kernel - SWAP CACHE part 8/many - Add the swap cache read intercept, rate ctl
[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>
5d5c5831 131#include <vm/vnode_pager.h>
984263bc 132
3020e3be 133#include <sys/buf2.h>
12e4aaff 134#include <vm/vm_page2.h>
3020e3be 135
984263bc
MD
136#define SWM_FREE 0x02 /* free, period */
137#define SWM_POP 0x04 /* pop out */
138
8aa92e4b
MD
139#define SWBIO_READ 0x01
140#define SWBIO_WRITE 0x02
141#define SWBIO_SYNC 0x04
142
8d292090
MD
143struct swfreeinfo {
144 vm_object_t object;
145 vm_pindex_t basei;
146 vm_pindex_t begi;
147 vm_pindex_t endi; /* inclusive */
148};
149
984263bc
MD
150/*
151 * vm_swap_size is in page-sized chunks now. It was DEV_BSIZE'd chunks
152 * in the old system.
153 */
154
984263bc 155int swap_pager_full; /* swap space exhaustion (task killing) */
096e95c0
MD
156int vm_swap_cache_use;
157int vm_swap_anon_use;
158
984263bc
MD
159static int swap_pager_almost_full; /* swap space exhaustion (w/ hysteresis)*/
160static int nsw_rcount; /* free read buffers */
161static int nsw_wcount_sync; /* limit write buffers / synchronous */
162static int nsw_wcount_async; /* limit write buffers / asynchronous */
163static int nsw_wcount_async_max;/* assigned maximum */
164static int nsw_cluster_max; /* maximum VOP I/O allowed */
984263bc
MD
165
166struct blist *swapblist;
984263bc 167static int swap_async_max = 4; /* maximum in-progress async I/O's */
5d5c5831 168static int swap_burst_read = 0; /* allow burst reading */
984263bc
MD
169
170extern struct vnode *swapdev_vp; /* from vm_swap.c */
171
172SYSCTL_INT(_vm, OID_AUTO, swap_async_max,
173 CTLFLAG_RW, &swap_async_max, 0, "Maximum running async swap ops");
5d5c5831
MD
174SYSCTL_INT(_vm, OID_AUTO, swap_burst_read,
175 CTLFLAG_RW, &swap_burst_read, 0, "Allow burst reads for pageins");
984263bc 176
096e95c0
MD
177SYSCTL_INT(_vm, OID_AUTO, swap_cache_use,
178 CTLFLAG_RD, &vm_swap_cache_use, 0, "");
179SYSCTL_INT(_vm, OID_AUTO, swap_anon_use,
180 CTLFLAG_RD, &vm_swap_anon_use, 0, "");
181
984263bc
MD
182vm_zone_t swap_zone;
183
96adc753
MD
184/*
185 * Red-Black tree for swblock entries
186 */
187RB_GENERATE2(swblock_rb_tree, swblock, swb_entry, rb_swblock_compare,
188 vm_pindex_t, swb_index);
189
190int
191rb_swblock_compare(struct swblock *swb1, struct swblock *swb2)
192{
193 if (swb1->swb_index < swb2->swb_index)
194 return(-1);
195 if (swb1->swb_index > swb2->swb_index)
196 return(1);
197 return(0);
198}
199
8d292090
MD
200static
201int
202rb_swblock_scancmp(struct swblock *swb, void *data)
203{
204 struct swfreeinfo *info = data;
205
206 if (swb->swb_index < info->basei)
207 return(-1);
208 if (swb->swb_index > info->endi)
209 return(1);
210 return(0);
211}
212
984263bc
MD
213/*
214 * pagerops for OBJT_SWAP - "swap pager". Some ops are also global procedure
215 * calls hooked from other parts of the VM system and do not appear here.
216 * (see vm/swap_pager.h).
217 */
218
219static vm_object_t
57f7b636
MD
220 swap_pager_alloc (void *handle, off_t size,
221 vm_prot_t prot, off_t offset);
1388df65 222static void swap_pager_dealloc (vm_object_t object);
1b9d3514 223static int swap_pager_getpage (vm_object_t, vm_page_t *, int);
81b5c339 224static void swap_chain_iodone(struct bio *biox);
984263bc
MD
225
226struct pagerops swappagerops = {
984263bc
MD
227 swap_pager_alloc, /* allocate an OBJT_SWAP object */
228 swap_pager_dealloc, /* deallocate an OBJT_SWAP object */
1b9d3514 229 swap_pager_getpage, /* pagein */
984263bc 230 swap_pager_putpages, /* pageout */
107e9bcc 231 swap_pager_haspage /* get backing store status for page */
984263bc
MD
232};
233
234/*
235 * dmmax is in page-sized chunks with the new swap system. It was
236 * dev-bsized chunks in the old. dmmax is always a power of 2.
237 *
238 * swap_*() routines are externally accessible. swp_*() routines are
239 * internal.
240 */
241
242int dmmax;
243static int dmmax_mask;
244int nswap_lowat = 128; /* in pages, swap_pager_almost_full warn */
245int nswap_hiwat = 512; /* in pages, swap_pager_almost_full warn */
246
1388df65 247static __inline void swp_sizecheck (void);
81b5c339 248static void swp_pager_async_iodone (struct bio *bio);
984263bc
MD
249
250/*
251 * Swap bitmap functions
252 */
253
096e95c0
MD
254static __inline void swp_pager_freeswapspace (vm_object_t object, daddr_t blk, int npages);
255static __inline daddr_t swp_pager_getswapspace (vm_object_t object, int npages);
984263bc
MD
256
257/*
258 * Metadata functions
259 */
260
96adc753 261static void swp_pager_meta_convert (vm_object_t);
1388df65 262static void swp_pager_meta_build (vm_object_t, vm_pindex_t, daddr_t);
8d292090 263static void swp_pager_meta_free (vm_object_t, vm_pindex_t, vm_pindex_t);
1388df65
RG
264static void swp_pager_meta_free_all (vm_object_t);
265static daddr_t swp_pager_meta_ctl (vm_object_t, vm_pindex_t, int);
984263bc
MD
266
267/*
268 * SWP_SIZECHECK() - update swap_pager_full indication
269 *
270 * update the swap_pager_almost_full indication and warn when we are
271 * about to run out of swap space, using lowat/hiwat hysteresis.
272 *
273 * Clear swap_pager_full ( task killing ) indication when lowat is met.
274 *
275 * No restrictions on call
276 * This routine may not block.
277 * This routine must be called at splvm()
278 */
279
280static __inline void
57e43348 281swp_sizecheck(void)
984263bc
MD
282{
283 if (vm_swap_size < nswap_lowat) {
284 if (swap_pager_almost_full == 0) {
086c1d7e 285 kprintf("swap_pager: out of swap space\n");
984263bc
MD
286 swap_pager_almost_full = 1;
287 }
288 } else {
289 swap_pager_full = 0;
290 if (vm_swap_size > nswap_hiwat)
291 swap_pager_almost_full = 0;
292 }
293}
294
295/*
296 * SWAP_PAGER_INIT() - initialize the swap pager!
297 *
298 * Expected to be started from system init. NOTE: This code is run
299 * before much else so be careful what you depend on. Most of the VM
300 * system has yet to be initialized at this point.
301 */
984263bc 302static void
107e9bcc 303swap_pager_init(void *arg __unused)
984263bc 304{
984263bc
MD
305 /*
306 * Device Stripe, in PAGE_SIZE'd blocks
307 */
984263bc
MD
308 dmmax = SWB_NPAGES * 2;
309 dmmax_mask = ~(dmmax - 1);
310}
107e9bcc 311SYSINIT(vm_mem, SI_BOOT1_VM, SI_ORDER_THIRD, swap_pager_init, NULL)
984263bc
MD
312
313/*
314 * SWAP_PAGER_SWAP_INIT() - swap pager initialization from pageout process
315 *
316 * Expected to be started from pageout process once, prior to entering
317 * its main loop.
318 */
319
320void
57e43348 321swap_pager_swap_init(void)
984263bc
MD
322{
323 int n, n2;
324
325 /*
326 * Number of in-transit swap bp operations. Don't
327 * exhaust the pbufs completely. Make sure we
328 * initialize workable values (0 will work for hysteresis
329 * but it isn't very efficient).
330 *
54f51aeb
HP
331 * The nsw_cluster_max is constrained by the number of pages an XIO
332 * holds, i.e., (MAXPHYS/PAGE_SIZE) and our locally defined
984263bc
MD
333 * MAX_PAGEOUT_CLUSTER. Also be aware that swap ops are
334 * constrained by the swap device interleave stripe size.
335 *
336 * Currently we hardwire nsw_wcount_async to 4. This limit is
337 * designed to prevent other I/O from having high latencies due to
338 * our pageout I/O. The value 4 works well for one or two active swap
339 * devices but is probably a little low if you have more. Even so,
340 * a higher value would probably generate only a limited improvement
341 * with three or four active swap devices since the system does not
342 * typically have to pageout at extreme bandwidths. We will want
343 * at least 2 per swap devices, and 4 is a pretty good value if you
344 * have one NFS swap device due to the command/ack latency over NFS.
345 * So it all works out pretty well.
346 */
347
348 nsw_cluster_max = min((MAXPHYS/PAGE_SIZE), MAX_PAGEOUT_CLUSTER);
349
350 nsw_rcount = (nswbuf + 1) / 2;
351 nsw_wcount_sync = (nswbuf + 3) / 4;
352 nsw_wcount_async = 4;
353 nsw_wcount_async_max = nsw_wcount_async;
354
355 /*
79634a66
MD
356 * The zone is dynamically allocated so generally size it to
357 * maxswzone (32MB to 512MB of KVM). Set a minimum size based
358 * on physical memory of around 8x (each swblock can hold 16 pages).
359 *
360 * With the advent of SSDs (vs HDs) the practical (swap:memory) ratio
361 * has increased dramatically.
984263bc 362 */
12e4aaff 363 n = vmstats.v_page_count / 2;
79634a66 364 if (maxswzone && n < maxswzone / sizeof(struct swblock))
984263bc
MD
365 n = maxswzone / sizeof(struct swblock);
366 n2 = n;
367
368 do {
369 swap_zone = zinit(
370 "SWAPMETA",
371 sizeof(struct swblock),
372 n,
373 ZONE_INTERRUPT,
374 1);
375 if (swap_zone != NULL)
376 break;
377 /*
378 * if the allocation failed, try a zone two thirds the
379 * size of the previous attempt.
380 */
381 n -= ((n + 2) / 3);
382 } while (n > 0);
383
384 if (swap_zone == NULL)
385 panic("swap_pager_swap_init: swap_zone == NULL");
386 if (n2 != n)
086c1d7e 387 kprintf("Swap zone entries reduced from %d to %d.\n", n2, n);
984263bc
MD
388}
389
390/*
391 * SWAP_PAGER_ALLOC() - allocate a new OBJT_SWAP VM object and instantiate
392 * its metadata structures.
393 *
394 * This routine is called from the mmap and fork code to create a new
395 * OBJT_SWAP object. We do this by creating an OBJT_DEFAULT object
96adc753 396 * and then converting it with swp_pager_meta_convert().
984263bc
MD
397 *
398 * This routine may block in vm_object_allocate() and create a named
399 * object lookup race, so we must interlock. We must also run at
400 * splvm() for the object lookup to handle races with interrupts, but
401 * we do not have to maintain splvm() in between the lookup and the
402 * add because (I believe) it is not possible to attempt to create
403 * a new swap object w/handle when a default object with that handle
404 * already exists.
405 */
406
407static vm_object_t
57f7b636 408swap_pager_alloc(void *handle, off_t size, vm_prot_t prot, off_t offset)
984263bc
MD
409{
410 vm_object_t object;
411
d28e1355
MD
412 KKASSERT(handle == NULL);
413#if 0
984263bc
MD
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 */
984263bc
MD
421 while (sw_alloc_interlock) {
422 sw_alloc_interlock = -1;
377d4740 423 tsleep(&sw_alloc_interlock, 0, "swpalc", 0);
984263bc
MD
424 }
425 sw_alloc_interlock = 1;
426
427 object = vm_pager_object_lookup(NOBJLIST(handle), handle);
428
429 if (object != NULL) {
430 vm_object_reference(object);
431 } else {
432 object = vm_object_allocate(OBJT_DEFAULT,
433 OFF_TO_IDX(offset + PAGE_MASK + size));
434 object->handle = handle;
96adc753 435 swp_pager_meta_convert(object);
984263bc
MD
436 }
437
438 if (sw_alloc_interlock < 0)
439 wakeup(&sw_alloc_interlock);
984263bc 440 sw_alloc_interlock = 0;
d28e1355
MD
441 } else { ... }
442#endif
443 object = vm_object_allocate(OBJT_DEFAULT,
444 OFF_TO_IDX(offset + PAGE_MASK + size));
445 swp_pager_meta_convert(object);
984263bc
MD
446
447 return (object);
448}
449
450/*
451 * SWAP_PAGER_DEALLOC() - remove swap metadata from object
452 *
453 * The swap backing for the object is destroyed. The code is
454 * designed such that we can reinstantiate it later, but this
455 * routine is typically called only when the entire object is
456 * about to be destroyed.
457 *
458 * This routine may block, but no longer does.
459 *
460 * The object must be locked or unreferenceable.
461 */
462
463static void
57e43348 464swap_pager_dealloc(vm_object_t object)
984263bc 465{
984263bc
MD
466 vm_object_pip_wait(object, "swpdea");
467
468 /*
469 * Free all remaining metadata. We only bother to free it from
470 * the swap meta data. We do not attempt to free swapblk's still
471 * associated with vm_page_t's for this object. We do not care
472 * if paging is still in progress on some objects.
473 */
cdd46d2e 474 crit_enter();
984263bc 475 swp_pager_meta_free_all(object);
cdd46d2e 476 crit_exit();
984263bc
MD
477}
478
479/************************************************************************
480 * SWAP PAGER BITMAP ROUTINES *
481 ************************************************************************/
482
483/*
484 * SWP_PAGER_GETSWAPSPACE() - allocate raw swap space
485 *
486 * Allocate swap for the requested number of pages. The starting
487 * swap block number (a page index) is returned or SWAPBLK_NONE
488 * if the allocation failed.
489 *
490 * Also has the side effect of advising that somebody made a mistake
491 * when they configured swap and didn't configure enough.
492 *
493 * Must be called at splvm() to avoid races with bitmap frees from
494 * vm_page_remove() aka swap_pager_page_removed().
495 *
496 * This routine may not block
497 * This routine must be called at splvm().
498 */
984263bc 499static __inline daddr_t
096e95c0 500swp_pager_getswapspace(vm_object_t object, int npages)
984263bc
MD
501{
502 daddr_t blk;
503
504 if ((blk = blist_alloc(swapblist, npages)) == SWAPBLK_NONE) {
505 if (swap_pager_full != 2) {
086c1d7e 506 kprintf("swap_pager_getswapspace: failed\n");
984263bc
MD
507 swap_pager_full = 2;
508 swap_pager_almost_full = 1;
509 }
510 } else {
511 vm_swap_size -= npages;
096e95c0
MD
512 if (object->type == OBJT_SWAP)
513 vm_swap_anon_use += npages;
514 else
515 vm_swap_cache_use += npages;
984263bc
MD
516 swp_sizecheck();
517 }
518 return(blk);
519}
520
521/*
522 * SWP_PAGER_FREESWAPSPACE() - free raw swap space
523 *
524 * This routine returns the specified swap blocks back to the bitmap.
525 *
526 * Note: This routine may not block (it could in the old swap code),
527 * and through the use of the new blist routines it does not block.
528 *
529 * We must be called at splvm() to avoid races with bitmap frees from
530 * vm_page_remove() aka swap_pager_page_removed().
531 *
532 * This routine may not block
533 * This routine must be called at splvm().
534 */
535
536static __inline void
096e95c0 537swp_pager_freeswapspace(vm_object_t object, daddr_t blk, int npages)
984263bc
MD
538{
539 blist_free(swapblist, blk, npages);
540 vm_swap_size += npages;
096e95c0
MD
541 if (object->type == OBJT_SWAP)
542 vm_swap_anon_use -= npages;
543 else
544 vm_swap_cache_use -= npages;
984263bc
MD
545 swp_sizecheck();
546}
547
548/*
549 * SWAP_PAGER_FREESPACE() - frees swap blocks associated with a page
550 * range within an object.
551 *
552 * This is a globally accessible routine.
553 *
554 * This routine removes swapblk assignments from swap metadata.
555 *
556 * The external callers of this routine typically have already destroyed
557 * or renamed vm_page_t's associated with this range in the object so
558 * we should be ok.
559 *
96adc753
MD
560 * This routine may be called at any spl. We up our spl to splvm
561 * temporarily in order to perform the metadata removal.
984263bc 562 */
984263bc 563void
8d292090 564swap_pager_freespace(vm_object_t object, vm_pindex_t start, vm_pindex_t size)
984263bc 565{
cdd46d2e 566 crit_enter();
984263bc 567 swp_pager_meta_free(object, start, size);
cdd46d2e 568 crit_exit();
984263bc
MD
569}
570
8d292090
MD
571void
572swap_pager_freespace_all(vm_object_t object)
573{
574 crit_enter();
575 swp_pager_meta_free_all(object);
576 crit_exit();
577}
578
096e95c0
MD
579/*
580 * Called by vm_page_alloc() when a new VM page is inserted
581 * into a VM object. Checks whether swap has been assigned to
582 * the page and sets PG_SWAPPED as necessary.
583 */
584void
585swap_pager_page_inserted(vm_page_t m)
586{
587 if (m->object->swblock_count) {
588 crit_enter();
589 if (swp_pager_meta_ctl(m->object, m->pindex, 0) != SWAPBLK_NONE)
590 vm_page_flag_set(m, PG_SWAPPED);
591 crit_exit();
592 }
593}
594
984263bc
MD
595/*
596 * SWAP_PAGER_RESERVE() - reserve swap blocks in object
597 *
598 * Assigns swap blocks to the specified range within the object. The
599 * swap blocks are not zerod. Any previous swap assignment is destroyed.
600 *
601 * Returns 0 on success, -1 on failure.
602 */
984263bc
MD
603int
604swap_pager_reserve(vm_object_t object, vm_pindex_t start, vm_size_t size)
605{
984263bc
MD
606 int n = 0;
607 daddr_t blk = SWAPBLK_NONE;
608 vm_pindex_t beg = start; /* save start index */
609
cdd46d2e 610 crit_enter();
984263bc
MD
611 while (size) {
612 if (n == 0) {
613 n = BLIST_MAX_ALLOC;
096e95c0
MD
614 while ((blk = swp_pager_getswapspace(object, n)) ==
615 SWAPBLK_NONE)
616 {
984263bc
MD
617 n >>= 1;
618 if (n == 0) {
8d292090
MD
619 swp_pager_meta_free(object, beg,
620 start - beg);
cdd46d2e 621 crit_exit();
984263bc
MD
622 return(-1);
623 }
624 }
625 }
626 swp_pager_meta_build(object, start, blk);
627 --size;
628 ++start;
629 ++blk;
630 --n;
631 }
632 swp_pager_meta_free(object, start, n);
cdd46d2e 633 crit_exit();
984263bc
MD
634 return(0);
635}
636
637/*
638 * SWAP_PAGER_COPY() - copy blocks from source pager to destination pager
639 * and destroy the source.
640 *
641 * Copy any valid swapblks from the source to the destination. In
642 * cases where both the source and destination have a valid swapblk,
643 * we keep the destination's.
644 *
645 * This routine is allowed to block. It may block allocating metadata
646 * indirectly through swp_pager_meta_build() or if paging is still in
647 * progress on the source.
648 *
649 * This routine can be called at any spl
650 *
651 * XXX vm_page_collapse() kinda expects us not to block because we
652 * supposedly do not need to allocate memory, but for the moment we
653 * *may* have to get a little memory from the zone allocator, but
654 * it is taken from the interrupt memory. We should be ok.
655 *
656 * The source object contains no vm_page_t's (which is just as well)
657 *
658 * The source object is of type OBJT_SWAP.
659 *
660 * The source and destination objects must be locked or
661 * inaccessible (XXX are they ?)
662 */
663
664void
57e43348 665swap_pager_copy(vm_object_t srcobject, vm_object_t dstobject,
8d292090 666 vm_pindex_t base_index, int destroysource)
984263bc
MD
667{
668 vm_pindex_t i;
984263bc 669
cdd46d2e 670 crit_enter();
984263bc 671
984263bc
MD
672 /*
673 * transfer source to destination.
674 */
984263bc
MD
675 for (i = 0; i < dstobject->size; ++i) {
676 daddr_t dstaddr;
677
678 /*
679 * Locate (without changing) the swapblk on the destination,
680 * unless it is invalid in which case free it silently, or
681 * if the destination is a resident page, in which case the
682 * source is thrown away.
683 */
984263bc
MD
684 dstaddr = swp_pager_meta_ctl(dstobject, i, 0);
685
686 if (dstaddr == SWAPBLK_NONE) {
687 /*
688 * Destination has no swapblk and is not resident,
689 * copy source.
690 */
691 daddr_t srcaddr;
692
8d292090
MD
693 srcaddr = swp_pager_meta_ctl(srcobject,
694 base_index + i, SWM_POP);
984263bc
MD
695
696 if (srcaddr != SWAPBLK_NONE)
697 swp_pager_meta_build(dstobject, i, srcaddr);
698 } else {
699 /*
700 * Destination has valid swapblk or it is represented
701 * by a resident page. We destroy the sourceblock.
702 */
8d292090 703 swp_pager_meta_ctl(srcobject, base_index + i, SWM_FREE);
984263bc
MD
704 }
705 }
706
707 /*
708 * Free left over swap blocks in source.
709 *
710 * We have to revert the type to OBJT_DEFAULT so we do not accidently
711 * double-remove the object from the swap queues.
712 */
984263bc 713 if (destroysource) {
984263bc
MD
714 /*
715 * Reverting the type is not necessary, the caller is going
716 * to destroy srcobject directly, but I'm doing it here
717 * for consistency since we've removed the object from its
718 * queues.
719 */
96adc753 720 swp_pager_meta_free_all(srcobject);
8d292090
MD
721 if (srcobject->type == OBJT_SWAP)
722 srcobject->type = OBJT_DEFAULT;
984263bc 723 }
cdd46d2e 724 crit_exit();
984263bc
MD
725}
726
727/*
728 * SWAP_PAGER_HASPAGE() - determine if we have good backing store for
729 * the requested page.
730 *
731 * We determine whether good backing store exists for the requested
732 * page and return TRUE if it does, FALSE if it doesn't.
733 *
734 * If TRUE, we also try to determine how much valid, contiguous backing
735 * store exists before and after the requested page within a reasonable
736 * distance. We do not try to restrict it to the swap device stripe
737 * (that is handled in getpages/putpages). It probably isn't worth
738 * doing here.
739 */
740
741boolean_t
1b9d3514 742swap_pager_haspage(vm_object_t object, vm_pindex_t pindex)
984263bc
MD
743{
744 daddr_t blk0;
984263bc
MD
745
746 /*
747 * do we have good backing store at the requested index ?
748 */
749
cdd46d2e 750 crit_enter();
984263bc
MD
751 blk0 = swp_pager_meta_ctl(object, pindex, 0);
752
753 if (blk0 == SWAPBLK_NONE) {
cdd46d2e 754 crit_exit();
984263bc
MD
755 return (FALSE);
756 }
757
1b9d3514 758#if 0
984263bc
MD
759 /*
760 * find backwards-looking contiguous good backing store
761 */
984263bc
MD
762 if (before != NULL) {
763 int i;
764
765 for (i = 1; i < (SWB_NPAGES/2); ++i) {
766 daddr_t blk;
767
768 if (i > pindex)
769 break;
770 blk = swp_pager_meta_ctl(object, pindex - i, 0);
771 if (blk != blk0 - i)
772 break;
773 }
774 *before = (i - 1);
775 }
776
777 /*
778 * find forward-looking contiguous good backing store
779 */
780
781 if (after != NULL) {
782 int i;
783
784 for (i = 1; i < (SWB_NPAGES/2); ++i) {
785 daddr_t blk;
786
787 blk = swp_pager_meta_ctl(object, pindex + i, 0);
788 if (blk != blk0 + i)
789 break;
790 }
791 *after = (i - 1);
792 }
1b9d3514 793#endif
cdd46d2e 794 crit_exit();
984263bc
MD
795 return (TRUE);
796}
797
798/*
799 * SWAP_PAGER_PAGE_UNSWAPPED() - remove swap backing store related to page
800 *
107e9bcc
MD
801 * This removes any associated swap backing store, whether valid or
802 * not, from the page. This operates on any VM object, not just OBJT_SWAP
803 * objects.
984263bc 804 *
107e9bcc
MD
805 * This routine is typically called when a page is made dirty, at
806 * which point any associated swap can be freed. MADV_FREE also
807 * calls us in a special-case situation
984263bc 808 *
107e9bcc
MD
809 * NOTE!!! If the page is clean and the swap was valid, the caller
810 * should make the page dirty before calling this routine. This routine
811 * does NOT change the m->dirty status of the page. Also: MADV_FREE
812 * depends on it.
984263bc 813 *
107e9bcc
MD
814 * This routine may not block
815 * This routine must be called at splvm()
984263bc 816 */
107e9bcc 817void
57e43348 818swap_pager_unswapped(vm_page_t m)
984263bc 819{
67803f3e
MD
820 if (m->flags & PG_SWAPPED) {
821 swp_pager_meta_ctl(m->object, m->pindex, SWM_FREE);
822 vm_page_flag_clear(m, PG_SWAPPED);
823 }
984263bc
MD
824}
825
826/*
827 * SWAP_PAGER_STRATEGY() - read, write, free blocks
828 *
107e9bcc
MD
829 * This implements a VM OBJECT strategy function using swap backing store.
830 * This can operate on any VM OBJECT type, not necessarily just OBJT_SWAP
831 * types.
832 *
833 * This is intended to be a cacheless interface (i.e. caching occurs at
834 * higher levels), and is also used as a swap-based SSD cache for vnode
835 * and device objects.
836 *
837 * All I/O goes directly to and from the swap device.
984263bc 838 *
107e9bcc
MD
839 * We currently attempt to run I/O synchronously or asynchronously as
840 * the caller requests. This isn't perfect because we loose error
841 * sequencing when we run multiple ops in parallel to satisfy a request.
842 * But this is swap, so we let it all hang out.
984263bc 843 */
107e9bcc 844void
81b5c339 845swap_pager_strategy(vm_object_t object, struct bio *bio)
984263bc 846{
81b5c339
MD
847 struct buf *bp = bio->bio_buf;
848 struct bio *nbio;
984263bc 849 vm_pindex_t start;
54078292 850 vm_pindex_t biox_blkno = 0;
984263bc 851 int count;
984263bc 852 char *data;
ae8e83e6
MD
853 struct bio *biox;
854 struct buf *bufx;
81b5c339
MD
855 struct bio_track *track;
856
857 /*
858 * tracking for swapdev vnode I/Os
859 */
10f3fee5 860 if (bp->b_cmd == BUF_CMD_READ)
81b5c339
MD
861 track = &swapdev_vp->v_track_read;
862 else
863 track = &swapdev_vp->v_track_write;
984263bc
MD
864
865 if (bp->b_bcount & PAGE_MASK) {
866 bp->b_error = EINVAL;
867 bp->b_flags |= B_ERROR | B_INVAL;
81b5c339 868 biodone(bio);
973c11b9
MD
869 kprintf("swap_pager_strategy: bp %p offset %lld size %d, "
870 "not page bounded\n",
871 bp, (long long)bio->bio_offset, (int)bp->b_bcount);
984263bc
MD
872 return;
873 }
874
875 /*
876 * Clear error indication, initialize page index, count, data pointer.
877 */
984263bc
MD
878 bp->b_error = 0;
879 bp->b_flags &= ~B_ERROR;
880 bp->b_resid = bp->b_bcount;
881
54078292 882 start = (vm_pindex_t)(bio->bio_offset >> PAGE_SHIFT);
984263bc
MD
883 count = howmany(bp->b_bcount, PAGE_SIZE);
884 data = bp->b_data;
885
984263bc 886 /*
10f3fee5 887 * Deal with BUF_CMD_FREEBLKS
984263bc 888 */
10f3fee5 889 if (bp->b_cmd == BUF_CMD_FREEBLKS) {
984263bc
MD
890 /*
891 * FREE PAGE(s) - destroy underlying swap that is no longer
892 * needed.
893 */
894 swp_pager_meta_free(object, start, count);
984263bc 895 bp->b_resid = 0;
81b5c339 896 biodone(bio);
984263bc
MD
897 return;
898 }
899
81b5c339
MD
900 /*
901 * We need to be able to create a new cluster of I/O's. We cannot
902 * use the caller fields of the passed bio so push a new one.
903 *
904 * Because nbio is just a placeholder for the cluster links,
905 * we can biodone() the original bio instead of nbio to make
906 * things a bit more efficient.
907 */
908 nbio = push_bio(bio);
54078292 909 nbio->bio_offset = bio->bio_offset;
81b5c339
MD
910 nbio->bio_caller_info1.cluster_head = NULL;
911 nbio->bio_caller_info2.cluster_tail = NULL;
912
ae8e83e6
MD
913 biox = NULL;
914 bufx = NULL;
915
984263bc
MD
916 /*
917 * Execute read or write
918 */
984263bc
MD
919 while (count > 0) {
920 daddr_t blk;
921
922 /*
923 * Obtain block. If block not found and writing, allocate a
924 * new block and build it into the object.
925 */
984263bc 926 blk = swp_pager_meta_ctl(object, start, 0);
10f3fee5 927 if ((blk == SWAPBLK_NONE) && bp->b_cmd != BUF_CMD_READ) {
096e95c0 928 blk = swp_pager_getswapspace(object, 1);
984263bc
MD
929 if (blk == SWAPBLK_NONE) {
930 bp->b_error = ENOMEM;
931 bp->b_flags |= B_ERROR;
932 break;
933 }
934 swp_pager_meta_build(object, start, blk);
935 }
936
937 /*
938 * Do we have to flush our current collection? Yes if:
939 *
940 * - no swap block at this index
941 * - swap block is not contiguous
942 * - we cross a physical disk boundry in the
943 * stripe.
944 */
984263bc 945 if (
54078292
MD
946 biox && (biox_blkno + btoc(bufx->b_bcount) != blk ||
947 ((biox_blkno ^ blk) & dmmax_mask)
984263bc
MD
948 )
949 ) {
10f3fee5 950 if (bp->b_cmd == BUF_CMD_READ) {
12e4aaff 951 ++mycpu->gd_cnt.v_swapin;
81b5c339 952 mycpu->gd_cnt.v_swappgsin += btoc(bufx->b_bcount);
984263bc 953 } else {
12e4aaff 954 ++mycpu->gd_cnt.v_swapout;
81b5c339
MD
955 mycpu->gd_cnt.v_swappgsout += btoc(bufx->b_bcount);
956 bufx->b_dirtyend = bufx->b_bcount;
957 }
958
959 /*
ae8e83e6 960 * Finished with this buf.
81b5c339 961 */
ae8e83e6
MD
962 KKASSERT(bufx->b_bcount != 0);
963 if (bufx->b_cmd != BUF_CMD_READ)
964 bufx->b_dirtyend = bufx->b_bcount;
81b5c339
MD
965 biox = NULL;
966 bufx = NULL;
984263bc
MD
967 }
968
969 /*
81b5c339 970 * Add new swapblk to biox, instantiating biox if necessary.
984263bc
MD
971 * Zero-fill reads are able to take a shortcut.
972 */
984263bc
MD
973 if (blk == SWAPBLK_NONE) {
974 /*
975 * We can only get here if we are reading. Since
976 * we are at splvm() we can safely modify b_resid,
977 * even if chain ops are in progress.
978 */
979 bzero(data, PAGE_SIZE);
980 bp->b_resid -= PAGE_SIZE;
981 } else {
81b5c339
MD
982 if (biox == NULL) {
983 /* XXX chain count > 4, wait to <= 4 */
984
985 bufx = getpbuf(NULL);
986 biox = &bufx->b_bio1;
987 cluster_append(nbio, bufx);
ae8e83e6 988 bufx->b_flags |= (bufx->b_flags & B_ORDERED);
10f3fee5 989 bufx->b_cmd = bp->b_cmd;
81b5c339 990 biox->bio_done = swap_chain_iodone;
54078292 991 biox->bio_offset = (off_t)blk << PAGE_SHIFT;
81b5c339 992 biox->bio_caller_info1.cluster_parent = nbio;
54078292 993 biox_blkno = blk;
81b5c339
MD
994 bufx->b_bcount = 0;
995 bufx->b_data = data;
984263bc 996 }
81b5c339 997 bufx->b_bcount += PAGE_SIZE;
984263bc
MD
998 }
999 --count;
1000 ++start;
1001 data += PAGE_SIZE;
1002 }
1003
1004 /*
1005 * Flush out last buffer
1006 */
81b5c339 1007 if (biox) {
10f3fee5 1008 if (bufx->b_cmd == BUF_CMD_READ) {
12e4aaff 1009 ++mycpu->gd_cnt.v_swapin;
81b5c339 1010 mycpu->gd_cnt.v_swappgsin += btoc(bufx->b_bcount);
984263bc 1011 } else {
12e4aaff 1012 ++mycpu->gd_cnt.v_swapout;
81b5c339
MD
1013 mycpu->gd_cnt.v_swappgsout += btoc(bufx->b_bcount);
1014 bufx->b_dirtyend = bufx->b_bcount;
1015 }
ae8e83e6
MD
1016 KKASSERT(bufx->b_bcount);
1017 if (bufx->b_cmd != BUF_CMD_READ)
1018 bufx->b_dirtyend = bufx->b_bcount;
81b5c339 1019 /* biox, bufx = NULL */
984263bc
MD
1020 }
1021
1022 /*
ae8e83e6
MD
1023 * Now initiate all the I/O. Be careful looping on our chain as
1024 * I/O's may complete while we are still initiating them.
984263bc 1025 */
ae8e83e6
MD
1026 nbio->bio_caller_info2.cluster_tail = NULL;
1027 bufx = nbio->bio_caller_info1.cluster_head;
1028
1029 while (bufx) {
1030 biox = &bufx->b_bio1;
1031 BUF_KERNPROC(bufx);
1032 bufx = bufx->b_cluster_next;
1033 vn_strategy(swapdev_vp, biox);
984263bc 1034 }
ae8e83e6
MD
1035
1036 /*
1037 * Completion of the cluster will also call biodone_chain(nbio).
1038 * We never call biodone(nbio) so we don't have to worry about
1039 * setting up a bio_done callback. It's handled in the sub-IO.
1040 */
1041 /**/
984263bc
MD
1042}
1043
81b5c339
MD
1044static void
1045swap_chain_iodone(struct bio *biox)
1046{
1047 struct buf **nextp;
1048 struct buf *bufx; /* chained sub-buffer */
1049 struct bio *nbio; /* parent nbio with chain glue */
1050 struct buf *bp; /* original bp associated with nbio */
ae8e83e6 1051 int chain_empty;
81b5c339
MD
1052
1053 bufx = biox->bio_buf;
1054 nbio = biox->bio_caller_info1.cluster_parent;
1055 bp = nbio->bio_buf;
1056
1057 /*
1058 * Update the original buffer
1059 */
1060 KKASSERT(bp != NULL);
1061 if (bufx->b_flags & B_ERROR) {
ae8e83e6 1062 atomic_set_int(&bufx->b_flags, B_ERROR);
81b5c339
MD
1063 bp->b_error = bufx->b_error;
1064 } else if (bufx->b_resid != 0) {
ae8e83e6 1065 atomic_set_int(&bufx->b_flags, B_ERROR);
81b5c339
MD
1066 bp->b_error = EINVAL;
1067 } else {
ae8e83e6 1068 atomic_subtract_int(&bp->b_resid, bufx->b_bcount);
81b5c339
MD
1069 }
1070
1071 /*
ae8e83e6 1072 * Remove us from the chain.
81b5c339 1073 */
ae8e83e6 1074 spin_lock_wr(&bp->b_lock.lk_spinlock);
81b5c339
MD
1075 nextp = &nbio->bio_caller_info1.cluster_head;
1076 while (*nextp != bufx) {
1077 KKASSERT(*nextp != NULL);
1078 nextp = &(*nextp)->b_cluster_next;
1079 }
1080 *nextp = bufx->b_cluster_next;
ae8e83e6
MD
1081 chain_empty = (nbio->bio_caller_info1.cluster_head == NULL);
1082 spin_unlock_wr(&bp->b_lock.lk_spinlock);
81b5c339
MD
1083
1084 /*
ae8e83e6
MD
1085 * Clean up bufx. If the chain is now empty we finish out
1086 * the parent. Note that we may be racing other completions
1087 * so we must use the chain_empty status from above.
81b5c339 1088 */
ae8e83e6 1089 if (chain_empty) {
81b5c339 1090 if (bp->b_resid != 0 && !(bp->b_flags & B_ERROR)) {
ae8e83e6 1091 atomic_set_int(&bp->b_flags, B_ERROR);
81b5c339
MD
1092 bp->b_error = EINVAL;
1093 }
ae8e83e6 1094 biodone_chain(nbio);
81b5c339 1095 }
81b5c339
MD
1096 relpbuf(bufx, NULL);
1097}
1098
984263bc 1099/*
5d5c5831 1100 * SWAP_PAGER_GETPAGES() - bring page in from swap
984263bc 1101 *
5d5c5831
MD
1102 * The requested page may have to be brought in from swap. Calculate the
1103 * swap block and bring in additional pages if possible. All pages must
1104 * have contiguous swap block assignments and reside in the same object.
984263bc 1105 *
5d5c5831
MD
1106 * The caller has a single vm_object_pip_add() reference prior to
1107 * calling us and we should return with the same.
984263bc 1108 *
5d5c5831
MD
1109 * The caller has BUSY'd the page. We should return with (*mpp) left busy,
1110 * and any additinal pages unbusied.
984263bc 1111 *
5d5c5831
MD
1112 * If the caller encounters a PG_RAM page it will pass it to us even though
1113 * it may be valid and dirty. We cannot overwrite the page in this case!
1114 * The case is used to allow us to issue pure read-aheads.
1115 *
1116 * NOTE! XXX This code does not entirely pipeline yet due to the fact that
1117 * the PG_RAM page is validated at the same time as mreq. What we
1118 * really need to do is issue a separate read-ahead pbuf.
984263bc 1119 */
984263bc 1120static int
1b9d3514 1121swap_pager_getpage(vm_object_t object, vm_page_t *mpp, int seqaccess)
984263bc
MD
1122{
1123 struct buf *bp;
81b5c339 1124 struct bio *bio;
984263bc 1125 vm_page_t mreq;
5d5c5831
MD
1126 vm_page_t m;
1127 vm_offset_t kva;
1128 daddr_t blk;
984263bc
MD
1129 int i;
1130 int j;
5d5c5831
MD
1131 int raonly;
1132 vm_page_t marray[XIO_INTERNAL_PAGES];
984263bc 1133
1b9d3514 1134 mreq = *mpp;
984263bc
MD
1135
1136 if (mreq->object != object) {
1137 panic("swap_pager_getpages: object mismatch %p/%p",
1138 object,
1139 mreq->object
1140 );
1141 }
17cde63e 1142
984263bc 1143 /*
5d5c5831
MD
1144 * We don't want to overwrite a fully valid page as it might be
1145 * dirty. This case can occur when e.g. vm_fault hits a perfectly
1146 * valid page with PG_RAM set.
984263bc 1147 *
5d5c5831
MD
1148 * In this case we see if the next page is a suitable page-in
1149 * candidate and if it is we issue read-ahead. PG_RAM will be
1150 * set on the last page of the read-ahead to continue the pipeline.
1151 */
1152 if (mreq->valid == VM_PAGE_BITS_ALL) {
3bb7eedb 1153 if (swap_burst_read == 0 || mreq->pindex + 1 >= object->size)
5d5c5831
MD
1154 return(VM_PAGER_OK);
1155 crit_enter();
3bb7eedb
MD
1156 blk = swp_pager_meta_ctl(object, mreq->pindex + 1, 0);
1157 if (blk == SWAPBLK_NONE) {
5d5c5831
MD
1158 crit_exit();
1159 return(VM_PAGER_OK);
1160 }
1161 m = vm_page_lookup(object, mreq->pindex + 1);
1162 if (m == NULL) {
1163 m = vm_page_alloc(object, mreq->pindex + 1,
1164 VM_ALLOC_QUICK);
1165 if (m == NULL) {
1166 crit_exit();
1167 return(VM_PAGER_OK);
1168 }
1169 } else {
5d5c5831
MD
1170 if ((m->flags & PG_BUSY) || m->busy || m->valid) {
1171 crit_exit();
1172 return(VM_PAGER_OK);
1173 }
3bb7eedb 1174 vm_page_unqueue_nowakeup(m);
5d5c5831 1175 vm_page_busy(m);
5d5c5831
MD
1176 }
1177 mreq = m;
1178 raonly = 1;
1179 crit_exit();
1180 } else {
1181 raonly = 0;
1182 }
1183
1184 /*
1185 * Try to block-read contiguous pages from swap if sequential,
1186 * otherwise just read one page. Contiguous pages from swap must
1187 * reside within a single device stripe because the I/O cannot be
1188 * broken up across multiple stripes.
1189 *
1190 * Note that blk and iblk can be SWAPBLK_NONE but the loop is
1191 * set up such that the case(s) are handled implicitly.
984263bc 1192 */
cdd46d2e 1193 crit_enter();
984263bc 1194 blk = swp_pager_meta_ctl(mreq->object, mreq->pindex, 0);
5d5c5831 1195 marray[0] = mreq;
984263bc 1196
5d5c5831
MD
1197 for (i = 1; swap_burst_read &&
1198 i < XIO_INTERNAL_PAGES &&
1199 mreq->pindex + i < object->size; ++i) {
984263bc
MD
1200 daddr_t iblk;
1201
5d5c5831
MD
1202 iblk = swp_pager_meta_ctl(object, mreq->pindex + i, 0);
1203 if (iblk != blk + i)
984263bc
MD
1204 break;
1205 if ((blk ^ iblk) & dmmax_mask)
1206 break;
5d5c5831
MD
1207 m = vm_page_lookup(object, mreq->pindex + i);
1208 if (m == NULL) {
1209 m = vm_page_alloc(object, mreq->pindex + i,
1210 VM_ALLOC_QUICK);
1211 if (m == NULL)
1212 break;
1213 } else {
5d5c5831
MD
1214 if ((m->flags & PG_BUSY) || m->busy || m->valid)
1215 break;
3bb7eedb 1216 vm_page_unqueue_nowakeup(m);
5d5c5831 1217 vm_page_busy(m);
5d5c5831
MD
1218 }
1219 marray[i] = m;
984263bc 1220 }
5d5c5831
MD
1221 if (i > 1)
1222 vm_page_flag_set(marray[i - 1], PG_RAM);
984263bc 1223
cdd46d2e 1224 crit_exit();
984263bc 1225
984263bc 1226 /*
5d5c5831
MD
1227 * If mreq is the requested page and we have nothing to do return
1228 * VM_PAGER_FAIL. If raonly is set mreq is just another read-ahead
1229 * page and must be cleaned up.
984263bc 1230 */
5d5c5831
MD
1231 if (blk == SWAPBLK_NONE) {
1232 KKASSERT(i == 1);
1233 if (raonly) {
1234 vnode_pager_freepage(mreq);
1235 return(VM_PAGER_OK);
1236 } else {
1237 return(VM_PAGER_FAIL);
1238 }
1239 }
984263bc
MD
1240
1241 /*
5d5c5831 1242 * map our page(s) into kva for input
984263bc 1243 */
984263bc 1244 bp = getpbuf(&nsw_rcount);
81b5c339 1245 bio = &bp->b_bio1;
5d5c5831
MD
1246 kva = (vm_offset_t) bp->b_kvabase;
1247 bcopy(marray, bp->b_xio.xio_pages, i * sizeof(vm_page_t));
1248 pmap_qenter(kva, bp->b_xio.xio_pages, i);
984263bc 1249
5d5c5831
MD
1250 bp->b_data = (caddr_t)kva;
1251 bp->b_bcount = PAGE_SIZE * i;
1252 bp->b_xio.xio_npages = i;
81b5c339 1253 bio->bio_done = swp_pager_async_iodone;
5d5c5831 1254 bio->bio_offset = (off_t)blk << PAGE_SHIFT;
8aa92e4b 1255 bio->bio_caller_info1.index = SWBIO_READ;
984263bc 1256
5d5c5831
MD
1257 /*
1258 * Set index. If raonly set the index beyond the array so all
1259 * the pages are treated the same, otherwise the original mreq is
1260 * at index 0.
1261 */
1262 if (raonly)
1263 bio->bio_driver_info = (void *)(intptr_t)i;
1264 else
1265 bio->bio_driver_info = (void *)(intptr_t)0;
984263bc 1266
5d5c5831
MD
1267 for (j = 0; j < i; ++j)
1268 vm_page_flag_set(bp->b_xio.xio_pages[j], PG_SWAPINPROG);
984263bc 1269
12e4aaff 1270 mycpu->gd_cnt.v_swapin++;
54f51aeb 1271 mycpu->gd_cnt.v_swappgsin += bp->b_xio.xio_npages;
984263bc
MD
1272
1273 /*
1274 * We still hold the lock on mreq, and our automatic completion routine
1275 * does not remove it.
1276 */
3bb7eedb 1277 vm_object_pip_add(object, bp->b_xio.xio_npages);
984263bc
MD
1278
1279 /*
1280 * perform the I/O. NOTE!!! bp cannot be considered valid after
1281 * this point because we automatically release it on completion.
1282 * Instead, we look at the one page we are interested in which we
1283 * still hold a lock on even through the I/O completion.
1284 *
1285 * The other pages in our m[] array are also released on completion,
1286 * so we cannot assume they are valid anymore either.
984263bc 1287 */
10f3fee5 1288 bp->b_cmd = BUF_CMD_READ;
984263bc 1289 BUF_KERNPROC(bp);
81b5c339 1290 vn_strategy(swapdev_vp, bio);
984263bc
MD
1291
1292 /*
3bb7eedb 1293 * Wait for the page we want to complete. PG_SWAPINPROG is always
984263bc
MD
1294 * cleared on completion. If an I/O error occurs, SWAPBLK_NONE
1295 * is set in the meta-data.
5d5c5831
MD
1296 *
1297 * If this is a read-ahead only we return immediately without
1298 * waiting for I/O.
984263bc 1299 */
5d5c5831
MD
1300 if (raonly)
1301 return(VM_PAGER_OK);
984263bc 1302
3bb7eedb
MD
1303 /*
1304 * Read-ahead includes originally requested page case.
1305 */
cdd46d2e 1306 crit_enter();
984263bc
MD
1307 while ((mreq->flags & PG_SWAPINPROG) != 0) {
1308 vm_page_flag_set(mreq, PG_WANTED | PG_REFERENCED);
12e4aaff 1309 mycpu->gd_cnt.v_intrans++;
377d4740 1310 if (tsleep(mreq, 0, "swread", hz*20)) {
086c1d7e 1311 kprintf(
81b5c339 1312 "swap_pager: indefinite wait buffer: "
973c11b9
MD
1313 " offset: %lld, size: %ld\n",
1314 (long long)bio->bio_offset,
1315 (long)bp->b_bcount
984263bc
MD
1316 );
1317 }
1318 }
cdd46d2e 1319 crit_exit();
984263bc
MD
1320
1321 /*
1322 * mreq is left bussied after completion, but all the other pages
1323 * are freed. If we had an unrecoverable read error the page will
1324 * not be valid.
1325 */
5d5c5831 1326 if (mreq->valid != VM_PAGE_BITS_ALL)
984263bc 1327 return(VM_PAGER_ERROR);
5d5c5831 1328 else
984263bc 1329 return(VM_PAGER_OK);
984263bc
MD
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 */
96adc753
MD
1382 if (object->type == OBJT_DEFAULT)
1383 swp_pager_meta_convert(object);
984263bc 1384
bc6dffab 1385 if (curthread != pagethread)
984263bc
MD
1386 sync = TRUE;
1387
1388 /*
1389 * Step 2
1390 *
1391 * Update nsw parameters from swap_async_max sysctl values.
1392 * Do not let the sysop crash the machine with bogus numbers.
1393 */
1394
1395 if (swap_async_max != nsw_wcount_async_max) {
1396 int n;
984263bc
MD
1397
1398 /*
1399 * limit range
1400 */
1401 if ((n = swap_async_max) > nswbuf / 2)
1402 n = nswbuf / 2;
1403 if (n < 1)
1404 n = 1;
1405 swap_async_max = n;
1406
1407 /*
1408 * Adjust difference ( if possible ). If the current async
1409 * count is too low, we may not be able to make the adjustment
1410 * at this time.
1411 */
cdd46d2e 1412 crit_enter();
984263bc
MD
1413 n -= nsw_wcount_async_max;
1414 if (nsw_wcount_async + n >= 0) {
1415 nsw_wcount_async += n;
1416 nsw_wcount_async_max += n;
1417 wakeup(&nsw_wcount_async);
1418 }
cdd46d2e 1419 crit_exit();
984263bc
MD
1420 }
1421
1422 /*
1423 * Step 3
1424 *
1425 * Assign swap blocks and issue I/O. We reallocate swap on the fly.
1426 * The page is left dirty until the pageout operation completes
1427 * successfully.
1428 */
1429
1430 for (i = 0; i < count; i += n) {
984263bc 1431 struct buf *bp;
81b5c339 1432 struct bio *bio;
984263bc 1433 daddr_t blk;
81b5c339 1434 int j;
984263bc
MD
1435
1436 /*
1437 * Maximum I/O size is limited by a number of factors.
1438 */
1439
1440 n = min(BLIST_MAX_ALLOC, count - i);
1441 n = min(n, nsw_cluster_max);
1442
cdd46d2e 1443 crit_enter();
984263bc
MD
1444
1445 /*
1446 * Get biggest block of swap we can. If we fail, fall
1447 * back and try to allocate a smaller block. Don't go
1448 * overboard trying to allocate space if it would overly
1449 * fragment swap.
1450 */
1451 while (
096e95c0 1452 (blk = swp_pager_getswapspace(object, n)) == SWAPBLK_NONE &&
984263bc
MD
1453 n > 4
1454 ) {
1455 n >>= 1;
1456 }
1457 if (blk == SWAPBLK_NONE) {
1458 for (j = 0; j < n; ++j)
1459 rtvals[i+j] = VM_PAGER_FAIL;
cdd46d2e 1460 crit_exit();
984263bc
MD
1461 continue;
1462 }
1463
1464 /*
1465 * The I/O we are constructing cannot cross a physical
1466 * disk boundry in the swap stripe. Note: we are still
1467 * at splvm().
1468 */
1469 if ((blk ^ (blk + n)) & dmmax_mask) {
1470 j = ((blk + dmmax) & dmmax_mask) - blk;
096e95c0 1471 swp_pager_freeswapspace(object, blk + j, n - j);
984263bc
MD
1472 n = j;
1473 }
1474
1475 /*
1476 * All I/O parameters have been satisfied, build the I/O
1477 * request and assign the swap space.
984263bc 1478 */
10f3fee5 1479 if (sync == TRUE)
984263bc 1480 bp = getpbuf(&nsw_wcount_sync);
10f3fee5 1481 else
984263bc 1482 bp = getpbuf(&nsw_wcount_async);
81b5c339 1483 bio = &bp->b_bio1;
984263bc
MD
1484
1485 pmap_qenter((vm_offset_t)bp->b_data, &m[i], n);
1486
984263bc 1487 bp->b_bcount = PAGE_SIZE * n;
54078292 1488 bio->bio_offset = (off_t)blk << PAGE_SHIFT;
984263bc 1489
984263bc
MD
1490 for (j = 0; j < n; ++j) {
1491 vm_page_t mreq = m[i+j];
1492
096e95c0
MD
1493 swp_pager_meta_build(mreq->object, mreq->pindex,
1494 blk + j);
1495 if (object->type == OBJT_SWAP)
1496 vm_page_dirty(mreq);
984263bc
MD
1497 rtvals[i+j] = VM_PAGER_OK;
1498
1499 vm_page_flag_set(mreq, PG_SWAPINPROG);
54f51aeb 1500 bp->b_xio.xio_pages[j] = mreq;
984263bc 1501 }
54f51aeb 1502 bp->b_xio.xio_npages = n;
984263bc 1503
12e4aaff 1504 mycpu->gd_cnt.v_swapout++;
54f51aeb 1505 mycpu->gd_cnt.v_swappgsout += bp->b_xio.xio_npages;
984263bc 1506
cdd46d2e 1507 crit_exit();
984263bc 1508
10f3fee5
MD
1509 bp->b_dirtyoff = 0; /* req'd for NFS */
1510 bp->b_dirtyend = bp->b_bcount; /* req'd for NFS */
1511 bp->b_cmd = BUF_CMD_WRITE;
8aa92e4b 1512 bio->bio_caller_info1.index = SWBIO_WRITE;
10f3fee5 1513
984263bc
MD
1514 /*
1515 * asynchronous
984263bc 1516 */
984263bc 1517 if (sync == FALSE) {
81b5c339 1518 bio->bio_done = swp_pager_async_iodone;
984263bc 1519 BUF_KERNPROC(bp);
81b5c339 1520 vn_strategy(swapdev_vp, bio);
984263bc
MD
1521
1522 for (j = 0; j < n; ++j)
1523 rtvals[i+j] = VM_PAGER_PEND;
1524 continue;
1525 }
1526
1527 /*
ae8e83e6
MD
1528 * Issue synchrnously.
1529 *
984263bc
MD
1530 * Wait for the sync I/O to complete, then update rtvals.
1531 * We just set the rtvals[] to VM_PAGER_PEND so we can call
1532 * our async completion routine at the end, thus avoiding a
1533 * double-free.
1534 */
8aa92e4b 1535 bio->bio_caller_info1.index |= SWBIO_SYNC;
ae8e83e6
MD
1536 bio->bio_done = biodone_sync;
1537 bio->bio_flags |= BIO_SYNC;
1538 vn_strategy(swapdev_vp, bio);
1539 biowait(bio, "swwrt");
984263bc
MD
1540
1541 for (j = 0; j < n; ++j)
1542 rtvals[i+j] = VM_PAGER_PEND;
1543
1544 /*
1545 * Now that we are through with the bp, we can call the
1546 * normal async completion, which frees everything up.
1547 */
81b5c339 1548 swp_pager_async_iodone(bio);
984263bc
MD
1549 }
1550}
1551
c84c24da
MD
1552void
1553swap_pager_newswap(void)
1554{
1555 swp_sizecheck();
1556}
1557
984263bc
MD
1558/*
1559 * swp_pager_async_iodone:
1560 *
1561 * Completion routine for asynchronous reads and writes from/to swap.
1562 * Also called manually by synchronous code to finish up a bp.
1563 *
1564 * For READ operations, the pages are PG_BUSY'd. For WRITE operations,
1565 * the pages are vm_page_t->busy'd. For READ operations, we PG_BUSY
1566 * unbusy all pages except the 'main' request page. For WRITE
1567 * operations, we vm_page_t->busy'd unbusy all pages ( we can do this
1568 * because we marked them all VM_PAGER_PEND on return from putpages ).
1569 *
1570 * This routine may not block.
984263bc 1571 */
984263bc 1572static void
81b5c339 1573swp_pager_async_iodone(struct bio *bio)
984263bc 1574{
81b5c339 1575 struct buf *bp = bio->bio_buf;
984263bc 1576 vm_object_t object = NULL;
81b5c339 1577 int i;
10f3fee5 1578 int *nswptr;
984263bc
MD
1579
1580 /*
1581 * report error
1582 */
984263bc 1583 if (bp->b_flags & B_ERROR) {
086c1d7e 1584 kprintf(
54078292 1585 "swap_pager: I/O error - %s failed; offset %lld,"
984263bc 1586 "size %ld, error %d\n",
8aa92e4b
MD
1587 ((bio->bio_caller_info1.index & SWBIO_READ) ?
1588 "pagein" : "pageout"),
973c11b9 1589 (long long)bio->bio_offset,
984263bc
MD
1590 (long)bp->b_bcount,
1591 bp->b_error
1592 );
1593 }
1594
1595 /*
1596 * set object, raise to splvm().
1597 */
54f51aeb
HP
1598 if (bp->b_xio.xio_npages)
1599 object = bp->b_xio.xio_pages[0]->object;
cdd46d2e 1600 crit_enter();
984263bc
MD
1601
1602 /*
1603 * remove the mapping for kernel virtual
1604 */
54f51aeb 1605 pmap_qremove((vm_offset_t)bp->b_data, bp->b_xio.xio_npages);
984263bc
MD
1606
1607 /*
1608 * cleanup pages. If an error occurs writing to swap, we are in
1609 * very serious trouble. If it happens to be a disk error, though,
1610 * we may be able to recover by reassigning the swap later on. So
1611 * in this case we remove the m->swapblk assignment for the page
1612 * but do not free it in the rlist. The errornous block(s) are thus
1613 * never reallocated as swap. Redirty the page and continue.
1614 */
54f51aeb
HP
1615 for (i = 0; i < bp->b_xio.xio_npages; ++i) {
1616 vm_page_t m = bp->b_xio.xio_pages[i];
984263bc 1617
984263bc
MD
1618 if (bp->b_flags & B_ERROR) {
1619 /*
1620 * If an error occurs I'd love to throw the swapblk
1621 * away without freeing it back to swapspace, so it
1622 * can never be used again. But I can't from an
1623 * interrupt.
1624 */
1625
8aa92e4b 1626 if (bio->bio_caller_info1.index & SWBIO_READ) {
984263bc
MD
1627 /*
1628 * When reading, reqpage needs to stay
1629 * locked for the parent, but all other
1630 * pages can be freed. We still want to
1631 * wakeup the parent waiting on the page,
1632 * though. ( also: pg_reqpage can be -1 and
1633 * not match anything ).
1634 *
1635 * We have to wake specifically requested pages
1636 * up too because we cleared PG_SWAPINPROG and
1637 * someone may be waiting for that.
1638 *
1639 * NOTE: for reads, m->dirty will probably
1640 * be overridden by the original caller of
1641 * getpages so don't play cute tricks here.
1642 *
93afe6be
MD
1643 * NOTE: We can't actually free the page from
1644 * here, because this is an interrupt. It
1645 * is not legal to mess with object->memq
1646 * from an interrupt. Deactivate the page
1647 * instead.
984263bc
MD
1648 */
1649
1650 m->valid = 0;
1651 vm_page_flag_clear(m, PG_ZERO);
5d5c5831 1652 vm_page_flag_clear(m, PG_SWAPINPROG);
984263bc 1653
81b5c339
MD
1654 /*
1655 * bio_driver_info holds the requested page
1656 * index.
1657 */
973c11b9 1658 if (i != (int)(intptr_t)bio->bio_driver_info) {
93afe6be
MD
1659 vm_page_deactivate(m);
1660 vm_page_wakeup(m);
1661 } else {
984263bc 1662 vm_page_flash(m);
93afe6be 1663 }
984263bc
MD
1664 /*
1665 * If i == bp->b_pager.pg_reqpage, do not wake
1666 * the page up. The caller needs to.
1667 */
1668 } else {
1669 /*
1670 * If a write error occurs, reactivate page
1671 * so it doesn't clog the inactive list,
1672 * then finish the I/O.
096e95c0
MD
1673 *
1674 * Only for OBJT_SWAP. When using the swap
1675 * as a cache for clean vnode-backed pages
1676 * we don't mess with the page dirty state.
984263bc 1677 */
5d5c5831 1678 vm_page_flag_clear(m, PG_SWAPINPROG);
096e95c0
MD
1679 if (m->object->type == OBJT_SWAP) {
1680 vm_page_dirty(m);
1681 vm_page_activate(m);
1682 }
984263bc
MD
1683 vm_page_io_finish(m);
1684 }
8aa92e4b 1685 } else if (bio->bio_caller_info1.index & SWBIO_READ) {
984263bc 1686 /*
984263bc
MD
1687 * NOTE: for reads, m->dirty will probably be
1688 * overridden by the original caller of getpages so
1689 * we cannot set them in order to free the underlying
1690 * swap in a low-swap situation. I don't think we'd
1691 * want to do that anyway, but it was an optimization
1692 * that existed in the old swapper for a time before
1693 * it got ripped out due to precisely this problem.
1694 *
1695 * clear PG_ZERO in page.
1696 *
1697 * If not the requested page then deactivate it.
1698 *
1699 * Note that the requested page, reqpage, is left
1700 * busied, but we still have to wake it up. The
1701 * other pages are released (unbusied) by
1702 * vm_page_wakeup(). We do not set reqpage's
1703 * valid bits here, it is up to the caller.
1704 */
1705
4530a3aa
MD
1706 /*
1707 * NOTE: can't call pmap_clear_modify(m) from an
1708 * interrupt thread, the pmap code may have to map
1709 * non-kernel pmaps and currently asserts the case.
1710 */
1711 /*pmap_clear_modify(m);*/
984263bc
MD
1712 m->valid = VM_PAGE_BITS_ALL;
1713 vm_page_undirty(m);
5d5c5831 1714 vm_page_flag_clear(m, PG_ZERO | PG_SWAPINPROG);
67803f3e 1715 vm_page_flag_set(m, PG_SWAPPED);
984263bc
MD
1716
1717 /*
1718 * We have to wake specifically requested pages
1719 * up too because we cleared PG_SWAPINPROG and
1720 * could be waiting for it in getpages. However,
1721 * be sure to not unbusy getpages specifically
1722 * requested page - getpages expects it to be
1723 * left busy.
81b5c339
MD
1724 *
1725 * bio_driver_info holds the requested page
984263bc 1726 */
973c11b9 1727 if (i != (int)(intptr_t)bio->bio_driver_info) {
984263bc
MD
1728 vm_page_deactivate(m);
1729 vm_page_wakeup(m);
1730 } else {
1731 vm_page_flash(m);
1732 }
1733 } else {
1734 /*
93afe6be
MD
1735 * Mark the page clean but do not mess with the
1736 * pmap-layer's modified state. That state should
1737 * also be clear since the caller protected the
1738 * page VM_PROT_READ, but allow the case.
1739 *
1740 * We are in an interrupt, avoid pmap operations.
1741 *
1742 * If we have a severe page deficit, deactivate the
1743 * page. Do not try to cache it (which would also
1744 * involve a pmap op), because the page might still
1745 * be read-heavy.
096e95c0
MD
1746 *
1747 * When using the swap to cache clean vnode pages
1748 * we do not mess with the page dirty bits.
984263bc 1749 */
096e95c0
MD
1750 if (m->object->type == OBJT_SWAP)
1751 vm_page_undirty(m);
5d5c5831 1752 vm_page_flag_clear(m, PG_SWAPINPROG);
67803f3e 1753 vm_page_flag_set(m, PG_SWAPPED);
984263bc 1754 vm_page_io_finish(m);
93afe6be
MD
1755 if (vm_page_count_severe())
1756 vm_page_deactivate(m);
1757#if 0
984263bc
MD
1758 if (!vm_page_count_severe() || !vm_page_try_to_cache(m))
1759 vm_page_protect(m, VM_PROT_READ);
93afe6be 1760#endif
984263bc
MD
1761 }
1762 }
1763
1764 /*
1765 * adjust pip. NOTE: the original parent may still have its own
1766 * pip refs on the object.
1767 */
1768
1769 if (object)
54f51aeb 1770 vm_object_pip_wakeupn(object, bp->b_xio.xio_npages);
984263bc
MD
1771
1772 /*
8aa92e4b
MD
1773 * Release the physical I/O buffer.
1774 *
1775 * NOTE: Due to synchronous operations in the write case b_cmd may
1776 * already be set to BUF_CMD_DONE and BIO_SYNC may have already
1777 * been cleared.
984263bc 1778 */
8aa92e4b 1779 if (bio->bio_caller_info1.index & SWBIO_READ)
10f3fee5 1780 nswptr = &nsw_rcount;
8aa92e4b 1781 else if (bio->bio_caller_info1.index & SWBIO_SYNC)
10f3fee5 1782 nswptr = &nsw_wcount_sync;
ae8e83e6
MD
1783 else
1784 nswptr = &nsw_wcount_async;
10f3fee5
MD
1785 bp->b_cmd = BUF_CMD_DONE;
1786 relpbuf(bp, nswptr);
cdd46d2e 1787 crit_exit();
984263bc
MD
1788}
1789
1790/************************************************************************
1791 * SWAP META DATA *
1792 ************************************************************************
1793 *
1794 * These routines manipulate the swap metadata stored in the
1795 * OBJT_SWAP object. All swp_*() routines must be called at
1796 * splvm() because swap can be freed up by the low level vm_page
1797 * code which might be called from interrupts beyond what splbio() covers.
1798 *
1799 * Swap metadata is implemented with a global hash and not directly
1800 * linked into the object. Instead the object simply contains
1801 * appropriate tracking counters.
1802 */
1803
1804/*
96adc753 1805 * Lookup the swblock containing the specified swap block index.
984263bc 1806 */
96adc753
MD
1807static __inline
1808struct swblock *
1809swp_pager_lookup(vm_object_t object, vm_pindex_t index)
984263bc 1810{
984263bc 1811 index &= ~SWAP_META_MASK;
96adc753 1812 return (RB_LOOKUP(swblock_rb_tree, &object->swblock_root, index));
984263bc
MD
1813}
1814
1815/*
96adc753 1816 * Remove a swblock from the RB tree.
984263bc 1817 */
96adc753
MD
1818static __inline
1819void
1820swp_pager_remove(vm_object_t object, struct swblock *swap)
1821{
1822 RB_REMOVE(swblock_rb_tree, &object->swblock_root, swap);
1823}
984263bc 1824
96adc753
MD
1825/*
1826 * Convert default object to swap object if necessary
1827 */
984263bc 1828static void
96adc753
MD
1829swp_pager_meta_convert(vm_object_t object)
1830{
1831 if (object->type == OBJT_DEFAULT) {
984263bc 1832 object->type = OBJT_SWAP;
96adc753 1833 KKASSERT(object->swblock_count == 0);
984263bc 1834 }
96adc753
MD
1835}
1836
1837/*
1838 * SWP_PAGER_META_BUILD() - add swap block to swap meta data for object
1839 *
1840 * We first convert the object to a swap object if it is a default
1841 * object. Vnode objects do not need to be converted.
1842 *
1843 * The specified swapblk is added to the object's swap metadata. If
1844 * the swapblk is not valid, it is freed instead. Any previously
1845 * assigned swapblk is freed.
1846 */
1847static void
1848swp_pager_meta_build(vm_object_t object, vm_pindex_t index, daddr_t swapblk)
1849{
1850 struct swblock *swap;
1851 struct swblock *oswap;
1852
1853 KKASSERT(swapblk != SWAPBLK_NONE);
1854
1855 /*
1856 * Convert object if necessary
1857 */
1858 if (object->type == OBJT_DEFAULT)
1859 swp_pager_meta_convert(object);
984263bc
MD
1860
1861 /*
96adc753 1862 * Locate swblock. If not found create, but if we aren't adding
984263bc
MD
1863 * anything just return. If we run out of space in the map we wait
1864 * and, since the hash table may have changed, retry.
1865 */
984263bc 1866retry:
96adc753 1867 swap = swp_pager_lookup(object, index);
984263bc 1868
96adc753 1869 if (swap == NULL) {
984263bc
MD
1870 int i;
1871
96adc753 1872 swap = zalloc(swap_zone);
984263bc 1873 if (swap == NULL) {
4ecf7cc9 1874 vm_wait(0);
984263bc
MD
1875 goto retry;
1876 }
984263bc
MD
1877 swap->swb_index = index & ~SWAP_META_MASK;
1878 swap->swb_count = 0;
1879
96adc753 1880 ++object->swblock_count;
984263bc
MD
1881
1882 for (i = 0; i < SWAP_META_PAGES; ++i)
1883 swap->swb_pages[i] = SWAPBLK_NONE;
96adc753
MD
1884 oswap = RB_INSERT(swblock_rb_tree, &object->swblock_root, swap);
1885 KKASSERT(oswap == NULL);
984263bc
MD
1886 }
1887
1888 /*
1889 * Delete prior contents of metadata
1890 */
1891
1892 index &= SWAP_META_MASK;
1893
1894 if (swap->swb_pages[index] != SWAPBLK_NONE) {
096e95c0 1895 swp_pager_freeswapspace(object, swap->swb_pages[index], 1);
984263bc
MD
1896 --swap->swb_count;
1897 }
1898
1899 /*
1900 * Enter block into metadata
1901 */
984263bc
MD
1902 swap->swb_pages[index] = swapblk;
1903 if (swapblk != SWAPBLK_NONE)
1904 ++swap->swb_count;
1905}
1906
1907/*
1908 * SWP_PAGER_META_FREE() - free a range of blocks in the object's swap metadata
1909 *
1910 * The requested range of blocks is freed, with any associated swap
1911 * returned to the swap bitmap.
1912 *
1913 * This routine will free swap metadata structures as they are cleaned
1914 * out. This routine does *NOT* operate on swap metadata associated
1915 * with resident pages.
1916 *
1917 * This routine must be called at splvm()
1918 */
8d292090
MD
1919static int swp_pager_meta_free_callback(struct swblock *swb, void *data);
1920
984263bc 1921static void
8d292090 1922swp_pager_meta_free(vm_object_t object, vm_pindex_t index, vm_pindex_t count)
984263bc 1923{
8d292090 1924 struct swfreeinfo info;
96adc753 1925
8d292090
MD
1926 /*
1927 * Nothing to do
1928 */
1929 if (object->swblock_count == 0) {
1930 KKASSERT(RB_EMPTY(&object->swblock_root));
1931 return;
1932 }
1933 if (count == 0)
984263bc
MD
1934 return;
1935
8d292090
MD
1936 /*
1937 * Setup for RB tree scan. Note that the pindex range can be huge
1938 * due to the 64 bit page index space so we cannot safely iterate.
1939 */
1940 info.object = object;
1941 info.basei = index & ~SWAP_META_MASK;
1942 info.begi = index;
1943 info.endi = index + count - 1;
1944 swblock_rb_tree_RB_SCAN(&object->swblock_root, rb_swblock_scancmp,
1945 swp_pager_meta_free_callback, &info);
1946}
984263bc 1947
8d292090
MD
1948static
1949int
1950swp_pager_meta_free_callback(struct swblock *swap, void *data)
1951{
1952 struct swfreeinfo *info = data;
1953 vm_object_t object = info->object;
1954 int index;
1955 int eindex;
1956
1957 /*
1958 * Figure out the range within the swblock. The wider scan may
1959 * return edge-case swap blocks when the start and/or end points
1960 * are in the middle of a block.
1961 */
1962 if (swap->swb_index < info->begi)
1963 index = (int)info->begi & SWAP_META_MASK;
1964 else
1965 index = 0;
1966
1967 if (swap->swb_index + SWAP_META_PAGES > info->endi)
1968 eindex = (int)info->endi & SWAP_META_MASK;
1969 else
1970 eindex = SWAP_META_MASK;
1971
1972 /*
1973 * Scan and free the blocks. The loop terminates early
1974 * if (swap) runs out of blocks and could be freed.
1975 */
1976 while (index <= eindex) {
1977 daddr_t v = swap->swb_pages[index];
1978
1979 if (v != SWAPBLK_NONE) {
096e95c0 1980 swp_pager_freeswapspace(object, v, 1);
8d292090
MD
1981 swap->swb_pages[index] = SWAPBLK_NONE;
1982 if (--swap->swb_count == 0) {
1983 swp_pager_remove(object, swap);
1984 zfree(swap_zone, swap);
1985 --object->swblock_count;
1986 break;
984263bc 1987 }
984263bc 1988 }
8d292090 1989 ++index;
984263bc 1990 }
8d292090
MD
1991 /* swap may be invalid here due to zfree above */
1992 return(0);
984263bc
MD
1993}
1994
1995/*
1996 * SWP_PAGER_META_FREE_ALL() - destroy all swap metadata associated with object
1997 *
1998 * This routine locates and destroys all swap metadata associated with
1999 * an object.
2000 *
2001 * This routine must be called at splvm()
2002 */
984263bc
MD
2003static void
2004swp_pager_meta_free_all(vm_object_t object)
2005{
96adc753
MD
2006 struct swblock *swap;
2007 int i;
984263bc 2008
96adc753
MD
2009 while ((swap = RB_ROOT(&object->swblock_root)) != NULL) {
2010 swp_pager_remove(object, swap);
2011 for (i = 0; i < SWAP_META_PAGES; ++i) {
2012 daddr_t v = swap->swb_pages[i];
2013 if (v != SWAPBLK_NONE) {
2014 --swap->swb_count;
096e95c0 2015 swp_pager_freeswapspace(object, v, 1);
984263bc 2016 }
984263bc 2017 }
96adc753
MD
2018 if (swap->swb_count != 0)
2019 panic("swap_pager_meta_free_all: swb_count != 0");
2020 zfree(swap_zone, swap);
2021 --object->swblock_count;
984263bc 2022 }
96adc753 2023 KKASSERT(object->swblock_count == 0);
984263bc
MD
2024}
2025
2026/*
2027 * SWP_PAGER_METACTL() - misc control of swap and vm_page_t meta data.
2028 *
2029 * This routine is capable of looking up, popping, or freeing
2030 * swapblk assignments in the swap meta data or in the vm_page_t.
2031 * The routine typically returns the swapblk being looked-up, or popped,
2032 * or SWAPBLK_NONE if the block was freed, or SWAPBLK_NONE if the block
2033 * was invalid. This routine will automatically free any invalid
2034 * meta-data swapblks.
2035 *
2036 * It is not possible to store invalid swapblks in the swap meta data
2037 * (other then a literal 'SWAPBLK_NONE'), so we don't bother checking.
2038 *
2039 * When acting on a busy resident page and paging is in progress, we
2040 * have to wait until paging is complete but otherwise can act on the
2041 * busy page.
2042 *
2043 * This routine must be called at splvm().
2044 *
2045 * SWM_FREE remove and free swap block from metadata
2046 * SWM_POP remove from meta data but do not free.. pop it out
2047 */
984263bc 2048static daddr_t
96adc753
MD
2049swp_pager_meta_ctl(vm_object_t object, vm_pindex_t index, int flags)
2050{
984263bc
MD
2051 struct swblock *swap;
2052 daddr_t r1;
2053
8d292090 2054 if (object->swblock_count == 0)
984263bc
MD
2055 return(SWAPBLK_NONE);
2056
2057 r1 = SWAPBLK_NONE;
96adc753 2058 swap = swp_pager_lookup(object, index);
984263bc 2059
96adc753 2060 if (swap != NULL) {
984263bc
MD
2061 index &= SWAP_META_MASK;
2062 r1 = swap->swb_pages[index];
2063
2064 if (r1 != SWAPBLK_NONE) {
2065 if (flags & SWM_FREE) {
096e95c0 2066 swp_pager_freeswapspace(object, r1, 1);
984263bc
MD
2067 r1 = SWAPBLK_NONE;
2068 }
2069 if (flags & (SWM_FREE|SWM_POP)) {
2070 swap->swb_pages[index] = SWAPBLK_NONE;
2071 if (--swap->swb_count == 0) {
96adc753 2072 swp_pager_remove(object, swap);
984263bc 2073 zfree(swap_zone, swap);
96adc753 2074 --object->swblock_count;
984263bc
MD
2075 }
2076 }
2077 }
2078 }
2079 return(r1);
2080}