1 /*- 2 * SPDX-License-Identifier: BSD-4-Clause 3 * 4 * Copyright (c) 1998 Matthew Dillon, 5 * Copyright (c) 1994 John S. Dyson 6 * Copyright (c) 1990 University of Utah. 7 * Copyright (c) 1982, 1986, 1989, 1993 8 * The Regents of the University of California. All rights reserved. 9 * 10 * This code is derived from software contributed to Berkeley by 11 * the Systems Programming Group of the University of Utah Computer 12 * Science Department. 13 * 14 * Redistribution and use in source and binary forms, with or without 15 * modification, are permitted provided that the following conditions 16 * are met: 17 * 1. Redistributions of source code must retain the above copyright 18 * notice, this list of conditions and the following disclaimer. 19 * 2. Redistributions in binary form must reproduce the above copyright 20 * notice, this list of conditions and the following disclaimer in the 21 * documentation and/or other materials provided with the distribution. 22 * 3. All advertising materials mentioning features or use of this software 23 * must display the following acknowledgement: 24 * This product includes software developed by the University of 25 * California, Berkeley and its contributors. 26 * 4. Neither the name of the University nor the names of its contributors 27 * may be used to endorse or promote products derived from this software 28 * without specific prior written permission. 29 * 30 * THIS SOFTWARE IS PROVIDED BY THE REGENTS AND CONTRIBUTORS ``AS IS'' AND 31 * ANY EXPRESS OR IMPLIED WARRANTIES, INCLUDING, BUT NOT LIMITED TO, THE 32 * IMPLIED WARRANTIES OF MERCHANTABILITY AND FITNESS FOR A PARTICULAR PURPOSE 33 * ARE DISCLAIMED. IN NO EVENT SHALL THE REGENTS OR CONTRIBUTORS BE LIABLE 34 * FOR ANY DIRECT, INDIRECT, INCIDENTAL, SPECIAL, EXEMPLARY, OR CONSEQUENTIAL 35 * DAMAGES (INCLUDING, BUT NOT LIMITED TO, PROCUREMENT OF SUBSTITUTE GOODS 36 * OR SERVICES; LOSS OF USE, DATA, OR PROFITS; OR BUSINESS INTERRUPTION) 37 * HOWEVER CAUSED AND ON ANY THEORY OF LIABILITY, WHETHER IN CONTRACT, STRICT 38 * LIABILITY, OR TORT (INCLUDING NEGLIGENCE OR OTHERWISE) ARISING IN ANY WAY 39 * OUT OF THE USE OF THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF 40 * SUCH DAMAGE. 41 * 42 * New Swap System 43 * Matthew Dillon 44 * 45 * Radix Bitmap 'blists'. 46 * 47 * - The new swapper uses the new radix bitmap code. This should scale 48 * to arbitrarily small or arbitrarily large swap spaces and an almost 49 * arbitrary degree of fragmentation. 50 * 51 * Features: 52 * 53 * - on the fly reallocation of swap during putpages. The new system 54 * does not try to keep previously allocated swap blocks for dirty 55 * pages. 56 * 57 * - on the fly deallocation of swap 58 * 59 * - No more garbage collection required. Unnecessarily allocated swap 60 * blocks only exist for dirty vm_page_t's now and these are already 61 * cycled (in a high-load system) by the pager. We also do on-the-fly 62 * removal of invalidated swap blocks when a page is destroyed 63 * or renamed. 64 * 65 * from: Utah $Hdr: swap_pager.c 1.4 91/04/30$ 66 * 67 * @(#)swap_pager.c 8.9 (Berkeley) 3/21/94 68 * @(#)vm_swap.c 8.5 (Berkeley) 2/17/94 69 */ 70 71 #include <sys/cdefs.h> 72 __FBSDID("$FreeBSD$"); 73 74 #include "opt_vm.h" 75 76 #include <sys/param.h> 77 #include <sys/bio.h> 78 #include <sys/blist.h> 79 #include <sys/buf.h> 80 #include <sys/conf.h> 81 #include <sys/disk.h> 82 #include <sys/disklabel.h> 83 #include <sys/eventhandler.h> 84 #include <sys/fcntl.h> 85 #include <sys/lock.h> 86 #include <sys/kernel.h> 87 #include <sys/mount.h> 88 #include <sys/namei.h> 89 #include <sys/malloc.h> 90 #include <sys/pctrie.h> 91 #include <sys/priv.h> 92 #include <sys/proc.h> 93 #include <sys/racct.h> 94 #include <sys/resource.h> 95 #include <sys/resourcevar.h> 96 #include <sys/rwlock.h> 97 #include <sys/sbuf.h> 98 #include <sys/sysctl.h> 99 #include <sys/sysproto.h> 100 #include <sys/systm.h> 101 #include <sys/sx.h> 102 #include <sys/vmmeter.h> 103 #include <sys/vnode.h> 104 105 #include <security/mac/mac_framework.h> 106 107 #include <vm/vm.h> 108 #include <vm/pmap.h> 109 #include <vm/vm_map.h> 110 #include <vm/vm_kern.h> 111 #include <vm/vm_object.h> 112 #include <vm/vm_page.h> 113 #include <vm/vm_pager.h> 114 #include <vm/vm_pageout.h> 115 #include <vm/vm_param.h> 116 #include <vm/swap_pager.h> 117 #include <vm/vm_extern.h> 118 #include <vm/uma.h> 119 120 #include <geom/geom.h> 121 122 /* 123 * MAX_PAGEOUT_CLUSTER must be a power of 2 between 1 and 64. 124 * The 64-page limit is due to the radix code (kern/subr_blist.c). 125 */ 126 #ifndef MAX_PAGEOUT_CLUSTER 127 #define MAX_PAGEOUT_CLUSTER 32 128 #endif 129 130 #if !defined(SWB_NPAGES) 131 #define SWB_NPAGES MAX_PAGEOUT_CLUSTER 132 #endif 133 134 #define SWAP_META_PAGES PCTRIE_COUNT 135 136 /* 137 * A swblk structure maps each page index within a 138 * SWAP_META_PAGES-aligned and sized range to the address of an 139 * on-disk swap block (or SWAPBLK_NONE). The collection of these 140 * mappings for an entire vm object is implemented as a pc-trie. 141 */ 142 struct swblk { 143 vm_pindex_t p; 144 daddr_t d[SWAP_META_PAGES]; 145 }; 146 147 static MALLOC_DEFINE(M_VMPGDATA, "vm_pgdata", "swap pager private data"); 148 static struct mtx sw_dev_mtx; 149 static TAILQ_HEAD(, swdevt) swtailq = TAILQ_HEAD_INITIALIZER(swtailq); 150 static struct swdevt *swdevhd; /* Allocate from here next */ 151 static int nswapdev; /* Number of swap devices */ 152 int swap_pager_avail; 153 static struct sx swdev_syscall_lock; /* serialize swap(on|off) */ 154 155 static u_long swap_reserved; 156 static u_long swap_total; 157 static int sysctl_page_shift(SYSCTL_HANDLER_ARGS); 158 159 static SYSCTL_NODE(_vm_stats, OID_AUTO, swap, CTLFLAG_RD, 0, "VM swap stats"); 160 161 SYSCTL_PROC(_vm, OID_AUTO, swap_reserved, CTLTYPE_U64 | CTLFLAG_RD | CTLFLAG_MPSAFE, 162 &swap_reserved, 0, sysctl_page_shift, "A", 163 "Amount of swap storage needed to back all allocated anonymous memory."); 164 SYSCTL_PROC(_vm, OID_AUTO, swap_total, CTLTYPE_U64 | CTLFLAG_RD | CTLFLAG_MPSAFE, 165 &swap_total, 0, sysctl_page_shift, "A", 166 "Total amount of available swap storage."); 167 168 static int overcommit = 0; 169 SYSCTL_INT(_vm, VM_OVERCOMMIT, overcommit, CTLFLAG_RW, &overcommit, 0, 170 "Configure virtual memory overcommit behavior. See tuning(7) " 171 "for details."); 172 static unsigned long swzone; 173 SYSCTL_ULONG(_vm, OID_AUTO, swzone, CTLFLAG_RD, &swzone, 0, 174 "Actual size of swap metadata zone"); 175 static unsigned long swap_maxpages; 176 SYSCTL_ULONG(_vm, OID_AUTO, swap_maxpages, CTLFLAG_RD, &swap_maxpages, 0, 177 "Maximum amount of swap supported"); 178 179 static counter_u64_t swap_free_deferred; 180 SYSCTL_COUNTER_U64(_vm_stats_swap, OID_AUTO, free_deferred, 181 CTLFLAG_RD, &swap_free_deferred, 182 "Number of pages that deferred freeing swap space"); 183 184 static counter_u64_t swap_free_completed; 185 SYSCTL_COUNTER_U64(_vm_stats_swap, OID_AUTO, free_completed, 186 CTLFLAG_RD, &swap_free_completed, 187 "Number of deferred frees completed"); 188 189 /* bits from overcommit */ 190 #define SWAP_RESERVE_FORCE_ON (1 << 0) 191 #define SWAP_RESERVE_RLIMIT_ON (1 << 1) 192 #define SWAP_RESERVE_ALLOW_NONWIRED (1 << 2) 193 194 static int 195 sysctl_page_shift(SYSCTL_HANDLER_ARGS) 196 { 197 uint64_t newval; 198 u_long value = *(u_long *)arg1; 199 200 newval = ((uint64_t)value) << PAGE_SHIFT; 201 return (sysctl_handle_64(oidp, &newval, 0, req)); 202 } 203 204 int 205 swap_reserve(vm_ooffset_t incr) 206 { 207 208 return (swap_reserve_by_cred(incr, curthread->td_ucred)); 209 } 210 211 int 212 swap_reserve_by_cred(vm_ooffset_t incr, struct ucred *cred) 213 { 214 u_long r, s, prev, pincr; 215 int res, error; 216 static int curfail; 217 static struct timeval lastfail; 218 struct uidinfo *uip; 219 220 uip = cred->cr_ruidinfo; 221 222 KASSERT((incr & PAGE_MASK) == 0, ("%s: incr: %ju & PAGE_MASK", __func__, 223 (uintmax_t)incr)); 224 225 #ifdef RACCT 226 if (racct_enable) { 227 PROC_LOCK(curproc); 228 error = racct_add(curproc, RACCT_SWAP, incr); 229 PROC_UNLOCK(curproc); 230 if (error != 0) 231 return (0); 232 } 233 #endif 234 235 pincr = atop(incr); 236 res = 0; 237 prev = atomic_fetchadd_long(&swap_reserved, pincr); 238 r = prev + pincr; 239 if (overcommit & SWAP_RESERVE_ALLOW_NONWIRED) { 240 s = vm_cnt.v_page_count - vm_cnt.v_free_reserved - 241 vm_wire_count(); 242 } else 243 s = 0; 244 s += swap_total; 245 if ((overcommit & SWAP_RESERVE_FORCE_ON) == 0 || r <= s || 246 (error = priv_check(curthread, PRIV_VM_SWAP_NOQUOTA)) == 0) { 247 res = 1; 248 } else { 249 prev = atomic_fetchadd_long(&swap_reserved, -pincr); 250 if (prev < pincr) 251 panic("swap_reserved < incr on overcommit fail"); 252 } 253 if (res) { 254 prev = atomic_fetchadd_long(&uip->ui_vmsize, pincr); 255 if ((overcommit & SWAP_RESERVE_RLIMIT_ON) != 0 && 256 prev + pincr > lim_cur(curthread, RLIMIT_SWAP) && 257 priv_check(curthread, PRIV_VM_SWAP_NORLIMIT)) { 258 res = 0; 259 prev = atomic_fetchadd_long(&uip->ui_vmsize, -pincr); 260 if (prev < pincr) 261 panic("uip->ui_vmsize < incr on overcommit fail"); 262 } 263 } 264 if (!res && ppsratecheck(&lastfail, &curfail, 1)) { 265 printf("uid %d, pid %d: swap reservation for %jd bytes failed\n", 266 uip->ui_uid, curproc->p_pid, incr); 267 } 268 269 #ifdef RACCT 270 if (racct_enable && !res) { 271 PROC_LOCK(curproc); 272 racct_sub(curproc, RACCT_SWAP, incr); 273 PROC_UNLOCK(curproc); 274 } 275 #endif 276 277 return (res); 278 } 279 280 void 281 swap_reserve_force(vm_ooffset_t incr) 282 { 283 struct uidinfo *uip; 284 u_long pincr; 285 286 KASSERT((incr & PAGE_MASK) == 0, ("%s: incr: %ju & PAGE_MASK", __func__, 287 (uintmax_t)incr)); 288 289 PROC_LOCK(curproc); 290 #ifdef RACCT 291 if (racct_enable) 292 racct_add_force(curproc, RACCT_SWAP, incr); 293 #endif 294 pincr = atop(incr); 295 atomic_add_long(&swap_reserved, pincr); 296 uip = curproc->p_ucred->cr_ruidinfo; 297 atomic_add_long(&uip->ui_vmsize, pincr); 298 PROC_UNLOCK(curproc); 299 } 300 301 void 302 swap_release(vm_ooffset_t decr) 303 { 304 struct ucred *cred; 305 306 PROC_LOCK(curproc); 307 cred = curproc->p_ucred; 308 swap_release_by_cred(decr, cred); 309 PROC_UNLOCK(curproc); 310 } 311 312 void 313 swap_release_by_cred(vm_ooffset_t decr, struct ucred *cred) 314 { 315 u_long prev, pdecr; 316 struct uidinfo *uip; 317 318 uip = cred->cr_ruidinfo; 319 320 KASSERT((decr & PAGE_MASK) == 0, ("%s: decr: %ju & PAGE_MASK", __func__, 321 (uintmax_t)decr)); 322 323 pdecr = atop(decr); 324 prev = atomic_fetchadd_long(&swap_reserved, -pdecr); 325 if (prev < pdecr) 326 panic("swap_reserved < decr"); 327 328 prev = atomic_fetchadd_long(&uip->ui_vmsize, -pdecr); 329 if (prev < pdecr) 330 printf("negative vmsize for uid = %d\n", uip->ui_uid); 331 #ifdef RACCT 332 if (racct_enable) 333 racct_sub_cred(cred, RACCT_SWAP, decr); 334 #endif 335 } 336 337 static int swap_pager_full = 2; /* swap space exhaustion (task killing) */ 338 static int swap_pager_almost_full = 1; /* swap space exhaustion (w/hysteresis)*/ 339 static struct mtx swbuf_mtx; /* to sync nsw_wcount_async */ 340 static int nsw_wcount_async; /* limit async write buffers */ 341 static int nsw_wcount_async_max;/* assigned maximum */ 342 static int nsw_cluster_max; /* maximum VOP I/O allowed */ 343 344 static int sysctl_swap_async_max(SYSCTL_HANDLER_ARGS); 345 SYSCTL_PROC(_vm, OID_AUTO, swap_async_max, CTLTYPE_INT | CTLFLAG_RW | 346 CTLFLAG_MPSAFE, NULL, 0, sysctl_swap_async_max, "I", 347 "Maximum running async swap ops"); 348 static int sysctl_swap_fragmentation(SYSCTL_HANDLER_ARGS); 349 SYSCTL_PROC(_vm, OID_AUTO, swap_fragmentation, CTLTYPE_STRING | CTLFLAG_RD | 350 CTLFLAG_MPSAFE, NULL, 0, sysctl_swap_fragmentation, "A", 351 "Swap Fragmentation Info"); 352 353 static struct sx sw_alloc_sx; 354 355 /* 356 * "named" and "unnamed" anon region objects. Try to reduce the overhead 357 * of searching a named list by hashing it just a little. 358 */ 359 360 #define NOBJLISTS 8 361 362 #define NOBJLIST(handle) \ 363 (&swap_pager_object_list[((int)(intptr_t)handle >> 4) & (NOBJLISTS-1)]) 364 365 static struct pagerlst swap_pager_object_list[NOBJLISTS]; 366 static uma_zone_t swwbuf_zone; 367 static uma_zone_t swrbuf_zone; 368 static uma_zone_t swblk_zone; 369 static uma_zone_t swpctrie_zone; 370 371 /* 372 * pagerops for OBJT_SWAP - "swap pager". Some ops are also global procedure 373 * calls hooked from other parts of the VM system and do not appear here. 374 * (see vm/swap_pager.h). 375 */ 376 static vm_object_t 377 swap_pager_alloc(void *handle, vm_ooffset_t size, 378 vm_prot_t prot, vm_ooffset_t offset, struct ucred *); 379 static void swap_pager_dealloc(vm_object_t object); 380 static int swap_pager_getpages(vm_object_t, vm_page_t *, int, int *, 381 int *); 382 static int swap_pager_getpages_async(vm_object_t, vm_page_t *, int, int *, 383 int *, pgo_getpages_iodone_t, void *); 384 static void swap_pager_putpages(vm_object_t, vm_page_t *, int, boolean_t, int *); 385 static boolean_t 386 swap_pager_haspage(vm_object_t object, vm_pindex_t pindex, int *before, int *after); 387 static void swap_pager_init(void); 388 static void swap_pager_unswapped(vm_page_t); 389 static void swap_pager_swapoff(struct swdevt *sp); 390 static void swap_pager_update_writecount(vm_object_t object, 391 vm_offset_t start, vm_offset_t end); 392 static void swap_pager_release_writecount(vm_object_t object, 393 vm_offset_t start, vm_offset_t end); 394 395 struct pagerops swappagerops = { 396 .pgo_init = swap_pager_init, /* early system initialization of pager */ 397 .pgo_alloc = swap_pager_alloc, /* allocate an OBJT_SWAP object */ 398 .pgo_dealloc = swap_pager_dealloc, /* deallocate an OBJT_SWAP object */ 399 .pgo_getpages = swap_pager_getpages, /* pagein */ 400 .pgo_getpages_async = swap_pager_getpages_async, /* pagein (async) */ 401 .pgo_putpages = swap_pager_putpages, /* pageout */ 402 .pgo_haspage = swap_pager_haspage, /* get backing store status for page */ 403 .pgo_pageunswapped = swap_pager_unswapped, /* remove swap related to page */ 404 .pgo_update_writecount = swap_pager_update_writecount, 405 .pgo_release_writecount = swap_pager_release_writecount, 406 }; 407 408 /* 409 * swap_*() routines are externally accessible. swp_*() routines are 410 * internal. 411 */ 412 static int nswap_lowat = 128; /* in pages, swap_pager_almost_full warn */ 413 static int nswap_hiwat = 512; /* in pages, swap_pager_almost_full warn */ 414 415 SYSCTL_INT(_vm, OID_AUTO, dmmax, CTLFLAG_RD, &nsw_cluster_max, 0, 416 "Maximum size of a swap block in pages"); 417 418 static void swp_sizecheck(void); 419 static void swp_pager_async_iodone(struct buf *bp); 420 static bool swp_pager_swblk_empty(struct swblk *sb, int start, int limit); 421 static void swp_pager_free_empty_swblk(vm_object_t, struct swblk *sb); 422 static int swapongeom(struct vnode *); 423 static int swaponvp(struct thread *, struct vnode *, u_long); 424 static int swapoff_one(struct swdevt *sp, struct ucred *cred); 425 426 /* 427 * Swap bitmap functions 428 */ 429 static void swp_pager_freeswapspace(daddr_t blk, daddr_t npages); 430 static daddr_t swp_pager_getswapspace(int *npages, int limit); 431 432 /* 433 * Metadata functions 434 */ 435 static daddr_t swp_pager_meta_build(vm_object_t, vm_pindex_t, daddr_t); 436 static void swp_pager_meta_free(vm_object_t, vm_pindex_t, vm_pindex_t); 437 static void swp_pager_meta_transfer(vm_object_t src, vm_object_t dst, 438 vm_pindex_t pindex, vm_pindex_t count); 439 static void swp_pager_meta_free_all(vm_object_t); 440 static daddr_t swp_pager_meta_lookup(vm_object_t, vm_pindex_t); 441 442 static void 443 swp_pager_init_freerange(daddr_t *start, daddr_t *num) 444 { 445 446 *start = SWAPBLK_NONE; 447 *num = 0; 448 } 449 450 static void 451 swp_pager_update_freerange(daddr_t *start, daddr_t *num, daddr_t addr) 452 { 453 454 if (*start + *num == addr) { 455 (*num)++; 456 } else { 457 swp_pager_freeswapspace(*start, *num); 458 *start = addr; 459 *num = 1; 460 } 461 } 462 463 static void * 464 swblk_trie_alloc(struct pctrie *ptree) 465 { 466 467 return (uma_zalloc(swpctrie_zone, M_NOWAIT | (curproc == pageproc ? 468 M_USE_RESERVE : 0))); 469 } 470 471 static void 472 swblk_trie_free(struct pctrie *ptree, void *node) 473 { 474 475 uma_zfree(swpctrie_zone, node); 476 } 477 478 PCTRIE_DEFINE(SWAP, swblk, p, swblk_trie_alloc, swblk_trie_free); 479 480 /* 481 * SWP_SIZECHECK() - update swap_pager_full indication 482 * 483 * update the swap_pager_almost_full indication and warn when we are 484 * about to run out of swap space, using lowat/hiwat hysteresis. 485 * 486 * Clear swap_pager_full ( task killing ) indication when lowat is met. 487 * 488 * No restrictions on call 489 * This routine may not block. 490 */ 491 static void 492 swp_sizecheck(void) 493 { 494 495 if (swap_pager_avail < nswap_lowat) { 496 if (swap_pager_almost_full == 0) { 497 printf("swap_pager: out of swap space\n"); 498 swap_pager_almost_full = 1; 499 } 500 } else { 501 swap_pager_full = 0; 502 if (swap_pager_avail > nswap_hiwat) 503 swap_pager_almost_full = 0; 504 } 505 } 506 507 /* 508 * SWAP_PAGER_INIT() - initialize the swap pager! 509 * 510 * Expected to be started from system init. NOTE: This code is run 511 * before much else so be careful what you depend on. Most of the VM 512 * system has yet to be initialized at this point. 513 */ 514 static void 515 swap_pager_init(void) 516 { 517 /* 518 * Initialize object lists 519 */ 520 int i; 521 522 for (i = 0; i < NOBJLISTS; ++i) 523 TAILQ_INIT(&swap_pager_object_list[i]); 524 mtx_init(&sw_dev_mtx, "swapdev", NULL, MTX_DEF); 525 sx_init(&sw_alloc_sx, "swspsx"); 526 sx_init(&swdev_syscall_lock, "swsysc"); 527 } 528 529 static void 530 swap_pager_counters(void) 531 { 532 533 swap_free_deferred = counter_u64_alloc(M_WAITOK); 534 swap_free_completed = counter_u64_alloc(M_WAITOK); 535 } 536 SYSINIT(swap_counters, SI_SUB_CPU, SI_ORDER_ANY, swap_pager_counters, NULL); 537 538 /* 539 * SWAP_PAGER_SWAP_INIT() - swap pager initialization from pageout process 540 * 541 * Expected to be started from pageout process once, prior to entering 542 * its main loop. 543 */ 544 void 545 swap_pager_swap_init(void) 546 { 547 unsigned long n, n2; 548 549 /* 550 * Number of in-transit swap bp operations. Don't 551 * exhaust the pbufs completely. Make sure we 552 * initialize workable values (0 will work for hysteresis 553 * but it isn't very efficient). 554 * 555 * The nsw_cluster_max is constrained by the bp->b_pages[] 556 * array, which has MAXPHYS / PAGE_SIZE entries, and our locally 557 * defined MAX_PAGEOUT_CLUSTER. Also be aware that swap ops are 558 * constrained by the swap device interleave stripe size. 559 * 560 * Currently we hardwire nsw_wcount_async to 4. This limit is 561 * designed to prevent other I/O from having high latencies due to 562 * our pageout I/O. The value 4 works well for one or two active swap 563 * devices but is probably a little low if you have more. Even so, 564 * a higher value would probably generate only a limited improvement 565 * with three or four active swap devices since the system does not 566 * typically have to pageout at extreme bandwidths. We will want 567 * at least 2 per swap devices, and 4 is a pretty good value if you 568 * have one NFS swap device due to the command/ack latency over NFS. 569 * So it all works out pretty well. 570 */ 571 nsw_cluster_max = min(MAXPHYS / PAGE_SIZE, MAX_PAGEOUT_CLUSTER); 572 573 nsw_wcount_async = 4; 574 nsw_wcount_async_max = nsw_wcount_async; 575 mtx_init(&swbuf_mtx, "async swbuf mutex", NULL, MTX_DEF); 576 577 swwbuf_zone = pbuf_zsecond_create("swwbuf", nswbuf / 4); 578 swrbuf_zone = pbuf_zsecond_create("swrbuf", nswbuf / 2); 579 580 /* 581 * Initialize our zone, taking the user's requested size or 582 * estimating the number we need based on the number of pages 583 * in the system. 584 */ 585 n = maxswzone != 0 ? maxswzone / sizeof(struct swblk) : 586 vm_cnt.v_page_count / 2; 587 swpctrie_zone = uma_zcreate("swpctrie", pctrie_node_size(), NULL, NULL, 588 pctrie_zone_init, NULL, UMA_ALIGN_PTR, UMA_ZONE_VM); 589 if (swpctrie_zone == NULL) 590 panic("failed to create swap pctrie zone."); 591 swblk_zone = uma_zcreate("swblk", sizeof(struct swblk), NULL, NULL, 592 NULL, NULL, _Alignof(struct swblk) - 1, UMA_ZONE_VM); 593 if (swblk_zone == NULL) 594 panic("failed to create swap blk zone."); 595 n2 = n; 596 do { 597 if (uma_zone_reserve_kva(swblk_zone, n)) 598 break; 599 /* 600 * if the allocation failed, try a zone two thirds the 601 * size of the previous attempt. 602 */ 603 n -= ((n + 2) / 3); 604 } while (n > 0); 605 606 /* 607 * Often uma_zone_reserve_kva() cannot reserve exactly the 608 * requested size. Account for the difference when 609 * calculating swap_maxpages. 610 */ 611 n = uma_zone_get_max(swblk_zone); 612 613 if (n < n2) 614 printf("Swap blk zone entries changed from %lu to %lu.\n", 615 n2, n); 616 /* absolute maximum we can handle assuming 100% efficiency */ 617 swap_maxpages = n * SWAP_META_PAGES; 618 swzone = n * sizeof(struct swblk); 619 if (!uma_zone_reserve_kva(swpctrie_zone, n)) 620 printf("Cannot reserve swap pctrie zone, " 621 "reduce kern.maxswzone.\n"); 622 } 623 624 static vm_object_t 625 swap_pager_alloc_init(void *handle, struct ucred *cred, vm_ooffset_t size, 626 vm_ooffset_t offset) 627 { 628 vm_object_t object; 629 630 if (cred != NULL) { 631 if (!swap_reserve_by_cred(size, cred)) 632 return (NULL); 633 crhold(cred); 634 } 635 636 /* 637 * The un_pager.swp.swp_blks trie is initialized by 638 * vm_object_allocate() to ensure the correct order of 639 * visibility to other threads. 640 */ 641 object = vm_object_allocate(OBJT_SWAP, OFF_TO_IDX(offset + 642 PAGE_MASK + size)); 643 644 object->un_pager.swp.writemappings = 0; 645 object->handle = handle; 646 if (cred != NULL) { 647 object->cred = cred; 648 object->charge = size; 649 } 650 return (object); 651 } 652 653 /* 654 * SWAP_PAGER_ALLOC() - allocate a new OBJT_SWAP VM object and instantiate 655 * its metadata structures. 656 * 657 * This routine is called from the mmap and fork code to create a new 658 * OBJT_SWAP object. 659 * 660 * This routine must ensure that no live duplicate is created for 661 * the named object request, which is protected against by 662 * holding the sw_alloc_sx lock in case handle != NULL. 663 */ 664 static vm_object_t 665 swap_pager_alloc(void *handle, vm_ooffset_t size, vm_prot_t prot, 666 vm_ooffset_t offset, struct ucred *cred) 667 { 668 vm_object_t object; 669 670 if (handle != NULL) { 671 /* 672 * Reference existing named region or allocate new one. There 673 * should not be a race here against swp_pager_meta_build() 674 * as called from vm_page_remove() in regards to the lookup 675 * of the handle. 676 */ 677 sx_xlock(&sw_alloc_sx); 678 object = vm_pager_object_lookup(NOBJLIST(handle), handle); 679 if (object == NULL) { 680 object = swap_pager_alloc_init(handle, cred, size, 681 offset); 682 if (object != NULL) { 683 TAILQ_INSERT_TAIL(NOBJLIST(object->handle), 684 object, pager_object_list); 685 } 686 } 687 sx_xunlock(&sw_alloc_sx); 688 } else { 689 object = swap_pager_alloc_init(handle, cred, size, offset); 690 } 691 return (object); 692 } 693 694 /* 695 * SWAP_PAGER_DEALLOC() - remove swap metadata from object 696 * 697 * The swap backing for the object is destroyed. The code is 698 * designed such that we can reinstantiate it later, but this 699 * routine is typically called only when the entire object is 700 * about to be destroyed. 701 * 702 * The object must be locked. 703 */ 704 static void 705 swap_pager_dealloc(vm_object_t object) 706 { 707 708 VM_OBJECT_ASSERT_WLOCKED(object); 709 KASSERT((object->flags & OBJ_DEAD) != 0, ("dealloc of reachable obj")); 710 711 /* 712 * Remove from list right away so lookups will fail if we block for 713 * pageout completion. 714 */ 715 if ((object->flags & OBJ_ANON) == 0 && object->handle != NULL) { 716 VM_OBJECT_WUNLOCK(object); 717 sx_xlock(&sw_alloc_sx); 718 TAILQ_REMOVE(NOBJLIST(object->handle), object, 719 pager_object_list); 720 sx_xunlock(&sw_alloc_sx); 721 VM_OBJECT_WLOCK(object); 722 } 723 724 vm_object_pip_wait(object, "swpdea"); 725 726 /* 727 * Free all remaining metadata. We only bother to free it from 728 * the swap meta data. We do not attempt to free swapblk's still 729 * associated with vm_page_t's for this object. We do not care 730 * if paging is still in progress on some objects. 731 */ 732 swp_pager_meta_free_all(object); 733 object->handle = NULL; 734 object->type = OBJT_DEAD; 735 } 736 737 /************************************************************************ 738 * SWAP PAGER BITMAP ROUTINES * 739 ************************************************************************/ 740 741 /* 742 * SWP_PAGER_GETSWAPSPACE() - allocate raw swap space 743 * 744 * Allocate swap for up to the requested number of pages, and at 745 * least a minimum number of pages. The starting swap block number 746 * (a page index) is returned or SWAPBLK_NONE if the allocation 747 * failed. 748 * 749 * Also has the side effect of advising that somebody made a mistake 750 * when they configured swap and didn't configure enough. 751 * 752 * This routine may not sleep. 753 * 754 * We allocate in round-robin fashion from the configured devices. 755 */ 756 static daddr_t 757 swp_pager_getswapspace(int *io_npages, int limit) 758 { 759 daddr_t blk; 760 struct swdevt *sp; 761 int mpages, npages; 762 763 blk = SWAPBLK_NONE; 764 mpages = *io_npages; 765 npages = imin(BLIST_MAX_ALLOC, mpages); 766 mtx_lock(&sw_dev_mtx); 767 sp = swdevhd; 768 while (!TAILQ_EMPTY(&swtailq)) { 769 if (sp == NULL) 770 sp = TAILQ_FIRST(&swtailq); 771 if ((sp->sw_flags & SW_CLOSING) == 0) 772 blk = blist_alloc(sp->sw_blist, &npages, mpages); 773 if (blk != SWAPBLK_NONE) 774 break; 775 sp = TAILQ_NEXT(sp, sw_list); 776 if (swdevhd == sp) { 777 if (npages <= limit) 778 break; 779 mpages = npages - 1; 780 npages >>= 1; 781 } 782 } 783 if (blk != SWAPBLK_NONE) { 784 *io_npages = npages; 785 blk += sp->sw_first; 786 sp->sw_used += npages; 787 swap_pager_avail -= npages; 788 swp_sizecheck(); 789 swdevhd = TAILQ_NEXT(sp, sw_list); 790 } else { 791 if (swap_pager_full != 2) { 792 printf("swp_pager_getswapspace(%d): failed\n", 793 *io_npages); 794 swap_pager_full = 2; 795 swap_pager_almost_full = 1; 796 } 797 swdevhd = NULL; 798 } 799 mtx_unlock(&sw_dev_mtx); 800 return (blk); 801 } 802 803 static bool 804 swp_pager_isondev(daddr_t blk, struct swdevt *sp) 805 { 806 807 return (blk >= sp->sw_first && blk < sp->sw_end); 808 } 809 810 static void 811 swp_pager_strategy(struct buf *bp) 812 { 813 struct swdevt *sp; 814 815 mtx_lock(&sw_dev_mtx); 816 TAILQ_FOREACH(sp, &swtailq, sw_list) { 817 if (swp_pager_isondev(bp->b_blkno, sp)) { 818 mtx_unlock(&sw_dev_mtx); 819 if ((sp->sw_flags & SW_UNMAPPED) != 0 && 820 unmapped_buf_allowed) { 821 bp->b_data = unmapped_buf; 822 bp->b_offset = 0; 823 } else { 824 pmap_qenter((vm_offset_t)bp->b_data, 825 &bp->b_pages[0], bp->b_bcount / PAGE_SIZE); 826 } 827 sp->sw_strategy(bp, sp); 828 return; 829 } 830 } 831 panic("Swapdev not found"); 832 } 833 834 835 /* 836 * SWP_PAGER_FREESWAPSPACE() - free raw swap space 837 * 838 * This routine returns the specified swap blocks back to the bitmap. 839 * 840 * This routine may not sleep. 841 */ 842 static void 843 swp_pager_freeswapspace(daddr_t blk, daddr_t npages) 844 { 845 struct swdevt *sp; 846 847 if (npages == 0) 848 return; 849 mtx_lock(&sw_dev_mtx); 850 TAILQ_FOREACH(sp, &swtailq, sw_list) { 851 if (swp_pager_isondev(blk, sp)) { 852 sp->sw_used -= npages; 853 /* 854 * If we are attempting to stop swapping on 855 * this device, we don't want to mark any 856 * blocks free lest they be reused. 857 */ 858 if ((sp->sw_flags & SW_CLOSING) == 0) { 859 blist_free(sp->sw_blist, blk - sp->sw_first, 860 npages); 861 swap_pager_avail += npages; 862 swp_sizecheck(); 863 } 864 mtx_unlock(&sw_dev_mtx); 865 return; 866 } 867 } 868 panic("Swapdev not found"); 869 } 870 871 /* 872 * SYSCTL_SWAP_FRAGMENTATION() - produce raw swap space stats 873 */ 874 static int 875 sysctl_swap_fragmentation(SYSCTL_HANDLER_ARGS) 876 { 877 struct sbuf sbuf; 878 struct swdevt *sp; 879 const char *devname; 880 int error; 881 882 error = sysctl_wire_old_buffer(req, 0); 883 if (error != 0) 884 return (error); 885 sbuf_new_for_sysctl(&sbuf, NULL, 128, req); 886 mtx_lock(&sw_dev_mtx); 887 TAILQ_FOREACH(sp, &swtailq, sw_list) { 888 if (vn_isdisk(sp->sw_vp, NULL)) 889 devname = devtoname(sp->sw_vp->v_rdev); 890 else 891 devname = "[file]"; 892 sbuf_printf(&sbuf, "\nFree space on device %s:\n", devname); 893 blist_stats(sp->sw_blist, &sbuf); 894 } 895 mtx_unlock(&sw_dev_mtx); 896 error = sbuf_finish(&sbuf); 897 sbuf_delete(&sbuf); 898 return (error); 899 } 900 901 /* 902 * SWAP_PAGER_FREESPACE() - frees swap blocks associated with a page 903 * range within an object. 904 * 905 * This is a globally accessible routine. 906 * 907 * This routine removes swapblk assignments from swap metadata. 908 * 909 * The external callers of this routine typically have already destroyed 910 * or renamed vm_page_t's associated with this range in the object so 911 * we should be ok. 912 * 913 * The object must be locked. 914 */ 915 void 916 swap_pager_freespace(vm_object_t object, vm_pindex_t start, vm_size_t size) 917 { 918 919 swp_pager_meta_free(object, start, size); 920 } 921 922 /* 923 * SWAP_PAGER_RESERVE() - reserve swap blocks in object 924 * 925 * Assigns swap blocks to the specified range within the object. The 926 * swap blocks are not zeroed. Any previous swap assignment is destroyed. 927 * 928 * Returns 0 on success, -1 on failure. 929 */ 930 int 931 swap_pager_reserve(vm_object_t object, vm_pindex_t start, vm_size_t size) 932 { 933 daddr_t addr, blk, n_free, s_free; 934 int i, j, n; 935 936 swp_pager_init_freerange(&s_free, &n_free); 937 VM_OBJECT_WLOCK(object); 938 for (i = 0; i < size; i += n) { 939 n = size - i; 940 blk = swp_pager_getswapspace(&n, 1); 941 if (blk == SWAPBLK_NONE) { 942 swp_pager_meta_free(object, start, i); 943 VM_OBJECT_WUNLOCK(object); 944 return (-1); 945 } 946 for (j = 0; j < n; ++j) { 947 addr = swp_pager_meta_build(object, 948 start + i + j, blk + j); 949 if (addr != SWAPBLK_NONE) 950 swp_pager_update_freerange(&s_free, &n_free, 951 addr); 952 } 953 } 954 swp_pager_freeswapspace(s_free, n_free); 955 VM_OBJECT_WUNLOCK(object); 956 return (0); 957 } 958 959 static bool 960 swp_pager_xfer_source(vm_object_t srcobject, vm_object_t dstobject, 961 vm_pindex_t pindex, daddr_t addr) 962 { 963 daddr_t dstaddr; 964 965 KASSERT(srcobject->type == OBJT_SWAP, 966 ("%s: Srcobject not swappable", __func__)); 967 if (dstobject->type == OBJT_SWAP && 968 swp_pager_meta_lookup(dstobject, pindex) != SWAPBLK_NONE) { 969 /* Caller should destroy the source block. */ 970 return (false); 971 } 972 973 /* 974 * Destination has no swapblk and is not resident, transfer source. 975 * swp_pager_meta_build() can sleep. 976 */ 977 VM_OBJECT_WUNLOCK(srcobject); 978 dstaddr = swp_pager_meta_build(dstobject, pindex, addr); 979 KASSERT(dstaddr == SWAPBLK_NONE, 980 ("Unexpected destination swapblk")); 981 VM_OBJECT_WLOCK(srcobject); 982 983 return (true); 984 } 985 986 /* 987 * SWAP_PAGER_COPY() - copy blocks from source pager to destination pager 988 * and destroy the source. 989 * 990 * Copy any valid swapblks from the source to the destination. In 991 * cases where both the source and destination have a valid swapblk, 992 * we keep the destination's. 993 * 994 * This routine is allowed to sleep. It may sleep allocating metadata 995 * indirectly through swp_pager_meta_build(). 996 * 997 * The source object contains no vm_page_t's (which is just as well) 998 * 999 * The source object is of type OBJT_SWAP. 1000 * 1001 * The source and destination objects must be locked. 1002 * Both object locks may temporarily be released. 1003 */ 1004 void 1005 swap_pager_copy(vm_object_t srcobject, vm_object_t dstobject, 1006 vm_pindex_t offset, int destroysource) 1007 { 1008 1009 VM_OBJECT_ASSERT_WLOCKED(srcobject); 1010 VM_OBJECT_ASSERT_WLOCKED(dstobject); 1011 1012 /* 1013 * If destroysource is set, we remove the source object from the 1014 * swap_pager internal queue now. 1015 */ 1016 if (destroysource && (srcobject->flags & OBJ_ANON) == 0 && 1017 srcobject->handle != NULL) { 1018 VM_OBJECT_WUNLOCK(srcobject); 1019 VM_OBJECT_WUNLOCK(dstobject); 1020 sx_xlock(&sw_alloc_sx); 1021 TAILQ_REMOVE(NOBJLIST(srcobject->handle), srcobject, 1022 pager_object_list); 1023 sx_xunlock(&sw_alloc_sx); 1024 VM_OBJECT_WLOCK(dstobject); 1025 VM_OBJECT_WLOCK(srcobject); 1026 } 1027 1028 /* 1029 * Transfer source to destination. 1030 */ 1031 swp_pager_meta_transfer(srcobject, dstobject, offset, dstobject->size); 1032 1033 /* 1034 * Free left over swap blocks in source. 1035 * 1036 * We have to revert the type to OBJT_DEFAULT so we do not accidentally 1037 * double-remove the object from the swap queues. 1038 */ 1039 if (destroysource) { 1040 swp_pager_meta_free_all(srcobject); 1041 /* 1042 * Reverting the type is not necessary, the caller is going 1043 * to destroy srcobject directly, but I'm doing it here 1044 * for consistency since we've removed the object from its 1045 * queues. 1046 */ 1047 srcobject->type = OBJT_DEFAULT; 1048 } 1049 } 1050 1051 /* 1052 * SWAP_PAGER_HASPAGE() - determine if we have good backing store for 1053 * the requested page. 1054 * 1055 * We determine whether good backing store exists for the requested 1056 * page and return TRUE if it does, FALSE if it doesn't. 1057 * 1058 * If TRUE, we also try to determine how much valid, contiguous backing 1059 * store exists before and after the requested page. 1060 */ 1061 static boolean_t 1062 swap_pager_haspage(vm_object_t object, vm_pindex_t pindex, int *before, 1063 int *after) 1064 { 1065 daddr_t blk, blk0; 1066 int i; 1067 1068 VM_OBJECT_ASSERT_LOCKED(object); 1069 KASSERT(object->type == OBJT_SWAP, 1070 ("%s: object not swappable", __func__)); 1071 1072 /* 1073 * do we have good backing store at the requested index ? 1074 */ 1075 blk0 = swp_pager_meta_lookup(object, pindex); 1076 if (blk0 == SWAPBLK_NONE) { 1077 if (before) 1078 *before = 0; 1079 if (after) 1080 *after = 0; 1081 return (FALSE); 1082 } 1083 1084 /* 1085 * find backwards-looking contiguous good backing store 1086 */ 1087 if (before != NULL) { 1088 for (i = 1; i < SWB_NPAGES; i++) { 1089 if (i > pindex) 1090 break; 1091 blk = swp_pager_meta_lookup(object, pindex - i); 1092 if (blk != blk0 - i) 1093 break; 1094 } 1095 *before = i - 1; 1096 } 1097 1098 /* 1099 * find forward-looking contiguous good backing store 1100 */ 1101 if (after != NULL) { 1102 for (i = 1; i < SWB_NPAGES; i++) { 1103 blk = swp_pager_meta_lookup(object, pindex + i); 1104 if (blk != blk0 + i) 1105 break; 1106 } 1107 *after = i - 1; 1108 } 1109 return (TRUE); 1110 } 1111 1112 /* 1113 * SWAP_PAGER_PAGE_UNSWAPPED() - remove swap backing store related to page 1114 * 1115 * This removes any associated swap backing store, whether valid or 1116 * not, from the page. 1117 * 1118 * This routine is typically called when a page is made dirty, at 1119 * which point any associated swap can be freed. MADV_FREE also 1120 * calls us in a special-case situation 1121 * 1122 * NOTE!!! If the page is clean and the swap was valid, the caller 1123 * should make the page dirty before calling this routine. This routine 1124 * does NOT change the m->dirty status of the page. Also: MADV_FREE 1125 * depends on it. 1126 * 1127 * This routine may not sleep. 1128 * 1129 * The object containing the page may be locked. 1130 */ 1131 static void 1132 swap_pager_unswapped(vm_page_t m) 1133 { 1134 struct swblk *sb; 1135 vm_object_t obj; 1136 1137 /* 1138 * Handle enqueing deferred frees first. If we do not have the 1139 * object lock we wait for the page daemon to clear the space. 1140 */ 1141 obj = m->object; 1142 if (!VM_OBJECT_WOWNED(obj)) { 1143 VM_PAGE_OBJECT_BUSY_ASSERT(m); 1144 /* 1145 * The caller is responsible for synchronization but we 1146 * will harmlessly handle races. This is typically provided 1147 * by only calling unswapped() when a page transitions from 1148 * clean to dirty. 1149 */ 1150 if ((m->a.flags & (PGA_SWAP_SPACE | PGA_SWAP_FREE)) == 1151 PGA_SWAP_SPACE) { 1152 vm_page_aflag_set(m, PGA_SWAP_FREE); 1153 counter_u64_add(swap_free_deferred, 1); 1154 } 1155 return; 1156 } 1157 if ((m->a.flags & PGA_SWAP_FREE) != 0) 1158 counter_u64_add(swap_free_completed, 1); 1159 vm_page_aflag_clear(m, PGA_SWAP_FREE | PGA_SWAP_SPACE); 1160 1161 /* 1162 * The meta data only exists if the object is OBJT_SWAP 1163 * and even then might not be allocated yet. 1164 */ 1165 KASSERT(m->object->type == OBJT_SWAP, 1166 ("Free object not swappable")); 1167 1168 sb = SWAP_PCTRIE_LOOKUP(&m->object->un_pager.swp.swp_blks, 1169 rounddown(m->pindex, SWAP_META_PAGES)); 1170 if (sb == NULL) 1171 return; 1172 if (sb->d[m->pindex % SWAP_META_PAGES] == SWAPBLK_NONE) 1173 return; 1174 swp_pager_freeswapspace(sb->d[m->pindex % SWAP_META_PAGES], 1); 1175 sb->d[m->pindex % SWAP_META_PAGES] = SWAPBLK_NONE; 1176 swp_pager_free_empty_swblk(m->object, sb); 1177 } 1178 1179 /* 1180 * swap_pager_getpages() - bring pages in from swap 1181 * 1182 * Attempt to page in the pages in array "ma" of length "count". The 1183 * caller may optionally specify that additional pages preceding and 1184 * succeeding the specified range be paged in. The number of such pages 1185 * is returned in the "rbehind" and "rahead" parameters, and they will 1186 * be in the inactive queue upon return. 1187 * 1188 * The pages in "ma" must be busied and will remain busied upon return. 1189 */ 1190 static int 1191 swap_pager_getpages_locked(vm_object_t object, vm_page_t *ma, int count, 1192 int *rbehind, int *rahead) 1193 { 1194 struct buf *bp; 1195 vm_page_t bm, mpred, msucc, p; 1196 vm_pindex_t pindex; 1197 daddr_t blk; 1198 int i, maxahead, maxbehind, reqcount; 1199 1200 VM_OBJECT_ASSERT_WLOCKED(object); 1201 reqcount = count; 1202 1203 KASSERT(object->type == OBJT_SWAP, 1204 ("%s: object not swappable", __func__)); 1205 if (!swap_pager_haspage(object, ma[0]->pindex, &maxbehind, &maxahead)) { 1206 VM_OBJECT_WUNLOCK(object); 1207 return (VM_PAGER_FAIL); 1208 } 1209 1210 KASSERT(reqcount - 1 <= maxahead, 1211 ("page count %d extends beyond swap block", reqcount)); 1212 1213 /* 1214 * Do not transfer any pages other than those that are xbusied 1215 * when running during a split or collapse operation. This 1216 * prevents clustering from re-creating pages which are being 1217 * moved into another object. 1218 */ 1219 if ((object->flags & (OBJ_SPLIT | OBJ_DEAD)) != 0) { 1220 maxahead = reqcount - 1; 1221 maxbehind = 0; 1222 } 1223 1224 /* 1225 * Clip the readahead and readbehind ranges to exclude resident pages. 1226 */ 1227 if (rahead != NULL) { 1228 *rahead = imin(*rahead, maxahead - (reqcount - 1)); 1229 pindex = ma[reqcount - 1]->pindex; 1230 msucc = TAILQ_NEXT(ma[reqcount - 1], listq); 1231 if (msucc != NULL && msucc->pindex - pindex - 1 < *rahead) 1232 *rahead = msucc->pindex - pindex - 1; 1233 } 1234 if (rbehind != NULL) { 1235 *rbehind = imin(*rbehind, maxbehind); 1236 pindex = ma[0]->pindex; 1237 mpred = TAILQ_PREV(ma[0], pglist, listq); 1238 if (mpred != NULL && pindex - mpred->pindex - 1 < *rbehind) 1239 *rbehind = pindex - mpred->pindex - 1; 1240 } 1241 1242 bm = ma[0]; 1243 for (i = 0; i < count; i++) 1244 ma[i]->oflags |= VPO_SWAPINPROG; 1245 1246 /* 1247 * Allocate readahead and readbehind pages. 1248 */ 1249 if (rbehind != NULL) { 1250 for (i = 1; i <= *rbehind; i++) { 1251 p = vm_page_alloc(object, ma[0]->pindex - i, 1252 VM_ALLOC_NORMAL); 1253 if (p == NULL) 1254 break; 1255 p->oflags |= VPO_SWAPINPROG; 1256 bm = p; 1257 } 1258 *rbehind = i - 1; 1259 } 1260 if (rahead != NULL) { 1261 for (i = 0; i < *rahead; i++) { 1262 p = vm_page_alloc(object, 1263 ma[reqcount - 1]->pindex + i + 1, VM_ALLOC_NORMAL); 1264 if (p == NULL) 1265 break; 1266 p->oflags |= VPO_SWAPINPROG; 1267 } 1268 *rahead = i; 1269 } 1270 if (rbehind != NULL) 1271 count += *rbehind; 1272 if (rahead != NULL) 1273 count += *rahead; 1274 1275 vm_object_pip_add(object, count); 1276 1277 pindex = bm->pindex; 1278 blk = swp_pager_meta_lookup(object, pindex); 1279 KASSERT(blk != SWAPBLK_NONE, 1280 ("no swap blocking containing %p(%jx)", object, (uintmax_t)pindex)); 1281 1282 VM_OBJECT_WUNLOCK(object); 1283 bp = uma_zalloc(swrbuf_zone, M_WAITOK); 1284 /* Pages cannot leave the object while busy. */ 1285 for (i = 0, p = bm; i < count; i++, p = TAILQ_NEXT(p, listq)) { 1286 MPASS(p->pindex == bm->pindex + i); 1287 bp->b_pages[i] = p; 1288 } 1289 1290 bp->b_flags |= B_PAGING; 1291 bp->b_iocmd = BIO_READ; 1292 bp->b_iodone = swp_pager_async_iodone; 1293 bp->b_rcred = crhold(thread0.td_ucred); 1294 bp->b_wcred = crhold(thread0.td_ucred); 1295 bp->b_blkno = blk; 1296 bp->b_bcount = PAGE_SIZE * count; 1297 bp->b_bufsize = PAGE_SIZE * count; 1298 bp->b_npages = count; 1299 bp->b_pgbefore = rbehind != NULL ? *rbehind : 0; 1300 bp->b_pgafter = rahead != NULL ? *rahead : 0; 1301 1302 VM_CNT_INC(v_swapin); 1303 VM_CNT_ADD(v_swappgsin, count); 1304 1305 /* 1306 * perform the I/O. NOTE!!! bp cannot be considered valid after 1307 * this point because we automatically release it on completion. 1308 * Instead, we look at the one page we are interested in which we 1309 * still hold a lock on even through the I/O completion. 1310 * 1311 * The other pages in our ma[] array are also released on completion, 1312 * so we cannot assume they are valid anymore either. 1313 * 1314 * NOTE: b_blkno is destroyed by the call to swapdev_strategy 1315 */ 1316 BUF_KERNPROC(bp); 1317 swp_pager_strategy(bp); 1318 1319 /* 1320 * Wait for the pages we want to complete. VPO_SWAPINPROG is always 1321 * cleared on completion. If an I/O error occurs, SWAPBLK_NONE 1322 * is set in the metadata for each page in the request. 1323 */ 1324 VM_OBJECT_WLOCK(object); 1325 /* This could be implemented more efficiently with aflags */ 1326 while ((ma[0]->oflags & VPO_SWAPINPROG) != 0) { 1327 ma[0]->oflags |= VPO_SWAPSLEEP; 1328 VM_CNT_INC(v_intrans); 1329 if (VM_OBJECT_SLEEP(object, &object->handle, PSWP, 1330 "swread", hz * 20)) { 1331 printf( 1332 "swap_pager: indefinite wait buffer: bufobj: %p, blkno: %jd, size: %ld\n", 1333 bp->b_bufobj, (intmax_t)bp->b_blkno, bp->b_bcount); 1334 } 1335 } 1336 VM_OBJECT_WUNLOCK(object); 1337 1338 /* 1339 * If we had an unrecoverable read error pages will not be valid. 1340 */ 1341 for (i = 0; i < reqcount; i++) 1342 if (ma[i]->valid != VM_PAGE_BITS_ALL) 1343 return (VM_PAGER_ERROR); 1344 1345 return (VM_PAGER_OK); 1346 1347 /* 1348 * A final note: in a low swap situation, we cannot deallocate swap 1349 * and mark a page dirty here because the caller is likely to mark 1350 * the page clean when we return, causing the page to possibly revert 1351 * to all-zero's later. 1352 */ 1353 } 1354 1355 static int 1356 swap_pager_getpages(vm_object_t object, vm_page_t *ma, int count, 1357 int *rbehind, int *rahead) 1358 { 1359 1360 VM_OBJECT_WLOCK(object); 1361 return (swap_pager_getpages_locked(object, ma, count, rbehind, rahead)); 1362 } 1363 1364 /* 1365 * swap_pager_getpages_async(): 1366 * 1367 * Right now this is emulation of asynchronous operation on top of 1368 * swap_pager_getpages(). 1369 */ 1370 static int 1371 swap_pager_getpages_async(vm_object_t object, vm_page_t *ma, int count, 1372 int *rbehind, int *rahead, pgo_getpages_iodone_t iodone, void *arg) 1373 { 1374 int r, error; 1375 1376 r = swap_pager_getpages(object, ma, count, rbehind, rahead); 1377 switch (r) { 1378 case VM_PAGER_OK: 1379 error = 0; 1380 break; 1381 case VM_PAGER_ERROR: 1382 error = EIO; 1383 break; 1384 case VM_PAGER_FAIL: 1385 error = EINVAL; 1386 break; 1387 default: 1388 panic("unhandled swap_pager_getpages() error %d", r); 1389 } 1390 (iodone)(arg, ma, count, error); 1391 1392 return (r); 1393 } 1394 1395 /* 1396 * swap_pager_putpages: 1397 * 1398 * Assign swap (if necessary) and initiate I/O on the specified pages. 1399 * 1400 * We support both OBJT_DEFAULT and OBJT_SWAP objects. DEFAULT objects 1401 * are automatically converted to SWAP objects. 1402 * 1403 * In a low memory situation we may block in VOP_STRATEGY(), but the new 1404 * vm_page reservation system coupled with properly written VFS devices 1405 * should ensure that no low-memory deadlock occurs. This is an area 1406 * which needs work. 1407 * 1408 * The parent has N vm_object_pip_add() references prior to 1409 * calling us and will remove references for rtvals[] that are 1410 * not set to VM_PAGER_PEND. We need to remove the rest on I/O 1411 * completion. 1412 * 1413 * The parent has soft-busy'd the pages it passes us and will unbusy 1414 * those whose rtvals[] entry is not set to VM_PAGER_PEND on return. 1415 * We need to unbusy the rest on I/O completion. 1416 */ 1417 static void 1418 swap_pager_putpages(vm_object_t object, vm_page_t *ma, int count, 1419 int flags, int *rtvals) 1420 { 1421 struct buf *bp; 1422 daddr_t addr, blk, n_free, s_free; 1423 vm_page_t mreq; 1424 int i, j, n; 1425 bool async; 1426 1427 KASSERT(count == 0 || ma[0]->object == object, 1428 ("%s: object mismatch %p/%p", 1429 __func__, object, ma[0]->object)); 1430 1431 /* 1432 * Step 1 1433 * 1434 * Turn object into OBJT_SWAP. Force sync if not a pageout process. 1435 */ 1436 if (object->type != OBJT_SWAP) { 1437 addr = swp_pager_meta_build(object, 0, SWAPBLK_NONE); 1438 KASSERT(addr == SWAPBLK_NONE, 1439 ("unexpected object swap block")); 1440 } 1441 VM_OBJECT_WUNLOCK(object); 1442 async = curproc == pageproc && (flags & VM_PAGER_PUT_SYNC) == 0; 1443 swp_pager_init_freerange(&s_free, &n_free); 1444 1445 /* 1446 * Step 2 1447 * 1448 * Assign swap blocks and issue I/O. We reallocate swap on the fly. 1449 * The page is left dirty until the pageout operation completes 1450 * successfully. 1451 */ 1452 for (i = 0; i < count; i += n) { 1453 /* Maximum I/O size is limited by maximum swap block size. */ 1454 n = min(count - i, nsw_cluster_max); 1455 1456 /* Get a block of swap of size up to size n. */ 1457 blk = swp_pager_getswapspace(&n, 4); 1458 if (blk == SWAPBLK_NONE) { 1459 for (j = 0; j < n; ++j) 1460 rtvals[i + j] = VM_PAGER_FAIL; 1461 continue; 1462 } 1463 1464 /* 1465 * All I/O parameters have been satisfied. Build the I/O 1466 * request and assign the swap space. 1467 */ 1468 if (async) { 1469 mtx_lock(&swbuf_mtx); 1470 while (nsw_wcount_async == 0) 1471 msleep(&nsw_wcount_async, &swbuf_mtx, PVM, 1472 "swbufa", 0); 1473 nsw_wcount_async--; 1474 mtx_unlock(&swbuf_mtx); 1475 } 1476 bp = uma_zalloc(swwbuf_zone, M_WAITOK); 1477 if (async) 1478 bp->b_flags = B_ASYNC; 1479 bp->b_flags |= B_PAGING; 1480 bp->b_iocmd = BIO_WRITE; 1481 1482 bp->b_rcred = crhold(thread0.td_ucred); 1483 bp->b_wcred = crhold(thread0.td_ucred); 1484 bp->b_bcount = PAGE_SIZE * n; 1485 bp->b_bufsize = PAGE_SIZE * n; 1486 bp->b_blkno = blk; 1487 1488 VM_OBJECT_WLOCK(object); 1489 for (j = 0; j < n; ++j) { 1490 mreq = ma[i + j]; 1491 vm_page_aflag_clear(mreq, PGA_SWAP_FREE); 1492 addr = swp_pager_meta_build(mreq->object, mreq->pindex, 1493 blk + j); 1494 if (addr != SWAPBLK_NONE) 1495 swp_pager_update_freerange(&s_free, &n_free, 1496 addr); 1497 MPASS(mreq->dirty == VM_PAGE_BITS_ALL); 1498 mreq->oflags |= VPO_SWAPINPROG; 1499 bp->b_pages[j] = mreq; 1500 } 1501 VM_OBJECT_WUNLOCK(object); 1502 bp->b_npages = n; 1503 /* 1504 * Must set dirty range for NFS to work. 1505 */ 1506 bp->b_dirtyoff = 0; 1507 bp->b_dirtyend = bp->b_bcount; 1508 1509 VM_CNT_INC(v_swapout); 1510 VM_CNT_ADD(v_swappgsout, bp->b_npages); 1511 1512 /* 1513 * We unconditionally set rtvals[] to VM_PAGER_PEND so that we 1514 * can call the async completion routine at the end of a 1515 * synchronous I/O operation. Otherwise, our caller would 1516 * perform duplicate unbusy and wakeup operations on the page 1517 * and object, respectively. 1518 */ 1519 for (j = 0; j < n; j++) 1520 rtvals[i + j] = VM_PAGER_PEND; 1521 1522 /* 1523 * asynchronous 1524 * 1525 * NOTE: b_blkno is destroyed by the call to swapdev_strategy. 1526 */ 1527 if (async) { 1528 bp->b_iodone = swp_pager_async_iodone; 1529 BUF_KERNPROC(bp); 1530 swp_pager_strategy(bp); 1531 continue; 1532 } 1533 1534 /* 1535 * synchronous 1536 * 1537 * NOTE: b_blkno is destroyed by the call to swapdev_strategy. 1538 */ 1539 bp->b_iodone = bdone; 1540 swp_pager_strategy(bp); 1541 1542 /* 1543 * Wait for the sync I/O to complete. 1544 */ 1545 bwait(bp, PVM, "swwrt"); 1546 1547 /* 1548 * Now that we are through with the bp, we can call the 1549 * normal async completion, which frees everything up. 1550 */ 1551 swp_pager_async_iodone(bp); 1552 } 1553 swp_pager_freeswapspace(s_free, n_free); 1554 VM_OBJECT_WLOCK(object); 1555 } 1556 1557 /* 1558 * swp_pager_async_iodone: 1559 * 1560 * Completion routine for asynchronous reads and writes from/to swap. 1561 * Also called manually by synchronous code to finish up a bp. 1562 * 1563 * This routine may not sleep. 1564 */ 1565 static void 1566 swp_pager_async_iodone(struct buf *bp) 1567 { 1568 int i; 1569 vm_object_t object = NULL; 1570 1571 /* 1572 * Report error - unless we ran out of memory, in which case 1573 * we've already logged it in swapgeom_strategy(). 1574 */ 1575 if (bp->b_ioflags & BIO_ERROR && bp->b_error != ENOMEM) { 1576 printf( 1577 "swap_pager: I/O error - %s failed; blkno %ld," 1578 "size %ld, error %d\n", 1579 ((bp->b_iocmd == BIO_READ) ? "pagein" : "pageout"), 1580 (long)bp->b_blkno, 1581 (long)bp->b_bcount, 1582 bp->b_error 1583 ); 1584 } 1585 1586 /* 1587 * remove the mapping for kernel virtual 1588 */ 1589 if (buf_mapped(bp)) 1590 pmap_qremove((vm_offset_t)bp->b_data, bp->b_npages); 1591 else 1592 bp->b_data = bp->b_kvabase; 1593 1594 if (bp->b_npages) { 1595 object = bp->b_pages[0]->object; 1596 VM_OBJECT_WLOCK(object); 1597 } 1598 1599 /* 1600 * cleanup pages. If an error occurs writing to swap, we are in 1601 * very serious trouble. If it happens to be a disk error, though, 1602 * we may be able to recover by reassigning the swap later on. So 1603 * in this case we remove the m->swapblk assignment for the page 1604 * but do not free it in the rlist. The errornous block(s) are thus 1605 * never reallocated as swap. Redirty the page and continue. 1606 */ 1607 for (i = 0; i < bp->b_npages; ++i) { 1608 vm_page_t m = bp->b_pages[i]; 1609 1610 m->oflags &= ~VPO_SWAPINPROG; 1611 if (m->oflags & VPO_SWAPSLEEP) { 1612 m->oflags &= ~VPO_SWAPSLEEP; 1613 wakeup(&object->handle); 1614 } 1615 1616 /* We always have space after I/O, successful or not. */ 1617 vm_page_aflag_set(m, PGA_SWAP_SPACE); 1618 1619 if (bp->b_ioflags & BIO_ERROR) { 1620 /* 1621 * If an error occurs I'd love to throw the swapblk 1622 * away without freeing it back to swapspace, so it 1623 * can never be used again. But I can't from an 1624 * interrupt. 1625 */ 1626 if (bp->b_iocmd == BIO_READ) { 1627 /* 1628 * NOTE: for reads, m->dirty will probably 1629 * be overridden by the original caller of 1630 * getpages so don't play cute tricks here. 1631 */ 1632 vm_page_invalid(m); 1633 } else { 1634 /* 1635 * If a write error occurs, reactivate page 1636 * so it doesn't clog the inactive list, 1637 * then finish the I/O. 1638 */ 1639 MPASS(m->dirty == VM_PAGE_BITS_ALL); 1640 1641 /* PQ_UNSWAPPABLE? */ 1642 vm_page_activate(m); 1643 vm_page_sunbusy(m); 1644 } 1645 } else if (bp->b_iocmd == BIO_READ) { 1646 /* 1647 * NOTE: for reads, m->dirty will probably be 1648 * overridden by the original caller of getpages so 1649 * we cannot set them in order to free the underlying 1650 * swap in a low-swap situation. I don't think we'd 1651 * want to do that anyway, but it was an optimization 1652 * that existed in the old swapper for a time before 1653 * it got ripped out due to precisely this problem. 1654 */ 1655 KASSERT(!pmap_page_is_mapped(m), 1656 ("swp_pager_async_iodone: page %p is mapped", m)); 1657 KASSERT(m->dirty == 0, 1658 ("swp_pager_async_iodone: page %p is dirty", m)); 1659 1660 vm_page_valid(m); 1661 if (i < bp->b_pgbefore || 1662 i >= bp->b_npages - bp->b_pgafter) 1663 vm_page_readahead_finish(m); 1664 } else { 1665 /* 1666 * For write success, clear the dirty 1667 * status, then finish the I/O ( which decrements the 1668 * busy count and possibly wakes waiter's up ). 1669 * A page is only written to swap after a period of 1670 * inactivity. Therefore, we do not expect it to be 1671 * reused. 1672 */ 1673 KASSERT(!pmap_page_is_write_mapped(m), 1674 ("swp_pager_async_iodone: page %p is not write" 1675 " protected", m)); 1676 vm_page_undirty(m); 1677 vm_page_deactivate_noreuse(m); 1678 vm_page_sunbusy(m); 1679 } 1680 } 1681 1682 /* 1683 * adjust pip. NOTE: the original parent may still have its own 1684 * pip refs on the object. 1685 */ 1686 if (object != NULL) { 1687 vm_object_pip_wakeupn(object, bp->b_npages); 1688 VM_OBJECT_WUNLOCK(object); 1689 } 1690 1691 /* 1692 * swapdev_strategy() manually sets b_vp and b_bufobj before calling 1693 * bstrategy(). Set them back to NULL now we're done with it, or we'll 1694 * trigger a KASSERT in relpbuf(). 1695 */ 1696 if (bp->b_vp) { 1697 bp->b_vp = NULL; 1698 bp->b_bufobj = NULL; 1699 } 1700 /* 1701 * release the physical I/O buffer 1702 */ 1703 if (bp->b_flags & B_ASYNC) { 1704 mtx_lock(&swbuf_mtx); 1705 if (++nsw_wcount_async == 1) 1706 wakeup(&nsw_wcount_async); 1707 mtx_unlock(&swbuf_mtx); 1708 } 1709 uma_zfree((bp->b_iocmd == BIO_READ) ? swrbuf_zone : swwbuf_zone, bp); 1710 } 1711 1712 int 1713 swap_pager_nswapdev(void) 1714 { 1715 1716 return (nswapdev); 1717 } 1718 1719 static void 1720 swp_pager_force_dirty(vm_page_t m) 1721 { 1722 1723 vm_page_dirty(m); 1724 swap_pager_unswapped(m); 1725 vm_page_launder(m); 1726 } 1727 1728 /* 1729 * swap_pager_swapoff_object: 1730 * 1731 * Page in all of the pages that have been paged out for an object 1732 * to a swap device. 1733 */ 1734 static void 1735 swap_pager_swapoff_object(struct swdevt *sp, vm_object_t object) 1736 { 1737 struct swblk *sb; 1738 vm_page_t m; 1739 vm_pindex_t pi; 1740 daddr_t blk; 1741 int i, nv, rahead, rv; 1742 1743 KASSERT(object->type == OBJT_SWAP, 1744 ("%s: Object not swappable", __func__)); 1745 1746 for (pi = 0; (sb = SWAP_PCTRIE_LOOKUP_GE( 1747 &object->un_pager.swp.swp_blks, pi)) != NULL; ) { 1748 if ((object->flags & OBJ_DEAD) != 0) { 1749 /* 1750 * Make sure that pending writes finish before 1751 * returning. 1752 */ 1753 vm_object_pip_wait(object, "swpoff"); 1754 swp_pager_meta_free_all(object); 1755 break; 1756 } 1757 for (i = 0; i < SWAP_META_PAGES; i++) { 1758 /* 1759 * Count the number of contiguous valid blocks. 1760 */ 1761 for (nv = 0; nv < SWAP_META_PAGES - i; nv++) { 1762 blk = sb->d[i + nv]; 1763 if (!swp_pager_isondev(blk, sp) || 1764 blk == SWAPBLK_NONE) 1765 break; 1766 } 1767 if (nv == 0) 1768 continue; 1769 1770 /* 1771 * Look for a page corresponding to the first 1772 * valid block and ensure that any pending paging 1773 * operations on it are complete. If the page is valid, 1774 * mark it dirty and free the swap block. Try to batch 1775 * this operation since it may cause sp to be freed, 1776 * meaning that we must restart the scan. Avoid busying 1777 * valid pages since we may block forever on kernel 1778 * stack pages. 1779 */ 1780 m = vm_page_lookup(object, sb->p + i); 1781 if (m == NULL) { 1782 m = vm_page_alloc(object, sb->p + i, 1783 VM_ALLOC_NORMAL | VM_ALLOC_WAITFAIL); 1784 if (m == NULL) 1785 break; 1786 } else { 1787 if ((m->oflags & VPO_SWAPINPROG) != 0) { 1788 m->oflags |= VPO_SWAPSLEEP; 1789 VM_OBJECT_SLEEP(object, &object->handle, 1790 PSWP, "swpoff", 0); 1791 break; 1792 } 1793 if (vm_page_all_valid(m)) { 1794 do { 1795 swp_pager_force_dirty(m); 1796 } while (--nv > 0 && 1797 (m = vm_page_next(m)) != NULL && 1798 vm_page_all_valid(m) && 1799 (m->oflags & VPO_SWAPINPROG) == 0); 1800 break; 1801 } 1802 if (!vm_page_busy_acquire(m, VM_ALLOC_WAITFAIL)) 1803 break; 1804 } 1805 1806 vm_object_pip_add(object, 1); 1807 rahead = SWAP_META_PAGES; 1808 rv = swap_pager_getpages_locked(object, &m, 1, NULL, 1809 &rahead); 1810 if (rv != VM_PAGER_OK) 1811 panic("%s: read from swap failed: %d", 1812 __func__, rv); 1813 vm_object_pip_wakeupn(object, 1); 1814 VM_OBJECT_WLOCK(object); 1815 vm_page_xunbusy(m); 1816 1817 /* 1818 * The object lock was dropped so we must restart the 1819 * scan of this swap block. Pages paged in during this 1820 * iteration will be marked dirty in a future iteration. 1821 */ 1822 break; 1823 } 1824 if (i == SWAP_META_PAGES) 1825 pi = sb->p + SWAP_META_PAGES; 1826 } 1827 } 1828 1829 /* 1830 * swap_pager_swapoff: 1831 * 1832 * Page in all of the pages that have been paged out to the 1833 * given device. The corresponding blocks in the bitmap must be 1834 * marked as allocated and the device must be flagged SW_CLOSING. 1835 * There may be no processes swapped out to the device. 1836 * 1837 * This routine may block. 1838 */ 1839 static void 1840 swap_pager_swapoff(struct swdevt *sp) 1841 { 1842 vm_object_t object; 1843 int retries; 1844 1845 sx_assert(&swdev_syscall_lock, SA_XLOCKED); 1846 1847 retries = 0; 1848 full_rescan: 1849 mtx_lock(&vm_object_list_mtx); 1850 TAILQ_FOREACH(object, &vm_object_list, object_list) { 1851 if (object->type != OBJT_SWAP) 1852 continue; 1853 mtx_unlock(&vm_object_list_mtx); 1854 /* Depends on type-stability. */ 1855 VM_OBJECT_WLOCK(object); 1856 1857 /* 1858 * Dead objects are eventually terminated on their own. 1859 */ 1860 if ((object->flags & OBJ_DEAD) != 0) 1861 goto next_obj; 1862 1863 /* 1864 * Sync with fences placed after pctrie 1865 * initialization. We must not access pctrie below 1866 * unless we checked that our object is swap and not 1867 * dead. 1868 */ 1869 atomic_thread_fence_acq(); 1870 if (object->type != OBJT_SWAP) 1871 goto next_obj; 1872 1873 swap_pager_swapoff_object(sp, object); 1874 next_obj: 1875 VM_OBJECT_WUNLOCK(object); 1876 mtx_lock(&vm_object_list_mtx); 1877 } 1878 mtx_unlock(&vm_object_list_mtx); 1879 1880 if (sp->sw_used) { 1881 /* 1882 * Objects may be locked or paging to the device being 1883 * removed, so we will miss their pages and need to 1884 * make another pass. We have marked this device as 1885 * SW_CLOSING, so the activity should finish soon. 1886 */ 1887 retries++; 1888 if (retries > 100) { 1889 panic("swapoff: failed to locate %d swap blocks", 1890 sp->sw_used); 1891 } 1892 pause("swpoff", hz / 20); 1893 goto full_rescan; 1894 } 1895 EVENTHANDLER_INVOKE(swapoff, sp); 1896 } 1897 1898 /************************************************************************ 1899 * SWAP META DATA * 1900 ************************************************************************ 1901 * 1902 * These routines manipulate the swap metadata stored in the 1903 * OBJT_SWAP object. 1904 * 1905 * Swap metadata is implemented with a global hash and not directly 1906 * linked into the object. Instead the object simply contains 1907 * appropriate tracking counters. 1908 */ 1909 1910 /* 1911 * SWP_PAGER_SWBLK_EMPTY() - is a range of blocks free? 1912 */ 1913 static bool 1914 swp_pager_swblk_empty(struct swblk *sb, int start, int limit) 1915 { 1916 int i; 1917 1918 MPASS(0 <= start && start <= limit && limit <= SWAP_META_PAGES); 1919 for (i = start; i < limit; i++) { 1920 if (sb->d[i] != SWAPBLK_NONE) 1921 return (false); 1922 } 1923 return (true); 1924 } 1925 1926 /* 1927 * SWP_PAGER_FREE_EMPTY_SWBLK() - frees if a block is free 1928 * 1929 * Nothing is done if the block is still in use. 1930 */ 1931 static void 1932 swp_pager_free_empty_swblk(vm_object_t object, struct swblk *sb) 1933 { 1934 1935 if (swp_pager_swblk_empty(sb, 0, SWAP_META_PAGES)) { 1936 SWAP_PCTRIE_REMOVE(&object->un_pager.swp.swp_blks, sb->p); 1937 uma_zfree(swblk_zone, sb); 1938 } 1939 } 1940 1941 /* 1942 * SWP_PAGER_META_BUILD() - add swap block to swap meta data for object 1943 * 1944 * We first convert the object to a swap object if it is a default 1945 * object. 1946 * 1947 * The specified swapblk is added to the object's swap metadata. If 1948 * the swapblk is not valid, it is freed instead. Any previously 1949 * assigned swapblk is returned. 1950 */ 1951 static daddr_t 1952 swp_pager_meta_build(vm_object_t object, vm_pindex_t pindex, daddr_t swapblk) 1953 { 1954 static volatile int swblk_zone_exhausted, swpctrie_zone_exhausted; 1955 struct swblk *sb, *sb1; 1956 vm_pindex_t modpi, rdpi; 1957 daddr_t prev_swapblk; 1958 int error, i; 1959 1960 VM_OBJECT_ASSERT_WLOCKED(object); 1961 1962 /* 1963 * Convert default object to swap object if necessary 1964 */ 1965 if (object->type != OBJT_SWAP) { 1966 pctrie_init(&object->un_pager.swp.swp_blks); 1967 1968 /* 1969 * Ensure that swap_pager_swapoff()'s iteration over 1970 * object_list does not see a garbage pctrie. 1971 */ 1972 atomic_thread_fence_rel(); 1973 1974 object->type = OBJT_SWAP; 1975 object->un_pager.swp.writemappings = 0; 1976 KASSERT((object->flags & OBJ_ANON) != 0 || 1977 object->handle == NULL, 1978 ("default pager %p with handle %p", 1979 object, object->handle)); 1980 } 1981 1982 rdpi = rounddown(pindex, SWAP_META_PAGES); 1983 sb = SWAP_PCTRIE_LOOKUP(&object->un_pager.swp.swp_blks, rdpi); 1984 if (sb == NULL) { 1985 if (swapblk == SWAPBLK_NONE) 1986 return (SWAPBLK_NONE); 1987 for (;;) { 1988 sb = uma_zalloc(swblk_zone, M_NOWAIT | (curproc == 1989 pageproc ? M_USE_RESERVE : 0)); 1990 if (sb != NULL) { 1991 sb->p = rdpi; 1992 for (i = 0; i < SWAP_META_PAGES; i++) 1993 sb->d[i] = SWAPBLK_NONE; 1994 if (atomic_cmpset_int(&swblk_zone_exhausted, 1995 1, 0)) 1996 printf("swblk zone ok\n"); 1997 break; 1998 } 1999 VM_OBJECT_WUNLOCK(object); 2000 if (uma_zone_exhausted(swblk_zone)) { 2001 if (atomic_cmpset_int(&swblk_zone_exhausted, 2002 0, 1)) 2003 printf("swap blk zone exhausted, " 2004 "increase kern.maxswzone\n"); 2005 vm_pageout_oom(VM_OOM_SWAPZ); 2006 pause("swzonxb", 10); 2007 } else 2008 uma_zwait(swblk_zone); 2009 VM_OBJECT_WLOCK(object); 2010 sb = SWAP_PCTRIE_LOOKUP(&object->un_pager.swp.swp_blks, 2011 rdpi); 2012 if (sb != NULL) 2013 /* 2014 * Somebody swapped out a nearby page, 2015 * allocating swblk at the rdpi index, 2016 * while we dropped the object lock. 2017 */ 2018 goto allocated; 2019 } 2020 for (;;) { 2021 error = SWAP_PCTRIE_INSERT( 2022 &object->un_pager.swp.swp_blks, sb); 2023 if (error == 0) { 2024 if (atomic_cmpset_int(&swpctrie_zone_exhausted, 2025 1, 0)) 2026 printf("swpctrie zone ok\n"); 2027 break; 2028 } 2029 VM_OBJECT_WUNLOCK(object); 2030 if (uma_zone_exhausted(swpctrie_zone)) { 2031 if (atomic_cmpset_int(&swpctrie_zone_exhausted, 2032 0, 1)) 2033 printf("swap pctrie zone exhausted, " 2034 "increase kern.maxswzone\n"); 2035 vm_pageout_oom(VM_OOM_SWAPZ); 2036 pause("swzonxp", 10); 2037 } else 2038 uma_zwait(swpctrie_zone); 2039 VM_OBJECT_WLOCK(object); 2040 sb1 = SWAP_PCTRIE_LOOKUP(&object->un_pager.swp.swp_blks, 2041 rdpi); 2042 if (sb1 != NULL) { 2043 uma_zfree(swblk_zone, sb); 2044 sb = sb1; 2045 goto allocated; 2046 } 2047 } 2048 } 2049 allocated: 2050 MPASS(sb->p == rdpi); 2051 2052 modpi = pindex % SWAP_META_PAGES; 2053 /* Return prior contents of metadata. */ 2054 prev_swapblk = sb->d[modpi]; 2055 /* Enter block into metadata. */ 2056 sb->d[modpi] = swapblk; 2057 2058 /* 2059 * Free the swblk if we end up with the empty page run. 2060 */ 2061 if (swapblk == SWAPBLK_NONE) 2062 swp_pager_free_empty_swblk(object, sb); 2063 return (prev_swapblk); 2064 } 2065 2066 /* 2067 * SWP_PAGER_META_TRANSFER() - free a range of blocks in the srcobject's swap 2068 * metadata, or transfer it into dstobject. 2069 * 2070 * This routine will free swap metadata structures as they are cleaned 2071 * out. 2072 */ 2073 static void 2074 swp_pager_meta_transfer(vm_object_t srcobject, vm_object_t dstobject, 2075 vm_pindex_t pindex, vm_pindex_t count) 2076 { 2077 struct swblk *sb; 2078 daddr_t n_free, s_free; 2079 vm_pindex_t offset, last; 2080 int i, limit, start; 2081 2082 VM_OBJECT_ASSERT_WLOCKED(srcobject); 2083 if (srcobject->type != OBJT_SWAP || count == 0) 2084 return; 2085 2086 swp_pager_init_freerange(&s_free, &n_free); 2087 offset = pindex; 2088 last = pindex + count; 2089 for (;;) { 2090 sb = SWAP_PCTRIE_LOOKUP_GE(&srcobject->un_pager.swp.swp_blks, 2091 rounddown(pindex, SWAP_META_PAGES)); 2092 if (sb == NULL || sb->p >= last) 2093 break; 2094 start = pindex > sb->p ? pindex - sb->p : 0; 2095 limit = last - sb->p < SWAP_META_PAGES ? last - sb->p : 2096 SWAP_META_PAGES; 2097 for (i = start; i < limit; i++) { 2098 if (sb->d[i] == SWAPBLK_NONE) 2099 continue; 2100 if (dstobject == NULL || 2101 !swp_pager_xfer_source(srcobject, dstobject, 2102 sb->p + i - offset, sb->d[i])) { 2103 swp_pager_update_freerange(&s_free, &n_free, 2104 sb->d[i]); 2105 } 2106 sb->d[i] = SWAPBLK_NONE; 2107 } 2108 pindex = sb->p + SWAP_META_PAGES; 2109 if (swp_pager_swblk_empty(sb, 0, start) && 2110 swp_pager_swblk_empty(sb, limit, SWAP_META_PAGES)) { 2111 SWAP_PCTRIE_REMOVE(&srcobject->un_pager.swp.swp_blks, 2112 sb->p); 2113 uma_zfree(swblk_zone, sb); 2114 } 2115 } 2116 swp_pager_freeswapspace(s_free, n_free); 2117 } 2118 2119 /* 2120 * SWP_PAGER_META_FREE() - free a range of blocks in the object's swap metadata 2121 * 2122 * The requested range of blocks is freed, with any associated swap 2123 * returned to the swap bitmap. 2124 * 2125 * This routine will free swap metadata structures as they are cleaned 2126 * out. This routine does *NOT* operate on swap metadata associated 2127 * with resident pages. 2128 */ 2129 static void 2130 swp_pager_meta_free(vm_object_t object, vm_pindex_t pindex, vm_pindex_t count) 2131 { 2132 swp_pager_meta_transfer(object, NULL, pindex, count); 2133 } 2134 2135 /* 2136 * SWP_PAGER_META_FREE_ALL() - destroy all swap metadata associated with object 2137 * 2138 * This routine locates and destroys all swap metadata associated with 2139 * an object. 2140 */ 2141 static void 2142 swp_pager_meta_free_all(vm_object_t object) 2143 { 2144 struct swblk *sb; 2145 daddr_t n_free, s_free; 2146 vm_pindex_t pindex; 2147 int i; 2148 2149 VM_OBJECT_ASSERT_WLOCKED(object); 2150 if (object->type != OBJT_SWAP) 2151 return; 2152 2153 swp_pager_init_freerange(&s_free, &n_free); 2154 for (pindex = 0; (sb = SWAP_PCTRIE_LOOKUP_GE( 2155 &object->un_pager.swp.swp_blks, pindex)) != NULL;) { 2156 pindex = sb->p + SWAP_META_PAGES; 2157 for (i = 0; i < SWAP_META_PAGES; i++) { 2158 if (sb->d[i] == SWAPBLK_NONE) 2159 continue; 2160 swp_pager_update_freerange(&s_free, &n_free, sb->d[i]); 2161 } 2162 SWAP_PCTRIE_REMOVE(&object->un_pager.swp.swp_blks, sb->p); 2163 uma_zfree(swblk_zone, sb); 2164 } 2165 swp_pager_freeswapspace(s_free, n_free); 2166 } 2167 2168 /* 2169 * SWP_PAGER_METACTL() - misc control of swap meta data. 2170 * 2171 * This routine is capable of looking up, or removing swapblk 2172 * assignments in the swap meta data. It returns the swapblk being 2173 * looked-up, popped, or SWAPBLK_NONE if the block was invalid. 2174 * 2175 * When acting on a busy resident page and paging is in progress, we 2176 * have to wait until paging is complete but otherwise can act on the 2177 * busy page. 2178 */ 2179 static daddr_t 2180 swp_pager_meta_lookup(vm_object_t object, vm_pindex_t pindex) 2181 { 2182 struct swblk *sb; 2183 2184 VM_OBJECT_ASSERT_LOCKED(object); 2185 2186 /* 2187 * The meta data only exists if the object is OBJT_SWAP 2188 * and even then might not be allocated yet. 2189 */ 2190 KASSERT(object->type == OBJT_SWAP, 2191 ("Lookup object not swappable")); 2192 2193 sb = SWAP_PCTRIE_LOOKUP(&object->un_pager.swp.swp_blks, 2194 rounddown(pindex, SWAP_META_PAGES)); 2195 if (sb == NULL) 2196 return (SWAPBLK_NONE); 2197 return (sb->d[pindex % SWAP_META_PAGES]); 2198 } 2199 2200 /* 2201 * Returns the least page index which is greater than or equal to the 2202 * parameter pindex and for which there is a swap block allocated. 2203 * Returns object's size if the object's type is not swap or if there 2204 * are no allocated swap blocks for the object after the requested 2205 * pindex. 2206 */ 2207 vm_pindex_t 2208 swap_pager_find_least(vm_object_t object, vm_pindex_t pindex) 2209 { 2210 struct swblk *sb; 2211 int i; 2212 2213 VM_OBJECT_ASSERT_LOCKED(object); 2214 if (object->type != OBJT_SWAP) 2215 return (object->size); 2216 2217 sb = SWAP_PCTRIE_LOOKUP_GE(&object->un_pager.swp.swp_blks, 2218 rounddown(pindex, SWAP_META_PAGES)); 2219 if (sb == NULL) 2220 return (object->size); 2221 if (sb->p < pindex) { 2222 for (i = pindex % SWAP_META_PAGES; i < SWAP_META_PAGES; i++) { 2223 if (sb->d[i] != SWAPBLK_NONE) 2224 return (sb->p + i); 2225 } 2226 sb = SWAP_PCTRIE_LOOKUP_GE(&object->un_pager.swp.swp_blks, 2227 roundup(pindex, SWAP_META_PAGES)); 2228 if (sb == NULL) 2229 return (object->size); 2230 } 2231 for (i = 0; i < SWAP_META_PAGES; i++) { 2232 if (sb->d[i] != SWAPBLK_NONE) 2233 return (sb->p + i); 2234 } 2235 2236 /* 2237 * We get here if a swblk is present in the trie but it 2238 * doesn't map any blocks. 2239 */ 2240 MPASS(0); 2241 return (object->size); 2242 } 2243 2244 /* 2245 * System call swapon(name) enables swapping on device name, 2246 * which must be in the swdevsw. Return EBUSY 2247 * if already swapping on this device. 2248 */ 2249 #ifndef _SYS_SYSPROTO_H_ 2250 struct swapon_args { 2251 char *name; 2252 }; 2253 #endif 2254 2255 /* 2256 * MPSAFE 2257 */ 2258 /* ARGSUSED */ 2259 int 2260 sys_swapon(struct thread *td, struct swapon_args *uap) 2261 { 2262 struct vattr attr; 2263 struct vnode *vp; 2264 struct nameidata nd; 2265 int error; 2266 2267 error = priv_check(td, PRIV_SWAPON); 2268 if (error) 2269 return (error); 2270 2271 sx_xlock(&swdev_syscall_lock); 2272 2273 /* 2274 * Swap metadata may not fit in the KVM if we have physical 2275 * memory of >1GB. 2276 */ 2277 if (swblk_zone == NULL) { 2278 error = ENOMEM; 2279 goto done; 2280 } 2281 2282 NDINIT(&nd, LOOKUP, ISOPEN | FOLLOW | AUDITVNODE1, UIO_USERSPACE, 2283 uap->name, td); 2284 error = namei(&nd); 2285 if (error) 2286 goto done; 2287 2288 NDFREE(&nd, NDF_ONLY_PNBUF); 2289 vp = nd.ni_vp; 2290 2291 if (vn_isdisk(vp, &error)) { 2292 error = swapongeom(vp); 2293 } else if (vp->v_type == VREG && 2294 (vp->v_mount->mnt_vfc->vfc_flags & VFCF_NETWORK) != 0 && 2295 (error = VOP_GETATTR(vp, &attr, td->td_ucred)) == 0) { 2296 /* 2297 * Allow direct swapping to NFS regular files in the same 2298 * way that nfs_mountroot() sets up diskless swapping. 2299 */ 2300 error = swaponvp(td, vp, attr.va_size / DEV_BSIZE); 2301 } 2302 2303 if (error) 2304 vrele(vp); 2305 done: 2306 sx_xunlock(&swdev_syscall_lock); 2307 return (error); 2308 } 2309 2310 /* 2311 * Check that the total amount of swap currently configured does not 2312 * exceed half the theoretical maximum. If it does, print a warning 2313 * message. 2314 */ 2315 static void 2316 swapon_check_swzone(void) 2317 { 2318 2319 /* recommend using no more than half that amount */ 2320 if (swap_total > swap_maxpages / 2) { 2321 printf("warning: total configured swap (%lu pages) " 2322 "exceeds maximum recommended amount (%lu pages).\n", 2323 swap_total, swap_maxpages / 2); 2324 printf("warning: increase kern.maxswzone " 2325 "or reduce amount of swap.\n"); 2326 } 2327 } 2328 2329 static void 2330 swaponsomething(struct vnode *vp, void *id, u_long nblks, 2331 sw_strategy_t *strategy, sw_close_t *close, dev_t dev, int flags) 2332 { 2333 struct swdevt *sp, *tsp; 2334 swblk_t dvbase; 2335 u_long mblocks; 2336 2337 /* 2338 * nblks is in DEV_BSIZE'd chunks, convert to PAGE_SIZE'd chunks. 2339 * First chop nblks off to page-align it, then convert. 2340 * 2341 * sw->sw_nblks is in page-sized chunks now too. 2342 */ 2343 nblks &= ~(ctodb(1) - 1); 2344 nblks = dbtoc(nblks); 2345 2346 /* 2347 * If we go beyond this, we get overflows in the radix 2348 * tree bitmap code. 2349 */ 2350 mblocks = 0x40000000 / BLIST_META_RADIX; 2351 if (nblks > mblocks) { 2352 printf( 2353 "WARNING: reducing swap size to maximum of %luMB per unit\n", 2354 mblocks / 1024 / 1024 * PAGE_SIZE); 2355 nblks = mblocks; 2356 } 2357 2358 sp = malloc(sizeof *sp, M_VMPGDATA, M_WAITOK | M_ZERO); 2359 sp->sw_vp = vp; 2360 sp->sw_id = id; 2361 sp->sw_dev = dev; 2362 sp->sw_nblks = nblks; 2363 sp->sw_used = 0; 2364 sp->sw_strategy = strategy; 2365 sp->sw_close = close; 2366 sp->sw_flags = flags; 2367 2368 sp->sw_blist = blist_create(nblks, M_WAITOK); 2369 /* 2370 * Do not free the first blocks in order to avoid overwriting 2371 * any bsd label at the front of the partition 2372 */ 2373 blist_free(sp->sw_blist, howmany(BBSIZE, PAGE_SIZE), 2374 nblks - howmany(BBSIZE, PAGE_SIZE)); 2375 2376 dvbase = 0; 2377 mtx_lock(&sw_dev_mtx); 2378 TAILQ_FOREACH(tsp, &swtailq, sw_list) { 2379 if (tsp->sw_end >= dvbase) { 2380 /* 2381 * We put one uncovered page between the devices 2382 * in order to definitively prevent any cross-device 2383 * I/O requests 2384 */ 2385 dvbase = tsp->sw_end + 1; 2386 } 2387 } 2388 sp->sw_first = dvbase; 2389 sp->sw_end = dvbase + nblks; 2390 TAILQ_INSERT_TAIL(&swtailq, sp, sw_list); 2391 nswapdev++; 2392 swap_pager_avail += nblks - howmany(BBSIZE, PAGE_SIZE); 2393 swap_total += nblks; 2394 swapon_check_swzone(); 2395 swp_sizecheck(); 2396 mtx_unlock(&sw_dev_mtx); 2397 EVENTHANDLER_INVOKE(swapon, sp); 2398 } 2399 2400 /* 2401 * SYSCALL: swapoff(devname) 2402 * 2403 * Disable swapping on the given device. 2404 * 2405 * XXX: Badly designed system call: it should use a device index 2406 * rather than filename as specification. We keep sw_vp around 2407 * only to make this work. 2408 */ 2409 #ifndef _SYS_SYSPROTO_H_ 2410 struct swapoff_args { 2411 char *name; 2412 }; 2413 #endif 2414 2415 /* 2416 * MPSAFE 2417 */ 2418 /* ARGSUSED */ 2419 int 2420 sys_swapoff(struct thread *td, struct swapoff_args *uap) 2421 { 2422 struct vnode *vp; 2423 struct nameidata nd; 2424 struct swdevt *sp; 2425 int error; 2426 2427 error = priv_check(td, PRIV_SWAPOFF); 2428 if (error) 2429 return (error); 2430 2431 sx_xlock(&swdev_syscall_lock); 2432 2433 NDINIT(&nd, LOOKUP, FOLLOW | AUDITVNODE1, UIO_USERSPACE, uap->name, 2434 td); 2435 error = namei(&nd); 2436 if (error) 2437 goto done; 2438 NDFREE(&nd, NDF_ONLY_PNBUF); 2439 vp = nd.ni_vp; 2440 2441 mtx_lock(&sw_dev_mtx); 2442 TAILQ_FOREACH(sp, &swtailq, sw_list) { 2443 if (sp->sw_vp == vp) 2444 break; 2445 } 2446 mtx_unlock(&sw_dev_mtx); 2447 if (sp == NULL) { 2448 error = EINVAL; 2449 goto done; 2450 } 2451 error = swapoff_one(sp, td->td_ucred); 2452 done: 2453 sx_xunlock(&swdev_syscall_lock); 2454 return (error); 2455 } 2456 2457 static int 2458 swapoff_one(struct swdevt *sp, struct ucred *cred) 2459 { 2460 u_long nblks; 2461 #ifdef MAC 2462 int error; 2463 #endif 2464 2465 sx_assert(&swdev_syscall_lock, SA_XLOCKED); 2466 #ifdef MAC 2467 (void) vn_lock(sp->sw_vp, LK_EXCLUSIVE | LK_RETRY); 2468 error = mac_system_check_swapoff(cred, sp->sw_vp); 2469 (void) VOP_UNLOCK(sp->sw_vp); 2470 if (error != 0) 2471 return (error); 2472 #endif 2473 nblks = sp->sw_nblks; 2474 2475 /* 2476 * We can turn off this swap device safely only if the 2477 * available virtual memory in the system will fit the amount 2478 * of data we will have to page back in, plus an epsilon so 2479 * the system doesn't become critically low on swap space. 2480 */ 2481 if (vm_free_count() + swap_pager_avail < nblks + nswap_lowat) 2482 return (ENOMEM); 2483 2484 /* 2485 * Prevent further allocations on this device. 2486 */ 2487 mtx_lock(&sw_dev_mtx); 2488 sp->sw_flags |= SW_CLOSING; 2489 swap_pager_avail -= blist_fill(sp->sw_blist, 0, nblks); 2490 swap_total -= nblks; 2491 mtx_unlock(&sw_dev_mtx); 2492 2493 /* 2494 * Page in the contents of the device and close it. 2495 */ 2496 swap_pager_swapoff(sp); 2497 2498 sp->sw_close(curthread, sp); 2499 mtx_lock(&sw_dev_mtx); 2500 sp->sw_id = NULL; 2501 TAILQ_REMOVE(&swtailq, sp, sw_list); 2502 nswapdev--; 2503 if (nswapdev == 0) { 2504 swap_pager_full = 2; 2505 swap_pager_almost_full = 1; 2506 } 2507 if (swdevhd == sp) 2508 swdevhd = NULL; 2509 mtx_unlock(&sw_dev_mtx); 2510 blist_destroy(sp->sw_blist); 2511 free(sp, M_VMPGDATA); 2512 return (0); 2513 } 2514 2515 void 2516 swapoff_all(void) 2517 { 2518 struct swdevt *sp, *spt; 2519 const char *devname; 2520 int error; 2521 2522 sx_xlock(&swdev_syscall_lock); 2523 2524 mtx_lock(&sw_dev_mtx); 2525 TAILQ_FOREACH_SAFE(sp, &swtailq, sw_list, spt) { 2526 mtx_unlock(&sw_dev_mtx); 2527 if (vn_isdisk(sp->sw_vp, NULL)) 2528 devname = devtoname(sp->sw_vp->v_rdev); 2529 else 2530 devname = "[file]"; 2531 error = swapoff_one(sp, thread0.td_ucred); 2532 if (error != 0) { 2533 printf("Cannot remove swap device %s (error=%d), " 2534 "skipping.\n", devname, error); 2535 } else if (bootverbose) { 2536 printf("Swap device %s removed.\n", devname); 2537 } 2538 mtx_lock(&sw_dev_mtx); 2539 } 2540 mtx_unlock(&sw_dev_mtx); 2541 2542 sx_xunlock(&swdev_syscall_lock); 2543 } 2544 2545 void 2546 swap_pager_status(int *total, int *used) 2547 { 2548 struct swdevt *sp; 2549 2550 *total = 0; 2551 *used = 0; 2552 mtx_lock(&sw_dev_mtx); 2553 TAILQ_FOREACH(sp, &swtailq, sw_list) { 2554 *total += sp->sw_nblks; 2555 *used += sp->sw_used; 2556 } 2557 mtx_unlock(&sw_dev_mtx); 2558 } 2559 2560 int 2561 swap_dev_info(int name, struct xswdev *xs, char *devname, size_t len) 2562 { 2563 struct swdevt *sp; 2564 const char *tmp_devname; 2565 int error, n; 2566 2567 n = 0; 2568 error = ENOENT; 2569 mtx_lock(&sw_dev_mtx); 2570 TAILQ_FOREACH(sp, &swtailq, sw_list) { 2571 if (n != name) { 2572 n++; 2573 continue; 2574 } 2575 xs->xsw_version = XSWDEV_VERSION; 2576 xs->xsw_dev = sp->sw_dev; 2577 xs->xsw_flags = sp->sw_flags; 2578 xs->xsw_nblks = sp->sw_nblks; 2579 xs->xsw_used = sp->sw_used; 2580 if (devname != NULL) { 2581 if (vn_isdisk(sp->sw_vp, NULL)) 2582 tmp_devname = devtoname(sp->sw_vp->v_rdev); 2583 else 2584 tmp_devname = "[file]"; 2585 strncpy(devname, tmp_devname, len); 2586 } 2587 error = 0; 2588 break; 2589 } 2590 mtx_unlock(&sw_dev_mtx); 2591 return (error); 2592 } 2593 2594 #if defined(COMPAT_FREEBSD11) 2595 #define XSWDEV_VERSION_11 1 2596 struct xswdev11 { 2597 u_int xsw_version; 2598 uint32_t xsw_dev; 2599 int xsw_flags; 2600 int xsw_nblks; 2601 int xsw_used; 2602 }; 2603 #endif 2604 2605 #if defined(__amd64__) && defined(COMPAT_FREEBSD32) 2606 struct xswdev32 { 2607 u_int xsw_version; 2608 u_int xsw_dev1, xsw_dev2; 2609 int xsw_flags; 2610 int xsw_nblks; 2611 int xsw_used; 2612 }; 2613 #endif 2614 2615 static int 2616 sysctl_vm_swap_info(SYSCTL_HANDLER_ARGS) 2617 { 2618 struct xswdev xs; 2619 #if defined(__amd64__) && defined(COMPAT_FREEBSD32) 2620 struct xswdev32 xs32; 2621 #endif 2622 #if defined(COMPAT_FREEBSD11) 2623 struct xswdev11 xs11; 2624 #endif 2625 int error; 2626 2627 if (arg2 != 1) /* name length */ 2628 return (EINVAL); 2629 error = swap_dev_info(*(int *)arg1, &xs, NULL, 0); 2630 if (error != 0) 2631 return (error); 2632 #if defined(__amd64__) && defined(COMPAT_FREEBSD32) 2633 if (req->oldlen == sizeof(xs32)) { 2634 xs32.xsw_version = XSWDEV_VERSION; 2635 xs32.xsw_dev1 = xs.xsw_dev; 2636 xs32.xsw_dev2 = xs.xsw_dev >> 32; 2637 xs32.xsw_flags = xs.xsw_flags; 2638 xs32.xsw_nblks = xs.xsw_nblks; 2639 xs32.xsw_used = xs.xsw_used; 2640 error = SYSCTL_OUT(req, &xs32, sizeof(xs32)); 2641 return (error); 2642 } 2643 #endif 2644 #if defined(COMPAT_FREEBSD11) 2645 if (req->oldlen == sizeof(xs11)) { 2646 xs11.xsw_version = XSWDEV_VERSION_11; 2647 xs11.xsw_dev = xs.xsw_dev; /* truncation */ 2648 xs11.xsw_flags = xs.xsw_flags; 2649 xs11.xsw_nblks = xs.xsw_nblks; 2650 xs11.xsw_used = xs.xsw_used; 2651 error = SYSCTL_OUT(req, &xs11, sizeof(xs11)); 2652 return (error); 2653 } 2654 #endif 2655 error = SYSCTL_OUT(req, &xs, sizeof(xs)); 2656 return (error); 2657 } 2658 2659 SYSCTL_INT(_vm, OID_AUTO, nswapdev, CTLFLAG_RD, &nswapdev, 0, 2660 "Number of swap devices"); 2661 SYSCTL_NODE(_vm, OID_AUTO, swap_info, CTLFLAG_RD | CTLFLAG_MPSAFE, 2662 sysctl_vm_swap_info, 2663 "Swap statistics by device"); 2664 2665 /* 2666 * Count the approximate swap usage in pages for a vmspace. The 2667 * shadowed or not yet copied on write swap blocks are not accounted. 2668 * The map must be locked. 2669 */ 2670 long 2671 vmspace_swap_count(struct vmspace *vmspace) 2672 { 2673 vm_map_t map; 2674 vm_map_entry_t cur; 2675 vm_object_t object; 2676 struct swblk *sb; 2677 vm_pindex_t e, pi; 2678 long count; 2679 int i; 2680 2681 map = &vmspace->vm_map; 2682 count = 0; 2683 2684 VM_MAP_ENTRY_FOREACH(cur, map) { 2685 if ((cur->eflags & MAP_ENTRY_IS_SUB_MAP) != 0) 2686 continue; 2687 object = cur->object.vm_object; 2688 if (object == NULL || object->type != OBJT_SWAP) 2689 continue; 2690 VM_OBJECT_RLOCK(object); 2691 if (object->type != OBJT_SWAP) 2692 goto unlock; 2693 pi = OFF_TO_IDX(cur->offset); 2694 e = pi + OFF_TO_IDX(cur->end - cur->start); 2695 for (;; pi = sb->p + SWAP_META_PAGES) { 2696 sb = SWAP_PCTRIE_LOOKUP_GE( 2697 &object->un_pager.swp.swp_blks, pi); 2698 if (sb == NULL || sb->p >= e) 2699 break; 2700 for (i = 0; i < SWAP_META_PAGES; i++) { 2701 if (sb->p + i < e && 2702 sb->d[i] != SWAPBLK_NONE) 2703 count++; 2704 } 2705 } 2706 unlock: 2707 VM_OBJECT_RUNLOCK(object); 2708 } 2709 return (count); 2710 } 2711 2712 /* 2713 * GEOM backend 2714 * 2715 * Swapping onto disk devices. 2716 * 2717 */ 2718 2719 static g_orphan_t swapgeom_orphan; 2720 2721 static struct g_class g_swap_class = { 2722 .name = "SWAP", 2723 .version = G_VERSION, 2724 .orphan = swapgeom_orphan, 2725 }; 2726 2727 DECLARE_GEOM_CLASS(g_swap_class, g_class); 2728 2729 2730 static void 2731 swapgeom_close_ev(void *arg, int flags) 2732 { 2733 struct g_consumer *cp; 2734 2735 cp = arg; 2736 g_access(cp, -1, -1, 0); 2737 g_detach(cp); 2738 g_destroy_consumer(cp); 2739 } 2740 2741 /* 2742 * Add a reference to the g_consumer for an inflight transaction. 2743 */ 2744 static void 2745 swapgeom_acquire(struct g_consumer *cp) 2746 { 2747 2748 mtx_assert(&sw_dev_mtx, MA_OWNED); 2749 cp->index++; 2750 } 2751 2752 /* 2753 * Remove a reference from the g_consumer. Post a close event if all 2754 * references go away, since the function might be called from the 2755 * biodone context. 2756 */ 2757 static void 2758 swapgeom_release(struct g_consumer *cp, struct swdevt *sp) 2759 { 2760 2761 mtx_assert(&sw_dev_mtx, MA_OWNED); 2762 cp->index--; 2763 if (cp->index == 0) { 2764 if (g_post_event(swapgeom_close_ev, cp, M_NOWAIT, NULL) == 0) 2765 sp->sw_id = NULL; 2766 } 2767 } 2768 2769 static void 2770 swapgeom_done(struct bio *bp2) 2771 { 2772 struct swdevt *sp; 2773 struct buf *bp; 2774 struct g_consumer *cp; 2775 2776 bp = bp2->bio_caller2; 2777 cp = bp2->bio_from; 2778 bp->b_ioflags = bp2->bio_flags; 2779 if (bp2->bio_error) 2780 bp->b_ioflags |= BIO_ERROR; 2781 bp->b_resid = bp->b_bcount - bp2->bio_completed; 2782 bp->b_error = bp2->bio_error; 2783 bp->b_caller1 = NULL; 2784 bufdone(bp); 2785 sp = bp2->bio_caller1; 2786 mtx_lock(&sw_dev_mtx); 2787 swapgeom_release(cp, sp); 2788 mtx_unlock(&sw_dev_mtx); 2789 g_destroy_bio(bp2); 2790 } 2791 2792 static void 2793 swapgeom_strategy(struct buf *bp, struct swdevt *sp) 2794 { 2795 struct bio *bio; 2796 struct g_consumer *cp; 2797 2798 mtx_lock(&sw_dev_mtx); 2799 cp = sp->sw_id; 2800 if (cp == NULL) { 2801 mtx_unlock(&sw_dev_mtx); 2802 bp->b_error = ENXIO; 2803 bp->b_ioflags |= BIO_ERROR; 2804 bufdone(bp); 2805 return; 2806 } 2807 swapgeom_acquire(cp); 2808 mtx_unlock(&sw_dev_mtx); 2809 if (bp->b_iocmd == BIO_WRITE) 2810 bio = g_new_bio(); 2811 else 2812 bio = g_alloc_bio(); 2813 if (bio == NULL) { 2814 mtx_lock(&sw_dev_mtx); 2815 swapgeom_release(cp, sp); 2816 mtx_unlock(&sw_dev_mtx); 2817 bp->b_error = ENOMEM; 2818 bp->b_ioflags |= BIO_ERROR; 2819 printf("swap_pager: cannot allocate bio\n"); 2820 bufdone(bp); 2821 return; 2822 } 2823 2824 bp->b_caller1 = bio; 2825 bio->bio_caller1 = sp; 2826 bio->bio_caller2 = bp; 2827 bio->bio_cmd = bp->b_iocmd; 2828 bio->bio_offset = (bp->b_blkno - sp->sw_first) * PAGE_SIZE; 2829 bio->bio_length = bp->b_bcount; 2830 bio->bio_done = swapgeom_done; 2831 if (!buf_mapped(bp)) { 2832 bio->bio_ma = bp->b_pages; 2833 bio->bio_data = unmapped_buf; 2834 bio->bio_ma_offset = (vm_offset_t)bp->b_offset & PAGE_MASK; 2835 bio->bio_ma_n = bp->b_npages; 2836 bio->bio_flags |= BIO_UNMAPPED; 2837 } else { 2838 bio->bio_data = bp->b_data; 2839 bio->bio_ma = NULL; 2840 } 2841 g_io_request(bio, cp); 2842 return; 2843 } 2844 2845 static void 2846 swapgeom_orphan(struct g_consumer *cp) 2847 { 2848 struct swdevt *sp; 2849 int destroy; 2850 2851 mtx_lock(&sw_dev_mtx); 2852 TAILQ_FOREACH(sp, &swtailq, sw_list) { 2853 if (sp->sw_id == cp) { 2854 sp->sw_flags |= SW_CLOSING; 2855 break; 2856 } 2857 } 2858 /* 2859 * Drop reference we were created with. Do directly since we're in a 2860 * special context where we don't have to queue the call to 2861 * swapgeom_close_ev(). 2862 */ 2863 cp->index--; 2864 destroy = ((sp != NULL) && (cp->index == 0)); 2865 if (destroy) 2866 sp->sw_id = NULL; 2867 mtx_unlock(&sw_dev_mtx); 2868 if (destroy) 2869 swapgeom_close_ev(cp, 0); 2870 } 2871 2872 static void 2873 swapgeom_close(struct thread *td, struct swdevt *sw) 2874 { 2875 struct g_consumer *cp; 2876 2877 mtx_lock(&sw_dev_mtx); 2878 cp = sw->sw_id; 2879 sw->sw_id = NULL; 2880 mtx_unlock(&sw_dev_mtx); 2881 2882 /* 2883 * swapgeom_close() may be called from the biodone context, 2884 * where we cannot perform topology changes. Delegate the 2885 * work to the events thread. 2886 */ 2887 if (cp != NULL) 2888 g_waitfor_event(swapgeom_close_ev, cp, M_WAITOK, NULL); 2889 } 2890 2891 static int 2892 swapongeom_locked(struct cdev *dev, struct vnode *vp) 2893 { 2894 struct g_provider *pp; 2895 struct g_consumer *cp; 2896 static struct g_geom *gp; 2897 struct swdevt *sp; 2898 u_long nblks; 2899 int error; 2900 2901 pp = g_dev_getprovider(dev); 2902 if (pp == NULL) 2903 return (ENODEV); 2904 mtx_lock(&sw_dev_mtx); 2905 TAILQ_FOREACH(sp, &swtailq, sw_list) { 2906 cp = sp->sw_id; 2907 if (cp != NULL && cp->provider == pp) { 2908 mtx_unlock(&sw_dev_mtx); 2909 return (EBUSY); 2910 } 2911 } 2912 mtx_unlock(&sw_dev_mtx); 2913 if (gp == NULL) 2914 gp = g_new_geomf(&g_swap_class, "swap"); 2915 cp = g_new_consumer(gp); 2916 cp->index = 1; /* Number of active I/Os, plus one for being active. */ 2917 cp->flags |= G_CF_DIRECT_SEND | G_CF_DIRECT_RECEIVE; 2918 g_attach(cp, pp); 2919 /* 2920 * XXX: Every time you think you can improve the margin for 2921 * footshooting, somebody depends on the ability to do so: 2922 * savecore(8) wants to write to our swapdev so we cannot 2923 * set an exclusive count :-( 2924 */ 2925 error = g_access(cp, 1, 1, 0); 2926 if (error != 0) { 2927 g_detach(cp); 2928 g_destroy_consumer(cp); 2929 return (error); 2930 } 2931 nblks = pp->mediasize / DEV_BSIZE; 2932 swaponsomething(vp, cp, nblks, swapgeom_strategy, 2933 swapgeom_close, dev2udev(dev), 2934 (pp->flags & G_PF_ACCEPT_UNMAPPED) != 0 ? SW_UNMAPPED : 0); 2935 return (0); 2936 } 2937 2938 static int 2939 swapongeom(struct vnode *vp) 2940 { 2941 int error; 2942 2943 vn_lock(vp, LK_EXCLUSIVE | LK_RETRY); 2944 if (vp->v_type != VCHR || VN_IS_DOOMED(vp)) { 2945 error = ENOENT; 2946 } else { 2947 g_topology_lock(); 2948 error = swapongeom_locked(vp->v_rdev, vp); 2949 g_topology_unlock(); 2950 } 2951 VOP_UNLOCK(vp); 2952 return (error); 2953 } 2954 2955 /* 2956 * VNODE backend 2957 * 2958 * This is used mainly for network filesystem (read: probably only tested 2959 * with NFS) swapfiles. 2960 * 2961 */ 2962 2963 static void 2964 swapdev_strategy(struct buf *bp, struct swdevt *sp) 2965 { 2966 struct vnode *vp2; 2967 2968 bp->b_blkno = ctodb(bp->b_blkno - sp->sw_first); 2969 2970 vp2 = sp->sw_id; 2971 vhold(vp2); 2972 if (bp->b_iocmd == BIO_WRITE) { 2973 if (bp->b_bufobj) 2974 bufobj_wdrop(bp->b_bufobj); 2975 bufobj_wref(&vp2->v_bufobj); 2976 } 2977 if (bp->b_bufobj != &vp2->v_bufobj) 2978 bp->b_bufobj = &vp2->v_bufobj; 2979 bp->b_vp = vp2; 2980 bp->b_iooffset = dbtob(bp->b_blkno); 2981 bstrategy(bp); 2982 return; 2983 } 2984 2985 static void 2986 swapdev_close(struct thread *td, struct swdevt *sp) 2987 { 2988 2989 VOP_CLOSE(sp->sw_vp, FREAD | FWRITE, td->td_ucred, td); 2990 vrele(sp->sw_vp); 2991 } 2992 2993 2994 static int 2995 swaponvp(struct thread *td, struct vnode *vp, u_long nblks) 2996 { 2997 struct swdevt *sp; 2998 int error; 2999 3000 if (nblks == 0) 3001 return (ENXIO); 3002 mtx_lock(&sw_dev_mtx); 3003 TAILQ_FOREACH(sp, &swtailq, sw_list) { 3004 if (sp->sw_id == vp) { 3005 mtx_unlock(&sw_dev_mtx); 3006 return (EBUSY); 3007 } 3008 } 3009 mtx_unlock(&sw_dev_mtx); 3010 3011 (void) vn_lock(vp, LK_EXCLUSIVE | LK_RETRY); 3012 #ifdef MAC 3013 error = mac_system_check_swapon(td->td_ucred, vp); 3014 if (error == 0) 3015 #endif 3016 error = VOP_OPEN(vp, FREAD | FWRITE, td->td_ucred, td, NULL); 3017 (void) VOP_UNLOCK(vp); 3018 if (error) 3019 return (error); 3020 3021 swaponsomething(vp, vp, nblks, swapdev_strategy, swapdev_close, 3022 NODEV, 0); 3023 return (0); 3024 } 3025 3026 static int 3027 sysctl_swap_async_max(SYSCTL_HANDLER_ARGS) 3028 { 3029 int error, new, n; 3030 3031 new = nsw_wcount_async_max; 3032 error = sysctl_handle_int(oidp, &new, 0, req); 3033 if (error != 0 || req->newptr == NULL) 3034 return (error); 3035 3036 if (new > nswbuf / 2 || new < 1) 3037 return (EINVAL); 3038 3039 mtx_lock(&swbuf_mtx); 3040 while (nsw_wcount_async_max != new) { 3041 /* 3042 * Adjust difference. If the current async count is too low, 3043 * we will need to sqeeze our update slowly in. Sleep with a 3044 * higher priority than getpbuf() to finish faster. 3045 */ 3046 n = new - nsw_wcount_async_max; 3047 if (nsw_wcount_async + n >= 0) { 3048 nsw_wcount_async += n; 3049 nsw_wcount_async_max += n; 3050 wakeup(&nsw_wcount_async); 3051 } else { 3052 nsw_wcount_async_max -= nsw_wcount_async; 3053 nsw_wcount_async = 0; 3054 msleep(&nsw_wcount_async, &swbuf_mtx, PSWP, 3055 "swpsysctl", 0); 3056 } 3057 } 3058 mtx_unlock(&swbuf_mtx); 3059 3060 return (0); 3061 } 3062 3063 static void 3064 swap_pager_update_writecount(vm_object_t object, vm_offset_t start, 3065 vm_offset_t end) 3066 { 3067 3068 VM_OBJECT_WLOCK(object); 3069 KASSERT((object->flags & OBJ_ANON) == 0, 3070 ("Splittable object with writecount")); 3071 object->un_pager.swp.writemappings += (vm_ooffset_t)end - start; 3072 VM_OBJECT_WUNLOCK(object); 3073 } 3074 3075 static void 3076 swap_pager_release_writecount(vm_object_t object, vm_offset_t start, 3077 vm_offset_t end) 3078 { 3079 3080 VM_OBJECT_WLOCK(object); 3081 KASSERT((object->flags & OBJ_ANON) == 0, 3082 ("Splittable object with writecount")); 3083 object->un_pager.swp.writemappings -= (vm_ooffset_t)end - start; 3084 VM_OBJECT_WUNLOCK(object); 3085 } 3086