AMD64 - Fix many compile-time warnings. int/ptr type mismatches, %llx, etc.
[dragonfly.git] / sys / vm / swap_pager.c
CommitLineData
984263bc 1/*
8c10bfcf
MD
2 * Copyright (c) 1998,2004 The DragonFly Project. All rights reserved.
3 *
4 * This code is derived from software contributed to The DragonFly Project
5 * by Matthew Dillon <dillon@backplane.com>
6 *
7 * Redistribution and use in source and binary forms, with or without
8 * modification, are permitted provided that the following conditions
9 * are met:
10 *
11 * 1. Redistributions of source code must retain the above copyright
12 * notice, this list of conditions and the following disclaimer.
13 * 2. Redistributions in binary form must reproduce the above copyright
14 * notice, this list of conditions and the following disclaimer in
15 * the documentation and/or other materials provided with the
16 * distribution.
17 * 3. Neither the name of The DragonFly Project nor the names of its
18 * contributors may be used to endorse or promote products derived
19 * from this software without specific, prior written permission.
20 *
21 * THIS SOFTWARE IS PROVIDED BY THE COPYRIGHT HOLDERS AND CONTRIBUTORS
22 * ``AS IS'' AND ANY EXPRESS OR IMPLIED WARRANTIES, INCLUDING, BUT NOT
23 * LIMITED TO, THE IMPLIED WARRANTIES OF MERCHANTABILITY AND FITNESS
24 * FOR A PARTICULAR PURPOSE ARE DISCLAIMED. IN NO EVENT SHALL THE
25 * COPYRIGHT HOLDERS OR CONTRIBUTORS BE LIABLE FOR ANY DIRECT, INDIRECT,
26 * INCIDENTAL, SPECIAL, EXEMPLARY OR CONSEQUENTIAL DAMAGES (INCLUDING,
27 * BUT NOT LIMITED TO, PROCUREMENT OF SUBSTITUTE GOODS OR SERVICES;
28 * LOSS OF USE, DATA, OR PROFITS; OR BUSINESS INTERRUPTION) HOWEVER CAUSED
29 * AND ON ANY THEORY OF LIABILITY, WHETHER IN CONTRACT, STRICT LIABILITY,
30 * OR TORT (INCLUDING NEGLIGENCE OR OTHERWISE) ARISING IN ANY WAY OUT
31 * OF THE USE OF THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF
32 * SUCH DAMAGE.
33 *
984263bc
MD
34 * Copyright (c) 1994 John S. Dyson
35 * Copyright (c) 1990 University of Utah.
36 * Copyright (c) 1991, 1993
37 * The Regents of the University of California. All rights reserved.
38 *
39 * This code is derived from software contributed to Berkeley by
40 * the Systems Programming Group of the University of Utah Computer
41 * Science Department.
42 *
43 * Redistribution and use in source and binary forms, with or without
44 * modification, are permitted provided that the following conditions
45 * are met:
46 * 1. Redistributions of source code must retain the above copyright
47 * notice, this list of conditions and the following disclaimer.
48 * 2. Redistributions in binary form must reproduce the above copyright
49 * notice, this list of conditions and the following disclaimer in the
50 * documentation and/or other materials provided with the distribution.
51 * 3. All advertising materials mentioning features or use of this software
52 * must display the following acknowledgement:
53 * This product includes software developed by the University of
54 * California, Berkeley and its contributors.
55 * 4. Neither the name of the University nor the names of its contributors
56 * may be used to endorse or promote products derived from this software
57 * without specific prior written permission.
58 *
59 * THIS SOFTWARE IS PROVIDED BY THE REGENTS AND CONTRIBUTORS ``AS IS'' AND
60 * ANY EXPRESS OR IMPLIED WARRANTIES, INCLUDING, BUT NOT LIMITED TO, THE
61 * IMPLIED WARRANTIES OF MERCHANTABILITY AND FITNESS FOR A PARTICULAR PURPOSE
62 * ARE DISCLAIMED. IN NO EVENT SHALL THE REGENTS OR CONTRIBUTORS BE LIABLE
63 * FOR ANY DIRECT, INDIRECT, INCIDENTAL, SPECIAL, EXEMPLARY, OR CONSEQUENTIAL
64 * DAMAGES (INCLUDING, BUT NOT LIMITED TO, PROCUREMENT OF SUBSTITUTE GOODS
65 * OR SERVICES; LOSS OF USE, DATA, OR PROFITS; OR BUSINESS INTERRUPTION)
66 * HOWEVER CAUSED AND ON ANY THEORY OF LIABILITY, WHETHER IN CONTRACT, STRICT
67 * LIABILITY, OR TORT (INCLUDING NEGLIGENCE OR OTHERWISE) ARISING IN ANY WAY
68 * OUT OF THE USE OF THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF
69 * SUCH DAMAGE.
70 *
71 * New Swap System
72 * Matthew Dillon
73 *
74 * Radix Bitmap 'blists'.
75 *
76 * - The new swapper uses the new radix bitmap code. This should scale
77 * to arbitrarily small or arbitrarily large swap spaces and an almost
78 * arbitrary degree of fragmentation.
79 *
80 * Features:
81 *
82 * - on the fly reallocation of swap during putpages. The new system
83 * does not try to keep previously allocated swap blocks for dirty
84 * pages.
85 *
86 * - on the fly deallocation of swap
87 *
88 * - No more garbage collection required. Unnecessarily allocated swap
89 * blocks only exist for dirty vm_page_t's now and these are already
90 * cycled (in a high-load system) by the pager. We also do on-the-fly
91 * removal of invalidated swap blocks when a page is destroyed
92 * or renamed.
93 *
94 * from: Utah $Hdr: swap_pager.c 1.4 91/04/30$
95 *
96 * @(#)swap_pager.c 8.9 (Berkeley) 3/21/94
97 *
98 * $FreeBSD: src/sys/vm/swap_pager.c,v 1.130.2.12 2002/08/31 21:15:55 dillon Exp $
4ecf7cc9 99 * $DragonFly: src/sys/vm/swap_pager.c,v 1.32 2008/07/01 02:02:56 dillon Exp $
984263bc
MD
100 */
101
102#include <sys/param.h>
103#include <sys/systm.h>
104#include <sys/conf.h>
105#include <sys/kernel.h>
106#include <sys/proc.h>
107#include <sys/buf.h>
108#include <sys/vnode.h>
109#include <sys/malloc.h>
110#include <sys/vmmeter.h>
111#include <sys/sysctl.h>
112#include <sys/blist.h>
113#include <sys/lock.h>
cdd46d2e 114#include <sys/thread2.h>
984263bc
MD
115
116#ifndef MAX_PAGEOUT_CLUSTER
117#define MAX_PAGEOUT_CLUSTER 16
118#endif
119
120#define SWB_NPAGES MAX_PAGEOUT_CLUSTER
121
122#include "opt_swap.h"
123#include <vm/vm.h>
124#include <vm/vm_object.h>
125#include <vm/vm_page.h>
126#include <vm/vm_pager.h>
127#include <vm/vm_pageout.h>
128#include <vm/swap_pager.h>
129#include <vm/vm_extern.h>
130#include <vm/vm_zone.h>
131
3020e3be 132#include <sys/buf2.h>
12e4aaff 133#include <vm/vm_page2.h>
3020e3be 134
984263bc
MD
135#define SWM_FREE 0x02 /* free, period */
136#define SWM_POP 0x04 /* pop out */
137
9e45bec7
MD
138#define AUTOCHAINDONE ((struct buf *)(intptr_t)-1)
139
984263bc
MD
140/*
141 * vm_swap_size is in page-sized chunks now. It was DEV_BSIZE'd chunks
142 * in the old system.
143 */
144
145extern int vm_swap_size; /* number of free swap blocks, in pages */
146
147int swap_pager_full; /* swap space exhaustion (task killing) */
148static int swap_pager_almost_full; /* swap space exhaustion (w/ hysteresis)*/
149static int nsw_rcount; /* free read buffers */
150static int nsw_wcount_sync; /* limit write buffers / synchronous */
151static int nsw_wcount_async; /* limit write buffers / asynchronous */
152static int nsw_wcount_async_max;/* assigned maximum */
153static int nsw_cluster_max; /* maximum VOP I/O allowed */
154static int sw_alloc_interlock; /* swap pager allocation interlock */
155
156struct blist *swapblist;
157static struct swblock **swhash;
158static int swhash_mask;
159static int swap_async_max = 4; /* maximum in-progress async I/O's */
160
161extern struct vnode *swapdev_vp; /* from vm_swap.c */
162
163SYSCTL_INT(_vm, OID_AUTO, swap_async_max,
164 CTLFLAG_RW, &swap_async_max, 0, "Maximum running async swap ops");
165
166/*
167 * "named" and "unnamed" anon region objects. Try to reduce the overhead
168 * of searching a named list by hashing it just a little.
169 */
170
171#define NOBJLISTS 8
172
173#define NOBJLIST(handle) \
174 (&swap_pager_object_list[((int)(intptr_t)handle >> 4) & (NOBJLISTS-1)])
175
176static struct pagerlst swap_pager_object_list[NOBJLISTS];
177struct pagerlst swap_pager_un_object_list;
178vm_zone_t swap_zone;
179
180/*
181 * pagerops for OBJT_SWAP - "swap pager". Some ops are also global procedure
182 * calls hooked from other parts of the VM system and do not appear here.
183 * (see vm/swap_pager.h).
184 */
185
186static vm_object_t
57f7b636
MD
187 swap_pager_alloc (void *handle, off_t size,
188 vm_prot_t prot, off_t offset);
1388df65
RG
189static void swap_pager_dealloc (vm_object_t object);
190static int swap_pager_getpages (vm_object_t, vm_page_t *, int, int);
191static void swap_pager_init (void);
192static void swap_pager_unswapped (vm_page_t);
81b5c339
MD
193static void swap_pager_strategy (vm_object_t, struct bio *);
194static void swap_chain_iodone(struct bio *biox);
984263bc
MD
195
196struct pagerops swappagerops = {
197 swap_pager_init, /* early system initialization of pager */
198 swap_pager_alloc, /* allocate an OBJT_SWAP object */
199 swap_pager_dealloc, /* deallocate an OBJT_SWAP object */
200 swap_pager_getpages, /* pagein */
201 swap_pager_putpages, /* pageout */
202 swap_pager_haspage, /* get backing store status for page */
203 swap_pager_unswapped, /* remove swap related to page */
204 swap_pager_strategy /* pager strategy call */
205};
206
207/*
208 * dmmax is in page-sized chunks with the new swap system. It was
209 * dev-bsized chunks in the old. dmmax is always a power of 2.
210 *
211 * swap_*() routines are externally accessible. swp_*() routines are
212 * internal.
213 */
214
215int dmmax;
216static int dmmax_mask;
217int nswap_lowat = 128; /* in pages, swap_pager_almost_full warn */
218int nswap_hiwat = 512; /* in pages, swap_pager_almost_full warn */
219
1388df65 220static __inline void swp_sizecheck (void);
81b5c339
MD
221static void swp_pager_sync_iodone (struct bio *bio);
222static void swp_pager_async_iodone (struct bio *bio);
984263bc
MD
223
224/*
225 * Swap bitmap functions
226 */
227
1388df65
RG
228static __inline void swp_pager_freeswapspace (daddr_t blk, int npages);
229static __inline daddr_t swp_pager_getswapspace (int npages);
984263bc
MD
230
231/*
232 * Metadata functions
233 */
234
1388df65
RG
235static void swp_pager_meta_build (vm_object_t, vm_pindex_t, daddr_t);
236static void swp_pager_meta_free (vm_object_t, vm_pindex_t, daddr_t);
237static void swp_pager_meta_free_all (vm_object_t);
238static daddr_t swp_pager_meta_ctl (vm_object_t, vm_pindex_t, int);
984263bc
MD
239
240/*
241 * SWP_SIZECHECK() - update swap_pager_full indication
242 *
243 * update the swap_pager_almost_full indication and warn when we are
244 * about to run out of swap space, using lowat/hiwat hysteresis.
245 *
246 * Clear swap_pager_full ( task killing ) indication when lowat is met.
247 *
248 * No restrictions on call
249 * This routine may not block.
250 * This routine must be called at splvm()
251 */
252
253static __inline void
57e43348 254swp_sizecheck(void)
984263bc
MD
255{
256 if (vm_swap_size < nswap_lowat) {
257 if (swap_pager_almost_full == 0) {
086c1d7e 258 kprintf("swap_pager: out of swap space\n");
984263bc
MD
259 swap_pager_almost_full = 1;
260 }
261 } else {
262 swap_pager_full = 0;
263 if (vm_swap_size > nswap_hiwat)
264 swap_pager_almost_full = 0;
265 }
266}
267
268/*
269 * SWAP_PAGER_INIT() - initialize the swap pager!
270 *
271 * Expected to be started from system init. NOTE: This code is run
272 * before much else so be careful what you depend on. Most of the VM
273 * system has yet to be initialized at this point.
274 */
275
276static void
57e43348 277swap_pager_init(void)
984263bc
MD
278{
279 /*
280 * Initialize object lists
281 */
282 int i;
283
284 for (i = 0; i < NOBJLISTS; ++i)
285 TAILQ_INIT(&swap_pager_object_list[i]);
286 TAILQ_INIT(&swap_pager_un_object_list);
287
288 /*
289 * Device Stripe, in PAGE_SIZE'd blocks
290 */
291
292 dmmax = SWB_NPAGES * 2;
293 dmmax_mask = ~(dmmax - 1);
294}
295
296/*
297 * SWAP_PAGER_SWAP_INIT() - swap pager initialization from pageout process
298 *
299 * Expected to be started from pageout process once, prior to entering
300 * its main loop.
301 */
302
303void
57e43348 304swap_pager_swap_init(void)
984263bc
MD
305{
306 int n, n2;
307
308 /*
309 * Number of in-transit swap bp operations. Don't
310 * exhaust the pbufs completely. Make sure we
311 * initialize workable values (0 will work for hysteresis
312 * but it isn't very efficient).
313 *
54f51aeb
HP
314 * The nsw_cluster_max is constrained by the number of pages an XIO
315 * holds, i.e., (MAXPHYS/PAGE_SIZE) and our locally defined
984263bc
MD
316 * MAX_PAGEOUT_CLUSTER. Also be aware that swap ops are
317 * constrained by the swap device interleave stripe size.
318 *
319 * Currently we hardwire nsw_wcount_async to 4. This limit is
320 * designed to prevent other I/O from having high latencies due to
321 * our pageout I/O. The value 4 works well for one or two active swap
322 * devices but is probably a little low if you have more. Even so,
323 * a higher value would probably generate only a limited improvement
324 * with three or four active swap devices since the system does not
325 * typically have to pageout at extreme bandwidths. We will want
326 * at least 2 per swap devices, and 4 is a pretty good value if you
327 * have one NFS swap device due to the command/ack latency over NFS.
328 * So it all works out pretty well.
329 */
330
331 nsw_cluster_max = min((MAXPHYS/PAGE_SIZE), MAX_PAGEOUT_CLUSTER);
332
333 nsw_rcount = (nswbuf + 1) / 2;
334 nsw_wcount_sync = (nswbuf + 3) / 4;
335 nsw_wcount_async = 4;
336 nsw_wcount_async_max = nsw_wcount_async;
337
338 /*
339 * Initialize our zone. Right now I'm just guessing on the number
340 * we need based on the number of pages in the system. Each swblock
341 * can hold 16 pages, so this is probably overkill. This reservation
342 * is typically limited to around 32MB by default.
343 */
12e4aaff 344 n = vmstats.v_page_count / 2;
984263bc
MD
345 if (maxswzone && n > maxswzone / sizeof(struct swblock))
346 n = maxswzone / sizeof(struct swblock);
347 n2 = n;
348
349 do {
350 swap_zone = zinit(
351 "SWAPMETA",
352 sizeof(struct swblock),
353 n,
354 ZONE_INTERRUPT,
355 1);
356 if (swap_zone != NULL)
357 break;
358 /*
359 * if the allocation failed, try a zone two thirds the
360 * size of the previous attempt.
361 */
362 n -= ((n + 2) / 3);
363 } while (n > 0);
364
365 if (swap_zone == NULL)
366 panic("swap_pager_swap_init: swap_zone == NULL");
367 if (n2 != n)
086c1d7e 368 kprintf("Swap zone entries reduced from %d to %d.\n", n2, n);
984263bc
MD
369 n2 = n;
370
371 /*
372 * Initialize our meta-data hash table. The swapper does not need to
373 * be quite as efficient as the VM system, so we do not use an
374 * oversized hash table.
375 *
376 * n: size of hash table, must be power of 2
377 * swhash_mask: hash table index mask
378 */
379
380 for (n = 1; n < n2 / 8; n *= 2)
381 ;
382
e7b4468c
SW
383 swhash = kmalloc(sizeof(struct swblock *) * n, M_VMPGDATA,
384 M_WAITOK | M_ZERO);
984263bc
MD
385
386 swhash_mask = n - 1;
387}
388
389/*
390 * SWAP_PAGER_ALLOC() - allocate a new OBJT_SWAP VM object and instantiate
391 * its metadata structures.
392 *
393 * This routine is called from the mmap and fork code to create a new
394 * OBJT_SWAP object. We do this by creating an OBJT_DEFAULT object
395 * and then converting it with swp_pager_meta_build().
396 *
397 * This routine may block in vm_object_allocate() and create a named
398 * object lookup race, so we must interlock. We must also run at
399 * splvm() for the object lookup to handle races with interrupts, but
400 * we do not have to maintain splvm() in between the lookup and the
401 * add because (I believe) it is not possible to attempt to create
402 * a new swap object w/handle when a default object with that handle
403 * already exists.
404 */
405
406static vm_object_t
57f7b636 407swap_pager_alloc(void *handle, off_t size, vm_prot_t prot, off_t offset)
984263bc
MD
408{
409 vm_object_t object;
410
411 if (handle) {
412 /*
413 * Reference existing named region or allocate new one. There
414 * should not be a race here against swp_pager_meta_build()
415 * as called from vm_page_remove() in regards to the lookup
416 * of the handle.
417 */
418
419 while (sw_alloc_interlock) {
420 sw_alloc_interlock = -1;
377d4740 421 tsleep(&sw_alloc_interlock, 0, "swpalc", 0);
984263bc
MD
422 }
423 sw_alloc_interlock = 1;
424
425 object = vm_pager_object_lookup(NOBJLIST(handle), handle);
426
427 if (object != NULL) {
428 vm_object_reference(object);
429 } else {
430 object = vm_object_allocate(OBJT_DEFAULT,
431 OFF_TO_IDX(offset + PAGE_MASK + size));
432 object->handle = handle;
433
434 swp_pager_meta_build(object, 0, SWAPBLK_NONE);
435 }
436
437 if (sw_alloc_interlock < 0)
438 wakeup(&sw_alloc_interlock);
439
440 sw_alloc_interlock = 0;
441 } else {
442 object = vm_object_allocate(OBJT_DEFAULT,
443 OFF_TO_IDX(offset + PAGE_MASK + size));
444
445 swp_pager_meta_build(object, 0, SWAPBLK_NONE);
446 }
447
448 return (object);
449}
450
451/*
452 * SWAP_PAGER_DEALLOC() - remove swap metadata from object
453 *
454 * The swap backing for the object is destroyed. The code is
455 * designed such that we can reinstantiate it later, but this
456 * routine is typically called only when the entire object is
457 * about to be destroyed.
458 *
459 * This routine may block, but no longer does.
460 *
461 * The object must be locked or unreferenceable.
462 */
463
464static void
57e43348 465swap_pager_dealloc(vm_object_t object)
984263bc 466{
984263bc
MD
467 /*
468 * Remove from list right away so lookups will fail if we block for
469 * pageout completion.
470 */
471
472 if (object->handle == NULL) {
473 TAILQ_REMOVE(&swap_pager_un_object_list, object, pager_object_list);
474 } else {
475 TAILQ_REMOVE(NOBJLIST(object->handle), object, pager_object_list);
476 }
477
478 vm_object_pip_wait(object, "swpdea");
479
480 /*
481 * Free all remaining metadata. We only bother to free it from
482 * the swap meta data. We do not attempt to free swapblk's still
483 * associated with vm_page_t's for this object. We do not care
484 * if paging is still in progress on some objects.
485 */
cdd46d2e 486 crit_enter();
984263bc 487 swp_pager_meta_free_all(object);
cdd46d2e 488 crit_exit();
984263bc
MD
489}
490
491/************************************************************************
492 * SWAP PAGER BITMAP ROUTINES *
493 ************************************************************************/
494
495/*
496 * SWP_PAGER_GETSWAPSPACE() - allocate raw swap space
497 *
498 * Allocate swap for the requested number of pages. The starting
499 * swap block number (a page index) is returned or SWAPBLK_NONE
500 * if the allocation failed.
501 *
502 * Also has the side effect of advising that somebody made a mistake
503 * when they configured swap and didn't configure enough.
504 *
505 * Must be called at splvm() to avoid races with bitmap frees from
506 * vm_page_remove() aka swap_pager_page_removed().
507 *
508 * This routine may not block
509 * This routine must be called at splvm().
510 */
511
512static __inline daddr_t
57e43348 513swp_pager_getswapspace(int npages)
984263bc
MD
514{
515 daddr_t blk;
516
517 if ((blk = blist_alloc(swapblist, npages)) == SWAPBLK_NONE) {
518 if (swap_pager_full != 2) {
086c1d7e 519 kprintf("swap_pager_getswapspace: failed\n");
984263bc
MD
520 swap_pager_full = 2;
521 swap_pager_almost_full = 1;
522 }
523 } else {
524 vm_swap_size -= npages;
525 swp_sizecheck();
526 }
527 return(blk);
528}
529
530/*
531 * SWP_PAGER_FREESWAPSPACE() - free raw swap space
532 *
533 * This routine returns the specified swap blocks back to the bitmap.
534 *
535 * Note: This routine may not block (it could in the old swap code),
536 * and through the use of the new blist routines it does not block.
537 *
538 * We must be called at splvm() to avoid races with bitmap frees from
539 * vm_page_remove() aka swap_pager_page_removed().
540 *
541 * This routine may not block
542 * This routine must be called at splvm().
543 */
544
545static __inline void
57e43348 546swp_pager_freeswapspace(daddr_t blk, int npages)
984263bc
MD
547{
548 blist_free(swapblist, blk, npages);
549 vm_swap_size += npages;
550 swp_sizecheck();
551}
552
553/*
554 * SWAP_PAGER_FREESPACE() - frees swap blocks associated with a page
555 * range within an object.
556 *
557 * This is a globally accessible routine.
558 *
559 * This routine removes swapblk assignments from swap metadata.
560 *
561 * The external callers of this routine typically have already destroyed
562 * or renamed vm_page_t's associated with this range in the object so
563 * we should be ok.
564 *
565 * This routine may be called at any spl. We up our spl to splvm temporarily
566 * in order to perform the metadata removal.
567 */
568
569void
57e43348 570swap_pager_freespace(vm_object_t object, vm_pindex_t start, vm_size_t size)
984263bc 571{
cdd46d2e 572 crit_enter();
984263bc 573 swp_pager_meta_free(object, start, size);
cdd46d2e 574 crit_exit();
984263bc
MD
575}
576
577/*
578 * SWAP_PAGER_RESERVE() - reserve swap blocks in object
579 *
580 * Assigns swap blocks to the specified range within the object. The
581 * swap blocks are not zerod. Any previous swap assignment is destroyed.
582 *
583 * Returns 0 on success, -1 on failure.
584 */
585
586int
587swap_pager_reserve(vm_object_t object, vm_pindex_t start, vm_size_t size)
588{
984263bc
MD
589 int n = 0;
590 daddr_t blk = SWAPBLK_NONE;
591 vm_pindex_t beg = start; /* save start index */
592
cdd46d2e 593 crit_enter();
984263bc
MD
594 while (size) {
595 if (n == 0) {
596 n = BLIST_MAX_ALLOC;
597 while ((blk = swp_pager_getswapspace(n)) == SWAPBLK_NONE) {
598 n >>= 1;
599 if (n == 0) {
600 swp_pager_meta_free(object, beg, start - beg);
cdd46d2e 601 crit_exit();
984263bc
MD
602 return(-1);
603 }
604 }
605 }
606 swp_pager_meta_build(object, start, blk);
607 --size;
608 ++start;
609 ++blk;
610 --n;
611 }
612 swp_pager_meta_free(object, start, n);
cdd46d2e 613 crit_exit();
984263bc
MD
614 return(0);
615}
616
617/*
618 * SWAP_PAGER_COPY() - copy blocks from source pager to destination pager
619 * and destroy the source.
620 *
621 * Copy any valid swapblks from the source to the destination. In
622 * cases where both the source and destination have a valid swapblk,
623 * we keep the destination's.
624 *
625 * This routine is allowed to block. It may block allocating metadata
626 * indirectly through swp_pager_meta_build() or if paging is still in
627 * progress on the source.
628 *
629 * This routine can be called at any spl
630 *
631 * XXX vm_page_collapse() kinda expects us not to block because we
632 * supposedly do not need to allocate memory, but for the moment we
633 * *may* have to get a little memory from the zone allocator, but
634 * it is taken from the interrupt memory. We should be ok.
635 *
636 * The source object contains no vm_page_t's (which is just as well)
637 *
638 * The source object is of type OBJT_SWAP.
639 *
640 * The source and destination objects must be locked or
641 * inaccessible (XXX are they ?)
642 */
643
644void
57e43348
MD
645swap_pager_copy(vm_object_t srcobject, vm_object_t dstobject,
646 vm_pindex_t offset, int destroysource)
984263bc
MD
647{
648 vm_pindex_t i;
984263bc 649
cdd46d2e 650 crit_enter();
984263bc
MD
651
652 /*
653 * If destroysource is set, we remove the source object from the
654 * swap_pager internal queue now.
655 */
656
657 if (destroysource) {
658 if (srcobject->handle == NULL) {
659 TAILQ_REMOVE(
660 &swap_pager_un_object_list,
661 srcobject,
662 pager_object_list
663 );
664 } else {
665 TAILQ_REMOVE(
666 NOBJLIST(srcobject->handle),
667 srcobject,
668 pager_object_list
669 );
670 }
671 }
672
673 /*
674 * transfer source to destination.
675 */
676
677 for (i = 0; i < dstobject->size; ++i) {
678 daddr_t dstaddr;
679
680 /*
681 * Locate (without changing) the swapblk on the destination,
682 * unless it is invalid in which case free it silently, or
683 * if the destination is a resident page, in which case the
684 * source is thrown away.
685 */
686
687 dstaddr = swp_pager_meta_ctl(dstobject, i, 0);
688
689 if (dstaddr == SWAPBLK_NONE) {
690 /*
691 * Destination has no swapblk and is not resident,
692 * copy source.
693 */
694 daddr_t srcaddr;
695
696 srcaddr = swp_pager_meta_ctl(
697 srcobject,
698 i + offset,
699 SWM_POP
700 );
701
702 if (srcaddr != SWAPBLK_NONE)
703 swp_pager_meta_build(dstobject, i, srcaddr);
704 } else {
705 /*
706 * Destination has valid swapblk or it is represented
707 * by a resident page. We destroy the sourceblock.
708 */
709
710 swp_pager_meta_ctl(srcobject, i + offset, SWM_FREE);
711 }
712 }
713
714 /*
715 * Free left over swap blocks in source.
716 *
717 * We have to revert the type to OBJT_DEFAULT so we do not accidently
718 * double-remove the object from the swap queues.
719 */
720
721 if (destroysource) {
722 swp_pager_meta_free_all(srcobject);
723 /*
724 * Reverting the type is not necessary, the caller is going
725 * to destroy srcobject directly, but I'm doing it here
726 * for consistency since we've removed the object from its
727 * queues.
728 */
729 srcobject->type = OBJT_DEFAULT;
730 }
cdd46d2e 731 crit_exit();
984263bc
MD
732}
733
734/*
735 * SWAP_PAGER_HASPAGE() - determine if we have good backing store for
736 * the requested page.
737 *
738 * We determine whether good backing store exists for the requested
739 * page and return TRUE if it does, FALSE if it doesn't.
740 *
741 * If TRUE, we also try to determine how much valid, contiguous backing
742 * store exists before and after the requested page within a reasonable
743 * distance. We do not try to restrict it to the swap device stripe
744 * (that is handled in getpages/putpages). It probably isn't worth
745 * doing here.
746 */
747
748boolean_t
57e43348
MD
749swap_pager_haspage(vm_object_t object, vm_pindex_t pindex, int *before,
750 int *after)
984263bc
MD
751{
752 daddr_t blk0;
984263bc
MD
753
754 /*
755 * do we have good backing store at the requested index ?
756 */
757
cdd46d2e 758 crit_enter();
984263bc
MD
759 blk0 = swp_pager_meta_ctl(object, pindex, 0);
760
761 if (blk0 == SWAPBLK_NONE) {
cdd46d2e 762 crit_exit();
984263bc
MD
763 if (before)
764 *before = 0;
765 if (after)
766 *after = 0;
767 return (FALSE);
768 }
769
770 /*
771 * find backwards-looking contiguous good backing store
772 */
773
774 if (before != NULL) {
775 int i;
776
777 for (i = 1; i < (SWB_NPAGES/2); ++i) {
778 daddr_t blk;
779
780 if (i > pindex)
781 break;
782 blk = swp_pager_meta_ctl(object, pindex - i, 0);
783 if (blk != blk0 - i)
784 break;
785 }
786 *before = (i - 1);
787 }
788
789 /*
790 * find forward-looking contiguous good backing store
791 */
792
793 if (after != NULL) {
794 int i;
795
796 for (i = 1; i < (SWB_NPAGES/2); ++i) {
797 daddr_t blk;
798
799 blk = swp_pager_meta_ctl(object, pindex + i, 0);
800 if (blk != blk0 + i)
801 break;
802 }
803 *after = (i - 1);
804 }
cdd46d2e 805 crit_exit();
984263bc
MD
806 return (TRUE);
807}
808
809/*
810 * SWAP_PAGER_PAGE_UNSWAPPED() - remove swap backing store related to page
811 *
812 * This removes any associated swap backing store, whether valid or
813 * not, from the page.
814 *
815 * This routine is typically called when a page is made dirty, at
816 * which point any associated swap can be freed. MADV_FREE also
817 * calls us in a special-case situation
818 *
819 * NOTE!!! If the page is clean and the swap was valid, the caller
820 * should make the page dirty before calling this routine. This routine
821 * does NOT change the m->dirty status of the page. Also: MADV_FREE
822 * depends on it.
823 *
824 * This routine may not block
825 * This routine must be called at splvm()
826 */
827
828static void
57e43348 829swap_pager_unswapped(vm_page_t m)
984263bc
MD
830{
831 swp_pager_meta_ctl(m->object, m->pindex, SWM_FREE);
832}
833
834/*
835 * SWAP_PAGER_STRATEGY() - read, write, free blocks
836 *
837 * This implements the vm_pager_strategy() interface to swap and allows
838 * other parts of the system to directly access swap as backing store
839 * through vm_objects of type OBJT_SWAP. This is intended to be a
840 * cacheless interface ( i.e. caching occurs at higher levels ).
841 * Therefore we do not maintain any resident pages. All I/O goes
842 * directly to and from the swap device.
843 *
984263bc
MD
844 * We currently attempt to run I/O synchronously or asynchronously as
845 * the caller requests. This isn't perfect because we loose error
846 * sequencing when we run multiple ops in parallel to satisfy a request.
847 * But this is swap, so we let it all hang out.
848 */
849
850static void
81b5c339 851swap_pager_strategy(vm_object_t object, struct bio *bio)
984263bc 852{
81b5c339
MD
853 struct buf *bp = bio->bio_buf;
854 struct bio *nbio;
984263bc 855 vm_pindex_t start;
54078292 856 vm_pindex_t biox_blkno = 0;
984263bc 857 int count;
984263bc 858 char *data;
81b5c339
MD
859 struct bio *biox = NULL;
860 struct buf *bufx = NULL;
861 struct bio_track *track;
862
863 /*
864 * tracking for swapdev vnode I/Os
865 */
10f3fee5 866 if (bp->b_cmd == BUF_CMD_READ)
81b5c339
MD
867 track = &swapdev_vp->v_track_read;
868 else
869 track = &swapdev_vp->v_track_write;
984263bc
MD
870
871 if (bp->b_bcount & PAGE_MASK) {
872 bp->b_error = EINVAL;
873 bp->b_flags |= B_ERROR | B_INVAL;
81b5c339 874 biodone(bio);
973c11b9
MD
875 kprintf("swap_pager_strategy: bp %p offset %lld size %d, "
876 "not page bounded\n",
877 bp, (long long)bio->bio_offset, (int)bp->b_bcount);
984263bc
MD
878 return;
879 }
880
881 /*
882 * Clear error indication, initialize page index, count, data pointer.
883 */
984263bc
MD
884 bp->b_error = 0;
885 bp->b_flags &= ~B_ERROR;
886 bp->b_resid = bp->b_bcount;
887
54078292 888 start = (vm_pindex_t)(bio->bio_offset >> PAGE_SHIFT);
984263bc
MD
889 count = howmany(bp->b_bcount, PAGE_SIZE);
890 data = bp->b_data;
891
cdd46d2e 892 crit_enter();
984263bc
MD
893
894 /*
10f3fee5 895 * Deal with BUF_CMD_FREEBLKS
984263bc 896 */
10f3fee5 897 if (bp->b_cmd == BUF_CMD_FREEBLKS) {
984263bc
MD
898 /*
899 * FREE PAGE(s) - destroy underlying swap that is no longer
900 * needed.
901 */
902 swp_pager_meta_free(object, start, count);
cdd46d2e 903 crit_exit();
984263bc 904 bp->b_resid = 0;
81b5c339 905 biodone(bio);
984263bc
MD
906 return;
907 }
908
909 /*
81b5c339
MD
910 * We need to be able to create a new cluster of I/O's. We cannot
911 * use the caller fields of the passed bio so push a new one.
912 *
913 * Because nbio is just a placeholder for the cluster links,
914 * we can biodone() the original bio instead of nbio to make
915 * things a bit more efficient.
916 */
917 nbio = push_bio(bio);
54078292 918 nbio->bio_offset = bio->bio_offset;
81b5c339
MD
919 nbio->bio_caller_info1.cluster_head = NULL;
920 nbio->bio_caller_info2.cluster_tail = NULL;
921
922 /*
984263bc
MD
923 * Execute read or write
924 */
925
926 while (count > 0) {
927 daddr_t blk;
928
929 /*
930 * Obtain block. If block not found and writing, allocate a
931 * new block and build it into the object.
932 */
933
934 blk = swp_pager_meta_ctl(object, start, 0);
10f3fee5 935 if ((blk == SWAPBLK_NONE) && bp->b_cmd != BUF_CMD_READ) {
984263bc
MD
936 blk = swp_pager_getswapspace(1);
937 if (blk == SWAPBLK_NONE) {
938 bp->b_error = ENOMEM;
939 bp->b_flags |= B_ERROR;
940 break;
941 }
942 swp_pager_meta_build(object, start, blk);
943 }
944
945 /*
946 * Do we have to flush our current collection? Yes if:
947 *
948 * - no swap block at this index
949 * - swap block is not contiguous
950 * - we cross a physical disk boundry in the
951 * stripe.
952 */
953
954 if (
54078292
MD
955 biox && (biox_blkno + btoc(bufx->b_bcount) != blk ||
956 ((biox_blkno ^ blk) & dmmax_mask)
984263bc
MD
957 )
958 ) {
cdd46d2e 959 crit_exit();
10f3fee5 960 if (bp->b_cmd == BUF_CMD_READ) {
12e4aaff 961 ++mycpu->gd_cnt.v_swapin;
81b5c339 962 mycpu->gd_cnt.v_swappgsin += btoc(bufx->b_bcount);
984263bc 963 } else {
12e4aaff 964 ++mycpu->gd_cnt.v_swapout;
81b5c339
MD
965 mycpu->gd_cnt.v_swappgsout += btoc(bufx->b_bcount);
966 bufx->b_dirtyend = bufx->b_bcount;
967 }
968
969 /*
970 * Flush the biox to the swap device.
971 */
972 if (bufx->b_bcount) {
10f3fee5 973 if (bufx->b_cmd != BUF_CMD_READ)
81b5c339
MD
974 bufx->b_dirtyend = bufx->b_bcount;
975 BUF_KERNPROC(bufx);
a8f169e2 976 vn_strategy(swapdev_vp, biox);
81b5c339
MD
977 } else {
978 biodone(biox);
984263bc 979 }
cdd46d2e 980 crit_enter();
81b5c339
MD
981 biox = NULL;
982 bufx = NULL;
984263bc
MD
983 }
984
985 /*
81b5c339 986 * Add new swapblk to biox, instantiating biox if necessary.
984263bc
MD
987 * Zero-fill reads are able to take a shortcut.
988 */
984263bc
MD
989 if (blk == SWAPBLK_NONE) {
990 /*
991 * We can only get here if we are reading. Since
992 * we are at splvm() we can safely modify b_resid,
993 * even if chain ops are in progress.
994 */
995 bzero(data, PAGE_SIZE);
996 bp->b_resid -= PAGE_SIZE;
997 } else {
81b5c339
MD
998 if (biox == NULL) {
999 /* XXX chain count > 4, wait to <= 4 */
1000
1001 bufx = getpbuf(NULL);
1002 biox = &bufx->b_bio1;
1003 cluster_append(nbio, bufx);
a8f169e2 1004 bufx->b_flags |= (bufx->b_flags & B_ORDERED) |
81b5c339 1005 B_ASYNC;
10f3fee5 1006 bufx->b_cmd = bp->b_cmd;
81b5c339 1007 biox->bio_done = swap_chain_iodone;
54078292 1008 biox->bio_offset = (off_t)blk << PAGE_SHIFT;
81b5c339 1009 biox->bio_caller_info1.cluster_parent = nbio;
54078292 1010 biox_blkno = blk;
81b5c339
MD
1011 bufx->b_bcount = 0;
1012 bufx->b_data = data;
984263bc 1013 }
81b5c339 1014 bufx->b_bcount += PAGE_SIZE;
984263bc
MD
1015 }
1016 --count;
1017 ++start;
1018 data += PAGE_SIZE;
1019 }
1020
1021 /*
1022 * Flush out last buffer
1023 */
cdd46d2e 1024 crit_exit();
984263bc 1025
81b5c339 1026 if (biox) {
984263bc 1027 if ((bp->b_flags & B_ASYNC) == 0)
81b5c339 1028 bufx->b_flags &= ~B_ASYNC;
10f3fee5 1029 if (bufx->b_cmd == BUF_CMD_READ) {
12e4aaff 1030 ++mycpu->gd_cnt.v_swapin;
81b5c339 1031 mycpu->gd_cnt.v_swappgsin += btoc(bufx->b_bcount);
984263bc 1032 } else {
12e4aaff 1033 ++mycpu->gd_cnt.v_swapout;
81b5c339
MD
1034 mycpu->gd_cnt.v_swappgsout += btoc(bufx->b_bcount);
1035 bufx->b_dirtyend = bufx->b_bcount;
1036 }
1037 if (bufx->b_bcount) {
10f3fee5 1038 if (bufx->b_cmd != BUF_CMD_READ)
81b5c339
MD
1039 bufx->b_dirtyend = bufx->b_bcount;
1040 BUF_KERNPROC(bufx);
a8f169e2 1041 vn_strategy(swapdev_vp, biox);
81b5c339
MD
1042 } else {
1043 biodone(biox);
984263bc 1044 }
81b5c339 1045 /* biox, bufx = NULL */
984263bc
MD
1046 }
1047
1048 /*
9e45bec7
MD
1049 * Wait for completion. Now that we are no longer using
1050 * cluster_append, use the cluster_tail field to indicate
1051 * auto-completion if there are still I/O's in progress.
984263bc 1052 */
984263bc 1053 if (bp->b_flags & B_ASYNC) {
81b5c339
MD
1054 crit_enter();
1055 if (nbio->bio_caller_info1.cluster_head == NULL) {
1056 biodone(bio);
1057 } else {
9e45bec7 1058 nbio->bio_caller_info2.cluster_tail = AUTOCHAINDONE;
81b5c339
MD
1059 }
1060 crit_exit();
984263bc 1061 } else {
81b5c339
MD
1062 crit_enter();
1063 while (nbio->bio_caller_info1.cluster_head != NULL) {
1064 bp->b_flags |= B_WANT;
1065 tsleep(bp, 0, "bpchain", 0);
1066 }
1067 if (bp->b_resid != 0 && !(bp->b_flags & B_ERROR)) {
1068 bp->b_flags |= B_ERROR;
1069 bp->b_error = EINVAL;
1070 }
1071 biodone(bio);
1072 crit_exit();
984263bc
MD
1073 }
1074}
1075
81b5c339
MD
1076static void
1077swap_chain_iodone(struct bio *biox)
1078{
1079 struct buf **nextp;
1080 struct buf *bufx; /* chained sub-buffer */
1081 struct bio *nbio; /* parent nbio with chain glue */
1082 struct buf *bp; /* original bp associated with nbio */
1083
1084 bufx = biox->bio_buf;
1085 nbio = biox->bio_caller_info1.cluster_parent;
1086 bp = nbio->bio_buf;
1087
1088 /*
1089 * Update the original buffer
1090 */
1091 KKASSERT(bp != NULL);
1092 if (bufx->b_flags & B_ERROR) {
1093 bp->b_flags |= B_ERROR;
1094 bp->b_error = bufx->b_error;
1095 } else if (bufx->b_resid != 0) {
1096 bp->b_flags |= B_ERROR;
1097 bp->b_error = EINVAL;
1098 } else {
1099 bp->b_resid -= bufx->b_bcount;
1100 }
1101
1102 /*
1103 * Remove us from the chain. It is sufficient to clean up
9e45bec7
MD
1104 * cluster_head. Once the chain is operational cluster_tail
1105 * may be used to indicate AUTOCHAINDONE. Note that I/O's
1106 * can complete while the swap system is still appending new
1107 * BIOs to the chain.
81b5c339
MD
1108 */
1109 nextp = &nbio->bio_caller_info1.cluster_head;
1110 while (*nextp != bufx) {
1111 KKASSERT(*nextp != NULL);
1112 nextp = &(*nextp)->b_cluster_next;
1113 }
1114 *nextp = bufx->b_cluster_next;
1115 if (bp->b_flags & B_WANT) {
1116 bp->b_flags &= ~B_WANT;
1117 wakeup(bp);
1118 }
1119
1120 /*
1121 * Clean up bufx. If this was the last buffer in the chain
9e45bec7 1122 * and AUTOCHAINDONE was set, finish off the original I/O
81b5c339
MD
1123 * as well.
1124 *
1125 * nbio was just a fake BIO layer to hold the cluster links,
1126 * we can issue the biodone() on the layer above it.
1127 */
1128 if (nbio->bio_caller_info1.cluster_head == NULL &&
9e45bec7
MD
1129 nbio->bio_caller_info2.cluster_tail == AUTOCHAINDONE
1130 ) {
1131 nbio->bio_caller_info2.cluster_tail = NULL;
81b5c339
MD
1132 if (bp->b_resid != 0 && !(bp->b_flags & B_ERROR)) {
1133 bp->b_flags |= B_ERROR;
1134 bp->b_error = EINVAL;
1135 }
1136 biodone(nbio->bio_prev);
1137 }
81b5c339
MD
1138 bufx->b_flags &= ~B_ASYNC;
1139 relpbuf(bufx, NULL);
1140}
1141
984263bc
MD
1142/*
1143 * SWAP_PAGER_GETPAGES() - bring pages in from swap
1144 *
1145 * Attempt to retrieve (m, count) pages from backing store, but make
1146 * sure we retrieve at least m[reqpage]. We try to load in as large
1147 * a chunk surrounding m[reqpage] as is contiguous in swap and which
1148 * belongs to the same object.
1149 *
1150 * The code is designed for asynchronous operation and
1151 * immediate-notification of 'reqpage' but tends not to be
1152 * used that way. Please do not optimize-out this algorithmic
1153 * feature, I intend to improve on it in the future.
1154 *
1155 * The parent has a single vm_object_pip_add() reference prior to
1156 * calling us and we should return with the same.
1157 *
1158 * The parent has BUSY'd the pages. We should return with 'm'
1159 * left busy, but the others adjusted.
1160 */
1161
1162static int
57e43348 1163swap_pager_getpages(vm_object_t object, vm_page_t *m, int count, int reqpage)
984263bc
MD
1164{
1165 struct buf *bp;
81b5c339 1166 struct bio *bio;
984263bc 1167 vm_page_t mreq;
984263bc
MD
1168 int i;
1169 int j;
1170 daddr_t blk;
1171 vm_offset_t kva;
1172 vm_pindex_t lastpindex;
1173
1174 mreq = m[reqpage];
1175
1176 if (mreq->object != object) {
1177 panic("swap_pager_getpages: object mismatch %p/%p",
1178 object,
1179 mreq->object
1180 );
1181 }
17cde63e 1182
984263bc
MD
1183 /*
1184 * Calculate range to retrieve. The pages have already been assigned
1185 * their swapblks. We require a *contiguous* range that falls entirely
1186 * within a single device stripe. If we do not supply it, bad things
1187 * happen. Note that blk, iblk & jblk can be SWAPBLK_NONE, but the
1188 * loops are set up such that the case(s) are handled implicitly.
1189 *
1190 * The swp_*() calls must be made at splvm(). vm_page_free() does
1191 * not need to be, but it will go a little faster if it is.
1192 */
cdd46d2e 1193 crit_enter();
984263bc
MD
1194 blk = swp_pager_meta_ctl(mreq->object, mreq->pindex, 0);
1195
1196 for (i = reqpage - 1; i >= 0; --i) {
1197 daddr_t iblk;
1198
1199 iblk = swp_pager_meta_ctl(m[i]->object, m[i]->pindex, 0);
1200 if (blk != iblk + (reqpage - i))
1201 break;
1202 if ((blk ^ iblk) & dmmax_mask)
1203 break;
1204 }
1205 ++i;
1206
1207 for (j = reqpage + 1; j < count; ++j) {
1208 daddr_t jblk;
1209
1210 jblk = swp_pager_meta_ctl(m[j]->object, m[j]->pindex, 0);
1211 if (blk != jblk - (j - reqpage))
1212 break;
1213 if ((blk ^ jblk) & dmmax_mask)
1214 break;
1215 }
1216
1217 /*
1218 * free pages outside our collection range. Note: we never free
1219 * mreq, it must remain busy throughout.
1220 */
1221
1222 {
1223 int k;
1224
1225 for (k = 0; k < i; ++k)
1226 vm_page_free(m[k]);
1227 for (k = j; k < count; ++k)
1228 vm_page_free(m[k]);
1229 }
cdd46d2e 1230 crit_exit();
984263bc
MD
1231
1232
1233 /*
1234 * Return VM_PAGER_FAIL if we have nothing to do. Return mreq
1235 * still busy, but the others unbusied.
1236 */
1237
1238 if (blk == SWAPBLK_NONE)
1239 return(VM_PAGER_FAIL);
1240
1241 /*
1242 * Get a swap buffer header to perform the IO
1243 */
1244
1245 bp = getpbuf(&nsw_rcount);
81b5c339 1246 bio = &bp->b_bio1;
984263bc
MD
1247 kva = (vm_offset_t) bp->b_data;
1248
1249 /*
1250 * map our page(s) into kva for input
984263bc
MD
1251 */
1252
1253 pmap_qenter(kva, m + i, j - i);
1254
984263bc 1255 bp->b_data = (caddr_t) kva;
984263bc 1256 bp->b_bcount = PAGE_SIZE * (j - i);
81b5c339 1257 bio->bio_done = swp_pager_async_iodone;
54078292 1258 bio->bio_offset = (off_t)(blk - (reqpage - i)) << PAGE_SHIFT;
973c11b9 1259 bio->bio_driver_info = (void *)(intptr_t)(reqpage - i);
984263bc
MD
1260
1261 {
1262 int k;
1263
1264 for (k = i; k < j; ++k) {
54f51aeb 1265 bp->b_xio.xio_pages[k - i] = m[k];
984263bc
MD
1266 vm_page_flag_set(m[k], PG_SWAPINPROG);
1267 }
1268 }
54f51aeb 1269 bp->b_xio.xio_npages = j - i;
984263bc 1270
12e4aaff 1271 mycpu->gd_cnt.v_swapin++;
54f51aeb 1272 mycpu->gd_cnt.v_swappgsin += bp->b_xio.xio_npages;
984263bc
MD
1273
1274 /*
1275 * We still hold the lock on mreq, and our automatic completion routine
1276 * does not remove it.
1277 */
1278
54f51aeb 1279 vm_object_pip_add(mreq->object, bp->b_xio.xio_npages);
984263bc
MD
1280 lastpindex = m[j-1]->pindex;
1281
1282 /*
1283 * perform the I/O. NOTE!!! bp cannot be considered valid after
1284 * this point because we automatically release it on completion.
1285 * Instead, we look at the one page we are interested in which we
1286 * still hold a lock on even through the I/O completion.
1287 *
1288 * The other pages in our m[] array are also released on completion,
1289 * so we cannot assume they are valid anymore either.
984263bc
MD
1290 */
1291
10f3fee5 1292 bp->b_cmd = BUF_CMD_READ;
984263bc 1293 BUF_KERNPROC(bp);
81b5c339 1294 vn_strategy(swapdev_vp, bio);
984263bc
MD
1295
1296 /*
1297 * wait for the page we want to complete. PG_SWAPINPROG is always
1298 * cleared on completion. If an I/O error occurs, SWAPBLK_NONE
1299 * is set in the meta-data.
1300 */
1301
cdd46d2e 1302 crit_enter();
984263bc
MD
1303
1304 while ((mreq->flags & PG_SWAPINPROG) != 0) {
1305 vm_page_flag_set(mreq, PG_WANTED | PG_REFERENCED);
12e4aaff 1306 mycpu->gd_cnt.v_intrans++;
377d4740 1307 if (tsleep(mreq, 0, "swread", hz*20)) {
086c1d7e 1308 kprintf(
81b5c339 1309 "swap_pager: indefinite wait buffer: "
973c11b9
MD
1310 " offset: %lld, size: %ld\n",
1311 (long long)bio->bio_offset,
1312 (long)bp->b_bcount
984263bc
MD
1313 );
1314 }
1315 }
1316
cdd46d2e 1317 crit_exit();
984263bc
MD
1318
1319 /*
1320 * mreq is left bussied after completion, but all the other pages
1321 * are freed. If we had an unrecoverable read error the page will
1322 * not be valid.
1323 */
1324
1325 if (mreq->valid != VM_PAGE_BITS_ALL) {
1326 return(VM_PAGER_ERROR);
1327 } else {
1328 return(VM_PAGER_OK);
1329 }
1330
1331 /*
1332 * A final note: in a low swap situation, we cannot deallocate swap
1333 * and mark a page dirty here because the caller is likely to mark
1334 * the page clean when we return, causing the page to possibly revert
1335 * to all-zero's later.
1336 */
1337}
1338
1339/*
1340 * swap_pager_putpages:
1341 *
1342 * Assign swap (if necessary) and initiate I/O on the specified pages.
1343 *
1344 * We support both OBJT_DEFAULT and OBJT_SWAP objects. DEFAULT objects
1345 * are automatically converted to SWAP objects.
1346 *
81b5c339 1347 * In a low memory situation we may block in vn_strategy(), but the new
984263bc
MD
1348 * vm_page reservation system coupled with properly written VFS devices
1349 * should ensure that no low-memory deadlock occurs. This is an area
1350 * which needs work.
1351 *
1352 * The parent has N vm_object_pip_add() references prior to
1353 * calling us and will remove references for rtvals[] that are
1354 * not set to VM_PAGER_PEND. We need to remove the rest on I/O
1355 * completion.
1356 *
1357 * The parent has soft-busy'd the pages it passes us and will unbusy
1358 * those whos rtvals[] entry is not set to VM_PAGER_PEND on return.
1359 * We need to unbusy the rest on I/O completion.
1360 */
984263bc 1361void
17cde63e
MD
1362swap_pager_putpages(vm_object_t object, vm_page_t *m, int count,
1363 boolean_t sync, int *rtvals)
984263bc
MD
1364{
1365 int i;
1366 int n = 0;
1367
1368 if (count && m[0]->object != object) {
1369 panic("swap_pager_getpages: object mismatch %p/%p",
1370 object,
1371 m[0]->object
1372 );
1373 }
17cde63e 1374
984263bc
MD
1375 /*
1376 * Step 1
1377 *
1378 * Turn object into OBJT_SWAP
1379 * check for bogus sysops
1380 * force sync if not pageout process
1381 */
1382
1383 if (object->type != OBJT_SWAP)
1384 swp_pager_meta_build(object, 0, SWAPBLK_NONE);
1385
bc6dffab 1386 if (curthread != pagethread)
984263bc
MD
1387 sync = TRUE;
1388
1389 /*
1390 * Step 2
1391 *
1392 * Update nsw parameters from swap_async_max sysctl values.
1393 * Do not let the sysop crash the machine with bogus numbers.
1394 */
1395
1396 if (swap_async_max != nsw_wcount_async_max) {
1397 int n;
984263bc
MD
1398
1399 /*
1400 * limit range
1401 */
1402 if ((n = swap_async_max) > nswbuf / 2)
1403 n = nswbuf / 2;
1404 if (n < 1)
1405 n = 1;
1406 swap_async_max = n;
1407
1408 /*
1409 * Adjust difference ( if possible ). If the current async
1410 * count is too low, we may not be able to make the adjustment
1411 * at this time.
1412 */
cdd46d2e 1413 crit_enter();
984263bc
MD
1414 n -= nsw_wcount_async_max;
1415 if (nsw_wcount_async + n >= 0) {
1416 nsw_wcount_async += n;
1417 nsw_wcount_async_max += n;
1418 wakeup(&nsw_wcount_async);
1419 }
cdd46d2e 1420 crit_exit();
984263bc
MD
1421 }
1422
1423 /*
1424 * Step 3
1425 *
1426 * Assign swap blocks and issue I/O. We reallocate swap on the fly.
1427 * The page is left dirty until the pageout operation completes
1428 * successfully.
1429 */
1430
1431 for (i = 0; i < count; i += n) {
984263bc 1432 struct buf *bp;
81b5c339 1433 struct bio *bio;
984263bc 1434 daddr_t blk;
81b5c339 1435 int j;
984263bc
MD
1436
1437 /*
1438 * Maximum I/O size is limited by a number of factors.
1439 */
1440
1441 n = min(BLIST_MAX_ALLOC, count - i);
1442 n = min(n, nsw_cluster_max);
1443
cdd46d2e 1444 crit_enter();
984263bc
MD
1445
1446 /*
1447 * Get biggest block of swap we can. If we fail, fall
1448 * back and try to allocate a smaller block. Don't go
1449 * overboard trying to allocate space if it would overly
1450 * fragment swap.
1451 */
1452 while (
1453 (blk = swp_pager_getswapspace(n)) == SWAPBLK_NONE &&
1454 n > 4
1455 ) {
1456 n >>= 1;
1457 }
1458 if (blk == SWAPBLK_NONE) {
1459 for (j = 0; j < n; ++j)
1460 rtvals[i+j] = VM_PAGER_FAIL;
cdd46d2e 1461 crit_exit();
984263bc
MD
1462 continue;
1463 }
1464
1465 /*
1466 * The I/O we are constructing cannot cross a physical
1467 * disk boundry in the swap stripe. Note: we are still
1468 * at splvm().
1469 */
1470 if ((blk ^ (blk + n)) & dmmax_mask) {
1471 j = ((blk + dmmax) & dmmax_mask) - blk;
1472 swp_pager_freeswapspace(blk + j, n - j);
1473 n = j;
1474 }
1475
1476 /*
1477 * All I/O parameters have been satisfied, build the I/O
1478 * request and assign the swap space.
984263bc
MD
1479 */
1480
10f3fee5 1481 if (sync == TRUE)
984263bc 1482 bp = getpbuf(&nsw_wcount_sync);
10f3fee5 1483 else
984263bc 1484 bp = getpbuf(&nsw_wcount_async);
81b5c339 1485 bio = &bp->b_bio1;
984263bc
MD
1486
1487 pmap_qenter((vm_offset_t)bp->b_data, &m[i], n);
1488
984263bc 1489 bp->b_bcount = PAGE_SIZE * n;
54078292 1490 bio->bio_offset = (off_t)blk << PAGE_SHIFT;
984263bc 1491
984263bc
MD
1492 for (j = 0; j < n; ++j) {
1493 vm_page_t mreq = m[i+j];
1494
1495 swp_pager_meta_build(
1496 mreq->object,
1497 mreq->pindex,
1498 blk + j
1499 );
1500 vm_page_dirty(mreq);
1501 rtvals[i+j] = VM_PAGER_OK;
1502
1503 vm_page_flag_set(mreq, PG_SWAPINPROG);
54f51aeb 1504 bp->b_xio.xio_pages[j] = mreq;
984263bc 1505 }
54f51aeb 1506 bp->b_xio.xio_npages = n;
984263bc 1507
12e4aaff 1508 mycpu->gd_cnt.v_swapout++;
54f51aeb 1509 mycpu->gd_cnt.v_swappgsout += bp->b_xio.xio_npages;
984263bc 1510
cdd46d2e 1511 crit_exit();
984263bc 1512
10f3fee5
MD
1513 bp->b_dirtyoff = 0; /* req'd for NFS */
1514 bp->b_dirtyend = bp->b_bcount; /* req'd for NFS */
1515 bp->b_cmd = BUF_CMD_WRITE;
1516
984263bc
MD
1517 /*
1518 * asynchronous
984263bc 1519 */
984263bc 1520 if (sync == FALSE) {
10f3fee5 1521 bp->b_flags |= B_ASYNC;
81b5c339 1522 bio->bio_done = swp_pager_async_iodone;
984263bc 1523 BUF_KERNPROC(bp);
81b5c339 1524 vn_strategy(swapdev_vp, bio);
984263bc
MD
1525
1526 for (j = 0; j < n; ++j)
1527 rtvals[i+j] = VM_PAGER_PEND;
1528 continue;
1529 }
1530
1531 /*
1532 * synchronous
984263bc
MD
1533 */
1534
81b5c339
MD
1535 bio->bio_done = swp_pager_sync_iodone;
1536 vn_strategy(swapdev_vp, bio);
984263bc
MD
1537
1538 /*
1539 * Wait for the sync I/O to complete, then update rtvals.
1540 * We just set the rtvals[] to VM_PAGER_PEND so we can call
1541 * our async completion routine at the end, thus avoiding a
1542 * double-free.
1543 */
cdd46d2e 1544 crit_enter();
984263bc 1545
10f3fee5 1546 while (bp->b_cmd != BUF_CMD_DONE)
377d4740 1547 tsleep(bp, 0, "swwrt", 0);
984263bc
MD
1548
1549 for (j = 0; j < n; ++j)
1550 rtvals[i+j] = VM_PAGER_PEND;
1551
1552 /*
1553 * Now that we are through with the bp, we can call the
1554 * normal async completion, which frees everything up.
1555 */
1556
81b5c339 1557 swp_pager_async_iodone(bio);
984263bc 1558
cdd46d2e 1559 crit_exit();
984263bc
MD
1560 }
1561}
1562
c84c24da
MD
1563void
1564swap_pager_newswap(void)
1565{
1566 swp_sizecheck();
1567}
1568
984263bc
MD
1569/*
1570 * swap_pager_sync_iodone:
1571 *
1572 * Completion routine for synchronous reads and writes from/to swap.
1573 * We just mark the bp is complete and wake up anyone waiting on it.
1574 *
c84c24da
MD
1575 * This routine may not block. This routine is called at splbio()
1576 * or better.
984263bc
MD
1577 */
1578
1579static void
81b5c339 1580swp_pager_sync_iodone(struct bio *bio)
984263bc 1581{
81b5c339
MD
1582 struct buf *bp = bio->bio_buf;
1583
984263bc 1584 bp->b_flags &= ~B_ASYNC;
10f3fee5 1585 bp->b_cmd = BUF_CMD_DONE;
984263bc
MD
1586 wakeup(bp);
1587}
1588
1589/*
1590 * swp_pager_async_iodone:
1591 *
1592 * Completion routine for asynchronous reads and writes from/to swap.
1593 * Also called manually by synchronous code to finish up a bp.
1594 *
1595 * For READ operations, the pages are PG_BUSY'd. For WRITE operations,
1596 * the pages are vm_page_t->busy'd. For READ operations, we PG_BUSY
1597 * unbusy all pages except the 'main' request page. For WRITE
1598 * operations, we vm_page_t->busy'd unbusy all pages ( we can do this
1599 * because we marked them all VM_PAGER_PEND on return from putpages ).
1600 *
1601 * This routine may not block.
984263bc
MD
1602 */
1603
1604static void
81b5c339 1605swp_pager_async_iodone(struct bio *bio)
984263bc 1606{
81b5c339 1607 struct buf *bp = bio->bio_buf;
984263bc 1608 vm_object_t object = NULL;
81b5c339 1609 int i;
10f3fee5 1610 int *nswptr;
984263bc
MD
1611
1612 /*
1613 * report error
1614 */
984263bc 1615 if (bp->b_flags & B_ERROR) {
086c1d7e 1616 kprintf(
54078292 1617 "swap_pager: I/O error - %s failed; offset %lld,"
984263bc 1618 "size %ld, error %d\n",
10f3fee5 1619 ((bp->b_cmd == BUF_CMD_READ) ? "pagein" : "pageout"),
973c11b9 1620 (long long)bio->bio_offset,
984263bc
MD
1621 (long)bp->b_bcount,
1622 bp->b_error
1623 );
1624 }
1625
1626 /*
1627 * set object, raise to splvm().
1628 */
1629
54f51aeb
HP
1630 if (bp->b_xio.xio_npages)
1631 object = bp->b_xio.xio_pages[0]->object;
cdd46d2e 1632 crit_enter();
984263bc
MD
1633
1634 /*
1635 * remove the mapping for kernel virtual
1636 */
1637
54f51aeb 1638 pmap_qremove((vm_offset_t)bp->b_data, bp->b_xio.xio_npages);
984263bc
MD
1639
1640 /*
1641 * cleanup pages. If an error occurs writing to swap, we are in
1642 * very serious trouble. If it happens to be a disk error, though,
1643 * we may be able to recover by reassigning the swap later on. So
1644 * in this case we remove the m->swapblk assignment for the page
1645 * but do not free it in the rlist. The errornous block(s) are thus
1646 * never reallocated as swap. Redirty the page and continue.
1647 */
1648
54f51aeb
HP
1649 for (i = 0; i < bp->b_xio.xio_npages; ++i) {
1650 vm_page_t m = bp->b_xio.xio_pages[i];
984263bc
MD
1651
1652 vm_page_flag_clear(m, PG_SWAPINPROG);
1653
1654 if (bp->b_flags & B_ERROR) {
1655 /*
1656 * If an error occurs I'd love to throw the swapblk
1657 * away without freeing it back to swapspace, so it
1658 * can never be used again. But I can't from an
1659 * interrupt.
1660 */
1661
10f3fee5 1662 if (bp->b_cmd == BUF_CMD_READ) {
984263bc
MD
1663 /*
1664 * When reading, reqpage needs to stay
1665 * locked for the parent, but all other
1666 * pages can be freed. We still want to
1667 * wakeup the parent waiting on the page,
1668 * though. ( also: pg_reqpage can be -1 and
1669 * not match anything ).
1670 *
1671 * We have to wake specifically requested pages
1672 * up too because we cleared PG_SWAPINPROG and
1673 * someone may be waiting for that.
1674 *
1675 * NOTE: for reads, m->dirty will probably
1676 * be overridden by the original caller of
1677 * getpages so don't play cute tricks here.
1678 *
93afe6be
MD
1679 * NOTE: We can't actually free the page from
1680 * here, because this is an interrupt. It
1681 * is not legal to mess with object->memq
1682 * from an interrupt. Deactivate the page
1683 * instead.
984263bc
MD
1684 */
1685
1686 m->valid = 0;
1687 vm_page_flag_clear(m, PG_ZERO);
1688
81b5c339
MD
1689 /*
1690 * bio_driver_info holds the requested page
1691 * index.
1692 */
973c11b9 1693 if (i != (int)(intptr_t)bio->bio_driver_info) {
93afe6be
MD
1694 vm_page_deactivate(m);
1695 vm_page_wakeup(m);
1696 } else {
984263bc 1697 vm_page_flash(m);
93afe6be 1698 }
984263bc
MD
1699 /*
1700 * If i == bp->b_pager.pg_reqpage, do not wake
1701 * the page up. The caller needs to.
1702 */
1703 } else {
1704 /*
1705 * If a write error occurs, reactivate page
1706 * so it doesn't clog the inactive list,
1707 * then finish the I/O.
1708 */
1709 vm_page_dirty(m);
c84c24da 1710 kprintf("f");
984263bc
MD
1711 vm_page_activate(m);
1712 vm_page_io_finish(m);
1713 }
10f3fee5 1714 } else if (bp->b_cmd == BUF_CMD_READ) {
984263bc 1715 /*
984263bc
MD
1716 * NOTE: for reads, m->dirty will probably be
1717 * overridden by the original caller of getpages so
1718 * we cannot set them in order to free the underlying
1719 * swap in a low-swap situation. I don't think we'd
1720 * want to do that anyway, but it was an optimization
1721 * that existed in the old swapper for a time before
1722 * it got ripped out due to precisely this problem.
1723 *
1724 * clear PG_ZERO in page.
1725 *
1726 * If not the requested page then deactivate it.
1727 *
1728 * Note that the requested page, reqpage, is left
1729 * busied, but we still have to wake it up. The
1730 * other pages are released (unbusied) by
1731 * vm_page_wakeup(). We do not set reqpage's
1732 * valid bits here, it is up to the caller.
1733 */
1734
4530a3aa
MD
1735 /*
1736 * NOTE: can't call pmap_clear_modify(m) from an
1737 * interrupt thread, the pmap code may have to map
1738 * non-kernel pmaps and currently asserts the case.
1739 */
1740 /*pmap_clear_modify(m);*/
984263bc
MD
1741 m->valid = VM_PAGE_BITS_ALL;
1742 vm_page_undirty(m);
1743 vm_page_flag_clear(m, PG_ZERO);
1744
1745 /*
1746 * We have to wake specifically requested pages
1747 * up too because we cleared PG_SWAPINPROG and
1748 * could be waiting for it in getpages. However,
1749 * be sure to not unbusy getpages specifically
1750 * requested page - getpages expects it to be
1751 * left busy.
81b5c339
MD
1752 *
1753 * bio_driver_info holds the requested page
984263bc 1754 */
973c11b9 1755 if (i != (int)(intptr_t)bio->bio_driver_info) {
984263bc
MD
1756 vm_page_deactivate(m);
1757 vm_page_wakeup(m);
1758 } else {
1759 vm_page_flash(m);
1760 }
1761 } else {
1762 /*
93afe6be
MD
1763 * Mark the page clean but do not mess with the
1764 * pmap-layer's modified state. That state should
1765 * also be clear since the caller protected the
1766 * page VM_PROT_READ, but allow the case.
1767 *
1768 * We are in an interrupt, avoid pmap operations.
1769 *
1770 * If we have a severe page deficit, deactivate the
1771 * page. Do not try to cache it (which would also
1772 * involve a pmap op), because the page might still
1773 * be read-heavy.
984263bc 1774 */
984263bc
MD
1775 vm_page_undirty(m);
1776 vm_page_io_finish(m);
93afe6be
MD
1777 if (vm_page_count_severe())
1778 vm_page_deactivate(m);
1779#if 0
984263bc
MD
1780 if (!vm_page_count_severe() || !vm_page_try_to_cache(m))
1781 vm_page_protect(m, VM_PROT_READ);
93afe6be 1782#endif
984263bc
MD
1783 }
1784 }
1785
1786 /*
1787 * adjust pip. NOTE: the original parent may still have its own
1788 * pip refs on the object.
1789 */
1790
1791 if (object)
54f51aeb 1792 vm_object_pip_wakeupn(object, bp->b_xio.xio_npages);
984263bc
MD
1793
1794 /*
1795 * release the physical I/O buffer
1796 */
10f3fee5
MD
1797 if (bp->b_cmd == BUF_CMD_READ)
1798 nswptr = &nsw_rcount;
1799 else if (bp->b_flags & B_ASYNC)
1800 nswptr = &nsw_wcount_async;
1801 else
1802 nswptr = &nsw_wcount_sync;
1803 bp->b_cmd = BUF_CMD_DONE;
1804 relpbuf(bp, nswptr);
cdd46d2e 1805 crit_exit();
984263bc
MD
1806}
1807
1808/************************************************************************
1809 * SWAP META DATA *
1810 ************************************************************************
1811 *
1812 * These routines manipulate the swap metadata stored in the
1813 * OBJT_SWAP object. All swp_*() routines must be called at
1814 * splvm() because swap can be freed up by the low level vm_page
1815 * code which might be called from interrupts beyond what splbio() covers.
1816 *
1817 * Swap metadata is implemented with a global hash and not directly
1818 * linked into the object. Instead the object simply contains
1819 * appropriate tracking counters.
1820 */
1821
1822/*
1823 * SWP_PAGER_HASH() - hash swap meta data
1824 *
1825 * This is an inline helper function which hashes the swapblk given
1826 * the object and page index. It returns a pointer to a pointer
1827 * to the object, or a pointer to a NULL pointer if it could not
1828 * find a swapblk.
1829 *
1830 * This routine must be called at splvm().
1831 */
1832
1833static __inline struct swblock **
1834swp_pager_hash(vm_object_t object, vm_pindex_t index)
1835{
1836 struct swblock **pswap;
1837 struct swblock *swap;
1838
1839 index &= ~SWAP_META_MASK;
1840 pswap = &swhash[(index ^ (int)(intptr_t)object) & swhash_mask];
1841
1842 while ((swap = *pswap) != NULL) {
1843 if (swap->swb_object == object &&
1844 swap->swb_index == index
1845 ) {
1846 break;
1847 }
1848 pswap = &swap->swb_hnext;
1849 }
1850 return(pswap);
1851}
1852
1853/*
1854 * SWP_PAGER_META_BUILD() - add swap block to swap meta data for object
1855 *
1856 * We first convert the object to a swap object if it is a default
1857 * object.
1858 *
1859 * The specified swapblk is added to the object's swap metadata. If
1860 * the swapblk is not valid, it is freed instead. Any previously
1861 * assigned swapblk is freed.
1862 *
1863 * This routine must be called at splvm(), except when used to convert
1864 * an OBJT_DEFAULT object into an OBJT_SWAP object.
1865
1866 */
1867
1868static void
1869swp_pager_meta_build(
1870 vm_object_t object,
1871 vm_pindex_t index,
1872 daddr_t swapblk
1873) {
1874 struct swblock *swap;
1875 struct swblock **pswap;
1876
1877 /*
1878 * Convert default object to swap object if necessary
1879 */
1880
1881 if (object->type != OBJT_SWAP) {
1882 object->type = OBJT_SWAP;
1883 object->un_pager.swp.swp_bcount = 0;
1884
1885 if (object->handle != NULL) {
1886 TAILQ_INSERT_TAIL(
1887 NOBJLIST(object->handle),
1888 object,
1889 pager_object_list
1890 );
1891 } else {
1892 TAILQ_INSERT_TAIL(
1893 &swap_pager_un_object_list,
1894 object,
1895 pager_object_list
1896 );
1897 }
1898 }
1899
1900 /*
1901 * Locate hash entry. If not found create, but if we aren't adding
1902 * anything just return. If we run out of space in the map we wait
1903 * and, since the hash table may have changed, retry.
1904 */
1905
1906retry:
1907 pswap = swp_pager_hash(object, index);
1908
1909 if ((swap = *pswap) == NULL) {
1910 int i;
1911
1912 if (swapblk == SWAPBLK_NONE)
1913 return;
1914
1915 swap = *pswap = zalloc(swap_zone);
1916 if (swap == NULL) {
4ecf7cc9 1917 vm_wait(0);
984263bc
MD
1918 goto retry;
1919 }
1920 swap->swb_hnext = NULL;
1921 swap->swb_object = object;
1922 swap->swb_index = index & ~SWAP_META_MASK;
1923 swap->swb_count = 0;
1924
1925 ++object->un_pager.swp.swp_bcount;
1926
1927 for (i = 0; i < SWAP_META_PAGES; ++i)
1928 swap->swb_pages[i] = SWAPBLK_NONE;
1929 }
1930
1931 /*
1932 * Delete prior contents of metadata
1933 */
1934
1935 index &= SWAP_META_MASK;
1936
1937 if (swap->swb_pages[index] != SWAPBLK_NONE) {
1938 swp_pager_freeswapspace(swap->swb_pages[index], 1);
1939 --swap->swb_count;
1940 }
1941
1942 /*
1943 * Enter block into metadata
1944 */
1945
1946 swap->swb_pages[index] = swapblk;
1947 if (swapblk != SWAPBLK_NONE)
1948 ++swap->swb_count;
1949}
1950
1951/*
1952 * SWP_PAGER_META_FREE() - free a range of blocks in the object's swap metadata
1953 *
1954 * The requested range of blocks is freed, with any associated swap
1955 * returned to the swap bitmap.
1956 *
1957 * This routine will free swap metadata structures as they are cleaned
1958 * out. This routine does *NOT* operate on swap metadata associated
1959 * with resident pages.
1960 *
1961 * This routine must be called at splvm()
1962 */
1963
1964static void
1965swp_pager_meta_free(vm_object_t object, vm_pindex_t index, daddr_t count)
1966{
1967 if (object->type != OBJT_SWAP)
1968 return;
1969
1970 while (count > 0) {
1971 struct swblock **pswap;
1972 struct swblock *swap;
1973
1974 pswap = swp_pager_hash(object, index);
1975
1976 if ((swap = *pswap) != NULL) {
1977 daddr_t v = swap->swb_pages[index & SWAP_META_MASK];
1978
1979 if (v != SWAPBLK_NONE) {
1980 swp_pager_freeswapspace(v, 1);
1981 swap->swb_pages[index & SWAP_META_MASK] =
1982 SWAPBLK_NONE;
1983 if (--swap->swb_count == 0) {
1984 *pswap = swap->swb_hnext;
1985 zfree(swap_zone, swap);
1986 --object->un_pager.swp.swp_bcount;
1987 }
1988 }
1989 --count;
1990 ++index;
1991 } else {
1992 int n = SWAP_META_PAGES - (index & SWAP_META_MASK);
1993 count -= n;
1994 index += n;
1995 }
1996 }
1997}
1998
1999/*
2000 * SWP_PAGER_META_FREE_ALL() - destroy all swap metadata associated with object
2001 *
2002 * This routine locates and destroys all swap metadata associated with
2003 * an object.
2004 *
2005 * This routine must be called at splvm()
2006 */
2007
2008static void
2009swp_pager_meta_free_all(vm_object_t object)
2010{
2011 daddr_t index = 0;
2012
2013 if (object->type != OBJT_SWAP)
2014 return;
2015
2016 while (object->un_pager.swp.swp_bcount) {
2017 struct swblock **pswap;
2018 struct swblock *swap;
2019
2020 pswap = swp_pager_hash(object, index);
2021 if ((swap = *pswap) != NULL) {
2022 int i;
2023
2024 for (i = 0; i < SWAP_META_PAGES; ++i) {
2025 daddr_t v = swap->swb_pages[i];
2026 if (v != SWAPBLK_NONE) {
2027 --swap->swb_count;
2028 swp_pager_freeswapspace(v, 1);
2029 }
2030 }
2031 if (swap->swb_count != 0)
2032 panic("swap_pager_meta_free_all: swb_count != 0");
2033 *pswap = swap->swb_hnext;
2034 zfree(swap_zone, swap);
2035 --object->un_pager.swp.swp_bcount;
2036 }
2037 index += SWAP_META_PAGES;
2038 if (index > 0x20000000)
2039 panic("swp_pager_meta_free_all: failed to locate all swap meta blocks");
2040 }
2041}
2042
2043/*
2044 * SWP_PAGER_METACTL() - misc control of swap and vm_page_t meta data.
2045 *
2046 * This routine is capable of looking up, popping, or freeing
2047 * swapblk assignments in the swap meta data or in the vm_page_t.
2048 * The routine typically returns the swapblk being looked-up, or popped,
2049 * or SWAPBLK_NONE if the block was freed, or SWAPBLK_NONE if the block
2050 * was invalid. This routine will automatically free any invalid
2051 * meta-data swapblks.
2052 *
2053 * It is not possible to store invalid swapblks in the swap meta data
2054 * (other then a literal 'SWAPBLK_NONE'), so we don't bother checking.
2055 *
2056 * When acting on a busy resident page and paging is in progress, we
2057 * have to wait until paging is complete but otherwise can act on the
2058 * busy page.
2059 *
2060 * This routine must be called at splvm().
2061 *
2062 * SWM_FREE remove and free swap block from metadata
2063 * SWM_POP remove from meta data but do not free.. pop it out
2064 */
2065
2066static daddr_t
2067swp_pager_meta_ctl(
2068 vm_object_t object,
2069 vm_pindex_t index,
2070 int flags
2071) {
2072 struct swblock **pswap;
2073 struct swblock *swap;
2074 daddr_t r1;
2075
2076 /*
2077 * The meta data only exists of the object is OBJT_SWAP
2078 * and even then might not be allocated yet.
2079 */
2080
2081 if (object->type != OBJT_SWAP)
2082 return(SWAPBLK_NONE);
2083
2084 r1 = SWAPBLK_NONE;
2085 pswap = swp_pager_hash(object, index);
2086
2087 if ((swap = *pswap) != NULL) {
2088 index &= SWAP_META_MASK;
2089 r1 = swap->swb_pages[index];
2090
2091 if (r1 != SWAPBLK_NONE) {
2092 if (flags & SWM_FREE) {
2093 swp_pager_freeswapspace(r1, 1);
2094 r1 = SWAPBLK_NONE;
2095 }
2096 if (flags & (SWM_FREE|SWM_POP)) {
2097 swap->swb_pages[index] = SWAPBLK_NONE;
2098 if (--swap->swb_count == 0) {
2099 *pswap = swap->swb_hnext;
2100 zfree(swap_zone, swap);
2101 --object->un_pager.swp.swp_bcount;
2102 }
2103 }
2104 }
2105 }
2106 return(r1);
2107}