FreeBSD/Linux Kernel Cross Reference
sys/vm/swap_pager.c
1 /*-
2 * Copyright (c) 1998 Matthew Dillon,
3 * Copyright (c) 1994 John S. Dyson
4 * Copyright (c) 1990 University of Utah.
5 * Copyright (c) 1982, 1986, 1989, 1993
6 * The Regents of the University of California. All rights reserved.
7 *
8 * This code is derived from software contributed to Berkeley by
9 * the Systems Programming Group of the University of Utah Computer
10 * Science Department.
11 *
12 * Redistribution and use in source and binary forms, with or without
13 * modification, are permitted provided that the following conditions
14 * are met:
15 * 1. Redistributions of source code must retain the above copyright
16 * notice, this list of conditions and the following disclaimer.
17 * 2. Redistributions in binary form must reproduce the above copyright
18 * notice, this list of conditions and the following disclaimer in the
19 * documentation and/or other materials provided with the distribution.
20 * 3. All advertising materials mentioning features or use of this software
21 * must display the following acknowledgement:
22 * This product includes software developed by the University of
23 * California, Berkeley and its contributors.
24 * 4. Neither the name of the University nor the names of its contributors
25 * may be used to endorse or promote products derived from this software
26 * without specific prior written permission.
27 *
28 * THIS SOFTWARE IS PROVIDED BY THE REGENTS AND CONTRIBUTORS ``AS IS'' AND
29 * ANY EXPRESS OR IMPLIED WARRANTIES, INCLUDING, BUT NOT LIMITED TO, THE
30 * IMPLIED WARRANTIES OF MERCHANTABILITY AND FITNESS FOR A PARTICULAR PURPOSE
31 * ARE DISCLAIMED. IN NO EVENT SHALL THE REGENTS OR CONTRIBUTORS BE LIABLE
32 * FOR ANY DIRECT, INDIRECT, INCIDENTAL, SPECIAL, EXEMPLARY, OR CONSEQUENTIAL
33 * DAMAGES (INCLUDING, BUT NOT LIMITED TO, PROCUREMENT OF SUBSTITUTE GOODS
34 * OR SERVICES; LOSS OF USE, DATA, OR PROFITS; OR BUSINESS INTERRUPTION)
35 * HOWEVER CAUSED AND ON ANY THEORY OF LIABILITY, WHETHER IN CONTRACT, STRICT
36 * LIABILITY, OR TORT (INCLUDING NEGLIGENCE OR OTHERWISE) ARISING IN ANY WAY
37 * OUT OF THE USE OF THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF
38 * SUCH DAMAGE.
39 *
40 * New Swap System
41 * Matthew Dillon
42 *
43 * Radix Bitmap 'blists'.
44 *
45 * - The new swapper uses the new radix bitmap code. This should scale
46 * to arbitrarily small or arbitrarily large swap spaces and an almost
47 * arbitrary degree of fragmentation.
48 *
49 * Features:
50 *
51 * - on the fly reallocation of swap during putpages. The new system
52 * does not try to keep previously allocated swap blocks for dirty
53 * pages.
54 *
55 * - on the fly deallocation of swap
56 *
57 * - No more garbage collection required. Unnecessarily allocated swap
58 * blocks only exist for dirty vm_page_t's now and these are already
59 * cycled (in a high-load system) by the pager. We also do on-the-fly
60 * removal of invalidated swap blocks when a page is destroyed
61 * or renamed.
62 *
63 * from: Utah $Hdr: swap_pager.c 1.4 91/04/30$
64 *
65 * @(#)swap_pager.c 8.9 (Berkeley) 3/21/94
66 * @(#)vm_swap.c 8.5 (Berkeley) 2/17/94
67 */
68
69 #include <sys/cdefs.h>
70 __FBSDID("$FreeBSD: releng/10.1/sys/vm/swap_pager.c 265932 2014-05-12 20:48:04Z des $");
71
72 #include "opt_swap.h"
73 #include "opt_vm.h"
74
75 #include <sys/param.h>
76 #include <sys/systm.h>
77 #include <sys/conf.h>
78 #include <sys/kernel.h>
79 #include <sys/priv.h>
80 #include <sys/proc.h>
81 #include <sys/bio.h>
82 #include <sys/buf.h>
83 #include <sys/disk.h>
84 #include <sys/fcntl.h>
85 #include <sys/mount.h>
86 #include <sys/namei.h>
87 #include <sys/vnode.h>
88 #include <sys/malloc.h>
89 #include <sys/racct.h>
90 #include <sys/resource.h>
91 #include <sys/resourcevar.h>
92 #include <sys/rwlock.h>
93 #include <sys/sysctl.h>
94 #include <sys/sysproto.h>
95 #include <sys/blist.h>
96 #include <sys/lock.h>
97 #include <sys/sx.h>
98 #include <sys/vmmeter.h>
99
100 #include <security/mac/mac_framework.h>
101
102 #include <vm/vm.h>
103 #include <vm/pmap.h>
104 #include <vm/vm_map.h>
105 #include <vm/vm_kern.h>
106 #include <vm/vm_object.h>
107 #include <vm/vm_page.h>
108 #include <vm/vm_pager.h>
109 #include <vm/vm_pageout.h>
110 #include <vm/vm_param.h>
111 #include <vm/swap_pager.h>
112 #include <vm/vm_extern.h>
113 #include <vm/uma.h>
114
115 #include <geom/geom.h>
116
117 /*
118 * SWB_NPAGES must be a power of 2. It may be set to 1, 2, 4, 8, 16
119 * or 32 pages per allocation.
120 * The 32-page limit is due to the radix code (kern/subr_blist.c).
121 */
122 #ifndef MAX_PAGEOUT_CLUSTER
123 #define MAX_PAGEOUT_CLUSTER 16
124 #endif
125
126 #if !defined(SWB_NPAGES)
127 #define SWB_NPAGES MAX_PAGEOUT_CLUSTER
128 #endif
129
130 /*
131 * The swblock structure maps an object and a small, fixed-size range
132 * of page indices to disk addresses within a swap area.
133 * The collection of these mappings is implemented as a hash table.
134 * Unused disk addresses within a swap area are allocated and managed
135 * using a blist.
136 */
137 #define SWCORRECT(n) (sizeof(void *) * (n) / sizeof(daddr_t))
138 #define SWAP_META_PAGES (SWB_NPAGES * 2)
139 #define SWAP_META_MASK (SWAP_META_PAGES - 1)
140
141 struct swblock {
142 struct swblock *swb_hnext;
143 vm_object_t swb_object;
144 vm_pindex_t swb_index;
145 int swb_count;
146 daddr_t swb_pages[SWAP_META_PAGES];
147 };
148
149 static MALLOC_DEFINE(M_VMPGDATA, "vm_pgdata", "swap pager private data");
150 static struct mtx sw_dev_mtx;
151 static TAILQ_HEAD(, swdevt) swtailq = TAILQ_HEAD_INITIALIZER(swtailq);
152 static struct swdevt *swdevhd; /* Allocate from here next */
153 static int nswapdev; /* Number of swap devices */
154 int swap_pager_avail;
155 static int swdev_syscall_active = 0; /* serialize swap(on|off) */
156
157 static vm_ooffset_t swap_total;
158 SYSCTL_QUAD(_vm, OID_AUTO, swap_total, CTLFLAG_RD, &swap_total, 0,
159 "Total amount of available swap storage.");
160 static vm_ooffset_t swap_reserved;
161 SYSCTL_QUAD(_vm, OID_AUTO, swap_reserved, CTLFLAG_RD, &swap_reserved, 0,
162 "Amount of swap storage needed to back all allocated anonymous memory.");
163 static int overcommit = 0;
164 SYSCTL_INT(_vm, OID_AUTO, overcommit, CTLFLAG_RW, &overcommit, 0,
165 "Configure virtual memory overcommit behavior. See tuning(7) "
166 "for details.");
167 static unsigned long swzone;
168 SYSCTL_ULONG(_vm, OID_AUTO, swzone, CTLFLAG_RD, &swzone, 0,
169 "Actual size of swap metadata zone");
170 static unsigned long swap_maxpages;
171 SYSCTL_ULONG(_vm, OID_AUTO, swap_maxpages, CTLFLAG_RD, &swap_maxpages, 0,
172 "Maximum amount of swap supported");
173
174 /* bits from overcommit */
175 #define SWAP_RESERVE_FORCE_ON (1 << 0)
176 #define SWAP_RESERVE_RLIMIT_ON (1 << 1)
177 #define SWAP_RESERVE_ALLOW_NONWIRED (1 << 2)
178
179 int
180 swap_reserve(vm_ooffset_t incr)
181 {
182
183 return (swap_reserve_by_cred(incr, curthread->td_ucred));
184 }
185
186 int
187 swap_reserve_by_cred(vm_ooffset_t incr, struct ucred *cred)
188 {
189 vm_ooffset_t r, s;
190 int res, error;
191 static int curfail;
192 static struct timeval lastfail;
193 struct uidinfo *uip;
194
195 uip = cred->cr_ruidinfo;
196
197 if (incr & PAGE_MASK)
198 panic("swap_reserve: & PAGE_MASK");
199
200 #ifdef RACCT
201 PROC_LOCK(curproc);
202 error = racct_add(curproc, RACCT_SWAP, incr);
203 PROC_UNLOCK(curproc);
204 if (error != 0)
205 return (0);
206 #endif
207
208 res = 0;
209 mtx_lock(&sw_dev_mtx);
210 r = swap_reserved + incr;
211 if (overcommit & SWAP_RESERVE_ALLOW_NONWIRED) {
212 s = cnt.v_page_count - cnt.v_free_reserved - cnt.v_wire_count;
213 s *= PAGE_SIZE;
214 } else
215 s = 0;
216 s += swap_total;
217 if ((overcommit & SWAP_RESERVE_FORCE_ON) == 0 || r <= s ||
218 (error = priv_check(curthread, PRIV_VM_SWAP_NOQUOTA)) == 0) {
219 res = 1;
220 swap_reserved = r;
221 }
222 mtx_unlock(&sw_dev_mtx);
223
224 if (res) {
225 PROC_LOCK(curproc);
226 UIDINFO_VMSIZE_LOCK(uip);
227 if ((overcommit & SWAP_RESERVE_RLIMIT_ON) != 0 &&
228 uip->ui_vmsize + incr > lim_cur(curproc, RLIMIT_SWAP) &&
229 priv_check(curthread, PRIV_VM_SWAP_NORLIMIT))
230 res = 0;
231 else
232 uip->ui_vmsize += incr;
233 UIDINFO_VMSIZE_UNLOCK(uip);
234 PROC_UNLOCK(curproc);
235 if (!res) {
236 mtx_lock(&sw_dev_mtx);
237 swap_reserved -= incr;
238 mtx_unlock(&sw_dev_mtx);
239 }
240 }
241 if (!res && ppsratecheck(&lastfail, &curfail, 1)) {
242 printf("uid %d, pid %d: swap reservation for %jd bytes failed\n",
243 uip->ui_uid, curproc->p_pid, incr);
244 }
245
246 #ifdef RACCT
247 if (!res) {
248 PROC_LOCK(curproc);
249 racct_sub(curproc, RACCT_SWAP, incr);
250 PROC_UNLOCK(curproc);
251 }
252 #endif
253
254 return (res);
255 }
256
257 void
258 swap_reserve_force(vm_ooffset_t incr)
259 {
260 struct uidinfo *uip;
261
262 mtx_lock(&sw_dev_mtx);
263 swap_reserved += incr;
264 mtx_unlock(&sw_dev_mtx);
265
266 #ifdef RACCT
267 PROC_LOCK(curproc);
268 racct_add_force(curproc, RACCT_SWAP, incr);
269 PROC_UNLOCK(curproc);
270 #endif
271
272 uip = curthread->td_ucred->cr_ruidinfo;
273 PROC_LOCK(curproc);
274 UIDINFO_VMSIZE_LOCK(uip);
275 uip->ui_vmsize += incr;
276 UIDINFO_VMSIZE_UNLOCK(uip);
277 PROC_UNLOCK(curproc);
278 }
279
280 void
281 swap_release(vm_ooffset_t decr)
282 {
283 struct ucred *cred;
284
285 PROC_LOCK(curproc);
286 cred = curthread->td_ucred;
287 swap_release_by_cred(decr, cred);
288 PROC_UNLOCK(curproc);
289 }
290
291 void
292 swap_release_by_cred(vm_ooffset_t decr, struct ucred *cred)
293 {
294 struct uidinfo *uip;
295
296 uip = cred->cr_ruidinfo;
297
298 if (decr & PAGE_MASK)
299 panic("swap_release: & PAGE_MASK");
300
301 mtx_lock(&sw_dev_mtx);
302 if (swap_reserved < decr)
303 panic("swap_reserved < decr");
304 swap_reserved -= decr;
305 mtx_unlock(&sw_dev_mtx);
306
307 UIDINFO_VMSIZE_LOCK(uip);
308 if (uip->ui_vmsize < decr)
309 printf("negative vmsize for uid = %d\n", uip->ui_uid);
310 uip->ui_vmsize -= decr;
311 UIDINFO_VMSIZE_UNLOCK(uip);
312
313 racct_sub_cred(cred, RACCT_SWAP, decr);
314 }
315
316 static void swapdev_strategy(struct buf *, struct swdevt *sw);
317
318 #define SWM_FREE 0x02 /* free, period */
319 #define SWM_POP 0x04 /* pop out */
320
321 int swap_pager_full = 2; /* swap space exhaustion (task killing) */
322 static int swap_pager_almost_full = 1; /* swap space exhaustion (w/hysteresis)*/
323 static int nsw_rcount; /* free read buffers */
324 static int nsw_wcount_sync; /* limit write buffers / synchronous */
325 static int nsw_wcount_async; /* limit write buffers / asynchronous */
326 static int nsw_wcount_async_max;/* assigned maximum */
327 static int nsw_cluster_max; /* maximum VOP I/O allowed */
328
329 static struct swblock **swhash;
330 static int swhash_mask;
331 static struct mtx swhash_mtx;
332
333 static int swap_async_max = 4; /* maximum in-progress async I/O's */
334 static struct sx sw_alloc_sx;
335
336
337 SYSCTL_INT(_vm, OID_AUTO, swap_async_max,
338 CTLFLAG_RW, &swap_async_max, 0, "Maximum running async swap ops");
339
340 /*
341 * "named" and "unnamed" anon region objects. Try to reduce the overhead
342 * of searching a named list by hashing it just a little.
343 */
344
345 #define NOBJLISTS 8
346
347 #define NOBJLIST(handle) \
348 (&swap_pager_object_list[((int)(intptr_t)handle >> 4) & (NOBJLISTS-1)])
349
350 static struct mtx sw_alloc_mtx; /* protect list manipulation */
351 static struct pagerlst swap_pager_object_list[NOBJLISTS];
352 static uma_zone_t swap_zone;
353
354 /*
355 * pagerops for OBJT_SWAP - "swap pager". Some ops are also global procedure
356 * calls hooked from other parts of the VM system and do not appear here.
357 * (see vm/swap_pager.h).
358 */
359 static vm_object_t
360 swap_pager_alloc(void *handle, vm_ooffset_t size,
361 vm_prot_t prot, vm_ooffset_t offset, struct ucred *);
362 static void swap_pager_dealloc(vm_object_t object);
363 static int swap_pager_getpages(vm_object_t, vm_page_t *, int, int);
364 static void swap_pager_putpages(vm_object_t, vm_page_t *, int, boolean_t, int *);
365 static boolean_t
366 swap_pager_haspage(vm_object_t object, vm_pindex_t pindex, int *before, int *after);
367 static void swap_pager_init(void);
368 static void swap_pager_unswapped(vm_page_t);
369 static void swap_pager_swapoff(struct swdevt *sp);
370
371 struct pagerops swappagerops = {
372 .pgo_init = swap_pager_init, /* early system initialization of pager */
373 .pgo_alloc = swap_pager_alloc, /* allocate an OBJT_SWAP object */
374 .pgo_dealloc = swap_pager_dealloc, /* deallocate an OBJT_SWAP object */
375 .pgo_getpages = swap_pager_getpages, /* pagein */
376 .pgo_putpages = swap_pager_putpages, /* pageout */
377 .pgo_haspage = swap_pager_haspage, /* get backing store status for page */
378 .pgo_pageunswapped = swap_pager_unswapped, /* remove swap related to page */
379 };
380
381 /*
382 * dmmax is in page-sized chunks with the new swap system. It was
383 * dev-bsized chunks in the old. dmmax is always a power of 2.
384 *
385 * swap_*() routines are externally accessible. swp_*() routines are
386 * internal.
387 */
388 static int dmmax;
389 static int nswap_lowat = 128; /* in pages, swap_pager_almost_full warn */
390 static int nswap_hiwat = 512; /* in pages, swap_pager_almost_full warn */
391
392 SYSCTL_INT(_vm, OID_AUTO, dmmax,
393 CTLFLAG_RD, &dmmax, 0, "Maximum size of a swap block");
394
395 static void swp_sizecheck(void);
396 static void swp_pager_async_iodone(struct buf *bp);
397 static int swapongeom(struct thread *, struct vnode *);
398 static int swaponvp(struct thread *, struct vnode *, u_long);
399 static int swapoff_one(struct swdevt *sp, struct ucred *cred);
400
401 /*
402 * Swap bitmap functions
403 */
404 static void swp_pager_freeswapspace(daddr_t blk, int npages);
405 static daddr_t swp_pager_getswapspace(int npages);
406
407 /*
408 * Metadata functions
409 */
410 static struct swblock **swp_pager_hash(vm_object_t object, vm_pindex_t index);
411 static void swp_pager_meta_build(vm_object_t, vm_pindex_t, daddr_t);
412 static void swp_pager_meta_free(vm_object_t, vm_pindex_t, daddr_t);
413 static void swp_pager_meta_free_all(vm_object_t);
414 static daddr_t swp_pager_meta_ctl(vm_object_t, vm_pindex_t, int);
415
416 static void
417 swp_pager_free_nrpage(vm_page_t m)
418 {
419
420 vm_page_lock(m);
421 if (m->wire_count == 0)
422 vm_page_free(m);
423 vm_page_unlock(m);
424 }
425
426 /*
427 * SWP_SIZECHECK() - update swap_pager_full indication
428 *
429 * update the swap_pager_almost_full indication and warn when we are
430 * about to run out of swap space, using lowat/hiwat hysteresis.
431 *
432 * Clear swap_pager_full ( task killing ) indication when lowat is met.
433 *
434 * No restrictions on call
435 * This routine may not block.
436 */
437 static void
438 swp_sizecheck(void)
439 {
440
441 if (swap_pager_avail < nswap_lowat) {
442 if (swap_pager_almost_full == 0) {
443 printf("swap_pager: out of swap space\n");
444 swap_pager_almost_full = 1;
445 }
446 } else {
447 swap_pager_full = 0;
448 if (swap_pager_avail > nswap_hiwat)
449 swap_pager_almost_full = 0;
450 }
451 }
452
453 /*
454 * SWP_PAGER_HASH() - hash swap meta data
455 *
456 * This is an helper function which hashes the swapblk given
457 * the object and page index. It returns a pointer to a pointer
458 * to the object, or a pointer to a NULL pointer if it could not
459 * find a swapblk.
460 */
461 static struct swblock **
462 swp_pager_hash(vm_object_t object, vm_pindex_t index)
463 {
464 struct swblock **pswap;
465 struct swblock *swap;
466
467 index &= ~(vm_pindex_t)SWAP_META_MASK;
468 pswap = &swhash[(index ^ (int)(intptr_t)object) & swhash_mask];
469 while ((swap = *pswap) != NULL) {
470 if (swap->swb_object == object &&
471 swap->swb_index == index
472 ) {
473 break;
474 }
475 pswap = &swap->swb_hnext;
476 }
477 return (pswap);
478 }
479
480 /*
481 * SWAP_PAGER_INIT() - initialize the swap pager!
482 *
483 * Expected to be started from system init. NOTE: This code is run
484 * before much else so be careful what you depend on. Most of the VM
485 * system has yet to be initialized at this point.
486 */
487 static void
488 swap_pager_init(void)
489 {
490 /*
491 * Initialize object lists
492 */
493 int i;
494
495 for (i = 0; i < NOBJLISTS; ++i)
496 TAILQ_INIT(&swap_pager_object_list[i]);
497 mtx_init(&sw_alloc_mtx, "swap_pager list", NULL, MTX_DEF);
498 mtx_init(&sw_dev_mtx, "swapdev", NULL, MTX_DEF);
499
500 /*
501 * Device Stripe, in PAGE_SIZE'd blocks
502 */
503 dmmax = SWB_NPAGES * 2;
504 }
505
506 /*
507 * SWAP_PAGER_SWAP_INIT() - swap pager initialization from pageout process
508 *
509 * Expected to be started from pageout process once, prior to entering
510 * its main loop.
511 */
512 void
513 swap_pager_swap_init(void)
514 {
515 unsigned long n, n2;
516
517 /*
518 * Number of in-transit swap bp operations. Don't
519 * exhaust the pbufs completely. Make sure we
520 * initialize workable values (0 will work for hysteresis
521 * but it isn't very efficient).
522 *
523 * The nsw_cluster_max is constrained by the bp->b_pages[]
524 * array (MAXPHYS/PAGE_SIZE) and our locally defined
525 * MAX_PAGEOUT_CLUSTER. Also be aware that swap ops are
526 * constrained by the swap device interleave stripe size.
527 *
528 * Currently we hardwire nsw_wcount_async to 4. This limit is
529 * designed to prevent other I/O from having high latencies due to
530 * our pageout I/O. The value 4 works well for one or two active swap
531 * devices but is probably a little low if you have more. Even so,
532 * a higher value would probably generate only a limited improvement
533 * with three or four active swap devices since the system does not
534 * typically have to pageout at extreme bandwidths. We will want
535 * at least 2 per swap devices, and 4 is a pretty good value if you
536 * have one NFS swap device due to the command/ack latency over NFS.
537 * So it all works out pretty well.
538 */
539 nsw_cluster_max = min((MAXPHYS/PAGE_SIZE), MAX_PAGEOUT_CLUSTER);
540
541 mtx_lock(&pbuf_mtx);
542 nsw_rcount = (nswbuf + 1) / 2;
543 nsw_wcount_sync = (nswbuf + 3) / 4;
544 nsw_wcount_async = 4;
545 nsw_wcount_async_max = nsw_wcount_async;
546 mtx_unlock(&pbuf_mtx);
547
548 /*
549 * Initialize our zone. Right now I'm just guessing on the number
550 * we need based on the number of pages in the system. Each swblock
551 * can hold 32 pages, so this is probably overkill. This reservation
552 * is typically limited to around 32MB by default.
553 */
554 n = cnt.v_page_count / 2;
555 if (maxswzone && n > maxswzone / sizeof(struct swblock))
556 n = maxswzone / sizeof(struct swblock);
557 n2 = n;
558 swap_zone = uma_zcreate("SWAPMETA", sizeof(struct swblock), NULL, NULL,
559 NULL, NULL, UMA_ALIGN_PTR, UMA_ZONE_NOFREE | UMA_ZONE_VM);
560 if (swap_zone == NULL)
561 panic("failed to create swap_zone.");
562 do {
563 if (uma_zone_reserve_kva(swap_zone, n))
564 break;
565 /*
566 * if the allocation failed, try a zone two thirds the
567 * size of the previous attempt.
568 */
569 n -= ((n + 2) / 3);
570 } while (n > 0);
571 if (n2 != n)
572 printf("Swap zone entries reduced from %lu to %lu.\n", n2, n);
573 swap_maxpages = n * SWAP_META_PAGES;
574 swzone = n * sizeof(struct swblock);
575 n2 = n;
576
577 /*
578 * Initialize our meta-data hash table. The swapper does not need to
579 * be quite as efficient as the VM system, so we do not use an
580 * oversized hash table.
581 *
582 * n: size of hash table, must be power of 2
583 * swhash_mask: hash table index mask
584 */
585 for (n = 1; n < n2 / 8; n *= 2)
586 ;
587 swhash = malloc(sizeof(struct swblock *) * n, M_VMPGDATA, M_WAITOK | M_ZERO);
588 swhash_mask = n - 1;
589 mtx_init(&swhash_mtx, "swap_pager swhash", NULL, MTX_DEF);
590 }
591
592 /*
593 * SWAP_PAGER_ALLOC() - allocate a new OBJT_SWAP VM object and instantiate
594 * its metadata structures.
595 *
596 * This routine is called from the mmap and fork code to create a new
597 * OBJT_SWAP object. We do this by creating an OBJT_DEFAULT object
598 * and then converting it with swp_pager_meta_build().
599 *
600 * This routine may block in vm_object_allocate() and create a named
601 * object lookup race, so we must interlock.
602 *
603 * MPSAFE
604 */
605 static vm_object_t
606 swap_pager_alloc(void *handle, vm_ooffset_t size, vm_prot_t prot,
607 vm_ooffset_t offset, struct ucred *cred)
608 {
609 vm_object_t object;
610 vm_pindex_t pindex;
611
612 pindex = OFF_TO_IDX(offset + PAGE_MASK + size);
613 if (handle) {
614 mtx_lock(&Giant);
615 /*
616 * Reference existing named region or allocate new one. There
617 * should not be a race here against swp_pager_meta_build()
618 * as called from vm_page_remove() in regards to the lookup
619 * of the handle.
620 */
621 sx_xlock(&sw_alloc_sx);
622 object = vm_pager_object_lookup(NOBJLIST(handle), handle);
623 if (object == NULL) {
624 if (cred != NULL) {
625 if (!swap_reserve_by_cred(size, cred)) {
626 sx_xunlock(&sw_alloc_sx);
627 mtx_unlock(&Giant);
628 return (NULL);
629 }
630 crhold(cred);
631 }
632 object = vm_object_allocate(OBJT_DEFAULT, pindex);
633 VM_OBJECT_WLOCK(object);
634 object->handle = handle;
635 if (cred != NULL) {
636 object->cred = cred;
637 object->charge = size;
638 }
639 swp_pager_meta_build(object, 0, SWAPBLK_NONE);
640 VM_OBJECT_WUNLOCK(object);
641 }
642 sx_xunlock(&sw_alloc_sx);
643 mtx_unlock(&Giant);
644 } else {
645 if (cred != NULL) {
646 if (!swap_reserve_by_cred(size, cred))
647 return (NULL);
648 crhold(cred);
649 }
650 object = vm_object_allocate(OBJT_DEFAULT, pindex);
651 VM_OBJECT_WLOCK(object);
652 if (cred != NULL) {
653 object->cred = cred;
654 object->charge = size;
655 }
656 swp_pager_meta_build(object, 0, SWAPBLK_NONE);
657 VM_OBJECT_WUNLOCK(object);
658 }
659 return (object);
660 }
661
662 /*
663 * SWAP_PAGER_DEALLOC() - remove swap metadata from object
664 *
665 * The swap backing for the object is destroyed. The code is
666 * designed such that we can reinstantiate it later, but this
667 * routine is typically called only when the entire object is
668 * about to be destroyed.
669 *
670 * The object must be locked.
671 */
672 static void
673 swap_pager_dealloc(vm_object_t object)
674 {
675
676 /*
677 * Remove from list right away so lookups will fail if we block for
678 * pageout completion.
679 */
680 if (object->handle != NULL) {
681 mtx_lock(&sw_alloc_mtx);
682 TAILQ_REMOVE(NOBJLIST(object->handle), object, pager_object_list);
683 mtx_unlock(&sw_alloc_mtx);
684 }
685
686 VM_OBJECT_ASSERT_WLOCKED(object);
687 vm_object_pip_wait(object, "swpdea");
688
689 /*
690 * Free all remaining metadata. We only bother to free it from
691 * the swap meta data. We do not attempt to free swapblk's still
692 * associated with vm_page_t's for this object. We do not care
693 * if paging is still in progress on some objects.
694 */
695 swp_pager_meta_free_all(object);
696 }
697
698 /************************************************************************
699 * SWAP PAGER BITMAP ROUTINES *
700 ************************************************************************/
701
702 /*
703 * SWP_PAGER_GETSWAPSPACE() - allocate raw swap space
704 *
705 * Allocate swap for the requested number of pages. The starting
706 * swap block number (a page index) is returned or SWAPBLK_NONE
707 * if the allocation failed.
708 *
709 * Also has the side effect of advising that somebody made a mistake
710 * when they configured swap and didn't configure enough.
711 *
712 * This routine may not sleep.
713 *
714 * We allocate in round-robin fashion from the configured devices.
715 */
716 static daddr_t
717 swp_pager_getswapspace(int npages)
718 {
719 daddr_t blk;
720 struct swdevt *sp;
721 int i;
722
723 blk = SWAPBLK_NONE;
724 mtx_lock(&sw_dev_mtx);
725 sp = swdevhd;
726 for (i = 0; i < nswapdev; i++) {
727 if (sp == NULL)
728 sp = TAILQ_FIRST(&swtailq);
729 if (!(sp->sw_flags & SW_CLOSING)) {
730 blk = blist_alloc(sp->sw_blist, npages);
731 if (blk != SWAPBLK_NONE) {
732 blk += sp->sw_first;
733 sp->sw_used += npages;
734 swap_pager_avail -= npages;
735 swp_sizecheck();
736 swdevhd = TAILQ_NEXT(sp, sw_list);
737 goto done;
738 }
739 }
740 sp = TAILQ_NEXT(sp, sw_list);
741 }
742 if (swap_pager_full != 2) {
743 printf("swap_pager_getswapspace(%d): failed\n", npages);
744 swap_pager_full = 2;
745 swap_pager_almost_full = 1;
746 }
747 swdevhd = NULL;
748 done:
749 mtx_unlock(&sw_dev_mtx);
750 return (blk);
751 }
752
753 static int
754 swp_pager_isondev(daddr_t blk, struct swdevt *sp)
755 {
756
757 return (blk >= sp->sw_first && blk < sp->sw_end);
758 }
759
760 static void
761 swp_pager_strategy(struct buf *bp)
762 {
763 struct swdevt *sp;
764
765 mtx_lock(&sw_dev_mtx);
766 TAILQ_FOREACH(sp, &swtailq, sw_list) {
767 if (bp->b_blkno >= sp->sw_first && bp->b_blkno < sp->sw_end) {
768 mtx_unlock(&sw_dev_mtx);
769 if ((sp->sw_flags & SW_UNMAPPED) != 0 &&
770 unmapped_buf_allowed) {
771 bp->b_kvaalloc = bp->b_data;
772 bp->b_data = unmapped_buf;
773 bp->b_kvabase = unmapped_buf;
774 bp->b_offset = 0;
775 bp->b_flags |= B_UNMAPPED;
776 } else {
777 pmap_qenter((vm_offset_t)bp->b_data,
778 &bp->b_pages[0], bp->b_bcount / PAGE_SIZE);
779 }
780 sp->sw_strategy(bp, sp);
781 return;
782 }
783 }
784 panic("Swapdev not found");
785 }
786
787
788 /*
789 * SWP_PAGER_FREESWAPSPACE() - free raw swap space
790 *
791 * This routine returns the specified swap blocks back to the bitmap.
792 *
793 * This routine may not sleep.
794 */
795 static void
796 swp_pager_freeswapspace(daddr_t blk, int npages)
797 {
798 struct swdevt *sp;
799
800 mtx_lock(&sw_dev_mtx);
801 TAILQ_FOREACH(sp, &swtailq, sw_list) {
802 if (blk >= sp->sw_first && blk < sp->sw_end) {
803 sp->sw_used -= npages;
804 /*
805 * If we are attempting to stop swapping on
806 * this device, we don't want to mark any
807 * blocks free lest they be reused.
808 */
809 if ((sp->sw_flags & SW_CLOSING) == 0) {
810 blist_free(sp->sw_blist, blk - sp->sw_first,
811 npages);
812 swap_pager_avail += npages;
813 swp_sizecheck();
814 }
815 mtx_unlock(&sw_dev_mtx);
816 return;
817 }
818 }
819 panic("Swapdev not found");
820 }
821
822 /*
823 * SWAP_PAGER_FREESPACE() - frees swap blocks associated with a page
824 * range within an object.
825 *
826 * This is a globally accessible routine.
827 *
828 * This routine removes swapblk assignments from swap metadata.
829 *
830 * The external callers of this routine typically have already destroyed
831 * or renamed vm_page_t's associated with this range in the object so
832 * we should be ok.
833 *
834 * The object must be locked.
835 */
836 void
837 swap_pager_freespace(vm_object_t object, vm_pindex_t start, vm_size_t size)
838 {
839
840 swp_pager_meta_free(object, start, size);
841 }
842
843 /*
844 * SWAP_PAGER_RESERVE() - reserve swap blocks in object
845 *
846 * Assigns swap blocks to the specified range within the object. The
847 * swap blocks are not zeroed. Any previous swap assignment is destroyed.
848 *
849 * Returns 0 on success, -1 on failure.
850 */
851 int
852 swap_pager_reserve(vm_object_t object, vm_pindex_t start, vm_size_t size)
853 {
854 int n = 0;
855 daddr_t blk = SWAPBLK_NONE;
856 vm_pindex_t beg = start; /* save start index */
857
858 VM_OBJECT_WLOCK(object);
859 while (size) {
860 if (n == 0) {
861 n = BLIST_MAX_ALLOC;
862 while ((blk = swp_pager_getswapspace(n)) == SWAPBLK_NONE) {
863 n >>= 1;
864 if (n == 0) {
865 swp_pager_meta_free(object, beg, start - beg);
866 VM_OBJECT_WUNLOCK(object);
867 return (-1);
868 }
869 }
870 }
871 swp_pager_meta_build(object, start, blk);
872 --size;
873 ++start;
874 ++blk;
875 --n;
876 }
877 swp_pager_meta_free(object, start, n);
878 VM_OBJECT_WUNLOCK(object);
879 return (0);
880 }
881
882 /*
883 * SWAP_PAGER_COPY() - copy blocks from source pager to destination pager
884 * and destroy the source.
885 *
886 * Copy any valid swapblks from the source to the destination. In
887 * cases where both the source and destination have a valid swapblk,
888 * we keep the destination's.
889 *
890 * This routine is allowed to sleep. It may sleep allocating metadata
891 * indirectly through swp_pager_meta_build() or if paging is still in
892 * progress on the source.
893 *
894 * The source object contains no vm_page_t's (which is just as well)
895 *
896 * The source object is of type OBJT_SWAP.
897 *
898 * The source and destination objects must be locked.
899 * Both object locks may temporarily be released.
900 */
901 void
902 swap_pager_copy(vm_object_t srcobject, vm_object_t dstobject,
903 vm_pindex_t offset, int destroysource)
904 {
905 vm_pindex_t i;
906
907 VM_OBJECT_ASSERT_WLOCKED(srcobject);
908 VM_OBJECT_ASSERT_WLOCKED(dstobject);
909
910 /*
911 * If destroysource is set, we remove the source object from the
912 * swap_pager internal queue now.
913 */
914 if (destroysource) {
915 if (srcobject->handle != NULL) {
916 mtx_lock(&sw_alloc_mtx);
917 TAILQ_REMOVE(
918 NOBJLIST(srcobject->handle),
919 srcobject,
920 pager_object_list
921 );
922 mtx_unlock(&sw_alloc_mtx);
923 }
924 }
925
926 /*
927 * transfer source to destination.
928 */
929 for (i = 0; i < dstobject->size; ++i) {
930 daddr_t dstaddr;
931
932 /*
933 * Locate (without changing) the swapblk on the destination,
934 * unless it is invalid in which case free it silently, or
935 * if the destination is a resident page, in which case the
936 * source is thrown away.
937 */
938 dstaddr = swp_pager_meta_ctl(dstobject, i, 0);
939
940 if (dstaddr == SWAPBLK_NONE) {
941 /*
942 * Destination has no swapblk and is not resident,
943 * copy source.
944 */
945 daddr_t srcaddr;
946
947 srcaddr = swp_pager_meta_ctl(
948 srcobject,
949 i + offset,
950 SWM_POP
951 );
952
953 if (srcaddr != SWAPBLK_NONE) {
954 /*
955 * swp_pager_meta_build() can sleep.
956 */
957 vm_object_pip_add(srcobject, 1);
958 VM_OBJECT_WUNLOCK(srcobject);
959 vm_object_pip_add(dstobject, 1);
960 swp_pager_meta_build(dstobject, i, srcaddr);
961 vm_object_pip_wakeup(dstobject);
962 VM_OBJECT_WLOCK(srcobject);
963 vm_object_pip_wakeup(srcobject);
964 }
965 } else {
966 /*
967 * Destination has valid swapblk or it is represented
968 * by a resident page. We destroy the sourceblock.
969 */
970
971 swp_pager_meta_ctl(srcobject, i + offset, SWM_FREE);
972 }
973 }
974
975 /*
976 * Free left over swap blocks in source.
977 *
978 * We have to revert the type to OBJT_DEFAULT so we do not accidently
979 * double-remove the object from the swap queues.
980 */
981 if (destroysource) {
982 swp_pager_meta_free_all(srcobject);
983 /*
984 * Reverting the type is not necessary, the caller is going
985 * to destroy srcobject directly, but I'm doing it here
986 * for consistency since we've removed the object from its
987 * queues.
988 */
989 srcobject->type = OBJT_DEFAULT;
990 }
991 }
992
993 /*
994 * SWAP_PAGER_HASPAGE() - determine if we have good backing store for
995 * the requested page.
996 *
997 * We determine whether good backing store exists for the requested
998 * page and return TRUE if it does, FALSE if it doesn't.
999 *
1000 * If TRUE, we also try to determine how much valid, contiguous backing
1001 * store exists before and after the requested page within a reasonable
1002 * distance. We do not try to restrict it to the swap device stripe
1003 * (that is handled in getpages/putpages). It probably isn't worth
1004 * doing here.
1005 */
1006 static boolean_t
1007 swap_pager_haspage(vm_object_t object, vm_pindex_t pindex, int *before, int *after)
1008 {
1009 daddr_t blk0;
1010
1011 VM_OBJECT_ASSERT_LOCKED(object);
1012 /*
1013 * do we have good backing store at the requested index ?
1014 */
1015 blk0 = swp_pager_meta_ctl(object, pindex, 0);
1016
1017 if (blk0 == SWAPBLK_NONE) {
1018 if (before)
1019 *before = 0;
1020 if (after)
1021 *after = 0;
1022 return (FALSE);
1023 }
1024
1025 /*
1026 * find backwards-looking contiguous good backing store
1027 */
1028 if (before != NULL) {
1029 int i;
1030
1031 for (i = 1; i < (SWB_NPAGES/2); ++i) {
1032 daddr_t blk;
1033
1034 if (i > pindex)
1035 break;
1036 blk = swp_pager_meta_ctl(object, pindex - i, 0);
1037 if (blk != blk0 - i)
1038 break;
1039 }
1040 *before = (i - 1);
1041 }
1042
1043 /*
1044 * find forward-looking contiguous good backing store
1045 */
1046 if (after != NULL) {
1047 int i;
1048
1049 for (i = 1; i < (SWB_NPAGES/2); ++i) {
1050 daddr_t blk;
1051
1052 blk = swp_pager_meta_ctl(object, pindex + i, 0);
1053 if (blk != blk0 + i)
1054 break;
1055 }
1056 *after = (i - 1);
1057 }
1058 return (TRUE);
1059 }
1060
1061 /*
1062 * SWAP_PAGER_PAGE_UNSWAPPED() - remove swap backing store related to page
1063 *
1064 * This removes any associated swap backing store, whether valid or
1065 * not, from the page.
1066 *
1067 * This routine is typically called when a page is made dirty, at
1068 * which point any associated swap can be freed. MADV_FREE also
1069 * calls us in a special-case situation
1070 *
1071 * NOTE!!! If the page is clean and the swap was valid, the caller
1072 * should make the page dirty before calling this routine. This routine
1073 * does NOT change the m->dirty status of the page. Also: MADV_FREE
1074 * depends on it.
1075 *
1076 * This routine may not sleep.
1077 *
1078 * The object containing the page must be locked.
1079 */
1080 static void
1081 swap_pager_unswapped(vm_page_t m)
1082 {
1083
1084 swp_pager_meta_ctl(m->object, m->pindex, SWM_FREE);
1085 }
1086
1087 /*
1088 * SWAP_PAGER_GETPAGES() - bring pages in from swap
1089 *
1090 * Attempt to retrieve (m, count) pages from backing store, but make
1091 * sure we retrieve at least m[reqpage]. We try to load in as large
1092 * a chunk surrounding m[reqpage] as is contiguous in swap and which
1093 * belongs to the same object.
1094 *
1095 * The code is designed for asynchronous operation and
1096 * immediate-notification of 'reqpage' but tends not to be
1097 * used that way. Please do not optimize-out this algorithmic
1098 * feature, I intend to improve on it in the future.
1099 *
1100 * The parent has a single vm_object_pip_add() reference prior to
1101 * calling us and we should return with the same.
1102 *
1103 * The parent has BUSY'd the pages. We should return with 'm'
1104 * left busy, but the others adjusted.
1105 */
1106 static int
1107 swap_pager_getpages(vm_object_t object, vm_page_t *m, int count, int reqpage)
1108 {
1109 struct buf *bp;
1110 vm_page_t mreq;
1111 int i;
1112 int j;
1113 daddr_t blk;
1114
1115 mreq = m[reqpage];
1116
1117 KASSERT(mreq->object == object,
1118 ("swap_pager_getpages: object mismatch %p/%p",
1119 object, mreq->object));
1120
1121 /*
1122 * Calculate range to retrieve. The pages have already been assigned
1123 * their swapblks. We require a *contiguous* range but we know it to
1124 * not span devices. If we do not supply it, bad things
1125 * happen. Note that blk, iblk & jblk can be SWAPBLK_NONE, but the
1126 * loops are set up such that the case(s) are handled implicitly.
1127 *
1128 * The swp_*() calls must be made with the object locked.
1129 */
1130 blk = swp_pager_meta_ctl(mreq->object, mreq->pindex, 0);
1131
1132 for (i = reqpage - 1; i >= 0; --i) {
1133 daddr_t iblk;
1134
1135 iblk = swp_pager_meta_ctl(m[i]->object, m[i]->pindex, 0);
1136 if (blk != iblk + (reqpage - i))
1137 break;
1138 }
1139 ++i;
1140
1141 for (j = reqpage + 1; j < count; ++j) {
1142 daddr_t jblk;
1143
1144 jblk = swp_pager_meta_ctl(m[j]->object, m[j]->pindex, 0);
1145 if (blk != jblk - (j - reqpage))
1146 break;
1147 }
1148
1149 /*
1150 * free pages outside our collection range. Note: we never free
1151 * mreq, it must remain busy throughout.
1152 */
1153 if (0 < i || j < count) {
1154 int k;
1155
1156 for (k = 0; k < i; ++k)
1157 swp_pager_free_nrpage(m[k]);
1158 for (k = j; k < count; ++k)
1159 swp_pager_free_nrpage(m[k]);
1160 }
1161
1162 /*
1163 * Return VM_PAGER_FAIL if we have nothing to do. Return mreq
1164 * still busy, but the others unbusied.
1165 */
1166 if (blk == SWAPBLK_NONE)
1167 return (VM_PAGER_FAIL);
1168
1169 /*
1170 * Getpbuf() can sleep.
1171 */
1172 VM_OBJECT_WUNLOCK(object);
1173 /*
1174 * Get a swap buffer header to perform the IO
1175 */
1176 bp = getpbuf(&nsw_rcount);
1177 bp->b_flags |= B_PAGING;
1178
1179 bp->b_iocmd = BIO_READ;
1180 bp->b_iodone = swp_pager_async_iodone;
1181 bp->b_rcred = crhold(thread0.td_ucred);
1182 bp->b_wcred = crhold(thread0.td_ucred);
1183 bp->b_blkno = blk - (reqpage - i);
1184 bp->b_bcount = PAGE_SIZE * (j - i);
1185 bp->b_bufsize = PAGE_SIZE * (j - i);
1186 bp->b_pager.pg_reqpage = reqpage - i;
1187
1188 VM_OBJECT_WLOCK(object);
1189 {
1190 int k;
1191
1192 for (k = i; k < j; ++k) {
1193 bp->b_pages[k - i] = m[k];
1194 m[k]->oflags |= VPO_SWAPINPROG;
1195 }
1196 }
1197 bp->b_npages = j - i;
1198
1199 PCPU_INC(cnt.v_swapin);
1200 PCPU_ADD(cnt.v_swappgsin, bp->b_npages);
1201
1202 /*
1203 * We still hold the lock on mreq, and our automatic completion routine
1204 * does not remove it.
1205 */
1206 vm_object_pip_add(object, bp->b_npages);
1207 VM_OBJECT_WUNLOCK(object);
1208
1209 /*
1210 * perform the I/O. NOTE!!! bp cannot be considered valid after
1211 * this point because we automatically release it on completion.
1212 * Instead, we look at the one page we are interested in which we
1213 * still hold a lock on even through the I/O completion.
1214 *
1215 * The other pages in our m[] array are also released on completion,
1216 * so we cannot assume they are valid anymore either.
1217 *
1218 * NOTE: b_blkno is destroyed by the call to swapdev_strategy
1219 */
1220 BUF_KERNPROC(bp);
1221 swp_pager_strategy(bp);
1222
1223 /*
1224 * wait for the page we want to complete. VPO_SWAPINPROG is always
1225 * cleared on completion. If an I/O error occurs, SWAPBLK_NONE
1226 * is set in the meta-data.
1227 */
1228 VM_OBJECT_WLOCK(object);
1229 while ((mreq->oflags & VPO_SWAPINPROG) != 0) {
1230 mreq->oflags |= VPO_SWAPSLEEP;
1231 PCPU_INC(cnt.v_intrans);
1232 if (VM_OBJECT_SLEEP(object, &object->paging_in_progress, PSWP,
1233 "swread", hz * 20)) {
1234 printf(
1235 "swap_pager: indefinite wait buffer: bufobj: %p, blkno: %jd, size: %ld\n",
1236 bp->b_bufobj, (intmax_t)bp->b_blkno, bp->b_bcount);
1237 }
1238 }
1239
1240 /*
1241 * mreq is left busied after completion, but all the other pages
1242 * are freed. If we had an unrecoverable read error the page will
1243 * not be valid.
1244 */
1245 if (mreq->valid != VM_PAGE_BITS_ALL) {
1246 return (VM_PAGER_ERROR);
1247 } else {
1248 return (VM_PAGER_OK);
1249 }
1250
1251 /*
1252 * A final note: in a low swap situation, we cannot deallocate swap
1253 * and mark a page dirty here because the caller is likely to mark
1254 * the page clean when we return, causing the page to possibly revert
1255 * to all-zero's later.
1256 */
1257 }
1258
1259 /*
1260 * swap_pager_putpages:
1261 *
1262 * Assign swap (if necessary) and initiate I/O on the specified pages.
1263 *
1264 * We support both OBJT_DEFAULT and OBJT_SWAP objects. DEFAULT objects
1265 * are automatically converted to SWAP objects.
1266 *
1267 * In a low memory situation we may block in VOP_STRATEGY(), but the new
1268 * vm_page reservation system coupled with properly written VFS devices
1269 * should ensure that no low-memory deadlock occurs. This is an area
1270 * which needs work.
1271 *
1272 * The parent has N vm_object_pip_add() references prior to
1273 * calling us and will remove references for rtvals[] that are
1274 * not set to VM_PAGER_PEND. We need to remove the rest on I/O
1275 * completion.
1276 *
1277 * The parent has soft-busy'd the pages it passes us and will unbusy
1278 * those whos rtvals[] entry is not set to VM_PAGER_PEND on return.
1279 * We need to unbusy the rest on I/O completion.
1280 */
1281 void
1282 swap_pager_putpages(vm_object_t object, vm_page_t *m, int count,
1283 boolean_t sync, int *rtvals)
1284 {
1285 int i;
1286 int n = 0;
1287
1288 if (count && m[0]->object != object) {
1289 panic("swap_pager_putpages: object mismatch %p/%p",
1290 object,
1291 m[0]->object
1292 );
1293 }
1294
1295 /*
1296 * Step 1
1297 *
1298 * Turn object into OBJT_SWAP
1299 * check for bogus sysops
1300 * force sync if not pageout process
1301 */
1302 if (object->type != OBJT_SWAP)
1303 swp_pager_meta_build(object, 0, SWAPBLK_NONE);
1304 VM_OBJECT_WUNLOCK(object);
1305
1306 if (curproc != pageproc)
1307 sync = TRUE;
1308
1309 /*
1310 * Step 2
1311 *
1312 * Update nsw parameters from swap_async_max sysctl values.
1313 * Do not let the sysop crash the machine with bogus numbers.
1314 */
1315 mtx_lock(&pbuf_mtx);
1316 if (swap_async_max != nsw_wcount_async_max) {
1317 int n;
1318
1319 /*
1320 * limit range
1321 */
1322 if ((n = swap_async_max) > nswbuf / 2)
1323 n = nswbuf / 2;
1324 if (n < 1)
1325 n = 1;
1326 swap_async_max = n;
1327
1328 /*
1329 * Adjust difference ( if possible ). If the current async
1330 * count is too low, we may not be able to make the adjustment
1331 * at this time.
1332 */
1333 n -= nsw_wcount_async_max;
1334 if (nsw_wcount_async + n >= 0) {
1335 nsw_wcount_async += n;
1336 nsw_wcount_async_max += n;
1337 wakeup(&nsw_wcount_async);
1338 }
1339 }
1340 mtx_unlock(&pbuf_mtx);
1341
1342 /*
1343 * Step 3
1344 *
1345 * Assign swap blocks and issue I/O. We reallocate swap on the fly.
1346 * The page is left dirty until the pageout operation completes
1347 * successfully.
1348 */
1349 for (i = 0; i < count; i += n) {
1350 int j;
1351 struct buf *bp;
1352 daddr_t blk;
1353
1354 /*
1355 * Maximum I/O size is limited by a number of factors.
1356 */
1357 n = min(BLIST_MAX_ALLOC, count - i);
1358 n = min(n, nsw_cluster_max);
1359
1360 /*
1361 * Get biggest block of swap we can. If we fail, fall
1362 * back and try to allocate a smaller block. Don't go
1363 * overboard trying to allocate space if it would overly
1364 * fragment swap.
1365 */
1366 while (
1367 (blk = swp_pager_getswapspace(n)) == SWAPBLK_NONE &&
1368 n > 4
1369 ) {
1370 n >>= 1;
1371 }
1372 if (blk == SWAPBLK_NONE) {
1373 for (j = 0; j < n; ++j)
1374 rtvals[i+j] = VM_PAGER_FAIL;
1375 continue;
1376 }
1377
1378 /*
1379 * All I/O parameters have been satisfied, build the I/O
1380 * request and assign the swap space.
1381 */
1382 if (sync == TRUE) {
1383 bp = getpbuf(&nsw_wcount_sync);
1384 } else {
1385 bp = getpbuf(&nsw_wcount_async);
1386 bp->b_flags = B_ASYNC;
1387 }
1388 bp->b_flags |= B_PAGING;
1389 bp->b_iocmd = BIO_WRITE;
1390
1391 bp->b_rcred = crhold(thread0.td_ucred);
1392 bp->b_wcred = crhold(thread0.td_ucred);
1393 bp->b_bcount = PAGE_SIZE * n;
1394 bp->b_bufsize = PAGE_SIZE * n;
1395 bp->b_blkno = blk;
1396
1397 VM_OBJECT_WLOCK(object);
1398 for (j = 0; j < n; ++j) {
1399 vm_page_t mreq = m[i+j];
1400
1401 swp_pager_meta_build(
1402 mreq->object,
1403 mreq->pindex,
1404 blk + j
1405 );
1406 vm_page_dirty(mreq);
1407 rtvals[i+j] = VM_PAGER_OK;
1408
1409 mreq->oflags |= VPO_SWAPINPROG;
1410 bp->b_pages[j] = mreq;
1411 }
1412 VM_OBJECT_WUNLOCK(object);
1413 bp->b_npages = n;
1414 /*
1415 * Must set dirty range for NFS to work.
1416 */
1417 bp->b_dirtyoff = 0;
1418 bp->b_dirtyend = bp->b_bcount;
1419
1420 PCPU_INC(cnt.v_swapout);
1421 PCPU_ADD(cnt.v_swappgsout, bp->b_npages);
1422
1423 /*
1424 * asynchronous
1425 *
1426 * NOTE: b_blkno is destroyed by the call to swapdev_strategy
1427 */
1428 if (sync == FALSE) {
1429 bp->b_iodone = swp_pager_async_iodone;
1430 BUF_KERNPROC(bp);
1431 swp_pager_strategy(bp);
1432
1433 for (j = 0; j < n; ++j)
1434 rtvals[i+j] = VM_PAGER_PEND;
1435 /* restart outter loop */
1436 continue;
1437 }
1438
1439 /*
1440 * synchronous
1441 *
1442 * NOTE: b_blkno is destroyed by the call to swapdev_strategy
1443 */
1444 bp->b_iodone = bdone;
1445 swp_pager_strategy(bp);
1446
1447 /*
1448 * Wait for the sync I/O to complete, then update rtvals.
1449 * We just set the rtvals[] to VM_PAGER_PEND so we can call
1450 * our async completion routine at the end, thus avoiding a
1451 * double-free.
1452 */
1453 bwait(bp, PVM, "swwrt");
1454 for (j = 0; j < n; ++j)
1455 rtvals[i+j] = VM_PAGER_PEND;
1456 /*
1457 * Now that we are through with the bp, we can call the
1458 * normal async completion, which frees everything up.
1459 */
1460 swp_pager_async_iodone(bp);
1461 }
1462 VM_OBJECT_WLOCK(object);
1463 }
1464
1465 /*
1466 * swp_pager_async_iodone:
1467 *
1468 * Completion routine for asynchronous reads and writes from/to swap.
1469 * Also called manually by synchronous code to finish up a bp.
1470 *
1471 * This routine may not sleep.
1472 */
1473 static void
1474 swp_pager_async_iodone(struct buf *bp)
1475 {
1476 int i;
1477 vm_object_t object = NULL;
1478
1479 /*
1480 * report error
1481 */
1482 if (bp->b_ioflags & BIO_ERROR) {
1483 printf(
1484 "swap_pager: I/O error - %s failed; blkno %ld,"
1485 "size %ld, error %d\n",
1486 ((bp->b_iocmd == BIO_READ) ? "pagein" : "pageout"),
1487 (long)bp->b_blkno,
1488 (long)bp->b_bcount,
1489 bp->b_error
1490 );
1491 }
1492
1493 /*
1494 * remove the mapping for kernel virtual
1495 */
1496 if ((bp->b_flags & B_UNMAPPED) != 0) {
1497 bp->b_data = bp->b_kvaalloc;
1498 bp->b_kvabase = bp->b_kvaalloc;
1499 bp->b_flags &= ~B_UNMAPPED;
1500 } else
1501 pmap_qremove((vm_offset_t)bp->b_data, bp->b_npages);
1502
1503 if (bp->b_npages) {
1504 object = bp->b_pages[0]->object;
1505 VM_OBJECT_WLOCK(object);
1506 }
1507
1508 /*
1509 * cleanup pages. If an error occurs writing to swap, we are in
1510 * very serious trouble. If it happens to be a disk error, though,
1511 * we may be able to recover by reassigning the swap later on. So
1512 * in this case we remove the m->swapblk assignment for the page
1513 * but do not free it in the rlist. The errornous block(s) are thus
1514 * never reallocated as swap. Redirty the page and continue.
1515 */
1516 for (i = 0; i < bp->b_npages; ++i) {
1517 vm_page_t m = bp->b_pages[i];
1518
1519 m->oflags &= ~VPO_SWAPINPROG;
1520 if (m->oflags & VPO_SWAPSLEEP) {
1521 m->oflags &= ~VPO_SWAPSLEEP;
1522 wakeup(&object->paging_in_progress);
1523 }
1524
1525 if (bp->b_ioflags & BIO_ERROR) {
1526 /*
1527 * If an error occurs I'd love to throw the swapblk
1528 * away without freeing it back to swapspace, so it
1529 * can never be used again. But I can't from an
1530 * interrupt.
1531 */
1532 if (bp->b_iocmd == BIO_READ) {
1533 /*
1534 * When reading, reqpage needs to stay
1535 * locked for the parent, but all other
1536 * pages can be freed. We still want to
1537 * wakeup the parent waiting on the page,
1538 * though. ( also: pg_reqpage can be -1 and
1539 * not match anything ).
1540 *
1541 * We have to wake specifically requested pages
1542 * up too because we cleared VPO_SWAPINPROG and
1543 * someone may be waiting for that.
1544 *
1545 * NOTE: for reads, m->dirty will probably
1546 * be overridden by the original caller of
1547 * getpages so don't play cute tricks here.
1548 */
1549 m->valid = 0;
1550 if (i != bp->b_pager.pg_reqpage)
1551 swp_pager_free_nrpage(m);
1552 else {
1553 vm_page_lock(m);
1554 vm_page_flash(m);
1555 vm_page_unlock(m);
1556 }
1557 /*
1558 * If i == bp->b_pager.pg_reqpage, do not wake
1559 * the page up. The caller needs to.
1560 */
1561 } else {
1562 /*
1563 * If a write error occurs, reactivate page
1564 * so it doesn't clog the inactive list,
1565 * then finish the I/O.
1566 */
1567 vm_page_dirty(m);
1568 vm_page_lock(m);
1569 vm_page_activate(m);
1570 vm_page_unlock(m);
1571 vm_page_sunbusy(m);
1572 }
1573 } else if (bp->b_iocmd == BIO_READ) {
1574 /*
1575 * NOTE: for reads, m->dirty will probably be
1576 * overridden by the original caller of getpages so
1577 * we cannot set them in order to free the underlying
1578 * swap in a low-swap situation. I don't think we'd
1579 * want to do that anyway, but it was an optimization
1580 * that existed in the old swapper for a time before
1581 * it got ripped out due to precisely this problem.
1582 *
1583 * If not the requested page then deactivate it.
1584 *
1585 * Note that the requested page, reqpage, is left
1586 * busied, but we still have to wake it up. The
1587 * other pages are released (unbusied) by
1588 * vm_page_xunbusy().
1589 */
1590 KASSERT(!pmap_page_is_mapped(m),
1591 ("swp_pager_async_iodone: page %p is mapped", m));
1592 m->valid = VM_PAGE_BITS_ALL;
1593 KASSERT(m->dirty == 0,
1594 ("swp_pager_async_iodone: page %p is dirty", m));
1595
1596 /*
1597 * We have to wake specifically requested pages
1598 * up too because we cleared VPO_SWAPINPROG and
1599 * could be waiting for it in getpages. However,
1600 * be sure to not unbusy getpages specifically
1601 * requested page - getpages expects it to be
1602 * left busy.
1603 */
1604 if (i != bp->b_pager.pg_reqpage) {
1605 vm_page_lock(m);
1606 vm_page_deactivate(m);
1607 vm_page_unlock(m);
1608 vm_page_xunbusy(m);
1609 } else {
1610 vm_page_lock(m);
1611 vm_page_flash(m);
1612 vm_page_unlock(m);
1613 }
1614 } else {
1615 /*
1616 * For write success, clear the dirty
1617 * status, then finish the I/O ( which decrements the
1618 * busy count and possibly wakes waiter's up ).
1619 */
1620 KASSERT(!pmap_page_is_write_mapped(m),
1621 ("swp_pager_async_iodone: page %p is not write"
1622 " protected", m));
1623 vm_page_undirty(m);
1624 vm_page_sunbusy(m);
1625 if (vm_page_count_severe()) {
1626 vm_page_lock(m);
1627 vm_page_try_to_cache(m);
1628 vm_page_unlock(m);
1629 }
1630 }
1631 }
1632
1633 /*
1634 * adjust pip. NOTE: the original parent may still have its own
1635 * pip refs on the object.
1636 */
1637 if (object != NULL) {
1638 vm_object_pip_wakeupn(object, bp->b_npages);
1639 VM_OBJECT_WUNLOCK(object);
1640 }
1641
1642 /*
1643 * swapdev_strategy() manually sets b_vp and b_bufobj before calling
1644 * bstrategy(). Set them back to NULL now we're done with it, or we'll
1645 * trigger a KASSERT in relpbuf().
1646 */
1647 if (bp->b_vp) {
1648 bp->b_vp = NULL;
1649 bp->b_bufobj = NULL;
1650 }
1651 /*
1652 * release the physical I/O buffer
1653 */
1654 relpbuf(
1655 bp,
1656 ((bp->b_iocmd == BIO_READ) ? &nsw_rcount :
1657 ((bp->b_flags & B_ASYNC) ?
1658 &nsw_wcount_async :
1659 &nsw_wcount_sync
1660 )
1661 )
1662 );
1663 }
1664
1665 /*
1666 * swap_pager_isswapped:
1667 *
1668 * Return 1 if at least one page in the given object is paged
1669 * out to the given swap device.
1670 *
1671 * This routine may not sleep.
1672 */
1673 int
1674 swap_pager_isswapped(vm_object_t object, struct swdevt *sp)
1675 {
1676 daddr_t index = 0;
1677 int bcount;
1678 int i;
1679
1680 VM_OBJECT_ASSERT_WLOCKED(object);
1681 if (object->type != OBJT_SWAP)
1682 return (0);
1683
1684 mtx_lock(&swhash_mtx);
1685 for (bcount = 0; bcount < object->un_pager.swp.swp_bcount; bcount++) {
1686 struct swblock *swap;
1687
1688 if ((swap = *swp_pager_hash(object, index)) != NULL) {
1689 for (i = 0; i < SWAP_META_PAGES; ++i) {
1690 if (swp_pager_isondev(swap->swb_pages[i], sp)) {
1691 mtx_unlock(&swhash_mtx);
1692 return (1);
1693 }
1694 }
1695 }
1696 index += SWAP_META_PAGES;
1697 }
1698 mtx_unlock(&swhash_mtx);
1699 return (0);
1700 }
1701
1702 /*
1703 * SWP_PAGER_FORCE_PAGEIN() - force a swap block to be paged in
1704 *
1705 * This routine dissociates the page at the given index within a
1706 * swap block from its backing store, paging it in if necessary.
1707 * If the page is paged in, it is placed in the inactive queue,
1708 * since it had its backing store ripped out from under it.
1709 * We also attempt to swap in all other pages in the swap block,
1710 * we only guarantee that the one at the specified index is
1711 * paged in.
1712 *
1713 * XXX - The code to page the whole block in doesn't work, so we
1714 * revert to the one-by-one behavior for now. Sigh.
1715 */
1716 static inline void
1717 swp_pager_force_pagein(vm_object_t object, vm_pindex_t pindex)
1718 {
1719 vm_page_t m;
1720
1721 vm_object_pip_add(object, 1);
1722 m = vm_page_grab(object, pindex, VM_ALLOC_NORMAL);
1723 if (m->valid == VM_PAGE_BITS_ALL) {
1724 vm_object_pip_subtract(object, 1);
1725 vm_page_dirty(m);
1726 vm_page_lock(m);
1727 vm_page_activate(m);
1728 vm_page_unlock(m);
1729 vm_page_xunbusy(m);
1730 vm_pager_page_unswapped(m);
1731 return;
1732 }
1733
1734 if (swap_pager_getpages(object, &m, 1, 0) != VM_PAGER_OK)
1735 panic("swap_pager_force_pagein: read from swap failed");/*XXX*/
1736 vm_object_pip_subtract(object, 1);
1737 vm_page_dirty(m);
1738 vm_page_lock(m);
1739 vm_page_deactivate(m);
1740 vm_page_unlock(m);
1741 vm_page_xunbusy(m);
1742 vm_pager_page_unswapped(m);
1743 }
1744
1745 /*
1746 * swap_pager_swapoff:
1747 *
1748 * Page in all of the pages that have been paged out to the
1749 * given device. The corresponding blocks in the bitmap must be
1750 * marked as allocated and the device must be flagged SW_CLOSING.
1751 * There may be no processes swapped out to the device.
1752 *
1753 * This routine may block.
1754 */
1755 static void
1756 swap_pager_swapoff(struct swdevt *sp)
1757 {
1758 struct swblock *swap;
1759 int i, j, retries;
1760
1761 GIANT_REQUIRED;
1762
1763 retries = 0;
1764 full_rescan:
1765 mtx_lock(&swhash_mtx);
1766 for (i = 0; i <= swhash_mask; i++) { /* '<=' is correct here */
1767 restart:
1768 for (swap = swhash[i]; swap != NULL; swap = swap->swb_hnext) {
1769 vm_object_t object = swap->swb_object;
1770 vm_pindex_t pindex = swap->swb_index;
1771 for (j = 0; j < SWAP_META_PAGES; ++j) {
1772 if (swp_pager_isondev(swap->swb_pages[j], sp)) {
1773 /* avoid deadlock */
1774 if (!VM_OBJECT_TRYWLOCK(object)) {
1775 break;
1776 } else {
1777 mtx_unlock(&swhash_mtx);
1778 swp_pager_force_pagein(object,
1779 pindex + j);
1780 VM_OBJECT_WUNLOCK(object);
1781 mtx_lock(&swhash_mtx);
1782 goto restart;
1783 }
1784 }
1785 }
1786 }
1787 }
1788 mtx_unlock(&swhash_mtx);
1789 if (sp->sw_used) {
1790 /*
1791 * Objects may be locked or paging to the device being
1792 * removed, so we will miss their pages and need to
1793 * make another pass. We have marked this device as
1794 * SW_CLOSING, so the activity should finish soon.
1795 */
1796 retries++;
1797 if (retries > 100) {
1798 panic("swapoff: failed to locate %d swap blocks",
1799 sp->sw_used);
1800 }
1801 pause("swpoff", hz / 20);
1802 goto full_rescan;
1803 }
1804 }
1805
1806 /************************************************************************
1807 * SWAP META DATA *
1808 ************************************************************************
1809 *
1810 * These routines manipulate the swap metadata stored in the
1811 * OBJT_SWAP object.
1812 *
1813 * Swap metadata is implemented with a global hash and not directly
1814 * linked into the object. Instead the object simply contains
1815 * appropriate tracking counters.
1816 */
1817
1818 /*
1819 * SWP_PAGER_META_BUILD() - add swap block to swap meta data for object
1820 *
1821 * We first convert the object to a swap object if it is a default
1822 * object.
1823 *
1824 * The specified swapblk is added to the object's swap metadata. If
1825 * the swapblk is not valid, it is freed instead. Any previously
1826 * assigned swapblk is freed.
1827 */
1828 static void
1829 swp_pager_meta_build(vm_object_t object, vm_pindex_t pindex, daddr_t swapblk)
1830 {
1831 static volatile int exhausted;
1832 struct swblock *swap;
1833 struct swblock **pswap;
1834 int idx;
1835
1836 VM_OBJECT_ASSERT_WLOCKED(object);
1837 /*
1838 * Convert default object to swap object if necessary
1839 */
1840 if (object->type != OBJT_SWAP) {
1841 object->type = OBJT_SWAP;
1842 object->un_pager.swp.swp_bcount = 0;
1843
1844 if (object->handle != NULL) {
1845 mtx_lock(&sw_alloc_mtx);
1846 TAILQ_INSERT_TAIL(
1847 NOBJLIST(object->handle),
1848 object,
1849 pager_object_list
1850 );
1851 mtx_unlock(&sw_alloc_mtx);
1852 }
1853 }
1854
1855 /*
1856 * Locate hash entry. If not found create, but if we aren't adding
1857 * anything just return. If we run out of space in the map we wait
1858 * and, since the hash table may have changed, retry.
1859 */
1860 retry:
1861 mtx_lock(&swhash_mtx);
1862 pswap = swp_pager_hash(object, pindex);
1863
1864 if ((swap = *pswap) == NULL) {
1865 int i;
1866
1867 if (swapblk == SWAPBLK_NONE)
1868 goto done;
1869
1870 swap = *pswap = uma_zalloc(swap_zone, M_NOWAIT |
1871 (curproc == pageproc ? M_USE_RESERVE : 0));
1872 if (swap == NULL) {
1873 mtx_unlock(&swhash_mtx);
1874 VM_OBJECT_WUNLOCK(object);
1875 if (uma_zone_exhausted(swap_zone)) {
1876 if (atomic_cmpset_int(&exhausted, 0, 1))
1877 printf("swap zone exhausted, "
1878 "increase kern.maxswzone\n");
1879 vm_pageout_oom(VM_OOM_SWAPZ);
1880 pause("swzonex", 10);
1881 } else
1882 VM_WAIT;
1883 VM_OBJECT_WLOCK(object);
1884 goto retry;
1885 }
1886
1887 if (atomic_cmpset_int(&exhausted, 1, 0))
1888 printf("swap zone ok\n");
1889
1890 swap->swb_hnext = NULL;
1891 swap->swb_object = object;
1892 swap->swb_index = pindex & ~(vm_pindex_t)SWAP_META_MASK;
1893 swap->swb_count = 0;
1894
1895 ++object->un_pager.swp.swp_bcount;
1896
1897 for (i = 0; i < SWAP_META_PAGES; ++i)
1898 swap->swb_pages[i] = SWAPBLK_NONE;
1899 }
1900
1901 /*
1902 * Delete prior contents of metadata
1903 */
1904 idx = pindex & SWAP_META_MASK;
1905
1906 if (swap->swb_pages[idx] != SWAPBLK_NONE) {
1907 swp_pager_freeswapspace(swap->swb_pages[idx], 1);
1908 --swap->swb_count;
1909 }
1910
1911 /*
1912 * Enter block into metadata
1913 */
1914 swap->swb_pages[idx] = swapblk;
1915 if (swapblk != SWAPBLK_NONE)
1916 ++swap->swb_count;
1917 done:
1918 mtx_unlock(&swhash_mtx);
1919 }
1920
1921 /*
1922 * SWP_PAGER_META_FREE() - free a range of blocks in the object's swap metadata
1923 *
1924 * The requested range of blocks is freed, with any associated swap
1925 * returned to the swap bitmap.
1926 *
1927 * This routine will free swap metadata structures as they are cleaned
1928 * out. This routine does *NOT* operate on swap metadata associated
1929 * with resident pages.
1930 */
1931 static void
1932 swp_pager_meta_free(vm_object_t object, vm_pindex_t index, daddr_t count)
1933 {
1934
1935 VM_OBJECT_ASSERT_LOCKED(object);
1936 if (object->type != OBJT_SWAP)
1937 return;
1938
1939 while (count > 0) {
1940 struct swblock **pswap;
1941 struct swblock *swap;
1942
1943 mtx_lock(&swhash_mtx);
1944 pswap = swp_pager_hash(object, index);
1945
1946 if ((swap = *pswap) != NULL) {
1947 daddr_t v = swap->swb_pages[index & SWAP_META_MASK];
1948
1949 if (v != SWAPBLK_NONE) {
1950 swp_pager_freeswapspace(v, 1);
1951 swap->swb_pages[index & SWAP_META_MASK] =
1952 SWAPBLK_NONE;
1953 if (--swap->swb_count == 0) {
1954 *pswap = swap->swb_hnext;
1955 uma_zfree(swap_zone, swap);
1956 --object->un_pager.swp.swp_bcount;
1957 }
1958 }
1959 --count;
1960 ++index;
1961 } else {
1962 int n = SWAP_META_PAGES - (index & SWAP_META_MASK);
1963 count -= n;
1964 index += n;
1965 }
1966 mtx_unlock(&swhash_mtx);
1967 }
1968 }
1969
1970 /*
1971 * SWP_PAGER_META_FREE_ALL() - destroy all swap metadata associated with object
1972 *
1973 * This routine locates and destroys all swap metadata associated with
1974 * an object.
1975 */
1976 static void
1977 swp_pager_meta_free_all(vm_object_t object)
1978 {
1979 daddr_t index = 0;
1980
1981 VM_OBJECT_ASSERT_WLOCKED(object);
1982 if (object->type != OBJT_SWAP)
1983 return;
1984
1985 while (object->un_pager.swp.swp_bcount) {
1986 struct swblock **pswap;
1987 struct swblock *swap;
1988
1989 mtx_lock(&swhash_mtx);
1990 pswap = swp_pager_hash(object, index);
1991 if ((swap = *pswap) != NULL) {
1992 int i;
1993
1994 for (i = 0; i < SWAP_META_PAGES; ++i) {
1995 daddr_t v = swap->swb_pages[i];
1996 if (v != SWAPBLK_NONE) {
1997 --swap->swb_count;
1998 swp_pager_freeswapspace(v, 1);
1999 }
2000 }
2001 if (swap->swb_count != 0)
2002 panic("swap_pager_meta_free_all: swb_count != 0");
2003 *pswap = swap->swb_hnext;
2004 uma_zfree(swap_zone, swap);
2005 --object->un_pager.swp.swp_bcount;
2006 }
2007 mtx_unlock(&swhash_mtx);
2008 index += SWAP_META_PAGES;
2009 }
2010 }
2011
2012 /*
2013 * SWP_PAGER_METACTL() - misc control of swap and vm_page_t meta data.
2014 *
2015 * This routine is capable of looking up, popping, or freeing
2016 * swapblk assignments in the swap meta data or in the vm_page_t.
2017 * The routine typically returns the swapblk being looked-up, or popped,
2018 * or SWAPBLK_NONE if the block was freed, or SWAPBLK_NONE if the block
2019 * was invalid. This routine will automatically free any invalid
2020 * meta-data swapblks.
2021 *
2022 * It is not possible to store invalid swapblks in the swap meta data
2023 * (other then a literal 'SWAPBLK_NONE'), so we don't bother checking.
2024 *
2025 * When acting on a busy resident page and paging is in progress, we
2026 * have to wait until paging is complete but otherwise can act on the
2027 * busy page.
2028 *
2029 * SWM_FREE remove and free swap block from metadata
2030 * SWM_POP remove from meta data but do not free.. pop it out
2031 */
2032 static daddr_t
2033 swp_pager_meta_ctl(vm_object_t object, vm_pindex_t pindex, int flags)
2034 {
2035 struct swblock **pswap;
2036 struct swblock *swap;
2037 daddr_t r1;
2038 int idx;
2039
2040 VM_OBJECT_ASSERT_LOCKED(object);
2041 /*
2042 * The meta data only exists of the object is OBJT_SWAP
2043 * and even then might not be allocated yet.
2044 */
2045 if (object->type != OBJT_SWAP)
2046 return (SWAPBLK_NONE);
2047
2048 r1 = SWAPBLK_NONE;
2049 mtx_lock(&swhash_mtx);
2050 pswap = swp_pager_hash(object, pindex);
2051
2052 if ((swap = *pswap) != NULL) {
2053 idx = pindex & SWAP_META_MASK;
2054 r1 = swap->swb_pages[idx];
2055
2056 if (r1 != SWAPBLK_NONE) {
2057 if (flags & SWM_FREE) {
2058 swp_pager_freeswapspace(r1, 1);
2059 r1 = SWAPBLK_NONE;
2060 }
2061 if (flags & (SWM_FREE|SWM_POP)) {
2062 swap->swb_pages[idx] = SWAPBLK_NONE;
2063 if (--swap->swb_count == 0) {
2064 *pswap = swap->swb_hnext;
2065 uma_zfree(swap_zone, swap);
2066 --object->un_pager.swp.swp_bcount;
2067 }
2068 }
2069 }
2070 }
2071 mtx_unlock(&swhash_mtx);
2072 return (r1);
2073 }
2074
2075 /*
2076 * System call swapon(name) enables swapping on device name,
2077 * which must be in the swdevsw. Return EBUSY
2078 * if already swapping on this device.
2079 */
2080 #ifndef _SYS_SYSPROTO_H_
2081 struct swapon_args {
2082 char *name;
2083 };
2084 #endif
2085
2086 /*
2087 * MPSAFE
2088 */
2089 /* ARGSUSED */
2090 int
2091 sys_swapon(struct thread *td, struct swapon_args *uap)
2092 {
2093 struct vattr attr;
2094 struct vnode *vp;
2095 struct nameidata nd;
2096 int error;
2097
2098 error = priv_check(td, PRIV_SWAPON);
2099 if (error)
2100 return (error);
2101
2102 mtx_lock(&Giant);
2103 while (swdev_syscall_active)
2104 tsleep(&swdev_syscall_active, PUSER - 1, "swpon", 0);
2105 swdev_syscall_active = 1;
2106
2107 /*
2108 * Swap metadata may not fit in the KVM if we have physical
2109 * memory of >1GB.
2110 */
2111 if (swap_zone == NULL) {
2112 error = ENOMEM;
2113 goto done;
2114 }
2115
2116 NDINIT(&nd, LOOKUP, ISOPEN | FOLLOW | AUDITVNODE1, UIO_USERSPACE,
2117 uap->name, td);
2118 error = namei(&nd);
2119 if (error)
2120 goto done;
2121
2122 NDFREE(&nd, NDF_ONLY_PNBUF);
2123 vp = nd.ni_vp;
2124
2125 if (vn_isdisk(vp, &error)) {
2126 error = swapongeom(td, vp);
2127 } else if (vp->v_type == VREG &&
2128 (vp->v_mount->mnt_vfc->vfc_flags & VFCF_NETWORK) != 0 &&
2129 (error = VOP_GETATTR(vp, &attr, td->td_ucred)) == 0) {
2130 /*
2131 * Allow direct swapping to NFS regular files in the same
2132 * way that nfs_mountroot() sets up diskless swapping.
2133 */
2134 error = swaponvp(td, vp, attr.va_size / DEV_BSIZE);
2135 }
2136
2137 if (error)
2138 vrele(vp);
2139 done:
2140 swdev_syscall_active = 0;
2141 wakeup_one(&swdev_syscall_active);
2142 mtx_unlock(&Giant);
2143 return (error);
2144 }
2145
2146 /*
2147 * Check that the total amount of swap currently configured does not
2148 * exceed half the theoretical maximum. If it does, print a warning
2149 * message and return -1; otherwise, return 0.
2150 */
2151 static int
2152 swapon_check_swzone(unsigned long npages)
2153 {
2154 unsigned long maxpages;
2155
2156 /* absolute maximum we can handle assuming 100% efficiency */
2157 maxpages = uma_zone_get_max(swap_zone) * SWAP_META_PAGES;
2158
2159 /* recommend using no more than half that amount */
2160 if (npages > maxpages / 2) {
2161 printf("warning: total configured swap (%lu pages) "
2162 "exceeds maximum recommended amount (%lu pages).\n",
2163 npages, maxpages / 2);
2164 printf("warning: increase kern.maxswzone "
2165 "or reduce amount of swap.\n");
2166 return (-1);
2167 }
2168 return (0);
2169 }
2170
2171 static void
2172 swaponsomething(struct vnode *vp, void *id, u_long nblks,
2173 sw_strategy_t *strategy, sw_close_t *close, dev_t dev, int flags)
2174 {
2175 struct swdevt *sp, *tsp;
2176 swblk_t dvbase;
2177 u_long mblocks;
2178
2179 /*
2180 * nblks is in DEV_BSIZE'd chunks, convert to PAGE_SIZE'd chunks.
2181 * First chop nblks off to page-align it, then convert.
2182 *
2183 * sw->sw_nblks is in page-sized chunks now too.
2184 */
2185 nblks &= ~(ctodb(1) - 1);
2186 nblks = dbtoc(nblks);
2187
2188 /*
2189 * If we go beyond this, we get overflows in the radix
2190 * tree bitmap code.
2191 */
2192 mblocks = 0x40000000 / BLIST_META_RADIX;
2193 if (nblks > mblocks) {
2194 printf(
2195 "WARNING: reducing swap size to maximum of %luMB per unit\n",
2196 mblocks / 1024 / 1024 * PAGE_SIZE);
2197 nblks = mblocks;
2198 }
2199
2200 sp = malloc(sizeof *sp, M_VMPGDATA, M_WAITOK | M_ZERO);
2201 sp->sw_vp = vp;
2202 sp->sw_id = id;
2203 sp->sw_dev = dev;
2204 sp->sw_flags = 0;
2205 sp->sw_nblks = nblks;
2206 sp->sw_used = 0;
2207 sp->sw_strategy = strategy;
2208 sp->sw_close = close;
2209 sp->sw_flags = flags;
2210
2211 sp->sw_blist = blist_create(nblks, M_WAITOK);
2212 /*
2213 * Do not free the first two block in order to avoid overwriting
2214 * any bsd label at the front of the partition
2215 */
2216 blist_free(sp->sw_blist, 2, nblks - 2);
2217
2218 dvbase = 0;
2219 mtx_lock(&sw_dev_mtx);
2220 TAILQ_FOREACH(tsp, &swtailq, sw_list) {
2221 if (tsp->sw_end >= dvbase) {
2222 /*
2223 * We put one uncovered page between the devices
2224 * in order to definitively prevent any cross-device
2225 * I/O requests
2226 */
2227 dvbase = tsp->sw_end + 1;
2228 }
2229 }
2230 sp->sw_first = dvbase;
2231 sp->sw_end = dvbase + nblks;
2232 TAILQ_INSERT_TAIL(&swtailq, sp, sw_list);
2233 nswapdev++;
2234 swap_pager_avail += nblks;
2235 swap_total += (vm_ooffset_t)nblks * PAGE_SIZE;
2236 swapon_check_swzone(swap_total / PAGE_SIZE);
2237 swp_sizecheck();
2238 mtx_unlock(&sw_dev_mtx);
2239 }
2240
2241 /*
2242 * SYSCALL: swapoff(devname)
2243 *
2244 * Disable swapping on the given device.
2245 *
2246 * XXX: Badly designed system call: it should use a device index
2247 * rather than filename as specification. We keep sw_vp around
2248 * only to make this work.
2249 */
2250 #ifndef _SYS_SYSPROTO_H_
2251 struct swapoff_args {
2252 char *name;
2253 };
2254 #endif
2255
2256 /*
2257 * MPSAFE
2258 */
2259 /* ARGSUSED */
2260 int
2261 sys_swapoff(struct thread *td, struct swapoff_args *uap)
2262 {
2263 struct vnode *vp;
2264 struct nameidata nd;
2265 struct swdevt *sp;
2266 int error;
2267
2268 error = priv_check(td, PRIV_SWAPOFF);
2269 if (error)
2270 return (error);
2271
2272 mtx_lock(&Giant);
2273 while (swdev_syscall_active)
2274 tsleep(&swdev_syscall_active, PUSER - 1, "swpoff", 0);
2275 swdev_syscall_active = 1;
2276
2277 NDINIT(&nd, LOOKUP, FOLLOW | AUDITVNODE1, UIO_USERSPACE, uap->name,
2278 td);
2279 error = namei(&nd);
2280 if (error)
2281 goto done;
2282 NDFREE(&nd, NDF_ONLY_PNBUF);
2283 vp = nd.ni_vp;
2284
2285 mtx_lock(&sw_dev_mtx);
2286 TAILQ_FOREACH(sp, &swtailq, sw_list) {
2287 if (sp->sw_vp == vp)
2288 break;
2289 }
2290 mtx_unlock(&sw_dev_mtx);
2291 if (sp == NULL) {
2292 error = EINVAL;
2293 goto done;
2294 }
2295 error = swapoff_one(sp, td->td_ucred);
2296 done:
2297 swdev_syscall_active = 0;
2298 wakeup_one(&swdev_syscall_active);
2299 mtx_unlock(&Giant);
2300 return (error);
2301 }
2302
2303 static int
2304 swapoff_one(struct swdevt *sp, struct ucred *cred)
2305 {
2306 u_long nblks, dvbase;
2307 #ifdef MAC
2308 int error;
2309 #endif
2310
2311 mtx_assert(&Giant, MA_OWNED);
2312 #ifdef MAC
2313 (void) vn_lock(sp->sw_vp, LK_EXCLUSIVE | LK_RETRY);
2314 error = mac_system_check_swapoff(cred, sp->sw_vp);
2315 (void) VOP_UNLOCK(sp->sw_vp, 0);
2316 if (error != 0)
2317 return (error);
2318 #endif
2319 nblks = sp->sw_nblks;
2320
2321 /*
2322 * We can turn off this swap device safely only if the
2323 * available virtual memory in the system will fit the amount
2324 * of data we will have to page back in, plus an epsilon so
2325 * the system doesn't become critically low on swap space.
2326 */
2327 if (cnt.v_free_count + cnt.v_cache_count + swap_pager_avail <
2328 nblks + nswap_lowat) {
2329 return (ENOMEM);
2330 }
2331
2332 /*
2333 * Prevent further allocations on this device.
2334 */
2335 mtx_lock(&sw_dev_mtx);
2336 sp->sw_flags |= SW_CLOSING;
2337 for (dvbase = 0; dvbase < sp->sw_end; dvbase += dmmax) {
2338 swap_pager_avail -= blist_fill(sp->sw_blist,
2339 dvbase, dmmax);
2340 }
2341 swap_total -= (vm_ooffset_t)nblks * PAGE_SIZE;
2342 mtx_unlock(&sw_dev_mtx);
2343
2344 /*
2345 * Page in the contents of the device and close it.
2346 */
2347 swap_pager_swapoff(sp);
2348
2349 sp->sw_close(curthread, sp);
2350 sp->sw_id = NULL;
2351 mtx_lock(&sw_dev_mtx);
2352 TAILQ_REMOVE(&swtailq, sp, sw_list);
2353 nswapdev--;
2354 if (nswapdev == 0) {
2355 swap_pager_full = 2;
2356 swap_pager_almost_full = 1;
2357 }
2358 if (swdevhd == sp)
2359 swdevhd = NULL;
2360 mtx_unlock(&sw_dev_mtx);
2361 blist_destroy(sp->sw_blist);
2362 free(sp, M_VMPGDATA);
2363 return (0);
2364 }
2365
2366 void
2367 swapoff_all(void)
2368 {
2369 struct swdevt *sp, *spt;
2370 const char *devname;
2371 int error;
2372
2373 mtx_lock(&Giant);
2374 while (swdev_syscall_active)
2375 tsleep(&swdev_syscall_active, PUSER - 1, "swpoff", 0);
2376 swdev_syscall_active = 1;
2377
2378 mtx_lock(&sw_dev_mtx);
2379 TAILQ_FOREACH_SAFE(sp, &swtailq, sw_list, spt) {
2380 mtx_unlock(&sw_dev_mtx);
2381 if (vn_isdisk(sp->sw_vp, NULL))
2382 devname = devtoname(sp->sw_vp->v_rdev);
2383 else
2384 devname = "[file]";
2385 error = swapoff_one(sp, thread0.td_ucred);
2386 if (error != 0) {
2387 printf("Cannot remove swap device %s (error=%d), "
2388 "skipping.\n", devname, error);
2389 } else if (bootverbose) {
2390 printf("Swap device %s removed.\n", devname);
2391 }
2392 mtx_lock(&sw_dev_mtx);
2393 }
2394 mtx_unlock(&sw_dev_mtx);
2395
2396 swdev_syscall_active = 0;
2397 wakeup_one(&swdev_syscall_active);
2398 mtx_unlock(&Giant);
2399 }
2400
2401 void
2402 swap_pager_status(int *total, int *used)
2403 {
2404 struct swdevt *sp;
2405
2406 *total = 0;
2407 *used = 0;
2408 mtx_lock(&sw_dev_mtx);
2409 TAILQ_FOREACH(sp, &swtailq, sw_list) {
2410 *total += sp->sw_nblks;
2411 *used += sp->sw_used;
2412 }
2413 mtx_unlock(&sw_dev_mtx);
2414 }
2415
2416 int
2417 swap_dev_info(int name, struct xswdev *xs, char *devname, size_t len)
2418 {
2419 struct swdevt *sp;
2420 const char *tmp_devname;
2421 int error, n;
2422
2423 n = 0;
2424 error = ENOENT;
2425 mtx_lock(&sw_dev_mtx);
2426 TAILQ_FOREACH(sp, &swtailq, sw_list) {
2427 if (n != name) {
2428 n++;
2429 continue;
2430 }
2431 xs->xsw_version = XSWDEV_VERSION;
2432 xs->xsw_dev = sp->sw_dev;
2433 xs->xsw_flags = sp->sw_flags;
2434 xs->xsw_nblks = sp->sw_nblks;
2435 xs->xsw_used = sp->sw_used;
2436 if (devname != NULL) {
2437 if (vn_isdisk(sp->sw_vp, NULL))
2438 tmp_devname = devtoname(sp->sw_vp->v_rdev);
2439 else
2440 tmp_devname = "[file]";
2441 strncpy(devname, tmp_devname, len);
2442 }
2443 error = 0;
2444 break;
2445 }
2446 mtx_unlock(&sw_dev_mtx);
2447 return (error);
2448 }
2449
2450 static int
2451 sysctl_vm_swap_info(SYSCTL_HANDLER_ARGS)
2452 {
2453 struct xswdev xs;
2454 int error;
2455
2456 if (arg2 != 1) /* name length */
2457 return (EINVAL);
2458 error = swap_dev_info(*(int *)arg1, &xs, NULL, 0);
2459 if (error != 0)
2460 return (error);
2461 error = SYSCTL_OUT(req, &xs, sizeof(xs));
2462 return (error);
2463 }
2464
2465 SYSCTL_INT(_vm, OID_AUTO, nswapdev, CTLFLAG_RD, &nswapdev, 0,
2466 "Number of swap devices");
2467 SYSCTL_NODE(_vm, OID_AUTO, swap_info, CTLFLAG_RD, sysctl_vm_swap_info,
2468 "Swap statistics by device");
2469
2470 /*
2471 * vmspace_swap_count() - count the approximate swap usage in pages for a
2472 * vmspace.
2473 *
2474 * The map must be locked.
2475 *
2476 * Swap usage is determined by taking the proportional swap used by
2477 * VM objects backing the VM map. To make up for fractional losses,
2478 * if the VM object has any swap use at all the associated map entries
2479 * count for at least 1 swap page.
2480 */
2481 long
2482 vmspace_swap_count(struct vmspace *vmspace)
2483 {
2484 vm_map_t map;
2485 vm_map_entry_t cur;
2486 vm_object_t object;
2487 long count, n;
2488
2489 map = &vmspace->vm_map;
2490 count = 0;
2491
2492 for (cur = map->header.next; cur != &map->header; cur = cur->next) {
2493 if ((cur->eflags & MAP_ENTRY_IS_SUB_MAP) == 0 &&
2494 (object = cur->object.vm_object) != NULL) {
2495 VM_OBJECT_WLOCK(object);
2496 if (object->type == OBJT_SWAP &&
2497 object->un_pager.swp.swp_bcount != 0) {
2498 n = (cur->end - cur->start) / PAGE_SIZE;
2499 count += object->un_pager.swp.swp_bcount *
2500 SWAP_META_PAGES * n / object->size + 1;
2501 }
2502 VM_OBJECT_WUNLOCK(object);
2503 }
2504 }
2505 return (count);
2506 }
2507
2508 /*
2509 * GEOM backend
2510 *
2511 * Swapping onto disk devices.
2512 *
2513 */
2514
2515 static g_orphan_t swapgeom_orphan;
2516
2517 static struct g_class g_swap_class = {
2518 .name = "SWAP",
2519 .version = G_VERSION,
2520 .orphan = swapgeom_orphan,
2521 };
2522
2523 DECLARE_GEOM_CLASS(g_swap_class, g_class);
2524
2525
2526 static void
2527 swapgeom_done(struct bio *bp2)
2528 {
2529 struct buf *bp;
2530
2531 bp = bp2->bio_caller2;
2532 bp->b_ioflags = bp2->bio_flags;
2533 if (bp2->bio_error)
2534 bp->b_ioflags |= BIO_ERROR;
2535 bp->b_resid = bp->b_bcount - bp2->bio_completed;
2536 bp->b_error = bp2->bio_error;
2537 bufdone(bp);
2538 g_destroy_bio(bp2);
2539 }
2540
2541 static void
2542 swapgeom_strategy(struct buf *bp, struct swdevt *sp)
2543 {
2544 struct bio *bio;
2545 struct g_consumer *cp;
2546
2547 cp = sp->sw_id;
2548 if (cp == NULL) {
2549 bp->b_error = ENXIO;
2550 bp->b_ioflags |= BIO_ERROR;
2551 bufdone(bp);
2552 return;
2553 }
2554 if (bp->b_iocmd == BIO_WRITE)
2555 bio = g_new_bio();
2556 else
2557 bio = g_alloc_bio();
2558 if (bio == NULL) {
2559 bp->b_error = ENOMEM;
2560 bp->b_ioflags |= BIO_ERROR;
2561 bufdone(bp);
2562 return;
2563 }
2564
2565 bio->bio_caller2 = bp;
2566 bio->bio_cmd = bp->b_iocmd;
2567 bio->bio_offset = (bp->b_blkno - sp->sw_first) * PAGE_SIZE;
2568 bio->bio_length = bp->b_bcount;
2569 bio->bio_done = swapgeom_done;
2570 if ((bp->b_flags & B_UNMAPPED) != 0) {
2571 bio->bio_ma = bp->b_pages;
2572 bio->bio_data = unmapped_buf;
2573 bio->bio_ma_offset = (vm_offset_t)bp->b_offset & PAGE_MASK;
2574 bio->bio_ma_n = bp->b_npages;
2575 bio->bio_flags |= BIO_UNMAPPED;
2576 } else {
2577 bio->bio_data = bp->b_data;
2578 bio->bio_ma = NULL;
2579 }
2580 g_io_request(bio, cp);
2581 return;
2582 }
2583
2584 static void
2585 swapgeom_orphan(struct g_consumer *cp)
2586 {
2587 struct swdevt *sp;
2588
2589 mtx_lock(&sw_dev_mtx);
2590 TAILQ_FOREACH(sp, &swtailq, sw_list)
2591 if (sp->sw_id == cp)
2592 sp->sw_flags |= SW_CLOSING;
2593 mtx_unlock(&sw_dev_mtx);
2594 }
2595
2596 static void
2597 swapgeom_close_ev(void *arg, int flags)
2598 {
2599 struct g_consumer *cp;
2600
2601 cp = arg;
2602 g_access(cp, -1, -1, 0);
2603 g_detach(cp);
2604 g_destroy_consumer(cp);
2605 }
2606
2607 static void
2608 swapgeom_close(struct thread *td, struct swdevt *sw)
2609 {
2610
2611 /* XXX: direct call when Giant untangled */
2612 g_waitfor_event(swapgeom_close_ev, sw->sw_id, M_WAITOK, NULL);
2613 }
2614
2615
2616 struct swh0h0 {
2617 struct cdev *dev;
2618 struct vnode *vp;
2619 int error;
2620 };
2621
2622 static void
2623 swapongeom_ev(void *arg, int flags)
2624 {
2625 struct swh0h0 *swh;
2626 struct g_provider *pp;
2627 struct g_consumer *cp;
2628 static struct g_geom *gp;
2629 struct swdevt *sp;
2630 u_long nblks;
2631 int error;
2632
2633 swh = arg;
2634 swh->error = 0;
2635 pp = g_dev_getprovider(swh->dev);
2636 if (pp == NULL) {
2637 swh->error = ENODEV;
2638 return;
2639 }
2640 mtx_lock(&sw_dev_mtx);
2641 TAILQ_FOREACH(sp, &swtailq, sw_list) {
2642 cp = sp->sw_id;
2643 if (cp != NULL && cp->provider == pp) {
2644 mtx_unlock(&sw_dev_mtx);
2645 swh->error = EBUSY;
2646 return;
2647 }
2648 }
2649 mtx_unlock(&sw_dev_mtx);
2650 if (gp == NULL)
2651 gp = g_new_geomf(&g_swap_class, "swap");
2652 cp = g_new_consumer(gp);
2653 g_attach(cp, pp);
2654 /*
2655 * XXX: Everytime you think you can improve the margin for
2656 * footshooting, somebody depends on the ability to do so:
2657 * savecore(8) wants to write to our swapdev so we cannot
2658 * set an exclusive count :-(
2659 */
2660 error = g_access(cp, 1, 1, 0);
2661 if (error) {
2662 g_detach(cp);
2663 g_destroy_consumer(cp);
2664 swh->error = error;
2665 return;
2666 }
2667 nblks = pp->mediasize / DEV_BSIZE;
2668 swaponsomething(swh->vp, cp, nblks, swapgeom_strategy,
2669 swapgeom_close, dev2udev(swh->dev),
2670 (pp->flags & G_PF_ACCEPT_UNMAPPED) != 0 ? SW_UNMAPPED : 0);
2671 swh->error = 0;
2672 }
2673
2674 static int
2675 swapongeom(struct thread *td, struct vnode *vp)
2676 {
2677 int error;
2678 struct swh0h0 swh;
2679
2680 vn_lock(vp, LK_EXCLUSIVE | LK_RETRY);
2681
2682 swh.dev = vp->v_rdev;
2683 swh.vp = vp;
2684 swh.error = 0;
2685 /* XXX: direct call when Giant untangled */
2686 error = g_waitfor_event(swapongeom_ev, &swh, M_WAITOK, NULL);
2687 if (!error)
2688 error = swh.error;
2689 VOP_UNLOCK(vp, 0);
2690 return (error);
2691 }
2692
2693 /*
2694 * VNODE backend
2695 *
2696 * This is used mainly for network filesystem (read: probably only tested
2697 * with NFS) swapfiles.
2698 *
2699 */
2700
2701 static void
2702 swapdev_strategy(struct buf *bp, struct swdevt *sp)
2703 {
2704 struct vnode *vp2;
2705
2706 bp->b_blkno = ctodb(bp->b_blkno - sp->sw_first);
2707
2708 vp2 = sp->sw_id;
2709 vhold(vp2);
2710 if (bp->b_iocmd == BIO_WRITE) {
2711 if (bp->b_bufobj)
2712 bufobj_wdrop(bp->b_bufobj);
2713 bufobj_wref(&vp2->v_bufobj);
2714 }
2715 if (bp->b_bufobj != &vp2->v_bufobj)
2716 bp->b_bufobj = &vp2->v_bufobj;
2717 bp->b_vp = vp2;
2718 bp->b_iooffset = dbtob(bp->b_blkno);
2719 bstrategy(bp);
2720 return;
2721 }
2722
2723 static void
2724 swapdev_close(struct thread *td, struct swdevt *sp)
2725 {
2726
2727 VOP_CLOSE(sp->sw_vp, FREAD | FWRITE, td->td_ucred, td);
2728 vrele(sp->sw_vp);
2729 }
2730
2731
2732 static int
2733 swaponvp(struct thread *td, struct vnode *vp, u_long nblks)
2734 {
2735 struct swdevt *sp;
2736 int error;
2737
2738 if (nblks == 0)
2739 return (ENXIO);
2740 mtx_lock(&sw_dev_mtx);
2741 TAILQ_FOREACH(sp, &swtailq, sw_list) {
2742 if (sp->sw_id == vp) {
2743 mtx_unlock(&sw_dev_mtx);
2744 return (EBUSY);
2745 }
2746 }
2747 mtx_unlock(&sw_dev_mtx);
2748
2749 (void) vn_lock(vp, LK_EXCLUSIVE | LK_RETRY);
2750 #ifdef MAC
2751 error = mac_system_check_swapon(td->td_ucred, vp);
2752 if (error == 0)
2753 #endif
2754 error = VOP_OPEN(vp, FREAD | FWRITE, td->td_ucred, td, NULL);
2755 (void) VOP_UNLOCK(vp, 0);
2756 if (error)
2757 return (error);
2758
2759 swaponsomething(vp, vp, nblks, swapdev_strategy, swapdev_close,
2760 NODEV, 0);
2761 return (0);
2762 }
Cache object: 77c2082bfcc355f688c295cf074180ac
|