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