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_swap.h" 75 #include "opt_vm.h" 76 77 #include <sys/param.h> 78 #include <sys/systm.h> 79 #include <sys/conf.h> 80 #include <sys/kernel.h> 81 #include <sys/priv.h> 82 #include <sys/proc.h> 83 #include <sys/bio.h> 84 #include <sys/buf.h> 85 #include <sys/disk.h> 86 #include <sys/fcntl.h> 87 #include <sys/mount.h> 88 #include <sys/namei.h> 89 #include <sys/vnode.h> 90 #include <sys/malloc.h> 91 #include <sys/pctrie.h> 92 #include <sys/racct.h> 93 #include <sys/resource.h> 94 #include <sys/resourcevar.h> 95 #include <sys/rwlock.h> 96 #include <sys/sbuf.h> 97 #include <sys/sysctl.h> 98 #include <sys/sysproto.h> 99 #include <sys/blist.h> 100 #include <sys/lock.h> 101 #include <sys/sx.h> 102 #include <sys/vmmeter.h> 103 104 #include <security/mac/mac_framework.h> 105 106 #include <vm/vm.h> 107 #include <vm/pmap.h> 108 #include <vm/vm_map.h> 109 #include <vm/vm_kern.h> 110 #include <vm/vm_object.h> 111 #include <vm/vm_page.h> 112 #include <vm/vm_pager.h> 113 #include <vm/vm_pageout.h> 114 #include <vm/vm_param.h> 115 #include <vm/swap_pager.h> 116 #include <vm/vm_extern.h> 117 #include <vm/uma.h> 118 119 #include <geom/geom.h> 120 121 /* 122 * MAX_PAGEOUT_CLUSTER must be a power of 2 between 1 and 64. 123 * The 64-page limit is due to the radix code (kern/subr_blist.c). 124 */ 125 #ifndef MAX_PAGEOUT_CLUSTER 126 #define MAX_PAGEOUT_CLUSTER 32 127 #endif 128 129 #if !defined(SWB_NPAGES) 130 #define SWB_NPAGES MAX_PAGEOUT_CLUSTER 131 #endif 132 133 #define SWAP_META_PAGES PCTRIE_COUNT 134 135 /* 136 * A swblk structure maps each page index within a 137 * SWAP_META_PAGES-aligned and sized range to the address of an 138 * on-disk swap block (or SWAPBLK_NONE). The collection of these 139 * mappings for an entire vm object is implemented as a pc-trie. 140 */ 141 struct swblk { 142 vm_pindex_t p; 143 daddr_t d[SWAP_META_PAGES]; 144 }; 145 146 static MALLOC_DEFINE(M_VMPGDATA, "vm_pgdata", "swap pager private data"); 147 static struct mtx sw_dev_mtx; 148 static TAILQ_HEAD(, swdevt) swtailq = TAILQ_HEAD_INITIALIZER(swtailq); 149 static struct swdevt *swdevhd; /* Allocate from here next */ 150 static int nswapdev; /* Number of swap devices */ 151 int swap_pager_avail; 152 static struct sx swdev_syscall_lock; /* serialize swap(on|off) */ 153 154 static vm_ooffset_t swap_total; 155 SYSCTL_QUAD(_vm, OID_AUTO, swap_total, CTLFLAG_RD, &swap_total, 0, 156 "Total amount of available swap storage."); 157 static vm_ooffset_t swap_reserved; 158 SYSCTL_QUAD(_vm, OID_AUTO, swap_reserved, CTLFLAG_RD, &swap_reserved, 0, 159 "Amount of swap storage needed to back all allocated anonymous memory."); 160 static int overcommit = 0; 161 SYSCTL_INT(_vm, VM_OVERCOMMIT, overcommit, CTLFLAG_RW, &overcommit, 0, 162 "Configure virtual memory overcommit behavior. See tuning(7) " 163 "for details."); 164 static unsigned long swzone; 165 SYSCTL_ULONG(_vm, OID_AUTO, swzone, CTLFLAG_RD, &swzone, 0, 166 "Actual size of swap metadata zone"); 167 static unsigned long swap_maxpages; 168 SYSCTL_ULONG(_vm, OID_AUTO, swap_maxpages, CTLFLAG_RD, &swap_maxpages, 0, 169 "Maximum amount of swap supported"); 170 171 /* bits from overcommit */ 172 #define SWAP_RESERVE_FORCE_ON (1 << 0) 173 #define SWAP_RESERVE_RLIMIT_ON (1 << 1) 174 #define SWAP_RESERVE_ALLOW_NONWIRED (1 << 2) 175 176 int 177 swap_reserve(vm_ooffset_t incr) 178 { 179 180 return (swap_reserve_by_cred(incr, curthread->td_ucred)); 181 } 182 183 int 184 swap_reserve_by_cred(vm_ooffset_t incr, struct ucred *cred) 185 { 186 vm_ooffset_t r, s; 187 int res, error; 188 static int curfail; 189 static struct timeval lastfail; 190 struct uidinfo *uip; 191 192 uip = cred->cr_ruidinfo; 193 194 if (incr & PAGE_MASK) 195 panic("swap_reserve: & PAGE_MASK"); 196 197 #ifdef RACCT 198 if (racct_enable) { 199 PROC_LOCK(curproc); 200 error = racct_add(curproc, RACCT_SWAP, incr); 201 PROC_UNLOCK(curproc); 202 if (error != 0) 203 return (0); 204 } 205 #endif 206 207 res = 0; 208 mtx_lock(&sw_dev_mtx); 209 r = swap_reserved + incr; 210 if (overcommit & SWAP_RESERVE_ALLOW_NONWIRED) { 211 s = vm_cnt.v_page_count - vm_cnt.v_free_reserved - 212 vm_wire_count(); 213 s *= PAGE_SIZE; 214 } else 215 s = 0; 216 s += swap_total; 217 if ((overcommit & SWAP_RESERVE_FORCE_ON) == 0 || r <= s || 218 (error = priv_check(curthread, PRIV_VM_SWAP_NOQUOTA)) == 0) { 219 res = 1; 220 swap_reserved = r; 221 } 222 mtx_unlock(&sw_dev_mtx); 223 224 if (res) { 225 UIDINFO_VMSIZE_LOCK(uip); 226 if ((overcommit & SWAP_RESERVE_RLIMIT_ON) != 0 && 227 uip->ui_vmsize + incr > lim_cur(curthread, RLIMIT_SWAP) && 228 priv_check(curthread, PRIV_VM_SWAP_NORLIMIT)) 229 res = 0; 230 else 231 uip->ui_vmsize += incr; 232 UIDINFO_VMSIZE_UNLOCK(uip); 233 if (!res) { 234 mtx_lock(&sw_dev_mtx); 235 swap_reserved -= incr; 236 mtx_unlock(&sw_dev_mtx); 237 } 238 } 239 if (!res && ppsratecheck(&lastfail, &curfail, 1)) { 240 printf("uid %d, pid %d: swap reservation for %jd bytes failed\n", 241 uip->ui_uid, curproc->p_pid, incr); 242 } 243 244 #ifdef RACCT 245 if (!res) { 246 PROC_LOCK(curproc); 247 racct_sub(curproc, RACCT_SWAP, incr); 248 PROC_UNLOCK(curproc); 249 } 250 #endif 251 252 return (res); 253 } 254 255 void 256 swap_reserve_force(vm_ooffset_t incr) 257 { 258 struct uidinfo *uip; 259 260 mtx_lock(&sw_dev_mtx); 261 swap_reserved += incr; 262 mtx_unlock(&sw_dev_mtx); 263 264 #ifdef RACCT 265 PROC_LOCK(curproc); 266 racct_add_force(curproc, RACCT_SWAP, incr); 267 PROC_UNLOCK(curproc); 268 #endif 269 270 uip = curthread->td_ucred->cr_ruidinfo; 271 PROC_LOCK(curproc); 272 UIDINFO_VMSIZE_LOCK(uip); 273 uip->ui_vmsize += incr; 274 UIDINFO_VMSIZE_UNLOCK(uip); 275 PROC_UNLOCK(curproc); 276 } 277 278 void 279 swap_release(vm_ooffset_t decr) 280 { 281 struct ucred *cred; 282 283 PROC_LOCK(curproc); 284 cred = curthread->td_ucred; 285 swap_release_by_cred(decr, cred); 286 PROC_UNLOCK(curproc); 287 } 288 289 void 290 swap_release_by_cred(vm_ooffset_t decr, struct ucred *cred) 291 { 292 struct uidinfo *uip; 293 294 uip = cred->cr_ruidinfo; 295 296 if (decr & PAGE_MASK) 297 panic("swap_release: & PAGE_MASK"); 298 299 mtx_lock(&sw_dev_mtx); 300 if (swap_reserved < decr) 301 panic("swap_reserved < decr"); 302 swap_reserved -= decr; 303 mtx_unlock(&sw_dev_mtx); 304 305 UIDINFO_VMSIZE_LOCK(uip); 306 if (uip->ui_vmsize < decr) 307 printf("negative vmsize for uid = %d\n", uip->ui_uid); 308 uip->ui_vmsize -= decr; 309 UIDINFO_VMSIZE_UNLOCK(uip); 310 311 racct_sub_cred(cred, RACCT_SWAP, decr); 312 } 313 314 #define SWM_POP 0x01 /* pop out */ 315 316 static int swap_pager_full = 2; /* swap space exhaustion (task killing) */ 317 static int swap_pager_almost_full = 1; /* swap space exhaustion (w/hysteresis)*/ 318 static int nsw_rcount; /* free read buffers */ 319 static int nsw_wcount_sync; /* limit write buffers / synchronous */ 320 static int nsw_wcount_async; /* limit write buffers / asynchronous */ 321 static int nsw_wcount_async_max;/* assigned maximum */ 322 static int nsw_cluster_max; /* maximum VOP I/O allowed */ 323 324 static int sysctl_swap_async_max(SYSCTL_HANDLER_ARGS); 325 SYSCTL_PROC(_vm, OID_AUTO, swap_async_max, CTLTYPE_INT | CTLFLAG_RW | 326 CTLFLAG_MPSAFE, NULL, 0, sysctl_swap_async_max, "I", 327 "Maximum running async swap ops"); 328 static int sysctl_swap_fragmentation(SYSCTL_HANDLER_ARGS); 329 SYSCTL_PROC(_vm, OID_AUTO, swap_fragmentation, CTLTYPE_STRING | CTLFLAG_RD | 330 CTLFLAG_MPSAFE, NULL, 0, sysctl_swap_fragmentation, "A", 331 "Swap Fragmentation Info"); 332 333 static struct sx sw_alloc_sx; 334 335 /* 336 * "named" and "unnamed" anon region objects. Try to reduce the overhead 337 * of searching a named list by hashing it just a little. 338 */ 339 340 #define NOBJLISTS 8 341 342 #define NOBJLIST(handle) \ 343 (&swap_pager_object_list[((int)(intptr_t)handle >> 4) & (NOBJLISTS-1)]) 344 345 static struct pagerlst swap_pager_object_list[NOBJLISTS]; 346 static uma_zone_t swblk_zone; 347 static uma_zone_t swpctrie_zone; 348 349 /* 350 * pagerops for OBJT_SWAP - "swap pager". Some ops are also global procedure 351 * calls hooked from other parts of the VM system and do not appear here. 352 * (see vm/swap_pager.h). 353 */ 354 static vm_object_t 355 swap_pager_alloc(void *handle, vm_ooffset_t size, 356 vm_prot_t prot, vm_ooffset_t offset, struct ucred *); 357 static void swap_pager_dealloc(vm_object_t object); 358 static int swap_pager_getpages(vm_object_t, vm_page_t *, int, int *, 359 int *); 360 static int swap_pager_getpages_async(vm_object_t, vm_page_t *, int, int *, 361 int *, pgo_getpages_iodone_t, void *); 362 static void swap_pager_putpages(vm_object_t, vm_page_t *, int, boolean_t, int *); 363 static boolean_t 364 swap_pager_haspage(vm_object_t object, vm_pindex_t pindex, int *before, int *after); 365 static void swap_pager_init(void); 366 static void swap_pager_unswapped(vm_page_t); 367 static void swap_pager_swapoff(struct swdevt *sp); 368 369 struct pagerops swappagerops = { 370 .pgo_init = swap_pager_init, /* early system initialization of pager */ 371 .pgo_alloc = swap_pager_alloc, /* allocate an OBJT_SWAP object */ 372 .pgo_dealloc = swap_pager_dealloc, /* deallocate an OBJT_SWAP object */ 373 .pgo_getpages = swap_pager_getpages, /* pagein */ 374 .pgo_getpages_async = swap_pager_getpages_async, /* pagein (async) */ 375 .pgo_putpages = swap_pager_putpages, /* pageout */ 376 .pgo_haspage = swap_pager_haspage, /* get backing store status for page */ 377 .pgo_pageunswapped = swap_pager_unswapped, /* remove swap related to page */ 378 }; 379 380 /* 381 * swap_*() routines are externally accessible. swp_*() routines are 382 * internal. 383 */ 384 static int nswap_lowat = 128; /* in pages, swap_pager_almost_full warn */ 385 static int nswap_hiwat = 512; /* in pages, swap_pager_almost_full warn */ 386 387 SYSCTL_INT(_vm, OID_AUTO, dmmax, CTLFLAG_RD, &nsw_cluster_max, 0, 388 "Maximum size of a swap block in pages"); 389 390 static void swp_sizecheck(void); 391 static void swp_pager_async_iodone(struct buf *bp); 392 static bool swp_pager_swblk_empty(struct swblk *sb, int start, int limit); 393 static int swapongeom(struct vnode *); 394 static int swaponvp(struct thread *, struct vnode *, u_long); 395 static int swapoff_one(struct swdevt *sp, struct ucred *cred); 396 397 /* 398 * Swap bitmap functions 399 */ 400 static void swp_pager_freeswapspace(daddr_t blk, daddr_t npages); 401 static daddr_t swp_pager_getswapspace(int npages); 402 403 /* 404 * Metadata functions 405 */ 406 static void swp_pager_meta_build(vm_object_t, vm_pindex_t, daddr_t); 407 static void swp_pager_meta_free(vm_object_t, vm_pindex_t, vm_pindex_t); 408 static void swp_pager_meta_free_all(vm_object_t); 409 static daddr_t swp_pager_meta_ctl(vm_object_t, vm_pindex_t, int); 410 411 static void * 412 swblk_trie_alloc(struct pctrie *ptree) 413 { 414 415 return (uma_zalloc(swpctrie_zone, M_NOWAIT | (curproc == pageproc ? 416 M_USE_RESERVE : 0))); 417 } 418 419 static void 420 swblk_trie_free(struct pctrie *ptree, void *node) 421 { 422 423 uma_zfree(swpctrie_zone, node); 424 } 425 426 PCTRIE_DEFINE(SWAP, swblk, p, swblk_trie_alloc, swblk_trie_free); 427 428 /* 429 * SWP_SIZECHECK() - update swap_pager_full indication 430 * 431 * update the swap_pager_almost_full indication and warn when we are 432 * about to run out of swap space, using lowat/hiwat hysteresis. 433 * 434 * Clear swap_pager_full ( task killing ) indication when lowat is met. 435 * 436 * No restrictions on call 437 * This routine may not block. 438 */ 439 static void 440 swp_sizecheck(void) 441 { 442 443 if (swap_pager_avail < nswap_lowat) { 444 if (swap_pager_almost_full == 0) { 445 printf("swap_pager: out of swap space\n"); 446 swap_pager_almost_full = 1; 447 } 448 } else { 449 swap_pager_full = 0; 450 if (swap_pager_avail > nswap_hiwat) 451 swap_pager_almost_full = 0; 452 } 453 } 454 455 /* 456 * SWAP_PAGER_INIT() - initialize the swap pager! 457 * 458 * Expected to be started from system init. NOTE: This code is run 459 * before much else so be careful what you depend on. Most of the VM 460 * system has yet to be initialized at this point. 461 */ 462 static void 463 swap_pager_init(void) 464 { 465 /* 466 * Initialize object lists 467 */ 468 int i; 469 470 for (i = 0; i < NOBJLISTS; ++i) 471 TAILQ_INIT(&swap_pager_object_list[i]); 472 mtx_init(&sw_dev_mtx, "swapdev", NULL, MTX_DEF); 473 sx_init(&sw_alloc_sx, "swspsx"); 474 sx_init(&swdev_syscall_lock, "swsysc"); 475 } 476 477 /* 478 * SWAP_PAGER_SWAP_INIT() - swap pager initialization from pageout process 479 * 480 * Expected to be started from pageout process once, prior to entering 481 * its main loop. 482 */ 483 void 484 swap_pager_swap_init(void) 485 { 486 unsigned long n, n2; 487 488 /* 489 * Number of in-transit swap bp operations. Don't 490 * exhaust the pbufs completely. Make sure we 491 * initialize workable values (0 will work for hysteresis 492 * but it isn't very efficient). 493 * 494 * The nsw_cluster_max is constrained by the bp->b_pages[] 495 * array (MAXPHYS/PAGE_SIZE) and our locally defined 496 * MAX_PAGEOUT_CLUSTER. Also be aware that swap ops are 497 * constrained by the swap device interleave stripe size. 498 * 499 * Currently we hardwire nsw_wcount_async to 4. This limit is 500 * designed to prevent other I/O from having high latencies due to 501 * our pageout I/O. The value 4 works well for one or two active swap 502 * devices but is probably a little low if you have more. Even so, 503 * a higher value would probably generate only a limited improvement 504 * with three or four active swap devices since the system does not 505 * typically have to pageout at extreme bandwidths. We will want 506 * at least 2 per swap devices, and 4 is a pretty good value if you 507 * have one NFS swap device due to the command/ack latency over NFS. 508 * So it all works out pretty well. 509 */ 510 nsw_cluster_max = min((MAXPHYS/PAGE_SIZE), MAX_PAGEOUT_CLUSTER); 511 512 mtx_lock(&pbuf_mtx); 513 nsw_rcount = (nswbuf + 1) / 2; 514 nsw_wcount_sync = (nswbuf + 3) / 4; 515 nsw_wcount_async = 4; 516 nsw_wcount_async_max = nsw_wcount_async; 517 mtx_unlock(&pbuf_mtx); 518 519 /* 520 * Initialize our zone, guessing on the number we need based 521 * on the number of pages in the system. 522 */ 523 n = vm_cnt.v_page_count / 2; 524 if (maxswzone && n > maxswzone / sizeof(struct swblk)) 525 n = maxswzone / sizeof(struct swblk); 526 swpctrie_zone = uma_zcreate("swpctrie", pctrie_node_size(), NULL, NULL, 527 pctrie_zone_init, NULL, UMA_ALIGN_PTR, 528 UMA_ZONE_NOFREE | UMA_ZONE_VM); 529 if (swpctrie_zone == NULL) 530 panic("failed to create swap pctrie zone."); 531 swblk_zone = uma_zcreate("swblk", sizeof(struct swblk), NULL, NULL, 532 NULL, NULL, _Alignof(struct swblk) - 1, 533 UMA_ZONE_NOFREE | UMA_ZONE_VM); 534 if (swblk_zone == NULL) 535 panic("failed to create swap blk zone."); 536 n2 = n; 537 do { 538 if (uma_zone_reserve_kva(swblk_zone, n)) 539 break; 540 /* 541 * if the allocation failed, try a zone two thirds the 542 * size of the previous attempt. 543 */ 544 n -= ((n + 2) / 3); 545 } while (n > 0); 546 547 /* 548 * Often uma_zone_reserve_kva() cannot reserve exactly the 549 * requested size. Account for the difference when 550 * calculating swap_maxpages. 551 */ 552 n = uma_zone_get_max(swblk_zone); 553 554 if (n < n2) 555 printf("Swap blk zone entries reduced from %lu to %lu.\n", 556 n2, n); 557 swap_maxpages = n * SWAP_META_PAGES; 558 swzone = n * sizeof(struct swblk); 559 if (!uma_zone_reserve_kva(swpctrie_zone, n)) 560 printf("Cannot reserve swap pctrie zone, " 561 "reduce kern.maxswzone.\n"); 562 } 563 564 static vm_object_t 565 swap_pager_alloc_init(void *handle, struct ucred *cred, vm_ooffset_t size, 566 vm_ooffset_t offset) 567 { 568 vm_object_t object; 569 570 if (cred != NULL) { 571 if (!swap_reserve_by_cred(size, cred)) 572 return (NULL); 573 crhold(cred); 574 } 575 576 /* 577 * The un_pager.swp.swp_blks trie is initialized by 578 * vm_object_allocate() to ensure the correct order of 579 * visibility to other threads. 580 */ 581 object = vm_object_allocate(OBJT_SWAP, OFF_TO_IDX(offset + 582 PAGE_MASK + size)); 583 584 object->handle = handle; 585 if (cred != NULL) { 586 object->cred = cred; 587 object->charge = size; 588 } 589 return (object); 590 } 591 592 /* 593 * SWAP_PAGER_ALLOC() - allocate a new OBJT_SWAP VM object and instantiate 594 * its metadata structures. 595 * 596 * This routine is called from the mmap and fork code to create a new 597 * OBJT_SWAP object. 598 * 599 * This routine must ensure that no live duplicate is created for 600 * the named object request, which is protected against by 601 * holding the sw_alloc_sx lock in case handle != NULL. 602 */ 603 static vm_object_t 604 swap_pager_alloc(void *handle, vm_ooffset_t size, vm_prot_t prot, 605 vm_ooffset_t offset, struct ucred *cred) 606 { 607 vm_object_t object; 608 609 if (handle != NULL) { 610 /* 611 * Reference existing named region or allocate new one. There 612 * should not be a race here against swp_pager_meta_build() 613 * as called from vm_page_remove() in regards to the lookup 614 * of the handle. 615 */ 616 sx_xlock(&sw_alloc_sx); 617 object = vm_pager_object_lookup(NOBJLIST(handle), handle); 618 if (object == NULL) { 619 object = swap_pager_alloc_init(handle, cred, size, 620 offset); 621 if (object != NULL) { 622 TAILQ_INSERT_TAIL(NOBJLIST(object->handle), 623 object, pager_object_list); 624 } 625 } 626 sx_xunlock(&sw_alloc_sx); 627 } else { 628 object = swap_pager_alloc_init(handle, cred, size, offset); 629 } 630 return (object); 631 } 632 633 /* 634 * SWAP_PAGER_DEALLOC() - remove swap metadata from object 635 * 636 * The swap backing for the object is destroyed. The code is 637 * designed such that we can reinstantiate it later, but this 638 * routine is typically called only when the entire object is 639 * about to be destroyed. 640 * 641 * The object must be locked. 642 */ 643 static void 644 swap_pager_dealloc(vm_object_t object) 645 { 646 647 VM_OBJECT_ASSERT_WLOCKED(object); 648 KASSERT((object->flags & OBJ_DEAD) != 0, ("dealloc of reachable obj")); 649 650 /* 651 * Remove from list right away so lookups will fail if we block for 652 * pageout completion. 653 */ 654 if (object->handle != NULL) { 655 VM_OBJECT_WUNLOCK(object); 656 sx_xlock(&sw_alloc_sx); 657 TAILQ_REMOVE(NOBJLIST(object->handle), object, 658 pager_object_list); 659 sx_xunlock(&sw_alloc_sx); 660 VM_OBJECT_WLOCK(object); 661 } 662 663 vm_object_pip_wait(object, "swpdea"); 664 665 /* 666 * Free all remaining metadata. We only bother to free it from 667 * the swap meta data. We do not attempt to free swapblk's still 668 * associated with vm_page_t's for this object. We do not care 669 * if paging is still in progress on some objects. 670 */ 671 swp_pager_meta_free_all(object); 672 object->handle = NULL; 673 object->type = OBJT_DEAD; 674 } 675 676 /************************************************************************ 677 * SWAP PAGER BITMAP ROUTINES * 678 ************************************************************************/ 679 680 /* 681 * SWP_PAGER_GETSWAPSPACE() - allocate raw swap space 682 * 683 * Allocate swap for the requested number of pages. The starting 684 * swap block number (a page index) is returned or SWAPBLK_NONE 685 * if the allocation failed. 686 * 687 * Also has the side effect of advising that somebody made a mistake 688 * when they configured swap and didn't configure enough. 689 * 690 * This routine may not sleep. 691 * 692 * We allocate in round-robin fashion from the configured devices. 693 */ 694 static daddr_t 695 swp_pager_getswapspace(int npages) 696 { 697 daddr_t blk; 698 struct swdevt *sp; 699 int i; 700 701 blk = SWAPBLK_NONE; 702 mtx_lock(&sw_dev_mtx); 703 sp = swdevhd; 704 for (i = 0; i < nswapdev; i++) { 705 if (sp == NULL) 706 sp = TAILQ_FIRST(&swtailq); 707 if (!(sp->sw_flags & SW_CLOSING)) { 708 blk = blist_alloc(sp->sw_blist, npages); 709 if (blk != SWAPBLK_NONE) { 710 blk += sp->sw_first; 711 sp->sw_used += npages; 712 swap_pager_avail -= npages; 713 swp_sizecheck(); 714 swdevhd = TAILQ_NEXT(sp, sw_list); 715 goto done; 716 } 717 } 718 sp = TAILQ_NEXT(sp, sw_list); 719 } 720 if (swap_pager_full != 2) { 721 printf("swap_pager_getswapspace(%d): failed\n", npages); 722 swap_pager_full = 2; 723 swap_pager_almost_full = 1; 724 } 725 swdevhd = NULL; 726 done: 727 mtx_unlock(&sw_dev_mtx); 728 return (blk); 729 } 730 731 static int 732 swp_pager_isondev(daddr_t blk, struct swdevt *sp) 733 { 734 735 return (blk >= sp->sw_first && blk < sp->sw_end); 736 } 737 738 static void 739 swp_pager_strategy(struct buf *bp) 740 { 741 struct swdevt *sp; 742 743 mtx_lock(&sw_dev_mtx); 744 TAILQ_FOREACH(sp, &swtailq, sw_list) { 745 if (bp->b_blkno >= sp->sw_first && bp->b_blkno < sp->sw_end) { 746 mtx_unlock(&sw_dev_mtx); 747 if ((sp->sw_flags & SW_UNMAPPED) != 0 && 748 unmapped_buf_allowed) { 749 bp->b_data = unmapped_buf; 750 bp->b_offset = 0; 751 } else { 752 pmap_qenter((vm_offset_t)bp->b_data, 753 &bp->b_pages[0], bp->b_bcount / PAGE_SIZE); 754 } 755 sp->sw_strategy(bp, sp); 756 return; 757 } 758 } 759 panic("Swapdev not found"); 760 } 761 762 763 /* 764 * SWP_PAGER_FREESWAPSPACE() - free raw swap space 765 * 766 * This routine returns the specified swap blocks back to the bitmap. 767 * 768 * This routine may not sleep. 769 */ 770 static void 771 swp_pager_freeswapspace(daddr_t blk, daddr_t npages) 772 { 773 struct swdevt *sp; 774 775 if (npages == 0) 776 return; 777 mtx_lock(&sw_dev_mtx); 778 TAILQ_FOREACH(sp, &swtailq, sw_list) { 779 if (blk >= sp->sw_first && blk < sp->sw_end) { 780 sp->sw_used -= npages; 781 /* 782 * If we are attempting to stop swapping on 783 * this device, we don't want to mark any 784 * blocks free lest they be reused. 785 */ 786 if ((sp->sw_flags & SW_CLOSING) == 0) { 787 blist_free(sp->sw_blist, blk - sp->sw_first, 788 npages); 789 swap_pager_avail += npages; 790 swp_sizecheck(); 791 } 792 mtx_unlock(&sw_dev_mtx); 793 return; 794 } 795 } 796 panic("Swapdev not found"); 797 } 798 799 /* 800 * SYSCTL_SWAP_FRAGMENTATION() - produce raw swap space stats 801 */ 802 static int 803 sysctl_swap_fragmentation(SYSCTL_HANDLER_ARGS) 804 { 805 struct sbuf sbuf; 806 struct swdevt *sp; 807 const char *devname; 808 int error; 809 810 error = sysctl_wire_old_buffer(req, 0); 811 if (error != 0) 812 return (error); 813 sbuf_new_for_sysctl(&sbuf, NULL, 128, req); 814 mtx_lock(&sw_dev_mtx); 815 TAILQ_FOREACH(sp, &swtailq, sw_list) { 816 if (vn_isdisk(sp->sw_vp, NULL)) 817 devname = devtoname(sp->sw_vp->v_rdev); 818 else 819 devname = "[file]"; 820 sbuf_printf(&sbuf, "\nFree space on device %s:\n", devname); 821 blist_stats(sp->sw_blist, &sbuf); 822 } 823 mtx_unlock(&sw_dev_mtx); 824 error = sbuf_finish(&sbuf); 825 sbuf_delete(&sbuf); 826 return (error); 827 } 828 829 /* 830 * SWAP_PAGER_FREESPACE() - frees swap blocks associated with a page 831 * range within an object. 832 * 833 * This is a globally accessible routine. 834 * 835 * This routine removes swapblk assignments from swap metadata. 836 * 837 * The external callers of this routine typically have already destroyed 838 * or renamed vm_page_t's associated with this range in the object so 839 * we should be ok. 840 * 841 * The object must be locked. 842 */ 843 void 844 swap_pager_freespace(vm_object_t object, vm_pindex_t start, vm_size_t size) 845 { 846 847 swp_pager_meta_free(object, start, size); 848 } 849 850 /* 851 * SWAP_PAGER_RESERVE() - reserve swap blocks in object 852 * 853 * Assigns swap blocks to the specified range within the object. The 854 * swap blocks are not zeroed. Any previous swap assignment is destroyed. 855 * 856 * Returns 0 on success, -1 on failure. 857 */ 858 int 859 swap_pager_reserve(vm_object_t object, vm_pindex_t start, vm_size_t size) 860 { 861 int n = 0; 862 daddr_t blk = SWAPBLK_NONE; 863 vm_pindex_t beg = start; /* save start index */ 864 865 VM_OBJECT_WLOCK(object); 866 while (size) { 867 if (n == 0) { 868 n = BLIST_MAX_ALLOC; 869 while ((blk = swp_pager_getswapspace(n)) == SWAPBLK_NONE) { 870 n >>= 1; 871 if (n == 0) { 872 swp_pager_meta_free(object, beg, start - beg); 873 VM_OBJECT_WUNLOCK(object); 874 return (-1); 875 } 876 } 877 } 878 swp_pager_meta_build(object, start, blk); 879 --size; 880 ++start; 881 ++blk; 882 --n; 883 } 884 swp_pager_meta_free(object, start, n); 885 VM_OBJECT_WUNLOCK(object); 886 return (0); 887 } 888 889 /* 890 * SWAP_PAGER_COPY() - copy blocks from source pager to destination pager 891 * and destroy the source. 892 * 893 * Copy any valid swapblks from the source to the destination. In 894 * cases where both the source and destination have a valid swapblk, 895 * we keep the destination's. 896 * 897 * This routine is allowed to sleep. It may sleep allocating metadata 898 * indirectly through swp_pager_meta_build() or if paging is still in 899 * progress on the source. 900 * 901 * The source object contains no vm_page_t's (which is just as well) 902 * 903 * The source object is of type OBJT_SWAP. 904 * 905 * The source and destination objects must be locked. 906 * Both object locks may temporarily be released. 907 */ 908 void 909 swap_pager_copy(vm_object_t srcobject, vm_object_t dstobject, 910 vm_pindex_t offset, int destroysource) 911 { 912 vm_pindex_t i; 913 daddr_t dstaddr, first_free, num_free, srcaddr; 914 915 VM_OBJECT_ASSERT_WLOCKED(srcobject); 916 VM_OBJECT_ASSERT_WLOCKED(dstobject); 917 918 /* 919 * If destroysource is set, we remove the source object from the 920 * swap_pager internal queue now. 921 */ 922 if (destroysource && srcobject->handle != NULL) { 923 vm_object_pip_add(srcobject, 1); 924 VM_OBJECT_WUNLOCK(srcobject); 925 vm_object_pip_add(dstobject, 1); 926 VM_OBJECT_WUNLOCK(dstobject); 927 sx_xlock(&sw_alloc_sx); 928 TAILQ_REMOVE(NOBJLIST(srcobject->handle), srcobject, 929 pager_object_list); 930 sx_xunlock(&sw_alloc_sx); 931 VM_OBJECT_WLOCK(dstobject); 932 vm_object_pip_wakeup(dstobject); 933 VM_OBJECT_WLOCK(srcobject); 934 vm_object_pip_wakeup(srcobject); 935 } 936 937 /* 938 * Transfer source to destination. 939 */ 940 first_free = SWAPBLK_NONE; 941 num_free = 0; 942 for (i = 0; i < dstobject->size; ++i) { 943 srcaddr = swp_pager_meta_ctl(srcobject, i + offset, SWM_POP); 944 if (srcaddr == SWAPBLK_NONE) 945 continue; 946 dstaddr = swp_pager_meta_ctl(dstobject, i, 0); 947 if (dstaddr == SWAPBLK_NONE) { 948 /* 949 * Destination has no swapblk and is not resident, 950 * copy source. 951 * 952 * swp_pager_meta_build() can sleep. 953 */ 954 vm_object_pip_add(srcobject, 1); 955 VM_OBJECT_WUNLOCK(srcobject); 956 vm_object_pip_add(dstobject, 1); 957 swp_pager_meta_build(dstobject, i, srcaddr); 958 vm_object_pip_wakeup(dstobject); 959 VM_OBJECT_WLOCK(srcobject); 960 vm_object_pip_wakeup(srcobject); 961 } else { 962 /* 963 * Destination has valid swapblk or it is represented 964 * by a resident page. We destroy the sourceblock. 965 */ 966 if (first_free + num_free == srcaddr) 967 num_free++; 968 else { 969 swp_pager_freeswapspace(first_free, num_free); 970 first_free = srcaddr; 971 num_free = 1; 972 } 973 } 974 } 975 swp_pager_freeswapspace(first_free, num_free); 976 977 /* 978 * Free left over swap blocks in source. 979 * 980 * We have to revert the type to OBJT_DEFAULT so we do not accidentally 981 * double-remove the object from the swap queues. 982 */ 983 if (destroysource) { 984 swp_pager_meta_free_all(srcobject); 985 /* 986 * Reverting the type is not necessary, the caller is going 987 * to destroy srcobject directly, but I'm doing it here 988 * for consistency since we've removed the object from its 989 * queues. 990 */ 991 srcobject->type = OBJT_DEFAULT; 992 } 993 } 994 995 /* 996 * SWAP_PAGER_HASPAGE() - determine if we have good backing store for 997 * the requested page. 998 * 999 * We determine whether good backing store exists for the requested 1000 * page and return TRUE if it does, FALSE if it doesn't. 1001 * 1002 * If TRUE, we also try to determine how much valid, contiguous backing 1003 * store exists before and after the requested page. 1004 */ 1005 static boolean_t 1006 swap_pager_haspage(vm_object_t object, vm_pindex_t pindex, int *before, 1007 int *after) 1008 { 1009 daddr_t blk, blk0; 1010 int i; 1011 1012 VM_OBJECT_ASSERT_LOCKED(object); 1013 1014 /* 1015 * do we have good backing store at the requested index ? 1016 */ 1017 blk0 = swp_pager_meta_ctl(object, pindex, 0); 1018 if (blk0 == SWAPBLK_NONE) { 1019 if (before) 1020 *before = 0; 1021 if (after) 1022 *after = 0; 1023 return (FALSE); 1024 } 1025 1026 /* 1027 * find backwards-looking contiguous good backing store 1028 */ 1029 if (before != NULL) { 1030 for (i = 1; i < SWB_NPAGES; i++) { 1031 if (i > pindex) 1032 break; 1033 blk = swp_pager_meta_ctl(object, pindex - i, 0); 1034 if (blk != blk0 - i) 1035 break; 1036 } 1037 *before = i - 1; 1038 } 1039 1040 /* 1041 * find forward-looking contiguous good backing store 1042 */ 1043 if (after != NULL) { 1044 for (i = 1; i < SWB_NPAGES; i++) { 1045 blk = swp_pager_meta_ctl(object, pindex + i, 0); 1046 if (blk != blk0 + i) 1047 break; 1048 } 1049 *after = i - 1; 1050 } 1051 return (TRUE); 1052 } 1053 1054 /* 1055 * SWAP_PAGER_PAGE_UNSWAPPED() - remove swap backing store related to page 1056 * 1057 * This removes any associated swap backing store, whether valid or 1058 * not, from the page. 1059 * 1060 * This routine is typically called when a page is made dirty, at 1061 * which point any associated swap can be freed. MADV_FREE also 1062 * calls us in a special-case situation 1063 * 1064 * NOTE!!! If the page is clean and the swap was valid, the caller 1065 * should make the page dirty before calling this routine. This routine 1066 * does NOT change the m->dirty status of the page. Also: MADV_FREE 1067 * depends on it. 1068 * 1069 * This routine may not sleep. 1070 * 1071 * The object containing the page must be locked. 1072 */ 1073 static void 1074 swap_pager_unswapped(vm_page_t m) 1075 { 1076 daddr_t srcaddr; 1077 1078 srcaddr = swp_pager_meta_ctl(m->object, m->pindex, SWM_POP); 1079 if (srcaddr != SWAPBLK_NONE) 1080 swp_pager_freeswapspace(srcaddr, 1); 1081 } 1082 1083 /* 1084 * swap_pager_getpages() - bring pages in from swap 1085 * 1086 * Attempt to page in the pages in array "ma" of length "count". The 1087 * caller may optionally specify that additional pages preceding and 1088 * succeeding the specified range be paged in. The number of such pages 1089 * is returned in the "rbehind" and "rahead" parameters, and they will 1090 * be in the inactive queue upon return. 1091 * 1092 * The pages in "ma" must be busied and will remain busied upon return. 1093 */ 1094 static int 1095 swap_pager_getpages(vm_object_t object, vm_page_t *ma, int count, int *rbehind, 1096 int *rahead) 1097 { 1098 struct buf *bp; 1099 vm_page_t mpred, msucc, p; 1100 vm_pindex_t pindex; 1101 daddr_t blk; 1102 int i, j, maxahead, maxbehind, reqcount, shift; 1103 1104 reqcount = count; 1105 1106 VM_OBJECT_WUNLOCK(object); 1107 bp = getpbuf(&nsw_rcount); 1108 VM_OBJECT_WLOCK(object); 1109 1110 if (!swap_pager_haspage(object, ma[0]->pindex, &maxbehind, &maxahead)) { 1111 relpbuf(bp, &nsw_rcount); 1112 return (VM_PAGER_FAIL); 1113 } 1114 1115 /* 1116 * Clip the readahead and readbehind ranges to exclude resident pages. 1117 */ 1118 if (rahead != NULL) { 1119 KASSERT(reqcount - 1 <= maxahead, 1120 ("page count %d extends beyond swap block", reqcount)); 1121 *rahead = imin(*rahead, maxahead - (reqcount - 1)); 1122 pindex = ma[reqcount - 1]->pindex; 1123 msucc = TAILQ_NEXT(ma[reqcount - 1], listq); 1124 if (msucc != NULL && msucc->pindex - pindex - 1 < *rahead) 1125 *rahead = msucc->pindex - pindex - 1; 1126 } 1127 if (rbehind != NULL) { 1128 *rbehind = imin(*rbehind, maxbehind); 1129 pindex = ma[0]->pindex; 1130 mpred = TAILQ_PREV(ma[0], pglist, listq); 1131 if (mpred != NULL && pindex - mpred->pindex - 1 < *rbehind) 1132 *rbehind = pindex - mpred->pindex - 1; 1133 } 1134 1135 /* 1136 * Allocate readahead and readbehind pages. 1137 */ 1138 shift = rbehind != NULL ? *rbehind : 0; 1139 if (shift != 0) { 1140 for (i = 1; i <= shift; i++) { 1141 p = vm_page_alloc(object, ma[0]->pindex - i, 1142 VM_ALLOC_NORMAL); 1143 if (p == NULL) { 1144 /* Shift allocated pages to the left. */ 1145 for (j = 0; j < i - 1; j++) 1146 bp->b_pages[j] = 1147 bp->b_pages[j + shift - i + 1]; 1148 break; 1149 } 1150 bp->b_pages[shift - i] = p; 1151 } 1152 shift = i - 1; 1153 *rbehind = shift; 1154 } 1155 for (i = 0; i < reqcount; i++) 1156 bp->b_pages[i + shift] = ma[i]; 1157 if (rahead != NULL) { 1158 for (i = 0; i < *rahead; i++) { 1159 p = vm_page_alloc(object, 1160 ma[reqcount - 1]->pindex + i + 1, VM_ALLOC_NORMAL); 1161 if (p == NULL) 1162 break; 1163 bp->b_pages[shift + reqcount + i] = p; 1164 } 1165 *rahead = i; 1166 } 1167 if (rbehind != NULL) 1168 count += *rbehind; 1169 if (rahead != NULL) 1170 count += *rahead; 1171 1172 vm_object_pip_add(object, count); 1173 1174 for (i = 0; i < count; i++) 1175 bp->b_pages[i]->oflags |= VPO_SWAPINPROG; 1176 1177 pindex = bp->b_pages[0]->pindex; 1178 blk = swp_pager_meta_ctl(object, pindex, 0); 1179 KASSERT(blk != SWAPBLK_NONE, 1180 ("no swap blocking containing %p(%jx)", object, (uintmax_t)pindex)); 1181 1182 VM_OBJECT_WUNLOCK(object); 1183 1184 bp->b_flags |= B_PAGING; 1185 bp->b_iocmd = BIO_READ; 1186 bp->b_iodone = swp_pager_async_iodone; 1187 bp->b_rcred = crhold(thread0.td_ucred); 1188 bp->b_wcred = crhold(thread0.td_ucred); 1189 bp->b_blkno = blk; 1190 bp->b_bcount = PAGE_SIZE * count; 1191 bp->b_bufsize = PAGE_SIZE * count; 1192 bp->b_npages = count; 1193 bp->b_pgbefore = rbehind != NULL ? *rbehind : 0; 1194 bp->b_pgafter = rahead != NULL ? *rahead : 0; 1195 1196 VM_CNT_INC(v_swapin); 1197 VM_CNT_ADD(v_swappgsin, count); 1198 1199 /* 1200 * perform the I/O. NOTE!!! bp cannot be considered valid after 1201 * this point because we automatically release it on completion. 1202 * Instead, we look at the one page we are interested in which we 1203 * still hold a lock on even through the I/O completion. 1204 * 1205 * The other pages in our ma[] array are also released on completion, 1206 * so we cannot assume they are valid anymore either. 1207 * 1208 * NOTE: b_blkno is destroyed by the call to swapdev_strategy 1209 */ 1210 BUF_KERNPROC(bp); 1211 swp_pager_strategy(bp); 1212 1213 /* 1214 * Wait for the pages we want to complete. VPO_SWAPINPROG is always 1215 * cleared on completion. If an I/O error occurs, SWAPBLK_NONE 1216 * is set in the metadata for each page in the request. 1217 */ 1218 VM_OBJECT_WLOCK(object); 1219 while ((ma[0]->oflags & VPO_SWAPINPROG) != 0) { 1220 ma[0]->oflags |= VPO_SWAPSLEEP; 1221 VM_CNT_INC(v_intrans); 1222 if (VM_OBJECT_SLEEP(object, &object->paging_in_progress, PSWP, 1223 "swread", hz * 20)) { 1224 printf( 1225 "swap_pager: indefinite wait buffer: bufobj: %p, blkno: %jd, size: %ld\n", 1226 bp->b_bufobj, (intmax_t)bp->b_blkno, bp->b_bcount); 1227 } 1228 } 1229 1230 /* 1231 * If we had an unrecoverable read error pages will not be valid. 1232 */ 1233 for (i = 0; i < reqcount; i++) 1234 if (ma[i]->valid != VM_PAGE_BITS_ALL) 1235 return (VM_PAGER_ERROR); 1236 1237 return (VM_PAGER_OK); 1238 1239 /* 1240 * A final note: in a low swap situation, we cannot deallocate swap 1241 * and mark a page dirty here because the caller is likely to mark 1242 * the page clean when we return, causing the page to possibly revert 1243 * to all-zero's later. 1244 */ 1245 } 1246 1247 /* 1248 * swap_pager_getpages_async(): 1249 * 1250 * Right now this is emulation of asynchronous operation on top of 1251 * swap_pager_getpages(). 1252 */ 1253 static int 1254 swap_pager_getpages_async(vm_object_t object, vm_page_t *ma, int count, 1255 int *rbehind, int *rahead, pgo_getpages_iodone_t iodone, void *arg) 1256 { 1257 int r, error; 1258 1259 r = swap_pager_getpages(object, ma, count, rbehind, rahead); 1260 VM_OBJECT_WUNLOCK(object); 1261 switch (r) { 1262 case VM_PAGER_OK: 1263 error = 0; 1264 break; 1265 case VM_PAGER_ERROR: 1266 error = EIO; 1267 break; 1268 case VM_PAGER_FAIL: 1269 error = EINVAL; 1270 break; 1271 default: 1272 panic("unhandled swap_pager_getpages() error %d", r); 1273 } 1274 (iodone)(arg, ma, count, error); 1275 VM_OBJECT_WLOCK(object); 1276 1277 return (r); 1278 } 1279 1280 /* 1281 * swap_pager_putpages: 1282 * 1283 * Assign swap (if necessary) and initiate I/O on the specified pages. 1284 * 1285 * We support both OBJT_DEFAULT and OBJT_SWAP objects. DEFAULT objects 1286 * are automatically converted to SWAP objects. 1287 * 1288 * In a low memory situation we may block in VOP_STRATEGY(), but the new 1289 * vm_page reservation system coupled with properly written VFS devices 1290 * should ensure that no low-memory deadlock occurs. This is an area 1291 * which needs work. 1292 * 1293 * The parent has N vm_object_pip_add() references prior to 1294 * calling us and will remove references for rtvals[] that are 1295 * not set to VM_PAGER_PEND. We need to remove the rest on I/O 1296 * completion. 1297 * 1298 * The parent has soft-busy'd the pages it passes us and will unbusy 1299 * those whos rtvals[] entry is not set to VM_PAGER_PEND on return. 1300 * We need to unbusy the rest on I/O completion. 1301 */ 1302 static void 1303 swap_pager_putpages(vm_object_t object, vm_page_t *ma, int count, 1304 int flags, int *rtvals) 1305 { 1306 int i, n; 1307 boolean_t sync; 1308 1309 if (count && ma[0]->object != object) { 1310 panic("swap_pager_putpages: object mismatch %p/%p", 1311 object, 1312 ma[0]->object 1313 ); 1314 } 1315 1316 /* 1317 * Step 1 1318 * 1319 * Turn object into OBJT_SWAP 1320 * check for bogus sysops 1321 * force sync if not pageout process 1322 */ 1323 if (object->type != OBJT_SWAP) 1324 swp_pager_meta_build(object, 0, SWAPBLK_NONE); 1325 VM_OBJECT_WUNLOCK(object); 1326 1327 n = 0; 1328 if (curproc != pageproc) 1329 sync = TRUE; 1330 else 1331 sync = (flags & VM_PAGER_PUT_SYNC) != 0; 1332 1333 /* 1334 * Step 2 1335 * 1336 * Assign swap blocks and issue I/O. We reallocate swap on the fly. 1337 * The page is left dirty until the pageout operation completes 1338 * successfully. 1339 */ 1340 for (i = 0; i < count; i += n) { 1341 int j; 1342 struct buf *bp; 1343 daddr_t blk; 1344 1345 /* 1346 * Maximum I/O size is limited by a number of factors. 1347 */ 1348 n = min(BLIST_MAX_ALLOC, count - i); 1349 n = min(n, nsw_cluster_max); 1350 1351 /* 1352 * Get biggest block of swap we can. If we fail, fall 1353 * back and try to allocate a smaller block. Don't go 1354 * overboard trying to allocate space if it would overly 1355 * fragment swap. 1356 */ 1357 while ( 1358 (blk = swp_pager_getswapspace(n)) == SWAPBLK_NONE && 1359 n > 4 1360 ) { 1361 n >>= 1; 1362 } 1363 if (blk == SWAPBLK_NONE) { 1364 for (j = 0; j < n; ++j) 1365 rtvals[i+j] = VM_PAGER_FAIL; 1366 continue; 1367 } 1368 1369 /* 1370 * All I/O parameters have been satisfied, build the I/O 1371 * request and assign the swap space. 1372 */ 1373 if (sync == TRUE) { 1374 bp = getpbuf(&nsw_wcount_sync); 1375 } else { 1376 bp = getpbuf(&nsw_wcount_async); 1377 bp->b_flags = B_ASYNC; 1378 } 1379 bp->b_flags |= B_PAGING; 1380 bp->b_iocmd = BIO_WRITE; 1381 1382 bp->b_rcred = crhold(thread0.td_ucred); 1383 bp->b_wcred = crhold(thread0.td_ucred); 1384 bp->b_bcount = PAGE_SIZE * n; 1385 bp->b_bufsize = PAGE_SIZE * n; 1386 bp->b_blkno = blk; 1387 1388 VM_OBJECT_WLOCK(object); 1389 for (j = 0; j < n; ++j) { 1390 vm_page_t mreq = ma[i+j]; 1391 1392 swp_pager_meta_build( 1393 mreq->object, 1394 mreq->pindex, 1395 blk + j 1396 ); 1397 MPASS(mreq->dirty == VM_PAGE_BITS_ALL); 1398 mreq->oflags |= VPO_SWAPINPROG; 1399 bp->b_pages[j] = mreq; 1400 } 1401 VM_OBJECT_WUNLOCK(object); 1402 bp->b_npages = n; 1403 /* 1404 * Must set dirty range for NFS to work. 1405 */ 1406 bp->b_dirtyoff = 0; 1407 bp->b_dirtyend = bp->b_bcount; 1408 1409 VM_CNT_INC(v_swapout); 1410 VM_CNT_ADD(v_swappgsout, bp->b_npages); 1411 1412 /* 1413 * We unconditionally set rtvals[] to VM_PAGER_PEND so that we 1414 * can call the async completion routine at the end of a 1415 * synchronous I/O operation. Otherwise, our caller would 1416 * perform duplicate unbusy and wakeup operations on the page 1417 * and object, respectively. 1418 */ 1419 for (j = 0; j < n; j++) 1420 rtvals[i + j] = VM_PAGER_PEND; 1421 1422 /* 1423 * asynchronous 1424 * 1425 * NOTE: b_blkno is destroyed by the call to swapdev_strategy 1426 */ 1427 if (sync == FALSE) { 1428 bp->b_iodone = swp_pager_async_iodone; 1429 BUF_KERNPROC(bp); 1430 swp_pager_strategy(bp); 1431 continue; 1432 } 1433 1434 /* 1435 * synchronous 1436 * 1437 * NOTE: b_blkno is destroyed by the call to swapdev_strategy 1438 */ 1439 bp->b_iodone = bdone; 1440 swp_pager_strategy(bp); 1441 1442 /* 1443 * Wait for the sync I/O to complete. 1444 */ 1445 bwait(bp, PVM, "swwrt"); 1446 1447 /* 1448 * Now that we are through with the bp, we can call the 1449 * normal async completion, which frees everything up. 1450 */ 1451 swp_pager_async_iodone(bp); 1452 } 1453 VM_OBJECT_WLOCK(object); 1454 } 1455 1456 /* 1457 * swp_pager_async_iodone: 1458 * 1459 * Completion routine for asynchronous reads and writes from/to swap. 1460 * Also called manually by synchronous code to finish up a bp. 1461 * 1462 * This routine may not sleep. 1463 */ 1464 static void 1465 swp_pager_async_iodone(struct buf *bp) 1466 { 1467 int i; 1468 vm_object_t object = NULL; 1469 1470 /* 1471 * report error 1472 */ 1473 if (bp->b_ioflags & BIO_ERROR) { 1474 printf( 1475 "swap_pager: I/O error - %s failed; blkno %ld," 1476 "size %ld, error %d\n", 1477 ((bp->b_iocmd == BIO_READ) ? "pagein" : "pageout"), 1478 (long)bp->b_blkno, 1479 (long)bp->b_bcount, 1480 bp->b_error 1481 ); 1482 } 1483 1484 /* 1485 * remove the mapping for kernel virtual 1486 */ 1487 if (buf_mapped(bp)) 1488 pmap_qremove((vm_offset_t)bp->b_data, bp->b_npages); 1489 else 1490 bp->b_data = bp->b_kvabase; 1491 1492 if (bp->b_npages) { 1493 object = bp->b_pages[0]->object; 1494 VM_OBJECT_WLOCK(object); 1495 } 1496 1497 /* 1498 * cleanup pages. If an error occurs writing to swap, we are in 1499 * very serious trouble. If it happens to be a disk error, though, 1500 * we may be able to recover by reassigning the swap later on. So 1501 * in this case we remove the m->swapblk assignment for the page 1502 * but do not free it in the rlist. The errornous block(s) are thus 1503 * never reallocated as swap. Redirty the page and continue. 1504 */ 1505 for (i = 0; i < bp->b_npages; ++i) { 1506 vm_page_t m = bp->b_pages[i]; 1507 1508 m->oflags &= ~VPO_SWAPINPROG; 1509 if (m->oflags & VPO_SWAPSLEEP) { 1510 m->oflags &= ~VPO_SWAPSLEEP; 1511 wakeup(&object->paging_in_progress); 1512 } 1513 1514 if (bp->b_ioflags & BIO_ERROR) { 1515 /* 1516 * If an error occurs I'd love to throw the swapblk 1517 * away without freeing it back to swapspace, so it 1518 * can never be used again. But I can't from an 1519 * interrupt. 1520 */ 1521 if (bp->b_iocmd == BIO_READ) { 1522 /* 1523 * NOTE: for reads, m->dirty will probably 1524 * be overridden by the original caller of 1525 * getpages so don't play cute tricks here. 1526 */ 1527 m->valid = 0; 1528 } else { 1529 /* 1530 * If a write error occurs, reactivate page 1531 * so it doesn't clog the inactive list, 1532 * then finish the I/O. 1533 */ 1534 MPASS(m->dirty == VM_PAGE_BITS_ALL); 1535 vm_page_lock(m); 1536 vm_page_activate(m); 1537 vm_page_unlock(m); 1538 vm_page_sunbusy(m); 1539 } 1540 } else if (bp->b_iocmd == BIO_READ) { 1541 /* 1542 * NOTE: for reads, m->dirty will probably be 1543 * overridden by the original caller of getpages so 1544 * we cannot set them in order to free the underlying 1545 * swap in a low-swap situation. I don't think we'd 1546 * want to do that anyway, but it was an optimization 1547 * that existed in the old swapper for a time before 1548 * it got ripped out due to precisely this problem. 1549 */ 1550 KASSERT(!pmap_page_is_mapped(m), 1551 ("swp_pager_async_iodone: page %p is mapped", m)); 1552 KASSERT(m->dirty == 0, 1553 ("swp_pager_async_iodone: page %p is dirty", m)); 1554 1555 m->valid = VM_PAGE_BITS_ALL; 1556 if (i < bp->b_pgbefore || 1557 i >= bp->b_npages - bp->b_pgafter) 1558 vm_page_readahead_finish(m); 1559 } else { 1560 /* 1561 * For write success, clear the dirty 1562 * status, then finish the I/O ( which decrements the 1563 * busy count and possibly wakes waiter's up ). 1564 * A page is only written to swap after a period of 1565 * inactivity. Therefore, we do not expect it to be 1566 * reused. 1567 */ 1568 KASSERT(!pmap_page_is_write_mapped(m), 1569 ("swp_pager_async_iodone: page %p is not write" 1570 " protected", m)); 1571 vm_page_undirty(m); 1572 vm_page_lock(m); 1573 vm_page_deactivate_noreuse(m); 1574 vm_page_unlock(m); 1575 vm_page_sunbusy(m); 1576 } 1577 } 1578 1579 /* 1580 * adjust pip. NOTE: the original parent may still have its own 1581 * pip refs on the object. 1582 */ 1583 if (object != NULL) { 1584 vm_object_pip_wakeupn(object, bp->b_npages); 1585 VM_OBJECT_WUNLOCK(object); 1586 } 1587 1588 /* 1589 * swapdev_strategy() manually sets b_vp and b_bufobj before calling 1590 * bstrategy(). Set them back to NULL now we're done with it, or we'll 1591 * trigger a KASSERT in relpbuf(). 1592 */ 1593 if (bp->b_vp) { 1594 bp->b_vp = NULL; 1595 bp->b_bufobj = NULL; 1596 } 1597 /* 1598 * release the physical I/O buffer 1599 */ 1600 relpbuf( 1601 bp, 1602 ((bp->b_iocmd == BIO_READ) ? &nsw_rcount : 1603 ((bp->b_flags & B_ASYNC) ? 1604 &nsw_wcount_async : 1605 &nsw_wcount_sync 1606 ) 1607 ) 1608 ); 1609 } 1610 1611 int 1612 swap_pager_nswapdev(void) 1613 { 1614 1615 return (nswapdev); 1616 } 1617 1618 /* 1619 * SWP_PAGER_FORCE_PAGEIN() - force a swap block to be paged in 1620 * 1621 * This routine dissociates the page at the given index within an object 1622 * from its backing store, paging it in if it does not reside in memory. 1623 * If the page is paged in, it is marked dirty and placed in the laundry 1624 * queue. The page is marked dirty because it no longer has backing 1625 * store. It is placed in the laundry queue because it has not been 1626 * accessed recently. Otherwise, it would already reside in memory. 1627 * 1628 * We also attempt to swap in all other pages in the swap block. 1629 * However, we only guarantee that the one at the specified index is 1630 * paged in. 1631 * 1632 * XXX - The code to page the whole block in doesn't work, so we 1633 * revert to the one-by-one behavior for now. Sigh. 1634 */ 1635 static inline void 1636 swp_pager_force_pagein(vm_object_t object, vm_pindex_t pindex) 1637 { 1638 vm_page_t m; 1639 1640 vm_object_pip_add(object, 1); 1641 m = vm_page_grab(object, pindex, VM_ALLOC_NORMAL); 1642 if (m->valid == VM_PAGE_BITS_ALL) { 1643 vm_object_pip_wakeup(object); 1644 vm_page_dirty(m); 1645 #ifdef INVARIANTS 1646 vm_page_lock(m); 1647 if (m->wire_count == 0 && m->queue == PQ_NONE) 1648 panic("page %p is neither wired nor queued", m); 1649 vm_page_unlock(m); 1650 #endif 1651 vm_page_xunbusy(m); 1652 vm_pager_page_unswapped(m); 1653 return; 1654 } 1655 1656 if (swap_pager_getpages(object, &m, 1, NULL, NULL) != VM_PAGER_OK) 1657 panic("swap_pager_force_pagein: read from swap failed");/*XXX*/ 1658 vm_object_pip_wakeup(object); 1659 vm_page_dirty(m); 1660 vm_page_lock(m); 1661 vm_page_launder(m); 1662 vm_page_unlock(m); 1663 vm_page_xunbusy(m); 1664 vm_pager_page_unswapped(m); 1665 } 1666 1667 /* 1668 * swap_pager_swapoff: 1669 * 1670 * Page in all of the pages that have been paged out to the 1671 * given device. The corresponding blocks in the bitmap must be 1672 * marked as allocated and the device must be flagged SW_CLOSING. 1673 * There may be no processes swapped out to the device. 1674 * 1675 * This routine may block. 1676 */ 1677 static void 1678 swap_pager_swapoff(struct swdevt *sp) 1679 { 1680 struct swblk *sb; 1681 vm_object_t object; 1682 vm_pindex_t pi; 1683 int i, retries; 1684 1685 sx_assert(&swdev_syscall_lock, SA_XLOCKED); 1686 1687 retries = 0; 1688 full_rescan: 1689 mtx_lock(&vm_object_list_mtx); 1690 TAILQ_FOREACH(object, &vm_object_list, object_list) { 1691 if (object->type != OBJT_SWAP) 1692 continue; 1693 mtx_unlock(&vm_object_list_mtx); 1694 /* Depends on type-stability. */ 1695 VM_OBJECT_WLOCK(object); 1696 1697 /* 1698 * Dead objects are eventually terminated on their own. 1699 */ 1700 if ((object->flags & OBJ_DEAD) != 0) 1701 goto next_obj; 1702 1703 /* 1704 * Sync with fences placed after pctrie 1705 * initialization. We must not access pctrie below 1706 * unless we checked that our object is swap and not 1707 * dead. 1708 */ 1709 atomic_thread_fence_acq(); 1710 if (object->type != OBJT_SWAP) 1711 goto next_obj; 1712 1713 for (pi = 0; (sb = SWAP_PCTRIE_LOOKUP_GE( 1714 &object->un_pager.swp.swp_blks, pi)) != NULL; ) { 1715 pi = sb->p + SWAP_META_PAGES; 1716 for (i = 0; i < SWAP_META_PAGES; i++) { 1717 if (sb->d[i] == SWAPBLK_NONE) 1718 continue; 1719 if (swp_pager_isondev(sb->d[i], sp)) 1720 swp_pager_force_pagein(object, 1721 sb->p + i); 1722 } 1723 } 1724 next_obj: 1725 VM_OBJECT_WUNLOCK(object); 1726 mtx_lock(&vm_object_list_mtx); 1727 } 1728 mtx_unlock(&vm_object_list_mtx); 1729 1730 if (sp->sw_used) { 1731 /* 1732 * Objects may be locked or paging to the device being 1733 * removed, so we will miss their pages and need to 1734 * make another pass. We have marked this device as 1735 * SW_CLOSING, so the activity should finish soon. 1736 */ 1737 retries++; 1738 if (retries > 100) { 1739 panic("swapoff: failed to locate %d swap blocks", 1740 sp->sw_used); 1741 } 1742 pause("swpoff", hz / 20); 1743 goto full_rescan; 1744 } 1745 EVENTHANDLER_INVOKE(swapoff, sp); 1746 } 1747 1748 /************************************************************************ 1749 * SWAP META DATA * 1750 ************************************************************************ 1751 * 1752 * These routines manipulate the swap metadata stored in the 1753 * OBJT_SWAP object. 1754 * 1755 * Swap metadata is implemented with a global hash and not directly 1756 * linked into the object. Instead the object simply contains 1757 * appropriate tracking counters. 1758 */ 1759 1760 /* 1761 * SWP_PAGER_SWBLK_EMPTY() - is a range of blocks free? 1762 */ 1763 static bool 1764 swp_pager_swblk_empty(struct swblk *sb, int start, int limit) 1765 { 1766 int i; 1767 1768 MPASS(0 <= start && start <= limit && limit <= SWAP_META_PAGES); 1769 for (i = start; i < limit; i++) { 1770 if (sb->d[i] != SWAPBLK_NONE) 1771 return (false); 1772 } 1773 return (true); 1774 } 1775 1776 /* 1777 * SWP_PAGER_META_BUILD() - add swap block to swap meta data for object 1778 * 1779 * We first convert the object to a swap object if it is a default 1780 * object. 1781 * 1782 * The specified swapblk is added to the object's swap metadata. If 1783 * the swapblk is not valid, it is freed instead. Any previously 1784 * assigned swapblk is freed. 1785 */ 1786 static void 1787 swp_pager_meta_build(vm_object_t object, vm_pindex_t pindex, daddr_t swapblk) 1788 { 1789 static volatile int swblk_zone_exhausted, swpctrie_zone_exhausted; 1790 struct swblk *sb, *sb1; 1791 vm_pindex_t modpi, rdpi; 1792 int error, i; 1793 1794 VM_OBJECT_ASSERT_WLOCKED(object); 1795 1796 /* 1797 * Convert default object to swap object if necessary 1798 */ 1799 if (object->type != OBJT_SWAP) { 1800 pctrie_init(&object->un_pager.swp.swp_blks); 1801 1802 /* 1803 * Ensure that swap_pager_swapoff()'s iteration over 1804 * object_list does not see a garbage pctrie. 1805 */ 1806 atomic_thread_fence_rel(); 1807 1808 object->type = OBJT_SWAP; 1809 KASSERT(object->handle == NULL, ("default pager with handle")); 1810 } 1811 1812 rdpi = rounddown(pindex, SWAP_META_PAGES); 1813 sb = SWAP_PCTRIE_LOOKUP(&object->un_pager.swp.swp_blks, rdpi); 1814 if (sb == NULL) { 1815 if (swapblk == SWAPBLK_NONE) 1816 return; 1817 for (;;) { 1818 sb = uma_zalloc(swblk_zone, M_NOWAIT | (curproc == 1819 pageproc ? M_USE_RESERVE : 0)); 1820 if (sb != NULL) { 1821 sb->p = rdpi; 1822 for (i = 0; i < SWAP_META_PAGES; i++) 1823 sb->d[i] = SWAPBLK_NONE; 1824 if (atomic_cmpset_int(&swblk_zone_exhausted, 1825 1, 0)) 1826 printf("swblk zone ok\n"); 1827 break; 1828 } 1829 VM_OBJECT_WUNLOCK(object); 1830 if (uma_zone_exhausted(swblk_zone)) { 1831 if (atomic_cmpset_int(&swblk_zone_exhausted, 1832 0, 1)) 1833 printf("swap blk zone exhausted, " 1834 "increase kern.maxswzone\n"); 1835 vm_pageout_oom(VM_OOM_SWAPZ); 1836 pause("swzonxb", 10); 1837 } else 1838 uma_zwait(swblk_zone); 1839 VM_OBJECT_WLOCK(object); 1840 sb = SWAP_PCTRIE_LOOKUP(&object->un_pager.swp.swp_blks, 1841 rdpi); 1842 if (sb != NULL) 1843 /* 1844 * Somebody swapped out a nearby page, 1845 * allocating swblk at the rdpi index, 1846 * while we dropped the object lock. 1847 */ 1848 goto allocated; 1849 } 1850 for (;;) { 1851 error = SWAP_PCTRIE_INSERT( 1852 &object->un_pager.swp.swp_blks, sb); 1853 if (error == 0) { 1854 if (atomic_cmpset_int(&swpctrie_zone_exhausted, 1855 1, 0)) 1856 printf("swpctrie zone ok\n"); 1857 break; 1858 } 1859 VM_OBJECT_WUNLOCK(object); 1860 if (uma_zone_exhausted(swpctrie_zone)) { 1861 if (atomic_cmpset_int(&swpctrie_zone_exhausted, 1862 0, 1)) 1863 printf("swap pctrie zone exhausted, " 1864 "increase kern.maxswzone\n"); 1865 vm_pageout_oom(VM_OOM_SWAPZ); 1866 pause("swzonxp", 10); 1867 } else 1868 uma_zwait(swpctrie_zone); 1869 VM_OBJECT_WLOCK(object); 1870 sb1 = SWAP_PCTRIE_LOOKUP(&object->un_pager.swp.swp_blks, 1871 rdpi); 1872 if (sb1 != NULL) { 1873 uma_zfree(swblk_zone, sb); 1874 sb = sb1; 1875 goto allocated; 1876 } 1877 } 1878 } 1879 allocated: 1880 MPASS(sb->p == rdpi); 1881 1882 modpi = pindex % SWAP_META_PAGES; 1883 /* Delete prior contents of metadata. */ 1884 if (sb->d[modpi] != SWAPBLK_NONE) 1885 swp_pager_freeswapspace(sb->d[modpi], 1); 1886 /* Enter block into metadata. */ 1887 sb->d[modpi] = swapblk; 1888 1889 /* 1890 * Free the swblk if we end up with the empty page run. 1891 */ 1892 if (swapblk == SWAPBLK_NONE && 1893 swp_pager_swblk_empty(sb, 0, SWAP_META_PAGES)) { 1894 SWAP_PCTRIE_REMOVE(&object->un_pager.swp.swp_blks, rdpi); 1895 uma_zfree(swblk_zone, sb); 1896 } 1897 } 1898 1899 /* 1900 * SWP_PAGER_META_FREE() - free a range of blocks in the object's swap metadata 1901 * 1902 * The requested range of blocks is freed, with any associated swap 1903 * returned to the swap bitmap. 1904 * 1905 * This routine will free swap metadata structures as they are cleaned 1906 * out. This routine does *NOT* operate on swap metadata associated 1907 * with resident pages. 1908 */ 1909 static void 1910 swp_pager_meta_free(vm_object_t object, vm_pindex_t pindex, vm_pindex_t count) 1911 { 1912 struct swblk *sb; 1913 daddr_t first_free, num_free; 1914 vm_pindex_t last; 1915 int i, limit, start; 1916 1917 VM_OBJECT_ASSERT_WLOCKED(object); 1918 if (object->type != OBJT_SWAP || count == 0) 1919 return; 1920 1921 first_free = SWAPBLK_NONE; 1922 num_free = 0; 1923 last = pindex + count; 1924 for (;;) { 1925 sb = SWAP_PCTRIE_LOOKUP_GE(&object->un_pager.swp.swp_blks, 1926 rounddown(pindex, SWAP_META_PAGES)); 1927 if (sb == NULL || sb->p >= last) 1928 break; 1929 start = pindex > sb->p ? pindex - sb->p : 0; 1930 limit = last - sb->p < SWAP_META_PAGES ? last - sb->p : 1931 SWAP_META_PAGES; 1932 for (i = start; i < limit; i++) { 1933 if (sb->d[i] == SWAPBLK_NONE) 1934 continue; 1935 if (first_free + num_free == sb->d[i]) 1936 num_free++; 1937 else { 1938 swp_pager_freeswapspace(first_free, num_free); 1939 first_free = sb->d[i]; 1940 num_free = 1; 1941 } 1942 sb->d[i] = SWAPBLK_NONE; 1943 } 1944 if (swp_pager_swblk_empty(sb, 0, start) && 1945 swp_pager_swblk_empty(sb, limit, SWAP_META_PAGES)) { 1946 SWAP_PCTRIE_REMOVE(&object->un_pager.swp.swp_blks, 1947 sb->p); 1948 uma_zfree(swblk_zone, sb); 1949 } 1950 pindex = sb->p + SWAP_META_PAGES; 1951 } 1952 swp_pager_freeswapspace(first_free, num_free); 1953 } 1954 1955 /* 1956 * SWP_PAGER_META_FREE_ALL() - destroy all swap metadata associated with object 1957 * 1958 * This routine locates and destroys all swap metadata associated with 1959 * an object. 1960 */ 1961 static void 1962 swp_pager_meta_free_all(vm_object_t object) 1963 { 1964 struct swblk *sb; 1965 daddr_t first_free, num_free; 1966 vm_pindex_t pindex; 1967 int i; 1968 1969 VM_OBJECT_ASSERT_WLOCKED(object); 1970 if (object->type != OBJT_SWAP) 1971 return; 1972 1973 first_free = SWAPBLK_NONE; 1974 num_free = 0; 1975 for (pindex = 0; (sb = SWAP_PCTRIE_LOOKUP_GE( 1976 &object->un_pager.swp.swp_blks, pindex)) != NULL;) { 1977 pindex = sb->p + SWAP_META_PAGES; 1978 for (i = 0; i < SWAP_META_PAGES; i++) { 1979 if (sb->d[i] == SWAPBLK_NONE) 1980 continue; 1981 if (first_free + num_free == sb->d[i]) 1982 num_free++; 1983 else { 1984 swp_pager_freeswapspace(first_free, num_free); 1985 first_free = sb->d[i]; 1986 num_free = 1; 1987 } 1988 } 1989 SWAP_PCTRIE_REMOVE(&object->un_pager.swp.swp_blks, sb->p); 1990 uma_zfree(swblk_zone, sb); 1991 } 1992 swp_pager_freeswapspace(first_free, num_free); 1993 } 1994 1995 /* 1996 * SWP_PAGER_METACTL() - misc control of swap meta data. 1997 * 1998 * This routine is capable of looking up, or removing swapblk 1999 * assignments in the swap meta data. It returns the swapblk being 2000 * looked-up, popped, or SWAPBLK_NONE if the block was invalid. 2001 * 2002 * When acting on a busy resident page and paging is in progress, we 2003 * have to wait until paging is complete but otherwise can act on the 2004 * busy page. 2005 * 2006 * SWM_POP remove from meta data but do not free it 2007 */ 2008 static daddr_t 2009 swp_pager_meta_ctl(vm_object_t object, vm_pindex_t pindex, int flags) 2010 { 2011 struct swblk *sb; 2012 daddr_t r1; 2013 2014 if ((flags & SWM_POP) != 0) 2015 VM_OBJECT_ASSERT_WLOCKED(object); 2016 else 2017 VM_OBJECT_ASSERT_LOCKED(object); 2018 2019 /* 2020 * The meta data only exists if the object is OBJT_SWAP 2021 * and even then might not be allocated yet. 2022 */ 2023 if (object->type != OBJT_SWAP) 2024 return (SWAPBLK_NONE); 2025 2026 sb = SWAP_PCTRIE_LOOKUP(&object->un_pager.swp.swp_blks, 2027 rounddown(pindex, SWAP_META_PAGES)); 2028 if (sb == NULL) 2029 return (SWAPBLK_NONE); 2030 r1 = sb->d[pindex % SWAP_META_PAGES]; 2031 if (r1 == SWAPBLK_NONE) 2032 return (SWAPBLK_NONE); 2033 if ((flags & SWM_POP) != 0) { 2034 sb->d[pindex % SWAP_META_PAGES] = SWAPBLK_NONE; 2035 if (swp_pager_swblk_empty(sb, 0, SWAP_META_PAGES)) { 2036 SWAP_PCTRIE_REMOVE(&object->un_pager.swp.swp_blks, 2037 rounddown(pindex, SWAP_META_PAGES)); 2038 uma_zfree(swblk_zone, sb); 2039 } 2040 } 2041 return (r1); 2042 } 2043 2044 /* 2045 * Returns the least page index which is greater than or equal to the 2046 * parameter pindex and for which there is a swap block allocated. 2047 * Returns object's size if the object's type is not swap or if there 2048 * are no allocated swap blocks for the object after the requested 2049 * pindex. 2050 */ 2051 vm_pindex_t 2052 swap_pager_find_least(vm_object_t object, vm_pindex_t pindex) 2053 { 2054 struct swblk *sb; 2055 int i; 2056 2057 VM_OBJECT_ASSERT_LOCKED(object); 2058 if (object->type != OBJT_SWAP) 2059 return (object->size); 2060 2061 sb = SWAP_PCTRIE_LOOKUP_GE(&object->un_pager.swp.swp_blks, 2062 rounddown(pindex, SWAP_META_PAGES)); 2063 if (sb == NULL) 2064 return (object->size); 2065 if (sb->p < pindex) { 2066 for (i = pindex % SWAP_META_PAGES; i < SWAP_META_PAGES; i++) { 2067 if (sb->d[i] != SWAPBLK_NONE) 2068 return (sb->p + i); 2069 } 2070 sb = SWAP_PCTRIE_LOOKUP_GE(&object->un_pager.swp.swp_blks, 2071 roundup(pindex, SWAP_META_PAGES)); 2072 if (sb == NULL) 2073 return (object->size); 2074 } 2075 for (i = 0; i < SWAP_META_PAGES; i++) { 2076 if (sb->d[i] != SWAPBLK_NONE) 2077 return (sb->p + i); 2078 } 2079 2080 /* 2081 * We get here if a swblk is present in the trie but it 2082 * doesn't map any blocks. 2083 */ 2084 MPASS(0); 2085 return (object->size); 2086 } 2087 2088 /* 2089 * System call swapon(name) enables swapping on device name, 2090 * which must be in the swdevsw. Return EBUSY 2091 * if already swapping on this device. 2092 */ 2093 #ifndef _SYS_SYSPROTO_H_ 2094 struct swapon_args { 2095 char *name; 2096 }; 2097 #endif 2098 2099 /* 2100 * MPSAFE 2101 */ 2102 /* ARGSUSED */ 2103 int 2104 sys_swapon(struct thread *td, struct swapon_args *uap) 2105 { 2106 struct vattr attr; 2107 struct vnode *vp; 2108 struct nameidata nd; 2109 int error; 2110 2111 error = priv_check(td, PRIV_SWAPON); 2112 if (error) 2113 return (error); 2114 2115 sx_xlock(&swdev_syscall_lock); 2116 2117 /* 2118 * Swap metadata may not fit in the KVM if we have physical 2119 * memory of >1GB. 2120 */ 2121 if (swblk_zone == NULL) { 2122 error = ENOMEM; 2123 goto done; 2124 } 2125 2126 NDINIT(&nd, LOOKUP, ISOPEN | FOLLOW | AUDITVNODE1, UIO_USERSPACE, 2127 uap->name, td); 2128 error = namei(&nd); 2129 if (error) 2130 goto done; 2131 2132 NDFREE(&nd, NDF_ONLY_PNBUF); 2133 vp = nd.ni_vp; 2134 2135 if (vn_isdisk(vp, &error)) { 2136 error = swapongeom(vp); 2137 } else if (vp->v_type == VREG && 2138 (vp->v_mount->mnt_vfc->vfc_flags & VFCF_NETWORK) != 0 && 2139 (error = VOP_GETATTR(vp, &attr, td->td_ucred)) == 0) { 2140 /* 2141 * Allow direct swapping to NFS regular files in the same 2142 * way that nfs_mountroot() sets up diskless swapping. 2143 */ 2144 error = swaponvp(td, vp, attr.va_size / DEV_BSIZE); 2145 } 2146 2147 if (error) 2148 vrele(vp); 2149 done: 2150 sx_xunlock(&swdev_syscall_lock); 2151 return (error); 2152 } 2153 2154 /* 2155 * Check that the total amount of swap currently configured does not 2156 * exceed half the theoretical maximum. If it does, print a warning 2157 * message. 2158 */ 2159 static void 2160 swapon_check_swzone(void) 2161 { 2162 unsigned long maxpages, npages; 2163 2164 npages = swap_total / PAGE_SIZE; 2165 /* absolute maximum we can handle assuming 100% efficiency */ 2166 maxpages = uma_zone_get_max(swblk_zone) * SWAP_META_PAGES; 2167 2168 /* recommend using no more than half that amount */ 2169 if (npages > maxpages / 2) { 2170 printf("warning: total configured swap (%lu pages) " 2171 "exceeds maximum recommended amount (%lu pages).\n", 2172 npages, maxpages / 2); 2173 printf("warning: increase kern.maxswzone " 2174 "or reduce amount of swap.\n"); 2175 } 2176 } 2177 2178 static void 2179 swaponsomething(struct vnode *vp, void *id, u_long nblks, 2180 sw_strategy_t *strategy, sw_close_t *close, dev_t dev, int flags) 2181 { 2182 struct swdevt *sp, *tsp; 2183 swblk_t dvbase; 2184 u_long mblocks; 2185 2186 /* 2187 * nblks is in DEV_BSIZE'd chunks, convert to PAGE_SIZE'd chunks. 2188 * First chop nblks off to page-align it, then convert. 2189 * 2190 * sw->sw_nblks is in page-sized chunks now too. 2191 */ 2192 nblks &= ~(ctodb(1) - 1); 2193 nblks = dbtoc(nblks); 2194 2195 /* 2196 * If we go beyond this, we get overflows in the radix 2197 * tree bitmap code. 2198 */ 2199 mblocks = 0x40000000 / BLIST_META_RADIX; 2200 if (nblks > mblocks) { 2201 printf( 2202 "WARNING: reducing swap size to maximum of %luMB per unit\n", 2203 mblocks / 1024 / 1024 * PAGE_SIZE); 2204 nblks = mblocks; 2205 } 2206 2207 sp = malloc(sizeof *sp, M_VMPGDATA, M_WAITOK | M_ZERO); 2208 sp->sw_vp = vp; 2209 sp->sw_id = id; 2210 sp->sw_dev = dev; 2211 sp->sw_flags = 0; 2212 sp->sw_nblks = nblks; 2213 sp->sw_used = 0; 2214 sp->sw_strategy = strategy; 2215 sp->sw_close = close; 2216 sp->sw_flags = flags; 2217 2218 sp->sw_blist = blist_create(nblks, M_WAITOK); 2219 /* 2220 * Do not free the first two block in order to avoid overwriting 2221 * any bsd label at the front of the partition 2222 */ 2223 blist_free(sp->sw_blist, 2, nblks - 2); 2224 2225 dvbase = 0; 2226 mtx_lock(&sw_dev_mtx); 2227 TAILQ_FOREACH(tsp, &swtailq, sw_list) { 2228 if (tsp->sw_end >= dvbase) { 2229 /* 2230 * We put one uncovered page between the devices 2231 * in order to definitively prevent any cross-device 2232 * I/O requests 2233 */ 2234 dvbase = tsp->sw_end + 1; 2235 } 2236 } 2237 sp->sw_first = dvbase; 2238 sp->sw_end = dvbase + nblks; 2239 TAILQ_INSERT_TAIL(&swtailq, sp, sw_list); 2240 nswapdev++; 2241 swap_pager_avail += nblks - 2; 2242 swap_total += (vm_ooffset_t)nblks * PAGE_SIZE; 2243 swapon_check_swzone(); 2244 swp_sizecheck(); 2245 mtx_unlock(&sw_dev_mtx); 2246 EVENTHANDLER_INVOKE(swapon, sp); 2247 } 2248 2249 /* 2250 * SYSCALL: swapoff(devname) 2251 * 2252 * Disable swapping on the given device. 2253 * 2254 * XXX: Badly designed system call: it should use a device index 2255 * rather than filename as specification. We keep sw_vp around 2256 * only to make this work. 2257 */ 2258 #ifndef _SYS_SYSPROTO_H_ 2259 struct swapoff_args { 2260 char *name; 2261 }; 2262 #endif 2263 2264 /* 2265 * MPSAFE 2266 */ 2267 /* ARGSUSED */ 2268 int 2269 sys_swapoff(struct thread *td, struct swapoff_args *uap) 2270 { 2271 struct vnode *vp; 2272 struct nameidata nd; 2273 struct swdevt *sp; 2274 int error; 2275 2276 error = priv_check(td, PRIV_SWAPOFF); 2277 if (error) 2278 return (error); 2279 2280 sx_xlock(&swdev_syscall_lock); 2281 2282 NDINIT(&nd, LOOKUP, FOLLOW | AUDITVNODE1, UIO_USERSPACE, uap->name, 2283 td); 2284 error = namei(&nd); 2285 if (error) 2286 goto done; 2287 NDFREE(&nd, NDF_ONLY_PNBUF); 2288 vp = nd.ni_vp; 2289 2290 mtx_lock(&sw_dev_mtx); 2291 TAILQ_FOREACH(sp, &swtailq, sw_list) { 2292 if (sp->sw_vp == vp) 2293 break; 2294 } 2295 mtx_unlock(&sw_dev_mtx); 2296 if (sp == NULL) { 2297 error = EINVAL; 2298 goto done; 2299 } 2300 error = swapoff_one(sp, td->td_ucred); 2301 done: 2302 sx_xunlock(&swdev_syscall_lock); 2303 return (error); 2304 } 2305 2306 static int 2307 swapoff_one(struct swdevt *sp, struct ucred *cred) 2308 { 2309 u_long nblks; 2310 #ifdef MAC 2311 int error; 2312 #endif 2313 2314 sx_assert(&swdev_syscall_lock, SA_XLOCKED); 2315 #ifdef MAC 2316 (void) vn_lock(sp->sw_vp, LK_EXCLUSIVE | LK_RETRY); 2317 error = mac_system_check_swapoff(cred, sp->sw_vp); 2318 (void) VOP_UNLOCK(sp->sw_vp, 0); 2319 if (error != 0) 2320 return (error); 2321 #endif 2322 nblks = sp->sw_nblks; 2323 2324 /* 2325 * We can turn off this swap device safely only if the 2326 * available virtual memory in the system will fit the amount 2327 * of data we will have to page back in, plus an epsilon so 2328 * the system doesn't become critically low on swap space. 2329 */ 2330 if (vm_free_count() + swap_pager_avail < nblks + nswap_lowat) 2331 return (ENOMEM); 2332 2333 /* 2334 * Prevent further allocations on this device. 2335 */ 2336 mtx_lock(&sw_dev_mtx); 2337 sp->sw_flags |= SW_CLOSING; 2338 swap_pager_avail -= blist_fill(sp->sw_blist, 0, nblks); 2339 swap_total -= (vm_ooffset_t)nblks * PAGE_SIZE; 2340 mtx_unlock(&sw_dev_mtx); 2341 2342 /* 2343 * Page in the contents of the device and close it. 2344 */ 2345 swap_pager_swapoff(sp); 2346 2347 sp->sw_close(curthread, sp); 2348 mtx_lock(&sw_dev_mtx); 2349 sp->sw_id = NULL; 2350 TAILQ_REMOVE(&swtailq, sp, sw_list); 2351 nswapdev--; 2352 if (nswapdev == 0) { 2353 swap_pager_full = 2; 2354 swap_pager_almost_full = 1; 2355 } 2356 if (swdevhd == sp) 2357 swdevhd = NULL; 2358 mtx_unlock(&sw_dev_mtx); 2359 blist_destroy(sp->sw_blist); 2360 free(sp, M_VMPGDATA); 2361 return (0); 2362 } 2363 2364 void 2365 swapoff_all(void) 2366 { 2367 struct swdevt *sp, *spt; 2368 const char *devname; 2369 int error; 2370 2371 sx_xlock(&swdev_syscall_lock); 2372 2373 mtx_lock(&sw_dev_mtx); 2374 TAILQ_FOREACH_SAFE(sp, &swtailq, sw_list, spt) { 2375 mtx_unlock(&sw_dev_mtx); 2376 if (vn_isdisk(sp->sw_vp, NULL)) 2377 devname = devtoname(sp->sw_vp->v_rdev); 2378 else 2379 devname = "[file]"; 2380 error = swapoff_one(sp, thread0.td_ucred); 2381 if (error != 0) { 2382 printf("Cannot remove swap device %s (error=%d), " 2383 "skipping.\n", devname, error); 2384 } else if (bootverbose) { 2385 printf("Swap device %s removed.\n", devname); 2386 } 2387 mtx_lock(&sw_dev_mtx); 2388 } 2389 mtx_unlock(&sw_dev_mtx); 2390 2391 sx_xunlock(&swdev_syscall_lock); 2392 } 2393 2394 void 2395 swap_pager_status(int *total, int *used) 2396 { 2397 struct swdevt *sp; 2398 2399 *total = 0; 2400 *used = 0; 2401 mtx_lock(&sw_dev_mtx); 2402 TAILQ_FOREACH(sp, &swtailq, sw_list) { 2403 *total += sp->sw_nblks; 2404 *used += sp->sw_used; 2405 } 2406 mtx_unlock(&sw_dev_mtx); 2407 } 2408 2409 int 2410 swap_dev_info(int name, struct xswdev *xs, char *devname, size_t len) 2411 { 2412 struct swdevt *sp; 2413 const char *tmp_devname; 2414 int error, n; 2415 2416 n = 0; 2417 error = ENOENT; 2418 mtx_lock(&sw_dev_mtx); 2419 TAILQ_FOREACH(sp, &swtailq, sw_list) { 2420 if (n != name) { 2421 n++; 2422 continue; 2423 } 2424 xs->xsw_version = XSWDEV_VERSION; 2425 xs->xsw_dev = sp->sw_dev; 2426 xs->xsw_flags = sp->sw_flags; 2427 xs->xsw_nblks = sp->sw_nblks; 2428 xs->xsw_used = sp->sw_used; 2429 if (devname != NULL) { 2430 if (vn_isdisk(sp->sw_vp, NULL)) 2431 tmp_devname = devtoname(sp->sw_vp->v_rdev); 2432 else 2433 tmp_devname = "[file]"; 2434 strncpy(devname, tmp_devname, len); 2435 } 2436 error = 0; 2437 break; 2438 } 2439 mtx_unlock(&sw_dev_mtx); 2440 return (error); 2441 } 2442 2443 #if defined(COMPAT_FREEBSD11) 2444 #define XSWDEV_VERSION_11 1 2445 struct xswdev11 { 2446 u_int xsw_version; 2447 uint32_t xsw_dev; 2448 int xsw_flags; 2449 int xsw_nblks; 2450 int xsw_used; 2451 }; 2452 #endif 2453 2454 static int 2455 sysctl_vm_swap_info(SYSCTL_HANDLER_ARGS) 2456 { 2457 struct xswdev xs; 2458 #if defined(COMPAT_FREEBSD11) 2459 struct xswdev11 xs11; 2460 #endif 2461 int error; 2462 2463 if (arg2 != 1) /* name length */ 2464 return (EINVAL); 2465 error = swap_dev_info(*(int *)arg1, &xs, NULL, 0); 2466 if (error != 0) 2467 return (error); 2468 #if defined(COMPAT_FREEBSD11) 2469 if (req->oldlen == sizeof(xs11)) { 2470 xs11.xsw_version = XSWDEV_VERSION_11; 2471 xs11.xsw_dev = xs.xsw_dev; /* truncation */ 2472 xs11.xsw_flags = xs.xsw_flags; 2473 xs11.xsw_nblks = xs.xsw_nblks; 2474 xs11.xsw_used = xs.xsw_used; 2475 error = SYSCTL_OUT(req, &xs11, sizeof(xs11)); 2476 } else 2477 #endif 2478 error = SYSCTL_OUT(req, &xs, sizeof(xs)); 2479 return (error); 2480 } 2481 2482 SYSCTL_INT(_vm, OID_AUTO, nswapdev, CTLFLAG_RD, &nswapdev, 0, 2483 "Number of swap devices"); 2484 SYSCTL_NODE(_vm, OID_AUTO, swap_info, CTLFLAG_RD | CTLFLAG_MPSAFE, 2485 sysctl_vm_swap_info, 2486 "Swap statistics by device"); 2487 2488 /* 2489 * Count the approximate swap usage in pages for a vmspace. The 2490 * shadowed or not yet copied on write swap blocks are not accounted. 2491 * The map must be locked. 2492 */ 2493 long 2494 vmspace_swap_count(struct vmspace *vmspace) 2495 { 2496 vm_map_t map; 2497 vm_map_entry_t cur; 2498 vm_object_t object; 2499 struct swblk *sb; 2500 vm_pindex_t e, pi; 2501 long count; 2502 int i; 2503 2504 map = &vmspace->vm_map; 2505 count = 0; 2506 2507 for (cur = map->header.next; cur != &map->header; cur = cur->next) { 2508 if ((cur->eflags & MAP_ENTRY_IS_SUB_MAP) != 0) 2509 continue; 2510 object = cur->object.vm_object; 2511 if (object == NULL || object->type != OBJT_SWAP) 2512 continue; 2513 VM_OBJECT_RLOCK(object); 2514 if (object->type != OBJT_SWAP) 2515 goto unlock; 2516 pi = OFF_TO_IDX(cur->offset); 2517 e = pi + OFF_TO_IDX(cur->end - cur->start); 2518 for (;; pi = sb->p + SWAP_META_PAGES) { 2519 sb = SWAP_PCTRIE_LOOKUP_GE( 2520 &object->un_pager.swp.swp_blks, pi); 2521 if (sb == NULL || sb->p >= e) 2522 break; 2523 for (i = 0; i < SWAP_META_PAGES; i++) { 2524 if (sb->p + i < e && 2525 sb->d[i] != SWAPBLK_NONE) 2526 count++; 2527 } 2528 } 2529 unlock: 2530 VM_OBJECT_RUNLOCK(object); 2531 } 2532 return (count); 2533 } 2534 2535 /* 2536 * GEOM backend 2537 * 2538 * Swapping onto disk devices. 2539 * 2540 */ 2541 2542 static g_orphan_t swapgeom_orphan; 2543 2544 static struct g_class g_swap_class = { 2545 .name = "SWAP", 2546 .version = G_VERSION, 2547 .orphan = swapgeom_orphan, 2548 }; 2549 2550 DECLARE_GEOM_CLASS(g_swap_class, g_class); 2551 2552 2553 static void 2554 swapgeom_close_ev(void *arg, int flags) 2555 { 2556 struct g_consumer *cp; 2557 2558 cp = arg; 2559 g_access(cp, -1, -1, 0); 2560 g_detach(cp); 2561 g_destroy_consumer(cp); 2562 } 2563 2564 /* 2565 * Add a reference to the g_consumer for an inflight transaction. 2566 */ 2567 static void 2568 swapgeom_acquire(struct g_consumer *cp) 2569 { 2570 2571 mtx_assert(&sw_dev_mtx, MA_OWNED); 2572 cp->index++; 2573 } 2574 2575 /* 2576 * Remove a reference from the g_consumer. Post a close event if all 2577 * references go away, since the function might be called from the 2578 * biodone context. 2579 */ 2580 static void 2581 swapgeom_release(struct g_consumer *cp, struct swdevt *sp) 2582 { 2583 2584 mtx_assert(&sw_dev_mtx, MA_OWNED); 2585 cp->index--; 2586 if (cp->index == 0) { 2587 if (g_post_event(swapgeom_close_ev, cp, M_NOWAIT, NULL) == 0) 2588 sp->sw_id = NULL; 2589 } 2590 } 2591 2592 static void 2593 swapgeom_done(struct bio *bp2) 2594 { 2595 struct swdevt *sp; 2596 struct buf *bp; 2597 struct g_consumer *cp; 2598 2599 bp = bp2->bio_caller2; 2600 cp = bp2->bio_from; 2601 bp->b_ioflags = bp2->bio_flags; 2602 if (bp2->bio_error) 2603 bp->b_ioflags |= BIO_ERROR; 2604 bp->b_resid = bp->b_bcount - bp2->bio_completed; 2605 bp->b_error = bp2->bio_error; 2606 bufdone(bp); 2607 sp = bp2->bio_caller1; 2608 mtx_lock(&sw_dev_mtx); 2609 swapgeom_release(cp, sp); 2610 mtx_unlock(&sw_dev_mtx); 2611 g_destroy_bio(bp2); 2612 } 2613 2614 static void 2615 swapgeom_strategy(struct buf *bp, struct swdevt *sp) 2616 { 2617 struct bio *bio; 2618 struct g_consumer *cp; 2619 2620 mtx_lock(&sw_dev_mtx); 2621 cp = sp->sw_id; 2622 if (cp == NULL) { 2623 mtx_unlock(&sw_dev_mtx); 2624 bp->b_error = ENXIO; 2625 bp->b_ioflags |= BIO_ERROR; 2626 bufdone(bp); 2627 return; 2628 } 2629 swapgeom_acquire(cp); 2630 mtx_unlock(&sw_dev_mtx); 2631 if (bp->b_iocmd == BIO_WRITE) 2632 bio = g_new_bio(); 2633 else 2634 bio = g_alloc_bio(); 2635 if (bio == NULL) { 2636 mtx_lock(&sw_dev_mtx); 2637 swapgeom_release(cp, sp); 2638 mtx_unlock(&sw_dev_mtx); 2639 bp->b_error = ENOMEM; 2640 bp->b_ioflags |= BIO_ERROR; 2641 bufdone(bp); 2642 return; 2643 } 2644 2645 bio->bio_caller1 = sp; 2646 bio->bio_caller2 = bp; 2647 bio->bio_cmd = bp->b_iocmd; 2648 bio->bio_offset = (bp->b_blkno - sp->sw_first) * PAGE_SIZE; 2649 bio->bio_length = bp->b_bcount; 2650 bio->bio_done = swapgeom_done; 2651 if (!buf_mapped(bp)) { 2652 bio->bio_ma = bp->b_pages; 2653 bio->bio_data = unmapped_buf; 2654 bio->bio_ma_offset = (vm_offset_t)bp->b_offset & PAGE_MASK; 2655 bio->bio_ma_n = bp->b_npages; 2656 bio->bio_flags |= BIO_UNMAPPED; 2657 } else { 2658 bio->bio_data = bp->b_data; 2659 bio->bio_ma = NULL; 2660 } 2661 g_io_request(bio, cp); 2662 return; 2663 } 2664 2665 static void 2666 swapgeom_orphan(struct g_consumer *cp) 2667 { 2668 struct swdevt *sp; 2669 int destroy; 2670 2671 mtx_lock(&sw_dev_mtx); 2672 TAILQ_FOREACH(sp, &swtailq, sw_list) { 2673 if (sp->sw_id == cp) { 2674 sp->sw_flags |= SW_CLOSING; 2675 break; 2676 } 2677 } 2678 /* 2679 * Drop reference we were created with. Do directly since we're in a 2680 * special context where we don't have to queue the call to 2681 * swapgeom_close_ev(). 2682 */ 2683 cp->index--; 2684 destroy = ((sp != NULL) && (cp->index == 0)); 2685 if (destroy) 2686 sp->sw_id = NULL; 2687 mtx_unlock(&sw_dev_mtx); 2688 if (destroy) 2689 swapgeom_close_ev(cp, 0); 2690 } 2691 2692 static void 2693 swapgeom_close(struct thread *td, struct swdevt *sw) 2694 { 2695 struct g_consumer *cp; 2696 2697 mtx_lock(&sw_dev_mtx); 2698 cp = sw->sw_id; 2699 sw->sw_id = NULL; 2700 mtx_unlock(&sw_dev_mtx); 2701 2702 /* 2703 * swapgeom_close() may be called from the biodone context, 2704 * where we cannot perform topology changes. Delegate the 2705 * work to the events thread. 2706 */ 2707 if (cp != NULL) 2708 g_waitfor_event(swapgeom_close_ev, cp, M_WAITOK, NULL); 2709 } 2710 2711 static int 2712 swapongeom_locked(struct cdev *dev, struct vnode *vp) 2713 { 2714 struct g_provider *pp; 2715 struct g_consumer *cp; 2716 static struct g_geom *gp; 2717 struct swdevt *sp; 2718 u_long nblks; 2719 int error; 2720 2721 pp = g_dev_getprovider(dev); 2722 if (pp == NULL) 2723 return (ENODEV); 2724 mtx_lock(&sw_dev_mtx); 2725 TAILQ_FOREACH(sp, &swtailq, sw_list) { 2726 cp = sp->sw_id; 2727 if (cp != NULL && cp->provider == pp) { 2728 mtx_unlock(&sw_dev_mtx); 2729 return (EBUSY); 2730 } 2731 } 2732 mtx_unlock(&sw_dev_mtx); 2733 if (gp == NULL) 2734 gp = g_new_geomf(&g_swap_class, "swap"); 2735 cp = g_new_consumer(gp); 2736 cp->index = 1; /* Number of active I/Os, plus one for being active. */ 2737 cp->flags |= G_CF_DIRECT_SEND | G_CF_DIRECT_RECEIVE; 2738 g_attach(cp, pp); 2739 /* 2740 * XXX: Every time you think you can improve the margin for 2741 * footshooting, somebody depends on the ability to do so: 2742 * savecore(8) wants to write to our swapdev so we cannot 2743 * set an exclusive count :-( 2744 */ 2745 error = g_access(cp, 1, 1, 0); 2746 if (error != 0) { 2747 g_detach(cp); 2748 g_destroy_consumer(cp); 2749 return (error); 2750 } 2751 nblks = pp->mediasize / DEV_BSIZE; 2752 swaponsomething(vp, cp, nblks, swapgeom_strategy, 2753 swapgeom_close, dev2udev(dev), 2754 (pp->flags & G_PF_ACCEPT_UNMAPPED) != 0 ? SW_UNMAPPED : 0); 2755 return (0); 2756 } 2757 2758 static int 2759 swapongeom(struct vnode *vp) 2760 { 2761 int error; 2762 2763 vn_lock(vp, LK_EXCLUSIVE | LK_RETRY); 2764 if (vp->v_type != VCHR || (vp->v_iflag & VI_DOOMED) != 0) { 2765 error = ENOENT; 2766 } else { 2767 g_topology_lock(); 2768 error = swapongeom_locked(vp->v_rdev, vp); 2769 g_topology_unlock(); 2770 } 2771 VOP_UNLOCK(vp, 0); 2772 return (error); 2773 } 2774 2775 /* 2776 * VNODE backend 2777 * 2778 * This is used mainly for network filesystem (read: probably only tested 2779 * with NFS) swapfiles. 2780 * 2781 */ 2782 2783 static void 2784 swapdev_strategy(struct buf *bp, struct swdevt *sp) 2785 { 2786 struct vnode *vp2; 2787 2788 bp->b_blkno = ctodb(bp->b_blkno - sp->sw_first); 2789 2790 vp2 = sp->sw_id; 2791 vhold(vp2); 2792 if (bp->b_iocmd == BIO_WRITE) { 2793 if (bp->b_bufobj) 2794 bufobj_wdrop(bp->b_bufobj); 2795 bufobj_wref(&vp2->v_bufobj); 2796 } 2797 if (bp->b_bufobj != &vp2->v_bufobj) 2798 bp->b_bufobj = &vp2->v_bufobj; 2799 bp->b_vp = vp2; 2800 bp->b_iooffset = dbtob(bp->b_blkno); 2801 bstrategy(bp); 2802 return; 2803 } 2804 2805 static void 2806 swapdev_close(struct thread *td, struct swdevt *sp) 2807 { 2808 2809 VOP_CLOSE(sp->sw_vp, FREAD | FWRITE, td->td_ucred, td); 2810 vrele(sp->sw_vp); 2811 } 2812 2813 2814 static int 2815 swaponvp(struct thread *td, struct vnode *vp, u_long nblks) 2816 { 2817 struct swdevt *sp; 2818 int error; 2819 2820 if (nblks == 0) 2821 return (ENXIO); 2822 mtx_lock(&sw_dev_mtx); 2823 TAILQ_FOREACH(sp, &swtailq, sw_list) { 2824 if (sp->sw_id == vp) { 2825 mtx_unlock(&sw_dev_mtx); 2826 return (EBUSY); 2827 } 2828 } 2829 mtx_unlock(&sw_dev_mtx); 2830 2831 (void) vn_lock(vp, LK_EXCLUSIVE | LK_RETRY); 2832 #ifdef MAC 2833 error = mac_system_check_swapon(td->td_ucred, vp); 2834 if (error == 0) 2835 #endif 2836 error = VOP_OPEN(vp, FREAD | FWRITE, td->td_ucred, td, NULL); 2837 (void) VOP_UNLOCK(vp, 0); 2838 if (error) 2839 return (error); 2840 2841 swaponsomething(vp, vp, nblks, swapdev_strategy, swapdev_close, 2842 NODEV, 0); 2843 return (0); 2844 } 2845 2846 static int 2847 sysctl_swap_async_max(SYSCTL_HANDLER_ARGS) 2848 { 2849 int error, new, n; 2850 2851 new = nsw_wcount_async_max; 2852 error = sysctl_handle_int(oidp, &new, 0, req); 2853 if (error != 0 || req->newptr == NULL) 2854 return (error); 2855 2856 if (new > nswbuf / 2 || new < 1) 2857 return (EINVAL); 2858 2859 mtx_lock(&pbuf_mtx); 2860 while (nsw_wcount_async_max != new) { 2861 /* 2862 * Adjust difference. If the current async count is too low, 2863 * we will need to sqeeze our update slowly in. Sleep with a 2864 * higher priority than getpbuf() to finish faster. 2865 */ 2866 n = new - nsw_wcount_async_max; 2867 if (nsw_wcount_async + n >= 0) { 2868 nsw_wcount_async += n; 2869 nsw_wcount_async_max += n; 2870 wakeup(&nsw_wcount_async); 2871 } else { 2872 nsw_wcount_async_max -= nsw_wcount_async; 2873 nsw_wcount_async = 0; 2874 msleep(&nsw_wcount_async, &pbuf_mtx, PSWP, 2875 "swpsysctl", 0); 2876 } 2877 } 2878 mtx_unlock(&pbuf_mtx); 2879 2880 return (0); 2881 } 2882