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