FreeBSD/Linux Kernel Cross Reference
sys/vm/swap_pager.c
1 /*-
2 * SPDX-License-Identifier: BSD-4-Clause
3 *
4 * Copyright (c) 1998 Matthew Dillon,
5 * Copyright (c) 1994 John S. Dyson
6 * Copyright (c) 1990 University of Utah.
7 * Copyright (c) 1982, 1986, 1989, 1993
8 * The Regents of the University of California. All rights reserved.
9 *
10 * This code is derived from software contributed to Berkeley by
11 * the Systems Programming Group of the University of Utah Computer
12 * Science Department.
13 *
14 * Redistribution and use in source and binary forms, with or without
15 * modification, are permitted provided that the following conditions
16 * are met:
17 * 1. Redistributions of source code must retain the above copyright
18 * notice, this list of conditions and the following disclaimer.
19 * 2. Redistributions in binary form must reproduce the above copyright
20 * notice, this list of conditions and the following disclaimer in the
21 * documentation and/or other materials provided with the distribution.
22 * 3. All advertising materials mentioning features or use of this software
23 * must display the following acknowledgement:
24 * This product includes software developed by the University of
25 * California, Berkeley and its contributors.
26 * 4. Neither the name of the University nor the names of its contributors
27 * may be used to endorse or promote products derived from this software
28 * without specific prior written permission.
29 *
30 * THIS SOFTWARE IS PROVIDED BY THE REGENTS AND CONTRIBUTORS ``AS IS'' AND
31 * ANY EXPRESS OR IMPLIED WARRANTIES, INCLUDING, BUT NOT LIMITED TO, THE
32 * IMPLIED WARRANTIES OF MERCHANTABILITY AND FITNESS FOR A PARTICULAR PURPOSE
33 * ARE DISCLAIMED. IN NO EVENT SHALL THE REGENTS OR CONTRIBUTORS BE LIABLE
34 * FOR ANY DIRECT, INDIRECT, INCIDENTAL, SPECIAL, EXEMPLARY, OR CONSEQUENTIAL
35 * DAMAGES (INCLUDING, BUT NOT LIMITED TO, PROCUREMENT OF SUBSTITUTE GOODS
36 * OR SERVICES; LOSS OF USE, DATA, OR PROFITS; OR BUSINESS INTERRUPTION)
37 * HOWEVER CAUSED AND ON ANY THEORY OF LIABILITY, WHETHER IN CONTRACT, STRICT
38 * LIABILITY, OR TORT (INCLUDING NEGLIGENCE OR OTHERWISE) ARISING IN ANY WAY
39 * OUT OF THE USE OF THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF
40 * SUCH DAMAGE.
41 *
42 * New Swap System
43 * Matthew Dillon
44 *
45 * Radix Bitmap 'blists'.
46 *
47 * - The new swapper uses the new radix bitmap code. This should scale
48 * to arbitrarily small or arbitrarily large swap spaces and an almost
49 * arbitrary degree of fragmentation.
50 *
51 * Features:
52 *
53 * - on the fly reallocation of swap during putpages. The new system
54 * does not try to keep previously allocated swap blocks for dirty
55 * pages.
56 *
57 * - on the fly deallocation of swap
58 *
59 * - No more garbage collection required. Unnecessarily allocated swap
60 * blocks only exist for dirty vm_page_t's now and these are already
61 * cycled (in a high-load system) by the pager. We also do on-the-fly
62 * removal of invalidated swap blocks when a page is destroyed
63 * or renamed.
64 *
65 * from: Utah $Hdr: swap_pager.c 1.4 91/04/30$
66 *
67 * @(#)swap_pager.c 8.9 (Berkeley) 3/21/94
68 * @(#)vm_swap.c 8.5 (Berkeley) 2/17/94
69 */
70
71 #include <sys/cdefs.h>
72 __FBSDID("$FreeBSD$");
73
74 #include "opt_vm.h"
75
76 #include <sys/param.h>
77 #include <sys/bio.h>
78 #include <sys/blist.h>
79 #include <sys/buf.h>
80 #include <sys/conf.h>
81 #include <sys/disk.h>
82 #include <sys/disklabel.h>
83 #include <sys/eventhandler.h>
84 #include <sys/fcntl.h>
85 #include <sys/lock.h>
86 #include <sys/kernel.h>
87 #include <sys/mount.h>
88 #include <sys/namei.h>
89 #include <sys/malloc.h>
90 #include <sys/pctrie.h>
91 #include <sys/priv.h>
92 #include <sys/proc.h>
93 #include <sys/racct.h>
94 #include <sys/resource.h>
95 #include <sys/resourcevar.h>
96 #include <sys/rwlock.h>
97 #include <sys/sbuf.h>
98 #include <sys/sysctl.h>
99 #include <sys/sysproto.h>
100 #include <sys/systm.h>
101 #include <sys/sx.h>
102 #include <sys/vmmeter.h>
103 #include <sys/vnode.h>
104
105 #include <security/mac/mac_framework.h>
106
107 #include <vm/vm.h>
108 #include <vm/pmap.h>
109 #include <vm/vm_map.h>
110 #include <vm/vm_kern.h>
111 #include <vm/vm_object.h>
112 #include <vm/vm_page.h>
113 #include <vm/vm_pager.h>
114 #include <vm/vm_pageout.h>
115 #include <vm/vm_param.h>
116 #include <vm/swap_pager.h>
117 #include <vm/vm_extern.h>
118 #include <vm/uma.h>
119
120 #include <geom/geom.h>
121
122 /*
123 * MAX_PAGEOUT_CLUSTER must be a power of 2 between 1 and 64.
124 * The 64-page limit is due to the radix code (kern/subr_blist.c).
125 */
126 #ifndef MAX_PAGEOUT_CLUSTER
127 #define MAX_PAGEOUT_CLUSTER 32
128 #endif
129
130 #if !defined(SWB_NPAGES)
131 #define SWB_NPAGES MAX_PAGEOUT_CLUSTER
132 #endif
133
134 #define SWAP_META_PAGES PCTRIE_COUNT
135
136 /*
137 * A swblk structure maps each page index within a
138 * SWAP_META_PAGES-aligned and sized range to the address of an
139 * on-disk swap block (or SWAPBLK_NONE). The collection of these
140 * mappings for an entire vm object is implemented as a pc-trie.
141 */
142 struct swblk {
143 vm_pindex_t p;
144 daddr_t d[SWAP_META_PAGES];
145 };
146
147 static MALLOC_DEFINE(M_VMPGDATA, "vm_pgdata", "swap pager private data");
148 static struct mtx sw_dev_mtx;
149 static TAILQ_HEAD(, swdevt) swtailq = TAILQ_HEAD_INITIALIZER(swtailq);
150 static struct swdevt *swdevhd; /* Allocate from here next */
151 static int nswapdev; /* Number of swap devices */
152 int swap_pager_avail;
153 static struct sx swdev_syscall_lock; /* serialize swap(on|off) */
154
155 static __exclusive_cache_line u_long swap_reserved;
156 static u_long swap_total;
157 static int sysctl_page_shift(SYSCTL_HANDLER_ARGS);
158
159 static SYSCTL_NODE(_vm_stats, OID_AUTO, swap, CTLFLAG_RD | CTLFLAG_MPSAFE, 0,
160 "VM swap stats");
161
162 SYSCTL_PROC(_vm, OID_AUTO, swap_reserved, CTLTYPE_U64 | CTLFLAG_RD | CTLFLAG_MPSAFE,
163 &swap_reserved, 0, sysctl_page_shift, "A",
164 "Amount of swap storage needed to back all allocated anonymous memory.");
165 SYSCTL_PROC(_vm, OID_AUTO, swap_total, CTLTYPE_U64 | CTLFLAG_RD | CTLFLAG_MPSAFE,
166 &swap_total, 0, sysctl_page_shift, "A",
167 "Total amount of available swap storage.");
168
169 static int overcommit = 0;
170 SYSCTL_INT(_vm, VM_OVERCOMMIT, overcommit, CTLFLAG_RW, &overcommit, 0,
171 "Configure virtual memory overcommit behavior. See tuning(7) "
172 "for details.");
173 static unsigned long swzone;
174 SYSCTL_ULONG(_vm, OID_AUTO, swzone, CTLFLAG_RD, &swzone, 0,
175 "Actual size of swap metadata zone");
176 static unsigned long swap_maxpages;
177 SYSCTL_ULONG(_vm, OID_AUTO, swap_maxpages, CTLFLAG_RD, &swap_maxpages, 0,
178 "Maximum amount of swap supported");
179
180 static COUNTER_U64_DEFINE_EARLY(swap_free_deferred);
181 SYSCTL_COUNTER_U64(_vm_stats_swap, OID_AUTO, free_deferred,
182 CTLFLAG_RD, &swap_free_deferred,
183 "Number of pages that deferred freeing swap space");
184
185 static COUNTER_U64_DEFINE_EARLY(swap_free_completed);
186 SYSCTL_COUNTER_U64(_vm_stats_swap, OID_AUTO, free_completed,
187 CTLFLAG_RD, &swap_free_completed,
188 "Number of deferred frees completed");
189
190 /* bits from overcommit */
191 #define SWAP_RESERVE_FORCE_ON (1 << 0)
192 #define SWAP_RESERVE_RLIMIT_ON (1 << 1)
193 #define SWAP_RESERVE_ALLOW_NONWIRED (1 << 2)
194
195 static int
196 sysctl_page_shift(SYSCTL_HANDLER_ARGS)
197 {
198 uint64_t newval;
199 u_long value = *(u_long *)arg1;
200
201 newval = ((uint64_t)value) << PAGE_SHIFT;
202 return (sysctl_handle_64(oidp, &newval, 0, req));
203 }
204
205 static bool
206 swap_reserve_by_cred_rlimit(u_long pincr, struct ucred *cred, int oc)
207 {
208 struct uidinfo *uip;
209 u_long prev;
210
211 uip = cred->cr_ruidinfo;
212
213 prev = atomic_fetchadd_long(&uip->ui_vmsize, pincr);
214 if ((oc & SWAP_RESERVE_RLIMIT_ON) != 0 &&
215 prev + pincr > lim_cur(curthread, RLIMIT_SWAP) &&
216 priv_check(curthread, PRIV_VM_SWAP_NORLIMIT) != 0) {
217 prev = atomic_fetchadd_long(&uip->ui_vmsize, -pincr);
218 KASSERT(prev >= pincr, ("negative vmsize for uid = %d\n", uip->ui_uid));
219 return (false);
220 }
221 return (true);
222 }
223
224 static void
225 swap_release_by_cred_rlimit(u_long pdecr, struct ucred *cred)
226 {
227 struct uidinfo *uip;
228 #ifdef INVARIANTS
229 u_long prev;
230 #endif
231
232 uip = cred->cr_ruidinfo;
233
234 #ifdef INVARIANTS
235 prev = atomic_fetchadd_long(&uip->ui_vmsize, -pdecr);
236 KASSERT(prev >= pdecr, ("negative vmsize for uid = %d\n", uip->ui_uid));
237 #else
238 atomic_subtract_long(&uip->ui_vmsize, pdecr);
239 #endif
240 }
241
242 static void
243 swap_reserve_force_rlimit(u_long pincr, struct ucred *cred)
244 {
245 struct uidinfo *uip;
246
247 uip = cred->cr_ruidinfo;
248 atomic_add_long(&uip->ui_vmsize, pincr);
249 }
250
251 bool
252 swap_reserve(vm_ooffset_t incr)
253 {
254
255 return (swap_reserve_by_cred(incr, curthread->td_ucred));
256 }
257
258 bool
259 swap_reserve_by_cred(vm_ooffset_t incr, struct ucred *cred)
260 {
261 u_long r, s, prev, pincr;
262 #ifdef RACCT
263 int error;
264 #endif
265 int oc;
266 static int curfail;
267 static struct timeval lastfail;
268
269 KASSERT((incr & PAGE_MASK) == 0, ("%s: incr: %ju & PAGE_MASK", __func__,
270 (uintmax_t)incr));
271
272 #ifdef RACCT
273 if (RACCT_ENABLED()) {
274 PROC_LOCK(curproc);
275 error = racct_add(curproc, RACCT_SWAP, incr);
276 PROC_UNLOCK(curproc);
277 if (error != 0)
278 return (false);
279 }
280 #endif
281
282 pincr = atop(incr);
283 prev = atomic_fetchadd_long(&swap_reserved, pincr);
284 r = prev + pincr;
285 s = swap_total;
286 oc = atomic_load_int(&overcommit);
287 if (r > s && (oc & SWAP_RESERVE_ALLOW_NONWIRED) != 0) {
288 s += vm_cnt.v_page_count - vm_cnt.v_free_reserved -
289 vm_wire_count();
290 }
291 if ((oc & SWAP_RESERVE_FORCE_ON) != 0 && r > s &&
292 priv_check(curthread, PRIV_VM_SWAP_NOQUOTA) != 0) {
293 prev = atomic_fetchadd_long(&swap_reserved, -pincr);
294 KASSERT(prev >= pincr, ("swap_reserved < incr on overcommit fail"));
295 goto out_error;
296 }
297
298 if (!swap_reserve_by_cred_rlimit(pincr, cred, oc)) {
299 prev = atomic_fetchadd_long(&swap_reserved, -pincr);
300 KASSERT(prev >= pincr, ("swap_reserved < incr on overcommit fail"));
301 goto out_error;
302 }
303
304 return (true);
305
306 out_error:
307 if (ppsratecheck(&lastfail, &curfail, 1)) {
308 printf("uid %d, pid %d: swap reservation for %jd bytes failed\n",
309 cred->cr_ruidinfo->ui_uid, curproc->p_pid, incr);
310 }
311 #ifdef RACCT
312 if (RACCT_ENABLED()) {
313 PROC_LOCK(curproc);
314 racct_sub(curproc, RACCT_SWAP, incr);
315 PROC_UNLOCK(curproc);
316 }
317 #endif
318
319 return (false);
320 }
321
322 void
323 swap_reserve_force(vm_ooffset_t incr)
324 {
325 u_long pincr;
326
327 KASSERT((incr & PAGE_MASK) == 0, ("%s: incr: %ju & PAGE_MASK", __func__,
328 (uintmax_t)incr));
329
330 #ifdef RACCT
331 if (RACCT_ENABLED()) {
332 PROC_LOCK(curproc);
333 racct_add_force(curproc, RACCT_SWAP, incr);
334 PROC_UNLOCK(curproc);
335 }
336 #endif
337 pincr = atop(incr);
338 atomic_add_long(&swap_reserved, pincr);
339 swap_reserve_force_rlimit(pincr, curthread->td_ucred);
340 }
341
342 void
343 swap_release(vm_ooffset_t decr)
344 {
345 struct ucred *cred;
346
347 PROC_LOCK(curproc);
348 cred = curproc->p_ucred;
349 swap_release_by_cred(decr, cred);
350 PROC_UNLOCK(curproc);
351 }
352
353 void
354 swap_release_by_cred(vm_ooffset_t decr, struct ucred *cred)
355 {
356 u_long pdecr;
357 #ifdef INVARIANTS
358 u_long prev;
359 #endif
360
361 KASSERT((decr & PAGE_MASK) == 0, ("%s: decr: %ju & PAGE_MASK", __func__,
362 (uintmax_t)decr));
363
364 pdecr = atop(decr);
365 #ifdef INVARIANTS
366 prev = atomic_fetchadd_long(&swap_reserved, -pdecr);
367 KASSERT(prev >= pdecr, ("swap_reserved < decr"));
368 #else
369 atomic_subtract_long(&swap_reserved, pdecr);
370 #endif
371
372 swap_release_by_cred_rlimit(pdecr, cred);
373 #ifdef RACCT
374 if (racct_enable)
375 racct_sub_cred(cred, RACCT_SWAP, decr);
376 #endif
377 }
378
379 static int swap_pager_full = 2; /* swap space exhaustion (task killing) */
380 static int swap_pager_almost_full = 1; /* swap space exhaustion (w/hysteresis)*/
381 static struct mtx swbuf_mtx; /* to sync nsw_wcount_async */
382 static int nsw_wcount_async; /* limit async write buffers */
383 static int nsw_wcount_async_max;/* assigned maximum */
384 static int nsw_cluster_max; /* maximum VOP I/O allowed */
385
386 static int sysctl_swap_async_max(SYSCTL_HANDLER_ARGS);
387 SYSCTL_PROC(_vm, OID_AUTO, swap_async_max, CTLTYPE_INT | CTLFLAG_RW |
388 CTLFLAG_MPSAFE, NULL, 0, sysctl_swap_async_max, "I",
389 "Maximum running async swap ops");
390 static int sysctl_swap_fragmentation(SYSCTL_HANDLER_ARGS);
391 SYSCTL_PROC(_vm, OID_AUTO, swap_fragmentation, CTLTYPE_STRING | CTLFLAG_RD |
392 CTLFLAG_MPSAFE, NULL, 0, sysctl_swap_fragmentation, "A",
393 "Swap Fragmentation Info");
394
395 static struct sx sw_alloc_sx;
396
397 /*
398 * "named" and "unnamed" anon region objects. Try to reduce the overhead
399 * of searching a named list by hashing it just a little.
400 */
401
402 #define NOBJLISTS 8
403
404 #define NOBJLIST(handle) \
405 (&swap_pager_object_list[((int)(intptr_t)handle >> 4) & (NOBJLISTS-1)])
406
407 static struct pagerlst swap_pager_object_list[NOBJLISTS];
408 static uma_zone_t swwbuf_zone;
409 static uma_zone_t swrbuf_zone;
410 static uma_zone_t swblk_zone;
411 static uma_zone_t swpctrie_zone;
412
413 /*
414 * pagerops for OBJT_SWAP - "swap pager". Some ops are also global procedure
415 * calls hooked from other parts of the VM system and do not appear here.
416 * (see vm/swap_pager.h).
417 */
418 static vm_object_t
419 swap_pager_alloc(void *handle, vm_ooffset_t size,
420 vm_prot_t prot, vm_ooffset_t offset, struct ucred *);
421 static void swap_pager_dealloc(vm_object_t object);
422 static int swap_pager_getpages(vm_object_t, vm_page_t *, int, int *,
423 int *);
424 static int swap_pager_getpages_async(vm_object_t, vm_page_t *, int, int *,
425 int *, pgo_getpages_iodone_t, void *);
426 static void swap_pager_putpages(vm_object_t, vm_page_t *, int, boolean_t, int *);
427 static boolean_t
428 swap_pager_haspage(vm_object_t object, vm_pindex_t pindex, int *before, int *after);
429 static void swap_pager_init(void);
430 static void swap_pager_unswapped(vm_page_t);
431 static void swap_pager_swapoff(struct swdevt *sp);
432 static void swap_pager_update_writecount(vm_object_t object,
433 vm_offset_t start, vm_offset_t end);
434 static void swap_pager_release_writecount(vm_object_t object,
435 vm_offset_t start, vm_offset_t end);
436
437 struct pagerops swappagerops = {
438 .pgo_init = swap_pager_init, /* early system initialization of pager */
439 .pgo_alloc = swap_pager_alloc, /* allocate an OBJT_SWAP object */
440 .pgo_dealloc = swap_pager_dealloc, /* deallocate an OBJT_SWAP object */
441 .pgo_getpages = swap_pager_getpages, /* pagein */
442 .pgo_getpages_async = swap_pager_getpages_async, /* pagein (async) */
443 .pgo_putpages = swap_pager_putpages, /* pageout */
444 .pgo_haspage = swap_pager_haspage, /* get backing store status for page */
445 .pgo_pageunswapped = swap_pager_unswapped, /* remove swap related to page */
446 .pgo_update_writecount = swap_pager_update_writecount,
447 .pgo_release_writecount = swap_pager_release_writecount,
448 };
449
450 /*
451 * swap_*() routines are externally accessible. swp_*() routines are
452 * internal.
453 */
454 static int nswap_lowat = 128; /* in pages, swap_pager_almost_full warn */
455 static int nswap_hiwat = 512; /* in pages, swap_pager_almost_full warn */
456
457 SYSCTL_INT(_vm, OID_AUTO, dmmax, CTLFLAG_RD, &nsw_cluster_max, 0,
458 "Maximum size of a swap block in pages");
459
460 static void swp_sizecheck(void);
461 static void swp_pager_async_iodone(struct buf *bp);
462 static bool swp_pager_swblk_empty(struct swblk *sb, int start, int limit);
463 static void swp_pager_free_empty_swblk(vm_object_t, struct swblk *sb);
464 static int swapongeom(struct vnode *);
465 static int swaponvp(struct thread *, struct vnode *, u_long);
466 static int swapoff_one(struct swdevt *sp, struct ucred *cred);
467
468 /*
469 * Swap bitmap functions
470 */
471 static void swp_pager_freeswapspace(daddr_t blk, daddr_t npages);
472 static daddr_t swp_pager_getswapspace(int *npages);
473
474 /*
475 * Metadata functions
476 */
477 static daddr_t swp_pager_meta_build(vm_object_t, vm_pindex_t, daddr_t);
478 static void swp_pager_meta_free(vm_object_t, vm_pindex_t, vm_pindex_t);
479 static void swp_pager_meta_transfer(vm_object_t src, vm_object_t dst,
480 vm_pindex_t pindex, vm_pindex_t count);
481 static void swp_pager_meta_free_all(vm_object_t);
482 static daddr_t swp_pager_meta_lookup(vm_object_t, vm_pindex_t);
483
484 static void
485 swp_pager_init_freerange(daddr_t *start, daddr_t *num)
486 {
487
488 *start = SWAPBLK_NONE;
489 *num = 0;
490 }
491
492 static void
493 swp_pager_update_freerange(daddr_t *start, daddr_t *num, daddr_t addr)
494 {
495
496 if (*start + *num == addr) {
497 (*num)++;
498 } else {
499 swp_pager_freeswapspace(*start, *num);
500 *start = addr;
501 *num = 1;
502 }
503 }
504
505 static void *
506 swblk_trie_alloc(struct pctrie *ptree)
507 {
508
509 return (uma_zalloc(swpctrie_zone, M_NOWAIT | (curproc == pageproc ?
510 M_USE_RESERVE : 0)));
511 }
512
513 static void
514 swblk_trie_free(struct pctrie *ptree, void *node)
515 {
516
517 uma_zfree(swpctrie_zone, node);
518 }
519
520 PCTRIE_DEFINE(SWAP, swblk, p, swblk_trie_alloc, swblk_trie_free);
521
522 /*
523 * SWP_SIZECHECK() - update swap_pager_full indication
524 *
525 * update the swap_pager_almost_full indication and warn when we are
526 * about to run out of swap space, using lowat/hiwat hysteresis.
527 *
528 * Clear swap_pager_full ( task killing ) indication when lowat is met.
529 *
530 * No restrictions on call
531 * This routine may not block.
532 */
533 static void
534 swp_sizecheck(void)
535 {
536
537 if (swap_pager_avail < nswap_lowat) {
538 if (swap_pager_almost_full == 0) {
539 printf("swap_pager: out of swap space\n");
540 swap_pager_almost_full = 1;
541 }
542 } else {
543 swap_pager_full = 0;
544 if (swap_pager_avail > nswap_hiwat)
545 swap_pager_almost_full = 0;
546 }
547 }
548
549 /*
550 * SWAP_PAGER_INIT() - initialize the swap pager!
551 *
552 * Expected to be started from system init. NOTE: This code is run
553 * before much else so be careful what you depend on. Most of the VM
554 * system has yet to be initialized at this point.
555 */
556 static void
557 swap_pager_init(void)
558 {
559 /*
560 * Initialize object lists
561 */
562 int i;
563
564 for (i = 0; i < NOBJLISTS; ++i)
565 TAILQ_INIT(&swap_pager_object_list[i]);
566 mtx_init(&sw_dev_mtx, "swapdev", NULL, MTX_DEF);
567 sx_init(&sw_alloc_sx, "swspsx");
568 sx_init(&swdev_syscall_lock, "swsysc");
569 }
570
571 /*
572 * SWAP_PAGER_SWAP_INIT() - swap pager initialization from pageout process
573 *
574 * Expected to be started from pageout process once, prior to entering
575 * its main loop.
576 */
577 void
578 swap_pager_swap_init(void)
579 {
580 unsigned long n, n2;
581
582 /*
583 * Number of in-transit swap bp operations. Don't
584 * exhaust the pbufs completely. Make sure we
585 * initialize workable values (0 will work for hysteresis
586 * but it isn't very efficient).
587 *
588 * The nsw_cluster_max is constrained by the bp->b_pages[]
589 * array, which has maxphys / PAGE_SIZE entries, and our locally
590 * defined MAX_PAGEOUT_CLUSTER. Also be aware that swap ops are
591 * constrained by the swap device interleave stripe size.
592 *
593 * Currently we hardwire nsw_wcount_async to 4. This limit is
594 * designed to prevent other I/O from having high latencies due to
595 * our pageout I/O. The value 4 works well for one or two active swap
596 * devices but is probably a little low if you have more. Even so,
597 * a higher value would probably generate only a limited improvement
598 * with three or four active swap devices since the system does not
599 * typically have to pageout at extreme bandwidths. We will want
600 * at least 2 per swap devices, and 4 is a pretty good value if you
601 * have one NFS swap device due to the command/ack latency over NFS.
602 * So it all works out pretty well.
603 */
604 nsw_cluster_max = min(maxphys / PAGE_SIZE, MAX_PAGEOUT_CLUSTER);
605
606 nsw_wcount_async = 4;
607 nsw_wcount_async_max = nsw_wcount_async;
608 mtx_init(&swbuf_mtx, "async swbuf mutex", NULL, MTX_DEF);
609
610 swwbuf_zone = pbuf_zsecond_create("swwbuf", nswbuf / 4);
611 swrbuf_zone = pbuf_zsecond_create("swrbuf", nswbuf / 2);
612
613 /*
614 * Initialize our zone, taking the user's requested size or
615 * estimating the number we need based on the number of pages
616 * in the system.
617 */
618 n = maxswzone != 0 ? maxswzone / sizeof(struct swblk) :
619 vm_cnt.v_page_count / 2;
620 swpctrie_zone = uma_zcreate("swpctrie", pctrie_node_size(), NULL, NULL,
621 pctrie_zone_init, NULL, UMA_ALIGN_PTR, 0);
622 if (swpctrie_zone == NULL)
623 panic("failed to create swap pctrie zone.");
624 swblk_zone = uma_zcreate("swblk", sizeof(struct swblk), NULL, NULL,
625 NULL, NULL, _Alignof(struct swblk) - 1, 0);
626 if (swblk_zone == NULL)
627 panic("failed to create swap blk zone.");
628 n2 = n;
629 do {
630 if (uma_zone_reserve_kva(swblk_zone, n))
631 break;
632 /*
633 * if the allocation failed, try a zone two thirds the
634 * size of the previous attempt.
635 */
636 n -= ((n + 2) / 3);
637 } while (n > 0);
638
639 /*
640 * Often uma_zone_reserve_kva() cannot reserve exactly the
641 * requested size. Account for the difference when
642 * calculating swap_maxpages.
643 */
644 n = uma_zone_get_max(swblk_zone);
645
646 if (n < n2)
647 printf("Swap blk zone entries changed from %lu to %lu.\n",
648 n2, n);
649 /* absolute maximum we can handle assuming 100% efficiency */
650 swap_maxpages = n * SWAP_META_PAGES;
651 swzone = n * sizeof(struct swblk);
652 if (!uma_zone_reserve_kva(swpctrie_zone, n))
653 printf("Cannot reserve swap pctrie zone, "
654 "reduce kern.maxswzone.\n");
655 }
656
657 static vm_object_t
658 swap_pager_alloc_init(void *handle, struct ucred *cred, vm_ooffset_t size,
659 vm_ooffset_t offset)
660 {
661 vm_object_t object;
662
663 if (cred != NULL) {
664 if (!swap_reserve_by_cred(size, cred))
665 return (NULL);
666 crhold(cred);
667 }
668
669 /*
670 * The un_pager.swp.swp_blks trie is initialized by
671 * vm_object_allocate() to ensure the correct order of
672 * visibility to other threads.
673 */
674 object = vm_object_allocate(OBJT_SWAP, OFF_TO_IDX(offset +
675 PAGE_MASK + size));
676
677 object->un_pager.swp.writemappings = 0;
678 object->handle = handle;
679 if (cred != NULL) {
680 object->cred = cred;
681 object->charge = size;
682 }
683 return (object);
684 }
685
686 /*
687 * SWAP_PAGER_ALLOC() - allocate a new OBJT_SWAP VM object and instantiate
688 * its metadata structures.
689 *
690 * This routine is called from the mmap and fork code to create a new
691 * OBJT_SWAP object.
692 *
693 * This routine must ensure that no live duplicate is created for
694 * the named object request, which is protected against by
695 * holding the sw_alloc_sx lock in case handle != NULL.
696 */
697 static vm_object_t
698 swap_pager_alloc(void *handle, vm_ooffset_t size, vm_prot_t prot,
699 vm_ooffset_t offset, struct ucred *cred)
700 {
701 vm_object_t object;
702
703 if (handle != NULL) {
704 /*
705 * Reference existing named region or allocate new one. There
706 * should not be a race here against swp_pager_meta_build()
707 * as called from vm_page_remove() in regards to the lookup
708 * of the handle.
709 */
710 sx_xlock(&sw_alloc_sx);
711 object = vm_pager_object_lookup(NOBJLIST(handle), handle);
712 if (object == NULL) {
713 object = swap_pager_alloc_init(handle, cred, size,
714 offset);
715 if (object != NULL) {
716 TAILQ_INSERT_TAIL(NOBJLIST(object->handle),
717 object, pager_object_list);
718 }
719 }
720 sx_xunlock(&sw_alloc_sx);
721 } else {
722 object = swap_pager_alloc_init(handle, cred, size, offset);
723 }
724 return (object);
725 }
726
727 /*
728 * SWAP_PAGER_DEALLOC() - remove swap metadata from object
729 *
730 * The swap backing for the object is destroyed. The code is
731 * designed such that we can reinstantiate it later, but this
732 * routine is typically called only when the entire object is
733 * about to be destroyed.
734 *
735 * The object must be locked.
736 */
737 static void
738 swap_pager_dealloc(vm_object_t object)
739 {
740
741 VM_OBJECT_ASSERT_WLOCKED(object);
742 KASSERT((object->flags & OBJ_DEAD) != 0, ("dealloc of reachable obj"));
743
744 /*
745 * Remove from list right away so lookups will fail if we block for
746 * pageout completion.
747 */
748 if ((object->flags & OBJ_ANON) == 0 && object->handle != NULL) {
749 VM_OBJECT_WUNLOCK(object);
750 sx_xlock(&sw_alloc_sx);
751 TAILQ_REMOVE(NOBJLIST(object->handle), object,
752 pager_object_list);
753 sx_xunlock(&sw_alloc_sx);
754 VM_OBJECT_WLOCK(object);
755 }
756
757 vm_object_pip_wait(object, "swpdea");
758
759 /*
760 * Free all remaining metadata. We only bother to free it from
761 * the swap meta data. We do not attempt to free swapblk's still
762 * associated with vm_page_t's for this object. We do not care
763 * if paging is still in progress on some objects.
764 */
765 swp_pager_meta_free_all(object);
766 object->handle = NULL;
767 object->type = OBJT_DEAD;
768 }
769
770 /************************************************************************
771 * SWAP PAGER BITMAP ROUTINES *
772 ************************************************************************/
773
774 /*
775 * SWP_PAGER_GETSWAPSPACE() - allocate raw swap space
776 *
777 * Allocate swap for up to the requested number of pages. The
778 * starting swap block number (a page index) is returned or
779 * SWAPBLK_NONE if the allocation failed.
780 *
781 * Also has the side effect of advising that somebody made a mistake
782 * when they configured swap and didn't configure enough.
783 *
784 * This routine may not sleep.
785 *
786 * We allocate in round-robin fashion from the configured devices.
787 */
788 static daddr_t
789 swp_pager_getswapspace(int *io_npages)
790 {
791 daddr_t blk;
792 struct swdevt *sp;
793 int mpages, npages;
794
795 KASSERT(*io_npages >= 1,
796 ("%s: npages not positive", __func__));
797 blk = SWAPBLK_NONE;
798 mpages = *io_npages;
799 npages = imin(BLIST_MAX_ALLOC, mpages);
800 mtx_lock(&sw_dev_mtx);
801 sp = swdevhd;
802 while (!TAILQ_EMPTY(&swtailq)) {
803 if (sp == NULL)
804 sp = TAILQ_FIRST(&swtailq);
805 if ((sp->sw_flags & SW_CLOSING) == 0)
806 blk = blist_alloc(sp->sw_blist, &npages, mpages);
807 if (blk != SWAPBLK_NONE)
808 break;
809 sp = TAILQ_NEXT(sp, sw_list);
810 if (swdevhd == sp) {
811 if (npages == 1)
812 break;
813 mpages = npages - 1;
814 npages >>= 1;
815 }
816 }
817 if (blk != SWAPBLK_NONE) {
818 *io_npages = npages;
819 blk += sp->sw_first;
820 sp->sw_used += npages;
821 swap_pager_avail -= npages;
822 swp_sizecheck();
823 swdevhd = TAILQ_NEXT(sp, sw_list);
824 } else {
825 if (swap_pager_full != 2) {
826 printf("swp_pager_getswapspace(%d): failed\n",
827 *io_npages);
828 swap_pager_full = 2;
829 swap_pager_almost_full = 1;
830 }
831 swdevhd = NULL;
832 }
833 mtx_unlock(&sw_dev_mtx);
834 return (blk);
835 }
836
837 static bool
838 swp_pager_isondev(daddr_t blk, struct swdevt *sp)
839 {
840
841 return (blk >= sp->sw_first && blk < sp->sw_end);
842 }
843
844 static void
845 swp_pager_strategy(struct buf *bp)
846 {
847 struct swdevt *sp;
848
849 mtx_lock(&sw_dev_mtx);
850 TAILQ_FOREACH(sp, &swtailq, sw_list) {
851 if (swp_pager_isondev(bp->b_blkno, sp)) {
852 mtx_unlock(&sw_dev_mtx);
853 if ((sp->sw_flags & SW_UNMAPPED) != 0 &&
854 unmapped_buf_allowed) {
855 bp->b_data = unmapped_buf;
856 bp->b_offset = 0;
857 } else {
858 pmap_qenter((vm_offset_t)bp->b_data,
859 &bp->b_pages[0], bp->b_bcount / PAGE_SIZE);
860 }
861 sp->sw_strategy(bp, sp);
862 return;
863 }
864 }
865 panic("Swapdev not found");
866 }
867
868 /*
869 * SWP_PAGER_FREESWAPSPACE() - free raw swap space
870 *
871 * This routine returns the specified swap blocks back to the bitmap.
872 *
873 * This routine may not sleep.
874 */
875 static void
876 swp_pager_freeswapspace(daddr_t blk, daddr_t npages)
877 {
878 struct swdevt *sp;
879
880 if (npages == 0)
881 return;
882 mtx_lock(&sw_dev_mtx);
883 TAILQ_FOREACH(sp, &swtailq, sw_list) {
884 if (swp_pager_isondev(blk, sp)) {
885 sp->sw_used -= npages;
886 /*
887 * If we are attempting to stop swapping on
888 * this device, we don't want to mark any
889 * blocks free lest they be reused.
890 */
891 if ((sp->sw_flags & SW_CLOSING) == 0) {
892 blist_free(sp->sw_blist, blk - sp->sw_first,
893 npages);
894 swap_pager_avail += npages;
895 swp_sizecheck();
896 }
897 mtx_unlock(&sw_dev_mtx);
898 return;
899 }
900 }
901 panic("Swapdev not found");
902 }
903
904 /*
905 * SYSCTL_SWAP_FRAGMENTATION() - produce raw swap space stats
906 */
907 static int
908 sysctl_swap_fragmentation(SYSCTL_HANDLER_ARGS)
909 {
910 struct sbuf sbuf;
911 struct swdevt *sp;
912 const char *devname;
913 int error;
914
915 error = sysctl_wire_old_buffer(req, 0);
916 if (error != 0)
917 return (error);
918 sbuf_new_for_sysctl(&sbuf, NULL, 128, req);
919 mtx_lock(&sw_dev_mtx);
920 TAILQ_FOREACH(sp, &swtailq, sw_list) {
921 if (vn_isdisk(sp->sw_vp))
922 devname = devtoname(sp->sw_vp->v_rdev);
923 else
924 devname = "[file]";
925 sbuf_printf(&sbuf, "\nFree space on device %s:\n", devname);
926 blist_stats(sp->sw_blist, &sbuf);
927 }
928 mtx_unlock(&sw_dev_mtx);
929 error = sbuf_finish(&sbuf);
930 sbuf_delete(&sbuf);
931 return (error);
932 }
933
934 /*
935 * SWAP_PAGER_FREESPACE() - frees swap blocks associated with a page
936 * range within an object.
937 *
938 * This is a globally accessible routine.
939 *
940 * This routine removes swapblk assignments from swap metadata.
941 *
942 * The external callers of this routine typically have already destroyed
943 * or renamed vm_page_t's associated with this range in the object so
944 * we should be ok.
945 *
946 * The object must be locked.
947 */
948 void
949 swap_pager_freespace(vm_object_t object, vm_pindex_t start, vm_size_t size)
950 {
951
952 swp_pager_meta_free(object, start, size);
953 }
954
955 /*
956 * SWAP_PAGER_RESERVE() - reserve swap blocks in object
957 *
958 * Assigns swap blocks to the specified range within the object. The
959 * swap blocks are not zeroed. Any previous swap assignment is destroyed.
960 *
961 * Returns 0 on success, -1 on failure.
962 */
963 int
964 swap_pager_reserve(vm_object_t object, vm_pindex_t start, vm_size_t size)
965 {
966 daddr_t addr, blk, n_free, s_free;
967 int i, j, n;
968
969 swp_pager_init_freerange(&s_free, &n_free);
970 VM_OBJECT_WLOCK(object);
971 for (i = 0; i < size; i += n) {
972 n = size - i;
973 blk = swp_pager_getswapspace(&n);
974 if (blk == SWAPBLK_NONE) {
975 swp_pager_meta_free(object, start, i);
976 VM_OBJECT_WUNLOCK(object);
977 return (-1);
978 }
979 for (j = 0; j < n; ++j) {
980 addr = swp_pager_meta_build(object,
981 start + i + j, blk + j);
982 if (addr != SWAPBLK_NONE)
983 swp_pager_update_freerange(&s_free, &n_free,
984 addr);
985 }
986 }
987 swp_pager_freeswapspace(s_free, n_free);
988 VM_OBJECT_WUNLOCK(object);
989 return (0);
990 }
991
992 static bool
993 swp_pager_xfer_source(vm_object_t srcobject, vm_object_t dstobject,
994 vm_pindex_t pindex, daddr_t addr)
995 {
996 daddr_t dstaddr;
997
998 KASSERT(srcobject->type == OBJT_SWAP,
999 ("%s: Srcobject not swappable", __func__));
1000 if (dstobject->type == OBJT_SWAP &&
1001 swp_pager_meta_lookup(dstobject, pindex) != SWAPBLK_NONE) {
1002 /* Caller should destroy the source block. */
1003 return (false);
1004 }
1005
1006 /*
1007 * Destination has no swapblk and is not resident, transfer source.
1008 * swp_pager_meta_build() can sleep.
1009 */
1010 VM_OBJECT_WUNLOCK(srcobject);
1011 dstaddr = swp_pager_meta_build(dstobject, pindex, addr);
1012 KASSERT(dstaddr == SWAPBLK_NONE,
1013 ("Unexpected destination swapblk"));
1014 VM_OBJECT_WLOCK(srcobject);
1015
1016 return (true);
1017 }
1018
1019 /*
1020 * SWAP_PAGER_COPY() - copy blocks from source pager to destination pager
1021 * and destroy the source.
1022 *
1023 * Copy any valid swapblks from the source to the destination. In
1024 * cases where both the source and destination have a valid swapblk,
1025 * we keep the destination's.
1026 *
1027 * This routine is allowed to sleep. It may sleep allocating metadata
1028 * indirectly through swp_pager_meta_build().
1029 *
1030 * The source object contains no vm_page_t's (which is just as well)
1031 *
1032 * The source object is of type OBJT_SWAP.
1033 *
1034 * The source and destination objects must be locked.
1035 * Both object locks may temporarily be released.
1036 */
1037 void
1038 swap_pager_copy(vm_object_t srcobject, vm_object_t dstobject,
1039 vm_pindex_t offset, int destroysource)
1040 {
1041
1042 VM_OBJECT_ASSERT_WLOCKED(srcobject);
1043 VM_OBJECT_ASSERT_WLOCKED(dstobject);
1044
1045 /*
1046 * If destroysource is set, we remove the source object from the
1047 * swap_pager internal queue now.
1048 */
1049 if (destroysource && (srcobject->flags & OBJ_ANON) == 0 &&
1050 srcobject->handle != NULL) {
1051 VM_OBJECT_WUNLOCK(srcobject);
1052 VM_OBJECT_WUNLOCK(dstobject);
1053 sx_xlock(&sw_alloc_sx);
1054 TAILQ_REMOVE(NOBJLIST(srcobject->handle), srcobject,
1055 pager_object_list);
1056 sx_xunlock(&sw_alloc_sx);
1057 VM_OBJECT_WLOCK(dstobject);
1058 VM_OBJECT_WLOCK(srcobject);
1059 }
1060
1061 /*
1062 * Transfer source to destination.
1063 */
1064 swp_pager_meta_transfer(srcobject, dstobject, offset, dstobject->size);
1065
1066 /*
1067 * Free left over swap blocks in source.
1068 *
1069 * We have to revert the type to OBJT_DEFAULT so we do not accidentally
1070 * double-remove the object from the swap queues.
1071 */
1072 if (destroysource) {
1073 swp_pager_meta_free_all(srcobject);
1074 /*
1075 * Reverting the type is not necessary, the caller is going
1076 * to destroy srcobject directly, but I'm doing it here
1077 * for consistency since we've removed the object from its
1078 * queues.
1079 */
1080 srcobject->type = OBJT_DEFAULT;
1081 }
1082 }
1083
1084 /*
1085 * SWAP_PAGER_HASPAGE() - determine if we have good backing store for
1086 * the requested page.
1087 *
1088 * We determine whether good backing store exists for the requested
1089 * page and return TRUE if it does, FALSE if it doesn't.
1090 *
1091 * If TRUE, we also try to determine how much valid, contiguous backing
1092 * store exists before and after the requested page.
1093 */
1094 static boolean_t
1095 swap_pager_haspage(vm_object_t object, vm_pindex_t pindex, int *before,
1096 int *after)
1097 {
1098 daddr_t blk, blk0;
1099 int i;
1100
1101 VM_OBJECT_ASSERT_LOCKED(object);
1102 KASSERT(object->type == OBJT_SWAP,
1103 ("%s: object not swappable", __func__));
1104
1105 /*
1106 * do we have good backing store at the requested index ?
1107 */
1108 blk0 = swp_pager_meta_lookup(object, pindex);
1109 if (blk0 == SWAPBLK_NONE) {
1110 if (before)
1111 *before = 0;
1112 if (after)
1113 *after = 0;
1114 return (FALSE);
1115 }
1116
1117 /*
1118 * find backwards-looking contiguous good backing store
1119 */
1120 if (before != NULL) {
1121 for (i = 1; i < SWB_NPAGES; i++) {
1122 if (i > pindex)
1123 break;
1124 blk = swp_pager_meta_lookup(object, pindex - i);
1125 if (blk != blk0 - i)
1126 break;
1127 }
1128 *before = i - 1;
1129 }
1130
1131 /*
1132 * find forward-looking contiguous good backing store
1133 */
1134 if (after != NULL) {
1135 for (i = 1; i < SWB_NPAGES; i++) {
1136 blk = swp_pager_meta_lookup(object, pindex + i);
1137 if (blk != blk0 + i)
1138 break;
1139 }
1140 *after = i - 1;
1141 }
1142 return (TRUE);
1143 }
1144
1145 /*
1146 * SWAP_PAGER_PAGE_UNSWAPPED() - remove swap backing store related to page
1147 *
1148 * This removes any associated swap backing store, whether valid or
1149 * not, from the page.
1150 *
1151 * This routine is typically called when a page is made dirty, at
1152 * which point any associated swap can be freed. MADV_FREE also
1153 * calls us in a special-case situation
1154 *
1155 * NOTE!!! If the page is clean and the swap was valid, the caller
1156 * should make the page dirty before calling this routine. This routine
1157 * does NOT change the m->dirty status of the page. Also: MADV_FREE
1158 * depends on it.
1159 *
1160 * This routine may not sleep.
1161 *
1162 * The object containing the page may be locked.
1163 */
1164 static void
1165 swap_pager_unswapped(vm_page_t m)
1166 {
1167 struct swblk *sb;
1168 vm_object_t obj;
1169
1170 /*
1171 * Handle enqueing deferred frees first. If we do not have the
1172 * object lock we wait for the page daemon to clear the space.
1173 */
1174 obj = m->object;
1175 if (!VM_OBJECT_WOWNED(obj)) {
1176 VM_PAGE_OBJECT_BUSY_ASSERT(m);
1177 /*
1178 * The caller is responsible for synchronization but we
1179 * will harmlessly handle races. This is typically provided
1180 * by only calling unswapped() when a page transitions from
1181 * clean to dirty.
1182 */
1183 if ((m->a.flags & (PGA_SWAP_SPACE | PGA_SWAP_FREE)) ==
1184 PGA_SWAP_SPACE) {
1185 vm_page_aflag_set(m, PGA_SWAP_FREE);
1186 counter_u64_add(swap_free_deferred, 1);
1187 }
1188 return;
1189 }
1190 if ((m->a.flags & PGA_SWAP_FREE) != 0)
1191 counter_u64_add(swap_free_completed, 1);
1192 vm_page_aflag_clear(m, PGA_SWAP_FREE | PGA_SWAP_SPACE);
1193
1194 /*
1195 * The meta data only exists if the object is OBJT_SWAP
1196 * and even then might not be allocated yet.
1197 */
1198 KASSERT(m->object->type == OBJT_SWAP,
1199 ("Free object not swappable"));
1200
1201 sb = SWAP_PCTRIE_LOOKUP(&m->object->un_pager.swp.swp_blks,
1202 rounddown(m->pindex, SWAP_META_PAGES));
1203 if (sb == NULL)
1204 return;
1205 if (sb->d[m->pindex % SWAP_META_PAGES] == SWAPBLK_NONE)
1206 return;
1207 swp_pager_freeswapspace(sb->d[m->pindex % SWAP_META_PAGES], 1);
1208 sb->d[m->pindex % SWAP_META_PAGES] = SWAPBLK_NONE;
1209 swp_pager_free_empty_swblk(m->object, sb);
1210 }
1211
1212 /*
1213 * swap_pager_getpages() - bring pages in from swap
1214 *
1215 * Attempt to page in the pages in array "ma" of length "count". The
1216 * caller may optionally specify that additional pages preceding and
1217 * succeeding the specified range be paged in. The number of such pages
1218 * is returned in the "rbehind" and "rahead" parameters, and they will
1219 * be in the inactive queue upon return.
1220 *
1221 * The pages in "ma" must be busied and will remain busied upon return.
1222 */
1223 static int
1224 swap_pager_getpages_locked(vm_object_t object, vm_page_t *ma, int count,
1225 int *rbehind, int *rahead)
1226 {
1227 struct buf *bp;
1228 vm_page_t bm, mpred, msucc, p;
1229 vm_pindex_t pindex;
1230 daddr_t blk;
1231 int i, maxahead, maxbehind, reqcount;
1232
1233 VM_OBJECT_ASSERT_WLOCKED(object);
1234 reqcount = count;
1235
1236 KASSERT(object->type == OBJT_SWAP,
1237 ("%s: object not swappable", __func__));
1238 if (!swap_pager_haspage(object, ma[0]->pindex, &maxbehind, &maxahead)) {
1239 VM_OBJECT_WUNLOCK(object);
1240 return (VM_PAGER_FAIL);
1241 }
1242
1243 KASSERT(reqcount - 1 <= maxahead,
1244 ("page count %d extends beyond swap block", reqcount));
1245
1246 /*
1247 * Do not transfer any pages other than those that are xbusied
1248 * when running during a split or collapse operation. This
1249 * prevents clustering from re-creating pages which are being
1250 * moved into another object.
1251 */
1252 if ((object->flags & (OBJ_SPLIT | OBJ_DEAD)) != 0) {
1253 maxahead = reqcount - 1;
1254 maxbehind = 0;
1255 }
1256
1257 /*
1258 * Clip the readahead and readbehind ranges to exclude resident pages.
1259 */
1260 if (rahead != NULL) {
1261 *rahead = imin(*rahead, maxahead - (reqcount - 1));
1262 pindex = ma[reqcount - 1]->pindex;
1263 msucc = TAILQ_NEXT(ma[reqcount - 1], listq);
1264 if (msucc != NULL && msucc->pindex - pindex - 1 < *rahead)
1265 *rahead = msucc->pindex - pindex - 1;
1266 }
1267 if (rbehind != NULL) {
1268 *rbehind = imin(*rbehind, maxbehind);
1269 pindex = ma[0]->pindex;
1270 mpred = TAILQ_PREV(ma[0], pglist, listq);
1271 if (mpred != NULL && pindex - mpred->pindex - 1 < *rbehind)
1272 *rbehind = pindex - mpred->pindex - 1;
1273 }
1274
1275 bm = ma[0];
1276 for (i = 0; i < count; i++)
1277 ma[i]->oflags |= VPO_SWAPINPROG;
1278
1279 /*
1280 * Allocate readahead and readbehind pages.
1281 */
1282 if (rbehind != NULL) {
1283 for (i = 1; i <= *rbehind; i++) {
1284 p = vm_page_alloc(object, ma[0]->pindex - i,
1285 VM_ALLOC_NORMAL);
1286 if (p == NULL)
1287 break;
1288 p->oflags |= VPO_SWAPINPROG;
1289 bm = p;
1290 }
1291 *rbehind = i - 1;
1292 }
1293 if (rahead != NULL) {
1294 for (i = 0; i < *rahead; i++) {
1295 p = vm_page_alloc(object,
1296 ma[reqcount - 1]->pindex + i + 1, VM_ALLOC_NORMAL);
1297 if (p == NULL)
1298 break;
1299 p->oflags |= VPO_SWAPINPROG;
1300 }
1301 *rahead = i;
1302 }
1303 if (rbehind != NULL)
1304 count += *rbehind;
1305 if (rahead != NULL)
1306 count += *rahead;
1307
1308 vm_object_pip_add(object, count);
1309
1310 pindex = bm->pindex;
1311 blk = swp_pager_meta_lookup(object, pindex);
1312 KASSERT(blk != SWAPBLK_NONE,
1313 ("no swap blocking containing %p(%jx)", object, (uintmax_t)pindex));
1314
1315 VM_OBJECT_WUNLOCK(object);
1316 bp = uma_zalloc(swrbuf_zone, M_WAITOK);
1317 MPASS((bp->b_flags & B_MAXPHYS) != 0);
1318 /* Pages cannot leave the object while busy. */
1319 for (i = 0, p = bm; i < count; i++, p = TAILQ_NEXT(p, listq)) {
1320 MPASS(p->pindex == bm->pindex + i);
1321 bp->b_pages[i] = p;
1322 }
1323
1324 bp->b_flags |= B_PAGING;
1325 bp->b_iocmd = BIO_READ;
1326 bp->b_iodone = swp_pager_async_iodone;
1327 bp->b_rcred = crhold(thread0.td_ucred);
1328 bp->b_wcred = crhold(thread0.td_ucred);
1329 bp->b_blkno = blk;
1330 bp->b_bcount = PAGE_SIZE * count;
1331 bp->b_bufsize = PAGE_SIZE * count;
1332 bp->b_npages = count;
1333 bp->b_pgbefore = rbehind != NULL ? *rbehind : 0;
1334 bp->b_pgafter = rahead != NULL ? *rahead : 0;
1335
1336 VM_CNT_INC(v_swapin);
1337 VM_CNT_ADD(v_swappgsin, count);
1338
1339 /*
1340 * perform the I/O. NOTE!!! bp cannot be considered valid after
1341 * this point because we automatically release it on completion.
1342 * Instead, we look at the one page we are interested in which we
1343 * still hold a lock on even through the I/O completion.
1344 *
1345 * The other pages in our ma[] array are also released on completion,
1346 * so we cannot assume they are valid anymore either.
1347 *
1348 * NOTE: b_blkno is destroyed by the call to swapdev_strategy
1349 */
1350 BUF_KERNPROC(bp);
1351 swp_pager_strategy(bp);
1352
1353 /*
1354 * Wait for the pages we want to complete. VPO_SWAPINPROG is always
1355 * cleared on completion. If an I/O error occurs, SWAPBLK_NONE
1356 * is set in the metadata for each page in the request.
1357 */
1358 VM_OBJECT_WLOCK(object);
1359 /* This could be implemented more efficiently with aflags */
1360 while ((ma[0]->oflags & VPO_SWAPINPROG) != 0) {
1361 ma[0]->oflags |= VPO_SWAPSLEEP;
1362 VM_CNT_INC(v_intrans);
1363 if (VM_OBJECT_SLEEP(object, &object->handle, PSWP,
1364 "swread", hz * 20)) {
1365 printf(
1366 "swap_pager: indefinite wait buffer: bufobj: %p, blkno: %jd, size: %ld\n",
1367 bp->b_bufobj, (intmax_t)bp->b_blkno, bp->b_bcount);
1368 }
1369 }
1370 VM_OBJECT_WUNLOCK(object);
1371
1372 /*
1373 * If we had an unrecoverable read error pages will not be valid.
1374 */
1375 for (i = 0; i < reqcount; i++)
1376 if (ma[i]->valid != VM_PAGE_BITS_ALL)
1377 return (VM_PAGER_ERROR);
1378
1379 return (VM_PAGER_OK);
1380
1381 /*
1382 * A final note: in a low swap situation, we cannot deallocate swap
1383 * and mark a page dirty here because the caller is likely to mark
1384 * the page clean when we return, causing the page to possibly revert
1385 * to all-zero's later.
1386 */
1387 }
1388
1389 static int
1390 swap_pager_getpages(vm_object_t object, vm_page_t *ma, int count,
1391 int *rbehind, int *rahead)
1392 {
1393
1394 VM_OBJECT_WLOCK(object);
1395 return (swap_pager_getpages_locked(object, ma, count, rbehind, rahead));
1396 }
1397
1398 /*
1399 * swap_pager_getpages_async():
1400 *
1401 * Right now this is emulation of asynchronous operation on top of
1402 * swap_pager_getpages().
1403 */
1404 static int
1405 swap_pager_getpages_async(vm_object_t object, vm_page_t *ma, int count,
1406 int *rbehind, int *rahead, pgo_getpages_iodone_t iodone, void *arg)
1407 {
1408 int r, error;
1409
1410 r = swap_pager_getpages(object, ma, count, rbehind, rahead);
1411 switch (r) {
1412 case VM_PAGER_OK:
1413 error = 0;
1414 break;
1415 case VM_PAGER_ERROR:
1416 error = EIO;
1417 break;
1418 case VM_PAGER_FAIL:
1419 error = EINVAL;
1420 break;
1421 default:
1422 panic("unhandled swap_pager_getpages() error %d", r);
1423 }
1424 (iodone)(arg, ma, count, error);
1425
1426 return (r);
1427 }
1428
1429 /*
1430 * swap_pager_putpages:
1431 *
1432 * Assign swap (if necessary) and initiate I/O on the specified pages.
1433 *
1434 * We support both OBJT_DEFAULT and OBJT_SWAP objects. DEFAULT objects
1435 * are automatically converted to SWAP objects.
1436 *
1437 * In a low memory situation we may block in VOP_STRATEGY(), but the new
1438 * vm_page reservation system coupled with properly written VFS devices
1439 * should ensure that no low-memory deadlock occurs. This is an area
1440 * which needs work.
1441 *
1442 * The parent has N vm_object_pip_add() references prior to
1443 * calling us and will remove references for rtvals[] that are
1444 * not set to VM_PAGER_PEND. We need to remove the rest on I/O
1445 * completion.
1446 *
1447 * The parent has soft-busy'd the pages it passes us and will unbusy
1448 * those whose rtvals[] entry is not set to VM_PAGER_PEND on return.
1449 * We need to unbusy the rest on I/O completion.
1450 */
1451 static void
1452 swap_pager_putpages(vm_object_t object, vm_page_t *ma, int count,
1453 int flags, int *rtvals)
1454 {
1455 struct buf *bp;
1456 daddr_t addr, blk, n_free, s_free;
1457 vm_page_t mreq;
1458 int i, j, n;
1459 bool async;
1460
1461 KASSERT(count == 0 || ma[0]->object == object,
1462 ("%s: object mismatch %p/%p",
1463 __func__, object, ma[0]->object));
1464
1465 /*
1466 * Step 1
1467 *
1468 * Turn object into OBJT_SWAP. Force sync if not a pageout process.
1469 */
1470 if (object->type != OBJT_SWAP) {
1471 addr = swp_pager_meta_build(object, 0, SWAPBLK_NONE);
1472 KASSERT(addr == SWAPBLK_NONE,
1473 ("unexpected object swap block"));
1474 }
1475 VM_OBJECT_WUNLOCK(object);
1476 async = curproc == pageproc && (flags & VM_PAGER_PUT_SYNC) == 0;
1477 swp_pager_init_freerange(&s_free, &n_free);
1478
1479 /*
1480 * Step 2
1481 *
1482 * Assign swap blocks and issue I/O. We reallocate swap on the fly.
1483 * The page is left dirty until the pageout operation completes
1484 * successfully.
1485 */
1486 for (i = 0; i < count; i += n) {
1487 /* Maximum I/O size is limited by maximum swap block size. */
1488 n = min(count - i, nsw_cluster_max);
1489
1490 if (async) {
1491 mtx_lock(&swbuf_mtx);
1492 while (nsw_wcount_async == 0)
1493 msleep(&nsw_wcount_async, &swbuf_mtx, PVM,
1494 "swbufa", 0);
1495 nsw_wcount_async--;
1496 mtx_unlock(&swbuf_mtx);
1497 }
1498
1499 /* Get a block of swap of size up to size n. */
1500 VM_OBJECT_WLOCK(object);
1501 blk = swp_pager_getswapspace(&n);
1502 if (blk == SWAPBLK_NONE) {
1503 VM_OBJECT_WUNLOCK(object);
1504 mtx_lock(&swbuf_mtx);
1505 if (++nsw_wcount_async == 1)
1506 wakeup(&nsw_wcount_async);
1507 mtx_unlock(&swbuf_mtx);
1508 for (j = 0; j < n; ++j)
1509 rtvals[i + j] = VM_PAGER_FAIL;
1510 continue;
1511 }
1512 for (j = 0; j < n; ++j) {
1513 mreq = ma[i + j];
1514 vm_page_aflag_clear(mreq, PGA_SWAP_FREE);
1515 addr = swp_pager_meta_build(mreq->object, mreq->pindex,
1516 blk + j);
1517 if (addr != SWAPBLK_NONE)
1518 swp_pager_update_freerange(&s_free, &n_free,
1519 addr);
1520 MPASS(mreq->dirty == VM_PAGE_BITS_ALL);
1521 mreq->oflags |= VPO_SWAPINPROG;
1522 }
1523 VM_OBJECT_WUNLOCK(object);
1524
1525 bp = uma_zalloc(swwbuf_zone, M_WAITOK);
1526 MPASS((bp->b_flags & B_MAXPHYS) != 0);
1527 if (async)
1528 bp->b_flags |= B_ASYNC;
1529 bp->b_flags |= B_PAGING;
1530 bp->b_iocmd = BIO_WRITE;
1531
1532 bp->b_rcred = crhold(thread0.td_ucred);
1533 bp->b_wcred = crhold(thread0.td_ucred);
1534 bp->b_bcount = PAGE_SIZE * n;
1535 bp->b_bufsize = PAGE_SIZE * n;
1536 bp->b_blkno = blk;
1537 for (j = 0; j < n; j++)
1538 bp->b_pages[j] = ma[i + j];
1539 bp->b_npages = n;
1540
1541 /*
1542 * Must set dirty range for NFS to work.
1543 */
1544 bp->b_dirtyoff = 0;
1545 bp->b_dirtyend = bp->b_bcount;
1546
1547 VM_CNT_INC(v_swapout);
1548 VM_CNT_ADD(v_swappgsout, bp->b_npages);
1549
1550 /*
1551 * We unconditionally set rtvals[] to VM_PAGER_PEND so that we
1552 * can call the async completion routine at the end of a
1553 * synchronous I/O operation. Otherwise, our caller would
1554 * perform duplicate unbusy and wakeup operations on the page
1555 * and object, respectively.
1556 */
1557 for (j = 0; j < n; j++)
1558 rtvals[i + j] = VM_PAGER_PEND;
1559
1560 /*
1561 * asynchronous
1562 *
1563 * NOTE: b_blkno is destroyed by the call to swapdev_strategy.
1564 */
1565 if (async) {
1566 bp->b_iodone = swp_pager_async_iodone;
1567 BUF_KERNPROC(bp);
1568 swp_pager_strategy(bp);
1569 continue;
1570 }
1571
1572 /*
1573 * synchronous
1574 *
1575 * NOTE: b_blkno is destroyed by the call to swapdev_strategy.
1576 */
1577 bp->b_iodone = bdone;
1578 swp_pager_strategy(bp);
1579
1580 /*
1581 * Wait for the sync I/O to complete.
1582 */
1583 bwait(bp, PVM, "swwrt");
1584
1585 /*
1586 * Now that we are through with the bp, we can call the
1587 * normal async completion, which frees everything up.
1588 */
1589 swp_pager_async_iodone(bp);
1590 }
1591 swp_pager_freeswapspace(s_free, n_free);
1592 VM_OBJECT_WLOCK(object);
1593 }
1594
1595 /*
1596 * swp_pager_async_iodone:
1597 *
1598 * Completion routine for asynchronous reads and writes from/to swap.
1599 * Also called manually by synchronous code to finish up a bp.
1600 *
1601 * This routine may not sleep.
1602 */
1603 static void
1604 swp_pager_async_iodone(struct buf *bp)
1605 {
1606 int i;
1607 vm_object_t object = NULL;
1608
1609 /*
1610 * Report error - unless we ran out of memory, in which case
1611 * we've already logged it in swapgeom_strategy().
1612 */
1613 if (bp->b_ioflags & BIO_ERROR && bp->b_error != ENOMEM) {
1614 printf(
1615 "swap_pager: I/O error - %s failed; blkno %ld,"
1616 "size %ld, error %d\n",
1617 ((bp->b_iocmd == BIO_READ) ? "pagein" : "pageout"),
1618 (long)bp->b_blkno,
1619 (long)bp->b_bcount,
1620 bp->b_error
1621 );
1622 }
1623
1624 /*
1625 * remove the mapping for kernel virtual
1626 */
1627 if (buf_mapped(bp))
1628 pmap_qremove((vm_offset_t)bp->b_data, bp->b_npages);
1629 else
1630 bp->b_data = bp->b_kvabase;
1631
1632 if (bp->b_npages) {
1633 object = bp->b_pages[0]->object;
1634 VM_OBJECT_WLOCK(object);
1635 }
1636
1637 /*
1638 * cleanup pages. If an error occurs writing to swap, we are in
1639 * very serious trouble. If it happens to be a disk error, though,
1640 * we may be able to recover by reassigning the swap later on. So
1641 * in this case we remove the m->swapblk assignment for the page
1642 * but do not free it in the rlist. The errornous block(s) are thus
1643 * never reallocated as swap. Redirty the page and continue.
1644 */
1645 for (i = 0; i < bp->b_npages; ++i) {
1646 vm_page_t m = bp->b_pages[i];
1647
1648 m->oflags &= ~VPO_SWAPINPROG;
1649 if (m->oflags & VPO_SWAPSLEEP) {
1650 m->oflags &= ~VPO_SWAPSLEEP;
1651 wakeup(&object->handle);
1652 }
1653
1654 /* We always have space after I/O, successful or not. */
1655 vm_page_aflag_set(m, PGA_SWAP_SPACE);
1656
1657 if (bp->b_ioflags & BIO_ERROR) {
1658 /*
1659 * If an error occurs I'd love to throw the swapblk
1660 * away without freeing it back to swapspace, so it
1661 * can never be used again. But I can't from an
1662 * interrupt.
1663 */
1664 if (bp->b_iocmd == BIO_READ) {
1665 /*
1666 * NOTE: for reads, m->dirty will probably
1667 * be overridden by the original caller of
1668 * getpages so don't play cute tricks here.
1669 */
1670 vm_page_invalid(m);
1671 } else {
1672 /*
1673 * If a write error occurs, reactivate page
1674 * so it doesn't clog the inactive list,
1675 * then finish the I/O.
1676 */
1677 MPASS(m->dirty == VM_PAGE_BITS_ALL);
1678
1679 /* PQ_UNSWAPPABLE? */
1680 vm_page_activate(m);
1681 vm_page_sunbusy(m);
1682 }
1683 } else if (bp->b_iocmd == BIO_READ) {
1684 /*
1685 * NOTE: for reads, m->dirty will probably be
1686 * overridden by the original caller of getpages so
1687 * we cannot set them in order to free the underlying
1688 * swap in a low-swap situation. I don't think we'd
1689 * want to do that anyway, but it was an optimization
1690 * that existed in the old swapper for a time before
1691 * it got ripped out due to precisely this problem.
1692 */
1693 KASSERT(!pmap_page_is_mapped(m),
1694 ("swp_pager_async_iodone: page %p is mapped", m));
1695 KASSERT(m->dirty == 0,
1696 ("swp_pager_async_iodone: page %p is dirty", m));
1697
1698 vm_page_valid(m);
1699 if (i < bp->b_pgbefore ||
1700 i >= bp->b_npages - bp->b_pgafter)
1701 vm_page_readahead_finish(m);
1702 } else {
1703 /*
1704 * For write success, clear the dirty
1705 * status, then finish the I/O ( which decrements the
1706 * busy count and possibly wakes waiter's up ).
1707 * A page is only written to swap after a period of
1708 * inactivity. Therefore, we do not expect it to be
1709 * reused.
1710 */
1711 KASSERT(!pmap_page_is_write_mapped(m),
1712 ("swp_pager_async_iodone: page %p is not write"
1713 " protected", m));
1714 vm_page_undirty(m);
1715 vm_page_deactivate_noreuse(m);
1716 vm_page_sunbusy(m);
1717 }
1718 }
1719
1720 /*
1721 * adjust pip. NOTE: the original parent may still have its own
1722 * pip refs on the object.
1723 */
1724 if (object != NULL) {
1725 vm_object_pip_wakeupn(object, bp->b_npages);
1726 VM_OBJECT_WUNLOCK(object);
1727 }
1728
1729 /*
1730 * swapdev_strategy() manually sets b_vp and b_bufobj before calling
1731 * bstrategy(). Set them back to NULL now we're done with it, or we'll
1732 * trigger a KASSERT in relpbuf().
1733 */
1734 if (bp->b_vp) {
1735 bp->b_vp = NULL;
1736 bp->b_bufobj = NULL;
1737 }
1738 /*
1739 * release the physical I/O buffer
1740 */
1741 if (bp->b_flags & B_ASYNC) {
1742 mtx_lock(&swbuf_mtx);
1743 if (++nsw_wcount_async == 1)
1744 wakeup(&nsw_wcount_async);
1745 mtx_unlock(&swbuf_mtx);
1746 }
1747 uma_zfree((bp->b_iocmd == BIO_READ) ? swrbuf_zone : swwbuf_zone, bp);
1748 }
1749
1750 int
1751 swap_pager_nswapdev(void)
1752 {
1753
1754 return (nswapdev);
1755 }
1756
1757 static void
1758 swp_pager_force_dirty(vm_page_t m)
1759 {
1760
1761 vm_page_dirty(m);
1762 swap_pager_unswapped(m);
1763 vm_page_launder(m);
1764 }
1765
1766 /*
1767 * swap_pager_swapoff_object:
1768 *
1769 * Page in all of the pages that have been paged out for an object
1770 * to a swap device.
1771 */
1772 static void
1773 swap_pager_swapoff_object(struct swdevt *sp, vm_object_t object)
1774 {
1775 struct swblk *sb;
1776 vm_page_t m;
1777 vm_pindex_t pi;
1778 daddr_t blk;
1779 int i, nv, rahead, rv;
1780
1781 KASSERT(object->type == OBJT_SWAP,
1782 ("%s: Object not swappable", __func__));
1783
1784 for (pi = 0; (sb = SWAP_PCTRIE_LOOKUP_GE(
1785 &object->un_pager.swp.swp_blks, pi)) != NULL; ) {
1786 if ((object->flags & OBJ_DEAD) != 0) {
1787 /*
1788 * Make sure that pending writes finish before
1789 * returning.
1790 */
1791 vm_object_pip_wait(object, "swpoff");
1792 swp_pager_meta_free_all(object);
1793 break;
1794 }
1795 for (i = 0; i < SWAP_META_PAGES; i++) {
1796 /*
1797 * Count the number of contiguous valid blocks.
1798 */
1799 for (nv = 0; nv < SWAP_META_PAGES - i; nv++) {
1800 blk = sb->d[i + nv];
1801 if (!swp_pager_isondev(blk, sp) ||
1802 blk == SWAPBLK_NONE)
1803 break;
1804 }
1805 if (nv == 0)
1806 continue;
1807
1808 /*
1809 * Look for a page corresponding to the first
1810 * valid block and ensure that any pending paging
1811 * operations on it are complete. If the page is valid,
1812 * mark it dirty and free the swap block. Try to batch
1813 * this operation since it may cause sp to be freed,
1814 * meaning that we must restart the scan. Avoid busying
1815 * valid pages since we may block forever on kernel
1816 * stack pages.
1817 */
1818 m = vm_page_lookup(object, sb->p + i);
1819 if (m == NULL) {
1820 m = vm_page_alloc(object, sb->p + i,
1821 VM_ALLOC_NORMAL | VM_ALLOC_WAITFAIL);
1822 if (m == NULL)
1823 break;
1824 } else {
1825 if ((m->oflags & VPO_SWAPINPROG) != 0) {
1826 m->oflags |= VPO_SWAPSLEEP;
1827 VM_OBJECT_SLEEP(object, &object->handle,
1828 PSWP, "swpoff", 0);
1829 break;
1830 }
1831 if (vm_page_all_valid(m)) {
1832 do {
1833 swp_pager_force_dirty(m);
1834 } while (--nv > 0 &&
1835 (m = vm_page_next(m)) != NULL &&
1836 vm_page_all_valid(m) &&
1837 (m->oflags & VPO_SWAPINPROG) == 0);
1838 break;
1839 }
1840 if (!vm_page_busy_acquire(m, VM_ALLOC_WAITFAIL))
1841 break;
1842 }
1843
1844 vm_object_pip_add(object, 1);
1845 rahead = SWAP_META_PAGES;
1846 rv = swap_pager_getpages_locked(object, &m, 1, NULL,
1847 &rahead);
1848 if (rv != VM_PAGER_OK)
1849 panic("%s: read from swap failed: %d",
1850 __func__, rv);
1851 vm_object_pip_wakeupn(object, 1);
1852 VM_OBJECT_WLOCK(object);
1853 vm_page_xunbusy(m);
1854
1855 /*
1856 * The object lock was dropped so we must restart the
1857 * scan of this swap block. Pages paged in during this
1858 * iteration will be marked dirty in a future iteration.
1859 */
1860 break;
1861 }
1862 if (i == SWAP_META_PAGES)
1863 pi = sb->p + SWAP_META_PAGES;
1864 }
1865 }
1866
1867 /*
1868 * swap_pager_swapoff:
1869 *
1870 * Page in all of the pages that have been paged out to the
1871 * given device. The corresponding blocks in the bitmap must be
1872 * marked as allocated and the device must be flagged SW_CLOSING.
1873 * There may be no processes swapped out to the device.
1874 *
1875 * This routine may block.
1876 */
1877 static void
1878 swap_pager_swapoff(struct swdevt *sp)
1879 {
1880 vm_object_t object;
1881 int retries;
1882
1883 sx_assert(&swdev_syscall_lock, SA_XLOCKED);
1884
1885 retries = 0;
1886 full_rescan:
1887 mtx_lock(&vm_object_list_mtx);
1888 TAILQ_FOREACH(object, &vm_object_list, object_list) {
1889 if (object->type != OBJT_SWAP)
1890 continue;
1891 mtx_unlock(&vm_object_list_mtx);
1892 /* Depends on type-stability. */
1893 VM_OBJECT_WLOCK(object);
1894
1895 /*
1896 * Dead objects are eventually terminated on their own.
1897 */
1898 if ((object->flags & OBJ_DEAD) != 0)
1899 goto next_obj;
1900
1901 /*
1902 * Sync with fences placed after pctrie
1903 * initialization. We must not access pctrie below
1904 * unless we checked that our object is swap and not
1905 * dead.
1906 */
1907 atomic_thread_fence_acq();
1908 if (object->type != OBJT_SWAP)
1909 goto next_obj;
1910
1911 swap_pager_swapoff_object(sp, object);
1912 next_obj:
1913 VM_OBJECT_WUNLOCK(object);
1914 mtx_lock(&vm_object_list_mtx);
1915 }
1916 mtx_unlock(&vm_object_list_mtx);
1917
1918 if (sp->sw_used) {
1919 /*
1920 * Objects may be locked or paging to the device being
1921 * removed, so we will miss their pages and need to
1922 * make another pass. We have marked this device as
1923 * SW_CLOSING, so the activity should finish soon.
1924 */
1925 retries++;
1926 if (retries > 100) {
1927 panic("swapoff: failed to locate %d swap blocks",
1928 sp->sw_used);
1929 }
1930 pause("swpoff", hz / 20);
1931 goto full_rescan;
1932 }
1933 EVENTHANDLER_INVOKE(swapoff, sp);
1934 }
1935
1936 /************************************************************************
1937 * SWAP META DATA *
1938 ************************************************************************
1939 *
1940 * These routines manipulate the swap metadata stored in the
1941 * OBJT_SWAP object.
1942 *
1943 * Swap metadata is implemented with a global hash and not directly
1944 * linked into the object. Instead the object simply contains
1945 * appropriate tracking counters.
1946 */
1947
1948 /*
1949 * SWP_PAGER_SWBLK_EMPTY() - is a range of blocks free?
1950 */
1951 static bool
1952 swp_pager_swblk_empty(struct swblk *sb, int start, int limit)
1953 {
1954 int i;
1955
1956 MPASS(0 <= start && start <= limit && limit <= SWAP_META_PAGES);
1957 for (i = start; i < limit; i++) {
1958 if (sb->d[i] != SWAPBLK_NONE)
1959 return (false);
1960 }
1961 return (true);
1962 }
1963
1964 /*
1965 * SWP_PAGER_FREE_EMPTY_SWBLK() - frees if a block is free
1966 *
1967 * Nothing is done if the block is still in use.
1968 */
1969 static void
1970 swp_pager_free_empty_swblk(vm_object_t object, struct swblk *sb)
1971 {
1972
1973 if (swp_pager_swblk_empty(sb, 0, SWAP_META_PAGES)) {
1974 SWAP_PCTRIE_REMOVE(&object->un_pager.swp.swp_blks, sb->p);
1975 uma_zfree(swblk_zone, sb);
1976 }
1977 }
1978
1979 /*
1980 * SWP_PAGER_META_BUILD() - add swap block to swap meta data for object
1981 *
1982 * We first convert the object to a swap object if it is a default
1983 * object.
1984 *
1985 * The specified swapblk is added to the object's swap metadata. If
1986 * the swapblk is not valid, it is freed instead. Any previously
1987 * assigned swapblk is returned.
1988 */
1989 static daddr_t
1990 swp_pager_meta_build(vm_object_t object, vm_pindex_t pindex, daddr_t swapblk)
1991 {
1992 static volatile int swblk_zone_exhausted, swpctrie_zone_exhausted;
1993 struct swblk *sb, *sb1;
1994 vm_pindex_t modpi, rdpi;
1995 daddr_t prev_swapblk;
1996 int error, i;
1997
1998 VM_OBJECT_ASSERT_WLOCKED(object);
1999
2000 /*
2001 * Convert default object to swap object if necessary
2002 */
2003 if (object->type != OBJT_SWAP) {
2004 pctrie_init(&object->un_pager.swp.swp_blks);
2005
2006 /*
2007 * Ensure that swap_pager_swapoff()'s iteration over
2008 * object_list does not see a garbage pctrie.
2009 */
2010 atomic_thread_fence_rel();
2011
2012 object->type = OBJT_SWAP;
2013 object->un_pager.swp.writemappings = 0;
2014 KASSERT((object->flags & OBJ_ANON) != 0 ||
2015 object->handle == NULL,
2016 ("default pager %p with handle %p",
2017 object, object->handle));
2018 }
2019
2020 rdpi = rounddown(pindex, SWAP_META_PAGES);
2021 sb = SWAP_PCTRIE_LOOKUP(&object->un_pager.swp.swp_blks, rdpi);
2022 if (sb == NULL) {
2023 if (swapblk == SWAPBLK_NONE)
2024 return (SWAPBLK_NONE);
2025 for (;;) {
2026 sb = uma_zalloc(swblk_zone, M_NOWAIT | (curproc ==
2027 pageproc ? M_USE_RESERVE : 0));
2028 if (sb != NULL) {
2029 sb->p = rdpi;
2030 for (i = 0; i < SWAP_META_PAGES; i++)
2031 sb->d[i] = SWAPBLK_NONE;
2032 if (atomic_cmpset_int(&swblk_zone_exhausted,
2033 1, 0))
2034 printf("swblk zone ok\n");
2035 break;
2036 }
2037 VM_OBJECT_WUNLOCK(object);
2038 if (uma_zone_exhausted(swblk_zone)) {
2039 if (atomic_cmpset_int(&swblk_zone_exhausted,
2040 0, 1))
2041 printf("swap blk zone exhausted, "
2042 "increase kern.maxswzone\n");
2043 vm_pageout_oom(VM_OOM_SWAPZ);
2044 pause("swzonxb", 10);
2045 } else
2046 uma_zwait(swblk_zone);
2047 VM_OBJECT_WLOCK(object);
2048 sb = SWAP_PCTRIE_LOOKUP(&object->un_pager.swp.swp_blks,
2049 rdpi);
2050 if (sb != NULL)
2051 /*
2052 * Somebody swapped out a nearby page,
2053 * allocating swblk at the rdpi index,
2054 * while we dropped the object lock.
2055 */
2056 goto allocated;
2057 }
2058 for (;;) {
2059 error = SWAP_PCTRIE_INSERT(
2060 &object->un_pager.swp.swp_blks, sb);
2061 if (error == 0) {
2062 if (atomic_cmpset_int(&swpctrie_zone_exhausted,
2063 1, 0))
2064 printf("swpctrie zone ok\n");
2065 break;
2066 }
2067 VM_OBJECT_WUNLOCK(object);
2068 if (uma_zone_exhausted(swpctrie_zone)) {
2069 if (atomic_cmpset_int(&swpctrie_zone_exhausted,
2070 0, 1))
2071 printf("swap pctrie zone exhausted, "
2072 "increase kern.maxswzone\n");
2073 vm_pageout_oom(VM_OOM_SWAPZ);
2074 pause("swzonxp", 10);
2075 } else
2076 uma_zwait(swpctrie_zone);
2077 VM_OBJECT_WLOCK(object);
2078 sb1 = SWAP_PCTRIE_LOOKUP(&object->un_pager.swp.swp_blks,
2079 rdpi);
2080 if (sb1 != NULL) {
2081 uma_zfree(swblk_zone, sb);
2082 sb = sb1;
2083 goto allocated;
2084 }
2085 }
2086 }
2087 allocated:
2088 MPASS(sb->p == rdpi);
2089
2090 modpi = pindex % SWAP_META_PAGES;
2091 /* Return prior contents of metadata. */
2092 prev_swapblk = sb->d[modpi];
2093 /* Enter block into metadata. */
2094 sb->d[modpi] = swapblk;
2095
2096 /*
2097 * Free the swblk if we end up with the empty page run.
2098 */
2099 if (swapblk == SWAPBLK_NONE)
2100 swp_pager_free_empty_swblk(object, sb);
2101 return (prev_swapblk);
2102 }
2103
2104 /*
2105 * SWP_PAGER_META_TRANSFER() - free a range of blocks in the srcobject's swap
2106 * metadata, or transfer it into dstobject.
2107 *
2108 * This routine will free swap metadata structures as they are cleaned
2109 * out.
2110 */
2111 static void
2112 swp_pager_meta_transfer(vm_object_t srcobject, vm_object_t dstobject,
2113 vm_pindex_t pindex, vm_pindex_t count)
2114 {
2115 struct swblk *sb;
2116 daddr_t n_free, s_free;
2117 vm_pindex_t offset, last;
2118 int i, limit, start;
2119
2120 VM_OBJECT_ASSERT_WLOCKED(srcobject);
2121 if (srcobject->type != OBJT_SWAP || count == 0)
2122 return;
2123
2124 swp_pager_init_freerange(&s_free, &n_free);
2125 offset = pindex;
2126 last = pindex + count;
2127 for (;;) {
2128 sb = SWAP_PCTRIE_LOOKUP_GE(&srcobject->un_pager.swp.swp_blks,
2129 rounddown(pindex, SWAP_META_PAGES));
2130 if (sb == NULL || sb->p >= last)
2131 break;
2132 start = pindex > sb->p ? pindex - sb->p : 0;
2133 limit = last - sb->p < SWAP_META_PAGES ? last - sb->p :
2134 SWAP_META_PAGES;
2135 for (i = start; i < limit; i++) {
2136 if (sb->d[i] == SWAPBLK_NONE)
2137 continue;
2138 if (dstobject == NULL ||
2139 !swp_pager_xfer_source(srcobject, dstobject,
2140 sb->p + i - offset, sb->d[i])) {
2141 swp_pager_update_freerange(&s_free, &n_free,
2142 sb->d[i]);
2143 }
2144 sb->d[i] = SWAPBLK_NONE;
2145 }
2146 pindex = sb->p + SWAP_META_PAGES;
2147 if (swp_pager_swblk_empty(sb, 0, start) &&
2148 swp_pager_swblk_empty(sb, limit, SWAP_META_PAGES)) {
2149 SWAP_PCTRIE_REMOVE(&srcobject->un_pager.swp.swp_blks,
2150 sb->p);
2151 uma_zfree(swblk_zone, sb);
2152 }
2153 }
2154 swp_pager_freeswapspace(s_free, n_free);
2155 }
2156
2157 /*
2158 * SWP_PAGER_META_FREE() - free a range of blocks in the object's swap metadata
2159 *
2160 * The requested range of blocks is freed, with any associated swap
2161 * returned to the swap bitmap.
2162 *
2163 * This routine will free swap metadata structures as they are cleaned
2164 * out. This routine does *NOT* operate on swap metadata associated
2165 * with resident pages.
2166 */
2167 static void
2168 swp_pager_meta_free(vm_object_t object, vm_pindex_t pindex, vm_pindex_t count)
2169 {
2170 swp_pager_meta_transfer(object, NULL, pindex, count);
2171 }
2172
2173 /*
2174 * SWP_PAGER_META_FREE_ALL() - destroy all swap metadata associated with object
2175 *
2176 * This routine locates and destroys all swap metadata associated with
2177 * an object.
2178 */
2179 static void
2180 swp_pager_meta_free_all(vm_object_t object)
2181 {
2182 struct swblk *sb;
2183 daddr_t n_free, s_free;
2184 vm_pindex_t pindex;
2185 int i;
2186
2187 VM_OBJECT_ASSERT_WLOCKED(object);
2188 if (object->type != OBJT_SWAP)
2189 return;
2190
2191 swp_pager_init_freerange(&s_free, &n_free);
2192 for (pindex = 0; (sb = SWAP_PCTRIE_LOOKUP_GE(
2193 &object->un_pager.swp.swp_blks, pindex)) != NULL;) {
2194 pindex = sb->p + SWAP_META_PAGES;
2195 for (i = 0; i < SWAP_META_PAGES; i++) {
2196 if (sb->d[i] == SWAPBLK_NONE)
2197 continue;
2198 swp_pager_update_freerange(&s_free, &n_free, sb->d[i]);
2199 }
2200 SWAP_PCTRIE_REMOVE(&object->un_pager.swp.swp_blks, sb->p);
2201 uma_zfree(swblk_zone, sb);
2202 }
2203 swp_pager_freeswapspace(s_free, n_free);
2204 }
2205
2206 /*
2207 * SWP_PAGER_METACTL() - misc control of swap meta data.
2208 *
2209 * This routine is capable of looking up, or removing swapblk
2210 * assignments in the swap meta data. It returns the swapblk being
2211 * looked-up, popped, or SWAPBLK_NONE if the block was invalid.
2212 *
2213 * When acting on a busy resident page and paging is in progress, we
2214 * have to wait until paging is complete but otherwise can act on the
2215 * busy page.
2216 */
2217 static daddr_t
2218 swp_pager_meta_lookup(vm_object_t object, vm_pindex_t pindex)
2219 {
2220 struct swblk *sb;
2221
2222 VM_OBJECT_ASSERT_LOCKED(object);
2223
2224 /*
2225 * The meta data only exists if the object is OBJT_SWAP
2226 * and even then might not be allocated yet.
2227 */
2228 KASSERT(object->type == OBJT_SWAP,
2229 ("Lookup object not swappable"));
2230
2231 sb = SWAP_PCTRIE_LOOKUP(&object->un_pager.swp.swp_blks,
2232 rounddown(pindex, SWAP_META_PAGES));
2233 if (sb == NULL)
2234 return (SWAPBLK_NONE);
2235 return (sb->d[pindex % SWAP_META_PAGES]);
2236 }
2237
2238 /*
2239 * Returns the least page index which is greater than or equal to the
2240 * parameter pindex and for which there is a swap block allocated.
2241 * Returns object's size if the object's type is not swap or if there
2242 * are no allocated swap blocks for the object after the requested
2243 * pindex.
2244 */
2245 vm_pindex_t
2246 swap_pager_find_least(vm_object_t object, vm_pindex_t pindex)
2247 {
2248 struct swblk *sb;
2249 int i;
2250
2251 VM_OBJECT_ASSERT_LOCKED(object);
2252 if (object->type != OBJT_SWAP)
2253 return (object->size);
2254
2255 sb = SWAP_PCTRIE_LOOKUP_GE(&object->un_pager.swp.swp_blks,
2256 rounddown(pindex, SWAP_META_PAGES));
2257 if (sb == NULL)
2258 return (object->size);
2259 if (sb->p < pindex) {
2260 for (i = pindex % SWAP_META_PAGES; i < SWAP_META_PAGES; i++) {
2261 if (sb->d[i] != SWAPBLK_NONE)
2262 return (sb->p + i);
2263 }
2264 sb = SWAP_PCTRIE_LOOKUP_GE(&object->un_pager.swp.swp_blks,
2265 roundup(pindex, SWAP_META_PAGES));
2266 if (sb == NULL)
2267 return (object->size);
2268 }
2269 for (i = 0; i < SWAP_META_PAGES; i++) {
2270 if (sb->d[i] != SWAPBLK_NONE)
2271 return (sb->p + i);
2272 }
2273
2274 /*
2275 * We get here if a swblk is present in the trie but it
2276 * doesn't map any blocks.
2277 */
2278 MPASS(0);
2279 return (object->size);
2280 }
2281
2282 /*
2283 * System call swapon(name) enables swapping on device name,
2284 * which must be in the swdevsw. Return EBUSY
2285 * if already swapping on this device.
2286 */
2287 #ifndef _SYS_SYSPROTO_H_
2288 struct swapon_args {
2289 char *name;
2290 };
2291 #endif
2292
2293 /*
2294 * MPSAFE
2295 */
2296 /* ARGSUSED */
2297 int
2298 sys_swapon(struct thread *td, struct swapon_args *uap)
2299 {
2300 struct vattr attr;
2301 struct vnode *vp;
2302 struct nameidata nd;
2303 int error;
2304
2305 error = priv_check(td, PRIV_SWAPON);
2306 if (error)
2307 return (error);
2308
2309 sx_xlock(&swdev_syscall_lock);
2310
2311 /*
2312 * Swap metadata may not fit in the KVM if we have physical
2313 * memory of >1GB.
2314 */
2315 if (swblk_zone == NULL) {
2316 error = ENOMEM;
2317 goto done;
2318 }
2319
2320 NDINIT(&nd, LOOKUP, ISOPEN | FOLLOW | AUDITVNODE1, UIO_USERSPACE,
2321 uap->name, td);
2322 error = namei(&nd);
2323 if (error)
2324 goto done;
2325
2326 NDFREE(&nd, NDF_ONLY_PNBUF);
2327 vp = nd.ni_vp;
2328
2329 if (vn_isdisk_error(vp, &error)) {
2330 error = swapongeom(vp);
2331 } else if (vp->v_type == VREG &&
2332 (vp->v_mount->mnt_vfc->vfc_flags & VFCF_NETWORK) != 0 &&
2333 (error = VOP_GETATTR(vp, &attr, td->td_ucred)) == 0) {
2334 /*
2335 * Allow direct swapping to NFS regular files in the same
2336 * way that nfs_mountroot() sets up diskless swapping.
2337 */
2338 error = swaponvp(td, vp, attr.va_size / DEV_BSIZE);
2339 }
2340
2341 if (error)
2342 vrele(vp);
2343 done:
2344 sx_xunlock(&swdev_syscall_lock);
2345 return (error);
2346 }
2347
2348 /*
2349 * Check that the total amount of swap currently configured does not
2350 * exceed half the theoretical maximum. If it does, print a warning
2351 * message.
2352 */
2353 static void
2354 swapon_check_swzone(void)
2355 {
2356
2357 /* recommend using no more than half that amount */
2358 if (swap_total > swap_maxpages / 2) {
2359 printf("warning: total configured swap (%lu pages) "
2360 "exceeds maximum recommended amount (%lu pages).\n",
2361 swap_total, swap_maxpages / 2);
2362 printf("warning: increase kern.maxswzone "
2363 "or reduce amount of swap.\n");
2364 }
2365 }
2366
2367 static void
2368 swaponsomething(struct vnode *vp, void *id, u_long nblks,
2369 sw_strategy_t *strategy, sw_close_t *close, dev_t dev, int flags)
2370 {
2371 struct swdevt *sp, *tsp;
2372 daddr_t dvbase;
2373
2374 /*
2375 * nblks is in DEV_BSIZE'd chunks, convert to PAGE_SIZE'd chunks.
2376 * First chop nblks off to page-align it, then convert.
2377 *
2378 * sw->sw_nblks is in page-sized chunks now too.
2379 */
2380 nblks &= ~(ctodb(1) - 1);
2381 nblks = dbtoc(nblks);
2382
2383 sp = malloc(sizeof *sp, M_VMPGDATA, M_WAITOK | M_ZERO);
2384 sp->sw_blist = blist_create(nblks, M_WAITOK);
2385 sp->sw_vp = vp;
2386 sp->sw_id = id;
2387 sp->sw_dev = dev;
2388 sp->sw_nblks = nblks;
2389 sp->sw_used = 0;
2390 sp->sw_strategy = strategy;
2391 sp->sw_close = close;
2392 sp->sw_flags = flags;
2393
2394 /*
2395 * Do not free the first blocks in order to avoid overwriting
2396 * any bsd label at the front of the partition
2397 */
2398 blist_free(sp->sw_blist, howmany(BBSIZE, PAGE_SIZE),
2399 nblks - howmany(BBSIZE, PAGE_SIZE));
2400
2401 dvbase = 0;
2402 mtx_lock(&sw_dev_mtx);
2403 TAILQ_FOREACH(tsp, &swtailq, sw_list) {
2404 if (tsp->sw_end >= dvbase) {
2405 /*
2406 * We put one uncovered page between the devices
2407 * in order to definitively prevent any cross-device
2408 * I/O requests
2409 */
2410 dvbase = tsp->sw_end + 1;
2411 }
2412 }
2413 sp->sw_first = dvbase;
2414 sp->sw_end = dvbase + nblks;
2415 TAILQ_INSERT_TAIL(&swtailq, sp, sw_list);
2416 nswapdev++;
2417 swap_pager_avail += nblks - howmany(BBSIZE, PAGE_SIZE);
2418 swap_total += nblks;
2419 swapon_check_swzone();
2420 swp_sizecheck();
2421 mtx_unlock(&sw_dev_mtx);
2422 EVENTHANDLER_INVOKE(swapon, sp);
2423 }
2424
2425 /*
2426 * SYSCALL: swapoff(devname)
2427 *
2428 * Disable swapping on the given device.
2429 *
2430 * XXX: Badly designed system call: it should use a device index
2431 * rather than filename as specification. We keep sw_vp around
2432 * only to make this work.
2433 */
2434 #ifndef _SYS_SYSPROTO_H_
2435 struct swapoff_args {
2436 char *name;
2437 };
2438 #endif
2439
2440 /*
2441 * MPSAFE
2442 */
2443 /* ARGSUSED */
2444 int
2445 sys_swapoff(struct thread *td, struct swapoff_args *uap)
2446 {
2447 struct vnode *vp;
2448 struct nameidata nd;
2449 struct swdevt *sp;
2450 int error;
2451
2452 error = priv_check(td, PRIV_SWAPOFF);
2453 if (error)
2454 return (error);
2455
2456 sx_xlock(&swdev_syscall_lock);
2457
2458 NDINIT(&nd, LOOKUP, FOLLOW | AUDITVNODE1, UIO_USERSPACE, uap->name,
2459 td);
2460 error = namei(&nd);
2461 if (error)
2462 goto done;
2463 NDFREE(&nd, NDF_ONLY_PNBUF);
2464 vp = nd.ni_vp;
2465
2466 mtx_lock(&sw_dev_mtx);
2467 TAILQ_FOREACH(sp, &swtailq, sw_list) {
2468 if (sp->sw_vp == vp)
2469 break;
2470 }
2471 mtx_unlock(&sw_dev_mtx);
2472 if (sp == NULL) {
2473 error = EINVAL;
2474 goto done;
2475 }
2476 error = swapoff_one(sp, td->td_ucred);
2477 done:
2478 sx_xunlock(&swdev_syscall_lock);
2479 return (error);
2480 }
2481
2482 static int
2483 swapoff_one(struct swdevt *sp, struct ucred *cred)
2484 {
2485 u_long nblks;
2486 #ifdef MAC
2487 int error;
2488 #endif
2489
2490 sx_assert(&swdev_syscall_lock, SA_XLOCKED);
2491 #ifdef MAC
2492 (void) vn_lock(sp->sw_vp, LK_EXCLUSIVE | LK_RETRY);
2493 error = mac_system_check_swapoff(cred, sp->sw_vp);
2494 (void) VOP_UNLOCK(sp->sw_vp);
2495 if (error != 0)
2496 return (error);
2497 #endif
2498 nblks = sp->sw_nblks;
2499
2500 /*
2501 * We can turn off this swap device safely only if the
2502 * available virtual memory in the system will fit the amount
2503 * of data we will have to page back in, plus an epsilon so
2504 * the system doesn't become critically low on swap space.
2505 */
2506 if (vm_free_count() + swap_pager_avail < nblks + nswap_lowat)
2507 return (ENOMEM);
2508
2509 /*
2510 * Prevent further allocations on this device.
2511 */
2512 mtx_lock(&sw_dev_mtx);
2513 sp->sw_flags |= SW_CLOSING;
2514 swap_pager_avail -= blist_fill(sp->sw_blist, 0, nblks);
2515 swap_total -= nblks;
2516 mtx_unlock(&sw_dev_mtx);
2517
2518 /*
2519 * Page in the contents of the device and close it.
2520 */
2521 swap_pager_swapoff(sp);
2522
2523 sp->sw_close(curthread, sp);
2524 mtx_lock(&sw_dev_mtx);
2525 sp->sw_id = NULL;
2526 TAILQ_REMOVE(&swtailq, sp, sw_list);
2527 nswapdev--;
2528 if (nswapdev == 0) {
2529 swap_pager_full = 2;
2530 swap_pager_almost_full = 1;
2531 }
2532 if (swdevhd == sp)
2533 swdevhd = NULL;
2534 mtx_unlock(&sw_dev_mtx);
2535 blist_destroy(sp->sw_blist);
2536 free(sp, M_VMPGDATA);
2537 return (0);
2538 }
2539
2540 void
2541 swapoff_all(void)
2542 {
2543 struct swdevt *sp, *spt;
2544 const char *devname;
2545 int error;
2546
2547 sx_xlock(&swdev_syscall_lock);
2548
2549 mtx_lock(&sw_dev_mtx);
2550 TAILQ_FOREACH_SAFE(sp, &swtailq, sw_list, spt) {
2551 mtx_unlock(&sw_dev_mtx);
2552 if (vn_isdisk(sp->sw_vp))
2553 devname = devtoname(sp->sw_vp->v_rdev);
2554 else
2555 devname = "[file]";
2556 error = swapoff_one(sp, thread0.td_ucred);
2557 if (error != 0) {
2558 printf("Cannot remove swap device %s (error=%d), "
2559 "skipping.\n", devname, error);
2560 } else if (bootverbose) {
2561 printf("Swap device %s removed.\n", devname);
2562 }
2563 mtx_lock(&sw_dev_mtx);
2564 }
2565 mtx_unlock(&sw_dev_mtx);
2566
2567 sx_xunlock(&swdev_syscall_lock);
2568 }
2569
2570 void
2571 swap_pager_status(int *total, int *used)
2572 {
2573
2574 *total = swap_total;
2575 *used = swap_total - swap_pager_avail -
2576 nswapdev * howmany(BBSIZE, PAGE_SIZE);
2577 }
2578
2579 int
2580 swap_dev_info(int name, struct xswdev *xs, char *devname, size_t len)
2581 {
2582 struct swdevt *sp;
2583 const char *tmp_devname;
2584 int error, n;
2585
2586 n = 0;
2587 error = ENOENT;
2588 mtx_lock(&sw_dev_mtx);
2589 TAILQ_FOREACH(sp, &swtailq, sw_list) {
2590 if (n != name) {
2591 n++;
2592 continue;
2593 }
2594 xs->xsw_version = XSWDEV_VERSION;
2595 xs->xsw_dev = sp->sw_dev;
2596 xs->xsw_flags = sp->sw_flags;
2597 xs->xsw_nblks = sp->sw_nblks;
2598 xs->xsw_used = sp->sw_used;
2599 if (devname != NULL) {
2600 if (vn_isdisk(sp->sw_vp))
2601 tmp_devname = devtoname(sp->sw_vp->v_rdev);
2602 else
2603 tmp_devname = "[file]";
2604 strncpy(devname, tmp_devname, len);
2605 }
2606 error = 0;
2607 break;
2608 }
2609 mtx_unlock(&sw_dev_mtx);
2610 return (error);
2611 }
2612
2613 #if defined(COMPAT_FREEBSD11)
2614 #define XSWDEV_VERSION_11 1
2615 struct xswdev11 {
2616 u_int xsw_version;
2617 uint32_t xsw_dev;
2618 int xsw_flags;
2619 int xsw_nblks;
2620 int xsw_used;
2621 };
2622 #endif
2623
2624 #if defined(__amd64__) && defined(COMPAT_FREEBSD32)
2625 struct xswdev32 {
2626 u_int xsw_version;
2627 u_int xsw_dev1, xsw_dev2;
2628 int xsw_flags;
2629 int xsw_nblks;
2630 int xsw_used;
2631 };
2632 #endif
2633
2634 static int
2635 sysctl_vm_swap_info(SYSCTL_HANDLER_ARGS)
2636 {
2637 struct xswdev xs;
2638 #if defined(__amd64__) && defined(COMPAT_FREEBSD32)
2639 struct xswdev32 xs32;
2640 #endif
2641 #if defined(COMPAT_FREEBSD11)
2642 struct xswdev11 xs11;
2643 #endif
2644 int error;
2645
2646 if (arg2 != 1) /* name length */
2647 return (EINVAL);
2648 error = swap_dev_info(*(int *)arg1, &xs, NULL, 0);
2649 if (error != 0)
2650 return (error);
2651 #if defined(__amd64__) && defined(COMPAT_FREEBSD32)
2652 if (req->oldlen == sizeof(xs32)) {
2653 xs32.xsw_version = XSWDEV_VERSION;
2654 xs32.xsw_dev1 = xs.xsw_dev;
2655 xs32.xsw_dev2 = xs.xsw_dev >> 32;
2656 xs32.xsw_flags = xs.xsw_flags;
2657 xs32.xsw_nblks = xs.xsw_nblks;
2658 xs32.xsw_used = xs.xsw_used;
2659 error = SYSCTL_OUT(req, &xs32, sizeof(xs32));
2660 return (error);
2661 }
2662 #endif
2663 #if defined(COMPAT_FREEBSD11)
2664 if (req->oldlen == sizeof(xs11)) {
2665 xs11.xsw_version = XSWDEV_VERSION_11;
2666 xs11.xsw_dev = xs.xsw_dev; /* truncation */
2667 xs11.xsw_flags = xs.xsw_flags;
2668 xs11.xsw_nblks = xs.xsw_nblks;
2669 xs11.xsw_used = xs.xsw_used;
2670 error = SYSCTL_OUT(req, &xs11, sizeof(xs11));
2671 return (error);
2672 }
2673 #endif
2674 error = SYSCTL_OUT(req, &xs, sizeof(xs));
2675 return (error);
2676 }
2677
2678 SYSCTL_INT(_vm, OID_AUTO, nswapdev, CTLFLAG_RD, &nswapdev, 0,
2679 "Number of swap devices");
2680 SYSCTL_NODE(_vm, OID_AUTO, swap_info, CTLFLAG_RD | CTLFLAG_MPSAFE,
2681 sysctl_vm_swap_info,
2682 "Swap statistics by device");
2683
2684 /*
2685 * Count the approximate swap usage in pages for a vmspace. The
2686 * shadowed or not yet copied on write swap blocks are not accounted.
2687 * The map must be locked.
2688 */
2689 long
2690 vmspace_swap_count(struct vmspace *vmspace)
2691 {
2692 vm_map_t map;
2693 vm_map_entry_t cur;
2694 vm_object_t object;
2695 struct swblk *sb;
2696 vm_pindex_t e, pi;
2697 long count;
2698 int i;
2699
2700 map = &vmspace->vm_map;
2701 count = 0;
2702
2703 VM_MAP_ENTRY_FOREACH(cur, map) {
2704 if ((cur->eflags & MAP_ENTRY_IS_SUB_MAP) != 0)
2705 continue;
2706 object = cur->object.vm_object;
2707 if (object == NULL || object->type != OBJT_SWAP)
2708 continue;
2709 VM_OBJECT_RLOCK(object);
2710 if (object->type != OBJT_SWAP)
2711 goto unlock;
2712 pi = OFF_TO_IDX(cur->offset);
2713 e = pi + OFF_TO_IDX(cur->end - cur->start);
2714 for (;; pi = sb->p + SWAP_META_PAGES) {
2715 sb = SWAP_PCTRIE_LOOKUP_GE(
2716 &object->un_pager.swp.swp_blks, pi);
2717 if (sb == NULL || sb->p >= e)
2718 break;
2719 for (i = 0; i < SWAP_META_PAGES; i++) {
2720 if (sb->p + i < e &&
2721 sb->d[i] != SWAPBLK_NONE)
2722 count++;
2723 }
2724 }
2725 unlock:
2726 VM_OBJECT_RUNLOCK(object);
2727 }
2728 return (count);
2729 }
2730
2731 /*
2732 * GEOM backend
2733 *
2734 * Swapping onto disk devices.
2735 *
2736 */
2737
2738 static g_orphan_t swapgeom_orphan;
2739
2740 static struct g_class g_swap_class = {
2741 .name = "SWAP",
2742 .version = G_VERSION,
2743 .orphan = swapgeom_orphan,
2744 };
2745
2746 DECLARE_GEOM_CLASS(g_swap_class, g_class);
2747
2748 static void
2749 swapgeom_close_ev(void *arg, int flags)
2750 {
2751 struct g_consumer *cp;
2752
2753 cp = arg;
2754 g_access(cp, -1, -1, 0);
2755 g_detach(cp);
2756 g_destroy_consumer(cp);
2757 }
2758
2759 /*
2760 * Add a reference to the g_consumer for an inflight transaction.
2761 */
2762 static void
2763 swapgeom_acquire(struct g_consumer *cp)
2764 {
2765
2766 mtx_assert(&sw_dev_mtx, MA_OWNED);
2767 cp->index++;
2768 }
2769
2770 /*
2771 * Remove a reference from the g_consumer. Post a close event if all
2772 * references go away, since the function might be called from the
2773 * biodone context.
2774 */
2775 static void
2776 swapgeom_release(struct g_consumer *cp, struct swdevt *sp)
2777 {
2778
2779 mtx_assert(&sw_dev_mtx, MA_OWNED);
2780 cp->index--;
2781 if (cp->index == 0) {
2782 if (g_post_event(swapgeom_close_ev, cp, M_NOWAIT, NULL) == 0)
2783 sp->sw_id = NULL;
2784 }
2785 }
2786
2787 static void
2788 swapgeom_done(struct bio *bp2)
2789 {
2790 struct swdevt *sp;
2791 struct buf *bp;
2792 struct g_consumer *cp;
2793
2794 bp = bp2->bio_caller2;
2795 cp = bp2->bio_from;
2796 bp->b_ioflags = bp2->bio_flags;
2797 if (bp2->bio_error)
2798 bp->b_ioflags |= BIO_ERROR;
2799 bp->b_resid = bp->b_bcount - bp2->bio_completed;
2800 bp->b_error = bp2->bio_error;
2801 bp->b_caller1 = NULL;
2802 bufdone(bp);
2803 sp = bp2->bio_caller1;
2804 mtx_lock(&sw_dev_mtx);
2805 swapgeom_release(cp, sp);
2806 mtx_unlock(&sw_dev_mtx);
2807 g_destroy_bio(bp2);
2808 }
2809
2810 static void
2811 swapgeom_strategy(struct buf *bp, struct swdevt *sp)
2812 {
2813 struct bio *bio;
2814 struct g_consumer *cp;
2815
2816 mtx_lock(&sw_dev_mtx);
2817 cp = sp->sw_id;
2818 if (cp == NULL) {
2819 mtx_unlock(&sw_dev_mtx);
2820 bp->b_error = ENXIO;
2821 bp->b_ioflags |= BIO_ERROR;
2822 bufdone(bp);
2823 return;
2824 }
2825 swapgeom_acquire(cp);
2826 mtx_unlock(&sw_dev_mtx);
2827 if (bp->b_iocmd == BIO_WRITE)
2828 bio = g_new_bio();
2829 else
2830 bio = g_alloc_bio();
2831 if (bio == NULL) {
2832 mtx_lock(&sw_dev_mtx);
2833 swapgeom_release(cp, sp);
2834 mtx_unlock(&sw_dev_mtx);
2835 bp->b_error = ENOMEM;
2836 bp->b_ioflags |= BIO_ERROR;
2837 printf("swap_pager: cannot allocate bio\n");
2838 bufdone(bp);
2839 return;
2840 }
2841
2842 bp->b_caller1 = bio;
2843 bio->bio_caller1 = sp;
2844 bio->bio_caller2 = bp;
2845 bio->bio_cmd = bp->b_iocmd;
2846 bio->bio_offset = (bp->b_blkno - sp->sw_first) * PAGE_SIZE;
2847 bio->bio_length = bp->b_bcount;
2848 bio->bio_done = swapgeom_done;
2849 if (!buf_mapped(bp)) {
2850 bio->bio_ma = bp->b_pages;
2851 bio->bio_data = unmapped_buf;
2852 bio->bio_ma_offset = (vm_offset_t)bp->b_offset & PAGE_MASK;
2853 bio->bio_ma_n = bp->b_npages;
2854 bio->bio_flags |= BIO_UNMAPPED;
2855 } else {
2856 bio->bio_data = bp->b_data;
2857 bio->bio_ma = NULL;
2858 }
2859 g_io_request(bio, cp);
2860 return;
2861 }
2862
2863 static void
2864 swapgeom_orphan(struct g_consumer *cp)
2865 {
2866 struct swdevt *sp;
2867 int destroy;
2868
2869 mtx_lock(&sw_dev_mtx);
2870 TAILQ_FOREACH(sp, &swtailq, sw_list) {
2871 if (sp->sw_id == cp) {
2872 sp->sw_flags |= SW_CLOSING;
2873 break;
2874 }
2875 }
2876 /*
2877 * Drop reference we were created with. Do directly since we're in a
2878 * special context where we don't have to queue the call to
2879 * swapgeom_close_ev().
2880 */
2881 cp->index--;
2882 destroy = ((sp != NULL) && (cp->index == 0));
2883 if (destroy)
2884 sp->sw_id = NULL;
2885 mtx_unlock(&sw_dev_mtx);
2886 if (destroy)
2887 swapgeom_close_ev(cp, 0);
2888 }
2889
2890 static void
2891 swapgeom_close(struct thread *td, struct swdevt *sw)
2892 {
2893 struct g_consumer *cp;
2894
2895 mtx_lock(&sw_dev_mtx);
2896 cp = sw->sw_id;
2897 sw->sw_id = NULL;
2898 mtx_unlock(&sw_dev_mtx);
2899
2900 /*
2901 * swapgeom_close() may be called from the biodone context,
2902 * where we cannot perform topology changes. Delegate the
2903 * work to the events thread.
2904 */
2905 if (cp != NULL)
2906 g_waitfor_event(swapgeom_close_ev, cp, M_WAITOK, NULL);
2907 }
2908
2909 static int
2910 swapongeom_locked(struct cdev *dev, struct vnode *vp)
2911 {
2912 struct g_provider *pp;
2913 struct g_consumer *cp;
2914 static struct g_geom *gp;
2915 struct swdevt *sp;
2916 u_long nblks;
2917 int error;
2918
2919 pp = g_dev_getprovider(dev);
2920 if (pp == NULL)
2921 return (ENODEV);
2922 mtx_lock(&sw_dev_mtx);
2923 TAILQ_FOREACH(sp, &swtailq, sw_list) {
2924 cp = sp->sw_id;
2925 if (cp != NULL && cp->provider == pp) {
2926 mtx_unlock(&sw_dev_mtx);
2927 return (EBUSY);
2928 }
2929 }
2930 mtx_unlock(&sw_dev_mtx);
2931 if (gp == NULL)
2932 gp = g_new_geomf(&g_swap_class, "swap");
2933 cp = g_new_consumer(gp);
2934 cp->index = 1; /* Number of active I/Os, plus one for being active. */
2935 cp->flags |= G_CF_DIRECT_SEND | G_CF_DIRECT_RECEIVE;
2936 g_attach(cp, pp);
2937 /*
2938 * XXX: Every time you think you can improve the margin for
2939 * footshooting, somebody depends on the ability to do so:
2940 * savecore(8) wants to write to our swapdev so we cannot
2941 * set an exclusive count :-(
2942 */
2943 error = g_access(cp, 1, 1, 0);
2944 if (error != 0) {
2945 g_detach(cp);
2946 g_destroy_consumer(cp);
2947 return (error);
2948 }
2949 nblks = pp->mediasize / DEV_BSIZE;
2950 swaponsomething(vp, cp, nblks, swapgeom_strategy,
2951 swapgeom_close, dev2udev(dev),
2952 (pp->flags & G_PF_ACCEPT_UNMAPPED) != 0 ? SW_UNMAPPED : 0);
2953 return (0);
2954 }
2955
2956 static int
2957 swapongeom(struct vnode *vp)
2958 {
2959 int error;
2960
2961 vn_lock(vp, LK_EXCLUSIVE | LK_RETRY);
2962 if (vp->v_type != VCHR || VN_IS_DOOMED(vp)) {
2963 error = ENOENT;
2964 } else {
2965 g_topology_lock();
2966 error = swapongeom_locked(vp->v_rdev, vp);
2967 g_topology_unlock();
2968 }
2969 VOP_UNLOCK(vp);
2970 return (error);
2971 }
2972
2973 /*
2974 * VNODE backend
2975 *
2976 * This is used mainly for network filesystem (read: probably only tested
2977 * with NFS) swapfiles.
2978 *
2979 */
2980
2981 static void
2982 swapdev_strategy(struct buf *bp, struct swdevt *sp)
2983 {
2984 struct vnode *vp2;
2985
2986 bp->b_blkno = ctodb(bp->b_blkno - sp->sw_first);
2987
2988 vp2 = sp->sw_id;
2989 vhold(vp2);
2990 if (bp->b_iocmd == BIO_WRITE) {
2991 if (bp->b_bufobj)
2992 bufobj_wdrop(bp->b_bufobj);
2993 bufobj_wref(&vp2->v_bufobj);
2994 }
2995 if (bp->b_bufobj != &vp2->v_bufobj)
2996 bp->b_bufobj = &vp2->v_bufobj;
2997 bp->b_vp = vp2;
2998 bp->b_iooffset = dbtob(bp->b_blkno);
2999 bstrategy(bp);
3000 return;
3001 }
3002
3003 static void
3004 swapdev_close(struct thread *td, struct swdevt *sp)
3005 {
3006
3007 VOP_CLOSE(sp->sw_vp, FREAD | FWRITE, td->td_ucred, td);
3008 vrele(sp->sw_vp);
3009 }
3010
3011 static int
3012 swaponvp(struct thread *td, struct vnode *vp, u_long nblks)
3013 {
3014 struct swdevt *sp;
3015 int error;
3016
3017 if (nblks == 0)
3018 return (ENXIO);
3019 mtx_lock(&sw_dev_mtx);
3020 TAILQ_FOREACH(sp, &swtailq, sw_list) {
3021 if (sp->sw_id == vp) {
3022 mtx_unlock(&sw_dev_mtx);
3023 return (EBUSY);
3024 }
3025 }
3026 mtx_unlock(&sw_dev_mtx);
3027
3028 (void) vn_lock(vp, LK_EXCLUSIVE | LK_RETRY);
3029 #ifdef MAC
3030 error = mac_system_check_swapon(td->td_ucred, vp);
3031 if (error == 0)
3032 #endif
3033 error = VOP_OPEN(vp, FREAD | FWRITE, td->td_ucred, td, NULL);
3034 (void) VOP_UNLOCK(vp);
3035 if (error)
3036 return (error);
3037
3038 swaponsomething(vp, vp, nblks, swapdev_strategy, swapdev_close,
3039 NODEV, 0);
3040 return (0);
3041 }
3042
3043 static int
3044 sysctl_swap_async_max(SYSCTL_HANDLER_ARGS)
3045 {
3046 int error, new, n;
3047
3048 new = nsw_wcount_async_max;
3049 error = sysctl_handle_int(oidp, &new, 0, req);
3050 if (error != 0 || req->newptr == NULL)
3051 return (error);
3052
3053 if (new > nswbuf / 2 || new < 1)
3054 return (EINVAL);
3055
3056 mtx_lock(&swbuf_mtx);
3057 while (nsw_wcount_async_max != new) {
3058 /*
3059 * Adjust difference. If the current async count is too low,
3060 * we will need to sqeeze our update slowly in. Sleep with a
3061 * higher priority than getpbuf() to finish faster.
3062 */
3063 n = new - nsw_wcount_async_max;
3064 if (nsw_wcount_async + n >= 0) {
3065 nsw_wcount_async += n;
3066 nsw_wcount_async_max += n;
3067 wakeup(&nsw_wcount_async);
3068 } else {
3069 nsw_wcount_async_max -= nsw_wcount_async;
3070 nsw_wcount_async = 0;
3071 msleep(&nsw_wcount_async, &swbuf_mtx, PSWP,
3072 "swpsysctl", 0);
3073 }
3074 }
3075 mtx_unlock(&swbuf_mtx);
3076
3077 return (0);
3078 }
3079
3080 static void
3081 swap_pager_update_writecount(vm_object_t object, vm_offset_t start,
3082 vm_offset_t end)
3083 {
3084
3085 VM_OBJECT_WLOCK(object);
3086 KASSERT((object->flags & OBJ_ANON) == 0,
3087 ("Splittable object with writecount"));
3088 object->un_pager.swp.writemappings += (vm_ooffset_t)end - start;
3089 VM_OBJECT_WUNLOCK(object);
3090 }
3091
3092 static void
3093 swap_pager_release_writecount(vm_object_t object, vm_offset_t start,
3094 vm_offset_t end)
3095 {
3096
3097 VM_OBJECT_WLOCK(object);
3098 KASSERT((object->flags & OBJ_ANON) == 0,
3099 ("Splittable object with writecount"));
3100 object->un_pager.swp.writemappings -= (vm_ooffset_t)end - start;
3101 VM_OBJECT_WUNLOCK(object);
3102 }
Cache object: 01cbe0c2eec3dfa0f46a33b5917e1aba
|