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 68 #include <sys/cdefs.h> 69 #include "opt_vm.h" 70 71 #include <sys/param.h> 72 #include <sys/bio.h> 73 #include <sys/blist.h> 74 #include <sys/buf.h> 75 #include <sys/conf.h> 76 #include <sys/disk.h> 77 #include <sys/disklabel.h> 78 #include <sys/eventhandler.h> 79 #include <sys/fcntl.h> 80 #include <sys/limits.h> 81 #include <sys/lock.h> 82 #include <sys/kernel.h> 83 #include <sys/mount.h> 84 #include <sys/namei.h> 85 #include <sys/malloc.h> 86 #include <sys/pctrie.h> 87 #include <sys/priv.h> 88 #include <sys/proc.h> 89 #include <sys/racct.h> 90 #include <sys/resource.h> 91 #include <sys/resourcevar.h> 92 #include <sys/rwlock.h> 93 #include <sys/sbuf.h> 94 #include <sys/sysctl.h> 95 #include <sys/sysproto.h> 96 #include <sys/systm.h> 97 #include <sys/sx.h> 98 #include <sys/unistd.h> 99 #include <sys/user.h> 100 #include <sys/vmmeter.h> 101 #include <sys/vnode.h> 102 103 #include <security/mac/mac_framework.h> 104 105 #include <vm/vm.h> 106 #include <vm/pmap.h> 107 #include <vm/vm_map.h> 108 #include <vm/vm_kern.h> 109 #include <vm/vm_object.h> 110 #include <vm/vm_page.h> 111 #include <vm/vm_pager.h> 112 #include <vm/vm_pageout.h> 113 #include <vm/vm_param.h> 114 #include <vm/vm_radix.h> 115 #include <vm/swap_pager.h> 116 #include <vm/vm_extern.h> 117 #include <vm/uma.h> 118 119 #include <geom/geom.h> 120 121 /* 122 * MAX_PAGEOUT_CLUSTER must be a power of 2 between 1 and 64. 123 * The 64-page limit is due to the radix code (kern/subr_blist.c). 124 */ 125 #ifndef MAX_PAGEOUT_CLUSTER 126 #define MAX_PAGEOUT_CLUSTER 32 127 #endif 128 129 #if !defined(SWB_NPAGES) 130 #define SWB_NPAGES MAX_PAGEOUT_CLUSTER 131 #endif 132 133 #define SWAP_META_PAGES PCTRIE_COUNT 134 135 /* 136 * A swblk structure maps each page index within a 137 * SWAP_META_PAGES-aligned and sized range to the address of an 138 * on-disk swap block (or SWAPBLK_NONE). The collection of these 139 * mappings for an entire vm object is implemented as a pc-trie. 140 */ 141 struct swblk { 142 vm_pindex_t p; 143 daddr_t d[SWAP_META_PAGES]; 144 }; 145 146 /* 147 * A page_range structure records the start address and length of a sequence of 148 * mapped page addresses. 149 */ 150 struct page_range { 151 daddr_t start; 152 daddr_t num; 153 }; 154 155 static MALLOC_DEFINE(M_VMPGDATA, "vm_pgdata", "swap pager private data"); 156 static struct mtx sw_dev_mtx; 157 static TAILQ_HEAD(, swdevt) swtailq = TAILQ_HEAD_INITIALIZER(swtailq); 158 static struct swdevt *swdevhd; /* Allocate from here next */ 159 static int nswapdev; /* Number of swap devices */ 160 int swap_pager_avail; 161 static struct sx swdev_syscall_lock; /* serialize swap(on|off) */ 162 163 static __exclusive_cache_line u_long swap_reserved; 164 static u_long swap_total; 165 static int sysctl_page_shift(SYSCTL_HANDLER_ARGS); 166 167 static SYSCTL_NODE(_vm_stats, OID_AUTO, swap, CTLFLAG_RD | CTLFLAG_MPSAFE, 0, 168 "VM swap stats"); 169 170 SYSCTL_PROC(_vm, OID_AUTO, swap_reserved, CTLTYPE_U64 | CTLFLAG_RD | CTLFLAG_MPSAFE, 171 &swap_reserved, 0, sysctl_page_shift, "QU", 172 "Amount of swap storage needed to back all allocated anonymous memory."); 173 SYSCTL_PROC(_vm, OID_AUTO, swap_total, CTLTYPE_U64 | CTLFLAG_RD | CTLFLAG_MPSAFE, 174 &swap_total, 0, sysctl_page_shift, "QU", 175 "Total amount of available swap storage."); 176 177 int vm_overcommit __read_mostly = 0; 178 SYSCTL_INT(_vm, VM_OVERCOMMIT, overcommit, CTLFLAG_RW, &vm_overcommit, 0, 179 "Configure virtual memory overcommit behavior. See tuning(7) " 180 "for details."); 181 static unsigned long swzone; 182 SYSCTL_ULONG(_vm, OID_AUTO, swzone, CTLFLAG_RD, &swzone, 0, 183 "Actual size of swap metadata zone"); 184 static unsigned long swap_maxpages; 185 SYSCTL_ULONG(_vm, OID_AUTO, swap_maxpages, CTLFLAG_RD, &swap_maxpages, 0, 186 "Maximum amount of swap supported"); 187 188 static COUNTER_U64_DEFINE_EARLY(swap_free_deferred); 189 SYSCTL_COUNTER_U64(_vm_stats_swap, OID_AUTO, free_deferred, 190 CTLFLAG_RD, &swap_free_deferred, 191 "Number of pages that deferred freeing swap space"); 192 193 static COUNTER_U64_DEFINE_EARLY(swap_free_completed); 194 SYSCTL_COUNTER_U64(_vm_stats_swap, OID_AUTO, free_completed, 195 CTLFLAG_RD, &swap_free_completed, 196 "Number of deferred frees completed"); 197 198 static int 199 sysctl_page_shift(SYSCTL_HANDLER_ARGS) 200 { 201 uint64_t newval; 202 u_long value = *(u_long *)arg1; 203 204 newval = ((uint64_t)value) << PAGE_SHIFT; 205 return (sysctl_handle_64(oidp, &newval, 0, req)); 206 } 207 208 static bool 209 swap_reserve_by_cred_rlimit(u_long pincr, struct ucred *cred, int oc) 210 { 211 struct uidinfo *uip; 212 u_long prev; 213 214 uip = cred->cr_ruidinfo; 215 216 prev = atomic_fetchadd_long(&uip->ui_vmsize, pincr); 217 if ((oc & SWAP_RESERVE_RLIMIT_ON) != 0 && 218 prev + pincr > lim_cur(curthread, RLIMIT_SWAP) && 219 priv_check(curthread, PRIV_VM_SWAP_NORLIMIT) != 0) { 220 prev = atomic_fetchadd_long(&uip->ui_vmsize, -pincr); 221 KASSERT(prev >= pincr, 222 ("negative vmsize for uid %d\n", uip->ui_uid)); 223 return (false); 224 } 225 return (true); 226 } 227 228 static void 229 swap_release_by_cred_rlimit(u_long pdecr, struct ucred *cred) 230 { 231 struct uidinfo *uip; 232 #ifdef INVARIANTS 233 u_long prev; 234 #endif 235 236 uip = cred->cr_ruidinfo; 237 238 #ifdef INVARIANTS 239 prev = atomic_fetchadd_long(&uip->ui_vmsize, -pdecr); 240 KASSERT(prev >= pdecr, 241 ("negative vmsize for uid %d\n", uip->ui_uid)); 242 #else 243 atomic_subtract_long(&uip->ui_vmsize, pdecr); 244 #endif 245 } 246 247 static void 248 swap_reserve_force_rlimit(u_long pincr, struct ucred *cred) 249 { 250 struct uidinfo *uip; 251 252 uip = cred->cr_ruidinfo; 253 atomic_add_long(&uip->ui_vmsize, pincr); 254 } 255 256 bool 257 swap_reserve(vm_ooffset_t incr) 258 { 259 260 return (swap_reserve_by_cred(incr, curthread->td_ucred)); 261 } 262 263 bool 264 swap_reserve_by_cred(vm_ooffset_t incr, struct ucred *cred) 265 { 266 u_long r, s, prev, pincr; 267 #ifdef RACCT 268 int error; 269 #endif 270 int oc; 271 static int curfail; 272 static struct timeval lastfail; 273 274 KASSERT((incr & PAGE_MASK) == 0, ("%s: incr: %ju & PAGE_MASK", 275 __func__, (uintmax_t)incr)); 276 277 #ifdef RACCT 278 if (RACCT_ENABLED()) { 279 PROC_LOCK(curproc); 280 error = racct_add(curproc, RACCT_SWAP, incr); 281 PROC_UNLOCK(curproc); 282 if (error != 0) 283 return (false); 284 } 285 #endif 286 287 pincr = atop(incr); 288 prev = atomic_fetchadd_long(&swap_reserved, pincr); 289 r = prev + pincr; 290 s = swap_total; 291 oc = atomic_load_int(&vm_overcommit); 292 if (r > s && (oc & SWAP_RESERVE_ALLOW_NONWIRED) != 0) { 293 s += vm_cnt.v_page_count - vm_cnt.v_free_reserved - 294 vm_wire_count(); 295 } 296 if ((oc & SWAP_RESERVE_FORCE_ON) != 0 && r > s && 297 priv_check(curthread, PRIV_VM_SWAP_NOQUOTA) != 0) { 298 prev = atomic_fetchadd_long(&swap_reserved, -pincr); 299 KASSERT(prev >= pincr, 300 ("swap_reserved < incr on overcommit fail")); 301 goto out_error; 302 } 303 304 if (!swap_reserve_by_cred_rlimit(pincr, cred, oc)) { 305 prev = atomic_fetchadd_long(&swap_reserved, -pincr); 306 KASSERT(prev >= pincr, 307 ("swap_reserved < incr on overcommit fail")); 308 goto out_error; 309 } 310 311 return (true); 312 313 out_error: 314 if (ppsratecheck(&lastfail, &curfail, 1)) { 315 printf("uid %d, pid %d: swap reservation " 316 "for %jd bytes failed\n", 317 cred->cr_ruidinfo->ui_uid, curproc->p_pid, incr); 318 } 319 #ifdef RACCT 320 if (RACCT_ENABLED()) { 321 PROC_LOCK(curproc); 322 racct_sub(curproc, RACCT_SWAP, incr); 323 PROC_UNLOCK(curproc); 324 } 325 #endif 326 327 return (false); 328 } 329 330 void 331 swap_reserve_force(vm_ooffset_t incr) 332 { 333 u_long pincr; 334 335 KASSERT((incr & PAGE_MASK) == 0, ("%s: incr: %ju & PAGE_MASK", 336 __func__, (uintmax_t)incr)); 337 338 #ifdef RACCT 339 if (RACCT_ENABLED()) { 340 PROC_LOCK(curproc); 341 racct_add_force(curproc, RACCT_SWAP, incr); 342 PROC_UNLOCK(curproc); 343 } 344 #endif 345 pincr = atop(incr); 346 atomic_add_long(&swap_reserved, pincr); 347 swap_reserve_force_rlimit(pincr, curthread->td_ucred); 348 } 349 350 void 351 swap_release(vm_ooffset_t decr) 352 { 353 struct ucred *cred; 354 355 PROC_LOCK(curproc); 356 cred = curproc->p_ucred; 357 swap_release_by_cred(decr, cred); 358 PROC_UNLOCK(curproc); 359 } 360 361 void 362 swap_release_by_cred(vm_ooffset_t decr, struct ucred *cred) 363 { 364 u_long pdecr; 365 #ifdef INVARIANTS 366 u_long prev; 367 #endif 368 369 KASSERT((decr & PAGE_MASK) == 0, ("%s: decr: %ju & PAGE_MASK", 370 __func__, (uintmax_t)decr)); 371 372 pdecr = atop(decr); 373 #ifdef INVARIANTS 374 prev = atomic_fetchadd_long(&swap_reserved, -pdecr); 375 KASSERT(prev >= pdecr, ("swap_reserved < decr")); 376 #else 377 atomic_subtract_long(&swap_reserved, pdecr); 378 #endif 379 380 swap_release_by_cred_rlimit(pdecr, cred); 381 #ifdef RACCT 382 if (racct_enable) 383 racct_sub_cred(cred, RACCT_SWAP, decr); 384 #endif 385 } 386 387 static int swap_pager_full = 2; /* swap space exhaustion (task killing) */ 388 static int swap_pager_almost_full = 1; /* swap space exhaustion (w/hysteresis)*/ 389 static struct mtx swbuf_mtx; /* to sync nsw_wcount_async */ 390 static int nsw_wcount_async; /* limit async write buffers */ 391 static int nsw_wcount_async_max;/* assigned maximum */ 392 int nsw_cluster_max; /* maximum VOP I/O allowed */ 393 394 static int sysctl_swap_async_max(SYSCTL_HANDLER_ARGS); 395 SYSCTL_PROC(_vm, OID_AUTO, swap_async_max, CTLTYPE_INT | CTLFLAG_RW | 396 CTLFLAG_MPSAFE, NULL, 0, sysctl_swap_async_max, "I", 397 "Maximum running async swap ops"); 398 static int sysctl_swap_fragmentation(SYSCTL_HANDLER_ARGS); 399 SYSCTL_PROC(_vm, OID_AUTO, swap_fragmentation, CTLTYPE_STRING | CTLFLAG_RD | 400 CTLFLAG_MPSAFE, NULL, 0, sysctl_swap_fragmentation, "A", 401 "Swap Fragmentation Info"); 402 403 static struct sx sw_alloc_sx; 404 405 /* 406 * "named" and "unnamed" anon region objects. Try to reduce the overhead 407 * of searching a named list by hashing it just a little. 408 */ 409 410 #define NOBJLISTS 8 411 412 #define NOBJLIST(handle) \ 413 (&swap_pager_object_list[((int)(intptr_t)handle >> 4) & (NOBJLISTS-1)]) 414 415 static struct pagerlst swap_pager_object_list[NOBJLISTS]; 416 static uma_zone_t swwbuf_zone; 417 static uma_zone_t swrbuf_zone; 418 static uma_zone_t swblk_zone; 419 static uma_zone_t swpctrie_zone; 420 421 /* 422 * pagerops for OBJT_SWAP - "swap pager". Some ops are also global procedure 423 * calls hooked from other parts of the VM system and do not appear here. 424 * (see vm/swap_pager.h). 425 */ 426 static vm_object_t 427 swap_pager_alloc(void *handle, vm_ooffset_t size, 428 vm_prot_t prot, vm_ooffset_t offset, struct ucred *); 429 static void swap_pager_dealloc(vm_object_t object); 430 static int swap_pager_getpages(vm_object_t, vm_page_t *, int, int *, 431 int *); 432 static int swap_pager_getpages_async(vm_object_t, vm_page_t *, int, int *, 433 int *, pgo_getpages_iodone_t, void *); 434 static void swap_pager_putpages(vm_object_t, vm_page_t *, int, int, int *); 435 static boolean_t 436 swap_pager_haspage(vm_object_t object, vm_pindex_t pindex, int *before, int *after); 437 static void swap_pager_init(void); 438 static void swap_pager_unswapped(vm_page_t); 439 static void swap_pager_swapoff(struct swdevt *sp); 440 static void swap_pager_update_writecount(vm_object_t object, 441 vm_offset_t start, vm_offset_t end); 442 static void swap_pager_release_writecount(vm_object_t object, 443 vm_offset_t start, vm_offset_t end); 444 static void swap_pager_freespace_pgo(vm_object_t object, vm_pindex_t start, 445 vm_size_t size); 446 447 const struct pagerops swappagerops = { 448 .pgo_kvme_type = KVME_TYPE_SWAP, 449 .pgo_init = swap_pager_init, /* early system initialization of pager */ 450 .pgo_alloc = swap_pager_alloc, /* allocate an OBJT_SWAP object */ 451 .pgo_dealloc = swap_pager_dealloc, /* deallocate an OBJT_SWAP object */ 452 .pgo_getpages = swap_pager_getpages, /* pagein */ 453 .pgo_getpages_async = swap_pager_getpages_async, /* pagein (async) */ 454 .pgo_putpages = swap_pager_putpages, /* pageout */ 455 .pgo_haspage = swap_pager_haspage, /* get backing store status for page */ 456 .pgo_pageunswapped = swap_pager_unswapped, /* remove swap related to page */ 457 .pgo_update_writecount = swap_pager_update_writecount, 458 .pgo_release_writecount = swap_pager_release_writecount, 459 .pgo_freespace = swap_pager_freespace_pgo, 460 }; 461 462 /* 463 * swap_*() routines are externally accessible. swp_*() routines are 464 * internal. 465 */ 466 static int nswap_lowat = 128; /* in pages, swap_pager_almost_full warn */ 467 static int nswap_hiwat = 512; /* in pages, swap_pager_almost_full warn */ 468 469 SYSCTL_INT(_vm, OID_AUTO, dmmax, CTLFLAG_RD, &nsw_cluster_max, 0, 470 "Maximum size of a swap block in pages"); 471 472 static void swp_sizecheck(void); 473 static void swp_pager_async_iodone(struct buf *bp); 474 static bool swp_pager_swblk_empty(struct swblk *sb, int start, int limit); 475 static void swp_pager_free_empty_swblk(vm_object_t, struct swblk *sb); 476 static int swapongeom(struct vnode *); 477 static int swaponvp(struct thread *, struct vnode *, u_long); 478 static int swapoff_one(struct swdevt *sp, struct ucred *cred, 479 u_int flags); 480 481 /* 482 * Swap bitmap functions 483 */ 484 static void swp_pager_freeswapspace(const struct page_range *range); 485 static daddr_t swp_pager_getswapspace(int *npages); 486 487 /* 488 * Metadata functions 489 */ 490 static daddr_t swp_pager_meta_build(struct pctrie_iter *, vm_object_t object, 491 vm_pindex_t, daddr_t, bool); 492 static void swp_pager_meta_free(vm_object_t, vm_pindex_t, vm_pindex_t, 493 vm_size_t *); 494 static void swp_pager_meta_transfer(vm_object_t src, vm_object_t dst, 495 vm_pindex_t pindex, vm_pindex_t count); 496 static void swp_pager_meta_free_all(vm_object_t); 497 static daddr_t swp_pager_meta_lookup(struct pctrie_iter *, vm_pindex_t); 498 499 static void 500 swp_pager_init_freerange(struct page_range *range) 501 { 502 range->start = SWAPBLK_NONE; 503 range->num = 0; 504 } 505 506 static void 507 swp_pager_update_freerange(struct page_range *range, daddr_t addr) 508 { 509 if (range->start + range->num == addr) { 510 range->num++; 511 } else { 512 swp_pager_freeswapspace(range); 513 range->start = addr; 514 range->num = 1; 515 } 516 } 517 518 static void * 519 swblk_trie_alloc(struct pctrie *ptree) 520 { 521 522 return (uma_zalloc(swpctrie_zone, M_NOWAIT | (curproc == pageproc ? 523 M_USE_RESERVE : 0))); 524 } 525 526 static void 527 swblk_trie_free(struct pctrie *ptree, void *node) 528 { 529 530 uma_zfree(swpctrie_zone, node); 531 } 532 533 static int 534 swblk_start(struct swblk *sb, vm_pindex_t pindex) 535 { 536 return (sb == NULL || sb->p >= pindex ? 537 0 : pindex % SWAP_META_PAGES); 538 } 539 540 PCTRIE_DEFINE(SWAP, swblk, p, swblk_trie_alloc, swblk_trie_free); 541 542 static struct swblk * 543 swblk_lookup(vm_object_t object, vm_pindex_t pindex) 544 { 545 return (SWAP_PCTRIE_LOOKUP(&object->un_pager.swp.swp_blks, 546 rounddown(pindex, SWAP_META_PAGES))); 547 } 548 549 static void 550 swblk_lookup_remove(vm_object_t object, struct swblk *sb) 551 { 552 SWAP_PCTRIE_REMOVE(&object->un_pager.swp.swp_blks, sb->p); 553 } 554 555 static bool 556 swblk_is_empty(vm_object_t object) 557 { 558 return (pctrie_is_empty(&object->un_pager.swp.swp_blks)); 559 } 560 561 static struct swblk * 562 swblk_iter_lookup_ge(struct pctrie_iter *blks, vm_pindex_t pindex) 563 { 564 return (SWAP_PCTRIE_ITER_LOOKUP_GE(blks, 565 rounddown(pindex, SWAP_META_PAGES))); 566 } 567 568 static void 569 swblk_iter_init_only(struct pctrie_iter *blks, vm_object_t object) 570 { 571 VM_OBJECT_ASSERT_LOCKED(object); 572 MPASS((object->flags & OBJ_SWAP) != 0); 573 pctrie_iter_init(blks, &object->un_pager.swp.swp_blks); 574 } 575 576 577 static struct swblk * 578 swblk_iter_init(struct pctrie_iter *blks, vm_object_t object, 579 vm_pindex_t pindex) 580 { 581 swblk_iter_init_only(blks, object); 582 return (swblk_iter_lookup_ge(blks, pindex)); 583 } 584 585 static struct swblk * 586 swblk_iter_reinit(struct pctrie_iter *blks, vm_object_t object, 587 vm_pindex_t pindex) 588 { 589 swblk_iter_init_only(blks, object); 590 return (SWAP_PCTRIE_ITER_LOOKUP(blks, 591 rounddown(pindex, SWAP_META_PAGES))); 592 } 593 594 static struct swblk * 595 swblk_iter_limit_init(struct pctrie_iter *blks, vm_object_t object, 596 vm_pindex_t pindex, vm_pindex_t limit) 597 { 598 VM_OBJECT_ASSERT_LOCKED(object); 599 MPASS((object->flags & OBJ_SWAP) != 0); 600 pctrie_iter_limit_init(blks, &object->un_pager.swp.swp_blks, limit); 601 return (swblk_iter_lookup_ge(blks, pindex)); 602 } 603 604 static struct swblk * 605 swblk_iter_next(struct pctrie_iter *blks) 606 { 607 return (SWAP_PCTRIE_ITER_JUMP_GE(blks, SWAP_META_PAGES)); 608 } 609 610 static struct swblk * 611 swblk_iter_lookup(struct pctrie_iter *blks, vm_pindex_t pindex) 612 { 613 return (SWAP_PCTRIE_ITER_LOOKUP(blks, 614 rounddown(pindex, SWAP_META_PAGES))); 615 } 616 617 static int 618 swblk_iter_insert(struct pctrie_iter *blks, struct swblk *sb) 619 { 620 return (SWAP_PCTRIE_ITER_INSERT(blks, sb)); 621 } 622 623 static void 624 swblk_iter_remove(struct pctrie_iter *blks) 625 { 626 SWAP_PCTRIE_ITER_REMOVE(blks); 627 } 628 629 /* 630 * SWP_SIZECHECK() - update swap_pager_full indication 631 * 632 * update the swap_pager_almost_full indication and warn when we are 633 * about to run out of swap space, using lowat/hiwat hysteresis. 634 * 635 * Clear swap_pager_full ( task killing ) indication when lowat is met. 636 * 637 * No restrictions on call 638 * This routine may not block. 639 */ 640 static void 641 swp_sizecheck(void) 642 { 643 644 if (swap_pager_avail < nswap_lowat) { 645 if (swap_pager_almost_full == 0) { 646 printf("swap_pager: out of swap space\n"); 647 swap_pager_almost_full = 1; 648 } 649 } else { 650 swap_pager_full = 0; 651 if (swap_pager_avail > nswap_hiwat) 652 swap_pager_almost_full = 0; 653 } 654 } 655 656 /* 657 * SWAP_PAGER_INIT() - initialize the swap pager! 658 * 659 * Expected to be started from system init. NOTE: This code is run 660 * before much else so be careful what you depend on. Most of the VM 661 * system has yet to be initialized at this point. 662 */ 663 static void 664 swap_pager_init(void) 665 { 666 /* 667 * Initialize object lists 668 */ 669 int i; 670 671 for (i = 0; i < NOBJLISTS; ++i) 672 TAILQ_INIT(&swap_pager_object_list[i]); 673 mtx_init(&sw_dev_mtx, "swapdev", NULL, MTX_DEF); 674 sx_init(&sw_alloc_sx, "swspsx"); 675 sx_init(&swdev_syscall_lock, "swsysc"); 676 677 /* 678 * The nsw_cluster_max is constrained by the bp->b_pages[] 679 * array, which has maxphys / PAGE_SIZE entries, and our locally 680 * defined MAX_PAGEOUT_CLUSTER. Also be aware that swap ops are 681 * constrained by the swap device interleave stripe size. 682 * 683 * Initialized early so that GEOM_ELI can see it. 684 */ 685 nsw_cluster_max = min(maxphys / PAGE_SIZE, MAX_PAGEOUT_CLUSTER); 686 } 687 688 /* 689 * SWAP_PAGER_SWAP_INIT() - swap pager initialization from pageout process 690 * 691 * Expected to be started from pageout process once, prior to entering 692 * its main loop. 693 */ 694 void 695 swap_pager_swap_init(void) 696 { 697 unsigned long n, n2; 698 699 /* 700 * Number of in-transit swap bp operations. Don't 701 * exhaust the pbufs completely. Make sure we 702 * initialize workable values (0 will work for hysteresis 703 * but it isn't very efficient). 704 * 705 * Currently we hardwire nsw_wcount_async to 4. This limit is 706 * designed to prevent other I/O from having high latencies due to 707 * our pageout I/O. The value 4 works well for one or two active swap 708 * devices but is probably a little low if you have more. Even so, 709 * a higher value would probably generate only a limited improvement 710 * with three or four active swap devices since the system does not 711 * typically have to pageout at extreme bandwidths. We will want 712 * at least 2 per swap devices, and 4 is a pretty good value if you 713 * have one NFS swap device due to the command/ack latency over NFS. 714 * So it all works out pretty well. 715 * 716 * nsw_cluster_max is initialized in swap_pager_init(). 717 */ 718 719 nsw_wcount_async = 4; 720 nsw_wcount_async_max = nsw_wcount_async; 721 mtx_init(&swbuf_mtx, "async swbuf mutex", NULL, MTX_DEF); 722 723 swwbuf_zone = pbuf_zsecond_create("swwbuf", nswbuf / 4); 724 swrbuf_zone = pbuf_zsecond_create("swrbuf", nswbuf / 2); 725 726 /* 727 * Initialize our zone, taking the user's requested size or 728 * estimating the number we need based on the number of pages 729 * in the system. 730 */ 731 n = maxswzone != 0 ? maxswzone / sizeof(struct swblk) : 732 vm_cnt.v_page_count / 2; 733 swpctrie_zone = uma_zcreate("swpctrie", pctrie_node_size(), NULL, NULL, 734 pctrie_zone_init, NULL, UMA_ALIGN_PTR, 0); 735 swblk_zone = uma_zcreate("swblk", sizeof(struct swblk), NULL, NULL, 736 NULL, NULL, _Alignof(struct swblk) - 1, 0); 737 n2 = n; 738 do { 739 if (uma_zone_reserve_kva(swblk_zone, n)) 740 break; 741 /* 742 * if the allocation failed, try a zone two thirds the 743 * size of the previous attempt. 744 */ 745 n -= ((n + 2) / 3); 746 } while (n > 0); 747 748 /* 749 * Often uma_zone_reserve_kva() cannot reserve exactly the 750 * requested size. Account for the difference when 751 * calculating swap_maxpages. 752 */ 753 n = uma_zone_get_max(swblk_zone); 754 755 if (n < n2) 756 printf("Swap blk zone entries changed from %lu to %lu.\n", 757 n2, n); 758 /* absolute maximum we can handle assuming 100% efficiency */ 759 swap_maxpages = n * SWAP_META_PAGES; 760 swzone = n * sizeof(struct swblk); 761 if (!uma_zone_reserve_kva(swpctrie_zone, n)) 762 printf("Cannot reserve swap pctrie zone, " 763 "reduce kern.maxswzone.\n"); 764 } 765 766 bool 767 swap_pager_init_object(vm_object_t object, void *handle, struct ucred *cred, 768 vm_ooffset_t size, vm_ooffset_t offset) 769 { 770 if (cred != NULL) { 771 if (!swap_reserve_by_cred(size, cred)) 772 return (false); 773 crhold(cred); 774 } 775 776 object->un_pager.swp.writemappings = 0; 777 object->handle = handle; 778 if (cred != NULL) { 779 object->cred = cred; 780 object->charge = size; 781 } 782 return (true); 783 } 784 785 static vm_object_t 786 swap_pager_alloc_init(objtype_t otype, void *handle, struct ucred *cred, 787 vm_ooffset_t size, vm_ooffset_t offset) 788 { 789 vm_object_t object; 790 791 /* 792 * The un_pager.swp.swp_blks trie is initialized by 793 * vm_object_allocate() to ensure the correct order of 794 * visibility to other threads. 795 */ 796 object = vm_object_allocate(otype, OFF_TO_IDX(offset + 797 PAGE_MASK + size)); 798 799 if (!swap_pager_init_object(object, handle, cred, size, offset)) { 800 vm_object_deallocate(object); 801 return (NULL); 802 } 803 return (object); 804 } 805 806 /* 807 * SWAP_PAGER_ALLOC() - allocate a new OBJT_SWAP VM object and instantiate 808 * its metadata structures. 809 * 810 * This routine is called from the mmap and fork code to create a new 811 * OBJT_SWAP object. 812 * 813 * This routine must ensure that no live duplicate is created for 814 * the named object request, which is protected against by 815 * holding the sw_alloc_sx lock in case handle != NULL. 816 */ 817 static vm_object_t 818 swap_pager_alloc(void *handle, vm_ooffset_t size, vm_prot_t prot, 819 vm_ooffset_t offset, struct ucred *cred) 820 { 821 vm_object_t object; 822 823 if (handle != NULL) { 824 /* 825 * Reference existing named region or allocate new one. There 826 * should not be a race here against swp_pager_meta_build() 827 * as called from vm_page_remove() in regards to the lookup 828 * of the handle. 829 */ 830 sx_xlock(&sw_alloc_sx); 831 object = vm_pager_object_lookup(NOBJLIST(handle), handle); 832 if (object == NULL) { 833 object = swap_pager_alloc_init(OBJT_SWAP, handle, cred, 834 size, offset); 835 if (object != NULL) { 836 TAILQ_INSERT_TAIL(NOBJLIST(object->handle), 837 object, pager_object_list); 838 } 839 } 840 sx_xunlock(&sw_alloc_sx); 841 } else { 842 object = swap_pager_alloc_init(OBJT_SWAP, handle, cred, 843 size, offset); 844 } 845 return (object); 846 } 847 848 /* 849 * SWAP_PAGER_DEALLOC() - remove swap metadata from object 850 * 851 * The swap backing for the object is destroyed. The code is 852 * designed such that we can reinstantiate it later, but this 853 * routine is typically called only when the entire object is 854 * about to be destroyed. 855 * 856 * The object must be locked. 857 */ 858 static void 859 swap_pager_dealloc(vm_object_t object) 860 { 861 862 VM_OBJECT_ASSERT_WLOCKED(object); 863 KASSERT((object->flags & OBJ_DEAD) != 0, ("dealloc of reachable obj")); 864 865 /* 866 * Remove from list right away so lookups will fail if we block for 867 * pageout completion. 868 */ 869 if ((object->flags & OBJ_ANON) == 0 && object->handle != NULL) { 870 VM_OBJECT_WUNLOCK(object); 871 sx_xlock(&sw_alloc_sx); 872 TAILQ_REMOVE(NOBJLIST(object->handle), object, 873 pager_object_list); 874 sx_xunlock(&sw_alloc_sx); 875 VM_OBJECT_WLOCK(object); 876 } 877 878 vm_object_pip_wait(object, "swpdea"); 879 880 /* 881 * Free all remaining metadata. We only bother to free it from 882 * the swap meta data. We do not attempt to free swapblk's still 883 * associated with vm_page_t's for this object. We do not care 884 * if paging is still in progress on some objects. 885 */ 886 swp_pager_meta_free_all(object); 887 object->handle = NULL; 888 object->type = OBJT_DEAD; 889 890 /* 891 * Release the allocation charge. 892 */ 893 if (object->cred != NULL) { 894 swap_release_by_cred(object->charge, object->cred); 895 object->charge = 0; 896 crfree(object->cred); 897 object->cred = NULL; 898 } 899 900 /* 901 * Hide the object from swap_pager_swapoff(). 902 */ 903 vm_object_clear_flag(object, OBJ_SWAP); 904 } 905 906 /************************************************************************ 907 * SWAP PAGER BITMAP ROUTINES * 908 ************************************************************************/ 909 910 /* 911 * SWP_PAGER_GETSWAPSPACE() - allocate raw swap space 912 * 913 * Allocate swap for up to the requested number of pages. The 914 * starting swap block number (a page index) is returned or 915 * SWAPBLK_NONE if the allocation failed. 916 * 917 * Also has the side effect of advising that somebody made a mistake 918 * when they configured swap and didn't configure enough. 919 * 920 * This routine may not sleep. 921 * 922 * We allocate in round-robin fashion from the configured devices. 923 */ 924 static daddr_t 925 swp_pager_getswapspace(int *io_npages) 926 { 927 daddr_t blk; 928 struct swdevt *sp; 929 int mpages, npages; 930 931 KASSERT(*io_npages >= 1, 932 ("%s: npages not positive", __func__)); 933 blk = SWAPBLK_NONE; 934 mpages = *io_npages; 935 npages = imin(BLIST_MAX_ALLOC, mpages); 936 mtx_lock(&sw_dev_mtx); 937 sp = swdevhd; 938 while (!TAILQ_EMPTY(&swtailq)) { 939 if (sp == NULL) 940 sp = TAILQ_FIRST(&swtailq); 941 if ((sp->sw_flags & SW_CLOSING) == 0) 942 blk = blist_alloc(sp->sw_blist, &npages, mpages); 943 if (blk != SWAPBLK_NONE) 944 break; 945 sp = TAILQ_NEXT(sp, sw_list); 946 if (swdevhd == sp) { 947 if (npages == 1) 948 break; 949 mpages = npages - 1; 950 npages >>= 1; 951 } 952 } 953 if (blk != SWAPBLK_NONE) { 954 *io_npages = npages; 955 blk += sp->sw_first; 956 sp->sw_used += npages; 957 swap_pager_avail -= npages; 958 swp_sizecheck(); 959 swdevhd = TAILQ_NEXT(sp, sw_list); 960 } else { 961 if (swap_pager_full != 2) { 962 printf("swp_pager_getswapspace(%d): failed\n", 963 *io_npages); 964 swap_pager_full = 2; 965 swap_pager_almost_full = 1; 966 } 967 swdevhd = NULL; 968 } 969 mtx_unlock(&sw_dev_mtx); 970 return (blk); 971 } 972 973 static bool 974 swp_pager_isondev(daddr_t blk, struct swdevt *sp) 975 { 976 977 return (blk >= sp->sw_first && blk < sp->sw_end); 978 } 979 980 static void 981 swp_pager_strategy(struct buf *bp) 982 { 983 struct swdevt *sp; 984 985 mtx_lock(&sw_dev_mtx); 986 TAILQ_FOREACH(sp, &swtailq, sw_list) { 987 if (swp_pager_isondev(bp->b_blkno, sp)) { 988 mtx_unlock(&sw_dev_mtx); 989 if ((sp->sw_flags & SW_UNMAPPED) != 0 && 990 unmapped_buf_allowed) { 991 bp->b_data = unmapped_buf; 992 bp->b_offset = 0; 993 } else { 994 pmap_qenter((vm_offset_t)bp->b_data, 995 &bp->b_pages[0], bp->b_bcount / PAGE_SIZE); 996 } 997 sp->sw_strategy(bp, sp); 998 return; 999 } 1000 } 1001 panic("Swapdev not found"); 1002 } 1003 1004 /* 1005 * SWP_PAGER_FREESWAPSPACE() - free raw swap space 1006 * 1007 * This routine returns the specified swap blocks back to the bitmap. 1008 * 1009 * This routine may not sleep. 1010 */ 1011 static void 1012 swp_pager_freeswapspace(const struct page_range *range) 1013 { 1014 daddr_t blk, npages; 1015 struct swdevt *sp; 1016 1017 blk = range->start; 1018 npages = range->num; 1019 if (npages == 0) 1020 return; 1021 mtx_lock(&sw_dev_mtx); 1022 TAILQ_FOREACH(sp, &swtailq, sw_list) { 1023 if (swp_pager_isondev(blk, sp)) { 1024 sp->sw_used -= npages; 1025 /* 1026 * If we are attempting to stop swapping on 1027 * this device, we don't want to mark any 1028 * blocks free lest they be reused. 1029 */ 1030 if ((sp->sw_flags & SW_CLOSING) == 0) { 1031 blist_free(sp->sw_blist, blk - sp->sw_first, 1032 npages); 1033 swap_pager_avail += npages; 1034 swp_sizecheck(); 1035 } 1036 mtx_unlock(&sw_dev_mtx); 1037 return; 1038 } 1039 } 1040 panic("Swapdev not found"); 1041 } 1042 1043 /* 1044 * SYSCTL_SWAP_FRAGMENTATION() - produce raw swap space stats 1045 */ 1046 static int 1047 sysctl_swap_fragmentation(SYSCTL_HANDLER_ARGS) 1048 { 1049 struct sbuf sbuf; 1050 struct swdevt *sp; 1051 const char *devname; 1052 int error; 1053 1054 error = sysctl_wire_old_buffer(req, 0); 1055 if (error != 0) 1056 return (error); 1057 sbuf_new_for_sysctl(&sbuf, NULL, 128, req); 1058 mtx_lock(&sw_dev_mtx); 1059 TAILQ_FOREACH(sp, &swtailq, sw_list) { 1060 if (vn_isdisk(sp->sw_vp)) 1061 devname = devtoname(sp->sw_vp->v_rdev); 1062 else 1063 devname = "[file]"; 1064 sbuf_printf(&sbuf, "\nFree space on device %s:\n", devname); 1065 blist_stats(sp->sw_blist, &sbuf); 1066 } 1067 mtx_unlock(&sw_dev_mtx); 1068 error = sbuf_finish(&sbuf); 1069 sbuf_delete(&sbuf); 1070 return (error); 1071 } 1072 1073 /* 1074 * SWAP_PAGER_FREESPACE() - frees swap blocks associated with a page 1075 * range within an object. 1076 * 1077 * This routine removes swapblk assignments from swap metadata. 1078 * 1079 * The external callers of this routine typically have already destroyed 1080 * or renamed vm_page_t's associated with this range in the object so 1081 * we should be ok. 1082 * 1083 * The object must be locked. 1084 */ 1085 void 1086 swap_pager_freespace(vm_object_t object, vm_pindex_t start, vm_size_t size, 1087 vm_size_t *freed) 1088 { 1089 MPASS((object->flags & OBJ_SWAP) != 0); 1090 1091 swp_pager_meta_free(object, start, size, freed); 1092 } 1093 1094 static void 1095 swap_pager_freespace_pgo(vm_object_t object, vm_pindex_t start, vm_size_t size) 1096 { 1097 MPASS((object->flags & OBJ_SWAP) != 0); 1098 1099 swp_pager_meta_free(object, start, size, NULL); 1100 } 1101 1102 /* 1103 * SWAP_PAGER_RESERVE() - reserve swap blocks in object 1104 * 1105 * Assigns swap blocks to the specified range within the object. The 1106 * swap blocks are not zeroed. Any previous swap assignment is destroyed. 1107 * 1108 * Returns 0 on success, -1 on failure. 1109 */ 1110 int 1111 swap_pager_reserve(vm_object_t object, vm_pindex_t start, vm_pindex_t size) 1112 { 1113 struct pctrie_iter blks; 1114 struct page_range range; 1115 daddr_t addr, blk; 1116 vm_pindex_t i, j; 1117 int n; 1118 1119 swp_pager_init_freerange(&range); 1120 VM_OBJECT_WLOCK(object); 1121 swblk_iter_init_only(&blks, object); 1122 for (i = 0; i < size; i += n) { 1123 n = MIN(size - i, INT_MAX); 1124 blk = swp_pager_getswapspace(&n); 1125 if (blk == SWAPBLK_NONE) { 1126 swp_pager_meta_free(object, start, i, NULL); 1127 VM_OBJECT_WUNLOCK(object); 1128 return (-1); 1129 } 1130 for (j = 0; j < n; ++j) { 1131 addr = swp_pager_meta_build(&blks, object, 1132 start + i + j, blk + j, false); 1133 if (addr != SWAPBLK_NONE) 1134 swp_pager_update_freerange(&range, addr); 1135 } 1136 } 1137 swp_pager_freeswapspace(&range); 1138 VM_OBJECT_WUNLOCK(object); 1139 return (0); 1140 } 1141 1142 /* 1143 * SWAP_PAGER_COPY() - copy blocks from source pager to destination pager 1144 * and destroy the source. 1145 * 1146 * Copy any valid swapblks from the source to the destination. In 1147 * cases where both the source and destination have a valid swapblk, 1148 * we keep the destination's. 1149 * 1150 * This routine is allowed to sleep. It may sleep allocating metadata 1151 * indirectly through swp_pager_meta_build(). 1152 * 1153 * The source object contains no vm_page_t's (which is just as well) 1154 * 1155 * The source and destination objects must be locked. 1156 * Both object locks may temporarily be released. 1157 */ 1158 void 1159 swap_pager_copy(vm_object_t srcobject, vm_object_t dstobject, 1160 vm_pindex_t offset, int destroysource) 1161 { 1162 VM_OBJECT_ASSERT_WLOCKED(srcobject); 1163 VM_OBJECT_ASSERT_WLOCKED(dstobject); 1164 1165 /* 1166 * If destroysource is set, we remove the source object from the 1167 * swap_pager internal queue now. 1168 */ 1169 if (destroysource && (srcobject->flags & OBJ_ANON) == 0 && 1170 srcobject->handle != NULL) { 1171 VM_OBJECT_WUNLOCK(srcobject); 1172 VM_OBJECT_WUNLOCK(dstobject); 1173 sx_xlock(&sw_alloc_sx); 1174 TAILQ_REMOVE(NOBJLIST(srcobject->handle), srcobject, 1175 pager_object_list); 1176 sx_xunlock(&sw_alloc_sx); 1177 VM_OBJECT_WLOCK(dstobject); 1178 VM_OBJECT_WLOCK(srcobject); 1179 } 1180 1181 /* 1182 * Transfer source to destination. 1183 */ 1184 swp_pager_meta_transfer(srcobject, dstobject, offset, dstobject->size); 1185 1186 /* 1187 * Free left over swap blocks in source. 1188 */ 1189 if (destroysource) 1190 swp_pager_meta_free_all(srcobject); 1191 } 1192 1193 /* 1194 * SWP_PAGER_HASPAGE_ITER() - determine if we have good backing store for 1195 * the requested page, accessed with the given 1196 * iterator. 1197 * 1198 * We determine whether good backing store exists for the requested 1199 * page and return TRUE if it does, FALSE if it doesn't. 1200 * 1201 * If TRUE, we also try to determine how much valid, contiguous backing 1202 * store exists before and after the requested page. 1203 */ 1204 static boolean_t 1205 swp_pager_haspage_iter(struct pctrie_iter *blks, vm_pindex_t pindex, 1206 int *before, int *after) 1207 { 1208 daddr_t blk, blk0; 1209 int i; 1210 1211 /* 1212 * do we have good backing store at the requested index ? 1213 */ 1214 blk0 = swp_pager_meta_lookup(blks, pindex); 1215 if (blk0 == SWAPBLK_NONE) { 1216 if (before) 1217 *before = 0; 1218 if (after) 1219 *after = 0; 1220 return (FALSE); 1221 } 1222 1223 /* 1224 * find backwards-looking contiguous good backing store 1225 */ 1226 if (before != NULL) { 1227 for (i = 1; i < SWB_NPAGES; i++) { 1228 if (i > pindex) 1229 break; 1230 blk = swp_pager_meta_lookup(blks, pindex - i); 1231 if (blk != blk0 - i) 1232 break; 1233 } 1234 *before = i - 1; 1235 } 1236 1237 /* 1238 * find forward-looking contiguous good backing store 1239 */ 1240 if (after != NULL) { 1241 for (i = 1; i < SWB_NPAGES; i++) { 1242 blk = swp_pager_meta_lookup(blks, pindex + i); 1243 if (blk != blk0 + i) 1244 break; 1245 } 1246 *after = i - 1; 1247 } 1248 return (TRUE); 1249 } 1250 1251 /* 1252 * SWAP_PAGER_HASPAGE() - determine if we have good backing store for 1253 * the requested page, in the given object. 1254 * 1255 * We determine whether good backing store exists for the requested 1256 * page and return TRUE if it does, FALSE if it doesn't. 1257 * 1258 * If TRUE, we also try to determine how much valid, contiguous backing 1259 * store exists before and after the requested page. 1260 */ 1261 static boolean_t 1262 swap_pager_haspage(vm_object_t object, vm_pindex_t pindex, int *before, 1263 int *after) 1264 { 1265 struct pctrie_iter blks; 1266 1267 swblk_iter_init_only(&blks, object); 1268 return (swp_pager_haspage_iter(&blks, pindex, before, after)); 1269 } 1270 1271 static void 1272 swap_pager_unswapped_acct(vm_page_t m) 1273 { 1274 KASSERT((m->object->flags & OBJ_SWAP) != 0, 1275 ("Free object not swappable")); 1276 if ((m->a.flags & PGA_SWAP_FREE) != 0) 1277 counter_u64_add(swap_free_completed, 1); 1278 vm_page_aflag_clear(m, PGA_SWAP_FREE | PGA_SWAP_SPACE); 1279 1280 /* 1281 * The meta data only exists if the object is OBJT_SWAP 1282 * and even then might not be allocated yet. 1283 */ 1284 } 1285 1286 /* 1287 * SWAP_PAGER_PAGE_UNSWAPPED() - remove swap backing store related to page 1288 * 1289 * This removes any associated swap backing store, whether valid or 1290 * not, from the page. 1291 * 1292 * This routine is typically called when a page is made dirty, at 1293 * which point any associated swap can be freed. MADV_FREE also 1294 * calls us in a special-case situation 1295 * 1296 * NOTE!!! If the page is clean and the swap was valid, the caller 1297 * should make the page dirty before calling this routine. This routine 1298 * does NOT change the m->dirty status of the page. Also: MADV_FREE 1299 * depends on it. 1300 * 1301 * This routine may not sleep. 1302 * 1303 * The object containing the page may be locked. 1304 */ 1305 static void 1306 swap_pager_unswapped(vm_page_t m) 1307 { 1308 struct page_range range; 1309 struct swblk *sb; 1310 vm_object_t obj; 1311 1312 /* 1313 * Handle enqueing deferred frees first. If we do not have the 1314 * object lock we wait for the page daemon to clear the space. 1315 */ 1316 obj = m->object; 1317 if (!VM_OBJECT_WOWNED(obj)) { 1318 VM_PAGE_OBJECT_BUSY_ASSERT(m); 1319 /* 1320 * The caller is responsible for synchronization but we 1321 * will harmlessly handle races. This is typically provided 1322 * by only calling unswapped() when a page transitions from 1323 * clean to dirty. 1324 */ 1325 if ((m->a.flags & (PGA_SWAP_SPACE | PGA_SWAP_FREE)) == 1326 PGA_SWAP_SPACE) { 1327 vm_page_aflag_set(m, PGA_SWAP_FREE); 1328 counter_u64_add(swap_free_deferred, 1); 1329 } 1330 return; 1331 } 1332 swap_pager_unswapped_acct(m); 1333 1334 sb = swblk_lookup(m->object, m->pindex); 1335 if (sb == NULL) 1336 return; 1337 range.start = sb->d[m->pindex % SWAP_META_PAGES]; 1338 if (range.start == SWAPBLK_NONE) 1339 return; 1340 range.num = 1; 1341 swp_pager_freeswapspace(&range); 1342 sb->d[m->pindex % SWAP_META_PAGES] = SWAPBLK_NONE; 1343 swp_pager_free_empty_swblk(m->object, sb); 1344 } 1345 1346 /* 1347 * swap_pager_getpages() - bring pages in from swap 1348 * 1349 * Attempt to page in the pages in array "ma" of length "count". The 1350 * caller may optionally specify that additional pages preceding and 1351 * succeeding the specified range be paged in. The number of such pages 1352 * is returned in the "rbehind" and "rahead" parameters, and they will 1353 * be in the inactive queue upon return. 1354 * 1355 * The pages in "ma" must be busied and will remain busied upon return. 1356 */ 1357 static int 1358 swap_pager_getpages_locked(struct pctrie_iter *blks, vm_object_t object, 1359 vm_page_t *ma, int count, int *rbehind, int *rahead) 1360 { 1361 struct buf *bp; 1362 vm_page_t bm, mpred, msucc, p; 1363 vm_pindex_t pindex; 1364 daddr_t blk; 1365 int i, maxahead, maxbehind, reqcount; 1366 1367 VM_OBJECT_ASSERT_WLOCKED(object); 1368 reqcount = count; 1369 1370 KASSERT((object->flags & OBJ_SWAP) != 0, 1371 ("%s: object not swappable", __func__)); 1372 if (!swp_pager_haspage_iter(blks, ma[0]->pindex, &maxbehind, 1373 &maxahead)) { 1374 VM_OBJECT_WUNLOCK(object); 1375 return (VM_PAGER_FAIL); 1376 } 1377 1378 KASSERT(reqcount - 1 <= maxahead, 1379 ("page count %d extends beyond swap block", reqcount)); 1380 1381 /* 1382 * Do not transfer any pages other than those that are xbusied 1383 * when running during a split or collapse operation. This 1384 * prevents clustering from re-creating pages which are being 1385 * moved into another object. 1386 */ 1387 if ((object->flags & (OBJ_SPLIT | OBJ_DEAD)) != 0) { 1388 maxahead = reqcount - 1; 1389 maxbehind = 0; 1390 } 1391 1392 /* 1393 * Clip the readahead and readbehind ranges to exclude resident pages. 1394 */ 1395 if (rahead != NULL) { 1396 *rahead = imin(*rahead, maxahead - (reqcount - 1)); 1397 pindex = ma[reqcount - 1]->pindex; 1398 msucc = TAILQ_NEXT(ma[reqcount - 1], listq); 1399 if (msucc != NULL && msucc->pindex - pindex - 1 < *rahead) 1400 *rahead = msucc->pindex - pindex - 1; 1401 } 1402 if (rbehind != NULL) { 1403 *rbehind = imin(*rbehind, maxbehind); 1404 pindex = ma[0]->pindex; 1405 mpred = TAILQ_PREV(ma[0], pglist, listq); 1406 if (mpred != NULL && pindex - mpred->pindex - 1 < *rbehind) 1407 *rbehind = pindex - mpred->pindex - 1; 1408 } 1409 1410 bm = ma[0]; 1411 for (i = 0; i < count; i++) 1412 ma[i]->oflags |= VPO_SWAPINPROG; 1413 1414 /* 1415 * Allocate readahead and readbehind pages. 1416 */ 1417 if (rbehind != NULL) { 1418 for (i = 1; i <= *rbehind; i++) { 1419 p = vm_page_alloc(object, ma[0]->pindex - i, 1420 VM_ALLOC_NORMAL); 1421 if (p == NULL) 1422 break; 1423 p->oflags |= VPO_SWAPINPROG; 1424 bm = p; 1425 } 1426 *rbehind = i - 1; 1427 } 1428 if (rahead != NULL) { 1429 for (i = 0; i < *rahead; i++) { 1430 p = vm_page_alloc(object, 1431 ma[reqcount - 1]->pindex + i + 1, VM_ALLOC_NORMAL); 1432 if (p == NULL) 1433 break; 1434 p->oflags |= VPO_SWAPINPROG; 1435 } 1436 *rahead = i; 1437 } 1438 if (rbehind != NULL) 1439 count += *rbehind; 1440 if (rahead != NULL) 1441 count += *rahead; 1442 1443 vm_object_pip_add(object, count); 1444 1445 pindex = bm->pindex; 1446 blk = swp_pager_meta_lookup(blks, pindex); 1447 KASSERT(blk != SWAPBLK_NONE, 1448 ("no swap blocking containing %p(%jx)", object, (uintmax_t)pindex)); 1449 1450 VM_OBJECT_WUNLOCK(object); 1451 bp = uma_zalloc(swrbuf_zone, M_WAITOK); 1452 MPASS((bp->b_flags & B_MAXPHYS) != 0); 1453 /* Pages cannot leave the object while busy. */ 1454 for (i = 0, p = bm; i < count; i++, p = TAILQ_NEXT(p, listq)) { 1455 MPASS(p->pindex == bm->pindex + i); 1456 bp->b_pages[i] = p; 1457 } 1458 1459 bp->b_flags |= B_PAGING; 1460 bp->b_iocmd = BIO_READ; 1461 bp->b_iodone = swp_pager_async_iodone; 1462 bp->b_rcred = crhold(thread0.td_ucred); 1463 bp->b_wcred = crhold(thread0.td_ucred); 1464 bp->b_blkno = blk; 1465 bp->b_bcount = PAGE_SIZE * count; 1466 bp->b_bufsize = PAGE_SIZE * count; 1467 bp->b_npages = count; 1468 bp->b_pgbefore = rbehind != NULL ? *rbehind : 0; 1469 bp->b_pgafter = rahead != NULL ? *rahead : 0; 1470 1471 VM_CNT_INC(v_swapin); 1472 VM_CNT_ADD(v_swappgsin, count); 1473 1474 /* 1475 * perform the I/O. NOTE!!! bp cannot be considered valid after 1476 * this point because we automatically release it on completion. 1477 * Instead, we look at the one page we are interested in which we 1478 * still hold a lock on even through the I/O completion. 1479 * 1480 * The other pages in our ma[] array are also released on completion, 1481 * so we cannot assume they are valid anymore either. 1482 * 1483 * NOTE: b_blkno is destroyed by the call to swapdev_strategy 1484 */ 1485 BUF_KERNPROC(bp); 1486 swp_pager_strategy(bp); 1487 1488 /* 1489 * Wait for the pages we want to complete. VPO_SWAPINPROG is always 1490 * cleared on completion. If an I/O error occurs, SWAPBLK_NONE 1491 * is set in the metadata for each page in the request. 1492 */ 1493 VM_OBJECT_WLOCK(object); 1494 /* This could be implemented more efficiently with aflags */ 1495 while ((ma[0]->oflags & VPO_SWAPINPROG) != 0) { 1496 ma[0]->oflags |= VPO_SWAPSLEEP; 1497 VM_CNT_INC(v_intrans); 1498 if (VM_OBJECT_SLEEP(object, &object->handle, PSWP, 1499 "swread", hz * 20)) { 1500 printf( 1501 "swap_pager: indefinite wait buffer: bufobj: %p, blkno: %jd, size: %ld\n", 1502 bp->b_bufobj, (intmax_t)bp->b_blkno, bp->b_bcount); 1503 } 1504 } 1505 VM_OBJECT_WUNLOCK(object); 1506 1507 /* 1508 * If we had an unrecoverable read error pages will not be valid. 1509 */ 1510 for (i = 0; i < reqcount; i++) 1511 if (ma[i]->valid != VM_PAGE_BITS_ALL) 1512 return (VM_PAGER_ERROR); 1513 1514 return (VM_PAGER_OK); 1515 1516 /* 1517 * A final note: in a low swap situation, we cannot deallocate swap 1518 * and mark a page dirty here because the caller is likely to mark 1519 * the page clean when we return, causing the page to possibly revert 1520 * to all-zero's later. 1521 */ 1522 } 1523 1524 static int 1525 swap_pager_getpages(vm_object_t object, vm_page_t *ma, int count, 1526 int *rbehind, int *rahead) 1527 { 1528 struct pctrie_iter blks; 1529 1530 VM_OBJECT_WLOCK(object); 1531 swblk_iter_init_only(&blks, object); 1532 return (swap_pager_getpages_locked(&blks, object, ma, count, rbehind, 1533 rahead)); 1534 } 1535 1536 /* 1537 * swap_pager_getpages_async(): 1538 * 1539 * Right now this is emulation of asynchronous operation on top of 1540 * swap_pager_getpages(). 1541 */ 1542 static int 1543 swap_pager_getpages_async(vm_object_t object, vm_page_t *ma, int count, 1544 int *rbehind, int *rahead, pgo_getpages_iodone_t iodone, void *arg) 1545 { 1546 int r, error; 1547 1548 r = swap_pager_getpages(object, ma, count, rbehind, rahead); 1549 switch (r) { 1550 case VM_PAGER_OK: 1551 error = 0; 1552 break; 1553 case VM_PAGER_ERROR: 1554 error = EIO; 1555 break; 1556 case VM_PAGER_FAIL: 1557 error = EINVAL; 1558 break; 1559 default: 1560 panic("unhandled swap_pager_getpages() error %d", r); 1561 } 1562 (iodone)(arg, ma, count, error); 1563 1564 return (r); 1565 } 1566 1567 /* 1568 * swap_pager_putpages: 1569 * 1570 * Assign swap (if necessary) and initiate I/O on the specified pages. 1571 * 1572 * In a low memory situation we may block in VOP_STRATEGY(), but the new 1573 * vm_page reservation system coupled with properly written VFS devices 1574 * should ensure that no low-memory deadlock occurs. This is an area 1575 * which needs work. 1576 * 1577 * The parent has N vm_object_pip_add() references prior to 1578 * calling us and will remove references for rtvals[] that are 1579 * not set to VM_PAGER_PEND. We need to remove the rest on I/O 1580 * completion. 1581 * 1582 * The parent has soft-busy'd the pages it passes us and will unbusy 1583 * those whose rtvals[] entry is not set to VM_PAGER_PEND on return. 1584 * We need to unbusy the rest on I/O completion. 1585 */ 1586 static void 1587 swap_pager_putpages(vm_object_t object, vm_page_t *ma, int count, 1588 int flags, int *rtvals) 1589 { 1590 struct pctrie_iter blks; 1591 struct page_range range; 1592 struct buf *bp; 1593 daddr_t addr, blk; 1594 vm_page_t mreq; 1595 int i, j, n; 1596 bool async; 1597 1598 KASSERT(count == 0 || ma[0]->object == object, 1599 ("%s: object mismatch %p/%p", 1600 __func__, object, ma[0]->object)); 1601 1602 VM_OBJECT_WUNLOCK(object); 1603 async = curproc == pageproc && (flags & VM_PAGER_PUT_SYNC) == 0; 1604 swp_pager_init_freerange(&range); 1605 1606 /* 1607 * Assign swap blocks and issue I/O. We reallocate swap on the fly. 1608 * The page is left dirty until the pageout operation completes 1609 * successfully. 1610 */ 1611 for (i = 0; i < count; i += n) { 1612 /* Maximum I/O size is limited by maximum swap block size. */ 1613 n = min(count - i, nsw_cluster_max); 1614 1615 if (async) { 1616 mtx_lock(&swbuf_mtx); 1617 while (nsw_wcount_async == 0) 1618 msleep(&nsw_wcount_async, &swbuf_mtx, PVM, 1619 "swbufa", 0); 1620 nsw_wcount_async--; 1621 mtx_unlock(&swbuf_mtx); 1622 } 1623 1624 /* Get a block of swap of size up to size n. */ 1625 blk = swp_pager_getswapspace(&n); 1626 if (blk == SWAPBLK_NONE) { 1627 mtx_lock(&swbuf_mtx); 1628 if (++nsw_wcount_async == 1) 1629 wakeup(&nsw_wcount_async); 1630 mtx_unlock(&swbuf_mtx); 1631 for (j = 0; j < n; ++j) 1632 rtvals[i + j] = VM_PAGER_FAIL; 1633 continue; 1634 } 1635 VM_OBJECT_WLOCK(object); 1636 swblk_iter_init_only(&blks, object); 1637 for (j = 0; j < n; ++j) { 1638 mreq = ma[i + j]; 1639 vm_page_aflag_clear(mreq, PGA_SWAP_FREE); 1640 KASSERT(mreq->object == object, 1641 ("%s: object mismatch %p/%p", 1642 __func__, mreq->object, object)); 1643 addr = swp_pager_meta_build(&blks, object, 1644 mreq->pindex, blk + j, false); 1645 if (addr != SWAPBLK_NONE) 1646 swp_pager_update_freerange(&range, addr); 1647 MPASS(mreq->dirty == VM_PAGE_BITS_ALL); 1648 mreq->oflags |= VPO_SWAPINPROG; 1649 } 1650 VM_OBJECT_WUNLOCK(object); 1651 1652 bp = uma_zalloc(swwbuf_zone, M_WAITOK); 1653 MPASS((bp->b_flags & B_MAXPHYS) != 0); 1654 if (async) 1655 bp->b_flags |= B_ASYNC; 1656 bp->b_flags |= B_PAGING; 1657 bp->b_iocmd = BIO_WRITE; 1658 1659 bp->b_rcred = crhold(thread0.td_ucred); 1660 bp->b_wcred = crhold(thread0.td_ucred); 1661 bp->b_bcount = PAGE_SIZE * n; 1662 bp->b_bufsize = PAGE_SIZE * n; 1663 bp->b_blkno = blk; 1664 for (j = 0; j < n; j++) 1665 bp->b_pages[j] = ma[i + j]; 1666 bp->b_npages = n; 1667 1668 /* 1669 * Must set dirty range for NFS to work. 1670 */ 1671 bp->b_dirtyoff = 0; 1672 bp->b_dirtyend = bp->b_bcount; 1673 1674 VM_CNT_INC(v_swapout); 1675 VM_CNT_ADD(v_swappgsout, bp->b_npages); 1676 1677 /* 1678 * We unconditionally set rtvals[] to VM_PAGER_PEND so that we 1679 * can call the async completion routine at the end of a 1680 * synchronous I/O operation. Otherwise, our caller would 1681 * perform duplicate unbusy and wakeup operations on the page 1682 * and object, respectively. 1683 */ 1684 for (j = 0; j < n; j++) 1685 rtvals[i + j] = VM_PAGER_PEND; 1686 1687 /* 1688 * asynchronous 1689 * 1690 * NOTE: b_blkno is destroyed by the call to swapdev_strategy. 1691 */ 1692 if (async) { 1693 bp->b_iodone = swp_pager_async_iodone; 1694 BUF_KERNPROC(bp); 1695 swp_pager_strategy(bp); 1696 continue; 1697 } 1698 1699 /* 1700 * synchronous 1701 * 1702 * NOTE: b_blkno is destroyed by the call to swapdev_strategy. 1703 */ 1704 bp->b_iodone = bdone; 1705 swp_pager_strategy(bp); 1706 1707 /* 1708 * Wait for the sync I/O to complete. 1709 */ 1710 bwait(bp, PVM, "swwrt"); 1711 1712 /* 1713 * Now that we are through with the bp, we can call the 1714 * normal async completion, which frees everything up. 1715 */ 1716 swp_pager_async_iodone(bp); 1717 } 1718 swp_pager_freeswapspace(&range); 1719 VM_OBJECT_WLOCK(object); 1720 } 1721 1722 /* 1723 * swp_pager_async_iodone: 1724 * 1725 * Completion routine for asynchronous reads and writes from/to swap. 1726 * Also called manually by synchronous code to finish up a bp. 1727 * 1728 * This routine may not sleep. 1729 */ 1730 static void 1731 swp_pager_async_iodone(struct buf *bp) 1732 { 1733 int i; 1734 vm_object_t object = NULL; 1735 1736 /* 1737 * Report error - unless we ran out of memory, in which case 1738 * we've already logged it in swapgeom_strategy(). 1739 */ 1740 if (bp->b_ioflags & BIO_ERROR && bp->b_error != ENOMEM) { 1741 printf( 1742 "swap_pager: I/O error - %s failed; blkno %ld," 1743 "size %ld, error %d\n", 1744 ((bp->b_iocmd == BIO_READ) ? "pagein" : "pageout"), 1745 (long)bp->b_blkno, 1746 (long)bp->b_bcount, 1747 bp->b_error 1748 ); 1749 } 1750 1751 /* 1752 * remove the mapping for kernel virtual 1753 */ 1754 if (buf_mapped(bp)) 1755 pmap_qremove((vm_offset_t)bp->b_data, bp->b_npages); 1756 else 1757 bp->b_data = bp->b_kvabase; 1758 1759 if (bp->b_npages) { 1760 object = bp->b_pages[0]->object; 1761 VM_OBJECT_WLOCK(object); 1762 } 1763 1764 /* 1765 * cleanup pages. If an error occurs writing to swap, we are in 1766 * very serious trouble. If it happens to be a disk error, though, 1767 * we may be able to recover by reassigning the swap later on. So 1768 * in this case we remove the m->swapblk assignment for the page 1769 * but do not free it in the rlist. The errornous block(s) are thus 1770 * never reallocated as swap. Redirty the page and continue. 1771 */ 1772 for (i = 0; i < bp->b_npages; ++i) { 1773 vm_page_t m = bp->b_pages[i]; 1774 1775 m->oflags &= ~VPO_SWAPINPROG; 1776 if (m->oflags & VPO_SWAPSLEEP) { 1777 m->oflags &= ~VPO_SWAPSLEEP; 1778 wakeup(&object->handle); 1779 } 1780 1781 /* We always have space after I/O, successful or not. */ 1782 vm_page_aflag_set(m, PGA_SWAP_SPACE); 1783 1784 if (bp->b_ioflags & BIO_ERROR) { 1785 /* 1786 * If an error occurs I'd love to throw the swapblk 1787 * away without freeing it back to swapspace, so it 1788 * can never be used again. But I can't from an 1789 * interrupt. 1790 */ 1791 if (bp->b_iocmd == BIO_READ) { 1792 /* 1793 * NOTE: for reads, m->dirty will probably 1794 * be overridden by the original caller of 1795 * getpages so don't play cute tricks here. 1796 */ 1797 vm_page_invalid(m); 1798 if (i < bp->b_pgbefore || 1799 i >= bp->b_npages - bp->b_pgafter) 1800 vm_page_free_invalid(m); 1801 } else { 1802 /* 1803 * If a write error occurs, reactivate page 1804 * so it doesn't clog the inactive list, 1805 * then finish the I/O. 1806 */ 1807 MPASS(m->dirty == VM_PAGE_BITS_ALL); 1808 1809 /* PQ_UNSWAPPABLE? */ 1810 vm_page_activate(m); 1811 vm_page_sunbusy(m); 1812 } 1813 } else if (bp->b_iocmd == BIO_READ) { 1814 /* 1815 * NOTE: for reads, m->dirty will probably be 1816 * overridden by the original caller of getpages so 1817 * we cannot set them in order to free the underlying 1818 * swap in a low-swap situation. I don't think we'd 1819 * want to do that anyway, but it was an optimization 1820 * that existed in the old swapper for a time before 1821 * it got ripped out due to precisely this problem. 1822 */ 1823 KASSERT(!pmap_page_is_mapped(m), 1824 ("swp_pager_async_iodone: page %p is mapped", m)); 1825 KASSERT(m->dirty == 0, 1826 ("swp_pager_async_iodone: page %p is dirty", m)); 1827 1828 vm_page_valid(m); 1829 if (i < bp->b_pgbefore || 1830 i >= bp->b_npages - bp->b_pgafter) 1831 vm_page_readahead_finish(m); 1832 } else { 1833 /* 1834 * For write success, clear the dirty 1835 * status, then finish the I/O ( which decrements the 1836 * busy count and possibly wakes waiter's up ). 1837 * A page is only written to swap after a period of 1838 * inactivity. Therefore, we do not expect it to be 1839 * reused. 1840 */ 1841 KASSERT(!pmap_page_is_write_mapped(m), 1842 ("swp_pager_async_iodone: page %p is not write" 1843 " protected", m)); 1844 vm_page_undirty(m); 1845 vm_page_deactivate_noreuse(m); 1846 vm_page_sunbusy(m); 1847 } 1848 } 1849 1850 /* 1851 * adjust pip. NOTE: the original parent may still have its own 1852 * pip refs on the object. 1853 */ 1854 if (object != NULL) { 1855 vm_object_pip_wakeupn(object, bp->b_npages); 1856 VM_OBJECT_WUNLOCK(object); 1857 } 1858 1859 /* 1860 * swapdev_strategy() manually sets b_vp and b_bufobj before calling 1861 * bstrategy(). Set them back to NULL now we're done with it, or we'll 1862 * trigger a KASSERT in relpbuf(). 1863 */ 1864 if (bp->b_vp) { 1865 bp->b_vp = NULL; 1866 bp->b_bufobj = NULL; 1867 } 1868 /* 1869 * release the physical I/O buffer 1870 */ 1871 if (bp->b_flags & B_ASYNC) { 1872 mtx_lock(&swbuf_mtx); 1873 if (++nsw_wcount_async == 1) 1874 wakeup(&nsw_wcount_async); 1875 mtx_unlock(&swbuf_mtx); 1876 } 1877 uma_zfree((bp->b_iocmd == BIO_READ) ? swrbuf_zone : swwbuf_zone, bp); 1878 } 1879 1880 int 1881 swap_pager_nswapdev(void) 1882 { 1883 1884 return (nswapdev); 1885 } 1886 1887 static void 1888 swp_pager_force_dirty(struct page_range *range, vm_page_t m, daddr_t *blk) 1889 { 1890 vm_page_dirty(m); 1891 swap_pager_unswapped_acct(m); 1892 swp_pager_update_freerange(range, *blk); 1893 *blk = SWAPBLK_NONE; 1894 vm_page_launder(m); 1895 } 1896 1897 u_long 1898 swap_pager_swapped_pages(vm_object_t object) 1899 { 1900 struct pctrie_iter blks; 1901 struct swblk *sb; 1902 u_long res; 1903 int i; 1904 1905 VM_OBJECT_ASSERT_LOCKED(object); 1906 1907 if (swblk_is_empty(object)) 1908 return (0); 1909 1910 res = 0; 1911 for (sb = swblk_iter_init(&blks, object, 0); sb != NULL; 1912 sb = swblk_iter_next(&blks)) { 1913 for (i = 0; i < SWAP_META_PAGES; i++) { 1914 if (sb->d[i] != SWAPBLK_NONE) 1915 res++; 1916 } 1917 } 1918 return (res); 1919 } 1920 1921 /* 1922 * swap_pager_swapoff_object: 1923 * 1924 * Page in all of the pages that have been paged out for an object 1925 * to a swap device. 1926 */ 1927 static void 1928 swap_pager_swapoff_object(struct swdevt *sp, vm_object_t object) 1929 { 1930 struct pctrie_iter blks, pages; 1931 struct page_range range; 1932 struct swblk *sb; 1933 vm_page_t m; 1934 int i, rahead, rv; 1935 bool sb_empty; 1936 1937 VM_OBJECT_ASSERT_WLOCKED(object); 1938 KASSERT((object->flags & OBJ_SWAP) != 0, 1939 ("%s: Object not swappable", __func__)); 1940 KASSERT((object->flags & OBJ_DEAD) == 0, 1941 ("%s: Object already dead", __func__)); 1942 KASSERT((sp->sw_flags & SW_CLOSING) != 0, 1943 ("%s: Device not blocking further allocations", __func__)); 1944 1945 vm_page_iter_init(&pages, object); 1946 swp_pager_init_freerange(&range); 1947 sb = swblk_iter_init(&blks, object, 0); 1948 while (sb != NULL) { 1949 sb_empty = true; 1950 for (i = 0; i < SWAP_META_PAGES; i++) { 1951 /* Skip an invalid block. */ 1952 if (sb->d[i] == SWAPBLK_NONE) 1953 continue; 1954 /* Skip a block not of this device. */ 1955 if (!swp_pager_isondev(sb->d[i], sp)) { 1956 sb_empty = false; 1957 continue; 1958 } 1959 1960 /* 1961 * Look for a page corresponding to this block. If the 1962 * found page has pending operations, sleep and restart 1963 * the scan. 1964 */ 1965 m = vm_page_iter_lookup(&pages, blks.index + i); 1966 if (m != NULL && (m->oflags & VPO_SWAPINPROG) != 0) { 1967 m->oflags |= VPO_SWAPSLEEP; 1968 VM_OBJECT_SLEEP(object, &object->handle, PSWP, 1969 "swpoff", 0); 1970 break; 1971 } 1972 1973 /* 1974 * If the found page is valid, mark it dirty and free 1975 * the swap block. 1976 */ 1977 if (m != NULL && vm_page_all_valid(m)) { 1978 swp_pager_force_dirty(&range, m, &sb->d[i]); 1979 continue; 1980 } 1981 /* Is there a page we can acquire or allocate? */ 1982 if (m != NULL) { 1983 if (!vm_page_busy_acquire(m, VM_ALLOC_WAITFAIL)) 1984 break; 1985 } else if ((m = vm_page_alloc(object, blks.index + i, 1986 VM_ALLOC_NORMAL | VM_ALLOC_WAITFAIL)) == NULL) 1987 break; 1988 1989 /* Get the page from swap, and restart the scan. */ 1990 vm_object_pip_add(object, 1); 1991 rahead = SWAP_META_PAGES; 1992 rv = swap_pager_getpages_locked(&blks, object, &m, 1, 1993 NULL, &rahead); 1994 if (rv != VM_PAGER_OK) 1995 panic("%s: read from swap failed: %d", 1996 __func__, rv); 1997 VM_OBJECT_WLOCK(object); 1998 vm_object_pip_wakeupn(object, 1); 1999 KASSERT(vm_page_all_valid(m), 2000 ("%s: Page %p not all valid", __func__, m)); 2001 vm_page_deactivate_noreuse(m); 2002 vm_page_xunbusy(m); 2003 break; 2004 } 2005 if (i < SWAP_META_PAGES) { 2006 /* 2007 * The object lock has been released and regained. 2008 * Perhaps the object is now dead. 2009 */ 2010 if ((object->flags & OBJ_DEAD) != 0) { 2011 /* 2012 * Make sure that pending writes finish before 2013 * returning. 2014 */ 2015 vm_object_pip_wait(object, "swpoff"); 2016 swp_pager_meta_free_all(object); 2017 break; 2018 } 2019 2020 /* 2021 * The swapblk could have been freed, so reset the pages 2022 * iterator and search again for the first swblk at or 2023 * after blks.index. 2024 */ 2025 pctrie_iter_reset(&pages); 2026 sb = swblk_iter_init(&blks, object, blks.index); 2027 continue; 2028 } 2029 if (sb_empty) { 2030 swblk_iter_remove(&blks); 2031 uma_zfree(swblk_zone, sb); 2032 } 2033 2034 /* 2035 * It is safe to advance to the next block. No allocations 2036 * before blk.index have happened, even with the lock released, 2037 * because allocations on this device are blocked. 2038 */ 2039 sb = swblk_iter_next(&blks); 2040 } 2041 swp_pager_freeswapspace(&range); 2042 } 2043 2044 /* 2045 * swap_pager_swapoff: 2046 * 2047 * Page in all of the pages that have been paged out to the 2048 * given device. The corresponding blocks in the bitmap must be 2049 * marked as allocated and the device must be flagged SW_CLOSING. 2050 * There may be no processes swapped out to the device. 2051 * 2052 * This routine may block. 2053 */ 2054 static void 2055 swap_pager_swapoff(struct swdevt *sp) 2056 { 2057 vm_object_t object; 2058 int retries; 2059 2060 sx_assert(&swdev_syscall_lock, SA_XLOCKED); 2061 2062 retries = 0; 2063 full_rescan: 2064 mtx_lock(&vm_object_list_mtx); 2065 TAILQ_FOREACH(object, &vm_object_list, object_list) { 2066 if ((object->flags & OBJ_SWAP) == 0) 2067 continue; 2068 mtx_unlock(&vm_object_list_mtx); 2069 /* Depends on type-stability. */ 2070 VM_OBJECT_WLOCK(object); 2071 2072 /* 2073 * Dead objects are eventually terminated on their own. 2074 */ 2075 if ((object->flags & OBJ_DEAD) != 0) 2076 goto next_obj; 2077 2078 /* 2079 * Sync with fences placed after pctrie 2080 * initialization. We must not access pctrie below 2081 * unless we checked that our object is swap and not 2082 * dead. 2083 */ 2084 atomic_thread_fence_acq(); 2085 if ((object->flags & OBJ_SWAP) == 0) 2086 goto next_obj; 2087 2088 swap_pager_swapoff_object(sp, object); 2089 next_obj: 2090 VM_OBJECT_WUNLOCK(object); 2091 mtx_lock(&vm_object_list_mtx); 2092 } 2093 mtx_unlock(&vm_object_list_mtx); 2094 2095 if (sp->sw_used) { 2096 /* 2097 * Objects may be locked or paging to the device being 2098 * removed, so we will miss their pages and need to 2099 * make another pass. We have marked this device as 2100 * SW_CLOSING, so the activity should finish soon. 2101 */ 2102 retries++; 2103 if (retries > 100) { 2104 panic("swapoff: failed to locate %d swap blocks", 2105 sp->sw_used); 2106 } 2107 pause("swpoff", hz / 20); 2108 goto full_rescan; 2109 } 2110 EVENTHANDLER_INVOKE(swapoff, sp); 2111 } 2112 2113 /************************************************************************ 2114 * SWAP META DATA * 2115 ************************************************************************ 2116 * 2117 * These routines manipulate the swap metadata stored in the 2118 * OBJT_SWAP object. 2119 * 2120 * Swap metadata is implemented with a global hash and not directly 2121 * linked into the object. Instead the object simply contains 2122 * appropriate tracking counters. 2123 */ 2124 2125 /* 2126 * SWP_PAGER_SWBLK_EMPTY() - is a range of blocks free? 2127 */ 2128 static bool 2129 swp_pager_swblk_empty(struct swblk *sb, int start, int limit) 2130 { 2131 int i; 2132 2133 MPASS(0 <= start && start <= limit && limit <= SWAP_META_PAGES); 2134 for (i = start; i < limit; i++) { 2135 if (sb->d[i] != SWAPBLK_NONE) 2136 return (false); 2137 } 2138 return (true); 2139 } 2140 2141 /* 2142 * SWP_PAGER_FREE_EMPTY_SWBLK() - frees if a block is free 2143 * 2144 * Nothing is done if the block is still in use. 2145 */ 2146 static void 2147 swp_pager_free_empty_swblk(vm_object_t object, struct swblk *sb) 2148 { 2149 2150 if (swp_pager_swblk_empty(sb, 0, SWAP_META_PAGES)) { 2151 swblk_lookup_remove(object, sb); 2152 uma_zfree(swblk_zone, sb); 2153 } 2154 } 2155 2156 /* 2157 * SWP_PAGER_META_BUILD() - add swap block to swap meta data for object 2158 * 2159 * Try to add the specified swapblk to the object's swap metadata. If 2160 * nowait_noreplace is set, add the specified swapblk only if there is no 2161 * previously assigned swapblk at pindex. If the swapblk is invalid, and 2162 * replaces a valid swapblk, empty swap metadata is freed. If memory 2163 * allocation fails, and nowait_noreplace is set, return the specified 2164 * swapblk immediately to indicate failure; otherwise, wait and retry until 2165 * memory allocation succeeds. Return the previously assigned swapblk, if 2166 * any. 2167 */ 2168 static daddr_t 2169 swp_pager_meta_build(struct pctrie_iter *blks, vm_object_t object, 2170 vm_pindex_t pindex, daddr_t swapblk, bool nowait_noreplace) 2171 { 2172 static volatile int swblk_zone_exhausted, swpctrie_zone_exhausted; 2173 struct swblk *sb, *sb1; 2174 vm_pindex_t modpi; 2175 daddr_t prev_swapblk; 2176 int error, i; 2177 2178 VM_OBJECT_ASSERT_WLOCKED(object); 2179 2180 sb = swblk_iter_lookup(blks, pindex); 2181 if (sb == NULL) { 2182 if (swapblk == SWAPBLK_NONE) 2183 return (SWAPBLK_NONE); 2184 for (;;) { 2185 sb = uma_zalloc(swblk_zone, M_NOWAIT | (curproc == 2186 pageproc ? M_USE_RESERVE : 0)); 2187 if (sb != NULL) { 2188 sb->p = rounddown(pindex, SWAP_META_PAGES); 2189 for (i = 0; i < SWAP_META_PAGES; i++) 2190 sb->d[i] = SWAPBLK_NONE; 2191 if (atomic_cmpset_int(&swblk_zone_exhausted, 2192 1, 0)) 2193 printf("swblk zone ok\n"); 2194 break; 2195 } 2196 if (nowait_noreplace) 2197 return (swapblk); 2198 VM_OBJECT_WUNLOCK(object); 2199 if (uma_zone_exhausted(swblk_zone)) { 2200 if (atomic_cmpset_int(&swblk_zone_exhausted, 2201 0, 1)) 2202 printf("swap blk zone exhausted, " 2203 "increase kern.maxswzone\n"); 2204 vm_pageout_oom(VM_OOM_SWAPZ); 2205 pause("swzonxb", 10); 2206 } else 2207 uma_zwait(swblk_zone); 2208 VM_OBJECT_WLOCK(object); 2209 sb = swblk_iter_reinit(blks, object, pindex); 2210 if (sb != NULL) 2211 /* 2212 * Somebody swapped out a nearby page, 2213 * allocating swblk at the pindex index, 2214 * while we dropped the object lock. 2215 */ 2216 goto allocated; 2217 } 2218 for (;;) { 2219 error = swblk_iter_insert(blks, sb); 2220 if (error == 0) { 2221 if (atomic_cmpset_int(&swpctrie_zone_exhausted, 2222 1, 0)) 2223 printf("swpctrie zone ok\n"); 2224 break; 2225 } 2226 if (nowait_noreplace) { 2227 uma_zfree(swblk_zone, sb); 2228 return (swapblk); 2229 } 2230 VM_OBJECT_WUNLOCK(object); 2231 if (uma_zone_exhausted(swpctrie_zone)) { 2232 if (atomic_cmpset_int(&swpctrie_zone_exhausted, 2233 0, 1)) 2234 printf("swap pctrie zone exhausted, " 2235 "increase kern.maxswzone\n"); 2236 vm_pageout_oom(VM_OOM_SWAPZ); 2237 pause("swzonxp", 10); 2238 } else 2239 uma_zwait(swpctrie_zone); 2240 VM_OBJECT_WLOCK(object); 2241 sb1 = swblk_iter_reinit(blks, object, pindex); 2242 if (sb1 != NULL) { 2243 uma_zfree(swblk_zone, sb); 2244 sb = sb1; 2245 goto allocated; 2246 } 2247 } 2248 } 2249 allocated: 2250 MPASS(sb->p == rounddown(pindex, SWAP_META_PAGES)); 2251 2252 modpi = pindex % SWAP_META_PAGES; 2253 /* Return prior contents of metadata. */ 2254 prev_swapblk = sb->d[modpi]; 2255 if (!nowait_noreplace || prev_swapblk == SWAPBLK_NONE) { 2256 /* Enter block into metadata. */ 2257 sb->d[modpi] = swapblk; 2258 2259 /* 2260 * Free the swblk if we end up with the empty page run. 2261 */ 2262 if (swapblk == SWAPBLK_NONE && 2263 swp_pager_swblk_empty(sb, 0, SWAP_META_PAGES)) { 2264 swblk_iter_remove(blks); 2265 uma_zfree(swblk_zone, sb); 2266 } 2267 } 2268 return (prev_swapblk); 2269 } 2270 2271 /* 2272 * SWP_PAGER_META_TRANSFER() - transfer a range of blocks in the srcobject's 2273 * swap metadata into dstobject. 2274 * 2275 * Blocks in src that correspond to holes in dst are transferred. Blocks 2276 * in src that correspond to blocks in dst are freed. 2277 */ 2278 static void 2279 swp_pager_meta_transfer(vm_object_t srcobject, vm_object_t dstobject, 2280 vm_pindex_t pindex, vm_pindex_t count) 2281 { 2282 struct pctrie_iter dstblks, srcblks; 2283 struct page_range range; 2284 struct swblk *sb; 2285 daddr_t blk, d[SWAP_META_PAGES]; 2286 vm_pindex_t last; 2287 int d_mask, i, limit, start; 2288 _Static_assert(8 * sizeof(d_mask) >= SWAP_META_PAGES, 2289 "d_mask not big enough"); 2290 2291 VM_OBJECT_ASSERT_WLOCKED(srcobject); 2292 VM_OBJECT_ASSERT_WLOCKED(dstobject); 2293 2294 if (count == 0 || swblk_is_empty(srcobject)) 2295 return; 2296 2297 swp_pager_init_freerange(&range); 2298 d_mask = 0; 2299 last = pindex + count; 2300 swblk_iter_init_only(&dstblks, dstobject); 2301 for (sb = swblk_iter_limit_init(&srcblks, srcobject, pindex, last), 2302 start = swblk_start(sb, pindex); 2303 sb != NULL; sb = swblk_iter_next(&srcblks), start = 0) { 2304 limit = MIN(last - srcblks.index, SWAP_META_PAGES); 2305 for (i = start; i < limit; i++) { 2306 if (sb->d[i] == SWAPBLK_NONE) 2307 continue; 2308 blk = swp_pager_meta_build(&dstblks, dstobject, 2309 srcblks.index + i - pindex, sb->d[i], true); 2310 if (blk == sb->d[i]) { 2311 /* 2312 * Failed memory allocation stopped transfer; 2313 * save this block for transfer with lock 2314 * released. 2315 */ 2316 d[i] = blk; 2317 d_mask |= 1 << i; 2318 } else if (blk != SWAPBLK_NONE) { 2319 /* Dst has a block at pindex, so free block. */ 2320 swp_pager_update_freerange(&range, sb->d[i]); 2321 } 2322 sb->d[i] = SWAPBLK_NONE; 2323 } 2324 if (swp_pager_swblk_empty(sb, 0, start) && 2325 swp_pager_swblk_empty(sb, limit, SWAP_META_PAGES)) { 2326 swblk_iter_remove(&srcblks); 2327 uma_zfree(swblk_zone, sb); 2328 } 2329 if (d_mask != 0) { 2330 /* Finish block transfer, with the lock released. */ 2331 VM_OBJECT_WUNLOCK(srcobject); 2332 do { 2333 i = ffs(d_mask) - 1; 2334 swp_pager_meta_build(&dstblks, dstobject, 2335 srcblks.index + i - pindex, d[i], false); 2336 d_mask &= ~(1 << i); 2337 } while (d_mask != 0); 2338 VM_OBJECT_WLOCK(srcobject); 2339 2340 /* 2341 * While the lock was not held, the iterator path could 2342 * have become stale, so discard it. 2343 */ 2344 pctrie_iter_reset(&srcblks); 2345 } 2346 } 2347 swp_pager_freeswapspace(&range); 2348 } 2349 2350 /* 2351 * SWP_PAGER_META_FREE() - free a range of blocks in the object's swap metadata 2352 * 2353 * Return freed swap blocks to the swap bitmap, and free emptied swblk 2354 * metadata. With 'freed' set, provide a count of freed blocks that were 2355 * not associated with valid resident pages. 2356 */ 2357 static void 2358 swp_pager_meta_free(vm_object_t object, vm_pindex_t pindex, vm_pindex_t count, 2359 vm_size_t *freed) 2360 { 2361 struct pctrie_iter blks, pages; 2362 struct page_range range; 2363 struct swblk *sb; 2364 vm_page_t m; 2365 vm_pindex_t last; 2366 vm_size_t fc; 2367 int i, limit, start; 2368 2369 VM_OBJECT_ASSERT_WLOCKED(object); 2370 2371 fc = 0; 2372 if (count == 0 || swblk_is_empty(object)) 2373 goto out; 2374 2375 swp_pager_init_freerange(&range); 2376 vm_page_iter_init(&pages, object); 2377 last = pindex + count; 2378 for (sb = swblk_iter_limit_init(&blks, object, pindex, last), 2379 start = swblk_start(sb, pindex); 2380 sb != NULL; sb = swblk_iter_next(&blks), start = 0) { 2381 limit = MIN(last - blks.index, SWAP_META_PAGES); 2382 for (i = start; i < limit; i++) { 2383 if (sb->d[i] == SWAPBLK_NONE) 2384 continue; 2385 swp_pager_update_freerange(&range, sb->d[i]); 2386 if (freed != NULL) { 2387 m = vm_page_iter_lookup(&pages, blks.index + i); 2388 if (m == NULL || vm_page_none_valid(m)) 2389 fc++; 2390 } 2391 sb->d[i] = SWAPBLK_NONE; 2392 } 2393 if (swp_pager_swblk_empty(sb, 0, start) && 2394 swp_pager_swblk_empty(sb, limit, SWAP_META_PAGES)) { 2395 swblk_iter_remove(&blks); 2396 uma_zfree(swblk_zone, sb); 2397 } 2398 } 2399 swp_pager_freeswapspace(&range); 2400 out: 2401 if (freed != NULL) 2402 *freed = fc; 2403 } 2404 2405 static void 2406 swp_pager_meta_free_block(struct swblk *sb, void *rangev) 2407 { 2408 struct page_range *range = rangev; 2409 2410 for (int i = 0; i < SWAP_META_PAGES; i++) { 2411 if (sb->d[i] != SWAPBLK_NONE) 2412 swp_pager_update_freerange(range, sb->d[i]); 2413 } 2414 uma_zfree(swblk_zone, sb); 2415 } 2416 2417 /* 2418 * SWP_PAGER_META_FREE_ALL() - destroy all swap metadata associated with object 2419 * 2420 * This routine locates and destroys all swap metadata associated with 2421 * an object. 2422 */ 2423 static void 2424 swp_pager_meta_free_all(vm_object_t object) 2425 { 2426 struct page_range range; 2427 2428 VM_OBJECT_ASSERT_WLOCKED(object); 2429 2430 swp_pager_init_freerange(&range); 2431 SWAP_PCTRIE_RECLAIM_CALLBACK(&object->un_pager.swp.swp_blks, 2432 swp_pager_meta_free_block, &range); 2433 swp_pager_freeswapspace(&range); 2434 } 2435 2436 /* 2437 * SWP_PAGER_METACTL() - misc control of swap meta data. 2438 * 2439 * This routine is capable of looking up, or removing swapblk 2440 * assignments in the swap meta data. It returns the swapblk being 2441 * looked-up, popped, or SWAPBLK_NONE if the block was invalid. 2442 * 2443 * When acting on a busy resident page and paging is in progress, we 2444 * have to wait until paging is complete but otherwise can act on the 2445 * busy page. 2446 */ 2447 static daddr_t 2448 swp_pager_meta_lookup(struct pctrie_iter *blks, vm_pindex_t pindex) 2449 { 2450 struct swblk *sb; 2451 2452 sb = swblk_iter_lookup(blks, pindex); 2453 if (sb == NULL) 2454 return (SWAPBLK_NONE); 2455 return (sb->d[pindex % SWAP_META_PAGES]); 2456 } 2457 2458 /* 2459 * Returns the least page index which is greater than or equal to the parameter 2460 * pindex and for which there is a swap block allocated. Returns OBJ_MAX_SIZE 2461 * if are no allocated swap blocks for the object after the requested pindex. 2462 */ 2463 static vm_pindex_t 2464 swap_pager_iter_find_least(struct pctrie_iter *blks, vm_pindex_t pindex) 2465 { 2466 struct swblk *sb; 2467 int i; 2468 2469 if ((sb = swblk_iter_lookup_ge(blks, pindex)) == NULL) 2470 return (OBJ_MAX_SIZE); 2471 if (blks->index < pindex) { 2472 for (i = pindex % SWAP_META_PAGES; i < SWAP_META_PAGES; i++) { 2473 if (sb->d[i] != SWAPBLK_NONE) 2474 return (blks->index + i); 2475 } 2476 if ((sb = swblk_iter_next(blks)) == NULL) 2477 return (OBJ_MAX_SIZE); 2478 } 2479 for (i = 0; i < SWAP_META_PAGES; i++) { 2480 if (sb->d[i] != SWAPBLK_NONE) 2481 return (blks->index + i); 2482 } 2483 2484 /* 2485 * We get here if a swblk is present in the trie but it 2486 * doesn't map any blocks. 2487 */ 2488 MPASS(0); 2489 return (OBJ_MAX_SIZE); 2490 } 2491 2492 /* 2493 * Find the first index >= pindex that has either a valid page or a swap 2494 * block. 2495 */ 2496 vm_pindex_t 2497 swap_pager_seek_data(vm_object_t object, vm_pindex_t pindex) 2498 { 2499 struct pctrie_iter blks, pages; 2500 vm_page_t m; 2501 vm_pindex_t swap_index; 2502 2503 VM_OBJECT_ASSERT_RLOCKED(object); 2504 vm_page_iter_init(&pages, object); 2505 m = vm_page_iter_lookup_ge(&pages, pindex); 2506 if (m != NULL && pages.index == pindex && vm_page_any_valid(m)) 2507 return (pages.index); 2508 swblk_iter_init_only(&blks, object); 2509 swap_index = swap_pager_iter_find_least(&blks, pindex); 2510 if (swap_index == pindex) 2511 return (swap_index); 2512 2513 /* 2514 * Find the first resident page after m, before swap_index. 2515 */ 2516 while (m != NULL && pages.index < swap_index) { 2517 if (vm_page_any_valid(m)) 2518 return (pages.index); 2519 m = vm_radix_iter_step(&pages); 2520 } 2521 if (swap_index == OBJ_MAX_SIZE) 2522 swap_index = object->size; 2523 return (swap_index); 2524 } 2525 2526 /* 2527 * Find the first index >= pindex that has neither a valid page nor a swap 2528 * block. 2529 */ 2530 vm_pindex_t 2531 swap_pager_seek_hole(vm_object_t object, vm_pindex_t pindex) 2532 { 2533 struct pctrie_iter blks, pages; 2534 struct swblk *sb; 2535 vm_page_t m; 2536 2537 VM_OBJECT_ASSERT_RLOCKED(object); 2538 vm_page_iter_init(&pages, object); 2539 swblk_iter_init_only(&blks, object); 2540 while (((m = vm_page_iter_lookup(&pages, pindex)) != NULL && 2541 vm_page_any_valid(m)) || 2542 ((sb = swblk_iter_lookup(&blks, pindex)) != NULL && 2543 sb->d[pindex % SWAP_META_PAGES] != SWAPBLK_NONE)) 2544 pindex++; 2545 return (pindex); 2546 } 2547 2548 /* 2549 * Is every page in the backing object or swap shadowed in the parent, and 2550 * unbusy and valid in swap? 2551 */ 2552 bool 2553 swap_pager_scan_all_shadowed(vm_object_t object) 2554 { 2555 struct pctrie_iter backing_blks, backing_pages, blks, pages; 2556 vm_object_t backing_object; 2557 vm_page_t p, pp; 2558 vm_pindex_t backing_offset_index, new_pindex, pi, pi_ubound, ps, pv; 2559 2560 VM_OBJECT_ASSERT_WLOCKED(object); 2561 VM_OBJECT_ASSERT_WLOCKED(object->backing_object); 2562 2563 backing_object = object->backing_object; 2564 2565 if ((backing_object->flags & OBJ_ANON) == 0) 2566 return (false); 2567 2568 KASSERT((object->flags & OBJ_ANON) != 0, 2569 ("Shadow object is not anonymous")); 2570 backing_offset_index = OFF_TO_IDX(object->backing_object_offset); 2571 pi_ubound = MIN(backing_object->size, 2572 backing_offset_index + object->size); 2573 vm_page_iter_init(&pages, object); 2574 vm_page_iter_init(&backing_pages, backing_object); 2575 swblk_iter_init_only(&blks, object); 2576 swblk_iter_init_only(&backing_blks, backing_object); 2577 2578 /* 2579 * Only check pages inside the parent object's range and inside the 2580 * parent object's mapping of the backing object. 2581 */ 2582 pv = ps = pi = backing_offset_index - 1; 2583 for (;;) { 2584 if (pi == pv) { 2585 p = vm_page_iter_lookup_ge(&backing_pages, pv + 1); 2586 pv = p != NULL ? p->pindex : backing_object->size; 2587 } 2588 if (pi == ps) 2589 ps = swap_pager_iter_find_least(&backing_blks, ps + 1); 2590 pi = MIN(pv, ps); 2591 if (pi >= pi_ubound) 2592 break; 2593 2594 if (pi == pv) { 2595 /* 2596 * If the backing object page is busy a grandparent or 2597 * older page may still be undergoing CoW. It is not 2598 * safe to collapse the backing object until it is 2599 * quiesced. 2600 */ 2601 if (vm_page_tryxbusy(p) == 0) 2602 return (false); 2603 2604 /* 2605 * We raced with the fault handler that left newly 2606 * allocated invalid page on the object queue and 2607 * retried. 2608 */ 2609 if (!vm_page_all_valid(p)) 2610 break; 2611 2612 /* 2613 * Busy of p disallows fault handler to validate parent 2614 * page (pp, below). 2615 */ 2616 } 2617 2618 /* 2619 * See if the parent has the page or if the parent's object 2620 * pager has the page. If the parent has the page but the page 2621 * is not valid, the parent's object pager must have the page. 2622 * 2623 * If this fails, the parent does not completely shadow the 2624 * object and we might as well give up now. 2625 */ 2626 new_pindex = pi - backing_offset_index; 2627 pp = vm_page_iter_lookup(&pages, new_pindex); 2628 2629 /* 2630 * The valid check here is stable due to object lock being 2631 * required to clear valid and initiate paging. 2632 */ 2633 if ((pp == NULL || vm_page_none_valid(pp)) && 2634 !swp_pager_haspage_iter(&blks, new_pindex, NULL, 2635 NULL)) 2636 break; 2637 if (pi == pv) 2638 vm_page_xunbusy(p); 2639 } 2640 if (pi < pi_ubound) { 2641 if (pi == pv) 2642 vm_page_xunbusy(p); 2643 return (false); 2644 } 2645 return (true); 2646 } 2647 2648 /* 2649 * System call swapon(name) enables swapping on device name, 2650 * which must be in the swdevsw. Return EBUSY 2651 * if already swapping on this device. 2652 */ 2653 #ifndef _SYS_SYSPROTO_H_ 2654 struct swapon_args { 2655 char *name; 2656 }; 2657 #endif 2658 2659 int 2660 sys_swapon(struct thread *td, struct swapon_args *uap) 2661 { 2662 struct vattr attr; 2663 struct vnode *vp; 2664 struct nameidata nd; 2665 int error; 2666 2667 error = priv_check(td, PRIV_SWAPON); 2668 if (error) 2669 return (error); 2670 2671 sx_xlock(&swdev_syscall_lock); 2672 2673 /* 2674 * Swap metadata may not fit in the KVM if we have physical 2675 * memory of >1GB. 2676 */ 2677 if (swblk_zone == NULL) { 2678 error = ENOMEM; 2679 goto done; 2680 } 2681 2682 NDINIT(&nd, LOOKUP, ISOPEN | FOLLOW | LOCKLEAF | AUDITVNODE1, 2683 UIO_USERSPACE, uap->name); 2684 error = namei(&nd); 2685 if (error) 2686 goto done; 2687 2688 NDFREE_PNBUF(&nd); 2689 vp = nd.ni_vp; 2690 2691 if (vn_isdisk_error(vp, &error)) { 2692 error = swapongeom(vp); 2693 } else if (vp->v_type == VREG && 2694 (vp->v_mount->mnt_vfc->vfc_flags & VFCF_NETWORK) != 0 && 2695 (error = VOP_GETATTR(vp, &attr, td->td_ucred)) == 0) { 2696 /* 2697 * Allow direct swapping to NFS regular files in the same 2698 * way that nfs_mountroot() sets up diskless swapping. 2699 */ 2700 error = swaponvp(td, vp, attr.va_size / DEV_BSIZE); 2701 } 2702 2703 if (error != 0) 2704 vput(vp); 2705 else 2706 VOP_UNLOCK(vp); 2707 done: 2708 sx_xunlock(&swdev_syscall_lock); 2709 return (error); 2710 } 2711 2712 /* 2713 * Check that the total amount of swap currently configured does not 2714 * exceed half the theoretical maximum. If it does, print a warning 2715 * message. 2716 */ 2717 static void 2718 swapon_check_swzone(void) 2719 { 2720 2721 /* recommend using no more than half that amount */ 2722 if (swap_total > swap_maxpages / 2) { 2723 printf("warning: total configured swap (%lu pages) " 2724 "exceeds maximum recommended amount (%lu pages).\n", 2725 swap_total, swap_maxpages / 2); 2726 printf("warning: increase kern.maxswzone " 2727 "or reduce amount of swap.\n"); 2728 } 2729 } 2730 2731 static void 2732 swaponsomething(struct vnode *vp, void *id, u_long nblks, 2733 sw_strategy_t *strategy, sw_close_t *close, dev_t dev, int flags) 2734 { 2735 struct swdevt *sp, *tsp; 2736 daddr_t dvbase; 2737 2738 /* 2739 * nblks is in DEV_BSIZE'd chunks, convert to PAGE_SIZE'd chunks. 2740 * First chop nblks off to page-align it, then convert. 2741 * 2742 * sw->sw_nblks is in page-sized chunks now too. 2743 */ 2744 nblks &= ~(ctodb(1) - 1); 2745 nblks = dbtoc(nblks); 2746 2747 sp = malloc(sizeof *sp, M_VMPGDATA, M_WAITOK | M_ZERO); 2748 sp->sw_blist = blist_create(nblks, M_WAITOK); 2749 sp->sw_vp = vp; 2750 sp->sw_id = id; 2751 sp->sw_dev = dev; 2752 sp->sw_nblks = nblks; 2753 sp->sw_used = 0; 2754 sp->sw_strategy = strategy; 2755 sp->sw_close = close; 2756 sp->sw_flags = flags; 2757 2758 /* 2759 * Do not free the first blocks in order to avoid overwriting 2760 * any bsd label at the front of the partition 2761 */ 2762 blist_free(sp->sw_blist, howmany(BBSIZE, PAGE_SIZE), 2763 nblks - howmany(BBSIZE, PAGE_SIZE)); 2764 2765 dvbase = 0; 2766 mtx_lock(&sw_dev_mtx); 2767 TAILQ_FOREACH(tsp, &swtailq, sw_list) { 2768 if (tsp->sw_end >= dvbase) { 2769 /* 2770 * We put one uncovered page between the devices 2771 * in order to definitively prevent any cross-device 2772 * I/O requests 2773 */ 2774 dvbase = tsp->sw_end + 1; 2775 } 2776 } 2777 sp->sw_first = dvbase; 2778 sp->sw_end = dvbase + nblks; 2779 TAILQ_INSERT_TAIL(&swtailq, sp, sw_list); 2780 nswapdev++; 2781 swap_pager_avail += nblks - howmany(BBSIZE, PAGE_SIZE); 2782 swap_total += nblks; 2783 swapon_check_swzone(); 2784 swp_sizecheck(); 2785 mtx_unlock(&sw_dev_mtx); 2786 EVENTHANDLER_INVOKE(swapon, sp); 2787 } 2788 2789 /* 2790 * SYSCALL: swapoff(devname) 2791 * 2792 * Disable swapping on the given device. 2793 * 2794 * XXX: Badly designed system call: it should use a device index 2795 * rather than filename as specification. We keep sw_vp around 2796 * only to make this work. 2797 */ 2798 static int 2799 kern_swapoff(struct thread *td, const char *name, enum uio_seg name_seg, 2800 u_int flags) 2801 { 2802 struct vnode *vp; 2803 struct nameidata nd; 2804 struct swdevt *sp; 2805 int error; 2806 2807 error = priv_check(td, PRIV_SWAPOFF); 2808 if (error != 0) 2809 return (error); 2810 if ((flags & ~(SWAPOFF_FORCE)) != 0) 2811 return (EINVAL); 2812 2813 sx_xlock(&swdev_syscall_lock); 2814 2815 NDINIT(&nd, LOOKUP, FOLLOW | AUDITVNODE1, name_seg, name); 2816 error = namei(&nd); 2817 if (error) 2818 goto done; 2819 NDFREE_PNBUF(&nd); 2820 vp = nd.ni_vp; 2821 2822 mtx_lock(&sw_dev_mtx); 2823 TAILQ_FOREACH(sp, &swtailq, sw_list) { 2824 if (sp->sw_vp == vp) 2825 break; 2826 } 2827 mtx_unlock(&sw_dev_mtx); 2828 if (sp == NULL) { 2829 error = EINVAL; 2830 goto done; 2831 } 2832 error = swapoff_one(sp, td->td_ucred, flags); 2833 done: 2834 sx_xunlock(&swdev_syscall_lock); 2835 return (error); 2836 } 2837 2838 2839 #ifdef COMPAT_FREEBSD13 2840 int 2841 freebsd13_swapoff(struct thread *td, struct freebsd13_swapoff_args *uap) 2842 { 2843 return (kern_swapoff(td, uap->name, UIO_USERSPACE, 0)); 2844 } 2845 #endif 2846 2847 int 2848 sys_swapoff(struct thread *td, struct swapoff_args *uap) 2849 { 2850 return (kern_swapoff(td, uap->name, UIO_USERSPACE, uap->flags)); 2851 } 2852 2853 static int 2854 swapoff_one(struct swdevt *sp, struct ucred *cred, u_int flags) 2855 { 2856 u_long nblks; 2857 #ifdef MAC 2858 int error; 2859 #endif 2860 2861 sx_assert(&swdev_syscall_lock, SA_XLOCKED); 2862 #ifdef MAC 2863 (void) vn_lock(sp->sw_vp, LK_EXCLUSIVE | LK_RETRY); 2864 error = mac_system_check_swapoff(cred, sp->sw_vp); 2865 (void) VOP_UNLOCK(sp->sw_vp); 2866 if (error != 0) 2867 return (error); 2868 #endif 2869 nblks = sp->sw_nblks; 2870 2871 /* 2872 * We can turn off this swap device safely only if the 2873 * available virtual memory in the system will fit the amount 2874 * of data we will have to page back in, plus an epsilon so 2875 * the system doesn't become critically low on swap space. 2876 * The vm_free_count() part does not account e.g. for clean 2877 * pages that can be immediately reclaimed without paging, so 2878 * this is a very rough estimation. 2879 * 2880 * On the other hand, not turning swap off on swapoff_all() 2881 * means that we can lose swap data when filesystems go away, 2882 * which is arguably worse. 2883 */ 2884 if ((flags & SWAPOFF_FORCE) == 0 && 2885 vm_free_count() + swap_pager_avail < nblks + nswap_lowat) 2886 return (ENOMEM); 2887 2888 /* 2889 * Prevent further allocations on this device. 2890 */ 2891 mtx_lock(&sw_dev_mtx); 2892 sp->sw_flags |= SW_CLOSING; 2893 swap_pager_avail -= blist_fill(sp->sw_blist, 0, nblks); 2894 swap_total -= nblks; 2895 mtx_unlock(&sw_dev_mtx); 2896 2897 /* 2898 * Page in the contents of the device and close it. 2899 */ 2900 swap_pager_swapoff(sp); 2901 2902 sp->sw_close(curthread, sp); 2903 mtx_lock(&sw_dev_mtx); 2904 sp->sw_id = NULL; 2905 TAILQ_REMOVE(&swtailq, sp, sw_list); 2906 nswapdev--; 2907 if (nswapdev == 0) { 2908 swap_pager_full = 2; 2909 swap_pager_almost_full = 1; 2910 } 2911 if (swdevhd == sp) 2912 swdevhd = NULL; 2913 mtx_unlock(&sw_dev_mtx); 2914 blist_destroy(sp->sw_blist); 2915 free(sp, M_VMPGDATA); 2916 return (0); 2917 } 2918 2919 void 2920 swapoff_all(void) 2921 { 2922 struct swdevt *sp, *spt; 2923 const char *devname; 2924 int error; 2925 2926 sx_xlock(&swdev_syscall_lock); 2927 2928 mtx_lock(&sw_dev_mtx); 2929 TAILQ_FOREACH_SAFE(sp, &swtailq, sw_list, spt) { 2930 mtx_unlock(&sw_dev_mtx); 2931 if (vn_isdisk(sp->sw_vp)) 2932 devname = devtoname(sp->sw_vp->v_rdev); 2933 else 2934 devname = "[file]"; 2935 error = swapoff_one(sp, thread0.td_ucred, SWAPOFF_FORCE); 2936 if (error != 0) { 2937 printf("Cannot remove swap device %s (error=%d), " 2938 "skipping.\n", devname, error); 2939 } else if (bootverbose) { 2940 printf("Swap device %s removed.\n", devname); 2941 } 2942 mtx_lock(&sw_dev_mtx); 2943 } 2944 mtx_unlock(&sw_dev_mtx); 2945 2946 sx_xunlock(&swdev_syscall_lock); 2947 } 2948 2949 void 2950 swap_pager_status(int *total, int *used) 2951 { 2952 2953 *total = swap_total; 2954 *used = swap_total - swap_pager_avail - 2955 nswapdev * howmany(BBSIZE, PAGE_SIZE); 2956 } 2957 2958 int 2959 swap_dev_info(int name, struct xswdev *xs, char *devname, size_t len) 2960 { 2961 struct swdevt *sp; 2962 const char *tmp_devname; 2963 int error, n; 2964 2965 n = 0; 2966 error = ENOENT; 2967 mtx_lock(&sw_dev_mtx); 2968 TAILQ_FOREACH(sp, &swtailq, sw_list) { 2969 if (n != name) { 2970 n++; 2971 continue; 2972 } 2973 xs->xsw_version = XSWDEV_VERSION; 2974 xs->xsw_dev = sp->sw_dev; 2975 xs->xsw_flags = sp->sw_flags; 2976 xs->xsw_nblks = sp->sw_nblks; 2977 xs->xsw_used = sp->sw_used; 2978 if (devname != NULL) { 2979 if (vn_isdisk(sp->sw_vp)) 2980 tmp_devname = devtoname(sp->sw_vp->v_rdev); 2981 else 2982 tmp_devname = "[file]"; 2983 strncpy(devname, tmp_devname, len); 2984 } 2985 error = 0; 2986 break; 2987 } 2988 mtx_unlock(&sw_dev_mtx); 2989 return (error); 2990 } 2991 2992 #if defined(COMPAT_FREEBSD11) 2993 #define XSWDEV_VERSION_11 1 2994 struct xswdev11 { 2995 u_int xsw_version; 2996 uint32_t xsw_dev; 2997 int xsw_flags; 2998 int xsw_nblks; 2999 int xsw_used; 3000 }; 3001 #endif 3002 3003 #if defined(__amd64__) && defined(COMPAT_FREEBSD32) 3004 struct xswdev32 { 3005 u_int xsw_version; 3006 u_int xsw_dev1, xsw_dev2; 3007 int xsw_flags; 3008 int xsw_nblks; 3009 int xsw_used; 3010 }; 3011 #endif 3012 3013 static int 3014 sysctl_vm_swap_info(SYSCTL_HANDLER_ARGS) 3015 { 3016 struct xswdev xs; 3017 #if defined(__amd64__) && defined(COMPAT_FREEBSD32) 3018 struct xswdev32 xs32; 3019 #endif 3020 #if defined(COMPAT_FREEBSD11) 3021 struct xswdev11 xs11; 3022 #endif 3023 int error; 3024 3025 if (arg2 != 1) /* name length */ 3026 return (EINVAL); 3027 3028 memset(&xs, 0, sizeof(xs)); 3029 error = swap_dev_info(*(int *)arg1, &xs, NULL, 0); 3030 if (error != 0) 3031 return (error); 3032 #if defined(__amd64__) && defined(COMPAT_FREEBSD32) 3033 if (req->oldlen == sizeof(xs32)) { 3034 memset(&xs32, 0, sizeof(xs32)); 3035 xs32.xsw_version = XSWDEV_VERSION; 3036 xs32.xsw_dev1 = xs.xsw_dev; 3037 xs32.xsw_dev2 = xs.xsw_dev >> 32; 3038 xs32.xsw_flags = xs.xsw_flags; 3039 xs32.xsw_nblks = xs.xsw_nblks; 3040 xs32.xsw_used = xs.xsw_used; 3041 error = SYSCTL_OUT(req, &xs32, sizeof(xs32)); 3042 return (error); 3043 } 3044 #endif 3045 #if defined(COMPAT_FREEBSD11) 3046 if (req->oldlen == sizeof(xs11)) { 3047 memset(&xs11, 0, sizeof(xs11)); 3048 xs11.xsw_version = XSWDEV_VERSION_11; 3049 xs11.xsw_dev = xs.xsw_dev; /* truncation */ 3050 xs11.xsw_flags = xs.xsw_flags; 3051 xs11.xsw_nblks = xs.xsw_nblks; 3052 xs11.xsw_used = xs.xsw_used; 3053 error = SYSCTL_OUT(req, &xs11, sizeof(xs11)); 3054 return (error); 3055 } 3056 #endif 3057 error = SYSCTL_OUT(req, &xs, sizeof(xs)); 3058 return (error); 3059 } 3060 3061 SYSCTL_INT(_vm, OID_AUTO, nswapdev, CTLFLAG_RD, &nswapdev, 0, 3062 "Number of swap devices"); 3063 SYSCTL_NODE(_vm, OID_AUTO, swap_info, CTLFLAG_RD | CTLFLAG_MPSAFE, 3064 sysctl_vm_swap_info, 3065 "Swap statistics by device"); 3066 3067 /* 3068 * Count the approximate swap usage in pages for a vmspace. The 3069 * shadowed or not yet copied on write swap blocks are not accounted. 3070 * The map must be locked. 3071 */ 3072 long 3073 vmspace_swap_count(struct vmspace *vmspace) 3074 { 3075 struct pctrie_iter blks; 3076 vm_map_t map; 3077 vm_map_entry_t cur; 3078 vm_object_t object; 3079 struct swblk *sb; 3080 vm_pindex_t e, pi; 3081 long count; 3082 int i, limit, start; 3083 3084 map = &vmspace->vm_map; 3085 count = 0; 3086 3087 VM_MAP_ENTRY_FOREACH(cur, map) { 3088 if ((cur->eflags & MAP_ENTRY_IS_SUB_MAP) != 0) 3089 continue; 3090 object = cur->object.vm_object; 3091 if (object == NULL || (object->flags & OBJ_SWAP) == 0) 3092 continue; 3093 VM_OBJECT_RLOCK(object); 3094 if ((object->flags & OBJ_SWAP) == 0) 3095 goto unlock; 3096 pi = OFF_TO_IDX(cur->offset); 3097 e = pi + OFF_TO_IDX(cur->end - cur->start); 3098 for (sb = swblk_iter_limit_init(&blks, object, pi, e), 3099 start = swblk_start(sb, pi); 3100 sb != NULL; sb = swblk_iter_next(&blks), start = 0) { 3101 limit = MIN(e - blks.index, SWAP_META_PAGES); 3102 for (i = start; i < limit; i++) { 3103 if (sb->d[i] != SWAPBLK_NONE) 3104 count++; 3105 } 3106 } 3107 unlock: 3108 VM_OBJECT_RUNLOCK(object); 3109 } 3110 return (count); 3111 } 3112 3113 /* 3114 * GEOM backend 3115 * 3116 * Swapping onto disk devices. 3117 * 3118 */ 3119 3120 static g_orphan_t swapgeom_orphan; 3121 3122 static struct g_class g_swap_class = { 3123 .name = "SWAP", 3124 .version = G_VERSION, 3125 .orphan = swapgeom_orphan, 3126 }; 3127 3128 DECLARE_GEOM_CLASS(g_swap_class, g_class); 3129 3130 static void 3131 swapgeom_close_ev(void *arg, int flags) 3132 { 3133 struct g_consumer *cp; 3134 3135 cp = arg; 3136 g_access(cp, -1, -1, 0); 3137 g_detach(cp); 3138 g_destroy_consumer(cp); 3139 } 3140 3141 /* 3142 * Add a reference to the g_consumer for an inflight transaction. 3143 */ 3144 static void 3145 swapgeom_acquire(struct g_consumer *cp) 3146 { 3147 3148 mtx_assert(&sw_dev_mtx, MA_OWNED); 3149 cp->index++; 3150 } 3151 3152 /* 3153 * Remove a reference from the g_consumer. Post a close event if all 3154 * references go away, since the function might be called from the 3155 * biodone context. 3156 */ 3157 static void 3158 swapgeom_release(struct g_consumer *cp, struct swdevt *sp) 3159 { 3160 3161 mtx_assert(&sw_dev_mtx, MA_OWNED); 3162 cp->index--; 3163 if (cp->index == 0) { 3164 if (g_post_event(swapgeom_close_ev, cp, M_NOWAIT, NULL) == 0) 3165 sp->sw_id = NULL; 3166 } 3167 } 3168 3169 static void 3170 swapgeom_done(struct bio *bp2) 3171 { 3172 struct swdevt *sp; 3173 struct buf *bp; 3174 struct g_consumer *cp; 3175 3176 bp = bp2->bio_caller2; 3177 cp = bp2->bio_from; 3178 bp->b_ioflags = bp2->bio_flags; 3179 if (bp2->bio_error) 3180 bp->b_ioflags |= BIO_ERROR; 3181 bp->b_resid = bp->b_bcount - bp2->bio_completed; 3182 bp->b_error = bp2->bio_error; 3183 bp->b_caller1 = NULL; 3184 bufdone(bp); 3185 sp = bp2->bio_caller1; 3186 mtx_lock(&sw_dev_mtx); 3187 swapgeom_release(cp, sp); 3188 mtx_unlock(&sw_dev_mtx); 3189 g_destroy_bio(bp2); 3190 } 3191 3192 static void 3193 swapgeom_strategy(struct buf *bp, struct swdevt *sp) 3194 { 3195 struct bio *bio; 3196 struct g_consumer *cp; 3197 3198 mtx_lock(&sw_dev_mtx); 3199 cp = sp->sw_id; 3200 if (cp == NULL) { 3201 mtx_unlock(&sw_dev_mtx); 3202 bp->b_error = ENXIO; 3203 bp->b_ioflags |= BIO_ERROR; 3204 bufdone(bp); 3205 return; 3206 } 3207 swapgeom_acquire(cp); 3208 mtx_unlock(&sw_dev_mtx); 3209 if (bp->b_iocmd == BIO_WRITE) 3210 bio = g_new_bio(); 3211 else 3212 bio = g_alloc_bio(); 3213 if (bio == NULL) { 3214 mtx_lock(&sw_dev_mtx); 3215 swapgeom_release(cp, sp); 3216 mtx_unlock(&sw_dev_mtx); 3217 bp->b_error = ENOMEM; 3218 bp->b_ioflags |= BIO_ERROR; 3219 printf("swap_pager: cannot allocate bio\n"); 3220 bufdone(bp); 3221 return; 3222 } 3223 3224 bp->b_caller1 = bio; 3225 bio->bio_caller1 = sp; 3226 bio->bio_caller2 = bp; 3227 bio->bio_cmd = bp->b_iocmd; 3228 bio->bio_offset = (bp->b_blkno - sp->sw_first) * PAGE_SIZE; 3229 bio->bio_length = bp->b_bcount; 3230 bio->bio_done = swapgeom_done; 3231 bio->bio_flags |= BIO_SWAP; 3232 if (!buf_mapped(bp)) { 3233 bio->bio_ma = bp->b_pages; 3234 bio->bio_data = unmapped_buf; 3235 bio->bio_ma_offset = (vm_offset_t)bp->b_offset & PAGE_MASK; 3236 bio->bio_ma_n = bp->b_npages; 3237 bio->bio_flags |= BIO_UNMAPPED; 3238 } else { 3239 bio->bio_data = bp->b_data; 3240 bio->bio_ma = NULL; 3241 } 3242 g_io_request(bio, cp); 3243 return; 3244 } 3245 3246 static void 3247 swapgeom_orphan(struct g_consumer *cp) 3248 { 3249 struct swdevt *sp; 3250 int destroy; 3251 3252 mtx_lock(&sw_dev_mtx); 3253 TAILQ_FOREACH(sp, &swtailq, sw_list) { 3254 if (sp->sw_id == cp) { 3255 sp->sw_flags |= SW_CLOSING; 3256 break; 3257 } 3258 } 3259 /* 3260 * Drop reference we were created with. Do directly since we're in a 3261 * special context where we don't have to queue the call to 3262 * swapgeom_close_ev(). 3263 */ 3264 cp->index--; 3265 destroy = ((sp != NULL) && (cp->index == 0)); 3266 if (destroy) 3267 sp->sw_id = NULL; 3268 mtx_unlock(&sw_dev_mtx); 3269 if (destroy) 3270 swapgeom_close_ev(cp, 0); 3271 } 3272 3273 static void 3274 swapgeom_close(struct thread *td, struct swdevt *sw) 3275 { 3276 struct g_consumer *cp; 3277 3278 mtx_lock(&sw_dev_mtx); 3279 cp = sw->sw_id; 3280 sw->sw_id = NULL; 3281 mtx_unlock(&sw_dev_mtx); 3282 3283 /* 3284 * swapgeom_close() may be called from the biodone context, 3285 * where we cannot perform topology changes. Delegate the 3286 * work to the events thread. 3287 */ 3288 if (cp != NULL) 3289 g_waitfor_event(swapgeom_close_ev, cp, M_WAITOK, NULL); 3290 } 3291 3292 static int 3293 swapongeom_locked(struct cdev *dev, struct vnode *vp) 3294 { 3295 struct g_provider *pp; 3296 struct g_consumer *cp; 3297 static struct g_geom *gp; 3298 struct swdevt *sp; 3299 u_long nblks; 3300 int error; 3301 3302 pp = g_dev_getprovider(dev); 3303 if (pp == NULL) 3304 return (ENODEV); 3305 mtx_lock(&sw_dev_mtx); 3306 TAILQ_FOREACH(sp, &swtailq, sw_list) { 3307 cp = sp->sw_id; 3308 if (cp != NULL && cp->provider == pp) { 3309 mtx_unlock(&sw_dev_mtx); 3310 return (EBUSY); 3311 } 3312 } 3313 mtx_unlock(&sw_dev_mtx); 3314 if (gp == NULL) 3315 gp = g_new_geomf(&g_swap_class, "swap"); 3316 cp = g_new_consumer(gp); 3317 cp->index = 1; /* Number of active I/Os, plus one for being active. */ 3318 cp->flags |= G_CF_DIRECT_SEND | G_CF_DIRECT_RECEIVE; 3319 g_attach(cp, pp); 3320 /* 3321 * XXX: Every time you think you can improve the margin for 3322 * footshooting, somebody depends on the ability to do so: 3323 * savecore(8) wants to write to our swapdev so we cannot 3324 * set an exclusive count :-( 3325 */ 3326 error = g_access(cp, 1, 1, 0); 3327 if (error != 0) { 3328 g_detach(cp); 3329 g_destroy_consumer(cp); 3330 return (error); 3331 } 3332 nblks = pp->mediasize / DEV_BSIZE; 3333 swaponsomething(vp, cp, nblks, swapgeom_strategy, 3334 swapgeom_close, dev2udev(dev), 3335 (pp->flags & G_PF_ACCEPT_UNMAPPED) != 0 ? SW_UNMAPPED : 0); 3336 return (0); 3337 } 3338 3339 static int 3340 swapongeom(struct vnode *vp) 3341 { 3342 int error; 3343 3344 ASSERT_VOP_ELOCKED(vp, "swapongeom"); 3345 if (vp->v_type != VCHR || VN_IS_DOOMED(vp)) { 3346 error = ENOENT; 3347 } else { 3348 g_topology_lock(); 3349 error = swapongeom_locked(vp->v_rdev, vp); 3350 g_topology_unlock(); 3351 } 3352 return (error); 3353 } 3354 3355 /* 3356 * VNODE backend 3357 * 3358 * This is used mainly for network filesystem (read: probably only tested 3359 * with NFS) swapfiles. 3360 * 3361 */ 3362 3363 static void 3364 swapdev_strategy(struct buf *bp, struct swdevt *sp) 3365 { 3366 struct vnode *vp2; 3367 3368 bp->b_blkno = ctodb(bp->b_blkno - sp->sw_first); 3369 3370 vp2 = sp->sw_id; 3371 vhold(vp2); 3372 if (bp->b_iocmd == BIO_WRITE) { 3373 vn_lock(vp2, LK_EXCLUSIVE | LK_RETRY); 3374 if (bp->b_bufobj) 3375 bufobj_wdrop(bp->b_bufobj); 3376 bufobj_wref(&vp2->v_bufobj); 3377 } else { 3378 vn_lock(vp2, LK_SHARED | LK_RETRY); 3379 } 3380 if (bp->b_bufobj != &vp2->v_bufobj) 3381 bp->b_bufobj = &vp2->v_bufobj; 3382 bp->b_vp = vp2; 3383 bp->b_iooffset = dbtob(bp->b_blkno); 3384 bstrategy(bp); 3385 VOP_UNLOCK(vp2); 3386 } 3387 3388 static void 3389 swapdev_close(struct thread *td, struct swdevt *sp) 3390 { 3391 struct vnode *vp; 3392 3393 vp = sp->sw_vp; 3394 vn_lock(vp, LK_EXCLUSIVE | LK_RETRY); 3395 VOP_CLOSE(vp, FREAD | FWRITE, td->td_ucred, td); 3396 vput(vp); 3397 } 3398 3399 static int 3400 swaponvp(struct thread *td, struct vnode *vp, u_long nblks) 3401 { 3402 struct swdevt *sp; 3403 int error; 3404 3405 ASSERT_VOP_ELOCKED(vp, "swaponvp"); 3406 if (nblks == 0) 3407 return (ENXIO); 3408 mtx_lock(&sw_dev_mtx); 3409 TAILQ_FOREACH(sp, &swtailq, sw_list) { 3410 if (sp->sw_id == vp) { 3411 mtx_unlock(&sw_dev_mtx); 3412 return (EBUSY); 3413 } 3414 } 3415 mtx_unlock(&sw_dev_mtx); 3416 3417 #ifdef MAC 3418 error = mac_system_check_swapon(td->td_ucred, vp); 3419 if (error == 0) 3420 #endif 3421 error = VOP_OPEN(vp, FREAD | FWRITE, td->td_ucred, td, NULL); 3422 if (error != 0) 3423 return (error); 3424 3425 swaponsomething(vp, vp, nblks, swapdev_strategy, swapdev_close, 3426 NODEV, 0); 3427 return (0); 3428 } 3429 3430 static int 3431 sysctl_swap_async_max(SYSCTL_HANDLER_ARGS) 3432 { 3433 int error, new, n; 3434 3435 new = nsw_wcount_async_max; 3436 error = sysctl_handle_int(oidp, &new, 0, req); 3437 if (error != 0 || req->newptr == NULL) 3438 return (error); 3439 3440 if (new > nswbuf / 2 || new < 1) 3441 return (EINVAL); 3442 3443 mtx_lock(&swbuf_mtx); 3444 while (nsw_wcount_async_max != new) { 3445 /* 3446 * Adjust difference. If the current async count is too low, 3447 * we will need to sqeeze our update slowly in. Sleep with a 3448 * higher priority than getpbuf() to finish faster. 3449 */ 3450 n = new - nsw_wcount_async_max; 3451 if (nsw_wcount_async + n >= 0) { 3452 nsw_wcount_async += n; 3453 nsw_wcount_async_max += n; 3454 wakeup(&nsw_wcount_async); 3455 } else { 3456 nsw_wcount_async_max -= nsw_wcount_async; 3457 nsw_wcount_async = 0; 3458 msleep(&nsw_wcount_async, &swbuf_mtx, PSWP, 3459 "swpsysctl", 0); 3460 } 3461 } 3462 mtx_unlock(&swbuf_mtx); 3463 3464 return (0); 3465 } 3466 3467 static void 3468 swap_pager_update_writecount(vm_object_t object, vm_offset_t start, 3469 vm_offset_t end) 3470 { 3471 3472 VM_OBJECT_WLOCK(object); 3473 KASSERT((object->flags & OBJ_ANON) == 0, 3474 ("Splittable object with writecount")); 3475 object->un_pager.swp.writemappings += (vm_ooffset_t)end - start; 3476 VM_OBJECT_WUNLOCK(object); 3477 } 3478 3479 static void 3480 swap_pager_release_writecount(vm_object_t object, vm_offset_t start, 3481 vm_offset_t end) 3482 { 3483 3484 VM_OBJECT_WLOCK(object); 3485 KASSERT((object->flags & OBJ_ANON) == 0, 3486 ("Splittable object with writecount")); 3487 KASSERT(object->un_pager.swp.writemappings >= (vm_ooffset_t)end - start, 3488 ("swap obj %p writecount %jx dec %jx", object, 3489 (uintmax_t)object->un_pager.swp.writemappings, 3490 (uintmax_t)((vm_ooffset_t)end - start))); 3491 object->un_pager.swp.writemappings -= (vm_ooffset_t)end - start; 3492 VM_OBJECT_WUNLOCK(object); 3493 } 3494