyed.
854 * Returns 0 on success, -1 on failure.
857 swap_pager_reserve(vm_object_t object, vm_pindex_t start, vm_size_t size)
860 daddr_t blk = SWAPBLK_NONE;
861 vm_pindex_t beg = start; /* save start index */
863 VM_OBJECT_WLOCK(object);
867 while ((blk = swp_pager_getswapspace(n)) == SWAPBLK_NONE) {
870 swp_pager_meta_free(object, beg, start - beg);
871 VM_OBJECT_WUNLOCK(object);
876 swp_pager_meta_build(object, start, blk);
882 swp_pager_meta_free(object, start, n);
883 VM_OBJECT_WUNLOCK(object);
888 * SWAP_PAGER_COPY() - copy blocks from source pager to destination pager
889 * and destroy the source.
891 * Copy any valid swapblks from the source to the destination. In
892 * cases where both the source and destination have a valid swapblk,
893 * we keep the destination's.
895 * This routine is allowed to sleep. It may sleep allocating metadata
896 * indirectly through swp_pager_meta_build() or if paging is still in
897 * progress on the source.
899 * The source object contains no vm_page_t's (which is just as well)
901 * The source object is of type OBJT_SWAP.
903 * The source and destination objects must be locked.
904 * Both object locks may temporarily be released.
907 swap_pager_copy(vm_object_t srcobject, vm_object_t dstobject,
908 vm_pindex_t offset, int destroysource)
912 VM_OBJECT_ASSERT_WLOCKED(srcobject);
913 VM_OBJECT_ASSERT_WLOCKED(dstobject);
916 * If destroysource is set, we remove the source object from the
917 * swap_pager internal queue now.
920 if (srcobject->handle != NULL) {
921 mtx_lock(&sw_alloc_mtx);
923 NOBJLIST(srcobject->handle),
927 mtx_unlock(&sw_alloc_mtx);
932 * transfer source to destination.
934 for (i = 0; i < dstobject->size; ++i) {
938 * Locate (without changing) the swapblk on the destination,
939 * unless it is invalid in which case free it silently, or
940 * if the destination is a resident page, in which case the
941 * source is thrown away.
943 dstaddr = swp_pager_meta_ctl(dstobject, i, 0);
945 if (dstaddr == SWAPBLK_NONE) {
947 * Destination has no swapblk and is not resident,
952 srcaddr = swp_pager_meta_ctl(
958 if (srcaddr != SWAPBLK_NONE) {
960 * swp_pager_meta_build() can sleep.
962 vm_object_pip_add(srcobject, 1);
963 VM_OBJECT_WUNLOCK(srcobject);
964 vm_object_pip_add(dstobject, 1);
965 swp_pager_meta_build(dstobject, i, srcaddr);
966 vm_object_pip_wakeup(dstobject);
967 VM_OBJECT_WLOCK(srcobject);
968 vm_object_pip_wakeup(srcobject);
972 * Destination has valid swapblk or it is represented
973 * by a resident page. We destroy the sourceblock.
976 swp_pager_meta_ctl(srcobject, i + offset, SWM_FREE);
981 * Free left over swap blocks in source.
983 * We have to revert the type to OBJT_DEFAULT so we do not accidently
984 * double-remove the object from the swap queues.
987 swp_pager_meta_free_all(srcobject);
989 * Reverting the type is not necessary, the caller is going
990 * to destroy srcobject directly, but I'm doing it here
991 * for consistency since we've removed the object from its
994 srcobject->type = OBJT_DEFAULT;
999 * SWAP_PAGER_HASPAGE() - determine if we have good backing store for
1000 * the requested page.
1002 * We determine whether good backing store exists for the requested
1003 * page and return TRUE if it does, FALSE if it doesn't.
1005 * If TRUE, we also try to determine how much valid, contiguous backing
1006 * store exists before and after the requested page within a reasonable
1007 * distance. We do not try to restrict it to the swap device stripe
1008 * (that is handled in getpages/putpages). It probably isn't worth
1012 swap_pager_haspage(vm_object_t object, vm_pindex_t pindex, int *before, int *after)
1016 VM_OBJECT_ASSERT_LOCKED(object);
1018 * do we have good backing store at the requested index ?
1020 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) {
1036 for (i = 1; i < (SWB_NPAGES/2); ++i) {
1041 blk = swp_pager_meta_ctl(object, pindex - i, 0);
1042 if (blk != blk0 - i)
1049 * find forward-looking contiguous good backing store
1051 if (after != NULL) {
1054 for (i = 1; i < (SWB_NPAGES/2); ++i) {
1057 blk = swp_pager_meta_ctl(object, pindex + i, 0);
1058 if (blk != blk0 + i)
1067 * SWAP_PAGER_PAGE_UNSWAPPED() - remove swap backing store related to page
1069 * This removes any associated swap backing store, whether valid or
1070 * not, from the page.
1072 * This routine is typically called when a page is made dirty, at
1073 * which point any associated swap can be freed. MADV_FREE also
1074 * calls us in a special-case situation
1076 * NOTE!!! If the page is clean and the swap was valid, the caller
1077 * should make the page dirty before calling this routine. This routine
1078 * does NOT change the m->dirty status of the page. Also: MADV_FREE
1081 * This routine may not sleep.
1083 * The object containing the page must be locked.
1086 swap_pager_unswapped(vm_page_t m)
1089 swp_pager_meta_ctl(m->object, m->pindex, SWM_FREE);
1093 * SWAP_PAGER_GETPAGES() - bring pages in from swap
1095 * Attempt to retrieve (m, count) pages from backing store, but make
1096 * sure we retrieve at least m[reqpage]. We try to load in as large
1097 * a chunk surrounding m[reqpage] as is contiguous in swap and which
1098 * belongs to the same object.
1100 * The code is designed for asynchronous operation and
1101 * immediate-notification of 'reqpage' but tends not to be
1102 * used that way. Please do not optimize-out this algorithmic
1103 * feature, I intend to improve on it in the future.
1105 * The parent has a single vm_object_pip_add() reference prior to
1106 * calling us and we should return with the same.
1108 * The parent has BUSY'd the pages. We should return with 'm'
1109 * left busy, but the others adjusted.
1112 swap_pager_getpages(vm_object_t object, vm_page_t *m, int count, int reqpage)
1122 KASSERT(mreq->object == object,
1123 ("swap_pager_getpages: object mismatch %p/%p",
1124 object, mreq->object));
1127 * Calculate range to retrieve. The pages have already been assigned
1128 * their swapblks. We require a *contiguous* range but we know it to
1129 * not span devices. If we do not supply it, bad things
1130 * happen. Note that blk, iblk & jblk can be SWAPBLK_NONE, but the
1131 * loops are set up such that the case(s) are handled implicitly.
1133 * The swp_*() calls must be made with the object locked.
1135 blk = swp_pager_meta_ctl(mreq->object, mreq->pindex, 0);
1137 for (i = reqpage - 1; i >= 0; --i) {
1140 iblk = swp_pager_meta_ctl(m[i]->object, m[i]->pindex, 0);
1141 if (blk != iblk + (reqpage - i))
1146 for (j = reqpage + 1; j < count; ++j) {
1149 jblk = swp_pager_meta_ctl(m[j]->object, m[j]->pindex, 0);
1150 if (blk != jblk - (j - reqpage))
1155 * free pages outside our collection range. Note: we never free
1156 * mreq, it must remain busy throughout.
1158 if (0 < i || j < count) {
1161 for (k = 0; k < i; ++k)
1162 swp_pager_free_nrpage(m[k]);
1163 for (k = j; k < count; ++k)
1164 swp_pager_free_nrpage(m[k]);
1168 * Return VM_PAGER_FAIL if we have nothing to do. Return mreq
1169 * still busy, but the others unbusied.
1171 if (blk == SWAPBLK_NONE)
1172 return (VM_PAGER_FAIL);
1175 * Getpbuf() can sleep.
1177 VM_OBJECT_WUNLOCK(object);
1179 * Get a swap buffer header to perform the IO
1181 bp = getpbuf(&nsw_rcount);
1182 bp->b_flags |= B_PAGING;
1184 bp->b_iocmd = BIO_READ;
1185 bp->b_iodone = swp_pager_async_iodone;
1186 bp->b_rcred = crhold(thread0.td_ucred);
1187 bp->b_wcred = crhold(thread0.td_ucred);
1188 bp->b_blkno = blk - (reqpage - i);
1189 bp->b_bcount = PAGE_SIZE * (j - i);
1190 bp->b_bufsize = PAGE_SIZE * (j - i);
1191 bp->b_pager.pg_reqpage = reqpage - i;
1193 VM_OBJECT_WLOCK(object);
1197 for (k = i; k < j; ++k) {
1198 bp->b_pages[k - i] = m[k];
1199 m[k]->oflags |= VPO_SWAPINPROG;
1202 bp->b_npages = j - i;
1204 PCPU_INC(cnt.v_swapin);
1205 PCPU_ADD(cnt.v_swappgsin, bp->b_npages);
1208 * We still hold the lock on mreq, and our automatic completion routine
1209 * does not remove it.
1211 vm_object_pip_add(object, bp->b_npages);
1212 VM_OBJECT_WUNLOCK(object);
1215 * perform the I/O. NOTE!!! bp cannot be considered valid after
1216 * this point because we automatically release it on completion.
1217 * Instead, we look at the one page we are interested in which we
1218 * still hold a lock on even through the I/O completion.
1220 * The other pages in our m[] array are also released on completion,
1221 * so we cannot assume they are valid anymore either.
1223 * NOTE: b_blkno is destroyed by the call to swapdev_strategy
1226 swp_pager_strategy(bp);
1229 * wait for the page we want to complete. VPO_SWAPINPROG is always
1230 * cleared on completion. If an I/O error occurs, SWAPBLK_NONE
1231 * is set in the meta-data.
1233 VM_OBJECT_WLOCK(object);
1234 while ((mreq->oflags & VPO_SWAPINPROG) != 0) {
1235 mreq->oflags |= VPO_SWAPSLEEP;
1236 PCPU_INC(cnt.v_intrans);
1237 if (VM_OBJECT_SLEEP(object, &object->paging_in_progress, PSWP,
1238 "swread", hz * 20)) {
1240 "swap_pager: indefinite wait buffer: bufobj: %p, blkno: %jd, size: %ld\n",
1241 bp->b_bufobj, (intmax_t)bp->b_blkno, bp->b_bcount);
1246 * mreq is left busied after completion, but all the other pages
1247 * are freed. If we had an unrecoverable read error the page will
1250 if (mreq->valid != VM_PAGE_BITS_ALL) {
1251 return (VM_PAGER_ERROR);
1253 return (VM_PAGER_OK);
1257 * A final note: in a low swap situation, we cannot deallocate swap
1258 * and mark a page dirty here because the caller is likely to mark
1259 * the page clean when we return, causing the page to possibly revert
1260 * to all-zero's later.
1265 * swap_pager_putpages:
1267 * Assign swap (if necessary) and initiate I/O on the specified pages.
1269 * We support both OBJT_DEFAULT and OBJT_SWAP objects. DEFAULT objects
1270 * are automatically converted to SWAP objects.
1272 * In a low memory situation we may block in VOP_STRATEGY(), but the new
1273 * vm_page reservation system coupled with properly written VFS devices
1274 * should ensure that no low-memory deadlock occurs. This is an area
1277 * The parent has N vm_object_pip_add() references prior to
1278 * calling us and will remove references for rtvals[] that are
1279 * not set to VM_PAGER_PEND. We need to remove the rest on I/O
1282 * The parent has soft-busy'd the pages it passes us and will unbusy
1283 * those whos rtvals[] entry is not set to VM_PAGER_PEND on return.
1284 * We need to unbusy the rest on I/O completion.
1287 swap_pager_putpages(vm_object_t object, vm_page_t *m, int count,
1288 int flags, int *rtvals)
1293 if (count && m[0]->object != object) {
1294 panic("swap_pager_putpages: object mismatch %p/%p",
1303 * Turn object into OBJT_SWAP
1304 * check for bogus sysops
1305 * force sync if not pageout process
1307 if (object->type != OBJT_SWAP)
1308 swp_pager_meta_build(object, 0, SWAPBLK_NONE);
1309 VM_OBJECT_WUNLOCK(object);
1312 if (curproc != pageproc)
1315 sync = (flags & VM_PAGER_PUT_SYNC) != 0;
1320 * Update nsw parameters from swap_async_max sysctl values.
1321 * Do not let the sysop crash the machine with bogus numbers.
1323 mtx_lock(&pbuf_mtx);
1324 if (swap_async_max != nsw_wcount_async_max) {
1330 if ((n = swap_async_max) > nswbuf / 2)
1337 * Adjust difference ( if possible ). If the current async
1338 * count is too low, we may not be able to make the adjustment
1341 n -= nsw_wcount_async_max;
1342 if (nsw_wcount_async + n >= 0) {
1343 nsw_wcount_async += n;
1344 nsw_wcount_async_max += n;
1345 wakeup(&nsw_wcount_async);
1348 mtx_unlock(&pbuf_mtx);
1353 * Assign swap blocks and issue I/O. We reallocate swap on the fly.
1354 * The page is left dirty until the pageout operation completes
1357 for (i = 0; i < count; i += n) {
1363 * Maximum I/O size is limited by a number of factors.
1365 n = min(BLIST_MAX_ALLOC, count - i);
1366 n = min(n, nsw_cluster_max);
1369 * Get biggest block of swap we can. If we fail, fall
1370 * back and try to allocate a smaller block. Don't go
1371 * overboard trying to allocate space if it would overly
1375 (blk = swp_pager_getswapspace(n)) == SWAPBLK_NONE &&
1380 if (blk == SWAPBLK_NONE) {
1381 for (j = 0; j < n; ++j)
1382 rtvals[i+j] = VM_PAGER_FAIL;
1387 * All I/O parameters have been satisfied, build the I/O
1388 * request and assign the swap space.
1391 bp = getpbuf(&nsw_wcount_sync);
1393 bp = getpbuf(&nsw_wcount_async);
1394 bp->b_flags = B_ASYNC;
1396 bp->b_flags |= B_PAGING;
1397 bp->b_iocmd = BIO_WRITE;
1399 bp->b_rcred = crhold(thread0.td_ucred);
1400 bp->b_wcred = crhold(thread0.td_ucred);
1401 bp->b_bcount = PAGE_SIZE * n;
1402 bp->b_bufsize = PAGE_SIZE * n;
1405 VM_OBJECT_WLOCK(object);
1406 for (j = 0; j < n; ++j) {
1407 vm_page_t mreq = m[i+j];
1409 swp_pager_meta_build(
1414 vm_page_dirty(mreq);
1415 rtvals[i+j] = VM_PAGER_OK;
1417 mreq->oflags |= VPO_SWAPINPROG;
1418 bp->b_pages[j] = mreq;
1420 VM_OBJECT_WUNLOCK(object);
1423 * Must set dirty range for NFS to work.
1426 bp->b_dirtyend = bp->b_bcount;
1428 PCPU_INC(cnt.v_swapout);
1429 PCPU_ADD(cnt.v_swappgsout, bp->b_npages);
1434 * NOTE: b_blkno is destroyed by the call to swapdev_strategy
1436 if (sync == FALSE) {
1437 bp->b_iodone = swp_pager_async_iodone;
1439 swp_pager_strategy(bp);
1441 for (j = 0; j < n; ++j)
1442 rtvals[i+j] = VM_PAGER_PEND;
1443 /* restart outter loop */
1450 * NOTE: b_blkno is destroyed by the call to swapdev_strategy
1452 bp->b_iodone = bdone;
1453 swp_pager_strategy(bp);
1456 * Wait for the sync I/O to complete, then update rtvals.
1457 * We just set the rtvals[] to VM_PAGER_PEND so we can call
1458 * our async completion routine at the end, thus avoiding a
1461 bwait(bp, PVM, "swwrt");
1462 for (j = 0; j < n; ++j)
1463 rtvals[i+j] = VM_PAGER_PEND;
1465 * Now that we are through with the bp, we can call the
1466 * normal async completion, which frees everything up.
1468 swp_pager_async_iodone(bp);
1470 VM_OBJECT_WLOCK(object);
1474 * swp_pager_async_iodone:
1476 * Completion routine for asynchronous reads and writes from/to swap.
1477 * Also called manually by synchronous code to finish up a bp.
1479 * This routine may not sleep.
1482 swp_pager_async_iodone(struct buf *bp)
1485 vm_object_t object = NULL;
1490 if (bp->b_ioflags & BIO_ERROR) {
1492 "swap_pager: I/O error - %s failed; blkno %ld,"
1493 "size %ld, error %d\n",
1494 ((bp->b_iocmd == BIO_READ) ? "pagein" : "pageout"),
1502 * remove the mapping for kernel virtual
1504 if ((bp->b_flags & B_UNMAPPED) != 0) {
1505 bp->b_data = bp->b_kvaalloc;
1506 bp->b_kvabase = bp->b_kvaalloc;
1507 bp->b_flags &= ~B_UNMAPPED;
1509 pmap_qremove((vm_offset_t)bp->b_data, bp->b_npages);
1512 object = bp->b_pages[0]->object;
1513 VM_OBJECT_WLOCK(object);
1517 * cleanup pages. If an error occurs writing to swap, we are in
1518 * very serious trouble. If it happens to be a disk error, though,
1519 * we may be able to recover by reassigning the swap later on. So
1520 * in this case we remove the m->swapblk assignment for the page
1521 * but do not free it in the rlist. The errornous block(s) are thus
1522 * never reallocated as swap. Redirty the page and continue.
1524 for (i = 0; i < bp->b_npages; ++i) {
1525 vm_page_t m = bp->b_pages[i];
1527 m->oflags &= ~VPO_SWAPINPROG;
1528 if (m->oflags & VPO_SWAPSLEEP) {
1529 m->oflags &= ~VPO_SWAPSLEEP;
1530 wakeup(&object->paging_in_progress);
1533 if (bp->b_ioflags & BIO_ERROR) {
1535 * If an error occurs I'd love to throw the swapblk
1536 * away without freeing it back to swapspace, so it
1537 * can never be used again. But I can't from an
1540 if (bp->b_iocmd == BIO_READ) {
1542 * When reading, reqpage needs to stay
1543 * locked for the parent, but all other
1544 * pages can be freed. We still want to
1545 * wakeup the parent waiting on the page,
1546 * though. ( also: pg_reqpage can be -1 and
1547 * not match anything ).
1549 * We have to wake specifically requested pages
1550 * up too because we cleared VPO_SWAPINPROG and
1551 * someone may be waiting for that.
1553 * NOTE: for reads, m->dirty will probably
1554 * be overridden by the original caller of
1555 * getpages so don't play cute tricks here.
1558 if (i != bp->b_pager.pg_reqpage)
1559 swp_pager_free_nrpage(m);
1566 * If i == bp->b_pager.pg_reqpage, do not wake
1567 * the page up. The caller needs to.
1571 * If a write error occurs, reactivate page
1572 * so it doesn't clog the inactive list,
1573 * then finish the I/O.
1577 vm_page_activate(m);
1581 } else if (bp->b_iocmd == BIO_READ) {
1583 * NOTE: for reads, m->dirty will probably be
1584 * overridden by the original caller of getpages so
1585 * we cannot set them in order to free the underlying
1586 * swap in a low-swap situation. I don't think we'd
1587 * want to do that anyway, but it was an optimization
1588 * that existed in the old swapper for a time before
1589 * it got ripped out due to precisely this problem.
1591 * If not the requested page then deactivate it.
1593 * Note that the requested page, reqpage, is left
1594 * busied, but we still have to wake it up. The
1595 * other pages are released (unbusied) by
1596 * vm_page_xunbusy().
1598 KASSERT(!pmap_page_is_mapped(m),
1599 ("swp_pager_async_iodone: page %p is mapped", m));
1600 m->valid = VM_PAGE_BITS_ALL;
1601 KASSERT(m->dirty == 0,
1602 ("swp_pager_async_iodone: page %p is dirty", m));
1605 * We have to wake specifically requested pages
1606 * up too because we cleared VPO_SWAPINPROG and
1607 * could be waiting for it in getpages. However,
1608 * be sure to not unbusy getpages specifically
1609 * requested page - getpages expects it to be
1612 if (i != bp->b_pager.pg_reqpage) {
1614 vm_page_deactivate(m);
1624 * For write success, clear the dirty
1625 * status, then finish the I/O ( which decrements the
1626 * busy count and possibly wakes waiter's up ).
1628 KASSERT(!pmap_page_is_write_mapped(m),
1629 ("swp_pager_async_iodone: page %p is not write"
1633 if (vm_page_count_severe()) {
1635 vm_page_try_to_cache(m);
1642 * adjust pip. NOTE: the original parent may still have its own
1643 * pip refs on the object.
1645 if (object != NULL) {
1646 vm_object_pip_wakeupn(object, bp->b_npages);
1647 VM_OBJECT_WUNLOCK(object);
1651 * swapdev_strategy() manually sets b_vp and b_bufobj before calling
1652 * bstrategy(). Set them back to NULL now we're done with it, or we'll
1653 * trigger a KASSERT in relpbuf().
1657 bp->b_bufobj = NULL;
1660 * release the physical I/O buffer
1664 ((bp->b_iocmd == BIO_READ) ? &nsw_rcount :
1665 ((bp->b_flags & B_ASYNC) ?
1674 * swap_pager_isswapped:
1676 * Return 1 if at least one page in the given object is paged
1677 * out to the given swap device.
1679 * This routine may not sleep.
1682 swap_pager_isswapped(vm_object_t object, struct swdevt *sp)
1688 VM_OBJECT_ASSERT_WLOCKED(object);
1689 if (object->type != OBJT_SWAP)
1692 mtx_lock(&swhash_mtx);
1693 for (bcount = 0; bcount < object->un_pager.swp.swp_bcount; bcount++) {
1694 struct swblock *swap;
1696 if ((swap = *swp_pager_hash(object, index)) != NULL) {
1697 for (i = 0; i < SWAP_META_PAGES; ++i) {
1698 if (swp_pager_isondev(swap->swb_pages[i], sp)) {
1699 mtx_unlock(&swhash_mtx);
1704 index += SWAP_META_PAGES;
1706 mtx_unlock(&swhash_mtx);
1711 * SWP_PAGER_FORCE_PAGEIN() - force a swap block to be paged in
1713 * This routine dissociates the page at the given index within a
1714 * swap block from its backing store, paging it in if necessary.
1715 * If the page is paged in, it is placed in the inactive queue,
1716 * since it had its backing store ripped out from under it.
1717 * We also attempt to swap in all other pages in the swap block,
1718 * we only guarantee that the one at the specified index is
1721 * XXX - The code to page the whole block in doesn't work, so we
1722 * revert to the one-by-one behavior for now. Sigh.
1725 swp_pager_force_pagein(vm_object_t object, vm_pindex_t pindex)
1729 vm_object_pip_add(object, 1);
1730 m = vm_page_grab(object, pindex, VM_ALLOC_NORMAL);
1731 if (m->valid == VM_PAGE_BITS_ALL) {
1732 vm_object_pip_subtract(object, 1);
1735 vm_page_activate(m);
1738 vm_pager_page_unswapped(m);
1742 if (swap_pager_getpages(object, &m, 1, 0) != VM_PAGER_OK)
1743 panic("swap_pager_force_pagein: read from swap failed");/*XXX*/
1744 vm_object_pip_subtract(object, 1);
1747 vm_page_deactivate(m);
1750 vm_pager_page_unswapped(m);
1754 * swap_pager_swapoff:
1756 * Page in all of the pages that have been paged out to the
1757 * given device. The corresponding blocks in the bitmap must be
1758 * marked as allocated and the device must be flagged SW_CLOSING.
1759 * There may be no processes swapped out to the device.
1761 * This routine may block.
1764 swap_pager_swapoff(struct swdevt *sp)
1766 struct swblock *swap;
1767 vm_object_t locked_obj, object;
1776 mtx_lock(&swhash_mtx);
1777 for (i = 0; i <= swhash_mask; i++) { /* '<=' is correct here */
1779 for (swap = swhash[i]; swap != NULL; swap = swap->swb_hnext) {
1780 object = swap->swb_object;
1781 pindex = swap->swb_index;
1782 for (j = 0; j < SWAP_META_PAGES; ++j) {
1783 if (!swp_pager_isondev(swap->swb_pages[j], sp))
1785 if (locked_obj != object) {
1786 if (locked_obj != NULL)
1787 VM_OBJECT_WUNLOCK(locked_obj);
1788 locked_obj = object;
1789 if (!VM_OBJECT_TRYWLOCK(object)) {
1790 mtx_unlock(&swhash_mtx);
1791 /* Depends on type-stability. */
1792 VM_OBJECT_WLOCK(object);
1793 mtx_lock(&swhash_mtx);
1797 MPASS(locked_obj == object);
1798 mtx_unlock(&swhash_mtx);
1799 swp_pager_force_pagein(object, pindex + j);
1800 mtx_lock(&swhash_mtx);
1805 mtx_unlock(&swhash_mtx);
1806 if (locked_obj != NULL) {
1807 VM_OBJECT_WUNLOCK(locked_obj);
1812 * Objects may be locked or paging to the device being
1813 * removed, so we will miss their pages and need to
1814 * make another pass. We have marked this device as
1815 * SW_CLOSING, so the activity should finish soon.
1818 if (retries > 100) {
1819 panic("swapoff: failed to locate %d swap blocks",
1822 pause("swpoff", hz / 20);
1827 /************************************************************************
1829 ************************************************************************
1831 * These routines manipulate the swap metadata stored in the
1834 * Swap metadata is implemented with a global hash and not directly
1835 * linked into the object. Instead the object simply contains
1836 * appropriate tracking counters.
1840 * SWP_PAGER_META_BUILD() - add swap block to swap meta data for object
1842 * We first convert the object to a swap object if it is a default
1845 * The specified swapblk is added to the object's swap metadata. If
1846 * the swapblk is not valid, it is freed instead. Any previously
1847 * assigned swapblk is freed.
1850 swp_pager_meta_build(vm_object_t object, vm_pindex_t pindex, daddr_t swapblk)
1852 static volatile int exhausted;
1853 struct swblock *swap;
1854 struct swblock **pswap;
1857 VM_OBJECT_ASSERT_WLOCKED(object);
1859 * Convert default object to swap object if necessary
1861 if (object->type != OBJT_SWAP) {
1862 object->type = OBJT_SWAP;
1863 object->un_pager.swp.swp_bcount = 0;
1865 if (object->handle != NULL) {
1866 mtx_lock(&sw_alloc_mtx);
1868 NOBJLIST(object->handle),
1872 mtx_unlock(&sw_alloc_mtx);
1877 * Locate hash entry. If not found create, but if we aren't adding
1878 * anything just return. If we run out of space in the map we wait
1879 * and, since the hash table may have changed, retry.
1882 mtx_lock(&swhash_mtx);
1883 pswap = swp_pager_hash(object, pindex);
1885 if ((swap = *pswap) == NULL) {
1888 if (swapblk == SWAPBLK_NONE)
1891 swap = *pswap = uma_zalloc(swap_zone, M_NOWAIT |
1892 (curproc == pageproc ? M_USE_RESERVE : 0));
1894 mtx_unlock(&swhash_mtx);
1895 VM_OBJECT_WUNLOCK(object);
1896 if (uma_zone_exhausted(swap_zone)) {
1897 if (atomic_cmpset_int(&exhausted, 0, 1))
1898 printf("swap zone exhausted, "
1899 "increase kern.maxswzone\n");
1900 vm_pageout_oom(VM_OOM_SWAPZ);
1901 pause("swzonex", 10);
1904 VM_OBJECT_WLOCK(object);
1908 if (atomic_cmpset_int(&exhausted, 1, 0))
1909 printf("swap zone ok\n");
1911 swap->swb_hnext = NULL;
1912 swap->swb_object = object;
1913 swap->swb_index = pindex & ~(vm_pindex_t)SWAP_META_MASK;
1914 swap->swb_count = 0;
1916 ++object->un_pager.swp.swp_bcount;
1918 for (i = 0; i < SWAP_META_PAGES; ++i)
1919 swap->swb_pages[i] = SWAPBLK_NONE;
1923 * Delete prior contents of metadata
1925 idx = pindex & SWAP_META_MASK;
1927 if (swap->swb_pages[idx] != SWAPBLK_NONE) {
1928 swp_pager_freeswapspace(swap->swb_pages[idx], 1);
1933 * Enter block into metadata
1935 swap->swb_pages[idx] = swapblk;
1936 if (swapblk != SWAPBLK_NONE)
1939 mtx_unlock(&swhash_mtx);
1943 * SWP_PAGER_META_FREE() - free a range of blocks in the object's swap metadata
1945 * The requested range of blocks is freed, with any associated swap
1946 * returned to the swap bitmap.
1948 * This routine will free swap metadata structures as they are cleaned
1949 * out. This routine does *NOT* operate on swap metadata associated
1950 * with resident pages.
1953 swp_pager_meta_free(vm_object_t object, vm_pindex_t index, daddr_t count)
1956 VM_OBJECT_ASSERT_LOCKED(object);
1957 if (object->type != OBJT_SWAP)
1961 struct swblock **pswap;
1962 struct swblock *swap;
1964 mtx_lock(&swhash_mtx);
1965 pswap = swp_pager_hash(object, index);
1967 if ((swap = *pswap) != NULL) {
1968 daddr_t v = swap->swb_pages[index & SWAP_META_MASK];
1970 if (v != SWAPBLK_NONE) {
1971 swp_pager_freeswapspace(v, 1);
1972 swap->swb_pages[index & SWAP_META_MASK] =
1974 if (--swap->swb_count == 0) {
1975 *pswap = swap->swb_hnext;
1976 uma_zfree(swap_zone, swap);
1977 --object->un_pager.swp.swp_bcount;
1983 int n = SWAP_META_PAGES - (index & SWAP_META_MASK);
1987 mtx_unlock(&swhash_mtx);
1992 * SWP_PAGER_META_FREE_ALL() - destroy all swap metadata associated with object
1994 * This routine locates and destroys all swap metadata associated with
1998 swp_pager_meta_free_all(vm_object_t object)
2000 struct swblock **pswap, *swap;
2005 VM_OBJECT_ASSERT_WLOCKED(object);
2006 if (object->type != OBJT_SWAP)
2010 while (object->un_pager.swp.swp_bcount != 0) {
2011 mtx_lock(&swhash_mtx);
2012 pswap = swp_pager_hash(object, index);
2013 if ((swap = *pswap) != NULL) {
2014 for (i = 0; i < SWAP_META_PAGES; ++i) {
2015 v = swap->swb_pages[i];
2016 if (v != SWAPBLK_NONE) {
2018 swp_pager_freeswapspace(v, 1);
2021 if (swap->swb_count != 0)
2023 "swap_pager_meta_free_all: swb_count != 0");
2024 *pswap = swap->swb_hnext;
2025 uma_zfree(swap_zone, swap);
2026 --object->un_pager.swp.swp_bcount;
2028 mtx_unlock(&swhash_mtx);
2029 index += SWAP_META_PAGES;
2034 * SWP_PAGER_METACTL() - misc control of swap and vm_page_t meta data.
2036 * This routine is capable of looking up, popping, or freeing
2037 * swapblk assignments in the swap meta data or in the vm_page_t.
2038 * The routine typically returns the swapblk being looked-up, or popped,
2039 * or SWAPBLK_NONE if the block was freed, or SWAPBLK_NONE if the block
2040 * was invalid. This routine will automatically free any invalid
2041 * meta-data swapblks.
2043 * It is not possible to store invalid swapblks in the swap meta data
2044 * (other then a literal 'SWAPBLK_NONE'), so we don't bother checking.
2046 * When acting on a busy resident page and paging is in progress, we
2047 * have to wait until paging is complete but otherwise can act on the
2050 * SWM_FREE remove and free swap block from metadata
2051 * SWM_POP remove from meta data but do not free.. pop it out
2054 swp_pager_meta_ctl(vm_object_t object, vm_pindex_t pindex, int flags)
2056 struct swblock **pswap;
2057 struct swblock *swap;
2061 VM_OBJECT_ASSERT_LOCKED(object);
2063 * The meta data only exists of the object is OBJT_SWAP
2064 * and even then might not be allocated yet.
2066 if (object->type != OBJT_SWAP)
2067 return (SWAPBLK_NONE);
2070 mtx_lock(&swhash_mtx);
2071 pswap = swp_pager_hash(object, pindex);
2073 if ((swap = *pswap) != NULL) {
2074 idx = pindex & SWAP_META_MASK;
2075 r1 = swap->swb_pages[idx];
2077 if (r1 != SWAPBLK_NONE) {
2078 if (flags & SWM_FREE) {
2079 swp_pager_freeswapspace(r1, 1);
2082 if (flags & (SWM_FREE|SWM_POP)) {
2083 swap->swb_pages[idx] = SWAPBLK_NONE;
2084 if (--swap->swb_count == 0) {
2085 *pswap = swap->swb_hnext;
2086 uma_zfree(swap_zone, swap);
2087 --object->un_pager.swp.swp_bcount;
2092 mtx_unlock(&swhash_mtx);
2097 * System call swapon(name) enables swapping on device name,
2098 * which must be in the swdevsw. Return EBUSY
2099 * if already swapping on this device.
2101 #ifndef _SYS_SYSPROTO_H_
2102 struct swapon_args {
2112 sys_swapon(struct thread *td, struct swapon_args *uap)
2116 struct nameidata nd;
2119 error = priv_check(td, PRIV_SWAPON);
2124 while (swdev_syscall_active)
2125 tsleep(&swdev_syscall_active, PUSER - 1, "swpon", 0);
2126 swdev_syscall_active = 1;
2129 * Swap metadata may not fit in the KVM if we have physical
2132 if (swap_zone == NULL) {
2137 NDINIT(&nd, LOOKUP, ISOPEN | FOLLOW | AUDITVNODE1, UIO_USERSPACE,
2143 NDFREE(&nd, NDF_ONLY_PNBUF);
2146 if (vn_isdisk(vp, &error)) {
2147 error = swapongeom(td, vp);
2148 } else if (vp->v_type == VREG &&
2149 (vp->v_mount->mnt_vfc->vfc_flags & VFCF_NETWORK) != 0 &&
2150 (error = VOP_GETATTR(vp, &attr, td->td_ucred)) == 0) {
2152 * Allow direct swapping to NFS regular files in the same
2153 * way that nfs_mountroot() sets up diskless swapping.
2155 error = swaponvp(td, vp, attr.va_size / DEV_BSIZE);
2161 swdev_syscall_active = 0;
2162 wakeup_one(&swdev_syscall_active);
2168 * Check that the total amount of swap currently configured does not
2169 * exceed half the theoretical maximum. If it does, print a warning
2170 * message and return -1; otherwise, return 0.
2173 swapon_check_swzone(unsigned long npages)
2175 unsigned long maxpages;
2177 /* absolute maximum we can handle assuming 100% efficiency */
2178 maxpages = uma_zone_get_max(swap_zone) * SWAP_META_PAGES;
2180 /* recommend using no more than half that amount */
2181 if (npages > maxpages / 2) {
2182 printf("warning: total configured swap (%lu pages) "
2183 "exceeds maximum recommended amount (%lu pages).\n",
2184 npages, maxpages / 2);
2185 printf("warning: increase kern.maxswzone "
2186 "or reduce amount of swap.\n");
2193 swaponsomething(struct vnode *vp, void *id, u_long nblks,
2194 sw_strategy_t *strategy, sw_close_t *close, dev_t dev, int flags)
2196 struct swdevt *sp, *tsp;
2201 * nblks is in DEV_BSIZE'd chunks, convert to PAGE_SIZE'd chunks.
2202 * First chop nblks off to page-align it, then convert.
2204 * sw->sw_nblks is in page-sized chunks now too.
2206 nblks &= ~(ctodb(1) - 1);
2207 nblks = dbtoc(nblks);
2210 * If we go beyond this, we get overflows in the radix
2213 mblocks = 0x40000000 / BLIST_META_RADIX;
2214 if (nblks > mblocks) {
2216 "WARNING: reducing swap size to maximum of %luMB per unit\n",
2217 mblocks / 1024 / 1024 * PAGE_SIZE);
2221 sp = malloc(sizeof *sp, M_VMPGDATA, M_WAITOK | M_ZERO);
2226 sp->sw_nblks = nblks;
2228 sp->sw_strategy = strategy;
2229 sp->sw_close = close;
2230 sp->sw_flags = flags;
2232 sp->sw_blist = blist_create(nblks, M_WAITOK);
2234 * Do not free the first two block in order to avoid overwriting
2235 * any bsd label at the front of the partition
2237 blist_free(sp->sw_blist, 2, nblks - 2);
2240 mtx_lock(&sw_dev_mtx);
2241 TAILQ_FOREACH(tsp, &swtailq, sw_list) {
2242 if (tsp->sw_end >= dvbase) {
2244 * We put one uncovered page between the devices
2245 * in order to definitively prevent any cross-device
2248 dvbase = tsp->sw_end + 1;
2251 sp->sw_first = dvbase;
2252 sp->sw_end = dvbase + nblks;
2253 TAILQ_INSERT_TAIL(&swtailq, sp, sw_list);
2255 swap_pager_avail += nblks;
2256 swap_total += (vm_ooffset_t)nblks * PAGE_SIZE;
2257 swapon_check_swzone(swap_total / PAGE_SIZE);
2259 mtx_unlock(&sw_dev_mtx);
2263 * SYSCALL: swapoff(devname)
2265 * Disable swapping on the given device.
2267 * XXX: Badly designed system call: it should use a device index
2268 * rather than filename as specification. We keep sw_vp around
2269 * only to make this work.
2271 #ifndef _SYS_SYSPROTO_H_
2272 struct swapoff_args {
2282 sys_swapoff(struct thread *td, struct swapoff_args *uap)
2285 struct nameidata nd;
2289 error = priv_check(td, PRIV_SWAPOFF);
2294 while (swdev_syscall_active)
2295 tsleep(&swdev_syscall_active, PUSER - 1, "swpoff", 0);
2296 swdev_syscall_active = 1;
2298 NDINIT(&nd, LOOKUP, FOLLOW | AUDITVNODE1, UIO_USERSPACE, uap->name,
2303 NDFREE(&nd, NDF_ONLY_PNBUF);
2306 mtx_lock(&sw_dev_mtx);
2307 TAILQ_FOREACH(sp, &swtailq, sw_list) {
2308 if (sp->sw_vp == vp)
2311 mtx_unlock(&sw_dev_mtx);
2316 error = swapoff_one(sp, td->td_ucred);
2318 swdev_syscall_active = 0;
2319 wakeup_one(&swdev_syscall_active);
2325 swapoff_one(struct swdevt *sp, struct ucred *cred)
2327 u_long nblks, dvbase;
2332 mtx_assert(&Giant, MA_OWNED);
2334 (void) vn_lock(sp->sw_vp, LK_EXCLUSIVE | LK_RETRY);
2335 error = mac_system_check_swapoff(cred, sp->sw_vp);
2336 (void) VOP_UNLOCK(sp->sw_vp, 0);
2340 nblks = sp->sw_nblks;
2343 * We can turn off this swap device safely only if the
2344 * available virtual memory in the system will fit the amount
2345 * of data we will have to page back in, plus an epsilon so
2346 * the system doesn't become critically low on swap space.
2348 if (cnt.v_free_count + cnt.v_cache_count + swap_pager_avail <
2349 nblks + nswap_lowat) {
2354 * Prevent further allocations on this device.
2356 mtx_lock(&sw_dev_mtx);
2357 sp->sw_flags |= SW_CLOSING;
2358 for (dvbase = 0; dvbase < sp->sw_end; dvbase += dmmax) {
2359 swap_pager_avail -= blist_fill(sp->sw_blist,
2362 swap_total -= (vm_ooffset_t)nblks * PAGE_SIZE;
2363 mtx_unlock(&sw_dev_mtx);
2366 * Page in the contents of the device and close it.
2368 swap_pager_swapoff(sp);
2370 sp->sw_close(curthread, sp);
2371 mtx_lock(&sw_dev_mtx);
2373 TAILQ_REMOVE(&swtailq, sp, sw_list);
2375 if (nswapdev == 0) {
2376 swap_pager_full = 2;
2377 swap_pager_almost_full = 1;
2381 mtx_unlock(&sw_dev_mtx);
2382 blist_destroy(sp->sw_blist);
2383 free(sp, M_VMPGDATA);
2390 struct swdevt *sp, *spt;
2391 const char *devname;
2395 while (swdev_syscall_active)
2396 tsleep(&swdev_syscall_active, PUSER - 1, "swpoff", 0);
2397 swdev_syscall_active = 1;
2399 mtx_lock(&sw_dev_mtx);
2400 TAILQ_FOREACH_SAFE(sp, &swtailq, sw_list, spt) {
2401 mtx_unlock(&sw_dev_mtx);
2402 if (vn_isdisk(sp->sw_vp, NULL))
2403 devname = devtoname(sp->sw_vp->v_rdev);
2406 error = swapoff_one(sp, thread0.td_ucred);
2408 printf("Cannot remove swap device %s (error=%d), "
2409 "skipping.\n", devname, error);
2410 } else if (bootverbose) {
2411 printf("Swap device %s removed.\n", devname);
2413 mtx_lock(&sw_dev_mtx);
2415 mtx_unlock(&sw_dev_mtx);
2417 swdev_syscall_active = 0;
2418 wakeup_one(&swdev_syscall_active);
2423 swap_pager_status(int *total, int *used)
2429 mtx_lock(&sw_dev_mtx);
2430 TAILQ_FOREACH(sp, &swtailq, sw_list) {
2431 *total += sp->sw_nblks;
2432 *used += sp->sw_used;
2434 mtx_unlock(&sw_dev_mtx);
2438 swap_dev_info(int name, struct xswdev *xs, char *devname, size_t len)
2441 const char *tmp_devname;
2446 mtx_lock(&sw_dev_mtx);
2447 TAILQ_FOREACH(sp, &swtailq, sw_list) {
2452 xs->xsw_version = XSWDEV_VERSION;
2453 xs->xsw_dev = sp->sw_dev;
2454 xs->xsw_flags = sp->sw_flags;
2455 xs->xsw_nblks = sp->sw_nblks;
2456 xs->xsw_used = sp->sw_used;
2457 if (devname != NULL) {
2458 if (vn_isdisk(sp->sw_vp, NULL))
2459 tmp_devname = devtoname(sp->sw_vp->v_rdev);
2461 tmp_devname = "[file]";
2462 strncpy(devname, tmp_devname, len);
2467 mtx_unlock(&sw_dev_mtx);
2472 sysctl_vm_swap_info(SYSCTL_HANDLER_ARGS)
2477 if (arg2 != 1) /* name length */
2479 error = swap_dev_info(*(int *)arg1, &xs, NULL, 0);
2482 error = SYSCTL_OUT(req, &xs, sizeof(xs));
2486 SYSCTL_INT(_vm, OID_AUTO, nswapdev, CTLFLAG_RD, &nswapdev, 0,
2487 "Number of swap devices");
2488 SYSCTL_NODE(_vm, OID_AUTO, swap_info, CTLFLAG_RD, sysctl_vm_swap_info,
2489 "Swap statistics by device");
2492 * vmspace_swap_count() - count the approximate swap usage in pages for a
2495 * The map must be locked.
2497 * Swap usage is determined by taking the proportional swap used by
2498 * VM objects backing the VM map. To make up for fractional losses,
2499 * if the VM object has any swap use at all the associated map entries
2500 * count for at least 1 swap page.
2503 vmspace_swap_count(struct vmspace *vmspace)
2510 map = &vmspace->vm_map;
2513 for (cur = map->header.next; cur != &map->header; cur = cur->next) {
2514 if ((cur->eflags & MAP_ENTRY_IS_SUB_MAP) == 0 &&
2515 (object = cur->object.vm_object) != NULL) {
2516 VM_OBJECT_WLOCK(object);
2517 if (object->type == OBJT_SWAP &&
2518 object->un_pager.swp.swp_bcount != 0) {
2519 n = (cur->end - cur->start) / PAGE_SIZE;
2520 count += object->un_pager.swp.swp_bcount *
2521 SWAP_META_PAGES * n / object->size + 1;
2523 VM_OBJECT_WUNLOCK(object);
2532 * Swapping onto disk devices.
2536 static g_orphan_t swapgeom_orphan;
2538 static struct g_class g_swap_class = {
2540 .version = G_VERSION,
2541 .orphan = swapgeom_orphan,
2544 DECLARE_GEOM_CLASS(g_swap_class, g_class);
2548 swapgeom_close_ev(void *arg, int flags)
2550 struct g_consumer *cp;
2553 g_access(cp, -1, -1, 0);
2555 g_destroy_consumer(cp);
2559 * Add a reference to the g_consumer for an inflight transaction.
2562 swapgeom_acquire(struct g_consumer *cp)
2565 mtx_assert(&sw_dev_mtx, MA_OWNED);
2570 * Remove a reference from the g_consumer. Post a close event if
2571 * all referneces go away.
2574 swapgeom_release(struct g_consumer *cp, struct swdevt *sp)
2577 mtx_assert(&sw_dev_mtx, MA_OWNED);
2579 if (cp->index == 0) {
2580 if (g_post_event(swapgeom_close_ev, cp, M_NOWAIT, NULL) == 0)
2586 swapgeom_done(struct bio *bp2)
2590 struct g_consumer *cp;
2592 bp = bp2->bio_caller2;
2594 bp->b_ioflags = bp2->bio_flags;
2596 bp->b_ioflags |= BIO_ERROR;
2597 bp->b_resid = bp->b_bcount - bp2->bio_completed;
2598 bp->b_error = bp2->bio_error;
2600 sp = bp2->bio_caller1;
2601 mtx_lock(&sw_dev_mtx);
2602 swapgeom_release(cp, sp);
2603 mtx_unlock(&sw_dev_mtx);
2608 swapgeom_strategy(struct buf *bp, struct swdevt *sp)
2611 struct g_consumer *cp;
2613 mtx_lock(&sw_dev_mtx);
2616 mtx_unlock(&sw_dev_mtx);
2617 bp->b_error = ENXIO;
2618 bp->b_ioflags |= BIO_ERROR;
2622 swapgeom_acquire(cp);
2623 mtx_unlock(&sw_dev_mtx);
2624 if (bp->b_iocmd == BIO_WRITE)
2627 bio = g_alloc_bio();
2629 mtx_lock(&sw_dev_mtx);
2630 swapgeom_release(cp, sp);
2631 mtx_unlock(&sw_dev_mtx);
2632 bp->b_error = ENOMEM;
2633 bp->b_ioflags |= BIO_ERROR;
2638 bio->bio_caller1 = sp;
2639 bio->bio_caller2 = bp;
2640 bio->bio_cmd = bp->b_iocmd;
2641 bio->bio_offset = (bp->b_blkno - sp->sw_first) * PAGE_SIZE;
2642 bio->bio_length = bp->b_bcount;
2643 bio->bio_done = swapgeom_done;
2644 if ((bp->b_flags & B_UNMAPPED) != 0) {
2645 bio->bio_ma = bp->b_pages;
2646 bio->bio_data = unmapped_buf;
2647 bio->bio_ma_offset = (vm_offset_t)bp->b_offset & PAGE_MASK;
2648 bio->bio_ma_n = bp->b_npages;
2649 bio->bio_flags |= BIO_UNMAPPED;
2651 bio->bio_data = bp->b_data;
2654 g_io_request(bio, cp);
2659 swapgeom_orphan(struct g_consumer *cp)
2664 mtx_lock(&sw_dev_mtx);
2665 TAILQ_FOREACH(sp, &swtailq, sw_list) {
2666 if (sp->sw_id == cp) {
2667 sp->sw_flags |= SW_CLOSING;
2672 * Drop reference we were created with. Do directly since we're in a
2673 * special context where we don't have to queue the call to
2674 * swapgeom_close_ev().
2677 destroy = ((sp != NULL) && (cp->index == 0));
2680 mtx_unlock(&sw_dev_mtx);
2682 swapgeom_close_ev(cp, 0);
2686 swapgeom_close(struct thread *td, struct swdevt *sw)
2688 struct g_consumer *cp;
2690 mtx_lock(&sw_dev_mtx);
2693 mtx_unlock(&sw_dev_mtx);
2694 /* XXX: direct call when Giant untangled */
2696 g_waitfor_event(swapgeom_close_ev, cp, M_WAITOK, NULL);
2707 swapongeom_ev(void *arg, int flags)
2710 struct g_provider *pp;
2711 struct g_consumer *cp;
2712 static struct g_geom *gp;
2719 pp = g_dev_getprovider(swh->dev);
2721 swh->error = ENODEV;
2724 mtx_lock(&sw_dev_mtx);
2725 TAILQ_FOREACH(sp, &swtailq, sw_list) {
2727 if (cp != NULL && cp->provider == pp) {
2728 mtx_unlock(&sw_dev_mtx);
2733 mtx_unlock(&sw_dev_mtx);
2735 gp = g_new_geomf(&g_swap_class, "swap");
2736 cp = g_new_consumer(gp);
2737 cp->index = 1; /* Number of active I/Os, plus one for being active. */
2738 cp->flags |= G_CF_DIRECT_SEND | G_CF_DIRECT_RECEIVE;
2741 * XXX: Everytime you think you can improve the margin for
2742 * footshooting, somebody depends on the ability to do so:
2743 * savecore(8) wants to write to our swapdev so we cannot
2744 * set an exclusive count :-(
2746 error = g_access(cp, 1, 1, 0);
2749 g_destroy_consumer(cp);
2753 nblks = pp->mediasize / DEV_BSIZE;
2754 swaponsomething(swh->vp, cp, nblks, swapgeom_strategy,
2755 swapgeom_close, dev2udev(swh->dev),
2756 (pp->flags & G_PF_ACCEPT_UNMAPPED) != 0 ? SW_UNMAPPED : 0);
2761 swapongeom(struct thread *td, struct vnode *vp)
2766 vn_lock(vp, LK_EXCLUSIVE | LK_RETRY);
2768 swh.dev = vp->v_rdev;
2771 /* XXX: direct call when Giant untangled */
2772 error = g_waitfor_event(swapongeom_ev, &swh, M_WAITOK, NULL);
2782 * This is used mainly for network filesystem (read: probably only tested
2783 * with NFS) swapfiles.
2788 swapdev_strategy(struct buf *bp, struct swdevt *sp)
2792 bp->b_blkno = ctodb(bp->b_blkno - sp->sw_first);
2796 if (bp->b_iocmd == BIO_WRITE) {
2798 bufobj_wdrop(bp->b_bufobj);
2799 bufobj_wref(&vp2->v_bufobj);
2801 if (bp->b_bufobj != &vp2->v_bufobj)
2802 bp->b_bufobj = &vp2->v_bufobj;
2804 bp->b_iooffset = dbtob(bp->b_blkno);
2810 swapdev_close(struct thread *td, struct swdevt *sp)
2813 VOP_CLOSE(sp->sw_vp, FREAD | FWRITE, td->td_ucred, td);
2819 swaponvp(struct thread *td, struct vnode *vp, u_long nblks)
2826 mtx_lock(&sw_dev_mtx);
2827 TAILQ_FOREACH(sp, &swtailq, sw_list) {
2828 if (sp->sw_id == vp) {
2829 mtx_unlock(&sw_dev_mtx);
2833 mtx_unlock(&sw_dev_mtx);
2835 (void) vn_lock(vp, LK_EXCLUSIVE | LK_RETRY);
2837 error = mac_system_check_swapon(td->td_ucred, vp);
2840 error = VOP_OPEN(vp, FREAD | FWRITE, td->td_ucred, td, NULL);
2841 (void) VOP_UNLOCK(vp, 0);
2845 swaponsomething(vp, vp, nblks, swapdev_strategy, swapdev_close,