1 /*- 2 * SPDX-License-Identifier: BSD-4-Clause 3 * 4 * Copyright (c) 1998 Matthew Dillon, 5 * Copyright (c) 1994 John S. Dyson 6 * Copyright (c) 1990 University of Utah. 7 * Copyright (c) 1982, 1986, 1989, 1993 8 * The Regents of the University of California. All rights reserved. 9 * 10 * This code is derived from software contributed to Berkeley by 11 * the Systems Programming Group of the University of Utah Computer 12 * Science Department. 13 * 14 * Redistribution and use in source and binary forms, with or without 15 * modification, are permitted provided that the following conditions 16 * are met: 17 * 1. Redistributions of source code must retain the above copyright 18 * notice, this list of conditions and the following disclaimer. 19 * 2. Redistributions in binary form must reproduce the above copyright 20 * notice, this list of conditions and the following disclaimer in the 21 * documentation and/or other materials provided with the distribution. 22 * 3. All advertising materials mentioning features or use of this software 23 * must display the following acknowledgement: 24 * This product includes software developed by the University of 25 * California, Berkeley and its contributors. 26 * 4. Neither the name of the University nor the names of its contributors 27 * may be used to endorse or promote products derived from this software 28 * without specific prior written permission. 29 * 30 * THIS SOFTWARE IS PROVIDED BY THE REGENTS AND CONTRIBUTORS ``AS IS'' AND 31 * ANY EXPRESS OR IMPLIED WARRANTIES, INCLUDING, BUT NOT LIMITED TO, THE 32 * IMPLIED WARRANTIES OF MERCHANTABILITY AND FITNESS FOR A PARTICULAR PURPOSE 33 * ARE DISCLAIMED. IN NO EVENT SHALL THE REGENTS OR CONTRIBUTORS BE LIABLE 34 * FOR ANY DIRECT, INDIRECT, INCIDENTAL, SPECIAL, EXEMPLARY, OR CONSEQUENTIAL 35 * DAMAGES (INCLUDING, BUT NOT LIMITED TO, PROCUREMENT OF SUBSTITUTE GOODS 36 * OR SERVICES; LOSS OF USE, DATA, OR PROFITS; OR BUSINESS INTERRUPTION) 37 * HOWEVER CAUSED AND ON ANY THEORY OF LIABILITY, WHETHER IN CONTRACT, STRICT 38 * LIABILITY, OR TORT (INCLUDING NEGLIGENCE OR OTHERWISE) ARISING IN ANY WAY 39 * OUT OF THE USE OF THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF 40 * SUCH DAMAGE. 41 * 42 * New Swap System 43 * Matthew Dillon 44 * 45 * Radix Bitmap 'blists'. 46 * 47 * - The new swapper uses the new radix bitmap code. This should scale 48 * to arbitrarily small or arbitrarily large swap spaces and an almost 49 * arbitrary degree of fragmentation. 50 * 51 * Features: 52 * 53 * - on the fly reallocation of swap during putpages. The new system 54 * does not try to keep previously allocated swap blocks for dirty 55 * pages. 56 * 57 * - on the fly deallocation of swap 58 * 59 * - No more garbage collection required. Unnecessarily allocated swap 60 * blocks only exist for dirty vm_page_t's now and these are already 61 * cycled (in a high-load system) by the pager. We also do on-the-fly 62 * removal of invalidated swap blocks when a page is destroyed 63 * or renamed. 64 * 65 * from: Utah $Hdr: swap_pager.c 1.4 91/04/30$ 66 * 67 * @(#)swap_pager.c 8.9 (Berkeley) 3/21/94 68 * @(#)vm_swap.c 8.5 (Berkeley) 2/17/94 69 */ 70 71 #include <sys/cdefs.h> 72 __FBSDID("$FreeBSD$"); 73 74 #include "opt_vm.h" 75 76 #include <sys/param.h> 77 #include <sys/bio.h> 78 #include <sys/blist.h> 79 #include <sys/buf.h> 80 #include <sys/conf.h> 81 #include <sys/disk.h> 82 #include <sys/disklabel.h> 83 #include <sys/eventhandler.h> 84 #include <sys/fcntl.h> 85 #include <sys/lock.h> 86 #include <sys/kernel.h> 87 #include <sys/mount.h> 88 #include <sys/namei.h> 89 #include <sys/malloc.h> 90 #include <sys/pctrie.h> 91 #include <sys/priv.h> 92 #include <sys/proc.h> 93 #include <sys/racct.h> 94 #include <sys/resource.h> 95 #include <sys/resourcevar.h> 96 #include <sys/rwlock.h> 97 #include <sys/sbuf.h> 98 #include <sys/sysctl.h> 99 #include <sys/sysproto.h> 100 #include <sys/systm.h> 101 #include <sys/sx.h> 102 #include <sys/vmmeter.h> 103 #include <sys/vnode.h> 104 105 #include <security/mac/mac_framework.h> 106 107 #include <vm/vm.h> 108 #include <vm/pmap.h> 109 #include <vm/vm_map.h> 110 #include <vm/vm_kern.h> 111 #include <vm/vm_object.h> 112 #include <vm/vm_page.h> 113 #include <vm/vm_pager.h> 114 #include <vm/vm_pageout.h> 115 #include <vm/vm_param.h> 116 #include <vm/swap_pager.h> 117 #include <vm/vm_extern.h> 118 #include <vm/uma.h> 119 120 #include <geom/geom.h> 121 122 /* 123 * MAX_PAGEOUT_CLUSTER must be a power of 2 between 1 and 64. 124 * The 64-page limit is due to the radix code (kern/subr_blist.c). 125 */ 126 #ifndef MAX_PAGEOUT_CLUSTER 127 #define MAX_PAGEOUT_CLUSTER 32 128 #endif 129 130 #if !defined(SWB_NPAGES) 131 #define SWB_NPAGES MAX_PAGEOUT_CLUSTER 132 #endif 133 134 #define SWAP_META_PAGES PCTRIE_COUNT 135 136 /* 137 * A swblk structure maps each page index within a 138 * SWAP_META_PAGES-aligned and sized range to the address of an 139 * on-disk swap block (or SWAPBLK_NONE). The collection of these 140 * mappings for an entire vm object is implemented as a pc-trie. 141 */ 142 struct swblk { 143 vm_pindex_t p; 144 daddr_t d[SWAP_META_PAGES]; 145 }; 146 147 static MALLOC_DEFINE(M_VMPGDATA, "vm_pgdata", "swap pager private data"); 148 static struct mtx sw_dev_mtx; 149 static TAILQ_HEAD(, swdevt) swtailq = TAILQ_HEAD_INITIALIZER(swtailq); 150 static struct swdevt *swdevhd; /* Allocate from here next */ 151 static int nswapdev; /* Number of swap devices */ 152 int swap_pager_avail; 153 static struct sx swdev_syscall_lock; /* serialize swap(on|off) */ 154 155 static u_long swap_reserved; 156 static u_long swap_total; 157 static int sysctl_page_shift(SYSCTL_HANDLER_ARGS); 158 159 static SYSCTL_NODE(_vm_stats, OID_AUTO, swap, CTLFLAG_RD, 0, "VM swap stats"); 160 161 SYSCTL_PROC(_vm, OID_AUTO, swap_reserved, CTLTYPE_U64 | CTLFLAG_RD | CTLFLAG_MPSAFE, 162 &swap_reserved, 0, sysctl_page_shift, "A", 163 "Amount of swap storage needed to back all allocated anonymous memory."); 164 SYSCTL_PROC(_vm, OID_AUTO, swap_total, CTLTYPE_U64 | CTLFLAG_RD | CTLFLAG_MPSAFE, 165 &swap_total, 0, sysctl_page_shift, "A", 166 "Total amount of available swap storage."); 167 168 static int overcommit = 0; 169 SYSCTL_INT(_vm, VM_OVERCOMMIT, overcommit, CTLFLAG_RW, &overcommit, 0, 170 "Configure virtual memory overcommit behavior. See tuning(7) " 171 "for details."); 172 static unsigned long swzone; 173 SYSCTL_ULONG(_vm, OID_AUTO, swzone, CTLFLAG_RD, &swzone, 0, 174 "Actual size of swap metadata zone"); 175 static unsigned long swap_maxpages; 176 SYSCTL_ULONG(_vm, OID_AUTO, swap_maxpages, CTLFLAG_RD, &swap_maxpages, 0, 177 "Maximum amount of swap supported"); 178 179 static counter_u64_t swap_free_deferred; 180 SYSCTL_COUNTER_U64(_vm_stats_swap, OID_AUTO, free_deferred, 181 CTLFLAG_RD, &swap_free_deferred, 182 "Number of pages that deferred freeing swap space"); 183 184 static counter_u64_t swap_free_completed; 185 SYSCTL_COUNTER_U64(_vm_stats_swap, OID_AUTO, free_completed, 186 CTLFLAG_RD, &swap_free_completed, 187 "Number of deferred frees completed"); 188 189 /* bits from overcommit */ 190 #define SWAP_RESERVE_FORCE_ON (1 << 0) 191 #define SWAP_RESERVE_RLIMIT_ON (1 << 1) 192 #define SWAP_RESERVE_ALLOW_NONWIRED (1 << 2) 193 194 static int 195 sysctl_page_shift(SYSCTL_HANDLER_ARGS) 196 { 197 uint64_t newval; 198 u_long value = *(u_long *)arg1; 199 200 newval = ((uint64_t)value) << PAGE_SHIFT; 201 return (sysctl_handle_64(oidp, &newval, 0, req)); 202 } 203 204 int 205 swap_reserve(vm_ooffset_t incr) 206 { 207 208 return (swap_reserve_by_cred(incr, curthread->td_ucred)); 209 } 210 211 int 212 swap_reserve_by_cred(vm_ooffset_t incr, struct ucred *cred) 213 { 214 u_long r, s, prev, pincr; 215 int res, error; 216 static int curfail; 217 static struct timeval lastfail; 218 struct uidinfo *uip; 219 220 uip = cred->cr_ruidinfo; 221 222 KASSERT((incr & PAGE_MASK) == 0, ("%s: incr: %ju & PAGE_MASK", __func__, 223 (uintmax_t)incr)); 224 225 #ifdef RACCT 226 if (racct_enable) { 227 PROC_LOCK(curproc); 228 error = racct_add(curproc, RACCT_SWAP, incr); 229 PROC_UNLOCK(curproc); 230 if (error != 0) 231 return (0); 232 } 233 #endif 234 235 pincr = atop(incr); 236 res = 0; 237 prev = atomic_fetchadd_long(&swap_reserved, pincr); 238 r = prev + pincr; 239 if (overcommit & SWAP_RESERVE_ALLOW_NONWIRED) { 240 s = vm_cnt.v_page_count - vm_cnt.v_free_reserved - 241 vm_wire_count(); 242 } else 243 s = 0; 244 s += swap_total; 245 if ((overcommit & SWAP_RESERVE_FORCE_ON) == 0 || r <= s || 246 (error = priv_check(curthread, PRIV_VM_SWAP_NOQUOTA)) == 0) { 247 res = 1; 248 } else { 249 prev = atomic_fetchadd_long(&swap_reserved, -pincr); 250 if (prev < pincr) 251 panic("swap_reserved < incr on overcommit fail"); 252 } 253 if (res) { 254 prev = atomic_fetchadd_long(&uip->ui_vmsize, pincr); 255 if ((overcommit & SWAP_RESERVE_RLIMIT_ON) != 0 && 256 prev + pincr > lim_cur(curthread, RLIMIT_SWAP) && 257 priv_check(curthread, PRIV_VM_SWAP_NORLIMIT)) { 258 res = 0; 259 prev = atomic_fetchadd_long(&uip->ui_vmsize, -pincr); 260 if (prev < pincr) 261 panic("uip->ui_vmsize < incr on overcommit fail"); 262 } 263 } 264 if (!res && ppsratecheck(&lastfail, &curfail, 1)) { 265 printf("uid %d, pid %d: swap reservation for %jd bytes failed\n", 266 uip->ui_uid, curproc->p_pid, incr); 267 } 268 269 #ifdef RACCT 270 if (racct_enable && !res) { 271 PROC_LOCK(curproc); 272 racct_sub(curproc, RACCT_SWAP, incr); 273 PROC_UNLOCK(curproc); 274 } 275 #endif 276 277 return (res); 278 } 279 280 void 281 swap_reserve_force(vm_ooffset_t incr) 282 { 283 struct uidinfo *uip; 284 u_long pincr; 285 286 KASSERT((incr & PAGE_MASK) == 0, ("%s: incr: %ju & PAGE_MASK", __func__, 287 (uintmax_t)incr)); 288 289 PROC_LOCK(curproc); 290 #ifdef RACCT 291 if (racct_enable) 292 racct_add_force(curproc, RACCT_SWAP, incr); 293 #endif 294 pincr = atop(incr); 295 atomic_add_long(&swap_reserved, pincr); 296 uip = curproc->p_ucred->cr_ruidinfo; 297 atomic_add_long(&uip->ui_vmsize, pincr); 298 PROC_UNLOCK(curproc); 299 } 300 301 void 302 swap_release(vm_ooffset_t decr) 303 { 304 struct ucred *cred; 305 306 PROC_LOCK(curproc); 307 cred = curproc->p_ucred; 308 swap_release_by_cred(decr, cred); 309 PROC_UNLOCK(curproc); 310 } 311 312 void 313 swap_release_by_cred(vm_ooffset_t decr, struct ucred *cred) 314 { 315 u_long prev, pdecr; 316 struct uidinfo *uip; 317 318 uip = cred->cr_ruidinfo; 319 320 KASSERT((decr & PAGE_MASK) == 0, ("%s: decr: %ju & PAGE_MASK", __func__, 321 (uintmax_t)decr)); 322 323 pdecr = atop(decr); 324 prev = atomic_fetchadd_long(&swap_reserved, -pdecr); 325 if (prev < pdecr) 326 panic("swap_reserved < decr"); 327 328 prev = atomic_fetchadd_long(&uip->ui_vmsize, -pdecr); 329 if (prev < pdecr) 330 printf("negative vmsize for uid = %d\n", uip->ui_uid); 331 #ifdef RACCT 332 if (racct_enable) 333 racct_sub_cred(cred, RACCT_SWAP, decr); 334 #endif 335 } 336 337 static int swap_pager_full = 2; /* swap space exhaustion (task killing) */ 338 static int swap_pager_almost_full = 1; /* swap space exhaustion (w/hysteresis)*/ 339 static struct mtx swbuf_mtx; /* to sync nsw_wcount_async */ 340 static int nsw_wcount_async; /* limit async write buffers */ 341 static int nsw_wcount_async_max;/* assigned maximum */ 342 static int nsw_cluster_max; /* maximum VOP I/O allowed */ 343 344 static int sysctl_swap_async_max(SYSCTL_HANDLER_ARGS); 345 SYSCTL_PROC(_vm, OID_AUTO, swap_async_max, CTLTYPE_INT | CTLFLAG_RW | 346 CTLFLAG_MPSAFE, NULL, 0, sysctl_swap_async_max, "I", 347 "Maximum running async swap ops"); 348 static int sysctl_swap_fragmentation(SYSCTL_HANDLER_ARGS); 349 SYSCTL_PROC(_vm, OID_AUTO, swap_fragmentation, CTLTYPE_STRING | CTLFLAG_RD | 350 CTLFLAG_MPSAFE, NULL, 0, sysctl_swap_fragmentation, "A", 351 "Swap Fragmentation Info"); 352 353 static struct sx sw_alloc_sx; 354 355 /* 356 * "named" and "unnamed" anon region objects. Try to reduce the overhead 357 * of searching a named list by hashing it just a little. 358 */ 359 360 #define NOBJLISTS 8 361 362 #define NOBJLIST(handle) \ 363 (&swap_pager_object_list[((int)(intptr_t)handle >> 4) & (NOBJLISTS-1)]) 364 365 static struct pagerlst swap_pager_object_list[NOBJLISTS]; 366 static uma_zone_t swwbuf_zone; 367 static uma_zone_t swrbuf_zone; 368 static uma_zone_t swblk_zone; 369 static uma_zone_t swpctrie_zone; 370 371 /* 372 * pagerops for OBJT_SWAP - "swap pager". Some ops are also global procedure 373 * calls hooked from other parts of the VM system and do not appear here. 374 * (see vm/swap_pager.h). 375 */ 376 static vm_object_t 377 swap_pager_alloc(void *handle, vm_ooffset_t size, 378 vm_prot_t prot, vm_ooffset_t offset, struct ucred *); 379 static void swap_pager_dealloc(vm_object_t object); 380 static int swap_pager_getpages(vm_object_t, vm_page_t *, int, int *, 381 int *); 382 static int swap_pager_getpages_async(vm_object_t, vm_page_t *, int, int *, 383 int *, pgo_getpages_iodone_t, void *); 384 static void swap_pager_putpages(vm_object_t, vm_page_t *, int, boolean_t, int *); 385 static boolean_t 386 swap_pager_haspage(vm_object_t object, vm_pindex_t pindex, int *before, int *after); 387 static void swap_pager_init(void); 388 static void swap_pager_unswapped(vm_page_t); 389 static void swap_pager_swapoff(struct swdevt *sp); 390 static void swap_pager_update_writecount(vm_object_t object, 391 vm_offset_t start, vm_offset_t end); 392 static void swap_pager_release_writecount(vm_object_t object, 393 vm_offset_t start, vm_offset_t end); 394 395 struct pagerops swappagerops = { 396 .pgo_init = swap_pager_init, /* early system initialization of pager */ 397 .pgo_alloc = swap_pager_alloc, /* allocate an OBJT_SWAP object */ 398 .pgo_dealloc = swap_pager_dealloc, /* deallocate an OBJT_SWAP object */ 399 .pgo_getpages = swap_pager_getpages, /* pagein */ 400 .pgo_getpages_async = swap_pager_getpages_async, /* pagein (async) */ 401 .pgo_putpages = swap_pager_putpages, /* pageout */ 402 .pgo_haspage = swap_pager_haspage, /* get backing store status for page */ 403 .pgo_pageunswapped = swap_pager_unswapped, /* remove swap related to page */ 404 .pgo_update_writecount = swap_pager_update_writecount, 405 .pgo_release_writecount = swap_pager_release_writecount, 406 }; 407 408 /* 409 * swap_*() routines are externally accessible. swp_*() routines are 410 * internal. 411 */ 412 static int nswap_lowat = 128; /* in pages, swap_pager_almost_full warn */ 413 static int nswap_hiwat = 512; /* in pages, swap_pager_almost_full warn */ 414 415 SYSCTL_INT(_vm, OID_AUTO, dmmax, CTLFLAG_RD, &nsw_cluster_max, 0, 416 "Maximum size of a swap block in pages"); 417 418 static void swp_sizecheck(void); 419 static void swp_pager_async_iodone(struct buf *bp); 420 static bool swp_pager_swblk_empty(struct swblk *sb, int start, int limit); 421 static void swp_pager_free_empty_swblk(vm_object_t, struct swblk *sb); 422 static int swapongeom(struct vnode *); 423 static int swaponvp(struct thread *, struct vnode *, u_long); 424 static int swapoff_one(struct swdevt *sp, struct ucred *cred); 425 426 /* 427 * Swap bitmap functions 428 */ 429 static void swp_pager_freeswapspace(daddr_t blk, daddr_t npages); 430 static daddr_t swp_pager_getswapspace(int *npages, int limit); 431 432 /* 433 * Metadata functions 434 */ 435 static daddr_t swp_pager_meta_build(vm_object_t, vm_pindex_t, daddr_t); 436 static void swp_pager_meta_free(vm_object_t, vm_pindex_t, vm_pindex_t); 437 static void swp_pager_meta_transfer(vm_object_t src, vm_object_t dst, 438 vm_pindex_t pindex, vm_pindex_t count); 439 static void swp_pager_meta_free_all(vm_object_t); 440 static daddr_t swp_pager_meta_lookup(vm_object_t, vm_pindex_t); 441 442 static void 443 swp_pager_init_freerange(daddr_t *start, daddr_t *num) 444 { 445 446 *start = SWAPBLK_NONE; 447 *num = 0; 448 } 449 450 static void 451 swp_pager_update_freerange(daddr_t *start, daddr_t *num, daddr_t addr) 452 { 453 454 if (*start + *num == addr) { 455 (*num)++; 456 } else { 457 swp_pager_freeswapspace(*start, *num); 458 *start = addr; 459 *num = 1; 460 } 461 } 462 463 static void * 464 swblk_trie_alloc(struct pctrie *ptree) 465 { 466 467 return (uma_zalloc(swpctrie_zone, M_NOWAIT | (curproc == pageproc ? 468 M_USE_RESERVE : 0))); 469 } 470 471 static void 472 swblk_trie_free(struct pctrie *ptree, void *node) 473 { 474 475 uma_zfree(swpctrie_zone, node); 476 } 477 478 PCTRIE_DEFINE(SWAP, swblk, p, swblk_trie_alloc, swblk_trie_free); 479 480 /* 481 * SWP_SIZECHECK() - update swap_pager_full indication 482 * 483 * update the swap_pager_almost_full indication and warn when we are 484 * about to run out of swap space, using lowat/hiwat hysteresis. 485 * 486 * Clear swap_pager_full ( task killing ) indication when lowat is met. 487 * 488 * No restrictions on call 489 * This routine may not block. 490 */ 491 static void 492 swp_sizecheck(void) 493 { 494 495 if (swap_pager_avail < nswap_lowat) { 496 if (swap_pager_almost_full == 0) { 497 printf("swap_pager: out of swap space\n"); 498 swap_pager_almost_full = 1; 499 } 500 } else { 501 swap_pager_full = 0; 502 if (swap_pager_avail > nswap_hiwat) 503 swap_pager_almost_full = 0; 504 } 505 } 506 507 /* 508 * SWAP_PAGER_INIT() - initialize the swap pager! 509 * 510 * Expected to be started from system init. NOTE: This code is run 511 * before much else so be careful what you depend on. Most of the VM 512 * system has yet to be initialized at this point. 513 */ 514 static void 515 swap_pager_init(void) 516 { 517 /* 518 * Initialize object lists 519 */ 520 int i; 521 522 for (i = 0; i < NOBJLISTS; ++i) 523 TAILQ_INIT(&swap_pager_object_list[i]); 524 mtx_init(&sw_dev_mtx, "swapdev", NULL, MTX_DEF); 525 sx_init(&sw_alloc_sx, "swspsx"); 526 sx_init(&swdev_syscall_lock, "swsysc"); 527 } 528 529 static void 530 swap_pager_counters(void) 531 { 532 533 swap_free_deferred = counter_u64_alloc(M_WAITOK); 534 swap_free_completed = counter_u64_alloc(M_WAITOK); 535 } 536 SYSINIT(swap_counters, SI_SUB_CPU, SI_ORDER_ANY, swap_pager_counters, NULL); 537 538 /* 539 * SWAP_PAGER_SWAP_INIT() - swap pager initialization from pageout process 540 * 541 * Expected to be started from pageout process once, prior to entering 542 * its main loop. 543 */ 544 void 545 swap_pager_swap_init(void) 546 { 547 unsigned long n, n2; 548 549 /* 550 * Number of in-transit swap bp operations. Don't 551 * exhaust the pbufs completely. Make sure we 552 * initialize workable values (0 will work for hysteresis 553 * but it isn't very efficient). 554 * 555 * The nsw_cluster_max is constrained by the bp->b_pages[] 556 * array, which has MAXPHYS / PAGE_SIZE entries, and our locally 557 * defined MAX_PAGEOUT_CLUSTER. Also be aware that swap ops are 558 * constrained by the swap device interleave stripe size. 559 * 560 * Currently we hardwire nsw_wcount_async to 4. This limit is 561 * designed to prevent other I/O from having high latencies due to 562 * our pageout I/O. The value 4 works well for one or two active swap 563 * devices but is probably a little low if you have more. Even so, 564 * a higher value would probably generate only a limited improvement 565 * with three or four active swap devices since the system does not 566 * typically have to pageout at extreme bandwidths. We will want 567 * at least 2 per swap devices, and 4 is a pretty good value if you 568 * have one NFS swap device due to the command/ack latency over NFS. 569 * So it all works out pretty well. 570 */ 571 nsw_cluster_max = min(MAXPHYS / PAGE_SIZE, MAX_PAGEOUT_CLUSTER); 572 573 nsw_wcount_async = 4; 574 nsw_wcount_async_max = nsw_wcount_async; 575 mtx_init(&swbuf_mtx, "async swbuf mutex", NULL, MTX_DEF); 576 577 swwbuf_zone = pbuf_zsecond_create("swwbuf", nswbuf / 4); 578 swrbuf_zone = pbuf_zsecond_create("swrbuf", nswbuf / 2); 579 580 /* 581 * Initialize our zone, taking the user's requested size or 582 * estimating the number we need based on the number of pages 583 * in the system. 584 */ 585 n = maxswzone != 0 ? maxswzone / sizeof(struct swblk) : 586 vm_cnt.v_page_count / 2; 587 swpctrie_zone = uma_zcreate("swpctrie", pctrie_node_size(), NULL, NULL, 588 pctrie_zone_init, NULL, UMA_ALIGN_PTR, UMA_ZONE_VM); 589 if (swpctrie_zone == NULL) 590 panic("failed to create swap pctrie zone."); 591 swblk_zone = uma_zcreate("swblk", sizeof(struct swblk), NULL, NULL, 592 NULL, NULL, _Alignof(struct swblk) - 1, UMA_ZONE_VM); 593 if (swblk_zone == NULL) 594 panic("failed to create swap blk zone."); 595 n2 = n; 596 do { 597 if (uma_zone_reserve_kva(swblk_zone, n)) 598 break; 599 /* 600 * if the allocation failed, try a zone two thirds the 601 * size of the previous attempt. 602 */ 603 n -= ((n + 2) / 3); 604 } while (n > 0); 605 606 /* 607 * Often uma_zone_reserve_kva() cannot reserve exactly the 608 * requested size. Account for the difference when 609 * calculating swap_maxpages. 610 */ 611 n = uma_zone_get_max(swblk_zone); 612 613 if (n < n2) 614 printf("Swap blk zone entries changed from %lu to %lu.\n", 615 n2, n); 616 /* absolute maximum we can handle assuming 100% efficiency */ 617 swap_maxpages = n * SWAP_META_PAGES; 618 swzone = n * sizeof(struct swblk); 619 if (!uma_zone_reserve_kva(swpctrie_zone, n)) 620 printf("Cannot reserve swap pctrie zone, " 621 "reduce kern.maxswzone.\n"); 622 } 623 624 static vm_object_t 625 swap_pager_alloc_init(void *handle, struct ucred *cred, vm_ooffset_t size, 626 vm_ooffset_t offset) 627 { 628 vm_object_t object; 629 630 if (cred != NULL) { 631 if (!swap_reserve_by_cred(size, cred)) 632 return (NULL); 633 crhold(cred); 634 } 635 636 /* 637 * The un_pager.swp.swp_blks trie is initialized by 638 * vm_object_allocate() to ensure the correct order of 639 * visibility to other threads. 640 */ 641 object = vm_object_allocate(OBJT_SWAP, OFF_TO_IDX(offset + 642 PAGE_MASK + size)); 643 644 object->un_pager.swp.writemappings = 0; 645 object->handle = handle; 646 if (cred != NULL) { 647 object->cred = cred; 648 object->charge = size; 649 } 650 return (object); 651 } 652 653 /* 654 * SWAP_PAGER_ALLOC() - allocate a new OBJT_SWAP VM object and instantiate 655 * its metadata structures. 656 * 657 * This routine is called from the mmap and fork code to create a new 658 * OBJT_SWAP object. 659 * 660 * This routine must ensure that no live duplicate is created for 661 * the named object request, which is protected against by 662 * holding the sw_alloc_sx lock in case handle != NULL. 663 */ 664 static vm_object_t 665 swap_pager_alloc(void *handle, vm_ooffset_t size, vm_prot_t prot, 666 vm_ooffset_t offset, struct ucred *cred) 667 { 668 vm_object_t object; 669 670 if (handle != NULL) { 671 /* 672 * Reference existing named region or allocate new one. There 673 * should not be a race here against swp_pager_meta_build() 674 * as called from vm_page_remove() in regards to the lookup 675 * of the handle. 676 */ 677 sx_xlock(&sw_alloc_sx); 678 object = vm_pager_object_lookup(NOBJLIST(handle), handle); 679 if (object == NULL) { 680 object = swap_pager_alloc_init(handle, cred, size, 681 offset); 682 if (object != NULL) { 683 TAILQ_INSERT_TAIL(NOBJLIST(object->handle), 684 object, pager_object_list); 685 } 686 } 687 sx_xunlock(&sw_alloc_sx); 688 } else { 689 object = swap_pager_alloc_init(handle, cred, size, offset); 690 } 691 return (object); 692 } 693 694 /* 695 * SWAP_PAGER_DEALLOC() - remove swap metadata from object 696 * 697 * The swap backing for the object is destroyed. The code is 698 * designed such that we can reinstantiate it later, but this 699 * routine is typically called only when the entire object is 700 * about to be destroyed. 701 * 702 * The object must be locked. 703 */ 704 static void 705 swap_pager_dealloc(vm_object_t object) 706 { 707 708 VM_OBJECT_ASSERT_WLOCKED(object); 709 KASSERT((object->flags & OBJ_DEAD) != 0, ("dealloc of reachable obj")); 710 711 /* 712 * Remove from list right away so lookups will fail if we block for 713 * pageout completion. 714 */ 715 if ((object->flags & OBJ_ANON) == 0 && object->handle != NULL) { 716 VM_OBJECT_WUNLOCK(object); 717 sx_xlock(&sw_alloc_sx); 718 TAILQ_REMOVE(NOBJLIST(object->handle), object, 719 pager_object_list); 720 sx_xunlock(&sw_alloc_sx); 721 VM_OBJECT_WLOCK(object); 722 } 723 724 vm_object_pip_wait(object, "swpdea"); 725 726 /* 727 * Free all remaining metadata. We only bother to free it from 728 * the swap meta data. We do not attempt to free swapblk's still 729 * associated with vm_page_t's for this object. We do not care 730 * if paging is still in progress on some objects. 731 */ 732 swp_pager_meta_free_all(object); 733 object->handle = NULL; 734 object->type = OBJT_DEAD; 735 } 736 737 /************************************************************************ 738 * SWAP PAGER BITMAP ROUTINES * 739 ************************************************************************/ 740 741 /* 742 * SWP_PAGER_GETSWAPSPACE() - allocate raw swap space 743 * 744 * Allocate swap for up to the requested number of pages, and at 745 * least a minimum number of pages. The starting swap block number 746 * (a page index) is returned or SWAPBLK_NONE if the allocation 747 * failed. 748 * 749 * Also has the side effect of advising that somebody made a mistake 750 * when they configured swap and didn't configure enough. 751 * 752 * This routine may not sleep. 753 * 754 * We allocate in round-robin fashion from the configured devices. 755 */ 756 static daddr_t 757 swp_pager_getswapspace(int *io_npages, int limit) 758 { 759 daddr_t blk; 760 struct swdevt *sp; 761 int mpages, npages; 762 763 blk = SWAPBLK_NONE; 764 mpages = *io_npages; 765 npages = imin(BLIST_MAX_ALLOC, mpages); 766 mtx_lock(&sw_dev_mtx); 767 sp = swdevhd; 768 while (!TAILQ_EMPTY(&swtailq)) { 769 if (sp == NULL) 770 sp = TAILQ_FIRST(&swtailq); 771 if ((sp->sw_flags & SW_CLOSING) == 0) 772 blk = blist_alloc(sp->sw_blist, &npages, mpages); 773 if (blk != SWAPBLK_NONE) 774 break; 775 sp = TAILQ_NEXT(sp, sw_list); 776 if (swdevhd == sp) { 777 if (npages <= limit) 778 break; 779 mpages = npages - 1; 780 npages >>= 1; 781 } 782 } 783 if (blk != SWAPBLK_NONE) { 784 *io_npages = npages; 785 blk += sp->sw_first; 786 sp->sw_used += npages; 787 swap_pager_avail -= npages; 788 swp_sizecheck(); 789 swdevhd = TAILQ_NEXT(sp, sw_list); 790 } else { 791 if (swap_pager_full != 2) { 792 printf("swp_pager_getswapspace(%d): failed\n", 793 *io_npages); 794 swap_pager_full = 2; 795 swap_pager_almost_full = 1; 796 } 797 swdevhd = NULL; 798 } 799 mtx_unlock(&sw_dev_mtx); 800 return (blk); 801 } 802 803 static bool 804 swp_pager_isondev(daddr_t blk, struct swdevt *sp) 805 { 806 807 return (blk >= sp->sw_first && blk < sp->sw_end); 808 } 809 810 static void 811 swp_pager_strategy(struct buf *bp) 812 { 813 struct swdevt *sp; 814 815 mtx_lock(&sw_dev_mtx); 816 TAILQ_FOREACH(sp, &swtailq, sw_list) { 817 if (swp_pager_isondev(bp->b_blkno, sp)) { 818 mtx_unlock(&sw_dev_mtx); 819 if ((sp->sw_flags & SW_UNMAPPED) != 0 && 820 unmapped_buf_allowed) { 821 bp->b_data = unmapped_buf; 822 bp->b_offset = 0; 823 } else { 824 pmap_qenter((vm_offset_t)bp->b_data, 825 &bp->b_pages[0], bp->b_bcount / PAGE_SIZE); 826 } 827 sp->sw_strategy(bp, sp); 828 return; 829 } 830 } 831 panic("Swapdev not found"); 832 } 833 834 835 /* 836 * SWP_PAGER_FREESWAPSPACE() - free raw swap space 837 * 838 * This routine returns the specified swap blocks back to the bitmap. 839 * 840 * This routine may not sleep. 841 */ 842 static void 843 swp_pager_freeswapspace(daddr_t blk, daddr_t npages) 844 { 845 struct swdevt *sp; 846 847 if (npages == 0) 848 return; 849 mtx_lock(&sw_dev_mtx); 850 TAILQ_FOREACH(sp, &swtailq, sw_list) { 851 if (swp_pager_isondev(blk, sp)) { 852 sp->sw_used -= npages; 853 /* 854 * If we are attempting to stop swapping on 855 * this device, we don't want to mark any 856 * blocks free lest they be reused. 857 */ 858 if ((sp->sw_flags & SW_CLOSING) == 0) { 859 blist_free(sp->sw_blist, blk - sp->sw_first, 860 npages); 861 swap_pager_avail += npages; 862 swp_sizecheck(); 863 } 864 mtx_unlock(&sw_dev_mtx); 865 return; 866 } 867 } 868 panic("Swapdev not found"); 869 } 870 871 /* 872 * SYSCTL_SWAP_FRAGMENTATION() - produce raw swap space stats 873 */ 874 static int 875 sysctl_swap_fragmentation(SYSCTL_HANDLER_ARGS) 876 { 877 struct sbuf sbuf; 878 struct swdevt *sp; 879 const char *devname; 880 int error; 881 882 error = sysctl_wire_old_buffer(req, 0); 883 if (error != 0) 884 return (error); 885 sbuf_new_for_sysctl(&sbuf, NULL, 128, req); 886 mtx_lock(&sw_dev_mtx); 887 TAILQ_FOREACH(sp, &swtailq, sw_list) { 888 if (vn_isdisk(sp->sw_vp, NULL)) 889 devname = devtoname(sp->sw_vp->v_rdev); 890 else 891 devname = "[file]"; 892 sbuf_printf(&sbuf, "\nFree space on device %s:\n", devname); 893 blist_stats(sp->sw_blist, &sbuf); 894 } 895 mtx_unlock(&sw_dev_mtx); 896 error = sbuf_finish(&sbuf); 897 sbuf_delete(&sbuf); 898 return (error); 899 } 900 901 /* 902 * SWAP_PAGER_FREESPACE() - frees swap blocks associated with a page 903 * range within an object. 904 * 905 * This is a globally accessible routine. 906 * 907 * This routine removes swapblk assignments from swap metadata. 908 * 909 * The external callers of this routine typically have already destroyed 910 * or renamed vm_page_t's associated with this range in the object so 911 * we should be ok. 912 * 913 * The object must be locked. 914 */ 915 void 916 swap_pager_freespace(vm_object_t object, vm_pindex_t start, vm_size_t size) 917 { 918 919 swp_pager_meta_free(object, start, size); 920 } 921 922 /* 923 * SWAP_PAGER_RESERVE() - reserve swap blocks in object 924 * 925 * Assigns swap blocks to the specified range within the object. The 926 * swap blocks are not zeroed. Any previous swap assignment is destroyed. 927 * 928 * Returns 0 on success, -1 on failure. 929 */ 930 int 931 swap_pager_reserve(vm_object_t object, vm_pindex_t start, vm_size_t size) 932 { 933 daddr_t addr, blk, n_free, s_free; 934 int i, j, n; 935 936 swp_pager_init_freerange(&s_free, &n_free); 937 VM_OBJECT_WLOCK(object); 938 for (i = 0; i < size; i += n) { 939 n = size - i; 940 blk = swp_pager_getswapspace(&n, 1); 941 if (blk == SWAPBLK_NONE) { 942 swp_pager_meta_free(object, start, i); 943 VM_OBJECT_WUNLOCK(object); 944 return (-1); 945 } 946 for (j = 0; j < n; ++j) { 947 addr = swp_pager_meta_build(object, 948 start + i + j, blk + j); 949 if (addr != SWAPBLK_NONE) 950 swp_pager_update_freerange(&s_free, &n_free, 951 addr); 952 } 953 } 954 swp_pager_freeswapspace(s_free, n_free); 955 VM_OBJECT_WUNLOCK(object); 956 return (0); 957 } 958 959 static bool 960 swp_pager_xfer_source(vm_object_t srcobject, vm_object_t dstobject, 961 vm_pindex_t pindex, daddr_t addr) 962 { 963 daddr_t dstaddr; 964 965 KASSERT(srcobject->type == OBJT_SWAP, 966 ("%s: Srcobject not swappable", __func__)); 967 if (dstobject->type == OBJT_SWAP && 968 swp_pager_meta_lookup(dstobject, pindex) != SWAPBLK_NONE) { 969 /* Caller should destroy the source block. */ 970 return (false); 971 } 972 973 /* 974 * Destination has no swapblk and is not resident, transfer source. 975 * swp_pager_meta_build() can sleep. 976 */ 977 VM_OBJECT_WUNLOCK(srcobject); 978 dstaddr = swp_pager_meta_build(dstobject, pindex, addr); 979 KASSERT(dstaddr == SWAPBLK_NONE, 980 ("Unexpected destination swapblk")); 981 VM_OBJECT_WLOCK(srcobject); 982 983 return (true); 984 } 985 986 /* 987 * SWAP_PAGER_COPY() - copy blocks from source pager to destination pager 988 * and destroy the source. 989 * 990 * Copy any valid swapblks from the source to the destination. In 991 * cases where both the source and destination have a valid swapblk, 992 * we keep the destination's. 993 * 994 * This routine is allowed to sleep. It may sleep allocating metadata 995 * indirectly through swp_pager_meta_build(). 996 * 997 * The source object contains no vm_page_t's (which is just as well) 998 * 999 * The source object is of type OBJT_SWAP. 1000 * 1001 * The source and destination objects must be locked. 1002 * Both object locks may temporarily be released. 1003 */ 1004 void 1005 swap_pager_copy(vm_object_t srcobject, vm_object_t dstobject, 1006 vm_pindex_t offset, int destroysource) 1007 { 1008 1009 VM_OBJECT_ASSERT_WLOCKED(srcobject); 1010 VM_OBJECT_ASSERT_WLOCKED(dstobject); 1011 1012 /* 1013 * If destroysource is set, we remove the source object from the 1014 * swap_pager internal queue now. 1015 */ 1016 if (destroysource && (srcobject->flags & OBJ_ANON) == 0 && 1017 srcobject->handle != NULL) { 1018 VM_OBJECT_WUNLOCK(srcobject); 1019 VM_OBJECT_WUNLOCK(dstobject); 1020 sx_xlock(&sw_alloc_sx); 1021 TAILQ_REMOVE(NOBJLIST(srcobject->handle), srcobject, 1022 pager_object_list); 1023 sx_xunlock(&sw_alloc_sx); 1024 VM_OBJECT_WLOCK(dstobject); 1025 VM_OBJECT_WLOCK(srcobject); 1026 } 1027 1028 /* 1029 * Transfer source to destination. 1030 */ 1031 swp_pager_meta_transfer(srcobject, dstobject, offset, dstobject->size); 1032 1033 /* 1034 * Free left over swap blocks in source. 1035 * 1036 * We have to revert the type to OBJT_DEFAULT so we do not accidentally 1037 * double-remove the object from the swap queues. 1038 */ 1039 if (destroysource) { 1040 swp_pager_meta_free_all(srcobject); 1041 /* 1042 * Reverting the type is not necessary, the caller is going 1043 * to destroy srcobject directly, but I'm doing it here 1044 * for consistency since we've removed the object from its 1045 * queues. 1046 */ 1047 srcobject->type = OBJT_DEFAULT; 1048 } 1049 } 1050 1051 /* 1052 * SWAP_PAGER_HASPAGE() - determine if we have good backing store for 1053 * the requested page. 1054 * 1055 * We determine whether good backing store exists for the requested 1056 * page and return TRUE if it does, FALSE if it doesn't. 1057 * 1058 * If TRUE, we also try to determine how much valid, contiguous backing 1059 * store exists before and after the requested page. 1060 */ 1061 static boolean_t 1062 swap_pager_haspage(vm_object_t object, vm_pindex_t pindex, int *before, 1063 int *after) 1064 { 1065 daddr_t blk, blk0; 1066 int i; 1067 1068 VM_OBJECT_ASSERT_LOCKED(object); 1069 KASSERT(object->type == OBJT_SWAP, 1070 ("%s: object not swappable", __func__)); 1071 1072 /* 1073 * do we have good backing store at the requested index ? 1074 */ 1075 blk0 = swp_pager_meta_lookup(object, pindex); 1076 if (blk0 == SWAPBLK_NONE) { 1077 if (before) 1078 *before = 0; 1079 if (after) 1080 *after = 0; 1081 return (FALSE); 1082 } 1083 1084 /* 1085 * find backwards-looking contiguous good backing store 1086 */ 1087 if (before != NULL) { 1088 for (i = 1; i < SWB_NPAGES; i++) { 1089 if (i > pindex) 1090 break; 1091 blk = swp_pager_meta_lookup(object, pindex - i); 1092 if (blk != blk0 - i) 1093 break; 1094 } 1095 *before = i - 1; 1096 } 1097 1098 /* 1099 * find forward-looking contiguous good backing store 1100 */ 1101 if (after != NULL) { 1102 for (i = 1; i < SWB_NPAGES; i++) { 1103 blk = swp_pager_meta_lookup(object, pindex + i); 1104 if (blk != blk0 + i) 1105 break; 1106 } 1107 *after = i - 1; 1108 } 1109 return (TRUE); 1110 } 1111 1112 /* 1113 * SWAP_PAGER_PAGE_UNSWAPPED() - remove swap backing store related to page 1114 * 1115 * This removes any associated swap backing store, whether valid or 1116 * not, from the page. 1117 * 1118 * This routine is typically called when a page is made dirty, at 1119 * which point any associated swap can be freed. MADV_FREE also 1120 * calls us in a special-case situation 1121 * 1122 * NOTE!!! If the page is clean and the swap was valid, the caller 1123 * should make the page dirty before calling this routine. This routine 1124 * does NOT change the m->dirty status of the page. Also: MADV_FREE 1125 * depends on it. 1126 * 1127 * This routine may not sleep. 1128 * 1129 * The object containing the page may be locked. 1130 */ 1131 static void 1132 swap_pager_unswapped(vm_page_t m) 1133 { 1134 struct swblk *sb; 1135 vm_object_t obj; 1136 1137 /* 1138 * Handle enqueing deferred frees first. If we do not have the 1139 * object lock we wait for the page daemon to clear the space. 1140 */ 1141 obj = m->object; 1142 if (!VM_OBJECT_WOWNED(obj)) { 1143 VM_PAGE_OBJECT_BUSY_ASSERT(m); 1144 /* 1145 * The caller is responsible for synchronization but we 1146 * will harmlessly handle races. This is typically provided 1147 * by only calling unswapped() when a page transitions from 1148 * clean to dirty. 1149 */ 1150 if ((m->a.flags & (PGA_SWAP_SPACE | PGA_SWAP_FREE)) == 1151 PGA_SWAP_SPACE) { 1152 vm_page_aflag_set(m, PGA_SWAP_FREE); 1153 counter_u64_add(swap_free_deferred, 1); 1154 } 1155 return; 1156 } 1157 if ((m->a.flags & PGA_SWAP_FREE) != 0) 1158 counter_u64_add(swap_free_completed, 1); 1159 vm_page_aflag_clear(m, PGA_SWAP_FREE | PGA_SWAP_SPACE); 1160 1161 /* 1162 * The meta data only exists if the object is OBJT_SWAP 1163 * and even then might not be allocated yet. 1164 */ 1165 KASSERT(m->object->type == OBJT_SWAP, 1166 ("Free object not swappable")); 1167 1168 sb = SWAP_PCTRIE_LOOKUP(&m->object->un_pager.swp.swp_blks, 1169 rounddown(m->pindex, SWAP_META_PAGES)); 1170 if (sb == NULL) 1171 return; 1172 if (sb->d[m->pindex % SWAP_META_PAGES] == SWAPBLK_NONE) 1173 return; 1174 swp_pager_freeswapspace(sb->d[m->pindex % SWAP_META_PAGES], 1); 1175 sb->d[m->pindex % SWAP_META_PAGES] = SWAPBLK_NONE; 1176 swp_pager_free_empty_swblk(m->object, sb); 1177 } 1178 1179 /* 1180 * swap_pager_getpages() - bring pages in from swap 1181 * 1182 * Attempt to page in the pages in array "ma" of length "count". The 1183 * caller may optionally specify that additional pages preceding and 1184 * succeeding the specified range be paged in. The number of such pages 1185 * is returned in the "rbehind" and "rahead" parameters, and they will 1186 * be in the inactive queue upon return. 1187 * 1188 * The pages in "ma" must be busied and will remain busied upon return. 1189 */ 1190 static int 1191 swap_pager_getpages(vm_object_t object, vm_page_t *ma, int count, int *rbehind, 1192 int *rahead) 1193 { 1194 struct buf *bp; 1195 vm_page_t bm, mpred, msucc, p; 1196 vm_pindex_t pindex; 1197 daddr_t blk; 1198 int i, maxahead, maxbehind, reqcount; 1199 1200 VM_OBJECT_WLOCK(object); 1201 reqcount = count; 1202 1203 KASSERT(object->type == OBJT_SWAP, 1204 ("%s: object not swappable", __func__)); 1205 if (!swap_pager_haspage(object, ma[0]->pindex, &maxbehind, &maxahead)) { 1206 VM_OBJECT_WUNLOCK(object); 1207 return (VM_PAGER_FAIL); 1208 } 1209 1210 KASSERT(reqcount - 1 <= maxahead, 1211 ("page count %d extends beyond swap block", reqcount)); 1212 1213 /* 1214 * Do not transfer any pages other than those that are xbusied 1215 * when running during a split or collapse operation. This 1216 * prevents clustering from re-creating pages which are being 1217 * moved into another object. 1218 */ 1219 if ((object->flags & (OBJ_SPLIT | OBJ_DEAD)) != 0) { 1220 maxahead = reqcount - 1; 1221 maxbehind = 0; 1222 } 1223 1224 /* 1225 * Clip the readahead and readbehind ranges to exclude resident pages. 1226 */ 1227 if (rahead != NULL) { 1228 *rahead = imin(*rahead, maxahead - (reqcount - 1)); 1229 pindex = ma[reqcount - 1]->pindex; 1230 msucc = TAILQ_NEXT(ma[reqcount - 1], listq); 1231 if (msucc != NULL && msucc->pindex - pindex - 1 < *rahead) 1232 *rahead = msucc->pindex - pindex - 1; 1233 } 1234 if (rbehind != NULL) { 1235 *rbehind = imin(*rbehind, maxbehind); 1236 pindex = ma[0]->pindex; 1237 mpred = TAILQ_PREV(ma[0], pglist, listq); 1238 if (mpred != NULL && pindex - mpred->pindex - 1 < *rbehind) 1239 *rbehind = pindex - mpred->pindex - 1; 1240 } 1241 1242 bm = ma[0]; 1243 for (i = 0; i < count; i++) 1244 ma[i]->oflags |= VPO_SWAPINPROG; 1245 1246 /* 1247 * Allocate readahead and readbehind pages. 1248 */ 1249 if (rbehind != NULL) { 1250 for (i = 1; i <= *rbehind; i++) { 1251 p = vm_page_alloc(object, ma[0]->pindex - i, 1252 VM_ALLOC_NORMAL); 1253 if (p == NULL) 1254 break; 1255 p->oflags |= VPO_SWAPINPROG; 1256 bm = p; 1257 } 1258 *rbehind = i - 1; 1259 } 1260 if (rahead != NULL) { 1261 for (i = 0; i < *rahead; i++) { 1262 p = vm_page_alloc(object, 1263 ma[reqcount - 1]->pindex + i + 1, VM_ALLOC_NORMAL); 1264 if (p == NULL) 1265 break; 1266 p->oflags |= VPO_SWAPINPROG; 1267 } 1268 *rahead = i; 1269 } 1270 if (rbehind != NULL) 1271 count += *rbehind; 1272 if (rahead != NULL) 1273 count += *rahead; 1274 1275 vm_object_pip_add(object, count); 1276 1277 pindex = bm->pindex; 1278 blk = swp_pager_meta_lookup(object, pindex); 1279 KASSERT(blk != SWAPBLK_NONE, 1280 ("no swap blocking containing %p(%jx)", object, (uintmax_t)pindex)); 1281 1282 VM_OBJECT_WUNLOCK(object); 1283 bp = uma_zalloc(swrbuf_zone, M_WAITOK); 1284 /* Pages cannot leave the object while busy. */ 1285 for (i = 0, p = bm; i < count; i++, p = TAILQ_NEXT(p, listq)) { 1286 MPASS(p->pindex == bm->pindex + i); 1287 bp->b_pages[i] = p; 1288 } 1289 1290 bp->b_flags |= B_PAGING; 1291 bp->b_iocmd = BIO_READ; 1292 bp->b_iodone = swp_pager_async_iodone; 1293 bp->b_rcred = crhold(thread0.td_ucred); 1294 bp->b_wcred = crhold(thread0.td_ucred); 1295 bp->b_blkno = blk; 1296 bp->b_bcount = PAGE_SIZE * count; 1297 bp->b_bufsize = PAGE_SIZE * count; 1298 bp->b_npages = count; 1299 bp->b_pgbefore = rbehind != NULL ? *rbehind : 0; 1300 bp->b_pgafter = rahead != NULL ? *rahead : 0; 1301 1302 VM_CNT_INC(v_swapin); 1303 VM_CNT_ADD(v_swappgsin, count); 1304 1305 /* 1306 * perform the I/O. NOTE!!! bp cannot be considered valid after 1307 * this point because we automatically release it on completion. 1308 * Instead, we look at the one page we are interested in which we 1309 * still hold a lock on even through the I/O completion. 1310 * 1311 * The other pages in our ma[] array are also released on completion, 1312 * so we cannot assume they are valid anymore either. 1313 * 1314 * NOTE: b_blkno is destroyed by the call to swapdev_strategy 1315 */ 1316 BUF_KERNPROC(bp); 1317 swp_pager_strategy(bp); 1318 1319 /* 1320 * Wait for the pages we want to complete. VPO_SWAPINPROG is always 1321 * cleared on completion. If an I/O error occurs, SWAPBLK_NONE 1322 * is set in the metadata for each page in the request. 1323 */ 1324 VM_OBJECT_WLOCK(object); 1325 /* This could be implemented more efficiently with aflags */ 1326 while ((ma[0]->oflags & VPO_SWAPINPROG) != 0) { 1327 ma[0]->oflags |= VPO_SWAPSLEEP; 1328 VM_CNT_INC(v_intrans); 1329 if (VM_OBJECT_SLEEP(object, &object->handle, PSWP, 1330 "swread", hz * 20)) { 1331 printf( 1332 "swap_pager: indefinite wait buffer: bufobj: %p, blkno: %jd, size: %ld\n", 1333 bp->b_bufobj, (intmax_t)bp->b_blkno, bp->b_bcount); 1334 } 1335 } 1336 VM_OBJECT_WUNLOCK(object); 1337 1338 /* 1339 * If we had an unrecoverable read error pages will not be valid. 1340 */ 1341 for (i = 0; i < reqcount; i++) 1342 if (ma[i]->valid != VM_PAGE_BITS_ALL) 1343 return (VM_PAGER_ERROR); 1344 1345 return (VM_PAGER_OK); 1346 1347 /* 1348 * A final note: in a low swap situation, we cannot deallocate swap 1349 * and mark a page dirty here because the caller is likely to mark 1350 * the page clean when we return, causing the page to possibly revert 1351 * to all-zero's later. 1352 */ 1353 } 1354 1355 /* 1356 * swap_pager_getpages_async(): 1357 * 1358 * Right now this is emulation of asynchronous operation on top of 1359 * swap_pager_getpages(). 1360 */ 1361 static int 1362 swap_pager_getpages_async(vm_object_t object, vm_page_t *ma, int count, 1363 int *rbehind, int *rahead, pgo_getpages_iodone_t iodone, void *arg) 1364 { 1365 int r, error; 1366 1367 r = swap_pager_getpages(object, ma, count, rbehind, rahead); 1368 switch (r) { 1369 case VM_PAGER_OK: 1370 error = 0; 1371 break; 1372 case VM_PAGER_ERROR: 1373 error = EIO; 1374 break; 1375 case VM_PAGER_FAIL: 1376 error = EINVAL; 1377 break; 1378 default: 1379 panic("unhandled swap_pager_getpages() error %d", r); 1380 } 1381 (iodone)(arg, ma, count, error); 1382 1383 return (r); 1384 } 1385 1386 /* 1387 * swap_pager_putpages: 1388 * 1389 * Assign swap (if necessary) and initiate I/O on the specified pages. 1390 * 1391 * We support both OBJT_DEFAULT and OBJT_SWAP objects. DEFAULT objects 1392 * are automatically converted to SWAP objects. 1393 * 1394 * In a low memory situation we may block in VOP_STRATEGY(), but the new 1395 * vm_page reservation system coupled with properly written VFS devices 1396 * should ensure that no low-memory deadlock occurs. This is an area 1397 * which needs work. 1398 * 1399 * The parent has N vm_object_pip_add() references prior to 1400 * calling us and will remove references for rtvals[] that are 1401 * not set to VM_PAGER_PEND. We need to remove the rest on I/O 1402 * completion. 1403 * 1404 * The parent has soft-busy'd the pages it passes us and will unbusy 1405 * those whose rtvals[] entry is not set to VM_PAGER_PEND on return. 1406 * We need to unbusy the rest on I/O completion. 1407 */ 1408 static void 1409 swap_pager_putpages(vm_object_t object, vm_page_t *ma, int count, 1410 int flags, int *rtvals) 1411 { 1412 struct buf *bp; 1413 daddr_t addr, blk, n_free, s_free; 1414 vm_page_t mreq; 1415 int i, j, n; 1416 bool async; 1417 1418 KASSERT(count == 0 || ma[0]->object == object, 1419 ("%s: object mismatch %p/%p", 1420 __func__, object, ma[0]->object)); 1421 1422 /* 1423 * Step 1 1424 * 1425 * Turn object into OBJT_SWAP. Force sync if not a pageout process. 1426 */ 1427 if (object->type != OBJT_SWAP) { 1428 addr = swp_pager_meta_build(object, 0, SWAPBLK_NONE); 1429 KASSERT(addr == SWAPBLK_NONE, 1430 ("unexpected object swap block")); 1431 } 1432 VM_OBJECT_WUNLOCK(object); 1433 async = curproc == pageproc && (flags & VM_PAGER_PUT_SYNC) == 0; 1434 swp_pager_init_freerange(&s_free, &n_free); 1435 1436 /* 1437 * Step 2 1438 * 1439 * Assign swap blocks and issue I/O. We reallocate swap on the fly. 1440 * The page is left dirty until the pageout operation completes 1441 * successfully. 1442 */ 1443 for (i = 0; i < count; i += n) { 1444 /* Maximum I/O size is limited by maximum swap block size. */ 1445 n = min(count - i, nsw_cluster_max); 1446 1447 /* Get a block of swap of size up to size n. */ 1448 blk = swp_pager_getswapspace(&n, 4); 1449 if (blk == SWAPBLK_NONE) { 1450 for (j = 0; j < n; ++j) 1451 rtvals[i + j] = VM_PAGER_FAIL; 1452 continue; 1453 } 1454 1455 /* 1456 * All I/O parameters have been satisfied. Build the I/O 1457 * request and assign the swap space. 1458 */ 1459 if (async) { 1460 mtx_lock(&swbuf_mtx); 1461 while (nsw_wcount_async == 0) 1462 msleep(&nsw_wcount_async, &swbuf_mtx, PVM, 1463 "swbufa", 0); 1464 nsw_wcount_async--; 1465 mtx_unlock(&swbuf_mtx); 1466 } 1467 bp = uma_zalloc(swwbuf_zone, M_WAITOK); 1468 if (async) 1469 bp->b_flags = B_ASYNC; 1470 bp->b_flags |= B_PAGING; 1471 bp->b_iocmd = BIO_WRITE; 1472 1473 bp->b_rcred = crhold(thread0.td_ucred); 1474 bp->b_wcred = crhold(thread0.td_ucred); 1475 bp->b_bcount = PAGE_SIZE * n; 1476 bp->b_bufsize = PAGE_SIZE * n; 1477 bp->b_blkno = blk; 1478 1479 VM_OBJECT_WLOCK(object); 1480 for (j = 0; j < n; ++j) { 1481 mreq = ma[i + j]; 1482 vm_page_aflag_clear(mreq, PGA_SWAP_FREE); 1483 addr = swp_pager_meta_build(mreq->object, mreq->pindex, 1484 blk + j); 1485 if (addr != SWAPBLK_NONE) 1486 swp_pager_update_freerange(&s_free, &n_free, 1487 addr); 1488 MPASS(mreq->dirty == VM_PAGE_BITS_ALL); 1489 mreq->oflags |= VPO_SWAPINPROG; 1490 bp->b_pages[j] = mreq; 1491 } 1492 VM_OBJECT_WUNLOCK(object); 1493 bp->b_npages = n; 1494 /* 1495 * Must set dirty range for NFS to work. 1496 */ 1497 bp->b_dirtyoff = 0; 1498 bp->b_dirtyend = bp->b_bcount; 1499 1500 VM_CNT_INC(v_swapout); 1501 VM_CNT_ADD(v_swappgsout, bp->b_npages); 1502 1503 /* 1504 * We unconditionally set rtvals[] to VM_PAGER_PEND so that we 1505 * can call the async completion routine at the end of a 1506 * synchronous I/O operation. Otherwise, our caller would 1507 * perform duplicate unbusy and wakeup operations on the page 1508 * and object, respectively. 1509 */ 1510 for (j = 0; j < n; j++) 1511 rtvals[i + j] = VM_PAGER_PEND; 1512 1513 /* 1514 * asynchronous 1515 * 1516 * NOTE: b_blkno is destroyed by the call to swapdev_strategy. 1517 */ 1518 if (async) { 1519 bp->b_iodone = swp_pager_async_iodone; 1520 BUF_KERNPROC(bp); 1521 swp_pager_strategy(bp); 1522 continue; 1523 } 1524 1525 /* 1526 * synchronous 1527 * 1528 * NOTE: b_blkno is destroyed by the call to swapdev_strategy. 1529 */ 1530 bp->b_iodone = bdone; 1531 swp_pager_strategy(bp); 1532 1533 /* 1534 * Wait for the sync I/O to complete. 1535 */ 1536 bwait(bp, PVM, "swwrt"); 1537 1538 /* 1539 * Now that we are through with the bp, we can call the 1540 * normal async completion, which frees everything up. 1541 */ 1542 swp_pager_async_iodone(bp); 1543 } 1544 swp_pager_freeswapspace(s_free, n_free); 1545 VM_OBJECT_WLOCK(object); 1546 } 1547 1548 /* 1549 * swp_pager_async_iodone: 1550 * 1551 * Completion routine for asynchronous reads and writes from/to swap. 1552 * Also called manually by synchronous code to finish up a bp. 1553 * 1554 * This routine may not sleep. 1555 */ 1556 static void 1557 swp_pager_async_iodone(struct buf *bp) 1558 { 1559 int i; 1560 vm_object_t object = NULL; 1561 1562 /* 1563 * Report error - unless we ran out of memory, in which case 1564 * we've already logged it in swapgeom_strategy(). 1565 */ 1566 if (bp->b_ioflags & BIO_ERROR && bp->b_error != ENOMEM) { 1567 printf( 1568 "swap_pager: I/O error - %s failed; blkno %ld," 1569 "size %ld, error %d\n", 1570 ((bp->b_iocmd == BIO_READ) ? "pagein" : "pageout"), 1571 (long)bp->b_blkno, 1572 (long)bp->b_bcount, 1573 bp->b_error 1574 ); 1575 } 1576 1577 /* 1578 * remove the mapping for kernel virtual 1579 */ 1580 if (buf_mapped(bp)) 1581 pmap_qremove((vm_offset_t)bp->b_data, bp->b_npages); 1582 else 1583 bp->b_data = bp->b_kvabase; 1584 1585 if (bp->b_npages) { 1586 object = bp->b_pages[0]->object; 1587 VM_OBJECT_WLOCK(object); 1588 } 1589 1590 /* 1591 * cleanup pages. If an error occurs writing to swap, we are in 1592 * very serious trouble. If it happens to be a disk error, though, 1593 * we may be able to recover by reassigning the swap later on. So 1594 * in this case we remove the m->swapblk assignment for the page 1595 * but do not free it in the rlist. The errornous block(s) are thus 1596 * never reallocated as swap. Redirty the page and continue. 1597 */ 1598 for (i = 0; i < bp->b_npages; ++i) { 1599 vm_page_t m = bp->b_pages[i]; 1600 1601 m->oflags &= ~VPO_SWAPINPROG; 1602 if (m->oflags & VPO_SWAPSLEEP) { 1603 m->oflags &= ~VPO_SWAPSLEEP; 1604 wakeup(&object->handle); 1605 } 1606 1607 /* We always have space after I/O, successful or not. */ 1608 vm_page_aflag_set(m, PGA_SWAP_SPACE); 1609 1610 if (bp->b_ioflags & BIO_ERROR) { 1611 /* 1612 * If an error occurs I'd love to throw the swapblk 1613 * away without freeing it back to swapspace, so it 1614 * can never be used again. But I can't from an 1615 * interrupt. 1616 */ 1617 if (bp->b_iocmd == BIO_READ) { 1618 /* 1619 * NOTE: for reads, m->dirty will probably 1620 * be overridden by the original caller of 1621 * getpages so don't play cute tricks here. 1622 */ 1623 vm_page_invalid(m); 1624 } else { 1625 /* 1626 * If a write error occurs, reactivate page 1627 * so it doesn't clog the inactive list, 1628 * then finish the I/O. 1629 */ 1630 MPASS(m->dirty == VM_PAGE_BITS_ALL); 1631 1632 /* PQ_UNSWAPPABLE? */ 1633 vm_page_activate(m); 1634 vm_page_sunbusy(m); 1635 } 1636 } else if (bp->b_iocmd == BIO_READ) { 1637 /* 1638 * NOTE: for reads, m->dirty will probably be 1639 * overridden by the original caller of getpages so 1640 * we cannot set them in order to free the underlying 1641 * swap in a low-swap situation. I don't think we'd 1642 * want to do that anyway, but it was an optimization 1643 * that existed in the old swapper for a time before 1644 * it got ripped out due to precisely this problem. 1645 */ 1646 KASSERT(!pmap_page_is_mapped(m), 1647 ("swp_pager_async_iodone: page %p is mapped", m)); 1648 KASSERT(m->dirty == 0, 1649 ("swp_pager_async_iodone: page %p is dirty", m)); 1650 1651 vm_page_valid(m); 1652 if (i < bp->b_pgbefore || 1653 i >= bp->b_npages - bp->b_pgafter) 1654 vm_page_readahead_finish(m); 1655 } else { 1656 /* 1657 * For write success, clear the dirty 1658 * status, then finish the I/O ( which decrements the 1659 * busy count and possibly wakes waiter's up ). 1660 * A page is only written to swap after a period of 1661 * inactivity. Therefore, we do not expect it to be 1662 * reused. 1663 */ 1664 KASSERT(!pmap_page_is_write_mapped(m), 1665 ("swp_pager_async_iodone: page %p is not write" 1666 " protected", m)); 1667 vm_page_undirty(m); 1668 vm_page_deactivate_noreuse(m); 1669 vm_page_sunbusy(m); 1670 } 1671 } 1672 1673 /* 1674 * adjust pip. NOTE: the original parent may still have its own 1675 * pip refs on the object. 1676 */ 1677 if (object != NULL) { 1678 vm_object_pip_wakeupn(object, bp->b_npages); 1679 VM_OBJECT_WUNLOCK(object); 1680 } 1681 1682 /* 1683 * swapdev_strategy() manually sets b_vp and b_bufobj before calling 1684 * bstrategy(). Set them back to NULL now we're done with it, or we'll 1685 * trigger a KASSERT in relpbuf(). 1686 */ 1687 if (bp->b_vp) { 1688 bp->b_vp = NULL; 1689 bp->b_bufobj = NULL; 1690 } 1691 /* 1692 * release the physical I/O buffer 1693 */ 1694 if (bp->b_flags & B_ASYNC) { 1695 mtx_lock(&swbuf_mtx); 1696 if (++nsw_wcount_async == 1) 1697 wakeup(&nsw_wcount_async); 1698 mtx_unlock(&swbuf_mtx); 1699 } 1700 uma_zfree((bp->b_iocmd == BIO_READ) ? swrbuf_zone : swwbuf_zone, bp); 1701 } 1702 1703 int 1704 swap_pager_nswapdev(void) 1705 { 1706 1707 return (nswapdev); 1708 } 1709 1710 static void 1711 swp_pager_force_dirty(vm_page_t m) 1712 { 1713 1714 vm_page_dirty(m); 1715 #ifdef INVARIANTS 1716 if (!vm_page_wired(m) && m->a.queue == PQ_NONE) 1717 panic("page %p is neither wired nor queued", m); 1718 #endif 1719 vm_page_xunbusy(m); 1720 swap_pager_unswapped(m); 1721 } 1722 1723 static void 1724 swp_pager_force_launder(vm_page_t m) 1725 { 1726 1727 vm_page_dirty(m); 1728 vm_page_launder(m); 1729 vm_page_xunbusy(m); 1730 swap_pager_unswapped(m); 1731 } 1732 1733 /* 1734 * SWP_PAGER_FORCE_PAGEIN() - force swap blocks to be paged in 1735 * 1736 * This routine dissociates pages starting at the given index within an 1737 * object from their backing store, paging them in if they do not reside 1738 * in memory. Pages that are paged in are marked dirty and placed in the 1739 * laundry queue. Pages are marked dirty because they no longer have 1740 * backing store. They are placed in the laundry queue because they have 1741 * not been accessed recently. Otherwise, they would already reside in 1742 * memory. 1743 */ 1744 static void 1745 swp_pager_force_pagein(vm_object_t object, vm_pindex_t pindex, int npages) 1746 { 1747 vm_page_t ma[npages]; 1748 int i, j; 1749 1750 KASSERT(npages > 0, ("%s: No pages", __func__)); 1751 KASSERT(npages <= MAXPHYS / PAGE_SIZE, 1752 ("%s: Too many pages: %d", __func__, npages)); 1753 KASSERT(object->type == OBJT_SWAP, 1754 ("%s: Object not swappable", __func__)); 1755 vm_object_pip_add(object, npages); 1756 vm_page_grab_pages(object, pindex, VM_ALLOC_NORMAL, ma, npages); 1757 for (i = j = 0;; i++) { 1758 /* Count nonresident pages, to page-in all at once. */ 1759 if (i < npages && ma[i]->valid != VM_PAGE_BITS_ALL) 1760 continue; 1761 if (j < i) { 1762 VM_OBJECT_WUNLOCK(object); 1763 /* Page-in nonresident pages. Mark for laundering. */ 1764 if (swap_pager_getpages(object, &ma[j], i - j, NULL, 1765 NULL) != VM_PAGER_OK) 1766 panic("%s: read from swap failed", __func__); 1767 VM_OBJECT_WLOCK(object); 1768 do { 1769 swp_pager_force_launder(ma[j]); 1770 } while (++j < i); 1771 } 1772 if (i == npages) 1773 break; 1774 /* Mark dirty a resident page. */ 1775 swp_pager_force_dirty(ma[j++]); 1776 } 1777 vm_object_pip_wakeupn(object, npages); 1778 } 1779 1780 /* 1781 * swap_pager_swapoff_object: 1782 * 1783 * Page in all of the pages that have been paged out for an object 1784 * to a swap device. 1785 */ 1786 static void 1787 swap_pager_swapoff_object(struct swdevt *sp, vm_object_t object) 1788 { 1789 struct swblk *sb; 1790 vm_pindex_t pi, s_pindex; 1791 daddr_t blk, n_blks, s_blk; 1792 int i; 1793 1794 KASSERT(object->type == OBJT_SWAP, 1795 ("%s: Object not swappable", __func__)); 1796 n_blks = 0; 1797 for (pi = 0; (sb = SWAP_PCTRIE_LOOKUP_GE( 1798 &object->un_pager.swp.swp_blks, pi)) != NULL; ) { 1799 for (i = 0; i < SWAP_META_PAGES; i++) { 1800 blk = sb->d[i]; 1801 if (!swp_pager_isondev(blk, sp)) 1802 blk = SWAPBLK_NONE; 1803 1804 /* 1805 * If there are no blocks/pages accumulated, start a new 1806 * accumulation here. 1807 */ 1808 if (n_blks == 0) { 1809 if (blk != SWAPBLK_NONE) { 1810 s_blk = blk; 1811 s_pindex = sb->p + i; 1812 n_blks = 1; 1813 } 1814 continue; 1815 } 1816 1817 /* 1818 * If the accumulation can be extended without breaking 1819 * the sequence of consecutive blocks and pages that 1820 * swp_pager_force_pagein() depends on, do so. 1821 */ 1822 if (n_blks < MAXPHYS / PAGE_SIZE && 1823 s_blk + n_blks == blk && 1824 s_pindex + n_blks == sb->p + i) { 1825 ++n_blks; 1826 continue; 1827 } 1828 1829 /* 1830 * The sequence of consecutive blocks and pages cannot 1831 * be extended, so page them all in here. Then, 1832 * because doing so involves releasing and reacquiring 1833 * a lock that protects the swap block pctrie, do not 1834 * rely on the current swap block. Break this loop and 1835 * re-fetch the same pindex from the pctrie again. 1836 */ 1837 swp_pager_force_pagein(object, s_pindex, n_blks); 1838 n_blks = 0; 1839 break; 1840 } 1841 if (i == SWAP_META_PAGES) 1842 pi = sb->p + SWAP_META_PAGES; 1843 } 1844 if (n_blks > 0) 1845 swp_pager_force_pagein(object, s_pindex, n_blks); 1846 } 1847 1848 /* 1849 * swap_pager_swapoff: 1850 * 1851 * Page in all of the pages that have been paged out to the 1852 * given device. The corresponding blocks in the bitmap must be 1853 * marked as allocated and the device must be flagged SW_CLOSING. 1854 * There may be no processes swapped out to the device. 1855 * 1856 * This routine may block. 1857 */ 1858 static void 1859 swap_pager_swapoff(struct swdevt *sp) 1860 { 1861 vm_object_t object; 1862 int retries; 1863 1864 sx_assert(&swdev_syscall_lock, SA_XLOCKED); 1865 1866 retries = 0; 1867 full_rescan: 1868 mtx_lock(&vm_object_list_mtx); 1869 TAILQ_FOREACH(object, &vm_object_list, object_list) { 1870 if (object->type != OBJT_SWAP) 1871 continue; 1872 mtx_unlock(&vm_object_list_mtx); 1873 /* Depends on type-stability. */ 1874 VM_OBJECT_WLOCK(object); 1875 1876 /* 1877 * Dead objects are eventually terminated on their own. 1878 */ 1879 if ((object->flags & OBJ_DEAD) != 0) 1880 goto next_obj; 1881 1882 /* 1883 * Sync with fences placed after pctrie 1884 * initialization. We must not access pctrie below 1885 * unless we checked that our object is swap and not 1886 * dead. 1887 */ 1888 atomic_thread_fence_acq(); 1889 if (object->type != OBJT_SWAP) 1890 goto next_obj; 1891 1892 swap_pager_swapoff_object(sp, object); 1893 next_obj: 1894 VM_OBJECT_WUNLOCK(object); 1895 mtx_lock(&vm_object_list_mtx); 1896 } 1897 mtx_unlock(&vm_object_list_mtx); 1898 1899 if (sp->sw_used) { 1900 /* 1901 * Objects may be locked or paging to the device being 1902 * removed, so we will miss their pages and need to 1903 * make another pass. We have marked this device as 1904 * SW_CLOSING, so the activity should finish soon. 1905 */ 1906 retries++; 1907 if (retries > 100) { 1908 panic("swapoff: failed to locate %d swap blocks", 1909 sp->sw_used); 1910 } 1911 pause("swpoff", hz / 20); 1912 goto full_rescan; 1913 } 1914 EVENTHANDLER_INVOKE(swapoff, sp); 1915 } 1916 1917 /************************************************************************ 1918 * SWAP META DATA * 1919 ************************************************************************ 1920 * 1921 * These routines manipulate the swap metadata stored in the 1922 * OBJT_SWAP object. 1923 * 1924 * Swap metadata is implemented with a global hash and not directly 1925 * linked into the object. Instead the object simply contains 1926 * appropriate tracking counters. 1927 */ 1928 1929 /* 1930 * SWP_PAGER_SWBLK_EMPTY() - is a range of blocks free? 1931 */ 1932 static bool 1933 swp_pager_swblk_empty(struct swblk *sb, int start, int limit) 1934 { 1935 int i; 1936 1937 MPASS(0 <= start && start <= limit && limit <= SWAP_META_PAGES); 1938 for (i = start; i < limit; i++) { 1939 if (sb->d[i] != SWAPBLK_NONE) 1940 return (false); 1941 } 1942 return (true); 1943 } 1944 1945 /* 1946 * SWP_PAGER_FREE_EMPTY_SWBLK() - frees if a block is free 1947 * 1948 * Nothing is done if the block is still in use. 1949 */ 1950 static void 1951 swp_pager_free_empty_swblk(vm_object_t object, struct swblk *sb) 1952 { 1953 1954 if (swp_pager_swblk_empty(sb, 0, SWAP_META_PAGES)) { 1955 SWAP_PCTRIE_REMOVE(&object->un_pager.swp.swp_blks, sb->p); 1956 uma_zfree(swblk_zone, sb); 1957 } 1958 } 1959 1960 /* 1961 * SWP_PAGER_META_BUILD() - add swap block to swap meta data for object 1962 * 1963 * We first convert the object to a swap object if it is a default 1964 * object. 1965 * 1966 * The specified swapblk is added to the object's swap metadata. If 1967 * the swapblk is not valid, it is freed instead. Any previously 1968 * assigned swapblk is returned. 1969 */ 1970 static daddr_t 1971 swp_pager_meta_build(vm_object_t object, vm_pindex_t pindex, daddr_t swapblk) 1972 { 1973 static volatile int swblk_zone_exhausted, swpctrie_zone_exhausted; 1974 struct swblk *sb, *sb1; 1975 vm_pindex_t modpi, rdpi; 1976 daddr_t prev_swapblk; 1977 int error, i; 1978 1979 VM_OBJECT_ASSERT_WLOCKED(object); 1980 1981 /* 1982 * Convert default object to swap object if necessary 1983 */ 1984 if (object->type != OBJT_SWAP) { 1985 pctrie_init(&object->un_pager.swp.swp_blks); 1986 1987 /* 1988 * Ensure that swap_pager_swapoff()'s iteration over 1989 * object_list does not see a garbage pctrie. 1990 */ 1991 atomic_thread_fence_rel(); 1992 1993 object->type = OBJT_SWAP; 1994 object->un_pager.swp.writemappings = 0; 1995 KASSERT((object->flags & OBJ_ANON) != 0 || 1996 object->handle == NULL, 1997 ("default pager %p with handle %p", 1998 object, object->handle)); 1999 } 2000 2001 rdpi = rounddown(pindex, SWAP_META_PAGES); 2002 sb = SWAP_PCTRIE_LOOKUP(&object->un_pager.swp.swp_blks, rdpi); 2003 if (sb == NULL) { 2004 if (swapblk == SWAPBLK_NONE) 2005 return (SWAPBLK_NONE); 2006 for (;;) { 2007 sb = uma_zalloc(swblk_zone, M_NOWAIT | (curproc == 2008 pageproc ? M_USE_RESERVE : 0)); 2009 if (sb != NULL) { 2010 sb->p = rdpi; 2011 for (i = 0; i < SWAP_META_PAGES; i++) 2012 sb->d[i] = SWAPBLK_NONE; 2013 if (atomic_cmpset_int(&swblk_zone_exhausted, 2014 1, 0)) 2015 printf("swblk zone ok\n"); 2016 break; 2017 } 2018 VM_OBJECT_WUNLOCK(object); 2019 if (uma_zone_exhausted(swblk_zone)) { 2020 if (atomic_cmpset_int(&swblk_zone_exhausted, 2021 0, 1)) 2022 printf("swap blk zone exhausted, " 2023 "increase kern.maxswzone\n"); 2024 vm_pageout_oom(VM_OOM_SWAPZ); 2025 pause("swzonxb", 10); 2026 } else 2027 uma_zwait(swblk_zone); 2028 VM_OBJECT_WLOCK(object); 2029 sb = SWAP_PCTRIE_LOOKUP(&object->un_pager.swp.swp_blks, 2030 rdpi); 2031 if (sb != NULL) 2032 /* 2033 * Somebody swapped out a nearby page, 2034 * allocating swblk at the rdpi index, 2035 * while we dropped the object lock. 2036 */ 2037 goto allocated; 2038 } 2039 for (;;) { 2040 error = SWAP_PCTRIE_INSERT( 2041 &object->un_pager.swp.swp_blks, sb); 2042 if (error == 0) { 2043 if (atomic_cmpset_int(&swpctrie_zone_exhausted, 2044 1, 0)) 2045 printf("swpctrie zone ok\n"); 2046 break; 2047 } 2048 VM_OBJECT_WUNLOCK(object); 2049 if (uma_zone_exhausted(swpctrie_zone)) { 2050 if (atomic_cmpset_int(&swpctrie_zone_exhausted, 2051 0, 1)) 2052 printf("swap pctrie zone exhausted, " 2053 "increase kern.maxswzone\n"); 2054 vm_pageout_oom(VM_OOM_SWAPZ); 2055 pause("swzonxp", 10); 2056 } else 2057 uma_zwait(swpctrie_zone); 2058 VM_OBJECT_WLOCK(object); 2059 sb1 = SWAP_PCTRIE_LOOKUP(&object->un_pager.swp.swp_blks, 2060 rdpi); 2061 if (sb1 != NULL) { 2062 uma_zfree(swblk_zone, sb); 2063 sb = sb1; 2064 goto allocated; 2065 } 2066 } 2067 } 2068 allocated: 2069 MPASS(sb->p == rdpi); 2070 2071 modpi = pindex % SWAP_META_PAGES; 2072 /* Return prior contents of metadata. */ 2073 prev_swapblk = sb->d[modpi]; 2074 /* Enter block into metadata. */ 2075 sb->d[modpi] = swapblk; 2076 2077 /* 2078 * Free the swblk if we end up with the empty page run. 2079 */ 2080 if (swapblk == SWAPBLK_NONE) 2081 swp_pager_free_empty_swblk(object, sb); 2082 return (prev_swapblk); 2083 } 2084 2085 /* 2086 * SWP_PAGER_META_TRANSFER() - free a range of blocks in the srcobject's swap 2087 * metadata, or transfer it into dstobject. 2088 * 2089 * This routine will free swap metadata structures as they are cleaned 2090 * out. 2091 */ 2092 static void 2093 swp_pager_meta_transfer(vm_object_t srcobject, vm_object_t dstobject, 2094 vm_pindex_t pindex, vm_pindex_t count) 2095 { 2096 struct swblk *sb; 2097 daddr_t n_free, s_free; 2098 vm_pindex_t offset, last; 2099 int i, limit, start; 2100 2101 VM_OBJECT_ASSERT_WLOCKED(srcobject); 2102 if (srcobject->type != OBJT_SWAP || count == 0) 2103 return; 2104 2105 swp_pager_init_freerange(&s_free, &n_free); 2106 offset = pindex; 2107 last = pindex + count; 2108 for (;;) { 2109 sb = SWAP_PCTRIE_LOOKUP_GE(&srcobject->un_pager.swp.swp_blks, 2110 rounddown(pindex, SWAP_META_PAGES)); 2111 if (sb == NULL || sb->p >= last) 2112 break; 2113 start = pindex > sb->p ? pindex - sb->p : 0; 2114 limit = last - sb->p < SWAP_META_PAGES ? last - sb->p : 2115 SWAP_META_PAGES; 2116 for (i = start; i < limit; i++) { 2117 if (sb->d[i] == SWAPBLK_NONE) 2118 continue; 2119 if (dstobject == NULL || 2120 !swp_pager_xfer_source(srcobject, dstobject, 2121 sb->p + i - offset, sb->d[i])) { 2122 swp_pager_update_freerange(&s_free, &n_free, 2123 sb->d[i]); 2124 } 2125 sb->d[i] = SWAPBLK_NONE; 2126 } 2127 pindex = sb->p + SWAP_META_PAGES; 2128 if (swp_pager_swblk_empty(sb, 0, start) && 2129 swp_pager_swblk_empty(sb, limit, SWAP_META_PAGES)) { 2130 SWAP_PCTRIE_REMOVE(&srcobject->un_pager.swp.swp_blks, 2131 sb->p); 2132 uma_zfree(swblk_zone, sb); 2133 } 2134 } 2135 swp_pager_freeswapspace(s_free, n_free); 2136 } 2137 2138 /* 2139 * SWP_PAGER_META_FREE() - free a range of blocks in the object's swap metadata 2140 * 2141 * The requested range of blocks is freed, with any associated swap 2142 * returned to the swap bitmap. 2143 * 2144 * This routine will free swap metadata structures as they are cleaned 2145 * out. This routine does *NOT* operate on swap metadata associated 2146 * with resident pages. 2147 */ 2148 static void 2149 swp_pager_meta_free(vm_object_t object, vm_pindex_t pindex, vm_pindex_t count) 2150 { 2151 swp_pager_meta_transfer(object, NULL, pindex, count); 2152 } 2153 2154 /* 2155 * SWP_PAGER_META_FREE_ALL() - destroy all swap metadata associated with object 2156 * 2157 * This routine locates and destroys all swap metadata associated with 2158 * an object. 2159 */ 2160 static void 2161 swp_pager_meta_free_all(vm_object_t object) 2162 { 2163 struct swblk *sb; 2164 daddr_t n_free, s_free; 2165 vm_pindex_t pindex; 2166 int i; 2167 2168 VM_OBJECT_ASSERT_WLOCKED(object); 2169 if (object->type != OBJT_SWAP) 2170 return; 2171 2172 swp_pager_init_freerange(&s_free, &n_free); 2173 for (pindex = 0; (sb = SWAP_PCTRIE_LOOKUP_GE( 2174 &object->un_pager.swp.swp_blks, pindex)) != NULL;) { 2175 pindex = sb->p + SWAP_META_PAGES; 2176 for (i = 0; i < SWAP_META_PAGES; i++) { 2177 if (sb->d[i] == SWAPBLK_NONE) 2178 continue; 2179 swp_pager_update_freerange(&s_free, &n_free, sb->d[i]); 2180 } 2181 SWAP_PCTRIE_REMOVE(&object->un_pager.swp.swp_blks, sb->p); 2182 uma_zfree(swblk_zone, sb); 2183 } 2184 swp_pager_freeswapspace(s_free, n_free); 2185 } 2186 2187 /* 2188 * SWP_PAGER_METACTL() - misc control of swap meta data. 2189 * 2190 * This routine is capable of looking up, or removing swapblk 2191 * assignments in the swap meta data. It returns the swapblk being 2192 * looked-up, popped, or SWAPBLK_NONE if the block was invalid. 2193 * 2194 * When acting on a busy resident page and paging is in progress, we 2195 * have to wait until paging is complete but otherwise can act on the 2196 * busy page. 2197 */ 2198 static daddr_t 2199 swp_pager_meta_lookup(vm_object_t object, vm_pindex_t pindex) 2200 { 2201 struct swblk *sb; 2202 2203 VM_OBJECT_ASSERT_LOCKED(object); 2204 2205 /* 2206 * The meta data only exists if the object is OBJT_SWAP 2207 * and even then might not be allocated yet. 2208 */ 2209 KASSERT(object->type == OBJT_SWAP, 2210 ("Lookup object not swappable")); 2211 2212 sb = SWAP_PCTRIE_LOOKUP(&object->un_pager.swp.swp_blks, 2213 rounddown(pindex, SWAP_META_PAGES)); 2214 if (sb == NULL) 2215 return (SWAPBLK_NONE); 2216 return (sb->d[pindex % SWAP_META_PAGES]); 2217 } 2218 2219 /* 2220 * Returns the least page index which is greater than or equal to the 2221 * parameter pindex and for which there is a swap block allocated. 2222 * Returns object's size if the object's type is not swap or if there 2223 * are no allocated swap blocks for the object after the requested 2224 * pindex. 2225 */ 2226 vm_pindex_t 2227 swap_pager_find_least(vm_object_t object, vm_pindex_t pindex) 2228 { 2229 struct swblk *sb; 2230 int i; 2231 2232 VM_OBJECT_ASSERT_LOCKED(object); 2233 if (object->type != OBJT_SWAP) 2234 return (object->size); 2235 2236 sb = SWAP_PCTRIE_LOOKUP_GE(&object->un_pager.swp.swp_blks, 2237 rounddown(pindex, SWAP_META_PAGES)); 2238 if (sb == NULL) 2239 return (object->size); 2240 if (sb->p < pindex) { 2241 for (i = pindex % SWAP_META_PAGES; i < SWAP_META_PAGES; i++) { 2242 if (sb->d[i] != SWAPBLK_NONE) 2243 return (sb->p + i); 2244 } 2245 sb = SWAP_PCTRIE_LOOKUP_GE(&object->un_pager.swp.swp_blks, 2246 roundup(pindex, SWAP_META_PAGES)); 2247 if (sb == NULL) 2248 return (object->size); 2249 } 2250 for (i = 0; i < SWAP_META_PAGES; i++) { 2251 if (sb->d[i] != SWAPBLK_NONE) 2252 return (sb->p + i); 2253 } 2254 2255 /* 2256 * We get here if a swblk is present in the trie but it 2257 * doesn't map any blocks. 2258 */ 2259 MPASS(0); 2260 return (object->size); 2261 } 2262 2263 /* 2264 * System call swapon(name) enables swapping on device name, 2265 * which must be in the swdevsw. Return EBUSY 2266 * if already swapping on this device. 2267 */ 2268 #ifndef _SYS_SYSPROTO_H_ 2269 struct swapon_args { 2270 char *name; 2271 }; 2272 #endif 2273 2274 /* 2275 * MPSAFE 2276 */ 2277 /* ARGSUSED */ 2278 int 2279 sys_swapon(struct thread *td, struct swapon_args *uap) 2280 { 2281 struct vattr attr; 2282 struct vnode *vp; 2283 struct nameidata nd; 2284 int error; 2285 2286 error = priv_check(td, PRIV_SWAPON); 2287 if (error) 2288 return (error); 2289 2290 sx_xlock(&swdev_syscall_lock); 2291 2292 /* 2293 * Swap metadata may not fit in the KVM if we have physical 2294 * memory of >1GB. 2295 */ 2296 if (swblk_zone == NULL) { 2297 error = ENOMEM; 2298 goto done; 2299 } 2300 2301 NDINIT(&nd, LOOKUP, ISOPEN | FOLLOW | AUDITVNODE1, UIO_USERSPACE, 2302 uap->name, td); 2303 error = namei(&nd); 2304 if (error) 2305 goto done; 2306 2307 NDFREE(&nd, NDF_ONLY_PNBUF); 2308 vp = nd.ni_vp; 2309 2310 if (vn_isdisk(vp, &error)) { 2311 error = swapongeom(vp); 2312 } else if (vp->v_type == VREG && 2313 (vp->v_mount->mnt_vfc->vfc_flags & VFCF_NETWORK) != 0 && 2314 (error = VOP_GETATTR(vp, &attr, td->td_ucred)) == 0) { 2315 /* 2316 * Allow direct swapping to NFS regular files in the same 2317 * way that nfs_mountroot() sets up diskless swapping. 2318 */ 2319 error = swaponvp(td, vp, attr.va_size / DEV_BSIZE); 2320 } 2321 2322 if (error) 2323 vrele(vp); 2324 done: 2325 sx_xunlock(&swdev_syscall_lock); 2326 return (error); 2327 } 2328 2329 /* 2330 * Check that the total amount of swap currently configured does not 2331 * exceed half the theoretical maximum. If it does, print a warning 2332 * message. 2333 */ 2334 static void 2335 swapon_check_swzone(void) 2336 { 2337 2338 /* recommend using no more than half that amount */ 2339 if (swap_total > swap_maxpages / 2) { 2340 printf("warning: total configured swap (%lu pages) " 2341 "exceeds maximum recommended amount (%lu pages).\n", 2342 swap_total, swap_maxpages / 2); 2343 printf("warning: increase kern.maxswzone " 2344 "or reduce amount of swap.\n"); 2345 } 2346 } 2347 2348 static void 2349 swaponsomething(struct vnode *vp, void *id, u_long nblks, 2350 sw_strategy_t *strategy, sw_close_t *close, dev_t dev, int flags) 2351 { 2352 struct swdevt *sp, *tsp; 2353 swblk_t dvbase; 2354 u_long mblocks; 2355 2356 /* 2357 * nblks is in DEV_BSIZE'd chunks, convert to PAGE_SIZE'd chunks. 2358 * First chop nblks off to page-align it, then convert. 2359 * 2360 * sw->sw_nblks is in page-sized chunks now too. 2361 */ 2362 nblks &= ~(ctodb(1) - 1); 2363 nblks = dbtoc(nblks); 2364 2365 /* 2366 * If we go beyond this, we get overflows in the radix 2367 * tree bitmap code. 2368 */ 2369 mblocks = 0x40000000 / BLIST_META_RADIX; 2370 if (nblks > mblocks) { 2371 printf( 2372 "WARNING: reducing swap size to maximum of %luMB per unit\n", 2373 mblocks / 1024 / 1024 * PAGE_SIZE); 2374 nblks = mblocks; 2375 } 2376 2377 sp = malloc(sizeof *sp, M_VMPGDATA, M_WAITOK | M_ZERO); 2378 sp->sw_vp = vp; 2379 sp->sw_id = id; 2380 sp->sw_dev = dev; 2381 sp->sw_nblks = nblks; 2382 sp->sw_used = 0; 2383 sp->sw_strategy = strategy; 2384 sp->sw_close = close; 2385 sp->sw_flags = flags; 2386 2387 sp->sw_blist = blist_create(nblks, M_WAITOK); 2388 /* 2389 * Do not free the first blocks in order to avoid overwriting 2390 * any bsd label at the front of the partition 2391 */ 2392 blist_free(sp->sw_blist, howmany(BBSIZE, PAGE_SIZE), 2393 nblks - howmany(BBSIZE, PAGE_SIZE)); 2394 2395 dvbase = 0; 2396 mtx_lock(&sw_dev_mtx); 2397 TAILQ_FOREACH(tsp, &swtailq, sw_list) { 2398 if (tsp->sw_end >= dvbase) { 2399 /* 2400 * We put one uncovered page between the devices 2401 * in order to definitively prevent any cross-device 2402 * I/O requests 2403 */ 2404 dvbase = tsp->sw_end + 1; 2405 } 2406 } 2407 sp->sw_first = dvbase; 2408 sp->sw_end = dvbase + nblks; 2409 TAILQ_INSERT_TAIL(&swtailq, sp, sw_list); 2410 nswapdev++; 2411 swap_pager_avail += nblks - howmany(BBSIZE, PAGE_SIZE); 2412 swap_total += nblks; 2413 swapon_check_swzone(); 2414 swp_sizecheck(); 2415 mtx_unlock(&sw_dev_mtx); 2416 EVENTHANDLER_INVOKE(swapon, sp); 2417 } 2418 2419 /* 2420 * SYSCALL: swapoff(devname) 2421 * 2422 * Disable swapping on the given device. 2423 * 2424 * XXX: Badly designed system call: it should use a device index 2425 * rather than filename as specification. We keep sw_vp around 2426 * only to make this work. 2427 */ 2428 #ifndef _SYS_SYSPROTO_H_ 2429 struct swapoff_args { 2430 char *name; 2431 }; 2432 #endif 2433 2434 /* 2435 * MPSAFE 2436 */ 2437 /* ARGSUSED */ 2438 int 2439 sys_swapoff(struct thread *td, struct swapoff_args *uap) 2440 { 2441 struct vnode *vp; 2442 struct nameidata nd; 2443 struct swdevt *sp; 2444 int error; 2445 2446 error = priv_check(td, PRIV_SWAPOFF); 2447 if (error) 2448 return (error); 2449 2450 sx_xlock(&swdev_syscall_lock); 2451 2452 NDINIT(&nd, LOOKUP, FOLLOW | AUDITVNODE1, UIO_USERSPACE, uap->name, 2453 td); 2454 error = namei(&nd); 2455 if (error) 2456 goto done; 2457 NDFREE(&nd, NDF_ONLY_PNBUF); 2458 vp = nd.ni_vp; 2459 2460 mtx_lock(&sw_dev_mtx); 2461 TAILQ_FOREACH(sp, &swtailq, sw_list) { 2462 if (sp->sw_vp == vp) 2463 break; 2464 } 2465 mtx_unlock(&sw_dev_mtx); 2466 if (sp == NULL) { 2467 error = EINVAL; 2468 goto done; 2469 } 2470 error = swapoff_one(sp, td->td_ucred); 2471 done: 2472 sx_xunlock(&swdev_syscall_lock); 2473 return (error); 2474 } 2475 2476 static int 2477 swapoff_one(struct swdevt *sp, struct ucred *cred) 2478 { 2479 u_long nblks; 2480 #ifdef MAC 2481 int error; 2482 #endif 2483 2484 sx_assert(&swdev_syscall_lock, SA_XLOCKED); 2485 #ifdef MAC 2486 (void) vn_lock(sp->sw_vp, LK_EXCLUSIVE | LK_RETRY); 2487 error = mac_system_check_swapoff(cred, sp->sw_vp); 2488 (void) VOP_UNLOCK(sp->sw_vp); 2489 if (error != 0) 2490 return (error); 2491 #endif 2492 nblks = sp->sw_nblks; 2493 2494 /* 2495 * We can turn off this swap device safely only if the 2496 * available virtual memory in the system will fit the amount 2497 * of data we will have to page back in, plus an epsilon so 2498 * the system doesn't become critically low on swap space. 2499 */ 2500 if (vm_free_count() + swap_pager_avail < nblks + nswap_lowat) 2501 return (ENOMEM); 2502 2503 /* 2504 * Prevent further allocations on this device. 2505 */ 2506 mtx_lock(&sw_dev_mtx); 2507 sp->sw_flags |= SW_CLOSING; 2508 swap_pager_avail -= blist_fill(sp->sw_blist, 0, nblks); 2509 swap_total -= nblks; 2510 mtx_unlock(&sw_dev_mtx); 2511 2512 /* 2513 * Page in the contents of the device and close it. 2514 */ 2515 swap_pager_swapoff(sp); 2516 2517 sp->sw_close(curthread, sp); 2518 mtx_lock(&sw_dev_mtx); 2519 sp->sw_id = NULL; 2520 TAILQ_REMOVE(&swtailq, sp, sw_list); 2521 nswapdev--; 2522 if (nswapdev == 0) { 2523 swap_pager_full = 2; 2524 swap_pager_almost_full = 1; 2525 } 2526 if (swdevhd == sp) 2527 swdevhd = NULL; 2528 mtx_unlock(&sw_dev_mtx); 2529 blist_destroy(sp->sw_blist); 2530 free(sp, M_VMPGDATA); 2531 return (0); 2532 } 2533 2534 void 2535 swapoff_all(void) 2536 { 2537 struct swdevt *sp, *spt; 2538 const char *devname; 2539 int error; 2540 2541 sx_xlock(&swdev_syscall_lock); 2542 2543 mtx_lock(&sw_dev_mtx); 2544 TAILQ_FOREACH_SAFE(sp, &swtailq, sw_list, spt) { 2545 mtx_unlock(&sw_dev_mtx); 2546 if (vn_isdisk(sp->sw_vp, NULL)) 2547 devname = devtoname(sp->sw_vp->v_rdev); 2548 else 2549 devname = "[file]"; 2550 error = swapoff_one(sp, thread0.td_ucred); 2551 if (error != 0) { 2552 printf("Cannot remove swap device %s (error=%d), " 2553 "skipping.\n", devname, error); 2554 } else if (bootverbose) { 2555 printf("Swap device %s removed.\n", devname); 2556 } 2557 mtx_lock(&sw_dev_mtx); 2558 } 2559 mtx_unlock(&sw_dev_mtx); 2560 2561 sx_xunlock(&swdev_syscall_lock); 2562 } 2563 2564 void 2565 swap_pager_status(int *total, int *used) 2566 { 2567 struct swdevt *sp; 2568 2569 *total = 0; 2570 *used = 0; 2571 mtx_lock(&sw_dev_mtx); 2572 TAILQ_FOREACH(sp, &swtailq, sw_list) { 2573 *total += sp->sw_nblks; 2574 *used += sp->sw_used; 2575 } 2576 mtx_unlock(&sw_dev_mtx); 2577 } 2578 2579 int 2580 swap_dev_info(int name, struct xswdev *xs, char *devname, size_t len) 2581 { 2582 struct swdevt *sp; 2583 const char *tmp_devname; 2584 int error, n; 2585 2586 n = 0; 2587 error = ENOENT; 2588 mtx_lock(&sw_dev_mtx); 2589 TAILQ_FOREACH(sp, &swtailq, sw_list) { 2590 if (n != name) { 2591 n++; 2592 continue; 2593 } 2594 xs->xsw_version = XSWDEV_VERSION; 2595 xs->xsw_dev = sp->sw_dev; 2596 xs->xsw_flags = sp->sw_flags; 2597 xs->xsw_nblks = sp->sw_nblks; 2598 xs->xsw_used = sp->sw_used; 2599 if (devname != NULL) { 2600 if (vn_isdisk(sp->sw_vp, NULL)) 2601 tmp_devname = devtoname(sp->sw_vp->v_rdev); 2602 else 2603 tmp_devname = "[file]"; 2604 strncpy(devname, tmp_devname, len); 2605 } 2606 error = 0; 2607 break; 2608 } 2609 mtx_unlock(&sw_dev_mtx); 2610 return (error); 2611 } 2612 2613 #if defined(COMPAT_FREEBSD11) 2614 #define XSWDEV_VERSION_11 1 2615 struct xswdev11 { 2616 u_int xsw_version; 2617 uint32_t xsw_dev; 2618 int xsw_flags; 2619 int xsw_nblks; 2620 int xsw_used; 2621 }; 2622 #endif 2623 2624 #if defined(__amd64__) && defined(COMPAT_FREEBSD32) 2625 struct xswdev32 { 2626 u_int xsw_version; 2627 u_int xsw_dev1, xsw_dev2; 2628 int xsw_flags; 2629 int xsw_nblks; 2630 int xsw_used; 2631 }; 2632 #endif 2633 2634 static int 2635 sysctl_vm_swap_info(SYSCTL_HANDLER_ARGS) 2636 { 2637 struct xswdev xs; 2638 #if defined(__amd64__) && defined(COMPAT_FREEBSD32) 2639 struct xswdev32 xs32; 2640 #endif 2641 #if defined(COMPAT_FREEBSD11) 2642 struct xswdev11 xs11; 2643 #endif 2644 int error; 2645 2646 if (arg2 != 1) /* name length */ 2647 return (EINVAL); 2648 error = swap_dev_info(*(int *)arg1, &xs, NULL, 0); 2649 if (error != 0) 2650 return (error); 2651 #if defined(__amd64__) && defined(COMPAT_FREEBSD32) 2652 if (req->oldlen == sizeof(xs32)) { 2653 xs32.xsw_version = XSWDEV_VERSION; 2654 xs32.xsw_dev1 = xs.xsw_dev; 2655 xs32.xsw_dev2 = xs.xsw_dev >> 32; 2656 xs32.xsw_flags = xs.xsw_flags; 2657 xs32.xsw_nblks = xs.xsw_nblks; 2658 xs32.xsw_used = xs.xsw_used; 2659 error = SYSCTL_OUT(req, &xs32, sizeof(xs32)); 2660 return (error); 2661 } 2662 #endif 2663 #if defined(COMPAT_FREEBSD11) 2664 if (req->oldlen == sizeof(xs11)) { 2665 xs11.xsw_version = XSWDEV_VERSION_11; 2666 xs11.xsw_dev = xs.xsw_dev; /* truncation */ 2667 xs11.xsw_flags = xs.xsw_flags; 2668 xs11.xsw_nblks = xs.xsw_nblks; 2669 xs11.xsw_used = xs.xsw_used; 2670 error = SYSCTL_OUT(req, &xs11, sizeof(xs11)); 2671 return (error); 2672 } 2673 #endif 2674 error = SYSCTL_OUT(req, &xs, sizeof(xs)); 2675 return (error); 2676 } 2677 2678 SYSCTL_INT(_vm, OID_AUTO, nswapdev, CTLFLAG_RD, &nswapdev, 0, 2679 "Number of swap devices"); 2680 SYSCTL_NODE(_vm, OID_AUTO, swap_info, CTLFLAG_RD | CTLFLAG_MPSAFE, 2681 sysctl_vm_swap_info, 2682 "Swap statistics by device"); 2683 2684 /* 2685 * Count the approximate swap usage in pages for a vmspace. The 2686 * shadowed or not yet copied on write swap blocks are not accounted. 2687 * The map must be locked. 2688 */ 2689 long 2690 vmspace_swap_count(struct vmspace *vmspace) 2691 { 2692 vm_map_t map; 2693 vm_map_entry_t cur; 2694 vm_object_t object; 2695 struct swblk *sb; 2696 vm_pindex_t e, pi; 2697 long count; 2698 int i; 2699 2700 map = &vmspace->vm_map; 2701 count = 0; 2702 2703 VM_MAP_ENTRY_FOREACH(cur, map) { 2704 if ((cur->eflags & MAP_ENTRY_IS_SUB_MAP) != 0) 2705 continue; 2706 object = cur->object.vm_object; 2707 if (object == NULL || object->type != OBJT_SWAP) 2708 continue; 2709 VM_OBJECT_RLOCK(object); 2710 if (object->type != OBJT_SWAP) 2711 goto unlock; 2712 pi = OFF_TO_IDX(cur->offset); 2713 e = pi + OFF_TO_IDX(cur->end - cur->start); 2714 for (;; pi = sb->p + SWAP_META_PAGES) { 2715 sb = SWAP_PCTRIE_LOOKUP_GE( 2716 &object->un_pager.swp.swp_blks, pi); 2717 if (sb == NULL || sb->p >= e) 2718 break; 2719 for (i = 0; i < SWAP_META_PAGES; i++) { 2720 if (sb->p + i < e && 2721 sb->d[i] != SWAPBLK_NONE) 2722 count++; 2723 } 2724 } 2725 unlock: 2726 VM_OBJECT_RUNLOCK(object); 2727 } 2728 return (count); 2729 } 2730 2731 /* 2732 * GEOM backend 2733 * 2734 * Swapping onto disk devices. 2735 * 2736 */ 2737 2738 static g_orphan_t swapgeom_orphan; 2739 2740 static struct g_class g_swap_class = { 2741 .name = "SWAP", 2742 .version = G_VERSION, 2743 .orphan = swapgeom_orphan, 2744 }; 2745 2746 DECLARE_GEOM_CLASS(g_swap_class, g_class); 2747 2748 2749 static void 2750 swapgeom_close_ev(void *arg, int flags) 2751 { 2752 struct g_consumer *cp; 2753 2754 cp = arg; 2755 g_access(cp, -1, -1, 0); 2756 g_detach(cp); 2757 g_destroy_consumer(cp); 2758 } 2759 2760 /* 2761 * Add a reference to the g_consumer for an inflight transaction. 2762 */ 2763 static void 2764 swapgeom_acquire(struct g_consumer *cp) 2765 { 2766 2767 mtx_assert(&sw_dev_mtx, MA_OWNED); 2768 cp->index++; 2769 } 2770 2771 /* 2772 * Remove a reference from the g_consumer. Post a close event if all 2773 * references go away, since the function might be called from the 2774 * biodone context. 2775 */ 2776 static void 2777 swapgeom_release(struct g_consumer *cp, struct swdevt *sp) 2778 { 2779 2780 mtx_assert(&sw_dev_mtx, MA_OWNED); 2781 cp->index--; 2782 if (cp->index == 0) { 2783 if (g_post_event(swapgeom_close_ev, cp, M_NOWAIT, NULL) == 0) 2784 sp->sw_id = NULL; 2785 } 2786 } 2787 2788 static void 2789 swapgeom_done(struct bio *bp2) 2790 { 2791 struct swdevt *sp; 2792 struct buf *bp; 2793 struct g_consumer *cp; 2794 2795 bp = bp2->bio_caller2; 2796 cp = bp2->bio_from; 2797 bp->b_ioflags = bp2->bio_flags; 2798 if (bp2->bio_error) 2799 bp->b_ioflags |= BIO_ERROR; 2800 bp->b_resid = bp->b_bcount - bp2->bio_completed; 2801 bp->b_error = bp2->bio_error; 2802 bp->b_caller1 = NULL; 2803 bufdone(bp); 2804 sp = bp2->bio_caller1; 2805 mtx_lock(&sw_dev_mtx); 2806 swapgeom_release(cp, sp); 2807 mtx_unlock(&sw_dev_mtx); 2808 g_destroy_bio(bp2); 2809 } 2810 2811 static void 2812 swapgeom_strategy(struct buf *bp, struct swdevt *sp) 2813 { 2814 struct bio *bio; 2815 struct g_consumer *cp; 2816 2817 mtx_lock(&sw_dev_mtx); 2818 cp = sp->sw_id; 2819 if (cp == NULL) { 2820 mtx_unlock(&sw_dev_mtx); 2821 bp->b_error = ENXIO; 2822 bp->b_ioflags |= BIO_ERROR; 2823 bufdone(bp); 2824 return; 2825 } 2826 swapgeom_acquire(cp); 2827 mtx_unlock(&sw_dev_mtx); 2828 if (bp->b_iocmd == BIO_WRITE) 2829 bio = g_new_bio(); 2830 else 2831 bio = g_alloc_bio(); 2832 if (bio == NULL) { 2833 mtx_lock(&sw_dev_mtx); 2834 swapgeom_release(cp, sp); 2835 mtx_unlock(&sw_dev_mtx); 2836 bp->b_error = ENOMEM; 2837 bp->b_ioflags |= BIO_ERROR; 2838 printf("swap_pager: cannot allocate bio\n"); 2839 bufdone(bp); 2840 return; 2841 } 2842 2843 bp->b_caller1 = bio; 2844 bio->bio_caller1 = sp; 2845 bio->bio_caller2 = bp; 2846 bio->bio_cmd = bp->b_iocmd; 2847 bio->bio_offset = (bp->b_blkno - sp->sw_first) * PAGE_SIZE; 2848 bio->bio_length = bp->b_bcount; 2849 bio->bio_done = swapgeom_done; 2850 if (!buf_mapped(bp)) { 2851 bio->bio_ma = bp->b_pages; 2852 bio->bio_data = unmapped_buf; 2853 bio->bio_ma_offset = (vm_offset_t)bp->b_offset & PAGE_MASK; 2854 bio->bio_ma_n = bp->b_npages; 2855 bio->bio_flags |= BIO_UNMAPPED; 2856 } else { 2857 bio->bio_data = bp->b_data; 2858 bio->bio_ma = NULL; 2859 } 2860 g_io_request(bio, cp); 2861 return; 2862 } 2863 2864 static void 2865 swapgeom_orphan(struct g_consumer *cp) 2866 { 2867 struct swdevt *sp; 2868 int destroy; 2869 2870 mtx_lock(&sw_dev_mtx); 2871 TAILQ_FOREACH(sp, &swtailq, sw_list) { 2872 if (sp->sw_id == cp) { 2873 sp->sw_flags |= SW_CLOSING; 2874 break; 2875 } 2876 } 2877 /* 2878 * Drop reference we were created with. Do directly since we're in a 2879 * special context where we don't have to queue the call to 2880 * swapgeom_close_ev(). 2881 */ 2882 cp->index--; 2883 destroy = ((sp != NULL) && (cp->index == 0)); 2884 if (destroy) 2885 sp->sw_id = NULL; 2886 mtx_unlock(&sw_dev_mtx); 2887 if (destroy) 2888 swapgeom_close_ev(cp, 0); 2889 } 2890 2891 static void 2892 swapgeom_close(struct thread *td, struct swdevt *sw) 2893 { 2894 struct g_consumer *cp; 2895 2896 mtx_lock(&sw_dev_mtx); 2897 cp = sw->sw_id; 2898 sw->sw_id = NULL; 2899 mtx_unlock(&sw_dev_mtx); 2900 2901 /* 2902 * swapgeom_close() may be called from the biodone context, 2903 * where we cannot perform topology changes. Delegate the 2904 * work to the events thread. 2905 */ 2906 if (cp != NULL) 2907 g_waitfor_event(swapgeom_close_ev, cp, M_WAITOK, NULL); 2908 } 2909 2910 static int 2911 swapongeom_locked(struct cdev *dev, struct vnode *vp) 2912 { 2913 struct g_provider *pp; 2914 struct g_consumer *cp; 2915 static struct g_geom *gp; 2916 struct swdevt *sp; 2917 u_long nblks; 2918 int error; 2919 2920 pp = g_dev_getprovider(dev); 2921 if (pp == NULL) 2922 return (ENODEV); 2923 mtx_lock(&sw_dev_mtx); 2924 TAILQ_FOREACH(sp, &swtailq, sw_list) { 2925 cp = sp->sw_id; 2926 if (cp != NULL && cp->provider == pp) { 2927 mtx_unlock(&sw_dev_mtx); 2928 return (EBUSY); 2929 } 2930 } 2931 mtx_unlock(&sw_dev_mtx); 2932 if (gp == NULL) 2933 gp = g_new_geomf(&g_swap_class, "swap"); 2934 cp = g_new_consumer(gp); 2935 cp->index = 1; /* Number of active I/Os, plus one for being active. */ 2936 cp->flags |= G_CF_DIRECT_SEND | G_CF_DIRECT_RECEIVE; 2937 g_attach(cp, pp); 2938 /* 2939 * XXX: Every time you think you can improve the margin for 2940 * footshooting, somebody depends on the ability to do so: 2941 * savecore(8) wants to write to our swapdev so we cannot 2942 * set an exclusive count :-( 2943 */ 2944 error = g_access(cp, 1, 1, 0); 2945 if (error != 0) { 2946 g_detach(cp); 2947 g_destroy_consumer(cp); 2948 return (error); 2949 } 2950 nblks = pp->mediasize / DEV_BSIZE; 2951 swaponsomething(vp, cp, nblks, swapgeom_strategy, 2952 swapgeom_close, dev2udev(dev), 2953 (pp->flags & G_PF_ACCEPT_UNMAPPED) != 0 ? SW_UNMAPPED : 0); 2954 return (0); 2955 } 2956 2957 static int 2958 swapongeom(struct vnode *vp) 2959 { 2960 int error; 2961 2962 vn_lock(vp, LK_EXCLUSIVE | LK_RETRY); 2963 if (vp->v_type != VCHR || VN_IS_DOOMED(vp)) { 2964 error = ENOENT; 2965 } else { 2966 g_topology_lock(); 2967 error = swapongeom_locked(vp->v_rdev, vp); 2968 g_topology_unlock(); 2969 } 2970 VOP_UNLOCK(vp); 2971 return (error); 2972 } 2973 2974 /* 2975 * VNODE backend 2976 * 2977 * This is used mainly for network filesystem (read: probably only tested 2978 * with NFS) swapfiles. 2979 * 2980 */ 2981 2982 static void 2983 swapdev_strategy(struct buf *bp, struct swdevt *sp) 2984 { 2985 struct vnode *vp2; 2986 2987 bp->b_blkno = ctodb(bp->b_blkno - sp->sw_first); 2988 2989 vp2 = sp->sw_id; 2990 vhold(vp2); 2991 if (bp->b_iocmd == BIO_WRITE) { 2992 if (bp->b_bufobj) 2993 bufobj_wdrop(bp->b_bufobj); 2994 bufobj_wref(&vp2->v_bufobj); 2995 } 2996 if (bp->b_bufobj != &vp2->v_bufobj) 2997 bp->b_bufobj = &vp2->v_bufobj; 2998 bp->b_vp = vp2; 2999 bp->b_iooffset = dbtob(bp->b_blkno); 3000 bstrategy(bp); 3001 return; 3002 } 3003 3004 static void 3005 swapdev_close(struct thread *td, struct swdevt *sp) 3006 { 3007 3008 VOP_CLOSE(sp->sw_vp, FREAD | FWRITE, td->td_ucred, td); 3009 vrele(sp->sw_vp); 3010 } 3011 3012 3013 static int 3014 swaponvp(struct thread *td, struct vnode *vp, u_long nblks) 3015 { 3016 struct swdevt *sp; 3017 int error; 3018 3019 if (nblks == 0) 3020 return (ENXIO); 3021 mtx_lock(&sw_dev_mtx); 3022 TAILQ_FOREACH(sp, &swtailq, sw_list) { 3023 if (sp->sw_id == vp) { 3024 mtx_unlock(&sw_dev_mtx); 3025 return (EBUSY); 3026 } 3027 } 3028 mtx_unlock(&sw_dev_mtx); 3029 3030 (void) vn_lock(vp, LK_EXCLUSIVE | LK_RETRY); 3031 #ifdef MAC 3032 error = mac_system_check_swapon(td->td_ucred, vp); 3033 if (error == 0) 3034 #endif 3035 error = VOP_OPEN(vp, FREAD | FWRITE, td->td_ucred, td, NULL); 3036 (void) VOP_UNLOCK(vp); 3037 if (error) 3038 return (error); 3039 3040 swaponsomething(vp, vp, nblks, swapdev_strategy, swapdev_close, 3041 NODEV, 0); 3042 return (0); 3043 } 3044 3045 static int 3046 sysctl_swap_async_max(SYSCTL_HANDLER_ARGS) 3047 { 3048 int error, new, n; 3049 3050 new = nsw_wcount_async_max; 3051 error = sysctl_handle_int(oidp, &new, 0, req); 3052 if (error != 0 || req->newptr == NULL) 3053 return (error); 3054 3055 if (new > nswbuf / 2 || new < 1) 3056 return (EINVAL); 3057 3058 mtx_lock(&swbuf_mtx); 3059 while (nsw_wcount_async_max != new) { 3060 /* 3061 * Adjust difference. If the current async count is too low, 3062 * we will need to sqeeze our update slowly in. Sleep with a 3063 * higher priority than getpbuf() to finish faster. 3064 */ 3065 n = new - nsw_wcount_async_max; 3066 if (nsw_wcount_async + n >= 0) { 3067 nsw_wcount_async += n; 3068 nsw_wcount_async_max += n; 3069 wakeup(&nsw_wcount_async); 3070 } else { 3071 nsw_wcount_async_max -= nsw_wcount_async; 3072 nsw_wcount_async = 0; 3073 msleep(&nsw_wcount_async, &swbuf_mtx, PSWP, 3074 "swpsysctl", 0); 3075 } 3076 } 3077 mtx_unlock(&swbuf_mtx); 3078 3079 return (0); 3080 } 3081 3082 static void 3083 swap_pager_update_writecount(vm_object_t object, vm_offset_t start, 3084 vm_offset_t end) 3085 { 3086 3087 VM_OBJECT_WLOCK(object); 3088 KASSERT((object->flags & OBJ_ANON) == 0, 3089 ("Splittable object with writecount")); 3090 object->un_pager.swp.writemappings += (vm_ooffset_t)end - start; 3091 VM_OBJECT_WUNLOCK(object); 3092 } 3093 3094 static void 3095 swap_pager_release_writecount(vm_object_t object, vm_offset_t start, 3096 vm_offset_t end) 3097 { 3098 3099 VM_OBJECT_WLOCK(object); 3100 KASSERT((object->flags & OBJ_ANON) == 0, 3101 ("Splittable object with writecount")); 3102 object->un_pager.swp.writemappings -= (vm_ooffset_t)end - start; 3103 VM_OBJECT_WUNLOCK(object); 3104 } 3105