kernel - Fix missing lock in pbuf update (minor)
[dragonfly.git] / sys / vm / swap_pager.c
CommitLineData
984263bc 1/*
8e7c4729
MD
2 * (MPSAFE)
3 *
4 * Copyright (c) 1998-2010 The DragonFly Project. All rights reserved.
8c10bfcf
MD
5 *
6 * This code is derived from software contributed to The DragonFly Project
7 * by Matthew Dillon <dillon@backplane.com>
8 *
9 * Redistribution and use in source and binary forms, with or without
10 * modification, are permitted provided that the following conditions
11 * are met:
12 *
13 * 1. Redistributions of source code must retain the above copyright
14 * notice, this list of conditions and the following disclaimer.
15 * 2. Redistributions in binary form must reproduce the above copyright
16 * notice, this list of conditions and the following disclaimer in
17 * the documentation and/or other materials provided with the
18 * distribution.
19 * 3. Neither the name of The DragonFly Project nor the names of its
20 * contributors may be used to endorse or promote products derived
21 * from this software without specific, prior written permission.
22 *
23 * THIS SOFTWARE IS PROVIDED BY THE COPYRIGHT HOLDERS AND CONTRIBUTORS
24 * ``AS IS'' AND ANY EXPRESS OR IMPLIED WARRANTIES, INCLUDING, BUT NOT
25 * LIMITED TO, THE IMPLIED WARRANTIES OF MERCHANTABILITY AND FITNESS
26 * FOR A PARTICULAR PURPOSE ARE DISCLAIMED. IN NO EVENT SHALL THE
27 * COPYRIGHT HOLDERS OR CONTRIBUTORS BE LIABLE FOR ANY DIRECT, INDIRECT,
28 * INCIDENTAL, SPECIAL, EXEMPLARY OR CONSEQUENTIAL DAMAGES (INCLUDING,
29 * BUT NOT LIMITED TO, PROCUREMENT OF SUBSTITUTE GOODS OR SERVICES;
30 * LOSS OF USE, DATA, OR PROFITS; OR BUSINESS INTERRUPTION) HOWEVER CAUSED
31 * AND ON ANY THEORY OF LIABILITY, WHETHER IN CONTRACT, STRICT LIABILITY,
32 * OR TORT (INCLUDING NEGLIGENCE OR OTHERWISE) ARISING IN ANY WAY OUT
33 * OF THE USE OF THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF
34 * SUCH DAMAGE.
35 *
984263bc
MD
36 * Copyright (c) 1994 John S. Dyson
37 * Copyright (c) 1990 University of Utah.
38 * Copyright (c) 1991, 1993
39 * The Regents of the University of California. All rights reserved.
40 *
41 * This code is derived from software contributed to Berkeley by
42 * the Systems Programming Group of the University of Utah Computer
43 * Science Department.
44 *
45 * Redistribution and use in source and binary forms, with or without
46 * modification, are permitted provided that the following conditions
47 * are met:
48 * 1. Redistributions of source code must retain the above copyright
49 * notice, this list of conditions and the following disclaimer.
50 * 2. Redistributions in binary form must reproduce the above copyright
51 * notice, this list of conditions and the following disclaimer in the
52 * documentation and/or other materials provided with the distribution.
53 * 3. All advertising materials mentioning features or use of this software
54 * must display the following acknowledgement:
55 * This product includes software developed by the University of
56 * California, Berkeley and its contributors.
57 * 4. Neither the name of the University nor the names of its contributors
58 * may be used to endorse or promote products derived from this software
59 * without specific prior written permission.
60 *
61 * THIS SOFTWARE IS PROVIDED BY THE REGENTS AND CONTRIBUTORS ``AS IS'' AND
62 * ANY EXPRESS OR IMPLIED WARRANTIES, INCLUDING, BUT NOT LIMITED TO, THE
63 * IMPLIED WARRANTIES OF MERCHANTABILITY AND FITNESS FOR A PARTICULAR PURPOSE
64 * ARE DISCLAIMED. IN NO EVENT SHALL THE REGENTS OR CONTRIBUTORS BE LIABLE
65 * FOR ANY DIRECT, INDIRECT, INCIDENTAL, SPECIAL, EXEMPLARY, OR CONSEQUENTIAL
66 * DAMAGES (INCLUDING, BUT NOT LIMITED TO, PROCUREMENT OF SUBSTITUTE GOODS
67 * OR SERVICES; LOSS OF USE, DATA, OR PROFITS; OR BUSINESS INTERRUPTION)
68 * HOWEVER CAUSED AND ON ANY THEORY OF LIABILITY, WHETHER IN CONTRACT, STRICT
69 * LIABILITY, OR TORT (INCLUDING NEGLIGENCE OR OTHERWISE) ARISING IN ANY WAY
70 * OUT OF THE USE OF THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF
71 * SUCH DAMAGE.
72 *
73 * New Swap System
74 * Matthew Dillon
75 *
76 * Radix Bitmap 'blists'.
77 *
78 * - The new swapper uses the new radix bitmap code. This should scale
79 * to arbitrarily small or arbitrarily large swap spaces and an almost
80 * arbitrary degree of fragmentation.
81 *
82 * Features:
83 *
84 * - on the fly reallocation of swap during putpages. The new system
85 * does not try to keep previously allocated swap blocks for dirty
86 * pages.
87 *
88 * - on the fly deallocation of swap
89 *
90 * - No more garbage collection required. Unnecessarily allocated swap
91 * blocks only exist for dirty vm_page_t's now and these are already
92 * cycled (in a high-load system) by the pager. We also do on-the-fly
93 * removal of invalidated swap blocks when a page is destroyed
94 * or renamed.
95 *
96 * from: Utah $Hdr: swap_pager.c 1.4 91/04/30$
8e7c4729 97 * @(#)swap_pager.c 8.9 (Berkeley) 3/21/94
984263bc
MD
98 * $FreeBSD: src/sys/vm/swap_pager.c,v 1.130.2.12 2002/08/31 21:15:55 dillon Exp $
99 */
100
101#include <sys/param.h>
102#include <sys/systm.h>
103#include <sys/conf.h>
104#include <sys/kernel.h>
105#include <sys/proc.h>
106#include <sys/buf.h>
107#include <sys/vnode.h>
108#include <sys/malloc.h>
109#include <sys/vmmeter.h>
110#include <sys/sysctl.h>
111#include <sys/blist.h>
112#include <sys/lock.h>
cdd46d2e 113#include <sys/thread2.h>
984263bc
MD
114
115#ifndef MAX_PAGEOUT_CLUSTER
116#define MAX_PAGEOUT_CLUSTER 16
117#endif
118
119#define SWB_NPAGES MAX_PAGEOUT_CLUSTER
120
121#include "opt_swap.h"
122#include <vm/vm.h>
123#include <vm/vm_object.h>
124#include <vm/vm_page.h>
125#include <vm/vm_pager.h>
126#include <vm/vm_pageout.h>
127#include <vm/swap_pager.h>
128#include <vm/vm_extern.h>
129#include <vm/vm_zone.h>
5d5c5831 130#include <vm/vnode_pager.h>
984263bc 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
8d292090
MD
142struct swfreeinfo {
143 vm_object_t object;
144 vm_pindex_t basei;
145 vm_pindex_t begi;
146 vm_pindex_t endi; /* inclusive */
147};
148
984263bc
MD
149/*
150 * vm_swap_size is in page-sized chunks now. It was DEV_BSIZE'd chunks
151 * in the old system.
152 */
153
984263bc 154int swap_pager_full; /* swap space exhaustion (task killing) */
096e95c0
MD
155int vm_swap_cache_use;
156int vm_swap_anon_use;
157
984263bc
MD
158static int swap_pager_almost_full; /* swap space exhaustion (w/ hysteresis)*/
159static int nsw_rcount; /* free read buffers */
160static int nsw_wcount_sync; /* limit write buffers / synchronous */
161static int nsw_wcount_async; /* limit write buffers / asynchronous */
162static int nsw_wcount_async_max;/* assigned maximum */
163static int nsw_cluster_max; /* maximum VOP I/O allowed */
984263bc
MD
164
165struct blist *swapblist;
984263bc 166static int swap_async_max = 4; /* maximum in-progress async I/O's */
5d5c5831 167static int swap_burst_read = 0; /* allow burst reading */
984263bc 168
9f3543c6
MD
169/* from vm_swap.c */
170extern struct vnode *swapdev_vp;
171extern struct swdevt *swdevt;
172extern int nswdev;
173
174#define BLK2DEVIDX(blk) (nswdev > 1 ? blk / dmmax % nswdev : 0)
984263bc
MD
175
176SYSCTL_INT(_vm, OID_AUTO, swap_async_max,
177 CTLFLAG_RW, &swap_async_max, 0, "Maximum running async swap ops");
5d5c5831
MD
178SYSCTL_INT(_vm, OID_AUTO, swap_burst_read,
179 CTLFLAG_RW, &swap_burst_read, 0, "Allow burst reads for pageins");
984263bc 180
096e95c0
MD
181SYSCTL_INT(_vm, OID_AUTO, swap_cache_use,
182 CTLFLAG_RD, &vm_swap_cache_use, 0, "");
183SYSCTL_INT(_vm, OID_AUTO, swap_anon_use,
184 CTLFLAG_RD, &vm_swap_anon_use, 0, "");
97a3ace7
MD
185SYSCTL_INT(_vm, OID_AUTO, swap_size,
186 CTLFLAG_RD, &vm_swap_size, 0, "");
096e95c0 187
984263bc
MD
188vm_zone_t swap_zone;
189
190/*
96adc753 191 * Red-Black tree for swblock entries
8e7c4729
MD
192 *
193 * The caller must hold vm_token
96adc753
MD
194 */
195RB_GENERATE2(swblock_rb_tree, swblock, swb_entry, rb_swblock_compare,
196 vm_pindex_t, swb_index);
197
198int
199rb_swblock_compare(struct swblock *swb1, struct swblock *swb2)
200{
201 if (swb1->swb_index < swb2->swb_index)
202 return(-1);
203 if (swb1->swb_index > swb2->swb_index)
204 return(1);
205 return(0);
206}
207
8d292090
MD
208static
209int
210rb_swblock_scancmp(struct swblock *swb, void *data)
211{
212 struct swfreeinfo *info = data;
213
214 if (swb->swb_index < info->basei)
215 return(-1);
216 if (swb->swb_index > info->endi)
217 return(1);
218 return(0);
219}
220
00a3fdca
MD
221static
222int
223rb_swblock_condcmp(struct swblock *swb, void *data)
224{
225 struct swfreeinfo *info = data;
226
227 if (swb->swb_index < info->basei)
228 return(-1);
229 return(0);
230}
231
96adc753 232/*
984263bc
MD
233 * pagerops for OBJT_SWAP - "swap pager". Some ops are also global procedure
234 * calls hooked from other parts of the VM system and do not appear here.
235 * (see vm/swap_pager.h).
236 */
237
1388df65 238static void swap_pager_dealloc (vm_object_t object);
1b9d3514 239static int swap_pager_getpage (vm_object_t, vm_page_t *, int);
81b5c339 240static void swap_chain_iodone(struct bio *biox);
984263bc
MD
241
242struct pagerops swappagerops = {
984263bc 243 swap_pager_dealloc, /* deallocate an OBJT_SWAP object */
1b9d3514 244 swap_pager_getpage, /* pagein */
984263bc 245 swap_pager_putpages, /* pageout */
107e9bcc 246 swap_pager_haspage /* get backing store status for page */
984263bc
MD
247};
248
249/*
250 * dmmax is in page-sized chunks with the new swap system. It was
251 * dev-bsized chunks in the old. dmmax is always a power of 2.
252 *
253 * swap_*() routines are externally accessible. swp_*() routines are
254 * internal.
255 */
256
257int dmmax;
258static int dmmax_mask;
259int nswap_lowat = 128; /* in pages, swap_pager_almost_full warn */
260int nswap_hiwat = 512; /* in pages, swap_pager_almost_full warn */
261
1388df65 262static __inline void swp_sizecheck (void);
81b5c339 263static void swp_pager_async_iodone (struct bio *bio);
984263bc
MD
264
265/*
266 * Swap bitmap functions
267 */
268
8e7c4729 269static __inline void swp_pager_freeswapspace(vm_object_t object,
651d8e75
MD
270 swblk_t blk, int npages);
271static __inline swblk_t swp_pager_getswapspace(vm_object_t object, int npages);
984263bc
MD
272
273/*
274 * Metadata functions
275 */
276
8e7c4729 277static void swp_pager_meta_convert(vm_object_t);
651d8e75 278static void swp_pager_meta_build(vm_object_t, vm_pindex_t, swblk_t);
8e7c4729
MD
279static void swp_pager_meta_free(vm_object_t, vm_pindex_t, vm_pindex_t);
280static void swp_pager_meta_free_all(vm_object_t);
651d8e75 281static swblk_t swp_pager_meta_ctl(vm_object_t, vm_pindex_t, int);
984263bc
MD
282
283/*
284 * SWP_SIZECHECK() - update swap_pager_full indication
285 *
286 * update the swap_pager_almost_full indication and warn when we are
287 * about to run out of swap space, using lowat/hiwat hysteresis.
288 *
289 * Clear swap_pager_full ( task killing ) indication when lowat is met.
290 *
8e7c4729
MD
291 * No restrictions on call
292 * This routine may not block.
293 * SMP races are ok.
984263bc 294 */
984263bc 295static __inline void
57e43348 296swp_sizecheck(void)
984263bc
MD
297{
298 if (vm_swap_size < nswap_lowat) {
299 if (swap_pager_almost_full == 0) {
086c1d7e 300 kprintf("swap_pager: out of swap space\n");
984263bc
MD
301 swap_pager_almost_full = 1;
302 }
303 } else {
304 swap_pager_full = 0;
305 if (vm_swap_size > nswap_hiwat)
306 swap_pager_almost_full = 0;
307 }
308}
309
310/*
311 * SWAP_PAGER_INIT() - initialize the swap pager!
312 *
313 * Expected to be started from system init. NOTE: This code is run
314 * before much else so be careful what you depend on. Most of the VM
315 * system has yet to be initialized at this point.
8e7c4729
MD
316 *
317 * Called from the low level boot code only.
984263bc 318 */
984263bc 319static void
107e9bcc 320swap_pager_init(void *arg __unused)
984263bc
MD
321{
322 /*
984263bc
MD
323 * Device Stripe, in PAGE_SIZE'd blocks
324 */
984263bc
MD
325 dmmax = SWB_NPAGES * 2;
326 dmmax_mask = ~(dmmax - 1);
327}
107e9bcc 328SYSINIT(vm_mem, SI_BOOT1_VM, SI_ORDER_THIRD, swap_pager_init, NULL)
984263bc
MD
329
330/*
331 * SWAP_PAGER_SWAP_INIT() - swap pager initialization from pageout process
332 *
333 * Expected to be started from pageout process once, prior to entering
334 * its main loop.
8e7c4729
MD
335 *
336 * Called from the low level boot code only.
984263bc 337 */
984263bc 338void
57e43348 339swap_pager_swap_init(void)
984263bc
MD
340{
341 int n, n2;
342
343 /*
344 * Number of in-transit swap bp operations. Don't
345 * exhaust the pbufs completely. Make sure we
346 * initialize workable values (0 will work for hysteresis
347 * but it isn't very efficient).
348 *
54f51aeb
HP
349 * The nsw_cluster_max is constrained by the number of pages an XIO
350 * holds, i.e., (MAXPHYS/PAGE_SIZE) and our locally defined
984263bc
MD
351 * MAX_PAGEOUT_CLUSTER. Also be aware that swap ops are
352 * constrained by the swap device interleave stripe size.
353 *
354 * Currently we hardwire nsw_wcount_async to 4. This limit is
355 * designed to prevent other I/O from having high latencies due to
356 * our pageout I/O. The value 4 works well for one or two active swap
357 * devices but is probably a little low if you have more. Even so,
358 * a higher value would probably generate only a limited improvement
359 * with three or four active swap devices since the system does not
360 * typically have to pageout at extreme bandwidths. We will want
361 * at least 2 per swap devices, and 4 is a pretty good value if you
362 * have one NFS swap device due to the command/ack latency over NFS.
363 * So it all works out pretty well.
364 */
365
366 nsw_cluster_max = min((MAXPHYS/PAGE_SIZE), MAX_PAGEOUT_CLUSTER);
367
368 nsw_rcount = (nswbuf + 1) / 2;
369 nsw_wcount_sync = (nswbuf + 3) / 4;
370 nsw_wcount_async = 4;
371 nsw_wcount_async_max = nsw_wcount_async;
372
373 /*
79634a66
MD
374 * The zone is dynamically allocated so generally size it to
375 * maxswzone (32MB to 512MB of KVM). Set a minimum size based
376 * on physical memory of around 8x (each swblock can hold 16 pages).
377 *
378 * With the advent of SSDs (vs HDs) the practical (swap:memory) ratio
379 * has increased dramatically.
984263bc 380 */
12e4aaff 381 n = vmstats.v_page_count / 2;
79634a66 382 if (maxswzone && n < maxswzone / sizeof(struct swblock))
984263bc
MD
383 n = maxswzone / sizeof(struct swblock);
384 n2 = n;
385
386 do {
387 swap_zone = zinit(
388 "SWAPMETA",
389 sizeof(struct swblock),
390 n,
391 ZONE_INTERRUPT,
392 1);
393 if (swap_zone != NULL)
394 break;
395 /*
396 * if the allocation failed, try a zone two thirds the
397 * size of the previous attempt.
398 */
399 n -= ((n + 2) / 3);
400 } while (n > 0);
401
402 if (swap_zone == NULL)
403 panic("swap_pager_swap_init: swap_zone == NULL");
404 if (n2 != n)
086c1d7e 405 kprintf("Swap zone entries reduced from %d to %d.\n", n2, n);
984263bc
MD
406}
407
408/*
409 * SWAP_PAGER_ALLOC() - allocate a new OBJT_SWAP VM object and instantiate
410 * its metadata structures.
411 *
412 * This routine is called from the mmap and fork code to create a new
413 * OBJT_SWAP object. We do this by creating an OBJT_DEFAULT object
96adc753 414 * and then converting it with swp_pager_meta_convert().
984263bc 415 *
8e7c4729
MD
416 * We only support unnamed objects.
417 *
418 * No restrictions.
984263bc 419 */
5a648714 420vm_object_t
57f7b636 421swap_pager_alloc(void *handle, off_t size, vm_prot_t prot, off_t offset)
984263bc
MD
422{
423 vm_object_t object;
424
d28e1355 425 KKASSERT(handle == NULL);
a2ee730d
MD
426 object = vm_object_allocate_hold(OBJT_DEFAULT,
427 OFF_TO_IDX(offset + PAGE_MASK + size));
d28e1355 428 swp_pager_meta_convert(object);
b12defdc 429 vm_object_drop(object);
984263bc
MD
430
431 return (object);
432}
433
434/*
435 * SWAP_PAGER_DEALLOC() - remove swap metadata from object
436 *
437 * The swap backing for the object is destroyed. The code is
438 * designed such that we can reinstantiate it later, but this
439 * routine is typically called only when the entire object is
440 * about to be destroyed.
441 *
8e7c4729
MD
442 * The object must be locked or unreferenceable.
443 * No other requirements.
984263bc 444 */
984263bc 445static void
57e43348 446swap_pager_dealloc(vm_object_t object)
984263bc 447{
b12defdc 448 vm_object_hold(object);
984263bc
MD
449 vm_object_pip_wait(object, "swpdea");
450
451 /*
452 * Free all remaining metadata. We only bother to free it from
453 * the swap meta data. We do not attempt to free swapblk's still
454 * associated with vm_page_t's for this object. We do not care
455 * if paging is still in progress on some objects.
456 */
984263bc 457 swp_pager_meta_free_all(object);
b12defdc 458 vm_object_drop(object);
984263bc
MD
459}
460
461/************************************************************************
462 * SWAP PAGER BITMAP ROUTINES *
463 ************************************************************************/
464
465/*
466 * SWP_PAGER_GETSWAPSPACE() - allocate raw swap space
467 *
468 * Allocate swap for the requested number of pages. The starting
469 * swap block number (a page index) is returned or SWAPBLK_NONE
470 * if the allocation failed.
471 *
472 * Also has the side effect of advising that somebody made a mistake
473 * when they configured swap and didn't configure enough.
474 *
b12defdc 475 * The caller must hold the object.
8e7c4729 476 * This routine may not block.
984263bc 477 */
651d8e75 478static __inline swblk_t
096e95c0 479swp_pager_getswapspace(vm_object_t object, int npages)
984263bc 480{
651d8e75 481 swblk_t blk;
984263bc 482
b12defdc 483 lwkt_gettoken(&vm_token);
984263bc
MD
484 if ((blk = blist_alloc(swapblist, npages)) == SWAPBLK_NONE) {
485 if (swap_pager_full != 2) {
b5e19bf6
MD
486 kprintf("swap_pager_getswapspace: failed alloc=%d\n",
487 npages);
984263bc
MD
488 swap_pager_full = 2;
489 swap_pager_almost_full = 1;
490 }
491 } else {
099f3e5e 492 swapacctspace(blk, -npages);
096e95c0
MD
493 if (object->type == OBJT_SWAP)
494 vm_swap_anon_use += npages;
495 else
496 vm_swap_cache_use += npages;
984263bc
MD
497 swp_sizecheck();
498 }
b12defdc 499 lwkt_reltoken(&vm_token);
984263bc
MD
500 return(blk);
501}
502
503/*
504 * SWP_PAGER_FREESWAPSPACE() - free raw swap space
505 *
506 * This routine returns the specified swap blocks back to the bitmap.
507 *
508 * Note: This routine may not block (it could in the old swap code),
509 * and through the use of the new blist routines it does not block.
510 *
511 * We must be called at splvm() to avoid races with bitmap frees from
512 * vm_page_remove() aka swap_pager_page_removed().
513 *
8e7c4729 514 * This routine may not block.
984263bc
MD
515 */
516
517static __inline void
651d8e75 518swp_pager_freeswapspace(vm_object_t object, swblk_t blk, int npages)
984263bc 519{
9f3543c6
MD
520 struct swdevt *sp = &swdevt[BLK2DEVIDX(blk)];
521
b12defdc 522 lwkt_gettoken(&vm_token);
9f3543c6 523 sp->sw_nused -= npages;
096e95c0
MD
524 if (object->type == OBJT_SWAP)
525 vm_swap_anon_use -= npages;
526 else
527 vm_swap_cache_use -= npages;
9f3543c6 528
b12defdc
MD
529 if (sp->sw_flags & SW_CLOSING) {
530 lwkt_reltoken(&vm_token);
9f3543c6 531 return;
b12defdc 532 }
9f3543c6
MD
533
534 blist_free(swapblist, blk, npages);
535 vm_swap_size += npages;
984263bc 536 swp_sizecheck();
b12defdc 537 lwkt_reltoken(&vm_token);
984263bc
MD
538}
539
540/*
541 * SWAP_PAGER_FREESPACE() - frees swap blocks associated with a page
542 * range within an object.
543 *
544 * This is a globally accessible routine.
545 *
546 * This routine removes swapblk assignments from swap metadata.
547 *
548 * The external callers of this routine typically have already destroyed
549 * or renamed vm_page_t's associated with this range in the object so
550 * we should be ok.
551 *
8e7c4729 552 * No requirements.
984263bc 553 */
984263bc 554void
8d292090 555swap_pager_freespace(vm_object_t object, vm_pindex_t start, vm_pindex_t size)
984263bc 556{
b12defdc 557 vm_object_hold(object);
984263bc 558 swp_pager_meta_free(object, start, size);
b12defdc 559 vm_object_drop(object);
984263bc
MD
560}
561
8e7c4729
MD
562/*
563 * No requirements.
564 */
8d292090
MD
565void
566swap_pager_freespace_all(vm_object_t object)
567{
b12defdc 568 vm_object_hold(object);
8d292090 569 swp_pager_meta_free_all(object);
b12defdc 570 vm_object_drop(object);
8d292090
MD
571}
572
984263bc 573/*
00a3fdca
MD
574 * This function conditionally frees swap cache swap starting at
575 * (*basei) in the object. (count) swap blocks will be nominally freed.
576 * The actual number of blocks freed can be more or less than the
577 * requested number.
578 *
579 * This function nominally returns the number of blocks freed. However,
580 * the actual number of blocks freed may be less then the returned value.
581 * If the function is unable to exhaust the object or if it is able to
582 * free (approximately) the requested number of blocks it returns
583 * a value n > count.
584 *
585 * If we exhaust the object we will return a value n <= count.
135b4b20 586 *
b12defdc 587 * The caller must hold the object.
34542daf
MD
588 *
589 * WARNING! If count == 0 then -1 can be returned as a degenerate case,
590 * callers should always pass a count value > 0.
00a3fdca
MD
591 */
592static int swap_pager_condfree_callback(struct swblock *swap, void *data);
593
594int
aecf2182 595swap_pager_condfree(vm_object_t object, vm_pindex_t *basei, int count)
00a3fdca
MD
596{
597 struct swfreeinfo info;
08fb7a9d
MD
598 int n;
599 int t;
00a3fdca 600
b12defdc 601 ASSERT_LWKT_TOKEN_HELD(vm_object_token(object));
8e7c4729 602
00a3fdca
MD
603 info.object = object;
604 info.basei = *basei; /* skip up to this page index */
605 info.begi = count; /* max swap pages to destroy */
606 info.endi = count * 8; /* max swblocks to scan */
607
608 swblock_rb_tree_RB_SCAN(&object->swblock_root, rb_swblock_condcmp,
609 swap_pager_condfree_callback, &info);
610 *basei = info.basei;
08fb7a9d
MD
611
612 /*
613 * Take the higher difference swblocks vs pages
614 */
615 n = count - (int)info.begi;
616 t = count * 8 - (int)info.endi;
617 if (n < t)
618 n = t;
619 if (n < 1)
620 n = 1;
621 return(n);
00a3fdca
MD
622}
623
624/*
625 * The idea is to free whole meta-block to avoid fragmenting
626 * the swap space or disk I/O. We only do this if NO VM pages
627 * are present.
628 *
629 * We do not have to deal with clearing PG_SWAPPED in related VM
630 * pages because there are no related VM pages.
8e7c4729 631 *
b12defdc 632 * The caller must hold the object.
00a3fdca
MD
633 */
634static int
635swap_pager_condfree_callback(struct swblock *swap, void *data)
636{
637 struct swfreeinfo *info = data;
638 vm_object_t object = info->object;
639 int i;
640
641 for (i = 0; i < SWAP_META_PAGES; ++i) {
642 if (vm_page_lookup(object, swap->swb_index + i))
643 break;
644 }
645 info->basei = swap->swb_index + SWAP_META_PAGES;
646 if (i == SWAP_META_PAGES) {
647 info->begi -= swap->swb_count;
648 swap_pager_freespace(object, swap->swb_index, SWAP_META_PAGES);
649 }
650 --info->endi;
651 if ((int)info->begi < 0 || (int)info->endi < 0)
652 return(-1);
d2d8515b 653 lwkt_yield();
00a3fdca
MD
654 return(0);
655}
656
657/*
096e95c0
MD
658 * Called by vm_page_alloc() when a new VM page is inserted
659 * into a VM object. Checks whether swap has been assigned to
660 * the page and sets PG_SWAPPED as necessary.
8e7c4729
MD
661 *
662 * No requirements.
096e95c0
MD
663 */
664void
665swap_pager_page_inserted(vm_page_t m)
666{
667 if (m->object->swblock_count) {
b12defdc 668 vm_object_hold(m->object);
096e95c0
MD
669 if (swp_pager_meta_ctl(m->object, m->pindex, 0) != SWAPBLK_NONE)
670 vm_page_flag_set(m, PG_SWAPPED);
b12defdc 671 vm_object_drop(m->object);
096e95c0
MD
672 }
673}
674
675/*
984263bc
MD
676 * SWAP_PAGER_RESERVE() - reserve swap blocks in object
677 *
678 * Assigns swap blocks to the specified range within the object. The
679 * swap blocks are not zerod. Any previous swap assignment is destroyed.
680 *
681 * Returns 0 on success, -1 on failure.
8e7c4729
MD
682 *
683 * The caller is responsible for avoiding races in the specified range.
684 * No other requirements.
984263bc 685 */
984263bc
MD
686int
687swap_pager_reserve(vm_object_t object, vm_pindex_t start, vm_size_t size)
688{
984263bc 689 int n = 0;
651d8e75 690 swblk_t blk = SWAPBLK_NONE;
984263bc
MD
691 vm_pindex_t beg = start; /* save start index */
692
b12defdc
MD
693 vm_object_hold(object);
694
984263bc
MD
695 while (size) {
696 if (n == 0) {
697 n = BLIST_MAX_ALLOC;
096e95c0
MD
698 while ((blk = swp_pager_getswapspace(object, n)) ==
699 SWAPBLK_NONE)
700 {
984263bc
MD
701 n >>= 1;
702 if (n == 0) {
8d292090
MD
703 swp_pager_meta_free(object, beg,
704 start - beg);
b12defdc 705 vm_object_drop(object);
984263bc
MD
706 return(-1);
707 }
708 }
709 }
710 swp_pager_meta_build(object, start, blk);
711 --size;
712 ++start;
713 ++blk;
714 --n;
715 }
716 swp_pager_meta_free(object, start, n);
b12defdc 717 vm_object_drop(object);
984263bc
MD
718 return(0);
719}
720
721/*
722 * SWAP_PAGER_COPY() - copy blocks from source pager to destination pager
723 * and destroy the source.
724 *
725 * Copy any valid swapblks from the source to the destination. In
726 * cases where both the source and destination have a valid swapblk,
727 * we keep the destination's.
728 *
729 * This routine is allowed to block. It may block allocating metadata
730 * indirectly through swp_pager_meta_build() or if paging is still in
731 * progress on the source.
732 *
984263bc
MD
733 * XXX vm_page_collapse() kinda expects us not to block because we
734 * supposedly do not need to allocate memory, but for the moment we
735 * *may* have to get a little memory from the zone allocator, but
736 * it is taken from the interrupt memory. We should be ok.
737 *
738 * The source object contains no vm_page_t's (which is just as well)
984263bc
MD
739 * The source object is of type OBJT_SWAP.
740 *
b12defdc 741 * The source and destination objects must be held by the caller.
984263bc 742 */
984263bc 743void
57e43348 744swap_pager_copy(vm_object_t srcobject, vm_object_t dstobject,
8d292090 745 vm_pindex_t base_index, int destroysource)
984263bc
MD
746{
747 vm_pindex_t i;
984263bc 748
b12defdc
MD
749 ASSERT_LWKT_TOKEN_HELD(vm_object_token(srcobject));
750 ASSERT_LWKT_TOKEN_HELD(vm_object_token(dstobject));
984263bc
MD
751
752 /*
984263bc
MD
753 * transfer source to destination.
754 */
984263bc 755 for (i = 0; i < dstobject->size; ++i) {
651d8e75 756 swblk_t dstaddr;
984263bc
MD
757
758 /*
759 * Locate (without changing) the swapblk on the destination,
760 * unless it is invalid in which case free it silently, or
761 * if the destination is a resident page, in which case the
762 * source is thrown away.
763 */
984263bc
MD
764 dstaddr = swp_pager_meta_ctl(dstobject, i, 0);
765
766 if (dstaddr == SWAPBLK_NONE) {
767 /*
768 * Destination has no swapblk and is not resident,
769 * copy source.
770 */
651d8e75 771 swblk_t srcaddr;
984263bc 772
8d292090
MD
773 srcaddr = swp_pager_meta_ctl(srcobject,
774 base_index + i, SWM_POP);
984263bc
MD
775
776 if (srcaddr != SWAPBLK_NONE)
777 swp_pager_meta_build(dstobject, i, srcaddr);
778 } else {
779 /*
780 * Destination has valid swapblk or it is represented
781 * by a resident page. We destroy the sourceblock.
782 */
8d292090 783 swp_pager_meta_ctl(srcobject, base_index + i, SWM_FREE);
984263bc
MD
784 }
785 }
786
787 /*
788 * Free left over swap blocks in source.
789 *
790 * We have to revert the type to OBJT_DEFAULT so we do not accidently
791 * double-remove the object from the swap queues.
792 */
984263bc 793 if (destroysource) {
984263bc
MD
794 /*
795 * Reverting the type is not necessary, the caller is going
796 * to destroy srcobject directly, but I'm doing it here
797 * for consistency since we've removed the object from its
798 * queues.
799 */
96adc753 800 swp_pager_meta_free_all(srcobject);
8d292090
MD
801 if (srcobject->type == OBJT_SWAP)
802 srcobject->type = OBJT_DEFAULT;
984263bc 803 }
984263bc
MD
804}
805
806/*
807 * SWAP_PAGER_HASPAGE() - determine if we have good backing store for
808 * the requested page.
809 *
810 * We determine whether good backing store exists for the requested
811 * page and return TRUE if it does, FALSE if it doesn't.
812 *
813 * If TRUE, we also try to determine how much valid, contiguous backing
814 * store exists before and after the requested page within a reasonable
815 * distance. We do not try to restrict it to the swap device stripe
816 * (that is handled in getpages/putpages). It probably isn't worth
817 * doing here.
8e7c4729
MD
818 *
819 * No requirements.
984263bc 820 */
984263bc 821boolean_t
1b9d3514 822swap_pager_haspage(vm_object_t object, vm_pindex_t pindex)
984263bc 823{
651d8e75 824 swblk_t blk0;
984263bc
MD
825
826 /*
827 * do we have good backing store at the requested index ?
828 */
b12defdc 829 vm_object_hold(object);
984263bc
MD
830 blk0 = swp_pager_meta_ctl(object, pindex, 0);
831
832 if (blk0 == SWAPBLK_NONE) {
b12defdc 833 vm_object_drop(object);
984263bc
MD
834 return (FALSE);
835 }
b12defdc 836 vm_object_drop(object);
984263bc
MD
837 return (TRUE);
838}
839
840/*
841 * SWAP_PAGER_PAGE_UNSWAPPED() - remove swap backing store related to page
842 *
107e9bcc
MD
843 * This removes any associated swap backing store, whether valid or
844 * not, from the page. This operates on any VM object, not just OBJT_SWAP
845 * objects.
984263bc 846 *
107e9bcc
MD
847 * This routine is typically called when a page is made dirty, at
848 * which point any associated swap can be freed. MADV_FREE also
849 * calls us in a special-case situation
984263bc 850 *
107e9bcc
MD
851 * NOTE!!! If the page is clean and the swap was valid, the caller
852 * should make the page dirty before calling this routine. This routine
853 * does NOT change the m->dirty status of the page. Also: MADV_FREE
854 * depends on it.
984263bc 855 *
135b4b20 856 * The page must be busied or soft-busied.
b12defdc 857 * The caller can hold the object to avoid blocking, else we might block.
8e7c4729 858 * No other requirements.
984263bc 859 */
107e9bcc 860void
57e43348 861swap_pager_unswapped(vm_page_t m)
984263bc 862{
67803f3e 863 if (m->flags & PG_SWAPPED) {
b12defdc 864 vm_object_hold(m->object);
135b4b20 865 KKASSERT(m->flags & PG_SWAPPED);
67803f3e
MD
866 swp_pager_meta_ctl(m->object, m->pindex, SWM_FREE);
867 vm_page_flag_clear(m, PG_SWAPPED);
b12defdc 868 vm_object_drop(m->object);
67803f3e 869 }
984263bc
MD
870}
871
872/*
873 * SWAP_PAGER_STRATEGY() - read, write, free blocks
874 *
107e9bcc
MD
875 * This implements a VM OBJECT strategy function using swap backing store.
876 * This can operate on any VM OBJECT type, not necessarily just OBJT_SWAP
877 * types.
878 *
879 * This is intended to be a cacheless interface (i.e. caching occurs at
880 * higher levels), and is also used as a swap-based SSD cache for vnode
881 * and device objects.
882 *
883 * All I/O goes directly to and from the swap device.
984263bc 884 *
107e9bcc
MD
885 * We currently attempt to run I/O synchronously or asynchronously as
886 * the caller requests. This isn't perfect because we loose error
887 * sequencing when we run multiple ops in parallel to satisfy a request.
888 * But this is swap, so we let it all hang out.
8e7c4729
MD
889 *
890 * No requirements.
984263bc 891 */
107e9bcc 892void
81b5c339 893swap_pager_strategy(vm_object_t object, struct bio *bio)
984263bc 894{
81b5c339
MD
895 struct buf *bp = bio->bio_buf;
896 struct bio *nbio;
984263bc 897 vm_pindex_t start;
54078292 898 vm_pindex_t biox_blkno = 0;
984263bc 899 int count;
984263bc 900 char *data;
ae8e83e6
MD
901 struct bio *biox;
902 struct buf *bufx;
81b5c339
MD
903 struct bio_track *track;
904
905 /*
906 * tracking for swapdev vnode I/Os
907 */
10f3fee5 908 if (bp->b_cmd == BUF_CMD_READ)
81b5c339
MD
909 track = &swapdev_vp->v_track_read;
910 else
911 track = &swapdev_vp->v_track_write;
984263bc
MD
912
913 if (bp->b_bcount & PAGE_MASK) {
914 bp->b_error = EINVAL;
915 bp->b_flags |= B_ERROR | B_INVAL;
81b5c339 916 biodone(bio);
973c11b9
MD
917 kprintf("swap_pager_strategy: bp %p offset %lld size %d, "
918 "not page bounded\n",
919 bp, (long long)bio->bio_offset, (int)bp->b_bcount);
984263bc
MD
920 return;
921 }
922
923 /*
924 * Clear error indication, initialize page index, count, data pointer.
925 */
984263bc
MD
926 bp->b_error = 0;
927 bp->b_flags &= ~B_ERROR;
928 bp->b_resid = bp->b_bcount;
929
54078292 930 start = (vm_pindex_t)(bio->bio_offset >> PAGE_SHIFT);
984263bc
MD
931 count = howmany(bp->b_bcount, PAGE_SIZE);
932 data = bp->b_data;
933
984263bc 934 /*
10f3fee5 935 * Deal with BUF_CMD_FREEBLKS
984263bc 936 */
10f3fee5 937 if (bp->b_cmd == BUF_CMD_FREEBLKS) {
984263bc
MD
938 /*
939 * FREE PAGE(s) - destroy underlying swap that is no longer
940 * needed.
941 */
b12defdc 942 vm_object_hold(object);
984263bc 943 swp_pager_meta_free(object, start, count);
b12defdc 944 vm_object_drop(object);
984263bc 945 bp->b_resid = 0;
81b5c339 946 biodone(bio);
984263bc
MD
947 return;
948 }
949
950 /*
81b5c339
MD
951 * We need to be able to create a new cluster of I/O's. We cannot
952 * use the caller fields of the passed bio so push a new one.
953 *
954 * Because nbio is just a placeholder for the cluster links,
955 * we can biodone() the original bio instead of nbio to make
956 * things a bit more efficient.
957 */
958 nbio = push_bio(bio);
54078292 959 nbio->bio_offset = bio->bio_offset;
81b5c339
MD
960 nbio->bio_caller_info1.cluster_head = NULL;
961 nbio->bio_caller_info2.cluster_tail = NULL;
962
ae8e83e6
MD
963 biox = NULL;
964 bufx = NULL;
965
81b5c339 966 /*
984263bc
MD
967 * Execute read or write
968 */
b12defdc
MD
969 vm_object_hold(object);
970
984263bc 971 while (count > 0) {
651d8e75 972 swblk_t blk;
984263bc
MD
973
974 /*
975 * Obtain block. If block not found and writing, allocate a
976 * new block and build it into the object.
977 */
984263bc 978 blk = swp_pager_meta_ctl(object, start, 0);
10f3fee5 979 if ((blk == SWAPBLK_NONE) && bp->b_cmd != BUF_CMD_READ) {
096e95c0 980 blk = swp_pager_getswapspace(object, 1);
984263bc
MD
981 if (blk == SWAPBLK_NONE) {
982 bp->b_error = ENOMEM;
983 bp->b_flags |= B_ERROR;
984 break;
985 }
986 swp_pager_meta_build(object, start, blk);
987 }
988
989 /*
990 * Do we have to flush our current collection? Yes if:
991 *
992 * - no swap block at this index
993 * - swap block is not contiguous
994 * - we cross a physical disk boundry in the
995 * stripe.
996 */
984263bc 997 if (
54078292
MD
998 biox && (biox_blkno + btoc(bufx->b_bcount) != blk ||
999 ((biox_blkno ^ blk) & dmmax_mask)
984263bc
MD
1000 )
1001 ) {
10f3fee5 1002 if (bp->b_cmd == BUF_CMD_READ) {
12e4aaff 1003 ++mycpu->gd_cnt.v_swapin;
81b5c339 1004 mycpu->gd_cnt.v_swappgsin += btoc(bufx->b_bcount);
984263bc 1005 } else {
12e4aaff 1006 ++mycpu->gd_cnt.v_swapout;
81b5c339
MD
1007 mycpu->gd_cnt.v_swappgsout += btoc(bufx->b_bcount);
1008 bufx->b_dirtyend = bufx->b_bcount;
1009 }
1010
1011 /*
ae8e83e6 1012 * Finished with this buf.
81b5c339 1013 */
ae8e83e6
MD
1014 KKASSERT(bufx->b_bcount != 0);
1015 if (bufx->b_cmd != BUF_CMD_READ)
1016 bufx->b_dirtyend = bufx->b_bcount;
81b5c339
MD
1017 biox = NULL;
1018 bufx = NULL;
984263bc
MD
1019 }
1020
1021 /*
81b5c339 1022 * Add new swapblk to biox, instantiating biox if necessary.
984263bc
MD
1023 * Zero-fill reads are able to take a shortcut.
1024 */
984263bc
MD
1025 if (blk == SWAPBLK_NONE) {
1026 /*
1027 * We can only get here if we are reading. Since
1028 * we are at splvm() we can safely modify b_resid,
1029 * even if chain ops are in progress.
1030 */
1031 bzero(data, PAGE_SIZE);
1032 bp->b_resid -= PAGE_SIZE;
1033 } else {
81b5c339
MD
1034 if (biox == NULL) {
1035 /* XXX chain count > 4, wait to <= 4 */
1036
1037 bufx = getpbuf(NULL);
1038 biox = &bufx->b_bio1;
1039 cluster_append(nbio, bufx);
ae8e83e6 1040 bufx->b_flags |= (bufx->b_flags & B_ORDERED);
10f3fee5 1041 bufx->b_cmd = bp->b_cmd;
81b5c339 1042 biox->bio_done = swap_chain_iodone;
54078292 1043 biox->bio_offset = (off_t)blk << PAGE_SHIFT;
81b5c339 1044 biox->bio_caller_info1.cluster_parent = nbio;
54078292 1045 biox_blkno = blk;
81b5c339
MD
1046 bufx->b_bcount = 0;
1047 bufx->b_data = data;
984263bc 1048 }
81b5c339 1049 bufx->b_bcount += PAGE_SIZE;
984263bc
MD
1050 }
1051 --count;
1052 ++start;
1053 data += PAGE_SIZE;
1054 }
b12defdc
MD
1055
1056 vm_object_drop(object);
984263bc
MD
1057
1058 /*
1059 * Flush out last buffer
1060 */
81b5c339 1061 if (biox) {
10f3fee5 1062 if (bufx->b_cmd == BUF_CMD_READ) {
12e4aaff 1063 ++mycpu->gd_cnt.v_swapin;
81b5c339 1064 mycpu->gd_cnt.v_swappgsin += btoc(bufx->b_bcount);
984263bc 1065 } else {
12e4aaff 1066 ++mycpu->gd_cnt.v_swapout;
81b5c339
MD
1067 mycpu->gd_cnt.v_swappgsout += btoc(bufx->b_bcount);
1068 bufx->b_dirtyend = bufx->b_bcount;
1069 }
ae8e83e6
MD
1070 KKASSERT(bufx->b_bcount);
1071 if (bufx->b_cmd != BUF_CMD_READ)
1072 bufx->b_dirtyend = bufx->b_bcount;
81b5c339 1073 /* biox, bufx = NULL */
984263bc
MD
1074 }
1075
1076 /*
ae8e83e6
MD
1077 * Now initiate all the I/O. Be careful looping on our chain as
1078 * I/O's may complete while we are still initiating them.
0a8aee15
MD
1079 *
1080 * If the request is a 100% sparse read no bios will be present
1081 * and we just biodone() the buffer.
984263bc 1082 */
ae8e83e6
MD
1083 nbio->bio_caller_info2.cluster_tail = NULL;
1084 bufx = nbio->bio_caller_info1.cluster_head;
1085
0a8aee15
MD
1086 if (bufx) {
1087 while (bufx) {
1088 biox = &bufx->b_bio1;
1089 BUF_KERNPROC(bufx);
1090 bufx = bufx->b_cluster_next;
1091 vn_strategy(swapdev_vp, biox);
1092 }
1093 } else {
1094 biodone(bio);
984263bc 1095 }
ae8e83e6
MD
1096
1097 /*
1098 * Completion of the cluster will also call biodone_chain(nbio).
1099 * We never call biodone(nbio) so we don't have to worry about
1100 * setting up a bio_done callback. It's handled in the sub-IO.
1101 */
1102 /**/
984263bc
MD
1103}
1104
8e7c4729
MD
1105/*
1106 * biodone callback
1107 *
1108 * No requirements.
1109 */
81b5c339
MD
1110static void
1111swap_chain_iodone(struct bio *biox)
1112{
1113 struct buf **nextp;
1114 struct buf *bufx; /* chained sub-buffer */
1115 struct bio *nbio; /* parent nbio with chain glue */
1116 struct buf *bp; /* original bp associated with nbio */
ae8e83e6 1117 int chain_empty;
81b5c339
MD
1118
1119 bufx = biox->bio_buf;
1120 nbio = biox->bio_caller_info1.cluster_parent;
1121 bp = nbio->bio_buf;
1122
1123 /*
1124 * Update the original buffer
1125 */
1126 KKASSERT(bp != NULL);
1127 if (bufx->b_flags & B_ERROR) {
ae8e83e6 1128 atomic_set_int(&bufx->b_flags, B_ERROR);
77912481 1129 bp->b_error = bufx->b_error; /* race ok */
81b5c339 1130 } else if (bufx->b_resid != 0) {
ae8e83e6 1131 atomic_set_int(&bufx->b_flags, B_ERROR);
77912481 1132 bp->b_error = EINVAL; /* race ok */
81b5c339 1133 } else {
ae8e83e6 1134 atomic_subtract_int(&bp->b_resid, bufx->b_bcount);
81b5c339
MD
1135 }
1136
1137 /*
ae8e83e6 1138 * Remove us from the chain.
81b5c339 1139 */
287a8577 1140 spin_lock(&bp->b_lock.lk_spinlock);
81b5c339
MD
1141 nextp = &nbio->bio_caller_info1.cluster_head;
1142 while (*nextp != bufx) {
1143 KKASSERT(*nextp != NULL);
1144 nextp = &(*nextp)->b_cluster_next;
1145 }
1146 *nextp = bufx->b_cluster_next;
ae8e83e6 1147 chain_empty = (nbio->bio_caller_info1.cluster_head == NULL);
287a8577 1148 spin_unlock(&bp->b_lock.lk_spinlock);
81b5c339
MD
1149
1150 /*
ae8e83e6
MD
1151 * Clean up bufx. If the chain is now empty we finish out
1152 * the parent. Note that we may be racing other completions
1153 * so we must use the chain_empty status from above.
81b5c339 1154 */
ae8e83e6 1155 if (chain_empty) {
81b5c339 1156 if (bp->b_resid != 0 && !(bp->b_flags & B_ERROR)) {
ae8e83e6 1157 atomic_set_int(&bp->b_flags, B_ERROR);
81b5c339
MD
1158 bp->b_error = EINVAL;
1159 }
ae8e83e6 1160 biodone_chain(nbio);
81b5c339 1161 }
81b5c339
MD
1162 relpbuf(bufx, NULL);
1163}
1164
984263bc 1165/*
5d5c5831 1166 * SWAP_PAGER_GETPAGES() - bring page in from swap
984263bc 1167 *
5d5c5831
MD
1168 * The requested page may have to be brought in from swap. Calculate the
1169 * swap block and bring in additional pages if possible. All pages must
1170 * have contiguous swap block assignments and reside in the same object.
984263bc 1171 *
5d5c5831
MD
1172 * The caller has a single vm_object_pip_add() reference prior to
1173 * calling us and we should return with the same.
984263bc 1174 *
5d5c5831
MD
1175 * The caller has BUSY'd the page. We should return with (*mpp) left busy,
1176 * and any additinal pages unbusied.
984263bc 1177 *
5d5c5831
MD
1178 * If the caller encounters a PG_RAM page it will pass it to us even though
1179 * it may be valid and dirty. We cannot overwrite the page in this case!
1180 * The case is used to allow us to issue pure read-aheads.
1181 *
1182 * NOTE! XXX This code does not entirely pipeline yet due to the fact that
1183 * the PG_RAM page is validated at the same time as mreq. What we
1184 * really need to do is issue a separate read-ahead pbuf.
8e7c4729
MD
1185 *
1186 * No requirements.
984263bc 1187 */
984263bc 1188static int
1b9d3514 1189swap_pager_getpage(vm_object_t object, vm_page_t *mpp, int seqaccess)
984263bc
MD
1190{
1191 struct buf *bp;
81b5c339 1192 struct bio *bio;
984263bc 1193 vm_page_t mreq;
5d5c5831
MD
1194 vm_page_t m;
1195 vm_offset_t kva;
651d8e75 1196 swblk_t blk;
984263bc
MD
1197 int i;
1198 int j;
5d5c5831 1199 int raonly;
b12defdc
MD
1200 int error;
1201 u_int32_t flags;
5d5c5831 1202 vm_page_t marray[XIO_INTERNAL_PAGES];
984263bc 1203
1b9d3514 1204 mreq = *mpp;
984263bc 1205
b12defdc 1206 vm_object_hold(object);
984263bc
MD
1207 if (mreq->object != object) {
1208 panic("swap_pager_getpages: object mismatch %p/%p",
1209 object,
1210 mreq->object
1211 );
1212 }
17cde63e 1213
984263bc 1214 /*
5d5c5831
MD
1215 * We don't want to overwrite a fully valid page as it might be
1216 * dirty. This case can occur when e.g. vm_fault hits a perfectly
1217 * valid page with PG_RAM set.
984263bc 1218 *
5d5c5831
MD
1219 * In this case we see if the next page is a suitable page-in
1220 * candidate and if it is we issue read-ahead. PG_RAM will be
1221 * set on the last page of the read-ahead to continue the pipeline.
1222 */
1223 if (mreq->valid == VM_PAGE_BITS_ALL) {
b12defdc
MD
1224 if (swap_burst_read == 0 || mreq->pindex + 1 >= object->size) {
1225 vm_object_drop(object);
5d5c5831 1226 return(VM_PAGER_OK);
b12defdc 1227 }
3bb7eedb
MD
1228 blk = swp_pager_meta_ctl(object, mreq->pindex + 1, 0);
1229 if (blk == SWAPBLK_NONE) {
b12defdc 1230 vm_object_drop(object);
5d5c5831
MD
1231 return(VM_PAGER_OK);
1232 }
b12defdc
MD
1233 m = vm_page_lookup_busy_try(object, mreq->pindex + 1,
1234 TRUE, &error);
1235 if (error) {
1236 vm_object_drop(object);
1237 return(VM_PAGER_OK);
1238 } else if (m == NULL) {
d2d8515b
MD
1239 /*
1240 * Use VM_ALLOC_QUICK to avoid blocking on cache
1241 * page reuse.
1242 */
5d5c5831
MD
1243 m = vm_page_alloc(object, mreq->pindex + 1,
1244 VM_ALLOC_QUICK);
1245 if (m == NULL) {
b12defdc 1246 vm_object_drop(object);
5d5c5831
MD
1247 return(VM_PAGER_OK);
1248 }
1249 } else {
b12defdc
MD
1250 if (m->valid) {
1251 vm_page_wakeup(m);
1252 vm_object_drop(object);
5d5c5831
MD
1253 return(VM_PAGER_OK);
1254 }
3bb7eedb 1255 vm_page_unqueue_nowakeup(m);
5d5c5831 1256 }
b12defdc 1257 /* page is busy */
5d5c5831
MD
1258 mreq = m;
1259 raonly = 1;
5d5c5831
MD
1260 } else {
1261 raonly = 0;
1262 }
1263
1264 /*
1265 * Try to block-read contiguous pages from swap if sequential,
1266 * otherwise just read one page. Contiguous pages from swap must
1267 * reside within a single device stripe because the I/O cannot be
1268 * broken up across multiple stripes.
1269 *
1270 * Note that blk and iblk can be SWAPBLK_NONE but the loop is
1271 * set up such that the case(s) are handled implicitly.
984263bc 1272 */
984263bc 1273 blk = swp_pager_meta_ctl(mreq->object, mreq->pindex, 0);
5d5c5831 1274 marray[0] = mreq;
984263bc 1275
5d5c5831
MD
1276 for (i = 1; swap_burst_read &&
1277 i < XIO_INTERNAL_PAGES &&
1278 mreq->pindex + i < object->size; ++i) {
651d8e75 1279 swblk_t iblk;
984263bc 1280
5d5c5831
MD
1281 iblk = swp_pager_meta_ctl(object, mreq->pindex + i, 0);
1282 if (iblk != blk + i)
984263bc
MD
1283 break;
1284 if ((blk ^ iblk) & dmmax_mask)
1285 break;
b12defdc
MD
1286 m = vm_page_lookup_busy_try(object, mreq->pindex + i,
1287 TRUE, &error);
1288 if (error) {
1289 break;
1290 } else if (m == NULL) {
d2d8515b
MD
1291 /*
1292 * Use VM_ALLOC_QUICK to avoid blocking on cache
1293 * page reuse.
1294 */
5d5c5831
MD
1295 m = vm_page_alloc(object, mreq->pindex + i,
1296 VM_ALLOC_QUICK);
1297 if (m == NULL)
1298 break;
1299 } else {
b12defdc
MD
1300 if (m->valid) {
1301 vm_page_wakeup(m);
5d5c5831 1302 break;
b12defdc 1303 }
3bb7eedb 1304 vm_page_unqueue_nowakeup(m);
5d5c5831 1305 }
b12defdc 1306 /* page is busy */
5d5c5831 1307 marray[i] = m;
984263bc 1308 }
5d5c5831
MD
1309 if (i > 1)
1310 vm_page_flag_set(marray[i - 1], PG_RAM);
984263bc 1311
984263bc 1312 /*
5d5c5831
MD
1313 * If mreq is the requested page and we have nothing to do return
1314 * VM_PAGER_FAIL. If raonly is set mreq is just another read-ahead
1315 * page and must be cleaned up.
984263bc 1316 */
5d5c5831
MD
1317 if (blk == SWAPBLK_NONE) {
1318 KKASSERT(i == 1);
1319 if (raonly) {
1320 vnode_pager_freepage(mreq);
b12defdc 1321 vm_object_drop(object);
5d5c5831
MD
1322 return(VM_PAGER_OK);
1323 } else {
b12defdc 1324 vm_object_drop(object);
5d5c5831
MD
1325 return(VM_PAGER_FAIL);
1326 }
1327 }
984263bc
MD
1328
1329 /*
5d5c5831 1330 * map our page(s) into kva for input
984263bc 1331 */
9a82e536 1332 bp = getpbuf_kva(&nsw_rcount);
81b5c339 1333 bio = &bp->b_bio1;
5d5c5831
MD
1334 kva = (vm_offset_t) bp->b_kvabase;
1335 bcopy(marray, bp->b_xio.xio_pages, i * sizeof(vm_page_t));
1336 pmap_qenter(kva, bp->b_xio.xio_pages, i);
984263bc 1337
5d5c5831
MD
1338 bp->b_data = (caddr_t)kva;
1339 bp->b_bcount = PAGE_SIZE * i;
1340 bp->b_xio.xio_npages = i;
81b5c339 1341 bio->bio_done = swp_pager_async_iodone;
5d5c5831 1342 bio->bio_offset = (off_t)blk << PAGE_SHIFT;
8aa92e4b 1343 bio->bio_caller_info1.index = SWBIO_READ;
984263bc 1344
5d5c5831
MD
1345 /*
1346 * Set index. If raonly set the index beyond the array so all
1347 * the pages are treated the same, otherwise the original mreq is
1348 * at index 0.
1349 */
1350 if (raonly)
1351 bio->bio_driver_info = (void *)(intptr_t)i;
1352 else
1353 bio->bio_driver_info = (void *)(intptr_t)0;
984263bc 1354
5d5c5831
MD
1355 for (j = 0; j < i; ++j)
1356 vm_page_flag_set(bp->b_xio.xio_pages[j], PG_SWAPINPROG);
984263bc 1357
12e4aaff 1358 mycpu->gd_cnt.v_swapin++;
54f51aeb 1359 mycpu->gd_cnt.v_swappgsin += bp->b_xio.xio_npages;
984263bc
MD
1360
1361 /*
1362 * We still hold the lock on mreq, and our automatic completion routine
1363 * does not remove it.
1364 */
3bb7eedb 1365 vm_object_pip_add(object, bp->b_xio.xio_npages);
984263bc
MD
1366
1367 /*
1368 * perform the I/O. NOTE!!! bp cannot be considered valid after
1369 * this point because we automatically release it on completion.
1370 * Instead, we look at the one page we are interested in which we
1371 * still hold a lock on even through the I/O completion.
1372 *
1373 * The other pages in our m[] array are also released on completion,
1374 * so we cannot assume they are valid anymore either.
984263bc 1375 */
10f3fee5 1376 bp->b_cmd = BUF_CMD_READ;
984263bc 1377 BUF_KERNPROC(bp);
81b5c339 1378 vn_strategy(swapdev_vp, bio);
984263bc
MD
1379
1380 /*
3bb7eedb 1381 * Wait for the page we want to complete. PG_SWAPINPROG is always
984263bc
MD
1382 * cleared on completion. If an I/O error occurs, SWAPBLK_NONE
1383 * is set in the meta-data.
5d5c5831
MD
1384 *
1385 * If this is a read-ahead only we return immediately without
1386 * waiting for I/O.
984263bc 1387 */
b12defdc
MD
1388 if (raonly) {
1389 vm_object_drop(object);
5d5c5831 1390 return(VM_PAGER_OK);
b12defdc 1391 }
984263bc 1392
3bb7eedb
MD
1393 /*
1394 * Read-ahead includes originally requested page case.
1395 */
b12defdc
MD
1396 for (;;) {
1397 flags = mreq->flags;
1398 cpu_ccfence();
1399 if ((flags & PG_SWAPINPROG) == 0)
1400 break;
1401 tsleep_interlock(mreq, 0);
1402 if (!atomic_cmpset_int(&mreq->flags, flags,
1403 flags | PG_WANTED | PG_REFERENCED)) {
1404 continue;
1405 }
12e4aaff 1406 mycpu->gd_cnt.v_intrans++;
b12defdc 1407 if (tsleep(mreq, PINTERLOCKED, "swread", hz*20)) {
086c1d7e 1408 kprintf(
81b5c339 1409 "swap_pager: indefinite wait buffer: "
973c11b9
MD
1410 " offset: %lld, size: %ld\n",
1411 (long long)bio->bio_offset,
1412 (long)bp->b_bcount
984263bc
MD
1413 );
1414 }
1415 }
984263bc
MD
1416
1417 /*
1418 * mreq is left bussied after completion, but all the other pages
1419 * are freed. If we had an unrecoverable read error the page will
1420 * not be valid.
1421 */
b12defdc 1422 vm_object_drop(object);
5d5c5831 1423 if (mreq->valid != VM_PAGE_BITS_ALL)
984263bc 1424 return(VM_PAGER_ERROR);
5d5c5831 1425 else
984263bc 1426 return(VM_PAGER_OK);
984263bc
MD
1427
1428 /*
1429 * A final note: in a low swap situation, we cannot deallocate swap
1430 * and mark a page dirty here because the caller is likely to mark
1431 * the page clean when we return, causing the page to possibly revert
1432 * to all-zero's later.
1433 */
1434}
1435
1436/*
1437 * swap_pager_putpages:
1438 *
1439 * Assign swap (if necessary) and initiate I/O on the specified pages.
1440 *
1441 * We support both OBJT_DEFAULT and OBJT_SWAP objects. DEFAULT objects
1442 * are automatically converted to SWAP objects.
1443 *
81b5c339 1444 * In a low memory situation we may block in vn_strategy(), but the new
984263bc
MD
1445 * vm_page reservation system coupled with properly written VFS devices
1446 * should ensure that no low-memory deadlock occurs. This is an area
1447 * which needs work.
1448 *
1449 * The parent has N vm_object_pip_add() references prior to
1450 * calling us and will remove references for rtvals[] that are
1451 * not set to VM_PAGER_PEND. We need to remove the rest on I/O
1452 * completion.
1453 *
1454 * The parent has soft-busy'd the pages it passes us and will unbusy
1455 * those whos rtvals[] entry is not set to VM_PAGER_PEND on return.
1456 * We need to unbusy the rest on I/O completion.
8e7c4729
MD
1457 *
1458 * No requirements.
984263bc 1459 */
984263bc 1460void
17cde63e
MD
1461swap_pager_putpages(vm_object_t object, vm_page_t *m, int count,
1462 boolean_t sync, int *rtvals)
984263bc
MD
1463{
1464 int i;
1465 int n = 0;
1466
b12defdc
MD
1467 vm_object_hold(object);
1468
984263bc
MD
1469 if (count && m[0]->object != object) {
1470 panic("swap_pager_getpages: object mismatch %p/%p",
1471 object,
1472 m[0]->object
1473 );
1474 }
17cde63e 1475
984263bc
MD
1476 /*
1477 * Step 1
1478 *
1479 * Turn object into OBJT_SWAP
1480 * check for bogus sysops
1481 * force sync if not pageout process
1482 */
8e7c4729 1483 if (object->type == OBJT_DEFAULT) {
8e7c4729
MD
1484 if (object->type == OBJT_DEFAULT)
1485 swp_pager_meta_convert(object);
8e7c4729 1486 }
984263bc 1487
bc6dffab 1488 if (curthread != pagethread)
984263bc
MD
1489 sync = TRUE;
1490
1491 /*
1492 * Step 2
1493 *
1494 * Update nsw parameters from swap_async_max sysctl values.
1495 * Do not let the sysop crash the machine with bogus numbers.
1496 */
984263bc
MD
1497 if (swap_async_max != nsw_wcount_async_max) {
1498 int n;
984263bc
MD
1499
1500 /*
1501 * limit range
1502 */
1503 if ((n = swap_async_max) > nswbuf / 2)
1504 n = nswbuf / 2;
1505 if (n < 1)
1506 n = 1;
1507 swap_async_max = n;
1508
1509 /*
1510 * Adjust difference ( if possible ). If the current async
1511 * count is too low, we may not be able to make the adjustment
1512 * at this time.
b12defdc
MD
1513 *
1514 * vm_token needed for nsw_wcount sleep interlock
984263bc 1515 */
8e7c4729 1516 lwkt_gettoken(&vm_token);
984263bc
MD
1517 n -= nsw_wcount_async_max;
1518 if (nsw_wcount_async + n >= 0) {
984263bc 1519 nsw_wcount_async_max += n;
7f86d367 1520 pbuf_adjcount(&nsw_wcount_async, n);
984263bc 1521 }
8e7c4729 1522 lwkt_reltoken(&vm_token);
984263bc
MD
1523 }
1524
1525 /*
1526 * Step 3
1527 *
1528 * Assign swap blocks and issue I/O. We reallocate swap on the fly.
1529 * The page is left dirty until the pageout operation completes
1530 * successfully.
1531 */
1532
1533 for (i = 0; i < count; i += n) {
984263bc 1534 struct buf *bp;
81b5c339 1535 struct bio *bio;
651d8e75 1536 swblk_t blk;
81b5c339 1537 int j;
984263bc
MD
1538
1539 /*
1540 * Maximum I/O size is limited by a number of factors.
1541 */
1542
1543 n = min(BLIST_MAX_ALLOC, count - i);
1544 n = min(n, nsw_cluster_max);
1545
8e7c4729 1546 lwkt_gettoken(&vm_token);
984263bc
MD
1547
1548 /*
1549 * Get biggest block of swap we can. If we fail, fall
1550 * back and try to allocate a smaller block. Don't go
1551 * overboard trying to allocate space if it would overly
1552 * fragment swap.
1553 */
1554 while (
096e95c0 1555 (blk = swp_pager_getswapspace(object, n)) == SWAPBLK_NONE &&
984263bc
MD
1556 n > 4
1557 ) {
1558 n >>= 1;
1559 }
1560 if (blk == SWAPBLK_NONE) {
1561 for (j = 0; j < n; ++j)
1562 rtvals[i+j] = VM_PAGER_FAIL;
8e7c4729 1563 lwkt_reltoken(&vm_token);
984263bc
MD
1564 continue;
1565 }
1566
1567 /*
1568 * The I/O we are constructing cannot cross a physical
1569 * disk boundry in the swap stripe. Note: we are still
1570 * at splvm().
1571 */
1572 if ((blk ^ (blk + n)) & dmmax_mask) {
1573 j = ((blk + dmmax) & dmmax_mask) - blk;
096e95c0 1574 swp_pager_freeswapspace(object, blk + j, n - j);
984263bc
MD
1575 n = j;
1576 }
1577
1578 /*
1579 * All I/O parameters have been satisfied, build the I/O
1580 * request and assign the swap space.
984263bc 1581 */
10f3fee5 1582 if (sync == TRUE)
9a82e536 1583 bp = getpbuf_kva(&nsw_wcount_sync);
10f3fee5 1584 else
9a82e536 1585 bp = getpbuf_kva(&nsw_wcount_async);
81b5c339 1586 bio = &bp->b_bio1;
984263bc 1587
b12defdc
MD
1588 lwkt_reltoken(&vm_token);
1589
984263bc
MD
1590 pmap_qenter((vm_offset_t)bp->b_data, &m[i], n);
1591
984263bc 1592 bp->b_bcount = PAGE_SIZE * n;
54078292 1593 bio->bio_offset = (off_t)blk << PAGE_SHIFT;
984263bc 1594
984263bc
MD
1595 for (j = 0; j < n; ++j) {
1596 vm_page_t mreq = m[i+j];
1597
096e95c0
MD
1598 swp_pager_meta_build(mreq->object, mreq->pindex,
1599 blk + j);
1600 if (object->type == OBJT_SWAP)
1601 vm_page_dirty(mreq);
984263bc
MD
1602 rtvals[i+j] = VM_PAGER_OK;
1603
1604 vm_page_flag_set(mreq, PG_SWAPINPROG);
54f51aeb 1605 bp->b_xio.xio_pages[j] = mreq;
984263bc 1606 }
54f51aeb 1607 bp->b_xio.xio_npages = n;
984263bc 1608
12e4aaff 1609 mycpu->gd_cnt.v_swapout++;
54f51aeb 1610 mycpu->gd_cnt.v_swappgsout += bp->b_xio.xio_npages;
984263bc 1611
10f3fee5
MD
1612 bp->b_dirtyoff = 0; /* req'd for NFS */
1613 bp->b_dirtyend = bp->b_bcount; /* req'd for NFS */
1614 bp->b_cmd = BUF_CMD_WRITE;
8aa92e4b 1615 bio->bio_caller_info1.index = SWBIO_WRITE;
10f3fee5 1616
984263bc
MD
1617 /*
1618 * asynchronous
984263bc 1619 */
984263bc 1620 if (sync == FALSE) {
81b5c339 1621 bio->bio_done = swp_pager_async_iodone;
984263bc 1622 BUF_KERNPROC(bp);
81b5c339 1623 vn_strategy(swapdev_vp, bio);
984263bc
MD
1624
1625 for (j = 0; j < n; ++j)
1626 rtvals[i+j] = VM_PAGER_PEND;
1627 continue;
1628 }
1629
1630 /*
ae8e83e6
MD
1631 * Issue synchrnously.
1632 *
984263bc
MD
1633 * Wait for the sync I/O to complete, then update rtvals.
1634 * We just set the rtvals[] to VM_PAGER_PEND so we can call
1635 * our async completion routine at the end, thus avoiding a
1636 * double-free.
1637 */
8aa92e4b 1638 bio->bio_caller_info1.index |= SWBIO_SYNC;
ae8e83e6
MD
1639 bio->bio_done = biodone_sync;
1640 bio->bio_flags |= BIO_SYNC;
1641 vn_strategy(swapdev_vp, bio);
1642 biowait(bio, "swwrt");
984263bc
MD
1643
1644 for (j = 0; j < n; ++j)
1645 rtvals[i+j] = VM_PAGER_PEND;
1646
1647 /*
1648 * Now that we are through with the bp, we can call the
1649 * normal async completion, which frees everything up.
1650 */
81b5c339 1651 swp_pager_async_iodone(bio);
984263bc 1652 }
b12defdc 1653 vm_object_drop(object);
984263bc
MD
1654}
1655
8e7c4729
MD
1656/*
1657 * No requirements.
1658 */
c84c24da
MD
1659void
1660swap_pager_newswap(void)
1661{
1662 swp_sizecheck();
1663}
1664
984263bc 1665/*
984263bc
MD
1666 * swp_pager_async_iodone:
1667 *
1668 * Completion routine for asynchronous reads and writes from/to swap.
1669 * Also called manually by synchronous code to finish up a bp.
1670 *
1671 * For READ operations, the pages are PG_BUSY'd. For WRITE operations,
1672 * the pages are vm_page_t->busy'd. For READ operations, we PG_BUSY
1673 * unbusy all pages except the 'main' request page. For WRITE
1674 * operations, we vm_page_t->busy'd unbusy all pages ( we can do this
1675 * because we marked them all VM_PAGER_PEND on return from putpages ).
1676 *
1677 * This routine may not block.
8e7c4729
MD
1678 *
1679 * No requirements.
984263bc 1680 */
984263bc 1681static void
81b5c339 1682swp_pager_async_iodone(struct bio *bio)
984263bc 1683{
81b5c339 1684 struct buf *bp = bio->bio_buf;
984263bc 1685 vm_object_t object = NULL;
81b5c339 1686 int i;
10f3fee5 1687 int *nswptr;
984263bc
MD
1688
1689 /*
1690 * report error
1691 */
984263bc 1692 if (bp->b_flags & B_ERROR) {
086c1d7e 1693 kprintf(
54078292 1694 "swap_pager: I/O error - %s failed; offset %lld,"
984263bc 1695 "size %ld, error %d\n",
8aa92e4b
MD
1696 ((bio->bio_caller_info1.index & SWBIO_READ) ?
1697 "pagein" : "pageout"),
973c11b9 1698 (long long)bio->bio_offset,
984263bc
MD
1699 (long)bp->b_bcount,
1700 bp->b_error
1701 );
1702 }
1703
1704 /*
1705 * set object, raise to splvm().
1706 */
54f51aeb
HP
1707 if (bp->b_xio.xio_npages)
1708 object = bp->b_xio.xio_pages[0]->object;
984263bc
MD
1709
1710 /*
1711 * remove the mapping for kernel virtual
1712 */
54f51aeb 1713 pmap_qremove((vm_offset_t)bp->b_data, bp->b_xio.xio_npages);
984263bc
MD
1714
1715 /*
1716 * cleanup pages. If an error occurs writing to swap, we are in
1717 * very serious trouble. If it happens to be a disk error, though,
1718 * we may be able to recover by reassigning the swap later on. So
1719 * in this case we remove the m->swapblk assignment for the page
1720 * but do not free it in the rlist. The errornous block(s) are thus
1721 * never reallocated as swap. Redirty the page and continue.
1722 */
54f51aeb
HP
1723 for (i = 0; i < bp->b_xio.xio_npages; ++i) {
1724 vm_page_t m = bp->b_xio.xio_pages[i];
984263bc 1725
984263bc
MD
1726 if (bp->b_flags & B_ERROR) {
1727 /*
1728 * If an error occurs I'd love to throw the swapblk
1729 * away without freeing it back to swapspace, so it
1730 * can never be used again. But I can't from an
1731 * interrupt.
1732 */
1733
8aa92e4b 1734 if (bio->bio_caller_info1.index & SWBIO_READ) {
984263bc
MD
1735 /*
1736 * When reading, reqpage needs to stay
1737 * locked for the parent, but all other
1738 * pages can be freed. We still want to
1739 * wakeup the parent waiting on the page,
1740 * though. ( also: pg_reqpage can be -1 and
1741 * not match anything ).
1742 *
1743 * We have to wake specifically requested pages
1744 * up too because we cleared PG_SWAPINPROG and
1745 * someone may be waiting for that.
1746 *
1747 * NOTE: for reads, m->dirty will probably
1748 * be overridden by the original caller of
1749 * getpages so don't play cute tricks here.
1750 *
93afe6be
MD
1751 * NOTE: We can't actually free the page from
1752 * here, because this is an interrupt. It
1753 * is not legal to mess with object->memq
1754 * from an interrupt. Deactivate the page
1755 * instead.
984263bc
MD
1756 */
1757
1758 m->valid = 0;
1759 vm_page_flag_clear(m, PG_ZERO);
5d5c5831 1760 vm_page_flag_clear(m, PG_SWAPINPROG);
984263bc 1761
81b5c339
MD
1762 /*
1763 * bio_driver_info holds the requested page
1764 * index.
1765 */
973c11b9 1766 if (i != (int)(intptr_t)bio->bio_driver_info) {
93afe6be
MD
1767 vm_page_deactivate(m);
1768 vm_page_wakeup(m);
1769 } else {
984263bc 1770 vm_page_flash(m);
93afe6be 1771 }
984263bc
MD
1772 /*
1773 * If i == bp->b_pager.pg_reqpage, do not wake
1774 * the page up. The caller needs to.
1775 */
1776 } else {
1777 /*
3ffc7051
MD
1778 * If a write error occurs remove the swap
1779 * assignment (note that PG_SWAPPED may or
1780 * may not be set depending on prior activity).
096e95c0 1781 *
3ffc7051
MD
1782 * Re-dirty OBJT_SWAP pages as there is no
1783 * other backing store, we can't throw the
1784 * page away.
1785 *
1786 * Non-OBJT_SWAP pages (aka swapcache) must
1787 * not be dirtied since they may not have
1788 * been dirty in the first place, and they
1789 * do have backing store (the vnode).
984263bc 1790 */
b12defdc 1791 vm_page_busy_wait(m, FALSE, "swadpg");
3ffc7051
MD
1792 swp_pager_meta_ctl(m->object, m->pindex,
1793 SWM_FREE);
1794 vm_page_flag_clear(m, PG_SWAPPED);
096e95c0
MD
1795 if (m->object->type == OBJT_SWAP) {
1796 vm_page_dirty(m);
1797 vm_page_activate(m);
1798 }
3ffc7051 1799 vm_page_flag_clear(m, PG_SWAPINPROG);
984263bc 1800 vm_page_io_finish(m);
b12defdc 1801 vm_page_wakeup(m);
984263bc 1802 }
8aa92e4b 1803 } else if (bio->bio_caller_info1.index & SWBIO_READ) {
984263bc 1804 /*
984263bc
MD
1805 * NOTE: for reads, m->dirty will probably be
1806 * overridden by the original caller of getpages so
1807 * we cannot set them in order to free the underlying
1808 * swap in a low-swap situation. I don't think we'd
1809 * want to do that anyway, but it was an optimization
1810 * that existed in the old swapper for a time before
1811 * it got ripped out due to precisely this problem.
1812 *
1813 * clear PG_ZERO in page.
1814 *
1815 * If not the requested page then deactivate it.
1816 *
1817 * Note that the requested page, reqpage, is left
1818 * busied, but we still have to wake it up. The
1819 * other pages are released (unbusied) by
1820 * vm_page_wakeup(). We do not set reqpage's
1821 * valid bits here, it is up to the caller.
1822 */
1823
4530a3aa
MD
1824 /*
1825 * NOTE: can't call pmap_clear_modify(m) from an
1826 * interrupt thread, the pmap code may have to map
1827 * non-kernel pmaps and currently asserts the case.
1828 */
1829 /*pmap_clear_modify(m);*/
984263bc
MD
1830 m->valid = VM_PAGE_BITS_ALL;
1831 vm_page_undirty(m);
5d5c5831 1832 vm_page_flag_clear(m, PG_ZERO | PG_SWAPINPROG);
67803f3e 1833 vm_page_flag_set(m, PG_SWAPPED);
984263bc
MD
1834
1835 /*
1836 * We have to wake specifically requested pages
1837 * up too because we cleared PG_SWAPINPROG and
1838 * could be waiting for it in getpages. However,
1839 * be sure to not unbusy getpages specifically
1840 * requested page - getpages expects it to be
1841 * left busy.
81b5c339
MD
1842 *
1843 * bio_driver_info holds the requested page
984263bc 1844 */
973c11b9 1845 if (i != (int)(intptr_t)bio->bio_driver_info) {
984263bc
MD
1846 vm_page_deactivate(m);
1847 vm_page_wakeup(m);
1848 } else {
1849 vm_page_flash(m);
1850 }
1851 } else {
1852 /*
93afe6be
MD
1853 * Mark the page clean but do not mess with the
1854 * pmap-layer's modified state. That state should
1855 * also be clear since the caller protected the
1856 * page VM_PROT_READ, but allow the case.
1857 *
1858 * We are in an interrupt, avoid pmap operations.
1859 *
1860 * If we have a severe page deficit, deactivate the
1861 * page. Do not try to cache it (which would also
1862 * involve a pmap op), because the page might still
1863 * be read-heavy.
096e95c0
MD
1864 *
1865 * When using the swap to cache clean vnode pages
1866 * we do not mess with the page dirty bits.
984263bc 1867 */
b12defdc 1868 vm_page_busy_wait(m, FALSE, "swadpg");
096e95c0
MD
1869 if (m->object->type == OBJT_SWAP)
1870 vm_page_undirty(m);
5d5c5831 1871 vm_page_flag_clear(m, PG_SWAPINPROG);
67803f3e 1872 vm_page_flag_set(m, PG_SWAPPED);
93afe6be
MD
1873 if (vm_page_count_severe())
1874 vm_page_deactivate(m);
1875#if 0
984263bc
MD
1876 if (!vm_page_count_severe() || !vm_page_try_to_cache(m))
1877 vm_page_protect(m, VM_PROT_READ);
93afe6be 1878#endif
a491077e 1879 vm_page_io_finish(m);
b12defdc 1880 vm_page_wakeup(m);
984263bc
MD
1881 }
1882 }
1883
1884 /*
1885 * adjust pip. NOTE: the original parent may still have its own
1886 * pip refs on the object.
1887 */
1888
1889 if (object)
00db03f1 1890 vm_object_pip_wakeup_n(object, bp->b_xio.xio_npages);
984263bc
MD
1891
1892 /*
8aa92e4b
MD
1893 * Release the physical I/O buffer.
1894 *
1895 * NOTE: Due to synchronous operations in the write case b_cmd may
1896 * already be set to BUF_CMD_DONE and BIO_SYNC may have already
1897 * been cleared.
b12defdc
MD
1898 *
1899 * Use vm_token to interlock nsw_rcount/wcount wakeup?
984263bc 1900 */
b12defdc 1901 lwkt_gettoken(&vm_token);
8aa92e4b 1902 if (bio->bio_caller_info1.index & SWBIO_READ)
10f3fee5 1903 nswptr = &nsw_rcount;
8aa92e4b 1904 else if (bio->bio_caller_info1.index & SWBIO_SYNC)
10f3fee5 1905 nswptr = &nsw_wcount_sync;
ae8e83e6
MD
1906 else
1907 nswptr = &nsw_wcount_async;
10f3fee5
MD
1908 bp->b_cmd = BUF_CMD_DONE;
1909 relpbuf(bp, nswptr);
8e7c4729 1910 lwkt_reltoken(&vm_token);
984263bc
MD
1911}
1912
9f3543c6
MD
1913/*
1914 * Fault-in a potentially swapped page and remove the swap reference.
b12defdc
MD
1915 *
1916 * object must be held.
9f3543c6
MD
1917 */
1918static __inline void
1919swp_pager_fault_page(vm_object_t object, vm_pindex_t pindex)
1920{
1921 struct vnode *vp;
1922 vm_page_t m;
1923 int error;
1924
b12defdc
MD
1925 ASSERT_LWKT_TOKEN_HELD(vm_object_token(object));
1926
9f3543c6
MD
1927 if (object->type == OBJT_VNODE) {
1928 /*
1929 * Any swap related to a vnode is due to swapcache. We must
1930 * vget() the vnode in case it is not active (otherwise
1931 * vref() will panic). Calling vm_object_page_remove() will
1932 * ensure that any swap ref is removed interlocked with the
1933 * page. clean_only is set to TRUE so we don't throw away
1934 * dirty pages.
1935 */
1936 vp = object->handle;
1937 error = vget(vp, LK_SHARED | LK_RETRY | LK_CANRECURSE);
1938 if (error == 0) {
1939 vm_object_page_remove(object, pindex, pindex + 1, TRUE);
1940 vput(vp);
1941 }
1942 } else {
1943 /*
1944 * Otherwise it is a normal OBJT_SWAP object and we can
1945 * fault the page in and remove the swap.
1946 */
1947 m = vm_fault_object_page(object, IDX_TO_OFF(pindex),
1948 VM_PROT_NONE,
1949 VM_FAULT_DIRTY | VM_FAULT_UNSWAP,
1950 &error);
1951 if (m)
1952 vm_page_unhold(m);
1953 }
1954}
1955
1956int
1957swap_pager_swapoff(int devidx)
1958{
1959 vm_object_t object;
1960 struct swblock *swap;
1961 swblk_t v;
1962 int i;
1963
9f3543c6
MD
1964 lwkt_gettoken(&vmobj_token);
1965rescan:
1966 TAILQ_FOREACH(object, &vm_object_list, object_list) {
b12defdc
MD
1967 if (object->type != OBJT_SWAP && object->type != OBJT_VNODE)
1968 continue;
1969 vm_object_hold(object);
9f3543c6 1970 if (object->type == OBJT_SWAP || object->type == OBJT_VNODE) {
b12defdc
MD
1971 RB_FOREACH(swap,
1972 swblock_rb_tree, &object->swblock_root) {
9f3543c6
MD
1973 for (i = 0; i < SWAP_META_PAGES; ++i) {
1974 v = swap->swb_pages[i];
1975 if (v != SWAPBLK_NONE &&
1976 BLK2DEVIDX(v) == devidx) {
1977 swp_pager_fault_page(
1978 object,
1979 swap->swb_index + i);
b12defdc 1980 vm_object_drop(object);
9f3543c6
MD
1981 goto rescan;
1982 }
1983 }
1984 }
1985 }
b12defdc 1986 vm_object_drop(object);
9f3543c6
MD
1987 }
1988 lwkt_reltoken(&vmobj_token);
9f3543c6
MD
1989
1990 /*
1991 * If we fail to locate all swblocks we just fail gracefully and
1992 * do not bother to restore paging on the swap device. If the
1993 * user wants to retry the user can retry.
1994 */
1995 if (swdevt[devidx].sw_nused)
1996 return (1);
1997 else
1998 return (0);
1999}
2000
984263bc
MD
2001/************************************************************************
2002 * SWAP META DATA *
2003 ************************************************************************
2004 *
2005 * These routines manipulate the swap metadata stored in the
2006 * OBJT_SWAP object. All swp_*() routines must be called at
2007 * splvm() because swap can be freed up by the low level vm_page
2008 * code which might be called from interrupts beyond what splbio() covers.
2009 *
2010 * Swap metadata is implemented with a global hash and not directly
2011 * linked into the object. Instead the object simply contains
2012 * appropriate tracking counters.
2013 */
2014
2015/*
96adc753 2016 * Lookup the swblock containing the specified swap block index.
8e7c4729 2017 *
b12defdc 2018 * The caller must hold the object.
984263bc 2019 */
96adc753
MD
2020static __inline
2021struct swblock *
2022swp_pager_lookup(vm_object_t object, vm_pindex_t index)
984263bc 2023{
b12defdc 2024 ASSERT_LWKT_TOKEN_HELD(vm_object_token(object));
0301c407 2025 index &= ~(vm_pindex_t)SWAP_META_MASK;
96adc753 2026 return (RB_LOOKUP(swblock_rb_tree, &object->swblock_root, index));
984263bc
MD
2027}
2028
2029/*
96adc753 2030 * Remove a swblock from the RB tree.
8e7c4729 2031 *
b12defdc 2032 * The caller must hold the object.
984263bc 2033 */
96adc753
MD
2034static __inline
2035void
2036swp_pager_remove(vm_object_t object, struct swblock *swap)
2037{
b12defdc 2038 ASSERT_LWKT_TOKEN_HELD(vm_object_token(object));
96adc753
MD
2039 RB_REMOVE(swblock_rb_tree, &object->swblock_root, swap);
2040}
984263bc 2041
96adc753
MD
2042/*
2043 * Convert default object to swap object if necessary
8e7c4729 2044 *
b12defdc 2045 * The caller must hold the object.
96adc753 2046 */
984263bc 2047static void
96adc753
MD
2048swp_pager_meta_convert(vm_object_t object)
2049{
2050 if (object->type == OBJT_DEFAULT) {
984263bc 2051 object->type = OBJT_SWAP;
96adc753 2052 KKASSERT(object->swblock_count == 0);
984263bc 2053 }
96adc753
MD
2054}
2055
2056/*
2057 * SWP_PAGER_META_BUILD() - add swap block to swap meta data for object
2058 *
2059 * We first convert the object to a swap object if it is a default
2060 * object. Vnode objects do not need to be converted.
2061 *
2062 * The specified swapblk is added to the object's swap metadata. If
2063 * the swapblk is not valid, it is freed instead. Any previously
2064 * assigned swapblk is freed.
8e7c4729 2065 *
b12defdc 2066 * The caller must hold the object.
96adc753
MD
2067 */
2068static void
651d8e75 2069swp_pager_meta_build(vm_object_t object, vm_pindex_t index, swblk_t swapblk)
96adc753
MD
2070{
2071 struct swblock *swap;
2072 struct swblock *oswap;
0301c407 2073 vm_pindex_t v;
96adc753
MD
2074
2075 KKASSERT(swapblk != SWAPBLK_NONE);
b12defdc 2076 ASSERT_LWKT_TOKEN_HELD(vm_object_token(object));
96adc753
MD
2077
2078 /*
2079 * Convert object if necessary
2080 */
2081 if (object->type == OBJT_DEFAULT)
2082 swp_pager_meta_convert(object);
984263bc
MD
2083
2084 /*
96adc753 2085 * Locate swblock. If not found create, but if we aren't adding
984263bc
MD
2086 * anything just return. If we run out of space in the map we wait
2087 * and, since the hash table may have changed, retry.
2088 */
984263bc 2089retry:
96adc753 2090 swap = swp_pager_lookup(object, index);
984263bc 2091
96adc753 2092 if (swap == NULL) {
984263bc
MD
2093 int i;
2094
96adc753 2095 swap = zalloc(swap_zone);
984263bc 2096 if (swap == NULL) {
4ecf7cc9 2097 vm_wait(0);
984263bc
MD
2098 goto retry;
2099 }
0301c407 2100 swap->swb_index = index & ~(vm_pindex_t)SWAP_META_MASK;
984263bc
MD
2101 swap->swb_count = 0;
2102
96adc753 2103 ++object->swblock_count;
984263bc
MD
2104
2105 for (i = 0; i < SWAP_META_PAGES; ++i)
2106 swap->swb_pages[i] = SWAPBLK_NONE;
96adc753
MD
2107 oswap = RB_INSERT(swblock_rb_tree, &object->swblock_root, swap);
2108 KKASSERT(oswap == NULL);
984263bc
MD
2109 }
2110
2111 /*
0301c407
MD
2112 * Delete prior contents of metadata.
2113 *
2114 * NOTE: Decrement swb_count after the freeing operation (which
2115 * might block) to prevent racing destruction of the swblock.
984263bc 2116 */
984263bc
MD
2117 index &= SWAP_META_MASK;
2118
0301c407
MD
2119 while ((v = swap->swb_pages[index]) != SWAPBLK_NONE) {
2120 swap->swb_pages[index] = SWAPBLK_NONE;
2121 /* can block */
2122 swp_pager_freeswapspace(object, v, 1);
984263bc
MD
2123 --swap->swb_count;
2124 }
2125
2126 /*
2127 * Enter block into metadata
2128 */
984263bc
MD
2129 swap->swb_pages[index] = swapblk;
2130 if (swapblk != SWAPBLK_NONE)
2131 ++swap->swb_count;
2132}
2133
2134/*
2135 * SWP_PAGER_META_FREE() - free a range of blocks in the object's swap metadata
2136 *
2137 * The requested range of blocks is freed, with any associated swap
2138 * returned to the swap bitmap.
2139 *
2140 * This routine will free swap metadata structures as they are cleaned
2141 * out. This routine does *NOT* operate on swap metadata associated
2142 * with resident pages.
2143 *
b12defdc 2144 * The caller must hold the object.
984263bc 2145 */
8d292090
MD
2146static int swp_pager_meta_free_callback(struct swblock *swb, void *data);
2147
984263bc 2148static void
8d292090 2149swp_pager_meta_free(vm_object_t object, vm_pindex_t index, vm_pindex_t count)
984263bc 2150{
8d292090 2151 struct swfreeinfo info;
96adc753 2152
b12defdc
MD
2153 ASSERT_LWKT_TOKEN_HELD(vm_object_token(object));
2154
8d292090
MD
2155 /*
2156 * Nothing to do
2157 */
2158 if (object->swblock_count == 0) {
2159 KKASSERT(RB_EMPTY(&object->swblock_root));
2160 return;
2161 }
2162 if (count == 0)
984263bc
MD
2163 return;
2164
8d292090
MD
2165 /*
2166 * Setup for RB tree scan. Note that the pindex range can be huge
2167 * due to the 64 bit page index space so we cannot safely iterate.
2168 */
2169 info.object = object;
0301c407 2170 info.basei = index & ~(vm_pindex_t)SWAP_META_MASK;
8d292090
MD
2171 info.begi = index;
2172 info.endi = index + count - 1;
2173 swblock_rb_tree_RB_SCAN(&object->swblock_root, rb_swblock_scancmp,
2174 swp_pager_meta_free_callback, &info);
2175}
984263bc 2176
8e7c4729 2177/*
b12defdc 2178 * The caller must hold the object.
8e7c4729 2179 */
8d292090
MD
2180static
2181int
2182swp_pager_meta_free_callback(struct swblock *swap, void *data)
2183{
2184 struct swfreeinfo *info = data;
2185 vm_object_t object = info->object;
2186 int index;
2187 int eindex;
2188
2189 /*
2190 * Figure out the range within the swblock. The wider scan may
2191 * return edge-case swap blocks when the start and/or end points
2192 * are in the middle of a block.
2193 */
2194 if (swap->swb_index < info->begi)
2195 index = (int)info->begi & SWAP_META_MASK;
2196 else
2197 index = 0;
2198
2199 if (swap->swb_index + SWAP_META_PAGES > info->endi)
2200 eindex = (int)info->endi & SWAP_META_MASK;
2201 else
2202 eindex = SWAP_META_MASK;
2203
2204 /*
2205 * Scan and free the blocks. The loop terminates early
2206 * if (swap) runs out of blocks and could be freed.
0301c407
MD
2207 *
2208 * NOTE: Decrement swb_count after swp_pager_freeswapspace()
2209 * to deal with a zfree race.
8d292090
MD
2210 */
2211 while (index <= eindex) {
651d8e75 2212 swblk_t v = swap->swb_pages[index];
8d292090
MD
2213
2214 if (v != SWAPBLK_NONE) {
8d292090 2215 swap->swb_pages[index] = SWAPBLK_NONE;
0301c407
MD
2216 /* can block */
2217 swp_pager_freeswapspace(object, v, 1);
8d292090
MD
2218 if (--swap->swb_count == 0) {
2219 swp_pager_remove(object, swap);
2220 zfree(swap_zone, swap);
2221 --object->swblock_count;
2222 break;
984263bc 2223 }
984263bc 2224 }
8d292090 2225 ++index;
984263bc 2226 }
8d292090
MD
2227 /* swap may be invalid here due to zfree above */
2228 return(0);
984263bc
MD
2229}
2230
2231/*
2232 * SWP_PAGER_META_FREE_ALL() - destroy all swap metadata associated with object
2233 *
2234 * This routine locates and destroys all swap metadata associated with
2235 * an object.
2236 *
0301c407
MD
2237 * NOTE: Decrement swb_count after the freeing operation (which
2238 * might block) to prevent racing destruction of the swblock.
2239 *
b12defdc 2240 * The caller must hold the object.
984263bc 2241 */
984263bc
MD
2242static void
2243swp_pager_meta_free_all(vm_object_t object)
2244{
96adc753
MD
2245 struct swblock *swap;
2246 int i;
984263bc 2247
b12defdc
MD
2248 ASSERT_LWKT_TOKEN_HELD(vm_object_token(object));
2249
96adc753
MD
2250 while ((swap = RB_ROOT(&object->swblock_root)) != NULL) {
2251 swp_pager_remove(object, swap);
2252 for (i = 0; i < SWAP_META_PAGES; ++i) {
651d8e75 2253 swblk_t v = swap->swb_pages[i];
96adc753 2254 if (v != SWAPBLK_NONE) {
0301c407 2255 /* can block */
096e95c0 2256 swp_pager_freeswapspace(object, v, 1);
0301c407 2257 --swap->swb_count;
984263bc 2258 }
984263bc 2259 }
96adc753
MD
2260 if (swap->swb_count != 0)
2261 panic("swap_pager_meta_free_all: swb_count != 0");
2262 zfree(swap_zone, swap);
2263 --object->swblock_count;
984263bc 2264 }
96adc753 2265 KKASSERT(object->swblock_count == 0);
984263bc
MD
2266}
2267
2268/*
2269 * SWP_PAGER_METACTL() - misc control of swap and vm_page_t meta data.
2270 *
2271 * This routine is capable of looking up, popping, or freeing
2272 * swapblk assignments in the swap meta data or in the vm_page_t.
2273 * The routine typically returns the swapblk being looked-up, or popped,
2274 * or SWAPBLK_NONE if the block was freed, or SWAPBLK_NONE if the block
2275 * was invalid. This routine will automatically free any invalid
2276 * meta-data swapblks.
2277 *
2278 * It is not possible to store invalid swapblks in the swap meta data
2279 * (other then a literal 'SWAPBLK_NONE'), so we don't bother checking.
2280 *
2281 * When acting on a busy resident page and paging is in progress, we
2282 * have to wait until paging is complete but otherwise can act on the
2283 * busy page.
2284 *
984263bc
MD
2285 * SWM_FREE remove and free swap block from metadata
2286 * SWM_POP remove from meta data but do not free.. pop it out
8e7c4729 2287 *
b12defdc 2288 * The caller must hold the object.
984263bc 2289 */
651d8e75 2290static swblk_t
96adc753
MD
2291swp_pager_meta_ctl(vm_object_t object, vm_pindex_t index, int flags)
2292{
984263bc 2293 struct swblock *swap;
651d8e75 2294 swblk_t r1;
984263bc 2295
8d292090 2296 if (object->swblock_count == 0)
984263bc
MD
2297 return(SWAPBLK_NONE);
2298
2299 r1 = SWAPBLK_NONE;
96adc753 2300 swap = swp_pager_lookup(object, index);
984263bc 2301
96adc753 2302 if (swap != NULL) {
984263bc
MD
2303 index &= SWAP_META_MASK;
2304 r1 = swap->swb_pages[index];
2305
2306 if (r1 != SWAPBLK_NONE) {
984263bc
MD
2307 if (flags & (SWM_FREE|SWM_POP)) {
2308 swap->swb_pages[index] = SWAPBLK_NONE;
2309 if (--swap->swb_count == 0) {
96adc753 2310 swp_pager_remove(object, swap);
984263bc 2311 zfree(swap_zone, swap);
96adc753 2312 --object->swblock_count;
984263bc
MD
2313 }
2314 }
0301c407 2315 /* swap ptr may be invalid */
b12defdc
MD
2316 if (flags & SWM_FREE) {
2317 swp_pager_freeswapspace(object, r1, 1);
2318 r1 = SWAPBLK_NONE;
2319 }
984263bc 2320 }
0301c407 2321 /* swap ptr may be invalid */
984263bc
MD
2322 }
2323 return(r1);
2324}