1 /*- 2 * SPDX-License-Identifier: BSD-4-Clause 3 * 4 * Copyright (c) 1998 Matthew Dillon, 5 * Copyright (c) 1994 John S. Dyson 6 * Copyright (c) 1990 University of Utah. 7 * Copyright (c) 1982, 1986, 1989, 1993 8 * The Regents of the University of California. All rights reserved. 9 * 10 * This code is derived from software contributed to Berkeley by 11 * the Systems Programming Group of the University of Utah Computer 12 * Science Department. 13 * 14 * Redistribution and use in source and binary forms, with or without 15 * modification, are permitted provided that the following conditions 16 * are met: 17 * 1. Redistributions of source code must retain the above copyright 18 * notice, this list of conditions and the following disclaimer. 19 * 2. Redistributions in binary form must reproduce the above copyright 20 * notice, this list of conditions and the following disclaimer in the 21 * documentation and/or other materials provided with the distribution. 22 * 3. All advertising materials mentioning features or use of this software 23 * must display the following acknowledgement: 24 * This product includes software developed by the University of 25 * California, Berkeley and its contributors. 26 * 4. Neither the name of the University nor the names of its contributors 27 * may be used to endorse or promote products derived from this software 28 * without specific prior written permission. 29 * 30 * THIS SOFTWARE IS PROVIDED BY THE REGENTS AND CONTRIBUTORS ``AS IS'' AND 31 * ANY EXPRESS OR IMPLIED WARRANTIES, INCLUDING, BUT NOT LIMITED TO, THE 32 * IMPLIED WARRANTIES OF MERCHANTABILITY AND FITNESS FOR A PARTICULAR PURPOSE 33 * ARE DISCLAIMED. IN NO EVENT SHALL THE REGENTS OR CONTRIBUTORS BE LIABLE 34 * FOR ANY DIRECT, INDIRECT, INCIDENTAL, SPECIAL, EXEMPLARY, OR CONSEQUENTIAL 35 * DAMAGES (INCLUDING, BUT NOT LIMITED TO, PROCUREMENT OF SUBSTITUTE GOODS 36 * OR SERVICES; LOSS OF USE, DATA, OR PROFITS; OR BUSINESS INTERRUPTION) 37 * HOWEVER CAUSED AND ON ANY THEORY OF LIABILITY, WHETHER IN CONTRACT, STRICT 38 * LIABILITY, OR TORT (INCLUDING NEGLIGENCE OR OTHERWISE) ARISING IN ANY WAY 39 * OUT OF THE USE OF THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF 40 * SUCH DAMAGE. 41 * 42 * New Swap System 43 * Matthew Dillon 44 * 45 * Radix Bitmap 'blists'. 46 * 47 * - The new swapper uses the new radix bitmap code. This should scale 48 * to arbitrarily small or arbitrarily large swap spaces and an almost 49 * arbitrary degree of fragmentation. 50 * 51 * Features: 52 * 53 * - on the fly reallocation of swap during putpages. The new system 54 * does not try to keep previously allocated swap blocks for dirty 55 * pages. 56 * 57 * - on the fly deallocation of swap 58 * 59 * - No more garbage collection required. Unnecessarily allocated swap 60 * blocks only exist for dirty vm_page_t's now and these are already 61 * cycled (in a high-load system) by the pager. We also do on-the-fly 62 * removal of invalidated swap blocks when a page is destroyed 63 * or renamed. 64 * 65 * from: Utah $Hdr: swap_pager.c 1.4 91/04/30$ 66 * 67 * @(#)swap_pager.c 8.9 (Berkeley) 3/21/94 68 * @(#)vm_swap.c 8.5 (Berkeley) 2/17/94 69 */ 70 71 #include <sys/cdefs.h> 72 __FBSDID("$FreeBSD$"); 73 74 #include "opt_vm.h" 75 76 #include <sys/param.h> 77 #include <sys/bio.h> 78 #include <sys/blist.h> 79 #include <sys/buf.h> 80 #include <sys/conf.h> 81 #include <sys/disk.h> 82 #include <sys/disklabel.h> 83 #include <sys/eventhandler.h> 84 #include <sys/fcntl.h> 85 #include <sys/lock.h> 86 #include <sys/kernel.h> 87 #include <sys/mount.h> 88 #include <sys/namei.h> 89 #include <sys/malloc.h> 90 #include <sys/pctrie.h> 91 #include <sys/priv.h> 92 #include <sys/proc.h> 93 #include <sys/racct.h> 94 #include <sys/resource.h> 95 #include <sys/resourcevar.h> 96 #include <sys/rwlock.h> 97 #include <sys/sbuf.h> 98 #include <sys/sysctl.h> 99 #include <sys/sysproto.h> 100 #include <sys/systm.h> 101 #include <sys/sx.h> 102 #include <sys/vmmeter.h> 103 #include <sys/vnode.h> 104 105 #include <security/mac/mac_framework.h> 106 107 #include <vm/vm.h> 108 #include <vm/pmap.h> 109 #include <vm/vm_map.h> 110 #include <vm/vm_kern.h> 111 #include <vm/vm_object.h> 112 #include <vm/vm_page.h> 113 #include <vm/vm_pager.h> 114 #include <vm/vm_pageout.h> 115 #include <vm/vm_param.h> 116 #include <vm/swap_pager.h> 117 #include <vm/vm_extern.h> 118 #include <vm/uma.h> 119 120 #include <geom/geom.h> 121 122 /* 123 * MAX_PAGEOUT_CLUSTER must be a power of 2 between 1 and 64. 124 * The 64-page limit is due to the radix code (kern/subr_blist.c). 125 */ 126 #ifndef MAX_PAGEOUT_CLUSTER 127 #define MAX_PAGEOUT_CLUSTER 32 128 #endif 129 130 #if !defined(SWB_NPAGES) 131 #define SWB_NPAGES MAX_PAGEOUT_CLUSTER 132 #endif 133 134 #define SWAP_META_PAGES PCTRIE_COUNT 135 136 /* 137 * A swblk structure maps each page index within a 138 * SWAP_META_PAGES-aligned and sized range to the address of an 139 * on-disk swap block (or SWAPBLK_NONE). The collection of these 140 * mappings for an entire vm object is implemented as a pc-trie. 141 */ 142 struct swblk { 143 vm_pindex_t p; 144 daddr_t d[SWAP_META_PAGES]; 145 }; 146 147 static MALLOC_DEFINE(M_VMPGDATA, "vm_pgdata", "swap pager private data"); 148 static struct mtx sw_dev_mtx; 149 static TAILQ_HEAD(, swdevt) swtailq = TAILQ_HEAD_INITIALIZER(swtailq); 150 static struct swdevt *swdevhd; /* Allocate from here next */ 151 static int nswapdev; /* Number of swap devices */ 152 int swap_pager_avail; 153 static struct sx swdev_syscall_lock; /* serialize swap(on|off) */ 154 155 static __exclusive_cache_line u_long swap_reserved; 156 static u_long swap_total; 157 static int sysctl_page_shift(SYSCTL_HANDLER_ARGS); 158 159 static SYSCTL_NODE(_vm_stats, OID_AUTO, swap, CTLFLAG_RD | CTLFLAG_MPSAFE, 0, 160 "VM swap stats"); 161 162 SYSCTL_PROC(_vm, OID_AUTO, swap_reserved, CTLTYPE_U64 | CTLFLAG_RD | CTLFLAG_MPSAFE, 163 &swap_reserved, 0, sysctl_page_shift, "A", 164 "Amount of swap storage needed to back all allocated anonymous memory."); 165 SYSCTL_PROC(_vm, OID_AUTO, swap_total, CTLTYPE_U64 | CTLFLAG_RD | CTLFLAG_MPSAFE, 166 &swap_total, 0, sysctl_page_shift, "A", 167 "Total amount of available swap storage."); 168 169 static int overcommit = 0; 170 SYSCTL_INT(_vm, VM_OVERCOMMIT, overcommit, CTLFLAG_RW, &overcommit, 0, 171 "Configure virtual memory overcommit behavior. See tuning(7) " 172 "for details."); 173 static unsigned long swzone; 174 SYSCTL_ULONG(_vm, OID_AUTO, swzone, CTLFLAG_RD, &swzone, 0, 175 "Actual size of swap metadata zone"); 176 static unsigned long swap_maxpages; 177 SYSCTL_ULONG(_vm, OID_AUTO, swap_maxpages, CTLFLAG_RD, &swap_maxpages, 0, 178 "Maximum amount of swap supported"); 179 180 static COUNTER_U64_DEFINE_EARLY(swap_free_deferred); 181 SYSCTL_COUNTER_U64(_vm_stats_swap, OID_AUTO, free_deferred, 182 CTLFLAG_RD, &swap_free_deferred, 183 "Number of pages that deferred freeing swap space"); 184 185 static COUNTER_U64_DEFINE_EARLY(swap_free_completed); 186 SYSCTL_COUNTER_U64(_vm_stats_swap, OID_AUTO, free_completed, 187 CTLFLAG_RD, &swap_free_completed, 188 "Number of deferred frees completed"); 189 190 /* bits from overcommit */ 191 #define SWAP_RESERVE_FORCE_ON (1 << 0) 192 #define SWAP_RESERVE_RLIMIT_ON (1 << 1) 193 #define SWAP_RESERVE_ALLOW_NONWIRED (1 << 2) 194 195 static int 196 sysctl_page_shift(SYSCTL_HANDLER_ARGS) 197 { 198 uint64_t newval; 199 u_long value = *(u_long *)arg1; 200 201 newval = ((uint64_t)value) << PAGE_SHIFT; 202 return (sysctl_handle_64(oidp, &newval, 0, req)); 203 } 204 205 static bool 206 swap_reserve_by_cred_rlimit(u_long pincr, struct ucred *cred, int oc) 207 { 208 struct uidinfo *uip; 209 u_long prev; 210 211 uip = cred->cr_ruidinfo; 212 213 prev = atomic_fetchadd_long(&uip->ui_vmsize, pincr); 214 if ((oc & SWAP_RESERVE_RLIMIT_ON) != 0 && 215 prev + pincr > lim_cur(curthread, RLIMIT_SWAP) && 216 priv_check(curthread, PRIV_VM_SWAP_NORLIMIT) != 0) { 217 prev = atomic_fetchadd_long(&uip->ui_vmsize, -pincr); 218 KASSERT(prev >= pincr, ("negative vmsize for uid = %d\n", uip->ui_uid)); 219 return (false); 220 } 221 return (true); 222 } 223 224 static void 225 swap_release_by_cred_rlimit(u_long pdecr, struct ucred *cred) 226 { 227 struct uidinfo *uip; 228 #ifdef INVARIANTS 229 u_long prev; 230 #endif 231 232 uip = cred->cr_ruidinfo; 233 234 #ifdef INVARIANTS 235 prev = atomic_fetchadd_long(&uip->ui_vmsize, -pdecr); 236 KASSERT(prev >= pdecr, ("negative vmsize for uid = %d\n", uip->ui_uid)); 237 #else 238 atomic_subtract_long(&uip->ui_vmsize, pdecr); 239 #endif 240 } 241 242 static void 243 swap_reserve_force_rlimit(u_long pincr, struct ucred *cred) 244 { 245 struct uidinfo *uip; 246 247 uip = cred->cr_ruidinfo; 248 atomic_add_long(&uip->ui_vmsize, pincr); 249 } 250 251 bool 252 swap_reserve(vm_ooffset_t incr) 253 { 254 255 return (swap_reserve_by_cred(incr, curthread->td_ucred)); 256 } 257 258 bool 259 swap_reserve_by_cred(vm_ooffset_t incr, struct ucred *cred) 260 { 261 u_long r, s, prev, pincr; 262 #ifdef RACCT 263 int error; 264 #endif 265 int oc; 266 static int curfail; 267 static struct timeval lastfail; 268 269 KASSERT((incr & PAGE_MASK) == 0, ("%s: incr: %ju & PAGE_MASK", __func__, 270 (uintmax_t)incr)); 271 272 #ifdef RACCT 273 if (RACCT_ENABLED()) { 274 PROC_LOCK(curproc); 275 error = racct_add(curproc, RACCT_SWAP, incr); 276 PROC_UNLOCK(curproc); 277 if (error != 0) 278 return (false); 279 } 280 #endif 281 282 pincr = atop(incr); 283 prev = atomic_fetchadd_long(&swap_reserved, pincr); 284 r = prev + pincr; 285 s = swap_total; 286 oc = atomic_load_int(&overcommit); 287 if (r > s && (oc & SWAP_RESERVE_ALLOW_NONWIRED) != 0) { 288 s += vm_cnt.v_page_count - vm_cnt.v_free_reserved - 289 vm_wire_count(); 290 } 291 if ((oc & SWAP_RESERVE_FORCE_ON) != 0 && r > s && 292 priv_check(curthread, PRIV_VM_SWAP_NOQUOTA) != 0) { 293 prev = atomic_fetchadd_long(&swap_reserved, -pincr); 294 KASSERT(prev >= pincr, ("swap_reserved < incr on overcommit fail")); 295 goto out_error; 296 } 297 298 if (!swap_reserve_by_cred_rlimit(pincr, cred, oc)) { 299 prev = atomic_fetchadd_long(&swap_reserved, -pincr); 300 KASSERT(prev >= pincr, ("swap_reserved < incr on overcommit fail")); 301 goto out_error; 302 } 303 304 return (true); 305 306 out_error: 307 if (ppsratecheck(&lastfail, &curfail, 1)) { 308 printf("uid %d, pid %d: swap reservation for %jd bytes failed\n", 309 cred->cr_ruidinfo->ui_uid, curproc->p_pid, incr); 310 } 311 #ifdef RACCT 312 if (RACCT_ENABLED()) { 313 PROC_LOCK(curproc); 314 racct_sub(curproc, RACCT_SWAP, incr); 315 PROC_UNLOCK(curproc); 316 } 317 #endif 318 319 return (false); 320 } 321 322 void 323 swap_reserve_force(vm_ooffset_t incr) 324 { 325 u_long pincr; 326 327 KASSERT((incr & PAGE_MASK) == 0, ("%s: incr: %ju & PAGE_MASK", __func__, 328 (uintmax_t)incr)); 329 330 #ifdef RACCT 331 if (RACCT_ENABLED()) { 332 PROC_LOCK(curproc); 333 racct_add_force(curproc, RACCT_SWAP, incr); 334 PROC_UNLOCK(curproc); 335 } 336 #endif 337 pincr = atop(incr); 338 atomic_add_long(&swap_reserved, pincr); 339 swap_reserve_force_rlimit(pincr, curthread->td_ucred); 340 } 341 342 void 343 swap_release(vm_ooffset_t decr) 344 { 345 struct ucred *cred; 346 347 PROC_LOCK(curproc); 348 cred = curproc->p_ucred; 349 swap_release_by_cred(decr, cred); 350 PROC_UNLOCK(curproc); 351 } 352 353 void 354 swap_release_by_cred(vm_ooffset_t decr, struct ucred *cred) 355 { 356 u_long pdecr; 357 #ifdef INVARIANTS 358 u_long prev; 359 #endif 360 361 KASSERT((decr & PAGE_MASK) == 0, ("%s: decr: %ju & PAGE_MASK", __func__, 362 (uintmax_t)decr)); 363 364 pdecr = atop(decr); 365 #ifdef INVARIANTS 366 prev = atomic_fetchadd_long(&swap_reserved, -pdecr); 367 KASSERT(prev >= pdecr, ("swap_reserved < decr")); 368 #else 369 atomic_subtract_long(&swap_reserved, pdecr); 370 #endif 371 372 swap_release_by_cred_rlimit(pdecr, cred); 373 #ifdef RACCT 374 if (racct_enable) 375 racct_sub_cred(cred, RACCT_SWAP, decr); 376 #endif 377 } 378 379 static int swap_pager_full = 2; /* swap space exhaustion (task killing) */ 380 static int swap_pager_almost_full = 1; /* swap space exhaustion (w/hysteresis)*/ 381 static struct mtx swbuf_mtx; /* to sync nsw_wcount_async */ 382 static int nsw_wcount_async; /* limit async write buffers */ 383 static int nsw_wcount_async_max;/* assigned maximum */ 384 static int nsw_cluster_max; /* maximum VOP I/O allowed */ 385 386 static int sysctl_swap_async_max(SYSCTL_HANDLER_ARGS); 387 SYSCTL_PROC(_vm, OID_AUTO, swap_async_max, CTLTYPE_INT | CTLFLAG_RW | 388 CTLFLAG_MPSAFE, NULL, 0, sysctl_swap_async_max, "I", 389 "Maximum running async swap ops"); 390 static int sysctl_swap_fragmentation(SYSCTL_HANDLER_ARGS); 391 SYSCTL_PROC(_vm, OID_AUTO, swap_fragmentation, CTLTYPE_STRING | CTLFLAG_RD | 392 CTLFLAG_MPSAFE, NULL, 0, sysctl_swap_fragmentation, "A", 393 "Swap Fragmentation Info"); 394 395 static struct sx sw_alloc_sx; 396 397 /* 398 * "named" and "unnamed" anon region objects. Try to reduce the overhead 399 * of searching a named list by hashing it just a little. 400 */ 401 402 #define NOBJLISTS 8 403 404 #define NOBJLIST(handle) \ 405 (&swap_pager_object_list[((int)(intptr_t)handle >> 4) & (NOBJLISTS-1)]) 406 407 static struct pagerlst swap_pager_object_list[NOBJLISTS]; 408 static uma_zone_t swwbuf_zone; 409 static uma_zone_t swrbuf_zone; 410 static uma_zone_t swblk_zone; 411 static uma_zone_t swpctrie_zone; 412 413 /* 414 * pagerops for OBJT_SWAP - "swap pager". Some ops are also global procedure 415 * calls hooked from other parts of the VM system and do not appear here. 416 * (see vm/swap_pager.h). 417 */ 418 static vm_object_t 419 swap_pager_alloc(void *handle, vm_ooffset_t size, 420 vm_prot_t prot, vm_ooffset_t offset, struct ucred *); 421 static vm_object_t 422 swap_tmpfs_pager_alloc(void *handle, vm_ooffset_t size, 423 vm_prot_t prot, vm_ooffset_t offset, struct ucred *); 424 static void swap_pager_dealloc(vm_object_t object); 425 static int swap_pager_getpages(vm_object_t, vm_page_t *, int, int *, 426 int *); 427 static int swap_pager_getpages_async(vm_object_t, vm_page_t *, int, int *, 428 int *, pgo_getpages_iodone_t, void *); 429 static void swap_pager_putpages(vm_object_t, vm_page_t *, int, boolean_t, int *); 430 static boolean_t 431 swap_pager_haspage(vm_object_t object, vm_pindex_t pindex, int *before, int *after); 432 static void swap_pager_init(void); 433 static void swap_pager_unswapped(vm_page_t); 434 static void swap_pager_swapoff(struct swdevt *sp); 435 static void swap_pager_update_writecount(vm_object_t object, 436 vm_offset_t start, vm_offset_t end); 437 static void swap_pager_release_writecount(vm_object_t object, 438 vm_offset_t start, vm_offset_t end); 439 static void swap_tmpfs_pager_getvp(vm_object_t object, struct vnode **vpp, 440 bool *vp_heldp); 441 static void swap_pager_freespace(vm_object_t object, vm_pindex_t start, 442 vm_size_t size); 443 444 const struct pagerops swappagerops = { 445 .pgo_init = swap_pager_init, /* early system initialization of pager */ 446 .pgo_alloc = swap_pager_alloc, /* allocate an OBJT_SWAP object */ 447 .pgo_dealloc = swap_pager_dealloc, /* deallocate an OBJT_SWAP object */ 448 .pgo_getpages = swap_pager_getpages, /* pagein */ 449 .pgo_getpages_async = swap_pager_getpages_async, /* pagein (async) */ 450 .pgo_putpages = swap_pager_putpages, /* pageout */ 451 .pgo_haspage = swap_pager_haspage, /* get backing store status for page */ 452 .pgo_pageunswapped = swap_pager_unswapped, /* remove swap related to page */ 453 .pgo_update_writecount = swap_pager_update_writecount, 454 .pgo_release_writecount = swap_pager_release_writecount, 455 .pgo_freespace = swap_pager_freespace, 456 }; 457 458 const struct pagerops swaptmpfspagerops = { 459 .pgo_alloc = swap_tmpfs_pager_alloc, 460 .pgo_dealloc = swap_pager_dealloc, 461 .pgo_getpages = swap_pager_getpages, 462 .pgo_getpages_async = swap_pager_getpages_async, 463 .pgo_putpages = swap_pager_putpages, 464 .pgo_haspage = swap_pager_haspage, 465 .pgo_pageunswapped = swap_pager_unswapped, 466 .pgo_update_writecount = swap_pager_update_writecount, 467 .pgo_release_writecount = swap_pager_release_writecount, 468 .pgo_set_writeable_dirty = vm_object_set_writeable_dirty_, 469 .pgo_mightbedirty = vm_object_mightbedirty_, 470 .pgo_getvp = swap_tmpfs_pager_getvp, 471 .pgo_freespace = swap_pager_freespace, 472 }; 473 474 /* 475 * swap_*() routines are externally accessible. swp_*() routines are 476 * internal. 477 */ 478 static int nswap_lowat = 128; /* in pages, swap_pager_almost_full warn */ 479 static int nswap_hiwat = 512; /* in pages, swap_pager_almost_full warn */ 480 481 SYSCTL_INT(_vm, OID_AUTO, dmmax, CTLFLAG_RD, &nsw_cluster_max, 0, 482 "Maximum size of a swap block in pages"); 483 484 static void swp_sizecheck(void); 485 static void swp_pager_async_iodone(struct buf *bp); 486 static bool swp_pager_swblk_empty(struct swblk *sb, int start, int limit); 487 static void swp_pager_free_empty_swblk(vm_object_t, struct swblk *sb); 488 static int swapongeom(struct vnode *); 489 static int swaponvp(struct thread *, struct vnode *, u_long); 490 static int swapoff_one(struct swdevt *sp, struct ucred *cred); 491 492 /* 493 * Swap bitmap functions 494 */ 495 static void swp_pager_freeswapspace(daddr_t blk, daddr_t npages); 496 static daddr_t swp_pager_getswapspace(int *npages); 497 498 /* 499 * Metadata functions 500 */ 501 static daddr_t swp_pager_meta_build(vm_object_t, vm_pindex_t, daddr_t); 502 static void swp_pager_meta_free(vm_object_t, vm_pindex_t, vm_pindex_t); 503 static void swp_pager_meta_transfer(vm_object_t src, vm_object_t dst, 504 vm_pindex_t pindex, vm_pindex_t count); 505 static void swp_pager_meta_free_all(vm_object_t); 506 static daddr_t swp_pager_meta_lookup(vm_object_t, vm_pindex_t); 507 508 static void 509 swp_pager_init_freerange(daddr_t *start, daddr_t *num) 510 { 511 512 *start = SWAPBLK_NONE; 513 *num = 0; 514 } 515 516 static void 517 swp_pager_update_freerange(daddr_t *start, daddr_t *num, daddr_t addr) 518 { 519 520 if (*start + *num == addr) { 521 (*num)++; 522 } else { 523 swp_pager_freeswapspace(*start, *num); 524 *start = addr; 525 *num = 1; 526 } 527 } 528 529 static void * 530 swblk_trie_alloc(struct pctrie *ptree) 531 { 532 533 return (uma_zalloc(swpctrie_zone, M_NOWAIT | (curproc == pageproc ? 534 M_USE_RESERVE : 0))); 535 } 536 537 static void 538 swblk_trie_free(struct pctrie *ptree, void *node) 539 { 540 541 uma_zfree(swpctrie_zone, node); 542 } 543 544 PCTRIE_DEFINE(SWAP, swblk, p, swblk_trie_alloc, swblk_trie_free); 545 546 /* 547 * SWP_SIZECHECK() - update swap_pager_full indication 548 * 549 * update the swap_pager_almost_full indication and warn when we are 550 * about to run out of swap space, using lowat/hiwat hysteresis. 551 * 552 * Clear swap_pager_full ( task killing ) indication when lowat is met. 553 * 554 * No restrictions on call 555 * This routine may not block. 556 */ 557 static void 558 swp_sizecheck(void) 559 { 560 561 if (swap_pager_avail < nswap_lowat) { 562 if (swap_pager_almost_full == 0) { 563 printf("swap_pager: out of swap space\n"); 564 swap_pager_almost_full = 1; 565 } 566 } else { 567 swap_pager_full = 0; 568 if (swap_pager_avail > nswap_hiwat) 569 swap_pager_almost_full = 0; 570 } 571 } 572 573 /* 574 * SWAP_PAGER_INIT() - initialize the swap pager! 575 * 576 * Expected to be started from system init. NOTE: This code is run 577 * before much else so be careful what you depend on. Most of the VM 578 * system has yet to be initialized at this point. 579 */ 580 static void 581 swap_pager_init(void) 582 { 583 /* 584 * Initialize object lists 585 */ 586 int i; 587 588 for (i = 0; i < NOBJLISTS; ++i) 589 TAILQ_INIT(&swap_pager_object_list[i]); 590 mtx_init(&sw_dev_mtx, "swapdev", NULL, MTX_DEF); 591 sx_init(&sw_alloc_sx, "swspsx"); 592 sx_init(&swdev_syscall_lock, "swsysc"); 593 } 594 595 /* 596 * SWAP_PAGER_SWAP_INIT() - swap pager initialization from pageout process 597 * 598 * Expected to be started from pageout process once, prior to entering 599 * its main loop. 600 */ 601 void 602 swap_pager_swap_init(void) 603 { 604 unsigned long n, n2; 605 606 /* 607 * Number of in-transit swap bp operations. Don't 608 * exhaust the pbufs completely. Make sure we 609 * initialize workable values (0 will work for hysteresis 610 * but it isn't very efficient). 611 * 612 * The nsw_cluster_max is constrained by the bp->b_pages[] 613 * array, which has maxphys / PAGE_SIZE entries, and our locally 614 * defined MAX_PAGEOUT_CLUSTER. Also be aware that swap ops are 615 * constrained by the swap device interleave stripe size. 616 * 617 * Currently we hardwire nsw_wcount_async to 4. This limit is 618 * designed to prevent other I/O from having high latencies due to 619 * our pageout I/O. The value 4 works well for one or two active swap 620 * devices but is probably a little low if you have more. Even so, 621 * a higher value would probably generate only a limited improvement 622 * with three or four active swap devices since the system does not 623 * typically have to pageout at extreme bandwidths. We will want 624 * at least 2 per swap devices, and 4 is a pretty good value if you 625 * have one NFS swap device due to the command/ack latency over NFS. 626 * So it all works out pretty well. 627 */ 628 nsw_cluster_max = min(maxphys / PAGE_SIZE, MAX_PAGEOUT_CLUSTER); 629 630 nsw_wcount_async = 4; 631 nsw_wcount_async_max = nsw_wcount_async; 632 mtx_init(&swbuf_mtx, "async swbuf mutex", NULL, MTX_DEF); 633 634 swwbuf_zone = pbuf_zsecond_create("swwbuf", nswbuf / 4); 635 swrbuf_zone = pbuf_zsecond_create("swrbuf", nswbuf / 2); 636 637 /* 638 * Initialize our zone, taking the user's requested size or 639 * estimating the number we need based on the number of pages 640 * in the system. 641 */ 642 n = maxswzone != 0 ? maxswzone / sizeof(struct swblk) : 643 vm_cnt.v_page_count / 2; 644 swpctrie_zone = uma_zcreate("swpctrie", pctrie_node_size(), NULL, NULL, 645 pctrie_zone_init, NULL, UMA_ALIGN_PTR, 0); 646 if (swpctrie_zone == NULL) 647 panic("failed to create swap pctrie zone."); 648 swblk_zone = uma_zcreate("swblk", sizeof(struct swblk), NULL, NULL, 649 NULL, NULL, _Alignof(struct swblk) - 1, 0); 650 if (swblk_zone == NULL) 651 panic("failed to create swap blk zone."); 652 n2 = n; 653 do { 654 if (uma_zone_reserve_kva(swblk_zone, n)) 655 break; 656 /* 657 * if the allocation failed, try a zone two thirds the 658 * size of the previous attempt. 659 */ 660 n -= ((n + 2) / 3); 661 } while (n > 0); 662 663 /* 664 * Often uma_zone_reserve_kva() cannot reserve exactly the 665 * requested size. Account for the difference when 666 * calculating swap_maxpages. 667 */ 668 n = uma_zone_get_max(swblk_zone); 669 670 if (n < n2) 671 printf("Swap blk zone entries changed from %lu to %lu.\n", 672 n2, n); 673 /* absolute maximum we can handle assuming 100% efficiency */ 674 swap_maxpages = n * SWAP_META_PAGES; 675 swzone = n * sizeof(struct swblk); 676 if (!uma_zone_reserve_kva(swpctrie_zone, n)) 677 printf("Cannot reserve swap pctrie zone, " 678 "reduce kern.maxswzone.\n"); 679 } 680 681 static vm_object_t 682 swap_pager_alloc_init(objtype_t otype, void *handle, struct ucred *cred, 683 vm_ooffset_t size, vm_ooffset_t offset) 684 { 685 vm_object_t object; 686 687 if (cred != NULL) { 688 if (!swap_reserve_by_cred(size, cred)) 689 return (NULL); 690 crhold(cred); 691 } 692 693 /* 694 * The un_pager.swp.swp_blks trie is initialized by 695 * vm_object_allocate() to ensure the correct order of 696 * visibility to other threads. 697 */ 698 object = vm_object_allocate(otype, OFF_TO_IDX(offset + 699 PAGE_MASK + size)); 700 701 object->un_pager.swp.writemappings = 0; 702 object->handle = handle; 703 if (cred != NULL) { 704 object->cred = cred; 705 object->charge = size; 706 } 707 return (object); 708 } 709 710 /* 711 * SWAP_PAGER_ALLOC() - allocate a new OBJT_SWAP VM object and instantiate 712 * its metadata structures. 713 * 714 * This routine is called from the mmap and fork code to create a new 715 * OBJT_SWAP object. 716 * 717 * This routine must ensure that no live duplicate is created for 718 * the named object request, which is protected against by 719 * holding the sw_alloc_sx lock in case handle != NULL. 720 */ 721 static vm_object_t 722 swap_pager_alloc(void *handle, vm_ooffset_t size, vm_prot_t prot, 723 vm_ooffset_t offset, struct ucred *cred) 724 { 725 vm_object_t object; 726 727 if (handle != NULL) { 728 /* 729 * Reference existing named region or allocate new one. There 730 * should not be a race here against swp_pager_meta_build() 731 * as called from vm_page_remove() in regards to the lookup 732 * of the handle. 733 */ 734 sx_xlock(&sw_alloc_sx); 735 object = vm_pager_object_lookup(NOBJLIST(handle), handle); 736 if (object == NULL) { 737 object = swap_pager_alloc_init(OBJT_SWAP, handle, cred, 738 size, offset); 739 if (object != NULL) { 740 TAILQ_INSERT_TAIL(NOBJLIST(object->handle), 741 object, pager_object_list); 742 } 743 } 744 sx_xunlock(&sw_alloc_sx); 745 } else { 746 object = swap_pager_alloc_init(OBJT_SWAP, handle, cred, 747 size, offset); 748 } 749 return (object); 750 } 751 752 static vm_object_t 753 swap_tmpfs_pager_alloc(void *handle, vm_ooffset_t size, vm_prot_t prot, 754 vm_ooffset_t offset, struct ucred *cred) 755 { 756 vm_object_t object; 757 758 MPASS(handle == NULL); 759 object = swap_pager_alloc_init(OBJT_SWAP_TMPFS, handle, cred, 760 size, offset); 761 return (object); 762 } 763 764 /* 765 * SWAP_PAGER_DEALLOC() - remove swap metadata from object 766 * 767 * The swap backing for the object is destroyed. The code is 768 * designed such that we can reinstantiate it later, but this 769 * routine is typically called only when the entire object is 770 * about to be destroyed. 771 * 772 * The object must be locked. 773 */ 774 static void 775 swap_pager_dealloc(vm_object_t object) 776 { 777 778 VM_OBJECT_ASSERT_WLOCKED(object); 779 KASSERT((object->flags & OBJ_DEAD) != 0, ("dealloc of reachable obj")); 780 781 /* 782 * Remove from list right away so lookups will fail if we block for 783 * pageout completion. 784 */ 785 if ((object->flags & OBJ_ANON) == 0 && object->handle != NULL) { 786 VM_OBJECT_WUNLOCK(object); 787 sx_xlock(&sw_alloc_sx); 788 TAILQ_REMOVE(NOBJLIST(object->handle), object, 789 pager_object_list); 790 sx_xunlock(&sw_alloc_sx); 791 VM_OBJECT_WLOCK(object); 792 } 793 794 vm_object_pip_wait(object, "swpdea"); 795 796 /* 797 * Free all remaining metadata. We only bother to free it from 798 * the swap meta data. We do not attempt to free swapblk's still 799 * associated with vm_page_t's for this object. We do not care 800 * if paging is still in progress on some objects. 801 */ 802 swp_pager_meta_free_all(object); 803 object->handle = NULL; 804 object->type = OBJT_DEAD; 805 vm_object_clear_flag(object, OBJ_SWAP); 806 } 807 808 /************************************************************************ 809 * SWAP PAGER BITMAP ROUTINES * 810 ************************************************************************/ 811 812 /* 813 * SWP_PAGER_GETSWAPSPACE() - allocate raw swap space 814 * 815 * Allocate swap for up to the requested number of pages. The 816 * starting swap block number (a page index) is returned or 817 * SWAPBLK_NONE if the allocation failed. 818 * 819 * Also has the side effect of advising that somebody made a mistake 820 * when they configured swap and didn't configure enough. 821 * 822 * This routine may not sleep. 823 * 824 * We allocate in round-robin fashion from the configured devices. 825 */ 826 static daddr_t 827 swp_pager_getswapspace(int *io_npages) 828 { 829 daddr_t blk; 830 struct swdevt *sp; 831 int mpages, npages; 832 833 KASSERT(*io_npages >= 1, 834 ("%s: npages not positive", __func__)); 835 blk = SWAPBLK_NONE; 836 mpages = *io_npages; 837 npages = imin(BLIST_MAX_ALLOC, mpages); 838 mtx_lock(&sw_dev_mtx); 839 sp = swdevhd; 840 while (!TAILQ_EMPTY(&swtailq)) { 841 if (sp == NULL) 842 sp = TAILQ_FIRST(&swtailq); 843 if ((sp->sw_flags & SW_CLOSING) == 0) 844 blk = blist_alloc(sp->sw_blist, &npages, mpages); 845 if (blk != SWAPBLK_NONE) 846 break; 847 sp = TAILQ_NEXT(sp, sw_list); 848 if (swdevhd == sp) { 849 if (npages == 1) 850 break; 851 mpages = npages - 1; 852 npages >>= 1; 853 } 854 } 855 if (blk != SWAPBLK_NONE) { 856 *io_npages = npages; 857 blk += sp->sw_first; 858 sp->sw_used += npages; 859 swap_pager_avail -= npages; 860 swp_sizecheck(); 861 swdevhd = TAILQ_NEXT(sp, sw_list); 862 } else { 863 if (swap_pager_full != 2) { 864 printf("swp_pager_getswapspace(%d): failed\n", 865 *io_npages); 866 swap_pager_full = 2; 867 swap_pager_almost_full = 1; 868 } 869 swdevhd = NULL; 870 } 871 mtx_unlock(&sw_dev_mtx); 872 return (blk); 873 } 874 875 static bool 876 swp_pager_isondev(daddr_t blk, struct swdevt *sp) 877 { 878 879 return (blk >= sp->sw_first && blk < sp->sw_end); 880 } 881 882 static void 883 swp_pager_strategy(struct buf *bp) 884 { 885 struct swdevt *sp; 886 887 mtx_lock(&sw_dev_mtx); 888 TAILQ_FOREACH(sp, &swtailq, sw_list) { 889 if (swp_pager_isondev(bp->b_blkno, sp)) { 890 mtx_unlock(&sw_dev_mtx); 891 if ((sp->sw_flags & SW_UNMAPPED) != 0 && 892 unmapped_buf_allowed) { 893 bp->b_data = unmapped_buf; 894 bp->b_offset = 0; 895 } else { 896 pmap_qenter((vm_offset_t)bp->b_data, 897 &bp->b_pages[0], bp->b_bcount / PAGE_SIZE); 898 } 899 sp->sw_strategy(bp, sp); 900 return; 901 } 902 } 903 panic("Swapdev not found"); 904 } 905 906 /* 907 * SWP_PAGER_FREESWAPSPACE() - free raw swap space 908 * 909 * This routine returns the specified swap blocks back to the bitmap. 910 * 911 * This routine may not sleep. 912 */ 913 static void 914 swp_pager_freeswapspace(daddr_t blk, daddr_t npages) 915 { 916 struct swdevt *sp; 917 918 if (npages == 0) 919 return; 920 mtx_lock(&sw_dev_mtx); 921 TAILQ_FOREACH(sp, &swtailq, sw_list) { 922 if (swp_pager_isondev(blk, sp)) { 923 sp->sw_used -= npages; 924 /* 925 * If we are attempting to stop swapping on 926 * this device, we don't want to mark any 927 * blocks free lest they be reused. 928 */ 929 if ((sp->sw_flags & SW_CLOSING) == 0) { 930 blist_free(sp->sw_blist, blk - sp->sw_first, 931 npages); 932 swap_pager_avail += npages; 933 swp_sizecheck(); 934 } 935 mtx_unlock(&sw_dev_mtx); 936 return; 937 } 938 } 939 panic("Swapdev not found"); 940 } 941 942 /* 943 * SYSCTL_SWAP_FRAGMENTATION() - produce raw swap space stats 944 */ 945 static int 946 sysctl_swap_fragmentation(SYSCTL_HANDLER_ARGS) 947 { 948 struct sbuf sbuf; 949 struct swdevt *sp; 950 const char *devname; 951 int error; 952 953 error = sysctl_wire_old_buffer(req, 0); 954 if (error != 0) 955 return (error); 956 sbuf_new_for_sysctl(&sbuf, NULL, 128, req); 957 mtx_lock(&sw_dev_mtx); 958 TAILQ_FOREACH(sp, &swtailq, sw_list) { 959 if (vn_isdisk(sp->sw_vp)) 960 devname = devtoname(sp->sw_vp->v_rdev); 961 else 962 devname = "[file]"; 963 sbuf_printf(&sbuf, "\nFree space on device %s:\n", devname); 964 blist_stats(sp->sw_blist, &sbuf); 965 } 966 mtx_unlock(&sw_dev_mtx); 967 error = sbuf_finish(&sbuf); 968 sbuf_delete(&sbuf); 969 return (error); 970 } 971 972 /* 973 * SWAP_PAGER_FREESPACE() - frees swap blocks associated with a page 974 * range within an object. 975 * 976 * This routine removes swapblk assignments from swap metadata. 977 * 978 * The external callers of this routine typically have already destroyed 979 * or renamed vm_page_t's associated with this range in the object so 980 * we should be ok. 981 * 982 * The object must be locked. 983 */ 984 static void 985 swap_pager_freespace(vm_object_t object, vm_pindex_t start, vm_size_t size) 986 { 987 988 swp_pager_meta_free(object, start, size); 989 } 990 991 /* 992 * SWAP_PAGER_RESERVE() - reserve swap blocks in object 993 * 994 * Assigns swap blocks to the specified range within the object. The 995 * swap blocks are not zeroed. Any previous swap assignment is destroyed. 996 * 997 * Returns 0 on success, -1 on failure. 998 */ 999 int 1000 swap_pager_reserve(vm_object_t object, vm_pindex_t start, vm_size_t size) 1001 { 1002 daddr_t addr, blk, n_free, s_free; 1003 int i, j, n; 1004 1005 swp_pager_init_freerange(&s_free, &n_free); 1006 VM_OBJECT_WLOCK(object); 1007 for (i = 0; i < size; i += n) { 1008 n = size - i; 1009 blk = swp_pager_getswapspace(&n); 1010 if (blk == SWAPBLK_NONE) { 1011 swp_pager_meta_free(object, start, i); 1012 VM_OBJECT_WUNLOCK(object); 1013 return (-1); 1014 } 1015 for (j = 0; j < n; ++j) { 1016 addr = swp_pager_meta_build(object, 1017 start + i + j, blk + j); 1018 if (addr != SWAPBLK_NONE) 1019 swp_pager_update_freerange(&s_free, &n_free, 1020 addr); 1021 } 1022 } 1023 swp_pager_freeswapspace(s_free, n_free); 1024 VM_OBJECT_WUNLOCK(object); 1025 return (0); 1026 } 1027 1028 static bool 1029 swp_pager_xfer_source(vm_object_t srcobject, vm_object_t dstobject, 1030 vm_pindex_t pindex, daddr_t addr) 1031 { 1032 daddr_t dstaddr; 1033 1034 KASSERT((srcobject->flags & OBJ_SWAP) != 0, 1035 ("%s: Srcobject not swappable", __func__)); 1036 if ((dstobject->flags & OBJ_SWAP) != 0 && 1037 swp_pager_meta_lookup(dstobject, pindex) != SWAPBLK_NONE) { 1038 /* Caller should destroy the source block. */ 1039 return (false); 1040 } 1041 1042 /* 1043 * Destination has no swapblk and is not resident, transfer source. 1044 * swp_pager_meta_build() can sleep. 1045 */ 1046 VM_OBJECT_WUNLOCK(srcobject); 1047 dstaddr = swp_pager_meta_build(dstobject, pindex, addr); 1048 KASSERT(dstaddr == SWAPBLK_NONE, 1049 ("Unexpected destination swapblk")); 1050 VM_OBJECT_WLOCK(srcobject); 1051 1052 return (true); 1053 } 1054 1055 /* 1056 * SWAP_PAGER_COPY() - copy blocks from source pager to destination pager 1057 * and destroy the source. 1058 * 1059 * Copy any valid swapblks from the source to the destination. In 1060 * cases where both the source and destination have a valid swapblk, 1061 * we keep the destination's. 1062 * 1063 * This routine is allowed to sleep. It may sleep allocating metadata 1064 * indirectly through swp_pager_meta_build(). 1065 * 1066 * The source object contains no vm_page_t's (which is just as well) 1067 * 1068 * The source object is of type OBJT_SWAP. 1069 * 1070 * The source and destination objects must be locked. 1071 * Both object locks may temporarily be released. 1072 */ 1073 void 1074 swap_pager_copy(vm_object_t srcobject, vm_object_t dstobject, 1075 vm_pindex_t offset, int destroysource) 1076 { 1077 1078 VM_OBJECT_ASSERT_WLOCKED(srcobject); 1079 VM_OBJECT_ASSERT_WLOCKED(dstobject); 1080 1081 /* 1082 * If destroysource is set, we remove the source object from the 1083 * swap_pager internal queue now. 1084 */ 1085 if (destroysource && (srcobject->flags & OBJ_ANON) == 0 && 1086 srcobject->handle != NULL) { 1087 VM_OBJECT_WUNLOCK(srcobject); 1088 VM_OBJECT_WUNLOCK(dstobject); 1089 sx_xlock(&sw_alloc_sx); 1090 TAILQ_REMOVE(NOBJLIST(srcobject->handle), srcobject, 1091 pager_object_list); 1092 sx_xunlock(&sw_alloc_sx); 1093 VM_OBJECT_WLOCK(dstobject); 1094 VM_OBJECT_WLOCK(srcobject); 1095 } 1096 1097 /* 1098 * Transfer source to destination. 1099 */ 1100 swp_pager_meta_transfer(srcobject, dstobject, offset, dstobject->size); 1101 1102 /* 1103 * Free left over swap blocks in source. 1104 * 1105 * We have to revert the type to OBJT_DEFAULT so we do not accidentally 1106 * double-remove the object from the swap queues. 1107 */ 1108 if (destroysource) { 1109 swp_pager_meta_free_all(srcobject); 1110 /* 1111 * Reverting the type is not necessary, the caller is going 1112 * to destroy srcobject directly, but I'm doing it here 1113 * for consistency since we've removed the object from its 1114 * queues. 1115 */ 1116 srcobject->type = OBJT_DEFAULT; 1117 vm_object_clear_flag(srcobject, OBJ_SWAP); 1118 } 1119 } 1120 1121 /* 1122 * SWAP_PAGER_HASPAGE() - determine if we have good backing store for 1123 * the requested page. 1124 * 1125 * We determine whether good backing store exists for the requested 1126 * page and return TRUE if it does, FALSE if it doesn't. 1127 * 1128 * If TRUE, we also try to determine how much valid, contiguous backing 1129 * store exists before and after the requested page. 1130 */ 1131 static boolean_t 1132 swap_pager_haspage(vm_object_t object, vm_pindex_t pindex, int *before, 1133 int *after) 1134 { 1135 daddr_t blk, blk0; 1136 int i; 1137 1138 VM_OBJECT_ASSERT_LOCKED(object); 1139 KASSERT((object->flags & OBJ_SWAP) != 0, 1140 ("%s: object not swappable", __func__)); 1141 1142 /* 1143 * do we have good backing store at the requested index ? 1144 */ 1145 blk0 = swp_pager_meta_lookup(object, pindex); 1146 if (blk0 == SWAPBLK_NONE) { 1147 if (before) 1148 *before = 0; 1149 if (after) 1150 *after = 0; 1151 return (FALSE); 1152 } 1153 1154 /* 1155 * find backwards-looking contiguous good backing store 1156 */ 1157 if (before != NULL) { 1158 for (i = 1; i < SWB_NPAGES; i++) { 1159 if (i > pindex) 1160 break; 1161 blk = swp_pager_meta_lookup(object, pindex - i); 1162 if (blk != blk0 - i) 1163 break; 1164 } 1165 *before = i - 1; 1166 } 1167 1168 /* 1169 * find forward-looking contiguous good backing store 1170 */ 1171 if (after != NULL) { 1172 for (i = 1; i < SWB_NPAGES; i++) { 1173 blk = swp_pager_meta_lookup(object, pindex + i); 1174 if (blk != blk0 + i) 1175 break; 1176 } 1177 *after = i - 1; 1178 } 1179 return (TRUE); 1180 } 1181 1182 /* 1183 * SWAP_PAGER_PAGE_UNSWAPPED() - remove swap backing store related to page 1184 * 1185 * This removes any associated swap backing store, whether valid or 1186 * not, from the page. 1187 * 1188 * This routine is typically called when a page is made dirty, at 1189 * which point any associated swap can be freed. MADV_FREE also 1190 * calls us in a special-case situation 1191 * 1192 * NOTE!!! If the page is clean and the swap was valid, the caller 1193 * should make the page dirty before calling this routine. This routine 1194 * does NOT change the m->dirty status of the page. Also: MADV_FREE 1195 * depends on it. 1196 * 1197 * This routine may not sleep. 1198 * 1199 * The object containing the page may be locked. 1200 */ 1201 static void 1202 swap_pager_unswapped(vm_page_t m) 1203 { 1204 struct swblk *sb; 1205 vm_object_t obj; 1206 1207 /* 1208 * Handle enqueing deferred frees first. If we do not have the 1209 * object lock we wait for the page daemon to clear the space. 1210 */ 1211 obj = m->object; 1212 if (!VM_OBJECT_WOWNED(obj)) { 1213 VM_PAGE_OBJECT_BUSY_ASSERT(m); 1214 /* 1215 * The caller is responsible for synchronization but we 1216 * will harmlessly handle races. This is typically provided 1217 * by only calling unswapped() when a page transitions from 1218 * clean to dirty. 1219 */ 1220 if ((m->a.flags & (PGA_SWAP_SPACE | PGA_SWAP_FREE)) == 1221 PGA_SWAP_SPACE) { 1222 vm_page_aflag_set(m, PGA_SWAP_FREE); 1223 counter_u64_add(swap_free_deferred, 1); 1224 } 1225 return; 1226 } 1227 if ((m->a.flags & PGA_SWAP_FREE) != 0) 1228 counter_u64_add(swap_free_completed, 1); 1229 vm_page_aflag_clear(m, PGA_SWAP_FREE | PGA_SWAP_SPACE); 1230 1231 /* 1232 * The meta data only exists if the object is OBJT_SWAP 1233 * and even then might not be allocated yet. 1234 */ 1235 KASSERT((m->object->flags & OBJ_SWAP) != 0, 1236 ("Free object not swappable")); 1237 1238 sb = SWAP_PCTRIE_LOOKUP(&m->object->un_pager.swp.swp_blks, 1239 rounddown(m->pindex, SWAP_META_PAGES)); 1240 if (sb == NULL) 1241 return; 1242 if (sb->d[m->pindex % SWAP_META_PAGES] == SWAPBLK_NONE) 1243 return; 1244 swp_pager_freeswapspace(sb->d[m->pindex % SWAP_META_PAGES], 1); 1245 sb->d[m->pindex % SWAP_META_PAGES] = SWAPBLK_NONE; 1246 swp_pager_free_empty_swblk(m->object, sb); 1247 } 1248 1249 /* 1250 * swap_pager_getpages() - bring pages in from swap 1251 * 1252 * Attempt to page in the pages in array "ma" of length "count". The 1253 * caller may optionally specify that additional pages preceding and 1254 * succeeding the specified range be paged in. The number of such pages 1255 * is returned in the "rbehind" and "rahead" parameters, and they will 1256 * be in the inactive queue upon return. 1257 * 1258 * The pages in "ma" must be busied and will remain busied upon return. 1259 */ 1260 static int 1261 swap_pager_getpages_locked(vm_object_t object, vm_page_t *ma, int count, 1262 int *rbehind, int *rahead) 1263 { 1264 struct buf *bp; 1265 vm_page_t bm, mpred, msucc, p; 1266 vm_pindex_t pindex; 1267 daddr_t blk; 1268 int i, maxahead, maxbehind, reqcount; 1269 1270 VM_OBJECT_ASSERT_WLOCKED(object); 1271 reqcount = count; 1272 1273 KASSERT((object->flags & OBJ_SWAP) != 0, 1274 ("%s: object not swappable", __func__)); 1275 if (!swap_pager_haspage(object, ma[0]->pindex, &maxbehind, &maxahead)) { 1276 VM_OBJECT_WUNLOCK(object); 1277 return (VM_PAGER_FAIL); 1278 } 1279 1280 KASSERT(reqcount - 1 <= maxahead, 1281 ("page count %d extends beyond swap block", reqcount)); 1282 1283 /* 1284 * Do not transfer any pages other than those that are xbusied 1285 * when running during a split or collapse operation. This 1286 * prevents clustering from re-creating pages which are being 1287 * moved into another object. 1288 */ 1289 if ((object->flags & (OBJ_SPLIT | OBJ_DEAD)) != 0) { 1290 maxahead = reqcount - 1; 1291 maxbehind = 0; 1292 } 1293 1294 /* 1295 * Clip the readahead and readbehind ranges to exclude resident pages. 1296 */ 1297 if (rahead != NULL) { 1298 *rahead = imin(*rahead, maxahead - (reqcount - 1)); 1299 pindex = ma[reqcount - 1]->pindex; 1300 msucc = TAILQ_NEXT(ma[reqcount - 1], listq); 1301 if (msucc != NULL && msucc->pindex - pindex - 1 < *rahead) 1302 *rahead = msucc->pindex - pindex - 1; 1303 } 1304 if (rbehind != NULL) { 1305 *rbehind = imin(*rbehind, maxbehind); 1306 pindex = ma[0]->pindex; 1307 mpred = TAILQ_PREV(ma[0], pglist, listq); 1308 if (mpred != NULL && pindex - mpred->pindex - 1 < *rbehind) 1309 *rbehind = pindex - mpred->pindex - 1; 1310 } 1311 1312 bm = ma[0]; 1313 for (i = 0; i < count; i++) 1314 ma[i]->oflags |= VPO_SWAPINPROG; 1315 1316 /* 1317 * Allocate readahead and readbehind pages. 1318 */ 1319 if (rbehind != NULL) { 1320 for (i = 1; i <= *rbehind; i++) { 1321 p = vm_page_alloc(object, ma[0]->pindex - i, 1322 VM_ALLOC_NORMAL); 1323 if (p == NULL) 1324 break; 1325 p->oflags |= VPO_SWAPINPROG; 1326 bm = p; 1327 } 1328 *rbehind = i - 1; 1329 } 1330 if (rahead != NULL) { 1331 for (i = 0; i < *rahead; i++) { 1332 p = vm_page_alloc(object, 1333 ma[reqcount - 1]->pindex + i + 1, VM_ALLOC_NORMAL); 1334 if (p == NULL) 1335 break; 1336 p->oflags |= VPO_SWAPINPROG; 1337 } 1338 *rahead = i; 1339 } 1340 if (rbehind != NULL) 1341 count += *rbehind; 1342 if (rahead != NULL) 1343 count += *rahead; 1344 1345 vm_object_pip_add(object, count); 1346 1347 pindex = bm->pindex; 1348 blk = swp_pager_meta_lookup(object, pindex); 1349 KASSERT(blk != SWAPBLK_NONE, 1350 ("no swap blocking containing %p(%jx)", object, (uintmax_t)pindex)); 1351 1352 VM_OBJECT_WUNLOCK(object); 1353 bp = uma_zalloc(swrbuf_zone, M_WAITOK); 1354 MPASS((bp->b_flags & B_MAXPHYS) != 0); 1355 /* Pages cannot leave the object while busy. */ 1356 for (i = 0, p = bm; i < count; i++, p = TAILQ_NEXT(p, listq)) { 1357 MPASS(p->pindex == bm->pindex + i); 1358 bp->b_pages[i] = p; 1359 } 1360 1361 bp->b_flags |= B_PAGING; 1362 bp->b_iocmd = BIO_READ; 1363 bp->b_iodone = swp_pager_async_iodone; 1364 bp->b_rcred = crhold(thread0.td_ucred); 1365 bp->b_wcred = crhold(thread0.td_ucred); 1366 bp->b_blkno = blk; 1367 bp->b_bcount = PAGE_SIZE * count; 1368 bp->b_bufsize = PAGE_SIZE * count; 1369 bp->b_npages = count; 1370 bp->b_pgbefore = rbehind != NULL ? *rbehind : 0; 1371 bp->b_pgafter = rahead != NULL ? *rahead : 0; 1372 1373 VM_CNT_INC(v_swapin); 1374 VM_CNT_ADD(v_swappgsin, count); 1375 1376 /* 1377 * perform the I/O. NOTE!!! bp cannot be considered valid after 1378 * this point because we automatically release it on completion. 1379 * Instead, we look at the one page we are interested in which we 1380 * still hold a lock on even through the I/O completion. 1381 * 1382 * The other pages in our ma[] array are also released on completion, 1383 * so we cannot assume they are valid anymore either. 1384 * 1385 * NOTE: b_blkno is destroyed by the call to swapdev_strategy 1386 */ 1387 BUF_KERNPROC(bp); 1388 swp_pager_strategy(bp); 1389 1390 /* 1391 * Wait for the pages we want to complete. VPO_SWAPINPROG is always 1392 * cleared on completion. If an I/O error occurs, SWAPBLK_NONE 1393 * is set in the metadata for each page in the request. 1394 */ 1395 VM_OBJECT_WLOCK(object); 1396 /* This could be implemented more efficiently with aflags */ 1397 while ((ma[0]->oflags & VPO_SWAPINPROG) != 0) { 1398 ma[0]->oflags |= VPO_SWAPSLEEP; 1399 VM_CNT_INC(v_intrans); 1400 if (VM_OBJECT_SLEEP(object, &object->handle, PSWP, 1401 "swread", hz * 20)) { 1402 printf( 1403 "swap_pager: indefinite wait buffer: bufobj: %p, blkno: %jd, size: %ld\n", 1404 bp->b_bufobj, (intmax_t)bp->b_blkno, bp->b_bcount); 1405 } 1406 } 1407 VM_OBJECT_WUNLOCK(object); 1408 1409 /* 1410 * If we had an unrecoverable read error pages will not be valid. 1411 */ 1412 for (i = 0; i < reqcount; i++) 1413 if (ma[i]->valid != VM_PAGE_BITS_ALL) 1414 return (VM_PAGER_ERROR); 1415 1416 return (VM_PAGER_OK); 1417 1418 /* 1419 * A final note: in a low swap situation, we cannot deallocate swap 1420 * and mark a page dirty here because the caller is likely to mark 1421 * the page clean when we return, causing the page to possibly revert 1422 * to all-zero's later. 1423 */ 1424 } 1425 1426 static int 1427 swap_pager_getpages(vm_object_t object, vm_page_t *ma, int count, 1428 int *rbehind, int *rahead) 1429 { 1430 1431 VM_OBJECT_WLOCK(object); 1432 return (swap_pager_getpages_locked(object, ma, count, rbehind, rahead)); 1433 } 1434 1435 /* 1436 * swap_pager_getpages_async(): 1437 * 1438 * Right now this is emulation of asynchronous operation on top of 1439 * swap_pager_getpages(). 1440 */ 1441 static int 1442 swap_pager_getpages_async(vm_object_t object, vm_page_t *ma, int count, 1443 int *rbehind, int *rahead, pgo_getpages_iodone_t iodone, void *arg) 1444 { 1445 int r, error; 1446 1447 r = swap_pager_getpages(object, ma, count, rbehind, rahead); 1448 switch (r) { 1449 case VM_PAGER_OK: 1450 error = 0; 1451 break; 1452 case VM_PAGER_ERROR: 1453 error = EIO; 1454 break; 1455 case VM_PAGER_FAIL: 1456 error = EINVAL; 1457 break; 1458 default: 1459 panic("unhandled swap_pager_getpages() error %d", r); 1460 } 1461 (iodone)(arg, ma, count, error); 1462 1463 return (r); 1464 } 1465 1466 /* 1467 * swap_pager_putpages: 1468 * 1469 * Assign swap (if necessary) and initiate I/O on the specified pages. 1470 * 1471 * We support both OBJT_DEFAULT and OBJT_SWAP objects. DEFAULT objects 1472 * are automatically converted to SWAP objects. 1473 * 1474 * In a low memory situation we may block in VOP_STRATEGY(), but the new 1475 * vm_page reservation system coupled with properly written VFS devices 1476 * should ensure that no low-memory deadlock occurs. This is an area 1477 * which needs work. 1478 * 1479 * The parent has N vm_object_pip_add() references prior to 1480 * calling us and will remove references for rtvals[] that are 1481 * not set to VM_PAGER_PEND. We need to remove the rest on I/O 1482 * completion. 1483 * 1484 * The parent has soft-busy'd the pages it passes us and will unbusy 1485 * those whose rtvals[] entry is not set to VM_PAGER_PEND on return. 1486 * We need to unbusy the rest on I/O completion. 1487 */ 1488 static void 1489 swap_pager_putpages(vm_object_t object, vm_page_t *ma, int count, 1490 int flags, int *rtvals) 1491 { 1492 struct buf *bp; 1493 daddr_t addr, blk, n_free, s_free; 1494 vm_page_t mreq; 1495 int i, j, n; 1496 bool async; 1497 1498 KASSERT(count == 0 || ma[0]->object == object, 1499 ("%s: object mismatch %p/%p", 1500 __func__, object, ma[0]->object)); 1501 1502 /* 1503 * Step 1 1504 * 1505 * Turn object into OBJT_SWAP. Force sync if not a pageout process. 1506 */ 1507 if ((object->flags & OBJ_SWAP) == 0) { 1508 addr = swp_pager_meta_build(object, 0, SWAPBLK_NONE); 1509 KASSERT(addr == SWAPBLK_NONE, 1510 ("unexpected object swap block")); 1511 } 1512 VM_OBJECT_WUNLOCK(object); 1513 async = curproc == pageproc && (flags & VM_PAGER_PUT_SYNC) == 0; 1514 swp_pager_init_freerange(&s_free, &n_free); 1515 1516 /* 1517 * Step 2 1518 * 1519 * Assign swap blocks and issue I/O. We reallocate swap on the fly. 1520 * The page is left dirty until the pageout operation completes 1521 * successfully. 1522 */ 1523 for (i = 0; i < count; i += n) { 1524 /* Maximum I/O size is limited by maximum swap block size. */ 1525 n = min(count - i, nsw_cluster_max); 1526 1527 if (async) { 1528 mtx_lock(&swbuf_mtx); 1529 while (nsw_wcount_async == 0) 1530 msleep(&nsw_wcount_async, &swbuf_mtx, PVM, 1531 "swbufa", 0); 1532 nsw_wcount_async--; 1533 mtx_unlock(&swbuf_mtx); 1534 } 1535 1536 /* Get a block of swap of size up to size n. */ 1537 VM_OBJECT_WLOCK(object); 1538 blk = swp_pager_getswapspace(&n); 1539 if (blk == SWAPBLK_NONE) { 1540 VM_OBJECT_WUNLOCK(object); 1541 mtx_lock(&swbuf_mtx); 1542 if (++nsw_wcount_async == 1) 1543 wakeup(&nsw_wcount_async); 1544 mtx_unlock(&swbuf_mtx); 1545 for (j = 0; j < n; ++j) 1546 rtvals[i + j] = VM_PAGER_FAIL; 1547 continue; 1548 } 1549 for (j = 0; j < n; ++j) { 1550 mreq = ma[i + j]; 1551 vm_page_aflag_clear(mreq, PGA_SWAP_FREE); 1552 addr = swp_pager_meta_build(mreq->object, mreq->pindex, 1553 blk + j); 1554 if (addr != SWAPBLK_NONE) 1555 swp_pager_update_freerange(&s_free, &n_free, 1556 addr); 1557 MPASS(mreq->dirty == VM_PAGE_BITS_ALL); 1558 mreq->oflags |= VPO_SWAPINPROG; 1559 } 1560 VM_OBJECT_WUNLOCK(object); 1561 1562 bp = uma_zalloc(swwbuf_zone, M_WAITOK); 1563 MPASS((bp->b_flags & B_MAXPHYS) != 0); 1564 if (async) 1565 bp->b_flags |= B_ASYNC; 1566 bp->b_flags |= B_PAGING; 1567 bp->b_iocmd = BIO_WRITE; 1568 1569 bp->b_rcred = crhold(thread0.td_ucred); 1570 bp->b_wcred = crhold(thread0.td_ucred); 1571 bp->b_bcount = PAGE_SIZE * n; 1572 bp->b_bufsize = PAGE_SIZE * n; 1573 bp->b_blkno = blk; 1574 for (j = 0; j < n; j++) 1575 bp->b_pages[j] = ma[i + j]; 1576 bp->b_npages = n; 1577 1578 /* 1579 * Must set dirty range for NFS to work. 1580 */ 1581 bp->b_dirtyoff = 0; 1582 bp->b_dirtyend = bp->b_bcount; 1583 1584 VM_CNT_INC(v_swapout); 1585 VM_CNT_ADD(v_swappgsout, bp->b_npages); 1586 1587 /* 1588 * We unconditionally set rtvals[] to VM_PAGER_PEND so that we 1589 * can call the async completion routine at the end of a 1590 * synchronous I/O operation. Otherwise, our caller would 1591 * perform duplicate unbusy and wakeup operations on the page 1592 * and object, respectively. 1593 */ 1594 for (j = 0; j < n; j++) 1595 rtvals[i + j] = VM_PAGER_PEND; 1596 1597 /* 1598 * asynchronous 1599 * 1600 * NOTE: b_blkno is destroyed by the call to swapdev_strategy. 1601 */ 1602 if (async) { 1603 bp->b_iodone = swp_pager_async_iodone; 1604 BUF_KERNPROC(bp); 1605 swp_pager_strategy(bp); 1606 continue; 1607 } 1608 1609 /* 1610 * synchronous 1611 * 1612 * NOTE: b_blkno is destroyed by the call to swapdev_strategy. 1613 */ 1614 bp->b_iodone = bdone; 1615 swp_pager_strategy(bp); 1616 1617 /* 1618 * Wait for the sync I/O to complete. 1619 */ 1620 bwait(bp, PVM, "swwrt"); 1621 1622 /* 1623 * Now that we are through with the bp, we can call the 1624 * normal async completion, which frees everything up. 1625 */ 1626 swp_pager_async_iodone(bp); 1627 } 1628 swp_pager_freeswapspace(s_free, n_free); 1629 VM_OBJECT_WLOCK(object); 1630 } 1631 1632 /* 1633 * swp_pager_async_iodone: 1634 * 1635 * Completion routine for asynchronous reads and writes from/to swap. 1636 * Also called manually by synchronous code to finish up a bp. 1637 * 1638 * This routine may not sleep. 1639 */ 1640 static void 1641 swp_pager_async_iodone(struct buf *bp) 1642 { 1643 int i; 1644 vm_object_t object = NULL; 1645 1646 /* 1647 * Report error - unless we ran out of memory, in which case 1648 * we've already logged it in swapgeom_strategy(). 1649 */ 1650 if (bp->b_ioflags & BIO_ERROR && bp->b_error != ENOMEM) { 1651 printf( 1652 "swap_pager: I/O error - %s failed; blkno %ld," 1653 "size %ld, error %d\n", 1654 ((bp->b_iocmd == BIO_READ) ? "pagein" : "pageout"), 1655 (long)bp->b_blkno, 1656 (long)bp->b_bcount, 1657 bp->b_error 1658 ); 1659 } 1660 1661 /* 1662 * remove the mapping for kernel virtual 1663 */ 1664 if (buf_mapped(bp)) 1665 pmap_qremove((vm_offset_t)bp->b_data, bp->b_npages); 1666 else 1667 bp->b_data = bp->b_kvabase; 1668 1669 if (bp->b_npages) { 1670 object = bp->b_pages[0]->object; 1671 VM_OBJECT_WLOCK(object); 1672 } 1673 1674 /* 1675 * cleanup pages. If an error occurs writing to swap, we are in 1676 * very serious trouble. If it happens to be a disk error, though, 1677 * we may be able to recover by reassigning the swap later on. So 1678 * in this case we remove the m->swapblk assignment for the page 1679 * but do not free it in the rlist. The errornous block(s) are thus 1680 * never reallocated as swap. Redirty the page and continue. 1681 */ 1682 for (i = 0; i < bp->b_npages; ++i) { 1683 vm_page_t m = bp->b_pages[i]; 1684 1685 m->oflags &= ~VPO_SWAPINPROG; 1686 if (m->oflags & VPO_SWAPSLEEP) { 1687 m->oflags &= ~VPO_SWAPSLEEP; 1688 wakeup(&object->handle); 1689 } 1690 1691 /* We always have space after I/O, successful or not. */ 1692 vm_page_aflag_set(m, PGA_SWAP_SPACE); 1693 1694 if (bp->b_ioflags & BIO_ERROR) { 1695 /* 1696 * If an error occurs I'd love to throw the swapblk 1697 * away without freeing it back to swapspace, so it 1698 * can never be used again. But I can't from an 1699 * interrupt. 1700 */ 1701 if (bp->b_iocmd == BIO_READ) { 1702 /* 1703 * NOTE: for reads, m->dirty will probably 1704 * be overridden by the original caller of 1705 * getpages so don't play cute tricks here. 1706 */ 1707 vm_page_invalid(m); 1708 } else { 1709 /* 1710 * If a write error occurs, reactivate page 1711 * so it doesn't clog the inactive list, 1712 * then finish the I/O. 1713 */ 1714 MPASS(m->dirty == VM_PAGE_BITS_ALL); 1715 1716 /* PQ_UNSWAPPABLE? */ 1717 vm_page_activate(m); 1718 vm_page_sunbusy(m); 1719 } 1720 } else if (bp->b_iocmd == BIO_READ) { 1721 /* 1722 * NOTE: for reads, m->dirty will probably be 1723 * overridden by the original caller of getpages so 1724 * we cannot set them in order to free the underlying 1725 * swap in a low-swap situation. I don't think we'd 1726 * want to do that anyway, but it was an optimization 1727 * that existed in the old swapper for a time before 1728 * it got ripped out due to precisely this problem. 1729 */ 1730 KASSERT(!pmap_page_is_mapped(m), 1731 ("swp_pager_async_iodone: page %p is mapped", m)); 1732 KASSERT(m->dirty == 0, 1733 ("swp_pager_async_iodone: page %p is dirty", m)); 1734 1735 vm_page_valid(m); 1736 if (i < bp->b_pgbefore || 1737 i >= bp->b_npages - bp->b_pgafter) 1738 vm_page_readahead_finish(m); 1739 } else { 1740 /* 1741 * For write success, clear the dirty 1742 * status, then finish the I/O ( which decrements the 1743 * busy count and possibly wakes waiter's up ). 1744 * A page is only written to swap after a period of 1745 * inactivity. Therefore, we do not expect it to be 1746 * reused. 1747 */ 1748 KASSERT(!pmap_page_is_write_mapped(m), 1749 ("swp_pager_async_iodone: page %p is not write" 1750 " protected", m)); 1751 vm_page_undirty(m); 1752 vm_page_deactivate_noreuse(m); 1753 vm_page_sunbusy(m); 1754 } 1755 } 1756 1757 /* 1758 * adjust pip. NOTE: the original parent may still have its own 1759 * pip refs on the object. 1760 */ 1761 if (object != NULL) { 1762 vm_object_pip_wakeupn(object, bp->b_npages); 1763 VM_OBJECT_WUNLOCK(object); 1764 } 1765 1766 /* 1767 * swapdev_strategy() manually sets b_vp and b_bufobj before calling 1768 * bstrategy(). Set them back to NULL now we're done with it, or we'll 1769 * trigger a KASSERT in relpbuf(). 1770 */ 1771 if (bp->b_vp) { 1772 bp->b_vp = NULL; 1773 bp->b_bufobj = NULL; 1774 } 1775 /* 1776 * release the physical I/O buffer 1777 */ 1778 if (bp->b_flags & B_ASYNC) { 1779 mtx_lock(&swbuf_mtx); 1780 if (++nsw_wcount_async == 1) 1781 wakeup(&nsw_wcount_async); 1782 mtx_unlock(&swbuf_mtx); 1783 } 1784 uma_zfree((bp->b_iocmd == BIO_READ) ? swrbuf_zone : swwbuf_zone, bp); 1785 } 1786 1787 int 1788 swap_pager_nswapdev(void) 1789 { 1790 1791 return (nswapdev); 1792 } 1793 1794 static void 1795 swp_pager_force_dirty(vm_page_t m) 1796 { 1797 1798 vm_page_dirty(m); 1799 swap_pager_unswapped(m); 1800 vm_page_launder(m); 1801 } 1802 1803 u_long 1804 swap_pager_swapped_pages(vm_object_t object) 1805 { 1806 struct swblk *sb; 1807 vm_pindex_t pi; 1808 u_long res; 1809 int i; 1810 1811 VM_OBJECT_ASSERT_LOCKED(object); 1812 if ((object->flags & OBJ_SWAP) == 0) 1813 return (0); 1814 1815 for (res = 0, pi = 0; (sb = SWAP_PCTRIE_LOOKUP_GE( 1816 &object->un_pager.swp.swp_blks, pi)) != NULL; 1817 pi = sb->p + SWAP_META_PAGES) { 1818 for (i = 0; i < SWAP_META_PAGES; i++) { 1819 if (sb->d[i] != SWAPBLK_NONE) 1820 res++; 1821 } 1822 } 1823 return (res); 1824 } 1825 1826 /* 1827 * swap_pager_swapoff_object: 1828 * 1829 * Page in all of the pages that have been paged out for an object 1830 * to a swap device. 1831 */ 1832 static void 1833 swap_pager_swapoff_object(struct swdevt *sp, vm_object_t object) 1834 { 1835 struct swblk *sb; 1836 vm_page_t m; 1837 vm_pindex_t pi; 1838 daddr_t blk; 1839 int i, nv, rahead, rv; 1840 1841 KASSERT((object->flags & OBJ_SWAP) != 0, 1842 ("%s: Object not swappable", __func__)); 1843 1844 for (pi = 0; (sb = SWAP_PCTRIE_LOOKUP_GE( 1845 &object->un_pager.swp.swp_blks, pi)) != NULL; ) { 1846 if ((object->flags & OBJ_DEAD) != 0) { 1847 /* 1848 * Make sure that pending writes finish before 1849 * returning. 1850 */ 1851 vm_object_pip_wait(object, "swpoff"); 1852 swp_pager_meta_free_all(object); 1853 break; 1854 } 1855 for (i = 0; i < SWAP_META_PAGES; i++) { 1856 /* 1857 * Count the number of contiguous valid blocks. 1858 */ 1859 for (nv = 0; nv < SWAP_META_PAGES - i; nv++) { 1860 blk = sb->d[i + nv]; 1861 if (!swp_pager_isondev(blk, sp) || 1862 blk == SWAPBLK_NONE) 1863 break; 1864 } 1865 if (nv == 0) 1866 continue; 1867 1868 /* 1869 * Look for a page corresponding to the first 1870 * valid block and ensure that any pending paging 1871 * operations on it are complete. If the page is valid, 1872 * mark it dirty and free the swap block. Try to batch 1873 * this operation since it may cause sp to be freed, 1874 * meaning that we must restart the scan. Avoid busying 1875 * valid pages since we may block forever on kernel 1876 * stack pages. 1877 */ 1878 m = vm_page_lookup(object, sb->p + i); 1879 if (m == NULL) { 1880 m = vm_page_alloc(object, sb->p + i, 1881 VM_ALLOC_NORMAL | VM_ALLOC_WAITFAIL); 1882 if (m == NULL) 1883 break; 1884 } else { 1885 if ((m->oflags & VPO_SWAPINPROG) != 0) { 1886 m->oflags |= VPO_SWAPSLEEP; 1887 VM_OBJECT_SLEEP(object, &object->handle, 1888 PSWP, "swpoff", 0); 1889 break; 1890 } 1891 if (vm_page_all_valid(m)) { 1892 do { 1893 swp_pager_force_dirty(m); 1894 } while (--nv > 0 && 1895 (m = vm_page_next(m)) != NULL && 1896 vm_page_all_valid(m) && 1897 (m->oflags & VPO_SWAPINPROG) == 0); 1898 break; 1899 } 1900 if (!vm_page_busy_acquire(m, VM_ALLOC_WAITFAIL)) 1901 break; 1902 } 1903 1904 vm_object_pip_add(object, 1); 1905 rahead = SWAP_META_PAGES; 1906 rv = swap_pager_getpages_locked(object, &m, 1, NULL, 1907 &rahead); 1908 if (rv != VM_PAGER_OK) 1909 panic("%s: read from swap failed: %d", 1910 __func__, rv); 1911 vm_object_pip_wakeupn(object, 1); 1912 VM_OBJECT_WLOCK(object); 1913 vm_page_xunbusy(m); 1914 1915 /* 1916 * The object lock was dropped so we must restart the 1917 * scan of this swap block. Pages paged in during this 1918 * iteration will be marked dirty in a future iteration. 1919 */ 1920 break; 1921 } 1922 if (i == SWAP_META_PAGES) 1923 pi = sb->p + SWAP_META_PAGES; 1924 } 1925 } 1926 1927 /* 1928 * swap_pager_swapoff: 1929 * 1930 * Page in all of the pages that have been paged out to the 1931 * given device. The corresponding blocks in the bitmap must be 1932 * marked as allocated and the device must be flagged SW_CLOSING. 1933 * There may be no processes swapped out to the device. 1934 * 1935 * This routine may block. 1936 */ 1937 static void 1938 swap_pager_swapoff(struct swdevt *sp) 1939 { 1940 vm_object_t object; 1941 int retries; 1942 1943 sx_assert(&swdev_syscall_lock, SA_XLOCKED); 1944 1945 retries = 0; 1946 full_rescan: 1947 mtx_lock(&vm_object_list_mtx); 1948 TAILQ_FOREACH(object, &vm_object_list, object_list) { 1949 if ((object->flags & OBJ_SWAP) == 0) 1950 continue; 1951 mtx_unlock(&vm_object_list_mtx); 1952 /* Depends on type-stability. */ 1953 VM_OBJECT_WLOCK(object); 1954 1955 /* 1956 * Dead objects are eventually terminated on their own. 1957 */ 1958 if ((object->flags & OBJ_DEAD) != 0) 1959 goto next_obj; 1960 1961 /* 1962 * Sync with fences placed after pctrie 1963 * initialization. We must not access pctrie below 1964 * unless we checked that our object is swap and not 1965 * dead. 1966 */ 1967 atomic_thread_fence_acq(); 1968 if ((object->flags & OBJ_SWAP) == 0) 1969 goto next_obj; 1970 1971 swap_pager_swapoff_object(sp, object); 1972 next_obj: 1973 VM_OBJECT_WUNLOCK(object); 1974 mtx_lock(&vm_object_list_mtx); 1975 } 1976 mtx_unlock(&vm_object_list_mtx); 1977 1978 if (sp->sw_used) { 1979 /* 1980 * Objects may be locked or paging to the device being 1981 * removed, so we will miss their pages and need to 1982 * make another pass. We have marked this device as 1983 * SW_CLOSING, so the activity should finish soon. 1984 */ 1985 retries++; 1986 if (retries > 100) { 1987 panic("swapoff: failed to locate %d swap blocks", 1988 sp->sw_used); 1989 } 1990 pause("swpoff", hz / 20); 1991 goto full_rescan; 1992 } 1993 EVENTHANDLER_INVOKE(swapoff, sp); 1994 } 1995 1996 /************************************************************************ 1997 * SWAP META DATA * 1998 ************************************************************************ 1999 * 2000 * These routines manipulate the swap metadata stored in the 2001 * OBJT_SWAP object. 2002 * 2003 * Swap metadata is implemented with a global hash and not directly 2004 * linked into the object. Instead the object simply contains 2005 * appropriate tracking counters. 2006 */ 2007 2008 /* 2009 * SWP_PAGER_SWBLK_EMPTY() - is a range of blocks free? 2010 */ 2011 static bool 2012 swp_pager_swblk_empty(struct swblk *sb, int start, int limit) 2013 { 2014 int i; 2015 2016 MPASS(0 <= start && start <= limit && limit <= SWAP_META_PAGES); 2017 for (i = start; i < limit; i++) { 2018 if (sb->d[i] != SWAPBLK_NONE) 2019 return (false); 2020 } 2021 return (true); 2022 } 2023 2024 /* 2025 * SWP_PAGER_FREE_EMPTY_SWBLK() - frees if a block is free 2026 * 2027 * Nothing is done if the block is still in use. 2028 */ 2029 static void 2030 swp_pager_free_empty_swblk(vm_object_t object, struct swblk *sb) 2031 { 2032 2033 if (swp_pager_swblk_empty(sb, 0, SWAP_META_PAGES)) { 2034 SWAP_PCTRIE_REMOVE(&object->un_pager.swp.swp_blks, sb->p); 2035 uma_zfree(swblk_zone, sb); 2036 } 2037 } 2038 2039 /* 2040 * SWP_PAGER_META_BUILD() - add swap block to swap meta data for object 2041 * 2042 * We first convert the object to a swap object if it is a default 2043 * object. 2044 * 2045 * The specified swapblk is added to the object's swap metadata. If 2046 * the swapblk is not valid, it is freed instead. Any previously 2047 * assigned swapblk is returned. 2048 */ 2049 static daddr_t 2050 swp_pager_meta_build(vm_object_t object, vm_pindex_t pindex, daddr_t swapblk) 2051 { 2052 static volatile int swblk_zone_exhausted, swpctrie_zone_exhausted; 2053 struct swblk *sb, *sb1; 2054 vm_pindex_t modpi, rdpi; 2055 daddr_t prev_swapblk; 2056 int error, i; 2057 2058 VM_OBJECT_ASSERT_WLOCKED(object); 2059 2060 /* 2061 * Convert default object to swap object if necessary 2062 */ 2063 if ((object->flags & OBJ_SWAP) == 0) { 2064 pctrie_init(&object->un_pager.swp.swp_blks); 2065 2066 /* 2067 * Ensure that swap_pager_swapoff()'s iteration over 2068 * object_list does not see a garbage pctrie. 2069 */ 2070 atomic_thread_fence_rel(); 2071 2072 object->type = OBJT_SWAP; 2073 vm_object_set_flag(object, OBJ_SWAP); 2074 object->un_pager.swp.writemappings = 0; 2075 KASSERT((object->flags & OBJ_ANON) != 0 || 2076 object->handle == NULL, 2077 ("default pager %p with handle %p", 2078 object, object->handle)); 2079 } 2080 2081 rdpi = rounddown(pindex, SWAP_META_PAGES); 2082 sb = SWAP_PCTRIE_LOOKUP(&object->un_pager.swp.swp_blks, rdpi); 2083 if (sb == NULL) { 2084 if (swapblk == SWAPBLK_NONE) 2085 return (SWAPBLK_NONE); 2086 for (;;) { 2087 sb = uma_zalloc(swblk_zone, M_NOWAIT | (curproc == 2088 pageproc ? M_USE_RESERVE : 0)); 2089 if (sb != NULL) { 2090 sb->p = rdpi; 2091 for (i = 0; i < SWAP_META_PAGES; i++) 2092 sb->d[i] = SWAPBLK_NONE; 2093 if (atomic_cmpset_int(&swblk_zone_exhausted, 2094 1, 0)) 2095 printf("swblk zone ok\n"); 2096 break; 2097 } 2098 VM_OBJECT_WUNLOCK(object); 2099 if (uma_zone_exhausted(swblk_zone)) { 2100 if (atomic_cmpset_int(&swblk_zone_exhausted, 2101 0, 1)) 2102 printf("swap blk zone exhausted, " 2103 "increase kern.maxswzone\n"); 2104 vm_pageout_oom(VM_OOM_SWAPZ); 2105 pause("swzonxb", 10); 2106 } else 2107 uma_zwait(swblk_zone); 2108 VM_OBJECT_WLOCK(object); 2109 sb = SWAP_PCTRIE_LOOKUP(&object->un_pager.swp.swp_blks, 2110 rdpi); 2111 if (sb != NULL) 2112 /* 2113 * Somebody swapped out a nearby page, 2114 * allocating swblk at the rdpi index, 2115 * while we dropped the object lock. 2116 */ 2117 goto allocated; 2118 } 2119 for (;;) { 2120 error = SWAP_PCTRIE_INSERT( 2121 &object->un_pager.swp.swp_blks, sb); 2122 if (error == 0) { 2123 if (atomic_cmpset_int(&swpctrie_zone_exhausted, 2124 1, 0)) 2125 printf("swpctrie zone ok\n"); 2126 break; 2127 } 2128 VM_OBJECT_WUNLOCK(object); 2129 if (uma_zone_exhausted(swpctrie_zone)) { 2130 if (atomic_cmpset_int(&swpctrie_zone_exhausted, 2131 0, 1)) 2132 printf("swap pctrie zone exhausted, " 2133 "increase kern.maxswzone\n"); 2134 vm_pageout_oom(VM_OOM_SWAPZ); 2135 pause("swzonxp", 10); 2136 } else 2137 uma_zwait(swpctrie_zone); 2138 VM_OBJECT_WLOCK(object); 2139 sb1 = SWAP_PCTRIE_LOOKUP(&object->un_pager.swp.swp_blks, 2140 rdpi); 2141 if (sb1 != NULL) { 2142 uma_zfree(swblk_zone, sb); 2143 sb = sb1; 2144 goto allocated; 2145 } 2146 } 2147 } 2148 allocated: 2149 MPASS(sb->p == rdpi); 2150 2151 modpi = pindex % SWAP_META_PAGES; 2152 /* Return prior contents of metadata. */ 2153 prev_swapblk = sb->d[modpi]; 2154 /* Enter block into metadata. */ 2155 sb->d[modpi] = swapblk; 2156 2157 /* 2158 * Free the swblk if we end up with the empty page run. 2159 */ 2160 if (swapblk == SWAPBLK_NONE) 2161 swp_pager_free_empty_swblk(object, sb); 2162 return (prev_swapblk); 2163 } 2164 2165 /* 2166 * SWP_PAGER_META_TRANSFER() - free a range of blocks in the srcobject's swap 2167 * metadata, or transfer it into dstobject. 2168 * 2169 * This routine will free swap metadata structures as they are cleaned 2170 * out. 2171 */ 2172 static void 2173 swp_pager_meta_transfer(vm_object_t srcobject, vm_object_t dstobject, 2174 vm_pindex_t pindex, vm_pindex_t count) 2175 { 2176 struct swblk *sb; 2177 daddr_t n_free, s_free; 2178 vm_pindex_t offset, last; 2179 int i, limit, start; 2180 2181 VM_OBJECT_ASSERT_WLOCKED(srcobject); 2182 if ((srcobject->flags & OBJ_SWAP) == 0 || count == 0) 2183 return; 2184 2185 swp_pager_init_freerange(&s_free, &n_free); 2186 offset = pindex; 2187 last = pindex + count; 2188 for (;;) { 2189 sb = SWAP_PCTRIE_LOOKUP_GE(&srcobject->un_pager.swp.swp_blks, 2190 rounddown(pindex, SWAP_META_PAGES)); 2191 if (sb == NULL || sb->p >= last) 2192 break; 2193 start = pindex > sb->p ? pindex - sb->p : 0; 2194 limit = last - sb->p < SWAP_META_PAGES ? last - sb->p : 2195 SWAP_META_PAGES; 2196 for (i = start; i < limit; i++) { 2197 if (sb->d[i] == SWAPBLK_NONE) 2198 continue; 2199 if (dstobject == NULL || 2200 !swp_pager_xfer_source(srcobject, dstobject, 2201 sb->p + i - offset, sb->d[i])) { 2202 swp_pager_update_freerange(&s_free, &n_free, 2203 sb->d[i]); 2204 } 2205 sb->d[i] = SWAPBLK_NONE; 2206 } 2207 pindex = sb->p + SWAP_META_PAGES; 2208 if (swp_pager_swblk_empty(sb, 0, start) && 2209 swp_pager_swblk_empty(sb, limit, SWAP_META_PAGES)) { 2210 SWAP_PCTRIE_REMOVE(&srcobject->un_pager.swp.swp_blks, 2211 sb->p); 2212 uma_zfree(swblk_zone, sb); 2213 } 2214 } 2215 swp_pager_freeswapspace(s_free, n_free); 2216 } 2217 2218 /* 2219 * SWP_PAGER_META_FREE() - free a range of blocks in the object's swap metadata 2220 * 2221 * The requested range of blocks is freed, with any associated swap 2222 * returned to the swap bitmap. 2223 * 2224 * This routine will free swap metadata structures as they are cleaned 2225 * out. This routine does *NOT* operate on swap metadata associated 2226 * with resident pages. 2227 */ 2228 static void 2229 swp_pager_meta_free(vm_object_t object, vm_pindex_t pindex, vm_pindex_t count) 2230 { 2231 swp_pager_meta_transfer(object, NULL, pindex, count); 2232 } 2233 2234 /* 2235 * SWP_PAGER_META_FREE_ALL() - destroy all swap metadata associated with object 2236 * 2237 * This routine locates and destroys all swap metadata associated with 2238 * an object. 2239 */ 2240 static void 2241 swp_pager_meta_free_all(vm_object_t object) 2242 { 2243 struct swblk *sb; 2244 daddr_t n_free, s_free; 2245 vm_pindex_t pindex; 2246 int i; 2247 2248 VM_OBJECT_ASSERT_WLOCKED(object); 2249 if ((object->flags & OBJ_SWAP) == 0) 2250 return; 2251 2252 swp_pager_init_freerange(&s_free, &n_free); 2253 for (pindex = 0; (sb = SWAP_PCTRIE_LOOKUP_GE( 2254 &object->un_pager.swp.swp_blks, pindex)) != NULL;) { 2255 pindex = sb->p + SWAP_META_PAGES; 2256 for (i = 0; i < SWAP_META_PAGES; i++) { 2257 if (sb->d[i] == SWAPBLK_NONE) 2258 continue; 2259 swp_pager_update_freerange(&s_free, &n_free, sb->d[i]); 2260 } 2261 SWAP_PCTRIE_REMOVE(&object->un_pager.swp.swp_blks, sb->p); 2262 uma_zfree(swblk_zone, sb); 2263 } 2264 swp_pager_freeswapspace(s_free, n_free); 2265 } 2266 2267 /* 2268 * SWP_PAGER_METACTL() - misc control of swap meta data. 2269 * 2270 * This routine is capable of looking up, or removing swapblk 2271 * assignments in the swap meta data. It returns the swapblk being 2272 * looked-up, popped, or SWAPBLK_NONE if the block was invalid. 2273 * 2274 * When acting on a busy resident page and paging is in progress, we 2275 * have to wait until paging is complete but otherwise can act on the 2276 * busy page. 2277 */ 2278 static daddr_t 2279 swp_pager_meta_lookup(vm_object_t object, vm_pindex_t pindex) 2280 { 2281 struct swblk *sb; 2282 2283 VM_OBJECT_ASSERT_LOCKED(object); 2284 2285 /* 2286 * The meta data only exists if the object is OBJT_SWAP 2287 * and even then might not be allocated yet. 2288 */ 2289 KASSERT((object->flags & OBJ_SWAP) != 0, 2290 ("Lookup object not swappable")); 2291 2292 sb = SWAP_PCTRIE_LOOKUP(&object->un_pager.swp.swp_blks, 2293 rounddown(pindex, SWAP_META_PAGES)); 2294 if (sb == NULL) 2295 return (SWAPBLK_NONE); 2296 return (sb->d[pindex % SWAP_META_PAGES]); 2297 } 2298 2299 /* 2300 * Returns the least page index which is greater than or equal to the 2301 * parameter pindex and for which there is a swap block allocated. 2302 * Returns object's size if the object's type is not swap or if there 2303 * are no allocated swap blocks for the object after the requested 2304 * pindex. 2305 */ 2306 vm_pindex_t 2307 swap_pager_find_least(vm_object_t object, vm_pindex_t pindex) 2308 { 2309 struct swblk *sb; 2310 int i; 2311 2312 VM_OBJECT_ASSERT_LOCKED(object); 2313 if ((object->flags & OBJ_SWAP) == 0) 2314 return (object->size); 2315 2316 sb = SWAP_PCTRIE_LOOKUP_GE(&object->un_pager.swp.swp_blks, 2317 rounddown(pindex, SWAP_META_PAGES)); 2318 if (sb == NULL) 2319 return (object->size); 2320 if (sb->p < pindex) { 2321 for (i = pindex % SWAP_META_PAGES; i < SWAP_META_PAGES; i++) { 2322 if (sb->d[i] != SWAPBLK_NONE) 2323 return (sb->p + i); 2324 } 2325 sb = SWAP_PCTRIE_LOOKUP_GE(&object->un_pager.swp.swp_blks, 2326 roundup(pindex, SWAP_META_PAGES)); 2327 if (sb == NULL) 2328 return (object->size); 2329 } 2330 for (i = 0; i < SWAP_META_PAGES; i++) { 2331 if (sb->d[i] != SWAPBLK_NONE) 2332 return (sb->p + i); 2333 } 2334 2335 /* 2336 * We get here if a swblk is present in the trie but it 2337 * doesn't map any blocks. 2338 */ 2339 MPASS(0); 2340 return (object->size); 2341 } 2342 2343 /* 2344 * System call swapon(name) enables swapping on device name, 2345 * which must be in the swdevsw. Return EBUSY 2346 * if already swapping on this device. 2347 */ 2348 #ifndef _SYS_SYSPROTO_H_ 2349 struct swapon_args { 2350 char *name; 2351 }; 2352 #endif 2353 2354 /* 2355 * MPSAFE 2356 */ 2357 /* ARGSUSED */ 2358 int 2359 sys_swapon(struct thread *td, struct swapon_args *uap) 2360 { 2361 struct vattr attr; 2362 struct vnode *vp; 2363 struct nameidata nd; 2364 int error; 2365 2366 error = priv_check(td, PRIV_SWAPON); 2367 if (error) 2368 return (error); 2369 2370 sx_xlock(&swdev_syscall_lock); 2371 2372 /* 2373 * Swap metadata may not fit in the KVM if we have physical 2374 * memory of >1GB. 2375 */ 2376 if (swblk_zone == NULL) { 2377 error = ENOMEM; 2378 goto done; 2379 } 2380 2381 NDINIT(&nd, LOOKUP, ISOPEN | FOLLOW | AUDITVNODE1, UIO_USERSPACE, 2382 uap->name, td); 2383 error = namei(&nd); 2384 if (error) 2385 goto done; 2386 2387 NDFREE(&nd, NDF_ONLY_PNBUF); 2388 vp = nd.ni_vp; 2389 2390 if (vn_isdisk_error(vp, &error)) { 2391 error = swapongeom(vp); 2392 } else if (vp->v_type == VREG && 2393 (vp->v_mount->mnt_vfc->vfc_flags & VFCF_NETWORK) != 0 && 2394 (error = VOP_GETATTR(vp, &attr, td->td_ucred)) == 0) { 2395 /* 2396 * Allow direct swapping to NFS regular files in the same 2397 * way that nfs_mountroot() sets up diskless swapping. 2398 */ 2399 error = swaponvp(td, vp, attr.va_size / DEV_BSIZE); 2400 } 2401 2402 if (error) 2403 vrele(vp); 2404 done: 2405 sx_xunlock(&swdev_syscall_lock); 2406 return (error); 2407 } 2408 2409 /* 2410 * Check that the total amount of swap currently configured does not 2411 * exceed half the theoretical maximum. If it does, print a warning 2412 * message. 2413 */ 2414 static void 2415 swapon_check_swzone(void) 2416 { 2417 2418 /* recommend using no more than half that amount */ 2419 if (swap_total > swap_maxpages / 2) { 2420 printf("warning: total configured swap (%lu pages) " 2421 "exceeds maximum recommended amount (%lu pages).\n", 2422 swap_total, swap_maxpages / 2); 2423 printf("warning: increase kern.maxswzone " 2424 "or reduce amount of swap.\n"); 2425 } 2426 } 2427 2428 static void 2429 swaponsomething(struct vnode *vp, void *id, u_long nblks, 2430 sw_strategy_t *strategy, sw_close_t *close, dev_t dev, int flags) 2431 { 2432 struct swdevt *sp, *tsp; 2433 daddr_t dvbase; 2434 2435 /* 2436 * nblks is in DEV_BSIZE'd chunks, convert to PAGE_SIZE'd chunks. 2437 * First chop nblks off to page-align it, then convert. 2438 * 2439 * sw->sw_nblks is in page-sized chunks now too. 2440 */ 2441 nblks &= ~(ctodb(1) - 1); 2442 nblks = dbtoc(nblks); 2443 2444 sp = malloc(sizeof *sp, M_VMPGDATA, M_WAITOK | M_ZERO); 2445 sp->sw_blist = blist_create(nblks, M_WAITOK); 2446 sp->sw_vp = vp; 2447 sp->sw_id = id; 2448 sp->sw_dev = dev; 2449 sp->sw_nblks = nblks; 2450 sp->sw_used = 0; 2451 sp->sw_strategy = strategy; 2452 sp->sw_close = close; 2453 sp->sw_flags = flags; 2454 2455 /* 2456 * Do not free the first blocks in order to avoid overwriting 2457 * any bsd label at the front of the partition 2458 */ 2459 blist_free(sp->sw_blist, howmany(BBSIZE, PAGE_SIZE), 2460 nblks - howmany(BBSIZE, PAGE_SIZE)); 2461 2462 dvbase = 0; 2463 mtx_lock(&sw_dev_mtx); 2464 TAILQ_FOREACH(tsp, &swtailq, sw_list) { 2465 if (tsp->sw_end >= dvbase) { 2466 /* 2467 * We put one uncovered page between the devices 2468 * in order to definitively prevent any cross-device 2469 * I/O requests 2470 */ 2471 dvbase = tsp->sw_end + 1; 2472 } 2473 } 2474 sp->sw_first = dvbase; 2475 sp->sw_end = dvbase + nblks; 2476 TAILQ_INSERT_TAIL(&swtailq, sp, sw_list); 2477 nswapdev++; 2478 swap_pager_avail += nblks - howmany(BBSIZE, PAGE_SIZE); 2479 swap_total += nblks; 2480 swapon_check_swzone(); 2481 swp_sizecheck(); 2482 mtx_unlock(&sw_dev_mtx); 2483 EVENTHANDLER_INVOKE(swapon, sp); 2484 } 2485 2486 /* 2487 * SYSCALL: swapoff(devname) 2488 * 2489 * Disable swapping on the given device. 2490 * 2491 * XXX: Badly designed system call: it should use a device index 2492 * rather than filename as specification. We keep sw_vp around 2493 * only to make this work. 2494 */ 2495 #ifndef _SYS_SYSPROTO_H_ 2496 struct swapoff_args { 2497 char *name; 2498 }; 2499 #endif 2500 2501 /* 2502 * MPSAFE 2503 */ 2504 /* ARGSUSED */ 2505 int 2506 sys_swapoff(struct thread *td, struct swapoff_args *uap) 2507 { 2508 struct vnode *vp; 2509 struct nameidata nd; 2510 struct swdevt *sp; 2511 int error; 2512 2513 error = priv_check(td, PRIV_SWAPOFF); 2514 if (error) 2515 return (error); 2516 2517 sx_xlock(&swdev_syscall_lock); 2518 2519 NDINIT(&nd, LOOKUP, FOLLOW | AUDITVNODE1, UIO_USERSPACE, uap->name, 2520 td); 2521 error = namei(&nd); 2522 if (error) 2523 goto done; 2524 NDFREE(&nd, NDF_ONLY_PNBUF); 2525 vp = nd.ni_vp; 2526 2527 mtx_lock(&sw_dev_mtx); 2528 TAILQ_FOREACH(sp, &swtailq, sw_list) { 2529 if (sp->sw_vp == vp) 2530 break; 2531 } 2532 mtx_unlock(&sw_dev_mtx); 2533 if (sp == NULL) { 2534 error = EINVAL; 2535 goto done; 2536 } 2537 error = swapoff_one(sp, td->td_ucred); 2538 done: 2539 sx_xunlock(&swdev_syscall_lock); 2540 return (error); 2541 } 2542 2543 static int 2544 swapoff_one(struct swdevt *sp, struct ucred *cred) 2545 { 2546 u_long nblks; 2547 #ifdef MAC 2548 int error; 2549 #endif 2550 2551 sx_assert(&swdev_syscall_lock, SA_XLOCKED); 2552 #ifdef MAC 2553 (void) vn_lock(sp->sw_vp, LK_EXCLUSIVE | LK_RETRY); 2554 error = mac_system_check_swapoff(cred, sp->sw_vp); 2555 (void) VOP_UNLOCK(sp->sw_vp); 2556 if (error != 0) 2557 return (error); 2558 #endif 2559 nblks = sp->sw_nblks; 2560 2561 /* 2562 * We can turn off this swap device safely only if the 2563 * available virtual memory in the system will fit the amount 2564 * of data we will have to page back in, plus an epsilon so 2565 * the system doesn't become critically low on swap space. 2566 */ 2567 if (vm_free_count() + swap_pager_avail < nblks + nswap_lowat) 2568 return (ENOMEM); 2569 2570 /* 2571 * Prevent further allocations on this device. 2572 */ 2573 mtx_lock(&sw_dev_mtx); 2574 sp->sw_flags |= SW_CLOSING; 2575 swap_pager_avail -= blist_fill(sp->sw_blist, 0, nblks); 2576 swap_total -= nblks; 2577 mtx_unlock(&sw_dev_mtx); 2578 2579 /* 2580 * Page in the contents of the device and close it. 2581 */ 2582 swap_pager_swapoff(sp); 2583 2584 sp->sw_close(curthread, sp); 2585 mtx_lock(&sw_dev_mtx); 2586 sp->sw_id = NULL; 2587 TAILQ_REMOVE(&swtailq, sp, sw_list); 2588 nswapdev--; 2589 if (nswapdev == 0) { 2590 swap_pager_full = 2; 2591 swap_pager_almost_full = 1; 2592 } 2593 if (swdevhd == sp) 2594 swdevhd = NULL; 2595 mtx_unlock(&sw_dev_mtx); 2596 blist_destroy(sp->sw_blist); 2597 free(sp, M_VMPGDATA); 2598 return (0); 2599 } 2600 2601 void 2602 swapoff_all(void) 2603 { 2604 struct swdevt *sp, *spt; 2605 const char *devname; 2606 int error; 2607 2608 sx_xlock(&swdev_syscall_lock); 2609 2610 mtx_lock(&sw_dev_mtx); 2611 TAILQ_FOREACH_SAFE(sp, &swtailq, sw_list, spt) { 2612 mtx_unlock(&sw_dev_mtx); 2613 if (vn_isdisk(sp->sw_vp)) 2614 devname = devtoname(sp->sw_vp->v_rdev); 2615 else 2616 devname = "[file]"; 2617 error = swapoff_one(sp, thread0.td_ucred); 2618 if (error != 0) { 2619 printf("Cannot remove swap device %s (error=%d), " 2620 "skipping.\n", devname, error); 2621 } else if (bootverbose) { 2622 printf("Swap device %s removed.\n", devname); 2623 } 2624 mtx_lock(&sw_dev_mtx); 2625 } 2626 mtx_unlock(&sw_dev_mtx); 2627 2628 sx_xunlock(&swdev_syscall_lock); 2629 } 2630 2631 void 2632 swap_pager_status(int *total, int *used) 2633 { 2634 2635 *total = swap_total; 2636 *used = swap_total - swap_pager_avail - 2637 nswapdev * howmany(BBSIZE, PAGE_SIZE); 2638 } 2639 2640 int 2641 swap_dev_info(int name, struct xswdev *xs, char *devname, size_t len) 2642 { 2643 struct swdevt *sp; 2644 const char *tmp_devname; 2645 int error, n; 2646 2647 n = 0; 2648 error = ENOENT; 2649 mtx_lock(&sw_dev_mtx); 2650 TAILQ_FOREACH(sp, &swtailq, sw_list) { 2651 if (n != name) { 2652 n++; 2653 continue; 2654 } 2655 xs->xsw_version = XSWDEV_VERSION; 2656 xs->xsw_dev = sp->sw_dev; 2657 xs->xsw_flags = sp->sw_flags; 2658 xs->xsw_nblks = sp->sw_nblks; 2659 xs->xsw_used = sp->sw_used; 2660 if (devname != NULL) { 2661 if (vn_isdisk(sp->sw_vp)) 2662 tmp_devname = devtoname(sp->sw_vp->v_rdev); 2663 else 2664 tmp_devname = "[file]"; 2665 strncpy(devname, tmp_devname, len); 2666 } 2667 error = 0; 2668 break; 2669 } 2670 mtx_unlock(&sw_dev_mtx); 2671 return (error); 2672 } 2673 2674 #if defined(COMPAT_FREEBSD11) 2675 #define XSWDEV_VERSION_11 1 2676 struct xswdev11 { 2677 u_int xsw_version; 2678 uint32_t xsw_dev; 2679 int xsw_flags; 2680 int xsw_nblks; 2681 int xsw_used; 2682 }; 2683 #endif 2684 2685 #if defined(__amd64__) && defined(COMPAT_FREEBSD32) 2686 struct xswdev32 { 2687 u_int xsw_version; 2688 u_int xsw_dev1, xsw_dev2; 2689 int xsw_flags; 2690 int xsw_nblks; 2691 int xsw_used; 2692 }; 2693 #endif 2694 2695 static int 2696 sysctl_vm_swap_info(SYSCTL_HANDLER_ARGS) 2697 { 2698 struct xswdev xs; 2699 #if defined(__amd64__) && defined(COMPAT_FREEBSD32) 2700 struct xswdev32 xs32; 2701 #endif 2702 #if defined(COMPAT_FREEBSD11) 2703 struct xswdev11 xs11; 2704 #endif 2705 int error; 2706 2707 if (arg2 != 1) /* name length */ 2708 return (EINVAL); 2709 2710 memset(&xs, 0, sizeof(xs)); 2711 error = swap_dev_info(*(int *)arg1, &xs, NULL, 0); 2712 if (error != 0) 2713 return (error); 2714 #if defined(__amd64__) && defined(COMPAT_FREEBSD32) 2715 if (req->oldlen == sizeof(xs32)) { 2716 memset(&xs32, 0, sizeof(xs32)); 2717 xs32.xsw_version = XSWDEV_VERSION; 2718 xs32.xsw_dev1 = xs.xsw_dev; 2719 xs32.xsw_dev2 = xs.xsw_dev >> 32; 2720 xs32.xsw_flags = xs.xsw_flags; 2721 xs32.xsw_nblks = xs.xsw_nblks; 2722 xs32.xsw_used = xs.xsw_used; 2723 error = SYSCTL_OUT(req, &xs32, sizeof(xs32)); 2724 return (error); 2725 } 2726 #endif 2727 #if defined(COMPAT_FREEBSD11) 2728 if (req->oldlen == sizeof(xs11)) { 2729 memset(&xs11, 0, sizeof(xs11)); 2730 xs11.xsw_version = XSWDEV_VERSION_11; 2731 xs11.xsw_dev = xs.xsw_dev; /* truncation */ 2732 xs11.xsw_flags = xs.xsw_flags; 2733 xs11.xsw_nblks = xs.xsw_nblks; 2734 xs11.xsw_used = xs.xsw_used; 2735 error = SYSCTL_OUT(req, &xs11, sizeof(xs11)); 2736 return (error); 2737 } 2738 #endif 2739 error = SYSCTL_OUT(req, &xs, sizeof(xs)); 2740 return (error); 2741 } 2742 2743 SYSCTL_INT(_vm, OID_AUTO, nswapdev, CTLFLAG_RD, &nswapdev, 0, 2744 "Number of swap devices"); 2745 SYSCTL_NODE(_vm, OID_AUTO, swap_info, CTLFLAG_RD | CTLFLAG_MPSAFE, 2746 sysctl_vm_swap_info, 2747 "Swap statistics by device"); 2748 2749 /* 2750 * Count the approximate swap usage in pages for a vmspace. The 2751 * shadowed or not yet copied on write swap blocks are not accounted. 2752 * The map must be locked. 2753 */ 2754 long 2755 vmspace_swap_count(struct vmspace *vmspace) 2756 { 2757 vm_map_t map; 2758 vm_map_entry_t cur; 2759 vm_object_t object; 2760 struct swblk *sb; 2761 vm_pindex_t e, pi; 2762 long count; 2763 int i; 2764 2765 map = &vmspace->vm_map; 2766 count = 0; 2767 2768 VM_MAP_ENTRY_FOREACH(cur, map) { 2769 if ((cur->eflags & MAP_ENTRY_IS_SUB_MAP) != 0) 2770 continue; 2771 object = cur->object.vm_object; 2772 if (object == NULL || (object->flags & OBJ_SWAP) == 0) 2773 continue; 2774 VM_OBJECT_RLOCK(object); 2775 if ((object->flags & OBJ_SWAP) == 0) 2776 goto unlock; 2777 pi = OFF_TO_IDX(cur->offset); 2778 e = pi + OFF_TO_IDX(cur->end - cur->start); 2779 for (;; pi = sb->p + SWAP_META_PAGES) { 2780 sb = SWAP_PCTRIE_LOOKUP_GE( 2781 &object->un_pager.swp.swp_blks, pi); 2782 if (sb == NULL || sb->p >= e) 2783 break; 2784 for (i = 0; i < SWAP_META_PAGES; i++) { 2785 if (sb->p + i < e && 2786 sb->d[i] != SWAPBLK_NONE) 2787 count++; 2788 } 2789 } 2790 unlock: 2791 VM_OBJECT_RUNLOCK(object); 2792 } 2793 return (count); 2794 } 2795 2796 /* 2797 * GEOM backend 2798 * 2799 * Swapping onto disk devices. 2800 * 2801 */ 2802 2803 static g_orphan_t swapgeom_orphan; 2804 2805 static struct g_class g_swap_class = { 2806 .name = "SWAP", 2807 .version = G_VERSION, 2808 .orphan = swapgeom_orphan, 2809 }; 2810 2811 DECLARE_GEOM_CLASS(g_swap_class, g_class); 2812 2813 static void 2814 swapgeom_close_ev(void *arg, int flags) 2815 { 2816 struct g_consumer *cp; 2817 2818 cp = arg; 2819 g_access(cp, -1, -1, 0); 2820 g_detach(cp); 2821 g_destroy_consumer(cp); 2822 } 2823 2824 /* 2825 * Add a reference to the g_consumer for an inflight transaction. 2826 */ 2827 static void 2828 swapgeom_acquire(struct g_consumer *cp) 2829 { 2830 2831 mtx_assert(&sw_dev_mtx, MA_OWNED); 2832 cp->index++; 2833 } 2834 2835 /* 2836 * Remove a reference from the g_consumer. Post a close event if all 2837 * references go away, since the function might be called from the 2838 * biodone context. 2839 */ 2840 static void 2841 swapgeom_release(struct g_consumer *cp, struct swdevt *sp) 2842 { 2843 2844 mtx_assert(&sw_dev_mtx, MA_OWNED); 2845 cp->index--; 2846 if (cp->index == 0) { 2847 if (g_post_event(swapgeom_close_ev, cp, M_NOWAIT, NULL) == 0) 2848 sp->sw_id = NULL; 2849 } 2850 } 2851 2852 static void 2853 swapgeom_done(struct bio *bp2) 2854 { 2855 struct swdevt *sp; 2856 struct buf *bp; 2857 struct g_consumer *cp; 2858 2859 bp = bp2->bio_caller2; 2860 cp = bp2->bio_from; 2861 bp->b_ioflags = bp2->bio_flags; 2862 if (bp2->bio_error) 2863 bp->b_ioflags |= BIO_ERROR; 2864 bp->b_resid = bp->b_bcount - bp2->bio_completed; 2865 bp->b_error = bp2->bio_error; 2866 bp->b_caller1 = NULL; 2867 bufdone(bp); 2868 sp = bp2->bio_caller1; 2869 mtx_lock(&sw_dev_mtx); 2870 swapgeom_release(cp, sp); 2871 mtx_unlock(&sw_dev_mtx); 2872 g_destroy_bio(bp2); 2873 } 2874 2875 static void 2876 swapgeom_strategy(struct buf *bp, struct swdevt *sp) 2877 { 2878 struct bio *bio; 2879 struct g_consumer *cp; 2880 2881 mtx_lock(&sw_dev_mtx); 2882 cp = sp->sw_id; 2883 if (cp == NULL) { 2884 mtx_unlock(&sw_dev_mtx); 2885 bp->b_error = ENXIO; 2886 bp->b_ioflags |= BIO_ERROR; 2887 bufdone(bp); 2888 return; 2889 } 2890 swapgeom_acquire(cp); 2891 mtx_unlock(&sw_dev_mtx); 2892 if (bp->b_iocmd == BIO_WRITE) 2893 bio = g_new_bio(); 2894 else 2895 bio = g_alloc_bio(); 2896 if (bio == NULL) { 2897 mtx_lock(&sw_dev_mtx); 2898 swapgeom_release(cp, sp); 2899 mtx_unlock(&sw_dev_mtx); 2900 bp->b_error = ENOMEM; 2901 bp->b_ioflags |= BIO_ERROR; 2902 printf("swap_pager: cannot allocate bio\n"); 2903 bufdone(bp); 2904 return; 2905 } 2906 2907 bp->b_caller1 = bio; 2908 bio->bio_caller1 = sp; 2909 bio->bio_caller2 = bp; 2910 bio->bio_cmd = bp->b_iocmd; 2911 bio->bio_offset = (bp->b_blkno - sp->sw_first) * PAGE_SIZE; 2912 bio->bio_length = bp->b_bcount; 2913 bio->bio_done = swapgeom_done; 2914 if (!buf_mapped(bp)) { 2915 bio->bio_ma = bp->b_pages; 2916 bio->bio_data = unmapped_buf; 2917 bio->bio_ma_offset = (vm_offset_t)bp->b_offset & PAGE_MASK; 2918 bio->bio_ma_n = bp->b_npages; 2919 bio->bio_flags |= BIO_UNMAPPED; 2920 } else { 2921 bio->bio_data = bp->b_data; 2922 bio->bio_ma = NULL; 2923 } 2924 g_io_request(bio, cp); 2925 return; 2926 } 2927 2928 static void 2929 swapgeom_orphan(struct g_consumer *cp) 2930 { 2931 struct swdevt *sp; 2932 int destroy; 2933 2934 mtx_lock(&sw_dev_mtx); 2935 TAILQ_FOREACH(sp, &swtailq, sw_list) { 2936 if (sp->sw_id == cp) { 2937 sp->sw_flags |= SW_CLOSING; 2938 break; 2939 } 2940 } 2941 /* 2942 * Drop reference we were created with. Do directly since we're in a 2943 * special context where we don't have to queue the call to 2944 * swapgeom_close_ev(). 2945 */ 2946 cp->index--; 2947 destroy = ((sp != NULL) && (cp->index == 0)); 2948 if (destroy) 2949 sp->sw_id = NULL; 2950 mtx_unlock(&sw_dev_mtx); 2951 if (destroy) 2952 swapgeom_close_ev(cp, 0); 2953 } 2954 2955 static void 2956 swapgeom_close(struct thread *td, struct swdevt *sw) 2957 { 2958 struct g_consumer *cp; 2959 2960 mtx_lock(&sw_dev_mtx); 2961 cp = sw->sw_id; 2962 sw->sw_id = NULL; 2963 mtx_unlock(&sw_dev_mtx); 2964 2965 /* 2966 * swapgeom_close() may be called from the biodone context, 2967 * where we cannot perform topology changes. Delegate the 2968 * work to the events thread. 2969 */ 2970 if (cp != NULL) 2971 g_waitfor_event(swapgeom_close_ev, cp, M_WAITOK, NULL); 2972 } 2973 2974 static int 2975 swapongeom_locked(struct cdev *dev, struct vnode *vp) 2976 { 2977 struct g_provider *pp; 2978 struct g_consumer *cp; 2979 static struct g_geom *gp; 2980 struct swdevt *sp; 2981 u_long nblks; 2982 int error; 2983 2984 pp = g_dev_getprovider(dev); 2985 if (pp == NULL) 2986 return (ENODEV); 2987 mtx_lock(&sw_dev_mtx); 2988 TAILQ_FOREACH(sp, &swtailq, sw_list) { 2989 cp = sp->sw_id; 2990 if (cp != NULL && cp->provider == pp) { 2991 mtx_unlock(&sw_dev_mtx); 2992 return (EBUSY); 2993 } 2994 } 2995 mtx_unlock(&sw_dev_mtx); 2996 if (gp == NULL) 2997 gp = g_new_geomf(&g_swap_class, "swap"); 2998 cp = g_new_consumer(gp); 2999 cp->index = 1; /* Number of active I/Os, plus one for being active. */ 3000 cp->flags |= G_CF_DIRECT_SEND | G_CF_DIRECT_RECEIVE; 3001 g_attach(cp, pp); 3002 /* 3003 * XXX: Every time you think you can improve the margin for 3004 * footshooting, somebody depends on the ability to do so: 3005 * savecore(8) wants to write to our swapdev so we cannot 3006 * set an exclusive count :-( 3007 */ 3008 error = g_access(cp, 1, 1, 0); 3009 if (error != 0) { 3010 g_detach(cp); 3011 g_destroy_consumer(cp); 3012 return (error); 3013 } 3014 nblks = pp->mediasize / DEV_BSIZE; 3015 swaponsomething(vp, cp, nblks, swapgeom_strategy, 3016 swapgeom_close, dev2udev(dev), 3017 (pp->flags & G_PF_ACCEPT_UNMAPPED) != 0 ? SW_UNMAPPED : 0); 3018 return (0); 3019 } 3020 3021 static int 3022 swapongeom(struct vnode *vp) 3023 { 3024 int error; 3025 3026 vn_lock(vp, LK_EXCLUSIVE | LK_RETRY); 3027 if (vp->v_type != VCHR || VN_IS_DOOMED(vp)) { 3028 error = ENOENT; 3029 } else { 3030 g_topology_lock(); 3031 error = swapongeom_locked(vp->v_rdev, vp); 3032 g_topology_unlock(); 3033 } 3034 VOP_UNLOCK(vp); 3035 return (error); 3036 } 3037 3038 /* 3039 * VNODE backend 3040 * 3041 * This is used mainly for network filesystem (read: probably only tested 3042 * with NFS) swapfiles. 3043 * 3044 */ 3045 3046 static void 3047 swapdev_strategy(struct buf *bp, struct swdevt *sp) 3048 { 3049 struct vnode *vp2; 3050 3051 bp->b_blkno = ctodb(bp->b_blkno - sp->sw_first); 3052 3053 vp2 = sp->sw_id; 3054 vhold(vp2); 3055 if (bp->b_iocmd == BIO_WRITE) { 3056 if (bp->b_bufobj) 3057 bufobj_wdrop(bp->b_bufobj); 3058 bufobj_wref(&vp2->v_bufobj); 3059 } 3060 if (bp->b_bufobj != &vp2->v_bufobj) 3061 bp->b_bufobj = &vp2->v_bufobj; 3062 bp->b_vp = vp2; 3063 bp->b_iooffset = dbtob(bp->b_blkno); 3064 bstrategy(bp); 3065 return; 3066 } 3067 3068 static void 3069 swapdev_close(struct thread *td, struct swdevt *sp) 3070 { 3071 3072 VOP_CLOSE(sp->sw_vp, FREAD | FWRITE, td->td_ucred, td); 3073 vrele(sp->sw_vp); 3074 } 3075 3076 static int 3077 swaponvp(struct thread *td, struct vnode *vp, u_long nblks) 3078 { 3079 struct swdevt *sp; 3080 int error; 3081 3082 if (nblks == 0) 3083 return (ENXIO); 3084 mtx_lock(&sw_dev_mtx); 3085 TAILQ_FOREACH(sp, &swtailq, sw_list) { 3086 if (sp->sw_id == vp) { 3087 mtx_unlock(&sw_dev_mtx); 3088 return (EBUSY); 3089 } 3090 } 3091 mtx_unlock(&sw_dev_mtx); 3092 3093 (void) vn_lock(vp, LK_EXCLUSIVE | LK_RETRY); 3094 #ifdef MAC 3095 error = mac_system_check_swapon(td->td_ucred, vp); 3096 if (error == 0) 3097 #endif 3098 error = VOP_OPEN(vp, FREAD | FWRITE, td->td_ucred, td, NULL); 3099 (void) VOP_UNLOCK(vp); 3100 if (error) 3101 return (error); 3102 3103 swaponsomething(vp, vp, nblks, swapdev_strategy, swapdev_close, 3104 NODEV, 0); 3105 return (0); 3106 } 3107 3108 static int 3109 sysctl_swap_async_max(SYSCTL_HANDLER_ARGS) 3110 { 3111 int error, new, n; 3112 3113 new = nsw_wcount_async_max; 3114 error = sysctl_handle_int(oidp, &new, 0, req); 3115 if (error != 0 || req->newptr == NULL) 3116 return (error); 3117 3118 if (new > nswbuf / 2 || new < 1) 3119 return (EINVAL); 3120 3121 mtx_lock(&swbuf_mtx); 3122 while (nsw_wcount_async_max != new) { 3123 /* 3124 * Adjust difference. If the current async count is too low, 3125 * we will need to sqeeze our update slowly in. Sleep with a 3126 * higher priority than getpbuf() to finish faster. 3127 */ 3128 n = new - nsw_wcount_async_max; 3129 if (nsw_wcount_async + n >= 0) { 3130 nsw_wcount_async += n; 3131 nsw_wcount_async_max += n; 3132 wakeup(&nsw_wcount_async); 3133 } else { 3134 nsw_wcount_async_max -= nsw_wcount_async; 3135 nsw_wcount_async = 0; 3136 msleep(&nsw_wcount_async, &swbuf_mtx, PSWP, 3137 "swpsysctl", 0); 3138 } 3139 } 3140 mtx_unlock(&swbuf_mtx); 3141 3142 return (0); 3143 } 3144 3145 static void 3146 swap_pager_update_writecount(vm_object_t object, vm_offset_t start, 3147 vm_offset_t end) 3148 { 3149 3150 VM_OBJECT_WLOCK(object); 3151 KASSERT((object->flags & OBJ_ANON) == 0, 3152 ("Splittable object with writecount")); 3153 object->un_pager.swp.writemappings += (vm_ooffset_t)end - start; 3154 VM_OBJECT_WUNLOCK(object); 3155 } 3156 3157 static void 3158 swap_pager_release_writecount(vm_object_t object, vm_offset_t start, 3159 vm_offset_t end) 3160 { 3161 3162 VM_OBJECT_WLOCK(object); 3163 KASSERT((object->flags & OBJ_ANON) == 0, 3164 ("Splittable object with writecount")); 3165 object->un_pager.swp.writemappings -= (vm_ooffset_t)end - start; 3166 VM_OBJECT_WUNLOCK(object); 3167 } 3168 3169 static void 3170 swap_tmpfs_pager_getvp(vm_object_t object, struct vnode **vpp, bool *vp_heldp) 3171 { 3172 struct vnode *vp; 3173 3174 /* 3175 * Tmpfs VREG node, which was reclaimed, has OBJT_SWAP_TMPFS 3176 * type, but not OBJ_TMPFS flag. In this case there is no 3177 * v_writecount to adjust. 3178 */ 3179 if (vp_heldp != NULL) 3180 VM_OBJECT_RLOCK(object); 3181 else 3182 VM_OBJECT_ASSERT_LOCKED(object); 3183 if ((object->flags & OBJ_TMPFS) != 0) { 3184 vp = object->un_pager.swp.swp_tmpfs; 3185 if (vp != NULL) { 3186 *vpp = vp; 3187 if (vp_heldp != NULL) { 3188 vhold(vp); 3189 *vp_heldp = true; 3190 } 3191 } 3192 } 3193 if (vp_heldp != NULL) 3194 VM_OBJECT_RUNLOCK(object); 3195 } 3196