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