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/limits.h> 86 #include <sys/lock.h> 87 #include <sys/kernel.h> 88 #include <sys/mount.h> 89 #include <sys/namei.h> 90 #include <sys/malloc.h> 91 #include <sys/pctrie.h> 92 #include <sys/priv.h> 93 #include <sys/proc.h> 94 #include <sys/racct.h> 95 #include <sys/resource.h> 96 #include <sys/resourcevar.h> 97 #include <sys/rwlock.h> 98 #include <sys/sbuf.h> 99 #include <sys/sysctl.h> 100 #include <sys/sysproto.h> 101 #include <sys/systm.h> 102 #include <sys/sx.h> 103 #include <sys/unistd.h> 104 #include <sys/user.h> 105 #include <sys/vmmeter.h> 106 #include <sys/vnode.h> 107 108 #include <security/mac/mac_framework.h> 109 110 #include <vm/vm.h> 111 #include <vm/pmap.h> 112 #include <vm/vm_map.h> 113 #include <vm/vm_kern.h> 114 #include <vm/vm_object.h> 115 #include <vm/vm_page.h> 116 #include <vm/vm_pager.h> 117 #include <vm/vm_pageout.h> 118 #include <vm/vm_param.h> 119 #include <vm/swap_pager.h> 120 #include <vm/vm_extern.h> 121 #include <vm/uma.h> 122 123 #include <geom/geom.h> 124 125 /* 126 * MAX_PAGEOUT_CLUSTER must be a power of 2 between 1 and 64. 127 * The 64-page limit is due to the radix code (kern/subr_blist.c). 128 */ 129 #ifndef MAX_PAGEOUT_CLUSTER 130 #define MAX_PAGEOUT_CLUSTER 32 131 #endif 132 133 #if !defined(SWB_NPAGES) 134 #define SWB_NPAGES MAX_PAGEOUT_CLUSTER 135 #endif 136 137 #define SWAP_META_PAGES PCTRIE_COUNT 138 139 /* 140 * A swblk structure maps each page index within a 141 * SWAP_META_PAGES-aligned and sized range to the address of an 142 * on-disk swap block (or SWAPBLK_NONE). The collection of these 143 * mappings for an entire vm object is implemented as a pc-trie. 144 */ 145 struct swblk { 146 vm_pindex_t p; 147 daddr_t d[SWAP_META_PAGES]; 148 }; 149 150 static MALLOC_DEFINE(M_VMPGDATA, "vm_pgdata", "swap pager private data"); 151 static struct mtx sw_dev_mtx; 152 static TAILQ_HEAD(, swdevt) swtailq = TAILQ_HEAD_INITIALIZER(swtailq); 153 static struct swdevt *swdevhd; /* Allocate from here next */ 154 static int nswapdev; /* Number of swap devices */ 155 int swap_pager_avail; 156 static struct sx swdev_syscall_lock; /* serialize swap(on|off) */ 157 158 static __exclusive_cache_line u_long swap_reserved; 159 static u_long swap_total; 160 static int sysctl_page_shift(SYSCTL_HANDLER_ARGS); 161 162 static SYSCTL_NODE(_vm_stats, OID_AUTO, swap, CTLFLAG_RD | CTLFLAG_MPSAFE, 0, 163 "VM swap stats"); 164 165 SYSCTL_PROC(_vm, OID_AUTO, swap_reserved, CTLTYPE_U64 | CTLFLAG_RD | CTLFLAG_MPSAFE, 166 &swap_reserved, 0, sysctl_page_shift, "QU", 167 "Amount of swap storage needed to back all allocated anonymous memory."); 168 SYSCTL_PROC(_vm, OID_AUTO, swap_total, CTLTYPE_U64 | CTLFLAG_RD | CTLFLAG_MPSAFE, 169 &swap_total, 0, sysctl_page_shift, "QU", 170 "Total amount of available swap storage."); 171 172 int vm_overcommit __read_mostly = 0; 173 SYSCTL_INT(_vm, VM_OVERCOMMIT, overcommit, CTLFLAG_RW, &vm_overcommit, 0, 174 "Configure virtual memory overcommit behavior. See tuning(7) " 175 "for details."); 176 static unsigned long swzone; 177 SYSCTL_ULONG(_vm, OID_AUTO, swzone, CTLFLAG_RD, &swzone, 0, 178 "Actual size of swap metadata zone"); 179 static unsigned long swap_maxpages; 180 SYSCTL_ULONG(_vm, OID_AUTO, swap_maxpages, CTLFLAG_RD, &swap_maxpages, 0, 181 "Maximum amount of swap supported"); 182 183 static COUNTER_U64_DEFINE_EARLY(swap_free_deferred); 184 SYSCTL_COUNTER_U64(_vm_stats_swap, OID_AUTO, free_deferred, 185 CTLFLAG_RD, &swap_free_deferred, 186 "Number of pages that deferred freeing swap space"); 187 188 static COUNTER_U64_DEFINE_EARLY(swap_free_completed); 189 SYSCTL_COUNTER_U64(_vm_stats_swap, OID_AUTO, free_completed, 190 CTLFLAG_RD, &swap_free_completed, 191 "Number of deferred frees completed"); 192 193 static int 194 sysctl_page_shift(SYSCTL_HANDLER_ARGS) 195 { 196 uint64_t newval; 197 u_long value = *(u_long *)arg1; 198 199 newval = ((uint64_t)value) << PAGE_SHIFT; 200 return (sysctl_handle_64(oidp, &newval, 0, req)); 201 } 202 203 static bool 204 swap_reserve_by_cred_rlimit(u_long pincr, struct ucred *cred, int oc) 205 { 206 struct uidinfo *uip; 207 u_long prev; 208 209 uip = cred->cr_ruidinfo; 210 211 prev = atomic_fetchadd_long(&uip->ui_vmsize, pincr); 212 if ((oc & SWAP_RESERVE_RLIMIT_ON) != 0 && 213 prev + pincr > lim_cur(curthread, RLIMIT_SWAP) && 214 priv_check(curthread, PRIV_VM_SWAP_NORLIMIT) != 0) { 215 prev = atomic_fetchadd_long(&uip->ui_vmsize, -pincr); 216 KASSERT(prev >= pincr, 217 ("negative vmsize for uid %d\n", uip->ui_uid)); 218 return (false); 219 } 220 return (true); 221 } 222 223 static void 224 swap_release_by_cred_rlimit(u_long pdecr, struct ucred *cred) 225 { 226 struct uidinfo *uip; 227 #ifdef INVARIANTS 228 u_long prev; 229 #endif 230 231 uip = cred->cr_ruidinfo; 232 233 #ifdef INVARIANTS 234 prev = atomic_fetchadd_long(&uip->ui_vmsize, -pdecr); 235 KASSERT(prev >= pdecr, 236 ("negative vmsize for uid %d\n", uip->ui_uid)); 237 #else 238 atomic_subtract_long(&uip->ui_vmsize, pdecr); 239 #endif 240 } 241 242 static void 243 swap_reserve_force_rlimit(u_long pincr, struct ucred *cred) 244 { 245 struct uidinfo *uip; 246 247 uip = cred->cr_ruidinfo; 248 atomic_add_long(&uip->ui_vmsize, pincr); 249 } 250 251 bool 252 swap_reserve(vm_ooffset_t incr) 253 { 254 255 return (swap_reserve_by_cred(incr, curthread->td_ucred)); 256 } 257 258 bool 259 swap_reserve_by_cred(vm_ooffset_t incr, struct ucred *cred) 260 { 261 u_long r, s, prev, pincr; 262 #ifdef RACCT 263 int error; 264 #endif 265 int oc; 266 static int curfail; 267 static struct timeval lastfail; 268 269 KASSERT((incr & PAGE_MASK) == 0, ("%s: incr: %ju & PAGE_MASK", 270 __func__, (uintmax_t)incr)); 271 272 #ifdef RACCT 273 if (RACCT_ENABLED()) { 274 PROC_LOCK(curproc); 275 error = racct_add(curproc, RACCT_SWAP, incr); 276 PROC_UNLOCK(curproc); 277 if (error != 0) 278 return (false); 279 } 280 #endif 281 282 pincr = atop(incr); 283 prev = atomic_fetchadd_long(&swap_reserved, pincr); 284 r = prev + pincr; 285 s = swap_total; 286 oc = atomic_load_int(&vm_overcommit); 287 if (r > s && (oc & SWAP_RESERVE_ALLOW_NONWIRED) != 0) { 288 s += vm_cnt.v_page_count - vm_cnt.v_free_reserved - 289 vm_wire_count(); 290 } 291 if ((oc & SWAP_RESERVE_FORCE_ON) != 0 && r > s && 292 priv_check(curthread, PRIV_VM_SWAP_NOQUOTA) != 0) { 293 prev = atomic_fetchadd_long(&swap_reserved, -pincr); 294 KASSERT(prev >= pincr, 295 ("swap_reserved < incr on overcommit fail")); 296 goto out_error; 297 } 298 299 if (!swap_reserve_by_cred_rlimit(pincr, cred, oc)) { 300 prev = atomic_fetchadd_long(&swap_reserved, -pincr); 301 KASSERT(prev >= pincr, 302 ("swap_reserved < incr on overcommit fail")); 303 goto out_error; 304 } 305 306 return (true); 307 308 out_error: 309 if (ppsratecheck(&lastfail, &curfail, 1)) { 310 printf("uid %d, pid %d: swap reservation " 311 "for %jd bytes failed\n", 312 cred->cr_ruidinfo->ui_uid, curproc->p_pid, incr); 313 } 314 #ifdef RACCT 315 if (RACCT_ENABLED()) { 316 PROC_LOCK(curproc); 317 racct_sub(curproc, RACCT_SWAP, incr); 318 PROC_UNLOCK(curproc); 319 } 320 #endif 321 322 return (false); 323 } 324 325 void 326 swap_reserve_force(vm_ooffset_t incr) 327 { 328 u_long pincr; 329 330 KASSERT((incr & PAGE_MASK) == 0, ("%s: incr: %ju & PAGE_MASK", 331 __func__, (uintmax_t)incr)); 332 333 #ifdef RACCT 334 if (RACCT_ENABLED()) { 335 PROC_LOCK(curproc); 336 racct_add_force(curproc, RACCT_SWAP, incr); 337 PROC_UNLOCK(curproc); 338 } 339 #endif 340 pincr = atop(incr); 341 atomic_add_long(&swap_reserved, pincr); 342 swap_reserve_force_rlimit(pincr, curthread->td_ucred); 343 } 344 345 void 346 swap_release(vm_ooffset_t decr) 347 { 348 struct ucred *cred; 349 350 PROC_LOCK(curproc); 351 cred = curproc->p_ucred; 352 swap_release_by_cred(decr, cred); 353 PROC_UNLOCK(curproc); 354 } 355 356 void 357 swap_release_by_cred(vm_ooffset_t decr, struct ucred *cred) 358 { 359 u_long pdecr; 360 #ifdef INVARIANTS 361 u_long prev; 362 #endif 363 364 KASSERT((decr & PAGE_MASK) == 0, ("%s: decr: %ju & PAGE_MASK", 365 __func__, (uintmax_t)decr)); 366 367 pdecr = atop(decr); 368 #ifdef INVARIANTS 369 prev = atomic_fetchadd_long(&swap_reserved, -pdecr); 370 KASSERT(prev >= pdecr, ("swap_reserved < decr")); 371 #else 372 atomic_subtract_long(&swap_reserved, pdecr); 373 #endif 374 375 swap_release_by_cred_rlimit(pdecr, cred); 376 #ifdef RACCT 377 if (racct_enable) 378 racct_sub_cred(cred, RACCT_SWAP, decr); 379 #endif 380 } 381 382 static int swap_pager_full = 2; /* swap space exhaustion (task killing) */ 383 static int swap_pager_almost_full = 1; /* swap space exhaustion (w/hysteresis)*/ 384 static struct mtx swbuf_mtx; /* to sync nsw_wcount_async */ 385 static int nsw_wcount_async; /* limit async write buffers */ 386 static int nsw_wcount_async_max;/* assigned maximum */ 387 int nsw_cluster_max; /* maximum VOP I/O allowed */ 388 389 static int sysctl_swap_async_max(SYSCTL_HANDLER_ARGS); 390 SYSCTL_PROC(_vm, OID_AUTO, swap_async_max, CTLTYPE_INT | CTLFLAG_RW | 391 CTLFLAG_MPSAFE, NULL, 0, sysctl_swap_async_max, "I", 392 "Maximum running async swap ops"); 393 static int sysctl_swap_fragmentation(SYSCTL_HANDLER_ARGS); 394 SYSCTL_PROC(_vm, OID_AUTO, swap_fragmentation, CTLTYPE_STRING | CTLFLAG_RD | 395 CTLFLAG_MPSAFE, NULL, 0, sysctl_swap_fragmentation, "A", 396 "Swap Fragmentation Info"); 397 398 static struct sx sw_alloc_sx; 399 400 /* 401 * "named" and "unnamed" anon region objects. Try to reduce the overhead 402 * of searching a named list by hashing it just a little. 403 */ 404 405 #define NOBJLISTS 8 406 407 #define NOBJLIST(handle) \ 408 (&swap_pager_object_list[((int)(intptr_t)handle >> 4) & (NOBJLISTS-1)]) 409 410 static struct pagerlst swap_pager_object_list[NOBJLISTS]; 411 static uma_zone_t swwbuf_zone; 412 static uma_zone_t swrbuf_zone; 413 static uma_zone_t swblk_zone; 414 static uma_zone_t swpctrie_zone; 415 416 /* 417 * pagerops for OBJT_SWAP - "swap pager". Some ops are also global procedure 418 * calls hooked from other parts of the VM system and do not appear here. 419 * (see vm/swap_pager.h). 420 */ 421 static vm_object_t 422 swap_pager_alloc(void *handle, vm_ooffset_t size, 423 vm_prot_t prot, vm_ooffset_t offset, struct ucred *); 424 static void swap_pager_dealloc(vm_object_t object); 425 static int swap_pager_getpages(vm_object_t, vm_page_t *, int, int *, 426 int *); 427 static int swap_pager_getpages_async(vm_object_t, vm_page_t *, int, int *, 428 int *, pgo_getpages_iodone_t, void *); 429 static void swap_pager_putpages(vm_object_t, vm_page_t *, int, boolean_t, int *); 430 static boolean_t 431 swap_pager_haspage(vm_object_t object, vm_pindex_t pindex, int *before, int *after); 432 static void swap_pager_init(void); 433 static void swap_pager_unswapped(vm_page_t); 434 static void swap_pager_swapoff(struct swdevt *sp); 435 static void swap_pager_update_writecount(vm_object_t object, 436 vm_offset_t start, vm_offset_t end); 437 static void swap_pager_release_writecount(vm_object_t object, 438 vm_offset_t start, vm_offset_t end); 439 static void swap_pager_freespace_pgo(vm_object_t object, vm_pindex_t start, 440 vm_size_t size); 441 442 const struct pagerops swappagerops = { 443 .pgo_kvme_type = KVME_TYPE_SWAP, 444 .pgo_init = swap_pager_init, /* early system initialization of pager */ 445 .pgo_alloc = swap_pager_alloc, /* allocate an OBJT_SWAP object */ 446 .pgo_dealloc = swap_pager_dealloc, /* deallocate an OBJT_SWAP object */ 447 .pgo_getpages = swap_pager_getpages, /* pagein */ 448 .pgo_getpages_async = swap_pager_getpages_async, /* pagein (async) */ 449 .pgo_putpages = swap_pager_putpages, /* pageout */ 450 .pgo_haspage = swap_pager_haspage, /* get backing store status for page */ 451 .pgo_pageunswapped = swap_pager_unswapped, /* remove swap related to page */ 452 .pgo_update_writecount = swap_pager_update_writecount, 453 .pgo_release_writecount = swap_pager_release_writecount, 454 .pgo_freespace = swap_pager_freespace_pgo, 455 }; 456 457 /* 458 * swap_*() routines are externally accessible. swp_*() routines are 459 * internal. 460 */ 461 static int nswap_lowat = 128; /* in pages, swap_pager_almost_full warn */ 462 static int nswap_hiwat = 512; /* in pages, swap_pager_almost_full warn */ 463 464 SYSCTL_INT(_vm, OID_AUTO, dmmax, CTLFLAG_RD, &nsw_cluster_max, 0, 465 "Maximum size of a swap block in pages"); 466 467 static void swp_sizecheck(void); 468 static void swp_pager_async_iodone(struct buf *bp); 469 static bool swp_pager_swblk_empty(struct swblk *sb, int start, int limit); 470 static void swp_pager_free_empty_swblk(vm_object_t, struct swblk *sb); 471 static int swapongeom(struct vnode *); 472 static int swaponvp(struct thread *, struct vnode *, u_long); 473 static int swapoff_one(struct swdevt *sp, struct ucred *cred, 474 u_int flags); 475 476 /* 477 * Swap bitmap functions 478 */ 479 static void swp_pager_freeswapspace(daddr_t blk, daddr_t npages); 480 static daddr_t swp_pager_getswapspace(int *npages); 481 482 /* 483 * Metadata functions 484 */ 485 static daddr_t swp_pager_meta_build(vm_object_t, vm_pindex_t, daddr_t); 486 static void swp_pager_meta_free(vm_object_t, vm_pindex_t, vm_pindex_t, 487 vm_size_t *); 488 static void swp_pager_meta_transfer(vm_object_t src, vm_object_t dst, 489 vm_pindex_t pindex, vm_pindex_t count, vm_size_t *freed); 490 static void swp_pager_meta_free_all(vm_object_t); 491 static daddr_t swp_pager_meta_lookup(vm_object_t, vm_pindex_t); 492 493 static void 494 swp_pager_init_freerange(daddr_t *start, daddr_t *num) 495 { 496 497 *start = SWAPBLK_NONE; 498 *num = 0; 499 } 500 501 static void 502 swp_pager_update_freerange(daddr_t *start, daddr_t *num, daddr_t addr) 503 { 504 505 if (*start + *num == addr) { 506 (*num)++; 507 } else { 508 swp_pager_freeswapspace(*start, *num); 509 *start = addr; 510 *num = 1; 511 } 512 } 513 514 static void * 515 swblk_trie_alloc(struct pctrie *ptree) 516 { 517 518 return (uma_zalloc(swpctrie_zone, M_NOWAIT | (curproc == pageproc ? 519 M_USE_RESERVE : 0))); 520 } 521 522 static void 523 swblk_trie_free(struct pctrie *ptree, void *node) 524 { 525 526 uma_zfree(swpctrie_zone, node); 527 } 528 529 PCTRIE_DEFINE(SWAP, swblk, p, swblk_trie_alloc, swblk_trie_free); 530 531 /* 532 * SWP_SIZECHECK() - update swap_pager_full indication 533 * 534 * update the swap_pager_almost_full indication and warn when we are 535 * about to run out of swap space, using lowat/hiwat hysteresis. 536 * 537 * Clear swap_pager_full ( task killing ) indication when lowat is met. 538 * 539 * No restrictions on call 540 * This routine may not block. 541 */ 542 static void 543 swp_sizecheck(void) 544 { 545 546 if (swap_pager_avail < nswap_lowat) { 547 if (swap_pager_almost_full == 0) { 548 printf("swap_pager: out of swap space\n"); 549 swap_pager_almost_full = 1; 550 } 551 } else { 552 swap_pager_full = 0; 553 if (swap_pager_avail > nswap_hiwat) 554 swap_pager_almost_full = 0; 555 } 556 } 557 558 /* 559 * SWAP_PAGER_INIT() - initialize the swap pager! 560 * 561 * Expected to be started from system init. NOTE: This code is run 562 * before much else so be careful what you depend on. Most of the VM 563 * system has yet to be initialized at this point. 564 */ 565 static void 566 swap_pager_init(void) 567 { 568 /* 569 * Initialize object lists 570 */ 571 int i; 572 573 for (i = 0; i < NOBJLISTS; ++i) 574 TAILQ_INIT(&swap_pager_object_list[i]); 575 mtx_init(&sw_dev_mtx, "swapdev", NULL, MTX_DEF); 576 sx_init(&sw_alloc_sx, "swspsx"); 577 sx_init(&swdev_syscall_lock, "swsysc"); 578 579 /* 580 * The nsw_cluster_max is constrained by the bp->b_pages[] 581 * array, which has maxphys / PAGE_SIZE entries, and our locally 582 * defined MAX_PAGEOUT_CLUSTER. Also be aware that swap ops are 583 * constrained by the swap device interleave stripe size. 584 * 585 * Initialized early so that GEOM_ELI can see it. 586 */ 587 nsw_cluster_max = min(maxphys / PAGE_SIZE, MAX_PAGEOUT_CLUSTER); 588 } 589 590 /* 591 * SWAP_PAGER_SWAP_INIT() - swap pager initialization from pageout process 592 * 593 * Expected to be started from pageout process once, prior to entering 594 * its main loop. 595 */ 596 void 597 swap_pager_swap_init(void) 598 { 599 unsigned long n, n2; 600 601 /* 602 * Number of in-transit swap bp operations. Don't 603 * exhaust the pbufs completely. Make sure we 604 * initialize workable values (0 will work for hysteresis 605 * but it isn't very efficient). 606 * 607 * Currently we hardwire nsw_wcount_async to 4. This limit is 608 * designed to prevent other I/O from having high latencies due to 609 * our pageout I/O. The value 4 works well for one or two active swap 610 * devices but is probably a little low if you have more. Even so, 611 * a higher value would probably generate only a limited improvement 612 * with three or four active swap devices since the system does not 613 * typically have to pageout at extreme bandwidths. We will want 614 * at least 2 per swap devices, and 4 is a pretty good value if you 615 * have one NFS swap device due to the command/ack latency over NFS. 616 * So it all works out pretty well. 617 * 618 * nsw_cluster_max is initialized in swap_pager_init(). 619 */ 620 621 nsw_wcount_async = 4; 622 nsw_wcount_async_max = nsw_wcount_async; 623 mtx_init(&swbuf_mtx, "async swbuf mutex", NULL, MTX_DEF); 624 625 swwbuf_zone = pbuf_zsecond_create("swwbuf", nswbuf / 4); 626 swrbuf_zone = pbuf_zsecond_create("swrbuf", nswbuf / 2); 627 628 /* 629 * Initialize our zone, taking the user's requested size or 630 * estimating the number we need based on the number of pages 631 * in the system. 632 */ 633 n = maxswzone != 0 ? maxswzone / sizeof(struct swblk) : 634 vm_cnt.v_page_count / 2; 635 swpctrie_zone = uma_zcreate("swpctrie", pctrie_node_size(), NULL, NULL, 636 pctrie_zone_init, NULL, UMA_ALIGN_PTR, 0); 637 swblk_zone = uma_zcreate("swblk", sizeof(struct swblk), NULL, NULL, 638 NULL, NULL, _Alignof(struct swblk) - 1, 0); 639 n2 = n; 640 do { 641 if (uma_zone_reserve_kva(swblk_zone, n)) 642 break; 643 /* 644 * if the allocation failed, try a zone two thirds the 645 * size of the previous attempt. 646 */ 647 n -= ((n + 2) / 3); 648 } while (n > 0); 649 650 /* 651 * Often uma_zone_reserve_kva() cannot reserve exactly the 652 * requested size. Account for the difference when 653 * calculating swap_maxpages. 654 */ 655 n = uma_zone_get_max(swblk_zone); 656 657 if (n < n2) 658 printf("Swap blk zone entries changed from %lu to %lu.\n", 659 n2, n); 660 /* absolute maximum we can handle assuming 100% efficiency */ 661 swap_maxpages = n * SWAP_META_PAGES; 662 swzone = n * sizeof(struct swblk); 663 if (!uma_zone_reserve_kva(swpctrie_zone, n)) 664 printf("Cannot reserve swap pctrie zone, " 665 "reduce kern.maxswzone.\n"); 666 } 667 668 bool 669 swap_pager_init_object(vm_object_t object, void *handle, struct ucred *cred, 670 vm_ooffset_t size, vm_ooffset_t offset) 671 { 672 if (cred != NULL) { 673 if (!swap_reserve_by_cred(size, cred)) 674 return (false); 675 crhold(cred); 676 } 677 678 object->un_pager.swp.writemappings = 0; 679 object->handle = handle; 680 if (cred != NULL) { 681 object->cred = cred; 682 object->charge = size; 683 } 684 return (true); 685 } 686 687 static vm_object_t 688 swap_pager_alloc_init(objtype_t otype, void *handle, struct ucred *cred, 689 vm_ooffset_t size, vm_ooffset_t offset) 690 { 691 vm_object_t object; 692 693 /* 694 * The un_pager.swp.swp_blks trie is initialized by 695 * vm_object_allocate() to ensure the correct order of 696 * visibility to other threads. 697 */ 698 object = vm_object_allocate(otype, OFF_TO_IDX(offset + 699 PAGE_MASK + size)); 700 701 if (!swap_pager_init_object(object, handle, cred, size, offset)) { 702 vm_object_deallocate(object); 703 return (NULL); 704 } 705 return (object); 706 } 707 708 /* 709 * SWAP_PAGER_ALLOC() - allocate a new OBJT_SWAP VM object and instantiate 710 * its metadata structures. 711 * 712 * This routine is called from the mmap and fork code to create a new 713 * OBJT_SWAP object. 714 * 715 * This routine must ensure that no live duplicate is created for 716 * the named object request, which is protected against by 717 * holding the sw_alloc_sx lock in case handle != NULL. 718 */ 719 static vm_object_t 720 swap_pager_alloc(void *handle, vm_ooffset_t size, vm_prot_t prot, 721 vm_ooffset_t offset, struct ucred *cred) 722 { 723 vm_object_t object; 724 725 if (handle != NULL) { 726 /* 727 * Reference existing named region or allocate new one. There 728 * should not be a race here against swp_pager_meta_build() 729 * as called from vm_page_remove() in regards to the lookup 730 * of the handle. 731 */ 732 sx_xlock(&sw_alloc_sx); 733 object = vm_pager_object_lookup(NOBJLIST(handle), handle); 734 if (object == NULL) { 735 object = swap_pager_alloc_init(OBJT_SWAP, handle, cred, 736 size, offset); 737 if (object != NULL) { 738 TAILQ_INSERT_TAIL(NOBJLIST(object->handle), 739 object, pager_object_list); 740 } 741 } 742 sx_xunlock(&sw_alloc_sx); 743 } else { 744 object = swap_pager_alloc_init(OBJT_SWAP, handle, cred, 745 size, offset); 746 } 747 return (object); 748 } 749 750 /* 751 * SWAP_PAGER_DEALLOC() - remove swap metadata from object 752 * 753 * The swap backing for the object is destroyed. The code is 754 * designed such that we can reinstantiate it later, but this 755 * routine is typically called only when the entire object is 756 * about to be destroyed. 757 * 758 * The object must be locked. 759 */ 760 static void 761 swap_pager_dealloc(vm_object_t object) 762 { 763 764 VM_OBJECT_ASSERT_WLOCKED(object); 765 KASSERT((object->flags & OBJ_DEAD) != 0, ("dealloc of reachable obj")); 766 767 /* 768 * Remove from list right away so lookups will fail if we block for 769 * pageout completion. 770 */ 771 if ((object->flags & OBJ_ANON) == 0 && object->handle != NULL) { 772 VM_OBJECT_WUNLOCK(object); 773 sx_xlock(&sw_alloc_sx); 774 TAILQ_REMOVE(NOBJLIST(object->handle), object, 775 pager_object_list); 776 sx_xunlock(&sw_alloc_sx); 777 VM_OBJECT_WLOCK(object); 778 } 779 780 vm_object_pip_wait(object, "swpdea"); 781 782 /* 783 * Free all remaining metadata. We only bother to free it from 784 * the swap meta data. We do not attempt to free swapblk's still 785 * associated with vm_page_t's for this object. We do not care 786 * if paging is still in progress on some objects. 787 */ 788 swp_pager_meta_free_all(object); 789 object->handle = NULL; 790 object->type = OBJT_DEAD; 791 792 /* 793 * Release the allocation charge. 794 */ 795 if (object->cred != NULL) { 796 swap_release_by_cred(object->charge, object->cred); 797 object->charge = 0; 798 crfree(object->cred); 799 object->cred = NULL; 800 } 801 802 /* 803 * Hide the object from swap_pager_swapoff(). 804 */ 805 vm_object_clear_flag(object, OBJ_SWAP); 806 } 807 808 /************************************************************************ 809 * SWAP PAGER BITMAP ROUTINES * 810 ************************************************************************/ 811 812 /* 813 * SWP_PAGER_GETSWAPSPACE() - allocate raw swap space 814 * 815 * Allocate swap for up to the requested number of pages. The 816 * starting swap block number (a page index) is returned or 817 * SWAPBLK_NONE if the allocation failed. 818 * 819 * Also has the side effect of advising that somebody made a mistake 820 * when they configured swap and didn't configure enough. 821 * 822 * This routine may not sleep. 823 * 824 * We allocate in round-robin fashion from the configured devices. 825 */ 826 static daddr_t 827 swp_pager_getswapspace(int *io_npages) 828 { 829 daddr_t blk; 830 struct swdevt *sp; 831 int mpages, npages; 832 833 KASSERT(*io_npages >= 1, 834 ("%s: npages not positive", __func__)); 835 blk = SWAPBLK_NONE; 836 mpages = *io_npages; 837 npages = imin(BLIST_MAX_ALLOC, mpages); 838 mtx_lock(&sw_dev_mtx); 839 sp = swdevhd; 840 while (!TAILQ_EMPTY(&swtailq)) { 841 if (sp == NULL) 842 sp = TAILQ_FIRST(&swtailq); 843 if ((sp->sw_flags & SW_CLOSING) == 0) 844 blk = blist_alloc(sp->sw_blist, &npages, mpages); 845 if (blk != SWAPBLK_NONE) 846 break; 847 sp = TAILQ_NEXT(sp, sw_list); 848 if (swdevhd == sp) { 849 if (npages == 1) 850 break; 851 mpages = npages - 1; 852 npages >>= 1; 853 } 854 } 855 if (blk != SWAPBLK_NONE) { 856 *io_npages = npages; 857 blk += sp->sw_first; 858 sp->sw_used += npages; 859 swap_pager_avail -= npages; 860 swp_sizecheck(); 861 swdevhd = TAILQ_NEXT(sp, sw_list); 862 } else { 863 if (swap_pager_full != 2) { 864 printf("swp_pager_getswapspace(%d): failed\n", 865 *io_npages); 866 swap_pager_full = 2; 867 swap_pager_almost_full = 1; 868 } 869 swdevhd = NULL; 870 } 871 mtx_unlock(&sw_dev_mtx); 872 return (blk); 873 } 874 875 static bool 876 swp_pager_isondev(daddr_t blk, struct swdevt *sp) 877 { 878 879 return (blk >= sp->sw_first && blk < sp->sw_end); 880 } 881 882 static void 883 swp_pager_strategy(struct buf *bp) 884 { 885 struct swdevt *sp; 886 887 mtx_lock(&sw_dev_mtx); 888 TAILQ_FOREACH(sp, &swtailq, sw_list) { 889 if (swp_pager_isondev(bp->b_blkno, sp)) { 890 mtx_unlock(&sw_dev_mtx); 891 if ((sp->sw_flags & SW_UNMAPPED) != 0 && 892 unmapped_buf_allowed) { 893 bp->b_data = unmapped_buf; 894 bp->b_offset = 0; 895 } else { 896 pmap_qenter((vm_offset_t)bp->b_data, 897 &bp->b_pages[0], bp->b_bcount / PAGE_SIZE); 898 } 899 sp->sw_strategy(bp, sp); 900 return; 901 } 902 } 903 panic("Swapdev not found"); 904 } 905 906 /* 907 * SWP_PAGER_FREESWAPSPACE() - free raw swap space 908 * 909 * This routine returns the specified swap blocks back to the bitmap. 910 * 911 * This routine may not sleep. 912 */ 913 static void 914 swp_pager_freeswapspace(daddr_t blk, daddr_t npages) 915 { 916 struct swdevt *sp; 917 918 if (npages == 0) 919 return; 920 mtx_lock(&sw_dev_mtx); 921 TAILQ_FOREACH(sp, &swtailq, sw_list) { 922 if (swp_pager_isondev(blk, sp)) { 923 sp->sw_used -= npages; 924 /* 925 * If we are attempting to stop swapping on 926 * this device, we don't want to mark any 927 * blocks free lest they be reused. 928 */ 929 if ((sp->sw_flags & SW_CLOSING) == 0) { 930 blist_free(sp->sw_blist, blk - sp->sw_first, 931 npages); 932 swap_pager_avail += npages; 933 swp_sizecheck(); 934 } 935 mtx_unlock(&sw_dev_mtx); 936 return; 937 } 938 } 939 panic("Swapdev not found"); 940 } 941 942 /* 943 * SYSCTL_SWAP_FRAGMENTATION() - produce raw swap space stats 944 */ 945 static int 946 sysctl_swap_fragmentation(SYSCTL_HANDLER_ARGS) 947 { 948 struct sbuf sbuf; 949 struct swdevt *sp; 950 const char *devname; 951 int error; 952 953 error = sysctl_wire_old_buffer(req, 0); 954 if (error != 0) 955 return (error); 956 sbuf_new_for_sysctl(&sbuf, NULL, 128, req); 957 mtx_lock(&sw_dev_mtx); 958 TAILQ_FOREACH(sp, &swtailq, sw_list) { 959 if (vn_isdisk(sp->sw_vp)) 960 devname = devtoname(sp->sw_vp->v_rdev); 961 else 962 devname = "[file]"; 963 sbuf_printf(&sbuf, "\nFree space on device %s:\n", devname); 964 blist_stats(sp->sw_blist, &sbuf); 965 } 966 mtx_unlock(&sw_dev_mtx); 967 error = sbuf_finish(&sbuf); 968 sbuf_delete(&sbuf); 969 return (error); 970 } 971 972 /* 973 * SWAP_PAGER_FREESPACE() - frees swap blocks associated with a page 974 * range within an object. 975 * 976 * This routine removes swapblk assignments from swap metadata. 977 * 978 * The external callers of this routine typically have already destroyed 979 * or renamed vm_page_t's associated with this range in the object so 980 * we should be ok. 981 * 982 * The object must be locked. 983 */ 984 void 985 swap_pager_freespace(vm_object_t object, vm_pindex_t start, vm_size_t size, 986 vm_size_t *freed) 987 { 988 MPASS((object->flags & OBJ_SWAP) != 0); 989 990 swp_pager_meta_free(object, start, size, freed); 991 } 992 993 static void 994 swap_pager_freespace_pgo(vm_object_t object, vm_pindex_t start, vm_size_t size) 995 { 996 MPASS((object->flags & OBJ_SWAP) != 0); 997 998 swp_pager_meta_free(object, start, size, NULL); 999 } 1000 1001 /* 1002 * SWAP_PAGER_RESERVE() - reserve swap blocks in object 1003 * 1004 * Assigns swap blocks to the specified range within the object. The 1005 * swap blocks are not zeroed. Any previous swap assignment is destroyed. 1006 * 1007 * Returns 0 on success, -1 on failure. 1008 */ 1009 int 1010 swap_pager_reserve(vm_object_t object, vm_pindex_t start, vm_pindex_t size) 1011 { 1012 daddr_t addr, blk, n_free, s_free; 1013 vm_pindex_t i, j; 1014 int n; 1015 1016 swp_pager_init_freerange(&s_free, &n_free); 1017 VM_OBJECT_WLOCK(object); 1018 for (i = 0; i < size; i += n) { 1019 n = MIN(size - i, INT_MAX); 1020 blk = swp_pager_getswapspace(&n); 1021 if (blk == SWAPBLK_NONE) { 1022 swp_pager_meta_free(object, start, i, NULL); 1023 VM_OBJECT_WUNLOCK(object); 1024 return (-1); 1025 } 1026 for (j = 0; j < n; ++j) { 1027 addr = swp_pager_meta_build(object, 1028 start + i + j, blk + j); 1029 if (addr != SWAPBLK_NONE) 1030 swp_pager_update_freerange(&s_free, &n_free, 1031 addr); 1032 } 1033 } 1034 swp_pager_freeswapspace(s_free, n_free); 1035 VM_OBJECT_WUNLOCK(object); 1036 return (0); 1037 } 1038 1039 static bool 1040 swp_pager_xfer_source(vm_object_t srcobject, vm_object_t dstobject, 1041 vm_pindex_t pindex, daddr_t addr) 1042 { 1043 daddr_t dstaddr __diagused; 1044 1045 KASSERT((srcobject->flags & OBJ_SWAP) != 0, 1046 ("%s: srcobject not swappable", __func__)); 1047 KASSERT((dstobject->flags & OBJ_SWAP) != 0, 1048 ("%s: dstobject not swappable", __func__)); 1049 1050 if (swp_pager_meta_lookup(dstobject, pindex) != SWAPBLK_NONE) { 1051 /* Caller should destroy the source block. */ 1052 return (false); 1053 } 1054 1055 /* 1056 * Destination has no swapblk and is not resident, transfer source. 1057 * swp_pager_meta_build() can sleep. 1058 */ 1059 VM_OBJECT_WUNLOCK(srcobject); 1060 dstaddr = swp_pager_meta_build(dstobject, pindex, addr); 1061 KASSERT(dstaddr == SWAPBLK_NONE, 1062 ("Unexpected destination swapblk")); 1063 VM_OBJECT_WLOCK(srcobject); 1064 1065 return (true); 1066 } 1067 1068 /* 1069 * SWAP_PAGER_COPY() - copy blocks from source pager to destination pager 1070 * and destroy the source. 1071 * 1072 * Copy any valid swapblks from the source to the destination. In 1073 * cases where both the source and destination have a valid swapblk, 1074 * we keep the destination's. 1075 * 1076 * This routine is allowed to sleep. It may sleep allocating metadata 1077 * indirectly through swp_pager_meta_build(). 1078 * 1079 * The source object contains no vm_page_t's (which is just as well) 1080 * 1081 * The source and destination objects must be locked. 1082 * Both object locks may temporarily be released. 1083 */ 1084 void 1085 swap_pager_copy(vm_object_t srcobject, vm_object_t dstobject, 1086 vm_pindex_t offset, int destroysource) 1087 { 1088 VM_OBJECT_ASSERT_WLOCKED(srcobject); 1089 VM_OBJECT_ASSERT_WLOCKED(dstobject); 1090 1091 /* 1092 * If destroysource is set, we remove the source object from the 1093 * swap_pager internal queue now. 1094 */ 1095 if (destroysource && (srcobject->flags & OBJ_ANON) == 0 && 1096 srcobject->handle != NULL) { 1097 VM_OBJECT_WUNLOCK(srcobject); 1098 VM_OBJECT_WUNLOCK(dstobject); 1099 sx_xlock(&sw_alloc_sx); 1100 TAILQ_REMOVE(NOBJLIST(srcobject->handle), srcobject, 1101 pager_object_list); 1102 sx_xunlock(&sw_alloc_sx); 1103 VM_OBJECT_WLOCK(dstobject); 1104 VM_OBJECT_WLOCK(srcobject); 1105 } 1106 1107 /* 1108 * Transfer source to destination. 1109 */ 1110 swp_pager_meta_transfer(srcobject, dstobject, offset, dstobject->size, 1111 NULL); 1112 1113 /* 1114 * Free left over swap blocks in source. 1115 */ 1116 if (destroysource) 1117 swp_pager_meta_free_all(srcobject); 1118 } 1119 1120 /* 1121 * SWAP_PAGER_HASPAGE() - determine if we have good backing store for 1122 * the requested page. 1123 * 1124 * We determine whether good backing store exists for the requested 1125 * page and return TRUE if it does, FALSE if it doesn't. 1126 * 1127 * If TRUE, we also try to determine how much valid, contiguous backing 1128 * store exists before and after the requested page. 1129 */ 1130 static boolean_t 1131 swap_pager_haspage(vm_object_t object, vm_pindex_t pindex, int *before, 1132 int *after) 1133 { 1134 daddr_t blk, blk0; 1135 int i; 1136 1137 VM_OBJECT_ASSERT_LOCKED(object); 1138 KASSERT((object->flags & OBJ_SWAP) != 0, 1139 ("%s: object not swappable", __func__)); 1140 1141 /* 1142 * do we have good backing store at the requested index ? 1143 */ 1144 blk0 = swp_pager_meta_lookup(object, pindex); 1145 if (blk0 == SWAPBLK_NONE) { 1146 if (before) 1147 *before = 0; 1148 if (after) 1149 *after = 0; 1150 return (FALSE); 1151 } 1152 1153 /* 1154 * find backwards-looking contiguous good backing store 1155 */ 1156 if (before != NULL) { 1157 for (i = 1; i < SWB_NPAGES; i++) { 1158 if (i > pindex) 1159 break; 1160 blk = swp_pager_meta_lookup(object, pindex - i); 1161 if (blk != blk0 - i) 1162 break; 1163 } 1164 *before = i - 1; 1165 } 1166 1167 /* 1168 * find forward-looking contiguous good backing store 1169 */ 1170 if (after != NULL) { 1171 for (i = 1; i < SWB_NPAGES; i++) { 1172 blk = swp_pager_meta_lookup(object, pindex + i); 1173 if (blk != blk0 + i) 1174 break; 1175 } 1176 *after = i - 1; 1177 } 1178 return (TRUE); 1179 } 1180 1181 /* 1182 * SWAP_PAGER_PAGE_UNSWAPPED() - remove swap backing store related to page 1183 * 1184 * This removes any associated swap backing store, whether valid or 1185 * not, from the page. 1186 * 1187 * This routine is typically called when a page is made dirty, at 1188 * which point any associated swap can be freed. MADV_FREE also 1189 * calls us in a special-case situation 1190 * 1191 * NOTE!!! If the page is clean and the swap was valid, the caller 1192 * should make the page dirty before calling this routine. This routine 1193 * does NOT change the m->dirty status of the page. Also: MADV_FREE 1194 * depends on it. 1195 * 1196 * This routine may not sleep. 1197 * 1198 * The object containing the page may be locked. 1199 */ 1200 static void 1201 swap_pager_unswapped(vm_page_t m) 1202 { 1203 struct swblk *sb; 1204 vm_object_t obj; 1205 1206 /* 1207 * Handle enqueing deferred frees first. If we do not have the 1208 * object lock we wait for the page daemon to clear the space. 1209 */ 1210 obj = m->object; 1211 if (!VM_OBJECT_WOWNED(obj)) { 1212 VM_PAGE_OBJECT_BUSY_ASSERT(m); 1213 /* 1214 * The caller is responsible for synchronization but we 1215 * will harmlessly handle races. This is typically provided 1216 * by only calling unswapped() when a page transitions from 1217 * clean to dirty. 1218 */ 1219 if ((m->a.flags & (PGA_SWAP_SPACE | PGA_SWAP_FREE)) == 1220 PGA_SWAP_SPACE) { 1221 vm_page_aflag_set(m, PGA_SWAP_FREE); 1222 counter_u64_add(swap_free_deferred, 1); 1223 } 1224 return; 1225 } 1226 if ((m->a.flags & PGA_SWAP_FREE) != 0) 1227 counter_u64_add(swap_free_completed, 1); 1228 vm_page_aflag_clear(m, PGA_SWAP_FREE | PGA_SWAP_SPACE); 1229 1230 /* 1231 * The meta data only exists if the object is OBJT_SWAP 1232 * and even then might not be allocated yet. 1233 */ 1234 KASSERT((m->object->flags & OBJ_SWAP) != 0, 1235 ("Free object not swappable")); 1236 1237 sb = SWAP_PCTRIE_LOOKUP(&m->object->un_pager.swp.swp_blks, 1238 rounddown(m->pindex, SWAP_META_PAGES)); 1239 if (sb == NULL) 1240 return; 1241 if (sb->d[m->pindex % SWAP_META_PAGES] == SWAPBLK_NONE) 1242 return; 1243 swp_pager_freeswapspace(sb->d[m->pindex % SWAP_META_PAGES], 1); 1244 sb->d[m->pindex % SWAP_META_PAGES] = SWAPBLK_NONE; 1245 swp_pager_free_empty_swblk(m->object, sb); 1246 } 1247 1248 /* 1249 * swap_pager_getpages() - bring pages in from swap 1250 * 1251 * Attempt to page in the pages in array "ma" of length "count". The 1252 * caller may optionally specify that additional pages preceding and 1253 * succeeding the specified range be paged in. The number of such pages 1254 * is returned in the "rbehind" and "rahead" parameters, and they will 1255 * be in the inactive queue upon return. 1256 * 1257 * The pages in "ma" must be busied and will remain busied upon return. 1258 */ 1259 static int 1260 swap_pager_getpages_locked(vm_object_t object, vm_page_t *ma, int count, 1261 int *rbehind, int *rahead) 1262 { 1263 struct buf *bp; 1264 vm_page_t bm, mpred, msucc, p; 1265 vm_pindex_t pindex; 1266 daddr_t blk; 1267 int i, maxahead, maxbehind, reqcount; 1268 1269 VM_OBJECT_ASSERT_WLOCKED(object); 1270 reqcount = count; 1271 1272 KASSERT((object->flags & OBJ_SWAP) != 0, 1273 ("%s: object not swappable", __func__)); 1274 if (!swap_pager_haspage(object, ma[0]->pindex, &maxbehind, &maxahead)) { 1275 VM_OBJECT_WUNLOCK(object); 1276 return (VM_PAGER_FAIL); 1277 } 1278 1279 KASSERT(reqcount - 1 <= maxahead, 1280 ("page count %d extends beyond swap block", reqcount)); 1281 1282 /* 1283 * Do not transfer any pages other than those that are xbusied 1284 * when running during a split or collapse operation. This 1285 * prevents clustering from re-creating pages which are being 1286 * moved into another object. 1287 */ 1288 if ((object->flags & (OBJ_SPLIT | OBJ_DEAD)) != 0) { 1289 maxahead = reqcount - 1; 1290 maxbehind = 0; 1291 } 1292 1293 /* 1294 * Clip the readahead and readbehind ranges to exclude resident pages. 1295 */ 1296 if (rahead != NULL) { 1297 *rahead = imin(*rahead, maxahead - (reqcount - 1)); 1298 pindex = ma[reqcount - 1]->pindex; 1299 msucc = TAILQ_NEXT(ma[reqcount - 1], listq); 1300 if (msucc != NULL && msucc->pindex - pindex - 1 < *rahead) 1301 *rahead = msucc->pindex - pindex - 1; 1302 } 1303 if (rbehind != NULL) { 1304 *rbehind = imin(*rbehind, maxbehind); 1305 pindex = ma[0]->pindex; 1306 mpred = TAILQ_PREV(ma[0], pglist, listq); 1307 if (mpred != NULL && pindex - mpred->pindex - 1 < *rbehind) 1308 *rbehind = pindex - mpred->pindex - 1; 1309 } 1310 1311 bm = ma[0]; 1312 for (i = 0; i < count; i++) 1313 ma[i]->oflags |= VPO_SWAPINPROG; 1314 1315 /* 1316 * Allocate readahead and readbehind pages. 1317 */ 1318 if (rbehind != NULL) { 1319 for (i = 1; i <= *rbehind; i++) { 1320 p = vm_page_alloc(object, ma[0]->pindex - i, 1321 VM_ALLOC_NORMAL); 1322 if (p == NULL) 1323 break; 1324 p->oflags |= VPO_SWAPINPROG; 1325 bm = p; 1326 } 1327 *rbehind = i - 1; 1328 } 1329 if (rahead != NULL) { 1330 for (i = 0; i < *rahead; i++) { 1331 p = vm_page_alloc(object, 1332 ma[reqcount - 1]->pindex + i + 1, VM_ALLOC_NORMAL); 1333 if (p == NULL) 1334 break; 1335 p->oflags |= VPO_SWAPINPROG; 1336 } 1337 *rahead = i; 1338 } 1339 if (rbehind != NULL) 1340 count += *rbehind; 1341 if (rahead != NULL) 1342 count += *rahead; 1343 1344 vm_object_pip_add(object, count); 1345 1346 pindex = bm->pindex; 1347 blk = swp_pager_meta_lookup(object, pindex); 1348 KASSERT(blk != SWAPBLK_NONE, 1349 ("no swap blocking containing %p(%jx)", object, (uintmax_t)pindex)); 1350 1351 VM_OBJECT_WUNLOCK(object); 1352 bp = uma_zalloc(swrbuf_zone, M_WAITOK); 1353 MPASS((bp->b_flags & B_MAXPHYS) != 0); 1354 /* Pages cannot leave the object while busy. */ 1355 for (i = 0, p = bm; i < count; i++, p = TAILQ_NEXT(p, listq)) { 1356 MPASS(p->pindex == bm->pindex + i); 1357 bp->b_pages[i] = p; 1358 } 1359 1360 bp->b_flags |= B_PAGING; 1361 bp->b_iocmd = BIO_READ; 1362 bp->b_iodone = swp_pager_async_iodone; 1363 bp->b_rcred = crhold(thread0.td_ucred); 1364 bp->b_wcred = crhold(thread0.td_ucred); 1365 bp->b_blkno = blk; 1366 bp->b_bcount = PAGE_SIZE * count; 1367 bp->b_bufsize = PAGE_SIZE * count; 1368 bp->b_npages = count; 1369 bp->b_pgbefore = rbehind != NULL ? *rbehind : 0; 1370 bp->b_pgafter = rahead != NULL ? *rahead : 0; 1371 1372 VM_CNT_INC(v_swapin); 1373 VM_CNT_ADD(v_swappgsin, count); 1374 1375 /* 1376 * perform the I/O. NOTE!!! bp cannot be considered valid after 1377 * this point because we automatically release it on completion. 1378 * Instead, we look at the one page we are interested in which we 1379 * still hold a lock on even through the I/O completion. 1380 * 1381 * The other pages in our ma[] array are also released on completion, 1382 * so we cannot assume they are valid anymore either. 1383 * 1384 * NOTE: b_blkno is destroyed by the call to swapdev_strategy 1385 */ 1386 BUF_KERNPROC(bp); 1387 swp_pager_strategy(bp); 1388 1389 /* 1390 * Wait for the pages we want to complete. VPO_SWAPINPROG is always 1391 * cleared on completion. If an I/O error occurs, SWAPBLK_NONE 1392 * is set in the metadata for each page in the request. 1393 */ 1394 VM_OBJECT_WLOCK(object); 1395 /* This could be implemented more efficiently with aflags */ 1396 while ((ma[0]->oflags & VPO_SWAPINPROG) != 0) { 1397 ma[0]->oflags |= VPO_SWAPSLEEP; 1398 VM_CNT_INC(v_intrans); 1399 if (VM_OBJECT_SLEEP(object, &object->handle, PSWP, 1400 "swread", hz * 20)) { 1401 printf( 1402 "swap_pager: indefinite wait buffer: bufobj: %p, blkno: %jd, size: %ld\n", 1403 bp->b_bufobj, (intmax_t)bp->b_blkno, bp->b_bcount); 1404 } 1405 } 1406 VM_OBJECT_WUNLOCK(object); 1407 1408 /* 1409 * If we had an unrecoverable read error pages will not be valid. 1410 */ 1411 for (i = 0; i < reqcount; i++) 1412 if (ma[i]->valid != VM_PAGE_BITS_ALL) 1413 return (VM_PAGER_ERROR); 1414 1415 return (VM_PAGER_OK); 1416 1417 /* 1418 * A final note: in a low swap situation, we cannot deallocate swap 1419 * and mark a page dirty here because the caller is likely to mark 1420 * the page clean when we return, causing the page to possibly revert 1421 * to all-zero's later. 1422 */ 1423 } 1424 1425 static int 1426 swap_pager_getpages(vm_object_t object, vm_page_t *ma, int count, 1427 int *rbehind, int *rahead) 1428 { 1429 1430 VM_OBJECT_WLOCK(object); 1431 return (swap_pager_getpages_locked(object, ma, count, rbehind, rahead)); 1432 } 1433 1434 /* 1435 * swap_pager_getpages_async(): 1436 * 1437 * Right now this is emulation of asynchronous operation on top of 1438 * swap_pager_getpages(). 1439 */ 1440 static int 1441 swap_pager_getpages_async(vm_object_t object, vm_page_t *ma, int count, 1442 int *rbehind, int *rahead, pgo_getpages_iodone_t iodone, void *arg) 1443 { 1444 int r, error; 1445 1446 r = swap_pager_getpages(object, ma, count, rbehind, rahead); 1447 switch (r) { 1448 case VM_PAGER_OK: 1449 error = 0; 1450 break; 1451 case VM_PAGER_ERROR: 1452 error = EIO; 1453 break; 1454 case VM_PAGER_FAIL: 1455 error = EINVAL; 1456 break; 1457 default: 1458 panic("unhandled swap_pager_getpages() error %d", r); 1459 } 1460 (iodone)(arg, ma, count, error); 1461 1462 return (r); 1463 } 1464 1465 /* 1466 * swap_pager_putpages: 1467 * 1468 * Assign swap (if necessary) and initiate I/O on the specified pages. 1469 * 1470 * In a low memory situation we may block in VOP_STRATEGY(), but the new 1471 * vm_page reservation system coupled with properly written VFS devices 1472 * should ensure that no low-memory deadlock occurs. This is an area 1473 * which needs work. 1474 * 1475 * The parent has N vm_object_pip_add() references prior to 1476 * calling us and will remove references for rtvals[] that are 1477 * not set to VM_PAGER_PEND. We need to remove the rest on I/O 1478 * completion. 1479 * 1480 * The parent has soft-busy'd the pages it passes us and will unbusy 1481 * those whose rtvals[] entry is not set to VM_PAGER_PEND on return. 1482 * We need to unbusy the rest on I/O completion. 1483 */ 1484 static void 1485 swap_pager_putpages(vm_object_t object, vm_page_t *ma, int count, 1486 int flags, int *rtvals) 1487 { 1488 struct buf *bp; 1489 daddr_t addr, blk, n_free, s_free; 1490 vm_page_t mreq; 1491 int i, j, n; 1492 bool async; 1493 1494 KASSERT(count == 0 || ma[0]->object == object, 1495 ("%s: object mismatch %p/%p", 1496 __func__, object, ma[0]->object)); 1497 1498 VM_OBJECT_WUNLOCK(object); 1499 async = curproc == pageproc && (flags & VM_PAGER_PUT_SYNC) == 0; 1500 swp_pager_init_freerange(&s_free, &n_free); 1501 1502 /* 1503 * Assign swap blocks and issue I/O. We reallocate swap on the fly. 1504 * The page is left dirty until the pageout operation completes 1505 * successfully. 1506 */ 1507 for (i = 0; i < count; i += n) { 1508 /* Maximum I/O size is limited by maximum swap block size. */ 1509 n = min(count - i, nsw_cluster_max); 1510 1511 if (async) { 1512 mtx_lock(&swbuf_mtx); 1513 while (nsw_wcount_async == 0) 1514 msleep(&nsw_wcount_async, &swbuf_mtx, PVM, 1515 "swbufa", 0); 1516 nsw_wcount_async--; 1517 mtx_unlock(&swbuf_mtx); 1518 } 1519 1520 /* Get a block of swap of size up to size n. */ 1521 blk = swp_pager_getswapspace(&n); 1522 if (blk == SWAPBLK_NONE) { 1523 mtx_lock(&swbuf_mtx); 1524 if (++nsw_wcount_async == 1) 1525 wakeup(&nsw_wcount_async); 1526 mtx_unlock(&swbuf_mtx); 1527 for (j = 0; j < n; ++j) 1528 rtvals[i + j] = VM_PAGER_FAIL; 1529 continue; 1530 } 1531 VM_OBJECT_WLOCK(object); 1532 for (j = 0; j < n; ++j) { 1533 mreq = ma[i + j]; 1534 vm_page_aflag_clear(mreq, PGA_SWAP_FREE); 1535 addr = swp_pager_meta_build(mreq->object, mreq->pindex, 1536 blk + j); 1537 if (addr != SWAPBLK_NONE) 1538 swp_pager_update_freerange(&s_free, &n_free, 1539 addr); 1540 MPASS(mreq->dirty == VM_PAGE_BITS_ALL); 1541 mreq->oflags |= VPO_SWAPINPROG; 1542 } 1543 VM_OBJECT_WUNLOCK(object); 1544 1545 bp = uma_zalloc(swwbuf_zone, M_WAITOK); 1546 MPASS((bp->b_flags & B_MAXPHYS) != 0); 1547 if (async) 1548 bp->b_flags |= B_ASYNC; 1549 bp->b_flags |= B_PAGING; 1550 bp->b_iocmd = BIO_WRITE; 1551 1552 bp->b_rcred = crhold(thread0.td_ucred); 1553 bp->b_wcred = crhold(thread0.td_ucred); 1554 bp->b_bcount = PAGE_SIZE * n; 1555 bp->b_bufsize = PAGE_SIZE * n; 1556 bp->b_blkno = blk; 1557 for (j = 0; j < n; j++) 1558 bp->b_pages[j] = ma[i + j]; 1559 bp->b_npages = n; 1560 1561 /* 1562 * Must set dirty range for NFS to work. 1563 */ 1564 bp->b_dirtyoff = 0; 1565 bp->b_dirtyend = bp->b_bcount; 1566 1567 VM_CNT_INC(v_swapout); 1568 VM_CNT_ADD(v_swappgsout, bp->b_npages); 1569 1570 /* 1571 * We unconditionally set rtvals[] to VM_PAGER_PEND so that we 1572 * can call the async completion routine at the end of a 1573 * synchronous I/O operation. Otherwise, our caller would 1574 * perform duplicate unbusy and wakeup operations on the page 1575 * and object, respectively. 1576 */ 1577 for (j = 0; j < n; j++) 1578 rtvals[i + j] = VM_PAGER_PEND; 1579 1580 /* 1581 * asynchronous 1582 * 1583 * NOTE: b_blkno is destroyed by the call to swapdev_strategy. 1584 */ 1585 if (async) { 1586 bp->b_iodone = swp_pager_async_iodone; 1587 BUF_KERNPROC(bp); 1588 swp_pager_strategy(bp); 1589 continue; 1590 } 1591 1592 /* 1593 * synchronous 1594 * 1595 * NOTE: b_blkno is destroyed by the call to swapdev_strategy. 1596 */ 1597 bp->b_iodone = bdone; 1598 swp_pager_strategy(bp); 1599 1600 /* 1601 * Wait for the sync I/O to complete. 1602 */ 1603 bwait(bp, PVM, "swwrt"); 1604 1605 /* 1606 * Now that we are through with the bp, we can call the 1607 * normal async completion, which frees everything up. 1608 */ 1609 swp_pager_async_iodone(bp); 1610 } 1611 swp_pager_freeswapspace(s_free, n_free); 1612 VM_OBJECT_WLOCK(object); 1613 } 1614 1615 /* 1616 * swp_pager_async_iodone: 1617 * 1618 * Completion routine for asynchronous reads and writes from/to swap. 1619 * Also called manually by synchronous code to finish up a bp. 1620 * 1621 * This routine may not sleep. 1622 */ 1623 static void 1624 swp_pager_async_iodone(struct buf *bp) 1625 { 1626 int i; 1627 vm_object_t object = NULL; 1628 1629 /* 1630 * Report error - unless we ran out of memory, in which case 1631 * we've already logged it in swapgeom_strategy(). 1632 */ 1633 if (bp->b_ioflags & BIO_ERROR && bp->b_error != ENOMEM) { 1634 printf( 1635 "swap_pager: I/O error - %s failed; blkno %ld," 1636 "size %ld, error %d\n", 1637 ((bp->b_iocmd == BIO_READ) ? "pagein" : "pageout"), 1638 (long)bp->b_blkno, 1639 (long)bp->b_bcount, 1640 bp->b_error 1641 ); 1642 } 1643 1644 /* 1645 * remove the mapping for kernel virtual 1646 */ 1647 if (buf_mapped(bp)) 1648 pmap_qremove((vm_offset_t)bp->b_data, bp->b_npages); 1649 else 1650 bp->b_data = bp->b_kvabase; 1651 1652 if (bp->b_npages) { 1653 object = bp->b_pages[0]->object; 1654 VM_OBJECT_WLOCK(object); 1655 } 1656 1657 /* 1658 * cleanup pages. If an error occurs writing to swap, we are in 1659 * very serious trouble. If it happens to be a disk error, though, 1660 * we may be able to recover by reassigning the swap later on. So 1661 * in this case we remove the m->swapblk assignment for the page 1662 * but do not free it in the rlist. The errornous block(s) are thus 1663 * never reallocated as swap. Redirty the page and continue. 1664 */ 1665 for (i = 0; i < bp->b_npages; ++i) { 1666 vm_page_t m = bp->b_pages[i]; 1667 1668 m->oflags &= ~VPO_SWAPINPROG; 1669 if (m->oflags & VPO_SWAPSLEEP) { 1670 m->oflags &= ~VPO_SWAPSLEEP; 1671 wakeup(&object->handle); 1672 } 1673 1674 /* We always have space after I/O, successful or not. */ 1675 vm_page_aflag_set(m, PGA_SWAP_SPACE); 1676 1677 if (bp->b_ioflags & BIO_ERROR) { 1678 /* 1679 * If an error occurs I'd love to throw the swapblk 1680 * away without freeing it back to swapspace, so it 1681 * can never be used again. But I can't from an 1682 * interrupt. 1683 */ 1684 if (bp->b_iocmd == BIO_READ) { 1685 /* 1686 * NOTE: for reads, m->dirty will probably 1687 * be overridden by the original caller of 1688 * getpages so don't play cute tricks here. 1689 */ 1690 vm_page_invalid(m); 1691 } else { 1692 /* 1693 * If a write error occurs, reactivate page 1694 * so it doesn't clog the inactive list, 1695 * then finish the I/O. 1696 */ 1697 MPASS(m->dirty == VM_PAGE_BITS_ALL); 1698 1699 /* PQ_UNSWAPPABLE? */ 1700 vm_page_activate(m); 1701 vm_page_sunbusy(m); 1702 } 1703 } else if (bp->b_iocmd == BIO_READ) { 1704 /* 1705 * NOTE: for reads, m->dirty will probably be 1706 * overridden by the original caller of getpages so 1707 * we cannot set them in order to free the underlying 1708 * swap in a low-swap situation. I don't think we'd 1709 * want to do that anyway, but it was an optimization 1710 * that existed in the old swapper for a time before 1711 * it got ripped out due to precisely this problem. 1712 */ 1713 KASSERT(!pmap_page_is_mapped(m), 1714 ("swp_pager_async_iodone: page %p is mapped", m)); 1715 KASSERT(m->dirty == 0, 1716 ("swp_pager_async_iodone: page %p is dirty", m)); 1717 1718 vm_page_valid(m); 1719 if (i < bp->b_pgbefore || 1720 i >= bp->b_npages - bp->b_pgafter) 1721 vm_page_readahead_finish(m); 1722 } else { 1723 /* 1724 * For write success, clear the dirty 1725 * status, then finish the I/O ( which decrements the 1726 * busy count and possibly wakes waiter's up ). 1727 * A page is only written to swap after a period of 1728 * inactivity. Therefore, we do not expect it to be 1729 * reused. 1730 */ 1731 KASSERT(!pmap_page_is_write_mapped(m), 1732 ("swp_pager_async_iodone: page %p is not write" 1733 " protected", m)); 1734 vm_page_undirty(m); 1735 vm_page_deactivate_noreuse(m); 1736 vm_page_sunbusy(m); 1737 } 1738 } 1739 1740 /* 1741 * adjust pip. NOTE: the original parent may still have its own 1742 * pip refs on the object. 1743 */ 1744 if (object != NULL) { 1745 vm_object_pip_wakeupn(object, bp->b_npages); 1746 VM_OBJECT_WUNLOCK(object); 1747 } 1748 1749 /* 1750 * swapdev_strategy() manually sets b_vp and b_bufobj before calling 1751 * bstrategy(). Set them back to NULL now we're done with it, or we'll 1752 * trigger a KASSERT in relpbuf(). 1753 */ 1754 if (bp->b_vp) { 1755 bp->b_vp = NULL; 1756 bp->b_bufobj = NULL; 1757 } 1758 /* 1759 * release the physical I/O buffer 1760 */ 1761 if (bp->b_flags & B_ASYNC) { 1762 mtx_lock(&swbuf_mtx); 1763 if (++nsw_wcount_async == 1) 1764 wakeup(&nsw_wcount_async); 1765 mtx_unlock(&swbuf_mtx); 1766 } 1767 uma_zfree((bp->b_iocmd == BIO_READ) ? swrbuf_zone : swwbuf_zone, bp); 1768 } 1769 1770 int 1771 swap_pager_nswapdev(void) 1772 { 1773 1774 return (nswapdev); 1775 } 1776 1777 static void 1778 swp_pager_force_dirty(vm_page_t m) 1779 { 1780 1781 vm_page_dirty(m); 1782 swap_pager_unswapped(m); 1783 vm_page_launder(m); 1784 } 1785 1786 u_long 1787 swap_pager_swapped_pages(vm_object_t object) 1788 { 1789 struct swblk *sb; 1790 vm_pindex_t pi; 1791 u_long res; 1792 int i; 1793 1794 VM_OBJECT_ASSERT_LOCKED(object); 1795 1796 if (pctrie_is_empty(&object->un_pager.swp.swp_blks)) 1797 return (0); 1798 1799 for (res = 0, pi = 0; (sb = SWAP_PCTRIE_LOOKUP_GE( 1800 &object->un_pager.swp.swp_blks, pi)) != NULL; 1801 pi = sb->p + SWAP_META_PAGES) { 1802 for (i = 0; i < SWAP_META_PAGES; i++) { 1803 if (sb->d[i] != SWAPBLK_NONE) 1804 res++; 1805 } 1806 } 1807 return (res); 1808 } 1809 1810 /* 1811 * swap_pager_swapoff_object: 1812 * 1813 * Page in all of the pages that have been paged out for an object 1814 * to a swap device. 1815 */ 1816 static void 1817 swap_pager_swapoff_object(struct swdevt *sp, vm_object_t object) 1818 { 1819 struct swblk *sb; 1820 vm_page_t m; 1821 vm_pindex_t pi; 1822 daddr_t blk; 1823 int i, nv, rahead, rv; 1824 1825 KASSERT((object->flags & OBJ_SWAP) != 0, 1826 ("%s: Object not swappable", __func__)); 1827 1828 for (pi = 0; (sb = SWAP_PCTRIE_LOOKUP_GE( 1829 &object->un_pager.swp.swp_blks, pi)) != NULL; ) { 1830 if ((object->flags & OBJ_DEAD) != 0) { 1831 /* 1832 * Make sure that pending writes finish before 1833 * returning. 1834 */ 1835 vm_object_pip_wait(object, "swpoff"); 1836 swp_pager_meta_free_all(object); 1837 break; 1838 } 1839 for (i = 0; i < SWAP_META_PAGES; i++) { 1840 /* 1841 * Count the number of contiguous valid blocks. 1842 */ 1843 for (nv = 0; nv < SWAP_META_PAGES - i; nv++) { 1844 blk = sb->d[i + nv]; 1845 if (!swp_pager_isondev(blk, sp) || 1846 blk == SWAPBLK_NONE) 1847 break; 1848 } 1849 if (nv == 0) 1850 continue; 1851 1852 /* 1853 * Look for a page corresponding to the first 1854 * valid block and ensure that any pending paging 1855 * operations on it are complete. If the page is valid, 1856 * mark it dirty and free the swap block. Try to batch 1857 * this operation since it may cause sp to be freed, 1858 * meaning that we must restart the scan. Avoid busying 1859 * valid pages since we may block forever on kernel 1860 * stack pages. 1861 */ 1862 m = vm_page_lookup(object, sb->p + i); 1863 if (m == NULL) { 1864 m = vm_page_alloc(object, sb->p + i, 1865 VM_ALLOC_NORMAL | VM_ALLOC_WAITFAIL); 1866 if (m == NULL) 1867 break; 1868 } else { 1869 if ((m->oflags & VPO_SWAPINPROG) != 0) { 1870 m->oflags |= VPO_SWAPSLEEP; 1871 VM_OBJECT_SLEEP(object, &object->handle, 1872 PSWP, "swpoff", 0); 1873 break; 1874 } 1875 if (vm_page_all_valid(m)) { 1876 do { 1877 swp_pager_force_dirty(m); 1878 } while (--nv > 0 && 1879 (m = vm_page_next(m)) != NULL && 1880 vm_page_all_valid(m) && 1881 (m->oflags & VPO_SWAPINPROG) == 0); 1882 break; 1883 } 1884 if (!vm_page_busy_acquire(m, VM_ALLOC_WAITFAIL)) 1885 break; 1886 } 1887 1888 vm_object_pip_add(object, 1); 1889 rahead = SWAP_META_PAGES; 1890 rv = swap_pager_getpages_locked(object, &m, 1, NULL, 1891 &rahead); 1892 if (rv != VM_PAGER_OK) 1893 panic("%s: read from swap failed: %d", 1894 __func__, rv); 1895 vm_object_pip_wakeupn(object, 1); 1896 VM_OBJECT_WLOCK(object); 1897 vm_page_xunbusy(m); 1898 1899 /* 1900 * The object lock was dropped so we must restart the 1901 * scan of this swap block. Pages paged in during this 1902 * iteration will be marked dirty in a future iteration. 1903 */ 1904 break; 1905 } 1906 if (i == SWAP_META_PAGES) 1907 pi = sb->p + SWAP_META_PAGES; 1908 } 1909 } 1910 1911 /* 1912 * swap_pager_swapoff: 1913 * 1914 * Page in all of the pages that have been paged out to the 1915 * given device. The corresponding blocks in the bitmap must be 1916 * marked as allocated and the device must be flagged SW_CLOSING. 1917 * There may be no processes swapped out to the device. 1918 * 1919 * This routine may block. 1920 */ 1921 static void 1922 swap_pager_swapoff(struct swdevt *sp) 1923 { 1924 vm_object_t object; 1925 int retries; 1926 1927 sx_assert(&swdev_syscall_lock, SA_XLOCKED); 1928 1929 retries = 0; 1930 full_rescan: 1931 mtx_lock(&vm_object_list_mtx); 1932 TAILQ_FOREACH(object, &vm_object_list, object_list) { 1933 if ((object->flags & OBJ_SWAP) == 0) 1934 continue; 1935 mtx_unlock(&vm_object_list_mtx); 1936 /* Depends on type-stability. */ 1937 VM_OBJECT_WLOCK(object); 1938 1939 /* 1940 * Dead objects are eventually terminated on their own. 1941 */ 1942 if ((object->flags & OBJ_DEAD) != 0) 1943 goto next_obj; 1944 1945 /* 1946 * Sync with fences placed after pctrie 1947 * initialization. We must not access pctrie below 1948 * unless we checked that our object is swap and not 1949 * dead. 1950 */ 1951 atomic_thread_fence_acq(); 1952 if ((object->flags & OBJ_SWAP) == 0) 1953 goto next_obj; 1954 1955 swap_pager_swapoff_object(sp, object); 1956 next_obj: 1957 VM_OBJECT_WUNLOCK(object); 1958 mtx_lock(&vm_object_list_mtx); 1959 } 1960 mtx_unlock(&vm_object_list_mtx); 1961 1962 if (sp->sw_used) { 1963 /* 1964 * Objects may be locked or paging to the device being 1965 * removed, so we will miss their pages and need to 1966 * make another pass. We have marked this device as 1967 * SW_CLOSING, so the activity should finish soon. 1968 */ 1969 retries++; 1970 if (retries > 100) { 1971 panic("swapoff: failed to locate %d swap blocks", 1972 sp->sw_used); 1973 } 1974 pause("swpoff", hz / 20); 1975 goto full_rescan; 1976 } 1977 EVENTHANDLER_INVOKE(swapoff, sp); 1978 } 1979 1980 /************************************************************************ 1981 * SWAP META DATA * 1982 ************************************************************************ 1983 * 1984 * These routines manipulate the swap metadata stored in the 1985 * OBJT_SWAP object. 1986 * 1987 * Swap metadata is implemented with a global hash and not directly 1988 * linked into the object. Instead the object simply contains 1989 * appropriate tracking counters. 1990 */ 1991 1992 /* 1993 * SWP_PAGER_SWBLK_EMPTY() - is a range of blocks free? 1994 */ 1995 static bool 1996 swp_pager_swblk_empty(struct swblk *sb, int start, int limit) 1997 { 1998 int i; 1999 2000 MPASS(0 <= start && start <= limit && limit <= SWAP_META_PAGES); 2001 for (i = start; i < limit; i++) { 2002 if (sb->d[i] != SWAPBLK_NONE) 2003 return (false); 2004 } 2005 return (true); 2006 } 2007 2008 /* 2009 * SWP_PAGER_FREE_EMPTY_SWBLK() - frees if a block is free 2010 * 2011 * Nothing is done if the block is still in use. 2012 */ 2013 static void 2014 swp_pager_free_empty_swblk(vm_object_t object, struct swblk *sb) 2015 { 2016 2017 if (swp_pager_swblk_empty(sb, 0, SWAP_META_PAGES)) { 2018 SWAP_PCTRIE_REMOVE(&object->un_pager.swp.swp_blks, sb->p); 2019 uma_zfree(swblk_zone, sb); 2020 } 2021 } 2022 2023 /* 2024 * SWP_PAGER_META_BUILD() - add swap block to swap meta data for object 2025 * 2026 * The specified swapblk is added to the object's swap metadata. If 2027 * the swapblk is not valid, it is freed instead. Any previously 2028 * assigned swapblk is returned. 2029 */ 2030 static daddr_t 2031 swp_pager_meta_build(vm_object_t object, vm_pindex_t pindex, daddr_t swapblk) 2032 { 2033 static volatile int swblk_zone_exhausted, swpctrie_zone_exhausted; 2034 struct swblk *sb, *sb1; 2035 vm_pindex_t modpi, rdpi; 2036 daddr_t prev_swapblk; 2037 int error, i; 2038 2039 VM_OBJECT_ASSERT_WLOCKED(object); 2040 2041 rdpi = rounddown(pindex, SWAP_META_PAGES); 2042 sb = SWAP_PCTRIE_LOOKUP(&object->un_pager.swp.swp_blks, rdpi); 2043 if (sb == NULL) { 2044 if (swapblk == SWAPBLK_NONE) 2045 return (SWAPBLK_NONE); 2046 for (;;) { 2047 sb = uma_zalloc(swblk_zone, M_NOWAIT | (curproc == 2048 pageproc ? M_USE_RESERVE : 0)); 2049 if (sb != NULL) { 2050 sb->p = rdpi; 2051 for (i = 0; i < SWAP_META_PAGES; i++) 2052 sb->d[i] = SWAPBLK_NONE; 2053 if (atomic_cmpset_int(&swblk_zone_exhausted, 2054 1, 0)) 2055 printf("swblk zone ok\n"); 2056 break; 2057 } 2058 VM_OBJECT_WUNLOCK(object); 2059 if (uma_zone_exhausted(swblk_zone)) { 2060 if (atomic_cmpset_int(&swblk_zone_exhausted, 2061 0, 1)) 2062 printf("swap blk zone exhausted, " 2063 "increase kern.maxswzone\n"); 2064 vm_pageout_oom(VM_OOM_SWAPZ); 2065 pause("swzonxb", 10); 2066 } else 2067 uma_zwait(swblk_zone); 2068 VM_OBJECT_WLOCK(object); 2069 sb = SWAP_PCTRIE_LOOKUP(&object->un_pager.swp.swp_blks, 2070 rdpi); 2071 if (sb != NULL) 2072 /* 2073 * Somebody swapped out a nearby page, 2074 * allocating swblk at the rdpi index, 2075 * while we dropped the object lock. 2076 */ 2077 goto allocated; 2078 } 2079 for (;;) { 2080 error = SWAP_PCTRIE_INSERT( 2081 &object->un_pager.swp.swp_blks, sb); 2082 if (error == 0) { 2083 if (atomic_cmpset_int(&swpctrie_zone_exhausted, 2084 1, 0)) 2085 printf("swpctrie zone ok\n"); 2086 break; 2087 } 2088 VM_OBJECT_WUNLOCK(object); 2089 if (uma_zone_exhausted(swpctrie_zone)) { 2090 if (atomic_cmpset_int(&swpctrie_zone_exhausted, 2091 0, 1)) 2092 printf("swap pctrie zone exhausted, " 2093 "increase kern.maxswzone\n"); 2094 vm_pageout_oom(VM_OOM_SWAPZ); 2095 pause("swzonxp", 10); 2096 } else 2097 uma_zwait(swpctrie_zone); 2098 VM_OBJECT_WLOCK(object); 2099 sb1 = SWAP_PCTRIE_LOOKUP(&object->un_pager.swp.swp_blks, 2100 rdpi); 2101 if (sb1 != NULL) { 2102 uma_zfree(swblk_zone, sb); 2103 sb = sb1; 2104 goto allocated; 2105 } 2106 } 2107 } 2108 allocated: 2109 MPASS(sb->p == rdpi); 2110 2111 modpi = pindex % SWAP_META_PAGES; 2112 /* Return prior contents of metadata. */ 2113 prev_swapblk = sb->d[modpi]; 2114 /* Enter block into metadata. */ 2115 sb->d[modpi] = swapblk; 2116 2117 /* 2118 * Free the swblk if we end up with the empty page run. 2119 */ 2120 if (swapblk == SWAPBLK_NONE) 2121 swp_pager_free_empty_swblk(object, sb); 2122 return (prev_swapblk); 2123 } 2124 2125 /* 2126 * SWP_PAGER_META_TRANSFER() - free a range of blocks in the srcobject's swap 2127 * metadata, or transfer it into dstobject. 2128 * 2129 * This routine will free swap metadata structures as they are cleaned 2130 * out. 2131 */ 2132 static void 2133 swp_pager_meta_transfer(vm_object_t srcobject, vm_object_t dstobject, 2134 vm_pindex_t pindex, vm_pindex_t count, vm_size_t *moved) 2135 { 2136 struct swblk *sb; 2137 vm_page_t m; 2138 daddr_t n_free, s_free; 2139 vm_pindex_t offset, last; 2140 vm_size_t mc; 2141 int i, limit, start; 2142 2143 VM_OBJECT_ASSERT_WLOCKED(srcobject); 2144 MPASS(moved == NULL || dstobject == NULL); 2145 2146 mc = 0; 2147 m = NULL; 2148 if (count == 0 || pctrie_is_empty(&srcobject->un_pager.swp.swp_blks)) 2149 goto out; 2150 2151 swp_pager_init_freerange(&s_free, &n_free); 2152 offset = pindex; 2153 last = pindex + count; 2154 for (;;) { 2155 sb = SWAP_PCTRIE_LOOKUP_GE(&srcobject->un_pager.swp.swp_blks, 2156 rounddown(pindex, SWAP_META_PAGES)); 2157 if (sb == NULL || sb->p >= last) 2158 break; 2159 start = pindex > sb->p ? pindex - sb->p : 0; 2160 limit = last - sb->p < SWAP_META_PAGES ? last - sb->p : 2161 SWAP_META_PAGES; 2162 for (i = start; i < limit; i++) { 2163 if (sb->d[i] == SWAPBLK_NONE) 2164 continue; 2165 if (dstobject == NULL || 2166 !swp_pager_xfer_source(srcobject, dstobject, 2167 sb->p + i - offset, sb->d[i])) { 2168 swp_pager_update_freerange(&s_free, &n_free, 2169 sb->d[i]); 2170 } 2171 if (moved != NULL) { 2172 if (m != NULL && m->pindex != pindex + i - 1) 2173 m = NULL; 2174 m = m != NULL ? vm_page_next(m) : 2175 vm_page_lookup(srcobject, pindex + i); 2176 if (m == NULL || vm_page_none_valid(m)) 2177 mc++; 2178 } 2179 sb->d[i] = SWAPBLK_NONE; 2180 } 2181 pindex = sb->p + SWAP_META_PAGES; 2182 if (swp_pager_swblk_empty(sb, 0, start) && 2183 swp_pager_swblk_empty(sb, limit, SWAP_META_PAGES)) { 2184 SWAP_PCTRIE_REMOVE(&srcobject->un_pager.swp.swp_blks, 2185 sb->p); 2186 uma_zfree(swblk_zone, sb); 2187 } 2188 } 2189 swp_pager_freeswapspace(s_free, n_free); 2190 out: 2191 if (moved != NULL) 2192 *moved = mc; 2193 } 2194 2195 /* 2196 * SWP_PAGER_META_FREE() - free a range of blocks in the object's swap metadata 2197 * 2198 * The requested range of blocks is freed, with any associated swap 2199 * returned to the swap bitmap. 2200 * 2201 * This routine will free swap metadata structures as they are cleaned 2202 * out. This routine does *NOT* operate on swap metadata associated 2203 * with resident pages. 2204 */ 2205 static void 2206 swp_pager_meta_free(vm_object_t object, vm_pindex_t pindex, vm_pindex_t count, 2207 vm_size_t *freed) 2208 { 2209 swp_pager_meta_transfer(object, NULL, pindex, count, freed); 2210 } 2211 2212 /* 2213 * SWP_PAGER_META_FREE_ALL() - destroy all swap metadata associated with object 2214 * 2215 * This routine locates and destroys all swap metadata associated with 2216 * an object. 2217 */ 2218 static void 2219 swp_pager_meta_free_all(vm_object_t object) 2220 { 2221 struct swblk *sb; 2222 daddr_t n_free, s_free; 2223 vm_pindex_t pindex; 2224 int i; 2225 2226 VM_OBJECT_ASSERT_WLOCKED(object); 2227 2228 if (pctrie_is_empty(&object->un_pager.swp.swp_blks)) 2229 return; 2230 2231 swp_pager_init_freerange(&s_free, &n_free); 2232 for (pindex = 0; (sb = SWAP_PCTRIE_LOOKUP_GE( 2233 &object->un_pager.swp.swp_blks, pindex)) != NULL;) { 2234 pindex = sb->p + SWAP_META_PAGES; 2235 for (i = 0; i < SWAP_META_PAGES; i++) { 2236 if (sb->d[i] == SWAPBLK_NONE) 2237 continue; 2238 swp_pager_update_freerange(&s_free, &n_free, sb->d[i]); 2239 } 2240 SWAP_PCTRIE_REMOVE(&object->un_pager.swp.swp_blks, sb->p); 2241 uma_zfree(swblk_zone, sb); 2242 } 2243 swp_pager_freeswapspace(s_free, n_free); 2244 } 2245 2246 /* 2247 * SWP_PAGER_METACTL() - misc control of swap meta data. 2248 * 2249 * This routine is capable of looking up, or removing swapblk 2250 * assignments in the swap meta data. It returns the swapblk being 2251 * looked-up, popped, or SWAPBLK_NONE if the block was invalid. 2252 * 2253 * When acting on a busy resident page and paging is in progress, we 2254 * have to wait until paging is complete but otherwise can act on the 2255 * busy page. 2256 */ 2257 static daddr_t 2258 swp_pager_meta_lookup(vm_object_t object, vm_pindex_t pindex) 2259 { 2260 struct swblk *sb; 2261 2262 VM_OBJECT_ASSERT_LOCKED(object); 2263 2264 /* 2265 * The meta data only exists if the object is OBJT_SWAP 2266 * and even then might not be allocated yet. 2267 */ 2268 KASSERT((object->flags & OBJ_SWAP) != 0, 2269 ("Lookup object not swappable")); 2270 2271 sb = SWAP_PCTRIE_LOOKUP(&object->un_pager.swp.swp_blks, 2272 rounddown(pindex, SWAP_META_PAGES)); 2273 if (sb == NULL) 2274 return (SWAPBLK_NONE); 2275 return (sb->d[pindex % SWAP_META_PAGES]); 2276 } 2277 2278 /* 2279 * Returns the least page index which is greater than or equal to the 2280 * parameter pindex and for which there is a swap block allocated. 2281 * Returns object's size if the object's type is not swap or if there 2282 * are no allocated swap blocks for the object after the requested 2283 * pindex. 2284 */ 2285 vm_pindex_t 2286 swap_pager_find_least(vm_object_t object, vm_pindex_t pindex) 2287 { 2288 struct swblk *sb; 2289 int i; 2290 2291 VM_OBJECT_ASSERT_LOCKED(object); 2292 MPASS((object->flags & OBJ_SWAP) != 0); 2293 2294 if (pctrie_is_empty(&object->un_pager.swp.swp_blks)) 2295 return (object->size); 2296 sb = SWAP_PCTRIE_LOOKUP_GE(&object->un_pager.swp.swp_blks, 2297 rounddown(pindex, SWAP_META_PAGES)); 2298 if (sb == NULL) 2299 return (object->size); 2300 if (sb->p < pindex) { 2301 for (i = pindex % SWAP_META_PAGES; i < SWAP_META_PAGES; i++) { 2302 if (sb->d[i] != SWAPBLK_NONE) 2303 return (sb->p + i); 2304 } 2305 sb = SWAP_PCTRIE_LOOKUP_GE(&object->un_pager.swp.swp_blks, 2306 roundup(pindex, SWAP_META_PAGES)); 2307 if (sb == NULL) 2308 return (object->size); 2309 } 2310 for (i = 0; i < SWAP_META_PAGES; i++) { 2311 if (sb->d[i] != SWAPBLK_NONE) 2312 return (sb->p + i); 2313 } 2314 2315 /* 2316 * We get here if a swblk is present in the trie but it 2317 * doesn't map any blocks. 2318 */ 2319 MPASS(0); 2320 return (object->size); 2321 } 2322 2323 /* 2324 * System call swapon(name) enables swapping on device name, 2325 * which must be in the swdevsw. Return EBUSY 2326 * if already swapping on this device. 2327 */ 2328 #ifndef _SYS_SYSPROTO_H_ 2329 struct swapon_args { 2330 char *name; 2331 }; 2332 #endif 2333 2334 int 2335 sys_swapon(struct thread *td, struct swapon_args *uap) 2336 { 2337 struct vattr attr; 2338 struct vnode *vp; 2339 struct nameidata nd; 2340 int error; 2341 2342 error = priv_check(td, PRIV_SWAPON); 2343 if (error) 2344 return (error); 2345 2346 sx_xlock(&swdev_syscall_lock); 2347 2348 /* 2349 * Swap metadata may not fit in the KVM if we have physical 2350 * memory of >1GB. 2351 */ 2352 if (swblk_zone == NULL) { 2353 error = ENOMEM; 2354 goto done; 2355 } 2356 2357 NDINIT(&nd, LOOKUP, ISOPEN | FOLLOW | LOCKLEAF | AUDITVNODE1, 2358 UIO_USERSPACE, uap->name); 2359 error = namei(&nd); 2360 if (error) 2361 goto done; 2362 2363 NDFREE_PNBUF(&nd); 2364 vp = nd.ni_vp; 2365 2366 if (vn_isdisk_error(vp, &error)) { 2367 error = swapongeom(vp); 2368 } else if (vp->v_type == VREG && 2369 (vp->v_mount->mnt_vfc->vfc_flags & VFCF_NETWORK) != 0 && 2370 (error = VOP_GETATTR(vp, &attr, td->td_ucred)) == 0) { 2371 /* 2372 * Allow direct swapping to NFS regular files in the same 2373 * way that nfs_mountroot() sets up diskless swapping. 2374 */ 2375 error = swaponvp(td, vp, attr.va_size / DEV_BSIZE); 2376 } 2377 2378 if (error != 0) 2379 vput(vp); 2380 else 2381 VOP_UNLOCK(vp); 2382 done: 2383 sx_xunlock(&swdev_syscall_lock); 2384 return (error); 2385 } 2386 2387 /* 2388 * Check that the total amount of swap currently configured does not 2389 * exceed half the theoretical maximum. If it does, print a warning 2390 * message. 2391 */ 2392 static void 2393 swapon_check_swzone(void) 2394 { 2395 2396 /* recommend using no more than half that amount */ 2397 if (swap_total > swap_maxpages / 2) { 2398 printf("warning: total configured swap (%lu pages) " 2399 "exceeds maximum recommended amount (%lu pages).\n", 2400 swap_total, swap_maxpages / 2); 2401 printf("warning: increase kern.maxswzone " 2402 "or reduce amount of swap.\n"); 2403 } 2404 } 2405 2406 static void 2407 swaponsomething(struct vnode *vp, void *id, u_long nblks, 2408 sw_strategy_t *strategy, sw_close_t *close, dev_t dev, int flags) 2409 { 2410 struct swdevt *sp, *tsp; 2411 daddr_t dvbase; 2412 2413 /* 2414 * nblks is in DEV_BSIZE'd chunks, convert to PAGE_SIZE'd chunks. 2415 * First chop nblks off to page-align it, then convert. 2416 * 2417 * sw->sw_nblks is in page-sized chunks now too. 2418 */ 2419 nblks &= ~(ctodb(1) - 1); 2420 nblks = dbtoc(nblks); 2421 2422 sp = malloc(sizeof *sp, M_VMPGDATA, M_WAITOK | M_ZERO); 2423 sp->sw_blist = blist_create(nblks, M_WAITOK); 2424 sp->sw_vp = vp; 2425 sp->sw_id = id; 2426 sp->sw_dev = dev; 2427 sp->sw_nblks = nblks; 2428 sp->sw_used = 0; 2429 sp->sw_strategy = strategy; 2430 sp->sw_close = close; 2431 sp->sw_flags = flags; 2432 2433 /* 2434 * Do not free the first blocks in order to avoid overwriting 2435 * any bsd label at the front of the partition 2436 */ 2437 blist_free(sp->sw_blist, howmany(BBSIZE, PAGE_SIZE), 2438 nblks - howmany(BBSIZE, PAGE_SIZE)); 2439 2440 dvbase = 0; 2441 mtx_lock(&sw_dev_mtx); 2442 TAILQ_FOREACH(tsp, &swtailq, sw_list) { 2443 if (tsp->sw_end >= dvbase) { 2444 /* 2445 * We put one uncovered page between the devices 2446 * in order to definitively prevent any cross-device 2447 * I/O requests 2448 */ 2449 dvbase = tsp->sw_end + 1; 2450 } 2451 } 2452 sp->sw_first = dvbase; 2453 sp->sw_end = dvbase + nblks; 2454 TAILQ_INSERT_TAIL(&swtailq, sp, sw_list); 2455 nswapdev++; 2456 swap_pager_avail += nblks - howmany(BBSIZE, PAGE_SIZE); 2457 swap_total += nblks; 2458 swapon_check_swzone(); 2459 swp_sizecheck(); 2460 mtx_unlock(&sw_dev_mtx); 2461 EVENTHANDLER_INVOKE(swapon, sp); 2462 } 2463 2464 /* 2465 * SYSCALL: swapoff(devname) 2466 * 2467 * Disable swapping on the given device. 2468 * 2469 * XXX: Badly designed system call: it should use a device index 2470 * rather than filename as specification. We keep sw_vp around 2471 * only to make this work. 2472 */ 2473 static int 2474 kern_swapoff(struct thread *td, const char *name, enum uio_seg name_seg, 2475 u_int flags) 2476 { 2477 struct vnode *vp; 2478 struct nameidata nd; 2479 struct swdevt *sp; 2480 int error; 2481 2482 error = priv_check(td, PRIV_SWAPOFF); 2483 if (error != 0) 2484 return (error); 2485 if ((flags & ~(SWAPOFF_FORCE)) != 0) 2486 return (EINVAL); 2487 2488 sx_xlock(&swdev_syscall_lock); 2489 2490 NDINIT(&nd, LOOKUP, FOLLOW | AUDITVNODE1, name_seg, name); 2491 error = namei(&nd); 2492 if (error) 2493 goto done; 2494 NDFREE_PNBUF(&nd); 2495 vp = nd.ni_vp; 2496 2497 mtx_lock(&sw_dev_mtx); 2498 TAILQ_FOREACH(sp, &swtailq, sw_list) { 2499 if (sp->sw_vp == vp) 2500 break; 2501 } 2502 mtx_unlock(&sw_dev_mtx); 2503 if (sp == NULL) { 2504 error = EINVAL; 2505 goto done; 2506 } 2507 error = swapoff_one(sp, td->td_ucred, flags); 2508 done: 2509 sx_xunlock(&swdev_syscall_lock); 2510 return (error); 2511 } 2512 2513 2514 #ifdef COMPAT_FREEBSD13 2515 int 2516 freebsd13_swapoff(struct thread *td, struct freebsd13_swapoff_args *uap) 2517 { 2518 return (kern_swapoff(td, uap->name, UIO_USERSPACE, 0)); 2519 } 2520 #endif 2521 2522 int 2523 sys_swapoff(struct thread *td, struct swapoff_args *uap) 2524 { 2525 return (kern_swapoff(td, uap->name, UIO_USERSPACE, uap->flags)); 2526 } 2527 2528 static int 2529 swapoff_one(struct swdevt *sp, struct ucred *cred, u_int flags) 2530 { 2531 u_long nblks; 2532 #ifdef MAC 2533 int error; 2534 #endif 2535 2536 sx_assert(&swdev_syscall_lock, SA_XLOCKED); 2537 #ifdef MAC 2538 (void) vn_lock(sp->sw_vp, LK_EXCLUSIVE | LK_RETRY); 2539 error = mac_system_check_swapoff(cred, sp->sw_vp); 2540 (void) VOP_UNLOCK(sp->sw_vp); 2541 if (error != 0) 2542 return (error); 2543 #endif 2544 nblks = sp->sw_nblks; 2545 2546 /* 2547 * We can turn off this swap device safely only if the 2548 * available virtual memory in the system will fit the amount 2549 * of data we will have to page back in, plus an epsilon so 2550 * the system doesn't become critically low on swap space. 2551 * The vm_free_count() part does not account e.g. for clean 2552 * pages that can be immediately reclaimed without paging, so 2553 * this is a very rough estimation. 2554 * 2555 * On the other hand, not turning swap off on swapoff_all() 2556 * means that we can lose swap data when filesystems go away, 2557 * which is arguably worse. 2558 */ 2559 if ((flags & SWAPOFF_FORCE) == 0 && 2560 vm_free_count() + swap_pager_avail < nblks + nswap_lowat) 2561 return (ENOMEM); 2562 2563 /* 2564 * Prevent further allocations on this device. 2565 */ 2566 mtx_lock(&sw_dev_mtx); 2567 sp->sw_flags |= SW_CLOSING; 2568 swap_pager_avail -= blist_fill(sp->sw_blist, 0, nblks); 2569 swap_total -= nblks; 2570 mtx_unlock(&sw_dev_mtx); 2571 2572 /* 2573 * Page in the contents of the device and close it. 2574 */ 2575 swap_pager_swapoff(sp); 2576 2577 sp->sw_close(curthread, sp); 2578 mtx_lock(&sw_dev_mtx); 2579 sp->sw_id = NULL; 2580 TAILQ_REMOVE(&swtailq, sp, sw_list); 2581 nswapdev--; 2582 if (nswapdev == 0) { 2583 swap_pager_full = 2; 2584 swap_pager_almost_full = 1; 2585 } 2586 if (swdevhd == sp) 2587 swdevhd = NULL; 2588 mtx_unlock(&sw_dev_mtx); 2589 blist_destroy(sp->sw_blist); 2590 free(sp, M_VMPGDATA); 2591 return (0); 2592 } 2593 2594 void 2595 swapoff_all(void) 2596 { 2597 struct swdevt *sp, *spt; 2598 const char *devname; 2599 int error; 2600 2601 sx_xlock(&swdev_syscall_lock); 2602 2603 mtx_lock(&sw_dev_mtx); 2604 TAILQ_FOREACH_SAFE(sp, &swtailq, sw_list, spt) { 2605 mtx_unlock(&sw_dev_mtx); 2606 if (vn_isdisk(sp->sw_vp)) 2607 devname = devtoname(sp->sw_vp->v_rdev); 2608 else 2609 devname = "[file]"; 2610 error = swapoff_one(sp, thread0.td_ucred, SWAPOFF_FORCE); 2611 if (error != 0) { 2612 printf("Cannot remove swap device %s (error=%d), " 2613 "skipping.\n", devname, error); 2614 } else if (bootverbose) { 2615 printf("Swap device %s removed.\n", devname); 2616 } 2617 mtx_lock(&sw_dev_mtx); 2618 } 2619 mtx_unlock(&sw_dev_mtx); 2620 2621 sx_xunlock(&swdev_syscall_lock); 2622 } 2623 2624 void 2625 swap_pager_status(int *total, int *used) 2626 { 2627 2628 *total = swap_total; 2629 *used = swap_total - swap_pager_avail - 2630 nswapdev * howmany(BBSIZE, PAGE_SIZE); 2631 } 2632 2633 int 2634 swap_dev_info(int name, struct xswdev *xs, char *devname, size_t len) 2635 { 2636 struct swdevt *sp; 2637 const char *tmp_devname; 2638 int error, n; 2639 2640 n = 0; 2641 error = ENOENT; 2642 mtx_lock(&sw_dev_mtx); 2643 TAILQ_FOREACH(sp, &swtailq, sw_list) { 2644 if (n != name) { 2645 n++; 2646 continue; 2647 } 2648 xs->xsw_version = XSWDEV_VERSION; 2649 xs->xsw_dev = sp->sw_dev; 2650 xs->xsw_flags = sp->sw_flags; 2651 xs->xsw_nblks = sp->sw_nblks; 2652 xs->xsw_used = sp->sw_used; 2653 if (devname != NULL) { 2654 if (vn_isdisk(sp->sw_vp)) 2655 tmp_devname = devtoname(sp->sw_vp->v_rdev); 2656 else 2657 tmp_devname = "[file]"; 2658 strncpy(devname, tmp_devname, len); 2659 } 2660 error = 0; 2661 break; 2662 } 2663 mtx_unlock(&sw_dev_mtx); 2664 return (error); 2665 } 2666 2667 #if defined(COMPAT_FREEBSD11) 2668 #define XSWDEV_VERSION_11 1 2669 struct xswdev11 { 2670 u_int xsw_version; 2671 uint32_t xsw_dev; 2672 int xsw_flags; 2673 int xsw_nblks; 2674 int xsw_used; 2675 }; 2676 #endif 2677 2678 #if defined(__amd64__) && defined(COMPAT_FREEBSD32) 2679 struct xswdev32 { 2680 u_int xsw_version; 2681 u_int xsw_dev1, xsw_dev2; 2682 int xsw_flags; 2683 int xsw_nblks; 2684 int xsw_used; 2685 }; 2686 #endif 2687 2688 static int 2689 sysctl_vm_swap_info(SYSCTL_HANDLER_ARGS) 2690 { 2691 struct xswdev xs; 2692 #if defined(__amd64__) && defined(COMPAT_FREEBSD32) 2693 struct xswdev32 xs32; 2694 #endif 2695 #if defined(COMPAT_FREEBSD11) 2696 struct xswdev11 xs11; 2697 #endif 2698 int error; 2699 2700 if (arg2 != 1) /* name length */ 2701 return (EINVAL); 2702 2703 memset(&xs, 0, sizeof(xs)); 2704 error = swap_dev_info(*(int *)arg1, &xs, NULL, 0); 2705 if (error != 0) 2706 return (error); 2707 #if defined(__amd64__) && defined(COMPAT_FREEBSD32) 2708 if (req->oldlen == sizeof(xs32)) { 2709 memset(&xs32, 0, sizeof(xs32)); 2710 xs32.xsw_version = XSWDEV_VERSION; 2711 xs32.xsw_dev1 = xs.xsw_dev; 2712 xs32.xsw_dev2 = xs.xsw_dev >> 32; 2713 xs32.xsw_flags = xs.xsw_flags; 2714 xs32.xsw_nblks = xs.xsw_nblks; 2715 xs32.xsw_used = xs.xsw_used; 2716 error = SYSCTL_OUT(req, &xs32, sizeof(xs32)); 2717 return (error); 2718 } 2719 #endif 2720 #if defined(COMPAT_FREEBSD11) 2721 if (req->oldlen == sizeof(xs11)) { 2722 memset(&xs11, 0, sizeof(xs11)); 2723 xs11.xsw_version = XSWDEV_VERSION_11; 2724 xs11.xsw_dev = xs.xsw_dev; /* truncation */ 2725 xs11.xsw_flags = xs.xsw_flags; 2726 xs11.xsw_nblks = xs.xsw_nblks; 2727 xs11.xsw_used = xs.xsw_used; 2728 error = SYSCTL_OUT(req, &xs11, sizeof(xs11)); 2729 return (error); 2730 } 2731 #endif 2732 error = SYSCTL_OUT(req, &xs, sizeof(xs)); 2733 return (error); 2734 } 2735 2736 SYSCTL_INT(_vm, OID_AUTO, nswapdev, CTLFLAG_RD, &nswapdev, 0, 2737 "Number of swap devices"); 2738 SYSCTL_NODE(_vm, OID_AUTO, swap_info, CTLFLAG_RD | CTLFLAG_MPSAFE, 2739 sysctl_vm_swap_info, 2740 "Swap statistics by device"); 2741 2742 /* 2743 * Count the approximate swap usage in pages for a vmspace. The 2744 * shadowed or not yet copied on write swap blocks are not accounted. 2745 * The map must be locked. 2746 */ 2747 long 2748 vmspace_swap_count(struct vmspace *vmspace) 2749 { 2750 vm_map_t map; 2751 vm_map_entry_t cur; 2752 vm_object_t object; 2753 struct swblk *sb; 2754 vm_pindex_t e, pi; 2755 long count; 2756 int i; 2757 2758 map = &vmspace->vm_map; 2759 count = 0; 2760 2761 VM_MAP_ENTRY_FOREACH(cur, map) { 2762 if ((cur->eflags & MAP_ENTRY_IS_SUB_MAP) != 0) 2763 continue; 2764 object = cur->object.vm_object; 2765 if (object == NULL || (object->flags & OBJ_SWAP) == 0) 2766 continue; 2767 VM_OBJECT_RLOCK(object); 2768 if ((object->flags & OBJ_SWAP) == 0) 2769 goto unlock; 2770 pi = OFF_TO_IDX(cur->offset); 2771 e = pi + OFF_TO_IDX(cur->end - cur->start); 2772 for (;; pi = sb->p + SWAP_META_PAGES) { 2773 sb = SWAP_PCTRIE_LOOKUP_GE( 2774 &object->un_pager.swp.swp_blks, pi); 2775 if (sb == NULL || sb->p >= e) 2776 break; 2777 for (i = 0; i < SWAP_META_PAGES; i++) { 2778 if (sb->p + i < e && 2779 sb->d[i] != SWAPBLK_NONE) 2780 count++; 2781 } 2782 } 2783 unlock: 2784 VM_OBJECT_RUNLOCK(object); 2785 } 2786 return (count); 2787 } 2788 2789 /* 2790 * GEOM backend 2791 * 2792 * Swapping onto disk devices. 2793 * 2794 */ 2795 2796 static g_orphan_t swapgeom_orphan; 2797 2798 static struct g_class g_swap_class = { 2799 .name = "SWAP", 2800 .version = G_VERSION, 2801 .orphan = swapgeom_orphan, 2802 }; 2803 2804 DECLARE_GEOM_CLASS(g_swap_class, g_class); 2805 2806 static void 2807 swapgeom_close_ev(void *arg, int flags) 2808 { 2809 struct g_consumer *cp; 2810 2811 cp = arg; 2812 g_access(cp, -1, -1, 0); 2813 g_detach(cp); 2814 g_destroy_consumer(cp); 2815 } 2816 2817 /* 2818 * Add a reference to the g_consumer for an inflight transaction. 2819 */ 2820 static void 2821 swapgeom_acquire(struct g_consumer *cp) 2822 { 2823 2824 mtx_assert(&sw_dev_mtx, MA_OWNED); 2825 cp->index++; 2826 } 2827 2828 /* 2829 * Remove a reference from the g_consumer. Post a close event if all 2830 * references go away, since the function might be called from the 2831 * biodone context. 2832 */ 2833 static void 2834 swapgeom_release(struct g_consumer *cp, struct swdevt *sp) 2835 { 2836 2837 mtx_assert(&sw_dev_mtx, MA_OWNED); 2838 cp->index--; 2839 if (cp->index == 0) { 2840 if (g_post_event(swapgeom_close_ev, cp, M_NOWAIT, NULL) == 0) 2841 sp->sw_id = NULL; 2842 } 2843 } 2844 2845 static void 2846 swapgeom_done(struct bio *bp2) 2847 { 2848 struct swdevt *sp; 2849 struct buf *bp; 2850 struct g_consumer *cp; 2851 2852 bp = bp2->bio_caller2; 2853 cp = bp2->bio_from; 2854 bp->b_ioflags = bp2->bio_flags; 2855 if (bp2->bio_error) 2856 bp->b_ioflags |= BIO_ERROR; 2857 bp->b_resid = bp->b_bcount - bp2->bio_completed; 2858 bp->b_error = bp2->bio_error; 2859 bp->b_caller1 = NULL; 2860 bufdone(bp); 2861 sp = bp2->bio_caller1; 2862 mtx_lock(&sw_dev_mtx); 2863 swapgeom_release(cp, sp); 2864 mtx_unlock(&sw_dev_mtx); 2865 g_destroy_bio(bp2); 2866 } 2867 2868 static void 2869 swapgeom_strategy(struct buf *bp, struct swdevt *sp) 2870 { 2871 struct bio *bio; 2872 struct g_consumer *cp; 2873 2874 mtx_lock(&sw_dev_mtx); 2875 cp = sp->sw_id; 2876 if (cp == NULL) { 2877 mtx_unlock(&sw_dev_mtx); 2878 bp->b_error = ENXIO; 2879 bp->b_ioflags |= BIO_ERROR; 2880 bufdone(bp); 2881 return; 2882 } 2883 swapgeom_acquire(cp); 2884 mtx_unlock(&sw_dev_mtx); 2885 if (bp->b_iocmd == BIO_WRITE) 2886 bio = g_new_bio(); 2887 else 2888 bio = g_alloc_bio(); 2889 if (bio == NULL) { 2890 mtx_lock(&sw_dev_mtx); 2891 swapgeom_release(cp, sp); 2892 mtx_unlock(&sw_dev_mtx); 2893 bp->b_error = ENOMEM; 2894 bp->b_ioflags |= BIO_ERROR; 2895 printf("swap_pager: cannot allocate bio\n"); 2896 bufdone(bp); 2897 return; 2898 } 2899 2900 bp->b_caller1 = bio; 2901 bio->bio_caller1 = sp; 2902 bio->bio_caller2 = bp; 2903 bio->bio_cmd = bp->b_iocmd; 2904 bio->bio_offset = (bp->b_blkno - sp->sw_first) * PAGE_SIZE; 2905 bio->bio_length = bp->b_bcount; 2906 bio->bio_done = swapgeom_done; 2907 bio->bio_flags |= BIO_SWAP; 2908 if (!buf_mapped(bp)) { 2909 bio->bio_ma = bp->b_pages; 2910 bio->bio_data = unmapped_buf; 2911 bio->bio_ma_offset = (vm_offset_t)bp->b_offset & PAGE_MASK; 2912 bio->bio_ma_n = bp->b_npages; 2913 bio->bio_flags |= BIO_UNMAPPED; 2914 } else { 2915 bio->bio_data = bp->b_data; 2916 bio->bio_ma = NULL; 2917 } 2918 g_io_request(bio, cp); 2919 return; 2920 } 2921 2922 static void 2923 swapgeom_orphan(struct g_consumer *cp) 2924 { 2925 struct swdevt *sp; 2926 int destroy; 2927 2928 mtx_lock(&sw_dev_mtx); 2929 TAILQ_FOREACH(sp, &swtailq, sw_list) { 2930 if (sp->sw_id == cp) { 2931 sp->sw_flags |= SW_CLOSING; 2932 break; 2933 } 2934 } 2935 /* 2936 * Drop reference we were created with. Do directly since we're in a 2937 * special context where we don't have to queue the call to 2938 * swapgeom_close_ev(). 2939 */ 2940 cp->index--; 2941 destroy = ((sp != NULL) && (cp->index == 0)); 2942 if (destroy) 2943 sp->sw_id = NULL; 2944 mtx_unlock(&sw_dev_mtx); 2945 if (destroy) 2946 swapgeom_close_ev(cp, 0); 2947 } 2948 2949 static void 2950 swapgeom_close(struct thread *td, struct swdevt *sw) 2951 { 2952 struct g_consumer *cp; 2953 2954 mtx_lock(&sw_dev_mtx); 2955 cp = sw->sw_id; 2956 sw->sw_id = NULL; 2957 mtx_unlock(&sw_dev_mtx); 2958 2959 /* 2960 * swapgeom_close() may be called from the biodone context, 2961 * where we cannot perform topology changes. Delegate the 2962 * work to the events thread. 2963 */ 2964 if (cp != NULL) 2965 g_waitfor_event(swapgeom_close_ev, cp, M_WAITOK, NULL); 2966 } 2967 2968 static int 2969 swapongeom_locked(struct cdev *dev, struct vnode *vp) 2970 { 2971 struct g_provider *pp; 2972 struct g_consumer *cp; 2973 static struct g_geom *gp; 2974 struct swdevt *sp; 2975 u_long nblks; 2976 int error; 2977 2978 pp = g_dev_getprovider(dev); 2979 if (pp == NULL) 2980 return (ENODEV); 2981 mtx_lock(&sw_dev_mtx); 2982 TAILQ_FOREACH(sp, &swtailq, sw_list) { 2983 cp = sp->sw_id; 2984 if (cp != NULL && cp->provider == pp) { 2985 mtx_unlock(&sw_dev_mtx); 2986 return (EBUSY); 2987 } 2988 } 2989 mtx_unlock(&sw_dev_mtx); 2990 if (gp == NULL) 2991 gp = g_new_geomf(&g_swap_class, "swap"); 2992 cp = g_new_consumer(gp); 2993 cp->index = 1; /* Number of active I/Os, plus one for being active. */ 2994 cp->flags |= G_CF_DIRECT_SEND | G_CF_DIRECT_RECEIVE; 2995 g_attach(cp, pp); 2996 /* 2997 * XXX: Every time you think you can improve the margin for 2998 * footshooting, somebody depends on the ability to do so: 2999 * savecore(8) wants to write to our swapdev so we cannot 3000 * set an exclusive count :-( 3001 */ 3002 error = g_access(cp, 1, 1, 0); 3003 if (error != 0) { 3004 g_detach(cp); 3005 g_destroy_consumer(cp); 3006 return (error); 3007 } 3008 nblks = pp->mediasize / DEV_BSIZE; 3009 swaponsomething(vp, cp, nblks, swapgeom_strategy, 3010 swapgeom_close, dev2udev(dev), 3011 (pp->flags & G_PF_ACCEPT_UNMAPPED) != 0 ? SW_UNMAPPED : 0); 3012 return (0); 3013 } 3014 3015 static int 3016 swapongeom(struct vnode *vp) 3017 { 3018 int error; 3019 3020 ASSERT_VOP_ELOCKED(vp, "swapongeom"); 3021 if (vp->v_type != VCHR || VN_IS_DOOMED(vp)) { 3022 error = ENOENT; 3023 } else { 3024 g_topology_lock(); 3025 error = swapongeom_locked(vp->v_rdev, vp); 3026 g_topology_unlock(); 3027 } 3028 return (error); 3029 } 3030 3031 /* 3032 * VNODE backend 3033 * 3034 * This is used mainly for network filesystem (read: probably only tested 3035 * with NFS) swapfiles. 3036 * 3037 */ 3038 3039 static void 3040 swapdev_strategy(struct buf *bp, struct swdevt *sp) 3041 { 3042 struct vnode *vp2; 3043 3044 bp->b_blkno = ctodb(bp->b_blkno - sp->sw_first); 3045 3046 vp2 = sp->sw_id; 3047 vhold(vp2); 3048 if (bp->b_iocmd == BIO_WRITE) { 3049 vn_lock(vp2, LK_EXCLUSIVE | LK_RETRY); 3050 if (bp->b_bufobj) 3051 bufobj_wdrop(bp->b_bufobj); 3052 bufobj_wref(&vp2->v_bufobj); 3053 } else { 3054 vn_lock(vp2, LK_SHARED | LK_RETRY); 3055 } 3056 if (bp->b_bufobj != &vp2->v_bufobj) 3057 bp->b_bufobj = &vp2->v_bufobj; 3058 bp->b_vp = vp2; 3059 bp->b_iooffset = dbtob(bp->b_blkno); 3060 bstrategy(bp); 3061 VOP_UNLOCK(vp2); 3062 } 3063 3064 static void 3065 swapdev_close(struct thread *td, struct swdevt *sp) 3066 { 3067 struct vnode *vp; 3068 3069 vp = sp->sw_vp; 3070 vn_lock(vp, LK_EXCLUSIVE | LK_RETRY); 3071 VOP_CLOSE(vp, FREAD | FWRITE, td->td_ucred, td); 3072 vput(vp); 3073 } 3074 3075 static int 3076 swaponvp(struct thread *td, struct vnode *vp, u_long nblks) 3077 { 3078 struct swdevt *sp; 3079 int error; 3080 3081 ASSERT_VOP_ELOCKED(vp, "swaponvp"); 3082 if (nblks == 0) 3083 return (ENXIO); 3084 mtx_lock(&sw_dev_mtx); 3085 TAILQ_FOREACH(sp, &swtailq, sw_list) { 3086 if (sp->sw_id == vp) { 3087 mtx_unlock(&sw_dev_mtx); 3088 return (EBUSY); 3089 } 3090 } 3091 mtx_unlock(&sw_dev_mtx); 3092 3093 #ifdef MAC 3094 error = mac_system_check_swapon(td->td_ucred, vp); 3095 if (error == 0) 3096 #endif 3097 error = VOP_OPEN(vp, FREAD | FWRITE, td->td_ucred, td, NULL); 3098 if (error != 0) 3099 return (error); 3100 3101 swaponsomething(vp, vp, nblks, swapdev_strategy, swapdev_close, 3102 NODEV, 0); 3103 return (0); 3104 } 3105 3106 static int 3107 sysctl_swap_async_max(SYSCTL_HANDLER_ARGS) 3108 { 3109 int error, new, n; 3110 3111 new = nsw_wcount_async_max; 3112 error = sysctl_handle_int(oidp, &new, 0, req); 3113 if (error != 0 || req->newptr == NULL) 3114 return (error); 3115 3116 if (new > nswbuf / 2 || new < 1) 3117 return (EINVAL); 3118 3119 mtx_lock(&swbuf_mtx); 3120 while (nsw_wcount_async_max != new) { 3121 /* 3122 * Adjust difference. If the current async count is too low, 3123 * we will need to sqeeze our update slowly in. Sleep with a 3124 * higher priority than getpbuf() to finish faster. 3125 */ 3126 n = new - nsw_wcount_async_max; 3127 if (nsw_wcount_async + n >= 0) { 3128 nsw_wcount_async += n; 3129 nsw_wcount_async_max += n; 3130 wakeup(&nsw_wcount_async); 3131 } else { 3132 nsw_wcount_async_max -= nsw_wcount_async; 3133 nsw_wcount_async = 0; 3134 msleep(&nsw_wcount_async, &swbuf_mtx, PSWP, 3135 "swpsysctl", 0); 3136 } 3137 } 3138 mtx_unlock(&swbuf_mtx); 3139 3140 return (0); 3141 } 3142 3143 static void 3144 swap_pager_update_writecount(vm_object_t object, vm_offset_t start, 3145 vm_offset_t end) 3146 { 3147 3148 VM_OBJECT_WLOCK(object); 3149 KASSERT((object->flags & OBJ_ANON) == 0, 3150 ("Splittable object with writecount")); 3151 object->un_pager.swp.writemappings += (vm_ooffset_t)end - start; 3152 VM_OBJECT_WUNLOCK(object); 3153 } 3154 3155 static void 3156 swap_pager_release_writecount(vm_object_t object, vm_offset_t start, 3157 vm_offset_t end) 3158 { 3159 3160 VM_OBJECT_WLOCK(object); 3161 KASSERT((object->flags & OBJ_ANON) == 0, 3162 ("Splittable object with writecount")); 3163 object->un_pager.swp.writemappings -= (vm_ooffset_t)end - start; 3164 VM_OBJECT_WUNLOCK(object); 3165 } 3166