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/8.0/sys/vm/swap_pager.c 198330 2009-10-21 15:07:34Z kib $");
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/resource.h>
90 #include <sys/resourcevar.h>
91 #include <sys/sysctl.h>
92 #include <sys/sysproto.h>
93 #include <sys/blist.h>
94 #include <sys/lock.h>
95 #include <sys/sx.h>
96 #include <sys/vmmeter.h>
97
98 #include <security/mac/mac_framework.h>
99
100 #include <vm/vm.h>
101 #include <vm/pmap.h>
102 #include <vm/vm_map.h>
103 #include <vm/vm_kern.h>
104 #include <vm/vm_object.h>
105 #include <vm/vm_page.h>
106 #include <vm/vm_pager.h>
107 #include <vm/vm_pageout.h>
108 #include <vm/vm_param.h>
109 #include <vm/swap_pager.h>
110 #include <vm/vm_extern.h>
111 #include <vm/uma.h>
112
113 #include <geom/geom.h>
114
115 /*
116 * SWB_NPAGES must be a power of 2. It may be set to 1, 2, 4, 8, or 16
117 * pages per allocation. We recommend you stick with the default of 8.
118 * The 16-page limit is due to the radix code (kern/subr_blist.c).
119 */
120 #ifndef MAX_PAGEOUT_CLUSTER
121 #define MAX_PAGEOUT_CLUSTER 16
122 #endif
123
124 #if !defined(SWB_NPAGES)
125 #define SWB_NPAGES MAX_PAGEOUT_CLUSTER
126 #endif
127
128 /*
129 * Piecemeal swap metadata structure. Swap is stored in a radix tree.
130 *
131 * If SWB_NPAGES is 8 and sizeof(char *) == sizeof(daddr_t), our radix
132 * is basically 8. Assuming PAGE_SIZE == 4096, one tree level represents
133 * 32K worth of data, two levels represent 256K, three levels represent
134 * 2 MBytes. This is acceptable.
135 *
136 * Overall memory utilization is about the same as the old swap structure.
137 */
138 #define SWCORRECT(n) (sizeof(void *) * (n) / sizeof(daddr_t))
139 #define SWAP_META_PAGES (SWB_NPAGES * 2)
140 #define SWAP_META_MASK (SWAP_META_PAGES - 1)
141
142 struct swblock {
143 struct swblock *swb_hnext;
144 vm_object_t swb_object;
145 vm_pindex_t swb_index;
146 int swb_count;
147 daddr_t swb_pages[SWAP_META_PAGES];
148 };
149
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 static vm_ooffset_t swap_reserved;
160 SYSCTL_QUAD(_vm, OID_AUTO, swap_reserved, CTLFLAG_RD, &swap_reserved, 0, "");
161 static int overcommit = 0;
162 SYSCTL_INT(_vm, OID_AUTO, overcommit, CTLFLAG_RW, &overcommit, 0, "");
163
164 /* bits from overcommit */
165 #define SWAP_RESERVE_FORCE_ON (1 << 0)
166 #define SWAP_RESERVE_RLIMIT_ON (1 << 1)
167 #define SWAP_RESERVE_ALLOW_NONWIRED (1 << 2)
168
169 int
170 swap_reserve(vm_ooffset_t incr)
171 {
172
173 return (swap_reserve_by_uid(incr, curthread->td_ucred->cr_ruidinfo));
174 }
175
176 int
177 swap_reserve_by_uid(vm_ooffset_t incr, struct uidinfo *uip)
178 {
179 vm_ooffset_t r, s;
180 int res, error;
181 static int curfail;
182 static struct timeval lastfail;
183
184 if (incr & PAGE_MASK)
185 panic("swap_reserve: & PAGE_MASK");
186
187 res = 0;
188 mtx_lock(&sw_dev_mtx);
189 r = swap_reserved + incr;
190 if (overcommit & SWAP_RESERVE_ALLOW_NONWIRED) {
191 s = cnt.v_page_count - cnt.v_free_reserved - cnt.v_wire_count;
192 s *= PAGE_SIZE;
193 } else
194 s = 0;
195 s += swap_total;
196 if ((overcommit & SWAP_RESERVE_FORCE_ON) == 0 || r <= s ||
197 (error = priv_check(curthread, PRIV_VM_SWAP_NOQUOTA)) == 0) {
198 res = 1;
199 swap_reserved = r;
200 }
201 mtx_unlock(&sw_dev_mtx);
202
203 if (res) {
204 PROC_LOCK(curproc);
205 UIDINFO_VMSIZE_LOCK(uip);
206 if ((overcommit & SWAP_RESERVE_RLIMIT_ON) != 0 &&
207 uip->ui_vmsize + incr > lim_cur(curproc, RLIMIT_SWAP) &&
208 priv_check(curthread, PRIV_VM_SWAP_NORLIMIT))
209 res = 0;
210 else
211 uip->ui_vmsize += incr;
212 UIDINFO_VMSIZE_UNLOCK(uip);
213 PROC_UNLOCK(curproc);
214 if (!res) {
215 mtx_lock(&sw_dev_mtx);
216 swap_reserved -= incr;
217 mtx_unlock(&sw_dev_mtx);
218 }
219 }
220 if (!res && ppsratecheck(&lastfail, &curfail, 1)) {
221 printf("uid %d, pid %d: swap reservation for %jd bytes failed\n",
222 curproc->p_pid, uip->ui_uid, incr);
223 }
224
225 return (res);
226 }
227
228 void
229 swap_reserve_force(vm_ooffset_t incr)
230 {
231 struct uidinfo *uip;
232
233 mtx_lock(&sw_dev_mtx);
234 swap_reserved += incr;
235 mtx_unlock(&sw_dev_mtx);
236
237 uip = curthread->td_ucred->cr_ruidinfo;
238 PROC_LOCK(curproc);
239 UIDINFO_VMSIZE_LOCK(uip);
240 uip->ui_vmsize += incr;
241 UIDINFO_VMSIZE_UNLOCK(uip);
242 PROC_UNLOCK(curproc);
243 }
244
245 void
246 swap_release(vm_ooffset_t decr)
247 {
248 struct uidinfo *uip;
249
250 PROC_LOCK(curproc);
251 uip = curthread->td_ucred->cr_ruidinfo;
252 swap_release_by_uid(decr, uip);
253 PROC_UNLOCK(curproc);
254 }
255
256 void
257 swap_release_by_uid(vm_ooffset_t decr, struct uidinfo *uip)
258 {
259
260 if (decr & PAGE_MASK)
261 panic("swap_release: & PAGE_MASK");
262
263 mtx_lock(&sw_dev_mtx);
264 if (swap_reserved < decr)
265 panic("swap_reserved < decr");
266 swap_reserved -= decr;
267 mtx_unlock(&sw_dev_mtx);
268
269 UIDINFO_VMSIZE_LOCK(uip);
270 if (uip->ui_vmsize < decr)
271 printf("negative vmsize for uid = %d\n", uip->ui_uid);
272 uip->ui_vmsize -= decr;
273 UIDINFO_VMSIZE_UNLOCK(uip);
274 }
275
276 static void swapdev_strategy(struct buf *, struct swdevt *sw);
277
278 #define SWM_FREE 0x02 /* free, period */
279 #define SWM_POP 0x04 /* pop out */
280
281 int swap_pager_full = 2; /* swap space exhaustion (task killing) */
282 static int swap_pager_almost_full = 1; /* swap space exhaustion (w/hysteresis)*/
283 static int nsw_rcount; /* free read buffers */
284 static int nsw_wcount_sync; /* limit write buffers / synchronous */
285 static int nsw_wcount_async; /* limit write buffers / asynchronous */
286 static int nsw_wcount_async_max;/* assigned maximum */
287 static int nsw_cluster_max; /* maximum VOP I/O allowed */
288
289 static struct swblock **swhash;
290 static int swhash_mask;
291 static struct mtx swhash_mtx;
292
293 static int swap_async_max = 4; /* maximum in-progress async I/O's */
294 static struct sx sw_alloc_sx;
295
296
297 SYSCTL_INT(_vm, OID_AUTO, swap_async_max,
298 CTLFLAG_RW, &swap_async_max, 0, "Maximum running async swap ops");
299
300 /*
301 * "named" and "unnamed" anon region objects. Try to reduce the overhead
302 * of searching a named list by hashing it just a little.
303 */
304
305 #define NOBJLISTS 8
306
307 #define NOBJLIST(handle) \
308 (&swap_pager_object_list[((int)(intptr_t)handle >> 4) & (NOBJLISTS-1)])
309
310 static struct mtx sw_alloc_mtx; /* protect list manipulation */
311 static struct pagerlst swap_pager_object_list[NOBJLISTS];
312 static uma_zone_t swap_zone;
313 static struct vm_object swap_zone_obj;
314
315 /*
316 * pagerops for OBJT_SWAP - "swap pager". Some ops are also global procedure
317 * calls hooked from other parts of the VM system and do not appear here.
318 * (see vm/swap_pager.h).
319 */
320 static vm_object_t
321 swap_pager_alloc(void *handle, vm_ooffset_t size,
322 vm_prot_t prot, vm_ooffset_t offset, struct ucred *);
323 static void swap_pager_dealloc(vm_object_t object);
324 static int swap_pager_getpages(vm_object_t, vm_page_t *, int, int);
325 static void swap_pager_putpages(vm_object_t, vm_page_t *, int, boolean_t, int *);
326 static boolean_t
327 swap_pager_haspage(vm_object_t object, vm_pindex_t pindex, int *before, int *after);
328 static void swap_pager_init(void);
329 static void swap_pager_unswapped(vm_page_t);
330 static void swap_pager_swapoff(struct swdevt *sp);
331
332 struct pagerops swappagerops = {
333 .pgo_init = swap_pager_init, /* early system initialization of pager */
334 .pgo_alloc = swap_pager_alloc, /* allocate an OBJT_SWAP object */
335 .pgo_dealloc = swap_pager_dealloc, /* deallocate an OBJT_SWAP object */
336 .pgo_getpages = swap_pager_getpages, /* pagein */
337 .pgo_putpages = swap_pager_putpages, /* pageout */
338 .pgo_haspage = swap_pager_haspage, /* get backing store status for page */
339 .pgo_pageunswapped = swap_pager_unswapped, /* remove swap related to page */
340 };
341
342 /*
343 * dmmax is in page-sized chunks with the new swap system. It was
344 * dev-bsized chunks in the old. dmmax is always a power of 2.
345 *
346 * swap_*() routines are externally accessible. swp_*() routines are
347 * internal.
348 */
349 static int dmmax;
350 static int nswap_lowat = 128; /* in pages, swap_pager_almost_full warn */
351 static int nswap_hiwat = 512; /* in pages, swap_pager_almost_full warn */
352
353 SYSCTL_INT(_vm, OID_AUTO, dmmax,
354 CTLFLAG_RD, &dmmax, 0, "Maximum size of a swap block");
355
356 static void swp_sizecheck(void);
357 static void swp_pager_async_iodone(struct buf *bp);
358 static int swapongeom(struct thread *, struct vnode *);
359 static int swaponvp(struct thread *, struct vnode *, u_long);
360 static int swapoff_one(struct swdevt *sp, struct ucred *cred);
361
362 /*
363 * Swap bitmap functions
364 */
365 static void swp_pager_freeswapspace(daddr_t blk, int npages);
366 static daddr_t swp_pager_getswapspace(int npages);
367
368 /*
369 * Metadata functions
370 */
371 static struct swblock **swp_pager_hash(vm_object_t object, vm_pindex_t index);
372 static void swp_pager_meta_build(vm_object_t, vm_pindex_t, daddr_t);
373 static void swp_pager_meta_free(vm_object_t, vm_pindex_t, daddr_t);
374 static void swp_pager_meta_free_all(vm_object_t);
375 static daddr_t swp_pager_meta_ctl(vm_object_t, vm_pindex_t, int);
376
377 /*
378 * SWP_SIZECHECK() - update swap_pager_full indication
379 *
380 * update the swap_pager_almost_full indication and warn when we are
381 * about to run out of swap space, using lowat/hiwat hysteresis.
382 *
383 * Clear swap_pager_full ( task killing ) indication when lowat is met.
384 *
385 * No restrictions on call
386 * This routine may not block.
387 * This routine must be called at splvm()
388 */
389 static void
390 swp_sizecheck(void)
391 {
392
393 if (swap_pager_avail < nswap_lowat) {
394 if (swap_pager_almost_full == 0) {
395 printf("swap_pager: out of swap space\n");
396 swap_pager_almost_full = 1;
397 }
398 } else {
399 swap_pager_full = 0;
400 if (swap_pager_avail > nswap_hiwat)
401 swap_pager_almost_full = 0;
402 }
403 }
404
405 /*
406 * SWP_PAGER_HASH() - hash swap meta data
407 *
408 * This is an helper function which hashes the swapblk given
409 * the object and page index. It returns a pointer to a pointer
410 * to the object, or a pointer to a NULL pointer if it could not
411 * find a swapblk.
412 *
413 * This routine must be called at splvm().
414 */
415 static struct swblock **
416 swp_pager_hash(vm_object_t object, vm_pindex_t index)
417 {
418 struct swblock **pswap;
419 struct swblock *swap;
420
421 index &= ~(vm_pindex_t)SWAP_META_MASK;
422 pswap = &swhash[(index ^ (int)(intptr_t)object) & swhash_mask];
423 while ((swap = *pswap) != NULL) {
424 if (swap->swb_object == object &&
425 swap->swb_index == index
426 ) {
427 break;
428 }
429 pswap = &swap->swb_hnext;
430 }
431 return (pswap);
432 }
433
434 /*
435 * SWAP_PAGER_INIT() - initialize the swap pager!
436 *
437 * Expected to be started from system init. NOTE: This code is run
438 * before much else so be careful what you depend on. Most of the VM
439 * system has yet to be initialized at this point.
440 */
441 static void
442 swap_pager_init(void)
443 {
444 /*
445 * Initialize object lists
446 */
447 int i;
448
449 for (i = 0; i < NOBJLISTS; ++i)
450 TAILQ_INIT(&swap_pager_object_list[i]);
451 mtx_init(&sw_alloc_mtx, "swap_pager list", NULL, MTX_DEF);
452 mtx_init(&sw_dev_mtx, "swapdev", NULL, MTX_DEF);
453
454 /*
455 * Device Stripe, in PAGE_SIZE'd blocks
456 */
457 dmmax = SWB_NPAGES * 2;
458 }
459
460 /*
461 * SWAP_PAGER_SWAP_INIT() - swap pager initialization from pageout process
462 *
463 * Expected to be started from pageout process once, prior to entering
464 * its main loop.
465 */
466 void
467 swap_pager_swap_init(void)
468 {
469 int n, n2;
470
471 /*
472 * Number of in-transit swap bp operations. Don't
473 * exhaust the pbufs completely. Make sure we
474 * initialize workable values (0 will work for hysteresis
475 * but it isn't very efficient).
476 *
477 * The nsw_cluster_max is constrained by the bp->b_pages[]
478 * array (MAXPHYS/PAGE_SIZE) and our locally defined
479 * MAX_PAGEOUT_CLUSTER. Also be aware that swap ops are
480 * constrained by the swap device interleave stripe size.
481 *
482 * Currently we hardwire nsw_wcount_async to 4. This limit is
483 * designed to prevent other I/O from having high latencies due to
484 * our pageout I/O. The value 4 works well for one or two active swap
485 * devices but is probably a little low if you have more. Even so,
486 * a higher value would probably generate only a limited improvement
487 * with three or four active swap devices since the system does not
488 * typically have to pageout at extreme bandwidths. We will want
489 * at least 2 per swap devices, and 4 is a pretty good value if you
490 * have one NFS swap device due to the command/ack latency over NFS.
491 * So it all works out pretty well.
492 */
493 nsw_cluster_max = min((MAXPHYS/PAGE_SIZE), MAX_PAGEOUT_CLUSTER);
494
495 mtx_lock(&pbuf_mtx);
496 nsw_rcount = (nswbuf + 1) / 2;
497 nsw_wcount_sync = (nswbuf + 3) / 4;
498 nsw_wcount_async = 4;
499 nsw_wcount_async_max = nsw_wcount_async;
500 mtx_unlock(&pbuf_mtx);
501
502 /*
503 * Initialize our zone. Right now I'm just guessing on the number
504 * we need based on the number of pages in the system. Each swblock
505 * can hold 16 pages, so this is probably overkill. This reservation
506 * is typically limited to around 32MB by default.
507 */
508 n = cnt.v_page_count / 2;
509 if (maxswzone && n > maxswzone / sizeof(struct swblock))
510 n = maxswzone / sizeof(struct swblock);
511 n2 = n;
512 swap_zone = uma_zcreate("SWAPMETA", sizeof(struct swblock), NULL, NULL,
513 NULL, NULL, UMA_ALIGN_PTR, UMA_ZONE_NOFREE | UMA_ZONE_VM);
514 if (swap_zone == NULL)
515 panic("failed to create swap_zone.");
516 do {
517 if (uma_zone_set_obj(swap_zone, &swap_zone_obj, n))
518 break;
519 /*
520 * if the allocation failed, try a zone two thirds the
521 * size of the previous attempt.
522 */
523 n -= ((n + 2) / 3);
524 } while (n > 0);
525 if (n2 != n)
526 printf("Swap zone entries reduced from %d to %d.\n", n2, n);
527 n2 = n;
528
529 /*
530 * Initialize our meta-data hash table. The swapper does not need to
531 * be quite as efficient as the VM system, so we do not use an
532 * oversized hash table.
533 *
534 * n: size of hash table, must be power of 2
535 * swhash_mask: hash table index mask
536 */
537 for (n = 1; n < n2 / 8; n *= 2)
538 ;
539 swhash = malloc(sizeof(struct swblock *) * n, M_VMPGDATA, M_WAITOK | M_ZERO);
540 swhash_mask = n - 1;
541 mtx_init(&swhash_mtx, "swap_pager swhash", NULL, MTX_DEF);
542 }
543
544 /*
545 * SWAP_PAGER_ALLOC() - allocate a new OBJT_SWAP VM object and instantiate
546 * its metadata structures.
547 *
548 * This routine is called from the mmap and fork code to create a new
549 * OBJT_SWAP object. We do this by creating an OBJT_DEFAULT object
550 * and then converting it with swp_pager_meta_build().
551 *
552 * This routine may block in vm_object_allocate() and create a named
553 * object lookup race, so we must interlock. We must also run at
554 * splvm() for the object lookup to handle races with interrupts, but
555 * we do not have to maintain splvm() in between the lookup and the
556 * add because (I believe) it is not possible to attempt to create
557 * a new swap object w/handle when a default object with that handle
558 * already exists.
559 *
560 * MPSAFE
561 */
562 static vm_object_t
563 swap_pager_alloc(void *handle, vm_ooffset_t size, vm_prot_t prot,
564 vm_ooffset_t offset, struct ucred *cred)
565 {
566 vm_object_t object;
567 vm_pindex_t pindex;
568 struct uidinfo *uip;
569
570 uip = NULL;
571 pindex = OFF_TO_IDX(offset + PAGE_MASK + size);
572 if (handle) {
573 mtx_lock(&Giant);
574 /*
575 * Reference existing named region or allocate new one. There
576 * should not be a race here against swp_pager_meta_build()
577 * as called from vm_page_remove() in regards to the lookup
578 * of the handle.
579 */
580 sx_xlock(&sw_alloc_sx);
581 object = vm_pager_object_lookup(NOBJLIST(handle), handle);
582 if (object == NULL) {
583 if (cred != NULL) {
584 uip = cred->cr_ruidinfo;
585 if (!swap_reserve_by_uid(size, uip)) {
586 sx_xunlock(&sw_alloc_sx);
587 mtx_unlock(&Giant);
588 return (NULL);
589 }
590 uihold(uip);
591 }
592 object = vm_object_allocate(OBJT_DEFAULT, pindex);
593 VM_OBJECT_LOCK(object);
594 object->handle = handle;
595 if (cred != NULL) {
596 object->uip = uip;
597 object->charge = size;
598 }
599 swp_pager_meta_build(object, 0, SWAPBLK_NONE);
600 VM_OBJECT_UNLOCK(object);
601 }
602 sx_xunlock(&sw_alloc_sx);
603 mtx_unlock(&Giant);
604 } else {
605 if (cred != NULL) {
606 uip = cred->cr_ruidinfo;
607 if (!swap_reserve_by_uid(size, uip))
608 return (NULL);
609 uihold(uip);
610 }
611 object = vm_object_allocate(OBJT_DEFAULT, pindex);
612 VM_OBJECT_LOCK(object);
613 if (cred != NULL) {
614 object->uip = uip;
615 object->charge = size;
616 }
617 swp_pager_meta_build(object, 0, SWAPBLK_NONE);
618 VM_OBJECT_UNLOCK(object);
619 }
620 return (object);
621 }
622
623 /*
624 * SWAP_PAGER_DEALLOC() - remove swap metadata from object
625 *
626 * The swap backing for the object is destroyed. The code is
627 * designed such that we can reinstantiate it later, but this
628 * routine is typically called only when the entire object is
629 * about to be destroyed.
630 *
631 * This routine may block, but no longer does.
632 *
633 * The object must be locked or unreferenceable.
634 */
635 static void
636 swap_pager_dealloc(vm_object_t object)
637 {
638
639 /*
640 * Remove from list right away so lookups will fail if we block for
641 * pageout completion.
642 */
643 if (object->handle != NULL) {
644 mtx_lock(&sw_alloc_mtx);
645 TAILQ_REMOVE(NOBJLIST(object->handle), object, pager_object_list);
646 mtx_unlock(&sw_alloc_mtx);
647 }
648
649 VM_OBJECT_LOCK_ASSERT(object, MA_OWNED);
650 vm_object_pip_wait(object, "swpdea");
651
652 /*
653 * Free all remaining metadata. We only bother to free it from
654 * the swap meta data. We do not attempt to free swapblk's still
655 * associated with vm_page_t's for this object. We do not care
656 * if paging is still in progress on some objects.
657 */
658 swp_pager_meta_free_all(object);
659 }
660
661 /************************************************************************
662 * SWAP PAGER BITMAP ROUTINES *
663 ************************************************************************/
664
665 /*
666 * SWP_PAGER_GETSWAPSPACE() - allocate raw swap space
667 *
668 * Allocate swap for the requested number of pages. The starting
669 * swap block number (a page index) is returned or SWAPBLK_NONE
670 * if the allocation failed.
671 *
672 * Also has the side effect of advising that somebody made a mistake
673 * when they configured swap and didn't configure enough.
674 *
675 * Must be called at splvm() to avoid races with bitmap frees from
676 * vm_page_remove() aka swap_pager_page_removed().
677 *
678 * This routine may not block
679 * This routine must be called at splvm().
680 *
681 * We allocate in round-robin fashion from the configured devices.
682 */
683 static daddr_t
684 swp_pager_getswapspace(int npages)
685 {
686 daddr_t blk;
687 struct swdevt *sp;
688 int i;
689
690 blk = SWAPBLK_NONE;
691 mtx_lock(&sw_dev_mtx);
692 sp = swdevhd;
693 for (i = 0; i < nswapdev; i++) {
694 if (sp == NULL)
695 sp = TAILQ_FIRST(&swtailq);
696 if (!(sp->sw_flags & SW_CLOSING)) {
697 blk = blist_alloc(sp->sw_blist, npages);
698 if (blk != SWAPBLK_NONE) {
699 blk += sp->sw_first;
700 sp->sw_used += npages;
701 swap_pager_avail -= npages;
702 swp_sizecheck();
703 swdevhd = TAILQ_NEXT(sp, sw_list);
704 goto done;
705 }
706 }
707 sp = TAILQ_NEXT(sp, sw_list);
708 }
709 if (swap_pager_full != 2) {
710 printf("swap_pager_getswapspace(%d): failed\n", npages);
711 swap_pager_full = 2;
712 swap_pager_almost_full = 1;
713 }
714 swdevhd = NULL;
715 done:
716 mtx_unlock(&sw_dev_mtx);
717 return (blk);
718 }
719
720 static int
721 swp_pager_isondev(daddr_t blk, struct swdevt *sp)
722 {
723
724 return (blk >= sp->sw_first && blk < sp->sw_end);
725 }
726
727 static void
728 swp_pager_strategy(struct buf *bp)
729 {
730 struct swdevt *sp;
731
732 mtx_lock(&sw_dev_mtx);
733 TAILQ_FOREACH(sp, &swtailq, sw_list) {
734 if (bp->b_blkno >= sp->sw_first && bp->b_blkno < sp->sw_end) {
735 mtx_unlock(&sw_dev_mtx);
736 sp->sw_strategy(bp, sp);
737 return;
738 }
739 }
740 panic("Swapdev not found");
741 }
742
743
744 /*
745 * SWP_PAGER_FREESWAPSPACE() - free raw swap space
746 *
747 * This routine returns the specified swap blocks back to the bitmap.
748 *
749 * Note: This routine may not block (it could in the old swap code),
750 * and through the use of the new blist routines it does not block.
751 *
752 * We must be called at splvm() to avoid races with bitmap frees from
753 * vm_page_remove() aka swap_pager_page_removed().
754 *
755 * This routine may not block
756 * This routine must be called at splvm().
757 */
758 static void
759 swp_pager_freeswapspace(daddr_t blk, int npages)
760 {
761 struct swdevt *sp;
762
763 mtx_lock(&sw_dev_mtx);
764 TAILQ_FOREACH(sp, &swtailq, sw_list) {
765 if (blk >= sp->sw_first && blk < sp->sw_end) {
766 sp->sw_used -= npages;
767 /*
768 * If we are attempting to stop swapping on
769 * this device, we don't want to mark any
770 * blocks free lest they be reused.
771 */
772 if ((sp->sw_flags & SW_CLOSING) == 0) {
773 blist_free(sp->sw_blist, blk - sp->sw_first,
774 npages);
775 swap_pager_avail += npages;
776 swp_sizecheck();
777 }
778 mtx_unlock(&sw_dev_mtx);
779 return;
780 }
781 }
782 panic("Swapdev not found");
783 }
784
785 /*
786 * SWAP_PAGER_FREESPACE() - frees swap blocks associated with a page
787 * range within an object.
788 *
789 * This is a globally accessible routine.
790 *
791 * This routine removes swapblk assignments from swap metadata.
792 *
793 * The external callers of this routine typically have already destroyed
794 * or renamed vm_page_t's associated with this range in the object so
795 * we should be ok.
796 *
797 * This routine may be called at any spl. We up our spl to splvm temporarily
798 * in order to perform the metadata removal.
799 */
800 void
801 swap_pager_freespace(vm_object_t object, vm_pindex_t start, vm_size_t size)
802 {
803
804 VM_OBJECT_LOCK_ASSERT(object, MA_OWNED);
805 swp_pager_meta_free(object, start, size);
806 }
807
808 /*
809 * SWAP_PAGER_RESERVE() - reserve swap blocks in object
810 *
811 * Assigns swap blocks to the specified range within the object. The
812 * swap blocks are not zerod. Any previous swap assignment is destroyed.
813 *
814 * Returns 0 on success, -1 on failure.
815 */
816 int
817 swap_pager_reserve(vm_object_t object, vm_pindex_t start, vm_size_t size)
818 {
819 int n = 0;
820 daddr_t blk = SWAPBLK_NONE;
821 vm_pindex_t beg = start; /* save start index */
822
823 VM_OBJECT_LOCK(object);
824 while (size) {
825 if (n == 0) {
826 n = BLIST_MAX_ALLOC;
827 while ((blk = swp_pager_getswapspace(n)) == SWAPBLK_NONE) {
828 n >>= 1;
829 if (n == 0) {
830 swp_pager_meta_free(object, beg, start - beg);
831 VM_OBJECT_UNLOCK(object);
832 return (-1);
833 }
834 }
835 }
836 swp_pager_meta_build(object, start, blk);
837 --size;
838 ++start;
839 ++blk;
840 --n;
841 }
842 swp_pager_meta_free(object, start, n);
843 VM_OBJECT_UNLOCK(object);
844 return (0);
845 }
846
847 /*
848 * SWAP_PAGER_COPY() - copy blocks from source pager to destination pager
849 * and destroy the source.
850 *
851 * Copy any valid swapblks from the source to the destination. In
852 * cases where both the source and destination have a valid swapblk,
853 * we keep the destination's.
854 *
855 * This routine is allowed to block. It may block allocating metadata
856 * indirectly through swp_pager_meta_build() or if paging is still in
857 * progress on the source.
858 *
859 * This routine can be called at any spl
860 *
861 * XXX vm_page_collapse() kinda expects us not to block because we
862 * supposedly do not need to allocate memory, but for the moment we
863 * *may* have to get a little memory from the zone allocator, but
864 * it is taken from the interrupt memory. We should be ok.
865 *
866 * The source object contains no vm_page_t's (which is just as well)
867 *
868 * The source object is of type OBJT_SWAP.
869 *
870 * The source and destination objects must be locked or
871 * inaccessible (XXX are they ?)
872 */
873 void
874 swap_pager_copy(vm_object_t srcobject, vm_object_t dstobject,
875 vm_pindex_t offset, int destroysource)
876 {
877 vm_pindex_t i;
878
879 VM_OBJECT_LOCK_ASSERT(srcobject, MA_OWNED);
880 VM_OBJECT_LOCK_ASSERT(dstobject, MA_OWNED);
881
882 /*
883 * If destroysource is set, we remove the source object from the
884 * swap_pager internal queue now.
885 */
886 if (destroysource) {
887 if (srcobject->handle != NULL) {
888 mtx_lock(&sw_alloc_mtx);
889 TAILQ_REMOVE(
890 NOBJLIST(srcobject->handle),
891 srcobject,
892 pager_object_list
893 );
894 mtx_unlock(&sw_alloc_mtx);
895 }
896 }
897
898 /*
899 * transfer source to destination.
900 */
901 for (i = 0; i < dstobject->size; ++i) {
902 daddr_t dstaddr;
903
904 /*
905 * Locate (without changing) the swapblk on the destination,
906 * unless it is invalid in which case free it silently, or
907 * if the destination is a resident page, in which case the
908 * source is thrown away.
909 */
910 dstaddr = swp_pager_meta_ctl(dstobject, i, 0);
911
912 if (dstaddr == SWAPBLK_NONE) {
913 /*
914 * Destination has no swapblk and is not resident,
915 * copy source.
916 */
917 daddr_t srcaddr;
918
919 srcaddr = swp_pager_meta_ctl(
920 srcobject,
921 i + offset,
922 SWM_POP
923 );
924
925 if (srcaddr != SWAPBLK_NONE) {
926 /*
927 * swp_pager_meta_build() can sleep.
928 */
929 vm_object_pip_add(srcobject, 1);
930 VM_OBJECT_UNLOCK(srcobject);
931 vm_object_pip_add(dstobject, 1);
932 swp_pager_meta_build(dstobject, i, srcaddr);
933 vm_object_pip_wakeup(dstobject);
934 VM_OBJECT_LOCK(srcobject);
935 vm_object_pip_wakeup(srcobject);
936 }
937 } else {
938 /*
939 * Destination has valid swapblk or it is represented
940 * by a resident page. We destroy the sourceblock.
941 */
942
943 swp_pager_meta_ctl(srcobject, i + offset, SWM_FREE);
944 }
945 }
946
947 /*
948 * Free left over swap blocks in source.
949 *
950 * We have to revert the type to OBJT_DEFAULT so we do not accidently
951 * double-remove the object from the swap queues.
952 */
953 if (destroysource) {
954 swp_pager_meta_free_all(srcobject);
955 /*
956 * Reverting the type is not necessary, the caller is going
957 * to destroy srcobject directly, but I'm doing it here
958 * for consistency since we've removed the object from its
959 * queues.
960 */
961 srcobject->type = OBJT_DEFAULT;
962 }
963 }
964
965 /*
966 * SWAP_PAGER_HASPAGE() - determine if we have good backing store for
967 * the requested page.
968 *
969 * We determine whether good backing store exists for the requested
970 * page and return TRUE if it does, FALSE if it doesn't.
971 *
972 * If TRUE, we also try to determine how much valid, contiguous backing
973 * store exists before and after the requested page within a reasonable
974 * distance. We do not try to restrict it to the swap device stripe
975 * (that is handled in getpages/putpages). It probably isn't worth
976 * doing here.
977 */
978 static boolean_t
979 swap_pager_haspage(vm_object_t object, vm_pindex_t pindex, int *before, int *after)
980 {
981 daddr_t blk0;
982
983 VM_OBJECT_LOCK_ASSERT(object, MA_OWNED);
984 /*
985 * do we have good backing store at the requested index ?
986 */
987 blk0 = swp_pager_meta_ctl(object, pindex, 0);
988
989 if (blk0 == SWAPBLK_NONE) {
990 if (before)
991 *before = 0;
992 if (after)
993 *after = 0;
994 return (FALSE);
995 }
996
997 /*
998 * find backwards-looking contiguous good backing store
999 */
1000 if (before != NULL) {
1001 int i;
1002
1003 for (i = 1; i < (SWB_NPAGES/2); ++i) {
1004 daddr_t blk;
1005
1006 if (i > pindex)
1007 break;
1008 blk = swp_pager_meta_ctl(object, pindex - i, 0);
1009 if (blk != blk0 - i)
1010 break;
1011 }
1012 *before = (i - 1);
1013 }
1014
1015 /*
1016 * find forward-looking contiguous good backing store
1017 */
1018 if (after != NULL) {
1019 int i;
1020
1021 for (i = 1; i < (SWB_NPAGES/2); ++i) {
1022 daddr_t blk;
1023
1024 blk = swp_pager_meta_ctl(object, pindex + i, 0);
1025 if (blk != blk0 + i)
1026 break;
1027 }
1028 *after = (i - 1);
1029 }
1030 return (TRUE);
1031 }
1032
1033 /*
1034 * SWAP_PAGER_PAGE_UNSWAPPED() - remove swap backing store related to page
1035 *
1036 * This removes any associated swap backing store, whether valid or
1037 * not, from the page.
1038 *
1039 * This routine is typically called when a page is made dirty, at
1040 * which point any associated swap can be freed. MADV_FREE also
1041 * calls us in a special-case situation
1042 *
1043 * NOTE!!! If the page is clean and the swap was valid, the caller
1044 * should make the page dirty before calling this routine. This routine
1045 * does NOT change the m->dirty status of the page. Also: MADV_FREE
1046 * depends on it.
1047 *
1048 * This routine may not block
1049 * This routine must be called at splvm()
1050 */
1051 static void
1052 swap_pager_unswapped(vm_page_t m)
1053 {
1054
1055 VM_OBJECT_LOCK_ASSERT(m->object, MA_OWNED);
1056 swp_pager_meta_ctl(m->object, m->pindex, SWM_FREE);
1057 }
1058
1059 /*
1060 * SWAP_PAGER_GETPAGES() - bring pages in from swap
1061 *
1062 * Attempt to retrieve (m, count) pages from backing store, but make
1063 * sure we retrieve at least m[reqpage]. We try to load in as large
1064 * a chunk surrounding m[reqpage] as is contiguous in swap and which
1065 * belongs to the same object.
1066 *
1067 * The code is designed for asynchronous operation and
1068 * immediate-notification of 'reqpage' but tends not to be
1069 * used that way. Please do not optimize-out this algorithmic
1070 * feature, I intend to improve on it in the future.
1071 *
1072 * The parent has a single vm_object_pip_add() reference prior to
1073 * calling us and we should return with the same.
1074 *
1075 * The parent has BUSY'd the pages. We should return with 'm'
1076 * left busy, but the others adjusted.
1077 */
1078 static int
1079 swap_pager_getpages(vm_object_t object, vm_page_t *m, int count, int reqpage)
1080 {
1081 struct buf *bp;
1082 vm_page_t mreq;
1083 int i;
1084 int j;
1085 daddr_t blk;
1086
1087 mreq = m[reqpage];
1088
1089 KASSERT(mreq->object == object,
1090 ("swap_pager_getpages: object mismatch %p/%p",
1091 object, mreq->object));
1092
1093 /*
1094 * Calculate range to retrieve. The pages have already been assigned
1095 * their swapblks. We require a *contiguous* range but we know it to
1096 * not span devices. If we do not supply it, bad things
1097 * happen. Note that blk, iblk & jblk can be SWAPBLK_NONE, but the
1098 * loops are set up such that the case(s) are handled implicitly.
1099 *
1100 * The swp_*() calls must be made at splvm(). vm_page_free() does
1101 * not need to be, but it will go a little faster if it is.
1102 */
1103 blk = swp_pager_meta_ctl(mreq->object, mreq->pindex, 0);
1104
1105 for (i = reqpage - 1; i >= 0; --i) {
1106 daddr_t iblk;
1107
1108 iblk = swp_pager_meta_ctl(m[i]->object, m[i]->pindex, 0);
1109 if (blk != iblk + (reqpage - i))
1110 break;
1111 }
1112 ++i;
1113
1114 for (j = reqpage + 1; j < count; ++j) {
1115 daddr_t jblk;
1116
1117 jblk = swp_pager_meta_ctl(m[j]->object, m[j]->pindex, 0);
1118 if (blk != jblk - (j - reqpage))
1119 break;
1120 }
1121
1122 /*
1123 * free pages outside our collection range. Note: we never free
1124 * mreq, it must remain busy throughout.
1125 */
1126 if (0 < i || j < count) {
1127 int k;
1128
1129 vm_page_lock_queues();
1130 for (k = 0; k < i; ++k)
1131 vm_page_free(m[k]);
1132 for (k = j; k < count; ++k)
1133 vm_page_free(m[k]);
1134 vm_page_unlock_queues();
1135 }
1136
1137 /*
1138 * Return VM_PAGER_FAIL if we have nothing to do. Return mreq
1139 * still busy, but the others unbusied.
1140 */
1141 if (blk == SWAPBLK_NONE)
1142 return (VM_PAGER_FAIL);
1143
1144 /*
1145 * Getpbuf() can sleep.
1146 */
1147 VM_OBJECT_UNLOCK(object);
1148 /*
1149 * Get a swap buffer header to perform the IO
1150 */
1151 bp = getpbuf(&nsw_rcount);
1152 bp->b_flags |= B_PAGING;
1153
1154 /*
1155 * map our page(s) into kva for input
1156 */
1157 pmap_qenter((vm_offset_t)bp->b_data, m + i, j - i);
1158
1159 bp->b_iocmd = BIO_READ;
1160 bp->b_iodone = swp_pager_async_iodone;
1161 bp->b_rcred = crhold(thread0.td_ucred);
1162 bp->b_wcred = crhold(thread0.td_ucred);
1163 bp->b_blkno = blk - (reqpage - i);
1164 bp->b_bcount = PAGE_SIZE * (j - i);
1165 bp->b_bufsize = PAGE_SIZE * (j - i);
1166 bp->b_pager.pg_reqpage = reqpage - i;
1167
1168 VM_OBJECT_LOCK(object);
1169 {
1170 int k;
1171
1172 for (k = i; k < j; ++k) {
1173 bp->b_pages[k - i] = m[k];
1174 m[k]->oflags |= VPO_SWAPINPROG;
1175 }
1176 }
1177 bp->b_npages = j - i;
1178
1179 PCPU_INC(cnt.v_swapin);
1180 PCPU_ADD(cnt.v_swappgsin, bp->b_npages);
1181
1182 /*
1183 * We still hold the lock on mreq, and our automatic completion routine
1184 * does not remove it.
1185 */
1186 vm_object_pip_add(object, bp->b_npages);
1187 VM_OBJECT_UNLOCK(object);
1188
1189 /*
1190 * perform the I/O. NOTE!!! bp cannot be considered valid after
1191 * this point because we automatically release it on completion.
1192 * Instead, we look at the one page we are interested in which we
1193 * still hold a lock on even through the I/O completion.
1194 *
1195 * The other pages in our m[] array are also released on completion,
1196 * so we cannot assume they are valid anymore either.
1197 *
1198 * NOTE: b_blkno is destroyed by the call to swapdev_strategy
1199 */
1200 BUF_KERNPROC(bp);
1201 swp_pager_strategy(bp);
1202
1203 /*
1204 * wait for the page we want to complete. VPO_SWAPINPROG is always
1205 * cleared on completion. If an I/O error occurs, SWAPBLK_NONE
1206 * is set in the meta-data.
1207 */
1208 VM_OBJECT_LOCK(object);
1209 while ((mreq->oflags & VPO_SWAPINPROG) != 0) {
1210 mreq->oflags |= VPO_WANTED;
1211 vm_page_lock_queues();
1212 vm_page_flag_set(mreq, PG_REFERENCED);
1213 vm_page_unlock_queues();
1214 PCPU_INC(cnt.v_intrans);
1215 if (msleep(mreq, VM_OBJECT_MTX(object), PSWP, "swread", hz*20)) {
1216 printf(
1217 "swap_pager: indefinite wait buffer: bufobj: %p, blkno: %jd, size: %ld\n",
1218 bp->b_bufobj, (intmax_t)bp->b_blkno, bp->b_bcount);
1219 }
1220 }
1221
1222 /*
1223 * mreq is left busied after completion, but all the other pages
1224 * are freed. If we had an unrecoverable read error the page will
1225 * not be valid.
1226 */
1227 if (mreq->valid != VM_PAGE_BITS_ALL) {
1228 return (VM_PAGER_ERROR);
1229 } else {
1230 return (VM_PAGER_OK);
1231 }
1232
1233 /*
1234 * A final note: in a low swap situation, we cannot deallocate swap
1235 * and mark a page dirty here because the caller is likely to mark
1236 * the page clean when we return, causing the page to possibly revert
1237 * to all-zero's later.
1238 */
1239 }
1240
1241 /*
1242 * swap_pager_putpages:
1243 *
1244 * Assign swap (if necessary) and initiate I/O on the specified pages.
1245 *
1246 * We support both OBJT_DEFAULT and OBJT_SWAP objects. DEFAULT objects
1247 * are automatically converted to SWAP objects.
1248 *
1249 * In a low memory situation we may block in VOP_STRATEGY(), but the new
1250 * vm_page reservation system coupled with properly written VFS devices
1251 * should ensure that no low-memory deadlock occurs. This is an area
1252 * which needs work.
1253 *
1254 * The parent has N vm_object_pip_add() references prior to
1255 * calling us and will remove references for rtvals[] that are
1256 * not set to VM_PAGER_PEND. We need to remove the rest on I/O
1257 * completion.
1258 *
1259 * The parent has soft-busy'd the pages it passes us and will unbusy
1260 * those whos rtvals[] entry is not set to VM_PAGER_PEND on return.
1261 * We need to unbusy the rest on I/O completion.
1262 */
1263 void
1264 swap_pager_putpages(vm_object_t object, vm_page_t *m, int count,
1265 boolean_t sync, int *rtvals)
1266 {
1267 int i;
1268 int n = 0;
1269
1270 if (count && m[0]->object != object) {
1271 panic("swap_pager_putpages: object mismatch %p/%p",
1272 object,
1273 m[0]->object
1274 );
1275 }
1276
1277 /*
1278 * Step 1
1279 *
1280 * Turn object into OBJT_SWAP
1281 * check for bogus sysops
1282 * force sync if not pageout process
1283 */
1284 if (object->type != OBJT_SWAP)
1285 swp_pager_meta_build(object, 0, SWAPBLK_NONE);
1286 VM_OBJECT_UNLOCK(object);
1287
1288 if (curproc != pageproc)
1289 sync = TRUE;
1290
1291 /*
1292 * Step 2
1293 *
1294 * Update nsw parameters from swap_async_max sysctl values.
1295 * Do not let the sysop crash the machine with bogus numbers.
1296 */
1297 mtx_lock(&pbuf_mtx);
1298 if (swap_async_max != nsw_wcount_async_max) {
1299 int n;
1300
1301 /*
1302 * limit range
1303 */
1304 if ((n = swap_async_max) > nswbuf / 2)
1305 n = nswbuf / 2;
1306 if (n < 1)
1307 n = 1;
1308 swap_async_max = n;
1309
1310 /*
1311 * Adjust difference ( if possible ). If the current async
1312 * count is too low, we may not be able to make the adjustment
1313 * at this time.
1314 */
1315 n -= nsw_wcount_async_max;
1316 if (nsw_wcount_async + n >= 0) {
1317 nsw_wcount_async += n;
1318 nsw_wcount_async_max += n;
1319 wakeup(&nsw_wcount_async);
1320 }
1321 }
1322 mtx_unlock(&pbuf_mtx);
1323
1324 /*
1325 * Step 3
1326 *
1327 * Assign swap blocks and issue I/O. We reallocate swap on the fly.
1328 * The page is left dirty until the pageout operation completes
1329 * successfully.
1330 */
1331 for (i = 0; i < count; i += n) {
1332 int j;
1333 struct buf *bp;
1334 daddr_t blk;
1335
1336 /*
1337 * Maximum I/O size is limited by a number of factors.
1338 */
1339 n = min(BLIST_MAX_ALLOC, count - i);
1340 n = min(n, nsw_cluster_max);
1341
1342 /*
1343 * Get biggest block of swap we can. If we fail, fall
1344 * back and try to allocate a smaller block. Don't go
1345 * overboard trying to allocate space if it would overly
1346 * fragment swap.
1347 */
1348 while (
1349 (blk = swp_pager_getswapspace(n)) == SWAPBLK_NONE &&
1350 n > 4
1351 ) {
1352 n >>= 1;
1353 }
1354 if (blk == SWAPBLK_NONE) {
1355 for (j = 0; j < n; ++j)
1356 rtvals[i+j] = VM_PAGER_FAIL;
1357 continue;
1358 }
1359
1360 /*
1361 * All I/O parameters have been satisfied, build the I/O
1362 * request and assign the swap space.
1363 */
1364 if (sync == TRUE) {
1365 bp = getpbuf(&nsw_wcount_sync);
1366 } else {
1367 bp = getpbuf(&nsw_wcount_async);
1368 bp->b_flags = B_ASYNC;
1369 }
1370 bp->b_flags |= B_PAGING;
1371 bp->b_iocmd = BIO_WRITE;
1372
1373 pmap_qenter((vm_offset_t)bp->b_data, &m[i], n);
1374
1375 bp->b_rcred = crhold(thread0.td_ucred);
1376 bp->b_wcred = crhold(thread0.td_ucred);
1377 bp->b_bcount = PAGE_SIZE * n;
1378 bp->b_bufsize = PAGE_SIZE * n;
1379 bp->b_blkno = blk;
1380
1381 VM_OBJECT_LOCK(object);
1382 for (j = 0; j < n; ++j) {
1383 vm_page_t mreq = m[i+j];
1384
1385 swp_pager_meta_build(
1386 mreq->object,
1387 mreq->pindex,
1388 blk + j
1389 );
1390 vm_page_dirty(mreq);
1391 rtvals[i+j] = VM_PAGER_OK;
1392
1393 mreq->oflags |= VPO_SWAPINPROG;
1394 bp->b_pages[j] = mreq;
1395 }
1396 VM_OBJECT_UNLOCK(object);
1397 bp->b_npages = n;
1398 /*
1399 * Must set dirty range for NFS to work.
1400 */
1401 bp->b_dirtyoff = 0;
1402 bp->b_dirtyend = bp->b_bcount;
1403
1404 PCPU_INC(cnt.v_swapout);
1405 PCPU_ADD(cnt.v_swappgsout, bp->b_npages);
1406
1407 /*
1408 * asynchronous
1409 *
1410 * NOTE: b_blkno is destroyed by the call to swapdev_strategy
1411 */
1412 if (sync == FALSE) {
1413 bp->b_iodone = swp_pager_async_iodone;
1414 BUF_KERNPROC(bp);
1415 swp_pager_strategy(bp);
1416
1417 for (j = 0; j < n; ++j)
1418 rtvals[i+j] = VM_PAGER_PEND;
1419 /* restart outter loop */
1420 continue;
1421 }
1422
1423 /*
1424 * synchronous
1425 *
1426 * NOTE: b_blkno is destroyed by the call to swapdev_strategy
1427 */
1428 bp->b_iodone = bdone;
1429 swp_pager_strategy(bp);
1430
1431 /*
1432 * Wait for the sync I/O to complete, then update rtvals.
1433 * We just set the rtvals[] to VM_PAGER_PEND so we can call
1434 * our async completion routine at the end, thus avoiding a
1435 * double-free.
1436 */
1437 bwait(bp, PVM, "swwrt");
1438 for (j = 0; j < n; ++j)
1439 rtvals[i+j] = VM_PAGER_PEND;
1440 /*
1441 * Now that we are through with the bp, we can call the
1442 * normal async completion, which frees everything up.
1443 */
1444 swp_pager_async_iodone(bp);
1445 }
1446 VM_OBJECT_LOCK(object);
1447 }
1448
1449 /*
1450 * swp_pager_async_iodone:
1451 *
1452 * Completion routine for asynchronous reads and writes from/to swap.
1453 * Also called manually by synchronous code to finish up a bp.
1454 *
1455 * For READ operations, the pages are PG_BUSY'd. For WRITE operations,
1456 * the pages are vm_page_t->busy'd. For READ operations, we PG_BUSY
1457 * unbusy all pages except the 'main' request page. For WRITE
1458 * operations, we vm_page_t->busy'd unbusy all pages ( we can do this
1459 * because we marked them all VM_PAGER_PEND on return from putpages ).
1460 *
1461 * This routine may not block.
1462 */
1463 static void
1464 swp_pager_async_iodone(struct buf *bp)
1465 {
1466 int i;
1467 vm_object_t object = NULL;
1468
1469 /*
1470 * report error
1471 */
1472 if (bp->b_ioflags & BIO_ERROR) {
1473 printf(
1474 "swap_pager: I/O error - %s failed; blkno %ld,"
1475 "size %ld, error %d\n",
1476 ((bp->b_iocmd == BIO_READ) ? "pagein" : "pageout"),
1477 (long)bp->b_blkno,
1478 (long)bp->b_bcount,
1479 bp->b_error
1480 );
1481 }
1482
1483 /*
1484 * remove the mapping for kernel virtual
1485 */
1486 pmap_qremove((vm_offset_t)bp->b_data, bp->b_npages);
1487
1488 if (bp->b_npages) {
1489 object = bp->b_pages[0]->object;
1490 VM_OBJECT_LOCK(object);
1491 }
1492 vm_page_lock_queues();
1493 /*
1494 * cleanup pages. If an error occurs writing to swap, we are in
1495 * very serious trouble. If it happens to be a disk error, though,
1496 * we may be able to recover by reassigning the swap later on. So
1497 * in this case we remove the m->swapblk assignment for the page
1498 * but do not free it in the rlist. The errornous block(s) are thus
1499 * never reallocated as swap. Redirty the page and continue.
1500 */
1501 for (i = 0; i < bp->b_npages; ++i) {
1502 vm_page_t m = bp->b_pages[i];
1503
1504 m->oflags &= ~VPO_SWAPINPROG;
1505
1506 if (bp->b_ioflags & BIO_ERROR) {
1507 /*
1508 * If an error occurs I'd love to throw the swapblk
1509 * away without freeing it back to swapspace, so it
1510 * can never be used again. But I can't from an
1511 * interrupt.
1512 */
1513 if (bp->b_iocmd == BIO_READ) {
1514 /*
1515 * When reading, reqpage needs to stay
1516 * locked for the parent, but all other
1517 * pages can be freed. We still want to
1518 * wakeup the parent waiting on the page,
1519 * though. ( also: pg_reqpage can be -1 and
1520 * not match anything ).
1521 *
1522 * We have to wake specifically requested pages
1523 * up too because we cleared VPO_SWAPINPROG and
1524 * someone may be waiting for that.
1525 *
1526 * NOTE: for reads, m->dirty will probably
1527 * be overridden by the original caller of
1528 * getpages so don't play cute tricks here.
1529 */
1530 m->valid = 0;
1531 if (i != bp->b_pager.pg_reqpage)
1532 vm_page_free(m);
1533 else
1534 vm_page_flash(m);
1535 /*
1536 * If i == bp->b_pager.pg_reqpage, do not wake
1537 * the page up. The caller needs to.
1538 */
1539 } else {
1540 /*
1541 * If a write error occurs, reactivate page
1542 * so it doesn't clog the inactive list,
1543 * then finish the I/O.
1544 */
1545 vm_page_dirty(m);
1546 vm_page_activate(m);
1547 vm_page_io_finish(m);
1548 }
1549 } else if (bp->b_iocmd == BIO_READ) {
1550 /*
1551 * NOTE: for reads, m->dirty will probably be
1552 * overridden by the original caller of getpages so
1553 * we cannot set them in order to free the underlying
1554 * swap in a low-swap situation. I don't think we'd
1555 * want to do that anyway, but it was an optimization
1556 * that existed in the old swapper for a time before
1557 * it got ripped out due to precisely this problem.
1558 *
1559 * If not the requested page then deactivate it.
1560 *
1561 * Note that the requested page, reqpage, is left
1562 * busied, but we still have to wake it up. The
1563 * other pages are released (unbusied) by
1564 * vm_page_wakeup().
1565 */
1566 KASSERT(!pmap_page_is_mapped(m),
1567 ("swp_pager_async_iodone: page %p is mapped", m));
1568 m->valid = VM_PAGE_BITS_ALL;
1569 KASSERT(m->dirty == 0,
1570 ("swp_pager_async_iodone: page %p is dirty", m));
1571
1572 /*
1573 * We have to wake specifically requested pages
1574 * up too because we cleared VPO_SWAPINPROG and
1575 * could be waiting for it in getpages. However,
1576 * be sure to not unbusy getpages specifically
1577 * requested page - getpages expects it to be
1578 * left busy.
1579 */
1580 if (i != bp->b_pager.pg_reqpage) {
1581 vm_page_deactivate(m);
1582 vm_page_wakeup(m);
1583 } else {
1584 vm_page_flash(m);
1585 }
1586 } else {
1587 /*
1588 * For write success, clear the dirty
1589 * status, then finish the I/O ( which decrements the
1590 * busy count and possibly wakes waiter's up ).
1591 */
1592 KASSERT((m->flags & PG_WRITEABLE) == 0,
1593 ("swp_pager_async_iodone: page %p is not write"
1594 " protected", m));
1595 vm_page_undirty(m);
1596 vm_page_io_finish(m);
1597 if (vm_page_count_severe())
1598 vm_page_try_to_cache(m);
1599 }
1600 }
1601 vm_page_unlock_queues();
1602
1603 /*
1604 * adjust pip. NOTE: the original parent may still have its own
1605 * pip refs on the object.
1606 */
1607 if (object != NULL) {
1608 vm_object_pip_wakeupn(object, bp->b_npages);
1609 VM_OBJECT_UNLOCK(object);
1610 }
1611
1612 /*
1613 * swapdev_strategy() manually sets b_vp and b_bufobj before calling
1614 * bstrategy(). Set them back to NULL now we're done with it, or we'll
1615 * trigger a KASSERT in relpbuf().
1616 */
1617 if (bp->b_vp) {
1618 bp->b_vp = NULL;
1619 bp->b_bufobj = NULL;
1620 }
1621 /*
1622 * release the physical I/O buffer
1623 */
1624 relpbuf(
1625 bp,
1626 ((bp->b_iocmd == BIO_READ) ? &nsw_rcount :
1627 ((bp->b_flags & B_ASYNC) ?
1628 &nsw_wcount_async :
1629 &nsw_wcount_sync
1630 )
1631 )
1632 );
1633 }
1634
1635 /*
1636 * swap_pager_isswapped:
1637 *
1638 * Return 1 if at least one page in the given object is paged
1639 * out to the given swap device.
1640 *
1641 * This routine may not block.
1642 */
1643 int
1644 swap_pager_isswapped(vm_object_t object, struct swdevt *sp)
1645 {
1646 daddr_t index = 0;
1647 int bcount;
1648 int i;
1649
1650 VM_OBJECT_LOCK_ASSERT(object, MA_OWNED);
1651 if (object->type != OBJT_SWAP)
1652 return (0);
1653
1654 mtx_lock(&swhash_mtx);
1655 for (bcount = 0; bcount < object->un_pager.swp.swp_bcount; bcount++) {
1656 struct swblock *swap;
1657
1658 if ((swap = *swp_pager_hash(object, index)) != NULL) {
1659 for (i = 0; i < SWAP_META_PAGES; ++i) {
1660 if (swp_pager_isondev(swap->swb_pages[i], sp)) {
1661 mtx_unlock(&swhash_mtx);
1662 return (1);
1663 }
1664 }
1665 }
1666 index += SWAP_META_PAGES;
1667 if (index > 0x20000000)
1668 panic("swap_pager_isswapped: failed to locate all swap meta blocks");
1669 }
1670 mtx_unlock(&swhash_mtx);
1671 return (0);
1672 }
1673
1674 /*
1675 * SWP_PAGER_FORCE_PAGEIN() - force a swap block to be paged in
1676 *
1677 * This routine dissociates the page at the given index within a
1678 * swap block from its backing store, paging it in if necessary.
1679 * If the page is paged in, it is placed in the inactive queue,
1680 * since it had its backing store ripped out from under it.
1681 * We also attempt to swap in all other pages in the swap block,
1682 * we only guarantee that the one at the specified index is
1683 * paged in.
1684 *
1685 * XXX - The code to page the whole block in doesn't work, so we
1686 * revert to the one-by-one behavior for now. Sigh.
1687 */
1688 static inline void
1689 swp_pager_force_pagein(vm_object_t object, vm_pindex_t pindex)
1690 {
1691 vm_page_t m;
1692
1693 vm_object_pip_add(object, 1);
1694 m = vm_page_grab(object, pindex, VM_ALLOC_NORMAL|VM_ALLOC_RETRY);
1695 if (m->valid == VM_PAGE_BITS_ALL) {
1696 vm_object_pip_subtract(object, 1);
1697 vm_page_lock_queues();
1698 vm_page_activate(m);
1699 vm_page_dirty(m);
1700 vm_page_unlock_queues();
1701 vm_page_wakeup(m);
1702 vm_pager_page_unswapped(m);
1703 return;
1704 }
1705
1706 if (swap_pager_getpages(object, &m, 1, 0) != VM_PAGER_OK)
1707 panic("swap_pager_force_pagein: read from swap failed");/*XXX*/
1708 vm_object_pip_subtract(object, 1);
1709 vm_page_lock_queues();
1710 vm_page_dirty(m);
1711 vm_page_dontneed(m);
1712 vm_page_unlock_queues();
1713 vm_page_wakeup(m);
1714 vm_pager_page_unswapped(m);
1715 }
1716
1717 /*
1718 * swap_pager_swapoff:
1719 *
1720 * Page in all of the pages that have been paged out to the
1721 * given device. The corresponding blocks in the bitmap must be
1722 * marked as allocated and the device must be flagged SW_CLOSING.
1723 * There may be no processes swapped out to the device.
1724 *
1725 * This routine may block.
1726 */
1727 static void
1728 swap_pager_swapoff(struct swdevt *sp)
1729 {
1730 struct swblock *swap;
1731 int i, j, retries;
1732
1733 GIANT_REQUIRED;
1734
1735 retries = 0;
1736 full_rescan:
1737 mtx_lock(&swhash_mtx);
1738 for (i = 0; i <= swhash_mask; i++) { /* '<=' is correct here */
1739 restart:
1740 for (swap = swhash[i]; swap != NULL; swap = swap->swb_hnext) {
1741 vm_object_t object = swap->swb_object;
1742 vm_pindex_t pindex = swap->swb_index;
1743 for (j = 0; j < SWAP_META_PAGES; ++j) {
1744 if (swp_pager_isondev(swap->swb_pages[j], sp)) {
1745 /* avoid deadlock */
1746 if (!VM_OBJECT_TRYLOCK(object)) {
1747 break;
1748 } else {
1749 mtx_unlock(&swhash_mtx);
1750 swp_pager_force_pagein(object,
1751 pindex + j);
1752 VM_OBJECT_UNLOCK(object);
1753 mtx_lock(&swhash_mtx);
1754 goto restart;
1755 }
1756 }
1757 }
1758 }
1759 }
1760 mtx_unlock(&swhash_mtx);
1761 if (sp->sw_used) {
1762 /*
1763 * Objects may be locked or paging to the device being
1764 * removed, so we will miss their pages and need to
1765 * make another pass. We have marked this device as
1766 * SW_CLOSING, so the activity should finish soon.
1767 */
1768 retries++;
1769 if (retries > 100) {
1770 panic("swapoff: failed to locate %d swap blocks",
1771 sp->sw_used);
1772 }
1773 pause("swpoff", hz / 20);
1774 goto full_rescan;
1775 }
1776 }
1777
1778 /************************************************************************
1779 * SWAP META DATA *
1780 ************************************************************************
1781 *
1782 * These routines manipulate the swap metadata stored in the
1783 * OBJT_SWAP object. All swp_*() routines must be called at
1784 * splvm() because swap can be freed up by the low level vm_page
1785 * code which might be called from interrupts beyond what splbio() covers.
1786 *
1787 * Swap metadata is implemented with a global hash and not directly
1788 * linked into the object. Instead the object simply contains
1789 * appropriate tracking counters.
1790 */
1791
1792 /*
1793 * SWP_PAGER_META_BUILD() - add swap block to swap meta data for object
1794 *
1795 * We first convert the object to a swap object if it is a default
1796 * object.
1797 *
1798 * The specified swapblk is added to the object's swap metadata. If
1799 * the swapblk is not valid, it is freed instead. Any previously
1800 * assigned swapblk is freed.
1801 *
1802 * This routine must be called at splvm(), except when used to convert
1803 * an OBJT_DEFAULT object into an OBJT_SWAP object.
1804 */
1805 static void
1806 swp_pager_meta_build(vm_object_t object, vm_pindex_t pindex, daddr_t swapblk)
1807 {
1808 struct swblock *swap;
1809 struct swblock **pswap;
1810 int idx;
1811
1812 VM_OBJECT_LOCK_ASSERT(object, MA_OWNED);
1813 /*
1814 * Convert default object to swap object if necessary
1815 */
1816 if (object->type != OBJT_SWAP) {
1817 object->type = OBJT_SWAP;
1818 object->un_pager.swp.swp_bcount = 0;
1819
1820 if (object->handle != NULL) {
1821 mtx_lock(&sw_alloc_mtx);
1822 TAILQ_INSERT_TAIL(
1823 NOBJLIST(object->handle),
1824 object,
1825 pager_object_list
1826 );
1827 mtx_unlock(&sw_alloc_mtx);
1828 }
1829 }
1830
1831 /*
1832 * Locate hash entry. If not found create, but if we aren't adding
1833 * anything just return. If we run out of space in the map we wait
1834 * and, since the hash table may have changed, retry.
1835 */
1836 retry:
1837 mtx_lock(&swhash_mtx);
1838 pswap = swp_pager_hash(object, pindex);
1839
1840 if ((swap = *pswap) == NULL) {
1841 int i;
1842
1843 if (swapblk == SWAPBLK_NONE)
1844 goto done;
1845
1846 swap = *pswap = uma_zalloc(swap_zone, M_NOWAIT);
1847 if (swap == NULL) {
1848 mtx_unlock(&swhash_mtx);
1849 VM_OBJECT_UNLOCK(object);
1850 if (uma_zone_exhausted(swap_zone)) {
1851 printf("swap zone exhausted, increase kern.maxswzone\n");
1852 vm_pageout_oom(VM_OOM_SWAPZ);
1853 pause("swzonex", 10);
1854 } else
1855 VM_WAIT;
1856 VM_OBJECT_LOCK(object);
1857 goto retry;
1858 }
1859
1860 swap->swb_hnext = NULL;
1861 swap->swb_object = object;
1862 swap->swb_index = pindex & ~(vm_pindex_t)SWAP_META_MASK;
1863 swap->swb_count = 0;
1864
1865 ++object->un_pager.swp.swp_bcount;
1866
1867 for (i = 0; i < SWAP_META_PAGES; ++i)
1868 swap->swb_pages[i] = SWAPBLK_NONE;
1869 }
1870
1871 /*
1872 * Delete prior contents of metadata
1873 */
1874 idx = pindex & SWAP_META_MASK;
1875
1876 if (swap->swb_pages[idx] != SWAPBLK_NONE) {
1877 swp_pager_freeswapspace(swap->swb_pages[idx], 1);
1878 --swap->swb_count;
1879 }
1880
1881 /*
1882 * Enter block into metadata
1883 */
1884 swap->swb_pages[idx] = swapblk;
1885 if (swapblk != SWAPBLK_NONE)
1886 ++swap->swb_count;
1887 done:
1888 mtx_unlock(&swhash_mtx);
1889 }
1890
1891 /*
1892 * SWP_PAGER_META_FREE() - free a range of blocks in the object's swap metadata
1893 *
1894 * The requested range of blocks is freed, with any associated swap
1895 * returned to the swap bitmap.
1896 *
1897 * This routine will free swap metadata structures as they are cleaned
1898 * out. This routine does *NOT* operate on swap metadata associated
1899 * with resident pages.
1900 *
1901 * This routine must be called at splvm()
1902 */
1903 static void
1904 swp_pager_meta_free(vm_object_t object, vm_pindex_t index, daddr_t count)
1905 {
1906
1907 VM_OBJECT_LOCK_ASSERT(object, MA_OWNED);
1908 if (object->type != OBJT_SWAP)
1909 return;
1910
1911 while (count > 0) {
1912 struct swblock **pswap;
1913 struct swblock *swap;
1914
1915 mtx_lock(&swhash_mtx);
1916 pswap = swp_pager_hash(object, index);
1917
1918 if ((swap = *pswap) != NULL) {
1919 daddr_t v = swap->swb_pages[index & SWAP_META_MASK];
1920
1921 if (v != SWAPBLK_NONE) {
1922 swp_pager_freeswapspace(v, 1);
1923 swap->swb_pages[index & SWAP_META_MASK] =
1924 SWAPBLK_NONE;
1925 if (--swap->swb_count == 0) {
1926 *pswap = swap->swb_hnext;
1927 uma_zfree(swap_zone, swap);
1928 --object->un_pager.swp.swp_bcount;
1929 }
1930 }
1931 --count;
1932 ++index;
1933 } else {
1934 int n = SWAP_META_PAGES - (index & SWAP_META_MASK);
1935 count -= n;
1936 index += n;
1937 }
1938 mtx_unlock(&swhash_mtx);
1939 }
1940 }
1941
1942 /*
1943 * SWP_PAGER_META_FREE_ALL() - destroy all swap metadata associated with object
1944 *
1945 * This routine locates and destroys all swap metadata associated with
1946 * an object.
1947 *
1948 * This routine must be called at splvm()
1949 */
1950 static void
1951 swp_pager_meta_free_all(vm_object_t object)
1952 {
1953 daddr_t index = 0;
1954
1955 VM_OBJECT_LOCK_ASSERT(object, MA_OWNED);
1956 if (object->type != OBJT_SWAP)
1957 return;
1958
1959 while (object->un_pager.swp.swp_bcount) {
1960 struct swblock **pswap;
1961 struct swblock *swap;
1962
1963 mtx_lock(&swhash_mtx);
1964 pswap = swp_pager_hash(object, index);
1965 if ((swap = *pswap) != NULL) {
1966 int i;
1967
1968 for (i = 0; i < SWAP_META_PAGES; ++i) {
1969 daddr_t v = swap->swb_pages[i];
1970 if (v != SWAPBLK_NONE) {
1971 --swap->swb_count;
1972 swp_pager_freeswapspace(v, 1);
1973 }
1974 }
1975 if (swap->swb_count != 0)
1976 panic("swap_pager_meta_free_all: swb_count != 0");
1977 *pswap = swap->swb_hnext;
1978 uma_zfree(swap_zone, swap);
1979 --object->un_pager.swp.swp_bcount;
1980 }
1981 mtx_unlock(&swhash_mtx);
1982 index += SWAP_META_PAGES;
1983 if (index > 0x20000000)
1984 panic("swp_pager_meta_free_all: failed to locate all swap meta blocks");
1985 }
1986 }
1987
1988 /*
1989 * SWP_PAGER_METACTL() - misc control of swap and vm_page_t meta data.
1990 *
1991 * This routine is capable of looking up, popping, or freeing
1992 * swapblk assignments in the swap meta data or in the vm_page_t.
1993 * The routine typically returns the swapblk being looked-up, or popped,
1994 * or SWAPBLK_NONE if the block was freed, or SWAPBLK_NONE if the block
1995 * was invalid. This routine will automatically free any invalid
1996 * meta-data swapblks.
1997 *
1998 * It is not possible to store invalid swapblks in the swap meta data
1999 * (other then a literal 'SWAPBLK_NONE'), so we don't bother checking.
2000 *
2001 * When acting on a busy resident page and paging is in progress, we
2002 * have to wait until paging is complete but otherwise can act on the
2003 * busy page.
2004 *
2005 * This routine must be called at splvm().
2006 *
2007 * SWM_FREE remove and free swap block from metadata
2008 * SWM_POP remove from meta data but do not free.. pop it out
2009 */
2010 static daddr_t
2011 swp_pager_meta_ctl(vm_object_t object, vm_pindex_t pindex, int flags)
2012 {
2013 struct swblock **pswap;
2014 struct swblock *swap;
2015 daddr_t r1;
2016 int idx;
2017
2018 VM_OBJECT_LOCK_ASSERT(object, MA_OWNED);
2019 /*
2020 * The meta data only exists of the object is OBJT_SWAP
2021 * and even then might not be allocated yet.
2022 */
2023 if (object->type != OBJT_SWAP)
2024 return (SWAPBLK_NONE);
2025
2026 r1 = SWAPBLK_NONE;
2027 mtx_lock(&swhash_mtx);
2028 pswap = swp_pager_hash(object, pindex);
2029
2030 if ((swap = *pswap) != NULL) {
2031 idx = pindex & SWAP_META_MASK;
2032 r1 = swap->swb_pages[idx];
2033
2034 if (r1 != SWAPBLK_NONE) {
2035 if (flags & SWM_FREE) {
2036 swp_pager_freeswapspace(r1, 1);
2037 r1 = SWAPBLK_NONE;
2038 }
2039 if (flags & (SWM_FREE|SWM_POP)) {
2040 swap->swb_pages[idx] = SWAPBLK_NONE;
2041 if (--swap->swb_count == 0) {
2042 *pswap = swap->swb_hnext;
2043 uma_zfree(swap_zone, swap);
2044 --object->un_pager.swp.swp_bcount;
2045 }
2046 }
2047 }
2048 }
2049 mtx_unlock(&swhash_mtx);
2050 return (r1);
2051 }
2052
2053 /*
2054 * System call swapon(name) enables swapping on device name,
2055 * which must be in the swdevsw. Return EBUSY
2056 * if already swapping on this device.
2057 */
2058 #ifndef _SYS_SYSPROTO_H_
2059 struct swapon_args {
2060 char *name;
2061 };
2062 #endif
2063
2064 /*
2065 * MPSAFE
2066 */
2067 /* ARGSUSED */
2068 int
2069 swapon(struct thread *td, struct swapon_args *uap)
2070 {
2071 struct vattr attr;
2072 struct vnode *vp;
2073 struct nameidata nd;
2074 int error;
2075
2076 error = priv_check(td, PRIV_SWAPON);
2077 if (error)
2078 return (error);
2079
2080 mtx_lock(&Giant);
2081 while (swdev_syscall_active)
2082 tsleep(&swdev_syscall_active, PUSER - 1, "swpon", 0);
2083 swdev_syscall_active = 1;
2084
2085 /*
2086 * Swap metadata may not fit in the KVM if we have physical
2087 * memory of >1GB.
2088 */
2089 if (swap_zone == NULL) {
2090 error = ENOMEM;
2091 goto done;
2092 }
2093
2094 NDINIT(&nd, LOOKUP, ISOPEN | FOLLOW | AUDITVNODE1, UIO_USERSPACE,
2095 uap->name, td);
2096 error = namei(&nd);
2097 if (error)
2098 goto done;
2099
2100 NDFREE(&nd, NDF_ONLY_PNBUF);
2101 vp = nd.ni_vp;
2102
2103 if (vn_isdisk(vp, &error)) {
2104 error = swapongeom(td, vp);
2105 } else if (vp->v_type == VREG &&
2106 (vp->v_mount->mnt_vfc->vfc_flags & VFCF_NETWORK) != 0 &&
2107 (error = VOP_GETATTR(vp, &attr, td->td_ucred)) == 0) {
2108 /*
2109 * Allow direct swapping to NFS regular files in the same
2110 * way that nfs_mountroot() sets up diskless swapping.
2111 */
2112 error = swaponvp(td, vp, attr.va_size / DEV_BSIZE);
2113 }
2114
2115 if (error)
2116 vrele(vp);
2117 done:
2118 swdev_syscall_active = 0;
2119 wakeup_one(&swdev_syscall_active);
2120 mtx_unlock(&Giant);
2121 return (error);
2122 }
2123
2124 static void
2125 swaponsomething(struct vnode *vp, void *id, u_long nblks, sw_strategy_t *strategy, sw_close_t *close, dev_t dev)
2126 {
2127 struct swdevt *sp, *tsp;
2128 swblk_t dvbase;
2129 u_long mblocks;
2130
2131 /*
2132 * If we go beyond this, we get overflows in the radix
2133 * tree bitmap code.
2134 */
2135 mblocks = 0x40000000 / BLIST_META_RADIX;
2136 if (nblks > mblocks) {
2137 printf("WARNING: reducing size to maximum of %lu blocks per swap unit\n",
2138 mblocks);
2139 nblks = mblocks;
2140 }
2141 /*
2142 * nblks is in DEV_BSIZE'd chunks, convert to PAGE_SIZE'd chunks.
2143 * First chop nblks off to page-align it, then convert.
2144 *
2145 * sw->sw_nblks is in page-sized chunks now too.
2146 */
2147 nblks &= ~(ctodb(1) - 1);
2148 nblks = dbtoc(nblks);
2149
2150 sp = malloc(sizeof *sp, M_VMPGDATA, M_WAITOK | M_ZERO);
2151 sp->sw_vp = vp;
2152 sp->sw_id = id;
2153 sp->sw_dev = dev;
2154 sp->sw_flags = 0;
2155 sp->sw_nblks = nblks;
2156 sp->sw_used = 0;
2157 sp->sw_strategy = strategy;
2158 sp->sw_close = close;
2159
2160 sp->sw_blist = blist_create(nblks, M_WAITOK);
2161 /*
2162 * Do not free the first two block in order to avoid overwriting
2163 * any bsd label at the front of the partition
2164 */
2165 blist_free(sp->sw_blist, 2, nblks - 2);
2166
2167 dvbase = 0;
2168 mtx_lock(&sw_dev_mtx);
2169 TAILQ_FOREACH(tsp, &swtailq, sw_list) {
2170 if (tsp->sw_end >= dvbase) {
2171 /*
2172 * We put one uncovered page between the devices
2173 * in order to definitively prevent any cross-device
2174 * I/O requests
2175 */
2176 dvbase = tsp->sw_end + 1;
2177 }
2178 }
2179 sp->sw_first = dvbase;
2180 sp->sw_end = dvbase + nblks;
2181 TAILQ_INSERT_TAIL(&swtailq, sp, sw_list);
2182 nswapdev++;
2183 swap_pager_avail += nblks;
2184 swap_total += (vm_ooffset_t)nblks * PAGE_SIZE;
2185 swp_sizecheck();
2186 mtx_unlock(&sw_dev_mtx);
2187 }
2188
2189 /*
2190 * SYSCALL: swapoff(devname)
2191 *
2192 * Disable swapping on the given device.
2193 *
2194 * XXX: Badly designed system call: it should use a device index
2195 * rather than filename as specification. We keep sw_vp around
2196 * only to make this work.
2197 */
2198 #ifndef _SYS_SYSPROTO_H_
2199 struct swapoff_args {
2200 char *name;
2201 };
2202 #endif
2203
2204 /*
2205 * MPSAFE
2206 */
2207 /* ARGSUSED */
2208 int
2209 swapoff(struct thread *td, struct swapoff_args *uap)
2210 {
2211 struct vnode *vp;
2212 struct nameidata nd;
2213 struct swdevt *sp;
2214 int error;
2215
2216 error = priv_check(td, PRIV_SWAPOFF);
2217 if (error)
2218 return (error);
2219
2220 mtx_lock(&Giant);
2221 while (swdev_syscall_active)
2222 tsleep(&swdev_syscall_active, PUSER - 1, "swpoff", 0);
2223 swdev_syscall_active = 1;
2224
2225 NDINIT(&nd, LOOKUP, FOLLOW | AUDITVNODE1, UIO_USERSPACE, uap->name,
2226 td);
2227 error = namei(&nd);
2228 if (error)
2229 goto done;
2230 NDFREE(&nd, NDF_ONLY_PNBUF);
2231 vp = nd.ni_vp;
2232
2233 mtx_lock(&sw_dev_mtx);
2234 TAILQ_FOREACH(sp, &swtailq, sw_list) {
2235 if (sp->sw_vp == vp)
2236 break;
2237 }
2238 mtx_unlock(&sw_dev_mtx);
2239 if (sp == NULL) {
2240 error = EINVAL;
2241 goto done;
2242 }
2243 error = swapoff_one(sp, td->td_ucred);
2244 done:
2245 swdev_syscall_active = 0;
2246 wakeup_one(&swdev_syscall_active);
2247 mtx_unlock(&Giant);
2248 return (error);
2249 }
2250
2251 static int
2252 swapoff_one(struct swdevt *sp, struct ucred *cred)
2253 {
2254 u_long nblks, dvbase;
2255 #ifdef MAC
2256 int error;
2257 #endif
2258
2259 mtx_assert(&Giant, MA_OWNED);
2260 #ifdef MAC
2261 (void) vn_lock(sp->sw_vp, LK_EXCLUSIVE | LK_RETRY);
2262 error = mac_system_check_swapoff(cred, sp->sw_vp);
2263 (void) VOP_UNLOCK(sp->sw_vp, 0);
2264 if (error != 0)
2265 return (error);
2266 #endif
2267 nblks = sp->sw_nblks;
2268
2269 /*
2270 * We can turn off this swap device safely only if the
2271 * available virtual memory in the system will fit the amount
2272 * of data we will have to page back in, plus an epsilon so
2273 * the system doesn't become critically low on swap space.
2274 */
2275 if (cnt.v_free_count + cnt.v_cache_count + swap_pager_avail <
2276 nblks + nswap_lowat) {
2277 return (ENOMEM);
2278 }
2279
2280 /*
2281 * Prevent further allocations on this device.
2282 */
2283 mtx_lock(&sw_dev_mtx);
2284 sp->sw_flags |= SW_CLOSING;
2285 for (dvbase = 0; dvbase < sp->sw_end; dvbase += dmmax) {
2286 swap_pager_avail -= blist_fill(sp->sw_blist,
2287 dvbase, dmmax);
2288 }
2289 swap_total -= (vm_ooffset_t)nblks * PAGE_SIZE;
2290 mtx_unlock(&sw_dev_mtx);
2291
2292 /*
2293 * Page in the contents of the device and close it.
2294 */
2295 swap_pager_swapoff(sp);
2296
2297 sp->sw_close(curthread, sp);
2298 sp->sw_id = NULL;
2299 mtx_lock(&sw_dev_mtx);
2300 TAILQ_REMOVE(&swtailq, sp, sw_list);
2301 nswapdev--;
2302 if (nswapdev == 0) {
2303 swap_pager_full = 2;
2304 swap_pager_almost_full = 1;
2305 }
2306 if (swdevhd == sp)
2307 swdevhd = NULL;
2308 mtx_unlock(&sw_dev_mtx);
2309 blist_destroy(sp->sw_blist);
2310 free(sp, M_VMPGDATA);
2311 return (0);
2312 }
2313
2314 void
2315 swapoff_all(void)
2316 {
2317 struct swdevt *sp, *spt;
2318 const char *devname;
2319 int error;
2320
2321 mtx_lock(&Giant);
2322 while (swdev_syscall_active)
2323 tsleep(&swdev_syscall_active, PUSER - 1, "swpoff", 0);
2324 swdev_syscall_active = 1;
2325
2326 mtx_lock(&sw_dev_mtx);
2327 TAILQ_FOREACH_SAFE(sp, &swtailq, sw_list, spt) {
2328 mtx_unlock(&sw_dev_mtx);
2329 if (vn_isdisk(sp->sw_vp, NULL))
2330 devname = sp->sw_vp->v_rdev->si_name;
2331 else
2332 devname = "[file]";
2333 error = swapoff_one(sp, thread0.td_ucred);
2334 if (error != 0) {
2335 printf("Cannot remove swap device %s (error=%d), "
2336 "skipping.\n", devname, error);
2337 } else if (bootverbose) {
2338 printf("Swap device %s removed.\n", devname);
2339 }
2340 mtx_lock(&sw_dev_mtx);
2341 }
2342 mtx_unlock(&sw_dev_mtx);
2343
2344 swdev_syscall_active = 0;
2345 wakeup_one(&swdev_syscall_active);
2346 mtx_unlock(&Giant);
2347 }
2348
2349 void
2350 swap_pager_status(int *total, int *used)
2351 {
2352 struct swdevt *sp;
2353
2354 *total = 0;
2355 *used = 0;
2356 mtx_lock(&sw_dev_mtx);
2357 TAILQ_FOREACH(sp, &swtailq, sw_list) {
2358 *total += sp->sw_nblks;
2359 *used += sp->sw_used;
2360 }
2361 mtx_unlock(&sw_dev_mtx);
2362 }
2363
2364 static int
2365 sysctl_vm_swap_info(SYSCTL_HANDLER_ARGS)
2366 {
2367 int *name = (int *)arg1;
2368 int error, n;
2369 struct xswdev xs;
2370 struct swdevt *sp;
2371
2372 if (arg2 != 1) /* name length */
2373 return (EINVAL);
2374
2375 n = 0;
2376 mtx_lock(&sw_dev_mtx);
2377 TAILQ_FOREACH(sp, &swtailq, sw_list) {
2378 if (n == *name) {
2379 mtx_unlock(&sw_dev_mtx);
2380 xs.xsw_version = XSWDEV_VERSION;
2381 xs.xsw_dev = sp->sw_dev;
2382 xs.xsw_flags = sp->sw_flags;
2383 xs.xsw_nblks = sp->sw_nblks;
2384 xs.xsw_used = sp->sw_used;
2385
2386 error = SYSCTL_OUT(req, &xs, sizeof(xs));
2387 return (error);
2388 }
2389 n++;
2390 }
2391 mtx_unlock(&sw_dev_mtx);
2392 return (ENOENT);
2393 }
2394
2395 SYSCTL_INT(_vm, OID_AUTO, nswapdev, CTLFLAG_RD, &nswapdev, 0,
2396 "Number of swap devices");
2397 SYSCTL_NODE(_vm, OID_AUTO, swap_info, CTLFLAG_RD, sysctl_vm_swap_info,
2398 "Swap statistics by device");
2399
2400 /*
2401 * vmspace_swap_count() - count the approximate swap usage in pages for a
2402 * vmspace.
2403 *
2404 * The map must be locked.
2405 *
2406 * Swap usage is determined by taking the proportional swap used by
2407 * VM objects backing the VM map. To make up for fractional losses,
2408 * if the VM object has any swap use at all the associated map entries
2409 * count for at least 1 swap page.
2410 */
2411 int
2412 vmspace_swap_count(struct vmspace *vmspace)
2413 {
2414 vm_map_t map = &vmspace->vm_map;
2415 vm_map_entry_t cur;
2416 int count = 0;
2417
2418 for (cur = map->header.next; cur != &map->header; cur = cur->next) {
2419 vm_object_t object;
2420
2421 if ((cur->eflags & MAP_ENTRY_IS_SUB_MAP) == 0 &&
2422 (object = cur->object.vm_object) != NULL) {
2423 VM_OBJECT_LOCK(object);
2424 if (object->type == OBJT_SWAP &&
2425 object->un_pager.swp.swp_bcount != 0) {
2426 int n = (cur->end - cur->start) / PAGE_SIZE;
2427
2428 count += object->un_pager.swp.swp_bcount *
2429 SWAP_META_PAGES * n / object->size + 1;
2430 }
2431 VM_OBJECT_UNLOCK(object);
2432 }
2433 }
2434 return (count);
2435 }
2436
2437 /*
2438 * GEOM backend
2439 *
2440 * Swapping onto disk devices.
2441 *
2442 */
2443
2444 static g_orphan_t swapgeom_orphan;
2445
2446 static struct g_class g_swap_class = {
2447 .name = "SWAP",
2448 .version = G_VERSION,
2449 .orphan = swapgeom_orphan,
2450 };
2451
2452 DECLARE_GEOM_CLASS(g_swap_class, g_class);
2453
2454
2455 static void
2456 swapgeom_done(struct bio *bp2)
2457 {
2458 struct buf *bp;
2459
2460 bp = bp2->bio_caller2;
2461 bp->b_ioflags = bp2->bio_flags;
2462 if (bp2->bio_error)
2463 bp->b_ioflags |= BIO_ERROR;
2464 bp->b_resid = bp->b_bcount - bp2->bio_completed;
2465 bp->b_error = bp2->bio_error;
2466 bufdone(bp);
2467 g_destroy_bio(bp2);
2468 }
2469
2470 static void
2471 swapgeom_strategy(struct buf *bp, struct swdevt *sp)
2472 {
2473 struct bio *bio;
2474 struct g_consumer *cp;
2475
2476 cp = sp->sw_id;
2477 if (cp == NULL) {
2478 bp->b_error = ENXIO;
2479 bp->b_ioflags |= BIO_ERROR;
2480 bufdone(bp);
2481 return;
2482 }
2483 if (bp->b_iocmd == BIO_WRITE)
2484 bio = g_new_bio();
2485 else
2486 bio = g_alloc_bio();
2487 if (bio == NULL) {
2488 bp->b_error = ENOMEM;
2489 bp->b_ioflags |= BIO_ERROR;
2490 bufdone(bp);
2491 return;
2492 }
2493
2494 bio->bio_caller2 = bp;
2495 bio->bio_cmd = bp->b_iocmd;
2496 bio->bio_data = bp->b_data;
2497 bio->bio_offset = (bp->b_blkno - sp->sw_first) * PAGE_SIZE;
2498 bio->bio_length = bp->b_bcount;
2499 bio->bio_done = swapgeom_done;
2500 g_io_request(bio, cp);
2501 return;
2502 }
2503
2504 static void
2505 swapgeom_orphan(struct g_consumer *cp)
2506 {
2507 struct swdevt *sp;
2508
2509 mtx_lock(&sw_dev_mtx);
2510 TAILQ_FOREACH(sp, &swtailq, sw_list)
2511 if (sp->sw_id == cp)
2512 sp->sw_id = NULL;
2513 mtx_unlock(&sw_dev_mtx);
2514 }
2515
2516 static void
2517 swapgeom_close_ev(void *arg, int flags)
2518 {
2519 struct g_consumer *cp;
2520
2521 cp = arg;
2522 g_access(cp, -1, -1, 0);
2523 g_detach(cp);
2524 g_destroy_consumer(cp);
2525 }
2526
2527 static void
2528 swapgeom_close(struct thread *td, struct swdevt *sw)
2529 {
2530
2531 /* XXX: direct call when Giant untangled */
2532 g_waitfor_event(swapgeom_close_ev, sw->sw_id, M_WAITOK, NULL);
2533 }
2534
2535
2536 struct swh0h0 {
2537 struct cdev *dev;
2538 struct vnode *vp;
2539 int error;
2540 };
2541
2542 static void
2543 swapongeom_ev(void *arg, int flags)
2544 {
2545 struct swh0h0 *swh;
2546 struct g_provider *pp;
2547 struct g_consumer *cp;
2548 static struct g_geom *gp;
2549 struct swdevt *sp;
2550 u_long nblks;
2551 int error;
2552
2553 swh = arg;
2554 swh->error = 0;
2555 pp = g_dev_getprovider(swh->dev);
2556 if (pp == NULL) {
2557 swh->error = ENODEV;
2558 return;
2559 }
2560 mtx_lock(&sw_dev_mtx);
2561 TAILQ_FOREACH(sp, &swtailq, sw_list) {
2562 cp = sp->sw_id;
2563 if (cp != NULL && cp->provider == pp) {
2564 mtx_unlock(&sw_dev_mtx);
2565 swh->error = EBUSY;
2566 return;
2567 }
2568 }
2569 mtx_unlock(&sw_dev_mtx);
2570 if (gp == NULL)
2571 gp = g_new_geomf(&g_swap_class, "swap", NULL);
2572 cp = g_new_consumer(gp);
2573 g_attach(cp, pp);
2574 /*
2575 * XXX: Everytime you think you can improve the margin for
2576 * footshooting, somebody depends on the ability to do so:
2577 * savecore(8) wants to write to our swapdev so we cannot
2578 * set an exclusive count :-(
2579 */
2580 error = g_access(cp, 1, 1, 0);
2581 if (error) {
2582 g_detach(cp);
2583 g_destroy_consumer(cp);
2584 swh->error = error;
2585 return;
2586 }
2587 nblks = pp->mediasize / DEV_BSIZE;
2588 swaponsomething(swh->vp, cp, nblks, swapgeom_strategy,
2589 swapgeom_close, dev2udev(swh->dev));
2590 swh->error = 0;
2591 return;
2592 }
2593
2594 static int
2595 swapongeom(struct thread *td, struct vnode *vp)
2596 {
2597 int error;
2598 struct swh0h0 swh;
2599
2600 vn_lock(vp, LK_EXCLUSIVE | LK_RETRY);
2601
2602 swh.dev = vp->v_rdev;
2603 swh.vp = vp;
2604 swh.error = 0;
2605 /* XXX: direct call when Giant untangled */
2606 error = g_waitfor_event(swapongeom_ev, &swh, M_WAITOK, NULL);
2607 if (!error)
2608 error = swh.error;
2609 VOP_UNLOCK(vp, 0);
2610 return (error);
2611 }
2612
2613 /*
2614 * VNODE backend
2615 *
2616 * This is used mainly for network filesystem (read: probably only tested
2617 * with NFS) swapfiles.
2618 *
2619 */
2620
2621 static void
2622 swapdev_strategy(struct buf *bp, struct swdevt *sp)
2623 {
2624 struct vnode *vp2;
2625
2626 bp->b_blkno = ctodb(bp->b_blkno - sp->sw_first);
2627
2628 vp2 = sp->sw_id;
2629 vhold(vp2);
2630 if (bp->b_iocmd == BIO_WRITE) {
2631 if (bp->b_bufobj)
2632 bufobj_wdrop(bp->b_bufobj);
2633 bufobj_wref(&vp2->v_bufobj);
2634 }
2635 if (bp->b_bufobj != &vp2->v_bufobj)
2636 bp->b_bufobj = &vp2->v_bufobj;
2637 bp->b_vp = vp2;
2638 bp->b_iooffset = dbtob(bp->b_blkno);
2639 bstrategy(bp);
2640 return;
2641 }
2642
2643 static void
2644 swapdev_close(struct thread *td, struct swdevt *sp)
2645 {
2646
2647 VOP_CLOSE(sp->sw_vp, FREAD | FWRITE, td->td_ucred, td);
2648 vrele(sp->sw_vp);
2649 }
2650
2651
2652 static int
2653 swaponvp(struct thread *td, struct vnode *vp, u_long nblks)
2654 {
2655 struct swdevt *sp;
2656 int error;
2657
2658 if (nblks == 0)
2659 return (ENXIO);
2660 mtx_lock(&sw_dev_mtx);
2661 TAILQ_FOREACH(sp, &swtailq, sw_list) {
2662 if (sp->sw_id == vp) {
2663 mtx_unlock(&sw_dev_mtx);
2664 return (EBUSY);
2665 }
2666 }
2667 mtx_unlock(&sw_dev_mtx);
2668
2669 (void) vn_lock(vp, LK_EXCLUSIVE | LK_RETRY);
2670 #ifdef MAC
2671 error = mac_system_check_swapon(td->td_ucred, vp);
2672 if (error == 0)
2673 #endif
2674 error = VOP_OPEN(vp, FREAD | FWRITE, td->td_ucred, td, NULL);
2675 (void) VOP_UNLOCK(vp, 0);
2676 if (error)
2677 return (error);
2678
2679 swaponsomething(vp, vp, nblks, swapdev_strategy, swapdev_close,
2680 NODEV);
2681 return (0);
2682 }
Cache object: 7ab8c5930f3c48436bedb23917da1ad7
|