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