proc->thread stage 5: BUF/VFS clearance! Remove the ucred argument from
[dragonfly.git] / sys / vm / swap_pager.c
CommitLineData
984263bc
MD
1/*
2 * Copyright (c) 1998 Matthew Dillon,
3 * Copyright (c) 1994 John S. Dyson
4 * Copyright (c) 1990 University of Utah.
5 * Copyright (c) 1991, 1993
6 * The Regents of the University of California. All rights reserved.
7 *
8 * This code is derived from software contributed to Berkeley by
9 * the Systems Programming Group of the University of Utah Computer
10 * Science Department.
11 *
12 * Redistribution and use in source and binary forms, with or without
13 * modification, are permitted provided that the following conditions
14 * are met:
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.
27 *
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
38 * SUCH DAMAGE.
39 *
40 * New Swap System
41 * Matthew Dillon
42 *
43 * Radix Bitmap 'blists'.
44 *
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.
48 *
49 * Features:
50 *
51 * - on the fly reallocation of swap during putpages. The new system
52 * does not try to keep previously allocated swap blocks for dirty
53 * pages.
54 *
55 * - on the fly deallocation of swap
56 *
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
61 * or renamed.
62 *
63 * from: Utah $Hdr: swap_pager.c 1.4 91/04/30$
64 *
65 * @(#)swap_pager.c 8.9 (Berkeley) 3/21/94
66 *
67 * $FreeBSD: src/sys/vm/swap_pager.c,v 1.130.2.12 2002/08/31 21:15:55 dillon Exp $
3b568787 68 * $DragonFly: src/sys/vm/swap_pager.c,v 1.6 2003/06/26 05:55:21 dillon Exp $
984263bc
MD
69 */
70
71#include <sys/param.h>
72#include <sys/systm.h>
73#include <sys/conf.h>
74#include <sys/kernel.h>
75#include <sys/proc.h>
76#include <sys/buf.h>
77#include <sys/vnode.h>
78#include <sys/malloc.h>
79#include <sys/vmmeter.h>
80#include <sys/sysctl.h>
81#include <sys/blist.h>
82#include <sys/lock.h>
83#include <sys/vmmeter.h>
84
85#ifndef MAX_PAGEOUT_CLUSTER
86#define MAX_PAGEOUT_CLUSTER 16
87#endif
88
89#define SWB_NPAGES MAX_PAGEOUT_CLUSTER
90
91#include "opt_swap.h"
92#include <vm/vm.h>
93#include <vm/vm_object.h>
94#include <vm/vm_page.h>
95#include <vm/vm_pager.h>
96#include <vm/vm_pageout.h>
97#include <vm/swap_pager.h>
98#include <vm/vm_extern.h>
99#include <vm/vm_zone.h>
100
3020e3be
MD
101#include <sys/buf2.h>
102
984263bc
MD
103#define SWM_FREE 0x02 /* free, period */
104#define SWM_POP 0x04 /* pop out */
105
106/*
107 * vm_swap_size is in page-sized chunks now. It was DEV_BSIZE'd chunks
108 * in the old system.
109 */
110
111extern int vm_swap_size; /* number of free swap blocks, in pages */
112
113int swap_pager_full; /* swap space exhaustion (task killing) */
114static int swap_pager_almost_full; /* swap space exhaustion (w/ hysteresis)*/
115static int nsw_rcount; /* free read buffers */
116static int nsw_wcount_sync; /* limit write buffers / synchronous */
117static int nsw_wcount_async; /* limit write buffers / asynchronous */
118static int nsw_wcount_async_max;/* assigned maximum */
119static int nsw_cluster_max; /* maximum VOP I/O allowed */
120static int sw_alloc_interlock; /* swap pager allocation interlock */
121
122struct blist *swapblist;
123static struct swblock **swhash;
124static int swhash_mask;
125static int swap_async_max = 4; /* maximum in-progress async I/O's */
126
127extern struct vnode *swapdev_vp; /* from vm_swap.c */
128
129SYSCTL_INT(_vm, OID_AUTO, swap_async_max,
130 CTLFLAG_RW, &swap_async_max, 0, "Maximum running async swap ops");
131
132/*
133 * "named" and "unnamed" anon region objects. Try to reduce the overhead
134 * of searching a named list by hashing it just a little.
135 */
136
137#define NOBJLISTS 8
138
139#define NOBJLIST(handle) \
140 (&swap_pager_object_list[((int)(intptr_t)handle >> 4) & (NOBJLISTS-1)])
141
142static struct pagerlst swap_pager_object_list[NOBJLISTS];
143struct pagerlst swap_pager_un_object_list;
144vm_zone_t swap_zone;
145
146/*
147 * pagerops for OBJT_SWAP - "swap pager". Some ops are also global procedure
148 * calls hooked from other parts of the VM system and do not appear here.
149 * (see vm/swap_pager.h).
150 */
151
152static vm_object_t
153 swap_pager_alloc __P((void *handle, vm_ooffset_t size,
154 vm_prot_t prot, vm_ooffset_t offset));
155static void swap_pager_dealloc __P((vm_object_t object));
156static int swap_pager_getpages __P((vm_object_t, vm_page_t *, int, int));
157static void swap_pager_init __P((void));
158static void swap_pager_unswapped __P((vm_page_t));
159static void swap_pager_strategy __P((vm_object_t, struct buf *));
160
161struct pagerops swappagerops = {
162 swap_pager_init, /* early system initialization of pager */
163 swap_pager_alloc, /* allocate an OBJT_SWAP object */
164 swap_pager_dealloc, /* deallocate an OBJT_SWAP object */
165 swap_pager_getpages, /* pagein */
166 swap_pager_putpages, /* pageout */
167 swap_pager_haspage, /* get backing store status for page */
168 swap_pager_unswapped, /* remove swap related to page */
169 swap_pager_strategy /* pager strategy call */
170};
171
172/*
173 * dmmax is in page-sized chunks with the new swap system. It was
174 * dev-bsized chunks in the old. dmmax is always a power of 2.
175 *
176 * swap_*() routines are externally accessible. swp_*() routines are
177 * internal.
178 */
179
180int dmmax;
181static int dmmax_mask;
182int nswap_lowat = 128; /* in pages, swap_pager_almost_full warn */
183int nswap_hiwat = 512; /* in pages, swap_pager_almost_full warn */
184
185static __inline void swp_sizecheck __P((void));
186static void swp_pager_sync_iodone __P((struct buf *bp));
187static void swp_pager_async_iodone __P((struct buf *bp));
188
189/*
190 * Swap bitmap functions
191 */
192
193static __inline void swp_pager_freeswapspace __P((daddr_t blk, int npages));
194static __inline daddr_t swp_pager_getswapspace __P((int npages));
195
196/*
197 * Metadata functions
198 */
199
200static void swp_pager_meta_build __P((vm_object_t, vm_pindex_t, daddr_t));
201static void swp_pager_meta_free __P((vm_object_t, vm_pindex_t, daddr_t));
202static void swp_pager_meta_free_all __P((vm_object_t));
203static daddr_t swp_pager_meta_ctl __P((vm_object_t, vm_pindex_t, int));
204
205/*
206 * SWP_SIZECHECK() - update swap_pager_full indication
207 *
208 * update the swap_pager_almost_full indication and warn when we are
209 * about to run out of swap space, using lowat/hiwat hysteresis.
210 *
211 * Clear swap_pager_full ( task killing ) indication when lowat is met.
212 *
213 * No restrictions on call
214 * This routine may not block.
215 * This routine must be called at splvm()
216 */
217
218static __inline void
219swp_sizecheck()
220{
221 if (vm_swap_size < nswap_lowat) {
222 if (swap_pager_almost_full == 0) {
223 printf("swap_pager: out of swap space\n");
224 swap_pager_almost_full = 1;
225 }
226 } else {
227 swap_pager_full = 0;
228 if (vm_swap_size > nswap_hiwat)
229 swap_pager_almost_full = 0;
230 }
231}
232
233/*
234 * SWAP_PAGER_INIT() - initialize the swap pager!
235 *
236 * Expected to be started from system init. NOTE: This code is run
237 * before much else so be careful what you depend on. Most of the VM
238 * system has yet to be initialized at this point.
239 */
240
241static void
242swap_pager_init()
243{
244 /*
245 * Initialize object lists
246 */
247 int i;
248
249 for (i = 0; i < NOBJLISTS; ++i)
250 TAILQ_INIT(&swap_pager_object_list[i]);
251 TAILQ_INIT(&swap_pager_un_object_list);
252
253 /*
254 * Device Stripe, in PAGE_SIZE'd blocks
255 */
256
257 dmmax = SWB_NPAGES * 2;
258 dmmax_mask = ~(dmmax - 1);
259}
260
261/*
262 * SWAP_PAGER_SWAP_INIT() - swap pager initialization from pageout process
263 *
264 * Expected to be started from pageout process once, prior to entering
265 * its main loop.
266 */
267
268void
269swap_pager_swap_init()
270{
271 int n, n2;
272
273 /*
274 * Number of in-transit swap bp operations. Don't
275 * exhaust the pbufs completely. Make sure we
276 * initialize workable values (0 will work for hysteresis
277 * but it isn't very efficient).
278 *
279 * The nsw_cluster_max is constrained by the bp->b_pages[]
280 * array (MAXPHYS/PAGE_SIZE) and our locally defined
281 * MAX_PAGEOUT_CLUSTER. Also be aware that swap ops are
282 * constrained by the swap device interleave stripe size.
283 *
284 * Currently we hardwire nsw_wcount_async to 4. This limit is
285 * designed to prevent other I/O from having high latencies due to
286 * our pageout I/O. The value 4 works well for one or two active swap
287 * devices but is probably a little low if you have more. Even so,
288 * a higher value would probably generate only a limited improvement
289 * with three or four active swap devices since the system does not
290 * typically have to pageout at extreme bandwidths. We will want
291 * at least 2 per swap devices, and 4 is a pretty good value if you
292 * have one NFS swap device due to the command/ack latency over NFS.
293 * So it all works out pretty well.
294 */
295
296 nsw_cluster_max = min((MAXPHYS/PAGE_SIZE), MAX_PAGEOUT_CLUSTER);
297
298 nsw_rcount = (nswbuf + 1) / 2;
299 nsw_wcount_sync = (nswbuf + 3) / 4;
300 nsw_wcount_async = 4;
301 nsw_wcount_async_max = nsw_wcount_async;
302
303 /*
304 * Initialize our zone. Right now I'm just guessing on the number
305 * we need based on the number of pages in the system. Each swblock
306 * can hold 16 pages, so this is probably overkill. This reservation
307 * is typically limited to around 32MB by default.
308 */
309 n = cnt.v_page_count / 2;
310 if (maxswzone && n > maxswzone / sizeof(struct swblock))
311 n = maxswzone / sizeof(struct swblock);
312 n2 = n;
313
314 do {
315 swap_zone = zinit(
316 "SWAPMETA",
317 sizeof(struct swblock),
318 n,
319 ZONE_INTERRUPT,
320 1);
321 if (swap_zone != NULL)
322 break;
323 /*
324 * if the allocation failed, try a zone two thirds the
325 * size of the previous attempt.
326 */
327 n -= ((n + 2) / 3);
328 } while (n > 0);
329
330 if (swap_zone == NULL)
331 panic("swap_pager_swap_init: swap_zone == NULL");
332 if (n2 != n)
333 printf("Swap zone entries reduced from %d to %d.\n", n2, n);
334 n2 = n;
335
336 /*
337 * Initialize our meta-data hash table. The swapper does not need to
338 * be quite as efficient as the VM system, so we do not use an
339 * oversized hash table.
340 *
341 * n: size of hash table, must be power of 2
342 * swhash_mask: hash table index mask
343 */
344
345 for (n = 1; n < n2 / 8; n *= 2)
346 ;
347
348 swhash = malloc(sizeof(struct swblock *) * n, M_VMPGDATA, M_WAITOK);
349 bzero(swhash, sizeof(struct swblock *) * n);
350
351 swhash_mask = n - 1;
352}
353
354/*
355 * SWAP_PAGER_ALLOC() - allocate a new OBJT_SWAP VM object and instantiate
356 * its metadata structures.
357 *
358 * This routine is called from the mmap and fork code to create a new
359 * OBJT_SWAP object. We do this by creating an OBJT_DEFAULT object
360 * and then converting it with swp_pager_meta_build().
361 *
362 * This routine may block in vm_object_allocate() and create a named
363 * object lookup race, so we must interlock. We must also run at
364 * splvm() for the object lookup to handle races with interrupts, but
365 * we do not have to maintain splvm() in between the lookup and the
366 * add because (I believe) it is not possible to attempt to create
367 * a new swap object w/handle when a default object with that handle
368 * already exists.
369 */
370
371static vm_object_t
372swap_pager_alloc(void *handle, vm_ooffset_t size, vm_prot_t prot,
373 vm_ooffset_t offset)
374{
375 vm_object_t object;
376
377 if (handle) {
378 /*
379 * Reference existing named region or allocate new one. There
380 * should not be a race here against swp_pager_meta_build()
381 * as called from vm_page_remove() in regards to the lookup
382 * of the handle.
383 */
384
385 while (sw_alloc_interlock) {
386 sw_alloc_interlock = -1;
387 tsleep(&sw_alloc_interlock, PVM, "swpalc", 0);
388 }
389 sw_alloc_interlock = 1;
390
391 object = vm_pager_object_lookup(NOBJLIST(handle), handle);
392
393 if (object != NULL) {
394 vm_object_reference(object);
395 } else {
396 object = vm_object_allocate(OBJT_DEFAULT,
397 OFF_TO_IDX(offset + PAGE_MASK + size));
398 object->handle = handle;
399
400 swp_pager_meta_build(object, 0, SWAPBLK_NONE);
401 }
402
403 if (sw_alloc_interlock < 0)
404 wakeup(&sw_alloc_interlock);
405
406 sw_alloc_interlock = 0;
407 } else {
408 object = vm_object_allocate(OBJT_DEFAULT,
409 OFF_TO_IDX(offset + PAGE_MASK + size));
410
411 swp_pager_meta_build(object, 0, SWAPBLK_NONE);
412 }
413
414 return (object);
415}
416
417/*
418 * SWAP_PAGER_DEALLOC() - remove swap metadata from object
419 *
420 * The swap backing for the object is destroyed. The code is
421 * designed such that we can reinstantiate it later, but this
422 * routine is typically called only when the entire object is
423 * about to be destroyed.
424 *
425 * This routine may block, but no longer does.
426 *
427 * The object must be locked or unreferenceable.
428 */
429
430static void
431swap_pager_dealloc(object)
432 vm_object_t object;
433{
434 int s;
435
436 /*
437 * Remove from list right away so lookups will fail if we block for
438 * pageout completion.
439 */
440
441 if (object->handle == NULL) {
442 TAILQ_REMOVE(&swap_pager_un_object_list, object, pager_object_list);
443 } else {
444 TAILQ_REMOVE(NOBJLIST(object->handle), object, pager_object_list);
445 }
446
447 vm_object_pip_wait(object, "swpdea");
448
449 /*
450 * Free all remaining metadata. We only bother to free it from
451 * the swap meta data. We do not attempt to free swapblk's still
452 * associated with vm_page_t's for this object. We do not care
453 * if paging is still in progress on some objects.
454 */
455 s = splvm();
456 swp_pager_meta_free_all(object);
457 splx(s);
458}
459
460/************************************************************************
461 * SWAP PAGER BITMAP ROUTINES *
462 ************************************************************************/
463
464/*
465 * SWP_PAGER_GETSWAPSPACE() - allocate raw swap space
466 *
467 * Allocate swap for the requested number of pages. The starting
468 * swap block number (a page index) is returned or SWAPBLK_NONE
469 * if the allocation failed.
470 *
471 * Also has the side effect of advising that somebody made a mistake
472 * when they configured swap and didn't configure enough.
473 *
474 * Must be called at splvm() to avoid races with bitmap frees from
475 * vm_page_remove() aka swap_pager_page_removed().
476 *
477 * This routine may not block
478 * This routine must be called at splvm().
479 */
480
481static __inline daddr_t
482swp_pager_getswapspace(npages)
483 int npages;
484{
485 daddr_t blk;
486
487 if ((blk = blist_alloc(swapblist, npages)) == SWAPBLK_NONE) {
488 if (swap_pager_full != 2) {
489 printf("swap_pager_getswapspace: failed\n");
490 swap_pager_full = 2;
491 swap_pager_almost_full = 1;
492 }
493 } else {
494 vm_swap_size -= npages;
495 swp_sizecheck();
496 }
497 return(blk);
498}
499
500/*
501 * SWP_PAGER_FREESWAPSPACE() - free raw swap space
502 *
503 * This routine returns the specified swap blocks back to the bitmap.
504 *
505 * Note: This routine may not block (it could in the old swap code),
506 * and through the use of the new blist routines it does not block.
507 *
508 * We must be called at splvm() to avoid races with bitmap frees from
509 * vm_page_remove() aka swap_pager_page_removed().
510 *
511 * This routine may not block
512 * This routine must be called at splvm().
513 */
514
515static __inline void
516swp_pager_freeswapspace(blk, npages)
517 daddr_t blk;
518 int npages;
519{
520 blist_free(swapblist, blk, npages);
521 vm_swap_size += npages;
522 swp_sizecheck();
523}
524
525/*
526 * SWAP_PAGER_FREESPACE() - frees swap blocks associated with a page
527 * range within an object.
528 *
529 * This is a globally accessible routine.
530 *
531 * This routine removes swapblk assignments from swap metadata.
532 *
533 * The external callers of this routine typically have already destroyed
534 * or renamed vm_page_t's associated with this range in the object so
535 * we should be ok.
536 *
537 * This routine may be called at any spl. We up our spl to splvm temporarily
538 * in order to perform the metadata removal.
539 */
540
541void
542swap_pager_freespace(object, start, size)
543 vm_object_t object;
544 vm_pindex_t start;
545 vm_size_t size;
546{
547 int s = splvm();
548 swp_pager_meta_free(object, start, size);
549 splx(s);
550}
551
552/*
553 * SWAP_PAGER_RESERVE() - reserve swap blocks in object
554 *
555 * Assigns swap blocks to the specified range within the object. The
556 * swap blocks are not zerod. Any previous swap assignment is destroyed.
557 *
558 * Returns 0 on success, -1 on failure.
559 */
560
561int
562swap_pager_reserve(vm_object_t object, vm_pindex_t start, vm_size_t size)
563{
564 int s;
565 int n = 0;
566 daddr_t blk = SWAPBLK_NONE;
567 vm_pindex_t beg = start; /* save start index */
568
569 s = splvm();
570 while (size) {
571 if (n == 0) {
572 n = BLIST_MAX_ALLOC;
573 while ((blk = swp_pager_getswapspace(n)) == SWAPBLK_NONE) {
574 n >>= 1;
575 if (n == 0) {
576 swp_pager_meta_free(object, beg, start - beg);
577 splx(s);
578 return(-1);
579 }
580 }
581 }
582 swp_pager_meta_build(object, start, blk);
583 --size;
584 ++start;
585 ++blk;
586 --n;
587 }
588 swp_pager_meta_free(object, start, n);
589 splx(s);
590 return(0);
591}
592
593/*
594 * SWAP_PAGER_COPY() - copy blocks from source pager to destination pager
595 * and destroy the source.
596 *
597 * Copy any valid swapblks from the source to the destination. In
598 * cases where both the source and destination have a valid swapblk,
599 * we keep the destination's.
600 *
601 * This routine is allowed to block. It may block allocating metadata
602 * indirectly through swp_pager_meta_build() or if paging is still in
603 * progress on the source.
604 *
605 * This routine can be called at any spl
606 *
607 * XXX vm_page_collapse() kinda expects us not to block because we
608 * supposedly do not need to allocate memory, but for the moment we
609 * *may* have to get a little memory from the zone allocator, but
610 * it is taken from the interrupt memory. We should be ok.
611 *
612 * The source object contains no vm_page_t's (which is just as well)
613 *
614 * The source object is of type OBJT_SWAP.
615 *
616 * The source and destination objects must be locked or
617 * inaccessible (XXX are they ?)
618 */
619
620void
621swap_pager_copy(srcobject, dstobject, offset, destroysource)
622 vm_object_t srcobject;
623 vm_object_t dstobject;
624 vm_pindex_t offset;
625 int destroysource;
626{
627 vm_pindex_t i;
628 int s;
629
630 s = splvm();
631
632 /*
633 * If destroysource is set, we remove the source object from the
634 * swap_pager internal queue now.
635 */
636
637 if (destroysource) {
638 if (srcobject->handle == NULL) {
639 TAILQ_REMOVE(
640 &swap_pager_un_object_list,
641 srcobject,
642 pager_object_list
643 );
644 } else {
645 TAILQ_REMOVE(
646 NOBJLIST(srcobject->handle),
647 srcobject,
648 pager_object_list
649 );
650 }
651 }
652
653 /*
654 * transfer source to destination.
655 */
656
657 for (i = 0; i < dstobject->size; ++i) {
658 daddr_t dstaddr;
659
660 /*
661 * Locate (without changing) the swapblk on the destination,
662 * unless it is invalid in which case free it silently, or
663 * if the destination is a resident page, in which case the
664 * source is thrown away.
665 */
666
667 dstaddr = swp_pager_meta_ctl(dstobject, i, 0);
668
669 if (dstaddr == SWAPBLK_NONE) {
670 /*
671 * Destination has no swapblk and is not resident,
672 * copy source.
673 */
674 daddr_t srcaddr;
675
676 srcaddr = swp_pager_meta_ctl(
677 srcobject,
678 i + offset,
679 SWM_POP
680 );
681
682 if (srcaddr != SWAPBLK_NONE)
683 swp_pager_meta_build(dstobject, i, srcaddr);
684 } else {
685 /*
686 * Destination has valid swapblk or it is represented
687 * by a resident page. We destroy the sourceblock.
688 */
689
690 swp_pager_meta_ctl(srcobject, i + offset, SWM_FREE);
691 }
692 }
693
694 /*
695 * Free left over swap blocks in source.
696 *
697 * We have to revert the type to OBJT_DEFAULT so we do not accidently
698 * double-remove the object from the swap queues.
699 */
700
701 if (destroysource) {
702 swp_pager_meta_free_all(srcobject);
703 /*
704 * Reverting the type is not necessary, the caller is going
705 * to destroy srcobject directly, but I'm doing it here
706 * for consistency since we've removed the object from its
707 * queues.
708 */
709 srcobject->type = OBJT_DEFAULT;
710 }
711 splx(s);
712}
713
714/*
715 * SWAP_PAGER_HASPAGE() - determine if we have good backing store for
716 * the requested page.
717 *
718 * We determine whether good backing store exists for the requested
719 * page and return TRUE if it does, FALSE if it doesn't.
720 *
721 * If TRUE, we also try to determine how much valid, contiguous backing
722 * store exists before and after the requested page within a reasonable
723 * distance. We do not try to restrict it to the swap device stripe
724 * (that is handled in getpages/putpages). It probably isn't worth
725 * doing here.
726 */
727
728boolean_t
729swap_pager_haspage(object, pindex, before, after)
730 vm_object_t object;
731 vm_pindex_t pindex;
732 int *before;
733 int *after;
734{
735 daddr_t blk0;
736 int s;
737
738 /*
739 * do we have good backing store at the requested index ?
740 */
741
742 s = splvm();
743 blk0 = swp_pager_meta_ctl(object, pindex, 0);
744
745 if (blk0 == SWAPBLK_NONE) {
746 splx(s);
747 if (before)
748 *before = 0;
749 if (after)
750 *after = 0;
751 return (FALSE);
752 }
753
754 /*
755 * find backwards-looking contiguous good backing store
756 */
757
758 if (before != NULL) {
759 int i;
760
761 for (i = 1; i < (SWB_NPAGES/2); ++i) {
762 daddr_t blk;
763
764 if (i > pindex)
765 break;
766 blk = swp_pager_meta_ctl(object, pindex - i, 0);
767 if (blk != blk0 - i)
768 break;
769 }
770 *before = (i - 1);
771 }
772
773 /*
774 * find forward-looking contiguous good backing store
775 */
776
777 if (after != NULL) {
778 int i;
779
780 for (i = 1; i < (SWB_NPAGES/2); ++i) {
781 daddr_t blk;
782
783 blk = swp_pager_meta_ctl(object, pindex + i, 0);
784 if (blk != blk0 + i)
785 break;
786 }
787 *after = (i - 1);
788 }
789 splx(s);
790 return (TRUE);
791}
792
793/*
794 * SWAP_PAGER_PAGE_UNSWAPPED() - remove swap backing store related to page
795 *
796 * This removes any associated swap backing store, whether valid or
797 * not, from the page.
798 *
799 * This routine is typically called when a page is made dirty, at
800 * which point any associated swap can be freed. MADV_FREE also
801 * calls us in a special-case situation
802 *
803 * NOTE!!! If the page is clean and the swap was valid, the caller
804 * should make the page dirty before calling this routine. This routine
805 * does NOT change the m->dirty status of the page. Also: MADV_FREE
806 * depends on it.
807 *
808 * This routine may not block
809 * This routine must be called at splvm()
810 */
811
812static void
813swap_pager_unswapped(m)
814 vm_page_t m;
815{
816 swp_pager_meta_ctl(m->object, m->pindex, SWM_FREE);
817}
818
819/*
820 * SWAP_PAGER_STRATEGY() - read, write, free blocks
821 *
822 * This implements the vm_pager_strategy() interface to swap and allows
823 * other parts of the system to directly access swap as backing store
824 * through vm_objects of type OBJT_SWAP. This is intended to be a
825 * cacheless interface ( i.e. caching occurs at higher levels ).
826 * Therefore we do not maintain any resident pages. All I/O goes
827 * directly to and from the swap device.
828 *
829 * Note that b_blkno is scaled for PAGE_SIZE
830 *
831 * We currently attempt to run I/O synchronously or asynchronously as
832 * the caller requests. This isn't perfect because we loose error
833 * sequencing when we run multiple ops in parallel to satisfy a request.
834 * But this is swap, so we let it all hang out.
835 */
836
837static void
838swap_pager_strategy(vm_object_t object, struct buf *bp)
839{
840 vm_pindex_t start;
841 int count;
842 int s;
843 char *data;
844 struct buf *nbp = NULL;
845
846 if (bp->b_bcount & PAGE_MASK) {
847 bp->b_error = EINVAL;
848 bp->b_flags |= B_ERROR | B_INVAL;
849 biodone(bp);
850 printf("swap_pager_strategy: bp %p b_vp %p blk %d size %d, not page bounded\n", bp, bp->b_vp, (int)bp->b_pblkno, (int)bp->b_bcount);
851 return;
852 }
853
854 /*
855 * Clear error indication, initialize page index, count, data pointer.
856 */
857
858 bp->b_error = 0;
859 bp->b_flags &= ~B_ERROR;
860 bp->b_resid = bp->b_bcount;
861
862 start = bp->b_pblkno;
863 count = howmany(bp->b_bcount, PAGE_SIZE);
864 data = bp->b_data;
865
866 s = splvm();
867
868 /*
869 * Deal with B_FREEBUF
870 */
871
872 if (bp->b_flags & B_FREEBUF) {
873 /*
874 * FREE PAGE(s) - destroy underlying swap that is no longer
875 * needed.
876 */
877 swp_pager_meta_free(object, start, count);
878 splx(s);
879 bp->b_resid = 0;
880 biodone(bp);
881 return;
882 }
883
884 /*
885 * Execute read or write
886 */
887
888 while (count > 0) {
889 daddr_t blk;
890
891 /*
892 * Obtain block. If block not found and writing, allocate a
893 * new block and build it into the object.
894 */
895
896 blk = swp_pager_meta_ctl(object, start, 0);
897 if ((blk == SWAPBLK_NONE) && (bp->b_flags & B_READ) == 0) {
898 blk = swp_pager_getswapspace(1);
899 if (blk == SWAPBLK_NONE) {
900 bp->b_error = ENOMEM;
901 bp->b_flags |= B_ERROR;
902 break;
903 }
904 swp_pager_meta_build(object, start, blk);
905 }
906
907 /*
908 * Do we have to flush our current collection? Yes if:
909 *
910 * - no swap block at this index
911 * - swap block is not contiguous
912 * - we cross a physical disk boundry in the
913 * stripe.
914 */
915
916 if (
917 nbp && (nbp->b_blkno + btoc(nbp->b_bcount) != blk ||
918 ((nbp->b_blkno ^ blk) & dmmax_mask)
919 )
920 ) {
921 splx(s);
922 if (bp->b_flags & B_READ) {
923 ++cnt.v_swapin;
924 cnt.v_swappgsin += btoc(nbp->b_bcount);
925 } else {
926 ++cnt.v_swapout;
927 cnt.v_swappgsout += btoc(nbp->b_bcount);
928 nbp->b_dirtyend = nbp->b_bcount;
929 }
930 flushchainbuf(nbp);
931 s = splvm();
932 nbp = NULL;
933 }
934
935 /*
936 * Add new swapblk to nbp, instantiating nbp if necessary.
937 * Zero-fill reads are able to take a shortcut.
938 */
939
940 if (blk == SWAPBLK_NONE) {
941 /*
942 * We can only get here if we are reading. Since
943 * we are at splvm() we can safely modify b_resid,
944 * even if chain ops are in progress.
945 */
946 bzero(data, PAGE_SIZE);
947 bp->b_resid -= PAGE_SIZE;
948 } else {
949 if (nbp == NULL) {
950 nbp = getchainbuf(bp, swapdev_vp, (bp->b_flags & B_READ) | B_ASYNC);
951 nbp->b_blkno = blk;
952 nbp->b_bcount = 0;
953 nbp->b_data = data;
954 }
955 nbp->b_bcount += PAGE_SIZE;
956 }
957 --count;
958 ++start;
959 data += PAGE_SIZE;
960 }
961
962 /*
963 * Flush out last buffer
964 */
965
966 splx(s);
967
968 if (nbp) {
969 if ((bp->b_flags & B_ASYNC) == 0)
970 nbp->b_flags &= ~B_ASYNC;
971 if (nbp->b_flags & B_READ) {
972 ++cnt.v_swapin;
973 cnt.v_swappgsin += btoc(nbp->b_bcount);
974 } else {
975 ++cnt.v_swapout;
976 cnt.v_swappgsout += btoc(nbp->b_bcount);
977 nbp->b_dirtyend = nbp->b_bcount;
978 }
979 flushchainbuf(nbp);
980 /* nbp = NULL; */
981 }
982
983 /*
984 * Wait for completion.
985 */
986
987 if (bp->b_flags & B_ASYNC) {
988 autochaindone(bp);
989 } else {
990 waitchainbuf(bp, 0, 1);
991 }
992}
993
994/*
995 * SWAP_PAGER_GETPAGES() - bring pages in from swap
996 *
997 * Attempt to retrieve (m, count) pages from backing store, but make
998 * sure we retrieve at least m[reqpage]. We try to load in as large
999 * a chunk surrounding m[reqpage] as is contiguous in swap and which
1000 * belongs to the same object.
1001 *
1002 * The code is designed for asynchronous operation and
1003 * immediate-notification of 'reqpage' but tends not to be
1004 * used that way. Please do not optimize-out this algorithmic
1005 * feature, I intend to improve on it in the future.
1006 *
1007 * The parent has a single vm_object_pip_add() reference prior to
1008 * calling us and we should return with the same.
1009 *
1010 * The parent has BUSY'd the pages. We should return with 'm'
1011 * left busy, but the others adjusted.
1012 */
1013
1014static int
1015swap_pager_getpages(object, m, count, reqpage)
1016 vm_object_t object;
1017 vm_page_t *m;
1018 int count, reqpage;
1019{
1020 struct buf *bp;
1021 vm_page_t mreq;
1022 int s;
1023 int i;
1024 int j;
1025 daddr_t blk;
1026 vm_offset_t kva;
1027 vm_pindex_t lastpindex;
1028
1029 mreq = m[reqpage];
1030
1031 if (mreq->object != object) {
1032 panic("swap_pager_getpages: object mismatch %p/%p",
1033 object,
1034 mreq->object
1035 );
1036 }
1037 /*
1038 * Calculate range to retrieve. The pages have already been assigned
1039 * their swapblks. We require a *contiguous* range that falls entirely
1040 * within a single device stripe. If we do not supply it, bad things
1041 * happen. Note that blk, iblk & jblk can be SWAPBLK_NONE, but the
1042 * loops are set up such that the case(s) are handled implicitly.
1043 *
1044 * The swp_*() calls must be made at splvm(). vm_page_free() does
1045 * not need to be, but it will go a little faster if it is.
1046 */
1047
1048 s = splvm();
1049 blk = swp_pager_meta_ctl(mreq->object, mreq->pindex, 0);
1050
1051 for (i = reqpage - 1; i >= 0; --i) {
1052 daddr_t iblk;
1053
1054 iblk = swp_pager_meta_ctl(m[i]->object, m[i]->pindex, 0);
1055 if (blk != iblk + (reqpage - i))
1056 break;
1057 if ((blk ^ iblk) & dmmax_mask)
1058 break;
1059 }
1060 ++i;
1061
1062 for (j = reqpage + 1; j < count; ++j) {
1063 daddr_t jblk;
1064
1065 jblk = swp_pager_meta_ctl(m[j]->object, m[j]->pindex, 0);
1066 if (blk != jblk - (j - reqpage))
1067 break;
1068 if ((blk ^ jblk) & dmmax_mask)
1069 break;
1070 }
1071
1072 /*
1073 * free pages outside our collection range. Note: we never free
1074 * mreq, it must remain busy throughout.
1075 */
1076
1077 {
1078 int k;
1079
1080 for (k = 0; k < i; ++k)
1081 vm_page_free(m[k]);
1082 for (k = j; k < count; ++k)
1083 vm_page_free(m[k]);
1084 }
1085 splx(s);
1086
1087
1088 /*
1089 * Return VM_PAGER_FAIL if we have nothing to do. Return mreq
1090 * still busy, but the others unbusied.
1091 */
1092
1093 if (blk == SWAPBLK_NONE)
1094 return(VM_PAGER_FAIL);
1095
1096 /*
1097 * Get a swap buffer header to perform the IO
1098 */
1099
1100 bp = getpbuf(&nsw_rcount);
1101 kva = (vm_offset_t) bp->b_data;
1102
1103 /*
1104 * map our page(s) into kva for input
1105 *
1106 * NOTE: B_PAGING is set by pbgetvp()
1107 */
1108
1109 pmap_qenter(kva, m + i, j - i);
1110
1111 bp->b_flags = B_READ | B_CALL;
1112 bp->b_iodone = swp_pager_async_iodone;
984263bc 1113 bp->b_data = (caddr_t) kva;
984263bc
MD
1114 bp->b_blkno = blk - (reqpage - i);
1115 bp->b_bcount = PAGE_SIZE * (j - i);
1116 bp->b_bufsize = PAGE_SIZE * (j - i);
1117 bp->b_pager.pg_reqpage = reqpage - i;
1118
1119 {
1120 int k;
1121
1122 for (k = i; k < j; ++k) {
1123 bp->b_pages[k - i] = m[k];
1124 vm_page_flag_set(m[k], PG_SWAPINPROG);
1125 }
1126 }
1127 bp->b_npages = j - i;
1128
1129 pbgetvp(swapdev_vp, bp);
1130
1131 cnt.v_swapin++;
1132 cnt.v_swappgsin += bp->b_npages;
1133
1134 /*
1135 * We still hold the lock on mreq, and our automatic completion routine
1136 * does not remove it.
1137 */
1138
1139 vm_object_pip_add(mreq->object, bp->b_npages);
1140 lastpindex = m[j-1]->pindex;
1141
1142 /*
1143 * perform the I/O. NOTE!!! bp cannot be considered valid after
1144 * this point because we automatically release it on completion.
1145 * Instead, we look at the one page we are interested in which we
1146 * still hold a lock on even through the I/O completion.
1147 *
1148 * The other pages in our m[] array are also released on completion,
1149 * so we cannot assume they are valid anymore either.
1150 *
1151 * NOTE: b_blkno is destroyed by the call to VOP_STRATEGY
1152 */
1153
1154 BUF_KERNPROC(bp);
1155 VOP_STRATEGY(bp->b_vp, bp);
1156
1157 /*
1158 * wait for the page we want to complete. PG_SWAPINPROG is always
1159 * cleared on completion. If an I/O error occurs, SWAPBLK_NONE
1160 * is set in the meta-data.
1161 */
1162
1163 s = splvm();
1164
1165 while ((mreq->flags & PG_SWAPINPROG) != 0) {
1166 vm_page_flag_set(mreq, PG_WANTED | PG_REFERENCED);
1167 cnt.v_intrans++;
1168 if (tsleep(mreq, PSWP, "swread", hz*20)) {
1169 printf(
1170 "swap_pager: indefinite wait buffer: device:"
1171 " %s, blkno: %ld, size: %ld\n",
1172 devtoname(bp->b_dev), (long)bp->b_blkno,
1173 bp->b_bcount
1174 );
1175 }
1176 }
1177
1178 splx(s);
1179
1180 /*
1181 * mreq is left bussied after completion, but all the other pages
1182 * are freed. If we had an unrecoverable read error the page will
1183 * not be valid.
1184 */
1185
1186 if (mreq->valid != VM_PAGE_BITS_ALL) {
1187 return(VM_PAGER_ERROR);
1188 } else {
1189 return(VM_PAGER_OK);
1190 }
1191
1192 /*
1193 * A final note: in a low swap situation, we cannot deallocate swap
1194 * and mark a page dirty here because the caller is likely to mark
1195 * the page clean when we return, causing the page to possibly revert
1196 * to all-zero's later.
1197 */
1198}
1199
1200/*
1201 * swap_pager_putpages:
1202 *
1203 * Assign swap (if necessary) and initiate I/O on the specified pages.
1204 *
1205 * We support both OBJT_DEFAULT and OBJT_SWAP objects. DEFAULT objects
1206 * are automatically converted to SWAP objects.
1207 *
1208 * In a low memory situation we may block in VOP_STRATEGY(), but the new
1209 * vm_page reservation system coupled with properly written VFS devices
1210 * should ensure that no low-memory deadlock occurs. This is an area
1211 * which needs work.
1212 *
1213 * The parent has N vm_object_pip_add() references prior to
1214 * calling us and will remove references for rtvals[] that are
1215 * not set to VM_PAGER_PEND. We need to remove the rest on I/O
1216 * completion.
1217 *
1218 * The parent has soft-busy'd the pages it passes us and will unbusy
1219 * those whos rtvals[] entry is not set to VM_PAGER_PEND on return.
1220 * We need to unbusy the rest on I/O completion.
1221 */
1222
1223void
1224swap_pager_putpages(object, m, count, sync, rtvals)
1225 vm_object_t object;
1226 vm_page_t *m;
1227 int count;
1228 boolean_t sync;
1229 int *rtvals;
1230{
1231 int i;
1232 int n = 0;
1233
1234 if (count && m[0]->object != object) {
1235 panic("swap_pager_getpages: object mismatch %p/%p",
1236 object,
1237 m[0]->object
1238 );
1239 }
1240 /*
1241 * Step 1
1242 *
1243 * Turn object into OBJT_SWAP
1244 * check for bogus sysops
1245 * force sync if not pageout process
1246 */
1247
1248 if (object->type != OBJT_SWAP)
1249 swp_pager_meta_build(object, 0, SWAPBLK_NONE);
1250
bc6dffab 1251 if (curthread != pagethread)
984263bc
MD
1252 sync = TRUE;
1253
1254 /*
1255 * Step 2
1256 *
1257 * Update nsw parameters from swap_async_max sysctl values.
1258 * Do not let the sysop crash the machine with bogus numbers.
1259 */
1260
1261 if (swap_async_max != nsw_wcount_async_max) {
1262 int n;
1263 int s;
1264
1265 /*
1266 * limit range
1267 */
1268 if ((n = swap_async_max) > nswbuf / 2)
1269 n = nswbuf / 2;
1270 if (n < 1)
1271 n = 1;
1272 swap_async_max = n;
1273
1274 /*
1275 * Adjust difference ( if possible ). If the current async
1276 * count is too low, we may not be able to make the adjustment
1277 * at this time.
1278 */
1279 s = splvm();
1280 n -= nsw_wcount_async_max;
1281 if (nsw_wcount_async + n >= 0) {
1282 nsw_wcount_async += n;
1283 nsw_wcount_async_max += n;
1284 wakeup(&nsw_wcount_async);
1285 }
1286 splx(s);
1287 }
1288
1289 /*
1290 * Step 3
1291 *
1292 * Assign swap blocks and issue I/O. We reallocate swap on the fly.
1293 * The page is left dirty until the pageout operation completes
1294 * successfully.
1295 */
1296
1297 for (i = 0; i < count; i += n) {
1298 int s;
1299 int j;
1300 struct buf *bp;
1301 daddr_t blk;
1302
1303 /*
1304 * Maximum I/O size is limited by a number of factors.
1305 */
1306
1307 n = min(BLIST_MAX_ALLOC, count - i);
1308 n = min(n, nsw_cluster_max);
1309
1310 s = splvm();
1311
1312 /*
1313 * Get biggest block of swap we can. If we fail, fall
1314 * back and try to allocate a smaller block. Don't go
1315 * overboard trying to allocate space if it would overly
1316 * fragment swap.
1317 */
1318 while (
1319 (blk = swp_pager_getswapspace(n)) == SWAPBLK_NONE &&
1320 n > 4
1321 ) {
1322 n >>= 1;
1323 }
1324 if (blk == SWAPBLK_NONE) {
1325 for (j = 0; j < n; ++j)
1326 rtvals[i+j] = VM_PAGER_FAIL;
1327 splx(s);
1328 continue;
1329 }
1330
1331 /*
1332 * The I/O we are constructing cannot cross a physical
1333 * disk boundry in the swap stripe. Note: we are still
1334 * at splvm().
1335 */
1336 if ((blk ^ (blk + n)) & dmmax_mask) {
1337 j = ((blk + dmmax) & dmmax_mask) - blk;
1338 swp_pager_freeswapspace(blk + j, n - j);
1339 n = j;
1340 }
1341
1342 /*
1343 * All I/O parameters have been satisfied, build the I/O
1344 * request and assign the swap space.
1345 *
1346 * NOTE: B_PAGING is set by pbgetvp()
1347 */
1348
1349 if (sync == TRUE) {
1350 bp = getpbuf(&nsw_wcount_sync);
1351 bp->b_flags = B_CALL;
1352 } else {
1353 bp = getpbuf(&nsw_wcount_async);
1354 bp->b_flags = B_CALL | B_ASYNC;
1355 }
1356 bp->b_spc = NULL; /* not used, but NULL-out anyway */
1357
1358 pmap_qenter((vm_offset_t)bp->b_data, &m[i], n);
1359
984263bc
MD
1360 bp->b_bcount = PAGE_SIZE * n;
1361 bp->b_bufsize = PAGE_SIZE * n;
1362 bp->b_blkno = blk;
1363
984263bc
MD
1364 pbgetvp(swapdev_vp, bp);
1365
1366 for (j = 0; j < n; ++j) {
1367 vm_page_t mreq = m[i+j];
1368
1369 swp_pager_meta_build(
1370 mreq->object,
1371 mreq->pindex,
1372 blk + j
1373 );
1374 vm_page_dirty(mreq);
1375 rtvals[i+j] = VM_PAGER_OK;
1376
1377 vm_page_flag_set(mreq, PG_SWAPINPROG);
1378 bp->b_pages[j] = mreq;
1379 }
1380 bp->b_npages = n;
1381 /*
1382 * Must set dirty range for NFS to work.
1383 */
1384 bp->b_dirtyoff = 0;
1385 bp->b_dirtyend = bp->b_bcount;
1386
1387 cnt.v_swapout++;
1388 cnt.v_swappgsout += bp->b_npages;
1389 swapdev_vp->v_numoutput++;
1390
1391 splx(s);
1392
1393 /*
1394 * asynchronous
1395 *
1396 * NOTE: b_blkno is destroyed by the call to VOP_STRATEGY
1397 */
1398
1399 if (sync == FALSE) {
1400 bp->b_iodone = swp_pager_async_iodone;
1401 BUF_KERNPROC(bp);
1402 VOP_STRATEGY(bp->b_vp, bp);
1403
1404 for (j = 0; j < n; ++j)
1405 rtvals[i+j] = VM_PAGER_PEND;
1406 continue;
1407 }
1408
1409 /*
1410 * synchronous
1411 *
1412 * NOTE: b_blkno is destroyed by the call to VOP_STRATEGY
1413 */
1414
1415 bp->b_iodone = swp_pager_sync_iodone;
1416 VOP_STRATEGY(bp->b_vp, bp);
1417
1418 /*
1419 * Wait for the sync I/O to complete, then update rtvals.
1420 * We just set the rtvals[] to VM_PAGER_PEND so we can call
1421 * our async completion routine at the end, thus avoiding a
1422 * double-free.
1423 */
1424 s = splbio();
1425
1426 while ((bp->b_flags & B_DONE) == 0) {
1427 tsleep(bp, PVM, "swwrt", 0);
1428 }
1429
1430 for (j = 0; j < n; ++j)
1431 rtvals[i+j] = VM_PAGER_PEND;
1432
1433 /*
1434 * Now that we are through with the bp, we can call the
1435 * normal async completion, which frees everything up.
1436 */
1437
1438 swp_pager_async_iodone(bp);
1439
1440 splx(s);
1441 }
1442}
1443
1444/*
1445 * swap_pager_sync_iodone:
1446 *
1447 * Completion routine for synchronous reads and writes from/to swap.
1448 * We just mark the bp is complete and wake up anyone waiting on it.
1449 *
1450 * This routine may not block. This routine is called at splbio() or better.
1451 */
1452
1453static void
1454swp_pager_sync_iodone(bp)
1455 struct buf *bp;
1456{
1457 bp->b_flags |= B_DONE;
1458 bp->b_flags &= ~B_ASYNC;
1459 wakeup(bp);
1460}
1461
1462/*
1463 * swp_pager_async_iodone:
1464 *
1465 * Completion routine for asynchronous reads and writes from/to swap.
1466 * Also called manually by synchronous code to finish up a bp.
1467 *
1468 * For READ operations, the pages are PG_BUSY'd. For WRITE operations,
1469 * the pages are vm_page_t->busy'd. For READ operations, we PG_BUSY
1470 * unbusy all pages except the 'main' request page. For WRITE
1471 * operations, we vm_page_t->busy'd unbusy all pages ( we can do this
1472 * because we marked them all VM_PAGER_PEND on return from putpages ).
1473 *
1474 * This routine may not block.
1475 * This routine is called at splbio() or better
1476 *
1477 * We up ourselves to splvm() as required for various vm_page related
1478 * calls.
1479 */
1480
1481static void
1482swp_pager_async_iodone(bp)
1483 register struct buf *bp;
1484{
1485 int s;
1486 int i;
1487 vm_object_t object = NULL;
1488
1489 bp->b_flags |= B_DONE;
1490
1491 /*
1492 * report error
1493 */
1494
1495 if (bp->b_flags & B_ERROR) {
1496 printf(
1497 "swap_pager: I/O error - %s failed; blkno %ld,"
1498 "size %ld, error %d\n",
1499 ((bp->b_flags & B_READ) ? "pagein" : "pageout"),
1500 (long)bp->b_blkno,
1501 (long)bp->b_bcount,
1502 bp->b_error
1503 );
1504 }
1505
1506 /*
1507 * set object, raise to splvm().
1508 */
1509
1510 if (bp->b_npages)
1511 object = bp->b_pages[0]->object;
1512 s = splvm();
1513
1514 /*
1515 * remove the mapping for kernel virtual
1516 */
1517
1518 pmap_qremove((vm_offset_t)bp->b_data, bp->b_npages);
1519
1520 /*
1521 * cleanup pages. If an error occurs writing to swap, we are in
1522 * very serious trouble. If it happens to be a disk error, though,
1523 * we may be able to recover by reassigning the swap later on. So
1524 * in this case we remove the m->swapblk assignment for the page
1525 * but do not free it in the rlist. The errornous block(s) are thus
1526 * never reallocated as swap. Redirty the page and continue.
1527 */
1528
1529 for (i = 0; i < bp->b_npages; ++i) {
1530 vm_page_t m = bp->b_pages[i];
1531
1532 vm_page_flag_clear(m, PG_SWAPINPROG);
1533
1534 if (bp->b_flags & B_ERROR) {
1535 /*
1536 * If an error occurs I'd love to throw the swapblk
1537 * away without freeing it back to swapspace, so it
1538 * can never be used again. But I can't from an
1539 * interrupt.
1540 */
1541
1542 if (bp->b_flags & B_READ) {
1543 /*
1544 * When reading, reqpage needs to stay
1545 * locked for the parent, but all other
1546 * pages can be freed. We still want to
1547 * wakeup the parent waiting on the page,
1548 * though. ( also: pg_reqpage can be -1 and
1549 * not match anything ).
1550 *
1551 * We have to wake specifically requested pages
1552 * up too because we cleared PG_SWAPINPROG and
1553 * someone may be waiting for that.
1554 *
1555 * NOTE: for reads, m->dirty will probably
1556 * be overridden by the original caller of
1557 * getpages so don't play cute tricks here.
1558 *
1559 * XXX IT IS NOT LEGAL TO FREE THE PAGE HERE
1560 * AS THIS MESSES WITH object->memq, and it is
1561 * not legal to mess with object->memq from an
1562 * interrupt.
1563 */
1564
1565 m->valid = 0;
1566 vm_page_flag_clear(m, PG_ZERO);
1567
1568 if (i != bp->b_pager.pg_reqpage)
1569 vm_page_free(m);
1570 else
1571 vm_page_flash(m);
1572 /*
1573 * If i == bp->b_pager.pg_reqpage, do not wake
1574 * the page up. The caller needs to.
1575 */
1576 } else {
1577 /*
1578 * If a write error occurs, reactivate page
1579 * so it doesn't clog the inactive list,
1580 * then finish the I/O.
1581 */
1582 vm_page_dirty(m);
1583 vm_page_activate(m);
1584 vm_page_io_finish(m);
1585 }
1586 } else if (bp->b_flags & B_READ) {
1587 /*
1588 * For read success, clear dirty bits. Nobody should
1589 * have this page mapped but don't take any chances,
1590 * make sure the pmap modify bits are also cleared.
1591 *
1592 * NOTE: for reads, m->dirty will probably be
1593 * overridden by the original caller of getpages so
1594 * we cannot set them in order to free the underlying
1595 * swap in a low-swap situation. I don't think we'd
1596 * want to do that anyway, but it was an optimization
1597 * that existed in the old swapper for a time before
1598 * it got ripped out due to precisely this problem.
1599 *
1600 * clear PG_ZERO in page.
1601 *
1602 * If not the requested page then deactivate it.
1603 *
1604 * Note that the requested page, reqpage, is left
1605 * busied, but we still have to wake it up. The
1606 * other pages are released (unbusied) by
1607 * vm_page_wakeup(). We do not set reqpage's
1608 * valid bits here, it is up to the caller.
1609 */
1610
1611 pmap_clear_modify(m);
1612 m->valid = VM_PAGE_BITS_ALL;
1613 vm_page_undirty(m);
1614 vm_page_flag_clear(m, PG_ZERO);
1615
1616 /*
1617 * We have to wake specifically requested pages
1618 * up too because we cleared PG_SWAPINPROG and
1619 * could be waiting for it in getpages. However,
1620 * be sure to not unbusy getpages specifically
1621 * requested page - getpages expects it to be
1622 * left busy.
1623 */
1624 if (i != bp->b_pager.pg_reqpage) {
1625 vm_page_deactivate(m);
1626 vm_page_wakeup(m);
1627 } else {
1628 vm_page_flash(m);
1629 }
1630 } else {
1631 /*
1632 * For write success, clear the modify and dirty
1633 * status, then finish the I/O ( which decrements the
1634 * busy count and possibly wakes waiter's up ).
1635 */
1636 pmap_clear_modify(m);
1637 vm_page_undirty(m);
1638 vm_page_io_finish(m);
1639 if (!vm_page_count_severe() || !vm_page_try_to_cache(m))
1640 vm_page_protect(m, VM_PROT_READ);
1641 }
1642 }
1643
1644 /*
1645 * adjust pip. NOTE: the original parent may still have its own
1646 * pip refs on the object.
1647 */
1648
1649 if (object)
1650 vm_object_pip_wakeupn(object, bp->b_npages);
1651
1652 /*
1653 * release the physical I/O buffer
1654 */
1655
1656 relpbuf(
1657 bp,
1658 ((bp->b_flags & B_READ) ? &nsw_rcount :
1659 ((bp->b_flags & B_ASYNC) ?
1660 &nsw_wcount_async :
1661 &nsw_wcount_sync
1662 )
1663 )
1664 );
1665 splx(s);
1666}
1667
1668/************************************************************************
1669 * SWAP META DATA *
1670 ************************************************************************
1671 *
1672 * These routines manipulate the swap metadata stored in the
1673 * OBJT_SWAP object. All swp_*() routines must be called at
1674 * splvm() because swap can be freed up by the low level vm_page
1675 * code which might be called from interrupts beyond what splbio() covers.
1676 *
1677 * Swap metadata is implemented with a global hash and not directly
1678 * linked into the object. Instead the object simply contains
1679 * appropriate tracking counters.
1680 */
1681
1682/*
1683 * SWP_PAGER_HASH() - hash swap meta data
1684 *
1685 * This is an inline helper function which hashes the swapblk given
1686 * the object and page index. It returns a pointer to a pointer
1687 * to the object, or a pointer to a NULL pointer if it could not
1688 * find a swapblk.
1689 *
1690 * This routine must be called at splvm().
1691 */
1692
1693static __inline struct swblock **
1694swp_pager_hash(vm_object_t object, vm_pindex_t index)
1695{
1696 struct swblock **pswap;
1697 struct swblock *swap;
1698
1699 index &= ~SWAP_META_MASK;
1700 pswap = &swhash[(index ^ (int)(intptr_t)object) & swhash_mask];
1701
1702 while ((swap = *pswap) != NULL) {
1703 if (swap->swb_object == object &&
1704 swap->swb_index == index
1705 ) {
1706 break;
1707 }
1708 pswap = &swap->swb_hnext;
1709 }
1710 return(pswap);
1711}
1712
1713/*
1714 * SWP_PAGER_META_BUILD() - add swap block to swap meta data for object
1715 *
1716 * We first convert the object to a swap object if it is a default
1717 * object.
1718 *
1719 * The specified swapblk is added to the object's swap metadata. If
1720 * the swapblk is not valid, it is freed instead. Any previously
1721 * assigned swapblk is freed.
1722 *
1723 * This routine must be called at splvm(), except when used to convert
1724 * an OBJT_DEFAULT object into an OBJT_SWAP object.
1725
1726 */
1727
1728static void
1729swp_pager_meta_build(
1730 vm_object_t object,
1731 vm_pindex_t index,
1732 daddr_t swapblk
1733) {
1734 struct swblock *swap;
1735 struct swblock **pswap;
1736
1737 /*
1738 * Convert default object to swap object if necessary
1739 */
1740
1741 if (object->type != OBJT_SWAP) {
1742 object->type = OBJT_SWAP;
1743 object->un_pager.swp.swp_bcount = 0;
1744
1745 if (object->handle != NULL) {
1746 TAILQ_INSERT_TAIL(
1747 NOBJLIST(object->handle),
1748 object,
1749 pager_object_list
1750 );
1751 } else {
1752 TAILQ_INSERT_TAIL(
1753 &swap_pager_un_object_list,
1754 object,
1755 pager_object_list
1756 );
1757 }
1758 }
1759
1760 /*
1761 * Locate hash entry. If not found create, but if we aren't adding
1762 * anything just return. If we run out of space in the map we wait
1763 * and, since the hash table may have changed, retry.
1764 */
1765
1766retry:
1767 pswap = swp_pager_hash(object, index);
1768
1769 if ((swap = *pswap) == NULL) {
1770 int i;
1771
1772 if (swapblk == SWAPBLK_NONE)
1773 return;
1774
1775 swap = *pswap = zalloc(swap_zone);
1776 if (swap == NULL) {
1777 VM_WAIT;
1778 goto retry;
1779 }
1780 swap->swb_hnext = NULL;
1781 swap->swb_object = object;
1782 swap->swb_index = index & ~SWAP_META_MASK;
1783 swap->swb_count = 0;
1784
1785 ++object->un_pager.swp.swp_bcount;
1786
1787 for (i = 0; i < SWAP_META_PAGES; ++i)
1788 swap->swb_pages[i] = SWAPBLK_NONE;
1789 }
1790
1791 /*
1792 * Delete prior contents of metadata
1793 */
1794
1795 index &= SWAP_META_MASK;
1796
1797 if (swap->swb_pages[index] != SWAPBLK_NONE) {
1798 swp_pager_freeswapspace(swap->swb_pages[index], 1);
1799 --swap->swb_count;
1800 }
1801
1802 /*
1803 * Enter block into metadata
1804 */
1805
1806 swap->swb_pages[index] = swapblk;
1807 if (swapblk != SWAPBLK_NONE)
1808 ++swap->swb_count;
1809}
1810
1811/*
1812 * SWP_PAGER_META_FREE() - free a range of blocks in the object's swap metadata
1813 *
1814 * The requested range of blocks is freed, with any associated swap
1815 * returned to the swap bitmap.
1816 *
1817 * This routine will free swap metadata structures as they are cleaned
1818 * out. This routine does *NOT* operate on swap metadata associated
1819 * with resident pages.
1820 *
1821 * This routine must be called at splvm()
1822 */
1823
1824static void
1825swp_pager_meta_free(vm_object_t object, vm_pindex_t index, daddr_t count)
1826{
1827 if (object->type != OBJT_SWAP)
1828 return;
1829
1830 while (count > 0) {
1831 struct swblock **pswap;
1832 struct swblock *swap;
1833
1834 pswap = swp_pager_hash(object, index);
1835
1836 if ((swap = *pswap) != NULL) {
1837 daddr_t v = swap->swb_pages[index & SWAP_META_MASK];
1838
1839 if (v != SWAPBLK_NONE) {
1840 swp_pager_freeswapspace(v, 1);
1841 swap->swb_pages[index & SWAP_META_MASK] =
1842 SWAPBLK_NONE;
1843 if (--swap->swb_count == 0) {
1844 *pswap = swap->swb_hnext;
1845 zfree(swap_zone, swap);
1846 --object->un_pager.swp.swp_bcount;
1847 }
1848 }
1849 --count;
1850 ++index;
1851 } else {
1852 int n = SWAP_META_PAGES - (index & SWAP_META_MASK);
1853 count -= n;
1854 index += n;
1855 }
1856 }
1857}
1858
1859/*
1860 * SWP_PAGER_META_FREE_ALL() - destroy all swap metadata associated with object
1861 *
1862 * This routine locates and destroys all swap metadata associated with
1863 * an object.
1864 *
1865 * This routine must be called at splvm()
1866 */
1867
1868static void
1869swp_pager_meta_free_all(vm_object_t object)
1870{
1871 daddr_t index = 0;
1872
1873 if (object->type != OBJT_SWAP)
1874 return;
1875
1876 while (object->un_pager.swp.swp_bcount) {
1877 struct swblock **pswap;
1878 struct swblock *swap;
1879
1880 pswap = swp_pager_hash(object, index);
1881 if ((swap = *pswap) != NULL) {
1882 int i;
1883
1884 for (i = 0; i < SWAP_META_PAGES; ++i) {
1885 daddr_t v = swap->swb_pages[i];
1886 if (v != SWAPBLK_NONE) {
1887 --swap->swb_count;
1888 swp_pager_freeswapspace(v, 1);
1889 }
1890 }
1891 if (swap->swb_count != 0)
1892 panic("swap_pager_meta_free_all: swb_count != 0");
1893 *pswap = swap->swb_hnext;
1894 zfree(swap_zone, swap);
1895 --object->un_pager.swp.swp_bcount;
1896 }
1897 index += SWAP_META_PAGES;
1898 if (index > 0x20000000)
1899 panic("swp_pager_meta_free_all: failed to locate all swap meta blocks");
1900 }
1901}
1902
1903/*
1904 * SWP_PAGER_METACTL() - misc control of swap and vm_page_t meta data.
1905 *
1906 * This routine is capable of looking up, popping, or freeing
1907 * swapblk assignments in the swap meta data or in the vm_page_t.
1908 * The routine typically returns the swapblk being looked-up, or popped,
1909 * or SWAPBLK_NONE if the block was freed, or SWAPBLK_NONE if the block
1910 * was invalid. This routine will automatically free any invalid
1911 * meta-data swapblks.
1912 *
1913 * It is not possible to store invalid swapblks in the swap meta data
1914 * (other then a literal 'SWAPBLK_NONE'), so we don't bother checking.
1915 *
1916 * When acting on a busy resident page and paging is in progress, we
1917 * have to wait until paging is complete but otherwise can act on the
1918 * busy page.
1919 *
1920 * This routine must be called at splvm().
1921 *
1922 * SWM_FREE remove and free swap block from metadata
1923 * SWM_POP remove from meta data but do not free.. pop it out
1924 */
1925
1926static daddr_t
1927swp_pager_meta_ctl(
1928 vm_object_t object,
1929 vm_pindex_t index,
1930 int flags
1931) {
1932 struct swblock **pswap;
1933 struct swblock *swap;
1934 daddr_t r1;
1935
1936 /*
1937 * The meta data only exists of the object is OBJT_SWAP
1938 * and even then might not be allocated yet.
1939 */
1940
1941 if (object->type != OBJT_SWAP)
1942 return(SWAPBLK_NONE);
1943
1944 r1 = SWAPBLK_NONE;
1945 pswap = swp_pager_hash(object, index);
1946
1947 if ((swap = *pswap) != NULL) {
1948 index &= SWAP_META_MASK;
1949 r1 = swap->swb_pages[index];
1950
1951 if (r1 != SWAPBLK_NONE) {
1952 if (flags & SWM_FREE) {
1953 swp_pager_freeswapspace(r1, 1);
1954 r1 = SWAPBLK_NONE;
1955 }
1956 if (flags & (SWM_FREE|SWM_POP)) {
1957 swap->swb_pages[index] = SWAPBLK_NONE;
1958 if (--swap->swb_count == 0) {
1959 *pswap = swap->swb_hnext;
1960 zfree(swap_zone, swap);
1961 --object->un_pager.swp.swp_bcount;
1962 }
1963 }
1964 }
1965 }
1966 return(r1);
1967}
1968