1 /*- 2 * SPDX-License-Identifier: BSD-4-Clause 3 * 4 * Copyright (c) 1998 Matthew Dillon, 5 * Copyright (c) 1994 John S. Dyson 6 * Copyright (c) 1990 University of Utah. 7 * Copyright (c) 1982, 1986, 1989, 1993 8 * The Regents of the University of California. All rights reserved. 9 * 10 * This code is derived from software contributed to Berkeley by 11 * the Systems Programming Group of the University of Utah Computer 12 * Science Department. 13 * 14 * Redistribution and use in source and binary forms, with or without 15 * modification, are permitted provided that the following conditions 16 * are met: 17 * 1. Redistributions of source code must retain the above copyright 18 * notice, this list of conditions and the following disclaimer. 19 * 2. Redistributions in binary form must reproduce the above copyright 20 * notice, this list of conditions and the following disclaimer in the 21 * documentation and/or other materials provided with the distribution. 22 * 3. All advertising materials mentioning features or use of this software 23 * must display the following acknowledgement: 24 * This product includes software developed by the University of 25 * California, Berkeley and its contributors. 26 * 4. Neither the name of the University nor the names of its contributors 27 * may be used to endorse or promote products derived from this software 28 * without specific prior written permission. 29 * 30 * THIS SOFTWARE IS PROVIDED BY THE REGENTS AND CONTRIBUTORS ``AS IS'' AND 31 * ANY EXPRESS OR IMPLIED WARRANTIES, INCLUDING, BUT NOT LIMITED TO, THE 32 * IMPLIED WARRANTIES OF MERCHANTABILITY AND FITNESS FOR A PARTICULAR PURPOSE 33 * ARE DISCLAIMED. IN NO EVENT SHALL THE REGENTS OR CONTRIBUTORS BE LIABLE 34 * FOR ANY DIRECT, INDIRECT, INCIDENTAL, SPECIAL, EXEMPLARY, OR CONSEQUENTIAL 35 * DAMAGES (INCLUDING, BUT NOT LIMITED TO, PROCUREMENT OF SUBSTITUTE GOODS 36 * OR SERVICES; LOSS OF USE, DATA, OR PROFITS; OR BUSINESS INTERRUPTION) 37 * HOWEVER CAUSED AND ON ANY THEORY OF LIABILITY, WHETHER IN CONTRACT, STRICT 38 * LIABILITY, OR TORT (INCLUDING NEGLIGENCE OR OTHERWISE) ARISING IN ANY WAY 39 * OUT OF THE USE OF THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF 40 * SUCH DAMAGE. 41 * 42 * New Swap System 43 * Matthew Dillon 44 * 45 * Radix Bitmap 'blists'. 46 * 47 * - The new swapper uses the new radix bitmap code. This should scale 48 * to arbitrarily small or arbitrarily large swap spaces and an almost 49 * arbitrary degree of fragmentation. 50 * 51 * Features: 52 * 53 * - on the fly reallocation of swap during putpages. The new system 54 * does not try to keep previously allocated swap blocks for dirty 55 * pages. 56 * 57 * - on the fly deallocation of swap 58 * 59 * - No more garbage collection required. Unnecessarily allocated swap 60 * blocks only exist for dirty vm_page_t's now and these are already 61 * cycled (in a high-load system) by the pager. We also do on-the-fly 62 * removal of invalidated swap blocks when a page is destroyed 63 * or renamed. 64 * 65 * from: Utah $Hdr: swap_pager.c 1.4 91/04/30$ 66 */ 67 68 #include <sys/cdefs.h> 69 #include "opt_vm.h" 70 71 #include <sys/param.h> 72 #include <sys/bio.h> 73 #include <sys/blist.h> 74 #include <sys/buf.h> 75 #include <sys/conf.h> 76 #include <sys/disk.h> 77 #include <sys/disklabel.h> 78 #include <sys/eventhandler.h> 79 #include <sys/fcntl.h> 80 #include <sys/limits.h> 81 #include <sys/lock.h> 82 #include <sys/kernel.h> 83 #include <sys/mount.h> 84 #include <sys/namei.h> 85 #include <sys/malloc.h> 86 #include <sys/pctrie.h> 87 #include <sys/priv.h> 88 #include <sys/proc.h> 89 #include <sys/racct.h> 90 #include <sys/resource.h> 91 #include <sys/resourcevar.h> 92 #include <sys/rwlock.h> 93 #include <sys/sbuf.h> 94 #include <sys/sysctl.h> 95 #include <sys/sysproto.h> 96 #include <sys/systm.h> 97 #include <sys/sx.h> 98 #include <sys/unistd.h> 99 #include <sys/user.h> 100 #include <sys/vmmeter.h> 101 #include <sys/vnode.h> 102 103 #include <security/mac/mac_framework.h> 104 105 #include <vm/vm.h> 106 #include <vm/pmap.h> 107 #include <vm/vm_map.h> 108 #include <vm/vm_kern.h> 109 #include <vm/vm_object.h> 110 #include <vm/vm_page.h> 111 #include <vm/vm_pager.h> 112 #include <vm/vm_pageout.h> 113 #include <vm/vm_param.h> 114 #include <vm/swap_pager.h> 115 #include <vm/vm_extern.h> 116 #include <vm/uma.h> 117 118 #include <geom/geom.h> 119 120 /* 121 * MAX_PAGEOUT_CLUSTER must be a power of 2 between 1 and 64. 122 * The 64-page limit is due to the radix code (kern/subr_blist.c). 123 */ 124 #ifndef MAX_PAGEOUT_CLUSTER 125 #define MAX_PAGEOUT_CLUSTER 32 126 #endif 127 128 #if !defined(SWB_NPAGES) 129 #define SWB_NPAGES MAX_PAGEOUT_CLUSTER 130 #endif 131 132 #define SWAP_META_PAGES PCTRIE_COUNT 133 134 /* 135 * A swblk structure maps each page index within a 136 * SWAP_META_PAGES-aligned and sized range to the address of an 137 * on-disk swap block (or SWAPBLK_NONE). The collection of these 138 * mappings for an entire vm object is implemented as a pc-trie. 139 */ 140 struct swblk { 141 vm_pindex_t p; 142 daddr_t d[SWAP_META_PAGES]; 143 }; 144 145 /* 146 * A page_range structure records the start address and length of a sequence of 147 * mapped page addresses. 148 */ 149 struct page_range { 150 daddr_t start; 151 daddr_t num; 152 }; 153 154 static MALLOC_DEFINE(M_VMPGDATA, "vm_pgdata", "swap pager private data"); 155 static struct mtx sw_dev_mtx; 156 static TAILQ_HEAD(, swdevt) swtailq = TAILQ_HEAD_INITIALIZER(swtailq); 157 static struct swdevt *swdevhd; /* Allocate from here next */ 158 static int nswapdev; /* Number of swap devices */ 159 int swap_pager_avail; 160 static struct sx swdev_syscall_lock; /* serialize swap(on|off) */ 161 162 static __exclusive_cache_line u_long swap_reserved; 163 static u_long swap_total; 164 static int sysctl_page_shift(SYSCTL_HANDLER_ARGS); 165 166 static SYSCTL_NODE(_vm_stats, OID_AUTO, swap, CTLFLAG_RD | CTLFLAG_MPSAFE, 0, 167 "VM swap stats"); 168 169 SYSCTL_PROC(_vm, OID_AUTO, swap_reserved, CTLTYPE_U64 | CTLFLAG_RD | CTLFLAG_MPSAFE, 170 &swap_reserved, 0, sysctl_page_shift, "QU", 171 "Amount of swap storage needed to back all allocated anonymous memory."); 172 SYSCTL_PROC(_vm, OID_AUTO, swap_total, CTLTYPE_U64 | CTLFLAG_RD | CTLFLAG_MPSAFE, 173 &swap_total, 0, sysctl_page_shift, "QU", 174 "Total amount of available swap storage."); 175 176 int vm_overcommit __read_mostly = 0; 177 SYSCTL_INT(_vm, VM_OVERCOMMIT, overcommit, CTLFLAG_RW, &vm_overcommit, 0, 178 "Configure virtual memory overcommit behavior. See tuning(7) " 179 "for details."); 180 static unsigned long swzone; 181 SYSCTL_ULONG(_vm, OID_AUTO, swzone, CTLFLAG_RD, &swzone, 0, 182 "Actual size of swap metadata zone"); 183 static unsigned long swap_maxpages; 184 SYSCTL_ULONG(_vm, OID_AUTO, swap_maxpages, CTLFLAG_RD, &swap_maxpages, 0, 185 "Maximum amount of swap supported"); 186 187 static COUNTER_U64_DEFINE_EARLY(swap_free_deferred); 188 SYSCTL_COUNTER_U64(_vm_stats_swap, OID_AUTO, free_deferred, 189 CTLFLAG_RD, &swap_free_deferred, 190 "Number of pages that deferred freeing swap space"); 191 192 static COUNTER_U64_DEFINE_EARLY(swap_free_completed); 193 SYSCTL_COUNTER_U64(_vm_stats_swap, OID_AUTO, free_completed, 194 CTLFLAG_RD, &swap_free_completed, 195 "Number of deferred frees completed"); 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, 221 ("negative vmsize for uid %d\n", uip->ui_uid)); 222 return (false); 223 } 224 return (true); 225 } 226 227 static void 228 swap_release_by_cred_rlimit(u_long pdecr, struct ucred *cred) 229 { 230 struct uidinfo *uip; 231 #ifdef INVARIANTS 232 u_long prev; 233 #endif 234 235 uip = cred->cr_ruidinfo; 236 237 #ifdef INVARIANTS 238 prev = atomic_fetchadd_long(&uip->ui_vmsize, -pdecr); 239 KASSERT(prev >= pdecr, 240 ("negative vmsize for uid %d\n", uip->ui_uid)); 241 #else 242 atomic_subtract_long(&uip->ui_vmsize, pdecr); 243 #endif 244 } 245 246 static void 247 swap_reserve_force_rlimit(u_long pincr, struct ucred *cred) 248 { 249 struct uidinfo *uip; 250 251 uip = cred->cr_ruidinfo; 252 atomic_add_long(&uip->ui_vmsize, pincr); 253 } 254 255 bool 256 swap_reserve(vm_ooffset_t incr) 257 { 258 259 return (swap_reserve_by_cred(incr, curthread->td_ucred)); 260 } 261 262 bool 263 swap_reserve_by_cred(vm_ooffset_t incr, struct ucred *cred) 264 { 265 u_long r, s, prev, pincr; 266 #ifdef RACCT 267 int error; 268 #endif 269 int oc; 270 static int curfail; 271 static struct timeval lastfail; 272 273 KASSERT((incr & PAGE_MASK) == 0, ("%s: incr: %ju & PAGE_MASK", 274 __func__, (uintmax_t)incr)); 275 276 #ifdef RACCT 277 if (RACCT_ENABLED()) { 278 PROC_LOCK(curproc); 279 error = racct_add(curproc, RACCT_SWAP, incr); 280 PROC_UNLOCK(curproc); 281 if (error != 0) 282 return (false); 283 } 284 #endif 285 286 pincr = atop(incr); 287 prev = atomic_fetchadd_long(&swap_reserved, pincr); 288 r = prev + pincr; 289 s = swap_total; 290 oc = atomic_load_int(&vm_overcommit); 291 if (r > s && (oc & SWAP_RESERVE_ALLOW_NONWIRED) != 0) { 292 s += vm_cnt.v_page_count - vm_cnt.v_free_reserved - 293 vm_wire_count(); 294 } 295 if ((oc & SWAP_RESERVE_FORCE_ON) != 0 && r > s && 296 priv_check(curthread, PRIV_VM_SWAP_NOQUOTA) != 0) { 297 prev = atomic_fetchadd_long(&swap_reserved, -pincr); 298 KASSERT(prev >= pincr, 299 ("swap_reserved < incr on overcommit fail")); 300 goto out_error; 301 } 302 303 if (!swap_reserve_by_cred_rlimit(pincr, cred, oc)) { 304 prev = atomic_fetchadd_long(&swap_reserved, -pincr); 305 KASSERT(prev >= pincr, 306 ("swap_reserved < incr on overcommit fail")); 307 goto out_error; 308 } 309 310 return (true); 311 312 out_error: 313 if (ppsratecheck(&lastfail, &curfail, 1)) { 314 printf("uid %d, pid %d: swap reservation " 315 "for %jd bytes failed\n", 316 cred->cr_ruidinfo->ui_uid, curproc->p_pid, incr); 317 } 318 #ifdef RACCT 319 if (RACCT_ENABLED()) { 320 PROC_LOCK(curproc); 321 racct_sub(curproc, RACCT_SWAP, incr); 322 PROC_UNLOCK(curproc); 323 } 324 #endif 325 326 return (false); 327 } 328 329 void 330 swap_reserve_force(vm_ooffset_t incr) 331 { 332 u_long pincr; 333 334 KASSERT((incr & PAGE_MASK) == 0, ("%s: incr: %ju & PAGE_MASK", 335 __func__, (uintmax_t)incr)); 336 337 #ifdef RACCT 338 if (RACCT_ENABLED()) { 339 PROC_LOCK(curproc); 340 racct_add_force(curproc, RACCT_SWAP, incr); 341 PROC_UNLOCK(curproc); 342 } 343 #endif 344 pincr = atop(incr); 345 atomic_add_long(&swap_reserved, pincr); 346 swap_reserve_force_rlimit(pincr, curthread->td_ucred); 347 } 348 349 void 350 swap_release(vm_ooffset_t decr) 351 { 352 struct ucred *cred; 353 354 PROC_LOCK(curproc); 355 cred = curproc->p_ucred; 356 swap_release_by_cred(decr, cred); 357 PROC_UNLOCK(curproc); 358 } 359 360 void 361 swap_release_by_cred(vm_ooffset_t decr, struct ucred *cred) 362 { 363 u_long pdecr; 364 #ifdef INVARIANTS 365 u_long prev; 366 #endif 367 368 KASSERT((decr & PAGE_MASK) == 0, ("%s: decr: %ju & PAGE_MASK", 369 __func__, (uintmax_t)decr)); 370 371 pdecr = atop(decr); 372 #ifdef INVARIANTS 373 prev = atomic_fetchadd_long(&swap_reserved, -pdecr); 374 KASSERT(prev >= pdecr, ("swap_reserved < decr")); 375 #else 376 atomic_subtract_long(&swap_reserved, pdecr); 377 #endif 378 379 swap_release_by_cred_rlimit(pdecr, cred); 380 #ifdef RACCT 381 if (racct_enable) 382 racct_sub_cred(cred, RACCT_SWAP, decr); 383 #endif 384 } 385 386 static int swap_pager_full = 2; /* swap space exhaustion (task killing) */ 387 static int swap_pager_almost_full = 1; /* swap space exhaustion (w/hysteresis)*/ 388 static struct mtx swbuf_mtx; /* to sync nsw_wcount_async */ 389 static int nsw_wcount_async; /* limit async write buffers */ 390 static int nsw_wcount_async_max;/* assigned maximum */ 391 int nsw_cluster_max; /* maximum VOP I/O allowed */ 392 393 static int sysctl_swap_async_max(SYSCTL_HANDLER_ARGS); 394 SYSCTL_PROC(_vm, OID_AUTO, swap_async_max, CTLTYPE_INT | CTLFLAG_RW | 395 CTLFLAG_MPSAFE, NULL, 0, sysctl_swap_async_max, "I", 396 "Maximum running async swap ops"); 397 static int sysctl_swap_fragmentation(SYSCTL_HANDLER_ARGS); 398 SYSCTL_PROC(_vm, OID_AUTO, swap_fragmentation, CTLTYPE_STRING | CTLFLAG_RD | 399 CTLFLAG_MPSAFE, NULL, 0, sysctl_swap_fragmentation, "A", 400 "Swap Fragmentation Info"); 401 402 static struct sx sw_alloc_sx; 403 404 /* 405 * "named" and "unnamed" anon region objects. Try to reduce the overhead 406 * of searching a named list by hashing it just a little. 407 */ 408 409 #define NOBJLISTS 8 410 411 #define NOBJLIST(handle) \ 412 (&swap_pager_object_list[((int)(intptr_t)handle >> 4) & (NOBJLISTS-1)]) 413 414 static struct pagerlst swap_pager_object_list[NOBJLISTS]; 415 static uma_zone_t swwbuf_zone; 416 static uma_zone_t swrbuf_zone; 417 static uma_zone_t swblk_zone; 418 static uma_zone_t swpctrie_zone; 419 420 /* 421 * pagerops for OBJT_SWAP - "swap pager". Some ops are also global procedure 422 * calls hooked from other parts of the VM system and do not appear here. 423 * (see vm/swap_pager.h). 424 */ 425 static vm_object_t 426 swap_pager_alloc(void *handle, vm_ooffset_t size, 427 vm_prot_t prot, vm_ooffset_t offset, struct ucred *); 428 static void swap_pager_dealloc(vm_object_t object); 429 static int swap_pager_getpages(vm_object_t, vm_page_t *, int, int *, 430 int *); 431 static int swap_pager_getpages_async(vm_object_t, vm_page_t *, int, int *, 432 int *, pgo_getpages_iodone_t, void *); 433 static void swap_pager_putpages(vm_object_t, vm_page_t *, int, int, int *); 434 static boolean_t 435 swap_pager_haspage(vm_object_t object, vm_pindex_t pindex, int *before, int *after); 436 static void swap_pager_init(void); 437 static void swap_pager_unswapped(vm_page_t); 438 static void swap_pager_swapoff(struct swdevt *sp); 439 static void swap_pager_update_writecount(vm_object_t object, 440 vm_offset_t start, vm_offset_t end); 441 static void swap_pager_release_writecount(vm_object_t object, 442 vm_offset_t start, vm_offset_t end); 443 static void swap_pager_freespace_pgo(vm_object_t object, vm_pindex_t start, 444 vm_size_t size); 445 446 const struct pagerops swappagerops = { 447 .pgo_kvme_type = KVME_TYPE_SWAP, 448 .pgo_init = swap_pager_init, /* early system initialization of pager */ 449 .pgo_alloc = swap_pager_alloc, /* allocate an OBJT_SWAP object */ 450 .pgo_dealloc = swap_pager_dealloc, /* deallocate an OBJT_SWAP object */ 451 .pgo_getpages = swap_pager_getpages, /* pagein */ 452 .pgo_getpages_async = swap_pager_getpages_async, /* pagein (async) */ 453 .pgo_putpages = swap_pager_putpages, /* pageout */ 454 .pgo_haspage = swap_pager_haspage, /* get backing store status for page */ 455 .pgo_pageunswapped = swap_pager_unswapped, /* remove swap related to page */ 456 .pgo_update_writecount = swap_pager_update_writecount, 457 .pgo_release_writecount = swap_pager_release_writecount, 458 .pgo_freespace = swap_pager_freespace_pgo, 459 }; 460 461 /* 462 * swap_*() routines are externally accessible. swp_*() routines are 463 * internal. 464 */ 465 static int nswap_lowat = 128; /* in pages, swap_pager_almost_full warn */ 466 static int nswap_hiwat = 512; /* in pages, swap_pager_almost_full warn */ 467 468 SYSCTL_INT(_vm, OID_AUTO, dmmax, CTLFLAG_RD, &nsw_cluster_max, 0, 469 "Maximum size of a swap block in pages"); 470 471 static void swp_sizecheck(void); 472 static void swp_pager_async_iodone(struct buf *bp); 473 static bool swp_pager_swblk_empty(struct swblk *sb, int start, int limit); 474 static void swp_pager_free_empty_swblk(vm_object_t, struct swblk *sb); 475 static int swapongeom(struct vnode *); 476 static int swaponvp(struct thread *, struct vnode *, u_long); 477 static int swapoff_one(struct swdevt *sp, struct ucred *cred, 478 u_int flags); 479 480 /* 481 * Swap bitmap functions 482 */ 483 static void swp_pager_freeswapspace(const struct page_range *range); 484 static daddr_t swp_pager_getswapspace(int *npages); 485 486 /* 487 * Metadata functions 488 */ 489 static daddr_t swp_pager_meta_build(vm_object_t, vm_pindex_t, daddr_t); 490 static void swp_pager_meta_free(vm_object_t, vm_pindex_t, vm_pindex_t, 491 vm_size_t *); 492 static void swp_pager_meta_transfer(vm_object_t src, vm_object_t dst, 493 vm_pindex_t pindex, vm_pindex_t count, vm_size_t *freed); 494 static void swp_pager_meta_free_all(vm_object_t); 495 static daddr_t swp_pager_meta_lookup(vm_object_t, vm_pindex_t); 496 497 static void 498 swp_pager_init_freerange(struct page_range *range) 499 { 500 range->start = SWAPBLK_NONE; 501 range->num = 0; 502 } 503 504 static void 505 swp_pager_update_freerange(struct page_range *range, daddr_t addr) 506 { 507 if (range->start + range->num == addr) { 508 range->num++; 509 } else { 510 swp_pager_freeswapspace(range); 511 range->start = addr; 512 range->num = 1; 513 } 514 } 515 516 static void * 517 swblk_trie_alloc(struct pctrie *ptree) 518 { 519 520 return (uma_zalloc(swpctrie_zone, M_NOWAIT | (curproc == pageproc ? 521 M_USE_RESERVE : 0))); 522 } 523 524 static void 525 swblk_trie_free(struct pctrie *ptree, void *node) 526 { 527 528 uma_zfree(swpctrie_zone, node); 529 } 530 531 PCTRIE_DEFINE(SWAP, swblk, p, swblk_trie_alloc, swblk_trie_free); 532 533 /* 534 * SWP_SIZECHECK() - update swap_pager_full indication 535 * 536 * update the swap_pager_almost_full indication and warn when we are 537 * about to run out of swap space, using lowat/hiwat hysteresis. 538 * 539 * Clear swap_pager_full ( task killing ) indication when lowat is met. 540 * 541 * No restrictions on call 542 * This routine may not block. 543 */ 544 static void 545 swp_sizecheck(void) 546 { 547 548 if (swap_pager_avail < nswap_lowat) { 549 if (swap_pager_almost_full == 0) { 550 printf("swap_pager: out of swap space\n"); 551 swap_pager_almost_full = 1; 552 } 553 } else { 554 swap_pager_full = 0; 555 if (swap_pager_avail > nswap_hiwat) 556 swap_pager_almost_full = 0; 557 } 558 } 559 560 /* 561 * SWAP_PAGER_INIT() - initialize the swap pager! 562 * 563 * Expected to be started from system init. NOTE: This code is run 564 * before much else so be careful what you depend on. Most of the VM 565 * system has yet to be initialized at this point. 566 */ 567 static void 568 swap_pager_init(void) 569 { 570 /* 571 * Initialize object lists 572 */ 573 int i; 574 575 for (i = 0; i < NOBJLISTS; ++i) 576 TAILQ_INIT(&swap_pager_object_list[i]); 577 mtx_init(&sw_dev_mtx, "swapdev", NULL, MTX_DEF); 578 sx_init(&sw_alloc_sx, "swspsx"); 579 sx_init(&swdev_syscall_lock, "swsysc"); 580 581 /* 582 * The nsw_cluster_max is constrained by the bp->b_pages[] 583 * array, which has maxphys / PAGE_SIZE entries, and our locally 584 * defined MAX_PAGEOUT_CLUSTER. Also be aware that swap ops are 585 * constrained by the swap device interleave stripe size. 586 * 587 * Initialized early so that GEOM_ELI can see it. 588 */ 589 nsw_cluster_max = min(maxphys / PAGE_SIZE, MAX_PAGEOUT_CLUSTER); 590 } 591 592 /* 593 * SWAP_PAGER_SWAP_INIT() - swap pager initialization from pageout process 594 * 595 * Expected to be started from pageout process once, prior to entering 596 * its main loop. 597 */ 598 void 599 swap_pager_swap_init(void) 600 { 601 unsigned long n, n2; 602 603 /* 604 * Number of in-transit swap bp operations. Don't 605 * exhaust the pbufs completely. Make sure we 606 * initialize workable values (0 will work for hysteresis 607 * but it isn't very efficient). 608 * 609 * Currently we hardwire nsw_wcount_async to 4. This limit is 610 * designed to prevent other I/O from having high latencies due to 611 * our pageout I/O. The value 4 works well for one or two active swap 612 * devices but is probably a little low if you have more. Even so, 613 * a higher value would probably generate only a limited improvement 614 * with three or four active swap devices since the system does not 615 * typically have to pageout at extreme bandwidths. We will want 616 * at least 2 per swap devices, and 4 is a pretty good value if you 617 * have one NFS swap device due to the command/ack latency over NFS. 618 * So it all works out pretty well. 619 * 620 * nsw_cluster_max is initialized in swap_pager_init(). 621 */ 622 623 nsw_wcount_async = 4; 624 nsw_wcount_async_max = nsw_wcount_async; 625 mtx_init(&swbuf_mtx, "async swbuf mutex", NULL, MTX_DEF); 626 627 swwbuf_zone = pbuf_zsecond_create("swwbuf", nswbuf / 4); 628 swrbuf_zone = pbuf_zsecond_create("swrbuf", nswbuf / 2); 629 630 /* 631 * Initialize our zone, taking the user's requested size or 632 * estimating the number we need based on the number of pages 633 * in the system. 634 */ 635 n = maxswzone != 0 ? maxswzone / sizeof(struct swblk) : 636 vm_cnt.v_page_count / 2; 637 swpctrie_zone = uma_zcreate("swpctrie", pctrie_node_size(), NULL, NULL, 638 pctrie_zone_init, NULL, UMA_ALIGN_PTR, 0); 639 swblk_zone = uma_zcreate("swblk", sizeof(struct swblk), NULL, NULL, 640 NULL, NULL, _Alignof(struct swblk) - 1, 0); 641 n2 = n; 642 do { 643 if (uma_zone_reserve_kva(swblk_zone, n)) 644 break; 645 /* 646 * if the allocation failed, try a zone two thirds the 647 * size of the previous attempt. 648 */ 649 n -= ((n + 2) / 3); 650 } while (n > 0); 651 652 /* 653 * Often uma_zone_reserve_kva() cannot reserve exactly the 654 * requested size. Account for the difference when 655 * calculating swap_maxpages. 656 */ 657 n = uma_zone_get_max(swblk_zone); 658 659 if (n < n2) 660 printf("Swap blk zone entries changed from %lu to %lu.\n", 661 n2, n); 662 /* absolute maximum we can handle assuming 100% efficiency */ 663 swap_maxpages = n * SWAP_META_PAGES; 664 swzone = n * sizeof(struct swblk); 665 if (!uma_zone_reserve_kva(swpctrie_zone, n)) 666 printf("Cannot reserve swap pctrie zone, " 667 "reduce kern.maxswzone.\n"); 668 } 669 670 bool 671 swap_pager_init_object(vm_object_t object, void *handle, struct ucred *cred, 672 vm_ooffset_t size, vm_ooffset_t offset) 673 { 674 if (cred != NULL) { 675 if (!swap_reserve_by_cred(size, cred)) 676 return (false); 677 crhold(cred); 678 } 679 680 object->un_pager.swp.writemappings = 0; 681 object->handle = handle; 682 if (cred != NULL) { 683 object->cred = cred; 684 object->charge = size; 685 } 686 return (true); 687 } 688 689 static vm_object_t 690 swap_pager_alloc_init(objtype_t otype, void *handle, struct ucred *cred, 691 vm_ooffset_t size, vm_ooffset_t offset) 692 { 693 vm_object_t object; 694 695 /* 696 * The un_pager.swp.swp_blks trie is initialized by 697 * vm_object_allocate() to ensure the correct order of 698 * visibility to other threads. 699 */ 700 object = vm_object_allocate(otype, OFF_TO_IDX(offset + 701 PAGE_MASK + size)); 702 703 if (!swap_pager_init_object(object, handle, cred, size, offset)) { 704 vm_object_deallocate(object); 705 return (NULL); 706 } 707 return (object); 708 } 709 710 /* 711 * SWAP_PAGER_ALLOC() - allocate a new OBJT_SWAP VM object and instantiate 712 * its metadata structures. 713 * 714 * This routine is called from the mmap and fork code to create a new 715 * OBJT_SWAP object. 716 * 717 * This routine must ensure that no live duplicate is created for 718 * the named object request, which is protected against by 719 * holding the sw_alloc_sx lock in case handle != NULL. 720 */ 721 static vm_object_t 722 swap_pager_alloc(void *handle, vm_ooffset_t size, vm_prot_t prot, 723 vm_ooffset_t offset, struct ucred *cred) 724 { 725 vm_object_t object; 726 727 if (handle != NULL) { 728 /* 729 * Reference existing named region or allocate new one. There 730 * should not be a race here against swp_pager_meta_build() 731 * as called from vm_page_remove() in regards to the lookup 732 * of the handle. 733 */ 734 sx_xlock(&sw_alloc_sx); 735 object = vm_pager_object_lookup(NOBJLIST(handle), handle); 736 if (object == NULL) { 737 object = swap_pager_alloc_init(OBJT_SWAP, handle, cred, 738 size, offset); 739 if (object != NULL) { 740 TAILQ_INSERT_TAIL(NOBJLIST(object->handle), 741 object, pager_object_list); 742 } 743 } 744 sx_xunlock(&sw_alloc_sx); 745 } else { 746 object = swap_pager_alloc_init(OBJT_SWAP, handle, cred, 747 size, offset); 748 } 749 return (object); 750 } 751 752 /* 753 * SWAP_PAGER_DEALLOC() - remove swap metadata from object 754 * 755 * The swap backing for the object is destroyed. The code is 756 * designed such that we can reinstantiate it later, but this 757 * routine is typically called only when the entire object is 758 * about to be destroyed. 759 * 760 * The object must be locked. 761 */ 762 static void 763 swap_pager_dealloc(vm_object_t object) 764 { 765 766 VM_OBJECT_ASSERT_WLOCKED(object); 767 KASSERT((object->flags & OBJ_DEAD) != 0, ("dealloc of reachable obj")); 768 769 /* 770 * Remove from list right away so lookups will fail if we block for 771 * pageout completion. 772 */ 773 if ((object->flags & OBJ_ANON) == 0 && object->handle != NULL) { 774 VM_OBJECT_WUNLOCK(object); 775 sx_xlock(&sw_alloc_sx); 776 TAILQ_REMOVE(NOBJLIST(object->handle), object, 777 pager_object_list); 778 sx_xunlock(&sw_alloc_sx); 779 VM_OBJECT_WLOCK(object); 780 } 781 782 vm_object_pip_wait(object, "swpdea"); 783 784 /* 785 * Free all remaining metadata. We only bother to free it from 786 * the swap meta data. We do not attempt to free swapblk's still 787 * associated with vm_page_t's for this object. We do not care 788 * if paging is still in progress on some objects. 789 */ 790 swp_pager_meta_free_all(object); 791 object->handle = NULL; 792 object->type = OBJT_DEAD; 793 794 /* 795 * Release the allocation charge. 796 */ 797 if (object->cred != NULL) { 798 swap_release_by_cred(object->charge, object->cred); 799 object->charge = 0; 800 crfree(object->cred); 801 object->cred = NULL; 802 } 803 804 /* 805 * Hide the object from swap_pager_swapoff(). 806 */ 807 vm_object_clear_flag(object, OBJ_SWAP); 808 } 809 810 /************************************************************************ 811 * SWAP PAGER BITMAP ROUTINES * 812 ************************************************************************/ 813 814 /* 815 * SWP_PAGER_GETSWAPSPACE() - allocate raw swap space 816 * 817 * Allocate swap for up to the requested number of pages. The 818 * starting swap block number (a page index) is returned or 819 * SWAPBLK_NONE if the allocation failed. 820 * 821 * Also has the side effect of advising that somebody made a mistake 822 * when they configured swap and didn't configure enough. 823 * 824 * This routine may not sleep. 825 * 826 * We allocate in round-robin fashion from the configured devices. 827 */ 828 static daddr_t 829 swp_pager_getswapspace(int *io_npages) 830 { 831 daddr_t blk; 832 struct swdevt *sp; 833 int mpages, npages; 834 835 KASSERT(*io_npages >= 1, 836 ("%s: npages not positive", __func__)); 837 blk = SWAPBLK_NONE; 838 mpages = *io_npages; 839 npages = imin(BLIST_MAX_ALLOC, mpages); 840 mtx_lock(&sw_dev_mtx); 841 sp = swdevhd; 842 while (!TAILQ_EMPTY(&swtailq)) { 843 if (sp == NULL) 844 sp = TAILQ_FIRST(&swtailq); 845 if ((sp->sw_flags & SW_CLOSING) == 0) 846 blk = blist_alloc(sp->sw_blist, &npages, mpages); 847 if (blk != SWAPBLK_NONE) 848 break; 849 sp = TAILQ_NEXT(sp, sw_list); 850 if (swdevhd == sp) { 851 if (npages == 1) 852 break; 853 mpages = npages - 1; 854 npages >>= 1; 855 } 856 } 857 if (blk != SWAPBLK_NONE) { 858 *io_npages = npages; 859 blk += sp->sw_first; 860 sp->sw_used += npages; 861 swap_pager_avail -= npages; 862 swp_sizecheck(); 863 swdevhd = TAILQ_NEXT(sp, sw_list); 864 } else { 865 if (swap_pager_full != 2) { 866 printf("swp_pager_getswapspace(%d): failed\n", 867 *io_npages); 868 swap_pager_full = 2; 869 swap_pager_almost_full = 1; 870 } 871 swdevhd = NULL; 872 } 873 mtx_unlock(&sw_dev_mtx); 874 return (blk); 875 } 876 877 static bool 878 swp_pager_isondev(daddr_t blk, struct swdevt *sp) 879 { 880 881 return (blk >= sp->sw_first && blk < sp->sw_end); 882 } 883 884 static void 885 swp_pager_strategy(struct buf *bp) 886 { 887 struct swdevt *sp; 888 889 mtx_lock(&sw_dev_mtx); 890 TAILQ_FOREACH(sp, &swtailq, sw_list) { 891 if (swp_pager_isondev(bp->b_blkno, sp)) { 892 mtx_unlock(&sw_dev_mtx); 893 if ((sp->sw_flags & SW_UNMAPPED) != 0 && 894 unmapped_buf_allowed) { 895 bp->b_data = unmapped_buf; 896 bp->b_offset = 0; 897 } else { 898 pmap_qenter((vm_offset_t)bp->b_data, 899 &bp->b_pages[0], bp->b_bcount / PAGE_SIZE); 900 } 901 sp->sw_strategy(bp, sp); 902 return; 903 } 904 } 905 panic("Swapdev not found"); 906 } 907 908 /* 909 * SWP_PAGER_FREESWAPSPACE() - free raw swap space 910 * 911 * This routine returns the specified swap blocks back to the bitmap. 912 * 913 * This routine may not sleep. 914 */ 915 static void 916 swp_pager_freeswapspace(const struct page_range *range) 917 { 918 daddr_t blk, npages; 919 struct swdevt *sp; 920 921 blk = range->start; 922 npages = range->num; 923 if (npages == 0) 924 return; 925 mtx_lock(&sw_dev_mtx); 926 TAILQ_FOREACH(sp, &swtailq, sw_list) { 927 if (swp_pager_isondev(blk, sp)) { 928 sp->sw_used -= npages; 929 /* 930 * If we are attempting to stop swapping on 931 * this device, we don't want to mark any 932 * blocks free lest they be reused. 933 */ 934 if ((sp->sw_flags & SW_CLOSING) == 0) { 935 blist_free(sp->sw_blist, blk - sp->sw_first, 936 npages); 937 swap_pager_avail += npages; 938 swp_sizecheck(); 939 } 940 mtx_unlock(&sw_dev_mtx); 941 return; 942 } 943 } 944 panic("Swapdev not found"); 945 } 946 947 /* 948 * SYSCTL_SWAP_FRAGMENTATION() - produce raw swap space stats 949 */ 950 static int 951 sysctl_swap_fragmentation(SYSCTL_HANDLER_ARGS) 952 { 953 struct sbuf sbuf; 954 struct swdevt *sp; 955 const char *devname; 956 int error; 957 958 error = sysctl_wire_old_buffer(req, 0); 959 if (error != 0) 960 return (error); 961 sbuf_new_for_sysctl(&sbuf, NULL, 128, req); 962 mtx_lock(&sw_dev_mtx); 963 TAILQ_FOREACH(sp, &swtailq, sw_list) { 964 if (vn_isdisk(sp->sw_vp)) 965 devname = devtoname(sp->sw_vp->v_rdev); 966 else 967 devname = "[file]"; 968 sbuf_printf(&sbuf, "\nFree space on device %s:\n", devname); 969 blist_stats(sp->sw_blist, &sbuf); 970 } 971 mtx_unlock(&sw_dev_mtx); 972 error = sbuf_finish(&sbuf); 973 sbuf_delete(&sbuf); 974 return (error); 975 } 976 977 /* 978 * SWAP_PAGER_FREESPACE() - frees swap blocks associated with a page 979 * range within an object. 980 * 981 * This routine removes swapblk assignments from swap metadata. 982 * 983 * The external callers of this routine typically have already destroyed 984 * or renamed vm_page_t's associated with this range in the object so 985 * we should be ok. 986 * 987 * The object must be locked. 988 */ 989 void 990 swap_pager_freespace(vm_object_t object, vm_pindex_t start, vm_size_t size, 991 vm_size_t *freed) 992 { 993 MPASS((object->flags & OBJ_SWAP) != 0); 994 995 swp_pager_meta_free(object, start, size, freed); 996 } 997 998 static void 999 swap_pager_freespace_pgo(vm_object_t object, vm_pindex_t start, vm_size_t size) 1000 { 1001 MPASS((object->flags & OBJ_SWAP) != 0); 1002 1003 swp_pager_meta_free(object, start, size, NULL); 1004 } 1005 1006 /* 1007 * SWAP_PAGER_RESERVE() - reserve swap blocks in object 1008 * 1009 * Assigns swap blocks to the specified range within the object. The 1010 * swap blocks are not zeroed. Any previous swap assignment is destroyed. 1011 * 1012 * Returns 0 on success, -1 on failure. 1013 */ 1014 int 1015 swap_pager_reserve(vm_object_t object, vm_pindex_t start, vm_pindex_t size) 1016 { 1017 struct page_range range; 1018 daddr_t addr, blk; 1019 vm_pindex_t i, j; 1020 int n; 1021 1022 swp_pager_init_freerange(&range); 1023 VM_OBJECT_WLOCK(object); 1024 for (i = 0; i < size; i += n) { 1025 n = MIN(size - i, INT_MAX); 1026 blk = swp_pager_getswapspace(&n); 1027 if (blk == SWAPBLK_NONE) { 1028 swp_pager_meta_free(object, start, i, NULL); 1029 VM_OBJECT_WUNLOCK(object); 1030 return (-1); 1031 } 1032 for (j = 0; j < n; ++j) { 1033 addr = swp_pager_meta_build(object, 1034 start + i + j, blk + j); 1035 if (addr != SWAPBLK_NONE) 1036 swp_pager_update_freerange(&range, addr); 1037 } 1038 } 1039 swp_pager_freeswapspace(&range); 1040 VM_OBJECT_WUNLOCK(object); 1041 return (0); 1042 } 1043 1044 static bool 1045 swp_pager_xfer_source(vm_object_t srcobject, vm_object_t dstobject, 1046 vm_pindex_t pindex, daddr_t addr) 1047 { 1048 daddr_t dstaddr __diagused; 1049 1050 KASSERT((srcobject->flags & OBJ_SWAP) != 0, 1051 ("%s: srcobject not swappable", __func__)); 1052 KASSERT((dstobject->flags & OBJ_SWAP) != 0, 1053 ("%s: dstobject not swappable", __func__)); 1054 1055 if (swp_pager_meta_lookup(dstobject, pindex) != SWAPBLK_NONE) { 1056 /* Caller should destroy the source block. */ 1057 return (false); 1058 } 1059 1060 /* 1061 * Destination has no swapblk and is not resident, transfer source. 1062 * swp_pager_meta_build() can sleep. 1063 */ 1064 VM_OBJECT_WUNLOCK(srcobject); 1065 dstaddr = swp_pager_meta_build(dstobject, pindex, addr); 1066 KASSERT(dstaddr == SWAPBLK_NONE, 1067 ("Unexpected destination swapblk")); 1068 VM_OBJECT_WLOCK(srcobject); 1069 1070 return (true); 1071 } 1072 1073 /* 1074 * SWAP_PAGER_COPY() - copy blocks from source pager to destination pager 1075 * and destroy the source. 1076 * 1077 * Copy any valid swapblks from the source to the destination. In 1078 * cases where both the source and destination have a valid swapblk, 1079 * we keep the destination's. 1080 * 1081 * This routine is allowed to sleep. It may sleep allocating metadata 1082 * indirectly through swp_pager_meta_build(). 1083 * 1084 * The source object contains no vm_page_t's (which is just as well) 1085 * 1086 * The source and destination objects must be locked. 1087 * Both object locks may temporarily be released. 1088 */ 1089 void 1090 swap_pager_copy(vm_object_t srcobject, vm_object_t dstobject, 1091 vm_pindex_t offset, int destroysource) 1092 { 1093 VM_OBJECT_ASSERT_WLOCKED(srcobject); 1094 VM_OBJECT_ASSERT_WLOCKED(dstobject); 1095 1096 /* 1097 * If destroysource is set, we remove the source object from the 1098 * swap_pager internal queue now. 1099 */ 1100 if (destroysource && (srcobject->flags & OBJ_ANON) == 0 && 1101 srcobject->handle != NULL) { 1102 VM_OBJECT_WUNLOCK(srcobject); 1103 VM_OBJECT_WUNLOCK(dstobject); 1104 sx_xlock(&sw_alloc_sx); 1105 TAILQ_REMOVE(NOBJLIST(srcobject->handle), srcobject, 1106 pager_object_list); 1107 sx_xunlock(&sw_alloc_sx); 1108 VM_OBJECT_WLOCK(dstobject); 1109 VM_OBJECT_WLOCK(srcobject); 1110 } 1111 1112 /* 1113 * Transfer source to destination. 1114 */ 1115 swp_pager_meta_transfer(srcobject, dstobject, offset, dstobject->size, 1116 NULL); 1117 1118 /* 1119 * Free left over swap blocks in source. 1120 */ 1121 if (destroysource) 1122 swp_pager_meta_free_all(srcobject); 1123 } 1124 1125 /* 1126 * SWAP_PAGER_HASPAGE() - determine if we have good backing store for 1127 * the requested page. 1128 * 1129 * We determine whether good backing store exists for the requested 1130 * page and return TRUE if it does, FALSE if it doesn't. 1131 * 1132 * If TRUE, we also try to determine how much valid, contiguous backing 1133 * store exists before and after the requested page. 1134 */ 1135 static boolean_t 1136 swap_pager_haspage(vm_object_t object, vm_pindex_t pindex, int *before, 1137 int *after) 1138 { 1139 daddr_t blk, blk0; 1140 int i; 1141 1142 VM_OBJECT_ASSERT_LOCKED(object); 1143 KASSERT((object->flags & OBJ_SWAP) != 0, 1144 ("%s: object not swappable", __func__)); 1145 1146 /* 1147 * do we have good backing store at the requested index ? 1148 */ 1149 blk0 = swp_pager_meta_lookup(object, pindex); 1150 if (blk0 == SWAPBLK_NONE) { 1151 if (before) 1152 *before = 0; 1153 if (after) 1154 *after = 0; 1155 return (FALSE); 1156 } 1157 1158 /* 1159 * find backwards-looking contiguous good backing store 1160 */ 1161 if (before != NULL) { 1162 for (i = 1; i < SWB_NPAGES; i++) { 1163 if (i > pindex) 1164 break; 1165 blk = swp_pager_meta_lookup(object, pindex - i); 1166 if (blk != blk0 - i) 1167 break; 1168 } 1169 *before = i - 1; 1170 } 1171 1172 /* 1173 * find forward-looking contiguous good backing store 1174 */ 1175 if (after != NULL) { 1176 for (i = 1; i < SWB_NPAGES; i++) { 1177 blk = swp_pager_meta_lookup(object, pindex + i); 1178 if (blk != blk0 + i) 1179 break; 1180 } 1181 *after = i - 1; 1182 } 1183 return (TRUE); 1184 } 1185 1186 static void 1187 swap_pager_unswapped_acct(vm_page_t m) 1188 { 1189 KASSERT((m->object->flags & OBJ_SWAP) != 0, 1190 ("Free object not swappable")); 1191 if ((m->a.flags & PGA_SWAP_FREE) != 0) 1192 counter_u64_add(swap_free_completed, 1); 1193 vm_page_aflag_clear(m, PGA_SWAP_FREE | PGA_SWAP_SPACE); 1194 1195 /* 1196 * The meta data only exists if the object is OBJT_SWAP 1197 * and even then might not be allocated yet. 1198 */ 1199 } 1200 1201 /* 1202 * SWAP_PAGER_PAGE_UNSWAPPED() - remove swap backing store related to page 1203 * 1204 * This removes any associated swap backing store, whether valid or 1205 * not, from the page. 1206 * 1207 * This routine is typically called when a page is made dirty, at 1208 * which point any associated swap can be freed. MADV_FREE also 1209 * calls us in a special-case situation 1210 * 1211 * NOTE!!! If the page is clean and the swap was valid, the caller 1212 * should make the page dirty before calling this routine. This routine 1213 * does NOT change the m->dirty status of the page. Also: MADV_FREE 1214 * depends on it. 1215 * 1216 * This routine may not sleep. 1217 * 1218 * The object containing the page may be locked. 1219 */ 1220 static void 1221 swap_pager_unswapped(vm_page_t m) 1222 { 1223 struct page_range range; 1224 struct swblk *sb; 1225 vm_object_t obj; 1226 1227 /* 1228 * Handle enqueing deferred frees first. If we do not have the 1229 * object lock we wait for the page daemon to clear the space. 1230 */ 1231 obj = m->object; 1232 if (!VM_OBJECT_WOWNED(obj)) { 1233 VM_PAGE_OBJECT_BUSY_ASSERT(m); 1234 /* 1235 * The caller is responsible for synchronization but we 1236 * will harmlessly handle races. This is typically provided 1237 * by only calling unswapped() when a page transitions from 1238 * clean to dirty. 1239 */ 1240 if ((m->a.flags & (PGA_SWAP_SPACE | PGA_SWAP_FREE)) == 1241 PGA_SWAP_SPACE) { 1242 vm_page_aflag_set(m, PGA_SWAP_FREE); 1243 counter_u64_add(swap_free_deferred, 1); 1244 } 1245 return; 1246 } 1247 swap_pager_unswapped_acct(m); 1248 1249 sb = SWAP_PCTRIE_LOOKUP(&m->object->un_pager.swp.swp_blks, 1250 rounddown(m->pindex, SWAP_META_PAGES)); 1251 if (sb == NULL) 1252 return; 1253 range.start = sb->d[m->pindex % SWAP_META_PAGES]; 1254 if (range.start == SWAPBLK_NONE) 1255 return; 1256 range.num = 1; 1257 swp_pager_freeswapspace(&range); 1258 sb->d[m->pindex % SWAP_META_PAGES] = SWAPBLK_NONE; 1259 swp_pager_free_empty_swblk(m->object, sb); 1260 } 1261 1262 /* 1263 * swap_pager_getpages() - bring pages in from swap 1264 * 1265 * Attempt to page in the pages in array "ma" of length "count". The 1266 * caller may optionally specify that additional pages preceding and 1267 * succeeding the specified range be paged in. The number of such pages 1268 * is returned in the "rbehind" and "rahead" parameters, and they will 1269 * be in the inactive queue upon return. 1270 * 1271 * The pages in "ma" must be busied and will remain busied upon return. 1272 */ 1273 static int 1274 swap_pager_getpages_locked(vm_object_t object, vm_page_t *ma, int count, 1275 int *rbehind, int *rahead) 1276 { 1277 struct buf *bp; 1278 vm_page_t bm, mpred, msucc, p; 1279 vm_pindex_t pindex; 1280 daddr_t blk; 1281 int i, maxahead, maxbehind, reqcount; 1282 1283 VM_OBJECT_ASSERT_WLOCKED(object); 1284 reqcount = count; 1285 1286 KASSERT((object->flags & OBJ_SWAP) != 0, 1287 ("%s: object not swappable", __func__)); 1288 if (!swap_pager_haspage(object, ma[0]->pindex, &maxbehind, &maxahead)) { 1289 VM_OBJECT_WUNLOCK(object); 1290 return (VM_PAGER_FAIL); 1291 } 1292 1293 KASSERT(reqcount - 1 <= maxahead, 1294 ("page count %d extends beyond swap block", reqcount)); 1295 1296 /* 1297 * Do not transfer any pages other than those that are xbusied 1298 * when running during a split or collapse operation. This 1299 * prevents clustering from re-creating pages which are being 1300 * moved into another object. 1301 */ 1302 if ((object->flags & (OBJ_SPLIT | OBJ_DEAD)) != 0) { 1303 maxahead = reqcount - 1; 1304 maxbehind = 0; 1305 } 1306 1307 /* 1308 * Clip the readahead and readbehind ranges to exclude resident pages. 1309 */ 1310 if (rahead != NULL) { 1311 *rahead = imin(*rahead, maxahead - (reqcount - 1)); 1312 pindex = ma[reqcount - 1]->pindex; 1313 msucc = TAILQ_NEXT(ma[reqcount - 1], listq); 1314 if (msucc != NULL && msucc->pindex - pindex - 1 < *rahead) 1315 *rahead = msucc->pindex - pindex - 1; 1316 } 1317 if (rbehind != NULL) { 1318 *rbehind = imin(*rbehind, maxbehind); 1319 pindex = ma[0]->pindex; 1320 mpred = TAILQ_PREV(ma[0], pglist, listq); 1321 if (mpred != NULL && pindex - mpred->pindex - 1 < *rbehind) 1322 *rbehind = pindex - mpred->pindex - 1; 1323 } 1324 1325 bm = ma[0]; 1326 for (i = 0; i < count; i++) 1327 ma[i]->oflags |= VPO_SWAPINPROG; 1328 1329 /* 1330 * Allocate readahead and readbehind pages. 1331 */ 1332 if (rbehind != NULL) { 1333 for (i = 1; i <= *rbehind; i++) { 1334 p = vm_page_alloc(object, ma[0]->pindex - i, 1335 VM_ALLOC_NORMAL); 1336 if (p == NULL) 1337 break; 1338 p->oflags |= VPO_SWAPINPROG; 1339 bm = p; 1340 } 1341 *rbehind = i - 1; 1342 } 1343 if (rahead != NULL) { 1344 for (i = 0; i < *rahead; i++) { 1345 p = vm_page_alloc(object, 1346 ma[reqcount - 1]->pindex + i + 1, VM_ALLOC_NORMAL); 1347 if (p == NULL) 1348 break; 1349 p->oflags |= VPO_SWAPINPROG; 1350 } 1351 *rahead = i; 1352 } 1353 if (rbehind != NULL) 1354 count += *rbehind; 1355 if (rahead != NULL) 1356 count += *rahead; 1357 1358 vm_object_pip_add(object, count); 1359 1360 pindex = bm->pindex; 1361 blk = swp_pager_meta_lookup(object, pindex); 1362 KASSERT(blk != SWAPBLK_NONE, 1363 ("no swap blocking containing %p(%jx)", object, (uintmax_t)pindex)); 1364 1365 VM_OBJECT_WUNLOCK(object); 1366 bp = uma_zalloc(swrbuf_zone, M_WAITOK); 1367 MPASS((bp->b_flags & B_MAXPHYS) != 0); 1368 /* Pages cannot leave the object while busy. */ 1369 for (i = 0, p = bm; i < count; i++, p = TAILQ_NEXT(p, listq)) { 1370 MPASS(p->pindex == bm->pindex + i); 1371 bp->b_pages[i] = p; 1372 } 1373 1374 bp->b_flags |= B_PAGING; 1375 bp->b_iocmd = BIO_READ; 1376 bp->b_iodone = swp_pager_async_iodone; 1377 bp->b_rcred = crhold(thread0.td_ucred); 1378 bp->b_wcred = crhold(thread0.td_ucred); 1379 bp->b_blkno = blk; 1380 bp->b_bcount = PAGE_SIZE * count; 1381 bp->b_bufsize = PAGE_SIZE * count; 1382 bp->b_npages = count; 1383 bp->b_pgbefore = rbehind != NULL ? *rbehind : 0; 1384 bp->b_pgafter = rahead != NULL ? *rahead : 0; 1385 1386 VM_CNT_INC(v_swapin); 1387 VM_CNT_ADD(v_swappgsin, count); 1388 1389 /* 1390 * perform the I/O. NOTE!!! bp cannot be considered valid after 1391 * this point because we automatically release it on completion. 1392 * Instead, we look at the one page we are interested in which we 1393 * still hold a lock on even through the I/O completion. 1394 * 1395 * The other pages in our ma[] array are also released on completion, 1396 * so we cannot assume they are valid anymore either. 1397 * 1398 * NOTE: b_blkno is destroyed by the call to swapdev_strategy 1399 */ 1400 BUF_KERNPROC(bp); 1401 swp_pager_strategy(bp); 1402 1403 /* 1404 * Wait for the pages we want to complete. VPO_SWAPINPROG is always 1405 * cleared on completion. If an I/O error occurs, SWAPBLK_NONE 1406 * is set in the metadata for each page in the request. 1407 */ 1408 VM_OBJECT_WLOCK(object); 1409 /* This could be implemented more efficiently with aflags */ 1410 while ((ma[0]->oflags & VPO_SWAPINPROG) != 0) { 1411 ma[0]->oflags |= VPO_SWAPSLEEP; 1412 VM_CNT_INC(v_intrans); 1413 if (VM_OBJECT_SLEEP(object, &object->handle, PSWP, 1414 "swread", hz * 20)) { 1415 printf( 1416 "swap_pager: indefinite wait buffer: bufobj: %p, blkno: %jd, size: %ld\n", 1417 bp->b_bufobj, (intmax_t)bp->b_blkno, bp->b_bcount); 1418 } 1419 } 1420 VM_OBJECT_WUNLOCK(object); 1421 1422 /* 1423 * If we had an unrecoverable read error pages will not be valid. 1424 */ 1425 for (i = 0; i < reqcount; i++) 1426 if (ma[i]->valid != VM_PAGE_BITS_ALL) 1427 return (VM_PAGER_ERROR); 1428 1429 return (VM_PAGER_OK); 1430 1431 /* 1432 * A final note: in a low swap situation, we cannot deallocate swap 1433 * and mark a page dirty here because the caller is likely to mark 1434 * the page clean when we return, causing the page to possibly revert 1435 * to all-zero's later. 1436 */ 1437 } 1438 1439 static int 1440 swap_pager_getpages(vm_object_t object, vm_page_t *ma, int count, 1441 int *rbehind, int *rahead) 1442 { 1443 1444 VM_OBJECT_WLOCK(object); 1445 return (swap_pager_getpages_locked(object, ma, count, rbehind, rahead)); 1446 } 1447 1448 /* 1449 * swap_pager_getpages_async(): 1450 * 1451 * Right now this is emulation of asynchronous operation on top of 1452 * swap_pager_getpages(). 1453 */ 1454 static int 1455 swap_pager_getpages_async(vm_object_t object, vm_page_t *ma, int count, 1456 int *rbehind, int *rahead, pgo_getpages_iodone_t iodone, void *arg) 1457 { 1458 int r, error; 1459 1460 r = swap_pager_getpages(object, ma, count, rbehind, rahead); 1461 switch (r) { 1462 case VM_PAGER_OK: 1463 error = 0; 1464 break; 1465 case VM_PAGER_ERROR: 1466 error = EIO; 1467 break; 1468 case VM_PAGER_FAIL: 1469 error = EINVAL; 1470 break; 1471 default: 1472 panic("unhandled swap_pager_getpages() error %d", r); 1473 } 1474 (iodone)(arg, ma, count, error); 1475 1476 return (r); 1477 } 1478 1479 /* 1480 * swap_pager_putpages: 1481 * 1482 * Assign swap (if necessary) and initiate I/O on the specified pages. 1483 * 1484 * In a low memory situation we may block in VOP_STRATEGY(), but the new 1485 * vm_page reservation system coupled with properly written VFS devices 1486 * should ensure that no low-memory deadlock occurs. This is an area 1487 * which needs work. 1488 * 1489 * The parent has N vm_object_pip_add() references prior to 1490 * calling us and will remove references for rtvals[] that are 1491 * not set to VM_PAGER_PEND. We need to remove the rest on I/O 1492 * completion. 1493 * 1494 * The parent has soft-busy'd the pages it passes us and will unbusy 1495 * those whose rtvals[] entry is not set to VM_PAGER_PEND on return. 1496 * We need to unbusy the rest on I/O completion. 1497 */ 1498 static void 1499 swap_pager_putpages(vm_object_t object, vm_page_t *ma, int count, 1500 int flags, int *rtvals) 1501 { 1502 struct page_range range; 1503 struct buf *bp; 1504 daddr_t addr, blk; 1505 vm_page_t mreq; 1506 int i, j, n; 1507 bool async; 1508 1509 KASSERT(count == 0 || ma[0]->object == object, 1510 ("%s: object mismatch %p/%p", 1511 __func__, object, ma[0]->object)); 1512 1513 VM_OBJECT_WUNLOCK(object); 1514 async = curproc == pageproc && (flags & VM_PAGER_PUT_SYNC) == 0; 1515 swp_pager_init_freerange(&range); 1516 1517 /* 1518 * Assign swap blocks and issue I/O. We reallocate swap on the fly. 1519 * The page is left dirty until the pageout operation completes 1520 * successfully. 1521 */ 1522 for (i = 0; i < count; i += n) { 1523 /* Maximum I/O size is limited by maximum swap block size. */ 1524 n = min(count - i, nsw_cluster_max); 1525 1526 if (async) { 1527 mtx_lock(&swbuf_mtx); 1528 while (nsw_wcount_async == 0) 1529 msleep(&nsw_wcount_async, &swbuf_mtx, PVM, 1530 "swbufa", 0); 1531 nsw_wcount_async--; 1532 mtx_unlock(&swbuf_mtx); 1533 } 1534 1535 /* Get a block of swap of size up to size n. */ 1536 blk = swp_pager_getswapspace(&n); 1537 if (blk == SWAPBLK_NONE) { 1538 mtx_lock(&swbuf_mtx); 1539 if (++nsw_wcount_async == 1) 1540 wakeup(&nsw_wcount_async); 1541 mtx_unlock(&swbuf_mtx); 1542 for (j = 0; j < n; ++j) 1543 rtvals[i + j] = VM_PAGER_FAIL; 1544 continue; 1545 } 1546 VM_OBJECT_WLOCK(object); 1547 for (j = 0; j < n; ++j) { 1548 mreq = ma[i + j]; 1549 vm_page_aflag_clear(mreq, PGA_SWAP_FREE); 1550 addr = swp_pager_meta_build(mreq->object, mreq->pindex, 1551 blk + j); 1552 if (addr != SWAPBLK_NONE) 1553 swp_pager_update_freerange(&range, addr); 1554 MPASS(mreq->dirty == VM_PAGE_BITS_ALL); 1555 mreq->oflags |= VPO_SWAPINPROG; 1556 } 1557 VM_OBJECT_WUNLOCK(object); 1558 1559 bp = uma_zalloc(swwbuf_zone, M_WAITOK); 1560 MPASS((bp->b_flags & B_MAXPHYS) != 0); 1561 if (async) 1562 bp->b_flags |= B_ASYNC; 1563 bp->b_flags |= B_PAGING; 1564 bp->b_iocmd = BIO_WRITE; 1565 1566 bp->b_rcred = crhold(thread0.td_ucred); 1567 bp->b_wcred = crhold(thread0.td_ucred); 1568 bp->b_bcount = PAGE_SIZE * n; 1569 bp->b_bufsize = PAGE_SIZE * n; 1570 bp->b_blkno = blk; 1571 for (j = 0; j < n; j++) 1572 bp->b_pages[j] = ma[i + j]; 1573 bp->b_npages = n; 1574 1575 /* 1576 * Must set dirty range for NFS to work. 1577 */ 1578 bp->b_dirtyoff = 0; 1579 bp->b_dirtyend = bp->b_bcount; 1580 1581 VM_CNT_INC(v_swapout); 1582 VM_CNT_ADD(v_swappgsout, bp->b_npages); 1583 1584 /* 1585 * We unconditionally set rtvals[] to VM_PAGER_PEND so that we 1586 * can call the async completion routine at the end of a 1587 * synchronous I/O operation. Otherwise, our caller would 1588 * perform duplicate unbusy and wakeup operations on the page 1589 * and object, respectively. 1590 */ 1591 for (j = 0; j < n; j++) 1592 rtvals[i + j] = VM_PAGER_PEND; 1593 1594 /* 1595 * asynchronous 1596 * 1597 * NOTE: b_blkno is destroyed by the call to swapdev_strategy. 1598 */ 1599 if (async) { 1600 bp->b_iodone = swp_pager_async_iodone; 1601 BUF_KERNPROC(bp); 1602 swp_pager_strategy(bp); 1603 continue; 1604 } 1605 1606 /* 1607 * synchronous 1608 * 1609 * NOTE: b_blkno is destroyed by the call to swapdev_strategy. 1610 */ 1611 bp->b_iodone = bdone; 1612 swp_pager_strategy(bp); 1613 1614 /* 1615 * Wait for the sync I/O to complete. 1616 */ 1617 bwait(bp, PVM, "swwrt"); 1618 1619 /* 1620 * Now that we are through with the bp, we can call the 1621 * normal async completion, which frees everything up. 1622 */ 1623 swp_pager_async_iodone(bp); 1624 } 1625 swp_pager_freeswapspace(&range); 1626 VM_OBJECT_WLOCK(object); 1627 } 1628 1629 /* 1630 * swp_pager_async_iodone: 1631 * 1632 * Completion routine for asynchronous reads and writes from/to swap. 1633 * Also called manually by synchronous code to finish up a bp. 1634 * 1635 * This routine may not sleep. 1636 */ 1637 static void 1638 swp_pager_async_iodone(struct buf *bp) 1639 { 1640 int i; 1641 vm_object_t object = NULL; 1642 1643 /* 1644 * Report error - unless we ran out of memory, in which case 1645 * we've already logged it in swapgeom_strategy(). 1646 */ 1647 if (bp->b_ioflags & BIO_ERROR && bp->b_error != ENOMEM) { 1648 printf( 1649 "swap_pager: I/O error - %s failed; blkno %ld," 1650 "size %ld, error %d\n", 1651 ((bp->b_iocmd == BIO_READ) ? "pagein" : "pageout"), 1652 (long)bp->b_blkno, 1653 (long)bp->b_bcount, 1654 bp->b_error 1655 ); 1656 } 1657 1658 /* 1659 * remove the mapping for kernel virtual 1660 */ 1661 if (buf_mapped(bp)) 1662 pmap_qremove((vm_offset_t)bp->b_data, bp->b_npages); 1663 else 1664 bp->b_data = bp->b_kvabase; 1665 1666 if (bp->b_npages) { 1667 object = bp->b_pages[0]->object; 1668 VM_OBJECT_WLOCK(object); 1669 } 1670 1671 /* 1672 * cleanup pages. If an error occurs writing to swap, we are in 1673 * very serious trouble. If it happens to be a disk error, though, 1674 * we may be able to recover by reassigning the swap later on. So 1675 * in this case we remove the m->swapblk assignment for the page 1676 * but do not free it in the rlist. The errornous block(s) are thus 1677 * never reallocated as swap. Redirty the page and continue. 1678 */ 1679 for (i = 0; i < bp->b_npages; ++i) { 1680 vm_page_t m = bp->b_pages[i]; 1681 1682 m->oflags &= ~VPO_SWAPINPROG; 1683 if (m->oflags & VPO_SWAPSLEEP) { 1684 m->oflags &= ~VPO_SWAPSLEEP; 1685 wakeup(&object->handle); 1686 } 1687 1688 /* We always have space after I/O, successful or not. */ 1689 vm_page_aflag_set(m, PGA_SWAP_SPACE); 1690 1691 if (bp->b_ioflags & BIO_ERROR) { 1692 /* 1693 * If an error occurs I'd love to throw the swapblk 1694 * away without freeing it back to swapspace, so it 1695 * can never be used again. But I can't from an 1696 * interrupt. 1697 */ 1698 if (bp->b_iocmd == BIO_READ) { 1699 /* 1700 * NOTE: for reads, m->dirty will probably 1701 * be overridden by the original caller of 1702 * getpages so don't play cute tricks here. 1703 */ 1704 vm_page_invalid(m); 1705 if (i < bp->b_pgbefore || 1706 i >= bp->b_npages - bp->b_pgafter) 1707 vm_page_free_invalid(m); 1708 } else { 1709 /* 1710 * If a write error occurs, reactivate page 1711 * so it doesn't clog the inactive list, 1712 * then finish the I/O. 1713 */ 1714 MPASS(m->dirty == VM_PAGE_BITS_ALL); 1715 1716 /* PQ_UNSWAPPABLE? */ 1717 vm_page_activate(m); 1718 vm_page_sunbusy(m); 1719 } 1720 } else if (bp->b_iocmd == BIO_READ) { 1721 /* 1722 * NOTE: for reads, m->dirty will probably be 1723 * overridden by the original caller of getpages so 1724 * we cannot set them in order to free the underlying 1725 * swap in a low-swap situation. I don't think we'd 1726 * want to do that anyway, but it was an optimization 1727 * that existed in the old swapper for a time before 1728 * it got ripped out due to precisely this problem. 1729 */ 1730 KASSERT(!pmap_page_is_mapped(m), 1731 ("swp_pager_async_iodone: page %p is mapped", m)); 1732 KASSERT(m->dirty == 0, 1733 ("swp_pager_async_iodone: page %p is dirty", m)); 1734 1735 vm_page_valid(m); 1736 if (i < bp->b_pgbefore || 1737 i >= bp->b_npages - bp->b_pgafter) 1738 vm_page_readahead_finish(m); 1739 } else { 1740 /* 1741 * For write success, clear the dirty 1742 * status, then finish the I/O ( which decrements the 1743 * busy count and possibly wakes waiter's up ). 1744 * A page is only written to swap after a period of 1745 * inactivity. Therefore, we do not expect it to be 1746 * reused. 1747 */ 1748 KASSERT(!pmap_page_is_write_mapped(m), 1749 ("swp_pager_async_iodone: page %p is not write" 1750 " protected", m)); 1751 vm_page_undirty(m); 1752 vm_page_deactivate_noreuse(m); 1753 vm_page_sunbusy(m); 1754 } 1755 } 1756 1757 /* 1758 * adjust pip. NOTE: the original parent may still have its own 1759 * pip refs on the object. 1760 */ 1761 if (object != NULL) { 1762 vm_object_pip_wakeupn(object, bp->b_npages); 1763 VM_OBJECT_WUNLOCK(object); 1764 } 1765 1766 /* 1767 * swapdev_strategy() manually sets b_vp and b_bufobj before calling 1768 * bstrategy(). Set them back to NULL now we're done with it, or we'll 1769 * trigger a KASSERT in relpbuf(). 1770 */ 1771 if (bp->b_vp) { 1772 bp->b_vp = NULL; 1773 bp->b_bufobj = NULL; 1774 } 1775 /* 1776 * release the physical I/O buffer 1777 */ 1778 if (bp->b_flags & B_ASYNC) { 1779 mtx_lock(&swbuf_mtx); 1780 if (++nsw_wcount_async == 1) 1781 wakeup(&nsw_wcount_async); 1782 mtx_unlock(&swbuf_mtx); 1783 } 1784 uma_zfree((bp->b_iocmd == BIO_READ) ? swrbuf_zone : swwbuf_zone, bp); 1785 } 1786 1787 int 1788 swap_pager_nswapdev(void) 1789 { 1790 1791 return (nswapdev); 1792 } 1793 1794 static void 1795 swp_pager_force_dirty(struct page_range *range, vm_page_t m, daddr_t *blk) 1796 { 1797 vm_page_dirty(m); 1798 swap_pager_unswapped_acct(m); 1799 swp_pager_update_freerange(range, *blk); 1800 *blk = SWAPBLK_NONE; 1801 vm_page_launder(m); 1802 } 1803 1804 u_long 1805 swap_pager_swapped_pages(vm_object_t object) 1806 { 1807 struct swblk *sb; 1808 vm_pindex_t pi; 1809 u_long res; 1810 int i; 1811 1812 VM_OBJECT_ASSERT_LOCKED(object); 1813 1814 if (pctrie_is_empty(&object->un_pager.swp.swp_blks)) 1815 return (0); 1816 1817 for (res = 0, pi = 0; (sb = SWAP_PCTRIE_LOOKUP_GE( 1818 &object->un_pager.swp.swp_blks, pi)) != NULL; 1819 pi = sb->p + SWAP_META_PAGES) { 1820 for (i = 0; i < SWAP_META_PAGES; i++) { 1821 if (sb->d[i] != SWAPBLK_NONE) 1822 res++; 1823 } 1824 } 1825 return (res); 1826 } 1827 1828 /* 1829 * swap_pager_swapoff_object: 1830 * 1831 * Page in all of the pages that have been paged out for an object 1832 * to a swap device. 1833 */ 1834 static void 1835 swap_pager_swapoff_object(struct swdevt *sp, vm_object_t object) 1836 { 1837 struct page_range range; 1838 struct swblk *sb; 1839 vm_page_t m; 1840 vm_pindex_t pi; 1841 daddr_t blk; 1842 int i, rahead, rv; 1843 bool sb_empty; 1844 1845 VM_OBJECT_ASSERT_WLOCKED(object); 1846 KASSERT((object->flags & OBJ_SWAP) != 0, 1847 ("%s: Object not swappable", __func__)); 1848 1849 pi = 0; 1850 i = 0; 1851 swp_pager_init_freerange(&range); 1852 for (;;) { 1853 if (i == 0 && (object->flags & OBJ_DEAD) != 0) { 1854 /* 1855 * Make sure that pending writes finish before 1856 * returning. 1857 */ 1858 vm_object_pip_wait(object, "swpoff"); 1859 swp_pager_meta_free_all(object); 1860 break; 1861 } 1862 1863 if (i == SWAP_META_PAGES) { 1864 pi = sb->p + SWAP_META_PAGES; 1865 if (sb_empty) { 1866 SWAP_PCTRIE_REMOVE( 1867 &object->un_pager.swp.swp_blks, sb->p); 1868 uma_zfree(swblk_zone, sb); 1869 } 1870 i = 0; 1871 } 1872 1873 if (i == 0) { 1874 sb = SWAP_PCTRIE_LOOKUP_GE( 1875 &object->un_pager.swp.swp_blks, pi); 1876 if (sb == NULL) 1877 break; 1878 sb_empty = true; 1879 m = NULL; 1880 } 1881 1882 /* Skip an invalid block. */ 1883 blk = sb->d[i]; 1884 if (blk == SWAPBLK_NONE || !swp_pager_isondev(blk, sp)) { 1885 if (blk != SWAPBLK_NONE) 1886 sb_empty = false; 1887 m = NULL; 1888 i++; 1889 continue; 1890 } 1891 1892 /* 1893 * Look for a page corresponding to this block, If the found 1894 * page has pending operations, sleep and restart the scan. 1895 */ 1896 m = m != NULL ? vm_page_next(m) : 1897 vm_page_lookup(object, sb->p + i); 1898 if (m != NULL && (m->oflags & VPO_SWAPINPROG) != 0) { 1899 m->oflags |= VPO_SWAPSLEEP; 1900 VM_OBJECT_SLEEP(object, &object->handle, PSWP, "swpoff", 1901 0); 1902 i = 0; /* Restart scan after object lock dropped. */ 1903 continue; 1904 } 1905 1906 /* 1907 * If the found page is valid, mark it dirty and free the swap 1908 * block. 1909 */ 1910 if (m != NULL && vm_page_all_valid(m)) { 1911 swp_pager_force_dirty(&range, m, &sb->d[i]); 1912 i++; 1913 continue; 1914 } 1915 1916 /* Is there a page we can acquire or allocate? */ 1917 if (m == NULL) { 1918 m = vm_page_alloc(object, sb->p + i, 1919 VM_ALLOC_NORMAL | VM_ALLOC_WAITFAIL); 1920 } else if (!vm_page_busy_acquire(m, VM_ALLOC_WAITFAIL)) 1921 m = NULL; 1922 1923 /* If no page available, repeat this iteration. */ 1924 if (m == NULL) 1925 continue; 1926 1927 /* Get the page from swap, mark it dirty, restart the scan. */ 1928 vm_object_pip_add(object, 1); 1929 rahead = SWAP_META_PAGES; 1930 rv = swap_pager_getpages_locked(object, &m, 1, NULL, &rahead); 1931 if (rv != VM_PAGER_OK) 1932 panic("%s: read from swap failed: %d", __func__, rv); 1933 VM_OBJECT_WLOCK(object); 1934 vm_object_pip_wakeupn(object, 1); 1935 KASSERT(vm_page_all_valid(m), 1936 ("%s: Page %p not all valid", __func__, m)); 1937 swp_pager_force_dirty(&range, m, &sb->d[i]); 1938 vm_page_xunbusy(m); 1939 i = 0; /* Restart scan after object lock dropped. */ 1940 } 1941 swp_pager_freeswapspace(&range); 1942 } 1943 1944 /* 1945 * swap_pager_swapoff: 1946 * 1947 * Page in all of the pages that have been paged out to the 1948 * given device. The corresponding blocks in the bitmap must be 1949 * marked as allocated and the device must be flagged SW_CLOSING. 1950 * There may be no processes swapped out to the device. 1951 * 1952 * This routine may block. 1953 */ 1954 static void 1955 swap_pager_swapoff(struct swdevt *sp) 1956 { 1957 vm_object_t object; 1958 int retries; 1959 1960 sx_assert(&swdev_syscall_lock, SA_XLOCKED); 1961 1962 retries = 0; 1963 full_rescan: 1964 mtx_lock(&vm_object_list_mtx); 1965 TAILQ_FOREACH(object, &vm_object_list, object_list) { 1966 if ((object->flags & OBJ_SWAP) == 0) 1967 continue; 1968 mtx_unlock(&vm_object_list_mtx); 1969 /* Depends on type-stability. */ 1970 VM_OBJECT_WLOCK(object); 1971 1972 /* 1973 * Dead objects are eventually terminated on their own. 1974 */ 1975 if ((object->flags & OBJ_DEAD) != 0) 1976 goto next_obj; 1977 1978 /* 1979 * Sync with fences placed after pctrie 1980 * initialization. We must not access pctrie below 1981 * unless we checked that our object is swap and not 1982 * dead. 1983 */ 1984 atomic_thread_fence_acq(); 1985 if ((object->flags & OBJ_SWAP) == 0) 1986 goto next_obj; 1987 1988 swap_pager_swapoff_object(sp, object); 1989 next_obj: 1990 VM_OBJECT_WUNLOCK(object); 1991 mtx_lock(&vm_object_list_mtx); 1992 } 1993 mtx_unlock(&vm_object_list_mtx); 1994 1995 if (sp->sw_used) { 1996 /* 1997 * Objects may be locked or paging to the device being 1998 * removed, so we will miss their pages and need to 1999 * make another pass. We have marked this device as 2000 * SW_CLOSING, so the activity should finish soon. 2001 */ 2002 retries++; 2003 if (retries > 100) { 2004 panic("swapoff: failed to locate %d swap blocks", 2005 sp->sw_used); 2006 } 2007 pause("swpoff", hz / 20); 2008 goto full_rescan; 2009 } 2010 EVENTHANDLER_INVOKE(swapoff, sp); 2011 } 2012 2013 /************************************************************************ 2014 * SWAP META DATA * 2015 ************************************************************************ 2016 * 2017 * These routines manipulate the swap metadata stored in the 2018 * OBJT_SWAP object. 2019 * 2020 * Swap metadata is implemented with a global hash and not directly 2021 * linked into the object. Instead the object simply contains 2022 * appropriate tracking counters. 2023 */ 2024 2025 /* 2026 * SWP_PAGER_SWBLK_EMPTY() - is a range of blocks free? 2027 */ 2028 static bool 2029 swp_pager_swblk_empty(struct swblk *sb, int start, int limit) 2030 { 2031 int i; 2032 2033 MPASS(0 <= start && start <= limit && limit <= SWAP_META_PAGES); 2034 for (i = start; i < limit; i++) { 2035 if (sb->d[i] != SWAPBLK_NONE) 2036 return (false); 2037 } 2038 return (true); 2039 } 2040 2041 /* 2042 * SWP_PAGER_FREE_EMPTY_SWBLK() - frees if a block is free 2043 * 2044 * Nothing is done if the block is still in use. 2045 */ 2046 static void 2047 swp_pager_free_empty_swblk(vm_object_t object, struct swblk *sb) 2048 { 2049 2050 if (swp_pager_swblk_empty(sb, 0, SWAP_META_PAGES)) { 2051 SWAP_PCTRIE_REMOVE(&object->un_pager.swp.swp_blks, sb->p); 2052 uma_zfree(swblk_zone, sb); 2053 } 2054 } 2055 2056 /* 2057 * SWP_PAGER_META_BUILD() - add swap block to swap meta data for object 2058 * 2059 * The specified swapblk is added to the object's swap metadata. If 2060 * the swapblk is not valid, it is freed instead. Any previously 2061 * assigned swapblk is returned. 2062 */ 2063 static daddr_t 2064 swp_pager_meta_build(vm_object_t object, vm_pindex_t pindex, daddr_t swapblk) 2065 { 2066 static volatile int swblk_zone_exhausted, swpctrie_zone_exhausted; 2067 struct swblk *sb, *sb1; 2068 vm_pindex_t modpi, rdpi; 2069 daddr_t prev_swapblk; 2070 int error, i; 2071 2072 VM_OBJECT_ASSERT_WLOCKED(object); 2073 2074 rdpi = rounddown(pindex, SWAP_META_PAGES); 2075 sb = SWAP_PCTRIE_LOOKUP(&object->un_pager.swp.swp_blks, rdpi); 2076 if (sb == NULL) { 2077 if (swapblk == SWAPBLK_NONE) 2078 return (SWAPBLK_NONE); 2079 for (;;) { 2080 sb = uma_zalloc(swblk_zone, M_NOWAIT | (curproc == 2081 pageproc ? M_USE_RESERVE : 0)); 2082 if (sb != NULL) { 2083 sb->p = rdpi; 2084 for (i = 0; i < SWAP_META_PAGES; i++) 2085 sb->d[i] = SWAPBLK_NONE; 2086 if (atomic_cmpset_int(&swblk_zone_exhausted, 2087 1, 0)) 2088 printf("swblk zone ok\n"); 2089 break; 2090 } 2091 VM_OBJECT_WUNLOCK(object); 2092 if (uma_zone_exhausted(swblk_zone)) { 2093 if (atomic_cmpset_int(&swblk_zone_exhausted, 2094 0, 1)) 2095 printf("swap blk zone exhausted, " 2096 "increase kern.maxswzone\n"); 2097 vm_pageout_oom(VM_OOM_SWAPZ); 2098 pause("swzonxb", 10); 2099 } else 2100 uma_zwait(swblk_zone); 2101 VM_OBJECT_WLOCK(object); 2102 sb = SWAP_PCTRIE_LOOKUP(&object->un_pager.swp.swp_blks, 2103 rdpi); 2104 if (sb != NULL) 2105 /* 2106 * Somebody swapped out a nearby page, 2107 * allocating swblk at the rdpi index, 2108 * while we dropped the object lock. 2109 */ 2110 goto allocated; 2111 } 2112 for (;;) { 2113 error = SWAP_PCTRIE_INSERT( 2114 &object->un_pager.swp.swp_blks, sb); 2115 if (error == 0) { 2116 if (atomic_cmpset_int(&swpctrie_zone_exhausted, 2117 1, 0)) 2118 printf("swpctrie zone ok\n"); 2119 break; 2120 } 2121 VM_OBJECT_WUNLOCK(object); 2122 if (uma_zone_exhausted(swpctrie_zone)) { 2123 if (atomic_cmpset_int(&swpctrie_zone_exhausted, 2124 0, 1)) 2125 printf("swap pctrie zone exhausted, " 2126 "increase kern.maxswzone\n"); 2127 vm_pageout_oom(VM_OOM_SWAPZ); 2128 pause("swzonxp", 10); 2129 } else 2130 uma_zwait(swpctrie_zone); 2131 VM_OBJECT_WLOCK(object); 2132 sb1 = SWAP_PCTRIE_LOOKUP(&object->un_pager.swp.swp_blks, 2133 rdpi); 2134 if (sb1 != NULL) { 2135 uma_zfree(swblk_zone, sb); 2136 sb = sb1; 2137 goto allocated; 2138 } 2139 } 2140 } 2141 allocated: 2142 MPASS(sb->p == rdpi); 2143 2144 modpi = pindex % SWAP_META_PAGES; 2145 /* Return prior contents of metadata. */ 2146 prev_swapblk = sb->d[modpi]; 2147 /* Enter block into metadata. */ 2148 sb->d[modpi] = swapblk; 2149 2150 /* 2151 * Free the swblk if we end up with the empty page run. 2152 */ 2153 if (swapblk == SWAPBLK_NONE) 2154 swp_pager_free_empty_swblk(object, sb); 2155 return (prev_swapblk); 2156 } 2157 2158 /* 2159 * SWP_PAGER_META_TRANSFER() - free a range of blocks in the srcobject's swap 2160 * metadata, or transfer it into dstobject. 2161 * 2162 * This routine will free swap metadata structures as they are cleaned 2163 * out. 2164 */ 2165 static void 2166 swp_pager_meta_transfer(vm_object_t srcobject, vm_object_t dstobject, 2167 vm_pindex_t pindex, vm_pindex_t count, vm_size_t *moved) 2168 { 2169 struct page_range range; 2170 struct swblk *sb; 2171 vm_page_t m; 2172 vm_pindex_t offset, last; 2173 vm_size_t mc; 2174 int i, limit, start; 2175 2176 VM_OBJECT_ASSERT_WLOCKED(srcobject); 2177 MPASS(moved == NULL || dstobject == NULL); 2178 2179 mc = 0; 2180 m = NULL; 2181 if (count == 0 || pctrie_is_empty(&srcobject->un_pager.swp.swp_blks)) 2182 goto out; 2183 2184 swp_pager_init_freerange(&range); 2185 offset = pindex; 2186 last = pindex + count; 2187 for (;;) { 2188 sb = SWAP_PCTRIE_LOOKUP_GE(&srcobject->un_pager.swp.swp_blks, 2189 rounddown(pindex, SWAP_META_PAGES)); 2190 if (sb == NULL || sb->p >= last) 2191 break; 2192 start = pindex > sb->p ? pindex - sb->p : 0; 2193 limit = last - sb->p < SWAP_META_PAGES ? last - sb->p : 2194 SWAP_META_PAGES; 2195 for (i = start; i < limit; i++) { 2196 if (sb->d[i] == SWAPBLK_NONE) 2197 continue; 2198 if (dstobject == NULL || 2199 !swp_pager_xfer_source(srcobject, dstobject, 2200 sb->p + i - offset, sb->d[i])) { 2201 swp_pager_update_freerange(&range, sb->d[i]); 2202 } 2203 if (moved != NULL) { 2204 if (m != NULL && m->pindex != pindex + i - 1) 2205 m = NULL; 2206 m = m != NULL ? vm_page_next(m) : 2207 vm_page_lookup(srcobject, pindex + i); 2208 if (m == NULL || vm_page_none_valid(m)) 2209 mc++; 2210 } 2211 sb->d[i] = SWAPBLK_NONE; 2212 } 2213 pindex = sb->p + SWAP_META_PAGES; 2214 if (swp_pager_swblk_empty(sb, 0, start) && 2215 swp_pager_swblk_empty(sb, limit, SWAP_META_PAGES)) { 2216 SWAP_PCTRIE_REMOVE(&srcobject->un_pager.swp.swp_blks, 2217 sb->p); 2218 uma_zfree(swblk_zone, sb); 2219 } 2220 } 2221 swp_pager_freeswapspace(&range); 2222 out: 2223 if (moved != NULL) 2224 *moved = mc; 2225 } 2226 2227 /* 2228 * SWP_PAGER_META_FREE() - free a range of blocks in the object's swap metadata 2229 * 2230 * The requested range of blocks is freed, with any associated swap 2231 * returned to the swap bitmap. 2232 * 2233 * This routine will free swap metadata structures as they are cleaned 2234 * out. This routine does *NOT* operate on swap metadata associated 2235 * with resident pages. 2236 */ 2237 static void 2238 swp_pager_meta_free(vm_object_t object, vm_pindex_t pindex, vm_pindex_t count, 2239 vm_size_t *freed) 2240 { 2241 swp_pager_meta_transfer(object, NULL, pindex, count, freed); 2242 } 2243 2244 static void 2245 swp_pager_meta_free_block(struct swblk *sb, void *rangev) 2246 { 2247 struct page_range *range = rangev; 2248 2249 for (int i = 0; i < SWAP_META_PAGES; i++) { 2250 if (sb->d[i] != SWAPBLK_NONE) 2251 swp_pager_update_freerange(range, sb->d[i]); 2252 } 2253 uma_zfree(swblk_zone, sb); 2254 } 2255 2256 /* 2257 * SWP_PAGER_META_FREE_ALL() - destroy all swap metadata associated with object 2258 * 2259 * This routine locates and destroys all swap metadata associated with 2260 * an object. 2261 */ 2262 static void 2263 swp_pager_meta_free_all(vm_object_t object) 2264 { 2265 struct page_range range; 2266 2267 VM_OBJECT_ASSERT_WLOCKED(object); 2268 2269 swp_pager_init_freerange(&range); 2270 SWAP_PCTRIE_RECLAIM_CALLBACK(&object->un_pager.swp.swp_blks, 2271 swp_pager_meta_free_block, &range); 2272 swp_pager_freeswapspace(&range); 2273 } 2274 2275 /* 2276 * SWP_PAGER_METACTL() - misc control of swap meta data. 2277 * 2278 * This routine is capable of looking up, or removing swapblk 2279 * assignments in the swap meta data. It returns the swapblk being 2280 * looked-up, popped, or SWAPBLK_NONE if the block was invalid. 2281 * 2282 * When acting on a busy resident page and paging is in progress, we 2283 * have to wait until paging is complete but otherwise can act on the 2284 * busy page. 2285 */ 2286 static daddr_t 2287 swp_pager_meta_lookup(vm_object_t object, vm_pindex_t pindex) 2288 { 2289 struct swblk *sb; 2290 2291 VM_OBJECT_ASSERT_LOCKED(object); 2292 2293 /* 2294 * The meta data only exists if the object is OBJT_SWAP 2295 * and even then might not be allocated yet. 2296 */ 2297 KASSERT((object->flags & OBJ_SWAP) != 0, 2298 ("Lookup object not swappable")); 2299 2300 sb = SWAP_PCTRIE_LOOKUP(&object->un_pager.swp.swp_blks, 2301 rounddown(pindex, SWAP_META_PAGES)); 2302 if (sb == NULL) 2303 return (SWAPBLK_NONE); 2304 return (sb->d[pindex % SWAP_META_PAGES]); 2305 } 2306 2307 /* 2308 * Returns the least page index which is greater than or equal to the 2309 * parameter pindex and for which there is a swap block allocated. 2310 * Returns object's size if the object's type is not swap or if there 2311 * are no allocated swap blocks for the object after the requested 2312 * pindex. 2313 */ 2314 vm_pindex_t 2315 swap_pager_find_least(vm_object_t object, vm_pindex_t pindex) 2316 { 2317 struct swblk *sb; 2318 int i; 2319 2320 VM_OBJECT_ASSERT_LOCKED(object); 2321 MPASS((object->flags & OBJ_SWAP) != 0); 2322 2323 if (pctrie_is_empty(&object->un_pager.swp.swp_blks)) 2324 return (object->size); 2325 sb = SWAP_PCTRIE_LOOKUP_GE(&object->un_pager.swp.swp_blks, 2326 rounddown(pindex, SWAP_META_PAGES)); 2327 if (sb == NULL) 2328 return (object->size); 2329 if (sb->p < pindex) { 2330 for (i = pindex % SWAP_META_PAGES; i < SWAP_META_PAGES; i++) { 2331 if (sb->d[i] != SWAPBLK_NONE) 2332 return (sb->p + i); 2333 } 2334 sb = SWAP_PCTRIE_LOOKUP_GE(&object->un_pager.swp.swp_blks, 2335 roundup(pindex, SWAP_META_PAGES)); 2336 if (sb == NULL) 2337 return (object->size); 2338 } 2339 for (i = 0; i < SWAP_META_PAGES; i++) { 2340 if (sb->d[i] != SWAPBLK_NONE) 2341 return (sb->p + i); 2342 } 2343 2344 /* 2345 * We get here if a swblk is present in the trie but it 2346 * doesn't map any blocks. 2347 */ 2348 MPASS(0); 2349 return (object->size); 2350 } 2351 2352 /* 2353 * System call swapon(name) enables swapping on device name, 2354 * which must be in the swdevsw. Return EBUSY 2355 * if already swapping on this device. 2356 */ 2357 #ifndef _SYS_SYSPROTO_H_ 2358 struct swapon_args { 2359 char *name; 2360 }; 2361 #endif 2362 2363 int 2364 sys_swapon(struct thread *td, struct swapon_args *uap) 2365 { 2366 struct vattr attr; 2367 struct vnode *vp; 2368 struct nameidata nd; 2369 int error; 2370 2371 error = priv_check(td, PRIV_SWAPON); 2372 if (error) 2373 return (error); 2374 2375 sx_xlock(&swdev_syscall_lock); 2376 2377 /* 2378 * Swap metadata may not fit in the KVM if we have physical 2379 * memory of >1GB. 2380 */ 2381 if (swblk_zone == NULL) { 2382 error = ENOMEM; 2383 goto done; 2384 } 2385 2386 NDINIT(&nd, LOOKUP, ISOPEN | FOLLOW | LOCKLEAF | AUDITVNODE1, 2387 UIO_USERSPACE, uap->name); 2388 error = namei(&nd); 2389 if (error) 2390 goto done; 2391 2392 NDFREE_PNBUF(&nd); 2393 vp = nd.ni_vp; 2394 2395 if (vn_isdisk_error(vp, &error)) { 2396 error = swapongeom(vp); 2397 } else if (vp->v_type == VREG && 2398 (vp->v_mount->mnt_vfc->vfc_flags & VFCF_NETWORK) != 0 && 2399 (error = VOP_GETATTR(vp, &attr, td->td_ucred)) == 0) { 2400 /* 2401 * Allow direct swapping to NFS regular files in the same 2402 * way that nfs_mountroot() sets up diskless swapping. 2403 */ 2404 error = swaponvp(td, vp, attr.va_size / DEV_BSIZE); 2405 } 2406 2407 if (error != 0) 2408 vput(vp); 2409 else 2410 VOP_UNLOCK(vp); 2411 done: 2412 sx_xunlock(&swdev_syscall_lock); 2413 return (error); 2414 } 2415 2416 /* 2417 * Check that the total amount of swap currently configured does not 2418 * exceed half the theoretical maximum. If it does, print a warning 2419 * message. 2420 */ 2421 static void 2422 swapon_check_swzone(void) 2423 { 2424 2425 /* recommend using no more than half that amount */ 2426 if (swap_total > swap_maxpages / 2) { 2427 printf("warning: total configured swap (%lu pages) " 2428 "exceeds maximum recommended amount (%lu pages).\n", 2429 swap_total, swap_maxpages / 2); 2430 printf("warning: increase kern.maxswzone " 2431 "or reduce amount of swap.\n"); 2432 } 2433 } 2434 2435 static void 2436 swaponsomething(struct vnode *vp, void *id, u_long nblks, 2437 sw_strategy_t *strategy, sw_close_t *close, dev_t dev, int flags) 2438 { 2439 struct swdevt *sp, *tsp; 2440 daddr_t dvbase; 2441 2442 /* 2443 * nblks is in DEV_BSIZE'd chunks, convert to PAGE_SIZE'd chunks. 2444 * First chop nblks off to page-align it, then convert. 2445 * 2446 * sw->sw_nblks is in page-sized chunks now too. 2447 */ 2448 nblks &= ~(ctodb(1) - 1); 2449 nblks = dbtoc(nblks); 2450 2451 sp = malloc(sizeof *sp, M_VMPGDATA, M_WAITOK | M_ZERO); 2452 sp->sw_blist = blist_create(nblks, M_WAITOK); 2453 sp->sw_vp = vp; 2454 sp->sw_id = id; 2455 sp->sw_dev = dev; 2456 sp->sw_nblks = nblks; 2457 sp->sw_used = 0; 2458 sp->sw_strategy = strategy; 2459 sp->sw_close = close; 2460 sp->sw_flags = flags; 2461 2462 /* 2463 * Do not free the first blocks in order to avoid overwriting 2464 * any bsd label at the front of the partition 2465 */ 2466 blist_free(sp->sw_blist, howmany(BBSIZE, PAGE_SIZE), 2467 nblks - howmany(BBSIZE, PAGE_SIZE)); 2468 2469 dvbase = 0; 2470 mtx_lock(&sw_dev_mtx); 2471 TAILQ_FOREACH(tsp, &swtailq, sw_list) { 2472 if (tsp->sw_end >= dvbase) { 2473 /* 2474 * We put one uncovered page between the devices 2475 * in order to definitively prevent any cross-device 2476 * I/O requests 2477 */ 2478 dvbase = tsp->sw_end + 1; 2479 } 2480 } 2481 sp->sw_first = dvbase; 2482 sp->sw_end = dvbase + nblks; 2483 TAILQ_INSERT_TAIL(&swtailq, sp, sw_list); 2484 nswapdev++; 2485 swap_pager_avail += nblks - howmany(BBSIZE, PAGE_SIZE); 2486 swap_total += nblks; 2487 swapon_check_swzone(); 2488 swp_sizecheck(); 2489 mtx_unlock(&sw_dev_mtx); 2490 EVENTHANDLER_INVOKE(swapon, sp); 2491 } 2492 2493 /* 2494 * SYSCALL: swapoff(devname) 2495 * 2496 * Disable swapping on the given device. 2497 * 2498 * XXX: Badly designed system call: it should use a device index 2499 * rather than filename as specification. We keep sw_vp around 2500 * only to make this work. 2501 */ 2502 static int 2503 kern_swapoff(struct thread *td, const char *name, enum uio_seg name_seg, 2504 u_int flags) 2505 { 2506 struct vnode *vp; 2507 struct nameidata nd; 2508 struct swdevt *sp; 2509 int error; 2510 2511 error = priv_check(td, PRIV_SWAPOFF); 2512 if (error != 0) 2513 return (error); 2514 if ((flags & ~(SWAPOFF_FORCE)) != 0) 2515 return (EINVAL); 2516 2517 sx_xlock(&swdev_syscall_lock); 2518 2519 NDINIT(&nd, LOOKUP, FOLLOW | AUDITVNODE1, name_seg, name); 2520 error = namei(&nd); 2521 if (error) 2522 goto done; 2523 NDFREE_PNBUF(&nd); 2524 vp = nd.ni_vp; 2525 2526 mtx_lock(&sw_dev_mtx); 2527 TAILQ_FOREACH(sp, &swtailq, sw_list) { 2528 if (sp->sw_vp == vp) 2529 break; 2530 } 2531 mtx_unlock(&sw_dev_mtx); 2532 if (sp == NULL) { 2533 error = EINVAL; 2534 goto done; 2535 } 2536 error = swapoff_one(sp, td->td_ucred, flags); 2537 done: 2538 sx_xunlock(&swdev_syscall_lock); 2539 return (error); 2540 } 2541 2542 2543 #ifdef COMPAT_FREEBSD13 2544 int 2545 freebsd13_swapoff(struct thread *td, struct freebsd13_swapoff_args *uap) 2546 { 2547 return (kern_swapoff(td, uap->name, UIO_USERSPACE, 0)); 2548 } 2549 #endif 2550 2551 int 2552 sys_swapoff(struct thread *td, struct swapoff_args *uap) 2553 { 2554 return (kern_swapoff(td, uap->name, UIO_USERSPACE, uap->flags)); 2555 } 2556 2557 static int 2558 swapoff_one(struct swdevt *sp, struct ucred *cred, u_int flags) 2559 { 2560 u_long nblks; 2561 #ifdef MAC 2562 int error; 2563 #endif 2564 2565 sx_assert(&swdev_syscall_lock, SA_XLOCKED); 2566 #ifdef MAC 2567 (void) vn_lock(sp->sw_vp, LK_EXCLUSIVE | LK_RETRY); 2568 error = mac_system_check_swapoff(cred, sp->sw_vp); 2569 (void) VOP_UNLOCK(sp->sw_vp); 2570 if (error != 0) 2571 return (error); 2572 #endif 2573 nblks = sp->sw_nblks; 2574 2575 /* 2576 * We can turn off this swap device safely only if the 2577 * available virtual memory in the system will fit the amount 2578 * of data we will have to page back in, plus an epsilon so 2579 * the system doesn't become critically low on swap space. 2580 * The vm_free_count() part does not account e.g. for clean 2581 * pages that can be immediately reclaimed without paging, so 2582 * this is a very rough estimation. 2583 * 2584 * On the other hand, not turning swap off on swapoff_all() 2585 * means that we can lose swap data when filesystems go away, 2586 * which is arguably worse. 2587 */ 2588 if ((flags & SWAPOFF_FORCE) == 0 && 2589 vm_free_count() + swap_pager_avail < nblks + nswap_lowat) 2590 return (ENOMEM); 2591 2592 /* 2593 * Prevent further allocations on this device. 2594 */ 2595 mtx_lock(&sw_dev_mtx); 2596 sp->sw_flags |= SW_CLOSING; 2597 swap_pager_avail -= blist_fill(sp->sw_blist, 0, nblks); 2598 swap_total -= nblks; 2599 mtx_unlock(&sw_dev_mtx); 2600 2601 /* 2602 * Page in the contents of the device and close it. 2603 */ 2604 swap_pager_swapoff(sp); 2605 2606 sp->sw_close(curthread, sp); 2607 mtx_lock(&sw_dev_mtx); 2608 sp->sw_id = NULL; 2609 TAILQ_REMOVE(&swtailq, sp, sw_list); 2610 nswapdev--; 2611 if (nswapdev == 0) { 2612 swap_pager_full = 2; 2613 swap_pager_almost_full = 1; 2614 } 2615 if (swdevhd == sp) 2616 swdevhd = NULL; 2617 mtx_unlock(&sw_dev_mtx); 2618 blist_destroy(sp->sw_blist); 2619 free(sp, M_VMPGDATA); 2620 return (0); 2621 } 2622 2623 void 2624 swapoff_all(void) 2625 { 2626 struct swdevt *sp, *spt; 2627 const char *devname; 2628 int error; 2629 2630 sx_xlock(&swdev_syscall_lock); 2631 2632 mtx_lock(&sw_dev_mtx); 2633 TAILQ_FOREACH_SAFE(sp, &swtailq, sw_list, spt) { 2634 mtx_unlock(&sw_dev_mtx); 2635 if (vn_isdisk(sp->sw_vp)) 2636 devname = devtoname(sp->sw_vp->v_rdev); 2637 else 2638 devname = "[file]"; 2639 error = swapoff_one(sp, thread0.td_ucred, SWAPOFF_FORCE); 2640 if (error != 0) { 2641 printf("Cannot remove swap device %s (error=%d), " 2642 "skipping.\n", devname, error); 2643 } else if (bootverbose) { 2644 printf("Swap device %s removed.\n", devname); 2645 } 2646 mtx_lock(&sw_dev_mtx); 2647 } 2648 mtx_unlock(&sw_dev_mtx); 2649 2650 sx_xunlock(&swdev_syscall_lock); 2651 } 2652 2653 void 2654 swap_pager_status(int *total, int *used) 2655 { 2656 2657 *total = swap_total; 2658 *used = swap_total - swap_pager_avail - 2659 nswapdev * howmany(BBSIZE, PAGE_SIZE); 2660 } 2661 2662 int 2663 swap_dev_info(int name, struct xswdev *xs, char *devname, size_t len) 2664 { 2665 struct swdevt *sp; 2666 const char *tmp_devname; 2667 int error, n; 2668 2669 n = 0; 2670 error = ENOENT; 2671 mtx_lock(&sw_dev_mtx); 2672 TAILQ_FOREACH(sp, &swtailq, sw_list) { 2673 if (n != name) { 2674 n++; 2675 continue; 2676 } 2677 xs->xsw_version = XSWDEV_VERSION; 2678 xs->xsw_dev = sp->sw_dev; 2679 xs->xsw_flags = sp->sw_flags; 2680 xs->xsw_nblks = sp->sw_nblks; 2681 xs->xsw_used = sp->sw_used; 2682 if (devname != NULL) { 2683 if (vn_isdisk(sp->sw_vp)) 2684 tmp_devname = devtoname(sp->sw_vp->v_rdev); 2685 else 2686 tmp_devname = "[file]"; 2687 strncpy(devname, tmp_devname, len); 2688 } 2689 error = 0; 2690 break; 2691 } 2692 mtx_unlock(&sw_dev_mtx); 2693 return (error); 2694 } 2695 2696 #if defined(COMPAT_FREEBSD11) 2697 #define XSWDEV_VERSION_11 1 2698 struct xswdev11 { 2699 u_int xsw_version; 2700 uint32_t xsw_dev; 2701 int xsw_flags; 2702 int xsw_nblks; 2703 int xsw_used; 2704 }; 2705 #endif 2706 2707 #if defined(__amd64__) && defined(COMPAT_FREEBSD32) 2708 struct xswdev32 { 2709 u_int xsw_version; 2710 u_int xsw_dev1, xsw_dev2; 2711 int xsw_flags; 2712 int xsw_nblks; 2713 int xsw_used; 2714 }; 2715 #endif 2716 2717 static int 2718 sysctl_vm_swap_info(SYSCTL_HANDLER_ARGS) 2719 { 2720 struct xswdev xs; 2721 #if defined(__amd64__) && defined(COMPAT_FREEBSD32) 2722 struct xswdev32 xs32; 2723 #endif 2724 #if defined(COMPAT_FREEBSD11) 2725 struct xswdev11 xs11; 2726 #endif 2727 int error; 2728 2729 if (arg2 != 1) /* name length */ 2730 return (EINVAL); 2731 2732 memset(&xs, 0, sizeof(xs)); 2733 error = swap_dev_info(*(int *)arg1, &xs, NULL, 0); 2734 if (error != 0) 2735 return (error); 2736 #if defined(__amd64__) && defined(COMPAT_FREEBSD32) 2737 if (req->oldlen == sizeof(xs32)) { 2738 memset(&xs32, 0, sizeof(xs32)); 2739 xs32.xsw_version = XSWDEV_VERSION; 2740 xs32.xsw_dev1 = xs.xsw_dev; 2741 xs32.xsw_dev2 = xs.xsw_dev >> 32; 2742 xs32.xsw_flags = xs.xsw_flags; 2743 xs32.xsw_nblks = xs.xsw_nblks; 2744 xs32.xsw_used = xs.xsw_used; 2745 error = SYSCTL_OUT(req, &xs32, sizeof(xs32)); 2746 return (error); 2747 } 2748 #endif 2749 #if defined(COMPAT_FREEBSD11) 2750 if (req->oldlen == sizeof(xs11)) { 2751 memset(&xs11, 0, sizeof(xs11)); 2752 xs11.xsw_version = XSWDEV_VERSION_11; 2753 xs11.xsw_dev = xs.xsw_dev; /* truncation */ 2754 xs11.xsw_flags = xs.xsw_flags; 2755 xs11.xsw_nblks = xs.xsw_nblks; 2756 xs11.xsw_used = xs.xsw_used; 2757 error = SYSCTL_OUT(req, &xs11, sizeof(xs11)); 2758 return (error); 2759 } 2760 #endif 2761 error = SYSCTL_OUT(req, &xs, sizeof(xs)); 2762 return (error); 2763 } 2764 2765 SYSCTL_INT(_vm, OID_AUTO, nswapdev, CTLFLAG_RD, &nswapdev, 0, 2766 "Number of swap devices"); 2767 SYSCTL_NODE(_vm, OID_AUTO, swap_info, CTLFLAG_RD | CTLFLAG_MPSAFE, 2768 sysctl_vm_swap_info, 2769 "Swap statistics by device"); 2770 2771 /* 2772 * Count the approximate swap usage in pages for a vmspace. The 2773 * shadowed or not yet copied on write swap blocks are not accounted. 2774 * The map must be locked. 2775 */ 2776 long 2777 vmspace_swap_count(struct vmspace *vmspace) 2778 { 2779 vm_map_t map; 2780 vm_map_entry_t cur; 2781 vm_object_t object; 2782 struct swblk *sb; 2783 vm_pindex_t e, pi; 2784 long count; 2785 int i; 2786 2787 map = &vmspace->vm_map; 2788 count = 0; 2789 2790 VM_MAP_ENTRY_FOREACH(cur, map) { 2791 if ((cur->eflags & MAP_ENTRY_IS_SUB_MAP) != 0) 2792 continue; 2793 object = cur->object.vm_object; 2794 if (object == NULL || (object->flags & OBJ_SWAP) == 0) 2795 continue; 2796 VM_OBJECT_RLOCK(object); 2797 if ((object->flags & OBJ_SWAP) == 0) 2798 goto unlock; 2799 pi = OFF_TO_IDX(cur->offset); 2800 e = pi + OFF_TO_IDX(cur->end - cur->start); 2801 for (;; pi = sb->p + SWAP_META_PAGES) { 2802 sb = SWAP_PCTRIE_LOOKUP_GE( 2803 &object->un_pager.swp.swp_blks, pi); 2804 if (sb == NULL || sb->p >= e) 2805 break; 2806 for (i = 0; i < SWAP_META_PAGES; i++) { 2807 if (sb->p + i < e && 2808 sb->d[i] != SWAPBLK_NONE) 2809 count++; 2810 } 2811 } 2812 unlock: 2813 VM_OBJECT_RUNLOCK(object); 2814 } 2815 return (count); 2816 } 2817 2818 /* 2819 * GEOM backend 2820 * 2821 * Swapping onto disk devices. 2822 * 2823 */ 2824 2825 static g_orphan_t swapgeom_orphan; 2826 2827 static struct g_class g_swap_class = { 2828 .name = "SWAP", 2829 .version = G_VERSION, 2830 .orphan = swapgeom_orphan, 2831 }; 2832 2833 DECLARE_GEOM_CLASS(g_swap_class, g_class); 2834 2835 static void 2836 swapgeom_close_ev(void *arg, int flags) 2837 { 2838 struct g_consumer *cp; 2839 2840 cp = arg; 2841 g_access(cp, -1, -1, 0); 2842 g_detach(cp); 2843 g_destroy_consumer(cp); 2844 } 2845 2846 /* 2847 * Add a reference to the g_consumer for an inflight transaction. 2848 */ 2849 static void 2850 swapgeom_acquire(struct g_consumer *cp) 2851 { 2852 2853 mtx_assert(&sw_dev_mtx, MA_OWNED); 2854 cp->index++; 2855 } 2856 2857 /* 2858 * Remove a reference from the g_consumer. Post a close event if all 2859 * references go away, since the function might be called from the 2860 * biodone context. 2861 */ 2862 static void 2863 swapgeom_release(struct g_consumer *cp, struct swdevt *sp) 2864 { 2865 2866 mtx_assert(&sw_dev_mtx, MA_OWNED); 2867 cp->index--; 2868 if (cp->index == 0) { 2869 if (g_post_event(swapgeom_close_ev, cp, M_NOWAIT, NULL) == 0) 2870 sp->sw_id = NULL; 2871 } 2872 } 2873 2874 static void 2875 swapgeom_done(struct bio *bp2) 2876 { 2877 struct swdevt *sp; 2878 struct buf *bp; 2879 struct g_consumer *cp; 2880 2881 bp = bp2->bio_caller2; 2882 cp = bp2->bio_from; 2883 bp->b_ioflags = bp2->bio_flags; 2884 if (bp2->bio_error) 2885 bp->b_ioflags |= BIO_ERROR; 2886 bp->b_resid = bp->b_bcount - bp2->bio_completed; 2887 bp->b_error = bp2->bio_error; 2888 bp->b_caller1 = NULL; 2889 bufdone(bp); 2890 sp = bp2->bio_caller1; 2891 mtx_lock(&sw_dev_mtx); 2892 swapgeom_release(cp, sp); 2893 mtx_unlock(&sw_dev_mtx); 2894 g_destroy_bio(bp2); 2895 } 2896 2897 static void 2898 swapgeom_strategy(struct buf *bp, struct swdevt *sp) 2899 { 2900 struct bio *bio; 2901 struct g_consumer *cp; 2902 2903 mtx_lock(&sw_dev_mtx); 2904 cp = sp->sw_id; 2905 if (cp == NULL) { 2906 mtx_unlock(&sw_dev_mtx); 2907 bp->b_error = ENXIO; 2908 bp->b_ioflags |= BIO_ERROR; 2909 bufdone(bp); 2910 return; 2911 } 2912 swapgeom_acquire(cp); 2913 mtx_unlock(&sw_dev_mtx); 2914 if (bp->b_iocmd == BIO_WRITE) 2915 bio = g_new_bio(); 2916 else 2917 bio = g_alloc_bio(); 2918 if (bio == NULL) { 2919 mtx_lock(&sw_dev_mtx); 2920 swapgeom_release(cp, sp); 2921 mtx_unlock(&sw_dev_mtx); 2922 bp->b_error = ENOMEM; 2923 bp->b_ioflags |= BIO_ERROR; 2924 printf("swap_pager: cannot allocate bio\n"); 2925 bufdone(bp); 2926 return; 2927 } 2928 2929 bp->b_caller1 = bio; 2930 bio->bio_caller1 = sp; 2931 bio->bio_caller2 = bp; 2932 bio->bio_cmd = bp->b_iocmd; 2933 bio->bio_offset = (bp->b_blkno - sp->sw_first) * PAGE_SIZE; 2934 bio->bio_length = bp->b_bcount; 2935 bio->bio_done = swapgeom_done; 2936 bio->bio_flags |= BIO_SWAP; 2937 if (!buf_mapped(bp)) { 2938 bio->bio_ma = bp->b_pages; 2939 bio->bio_data = unmapped_buf; 2940 bio->bio_ma_offset = (vm_offset_t)bp->b_offset & PAGE_MASK; 2941 bio->bio_ma_n = bp->b_npages; 2942 bio->bio_flags |= BIO_UNMAPPED; 2943 } else { 2944 bio->bio_data = bp->b_data; 2945 bio->bio_ma = NULL; 2946 } 2947 g_io_request(bio, cp); 2948 return; 2949 } 2950 2951 static void 2952 swapgeom_orphan(struct g_consumer *cp) 2953 { 2954 struct swdevt *sp; 2955 int destroy; 2956 2957 mtx_lock(&sw_dev_mtx); 2958 TAILQ_FOREACH(sp, &swtailq, sw_list) { 2959 if (sp->sw_id == cp) { 2960 sp->sw_flags |= SW_CLOSING; 2961 break; 2962 } 2963 } 2964 /* 2965 * Drop reference we were created with. Do directly since we're in a 2966 * special context where we don't have to queue the call to 2967 * swapgeom_close_ev(). 2968 */ 2969 cp->index--; 2970 destroy = ((sp != NULL) && (cp->index == 0)); 2971 if (destroy) 2972 sp->sw_id = NULL; 2973 mtx_unlock(&sw_dev_mtx); 2974 if (destroy) 2975 swapgeom_close_ev(cp, 0); 2976 } 2977 2978 static void 2979 swapgeom_close(struct thread *td, struct swdevt *sw) 2980 { 2981 struct g_consumer *cp; 2982 2983 mtx_lock(&sw_dev_mtx); 2984 cp = sw->sw_id; 2985 sw->sw_id = NULL; 2986 mtx_unlock(&sw_dev_mtx); 2987 2988 /* 2989 * swapgeom_close() may be called from the biodone context, 2990 * where we cannot perform topology changes. Delegate the 2991 * work to the events thread. 2992 */ 2993 if (cp != NULL) 2994 g_waitfor_event(swapgeom_close_ev, cp, M_WAITOK, NULL); 2995 } 2996 2997 static int 2998 swapongeom_locked(struct cdev *dev, struct vnode *vp) 2999 { 3000 struct g_provider *pp; 3001 struct g_consumer *cp; 3002 static struct g_geom *gp; 3003 struct swdevt *sp; 3004 u_long nblks; 3005 int error; 3006 3007 pp = g_dev_getprovider(dev); 3008 if (pp == NULL) 3009 return (ENODEV); 3010 mtx_lock(&sw_dev_mtx); 3011 TAILQ_FOREACH(sp, &swtailq, sw_list) { 3012 cp = sp->sw_id; 3013 if (cp != NULL && cp->provider == pp) { 3014 mtx_unlock(&sw_dev_mtx); 3015 return (EBUSY); 3016 } 3017 } 3018 mtx_unlock(&sw_dev_mtx); 3019 if (gp == NULL) 3020 gp = g_new_geomf(&g_swap_class, "swap"); 3021 cp = g_new_consumer(gp); 3022 cp->index = 1; /* Number of active I/Os, plus one for being active. */ 3023 cp->flags |= G_CF_DIRECT_SEND | G_CF_DIRECT_RECEIVE; 3024 g_attach(cp, pp); 3025 /* 3026 * XXX: Every time you think you can improve the margin for 3027 * footshooting, somebody depends on the ability to do so: 3028 * savecore(8) wants to write to our swapdev so we cannot 3029 * set an exclusive count :-( 3030 */ 3031 error = g_access(cp, 1, 1, 0); 3032 if (error != 0) { 3033 g_detach(cp); 3034 g_destroy_consumer(cp); 3035 return (error); 3036 } 3037 nblks = pp->mediasize / DEV_BSIZE; 3038 swaponsomething(vp, cp, nblks, swapgeom_strategy, 3039 swapgeom_close, dev2udev(dev), 3040 (pp->flags & G_PF_ACCEPT_UNMAPPED) != 0 ? SW_UNMAPPED : 0); 3041 return (0); 3042 } 3043 3044 static int 3045 swapongeom(struct vnode *vp) 3046 { 3047 int error; 3048 3049 ASSERT_VOP_ELOCKED(vp, "swapongeom"); 3050 if (vp->v_type != VCHR || VN_IS_DOOMED(vp)) { 3051 error = ENOENT; 3052 } else { 3053 g_topology_lock(); 3054 error = swapongeom_locked(vp->v_rdev, vp); 3055 g_topology_unlock(); 3056 } 3057 return (error); 3058 } 3059 3060 /* 3061 * VNODE backend 3062 * 3063 * This is used mainly for network filesystem (read: probably only tested 3064 * with NFS) swapfiles. 3065 * 3066 */ 3067 3068 static void 3069 swapdev_strategy(struct buf *bp, struct swdevt *sp) 3070 { 3071 struct vnode *vp2; 3072 3073 bp->b_blkno = ctodb(bp->b_blkno - sp->sw_first); 3074 3075 vp2 = sp->sw_id; 3076 vhold(vp2); 3077 if (bp->b_iocmd == BIO_WRITE) { 3078 vn_lock(vp2, LK_EXCLUSIVE | LK_RETRY); 3079 if (bp->b_bufobj) 3080 bufobj_wdrop(bp->b_bufobj); 3081 bufobj_wref(&vp2->v_bufobj); 3082 } else { 3083 vn_lock(vp2, LK_SHARED | LK_RETRY); 3084 } 3085 if (bp->b_bufobj != &vp2->v_bufobj) 3086 bp->b_bufobj = &vp2->v_bufobj; 3087 bp->b_vp = vp2; 3088 bp->b_iooffset = dbtob(bp->b_blkno); 3089 bstrategy(bp); 3090 VOP_UNLOCK(vp2); 3091 } 3092 3093 static void 3094 swapdev_close(struct thread *td, struct swdevt *sp) 3095 { 3096 struct vnode *vp; 3097 3098 vp = sp->sw_vp; 3099 vn_lock(vp, LK_EXCLUSIVE | LK_RETRY); 3100 VOP_CLOSE(vp, FREAD | FWRITE, td->td_ucred, td); 3101 vput(vp); 3102 } 3103 3104 static int 3105 swaponvp(struct thread *td, struct vnode *vp, u_long nblks) 3106 { 3107 struct swdevt *sp; 3108 int error; 3109 3110 ASSERT_VOP_ELOCKED(vp, "swaponvp"); 3111 if (nblks == 0) 3112 return (ENXIO); 3113 mtx_lock(&sw_dev_mtx); 3114 TAILQ_FOREACH(sp, &swtailq, sw_list) { 3115 if (sp->sw_id == vp) { 3116 mtx_unlock(&sw_dev_mtx); 3117 return (EBUSY); 3118 } 3119 } 3120 mtx_unlock(&sw_dev_mtx); 3121 3122 #ifdef MAC 3123 error = mac_system_check_swapon(td->td_ucred, vp); 3124 if (error == 0) 3125 #endif 3126 error = VOP_OPEN(vp, FREAD | FWRITE, td->td_ucred, td, NULL); 3127 if (error != 0) 3128 return (error); 3129 3130 swaponsomething(vp, vp, nblks, swapdev_strategy, swapdev_close, 3131 NODEV, 0); 3132 return (0); 3133 } 3134 3135 static int 3136 sysctl_swap_async_max(SYSCTL_HANDLER_ARGS) 3137 { 3138 int error, new, n; 3139 3140 new = nsw_wcount_async_max; 3141 error = sysctl_handle_int(oidp, &new, 0, req); 3142 if (error != 0 || req->newptr == NULL) 3143 return (error); 3144 3145 if (new > nswbuf / 2 || new < 1) 3146 return (EINVAL); 3147 3148 mtx_lock(&swbuf_mtx); 3149 while (nsw_wcount_async_max != new) { 3150 /* 3151 * Adjust difference. If the current async count is too low, 3152 * we will need to sqeeze our update slowly in. Sleep with a 3153 * higher priority than getpbuf() to finish faster. 3154 */ 3155 n = new - nsw_wcount_async_max; 3156 if (nsw_wcount_async + n >= 0) { 3157 nsw_wcount_async += n; 3158 nsw_wcount_async_max += n; 3159 wakeup(&nsw_wcount_async); 3160 } else { 3161 nsw_wcount_async_max -= nsw_wcount_async; 3162 nsw_wcount_async = 0; 3163 msleep(&nsw_wcount_async, &swbuf_mtx, PSWP, 3164 "swpsysctl", 0); 3165 } 3166 } 3167 mtx_unlock(&swbuf_mtx); 3168 3169 return (0); 3170 } 3171 3172 static void 3173 swap_pager_update_writecount(vm_object_t object, vm_offset_t start, 3174 vm_offset_t end) 3175 { 3176 3177 VM_OBJECT_WLOCK(object); 3178 KASSERT((object->flags & OBJ_ANON) == 0, 3179 ("Splittable object with writecount")); 3180 object->un_pager.swp.writemappings += (vm_ooffset_t)end - start; 3181 VM_OBJECT_WUNLOCK(object); 3182 } 3183 3184 static void 3185 swap_pager_release_writecount(vm_object_t object, vm_offset_t start, 3186 vm_offset_t end) 3187 { 3188 3189 VM_OBJECT_WLOCK(object); 3190 KASSERT((object->flags & OBJ_ANON) == 0, 3191 ("Splittable object with writecount")); 3192 KASSERT(object->un_pager.swp.writemappings >= (vm_ooffset_t)end - start, 3193 ("swap obj %p writecount %jx dec %jx", object, 3194 (uintmax_t)object->un_pager.swp.writemappings, 3195 (uintmax_t)((vm_ooffset_t)end - start))); 3196 object->un_pager.swp.writemappings -= (vm_ooffset_t)end - start; 3197 VM_OBJECT_WUNLOCK(object); 3198 } 3199