1 /*- 2 * SPDX-License-Identifier: BSD-4-Clause 3 * 4 * Copyright (c) 1998 Matthew Dillon, 5 * Copyright (c) 1994 John S. Dyson 6 * Copyright (c) 1990 University of Utah. 7 * Copyright (c) 1982, 1986, 1989, 1993 8 * The Regents of the University of California. All rights reserved. 9 * 10 * This code is derived from software contributed to Berkeley by 11 * the Systems Programming Group of the University of Utah Computer 12 * Science Department. 13 * 14 * Redistribution and use in source and binary forms, with or without 15 * modification, are permitted provided that the following conditions 16 * are met: 17 * 1. Redistributions of source code must retain the above copyright 18 * notice, this list of conditions and the following disclaimer. 19 * 2. Redistributions in binary form must reproduce the above copyright 20 * notice, this list of conditions and the following disclaimer in the 21 * documentation and/or other materials provided with the distribution. 22 * 3. All advertising materials mentioning features or use of this software 23 * must display the following acknowledgement: 24 * This product includes software developed by the University of 25 * California, Berkeley and its contributors. 26 * 4. Neither the name of the University nor the names of its contributors 27 * may be used to endorse or promote products derived from this software 28 * without specific prior written permission. 29 * 30 * THIS SOFTWARE IS PROVIDED BY THE REGENTS AND CONTRIBUTORS ``AS IS'' AND 31 * ANY EXPRESS OR IMPLIED WARRANTIES, INCLUDING, BUT NOT LIMITED TO, THE 32 * IMPLIED WARRANTIES OF MERCHANTABILITY AND FITNESS FOR A PARTICULAR PURPOSE 33 * ARE DISCLAIMED. IN NO EVENT SHALL THE REGENTS OR CONTRIBUTORS BE LIABLE 34 * FOR ANY DIRECT, INDIRECT, INCIDENTAL, SPECIAL, EXEMPLARY, OR CONSEQUENTIAL 35 * DAMAGES (INCLUDING, BUT NOT LIMITED TO, PROCUREMENT OF SUBSTITUTE GOODS 36 * OR SERVICES; LOSS OF USE, DATA, OR PROFITS; OR BUSINESS INTERRUPTION) 37 * HOWEVER CAUSED AND ON ANY THEORY OF LIABILITY, WHETHER IN CONTRACT, STRICT 38 * LIABILITY, OR TORT (INCLUDING NEGLIGENCE OR OTHERWISE) ARISING IN ANY WAY 39 * OUT OF THE USE OF THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF 40 * SUCH DAMAGE. 41 * 42 * New Swap System 43 * Matthew Dillon 44 * 45 * Radix Bitmap 'blists'. 46 * 47 * - The new swapper uses the new radix bitmap code. This should scale 48 * to arbitrarily small or arbitrarily large swap spaces and an almost 49 * arbitrary degree of fragmentation. 50 * 51 * Features: 52 * 53 * - on the fly reallocation of swap during putpages. The new system 54 * does not try to keep previously allocated swap blocks for dirty 55 * pages. 56 * 57 * - on the fly deallocation of swap 58 * 59 * - No more garbage collection required. Unnecessarily allocated swap 60 * blocks only exist for dirty vm_page_t's now and these are already 61 * cycled (in a high-load system) by the pager. We also do on-the-fly 62 * removal of invalidated swap blocks when a page is destroyed 63 * or renamed. 64 * 65 * from: Utah $Hdr: swap_pager.c 1.4 91/04/30$ 66 * 67 * @(#)swap_pager.c 8.9 (Berkeley) 3/21/94 68 * @(#)vm_swap.c 8.5 (Berkeley) 2/17/94 69 */ 70 71 #include <sys/cdefs.h> 72 __FBSDID("$FreeBSD$"); 73 74 #include "opt_vm.h" 75 76 #include <sys/param.h> 77 #include <sys/bio.h> 78 #include <sys/blist.h> 79 #include <sys/buf.h> 80 #include <sys/conf.h> 81 #include <sys/disk.h> 82 #include <sys/disklabel.h> 83 #include <sys/eventhandler.h> 84 #include <sys/fcntl.h> 85 #include <sys/limits.h> 86 #include <sys/lock.h> 87 #include <sys/kernel.h> 88 #include <sys/mount.h> 89 #include <sys/namei.h> 90 #include <sys/malloc.h> 91 #include <sys/pctrie.h> 92 #include <sys/priv.h> 93 #include <sys/proc.h> 94 #include <sys/racct.h> 95 #include <sys/resource.h> 96 #include <sys/resourcevar.h> 97 #include <sys/rwlock.h> 98 #include <sys/sbuf.h> 99 #include <sys/sysctl.h> 100 #include <sys/sysproto.h> 101 #include <sys/systm.h> 102 #include <sys/sx.h> 103 #include <sys/unistd.h> 104 #include <sys/user.h> 105 #include <sys/vmmeter.h> 106 #include <sys/vnode.h> 107 108 #include <security/mac/mac_framework.h> 109 110 #include <vm/vm.h> 111 #include <vm/pmap.h> 112 #include <vm/vm_map.h> 113 #include <vm/vm_kern.h> 114 #include <vm/vm_object.h> 115 #include <vm/vm_page.h> 116 #include <vm/vm_pager.h> 117 #include <vm/vm_pageout.h> 118 #include <vm/vm_param.h> 119 #include <vm/swap_pager.h> 120 #include <vm/vm_extern.h> 121 #include <vm/uma.h> 122 123 #include <geom/geom.h> 124 125 /* 126 * MAX_PAGEOUT_CLUSTER must be a power of 2 between 1 and 64. 127 * The 64-page limit is due to the radix code (kern/subr_blist.c). 128 */ 129 #ifndef MAX_PAGEOUT_CLUSTER 130 #define MAX_PAGEOUT_CLUSTER 32 131 #endif 132 133 #if !defined(SWB_NPAGES) 134 #define SWB_NPAGES MAX_PAGEOUT_CLUSTER 135 #endif 136 137 #define SWAP_META_PAGES PCTRIE_COUNT 138 139 /* 140 * A swblk structure maps each page index within a 141 * SWAP_META_PAGES-aligned and sized range to the address of an 142 * on-disk swap block (or SWAPBLK_NONE). The collection of these 143 * mappings for an entire vm object is implemented as a pc-trie. 144 */ 145 struct swblk { 146 vm_pindex_t p; 147 daddr_t d[SWAP_META_PAGES]; 148 }; 149 150 static MALLOC_DEFINE(M_VMPGDATA, "vm_pgdata", "swap pager private data"); 151 static struct mtx sw_dev_mtx; 152 static TAILQ_HEAD(, swdevt) swtailq = TAILQ_HEAD_INITIALIZER(swtailq); 153 static struct swdevt *swdevhd; /* Allocate from here next */ 154 static int nswapdev; /* Number of swap devices */ 155 int swap_pager_avail; 156 static struct sx swdev_syscall_lock; /* serialize swap(on|off) */ 157 158 static __exclusive_cache_line u_long swap_reserved; 159 static u_long swap_total; 160 static int sysctl_page_shift(SYSCTL_HANDLER_ARGS); 161 162 static SYSCTL_NODE(_vm_stats, OID_AUTO, swap, CTLFLAG_RD | CTLFLAG_MPSAFE, 0, 163 "VM swap stats"); 164 165 SYSCTL_PROC(_vm, OID_AUTO, swap_reserved, CTLTYPE_U64 | CTLFLAG_RD | CTLFLAG_MPSAFE, 166 &swap_reserved, 0, sysctl_page_shift, "A", 167 "Amount of swap storage needed to back all allocated anonymous memory."); 168 SYSCTL_PROC(_vm, OID_AUTO, swap_total, CTLTYPE_U64 | CTLFLAG_RD | CTLFLAG_MPSAFE, 169 &swap_total, 0, sysctl_page_shift, "A", 170 "Total amount of available swap storage."); 171 172 static int overcommit = 0; 173 SYSCTL_INT(_vm, VM_OVERCOMMIT, overcommit, CTLFLAG_RW, &overcommit, 0, 174 "Configure virtual memory overcommit behavior. See tuning(7) " 175 "for details."); 176 static unsigned long swzone; 177 SYSCTL_ULONG(_vm, OID_AUTO, swzone, CTLFLAG_RD, &swzone, 0, 178 "Actual size of swap metadata zone"); 179 static unsigned long swap_maxpages; 180 SYSCTL_ULONG(_vm, OID_AUTO, swap_maxpages, CTLFLAG_RD, &swap_maxpages, 0, 181 "Maximum amount of swap supported"); 182 183 static COUNTER_U64_DEFINE_EARLY(swap_free_deferred); 184 SYSCTL_COUNTER_U64(_vm_stats_swap, OID_AUTO, free_deferred, 185 CTLFLAG_RD, &swap_free_deferred, 186 "Number of pages that deferred freeing swap space"); 187 188 static COUNTER_U64_DEFINE_EARLY(swap_free_completed); 189 SYSCTL_COUNTER_U64(_vm_stats_swap, OID_AUTO, free_completed, 190 CTLFLAG_RD, &swap_free_completed, 191 "Number of deferred frees completed"); 192 193 /* bits from overcommit */ 194 #define SWAP_RESERVE_FORCE_ON (1 << 0) 195 #define SWAP_RESERVE_RLIMIT_ON (1 << 1) 196 #define SWAP_RESERVE_ALLOW_NONWIRED (1 << 2) 197 198 static int 199 sysctl_page_shift(SYSCTL_HANDLER_ARGS) 200 { 201 uint64_t newval; 202 u_long value = *(u_long *)arg1; 203 204 newval = ((uint64_t)value) << PAGE_SHIFT; 205 return (sysctl_handle_64(oidp, &newval, 0, req)); 206 } 207 208 static bool 209 swap_reserve_by_cred_rlimit(u_long pincr, struct ucred *cred, int oc) 210 { 211 struct uidinfo *uip; 212 u_long prev; 213 214 uip = cred->cr_ruidinfo; 215 216 prev = atomic_fetchadd_long(&uip->ui_vmsize, pincr); 217 if ((oc & SWAP_RESERVE_RLIMIT_ON) != 0 && 218 prev + pincr > lim_cur(curthread, RLIMIT_SWAP) && 219 priv_check(curthread, PRIV_VM_SWAP_NORLIMIT) != 0) { 220 prev = atomic_fetchadd_long(&uip->ui_vmsize, -pincr); 221 KASSERT(prev >= pincr, ("negative vmsize for uid = %d\n", uip->ui_uid)); 222 return (false); 223 } 224 return (true); 225 } 226 227 static void 228 swap_release_by_cred_rlimit(u_long pdecr, struct ucred *cred) 229 { 230 struct uidinfo *uip; 231 #ifdef INVARIANTS 232 u_long prev; 233 #endif 234 235 uip = cred->cr_ruidinfo; 236 237 #ifdef INVARIANTS 238 prev = atomic_fetchadd_long(&uip->ui_vmsize, -pdecr); 239 KASSERT(prev >= pdecr, ("negative vmsize for uid = %d\n", uip->ui_uid)); 240 #else 241 atomic_subtract_long(&uip->ui_vmsize, pdecr); 242 #endif 243 } 244 245 static void 246 swap_reserve_force_rlimit(u_long pincr, struct ucred *cred) 247 { 248 struct uidinfo *uip; 249 250 uip = cred->cr_ruidinfo; 251 atomic_add_long(&uip->ui_vmsize, pincr); 252 } 253 254 bool 255 swap_reserve(vm_ooffset_t incr) 256 { 257 258 return (swap_reserve_by_cred(incr, curthread->td_ucred)); 259 } 260 261 bool 262 swap_reserve_by_cred(vm_ooffset_t incr, struct ucred *cred) 263 { 264 u_long r, s, prev, pincr; 265 #ifdef RACCT 266 int error; 267 #endif 268 int oc; 269 static int curfail; 270 static struct timeval lastfail; 271 272 KASSERT((incr & PAGE_MASK) == 0, ("%s: incr: %ju & PAGE_MASK", __func__, 273 (uintmax_t)incr)); 274 275 #ifdef RACCT 276 if (RACCT_ENABLED()) { 277 PROC_LOCK(curproc); 278 error = racct_add(curproc, RACCT_SWAP, incr); 279 PROC_UNLOCK(curproc); 280 if (error != 0) 281 return (false); 282 } 283 #endif 284 285 pincr = atop(incr); 286 prev = atomic_fetchadd_long(&swap_reserved, pincr); 287 r = prev + pincr; 288 s = swap_total; 289 oc = atomic_load_int(&overcommit); 290 if (r > s && (oc & SWAP_RESERVE_ALLOW_NONWIRED) != 0) { 291 s += vm_cnt.v_page_count - vm_cnt.v_free_reserved - 292 vm_wire_count(); 293 } 294 if ((oc & SWAP_RESERVE_FORCE_ON) != 0 && r > s && 295 priv_check(curthread, PRIV_VM_SWAP_NOQUOTA) != 0) { 296 prev = atomic_fetchadd_long(&swap_reserved, -pincr); 297 KASSERT(prev >= pincr, ("swap_reserved < incr on overcommit fail")); 298 goto out_error; 299 } 300 301 if (!swap_reserve_by_cred_rlimit(pincr, cred, oc)) { 302 prev = atomic_fetchadd_long(&swap_reserved, -pincr); 303 KASSERT(prev >= pincr, ("swap_reserved < incr on overcommit fail")); 304 goto out_error; 305 } 306 307 return (true); 308 309 out_error: 310 if (ppsratecheck(&lastfail, &curfail, 1)) { 311 printf("uid %d, pid %d: swap reservation for %jd bytes failed\n", 312 cred->cr_ruidinfo->ui_uid, curproc->p_pid, incr); 313 } 314 #ifdef RACCT 315 if (RACCT_ENABLED()) { 316 PROC_LOCK(curproc); 317 racct_sub(curproc, RACCT_SWAP, incr); 318 PROC_UNLOCK(curproc); 319 } 320 #endif 321 322 return (false); 323 } 324 325 void 326 swap_reserve_force(vm_ooffset_t incr) 327 { 328 u_long pincr; 329 330 KASSERT((incr & PAGE_MASK) == 0, ("%s: incr: %ju & PAGE_MASK", __func__, 331 (uintmax_t)incr)); 332 333 #ifdef RACCT 334 if (RACCT_ENABLED()) { 335 PROC_LOCK(curproc); 336 racct_add_force(curproc, RACCT_SWAP, incr); 337 PROC_UNLOCK(curproc); 338 } 339 #endif 340 pincr = atop(incr); 341 atomic_add_long(&swap_reserved, pincr); 342 swap_reserve_force_rlimit(pincr, curthread->td_ucred); 343 } 344 345 void 346 swap_release(vm_ooffset_t decr) 347 { 348 struct ucred *cred; 349 350 PROC_LOCK(curproc); 351 cred = curproc->p_ucred; 352 swap_release_by_cred(decr, cred); 353 PROC_UNLOCK(curproc); 354 } 355 356 void 357 swap_release_by_cred(vm_ooffset_t decr, struct ucred *cred) 358 { 359 u_long pdecr; 360 #ifdef INVARIANTS 361 u_long prev; 362 #endif 363 364 KASSERT((decr & PAGE_MASK) == 0, ("%s: decr: %ju & PAGE_MASK", __func__, 365 (uintmax_t)decr)); 366 367 pdecr = atop(decr); 368 #ifdef INVARIANTS 369 prev = atomic_fetchadd_long(&swap_reserved, -pdecr); 370 KASSERT(prev >= pdecr, ("swap_reserved < decr")); 371 #else 372 atomic_subtract_long(&swap_reserved, pdecr); 373 #endif 374 375 swap_release_by_cred_rlimit(pdecr, cred); 376 #ifdef RACCT 377 if (racct_enable) 378 racct_sub_cred(cred, RACCT_SWAP, decr); 379 #endif 380 } 381 382 static int swap_pager_full = 2; /* swap space exhaustion (task killing) */ 383 static int swap_pager_almost_full = 1; /* swap space exhaustion (w/hysteresis)*/ 384 static struct mtx swbuf_mtx; /* to sync nsw_wcount_async */ 385 static int nsw_wcount_async; /* limit async write buffers */ 386 static int nsw_wcount_async_max;/* assigned maximum */ 387 int nsw_cluster_max; /* maximum VOP I/O allowed */ 388 389 static int sysctl_swap_async_max(SYSCTL_HANDLER_ARGS); 390 SYSCTL_PROC(_vm, OID_AUTO, swap_async_max, CTLTYPE_INT | CTLFLAG_RW | 391 CTLFLAG_MPSAFE, NULL, 0, sysctl_swap_async_max, "I", 392 "Maximum running async swap ops"); 393 static int sysctl_swap_fragmentation(SYSCTL_HANDLER_ARGS); 394 SYSCTL_PROC(_vm, OID_AUTO, swap_fragmentation, CTLTYPE_STRING | CTLFLAG_RD | 395 CTLFLAG_MPSAFE, NULL, 0, sysctl_swap_fragmentation, "A", 396 "Swap Fragmentation Info"); 397 398 static struct sx sw_alloc_sx; 399 400 /* 401 * "named" and "unnamed" anon region objects. Try to reduce the overhead 402 * of searching a named list by hashing it just a little. 403 */ 404 405 #define NOBJLISTS 8 406 407 #define NOBJLIST(handle) \ 408 (&swap_pager_object_list[((int)(intptr_t)handle >> 4) & (NOBJLISTS-1)]) 409 410 static struct pagerlst swap_pager_object_list[NOBJLISTS]; 411 static uma_zone_t swwbuf_zone; 412 static uma_zone_t swrbuf_zone; 413 static uma_zone_t swblk_zone; 414 static uma_zone_t swpctrie_zone; 415 416 /* 417 * pagerops for OBJT_SWAP - "swap pager". Some ops are also global procedure 418 * calls hooked from other parts of the VM system and do not appear here. 419 * (see vm/swap_pager.h). 420 */ 421 static vm_object_t 422 swap_pager_alloc(void *handle, vm_ooffset_t size, 423 vm_prot_t prot, vm_ooffset_t offset, struct ucred *); 424 static void swap_pager_dealloc(vm_object_t object); 425 static int swap_pager_getpages(vm_object_t, vm_page_t *, int, int *, 426 int *); 427 static int swap_pager_getpages_async(vm_object_t, vm_page_t *, int, int *, 428 int *, pgo_getpages_iodone_t, void *); 429 static void swap_pager_putpages(vm_object_t, vm_page_t *, int, boolean_t, int *); 430 static boolean_t 431 swap_pager_haspage(vm_object_t object, vm_pindex_t pindex, int *before, int *after); 432 static void swap_pager_init(void); 433 static void swap_pager_unswapped(vm_page_t); 434 static void swap_pager_swapoff(struct swdevt *sp); 435 static void swap_pager_update_writecount(vm_object_t object, 436 vm_offset_t start, vm_offset_t end); 437 static void swap_pager_release_writecount(vm_object_t object, 438 vm_offset_t start, vm_offset_t end); 439 static void swap_pager_freespace(vm_object_t object, vm_pindex_t start, 440 vm_size_t size); 441 442 const struct pagerops swappagerops = { 443 .pgo_kvme_type = KVME_TYPE_SWAP, 444 .pgo_init = swap_pager_init, /* early system initialization of pager */ 445 .pgo_alloc = swap_pager_alloc, /* allocate an OBJT_SWAP object */ 446 .pgo_dealloc = swap_pager_dealloc, /* deallocate an OBJT_SWAP object */ 447 .pgo_getpages = swap_pager_getpages, /* pagein */ 448 .pgo_getpages_async = swap_pager_getpages_async, /* pagein (async) */ 449 .pgo_putpages = swap_pager_putpages, /* pageout */ 450 .pgo_haspage = swap_pager_haspage, /* get backing store status for page */ 451 .pgo_pageunswapped = swap_pager_unswapped, /* remove swap related to page */ 452 .pgo_update_writecount = swap_pager_update_writecount, 453 .pgo_release_writecount = swap_pager_release_writecount, 454 .pgo_freespace = swap_pager_freespace, 455 }; 456 457 /* 458 * swap_*() routines are externally accessible. swp_*() routines are 459 * internal. 460 */ 461 static int nswap_lowat = 128; /* in pages, swap_pager_almost_full warn */ 462 static int nswap_hiwat = 512; /* in pages, swap_pager_almost_full warn */ 463 464 SYSCTL_INT(_vm, OID_AUTO, dmmax, CTLFLAG_RD, &nsw_cluster_max, 0, 465 "Maximum size of a swap block in pages"); 466 467 static void swp_sizecheck(void); 468 static void swp_pager_async_iodone(struct buf *bp); 469 static bool swp_pager_swblk_empty(struct swblk *sb, int start, int limit); 470 static void swp_pager_free_empty_swblk(vm_object_t, struct swblk *sb); 471 static int swapongeom(struct vnode *); 472 static int swaponvp(struct thread *, struct vnode *, u_long); 473 static int swapoff_one(struct swdevt *sp, struct ucred *cred, 474 u_int flags); 475 476 /* 477 * Swap bitmap functions 478 */ 479 static void swp_pager_freeswapspace(daddr_t blk, daddr_t npages); 480 static daddr_t swp_pager_getswapspace(int *npages); 481 482 /* 483 * Metadata functions 484 */ 485 static daddr_t swp_pager_meta_build(vm_object_t, vm_pindex_t, daddr_t); 486 static void swp_pager_meta_free(vm_object_t, vm_pindex_t, vm_pindex_t); 487 static void swp_pager_meta_transfer(vm_object_t src, vm_object_t dst, 488 vm_pindex_t pindex, vm_pindex_t count); 489 static void swp_pager_meta_free_all(vm_object_t); 490 static daddr_t swp_pager_meta_lookup(vm_object_t, vm_pindex_t); 491 492 static void 493 swp_pager_init_freerange(daddr_t *start, daddr_t *num) 494 { 495 496 *start = SWAPBLK_NONE; 497 *num = 0; 498 } 499 500 static void 501 swp_pager_update_freerange(daddr_t *start, daddr_t *num, daddr_t addr) 502 { 503 504 if (*start + *num == addr) { 505 (*num)++; 506 } else { 507 swp_pager_freeswapspace(*start, *num); 508 *start = addr; 509 *num = 1; 510 } 511 } 512 513 static void * 514 swblk_trie_alloc(struct pctrie *ptree) 515 { 516 517 return (uma_zalloc(swpctrie_zone, M_NOWAIT | (curproc == pageproc ? 518 M_USE_RESERVE : 0))); 519 } 520 521 static void 522 swblk_trie_free(struct pctrie *ptree, void *node) 523 { 524 525 uma_zfree(swpctrie_zone, node); 526 } 527 528 PCTRIE_DEFINE(SWAP, swblk, p, swblk_trie_alloc, swblk_trie_free); 529 530 /* 531 * SWP_SIZECHECK() - update swap_pager_full indication 532 * 533 * update the swap_pager_almost_full indication and warn when we are 534 * about to run out of swap space, using lowat/hiwat hysteresis. 535 * 536 * Clear swap_pager_full ( task killing ) indication when lowat is met. 537 * 538 * No restrictions on call 539 * This routine may not block. 540 */ 541 static void 542 swp_sizecheck(void) 543 { 544 545 if (swap_pager_avail < nswap_lowat) { 546 if (swap_pager_almost_full == 0) { 547 printf("swap_pager: out of swap space\n"); 548 swap_pager_almost_full = 1; 549 } 550 } else { 551 swap_pager_full = 0; 552 if (swap_pager_avail > nswap_hiwat) 553 swap_pager_almost_full = 0; 554 } 555 } 556 557 /* 558 * SWAP_PAGER_INIT() - initialize the swap pager! 559 * 560 * Expected to be started from system init. NOTE: This code is run 561 * before much else so be careful what you depend on. Most of the VM 562 * system has yet to be initialized at this point. 563 */ 564 static void 565 swap_pager_init(void) 566 { 567 /* 568 * Initialize object lists 569 */ 570 int i; 571 572 for (i = 0; i < NOBJLISTS; ++i) 573 TAILQ_INIT(&swap_pager_object_list[i]); 574 mtx_init(&sw_dev_mtx, "swapdev", NULL, MTX_DEF); 575 sx_init(&sw_alloc_sx, "swspsx"); 576 sx_init(&swdev_syscall_lock, "swsysc"); 577 578 /* 579 * The nsw_cluster_max is constrained by the bp->b_pages[] 580 * array, which has maxphys / PAGE_SIZE entries, and our locally 581 * defined MAX_PAGEOUT_CLUSTER. Also be aware that swap ops are 582 * constrained by the swap device interleave stripe size. 583 * 584 * Initialized early so that GEOM_ELI can see it. 585 */ 586 nsw_cluster_max = min(maxphys / PAGE_SIZE, MAX_PAGEOUT_CLUSTER); 587 } 588 589 /* 590 * SWAP_PAGER_SWAP_INIT() - swap pager initialization from pageout process 591 * 592 * Expected to be started from pageout process once, prior to entering 593 * its main loop. 594 */ 595 void 596 swap_pager_swap_init(void) 597 { 598 unsigned long n, n2; 599 600 /* 601 * Number of in-transit swap bp operations. Don't 602 * exhaust the pbufs completely. Make sure we 603 * initialize workable values (0 will work for hysteresis 604 * but it isn't very efficient). 605 * 606 * Currently we hardwire nsw_wcount_async to 4. This limit is 607 * designed to prevent other I/O from having high latencies due to 608 * our pageout I/O. The value 4 works well for one or two active swap 609 * devices but is probably a little low if you have more. Even so, 610 * a higher value would probably generate only a limited improvement 611 * with three or four active swap devices since the system does not 612 * typically have to pageout at extreme bandwidths. We will want 613 * at least 2 per swap devices, and 4 is a pretty good value if you 614 * have one NFS swap device due to the command/ack latency over NFS. 615 * So it all works out pretty well. 616 * 617 * nsw_cluster_max is initialized in swap_pager_init(). 618 */ 619 620 nsw_wcount_async = 4; 621 nsw_wcount_async_max = nsw_wcount_async; 622 mtx_init(&swbuf_mtx, "async swbuf mutex", NULL, MTX_DEF); 623 624 swwbuf_zone = pbuf_zsecond_create("swwbuf", nswbuf / 4); 625 swrbuf_zone = pbuf_zsecond_create("swrbuf", nswbuf / 2); 626 627 /* 628 * Initialize our zone, taking the user's requested size or 629 * estimating the number we need based on the number of pages 630 * in the system. 631 */ 632 n = maxswzone != 0 ? maxswzone / sizeof(struct swblk) : 633 vm_cnt.v_page_count / 2; 634 swpctrie_zone = uma_zcreate("swpctrie", pctrie_node_size(), NULL, NULL, 635 pctrie_zone_init, NULL, UMA_ALIGN_PTR, 0); 636 if (swpctrie_zone == NULL) 637 panic("failed to create swap pctrie zone."); 638 swblk_zone = uma_zcreate("swblk", sizeof(struct swblk), NULL, NULL, 639 NULL, NULL, _Alignof(struct swblk) - 1, 0); 640 if (swblk_zone == NULL) 641 panic("failed to create swap blk zone."); 642 n2 = n; 643 do { 644 if (uma_zone_reserve_kva(swblk_zone, n)) 645 break; 646 /* 647 * if the allocation failed, try a zone two thirds the 648 * size of the previous attempt. 649 */ 650 n -= ((n + 2) / 3); 651 } while (n > 0); 652 653 /* 654 * Often uma_zone_reserve_kva() cannot reserve exactly the 655 * requested size. Account for the difference when 656 * calculating swap_maxpages. 657 */ 658 n = uma_zone_get_max(swblk_zone); 659 660 if (n < n2) 661 printf("Swap blk zone entries changed from %lu to %lu.\n", 662 n2, n); 663 /* absolute maximum we can handle assuming 100% efficiency */ 664 swap_maxpages = n * SWAP_META_PAGES; 665 swzone = n * sizeof(struct swblk); 666 if (!uma_zone_reserve_kva(swpctrie_zone, n)) 667 printf("Cannot reserve swap pctrie zone, " 668 "reduce kern.maxswzone.\n"); 669 } 670 671 bool 672 swap_pager_init_object(vm_object_t object, void *handle, struct ucred *cred, 673 vm_ooffset_t size, vm_ooffset_t offset) 674 { 675 if (cred != NULL) { 676 if (!swap_reserve_by_cred(size, cred)) 677 return (false); 678 crhold(cred); 679 } 680 681 object->un_pager.swp.writemappings = 0; 682 object->handle = handle; 683 if (cred != NULL) { 684 object->cred = cred; 685 object->charge = size; 686 } 687 return (true); 688 } 689 690 static vm_object_t 691 swap_pager_alloc_init(objtype_t otype, void *handle, struct ucred *cred, 692 vm_ooffset_t size, vm_ooffset_t offset) 693 { 694 vm_object_t object; 695 696 /* 697 * The un_pager.swp.swp_blks trie is initialized by 698 * vm_object_allocate() to ensure the correct order of 699 * visibility to other threads. 700 */ 701 object = vm_object_allocate(otype, OFF_TO_IDX(offset + 702 PAGE_MASK + size)); 703 704 if (!swap_pager_init_object(object, handle, cred, size, offset)) { 705 vm_object_deallocate(object); 706 return (NULL); 707 } 708 return (object); 709 } 710 711 /* 712 * SWAP_PAGER_ALLOC() - allocate a new OBJT_SWAP VM object and instantiate 713 * its metadata structures. 714 * 715 * This routine is called from the mmap and fork code to create a new 716 * OBJT_SWAP object. 717 * 718 * This routine must ensure that no live duplicate is created for 719 * the named object request, which is protected against by 720 * holding the sw_alloc_sx lock in case handle != NULL. 721 */ 722 static vm_object_t 723 swap_pager_alloc(void *handle, vm_ooffset_t size, vm_prot_t prot, 724 vm_ooffset_t offset, struct ucred *cred) 725 { 726 vm_object_t object; 727 728 if (handle != NULL) { 729 /* 730 * Reference existing named region or allocate new one. There 731 * should not be a race here against swp_pager_meta_build() 732 * as called from vm_page_remove() in regards to the lookup 733 * of the handle. 734 */ 735 sx_xlock(&sw_alloc_sx); 736 object = vm_pager_object_lookup(NOBJLIST(handle), handle); 737 if (object == NULL) { 738 object = swap_pager_alloc_init(OBJT_SWAP, handle, cred, 739 size, offset); 740 if (object != NULL) { 741 TAILQ_INSERT_TAIL(NOBJLIST(object->handle), 742 object, pager_object_list); 743 } 744 } 745 sx_xunlock(&sw_alloc_sx); 746 } else { 747 object = swap_pager_alloc_init(OBJT_SWAP, handle, cred, 748 size, offset); 749 } 750 return (object); 751 } 752 753 /* 754 * SWAP_PAGER_DEALLOC() - remove swap metadata from object 755 * 756 * The swap backing for the object is destroyed. The code is 757 * designed such that we can reinstantiate it later, but this 758 * routine is typically called only when the entire object is 759 * about to be destroyed. 760 * 761 * The object must be locked. 762 */ 763 static void 764 swap_pager_dealloc(vm_object_t object) 765 { 766 767 VM_OBJECT_ASSERT_WLOCKED(object); 768 KASSERT((object->flags & OBJ_DEAD) != 0, ("dealloc of reachable obj")); 769 770 /* 771 * Remove from list right away so lookups will fail if we block for 772 * pageout completion. 773 */ 774 if ((object->flags & OBJ_ANON) == 0 && object->handle != NULL) { 775 VM_OBJECT_WUNLOCK(object); 776 sx_xlock(&sw_alloc_sx); 777 TAILQ_REMOVE(NOBJLIST(object->handle), object, 778 pager_object_list); 779 sx_xunlock(&sw_alloc_sx); 780 VM_OBJECT_WLOCK(object); 781 } 782 783 vm_object_pip_wait(object, "swpdea"); 784 785 /* 786 * Free all remaining metadata. We only bother to free it from 787 * the swap meta data. We do not attempt to free swapblk's still 788 * associated with vm_page_t's for this object. We do not care 789 * if paging is still in progress on some objects. 790 */ 791 swp_pager_meta_free_all(object); 792 object->handle = NULL; 793 object->type = OBJT_DEAD; 794 vm_object_clear_flag(object, OBJ_SWAP); 795 } 796 797 /************************************************************************ 798 * SWAP PAGER BITMAP ROUTINES * 799 ************************************************************************/ 800 801 /* 802 * SWP_PAGER_GETSWAPSPACE() - allocate raw swap space 803 * 804 * Allocate swap for up to the requested number of pages. The 805 * starting swap block number (a page index) is returned or 806 * SWAPBLK_NONE if the allocation failed. 807 * 808 * Also has the side effect of advising that somebody made a mistake 809 * when they configured swap and didn't configure enough. 810 * 811 * This routine may not sleep. 812 * 813 * We allocate in round-robin fashion from the configured devices. 814 */ 815 static daddr_t 816 swp_pager_getswapspace(int *io_npages) 817 { 818 daddr_t blk; 819 struct swdevt *sp; 820 int mpages, npages; 821 822 KASSERT(*io_npages >= 1, 823 ("%s: npages not positive", __func__)); 824 blk = SWAPBLK_NONE; 825 mpages = *io_npages; 826 npages = imin(BLIST_MAX_ALLOC, mpages); 827 mtx_lock(&sw_dev_mtx); 828 sp = swdevhd; 829 while (!TAILQ_EMPTY(&swtailq)) { 830 if (sp == NULL) 831 sp = TAILQ_FIRST(&swtailq); 832 if ((sp->sw_flags & SW_CLOSING) == 0) 833 blk = blist_alloc(sp->sw_blist, &npages, mpages); 834 if (blk != SWAPBLK_NONE) 835 break; 836 sp = TAILQ_NEXT(sp, sw_list); 837 if (swdevhd == sp) { 838 if (npages == 1) 839 break; 840 mpages = npages - 1; 841 npages >>= 1; 842 } 843 } 844 if (blk != SWAPBLK_NONE) { 845 *io_npages = npages; 846 blk += sp->sw_first; 847 sp->sw_used += npages; 848 swap_pager_avail -= npages; 849 swp_sizecheck(); 850 swdevhd = TAILQ_NEXT(sp, sw_list); 851 } else { 852 if (swap_pager_full != 2) { 853 printf("swp_pager_getswapspace(%d): failed\n", 854 *io_npages); 855 swap_pager_full = 2; 856 swap_pager_almost_full = 1; 857 } 858 swdevhd = NULL; 859 } 860 mtx_unlock(&sw_dev_mtx); 861 return (blk); 862 } 863 864 static bool 865 swp_pager_isondev(daddr_t blk, struct swdevt *sp) 866 { 867 868 return (blk >= sp->sw_first && blk < sp->sw_end); 869 } 870 871 static void 872 swp_pager_strategy(struct buf *bp) 873 { 874 struct swdevt *sp; 875 876 mtx_lock(&sw_dev_mtx); 877 TAILQ_FOREACH(sp, &swtailq, sw_list) { 878 if (swp_pager_isondev(bp->b_blkno, sp)) { 879 mtx_unlock(&sw_dev_mtx); 880 if ((sp->sw_flags & SW_UNMAPPED) != 0 && 881 unmapped_buf_allowed) { 882 bp->b_data = unmapped_buf; 883 bp->b_offset = 0; 884 } else { 885 pmap_qenter((vm_offset_t)bp->b_data, 886 &bp->b_pages[0], bp->b_bcount / PAGE_SIZE); 887 } 888 sp->sw_strategy(bp, sp); 889 return; 890 } 891 } 892 panic("Swapdev not found"); 893 } 894 895 /* 896 * SWP_PAGER_FREESWAPSPACE() - free raw swap space 897 * 898 * This routine returns the specified swap blocks back to the bitmap. 899 * 900 * This routine may not sleep. 901 */ 902 static void 903 swp_pager_freeswapspace(daddr_t blk, daddr_t npages) 904 { 905 struct swdevt *sp; 906 907 if (npages == 0) 908 return; 909 mtx_lock(&sw_dev_mtx); 910 TAILQ_FOREACH(sp, &swtailq, sw_list) { 911 if (swp_pager_isondev(blk, sp)) { 912 sp->sw_used -= npages; 913 /* 914 * If we are attempting to stop swapping on 915 * this device, we don't want to mark any 916 * blocks free lest they be reused. 917 */ 918 if ((sp->sw_flags & SW_CLOSING) == 0) { 919 blist_free(sp->sw_blist, blk - sp->sw_first, 920 npages); 921 swap_pager_avail += npages; 922 swp_sizecheck(); 923 } 924 mtx_unlock(&sw_dev_mtx); 925 return; 926 } 927 } 928 panic("Swapdev not found"); 929 } 930 931 /* 932 * SYSCTL_SWAP_FRAGMENTATION() - produce raw swap space stats 933 */ 934 static int 935 sysctl_swap_fragmentation(SYSCTL_HANDLER_ARGS) 936 { 937 struct sbuf sbuf; 938 struct swdevt *sp; 939 const char *devname; 940 int error; 941 942 error = sysctl_wire_old_buffer(req, 0); 943 if (error != 0) 944 return (error); 945 sbuf_new_for_sysctl(&sbuf, NULL, 128, req); 946 mtx_lock(&sw_dev_mtx); 947 TAILQ_FOREACH(sp, &swtailq, sw_list) { 948 if (vn_isdisk(sp->sw_vp)) 949 devname = devtoname(sp->sw_vp->v_rdev); 950 else 951 devname = "[file]"; 952 sbuf_printf(&sbuf, "\nFree space on device %s:\n", devname); 953 blist_stats(sp->sw_blist, &sbuf); 954 } 955 mtx_unlock(&sw_dev_mtx); 956 error = sbuf_finish(&sbuf); 957 sbuf_delete(&sbuf); 958 return (error); 959 } 960 961 /* 962 * SWAP_PAGER_FREESPACE() - frees swap blocks associated with a page 963 * range within an object. 964 * 965 * This routine removes swapblk assignments from swap metadata. 966 * 967 * The external callers of this routine typically have already destroyed 968 * or renamed vm_page_t's associated with this range in the object so 969 * we should be ok. 970 * 971 * The object must be locked. 972 */ 973 static void 974 swap_pager_freespace(vm_object_t object, vm_pindex_t start, vm_size_t size) 975 { 976 977 swp_pager_meta_free(object, start, size); 978 } 979 980 /* 981 * SWAP_PAGER_RESERVE() - reserve swap blocks in object 982 * 983 * Assigns swap blocks to the specified range within the object. The 984 * swap blocks are not zeroed. Any previous swap assignment is destroyed. 985 * 986 * Returns 0 on success, -1 on failure. 987 */ 988 int 989 swap_pager_reserve(vm_object_t object, vm_pindex_t start, vm_pindex_t size) 990 { 991 daddr_t addr, blk, n_free, s_free; 992 vm_pindex_t i, j; 993 int n; 994 995 swp_pager_init_freerange(&s_free, &n_free); 996 VM_OBJECT_WLOCK(object); 997 for (i = 0; i < size; i += n) { 998 n = MIN(size - i, INT_MAX); 999 blk = swp_pager_getswapspace(&n); 1000 if (blk == SWAPBLK_NONE) { 1001 swp_pager_meta_free(object, start, i); 1002 VM_OBJECT_WUNLOCK(object); 1003 return (-1); 1004 } 1005 for (j = 0; j < n; ++j) { 1006 addr = swp_pager_meta_build(object, 1007 start + i + j, blk + j); 1008 if (addr != SWAPBLK_NONE) 1009 swp_pager_update_freerange(&s_free, &n_free, 1010 addr); 1011 } 1012 } 1013 swp_pager_freeswapspace(s_free, n_free); 1014 VM_OBJECT_WUNLOCK(object); 1015 return (0); 1016 } 1017 1018 static bool 1019 swp_pager_xfer_source(vm_object_t srcobject, vm_object_t dstobject, 1020 vm_pindex_t pindex, daddr_t addr) 1021 { 1022 daddr_t dstaddr; 1023 1024 KASSERT((srcobject->flags & OBJ_SWAP) != 0, 1025 ("%s: Srcobject not swappable", __func__)); 1026 if ((dstobject->flags & OBJ_SWAP) != 0 && 1027 swp_pager_meta_lookup(dstobject, pindex) != SWAPBLK_NONE) { 1028 /* Caller should destroy the source block. */ 1029 return (false); 1030 } 1031 1032 /* 1033 * Destination has no swapblk and is not resident, transfer source. 1034 * swp_pager_meta_build() can sleep. 1035 */ 1036 VM_OBJECT_WUNLOCK(srcobject); 1037 dstaddr = swp_pager_meta_build(dstobject, pindex, addr); 1038 KASSERT(dstaddr == SWAPBLK_NONE, 1039 ("Unexpected destination swapblk")); 1040 VM_OBJECT_WLOCK(srcobject); 1041 1042 return (true); 1043 } 1044 1045 /* 1046 * SWAP_PAGER_COPY() - copy blocks from source pager to destination pager 1047 * and destroy the source. 1048 * 1049 * Copy any valid swapblks from the source to the destination. In 1050 * cases where both the source and destination have a valid swapblk, 1051 * we keep the destination's. 1052 * 1053 * This routine is allowed to sleep. It may sleep allocating metadata 1054 * indirectly through swp_pager_meta_build(). 1055 * 1056 * The source object contains no vm_page_t's (which is just as well) 1057 * 1058 * The source object is of type OBJT_SWAP. 1059 * 1060 * The source and destination objects must be locked. 1061 * Both object locks may temporarily be released. 1062 */ 1063 void 1064 swap_pager_copy(vm_object_t srcobject, vm_object_t dstobject, 1065 vm_pindex_t offset, int destroysource) 1066 { 1067 1068 VM_OBJECT_ASSERT_WLOCKED(srcobject); 1069 VM_OBJECT_ASSERT_WLOCKED(dstobject); 1070 1071 /* 1072 * If destroysource is set, we remove the source object from the 1073 * swap_pager internal queue now. 1074 */ 1075 if (destroysource && (srcobject->flags & OBJ_ANON) == 0 && 1076 srcobject->handle != NULL) { 1077 VM_OBJECT_WUNLOCK(srcobject); 1078 VM_OBJECT_WUNLOCK(dstobject); 1079 sx_xlock(&sw_alloc_sx); 1080 TAILQ_REMOVE(NOBJLIST(srcobject->handle), srcobject, 1081 pager_object_list); 1082 sx_xunlock(&sw_alloc_sx); 1083 VM_OBJECT_WLOCK(dstobject); 1084 VM_OBJECT_WLOCK(srcobject); 1085 } 1086 1087 /* 1088 * Transfer source to destination. 1089 */ 1090 swp_pager_meta_transfer(srcobject, dstobject, offset, dstobject->size); 1091 1092 /* 1093 * Free left over swap blocks in source. 1094 * 1095 * We have to revert the type to OBJT_DEFAULT so we do not accidentally 1096 * double-remove the object from the swap queues. 1097 */ 1098 if (destroysource) { 1099 swp_pager_meta_free_all(srcobject); 1100 /* 1101 * Reverting the type is not necessary, the caller is going 1102 * to destroy srcobject directly, but I'm doing it here 1103 * for consistency since we've removed the object from its 1104 * queues. 1105 */ 1106 srcobject->type = OBJT_DEFAULT; 1107 vm_object_clear_flag(srcobject, OBJ_SWAP); 1108 } 1109 } 1110 1111 /* 1112 * SWAP_PAGER_HASPAGE() - determine if we have good backing store for 1113 * the requested page. 1114 * 1115 * We determine whether good backing store exists for the requested 1116 * page and return TRUE if it does, FALSE if it doesn't. 1117 * 1118 * If TRUE, we also try to determine how much valid, contiguous backing 1119 * store exists before and after the requested page. 1120 */ 1121 static boolean_t 1122 swap_pager_haspage(vm_object_t object, vm_pindex_t pindex, int *before, 1123 int *after) 1124 { 1125 daddr_t blk, blk0; 1126 int i; 1127 1128 VM_OBJECT_ASSERT_LOCKED(object); 1129 KASSERT((object->flags & OBJ_SWAP) != 0, 1130 ("%s: object not swappable", __func__)); 1131 1132 /* 1133 * do we have good backing store at the requested index ? 1134 */ 1135 blk0 = swp_pager_meta_lookup(object, pindex); 1136 if (blk0 == SWAPBLK_NONE) { 1137 if (before) 1138 *before = 0; 1139 if (after) 1140 *after = 0; 1141 return (FALSE); 1142 } 1143 1144 /* 1145 * find backwards-looking contiguous good backing store 1146 */ 1147 if (before != NULL) { 1148 for (i = 1; i < SWB_NPAGES; i++) { 1149 if (i > pindex) 1150 break; 1151 blk = swp_pager_meta_lookup(object, pindex - i); 1152 if (blk != blk0 - i) 1153 break; 1154 } 1155 *before = i - 1; 1156 } 1157 1158 /* 1159 * find forward-looking contiguous good backing store 1160 */ 1161 if (after != NULL) { 1162 for (i = 1; i < SWB_NPAGES; i++) { 1163 blk = swp_pager_meta_lookup(object, pindex + i); 1164 if (blk != blk0 + i) 1165 break; 1166 } 1167 *after = i - 1; 1168 } 1169 return (TRUE); 1170 } 1171 1172 /* 1173 * SWAP_PAGER_PAGE_UNSWAPPED() - remove swap backing store related to page 1174 * 1175 * This removes any associated swap backing store, whether valid or 1176 * not, from the page. 1177 * 1178 * This routine is typically called when a page is made dirty, at 1179 * which point any associated swap can be freed. MADV_FREE also 1180 * calls us in a special-case situation 1181 * 1182 * NOTE!!! If the page is clean and the swap was valid, the caller 1183 * should make the page dirty before calling this routine. This routine 1184 * does NOT change the m->dirty status of the page. Also: MADV_FREE 1185 * depends on it. 1186 * 1187 * This routine may not sleep. 1188 * 1189 * The object containing the page may be locked. 1190 */ 1191 static void 1192 swap_pager_unswapped(vm_page_t m) 1193 { 1194 struct swblk *sb; 1195 vm_object_t obj; 1196 1197 /* 1198 * Handle enqueing deferred frees first. If we do not have the 1199 * object lock we wait for the page daemon to clear the space. 1200 */ 1201 obj = m->object; 1202 if (!VM_OBJECT_WOWNED(obj)) { 1203 VM_PAGE_OBJECT_BUSY_ASSERT(m); 1204 /* 1205 * The caller is responsible for synchronization but we 1206 * will harmlessly handle races. This is typically provided 1207 * by only calling unswapped() when a page transitions from 1208 * clean to dirty. 1209 */ 1210 if ((m->a.flags & (PGA_SWAP_SPACE | PGA_SWAP_FREE)) == 1211 PGA_SWAP_SPACE) { 1212 vm_page_aflag_set(m, PGA_SWAP_FREE); 1213 counter_u64_add(swap_free_deferred, 1); 1214 } 1215 return; 1216 } 1217 if ((m->a.flags & PGA_SWAP_FREE) != 0) 1218 counter_u64_add(swap_free_completed, 1); 1219 vm_page_aflag_clear(m, PGA_SWAP_FREE | PGA_SWAP_SPACE); 1220 1221 /* 1222 * The meta data only exists if the object is OBJT_SWAP 1223 * and even then might not be allocated yet. 1224 */ 1225 KASSERT((m->object->flags & OBJ_SWAP) != 0, 1226 ("Free object not swappable")); 1227 1228 sb = SWAP_PCTRIE_LOOKUP(&m->object->un_pager.swp.swp_blks, 1229 rounddown(m->pindex, SWAP_META_PAGES)); 1230 if (sb == NULL) 1231 return; 1232 if (sb->d[m->pindex % SWAP_META_PAGES] == SWAPBLK_NONE) 1233 return; 1234 swp_pager_freeswapspace(sb->d[m->pindex % SWAP_META_PAGES], 1); 1235 sb->d[m->pindex % SWAP_META_PAGES] = SWAPBLK_NONE; 1236 swp_pager_free_empty_swblk(m->object, sb); 1237 } 1238 1239 /* 1240 * swap_pager_getpages() - bring pages in from swap 1241 * 1242 * Attempt to page in the pages in array "ma" of length "count". The 1243 * caller may optionally specify that additional pages preceding and 1244 * succeeding the specified range be paged in. The number of such pages 1245 * is returned in the "rbehind" and "rahead" parameters, and they will 1246 * be in the inactive queue upon return. 1247 * 1248 * The pages in "ma" must be busied and will remain busied upon return. 1249 */ 1250 static int 1251 swap_pager_getpages_locked(vm_object_t object, vm_page_t *ma, int count, 1252 int *rbehind, int *rahead) 1253 { 1254 struct buf *bp; 1255 vm_page_t bm, mpred, msucc, p; 1256 vm_pindex_t pindex; 1257 daddr_t blk; 1258 int i, maxahead, maxbehind, reqcount; 1259 1260 VM_OBJECT_ASSERT_WLOCKED(object); 1261 reqcount = count; 1262 1263 KASSERT((object->flags & OBJ_SWAP) != 0, 1264 ("%s: object not swappable", __func__)); 1265 if (!swap_pager_haspage(object, ma[0]->pindex, &maxbehind, &maxahead)) { 1266 VM_OBJECT_WUNLOCK(object); 1267 return (VM_PAGER_FAIL); 1268 } 1269 1270 KASSERT(reqcount - 1 <= maxahead, 1271 ("page count %d extends beyond swap block", reqcount)); 1272 1273 /* 1274 * Do not transfer any pages other than those that are xbusied 1275 * when running during a split or collapse operation. This 1276 * prevents clustering from re-creating pages which are being 1277 * moved into another object. 1278 */ 1279 if ((object->flags & (OBJ_SPLIT | OBJ_DEAD)) != 0) { 1280 maxahead = reqcount - 1; 1281 maxbehind = 0; 1282 } 1283 1284 /* 1285 * Clip the readahead and readbehind ranges to exclude resident pages. 1286 */ 1287 if (rahead != NULL) { 1288 *rahead = imin(*rahead, maxahead - (reqcount - 1)); 1289 pindex = ma[reqcount - 1]->pindex; 1290 msucc = TAILQ_NEXT(ma[reqcount - 1], listq); 1291 if (msucc != NULL && msucc->pindex - pindex - 1 < *rahead) 1292 *rahead = msucc->pindex - pindex - 1; 1293 } 1294 if (rbehind != NULL) { 1295 *rbehind = imin(*rbehind, maxbehind); 1296 pindex = ma[0]->pindex; 1297 mpred = TAILQ_PREV(ma[0], pglist, listq); 1298 if (mpred != NULL && pindex - mpred->pindex - 1 < *rbehind) 1299 *rbehind = pindex - mpred->pindex - 1; 1300 } 1301 1302 bm = ma[0]; 1303 for (i = 0; i < count; i++) 1304 ma[i]->oflags |= VPO_SWAPINPROG; 1305 1306 /* 1307 * Allocate readahead and readbehind pages. 1308 */ 1309 if (rbehind != NULL) { 1310 for (i = 1; i <= *rbehind; i++) { 1311 p = vm_page_alloc(object, ma[0]->pindex - i, 1312 VM_ALLOC_NORMAL); 1313 if (p == NULL) 1314 break; 1315 p->oflags |= VPO_SWAPINPROG; 1316 bm = p; 1317 } 1318 *rbehind = i - 1; 1319 } 1320 if (rahead != NULL) { 1321 for (i = 0; i < *rahead; i++) { 1322 p = vm_page_alloc(object, 1323 ma[reqcount - 1]->pindex + i + 1, VM_ALLOC_NORMAL); 1324 if (p == NULL) 1325 break; 1326 p->oflags |= VPO_SWAPINPROG; 1327 } 1328 *rahead = i; 1329 } 1330 if (rbehind != NULL) 1331 count += *rbehind; 1332 if (rahead != NULL) 1333 count += *rahead; 1334 1335 vm_object_pip_add(object, count); 1336 1337 pindex = bm->pindex; 1338 blk = swp_pager_meta_lookup(object, pindex); 1339 KASSERT(blk != SWAPBLK_NONE, 1340 ("no swap blocking containing %p(%jx)", object, (uintmax_t)pindex)); 1341 1342 VM_OBJECT_WUNLOCK(object); 1343 bp = uma_zalloc(swrbuf_zone, M_WAITOK); 1344 MPASS((bp->b_flags & B_MAXPHYS) != 0); 1345 /* Pages cannot leave the object while busy. */ 1346 for (i = 0, p = bm; i < count; i++, p = TAILQ_NEXT(p, listq)) { 1347 MPASS(p->pindex == bm->pindex + i); 1348 bp->b_pages[i] = p; 1349 } 1350 1351 bp->b_flags |= B_PAGING; 1352 bp->b_iocmd = BIO_READ; 1353 bp->b_iodone = swp_pager_async_iodone; 1354 bp->b_rcred = crhold(thread0.td_ucred); 1355 bp->b_wcred = crhold(thread0.td_ucred); 1356 bp->b_blkno = blk; 1357 bp->b_bcount = PAGE_SIZE * count; 1358 bp->b_bufsize = PAGE_SIZE * count; 1359 bp->b_npages = count; 1360 bp->b_pgbefore = rbehind != NULL ? *rbehind : 0; 1361 bp->b_pgafter = rahead != NULL ? *rahead : 0; 1362 1363 VM_CNT_INC(v_swapin); 1364 VM_CNT_ADD(v_swappgsin, count); 1365 1366 /* 1367 * perform the I/O. NOTE!!! bp cannot be considered valid after 1368 * this point because we automatically release it on completion. 1369 * Instead, we look at the one page we are interested in which we 1370 * still hold a lock on even through the I/O completion. 1371 * 1372 * The other pages in our ma[] array are also released on completion, 1373 * so we cannot assume they are valid anymore either. 1374 * 1375 * NOTE: b_blkno is destroyed by the call to swapdev_strategy 1376 */ 1377 BUF_KERNPROC(bp); 1378 swp_pager_strategy(bp); 1379 1380 /* 1381 * Wait for the pages we want to complete. VPO_SWAPINPROG is always 1382 * cleared on completion. If an I/O error occurs, SWAPBLK_NONE 1383 * is set in the metadata for each page in the request. 1384 */ 1385 VM_OBJECT_WLOCK(object); 1386 /* This could be implemented more efficiently with aflags */ 1387 while ((ma[0]->oflags & VPO_SWAPINPROG) != 0) { 1388 ma[0]->oflags |= VPO_SWAPSLEEP; 1389 VM_CNT_INC(v_intrans); 1390 if (VM_OBJECT_SLEEP(object, &object->handle, PSWP, 1391 "swread", hz * 20)) { 1392 printf( 1393 "swap_pager: indefinite wait buffer: bufobj: %p, blkno: %jd, size: %ld\n", 1394 bp->b_bufobj, (intmax_t)bp->b_blkno, bp->b_bcount); 1395 } 1396 } 1397 VM_OBJECT_WUNLOCK(object); 1398 1399 /* 1400 * If we had an unrecoverable read error pages will not be valid. 1401 */ 1402 for (i = 0; i < reqcount; i++) 1403 if (ma[i]->valid != VM_PAGE_BITS_ALL) 1404 return (VM_PAGER_ERROR); 1405 1406 return (VM_PAGER_OK); 1407 1408 /* 1409 * A final note: in a low swap situation, we cannot deallocate swap 1410 * and mark a page dirty here because the caller is likely to mark 1411 * the page clean when we return, causing the page to possibly revert 1412 * to all-zero's later. 1413 */ 1414 } 1415 1416 static int 1417 swap_pager_getpages(vm_object_t object, vm_page_t *ma, int count, 1418 int *rbehind, int *rahead) 1419 { 1420 1421 VM_OBJECT_WLOCK(object); 1422 return (swap_pager_getpages_locked(object, ma, count, rbehind, rahead)); 1423 } 1424 1425 /* 1426 * swap_pager_getpages_async(): 1427 * 1428 * Right now this is emulation of asynchronous operation on top of 1429 * swap_pager_getpages(). 1430 */ 1431 static int 1432 swap_pager_getpages_async(vm_object_t object, vm_page_t *ma, int count, 1433 int *rbehind, int *rahead, pgo_getpages_iodone_t iodone, void *arg) 1434 { 1435 int r, error; 1436 1437 r = swap_pager_getpages(object, ma, count, rbehind, rahead); 1438 switch (r) { 1439 case VM_PAGER_OK: 1440 error = 0; 1441 break; 1442 case VM_PAGER_ERROR: 1443 error = EIO; 1444 break; 1445 case VM_PAGER_FAIL: 1446 error = EINVAL; 1447 break; 1448 default: 1449 panic("unhandled swap_pager_getpages() error %d", r); 1450 } 1451 (iodone)(arg, ma, count, error); 1452 1453 return (r); 1454 } 1455 1456 /* 1457 * swap_pager_putpages: 1458 * 1459 * Assign swap (if necessary) and initiate I/O on the specified pages. 1460 * 1461 * We support both OBJT_DEFAULT and OBJT_SWAP objects. DEFAULT objects 1462 * are automatically converted to SWAP objects. 1463 * 1464 * In a low memory situation we may block in VOP_STRATEGY(), but the new 1465 * vm_page reservation system coupled with properly written VFS devices 1466 * should ensure that no low-memory deadlock occurs. This is an area 1467 * which needs work. 1468 * 1469 * The parent has N vm_object_pip_add() references prior to 1470 * calling us and will remove references for rtvals[] that are 1471 * not set to VM_PAGER_PEND. We need to remove the rest on I/O 1472 * completion. 1473 * 1474 * The parent has soft-busy'd the pages it passes us and will unbusy 1475 * those whose rtvals[] entry is not set to VM_PAGER_PEND on return. 1476 * We need to unbusy the rest on I/O completion. 1477 */ 1478 static void 1479 swap_pager_putpages(vm_object_t object, vm_page_t *ma, int count, 1480 int flags, int *rtvals) 1481 { 1482 struct buf *bp; 1483 daddr_t addr, blk, n_free, s_free; 1484 vm_page_t mreq; 1485 int i, j, n; 1486 bool async; 1487 1488 KASSERT(count == 0 || ma[0]->object == object, 1489 ("%s: object mismatch %p/%p", 1490 __func__, object, ma[0]->object)); 1491 1492 /* 1493 * Step 1 1494 * 1495 * Turn object into OBJT_SWAP. Force sync if not a pageout process. 1496 */ 1497 if ((object->flags & OBJ_SWAP) == 0) { 1498 addr = swp_pager_meta_build(object, 0, SWAPBLK_NONE); 1499 KASSERT(addr == SWAPBLK_NONE, 1500 ("unexpected object swap block")); 1501 } 1502 VM_OBJECT_WUNLOCK(object); 1503 async = curproc == pageproc && (flags & VM_PAGER_PUT_SYNC) == 0; 1504 swp_pager_init_freerange(&s_free, &n_free); 1505 1506 /* 1507 * Step 2 1508 * 1509 * Assign swap blocks and issue I/O. We reallocate swap on the fly. 1510 * The page is left dirty until the pageout operation completes 1511 * successfully. 1512 */ 1513 for (i = 0; i < count; i += n) { 1514 /* Maximum I/O size is limited by maximum swap block size. */ 1515 n = min(count - i, nsw_cluster_max); 1516 1517 if (async) { 1518 mtx_lock(&swbuf_mtx); 1519 while (nsw_wcount_async == 0) 1520 msleep(&nsw_wcount_async, &swbuf_mtx, PVM, 1521 "swbufa", 0); 1522 nsw_wcount_async--; 1523 mtx_unlock(&swbuf_mtx); 1524 } 1525 1526 /* Get a block of swap of size up to size n. */ 1527 VM_OBJECT_WLOCK(object); 1528 blk = swp_pager_getswapspace(&n); 1529 if (blk == SWAPBLK_NONE) { 1530 VM_OBJECT_WUNLOCK(object); 1531 mtx_lock(&swbuf_mtx); 1532 if (++nsw_wcount_async == 1) 1533 wakeup(&nsw_wcount_async); 1534 mtx_unlock(&swbuf_mtx); 1535 for (j = 0; j < n; ++j) 1536 rtvals[i + j] = VM_PAGER_FAIL; 1537 continue; 1538 } 1539 for (j = 0; j < n; ++j) { 1540 mreq = ma[i + j]; 1541 vm_page_aflag_clear(mreq, PGA_SWAP_FREE); 1542 addr = swp_pager_meta_build(mreq->object, mreq->pindex, 1543 blk + j); 1544 if (addr != SWAPBLK_NONE) 1545 swp_pager_update_freerange(&s_free, &n_free, 1546 addr); 1547 MPASS(mreq->dirty == VM_PAGE_BITS_ALL); 1548 mreq->oflags |= VPO_SWAPINPROG; 1549 } 1550 VM_OBJECT_WUNLOCK(object); 1551 1552 bp = uma_zalloc(swwbuf_zone, M_WAITOK); 1553 MPASS((bp->b_flags & B_MAXPHYS) != 0); 1554 if (async) 1555 bp->b_flags |= B_ASYNC; 1556 bp->b_flags |= B_PAGING; 1557 bp->b_iocmd = BIO_WRITE; 1558 1559 bp->b_rcred = crhold(thread0.td_ucred); 1560 bp->b_wcred = crhold(thread0.td_ucred); 1561 bp->b_bcount = PAGE_SIZE * n; 1562 bp->b_bufsize = PAGE_SIZE * n; 1563 bp->b_blkno = blk; 1564 for (j = 0; j < n; j++) 1565 bp->b_pages[j] = ma[i + j]; 1566 bp->b_npages = n; 1567 1568 /* 1569 * Must set dirty range for NFS to work. 1570 */ 1571 bp->b_dirtyoff = 0; 1572 bp->b_dirtyend = bp->b_bcount; 1573 1574 VM_CNT_INC(v_swapout); 1575 VM_CNT_ADD(v_swappgsout, bp->b_npages); 1576 1577 /* 1578 * We unconditionally set rtvals[] to VM_PAGER_PEND so that we 1579 * can call the async completion routine at the end of a 1580 * synchronous I/O operation. Otherwise, our caller would 1581 * perform duplicate unbusy and wakeup operations on the page 1582 * and object, respectively. 1583 */ 1584 for (j = 0; j < n; j++) 1585 rtvals[i + j] = VM_PAGER_PEND; 1586 1587 /* 1588 * asynchronous 1589 * 1590 * NOTE: b_blkno is destroyed by the call to swapdev_strategy. 1591 */ 1592 if (async) { 1593 bp->b_iodone = swp_pager_async_iodone; 1594 BUF_KERNPROC(bp); 1595 swp_pager_strategy(bp); 1596 continue; 1597 } 1598 1599 /* 1600 * synchronous 1601 * 1602 * NOTE: b_blkno is destroyed by the call to swapdev_strategy. 1603 */ 1604 bp->b_iodone = bdone; 1605 swp_pager_strategy(bp); 1606 1607 /* 1608 * Wait for the sync I/O to complete. 1609 */ 1610 bwait(bp, PVM, "swwrt"); 1611 1612 /* 1613 * Now that we are through with the bp, we can call the 1614 * normal async completion, which frees everything up. 1615 */ 1616 swp_pager_async_iodone(bp); 1617 } 1618 swp_pager_freeswapspace(s_free, n_free); 1619 VM_OBJECT_WLOCK(object); 1620 } 1621 1622 /* 1623 * swp_pager_async_iodone: 1624 * 1625 * Completion routine for asynchronous reads and writes from/to swap. 1626 * Also called manually by synchronous code to finish up a bp. 1627 * 1628 * This routine may not sleep. 1629 */ 1630 static void 1631 swp_pager_async_iodone(struct buf *bp) 1632 { 1633 int i; 1634 vm_object_t object = NULL; 1635 1636 /* 1637 * Report error - unless we ran out of memory, in which case 1638 * we've already logged it in swapgeom_strategy(). 1639 */ 1640 if (bp->b_ioflags & BIO_ERROR && bp->b_error != ENOMEM) { 1641 printf( 1642 "swap_pager: I/O error - %s failed; blkno %ld," 1643 "size %ld, error %d\n", 1644 ((bp->b_iocmd == BIO_READ) ? "pagein" : "pageout"), 1645 (long)bp->b_blkno, 1646 (long)bp->b_bcount, 1647 bp->b_error 1648 ); 1649 } 1650 1651 /* 1652 * remove the mapping for kernel virtual 1653 */ 1654 if (buf_mapped(bp)) 1655 pmap_qremove((vm_offset_t)bp->b_data, bp->b_npages); 1656 else 1657 bp->b_data = bp->b_kvabase; 1658 1659 if (bp->b_npages) { 1660 object = bp->b_pages[0]->object; 1661 VM_OBJECT_WLOCK(object); 1662 } 1663 1664 /* 1665 * cleanup pages. If an error occurs writing to swap, we are in 1666 * very serious trouble. If it happens to be a disk error, though, 1667 * we may be able to recover by reassigning the swap later on. So 1668 * in this case we remove the m->swapblk assignment for the page 1669 * but do not free it in the rlist. The errornous block(s) are thus 1670 * never reallocated as swap. Redirty the page and continue. 1671 */ 1672 for (i = 0; i < bp->b_npages; ++i) { 1673 vm_page_t m = bp->b_pages[i]; 1674 1675 m->oflags &= ~VPO_SWAPINPROG; 1676 if (m->oflags & VPO_SWAPSLEEP) { 1677 m->oflags &= ~VPO_SWAPSLEEP; 1678 wakeup(&object->handle); 1679 } 1680 1681 /* We always have space after I/O, successful or not. */ 1682 vm_page_aflag_set(m, PGA_SWAP_SPACE); 1683 1684 if (bp->b_ioflags & BIO_ERROR) { 1685 /* 1686 * If an error occurs I'd love to throw the swapblk 1687 * away without freeing it back to swapspace, so it 1688 * can never be used again. But I can't from an 1689 * interrupt. 1690 */ 1691 if (bp->b_iocmd == BIO_READ) { 1692 /* 1693 * NOTE: for reads, m->dirty will probably 1694 * be overridden by the original caller of 1695 * getpages so don't play cute tricks here. 1696 */ 1697 vm_page_invalid(m); 1698 } else { 1699 /* 1700 * If a write error occurs, reactivate page 1701 * so it doesn't clog the inactive list, 1702 * then finish the I/O. 1703 */ 1704 MPASS(m->dirty == VM_PAGE_BITS_ALL); 1705 1706 /* PQ_UNSWAPPABLE? */ 1707 vm_page_activate(m); 1708 vm_page_sunbusy(m); 1709 } 1710 } else if (bp->b_iocmd == BIO_READ) { 1711 /* 1712 * NOTE: for reads, m->dirty will probably be 1713 * overridden by the original caller of getpages so 1714 * we cannot set them in order to free the underlying 1715 * swap in a low-swap situation. I don't think we'd 1716 * want to do that anyway, but it was an optimization 1717 * that existed in the old swapper for a time before 1718 * it got ripped out due to precisely this problem. 1719 */ 1720 KASSERT(!pmap_page_is_mapped(m), 1721 ("swp_pager_async_iodone: page %p is mapped", m)); 1722 KASSERT(m->dirty == 0, 1723 ("swp_pager_async_iodone: page %p is dirty", m)); 1724 1725 vm_page_valid(m); 1726 if (i < bp->b_pgbefore || 1727 i >= bp->b_npages - bp->b_pgafter) 1728 vm_page_readahead_finish(m); 1729 } else { 1730 /* 1731 * For write success, clear the dirty 1732 * status, then finish the I/O ( which decrements the 1733 * busy count and possibly wakes waiter's up ). 1734 * A page is only written to swap after a period of 1735 * inactivity. Therefore, we do not expect it to be 1736 * reused. 1737 */ 1738 KASSERT(!pmap_page_is_write_mapped(m), 1739 ("swp_pager_async_iodone: page %p is not write" 1740 " protected", m)); 1741 vm_page_undirty(m); 1742 vm_page_deactivate_noreuse(m); 1743 vm_page_sunbusy(m); 1744 } 1745 } 1746 1747 /* 1748 * adjust pip. NOTE: the original parent may still have its own 1749 * pip refs on the object. 1750 */ 1751 if (object != NULL) { 1752 vm_object_pip_wakeupn(object, bp->b_npages); 1753 VM_OBJECT_WUNLOCK(object); 1754 } 1755 1756 /* 1757 * swapdev_strategy() manually sets b_vp and b_bufobj before calling 1758 * bstrategy(). Set them back to NULL now we're done with it, or we'll 1759 * trigger a KASSERT in relpbuf(). 1760 */ 1761 if (bp->b_vp) { 1762 bp->b_vp = NULL; 1763 bp->b_bufobj = NULL; 1764 } 1765 /* 1766 * release the physical I/O buffer 1767 */ 1768 if (bp->b_flags & B_ASYNC) { 1769 mtx_lock(&swbuf_mtx); 1770 if (++nsw_wcount_async == 1) 1771 wakeup(&nsw_wcount_async); 1772 mtx_unlock(&swbuf_mtx); 1773 } 1774 uma_zfree((bp->b_iocmd == BIO_READ) ? swrbuf_zone : swwbuf_zone, bp); 1775 } 1776 1777 int 1778 swap_pager_nswapdev(void) 1779 { 1780 1781 return (nswapdev); 1782 } 1783 1784 static void 1785 swp_pager_force_dirty(vm_page_t m) 1786 { 1787 1788 vm_page_dirty(m); 1789 swap_pager_unswapped(m); 1790 vm_page_launder(m); 1791 } 1792 1793 u_long 1794 swap_pager_swapped_pages(vm_object_t object) 1795 { 1796 struct swblk *sb; 1797 vm_pindex_t pi; 1798 u_long res; 1799 int i; 1800 1801 VM_OBJECT_ASSERT_LOCKED(object); 1802 if ((object->flags & OBJ_SWAP) == 0) 1803 return (0); 1804 1805 for (res = 0, pi = 0; (sb = SWAP_PCTRIE_LOOKUP_GE( 1806 &object->un_pager.swp.swp_blks, pi)) != NULL; 1807 pi = sb->p + SWAP_META_PAGES) { 1808 for (i = 0; i < SWAP_META_PAGES; i++) { 1809 if (sb->d[i] != SWAPBLK_NONE) 1810 res++; 1811 } 1812 } 1813 return (res); 1814 } 1815 1816 /* 1817 * swap_pager_swapoff_object: 1818 * 1819 * Page in all of the pages that have been paged out for an object 1820 * to a swap device. 1821 */ 1822 static void 1823 swap_pager_swapoff_object(struct swdevt *sp, vm_object_t object) 1824 { 1825 struct swblk *sb; 1826 vm_page_t m; 1827 vm_pindex_t pi; 1828 daddr_t blk; 1829 int i, nv, rahead, rv; 1830 1831 KASSERT((object->flags & OBJ_SWAP) != 0, 1832 ("%s: Object not swappable", __func__)); 1833 1834 for (pi = 0; (sb = SWAP_PCTRIE_LOOKUP_GE( 1835 &object->un_pager.swp.swp_blks, pi)) != NULL; ) { 1836 if ((object->flags & OBJ_DEAD) != 0) { 1837 /* 1838 * Make sure that pending writes finish before 1839 * returning. 1840 */ 1841 vm_object_pip_wait(object, "swpoff"); 1842 swp_pager_meta_free_all(object); 1843 break; 1844 } 1845 for (i = 0; i < SWAP_META_PAGES; i++) { 1846 /* 1847 * Count the number of contiguous valid blocks. 1848 */ 1849 for (nv = 0; nv < SWAP_META_PAGES - i; nv++) { 1850 blk = sb->d[i + nv]; 1851 if (!swp_pager_isondev(blk, sp) || 1852 blk == SWAPBLK_NONE) 1853 break; 1854 } 1855 if (nv == 0) 1856 continue; 1857 1858 /* 1859 * Look for a page corresponding to the first 1860 * valid block and ensure that any pending paging 1861 * operations on it are complete. If the page is valid, 1862 * mark it dirty and free the swap block. Try to batch 1863 * this operation since it may cause sp to be freed, 1864 * meaning that we must restart the scan. Avoid busying 1865 * valid pages since we may block forever on kernel 1866 * stack pages. 1867 */ 1868 m = vm_page_lookup(object, sb->p + i); 1869 if (m == NULL) { 1870 m = vm_page_alloc(object, sb->p + i, 1871 VM_ALLOC_NORMAL | VM_ALLOC_WAITFAIL); 1872 if (m == NULL) 1873 break; 1874 } else { 1875 if ((m->oflags & VPO_SWAPINPROG) != 0) { 1876 m->oflags |= VPO_SWAPSLEEP; 1877 VM_OBJECT_SLEEP(object, &object->handle, 1878 PSWP, "swpoff", 0); 1879 break; 1880 } 1881 if (vm_page_all_valid(m)) { 1882 do { 1883 swp_pager_force_dirty(m); 1884 } while (--nv > 0 && 1885 (m = vm_page_next(m)) != NULL && 1886 vm_page_all_valid(m) && 1887 (m->oflags & VPO_SWAPINPROG) == 0); 1888 break; 1889 } 1890 if (!vm_page_busy_acquire(m, VM_ALLOC_WAITFAIL)) 1891 break; 1892 } 1893 1894 vm_object_pip_add(object, 1); 1895 rahead = SWAP_META_PAGES; 1896 rv = swap_pager_getpages_locked(object, &m, 1, NULL, 1897 &rahead); 1898 if (rv != VM_PAGER_OK) 1899 panic("%s: read from swap failed: %d", 1900 __func__, rv); 1901 vm_object_pip_wakeupn(object, 1); 1902 VM_OBJECT_WLOCK(object); 1903 vm_page_xunbusy(m); 1904 1905 /* 1906 * The object lock was dropped so we must restart the 1907 * scan of this swap block. Pages paged in during this 1908 * iteration will be marked dirty in a future iteration. 1909 */ 1910 break; 1911 } 1912 if (i == SWAP_META_PAGES) 1913 pi = sb->p + SWAP_META_PAGES; 1914 } 1915 } 1916 1917 /* 1918 * swap_pager_swapoff: 1919 * 1920 * Page in all of the pages that have been paged out to the 1921 * given device. The corresponding blocks in the bitmap must be 1922 * marked as allocated and the device must be flagged SW_CLOSING. 1923 * There may be no processes swapped out to the device. 1924 * 1925 * This routine may block. 1926 */ 1927 static void 1928 swap_pager_swapoff(struct swdevt *sp) 1929 { 1930 vm_object_t object; 1931 int retries; 1932 1933 sx_assert(&swdev_syscall_lock, SA_XLOCKED); 1934 1935 retries = 0; 1936 full_rescan: 1937 mtx_lock(&vm_object_list_mtx); 1938 TAILQ_FOREACH(object, &vm_object_list, object_list) { 1939 if ((object->flags & OBJ_SWAP) == 0) 1940 continue; 1941 mtx_unlock(&vm_object_list_mtx); 1942 /* Depends on type-stability. */ 1943 VM_OBJECT_WLOCK(object); 1944 1945 /* 1946 * Dead objects are eventually terminated on their own. 1947 */ 1948 if ((object->flags & OBJ_DEAD) != 0) 1949 goto next_obj; 1950 1951 /* 1952 * Sync with fences placed after pctrie 1953 * initialization. We must not access pctrie below 1954 * unless we checked that our object is swap and not 1955 * dead. 1956 */ 1957 atomic_thread_fence_acq(); 1958 if ((object->flags & OBJ_SWAP) == 0) 1959 goto next_obj; 1960 1961 swap_pager_swapoff_object(sp, object); 1962 next_obj: 1963 VM_OBJECT_WUNLOCK(object); 1964 mtx_lock(&vm_object_list_mtx); 1965 } 1966 mtx_unlock(&vm_object_list_mtx); 1967 1968 if (sp->sw_used) { 1969 /* 1970 * Objects may be locked or paging to the device being 1971 * removed, so we will miss their pages and need to 1972 * make another pass. We have marked this device as 1973 * SW_CLOSING, so the activity should finish soon. 1974 */ 1975 retries++; 1976 if (retries > 100) { 1977 panic("swapoff: failed to locate %d swap blocks", 1978 sp->sw_used); 1979 } 1980 pause("swpoff", hz / 20); 1981 goto full_rescan; 1982 } 1983 EVENTHANDLER_INVOKE(swapoff, sp); 1984 } 1985 1986 /************************************************************************ 1987 * SWAP META DATA * 1988 ************************************************************************ 1989 * 1990 * These routines manipulate the swap metadata stored in the 1991 * OBJT_SWAP object. 1992 * 1993 * Swap metadata is implemented with a global hash and not directly 1994 * linked into the object. Instead the object simply contains 1995 * appropriate tracking counters. 1996 */ 1997 1998 /* 1999 * SWP_PAGER_SWBLK_EMPTY() - is a range of blocks free? 2000 */ 2001 static bool 2002 swp_pager_swblk_empty(struct swblk *sb, int start, int limit) 2003 { 2004 int i; 2005 2006 MPASS(0 <= start && start <= limit && limit <= SWAP_META_PAGES); 2007 for (i = start; i < limit; i++) { 2008 if (sb->d[i] != SWAPBLK_NONE) 2009 return (false); 2010 } 2011 return (true); 2012 } 2013 2014 /* 2015 * SWP_PAGER_FREE_EMPTY_SWBLK() - frees if a block is free 2016 * 2017 * Nothing is done if the block is still in use. 2018 */ 2019 static void 2020 swp_pager_free_empty_swblk(vm_object_t object, struct swblk *sb) 2021 { 2022 2023 if (swp_pager_swblk_empty(sb, 0, SWAP_META_PAGES)) { 2024 SWAP_PCTRIE_REMOVE(&object->un_pager.swp.swp_blks, sb->p); 2025 uma_zfree(swblk_zone, sb); 2026 } 2027 } 2028 2029 /* 2030 * SWP_PAGER_META_BUILD() - add swap block to swap meta data for object 2031 * 2032 * We first convert the object to a swap object if it is a default 2033 * object. 2034 * 2035 * The specified swapblk is added to the object's swap metadata. If 2036 * the swapblk is not valid, it is freed instead. Any previously 2037 * assigned swapblk is returned. 2038 */ 2039 static daddr_t 2040 swp_pager_meta_build(vm_object_t object, vm_pindex_t pindex, daddr_t swapblk) 2041 { 2042 static volatile int swblk_zone_exhausted, swpctrie_zone_exhausted; 2043 struct swblk *sb, *sb1; 2044 vm_pindex_t modpi, rdpi; 2045 daddr_t prev_swapblk; 2046 int error, i; 2047 2048 VM_OBJECT_ASSERT_WLOCKED(object); 2049 2050 /* 2051 * Convert default object to swap object if necessary 2052 */ 2053 if ((object->flags & OBJ_SWAP) == 0) { 2054 pctrie_init(&object->un_pager.swp.swp_blks); 2055 2056 /* 2057 * Ensure that swap_pager_swapoff()'s iteration over 2058 * object_list does not see a garbage pctrie. 2059 */ 2060 atomic_thread_fence_rel(); 2061 2062 object->type = OBJT_SWAP; 2063 vm_object_set_flag(object, OBJ_SWAP); 2064 object->un_pager.swp.writemappings = 0; 2065 KASSERT((object->flags & OBJ_ANON) != 0 || 2066 object->handle == NULL, 2067 ("default pager %p with handle %p", 2068 object, object->handle)); 2069 } 2070 2071 rdpi = rounddown(pindex, SWAP_META_PAGES); 2072 sb = SWAP_PCTRIE_LOOKUP(&object->un_pager.swp.swp_blks, rdpi); 2073 if (sb == NULL) { 2074 if (swapblk == SWAPBLK_NONE) 2075 return (SWAPBLK_NONE); 2076 for (;;) { 2077 sb = uma_zalloc(swblk_zone, M_NOWAIT | (curproc == 2078 pageproc ? M_USE_RESERVE : 0)); 2079 if (sb != NULL) { 2080 sb->p = rdpi; 2081 for (i = 0; i < SWAP_META_PAGES; i++) 2082 sb->d[i] = SWAPBLK_NONE; 2083 if (atomic_cmpset_int(&swblk_zone_exhausted, 2084 1, 0)) 2085 printf("swblk zone ok\n"); 2086 break; 2087 } 2088 VM_OBJECT_WUNLOCK(object); 2089 if (uma_zone_exhausted(swblk_zone)) { 2090 if (atomic_cmpset_int(&swblk_zone_exhausted, 2091 0, 1)) 2092 printf("swap blk zone exhausted, " 2093 "increase kern.maxswzone\n"); 2094 vm_pageout_oom(VM_OOM_SWAPZ); 2095 pause("swzonxb", 10); 2096 } else 2097 uma_zwait(swblk_zone); 2098 VM_OBJECT_WLOCK(object); 2099 sb = SWAP_PCTRIE_LOOKUP(&object->un_pager.swp.swp_blks, 2100 rdpi); 2101 if (sb != NULL) 2102 /* 2103 * Somebody swapped out a nearby page, 2104 * allocating swblk at the rdpi index, 2105 * while we dropped the object lock. 2106 */ 2107 goto allocated; 2108 } 2109 for (;;) { 2110 error = SWAP_PCTRIE_INSERT( 2111 &object->un_pager.swp.swp_blks, sb); 2112 if (error == 0) { 2113 if (atomic_cmpset_int(&swpctrie_zone_exhausted, 2114 1, 0)) 2115 printf("swpctrie zone ok\n"); 2116 break; 2117 } 2118 VM_OBJECT_WUNLOCK(object); 2119 if (uma_zone_exhausted(swpctrie_zone)) { 2120 if (atomic_cmpset_int(&swpctrie_zone_exhausted, 2121 0, 1)) 2122 printf("swap pctrie zone exhausted, " 2123 "increase kern.maxswzone\n"); 2124 vm_pageout_oom(VM_OOM_SWAPZ); 2125 pause("swzonxp", 10); 2126 } else 2127 uma_zwait(swpctrie_zone); 2128 VM_OBJECT_WLOCK(object); 2129 sb1 = SWAP_PCTRIE_LOOKUP(&object->un_pager.swp.swp_blks, 2130 rdpi); 2131 if (sb1 != NULL) { 2132 uma_zfree(swblk_zone, sb); 2133 sb = sb1; 2134 goto allocated; 2135 } 2136 } 2137 } 2138 allocated: 2139 MPASS(sb->p == rdpi); 2140 2141 modpi = pindex % SWAP_META_PAGES; 2142 /* Return prior contents of metadata. */ 2143 prev_swapblk = sb->d[modpi]; 2144 /* Enter block into metadata. */ 2145 sb->d[modpi] = swapblk; 2146 2147 /* 2148 * Free the swblk if we end up with the empty page run. 2149 */ 2150 if (swapblk == SWAPBLK_NONE) 2151 swp_pager_free_empty_swblk(object, sb); 2152 return (prev_swapblk); 2153 } 2154 2155 /* 2156 * SWP_PAGER_META_TRANSFER() - free a range of blocks in the srcobject's swap 2157 * metadata, or transfer it into dstobject. 2158 * 2159 * This routine will free swap metadata structures as they are cleaned 2160 * out. 2161 */ 2162 static void 2163 swp_pager_meta_transfer(vm_object_t srcobject, vm_object_t dstobject, 2164 vm_pindex_t pindex, vm_pindex_t count) 2165 { 2166 struct swblk *sb; 2167 daddr_t n_free, s_free; 2168 vm_pindex_t offset, last; 2169 int i, limit, start; 2170 2171 VM_OBJECT_ASSERT_WLOCKED(srcobject); 2172 if ((srcobject->flags & OBJ_SWAP) == 0 || count == 0) 2173 return; 2174 2175 swp_pager_init_freerange(&s_free, &n_free); 2176 offset = pindex; 2177 last = pindex + count; 2178 for (;;) { 2179 sb = SWAP_PCTRIE_LOOKUP_GE(&srcobject->un_pager.swp.swp_blks, 2180 rounddown(pindex, SWAP_META_PAGES)); 2181 if (sb == NULL || sb->p >= last) 2182 break; 2183 start = pindex > sb->p ? pindex - sb->p : 0; 2184 limit = last - sb->p < SWAP_META_PAGES ? last - sb->p : 2185 SWAP_META_PAGES; 2186 for (i = start; i < limit; i++) { 2187 if (sb->d[i] == SWAPBLK_NONE) 2188 continue; 2189 if (dstobject == NULL || 2190 !swp_pager_xfer_source(srcobject, dstobject, 2191 sb->p + i - offset, sb->d[i])) { 2192 swp_pager_update_freerange(&s_free, &n_free, 2193 sb->d[i]); 2194 } 2195 sb->d[i] = SWAPBLK_NONE; 2196 } 2197 pindex = sb->p + SWAP_META_PAGES; 2198 if (swp_pager_swblk_empty(sb, 0, start) && 2199 swp_pager_swblk_empty(sb, limit, SWAP_META_PAGES)) { 2200 SWAP_PCTRIE_REMOVE(&srcobject->un_pager.swp.swp_blks, 2201 sb->p); 2202 uma_zfree(swblk_zone, sb); 2203 } 2204 } 2205 swp_pager_freeswapspace(s_free, n_free); 2206 } 2207 2208 /* 2209 * SWP_PAGER_META_FREE() - free a range of blocks in the object's swap metadata 2210 * 2211 * The requested range of blocks is freed, with any associated swap 2212 * returned to the swap bitmap. 2213 * 2214 * This routine will free swap metadata structures as they are cleaned 2215 * out. This routine does *NOT* operate on swap metadata associated 2216 * with resident pages. 2217 */ 2218 static void 2219 swp_pager_meta_free(vm_object_t object, vm_pindex_t pindex, vm_pindex_t count) 2220 { 2221 swp_pager_meta_transfer(object, NULL, pindex, count); 2222 } 2223 2224 /* 2225 * SWP_PAGER_META_FREE_ALL() - destroy all swap metadata associated with object 2226 * 2227 * This routine locates and destroys all swap metadata associated with 2228 * an object. 2229 */ 2230 static void 2231 swp_pager_meta_free_all(vm_object_t object) 2232 { 2233 struct swblk *sb; 2234 daddr_t n_free, s_free; 2235 vm_pindex_t pindex; 2236 int i; 2237 2238 VM_OBJECT_ASSERT_WLOCKED(object); 2239 if ((object->flags & OBJ_SWAP) == 0) 2240 return; 2241 2242 swp_pager_init_freerange(&s_free, &n_free); 2243 for (pindex = 0; (sb = SWAP_PCTRIE_LOOKUP_GE( 2244 &object->un_pager.swp.swp_blks, pindex)) != NULL;) { 2245 pindex = sb->p + SWAP_META_PAGES; 2246 for (i = 0; i < SWAP_META_PAGES; i++) { 2247 if (sb->d[i] == SWAPBLK_NONE) 2248 continue; 2249 swp_pager_update_freerange(&s_free, &n_free, sb->d[i]); 2250 } 2251 SWAP_PCTRIE_REMOVE(&object->un_pager.swp.swp_blks, sb->p); 2252 uma_zfree(swblk_zone, sb); 2253 } 2254 swp_pager_freeswapspace(s_free, n_free); 2255 } 2256 2257 /* 2258 * SWP_PAGER_METACTL() - misc control of swap meta data. 2259 * 2260 * This routine is capable of looking up, or removing swapblk 2261 * assignments in the swap meta data. It returns the swapblk being 2262 * looked-up, popped, or SWAPBLK_NONE if the block was invalid. 2263 * 2264 * When acting on a busy resident page and paging is in progress, we 2265 * have to wait until paging is complete but otherwise can act on the 2266 * busy page. 2267 */ 2268 static daddr_t 2269 swp_pager_meta_lookup(vm_object_t object, vm_pindex_t pindex) 2270 { 2271 struct swblk *sb; 2272 2273 VM_OBJECT_ASSERT_LOCKED(object); 2274 2275 /* 2276 * The meta data only exists if the object is OBJT_SWAP 2277 * and even then might not be allocated yet. 2278 */ 2279 KASSERT((object->flags & OBJ_SWAP) != 0, 2280 ("Lookup object not swappable")); 2281 2282 sb = SWAP_PCTRIE_LOOKUP(&object->un_pager.swp.swp_blks, 2283 rounddown(pindex, SWAP_META_PAGES)); 2284 if (sb == NULL) 2285 return (SWAPBLK_NONE); 2286 return (sb->d[pindex % SWAP_META_PAGES]); 2287 } 2288 2289 /* 2290 * Returns the least page index which is greater than or equal to the 2291 * parameter pindex and for which there is a swap block allocated. 2292 * Returns object's size if the object's type is not swap or if there 2293 * are no allocated swap blocks for the object after the requested 2294 * pindex. 2295 */ 2296 vm_pindex_t 2297 swap_pager_find_least(vm_object_t object, vm_pindex_t pindex) 2298 { 2299 struct swblk *sb; 2300 int i; 2301 2302 VM_OBJECT_ASSERT_LOCKED(object); 2303 if ((object->flags & OBJ_SWAP) == 0) 2304 return (object->size); 2305 2306 sb = SWAP_PCTRIE_LOOKUP_GE(&object->un_pager.swp.swp_blks, 2307 rounddown(pindex, SWAP_META_PAGES)); 2308 if (sb == NULL) 2309 return (object->size); 2310 if (sb->p < pindex) { 2311 for (i = pindex % SWAP_META_PAGES; i < SWAP_META_PAGES; i++) { 2312 if (sb->d[i] != SWAPBLK_NONE) 2313 return (sb->p + i); 2314 } 2315 sb = SWAP_PCTRIE_LOOKUP_GE(&object->un_pager.swp.swp_blks, 2316 roundup(pindex, SWAP_META_PAGES)); 2317 if (sb == NULL) 2318 return (object->size); 2319 } 2320 for (i = 0; i < SWAP_META_PAGES; i++) { 2321 if (sb->d[i] != SWAPBLK_NONE) 2322 return (sb->p + i); 2323 } 2324 2325 /* 2326 * We get here if a swblk is present in the trie but it 2327 * doesn't map any blocks. 2328 */ 2329 MPASS(0); 2330 return (object->size); 2331 } 2332 2333 /* 2334 * System call swapon(name) enables swapping on device name, 2335 * which must be in the swdevsw. Return EBUSY 2336 * if already swapping on this device. 2337 */ 2338 #ifndef _SYS_SYSPROTO_H_ 2339 struct swapon_args { 2340 char *name; 2341 }; 2342 #endif 2343 2344 int 2345 sys_swapon(struct thread *td, struct swapon_args *uap) 2346 { 2347 struct vattr attr; 2348 struct vnode *vp; 2349 struct nameidata nd; 2350 int error; 2351 2352 error = priv_check(td, PRIV_SWAPON); 2353 if (error) 2354 return (error); 2355 2356 sx_xlock(&swdev_syscall_lock); 2357 2358 /* 2359 * Swap metadata may not fit in the KVM if we have physical 2360 * memory of >1GB. 2361 */ 2362 if (swblk_zone == NULL) { 2363 error = ENOMEM; 2364 goto done; 2365 } 2366 2367 NDINIT(&nd, LOOKUP, ISOPEN | FOLLOW | LOCKLEAF | AUDITVNODE1, 2368 UIO_USERSPACE, uap->name); 2369 error = namei(&nd); 2370 if (error) 2371 goto done; 2372 2373 NDFREE(&nd, NDF_ONLY_PNBUF); 2374 vp = nd.ni_vp; 2375 2376 if (vn_isdisk_error(vp, &error)) { 2377 error = swapongeom(vp); 2378 } else if (vp->v_type == VREG && 2379 (vp->v_mount->mnt_vfc->vfc_flags & VFCF_NETWORK) != 0 && 2380 (error = VOP_GETATTR(vp, &attr, td->td_ucred)) == 0) { 2381 /* 2382 * Allow direct swapping to NFS regular files in the same 2383 * way that nfs_mountroot() sets up diskless swapping. 2384 */ 2385 error = swaponvp(td, vp, attr.va_size / DEV_BSIZE); 2386 } 2387 2388 if (error != 0) 2389 vput(vp); 2390 else 2391 VOP_UNLOCK(vp); 2392 done: 2393 sx_xunlock(&swdev_syscall_lock); 2394 return (error); 2395 } 2396 2397 /* 2398 * Check that the total amount of swap currently configured does not 2399 * exceed half the theoretical maximum. If it does, print a warning 2400 * message. 2401 */ 2402 static void 2403 swapon_check_swzone(void) 2404 { 2405 2406 /* recommend using no more than half that amount */ 2407 if (swap_total > swap_maxpages / 2) { 2408 printf("warning: total configured swap (%lu pages) " 2409 "exceeds maximum recommended amount (%lu pages).\n", 2410 swap_total, swap_maxpages / 2); 2411 printf("warning: increase kern.maxswzone " 2412 "or reduce amount of swap.\n"); 2413 } 2414 } 2415 2416 static void 2417 swaponsomething(struct vnode *vp, void *id, u_long nblks, 2418 sw_strategy_t *strategy, sw_close_t *close, dev_t dev, int flags) 2419 { 2420 struct swdevt *sp, *tsp; 2421 daddr_t dvbase; 2422 2423 /* 2424 * nblks is in DEV_BSIZE'd chunks, convert to PAGE_SIZE'd chunks. 2425 * First chop nblks off to page-align it, then convert. 2426 * 2427 * sw->sw_nblks is in page-sized chunks now too. 2428 */ 2429 nblks &= ~(ctodb(1) - 1); 2430 nblks = dbtoc(nblks); 2431 2432 sp = malloc(sizeof *sp, M_VMPGDATA, M_WAITOK | M_ZERO); 2433 sp->sw_blist = blist_create(nblks, M_WAITOK); 2434 sp->sw_vp = vp; 2435 sp->sw_id = id; 2436 sp->sw_dev = dev; 2437 sp->sw_nblks = nblks; 2438 sp->sw_used = 0; 2439 sp->sw_strategy = strategy; 2440 sp->sw_close = close; 2441 sp->sw_flags = flags; 2442 2443 /* 2444 * Do not free the first blocks in order to avoid overwriting 2445 * any bsd label at the front of the partition 2446 */ 2447 blist_free(sp->sw_blist, howmany(BBSIZE, PAGE_SIZE), 2448 nblks - howmany(BBSIZE, PAGE_SIZE)); 2449 2450 dvbase = 0; 2451 mtx_lock(&sw_dev_mtx); 2452 TAILQ_FOREACH(tsp, &swtailq, sw_list) { 2453 if (tsp->sw_end >= dvbase) { 2454 /* 2455 * We put one uncovered page between the devices 2456 * in order to definitively prevent any cross-device 2457 * I/O requests 2458 */ 2459 dvbase = tsp->sw_end + 1; 2460 } 2461 } 2462 sp->sw_first = dvbase; 2463 sp->sw_end = dvbase + nblks; 2464 TAILQ_INSERT_TAIL(&swtailq, sp, sw_list); 2465 nswapdev++; 2466 swap_pager_avail += nblks - howmany(BBSIZE, PAGE_SIZE); 2467 swap_total += nblks; 2468 swapon_check_swzone(); 2469 swp_sizecheck(); 2470 mtx_unlock(&sw_dev_mtx); 2471 EVENTHANDLER_INVOKE(swapon, sp); 2472 } 2473 2474 /* 2475 * SYSCALL: swapoff(devname) 2476 * 2477 * Disable swapping on the given device. 2478 * 2479 * XXX: Badly designed system call: it should use a device index 2480 * rather than filename as specification. We keep sw_vp around 2481 * only to make this work. 2482 */ 2483 static int 2484 kern_swapoff(struct thread *td, const char *name, enum uio_seg name_seg, 2485 u_int flags) 2486 { 2487 struct vnode *vp; 2488 struct nameidata nd; 2489 struct swdevt *sp; 2490 int error; 2491 2492 error = priv_check(td, PRIV_SWAPOFF); 2493 if (error != 0) 2494 return (error); 2495 if ((flags & ~(SWAPOFF_FORCE)) != 0) 2496 return (EINVAL); 2497 2498 sx_xlock(&swdev_syscall_lock); 2499 2500 NDINIT(&nd, LOOKUP, FOLLOW | AUDITVNODE1, name_seg, name); 2501 error = namei(&nd); 2502 if (error) 2503 goto done; 2504 NDFREE(&nd, NDF_ONLY_PNBUF); 2505 vp = nd.ni_vp; 2506 2507 mtx_lock(&sw_dev_mtx); 2508 TAILQ_FOREACH(sp, &swtailq, sw_list) { 2509 if (sp->sw_vp == vp) 2510 break; 2511 } 2512 mtx_unlock(&sw_dev_mtx); 2513 if (sp == NULL) { 2514 error = EINVAL; 2515 goto done; 2516 } 2517 error = swapoff_one(sp, td->td_ucred, flags); 2518 done: 2519 sx_xunlock(&swdev_syscall_lock); 2520 return (error); 2521 } 2522 2523 2524 #ifdef COMPAT_FREEBSD13 2525 int 2526 freebsd13_swapoff(struct thread *td, struct freebsd13_swapoff_args *uap) 2527 { 2528 return (kern_swapoff(td, uap->name, UIO_USERSPACE, 0)); 2529 } 2530 #endif 2531 2532 int 2533 sys_swapoff(struct thread *td, struct swapoff_args *uap) 2534 { 2535 return (kern_swapoff(td, uap->name, UIO_USERSPACE, uap->flags)); 2536 } 2537 2538 static int 2539 swapoff_one(struct swdevt *sp, struct ucred *cred, u_int flags) 2540 { 2541 u_long nblks; 2542 #ifdef MAC 2543 int error; 2544 #endif 2545 2546 sx_assert(&swdev_syscall_lock, SA_XLOCKED); 2547 #ifdef MAC 2548 (void) vn_lock(sp->sw_vp, LK_EXCLUSIVE | LK_RETRY); 2549 error = mac_system_check_swapoff(cred, sp->sw_vp); 2550 (void) VOP_UNLOCK(sp->sw_vp); 2551 if (error != 0) 2552 return (error); 2553 #endif 2554 nblks = sp->sw_nblks; 2555 2556 /* 2557 * We can turn off this swap device safely only if the 2558 * available virtual memory in the system will fit the amount 2559 * of data we will have to page back in, plus an epsilon so 2560 * the system doesn't become critically low on swap space. 2561 * The vm_free_count() part does not account e.g. for clean 2562 * pages that can be immediately reclaimed without paging, so 2563 * this is a very rough estimation. 2564 * 2565 * On the other hand, not turning swap off on swapoff_all() 2566 * means that we can lose swap data when filesystems go away, 2567 * which is arguably worse. 2568 */ 2569 if ((flags & SWAPOFF_FORCE) == 0 && 2570 vm_free_count() + swap_pager_avail < nblks + nswap_lowat) 2571 return (ENOMEM); 2572 2573 /* 2574 * Prevent further allocations on this device. 2575 */ 2576 mtx_lock(&sw_dev_mtx); 2577 sp->sw_flags |= SW_CLOSING; 2578 swap_pager_avail -= blist_fill(sp->sw_blist, 0, nblks); 2579 swap_total -= nblks; 2580 mtx_unlock(&sw_dev_mtx); 2581 2582 /* 2583 * Page in the contents of the device and close it. 2584 */ 2585 swap_pager_swapoff(sp); 2586 2587 sp->sw_close(curthread, sp); 2588 mtx_lock(&sw_dev_mtx); 2589 sp->sw_id = NULL; 2590 TAILQ_REMOVE(&swtailq, sp, sw_list); 2591 nswapdev--; 2592 if (nswapdev == 0) { 2593 swap_pager_full = 2; 2594 swap_pager_almost_full = 1; 2595 } 2596 if (swdevhd == sp) 2597 swdevhd = NULL; 2598 mtx_unlock(&sw_dev_mtx); 2599 blist_destroy(sp->sw_blist); 2600 free(sp, M_VMPGDATA); 2601 return (0); 2602 } 2603 2604 void 2605 swapoff_all(void) 2606 { 2607 struct swdevt *sp, *spt; 2608 const char *devname; 2609 int error; 2610 2611 sx_xlock(&swdev_syscall_lock); 2612 2613 mtx_lock(&sw_dev_mtx); 2614 TAILQ_FOREACH_SAFE(sp, &swtailq, sw_list, spt) { 2615 mtx_unlock(&sw_dev_mtx); 2616 if (vn_isdisk(sp->sw_vp)) 2617 devname = devtoname(sp->sw_vp->v_rdev); 2618 else 2619 devname = "[file]"; 2620 error = swapoff_one(sp, thread0.td_ucred, SWAPOFF_FORCE); 2621 if (error != 0) { 2622 printf("Cannot remove swap device %s (error=%d), " 2623 "skipping.\n", devname, error); 2624 } else if (bootverbose) { 2625 printf("Swap device %s removed.\n", devname); 2626 } 2627 mtx_lock(&sw_dev_mtx); 2628 } 2629 mtx_unlock(&sw_dev_mtx); 2630 2631 sx_xunlock(&swdev_syscall_lock); 2632 } 2633 2634 void 2635 swap_pager_status(int *total, int *used) 2636 { 2637 2638 *total = swap_total; 2639 *used = swap_total - swap_pager_avail - 2640 nswapdev * howmany(BBSIZE, PAGE_SIZE); 2641 } 2642 2643 int 2644 swap_dev_info(int name, struct xswdev *xs, char *devname, size_t len) 2645 { 2646 struct swdevt *sp; 2647 const char *tmp_devname; 2648 int error, n; 2649 2650 n = 0; 2651 error = ENOENT; 2652 mtx_lock(&sw_dev_mtx); 2653 TAILQ_FOREACH(sp, &swtailq, sw_list) { 2654 if (n != name) { 2655 n++; 2656 continue; 2657 } 2658 xs->xsw_version = XSWDEV_VERSION; 2659 xs->xsw_dev = sp->sw_dev; 2660 xs->xsw_flags = sp->sw_flags; 2661 xs->xsw_nblks = sp->sw_nblks; 2662 xs->xsw_used = sp->sw_used; 2663 if (devname != NULL) { 2664 if (vn_isdisk(sp->sw_vp)) 2665 tmp_devname = devtoname(sp->sw_vp->v_rdev); 2666 else 2667 tmp_devname = "[file]"; 2668 strncpy(devname, tmp_devname, len); 2669 } 2670 error = 0; 2671 break; 2672 } 2673 mtx_unlock(&sw_dev_mtx); 2674 return (error); 2675 } 2676 2677 #if defined(COMPAT_FREEBSD11) 2678 #define XSWDEV_VERSION_11 1 2679 struct xswdev11 { 2680 u_int xsw_version; 2681 uint32_t xsw_dev; 2682 int xsw_flags; 2683 int xsw_nblks; 2684 int xsw_used; 2685 }; 2686 #endif 2687 2688 #if defined(__amd64__) && defined(COMPAT_FREEBSD32) 2689 struct xswdev32 { 2690 u_int xsw_version; 2691 u_int xsw_dev1, xsw_dev2; 2692 int xsw_flags; 2693 int xsw_nblks; 2694 int xsw_used; 2695 }; 2696 #endif 2697 2698 static int 2699 sysctl_vm_swap_info(SYSCTL_HANDLER_ARGS) 2700 { 2701 struct xswdev xs; 2702 #if defined(__amd64__) && defined(COMPAT_FREEBSD32) 2703 struct xswdev32 xs32; 2704 #endif 2705 #if defined(COMPAT_FREEBSD11) 2706 struct xswdev11 xs11; 2707 #endif 2708 int error; 2709 2710 if (arg2 != 1) /* name length */ 2711 return (EINVAL); 2712 2713 memset(&xs, 0, sizeof(xs)); 2714 error = swap_dev_info(*(int *)arg1, &xs, NULL, 0); 2715 if (error != 0) 2716 return (error); 2717 #if defined(__amd64__) && defined(COMPAT_FREEBSD32) 2718 if (req->oldlen == sizeof(xs32)) { 2719 memset(&xs32, 0, sizeof(xs32)); 2720 xs32.xsw_version = XSWDEV_VERSION; 2721 xs32.xsw_dev1 = xs.xsw_dev; 2722 xs32.xsw_dev2 = xs.xsw_dev >> 32; 2723 xs32.xsw_flags = xs.xsw_flags; 2724 xs32.xsw_nblks = xs.xsw_nblks; 2725 xs32.xsw_used = xs.xsw_used; 2726 error = SYSCTL_OUT(req, &xs32, sizeof(xs32)); 2727 return (error); 2728 } 2729 #endif 2730 #if defined(COMPAT_FREEBSD11) 2731 if (req->oldlen == sizeof(xs11)) { 2732 memset(&xs11, 0, sizeof(xs11)); 2733 xs11.xsw_version = XSWDEV_VERSION_11; 2734 xs11.xsw_dev = xs.xsw_dev; /* truncation */ 2735 xs11.xsw_flags = xs.xsw_flags; 2736 xs11.xsw_nblks = xs.xsw_nblks; 2737 xs11.xsw_used = xs.xsw_used; 2738 error = SYSCTL_OUT(req, &xs11, sizeof(xs11)); 2739 return (error); 2740 } 2741 #endif 2742 error = SYSCTL_OUT(req, &xs, sizeof(xs)); 2743 return (error); 2744 } 2745 2746 SYSCTL_INT(_vm, OID_AUTO, nswapdev, CTLFLAG_RD, &nswapdev, 0, 2747 "Number of swap devices"); 2748 SYSCTL_NODE(_vm, OID_AUTO, swap_info, CTLFLAG_RD | CTLFLAG_MPSAFE, 2749 sysctl_vm_swap_info, 2750 "Swap statistics by device"); 2751 2752 /* 2753 * Count the approximate swap usage in pages for a vmspace. The 2754 * shadowed or not yet copied on write swap blocks are not accounted. 2755 * The map must be locked. 2756 */ 2757 long 2758 vmspace_swap_count(struct vmspace *vmspace) 2759 { 2760 vm_map_t map; 2761 vm_map_entry_t cur; 2762 vm_object_t object; 2763 struct swblk *sb; 2764 vm_pindex_t e, pi; 2765 long count; 2766 int i; 2767 2768 map = &vmspace->vm_map; 2769 count = 0; 2770 2771 VM_MAP_ENTRY_FOREACH(cur, map) { 2772 if ((cur->eflags & MAP_ENTRY_IS_SUB_MAP) != 0) 2773 continue; 2774 object = cur->object.vm_object; 2775 if (object == NULL || (object->flags & OBJ_SWAP) == 0) 2776 continue; 2777 VM_OBJECT_RLOCK(object); 2778 if ((object->flags & OBJ_SWAP) == 0) 2779 goto unlock; 2780 pi = OFF_TO_IDX(cur->offset); 2781 e = pi + OFF_TO_IDX(cur->end - cur->start); 2782 for (;; pi = sb->p + SWAP_META_PAGES) { 2783 sb = SWAP_PCTRIE_LOOKUP_GE( 2784 &object->un_pager.swp.swp_blks, pi); 2785 if (sb == NULL || sb->p >= e) 2786 break; 2787 for (i = 0; i < SWAP_META_PAGES; i++) { 2788 if (sb->p + i < e && 2789 sb->d[i] != SWAPBLK_NONE) 2790 count++; 2791 } 2792 } 2793 unlock: 2794 VM_OBJECT_RUNLOCK(object); 2795 } 2796 return (count); 2797 } 2798 2799 /* 2800 * GEOM backend 2801 * 2802 * Swapping onto disk devices. 2803 * 2804 */ 2805 2806 static g_orphan_t swapgeom_orphan; 2807 2808 static struct g_class g_swap_class = { 2809 .name = "SWAP", 2810 .version = G_VERSION, 2811 .orphan = swapgeom_orphan, 2812 }; 2813 2814 DECLARE_GEOM_CLASS(g_swap_class, g_class); 2815 2816 static void 2817 swapgeom_close_ev(void *arg, int flags) 2818 { 2819 struct g_consumer *cp; 2820 2821 cp = arg; 2822 g_access(cp, -1, -1, 0); 2823 g_detach(cp); 2824 g_destroy_consumer(cp); 2825 } 2826 2827 /* 2828 * Add a reference to the g_consumer for an inflight transaction. 2829 */ 2830 static void 2831 swapgeom_acquire(struct g_consumer *cp) 2832 { 2833 2834 mtx_assert(&sw_dev_mtx, MA_OWNED); 2835 cp->index++; 2836 } 2837 2838 /* 2839 * Remove a reference from the g_consumer. Post a close event if all 2840 * references go away, since the function might be called from the 2841 * biodone context. 2842 */ 2843 static void 2844 swapgeom_release(struct g_consumer *cp, struct swdevt *sp) 2845 { 2846 2847 mtx_assert(&sw_dev_mtx, MA_OWNED); 2848 cp->index--; 2849 if (cp->index == 0) { 2850 if (g_post_event(swapgeom_close_ev, cp, M_NOWAIT, NULL) == 0) 2851 sp->sw_id = NULL; 2852 } 2853 } 2854 2855 static void 2856 swapgeom_done(struct bio *bp2) 2857 { 2858 struct swdevt *sp; 2859 struct buf *bp; 2860 struct g_consumer *cp; 2861 2862 bp = bp2->bio_caller2; 2863 cp = bp2->bio_from; 2864 bp->b_ioflags = bp2->bio_flags; 2865 if (bp2->bio_error) 2866 bp->b_ioflags |= BIO_ERROR; 2867 bp->b_resid = bp->b_bcount - bp2->bio_completed; 2868 bp->b_error = bp2->bio_error; 2869 bp->b_caller1 = NULL; 2870 bufdone(bp); 2871 sp = bp2->bio_caller1; 2872 mtx_lock(&sw_dev_mtx); 2873 swapgeom_release(cp, sp); 2874 mtx_unlock(&sw_dev_mtx); 2875 g_destroy_bio(bp2); 2876 } 2877 2878 static void 2879 swapgeom_strategy(struct buf *bp, struct swdevt *sp) 2880 { 2881 struct bio *bio; 2882 struct g_consumer *cp; 2883 2884 mtx_lock(&sw_dev_mtx); 2885 cp = sp->sw_id; 2886 if (cp == NULL) { 2887 mtx_unlock(&sw_dev_mtx); 2888 bp->b_error = ENXIO; 2889 bp->b_ioflags |= BIO_ERROR; 2890 bufdone(bp); 2891 return; 2892 } 2893 swapgeom_acquire(cp); 2894 mtx_unlock(&sw_dev_mtx); 2895 if (bp->b_iocmd == BIO_WRITE) 2896 bio = g_new_bio(); 2897 else 2898 bio = g_alloc_bio(); 2899 if (bio == NULL) { 2900 mtx_lock(&sw_dev_mtx); 2901 swapgeom_release(cp, sp); 2902 mtx_unlock(&sw_dev_mtx); 2903 bp->b_error = ENOMEM; 2904 bp->b_ioflags |= BIO_ERROR; 2905 printf("swap_pager: cannot allocate bio\n"); 2906 bufdone(bp); 2907 return; 2908 } 2909 2910 bp->b_caller1 = bio; 2911 bio->bio_caller1 = sp; 2912 bio->bio_caller2 = bp; 2913 bio->bio_cmd = bp->b_iocmd; 2914 bio->bio_offset = (bp->b_blkno - sp->sw_first) * PAGE_SIZE; 2915 bio->bio_length = bp->b_bcount; 2916 bio->bio_done = swapgeom_done; 2917 bio->bio_flags |= BIO_SWAP; 2918 if (!buf_mapped(bp)) { 2919 bio->bio_ma = bp->b_pages; 2920 bio->bio_data = unmapped_buf; 2921 bio->bio_ma_offset = (vm_offset_t)bp->b_offset & PAGE_MASK; 2922 bio->bio_ma_n = bp->b_npages; 2923 bio->bio_flags |= BIO_UNMAPPED; 2924 } else { 2925 bio->bio_data = bp->b_data; 2926 bio->bio_ma = NULL; 2927 } 2928 g_io_request(bio, cp); 2929 return; 2930 } 2931 2932 static void 2933 swapgeom_orphan(struct g_consumer *cp) 2934 { 2935 struct swdevt *sp; 2936 int destroy; 2937 2938 mtx_lock(&sw_dev_mtx); 2939 TAILQ_FOREACH(sp, &swtailq, sw_list) { 2940 if (sp->sw_id == cp) { 2941 sp->sw_flags |= SW_CLOSING; 2942 break; 2943 } 2944 } 2945 /* 2946 * Drop reference we were created with. Do directly since we're in a 2947 * special context where we don't have to queue the call to 2948 * swapgeom_close_ev(). 2949 */ 2950 cp->index--; 2951 destroy = ((sp != NULL) && (cp->index == 0)); 2952 if (destroy) 2953 sp->sw_id = NULL; 2954 mtx_unlock(&sw_dev_mtx); 2955 if (destroy) 2956 swapgeom_close_ev(cp, 0); 2957 } 2958 2959 static void 2960 swapgeom_close(struct thread *td, struct swdevt *sw) 2961 { 2962 struct g_consumer *cp; 2963 2964 mtx_lock(&sw_dev_mtx); 2965 cp = sw->sw_id; 2966 sw->sw_id = NULL; 2967 mtx_unlock(&sw_dev_mtx); 2968 2969 /* 2970 * swapgeom_close() may be called from the biodone context, 2971 * where we cannot perform topology changes. Delegate the 2972 * work to the events thread. 2973 */ 2974 if (cp != NULL) 2975 g_waitfor_event(swapgeom_close_ev, cp, M_WAITOK, NULL); 2976 } 2977 2978 static int 2979 swapongeom_locked(struct cdev *dev, struct vnode *vp) 2980 { 2981 struct g_provider *pp; 2982 struct g_consumer *cp; 2983 static struct g_geom *gp; 2984 struct swdevt *sp; 2985 u_long nblks; 2986 int error; 2987 2988 pp = g_dev_getprovider(dev); 2989 if (pp == NULL) 2990 return (ENODEV); 2991 mtx_lock(&sw_dev_mtx); 2992 TAILQ_FOREACH(sp, &swtailq, sw_list) { 2993 cp = sp->sw_id; 2994 if (cp != NULL && cp->provider == pp) { 2995 mtx_unlock(&sw_dev_mtx); 2996 return (EBUSY); 2997 } 2998 } 2999 mtx_unlock(&sw_dev_mtx); 3000 if (gp == NULL) 3001 gp = g_new_geomf(&g_swap_class, "swap"); 3002 cp = g_new_consumer(gp); 3003 cp->index = 1; /* Number of active I/Os, plus one for being active. */ 3004 cp->flags |= G_CF_DIRECT_SEND | G_CF_DIRECT_RECEIVE; 3005 g_attach(cp, pp); 3006 /* 3007 * XXX: Every time you think you can improve the margin for 3008 * footshooting, somebody depends on the ability to do so: 3009 * savecore(8) wants to write to our swapdev so we cannot 3010 * set an exclusive count :-( 3011 */ 3012 error = g_access(cp, 1, 1, 0); 3013 if (error != 0) { 3014 g_detach(cp); 3015 g_destroy_consumer(cp); 3016 return (error); 3017 } 3018 nblks = pp->mediasize / DEV_BSIZE; 3019 swaponsomething(vp, cp, nblks, swapgeom_strategy, 3020 swapgeom_close, dev2udev(dev), 3021 (pp->flags & G_PF_ACCEPT_UNMAPPED) != 0 ? SW_UNMAPPED : 0); 3022 return (0); 3023 } 3024 3025 static int 3026 swapongeom(struct vnode *vp) 3027 { 3028 int error; 3029 3030 ASSERT_VOP_ELOCKED(vp, "swapongeom"); 3031 if (vp->v_type != VCHR || VN_IS_DOOMED(vp)) { 3032 error = ENOENT; 3033 } else { 3034 g_topology_lock(); 3035 error = swapongeom_locked(vp->v_rdev, vp); 3036 g_topology_unlock(); 3037 } 3038 return (error); 3039 } 3040 3041 /* 3042 * VNODE backend 3043 * 3044 * This is used mainly for network filesystem (read: probably only tested 3045 * with NFS) swapfiles. 3046 * 3047 */ 3048 3049 static void 3050 swapdev_strategy(struct buf *bp, struct swdevt *sp) 3051 { 3052 struct vnode *vp2; 3053 3054 bp->b_blkno = ctodb(bp->b_blkno - sp->sw_first); 3055 3056 vp2 = sp->sw_id; 3057 vhold(vp2); 3058 if (bp->b_iocmd == BIO_WRITE) { 3059 vn_lock(vp2, LK_EXCLUSIVE | LK_RETRY); 3060 if (bp->b_bufobj) 3061 bufobj_wdrop(bp->b_bufobj); 3062 bufobj_wref(&vp2->v_bufobj); 3063 } else { 3064 vn_lock(vp2, LK_SHARED | LK_RETRY); 3065 } 3066 if (bp->b_bufobj != &vp2->v_bufobj) 3067 bp->b_bufobj = &vp2->v_bufobj; 3068 bp->b_vp = vp2; 3069 bp->b_iooffset = dbtob(bp->b_blkno); 3070 bstrategy(bp); 3071 VOP_UNLOCK(vp2); 3072 } 3073 3074 static void 3075 swapdev_close(struct thread *td, struct swdevt *sp) 3076 { 3077 struct vnode *vp; 3078 3079 vp = sp->sw_vp; 3080 vn_lock(vp, LK_EXCLUSIVE | LK_RETRY); 3081 VOP_CLOSE(vp, FREAD | FWRITE, td->td_ucred, td); 3082 vput(vp); 3083 } 3084 3085 static int 3086 swaponvp(struct thread *td, struct vnode *vp, u_long nblks) 3087 { 3088 struct swdevt *sp; 3089 int error; 3090 3091 ASSERT_VOP_ELOCKED(vp, "swaponvp"); 3092 if (nblks == 0) 3093 return (ENXIO); 3094 mtx_lock(&sw_dev_mtx); 3095 TAILQ_FOREACH(sp, &swtailq, sw_list) { 3096 if (sp->sw_id == vp) { 3097 mtx_unlock(&sw_dev_mtx); 3098 return (EBUSY); 3099 } 3100 } 3101 mtx_unlock(&sw_dev_mtx); 3102 3103 #ifdef MAC 3104 error = mac_system_check_swapon(td->td_ucred, vp); 3105 if (error == 0) 3106 #endif 3107 error = VOP_OPEN(vp, FREAD | FWRITE, td->td_ucred, td, NULL); 3108 if (error != 0) 3109 return (error); 3110 3111 swaponsomething(vp, vp, nblks, swapdev_strategy, swapdev_close, 3112 NODEV, 0); 3113 return (0); 3114 } 3115 3116 static int 3117 sysctl_swap_async_max(SYSCTL_HANDLER_ARGS) 3118 { 3119 int error, new, n; 3120 3121 new = nsw_wcount_async_max; 3122 error = sysctl_handle_int(oidp, &new, 0, req); 3123 if (error != 0 || req->newptr == NULL) 3124 return (error); 3125 3126 if (new > nswbuf / 2 || new < 1) 3127 return (EINVAL); 3128 3129 mtx_lock(&swbuf_mtx); 3130 while (nsw_wcount_async_max != new) { 3131 /* 3132 * Adjust difference. If the current async count is too low, 3133 * we will need to sqeeze our update slowly in. Sleep with a 3134 * higher priority than getpbuf() to finish faster. 3135 */ 3136 n = new - nsw_wcount_async_max; 3137 if (nsw_wcount_async + n >= 0) { 3138 nsw_wcount_async += n; 3139 nsw_wcount_async_max += n; 3140 wakeup(&nsw_wcount_async); 3141 } else { 3142 nsw_wcount_async_max -= nsw_wcount_async; 3143 nsw_wcount_async = 0; 3144 msleep(&nsw_wcount_async, &swbuf_mtx, PSWP, 3145 "swpsysctl", 0); 3146 } 3147 } 3148 mtx_unlock(&swbuf_mtx); 3149 3150 return (0); 3151 } 3152 3153 static void 3154 swap_pager_update_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 3165 static void 3166 swap_pager_release_writecount(vm_object_t object, vm_offset_t start, 3167 vm_offset_t end) 3168 { 3169 3170 VM_OBJECT_WLOCK(object); 3171 KASSERT((object->flags & OBJ_ANON) == 0, 3172 ("Splittable object with writecount")); 3173 object->un_pager.swp.writemappings -= (vm_ooffset_t)end - start; 3174 VM_OBJECT_WUNLOCK(object); 3175 } 3176