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