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>
84 #include <sys/fcntl.h>
85 #include <sys/mount.h>
86 #include <sys/namei.h>
87 #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>
99 #include <vm/vm_map.h>
100 #include <vm/vm_kern.h>
101 #include <vm/vm_object.h>
102 #include <vm/vm_page.h>
103 #include <vm/vm_pager.h>
104 #include <vm/vm_pageout.h>
105 #include <vm/vm_param.h>
106 #include <vm/swap_pager.h>
107 #include <vm/vm_extern.h>
110 #include <geom/geom.h>
113 * SWB_NPAGES must be a power of 2. It may be set to 1, 2, 4, 8, or 16
114 * pages per allocation. We recommend you stick with the default of 8.
115 * The 16-page limit is due to the radix code (kern/subr_blist.c).
117 #ifndef MAX_PAGEOUT_CLUSTER
118 #define MAX_PAGEOUT_CLUSTER 16
121 #if !defined(SWB_NPAGES)
122 #define SWB_NPAGES MAX_PAGEOUT_CLUSTER
126 * Piecemeal swap metadata structure. Swap is stored in a radix tree.
128 * If SWB_NPAGES is 8 and sizeof(char *) == sizeof(daddr_t), our radix
129 * is basically 8. Assuming PAGE_SIZE == 4096, one tree level represents
130 * 32K worth of data, two levels represent 256K, three levels represent
131 * 2 MBytes. This is acceptable.
133 * Overall memory utilization is about the same as the old swap structure.
135 #define SWCORRECT(n) (sizeof(void *) * (n) / sizeof(daddr_t))
136 #define SWAP_META_PAGES (SWB_NPAGES * 2)
137 #define SWAP_META_MASK (SWAP_META_PAGES - 1)
140 struct swblock *swb_hnext;
141 vm_object_t swb_object;
142 vm_pindex_t swb_index;
144 daddr_t swb_pages[SWAP_META_PAGES];
147 static struct mtx sw_dev_mtx;
148 static TAILQ_HEAD(, swdevt) swtailq = TAILQ_HEAD_INITIALIZER(swtailq);
149 static struct swdevt *swdevhd; /* Allocate from here next */
150 static int nswapdev; /* Number of swap devices */
151 int swap_pager_avail;
152 static int swdev_syscall_active = 0; /* serialize swap(on|off) */
154 static void swapdev_strategy(struct buf *, struct swdevt *sw);
156 #define SWM_FREE 0x02 /* free, period */
157 #define SWM_POP 0x04 /* pop out */
159 int swap_pager_full = 2; /* swap space exhaustion (task killing) */
160 static int swap_pager_almost_full = 1; /* swap space exhaustion (w/hysteresis)*/
161 static int nsw_rcount; /* free read buffers */
162 static int nsw_wcount_sync; /* limit write buffers / synchronous */
163 static int nsw_wcount_async; /* limit write buffers / asynchronous */
164 static int nsw_wcount_async_max;/* assigned maximum */
165 static int nsw_cluster_max; /* maximum VOP I/O allowed */
167 static struct swblock **swhash;
168 static int swhash_mask;
169 static struct mtx swhash_mtx;
171 static int swap_async_max = 4; /* maximum in-progress async I/O's */
172 static struct sx sw_alloc_sx;
175 SYSCTL_INT(_vm, OID_AUTO, swap_async_max,
176 CTLFLAG_RW, &swap_async_max, 0, "Maximum running async swap ops");
179 * "named" and "unnamed" anon region objects. Try to reduce the overhead
180 * of searching a named list by hashing it just a little.
185 #define NOBJLIST(handle) \
186 (&swap_pager_object_list[((int)(intptr_t)handle >> 4) & (NOBJLISTS-1)])
188 static struct mtx sw_alloc_mtx; /* protect list manipulation */
189 static struct pagerlst swap_pager_object_list[NOBJLISTS];
190 static uma_zone_t swap_zone;
191 static struct vm_object swap_zone_obj;
194 * pagerops for OBJT_SWAP - "swap pager". Some ops are also global procedure
195 * calls hooked from other parts of the VM system and do not appear here.
196 * (see vm/swap_pager.h).
199 swap_pager_alloc(void *handle, vm_ooffset_t size,
200 vm_prot_t prot, vm_ooffset_t offset);
201 static void swap_pager_dealloc(vm_object_t object);
202 static int swap_pager_getpages(vm_object_t, vm_page_t *, int, int);
203 static void swap_pager_putpages(vm_object_t, vm_page_t *, int, boolean_t, int *);
205 swap_pager_haspage(vm_object_t object, vm_pindex_t pindex, int *before, int *after);
206 static void swap_pager_init(void);
207 static void swap_pager_unswapped(vm_page_t);
208 static void swap_pager_swapoff(struct swdevt *sp);
210 struct pagerops swappagerops = {
211 .pgo_init = swap_pager_init, /* early system initialization of pager */
212 .pgo_alloc = swap_pager_alloc, /* allocate an OBJT_SWAP object */
213 .pgo_dealloc = swap_pager_dealloc, /* deallocate an OBJT_SWAP object */
214 .pgo_getpages = swap_pager_getpages, /* pagein */
215 .pgo_putpages = swap_pager_putpages, /* pageout */
216 .pgo_haspage = swap_pager_haspage, /* get backing store status for page */
217 .pgo_pageunswapped = swap_pager_unswapped, /* remove swap related to page */
221 * dmmax is in page-sized chunks with the new swap system. It was
222 * dev-bsized chunks in the old. dmmax is always a power of 2.
224 * swap_*() routines are externally accessible. swp_*() routines are
228 static int nswap_lowat = 128; /* in pages, swap_pager_almost_full warn */
229 static int nswap_hiwat = 512; /* in pages, swap_pager_almost_full warn */
231 SYSCTL_INT(_vm, OID_AUTO, dmmax,
232 CTLFLAG_RD, &dmmax, 0, "Maximum size of a swap block");
234 static void swp_sizecheck(void);
235 static void swp_pager_async_iodone(struct buf *bp);
236 static int swapongeom(struct thread *, struct vnode *);
237 static int swaponvp(struct thread *, struct vnode *, u_long);
238 static int swapoff_one(struct swdevt *sp, struct thread *td);
241 * Swap bitmap functions
243 static void swp_pager_freeswapspace(daddr_t blk, int npages);
244 static daddr_t swp_pager_getswapspace(int npages);
249 static struct swblock **swp_pager_hash(vm_object_t object, vm_pindex_t index);
250 static void swp_pager_meta_build(vm_object_t, vm_pindex_t, daddr_t);
251 static void swp_pager_meta_free(vm_object_t, vm_pindex_t, daddr_t);
252 static void swp_pager_meta_free_all(vm_object_t);
253 static daddr_t swp_pager_meta_ctl(vm_object_t, vm_pindex_t, int);
256 * SWP_SIZECHECK() - update swap_pager_full indication
258 * update the swap_pager_almost_full indication and warn when we are
259 * about to run out of swap space, using lowat/hiwat hysteresis.
261 * Clear swap_pager_full ( task killing ) indication when lowat is met.
263 * No restrictions on call
264 * This routine may not block.
265 * This routine must be called at splvm()
271 if (swap_pager_avail < nswap_lowat) {
272 if (swap_pager_almost_full == 0) {
273 printf("swap_pager: out of swap space\n");
274 swap_pager_almost_full = 1;
278 if (swap_pager_avail > nswap_hiwat)
279 swap_pager_almost_full = 0;
284 * SWP_PAGER_HASH() - hash swap meta data
286 * This is an helper function which hashes the swapblk given
287 * the object and page index. It returns a pointer to a pointer
288 * to the object, or a pointer to a NULL pointer if it could not
291 * This routine must be called at splvm().
293 static struct swblock **
294 swp_pager_hash(vm_object_t object, vm_pindex_t index)
296 struct swblock **pswap;
297 struct swblock *swap;
299 index &= ~(vm_pindex_t)SWAP_META_MASK;
300 pswap = &swhash[(index ^ (int)(intptr_t)object) & swhash_mask];
301 while ((swap = *pswap) != NULL) {
302 if (swap->swb_object == object &&
303 swap->swb_index == index
307 pswap = &swap->swb_hnext;
313 * SWAP_PAGER_INIT() - initialize the swap pager!
315 * Expected to be started from system init. NOTE: This code is run
316 * before much else so be careful what you depend on. Most of the VM
317 * system has yet to be initialized at this point.
320 swap_pager_init(void)
323 * Initialize object lists
327 for (i = 0; i < NOBJLISTS; ++i)
328 TAILQ_INIT(&swap_pager_object_list[i]);
329 mtx_init(&sw_alloc_mtx, "swap_pager list", NULL, MTX_DEF);
330 mtx_init(&sw_dev_mtx, "swapdev", NULL, MTX_DEF);
333 * Device Stripe, in PAGE_SIZE'd blocks
335 dmmax = SWB_NPAGES * 2;
339 * SWAP_PAGER_SWAP_INIT() - swap pager initialization from pageout process
341 * Expected to be started from pageout process once, prior to entering
345 swap_pager_swap_init(void)
350 * Number of in-transit swap bp operations. Don't
351 * exhaust the pbufs completely. Make sure we
352 * initialize workable values (0 will work for hysteresis
353 * but it isn't very efficient).
355 * The nsw_cluster_max is constrained by the bp->b_pages[]
356 * array (MAXPHYS/PAGE_SIZE) and our locally defined
357 * MAX_PAGEOUT_CLUSTER. Also be aware that swap ops are
358 * constrained by the swap device interleave stripe size.
360 * Currently we hardwire nsw_wcount_async to 4. This limit is
361 * designed to prevent other I/O from having high latencies due to
362 * our pageout I/O. The value 4 works well for one or two active swap
363 * devices but is probably a little low if you have more. Even so,
364 * a higher value would probably generate only a limited improvement
365 * with three or four active swap devices since the system does not
366 * typically have to pageout at extreme bandwidths. We will want
367 * at least 2 per swap devices, and 4 is a pretty good value if you
368 * have one NFS swap device due to the command/ack latency over NFS.
369 * So it all works out pretty well.
371 nsw_cluster_max = min((MAXPHYS/PAGE_SIZE), MAX_PAGEOUT_CLUSTER);
374 nsw_rcount = (nswbuf + 1) / 2;
375 nsw_wcount_sync = (nswbuf + 3) / 4;
376 nsw_wcount_async = 4;
377 nsw_wcount_async_max = nsw_wcount_async;
378 mtx_unlock(&pbuf_mtx);
381 * Initialize our zone. Right now I'm just guessing on the number
382 * we need based on the number of pages in the system. Each swblock
383 * can hold 16 pages, so this is probably overkill. This reservation
384 * is typically limited to around 32MB by default.
386 n = cnt.v_page_count / 2;
387 if (maxswzone && n > maxswzone / sizeof(struct swblock))
388 n = maxswzone / sizeof(struct swblock);
390 swap_zone = uma_zcreate("SWAPMETA", sizeof(struct swblock), NULL, NULL,
391 NULL, NULL, UMA_ALIGN_PTR, UMA_ZONE_NOFREE | UMA_ZONE_VM);
392 if (swap_zone == NULL)
393 panic("failed to create swap_zone.");
395 if (uma_zone_set_obj(swap_zone, &swap_zone_obj, n))
398 * if the allocation failed, try a zone two thirds the
399 * size of the previous attempt.
404 printf("Swap zone entries reduced from %d to %d.\n", n2, n);
408 * Initialize our meta-data hash table. The swapper does not need to
409 * be quite as efficient as the VM system, so we do not use an
410 * oversized hash table.
412 * n: size of hash table, must be power of 2
413 * swhash_mask: hash table index mask
415 for (n = 1; n < n2 / 8; n *= 2)
417 swhash = malloc(sizeof(struct swblock *) * n, M_VMPGDATA, M_WAITOK | M_ZERO);
419 mtx_init(&swhash_mtx, "swap_pager swhash", NULL, MTX_DEF);
423 * SWAP_PAGER_ALLOC() - allocate a new OBJT_SWAP VM object and instantiate
424 * its metadata structures.
426 * This routine is called from the mmap and fork code to create a new
427 * OBJT_SWAP object. We do this by creating an OBJT_DEFAULT object
428 * and then converting it with swp_pager_meta_build().
430 * This routine may block in vm_object_allocate() and create a named
431 * object lookup race, so we must interlock. We must also run at
432 * splvm() for the object lookup to handle races with interrupts, but
433 * we do not have to maintain splvm() in between the lookup and the
434 * add because (I believe) it is not possible to attempt to create
435 * a new swap object w/handle when a default object with that handle
441 swap_pager_alloc(void *handle, vm_ooffset_t size, vm_prot_t prot,
447 pindex = OFF_TO_IDX(offset + PAGE_MASK + size);
452 * Reference existing named region or allocate new one. There
453 * should not be a race here against swp_pager_meta_build()
454 * as called from vm_page_remove() in regards to the lookup
457 sx_xlock(&sw_alloc_sx);
458 object = vm_pager_object_lookup(NOBJLIST(handle), handle);
460 if (object != NULL) {
461 vm_object_reference(object);
463 object = vm_object_allocate(OBJT_DEFAULT, pindex);
464 object->handle = handle;
466 VM_OBJECT_LOCK(object);
467 swp_pager_meta_build(object, 0, SWAPBLK_NONE);
468 VM_OBJECT_UNLOCK(object);
470 sx_xunlock(&sw_alloc_sx);
473 object = vm_object_allocate(OBJT_DEFAULT, pindex);
475 VM_OBJECT_LOCK(object);
476 swp_pager_meta_build(object, 0, SWAPBLK_NONE);
477 VM_OBJECT_UNLOCK(object);
483 * SWAP_PAGER_DEALLOC() - remove swap metadata from object
485 * The swap backing for the object is destroyed. The code is
486 * designed such that we can reinstantiate it later, but this
487 * routine is typically called only when the entire object is
488 * about to be destroyed.
490 * This routine may block, but no longer does.
492 * The object must be locked or unreferenceable.
495 swap_pager_dealloc(vm_object_t object)
499 * Remove from list right away so lookups will fail if we block for
500 * pageout completion.
502 if (object->handle != NULL) {
503 mtx_lock(&sw_alloc_mtx);
504 TAILQ_REMOVE(NOBJLIST(object->handle), object, pager_object_list);
505 mtx_unlock(&sw_alloc_mtx);
508 VM_OBJECT_LOCK_ASSERT(object, MA_OWNED);
509 vm_object_pip_wait(object, "swpdea");
512 * Free all remaining metadata. We only bother to free it from
513 * the swap meta data. We do not attempt to free swapblk's still
514 * associated with vm_page_t's for this object. We do not care
515 * if paging is still in progress on some objects.
517 swp_pager_meta_free_all(object);
520 /************************************************************************
521 * SWAP PAGER BITMAP ROUTINES *
522 ************************************************************************/
525 * SWP_PAGER_GETSWAPSPACE() - allocate raw swap space
527 * Allocate swap for the requested number of pages. The starting
528 * swap block number (a page index) is returned or SWAPBLK_NONE
529 * if the allocation failed.
531 * Also has the side effect of advising that somebody made a mistake
532 * when they configured swap and didn't configure enough.
534 * Must be called at splvm() to avoid races with bitmap frees from
535 * vm_page_remove() aka swap_pager_page_removed().
537 * This routine may not block
538 * This routine must be called at splvm().
540 * We allocate in round-robin fashion from the configured devices.
543 swp_pager_getswapspace(int npages)
550 mtx_lock(&sw_dev_mtx);
552 for (i = 0; i < nswapdev; i++) {
554 sp = TAILQ_FIRST(&swtailq);
555 if (!(sp->sw_flags & SW_CLOSING)) {
556 blk = blist_alloc(sp->sw_blist, npages);
557 if (blk != SWAPBLK_NONE) {
559 sp->sw_used += npages;
560 swap_pager_avail -= npages;
562 swdevhd = TAILQ_NEXT(sp, sw_list);
566 sp = TAILQ_NEXT(sp, sw_list);
568 if (swap_pager_full != 2) {
569 printf("swap_pager_getswapspace(%d): failed\n", npages);
571 swap_pager_almost_full = 1;
575 mtx_unlock(&sw_dev_mtx);
580 swp_pager_isondev(daddr_t blk, struct swdevt *sp)
583 return (blk >= sp->sw_first && blk < sp->sw_end);
587 swp_pager_strategy(struct buf *bp)
591 mtx_lock(&sw_dev_mtx);
592 TAILQ_FOREACH(sp, &swtailq, sw_list) {
593 if (bp->b_blkno >= sp->sw_first && bp->b_blkno < sp->sw_end) {
594 mtx_unlock(&sw_dev_mtx);
595 sp->sw_strategy(bp, sp);
599 panic("Swapdev not found");
604 * SWP_PAGER_FREESWAPSPACE() - free raw swap space
606 * This routine returns the specified swap blocks back to the bitmap.
608 * Note: This routine may not block (it could in the old swap code),
609 * and through the use of the new blist routines it does not block.
611 * We must be called at splvm() to avoid races with bitmap frees from
612 * vm_page_remove() aka swap_pager_page_removed().
614 * This routine may not block
615 * This routine must be called at splvm().
618 swp_pager_freeswapspace(daddr_t blk, int npages)
622 mtx_lock(&sw_dev_mtx);
623 TAILQ_FOREACH(sp, &swtailq, sw_list) {
624 if (blk >= sp->sw_first && blk < sp->sw_end) {
625 sp->sw_used -= npages;
627 * If we are attempting to stop swapping on
628 * this device, we don't want to mark any
629 * blocks free lest they be reused.
631 if ((sp->sw_flags & SW_CLOSING) == 0) {
632 blist_free(sp->sw_blist, blk - sp->sw_first,
634 swap_pager_avail += npages;
637 mtx_unlock(&sw_dev_mtx);
641 panic("Swapdev not found");
645 * SWAP_PAGER_FREESPACE() - frees swap blocks associated with a page
646 * range within an object.
648 * This is a globally accessible routine.
650 * This routine removes swapblk assignments from swap metadata.
652 * The external callers of this routine typically have already destroyed
653 * or renamed vm_page_t's associated with this range in the object so
656 * This routine may be called at any spl. We up our spl to splvm temporarily
657 * in order to perform the metadata removal.
660 swap_pager_freespace(vm_object_t object, vm_pindex_t start, vm_size_t size)
663 VM_OBJECT_LOCK_ASSERT(object, MA_OWNED);
664 swp_pager_meta_free(object, start, size);
668 * SWAP_PAGER_RESERVE() - reserve swap blocks in object
670 * Assigns swap blocks to the specified range within the object. The
671 * swap blocks are not zerod. Any previous swap assignment is destroyed.
673 * Returns 0 on success, -1 on failure.
676 swap_pager_reserve(vm_object_t object, vm_pindex_t start, vm_size_t size)
679 daddr_t blk = SWAPBLK_NONE;
680 vm_pindex_t beg = start; /* save start index */
682 VM_OBJECT_LOCK(object);
686 while ((blk = swp_pager_getswapspace(n)) == SWAPBLK_NONE) {
689 swp_pager_meta_free(object, beg, start - beg);
690 VM_OBJECT_UNLOCK(object);
695 swp_pager_meta_build(object, start, blk);
701 swp_pager_meta_free(object, start, n);
702 VM_OBJECT_UNLOCK(object);
707 * SWAP_PAGER_COPY() - copy blocks from source pager to destination pager
708 * and destroy the source.
710 * Copy any valid swapblks from the source to the destination. In
711 * cases where both the source and destination have a valid swapblk,
712 * we keep the destination's.
714 * This routine is allowed to block. It may block allocating metadata
715 * indirectly through swp_pager_meta_build() or if paging is still in
716 * progress on the source.
718 * This routine can be called at any spl
720 * XXX vm_page_collapse() kinda expects us not to block because we
721 * supposedly do not need to allocate memory, but for the moment we
722 * *may* have to get a little memory from the zone allocator, but
723 * it is taken from the interrupt memory. We should be ok.
725 * The source object contains no vm_page_t's (which is just as well)
727 * The source object is of type OBJT_SWAP.
729 * The source and destination objects must be locked or
730 * inaccessible (XXX are they ?)
733 swap_pager_copy(vm_object_t srcobject, vm_object_t dstobject,
734 vm_pindex_t offset, int destroysource)
738 VM_OBJECT_LOCK_ASSERT(srcobject, MA_OWNED);
739 VM_OBJECT_LOCK_ASSERT(dstobject, MA_OWNED);
742 * If destroysource is set, we remove the source object from the
743 * swap_pager internal queue now.
746 if (srcobject->handle != NULL) {
747 mtx_lock(&sw_alloc_mtx);
749 NOBJLIST(srcobject->handle),
753 mtx_unlock(&sw_alloc_mtx);
758 * transfer source to destination.
760 for (i = 0; i < dstobject->size; ++i) {
764 * Locate (without changing) the swapblk on the destination,
765 * unless it is invalid in which case free it silently, or
766 * if the destination is a resident page, in which case the
767 * source is thrown away.
769 dstaddr = swp_pager_meta_ctl(dstobject, i, 0);
771 if (dstaddr == SWAPBLK_NONE) {
773 * Destination has no swapblk and is not resident,
778 srcaddr = swp_pager_meta_ctl(
784 if (srcaddr != SWAPBLK_NONE) {
786 * swp_pager_meta_build() can sleep.
788 vm_object_pip_add(srcobject, 1);
789 VM_OBJECT_UNLOCK(srcobject);
790 vm_object_pip_add(dstobject, 1);
791 swp_pager_meta_build(dstobject, i, srcaddr);
792 vm_object_pip_wakeup(dstobject);
793 VM_OBJECT_LOCK(srcobject);
794 vm_object_pip_wakeup(srcobject);
798 * Destination has valid swapblk or it is represented
799 * by a resident page. We destroy the sourceblock.
802 swp_pager_meta_ctl(srcobject, i + offset, SWM_FREE);
807 * Free left over swap blocks in source.
809 * We have to revert the type to OBJT_DEFAULT so we do not accidently
810 * double-remove the object from the swap queues.
813 swp_pager_meta_free_all(srcobject);
815 * Reverting the type is not necessary, the caller is going
816 * to destroy srcobject directly, but I'm doing it here
817 * for consistency since we've removed the object from its
820 srcobject->type = OBJT_DEFAULT;
825 * SWAP_PAGER_HASPAGE() - determine if we have good backing store for
826 * the requested page.
828 * We determine whether good backing store exists for the requested
829 * page and return TRUE if it does, FALSE if it doesn't.
831 * If TRUE, we also try to determine how much valid, contiguous backing
832 * store exists before and after the requested page within a reasonable
833 * distance. We do not try to restrict it to the swap device stripe
834 * (that is handled in getpages/putpages). It probably isn't worth
838 swap_pager_haspage(vm_object_t object, vm_pindex_t pindex, int *before, int *after)
842 VM_OBJECT_LOCK_ASSERT(object, MA_OWNED);
844 * do we have good backing store at the requested index ?
846 blk0 = swp_pager_meta_ctl(object, pindex, 0);
848 if (blk0 == SWAPBLK_NONE) {
857 * find backwards-looking contiguous good backing store
859 if (before != NULL) {
862 for (i = 1; i < (SWB_NPAGES/2); ++i) {
867 blk = swp_pager_meta_ctl(object, pindex - i, 0);
875 * find forward-looking contiguous good backing store
880 for (i = 1; i < (SWB_NPAGES/2); ++i) {
883 blk = swp_pager_meta_ctl(object, pindex + i, 0);
893 * SWAP_PAGER_PAGE_UNSWAPPED() - remove swap backing store related to page
895 * This removes any associated swap backing store, whether valid or
896 * not, from the page.
898 * This routine is typically called when a page is made dirty, at
899 * which point any associated swap can be freed. MADV_FREE also
900 * calls us in a special-case situation
902 * NOTE!!! If the page is clean and the swap was valid, the caller
903 * should make the page dirty before calling this routine. This routine
904 * does NOT change the m->dirty status of the page. Also: MADV_FREE
907 * This routine may not block
908 * This routine must be called at splvm()
911 swap_pager_unswapped(vm_page_t m)
914 VM_OBJECT_LOCK_ASSERT(m->object, MA_OWNED);
915 swp_pager_meta_ctl(m->object, m->pindex, SWM_FREE);
919 * SWAP_PAGER_GETPAGES() - bring pages in from swap
921 * Attempt to retrieve (m, count) pages from backing store, but make
922 * sure we retrieve at least m[reqpage]. We try to load in as large
923 * a chunk surrounding m[reqpage] as is contiguous in swap and which
924 * belongs to the same object.
926 * The code is designed for asynchronous operation and
927 * immediate-notification of 'reqpage' but tends not to be
928 * used that way. Please do not optimize-out this algorithmic
929 * feature, I intend to improve on it in the future.
931 * The parent has a single vm_object_pip_add() reference prior to
932 * calling us and we should return with the same.
934 * The parent has BUSY'd the pages. We should return with 'm'
935 * left busy, but the others adjusted.
938 swap_pager_getpages(vm_object_t object, vm_page_t *m, int count, int reqpage)
948 KASSERT(mreq->object == object,
949 ("swap_pager_getpages: object mismatch %p/%p",
950 object, mreq->object));
953 * Calculate range to retrieve. The pages have already been assigned
954 * their swapblks. We require a *contiguous* range but we know it to
955 * not span devices. If we do not supply it, bad things
956 * happen. Note that blk, iblk & jblk can be SWAPBLK_NONE, but the
957 * loops are set up such that the case(s) are handled implicitly.
959 * The swp_*() calls must be made at splvm(). vm_page_free() does
960 * not need to be, but it will go a little faster if it is.
962 blk = swp_pager_meta_ctl(mreq->object, mreq->pindex, 0);
964 for (i = reqpage - 1; i >= 0; --i) {
967 iblk = swp_pager_meta_ctl(m[i]->object, m[i]->pindex, 0);
968 if (blk != iblk + (reqpage - i))
973 for (j = reqpage + 1; j < count; ++j) {
976 jblk = swp_pager_meta_ctl(m[j]->object, m[j]->pindex, 0);
977 if (blk != jblk - (j - reqpage))
982 * free pages outside our collection range. Note: we never free
983 * mreq, it must remain busy throughout.
985 if (0 < i || j < count) {
988 vm_page_lock_queues();
989 for (k = 0; k < i; ++k)
991 for (k = j; k < count; ++k)
993 vm_page_unlock_queues();
997 * Return VM_PAGER_FAIL if we have nothing to do. Return mreq
998 * still busy, but the others unbusied.
1000 if (blk == SWAPBLK_NONE)
1001 return (VM_PAGER_FAIL);
1004 * Getpbuf() can sleep.
1006 VM_OBJECT_UNLOCK(object);
1008 * Get a swap buffer header to perform the IO
1010 bp = getpbuf(&nsw_rcount);
1011 bp->b_flags |= B_PAGING;
1014 * map our page(s) into kva for input
1016 pmap_qenter((vm_offset_t)bp->b_data, m + i, j - i);
1018 bp->b_iocmd = BIO_READ;
1019 bp->b_iodone = swp_pager_async_iodone;
1020 bp->b_rcred = crhold(thread0.td_ucred);
1021 bp->b_wcred = crhold(thread0.td_ucred);
1022 bp->b_blkno = blk - (reqpage - i);
1023 bp->b_bcount = PAGE_SIZE * (j - i);
1024 bp->b_bufsize = PAGE_SIZE * (j - i);
1025 bp->b_pager.pg_reqpage = reqpage - i;
1027 VM_OBJECT_LOCK(object);
1028 vm_page_lock_queues();
1032 for (k = i; k < j; ++k) {
1033 bp->b_pages[k - i] = m[k];
1034 vm_page_flag_set(m[k], PG_SWAPINPROG);
1037 vm_page_unlock_queues();
1038 bp->b_npages = j - i;
1041 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. PG_SWAPINPROG is always
1066 * cleared on completion. If an I/O error occurs, SWAPBLK_NONE
1067 * is set in the meta-data.
1069 vm_page_lock_queues();
1070 while ((mreq->flags & PG_SWAPINPROG) != 0) {
1071 vm_page_flag_set(mreq, PG_WANTED | PG_REFERENCED);
1073 if (msleep(mreq, &vm_page_queue_mtx, PSWP, "swread", hz*20)) {
1075 "swap_pager: indefinite wait buffer: bufobj: %p, blkno: %jd, size: %ld\n",
1076 bp->b_bufobj, (intmax_t)bp->b_blkno, bp->b_bcount);
1079 vm_page_unlock_queues();
1081 VM_OBJECT_LOCK(object);
1083 * mreq is left busied after completion, but all the other pages
1084 * are freed. If we had an unrecoverable read error the page will
1087 if (mreq->valid != VM_PAGE_BITS_ALL) {
1088 return (VM_PAGER_ERROR);
1090 return (VM_PAGER_OK);
1094 * A final note: in a low swap situation, we cannot deallocate swap
1095 * and mark a page dirty here because the caller is likely to mark
1096 * the page clean when we return, causing the page to possibly revert
1097 * to all-zero's later.
1102 * swap_pager_putpages:
1104 * Assign swap (if necessary) and initiate I/O on the specified pages.
1106 * We support both OBJT_DEFAULT and OBJT_SWAP objects. DEFAULT objects
1107 * are automatically converted to SWAP objects.
1109 * In a low memory situation we may block in VOP_STRATEGY(), but the new
1110 * vm_page reservation system coupled with properly written VFS devices
1111 * should ensure that no low-memory deadlock occurs. This is an area
1114 * The parent has N vm_object_pip_add() references prior to
1115 * calling us and will remove references for rtvals[] that are
1116 * not set to VM_PAGER_PEND. We need to remove the rest on I/O
1119 * The parent has soft-busy'd the pages it passes us and will unbusy
1120 * those whos rtvals[] entry is not set to VM_PAGER_PEND on return.
1121 * We need to unbusy the rest on I/O completion.
1124 swap_pager_putpages(vm_object_t object, vm_page_t *m, int count,
1125 boolean_t sync, int *rtvals)
1130 if (count && m[0]->object != object) {
1131 panic("swap_pager_getpages: object mismatch %p/%p",
1140 * Turn object into OBJT_SWAP
1141 * check for bogus sysops
1142 * force sync if not pageout process
1144 if (object->type != OBJT_SWAP)
1145 swp_pager_meta_build(object, 0, SWAPBLK_NONE);
1146 VM_OBJECT_UNLOCK(object);
1148 if (curproc != pageproc)
1154 * Update nsw parameters from swap_async_max sysctl values.
1155 * Do not let the sysop crash the machine with bogus numbers.
1157 mtx_lock(&pbuf_mtx);
1158 if (swap_async_max != nsw_wcount_async_max) {
1164 if ((n = swap_async_max) > nswbuf / 2)
1171 * Adjust difference ( if possible ). If the current async
1172 * count is too low, we may not be able to make the adjustment
1175 n -= nsw_wcount_async_max;
1176 if (nsw_wcount_async + n >= 0) {
1177 nsw_wcount_async += n;
1178 nsw_wcount_async_max += n;
1179 wakeup(&nsw_wcount_async);
1182 mtx_unlock(&pbuf_mtx);
1187 * Assign swap blocks and issue I/O. We reallocate swap on the fly.
1188 * The page is left dirty until the pageout operation completes
1191 for (i = 0; i < count; i += n) {
1197 * Maximum I/O size is limited by a number of factors.
1199 n = min(BLIST_MAX_ALLOC, count - i);
1200 n = min(n, nsw_cluster_max);
1203 * Get biggest block of swap we can. If we fail, fall
1204 * back and try to allocate a smaller block. Don't go
1205 * overboard trying to allocate space if it would overly
1209 (blk = swp_pager_getswapspace(n)) == SWAPBLK_NONE &&
1214 if (blk == SWAPBLK_NONE) {
1215 for (j = 0; j < n; ++j)
1216 rtvals[i+j] = VM_PAGER_FAIL;
1221 * All I/O parameters have been satisfied, build the I/O
1222 * request and assign the swap space.
1225 bp = getpbuf(&nsw_wcount_sync);
1227 bp = getpbuf(&nsw_wcount_async);
1228 bp->b_flags = B_ASYNC;
1230 bp->b_flags |= B_PAGING;
1231 bp->b_iocmd = BIO_WRITE;
1233 pmap_qenter((vm_offset_t)bp->b_data, &m[i], n);
1235 bp->b_rcred = crhold(thread0.td_ucred);
1236 bp->b_wcred = crhold(thread0.td_ucred);
1237 bp->b_bcount = PAGE_SIZE * n;
1238 bp->b_bufsize = PAGE_SIZE * n;
1241 VM_OBJECT_LOCK(object);
1242 for (j = 0; j < n; ++j) {
1243 vm_page_t mreq = m[i+j];
1245 swp_pager_meta_build(
1250 vm_page_dirty(mreq);
1251 rtvals[i+j] = VM_PAGER_OK;
1253 vm_page_lock_queues();
1254 vm_page_flag_set(mreq, PG_SWAPINPROG);
1255 vm_page_unlock_queues();
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;
1267 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 vm_page_flag_clear(m, PG_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 PG_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.
1396 * XXX IT IS NOT LEGAL TO FREE THE PAGE HERE
1397 * AS THIS MESSES WITH object->memq, and it is
1398 * not legal to mess with object->memq from an
1402 if (i != bp->b_pager.pg_reqpage)
1407 * If i == bp->b_pager.pg_reqpage, do not wake
1408 * the page up. The caller needs to.
1412 * If a write error occurs, reactivate page
1413 * so it doesn't clog the inactive list,
1414 * then finish the I/O.
1417 vm_page_activate(m);
1418 vm_page_io_finish(m);
1420 } else if (bp->b_iocmd == BIO_READ) {
1422 * For read success, clear dirty bits. Nobody should
1423 * have this page mapped but don't take any chances,
1424 * make sure the pmap modify bits are also cleared.
1426 * NOTE: for reads, m->dirty will probably be
1427 * overridden by the original caller of getpages so
1428 * we cannot set them in order to free the underlying
1429 * swap in a low-swap situation. I don't think we'd
1430 * want to do that anyway, but it was an optimization
1431 * that existed in the old swapper for a time before
1432 * it got ripped out due to precisely this problem.
1434 * If not the requested page then deactivate it.
1436 * Note that the requested page, reqpage, is left
1437 * busied, but we still have to wake it up. The
1438 * other pages are released (unbusied) by
1439 * vm_page_wakeup(). We do not set reqpage's
1440 * valid bits here, it is up to the caller.
1442 pmap_clear_modify(m);
1443 m->valid = VM_PAGE_BITS_ALL;
1447 * We have to wake specifically requested pages
1448 * up too because we cleared PG_SWAPINPROG and
1449 * could be waiting for it in getpages. However,
1450 * be sure to not unbusy getpages specifically
1451 * requested page - getpages expects it to be
1454 if (i != bp->b_pager.pg_reqpage) {
1455 vm_page_deactivate(m);
1462 * For write success, clear the modify and dirty
1463 * status, then finish the I/O ( which decrements the
1464 * busy count and possibly wakes waiter's up ).
1466 pmap_clear_modify(m);
1468 vm_page_io_finish(m);
1469 if (vm_page_count_severe())
1470 vm_page_try_to_cache(m);
1473 vm_page_unlock_queues();
1476 * adjust pip. NOTE: the original parent may still have its own
1477 * pip refs on the object.
1479 if (object != NULL) {
1480 vm_object_pip_wakeupn(object, bp->b_npages);
1481 VM_OBJECT_UNLOCK(object);
1484 * swapdev_strategy() manually sets b_vp and b_bufobj before calling
1485 * bstrategy(). Set them back to NULL now we're done with it, or we'll
1486 * trigger a KASSERT in relpbuf().
1490 bp->b_bufobj = NULL;
1494 * release the physical I/O buffer
1498 ((bp->b_iocmd == BIO_READ) ? &nsw_rcount :
1499 ((bp->b_flags & B_ASYNC) ?
1508 * swap_pager_isswapped:
1510 * Return 1 if at least one page in the given object is paged
1511 * out to the given swap device.
1513 * This routine may not block.
1516 swap_pager_isswapped(vm_object_t object, struct swdevt *sp)
1522 VM_OBJECT_LOCK_ASSERT(object, MA_OWNED);
1523 if (object->type != OBJT_SWAP)
1526 mtx_lock(&swhash_mtx);
1527 for (bcount = 0; bcount < object->un_pager.swp.swp_bcount; bcount++) {
1528 struct swblock *swap;
1530 if ((swap = *swp_pager_hash(object, index)) != NULL) {
1531 for (i = 0; i < SWAP_META_PAGES; ++i) {
1532 if (swp_pager_isondev(swap->swb_pages[i], sp)) {
1533 mtx_unlock(&swhash_mtx);
1538 index += SWAP_META_PAGES;
1539 if (index > 0x20000000)
1540 panic("swap_pager_isswapped: failed to locate all swap meta blocks");
1542 mtx_unlock(&swhash_mtx);
1547 * SWP_PAGER_FORCE_PAGEIN() - force a swap block to be paged in
1549 * This routine dissociates the page at the given index within a
1550 * swap block from its backing store, paging it in if necessary.
1551 * If the page is paged in, it is placed in the inactive queue,
1552 * since it had its backing store ripped out from under it.
1553 * We also attempt to swap in all other pages in the swap block,
1554 * we only guarantee that the one at the specified index is
1557 * XXX - The code to page the whole block in doesn't work, so we
1558 * revert to the one-by-one behavior for now. Sigh.
1560 static __inline void
1561 swp_pager_force_pagein(vm_object_t object, vm_pindex_t pindex)
1565 vm_object_pip_add(object, 1);
1566 m = vm_page_grab(object, pindex, VM_ALLOC_NORMAL|VM_ALLOC_RETRY);
1567 if (m->valid == VM_PAGE_BITS_ALL) {
1568 vm_object_pip_subtract(object, 1);
1569 vm_page_lock_queues();
1570 vm_page_activate(m);
1573 vm_page_unlock_queues();
1574 vm_pager_page_unswapped(m);
1578 if (swap_pager_getpages(object, &m, 1, 0) != VM_PAGER_OK)
1579 panic("swap_pager_force_pagein: read from swap failed");/*XXX*/
1580 vm_object_pip_subtract(object, 1);
1581 vm_page_lock_queues();
1583 vm_page_dontneed(m);
1585 vm_page_unlock_queues();
1586 vm_pager_page_unswapped(m);
1590 * swap_pager_swapoff:
1592 * Page in all of the pages that have been paged out to the
1593 * given device. The corresponding blocks in the bitmap must be
1594 * marked as allocated and the device must be flagged SW_CLOSING.
1595 * There may be no processes swapped out to the device.
1597 * This routine may block.
1600 swap_pager_swapoff(struct swdevt *sp)
1602 struct swblock *swap;
1609 mtx_lock(&swhash_mtx);
1610 for (i = 0; i <= swhash_mask; i++) { /* '<=' is correct here */
1612 for (swap = swhash[i]; swap != NULL; swap = swap->swb_hnext) {
1613 vm_object_t object = swap->swb_object;
1614 vm_pindex_t pindex = swap->swb_index;
1615 for (j = 0; j < SWAP_META_PAGES; ++j) {
1616 if (swp_pager_isondev(swap->swb_pages[j], sp)) {
1617 /* avoid deadlock */
1618 if (!VM_OBJECT_TRYLOCK(object)) {
1621 mtx_unlock(&swhash_mtx);
1622 swp_pager_force_pagein(object,
1624 VM_OBJECT_UNLOCK(object);
1625 mtx_lock(&swhash_mtx);
1632 mtx_unlock(&swhash_mtx);
1636 * Objects may be locked or paging to the device being
1637 * removed, so we will miss their pages and need to
1638 * make another pass. We have marked this device as
1639 * SW_CLOSING, so the activity should finish soon.
1642 if (retries > 100) {
1643 panic("swapoff: failed to locate %d swap blocks",
1646 tsleep(&dummy, PVM, "swpoff", hz / 20);
1651 /************************************************************************
1653 ************************************************************************
1655 * These routines manipulate the swap metadata stored in the
1656 * OBJT_SWAP object. All swp_*() routines must be called at
1657 * splvm() because swap can be freed up by the low level vm_page
1658 * code which might be called from interrupts beyond what splbio() covers.
1660 * Swap metadata is implemented with a global hash and not directly
1661 * linked into the object. Instead the object simply contains
1662 * appropriate tracking counters.
1666 * SWP_PAGER_META_BUILD() - add swap block to swap meta data for object
1668 * We first convert the object to a swap object if it is a default
1671 * The specified swapblk is added to the object's swap metadata. If
1672 * the swapblk is not valid, it is freed instead. Any previously
1673 * assigned swapblk is freed.
1675 * This routine must be called at splvm(), except when used to convert
1676 * an OBJT_DEFAULT object into an OBJT_SWAP object.
1679 swp_pager_meta_build(vm_object_t object, vm_pindex_t pindex, daddr_t swapblk)
1681 struct swblock *swap;
1682 struct swblock **pswap;
1685 VM_OBJECT_LOCK_ASSERT(object, MA_OWNED);
1687 * Convert default object to swap object if necessary
1689 if (object->type != OBJT_SWAP) {
1690 object->type = OBJT_SWAP;
1691 object->un_pager.swp.swp_bcount = 0;
1693 if (object->handle != NULL) {
1694 mtx_lock(&sw_alloc_mtx);
1696 NOBJLIST(object->handle),
1700 mtx_unlock(&sw_alloc_mtx);
1705 * Locate hash entry. If not found create, but if we aren't adding
1706 * anything just return. If we run out of space in the map we wait
1707 * and, since the hash table may have changed, retry.
1710 mtx_lock(&swhash_mtx);
1711 pswap = swp_pager_hash(object, pindex);
1713 if ((swap = *pswap) == NULL) {
1716 if (swapblk == SWAPBLK_NONE)
1719 swap = *pswap = uma_zalloc(swap_zone, M_NOWAIT);
1721 mtx_unlock(&swhash_mtx);
1722 VM_OBJECT_UNLOCK(object);
1723 if (uma_zone_exhausted(swap_zone))
1724 printf("swap zone exhausted, increase kern.maxswzone\n");
1726 VM_OBJECT_LOCK(object);
1730 swap->swb_hnext = NULL;
1731 swap->swb_object = object;
1732 swap->swb_index = pindex & ~(vm_pindex_t)SWAP_META_MASK;
1733 swap->swb_count = 0;
1735 ++object->un_pager.swp.swp_bcount;
1737 for (i = 0; i < SWAP_META_PAGES; ++i)
1738 swap->swb_pages[i] = SWAPBLK_NONE;
1742 * Delete prior contents of metadata
1744 idx = pindex & SWAP_META_MASK;
1746 if (swap->swb_pages[idx] != SWAPBLK_NONE) {
1747 swp_pager_freeswapspace(swap->swb_pages[idx], 1);
1752 * Enter block into metadata
1754 swap->swb_pages[idx] = swapblk;
1755 if (swapblk != SWAPBLK_NONE)
1758 mtx_unlock(&swhash_mtx);
1762 * SWP_PAGER_META_FREE() - free a range of blocks in the object's swap metadata
1764 * The requested range of blocks is freed, with any associated swap
1765 * returned to the swap bitmap.
1767 * This routine will free swap metadata structures as they are cleaned
1768 * out. This routine does *NOT* operate on swap metadata associated
1769 * with resident pages.
1771 * This routine must be called at splvm()
1774 swp_pager_meta_free(vm_object_t object, vm_pindex_t index, daddr_t count)
1777 VM_OBJECT_LOCK_ASSERT(object, MA_OWNED);
1778 if (object->type != OBJT_SWAP)
1782 struct swblock **pswap;
1783 struct swblock *swap;
1785 mtx_lock(&swhash_mtx);
1786 pswap = swp_pager_hash(object, index);
1788 if ((swap = *pswap) != NULL) {
1789 daddr_t v = swap->swb_pages[index & SWAP_META_MASK];
1791 if (v != SWAPBLK_NONE) {
1792 swp_pager_freeswapspace(v, 1);
1793 swap->swb_pages[index & SWAP_META_MASK] =
1795 if (--swap->swb_count == 0) {
1796 *pswap = swap->swb_hnext;
1797 uma_zfree(swap_zone, swap);
1798 --object->un_pager.swp.swp_bcount;
1804 int n = SWAP_META_PAGES - (index & SWAP_META_MASK);
1808 mtx_unlock(&swhash_mtx);
1813 * SWP_PAGER_META_FREE_ALL() - destroy all swap metadata associated with object
1815 * This routine locates and destroys all swap metadata associated with
1818 * This routine must be called at splvm()
1821 swp_pager_meta_free_all(vm_object_t object)
1825 VM_OBJECT_LOCK_ASSERT(object, MA_OWNED);
1826 if (object->type != OBJT_SWAP)
1829 while (object->un_pager.swp.swp_bcount) {
1830 struct swblock **pswap;
1831 struct swblock *swap;
1833 mtx_lock(&swhash_mtx);
1834 pswap = swp_pager_hash(object, index);
1835 if ((swap = *pswap) != NULL) {
1838 for (i = 0; i < SWAP_META_PAGES; ++i) {
1839 daddr_t v = swap->swb_pages[i];
1840 if (v != SWAPBLK_NONE) {
1842 swp_pager_freeswapspace(v, 1);
1845 if (swap->swb_count != 0)
1846 panic("swap_pager_meta_free_all: swb_count != 0");
1847 *pswap = swap->swb_hnext;
1848 uma_zfree(swap_zone, swap);
1849 --object->un_pager.swp.swp_bcount;
1851 mtx_unlock(&swhash_mtx);
1852 index += SWAP_META_PAGES;
1853 if (index > 0x20000000)
1854 panic("swp_pager_meta_free_all: failed to locate all swap meta blocks");
1859 * SWP_PAGER_METACTL() - misc control of swap and vm_page_t meta data.
1861 * This routine is capable of looking up, popping, or freeing
1862 * swapblk assignments in the swap meta data or in the vm_page_t.
1863 * The routine typically returns the swapblk being looked-up, or popped,
1864 * or SWAPBLK_NONE if the block was freed, or SWAPBLK_NONE if the block
1865 * was invalid. This routine will automatically free any invalid
1866 * meta-data swapblks.
1868 * It is not possible to store invalid swapblks in the swap meta data
1869 * (other then a literal 'SWAPBLK_NONE'), so we don't bother checking.
1871 * When acting on a busy resident page and paging is in progress, we
1872 * have to wait until paging is complete but otherwise can act on the
1875 * This routine must be called at splvm().
1877 * SWM_FREE remove and free swap block from metadata
1878 * SWM_POP remove from meta data but do not free.. pop it out
1881 swp_pager_meta_ctl(vm_object_t object, vm_pindex_t pindex, int flags)
1883 struct swblock **pswap;
1884 struct swblock *swap;
1888 VM_OBJECT_LOCK_ASSERT(object, MA_OWNED);
1890 * The meta data only exists of the object is OBJT_SWAP
1891 * and even then might not be allocated yet.
1893 if (object->type != OBJT_SWAP)
1894 return (SWAPBLK_NONE);
1897 mtx_lock(&swhash_mtx);
1898 pswap = swp_pager_hash(object, pindex);
1900 if ((swap = *pswap) != NULL) {
1901 idx = pindex & SWAP_META_MASK;
1902 r1 = swap->swb_pages[idx];
1904 if (r1 != SWAPBLK_NONE) {
1905 if (flags & SWM_FREE) {
1906 swp_pager_freeswapspace(r1, 1);
1909 if (flags & (SWM_FREE|SWM_POP)) {
1910 swap->swb_pages[idx] = SWAPBLK_NONE;
1911 if (--swap->swb_count == 0) {
1912 *pswap = swap->swb_hnext;
1913 uma_zfree(swap_zone, swap);
1914 --object->un_pager.swp.swp_bcount;
1919 mtx_unlock(&swhash_mtx);
1924 * System call swapon(name) enables swapping on device name,
1925 * which must be in the swdevsw. Return EBUSY
1926 * if already swapping on this device.
1928 #ifndef _SYS_SYSPROTO_H_
1929 struct swapon_args {
1939 swapon(struct thread *td, struct swapon_args *uap)
1943 struct nameidata nd;
1951 while (swdev_syscall_active)
1952 tsleep(&swdev_syscall_active, PUSER - 1, "swpon", 0);
1953 swdev_syscall_active = 1;
1956 * Swap metadata may not fit in the KVM if we have physical
1959 if (swap_zone == NULL) {
1964 NDINIT(&nd, LOOKUP, ISOPEN | FOLLOW, UIO_USERSPACE, uap->name, td);
1969 NDFREE(&nd, NDF_ONLY_PNBUF);
1972 if (vn_isdisk(vp, &error)) {
1973 error = swapongeom(td, vp);
1974 } else if (vp->v_type == VREG &&
1975 (vp->v_mount->mnt_vfc->vfc_flags & VFCF_NETWORK) != 0 &&
1976 (error = VOP_GETATTR(vp, &attr, td->td_ucred, td)) == 0) {
1978 * Allow direct swapping to NFS regular files in the same
1979 * way that nfs_mountroot() sets up diskless swapping.
1981 error = swaponvp(td, vp, attr.va_size / DEV_BSIZE);
1987 swdev_syscall_active = 0;
1988 wakeup_one(&swdev_syscall_active);
1995 swaponsomething(struct vnode *vp, void *id, u_long nblks, sw_strategy_t *strategy, sw_close_t *close, dev_t dev)
1997 struct swdevt *sp, *tsp;
2002 * If we go beyond this, we get overflows in the radix
2005 mblocks = 0x40000000 / BLIST_META_RADIX;
2006 if (nblks > mblocks) {
2007 printf("WARNING: reducing size to maximum of %lu blocks per swap unit\n",
2012 * nblks is in DEV_BSIZE'd chunks, convert to PAGE_SIZE'd chunks.
2013 * First chop nblks off to page-align it, then convert.
2015 * sw->sw_nblks is in page-sized chunks now too.
2017 nblks &= ~(ctodb(1) - 1);
2018 nblks = dbtoc(nblks);
2020 sp = malloc(sizeof *sp, M_VMPGDATA, M_WAITOK | M_ZERO);
2025 sp->sw_nblks = nblks;
2027 sp->sw_strategy = strategy;
2028 sp->sw_close = close;
2030 sp->sw_blist = blist_create(nblks);
2032 * Do not free the first two block in order to avoid overwriting
2033 * any bsd label at the front of the partition
2035 blist_free(sp->sw_blist, 2, nblks - 2);
2038 mtx_lock(&sw_dev_mtx);
2039 TAILQ_FOREACH(tsp, &swtailq, sw_list) {
2040 if (tsp->sw_end >= dvbase) {
2042 * We put one uncovered page between the devices
2043 * in order to definitively prevent any cross-device
2046 dvbase = tsp->sw_end + 1;
2049 sp->sw_first = dvbase;
2050 sp->sw_end = dvbase + nblks;
2051 TAILQ_INSERT_TAIL(&swtailq, sp, sw_list);
2053 swap_pager_avail += nblks;
2055 mtx_unlock(&sw_dev_mtx);
2059 * SYSCALL: swapoff(devname)
2061 * Disable swapping on the given device.
2063 * XXX: Badly designed system call: it should use a device index
2064 * rather than filename as specification. We keep sw_vp around
2065 * only to make this work.
2067 #ifndef _SYS_SYSPROTO_H_
2068 struct swapoff_args {
2078 swapoff(struct thread *td, struct swapoff_args *uap)
2081 struct nameidata nd;
2090 while (swdev_syscall_active)
2091 tsleep(&swdev_syscall_active, PUSER - 1, "swpoff", 0);
2092 swdev_syscall_active = 1;
2094 NDINIT(&nd, LOOKUP, FOLLOW, UIO_USERSPACE, uap->name, td);
2098 NDFREE(&nd, NDF_ONLY_PNBUF);
2101 mtx_lock(&sw_dev_mtx);
2102 TAILQ_FOREACH(sp, &swtailq, sw_list) {
2103 if (sp->sw_vp == vp)
2106 mtx_unlock(&sw_dev_mtx);
2111 error = swapoff_one(sp, td);
2113 swdev_syscall_active = 0;
2114 wakeup_one(&swdev_syscall_active);
2120 swapoff_one(struct swdevt *sp, struct thread *td)
2122 u_long nblks, dvbase;
2127 mtx_assert(&Giant, MA_OWNED);
2129 (void) vn_lock(sp->sw_vp, LK_EXCLUSIVE | LK_RETRY, td);
2130 error = mac_check_system_swapoff(td->td_ucred, sp->sw_vp);
2131 (void) VOP_UNLOCK(sp->sw_vp, 0, td);
2135 nblks = sp->sw_nblks;
2138 * We can turn off this swap device safely only if the
2139 * available virtual memory in the system will fit the amount
2140 * of data we will have to page back in, plus an epsilon so
2141 * the system doesn't become critically low on swap space.
2143 if (cnt.v_free_count + cnt.v_cache_count + swap_pager_avail <
2144 nblks + nswap_lowat) {
2149 * Prevent further allocations on this device.
2151 mtx_lock(&sw_dev_mtx);
2152 sp->sw_flags |= SW_CLOSING;
2153 for (dvbase = 0; dvbase < sp->sw_end; dvbase += dmmax) {
2154 swap_pager_avail -= blist_fill(sp->sw_blist,
2157 mtx_unlock(&sw_dev_mtx);
2160 * Page in the contents of the device and close it.
2162 swap_pager_swapoff(sp);
2164 sp->sw_close(td, sp);
2166 mtx_lock(&sw_dev_mtx);
2167 TAILQ_REMOVE(&swtailq, sp, sw_list);
2169 if (nswapdev == 0) {
2170 swap_pager_full = 2;
2171 swap_pager_almost_full = 1;
2175 mtx_unlock(&sw_dev_mtx);
2176 blist_destroy(sp->sw_blist);
2177 free(sp, M_VMPGDATA);
2184 struct swdevt *sp, *spt;
2185 const char *devname;
2189 while (swdev_syscall_active)
2190 tsleep(&swdev_syscall_active, PUSER - 1, "swpoff", 0);
2191 swdev_syscall_active = 1;
2193 mtx_lock(&sw_dev_mtx);
2194 TAILQ_FOREACH_SAFE(sp, &swtailq, sw_list, spt) {
2195 mtx_unlock(&sw_dev_mtx);
2196 if (vn_isdisk(sp->sw_vp, NULL))
2197 devname = sp->sw_vp->v_rdev->si_name;
2200 error = swapoff_one(sp, &thread0);
2202 printf("Cannot remove swap device %s (error=%d), "
2203 "skipping.\n", devname, error);
2204 } else if (bootverbose) {
2205 printf("Swap device %s removed.\n", devname);
2207 mtx_lock(&sw_dev_mtx);
2209 mtx_unlock(&sw_dev_mtx);
2211 swdev_syscall_active = 0;
2212 wakeup_one(&swdev_syscall_active);
2217 swap_pager_status(int *total, int *used)
2223 mtx_lock(&sw_dev_mtx);
2224 TAILQ_FOREACH(sp, &swtailq, sw_list) {
2225 *total += sp->sw_nblks;
2226 *used += sp->sw_used;
2228 mtx_unlock(&sw_dev_mtx);
2232 sysctl_vm_swap_info(SYSCTL_HANDLER_ARGS)
2234 int *name = (int *)arg1;
2239 if (arg2 != 1) /* name length */
2243 mtx_lock(&sw_dev_mtx);
2244 TAILQ_FOREACH(sp, &swtailq, sw_list) {
2246 mtx_unlock(&sw_dev_mtx);
2247 xs.xsw_version = XSWDEV_VERSION;
2248 xs.xsw_dev = sp->sw_dev;
2249 xs.xsw_flags = sp->sw_flags;
2250 xs.xsw_nblks = sp->sw_nblks;
2251 xs.xsw_used = sp->sw_used;
2253 error = SYSCTL_OUT(req, &xs, sizeof(xs));
2258 mtx_unlock(&sw_dev_mtx);
2262 SYSCTL_INT(_vm, OID_AUTO, nswapdev, CTLFLAG_RD, &nswapdev, 0,
2263 "Number of swap devices");
2264 SYSCTL_NODE(_vm, OID_AUTO, swap_info, CTLFLAG_RD, sysctl_vm_swap_info,
2265 "Swap statistics by device");
2268 * vmspace_swap_count() - count the approximate swap useage in pages for a
2271 * The map must be locked.
2273 * Swap useage is determined by taking the proportional swap used by
2274 * VM objects backing the VM map. To make up for fractional losses,
2275 * if the VM object has any swap use at all the associated map entries
2276 * count for at least 1 swap page.
2279 vmspace_swap_count(struct vmspace *vmspace)
2281 vm_map_t map = &vmspace->vm_map;
2285 for (cur = map->header.next; cur != &map->header; cur = cur->next) {
2288 if ((cur->eflags & MAP_ENTRY_IS_SUB_MAP) == 0 &&
2289 (object = cur->object.vm_object) != NULL) {
2290 VM_OBJECT_LOCK(object);
2291 if (object->type == OBJT_SWAP &&
2292 object->un_pager.swp.swp_bcount != 0) {
2293 int n = (cur->end - cur->start) / PAGE_SIZE;
2295 count += object->un_pager.swp.swp_bcount *
2296 SWAP_META_PAGES * n / object->size + 1;
2298 VM_OBJECT_UNLOCK(object);
2307 * Swapping onto disk devices.
2311 static g_orphan_t swapgeom_orphan;
2313 static struct g_class g_swap_class = {
2315 .version = G_VERSION,
2316 .orphan = swapgeom_orphan,
2319 DECLARE_GEOM_CLASS(g_swap_class, g_class);
2323 swapgeom_done(struct bio *bp2)
2327 bp = bp2->bio_caller2;
2328 bp->b_ioflags = bp2->bio_flags;
2330 bp->b_ioflags |= BIO_ERROR;
2331 bp->b_resid = bp->b_bcount - bp2->bio_completed;
2332 bp->b_error = bp2->bio_error;
2338 swapgeom_strategy(struct buf *bp, struct swdevt *sp)
2341 struct g_consumer *cp;
2345 bp->b_error = ENXIO;
2346 bp->b_ioflags |= BIO_ERROR;
2350 bio = g_alloc_bio();
2353 * XXX: We shouldn't really sleep here when we run out of buffers
2354 * XXX: but the alternative is worse right now.
2357 bp->b_error = ENOMEM;
2358 bp->b_ioflags |= BIO_ERROR;
2363 bio->bio_caller2 = bp;
2364 bio->bio_cmd = bp->b_iocmd;
2365 bio->bio_data = bp->b_data;
2366 bio->bio_offset = (bp->b_blkno - sp->sw_first) * PAGE_SIZE;
2367 bio->bio_length = bp->b_bcount;
2368 bio->bio_done = swapgeom_done;
2369 g_io_request(bio, cp);
2374 swapgeom_orphan(struct g_consumer *cp)
2378 mtx_lock(&sw_dev_mtx);
2379 TAILQ_FOREACH(sp, &swtailq, sw_list)
2380 if (sp->sw_id == cp)
2382 mtx_unlock(&sw_dev_mtx);
2386 swapgeom_close_ev(void *arg, int flags)
2388 struct g_consumer *cp;
2391 g_access(cp, -1, -1, 0);
2393 g_destroy_consumer(cp);
2397 swapgeom_close(struct thread *td, struct swdevt *sw)
2400 /* XXX: direct call when Giant untangled */
2401 g_waitfor_event(swapgeom_close_ev, sw->sw_id, M_WAITOK, NULL);
2412 swapongeom_ev(void *arg, int flags)
2415 struct g_provider *pp;
2416 struct g_consumer *cp;
2417 static struct g_geom *gp;
2424 pp = g_dev_getprovider(swh->dev);
2426 swh->error = ENODEV;
2429 mtx_lock(&sw_dev_mtx);
2430 TAILQ_FOREACH(sp, &swtailq, sw_list) {
2432 if (cp != NULL && cp->provider == pp) {
2433 mtx_unlock(&sw_dev_mtx);
2438 mtx_unlock(&sw_dev_mtx);
2440 gp = g_new_geomf(&g_swap_class, "swap", NULL);
2441 cp = g_new_consumer(gp);
2444 * XXX: Everytime you think you can improve the margin for
2445 * footshooting, somebody depends on the ability to do so:
2446 * savecore(8) wants to write to our swapdev so we cannot
2447 * set an exclusive count :-(
2449 error = g_access(cp, 1, 1, 0);
2452 g_destroy_consumer(cp);
2456 nblks = pp->mediasize / DEV_BSIZE;
2457 swaponsomething(swh->vp, cp, nblks, swapgeom_strategy,
2458 swapgeom_close, dev2udev(swh->dev));
2464 swapongeom(struct thread *td, struct vnode *vp)
2469 vn_lock(vp, LK_EXCLUSIVE | LK_RETRY, td);
2471 swh.dev = vp->v_rdev;
2474 /* XXX: direct call when Giant untangled */
2475 error = g_waitfor_event(swapongeom_ev, &swh, M_WAITOK, NULL);
2478 VOP_UNLOCK(vp, 0, td);
2485 * This is used mainly for network filesystem (read: probably only tested
2486 * with NFS) swapfiles.
2491 swapdev_strategy(struct buf *bp, struct swdevt *sp)
2495 bp->b_blkno = ctodb(bp->b_blkno - sp->sw_first);
2499 if (bp->b_iocmd == BIO_WRITE) {
2500 if (bp->b_bufobj) /* XXX: should always be true /phk */
2501 bufobj_wdrop(bp->b_bufobj);
2502 bufobj_wref(&vp2->v_bufobj);
2504 if (bp->b_bufobj != &vp2->v_bufobj)
2505 bp->b_bufobj = &vp2->v_bufobj;
2507 bp->b_iooffset = dbtob(bp->b_blkno);
2513 swapdev_close(struct thread *td, struct swdevt *sp)
2516 VOP_CLOSE(sp->sw_vp, FREAD | FWRITE, td->td_ucred, td);
2522 swaponvp(struct thread *td, struct vnode *vp, u_long nblks)
2529 mtx_lock(&sw_dev_mtx);
2530 TAILQ_FOREACH(sp, &swtailq, sw_list) {
2531 if (sp->sw_id == vp) {
2532 mtx_unlock(&sw_dev_mtx);
2536 mtx_unlock(&sw_dev_mtx);
2538 (void) vn_lock(vp, LK_EXCLUSIVE | LK_RETRY, td);
2540 error = mac_check_system_swapon(td->td_ucred, vp);
2543 error = VOP_OPEN(vp, FREAD | FWRITE, td->td_ucred, td, -1);
2544 (void) VOP_UNLOCK(vp, 0, td);
2548 swaponsomething(vp, vp, nblks, swapdev_strategy, swapdev_close,