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("$FreeBSD$");
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/sysctl.h>
91 #include <sys/sysproto.h>
92 #include <sys/blist.h>
95 #include <sys/vmmeter.h>
97 #include <security/mac/mac_framework.h>
101 #include <vm/vm_map.h>
102 #include <vm/vm_kern.h>
103 #include <vm/vm_object.h>
104 #include <vm/vm_page.h>
105 #include <vm/vm_pager.h>
106 #include <vm/vm_pageout.h>
107 #include <vm/vm_param.h>
108 #include <vm/swap_pager.h>
109 #include <vm/vm_extern.h>
112 #include <geom/geom.h>
115 * SWB_NPAGES must be a power of 2. It may be set to 1, 2, 4, 8, or 16
116 * pages per allocation. We recommend you stick with the default of 8.
117 * The 16-page limit is due to the radix code (kern/subr_blist.c).
119 #ifndef MAX_PAGEOUT_CLUSTER
120 #define MAX_PAGEOUT_CLUSTER 16
123 #if !defined(SWB_NPAGES)
124 #define SWB_NPAGES MAX_PAGEOUT_CLUSTER
128 * Piecemeal swap metadata structure. Swap is stored in a radix tree.
130 * If SWB_NPAGES is 8 and sizeof(char *) == sizeof(daddr_t), our radix
131 * is basically 8. Assuming PAGE_SIZE == 4096, one tree level represents
132 * 32K worth of data, two levels represent 256K, three levels represent
133 * 2 MBytes. This is acceptable.
135 * Overall memory utilization is about the same as the old swap structure.
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 struct mtx sw_dev_mtx;
150 static TAILQ_HEAD(, swdevt) swtailq = TAILQ_HEAD_INITIALIZER(swtailq);
151 static struct swdevt *swdevhd; /* Allocate from here next */
152 static int nswapdev; /* Number of swap devices */
153 int swap_pager_avail;
154 static int swdev_syscall_active = 0; /* serialize swap(on|off) */
156 static void swapdev_strategy(struct buf *, struct swdevt *sw);
158 #define SWM_FREE 0x02 /* free, period */
159 #define SWM_POP 0x04 /* pop out */
161 int swap_pager_full = 2; /* swap space exhaustion (task killing) */
162 static int swap_pager_almost_full = 1; /* swap space exhaustion (w/hysteresis)*/
163 static int nsw_rcount; /* free read buffers */
164 static int nsw_wcount_sync; /* limit write buffers / synchronous */
165 static int nsw_wcount_async; /* limit write buffers / asynchronous */
166 static int nsw_wcount_async_max;/* assigned maximum */
167 static int nsw_cluster_max; /* maximum VOP I/O allowed */
169 static struct swblock **swhash;
170 static int swhash_mask;
171 static struct mtx swhash_mtx;
173 static int swap_async_max = 4; /* maximum in-progress async I/O's */
174 static struct sx sw_alloc_sx;
177 SYSCTL_INT(_vm, OID_AUTO, swap_async_max,
178 CTLFLAG_RW, &swap_async_max, 0, "Maximum running async swap ops");
181 * "named" and "unnamed" anon region objects. Try to reduce the overhead
182 * of searching a named list by hashing it just a little.
187 #define NOBJLIST(handle) \
188 (&swap_pager_object_list[((int)(intptr_t)handle >> 4) & (NOBJLISTS-1)])
190 static struct mtx sw_alloc_mtx; /* protect list manipulation */
191 static struct pagerlst swap_pager_object_list[NOBJLISTS];
192 static uma_zone_t swap_zone;
193 static struct vm_object swap_zone_obj;
196 * pagerops for OBJT_SWAP - "swap pager". Some ops are also global procedure
197 * calls hooked from other parts of the VM system and do not appear here.
198 * (see vm/swap_pager.h).
201 swap_pager_alloc(void *handle, vm_ooffset_t size,
202 vm_prot_t prot, vm_ooffset_t offset);
203 static void swap_pager_dealloc(vm_object_t object);
204 static int swap_pager_getpages(vm_object_t, vm_page_t *, int, int);
205 static void swap_pager_putpages(vm_object_t, vm_page_t *, int, boolean_t, int *);
207 swap_pager_haspage(vm_object_t object, vm_pindex_t pindex, int *before, int *after);
208 static void swap_pager_init(void);
209 static void swap_pager_unswapped(vm_page_t);
210 static void swap_pager_swapoff(struct swdevt *sp);
212 struct pagerops swappagerops = {
213 .pgo_init = swap_pager_init, /* early system initialization of pager */
214 .pgo_alloc = swap_pager_alloc, /* allocate an OBJT_SWAP object */
215 .pgo_dealloc = swap_pager_dealloc, /* deallocate an OBJT_SWAP object */
216 .pgo_getpages = swap_pager_getpages, /* pagein */
217 .pgo_putpages = swap_pager_putpages, /* pageout */
218 .pgo_haspage = swap_pager_haspage, /* get backing store status for page */
219 .pgo_pageunswapped = swap_pager_unswapped, /* remove swap related to page */
223 * dmmax is in page-sized chunks with the new swap system. It was
224 * dev-bsized chunks in the old. dmmax is always a power of 2.
226 * swap_*() routines are externally accessible. swp_*() routines are
230 static int nswap_lowat = 128; /* in pages, swap_pager_almost_full warn */
231 static int nswap_hiwat = 512; /* in pages, swap_pager_almost_full warn */
233 SYSCTL_INT(_vm, OID_AUTO, dmmax,
234 CTLFLAG_RD, &dmmax, 0, "Maximum size of a swap block");
236 static void swp_sizecheck(void);
237 static void swp_pager_async_iodone(struct buf *bp);
238 static int swapongeom(struct thread *, struct vnode *);
239 static int swaponvp(struct thread *, struct vnode *, u_long);
240 static int swapoff_one(struct swdevt *sp, struct thread *td);
243 * Swap bitmap functions
245 static void swp_pager_freeswapspace(daddr_t blk, int npages);
246 static daddr_t swp_pager_getswapspace(int npages);
251 static struct swblock **swp_pager_hash(vm_object_t object, vm_pindex_t index);
252 static void swp_pager_meta_build(vm_object_t, vm_pindex_t, daddr_t);
253 static void swp_pager_meta_free(vm_object_t, vm_pindex_t, daddr_t);
254 static void swp_pager_meta_free_all(vm_object_t);
255 static daddr_t swp_pager_meta_ctl(vm_object_t, vm_pindex_t, int);
258 * SWP_SIZECHECK() - update swap_pager_full indication
260 * update the swap_pager_almost_full indication and warn when we are
261 * about to run out of swap space, using lowat/hiwat hysteresis.
263 * Clear swap_pager_full ( task killing ) indication when lowat is met.
265 * No restrictions on call
266 * This routine may not block.
267 * This routine must be called at splvm()
273 if (swap_pager_avail < nswap_lowat) {
274 if (swap_pager_almost_full == 0) {
275 printf("swap_pager: out of swap space\n");
276 swap_pager_almost_full = 1;
280 if (swap_pager_avail > nswap_hiwat)
281 swap_pager_almost_full = 0;
286 * SWP_PAGER_HASH() - hash swap meta data
288 * This is an helper function which hashes the swapblk given
289 * the object and page index. It returns a pointer to a pointer
290 * to the object, or a pointer to a NULL pointer if it could not
293 * This routine must be called at splvm().
295 static struct swblock **
296 swp_pager_hash(vm_object_t object, vm_pindex_t index)
298 struct swblock **pswap;
299 struct swblock *swap;
301 index &= ~(vm_pindex_t)SWAP_META_MASK;
302 pswap = &swhash[(index ^ (int)(intptr_t)object) & swhash_mask];
303 while ((swap = *pswap) != NULL) {
304 if (swap->swb_object == object &&
305 swap->swb_index == index
309 pswap = &swap->swb_hnext;
315 * SWAP_PAGER_INIT() - initialize the swap pager!
317 * Expected to be started from system init. NOTE: This code is run
318 * before much else so be careful what you depend on. Most of the VM
319 * system has yet to be initialized at this point.
322 swap_pager_init(void)
325 * Initialize object lists
329 for (i = 0; i < NOBJLISTS; ++i)
330 TAILQ_INIT(&swap_pager_object_list[i]);
331 mtx_init(&sw_alloc_mtx, "swap_pager list", NULL, MTX_DEF);
332 mtx_init(&sw_dev_mtx, "swapdev", NULL, MTX_DEF);
335 * Device Stripe, in PAGE_SIZE'd blocks
337 dmmax = SWB_NPAGES * 2;
341 * SWAP_PAGER_SWAP_INIT() - swap pager initialization from pageout process
343 * Expected to be started from pageout process once, prior to entering
347 swap_pager_swap_init(void)
352 * Number of in-transit swap bp operations. Don't
353 * exhaust the pbufs completely. Make sure we
354 * initialize workable values (0 will work for hysteresis
355 * but it isn't very efficient).
357 * The nsw_cluster_max is constrained by the bp->b_pages[]
358 * array (MAXPHYS/PAGE_SIZE) and our locally defined
359 * MAX_PAGEOUT_CLUSTER. Also be aware that swap ops are
360 * constrained by the swap device interleave stripe size.
362 * Currently we hardwire nsw_wcount_async to 4. This limit is
363 * designed to prevent other I/O from having high latencies due to
364 * our pageout I/O. The value 4 works well for one or two active swap
365 * devices but is probably a little low if you have more. Even so,
366 * a higher value would probably generate only a limited improvement
367 * with three or four active swap devices since the system does not
368 * typically have to pageout at extreme bandwidths. We will want
369 * at least 2 per swap devices, and 4 is a pretty good value if you
370 * have one NFS swap device due to the command/ack latency over NFS.
371 * So it all works out pretty well.
373 nsw_cluster_max = min((MAXPHYS/PAGE_SIZE), MAX_PAGEOUT_CLUSTER);
376 nsw_rcount = (nswbuf + 1) / 2;
377 nsw_wcount_sync = (nswbuf + 3) / 4;
378 nsw_wcount_async = 4;
379 nsw_wcount_async_max = nsw_wcount_async;
380 mtx_unlock(&pbuf_mtx);
383 * Initialize our zone. Right now I'm just guessing on the number
384 * we need based on the number of pages in the system. Each swblock
385 * can hold 16 pages, so this is probably overkill. This reservation
386 * is typically limited to around 32MB by default.
388 n = cnt.v_page_count / 2;
389 if (maxswzone && n > maxswzone / sizeof(struct swblock))
390 n = maxswzone / sizeof(struct swblock);
392 swap_zone = uma_zcreate("SWAPMETA", sizeof(struct swblock), NULL, NULL,
393 NULL, NULL, UMA_ALIGN_PTR, UMA_ZONE_NOFREE | UMA_ZONE_VM);
394 if (swap_zone == NULL)
395 panic("failed to create swap_zone.");
397 if (uma_zone_set_obj(swap_zone, &swap_zone_obj, n))
400 * if the allocation failed, try a zone two thirds the
401 * size of the previous attempt.
406 printf("Swap zone entries reduced from %d to %d.\n", n2, n);
410 * Initialize our meta-data hash table. The swapper does not need to
411 * be quite as efficient as the VM system, so we do not use an
412 * oversized hash table.
414 * n: size of hash table, must be power of 2
415 * swhash_mask: hash table index mask
417 for (n = 1; n < n2 / 8; n *= 2)
419 swhash = malloc(sizeof(struct swblock *) * n, M_VMPGDATA, M_WAITOK | M_ZERO);
421 mtx_init(&swhash_mtx, "swap_pager swhash", NULL, MTX_DEF);
425 * SWAP_PAGER_ALLOC() - allocate a new OBJT_SWAP VM object and instantiate
426 * its metadata structures.
428 * This routine is called from the mmap and fork code to create a new
429 * OBJT_SWAP object. We do this by creating an OBJT_DEFAULT object
430 * and then converting it with swp_pager_meta_build().
432 * This routine may block in vm_object_allocate() and create a named
433 * object lookup race, so we must interlock. We must also run at
434 * splvm() for the object lookup to handle races with interrupts, but
435 * we do not have to maintain splvm() in between the lookup and the
436 * add because (I believe) it is not possible to attempt to create
437 * a new swap object w/handle when a default object with that handle
443 swap_pager_alloc(void *handle, vm_ooffset_t size, vm_prot_t prot,
449 pindex = OFF_TO_IDX(offset + PAGE_MASK + size);
454 * Reference existing named region or allocate new one. There
455 * should not be a race here against swp_pager_meta_build()
456 * as called from vm_page_remove() in regards to the lookup
459 sx_xlock(&sw_alloc_sx);
460 object = vm_pager_object_lookup(NOBJLIST(handle), handle);
462 if (object != NULL) {
463 vm_object_reference(object);
465 object = vm_object_allocate(OBJT_DEFAULT, pindex);
466 object->handle = handle;
468 VM_OBJECT_LOCK(object);
469 swp_pager_meta_build(object, 0, SWAPBLK_NONE);
470 VM_OBJECT_UNLOCK(object);
472 sx_xunlock(&sw_alloc_sx);
475 object = vm_object_allocate(OBJT_DEFAULT, pindex);
477 VM_OBJECT_LOCK(object);
478 swp_pager_meta_build(object, 0, SWAPBLK_NONE);
479 VM_OBJECT_UNLOCK(object);
485 * SWAP_PAGER_DEALLOC() - remove swap metadata from object
487 * The swap backing for the object is destroyed. The code is
488 * designed such that we can reinstantiate it later, but this
489 * routine is typically called only when the entire object is
490 * about to be destroyed.
492 * This routine may block, but no longer does.
494 * The object must be locked or unreferenceable.
497 swap_pager_dealloc(vm_object_t object)
501 * Remove from list right away so lookups will fail if we block for
502 * pageout completion.
504 if (object->handle != NULL) {
505 mtx_lock(&sw_alloc_mtx);
506 TAILQ_REMOVE(NOBJLIST(object->handle), object, pager_object_list);
507 mtx_unlock(&sw_alloc_mtx);
510 VM_OBJECT_LOCK_ASSERT(object, MA_OWNED);
511 vm_object_pip_wait(object, "swpdea");
514 * Free all remaining metadata. We only bother to free it from
515 * the swap meta data. We do not attempt to free swapblk's still
516 * associated with vm_page_t's for this object. We do not care
517 * if paging is still in progress on some objects.
519 swp_pager_meta_free_all(object);
522 /************************************************************************
523 * SWAP PAGER BITMAP ROUTINES *
524 ************************************************************************/
527 * SWP_PAGER_GETSWAPSPACE() - allocate raw swap space
529 * Allocate swap for the requested number of pages. The starting
530 * swap block number (a page index) is returned or SWAPBLK_NONE
531 * if the allocation failed.
533 * Also has the side effect of advising that somebody made a mistake
534 * when they configured swap and didn't configure enough.
536 * Must be called at splvm() to avoid races with bitmap frees from
537 * vm_page_remove() aka swap_pager_page_removed().
539 * This routine may not block
540 * This routine must be called at splvm().
542 * We allocate in round-robin fashion from the configured devices.
545 swp_pager_getswapspace(int npages)
552 mtx_lock(&sw_dev_mtx);
554 for (i = 0; i < nswapdev; i++) {
556 sp = TAILQ_FIRST(&swtailq);
557 if (!(sp->sw_flags & SW_CLOSING)) {
558 blk = blist_alloc(sp->sw_blist, npages);
559 if (blk != SWAPBLK_NONE) {
561 sp->sw_used += npages;
562 swap_pager_avail -= npages;
564 swdevhd = TAILQ_NEXT(sp, sw_list);
568 sp = TAILQ_NEXT(sp, sw_list);
570 if (swap_pager_full != 2) {
571 printf("swap_pager_getswapspace(%d): failed\n", npages);
573 swap_pager_almost_full = 1;
577 mtx_unlock(&sw_dev_mtx);
582 swp_pager_isondev(daddr_t blk, struct swdevt *sp)
585 return (blk >= sp->sw_first && blk < sp->sw_end);
589 swp_pager_strategy(struct buf *bp)
593 mtx_lock(&sw_dev_mtx);
594 TAILQ_FOREACH(sp, &swtailq, sw_list) {
595 if (bp->b_blkno >= sp->sw_first && bp->b_blkno < sp->sw_end) {
596 mtx_unlock(&sw_dev_mtx);
597 sp->sw_strategy(bp, sp);
601 panic("Swapdev not found");
606 * SWP_PAGER_FREESWAPSPACE() - free raw swap space
608 * This routine returns the specified swap blocks back to the bitmap.
610 * Note: This routine may not block (it could in the old swap code),
611 * and through the use of the new blist routines it does not block.
613 * We must be called at splvm() to avoid races with bitmap frees from
614 * vm_page_remove() aka swap_pager_page_removed().
616 * This routine may not block
617 * This routine must be called at splvm().
620 swp_pager_freeswapspace(daddr_t blk, int npages)
624 mtx_lock(&sw_dev_mtx);
625 TAILQ_FOREACH(sp, &swtailq, sw_list) {
626 if (blk >= sp->sw_first && blk < sp->sw_end) {
627 sp->sw_used -= npages;
629 * If we are attempting to stop swapping on
630 * this device, we don't want to mark any
631 * blocks free lest they be reused.
633 if ((sp->sw_flags & SW_CLOSING) == 0) {
634 blist_free(sp->sw_blist, blk - sp->sw_first,
636 swap_pager_avail += npages;
639 mtx_unlock(&sw_dev_mtx);
643 panic("Swapdev not found");
647 * SWAP_PAGER_FREESPACE() - frees swap blocks associated with a page
648 * range within an object.
650 * This is a globally accessible routine.
652 * This routine removes swapblk assignments from swap metadata.
654 * The external callers of this routine typically have already destroyed
655 * or renamed vm_page_t's associated with this range in the object so
658 * This routine may be called at any spl. We up our spl to splvm temporarily
659 * in order to perform the metadata removal.
662 swap_pager_freespace(vm_object_t object, vm_pindex_t start, vm_size_t size)
665 VM_OBJECT_LOCK_ASSERT(object, MA_OWNED);
666 swp_pager_meta_free(object, start, size);
670 * SWAP_PAGER_RESERVE() - reserve swap blocks in object
672 * Assigns swap blocks to the specified range within the object. The
673 * swap blocks are not zerod. Any previous swap assignment is destroyed.
675 * Returns 0 on success, -1 on failure.
678 swap_pager_reserve(vm_object_t object, vm_pindex_t start, vm_size_t size)
681 daddr_t blk = SWAPBLK_NONE;
682 vm_pindex_t beg = start; /* save start index */
684 VM_OBJECT_LOCK(object);
688 while ((blk = swp_pager_getswapspace(n)) == SWAPBLK_NONE) {
691 swp_pager_meta_free(object, beg, start - beg);
692 VM_OBJECT_UNLOCK(object);
697 swp_pager_meta_build(object, start, blk);
703 swp_pager_meta_free(object, start, n);
704 VM_OBJECT_UNLOCK(object);
709 * SWAP_PAGER_COPY() - copy blocks from source pager to destination pager
710 * and destroy the source.
712 * Copy any valid swapblks from the source to the destination. In
713 * cases where both the source and destination have a valid swapblk,
714 * we keep the destination's.
716 * This routine is allowed to block. It may block allocating metadata
717 * indirectly through swp_pager_meta_build() or if paging is still in
718 * progress on the source.
720 * This routine can be called at any spl
722 * XXX vm_page_collapse() kinda expects us not to block because we
723 * supposedly do not need to allocate memory, but for the moment we
724 * *may* have to get a little memory from the zone allocator, but
725 * it is taken from the interrupt memory. We should be ok.
727 * The source object contains no vm_page_t's (which is just as well)
729 * The source object is of type OBJT_SWAP.
731 * The source and destination objects must be locked or
732 * inaccessible (XXX are they ?)
735 swap_pager_copy(vm_object_t srcobject, vm_object_t dstobject,
736 vm_pindex_t offset, int destroysource)
740 VM_OBJECT_LOCK_ASSERT(srcobject, MA_OWNED);
741 VM_OBJECT_LOCK_ASSERT(dstobject, MA_OWNED);
744 * If destroysource is set, we remove the source object from the
745 * swap_pager internal queue now.
748 if (srcobject->handle != NULL) {
749 mtx_lock(&sw_alloc_mtx);
751 NOBJLIST(srcobject->handle),
755 mtx_unlock(&sw_alloc_mtx);
760 * transfer source to destination.
762 for (i = 0; i < dstobject->size; ++i) {
766 * Locate (without changing) the swapblk on the destination,
767 * unless it is invalid in which case free it silently, or
768 * if the destination is a resident page, in which case the
769 * source is thrown away.
771 dstaddr = swp_pager_meta_ctl(dstobject, i, 0);
773 if (dstaddr == SWAPBLK_NONE) {
775 * Destination has no swapblk and is not resident,
780 srcaddr = swp_pager_meta_ctl(
786 if (srcaddr != SWAPBLK_NONE) {
788 * swp_pager_meta_build() can sleep.
790 vm_object_pip_add(srcobject, 1);
791 VM_OBJECT_UNLOCK(srcobject);
792 vm_object_pip_add(dstobject, 1);
793 swp_pager_meta_build(dstobject, i, srcaddr);
794 vm_object_pip_wakeup(dstobject);
795 VM_OBJECT_LOCK(srcobject);
796 vm_object_pip_wakeup(srcobject);
800 * Destination has valid swapblk or it is represented
801 * by a resident page. We destroy the sourceblock.
804 swp_pager_meta_ctl(srcobject, i + offset, SWM_FREE);
809 * Free left over swap blocks in source.
811 * We have to revert the type to OBJT_DEFAULT so we do not accidently
812 * double-remove the object from the swap queues.
815 swp_pager_meta_free_all(srcobject);
817 * Reverting the type is not necessary, the caller is going
818 * to destroy srcobject directly, but I'm doing it here
819 * for consistency since we've removed the object from its
822 srcobject->type = OBJT_DEFAULT;
827 * SWAP_PAGER_HASPAGE() - determine if we have good backing store for
828 * the requested page.
830 * We determine whether good backing store exists for the requested
831 * page and return TRUE if it does, FALSE if it doesn't.
833 * If TRUE, we also try to determine how much valid, contiguous backing
834 * store exists before and after the requested page within a reasonable
835 * distance. We do not try to restrict it to the swap device stripe
836 * (that is handled in getpages/putpages). It probably isn't worth
840 swap_pager_haspage(vm_object_t object, vm_pindex_t pindex, int *before, int *after)
844 VM_OBJECT_LOCK_ASSERT(object, MA_OWNED);
846 * do we have good backing store at the requested index ?
848 blk0 = swp_pager_meta_ctl(object, pindex, 0);
850 if (blk0 == SWAPBLK_NONE) {
859 * find backwards-looking contiguous good backing store
861 if (before != NULL) {
864 for (i = 1; i < (SWB_NPAGES/2); ++i) {
869 blk = swp_pager_meta_ctl(object, pindex - i, 0);
877 * find forward-looking contiguous good backing store
882 for (i = 1; i < (SWB_NPAGES/2); ++i) {
885 blk = swp_pager_meta_ctl(object, pindex + i, 0);
895 * SWAP_PAGER_PAGE_UNSWAPPED() - remove swap backing store related to page
897 * This removes any associated swap backing store, whether valid or
898 * not, from the page.
900 * This routine is typically called when a page is made dirty, at
901 * which point any associated swap can be freed. MADV_FREE also
902 * calls us in a special-case situation
904 * NOTE!!! If the page is clean and the swap was valid, the caller
905 * should make the page dirty before calling this routine. This routine
906 * does NOT change the m->dirty status of the page. Also: MADV_FREE
909 * This routine may not block
910 * This routine must be called at splvm()
913 swap_pager_unswapped(vm_page_t m)
916 VM_OBJECT_LOCK_ASSERT(m->object, MA_OWNED);
917 swp_pager_meta_ctl(m->object, m->pindex, SWM_FREE);
921 * SWAP_PAGER_GETPAGES() - bring pages in from swap
923 * Attempt to retrieve (m, count) pages from backing store, but make
924 * sure we retrieve at least m[reqpage]. We try to load in as large
925 * a chunk surrounding m[reqpage] as is contiguous in swap and which
926 * belongs to the same object.
928 * The code is designed for asynchronous operation and
929 * immediate-notification of 'reqpage' but tends not to be
930 * used that way. Please do not optimize-out this algorithmic
931 * feature, I intend to improve on it in the future.
933 * The parent has a single vm_object_pip_add() reference prior to
934 * calling us and we should return with the same.
936 * The parent has BUSY'd the pages. We should return with 'm'
937 * left busy, but the others adjusted.
940 swap_pager_getpages(vm_object_t object, vm_page_t *m, int count, int reqpage)
950 KASSERT(mreq->object == object,
951 ("swap_pager_getpages: object mismatch %p/%p",
952 object, mreq->object));
955 * Calculate range to retrieve. The pages have already been assigned
956 * their swapblks. We require a *contiguous* range but we know it to
957 * not span devices. If we do not supply it, bad things
958 * happen. Note that blk, iblk & jblk can be SWAPBLK_NONE, but the
959 * loops are set up such that the case(s) are handled implicitly.
961 * The swp_*() calls must be made at splvm(). vm_page_free() does
962 * not need to be, but it will go a little faster if it is.
964 blk = swp_pager_meta_ctl(mreq->object, mreq->pindex, 0);
966 for (i = reqpage - 1; i >= 0; --i) {
969 iblk = swp_pager_meta_ctl(m[i]->object, m[i]->pindex, 0);
970 if (blk != iblk + (reqpage - i))
975 for (j = reqpage + 1; j < count; ++j) {
978 jblk = swp_pager_meta_ctl(m[j]->object, m[j]->pindex, 0);
979 if (blk != jblk - (j - reqpage))
984 * free pages outside our collection range. Note: we never free
985 * mreq, it must remain busy throughout.
987 if (0 < i || j < count) {
990 vm_page_lock_queues();
991 for (k = 0; k < i; ++k)
993 for (k = j; k < count; ++k)
995 vm_page_unlock_queues();
999 * Return VM_PAGER_FAIL if we have nothing to do. Return mreq
1000 * still busy, but the others unbusied.
1002 if (blk == SWAPBLK_NONE)
1003 return (VM_PAGER_FAIL);
1006 * Getpbuf() can sleep.
1008 VM_OBJECT_UNLOCK(object);
1010 * Get a swap buffer header to perform the IO
1012 bp = getpbuf(&nsw_rcount);
1013 bp->b_flags |= B_PAGING;
1016 * map our page(s) into kva for input
1018 pmap_qenter((vm_offset_t)bp->b_data, m + i, j - i);
1020 bp->b_iocmd = BIO_READ;
1021 bp->b_iodone = swp_pager_async_iodone;
1022 bp->b_rcred = crhold(thread0.td_ucred);
1023 bp->b_wcred = crhold(thread0.td_ucred);
1024 bp->b_blkno = blk - (reqpage - i);
1025 bp->b_bcount = PAGE_SIZE * (j - i);
1026 bp->b_bufsize = PAGE_SIZE * (j - i);
1027 bp->b_pager.pg_reqpage = reqpage - i;
1029 VM_OBJECT_LOCK(object);
1033 for (k = i; k < j; ++k) {
1034 bp->b_pages[k - i] = m[k];
1035 m[k]->oflags |= VPO_SWAPINPROG;
1038 bp->b_npages = j - i;
1040 PCPU_INC(cnt.v_swapin);
1041 PCPU_ADD(cnt.v_swappgsin, bp->b_npages);
1044 * We still hold the lock on mreq, and our automatic completion routine
1045 * does not remove it.
1047 vm_object_pip_add(object, bp->b_npages);
1048 VM_OBJECT_UNLOCK(object);
1051 * perform the I/O. NOTE!!! bp cannot be considered valid after
1052 * this point because we automatically release it on completion.
1053 * Instead, we look at the one page we are interested in which we
1054 * still hold a lock on even through the I/O completion.
1056 * The other pages in our m[] array are also released on completion,
1057 * so we cannot assume they are valid anymore either.
1059 * NOTE: b_blkno is destroyed by the call to swapdev_strategy
1062 swp_pager_strategy(bp);
1065 * wait for the page we want to complete. VPO_SWAPINPROG is always
1066 * cleared on completion. If an I/O error occurs, SWAPBLK_NONE
1067 * is set in the meta-data.
1069 VM_OBJECT_LOCK(object);
1070 while ((mreq->oflags & VPO_SWAPINPROG) != 0) {
1071 mreq->oflags |= VPO_WANTED;
1072 vm_page_lock_queues();
1073 vm_page_flag_set(mreq, PG_REFERENCED);
1074 vm_page_unlock_queues();
1075 PCPU_INC(cnt.v_intrans);
1076 if (msleep(mreq, VM_OBJECT_MTX(object), PSWP, "swread", hz*20)) {
1078 "swap_pager: indefinite wait buffer: bufobj: %p, blkno: %jd, size: %ld\n",
1079 bp->b_bufobj, (intmax_t)bp->b_blkno, bp->b_bcount);
1084 * mreq is left busied after completion, but all the other pages
1085 * are freed. If we had an unrecoverable read error the page will
1088 if (mreq->valid != VM_PAGE_BITS_ALL) {
1089 return (VM_PAGER_ERROR);
1091 return (VM_PAGER_OK);
1095 * A final note: in a low swap situation, we cannot deallocate swap
1096 * and mark a page dirty here because the caller is likely to mark
1097 * the page clean when we return, causing the page to possibly revert
1098 * to all-zero's later.
1103 * swap_pager_putpages:
1105 * Assign swap (if necessary) and initiate I/O on the specified pages.
1107 * We support both OBJT_DEFAULT and OBJT_SWAP objects. DEFAULT objects
1108 * are automatically converted to SWAP objects.
1110 * In a low memory situation we may block in VOP_STRATEGY(), but the new
1111 * vm_page reservation system coupled with properly written VFS devices
1112 * should ensure that no low-memory deadlock occurs. This is an area
1115 * The parent has N vm_object_pip_add() references prior to
1116 * calling us and will remove references for rtvals[] that are
1117 * not set to VM_PAGER_PEND. We need to remove the rest on I/O
1120 * The parent has soft-busy'd the pages it passes us and will unbusy
1121 * those whos rtvals[] entry is not set to VM_PAGER_PEND on return.
1122 * We need to unbusy the rest on I/O completion.
1125 swap_pager_putpages(vm_object_t object, vm_page_t *m, int count,
1126 boolean_t sync, int *rtvals)
1132 if (count && m[0]->object != object) {
1133 panic("swap_pager_getpages: object mismatch %p/%p",
1142 * Turn object into OBJT_SWAP
1143 * check for bogus sysops
1144 * force sync if not pageout process
1146 if (object->type != OBJT_SWAP)
1147 swp_pager_meta_build(object, 0, SWAPBLK_NONE);
1148 VM_OBJECT_UNLOCK(object);
1150 if (curproc != pageproc)
1156 * Update nsw parameters from swap_async_max sysctl values.
1157 * Do not let the sysop crash the machine with bogus numbers.
1159 mtx_lock(&pbuf_mtx);
1160 if (swap_async_max != nsw_wcount_async_max) {
1166 if ((n = swap_async_max) > nswbuf / 2)
1173 * Adjust difference ( if possible ). If the current async
1174 * count is too low, we may not be able to make the adjustment
1177 n -= nsw_wcount_async_max;
1178 if (nsw_wcount_async + n >= 0) {
1179 nsw_wcount_async += n;
1180 nsw_wcount_async_max += n;
1181 wakeup(&nsw_wcount_async);
1184 mtx_unlock(&pbuf_mtx);
1189 * Assign swap blocks and issue I/O. We reallocate swap on the fly.
1190 * The page is left dirty until the pageout operation completes
1193 for (i = 0; i < count; i += n) {
1199 * Maximum I/O size is limited by a number of factors.
1201 n = min(BLIST_MAX_ALLOC, count - i);
1202 n = min(n, nsw_cluster_max);
1205 * Get biggest block of swap we can. If we fail, fall
1206 * back and try to allocate a smaller block. Don't go
1207 * overboard trying to allocate space if it would overly
1211 (blk = swp_pager_getswapspace(n)) == SWAPBLK_NONE &&
1216 if (blk == SWAPBLK_NONE) {
1217 for (j = 0; j < n; ++j)
1218 rtvals[i+j] = VM_PAGER_FAIL;
1223 * All I/O parameters have been satisfied, build the I/O
1224 * request and assign the swap space.
1227 bp = getpbuf(&nsw_wcount_sync);
1229 bp = getpbuf(&nsw_wcount_async);
1230 bp->b_flags = B_ASYNC;
1232 bp->b_flags |= B_PAGING;
1233 bp->b_iocmd = BIO_WRITE;
1235 pmap_qenter((vm_offset_t)bp->b_data, &m[i], n);
1237 bp->b_rcred = crhold(thread0.td_ucred);
1238 bp->b_wcred = crhold(thread0.td_ucred);
1239 bp->b_bcount = PAGE_SIZE * n;
1240 bp->b_bufsize = PAGE_SIZE * n;
1243 VM_OBJECT_LOCK(object);
1244 for (j = 0; j < n; ++j) {
1245 vm_page_t mreq = m[i+j];
1247 swp_pager_meta_build(
1252 vm_page_dirty(mreq);
1253 rtvals[i+j] = VM_PAGER_OK;
1255 mreq->oflags |= VPO_SWAPINPROG;
1256 bp->b_pages[j] = mreq;
1258 VM_OBJECT_UNLOCK(object);
1261 * Must set dirty range for NFS to work.
1264 bp->b_dirtyend = bp->b_bcount;
1266 PCPU_INC(cnt.v_swapout);
1267 PCPU_ADD(cnt.v_swappgsout, bp->b_npages);
1272 * NOTE: b_blkno is destroyed by the call to swapdev_strategy
1274 if (sync == FALSE) {
1275 bp->b_iodone = swp_pager_async_iodone;
1277 swp_pager_strategy(bp);
1279 for (j = 0; j < n; ++j)
1280 rtvals[i+j] = VM_PAGER_PEND;
1281 /* restart outter loop */
1288 * NOTE: b_blkno is destroyed by the call to swapdev_strategy
1290 bp->b_iodone = bdone;
1291 swp_pager_strategy(bp);
1294 * Wait for the sync I/O to complete, then update rtvals.
1295 * We just set the rtvals[] to VM_PAGER_PEND so we can call
1296 * our async completion routine at the end, thus avoiding a
1299 bwait(bp, PVM, "swwrt");
1300 for (j = 0; j < n; ++j)
1301 rtvals[i+j] = VM_PAGER_PEND;
1303 * Now that we are through with the bp, we can call the
1304 * normal async completion, which frees everything up.
1306 swp_pager_async_iodone(bp);
1308 VM_OBJECT_LOCK(object);
1312 * swp_pager_async_iodone:
1314 * Completion routine for asynchronous reads and writes from/to swap.
1315 * Also called manually by synchronous code to finish up a bp.
1317 * For READ operations, the pages are PG_BUSY'd. For WRITE operations,
1318 * the pages are vm_page_t->busy'd. For READ operations, we PG_BUSY
1319 * unbusy all pages except the 'main' request page. For WRITE
1320 * operations, we vm_page_t->busy'd unbusy all pages ( we can do this
1321 * because we marked them all VM_PAGER_PEND on return from putpages ).
1323 * This routine may not block.
1324 * This routine is called at splbio() or better
1326 * We up ourselves to splvm() as required for various vm_page related
1330 swp_pager_async_iodone(struct buf *bp)
1333 vm_object_t object = NULL;
1338 if (bp->b_ioflags & BIO_ERROR) {
1340 "swap_pager: I/O error - %s failed; blkno %ld,"
1341 "size %ld, error %d\n",
1342 ((bp->b_iocmd == BIO_READ) ? "pagein" : "pageout"),
1350 * remove the mapping for kernel virtual
1352 pmap_qremove((vm_offset_t)bp->b_data, bp->b_npages);
1355 object = bp->b_pages[0]->object;
1356 VM_OBJECT_LOCK(object);
1358 vm_page_lock_queues();
1360 * cleanup pages. If an error occurs writing to swap, we are in
1361 * very serious trouble. If it happens to be a disk error, though,
1362 * we may be able to recover by reassigning the swap later on. So
1363 * in this case we remove the m->swapblk assignment for the page
1364 * but do not free it in the rlist. The errornous block(s) are thus
1365 * never reallocated as swap. Redirty the page and continue.
1367 for (i = 0; i < bp->b_npages; ++i) {
1368 vm_page_t m = bp->b_pages[i];
1370 m->oflags &= ~VPO_SWAPINPROG;
1372 if (bp->b_ioflags & BIO_ERROR) {
1374 * If an error occurs I'd love to throw the swapblk
1375 * away without freeing it back to swapspace, so it
1376 * can never be used again. But I can't from an
1379 if (bp->b_iocmd == BIO_READ) {
1381 * When reading, reqpage needs to stay
1382 * locked for the parent, but all other
1383 * pages can be freed. We still want to
1384 * wakeup the parent waiting on the page,
1385 * though. ( also: pg_reqpage can be -1 and
1386 * not match anything ).
1388 * We have to wake specifically requested pages
1389 * up too because we cleared VPO_SWAPINPROG and
1390 * someone may be waiting for that.
1392 * NOTE: for reads, m->dirty will probably
1393 * be overridden by the original caller of
1394 * getpages so don't play cute tricks here.
1397 if (i != bp->b_pager.pg_reqpage)
1402 * If i == bp->b_pager.pg_reqpage, do not wake
1403 * the page up. The caller needs to.
1407 * If a write error occurs, reactivate page
1408 * so it doesn't clog the inactive list,
1409 * then finish the I/O.
1412 vm_page_activate(m);
1413 vm_page_io_finish(m);
1415 } else if (bp->b_iocmd == BIO_READ) {
1417 * For read success, clear dirty bits. Nobody should
1418 * have this page mapped but don't take any chances,
1419 * make sure the pmap modify bits are also cleared.
1421 * NOTE: for reads, m->dirty will probably be
1422 * overridden by the original caller of getpages so
1423 * we cannot set them in order to free the underlying
1424 * swap in a low-swap situation. I don't think we'd
1425 * want to do that anyway, but it was an optimization
1426 * that existed in the old swapper for a time before
1427 * it got ripped out due to precisely this problem.
1429 * If not the requested page then deactivate it.
1431 * Note that the requested page, reqpage, is left
1432 * busied, but we still have to wake it up. The
1433 * other pages are released (unbusied) by
1434 * vm_page_wakeup(). We do not set reqpage's
1435 * valid bits here, it is up to the caller.
1437 pmap_clear_modify(m);
1438 m->valid = VM_PAGE_BITS_ALL;
1442 * We have to wake specifically requested pages
1443 * up too because we cleared VPO_SWAPINPROG and
1444 * could be waiting for it in getpages. However,
1445 * be sure to not unbusy getpages specifically
1446 * requested page - getpages expects it to be
1449 if (i != bp->b_pager.pg_reqpage) {
1450 vm_page_deactivate(m);
1457 * For write success, clear the modify and dirty
1458 * status, then finish the I/O ( which decrements the
1459 * busy count and possibly wakes waiter's up ).
1461 pmap_clear_modify(m);
1463 vm_page_io_finish(m);
1464 if (vm_page_count_severe())
1465 vm_page_try_to_cache(m);
1468 vm_page_unlock_queues();
1471 * adjust pip. NOTE: the original parent may still have its own
1472 * pip refs on the object.
1474 if (object != NULL) {
1475 vm_object_pip_wakeupn(object, bp->b_npages);
1476 VM_OBJECT_UNLOCK(object);
1480 * swapdev_strategy() manually sets b_vp and b_bufobj before calling
1481 * bstrategy(). Set them back to NULL now we're done with it, or we'll
1482 * trigger a KASSERT in relpbuf().
1486 bp->b_bufobj = NULL;
1489 * release the physical I/O buffer
1493 ((bp->b_iocmd == BIO_READ) ? &nsw_rcount :
1494 ((bp->b_flags & B_ASYNC) ?
1503 * swap_pager_isswapped:
1505 * Return 1 if at least one page in the given object is paged
1506 * out to the given swap device.
1508 * This routine may not block.
1511 swap_pager_isswapped(vm_object_t object, struct swdevt *sp)
1517 VM_OBJECT_LOCK_ASSERT(object, MA_OWNED);
1518 if (object->type != OBJT_SWAP)
1521 mtx_lock(&swhash_mtx);
1522 for (bcount = 0; bcount < object->un_pager.swp.swp_bcount; bcount++) {
1523 struct swblock *swap;
1525 if ((swap = *swp_pager_hash(object, index)) != NULL) {
1526 for (i = 0; i < SWAP_META_PAGES; ++i) {
1527 if (swp_pager_isondev(swap->swb_pages[i], sp)) {
1528 mtx_unlock(&swhash_mtx);
1533 index += SWAP_META_PAGES;
1534 if (index > 0x20000000)
1535 panic("swap_pager_isswapped: failed to locate all swap meta blocks");
1537 mtx_unlock(&swhash_mtx);
1542 * SWP_PAGER_FORCE_PAGEIN() - force a swap block to be paged in
1544 * This routine dissociates the page at the given index within a
1545 * swap block from its backing store, paging it in if necessary.
1546 * If the page is paged in, it is placed in the inactive queue,
1547 * since it had its backing store ripped out from under it.
1548 * We also attempt to swap in all other pages in the swap block,
1549 * we only guarantee that the one at the specified index is
1552 * XXX - The code to page the whole block in doesn't work, so we
1553 * revert to the one-by-one behavior for now. Sigh.
1556 swp_pager_force_pagein(vm_object_t object, vm_pindex_t pindex)
1560 vm_object_pip_add(object, 1);
1561 m = vm_page_grab(object, pindex, VM_ALLOC_NORMAL|VM_ALLOC_RETRY);
1562 if (m->valid == VM_PAGE_BITS_ALL) {
1563 vm_object_pip_subtract(object, 1);
1564 vm_page_lock_queues();
1565 vm_page_activate(m);
1567 vm_page_unlock_queues();
1569 vm_pager_page_unswapped(m);
1573 if (swap_pager_getpages(object, &m, 1, 0) != VM_PAGER_OK)
1574 panic("swap_pager_force_pagein: read from swap failed");/*XXX*/
1575 vm_object_pip_subtract(object, 1);
1576 vm_page_lock_queues();
1578 vm_page_dontneed(m);
1579 vm_page_unlock_queues();
1581 vm_pager_page_unswapped(m);
1585 * swap_pager_swapoff:
1587 * Page in all of the pages that have been paged out to the
1588 * given device. The corresponding blocks in the bitmap must be
1589 * marked as allocated and the device must be flagged SW_CLOSING.
1590 * There may be no processes swapped out to the device.
1592 * This routine may block.
1595 swap_pager_swapoff(struct swdevt *sp)
1597 struct swblock *swap;
1604 mtx_lock(&swhash_mtx);
1605 for (i = 0; i <= swhash_mask; i++) { /* '<=' is correct here */
1607 for (swap = swhash[i]; swap != NULL; swap = swap->swb_hnext) {
1608 vm_object_t object = swap->swb_object;
1609 vm_pindex_t pindex = swap->swb_index;
1610 for (j = 0; j < SWAP_META_PAGES; ++j) {
1611 if (swp_pager_isondev(swap->swb_pages[j], sp)) {
1612 /* avoid deadlock */
1613 if (!VM_OBJECT_TRYLOCK(object)) {
1616 mtx_unlock(&swhash_mtx);
1617 swp_pager_force_pagein(object,
1619 VM_OBJECT_UNLOCK(object);
1620 mtx_lock(&swhash_mtx);
1627 mtx_unlock(&swhash_mtx);
1630 * Objects may be locked or paging to the device being
1631 * removed, so we will miss their pages and need to
1632 * make another pass. We have marked this device as
1633 * SW_CLOSING, so the activity should finish soon.
1636 if (retries > 100) {
1637 panic("swapoff: failed to locate %d swap blocks",
1640 pause("swpoff", hz / 20);
1645 /************************************************************************
1647 ************************************************************************
1649 * These routines manipulate the swap metadata stored in the
1650 * OBJT_SWAP object. All swp_*() routines must be called at
1651 * splvm() because swap can be freed up by the low level vm_page
1652 * code which might be called from interrupts beyond what splbio() covers.
1654 * Swap metadata is implemented with a global hash and not directly
1655 * linked into the object. Instead the object simply contains
1656 * appropriate tracking counters.
1660 * SWP_PAGER_META_BUILD() - add swap block to swap meta data for object
1662 * We first convert the object to a swap object if it is a default
1665 * The specified swapblk is added to the object's swap metadata. If
1666 * the swapblk is not valid, it is freed instead. Any previously
1667 * assigned swapblk is freed.
1669 * This routine must be called at splvm(), except when used to convert
1670 * an OBJT_DEFAULT object into an OBJT_SWAP object.
1673 swp_pager_meta_build(vm_object_t object, vm_pindex_t pindex, daddr_t swapblk)
1675 struct swblock *swap;
1676 struct swblock **pswap;
1679 VM_OBJECT_LOCK_ASSERT(object, MA_OWNED);
1681 * Convert default object to swap object if necessary
1683 if (object->type != OBJT_SWAP) {
1684 object->type = OBJT_SWAP;
1685 object->un_pager.swp.swp_bcount = 0;
1687 if (object->handle != NULL) {
1688 mtx_lock(&sw_alloc_mtx);
1690 NOBJLIST(object->handle),
1694 mtx_unlock(&sw_alloc_mtx);
1699 * Locate hash entry. If not found create, but if we aren't adding
1700 * anything just return. If we run out of space in the map we wait
1701 * and, since the hash table may have changed, retry.
1704 mtx_lock(&swhash_mtx);
1705 pswap = swp_pager_hash(object, pindex);
1707 if ((swap = *pswap) == NULL) {
1710 if (swapblk == SWAPBLK_NONE)
1713 swap = *pswap = uma_zalloc(swap_zone, M_NOWAIT);
1715 mtx_unlock(&swhash_mtx);
1716 VM_OBJECT_UNLOCK(object);
1717 if (uma_zone_exhausted(swap_zone))
1718 printf("swap zone exhausted, increase kern.maxswzone\n");
1720 VM_OBJECT_LOCK(object);
1724 swap->swb_hnext = NULL;
1725 swap->swb_object = object;
1726 swap->swb_index = pindex & ~(vm_pindex_t)SWAP_META_MASK;
1727 swap->swb_count = 0;
1729 ++object->un_pager.swp.swp_bcount;
1731 for (i = 0; i < SWAP_META_PAGES; ++i)
1732 swap->swb_pages[i] = SWAPBLK_NONE;
1736 * Delete prior contents of metadata
1738 idx = pindex & SWAP_META_MASK;
1740 if (swap->swb_pages[idx] != SWAPBLK_NONE) {
1741 swp_pager_freeswapspace(swap->swb_pages[idx], 1);
1746 * Enter block into metadata
1748 swap->swb_pages[idx] = swapblk;
1749 if (swapblk != SWAPBLK_NONE)
1752 mtx_unlock(&swhash_mtx);
1756 * SWP_PAGER_META_FREE() - free a range of blocks in the object's swap metadata
1758 * The requested range of blocks is freed, with any associated swap
1759 * returned to the swap bitmap.
1761 * This routine will free swap metadata structures as they are cleaned
1762 * out. This routine does *NOT* operate on swap metadata associated
1763 * with resident pages.
1765 * This routine must be called at splvm()
1768 swp_pager_meta_free(vm_object_t object, vm_pindex_t index, daddr_t count)
1771 VM_OBJECT_LOCK_ASSERT(object, MA_OWNED);
1772 if (object->type != OBJT_SWAP)
1776 struct swblock **pswap;
1777 struct swblock *swap;
1779 mtx_lock(&swhash_mtx);
1780 pswap = swp_pager_hash(object, index);
1782 if ((swap = *pswap) != NULL) {
1783 daddr_t v = swap->swb_pages[index & SWAP_META_MASK];
1785 if (v != SWAPBLK_NONE) {
1786 swp_pager_freeswapspace(v, 1);
1787 swap->swb_pages[index & SWAP_META_MASK] =
1789 if (--swap->swb_count == 0) {
1790 *pswap = swap->swb_hnext;
1791 uma_zfree(swap_zone, swap);
1792 --object->un_pager.swp.swp_bcount;
1798 int n = SWAP_META_PAGES - (index & SWAP_META_MASK);
1802 mtx_unlock(&swhash_mtx);
1807 * SWP_PAGER_META_FREE_ALL() - destroy all swap metadata associated with object
1809 * This routine locates and destroys all swap metadata associated with
1812 * This routine must be called at splvm()
1815 swp_pager_meta_free_all(vm_object_t object)
1819 VM_OBJECT_LOCK_ASSERT(object, MA_OWNED);
1820 if (object->type != OBJT_SWAP)
1823 while (object->un_pager.swp.swp_bcount) {
1824 struct swblock **pswap;
1825 struct swblock *swap;
1827 mtx_lock(&swhash_mtx);
1828 pswap = swp_pager_hash(object, index);
1829 if ((swap = *pswap) != NULL) {
1832 for (i = 0; i < SWAP_META_PAGES; ++i) {
1833 daddr_t v = swap->swb_pages[i];
1834 if (v != SWAPBLK_NONE) {
1836 swp_pager_freeswapspace(v, 1);
1839 if (swap->swb_count != 0)
1840 panic("swap_pager_meta_free_all: swb_count != 0");
1841 *pswap = swap->swb_hnext;
1842 uma_zfree(swap_zone, swap);
1843 --object->un_pager.swp.swp_bcount;
1845 mtx_unlock(&swhash_mtx);
1846 index += SWAP_META_PAGES;
1847 if (index > 0x20000000)
1848 panic("swp_pager_meta_free_all: failed to locate all swap meta blocks");
1853 * SWP_PAGER_METACTL() - misc control of swap and vm_page_t meta data.
1855 * This routine is capable of looking up, popping, or freeing
1856 * swapblk assignments in the swap meta data or in the vm_page_t.
1857 * The routine typically returns the swapblk being looked-up, or popped,
1858 * or SWAPBLK_NONE if the block was freed, or SWAPBLK_NONE if the block
1859 * was invalid. This routine will automatically free any invalid
1860 * meta-data swapblks.
1862 * It is not possible to store invalid swapblks in the swap meta data
1863 * (other then a literal 'SWAPBLK_NONE'), so we don't bother checking.
1865 * When acting on a busy resident page and paging is in progress, we
1866 * have to wait until paging is complete but otherwise can act on the
1869 * This routine must be called at splvm().
1871 * SWM_FREE remove and free swap block from metadata
1872 * SWM_POP remove from meta data but do not free.. pop it out
1875 swp_pager_meta_ctl(vm_object_t object, vm_pindex_t pindex, int flags)
1877 struct swblock **pswap;
1878 struct swblock *swap;
1882 VM_OBJECT_LOCK_ASSERT(object, MA_OWNED);
1884 * The meta data only exists of the object is OBJT_SWAP
1885 * and even then might not be allocated yet.
1887 if (object->type != OBJT_SWAP)
1888 return (SWAPBLK_NONE);
1891 mtx_lock(&swhash_mtx);
1892 pswap = swp_pager_hash(object, pindex);
1894 if ((swap = *pswap) != NULL) {
1895 idx = pindex & SWAP_META_MASK;
1896 r1 = swap->swb_pages[idx];
1898 if (r1 != SWAPBLK_NONE) {
1899 if (flags & SWM_FREE) {
1900 swp_pager_freeswapspace(r1, 1);
1903 if (flags & (SWM_FREE|SWM_POP)) {
1904 swap->swb_pages[idx] = SWAPBLK_NONE;
1905 if (--swap->swb_count == 0) {
1906 *pswap = swap->swb_hnext;
1907 uma_zfree(swap_zone, swap);
1908 --object->un_pager.swp.swp_bcount;
1913 mtx_unlock(&swhash_mtx);
1918 * System call swapon(name) enables swapping on device name,
1919 * which must be in the swdevsw. Return EBUSY
1920 * if already swapping on this device.
1922 #ifndef _SYS_SYSPROTO_H_
1923 struct swapon_args {
1933 swapon(struct thread *td, struct swapon_args *uap)
1937 struct nameidata nd;
1940 error = priv_check(td, PRIV_SWAPON);
1945 while (swdev_syscall_active)
1946 tsleep(&swdev_syscall_active, PUSER - 1, "swpon", 0);
1947 swdev_syscall_active = 1;
1950 * Swap metadata may not fit in the KVM if we have physical
1953 if (swap_zone == NULL) {
1958 NDINIT(&nd, LOOKUP, ISOPEN | FOLLOW | AUDITVNODE1, UIO_USERSPACE,
1964 NDFREE(&nd, NDF_ONLY_PNBUF);
1967 if (vn_isdisk(vp, &error)) {
1968 error = swapongeom(td, vp);
1969 } else if (vp->v_type == VREG &&
1970 (vp->v_mount->mnt_vfc->vfc_flags & VFCF_NETWORK) != 0 &&
1971 (error = VOP_GETATTR(vp, &attr, td->td_ucred, td)) == 0) {
1973 * Allow direct swapping to NFS regular files in the same
1974 * way that nfs_mountroot() sets up diskless swapping.
1976 error = swaponvp(td, vp, attr.va_size / DEV_BSIZE);
1982 swdev_syscall_active = 0;
1983 wakeup_one(&swdev_syscall_active);
1989 swaponsomething(struct vnode *vp, void *id, u_long nblks, sw_strategy_t *strategy, sw_close_t *close, dev_t dev)
1991 struct swdevt *sp, *tsp;
1996 * If we go beyond this, we get overflows in the radix
1999 mblocks = 0x40000000 / BLIST_META_RADIX;
2000 if (nblks > mblocks) {
2001 printf("WARNING: reducing size to maximum of %lu blocks per swap unit\n",
2006 * nblks is in DEV_BSIZE'd chunks, convert to PAGE_SIZE'd chunks.
2007 * First chop nblks off to page-align it, then convert.
2009 * sw->sw_nblks is in page-sized chunks now too.
2011 nblks &= ~(ctodb(1) - 1);
2012 nblks = dbtoc(nblks);
2014 sp = malloc(sizeof *sp, M_VMPGDATA, M_WAITOK | M_ZERO);
2019 sp->sw_nblks = nblks;
2021 sp->sw_strategy = strategy;
2022 sp->sw_close = close;
2024 sp->sw_blist = blist_create(nblks);
2026 * Do not free the first two block in order to avoid overwriting
2027 * any bsd label at the front of the partition
2029 blist_free(sp->sw_blist, 2, nblks - 2);
2032 mtx_lock(&sw_dev_mtx);
2033 TAILQ_FOREACH(tsp, &swtailq, sw_list) {
2034 if (tsp->sw_end >= dvbase) {
2036 * We put one uncovered page between the devices
2037 * in order to definitively prevent any cross-device
2040 dvbase = tsp->sw_end + 1;
2043 sp->sw_first = dvbase;
2044 sp->sw_end = dvbase + nblks;
2045 TAILQ_INSERT_TAIL(&swtailq, sp, sw_list);
2047 swap_pager_avail += nblks;
2049 mtx_unlock(&sw_dev_mtx);
2053 * SYSCALL: swapoff(devname)
2055 * Disable swapping on the given device.
2057 * XXX: Badly designed system call: it should use a device index
2058 * rather than filename as specification. We keep sw_vp around
2059 * only to make this work.
2061 #ifndef _SYS_SYSPROTO_H_
2062 struct swapoff_args {
2072 swapoff(struct thread *td, struct swapoff_args *uap)
2075 struct nameidata nd;
2079 error = priv_check(td, PRIV_SWAPOFF);
2084 while (swdev_syscall_active)
2085 tsleep(&swdev_syscall_active, PUSER - 1, "swpoff", 0);
2086 swdev_syscall_active = 1;
2088 NDINIT(&nd, LOOKUP, FOLLOW | AUDITVNODE1, UIO_USERSPACE, uap->name,
2093 NDFREE(&nd, NDF_ONLY_PNBUF);
2096 mtx_lock(&sw_dev_mtx);
2097 TAILQ_FOREACH(sp, &swtailq, sw_list) {
2098 if (sp->sw_vp == vp)
2101 mtx_unlock(&sw_dev_mtx);
2106 error = swapoff_one(sp, td);
2108 swdev_syscall_active = 0;
2109 wakeup_one(&swdev_syscall_active);
2115 swapoff_one(struct swdevt *sp, struct thread *td)
2117 u_long nblks, dvbase;
2122 mtx_assert(&Giant, MA_OWNED);
2124 (void) vn_lock(sp->sw_vp, LK_EXCLUSIVE | LK_RETRY, td);
2125 error = mac_check_system_swapoff(td->td_ucred, sp->sw_vp);
2126 (void) VOP_UNLOCK(sp->sw_vp, 0, td);
2130 nblks = sp->sw_nblks;
2133 * We can turn off this swap device safely only if the
2134 * available virtual memory in the system will fit the amount
2135 * of data we will have to page back in, plus an epsilon so
2136 * the system doesn't become critically low on swap space.
2138 if (cnt.v_free_count + cnt.v_cache_count + swap_pager_avail <
2139 nblks + nswap_lowat) {
2144 * Prevent further allocations on this device.
2146 mtx_lock(&sw_dev_mtx);
2147 sp->sw_flags |= SW_CLOSING;
2148 for (dvbase = 0; dvbase < sp->sw_end; dvbase += dmmax) {
2149 swap_pager_avail -= blist_fill(sp->sw_blist,
2152 mtx_unlock(&sw_dev_mtx);
2155 * Page in the contents of the device and close it.
2157 swap_pager_swapoff(sp);
2159 sp->sw_close(td, sp);
2161 mtx_lock(&sw_dev_mtx);
2162 TAILQ_REMOVE(&swtailq, sp, sw_list);
2164 if (nswapdev == 0) {
2165 swap_pager_full = 2;
2166 swap_pager_almost_full = 1;
2170 mtx_unlock(&sw_dev_mtx);
2171 blist_destroy(sp->sw_blist);
2172 free(sp, M_VMPGDATA);
2179 struct swdevt *sp, *spt;
2180 const char *devname;
2184 while (swdev_syscall_active)
2185 tsleep(&swdev_syscall_active, PUSER - 1, "swpoff", 0);
2186 swdev_syscall_active = 1;
2188 mtx_lock(&sw_dev_mtx);
2189 TAILQ_FOREACH_SAFE(sp, &swtailq, sw_list, spt) {
2190 mtx_unlock(&sw_dev_mtx);
2191 if (vn_isdisk(sp->sw_vp, NULL))
2192 devname = sp->sw_vp->v_rdev->si_name;
2195 error = swapoff_one(sp, &thread0);
2197 printf("Cannot remove swap device %s (error=%d), "
2198 "skipping.\n", devname, error);
2199 } else if (bootverbose) {
2200 printf("Swap device %s removed.\n", devname);
2202 mtx_lock(&sw_dev_mtx);
2204 mtx_unlock(&sw_dev_mtx);
2206 swdev_syscall_active = 0;
2207 wakeup_one(&swdev_syscall_active);
2212 swap_pager_status(int *total, int *used)
2218 mtx_lock(&sw_dev_mtx);
2219 TAILQ_FOREACH(sp, &swtailq, sw_list) {
2220 *total += sp->sw_nblks;
2221 *used += sp->sw_used;
2223 mtx_unlock(&sw_dev_mtx);
2227 sysctl_vm_swap_info(SYSCTL_HANDLER_ARGS)
2229 int *name = (int *)arg1;
2234 if (arg2 != 1) /* name length */
2238 mtx_lock(&sw_dev_mtx);
2239 TAILQ_FOREACH(sp, &swtailq, sw_list) {
2241 mtx_unlock(&sw_dev_mtx);
2242 xs.xsw_version = XSWDEV_VERSION;
2243 xs.xsw_dev = sp->sw_dev;
2244 xs.xsw_flags = sp->sw_flags;
2245 xs.xsw_nblks = sp->sw_nblks;
2246 xs.xsw_used = sp->sw_used;
2248 error = SYSCTL_OUT(req, &xs, sizeof(xs));
2253 mtx_unlock(&sw_dev_mtx);
2257 SYSCTL_INT(_vm, OID_AUTO, nswapdev, CTLFLAG_RD, &nswapdev, 0,
2258 "Number of swap devices");
2259 SYSCTL_NODE(_vm, OID_AUTO, swap_info, CTLFLAG_RD, sysctl_vm_swap_info,
2260 "Swap statistics by device");
2263 * vmspace_swap_count() - count the approximate swap useage in pages for a
2266 * The map must be locked.
2268 * Swap useage is determined by taking the proportional swap used by
2269 * VM objects backing the VM map. To make up for fractional losses,
2270 * if the VM object has any swap use at all the associated map entries
2271 * count for at least 1 swap page.
2274 vmspace_swap_count(struct vmspace *vmspace)
2276 vm_map_t map = &vmspace->vm_map;
2280 for (cur = map->header.next; cur != &map->header; cur = cur->next) {
2283 if ((cur->eflags & MAP_ENTRY_IS_SUB_MAP) == 0 &&
2284 (object = cur->object.vm_object) != NULL) {
2285 VM_OBJECT_LOCK(object);
2286 if (object->type == OBJT_SWAP &&
2287 object->un_pager.swp.swp_bcount != 0) {
2288 int n = (cur->end - cur->start) / PAGE_SIZE;
2290 count += object->un_pager.swp.swp_bcount *
2291 SWAP_META_PAGES * n / object->size + 1;
2293 VM_OBJECT_UNLOCK(object);
2302 * Swapping onto disk devices.
2306 static g_orphan_t swapgeom_orphan;
2308 static struct g_class g_swap_class = {
2310 .version = G_VERSION,
2311 .orphan = swapgeom_orphan,
2314 DECLARE_GEOM_CLASS(g_swap_class, g_class);
2318 swapgeom_done(struct bio *bp2)
2322 bp = bp2->bio_caller2;
2323 bp->b_ioflags = bp2->bio_flags;
2325 bp->b_ioflags |= BIO_ERROR;
2326 bp->b_resid = bp->b_bcount - bp2->bio_completed;
2327 bp->b_error = bp2->bio_error;
2333 swapgeom_strategy(struct buf *bp, struct swdevt *sp)
2336 struct g_consumer *cp;
2340 bp->b_error = ENXIO;
2341 bp->b_ioflags |= BIO_ERROR;
2345 bio = g_alloc_bio();
2348 * XXX: We shouldn't really sleep here when we run out of buffers
2349 * XXX: but the alternative is worse right now.
2352 bp->b_error = ENOMEM;
2353 bp->b_ioflags |= BIO_ERROR;
2358 bio->bio_caller2 = bp;
2359 bio->bio_cmd = bp->b_iocmd;
2360 bio->bio_data = bp->b_data;
2361 bio->bio_offset = (bp->b_blkno - sp->sw_first) * PAGE_SIZE;
2362 bio->bio_length = bp->b_bcount;
2363 bio->bio_done = swapgeom_done;
2364 g_io_request(bio, cp);
2369 swapgeom_orphan(struct g_consumer *cp)
2373 mtx_lock(&sw_dev_mtx);
2374 TAILQ_FOREACH(sp, &swtailq, sw_list)
2375 if (sp->sw_id == cp)
2377 mtx_unlock(&sw_dev_mtx);
2381 swapgeom_close_ev(void *arg, int flags)
2383 struct g_consumer *cp;
2386 g_access(cp, -1, -1, 0);
2388 g_destroy_consumer(cp);
2392 swapgeom_close(struct thread *td, struct swdevt *sw)
2395 /* XXX: direct call when Giant untangled */
2396 g_waitfor_event(swapgeom_close_ev, sw->sw_id, M_WAITOK, NULL);
2407 swapongeom_ev(void *arg, int flags)
2410 struct g_provider *pp;
2411 struct g_consumer *cp;
2412 static struct g_geom *gp;
2419 pp = g_dev_getprovider(swh->dev);
2421 swh->error = ENODEV;
2424 mtx_lock(&sw_dev_mtx);
2425 TAILQ_FOREACH(sp, &swtailq, sw_list) {
2427 if (cp != NULL && cp->provider == pp) {
2428 mtx_unlock(&sw_dev_mtx);
2433 mtx_unlock(&sw_dev_mtx);
2435 gp = g_new_geomf(&g_swap_class, "swap", NULL);
2436 cp = g_new_consumer(gp);
2439 * XXX: Everytime you think you can improve the margin for
2440 * footshooting, somebody depends on the ability to do so:
2441 * savecore(8) wants to write to our swapdev so we cannot
2442 * set an exclusive count :-(
2444 error = g_access(cp, 1, 1, 0);
2447 g_destroy_consumer(cp);
2451 nblks = pp->mediasize / DEV_BSIZE;
2452 swaponsomething(swh->vp, cp, nblks, swapgeom_strategy,
2453 swapgeom_close, dev2udev(swh->dev));
2459 swapongeom(struct thread *td, struct vnode *vp)
2464 vn_lock(vp, LK_EXCLUSIVE | LK_RETRY, td);
2466 swh.dev = vp->v_rdev;
2469 /* XXX: direct call when Giant untangled */
2470 error = g_waitfor_event(swapongeom_ev, &swh, M_WAITOK, NULL);
2473 VOP_UNLOCK(vp, 0, td);
2480 * This is used mainly for network filesystem (read: probably only tested
2481 * with NFS) swapfiles.
2486 swapdev_strategy(struct buf *bp, struct swdevt *sp)
2490 bp->b_blkno = ctodb(bp->b_blkno - sp->sw_first);
2494 if (bp->b_iocmd == BIO_WRITE) {
2496 bufobj_wdrop(bp->b_bufobj);
2497 bufobj_wref(&vp2->v_bufobj);
2499 if (bp->b_bufobj != &vp2->v_bufobj)
2500 bp->b_bufobj = &vp2->v_bufobj;
2502 bp->b_iooffset = dbtob(bp->b_blkno);
2508 swapdev_close(struct thread *td, struct swdevt *sp)
2511 VOP_CLOSE(sp->sw_vp, FREAD | FWRITE, td->td_ucred, td);
2517 swaponvp(struct thread *td, struct vnode *vp, u_long nblks)
2524 mtx_lock(&sw_dev_mtx);
2525 TAILQ_FOREACH(sp, &swtailq, sw_list) {
2526 if (sp->sw_id == vp) {
2527 mtx_unlock(&sw_dev_mtx);
2531 mtx_unlock(&sw_dev_mtx);
2533 (void) vn_lock(vp, LK_EXCLUSIVE | LK_RETRY, td);
2535 error = mac_check_system_swapon(td->td_ucred, vp);
2538 error = VOP_OPEN(vp, FREAD | FWRITE, td->td_ucred, td, NULL);
2539 (void) VOP_UNLOCK(vp, 0, td);
2543 swaponsomething(vp, vp, nblks, swapdev_strategy, swapdev_close,