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/6.2/sys/vm/swap_pager.c 160454 2006-07-17 21:23:54Z cognet $");
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 static int swapoff_one(struct swdevt *sp, struct thread *td);
270
271 /*
272 * Swap bitmap functions
273 */
274 static void swp_pager_freeswapspace(daddr_t blk, int npages);
275 static daddr_t swp_pager_getswapspace(int npages);
276
277 /*
278 * Metadata functions
279 */
280 static struct swblock **swp_pager_hash(vm_object_t object, vm_pindex_t index);
281 static void swp_pager_meta_build(vm_object_t, vm_pindex_t, daddr_t);
282 static void swp_pager_meta_free(vm_object_t, vm_pindex_t, daddr_t);
283 static void swp_pager_meta_free_all(vm_object_t);
284 static daddr_t swp_pager_meta_ctl(vm_object_t, vm_pindex_t, int);
285
286 /*
287 * SWP_SIZECHECK() - update swap_pager_full indication
288 *
289 * update the swap_pager_almost_full indication and warn when we are
290 * about to run out of swap space, using lowat/hiwat hysteresis.
291 *
292 * Clear swap_pager_full ( task killing ) indication when lowat is met.
293 *
294 * No restrictions on call
295 * This routine may not block.
296 * This routine must be called at splvm()
297 */
298 static void
299 swp_sizecheck(void)
300 {
301
302 if (swap_pager_avail < nswap_lowat) {
303 if (swap_pager_almost_full == 0) {
304 printf("swap_pager: out of swap space\n");
305 swap_pager_almost_full = 1;
306 }
307 } else {
308 swap_pager_full = 0;
309 if (swap_pager_avail > nswap_hiwat)
310 swap_pager_almost_full = 0;
311 }
312 }
313
314 /*
315 * SWP_PAGER_HASH() - hash swap meta data
316 *
317 * This is an helper function which hashes the swapblk given
318 * the object and page index. It returns a pointer to a pointer
319 * to the object, or a pointer to a NULL pointer if it could not
320 * find a swapblk.
321 *
322 * This routine must be called at splvm().
323 */
324 static struct swblock **
325 swp_pager_hash(vm_object_t object, vm_pindex_t index)
326 {
327 struct swblock **pswap;
328 struct swblock *swap;
329
330 index &= ~(vm_pindex_t)SWAP_META_MASK;
331 pswap = &swhash[(index ^ (int)(intptr_t)object) & swhash_mask];
332 while ((swap = *pswap) != NULL) {
333 if (swap->swb_object == object &&
334 swap->swb_index == index
335 ) {
336 break;
337 }
338 pswap = &swap->swb_hnext;
339 }
340 return (pswap);
341 }
342
343 /*
344 * SWAP_PAGER_INIT() - initialize the swap pager!
345 *
346 * Expected to be started from system init. NOTE: This code is run
347 * before much else so be careful what you depend on. Most of the VM
348 * system has yet to be initialized at this point.
349 */
350 static void
351 swap_pager_init(void)
352 {
353 /*
354 * Initialize object lists
355 */
356 int i;
357
358 for (i = 0; i < NOBJLISTS; ++i)
359 TAILQ_INIT(&swap_pager_object_list[i]);
360 mtx_init(&sw_alloc_mtx, "swap_pager list", NULL, MTX_DEF);
361 mtx_init(&sw_dev_mtx, "swapdev", NULL, MTX_DEF);
362
363 /*
364 * Device Stripe, in PAGE_SIZE'd blocks
365 */
366 dmmax = SWB_NPAGES * 2;
367 }
368
369 /*
370 * SWAP_PAGER_SWAP_INIT() - swap pager initialization from pageout process
371 *
372 * Expected to be started from pageout process once, prior to entering
373 * its main loop.
374 */
375 void
376 swap_pager_swap_init(void)
377 {
378 int n, n2;
379
380 /*
381 * Number of in-transit swap bp operations. Don't
382 * exhaust the pbufs completely. Make sure we
383 * initialize workable values (0 will work for hysteresis
384 * but it isn't very efficient).
385 *
386 * The nsw_cluster_max is constrained by the bp->b_pages[]
387 * array (MAXPHYS/PAGE_SIZE) and our locally defined
388 * MAX_PAGEOUT_CLUSTER. Also be aware that swap ops are
389 * constrained by the swap device interleave stripe size.
390 *
391 * Currently we hardwire nsw_wcount_async to 4. This limit is
392 * designed to prevent other I/O from having high latencies due to
393 * our pageout I/O. The value 4 works well for one or two active swap
394 * devices but is probably a little low if you have more. Even so,
395 * a higher value would probably generate only a limited improvement
396 * with three or four active swap devices since the system does not
397 * typically have to pageout at extreme bandwidths. We will want
398 * at least 2 per swap devices, and 4 is a pretty good value if you
399 * have one NFS swap device due to the command/ack latency over NFS.
400 * So it all works out pretty well.
401 */
402 nsw_cluster_max = min((MAXPHYS/PAGE_SIZE), MAX_PAGEOUT_CLUSTER);
403
404 mtx_lock(&pbuf_mtx);
405 nsw_rcount = (nswbuf + 1) / 2;
406 nsw_wcount_sync = (nswbuf + 3) / 4;
407 nsw_wcount_async = 4;
408 nsw_wcount_async_max = nsw_wcount_async;
409 mtx_unlock(&pbuf_mtx);
410
411 /*
412 * Initialize our zone. Right now I'm just guessing on the number
413 * we need based on the number of pages in the system. Each swblock
414 * can hold 16 pages, so this is probably overkill. This reservation
415 * is typically limited to around 32MB by default.
416 */
417 n = cnt.v_page_count / 2;
418 if (maxswzone && n > maxswzone / sizeof(struct swblock))
419 n = maxswzone / sizeof(struct swblock);
420 n2 = n;
421 swap_zone = uma_zcreate("SWAPMETA", sizeof(struct swblock), NULL, NULL,
422 NULL, NULL, UMA_ALIGN_PTR, UMA_ZONE_NOFREE | UMA_ZONE_VM);
423 if (swap_zone == NULL)
424 panic("failed to create swap_zone.");
425 do {
426 if (uma_zone_set_obj(swap_zone, &swap_zone_obj, n))
427 break;
428 /*
429 * if the allocation failed, try a zone two thirds the
430 * size of the previous attempt.
431 */
432 n -= ((n + 2) / 3);
433 } while (n > 0);
434 if (n2 != n)
435 printf("Swap zone entries reduced from %d to %d.\n", n2, n);
436 n2 = n;
437
438 /*
439 * Initialize our meta-data hash table. The swapper does not need to
440 * be quite as efficient as the VM system, so we do not use an
441 * oversized hash table.
442 *
443 * n: size of hash table, must be power of 2
444 * swhash_mask: hash table index mask
445 */
446 for (n = 1; n < n2 / 8; n *= 2)
447 ;
448 swhash = malloc(sizeof(struct swblock *) * n, M_VMPGDATA, M_WAITOK | M_ZERO);
449 swhash_mask = n - 1;
450 mtx_init(&swhash_mtx, "swap_pager swhash", NULL, MTX_DEF);
451 }
452
453 /*
454 * SWAP_PAGER_ALLOC() - allocate a new OBJT_SWAP VM object and instantiate
455 * its metadata structures.
456 *
457 * This routine is called from the mmap and fork code to create a new
458 * OBJT_SWAP object. We do this by creating an OBJT_DEFAULT object
459 * and then converting it with swp_pager_meta_build().
460 *
461 * This routine may block in vm_object_allocate() and create a named
462 * object lookup race, so we must interlock. We must also run at
463 * splvm() for the object lookup to handle races with interrupts, but
464 * we do not have to maintain splvm() in between the lookup and the
465 * add because (I believe) it is not possible to attempt to create
466 * a new swap object w/handle when a default object with that handle
467 * already exists.
468 *
469 * MPSAFE
470 */
471 static vm_object_t
472 swap_pager_alloc(void *handle, vm_ooffset_t size, vm_prot_t prot,
473 vm_ooffset_t offset)
474 {
475 vm_object_t object;
476 vm_pindex_t pindex;
477
478 pindex = OFF_TO_IDX(offset + PAGE_MASK + size);
479
480 if (handle) {
481 mtx_lock(&Giant);
482 /*
483 * Reference existing named region or allocate new one. There
484 * should not be a race here against swp_pager_meta_build()
485 * as called from vm_page_remove() in regards to the lookup
486 * of the handle.
487 */
488 sx_xlock(&sw_alloc_sx);
489 object = vm_pager_object_lookup(NOBJLIST(handle), handle);
490
491 if (object != NULL) {
492 vm_object_reference(object);
493 } else {
494 object = vm_object_allocate(OBJT_DEFAULT, pindex);
495 object->handle = handle;
496
497 VM_OBJECT_LOCK(object);
498 swp_pager_meta_build(object, 0, SWAPBLK_NONE);
499 VM_OBJECT_UNLOCK(object);
500 }
501 sx_xunlock(&sw_alloc_sx);
502 mtx_unlock(&Giant);
503 } else {
504 object = vm_object_allocate(OBJT_DEFAULT, pindex);
505
506 VM_OBJECT_LOCK(object);
507 swp_pager_meta_build(object, 0, SWAPBLK_NONE);
508 VM_OBJECT_UNLOCK(object);
509 }
510 return (object);
511 }
512
513 /*
514 * SWAP_PAGER_DEALLOC() - remove swap metadata from object
515 *
516 * The swap backing for the object is destroyed. The code is
517 * designed such that we can reinstantiate it later, but this
518 * routine is typically called only when the entire object is
519 * about to be destroyed.
520 *
521 * This routine may block, but no longer does.
522 *
523 * The object must be locked or unreferenceable.
524 */
525 static void
526 swap_pager_dealloc(vm_object_t object)
527 {
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 swp_pager_meta_free_all(object);
549 }
550
551 /************************************************************************
552 * SWAP PAGER BITMAP ROUTINES *
553 ************************************************************************/
554
555 /*
556 * SWP_PAGER_GETSWAPSPACE() - allocate raw swap space
557 *
558 * Allocate swap for the requested number of pages. The starting
559 * swap block number (a page index) is returned or SWAPBLK_NONE
560 * if the allocation failed.
561 *
562 * Also has the side effect of advising that somebody made a mistake
563 * when they configured swap and didn't configure enough.
564 *
565 * Must be called at splvm() to avoid races with bitmap frees from
566 * vm_page_remove() aka swap_pager_page_removed().
567 *
568 * This routine may not block
569 * This routine must be called at splvm().
570 *
571 * We allocate in round-robin fashion from the configured devices.
572 */
573 static daddr_t
574 swp_pager_getswapspace(int npages)
575 {
576 daddr_t blk;
577 struct swdevt *sp;
578 int i;
579
580 blk = SWAPBLK_NONE;
581 mtx_lock(&sw_dev_mtx);
582 sp = swdevhd;
583 for (i = 0; i < nswapdev; i++) {
584 if (sp == NULL)
585 sp = TAILQ_FIRST(&swtailq);
586 if (!(sp->sw_flags & SW_CLOSING)) {
587 blk = blist_alloc(sp->sw_blist, npages);
588 if (blk != SWAPBLK_NONE) {
589 blk += sp->sw_first;
590 sp->sw_used += npages;
591 swap_pager_avail -= npages;
592 swp_sizecheck();
593 swdevhd = TAILQ_NEXT(sp, sw_list);
594 goto done;
595 }
596 }
597 sp = TAILQ_NEXT(sp, sw_list);
598 }
599 if (swap_pager_full != 2) {
600 printf("swap_pager_getswapspace(%d): failed\n", npages);
601 swap_pager_full = 2;
602 swap_pager_almost_full = 1;
603 }
604 swdevhd = NULL;
605 done:
606 mtx_unlock(&sw_dev_mtx);
607 return (blk);
608 }
609
610 static int
611 swp_pager_isondev(daddr_t blk, struct swdevt *sp)
612 {
613
614 return (blk >= sp->sw_first && blk < sp->sw_end);
615 }
616
617 static void
618 swp_pager_strategy(struct buf *bp)
619 {
620 struct swdevt *sp;
621
622 mtx_lock(&sw_dev_mtx);
623 TAILQ_FOREACH(sp, &swtailq, sw_list) {
624 if (bp->b_blkno >= sp->sw_first && bp->b_blkno < sp->sw_end) {
625 mtx_unlock(&sw_dev_mtx);
626 sp->sw_strategy(bp, sp);
627 return;
628 }
629 }
630 panic("Swapdev not found");
631 }
632
633
634 /*
635 * SWP_PAGER_FREESWAPSPACE() - free raw swap space
636 *
637 * This routine returns the specified swap blocks back to the bitmap.
638 *
639 * Note: This routine may not block (it could in the old swap code),
640 * and through the use of the new blist routines it does not block.
641 *
642 * We must be called at splvm() to avoid races with bitmap frees from
643 * vm_page_remove() aka swap_pager_page_removed().
644 *
645 * This routine may not block
646 * This routine must be called at splvm().
647 */
648 static void
649 swp_pager_freeswapspace(daddr_t blk, int npages)
650 {
651 struct swdevt *sp;
652
653 mtx_lock(&sw_dev_mtx);
654 TAILQ_FOREACH(sp, &swtailq, sw_list) {
655 if (blk >= sp->sw_first && blk < sp->sw_end) {
656 sp->sw_used -= npages;
657 /*
658 * If we are attempting to stop swapping on
659 * this device, we don't want to mark any
660 * blocks free lest they be reused.
661 */
662 if ((sp->sw_flags & SW_CLOSING) == 0) {
663 blist_free(sp->sw_blist, blk - sp->sw_first,
664 npages);
665 swap_pager_avail += npages;
666 swp_sizecheck();
667 }
668 mtx_unlock(&sw_dev_mtx);
669 return;
670 }
671 }
672 panic("Swapdev not found");
673 }
674
675 /*
676 * SWAP_PAGER_FREESPACE() - frees swap blocks associated with a page
677 * range within an object.
678 *
679 * This is a globally accessible routine.
680 *
681 * This routine removes swapblk assignments from swap metadata.
682 *
683 * The external callers of this routine typically have already destroyed
684 * or renamed vm_page_t's associated with this range in the object so
685 * we should be ok.
686 *
687 * This routine may be called at any spl. We up our spl to splvm temporarily
688 * in order to perform the metadata removal.
689 */
690 void
691 swap_pager_freespace(vm_object_t object, vm_pindex_t start, vm_size_t size)
692 {
693
694 VM_OBJECT_LOCK_ASSERT(object, MA_OWNED);
695 swp_pager_meta_free(object, start, size);
696 }
697
698 /*
699 * SWAP_PAGER_RESERVE() - reserve swap blocks in object
700 *
701 * Assigns swap blocks to the specified range within the object. The
702 * swap blocks are not zerod. Any previous swap assignment is destroyed.
703 *
704 * Returns 0 on success, -1 on failure.
705 */
706 int
707 swap_pager_reserve(vm_object_t object, vm_pindex_t start, vm_size_t size)
708 {
709 int n = 0;
710 daddr_t blk = SWAPBLK_NONE;
711 vm_pindex_t beg = start; /* save start index */
712
713 VM_OBJECT_LOCK(object);
714 while (size) {
715 if (n == 0) {
716 n = BLIST_MAX_ALLOC;
717 while ((blk = swp_pager_getswapspace(n)) == SWAPBLK_NONE) {
718 n >>= 1;
719 if (n == 0) {
720 swp_pager_meta_free(object, beg, start - beg);
721 VM_OBJECT_UNLOCK(object);
722 return (-1);
723 }
724 }
725 }
726 swp_pager_meta_build(object, start, blk);
727 --size;
728 ++start;
729 ++blk;
730 --n;
731 }
732 swp_pager_meta_free(object, start, n);
733 VM_OBJECT_UNLOCK(object);
734 return (0);
735 }
736
737 /*
738 * SWAP_PAGER_COPY() - copy blocks from source pager to destination pager
739 * and destroy the source.
740 *
741 * Copy any valid swapblks from the source to the destination. In
742 * cases where both the source and destination have a valid swapblk,
743 * we keep the destination's.
744 *
745 * This routine is allowed to block. It may block allocating metadata
746 * indirectly through swp_pager_meta_build() or if paging is still in
747 * progress on the source.
748 *
749 * This routine can be called at any spl
750 *
751 * XXX vm_page_collapse() kinda expects us not to block because we
752 * supposedly do not need to allocate memory, but for the moment we
753 * *may* have to get a little memory from the zone allocator, but
754 * it is taken from the interrupt memory. We should be ok.
755 *
756 * The source object contains no vm_page_t's (which is just as well)
757 *
758 * The source object is of type OBJT_SWAP.
759 *
760 * The source and destination objects must be locked or
761 * inaccessible (XXX are they ?)
762 */
763 void
764 swap_pager_copy(vm_object_t srcobject, vm_object_t dstobject,
765 vm_pindex_t offset, int destroysource)
766 {
767 vm_pindex_t i;
768
769 VM_OBJECT_LOCK_ASSERT(srcobject, MA_OWNED);
770 VM_OBJECT_LOCK_ASSERT(dstobject, MA_OWNED);
771
772 /*
773 * If destroysource is set, we remove the source object from the
774 * swap_pager internal queue now.
775 */
776 if (destroysource) {
777 if (srcobject->handle != NULL) {
778 mtx_lock(&sw_alloc_mtx);
779 TAILQ_REMOVE(
780 NOBJLIST(srcobject->handle),
781 srcobject,
782 pager_object_list
783 );
784 mtx_unlock(&sw_alloc_mtx);
785 }
786 }
787
788 /*
789 * transfer source to destination.
790 */
791 for (i = 0; i < dstobject->size; ++i) {
792 daddr_t dstaddr;
793
794 /*
795 * Locate (without changing) the swapblk on the destination,
796 * unless it is invalid in which case free it silently, or
797 * if the destination is a resident page, in which case the
798 * source is thrown away.
799 */
800 dstaddr = swp_pager_meta_ctl(dstobject, i, 0);
801
802 if (dstaddr == SWAPBLK_NONE) {
803 /*
804 * Destination has no swapblk and is not resident,
805 * copy source.
806 */
807 daddr_t srcaddr;
808
809 srcaddr = swp_pager_meta_ctl(
810 srcobject,
811 i + offset,
812 SWM_POP
813 );
814
815 if (srcaddr != SWAPBLK_NONE) {
816 /*
817 * swp_pager_meta_build() can sleep.
818 */
819 vm_object_pip_add(srcobject, 1);
820 VM_OBJECT_UNLOCK(srcobject);
821 vm_object_pip_add(dstobject, 1);
822 swp_pager_meta_build(dstobject, i, srcaddr);
823 vm_object_pip_wakeup(dstobject);
824 VM_OBJECT_LOCK(srcobject);
825 vm_object_pip_wakeup(srcobject);
826 }
827 } else {
828 /*
829 * Destination has valid swapblk or it is represented
830 * by a resident page. We destroy the sourceblock.
831 */
832
833 swp_pager_meta_ctl(srcobject, i + offset, SWM_FREE);
834 }
835 }
836
837 /*
838 * Free left over swap blocks in source.
839 *
840 * We have to revert the type to OBJT_DEFAULT so we do not accidently
841 * double-remove the object from the swap queues.
842 */
843 if (destroysource) {
844 swp_pager_meta_free_all(srcobject);
845 /*
846 * Reverting the type is not necessary, the caller is going
847 * to destroy srcobject directly, but I'm doing it here
848 * for consistency since we've removed the object from its
849 * queues.
850 */
851 srcobject->type = OBJT_DEFAULT;
852 }
853 }
854
855 /*
856 * SWAP_PAGER_HASPAGE() - determine if we have good backing store for
857 * the requested page.
858 *
859 * We determine whether good backing store exists for the requested
860 * page and return TRUE if it does, FALSE if it doesn't.
861 *
862 * If TRUE, we also try to determine how much valid, contiguous backing
863 * store exists before and after the requested page within a reasonable
864 * distance. We do not try to restrict it to the swap device stripe
865 * (that is handled in getpages/putpages). It probably isn't worth
866 * doing here.
867 */
868 static boolean_t
869 swap_pager_haspage(vm_object_t object, vm_pindex_t pindex, int *before, int *after)
870 {
871 daddr_t blk0;
872
873 VM_OBJECT_LOCK_ASSERT(object, MA_OWNED);
874 /*
875 * do we have good backing store at the requested index ?
876 */
877 blk0 = swp_pager_meta_ctl(object, pindex, 0);
878
879 if (blk0 == SWAPBLK_NONE) {
880 if (before)
881 *before = 0;
882 if (after)
883 *after = 0;
884 return (FALSE);
885 }
886
887 /*
888 * find backwards-looking contiguous good backing store
889 */
890 if (before != NULL) {
891 int i;
892
893 for (i = 1; i < (SWB_NPAGES/2); ++i) {
894 daddr_t blk;
895
896 if (i > pindex)
897 break;
898 blk = swp_pager_meta_ctl(object, pindex - i, 0);
899 if (blk != blk0 - i)
900 break;
901 }
902 *before = (i - 1);
903 }
904
905 /*
906 * find forward-looking contiguous good backing store
907 */
908 if (after != NULL) {
909 int i;
910
911 for (i = 1; i < (SWB_NPAGES/2); ++i) {
912 daddr_t blk;
913
914 blk = swp_pager_meta_ctl(object, pindex + i, 0);
915 if (blk != blk0 + i)
916 break;
917 }
918 *after = (i - 1);
919 }
920 return (TRUE);
921 }
922
923 /*
924 * SWAP_PAGER_PAGE_UNSWAPPED() - remove swap backing store related to page
925 *
926 * This removes any associated swap backing store, whether valid or
927 * not, from the page.
928 *
929 * This routine is typically called when a page is made dirty, at
930 * which point any associated swap can be freed. MADV_FREE also
931 * calls us in a special-case situation
932 *
933 * NOTE!!! If the page is clean and the swap was valid, the caller
934 * should make the page dirty before calling this routine. This routine
935 * does NOT change the m->dirty status of the page. Also: MADV_FREE
936 * depends on it.
937 *
938 * This routine may not block
939 * This routine must be called at splvm()
940 */
941 static void
942 swap_pager_unswapped(vm_page_t m)
943 {
944
945 VM_OBJECT_LOCK_ASSERT(m->object, MA_OWNED);
946 swp_pager_meta_ctl(m->object, m->pindex, SWM_FREE);
947 }
948
949 /*
950 * SWAP_PAGER_GETPAGES() - bring pages in from swap
951 *
952 * Attempt to retrieve (m, count) pages from backing store, but make
953 * sure we retrieve at least m[reqpage]. We try to load in as large
954 * a chunk surrounding m[reqpage] as is contiguous in swap and which
955 * belongs to the same object.
956 *
957 * The code is designed for asynchronous operation and
958 * immediate-notification of 'reqpage' but tends not to be
959 * used that way. Please do not optimize-out this algorithmic
960 * feature, I intend to improve on it in the future.
961 *
962 * The parent has a single vm_object_pip_add() reference prior to
963 * calling us and we should return with the same.
964 *
965 * The parent has BUSY'd the pages. We should return with 'm'
966 * left busy, but the others adjusted.
967 */
968 static int
969 swap_pager_getpages(vm_object_t object, vm_page_t *m, int count, int reqpage)
970 {
971 struct buf *bp;
972 vm_page_t mreq;
973 int i;
974 int j;
975 daddr_t blk;
976
977 mreq = m[reqpage];
978
979 KASSERT(mreq->object == object,
980 ("swap_pager_getpages: object mismatch %p/%p",
981 object, mreq->object));
982
983 /*
984 * Calculate range to retrieve. The pages have already been assigned
985 * their swapblks. We require a *contiguous* range but we know it to
986 * not span devices. If we do not supply it, bad things
987 * happen. Note that blk, iblk & jblk can be SWAPBLK_NONE, but the
988 * loops are set up such that the case(s) are handled implicitly.
989 *
990 * The swp_*() calls must be made at splvm(). vm_page_free() does
991 * not need to be, but it will go a little faster if it is.
992 */
993 blk = swp_pager_meta_ctl(mreq->object, mreq->pindex, 0);
994
995 for (i = reqpage - 1; i >= 0; --i) {
996 daddr_t iblk;
997
998 iblk = swp_pager_meta_ctl(m[i]->object, m[i]->pindex, 0);
999 if (blk != iblk + (reqpage - i))
1000 break;
1001 }
1002 ++i;
1003
1004 for (j = reqpage + 1; j < count; ++j) {
1005 daddr_t jblk;
1006
1007 jblk = swp_pager_meta_ctl(m[j]->object, m[j]->pindex, 0);
1008 if (blk != jblk - (j - reqpage))
1009 break;
1010 }
1011
1012 /*
1013 * free pages outside our collection range. Note: we never free
1014 * mreq, it must remain busy throughout.
1015 */
1016 if (0 < i || j < count) {
1017 int k;
1018
1019 vm_page_lock_queues();
1020 for (k = 0; k < i; ++k)
1021 vm_page_free(m[k]);
1022 for (k = j; k < count; ++k)
1023 vm_page_free(m[k]);
1024 vm_page_unlock_queues();
1025 }
1026
1027 /*
1028 * Return VM_PAGER_FAIL if we have nothing to do. Return mreq
1029 * still busy, but the others unbusied.
1030 */
1031 if (blk == SWAPBLK_NONE)
1032 return (VM_PAGER_FAIL);
1033
1034 /*
1035 * Getpbuf() can sleep.
1036 */
1037 VM_OBJECT_UNLOCK(object);
1038 /*
1039 * Get a swap buffer header to perform the IO
1040 */
1041 bp = getpbuf(&nsw_rcount);
1042 bp->b_flags |= B_PAGING;
1043
1044 /*
1045 * map our page(s) into kva for input
1046 */
1047 pmap_qenter((vm_offset_t)bp->b_data, m + i, j - i);
1048
1049 bp->b_iocmd = BIO_READ;
1050 bp->b_iodone = swp_pager_async_iodone;
1051 bp->b_rcred = crhold(thread0.td_ucred);
1052 bp->b_wcred = crhold(thread0.td_ucred);
1053 bp->b_blkno = blk - (reqpage - i);
1054 bp->b_bcount = PAGE_SIZE * (j - i);
1055 bp->b_bufsize = PAGE_SIZE * (j - i);
1056 bp->b_pager.pg_reqpage = reqpage - i;
1057
1058 VM_OBJECT_LOCK(object);
1059 vm_page_lock_queues();
1060 {
1061 int k;
1062
1063 for (k = i; k < j; ++k) {
1064 bp->b_pages[k - i] = m[k];
1065 vm_page_flag_set(m[k], PG_SWAPINPROG);
1066 }
1067 }
1068 vm_page_unlock_queues();
1069 bp->b_npages = j - i;
1070
1071 cnt.v_swapin++;
1072 cnt.v_swappgsin += bp->b_npages;
1073
1074 /*
1075 * We still hold the lock on mreq, and our automatic completion routine
1076 * does not remove it.
1077 */
1078 vm_object_pip_add(object, bp->b_npages);
1079 VM_OBJECT_UNLOCK(object);
1080
1081 /*
1082 * perform the I/O. NOTE!!! bp cannot be considered valid after
1083 * this point because we automatically release it on completion.
1084 * Instead, we look at the one page we are interested in which we
1085 * still hold a lock on even through the I/O completion.
1086 *
1087 * The other pages in our m[] array are also released on completion,
1088 * so we cannot assume they are valid anymore either.
1089 *
1090 * NOTE: b_blkno is destroyed by the call to swapdev_strategy
1091 */
1092 BUF_KERNPROC(bp);
1093 swp_pager_strategy(bp);
1094
1095 /*
1096 * wait for the page we want to complete. PG_SWAPINPROG is always
1097 * cleared on completion. If an I/O error occurs, SWAPBLK_NONE
1098 * is set in the meta-data.
1099 */
1100 vm_page_lock_queues();
1101 while ((mreq->flags & PG_SWAPINPROG) != 0) {
1102 vm_page_flag_set(mreq, PG_WANTED | PG_REFERENCED);
1103 cnt.v_intrans++;
1104 if (msleep(mreq, &vm_page_queue_mtx, PSWP, "swread", hz*20)) {
1105 printf(
1106 "swap_pager: indefinite wait buffer: bufobj: %p, blkno: %jd, size: %ld\n",
1107 bp->b_bufobj, (intmax_t)bp->b_blkno, bp->b_bcount);
1108 }
1109 }
1110 vm_page_unlock_queues();
1111
1112 VM_OBJECT_LOCK(object);
1113 /*
1114 * mreq is left busied after completion, but all the other pages
1115 * are freed. If we had an unrecoverable read error the page will
1116 * not be valid.
1117 */
1118 if (mreq->valid != VM_PAGE_BITS_ALL) {
1119 return (VM_PAGER_ERROR);
1120 } else {
1121 return (VM_PAGER_OK);
1122 }
1123
1124 /*
1125 * A final note: in a low swap situation, we cannot deallocate swap
1126 * and mark a page dirty here because the caller is likely to mark
1127 * the page clean when we return, causing the page to possibly revert
1128 * to all-zero's later.
1129 */
1130 }
1131
1132 /*
1133 * swap_pager_putpages:
1134 *
1135 * Assign swap (if necessary) and initiate I/O on the specified pages.
1136 *
1137 * We support both OBJT_DEFAULT and OBJT_SWAP objects. DEFAULT objects
1138 * are automatically converted to SWAP objects.
1139 *
1140 * In a low memory situation we may block in VOP_STRATEGY(), but the new
1141 * vm_page reservation system coupled with properly written VFS devices
1142 * should ensure that no low-memory deadlock occurs. This is an area
1143 * which needs work.
1144 *
1145 * The parent has N vm_object_pip_add() references prior to
1146 * calling us and will remove references for rtvals[] that are
1147 * not set to VM_PAGER_PEND. We need to remove the rest on I/O
1148 * completion.
1149 *
1150 * The parent has soft-busy'd the pages it passes us and will unbusy
1151 * those whos rtvals[] entry is not set to VM_PAGER_PEND on return.
1152 * We need to unbusy the rest on I/O completion.
1153 */
1154 void
1155 swap_pager_putpages(vm_object_t object, vm_page_t *m, int count,
1156 boolean_t sync, int *rtvals)
1157 {
1158 int i;
1159 int n = 0;
1160
1161 GIANT_REQUIRED;
1162 if (count && m[0]->object != object) {
1163 panic("swap_pager_getpages: object mismatch %p/%p",
1164 object,
1165 m[0]->object
1166 );
1167 }
1168
1169 /*
1170 * Step 1
1171 *
1172 * Turn object into OBJT_SWAP
1173 * check for bogus sysops
1174 * force sync if not pageout process
1175 */
1176 if (object->type != OBJT_SWAP)
1177 swp_pager_meta_build(object, 0, SWAPBLK_NONE);
1178 VM_OBJECT_UNLOCK(object);
1179
1180 if (curproc != pageproc)
1181 sync = TRUE;
1182
1183 /*
1184 * Step 2
1185 *
1186 * Update nsw parameters from swap_async_max sysctl values.
1187 * Do not let the sysop crash the machine with bogus numbers.
1188 */
1189 mtx_lock(&pbuf_mtx);
1190 if (swap_async_max != nsw_wcount_async_max) {
1191 int n;
1192
1193 /*
1194 * limit range
1195 */
1196 if ((n = swap_async_max) > nswbuf / 2)
1197 n = nswbuf / 2;
1198 if (n < 1)
1199 n = 1;
1200 swap_async_max = n;
1201
1202 /*
1203 * Adjust difference ( if possible ). If the current async
1204 * count is too low, we may not be able to make the adjustment
1205 * at this time.
1206 */
1207 n -= nsw_wcount_async_max;
1208 if (nsw_wcount_async + n >= 0) {
1209 nsw_wcount_async += n;
1210 nsw_wcount_async_max += n;
1211 wakeup(&nsw_wcount_async);
1212 }
1213 }
1214 mtx_unlock(&pbuf_mtx);
1215
1216 /*
1217 * Step 3
1218 *
1219 * Assign swap blocks and issue I/O. We reallocate swap on the fly.
1220 * The page is left dirty until the pageout operation completes
1221 * successfully.
1222 */
1223 for (i = 0; i < count; i += n) {
1224 int j;
1225 struct buf *bp;
1226 daddr_t blk;
1227
1228 /*
1229 * Maximum I/O size is limited by a number of factors.
1230 */
1231 n = min(BLIST_MAX_ALLOC, count - i);
1232 n = min(n, nsw_cluster_max);
1233
1234 /*
1235 * Get biggest block of swap we can. If we fail, fall
1236 * back and try to allocate a smaller block. Don't go
1237 * overboard trying to allocate space if it would overly
1238 * fragment swap.
1239 */
1240 while (
1241 (blk = swp_pager_getswapspace(n)) == SWAPBLK_NONE &&
1242 n > 4
1243 ) {
1244 n >>= 1;
1245 }
1246 if (blk == SWAPBLK_NONE) {
1247 for (j = 0; j < n; ++j)
1248 rtvals[i+j] = VM_PAGER_FAIL;
1249 continue;
1250 }
1251
1252 /*
1253 * All I/O parameters have been satisfied, build the I/O
1254 * request and assign the swap space.
1255 */
1256 if (sync == TRUE) {
1257 bp = getpbuf(&nsw_wcount_sync);
1258 } else {
1259 bp = getpbuf(&nsw_wcount_async);
1260 bp->b_flags = B_ASYNC;
1261 }
1262 bp->b_flags |= B_PAGING;
1263 bp->b_iocmd = BIO_WRITE;
1264
1265 pmap_qenter((vm_offset_t)bp->b_data, &m[i], n);
1266
1267 bp->b_rcred = crhold(thread0.td_ucred);
1268 bp->b_wcred = crhold(thread0.td_ucred);
1269 bp->b_bcount = PAGE_SIZE * n;
1270 bp->b_bufsize = PAGE_SIZE * n;
1271 bp->b_blkno = blk;
1272
1273 VM_OBJECT_LOCK(object);
1274 for (j = 0; j < n; ++j) {
1275 vm_page_t mreq = m[i+j];
1276
1277 swp_pager_meta_build(
1278 mreq->object,
1279 mreq->pindex,
1280 blk + j
1281 );
1282 vm_page_dirty(mreq);
1283 rtvals[i+j] = VM_PAGER_OK;
1284
1285 vm_page_lock_queues();
1286 vm_page_flag_set(mreq, PG_SWAPINPROG);
1287 vm_page_unlock_queues();
1288 bp->b_pages[j] = mreq;
1289 }
1290 VM_OBJECT_UNLOCK(object);
1291 bp->b_npages = n;
1292 /*
1293 * Must set dirty range for NFS to work.
1294 */
1295 bp->b_dirtyoff = 0;
1296 bp->b_dirtyend = bp->b_bcount;
1297
1298 cnt.v_swapout++;
1299 cnt.v_swappgsout += bp->b_npages;
1300
1301 /*
1302 * asynchronous
1303 *
1304 * NOTE: b_blkno is destroyed by the call to swapdev_strategy
1305 */
1306 if (sync == FALSE) {
1307 bp->b_iodone = swp_pager_async_iodone;
1308 BUF_KERNPROC(bp);
1309 swp_pager_strategy(bp);
1310
1311 for (j = 0; j < n; ++j)
1312 rtvals[i+j] = VM_PAGER_PEND;
1313 /* restart outter loop */
1314 continue;
1315 }
1316
1317 /*
1318 * synchronous
1319 *
1320 * NOTE: b_blkno is destroyed by the call to swapdev_strategy
1321 */
1322 bp->b_iodone = bdone;
1323 swp_pager_strategy(bp);
1324
1325 /*
1326 * Wait for the sync I/O to complete, then update rtvals.
1327 * We just set the rtvals[] to VM_PAGER_PEND so we can call
1328 * our async completion routine at the end, thus avoiding a
1329 * double-free.
1330 */
1331 bwait(bp, PVM, "swwrt");
1332 for (j = 0; j < n; ++j)
1333 rtvals[i+j] = VM_PAGER_PEND;
1334 /*
1335 * Now that we are through with the bp, we can call the
1336 * normal async completion, which frees everything up.
1337 */
1338 swp_pager_async_iodone(bp);
1339 }
1340 VM_OBJECT_LOCK(object);
1341 }
1342
1343 /*
1344 * swp_pager_async_iodone:
1345 *
1346 * Completion routine for asynchronous reads and writes from/to swap.
1347 * Also called manually by synchronous code to finish up a bp.
1348 *
1349 * For READ operations, the pages are PG_BUSY'd. For WRITE operations,
1350 * the pages are vm_page_t->busy'd. For READ operations, we PG_BUSY
1351 * unbusy all pages except the 'main' request page. For WRITE
1352 * operations, we vm_page_t->busy'd unbusy all pages ( we can do this
1353 * because we marked them all VM_PAGER_PEND on return from putpages ).
1354 *
1355 * This routine may not block.
1356 * This routine is called at splbio() or better
1357 *
1358 * We up ourselves to splvm() as required for various vm_page related
1359 * calls.
1360 */
1361 static void
1362 swp_pager_async_iodone(struct buf *bp)
1363 {
1364 int i;
1365 vm_object_t object = NULL;
1366
1367 /*
1368 * report error
1369 */
1370 if (bp->b_ioflags & BIO_ERROR) {
1371 printf(
1372 "swap_pager: I/O error - %s failed; blkno %ld,"
1373 "size %ld, error %d\n",
1374 ((bp->b_iocmd == BIO_READ) ? "pagein" : "pageout"),
1375 (long)bp->b_blkno,
1376 (long)bp->b_bcount,
1377 bp->b_error
1378 );
1379 }
1380
1381 /*
1382 * remove the mapping for kernel virtual
1383 */
1384 pmap_qremove((vm_offset_t)bp->b_data, bp->b_npages);
1385
1386 if (bp->b_npages) {
1387 object = bp->b_pages[0]->object;
1388 VM_OBJECT_LOCK(object);
1389 }
1390 vm_page_lock_queues();
1391 /*
1392 * cleanup pages. If an error occurs writing to swap, we are in
1393 * very serious trouble. If it happens to be a disk error, though,
1394 * we may be able to recover by reassigning the swap later on. So
1395 * in this case we remove the m->swapblk assignment for the page
1396 * but do not free it in the rlist. The errornous block(s) are thus
1397 * never reallocated as swap. Redirty the page and continue.
1398 */
1399 for (i = 0; i < bp->b_npages; ++i) {
1400 vm_page_t m = bp->b_pages[i];
1401
1402 vm_page_flag_clear(m, PG_SWAPINPROG);
1403
1404 if (bp->b_ioflags & BIO_ERROR) {
1405 /*
1406 * If an error occurs I'd love to throw the swapblk
1407 * away without freeing it back to swapspace, so it
1408 * can never be used again. But I can't from an
1409 * interrupt.
1410 */
1411 if (bp->b_iocmd == BIO_READ) {
1412 /*
1413 * When reading, reqpage needs to stay
1414 * locked for the parent, but all other
1415 * pages can be freed. We still want to
1416 * wakeup the parent waiting on the page,
1417 * though. ( also: pg_reqpage can be -1 and
1418 * not match anything ).
1419 *
1420 * We have to wake specifically requested pages
1421 * up too because we cleared PG_SWAPINPROG and
1422 * someone may be waiting for that.
1423 *
1424 * NOTE: for reads, m->dirty will probably
1425 * be overridden by the original caller of
1426 * getpages so don't play cute tricks here.
1427 *
1428 * XXX IT IS NOT LEGAL TO FREE THE PAGE HERE
1429 * AS THIS MESSES WITH object->memq, and it is
1430 * not legal to mess with object->memq from an
1431 * interrupt.
1432 */
1433 m->valid = 0;
1434 if (i != bp->b_pager.pg_reqpage)
1435 vm_page_free(m);
1436 else
1437 vm_page_flash(m);
1438 /*
1439 * If i == bp->b_pager.pg_reqpage, do not wake
1440 * the page up. The caller needs to.
1441 */
1442 } else {
1443 /*
1444 * If a write error occurs, reactivate page
1445 * so it doesn't clog the inactive list,
1446 * then finish the I/O.
1447 */
1448 vm_page_dirty(m);
1449 vm_page_activate(m);
1450 vm_page_io_finish(m);
1451 }
1452 } else if (bp->b_iocmd == BIO_READ) {
1453 /*
1454 * For read success, clear dirty bits. Nobody should
1455 * have this page mapped but don't take any chances,
1456 * make sure the pmap modify bits are also cleared.
1457 *
1458 * NOTE: for reads, m->dirty will probably be
1459 * overridden by the original caller of getpages so
1460 * we cannot set them in order to free the underlying
1461 * swap in a low-swap situation. I don't think we'd
1462 * want to do that anyway, but it was an optimization
1463 * that existed in the old swapper for a time before
1464 * it got ripped out due to precisely this problem.
1465 *
1466 * If not the requested page then deactivate it.
1467 *
1468 * Note that the requested page, reqpage, is left
1469 * busied, but we still have to wake it up. The
1470 * other pages are released (unbusied) by
1471 * vm_page_wakeup(). We do not set reqpage's
1472 * valid bits here, it is up to the caller.
1473 */
1474 pmap_clear_modify(m);
1475 m->valid = VM_PAGE_BITS_ALL;
1476 vm_page_undirty(m);
1477
1478 /*
1479 * We have to wake specifically requested pages
1480 * up too because we cleared PG_SWAPINPROG and
1481 * could be waiting for it in getpages. However,
1482 * be sure to not unbusy getpages specifically
1483 * requested page - getpages expects it to be
1484 * left busy.
1485 */
1486 if (i != bp->b_pager.pg_reqpage) {
1487 vm_page_deactivate(m);
1488 vm_page_wakeup(m);
1489 } else {
1490 vm_page_flash(m);
1491 }
1492 } else {
1493 /*
1494 * For write success, clear the modify and dirty
1495 * status, then finish the I/O ( which decrements the
1496 * busy count and possibly wakes waiter's up ).
1497 */
1498 pmap_clear_modify(m);
1499 vm_page_undirty(m);
1500 vm_page_io_finish(m);
1501 if (vm_page_count_severe())
1502 vm_page_try_to_cache(m);
1503 }
1504 }
1505 vm_page_unlock_queues();
1506
1507 /*
1508 * adjust pip. NOTE: the original parent may still have its own
1509 * pip refs on the object.
1510 */
1511 if (object != NULL) {
1512 vm_object_pip_wakeupn(object, bp->b_npages);
1513 VM_OBJECT_UNLOCK(object);
1514 }
1515 /*
1516 * swapdev_strategy() manually sets b_vp and b_bufobj before calling
1517 * bstrategy(). Set them back to NULL now we're done with it, or we'll
1518 * trigger a KASSERT in relpbuf().
1519 */
1520 if (bp->b_vp) {
1521 bp->b_vp = NULL;
1522 bp->b_bufobj = NULL;
1523 }
1524
1525 /*
1526 * release the physical I/O buffer
1527 */
1528 relpbuf(
1529 bp,
1530 ((bp->b_iocmd == BIO_READ) ? &nsw_rcount :
1531 ((bp->b_flags & B_ASYNC) ?
1532 &nsw_wcount_async :
1533 &nsw_wcount_sync
1534 )
1535 )
1536 );
1537 }
1538
1539 /*
1540 * swap_pager_isswapped:
1541 *
1542 * Return 1 if at least one page in the given object is paged
1543 * out to the given swap device.
1544 *
1545 * This routine may not block.
1546 */
1547 int
1548 swap_pager_isswapped(vm_object_t object, struct swdevt *sp)
1549 {
1550 daddr_t index = 0;
1551 int bcount;
1552 int i;
1553
1554 VM_OBJECT_LOCK_ASSERT(object, MA_OWNED);
1555 if (object->type != OBJT_SWAP)
1556 return (0);
1557
1558 mtx_lock(&swhash_mtx);
1559 for (bcount = 0; bcount < object->un_pager.swp.swp_bcount; bcount++) {
1560 struct swblock *swap;
1561
1562 if ((swap = *swp_pager_hash(object, index)) != NULL) {
1563 for (i = 0; i < SWAP_META_PAGES; ++i) {
1564 if (swp_pager_isondev(swap->swb_pages[i], sp)) {
1565 mtx_unlock(&swhash_mtx);
1566 return (1);
1567 }
1568 }
1569 }
1570 index += SWAP_META_PAGES;
1571 if (index > 0x20000000)
1572 panic("swap_pager_isswapped: failed to locate all swap meta blocks");
1573 }
1574 mtx_unlock(&swhash_mtx);
1575 return (0);
1576 }
1577
1578 /*
1579 * SWP_PAGER_FORCE_PAGEIN() - force a swap block to be paged in
1580 *
1581 * This routine dissociates the page at the given index within a
1582 * swap block from its backing store, paging it in if necessary.
1583 * If the page is paged in, it is placed in the inactive queue,
1584 * since it had its backing store ripped out from under it.
1585 * We also attempt to swap in all other pages in the swap block,
1586 * we only guarantee that the one at the specified index is
1587 * paged in.
1588 *
1589 * XXX - The code to page the whole block in doesn't work, so we
1590 * revert to the one-by-one behavior for now. Sigh.
1591 */
1592 static __inline void
1593 swp_pager_force_pagein(vm_object_t object, vm_pindex_t pindex)
1594 {
1595 vm_page_t m;
1596
1597 vm_object_pip_add(object, 1);
1598 m = vm_page_grab(object, pindex, VM_ALLOC_NORMAL|VM_ALLOC_RETRY);
1599 if (m->valid == VM_PAGE_BITS_ALL) {
1600 vm_object_pip_subtract(object, 1);
1601 vm_page_lock_queues();
1602 vm_page_activate(m);
1603 vm_page_dirty(m);
1604 vm_page_wakeup(m);
1605 vm_page_unlock_queues();
1606 vm_pager_page_unswapped(m);
1607 return;
1608 }
1609
1610 if (swap_pager_getpages(object, &m, 1, 0) != VM_PAGER_OK)
1611 panic("swap_pager_force_pagein: read from swap failed");/*XXX*/
1612 vm_object_pip_subtract(object, 1);
1613 vm_page_lock_queues();
1614 vm_page_dirty(m);
1615 vm_page_dontneed(m);
1616 vm_page_wakeup(m);
1617 vm_page_unlock_queues();
1618 vm_pager_page_unswapped(m);
1619 }
1620
1621 /*
1622 * swap_pager_swapoff:
1623 *
1624 * Page in all of the pages that have been paged out to the
1625 * given device. The corresponding blocks in the bitmap must be
1626 * marked as allocated and the device must be flagged SW_CLOSING.
1627 * There may be no processes swapped out to the device.
1628 *
1629 * This routine may block.
1630 */
1631 static void
1632 swap_pager_swapoff(struct swdevt *sp)
1633 {
1634 struct swblock *swap;
1635 int i, j, retries;
1636
1637 GIANT_REQUIRED;
1638
1639 retries = 0;
1640 full_rescan:
1641 mtx_lock(&swhash_mtx);
1642 for (i = 0; i <= swhash_mask; i++) { /* '<=' is correct here */
1643 restart:
1644 for (swap = swhash[i]; swap != NULL; swap = swap->swb_hnext) {
1645 vm_object_t object = swap->swb_object;
1646 vm_pindex_t pindex = swap->swb_index;
1647 for (j = 0; j < SWAP_META_PAGES; ++j) {
1648 if (swp_pager_isondev(swap->swb_pages[j], sp)) {
1649 /* avoid deadlock */
1650 if (!VM_OBJECT_TRYLOCK(object)) {
1651 break;
1652 } else {
1653 mtx_unlock(&swhash_mtx);
1654 swp_pager_force_pagein(object,
1655 pindex + j);
1656 VM_OBJECT_UNLOCK(object);
1657 mtx_lock(&swhash_mtx);
1658 goto restart;
1659 }
1660 }
1661 }
1662 }
1663 }
1664 mtx_unlock(&swhash_mtx);
1665 if (sp->sw_used) {
1666 int dummy;
1667 /*
1668 * Objects may be locked or paging to the device being
1669 * removed, so we will miss their pages and need to
1670 * make another pass. We have marked this device as
1671 * SW_CLOSING, so the activity should finish soon.
1672 */
1673 retries++;
1674 if (retries > 100) {
1675 panic("swapoff: failed to locate %d swap blocks",
1676 sp->sw_used);
1677 }
1678 tsleep(&dummy, PVM, "swpoff", hz / 20);
1679 goto full_rescan;
1680 }
1681 }
1682
1683 /************************************************************************
1684 * SWAP META DATA *
1685 ************************************************************************
1686 *
1687 * These routines manipulate the swap metadata stored in the
1688 * OBJT_SWAP object. All swp_*() routines must be called at
1689 * splvm() because swap can be freed up by the low level vm_page
1690 * code which might be called from interrupts beyond what splbio() covers.
1691 *
1692 * Swap metadata is implemented with a global hash and not directly
1693 * linked into the object. Instead the object simply contains
1694 * appropriate tracking counters.
1695 */
1696
1697 /*
1698 * SWP_PAGER_META_BUILD() - add swap block to swap meta data for object
1699 *
1700 * We first convert the object to a swap object if it is a default
1701 * object.
1702 *
1703 * The specified swapblk is added to the object's swap metadata. If
1704 * the swapblk is not valid, it is freed instead. Any previously
1705 * assigned swapblk is freed.
1706 *
1707 * This routine must be called at splvm(), except when used to convert
1708 * an OBJT_DEFAULT object into an OBJT_SWAP object.
1709 */
1710 static void
1711 swp_pager_meta_build(vm_object_t object, vm_pindex_t pindex, daddr_t swapblk)
1712 {
1713 struct swblock *swap;
1714 struct swblock **pswap;
1715 int idx;
1716
1717 VM_OBJECT_LOCK_ASSERT(object, MA_OWNED);
1718 /*
1719 * Convert default object to swap object if necessary
1720 */
1721 if (object->type != OBJT_SWAP) {
1722 object->type = OBJT_SWAP;
1723 object->un_pager.swp.swp_bcount = 0;
1724
1725 if (object->handle != NULL) {
1726 mtx_lock(&sw_alloc_mtx);
1727 TAILQ_INSERT_TAIL(
1728 NOBJLIST(object->handle),
1729 object,
1730 pager_object_list
1731 );
1732 mtx_unlock(&sw_alloc_mtx);
1733 }
1734 }
1735
1736 /*
1737 * Locate hash entry. If not found create, but if we aren't adding
1738 * anything just return. If we run out of space in the map we wait
1739 * and, since the hash table may have changed, retry.
1740 */
1741 retry:
1742 mtx_lock(&swhash_mtx);
1743 pswap = swp_pager_hash(object, pindex);
1744
1745 if ((swap = *pswap) == NULL) {
1746 int i;
1747
1748 if (swapblk == SWAPBLK_NONE)
1749 goto done;
1750
1751 swap = *pswap = uma_zalloc(swap_zone, M_NOWAIT);
1752 if (swap == NULL) {
1753 mtx_unlock(&swhash_mtx);
1754 VM_OBJECT_UNLOCK(object);
1755 VM_WAIT;
1756 VM_OBJECT_LOCK(object);
1757 goto retry;
1758 }
1759
1760 swap->swb_hnext = NULL;
1761 swap->swb_object = object;
1762 swap->swb_index = pindex & ~(vm_pindex_t)SWAP_META_MASK;
1763 swap->swb_count = 0;
1764
1765 ++object->un_pager.swp.swp_bcount;
1766
1767 for (i = 0; i < SWAP_META_PAGES; ++i)
1768 swap->swb_pages[i] = SWAPBLK_NONE;
1769 }
1770
1771 /*
1772 * Delete prior contents of metadata
1773 */
1774 idx = pindex & SWAP_META_MASK;
1775
1776 if (swap->swb_pages[idx] != SWAPBLK_NONE) {
1777 swp_pager_freeswapspace(swap->swb_pages[idx], 1);
1778 --swap->swb_count;
1779 }
1780
1781 /*
1782 * Enter block into metadata
1783 */
1784 swap->swb_pages[idx] = swapblk;
1785 if (swapblk != SWAPBLK_NONE)
1786 ++swap->swb_count;
1787 done:
1788 mtx_unlock(&swhash_mtx);
1789 }
1790
1791 /*
1792 * SWP_PAGER_META_FREE() - free a range of blocks in the object's swap metadata
1793 *
1794 * The requested range of blocks is freed, with any associated swap
1795 * returned to the swap bitmap.
1796 *
1797 * This routine will free swap metadata structures as they are cleaned
1798 * out. This routine does *NOT* operate on swap metadata associated
1799 * with resident pages.
1800 *
1801 * This routine must be called at splvm()
1802 */
1803 static void
1804 swp_pager_meta_free(vm_object_t object, vm_pindex_t index, daddr_t count)
1805 {
1806
1807 VM_OBJECT_LOCK_ASSERT(object, MA_OWNED);
1808 if (object->type != OBJT_SWAP)
1809 return;
1810
1811 while (count > 0) {
1812 struct swblock **pswap;
1813 struct swblock *swap;
1814
1815 mtx_lock(&swhash_mtx);
1816 pswap = swp_pager_hash(object, index);
1817
1818 if ((swap = *pswap) != NULL) {
1819 daddr_t v = swap->swb_pages[index & SWAP_META_MASK];
1820
1821 if (v != SWAPBLK_NONE) {
1822 swp_pager_freeswapspace(v, 1);
1823 swap->swb_pages[index & SWAP_META_MASK] =
1824 SWAPBLK_NONE;
1825 if (--swap->swb_count == 0) {
1826 *pswap = swap->swb_hnext;
1827 uma_zfree(swap_zone, swap);
1828 --object->un_pager.swp.swp_bcount;
1829 }
1830 }
1831 --count;
1832 ++index;
1833 } else {
1834 int n = SWAP_META_PAGES - (index & SWAP_META_MASK);
1835 count -= n;
1836 index += n;
1837 }
1838 mtx_unlock(&swhash_mtx);
1839 }
1840 }
1841
1842 /*
1843 * SWP_PAGER_META_FREE_ALL() - destroy all swap metadata associated with object
1844 *
1845 * This routine locates and destroys all swap metadata associated with
1846 * an object.
1847 *
1848 * This routine must be called at splvm()
1849 */
1850 static void
1851 swp_pager_meta_free_all(vm_object_t object)
1852 {
1853 daddr_t index = 0;
1854
1855 VM_OBJECT_LOCK_ASSERT(object, MA_OWNED);
1856 if (object->type != OBJT_SWAP)
1857 return;
1858
1859 while (object->un_pager.swp.swp_bcount) {
1860 struct swblock **pswap;
1861 struct swblock *swap;
1862
1863 mtx_lock(&swhash_mtx);
1864 pswap = swp_pager_hash(object, index);
1865 if ((swap = *pswap) != NULL) {
1866 int i;
1867
1868 for (i = 0; i < SWAP_META_PAGES; ++i) {
1869 daddr_t v = swap->swb_pages[i];
1870 if (v != SWAPBLK_NONE) {
1871 --swap->swb_count;
1872 swp_pager_freeswapspace(v, 1);
1873 }
1874 }
1875 if (swap->swb_count != 0)
1876 panic("swap_pager_meta_free_all: swb_count != 0");
1877 *pswap = swap->swb_hnext;
1878 uma_zfree(swap_zone, swap);
1879 --object->un_pager.swp.swp_bcount;
1880 }
1881 mtx_unlock(&swhash_mtx);
1882 index += SWAP_META_PAGES;
1883 if (index > 0x20000000)
1884 panic("swp_pager_meta_free_all: failed to locate all swap meta blocks");
1885 }
1886 }
1887
1888 /*
1889 * SWP_PAGER_METACTL() - misc control of swap and vm_page_t meta data.
1890 *
1891 * This routine is capable of looking up, popping, or freeing
1892 * swapblk assignments in the swap meta data or in the vm_page_t.
1893 * The routine typically returns the swapblk being looked-up, or popped,
1894 * or SWAPBLK_NONE if the block was freed, or SWAPBLK_NONE if the block
1895 * was invalid. This routine will automatically free any invalid
1896 * meta-data swapblks.
1897 *
1898 * It is not possible to store invalid swapblks in the swap meta data
1899 * (other then a literal 'SWAPBLK_NONE'), so we don't bother checking.
1900 *
1901 * When acting on a busy resident page and paging is in progress, we
1902 * have to wait until paging is complete but otherwise can act on the
1903 * busy page.
1904 *
1905 * This routine must be called at splvm().
1906 *
1907 * SWM_FREE remove and free swap block from metadata
1908 * SWM_POP remove from meta data but do not free.. pop it out
1909 */
1910 static daddr_t
1911 swp_pager_meta_ctl(vm_object_t object, vm_pindex_t pindex, int flags)
1912 {
1913 struct swblock **pswap;
1914 struct swblock *swap;
1915 daddr_t r1;
1916 int idx;
1917
1918 VM_OBJECT_LOCK_ASSERT(object, MA_OWNED);
1919 /*
1920 * The meta data only exists of the object is OBJT_SWAP
1921 * and even then might not be allocated yet.
1922 */
1923 if (object->type != OBJT_SWAP)
1924 return (SWAPBLK_NONE);
1925
1926 r1 = SWAPBLK_NONE;
1927 mtx_lock(&swhash_mtx);
1928 pswap = swp_pager_hash(object, pindex);
1929
1930 if ((swap = *pswap) != NULL) {
1931 idx = pindex & SWAP_META_MASK;
1932 r1 = swap->swb_pages[idx];
1933
1934 if (r1 != SWAPBLK_NONE) {
1935 if (flags & SWM_FREE) {
1936 swp_pager_freeswapspace(r1, 1);
1937 r1 = SWAPBLK_NONE;
1938 }
1939 if (flags & (SWM_FREE|SWM_POP)) {
1940 swap->swb_pages[idx] = SWAPBLK_NONE;
1941 if (--swap->swb_count == 0) {
1942 *pswap = swap->swb_hnext;
1943 uma_zfree(swap_zone, swap);
1944 --object->un_pager.swp.swp_bcount;
1945 }
1946 }
1947 }
1948 }
1949 mtx_unlock(&swhash_mtx);
1950 return (r1);
1951 }
1952
1953 /*
1954 * System call swapon(name) enables swapping on device name,
1955 * which must be in the swdevsw. Return EBUSY
1956 * if already swapping on this device.
1957 */
1958 #ifndef _SYS_SYSPROTO_H_
1959 struct swapon_args {
1960 char *name;
1961 };
1962 #endif
1963
1964 /*
1965 * MPSAFE
1966 */
1967 /* ARGSUSED */
1968 int
1969 swapon(struct thread *td, struct swapon_args *uap)
1970 {
1971 struct vattr attr;
1972 struct vnode *vp;
1973 struct nameidata nd;
1974 int error;
1975
1976 mtx_lock(&Giant);
1977 error = suser(td);
1978 if (error)
1979 goto done2;
1980
1981 while (swdev_syscall_active)
1982 tsleep(&swdev_syscall_active, PUSER - 1, "swpon", 0);
1983 swdev_syscall_active = 1;
1984
1985 /*
1986 * Swap metadata may not fit in the KVM if we have physical
1987 * memory of >1GB.
1988 */
1989 if (swap_zone == NULL) {
1990 error = ENOMEM;
1991 goto done;
1992 }
1993
1994 NDINIT(&nd, LOOKUP, ISOPEN | FOLLOW, UIO_USERSPACE, uap->name, td);
1995 error = namei(&nd);
1996 if (error)
1997 goto done;
1998
1999 NDFREE(&nd, NDF_ONLY_PNBUF);
2000 vp = nd.ni_vp;
2001
2002 if (vn_isdisk(vp, &error)) {
2003 error = swapongeom(td, vp);
2004 } else if (vp->v_type == VREG &&
2005 (vp->v_mount->mnt_vfc->vfc_flags & VFCF_NETWORK) != 0 &&
2006 (error = VOP_GETATTR(vp, &attr, td->td_ucred, td)) == 0) {
2007 /*
2008 * Allow direct swapping to NFS regular files in the same
2009 * way that nfs_mountroot() sets up diskless swapping.
2010 */
2011 error = swaponvp(td, vp, attr.va_size / DEV_BSIZE);
2012 }
2013
2014 if (error)
2015 vrele(vp);
2016 done:
2017 swdev_syscall_active = 0;
2018 wakeup_one(&swdev_syscall_active);
2019 done2:
2020 mtx_unlock(&Giant);
2021 return (error);
2022 }
2023
2024 static void
2025 swaponsomething(struct vnode *vp, void *id, u_long nblks, sw_strategy_t *strategy, sw_close_t *close, dev_t dev)
2026 {
2027 struct swdevt *sp, *tsp;
2028 swblk_t dvbase;
2029 u_long mblocks;
2030
2031 /*
2032 * If we go beyond this, we get overflows in the radix
2033 * tree bitmap code.
2034 */
2035 mblocks = 0x40000000 / BLIST_META_RADIX;
2036 if (nblks > mblocks) {
2037 printf("WARNING: reducing size to maximum of %lu blocks per swap unit\n",
2038 mblocks);
2039 nblks = mblocks;
2040 }
2041 /*
2042 * nblks is in DEV_BSIZE'd chunks, convert to PAGE_SIZE'd chunks.
2043 * First chop nblks off to page-align it, then convert.
2044 *
2045 * sw->sw_nblks is in page-sized chunks now too.
2046 */
2047 nblks &= ~(ctodb(1) - 1);
2048 nblks = dbtoc(nblks);
2049
2050 sp = malloc(sizeof *sp, M_VMPGDATA, M_WAITOK | M_ZERO);
2051 sp->sw_vp = vp;
2052 sp->sw_id = id;
2053 sp->sw_dev = dev;
2054 sp->sw_flags = 0;
2055 sp->sw_nblks = nblks;
2056 sp->sw_used = 0;
2057 sp->sw_strategy = strategy;
2058 sp->sw_close = close;
2059
2060 sp->sw_blist = blist_create(nblks);
2061 /*
2062 * Do not free the first two block in order to avoid overwriting
2063 * any bsd label at the front of the partition
2064 */
2065 blist_free(sp->sw_blist, 2, nblks - 2);
2066
2067 dvbase = 0;
2068 mtx_lock(&sw_dev_mtx);
2069 TAILQ_FOREACH(tsp, &swtailq, sw_list) {
2070 if (tsp->sw_end >= dvbase) {
2071 /*
2072 * We put one uncovered page between the devices
2073 * in order to definitively prevent any cross-device
2074 * I/O requests
2075 */
2076 dvbase = tsp->sw_end + 1;
2077 }
2078 }
2079 sp->sw_first = dvbase;
2080 sp->sw_end = dvbase + nblks;
2081 TAILQ_INSERT_TAIL(&swtailq, sp, sw_list);
2082 nswapdev++;
2083 swap_pager_avail += nblks;
2084 swp_sizecheck();
2085 mtx_unlock(&sw_dev_mtx);
2086 }
2087
2088 /*
2089 * SYSCALL: swapoff(devname)
2090 *
2091 * Disable swapping on the given device.
2092 *
2093 * XXX: Badly designed system call: it should use a device index
2094 * rather than filename as specification. We keep sw_vp around
2095 * only to make this work.
2096 */
2097 #ifndef _SYS_SYSPROTO_H_
2098 struct swapoff_args {
2099 char *name;
2100 };
2101 #endif
2102
2103 /*
2104 * MPSAFE
2105 */
2106 /* ARGSUSED */
2107 int
2108 swapoff(struct thread *td, struct swapoff_args *uap)
2109 {
2110 struct vnode *vp;
2111 struct nameidata nd;
2112 struct swdevt *sp;
2113 int error;
2114
2115 error = suser(td);
2116 if (error)
2117 return (error);
2118
2119 mtx_lock(&Giant);
2120 while (swdev_syscall_active)
2121 tsleep(&swdev_syscall_active, PUSER - 1, "swpoff", 0);
2122 swdev_syscall_active = 1;
2123
2124 NDINIT(&nd, LOOKUP, FOLLOW, UIO_USERSPACE, uap->name, td);
2125 error = namei(&nd);
2126 if (error)
2127 goto done;
2128 NDFREE(&nd, NDF_ONLY_PNBUF);
2129 vp = nd.ni_vp;
2130
2131 mtx_lock(&sw_dev_mtx);
2132 TAILQ_FOREACH(sp, &swtailq, sw_list) {
2133 if (sp->sw_vp == vp)
2134 break;
2135 }
2136 mtx_unlock(&sw_dev_mtx);
2137 if (sp == NULL) {
2138 error = EINVAL;
2139 goto done;
2140 }
2141 error = swapoff_one(sp, td);
2142 done:
2143 swdev_syscall_active = 0;
2144 wakeup_one(&swdev_syscall_active);
2145 mtx_unlock(&Giant);
2146 return (error);
2147 }
2148
2149 static int
2150 swapoff_one(struct swdevt *sp, struct thread *td)
2151 {
2152 u_long nblks, dvbase;
2153 #ifdef MAC
2154 int error;
2155 #endif
2156
2157 mtx_assert(&Giant, MA_OWNED);
2158 #ifdef MAC
2159 (void) vn_lock(sp->sw_vp, LK_EXCLUSIVE | LK_RETRY, td);
2160 error = mac_check_system_swapoff(td->td_ucred, sp->sw_vp);
2161 (void) VOP_UNLOCK(sp->sw_vp, 0, td);
2162 if (error != 0)
2163 return (error);
2164 #endif
2165 nblks = sp->sw_nblks;
2166
2167 /*
2168 * We can turn off this swap device safely only if the
2169 * available virtual memory in the system will fit the amount
2170 * of data we will have to page back in, plus an epsilon so
2171 * the system doesn't become critically low on swap space.
2172 */
2173 if (cnt.v_free_count + cnt.v_cache_count + swap_pager_avail <
2174 nblks + nswap_lowat) {
2175 return (ENOMEM);
2176 }
2177
2178 /*
2179 * Prevent further allocations on this device.
2180 */
2181 mtx_lock(&sw_dev_mtx);
2182 sp->sw_flags |= SW_CLOSING;
2183 for (dvbase = 0; dvbase < sp->sw_end; dvbase += dmmax) {
2184 swap_pager_avail -= blist_fill(sp->sw_blist,
2185 dvbase, dmmax);
2186 }
2187 mtx_unlock(&sw_dev_mtx);
2188
2189 /*
2190 * Page in the contents of the device and close it.
2191 */
2192 swap_pager_swapoff(sp);
2193
2194 sp->sw_close(td, sp);
2195 sp->sw_id = NULL;
2196 mtx_lock(&sw_dev_mtx);
2197 TAILQ_REMOVE(&swtailq, sp, sw_list);
2198 nswapdev--;
2199 if (nswapdev == 0) {
2200 swap_pager_full = 2;
2201 swap_pager_almost_full = 1;
2202 }
2203 if (swdevhd == sp)
2204 swdevhd = NULL;
2205 mtx_unlock(&sw_dev_mtx);
2206 blist_destroy(sp->sw_blist);
2207 free(sp, M_VMPGDATA);
2208 return (0);
2209 }
2210
2211 void
2212 swapoff_all(void)
2213 {
2214 struct swdevt *sp, *spt;
2215 const char *devname;
2216 int error;
2217
2218 mtx_lock(&Giant);
2219 while (swdev_syscall_active)
2220 tsleep(&swdev_syscall_active, PUSER - 1, "swpoff", 0);
2221 swdev_syscall_active = 1;
2222
2223 mtx_lock(&sw_dev_mtx);
2224 TAILQ_FOREACH_SAFE(sp, &swtailq, sw_list, spt) {
2225 mtx_unlock(&sw_dev_mtx);
2226 if (vn_isdisk(sp->sw_vp, NULL))
2227 devname = sp->sw_vp->v_rdev->si_name;
2228 else
2229 devname = "[file]";
2230 error = swapoff_one(sp, &thread0);
2231 if (error != 0) {
2232 printf("Cannot remove swap device %s (error=%d), "
2233 "skipping.\n", devname, error);
2234 } else if (bootverbose) {
2235 printf("Swap device %s removed.\n", devname);
2236 }
2237 mtx_lock(&sw_dev_mtx);
2238 }
2239 mtx_unlock(&sw_dev_mtx);
2240
2241 swdev_syscall_active = 0;
2242 wakeup_one(&swdev_syscall_active);
2243 mtx_unlock(&Giant);
2244 }
2245
2246 void
2247 swap_pager_status(int *total, int *used)
2248 {
2249 struct swdevt *sp;
2250
2251 *total = 0;
2252 *used = 0;
2253 mtx_lock(&sw_dev_mtx);
2254 TAILQ_FOREACH(sp, &swtailq, sw_list) {
2255 *total += sp->sw_nblks;
2256 *used += sp->sw_used;
2257 }
2258 mtx_unlock(&sw_dev_mtx);
2259 }
2260
2261 static int
2262 sysctl_vm_swap_info(SYSCTL_HANDLER_ARGS)
2263 {
2264 int *name = (int *)arg1;
2265 int error, n;
2266 struct xswdev xs;
2267 struct swdevt *sp;
2268
2269 if (arg2 != 1) /* name length */
2270 return (EINVAL);
2271
2272 n = 0;
2273 mtx_lock(&sw_dev_mtx);
2274 TAILQ_FOREACH(sp, &swtailq, sw_list) {
2275 if (n == *name) {
2276 mtx_unlock(&sw_dev_mtx);
2277 xs.xsw_version = XSWDEV_VERSION;
2278 xs.xsw_dev = sp->sw_dev;
2279 xs.xsw_flags = sp->sw_flags;
2280 xs.xsw_nblks = sp->sw_nblks;
2281 xs.xsw_used = sp->sw_used;
2282
2283 error = SYSCTL_OUT(req, &xs, sizeof(xs));
2284 return (error);
2285 }
2286 n++;
2287 }
2288 mtx_unlock(&sw_dev_mtx);
2289 return (ENOENT);
2290 }
2291
2292 SYSCTL_INT(_vm, OID_AUTO, nswapdev, CTLFLAG_RD, &nswapdev, 0,
2293 "Number of swap devices");
2294 SYSCTL_NODE(_vm, OID_AUTO, swap_info, CTLFLAG_RD, sysctl_vm_swap_info,
2295 "Swap statistics by device");
2296
2297 /*
2298 * vmspace_swap_count() - count the approximate swap useage in pages for a
2299 * vmspace.
2300 *
2301 * The map must be locked.
2302 *
2303 * Swap useage is determined by taking the proportional swap used by
2304 * VM objects backing the VM map. To make up for fractional losses,
2305 * if the VM object has any swap use at all the associated map entries
2306 * count for at least 1 swap page.
2307 */
2308 int
2309 vmspace_swap_count(struct vmspace *vmspace)
2310 {
2311 vm_map_t map = &vmspace->vm_map;
2312 vm_map_entry_t cur;
2313 int count = 0;
2314
2315 for (cur = map->header.next; cur != &map->header; cur = cur->next) {
2316 vm_object_t object;
2317
2318 if ((cur->eflags & MAP_ENTRY_IS_SUB_MAP) == 0 &&
2319 (object = cur->object.vm_object) != NULL) {
2320 VM_OBJECT_LOCK(object);
2321 if (object->type == OBJT_SWAP &&
2322 object->un_pager.swp.swp_bcount != 0) {
2323 int n = (cur->end - cur->start) / PAGE_SIZE;
2324
2325 count += object->un_pager.swp.swp_bcount *
2326 SWAP_META_PAGES * n / object->size + 1;
2327 }
2328 VM_OBJECT_UNLOCK(object);
2329 }
2330 }
2331 return (count);
2332 }
2333
2334 /*
2335 * GEOM backend
2336 *
2337 * Swapping onto disk devices.
2338 *
2339 */
2340
2341 static g_orphan_t swapgeom_orphan;
2342
2343 static struct g_class g_swap_class = {
2344 .name = "SWAP",
2345 .version = G_VERSION,
2346 .orphan = swapgeom_orphan,
2347 };
2348
2349 DECLARE_GEOM_CLASS(g_swap_class, g_class);
2350
2351
2352 static void
2353 swapgeom_done(struct bio *bp2)
2354 {
2355 struct buf *bp;
2356
2357 bp = bp2->bio_caller2;
2358 bp->b_ioflags = bp2->bio_flags;
2359 if (bp2->bio_error)
2360 bp->b_ioflags |= BIO_ERROR;
2361 bp->b_resid = bp->b_bcount - bp2->bio_completed;
2362 bp->b_error = bp2->bio_error;
2363 bufdone(bp);
2364 g_destroy_bio(bp2);
2365 }
2366
2367 static void
2368 swapgeom_strategy(struct buf *bp, struct swdevt *sp)
2369 {
2370 struct bio *bio;
2371 struct g_consumer *cp;
2372
2373 cp = sp->sw_id;
2374 if (cp == NULL) {
2375 bp->b_error = ENXIO;
2376 bp->b_ioflags |= BIO_ERROR;
2377 bufdone(bp);
2378 return;
2379 }
2380 bio = g_alloc_bio();
2381 #if 0
2382 /*
2383 * XXX: We shouldn't really sleep here when we run out of buffers
2384 * XXX: but the alternative is worse right now.
2385 */
2386 if (bio == NULL) {
2387 bp->b_error = ENOMEM;
2388 bp->b_ioflags |= BIO_ERROR;
2389 bufdone(bp);
2390 return;
2391 }
2392 #endif
2393 bio->bio_caller2 = bp;
2394 bio->bio_cmd = bp->b_iocmd;
2395 bio->bio_data = bp->b_data;
2396 bio->bio_offset = (bp->b_blkno - sp->sw_first) * PAGE_SIZE;
2397 bio->bio_length = bp->b_bcount;
2398 bio->bio_done = swapgeom_done;
2399 g_io_request(bio, cp);
2400 return;
2401 }
2402
2403 static void
2404 swapgeom_orphan(struct g_consumer *cp)
2405 {
2406 struct swdevt *sp;
2407
2408 mtx_lock(&sw_dev_mtx);
2409 TAILQ_FOREACH(sp, &swtailq, sw_list)
2410 if (sp->sw_id == cp)
2411 sp->sw_id = NULL;
2412 mtx_unlock(&sw_dev_mtx);
2413 }
2414
2415 static void
2416 swapgeom_close_ev(void *arg, int flags)
2417 {
2418 struct g_consumer *cp;
2419
2420 cp = arg;
2421 g_access(cp, -1, -1, 0);
2422 g_detach(cp);
2423 g_destroy_consumer(cp);
2424 }
2425
2426 static void
2427 swapgeom_close(struct thread *td, struct swdevt *sw)
2428 {
2429
2430 /* XXX: direct call when Giant untangled */
2431 g_waitfor_event(swapgeom_close_ev, sw->sw_id, M_WAITOK, NULL);
2432 }
2433
2434
2435 struct swh0h0 {
2436 struct cdev *dev;
2437 struct vnode *vp;
2438 int error;
2439 };
2440
2441 static void
2442 swapongeom_ev(void *arg, int flags)
2443 {
2444 struct swh0h0 *swh;
2445 struct g_provider *pp;
2446 struct g_consumer *cp;
2447 static struct g_geom *gp;
2448 struct swdevt *sp;
2449 u_long nblks;
2450 int error;
2451
2452 swh = arg;
2453 swh->error = 0;
2454 pp = g_dev_getprovider(swh->dev);
2455 if (pp == NULL) {
2456 swh->error = ENODEV;
2457 return;
2458 }
2459 mtx_lock(&sw_dev_mtx);
2460 TAILQ_FOREACH(sp, &swtailq, sw_list) {
2461 cp = sp->sw_id;
2462 if (cp != NULL && cp->provider == pp) {
2463 mtx_unlock(&sw_dev_mtx);
2464 swh->error = EBUSY;
2465 return;
2466 }
2467 }
2468 mtx_unlock(&sw_dev_mtx);
2469 if (gp == NULL)
2470 gp = g_new_geomf(&g_swap_class, "swap", NULL);
2471 cp = g_new_consumer(gp);
2472 g_attach(cp, pp);
2473 /*
2474 * XXX: Everytime you think you can improve the margin for
2475 * footshooting, somebody depends on the ability to do so:
2476 * savecore(8) wants to write to our swapdev so we cannot
2477 * set an exclusive count :-(
2478 */
2479 error = g_access(cp, 1, 1, 0);
2480 if (error) {
2481 g_detach(cp);
2482 g_destroy_consumer(cp);
2483 swh->error = error;
2484 return;
2485 }
2486 nblks = pp->mediasize / DEV_BSIZE;
2487 swaponsomething(swh->vp, cp, nblks, swapgeom_strategy,
2488 swapgeom_close, dev2udev(swh->dev));
2489 swh->error = 0;
2490 return;
2491 }
2492
2493 static int
2494 swapongeom(struct thread *td, struct vnode *vp)
2495 {
2496 int error;
2497 struct swh0h0 swh;
2498
2499 vn_lock(vp, LK_EXCLUSIVE | LK_RETRY, td);
2500
2501 swh.dev = vp->v_rdev;
2502 swh.vp = vp;
2503 swh.error = 0;
2504 /* XXX: direct call when Giant untangled */
2505 error = g_waitfor_event(swapongeom_ev, &swh, M_WAITOK, NULL);
2506 if (!error)
2507 error = swh.error;
2508 VOP_UNLOCK(vp, 0, td);
2509 return (error);
2510 }
2511
2512 /*
2513 * VNODE backend
2514 *
2515 * This is used mainly for network filesystem (read: probably only tested
2516 * with NFS) swapfiles.
2517 *
2518 */
2519
2520 static void
2521 swapdev_strategy(struct buf *bp, struct swdevt *sp)
2522 {
2523 struct vnode *vp2;
2524
2525 bp->b_blkno = ctodb(bp->b_blkno - sp->sw_first);
2526
2527 vp2 = sp->sw_id;
2528 vhold(vp2);
2529 if (bp->b_iocmd == BIO_WRITE) {
2530 if (bp->b_bufobj) /* XXX: should always be true /phk */
2531 bufobj_wdrop(bp->b_bufobj);
2532 bufobj_wref(&vp2->v_bufobj);
2533 }
2534 if (bp->b_bufobj != &vp2->v_bufobj)
2535 bp->b_bufobj = &vp2->v_bufobj;
2536 bp->b_vp = vp2;
2537 bp->b_iooffset = dbtob(bp->b_blkno);
2538 bstrategy(bp);
2539 return;
2540 }
2541
2542 static void
2543 swapdev_close(struct thread *td, struct swdevt *sp)
2544 {
2545
2546 VOP_CLOSE(sp->sw_vp, FREAD | FWRITE, td->td_ucred, td);
2547 vrele(sp->sw_vp);
2548 }
2549
2550
2551 static int
2552 swaponvp(struct thread *td, struct vnode *vp, u_long nblks)
2553 {
2554 struct swdevt *sp;
2555 int error;
2556
2557 if (nblks == 0)
2558 return (ENXIO);
2559 mtx_lock(&sw_dev_mtx);
2560 TAILQ_FOREACH(sp, &swtailq, sw_list) {
2561 if (sp->sw_id == vp) {
2562 mtx_unlock(&sw_dev_mtx);
2563 return (EBUSY);
2564 }
2565 }
2566 mtx_unlock(&sw_dev_mtx);
2567
2568 (void) vn_lock(vp, LK_EXCLUSIVE | LK_RETRY, td);
2569 #ifdef MAC
2570 error = mac_check_system_swapon(td->td_ucred, vp);
2571 if (error == 0)
2572 #endif
2573 error = VOP_OPEN(vp, FREAD | FWRITE, td->td_ucred, td, -1);
2574 (void) VOP_UNLOCK(vp, 0, td);
2575 if (error)
2576 return (error);
2577
2578 swaponsomething(vp, vp, nblks, swapdev_strategy, swapdev_close,
2579 NODEV);
2580 return (0);
2581 }
Cache object: 75522af69d42193a8e5116bf3ba876c1
|