kernel - simplify vm pager ops, add pre-faulting for zero-fill pages.
[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
8aa92e4b
MD
138#define SWBIO_READ 0x01
139#define SWBIO_WRITE 0x02
140#define SWBIO_SYNC 0x04
141
984263bc
MD
142/*
143 * vm_swap_size is in page-sized chunks now. It was DEV_BSIZE'd chunks
144 * in the old system.
145 */
146
147extern int vm_swap_size; /* number of free swap blocks, in pages */
148
149int swap_pager_full; /* swap space exhaustion (task killing) */
150static int swap_pager_almost_full; /* swap space exhaustion (w/ hysteresis)*/
151static int nsw_rcount; /* free read buffers */
152static int nsw_wcount_sync; /* limit write buffers / synchronous */
153static int nsw_wcount_async; /* limit write buffers / asynchronous */
154static int nsw_wcount_async_max;/* assigned maximum */
155static int nsw_cluster_max; /* maximum VOP I/O allowed */
156static int sw_alloc_interlock; /* swap pager allocation interlock */
157
158struct blist *swapblist;
159static struct swblock **swhash;
160static int swhash_mask;
161static int swap_async_max = 4; /* maximum in-progress async I/O's */
162
163extern struct vnode *swapdev_vp; /* from vm_swap.c */
164
165SYSCTL_INT(_vm, OID_AUTO, swap_async_max,
166 CTLFLAG_RW, &swap_async_max, 0, "Maximum running async swap ops");
167
168/*
169 * "named" and "unnamed" anon region objects. Try to reduce the overhead
170 * of searching a named list by hashing it just a little.
171 */
172
173#define NOBJLISTS 8
174
175#define NOBJLIST(handle) \
176 (&swap_pager_object_list[((int)(intptr_t)handle >> 4) & (NOBJLISTS-1)])
177
178static struct pagerlst swap_pager_object_list[NOBJLISTS];
179struct pagerlst swap_pager_un_object_list;
180vm_zone_t swap_zone;
181
182/*
183 * pagerops for OBJT_SWAP - "swap pager". Some ops are also global procedure
184 * calls hooked from other parts of the VM system and do not appear here.
185 * (see vm/swap_pager.h).
186 */
187
188static vm_object_t
57f7b636
MD
189 swap_pager_alloc (void *handle, off_t size,
190 vm_prot_t prot, off_t offset);
1388df65 191static void swap_pager_dealloc (vm_object_t object);
1b9d3514 192static int swap_pager_getpage (vm_object_t, vm_page_t *, int);
1388df65
RG
193static void swap_pager_init (void);
194static void swap_pager_unswapped (vm_page_t);
81b5c339
MD
195static void swap_pager_strategy (vm_object_t, struct bio *);
196static void swap_chain_iodone(struct bio *biox);
984263bc
MD
197
198struct pagerops swappagerops = {
199 swap_pager_init, /* early system initialization of pager */
200 swap_pager_alloc, /* allocate an OBJT_SWAP object */
201 swap_pager_dealloc, /* deallocate an OBJT_SWAP object */
1b9d3514 202 swap_pager_getpage, /* pagein */
984263bc
MD
203 swap_pager_putpages, /* pageout */
204 swap_pager_haspage, /* get backing store status for page */
205 swap_pager_unswapped, /* remove swap related to page */
206 swap_pager_strategy /* pager strategy call */
207};
208
209/*
210 * dmmax is in page-sized chunks with the new swap system. It was
211 * dev-bsized chunks in the old. dmmax is always a power of 2.
212 *
213 * swap_*() routines are externally accessible. swp_*() routines are
214 * internal.
215 */
216
217int dmmax;
218static int dmmax_mask;
219int nswap_lowat = 128; /* in pages, swap_pager_almost_full warn */
220int nswap_hiwat = 512; /* in pages, swap_pager_almost_full warn */
221
1388df65 222static __inline void swp_sizecheck (void);
81b5c339 223static void swp_pager_async_iodone (struct bio *bio);
984263bc
MD
224
225/*
226 * Swap bitmap functions
227 */
228
1388df65
RG
229static __inline void swp_pager_freeswapspace (daddr_t blk, int npages);
230static __inline daddr_t swp_pager_getswapspace (int npages);
984263bc
MD
231
232/*
233 * Metadata functions
234 */
235
1388df65
RG
236static void swp_pager_meta_build (vm_object_t, vm_pindex_t, daddr_t);
237static void swp_pager_meta_free (vm_object_t, vm_pindex_t, daddr_t);
238static void swp_pager_meta_free_all (vm_object_t);
239static daddr_t swp_pager_meta_ctl (vm_object_t, vm_pindex_t, int);
984263bc
MD
240
241/*
242 * SWP_SIZECHECK() - update swap_pager_full indication
243 *
244 * update the swap_pager_almost_full indication and warn when we are
245 * about to run out of swap space, using lowat/hiwat hysteresis.
246 *
247 * Clear swap_pager_full ( task killing ) indication when lowat is met.
248 *
249 * No restrictions on call
250 * This routine may not block.
251 * This routine must be called at splvm()
252 */
253
254static __inline void
57e43348 255swp_sizecheck(void)
984263bc
MD
256{
257 if (vm_swap_size < nswap_lowat) {
258 if (swap_pager_almost_full == 0) {
086c1d7e 259 kprintf("swap_pager: out of swap space\n");
984263bc
MD
260 swap_pager_almost_full = 1;
261 }
262 } else {
263 swap_pager_full = 0;
264 if (vm_swap_size > nswap_hiwat)
265 swap_pager_almost_full = 0;
266 }
267}
268
269/*
270 * SWAP_PAGER_INIT() - initialize the swap pager!
271 *
272 * Expected to be started from system init. NOTE: This code is run
273 * before much else so be careful what you depend on. Most of the VM
274 * system has yet to be initialized at this point.
275 */
276
277static void
57e43348 278swap_pager_init(void)
984263bc
MD
279{
280 /*
281 * Initialize object lists
282 */
283 int i;
284
285 for (i = 0; i < NOBJLISTS; ++i)
286 TAILQ_INIT(&swap_pager_object_list[i]);
287 TAILQ_INIT(&swap_pager_un_object_list);
288
289 /*
290 * Device Stripe, in PAGE_SIZE'd blocks
291 */
292
293 dmmax = SWB_NPAGES * 2;
294 dmmax_mask = ~(dmmax - 1);
295}
296
297/*
298 * SWAP_PAGER_SWAP_INIT() - swap pager initialization from pageout process
299 *
300 * Expected to be started from pageout process once, prior to entering
301 * its main loop.
302 */
303
304void
57e43348 305swap_pager_swap_init(void)
984263bc
MD
306{
307 int n, n2;
308
309 /*
310 * Number of in-transit swap bp operations. Don't
311 * exhaust the pbufs completely. Make sure we
312 * initialize workable values (0 will work for hysteresis
313 * but it isn't very efficient).
314 *
54f51aeb
HP
315 * The nsw_cluster_max is constrained by the number of pages an XIO
316 * holds, i.e., (MAXPHYS/PAGE_SIZE) and our locally defined
984263bc
MD
317 * MAX_PAGEOUT_CLUSTER. Also be aware that swap ops are
318 * constrained by the swap device interleave stripe size.
319 *
320 * Currently we hardwire nsw_wcount_async to 4. This limit is
321 * designed to prevent other I/O from having high latencies due to
322 * our pageout I/O. The value 4 works well for one or two active swap
323 * devices but is probably a little low if you have more. Even so,
324 * a higher value would probably generate only a limited improvement
325 * with three or four active swap devices since the system does not
326 * typically have to pageout at extreme bandwidths. We will want
327 * at least 2 per swap devices, and 4 is a pretty good value if you
328 * have one NFS swap device due to the command/ack latency over NFS.
329 * So it all works out pretty well.
330 */
331
332 nsw_cluster_max = min((MAXPHYS/PAGE_SIZE), MAX_PAGEOUT_CLUSTER);
333
334 nsw_rcount = (nswbuf + 1) / 2;
335 nsw_wcount_sync = (nswbuf + 3) / 4;
336 nsw_wcount_async = 4;
337 nsw_wcount_async_max = nsw_wcount_async;
338
339 /*
79634a66
MD
340 * The zone is dynamically allocated so generally size it to
341 * maxswzone (32MB to 512MB of KVM). Set a minimum size based
342 * on physical memory of around 8x (each swblock can hold 16 pages).
343 *
344 * With the advent of SSDs (vs HDs) the practical (swap:memory) ratio
345 * has increased dramatically.
984263bc 346 */
12e4aaff 347 n = vmstats.v_page_count / 2;
79634a66 348 if (maxswzone && n < maxswzone / sizeof(struct swblock))
984263bc
MD
349 n = maxswzone / sizeof(struct swblock);
350 n2 = n;
351
352 do {
353 swap_zone = zinit(
354 "SWAPMETA",
355 sizeof(struct swblock),
356 n,
357 ZONE_INTERRUPT,
358 1);
359 if (swap_zone != NULL)
360 break;
361 /*
362 * if the allocation failed, try a zone two thirds the
363 * size of the previous attempt.
364 */
365 n -= ((n + 2) / 3);
366 } while (n > 0);
367
368 if (swap_zone == NULL)
369 panic("swap_pager_swap_init: swap_zone == NULL");
370 if (n2 != n)
086c1d7e 371 kprintf("Swap zone entries reduced from %d to %d.\n", n2, n);
984263bc
MD
372 n2 = n;
373
374 /*
375 * Initialize our meta-data hash table. The swapper does not need to
376 * be quite as efficient as the VM system, so we do not use an
377 * oversized hash table.
378 *
379 * n: size of hash table, must be power of 2
380 * swhash_mask: hash table index mask
381 */
382
383 for (n = 1; n < n2 / 8; n *= 2)
384 ;
385
e7b4468c
SW
386 swhash = kmalloc(sizeof(struct swblock *) * n, M_VMPGDATA,
387 M_WAITOK | M_ZERO);
984263bc
MD
388
389 swhash_mask = n - 1;
390}
391
392/*
393 * SWAP_PAGER_ALLOC() - allocate a new OBJT_SWAP VM object and instantiate
394 * its metadata structures.
395 *
396 * This routine is called from the mmap and fork code to create a new
397 * OBJT_SWAP object. We do this by creating an OBJT_DEFAULT object
398 * and then converting it with swp_pager_meta_build().
399 *
400 * This routine may block in vm_object_allocate() and create a named
401 * object lookup race, so we must interlock. We must also run at
402 * splvm() for the object lookup to handle races with interrupts, but
403 * we do not have to maintain splvm() in between the lookup and the
404 * add because (I believe) it is not possible to attempt to create
405 * a new swap object w/handle when a default object with that handle
406 * already exists.
407 */
408
409static vm_object_t
57f7b636 410swap_pager_alloc(void *handle, off_t size, vm_prot_t prot, off_t offset)
984263bc
MD
411{
412 vm_object_t object;
413
414 if (handle) {
415 /*
416 * Reference existing named region or allocate new one. There
417 * should not be a race here against swp_pager_meta_build()
418 * as called from vm_page_remove() in regards to the lookup
419 * of the handle.
420 */
421
422 while (sw_alloc_interlock) {
423 sw_alloc_interlock = -1;
377d4740 424 tsleep(&sw_alloc_interlock, 0, "swpalc", 0);
984263bc
MD
425 }
426 sw_alloc_interlock = 1;
427
428 object = vm_pager_object_lookup(NOBJLIST(handle), handle);
429
430 if (object != NULL) {
431 vm_object_reference(object);
432 } else {
433 object = vm_object_allocate(OBJT_DEFAULT,
434 OFF_TO_IDX(offset + PAGE_MASK + size));
435 object->handle = handle;
436
437 swp_pager_meta_build(object, 0, SWAPBLK_NONE);
438 }
439
440 if (sw_alloc_interlock < 0)
441 wakeup(&sw_alloc_interlock);
442
443 sw_alloc_interlock = 0;
444 } else {
445 object = vm_object_allocate(OBJT_DEFAULT,
446 OFF_TO_IDX(offset + PAGE_MASK + size));
447
448 swp_pager_meta_build(object, 0, SWAPBLK_NONE);
449 }
450
451 return (object);
452}
453
454/*
455 * SWAP_PAGER_DEALLOC() - remove swap metadata from object
456 *
457 * The swap backing for the object is destroyed. The code is
458 * designed such that we can reinstantiate it later, but this
459 * routine is typically called only when the entire object is
460 * about to be destroyed.
461 *
462 * This routine may block, but no longer does.
463 *
464 * The object must be locked or unreferenceable.
465 */
466
467static void
57e43348 468swap_pager_dealloc(vm_object_t object)
984263bc 469{
984263bc
MD
470 /*
471 * Remove from list right away so lookups will fail if we block for
472 * pageout completion.
473 */
474
475 if (object->handle == NULL) {
476 TAILQ_REMOVE(&swap_pager_un_object_list, object, pager_object_list);
477 } else {
478 TAILQ_REMOVE(NOBJLIST(object->handle), object, pager_object_list);
479 }
480
481 vm_object_pip_wait(object, "swpdea");
482
483 /*
484 * Free all remaining metadata. We only bother to free it from
485 * the swap meta data. We do not attempt to free swapblk's still
486 * associated with vm_page_t's for this object. We do not care
487 * if paging is still in progress on some objects.
488 */
cdd46d2e 489 crit_enter();
984263bc 490 swp_pager_meta_free_all(object);
cdd46d2e 491 crit_exit();
984263bc
MD
492}
493
494/************************************************************************
495 * SWAP PAGER BITMAP ROUTINES *
496 ************************************************************************/
497
498/*
499 * SWP_PAGER_GETSWAPSPACE() - allocate raw swap space
500 *
501 * Allocate swap for the requested number of pages. The starting
502 * swap block number (a page index) is returned or SWAPBLK_NONE
503 * if the allocation failed.
504 *
505 * Also has the side effect of advising that somebody made a mistake
506 * when they configured swap and didn't configure enough.
507 *
508 * 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 daddr_t
57e43348 516swp_pager_getswapspace(int npages)
984263bc
MD
517{
518 daddr_t blk;
519
520 if ((blk = blist_alloc(swapblist, npages)) == SWAPBLK_NONE) {
521 if (swap_pager_full != 2) {
086c1d7e 522 kprintf("swap_pager_getswapspace: failed\n");
984263bc
MD
523 swap_pager_full = 2;
524 swap_pager_almost_full = 1;
525 }
526 } else {
527 vm_swap_size -= npages;
528 swp_sizecheck();
529 }
530 return(blk);
531}
532
533/*
534 * SWP_PAGER_FREESWAPSPACE() - free raw swap space
535 *
536 * This routine returns the specified swap blocks back to the bitmap.
537 *
538 * Note: This routine may not block (it could in the old swap code),
539 * and through the use of the new blist routines it does not block.
540 *
541 * We must be called at splvm() to avoid races with bitmap frees from
542 * vm_page_remove() aka swap_pager_page_removed().
543 *
544 * This routine may not block
545 * This routine must be called at splvm().
546 */
547
548static __inline void
57e43348 549swp_pager_freeswapspace(daddr_t blk, int npages)
984263bc
MD
550{
551 blist_free(swapblist, blk, npages);
552 vm_swap_size += npages;
553 swp_sizecheck();
554}
555
556/*
557 * SWAP_PAGER_FREESPACE() - frees swap blocks associated with a page
558 * range within an object.
559 *
560 * This is a globally accessible routine.
561 *
562 * This routine removes swapblk assignments from swap metadata.
563 *
564 * The external callers of this routine typically have already destroyed
565 * or renamed vm_page_t's associated with this range in the object so
566 * we should be ok.
567 *
568 * This routine may be called at any spl. We up our spl to splvm temporarily
569 * in order to perform the metadata removal.
570 */
571
572void
57e43348 573swap_pager_freespace(vm_object_t object, vm_pindex_t start, vm_size_t size)
984263bc 574{
cdd46d2e 575 crit_enter();
984263bc 576 swp_pager_meta_free(object, start, size);
cdd46d2e 577 crit_exit();
984263bc
MD
578}
579
580/*
581 * SWAP_PAGER_RESERVE() - reserve swap blocks in object
582 *
583 * Assigns swap blocks to the specified range within the object. The
584 * swap blocks are not zerod. Any previous swap assignment is destroyed.
585 *
586 * Returns 0 on success, -1 on failure.
587 */
588
589int
590swap_pager_reserve(vm_object_t object, vm_pindex_t start, vm_size_t size)
591{
984263bc
MD
592 int n = 0;
593 daddr_t blk = SWAPBLK_NONE;
594 vm_pindex_t beg = start; /* save start index */
595
cdd46d2e 596 crit_enter();
984263bc
MD
597 while (size) {
598 if (n == 0) {
599 n = BLIST_MAX_ALLOC;
600 while ((blk = swp_pager_getswapspace(n)) == SWAPBLK_NONE) {
601 n >>= 1;
602 if (n == 0) {
603 swp_pager_meta_free(object, beg, start - beg);
cdd46d2e 604 crit_exit();
984263bc
MD
605 return(-1);
606 }
607 }
608 }
609 swp_pager_meta_build(object, start, blk);
610 --size;
611 ++start;
612 ++blk;
613 --n;
614 }
615 swp_pager_meta_free(object, start, n);
cdd46d2e 616 crit_exit();
984263bc
MD
617 return(0);
618}
619
620/*
621 * SWAP_PAGER_COPY() - copy blocks from source pager to destination pager
622 * and destroy the source.
623 *
624 * Copy any valid swapblks from the source to the destination. In
625 * cases where both the source and destination have a valid swapblk,
626 * we keep the destination's.
627 *
628 * This routine is allowed to block. It may block allocating metadata
629 * indirectly through swp_pager_meta_build() or if paging is still in
630 * progress on the source.
631 *
632 * This routine can be called at any spl
633 *
634 * XXX vm_page_collapse() kinda expects us not to block because we
635 * supposedly do not need to allocate memory, but for the moment we
636 * *may* have to get a little memory from the zone allocator, but
637 * it is taken from the interrupt memory. We should be ok.
638 *
639 * The source object contains no vm_page_t's (which is just as well)
640 *
641 * The source object is of type OBJT_SWAP.
642 *
643 * The source and destination objects must be locked or
644 * inaccessible (XXX are they ?)
645 */
646
647void
57e43348
MD
648swap_pager_copy(vm_object_t srcobject, vm_object_t dstobject,
649 vm_pindex_t offset, int destroysource)
984263bc
MD
650{
651 vm_pindex_t i;
984263bc 652
cdd46d2e 653 crit_enter();
984263bc
MD
654
655 /*
656 * If destroysource is set, we remove the source object from the
657 * swap_pager internal queue now.
658 */
659
660 if (destroysource) {
661 if (srcobject->handle == NULL) {
662 TAILQ_REMOVE(
663 &swap_pager_un_object_list,
664 srcobject,
665 pager_object_list
666 );
667 } else {
668 TAILQ_REMOVE(
669 NOBJLIST(srcobject->handle),
670 srcobject,
671 pager_object_list
672 );
673 }
674 }
675
676 /*
677 * transfer source to destination.
678 */
679
680 for (i = 0; i < dstobject->size; ++i) {
681 daddr_t dstaddr;
682
683 /*
684 * Locate (without changing) the swapblk on the destination,
685 * unless it is invalid in which case free it silently, or
686 * if the destination is a resident page, in which case the
687 * source is thrown away.
688 */
689
690 dstaddr = swp_pager_meta_ctl(dstobject, i, 0);
691
692 if (dstaddr == SWAPBLK_NONE) {
693 /*
694 * Destination has no swapblk and is not resident,
695 * copy source.
696 */
697 daddr_t srcaddr;
698
699 srcaddr = swp_pager_meta_ctl(
700 srcobject,
701 i + offset,
702 SWM_POP
703 );
704
705 if (srcaddr != SWAPBLK_NONE)
706 swp_pager_meta_build(dstobject, i, srcaddr);
707 } else {
708 /*
709 * Destination has valid swapblk or it is represented
710 * by a resident page. We destroy the sourceblock.
711 */
712
713 swp_pager_meta_ctl(srcobject, i + offset, SWM_FREE);
714 }
715 }
716
717 /*
718 * Free left over swap blocks in source.
719 *
720 * We have to revert the type to OBJT_DEFAULT so we do not accidently
721 * double-remove the object from the swap queues.
722 */
723
724 if (destroysource) {
725 swp_pager_meta_free_all(srcobject);
726 /*
727 * Reverting the type is not necessary, the caller is going
728 * to destroy srcobject directly, but I'm doing it here
729 * for consistency since we've removed the object from its
730 * queues.
731 */
732 srcobject->type = OBJT_DEFAULT;
733 }
cdd46d2e 734 crit_exit();
984263bc
MD
735}
736
737/*
738 * SWAP_PAGER_HASPAGE() - determine if we have good backing store for
739 * the requested page.
740 *
741 * We determine whether good backing store exists for the requested
742 * page and return TRUE if it does, FALSE if it doesn't.
743 *
744 * If TRUE, we also try to determine how much valid, contiguous backing
745 * store exists before and after the requested page within a reasonable
746 * distance. We do not try to restrict it to the swap device stripe
747 * (that is handled in getpages/putpages). It probably isn't worth
748 * doing here.
749 */
750
751boolean_t
1b9d3514 752swap_pager_haspage(vm_object_t object, vm_pindex_t pindex)
984263bc
MD
753{
754 daddr_t blk0;
984263bc
MD
755
756 /*
757 * do we have good backing store at the requested index ?
758 */
759
cdd46d2e 760 crit_enter();
984263bc
MD
761 blk0 = swp_pager_meta_ctl(object, pindex, 0);
762
763 if (blk0 == SWAPBLK_NONE) {
cdd46d2e 764 crit_exit();
984263bc
MD
765 return (FALSE);
766 }
767
1b9d3514 768#if 0
984263bc
MD
769 /*
770 * find backwards-looking contiguous good backing store
771 */
984263bc
MD
772 if (before != NULL) {
773 int i;
774
775 for (i = 1; i < (SWB_NPAGES/2); ++i) {
776 daddr_t blk;
777
778 if (i > pindex)
779 break;
780 blk = swp_pager_meta_ctl(object, pindex - i, 0);
781 if (blk != blk0 - i)
782 break;
783 }
784 *before = (i - 1);
785 }
786
787 /*
788 * find forward-looking contiguous good backing store
789 */
790
791 if (after != NULL) {
792 int i;
793
794 for (i = 1; i < (SWB_NPAGES/2); ++i) {
795 daddr_t blk;
796
797 blk = swp_pager_meta_ctl(object, pindex + i, 0);
798 if (blk != blk0 + i)
799 break;
800 }
801 *after = (i - 1);
802 }
1b9d3514 803#endif
cdd46d2e 804 crit_exit();
984263bc
MD
805 return (TRUE);
806}
807
808/*
809 * SWAP_PAGER_PAGE_UNSWAPPED() - remove swap backing store related to page
810 *
811 * This removes any associated swap backing store, whether valid or
812 * not, from the page.
813 *
814 * This routine is typically called when a page is made dirty, at
815 * which point any associated swap can be freed. MADV_FREE also
816 * calls us in a special-case situation
817 *
818 * NOTE!!! If the page is clean and the swap was valid, the caller
819 * should make the page dirty before calling this routine. This routine
820 * does NOT change the m->dirty status of the page. Also: MADV_FREE
821 * depends on it.
822 *
823 * This routine may not block
824 * This routine must be called at splvm()
825 */
826
827static void
57e43348 828swap_pager_unswapped(vm_page_t m)
984263bc
MD
829{
830 swp_pager_meta_ctl(m->object, m->pindex, SWM_FREE);
831}
832
833/*
834 * SWAP_PAGER_STRATEGY() - read, write, free blocks
835 *
836 * This implements the vm_pager_strategy() interface to swap and allows
837 * other parts of the system to directly access swap as backing store
838 * through vm_objects of type OBJT_SWAP. This is intended to be a
839 * cacheless interface ( i.e. caching occurs at higher levels ).
840 * Therefore we do not maintain any resident pages. All I/O goes
841 * directly to and from the swap device.
842 *
984263bc
MD
843 * We currently attempt to run I/O synchronously or asynchronously as
844 * the caller requests. This isn't perfect because we loose error
845 * sequencing when we run multiple ops in parallel to satisfy a request.
846 * But this is swap, so we let it all hang out.
847 */
848
849static void
81b5c339 850swap_pager_strategy(vm_object_t object, struct bio *bio)
984263bc 851{
81b5c339
MD
852 struct buf *bp = bio->bio_buf;
853 struct bio *nbio;
984263bc 854 vm_pindex_t start;
54078292 855 vm_pindex_t biox_blkno = 0;
984263bc 856 int count;
984263bc 857 char *data;
ae8e83e6
MD
858 struct bio *biox;
859 struct buf *bufx;
81b5c339
MD
860 struct bio_track *track;
861
862 /*
863 * tracking for swapdev vnode I/Os
864 */
10f3fee5 865 if (bp->b_cmd == BUF_CMD_READ)
81b5c339
MD
866 track = &swapdev_vp->v_track_read;
867 else
868 track = &swapdev_vp->v_track_write;
984263bc
MD
869
870 if (bp->b_bcount & PAGE_MASK) {
871 bp->b_error = EINVAL;
872 bp->b_flags |= B_ERROR | B_INVAL;
81b5c339 873 biodone(bio);
973c11b9
MD
874 kprintf("swap_pager_strategy: bp %p offset %lld size %d, "
875 "not page bounded\n",
876 bp, (long long)bio->bio_offset, (int)bp->b_bcount);
984263bc
MD
877 return;
878 }
879
880 /*
881 * Clear error indication, initialize page index, count, data pointer.
882 */
984263bc
MD
883 bp->b_error = 0;
884 bp->b_flags &= ~B_ERROR;
885 bp->b_resid = bp->b_bcount;
886
54078292 887 start = (vm_pindex_t)(bio->bio_offset >> PAGE_SHIFT);
984263bc
MD
888 count = howmany(bp->b_bcount, PAGE_SIZE);
889 data = bp->b_data;
890
984263bc 891 /*
10f3fee5 892 * Deal with BUF_CMD_FREEBLKS
984263bc 893 */
10f3fee5 894 if (bp->b_cmd == BUF_CMD_FREEBLKS) {
984263bc
MD
895 /*
896 * FREE PAGE(s) - destroy underlying swap that is no longer
897 * needed.
898 */
899 swp_pager_meta_free(object, start, count);
984263bc 900 bp->b_resid = 0;
81b5c339 901 biodone(bio);
984263bc
MD
902 return;
903 }
904
81b5c339
MD
905 /*
906 * We need to be able to create a new cluster of I/O's. We cannot
907 * use the caller fields of the passed bio so push a new one.
908 *
909 * Because nbio is just a placeholder for the cluster links,
910 * we can biodone() the original bio instead of nbio to make
911 * things a bit more efficient.
912 */
913 nbio = push_bio(bio);
54078292 914 nbio->bio_offset = bio->bio_offset;
81b5c339
MD
915 nbio->bio_caller_info1.cluster_head = NULL;
916 nbio->bio_caller_info2.cluster_tail = NULL;
917
ae8e83e6
MD
918 biox = NULL;
919 bufx = NULL;
920
984263bc
MD
921 /*
922 * Execute read or write
923 */
984263bc
MD
924 while (count > 0) {
925 daddr_t blk;
926
927 /*
928 * Obtain block. If block not found and writing, allocate a
929 * new block and build it into the object.
930 */
984263bc 931 blk = swp_pager_meta_ctl(object, start, 0);
10f3fee5 932 if ((blk == SWAPBLK_NONE) && bp->b_cmd != BUF_CMD_READ) {
984263bc
MD
933 blk = swp_pager_getswapspace(1);
934 if (blk == SWAPBLK_NONE) {
935 bp->b_error = ENOMEM;
936 bp->b_flags |= B_ERROR;
937 break;
938 }
939 swp_pager_meta_build(object, start, blk);
940 }
941
942 /*
943 * Do we have to flush our current collection? Yes if:
944 *
945 * - no swap block at this index
946 * - swap block is not contiguous
947 * - we cross a physical disk boundry in the
948 * stripe.
949 */
984263bc 950 if (
54078292
MD
951 biox && (biox_blkno + btoc(bufx->b_bcount) != blk ||
952 ((biox_blkno ^ blk) & dmmax_mask)
984263bc
MD
953 )
954 ) {
10f3fee5 955 if (bp->b_cmd == BUF_CMD_READ) {
12e4aaff 956 ++mycpu->gd_cnt.v_swapin;
81b5c339 957 mycpu->gd_cnt.v_swappgsin += btoc(bufx->b_bcount);
984263bc 958 } else {
12e4aaff 959 ++mycpu->gd_cnt.v_swapout;
81b5c339
MD
960 mycpu->gd_cnt.v_swappgsout += btoc(bufx->b_bcount);
961 bufx->b_dirtyend = bufx->b_bcount;
962 }
963
964 /*
ae8e83e6 965 * Finished with this buf.
81b5c339 966 */
ae8e83e6
MD
967 KKASSERT(bufx->b_bcount != 0);
968 if (bufx->b_cmd != BUF_CMD_READ)
969 bufx->b_dirtyend = bufx->b_bcount;
81b5c339
MD
970 biox = NULL;
971 bufx = NULL;
984263bc
MD
972 }
973
974 /*
81b5c339 975 * Add new swapblk to biox, instantiating biox if necessary.
984263bc
MD
976 * Zero-fill reads are able to take a shortcut.
977 */
984263bc
MD
978 if (blk == SWAPBLK_NONE) {
979 /*
980 * We can only get here if we are reading. Since
981 * we are at splvm() we can safely modify b_resid,
982 * even if chain ops are in progress.
983 */
984 bzero(data, PAGE_SIZE);
985 bp->b_resid -= PAGE_SIZE;
986 } else {
81b5c339
MD
987 if (biox == NULL) {
988 /* XXX chain count > 4, wait to <= 4 */
989
990 bufx = getpbuf(NULL);
991 biox = &bufx->b_bio1;
992 cluster_append(nbio, bufx);
ae8e83e6 993 bufx->b_flags |= (bufx->b_flags & B_ORDERED);
10f3fee5 994 bufx->b_cmd = bp->b_cmd;
81b5c339 995 biox->bio_done = swap_chain_iodone;
54078292 996 biox->bio_offset = (off_t)blk << PAGE_SHIFT;
81b5c339 997 biox->bio_caller_info1.cluster_parent = nbio;
54078292 998 biox_blkno = blk;
81b5c339
MD
999 bufx->b_bcount = 0;
1000 bufx->b_data = data;
984263bc 1001 }
81b5c339 1002 bufx->b_bcount += PAGE_SIZE;
984263bc
MD
1003 }
1004 --count;
1005 ++start;
1006 data += PAGE_SIZE;
1007 }
1008
1009 /*
1010 * Flush out last buffer
1011 */
81b5c339 1012 if (biox) {
10f3fee5 1013 if (bufx->b_cmd == BUF_CMD_READ) {
12e4aaff 1014 ++mycpu->gd_cnt.v_swapin;
81b5c339 1015 mycpu->gd_cnt.v_swappgsin += btoc(bufx->b_bcount);
984263bc 1016 } else {
12e4aaff 1017 ++mycpu->gd_cnt.v_swapout;
81b5c339
MD
1018 mycpu->gd_cnt.v_swappgsout += btoc(bufx->b_bcount);
1019 bufx->b_dirtyend = bufx->b_bcount;
1020 }
ae8e83e6
MD
1021 KKASSERT(bufx->b_bcount);
1022 if (bufx->b_cmd != BUF_CMD_READ)
1023 bufx->b_dirtyend = bufx->b_bcount;
81b5c339 1024 /* biox, bufx = NULL */
984263bc
MD
1025 }
1026
1027 /*
ae8e83e6
MD
1028 * Now initiate all the I/O. Be careful looping on our chain as
1029 * I/O's may complete while we are still initiating them.
984263bc 1030 */
ae8e83e6
MD
1031 nbio->bio_caller_info2.cluster_tail = NULL;
1032 bufx = nbio->bio_caller_info1.cluster_head;
1033
1034 while (bufx) {
1035 biox = &bufx->b_bio1;
1036 BUF_KERNPROC(bufx);
1037 bufx = bufx->b_cluster_next;
1038 vn_strategy(swapdev_vp, biox);
984263bc 1039 }
ae8e83e6
MD
1040
1041 /*
1042 * Completion of the cluster will also call biodone_chain(nbio).
1043 * We never call biodone(nbio) so we don't have to worry about
1044 * setting up a bio_done callback. It's handled in the sub-IO.
1045 */
1046 /**/
984263bc
MD
1047}
1048
81b5c339
MD
1049static void
1050swap_chain_iodone(struct bio *biox)
1051{
1052 struct buf **nextp;
1053 struct buf *bufx; /* chained sub-buffer */
1054 struct bio *nbio; /* parent nbio with chain glue */
1055 struct buf *bp; /* original bp associated with nbio */
ae8e83e6 1056 int chain_empty;
81b5c339
MD
1057
1058 bufx = biox->bio_buf;
1059 nbio = biox->bio_caller_info1.cluster_parent;
1060 bp = nbio->bio_buf;
1061
1062 /*
1063 * Update the original buffer
1064 */
1065 KKASSERT(bp != NULL);
1066 if (bufx->b_flags & B_ERROR) {
ae8e83e6 1067 atomic_set_int(&bufx->b_flags, B_ERROR);
81b5c339
MD
1068 bp->b_error = bufx->b_error;
1069 } else if (bufx->b_resid != 0) {
ae8e83e6 1070 atomic_set_int(&bufx->b_flags, B_ERROR);
81b5c339
MD
1071 bp->b_error = EINVAL;
1072 } else {
ae8e83e6 1073 atomic_subtract_int(&bp->b_resid, bufx->b_bcount);
81b5c339
MD
1074 }
1075
1076 /*
ae8e83e6 1077 * Remove us from the chain.
81b5c339 1078 */
ae8e83e6 1079 spin_lock_wr(&bp->b_lock.lk_spinlock);
81b5c339
MD
1080 nextp = &nbio->bio_caller_info1.cluster_head;
1081 while (*nextp != bufx) {
1082 KKASSERT(*nextp != NULL);
1083 nextp = &(*nextp)->b_cluster_next;
1084 }
1085 *nextp = bufx->b_cluster_next;
ae8e83e6
MD
1086 chain_empty = (nbio->bio_caller_info1.cluster_head == NULL);
1087 spin_unlock_wr(&bp->b_lock.lk_spinlock);
81b5c339
MD
1088
1089 /*
ae8e83e6
MD
1090 * Clean up bufx. If the chain is now empty we finish out
1091 * the parent. Note that we may be racing other completions
1092 * so we must use the chain_empty status from above.
81b5c339 1093 */
ae8e83e6 1094 if (chain_empty) {
81b5c339 1095 if (bp->b_resid != 0 && !(bp->b_flags & B_ERROR)) {
ae8e83e6 1096 atomic_set_int(&bp->b_flags, B_ERROR);
81b5c339
MD
1097 bp->b_error = EINVAL;
1098 }
ae8e83e6 1099 biodone_chain(nbio);
81b5c339 1100 }
81b5c339
MD
1101 relpbuf(bufx, NULL);
1102}
1103
984263bc
MD
1104/*
1105 * SWAP_PAGER_GETPAGES() - bring pages in from swap
1106 *
1107 * Attempt to retrieve (m, count) pages from backing store, but make
1108 * sure we retrieve at least m[reqpage]. We try to load in as large
1109 * a chunk surrounding m[reqpage] as is contiguous in swap and which
1110 * belongs to the same object.
1111 *
1112 * The code is designed for asynchronous operation and
1113 * immediate-notification of 'reqpage' but tends not to be
1114 * used that way. Please do not optimize-out this algorithmic
1115 * feature, I intend to improve on it in the future.
1116 *
1117 * The parent has a single vm_object_pip_add() reference prior to
1118 * calling us and we should return with the same.
1119 *
1120 * The parent has BUSY'd the pages. We should return with 'm'
1121 * left busy, but the others adjusted.
1122 */
1123
1124static int
1b9d3514 1125swap_pager_getpage(vm_object_t object, vm_page_t *mpp, int seqaccess)
984263bc
MD
1126{
1127 struct buf *bp;
81b5c339 1128 struct bio *bio;
984263bc 1129 vm_page_t mreq;
984263bc
MD
1130 int i;
1131 int j;
1b9d3514 1132 int reqpage;
984263bc
MD
1133 daddr_t blk;
1134 vm_offset_t kva;
1135 vm_pindex_t lastpindex;
1136
1b9d3514 1137 mreq = *mpp;
984263bc
MD
1138
1139 if (mreq->object != object) {
1140 panic("swap_pager_getpages: object mismatch %p/%p",
1141 object,
1142 mreq->object
1143 );
1144 }
17cde63e 1145
984263bc
MD
1146 /*
1147 * Calculate range to retrieve. The pages have already been assigned
1148 * their swapblks. We require a *contiguous* range that falls entirely
1149 * within a single device stripe. If we do not supply it, bad things
1150 * happen. Note that blk, iblk & jblk can be SWAPBLK_NONE, but the
1151 * loops are set up such that the case(s) are handled implicitly.
1152 *
1153 * The swp_*() calls must be made at splvm(). vm_page_free() does
1154 * not need to be, but it will go a little faster if it is.
1155 */
cdd46d2e 1156 crit_enter();
984263bc
MD
1157 blk = swp_pager_meta_ctl(mreq->object, mreq->pindex, 0);
1158
1b9d3514 1159#if 0
984263bc
MD
1160 for (i = reqpage - 1; i >= 0; --i) {
1161 daddr_t iblk;
1162
1163 iblk = swp_pager_meta_ctl(m[i]->object, m[i]->pindex, 0);
1164 if (blk != iblk + (reqpage - i))
1165 break;
1166 if ((blk ^ iblk) & dmmax_mask)
1167 break;
1168 }
1169 ++i;
1170
1171 for (j = reqpage + 1; j < count; ++j) {
1172 daddr_t jblk;
1173
1174 jblk = swp_pager_meta_ctl(m[j]->object, m[j]->pindex, 0);
1175 if (blk != jblk - (j - reqpage))
1176 break;
1177 if ((blk ^ jblk) & dmmax_mask)
1178 break;
1179 }
1180
1181 /*
1182 * free pages outside our collection range. Note: we never free
1183 * mreq, it must remain busy throughout.
1184 */
1185
1186 {
1187 int k;
1188
1189 for (k = 0; k < i; ++k)
1190 vm_page_free(m[k]);
1191 for (k = j; k < count; ++k)
1192 vm_page_free(m[k]);
1193 }
1b9d3514 1194#endif
cdd46d2e 1195 crit_exit();
984263bc
MD
1196
1197
1198 /*
1199 * Return VM_PAGER_FAIL if we have nothing to do. Return mreq
1200 * still busy, but the others unbusied.
1201 */
1202
1203 if (blk == SWAPBLK_NONE)
1204 return(VM_PAGER_FAIL);
1205
1206 /*
1207 * Get a swap buffer header to perform the IO
1208 */
1209
1210 bp = getpbuf(&nsw_rcount);
81b5c339 1211 bio = &bp->b_bio1;
984263bc
MD
1212 kva = (vm_offset_t) bp->b_data;
1213
1214 /*
1215 * map our page(s) into kva for input
984263bc 1216 */
1b9d3514
MD
1217 i = 0;
1218 j = 1;
1219 reqpage = 0;
984263bc 1220
1b9d3514 1221 pmap_qenter(kva, mpp + i, j - i);
984263bc 1222
984263bc 1223 bp->b_data = (caddr_t) kva;
984263bc 1224 bp->b_bcount = PAGE_SIZE * (j - i);
81b5c339 1225 bio->bio_done = swp_pager_async_iodone;
54078292 1226 bio->bio_offset = (off_t)(blk - (reqpage - i)) << PAGE_SHIFT;
973c11b9 1227 bio->bio_driver_info = (void *)(intptr_t)(reqpage - i);
8aa92e4b 1228 bio->bio_caller_info1.index = SWBIO_READ;
984263bc
MD
1229
1230 {
1231 int k;
1232
1233 for (k = i; k < j; ++k) {
1b9d3514
MD
1234 bp->b_xio.xio_pages[k - i] = mpp[k];
1235 vm_page_flag_set(mpp[k], PG_SWAPINPROG);
984263bc
MD
1236 }
1237 }
54f51aeb 1238 bp->b_xio.xio_npages = j - i;
984263bc 1239
12e4aaff 1240 mycpu->gd_cnt.v_swapin++;
54f51aeb 1241 mycpu->gd_cnt.v_swappgsin += bp->b_xio.xio_npages;
984263bc
MD
1242
1243 /*
1244 * We still hold the lock on mreq, and our automatic completion routine
1245 * does not remove it.
1246 */
1247
54f51aeb 1248 vm_object_pip_add(mreq->object, bp->b_xio.xio_npages);
1b9d3514 1249 lastpindex = mpp[j-1]->pindex;
984263bc
MD
1250
1251 /*
1252 * perform the I/O. NOTE!!! bp cannot be considered valid after
1253 * this point because we automatically release it on completion.
1254 * Instead, we look at the one page we are interested in which we
1255 * still hold a lock on even through the I/O completion.
1256 *
1257 * The other pages in our m[] array are also released on completion,
1258 * so we cannot assume they are valid anymore either.
984263bc
MD
1259 */
1260
10f3fee5 1261 bp->b_cmd = BUF_CMD_READ;
984263bc 1262 BUF_KERNPROC(bp);
81b5c339 1263 vn_strategy(swapdev_vp, bio);
984263bc
MD
1264
1265 /*
1266 * wait for the page we want to complete. PG_SWAPINPROG is always
1267 * cleared on completion. If an I/O error occurs, SWAPBLK_NONE
1268 * is set in the meta-data.
1269 */
1270
cdd46d2e 1271 crit_enter();
984263bc
MD
1272
1273 while ((mreq->flags & PG_SWAPINPROG) != 0) {
1274 vm_page_flag_set(mreq, PG_WANTED | PG_REFERENCED);
12e4aaff 1275 mycpu->gd_cnt.v_intrans++;
377d4740 1276 if (tsleep(mreq, 0, "swread", hz*20)) {
086c1d7e 1277 kprintf(
81b5c339 1278 "swap_pager: indefinite wait buffer: "
973c11b9
MD
1279 " offset: %lld, size: %ld\n",
1280 (long long)bio->bio_offset,
1281 (long)bp->b_bcount
984263bc
MD
1282 );
1283 }
1284 }
1285
cdd46d2e 1286 crit_exit();
984263bc
MD
1287
1288 /*
1289 * mreq is left bussied after completion, but all the other pages
1290 * are freed. If we had an unrecoverable read error the page will
1291 * not be valid.
1292 */
1293
1294 if (mreq->valid != VM_PAGE_BITS_ALL) {
1295 return(VM_PAGER_ERROR);
1296 } else {
1297 return(VM_PAGER_OK);
1298 }
1299
1300 /*
1301 * A final note: in a low swap situation, we cannot deallocate swap
1302 * and mark a page dirty here because the caller is likely to mark
1303 * the page clean when we return, causing the page to possibly revert
1304 * to all-zero's later.
1305 */
1306}
1307
1308/*
1309 * swap_pager_putpages:
1310 *
1311 * Assign swap (if necessary) and initiate I/O on the specified pages.
1312 *
1313 * We support both OBJT_DEFAULT and OBJT_SWAP objects. DEFAULT objects
1314 * are automatically converted to SWAP objects.
1315 *
81b5c339 1316 * In a low memory situation we may block in vn_strategy(), but the new
984263bc
MD
1317 * vm_page reservation system coupled with properly written VFS devices
1318 * should ensure that no low-memory deadlock occurs. This is an area
1319 * which needs work.
1320 *
1321 * The parent has N vm_object_pip_add() references prior to
1322 * calling us and will remove references for rtvals[] that are
1323 * not set to VM_PAGER_PEND. We need to remove the rest on I/O
1324 * completion.
1325 *
1326 * The parent has soft-busy'd the pages it passes us and will unbusy
1327 * those whos rtvals[] entry is not set to VM_PAGER_PEND on return.
1328 * We need to unbusy the rest on I/O completion.
1329 */
984263bc 1330void
17cde63e
MD
1331swap_pager_putpages(vm_object_t object, vm_page_t *m, int count,
1332 boolean_t sync, int *rtvals)
984263bc
MD
1333{
1334 int i;
1335 int n = 0;
1336
1337 if (count && m[0]->object != object) {
1338 panic("swap_pager_getpages: object mismatch %p/%p",
1339 object,
1340 m[0]->object
1341 );
1342 }
17cde63e 1343
984263bc
MD
1344 /*
1345 * Step 1
1346 *
1347 * Turn object into OBJT_SWAP
1348 * check for bogus sysops
1349 * force sync if not pageout process
1350 */
1351
1352 if (object->type != OBJT_SWAP)
1353 swp_pager_meta_build(object, 0, SWAPBLK_NONE);
1354
bc6dffab 1355 if (curthread != pagethread)
984263bc
MD
1356 sync = TRUE;
1357
1358 /*
1359 * Step 2
1360 *
1361 * Update nsw parameters from swap_async_max sysctl values.
1362 * Do not let the sysop crash the machine with bogus numbers.
1363 */
1364
1365 if (swap_async_max != nsw_wcount_async_max) {
1366 int n;
984263bc
MD
1367
1368 /*
1369 * limit range
1370 */
1371 if ((n = swap_async_max) > nswbuf / 2)
1372 n = nswbuf / 2;
1373 if (n < 1)
1374 n = 1;
1375 swap_async_max = n;
1376
1377 /*
1378 * Adjust difference ( if possible ). If the current async
1379 * count is too low, we may not be able to make the adjustment
1380 * at this time.
1381 */
cdd46d2e 1382 crit_enter();
984263bc
MD
1383 n -= nsw_wcount_async_max;
1384 if (nsw_wcount_async + n >= 0) {
1385 nsw_wcount_async += n;
1386 nsw_wcount_async_max += n;
1387 wakeup(&nsw_wcount_async);
1388 }
cdd46d2e 1389 crit_exit();
984263bc
MD
1390 }
1391
1392 /*
1393 * Step 3
1394 *
1395 * Assign swap blocks and issue I/O. We reallocate swap on the fly.
1396 * The page is left dirty until the pageout operation completes
1397 * successfully.
1398 */
1399
1400 for (i = 0; i < count; i += n) {
984263bc 1401 struct buf *bp;
81b5c339 1402 struct bio *bio;
984263bc 1403 daddr_t blk;
81b5c339 1404 int j;
984263bc
MD
1405
1406 /*
1407 * Maximum I/O size is limited by a number of factors.
1408 */
1409
1410 n = min(BLIST_MAX_ALLOC, count - i);
1411 n = min(n, nsw_cluster_max);
1412
cdd46d2e 1413 crit_enter();
984263bc
MD
1414
1415 /*
1416 * Get biggest block of swap we can. If we fail, fall
1417 * back and try to allocate a smaller block. Don't go
1418 * overboard trying to allocate space if it would overly
1419 * fragment swap.
1420 */
1421 while (
1422 (blk = swp_pager_getswapspace(n)) == SWAPBLK_NONE &&
1423 n > 4
1424 ) {
1425 n >>= 1;
1426 }
1427 if (blk == SWAPBLK_NONE) {
1428 for (j = 0; j < n; ++j)
1429 rtvals[i+j] = VM_PAGER_FAIL;
cdd46d2e 1430 crit_exit();
984263bc
MD
1431 continue;
1432 }
1433
1434 /*
1435 * The I/O we are constructing cannot cross a physical
1436 * disk boundry in the swap stripe. Note: we are still
1437 * at splvm().
1438 */
1439 if ((blk ^ (blk + n)) & dmmax_mask) {
1440 j = ((blk + dmmax) & dmmax_mask) - blk;
1441 swp_pager_freeswapspace(blk + j, n - j);
1442 n = j;
1443 }
1444
1445 /*
1446 * All I/O parameters have been satisfied, build the I/O
1447 * request and assign the swap space.
984263bc
MD
1448 */
1449
10f3fee5 1450 if (sync == TRUE)
984263bc 1451 bp = getpbuf(&nsw_wcount_sync);
10f3fee5 1452 else
984263bc 1453 bp = getpbuf(&nsw_wcount_async);
81b5c339 1454 bio = &bp->b_bio1;
984263bc
MD
1455
1456 pmap_qenter((vm_offset_t)bp->b_data, &m[i], n);
1457
984263bc 1458 bp->b_bcount = PAGE_SIZE * n;
54078292 1459 bio->bio_offset = (off_t)blk << PAGE_SHIFT;
984263bc 1460
984263bc
MD
1461 for (j = 0; j < n; ++j) {
1462 vm_page_t mreq = m[i+j];
1463
1464 swp_pager_meta_build(
1465 mreq->object,
1466 mreq->pindex,
1467 blk + j
1468 );
1469 vm_page_dirty(mreq);
1470 rtvals[i+j] = VM_PAGER_OK;
1471
1472 vm_page_flag_set(mreq, PG_SWAPINPROG);
54f51aeb 1473 bp->b_xio.xio_pages[j] = mreq;
984263bc 1474 }
54f51aeb 1475 bp->b_xio.xio_npages = n;
984263bc 1476
12e4aaff 1477 mycpu->gd_cnt.v_swapout++;
54f51aeb 1478 mycpu->gd_cnt.v_swappgsout += bp->b_xio.xio_npages;
984263bc 1479
cdd46d2e 1480 crit_exit();
984263bc 1481
10f3fee5
MD
1482 bp->b_dirtyoff = 0; /* req'd for NFS */
1483 bp->b_dirtyend = bp->b_bcount; /* req'd for NFS */
1484 bp->b_cmd = BUF_CMD_WRITE;
8aa92e4b 1485 bio->bio_caller_info1.index = SWBIO_WRITE;
10f3fee5 1486
984263bc
MD
1487 /*
1488 * asynchronous
984263bc 1489 */
984263bc 1490 if (sync == FALSE) {
81b5c339 1491 bio->bio_done = swp_pager_async_iodone;
984263bc 1492 BUF_KERNPROC(bp);
81b5c339 1493 vn_strategy(swapdev_vp, bio);
984263bc
MD
1494
1495 for (j = 0; j < n; ++j)
1496 rtvals[i+j] = VM_PAGER_PEND;
1497 continue;
1498 }
1499
1500 /*
ae8e83e6
MD
1501 * Issue synchrnously.
1502 *
984263bc
MD
1503 * Wait for the sync I/O to complete, then update rtvals.
1504 * We just set the rtvals[] to VM_PAGER_PEND so we can call
1505 * our async completion routine at the end, thus avoiding a
1506 * double-free.
1507 */
8aa92e4b 1508 bio->bio_caller_info1.index |= SWBIO_SYNC;
ae8e83e6
MD
1509 bio->bio_done = biodone_sync;
1510 bio->bio_flags |= BIO_SYNC;
1511 vn_strategy(swapdev_vp, bio);
1512 biowait(bio, "swwrt");
984263bc
MD
1513
1514 for (j = 0; j < n; ++j)
1515 rtvals[i+j] = VM_PAGER_PEND;
1516
1517 /*
1518 * Now that we are through with the bp, we can call the
1519 * normal async completion, which frees everything up.
1520 */
81b5c339 1521 swp_pager_async_iodone(bio);
984263bc
MD
1522 }
1523}
1524
c84c24da
MD
1525void
1526swap_pager_newswap(void)
1527{
1528 swp_sizecheck();
1529}
1530
984263bc
MD
1531/*
1532 * swp_pager_async_iodone:
1533 *
1534 * Completion routine for asynchronous reads and writes from/to swap.
1535 * Also called manually by synchronous code to finish up a bp.
1536 *
1537 * For READ operations, the pages are PG_BUSY'd. For WRITE operations,
1538 * the pages are vm_page_t->busy'd. For READ operations, we PG_BUSY
1539 * unbusy all pages except the 'main' request page. For WRITE
1540 * operations, we vm_page_t->busy'd unbusy all pages ( we can do this
1541 * because we marked them all VM_PAGER_PEND on return from putpages ).
1542 *
1543 * This routine may not block.
984263bc 1544 */
984263bc 1545static void
81b5c339 1546swp_pager_async_iodone(struct bio *bio)
984263bc 1547{
81b5c339 1548 struct buf *bp = bio->bio_buf;
984263bc 1549 vm_object_t object = NULL;
81b5c339 1550 int i;
10f3fee5 1551 int *nswptr;
984263bc
MD
1552
1553 /*
1554 * report error
1555 */
984263bc 1556 if (bp->b_flags & B_ERROR) {
086c1d7e 1557 kprintf(
54078292 1558 "swap_pager: I/O error - %s failed; offset %lld,"
984263bc 1559 "size %ld, error %d\n",
8aa92e4b
MD
1560 ((bio->bio_caller_info1.index & SWBIO_READ) ?
1561 "pagein" : "pageout"),
973c11b9 1562 (long long)bio->bio_offset,
984263bc
MD
1563 (long)bp->b_bcount,
1564 bp->b_error
1565 );
1566 }
1567
1568 /*
1569 * set object, raise to splvm().
1570 */
54f51aeb
HP
1571 if (bp->b_xio.xio_npages)
1572 object = bp->b_xio.xio_pages[0]->object;
cdd46d2e 1573 crit_enter();
984263bc
MD
1574
1575 /*
1576 * remove the mapping for kernel virtual
1577 */
54f51aeb 1578 pmap_qremove((vm_offset_t)bp->b_data, bp->b_xio.xio_npages);
984263bc
MD
1579
1580 /*
1581 * cleanup pages. If an error occurs writing to swap, we are in
1582 * very serious trouble. If it happens to be a disk error, though,
1583 * we may be able to recover by reassigning the swap later on. So
1584 * in this case we remove the m->swapblk assignment for the page
1585 * but do not free it in the rlist. The errornous block(s) are thus
1586 * never reallocated as swap. Redirty the page and continue.
1587 */
54f51aeb
HP
1588 for (i = 0; i < bp->b_xio.xio_npages; ++i) {
1589 vm_page_t m = bp->b_xio.xio_pages[i];
984263bc
MD
1590
1591 vm_page_flag_clear(m, PG_SWAPINPROG);
1592
1593 if (bp->b_flags & B_ERROR) {
1594 /*
1595 * If an error occurs I'd love to throw the swapblk
1596 * away without freeing it back to swapspace, so it
1597 * can never be used again. But I can't from an
1598 * interrupt.
1599 */
1600
8aa92e4b 1601 if (bio->bio_caller_info1.index & SWBIO_READ) {
984263bc
MD
1602 /*
1603 * When reading, reqpage needs to stay
1604 * locked for the parent, but all other
1605 * pages can be freed. We still want to
1606 * wakeup the parent waiting on the page,
1607 * though. ( also: pg_reqpage can be -1 and
1608 * not match anything ).
1609 *
1610 * We have to wake specifically requested pages
1611 * up too because we cleared PG_SWAPINPROG and
1612 * someone may be waiting for that.
1613 *
1614 * NOTE: for reads, m->dirty will probably
1615 * be overridden by the original caller of
1616 * getpages so don't play cute tricks here.
1617 *
93afe6be
MD
1618 * NOTE: We can't actually free the page from
1619 * here, because this is an interrupt. It
1620 * is not legal to mess with object->memq
1621 * from an interrupt. Deactivate the page
1622 * instead.
984263bc
MD
1623 */
1624
1625 m->valid = 0;
1626 vm_page_flag_clear(m, PG_ZERO);
1627
81b5c339
MD
1628 /*
1629 * bio_driver_info holds the requested page
1630 * index.
1631 */
973c11b9 1632 if (i != (int)(intptr_t)bio->bio_driver_info) {
93afe6be
MD
1633 vm_page_deactivate(m);
1634 vm_page_wakeup(m);
1635 } else {
984263bc 1636 vm_page_flash(m);
93afe6be 1637 }
984263bc
MD
1638 /*
1639 * If i == bp->b_pager.pg_reqpage, do not wake
1640 * the page up. The caller needs to.
1641 */
1642 } else {
1643 /*
1644 * If a write error occurs, reactivate page
1645 * so it doesn't clog the inactive list,
1646 * then finish the I/O.
1647 */
1648 vm_page_dirty(m);
1649 vm_page_activate(m);
1650 vm_page_io_finish(m);
1651 }
8aa92e4b 1652 } else if (bio->bio_caller_info1.index & SWBIO_READ) {
984263bc 1653 /*
984263bc
MD
1654 * NOTE: for reads, m->dirty will probably be
1655 * overridden by the original caller of getpages so
1656 * we cannot set them in order to free the underlying
1657 * swap in a low-swap situation. I don't think we'd
1658 * want to do that anyway, but it was an optimization
1659 * that existed in the old swapper for a time before
1660 * it got ripped out due to precisely this problem.
1661 *
1662 * clear PG_ZERO in page.
1663 *
1664 * If not the requested page then deactivate it.
1665 *
1666 * Note that the requested page, reqpage, is left
1667 * busied, but we still have to wake it up. The
1668 * other pages are released (unbusied) by
1669 * vm_page_wakeup(). We do not set reqpage's
1670 * valid bits here, it is up to the caller.
1671 */
1672
4530a3aa
MD
1673 /*
1674 * NOTE: can't call pmap_clear_modify(m) from an
1675 * interrupt thread, the pmap code may have to map
1676 * non-kernel pmaps and currently asserts the case.
1677 */
1678 /*pmap_clear_modify(m);*/
984263bc
MD
1679 m->valid = VM_PAGE_BITS_ALL;
1680 vm_page_undirty(m);
1681 vm_page_flag_clear(m, PG_ZERO);
1682
1683 /*
1684 * We have to wake specifically requested pages
1685 * up too because we cleared PG_SWAPINPROG and
1686 * could be waiting for it in getpages. However,
1687 * be sure to not unbusy getpages specifically
1688 * requested page - getpages expects it to be
1689 * left busy.
81b5c339
MD
1690 *
1691 * bio_driver_info holds the requested page
984263bc 1692 */
973c11b9 1693 if (i != (int)(intptr_t)bio->bio_driver_info) {
984263bc
MD
1694 vm_page_deactivate(m);
1695 vm_page_wakeup(m);
1696 } else {
1697 vm_page_flash(m);
1698 }
1699 } else {
1700 /*
93afe6be
MD
1701 * Mark the page clean but do not mess with the
1702 * pmap-layer's modified state. That state should
1703 * also be clear since the caller protected the
1704 * page VM_PROT_READ, but allow the case.
1705 *
1706 * We are in an interrupt, avoid pmap operations.
1707 *
1708 * If we have a severe page deficit, deactivate the
1709 * page. Do not try to cache it (which would also
1710 * involve a pmap op), because the page might still
1711 * be read-heavy.
984263bc 1712 */
984263bc
MD
1713 vm_page_undirty(m);
1714 vm_page_io_finish(m);
93afe6be
MD
1715 if (vm_page_count_severe())
1716 vm_page_deactivate(m);
1717#if 0
984263bc
MD
1718 if (!vm_page_count_severe() || !vm_page_try_to_cache(m))
1719 vm_page_protect(m, VM_PROT_READ);
93afe6be 1720#endif
984263bc
MD
1721 }
1722 }
1723
1724 /*
1725 * adjust pip. NOTE: the original parent may still have its own
1726 * pip refs on the object.
1727 */
1728
1729 if (object)
54f51aeb 1730 vm_object_pip_wakeupn(object, bp->b_xio.xio_npages);
984263bc
MD
1731
1732 /*
8aa92e4b
MD
1733 * Release the physical I/O buffer.
1734 *
1735 * NOTE: Due to synchronous operations in the write case b_cmd may
1736 * already be set to BUF_CMD_DONE and BIO_SYNC may have already
1737 * been cleared.
984263bc 1738 */
8aa92e4b 1739 if (bio->bio_caller_info1.index & SWBIO_READ)
10f3fee5 1740 nswptr = &nsw_rcount;
8aa92e4b 1741 else if (bio->bio_caller_info1.index & SWBIO_SYNC)
10f3fee5 1742 nswptr = &nsw_wcount_sync;
ae8e83e6
MD
1743 else
1744 nswptr = &nsw_wcount_async;
10f3fee5
MD
1745 bp->b_cmd = BUF_CMD_DONE;
1746 relpbuf(bp, nswptr);
cdd46d2e 1747 crit_exit();
984263bc
MD
1748}
1749
1750/************************************************************************
1751 * SWAP META DATA *
1752 ************************************************************************
1753 *
1754 * These routines manipulate the swap metadata stored in the
1755 * OBJT_SWAP object. All swp_*() routines must be called at
1756 * splvm() because swap can be freed up by the low level vm_page
1757 * code which might be called from interrupts beyond what splbio() covers.
1758 *
1759 * Swap metadata is implemented with a global hash and not directly
1760 * linked into the object. Instead the object simply contains
1761 * appropriate tracking counters.
1762 */
1763
1764/*
1765 * SWP_PAGER_HASH() - hash swap meta data
1766 *
1767 * This is an inline helper function which hashes the swapblk given
1768 * the object and page index. It returns a pointer to a pointer
1769 * to the object, or a pointer to a NULL pointer if it could not
1770 * find a swapblk.
1771 *
1772 * This routine must be called at splvm().
1773 */
1774
1775static __inline struct swblock **
1776swp_pager_hash(vm_object_t object, vm_pindex_t index)
1777{
1778 struct swblock **pswap;
1779 struct swblock *swap;
1780
1781 index &= ~SWAP_META_MASK;
1782 pswap = &swhash[(index ^ (int)(intptr_t)object) & swhash_mask];
1783
1784 while ((swap = *pswap) != NULL) {
1785 if (swap->swb_object == object &&
1786 swap->swb_index == index
1787 ) {
1788 break;
1789 }
1790 pswap = &swap->swb_hnext;
1791 }
1792 return(pswap);
1793}
1794
1795/*
1796 * SWP_PAGER_META_BUILD() - add swap block to swap meta data for object
1797 *
1798 * We first convert the object to a swap object if it is a default
1799 * object.
1800 *
1801 * The specified swapblk is added to the object's swap metadata. If
1802 * the swapblk is not valid, it is freed instead. Any previously
1803 * assigned swapblk is freed.
1804 *
1805 * This routine must be called at splvm(), except when used to convert
1806 * an OBJT_DEFAULT object into an OBJT_SWAP object.
1807
1808 */
1809
1810static void
1811swp_pager_meta_build(
1812 vm_object_t object,
1813 vm_pindex_t index,
1814 daddr_t swapblk
1815) {
1816 struct swblock *swap;
1817 struct swblock **pswap;
1818
1819 /*
1820 * Convert default object to swap object if necessary
1821 */
1822
1823 if (object->type != OBJT_SWAP) {
1824 object->type = OBJT_SWAP;
1825 object->un_pager.swp.swp_bcount = 0;
1826
1827 if (object->handle != NULL) {
1828 TAILQ_INSERT_TAIL(
1829 NOBJLIST(object->handle),
1830 object,
1831 pager_object_list
1832 );
1833 } else {
1834 TAILQ_INSERT_TAIL(
1835 &swap_pager_un_object_list,
1836 object,
1837 pager_object_list
1838 );
1839 }
1840 }
1841
1842 /*
1843 * Locate hash entry. If not found create, but if we aren't adding
1844 * anything just return. If we run out of space in the map we wait
1845 * and, since the hash table may have changed, retry.
1846 */
1847
1848retry:
1849 pswap = swp_pager_hash(object, index);
1850
1851 if ((swap = *pswap) == NULL) {
1852 int i;
1853
1854 if (swapblk == SWAPBLK_NONE)
1855 return;
1856
1857 swap = *pswap = zalloc(swap_zone);
1858 if (swap == NULL) {
4ecf7cc9 1859 vm_wait(0);
984263bc
MD
1860 goto retry;
1861 }
1862 swap->swb_hnext = NULL;
1863 swap->swb_object = object;
1864 swap->swb_index = index & ~SWAP_META_MASK;
1865 swap->swb_count = 0;
1866
1867 ++object->un_pager.swp.swp_bcount;
1868
1869 for (i = 0; i < SWAP_META_PAGES; ++i)
1870 swap->swb_pages[i] = SWAPBLK_NONE;
1871 }
1872
1873 /*
1874 * Delete prior contents of metadata
1875 */
1876
1877 index &= SWAP_META_MASK;
1878
1879 if (swap->swb_pages[index] != SWAPBLK_NONE) {
1880 swp_pager_freeswapspace(swap->swb_pages[index], 1);
1881 --swap->swb_count;
1882 }
1883
1884 /*
1885 * Enter block into metadata
1886 */
1887
1888 swap->swb_pages[index] = swapblk;
1889 if (swapblk != SWAPBLK_NONE)
1890 ++swap->swb_count;
1891}
1892
1893/*
1894 * SWP_PAGER_META_FREE() - free a range of blocks in the object's swap metadata
1895 *
1896 * The requested range of blocks is freed, with any associated swap
1897 * returned to the swap bitmap.
1898 *
1899 * This routine will free swap metadata structures as they are cleaned
1900 * out. This routine does *NOT* operate on swap metadata associated
1901 * with resident pages.
1902 *
1903 * This routine must be called at splvm()
1904 */
1905
1906static void
1907swp_pager_meta_free(vm_object_t object, vm_pindex_t index, daddr_t count)
1908{
1909 if (object->type != OBJT_SWAP)
1910 return;
1911
1912 while (count > 0) {
1913 struct swblock **pswap;
1914 struct swblock *swap;
1915
1916 pswap = swp_pager_hash(object, index);
1917
1918 if ((swap = *pswap) != NULL) {
1919 daddr_t v = swap->swb_pages[index & SWAP_META_MASK];
1920
1921 if (v != SWAPBLK_NONE) {
1922 swp_pager_freeswapspace(v, 1);
1923 swap->swb_pages[index & SWAP_META_MASK] =
1924 SWAPBLK_NONE;
1925 if (--swap->swb_count == 0) {
1926 *pswap = swap->swb_hnext;
1927 zfree(swap_zone, swap);
1928 --object->un_pager.swp.swp_bcount;
1929 }
1930 }
1931 --count;
1932 ++index;
1933 } else {
1934 int n = SWAP_META_PAGES - (index & SWAP_META_MASK);
1935 count -= n;
1936 index += n;
1937 }
1938 }
1939}
1940
1941/*
1942 * SWP_PAGER_META_FREE_ALL() - destroy all swap metadata associated with object
1943 *
1944 * This routine locates and destroys all swap metadata associated with
1945 * an object.
1946 *
1947 * This routine must be called at splvm()
1948 */
1949
1950static void
1951swp_pager_meta_free_all(vm_object_t object)
1952{
1953 daddr_t index = 0;
1954
1955 if (object->type != OBJT_SWAP)
1956 return;
1957
1958 while (object->un_pager.swp.swp_bcount) {
1959 struct swblock **pswap;
1960 struct swblock *swap;
1961
1962 pswap = swp_pager_hash(object, index);
1963 if ((swap = *pswap) != NULL) {
1964 int i;
1965
1966 for (i = 0; i < SWAP_META_PAGES; ++i) {
1967 daddr_t v = swap->swb_pages[i];
1968 if (v != SWAPBLK_NONE) {
1969 --swap->swb_count;
1970 swp_pager_freeswapspace(v, 1);
1971 }
1972 }
1973 if (swap->swb_count != 0)
1974 panic("swap_pager_meta_free_all: swb_count != 0");
1975 *pswap = swap->swb_hnext;
1976 zfree(swap_zone, swap);
1977 --object->un_pager.swp.swp_bcount;
1978 }
1979 index += SWAP_META_PAGES;
1980 if (index > 0x20000000)
1981 panic("swp_pager_meta_free_all: failed to locate all swap meta blocks");
1982 }
1983}
1984
1985/*
1986 * SWP_PAGER_METACTL() - misc control of swap and vm_page_t meta data.
1987 *
1988 * This routine is capable of looking up, popping, or freeing
1989 * swapblk assignments in the swap meta data or in the vm_page_t.
1990 * The routine typically returns the swapblk being looked-up, or popped,
1991 * or SWAPBLK_NONE if the block was freed, or SWAPBLK_NONE if the block
1992 * was invalid. This routine will automatically free any invalid
1993 * meta-data swapblks.
1994 *
1995 * It is not possible to store invalid swapblks in the swap meta data
1996 * (other then a literal 'SWAPBLK_NONE'), so we don't bother checking.
1997 *
1998 * When acting on a busy resident page and paging is in progress, we
1999 * have to wait until paging is complete but otherwise can act on the
2000 * busy page.
2001 *
2002 * This routine must be called at splvm().
2003 *
2004 * SWM_FREE remove and free swap block from metadata
2005 * SWM_POP remove from meta data but do not free.. pop it out
2006 */
2007
2008static daddr_t
2009swp_pager_meta_ctl(
2010 vm_object_t object,
2011 vm_pindex_t index,
2012 int flags
2013) {
2014 struct swblock **pswap;
2015 struct swblock *swap;
2016 daddr_t r1;
2017
2018 /*
2019 * The meta data only exists of the object is OBJT_SWAP
2020 * and even then might not be allocated yet.
2021 */
2022
2023 if (object->type != OBJT_SWAP)
2024 return(SWAPBLK_NONE);
2025
2026 r1 = SWAPBLK_NONE;
2027 pswap = swp_pager_hash(object, index);
2028
2029 if ((swap = *pswap) != NULL) {
2030 index &= SWAP_META_MASK;
2031 r1 = swap->swb_pages[index];
2032
2033 if (r1 != SWAPBLK_NONE) {
2034 if (flags & SWM_FREE) {
2035 swp_pager_freeswapspace(r1, 1);
2036 r1 = SWAPBLK_NONE;
2037 }
2038 if (flags & (SWM_FREE|SWM_POP)) {
2039 swap->swb_pages[index] = SWAPBLK_NONE;
2040 if (--swap->swb_count == 0) {
2041 *pswap = swap->swb_hnext;
2042 zfree(swap_zone, swap);
2043 --object->un_pager.swp.swp_bcount;
2044 }
2045 }
2046 }
2047 }
2048 return(r1);
2049}