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("$FreeBSD$");
71
72 #include "opt_mac.h"
73 #include "opt_swap.h"
74 #include "opt_vm.h"
75
76 #include <sys/param.h>
77 #include <sys/systm.h>
78 #include <sys/conf.h>
79 #include <sys/kernel.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/mac.h>
89 #include <sys/malloc.h>
90 #include <sys/sysctl.h>
91 #include <sys/sysproto.h>
92 #include <sys/blist.h>
93 #include <sys/lock.h>
94 #include <sys/sx.h>
95 #include <sys/vmmeter.h>
96
97 #include <vm/vm.h>
98 #include <vm/pmap.h>
99 #include <vm/vm_map.h>
100 #include <vm/vm_kern.h>
101 #include <vm/vm_object.h>
102 #include <vm/vm_page.h>
103 #include <vm/vm_pager.h>
104 #include <vm/vm_pageout.h>
105 #include <vm/vm_param.h>
106 #include <vm/swap_pager.h>
107 #include <vm/vm_extern.h>
108 #include <vm/uma.h>
109
110 #include <geom/geom.h>
111
112 /*
113 * SWB_NPAGES must be a power of 2. It may be set to 1, 2, 4, 8, or 16
114 * pages per allocation. We recommend you stick with the default of 8.
115 * The 16-page limit is due to the radix code (kern/subr_blist.c).
116 */
117 #ifndef MAX_PAGEOUT_CLUSTER
118 #define MAX_PAGEOUT_CLUSTER 16
119 #endif
120
121 #if !defined(SWB_NPAGES)
122 #define SWB_NPAGES MAX_PAGEOUT_CLUSTER
123 #endif
124
125 /*
126 * Piecemeal swap metadata structure. Swap is stored in a radix tree.
127 *
128 * If SWB_NPAGES is 8 and sizeof(char *) == sizeof(daddr_t), our radix
129 * is basically 8. Assuming PAGE_SIZE == 4096, one tree level represents
130 * 32K worth of data, two levels represent 256K, three levels represent
131 * 2 MBytes. This is acceptable.
132 *
133 * Overall memory utilization is about the same as the old swap structure.
134 */
135 #define SWCORRECT(n) (sizeof(void *) * (n) / sizeof(daddr_t))
136 #define SWAP_META_PAGES (SWB_NPAGES * 2)
137 #define SWAP_META_MASK (SWAP_META_PAGES - 1)
138
139 typedef int32_t swblk_t; /*
140 * swap offset. This is the type used to
141 * address the "virtual swap device" and
142 * therefore the maximum swap space is
143 * 2^32 pages.
144 */
145
146 struct swdevt;
147 typedef void sw_strategy_t(struct buf *bp, struct swdevt *sw);
148 typedef void sw_close_t(struct thread *td, struct swdevt *sw);
149
150 /*
151 * Swap device table
152 */
153 struct swdevt {
154 int sw_flags;
155 int sw_nblks;
156 int sw_used;
157 dev_t sw_dev;
158 struct vnode *sw_vp;
159 void *sw_id;
160 swblk_t sw_first;
161 swblk_t sw_end;
162 struct blist *sw_blist;
163 TAILQ_ENTRY(swdevt) sw_list;
164 sw_strategy_t *sw_strategy;
165 sw_close_t *sw_close;
166 };
167
168 #define SW_CLOSING 0x04
169
170 struct swblock {
171 struct swblock *swb_hnext;
172 vm_object_t swb_object;
173 vm_pindex_t swb_index;
174 int swb_count;
175 daddr_t swb_pages[SWAP_META_PAGES];
176 };
177
178 static struct mtx sw_dev_mtx;
179 static TAILQ_HEAD(, swdevt) swtailq = TAILQ_HEAD_INITIALIZER(swtailq);
180 static struct swdevt *swdevhd; /* Allocate from here next */
181 static int nswapdev; /* Number of swap devices */
182 int swap_pager_avail;
183 static int swdev_syscall_active = 0; /* serialize swap(on|off) */
184
185 static void swapdev_strategy(struct buf *, struct swdevt *sw);
186
187 #define SWM_FREE 0x02 /* free, period */
188 #define SWM_POP 0x04 /* pop out */
189
190 int swap_pager_full = 2; /* swap space exhaustion (task killing) */
191 static int swap_pager_almost_full = 1; /* swap space exhaustion (w/hysteresis)*/
192 static int nsw_rcount; /* free read buffers */
193 static int nsw_wcount_sync; /* limit write buffers / synchronous */
194 static int nsw_wcount_async; /* limit write buffers / asynchronous */
195 static int nsw_wcount_async_max;/* assigned maximum */
196 static int nsw_cluster_max; /* maximum VOP I/O allowed */
197
198 static struct swblock **swhash;
199 static int swhash_mask;
200 static struct mtx swhash_mtx;
201
202 static int swap_async_max = 4; /* maximum in-progress async I/O's */
203 static struct sx sw_alloc_sx;
204
205
206 SYSCTL_INT(_vm, OID_AUTO, swap_async_max,
207 CTLFLAG_RW, &swap_async_max, 0, "Maximum running async swap ops");
208
209 /*
210 * "named" and "unnamed" anon region objects. Try to reduce the overhead
211 * of searching a named list by hashing it just a little.
212 */
213
214 #define NOBJLISTS 8
215
216 #define NOBJLIST(handle) \
217 (&swap_pager_object_list[((int)(intptr_t)handle >> 4) & (NOBJLISTS-1)])
218
219 static struct mtx sw_alloc_mtx; /* protect list manipulation */
220 static struct pagerlst swap_pager_object_list[NOBJLISTS];
221 static uma_zone_t swap_zone;
222 static struct vm_object swap_zone_obj;
223
224 /*
225 * pagerops for OBJT_SWAP - "swap pager". Some ops are also global procedure
226 * calls hooked from other parts of the VM system and do not appear here.
227 * (see vm/swap_pager.h).
228 */
229 static vm_object_t
230 swap_pager_alloc(void *handle, vm_ooffset_t size,
231 vm_prot_t prot, vm_ooffset_t offset);
232 static void swap_pager_dealloc(vm_object_t object);
233 static int swap_pager_getpages(vm_object_t, vm_page_t *, int, int);
234 static void swap_pager_putpages(vm_object_t, vm_page_t *, int, boolean_t, int *);
235 static boolean_t
236 swap_pager_haspage(vm_object_t object, vm_pindex_t pindex, int *before, int *after);
237 static void swap_pager_init(void);
238 static void swap_pager_unswapped(vm_page_t);
239 static void swap_pager_swapoff(struct swdevt *sp);
240
241 struct pagerops swappagerops = {
242 .pgo_init = swap_pager_init, /* early system initialization of pager */
243 .pgo_alloc = swap_pager_alloc, /* allocate an OBJT_SWAP object */
244 .pgo_dealloc = swap_pager_dealloc, /* deallocate an OBJT_SWAP object */
245 .pgo_getpages = swap_pager_getpages, /* pagein */
246 .pgo_putpages = swap_pager_putpages, /* pageout */
247 .pgo_haspage = swap_pager_haspage, /* get backing store status for page */
248 .pgo_pageunswapped = swap_pager_unswapped, /* remove swap related to page */
249 };
250
251 /*
252 * dmmax is in page-sized chunks with the new swap system. It was
253 * dev-bsized chunks in the old. dmmax is always a power of 2.
254 *
255 * swap_*() routines are externally accessible. swp_*() routines are
256 * internal.
257 */
258 static int dmmax;
259 static int nswap_lowat = 128; /* in pages, swap_pager_almost_full warn */
260 static int nswap_hiwat = 512; /* in pages, swap_pager_almost_full warn */
261
262 SYSCTL_INT(_vm, OID_AUTO, dmmax,
263 CTLFLAG_RD, &dmmax, 0, "Maximum size of a swap block");
264
265 static void swp_sizecheck(void);
266 static void swp_pager_async_iodone(struct buf *bp);
267 static int swapongeom(struct thread *, struct vnode *);
268 static int swaponvp(struct thread *, struct vnode *, u_long);
269
270 /*
271 * Swap bitmap functions
272 */
273 static void swp_pager_freeswapspace(daddr_t blk, int npages);
274 static daddr_t swp_pager_getswapspace(int npages);
275
276 /*
277 * Metadata functions
278 */
279 static struct swblock **swp_pager_hash(vm_object_t object, vm_pindex_t index);
280 static void swp_pager_meta_build(vm_object_t, vm_pindex_t, daddr_t);
281 static void swp_pager_meta_free(vm_object_t, vm_pindex_t, daddr_t);
282 static void swp_pager_meta_free_all(vm_object_t);
283 static daddr_t swp_pager_meta_ctl(vm_object_t, vm_pindex_t, int);
284
285 /*
286 * SWP_SIZECHECK() - update swap_pager_full indication
287 *
288 * update the swap_pager_almost_full indication and warn when we are
289 * about to run out of swap space, using lowat/hiwat hysteresis.
290 *
291 * Clear swap_pager_full ( task killing ) indication when lowat is met.
292 *
293 * No restrictions on call
294 * This routine may not block.
295 * This routine must be called at splvm()
296 */
297 static void
298 swp_sizecheck(void)
299 {
300
301 if (swap_pager_avail < nswap_lowat) {
302 if (swap_pager_almost_full == 0) {
303 printf("swap_pager: out of swap space\n");
304 swap_pager_almost_full = 1;
305 }
306 } else {
307 swap_pager_full = 0;
308 if (swap_pager_avail > nswap_hiwat)
309 swap_pager_almost_full = 0;
310 }
311 }
312
313 /*
314 * SWP_PAGER_HASH() - hash swap meta data
315 *
316 * This is an helper function which hashes the swapblk given
317 * the object and page index. It returns a pointer to a pointer
318 * to the object, or a pointer to a NULL pointer if it could not
319 * find a swapblk.
320 *
321 * This routine must be called at splvm().
322 */
323 static struct swblock **
324 swp_pager_hash(vm_object_t object, vm_pindex_t index)
325 {
326 struct swblock **pswap;
327 struct swblock *swap;
328
329 index &= ~(vm_pindex_t)SWAP_META_MASK;
330 pswap = &swhash[(index ^ (int)(intptr_t)object) & swhash_mask];
331 while ((swap = *pswap) != NULL) {
332 if (swap->swb_object == object &&
333 swap->swb_index == index
334 ) {
335 break;
336 }
337 pswap = &swap->swb_hnext;
338 }
339 return (pswap);
340 }
341
342 /*
343 * SWAP_PAGER_INIT() - initialize the swap pager!
344 *
345 * Expected to be started from system init. NOTE: This code is run
346 * before much else so be careful what you depend on. Most of the VM
347 * system has yet to be initialized at this point.
348 */
349 static void
350 swap_pager_init(void)
351 {
352 /*
353 * Initialize object lists
354 */
355 int i;
356
357 for (i = 0; i < NOBJLISTS; ++i)
358 TAILQ_INIT(&swap_pager_object_list[i]);
359 mtx_init(&sw_alloc_mtx, "swap_pager list", NULL, MTX_DEF);
360 mtx_init(&sw_dev_mtx, "swapdev", NULL, MTX_DEF);
361
362 /*
363 * Device Stripe, in PAGE_SIZE'd blocks
364 */
365 dmmax = SWB_NPAGES * 2;
366 }
367
368 /*
369 * SWAP_PAGER_SWAP_INIT() - swap pager initialization from pageout process
370 *
371 * Expected to be started from pageout process once, prior to entering
372 * its main loop.
373 */
374 void
375 swap_pager_swap_init(void)
376 {
377 int n, n2;
378
379 /*
380 * Number of in-transit swap bp operations. Don't
381 * exhaust the pbufs completely. Make sure we
382 * initialize workable values (0 will work for hysteresis
383 * but it isn't very efficient).
384 *
385 * The nsw_cluster_max is constrained by the bp->b_pages[]
386 * array (MAXPHYS/PAGE_SIZE) and our locally defined
387 * MAX_PAGEOUT_CLUSTER. Also be aware that swap ops are
388 * constrained by the swap device interleave stripe size.
389 *
390 * Currently we hardwire nsw_wcount_async to 4. This limit is
391 * designed to prevent other I/O from having high latencies due to
392 * our pageout I/O. The value 4 works well for one or two active swap
393 * devices but is probably a little low if you have more. Even so,
394 * a higher value would probably generate only a limited improvement
395 * with three or four active swap devices since the system does not
396 * typically have to pageout at extreme bandwidths. We will want
397 * at least 2 per swap devices, and 4 is a pretty good value if you
398 * have one NFS swap device due to the command/ack latency over NFS.
399 * So it all works out pretty well.
400 */
401 nsw_cluster_max = min((MAXPHYS/PAGE_SIZE), MAX_PAGEOUT_CLUSTER);
402
403 mtx_lock(&pbuf_mtx);
404 nsw_rcount = (nswbuf + 1) / 2;
405 nsw_wcount_sync = (nswbuf + 3) / 4;
406 nsw_wcount_async = 4;
407 nsw_wcount_async_max = nsw_wcount_async;
408 mtx_unlock(&pbuf_mtx);
409
410 /*
411 * Initialize our zone. Right now I'm just guessing on the number
412 * we need based on the number of pages in the system. Each swblock
413 * can hold 16 pages, so this is probably overkill. This reservation
414 * is typically limited to around 32MB by default.
415 */
416 n = cnt.v_page_count / 2;
417 if (maxswzone && n > maxswzone / sizeof(struct swblock))
418 n = maxswzone / sizeof(struct swblock);
419 n2 = n;
420 swap_zone = uma_zcreate("SWAPMETA", sizeof(struct swblock), NULL, NULL,
421 NULL, NULL, UMA_ALIGN_PTR, UMA_ZONE_NOFREE | UMA_ZONE_VM);
422 if (swap_zone == NULL)
423 panic("failed to create swap_zone.");
424 do {
425 if (uma_zone_set_obj(swap_zone, &swap_zone_obj, n))
426 break;
427 /*
428 * if the allocation failed, try a zone two thirds the
429 * size of the previous attempt.
430 */
431 n -= ((n + 2) / 3);
432 } while (n > 0);
433 if (n2 != n)
434 printf("Swap zone entries reduced from %d to %d.\n", n2, n);
435 n2 = n;
436
437 /*
438 * Initialize our meta-data hash table. The swapper does not need to
439 * be quite as efficient as the VM system, so we do not use an
440 * oversized hash table.
441 *
442 * n: size of hash table, must be power of 2
443 * swhash_mask: hash table index mask
444 */
445 for (n = 1; n < n2 / 8; n *= 2)
446 ;
447 swhash = malloc(sizeof(struct swblock *) * n, M_VMPGDATA, M_WAITOK | M_ZERO);
448 swhash_mask = n - 1;
449 mtx_init(&swhash_mtx, "swap_pager swhash", NULL, MTX_DEF);
450 }
451
452 /*
453 * SWAP_PAGER_ALLOC() - allocate a new OBJT_SWAP VM object and instantiate
454 * its metadata structures.
455 *
456 * This routine is called from the mmap and fork code to create a new
457 * OBJT_SWAP object. We do this by creating an OBJT_DEFAULT object
458 * and then converting it with swp_pager_meta_build().
459 *
460 * This routine may block in vm_object_allocate() and create a named
461 * object lookup race, so we must interlock. We must also run at
462 * splvm() for the object lookup to handle races with interrupts, but
463 * we do not have to maintain splvm() in between the lookup and the
464 * add because (I believe) it is not possible to attempt to create
465 * a new swap object w/handle when a default object with that handle
466 * already exists.
467 *
468 * MPSAFE
469 */
470 static vm_object_t
471 swap_pager_alloc(void *handle, vm_ooffset_t size, vm_prot_t prot,
472 vm_ooffset_t offset)
473 {
474 vm_object_t object;
475 vm_pindex_t pindex;
476
477 pindex = OFF_TO_IDX(offset + PAGE_MASK + size);
478
479 if (handle) {
480 mtx_lock(&Giant);
481 /*
482 * Reference existing named region or allocate new one. There
483 * should not be a race here against swp_pager_meta_build()
484 * as called from vm_page_remove() in regards to the lookup
485 * of the handle.
486 */
487 sx_xlock(&sw_alloc_sx);
488 object = vm_pager_object_lookup(NOBJLIST(handle), handle);
489
490 if (object != NULL) {
491 vm_object_reference(object);
492 } else {
493 object = vm_object_allocate(OBJT_DEFAULT, pindex);
494 object->handle = handle;
495
496 VM_OBJECT_LOCK(object);
497 swp_pager_meta_build(object, 0, SWAPBLK_NONE);
498 VM_OBJECT_UNLOCK(object);
499 }
500 sx_xunlock(&sw_alloc_sx);
501 mtx_unlock(&Giant);
502 } else {
503 object = vm_object_allocate(OBJT_DEFAULT, pindex);
504
505 VM_OBJECT_LOCK(object);
506 swp_pager_meta_build(object, 0, SWAPBLK_NONE);
507 VM_OBJECT_UNLOCK(object);
508 }
509 return (object);
510 }
511
512 /*
513 * SWAP_PAGER_DEALLOC() - remove swap metadata from object
514 *
515 * The swap backing for the object is destroyed. The code is
516 * designed such that we can reinstantiate it later, but this
517 * routine is typically called only when the entire object is
518 * about to be destroyed.
519 *
520 * This routine may block, but no longer does.
521 *
522 * The object must be locked or unreferenceable.
523 */
524 static void
525 swap_pager_dealloc(vm_object_t object)
526 {
527 int s;
528
529 /*
530 * Remove from list right away so lookups will fail if we block for
531 * pageout completion.
532 */
533 if (object->handle != NULL) {
534 mtx_lock(&sw_alloc_mtx);
535 TAILQ_REMOVE(NOBJLIST(object->handle), object, pager_object_list);
536 mtx_unlock(&sw_alloc_mtx);
537 }
538
539 VM_OBJECT_LOCK_ASSERT(object, MA_OWNED);
540 vm_object_pip_wait(object, "swpdea");
541
542 /*
543 * Free all remaining metadata. We only bother to free it from
544 * the swap meta data. We do not attempt to free swapblk's still
545 * associated with vm_page_t's for this object. We do not care
546 * if paging is still in progress on some objects.
547 */
548 s = splvm();
549 swp_pager_meta_free_all(object);
550 splx(s);
551 }
552
553 /************************************************************************
554 * SWAP PAGER BITMAP ROUTINES *
555 ************************************************************************/
556
557 /*
558 * SWP_PAGER_GETSWAPSPACE() - allocate raw swap space
559 *
560 * Allocate swap for the requested number of pages. The starting
561 * swap block number (a page index) is returned or SWAPBLK_NONE
562 * if the allocation failed.
563 *
564 * Also has the side effect of advising that somebody made a mistake
565 * when they configured swap and didn't configure enough.
566 *
567 * Must be called at splvm() to avoid races with bitmap frees from
568 * vm_page_remove() aka swap_pager_page_removed().
569 *
570 * This routine may not block
571 * This routine must be called at splvm().
572 *
573 * We allocate in round-robin fashion from the configured devices.
574 */
575 static daddr_t
576 swp_pager_getswapspace(int npages)
577 {
578 daddr_t blk;
579 struct swdevt *sp;
580 int i;
581
582 blk = SWAPBLK_NONE;
583 mtx_lock(&sw_dev_mtx);
584 sp = swdevhd;
585 for (i = 0; i < nswapdev; i++) {
586 if (sp == NULL)
587 sp = TAILQ_FIRST(&swtailq);
588 if (!(sp->sw_flags & SW_CLOSING)) {
589 blk = blist_alloc(sp->sw_blist, npages);
590 if (blk != SWAPBLK_NONE) {
591 blk += sp->sw_first;
592 sp->sw_used += npages;
593 swap_pager_avail -= npages;
594 swp_sizecheck();
595 swdevhd = TAILQ_NEXT(sp, sw_list);
596 goto done;
597 }
598 }
599 sp = TAILQ_NEXT(sp, sw_list);
600 }
601 if (swap_pager_full != 2) {
602 printf("swap_pager_getswapspace(%d): failed\n", npages);
603 swap_pager_full = 2;
604 swap_pager_almost_full = 1;
605 }
606 swdevhd = NULL;
607 done:
608 mtx_unlock(&sw_dev_mtx);
609 return (blk);
610 }
611
612 static int
613 swp_pager_isondev(daddr_t blk, struct swdevt *sp)
614 {
615
616 return (blk >= sp->sw_first && blk < sp->sw_end);
617 }
618
619 static void
620 swp_pager_strategy(struct buf *bp)
621 {
622 struct swdevt *sp;
623
624 mtx_lock(&sw_dev_mtx);
625 TAILQ_FOREACH(sp, &swtailq, sw_list) {
626 if (bp->b_blkno >= sp->sw_first && bp->b_blkno < sp->sw_end) {
627 mtx_unlock(&sw_dev_mtx);
628 sp->sw_strategy(bp, sp);
629 return;
630 }
631 }
632 panic("Swapdev not found");
633 }
634
635
636 /*
637 * SWP_PAGER_FREESWAPSPACE() - free raw swap space
638 *
639 * This routine returns the specified swap blocks back to the bitmap.
640 *
641 * Note: This routine may not block (it could in the old swap code),
642 * and through the use of the new blist routines it does not block.
643 *
644 * We must be called at splvm() to avoid races with bitmap frees from
645 * vm_page_remove() aka swap_pager_page_removed().
646 *
647 * This routine may not block
648 * This routine must be called at splvm().
649 */
650 static void
651 swp_pager_freeswapspace(daddr_t blk, int npages)
652 {
653 struct swdevt *sp;
654
655 mtx_lock(&sw_dev_mtx);
656 TAILQ_FOREACH(sp, &swtailq, sw_list) {
657 if (blk >= sp->sw_first && blk < sp->sw_end) {
658 sp->sw_used -= npages;
659 /*
660 * If we are attempting to stop swapping on
661 * this device, we don't want to mark any
662 * blocks free lest they be reused.
663 */
664 if ((sp->sw_flags & SW_CLOSING) == 0) {
665 blist_free(sp->sw_blist, blk - sp->sw_first,
666 npages);
667 swap_pager_avail += npages;
668 swp_sizecheck();
669 }
670 mtx_unlock(&sw_dev_mtx);
671 return;
672 }
673 }
674 panic("Swapdev not found");
675 }
676
677 /*
678 * SWAP_PAGER_FREESPACE() - frees swap blocks associated with a page
679 * range within an object.
680 *
681 * This is a globally accessible routine.
682 *
683 * This routine removes swapblk assignments from swap metadata.
684 *
685 * The external callers of this routine typically have already destroyed
686 * or renamed vm_page_t's associated with this range in the object so
687 * we should be ok.
688 *
689 * This routine may be called at any spl. We up our spl to splvm temporarily
690 * in order to perform the metadata removal.
691 */
692 void
693 swap_pager_freespace(vm_object_t object, vm_pindex_t start, vm_size_t size)
694 {
695 int s = splvm();
696
697 VM_OBJECT_LOCK_ASSERT(object, MA_OWNED);
698 swp_pager_meta_free(object, start, size);
699 splx(s);
700 }
701
702 /*
703 * SWAP_PAGER_RESERVE() - reserve swap blocks in object
704 *
705 * Assigns swap blocks to the specified range within the object. The
706 * swap blocks are not zerod. Any previous swap assignment is destroyed.
707 *
708 * Returns 0 on success, -1 on failure.
709 */
710 int
711 swap_pager_reserve(vm_object_t object, vm_pindex_t start, vm_size_t size)
712 {
713 int s;
714 int n = 0;
715 daddr_t blk = SWAPBLK_NONE;
716 vm_pindex_t beg = start; /* save start index */
717
718 s = splvm();
719 VM_OBJECT_LOCK(object);
720 while (size) {
721 if (n == 0) {
722 n = BLIST_MAX_ALLOC;
723 while ((blk = swp_pager_getswapspace(n)) == SWAPBLK_NONE) {
724 n >>= 1;
725 if (n == 0) {
726 swp_pager_meta_free(object, beg, start - beg);
727 VM_OBJECT_UNLOCK(object);
728 splx(s);
729 return (-1);
730 }
731 }
732 }
733 swp_pager_meta_build(object, start, blk);
734 --size;
735 ++start;
736 ++blk;
737 --n;
738 }
739 swp_pager_meta_free(object, start, n);
740 VM_OBJECT_UNLOCK(object);
741 splx(s);
742 return (0);
743 }
744
745 /*
746 * SWAP_PAGER_COPY() - copy blocks from source pager to destination pager
747 * and destroy the source.
748 *
749 * Copy any valid swapblks from the source to the destination. In
750 * cases where both the source and destination have a valid swapblk,
751 * we keep the destination's.
752 *
753 * This routine is allowed to block. It may block allocating metadata
754 * indirectly through swp_pager_meta_build() or if paging is still in
755 * progress on the source.
756 *
757 * This routine can be called at any spl
758 *
759 * XXX vm_page_collapse() kinda expects us not to block because we
760 * supposedly do not need to allocate memory, but for the moment we
761 * *may* have to get a little memory from the zone allocator, but
762 * it is taken from the interrupt memory. We should be ok.
763 *
764 * The source object contains no vm_page_t's (which is just as well)
765 *
766 * The source object is of type OBJT_SWAP.
767 *
768 * The source and destination objects must be locked or
769 * inaccessible (XXX are they ?)
770 */
771 void
772 swap_pager_copy(vm_object_t srcobject, vm_object_t dstobject,
773 vm_pindex_t offset, int destroysource)
774 {
775 vm_pindex_t i;
776 int s;
777
778 VM_OBJECT_LOCK_ASSERT(srcobject, MA_OWNED);
779 VM_OBJECT_LOCK_ASSERT(dstobject, MA_OWNED);
780
781 s = splvm();
782 /*
783 * If destroysource is set, we remove the source object from the
784 * swap_pager internal queue now.
785 */
786 if (destroysource) {
787 if (srcobject->handle != NULL) {
788 mtx_lock(&sw_alloc_mtx);
789 TAILQ_REMOVE(
790 NOBJLIST(srcobject->handle),
791 srcobject,
792 pager_object_list
793 );
794 mtx_unlock(&sw_alloc_mtx);
795 }
796 }
797
798 /*
799 * transfer source to destination.
800 */
801 for (i = 0; i < dstobject->size; ++i) {
802 daddr_t dstaddr;
803
804 /*
805 * Locate (without changing) the swapblk on the destination,
806 * unless it is invalid in which case free it silently, or
807 * if the destination is a resident page, in which case the
808 * source is thrown away.
809 */
810 dstaddr = swp_pager_meta_ctl(dstobject, i, 0);
811
812 if (dstaddr == SWAPBLK_NONE) {
813 /*
814 * Destination has no swapblk and is not resident,
815 * copy source.
816 */
817 daddr_t srcaddr;
818
819 srcaddr = swp_pager_meta_ctl(
820 srcobject,
821 i + offset,
822 SWM_POP
823 );
824
825 if (srcaddr != SWAPBLK_NONE) {
826 /*
827 * swp_pager_meta_build() can sleep.
828 */
829 vm_object_pip_add(srcobject, 1);
830 VM_OBJECT_UNLOCK(srcobject);
831 vm_object_pip_add(dstobject, 1);
832 swp_pager_meta_build(dstobject, i, srcaddr);
833 vm_object_pip_wakeup(dstobject);
834 VM_OBJECT_LOCK(srcobject);
835 vm_object_pip_wakeup(srcobject);
836 }
837 } else {
838 /*
839 * Destination has valid swapblk or it is represented
840 * by a resident page. We destroy the sourceblock.
841 */
842
843 swp_pager_meta_ctl(srcobject, i + offset, SWM_FREE);
844 }
845 }
846
847 /*
848 * Free left over swap blocks in source.
849 *
850 * We have to revert the type to OBJT_DEFAULT so we do not accidently
851 * double-remove the object from the swap queues.
852 */
853 if (destroysource) {
854 swp_pager_meta_free_all(srcobject);
855 /*
856 * Reverting the type is not necessary, the caller is going
857 * to destroy srcobject directly, but I'm doing it here
858 * for consistency since we've removed the object from its
859 * queues.
860 */
861 srcobject->type = OBJT_DEFAULT;
862 }
863 splx(s);
864 }
865
866 /*
867 * SWAP_PAGER_HASPAGE() - determine if we have good backing store for
868 * the requested page.
869 *
870 * We determine whether good backing store exists for the requested
871 * page and return TRUE if it does, FALSE if it doesn't.
872 *
873 * If TRUE, we also try to determine how much valid, contiguous backing
874 * store exists before and after the requested page within a reasonable
875 * distance. We do not try to restrict it to the swap device stripe
876 * (that is handled in getpages/putpages). It probably isn't worth
877 * doing here.
878 */
879 static boolean_t
880 swap_pager_haspage(vm_object_t object, vm_pindex_t pindex, int *before, int *after)
881 {
882 daddr_t blk0;
883 int s;
884
885 VM_OBJECT_LOCK_ASSERT(object, MA_OWNED);
886 /*
887 * do we have good backing store at the requested index ?
888 */
889 s = splvm();
890 blk0 = swp_pager_meta_ctl(object, pindex, 0);
891
892 if (blk0 == SWAPBLK_NONE) {
893 splx(s);
894 if (before)
895 *before = 0;
896 if (after)
897 *after = 0;
898 return (FALSE);
899 }
900
901 /*
902 * find backwards-looking contiguous good backing store
903 */
904 if (before != NULL) {
905 int i;
906
907 for (i = 1; i < (SWB_NPAGES/2); ++i) {
908 daddr_t blk;
909
910 if (i > pindex)
911 break;
912 blk = swp_pager_meta_ctl(object, pindex - i, 0);
913 if (blk != blk0 - i)
914 break;
915 }
916 *before = (i - 1);
917 }
918
919 /*
920 * find forward-looking contiguous good backing store
921 */
922 if (after != NULL) {
923 int i;
924
925 for (i = 1; i < (SWB_NPAGES/2); ++i) {
926 daddr_t blk;
927
928 blk = swp_pager_meta_ctl(object, pindex + i, 0);
929 if (blk != blk0 + i)
930 break;
931 }
932 *after = (i - 1);
933 }
934 splx(s);
935 return (TRUE);
936 }
937
938 /*
939 * SWAP_PAGER_PAGE_UNSWAPPED() - remove swap backing store related to page
940 *
941 * This removes any associated swap backing store, whether valid or
942 * not, from the page.
943 *
944 * This routine is typically called when a page is made dirty, at
945 * which point any associated swap can be freed. MADV_FREE also
946 * calls us in a special-case situation
947 *
948 * NOTE!!! If the page is clean and the swap was valid, the caller
949 * should make the page dirty before calling this routine. This routine
950 * does NOT change the m->dirty status of the page. Also: MADV_FREE
951 * depends on it.
952 *
953 * This routine may not block
954 * This routine must be called at splvm()
955 */
956 static void
957 swap_pager_unswapped(vm_page_t m)
958 {
959
960 VM_OBJECT_LOCK_ASSERT(m->object, MA_OWNED);
961 swp_pager_meta_ctl(m->object, m->pindex, SWM_FREE);
962 }
963
964 /*
965 * SWAP_PAGER_GETPAGES() - bring pages in from swap
966 *
967 * Attempt to retrieve (m, count) pages from backing store, but make
968 * sure we retrieve at least m[reqpage]. We try to load in as large
969 * a chunk surrounding m[reqpage] as is contiguous in swap and which
970 * belongs to the same object.
971 *
972 * The code is designed for asynchronous operation and
973 * immediate-notification of 'reqpage' but tends not to be
974 * used that way. Please do not optimize-out this algorithmic
975 * feature, I intend to improve on it in the future.
976 *
977 * The parent has a single vm_object_pip_add() reference prior to
978 * calling us and we should return with the same.
979 *
980 * The parent has BUSY'd the pages. We should return with 'm'
981 * left busy, but the others adjusted.
982 */
983 static int
984 swap_pager_getpages(vm_object_t object, vm_page_t *m, int count, int reqpage)
985 {
986 struct buf *bp;
987 vm_page_t mreq;
988 int s;
989 int i;
990 int j;
991 daddr_t blk;
992
993 mreq = m[reqpage];
994
995 KASSERT(mreq->object == object,
996 ("swap_pager_getpages: object mismatch %p/%p",
997 object, mreq->object));
998
999 /*
1000 * Calculate range to retrieve. The pages have already been assigned
1001 * their swapblks. We require a *contiguous* range but we know it to
1002 * not span devices. If we do not supply it, bad things
1003 * happen. Note that blk, iblk & jblk can be SWAPBLK_NONE, but the
1004 * loops are set up such that the case(s) are handled implicitly.
1005 *
1006 * The swp_*() calls must be made at splvm(). vm_page_free() does
1007 * not need to be, but it will go a little faster if it is.
1008 */
1009 s = splvm();
1010 blk = swp_pager_meta_ctl(mreq->object, mreq->pindex, 0);
1011
1012 for (i = reqpage - 1; i >= 0; --i) {
1013 daddr_t iblk;
1014
1015 iblk = swp_pager_meta_ctl(m[i]->object, m[i]->pindex, 0);
1016 if (blk != iblk + (reqpage - i))
1017 break;
1018 }
1019 ++i;
1020
1021 for (j = reqpage + 1; j < count; ++j) {
1022 daddr_t jblk;
1023
1024 jblk = swp_pager_meta_ctl(m[j]->object, m[j]->pindex, 0);
1025 if (blk != jblk - (j - reqpage))
1026 break;
1027 }
1028
1029 /*
1030 * free pages outside our collection range. Note: we never free
1031 * mreq, it must remain busy throughout.
1032 */
1033 vm_page_lock_queues();
1034 {
1035 int k;
1036
1037 for (k = 0; k < i; ++k)
1038 vm_page_free(m[k]);
1039 for (k = j; k < count; ++k)
1040 vm_page_free(m[k]);
1041 }
1042 vm_page_unlock_queues();
1043 splx(s);
1044
1045
1046 /*
1047 * Return VM_PAGER_FAIL if we have nothing to do. Return mreq
1048 * still busy, but the others unbusied.
1049 */
1050 if (blk == SWAPBLK_NONE)
1051 return (VM_PAGER_FAIL);
1052
1053 /*
1054 * Getpbuf() can sleep.
1055 */
1056 VM_OBJECT_UNLOCK(object);
1057 /*
1058 * Get a swap buffer header to perform the IO
1059 */
1060 bp = getpbuf(&nsw_rcount);
1061 bp->b_flags |= B_PAGING;
1062
1063 /*
1064 * map our page(s) into kva for input
1065 */
1066 pmap_qenter((vm_offset_t)bp->b_data, m + i, j - i);
1067
1068 bp->b_iocmd = BIO_READ;
1069 bp->b_iodone = swp_pager_async_iodone;
1070 bp->b_rcred = crhold(thread0.td_ucred);
1071 bp->b_wcred = crhold(thread0.td_ucred);
1072 bp->b_blkno = blk - (reqpage - i);
1073 bp->b_bcount = PAGE_SIZE * (j - i);
1074 bp->b_bufsize = PAGE_SIZE * (j - i);
1075 bp->b_pager.pg_reqpage = reqpage - i;
1076
1077 VM_OBJECT_LOCK(object);
1078 vm_page_lock_queues();
1079 {
1080 int k;
1081
1082 for (k = i; k < j; ++k) {
1083 bp->b_pages[k - i] = m[k];
1084 vm_page_flag_set(m[k], PG_SWAPINPROG);
1085 }
1086 }
1087 vm_page_unlock_queues();
1088 VM_OBJECT_UNLOCK(object);
1089 bp->b_npages = j - i;
1090
1091 cnt.v_swapin++;
1092 cnt.v_swappgsin += bp->b_npages;
1093
1094 /*
1095 * We still hold the lock on mreq, and our automatic completion routine
1096 * does not remove it.
1097 */
1098 VM_OBJECT_LOCK(mreq->object);
1099 vm_object_pip_add(mreq->object, bp->b_npages);
1100 VM_OBJECT_UNLOCK(mreq->object);
1101
1102 /*
1103 * perform the I/O. NOTE!!! bp cannot be considered valid after
1104 * this point because we automatically release it on completion.
1105 * Instead, we look at the one page we are interested in which we
1106 * still hold a lock on even through the I/O completion.
1107 *
1108 * The other pages in our m[] array are also released on completion,
1109 * so we cannot assume they are valid anymore either.
1110 *
1111 * NOTE: b_blkno is destroyed by the call to swapdev_strategy
1112 */
1113 BUF_KERNPROC(bp);
1114 swp_pager_strategy(bp);
1115
1116 /*
1117 * wait for the page we want to complete. PG_SWAPINPROG is always
1118 * cleared on completion. If an I/O error occurs, SWAPBLK_NONE
1119 * is set in the meta-data.
1120 */
1121 s = splvm();
1122 vm_page_lock_queues();
1123 while ((mreq->flags & PG_SWAPINPROG) != 0) {
1124 vm_page_flag_set(mreq, PG_WANTED | PG_REFERENCED);
1125 cnt.v_intrans++;
1126 if (msleep(mreq, &vm_page_queue_mtx, PSWP, "swread", hz*20)) {
1127 printf(
1128 "swap_pager: indefinite wait buffer: device: %s, blkno: %jd, size: %ld\n",
1129 bp->b_dev == NULL ? "[NULL]" : devtoname(bp->b_dev),
1130 (intmax_t)bp->b_blkno, bp->b_bcount);
1131 }
1132 }
1133 vm_page_unlock_queues();
1134 splx(s);
1135
1136 VM_OBJECT_LOCK(mreq->object);
1137 /*
1138 * mreq is left busied after completion, but all the other pages
1139 * are freed. If we had an unrecoverable read error the page will
1140 * not be valid.
1141 */
1142 if (mreq->valid != VM_PAGE_BITS_ALL) {
1143 return (VM_PAGER_ERROR);
1144 } else {
1145 return (VM_PAGER_OK);
1146 }
1147
1148 /*
1149 * A final note: in a low swap situation, we cannot deallocate swap
1150 * and mark a page dirty here because the caller is likely to mark
1151 * the page clean when we return, causing the page to possibly revert
1152 * to all-zero's later.
1153 */
1154 }
1155
1156 /*
1157 * swap_pager_putpages:
1158 *
1159 * Assign swap (if necessary) and initiate I/O on the specified pages.
1160 *
1161 * We support both OBJT_DEFAULT and OBJT_SWAP objects. DEFAULT objects
1162 * are automatically converted to SWAP objects.
1163 *
1164 * In a low memory situation we may block in VOP_STRATEGY(), but the new
1165 * vm_page reservation system coupled with properly written VFS devices
1166 * should ensure that no low-memory deadlock occurs. This is an area
1167 * which needs work.
1168 *
1169 * The parent has N vm_object_pip_add() references prior to
1170 * calling us and will remove references for rtvals[] that are
1171 * not set to VM_PAGER_PEND. We need to remove the rest on I/O
1172 * completion.
1173 *
1174 * The parent has soft-busy'd the pages it passes us and will unbusy
1175 * those whos rtvals[] entry is not set to VM_PAGER_PEND on return.
1176 * We need to unbusy the rest on I/O completion.
1177 */
1178 void
1179 swap_pager_putpages(vm_object_t object, vm_page_t *m, int count,
1180 boolean_t sync, int *rtvals)
1181 {
1182 int i;
1183 int n = 0;
1184
1185 GIANT_REQUIRED;
1186 if (count && m[0]->object != object) {
1187 panic("swap_pager_getpages: object mismatch %p/%p",
1188 object,
1189 m[0]->object
1190 );
1191 }
1192
1193 /*
1194 * Step 1
1195 *
1196 * Turn object into OBJT_SWAP
1197 * check for bogus sysops
1198 * force sync if not pageout process
1199 */
1200 if (object->type != OBJT_SWAP)
1201 swp_pager_meta_build(object, 0, SWAPBLK_NONE);
1202 VM_OBJECT_UNLOCK(object);
1203
1204 if (curproc != pageproc)
1205 sync = TRUE;
1206
1207 /*
1208 * Step 2
1209 *
1210 * Update nsw parameters from swap_async_max sysctl values.
1211 * Do not let the sysop crash the machine with bogus numbers.
1212 */
1213 mtx_lock(&pbuf_mtx);
1214 if (swap_async_max != nsw_wcount_async_max) {
1215 int n;
1216 int s;
1217
1218 /*
1219 * limit range
1220 */
1221 if ((n = swap_async_max) > nswbuf / 2)
1222 n = nswbuf / 2;
1223 if (n < 1)
1224 n = 1;
1225 swap_async_max = n;
1226
1227 /*
1228 * Adjust difference ( if possible ). If the current async
1229 * count is too low, we may not be able to make the adjustment
1230 * at this time.
1231 */
1232 s = splvm();
1233 n -= nsw_wcount_async_max;
1234 if (nsw_wcount_async + n >= 0) {
1235 nsw_wcount_async += n;
1236 nsw_wcount_async_max += n;
1237 wakeup(&nsw_wcount_async);
1238 }
1239 splx(s);
1240 }
1241 mtx_unlock(&pbuf_mtx);
1242
1243 /*
1244 * Step 3
1245 *
1246 * Assign swap blocks and issue I/O. We reallocate swap on the fly.
1247 * The page is left dirty until the pageout operation completes
1248 * successfully.
1249 */
1250 for (i = 0; i < count; i += n) {
1251 int s;
1252 int j;
1253 struct buf *bp;
1254 daddr_t blk;
1255
1256 /*
1257 * Maximum I/O size is limited by a number of factors.
1258 */
1259 n = min(BLIST_MAX_ALLOC, count - i);
1260 n = min(n, nsw_cluster_max);
1261
1262 s = splvm();
1263
1264 /*
1265 * Get biggest block of swap we can. If we fail, fall
1266 * back and try to allocate a smaller block. Don't go
1267 * overboard trying to allocate space if it would overly
1268 * fragment swap.
1269 */
1270 while (
1271 (blk = swp_pager_getswapspace(n)) == SWAPBLK_NONE &&
1272 n > 4
1273 ) {
1274 n >>= 1;
1275 }
1276 if (blk == SWAPBLK_NONE) {
1277 for (j = 0; j < n; ++j)
1278 rtvals[i+j] = VM_PAGER_FAIL;
1279 splx(s);
1280 continue;
1281 }
1282
1283 /*
1284 * All I/O parameters have been satisfied, build the I/O
1285 * request and assign the swap space.
1286 */
1287 if (sync == TRUE) {
1288 bp = getpbuf(&nsw_wcount_sync);
1289 } else {
1290 bp = getpbuf(&nsw_wcount_async);
1291 bp->b_flags = B_ASYNC;
1292 }
1293 bp->b_flags |= B_PAGING;
1294 bp->b_iocmd = BIO_WRITE;
1295
1296 pmap_qenter((vm_offset_t)bp->b_data, &m[i], n);
1297
1298 bp->b_rcred = crhold(thread0.td_ucred);
1299 bp->b_wcred = crhold(thread0.td_ucred);
1300 bp->b_bcount = PAGE_SIZE * n;
1301 bp->b_bufsize = PAGE_SIZE * n;
1302 bp->b_blkno = blk;
1303
1304 VM_OBJECT_LOCK(object);
1305 for (j = 0; j < n; ++j) {
1306 vm_page_t mreq = m[i+j];
1307
1308 swp_pager_meta_build(
1309 mreq->object,
1310 mreq->pindex,
1311 blk + j
1312 );
1313 vm_page_dirty(mreq);
1314 rtvals[i+j] = VM_PAGER_OK;
1315
1316 vm_page_lock_queues();
1317 vm_page_flag_set(mreq, PG_SWAPINPROG);
1318 vm_page_unlock_queues();
1319 bp->b_pages[j] = mreq;
1320 }
1321 VM_OBJECT_UNLOCK(object);
1322 bp->b_npages = n;
1323 /*
1324 * Must set dirty range for NFS to work.
1325 */
1326 bp->b_dirtyoff = 0;
1327 bp->b_dirtyend = bp->b_bcount;
1328
1329 cnt.v_swapout++;
1330 cnt.v_swappgsout += bp->b_npages;
1331
1332 splx(s);
1333
1334 /*
1335 * asynchronous
1336 *
1337 * NOTE: b_blkno is destroyed by the call to swapdev_strategy
1338 */
1339 if (sync == FALSE) {
1340 bp->b_iodone = swp_pager_async_iodone;
1341 BUF_KERNPROC(bp);
1342 swp_pager_strategy(bp);
1343
1344 for (j = 0; j < n; ++j)
1345 rtvals[i+j] = VM_PAGER_PEND;
1346 /* restart outter loop */
1347 continue;
1348 }
1349
1350 /*
1351 * synchronous
1352 *
1353 * NOTE: b_blkno is destroyed by the call to swapdev_strategy
1354 */
1355 bp->b_iodone = bdone;
1356 swp_pager_strategy(bp);
1357
1358 /*
1359 * Wait for the sync I/O to complete, then update rtvals.
1360 * We just set the rtvals[] to VM_PAGER_PEND so we can call
1361 * our async completion routine at the end, thus avoiding a
1362 * double-free.
1363 */
1364 s = splbio();
1365 bwait(bp, PVM, "swwrt");
1366 for (j = 0; j < n; ++j)
1367 rtvals[i+j] = VM_PAGER_PEND;
1368 /*
1369 * Now that we are through with the bp, we can call the
1370 * normal async completion, which frees everything up.
1371 */
1372 swp_pager_async_iodone(bp);
1373 splx(s);
1374 }
1375 VM_OBJECT_LOCK(object);
1376 }
1377
1378 /*
1379 * swp_pager_async_iodone:
1380 *
1381 * Completion routine for asynchronous reads and writes from/to swap.
1382 * Also called manually by synchronous code to finish up a bp.
1383 *
1384 * For READ operations, the pages are PG_BUSY'd. For WRITE operations,
1385 * the pages are vm_page_t->busy'd. For READ operations, we PG_BUSY
1386 * unbusy all pages except the 'main' request page. For WRITE
1387 * operations, we vm_page_t->busy'd unbusy all pages ( we can do this
1388 * because we marked them all VM_PAGER_PEND on return from putpages ).
1389 *
1390 * This routine may not block.
1391 * This routine is called at splbio() or better
1392 *
1393 * We up ourselves to splvm() as required for various vm_page related
1394 * calls.
1395 */
1396 static void
1397 swp_pager_async_iodone(struct buf *bp)
1398 {
1399 int s;
1400 int i;
1401 vm_object_t object = NULL;
1402
1403 bp->b_flags |= B_DONE;
1404
1405 /*
1406 * report error
1407 */
1408 if (bp->b_ioflags & BIO_ERROR) {
1409 printf(
1410 "swap_pager: I/O error - %s failed; blkno %ld,"
1411 "size %ld, error %d\n",
1412 ((bp->b_iocmd == BIO_READ) ? "pagein" : "pageout"),
1413 (long)bp->b_blkno,
1414 (long)bp->b_bcount,
1415 bp->b_error
1416 );
1417 }
1418
1419 /*
1420 * set object, raise to splvm().
1421 */
1422 s = splvm();
1423
1424 /*
1425 * remove the mapping for kernel virtual
1426 */
1427 pmap_qremove((vm_offset_t)bp->b_data, bp->b_npages);
1428
1429 if (bp->b_npages) {
1430 object = bp->b_pages[0]->object;
1431 VM_OBJECT_LOCK(object);
1432 }
1433 vm_page_lock_queues();
1434 /*
1435 * cleanup pages. If an error occurs writing to swap, we are in
1436 * very serious trouble. If it happens to be a disk error, though,
1437 * we may be able to recover by reassigning the swap later on. So
1438 * in this case we remove the m->swapblk assignment for the page
1439 * but do not free it in the rlist. The errornous block(s) are thus
1440 * never reallocated as swap. Redirty the page and continue.
1441 */
1442 for (i = 0; i < bp->b_npages; ++i) {
1443 vm_page_t m = bp->b_pages[i];
1444
1445 vm_page_flag_clear(m, PG_SWAPINPROG);
1446
1447 if (bp->b_ioflags & BIO_ERROR) {
1448 /*
1449 * If an error occurs I'd love to throw the swapblk
1450 * away without freeing it back to swapspace, so it
1451 * can never be used again. But I can't from an
1452 * interrupt.
1453 */
1454 if (bp->b_iocmd == BIO_READ) {
1455 /*
1456 * When reading, reqpage needs to stay
1457 * locked for the parent, but all other
1458 * pages can be freed. We still want to
1459 * wakeup the parent waiting on the page,
1460 * though. ( also: pg_reqpage can be -1 and
1461 * not match anything ).
1462 *
1463 * We have to wake specifically requested pages
1464 * up too because we cleared PG_SWAPINPROG and
1465 * someone may be waiting for that.
1466 *
1467 * NOTE: for reads, m->dirty will probably
1468 * be overridden by the original caller of
1469 * getpages so don't play cute tricks here.
1470 *
1471 * XXX IT IS NOT LEGAL TO FREE THE PAGE HERE
1472 * AS THIS MESSES WITH object->memq, and it is
1473 * not legal to mess with object->memq from an
1474 * interrupt.
1475 */
1476 m->valid = 0;
1477 if (i != bp->b_pager.pg_reqpage)
1478 vm_page_free(m);
1479 else
1480 vm_page_flash(m);
1481 /*
1482 * If i == bp->b_pager.pg_reqpage, do not wake
1483 * the page up. The caller needs to.
1484 */
1485 } else {
1486 /*
1487 * If a write error occurs, reactivate page
1488 * so it doesn't clog the inactive list,
1489 * then finish the I/O.
1490 */
1491 vm_page_dirty(m);
1492 vm_page_activate(m);
1493 vm_page_io_finish(m);
1494 }
1495 } else if (bp->b_iocmd == BIO_READ) {
1496 /*
1497 * For read success, clear dirty bits. Nobody should
1498 * have this page mapped but don't take any chances,
1499 * make sure the pmap modify bits are also cleared.
1500 *
1501 * NOTE: for reads, m->dirty will probably be
1502 * overridden by the original caller of getpages so
1503 * we cannot set them in order to free the underlying
1504 * swap in a low-swap situation. I don't think we'd
1505 * want to do that anyway, but it was an optimization
1506 * that existed in the old swapper for a time before
1507 * it got ripped out due to precisely this problem.
1508 *
1509 * If not the requested page then deactivate it.
1510 *
1511 * Note that the requested page, reqpage, is left
1512 * busied, but we still have to wake it up. The
1513 * other pages are released (unbusied) by
1514 * vm_page_wakeup(). We do not set reqpage's
1515 * valid bits here, it is up to the caller.
1516 */
1517 pmap_clear_modify(m);
1518 m->valid = VM_PAGE_BITS_ALL;
1519 vm_page_undirty(m);
1520
1521 /*
1522 * We have to wake specifically requested pages
1523 * up too because we cleared PG_SWAPINPROG and
1524 * could be waiting for it in getpages. However,
1525 * be sure to not unbusy getpages specifically
1526 * requested page - getpages expects it to be
1527 * left busy.
1528 */
1529 if (i != bp->b_pager.pg_reqpage) {
1530 vm_page_deactivate(m);
1531 vm_page_wakeup(m);
1532 } else {
1533 vm_page_flash(m);
1534 }
1535 } else {
1536 /*
1537 * For write success, clear the modify and dirty
1538 * status, then finish the I/O ( which decrements the
1539 * busy count and possibly wakes waiter's up ).
1540 */
1541 pmap_clear_modify(m);
1542 vm_page_undirty(m);
1543 vm_page_io_finish(m);
1544 if (vm_page_count_severe())
1545 vm_page_try_to_cache(m);
1546 }
1547 }
1548 vm_page_unlock_queues();
1549
1550 /*
1551 * adjust pip. NOTE: the original parent may still have its own
1552 * pip refs on the object.
1553 */
1554 if (object != NULL) {
1555 vm_object_pip_wakeupn(object, bp->b_npages);
1556 VM_OBJECT_UNLOCK(object);
1557 }
1558
1559 /*
1560 * release the physical I/O buffer
1561 */
1562 relpbuf(
1563 bp,
1564 ((bp->b_iocmd == BIO_READ) ? &nsw_rcount :
1565 ((bp->b_flags & B_ASYNC) ?
1566 &nsw_wcount_async :
1567 &nsw_wcount_sync
1568 )
1569 )
1570 );
1571 splx(s);
1572 }
1573
1574 /*
1575 * swap_pager_isswapped:
1576 *
1577 * Return 1 if at least one page in the given object is paged
1578 * out to the given swap device.
1579 *
1580 * This routine may not block.
1581 */
1582 int
1583 swap_pager_isswapped(vm_object_t object, struct swdevt *sp)
1584 {
1585 daddr_t index = 0;
1586 int bcount;
1587 int i;
1588
1589 VM_OBJECT_LOCK_ASSERT(object, MA_OWNED);
1590 if (object->type != OBJT_SWAP)
1591 return (0);
1592
1593 mtx_lock(&swhash_mtx);
1594 for (bcount = 0; bcount < object->un_pager.swp.swp_bcount; bcount++) {
1595 struct swblock *swap;
1596
1597 if ((swap = *swp_pager_hash(object, index)) != NULL) {
1598 for (i = 0; i < SWAP_META_PAGES; ++i) {
1599 if (swp_pager_isondev(swap->swb_pages[i], sp)) {
1600 mtx_unlock(&swhash_mtx);
1601 return (1);
1602 }
1603 }
1604 }
1605 index += SWAP_META_PAGES;
1606 if (index > 0x20000000)
1607 panic("swap_pager_isswapped: failed to locate all swap meta blocks");
1608 }
1609 mtx_unlock(&swhash_mtx);
1610 return (0);
1611 }
1612
1613 /*
1614 * SWP_PAGER_FORCE_PAGEIN() - force a swap block to be paged in
1615 *
1616 * This routine dissociates the page at the given index within a
1617 * swap block from its backing store, paging it in if necessary.
1618 * If the page is paged in, it is placed in the inactive queue,
1619 * since it had its backing store ripped out from under it.
1620 * We also attempt to swap in all other pages in the swap block,
1621 * we only guarantee that the one at the specified index is
1622 * paged in.
1623 *
1624 * XXX - The code to page the whole block in doesn't work, so we
1625 * revert to the one-by-one behavior for now. Sigh.
1626 */
1627 static __inline void
1628 swp_pager_force_pagein(vm_object_t object, vm_pindex_t pindex)
1629 {
1630 vm_page_t m;
1631
1632 vm_object_pip_add(object, 1);
1633 m = vm_page_grab(object, pindex, VM_ALLOC_NORMAL|VM_ALLOC_RETRY);
1634 if (m->valid == VM_PAGE_BITS_ALL) {
1635 vm_object_pip_subtract(object, 1);
1636 vm_page_lock_queues();
1637 vm_page_activate(m);
1638 vm_page_dirty(m);
1639 vm_page_wakeup(m);
1640 vm_page_unlock_queues();
1641 vm_pager_page_unswapped(m);
1642 return;
1643 }
1644
1645 if (swap_pager_getpages(object, &m, 1, 0) != VM_PAGER_OK)
1646 panic("swap_pager_force_pagein: read from swap failed");/*XXX*/
1647 vm_object_pip_subtract(object, 1);
1648 vm_page_lock_queues();
1649 vm_page_dirty(m);
1650 vm_page_dontneed(m);
1651 vm_page_wakeup(m);
1652 vm_page_unlock_queues();
1653 vm_pager_page_unswapped(m);
1654 }
1655
1656 /*
1657 * swap_pager_swapoff:
1658 *
1659 * Page in all of the pages that have been paged out to the
1660 * given device. The corresponding blocks in the bitmap must be
1661 * marked as allocated and the device must be flagged SW_CLOSING.
1662 * There may be no processes swapped out to the device.
1663 *
1664 * This routine may block.
1665 */
1666 static void
1667 swap_pager_swapoff(struct swdevt *sp)
1668 {
1669 struct swblock *swap;
1670 int i, j, retries;
1671
1672 GIANT_REQUIRED;
1673
1674 retries = 0;
1675 full_rescan:
1676 mtx_lock(&swhash_mtx);
1677 for (i = 0; i <= swhash_mask; i++) { /* '<=' is correct here */
1678 restart:
1679 for (swap = swhash[i]; swap != NULL; swap = swap->swb_hnext) {
1680 vm_object_t object = swap->swb_object;
1681 vm_pindex_t pindex = swap->swb_index;
1682 for (j = 0; j < SWAP_META_PAGES; ++j) {
1683 if (swp_pager_isondev(swap->swb_pages[j], sp)) {
1684 /* avoid deadlock */
1685 if (!VM_OBJECT_TRYLOCK(object)) {
1686 break;
1687 } else {
1688 mtx_unlock(&swhash_mtx);
1689 swp_pager_force_pagein(object,
1690 pindex + j);
1691 VM_OBJECT_UNLOCK(object);
1692 mtx_lock(&swhash_mtx);
1693 goto restart;
1694 }
1695 }
1696 }
1697 }
1698 }
1699 mtx_unlock(&swhash_mtx);
1700 if (sp->sw_used) {
1701 int dummy;
1702 /*
1703 * Objects may be locked or paging to the device being
1704 * removed, so we will miss their pages and need to
1705 * make another pass. We have marked this device as
1706 * SW_CLOSING, so the activity should finish soon.
1707 */
1708 retries++;
1709 if (retries > 100) {
1710 panic("swapoff: failed to locate %d swap blocks",
1711 sp->sw_used);
1712 }
1713 tsleep(&dummy, PVM, "swpoff", hz / 20);
1714 goto full_rescan;
1715 }
1716 }
1717
1718 /************************************************************************
1719 * SWAP META DATA *
1720 ************************************************************************
1721 *
1722 * These routines manipulate the swap metadata stored in the
1723 * OBJT_SWAP object. All swp_*() routines must be called at
1724 * splvm() because swap can be freed up by the low level vm_page
1725 * code which might be called from interrupts beyond what splbio() covers.
1726 *
1727 * Swap metadata is implemented with a global hash and not directly
1728 * linked into the object. Instead the object simply contains
1729 * appropriate tracking counters.
1730 */
1731
1732 /*
1733 * SWP_PAGER_META_BUILD() - add swap block to swap meta data for object
1734 *
1735 * We first convert the object to a swap object if it is a default
1736 * object.
1737 *
1738 * The specified swapblk is added to the object's swap metadata. If
1739 * the swapblk is not valid, it is freed instead. Any previously
1740 * assigned swapblk is freed.
1741 *
1742 * This routine must be called at splvm(), except when used to convert
1743 * an OBJT_DEFAULT object into an OBJT_SWAP object.
1744 */
1745 static void
1746 swp_pager_meta_build(vm_object_t object, vm_pindex_t pindex, daddr_t swapblk)
1747 {
1748 struct swblock *swap;
1749 struct swblock **pswap;
1750 int idx;
1751
1752 VM_OBJECT_LOCK_ASSERT(object, MA_OWNED);
1753 /*
1754 * Convert default object to swap object if necessary
1755 */
1756 if (object->type != OBJT_SWAP) {
1757 object->type = OBJT_SWAP;
1758 object->un_pager.swp.swp_bcount = 0;
1759
1760 if (object->handle != NULL) {
1761 mtx_lock(&sw_alloc_mtx);
1762 TAILQ_INSERT_TAIL(
1763 NOBJLIST(object->handle),
1764 object,
1765 pager_object_list
1766 );
1767 mtx_unlock(&sw_alloc_mtx);
1768 }
1769 }
1770
1771 /*
1772 * Locate hash entry. If not found create, but if we aren't adding
1773 * anything just return. If we run out of space in the map we wait
1774 * and, since the hash table may have changed, retry.
1775 */
1776 retry:
1777 mtx_lock(&swhash_mtx);
1778 pswap = swp_pager_hash(object, pindex);
1779
1780 if ((swap = *pswap) == NULL) {
1781 int i;
1782
1783 if (swapblk == SWAPBLK_NONE)
1784 goto done;
1785
1786 swap = *pswap = uma_zalloc(swap_zone, M_NOWAIT);
1787 if (swap == NULL) {
1788 mtx_unlock(&swhash_mtx);
1789 VM_OBJECT_UNLOCK(object);
1790 VM_WAIT;
1791 VM_OBJECT_LOCK(object);
1792 goto retry;
1793 }
1794
1795 swap->swb_hnext = NULL;
1796 swap->swb_object = object;
1797 swap->swb_index = pindex & ~(vm_pindex_t)SWAP_META_MASK;
1798 swap->swb_count = 0;
1799
1800 ++object->un_pager.swp.swp_bcount;
1801
1802 for (i = 0; i < SWAP_META_PAGES; ++i)
1803 swap->swb_pages[i] = SWAPBLK_NONE;
1804 }
1805
1806 /*
1807 * Delete prior contents of metadata
1808 */
1809 idx = pindex & SWAP_META_MASK;
1810
1811 if (swap->swb_pages[idx] != SWAPBLK_NONE) {
1812 swp_pager_freeswapspace(swap->swb_pages[idx], 1);
1813 --swap->swb_count;
1814 }
1815
1816 /*
1817 * Enter block into metadata
1818 */
1819 swap->swb_pages[idx] = swapblk;
1820 if (swapblk != SWAPBLK_NONE)
1821 ++swap->swb_count;
1822 done:
1823 mtx_unlock(&swhash_mtx);
1824 }
1825
1826 /*
1827 * SWP_PAGER_META_FREE() - free a range of blocks in the object's swap metadata
1828 *
1829 * The requested range of blocks is freed, with any associated swap
1830 * returned to the swap bitmap.
1831 *
1832 * This routine will free swap metadata structures as they are cleaned
1833 * out. This routine does *NOT* operate on swap metadata associated
1834 * with resident pages.
1835 *
1836 * This routine must be called at splvm()
1837 */
1838 static void
1839 swp_pager_meta_free(vm_object_t object, vm_pindex_t index, daddr_t count)
1840 {
1841
1842 VM_OBJECT_LOCK_ASSERT(object, MA_OWNED);
1843 if (object->type != OBJT_SWAP)
1844 return;
1845
1846 while (count > 0) {
1847 struct swblock **pswap;
1848 struct swblock *swap;
1849
1850 mtx_lock(&swhash_mtx);
1851 pswap = swp_pager_hash(object, index);
1852
1853 if ((swap = *pswap) != NULL) {
1854 daddr_t v = swap->swb_pages[index & SWAP_META_MASK];
1855
1856 if (v != SWAPBLK_NONE) {
1857 swp_pager_freeswapspace(v, 1);
1858 swap->swb_pages[index & SWAP_META_MASK] =
1859 SWAPBLK_NONE;
1860 if (--swap->swb_count == 0) {
1861 *pswap = swap->swb_hnext;
1862 uma_zfree(swap_zone, swap);
1863 --object->un_pager.swp.swp_bcount;
1864 }
1865 }
1866 --count;
1867 ++index;
1868 } else {
1869 int n = SWAP_META_PAGES - (index & SWAP_META_MASK);
1870 count -= n;
1871 index += n;
1872 }
1873 mtx_unlock(&swhash_mtx);
1874 }
1875 }
1876
1877 /*
1878 * SWP_PAGER_META_FREE_ALL() - destroy all swap metadata associated with object
1879 *
1880 * This routine locates and destroys all swap metadata associated with
1881 * an object.
1882 *
1883 * This routine must be called at splvm()
1884 */
1885 static void
1886 swp_pager_meta_free_all(vm_object_t object)
1887 {
1888 daddr_t index = 0;
1889
1890 VM_OBJECT_LOCK_ASSERT(object, MA_OWNED);
1891 if (object->type != OBJT_SWAP)
1892 return;
1893
1894 while (object->un_pager.swp.swp_bcount) {
1895 struct swblock **pswap;
1896 struct swblock *swap;
1897
1898 mtx_lock(&swhash_mtx);
1899 pswap = swp_pager_hash(object, index);
1900 if ((swap = *pswap) != NULL) {
1901 int i;
1902
1903 for (i = 0; i < SWAP_META_PAGES; ++i) {
1904 daddr_t v = swap->swb_pages[i];
1905 if (v != SWAPBLK_NONE) {
1906 --swap->swb_count;
1907 swp_pager_freeswapspace(v, 1);
1908 }
1909 }
1910 if (swap->swb_count != 0)
1911 panic("swap_pager_meta_free_all: swb_count != 0");
1912 *pswap = swap->swb_hnext;
1913 uma_zfree(swap_zone, swap);
1914 --object->un_pager.swp.swp_bcount;
1915 }
1916 mtx_unlock(&swhash_mtx);
1917 index += SWAP_META_PAGES;
1918 if (index > 0x20000000)
1919 panic("swp_pager_meta_free_all: failed to locate all swap meta blocks");
1920 }
1921 }
1922
1923 /*
1924 * SWP_PAGER_METACTL() - misc control of swap and vm_page_t meta data.
1925 *
1926 * This routine is capable of looking up, popping, or freeing
1927 * swapblk assignments in the swap meta data or in the vm_page_t.
1928 * The routine typically returns the swapblk being looked-up, or popped,
1929 * or SWAPBLK_NONE if the block was freed, or SWAPBLK_NONE if the block
1930 * was invalid. This routine will automatically free any invalid
1931 * meta-data swapblks.
1932 *
1933 * It is not possible to store invalid swapblks in the swap meta data
1934 * (other then a literal 'SWAPBLK_NONE'), so we don't bother checking.
1935 *
1936 * When acting on a busy resident page and paging is in progress, we
1937 * have to wait until paging is complete but otherwise can act on the
1938 * busy page.
1939 *
1940 * This routine must be called at splvm().
1941 *
1942 * SWM_FREE remove and free swap block from metadata
1943 * SWM_POP remove from meta data but do not free.. pop it out
1944 */
1945 static daddr_t
1946 swp_pager_meta_ctl(vm_object_t object, vm_pindex_t pindex, int flags)
1947 {
1948 struct swblock **pswap;
1949 struct swblock *swap;
1950 daddr_t r1;
1951 int idx;
1952
1953 VM_OBJECT_LOCK_ASSERT(object, MA_OWNED);
1954 /*
1955 * The meta data only exists of the object is OBJT_SWAP
1956 * and even then might not be allocated yet.
1957 */
1958 if (object->type != OBJT_SWAP)
1959 return (SWAPBLK_NONE);
1960
1961 r1 = SWAPBLK_NONE;
1962 mtx_lock(&swhash_mtx);
1963 pswap = swp_pager_hash(object, pindex);
1964
1965 if ((swap = *pswap) != NULL) {
1966 idx = pindex & SWAP_META_MASK;
1967 r1 = swap->swb_pages[idx];
1968
1969 if (r1 != SWAPBLK_NONE) {
1970 if (flags & SWM_FREE) {
1971 swp_pager_freeswapspace(r1, 1);
1972 r1 = SWAPBLK_NONE;
1973 }
1974 if (flags & (SWM_FREE|SWM_POP)) {
1975 swap->swb_pages[idx] = SWAPBLK_NONE;
1976 if (--swap->swb_count == 0) {
1977 *pswap = swap->swb_hnext;
1978 uma_zfree(swap_zone, swap);
1979 --object->un_pager.swp.swp_bcount;
1980 }
1981 }
1982 }
1983 }
1984 mtx_unlock(&swhash_mtx);
1985 return (r1);
1986 }
1987
1988 /*
1989 * System call swapon(name) enables swapping on device name,
1990 * which must be in the swdevsw. Return EBUSY
1991 * if already swapping on this device.
1992 */
1993 #ifndef _SYS_SYSPROTO_H_
1994 struct swapon_args {
1995 char *name;
1996 };
1997 #endif
1998
1999 /*
2000 * MPSAFE
2001 */
2002 /* ARGSUSED */
2003 int
2004 swapon(struct thread *td, struct swapon_args *uap)
2005 {
2006 struct vattr attr;
2007 struct vnode *vp;
2008 struct nameidata nd;
2009 int error;
2010
2011 mtx_lock(&Giant);
2012 error = suser(td);
2013 if (error)
2014 goto done2;
2015
2016 while (swdev_syscall_active)
2017 tsleep(&swdev_syscall_active, PUSER - 1, "swpon", 0);
2018 swdev_syscall_active = 1;
2019
2020 /*
2021 * Swap metadata may not fit in the KVM if we have physical
2022 * memory of >1GB.
2023 */
2024 if (swap_zone == NULL) {
2025 error = ENOMEM;
2026 goto done;
2027 }
2028
2029 NDINIT(&nd, LOOKUP, FOLLOW, UIO_USERSPACE, uap->name, td);
2030 error = namei(&nd);
2031 if (error)
2032 goto done;
2033
2034 NDFREE(&nd, NDF_ONLY_PNBUF);
2035 vp = nd.ni_vp;
2036
2037 if (vn_isdisk(vp, &error)) {
2038 error = swapongeom(td, vp);
2039 } else if (vp->v_type == VREG &&
2040 (vp->v_mount->mnt_vfc->vfc_flags & VFCF_NETWORK) != 0 &&
2041 (error = VOP_GETATTR(vp, &attr, td->td_ucred, td)) == 0) {
2042 /*
2043 * Allow direct swapping to NFS regular files in the same
2044 * way that nfs_mountroot() sets up diskless swapping.
2045 */
2046 error = swaponvp(td, vp, attr.va_size / DEV_BSIZE);
2047 }
2048
2049 if (error)
2050 vrele(vp);
2051 done:
2052 swdev_syscall_active = 0;
2053 wakeup_one(&swdev_syscall_active);
2054 done2:
2055 mtx_unlock(&Giant);
2056 return (error);
2057 }
2058
2059 static void
2060 swaponsomething(struct vnode *vp, void *id, u_long nblks, sw_strategy_t *strategy, sw_close_t *close, dev_t dev)
2061 {
2062 struct swdevt *sp, *tsp;
2063 swblk_t dvbase;
2064 u_long mblocks;
2065
2066 /*
2067 * If we go beyond this, we get overflows in the radix
2068 * tree bitmap code.
2069 */
2070 mblocks = 0x40000000 / BLIST_META_RADIX;
2071 if (nblks > mblocks) {
2072 printf("WARNING: reducing size to maximum of %lu blocks per swap unit\n",
2073 mblocks);
2074 nblks = mblocks;
2075 }
2076 /*
2077 * nblks is in DEV_BSIZE'd chunks, convert to PAGE_SIZE'd chunks.
2078 * First chop nblks off to page-align it, then convert.
2079 *
2080 * sw->sw_nblks is in page-sized chunks now too.
2081 */
2082 nblks &= ~(ctodb(1) - 1);
2083 nblks = dbtoc(nblks);
2084
2085 sp = malloc(sizeof *sp, M_VMPGDATA, M_WAITOK | M_ZERO);
2086 sp->sw_vp = vp;
2087 sp->sw_id = id;
2088 sp->sw_dev = dev;
2089 sp->sw_flags = 0;
2090 sp->sw_nblks = nblks;
2091 sp->sw_used = 0;
2092 sp->sw_strategy = strategy;
2093 sp->sw_close = close;
2094
2095 sp->sw_blist = blist_create(nblks);
2096 /*
2097 * Do not free the first two block in order to avoid overwriting
2098 * any bsd label at the front of the partition
2099 */
2100 blist_free(sp->sw_blist, 2, nblks - 2);
2101
2102 dvbase = 0;
2103 mtx_lock(&sw_dev_mtx);
2104 TAILQ_FOREACH(tsp, &swtailq, sw_list) {
2105 if (tsp->sw_end >= dvbase) {
2106 /*
2107 * We put one uncovered page between the devices
2108 * in order to definitively prevent any cross-device
2109 * I/O requests
2110 */
2111 dvbase = tsp->sw_end + 1;
2112 }
2113 }
2114 sp->sw_first = dvbase;
2115 sp->sw_end = dvbase + nblks;
2116 TAILQ_INSERT_TAIL(&swtailq, sp, sw_list);
2117 nswapdev++;
2118 swap_pager_avail += nblks;
2119 swp_sizecheck();
2120 mtx_unlock(&sw_dev_mtx);
2121 }
2122
2123 /*
2124 * SYSCALL: swapoff(devname)
2125 *
2126 * Disable swapping on the given device.
2127 *
2128 * XXX: Badly designed system call: it should use a device index
2129 * rather than filename as specification. We keep sw_vp around
2130 * only to make this work.
2131 */
2132 #ifndef _SYS_SYSPROTO_H_
2133 struct swapoff_args {
2134 char *name;
2135 };
2136 #endif
2137
2138 /*
2139 * MPSAFE
2140 */
2141 /* ARGSUSED */
2142 int
2143 swapoff(struct thread *td, struct swapoff_args *uap)
2144 {
2145 struct vnode *vp;
2146 struct nameidata nd;
2147 struct swdevt *sp;
2148 u_long nblks, dvbase;
2149 int error;
2150
2151 mtx_lock(&Giant);
2152
2153 error = suser(td);
2154 if (error)
2155 goto done2;
2156
2157 while (swdev_syscall_active)
2158 tsleep(&swdev_syscall_active, PUSER - 1, "swpoff", 0);
2159 swdev_syscall_active = 1;
2160
2161 NDINIT(&nd, LOOKUP, FOLLOW, UIO_USERSPACE, uap->name, td);
2162 error = namei(&nd);
2163 if (error)
2164 goto done;
2165 NDFREE(&nd, NDF_ONLY_PNBUF);
2166 vp = nd.ni_vp;
2167
2168 mtx_lock(&sw_dev_mtx);
2169 TAILQ_FOREACH(sp, &swtailq, sw_list) {
2170 if (sp->sw_vp == vp)
2171 goto found;
2172 }
2173 mtx_unlock(&sw_dev_mtx);
2174 error = EINVAL;
2175 goto done;
2176 found:
2177 mtx_unlock(&sw_dev_mtx);
2178 #ifdef MAC
2179 (void) vn_lock(vp, LK_EXCLUSIVE | LK_RETRY, td);
2180 error = mac_check_system_swapoff(td->td_ucred, vp);
2181 (void) VOP_UNLOCK(vp, 0, td);
2182 if (error != 0)
2183 goto done;
2184 #endif
2185
2186 nblks = sp->sw_nblks;
2187
2188 /*
2189 * We can turn off this swap device safely only if the
2190 * available virtual memory in the system will fit the amount
2191 * of data we will have to page back in, plus an epsilon so
2192 * the system doesn't become critically low on swap space.
2193 */
2194 if (cnt.v_free_count + cnt.v_cache_count + swap_pager_avail <
2195 nblks + nswap_lowat) {
2196 error = ENOMEM;
2197 goto done;
2198 }
2199
2200 /*
2201 * Prevent further allocations on this device.
2202 */
2203 mtx_lock(&sw_dev_mtx);
2204 sp->sw_flags |= SW_CLOSING;
2205 for (dvbase = 0; dvbase < sp->sw_end; dvbase += dmmax) {
2206 swap_pager_avail -= blist_fill(sp->sw_blist,
2207 dvbase, dmmax);
2208 }
2209 mtx_unlock(&sw_dev_mtx);
2210
2211 /*
2212 * Page in the contents of the device and close it.
2213 */
2214 swap_pager_swapoff(sp);
2215
2216 sp->sw_close(td, sp);
2217 sp->sw_id = NULL;
2218 mtx_lock(&sw_dev_mtx);
2219 TAILQ_REMOVE(&swtailq, sp, sw_list);
2220 nswapdev--;
2221 if (nswapdev == 0) {
2222 swap_pager_full = 2;
2223 swap_pager_almost_full = 1;
2224 }
2225 if (swdevhd == sp)
2226 swdevhd = NULL;
2227 mtx_unlock(&sw_dev_mtx);
2228 blist_destroy(sp->sw_blist);
2229 free(sp, M_VMPGDATA);
2230
2231 done:
2232 swdev_syscall_active = 0;
2233 wakeup_one(&swdev_syscall_active);
2234 done2:
2235 mtx_unlock(&Giant);
2236 return (error);
2237 }
2238
2239 void
2240 swap_pager_status(int *total, int *used)
2241 {
2242 struct swdevt *sp;
2243
2244 *total = 0;
2245 *used = 0;
2246 mtx_lock(&sw_dev_mtx);
2247 TAILQ_FOREACH(sp, &swtailq, sw_list) {
2248 *total += sp->sw_nblks;
2249 *used += sp->sw_used;
2250 }
2251 mtx_unlock(&sw_dev_mtx);
2252 }
2253
2254 static int
2255 sysctl_vm_swap_info(SYSCTL_HANDLER_ARGS)
2256 {
2257 int *name = (int *)arg1;
2258 int error, n;
2259 struct xswdev xs;
2260 struct swdevt *sp;
2261
2262 if (arg2 != 1) /* name length */
2263 return (EINVAL);
2264
2265 n = 0;
2266 mtx_lock(&sw_dev_mtx);
2267 TAILQ_FOREACH(sp, &swtailq, sw_list) {
2268 if (n == *name) {
2269 mtx_unlock(&sw_dev_mtx);
2270 xs.xsw_version = XSWDEV_VERSION;
2271 xs.xsw_dev = sp->sw_dev;
2272 xs.xsw_flags = sp->sw_flags;
2273 xs.xsw_nblks = sp->sw_nblks;
2274 xs.xsw_used = sp->sw_used;
2275
2276 error = SYSCTL_OUT(req, &xs, sizeof(xs));
2277 return (error);
2278 }
2279 n++;
2280 }
2281 mtx_unlock(&sw_dev_mtx);
2282 return (ENOENT);
2283 }
2284
2285 SYSCTL_INT(_vm, OID_AUTO, nswapdev, CTLFLAG_RD, &nswapdev, 0,
2286 "Number of swap devices");
2287 SYSCTL_NODE(_vm, OID_AUTO, swap_info, CTLFLAG_RD, sysctl_vm_swap_info,
2288 "Swap statistics by device");
2289
2290 /*
2291 * vmspace_swap_count() - count the approximate swap useage in pages for a
2292 * vmspace.
2293 *
2294 * The map must be locked.
2295 *
2296 * Swap useage is determined by taking the proportional swap used by
2297 * VM objects backing the VM map. To make up for fractional losses,
2298 * if the VM object has any swap use at all the associated map entries
2299 * count for at least 1 swap page.
2300 */
2301 int
2302 vmspace_swap_count(struct vmspace *vmspace)
2303 {
2304 vm_map_t map = &vmspace->vm_map;
2305 vm_map_entry_t cur;
2306 int count = 0;
2307
2308 for (cur = map->header.next; cur != &map->header; cur = cur->next) {
2309 vm_object_t object;
2310
2311 if ((cur->eflags & MAP_ENTRY_IS_SUB_MAP) == 0 &&
2312 (object = cur->object.vm_object) != NULL) {
2313 VM_OBJECT_LOCK(object);
2314 if (object->type == OBJT_SWAP &&
2315 object->un_pager.swp.swp_bcount != 0) {
2316 int n = (cur->end - cur->start) / PAGE_SIZE;
2317
2318 count += object->un_pager.swp.swp_bcount *
2319 SWAP_META_PAGES * n / object->size + 1;
2320 }
2321 VM_OBJECT_UNLOCK(object);
2322 }
2323 }
2324 return (count);
2325 }
2326
2327 /*
2328 * GEOM backend
2329 *
2330 * Swapping onto disk devices.
2331 *
2332 */
2333
2334 static g_orphan_t swapgeom_orphan;
2335
2336 static struct g_class g_swap_class = {
2337 .name = "SWAP",
2338 .version = G_VERSION,
2339 .orphan = swapgeom_orphan,
2340 };
2341
2342 DECLARE_GEOM_CLASS(g_swap_class, g_class);
2343
2344
2345 static void
2346 swapgeom_done(struct bio *bp2)
2347 {
2348 struct buf *bp;
2349
2350 bp = bp2->bio_caller2;
2351 if (bp2->bio_error)
2352 bp->b_ioflags |= BIO_ERROR;
2353 bufdone(bp);
2354 g_destroy_bio(bp2);
2355 }
2356
2357 static void
2358 swapgeom_strategy(struct buf *bp, struct swdevt *sp)
2359 {
2360 struct bio *bio;
2361 struct g_consumer *cp;
2362
2363 cp = sp->sw_id;
2364 if (cp == NULL) {
2365 bp->b_error = ENXIO;
2366 bp->b_ioflags |= BIO_ERROR;
2367 bufdone(bp);
2368 return;
2369 }
2370 bio = g_clone_bio(&bp->b_io);
2371 if (bio == NULL) {
2372 /*
2373 * XXX: This is better than panicing, but not much better.
2374 * XXX: Somehow this should be retried. A more generic
2375 * XXX: implementation of ENOMEM in geom may be able to cope.
2376 */
2377 bp->b_error = ENOMEM;
2378 bp->b_ioflags |= BIO_ERROR;
2379 bufdone(bp);
2380 return;
2381 }
2382 bio->bio_caller2 = bp;
2383 bio->bio_offset = (bp->b_blkno - sp->sw_first) * PAGE_SIZE;
2384 bio->bio_length = bp->b_bcount;
2385 bio->bio_done = swapgeom_done;
2386 g_io_request(bio, cp);
2387 return;
2388 }
2389
2390 static void
2391 swapgeom_orphan(struct g_consumer *cp)
2392 {
2393 struct swdevt *sp;
2394
2395 mtx_lock(&sw_dev_mtx);
2396 TAILQ_FOREACH(sp, &swtailq, sw_list)
2397 if (sp->sw_id == cp)
2398 sp->sw_id = NULL;
2399 mtx_unlock(&sw_dev_mtx);
2400 }
2401
2402 static void
2403 swapgeom_close_ev(void *arg, int flags)
2404 {
2405 struct g_consumer *cp;
2406
2407 cp = arg;
2408 g_access(cp, -1, -1, 0);
2409 g_detach(cp);
2410 g_destroy_consumer(cp);
2411 }
2412
2413 static void
2414 swapgeom_close(struct thread *td, struct swdevt *sw)
2415 {
2416
2417 /* XXX: direct call when Giant untangled */
2418 g_waitfor_event(swapgeom_close_ev, sw->sw_id, M_WAITOK, NULL);
2419 }
2420
2421
2422 struct swh0h0 {
2423 struct cdev *dev;
2424 struct vnode *vp;
2425 int error;
2426 };
2427
2428 static void
2429 swapongeom_ev(void *arg, int flags)
2430 {
2431 struct swh0h0 *swh;
2432 struct g_provider *pp;
2433 struct g_consumer *cp;
2434 static struct g_geom *gp;
2435 struct swdevt *sp;
2436 u_long nblks;
2437 int error;
2438
2439 swh = arg;
2440 swh->error = 0;
2441 pp = g_dev_getprovider(swh->dev);
2442 if (pp == NULL) {
2443 swh->error = ENODEV;
2444 return;
2445 }
2446 mtx_lock(&sw_dev_mtx);
2447 TAILQ_FOREACH(sp, &swtailq, sw_list) {
2448 cp = sp->sw_id;
2449 if (cp != NULL && cp->provider == pp) {
2450 mtx_unlock(&sw_dev_mtx);
2451 swh->error = EBUSY;
2452 return;
2453 }
2454 }
2455 mtx_unlock(&sw_dev_mtx);
2456 if (gp == NULL)
2457 gp = g_new_geomf(&g_swap_class, "swap", NULL);
2458 cp = g_new_consumer(gp);
2459 g_attach(cp, pp);
2460 /*
2461 * XXX: Everytime you think you can improve the margin for
2462 * footshooting, somebody depends on the ability to do so:
2463 * savecore(8) wants to write to our swapdev so we cannot
2464 * set an exclusive count :-(
2465 */
2466 error = g_access(cp, 1, 1, 0);
2467 if (error) {
2468 g_detach(cp);
2469 g_destroy_consumer(cp);
2470 swh->error = error;
2471 return;
2472 }
2473 nblks = pp->mediasize / DEV_BSIZE;
2474 swaponsomething(swh->vp, cp, nblks, swapgeom_strategy,
2475 swapgeom_close, dev2udev(swh->dev));
2476 swh->error = 0;
2477 return;
2478 }
2479
2480 static int
2481 swapongeom(struct thread *td, struct vnode *vp)
2482 {
2483 int error;
2484 struct swh0h0 swh;
2485
2486 vn_lock(vp, LK_EXCLUSIVE | LK_RETRY, td);
2487
2488 swh.dev = vp->v_rdev;
2489 swh.vp = vp;
2490 swh.error = 0;
2491 /* XXX: direct call when Giant untangled */
2492 error = g_waitfor_event(swapongeom_ev, &swh, M_WAITOK, NULL);
2493 if (!error)
2494 error = swh.error;
2495 VOP_UNLOCK(vp, 0, td);
2496 return (error);
2497 }
2498
2499 /*
2500 * VNODE backend
2501 *
2502 * This is used mainly for network filesystem (read: probably only tested
2503 * with NFS) swapfiles.
2504 *
2505 */
2506
2507 static void
2508 swapdev_strategy(struct buf *bp, struct swdevt *sp)
2509 {
2510 int s;
2511 struct vnode *vp, *vp2;
2512
2513 bp->b_dev = NULL;
2514 bp->b_blkno = ctodb(bp->b_blkno - sp->sw_first);
2515
2516 vp2 = sp->sw_id;
2517 vhold(vp2);
2518 s = splvm();
2519 if (bp->b_iocmd == BIO_WRITE) {
2520 vp = bp->b_vp;
2521 if (vp) {
2522 VI_LOCK(vp);
2523 vp->v_numoutput--;
2524 if ((vp->v_iflag & VI_BWAIT) && vp->v_numoutput <= 0) {
2525 vp->v_iflag &= ~VI_BWAIT;
2526 wakeup(&vp->v_numoutput);
2527 }
2528 VI_UNLOCK(vp);
2529 }
2530 VI_LOCK(vp2);
2531 vp2->v_numoutput++;
2532 VI_UNLOCK(vp2);
2533 }
2534 bp->b_vp = vp2;
2535 splx(s);
2536 bp->b_iooffset = dbtob(bp->b_blkno);
2537 VOP_STRATEGY(vp2, bp);
2538 return;
2539 }
2540
2541 static void
2542 swapdev_close(struct thread *td, struct swdevt *sp)
2543 {
2544
2545 VOP_CLOSE(sp->sw_vp, FREAD | FWRITE, td->td_ucred, td);
2546 vrele(sp->sw_vp);
2547 }
2548
2549
2550 static int
2551 swaponvp(struct thread *td, struct vnode *vp, u_long nblks)
2552 {
2553 struct swdevt *sp;
2554 int error;
2555
2556 if (nblks == 0)
2557 return (ENXIO);
2558 mtx_lock(&sw_dev_mtx);
2559 TAILQ_FOREACH(sp, &swtailq, sw_list) {
2560 if (sp->sw_id == vp) {
2561 mtx_unlock(&sw_dev_mtx);
2562 return (EBUSY);
2563 }
2564 }
2565 mtx_unlock(&sw_dev_mtx);
2566
2567 (void) vn_lock(vp, LK_EXCLUSIVE | LK_RETRY, td);
2568 #ifdef MAC
2569 error = mac_check_system_swapon(td->td_ucred, vp);
2570 if (error == 0)
2571 #endif
2572 error = VOP_OPEN(vp, FREAD | FWRITE, td->td_ucred, td, -1);
2573 (void) VOP_UNLOCK(vp, 0, td);
2574 if (error)
2575 return (error);
2576
2577 swaponsomething(vp, vp, nblks, swapdev_strategy, swapdev_close,
2578 NODEV);
2579 return (0);
2580 }
Cache object: af11d2fe8390ea4918c627ddcd07c3f1
|