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/10.2/sys/vm/swap_pager.c 284665 2015-06-21 06:28:26Z trasz $");
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 if (racct_enable) {
202 PROC_LOCK(curproc);
203 error = racct_add(curproc, RACCT_SWAP, incr);
204 PROC_UNLOCK(curproc);
205 if (error != 0)
206 return (0);
207 }
208 #endif
209
210 res = 0;
211 mtx_lock(&sw_dev_mtx);
212 r = swap_reserved + incr;
213 if (overcommit & SWAP_RESERVE_ALLOW_NONWIRED) {
214 s = cnt.v_page_count - cnt.v_free_reserved - cnt.v_wire_count;
215 s *= PAGE_SIZE;
216 } else
217 s = 0;
218 s += swap_total;
219 if ((overcommit & SWAP_RESERVE_FORCE_ON) == 0 || r <= s ||
220 (error = priv_check(curthread, PRIV_VM_SWAP_NOQUOTA)) == 0) {
221 res = 1;
222 swap_reserved = r;
223 }
224 mtx_unlock(&sw_dev_mtx);
225
226 if (res) {
227 PROC_LOCK(curproc);
228 UIDINFO_VMSIZE_LOCK(uip);
229 if ((overcommit & SWAP_RESERVE_RLIMIT_ON) != 0 &&
230 uip->ui_vmsize + incr > lim_cur(curproc, RLIMIT_SWAP) &&
231 priv_check(curthread, PRIV_VM_SWAP_NORLIMIT))
232 res = 0;
233 else
234 uip->ui_vmsize += incr;
235 UIDINFO_VMSIZE_UNLOCK(uip);
236 PROC_UNLOCK(curproc);
237 if (!res) {
238 mtx_lock(&sw_dev_mtx);
239 swap_reserved -= incr;
240 mtx_unlock(&sw_dev_mtx);
241 }
242 }
243 if (!res && ppsratecheck(&lastfail, &curfail, 1)) {
244 printf("uid %d, pid %d: swap reservation for %jd bytes failed\n",
245 uip->ui_uid, curproc->p_pid, incr);
246 }
247
248 #ifdef RACCT
249 if (!res) {
250 PROC_LOCK(curproc);
251 racct_sub(curproc, RACCT_SWAP, incr);
252 PROC_UNLOCK(curproc);
253 }
254 #endif
255
256 return (res);
257 }
258
259 void
260 swap_reserve_force(vm_ooffset_t incr)
261 {
262 struct uidinfo *uip;
263
264 mtx_lock(&sw_dev_mtx);
265 swap_reserved += incr;
266 mtx_unlock(&sw_dev_mtx);
267
268 #ifdef RACCT
269 PROC_LOCK(curproc);
270 racct_add_force(curproc, RACCT_SWAP, incr);
271 PROC_UNLOCK(curproc);
272 #endif
273
274 uip = curthread->td_ucred->cr_ruidinfo;
275 PROC_LOCK(curproc);
276 UIDINFO_VMSIZE_LOCK(uip);
277 uip->ui_vmsize += incr;
278 UIDINFO_VMSIZE_UNLOCK(uip);
279 PROC_UNLOCK(curproc);
280 }
281
282 void
283 swap_release(vm_ooffset_t decr)
284 {
285 struct ucred *cred;
286
287 PROC_LOCK(curproc);
288 cred = curthread->td_ucred;
289 swap_release_by_cred(decr, cred);
290 PROC_UNLOCK(curproc);
291 }
292
293 void
294 swap_release_by_cred(vm_ooffset_t decr, struct ucred *cred)
295 {
296 struct uidinfo *uip;
297
298 uip = cred->cr_ruidinfo;
299
300 if (decr & PAGE_MASK)
301 panic("swap_release: & PAGE_MASK");
302
303 mtx_lock(&sw_dev_mtx);
304 if (swap_reserved < decr)
305 panic("swap_reserved < decr");
306 swap_reserved -= decr;
307 mtx_unlock(&sw_dev_mtx);
308
309 UIDINFO_VMSIZE_LOCK(uip);
310 if (uip->ui_vmsize < decr)
311 printf("negative vmsize for uid = %d\n", uip->ui_uid);
312 uip->ui_vmsize -= decr;
313 UIDINFO_VMSIZE_UNLOCK(uip);
314
315 racct_sub_cred(cred, RACCT_SWAP, decr);
316 }
317
318 static void swapdev_strategy(struct buf *, struct swdevt *sw);
319
320 #define SWM_FREE 0x02 /* free, period */
321 #define SWM_POP 0x04 /* pop out */
322
323 int swap_pager_full = 2; /* swap space exhaustion (task killing) */
324 static int swap_pager_almost_full = 1; /* swap space exhaustion (w/hysteresis)*/
325 static int nsw_rcount; /* free read buffers */
326 static int nsw_wcount_sync; /* limit write buffers / synchronous */
327 static int nsw_wcount_async; /* limit write buffers / asynchronous */
328 static int nsw_wcount_async_max;/* assigned maximum */
329 static int nsw_cluster_max; /* maximum VOP I/O allowed */
330
331 static struct swblock **swhash;
332 static int swhash_mask;
333 static struct mtx swhash_mtx;
334
335 static int swap_async_max = 4; /* maximum in-progress async I/O's */
336 static struct sx sw_alloc_sx;
337
338
339 SYSCTL_INT(_vm, OID_AUTO, swap_async_max,
340 CTLFLAG_RW, &swap_async_max, 0, "Maximum running async swap ops");
341
342 /*
343 * "named" and "unnamed" anon region objects. Try to reduce the overhead
344 * of searching a named list by hashing it just a little.
345 */
346
347 #define NOBJLISTS 8
348
349 #define NOBJLIST(handle) \
350 (&swap_pager_object_list[((int)(intptr_t)handle >> 4) & (NOBJLISTS-1)])
351
352 static struct mtx sw_alloc_mtx; /* protect list manipulation */
353 static struct pagerlst swap_pager_object_list[NOBJLISTS];
354 static uma_zone_t swap_zone;
355
356 /*
357 * pagerops for OBJT_SWAP - "swap pager". Some ops are also global procedure
358 * calls hooked from other parts of the VM system and do not appear here.
359 * (see vm/swap_pager.h).
360 */
361 static vm_object_t
362 swap_pager_alloc(void *handle, vm_ooffset_t size,
363 vm_prot_t prot, vm_ooffset_t offset, struct ucred *);
364 static void swap_pager_dealloc(vm_object_t object);
365 static int swap_pager_getpages(vm_object_t, vm_page_t *, int, int);
366 static void swap_pager_putpages(vm_object_t, vm_page_t *, int, boolean_t, int *);
367 static boolean_t
368 swap_pager_haspage(vm_object_t object, vm_pindex_t pindex, int *before, int *after);
369 static void swap_pager_init(void);
370 static void swap_pager_unswapped(vm_page_t);
371 static void swap_pager_swapoff(struct swdevt *sp);
372
373 struct pagerops swappagerops = {
374 .pgo_init = swap_pager_init, /* early system initialization of pager */
375 .pgo_alloc = swap_pager_alloc, /* allocate an OBJT_SWAP object */
376 .pgo_dealloc = swap_pager_dealloc, /* deallocate an OBJT_SWAP object */
377 .pgo_getpages = swap_pager_getpages, /* pagein */
378 .pgo_putpages = swap_pager_putpages, /* pageout */
379 .pgo_haspage = swap_pager_haspage, /* get backing store status for page */
380 .pgo_pageunswapped = swap_pager_unswapped, /* remove swap related to page */
381 };
382
383 /*
384 * dmmax is in page-sized chunks with the new swap system. It was
385 * dev-bsized chunks in the old. dmmax is always a power of 2.
386 *
387 * swap_*() routines are externally accessible. swp_*() routines are
388 * internal.
389 */
390 static int dmmax;
391 static int nswap_lowat = 128; /* in pages, swap_pager_almost_full warn */
392 static int nswap_hiwat = 512; /* in pages, swap_pager_almost_full warn */
393
394 SYSCTL_INT(_vm, OID_AUTO, dmmax,
395 CTLFLAG_RD, &dmmax, 0, "Maximum size of a swap block");
396
397 static void swp_sizecheck(void);
398 static void swp_pager_async_iodone(struct buf *bp);
399 static int swapongeom(struct thread *, struct vnode *);
400 static int swaponvp(struct thread *, struct vnode *, u_long);
401 static int swapoff_one(struct swdevt *sp, struct ucred *cred);
402
403 /*
404 * Swap bitmap functions
405 */
406 static void swp_pager_freeswapspace(daddr_t blk, int npages);
407 static daddr_t swp_pager_getswapspace(int npages);
408
409 /*
410 * Metadata functions
411 */
412 static struct swblock **swp_pager_hash(vm_object_t object, vm_pindex_t index);
413 static void swp_pager_meta_build(vm_object_t, vm_pindex_t, daddr_t);
414 static void swp_pager_meta_free(vm_object_t, vm_pindex_t, daddr_t);
415 static void swp_pager_meta_free_all(vm_object_t);
416 static daddr_t swp_pager_meta_ctl(vm_object_t, vm_pindex_t, int);
417
418 static void
419 swp_pager_free_nrpage(vm_page_t m)
420 {
421
422 vm_page_lock(m);
423 if (m->wire_count == 0)
424 vm_page_free(m);
425 vm_page_unlock(m);
426 }
427
428 /*
429 * SWP_SIZECHECK() - update swap_pager_full indication
430 *
431 * update the swap_pager_almost_full indication and warn when we are
432 * about to run out of swap space, using lowat/hiwat hysteresis.
433 *
434 * Clear swap_pager_full ( task killing ) indication when lowat is met.
435 *
436 * No restrictions on call
437 * This routine may not block.
438 */
439 static void
440 swp_sizecheck(void)
441 {
442
443 if (swap_pager_avail < nswap_lowat) {
444 if (swap_pager_almost_full == 0) {
445 printf("swap_pager: out of swap space\n");
446 swap_pager_almost_full = 1;
447 }
448 } else {
449 swap_pager_full = 0;
450 if (swap_pager_avail > nswap_hiwat)
451 swap_pager_almost_full = 0;
452 }
453 }
454
455 /*
456 * SWP_PAGER_HASH() - hash swap meta data
457 *
458 * This is an helper function which hashes the swapblk given
459 * the object and page index. It returns a pointer to a pointer
460 * to the object, or a pointer to a NULL pointer if it could not
461 * find a swapblk.
462 */
463 static struct swblock **
464 swp_pager_hash(vm_object_t object, vm_pindex_t index)
465 {
466 struct swblock **pswap;
467 struct swblock *swap;
468
469 index &= ~(vm_pindex_t)SWAP_META_MASK;
470 pswap = &swhash[(index ^ (int)(intptr_t)object) & swhash_mask];
471 while ((swap = *pswap) != NULL) {
472 if (swap->swb_object == object &&
473 swap->swb_index == index
474 ) {
475 break;
476 }
477 pswap = &swap->swb_hnext;
478 }
479 return (pswap);
480 }
481
482 /*
483 * SWAP_PAGER_INIT() - initialize the swap pager!
484 *
485 * Expected to be started from system init. NOTE: This code is run
486 * before much else so be careful what you depend on. Most of the VM
487 * system has yet to be initialized at this point.
488 */
489 static void
490 swap_pager_init(void)
491 {
492 /*
493 * Initialize object lists
494 */
495 int i;
496
497 for (i = 0; i < NOBJLISTS; ++i)
498 TAILQ_INIT(&swap_pager_object_list[i]);
499 mtx_init(&sw_alloc_mtx, "swap_pager list", NULL, MTX_DEF);
500 mtx_init(&sw_dev_mtx, "swapdev", NULL, MTX_DEF);
501
502 /*
503 * Device Stripe, in PAGE_SIZE'd blocks
504 */
505 dmmax = SWB_NPAGES * 2;
506 }
507
508 /*
509 * SWAP_PAGER_SWAP_INIT() - swap pager initialization from pageout process
510 *
511 * Expected to be started from pageout process once, prior to entering
512 * its main loop.
513 */
514 void
515 swap_pager_swap_init(void)
516 {
517 unsigned long n, n2;
518
519 /*
520 * Number of in-transit swap bp operations. Don't
521 * exhaust the pbufs completely. Make sure we
522 * initialize workable values (0 will work for hysteresis
523 * but it isn't very efficient).
524 *
525 * The nsw_cluster_max is constrained by the bp->b_pages[]
526 * array (MAXPHYS/PAGE_SIZE) and our locally defined
527 * MAX_PAGEOUT_CLUSTER. Also be aware that swap ops are
528 * constrained by the swap device interleave stripe size.
529 *
530 * Currently we hardwire nsw_wcount_async to 4. This limit is
531 * designed to prevent other I/O from having high latencies due to
532 * our pageout I/O. The value 4 works well for one or two active swap
533 * devices but is probably a little low if you have more. Even so,
534 * a higher value would probably generate only a limited improvement
535 * with three or four active swap devices since the system does not
536 * typically have to pageout at extreme bandwidths. We will want
537 * at least 2 per swap devices, and 4 is a pretty good value if you
538 * have one NFS swap device due to the command/ack latency over NFS.
539 * So it all works out pretty well.
540 */
541 nsw_cluster_max = min((MAXPHYS/PAGE_SIZE), MAX_PAGEOUT_CLUSTER);
542
543 mtx_lock(&pbuf_mtx);
544 nsw_rcount = (nswbuf + 1) / 2;
545 nsw_wcount_sync = (nswbuf + 3) / 4;
546 nsw_wcount_async = 4;
547 nsw_wcount_async_max = nsw_wcount_async;
548 mtx_unlock(&pbuf_mtx);
549
550 /*
551 * Initialize our zone. Right now I'm just guessing on the number
552 * we need based on the number of pages in the system. Each swblock
553 * can hold 32 pages, so this is probably overkill. This reservation
554 * is typically limited to around 32MB by default.
555 */
556 n = cnt.v_page_count / 2;
557 if (maxswzone && n > maxswzone / sizeof(struct swblock))
558 n = maxswzone / sizeof(struct swblock);
559 n2 = n;
560 swap_zone = uma_zcreate("SWAPMETA", sizeof(struct swblock), NULL, NULL,
561 NULL, NULL, UMA_ALIGN_PTR, UMA_ZONE_NOFREE | UMA_ZONE_VM);
562 if (swap_zone == NULL)
563 panic("failed to create swap_zone.");
564 do {
565 if (uma_zone_reserve_kva(swap_zone, n))
566 break;
567 /*
568 * if the allocation failed, try a zone two thirds the
569 * size of the previous attempt.
570 */
571 n -= ((n + 2) / 3);
572 } while (n > 0);
573 if (n2 != n)
574 printf("Swap zone entries reduced from %lu to %lu.\n", n2, n);
575 swap_maxpages = n * SWAP_META_PAGES;
576 swzone = n * sizeof(struct swblock);
577 n2 = n;
578
579 /*
580 * Initialize our meta-data hash table. The swapper does not need to
581 * be quite as efficient as the VM system, so we do not use an
582 * oversized hash table.
583 *
584 * n: size of hash table, must be power of 2
585 * swhash_mask: hash table index mask
586 */
587 for (n = 1; n < n2 / 8; n *= 2)
588 ;
589 swhash = malloc(sizeof(struct swblock *) * n, M_VMPGDATA, M_WAITOK | M_ZERO);
590 swhash_mask = n - 1;
591 mtx_init(&swhash_mtx, "swap_pager swhash", NULL, MTX_DEF);
592 }
593
594 /*
595 * SWAP_PAGER_ALLOC() - allocate a new OBJT_SWAP VM object and instantiate
596 * its metadata structures.
597 *
598 * This routine is called from the mmap and fork code to create a new
599 * OBJT_SWAP object. We do this by creating an OBJT_DEFAULT object
600 * and then converting it with swp_pager_meta_build().
601 *
602 * This routine may block in vm_object_allocate() and create a named
603 * object lookup race, so we must interlock.
604 *
605 * MPSAFE
606 */
607 static vm_object_t
608 swap_pager_alloc(void *handle, vm_ooffset_t size, vm_prot_t prot,
609 vm_ooffset_t offset, struct ucred *cred)
610 {
611 vm_object_t object;
612 vm_pindex_t pindex;
613
614 pindex = OFF_TO_IDX(offset + PAGE_MASK + size);
615 if (handle) {
616 mtx_lock(&Giant);
617 /*
618 * Reference existing named region or allocate new one. There
619 * should not be a race here against swp_pager_meta_build()
620 * as called from vm_page_remove() in regards to the lookup
621 * of the handle.
622 */
623 sx_xlock(&sw_alloc_sx);
624 object = vm_pager_object_lookup(NOBJLIST(handle), handle);
625 if (object == NULL) {
626 if (cred != NULL) {
627 if (!swap_reserve_by_cred(size, cred)) {
628 sx_xunlock(&sw_alloc_sx);
629 mtx_unlock(&Giant);
630 return (NULL);
631 }
632 crhold(cred);
633 }
634 object = vm_object_allocate(OBJT_DEFAULT, pindex);
635 VM_OBJECT_WLOCK(object);
636 object->handle = handle;
637 if (cred != NULL) {
638 object->cred = cred;
639 object->charge = size;
640 }
641 swp_pager_meta_build(object, 0, SWAPBLK_NONE);
642 VM_OBJECT_WUNLOCK(object);
643 }
644 sx_xunlock(&sw_alloc_sx);
645 mtx_unlock(&Giant);
646 } else {
647 if (cred != NULL) {
648 if (!swap_reserve_by_cred(size, cred))
649 return (NULL);
650 crhold(cred);
651 }
652 object = vm_object_allocate(OBJT_DEFAULT, pindex);
653 VM_OBJECT_WLOCK(object);
654 if (cred != NULL) {
655 object->cred = cred;
656 object->charge = size;
657 }
658 swp_pager_meta_build(object, 0, SWAPBLK_NONE);
659 VM_OBJECT_WUNLOCK(object);
660 }
661 return (object);
662 }
663
664 /*
665 * SWAP_PAGER_DEALLOC() - remove swap metadata from object
666 *
667 * The swap backing for the object is destroyed. The code is
668 * designed such that we can reinstantiate it later, but this
669 * routine is typically called only when the entire object is
670 * about to be destroyed.
671 *
672 * The object must be locked.
673 */
674 static void
675 swap_pager_dealloc(vm_object_t object)
676 {
677
678 /*
679 * Remove from list right away so lookups will fail if we block for
680 * pageout completion.
681 */
682 if (object->handle != NULL) {
683 mtx_lock(&sw_alloc_mtx);
684 TAILQ_REMOVE(NOBJLIST(object->handle), object, pager_object_list);
685 mtx_unlock(&sw_alloc_mtx);
686 }
687
688 VM_OBJECT_ASSERT_WLOCKED(object);
689 vm_object_pip_wait(object, "swpdea");
690
691 /*
692 * Free all remaining metadata. We only bother to free it from
693 * the swap meta data. We do not attempt to free swapblk's still
694 * associated with vm_page_t's for this object. We do not care
695 * if paging is still in progress on some objects.
696 */
697 swp_pager_meta_free_all(object);
698 object->handle = NULL;
699 object->type = OBJT_DEAD;
700 }
701
702 /************************************************************************
703 * SWAP PAGER BITMAP ROUTINES *
704 ************************************************************************/
705
706 /*
707 * SWP_PAGER_GETSWAPSPACE() - allocate raw swap space
708 *
709 * Allocate swap for the requested number of pages. The starting
710 * swap block number (a page index) is returned or SWAPBLK_NONE
711 * if the allocation failed.
712 *
713 * Also has the side effect of advising that somebody made a mistake
714 * when they configured swap and didn't configure enough.
715 *
716 * This routine may not sleep.
717 *
718 * We allocate in round-robin fashion from the configured devices.
719 */
720 static daddr_t
721 swp_pager_getswapspace(int npages)
722 {
723 daddr_t blk;
724 struct swdevt *sp;
725 int i;
726
727 blk = SWAPBLK_NONE;
728 mtx_lock(&sw_dev_mtx);
729 sp = swdevhd;
730 for (i = 0; i < nswapdev; i++) {
731 if (sp == NULL)
732 sp = TAILQ_FIRST(&swtailq);
733 if (!(sp->sw_flags & SW_CLOSING)) {
734 blk = blist_alloc(sp->sw_blist, npages);
735 if (blk != SWAPBLK_NONE) {
736 blk += sp->sw_first;
737 sp->sw_used += npages;
738 swap_pager_avail -= npages;
739 swp_sizecheck();
740 swdevhd = TAILQ_NEXT(sp, sw_list);
741 goto done;
742 }
743 }
744 sp = TAILQ_NEXT(sp, sw_list);
745 }
746 if (swap_pager_full != 2) {
747 printf("swap_pager_getswapspace(%d): failed\n", npages);
748 swap_pager_full = 2;
749 swap_pager_almost_full = 1;
750 }
751 swdevhd = NULL;
752 done:
753 mtx_unlock(&sw_dev_mtx);
754 return (blk);
755 }
756
757 static int
758 swp_pager_isondev(daddr_t blk, struct swdevt *sp)
759 {
760
761 return (blk >= sp->sw_first && blk < sp->sw_end);
762 }
763
764 static void
765 swp_pager_strategy(struct buf *bp)
766 {
767 struct swdevt *sp;
768
769 mtx_lock(&sw_dev_mtx);
770 TAILQ_FOREACH(sp, &swtailq, sw_list) {
771 if (bp->b_blkno >= sp->sw_first && bp->b_blkno < sp->sw_end) {
772 mtx_unlock(&sw_dev_mtx);
773 if ((sp->sw_flags & SW_UNMAPPED) != 0 &&
774 unmapped_buf_allowed) {
775 bp->b_kvaalloc = bp->b_data;
776 bp->b_data = unmapped_buf;
777 bp->b_kvabase = unmapped_buf;
778 bp->b_offset = 0;
779 bp->b_flags |= B_UNMAPPED;
780 } else {
781 pmap_qenter((vm_offset_t)bp->b_data,
782 &bp->b_pages[0], bp->b_bcount / PAGE_SIZE);
783 }
784 sp->sw_strategy(bp, sp);
785 return;
786 }
787 }
788 panic("Swapdev not found");
789 }
790
791
792 /*
793 * SWP_PAGER_FREESWAPSPACE() - free raw swap space
794 *
795 * This routine returns the specified swap blocks back to the bitmap.
796 *
797 * This routine may not sleep.
798 */
799 static void
800 swp_pager_freeswapspace(daddr_t blk, int npages)
801 {
802 struct swdevt *sp;
803
804 mtx_lock(&sw_dev_mtx);
805 TAILQ_FOREACH(sp, &swtailq, sw_list) {
806 if (blk >= sp->sw_first && blk < sp->sw_end) {
807 sp->sw_used -= npages;
808 /*
809 * If we are attempting to stop swapping on
810 * this device, we don't want to mark any
811 * blocks free lest they be reused.
812 */
813 if ((sp->sw_flags & SW_CLOSING) == 0) {
814 blist_free(sp->sw_blist, blk - sp->sw_first,
815 npages);
816 swap_pager_avail += npages;
817 swp_sizecheck();
818 }
819 mtx_unlock(&sw_dev_mtx);
820 return;
821 }
822 }
823 panic("Swapdev not found");
824 }
825
826 /*
827 * SWAP_PAGER_FREESPACE() - frees swap blocks associated with a page
828 * range within an object.
829 *
830 * This is a globally accessible routine.
831 *
832 * This routine removes swapblk assignments from swap metadata.
833 *
834 * The external callers of this routine typically have already destroyed
835 * or renamed vm_page_t's associated with this range in the object so
836 * we should be ok.
837 *
838 * The object must be locked.
839 */
840 void
841 swap_pager_freespace(vm_object_t object, vm_pindex_t start, vm_size_t size)
842 {
843
844 swp_pager_meta_free(object, start, size);
845 }
846
847 /*
848 * SWAP_PAGER_RESERVE() - reserve swap blocks in object
849 *
850 * Assigns swap blocks to the specified range within the object. The
851 * swap blocks are not zeroed. Any previous swap assignment is destroyed.
852 *
853 * Returns 0 on success, -1 on failure.
854 */
855 int
856 swap_pager_reserve(vm_object_t object, vm_pindex_t start, vm_size_t size)
857 {
858 int n = 0;
859 daddr_t blk = SWAPBLK_NONE;
860 vm_pindex_t beg = start; /* save start index */
861
862 VM_OBJECT_WLOCK(object);
863 while (size) {
864 if (n == 0) {
865 n = BLIST_MAX_ALLOC;
866 while ((blk = swp_pager_getswapspace(n)) == SWAPBLK_NONE) {
867 n >>= 1;
868 if (n == 0) {
869 swp_pager_meta_free(object, beg, start - beg);
870 VM_OBJECT_WUNLOCK(object);
871 return (-1);
872 }
873 }
874 }
875 swp_pager_meta_build(object, start, blk);
876 --size;
877 ++start;
878 ++blk;
879 --n;
880 }
881 swp_pager_meta_free(object, start, n);
882 VM_OBJECT_WUNLOCK(object);
883 return (0);
884 }
885
886 /*
887 * SWAP_PAGER_COPY() - copy blocks from source pager to destination pager
888 * and destroy the source.
889 *
890 * Copy any valid swapblks from the source to the destination. In
891 * cases where both the source and destination have a valid swapblk,
892 * we keep the destination's.
893 *
894 * This routine is allowed to sleep. It may sleep allocating metadata
895 * indirectly through swp_pager_meta_build() or if paging is still in
896 * progress on the source.
897 *
898 * The source object contains no vm_page_t's (which is just as well)
899 *
900 * The source object is of type OBJT_SWAP.
901 *
902 * The source and destination objects must be locked.
903 * Both object locks may temporarily be released.
904 */
905 void
906 swap_pager_copy(vm_object_t srcobject, vm_object_t dstobject,
907 vm_pindex_t offset, int destroysource)
908 {
909 vm_pindex_t i;
910
911 VM_OBJECT_ASSERT_WLOCKED(srcobject);
912 VM_OBJECT_ASSERT_WLOCKED(dstobject);
913
914 /*
915 * If destroysource is set, we remove the source object from the
916 * swap_pager internal queue now.
917 */
918 if (destroysource) {
919 if (srcobject->handle != NULL) {
920 mtx_lock(&sw_alloc_mtx);
921 TAILQ_REMOVE(
922 NOBJLIST(srcobject->handle),
923 srcobject,
924 pager_object_list
925 );
926 mtx_unlock(&sw_alloc_mtx);
927 }
928 }
929
930 /*
931 * transfer source to destination.
932 */
933 for (i = 0; i < dstobject->size; ++i) {
934 daddr_t dstaddr;
935
936 /*
937 * Locate (without changing) the swapblk on the destination,
938 * unless it is invalid in which case free it silently, or
939 * if the destination is a resident page, in which case the
940 * source is thrown away.
941 */
942 dstaddr = swp_pager_meta_ctl(dstobject, i, 0);
943
944 if (dstaddr == SWAPBLK_NONE) {
945 /*
946 * Destination has no swapblk and is not resident,
947 * copy source.
948 */
949 daddr_t srcaddr;
950
951 srcaddr = swp_pager_meta_ctl(
952 srcobject,
953 i + offset,
954 SWM_POP
955 );
956
957 if (srcaddr != SWAPBLK_NONE) {
958 /*
959 * swp_pager_meta_build() can sleep.
960 */
961 vm_object_pip_add(srcobject, 1);
962 VM_OBJECT_WUNLOCK(srcobject);
963 vm_object_pip_add(dstobject, 1);
964 swp_pager_meta_build(dstobject, i, srcaddr);
965 vm_object_pip_wakeup(dstobject);
966 VM_OBJECT_WLOCK(srcobject);
967 vm_object_pip_wakeup(srcobject);
968 }
969 } else {
970 /*
971 * Destination has valid swapblk or it is represented
972 * by a resident page. We destroy the sourceblock.
973 */
974
975 swp_pager_meta_ctl(srcobject, i + offset, SWM_FREE);
976 }
977 }
978
979 /*
980 * Free left over swap blocks in source.
981 *
982 * We have to revert the type to OBJT_DEFAULT so we do not accidently
983 * double-remove the object from the swap queues.
984 */
985 if (destroysource) {
986 swp_pager_meta_free_all(srcobject);
987 /*
988 * Reverting the type is not necessary, the caller is going
989 * to destroy srcobject directly, but I'm doing it here
990 * for consistency since we've removed the object from its
991 * queues.
992 */
993 srcobject->type = OBJT_DEFAULT;
994 }
995 }
996
997 /*
998 * SWAP_PAGER_HASPAGE() - determine if we have good backing store for
999 * the requested page.
1000 *
1001 * We determine whether good backing store exists for the requested
1002 * page and return TRUE if it does, FALSE if it doesn't.
1003 *
1004 * If TRUE, we also try to determine how much valid, contiguous backing
1005 * store exists before and after the requested page within a reasonable
1006 * distance. We do not try to restrict it to the swap device stripe
1007 * (that is handled in getpages/putpages). It probably isn't worth
1008 * doing here.
1009 */
1010 static boolean_t
1011 swap_pager_haspage(vm_object_t object, vm_pindex_t pindex, int *before, int *after)
1012 {
1013 daddr_t blk0;
1014
1015 VM_OBJECT_ASSERT_LOCKED(object);
1016 /*
1017 * do we have good backing store at the requested index ?
1018 */
1019 blk0 = swp_pager_meta_ctl(object, pindex, 0);
1020
1021 if (blk0 == SWAPBLK_NONE) {
1022 if (before)
1023 *before = 0;
1024 if (after)
1025 *after = 0;
1026 return (FALSE);
1027 }
1028
1029 /*
1030 * find backwards-looking contiguous good backing store
1031 */
1032 if (before != NULL) {
1033 int i;
1034
1035 for (i = 1; i < (SWB_NPAGES/2); ++i) {
1036 daddr_t blk;
1037
1038 if (i > pindex)
1039 break;
1040 blk = swp_pager_meta_ctl(object, pindex - i, 0);
1041 if (blk != blk0 - i)
1042 break;
1043 }
1044 *before = (i - 1);
1045 }
1046
1047 /*
1048 * find forward-looking contiguous good backing store
1049 */
1050 if (after != NULL) {
1051 int i;
1052
1053 for (i = 1; i < (SWB_NPAGES/2); ++i) {
1054 daddr_t blk;
1055
1056 blk = swp_pager_meta_ctl(object, pindex + i, 0);
1057 if (blk != blk0 + i)
1058 break;
1059 }
1060 *after = (i - 1);
1061 }
1062 return (TRUE);
1063 }
1064
1065 /*
1066 * SWAP_PAGER_PAGE_UNSWAPPED() - remove swap backing store related to page
1067 *
1068 * This removes any associated swap backing store, whether valid or
1069 * not, from the page.
1070 *
1071 * This routine is typically called when a page is made dirty, at
1072 * which point any associated swap can be freed. MADV_FREE also
1073 * calls us in a special-case situation
1074 *
1075 * NOTE!!! If the page is clean and the swap was valid, the caller
1076 * should make the page dirty before calling this routine. This routine
1077 * does NOT change the m->dirty status of the page. Also: MADV_FREE
1078 * depends on it.
1079 *
1080 * This routine may not sleep.
1081 *
1082 * The object containing the page must be locked.
1083 */
1084 static void
1085 swap_pager_unswapped(vm_page_t m)
1086 {
1087
1088 swp_pager_meta_ctl(m->object, m->pindex, SWM_FREE);
1089 }
1090
1091 /*
1092 * SWAP_PAGER_GETPAGES() - bring pages in from swap
1093 *
1094 * Attempt to retrieve (m, count) pages from backing store, but make
1095 * sure we retrieve at least m[reqpage]. We try to load in as large
1096 * a chunk surrounding m[reqpage] as is contiguous in swap and which
1097 * belongs to the same object.
1098 *
1099 * The code is designed for asynchronous operation and
1100 * immediate-notification of 'reqpage' but tends not to be
1101 * used that way. Please do not optimize-out this algorithmic
1102 * feature, I intend to improve on it in the future.
1103 *
1104 * The parent has a single vm_object_pip_add() reference prior to
1105 * calling us and we should return with the same.
1106 *
1107 * The parent has BUSY'd the pages. We should return with 'm'
1108 * left busy, but the others adjusted.
1109 */
1110 static int
1111 swap_pager_getpages(vm_object_t object, vm_page_t *m, int count, int reqpage)
1112 {
1113 struct buf *bp;
1114 vm_page_t mreq;
1115 int i;
1116 int j;
1117 daddr_t blk;
1118
1119 mreq = m[reqpage];
1120
1121 KASSERT(mreq->object == object,
1122 ("swap_pager_getpages: object mismatch %p/%p",
1123 object, mreq->object));
1124
1125 /*
1126 * Calculate range to retrieve. The pages have already been assigned
1127 * their swapblks. We require a *contiguous* range but we know it to
1128 * not span devices. If we do not supply it, bad things
1129 * happen. Note that blk, iblk & jblk can be SWAPBLK_NONE, but the
1130 * loops are set up such that the case(s) are handled implicitly.
1131 *
1132 * The swp_*() calls must be made with the object locked.
1133 */
1134 blk = swp_pager_meta_ctl(mreq->object, mreq->pindex, 0);
1135
1136 for (i = reqpage - 1; i >= 0; --i) {
1137 daddr_t iblk;
1138
1139 iblk = swp_pager_meta_ctl(m[i]->object, m[i]->pindex, 0);
1140 if (blk != iblk + (reqpage - i))
1141 break;
1142 }
1143 ++i;
1144
1145 for (j = reqpage + 1; j < count; ++j) {
1146 daddr_t jblk;
1147
1148 jblk = swp_pager_meta_ctl(m[j]->object, m[j]->pindex, 0);
1149 if (blk != jblk - (j - reqpage))
1150 break;
1151 }
1152
1153 /*
1154 * free pages outside our collection range. Note: we never free
1155 * mreq, it must remain busy throughout.
1156 */
1157 if (0 < i || j < count) {
1158 int k;
1159
1160 for (k = 0; k < i; ++k)
1161 swp_pager_free_nrpage(m[k]);
1162 for (k = j; k < count; ++k)
1163 swp_pager_free_nrpage(m[k]);
1164 }
1165
1166 /*
1167 * Return VM_PAGER_FAIL if we have nothing to do. Return mreq
1168 * still busy, but the others unbusied.
1169 */
1170 if (blk == SWAPBLK_NONE)
1171 return (VM_PAGER_FAIL);
1172
1173 /*
1174 * Getpbuf() can sleep.
1175 */
1176 VM_OBJECT_WUNLOCK(object);
1177 /*
1178 * Get a swap buffer header to perform the IO
1179 */
1180 bp = getpbuf(&nsw_rcount);
1181 bp->b_flags |= B_PAGING;
1182
1183 bp->b_iocmd = BIO_READ;
1184 bp->b_iodone = swp_pager_async_iodone;
1185 bp->b_rcred = crhold(thread0.td_ucred);
1186 bp->b_wcred = crhold(thread0.td_ucred);
1187 bp->b_blkno = blk - (reqpage - i);
1188 bp->b_bcount = PAGE_SIZE * (j - i);
1189 bp->b_bufsize = PAGE_SIZE * (j - i);
1190 bp->b_pager.pg_reqpage = reqpage - i;
1191
1192 VM_OBJECT_WLOCK(object);
1193 {
1194 int k;
1195
1196 for (k = i; k < j; ++k) {
1197 bp->b_pages[k - i] = m[k];
1198 m[k]->oflags |= VPO_SWAPINPROG;
1199 }
1200 }
1201 bp->b_npages = j - i;
1202
1203 PCPU_INC(cnt.v_swapin);
1204 PCPU_ADD(cnt.v_swappgsin, bp->b_npages);
1205
1206 /*
1207 * We still hold the lock on mreq, and our automatic completion routine
1208 * does not remove it.
1209 */
1210 vm_object_pip_add(object, bp->b_npages);
1211 VM_OBJECT_WUNLOCK(object);
1212
1213 /*
1214 * perform the I/O. NOTE!!! bp cannot be considered valid after
1215 * this point because we automatically release it on completion.
1216 * Instead, we look at the one page we are interested in which we
1217 * still hold a lock on even through the I/O completion.
1218 *
1219 * The other pages in our m[] array are also released on completion,
1220 * so we cannot assume they are valid anymore either.
1221 *
1222 * NOTE: b_blkno is destroyed by the call to swapdev_strategy
1223 */
1224 BUF_KERNPROC(bp);
1225 swp_pager_strategy(bp);
1226
1227 /*
1228 * wait for the page we want to complete. VPO_SWAPINPROG is always
1229 * cleared on completion. If an I/O error occurs, SWAPBLK_NONE
1230 * is set in the meta-data.
1231 */
1232 VM_OBJECT_WLOCK(object);
1233 while ((mreq->oflags & VPO_SWAPINPROG) != 0) {
1234 mreq->oflags |= VPO_SWAPSLEEP;
1235 PCPU_INC(cnt.v_intrans);
1236 if (VM_OBJECT_SLEEP(object, &object->paging_in_progress, PSWP,
1237 "swread", hz * 20)) {
1238 printf(
1239 "swap_pager: indefinite wait buffer: bufobj: %p, blkno: %jd, size: %ld\n",
1240 bp->b_bufobj, (intmax_t)bp->b_blkno, bp->b_bcount);
1241 }
1242 }
1243
1244 /*
1245 * mreq is left busied after completion, but all the other pages
1246 * are freed. If we had an unrecoverable read error the page will
1247 * not be valid.
1248 */
1249 if (mreq->valid != VM_PAGE_BITS_ALL) {
1250 return (VM_PAGER_ERROR);
1251 } else {
1252 return (VM_PAGER_OK);
1253 }
1254
1255 /*
1256 * A final note: in a low swap situation, we cannot deallocate swap
1257 * and mark a page dirty here because the caller is likely to mark
1258 * the page clean when we return, causing the page to possibly revert
1259 * to all-zero's later.
1260 */
1261 }
1262
1263 /*
1264 * swap_pager_putpages:
1265 *
1266 * Assign swap (if necessary) and initiate I/O on the specified pages.
1267 *
1268 * We support both OBJT_DEFAULT and OBJT_SWAP objects. DEFAULT objects
1269 * are automatically converted to SWAP objects.
1270 *
1271 * In a low memory situation we may block in VOP_STRATEGY(), but the new
1272 * vm_page reservation system coupled with properly written VFS devices
1273 * should ensure that no low-memory deadlock occurs. This is an area
1274 * which needs work.
1275 *
1276 * The parent has N vm_object_pip_add() references prior to
1277 * calling us and will remove references for rtvals[] that are
1278 * not set to VM_PAGER_PEND. We need to remove the rest on I/O
1279 * completion.
1280 *
1281 * The parent has soft-busy'd the pages it passes us and will unbusy
1282 * those whos rtvals[] entry is not set to VM_PAGER_PEND on return.
1283 * We need to unbusy the rest on I/O completion.
1284 */
1285 void
1286 swap_pager_putpages(vm_object_t object, vm_page_t *m, int count,
1287 int flags, int *rtvals)
1288 {
1289 int i, n;
1290 boolean_t sync;
1291
1292 if (count && m[0]->object != object) {
1293 panic("swap_pager_putpages: object mismatch %p/%p",
1294 object,
1295 m[0]->object
1296 );
1297 }
1298
1299 /*
1300 * Step 1
1301 *
1302 * Turn object into OBJT_SWAP
1303 * check for bogus sysops
1304 * force sync if not pageout process
1305 */
1306 if (object->type != OBJT_SWAP)
1307 swp_pager_meta_build(object, 0, SWAPBLK_NONE);
1308 VM_OBJECT_WUNLOCK(object);
1309
1310 n = 0;
1311 if (curproc != pageproc)
1312 sync = TRUE;
1313 else
1314 sync = (flags & VM_PAGER_PUT_SYNC) != 0;
1315
1316 /*
1317 * Step 2
1318 *
1319 * Update nsw parameters from swap_async_max sysctl values.
1320 * Do not let the sysop crash the machine with bogus numbers.
1321 */
1322 mtx_lock(&pbuf_mtx);
1323 if (swap_async_max != nsw_wcount_async_max) {
1324 int n;
1325
1326 /*
1327 * limit range
1328 */
1329 if ((n = swap_async_max) > nswbuf / 2)
1330 n = nswbuf / 2;
1331 if (n < 1)
1332 n = 1;
1333 swap_async_max = n;
1334
1335 /*
1336 * Adjust difference ( if possible ). If the current async
1337 * count is too low, we may not be able to make the adjustment
1338 * at this time.
1339 */
1340 n -= nsw_wcount_async_max;
1341 if (nsw_wcount_async + n >= 0) {
1342 nsw_wcount_async += n;
1343 nsw_wcount_async_max += n;
1344 wakeup(&nsw_wcount_async);
1345 }
1346 }
1347 mtx_unlock(&pbuf_mtx);
1348
1349 /*
1350 * Step 3
1351 *
1352 * Assign swap blocks and issue I/O. We reallocate swap on the fly.
1353 * The page is left dirty until the pageout operation completes
1354 * successfully.
1355 */
1356 for (i = 0; i < count; i += n) {
1357 int j;
1358 struct buf *bp;
1359 daddr_t blk;
1360
1361 /*
1362 * Maximum I/O size is limited by a number of factors.
1363 */
1364 n = min(BLIST_MAX_ALLOC, count - i);
1365 n = min(n, nsw_cluster_max);
1366
1367 /*
1368 * Get biggest block of swap we can. If we fail, fall
1369 * back and try to allocate a smaller block. Don't go
1370 * overboard trying to allocate space if it would overly
1371 * fragment swap.
1372 */
1373 while (
1374 (blk = swp_pager_getswapspace(n)) == SWAPBLK_NONE &&
1375 n > 4
1376 ) {
1377 n >>= 1;
1378 }
1379 if (blk == SWAPBLK_NONE) {
1380 for (j = 0; j < n; ++j)
1381 rtvals[i+j] = VM_PAGER_FAIL;
1382 continue;
1383 }
1384
1385 /*
1386 * All I/O parameters have been satisfied, build the I/O
1387 * request and assign the swap space.
1388 */
1389 if (sync == TRUE) {
1390 bp = getpbuf(&nsw_wcount_sync);
1391 } else {
1392 bp = getpbuf(&nsw_wcount_async);
1393 bp->b_flags = B_ASYNC;
1394 }
1395 bp->b_flags |= B_PAGING;
1396 bp->b_iocmd = BIO_WRITE;
1397
1398 bp->b_rcred = crhold(thread0.td_ucred);
1399 bp->b_wcred = crhold(thread0.td_ucred);
1400 bp->b_bcount = PAGE_SIZE * n;
1401 bp->b_bufsize = PAGE_SIZE * n;
1402 bp->b_blkno = blk;
1403
1404 VM_OBJECT_WLOCK(object);
1405 for (j = 0; j < n; ++j) {
1406 vm_page_t mreq = m[i+j];
1407
1408 swp_pager_meta_build(
1409 mreq->object,
1410 mreq->pindex,
1411 blk + j
1412 );
1413 vm_page_dirty(mreq);
1414 rtvals[i+j] = VM_PAGER_OK;
1415
1416 mreq->oflags |= VPO_SWAPINPROG;
1417 bp->b_pages[j] = mreq;
1418 }
1419 VM_OBJECT_WUNLOCK(object);
1420 bp->b_npages = n;
1421 /*
1422 * Must set dirty range for NFS to work.
1423 */
1424 bp->b_dirtyoff = 0;
1425 bp->b_dirtyend = bp->b_bcount;
1426
1427 PCPU_INC(cnt.v_swapout);
1428 PCPU_ADD(cnt.v_swappgsout, bp->b_npages);
1429
1430 /*
1431 * asynchronous
1432 *
1433 * NOTE: b_blkno is destroyed by the call to swapdev_strategy
1434 */
1435 if (sync == FALSE) {
1436 bp->b_iodone = swp_pager_async_iodone;
1437 BUF_KERNPROC(bp);
1438 swp_pager_strategy(bp);
1439
1440 for (j = 0; j < n; ++j)
1441 rtvals[i+j] = VM_PAGER_PEND;
1442 /* restart outter loop */
1443 continue;
1444 }
1445
1446 /*
1447 * synchronous
1448 *
1449 * NOTE: b_blkno is destroyed by the call to swapdev_strategy
1450 */
1451 bp->b_iodone = bdone;
1452 swp_pager_strategy(bp);
1453
1454 /*
1455 * Wait for the sync I/O to complete, then update rtvals.
1456 * We just set the rtvals[] to VM_PAGER_PEND so we can call
1457 * our async completion routine at the end, thus avoiding a
1458 * double-free.
1459 */
1460 bwait(bp, PVM, "swwrt");
1461 for (j = 0; j < n; ++j)
1462 rtvals[i+j] = VM_PAGER_PEND;
1463 /*
1464 * Now that we are through with the bp, we can call the
1465 * normal async completion, which frees everything up.
1466 */
1467 swp_pager_async_iodone(bp);
1468 }
1469 VM_OBJECT_WLOCK(object);
1470 }
1471
1472 /*
1473 * swp_pager_async_iodone:
1474 *
1475 * Completion routine for asynchronous reads and writes from/to swap.
1476 * Also called manually by synchronous code to finish up a bp.
1477 *
1478 * This routine may not sleep.
1479 */
1480 static void
1481 swp_pager_async_iodone(struct buf *bp)
1482 {
1483 int i;
1484 vm_object_t object = NULL;
1485
1486 /*
1487 * report error
1488 */
1489 if (bp->b_ioflags & BIO_ERROR) {
1490 printf(
1491 "swap_pager: I/O error - %s failed; blkno %ld,"
1492 "size %ld, error %d\n",
1493 ((bp->b_iocmd == BIO_READ) ? "pagein" : "pageout"),
1494 (long)bp->b_blkno,
1495 (long)bp->b_bcount,
1496 bp->b_error
1497 );
1498 }
1499
1500 /*
1501 * remove the mapping for kernel virtual
1502 */
1503 if ((bp->b_flags & B_UNMAPPED) != 0) {
1504 bp->b_data = bp->b_kvaalloc;
1505 bp->b_kvabase = bp->b_kvaalloc;
1506 bp->b_flags &= ~B_UNMAPPED;
1507 } else
1508 pmap_qremove((vm_offset_t)bp->b_data, bp->b_npages);
1509
1510 if (bp->b_npages) {
1511 object = bp->b_pages[0]->object;
1512 VM_OBJECT_WLOCK(object);
1513 }
1514
1515 /*
1516 * cleanup pages. If an error occurs writing to swap, we are in
1517 * very serious trouble. If it happens to be a disk error, though,
1518 * we may be able to recover by reassigning the swap later on. So
1519 * in this case we remove the m->swapblk assignment for the page
1520 * but do not free it in the rlist. The errornous block(s) are thus
1521 * never reallocated as swap. Redirty the page and continue.
1522 */
1523 for (i = 0; i < bp->b_npages; ++i) {
1524 vm_page_t m = bp->b_pages[i];
1525
1526 m->oflags &= ~VPO_SWAPINPROG;
1527 if (m->oflags & VPO_SWAPSLEEP) {
1528 m->oflags &= ~VPO_SWAPSLEEP;
1529 wakeup(&object->paging_in_progress);
1530 }
1531
1532 if (bp->b_ioflags & BIO_ERROR) {
1533 /*
1534 * If an error occurs I'd love to throw the swapblk
1535 * away without freeing it back to swapspace, so it
1536 * can never be used again. But I can't from an
1537 * interrupt.
1538 */
1539 if (bp->b_iocmd == BIO_READ) {
1540 /*
1541 * When reading, reqpage needs to stay
1542 * locked for the parent, but all other
1543 * pages can be freed. We still want to
1544 * wakeup the parent waiting on the page,
1545 * though. ( also: pg_reqpage can be -1 and
1546 * not match anything ).
1547 *
1548 * We have to wake specifically requested pages
1549 * up too because we cleared VPO_SWAPINPROG and
1550 * someone may be waiting for that.
1551 *
1552 * NOTE: for reads, m->dirty will probably
1553 * be overridden by the original caller of
1554 * getpages so don't play cute tricks here.
1555 */
1556 m->valid = 0;
1557 if (i != bp->b_pager.pg_reqpage)
1558 swp_pager_free_nrpage(m);
1559 else {
1560 vm_page_lock(m);
1561 vm_page_flash(m);
1562 vm_page_unlock(m);
1563 }
1564 /*
1565 * If i == bp->b_pager.pg_reqpage, do not wake
1566 * the page up. The caller needs to.
1567 */
1568 } else {
1569 /*
1570 * If a write error occurs, reactivate page
1571 * so it doesn't clog the inactive list,
1572 * then finish the I/O.
1573 */
1574 vm_page_dirty(m);
1575 vm_page_lock(m);
1576 vm_page_activate(m);
1577 vm_page_unlock(m);
1578 vm_page_sunbusy(m);
1579 }
1580 } else if (bp->b_iocmd == BIO_READ) {
1581 /*
1582 * NOTE: for reads, m->dirty will probably be
1583 * overridden by the original caller of getpages so
1584 * we cannot set them in order to free the underlying
1585 * swap in a low-swap situation. I don't think we'd
1586 * want to do that anyway, but it was an optimization
1587 * that existed in the old swapper for a time before
1588 * it got ripped out due to precisely this problem.
1589 *
1590 * If not the requested page then deactivate it.
1591 *
1592 * Note that the requested page, reqpage, is left
1593 * busied, but we still have to wake it up. The
1594 * other pages are released (unbusied) by
1595 * vm_page_xunbusy().
1596 */
1597 KASSERT(!pmap_page_is_mapped(m),
1598 ("swp_pager_async_iodone: page %p is mapped", m));
1599 m->valid = VM_PAGE_BITS_ALL;
1600 KASSERT(m->dirty == 0,
1601 ("swp_pager_async_iodone: page %p is dirty", m));
1602
1603 /*
1604 * We have to wake specifically requested pages
1605 * up too because we cleared VPO_SWAPINPROG and
1606 * could be waiting for it in getpages. However,
1607 * be sure to not unbusy getpages specifically
1608 * requested page - getpages expects it to be
1609 * left busy.
1610 */
1611 if (i != bp->b_pager.pg_reqpage) {
1612 vm_page_lock(m);
1613 vm_page_deactivate(m);
1614 vm_page_unlock(m);
1615 vm_page_xunbusy(m);
1616 } else {
1617 vm_page_lock(m);
1618 vm_page_flash(m);
1619 vm_page_unlock(m);
1620 }
1621 } else {
1622 /*
1623 * For write success, clear the dirty
1624 * status, then finish the I/O ( which decrements the
1625 * busy count and possibly wakes waiter's up ).
1626 */
1627 KASSERT(!pmap_page_is_write_mapped(m),
1628 ("swp_pager_async_iodone: page %p is not write"
1629 " protected", m));
1630 vm_page_undirty(m);
1631 vm_page_sunbusy(m);
1632 if (vm_page_count_severe()) {
1633 vm_page_lock(m);
1634 vm_page_try_to_cache(m);
1635 vm_page_unlock(m);
1636 }
1637 }
1638 }
1639
1640 /*
1641 * adjust pip. NOTE: the original parent may still have its own
1642 * pip refs on the object.
1643 */
1644 if (object != NULL) {
1645 vm_object_pip_wakeupn(object, bp->b_npages);
1646 VM_OBJECT_WUNLOCK(object);
1647 }
1648
1649 /*
1650 * swapdev_strategy() manually sets b_vp and b_bufobj before calling
1651 * bstrategy(). Set them back to NULL now we're done with it, or we'll
1652 * trigger a KASSERT in relpbuf().
1653 */
1654 if (bp->b_vp) {
1655 bp->b_vp = NULL;
1656 bp->b_bufobj = NULL;
1657 }
1658 /*
1659 * release the physical I/O buffer
1660 */
1661 relpbuf(
1662 bp,
1663 ((bp->b_iocmd == BIO_READ) ? &nsw_rcount :
1664 ((bp->b_flags & B_ASYNC) ?
1665 &nsw_wcount_async :
1666 &nsw_wcount_sync
1667 )
1668 )
1669 );
1670 }
1671
1672 /*
1673 * swap_pager_isswapped:
1674 *
1675 * Return 1 if at least one page in the given object is paged
1676 * out to the given swap device.
1677 *
1678 * This routine may not sleep.
1679 */
1680 int
1681 swap_pager_isswapped(vm_object_t object, struct swdevt *sp)
1682 {
1683 daddr_t index = 0;
1684 int bcount;
1685 int i;
1686
1687 VM_OBJECT_ASSERT_WLOCKED(object);
1688 if (object->type != OBJT_SWAP)
1689 return (0);
1690
1691 mtx_lock(&swhash_mtx);
1692 for (bcount = 0; bcount < object->un_pager.swp.swp_bcount; bcount++) {
1693 struct swblock *swap;
1694
1695 if ((swap = *swp_pager_hash(object, index)) != NULL) {
1696 for (i = 0; i < SWAP_META_PAGES; ++i) {
1697 if (swp_pager_isondev(swap->swb_pages[i], sp)) {
1698 mtx_unlock(&swhash_mtx);
1699 return (1);
1700 }
1701 }
1702 }
1703 index += SWAP_META_PAGES;
1704 }
1705 mtx_unlock(&swhash_mtx);
1706 return (0);
1707 }
1708
1709 /*
1710 * SWP_PAGER_FORCE_PAGEIN() - force a swap block to be paged in
1711 *
1712 * This routine dissociates the page at the given index within a
1713 * swap block from its backing store, paging it in if necessary.
1714 * If the page is paged in, it is placed in the inactive queue,
1715 * since it had its backing store ripped out from under it.
1716 * We also attempt to swap in all other pages in the swap block,
1717 * we only guarantee that the one at the specified index is
1718 * paged in.
1719 *
1720 * XXX - The code to page the whole block in doesn't work, so we
1721 * revert to the one-by-one behavior for now. Sigh.
1722 */
1723 static inline void
1724 swp_pager_force_pagein(vm_object_t object, vm_pindex_t pindex)
1725 {
1726 vm_page_t m;
1727
1728 vm_object_pip_add(object, 1);
1729 m = vm_page_grab(object, pindex, VM_ALLOC_NORMAL);
1730 if (m->valid == VM_PAGE_BITS_ALL) {
1731 vm_object_pip_subtract(object, 1);
1732 vm_page_dirty(m);
1733 vm_page_lock(m);
1734 vm_page_activate(m);
1735 vm_page_unlock(m);
1736 vm_page_xunbusy(m);
1737 vm_pager_page_unswapped(m);
1738 return;
1739 }
1740
1741 if (swap_pager_getpages(object, &m, 1, 0) != VM_PAGER_OK)
1742 panic("swap_pager_force_pagein: read from swap failed");/*XXX*/
1743 vm_object_pip_subtract(object, 1);
1744 vm_page_dirty(m);
1745 vm_page_lock(m);
1746 vm_page_deactivate(m);
1747 vm_page_unlock(m);
1748 vm_page_xunbusy(m);
1749 vm_pager_page_unswapped(m);
1750 }
1751
1752 /*
1753 * swap_pager_swapoff:
1754 *
1755 * Page in all of the pages that have been paged out to the
1756 * given device. The corresponding blocks in the bitmap must be
1757 * marked as allocated and the device must be flagged SW_CLOSING.
1758 * There may be no processes swapped out to the device.
1759 *
1760 * This routine may block.
1761 */
1762 static void
1763 swap_pager_swapoff(struct swdevt *sp)
1764 {
1765 struct swblock *swap;
1766 int i, j, retries;
1767
1768 GIANT_REQUIRED;
1769
1770 retries = 0;
1771 full_rescan:
1772 mtx_lock(&swhash_mtx);
1773 for (i = 0; i <= swhash_mask; i++) { /* '<=' is correct here */
1774 restart:
1775 for (swap = swhash[i]; swap != NULL; swap = swap->swb_hnext) {
1776 vm_object_t object = swap->swb_object;
1777 vm_pindex_t pindex = swap->swb_index;
1778 for (j = 0; j < SWAP_META_PAGES; ++j) {
1779 if (swp_pager_isondev(swap->swb_pages[j], sp)) {
1780 /* avoid deadlock */
1781 if (!VM_OBJECT_TRYWLOCK(object)) {
1782 break;
1783 } else {
1784 mtx_unlock(&swhash_mtx);
1785 swp_pager_force_pagein(object,
1786 pindex + j);
1787 VM_OBJECT_WUNLOCK(object);
1788 mtx_lock(&swhash_mtx);
1789 goto restart;
1790 }
1791 }
1792 }
1793 }
1794 }
1795 mtx_unlock(&swhash_mtx);
1796 if (sp->sw_used) {
1797 /*
1798 * Objects may be locked or paging to the device being
1799 * removed, so we will miss their pages and need to
1800 * make another pass. We have marked this device as
1801 * SW_CLOSING, so the activity should finish soon.
1802 */
1803 retries++;
1804 if (retries > 100) {
1805 panic("swapoff: failed to locate %d swap blocks",
1806 sp->sw_used);
1807 }
1808 pause("swpoff", hz / 20);
1809 goto full_rescan;
1810 }
1811 }
1812
1813 /************************************************************************
1814 * SWAP META DATA *
1815 ************************************************************************
1816 *
1817 * These routines manipulate the swap metadata stored in the
1818 * OBJT_SWAP object.
1819 *
1820 * Swap metadata is implemented with a global hash and not directly
1821 * linked into the object. Instead the object simply contains
1822 * appropriate tracking counters.
1823 */
1824
1825 /*
1826 * SWP_PAGER_META_BUILD() - add swap block to swap meta data for object
1827 *
1828 * We first convert the object to a swap object if it is a default
1829 * object.
1830 *
1831 * The specified swapblk is added to the object's swap metadata. If
1832 * the swapblk is not valid, it is freed instead. Any previously
1833 * assigned swapblk is freed.
1834 */
1835 static void
1836 swp_pager_meta_build(vm_object_t object, vm_pindex_t pindex, daddr_t swapblk)
1837 {
1838 static volatile int exhausted;
1839 struct swblock *swap;
1840 struct swblock **pswap;
1841 int idx;
1842
1843 VM_OBJECT_ASSERT_WLOCKED(object);
1844 /*
1845 * Convert default object to swap object if necessary
1846 */
1847 if (object->type != OBJT_SWAP) {
1848 object->type = OBJT_SWAP;
1849 object->un_pager.swp.swp_bcount = 0;
1850
1851 if (object->handle != NULL) {
1852 mtx_lock(&sw_alloc_mtx);
1853 TAILQ_INSERT_TAIL(
1854 NOBJLIST(object->handle),
1855 object,
1856 pager_object_list
1857 );
1858 mtx_unlock(&sw_alloc_mtx);
1859 }
1860 }
1861
1862 /*
1863 * Locate hash entry. If not found create, but if we aren't adding
1864 * anything just return. If we run out of space in the map we wait
1865 * and, since the hash table may have changed, retry.
1866 */
1867 retry:
1868 mtx_lock(&swhash_mtx);
1869 pswap = swp_pager_hash(object, pindex);
1870
1871 if ((swap = *pswap) == NULL) {
1872 int i;
1873
1874 if (swapblk == SWAPBLK_NONE)
1875 goto done;
1876
1877 swap = *pswap = uma_zalloc(swap_zone, M_NOWAIT |
1878 (curproc == pageproc ? M_USE_RESERVE : 0));
1879 if (swap == NULL) {
1880 mtx_unlock(&swhash_mtx);
1881 VM_OBJECT_WUNLOCK(object);
1882 if (uma_zone_exhausted(swap_zone)) {
1883 if (atomic_cmpset_int(&exhausted, 0, 1))
1884 printf("swap zone exhausted, "
1885 "increase kern.maxswzone\n");
1886 vm_pageout_oom(VM_OOM_SWAPZ);
1887 pause("swzonex", 10);
1888 } else
1889 VM_WAIT;
1890 VM_OBJECT_WLOCK(object);
1891 goto retry;
1892 }
1893
1894 if (atomic_cmpset_int(&exhausted, 1, 0))
1895 printf("swap zone ok\n");
1896
1897 swap->swb_hnext = NULL;
1898 swap->swb_object = object;
1899 swap->swb_index = pindex & ~(vm_pindex_t)SWAP_META_MASK;
1900 swap->swb_count = 0;
1901
1902 ++object->un_pager.swp.swp_bcount;
1903
1904 for (i = 0; i < SWAP_META_PAGES; ++i)
1905 swap->swb_pages[i] = SWAPBLK_NONE;
1906 }
1907
1908 /*
1909 * Delete prior contents of metadata
1910 */
1911 idx = pindex & SWAP_META_MASK;
1912
1913 if (swap->swb_pages[idx] != SWAPBLK_NONE) {
1914 swp_pager_freeswapspace(swap->swb_pages[idx], 1);
1915 --swap->swb_count;
1916 }
1917
1918 /*
1919 * Enter block into metadata
1920 */
1921 swap->swb_pages[idx] = swapblk;
1922 if (swapblk != SWAPBLK_NONE)
1923 ++swap->swb_count;
1924 done:
1925 mtx_unlock(&swhash_mtx);
1926 }
1927
1928 /*
1929 * SWP_PAGER_META_FREE() - free a range of blocks in the object's swap metadata
1930 *
1931 * The requested range of blocks is freed, with any associated swap
1932 * returned to the swap bitmap.
1933 *
1934 * This routine will free swap metadata structures as they are cleaned
1935 * out. This routine does *NOT* operate on swap metadata associated
1936 * with resident pages.
1937 */
1938 static void
1939 swp_pager_meta_free(vm_object_t object, vm_pindex_t index, daddr_t count)
1940 {
1941
1942 VM_OBJECT_ASSERT_LOCKED(object);
1943 if (object->type != OBJT_SWAP)
1944 return;
1945
1946 while (count > 0) {
1947 struct swblock **pswap;
1948 struct swblock *swap;
1949
1950 mtx_lock(&swhash_mtx);
1951 pswap = swp_pager_hash(object, index);
1952
1953 if ((swap = *pswap) != NULL) {
1954 daddr_t v = swap->swb_pages[index & SWAP_META_MASK];
1955
1956 if (v != SWAPBLK_NONE) {
1957 swp_pager_freeswapspace(v, 1);
1958 swap->swb_pages[index & SWAP_META_MASK] =
1959 SWAPBLK_NONE;
1960 if (--swap->swb_count == 0) {
1961 *pswap = swap->swb_hnext;
1962 uma_zfree(swap_zone, swap);
1963 --object->un_pager.swp.swp_bcount;
1964 }
1965 }
1966 --count;
1967 ++index;
1968 } else {
1969 int n = SWAP_META_PAGES - (index & SWAP_META_MASK);
1970 count -= n;
1971 index += n;
1972 }
1973 mtx_unlock(&swhash_mtx);
1974 }
1975 }
1976
1977 /*
1978 * SWP_PAGER_META_FREE_ALL() - destroy all swap metadata associated with object
1979 *
1980 * This routine locates and destroys all swap metadata associated with
1981 * an object.
1982 */
1983 static void
1984 swp_pager_meta_free_all(vm_object_t object)
1985 {
1986 daddr_t index = 0;
1987
1988 VM_OBJECT_ASSERT_WLOCKED(object);
1989 if (object->type != OBJT_SWAP)
1990 return;
1991
1992 while (object->un_pager.swp.swp_bcount) {
1993 struct swblock **pswap;
1994 struct swblock *swap;
1995
1996 mtx_lock(&swhash_mtx);
1997 pswap = swp_pager_hash(object, index);
1998 if ((swap = *pswap) != NULL) {
1999 int i;
2000
2001 for (i = 0; i < SWAP_META_PAGES; ++i) {
2002 daddr_t v = swap->swb_pages[i];
2003 if (v != SWAPBLK_NONE) {
2004 --swap->swb_count;
2005 swp_pager_freeswapspace(v, 1);
2006 }
2007 }
2008 if (swap->swb_count != 0)
2009 panic("swap_pager_meta_free_all: swb_count != 0");
2010 *pswap = swap->swb_hnext;
2011 uma_zfree(swap_zone, swap);
2012 --object->un_pager.swp.swp_bcount;
2013 }
2014 mtx_unlock(&swhash_mtx);
2015 index += SWAP_META_PAGES;
2016 }
2017 }
2018
2019 /*
2020 * SWP_PAGER_METACTL() - misc control of swap and vm_page_t meta data.
2021 *
2022 * This routine is capable of looking up, popping, or freeing
2023 * swapblk assignments in the swap meta data or in the vm_page_t.
2024 * The routine typically returns the swapblk being looked-up, or popped,
2025 * or SWAPBLK_NONE if the block was freed, or SWAPBLK_NONE if the block
2026 * was invalid. This routine will automatically free any invalid
2027 * meta-data swapblks.
2028 *
2029 * It is not possible to store invalid swapblks in the swap meta data
2030 * (other then a literal 'SWAPBLK_NONE'), so we don't bother checking.
2031 *
2032 * When acting on a busy resident page and paging is in progress, we
2033 * have to wait until paging is complete but otherwise can act on the
2034 * busy page.
2035 *
2036 * SWM_FREE remove and free swap block from metadata
2037 * SWM_POP remove from meta data but do not free.. pop it out
2038 */
2039 static daddr_t
2040 swp_pager_meta_ctl(vm_object_t object, vm_pindex_t pindex, int flags)
2041 {
2042 struct swblock **pswap;
2043 struct swblock *swap;
2044 daddr_t r1;
2045 int idx;
2046
2047 VM_OBJECT_ASSERT_LOCKED(object);
2048 /*
2049 * The meta data only exists of the object is OBJT_SWAP
2050 * and even then might not be allocated yet.
2051 */
2052 if (object->type != OBJT_SWAP)
2053 return (SWAPBLK_NONE);
2054
2055 r1 = SWAPBLK_NONE;
2056 mtx_lock(&swhash_mtx);
2057 pswap = swp_pager_hash(object, pindex);
2058
2059 if ((swap = *pswap) != NULL) {
2060 idx = pindex & SWAP_META_MASK;
2061 r1 = swap->swb_pages[idx];
2062
2063 if (r1 != SWAPBLK_NONE) {
2064 if (flags & SWM_FREE) {
2065 swp_pager_freeswapspace(r1, 1);
2066 r1 = SWAPBLK_NONE;
2067 }
2068 if (flags & (SWM_FREE|SWM_POP)) {
2069 swap->swb_pages[idx] = SWAPBLK_NONE;
2070 if (--swap->swb_count == 0) {
2071 *pswap = swap->swb_hnext;
2072 uma_zfree(swap_zone, swap);
2073 --object->un_pager.swp.swp_bcount;
2074 }
2075 }
2076 }
2077 }
2078 mtx_unlock(&swhash_mtx);
2079 return (r1);
2080 }
2081
2082 /*
2083 * System call swapon(name) enables swapping on device name,
2084 * which must be in the swdevsw. Return EBUSY
2085 * if already swapping on this device.
2086 */
2087 #ifndef _SYS_SYSPROTO_H_
2088 struct swapon_args {
2089 char *name;
2090 };
2091 #endif
2092
2093 /*
2094 * MPSAFE
2095 */
2096 /* ARGSUSED */
2097 int
2098 sys_swapon(struct thread *td, struct swapon_args *uap)
2099 {
2100 struct vattr attr;
2101 struct vnode *vp;
2102 struct nameidata nd;
2103 int error;
2104
2105 error = priv_check(td, PRIV_SWAPON);
2106 if (error)
2107 return (error);
2108
2109 mtx_lock(&Giant);
2110 while (swdev_syscall_active)
2111 tsleep(&swdev_syscall_active, PUSER - 1, "swpon", 0);
2112 swdev_syscall_active = 1;
2113
2114 /*
2115 * Swap metadata may not fit in the KVM if we have physical
2116 * memory of >1GB.
2117 */
2118 if (swap_zone == NULL) {
2119 error = ENOMEM;
2120 goto done;
2121 }
2122
2123 NDINIT(&nd, LOOKUP, ISOPEN | FOLLOW | AUDITVNODE1, UIO_USERSPACE,
2124 uap->name, td);
2125 error = namei(&nd);
2126 if (error)
2127 goto done;
2128
2129 NDFREE(&nd, NDF_ONLY_PNBUF);
2130 vp = nd.ni_vp;
2131
2132 if (vn_isdisk(vp, &error)) {
2133 error = swapongeom(td, vp);
2134 } else if (vp->v_type == VREG &&
2135 (vp->v_mount->mnt_vfc->vfc_flags & VFCF_NETWORK) != 0 &&
2136 (error = VOP_GETATTR(vp, &attr, td->td_ucred)) == 0) {
2137 /*
2138 * Allow direct swapping to NFS regular files in the same
2139 * way that nfs_mountroot() sets up diskless swapping.
2140 */
2141 error = swaponvp(td, vp, attr.va_size / DEV_BSIZE);
2142 }
2143
2144 if (error)
2145 vrele(vp);
2146 done:
2147 swdev_syscall_active = 0;
2148 wakeup_one(&swdev_syscall_active);
2149 mtx_unlock(&Giant);
2150 return (error);
2151 }
2152
2153 /*
2154 * Check that the total amount of swap currently configured does not
2155 * exceed half the theoretical maximum. If it does, print a warning
2156 * message and return -1; otherwise, return 0.
2157 */
2158 static int
2159 swapon_check_swzone(unsigned long npages)
2160 {
2161 unsigned long maxpages;
2162
2163 /* absolute maximum we can handle assuming 100% efficiency */
2164 maxpages = uma_zone_get_max(swap_zone) * SWAP_META_PAGES;
2165
2166 /* recommend using no more than half that amount */
2167 if (npages > maxpages / 2) {
2168 printf("warning: total configured swap (%lu pages) "
2169 "exceeds maximum recommended amount (%lu pages).\n",
2170 npages, maxpages / 2);
2171 printf("warning: increase kern.maxswzone "
2172 "or reduce amount of swap.\n");
2173 return (-1);
2174 }
2175 return (0);
2176 }
2177
2178 static void
2179 swaponsomething(struct vnode *vp, void *id, u_long nblks,
2180 sw_strategy_t *strategy, sw_close_t *close, dev_t dev, int flags)
2181 {
2182 struct swdevt *sp, *tsp;
2183 swblk_t dvbase;
2184 u_long mblocks;
2185
2186 /*
2187 * nblks is in DEV_BSIZE'd chunks, convert to PAGE_SIZE'd chunks.
2188 * First chop nblks off to page-align it, then convert.
2189 *
2190 * sw->sw_nblks is in page-sized chunks now too.
2191 */
2192 nblks &= ~(ctodb(1) - 1);
2193 nblks = dbtoc(nblks);
2194
2195 /*
2196 * If we go beyond this, we get overflows in the radix
2197 * tree bitmap code.
2198 */
2199 mblocks = 0x40000000 / BLIST_META_RADIX;
2200 if (nblks > mblocks) {
2201 printf(
2202 "WARNING: reducing swap size to maximum of %luMB per unit\n",
2203 mblocks / 1024 / 1024 * PAGE_SIZE);
2204 nblks = mblocks;
2205 }
2206
2207 sp = malloc(sizeof *sp, M_VMPGDATA, M_WAITOK | M_ZERO);
2208 sp->sw_vp = vp;
2209 sp->sw_id = id;
2210 sp->sw_dev = dev;
2211 sp->sw_flags = 0;
2212 sp->sw_nblks = nblks;
2213 sp->sw_used = 0;
2214 sp->sw_strategy = strategy;
2215 sp->sw_close = close;
2216 sp->sw_flags = flags;
2217
2218 sp->sw_blist = blist_create(nblks, M_WAITOK);
2219 /*
2220 * Do not free the first two block in order to avoid overwriting
2221 * any bsd label at the front of the partition
2222 */
2223 blist_free(sp->sw_blist, 2, nblks - 2);
2224
2225 dvbase = 0;
2226 mtx_lock(&sw_dev_mtx);
2227 TAILQ_FOREACH(tsp, &swtailq, sw_list) {
2228 if (tsp->sw_end >= dvbase) {
2229 /*
2230 * We put one uncovered page between the devices
2231 * in order to definitively prevent any cross-device
2232 * I/O requests
2233 */
2234 dvbase = tsp->sw_end + 1;
2235 }
2236 }
2237 sp->sw_first = dvbase;
2238 sp->sw_end = dvbase + nblks;
2239 TAILQ_INSERT_TAIL(&swtailq, sp, sw_list);
2240 nswapdev++;
2241 swap_pager_avail += nblks;
2242 swap_total += (vm_ooffset_t)nblks * PAGE_SIZE;
2243 swapon_check_swzone(swap_total / PAGE_SIZE);
2244 swp_sizecheck();
2245 mtx_unlock(&sw_dev_mtx);
2246 }
2247
2248 /*
2249 * SYSCALL: swapoff(devname)
2250 *
2251 * Disable swapping on the given device.
2252 *
2253 * XXX: Badly designed system call: it should use a device index
2254 * rather than filename as specification. We keep sw_vp around
2255 * only to make this work.
2256 */
2257 #ifndef _SYS_SYSPROTO_H_
2258 struct swapoff_args {
2259 char *name;
2260 };
2261 #endif
2262
2263 /*
2264 * MPSAFE
2265 */
2266 /* ARGSUSED */
2267 int
2268 sys_swapoff(struct thread *td, struct swapoff_args *uap)
2269 {
2270 struct vnode *vp;
2271 struct nameidata nd;
2272 struct swdevt *sp;
2273 int error;
2274
2275 error = priv_check(td, PRIV_SWAPOFF);
2276 if (error)
2277 return (error);
2278
2279 mtx_lock(&Giant);
2280 while (swdev_syscall_active)
2281 tsleep(&swdev_syscall_active, PUSER - 1, "swpoff", 0);
2282 swdev_syscall_active = 1;
2283
2284 NDINIT(&nd, LOOKUP, FOLLOW | AUDITVNODE1, UIO_USERSPACE, uap->name,
2285 td);
2286 error = namei(&nd);
2287 if (error)
2288 goto done;
2289 NDFREE(&nd, NDF_ONLY_PNBUF);
2290 vp = nd.ni_vp;
2291
2292 mtx_lock(&sw_dev_mtx);
2293 TAILQ_FOREACH(sp, &swtailq, sw_list) {
2294 if (sp->sw_vp == vp)
2295 break;
2296 }
2297 mtx_unlock(&sw_dev_mtx);
2298 if (sp == NULL) {
2299 error = EINVAL;
2300 goto done;
2301 }
2302 error = swapoff_one(sp, td->td_ucred);
2303 done:
2304 swdev_syscall_active = 0;
2305 wakeup_one(&swdev_syscall_active);
2306 mtx_unlock(&Giant);
2307 return (error);
2308 }
2309
2310 static int
2311 swapoff_one(struct swdevt *sp, struct ucred *cred)
2312 {
2313 u_long nblks, dvbase;
2314 #ifdef MAC
2315 int error;
2316 #endif
2317
2318 mtx_assert(&Giant, MA_OWNED);
2319 #ifdef MAC
2320 (void) vn_lock(sp->sw_vp, LK_EXCLUSIVE | LK_RETRY);
2321 error = mac_system_check_swapoff(cred, sp->sw_vp);
2322 (void) VOP_UNLOCK(sp->sw_vp, 0);
2323 if (error != 0)
2324 return (error);
2325 #endif
2326 nblks = sp->sw_nblks;
2327
2328 /*
2329 * We can turn off this swap device safely only if the
2330 * available virtual memory in the system will fit the amount
2331 * of data we will have to page back in, plus an epsilon so
2332 * the system doesn't become critically low on swap space.
2333 */
2334 if (cnt.v_free_count + cnt.v_cache_count + swap_pager_avail <
2335 nblks + nswap_lowat) {
2336 return (ENOMEM);
2337 }
2338
2339 /*
2340 * Prevent further allocations on this device.
2341 */
2342 mtx_lock(&sw_dev_mtx);
2343 sp->sw_flags |= SW_CLOSING;
2344 for (dvbase = 0; dvbase < sp->sw_end; dvbase += dmmax) {
2345 swap_pager_avail -= blist_fill(sp->sw_blist,
2346 dvbase, dmmax);
2347 }
2348 swap_total -= (vm_ooffset_t)nblks * PAGE_SIZE;
2349 mtx_unlock(&sw_dev_mtx);
2350
2351 /*
2352 * Page in the contents of the device and close it.
2353 */
2354 swap_pager_swapoff(sp);
2355
2356 sp->sw_close(curthread, sp);
2357 sp->sw_id = NULL;
2358 mtx_lock(&sw_dev_mtx);
2359 TAILQ_REMOVE(&swtailq, sp, sw_list);
2360 nswapdev--;
2361 if (nswapdev == 0) {
2362 swap_pager_full = 2;
2363 swap_pager_almost_full = 1;
2364 }
2365 if (swdevhd == sp)
2366 swdevhd = NULL;
2367 mtx_unlock(&sw_dev_mtx);
2368 blist_destroy(sp->sw_blist);
2369 free(sp, M_VMPGDATA);
2370 return (0);
2371 }
2372
2373 void
2374 swapoff_all(void)
2375 {
2376 struct swdevt *sp, *spt;
2377 const char *devname;
2378 int error;
2379
2380 mtx_lock(&Giant);
2381 while (swdev_syscall_active)
2382 tsleep(&swdev_syscall_active, PUSER - 1, "swpoff", 0);
2383 swdev_syscall_active = 1;
2384
2385 mtx_lock(&sw_dev_mtx);
2386 TAILQ_FOREACH_SAFE(sp, &swtailq, sw_list, spt) {
2387 mtx_unlock(&sw_dev_mtx);
2388 if (vn_isdisk(sp->sw_vp, NULL))
2389 devname = devtoname(sp->sw_vp->v_rdev);
2390 else
2391 devname = "[file]";
2392 error = swapoff_one(sp, thread0.td_ucred);
2393 if (error != 0) {
2394 printf("Cannot remove swap device %s (error=%d), "
2395 "skipping.\n", devname, error);
2396 } else if (bootverbose) {
2397 printf("Swap device %s removed.\n", devname);
2398 }
2399 mtx_lock(&sw_dev_mtx);
2400 }
2401 mtx_unlock(&sw_dev_mtx);
2402
2403 swdev_syscall_active = 0;
2404 wakeup_one(&swdev_syscall_active);
2405 mtx_unlock(&Giant);
2406 }
2407
2408 void
2409 swap_pager_status(int *total, int *used)
2410 {
2411 struct swdevt *sp;
2412
2413 *total = 0;
2414 *used = 0;
2415 mtx_lock(&sw_dev_mtx);
2416 TAILQ_FOREACH(sp, &swtailq, sw_list) {
2417 *total += sp->sw_nblks;
2418 *used += sp->sw_used;
2419 }
2420 mtx_unlock(&sw_dev_mtx);
2421 }
2422
2423 int
2424 swap_dev_info(int name, struct xswdev *xs, char *devname, size_t len)
2425 {
2426 struct swdevt *sp;
2427 const char *tmp_devname;
2428 int error, n;
2429
2430 n = 0;
2431 error = ENOENT;
2432 mtx_lock(&sw_dev_mtx);
2433 TAILQ_FOREACH(sp, &swtailq, sw_list) {
2434 if (n != name) {
2435 n++;
2436 continue;
2437 }
2438 xs->xsw_version = XSWDEV_VERSION;
2439 xs->xsw_dev = sp->sw_dev;
2440 xs->xsw_flags = sp->sw_flags;
2441 xs->xsw_nblks = sp->sw_nblks;
2442 xs->xsw_used = sp->sw_used;
2443 if (devname != NULL) {
2444 if (vn_isdisk(sp->sw_vp, NULL))
2445 tmp_devname = devtoname(sp->sw_vp->v_rdev);
2446 else
2447 tmp_devname = "[file]";
2448 strncpy(devname, tmp_devname, len);
2449 }
2450 error = 0;
2451 break;
2452 }
2453 mtx_unlock(&sw_dev_mtx);
2454 return (error);
2455 }
2456
2457 static int
2458 sysctl_vm_swap_info(SYSCTL_HANDLER_ARGS)
2459 {
2460 struct xswdev xs;
2461 int error;
2462
2463 if (arg2 != 1) /* name length */
2464 return (EINVAL);
2465 error = swap_dev_info(*(int *)arg1, &xs, NULL, 0);
2466 if (error != 0)
2467 return (error);
2468 error = SYSCTL_OUT(req, &xs, sizeof(xs));
2469 return (error);
2470 }
2471
2472 SYSCTL_INT(_vm, OID_AUTO, nswapdev, CTLFLAG_RD, &nswapdev, 0,
2473 "Number of swap devices");
2474 SYSCTL_NODE(_vm, OID_AUTO, swap_info, CTLFLAG_RD, sysctl_vm_swap_info,
2475 "Swap statistics by device");
2476
2477 /*
2478 * vmspace_swap_count() - count the approximate swap usage in pages for a
2479 * vmspace.
2480 *
2481 * The map must be locked.
2482 *
2483 * Swap usage is determined by taking the proportional swap used by
2484 * VM objects backing the VM map. To make up for fractional losses,
2485 * if the VM object has any swap use at all the associated map entries
2486 * count for at least 1 swap page.
2487 */
2488 long
2489 vmspace_swap_count(struct vmspace *vmspace)
2490 {
2491 vm_map_t map;
2492 vm_map_entry_t cur;
2493 vm_object_t object;
2494 long count, n;
2495
2496 map = &vmspace->vm_map;
2497 count = 0;
2498
2499 for (cur = map->header.next; cur != &map->header; cur = cur->next) {
2500 if ((cur->eflags & MAP_ENTRY_IS_SUB_MAP) == 0 &&
2501 (object = cur->object.vm_object) != NULL) {
2502 VM_OBJECT_WLOCK(object);
2503 if (object->type == OBJT_SWAP &&
2504 object->un_pager.swp.swp_bcount != 0) {
2505 n = (cur->end - cur->start) / PAGE_SIZE;
2506 count += object->un_pager.swp.swp_bcount *
2507 SWAP_META_PAGES * n / object->size + 1;
2508 }
2509 VM_OBJECT_WUNLOCK(object);
2510 }
2511 }
2512 return (count);
2513 }
2514
2515 /*
2516 * GEOM backend
2517 *
2518 * Swapping onto disk devices.
2519 *
2520 */
2521
2522 static g_orphan_t swapgeom_orphan;
2523
2524 static struct g_class g_swap_class = {
2525 .name = "SWAP",
2526 .version = G_VERSION,
2527 .orphan = swapgeom_orphan,
2528 };
2529
2530 DECLARE_GEOM_CLASS(g_swap_class, g_class);
2531
2532
2533 static void
2534 swapgeom_close_ev(void *arg, int flags)
2535 {
2536 struct g_consumer *cp;
2537
2538 cp = arg;
2539 g_access(cp, -1, -1, 0);
2540 g_detach(cp);
2541 g_destroy_consumer(cp);
2542 }
2543
2544 static void
2545 swapgeom_done(struct bio *bp2)
2546 {
2547 struct swdevt *sp;
2548 struct buf *bp;
2549 struct g_consumer *cp;
2550 int destroy;
2551
2552 bp = bp2->bio_caller2;
2553 cp = bp2->bio_from;
2554 bp->b_ioflags = bp2->bio_flags;
2555 if (bp2->bio_error)
2556 bp->b_ioflags |= BIO_ERROR;
2557 bp->b_resid = bp->b_bcount - bp2->bio_completed;
2558 bp->b_error = bp2->bio_error;
2559 bufdone(bp);
2560 mtx_lock(&sw_dev_mtx);
2561 destroy = ((--cp->index) == 0 && cp->private);
2562 if (destroy) {
2563 sp = bp2->bio_caller1;
2564 sp->sw_id = NULL;
2565 }
2566 mtx_unlock(&sw_dev_mtx);
2567 g_destroy_bio(bp2);
2568 if (destroy)
2569 g_waitfor_event(swapgeom_close_ev, cp, M_WAITOK, NULL);
2570 }
2571
2572 static void
2573 swapgeom_strategy(struct buf *bp, struct swdevt *sp)
2574 {
2575 struct bio *bio;
2576 struct g_consumer *cp;
2577
2578 mtx_lock(&sw_dev_mtx);
2579 cp = sp->sw_id;
2580 if (cp == NULL) {
2581 mtx_unlock(&sw_dev_mtx);
2582 bp->b_error = ENXIO;
2583 bp->b_ioflags |= BIO_ERROR;
2584 bufdone(bp);
2585 return;
2586 }
2587 cp->index++;
2588 mtx_unlock(&sw_dev_mtx);
2589 if (bp->b_iocmd == BIO_WRITE)
2590 bio = g_new_bio();
2591 else
2592 bio = g_alloc_bio();
2593 if (bio == NULL) {
2594 bp->b_error = ENOMEM;
2595 bp->b_ioflags |= BIO_ERROR;
2596 bufdone(bp);
2597 return;
2598 }
2599
2600 bio->bio_caller1 = sp;
2601 bio->bio_caller2 = bp;
2602 bio->bio_cmd = bp->b_iocmd;
2603 bio->bio_offset = (bp->b_blkno - sp->sw_first) * PAGE_SIZE;
2604 bio->bio_length = bp->b_bcount;
2605 bio->bio_done = swapgeom_done;
2606 if ((bp->b_flags & B_UNMAPPED) != 0) {
2607 bio->bio_ma = bp->b_pages;
2608 bio->bio_data = unmapped_buf;
2609 bio->bio_ma_offset = (vm_offset_t)bp->b_offset & PAGE_MASK;
2610 bio->bio_ma_n = bp->b_npages;
2611 bio->bio_flags |= BIO_UNMAPPED;
2612 } else {
2613 bio->bio_data = bp->b_data;
2614 bio->bio_ma = NULL;
2615 }
2616 g_io_request(bio, cp);
2617 return;
2618 }
2619
2620 static void
2621 swapgeom_orphan(struct g_consumer *cp)
2622 {
2623 struct swdevt *sp;
2624 int destroy;
2625
2626 mtx_lock(&sw_dev_mtx);
2627 TAILQ_FOREACH(sp, &swtailq, sw_list) {
2628 if (sp->sw_id == cp) {
2629 sp->sw_flags |= SW_CLOSING;
2630 break;
2631 }
2632 }
2633 cp->private = (void *)(uintptr_t)1;
2634 destroy = ((sp != NULL) && (cp->index == 0));
2635 if (destroy)
2636 sp->sw_id = NULL;
2637 mtx_unlock(&sw_dev_mtx);
2638 if (destroy)
2639 swapgeom_close_ev(cp, 0);
2640 }
2641
2642 static void
2643 swapgeom_close(struct thread *td, struct swdevt *sw)
2644 {
2645 struct g_consumer *cp;
2646
2647 mtx_lock(&sw_dev_mtx);
2648 cp = sw->sw_id;
2649 sw->sw_id = NULL;
2650 mtx_unlock(&sw_dev_mtx);
2651 /* XXX: direct call when Giant untangled */
2652 if (cp != NULL)
2653 g_waitfor_event(swapgeom_close_ev, cp, M_WAITOK, NULL);
2654 }
2655
2656
2657 struct swh0h0 {
2658 struct cdev *dev;
2659 struct vnode *vp;
2660 int error;
2661 };
2662
2663 static void
2664 swapongeom_ev(void *arg, int flags)
2665 {
2666 struct swh0h0 *swh;
2667 struct g_provider *pp;
2668 struct g_consumer *cp;
2669 static struct g_geom *gp;
2670 struct swdevt *sp;
2671 u_long nblks;
2672 int error;
2673
2674 swh = arg;
2675 swh->error = 0;
2676 pp = g_dev_getprovider(swh->dev);
2677 if (pp == NULL) {
2678 swh->error = ENODEV;
2679 return;
2680 }
2681 mtx_lock(&sw_dev_mtx);
2682 TAILQ_FOREACH(sp, &swtailq, sw_list) {
2683 cp = sp->sw_id;
2684 if (cp != NULL && cp->provider == pp) {
2685 mtx_unlock(&sw_dev_mtx);
2686 swh->error = EBUSY;
2687 return;
2688 }
2689 }
2690 mtx_unlock(&sw_dev_mtx);
2691 if (gp == NULL)
2692 gp = g_new_geomf(&g_swap_class, "swap");
2693 cp = g_new_consumer(gp);
2694 cp->index = 0; /* Number of active I/Os. */
2695 cp->private = NULL; /* Orphanization flag */
2696 g_attach(cp, pp);
2697 /*
2698 * XXX: Everytime you think you can improve the margin for
2699 * footshooting, somebody depends on the ability to do so:
2700 * savecore(8) wants to write to our swapdev so we cannot
2701 * set an exclusive count :-(
2702 */
2703 error = g_access(cp, 1, 1, 0);
2704 if (error) {
2705 g_detach(cp);
2706 g_destroy_consumer(cp);
2707 swh->error = error;
2708 return;
2709 }
2710 nblks = pp->mediasize / DEV_BSIZE;
2711 swaponsomething(swh->vp, cp, nblks, swapgeom_strategy,
2712 swapgeom_close, dev2udev(swh->dev),
2713 (pp->flags & G_PF_ACCEPT_UNMAPPED) != 0 ? SW_UNMAPPED : 0);
2714 swh->error = 0;
2715 }
2716
2717 static int
2718 swapongeom(struct thread *td, struct vnode *vp)
2719 {
2720 int error;
2721 struct swh0h0 swh;
2722
2723 vn_lock(vp, LK_EXCLUSIVE | LK_RETRY);
2724
2725 swh.dev = vp->v_rdev;
2726 swh.vp = vp;
2727 swh.error = 0;
2728 /* XXX: direct call when Giant untangled */
2729 error = g_waitfor_event(swapongeom_ev, &swh, M_WAITOK, NULL);
2730 if (!error)
2731 error = swh.error;
2732 VOP_UNLOCK(vp, 0);
2733 return (error);
2734 }
2735
2736 /*
2737 * VNODE backend
2738 *
2739 * This is used mainly for network filesystem (read: probably only tested
2740 * with NFS) swapfiles.
2741 *
2742 */
2743
2744 static void
2745 swapdev_strategy(struct buf *bp, struct swdevt *sp)
2746 {
2747 struct vnode *vp2;
2748
2749 bp->b_blkno = ctodb(bp->b_blkno - sp->sw_first);
2750
2751 vp2 = sp->sw_id;
2752 vhold(vp2);
2753 if (bp->b_iocmd == BIO_WRITE) {
2754 if (bp->b_bufobj)
2755 bufobj_wdrop(bp->b_bufobj);
2756 bufobj_wref(&vp2->v_bufobj);
2757 }
2758 if (bp->b_bufobj != &vp2->v_bufobj)
2759 bp->b_bufobj = &vp2->v_bufobj;
2760 bp->b_vp = vp2;
2761 bp->b_iooffset = dbtob(bp->b_blkno);
2762 bstrategy(bp);
2763 return;
2764 }
2765
2766 static void
2767 swapdev_close(struct thread *td, struct swdevt *sp)
2768 {
2769
2770 VOP_CLOSE(sp->sw_vp, FREAD | FWRITE, td->td_ucred, td);
2771 vrele(sp->sw_vp);
2772 }
2773
2774
2775 static int
2776 swaponvp(struct thread *td, struct vnode *vp, u_long nblks)
2777 {
2778 struct swdevt *sp;
2779 int error;
2780
2781 if (nblks == 0)
2782 return (ENXIO);
2783 mtx_lock(&sw_dev_mtx);
2784 TAILQ_FOREACH(sp, &swtailq, sw_list) {
2785 if (sp->sw_id == vp) {
2786 mtx_unlock(&sw_dev_mtx);
2787 return (EBUSY);
2788 }
2789 }
2790 mtx_unlock(&sw_dev_mtx);
2791
2792 (void) vn_lock(vp, LK_EXCLUSIVE | LK_RETRY);
2793 #ifdef MAC
2794 error = mac_system_check_swapon(td->td_ucred, vp);
2795 if (error == 0)
2796 #endif
2797 error = VOP_OPEN(vp, FREAD | FWRITE, td->td_ucred, td, NULL);
2798 (void) VOP_UNLOCK(vp, 0);
2799 if (error)
2800 return (error);
2801
2802 swaponsomething(vp, vp, nblks, swapdev_strategy, swapdev_close,
2803 NODEV, 0);
2804 return (0);
2805 }
Cache object: da4443a7969479718a8d27a4ef5a0124
|