1 /*
2 * (MPSAFE)
3 *
4 * Copyright (c) 1998-2010 The DragonFly Project. All rights reserved.
5 *
6 * This code is derived from software contributed to The DragonFly Project
7 * by Matthew Dillon <dillon@backplane.com>
8 *
9 * Redistribution and use in source and binary forms, with or without
10 * modification, are permitted provided that the following conditions
11 * are met:
12 *
13 * 1. Redistributions of source code must retain the above copyright
14 * notice, this list of conditions and the following disclaimer.
15 * 2. Redistributions in binary form must reproduce the above copyright
16 * notice, this list of conditions and the following disclaimer in
17 * the documentation and/or other materials provided with the
18 * distribution.
19 * 3. Neither the name of The DragonFly Project nor the names of its
20 * contributors may be used to endorse or promote products derived
21 * from this software without specific, prior written permission.
22 *
23 * THIS SOFTWARE IS PROVIDED BY THE COPYRIGHT HOLDERS AND CONTRIBUTORS
24 * ``AS IS'' AND ANY EXPRESS OR IMPLIED WARRANTIES, INCLUDING, BUT NOT
25 * LIMITED TO, THE IMPLIED WARRANTIES OF MERCHANTABILITY AND FITNESS
26 * FOR A PARTICULAR PURPOSE ARE DISCLAIMED. IN NO EVENT SHALL THE
27 * COPYRIGHT HOLDERS OR CONTRIBUTORS BE LIABLE FOR ANY DIRECT, INDIRECT,
28 * INCIDENTAL, SPECIAL, EXEMPLARY OR CONSEQUENTIAL DAMAGES (INCLUDING,
29 * BUT NOT LIMITED TO, PROCUREMENT OF SUBSTITUTE GOODS OR SERVICES;
30 * LOSS OF USE, DATA, OR PROFITS; OR BUSINESS INTERRUPTION) HOWEVER CAUSED
31 * AND ON ANY THEORY OF LIABILITY, WHETHER IN CONTRACT, STRICT LIABILITY,
32 * OR TORT (INCLUDING NEGLIGENCE OR OTHERWISE) ARISING IN ANY WAY OUT
33 * OF THE USE OF THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF
34 * SUCH DAMAGE.
35 *
36 * Copyright (c) 1994 John S. Dyson
37 * Copyright (c) 1990 University of Utah.
38 * Copyright (c) 1991, 1993
39 * The Regents of the University of California. All rights reserved.
40 *
41 * This code is derived from software contributed to Berkeley by
42 * the Systems Programming Group of the University of Utah Computer
43 * Science Department.
44 *
45 * Redistribution and use in source and binary forms, with or without
46 * modification, are permitted provided that the following conditions
47 * are met:
48 * 1. Redistributions of source code must retain the above copyright
49 * notice, this list of conditions and the following disclaimer.
50 * 2. Redistributions in binary form must reproduce the above copyright
51 * notice, this list of conditions and the following disclaimer in the
52 * documentation and/or other materials provided with the distribution.
53 * 3. Neither the name of the University nor the names of its contributors
54 * may be used to endorse or promote products derived from this software
55 * without specific prior written permission.
56 *
57 * THIS SOFTWARE IS PROVIDED BY THE REGENTS AND CONTRIBUTORS ``AS IS'' AND
58 * ANY EXPRESS OR IMPLIED WARRANTIES, INCLUDING, BUT NOT LIMITED TO, THE
59 * IMPLIED WARRANTIES OF MERCHANTABILITY AND FITNESS FOR A PARTICULAR PURPOSE
60 * ARE DISCLAIMED. IN NO EVENT SHALL THE REGENTS OR CONTRIBUTORS BE LIABLE
61 * FOR ANY DIRECT, INDIRECT, INCIDENTAL, SPECIAL, EXEMPLARY, OR CONSEQUENTIAL
62 * DAMAGES (INCLUDING, BUT NOT LIMITED TO, PROCUREMENT OF SUBSTITUTE GOODS
63 * OR SERVICES; LOSS OF USE, DATA, OR PROFITS; OR BUSINESS INTERRUPTION)
64 * HOWEVER CAUSED AND ON ANY THEORY OF LIABILITY, WHETHER IN CONTRACT, STRICT
65 * LIABILITY, OR TORT (INCLUDING NEGLIGENCE OR OTHERWISE) ARISING IN ANY WAY
66 * OUT OF THE USE OF THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF
67 * SUCH DAMAGE.
68 *
69 * New Swap System
70 * Matthew Dillon
71 *
72 * Radix Bitmap 'blists'.
73 *
74 * - The new swapper uses the new radix bitmap code. This should scale
75 * to arbitrarily small or arbitrarily large swap spaces and an almost
76 * arbitrary degree of fragmentation.
77 *
78 * Features:
79 *
80 * - on the fly reallocation of swap during putpages. The new system
81 * does not try to keep previously allocated swap blocks for dirty
82 * pages.
83 *
84 * - on the fly deallocation of swap
85 *
86 * - No more garbage collection required. Unnecessarily allocated swap
87 * blocks only exist for dirty vm_page_t's now and these are already
88 * cycled (in a high-load system) by the pager. We also do on-the-fly
89 * removal of invalidated swap blocks when a page is destroyed
90 * or renamed.
91 *
92 * from: Utah $Hdr: swap_pager.c 1.4 91/04/30$
93 * @(#)swap_pager.c 8.9 (Berkeley) 3/21/94
94 * $FreeBSD: src/sys/vm/swap_pager.c,v 1.130.2.12 2002/08/31 21:15:55 dillon Exp $
95 */
96
97 #include "opt_swap.h"
98 #include <sys/param.h>
99 #include <sys/systm.h>
100 #include <sys/conf.h>
101 #include <sys/kernel.h>
102 #include <sys/proc.h>
103 #include <sys/buf.h>
104 #include <sys/vnode.h>
105 #include <sys/malloc.h>
106 #include <sys/vmmeter.h>
107 #include <sys/sysctl.h>
108 #include <sys/blist.h>
109 #include <sys/lock.h>
110 #include <sys/kcollect.h>
111
112 #include <vm/vm.h>
113 #include <vm/vm_object.h>
114 #include <vm/vm_page.h>
115 #include <vm/vm_pager.h>
116 #include <vm/vm_pageout.h>
117 #include <vm/swap_pager.h>
118 #include <vm/vm_extern.h>
119 #include <vm/vm_zone.h>
120 #include <vm/vnode_pager.h>
121
122 #include <sys/buf2.h>
123 #include <vm/vm_page2.h>
124
125 #ifndef MAX_PAGEOUT_CLUSTER
126 #define MAX_PAGEOUT_CLUSTER SWB_NPAGES
127 #endif
128
129 #define SWM_FREE 0x02 /* free, period */
130 #define SWM_POP 0x04 /* pop out */
131
132 #define SWBIO_READ 0x01
133 #define SWBIO_WRITE 0x02
134 #define SWBIO_SYNC 0x04
135 #define SWBIO_TTC 0x08 /* for OBJPC_TRY_TO_CACHE */
136
137 struct swfreeinfo {
138 vm_object_t object;
139 vm_pindex_t basei;
140 vm_pindex_t begi;
141 vm_pindex_t endi; /* inclusive */
142 };
143
144 struct swswapoffinfo {
145 vm_object_t object;
146 int devidx;
147 int shared;
148 };
149
150 /*
151 * vm_swap_size is in page-sized chunks now. It was DEV_BSIZE'd chunks
152 * in the old system.
153 */
154
155 int swap_pager_full; /* swap space exhaustion (task killing) */
156 int swap_fail_ticks; /* when we became exhausted */
157 int swap_pager_almost_full; /* swap space exhaustion (w/ hysteresis)*/
158 swblk_t vm_swap_cache_use;
159 swblk_t vm_swap_anon_use;
160 static int vm_report_swap_allocs;
161
162 static struct krate kswaprate = { 1 };
163 static int nsw_rcount; /* free read buffers */
164 static int nsw_wcount_sync; /* limit write buffers / synchronous */
165 static int nsw_wcount_async; /* limit write buffers / asynchronous */
166 static int nsw_wcount_async_max;/* assigned maximum */
167 static int nsw_cluster_max; /* maximum VOP I/O allowed */
168
169 struct blist *swapblist;
170 static int swap_async_max = 4; /* maximum in-progress async I/O's */
171 static int swap_burst_read = 0; /* allow burst reading */
172 static swblk_t swapiterator; /* linearize allocations */
173 int swap_user_async = 0; /* user swap pager operation can be async */
174
175 static struct spinlock swapbp_spin = SPINLOCK_INITIALIZER(&swapbp_spin, "swapbp_spin");
176
177 /* from vm_swap.c */
178 extern struct vnode *swapdev_vp;
179 extern struct swdevt *swdevt;
180 extern int nswdev;
181
182 #define BLK2DEVIDX(blk) (nswdev > 1 ? blk / SWB_DMMAX % nswdev : 0)
183
184 SYSCTL_INT(_vm, OID_AUTO, swap_async_max,
185 CTLFLAG_RW, &swap_async_max, 0, "Maximum running async swap ops");
186 SYSCTL_INT(_vm, OID_AUTO, swap_burst_read,
187 CTLFLAG_RW, &swap_burst_read, 0, "Allow burst reads for pageins");
188 SYSCTL_INT(_vm, OID_AUTO, swap_user_async,
189 CTLFLAG_RW, &swap_user_async, 0, "Allow async uuser swap write I/O");
190
191 #if SWBLK_BITS == 64
192 SYSCTL_LONG(_vm, OID_AUTO, swap_cache_use,
193 CTLFLAG_RD, &vm_swap_cache_use, 0, "");
194 SYSCTL_LONG(_vm, OID_AUTO, swap_anon_use,
195 CTLFLAG_RD, &vm_swap_anon_use, 0, "");
196 SYSCTL_LONG(_vm, OID_AUTO, swap_free,
197 CTLFLAG_RD, &vm_swap_size, 0, "");
198 SYSCTL_LONG(_vm, OID_AUTO, swap_size,
199 CTLFLAG_RD, &vm_swap_max, 0, "");
200 #else
201 SYSCTL_INT(_vm, OID_AUTO, swap_cache_use,
202 CTLFLAG_RD, &vm_swap_cache_use, 0, "");
203 SYSCTL_INT(_vm, OID_AUTO, swap_anon_use,
204 CTLFLAG_RD, &vm_swap_anon_use, 0, "");
205 SYSCTL_INT(_vm, OID_AUTO, swap_free,
206 CTLFLAG_RD, &vm_swap_size, 0, "");
207 SYSCTL_INT(_vm, OID_AUTO, swap_size,
208 CTLFLAG_RD, &vm_swap_max, 0, "");
209 #endif
210 SYSCTL_INT(_vm, OID_AUTO, report_swap_allocs,
211 CTLFLAG_RW, &vm_report_swap_allocs, 0, "");
212
213 __read_mostly vm_zone_t swap_zone;
214
215 /*
216 * Red-Black tree for swblock entries
217 *
218 * The caller must hold vm_token
219 */
220 RB_GENERATE2(swblock_rb_tree, swblock, swb_entry, rb_swblock_compare,
221 vm_pindex_t, swb_index);
222
223 int
rb_swblock_compare(struct swblock * swb1,struct swblock * swb2)224 rb_swblock_compare(struct swblock *swb1, struct swblock *swb2)
225 {
226 if (swb1->swb_index < swb2->swb_index)
227 return(-1);
228 if (swb1->swb_index > swb2->swb_index)
229 return(1);
230 return(0);
231 }
232
233 static
234 int
rb_swblock_scancmp(struct swblock * swb,void * data)235 rb_swblock_scancmp(struct swblock *swb, void *data)
236 {
237 struct swfreeinfo *info = data;
238
239 if (swb->swb_index < info->basei)
240 return(-1);
241 if (swb->swb_index > info->endi)
242 return(1);
243 return(0);
244 }
245
246 static
247 int
rb_swblock_condcmp(struct swblock * swb,void * data)248 rb_swblock_condcmp(struct swblock *swb, void *data)
249 {
250 struct swfreeinfo *info = data;
251
252 if (swb->swb_index < info->basei)
253 return(-1);
254 return(0);
255 }
256
257 /*
258 * pagerops for OBJT_SWAP - "swap pager". Some ops are also global procedure
259 * calls hooked from other parts of the VM system and do not appear here.
260 * (see vm/swap_pager.h).
261 */
262
263 static void swap_pager_dealloc (vm_object_t object);
264 static int swap_pager_getpage (vm_object_t, vm_pindex_t, vm_page_t *, int);
265 static void swap_chain_iodone(struct bio *biox);
266
267 struct pagerops swappagerops = {
268 swap_pager_dealloc, /* deallocate an OBJT_SWAP object */
269 swap_pager_getpage, /* pagein */
270 swap_pager_putpages, /* pageout */
271 swap_pager_haspage /* get backing store status for page */
272 };
273
274 /*
275 * SWB_DMMAX is in page-sized chunks with the new swap system. It was
276 * dev-bsized chunks in the old. SWB_DMMAX is always a power of 2.
277 *
278 * swap_*() routines are externally accessible. swp_*() routines are
279 * internal.
280 */
281
282 int nswap_lowat = 128; /* in pages, swap_pager_almost_full warn */
283 int nswap_hiwat = 512; /* in pages, swap_pager_almost_full warn */
284
285 static __inline void swp_sizecheck (void);
286 static void swp_pager_async_iodone (struct bio *bio);
287
288 /*
289 * Swap bitmap functions
290 */
291
292 static __inline void swp_pager_freeswapspace(vm_object_t object,
293 swblk_t blk, int npages);
294 static __inline swblk_t swp_pager_getswapspace(vm_object_t object, int npages);
295
296 /*
297 * Metadata functions
298 */
299
300 static void swp_pager_meta_convert(vm_object_t);
301 static void swp_pager_meta_build(vm_object_t, vm_pindex_t, swblk_t);
302 static void swp_pager_meta_free(vm_object_t, vm_pindex_t, vm_pindex_t);
303 static void swp_pager_meta_free_all(vm_object_t);
304 static swblk_t swp_pager_meta_ctl(vm_object_t, vm_pindex_t, int);
305
306 /*
307 * SWP_SIZECHECK() - update swap_pager_full indication
308 *
309 * update the swap_pager_almost_full indication and warn when we are
310 * about to run out of swap space, using lowat/hiwat hysteresis.
311 *
312 * Clear swap_pager_full ( task killing ) indication when lowat is met.
313 *
314 * No restrictions on call
315 * This routine may not block.
316 * SMP races are ok.
317 */
318 static __inline void
swp_sizecheck(void)319 swp_sizecheck(void)
320 {
321 if (vm_swap_size < nswap_lowat) {
322 if (swap_pager_almost_full == 0) {
323 kprintf("swap_pager: out of swap space\n");
324 swap_pager_almost_full = 1;
325 swap_fail_ticks = ticks;
326 }
327 } else {
328 swap_pager_full = 0;
329 if (vm_swap_size > nswap_hiwat)
330 swap_pager_almost_full = 0;
331 }
332 }
333
334 /*
335 * Long-term data collection on 10-second interval. Return the value
336 * for KCOLLECT_SWAPPCT and set the values for SWAPANO and SWAPCCAC.
337 *
338 * Return total swap in the scale field. This can change if swap is
339 * regularly added or removed and may cause some historical confusion
340 * in that case, but SWAPPCT will always be historically accurate.
341 */
342
343 #define PTOB(value) ((uint64_t)(value) << PAGE_SHIFT)
344
345 static uint64_t
collect_swap_callback(int n)346 collect_swap_callback(int n)
347 {
348 uint64_t total = vm_swap_max;
349 uint64_t anon = vm_swap_anon_use;
350 uint64_t cache = vm_swap_cache_use;
351
352 if (total == 0) /* avoid divide by zero */
353 total = 1;
354 kcollect_setvalue(KCOLLECT_SWAPANO, PTOB(anon));
355 kcollect_setvalue(KCOLLECT_SWAPCAC, PTOB(cache));
356 kcollect_setscale(KCOLLECT_SWAPANO,
357 KCOLLECT_SCALE(KCOLLECT_SWAPANO_FORMAT, PTOB(total)));
358 kcollect_setscale(KCOLLECT_SWAPCAC,
359 KCOLLECT_SCALE(KCOLLECT_SWAPCAC_FORMAT, PTOB(total)));
360 return (((anon + cache) * 10000 + (total >> 1)) / total);
361 }
362
363 /*
364 * SWAP_PAGER_INIT() - initialize the swap pager!
365 *
366 * Expected to be started from system init. NOTE: This code is run
367 * before much else so be careful what you depend on. Most of the VM
368 * system has yet to be initialized at this point.
369 *
370 * Called from the low level boot code only.
371 */
372 static void
swap_pager_init(void * arg __unused)373 swap_pager_init(void *arg __unused)
374 {
375 kcollect_register(KCOLLECT_SWAPPCT, "swapuse", collect_swap_callback,
376 KCOLLECT_SCALE(KCOLLECT_SWAPPCT_FORMAT, 0));
377 kcollect_register(KCOLLECT_SWAPANO, "swapano", NULL,
378 KCOLLECT_SCALE(KCOLLECT_SWAPANO_FORMAT, 0));
379 kcollect_register(KCOLLECT_SWAPCAC, "swapcac", NULL,
380 KCOLLECT_SCALE(KCOLLECT_SWAPCAC_FORMAT, 0));
381 }
382 SYSINIT(vm_mem, SI_BOOT1_VM, SI_ORDER_THIRD, swap_pager_init, NULL);
383
384 /*
385 * SWAP_PAGER_SWAP_INIT() - swap pager initialization from pageout process
386 *
387 * Expected to be started from pageout process once, prior to entering
388 * its main loop.
389 *
390 * Called from the low level boot code only.
391 */
392 void
swap_pager_swap_init(void)393 swap_pager_swap_init(void)
394 {
395 int n, n2;
396
397 /*
398 * Number of in-transit swap bp operations. Don't
399 * exhaust the pbufs completely. Make sure we
400 * initialize workable values (0 will work for hysteresis
401 * but it isn't very efficient).
402 *
403 * The nsw_cluster_max is constrained by the number of pages an XIO
404 * holds, i.e., (MAXPHYS/PAGE_SIZE) and our locally defined
405 * MAX_PAGEOUT_CLUSTER. Also be aware that swap ops are
406 * constrained by the swap device interleave stripe size.
407 *
408 * Currently we hardwire nsw_wcount_async to 4. This limit is
409 * designed to prevent other I/O from having high latencies due to
410 * our pageout I/O. The value 4 works well for one or two active swap
411 * devices but is probably a little low if you have more. Even so,
412 * a higher value would probably generate only a limited improvement
413 * with three or four active swap devices since the system does not
414 * typically have to pageout at extreme bandwidths. We will want
415 * at least 2 per swap devices, and 4 is a pretty good value if you
416 * have one NFS swap device due to the command/ack latency over NFS.
417 * So it all works out pretty well.
418 */
419
420 nsw_cluster_max = min((MAXPHYS/PAGE_SIZE), MAX_PAGEOUT_CLUSTER);
421
422 nsw_rcount = (nswbuf_kva + 1) / 2;
423 nsw_wcount_sync = (nswbuf_kva + 3) / 4;
424 nsw_wcount_async = 4;
425 nsw_wcount_async_max = nsw_wcount_async;
426
427 /*
428 * The zone is dynamically allocated so generally size it to
429 * maxswzone (32MB to 256GB of KVM). Set a minimum size based
430 * on physical memory of around 8x (each swblock can hold 16 pages).
431 *
432 * With the advent of SSDs (vs HDs) the practical (swap:memory) ratio
433 * has increased dramatically.
434 */
435 n = vmstats.v_page_count / 2;
436 if (maxswzone && n < maxswzone / sizeof(struct swblock))
437 n = maxswzone / sizeof(struct swblock);
438 n2 = n;
439
440 do {
441 swap_zone = zinit(
442 "SWAPMETA",
443 sizeof(struct swblock),
444 n,
445 ZONE_INTERRUPT);
446 if (swap_zone != NULL)
447 break;
448 /*
449 * if the allocation failed, try a zone two thirds the
450 * size of the previous attempt.
451 */
452 n -= ((n + 2) / 3);
453 } while (n > 0);
454
455 if (swap_zone == NULL)
456 panic("swap_pager_swap_init: swap_zone == NULL");
457 if (n2 != n)
458 kprintf("Swap zone entries reduced from %d to %d.\n", n2, n);
459 }
460
461 /*
462 * SWAP_PAGER_ALLOC() - allocate a new OBJT_SWAP VM object and instantiate
463 * its metadata structures.
464 *
465 * This routine is called from the mmap and fork code to create a new
466 * OBJT_SWAP object. We do this by creating an OBJT_DEFAULT object
467 * and then converting it with swp_pager_meta_convert().
468 *
469 * We only support unnamed objects.
470 *
471 * No restrictions.
472 */
473 vm_object_t
swap_pager_alloc(void * handle,off_t size,vm_prot_t prot,off_t offset)474 swap_pager_alloc(void *handle, off_t size, vm_prot_t prot, off_t offset)
475 {
476 vm_object_t object;
477
478 KKASSERT(handle == NULL);
479 object = vm_object_allocate_hold(OBJT_DEFAULT,
480 OFF_TO_IDX(offset + PAGE_MASK + size));
481 swp_pager_meta_convert(object);
482 vm_object_drop(object);
483
484 return (object);
485 }
486
487 /*
488 * SWAP_PAGER_DEALLOC() - remove swap metadata from object
489 *
490 * The swap backing for the object is destroyed. The code is
491 * designed such that we can reinstantiate it later, but this
492 * routine is typically called only when the entire object is
493 * about to be destroyed.
494 *
495 * The object must be locked or unreferenceable.
496 * No other requirements.
497 */
498 static void
swap_pager_dealloc(vm_object_t object)499 swap_pager_dealloc(vm_object_t object)
500 {
501 vm_object_hold(object);
502 vm_object_pip_wait(object, "swpdea");
503
504 /*
505 * Free all remaining metadata. We only bother to free it from
506 * the swap meta data. We do not attempt to free swapblk's still
507 * associated with vm_page_t's for this object. We do not care
508 * if paging is still in progress on some objects.
509 */
510 swp_pager_meta_free_all(object);
511 vm_object_drop(object);
512 }
513
514 /************************************************************************
515 * SWAP PAGER BITMAP ROUTINES *
516 ************************************************************************/
517
518 /*
519 * SWP_PAGER_GETSWAPSPACE() - allocate raw swap space
520 *
521 * Allocate swap for the requested number of pages. The starting
522 * swap block number (a page index) is returned or SWAPBLK_NONE
523 * if the allocation failed.
524 *
525 * Also has the side effect of advising that somebody made a mistake
526 * when they configured swap and didn't configure enough.
527 *
528 * The caller must hold the object.
529 * This routine may not block.
530 */
531 static __inline swblk_t
swp_pager_getswapspace(vm_object_t object,int npages)532 swp_pager_getswapspace(vm_object_t object, int npages)
533 {
534 swblk_t blk;
535
536 lwkt_gettoken(&vm_token);
537 blk = blist_allocat(swapblist, npages, swapiterator);
538 if (blk == SWAPBLK_NONE)
539 blk = blist_allocat(swapblist, npages, 0);
540 if (blk == SWAPBLK_NONE) {
541 if (swap_pager_full != 2) {
542 if (vm_swap_max == 0) {
543 krateprintf(&kswaprate,
544 "Warning: The system would like to "
545 "page to swap but no swap space "
546 "is configured!\n");
547 } else {
548 krateprintf(&kswaprate,
549 "swap_pager_getswapspace: "
550 "swap full allocating %d pages\n",
551 npages);
552 }
553 swap_pager_full = 2;
554 if (swap_pager_almost_full == 0)
555 swap_fail_ticks = ticks;
556 swap_pager_almost_full = 1;
557 }
558 } else {
559 /* swapiterator = blk; disable for now, doesn't work well */
560 swapacctspace(blk, -npages);
561 if (object->type == OBJT_SWAP)
562 vm_swap_anon_use += npages;
563 else
564 vm_swap_cache_use += npages;
565 swp_sizecheck();
566 }
567 lwkt_reltoken(&vm_token);
568 return(blk);
569 }
570
571 /*
572 * SWP_PAGER_FREESWAPSPACE() - free raw swap space
573 *
574 * This routine returns the specified swap blocks back to the bitmap.
575 *
576 * Note: This routine may not block (it could in the old swap code),
577 * and through the use of the new blist routines it does not block.
578 *
579 * This routine may not block.
580 */
581
582 static __inline void
swp_pager_freeswapspace(vm_object_t object,swblk_t blk,int npages)583 swp_pager_freeswapspace(vm_object_t object, swblk_t blk, int npages)
584 {
585 struct swdevt *sp = &swdevt[BLK2DEVIDX(blk)];
586
587 lwkt_gettoken(&vm_token);
588 sp->sw_nused -= npages;
589 if (object->type == OBJT_SWAP)
590 vm_swap_anon_use -= npages;
591 else
592 vm_swap_cache_use -= npages;
593
594 if (sp->sw_flags & SW_CLOSING) {
595 lwkt_reltoken(&vm_token);
596 return;
597 }
598
599 blist_free(swapblist, blk, npages);
600 vm_swap_size += npages;
601 swp_sizecheck();
602 lwkt_reltoken(&vm_token);
603 }
604
605 /*
606 * SWAP_PAGER_FREESPACE() - frees swap blocks associated with a page
607 * range within an object.
608 *
609 * This is a globally accessible routine.
610 *
611 * This routine removes swapblk assignments from swap metadata.
612 *
613 * The external callers of this routine typically have already destroyed
614 * or renamed vm_page_t's associated with this range in the object so
615 * we should be ok.
616 *
617 * No requirements.
618 */
619 void
swap_pager_freespace(vm_object_t object,vm_pindex_t start,vm_pindex_t size)620 swap_pager_freespace(vm_object_t object, vm_pindex_t start, vm_pindex_t size)
621 {
622 if (object->swblock_count == 0)
623 return;
624 vm_object_hold(object);
625 swp_pager_meta_free(object, start, size);
626 vm_object_drop(object);
627 }
628
629 /*
630 * No requirements.
631 */
632 void
swap_pager_freespace_all(vm_object_t object)633 swap_pager_freespace_all(vm_object_t object)
634 {
635 if (object->swblock_count == 0)
636 return;
637 vm_object_hold(object);
638 swp_pager_meta_free_all(object);
639 vm_object_drop(object);
640 }
641
642 /*
643 * This function conditionally frees swap cache swap starting at
644 * (*basei) in the object. (count) swap blocks will be nominally freed.
645 * The actual number of blocks freed can be more or less than the
646 * requested number.
647 *
648 * This function nominally returns the number of blocks freed. However,
649 * the actual number of blocks freed may be less then the returned value.
650 * If the function is unable to exhaust the object or if it is able to
651 * free (approximately) the requested number of blocks it returns
652 * a value n > count.
653 *
654 * If we exhaust the object we will return a value n <= count.
655 *
656 * The caller must hold the object.
657 *
658 * WARNING! If count == 0 then -1 can be returned as a degenerate case,
659 * callers should always pass a count value > 0.
660 */
661 static int swap_pager_condfree_callback(struct swblock *swap, void *data);
662
663 int
swap_pager_condfree(vm_object_t object,vm_pindex_t * basei,int count)664 swap_pager_condfree(vm_object_t object, vm_pindex_t *basei, int count)
665 {
666 struct swfreeinfo info;
667 int n;
668 int t;
669
670 ASSERT_LWKT_TOKEN_HELD(vm_object_token(object));
671
672 info.object = object;
673 info.basei = *basei; /* skip up to this page index */
674 info.begi = count; /* max swap pages to destroy */
675 info.endi = count * 8; /* max swblocks to scan */
676
677 swblock_rb_tree_RB_SCAN(&object->swblock_root, rb_swblock_condcmp,
678 swap_pager_condfree_callback, &info);
679 *basei = info.basei;
680
681 /*
682 * Take the higher difference swblocks vs pages
683 */
684 n = count - (int)info.begi;
685 t = count * 8 - (int)info.endi;
686 if (n < t)
687 n = t;
688 if (n < 1)
689 n = 1;
690 return(n);
691 }
692
693 /*
694 * The idea is to free whole meta-block to avoid fragmenting
695 * the swap space or disk I/O. We only do this if NO VM pages
696 * are present.
697 *
698 * We do not have to deal with clearing PG_SWAPPED in related VM
699 * pages because there are no related VM pages.
700 *
701 * The caller must hold the object.
702 */
703 static int
swap_pager_condfree_callback(struct swblock * swap,void * data)704 swap_pager_condfree_callback(struct swblock *swap, void *data)
705 {
706 struct swfreeinfo *info = data;
707 vm_object_t object = info->object;
708 int i;
709
710 for (i = 0; i < SWAP_META_PAGES; ++i) {
711 if (vm_page_lookup(object, swap->swb_index + i))
712 break;
713 }
714 info->basei = swap->swb_index + SWAP_META_PAGES;
715 if (i == SWAP_META_PAGES) {
716 info->begi -= swap->swb_count;
717 swap_pager_freespace(object, swap->swb_index, SWAP_META_PAGES);
718 }
719 --info->endi;
720 if ((int)info->begi < 0 || (int)info->endi < 0)
721 return(-1);
722 lwkt_yield();
723 return(0);
724 }
725
726 /*
727 * Called by vm_page_alloc() when a new VM page is inserted
728 * into a VM object. Checks whether swap has been assigned to
729 * the page and sets PG_SWAPPED as necessary.
730 *
731 * (m) must be busied by caller and remains busied on return.
732 */
733 void
swap_pager_page_inserted(vm_page_t m)734 swap_pager_page_inserted(vm_page_t m)
735 {
736 if (m->object->swblock_count) {
737 vm_object_hold(m->object);
738 if (swp_pager_meta_ctl(m->object, m->pindex, 0) != SWAPBLK_NONE)
739 vm_page_flag_set(m, PG_SWAPPED);
740 vm_object_drop(m->object);
741 }
742 }
743
744 /*
745 * SWAP_PAGER_RESERVE() - reserve swap blocks in object
746 *
747 * Assigns swap blocks to the specified range within the object. The
748 * swap blocks are not zerod. Any previous swap assignment is destroyed.
749 *
750 * Returns 0 on success, -1 on failure.
751 *
752 * The caller is responsible for avoiding races in the specified range.
753 * No other requirements.
754 */
755 int
swap_pager_reserve(vm_object_t object,vm_pindex_t start,vm_size_t size)756 swap_pager_reserve(vm_object_t object, vm_pindex_t start, vm_size_t size)
757 {
758 int n = 0;
759 swblk_t blk = SWAPBLK_NONE;
760 vm_pindex_t beg = start; /* save start index */
761
762 vm_object_hold(object);
763
764 while (size) {
765 if (n == 0) {
766 n = BLIST_MAX_ALLOC;
767 while ((blk = swp_pager_getswapspace(object, n)) ==
768 SWAPBLK_NONE)
769 {
770 n >>= 1;
771 if (n == 0) {
772 swp_pager_meta_free(object, beg,
773 start - beg);
774 vm_object_drop(object);
775 return(-1);
776 }
777 }
778 }
779 swp_pager_meta_build(object, start, blk);
780 --size;
781 ++start;
782 ++blk;
783 --n;
784 }
785 swp_pager_meta_free(object, start, n);
786 vm_object_drop(object);
787 return(0);
788 }
789
790 /*
791 * SWAP_PAGER_COPY() - copy blocks from source pager to destination pager
792 * and destroy the source.
793 *
794 * Copy any valid swapblks from the source to the destination. In
795 * cases where both the source and destination have a valid swapblk,
796 * we keep the destination's.
797 *
798 * This routine is allowed to block. It may block allocating metadata
799 * indirectly through swp_pager_meta_build() or if paging is still in
800 * progress on the source.
801 *
802 * XXX vm_page_collapse() kinda expects us not to block because we
803 * supposedly do not need to allocate memory, but for the moment we
804 * *may* have to get a little memory from the zone allocator, but
805 * it is taken from the interrupt memory. We should be ok.
806 *
807 * The source object contains no vm_page_t's (which is just as well)
808 * The source object is of type OBJT_SWAP.
809 *
810 * The source and destination objects must be held by the caller.
811 */
812 void
swap_pager_copy(vm_object_t srcobject,vm_object_t dstobject,vm_pindex_t base_index,int destroysource)813 swap_pager_copy(vm_object_t srcobject, vm_object_t dstobject,
814 vm_pindex_t base_index, int destroysource)
815 {
816 vm_pindex_t i;
817
818 ASSERT_LWKT_TOKEN_HELD(vm_object_token(srcobject));
819 ASSERT_LWKT_TOKEN_HELD(vm_object_token(dstobject));
820
821 /*
822 * transfer source to destination.
823 */
824 for (i = 0; i < dstobject->size; ++i) {
825 swblk_t dstaddr;
826
827 /*
828 * Locate (without changing) the swapblk on the destination,
829 * unless it is invalid in which case free it silently, or
830 * if the destination is a resident page, in which case the
831 * source is thrown away.
832 */
833 dstaddr = swp_pager_meta_ctl(dstobject, i, 0);
834
835 if (dstaddr == SWAPBLK_NONE) {
836 /*
837 * Destination has no swapblk and is not resident,
838 * copy source.
839 */
840 swblk_t srcaddr;
841
842 srcaddr = swp_pager_meta_ctl(srcobject,
843 base_index + i, SWM_POP);
844
845 if (srcaddr != SWAPBLK_NONE)
846 swp_pager_meta_build(dstobject, i, srcaddr);
847 } else {
848 /*
849 * Destination has valid swapblk or it is represented
850 * by a resident page. We destroy the sourceblock.
851 */
852 swp_pager_meta_ctl(srcobject, base_index + i, SWM_FREE);
853 }
854 }
855
856 /*
857 * Free left over swap blocks in source.
858 *
859 * We have to revert the type to OBJT_DEFAULT so we do not accidently
860 * double-remove the object from the swap queues.
861 */
862 if (destroysource) {
863 /*
864 * Reverting the type is not necessary, the caller is going
865 * to destroy srcobject directly, but I'm doing it here
866 * for consistency since we've removed the object from its
867 * queues.
868 */
869 swp_pager_meta_free_all(srcobject);
870 if (srcobject->type == OBJT_SWAP)
871 srcobject->type = OBJT_DEFAULT;
872 }
873 }
874
875 /*
876 * SWAP_PAGER_HASPAGE() - determine if we have good backing store for
877 * the requested page.
878 *
879 * We determine whether good backing store exists for the requested
880 * page and return TRUE if it does, FALSE if it doesn't.
881 *
882 * If TRUE, we also try to determine how much valid, contiguous backing
883 * store exists before and after the requested page within a reasonable
884 * distance. We do not try to restrict it to the swap device stripe
885 * (that is handled in getpages/putpages). It probably isn't worth
886 * doing here.
887 *
888 * No requirements.
889 */
890 boolean_t
swap_pager_haspage(vm_object_t object,vm_pindex_t pindex)891 swap_pager_haspage(vm_object_t object, vm_pindex_t pindex)
892 {
893 swblk_t blk0;
894
895 /*
896 * do we have good backing store at the requested index ?
897 */
898 vm_object_hold(object);
899 blk0 = swp_pager_meta_ctl(object, pindex, 0);
900
901 if (blk0 == SWAPBLK_NONE) {
902 vm_object_drop(object);
903 return (FALSE);
904 }
905 vm_object_drop(object);
906 return (TRUE);
907 }
908
909 /*
910 * Object must be held exclusive or shared by the caller.
911 */
912 boolean_t
swap_pager_haspage_locked(vm_object_t object,vm_pindex_t pindex)913 swap_pager_haspage_locked(vm_object_t object, vm_pindex_t pindex)
914 {
915 if (swp_pager_meta_ctl(object, pindex, 0) == SWAPBLK_NONE)
916 return FALSE;
917 return TRUE;
918 }
919
920 /*
921 * SWAP_PAGER_PAGE_UNSWAPPED() - remove swap backing store related to page
922 *
923 * This removes any associated swap backing store, whether valid or
924 * not, from the page. This operates on any VM object, not just OBJT_SWAP
925 * objects.
926 *
927 * This routine is typically called when a page is made dirty, at
928 * which point any associated swap can be freed. MADV_FREE also
929 * calls us in a special-case situation
930 *
931 * NOTE!!! If the page is clean and the swap was valid, the caller
932 * should make the page dirty before calling this routine.
933 * This routine does NOT change the m->dirty status of the page.
934 * Also: MADV_FREE depends on it.
935 *
936 * The page must be busied.
937 * The caller can hold the object to avoid blocking, else we might block.
938 * No other requirements.
939 */
940 void
swap_pager_unswapped(vm_page_t m)941 swap_pager_unswapped(vm_page_t m)
942 {
943 if (m->flags & PG_SWAPPED) {
944 vm_object_hold(m->object);
945 KKASSERT(m->flags & PG_SWAPPED);
946 swp_pager_meta_ctl(m->object, m->pindex, SWM_FREE);
947 vm_page_flag_clear(m, PG_SWAPPED);
948 vm_object_drop(m->object);
949 }
950 }
951
952 /*
953 * SWAP_PAGER_STRATEGY() - read, write, free blocks
954 *
955 * This implements a VM OBJECT strategy function using swap backing store.
956 * This can operate on any VM OBJECT type, not necessarily just OBJT_SWAP
957 * types. Only BUF_CMD_{READ,WRITE,FREEBLKS} is supported, any other
958 * requests will return EINVAL.
959 *
960 * This is intended to be a cacheless interface (i.e. caching occurs at
961 * higher levels), and is also used as a swap-based SSD cache for vnode
962 * and device objects.
963 *
964 * All I/O goes directly to and from the swap device.
965 *
966 * We currently attempt to run I/O synchronously or asynchronously as
967 * the caller requests. This isn't perfect because we loose error
968 * sequencing when we run multiple ops in parallel to satisfy a request.
969 * But this is swap, so we let it all hang out.
970 *
971 * NOTE: This function supports the KVABIO API wherein bp->b_data might
972 * not be synchronized to the current cpu.
973 *
974 * No requirements.
975 */
976 void
swap_pager_strategy(vm_object_t object,struct bio * bio)977 swap_pager_strategy(vm_object_t object, struct bio *bio)
978 {
979 struct buf *bp = bio->bio_buf;
980 struct bio *nbio;
981 vm_pindex_t start;
982 vm_pindex_t biox_blkno = 0;
983 int count;
984 char *data;
985 struct bio *biox;
986 struct buf *bufx;
987 #if 0
988 struct bio_track *track;
989 #endif
990
991 #if 0
992 /*
993 * tracking for swapdev vnode I/Os
994 */
995 if (bp->b_cmd == BUF_CMD_READ)
996 track = &swapdev_vp->v_track_read;
997 else
998 track = &swapdev_vp->v_track_write;
999 #endif
1000
1001 /*
1002 * Only supported commands
1003 */
1004 if (bp->b_cmd != BUF_CMD_FREEBLKS &&
1005 bp->b_cmd != BUF_CMD_READ &&
1006 bp->b_cmd != BUF_CMD_WRITE) {
1007 bp->b_error = EINVAL;
1008 bp->b_flags |= B_ERROR | B_INVAL;
1009 biodone(bio);
1010 return;
1011 }
1012
1013 /*
1014 * bcount must be an integral number of pages.
1015 */
1016 if (bp->b_bcount & PAGE_MASK) {
1017 bp->b_error = EINVAL;
1018 bp->b_flags |= B_ERROR | B_INVAL;
1019 biodone(bio);
1020 kprintf("swap_pager_strategy: bp %p offset %lld size %d, "
1021 "not page bounded\n",
1022 bp, (long long)bio->bio_offset, (int)bp->b_bcount);
1023 return;
1024 }
1025
1026 /*
1027 * Clear error indication, initialize page index, count, data pointer.
1028 */
1029 bp->b_error = 0;
1030 bp->b_flags &= ~B_ERROR;
1031 bp->b_resid = bp->b_bcount;
1032
1033 start = (vm_pindex_t)(bio->bio_offset >> PAGE_SHIFT);
1034 count = howmany(bp->b_bcount, PAGE_SIZE);
1035
1036 /*
1037 * WARNING! Do not dereference *data without issuing a bkvasync()
1038 */
1039 data = bp->b_data;
1040
1041 /*
1042 * Deal with BUF_CMD_FREEBLKS
1043 */
1044 if (bp->b_cmd == BUF_CMD_FREEBLKS) {
1045 /*
1046 * FREE PAGE(s) - destroy underlying swap that is no longer
1047 * needed.
1048 */
1049 vm_object_hold(object);
1050 swp_pager_meta_free(object, start, count);
1051 vm_object_drop(object);
1052 bp->b_resid = 0;
1053 biodone(bio);
1054 return;
1055 }
1056
1057 /*
1058 * We need to be able to create a new cluster of I/O's. We cannot
1059 * use the caller fields of the passed bio so push a new one.
1060 *
1061 * Because nbio is just a placeholder for the cluster links,
1062 * we can biodone() the original bio instead of nbio to make
1063 * things a bit more efficient.
1064 */
1065 nbio = push_bio(bio);
1066 nbio->bio_offset = bio->bio_offset;
1067 nbio->bio_caller_info1.cluster_head = NULL;
1068 nbio->bio_caller_info2.cluster_tail = NULL;
1069
1070 biox = NULL;
1071 bufx = NULL;
1072
1073 /*
1074 * Execute read or write
1075 */
1076 vm_object_hold(object);
1077
1078 while (count > 0) {
1079 swblk_t blk;
1080
1081 /*
1082 * Obtain block. If block not found and writing, allocate a
1083 * new block and build it into the object.
1084 */
1085 blk = swp_pager_meta_ctl(object, start, 0);
1086 if ((blk == SWAPBLK_NONE) && bp->b_cmd == BUF_CMD_WRITE) {
1087 blk = swp_pager_getswapspace(object, 1);
1088 if (blk == SWAPBLK_NONE) {
1089 bp->b_error = ENOMEM;
1090 bp->b_flags |= B_ERROR;
1091 break;
1092 }
1093 swp_pager_meta_build(object, start, blk);
1094 }
1095
1096 /*
1097 * Do we have to flush our current collection? Yes if:
1098 *
1099 * - no swap block at this index
1100 * - swap block is not contiguous
1101 * - we cross a physical disk boundry in the
1102 * stripe.
1103 */
1104 if (biox &&
1105 (biox_blkno + btoc(bufx->b_bcount) != blk ||
1106 ((biox_blkno ^ blk) & ~SWB_DMMASK))) {
1107 switch(bp->b_cmd) {
1108 case BUF_CMD_READ:
1109 ++mycpu->gd_cnt.v_swapin;
1110 mycpu->gd_cnt.v_swappgsin +=
1111 btoc(bufx->b_bcount);
1112 break;
1113 case BUF_CMD_WRITE:
1114 ++mycpu->gd_cnt.v_swapout;
1115 mycpu->gd_cnt.v_swappgsout +=
1116 btoc(bufx->b_bcount);
1117 bufx->b_dirtyend = bufx->b_bcount;
1118 break;
1119 default:
1120 /* NOT REACHED */
1121 break;
1122 }
1123
1124 /*
1125 * Finished with this buf.
1126 */
1127 KKASSERT(bufx->b_bcount != 0);
1128 if (bufx->b_cmd != BUF_CMD_READ)
1129 bufx->b_dirtyend = bufx->b_bcount;
1130 biox = NULL;
1131 bufx = NULL;
1132 }
1133
1134 /*
1135 * Add new swapblk to biox, instantiating biox if necessary.
1136 * Zero-fill reads are able to take a shortcut.
1137 */
1138 if (blk == SWAPBLK_NONE) {
1139 /*
1140 * We can only get here if we are reading.
1141 */
1142 bkvasync(bp);
1143 bzero(data, PAGE_SIZE);
1144 bp->b_resid -= PAGE_SIZE;
1145 } else {
1146 if (biox == NULL) {
1147 /* XXX chain count > 4, wait to <= 4 */
1148
1149 bufx = getpbuf(NULL);
1150 bufx->b_flags |= B_KVABIO;
1151 biox = &bufx->b_bio1;
1152 cluster_append(nbio, bufx);
1153 bufx->b_cmd = bp->b_cmd;
1154 biox->bio_done = swap_chain_iodone;
1155 biox->bio_offset = (off_t)blk << PAGE_SHIFT;
1156 biox->bio_caller_info1.cluster_parent = nbio;
1157 biox_blkno = blk;
1158 bufx->b_bcount = 0;
1159 bufx->b_data = data;
1160 }
1161 bufx->b_bcount += PAGE_SIZE;
1162 }
1163 --count;
1164 ++start;
1165 data += PAGE_SIZE;
1166 }
1167
1168 vm_object_drop(object);
1169
1170 /*
1171 * Flush out last buffer
1172 */
1173 if (biox) {
1174 if (bufx->b_cmd == BUF_CMD_READ) {
1175 ++mycpu->gd_cnt.v_swapin;
1176 mycpu->gd_cnt.v_swappgsin += btoc(bufx->b_bcount);
1177 } else {
1178 ++mycpu->gd_cnt.v_swapout;
1179 mycpu->gd_cnt.v_swappgsout += btoc(bufx->b_bcount);
1180 bufx->b_dirtyend = bufx->b_bcount;
1181 }
1182 KKASSERT(bufx->b_bcount);
1183 if (bufx->b_cmd != BUF_CMD_READ)
1184 bufx->b_dirtyend = bufx->b_bcount;
1185 /* biox, bufx = NULL */
1186 }
1187
1188 /*
1189 * Now initiate all the I/O. Be careful looping on our chain as
1190 * I/O's may complete while we are still initiating them.
1191 *
1192 * If the request is a 100% sparse read no bios will be present
1193 * and we just biodone() the buffer.
1194 */
1195 nbio->bio_caller_info2.cluster_tail = NULL;
1196 bufx = nbio->bio_caller_info1.cluster_head;
1197
1198 if (bufx) {
1199 while (bufx) {
1200 biox = &bufx->b_bio1;
1201 BUF_KERNPROC(bufx);
1202 bufx = bufx->b_cluster_next;
1203 vn_strategy(swapdev_vp, biox);
1204 }
1205 } else {
1206 biodone(bio);
1207 }
1208
1209 /*
1210 * Completion of the cluster will also call biodone_chain(nbio).
1211 * We never call biodone(nbio) so we don't have to worry about
1212 * setting up a bio_done callback. It's handled in the sub-IO.
1213 */
1214 /**/
1215 }
1216
1217 /*
1218 * biodone callback
1219 *
1220 * No requirements.
1221 */
1222 static void
swap_chain_iodone(struct bio * biox)1223 swap_chain_iodone(struct bio *biox)
1224 {
1225 struct buf **nextp;
1226 struct buf *bufx; /* chained sub-buffer */
1227 struct bio *nbio; /* parent nbio with chain glue */
1228 struct buf *bp; /* original bp associated with nbio */
1229 int chain_empty;
1230
1231 bufx = biox->bio_buf;
1232 nbio = biox->bio_caller_info1.cluster_parent;
1233 bp = nbio->bio_buf;
1234
1235 /*
1236 * Update the original buffer
1237 */
1238 KKASSERT(bp != NULL);
1239 if (bufx->b_flags & B_ERROR) {
1240 atomic_set_int(&bufx->b_flags, B_ERROR);
1241 bp->b_error = bufx->b_error; /* race ok */
1242 } else if (bufx->b_resid != 0) {
1243 atomic_set_int(&bufx->b_flags, B_ERROR);
1244 bp->b_error = EINVAL; /* race ok */
1245 } else {
1246 atomic_subtract_int(&bp->b_resid, bufx->b_bcount);
1247 }
1248
1249 /*
1250 * Remove us from the chain.
1251 */
1252 spin_lock(&swapbp_spin);
1253 nextp = &nbio->bio_caller_info1.cluster_head;
1254 while (*nextp != bufx) {
1255 KKASSERT(*nextp != NULL);
1256 nextp = &(*nextp)->b_cluster_next;
1257 }
1258 *nextp = bufx->b_cluster_next;
1259 chain_empty = (nbio->bio_caller_info1.cluster_head == NULL);
1260 spin_unlock(&swapbp_spin);
1261
1262 /*
1263 * Clean up bufx. If the chain is now empty we finish out
1264 * the parent. Note that we may be racing other completions
1265 * so we must use the chain_empty status from above.
1266 */
1267 if (chain_empty) {
1268 if (bp->b_resid != 0 && !(bp->b_flags & B_ERROR)) {
1269 atomic_set_int(&bp->b_flags, B_ERROR);
1270 bp->b_error = EINVAL;
1271 }
1272 biodone_chain(nbio);
1273 }
1274 relpbuf(bufx, NULL);
1275 }
1276
1277 /*
1278 * SWAP_PAGER_GETPAGES() - bring page in from swap
1279 *
1280 * The requested page may have to be brought in from swap. Calculate the
1281 * swap block and bring in additional pages if possible. All pages must
1282 * have contiguous swap block assignments and reside in the same object.
1283 *
1284 * The caller has a single vm_object_pip_add() reference prior to
1285 * calling us and we should return with the same.
1286 *
1287 * The caller has BUSY'd the page. We should return with (*mpp) left busy,
1288 * and any additinal pages unbusied.
1289 *
1290 * If the caller encounters a PG_RAM page it will pass it to us even though
1291 * it may be valid and dirty. We cannot overwrite the page in this case!
1292 * The case is used to allow us to issue pure read-aheads.
1293 *
1294 * NOTE! XXX This code does not entirely pipeline yet due to the fact that
1295 * the PG_RAM page is validated at the same time as mreq. What we
1296 * really need to do is issue a separate read-ahead pbuf.
1297 *
1298 * No requirements.
1299 */
1300 static int
swap_pager_getpage(vm_object_t object,vm_pindex_t pindex,vm_page_t * mpp,int seqaccess)1301 swap_pager_getpage(vm_object_t object, vm_pindex_t pindex,
1302 vm_page_t *mpp, int seqaccess)
1303 {
1304 struct buf *bp;
1305 struct bio *bio;
1306 vm_page_t mreq;
1307 vm_page_t m;
1308 vm_offset_t kva;
1309 swblk_t blk;
1310 int i;
1311 int j;
1312 int raonly;
1313 int error;
1314 u_int32_t busy_count;
1315 vm_page_t marray[XIO_INTERNAL_PAGES];
1316
1317 mreq = *mpp;
1318
1319 vm_object_hold(object);
1320 if (mreq->object != object) {
1321 panic("swap_pager_getpages: object mismatch %p/%p",
1322 object,
1323 mreq->object
1324 );
1325 }
1326
1327 /*
1328 * We don't want to overwrite a fully valid page as it might be
1329 * dirty. This case can occur when e.g. vm_fault hits a perfectly
1330 * valid page with PG_RAM set.
1331 *
1332 * In this case we see if the next page is a suitable page-in
1333 * candidate and if it is we issue read-ahead. PG_RAM will be
1334 * set on the last page of the read-ahead to continue the pipeline.
1335 */
1336 if (mreq->valid == VM_PAGE_BITS_ALL) {
1337 if (swap_burst_read == 0 || mreq->pindex + 1 >= object->size) {
1338 vm_object_drop(object);
1339 return(VM_PAGER_OK);
1340 }
1341 blk = swp_pager_meta_ctl(object, mreq->pindex + 1, 0);
1342 if (blk == SWAPBLK_NONE) {
1343 vm_object_drop(object);
1344 return(VM_PAGER_OK);
1345 }
1346 m = vm_page_lookup_busy_try(object, mreq->pindex + 1,
1347 TRUE, &error);
1348 if (error) {
1349 vm_object_drop(object);
1350 return(VM_PAGER_OK);
1351 } else if (m == NULL) {
1352 /*
1353 * Use VM_ALLOC_QUICK to avoid blocking on cache
1354 * page reuse.
1355 */
1356 m = vm_page_alloc(object, mreq->pindex + 1,
1357 VM_ALLOC_QUICK);
1358 if (m == NULL) {
1359 vm_object_drop(object);
1360 return(VM_PAGER_OK);
1361 }
1362 } else {
1363 if (m->valid) {
1364 vm_page_wakeup(m);
1365 vm_object_drop(object);
1366 return(VM_PAGER_OK);
1367 }
1368 vm_page_unqueue_nowakeup(m);
1369 }
1370 /* page is busy */
1371 mreq = m;
1372 raonly = 1;
1373 } else {
1374 raonly = 0;
1375 }
1376
1377 /*
1378 * Try to block-read contiguous pages from swap if sequential,
1379 * otherwise just read one page. Contiguous pages from swap must
1380 * reside within a single device stripe because the I/O cannot be
1381 * broken up across multiple stripes.
1382 *
1383 * Note that blk and iblk can be SWAPBLK_NONE but the loop is
1384 * set up such that the case(s) are handled implicitly.
1385 */
1386 blk = swp_pager_meta_ctl(mreq->object, mreq->pindex, 0);
1387 marray[0] = mreq;
1388
1389 for (i = 1; i <= swap_burst_read &&
1390 i < XIO_INTERNAL_PAGES &&
1391 mreq->pindex + i < object->size; ++i) {
1392 swblk_t iblk;
1393
1394 iblk = swp_pager_meta_ctl(object, mreq->pindex + i, 0);
1395 if (iblk != blk + i)
1396 break;
1397 if ((blk ^ iblk) & ~SWB_DMMASK)
1398 break;
1399 m = vm_page_lookup_busy_try(object, mreq->pindex + i,
1400 TRUE, &error);
1401 if (error) {
1402 break;
1403 } else if (m == NULL) {
1404 /*
1405 * Use VM_ALLOC_QUICK to avoid blocking on cache
1406 * page reuse.
1407 */
1408 m = vm_page_alloc(object, mreq->pindex + i,
1409 VM_ALLOC_QUICK);
1410 if (m == NULL)
1411 break;
1412 } else {
1413 if (m->valid) {
1414 vm_page_wakeup(m);
1415 break;
1416 }
1417 vm_page_unqueue_nowakeup(m);
1418 }
1419 /* page is busy */
1420 marray[i] = m;
1421 }
1422 if (i > 1)
1423 vm_page_flag_set(marray[i - 1], PG_RAM);
1424
1425 /*
1426 * If mreq is the requested page and we have nothing to do return
1427 * VM_PAGER_FAIL. If raonly is set mreq is just another read-ahead
1428 * page and must be cleaned up.
1429 */
1430 if (blk == SWAPBLK_NONE) {
1431 KKASSERT(i == 1);
1432 if (raonly) {
1433 vnode_pager_freepage(mreq);
1434 vm_object_drop(object);
1435 return(VM_PAGER_OK);
1436 } else {
1437 vm_object_drop(object);
1438 return(VM_PAGER_FAIL);
1439 }
1440 }
1441
1442 /*
1443 * Map our page(s) into kva for input
1444 *
1445 * Use the KVABIO API to avoid synchronizing the pmap.
1446 */
1447 bp = getpbuf_kva(&nsw_rcount);
1448 bio = &bp->b_bio1;
1449 kva = (vm_offset_t) bp->b_kvabase;
1450 bcopy(marray, bp->b_xio.xio_pages, i * sizeof(vm_page_t));
1451 pmap_qenter_noinval(kva, bp->b_xio.xio_pages, i);
1452
1453 bp->b_data = (caddr_t)kva;
1454 bp->b_bcount = PAGE_SIZE * i;
1455 bp->b_xio.xio_npages = i;
1456 bp->b_flags |= B_KVABIO;
1457 bio->bio_done = swp_pager_async_iodone;
1458 bio->bio_offset = (off_t)blk << PAGE_SHIFT;
1459 bio->bio_caller_info1.index = SWBIO_READ;
1460
1461 /*
1462 * Set index. If raonly set the index beyond the array so all
1463 * the pages are treated the same, otherwise the original mreq is
1464 * at index 0.
1465 */
1466 if (raonly)
1467 bio->bio_driver_info = (void *)(intptr_t)i;
1468 else
1469 bio->bio_driver_info = (void *)(intptr_t)0;
1470
1471 for (j = 0; j < i; ++j) {
1472 atomic_set_int(&bp->b_xio.xio_pages[j]->busy_count,
1473 PBUSY_SWAPINPROG);
1474 }
1475
1476 mycpu->gd_cnt.v_swapin++;
1477 mycpu->gd_cnt.v_swappgsin += bp->b_xio.xio_npages;
1478
1479 /*
1480 * We still hold the lock on mreq, and our automatic completion routine
1481 * does not remove it.
1482 */
1483 vm_object_pip_add(object, bp->b_xio.xio_npages);
1484
1485 /*
1486 * perform the I/O. NOTE!!! bp cannot be considered valid after
1487 * this point because we automatically release it on completion.
1488 * Instead, we look at the one page we are interested in which we
1489 * still hold a lock on even through the I/O completion.
1490 *
1491 * The other pages in our m[] array are also released on completion,
1492 * so we cannot assume they are valid anymore either.
1493 */
1494 bp->b_cmd = BUF_CMD_READ;
1495 BUF_KERNPROC(bp);
1496 vn_strategy(swapdev_vp, bio);
1497
1498 /*
1499 * Wait for the page we want to complete. PBUSY_SWAPINPROG is always
1500 * cleared on completion. If an I/O error occurs, SWAPBLK_NONE
1501 * is set in the meta-data.
1502 *
1503 * If this is a read-ahead only we return immediately without
1504 * waiting for I/O.
1505 */
1506 if (raonly) {
1507 vm_object_drop(object);
1508 return(VM_PAGER_OK);
1509 }
1510
1511 /*
1512 * Read-ahead includes originally requested page case.
1513 */
1514 for (;;) {
1515 busy_count = mreq->busy_count;
1516 cpu_ccfence();
1517 if ((busy_count & PBUSY_SWAPINPROG) == 0)
1518 break;
1519 tsleep_interlock(mreq, 0);
1520 if (!atomic_cmpset_int(&mreq->busy_count, busy_count,
1521 busy_count |
1522 PBUSY_SWAPINPROG | PBUSY_WANTED)) {
1523 continue;
1524 }
1525 atomic_set_int(&mreq->flags, PG_REFERENCED);
1526 mycpu->gd_cnt.v_intrans++;
1527 if (tsleep(mreq, PINTERLOCKED, "swread", hz*20)) {
1528 kprintf(
1529 "swap_pager: indefinite wait buffer: "
1530 " bp %p offset: %lld, size: %ld "
1531 " m=%p busy=%08x flags=%08x\n",
1532 bp,
1533 (long long)bio->bio_offset,
1534 (long)bp->b_bcount,
1535 mreq, mreq->busy_count, mreq->flags);
1536 }
1537 }
1538
1539 /*
1540 * Disallow speculative reads prior to the SWAPINPROG test.
1541 */
1542 cpu_lfence();
1543
1544 /*
1545 * mreq is left busied after completion, but all the other pages
1546 * are freed. If we had an unrecoverable read error the page will
1547 * not be valid.
1548 */
1549 vm_object_drop(object);
1550 if (mreq->valid != VM_PAGE_BITS_ALL)
1551 return(VM_PAGER_ERROR);
1552 else
1553 return(VM_PAGER_OK);
1554
1555 /*
1556 * A final note: in a low swap situation, we cannot deallocate swap
1557 * and mark a page dirty here because the caller is likely to mark
1558 * the page clean when we return, causing the page to possibly revert
1559 * to all-zero's later.
1560 */
1561 }
1562
1563 /*
1564 * swap_pager_putpages:
1565 *
1566 * Assign swap (if necessary) and initiate I/O on the specified pages.
1567 *
1568 * We support both OBJT_DEFAULT and OBJT_SWAP objects. DEFAULT objects
1569 * are automatically converted to SWAP objects.
1570 *
1571 * In a low memory situation we may block in vn_strategy(), but the new
1572 * vm_page reservation system coupled with properly written VFS devices
1573 * should ensure that no low-memory deadlock occurs. This is an area
1574 * which needs work.
1575 *
1576 * The parent has N vm_object_pip_add() references prior to
1577 * calling us and will remove references for rtvals[] that are
1578 * not set to VM_PAGER_PEND. We need to remove the rest on I/O
1579 * completion.
1580 *
1581 * The parent has soft-busy'd the pages it passes us and will unbusy
1582 * those whos rtvals[] entry is not set to VM_PAGER_PEND on return.
1583 * We need to unbusy the rest on I/O completion.
1584 *
1585 * No requirements.
1586 */
1587 void
swap_pager_putpages(vm_object_t object,vm_page_t * m,int count,int flags,int * rtvals)1588 swap_pager_putpages(vm_object_t object, vm_page_t *m, int count,
1589 int flags, int *rtvals)
1590 {
1591 int i;
1592 int n = 0;
1593
1594 vm_object_hold(object);
1595
1596 if (count && m[0]->object != object) {
1597 panic("swap_pager_getpages: object mismatch %p/%p",
1598 object,
1599 m[0]->object
1600 );
1601 }
1602
1603 /*
1604 * Step 1
1605 *
1606 * Turn object into OBJT_SWAP
1607 * Check for bogus sysops
1608 *
1609 * Force sync if not pageout process, we don't want any single
1610 * non-pageout process to be able to hog the I/O subsystem! This
1611 * can be overridden by setting.
1612 */
1613 if (object->type == OBJT_DEFAULT) {
1614 if (object->type == OBJT_DEFAULT)
1615 swp_pager_meta_convert(object);
1616 }
1617
1618 /*
1619 * Normally we force synchronous swap I/O if this is not the
1620 * pageout daemon to prevent any single user process limited
1621 * via RLIMIT_RSS from hogging swap write bandwidth.
1622 */
1623 if (curthread != pagethread &&
1624 curthread != emergpager &&
1625 swap_user_async == 0) {
1626 flags |= OBJPC_SYNC;
1627 }
1628
1629 /*
1630 * Step 2
1631 *
1632 * Update nsw parameters from swap_async_max sysctl values.
1633 * Do not let the sysop crash the machine with bogus numbers.
1634 */
1635 if (swap_async_max != nsw_wcount_async_max) {
1636 int n;
1637
1638 /*
1639 * limit range
1640 */
1641 if ((n = swap_async_max) > nswbuf_kva / 2)
1642 n = nswbuf_kva / 2;
1643 if (n < 1)
1644 n = 1;
1645 swap_async_max = n;
1646
1647 /*
1648 * Adjust difference ( if possible ). If the current async
1649 * count is too low, we may not be able to make the adjustment
1650 * at this time.
1651 *
1652 * vm_token needed for nsw_wcount sleep interlock
1653 */
1654 lwkt_gettoken(&vm_token);
1655 n -= nsw_wcount_async_max;
1656 if (nsw_wcount_async + n >= 0) {
1657 nsw_wcount_async_max += n;
1658 pbuf_adjcount(&nsw_wcount_async, n);
1659 }
1660 lwkt_reltoken(&vm_token);
1661 }
1662
1663 /*
1664 * Step 3
1665 *
1666 * Assign swap blocks and issue I/O. We reallocate swap on the fly.
1667 * The page is left dirty until the pageout operation completes
1668 * successfully.
1669 */
1670
1671 for (i = 0; i < count; i += n) {
1672 struct buf *bp;
1673 struct bio *bio;
1674 swblk_t blk;
1675 int j;
1676
1677 /*
1678 * Maximum I/O size is limited by a number of factors.
1679 */
1680
1681 n = min(BLIST_MAX_ALLOC, count - i);
1682 n = min(n, nsw_cluster_max);
1683
1684 lwkt_gettoken(&vm_token);
1685
1686 /*
1687 * Get biggest block of swap we can. If we fail, fall
1688 * back and try to allocate a smaller block. Don't go
1689 * overboard trying to allocate space if it would overly
1690 * fragment swap.
1691 */
1692 while (
1693 (blk = swp_pager_getswapspace(object, n)) == SWAPBLK_NONE &&
1694 n > 4
1695 ) {
1696 n >>= 1;
1697 }
1698 if (blk == SWAPBLK_NONE) {
1699 for (j = 0; j < n; ++j)
1700 rtvals[i+j] = VM_PAGER_FAIL;
1701 lwkt_reltoken(&vm_token);
1702 continue;
1703 }
1704 if (vm_report_swap_allocs > 0) {
1705 kprintf("swap_alloc %08jx,%d\n", (intmax_t)blk, n);
1706 --vm_report_swap_allocs;
1707 }
1708
1709 /*
1710 * The I/O we are constructing cannot cross a physical
1711 * disk boundry in the swap stripe.
1712 */
1713 if ((blk ^ (blk + n)) & ~SWB_DMMASK) {
1714 j = ((blk + SWB_DMMAX) & ~SWB_DMMASK) - blk;
1715 swp_pager_freeswapspace(object, blk + j, n - j);
1716 n = j;
1717 }
1718
1719 /*
1720 * All I/O parameters have been satisfied, build the I/O
1721 * request and assign the swap space.
1722 *
1723 * Use the KVABIO API to avoid synchronizing the pmap.
1724 */
1725 if ((flags & OBJPC_SYNC))
1726 bp = getpbuf_kva(&nsw_wcount_sync);
1727 else
1728 bp = getpbuf_kva(&nsw_wcount_async);
1729 bio = &bp->b_bio1;
1730
1731 lwkt_reltoken(&vm_token);
1732
1733 pmap_qenter_noinval((vm_offset_t)bp->b_data, &m[i], n);
1734
1735 bp->b_flags |= B_KVABIO;
1736 bp->b_bcount = PAGE_SIZE * n;
1737 bio->bio_offset = (off_t)blk << PAGE_SHIFT;
1738
1739 for (j = 0; j < n; ++j) {
1740 vm_page_t mreq = m[i+j];
1741
1742 swp_pager_meta_build(mreq->object, mreq->pindex,
1743 blk + j);
1744 if (object->type == OBJT_SWAP)
1745 vm_page_dirty(mreq);
1746 rtvals[i+j] = VM_PAGER_OK;
1747
1748 atomic_set_int(&mreq->busy_count, PBUSY_SWAPINPROG);
1749 bp->b_xio.xio_pages[j] = mreq;
1750 }
1751 bp->b_xio.xio_npages = n;
1752
1753 mycpu->gd_cnt.v_swapout++;
1754 mycpu->gd_cnt.v_swappgsout += bp->b_xio.xio_npages;
1755
1756 bp->b_dirtyoff = 0; /* req'd for NFS */
1757 bp->b_dirtyend = bp->b_bcount; /* req'd for NFS */
1758 bp->b_cmd = BUF_CMD_WRITE;
1759 bio->bio_caller_info1.index = SWBIO_WRITE;
1760
1761 /*
1762 * asynchronous
1763 */
1764 if ((flags & OBJPC_SYNC) == 0) {
1765 bio->bio_done = swp_pager_async_iodone;
1766 BUF_KERNPROC(bp);
1767 vn_strategy(swapdev_vp, bio);
1768
1769 for (j = 0; j < n; ++j)
1770 rtvals[i+j] = VM_PAGER_PEND;
1771 continue;
1772 }
1773
1774 /*
1775 * Issue synchrnously.
1776 *
1777 * Wait for the sync I/O to complete, then update rtvals.
1778 * We just set the rtvals[] to VM_PAGER_PEND so we can call
1779 * our async completion routine at the end, thus avoiding a
1780 * double-free.
1781 */
1782 bio->bio_caller_info1.index |= SWBIO_SYNC;
1783 if (flags & OBJPC_TRY_TO_CACHE)
1784 bio->bio_caller_info1.index |= SWBIO_TTC;
1785 bio->bio_done = biodone_sync;
1786 bio->bio_flags |= BIO_SYNC;
1787 vn_strategy(swapdev_vp, bio);
1788 biowait(bio, "swwrt");
1789
1790 for (j = 0; j < n; ++j)
1791 rtvals[i+j] = VM_PAGER_PEND;
1792
1793 /*
1794 * Now that we are through with the bp, we can call the
1795 * normal async completion, which frees everything up.
1796 */
1797 swp_pager_async_iodone(bio);
1798 }
1799 vm_object_drop(object);
1800 }
1801
1802 /*
1803 * No requirements.
1804 *
1805 * Recalculate the low and high-water marks.
1806 */
1807 void
swap_pager_newswap(void)1808 swap_pager_newswap(void)
1809 {
1810 /*
1811 * NOTE: vm_swap_max cannot exceed 1 billion blocks, which is the
1812 * limitation imposed by the blist code. Remember that this
1813 * will be divided by NSWAP_MAX (4), so each swap device is
1814 * limited to around a terrabyte.
1815 */
1816 if (vm_swap_max) {
1817 nswap_lowat = (int64_t)vm_swap_max * 4 / 100; /* 4% left */
1818 nswap_hiwat = (int64_t)vm_swap_max * 6 / 100; /* 6% left */
1819 kprintf("swap low/high-water marks set to %d/%d\n",
1820 nswap_lowat, nswap_hiwat);
1821 } else {
1822 nswap_lowat = 128;
1823 nswap_hiwat = 512;
1824 }
1825 swp_sizecheck();
1826 }
1827
1828 /*
1829 * swp_pager_async_iodone:
1830 *
1831 * Completion routine for asynchronous reads and writes from/to swap.
1832 * Also called manually by synchronous code to finish up a bp.
1833 *
1834 * For READ operations, the pages are BUSY'd. For WRITE operations,
1835 * the pages are vm_page_t->busy'd. For READ operations, we BUSY
1836 * unbusy all pages except the 'main' request page. For WRITE
1837 * operations, we vm_page_t->busy'd unbusy all pages ( we can do this
1838 * because we marked them all VM_PAGER_PEND on return from putpages ).
1839 *
1840 * This routine may not block.
1841 *
1842 * No requirements.
1843 */
1844 static void
swp_pager_async_iodone(struct bio * bio)1845 swp_pager_async_iodone(struct bio *bio)
1846 {
1847 struct buf *bp = bio->bio_buf;
1848 vm_object_t object = NULL;
1849 int i;
1850 int *nswptr;
1851
1852 /*
1853 * report error
1854 */
1855 if (bp->b_flags & B_ERROR) {
1856 kprintf(
1857 "swap_pager: I/O error - %s failed; offset %lld,"
1858 "size %ld, error %d\n",
1859 ((bio->bio_caller_info1.index & SWBIO_READ) ?
1860 "pagein" : "pageout"),
1861 (long long)bio->bio_offset,
1862 (long)bp->b_bcount,
1863 bp->b_error
1864 );
1865 }
1866
1867 /*
1868 * set object.
1869 */
1870 if (bp->b_xio.xio_npages)
1871 object = bp->b_xio.xio_pages[0]->object;
1872
1873 #if 0
1874 /* PMAP TESTING CODE (useful, keep it in but #if 0'd) */
1875 if (bio->bio_caller_info1.index & SWBIO_WRITE) {
1876 if (bio->bio_crc != iscsi_crc32(bp->b_data, bp->b_bcount)) {
1877 kprintf("SWAPOUT: BADCRC %08x %08x\n",
1878 bio->bio_crc,
1879 iscsi_crc32(bp->b_data, bp->b_bcount));
1880 for (i = 0; i < bp->b_xio.xio_npages; ++i) {
1881 vm_page_t m = bp->b_xio.xio_pages[i];
1882 if ((m->flags & PG_WRITEABLE) &&
1883 (pmap_mapped_sync(m) & PG_WRITEABLE)) {
1884 kprintf("SWAPOUT: "
1885 "%d/%d %p writable\n",
1886 i, bp->b_xio.xio_npages, m);
1887 }
1888 }
1889 }
1890 }
1891 #endif
1892
1893 /*
1894 * remove the mapping for kernel virtual
1895 */
1896 pmap_qremove((vm_offset_t)bp->b_data, bp->b_xio.xio_npages);
1897
1898 /*
1899 * cleanup pages. If an error occurs writing to swap, we are in
1900 * very serious trouble. If it happens to be a disk error, though,
1901 * we may be able to recover by reassigning the swap later on. So
1902 * in this case we remove the m->swapblk assignment for the page
1903 * but do not free it in the rlist. The errornous block(s) are thus
1904 * never reallocated as swap. Redirty the page and continue.
1905 */
1906 for (i = 0; i < bp->b_xio.xio_npages; ++i) {
1907 vm_page_t m = bp->b_xio.xio_pages[i];
1908
1909 if (bp->b_flags & B_ERROR) {
1910 /*
1911 * If an error occurs I'd love to throw the swapblk
1912 * away without freeing it back to swapspace, so it
1913 * can never be used again. But I can't from an
1914 * interrupt.
1915 */
1916
1917 if (bio->bio_caller_info1.index & SWBIO_READ) {
1918 /*
1919 * When reading, reqpage needs to stay
1920 * locked for the parent, but all other
1921 * pages can be freed. We still want to
1922 * wakeup the parent waiting on the page,
1923 * though. ( also: pg_reqpage can be -1 and
1924 * not match anything ).
1925 *
1926 * We have to wake specifically requested pages
1927 * up too because we cleared SWAPINPROG and
1928 * someone may be waiting for that.
1929 *
1930 * NOTE: For reads, m->dirty will probably
1931 * be overridden by the original caller
1932 * of getpages so don't play cute tricks
1933 * here.
1934 *
1935 * NOTE: We can't actually free the page from
1936 * here, because this is an interrupt.
1937 * It is not legal to mess with
1938 * object->memq from an interrupt.
1939 * Deactivate the page instead.
1940 *
1941 * WARNING! The instant SWAPINPROG is
1942 * cleared another cpu may start
1943 * using the mreq page (it will
1944 * check m->valid immediately).
1945 */
1946
1947 m->valid = 0;
1948 atomic_clear_int(&m->busy_count,
1949 PBUSY_SWAPINPROG);
1950
1951 /*
1952 * bio_driver_info holds the requested page
1953 * index.
1954 */
1955 if (i != (int)(intptr_t)bio->bio_driver_info) {
1956 vm_page_deactivate(m);
1957 vm_page_wakeup(m);
1958 } else {
1959 vm_page_flash(m);
1960 }
1961 /*
1962 * If i == bp->b_pager.pg_reqpage, do not wake
1963 * the page up. The caller needs to.
1964 */
1965 } else {
1966 /*
1967 * If a write error occurs remove the swap
1968 * assignment (note that PG_SWAPPED may or
1969 * may not be set depending on prior activity).
1970 *
1971 * Re-dirty OBJT_SWAP pages as there is no
1972 * other backing store, we can't throw the
1973 * page away.
1974 *
1975 * Non-OBJT_SWAP pages (aka swapcache) must
1976 * not be dirtied since they may not have
1977 * been dirty in the first place, and they
1978 * do have backing store (the vnode).
1979 */
1980 vm_page_busy_wait(m, FALSE, "swadpg");
1981 vm_object_hold(m->object);
1982 swp_pager_meta_ctl(m->object, m->pindex,
1983 SWM_FREE);
1984 vm_page_flag_clear(m, PG_SWAPPED);
1985 vm_object_drop(m->object);
1986 if (m->object->type == OBJT_SWAP) {
1987 vm_page_dirty(m);
1988 vm_page_activate(m);
1989 }
1990 vm_page_io_finish(m);
1991 atomic_clear_int(&m->busy_count,
1992 PBUSY_SWAPINPROG);
1993 vm_page_wakeup(m);
1994 }
1995 } else if (bio->bio_caller_info1.index & SWBIO_READ) {
1996 /*
1997 * NOTE: for reads, m->dirty will probably be
1998 * overridden by the original caller of getpages so
1999 * we cannot set them in order to free the underlying
2000 * swap in a low-swap situation. I don't think we'd
2001 * want to do that anyway, but it was an optimization
2002 * that existed in the old swapper for a time before
2003 * it got ripped out due to precisely this problem.
2004 *
2005 * If not the requested page then deactivate it.
2006 *
2007 * Note that the requested page, reqpage, is left
2008 * busied, but we still have to wake it up. The
2009 * other pages are released (unbusied) by
2010 * vm_page_wakeup(). We do not set reqpage's
2011 * valid bits here, it is up to the caller.
2012 */
2013
2014 /*
2015 * NOTE: Can't call pmap_clear_modify(m) from an
2016 * interrupt thread, the pmap code may have to
2017 * map non-kernel pmaps and currently asserts
2018 * the case.
2019 *
2020 * WARNING! The instant SWAPINPROG is
2021 * cleared another cpu may start
2022 * using the mreq page (it will
2023 * check m->valid immediately).
2024 */
2025 /*pmap_clear_modify(m);*/
2026 m->valid = VM_PAGE_BITS_ALL;
2027 vm_page_undirty(m);
2028 vm_page_flag_set(m, PG_SWAPPED);
2029 atomic_clear_int(&m->busy_count, PBUSY_SWAPINPROG);
2030
2031 /*
2032 * We have to wake specifically requested pages
2033 * up too because we cleared SWAPINPROG and
2034 * could be waiting for it in getpages. However,
2035 * be sure to not unbusy getpages specifically
2036 * requested page - getpages expects it to be
2037 * left busy.
2038 *
2039 * bio_driver_info holds the requested page
2040 */
2041 if (i != (int)(intptr_t)bio->bio_driver_info) {
2042 vm_page_deactivate(m);
2043 vm_page_wakeup(m);
2044 } else {
2045 vm_page_flash(m);
2046 }
2047 } else {
2048 /*
2049 * Mark the page clean but do not mess with the
2050 * pmap-layer's modified state. That state should
2051 * also be clear since the caller protected the
2052 * page VM_PROT_READ, but allow the case.
2053 *
2054 * We are in an interrupt, avoid pmap operations.
2055 *
2056 * If we have a severe page deficit, deactivate the
2057 * page. Do not try to cache it (which would also
2058 * involve a pmap op), because the page might still
2059 * be read-heavy.
2060 *
2061 * When using the swap to cache clean vnode pages
2062 * we do not mess with the page dirty bits.
2063 *
2064 * NOTE! Nobody is waiting for the key mreq page
2065 * on write completion.
2066 */
2067 vm_page_busy_wait(m, FALSE, "swadpg");
2068 if (m->object->type == OBJT_SWAP)
2069 vm_page_undirty(m);
2070 vm_page_flag_set(m, PG_SWAPPED);
2071 atomic_clear_int(&m->busy_count, PBUSY_SWAPINPROG);
2072 if (vm_paging_severe())
2073 vm_page_deactivate(m);
2074 vm_page_io_finish(m);
2075 if (bio->bio_caller_info1.index & SWBIO_TTC)
2076 vm_page_try_to_cache(m);
2077 else
2078 vm_page_wakeup(m);
2079 }
2080 }
2081
2082 /*
2083 * adjust pip. NOTE: the original parent may still have its own
2084 * pip refs on the object.
2085 */
2086
2087 if (object)
2088 vm_object_pip_wakeup_n(object, bp->b_xio.xio_npages);
2089
2090 /*
2091 * Release the physical I/O buffer.
2092 *
2093 * NOTE: Due to synchronous operations in the write case b_cmd may
2094 * already be set to BUF_CMD_DONE and BIO_SYNC may have already
2095 * been cleared.
2096 *
2097 * Use vm_token to interlock nsw_rcount/wcount wakeup?
2098 */
2099 lwkt_gettoken(&vm_token);
2100 if (bio->bio_caller_info1.index & SWBIO_READ)
2101 nswptr = &nsw_rcount;
2102 else if (bio->bio_caller_info1.index & SWBIO_SYNC)
2103 nswptr = &nsw_wcount_sync;
2104 else
2105 nswptr = &nsw_wcount_async;
2106 bp->b_cmd = BUF_CMD_DONE;
2107 relpbuf(bp, nswptr);
2108 lwkt_reltoken(&vm_token);
2109 }
2110
2111 /*
2112 * Fault-in a potentially swapped page and remove the swap reference.
2113 * (used by swapoff code)
2114 *
2115 * object must be held.
2116 */
2117 static __inline void
swp_pager_fault_page(vm_object_t object,int * sharedp,vm_pindex_t pindex)2118 swp_pager_fault_page(vm_object_t object, int *sharedp, vm_pindex_t pindex)
2119 {
2120 struct vnode *vp;
2121 vm_page_t m;
2122 int error;
2123
2124 ASSERT_LWKT_TOKEN_HELD(vm_object_token(object));
2125
2126 if (object->type == OBJT_VNODE) {
2127 /*
2128 * Any swap related to a vnode is due to swapcache. We must
2129 * vget() the vnode in case it is not active (otherwise
2130 * vref() will panic). Calling vm_object_page_remove() will
2131 * ensure that any swap ref is removed interlocked with the
2132 * page. clean_only is set to TRUE so we don't throw away
2133 * dirty pages.
2134 */
2135 vp = object->handle;
2136 error = vget(vp, LK_SHARED | LK_RETRY | LK_CANRECURSE);
2137 if (error == 0) {
2138 vm_object_page_remove(object, pindex, pindex + 1, TRUE);
2139 vput(vp);
2140 }
2141 } else {
2142 /*
2143 * Otherwise it is a normal OBJT_SWAP object and we can
2144 * fault the page in and remove the swap.
2145 */
2146 m = vm_fault_object_page(object, IDX_TO_OFF(pindex),
2147 VM_PROT_NONE,
2148 VM_FAULT_DIRTY | VM_FAULT_UNSWAP,
2149 sharedp, &error);
2150 if (m)
2151 vm_page_unhold(m);
2152 }
2153 }
2154
2155 /*
2156 * This removes all swap blocks related to a particular device. We have
2157 * to be careful of ripups during the scan.
2158 */
2159 static int swp_pager_swapoff_callback(struct swblock *swap, void *data);
2160
2161 int
swap_pager_swapoff(int devidx)2162 swap_pager_swapoff(int devidx)
2163 {
2164 struct vm_object_hash *hash;
2165 struct swswapoffinfo info;
2166 struct vm_object marker;
2167 vm_object_t object;
2168 int n;
2169
2170 bzero(&marker, sizeof(marker));
2171 marker.type = OBJT_MARKER;
2172
2173 for (n = 0; n < VMOBJ_HSIZE; ++n) {
2174 hash = &vm_object_hash[n];
2175
2176 lwkt_gettoken(&hash->token);
2177 TAILQ_INSERT_HEAD(&hash->list, &marker, object_entry);
2178
2179 while ((object = TAILQ_NEXT(&marker, object_entry)) != NULL) {
2180 if (object->type == OBJT_MARKER)
2181 goto skip;
2182 if (object->type != OBJT_SWAP &&
2183 object->type != OBJT_VNODE)
2184 goto skip;
2185 vm_object_hold(object);
2186 if (object->type != OBJT_SWAP &&
2187 object->type != OBJT_VNODE) {
2188 vm_object_drop(object);
2189 goto skip;
2190 }
2191
2192 /*
2193 * Object is special in that we can't just pagein
2194 * into vm_page's in it (tmpfs, vn).
2195 */
2196 if ((object->flags & OBJ_NOPAGEIN) &&
2197 RB_ROOT(&object->swblock_root)) {
2198 vm_object_drop(object);
2199 goto skip;
2200 }
2201
2202 info.object = object;
2203 info.shared = 0;
2204 info.devidx = devidx;
2205 swblock_rb_tree_RB_SCAN(&object->swblock_root,
2206 NULL, swp_pager_swapoff_callback,
2207 &info);
2208 vm_object_drop(object);
2209 skip:
2210 if (object == TAILQ_NEXT(&marker, object_entry)) {
2211 TAILQ_REMOVE(&hash->list, &marker,
2212 object_entry);
2213 TAILQ_INSERT_AFTER(&hash->list, object,
2214 &marker, object_entry);
2215 }
2216 }
2217 TAILQ_REMOVE(&hash->list, &marker, object_entry);
2218 lwkt_reltoken(&hash->token);
2219 }
2220
2221 /*
2222 * If we fail to locate all swblocks we just fail gracefully and
2223 * do not bother to restore paging on the swap device. If the
2224 * user wants to retry the user can retry.
2225 */
2226 if (swdevt[devidx].sw_nused)
2227 return (1);
2228 else
2229 return (0);
2230 }
2231
2232 static
2233 int
swp_pager_swapoff_callback(struct swblock * swap,void * data)2234 swp_pager_swapoff_callback(struct swblock *swap, void *data)
2235 {
2236 struct swswapoffinfo *info = data;
2237 vm_object_t object = info->object;
2238 vm_pindex_t index;
2239 swblk_t v;
2240 int i;
2241
2242 index = swap->swb_index;
2243 for (i = 0; i < SWAP_META_PAGES; ++i) {
2244 /*
2245 * Make sure we don't race a dying object. This will
2246 * kill the scan of the object's swap blocks entirely.
2247 */
2248 if (object->flags & OBJ_DEAD)
2249 return(-1);
2250
2251 /*
2252 * Fault the page, which can obviously block. If the swap
2253 * structure disappears break out.
2254 */
2255 v = swap->swb_pages[i];
2256 if (v != SWAPBLK_NONE && BLK2DEVIDX(v) == info->devidx) {
2257 swp_pager_fault_page(object, &info->shared,
2258 swap->swb_index + i);
2259 /* swap ptr might go away */
2260 if (RB_LOOKUP(swblock_rb_tree,
2261 &object->swblock_root, index) != swap) {
2262 break;
2263 }
2264 }
2265 }
2266 return(0);
2267 }
2268
2269 /************************************************************************
2270 * SWAP META DATA *
2271 ************************************************************************
2272 *
2273 * These routines manipulate the swap metadata stored in the
2274 * OBJT_SWAP object.
2275 *
2276 * Swap metadata is implemented with a global hash and not directly
2277 * linked into the object. Instead the object simply contains
2278 * appropriate tracking counters.
2279 */
2280
2281 /*
2282 * Lookup the swblock containing the specified swap block index.
2283 *
2284 * The caller must hold the object.
2285 */
2286 static __inline
2287 struct swblock *
swp_pager_lookup(vm_object_t object,vm_pindex_t index)2288 swp_pager_lookup(vm_object_t object, vm_pindex_t index)
2289 {
2290 ASSERT_LWKT_TOKEN_HELD(vm_object_token(object));
2291 index &= ~(vm_pindex_t)SWAP_META_MASK;
2292 return (RB_LOOKUP(swblock_rb_tree, &object->swblock_root, index));
2293 }
2294
2295 /*
2296 * Remove a swblock from the RB tree.
2297 *
2298 * The caller must hold the object.
2299 */
2300 static __inline
2301 void
swp_pager_remove(vm_object_t object,struct swblock * swap)2302 swp_pager_remove(vm_object_t object, struct swblock *swap)
2303 {
2304 ASSERT_LWKT_TOKEN_HELD(vm_object_token(object));
2305 RB_REMOVE(swblock_rb_tree, &object->swblock_root, swap);
2306 }
2307
2308 /*
2309 * Convert default object to swap object if necessary
2310 *
2311 * The caller must hold the object.
2312 */
2313 static void
swp_pager_meta_convert(vm_object_t object)2314 swp_pager_meta_convert(vm_object_t object)
2315 {
2316 if (object->type == OBJT_DEFAULT) {
2317 object->type = OBJT_SWAP;
2318 KKASSERT(object->swblock_count == 0);
2319 }
2320 }
2321
2322 /*
2323 * SWP_PAGER_META_BUILD() - add swap block to swap meta data for object
2324 *
2325 * We first convert the object to a swap object if it is a default
2326 * object. Vnode objects do not need to be converted.
2327 *
2328 * The specified swapblk is added to the object's swap metadata. If
2329 * the swapblk is not valid, it is freed instead. Any previously
2330 * assigned swapblk is freed.
2331 *
2332 * The caller must hold the object.
2333 */
2334 static void
swp_pager_meta_build(vm_object_t object,vm_pindex_t index,swblk_t swapblk)2335 swp_pager_meta_build(vm_object_t object, vm_pindex_t index, swblk_t swapblk)
2336 {
2337 struct swblock *swap;
2338 struct swblock *oswap;
2339 vm_pindex_t v;
2340
2341 KKASSERT(swapblk != SWAPBLK_NONE);
2342 ASSERT_LWKT_TOKEN_HELD(vm_object_token(object));
2343
2344 /*
2345 * Convert object if necessary
2346 */
2347 if (object->type == OBJT_DEFAULT)
2348 swp_pager_meta_convert(object);
2349
2350 /*
2351 * Locate swblock. If not found create, but if we aren't adding
2352 * anything just return. If we run out of space in the map we wait
2353 * and, since the hash table may have changed, retry.
2354 */
2355 retry:
2356 swap = swp_pager_lookup(object, index);
2357
2358 if (swap == NULL) {
2359 int i;
2360
2361 swap = zalloc(swap_zone);
2362 if (swap == NULL) {
2363 vm_wait(0);
2364 goto retry;
2365 }
2366 swap->swb_index = index & ~(vm_pindex_t)SWAP_META_MASK;
2367 swap->swb_count = 0;
2368
2369 ++object->swblock_count;
2370
2371 for (i = 0; i < SWAP_META_PAGES; ++i)
2372 swap->swb_pages[i] = SWAPBLK_NONE;
2373 oswap = RB_INSERT(swblock_rb_tree, &object->swblock_root, swap);
2374 KKASSERT(oswap == NULL);
2375 }
2376
2377 /*
2378 * Delete prior contents of metadata.
2379 *
2380 * NOTE: Decrement swb_count after the freeing operation (which
2381 * might block) to prevent racing destruction of the swblock.
2382 */
2383 index &= SWAP_META_MASK;
2384
2385 while ((v = swap->swb_pages[index]) != SWAPBLK_NONE) {
2386 swap->swb_pages[index] = SWAPBLK_NONE;
2387 /* can block */
2388 swp_pager_freeswapspace(object, v, 1);
2389 --swap->swb_count;
2390 --mycpu->gd_vmtotal.t_vm;
2391 }
2392
2393 /*
2394 * Enter block into metadata
2395 */
2396 swap->swb_pages[index] = swapblk;
2397 if (swapblk != SWAPBLK_NONE) {
2398 ++swap->swb_count;
2399 ++mycpu->gd_vmtotal.t_vm;
2400 }
2401 }
2402
2403 /*
2404 * SWP_PAGER_META_FREE() - free a range of blocks in the object's swap metadata
2405 *
2406 * The requested range of blocks is freed, with any associated swap
2407 * returned to the swap bitmap.
2408 *
2409 * This routine will free swap metadata structures as they are cleaned
2410 * out. This routine does *NOT* operate on swap metadata associated
2411 * with resident pages.
2412 *
2413 * The caller must hold the object.
2414 */
2415 static int swp_pager_meta_free_callback(struct swblock *swb, void *data);
2416
2417 static void
swp_pager_meta_free(vm_object_t object,vm_pindex_t index,vm_pindex_t count)2418 swp_pager_meta_free(vm_object_t object, vm_pindex_t index, vm_pindex_t count)
2419 {
2420 struct swfreeinfo info;
2421
2422 ASSERT_LWKT_TOKEN_HELD(vm_object_token(object));
2423
2424 /*
2425 * Nothing to do
2426 */
2427 if (object->swblock_count == 0) {
2428 KKASSERT(RB_EMPTY(&object->swblock_root));
2429 return;
2430 }
2431 if (count == 0)
2432 return;
2433
2434 /*
2435 * Setup for RB tree scan. Note that the pindex range can be huge
2436 * due to the 64 bit page index space so we cannot safely iterate.
2437 */
2438 info.object = object;
2439 info.basei = index & ~(vm_pindex_t)SWAP_META_MASK;
2440 info.begi = index;
2441 info.endi = index + count - 1;
2442 swblock_rb_tree_RB_SCAN(&object->swblock_root, rb_swblock_scancmp,
2443 swp_pager_meta_free_callback, &info);
2444 }
2445
2446 /*
2447 * The caller must hold the object.
2448 */
2449 static
2450 int
swp_pager_meta_free_callback(struct swblock * swap,void * data)2451 swp_pager_meta_free_callback(struct swblock *swap, void *data)
2452 {
2453 struct swfreeinfo *info = data;
2454 vm_object_t object = info->object;
2455 int index;
2456 int eindex;
2457
2458 /*
2459 * Figure out the range within the swblock. The wider scan may
2460 * return edge-case swap blocks when the start and/or end points
2461 * are in the middle of a block.
2462 */
2463 if (swap->swb_index < info->begi)
2464 index = (int)info->begi & SWAP_META_MASK;
2465 else
2466 index = 0;
2467
2468 if (swap->swb_index + SWAP_META_PAGES > info->endi)
2469 eindex = (int)info->endi & SWAP_META_MASK;
2470 else
2471 eindex = SWAP_META_MASK;
2472
2473 /*
2474 * Scan and free the blocks. The loop terminates early
2475 * if (swap) runs out of blocks and could be freed.
2476 *
2477 * NOTE: Decrement swb_count after swp_pager_freeswapspace()
2478 * to deal with a zfree race.
2479 */
2480 while (index <= eindex) {
2481 swblk_t v = swap->swb_pages[index];
2482
2483 if (v != SWAPBLK_NONE) {
2484 swap->swb_pages[index] = SWAPBLK_NONE;
2485 /* can block */
2486 swp_pager_freeswapspace(object, v, 1);
2487 --mycpu->gd_vmtotal.t_vm;
2488 if (--swap->swb_count == 0) {
2489 swp_pager_remove(object, swap);
2490 zfree(swap_zone, swap);
2491 --object->swblock_count;
2492 break;
2493 }
2494 }
2495 ++index;
2496 }
2497
2498 /* swap may be invalid here due to zfree above */
2499 lwkt_yield();
2500
2501 return(0);
2502 }
2503
2504 /*
2505 * SWP_PAGER_META_FREE_ALL() - destroy all swap metadata associated with object
2506 *
2507 * This routine locates and destroys all swap metadata associated with
2508 * an object.
2509 *
2510 * NOTE: Decrement swb_count after the freeing operation (which
2511 * might block) to prevent racing destruction of the swblock.
2512 *
2513 * The caller must hold the object.
2514 */
2515 static void
swp_pager_meta_free_all(vm_object_t object)2516 swp_pager_meta_free_all(vm_object_t object)
2517 {
2518 struct swblock *swap;
2519 int i;
2520
2521 ASSERT_LWKT_TOKEN_HELD(vm_object_token(object));
2522
2523 while ((swap = RB_ROOT(&object->swblock_root)) != NULL) {
2524 swp_pager_remove(object, swap);
2525 for (i = 0; i < SWAP_META_PAGES; ++i) {
2526 swblk_t v = swap->swb_pages[i];
2527 if (v != SWAPBLK_NONE) {
2528 /* can block */
2529 swp_pager_freeswapspace(object, v, 1);
2530 --swap->swb_count;
2531 --mycpu->gd_vmtotal.t_vm;
2532 }
2533 }
2534 if (swap->swb_count != 0)
2535 panic("swap_pager_meta_free_all: swb_count != 0");
2536 zfree(swap_zone, swap);
2537 --object->swblock_count;
2538 lwkt_yield();
2539 }
2540 KKASSERT(object->swblock_count == 0);
2541 }
2542
2543 /*
2544 * SWP_PAGER_METACTL() - misc control of swap and vm_page_t meta data.
2545 *
2546 * This routine is capable of looking up, popping, or freeing
2547 * swapblk assignments in the swap meta data or in the vm_page_t.
2548 * The routine typically returns the swapblk being looked-up, or popped,
2549 * or SWAPBLK_NONE if the block was freed, or SWAPBLK_NONE if the block
2550 * was invalid. This routine will automatically free any invalid
2551 * meta-data swapblks.
2552 *
2553 * It is not possible to store invalid swapblks in the swap meta data
2554 * (other then a literal 'SWAPBLK_NONE'), so we don't bother checking.
2555 *
2556 * When acting on a busy resident page and paging is in progress, we
2557 * have to wait until paging is complete but otherwise can act on the
2558 * busy page.
2559 *
2560 * SWM_FREE remove and free swap block from metadata
2561 * SWM_POP remove from meta data but do not free.. pop it out
2562 *
2563 * The caller must hold the object.
2564 */
2565 static swblk_t
swp_pager_meta_ctl(vm_object_t object,vm_pindex_t index,int flags)2566 swp_pager_meta_ctl(vm_object_t object, vm_pindex_t index, int flags)
2567 {
2568 struct swblock *swap;
2569 swblk_t r1;
2570
2571 if (object->swblock_count == 0)
2572 return(SWAPBLK_NONE);
2573
2574 r1 = SWAPBLK_NONE;
2575 swap = swp_pager_lookup(object, index);
2576
2577 if (swap != NULL) {
2578 index &= SWAP_META_MASK;
2579 r1 = swap->swb_pages[index];
2580
2581 if (r1 != SWAPBLK_NONE) {
2582 if (flags & (SWM_FREE|SWM_POP)) {
2583 swap->swb_pages[index] = SWAPBLK_NONE;
2584 --mycpu->gd_vmtotal.t_vm;
2585 if (--swap->swb_count == 0) {
2586 swp_pager_remove(object, swap);
2587 zfree(swap_zone, swap);
2588 --object->swblock_count;
2589 }
2590 }
2591 /* swap ptr may be invalid */
2592 if (flags & SWM_FREE) {
2593 swp_pager_freeswapspace(object, r1, 1);
2594 r1 = SWAPBLK_NONE;
2595 }
2596 }
2597 /* swap ptr may be invalid */
2598 }
2599 return(r1);
2600 }
2601