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 #include "opt_vm.h" 73 74 #include <sys/param.h> 75 #include <sys/bio.h> 76 #include <sys/blist.h> 77 #include <sys/buf.h> 78 #include <sys/conf.h> 79 #include <sys/disk.h> 80 #include <sys/disklabel.h> 81 #include <sys/eventhandler.h> 82 #include <sys/fcntl.h> 83 #include <sys/limits.h> 84 #include <sys/lock.h> 85 #include <sys/kernel.h> 86 #include <sys/mount.h> 87 #include <sys/namei.h> 88 #include <sys/malloc.h> 89 #include <sys/pctrie.h> 90 #include <sys/priv.h> 91 #include <sys/proc.h> 92 #include <sys/racct.h> 93 #include <sys/resource.h> 94 #include <sys/resourcevar.h> 95 #include <sys/rwlock.h> 96 #include <sys/sbuf.h> 97 #include <sys/sysctl.h> 98 #include <sys/sysproto.h> 99 #include <sys/systm.h> 100 #include <sys/sx.h> 101 #include <sys/unistd.h> 102 #include <sys/user.h> 103 #include <sys/vmmeter.h> 104 #include <sys/vnode.h> 105 106 #include <security/mac/mac_framework.h> 107 108 #include <vm/vm.h> 109 #include <vm/pmap.h> 110 #include <vm/vm_map.h> 111 #include <vm/vm_kern.h> 112 #include <vm/vm_object.h> 113 #include <vm/vm_page.h> 114 #include <vm/vm_pager.h> 115 #include <vm/vm_pageout.h> 116 #include <vm/vm_param.h> 117 #include <vm/swap_pager.h> 118 #include <vm/vm_extern.h> 119 #include <vm/uma.h> 120 121 #include <geom/geom.h> 122 123 /* 124 * MAX_PAGEOUT_CLUSTER must be a power of 2 between 1 and 64. 125 * The 64-page limit is due to the radix code (kern/subr_blist.c). 126 */ 127 #ifndef MAX_PAGEOUT_CLUSTER 128 #define MAX_PAGEOUT_CLUSTER 32 129 #endif 130 131 #if !defined(SWB_NPAGES) 132 #define SWB_NPAGES MAX_PAGEOUT_CLUSTER 133 #endif 134 135 #define SWAP_META_PAGES PCTRIE_COUNT 136 137 /* 138 * A swblk structure maps each page index within a 139 * SWAP_META_PAGES-aligned and sized range to the address of an 140 * on-disk swap block (or SWAPBLK_NONE). The collection of these 141 * mappings for an entire vm object is implemented as a pc-trie. 142 */ 143 struct swblk { 144 vm_pindex_t p; 145 daddr_t d[SWAP_META_PAGES]; 146 }; 147 148 static MALLOC_DEFINE(M_VMPGDATA, "vm_pgdata", "swap pager private data"); 149 static struct mtx sw_dev_mtx; 150 static TAILQ_HEAD(, swdevt) swtailq = TAILQ_HEAD_INITIALIZER(swtailq); 151 static struct swdevt *swdevhd; /* Allocate from here next */ 152 static int nswapdev; /* Number of swap devices */ 153 int swap_pager_avail; 154 static struct sx swdev_syscall_lock; /* serialize swap(on|off) */ 155 156 static __exclusive_cache_line u_long swap_reserved; 157 static u_long swap_total; 158 static int sysctl_page_shift(SYSCTL_HANDLER_ARGS); 159 160 static SYSCTL_NODE(_vm_stats, OID_AUTO, swap, CTLFLAG_RD | CTLFLAG_MPSAFE, 0, 161 "VM swap stats"); 162 163 SYSCTL_PROC(_vm, OID_AUTO, swap_reserved, CTLTYPE_U64 | CTLFLAG_RD | CTLFLAG_MPSAFE, 164 &swap_reserved, 0, sysctl_page_shift, "QU", 165 "Amount of swap storage needed to back all allocated anonymous memory."); 166 SYSCTL_PROC(_vm, OID_AUTO, swap_total, CTLTYPE_U64 | CTLFLAG_RD | CTLFLAG_MPSAFE, 167 &swap_total, 0, sysctl_page_shift, "QU", 168 "Total amount of available swap storage."); 169 170 int vm_overcommit __read_mostly = 0; 171 SYSCTL_INT(_vm, VM_OVERCOMMIT, overcommit, CTLFLAG_RW, &vm_overcommit, 0, 172 "Configure virtual memory overcommit behavior. See tuning(7) " 173 "for details."); 174 static unsigned long swzone; 175 SYSCTL_ULONG(_vm, OID_AUTO, swzone, CTLFLAG_RD, &swzone, 0, 176 "Actual size of swap metadata zone"); 177 static unsigned long swap_maxpages; 178 SYSCTL_ULONG(_vm, OID_AUTO, swap_maxpages, CTLFLAG_RD, &swap_maxpages, 0, 179 "Maximum amount of swap supported"); 180 181 static COUNTER_U64_DEFINE_EARLY(swap_free_deferred); 182 SYSCTL_COUNTER_U64(_vm_stats_swap, OID_AUTO, free_deferred, 183 CTLFLAG_RD, &swap_free_deferred, 184 "Number of pages that deferred freeing swap space"); 185 186 static COUNTER_U64_DEFINE_EARLY(swap_free_completed); 187 SYSCTL_COUNTER_U64(_vm_stats_swap, OID_AUTO, free_completed, 188 CTLFLAG_RD, &swap_free_completed, 189 "Number of deferred frees completed"); 190 191 static int 192 sysctl_page_shift(SYSCTL_HANDLER_ARGS) 193 { 194 uint64_t newval; 195 u_long value = *(u_long *)arg1; 196 197 newval = ((uint64_t)value) << PAGE_SHIFT; 198 return (sysctl_handle_64(oidp, &newval, 0, req)); 199 } 200 201 static bool 202 swap_reserve_by_cred_rlimit(u_long pincr, struct ucred *cred, int oc) 203 { 204 struct uidinfo *uip; 205 u_long prev; 206 207 uip = cred->cr_ruidinfo; 208 209 prev = atomic_fetchadd_long(&uip->ui_vmsize, pincr); 210 if ((oc & SWAP_RESERVE_RLIMIT_ON) != 0 && 211 prev + pincr > lim_cur(curthread, RLIMIT_SWAP) && 212 priv_check(curthread, PRIV_VM_SWAP_NORLIMIT) != 0) { 213 prev = atomic_fetchadd_long(&uip->ui_vmsize, -pincr); 214 KASSERT(prev >= pincr, 215 ("negative vmsize for uid %d\n", uip->ui_uid)); 216 return (false); 217 } 218 return (true); 219 } 220 221 static void 222 swap_release_by_cred_rlimit(u_long pdecr, struct ucred *cred) 223 { 224 struct uidinfo *uip; 225 #ifdef INVARIANTS 226 u_long prev; 227 #endif 228 229 uip = cred->cr_ruidinfo; 230 231 #ifdef INVARIANTS 232 prev = atomic_fetchadd_long(&uip->ui_vmsize, -pdecr); 233 KASSERT(prev >= pdecr, 234 ("negative vmsize for uid %d\n", uip->ui_uid)); 235 #else 236 atomic_subtract_long(&uip->ui_vmsize, pdecr); 237 #endif 238 } 239 240 static void 241 swap_reserve_force_rlimit(u_long pincr, struct ucred *cred) 242 { 243 struct uidinfo *uip; 244 245 uip = cred->cr_ruidinfo; 246 atomic_add_long(&uip->ui_vmsize, pincr); 247 } 248 249 bool 250 swap_reserve(vm_ooffset_t incr) 251 { 252 253 return (swap_reserve_by_cred(incr, curthread->td_ucred)); 254 } 255 256 bool 257 swap_reserve_by_cred(vm_ooffset_t incr, struct ucred *cred) 258 { 259 u_long r, s, prev, pincr; 260 #ifdef RACCT 261 int error; 262 #endif 263 int oc; 264 static int curfail; 265 static struct timeval lastfail; 266 267 KASSERT((incr & PAGE_MASK) == 0, ("%s: incr: %ju & PAGE_MASK", 268 __func__, (uintmax_t)incr)); 269 270 #ifdef RACCT 271 if (RACCT_ENABLED()) { 272 PROC_LOCK(curproc); 273 error = racct_add(curproc, RACCT_SWAP, incr); 274 PROC_UNLOCK(curproc); 275 if (error != 0) 276 return (false); 277 } 278 #endif 279 280 pincr = atop(incr); 281 prev = atomic_fetchadd_long(&swap_reserved, pincr); 282 r = prev + pincr; 283 s = swap_total; 284 oc = atomic_load_int(&vm_overcommit); 285 if (r > s && (oc & SWAP_RESERVE_ALLOW_NONWIRED) != 0) { 286 s += vm_cnt.v_page_count - vm_cnt.v_free_reserved - 287 vm_wire_count(); 288 } 289 if ((oc & SWAP_RESERVE_FORCE_ON) != 0 && r > s && 290 priv_check(curthread, PRIV_VM_SWAP_NOQUOTA) != 0) { 291 prev = atomic_fetchadd_long(&swap_reserved, -pincr); 292 KASSERT(prev >= pincr, 293 ("swap_reserved < incr on overcommit fail")); 294 goto out_error; 295 } 296 297 if (!swap_reserve_by_cred_rlimit(pincr, cred, oc)) { 298 prev = atomic_fetchadd_long(&swap_reserved, -pincr); 299 KASSERT(prev >= pincr, 300 ("swap_reserved < incr on overcommit fail")); 301 goto out_error; 302 } 303 304 return (true); 305 306 out_error: 307 if (ppsratecheck(&lastfail, &curfail, 1)) { 308 printf("uid %d, pid %d: swap reservation " 309 "for %jd bytes failed\n", 310 cred->cr_ruidinfo->ui_uid, curproc->p_pid, incr); 311 } 312 #ifdef RACCT 313 if (RACCT_ENABLED()) { 314 PROC_LOCK(curproc); 315 racct_sub(curproc, RACCT_SWAP, incr); 316 PROC_UNLOCK(curproc); 317 } 318 #endif 319 320 return (false); 321 } 322 323 void 324 swap_reserve_force(vm_ooffset_t incr) 325 { 326 u_long pincr; 327 328 KASSERT((incr & PAGE_MASK) == 0, ("%s: incr: %ju & PAGE_MASK", 329 __func__, (uintmax_t)incr)); 330 331 #ifdef RACCT 332 if (RACCT_ENABLED()) { 333 PROC_LOCK(curproc); 334 racct_add_force(curproc, RACCT_SWAP, incr); 335 PROC_UNLOCK(curproc); 336 } 337 #endif 338 pincr = atop(incr); 339 atomic_add_long(&swap_reserved, pincr); 340 swap_reserve_force_rlimit(pincr, curthread->td_ucred); 341 } 342 343 void 344 swap_release(vm_ooffset_t decr) 345 { 346 struct ucred *cred; 347 348 PROC_LOCK(curproc); 349 cred = curproc->p_ucred; 350 swap_release_by_cred(decr, cred); 351 PROC_UNLOCK(curproc); 352 } 353 354 void 355 swap_release_by_cred(vm_ooffset_t decr, struct ucred *cred) 356 { 357 u_long pdecr; 358 #ifdef INVARIANTS 359 u_long prev; 360 #endif 361 362 KASSERT((decr & PAGE_MASK) == 0, ("%s: decr: %ju & PAGE_MASK", 363 __func__, (uintmax_t)decr)); 364 365 pdecr = atop(decr); 366 #ifdef INVARIANTS 367 prev = atomic_fetchadd_long(&swap_reserved, -pdecr); 368 KASSERT(prev >= pdecr, ("swap_reserved < decr")); 369 #else 370 atomic_subtract_long(&swap_reserved, pdecr); 371 #endif 372 373 swap_release_by_cred_rlimit(pdecr, cred); 374 #ifdef RACCT 375 if (racct_enable) 376 racct_sub_cred(cred, RACCT_SWAP, decr); 377 #endif 378 } 379 380 static int swap_pager_full = 2; /* swap space exhaustion (task killing) */ 381 static int swap_pager_almost_full = 1; /* swap space exhaustion (w/hysteresis)*/ 382 static struct mtx swbuf_mtx; /* to sync nsw_wcount_async */ 383 static int nsw_wcount_async; /* limit async write buffers */ 384 static int nsw_wcount_async_max;/* assigned maximum */ 385 int nsw_cluster_max; /* maximum VOP I/O allowed */ 386 387 static int sysctl_swap_async_max(SYSCTL_HANDLER_ARGS); 388 SYSCTL_PROC(_vm, OID_AUTO, swap_async_max, CTLTYPE_INT | CTLFLAG_RW | 389 CTLFLAG_MPSAFE, NULL, 0, sysctl_swap_async_max, "I", 390 "Maximum running async swap ops"); 391 static int sysctl_swap_fragmentation(SYSCTL_HANDLER_ARGS); 392 SYSCTL_PROC(_vm, OID_AUTO, swap_fragmentation, CTLTYPE_STRING | CTLFLAG_RD | 393 CTLFLAG_MPSAFE, NULL, 0, sysctl_swap_fragmentation, "A", 394 "Swap Fragmentation Info"); 395 396 static struct sx sw_alloc_sx; 397 398 /* 399 * "named" and "unnamed" anon region objects. Try to reduce the overhead 400 * of searching a named list by hashing it just a little. 401 */ 402 403 #define NOBJLISTS 8 404 405 #define NOBJLIST(handle) \ 406 (&swap_pager_object_list[((int)(intptr_t)handle >> 4) & (NOBJLISTS-1)]) 407 408 static struct pagerlst swap_pager_object_list[NOBJLISTS]; 409 static uma_zone_t swwbuf_zone; 410 static uma_zone_t swrbuf_zone; 411 static uma_zone_t swblk_zone; 412 static uma_zone_t swpctrie_zone; 413 414 /* 415 * pagerops for OBJT_SWAP - "swap pager". Some ops are also global procedure 416 * calls hooked from other parts of the VM system and do not appear here. 417 * (see vm/swap_pager.h). 418 */ 419 static vm_object_t 420 swap_pager_alloc(void *handle, vm_ooffset_t size, 421 vm_prot_t prot, vm_ooffset_t offset, struct ucred *); 422 static void swap_pager_dealloc(vm_object_t object); 423 static int swap_pager_getpages(vm_object_t, vm_page_t *, int, int *, 424 int *); 425 static int swap_pager_getpages_async(vm_object_t, vm_page_t *, int, int *, 426 int *, pgo_getpages_iodone_t, void *); 427 static void swap_pager_putpages(vm_object_t, vm_page_t *, int, int, int *); 428 static boolean_t 429 swap_pager_haspage(vm_object_t object, vm_pindex_t pindex, int *before, int *after); 430 static void swap_pager_init(void); 431 static void swap_pager_unswapped(vm_page_t); 432 static void swap_pager_swapoff(struct swdevt *sp); 433 static void swap_pager_update_writecount(vm_object_t object, 434 vm_offset_t start, vm_offset_t end); 435 static void swap_pager_release_writecount(vm_object_t object, 436 vm_offset_t start, vm_offset_t end); 437 static void swap_pager_freespace_pgo(vm_object_t object, vm_pindex_t start, 438 vm_size_t size); 439 440 const struct pagerops swappagerops = { 441 .pgo_kvme_type = KVME_TYPE_SWAP, 442 .pgo_init = swap_pager_init, /* early system initialization of pager */ 443 .pgo_alloc = swap_pager_alloc, /* allocate an OBJT_SWAP object */ 444 .pgo_dealloc = swap_pager_dealloc, /* deallocate an OBJT_SWAP object */ 445 .pgo_getpages = swap_pager_getpages, /* pagein */ 446 .pgo_getpages_async = swap_pager_getpages_async, /* pagein (async) */ 447 .pgo_putpages = swap_pager_putpages, /* pageout */ 448 .pgo_haspage = swap_pager_haspage, /* get backing store status for page */ 449 .pgo_pageunswapped = swap_pager_unswapped, /* remove swap related to page */ 450 .pgo_update_writecount = swap_pager_update_writecount, 451 .pgo_release_writecount = swap_pager_release_writecount, 452 .pgo_freespace = swap_pager_freespace_pgo, 453 }; 454 455 /* 456 * swap_*() routines are externally accessible. swp_*() routines are 457 * internal. 458 */ 459 static int nswap_lowat = 128; /* in pages, swap_pager_almost_full warn */ 460 static int nswap_hiwat = 512; /* in pages, swap_pager_almost_full warn */ 461 462 SYSCTL_INT(_vm, OID_AUTO, dmmax, CTLFLAG_RD, &nsw_cluster_max, 0, 463 "Maximum size of a swap block in pages"); 464 465 static void swp_sizecheck(void); 466 static void swp_pager_async_iodone(struct buf *bp); 467 static bool swp_pager_swblk_empty(struct swblk *sb, int start, int limit); 468 static void swp_pager_free_empty_swblk(vm_object_t, struct swblk *sb); 469 static int swapongeom(struct vnode *); 470 static int swaponvp(struct thread *, struct vnode *, u_long); 471 static int swapoff_one(struct swdevt *sp, struct ucred *cred, 472 u_int flags); 473 474 /* 475 * Swap bitmap functions 476 */ 477 static void swp_pager_freeswapspace(daddr_t blk, daddr_t npages); 478 static daddr_t swp_pager_getswapspace(int *npages); 479 480 /* 481 * Metadata functions 482 */ 483 static daddr_t swp_pager_meta_build(vm_object_t, vm_pindex_t, daddr_t); 484 static void swp_pager_meta_free(vm_object_t, vm_pindex_t, vm_pindex_t, 485 vm_size_t *); 486 static void swp_pager_meta_transfer(vm_object_t src, vm_object_t dst, 487 vm_pindex_t pindex, vm_pindex_t count, vm_size_t *freed); 488 static void swp_pager_meta_free_all(vm_object_t); 489 static daddr_t swp_pager_meta_lookup(vm_object_t, vm_pindex_t); 490 491 static void 492 swp_pager_init_freerange(daddr_t *start, daddr_t *num) 493 { 494 495 *start = SWAPBLK_NONE; 496 *num = 0; 497 } 498 499 static void 500 swp_pager_update_freerange(daddr_t *start, daddr_t *num, daddr_t addr) 501 { 502 503 if (*start + *num == addr) { 504 (*num)++; 505 } else { 506 swp_pager_freeswapspace(*start, *num); 507 *start = addr; 508 *num = 1; 509 } 510 } 511 512 static void * 513 swblk_trie_alloc(struct pctrie *ptree) 514 { 515 516 return (uma_zalloc(swpctrie_zone, M_NOWAIT | (curproc == pageproc ? 517 M_USE_RESERVE : 0))); 518 } 519 520 static void 521 swblk_trie_free(struct pctrie *ptree, void *node) 522 { 523 524 uma_zfree(swpctrie_zone, node); 525 } 526 527 PCTRIE_DEFINE(SWAP, swblk, p, swblk_trie_alloc, swblk_trie_free); 528 529 /* 530 * SWP_SIZECHECK() - update swap_pager_full indication 531 * 532 * update the swap_pager_almost_full indication and warn when we are 533 * about to run out of swap space, using lowat/hiwat hysteresis. 534 * 535 * Clear swap_pager_full ( task killing ) indication when lowat is met. 536 * 537 * No restrictions on call 538 * This routine may not block. 539 */ 540 static void 541 swp_sizecheck(void) 542 { 543 544 if (swap_pager_avail < nswap_lowat) { 545 if (swap_pager_almost_full == 0) { 546 printf("swap_pager: out of swap space\n"); 547 swap_pager_almost_full = 1; 548 } 549 } else { 550 swap_pager_full = 0; 551 if (swap_pager_avail > nswap_hiwat) 552 swap_pager_almost_full = 0; 553 } 554 } 555 556 /* 557 * SWAP_PAGER_INIT() - initialize the swap pager! 558 * 559 * Expected to be started from system init. NOTE: This code is run 560 * before much else so be careful what you depend on. Most of the VM 561 * system has yet to be initialized at this point. 562 */ 563 static void 564 swap_pager_init(void) 565 { 566 /* 567 * Initialize object lists 568 */ 569 int i; 570 571 for (i = 0; i < NOBJLISTS; ++i) 572 TAILQ_INIT(&swap_pager_object_list[i]); 573 mtx_init(&sw_dev_mtx, "swapdev", NULL, MTX_DEF); 574 sx_init(&sw_alloc_sx, "swspsx"); 575 sx_init(&swdev_syscall_lock, "swsysc"); 576 577 /* 578 * The nsw_cluster_max is constrained by the bp->b_pages[] 579 * array, which has maxphys / PAGE_SIZE entries, and our locally 580 * defined MAX_PAGEOUT_CLUSTER. Also be aware that swap ops are 581 * constrained by the swap device interleave stripe size. 582 * 583 * Initialized early so that GEOM_ELI can see it. 584 */ 585 nsw_cluster_max = min(maxphys / PAGE_SIZE, MAX_PAGEOUT_CLUSTER); 586 } 587 588 /* 589 * SWAP_PAGER_SWAP_INIT() - swap pager initialization from pageout process 590 * 591 * Expected to be started from pageout process once, prior to entering 592 * its main loop. 593 */ 594 void 595 swap_pager_swap_init(void) 596 { 597 unsigned long n, n2; 598 599 /* 600 * Number of in-transit swap bp operations. Don't 601 * exhaust the pbufs completely. Make sure we 602 * initialize workable values (0 will work for hysteresis 603 * but it isn't very efficient). 604 * 605 * Currently we hardwire nsw_wcount_async to 4. This limit is 606 * designed to prevent other I/O from having high latencies due to 607 * our pageout I/O. The value 4 works well for one or two active swap 608 * devices but is probably a little low if you have more. Even so, 609 * a higher value would probably generate only a limited improvement 610 * with three or four active swap devices since the system does not 611 * typically have to pageout at extreme bandwidths. We will want 612 * at least 2 per swap devices, and 4 is a pretty good value if you 613 * have one NFS swap device due to the command/ack latency over NFS. 614 * So it all works out pretty well. 615 * 616 * nsw_cluster_max is initialized in swap_pager_init(). 617 */ 618 619 nsw_wcount_async = 4; 620 nsw_wcount_async_max = nsw_wcount_async; 621 mtx_init(&swbuf_mtx, "async swbuf mutex", NULL, MTX_DEF); 622 623 swwbuf_zone = pbuf_zsecond_create("swwbuf", nswbuf / 4); 624 swrbuf_zone = pbuf_zsecond_create("swrbuf", nswbuf / 2); 625 626 /* 627 * Initialize our zone, taking the user's requested size or 628 * estimating the number we need based on the number of pages 629 * in the system. 630 */ 631 n = maxswzone != 0 ? maxswzone / sizeof(struct swblk) : 632 vm_cnt.v_page_count / 2; 633 swpctrie_zone = uma_zcreate("swpctrie", pctrie_node_size(), NULL, NULL, 634 pctrie_zone_init, NULL, UMA_ALIGN_PTR, 0); 635 swblk_zone = uma_zcreate("swblk", sizeof(struct swblk), NULL, NULL, 636 NULL, NULL, _Alignof(struct swblk) - 1, 0); 637 n2 = n; 638 do { 639 if (uma_zone_reserve_kva(swblk_zone, n)) 640 break; 641 /* 642 * if the allocation failed, try a zone two thirds the 643 * size of the previous attempt. 644 */ 645 n -= ((n + 2) / 3); 646 } while (n > 0); 647 648 /* 649 * Often uma_zone_reserve_kva() cannot reserve exactly the 650 * requested size. Account for the difference when 651 * calculating swap_maxpages. 652 */ 653 n = uma_zone_get_max(swblk_zone); 654 655 if (n < n2) 656 printf("Swap blk zone entries changed from %lu to %lu.\n", 657 n2, n); 658 /* absolute maximum we can handle assuming 100% efficiency */ 659 swap_maxpages = n * SWAP_META_PAGES; 660 swzone = n * sizeof(struct swblk); 661 if (!uma_zone_reserve_kva(swpctrie_zone, n)) 662 printf("Cannot reserve swap pctrie zone, " 663 "reduce kern.maxswzone.\n"); 664 } 665 666 bool 667 swap_pager_init_object(vm_object_t object, void *handle, struct ucred *cred, 668 vm_ooffset_t size, vm_ooffset_t offset) 669 { 670 if (cred != NULL) { 671 if (!swap_reserve_by_cred(size, cred)) 672 return (false); 673 crhold(cred); 674 } 675 676 object->un_pager.swp.writemappings = 0; 677 object->handle = handle; 678 if (cred != NULL) { 679 object->cred = cred; 680 object->charge = size; 681 } 682 return (true); 683 } 684 685 static vm_object_t 686 swap_pager_alloc_init(objtype_t otype, void *handle, struct ucred *cred, 687 vm_ooffset_t size, vm_ooffset_t offset) 688 { 689 vm_object_t object; 690 691 /* 692 * The un_pager.swp.swp_blks trie is initialized by 693 * vm_object_allocate() to ensure the correct order of 694 * visibility to other threads. 695 */ 696 object = vm_object_allocate(otype, OFF_TO_IDX(offset + 697 PAGE_MASK + size)); 698 699 if (!swap_pager_init_object(object, handle, cred, size, offset)) { 700 vm_object_deallocate(object); 701 return (NULL); 702 } 703 return (object); 704 } 705 706 /* 707 * SWAP_PAGER_ALLOC() - allocate a new OBJT_SWAP VM object and instantiate 708 * its metadata structures. 709 * 710 * This routine is called from the mmap and fork code to create a new 711 * OBJT_SWAP object. 712 * 713 * This routine must ensure that no live duplicate is created for 714 * the named object request, which is protected against by 715 * holding the sw_alloc_sx lock in case handle != NULL. 716 */ 717 static vm_object_t 718 swap_pager_alloc(void *handle, vm_ooffset_t size, vm_prot_t prot, 719 vm_ooffset_t offset, struct ucred *cred) 720 { 721 vm_object_t object; 722 723 if (handle != NULL) { 724 /* 725 * Reference existing named region or allocate new one. There 726 * should not be a race here against swp_pager_meta_build() 727 * as called from vm_page_remove() in regards to the lookup 728 * of the handle. 729 */ 730 sx_xlock(&sw_alloc_sx); 731 object = vm_pager_object_lookup(NOBJLIST(handle), handle); 732 if (object == NULL) { 733 object = swap_pager_alloc_init(OBJT_SWAP, handle, cred, 734 size, offset); 735 if (object != NULL) { 736 TAILQ_INSERT_TAIL(NOBJLIST(object->handle), 737 object, pager_object_list); 738 } 739 } 740 sx_xunlock(&sw_alloc_sx); 741 } else { 742 object = swap_pager_alloc_init(OBJT_SWAP, handle, cred, 743 size, offset); 744 } 745 return (object); 746 } 747 748 /* 749 * SWAP_PAGER_DEALLOC() - remove swap metadata from object 750 * 751 * The swap backing for the object is destroyed. The code is 752 * designed such that we can reinstantiate it later, but this 753 * routine is typically called only when the entire object is 754 * about to be destroyed. 755 * 756 * The object must be locked. 757 */ 758 static void 759 swap_pager_dealloc(vm_object_t object) 760 { 761 762 VM_OBJECT_ASSERT_WLOCKED(object); 763 KASSERT((object->flags & OBJ_DEAD) != 0, ("dealloc of reachable obj")); 764 765 /* 766 * Remove from list right away so lookups will fail if we block for 767 * pageout completion. 768 */ 769 if ((object->flags & OBJ_ANON) == 0 && object->handle != NULL) { 770 VM_OBJECT_WUNLOCK(object); 771 sx_xlock(&sw_alloc_sx); 772 TAILQ_REMOVE(NOBJLIST(object->handle), object, 773 pager_object_list); 774 sx_xunlock(&sw_alloc_sx); 775 VM_OBJECT_WLOCK(object); 776 } 777 778 vm_object_pip_wait(object, "swpdea"); 779 780 /* 781 * Free all remaining metadata. We only bother to free it from 782 * the swap meta data. We do not attempt to free swapblk's still 783 * associated with vm_page_t's for this object. We do not care 784 * if paging is still in progress on some objects. 785 */ 786 swp_pager_meta_free_all(object); 787 object->handle = NULL; 788 object->type = OBJT_DEAD; 789 790 /* 791 * Release the allocation charge. 792 */ 793 if (object->cred != NULL) { 794 swap_release_by_cred(object->charge, object->cred); 795 object->charge = 0; 796 crfree(object->cred); 797 object->cred = NULL; 798 } 799 800 /* 801 * Hide the object from swap_pager_swapoff(). 802 */ 803 vm_object_clear_flag(object, OBJ_SWAP); 804 } 805 806 /************************************************************************ 807 * SWAP PAGER BITMAP ROUTINES * 808 ************************************************************************/ 809 810 /* 811 * SWP_PAGER_GETSWAPSPACE() - allocate raw swap space 812 * 813 * Allocate swap for up to the requested number of pages. The 814 * starting swap block number (a page index) is returned or 815 * SWAPBLK_NONE if the allocation failed. 816 * 817 * Also has the side effect of advising that somebody made a mistake 818 * when they configured swap and didn't configure enough. 819 * 820 * This routine may not sleep. 821 * 822 * We allocate in round-robin fashion from the configured devices. 823 */ 824 static daddr_t 825 swp_pager_getswapspace(int *io_npages) 826 { 827 daddr_t blk; 828 struct swdevt *sp; 829 int mpages, npages; 830 831 KASSERT(*io_npages >= 1, 832 ("%s: npages not positive", __func__)); 833 blk = SWAPBLK_NONE; 834 mpages = *io_npages; 835 npages = imin(BLIST_MAX_ALLOC, mpages); 836 mtx_lock(&sw_dev_mtx); 837 sp = swdevhd; 838 while (!TAILQ_EMPTY(&swtailq)) { 839 if (sp == NULL) 840 sp = TAILQ_FIRST(&swtailq); 841 if ((sp->sw_flags & SW_CLOSING) == 0) 842 blk = blist_alloc(sp->sw_blist, &npages, mpages); 843 if (blk != SWAPBLK_NONE) 844 break; 845 sp = TAILQ_NEXT(sp, sw_list); 846 if (swdevhd == sp) { 847 if (npages == 1) 848 break; 849 mpages = npages - 1; 850 npages >>= 1; 851 } 852 } 853 if (blk != SWAPBLK_NONE) { 854 *io_npages = npages; 855 blk += sp->sw_first; 856 sp->sw_used += npages; 857 swap_pager_avail -= npages; 858 swp_sizecheck(); 859 swdevhd = TAILQ_NEXT(sp, sw_list); 860 } else { 861 if (swap_pager_full != 2) { 862 printf("swp_pager_getswapspace(%d): failed\n", 863 *io_npages); 864 swap_pager_full = 2; 865 swap_pager_almost_full = 1; 866 } 867 swdevhd = NULL; 868 } 869 mtx_unlock(&sw_dev_mtx); 870 return (blk); 871 } 872 873 static bool 874 swp_pager_isondev(daddr_t blk, struct swdevt *sp) 875 { 876 877 return (blk >= sp->sw_first && blk < sp->sw_end); 878 } 879 880 static void 881 swp_pager_strategy(struct buf *bp) 882 { 883 struct swdevt *sp; 884 885 mtx_lock(&sw_dev_mtx); 886 TAILQ_FOREACH(sp, &swtailq, sw_list) { 887 if (swp_pager_isondev(bp->b_blkno, sp)) { 888 mtx_unlock(&sw_dev_mtx); 889 if ((sp->sw_flags & SW_UNMAPPED) != 0 && 890 unmapped_buf_allowed) { 891 bp->b_data = unmapped_buf; 892 bp->b_offset = 0; 893 } else { 894 pmap_qenter((vm_offset_t)bp->b_data, 895 &bp->b_pages[0], bp->b_bcount / PAGE_SIZE); 896 } 897 sp->sw_strategy(bp, sp); 898 return; 899 } 900 } 901 panic("Swapdev not found"); 902 } 903 904 /* 905 * SWP_PAGER_FREESWAPSPACE() - free raw swap space 906 * 907 * This routine returns the specified swap blocks back to the bitmap. 908 * 909 * This routine may not sleep. 910 */ 911 static void 912 swp_pager_freeswapspace(daddr_t blk, daddr_t npages) 913 { 914 struct swdevt *sp; 915 916 if (npages == 0) 917 return; 918 mtx_lock(&sw_dev_mtx); 919 TAILQ_FOREACH(sp, &swtailq, sw_list) { 920 if (swp_pager_isondev(blk, sp)) { 921 sp->sw_used -= npages; 922 /* 923 * If we are attempting to stop swapping on 924 * this device, we don't want to mark any 925 * blocks free lest they be reused. 926 */ 927 if ((sp->sw_flags & SW_CLOSING) == 0) { 928 blist_free(sp->sw_blist, blk - sp->sw_first, 929 npages); 930 swap_pager_avail += npages; 931 swp_sizecheck(); 932 } 933 mtx_unlock(&sw_dev_mtx); 934 return; 935 } 936 } 937 panic("Swapdev not found"); 938 } 939 940 /* 941 * SYSCTL_SWAP_FRAGMENTATION() - produce raw swap space stats 942 */ 943 static int 944 sysctl_swap_fragmentation(SYSCTL_HANDLER_ARGS) 945 { 946 struct sbuf sbuf; 947 struct swdevt *sp; 948 const char *devname; 949 int error; 950 951 error = sysctl_wire_old_buffer(req, 0); 952 if (error != 0) 953 return (error); 954 sbuf_new_for_sysctl(&sbuf, NULL, 128, req); 955 mtx_lock(&sw_dev_mtx); 956 TAILQ_FOREACH(sp, &swtailq, sw_list) { 957 if (vn_isdisk(sp->sw_vp)) 958 devname = devtoname(sp->sw_vp->v_rdev); 959 else 960 devname = "[file]"; 961 sbuf_printf(&sbuf, "\nFree space on device %s:\n", devname); 962 blist_stats(sp->sw_blist, &sbuf); 963 } 964 mtx_unlock(&sw_dev_mtx); 965 error = sbuf_finish(&sbuf); 966 sbuf_delete(&sbuf); 967 return (error); 968 } 969 970 /* 971 * SWAP_PAGER_FREESPACE() - frees swap blocks associated with a page 972 * range within an object. 973 * 974 * This routine removes swapblk assignments from swap metadata. 975 * 976 * The external callers of this routine typically have already destroyed 977 * or renamed vm_page_t's associated with this range in the object so 978 * we should be ok. 979 * 980 * The object must be locked. 981 */ 982 void 983 swap_pager_freespace(vm_object_t object, vm_pindex_t start, vm_size_t size, 984 vm_size_t *freed) 985 { 986 MPASS((object->flags & OBJ_SWAP) != 0); 987 988 swp_pager_meta_free(object, start, size, freed); 989 } 990 991 static void 992 swap_pager_freespace_pgo(vm_object_t object, vm_pindex_t start, vm_size_t size) 993 { 994 MPASS((object->flags & OBJ_SWAP) != 0); 995 996 swp_pager_meta_free(object, start, size, NULL); 997 } 998 999 /* 1000 * SWAP_PAGER_RESERVE() - reserve swap blocks in object 1001 * 1002 * Assigns swap blocks to the specified range within the object. The 1003 * swap blocks are not zeroed. Any previous swap assignment is destroyed. 1004 * 1005 * Returns 0 on success, -1 on failure. 1006 */ 1007 int 1008 swap_pager_reserve(vm_object_t object, vm_pindex_t start, vm_pindex_t size) 1009 { 1010 daddr_t addr, blk, n_free, s_free; 1011 vm_pindex_t i, j; 1012 int n; 1013 1014 swp_pager_init_freerange(&s_free, &n_free); 1015 VM_OBJECT_WLOCK(object); 1016 for (i = 0; i < size; i += n) { 1017 n = MIN(size - i, INT_MAX); 1018 blk = swp_pager_getswapspace(&n); 1019 if (blk == SWAPBLK_NONE) { 1020 swp_pager_meta_free(object, start, i, NULL); 1021 VM_OBJECT_WUNLOCK(object); 1022 return (-1); 1023 } 1024 for (j = 0; j < n; ++j) { 1025 addr = swp_pager_meta_build(object, 1026 start + i + j, blk + j); 1027 if (addr != SWAPBLK_NONE) 1028 swp_pager_update_freerange(&s_free, &n_free, 1029 addr); 1030 } 1031 } 1032 swp_pager_freeswapspace(s_free, n_free); 1033 VM_OBJECT_WUNLOCK(object); 1034 return (0); 1035 } 1036 1037 static bool 1038 swp_pager_xfer_source(vm_object_t srcobject, vm_object_t dstobject, 1039 vm_pindex_t pindex, daddr_t addr) 1040 { 1041 daddr_t dstaddr __diagused; 1042 1043 KASSERT((srcobject->flags & OBJ_SWAP) != 0, 1044 ("%s: srcobject not swappable", __func__)); 1045 KASSERT((dstobject->flags & OBJ_SWAP) != 0, 1046 ("%s: dstobject not swappable", __func__)); 1047 1048 if (swp_pager_meta_lookup(dstobject, pindex) != SWAPBLK_NONE) { 1049 /* Caller should destroy the source block. */ 1050 return (false); 1051 } 1052 1053 /* 1054 * Destination has no swapblk and is not resident, transfer source. 1055 * swp_pager_meta_build() can sleep. 1056 */ 1057 VM_OBJECT_WUNLOCK(srcobject); 1058 dstaddr = swp_pager_meta_build(dstobject, pindex, addr); 1059 KASSERT(dstaddr == SWAPBLK_NONE, 1060 ("Unexpected destination swapblk")); 1061 VM_OBJECT_WLOCK(srcobject); 1062 1063 return (true); 1064 } 1065 1066 /* 1067 * SWAP_PAGER_COPY() - copy blocks from source pager to destination pager 1068 * and destroy the source. 1069 * 1070 * Copy any valid swapblks from the source to the destination. In 1071 * cases where both the source and destination have a valid swapblk, 1072 * we keep the destination's. 1073 * 1074 * This routine is allowed to sleep. It may sleep allocating metadata 1075 * indirectly through swp_pager_meta_build(). 1076 * 1077 * The source object contains no vm_page_t's (which is just as well) 1078 * 1079 * The source and destination objects must be locked. 1080 * Both object locks may temporarily be released. 1081 */ 1082 void 1083 swap_pager_copy(vm_object_t srcobject, vm_object_t dstobject, 1084 vm_pindex_t offset, int destroysource) 1085 { 1086 VM_OBJECT_ASSERT_WLOCKED(srcobject); 1087 VM_OBJECT_ASSERT_WLOCKED(dstobject); 1088 1089 /* 1090 * If destroysource is set, we remove the source object from the 1091 * swap_pager internal queue now. 1092 */ 1093 if (destroysource && (srcobject->flags & OBJ_ANON) == 0 && 1094 srcobject->handle != NULL) { 1095 VM_OBJECT_WUNLOCK(srcobject); 1096 VM_OBJECT_WUNLOCK(dstobject); 1097 sx_xlock(&sw_alloc_sx); 1098 TAILQ_REMOVE(NOBJLIST(srcobject->handle), srcobject, 1099 pager_object_list); 1100 sx_xunlock(&sw_alloc_sx); 1101 VM_OBJECT_WLOCK(dstobject); 1102 VM_OBJECT_WLOCK(srcobject); 1103 } 1104 1105 /* 1106 * Transfer source to destination. 1107 */ 1108 swp_pager_meta_transfer(srcobject, dstobject, offset, dstobject->size, 1109 NULL); 1110 1111 /* 1112 * Free left over swap blocks in source. 1113 */ 1114 if (destroysource) 1115 swp_pager_meta_free_all(srcobject); 1116 } 1117 1118 /* 1119 * SWAP_PAGER_HASPAGE() - determine if we have good backing store for 1120 * the requested page. 1121 * 1122 * We determine whether good backing store exists for the requested 1123 * page and return TRUE if it does, FALSE if it doesn't. 1124 * 1125 * If TRUE, we also try to determine how much valid, contiguous backing 1126 * store exists before and after the requested page. 1127 */ 1128 static boolean_t 1129 swap_pager_haspage(vm_object_t object, vm_pindex_t pindex, int *before, 1130 int *after) 1131 { 1132 daddr_t blk, blk0; 1133 int i; 1134 1135 VM_OBJECT_ASSERT_LOCKED(object); 1136 KASSERT((object->flags & OBJ_SWAP) != 0, 1137 ("%s: object not swappable", __func__)); 1138 1139 /* 1140 * do we have good backing store at the requested index ? 1141 */ 1142 blk0 = swp_pager_meta_lookup(object, pindex); 1143 if (blk0 == SWAPBLK_NONE) { 1144 if (before) 1145 *before = 0; 1146 if (after) 1147 *after = 0; 1148 return (FALSE); 1149 } 1150 1151 /* 1152 * find backwards-looking contiguous good backing store 1153 */ 1154 if (before != NULL) { 1155 for (i = 1; i < SWB_NPAGES; i++) { 1156 if (i > pindex) 1157 break; 1158 blk = swp_pager_meta_lookup(object, pindex - i); 1159 if (blk != blk0 - i) 1160 break; 1161 } 1162 *before = i - 1; 1163 } 1164 1165 /* 1166 * find forward-looking contiguous good backing store 1167 */ 1168 if (after != NULL) { 1169 for (i = 1; i < SWB_NPAGES; i++) { 1170 blk = swp_pager_meta_lookup(object, pindex + i); 1171 if (blk != blk0 + i) 1172 break; 1173 } 1174 *after = i - 1; 1175 } 1176 return (TRUE); 1177 } 1178 1179 /* 1180 * SWAP_PAGER_PAGE_UNSWAPPED() - remove swap backing store related to page 1181 * 1182 * This removes any associated swap backing store, whether valid or 1183 * not, from the page. 1184 * 1185 * This routine is typically called when a page is made dirty, at 1186 * which point any associated swap can be freed. MADV_FREE also 1187 * calls us in a special-case situation 1188 * 1189 * NOTE!!! If the page is clean and the swap was valid, the caller 1190 * should make the page dirty before calling this routine. This routine 1191 * does NOT change the m->dirty status of the page. Also: MADV_FREE 1192 * depends on it. 1193 * 1194 * This routine may not sleep. 1195 * 1196 * The object containing the page may be locked. 1197 */ 1198 static void 1199 swap_pager_unswapped(vm_page_t m) 1200 { 1201 struct swblk *sb; 1202 vm_object_t obj; 1203 1204 /* 1205 * Handle enqueing deferred frees first. If we do not have the 1206 * object lock we wait for the page daemon to clear the space. 1207 */ 1208 obj = m->object; 1209 if (!VM_OBJECT_WOWNED(obj)) { 1210 VM_PAGE_OBJECT_BUSY_ASSERT(m); 1211 /* 1212 * The caller is responsible for synchronization but we 1213 * will harmlessly handle races. This is typically provided 1214 * by only calling unswapped() when a page transitions from 1215 * clean to dirty. 1216 */ 1217 if ((m->a.flags & (PGA_SWAP_SPACE | PGA_SWAP_FREE)) == 1218 PGA_SWAP_SPACE) { 1219 vm_page_aflag_set(m, PGA_SWAP_FREE); 1220 counter_u64_add(swap_free_deferred, 1); 1221 } 1222 return; 1223 } 1224 if ((m->a.flags & PGA_SWAP_FREE) != 0) 1225 counter_u64_add(swap_free_completed, 1); 1226 vm_page_aflag_clear(m, PGA_SWAP_FREE | PGA_SWAP_SPACE); 1227 1228 /* 1229 * The meta data only exists if the object is OBJT_SWAP 1230 * and even then might not be allocated yet. 1231 */ 1232 KASSERT((m->object->flags & OBJ_SWAP) != 0, 1233 ("Free object not swappable")); 1234 1235 sb = SWAP_PCTRIE_LOOKUP(&m->object->un_pager.swp.swp_blks, 1236 rounddown(m->pindex, SWAP_META_PAGES)); 1237 if (sb == NULL) 1238 return; 1239 if (sb->d[m->pindex % SWAP_META_PAGES] == SWAPBLK_NONE) 1240 return; 1241 swp_pager_freeswapspace(sb->d[m->pindex % SWAP_META_PAGES], 1); 1242 sb->d[m->pindex % SWAP_META_PAGES] = SWAPBLK_NONE; 1243 swp_pager_free_empty_swblk(m->object, sb); 1244 } 1245 1246 /* 1247 * swap_pager_getpages() - bring pages in from swap 1248 * 1249 * Attempt to page in the pages in array "ma" of length "count". The 1250 * caller may optionally specify that additional pages preceding and 1251 * succeeding the specified range be paged in. The number of such pages 1252 * is returned in the "rbehind" and "rahead" parameters, and they will 1253 * be in the inactive queue upon return. 1254 * 1255 * The pages in "ma" must be busied and will remain busied upon return. 1256 */ 1257 static int 1258 swap_pager_getpages_locked(vm_object_t object, vm_page_t *ma, int count, 1259 int *rbehind, int *rahead) 1260 { 1261 struct buf *bp; 1262 vm_page_t bm, mpred, msucc, p; 1263 vm_pindex_t pindex; 1264 daddr_t blk; 1265 int i, maxahead, maxbehind, reqcount; 1266 1267 VM_OBJECT_ASSERT_WLOCKED(object); 1268 reqcount = count; 1269 1270 KASSERT((object->flags & OBJ_SWAP) != 0, 1271 ("%s: object not swappable", __func__)); 1272 if (!swap_pager_haspage(object, ma[0]->pindex, &maxbehind, &maxahead)) { 1273 VM_OBJECT_WUNLOCK(object); 1274 return (VM_PAGER_FAIL); 1275 } 1276 1277 KASSERT(reqcount - 1 <= maxahead, 1278 ("page count %d extends beyond swap block", reqcount)); 1279 1280 /* 1281 * Do not transfer any pages other than those that are xbusied 1282 * when running during a split or collapse operation. This 1283 * prevents clustering from re-creating pages which are being 1284 * moved into another object. 1285 */ 1286 if ((object->flags & (OBJ_SPLIT | OBJ_DEAD)) != 0) { 1287 maxahead = reqcount - 1; 1288 maxbehind = 0; 1289 } 1290 1291 /* 1292 * Clip the readahead and readbehind ranges to exclude resident pages. 1293 */ 1294 if (rahead != NULL) { 1295 *rahead = imin(*rahead, maxahead - (reqcount - 1)); 1296 pindex = ma[reqcount - 1]->pindex; 1297 msucc = TAILQ_NEXT(ma[reqcount - 1], listq); 1298 if (msucc != NULL && msucc->pindex - pindex - 1 < *rahead) 1299 *rahead = msucc->pindex - pindex - 1; 1300 } 1301 if (rbehind != NULL) { 1302 *rbehind = imin(*rbehind, maxbehind); 1303 pindex = ma[0]->pindex; 1304 mpred = TAILQ_PREV(ma[0], pglist, listq); 1305 if (mpred != NULL && pindex - mpred->pindex - 1 < *rbehind) 1306 *rbehind = pindex - mpred->pindex - 1; 1307 } 1308 1309 bm = ma[0]; 1310 for (i = 0; i < count; i++) 1311 ma[i]->oflags |= VPO_SWAPINPROG; 1312 1313 /* 1314 * Allocate readahead and readbehind pages. 1315 */ 1316 if (rbehind != NULL) { 1317 for (i = 1; i <= *rbehind; i++) { 1318 p = vm_page_alloc(object, ma[0]->pindex - i, 1319 VM_ALLOC_NORMAL); 1320 if (p == NULL) 1321 break; 1322 p->oflags |= VPO_SWAPINPROG; 1323 bm = p; 1324 } 1325 *rbehind = i - 1; 1326 } 1327 if (rahead != NULL) { 1328 for (i = 0; i < *rahead; i++) { 1329 p = vm_page_alloc(object, 1330 ma[reqcount - 1]->pindex + i + 1, VM_ALLOC_NORMAL); 1331 if (p == NULL) 1332 break; 1333 p->oflags |= VPO_SWAPINPROG; 1334 } 1335 *rahead = i; 1336 } 1337 if (rbehind != NULL) 1338 count += *rbehind; 1339 if (rahead != NULL) 1340 count += *rahead; 1341 1342 vm_object_pip_add(object, count); 1343 1344 pindex = bm->pindex; 1345 blk = swp_pager_meta_lookup(object, pindex); 1346 KASSERT(blk != SWAPBLK_NONE, 1347 ("no swap blocking containing %p(%jx)", object, (uintmax_t)pindex)); 1348 1349 VM_OBJECT_WUNLOCK(object); 1350 bp = uma_zalloc(swrbuf_zone, M_WAITOK); 1351 MPASS((bp->b_flags & B_MAXPHYS) != 0); 1352 /* Pages cannot leave the object while busy. */ 1353 for (i = 0, p = bm; i < count; i++, p = TAILQ_NEXT(p, listq)) { 1354 MPASS(p->pindex == bm->pindex + i); 1355 bp->b_pages[i] = p; 1356 } 1357 1358 bp->b_flags |= B_PAGING; 1359 bp->b_iocmd = BIO_READ; 1360 bp->b_iodone = swp_pager_async_iodone; 1361 bp->b_rcred = crhold(thread0.td_ucred); 1362 bp->b_wcred = crhold(thread0.td_ucred); 1363 bp->b_blkno = blk; 1364 bp->b_bcount = PAGE_SIZE * count; 1365 bp->b_bufsize = PAGE_SIZE * count; 1366 bp->b_npages = count; 1367 bp->b_pgbefore = rbehind != NULL ? *rbehind : 0; 1368 bp->b_pgafter = rahead != NULL ? *rahead : 0; 1369 1370 VM_CNT_INC(v_swapin); 1371 VM_CNT_ADD(v_swappgsin, count); 1372 1373 /* 1374 * perform the I/O. NOTE!!! bp cannot be considered valid after 1375 * this point because we automatically release it on completion. 1376 * Instead, we look at the one page we are interested in which we 1377 * still hold a lock on even through the I/O completion. 1378 * 1379 * The other pages in our ma[] array are also released on completion, 1380 * so we cannot assume they are valid anymore either. 1381 * 1382 * NOTE: b_blkno is destroyed by the call to swapdev_strategy 1383 */ 1384 BUF_KERNPROC(bp); 1385 swp_pager_strategy(bp); 1386 1387 /* 1388 * Wait for the pages we want to complete. VPO_SWAPINPROG is always 1389 * cleared on completion. If an I/O error occurs, SWAPBLK_NONE 1390 * is set in the metadata for each page in the request. 1391 */ 1392 VM_OBJECT_WLOCK(object); 1393 /* This could be implemented more efficiently with aflags */ 1394 while ((ma[0]->oflags & VPO_SWAPINPROG) != 0) { 1395 ma[0]->oflags |= VPO_SWAPSLEEP; 1396 VM_CNT_INC(v_intrans); 1397 if (VM_OBJECT_SLEEP(object, &object->handle, PSWP, 1398 "swread", hz * 20)) { 1399 printf( 1400 "swap_pager: indefinite wait buffer: bufobj: %p, blkno: %jd, size: %ld\n", 1401 bp->b_bufobj, (intmax_t)bp->b_blkno, bp->b_bcount); 1402 } 1403 } 1404 VM_OBJECT_WUNLOCK(object); 1405 1406 /* 1407 * If we had an unrecoverable read error pages will not be valid. 1408 */ 1409 for (i = 0; i < reqcount; i++) 1410 if (ma[i]->valid != VM_PAGE_BITS_ALL) 1411 return (VM_PAGER_ERROR); 1412 1413 return (VM_PAGER_OK); 1414 1415 /* 1416 * A final note: in a low swap situation, we cannot deallocate swap 1417 * and mark a page dirty here because the caller is likely to mark 1418 * the page clean when we return, causing the page to possibly revert 1419 * to all-zero's later. 1420 */ 1421 } 1422 1423 static int 1424 swap_pager_getpages(vm_object_t object, vm_page_t *ma, int count, 1425 int *rbehind, int *rahead) 1426 { 1427 1428 VM_OBJECT_WLOCK(object); 1429 return (swap_pager_getpages_locked(object, ma, count, rbehind, rahead)); 1430 } 1431 1432 /* 1433 * swap_pager_getpages_async(): 1434 * 1435 * Right now this is emulation of asynchronous operation on top of 1436 * swap_pager_getpages(). 1437 */ 1438 static int 1439 swap_pager_getpages_async(vm_object_t object, vm_page_t *ma, int count, 1440 int *rbehind, int *rahead, pgo_getpages_iodone_t iodone, void *arg) 1441 { 1442 int r, error; 1443 1444 r = swap_pager_getpages(object, ma, count, rbehind, rahead); 1445 switch (r) { 1446 case VM_PAGER_OK: 1447 error = 0; 1448 break; 1449 case VM_PAGER_ERROR: 1450 error = EIO; 1451 break; 1452 case VM_PAGER_FAIL: 1453 error = EINVAL; 1454 break; 1455 default: 1456 panic("unhandled swap_pager_getpages() error %d", r); 1457 } 1458 (iodone)(arg, ma, count, error); 1459 1460 return (r); 1461 } 1462 1463 /* 1464 * swap_pager_putpages: 1465 * 1466 * Assign swap (if necessary) and initiate I/O on the specified pages. 1467 * 1468 * In a low memory situation we may block in VOP_STRATEGY(), but the new 1469 * vm_page reservation system coupled with properly written VFS devices 1470 * should ensure that no low-memory deadlock occurs. This is an area 1471 * which needs work. 1472 * 1473 * The parent has N vm_object_pip_add() references prior to 1474 * calling us and will remove references for rtvals[] that are 1475 * not set to VM_PAGER_PEND. We need to remove the rest on I/O 1476 * completion. 1477 * 1478 * The parent has soft-busy'd the pages it passes us and will unbusy 1479 * those whose rtvals[] entry is not set to VM_PAGER_PEND on return. 1480 * We need to unbusy the rest on I/O completion. 1481 */ 1482 static void 1483 swap_pager_putpages(vm_object_t object, vm_page_t *ma, int count, 1484 int flags, int *rtvals) 1485 { 1486 struct buf *bp; 1487 daddr_t addr, blk, n_free, s_free; 1488 vm_page_t mreq; 1489 int i, j, n; 1490 bool async; 1491 1492 KASSERT(count == 0 || ma[0]->object == object, 1493 ("%s: object mismatch %p/%p", 1494 __func__, object, ma[0]->object)); 1495 1496 VM_OBJECT_WUNLOCK(object); 1497 async = curproc == pageproc && (flags & VM_PAGER_PUT_SYNC) == 0; 1498 swp_pager_init_freerange(&s_free, &n_free); 1499 1500 /* 1501 * Assign swap blocks and issue I/O. We reallocate swap on the fly. 1502 * The page is left dirty until the pageout operation completes 1503 * successfully. 1504 */ 1505 for (i = 0; i < count; i += n) { 1506 /* Maximum I/O size is limited by maximum swap block size. */ 1507 n = min(count - i, nsw_cluster_max); 1508 1509 if (async) { 1510 mtx_lock(&swbuf_mtx); 1511 while (nsw_wcount_async == 0) 1512 msleep(&nsw_wcount_async, &swbuf_mtx, PVM, 1513 "swbufa", 0); 1514 nsw_wcount_async--; 1515 mtx_unlock(&swbuf_mtx); 1516 } 1517 1518 /* Get a block of swap of size up to size n. */ 1519 blk = swp_pager_getswapspace(&n); 1520 if (blk == SWAPBLK_NONE) { 1521 mtx_lock(&swbuf_mtx); 1522 if (++nsw_wcount_async == 1) 1523 wakeup(&nsw_wcount_async); 1524 mtx_unlock(&swbuf_mtx); 1525 for (j = 0; j < n; ++j) 1526 rtvals[i + j] = VM_PAGER_FAIL; 1527 continue; 1528 } 1529 VM_OBJECT_WLOCK(object); 1530 for (j = 0; j < n; ++j) { 1531 mreq = ma[i + j]; 1532 vm_page_aflag_clear(mreq, PGA_SWAP_FREE); 1533 addr = swp_pager_meta_build(mreq->object, mreq->pindex, 1534 blk + j); 1535 if (addr != SWAPBLK_NONE) 1536 swp_pager_update_freerange(&s_free, &n_free, 1537 addr); 1538 MPASS(mreq->dirty == VM_PAGE_BITS_ALL); 1539 mreq->oflags |= VPO_SWAPINPROG; 1540 } 1541 VM_OBJECT_WUNLOCK(object); 1542 1543 bp = uma_zalloc(swwbuf_zone, M_WAITOK); 1544 MPASS((bp->b_flags & B_MAXPHYS) != 0); 1545 if (async) 1546 bp->b_flags |= B_ASYNC; 1547 bp->b_flags |= B_PAGING; 1548 bp->b_iocmd = BIO_WRITE; 1549 1550 bp->b_rcred = crhold(thread0.td_ucred); 1551 bp->b_wcred = crhold(thread0.td_ucred); 1552 bp->b_bcount = PAGE_SIZE * n; 1553 bp->b_bufsize = PAGE_SIZE * n; 1554 bp->b_blkno = blk; 1555 for (j = 0; j < n; j++) 1556 bp->b_pages[j] = ma[i + j]; 1557 bp->b_npages = n; 1558 1559 /* 1560 * Must set dirty range for NFS to work. 1561 */ 1562 bp->b_dirtyoff = 0; 1563 bp->b_dirtyend = bp->b_bcount; 1564 1565 VM_CNT_INC(v_swapout); 1566 VM_CNT_ADD(v_swappgsout, bp->b_npages); 1567 1568 /* 1569 * We unconditionally set rtvals[] to VM_PAGER_PEND so that we 1570 * can call the async completion routine at the end of a 1571 * synchronous I/O operation. Otherwise, our caller would 1572 * perform duplicate unbusy and wakeup operations on the page 1573 * and object, respectively. 1574 */ 1575 for (j = 0; j < n; j++) 1576 rtvals[i + j] = VM_PAGER_PEND; 1577 1578 /* 1579 * asynchronous 1580 * 1581 * NOTE: b_blkno is destroyed by the call to swapdev_strategy. 1582 */ 1583 if (async) { 1584 bp->b_iodone = swp_pager_async_iodone; 1585 BUF_KERNPROC(bp); 1586 swp_pager_strategy(bp); 1587 continue; 1588 } 1589 1590 /* 1591 * synchronous 1592 * 1593 * NOTE: b_blkno is destroyed by the call to swapdev_strategy. 1594 */ 1595 bp->b_iodone = bdone; 1596 swp_pager_strategy(bp); 1597 1598 /* 1599 * Wait for the sync I/O to complete. 1600 */ 1601 bwait(bp, PVM, "swwrt"); 1602 1603 /* 1604 * Now that we are through with the bp, we can call the 1605 * normal async completion, which frees everything up. 1606 */ 1607 swp_pager_async_iodone(bp); 1608 } 1609 swp_pager_freeswapspace(s_free, n_free); 1610 VM_OBJECT_WLOCK(object); 1611 } 1612 1613 /* 1614 * swp_pager_async_iodone: 1615 * 1616 * Completion routine for asynchronous reads and writes from/to swap. 1617 * Also called manually by synchronous code to finish up a bp. 1618 * 1619 * This routine may not sleep. 1620 */ 1621 static void 1622 swp_pager_async_iodone(struct buf *bp) 1623 { 1624 int i; 1625 vm_object_t object = NULL; 1626 1627 /* 1628 * Report error - unless we ran out of memory, in which case 1629 * we've already logged it in swapgeom_strategy(). 1630 */ 1631 if (bp->b_ioflags & BIO_ERROR && bp->b_error != ENOMEM) { 1632 printf( 1633 "swap_pager: I/O error - %s failed; blkno %ld," 1634 "size %ld, error %d\n", 1635 ((bp->b_iocmd == BIO_READ) ? "pagein" : "pageout"), 1636 (long)bp->b_blkno, 1637 (long)bp->b_bcount, 1638 bp->b_error 1639 ); 1640 } 1641 1642 /* 1643 * remove the mapping for kernel virtual 1644 */ 1645 if (buf_mapped(bp)) 1646 pmap_qremove((vm_offset_t)bp->b_data, bp->b_npages); 1647 else 1648 bp->b_data = bp->b_kvabase; 1649 1650 if (bp->b_npages) { 1651 object = bp->b_pages[0]->object; 1652 VM_OBJECT_WLOCK(object); 1653 } 1654 1655 /* 1656 * cleanup pages. If an error occurs writing to swap, we are in 1657 * very serious trouble. If it happens to be a disk error, though, 1658 * we may be able to recover by reassigning the swap later on. So 1659 * in this case we remove the m->swapblk assignment for the page 1660 * but do not free it in the rlist. The errornous block(s) are thus 1661 * never reallocated as swap. Redirty the page and continue. 1662 */ 1663 for (i = 0; i < bp->b_npages; ++i) { 1664 vm_page_t m = bp->b_pages[i]; 1665 1666 m->oflags &= ~VPO_SWAPINPROG; 1667 if (m->oflags & VPO_SWAPSLEEP) { 1668 m->oflags &= ~VPO_SWAPSLEEP; 1669 wakeup(&object->handle); 1670 } 1671 1672 /* We always have space after I/O, successful or not. */ 1673 vm_page_aflag_set(m, PGA_SWAP_SPACE); 1674 1675 if (bp->b_ioflags & BIO_ERROR) { 1676 /* 1677 * If an error occurs I'd love to throw the swapblk 1678 * away without freeing it back to swapspace, so it 1679 * can never be used again. But I can't from an 1680 * interrupt. 1681 */ 1682 if (bp->b_iocmd == BIO_READ) { 1683 /* 1684 * NOTE: for reads, m->dirty will probably 1685 * be overridden by the original caller of 1686 * getpages so don't play cute tricks here. 1687 */ 1688 vm_page_invalid(m); 1689 } else { 1690 /* 1691 * If a write error occurs, reactivate page 1692 * so it doesn't clog the inactive list, 1693 * then finish the I/O. 1694 */ 1695 MPASS(m->dirty == VM_PAGE_BITS_ALL); 1696 1697 /* PQ_UNSWAPPABLE? */ 1698 vm_page_activate(m); 1699 vm_page_sunbusy(m); 1700 } 1701 } else if (bp->b_iocmd == BIO_READ) { 1702 /* 1703 * NOTE: for reads, m->dirty will probably be 1704 * overridden by the original caller of getpages so 1705 * we cannot set them in order to free the underlying 1706 * swap in a low-swap situation. I don't think we'd 1707 * want to do that anyway, but it was an optimization 1708 * that existed in the old swapper for a time before 1709 * it got ripped out due to precisely this problem. 1710 */ 1711 KASSERT(!pmap_page_is_mapped(m), 1712 ("swp_pager_async_iodone: page %p is mapped", m)); 1713 KASSERT(m->dirty == 0, 1714 ("swp_pager_async_iodone: page %p is dirty", m)); 1715 1716 vm_page_valid(m); 1717 if (i < bp->b_pgbefore || 1718 i >= bp->b_npages - bp->b_pgafter) 1719 vm_page_readahead_finish(m); 1720 } else { 1721 /* 1722 * For write success, clear the dirty 1723 * status, then finish the I/O ( which decrements the 1724 * busy count and possibly wakes waiter's up ). 1725 * A page is only written to swap after a period of 1726 * inactivity. Therefore, we do not expect it to be 1727 * reused. 1728 */ 1729 KASSERT(!pmap_page_is_write_mapped(m), 1730 ("swp_pager_async_iodone: page %p is not write" 1731 " protected", m)); 1732 vm_page_undirty(m); 1733 vm_page_deactivate_noreuse(m); 1734 vm_page_sunbusy(m); 1735 } 1736 } 1737 1738 /* 1739 * adjust pip. NOTE: the original parent may still have its own 1740 * pip refs on the object. 1741 */ 1742 if (object != NULL) { 1743 vm_object_pip_wakeupn(object, bp->b_npages); 1744 VM_OBJECT_WUNLOCK(object); 1745 } 1746 1747 /* 1748 * swapdev_strategy() manually sets b_vp and b_bufobj before calling 1749 * bstrategy(). Set them back to NULL now we're done with it, or we'll 1750 * trigger a KASSERT in relpbuf(). 1751 */ 1752 if (bp->b_vp) { 1753 bp->b_vp = NULL; 1754 bp->b_bufobj = NULL; 1755 } 1756 /* 1757 * release the physical I/O buffer 1758 */ 1759 if (bp->b_flags & B_ASYNC) { 1760 mtx_lock(&swbuf_mtx); 1761 if (++nsw_wcount_async == 1) 1762 wakeup(&nsw_wcount_async); 1763 mtx_unlock(&swbuf_mtx); 1764 } 1765 uma_zfree((bp->b_iocmd == BIO_READ) ? swrbuf_zone : swwbuf_zone, bp); 1766 } 1767 1768 int 1769 swap_pager_nswapdev(void) 1770 { 1771 1772 return (nswapdev); 1773 } 1774 1775 static void 1776 swp_pager_force_dirty(vm_page_t m) 1777 { 1778 1779 vm_page_dirty(m); 1780 swap_pager_unswapped(m); 1781 vm_page_launder(m); 1782 } 1783 1784 u_long 1785 swap_pager_swapped_pages(vm_object_t object) 1786 { 1787 struct swblk *sb; 1788 vm_pindex_t pi; 1789 u_long res; 1790 int i; 1791 1792 VM_OBJECT_ASSERT_LOCKED(object); 1793 1794 if (pctrie_is_empty(&object->un_pager.swp.swp_blks)) 1795 return (0); 1796 1797 for (res = 0, pi = 0; (sb = SWAP_PCTRIE_LOOKUP_GE( 1798 &object->un_pager.swp.swp_blks, pi)) != NULL; 1799 pi = sb->p + SWAP_META_PAGES) { 1800 for (i = 0; i < SWAP_META_PAGES; i++) { 1801 if (sb->d[i] != SWAPBLK_NONE) 1802 res++; 1803 } 1804 } 1805 return (res); 1806 } 1807 1808 /* 1809 * swap_pager_swapoff_object: 1810 * 1811 * Page in all of the pages that have been paged out for an object 1812 * to a swap device. 1813 */ 1814 static void 1815 swap_pager_swapoff_object(struct swdevt *sp, vm_object_t object) 1816 { 1817 struct swblk *sb; 1818 vm_page_t m; 1819 vm_pindex_t pi; 1820 daddr_t blk; 1821 int i, nv, rahead, rv; 1822 1823 KASSERT((object->flags & OBJ_SWAP) != 0, 1824 ("%s: Object not swappable", __func__)); 1825 1826 for (pi = 0; (sb = SWAP_PCTRIE_LOOKUP_GE( 1827 &object->un_pager.swp.swp_blks, pi)) != NULL; ) { 1828 if ((object->flags & OBJ_DEAD) != 0) { 1829 /* 1830 * Make sure that pending writes finish before 1831 * returning. 1832 */ 1833 vm_object_pip_wait(object, "swpoff"); 1834 swp_pager_meta_free_all(object); 1835 break; 1836 } 1837 for (i = 0; i < SWAP_META_PAGES; i++) { 1838 /* 1839 * Count the number of contiguous valid blocks. 1840 */ 1841 for (nv = 0; nv < SWAP_META_PAGES - i; nv++) { 1842 blk = sb->d[i + nv]; 1843 if (!swp_pager_isondev(blk, sp) || 1844 blk == SWAPBLK_NONE) 1845 break; 1846 } 1847 if (nv == 0) 1848 continue; 1849 1850 /* 1851 * Look for a page corresponding to the first 1852 * valid block and ensure that any pending paging 1853 * operations on it are complete. If the page is valid, 1854 * mark it dirty and free the swap block. Try to batch 1855 * this operation since it may cause sp to be freed, 1856 * meaning that we must restart the scan. Avoid busying 1857 * valid pages since we may block forever on kernel 1858 * stack pages. 1859 */ 1860 m = vm_page_lookup(object, sb->p + i); 1861 if (m == NULL) { 1862 m = vm_page_alloc(object, sb->p + i, 1863 VM_ALLOC_NORMAL | VM_ALLOC_WAITFAIL); 1864 if (m == NULL) 1865 break; 1866 } else { 1867 if ((m->oflags & VPO_SWAPINPROG) != 0) { 1868 m->oflags |= VPO_SWAPSLEEP; 1869 VM_OBJECT_SLEEP(object, &object->handle, 1870 PSWP, "swpoff", 0); 1871 break; 1872 } 1873 if (vm_page_all_valid(m)) { 1874 do { 1875 swp_pager_force_dirty(m); 1876 } while (--nv > 0 && 1877 (m = vm_page_next(m)) != NULL && 1878 vm_page_all_valid(m) && 1879 (m->oflags & VPO_SWAPINPROG) == 0); 1880 break; 1881 } 1882 if (!vm_page_busy_acquire(m, VM_ALLOC_WAITFAIL)) 1883 break; 1884 } 1885 1886 vm_object_pip_add(object, 1); 1887 rahead = SWAP_META_PAGES; 1888 rv = swap_pager_getpages_locked(object, &m, 1, NULL, 1889 &rahead); 1890 if (rv != VM_PAGER_OK) 1891 panic("%s: read from swap failed: %d", 1892 __func__, rv); 1893 vm_object_pip_wakeupn(object, 1); 1894 VM_OBJECT_WLOCK(object); 1895 vm_page_xunbusy(m); 1896 1897 /* 1898 * The object lock was dropped so we must restart the 1899 * scan of this swap block. Pages paged in during this 1900 * iteration will be marked dirty in a future iteration. 1901 */ 1902 break; 1903 } 1904 if (i == SWAP_META_PAGES) 1905 pi = sb->p + SWAP_META_PAGES; 1906 } 1907 } 1908 1909 /* 1910 * swap_pager_swapoff: 1911 * 1912 * Page in all of the pages that have been paged out to the 1913 * given device. The corresponding blocks in the bitmap must be 1914 * marked as allocated and the device must be flagged SW_CLOSING. 1915 * There may be no processes swapped out to the device. 1916 * 1917 * This routine may block. 1918 */ 1919 static void 1920 swap_pager_swapoff(struct swdevt *sp) 1921 { 1922 vm_object_t object; 1923 int retries; 1924 1925 sx_assert(&swdev_syscall_lock, SA_XLOCKED); 1926 1927 retries = 0; 1928 full_rescan: 1929 mtx_lock(&vm_object_list_mtx); 1930 TAILQ_FOREACH(object, &vm_object_list, object_list) { 1931 if ((object->flags & OBJ_SWAP) == 0) 1932 continue; 1933 mtx_unlock(&vm_object_list_mtx); 1934 /* Depends on type-stability. */ 1935 VM_OBJECT_WLOCK(object); 1936 1937 /* 1938 * Dead objects are eventually terminated on their own. 1939 */ 1940 if ((object->flags & OBJ_DEAD) != 0) 1941 goto next_obj; 1942 1943 /* 1944 * Sync with fences placed after pctrie 1945 * initialization. We must not access pctrie below 1946 * unless we checked that our object is swap and not 1947 * dead. 1948 */ 1949 atomic_thread_fence_acq(); 1950 if ((object->flags & OBJ_SWAP) == 0) 1951 goto next_obj; 1952 1953 swap_pager_swapoff_object(sp, object); 1954 next_obj: 1955 VM_OBJECT_WUNLOCK(object); 1956 mtx_lock(&vm_object_list_mtx); 1957 } 1958 mtx_unlock(&vm_object_list_mtx); 1959 1960 if (sp->sw_used) { 1961 /* 1962 * Objects may be locked or paging to the device being 1963 * removed, so we will miss their pages and need to 1964 * make another pass. We have marked this device as 1965 * SW_CLOSING, so the activity should finish soon. 1966 */ 1967 retries++; 1968 if (retries > 100) { 1969 panic("swapoff: failed to locate %d swap blocks", 1970 sp->sw_used); 1971 } 1972 pause("swpoff", hz / 20); 1973 goto full_rescan; 1974 } 1975 EVENTHANDLER_INVOKE(swapoff, sp); 1976 } 1977 1978 /************************************************************************ 1979 * SWAP META DATA * 1980 ************************************************************************ 1981 * 1982 * These routines manipulate the swap metadata stored in the 1983 * OBJT_SWAP object. 1984 * 1985 * Swap metadata is implemented with a global hash and not directly 1986 * linked into the object. Instead the object simply contains 1987 * appropriate tracking counters. 1988 */ 1989 1990 /* 1991 * SWP_PAGER_SWBLK_EMPTY() - is a range of blocks free? 1992 */ 1993 static bool 1994 swp_pager_swblk_empty(struct swblk *sb, int start, int limit) 1995 { 1996 int i; 1997 1998 MPASS(0 <= start && start <= limit && limit <= SWAP_META_PAGES); 1999 for (i = start; i < limit; i++) { 2000 if (sb->d[i] != SWAPBLK_NONE) 2001 return (false); 2002 } 2003 return (true); 2004 } 2005 2006 /* 2007 * SWP_PAGER_FREE_EMPTY_SWBLK() - frees if a block is free 2008 * 2009 * Nothing is done if the block is still in use. 2010 */ 2011 static void 2012 swp_pager_free_empty_swblk(vm_object_t object, struct swblk *sb) 2013 { 2014 2015 if (swp_pager_swblk_empty(sb, 0, SWAP_META_PAGES)) { 2016 SWAP_PCTRIE_REMOVE(&object->un_pager.swp.swp_blks, sb->p); 2017 uma_zfree(swblk_zone, sb); 2018 } 2019 } 2020 2021 /* 2022 * SWP_PAGER_META_BUILD() - add swap block to swap meta data for object 2023 * 2024 * The specified swapblk is added to the object's swap metadata. If 2025 * the swapblk is not valid, it is freed instead. Any previously 2026 * assigned swapblk is returned. 2027 */ 2028 static daddr_t 2029 swp_pager_meta_build(vm_object_t object, vm_pindex_t pindex, daddr_t swapblk) 2030 { 2031 static volatile int swblk_zone_exhausted, swpctrie_zone_exhausted; 2032 struct swblk *sb, *sb1; 2033 vm_pindex_t modpi, rdpi; 2034 daddr_t prev_swapblk; 2035 int error, i; 2036 2037 VM_OBJECT_ASSERT_WLOCKED(object); 2038 2039 rdpi = rounddown(pindex, SWAP_META_PAGES); 2040 sb = SWAP_PCTRIE_LOOKUP(&object->un_pager.swp.swp_blks, rdpi); 2041 if (sb == NULL) { 2042 if (swapblk == SWAPBLK_NONE) 2043 return (SWAPBLK_NONE); 2044 for (;;) { 2045 sb = uma_zalloc(swblk_zone, M_NOWAIT | (curproc == 2046 pageproc ? M_USE_RESERVE : 0)); 2047 if (sb != NULL) { 2048 sb->p = rdpi; 2049 for (i = 0; i < SWAP_META_PAGES; i++) 2050 sb->d[i] = SWAPBLK_NONE; 2051 if (atomic_cmpset_int(&swblk_zone_exhausted, 2052 1, 0)) 2053 printf("swblk zone ok\n"); 2054 break; 2055 } 2056 VM_OBJECT_WUNLOCK(object); 2057 if (uma_zone_exhausted(swblk_zone)) { 2058 if (atomic_cmpset_int(&swblk_zone_exhausted, 2059 0, 1)) 2060 printf("swap blk zone exhausted, " 2061 "increase kern.maxswzone\n"); 2062 vm_pageout_oom(VM_OOM_SWAPZ); 2063 pause("swzonxb", 10); 2064 } else 2065 uma_zwait(swblk_zone); 2066 VM_OBJECT_WLOCK(object); 2067 sb = SWAP_PCTRIE_LOOKUP(&object->un_pager.swp.swp_blks, 2068 rdpi); 2069 if (sb != NULL) 2070 /* 2071 * Somebody swapped out a nearby page, 2072 * allocating swblk at the rdpi index, 2073 * while we dropped the object lock. 2074 */ 2075 goto allocated; 2076 } 2077 for (;;) { 2078 error = SWAP_PCTRIE_INSERT( 2079 &object->un_pager.swp.swp_blks, sb); 2080 if (error == 0) { 2081 if (atomic_cmpset_int(&swpctrie_zone_exhausted, 2082 1, 0)) 2083 printf("swpctrie zone ok\n"); 2084 break; 2085 } 2086 VM_OBJECT_WUNLOCK(object); 2087 if (uma_zone_exhausted(swpctrie_zone)) { 2088 if (atomic_cmpset_int(&swpctrie_zone_exhausted, 2089 0, 1)) 2090 printf("swap pctrie zone exhausted, " 2091 "increase kern.maxswzone\n"); 2092 vm_pageout_oom(VM_OOM_SWAPZ); 2093 pause("swzonxp", 10); 2094 } else 2095 uma_zwait(swpctrie_zone); 2096 VM_OBJECT_WLOCK(object); 2097 sb1 = SWAP_PCTRIE_LOOKUP(&object->un_pager.swp.swp_blks, 2098 rdpi); 2099 if (sb1 != NULL) { 2100 uma_zfree(swblk_zone, sb); 2101 sb = sb1; 2102 goto allocated; 2103 } 2104 } 2105 } 2106 allocated: 2107 MPASS(sb->p == rdpi); 2108 2109 modpi = pindex % SWAP_META_PAGES; 2110 /* Return prior contents of metadata. */ 2111 prev_swapblk = sb->d[modpi]; 2112 /* Enter block into metadata. */ 2113 sb->d[modpi] = swapblk; 2114 2115 /* 2116 * Free the swblk if we end up with the empty page run. 2117 */ 2118 if (swapblk == SWAPBLK_NONE) 2119 swp_pager_free_empty_swblk(object, sb); 2120 return (prev_swapblk); 2121 } 2122 2123 /* 2124 * SWP_PAGER_META_TRANSFER() - free a range of blocks in the srcobject's swap 2125 * metadata, or transfer it into dstobject. 2126 * 2127 * This routine will free swap metadata structures as they are cleaned 2128 * out. 2129 */ 2130 static void 2131 swp_pager_meta_transfer(vm_object_t srcobject, vm_object_t dstobject, 2132 vm_pindex_t pindex, vm_pindex_t count, vm_size_t *moved) 2133 { 2134 struct swblk *sb; 2135 vm_page_t m; 2136 daddr_t n_free, s_free; 2137 vm_pindex_t offset, last; 2138 vm_size_t mc; 2139 int i, limit, start; 2140 2141 VM_OBJECT_ASSERT_WLOCKED(srcobject); 2142 MPASS(moved == NULL || dstobject == NULL); 2143 2144 mc = 0; 2145 m = NULL; 2146 if (count == 0 || pctrie_is_empty(&srcobject->un_pager.swp.swp_blks)) 2147 goto out; 2148 2149 swp_pager_init_freerange(&s_free, &n_free); 2150 offset = pindex; 2151 last = pindex + count; 2152 for (;;) { 2153 sb = SWAP_PCTRIE_LOOKUP_GE(&srcobject->un_pager.swp.swp_blks, 2154 rounddown(pindex, SWAP_META_PAGES)); 2155 if (sb == NULL || sb->p >= last) 2156 break; 2157 start = pindex > sb->p ? pindex - sb->p : 0; 2158 limit = last - sb->p < SWAP_META_PAGES ? last - sb->p : 2159 SWAP_META_PAGES; 2160 for (i = start; i < limit; i++) { 2161 if (sb->d[i] == SWAPBLK_NONE) 2162 continue; 2163 if (dstobject == NULL || 2164 !swp_pager_xfer_source(srcobject, dstobject, 2165 sb->p + i - offset, sb->d[i])) { 2166 swp_pager_update_freerange(&s_free, &n_free, 2167 sb->d[i]); 2168 } 2169 if (moved != NULL) { 2170 if (m != NULL && m->pindex != pindex + i - 1) 2171 m = NULL; 2172 m = m != NULL ? vm_page_next(m) : 2173 vm_page_lookup(srcobject, pindex + i); 2174 if (m == NULL || vm_page_none_valid(m)) 2175 mc++; 2176 } 2177 sb->d[i] = SWAPBLK_NONE; 2178 } 2179 pindex = sb->p + SWAP_META_PAGES; 2180 if (swp_pager_swblk_empty(sb, 0, start) && 2181 swp_pager_swblk_empty(sb, limit, SWAP_META_PAGES)) { 2182 SWAP_PCTRIE_REMOVE(&srcobject->un_pager.swp.swp_blks, 2183 sb->p); 2184 uma_zfree(swblk_zone, sb); 2185 } 2186 } 2187 swp_pager_freeswapspace(s_free, n_free); 2188 out: 2189 if (moved != NULL) 2190 *moved = mc; 2191 } 2192 2193 /* 2194 * SWP_PAGER_META_FREE() - free a range of blocks in the object's swap metadata 2195 * 2196 * The requested range of blocks is freed, with any associated swap 2197 * returned to the swap bitmap. 2198 * 2199 * This routine will free swap metadata structures as they are cleaned 2200 * out. This routine does *NOT* operate on swap metadata associated 2201 * with resident pages. 2202 */ 2203 static void 2204 swp_pager_meta_free(vm_object_t object, vm_pindex_t pindex, vm_pindex_t count, 2205 vm_size_t *freed) 2206 { 2207 swp_pager_meta_transfer(object, NULL, pindex, count, freed); 2208 } 2209 2210 /* 2211 * SWP_PAGER_META_FREE_ALL() - destroy all swap metadata associated with object 2212 * 2213 * This routine locates and destroys all swap metadata associated with 2214 * an object. 2215 */ 2216 static void 2217 swp_pager_meta_free_all(vm_object_t object) 2218 { 2219 struct swblk *sb; 2220 daddr_t n_free, s_free; 2221 vm_pindex_t pindex; 2222 int i; 2223 2224 VM_OBJECT_ASSERT_WLOCKED(object); 2225 2226 if (pctrie_is_empty(&object->un_pager.swp.swp_blks)) 2227 return; 2228 2229 swp_pager_init_freerange(&s_free, &n_free); 2230 for (pindex = 0; (sb = SWAP_PCTRIE_LOOKUP_GE( 2231 &object->un_pager.swp.swp_blks, pindex)) != NULL;) { 2232 pindex = sb->p + SWAP_META_PAGES; 2233 for (i = 0; i < SWAP_META_PAGES; i++) { 2234 if (sb->d[i] == SWAPBLK_NONE) 2235 continue; 2236 swp_pager_update_freerange(&s_free, &n_free, sb->d[i]); 2237 } 2238 SWAP_PCTRIE_REMOVE(&object->un_pager.swp.swp_blks, sb->p); 2239 uma_zfree(swblk_zone, sb); 2240 } 2241 swp_pager_freeswapspace(s_free, n_free); 2242 } 2243 2244 /* 2245 * SWP_PAGER_METACTL() - misc control of swap meta data. 2246 * 2247 * This routine is capable of looking up, or removing swapblk 2248 * assignments in the swap meta data. It returns the swapblk being 2249 * looked-up, popped, or SWAPBLK_NONE if the block was invalid. 2250 * 2251 * When acting on a busy resident page and paging is in progress, we 2252 * have to wait until paging is complete but otherwise can act on the 2253 * busy page. 2254 */ 2255 static daddr_t 2256 swp_pager_meta_lookup(vm_object_t object, vm_pindex_t pindex) 2257 { 2258 struct swblk *sb; 2259 2260 VM_OBJECT_ASSERT_LOCKED(object); 2261 2262 /* 2263 * The meta data only exists if the object is OBJT_SWAP 2264 * and even then might not be allocated yet. 2265 */ 2266 KASSERT((object->flags & OBJ_SWAP) != 0, 2267 ("Lookup object not swappable")); 2268 2269 sb = SWAP_PCTRIE_LOOKUP(&object->un_pager.swp.swp_blks, 2270 rounddown(pindex, SWAP_META_PAGES)); 2271 if (sb == NULL) 2272 return (SWAPBLK_NONE); 2273 return (sb->d[pindex % SWAP_META_PAGES]); 2274 } 2275 2276 /* 2277 * Returns the least page index which is greater than or equal to the 2278 * parameter pindex and for which there is a swap block allocated. 2279 * Returns object's size if the object's type is not swap or if there 2280 * are no allocated swap blocks for the object after the requested 2281 * pindex. 2282 */ 2283 vm_pindex_t 2284 swap_pager_find_least(vm_object_t object, vm_pindex_t pindex) 2285 { 2286 struct swblk *sb; 2287 int i; 2288 2289 VM_OBJECT_ASSERT_LOCKED(object); 2290 MPASS((object->flags & OBJ_SWAP) != 0); 2291 2292 if (pctrie_is_empty(&object->un_pager.swp.swp_blks)) 2293 return (object->size); 2294 sb = SWAP_PCTRIE_LOOKUP_GE(&object->un_pager.swp.swp_blks, 2295 rounddown(pindex, SWAP_META_PAGES)); 2296 if (sb == NULL) 2297 return (object->size); 2298 if (sb->p < pindex) { 2299 for (i = pindex % SWAP_META_PAGES; i < SWAP_META_PAGES; i++) { 2300 if (sb->d[i] != SWAPBLK_NONE) 2301 return (sb->p + i); 2302 } 2303 sb = SWAP_PCTRIE_LOOKUP_GE(&object->un_pager.swp.swp_blks, 2304 roundup(pindex, SWAP_META_PAGES)); 2305 if (sb == NULL) 2306 return (object->size); 2307 } 2308 for (i = 0; i < SWAP_META_PAGES; i++) { 2309 if (sb->d[i] != SWAPBLK_NONE) 2310 return (sb->p + i); 2311 } 2312 2313 /* 2314 * We get here if a swblk is present in the trie but it 2315 * doesn't map any blocks. 2316 */ 2317 MPASS(0); 2318 return (object->size); 2319 } 2320 2321 /* 2322 * System call swapon(name) enables swapping on device name, 2323 * which must be in the swdevsw. Return EBUSY 2324 * if already swapping on this device. 2325 */ 2326 #ifndef _SYS_SYSPROTO_H_ 2327 struct swapon_args { 2328 char *name; 2329 }; 2330 #endif 2331 2332 int 2333 sys_swapon(struct thread *td, struct swapon_args *uap) 2334 { 2335 struct vattr attr; 2336 struct vnode *vp; 2337 struct nameidata nd; 2338 int error; 2339 2340 error = priv_check(td, PRIV_SWAPON); 2341 if (error) 2342 return (error); 2343 2344 sx_xlock(&swdev_syscall_lock); 2345 2346 /* 2347 * Swap metadata may not fit in the KVM if we have physical 2348 * memory of >1GB. 2349 */ 2350 if (swblk_zone == NULL) { 2351 error = ENOMEM; 2352 goto done; 2353 } 2354 2355 NDINIT(&nd, LOOKUP, ISOPEN | FOLLOW | LOCKLEAF | AUDITVNODE1, 2356 UIO_USERSPACE, uap->name); 2357 error = namei(&nd); 2358 if (error) 2359 goto done; 2360 2361 NDFREE_PNBUF(&nd); 2362 vp = nd.ni_vp; 2363 2364 if (vn_isdisk_error(vp, &error)) { 2365 error = swapongeom(vp); 2366 } else if (vp->v_type == VREG && 2367 (vp->v_mount->mnt_vfc->vfc_flags & VFCF_NETWORK) != 0 && 2368 (error = VOP_GETATTR(vp, &attr, td->td_ucred)) == 0) { 2369 /* 2370 * Allow direct swapping to NFS regular files in the same 2371 * way that nfs_mountroot() sets up diskless swapping. 2372 */ 2373 error = swaponvp(td, vp, attr.va_size / DEV_BSIZE); 2374 } 2375 2376 if (error != 0) 2377 vput(vp); 2378 else 2379 VOP_UNLOCK(vp); 2380 done: 2381 sx_xunlock(&swdev_syscall_lock); 2382 return (error); 2383 } 2384 2385 /* 2386 * Check that the total amount of swap currently configured does not 2387 * exceed half the theoretical maximum. If it does, print a warning 2388 * message. 2389 */ 2390 static void 2391 swapon_check_swzone(void) 2392 { 2393 2394 /* recommend using no more than half that amount */ 2395 if (swap_total > swap_maxpages / 2) { 2396 printf("warning: total configured swap (%lu pages) " 2397 "exceeds maximum recommended amount (%lu pages).\n", 2398 swap_total, swap_maxpages / 2); 2399 printf("warning: increase kern.maxswzone " 2400 "or reduce amount of swap.\n"); 2401 } 2402 } 2403 2404 static void 2405 swaponsomething(struct vnode *vp, void *id, u_long nblks, 2406 sw_strategy_t *strategy, sw_close_t *close, dev_t dev, int flags) 2407 { 2408 struct swdevt *sp, *tsp; 2409 daddr_t dvbase; 2410 2411 /* 2412 * nblks is in DEV_BSIZE'd chunks, convert to PAGE_SIZE'd chunks. 2413 * First chop nblks off to page-align it, then convert. 2414 * 2415 * sw->sw_nblks is in page-sized chunks now too. 2416 */ 2417 nblks &= ~(ctodb(1) - 1); 2418 nblks = dbtoc(nblks); 2419 2420 sp = malloc(sizeof *sp, M_VMPGDATA, M_WAITOK | M_ZERO); 2421 sp->sw_blist = blist_create(nblks, M_WAITOK); 2422 sp->sw_vp = vp; 2423 sp->sw_id = id; 2424 sp->sw_dev = dev; 2425 sp->sw_nblks = nblks; 2426 sp->sw_used = 0; 2427 sp->sw_strategy = strategy; 2428 sp->sw_close = close; 2429 sp->sw_flags = flags; 2430 2431 /* 2432 * Do not free the first blocks in order to avoid overwriting 2433 * any bsd label at the front of the partition 2434 */ 2435 blist_free(sp->sw_blist, howmany(BBSIZE, PAGE_SIZE), 2436 nblks - howmany(BBSIZE, PAGE_SIZE)); 2437 2438 dvbase = 0; 2439 mtx_lock(&sw_dev_mtx); 2440 TAILQ_FOREACH(tsp, &swtailq, sw_list) { 2441 if (tsp->sw_end >= dvbase) { 2442 /* 2443 * We put one uncovered page between the devices 2444 * in order to definitively prevent any cross-device 2445 * I/O requests 2446 */ 2447 dvbase = tsp->sw_end + 1; 2448 } 2449 } 2450 sp->sw_first = dvbase; 2451 sp->sw_end = dvbase + nblks; 2452 TAILQ_INSERT_TAIL(&swtailq, sp, sw_list); 2453 nswapdev++; 2454 swap_pager_avail += nblks - howmany(BBSIZE, PAGE_SIZE); 2455 swap_total += nblks; 2456 swapon_check_swzone(); 2457 swp_sizecheck(); 2458 mtx_unlock(&sw_dev_mtx); 2459 EVENTHANDLER_INVOKE(swapon, sp); 2460 } 2461 2462 /* 2463 * SYSCALL: swapoff(devname) 2464 * 2465 * Disable swapping on the given device. 2466 * 2467 * XXX: Badly designed system call: it should use a device index 2468 * rather than filename as specification. We keep sw_vp around 2469 * only to make this work. 2470 */ 2471 static int 2472 kern_swapoff(struct thread *td, const char *name, enum uio_seg name_seg, 2473 u_int flags) 2474 { 2475 struct vnode *vp; 2476 struct nameidata nd; 2477 struct swdevt *sp; 2478 int error; 2479 2480 error = priv_check(td, PRIV_SWAPOFF); 2481 if (error != 0) 2482 return (error); 2483 if ((flags & ~(SWAPOFF_FORCE)) != 0) 2484 return (EINVAL); 2485 2486 sx_xlock(&swdev_syscall_lock); 2487 2488 NDINIT(&nd, LOOKUP, FOLLOW | AUDITVNODE1, name_seg, name); 2489 error = namei(&nd); 2490 if (error) 2491 goto done; 2492 NDFREE_PNBUF(&nd); 2493 vp = nd.ni_vp; 2494 2495 mtx_lock(&sw_dev_mtx); 2496 TAILQ_FOREACH(sp, &swtailq, sw_list) { 2497 if (sp->sw_vp == vp) 2498 break; 2499 } 2500 mtx_unlock(&sw_dev_mtx); 2501 if (sp == NULL) { 2502 error = EINVAL; 2503 goto done; 2504 } 2505 error = swapoff_one(sp, td->td_ucred, flags); 2506 done: 2507 sx_xunlock(&swdev_syscall_lock); 2508 return (error); 2509 } 2510 2511 2512 #ifdef COMPAT_FREEBSD13 2513 int 2514 freebsd13_swapoff(struct thread *td, struct freebsd13_swapoff_args *uap) 2515 { 2516 return (kern_swapoff(td, uap->name, UIO_USERSPACE, 0)); 2517 } 2518 #endif 2519 2520 int 2521 sys_swapoff(struct thread *td, struct swapoff_args *uap) 2522 { 2523 return (kern_swapoff(td, uap->name, UIO_USERSPACE, uap->flags)); 2524 } 2525 2526 static int 2527 swapoff_one(struct swdevt *sp, struct ucred *cred, u_int flags) 2528 { 2529 u_long nblks; 2530 #ifdef MAC 2531 int error; 2532 #endif 2533 2534 sx_assert(&swdev_syscall_lock, SA_XLOCKED); 2535 #ifdef MAC 2536 (void) vn_lock(sp->sw_vp, LK_EXCLUSIVE | LK_RETRY); 2537 error = mac_system_check_swapoff(cred, sp->sw_vp); 2538 (void) VOP_UNLOCK(sp->sw_vp); 2539 if (error != 0) 2540 return (error); 2541 #endif 2542 nblks = sp->sw_nblks; 2543 2544 /* 2545 * We can turn off this swap device safely only if the 2546 * available virtual memory in the system will fit the amount 2547 * of data we will have to page back in, plus an epsilon so 2548 * the system doesn't become critically low on swap space. 2549 * The vm_free_count() part does not account e.g. for clean 2550 * pages that can be immediately reclaimed without paging, so 2551 * this is a very rough estimation. 2552 * 2553 * On the other hand, not turning swap off on swapoff_all() 2554 * means that we can lose swap data when filesystems go away, 2555 * which is arguably worse. 2556 */ 2557 if ((flags & SWAPOFF_FORCE) == 0 && 2558 vm_free_count() + swap_pager_avail < nblks + nswap_lowat) 2559 return (ENOMEM); 2560 2561 /* 2562 * Prevent further allocations on this device. 2563 */ 2564 mtx_lock(&sw_dev_mtx); 2565 sp->sw_flags |= SW_CLOSING; 2566 swap_pager_avail -= blist_fill(sp->sw_blist, 0, nblks); 2567 swap_total -= nblks; 2568 mtx_unlock(&sw_dev_mtx); 2569 2570 /* 2571 * Page in the contents of the device and close it. 2572 */ 2573 swap_pager_swapoff(sp); 2574 2575 sp->sw_close(curthread, sp); 2576 mtx_lock(&sw_dev_mtx); 2577 sp->sw_id = NULL; 2578 TAILQ_REMOVE(&swtailq, sp, sw_list); 2579 nswapdev--; 2580 if (nswapdev == 0) { 2581 swap_pager_full = 2; 2582 swap_pager_almost_full = 1; 2583 } 2584 if (swdevhd == sp) 2585 swdevhd = NULL; 2586 mtx_unlock(&sw_dev_mtx); 2587 blist_destroy(sp->sw_blist); 2588 free(sp, M_VMPGDATA); 2589 return (0); 2590 } 2591 2592 void 2593 swapoff_all(void) 2594 { 2595 struct swdevt *sp, *spt; 2596 const char *devname; 2597 int error; 2598 2599 sx_xlock(&swdev_syscall_lock); 2600 2601 mtx_lock(&sw_dev_mtx); 2602 TAILQ_FOREACH_SAFE(sp, &swtailq, sw_list, spt) { 2603 mtx_unlock(&sw_dev_mtx); 2604 if (vn_isdisk(sp->sw_vp)) 2605 devname = devtoname(sp->sw_vp->v_rdev); 2606 else 2607 devname = "[file]"; 2608 error = swapoff_one(sp, thread0.td_ucred, SWAPOFF_FORCE); 2609 if (error != 0) { 2610 printf("Cannot remove swap device %s (error=%d), " 2611 "skipping.\n", devname, error); 2612 } else if (bootverbose) { 2613 printf("Swap device %s removed.\n", devname); 2614 } 2615 mtx_lock(&sw_dev_mtx); 2616 } 2617 mtx_unlock(&sw_dev_mtx); 2618 2619 sx_xunlock(&swdev_syscall_lock); 2620 } 2621 2622 void 2623 swap_pager_status(int *total, int *used) 2624 { 2625 2626 *total = swap_total; 2627 *used = swap_total - swap_pager_avail - 2628 nswapdev * howmany(BBSIZE, PAGE_SIZE); 2629 } 2630 2631 int 2632 swap_dev_info(int name, struct xswdev *xs, char *devname, size_t len) 2633 { 2634 struct swdevt *sp; 2635 const char *tmp_devname; 2636 int error, n; 2637 2638 n = 0; 2639 error = ENOENT; 2640 mtx_lock(&sw_dev_mtx); 2641 TAILQ_FOREACH(sp, &swtailq, sw_list) { 2642 if (n != name) { 2643 n++; 2644 continue; 2645 } 2646 xs->xsw_version = XSWDEV_VERSION; 2647 xs->xsw_dev = sp->sw_dev; 2648 xs->xsw_flags = sp->sw_flags; 2649 xs->xsw_nblks = sp->sw_nblks; 2650 xs->xsw_used = sp->sw_used; 2651 if (devname != NULL) { 2652 if (vn_isdisk(sp->sw_vp)) 2653 tmp_devname = devtoname(sp->sw_vp->v_rdev); 2654 else 2655 tmp_devname = "[file]"; 2656 strncpy(devname, tmp_devname, len); 2657 } 2658 error = 0; 2659 break; 2660 } 2661 mtx_unlock(&sw_dev_mtx); 2662 return (error); 2663 } 2664 2665 #if defined(COMPAT_FREEBSD11) 2666 #define XSWDEV_VERSION_11 1 2667 struct xswdev11 { 2668 u_int xsw_version; 2669 uint32_t xsw_dev; 2670 int xsw_flags; 2671 int xsw_nblks; 2672 int xsw_used; 2673 }; 2674 #endif 2675 2676 #if defined(__amd64__) && defined(COMPAT_FREEBSD32) 2677 struct xswdev32 { 2678 u_int xsw_version; 2679 u_int xsw_dev1, xsw_dev2; 2680 int xsw_flags; 2681 int xsw_nblks; 2682 int xsw_used; 2683 }; 2684 #endif 2685 2686 static int 2687 sysctl_vm_swap_info(SYSCTL_HANDLER_ARGS) 2688 { 2689 struct xswdev xs; 2690 #if defined(__amd64__) && defined(COMPAT_FREEBSD32) 2691 struct xswdev32 xs32; 2692 #endif 2693 #if defined(COMPAT_FREEBSD11) 2694 struct xswdev11 xs11; 2695 #endif 2696 int error; 2697 2698 if (arg2 != 1) /* name length */ 2699 return (EINVAL); 2700 2701 memset(&xs, 0, sizeof(xs)); 2702 error = swap_dev_info(*(int *)arg1, &xs, NULL, 0); 2703 if (error != 0) 2704 return (error); 2705 #if defined(__amd64__) && defined(COMPAT_FREEBSD32) 2706 if (req->oldlen == sizeof(xs32)) { 2707 memset(&xs32, 0, sizeof(xs32)); 2708 xs32.xsw_version = XSWDEV_VERSION; 2709 xs32.xsw_dev1 = xs.xsw_dev; 2710 xs32.xsw_dev2 = xs.xsw_dev >> 32; 2711 xs32.xsw_flags = xs.xsw_flags; 2712 xs32.xsw_nblks = xs.xsw_nblks; 2713 xs32.xsw_used = xs.xsw_used; 2714 error = SYSCTL_OUT(req, &xs32, sizeof(xs32)); 2715 return (error); 2716 } 2717 #endif 2718 #if defined(COMPAT_FREEBSD11) 2719 if (req->oldlen == sizeof(xs11)) { 2720 memset(&xs11, 0, sizeof(xs11)); 2721 xs11.xsw_version = XSWDEV_VERSION_11; 2722 xs11.xsw_dev = xs.xsw_dev; /* truncation */ 2723 xs11.xsw_flags = xs.xsw_flags; 2724 xs11.xsw_nblks = xs.xsw_nblks; 2725 xs11.xsw_used = xs.xsw_used; 2726 error = SYSCTL_OUT(req, &xs11, sizeof(xs11)); 2727 return (error); 2728 } 2729 #endif 2730 error = SYSCTL_OUT(req, &xs, sizeof(xs)); 2731 return (error); 2732 } 2733 2734 SYSCTL_INT(_vm, OID_AUTO, nswapdev, CTLFLAG_RD, &nswapdev, 0, 2735 "Number of swap devices"); 2736 SYSCTL_NODE(_vm, OID_AUTO, swap_info, CTLFLAG_RD | CTLFLAG_MPSAFE, 2737 sysctl_vm_swap_info, 2738 "Swap statistics by device"); 2739 2740 /* 2741 * Count the approximate swap usage in pages for a vmspace. The 2742 * shadowed or not yet copied on write swap blocks are not accounted. 2743 * The map must be locked. 2744 */ 2745 long 2746 vmspace_swap_count(struct vmspace *vmspace) 2747 { 2748 vm_map_t map; 2749 vm_map_entry_t cur; 2750 vm_object_t object; 2751 struct swblk *sb; 2752 vm_pindex_t e, pi; 2753 long count; 2754 int i; 2755 2756 map = &vmspace->vm_map; 2757 count = 0; 2758 2759 VM_MAP_ENTRY_FOREACH(cur, map) { 2760 if ((cur->eflags & MAP_ENTRY_IS_SUB_MAP) != 0) 2761 continue; 2762 object = cur->object.vm_object; 2763 if (object == NULL || (object->flags & OBJ_SWAP) == 0) 2764 continue; 2765 VM_OBJECT_RLOCK(object); 2766 if ((object->flags & OBJ_SWAP) == 0) 2767 goto unlock; 2768 pi = OFF_TO_IDX(cur->offset); 2769 e = pi + OFF_TO_IDX(cur->end - cur->start); 2770 for (;; pi = sb->p + SWAP_META_PAGES) { 2771 sb = SWAP_PCTRIE_LOOKUP_GE( 2772 &object->un_pager.swp.swp_blks, pi); 2773 if (sb == NULL || sb->p >= e) 2774 break; 2775 for (i = 0; i < SWAP_META_PAGES; i++) { 2776 if (sb->p + i < e && 2777 sb->d[i] != SWAPBLK_NONE) 2778 count++; 2779 } 2780 } 2781 unlock: 2782 VM_OBJECT_RUNLOCK(object); 2783 } 2784 return (count); 2785 } 2786 2787 /* 2788 * GEOM backend 2789 * 2790 * Swapping onto disk devices. 2791 * 2792 */ 2793 2794 static g_orphan_t swapgeom_orphan; 2795 2796 static struct g_class g_swap_class = { 2797 .name = "SWAP", 2798 .version = G_VERSION, 2799 .orphan = swapgeom_orphan, 2800 }; 2801 2802 DECLARE_GEOM_CLASS(g_swap_class, g_class); 2803 2804 static void 2805 swapgeom_close_ev(void *arg, int flags) 2806 { 2807 struct g_consumer *cp; 2808 2809 cp = arg; 2810 g_access(cp, -1, -1, 0); 2811 g_detach(cp); 2812 g_destroy_consumer(cp); 2813 } 2814 2815 /* 2816 * Add a reference to the g_consumer for an inflight transaction. 2817 */ 2818 static void 2819 swapgeom_acquire(struct g_consumer *cp) 2820 { 2821 2822 mtx_assert(&sw_dev_mtx, MA_OWNED); 2823 cp->index++; 2824 } 2825 2826 /* 2827 * Remove a reference from the g_consumer. Post a close event if all 2828 * references go away, since the function might be called from the 2829 * biodone context. 2830 */ 2831 static void 2832 swapgeom_release(struct g_consumer *cp, struct swdevt *sp) 2833 { 2834 2835 mtx_assert(&sw_dev_mtx, MA_OWNED); 2836 cp->index--; 2837 if (cp->index == 0) { 2838 if (g_post_event(swapgeom_close_ev, cp, M_NOWAIT, NULL) == 0) 2839 sp->sw_id = NULL; 2840 } 2841 } 2842 2843 static void 2844 swapgeom_done(struct bio *bp2) 2845 { 2846 struct swdevt *sp; 2847 struct buf *bp; 2848 struct g_consumer *cp; 2849 2850 bp = bp2->bio_caller2; 2851 cp = bp2->bio_from; 2852 bp->b_ioflags = bp2->bio_flags; 2853 if (bp2->bio_error) 2854 bp->b_ioflags |= BIO_ERROR; 2855 bp->b_resid = bp->b_bcount - bp2->bio_completed; 2856 bp->b_error = bp2->bio_error; 2857 bp->b_caller1 = NULL; 2858 bufdone(bp); 2859 sp = bp2->bio_caller1; 2860 mtx_lock(&sw_dev_mtx); 2861 swapgeom_release(cp, sp); 2862 mtx_unlock(&sw_dev_mtx); 2863 g_destroy_bio(bp2); 2864 } 2865 2866 static void 2867 swapgeom_strategy(struct buf *bp, struct swdevt *sp) 2868 { 2869 struct bio *bio; 2870 struct g_consumer *cp; 2871 2872 mtx_lock(&sw_dev_mtx); 2873 cp = sp->sw_id; 2874 if (cp == NULL) { 2875 mtx_unlock(&sw_dev_mtx); 2876 bp->b_error = ENXIO; 2877 bp->b_ioflags |= BIO_ERROR; 2878 bufdone(bp); 2879 return; 2880 } 2881 swapgeom_acquire(cp); 2882 mtx_unlock(&sw_dev_mtx); 2883 if (bp->b_iocmd == BIO_WRITE) 2884 bio = g_new_bio(); 2885 else 2886 bio = g_alloc_bio(); 2887 if (bio == NULL) { 2888 mtx_lock(&sw_dev_mtx); 2889 swapgeom_release(cp, sp); 2890 mtx_unlock(&sw_dev_mtx); 2891 bp->b_error = ENOMEM; 2892 bp->b_ioflags |= BIO_ERROR; 2893 printf("swap_pager: cannot allocate bio\n"); 2894 bufdone(bp); 2895 return; 2896 } 2897 2898 bp->b_caller1 = bio; 2899 bio->bio_caller1 = sp; 2900 bio->bio_caller2 = bp; 2901 bio->bio_cmd = bp->b_iocmd; 2902 bio->bio_offset = (bp->b_blkno - sp->sw_first) * PAGE_SIZE; 2903 bio->bio_length = bp->b_bcount; 2904 bio->bio_done = swapgeom_done; 2905 bio->bio_flags |= BIO_SWAP; 2906 if (!buf_mapped(bp)) { 2907 bio->bio_ma = bp->b_pages; 2908 bio->bio_data = unmapped_buf; 2909 bio->bio_ma_offset = (vm_offset_t)bp->b_offset & PAGE_MASK; 2910 bio->bio_ma_n = bp->b_npages; 2911 bio->bio_flags |= BIO_UNMAPPED; 2912 } else { 2913 bio->bio_data = bp->b_data; 2914 bio->bio_ma = NULL; 2915 } 2916 g_io_request(bio, cp); 2917 return; 2918 } 2919 2920 static void 2921 swapgeom_orphan(struct g_consumer *cp) 2922 { 2923 struct swdevt *sp; 2924 int destroy; 2925 2926 mtx_lock(&sw_dev_mtx); 2927 TAILQ_FOREACH(sp, &swtailq, sw_list) { 2928 if (sp->sw_id == cp) { 2929 sp->sw_flags |= SW_CLOSING; 2930 break; 2931 } 2932 } 2933 /* 2934 * Drop reference we were created with. Do directly since we're in a 2935 * special context where we don't have to queue the call to 2936 * swapgeom_close_ev(). 2937 */ 2938 cp->index--; 2939 destroy = ((sp != NULL) && (cp->index == 0)); 2940 if (destroy) 2941 sp->sw_id = NULL; 2942 mtx_unlock(&sw_dev_mtx); 2943 if (destroy) 2944 swapgeom_close_ev(cp, 0); 2945 } 2946 2947 static void 2948 swapgeom_close(struct thread *td, struct swdevt *sw) 2949 { 2950 struct g_consumer *cp; 2951 2952 mtx_lock(&sw_dev_mtx); 2953 cp = sw->sw_id; 2954 sw->sw_id = NULL; 2955 mtx_unlock(&sw_dev_mtx); 2956 2957 /* 2958 * swapgeom_close() may be called from the biodone context, 2959 * where we cannot perform topology changes. Delegate the 2960 * work to the events thread. 2961 */ 2962 if (cp != NULL) 2963 g_waitfor_event(swapgeom_close_ev, cp, M_WAITOK, NULL); 2964 } 2965 2966 static int 2967 swapongeom_locked(struct cdev *dev, struct vnode *vp) 2968 { 2969 struct g_provider *pp; 2970 struct g_consumer *cp; 2971 static struct g_geom *gp; 2972 struct swdevt *sp; 2973 u_long nblks; 2974 int error; 2975 2976 pp = g_dev_getprovider(dev); 2977 if (pp == NULL) 2978 return (ENODEV); 2979 mtx_lock(&sw_dev_mtx); 2980 TAILQ_FOREACH(sp, &swtailq, sw_list) { 2981 cp = sp->sw_id; 2982 if (cp != NULL && cp->provider == pp) { 2983 mtx_unlock(&sw_dev_mtx); 2984 return (EBUSY); 2985 } 2986 } 2987 mtx_unlock(&sw_dev_mtx); 2988 if (gp == NULL) 2989 gp = g_new_geomf(&g_swap_class, "swap"); 2990 cp = g_new_consumer(gp); 2991 cp->index = 1; /* Number of active I/Os, plus one for being active. */ 2992 cp->flags |= G_CF_DIRECT_SEND | G_CF_DIRECT_RECEIVE; 2993 g_attach(cp, pp); 2994 /* 2995 * XXX: Every time you think you can improve the margin for 2996 * footshooting, somebody depends on the ability to do so: 2997 * savecore(8) wants to write to our swapdev so we cannot 2998 * set an exclusive count :-( 2999 */ 3000 error = g_access(cp, 1, 1, 0); 3001 if (error != 0) { 3002 g_detach(cp); 3003 g_destroy_consumer(cp); 3004 return (error); 3005 } 3006 nblks = pp->mediasize / DEV_BSIZE; 3007 swaponsomething(vp, cp, nblks, swapgeom_strategy, 3008 swapgeom_close, dev2udev(dev), 3009 (pp->flags & G_PF_ACCEPT_UNMAPPED) != 0 ? SW_UNMAPPED : 0); 3010 return (0); 3011 } 3012 3013 static int 3014 swapongeom(struct vnode *vp) 3015 { 3016 int error; 3017 3018 ASSERT_VOP_ELOCKED(vp, "swapongeom"); 3019 if (vp->v_type != VCHR || VN_IS_DOOMED(vp)) { 3020 error = ENOENT; 3021 } else { 3022 g_topology_lock(); 3023 error = swapongeom_locked(vp->v_rdev, vp); 3024 g_topology_unlock(); 3025 } 3026 return (error); 3027 } 3028 3029 /* 3030 * VNODE backend 3031 * 3032 * This is used mainly for network filesystem (read: probably only tested 3033 * with NFS) swapfiles. 3034 * 3035 */ 3036 3037 static void 3038 swapdev_strategy(struct buf *bp, struct swdevt *sp) 3039 { 3040 struct vnode *vp2; 3041 3042 bp->b_blkno = ctodb(bp->b_blkno - sp->sw_first); 3043 3044 vp2 = sp->sw_id; 3045 vhold(vp2); 3046 if (bp->b_iocmd == BIO_WRITE) { 3047 vn_lock(vp2, LK_EXCLUSIVE | LK_RETRY); 3048 if (bp->b_bufobj) 3049 bufobj_wdrop(bp->b_bufobj); 3050 bufobj_wref(&vp2->v_bufobj); 3051 } else { 3052 vn_lock(vp2, LK_SHARED | LK_RETRY); 3053 } 3054 if (bp->b_bufobj != &vp2->v_bufobj) 3055 bp->b_bufobj = &vp2->v_bufobj; 3056 bp->b_vp = vp2; 3057 bp->b_iooffset = dbtob(bp->b_blkno); 3058 bstrategy(bp); 3059 VOP_UNLOCK(vp2); 3060 } 3061 3062 static void 3063 swapdev_close(struct thread *td, struct swdevt *sp) 3064 { 3065 struct vnode *vp; 3066 3067 vp = sp->sw_vp; 3068 vn_lock(vp, LK_EXCLUSIVE | LK_RETRY); 3069 VOP_CLOSE(vp, FREAD | FWRITE, td->td_ucred, td); 3070 vput(vp); 3071 } 3072 3073 static int 3074 swaponvp(struct thread *td, struct vnode *vp, u_long nblks) 3075 { 3076 struct swdevt *sp; 3077 int error; 3078 3079 ASSERT_VOP_ELOCKED(vp, "swaponvp"); 3080 if (nblks == 0) 3081 return (ENXIO); 3082 mtx_lock(&sw_dev_mtx); 3083 TAILQ_FOREACH(sp, &swtailq, sw_list) { 3084 if (sp->sw_id == vp) { 3085 mtx_unlock(&sw_dev_mtx); 3086 return (EBUSY); 3087 } 3088 } 3089 mtx_unlock(&sw_dev_mtx); 3090 3091 #ifdef MAC 3092 error = mac_system_check_swapon(td->td_ucred, vp); 3093 if (error == 0) 3094 #endif 3095 error = VOP_OPEN(vp, FREAD | FWRITE, td->td_ucred, td, NULL); 3096 if (error != 0) 3097 return (error); 3098 3099 swaponsomething(vp, vp, nblks, swapdev_strategy, swapdev_close, 3100 NODEV, 0); 3101 return (0); 3102 } 3103 3104 static int 3105 sysctl_swap_async_max(SYSCTL_HANDLER_ARGS) 3106 { 3107 int error, new, n; 3108 3109 new = nsw_wcount_async_max; 3110 error = sysctl_handle_int(oidp, &new, 0, req); 3111 if (error != 0 || req->newptr == NULL) 3112 return (error); 3113 3114 if (new > nswbuf / 2 || new < 1) 3115 return (EINVAL); 3116 3117 mtx_lock(&swbuf_mtx); 3118 while (nsw_wcount_async_max != new) { 3119 /* 3120 * Adjust difference. If the current async count is too low, 3121 * we will need to sqeeze our update slowly in. Sleep with a 3122 * higher priority than getpbuf() to finish faster. 3123 */ 3124 n = new - nsw_wcount_async_max; 3125 if (nsw_wcount_async + n >= 0) { 3126 nsw_wcount_async += n; 3127 nsw_wcount_async_max += n; 3128 wakeup(&nsw_wcount_async); 3129 } else { 3130 nsw_wcount_async_max -= nsw_wcount_async; 3131 nsw_wcount_async = 0; 3132 msleep(&nsw_wcount_async, &swbuf_mtx, PSWP, 3133 "swpsysctl", 0); 3134 } 3135 } 3136 mtx_unlock(&swbuf_mtx); 3137 3138 return (0); 3139 } 3140 3141 static void 3142 swap_pager_update_writecount(vm_object_t object, vm_offset_t start, 3143 vm_offset_t end) 3144 { 3145 3146 VM_OBJECT_WLOCK(object); 3147 KASSERT((object->flags & OBJ_ANON) == 0, 3148 ("Splittable object with writecount")); 3149 object->un_pager.swp.writemappings += (vm_ooffset_t)end - start; 3150 VM_OBJECT_WUNLOCK(object); 3151 } 3152 3153 static void 3154 swap_pager_release_writecount(vm_object_t object, vm_offset_t start, 3155 vm_offset_t end) 3156 { 3157 3158 VM_OBJECT_WLOCK(object); 3159 KASSERT((object->flags & OBJ_ANON) == 0, 3160 ("Splittable object with writecount")); 3161 object->un_pager.swp.writemappings -= (vm_ooffset_t)end - start; 3162 VM_OBJECT_WUNLOCK(object); 3163 } 3164