1 /*- 2 * Copyright (c) 1998 Matthew Dillon, 3 * Copyright (c) 1994 John S. Dyson 4 * Copyright (c) 1990 University of Utah. 5 * Copyright (c) 1982, 1986, 1989, 1993 6 * The Regents of the University of California. All rights reserved. 7 * 8 * This code is derived from software contributed to Berkeley by 9 * the Systems Programming Group of the University of Utah Computer 10 * Science Department. 11 * 12 * Redistribution and use in source and binary forms, with or without 13 * modification, are permitted provided that the following conditions 14 * are met: 15 * 1. Redistributions of source code must retain the above copyright 16 * notice, this list of conditions and the following disclaimer. 17 * 2. Redistributions in binary form must reproduce the above copyright 18 * notice, this list of conditions and the following disclaimer in the 19 * documentation and/or other materials provided with the distribution. 20 * 3. All advertising materials mentioning features or use of this software 21 * must display the following acknowledgement: 22 * This product includes software developed by the University of 23 * California, Berkeley and its contributors. 24 * 4. Neither the name of the University nor the names of its contributors 25 * may be used to endorse or promote products derived from this software 26 * without specific prior written permission. 27 * 28 * THIS SOFTWARE IS PROVIDED BY THE REGENTS AND CONTRIBUTORS ``AS IS'' AND 29 * ANY EXPRESS OR IMPLIED WARRANTIES, INCLUDING, BUT NOT LIMITED TO, THE 30 * IMPLIED WARRANTIES OF MERCHANTABILITY AND FITNESS FOR A PARTICULAR PURPOSE 31 * ARE DISCLAIMED. IN NO EVENT SHALL THE REGENTS OR CONTRIBUTORS BE LIABLE 32 * FOR ANY DIRECT, INDIRECT, INCIDENTAL, SPECIAL, EXEMPLARY, OR CONSEQUENTIAL 33 * DAMAGES (INCLUDING, BUT NOT LIMITED TO, PROCUREMENT OF SUBSTITUTE GOODS 34 * OR SERVICES; LOSS OF USE, DATA, OR PROFITS; OR BUSINESS INTERRUPTION) 35 * HOWEVER CAUSED AND ON ANY THEORY OF LIABILITY, WHETHER IN CONTRACT, STRICT 36 * LIABILITY, OR TORT (INCLUDING NEGLIGENCE OR OTHERWISE) ARISING IN ANY WAY 37 * OUT OF THE USE OF THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF 38 * SUCH DAMAGE. 39 * 40 * New Swap System 41 * Matthew Dillon 42 * 43 * Radix Bitmap 'blists'. 44 * 45 * - The new swapper uses the new radix bitmap code. This should scale 46 * to arbitrarily small or arbitrarily large swap spaces and an almost 47 * arbitrary degree of fragmentation. 48 * 49 * Features: 50 * 51 * - on the fly reallocation of swap during putpages. The new system 52 * does not try to keep previously allocated swap blocks for dirty 53 * pages. 54 * 55 * - on the fly deallocation of swap 56 * 57 * - No more garbage collection required. Unnecessarily allocated swap 58 * blocks only exist for dirty vm_page_t's now and these are already 59 * cycled (in a high-load system) by the pager. We also do on-the-fly 60 * removal of invalidated swap blocks when a page is destroyed 61 * or renamed. 62 * 63 * from: Utah $Hdr: swap_pager.c 1.4 91/04/30$ 64 * 65 * @(#)swap_pager.c 8.9 (Berkeley) 3/21/94 66 * @(#)vm_swap.c 8.5 (Berkeley) 2/17/94 67 */ 68 69 #include <sys/cdefs.h> 70 __FBSDID("$FreeBSD$"); 71 72 #include "opt_compat.h" 73 #include "opt_swap.h" 74 #include "opt_vm.h" 75 76 #include <sys/param.h> 77 #include <sys/systm.h> 78 #include <sys/conf.h> 79 #include <sys/kernel.h> 80 #include <sys/priv.h> 81 #include <sys/proc.h> 82 #include <sys/bio.h> 83 #include <sys/buf.h> 84 #include <sys/disk.h> 85 #include <sys/fcntl.h> 86 #include <sys/mount.h> 87 #include <sys/namei.h> 88 #include <sys/vnode.h> 89 #include <sys/malloc.h> 90 #include <sys/racct.h> 91 #include <sys/resource.h> 92 #include <sys/resourcevar.h> 93 #include <sys/rwlock.h> 94 #include <sys/sysctl.h> 95 #include <sys/sysproto.h> 96 #include <sys/blist.h> 97 #include <sys/lock.h> 98 #include <sys/sx.h> 99 #include <sys/vmmeter.h> 100 101 #include <security/mac/mac_framework.h> 102 103 #include <vm/vm.h> 104 #include <vm/pmap.h> 105 #include <vm/vm_map.h> 106 #include <vm/vm_kern.h> 107 #include <vm/vm_object.h> 108 #include <vm/vm_page.h> 109 #include <vm/vm_pager.h> 110 #include <vm/vm_pageout.h> 111 #include <vm/vm_param.h> 112 #include <vm/swap_pager.h> 113 #include <vm/vm_extern.h> 114 #include <vm/uma.h> 115 116 #include <geom/geom.h> 117 118 /* 119 * SWB_NPAGES must be a power of 2. It may be set to 1, 2, 4, 8, 16 120 * or 32 pages per allocation. 121 * The 32-page limit is due to the radix code (kern/subr_blist.c). 122 */ 123 #ifndef MAX_PAGEOUT_CLUSTER 124 #define MAX_PAGEOUT_CLUSTER 16 125 #endif 126 127 #if !defined(SWB_NPAGES) 128 #define SWB_NPAGES MAX_PAGEOUT_CLUSTER 129 #endif 130 131 /* 132 * The swblock structure maps an object and a small, fixed-size range 133 * of page indices to disk addresses within a swap area. 134 * The collection of these mappings is implemented as a hash table. 135 * Unused disk addresses within a swap area are allocated and managed 136 * using a blist. 137 */ 138 #define SWCORRECT(n) (sizeof(void *) * (n) / sizeof(daddr_t)) 139 #define SWAP_META_PAGES (SWB_NPAGES * 2) 140 #define SWAP_META_MASK (SWAP_META_PAGES - 1) 141 142 struct swblock { 143 struct swblock *swb_hnext; 144 vm_object_t swb_object; 145 vm_pindex_t swb_index; 146 int swb_count; 147 daddr_t swb_pages[SWAP_META_PAGES]; 148 }; 149 150 static MALLOC_DEFINE(M_VMPGDATA, "vm_pgdata", "swap pager private data"); 151 static struct mtx sw_dev_mtx; 152 static TAILQ_HEAD(, swdevt) swtailq = TAILQ_HEAD_INITIALIZER(swtailq); 153 static struct swdevt *swdevhd; /* Allocate from here next */ 154 static int nswapdev; /* Number of swap devices */ 155 int swap_pager_avail; 156 static struct sx swdev_syscall_lock; /* serialize swap(on|off) */ 157 158 static vm_ooffset_t swap_total; 159 SYSCTL_QUAD(_vm, OID_AUTO, swap_total, CTLFLAG_RD, &swap_total, 0, 160 "Total amount of available swap storage."); 161 static vm_ooffset_t swap_reserved; 162 SYSCTL_QUAD(_vm, OID_AUTO, swap_reserved, CTLFLAG_RD, &swap_reserved, 0, 163 "Amount of swap storage needed to back all allocated anonymous memory."); 164 static int overcommit = 0; 165 SYSCTL_INT(_vm, OID_AUTO, overcommit, CTLFLAG_RW, &overcommit, 0, 166 "Configure virtual memory overcommit behavior. See tuning(7) " 167 "for details."); 168 static unsigned long swzone; 169 SYSCTL_ULONG(_vm, OID_AUTO, swzone, CTLFLAG_RD, &swzone, 0, 170 "Actual size of swap metadata zone"); 171 static unsigned long swap_maxpages; 172 SYSCTL_ULONG(_vm, OID_AUTO, swap_maxpages, CTLFLAG_RD, &swap_maxpages, 0, 173 "Maximum amount of swap supported"); 174 175 /* bits from overcommit */ 176 #define SWAP_RESERVE_FORCE_ON (1 << 0) 177 #define SWAP_RESERVE_RLIMIT_ON (1 << 1) 178 #define SWAP_RESERVE_ALLOW_NONWIRED (1 << 2) 179 180 int 181 swap_reserve(vm_ooffset_t incr) 182 { 183 184 return (swap_reserve_by_cred(incr, curthread->td_ucred)); 185 } 186 187 int 188 swap_reserve_by_cred(vm_ooffset_t incr, struct ucred *cred) 189 { 190 vm_ooffset_t r, s; 191 int res, error; 192 static int curfail; 193 static struct timeval lastfail; 194 struct uidinfo *uip; 195 196 uip = cred->cr_ruidinfo; 197 198 if (incr & PAGE_MASK) 199 panic("swap_reserve: & PAGE_MASK"); 200 201 #ifdef RACCT 202 if (racct_enable) { 203 PROC_LOCK(curproc); 204 error = racct_add(curproc, RACCT_SWAP, incr); 205 PROC_UNLOCK(curproc); 206 if (error != 0) 207 return (0); 208 } 209 #endif 210 211 res = 0; 212 mtx_lock(&sw_dev_mtx); 213 r = swap_reserved + incr; 214 if (overcommit & SWAP_RESERVE_ALLOW_NONWIRED) { 215 s = vm_cnt.v_page_count - vm_cnt.v_free_reserved - vm_cnt.v_wire_count; 216 s *= PAGE_SIZE; 217 } else 218 s = 0; 219 s += swap_total; 220 if ((overcommit & SWAP_RESERVE_FORCE_ON) == 0 || r <= s || 221 (error = priv_check(curthread, PRIV_VM_SWAP_NOQUOTA)) == 0) { 222 res = 1; 223 swap_reserved = r; 224 } 225 mtx_unlock(&sw_dev_mtx); 226 227 if (res) { 228 UIDINFO_VMSIZE_LOCK(uip); 229 if ((overcommit & SWAP_RESERVE_RLIMIT_ON) != 0 && 230 uip->ui_vmsize + incr > lim_cur(curthread, RLIMIT_SWAP) && 231 priv_check(curthread, PRIV_VM_SWAP_NORLIMIT)) 232 res = 0; 233 else 234 uip->ui_vmsize += incr; 235 UIDINFO_VMSIZE_UNLOCK(uip); 236 if (!res) { 237 mtx_lock(&sw_dev_mtx); 238 swap_reserved -= incr; 239 mtx_unlock(&sw_dev_mtx); 240 } 241 } 242 if (!res && ppsratecheck(&lastfail, &curfail, 1)) { 243 printf("uid %d, pid %d: swap reservation for %jd bytes failed\n", 244 uip->ui_uid, curproc->p_pid, incr); 245 } 246 247 #ifdef RACCT 248 if (!res) { 249 PROC_LOCK(curproc); 250 racct_sub(curproc, RACCT_SWAP, incr); 251 PROC_UNLOCK(curproc); 252 } 253 #endif 254 255 return (res); 256 } 257 258 void 259 swap_reserve_force(vm_ooffset_t incr) 260 { 261 struct uidinfo *uip; 262 263 mtx_lock(&sw_dev_mtx); 264 swap_reserved += incr; 265 mtx_unlock(&sw_dev_mtx); 266 267 #ifdef RACCT 268 PROC_LOCK(curproc); 269 racct_add_force(curproc, RACCT_SWAP, incr); 270 PROC_UNLOCK(curproc); 271 #endif 272 273 uip = curthread->td_ucred->cr_ruidinfo; 274 PROC_LOCK(curproc); 275 UIDINFO_VMSIZE_LOCK(uip); 276 uip->ui_vmsize += incr; 277 UIDINFO_VMSIZE_UNLOCK(uip); 278 PROC_UNLOCK(curproc); 279 } 280 281 void 282 swap_release(vm_ooffset_t decr) 283 { 284 struct ucred *cred; 285 286 PROC_LOCK(curproc); 287 cred = curthread->td_ucred; 288 swap_release_by_cred(decr, cred); 289 PROC_UNLOCK(curproc); 290 } 291 292 void 293 swap_release_by_cred(vm_ooffset_t decr, struct ucred *cred) 294 { 295 struct uidinfo *uip; 296 297 uip = cred->cr_ruidinfo; 298 299 if (decr & PAGE_MASK) 300 panic("swap_release: & PAGE_MASK"); 301 302 mtx_lock(&sw_dev_mtx); 303 if (swap_reserved < decr) 304 panic("swap_reserved < decr"); 305 swap_reserved -= decr; 306 mtx_unlock(&sw_dev_mtx); 307 308 UIDINFO_VMSIZE_LOCK(uip); 309 if (uip->ui_vmsize < decr) 310 printf("negative vmsize for uid = %d\n", uip->ui_uid); 311 uip->ui_vmsize -= decr; 312 UIDINFO_VMSIZE_UNLOCK(uip); 313 314 racct_sub_cred(cred, RACCT_SWAP, decr); 315 } 316 317 #define SWM_FREE 0x02 /* free, period */ 318 #define SWM_POP 0x04 /* pop out */ 319 320 int swap_pager_full = 2; /* swap space exhaustion (task killing) */ 321 static int swap_pager_almost_full = 1; /* swap space exhaustion (w/hysteresis)*/ 322 static int nsw_rcount; /* free read buffers */ 323 static int nsw_wcount_sync; /* limit write buffers / synchronous */ 324 static int nsw_wcount_async; /* limit write buffers / asynchronous */ 325 static int nsw_wcount_async_max;/* assigned maximum */ 326 static int nsw_cluster_max; /* maximum VOP I/O allowed */ 327 328 static int sysctl_swap_async_max(SYSCTL_HANDLER_ARGS); 329 SYSCTL_PROC(_vm, OID_AUTO, swap_async_max, CTLTYPE_INT | CTLFLAG_RW | 330 CTLFLAG_MPSAFE, NULL, 0, sysctl_swap_async_max, "I", 331 "Maximum running async swap ops"); 332 333 static struct swblock **swhash; 334 static int swhash_mask; 335 static struct mtx swhash_mtx; 336 337 static struct sx sw_alloc_sx; 338 339 /* 340 * "named" and "unnamed" anon region objects. Try to reduce the overhead 341 * of searching a named list by hashing it just a little. 342 */ 343 344 #define NOBJLISTS 8 345 346 #define NOBJLIST(handle) \ 347 (&swap_pager_object_list[((int)(intptr_t)handle >> 4) & (NOBJLISTS-1)]) 348 349 static struct pagerlst swap_pager_object_list[NOBJLISTS]; 350 static uma_zone_t swap_zone; 351 352 /* 353 * pagerops for OBJT_SWAP - "swap pager". Some ops are also global procedure 354 * calls hooked from other parts of the VM system and do not appear here. 355 * (see vm/swap_pager.h). 356 */ 357 static vm_object_t 358 swap_pager_alloc(void *handle, vm_ooffset_t size, 359 vm_prot_t prot, vm_ooffset_t offset, struct ucred *); 360 static void swap_pager_dealloc(vm_object_t object); 361 static int swap_pager_getpages(vm_object_t, vm_page_t *, int, int *, 362 int *); 363 static int swap_pager_getpages_async(vm_object_t, vm_page_t *, int, int *, 364 int *, pgo_getpages_iodone_t, void *); 365 static void swap_pager_putpages(vm_object_t, vm_page_t *, int, boolean_t, int *); 366 static boolean_t 367 swap_pager_haspage(vm_object_t object, vm_pindex_t pindex, int *before, int *after); 368 static void swap_pager_init(void); 369 static void swap_pager_unswapped(vm_page_t); 370 static void swap_pager_swapoff(struct swdevt *sp); 371 372 struct pagerops swappagerops = { 373 .pgo_init = swap_pager_init, /* early system initialization of pager */ 374 .pgo_alloc = swap_pager_alloc, /* allocate an OBJT_SWAP object */ 375 .pgo_dealloc = swap_pager_dealloc, /* deallocate an OBJT_SWAP object */ 376 .pgo_getpages = swap_pager_getpages, /* pagein */ 377 .pgo_getpages_async = swap_pager_getpages_async, /* pagein (async) */ 378 .pgo_putpages = swap_pager_putpages, /* pageout */ 379 .pgo_haspage = swap_pager_haspage, /* get backing store status for page */ 380 .pgo_pageunswapped = swap_pager_unswapped, /* remove swap related to page */ 381 }; 382 383 /* 384 * dmmax is in page-sized chunks with the new swap system. It was 385 * dev-bsized chunks in the old. dmmax is always a power of 2. 386 * 387 * swap_*() routines are externally accessible. swp_*() routines are 388 * internal. 389 */ 390 static int dmmax; 391 static int nswap_lowat = 128; /* in pages, swap_pager_almost_full warn */ 392 static int nswap_hiwat = 512; /* in pages, swap_pager_almost_full warn */ 393 394 SYSCTL_INT(_vm, OID_AUTO, dmmax, CTLFLAG_RD, &dmmax, 0, 395 "Maximum size of a swap block"); 396 397 static void swp_sizecheck(void); 398 static void swp_pager_async_iodone(struct buf *bp); 399 static int swapongeom(struct vnode *); 400 static int swaponvp(struct thread *, struct vnode *, u_long); 401 static int swapoff_one(struct swdevt *sp, struct ucred *cred); 402 403 /* 404 * Swap bitmap functions 405 */ 406 static void swp_pager_freeswapspace(daddr_t blk, int npages); 407 static daddr_t swp_pager_getswapspace(int npages); 408 409 /* 410 * Metadata functions 411 */ 412 static struct swblock **swp_pager_hash(vm_object_t object, vm_pindex_t index); 413 static void swp_pager_meta_build(vm_object_t, vm_pindex_t, daddr_t); 414 static void swp_pager_meta_free(vm_object_t, vm_pindex_t, vm_pindex_t); 415 static void swp_pager_meta_free_all(vm_object_t); 416 static daddr_t swp_pager_meta_ctl(vm_object_t, vm_pindex_t, int); 417 418 /* 419 * SWP_SIZECHECK() - update swap_pager_full indication 420 * 421 * update the swap_pager_almost_full indication and warn when we are 422 * about to run out of swap space, using lowat/hiwat hysteresis. 423 * 424 * Clear swap_pager_full ( task killing ) indication when lowat is met. 425 * 426 * No restrictions on call 427 * This routine may not block. 428 */ 429 static void 430 swp_sizecheck(void) 431 { 432 433 if (swap_pager_avail < nswap_lowat) { 434 if (swap_pager_almost_full == 0) { 435 printf("swap_pager: out of swap space\n"); 436 swap_pager_almost_full = 1; 437 } 438 } else { 439 swap_pager_full = 0; 440 if (swap_pager_avail > nswap_hiwat) 441 swap_pager_almost_full = 0; 442 } 443 } 444 445 /* 446 * SWP_PAGER_HASH() - hash swap meta data 447 * 448 * This is an helper function which hashes the swapblk given 449 * the object and page index. It returns a pointer to a pointer 450 * to the object, or a pointer to a NULL pointer if it could not 451 * find a swapblk. 452 */ 453 static struct swblock ** 454 swp_pager_hash(vm_object_t object, vm_pindex_t index) 455 { 456 struct swblock **pswap; 457 struct swblock *swap; 458 459 index &= ~(vm_pindex_t)SWAP_META_MASK; 460 pswap = &swhash[(index ^ (int)(intptr_t)object) & swhash_mask]; 461 while ((swap = *pswap) != NULL) { 462 if (swap->swb_object == object && 463 swap->swb_index == index 464 ) { 465 break; 466 } 467 pswap = &swap->swb_hnext; 468 } 469 return (pswap); 470 } 471 472 /* 473 * SWAP_PAGER_INIT() - initialize the swap pager! 474 * 475 * Expected to be started from system init. NOTE: This code is run 476 * before much else so be careful what you depend on. Most of the VM 477 * system has yet to be initialized at this point. 478 */ 479 static void 480 swap_pager_init(void) 481 { 482 /* 483 * Initialize object lists 484 */ 485 int i; 486 487 for (i = 0; i < NOBJLISTS; ++i) 488 TAILQ_INIT(&swap_pager_object_list[i]); 489 mtx_init(&sw_dev_mtx, "swapdev", NULL, MTX_DEF); 490 sx_init(&sw_alloc_sx, "swspsx"); 491 sx_init(&swdev_syscall_lock, "swsysc"); 492 493 /* 494 * Device Stripe, in PAGE_SIZE'd blocks 495 */ 496 dmmax = SWB_NPAGES * 2; 497 } 498 499 /* 500 * SWAP_PAGER_SWAP_INIT() - swap pager initialization from pageout process 501 * 502 * Expected to be started from pageout process once, prior to entering 503 * its main loop. 504 */ 505 void 506 swap_pager_swap_init(void) 507 { 508 unsigned long n, n2; 509 510 /* 511 * Number of in-transit swap bp operations. Don't 512 * exhaust the pbufs completely. Make sure we 513 * initialize workable values (0 will work for hysteresis 514 * but it isn't very efficient). 515 * 516 * The nsw_cluster_max is constrained by the bp->b_pages[] 517 * array (MAXPHYS/PAGE_SIZE) and our locally defined 518 * MAX_PAGEOUT_CLUSTER. Also be aware that swap ops are 519 * constrained by the swap device interleave stripe size. 520 * 521 * Currently we hardwire nsw_wcount_async to 4. This limit is 522 * designed to prevent other I/O from having high latencies due to 523 * our pageout I/O. The value 4 works well for one or two active swap 524 * devices but is probably a little low if you have more. Even so, 525 * a higher value would probably generate only a limited improvement 526 * with three or four active swap devices since the system does not 527 * typically have to pageout at extreme bandwidths. We will want 528 * at least 2 per swap devices, and 4 is a pretty good value if you 529 * have one NFS swap device due to the command/ack latency over NFS. 530 * So it all works out pretty well. 531 */ 532 nsw_cluster_max = min((MAXPHYS/PAGE_SIZE), MAX_PAGEOUT_CLUSTER); 533 534 mtx_lock(&pbuf_mtx); 535 nsw_rcount = (nswbuf + 1) / 2; 536 nsw_wcount_sync = (nswbuf + 3) / 4; 537 nsw_wcount_async = 4; 538 nsw_wcount_async_max = nsw_wcount_async; 539 mtx_unlock(&pbuf_mtx); 540 541 /* 542 * Initialize our zone. Right now I'm just guessing on the number 543 * we need based on the number of pages in the system. Each swblock 544 * can hold 32 pages, so this is probably overkill. This reservation 545 * is typically limited to around 32MB by default. 546 */ 547 n = vm_cnt.v_page_count / 2; 548 if (maxswzone && n > maxswzone / sizeof(struct swblock)) 549 n = maxswzone / sizeof(struct swblock); 550 n2 = n; 551 swap_zone = uma_zcreate("SWAPMETA", sizeof(struct swblock), NULL, NULL, 552 NULL, NULL, UMA_ALIGN_PTR, UMA_ZONE_NOFREE | UMA_ZONE_VM); 553 if (swap_zone == NULL) 554 panic("failed to create swap_zone."); 555 do { 556 if (uma_zone_reserve_kva(swap_zone, n)) 557 break; 558 /* 559 * if the allocation failed, try a zone two thirds the 560 * size of the previous attempt. 561 */ 562 n -= ((n + 2) / 3); 563 } while (n > 0); 564 if (n2 != n) 565 printf("Swap zone entries reduced from %lu to %lu.\n", n2, n); 566 swap_maxpages = n * SWAP_META_PAGES; 567 swzone = n * sizeof(struct swblock); 568 n2 = n; 569 570 /* 571 * Initialize our meta-data hash table. The swapper does not need to 572 * be quite as efficient as the VM system, so we do not use an 573 * oversized hash table. 574 * 575 * n: size of hash table, must be power of 2 576 * swhash_mask: hash table index mask 577 */ 578 for (n = 1; n < n2 / 8; n *= 2) 579 ; 580 swhash = malloc(sizeof(struct swblock *) * n, M_VMPGDATA, M_WAITOK | M_ZERO); 581 swhash_mask = n - 1; 582 mtx_init(&swhash_mtx, "swap_pager swhash", NULL, MTX_DEF); 583 } 584 585 static vm_object_t 586 swap_pager_alloc_init(void *handle, struct ucred *cred, vm_ooffset_t size, 587 vm_ooffset_t offset) 588 { 589 vm_object_t object; 590 591 if (cred != NULL) { 592 if (!swap_reserve_by_cred(size, cred)) 593 return (NULL); 594 crhold(cred); 595 } 596 object = vm_object_allocate(OBJT_SWAP, OFF_TO_IDX(offset + 597 PAGE_MASK + size)); 598 object->handle = handle; 599 if (cred != NULL) { 600 object->cred = cred; 601 object->charge = size; 602 } 603 object->un_pager.swp.swp_bcount = 0; 604 return (object); 605 } 606 607 /* 608 * SWAP_PAGER_ALLOC() - allocate a new OBJT_SWAP VM object and instantiate 609 * its metadata structures. 610 * 611 * This routine is called from the mmap and fork code to create a new 612 * OBJT_SWAP object. 613 * 614 * This routine must ensure that no live duplicate is created for 615 * the named object request, which is protected against by 616 * holding the sw_alloc_sx lock in case handle != NULL. 617 */ 618 static vm_object_t 619 swap_pager_alloc(void *handle, vm_ooffset_t size, vm_prot_t prot, 620 vm_ooffset_t offset, struct ucred *cred) 621 { 622 vm_object_t object; 623 624 if (handle != NULL) { 625 /* 626 * Reference existing named region or allocate new one. There 627 * should not be a race here against swp_pager_meta_build() 628 * as called from vm_page_remove() in regards to the lookup 629 * of the handle. 630 */ 631 sx_xlock(&sw_alloc_sx); 632 object = vm_pager_object_lookup(NOBJLIST(handle), handle); 633 if (object == NULL) { 634 object = swap_pager_alloc_init(handle, cred, size, 635 offset); 636 if (object != NULL) { 637 TAILQ_INSERT_TAIL(NOBJLIST(object->handle), 638 object, pager_object_list); 639 } 640 } 641 sx_xunlock(&sw_alloc_sx); 642 } else { 643 object = swap_pager_alloc_init(handle, cred, size, offset); 644 } 645 return (object); 646 } 647 648 /* 649 * SWAP_PAGER_DEALLOC() - remove swap metadata from object 650 * 651 * The swap backing for the object is destroyed. The code is 652 * designed such that we can reinstantiate it later, but this 653 * routine is typically called only when the entire object is 654 * about to be destroyed. 655 * 656 * The object must be locked. 657 */ 658 static void 659 swap_pager_dealloc(vm_object_t object) 660 { 661 662 VM_OBJECT_ASSERT_WLOCKED(object); 663 KASSERT((object->flags & OBJ_DEAD) != 0, ("dealloc of reachable obj")); 664 665 /* 666 * Remove from list right away so lookups will fail if we block for 667 * pageout completion. 668 */ 669 if (object->handle != NULL) { 670 VM_OBJECT_WUNLOCK(object); 671 sx_xlock(&sw_alloc_sx); 672 TAILQ_REMOVE(NOBJLIST(object->handle), object, 673 pager_object_list); 674 sx_xunlock(&sw_alloc_sx); 675 VM_OBJECT_WLOCK(object); 676 } 677 678 vm_object_pip_wait(object, "swpdea"); 679 680 /* 681 * Free all remaining metadata. We only bother to free it from 682 * the swap meta data. We do not attempt to free swapblk's still 683 * associated with vm_page_t's for this object. We do not care 684 * if paging is still in progress on some objects. 685 */ 686 swp_pager_meta_free_all(object); 687 object->handle = NULL; 688 object->type = OBJT_DEAD; 689 } 690 691 /************************************************************************ 692 * SWAP PAGER BITMAP ROUTINES * 693 ************************************************************************/ 694 695 /* 696 * SWP_PAGER_GETSWAPSPACE() - allocate raw swap space 697 * 698 * Allocate swap for the requested number of pages. The starting 699 * swap block number (a page index) is returned or SWAPBLK_NONE 700 * if the allocation failed. 701 * 702 * Also has the side effect of advising that somebody made a mistake 703 * when they configured swap and didn't configure enough. 704 * 705 * This routine may not sleep. 706 * 707 * We allocate in round-robin fashion from the configured devices. 708 */ 709 static daddr_t 710 swp_pager_getswapspace(int npages) 711 { 712 daddr_t blk; 713 struct swdevt *sp; 714 int i; 715 716 blk = SWAPBLK_NONE; 717 mtx_lock(&sw_dev_mtx); 718 sp = swdevhd; 719 for (i = 0; i < nswapdev; i++) { 720 if (sp == NULL) 721 sp = TAILQ_FIRST(&swtailq); 722 if (!(sp->sw_flags & SW_CLOSING)) { 723 blk = blist_alloc(sp->sw_blist, npages); 724 if (blk != SWAPBLK_NONE) { 725 blk += sp->sw_first; 726 sp->sw_used += npages; 727 swap_pager_avail -= npages; 728 swp_sizecheck(); 729 swdevhd = TAILQ_NEXT(sp, sw_list); 730 goto done; 731 } 732 } 733 sp = TAILQ_NEXT(sp, sw_list); 734 } 735 if (swap_pager_full != 2) { 736 printf("swap_pager_getswapspace(%d): failed\n", npages); 737 swap_pager_full = 2; 738 swap_pager_almost_full = 1; 739 } 740 swdevhd = NULL; 741 done: 742 mtx_unlock(&sw_dev_mtx); 743 return (blk); 744 } 745 746 static int 747 swp_pager_isondev(daddr_t blk, struct swdevt *sp) 748 { 749 750 return (blk >= sp->sw_first && blk < sp->sw_end); 751 } 752 753 static void 754 swp_pager_strategy(struct buf *bp) 755 { 756 struct swdevt *sp; 757 758 mtx_lock(&sw_dev_mtx); 759 TAILQ_FOREACH(sp, &swtailq, sw_list) { 760 if (bp->b_blkno >= sp->sw_first && bp->b_blkno < sp->sw_end) { 761 mtx_unlock(&sw_dev_mtx); 762 if ((sp->sw_flags & SW_UNMAPPED) != 0 && 763 unmapped_buf_allowed) { 764 bp->b_data = unmapped_buf; 765 bp->b_offset = 0; 766 } else { 767 pmap_qenter((vm_offset_t)bp->b_data, 768 &bp->b_pages[0], bp->b_bcount / PAGE_SIZE); 769 } 770 sp->sw_strategy(bp, sp); 771 return; 772 } 773 } 774 panic("Swapdev not found"); 775 } 776 777 778 /* 779 * SWP_PAGER_FREESWAPSPACE() - free raw swap space 780 * 781 * This routine returns the specified swap blocks back to the bitmap. 782 * 783 * This routine may not sleep. 784 */ 785 static void 786 swp_pager_freeswapspace(daddr_t blk, int npages) 787 { 788 struct swdevt *sp; 789 790 mtx_lock(&sw_dev_mtx); 791 TAILQ_FOREACH(sp, &swtailq, sw_list) { 792 if (blk >= sp->sw_first && blk < sp->sw_end) { 793 sp->sw_used -= npages; 794 /* 795 * If we are attempting to stop swapping on 796 * this device, we don't want to mark any 797 * blocks free lest they be reused. 798 */ 799 if ((sp->sw_flags & SW_CLOSING) == 0) { 800 blist_free(sp->sw_blist, blk - sp->sw_first, 801 npages); 802 swap_pager_avail += npages; 803 swp_sizecheck(); 804 } 805 mtx_unlock(&sw_dev_mtx); 806 return; 807 } 808 } 809 panic("Swapdev not found"); 810 } 811 812 /* 813 * SWAP_PAGER_FREESPACE() - frees swap blocks associated with a page 814 * range within an object. 815 * 816 * This is a globally accessible routine. 817 * 818 * This routine removes swapblk assignments from swap metadata. 819 * 820 * The external callers of this routine typically have already destroyed 821 * or renamed vm_page_t's associated with this range in the object so 822 * we should be ok. 823 * 824 * The object must be locked. 825 */ 826 void 827 swap_pager_freespace(vm_object_t object, vm_pindex_t start, vm_size_t size) 828 { 829 830 swp_pager_meta_free(object, start, size); 831 } 832 833 /* 834 * SWAP_PAGER_RESERVE() - reserve swap blocks in object 835 * 836 * Assigns swap blocks to the specified range within the object. The 837 * swap blocks are not zeroed. Any previous swap assignment is destroyed. 838 * 839 * Returns 0 on success, -1 on failure. 840 */ 841 int 842 swap_pager_reserve(vm_object_t object, vm_pindex_t start, vm_size_t size) 843 { 844 int n = 0; 845 daddr_t blk = SWAPBLK_NONE; 846 vm_pindex_t beg = start; /* save start index */ 847 848 VM_OBJECT_WLOCK(object); 849 while (size) { 850 if (n == 0) { 851 n = BLIST_MAX_ALLOC; 852 while ((blk = swp_pager_getswapspace(n)) == SWAPBLK_NONE) { 853 n >>= 1; 854 if (n == 0) { 855 swp_pager_meta_free(object, beg, start - beg); 856 VM_OBJECT_WUNLOCK(object); 857 return (-1); 858 } 859 } 860 } 861 swp_pager_meta_build(object, start, blk); 862 --size; 863 ++start; 864 ++blk; 865 --n; 866 } 867 swp_pager_meta_free(object, start, n); 868 VM_OBJECT_WUNLOCK(object); 869 return (0); 870 } 871 872 /* 873 * SWAP_PAGER_COPY() - copy blocks from source pager to destination pager 874 * and destroy the source. 875 * 876 * Copy any valid swapblks from the source to the destination. In 877 * cases where both the source and destination have a valid swapblk, 878 * we keep the destination's. 879 * 880 * This routine is allowed to sleep. It may sleep allocating metadata 881 * indirectly through swp_pager_meta_build() or if paging is still in 882 * progress on the source. 883 * 884 * The source object contains no vm_page_t's (which is just as well) 885 * 886 * The source object is of type OBJT_SWAP. 887 * 888 * The source and destination objects must be locked. 889 * Both object locks may temporarily be released. 890 */ 891 void 892 swap_pager_copy(vm_object_t srcobject, vm_object_t dstobject, 893 vm_pindex_t offset, int destroysource) 894 { 895 vm_pindex_t i; 896 897 VM_OBJECT_ASSERT_WLOCKED(srcobject); 898 VM_OBJECT_ASSERT_WLOCKED(dstobject); 899 900 /* 901 * If destroysource is set, we remove the source object from the 902 * swap_pager internal queue now. 903 */ 904 if (destroysource && srcobject->handle != NULL) { 905 vm_object_pip_add(srcobject, 1); 906 VM_OBJECT_WUNLOCK(srcobject); 907 vm_object_pip_add(dstobject, 1); 908 VM_OBJECT_WUNLOCK(dstobject); 909 sx_xlock(&sw_alloc_sx); 910 TAILQ_REMOVE(NOBJLIST(srcobject->handle), srcobject, 911 pager_object_list); 912 sx_xunlock(&sw_alloc_sx); 913 VM_OBJECT_WLOCK(dstobject); 914 vm_object_pip_wakeup(dstobject); 915 VM_OBJECT_WLOCK(srcobject); 916 vm_object_pip_wakeup(srcobject); 917 } 918 919 /* 920 * transfer source to destination. 921 */ 922 for (i = 0; i < dstobject->size; ++i) { 923 daddr_t dstaddr; 924 925 /* 926 * Locate (without changing) the swapblk on the destination, 927 * unless it is invalid in which case free it silently, or 928 * if the destination is a resident page, in which case the 929 * source is thrown away. 930 */ 931 dstaddr = swp_pager_meta_ctl(dstobject, i, 0); 932 933 if (dstaddr == SWAPBLK_NONE) { 934 /* 935 * Destination has no swapblk and is not resident, 936 * copy source. 937 */ 938 daddr_t srcaddr; 939 940 srcaddr = swp_pager_meta_ctl( 941 srcobject, 942 i + offset, 943 SWM_POP 944 ); 945 946 if (srcaddr != SWAPBLK_NONE) { 947 /* 948 * swp_pager_meta_build() can sleep. 949 */ 950 vm_object_pip_add(srcobject, 1); 951 VM_OBJECT_WUNLOCK(srcobject); 952 vm_object_pip_add(dstobject, 1); 953 swp_pager_meta_build(dstobject, i, srcaddr); 954 vm_object_pip_wakeup(dstobject); 955 VM_OBJECT_WLOCK(srcobject); 956 vm_object_pip_wakeup(srcobject); 957 } 958 } else { 959 /* 960 * Destination has valid swapblk or it is represented 961 * by a resident page. We destroy the sourceblock. 962 */ 963 964 swp_pager_meta_ctl(srcobject, i + offset, SWM_FREE); 965 } 966 } 967 968 /* 969 * Free left over swap blocks in source. 970 * 971 * We have to revert the type to OBJT_DEFAULT so we do not accidentally 972 * double-remove the object from the swap queues. 973 */ 974 if (destroysource) { 975 swp_pager_meta_free_all(srcobject); 976 /* 977 * Reverting the type is not necessary, the caller is going 978 * to destroy srcobject directly, but I'm doing it here 979 * for consistency since we've removed the object from its 980 * queues. 981 */ 982 srcobject->type = OBJT_DEFAULT; 983 } 984 } 985 986 /* 987 * SWAP_PAGER_HASPAGE() - determine if we have good backing store for 988 * the requested page. 989 * 990 * We determine whether good backing store exists for the requested 991 * page and return TRUE if it does, FALSE if it doesn't. 992 * 993 * If TRUE, we also try to determine how much valid, contiguous backing 994 * store exists before and after the requested page. 995 */ 996 static boolean_t 997 swap_pager_haspage(vm_object_t object, vm_pindex_t pindex, int *before, 998 int *after) 999 { 1000 daddr_t blk, blk0; 1001 int i; 1002 1003 VM_OBJECT_ASSERT_LOCKED(object); 1004 1005 /* 1006 * do we have good backing store at the requested index ? 1007 */ 1008 blk0 = swp_pager_meta_ctl(object, pindex, 0); 1009 if (blk0 == SWAPBLK_NONE) { 1010 if (before) 1011 *before = 0; 1012 if (after) 1013 *after = 0; 1014 return (FALSE); 1015 } 1016 1017 /* 1018 * find backwards-looking contiguous good backing store 1019 */ 1020 if (before != NULL) { 1021 for (i = 1; i < SWB_NPAGES; i++) { 1022 if (i > pindex) 1023 break; 1024 blk = swp_pager_meta_ctl(object, pindex - i, 0); 1025 if (blk != blk0 - i) 1026 break; 1027 } 1028 *before = i - 1; 1029 } 1030 1031 /* 1032 * find forward-looking contiguous good backing store 1033 */ 1034 if (after != NULL) { 1035 for (i = 1; i < SWB_NPAGES; i++) { 1036 blk = swp_pager_meta_ctl(object, pindex + i, 0); 1037 if (blk != blk0 + i) 1038 break; 1039 } 1040 *after = i - 1; 1041 } 1042 return (TRUE); 1043 } 1044 1045 /* 1046 * SWAP_PAGER_PAGE_UNSWAPPED() - remove swap backing store related to page 1047 * 1048 * This removes any associated swap backing store, whether valid or 1049 * not, from the page. 1050 * 1051 * This routine is typically called when a page is made dirty, at 1052 * which point any associated swap can be freed. MADV_FREE also 1053 * calls us in a special-case situation 1054 * 1055 * NOTE!!! If the page is clean and the swap was valid, the caller 1056 * should make the page dirty before calling this routine. This routine 1057 * does NOT change the m->dirty status of the page. Also: MADV_FREE 1058 * depends on it. 1059 * 1060 * This routine may not sleep. 1061 * 1062 * The object containing the page must be locked. 1063 */ 1064 static void 1065 swap_pager_unswapped(vm_page_t m) 1066 { 1067 1068 swp_pager_meta_ctl(m->object, m->pindex, SWM_FREE); 1069 } 1070 1071 /* 1072 * swap_pager_getpages() - bring pages in from swap 1073 * 1074 * Attempt to page in the pages in array "m" of length "count". The caller 1075 * may optionally specify that additional pages preceding and succeeding 1076 * the specified range be paged in. The number of such pages is returned 1077 * in the "rbehind" and "rahead" parameters, and they will be in the 1078 * inactive queue upon return. 1079 * 1080 * The pages in "m" must be busied and will remain busied upon return. 1081 */ 1082 static int 1083 swap_pager_getpages(vm_object_t object, vm_page_t *m, int count, int *rbehind, 1084 int *rahead) 1085 { 1086 struct buf *bp; 1087 vm_page_t mpred, msucc, p; 1088 vm_pindex_t pindex; 1089 daddr_t blk; 1090 int i, j, maxahead, maxbehind, reqcount, shift; 1091 1092 reqcount = count; 1093 1094 VM_OBJECT_WUNLOCK(object); 1095 bp = getpbuf(&nsw_rcount); 1096 VM_OBJECT_WLOCK(object); 1097 1098 if (!swap_pager_haspage(object, m[0]->pindex, &maxbehind, &maxahead)) { 1099 relpbuf(bp, &nsw_rcount); 1100 return (VM_PAGER_FAIL); 1101 } 1102 1103 /* 1104 * Clip the readahead and readbehind ranges to exclude resident pages. 1105 */ 1106 if (rahead != NULL) { 1107 KASSERT(reqcount - 1 <= maxahead, 1108 ("page count %d extends beyond swap block", reqcount)); 1109 *rahead = imin(*rahead, maxahead - (reqcount - 1)); 1110 pindex = m[reqcount - 1]->pindex; 1111 msucc = TAILQ_NEXT(m[reqcount - 1], listq); 1112 if (msucc != NULL && msucc->pindex - pindex - 1 < *rahead) 1113 *rahead = msucc->pindex - pindex - 1; 1114 } 1115 if (rbehind != NULL) { 1116 *rbehind = imin(*rbehind, maxbehind); 1117 pindex = m[0]->pindex; 1118 mpred = TAILQ_PREV(m[0], pglist, listq); 1119 if (mpred != NULL && pindex - mpred->pindex - 1 < *rbehind) 1120 *rbehind = pindex - mpred->pindex - 1; 1121 } 1122 1123 /* 1124 * Allocate readahead and readbehind pages. 1125 */ 1126 shift = rbehind != NULL ? *rbehind : 0; 1127 if (shift != 0) { 1128 for (i = 1; i <= shift; i++) { 1129 p = vm_page_alloc(object, m[0]->pindex - i, 1130 VM_ALLOC_NORMAL); 1131 if (p == NULL) { 1132 /* Shift allocated pages to the left. */ 1133 for (j = 0; j < i - 1; j++) 1134 bp->b_pages[j] = 1135 bp->b_pages[j + shift - i + 1]; 1136 break; 1137 } 1138 bp->b_pages[shift - i] = p; 1139 } 1140 shift = i - 1; 1141 *rbehind = shift; 1142 } 1143 for (i = 0; i < reqcount; i++) 1144 bp->b_pages[i + shift] = m[i]; 1145 if (rahead != NULL) { 1146 for (i = 0; i < *rahead; i++) { 1147 p = vm_page_alloc(object, 1148 m[reqcount - 1]->pindex + i + 1, VM_ALLOC_NORMAL); 1149 if (p == NULL) 1150 break; 1151 bp->b_pages[shift + reqcount + i] = p; 1152 } 1153 *rahead = i; 1154 } 1155 if (rbehind != NULL) 1156 count += *rbehind; 1157 if (rahead != NULL) 1158 count += *rahead; 1159 1160 vm_object_pip_add(object, count); 1161 1162 for (i = 0; i < count; i++) 1163 bp->b_pages[i]->oflags |= VPO_SWAPINPROG; 1164 1165 pindex = bp->b_pages[0]->pindex; 1166 blk = swp_pager_meta_ctl(object, pindex, 0); 1167 KASSERT(blk != SWAPBLK_NONE, 1168 ("no swap blocking containing %p(%jx)", object, (uintmax_t)pindex)); 1169 1170 VM_OBJECT_WUNLOCK(object); 1171 1172 bp->b_flags |= B_PAGING; 1173 bp->b_iocmd = BIO_READ; 1174 bp->b_iodone = swp_pager_async_iodone; 1175 bp->b_rcred = crhold(thread0.td_ucred); 1176 bp->b_wcred = crhold(thread0.td_ucred); 1177 bp->b_blkno = blk; 1178 bp->b_bcount = PAGE_SIZE * count; 1179 bp->b_bufsize = PAGE_SIZE * count; 1180 bp->b_npages = count; 1181 bp->b_pgbefore = rbehind != NULL ? *rbehind : 0; 1182 bp->b_pgafter = rahead != NULL ? *rahead : 0; 1183 1184 VM_CNT_INC(v_swapin); 1185 VM_CNT_ADD(v_swappgsin, count); 1186 1187 /* 1188 * perform the I/O. NOTE!!! bp cannot be considered valid after 1189 * this point because we automatically release it on completion. 1190 * Instead, we look at the one page we are interested in which we 1191 * still hold a lock on even through the I/O completion. 1192 * 1193 * The other pages in our m[] array are also released on completion, 1194 * so we cannot assume they are valid anymore either. 1195 * 1196 * NOTE: b_blkno is destroyed by the call to swapdev_strategy 1197 */ 1198 BUF_KERNPROC(bp); 1199 swp_pager_strategy(bp); 1200 1201 /* 1202 * Wait for the pages we want to complete. VPO_SWAPINPROG is always 1203 * cleared on completion. If an I/O error occurs, SWAPBLK_NONE 1204 * is set in the metadata for each page in the request. 1205 */ 1206 VM_OBJECT_WLOCK(object); 1207 while ((m[0]->oflags & VPO_SWAPINPROG) != 0) { 1208 m[0]->oflags |= VPO_SWAPSLEEP; 1209 VM_CNT_INC(v_intrans); 1210 if (VM_OBJECT_SLEEP(object, &object->paging_in_progress, PSWP, 1211 "swread", hz * 20)) { 1212 printf( 1213 "swap_pager: indefinite wait buffer: bufobj: %p, blkno: %jd, size: %ld\n", 1214 bp->b_bufobj, (intmax_t)bp->b_blkno, bp->b_bcount); 1215 } 1216 } 1217 1218 /* 1219 * If we had an unrecoverable read error pages will not be valid. 1220 */ 1221 for (i = 0; i < reqcount; i++) 1222 if (m[i]->valid != VM_PAGE_BITS_ALL) 1223 return (VM_PAGER_ERROR); 1224 1225 return (VM_PAGER_OK); 1226 1227 /* 1228 * A final note: in a low swap situation, we cannot deallocate swap 1229 * and mark a page dirty here because the caller is likely to mark 1230 * the page clean when we return, causing the page to possibly revert 1231 * to all-zero's later. 1232 */ 1233 } 1234 1235 /* 1236 * swap_pager_getpages_async(): 1237 * 1238 * Right now this is emulation of asynchronous operation on top of 1239 * swap_pager_getpages(). 1240 */ 1241 static int 1242 swap_pager_getpages_async(vm_object_t object, vm_page_t *m, int count, 1243 int *rbehind, int *rahead, pgo_getpages_iodone_t iodone, void *arg) 1244 { 1245 int r, error; 1246 1247 r = swap_pager_getpages(object, m, count, rbehind, rahead); 1248 VM_OBJECT_WUNLOCK(object); 1249 switch (r) { 1250 case VM_PAGER_OK: 1251 error = 0; 1252 break; 1253 case VM_PAGER_ERROR: 1254 error = EIO; 1255 break; 1256 case VM_PAGER_FAIL: 1257 error = EINVAL; 1258 break; 1259 default: 1260 panic("unhandled swap_pager_getpages() error %d", r); 1261 } 1262 (iodone)(arg, m, count, error); 1263 VM_OBJECT_WLOCK(object); 1264 1265 return (r); 1266 } 1267 1268 /* 1269 * swap_pager_putpages: 1270 * 1271 * Assign swap (if necessary) and initiate I/O on the specified pages. 1272 * 1273 * We support both OBJT_DEFAULT and OBJT_SWAP objects. DEFAULT objects 1274 * are automatically converted to SWAP objects. 1275 * 1276 * In a low memory situation we may block in VOP_STRATEGY(), but the new 1277 * vm_page reservation system coupled with properly written VFS devices 1278 * should ensure that no low-memory deadlock occurs. This is an area 1279 * which needs work. 1280 * 1281 * The parent has N vm_object_pip_add() references prior to 1282 * calling us and will remove references for rtvals[] that are 1283 * not set to VM_PAGER_PEND. We need to remove the rest on I/O 1284 * completion. 1285 * 1286 * The parent has soft-busy'd the pages it passes us and will unbusy 1287 * those whos rtvals[] entry is not set to VM_PAGER_PEND on return. 1288 * We need to unbusy the rest on I/O completion. 1289 */ 1290 static void 1291 swap_pager_putpages(vm_object_t object, vm_page_t *m, int count, 1292 int flags, int *rtvals) 1293 { 1294 int i, n; 1295 boolean_t sync; 1296 1297 if (count && m[0]->object != object) { 1298 panic("swap_pager_putpages: object mismatch %p/%p", 1299 object, 1300 m[0]->object 1301 ); 1302 } 1303 1304 /* 1305 * Step 1 1306 * 1307 * Turn object into OBJT_SWAP 1308 * check for bogus sysops 1309 * force sync if not pageout process 1310 */ 1311 if (object->type != OBJT_SWAP) 1312 swp_pager_meta_build(object, 0, SWAPBLK_NONE); 1313 VM_OBJECT_WUNLOCK(object); 1314 1315 n = 0; 1316 if (curproc != pageproc) 1317 sync = TRUE; 1318 else 1319 sync = (flags & VM_PAGER_PUT_SYNC) != 0; 1320 1321 /* 1322 * Step 2 1323 * 1324 * Assign swap blocks and issue I/O. We reallocate swap on the fly. 1325 * The page is left dirty until the pageout operation completes 1326 * successfully. 1327 */ 1328 for (i = 0; i < count; i += n) { 1329 int j; 1330 struct buf *bp; 1331 daddr_t blk; 1332 1333 /* 1334 * Maximum I/O size is limited by a number of factors. 1335 */ 1336 n = min(BLIST_MAX_ALLOC, count - i); 1337 n = min(n, nsw_cluster_max); 1338 1339 /* 1340 * Get biggest block of swap we can. If we fail, fall 1341 * back and try to allocate a smaller block. Don't go 1342 * overboard trying to allocate space if it would overly 1343 * fragment swap. 1344 */ 1345 while ( 1346 (blk = swp_pager_getswapspace(n)) == SWAPBLK_NONE && 1347 n > 4 1348 ) { 1349 n >>= 1; 1350 } 1351 if (blk == SWAPBLK_NONE) { 1352 for (j = 0; j < n; ++j) 1353 rtvals[i+j] = VM_PAGER_FAIL; 1354 continue; 1355 } 1356 1357 /* 1358 * All I/O parameters have been satisfied, build the I/O 1359 * request and assign the swap space. 1360 */ 1361 if (sync == TRUE) { 1362 bp = getpbuf(&nsw_wcount_sync); 1363 } else { 1364 bp = getpbuf(&nsw_wcount_async); 1365 bp->b_flags = B_ASYNC; 1366 } 1367 bp->b_flags |= B_PAGING; 1368 bp->b_iocmd = BIO_WRITE; 1369 1370 bp->b_rcred = crhold(thread0.td_ucred); 1371 bp->b_wcred = crhold(thread0.td_ucred); 1372 bp->b_bcount = PAGE_SIZE * n; 1373 bp->b_bufsize = PAGE_SIZE * n; 1374 bp->b_blkno = blk; 1375 1376 VM_OBJECT_WLOCK(object); 1377 for (j = 0; j < n; ++j) { 1378 vm_page_t mreq = m[i+j]; 1379 1380 swp_pager_meta_build( 1381 mreq->object, 1382 mreq->pindex, 1383 blk + j 1384 ); 1385 vm_page_dirty(mreq); 1386 mreq->oflags |= VPO_SWAPINPROG; 1387 bp->b_pages[j] = mreq; 1388 } 1389 VM_OBJECT_WUNLOCK(object); 1390 bp->b_npages = n; 1391 /* 1392 * Must set dirty range for NFS to work. 1393 */ 1394 bp->b_dirtyoff = 0; 1395 bp->b_dirtyend = bp->b_bcount; 1396 1397 VM_CNT_INC(v_swapout); 1398 VM_CNT_ADD(v_swappgsout, bp->b_npages); 1399 1400 /* 1401 * We unconditionally set rtvals[] to VM_PAGER_PEND so that we 1402 * can call the async completion routine at the end of a 1403 * synchronous I/O operation. Otherwise, our caller would 1404 * perform duplicate unbusy and wakeup operations on the page 1405 * and object, respectively. 1406 */ 1407 for (j = 0; j < n; j++) 1408 rtvals[i + j] = VM_PAGER_PEND; 1409 1410 /* 1411 * asynchronous 1412 * 1413 * NOTE: b_blkno is destroyed by the call to swapdev_strategy 1414 */ 1415 if (sync == FALSE) { 1416 bp->b_iodone = swp_pager_async_iodone; 1417 BUF_KERNPROC(bp); 1418 swp_pager_strategy(bp); 1419 continue; 1420 } 1421 1422 /* 1423 * synchronous 1424 * 1425 * NOTE: b_blkno is destroyed by the call to swapdev_strategy 1426 */ 1427 bp->b_iodone = bdone; 1428 swp_pager_strategy(bp); 1429 1430 /* 1431 * Wait for the sync I/O to complete. 1432 */ 1433 bwait(bp, PVM, "swwrt"); 1434 1435 /* 1436 * Now that we are through with the bp, we can call the 1437 * normal async completion, which frees everything up. 1438 */ 1439 swp_pager_async_iodone(bp); 1440 } 1441 VM_OBJECT_WLOCK(object); 1442 } 1443 1444 /* 1445 * swp_pager_async_iodone: 1446 * 1447 * Completion routine for asynchronous reads and writes from/to swap. 1448 * Also called manually by synchronous code to finish up a bp. 1449 * 1450 * This routine may not sleep. 1451 */ 1452 static void 1453 swp_pager_async_iodone(struct buf *bp) 1454 { 1455 int i; 1456 vm_object_t object = NULL; 1457 1458 /* 1459 * report error 1460 */ 1461 if (bp->b_ioflags & BIO_ERROR) { 1462 printf( 1463 "swap_pager: I/O error - %s failed; blkno %ld," 1464 "size %ld, error %d\n", 1465 ((bp->b_iocmd == BIO_READ) ? "pagein" : "pageout"), 1466 (long)bp->b_blkno, 1467 (long)bp->b_bcount, 1468 bp->b_error 1469 ); 1470 } 1471 1472 /* 1473 * remove the mapping for kernel virtual 1474 */ 1475 if (buf_mapped(bp)) 1476 pmap_qremove((vm_offset_t)bp->b_data, bp->b_npages); 1477 else 1478 bp->b_data = bp->b_kvabase; 1479 1480 if (bp->b_npages) { 1481 object = bp->b_pages[0]->object; 1482 VM_OBJECT_WLOCK(object); 1483 } 1484 1485 /* 1486 * cleanup pages. If an error occurs writing to swap, we are in 1487 * very serious trouble. If it happens to be a disk error, though, 1488 * we may be able to recover by reassigning the swap later on. So 1489 * in this case we remove the m->swapblk assignment for the page 1490 * but do not free it in the rlist. The errornous block(s) are thus 1491 * never reallocated as swap. Redirty the page and continue. 1492 */ 1493 for (i = 0; i < bp->b_npages; ++i) { 1494 vm_page_t m = bp->b_pages[i]; 1495 1496 m->oflags &= ~VPO_SWAPINPROG; 1497 if (m->oflags & VPO_SWAPSLEEP) { 1498 m->oflags &= ~VPO_SWAPSLEEP; 1499 wakeup(&object->paging_in_progress); 1500 } 1501 1502 if (bp->b_ioflags & BIO_ERROR) { 1503 /* 1504 * If an error occurs I'd love to throw the swapblk 1505 * away without freeing it back to swapspace, so it 1506 * can never be used again. But I can't from an 1507 * interrupt. 1508 */ 1509 if (bp->b_iocmd == BIO_READ) { 1510 /* 1511 * NOTE: for reads, m->dirty will probably 1512 * be overridden by the original caller of 1513 * getpages so don't play cute tricks here. 1514 */ 1515 m->valid = 0; 1516 } else { 1517 /* 1518 * If a write error occurs, reactivate page 1519 * so it doesn't clog the inactive list, 1520 * then finish the I/O. 1521 */ 1522 vm_page_dirty(m); 1523 vm_page_lock(m); 1524 vm_page_activate(m); 1525 vm_page_unlock(m); 1526 vm_page_sunbusy(m); 1527 } 1528 } else if (bp->b_iocmd == BIO_READ) { 1529 /* 1530 * NOTE: for reads, m->dirty will probably be 1531 * overridden by the original caller of getpages so 1532 * we cannot set them in order to free the underlying 1533 * swap in a low-swap situation. I don't think we'd 1534 * want to do that anyway, but it was an optimization 1535 * that existed in the old swapper for a time before 1536 * it got ripped out due to precisely this problem. 1537 */ 1538 KASSERT(!pmap_page_is_mapped(m), 1539 ("swp_pager_async_iodone: page %p is mapped", m)); 1540 KASSERT(m->dirty == 0, 1541 ("swp_pager_async_iodone: page %p is dirty", m)); 1542 1543 m->valid = VM_PAGE_BITS_ALL; 1544 if (i < bp->b_pgbefore || 1545 i >= bp->b_npages - bp->b_pgafter) 1546 vm_page_readahead_finish(m); 1547 } else { 1548 /* 1549 * For write success, clear the dirty 1550 * status, then finish the I/O ( which decrements the 1551 * busy count and possibly wakes waiter's up ). 1552 * A page is only written to swap after a period of 1553 * inactivity. Therefore, we do not expect it to be 1554 * reused. 1555 */ 1556 KASSERT(!pmap_page_is_write_mapped(m), 1557 ("swp_pager_async_iodone: page %p is not write" 1558 " protected", m)); 1559 vm_page_undirty(m); 1560 vm_page_lock(m); 1561 vm_page_deactivate_noreuse(m); 1562 vm_page_unlock(m); 1563 vm_page_sunbusy(m); 1564 } 1565 } 1566 1567 /* 1568 * adjust pip. NOTE: the original parent may still have its own 1569 * pip refs on the object. 1570 */ 1571 if (object != NULL) { 1572 vm_object_pip_wakeupn(object, bp->b_npages); 1573 VM_OBJECT_WUNLOCK(object); 1574 } 1575 1576 /* 1577 * swapdev_strategy() manually sets b_vp and b_bufobj before calling 1578 * bstrategy(). Set them back to NULL now we're done with it, or we'll 1579 * trigger a KASSERT in relpbuf(). 1580 */ 1581 if (bp->b_vp) { 1582 bp->b_vp = NULL; 1583 bp->b_bufobj = NULL; 1584 } 1585 /* 1586 * release the physical I/O buffer 1587 */ 1588 relpbuf( 1589 bp, 1590 ((bp->b_iocmd == BIO_READ) ? &nsw_rcount : 1591 ((bp->b_flags & B_ASYNC) ? 1592 &nsw_wcount_async : 1593 &nsw_wcount_sync 1594 ) 1595 ) 1596 ); 1597 } 1598 1599 /* 1600 * swap_pager_isswapped: 1601 * 1602 * Return 1 if at least one page in the given object is paged 1603 * out to the given swap device. 1604 * 1605 * This routine may not sleep. 1606 */ 1607 int 1608 swap_pager_isswapped(vm_object_t object, struct swdevt *sp) 1609 { 1610 daddr_t index = 0; 1611 int bcount; 1612 int i; 1613 1614 VM_OBJECT_ASSERT_WLOCKED(object); 1615 if (object->type != OBJT_SWAP) 1616 return (0); 1617 1618 mtx_lock(&swhash_mtx); 1619 for (bcount = 0; bcount < object->un_pager.swp.swp_bcount; bcount++) { 1620 struct swblock *swap; 1621 1622 if ((swap = *swp_pager_hash(object, index)) != NULL) { 1623 for (i = 0; i < SWAP_META_PAGES; ++i) { 1624 if (swp_pager_isondev(swap->swb_pages[i], sp)) { 1625 mtx_unlock(&swhash_mtx); 1626 return (1); 1627 } 1628 } 1629 } 1630 index += SWAP_META_PAGES; 1631 } 1632 mtx_unlock(&swhash_mtx); 1633 return (0); 1634 } 1635 1636 int 1637 swap_pager_nswapdev(void) 1638 { 1639 1640 return (nswapdev); 1641 } 1642 1643 /* 1644 * SWP_PAGER_FORCE_PAGEIN() - force a swap block to be paged in 1645 * 1646 * This routine dissociates the page at the given index within an object 1647 * from its backing store, paging it in if it does not reside in memory. 1648 * If the page is paged in, it is marked dirty and placed in the laundry 1649 * queue. The page is marked dirty because it no longer has backing 1650 * store. It is placed in the laundry queue because it has not been 1651 * accessed recently. Otherwise, it would already reside in memory. 1652 * 1653 * We also attempt to swap in all other pages in the swap block. 1654 * However, we only guarantee that the one at the specified index is 1655 * paged in. 1656 * 1657 * XXX - The code to page the whole block in doesn't work, so we 1658 * revert to the one-by-one behavior for now. Sigh. 1659 */ 1660 static inline void 1661 swp_pager_force_pagein(vm_object_t object, vm_pindex_t pindex) 1662 { 1663 vm_page_t m; 1664 1665 vm_object_pip_add(object, 1); 1666 m = vm_page_grab(object, pindex, VM_ALLOC_NORMAL); 1667 if (m->valid == VM_PAGE_BITS_ALL) { 1668 vm_object_pip_wakeup(object); 1669 vm_page_dirty(m); 1670 vm_page_lock(m); 1671 vm_page_activate(m); 1672 vm_page_unlock(m); 1673 vm_page_xunbusy(m); 1674 vm_pager_page_unswapped(m); 1675 return; 1676 } 1677 1678 if (swap_pager_getpages(object, &m, 1, NULL, NULL) != VM_PAGER_OK) 1679 panic("swap_pager_force_pagein: read from swap failed");/*XXX*/ 1680 vm_object_pip_wakeup(object); 1681 vm_page_dirty(m); 1682 vm_page_lock(m); 1683 vm_page_launder(m); 1684 vm_page_unlock(m); 1685 vm_page_xunbusy(m); 1686 vm_pager_page_unswapped(m); 1687 } 1688 1689 /* 1690 * swap_pager_swapoff: 1691 * 1692 * Page in all of the pages that have been paged out to the 1693 * given device. The corresponding blocks in the bitmap must be 1694 * marked as allocated and the device must be flagged SW_CLOSING. 1695 * There may be no processes swapped out to the device. 1696 * 1697 * This routine may block. 1698 */ 1699 static void 1700 swap_pager_swapoff(struct swdevt *sp) 1701 { 1702 struct swblock *swap; 1703 vm_object_t locked_obj, object; 1704 vm_pindex_t pindex; 1705 int i, j, retries; 1706 1707 sx_assert(&swdev_syscall_lock, SA_XLOCKED); 1708 1709 retries = 0; 1710 locked_obj = NULL; 1711 full_rescan: 1712 mtx_lock(&swhash_mtx); 1713 for (i = 0; i <= swhash_mask; i++) { /* '<=' is correct here */ 1714 restart: 1715 for (swap = swhash[i]; swap != NULL; swap = swap->swb_hnext) { 1716 object = swap->swb_object; 1717 pindex = swap->swb_index; 1718 for (j = 0; j < SWAP_META_PAGES; ++j) { 1719 if (!swp_pager_isondev(swap->swb_pages[j], sp)) 1720 continue; 1721 if (locked_obj != object) { 1722 if (locked_obj != NULL) 1723 VM_OBJECT_WUNLOCK(locked_obj); 1724 locked_obj = object; 1725 if (!VM_OBJECT_TRYWLOCK(object)) { 1726 mtx_unlock(&swhash_mtx); 1727 /* Depends on type-stability. */ 1728 VM_OBJECT_WLOCK(object); 1729 mtx_lock(&swhash_mtx); 1730 goto restart; 1731 } 1732 } 1733 MPASS(locked_obj == object); 1734 mtx_unlock(&swhash_mtx); 1735 swp_pager_force_pagein(object, pindex + j); 1736 mtx_lock(&swhash_mtx); 1737 goto restart; 1738 } 1739 } 1740 } 1741 mtx_unlock(&swhash_mtx); 1742 if (locked_obj != NULL) { 1743 VM_OBJECT_WUNLOCK(locked_obj); 1744 locked_obj = NULL; 1745 } 1746 if (sp->sw_used) { 1747 /* 1748 * Objects may be locked or paging to the device being 1749 * removed, so we will miss their pages and need to 1750 * make another pass. We have marked this device as 1751 * SW_CLOSING, so the activity should finish soon. 1752 */ 1753 retries++; 1754 if (retries > 100) { 1755 panic("swapoff: failed to locate %d swap blocks", 1756 sp->sw_used); 1757 } 1758 pause("swpoff", hz / 20); 1759 goto full_rescan; 1760 } 1761 EVENTHANDLER_INVOKE(swapoff, sp); 1762 } 1763 1764 /************************************************************************ 1765 * SWAP META DATA * 1766 ************************************************************************ 1767 * 1768 * These routines manipulate the swap metadata stored in the 1769 * OBJT_SWAP object. 1770 * 1771 * Swap metadata is implemented with a global hash and not directly 1772 * linked into the object. Instead the object simply contains 1773 * appropriate tracking counters. 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 exhausted; 1790 struct swblock *swap; 1791 struct swblock **pswap; 1792 int idx; 1793 1794 VM_OBJECT_ASSERT_WLOCKED(object); 1795 /* 1796 * Convert default object to swap object if necessary 1797 */ 1798 if (object->type != OBJT_SWAP) { 1799 object->type = OBJT_SWAP; 1800 object->un_pager.swp.swp_bcount = 0; 1801 KASSERT(object->handle == NULL, ("default pager with handle")); 1802 } 1803 1804 /* 1805 * Locate hash entry. If not found create, but if we aren't adding 1806 * anything just return. If we run out of space in the map we wait 1807 * and, since the hash table may have changed, retry. 1808 */ 1809 retry: 1810 mtx_lock(&swhash_mtx); 1811 pswap = swp_pager_hash(object, pindex); 1812 1813 if ((swap = *pswap) == NULL) { 1814 int i; 1815 1816 if (swapblk == SWAPBLK_NONE) 1817 goto done; 1818 1819 swap = *pswap = uma_zalloc(swap_zone, M_NOWAIT | 1820 (curproc == pageproc ? M_USE_RESERVE : 0)); 1821 if (swap == NULL) { 1822 mtx_unlock(&swhash_mtx); 1823 VM_OBJECT_WUNLOCK(object); 1824 if (uma_zone_exhausted(swap_zone)) { 1825 if (atomic_cmpset_int(&exhausted, 0, 1)) 1826 printf("swap zone exhausted, " 1827 "increase kern.maxswzone\n"); 1828 vm_pageout_oom(VM_OOM_SWAPZ); 1829 pause("swzonex", 10); 1830 } else 1831 VM_WAIT; 1832 VM_OBJECT_WLOCK(object); 1833 goto retry; 1834 } 1835 1836 if (atomic_cmpset_int(&exhausted, 1, 0)) 1837 printf("swap zone ok\n"); 1838 1839 swap->swb_hnext = NULL; 1840 swap->swb_object = object; 1841 swap->swb_index = pindex & ~(vm_pindex_t)SWAP_META_MASK; 1842 swap->swb_count = 0; 1843 1844 ++object->un_pager.swp.swp_bcount; 1845 1846 for (i = 0; i < SWAP_META_PAGES; ++i) 1847 swap->swb_pages[i] = SWAPBLK_NONE; 1848 } 1849 1850 /* 1851 * Delete prior contents of metadata 1852 */ 1853 idx = pindex & SWAP_META_MASK; 1854 1855 if (swap->swb_pages[idx] != SWAPBLK_NONE) { 1856 swp_pager_freeswapspace(swap->swb_pages[idx], 1); 1857 --swap->swb_count; 1858 } 1859 1860 /* 1861 * Enter block into metadata 1862 */ 1863 swap->swb_pages[idx] = swapblk; 1864 if (swapblk != SWAPBLK_NONE) 1865 ++swap->swb_count; 1866 done: 1867 mtx_unlock(&swhash_mtx); 1868 } 1869 1870 /* 1871 * SWP_PAGER_META_FREE() - free a range of blocks in the object's swap metadata 1872 * 1873 * The requested range of blocks is freed, with any associated swap 1874 * returned to the swap bitmap. 1875 * 1876 * This routine will free swap metadata structures as they are cleaned 1877 * out. This routine does *NOT* operate on swap metadata associated 1878 * with resident pages. 1879 */ 1880 static void 1881 swp_pager_meta_free(vm_object_t object, vm_pindex_t index, vm_pindex_t count) 1882 { 1883 struct swblock **pswap, *swap; 1884 vm_pindex_t c; 1885 daddr_t v; 1886 int n, sidx; 1887 1888 VM_OBJECT_ASSERT_LOCKED(object); 1889 if (object->type != OBJT_SWAP || count == 0) 1890 return; 1891 1892 mtx_lock(&swhash_mtx); 1893 for (c = 0; c < count;) { 1894 pswap = swp_pager_hash(object, index); 1895 sidx = index & SWAP_META_MASK; 1896 n = SWAP_META_PAGES - sidx; 1897 index += n; 1898 if ((swap = *pswap) == NULL) { 1899 c += n; 1900 continue; 1901 } 1902 for (; c < count && sidx < SWAP_META_PAGES; ++c, ++sidx) { 1903 if ((v = swap->swb_pages[sidx]) == SWAPBLK_NONE) 1904 continue; 1905 swp_pager_freeswapspace(v, 1); 1906 swap->swb_pages[sidx] = SWAPBLK_NONE; 1907 if (--swap->swb_count == 0) { 1908 *pswap = swap->swb_hnext; 1909 uma_zfree(swap_zone, swap); 1910 --object->un_pager.swp.swp_bcount; 1911 c += SWAP_META_PAGES - sidx; 1912 break; 1913 } 1914 } 1915 } 1916 mtx_unlock(&swhash_mtx); 1917 } 1918 1919 /* 1920 * SWP_PAGER_META_FREE_ALL() - destroy all swap metadata associated with object 1921 * 1922 * This routine locates and destroys all swap metadata associated with 1923 * an object. 1924 */ 1925 static void 1926 swp_pager_meta_free_all(vm_object_t object) 1927 { 1928 struct swblock **pswap, *swap; 1929 vm_pindex_t index; 1930 daddr_t v; 1931 int i; 1932 1933 VM_OBJECT_ASSERT_WLOCKED(object); 1934 if (object->type != OBJT_SWAP) 1935 return; 1936 1937 index = 0; 1938 while (object->un_pager.swp.swp_bcount != 0) { 1939 mtx_lock(&swhash_mtx); 1940 pswap = swp_pager_hash(object, index); 1941 if ((swap = *pswap) != NULL) { 1942 for (i = 0; i < SWAP_META_PAGES; ++i) { 1943 v = swap->swb_pages[i]; 1944 if (v != SWAPBLK_NONE) { 1945 --swap->swb_count; 1946 swp_pager_freeswapspace(v, 1); 1947 } 1948 } 1949 if (swap->swb_count != 0) 1950 panic( 1951 "swap_pager_meta_free_all: swb_count != 0"); 1952 *pswap = swap->swb_hnext; 1953 uma_zfree(swap_zone, swap); 1954 --object->un_pager.swp.swp_bcount; 1955 } 1956 mtx_unlock(&swhash_mtx); 1957 index += SWAP_META_PAGES; 1958 } 1959 } 1960 1961 /* 1962 * SWP_PAGER_METACTL() - misc control of swap and vm_page_t meta data. 1963 * 1964 * This routine is capable of looking up, popping, or freeing 1965 * swapblk assignments in the swap meta data or in the vm_page_t. 1966 * The routine typically returns the swapblk being looked-up, or popped, 1967 * or SWAPBLK_NONE if the block was freed, or SWAPBLK_NONE if the block 1968 * was invalid. This routine will automatically free any invalid 1969 * meta-data swapblks. 1970 * 1971 * It is not possible to store invalid swapblks in the swap meta data 1972 * (other then a literal 'SWAPBLK_NONE'), so we don't bother checking. 1973 * 1974 * When acting on a busy resident page and paging is in progress, we 1975 * have to wait until paging is complete but otherwise can act on the 1976 * busy page. 1977 * 1978 * SWM_FREE remove and free swap block from metadata 1979 * SWM_POP remove from meta data but do not free.. pop it out 1980 */ 1981 static daddr_t 1982 swp_pager_meta_ctl(vm_object_t object, vm_pindex_t pindex, int flags) 1983 { 1984 struct swblock **pswap; 1985 struct swblock *swap; 1986 daddr_t r1; 1987 int idx; 1988 1989 VM_OBJECT_ASSERT_LOCKED(object); 1990 /* 1991 * The meta data only exists of the object is OBJT_SWAP 1992 * and even then might not be allocated yet. 1993 */ 1994 if (object->type != OBJT_SWAP) 1995 return (SWAPBLK_NONE); 1996 1997 r1 = SWAPBLK_NONE; 1998 mtx_lock(&swhash_mtx); 1999 pswap = swp_pager_hash(object, pindex); 2000 2001 if ((swap = *pswap) != NULL) { 2002 idx = pindex & SWAP_META_MASK; 2003 r1 = swap->swb_pages[idx]; 2004 2005 if (r1 != SWAPBLK_NONE) { 2006 if (flags & SWM_FREE) { 2007 swp_pager_freeswapspace(r1, 1); 2008 r1 = SWAPBLK_NONE; 2009 } 2010 if (flags & (SWM_FREE|SWM_POP)) { 2011 swap->swb_pages[idx] = SWAPBLK_NONE; 2012 if (--swap->swb_count == 0) { 2013 *pswap = swap->swb_hnext; 2014 uma_zfree(swap_zone, swap); 2015 --object->un_pager.swp.swp_bcount; 2016 } 2017 } 2018 } 2019 } 2020 mtx_unlock(&swhash_mtx); 2021 return (r1); 2022 } 2023 2024 /* 2025 * Returns the least page index which is greater than or equal to the 2026 * parameter pindex and for which there is a swap block allocated. 2027 * Returns object's size if the object's type is not swap or if there 2028 * are no allocated swap blocks for the object after the requested 2029 * pindex. 2030 */ 2031 vm_pindex_t 2032 swap_pager_find_least(vm_object_t object, vm_pindex_t pindex) 2033 { 2034 struct swblock **pswap, *swap; 2035 vm_pindex_t i, j, lim; 2036 int idx; 2037 2038 VM_OBJECT_ASSERT_LOCKED(object); 2039 if (object->type != OBJT_SWAP || object->un_pager.swp.swp_bcount == 0) 2040 return (object->size); 2041 2042 mtx_lock(&swhash_mtx); 2043 for (j = pindex; j < object->size; j = lim) { 2044 pswap = swp_pager_hash(object, j); 2045 lim = rounddown2(j + SWAP_META_PAGES, SWAP_META_PAGES); 2046 if (lim > object->size) 2047 lim = object->size; 2048 if ((swap = *pswap) != NULL) { 2049 for (idx = j & SWAP_META_MASK, i = j; i < lim; 2050 i++, idx++) { 2051 if (swap->swb_pages[idx] != SWAPBLK_NONE) 2052 goto found; 2053 } 2054 } 2055 } 2056 i = object->size; 2057 found: 2058 mtx_unlock(&swhash_mtx); 2059 return (i); 2060 } 2061 2062 /* 2063 * System call swapon(name) enables swapping on device name, 2064 * which must be in the swdevsw. Return EBUSY 2065 * if already swapping on this device. 2066 */ 2067 #ifndef _SYS_SYSPROTO_H_ 2068 struct swapon_args { 2069 char *name; 2070 }; 2071 #endif 2072 2073 /* 2074 * MPSAFE 2075 */ 2076 /* ARGSUSED */ 2077 int 2078 sys_swapon(struct thread *td, struct swapon_args *uap) 2079 { 2080 struct vattr attr; 2081 struct vnode *vp; 2082 struct nameidata nd; 2083 int error; 2084 2085 error = priv_check(td, PRIV_SWAPON); 2086 if (error) 2087 return (error); 2088 2089 sx_xlock(&swdev_syscall_lock); 2090 2091 /* 2092 * Swap metadata may not fit in the KVM if we have physical 2093 * memory of >1GB. 2094 */ 2095 if (swap_zone == NULL) { 2096 error = ENOMEM; 2097 goto done; 2098 } 2099 2100 NDINIT(&nd, LOOKUP, ISOPEN | FOLLOW | AUDITVNODE1, UIO_USERSPACE, 2101 uap->name, td); 2102 error = namei(&nd); 2103 if (error) 2104 goto done; 2105 2106 NDFREE(&nd, NDF_ONLY_PNBUF); 2107 vp = nd.ni_vp; 2108 2109 if (vn_isdisk(vp, &error)) { 2110 error = swapongeom(vp); 2111 } else if (vp->v_type == VREG && 2112 (vp->v_mount->mnt_vfc->vfc_flags & VFCF_NETWORK) != 0 && 2113 (error = VOP_GETATTR(vp, &attr, td->td_ucred)) == 0) { 2114 /* 2115 * Allow direct swapping to NFS regular files in the same 2116 * way that nfs_mountroot() sets up diskless swapping. 2117 */ 2118 error = swaponvp(td, vp, attr.va_size / DEV_BSIZE); 2119 } 2120 2121 if (error) 2122 vrele(vp); 2123 done: 2124 sx_xunlock(&swdev_syscall_lock); 2125 return (error); 2126 } 2127 2128 /* 2129 * Check that the total amount of swap currently configured does not 2130 * exceed half the theoretical maximum. If it does, print a warning 2131 * message and return -1; otherwise, return 0. 2132 */ 2133 static int 2134 swapon_check_swzone(unsigned long npages) 2135 { 2136 unsigned long maxpages; 2137 2138 /* absolute maximum we can handle assuming 100% efficiency */ 2139 maxpages = uma_zone_get_max(swap_zone) * SWAP_META_PAGES; 2140 2141 /* recommend using no more than half that amount */ 2142 if (npages > maxpages / 2) { 2143 printf("warning: total configured swap (%lu pages) " 2144 "exceeds maximum recommended amount (%lu pages).\n", 2145 npages, maxpages / 2); 2146 printf("warning: increase kern.maxswzone " 2147 "or reduce amount of swap.\n"); 2148 return (-1); 2149 } 2150 return (0); 2151 } 2152 2153 static void 2154 swaponsomething(struct vnode *vp, void *id, u_long nblks, 2155 sw_strategy_t *strategy, sw_close_t *close, dev_t dev, int flags) 2156 { 2157 struct swdevt *sp, *tsp; 2158 swblk_t dvbase; 2159 u_long mblocks; 2160 2161 /* 2162 * nblks is in DEV_BSIZE'd chunks, convert to PAGE_SIZE'd chunks. 2163 * First chop nblks off to page-align it, then convert. 2164 * 2165 * sw->sw_nblks is in page-sized chunks now too. 2166 */ 2167 nblks &= ~(ctodb(1) - 1); 2168 nblks = dbtoc(nblks); 2169 2170 /* 2171 * If we go beyond this, we get overflows in the radix 2172 * tree bitmap code. 2173 */ 2174 mblocks = 0x40000000 / BLIST_META_RADIX; 2175 if (nblks > mblocks) { 2176 printf( 2177 "WARNING: reducing swap size to maximum of %luMB per unit\n", 2178 mblocks / 1024 / 1024 * PAGE_SIZE); 2179 nblks = mblocks; 2180 } 2181 2182 sp = malloc(sizeof *sp, M_VMPGDATA, M_WAITOK | M_ZERO); 2183 sp->sw_vp = vp; 2184 sp->sw_id = id; 2185 sp->sw_dev = dev; 2186 sp->sw_flags = 0; 2187 sp->sw_nblks = nblks; 2188 sp->sw_used = 0; 2189 sp->sw_strategy = strategy; 2190 sp->sw_close = close; 2191 sp->sw_flags = flags; 2192 2193 sp->sw_blist = blist_create(nblks, M_WAITOK); 2194 /* 2195 * Do not free the first two block in order to avoid overwriting 2196 * any bsd label at the front of the partition 2197 */ 2198 blist_free(sp->sw_blist, 2, nblks - 2); 2199 2200 dvbase = 0; 2201 mtx_lock(&sw_dev_mtx); 2202 TAILQ_FOREACH(tsp, &swtailq, sw_list) { 2203 if (tsp->sw_end >= dvbase) { 2204 /* 2205 * We put one uncovered page between the devices 2206 * in order to definitively prevent any cross-device 2207 * I/O requests 2208 */ 2209 dvbase = tsp->sw_end + 1; 2210 } 2211 } 2212 sp->sw_first = dvbase; 2213 sp->sw_end = dvbase + nblks; 2214 TAILQ_INSERT_TAIL(&swtailq, sp, sw_list); 2215 nswapdev++; 2216 swap_pager_avail += nblks; 2217 swap_total += (vm_ooffset_t)nblks * PAGE_SIZE; 2218 swapon_check_swzone(swap_total / PAGE_SIZE); 2219 swp_sizecheck(); 2220 mtx_unlock(&sw_dev_mtx); 2221 EVENTHANDLER_INVOKE(swapon, sp); 2222 } 2223 2224 /* 2225 * SYSCALL: swapoff(devname) 2226 * 2227 * Disable swapping on the given device. 2228 * 2229 * XXX: Badly designed system call: it should use a device index 2230 * rather than filename as specification. We keep sw_vp around 2231 * only to make this work. 2232 */ 2233 #ifndef _SYS_SYSPROTO_H_ 2234 struct swapoff_args { 2235 char *name; 2236 }; 2237 #endif 2238 2239 /* 2240 * MPSAFE 2241 */ 2242 /* ARGSUSED */ 2243 int 2244 sys_swapoff(struct thread *td, struct swapoff_args *uap) 2245 { 2246 struct vnode *vp; 2247 struct nameidata nd; 2248 struct swdevt *sp; 2249 int error; 2250 2251 error = priv_check(td, PRIV_SWAPOFF); 2252 if (error) 2253 return (error); 2254 2255 sx_xlock(&swdev_syscall_lock); 2256 2257 NDINIT(&nd, LOOKUP, FOLLOW | AUDITVNODE1, UIO_USERSPACE, uap->name, 2258 td); 2259 error = namei(&nd); 2260 if (error) 2261 goto done; 2262 NDFREE(&nd, NDF_ONLY_PNBUF); 2263 vp = nd.ni_vp; 2264 2265 mtx_lock(&sw_dev_mtx); 2266 TAILQ_FOREACH(sp, &swtailq, sw_list) { 2267 if (sp->sw_vp == vp) 2268 break; 2269 } 2270 mtx_unlock(&sw_dev_mtx); 2271 if (sp == NULL) { 2272 error = EINVAL; 2273 goto done; 2274 } 2275 error = swapoff_one(sp, td->td_ucred); 2276 done: 2277 sx_xunlock(&swdev_syscall_lock); 2278 return (error); 2279 } 2280 2281 static int 2282 swapoff_one(struct swdevt *sp, struct ucred *cred) 2283 { 2284 u_long nblks, dvbase; 2285 #ifdef MAC 2286 int error; 2287 #endif 2288 2289 sx_assert(&swdev_syscall_lock, SA_XLOCKED); 2290 #ifdef MAC 2291 (void) vn_lock(sp->sw_vp, LK_EXCLUSIVE | LK_RETRY); 2292 error = mac_system_check_swapoff(cred, sp->sw_vp); 2293 (void) VOP_UNLOCK(sp->sw_vp, 0); 2294 if (error != 0) 2295 return (error); 2296 #endif 2297 nblks = sp->sw_nblks; 2298 2299 /* 2300 * We can turn off this swap device safely only if the 2301 * available virtual memory in the system will fit the amount 2302 * of data we will have to page back in, plus an epsilon so 2303 * the system doesn't become critically low on swap space. 2304 */ 2305 if (vm_cnt.v_free_count + swap_pager_avail < nblks + nswap_lowat) 2306 return (ENOMEM); 2307 2308 /* 2309 * Prevent further allocations on this device. 2310 */ 2311 mtx_lock(&sw_dev_mtx); 2312 sp->sw_flags |= SW_CLOSING; 2313 for (dvbase = 0; dvbase < sp->sw_end; dvbase += dmmax) { 2314 swap_pager_avail -= blist_fill(sp->sw_blist, 2315 dvbase, dmmax); 2316 } 2317 swap_total -= (vm_ooffset_t)nblks * PAGE_SIZE; 2318 mtx_unlock(&sw_dev_mtx); 2319 2320 /* 2321 * Page in the contents of the device and close it. 2322 */ 2323 swap_pager_swapoff(sp); 2324 2325 sp->sw_close(curthread, sp); 2326 mtx_lock(&sw_dev_mtx); 2327 sp->sw_id = NULL; 2328 TAILQ_REMOVE(&swtailq, sp, sw_list); 2329 nswapdev--; 2330 if (nswapdev == 0) { 2331 swap_pager_full = 2; 2332 swap_pager_almost_full = 1; 2333 } 2334 if (swdevhd == sp) 2335 swdevhd = NULL; 2336 mtx_unlock(&sw_dev_mtx); 2337 blist_destroy(sp->sw_blist); 2338 free(sp, M_VMPGDATA); 2339 return (0); 2340 } 2341 2342 void 2343 swapoff_all(void) 2344 { 2345 struct swdevt *sp, *spt; 2346 const char *devname; 2347 int error; 2348 2349 sx_xlock(&swdev_syscall_lock); 2350 2351 mtx_lock(&sw_dev_mtx); 2352 TAILQ_FOREACH_SAFE(sp, &swtailq, sw_list, spt) { 2353 mtx_unlock(&sw_dev_mtx); 2354 if (vn_isdisk(sp->sw_vp, NULL)) 2355 devname = devtoname(sp->sw_vp->v_rdev); 2356 else 2357 devname = "[file]"; 2358 error = swapoff_one(sp, thread0.td_ucred); 2359 if (error != 0) { 2360 printf("Cannot remove swap device %s (error=%d), " 2361 "skipping.\n", devname, error); 2362 } else if (bootverbose) { 2363 printf("Swap device %s removed.\n", devname); 2364 } 2365 mtx_lock(&sw_dev_mtx); 2366 } 2367 mtx_unlock(&sw_dev_mtx); 2368 2369 sx_xunlock(&swdev_syscall_lock); 2370 } 2371 2372 void 2373 swap_pager_status(int *total, int *used) 2374 { 2375 struct swdevt *sp; 2376 2377 *total = 0; 2378 *used = 0; 2379 mtx_lock(&sw_dev_mtx); 2380 TAILQ_FOREACH(sp, &swtailq, sw_list) { 2381 *total += sp->sw_nblks; 2382 *used += sp->sw_used; 2383 } 2384 mtx_unlock(&sw_dev_mtx); 2385 } 2386 2387 int 2388 swap_dev_info(int name, struct xswdev *xs, char *devname, size_t len) 2389 { 2390 struct swdevt *sp; 2391 const char *tmp_devname; 2392 int error, n; 2393 2394 n = 0; 2395 error = ENOENT; 2396 mtx_lock(&sw_dev_mtx); 2397 TAILQ_FOREACH(sp, &swtailq, sw_list) { 2398 if (n != name) { 2399 n++; 2400 continue; 2401 } 2402 xs->xsw_version = XSWDEV_VERSION; 2403 xs->xsw_dev = sp->sw_dev; 2404 xs->xsw_flags = sp->sw_flags; 2405 xs->xsw_nblks = sp->sw_nblks; 2406 xs->xsw_used = sp->sw_used; 2407 if (devname != NULL) { 2408 if (vn_isdisk(sp->sw_vp, NULL)) 2409 tmp_devname = devtoname(sp->sw_vp->v_rdev); 2410 else 2411 tmp_devname = "[file]"; 2412 strncpy(devname, tmp_devname, len); 2413 } 2414 error = 0; 2415 break; 2416 } 2417 mtx_unlock(&sw_dev_mtx); 2418 return (error); 2419 } 2420 2421 #if defined(COMPAT_FREEBSD11) 2422 #define XSWDEV_VERSION_11 1 2423 struct xswdev11 { 2424 u_int xsw_version; 2425 uint32_t xsw_dev; 2426 int xsw_flags; 2427 int xsw_nblks; 2428 int xsw_used; 2429 }; 2430 #endif 2431 2432 static int 2433 sysctl_vm_swap_info(SYSCTL_HANDLER_ARGS) 2434 { 2435 struct xswdev xs; 2436 #if defined(COMPAT_FREEBSD11) 2437 struct xswdev11 xs11; 2438 #endif 2439 int error; 2440 2441 if (arg2 != 1) /* name length */ 2442 return (EINVAL); 2443 error = swap_dev_info(*(int *)arg1, &xs, NULL, 0); 2444 if (error != 0) 2445 return (error); 2446 #if defined(COMPAT_FREEBSD11) 2447 if (req->oldlen == sizeof(xs11)) { 2448 xs11.xsw_version = XSWDEV_VERSION_11; 2449 xs11.xsw_dev = xs.xsw_dev; /* truncation */ 2450 xs11.xsw_flags = xs.xsw_flags; 2451 xs11.xsw_nblks = xs.xsw_nblks; 2452 xs11.xsw_used = xs.xsw_used; 2453 error = SYSCTL_OUT(req, &xs11, sizeof(xs11)); 2454 } else 2455 #endif 2456 error = SYSCTL_OUT(req, &xs, sizeof(xs)); 2457 return (error); 2458 } 2459 2460 SYSCTL_INT(_vm, OID_AUTO, nswapdev, CTLFLAG_RD, &nswapdev, 0, 2461 "Number of swap devices"); 2462 SYSCTL_NODE(_vm, OID_AUTO, swap_info, CTLFLAG_RD | CTLFLAG_MPSAFE, 2463 sysctl_vm_swap_info, 2464 "Swap statistics by device"); 2465 2466 /* 2467 * vmspace_swap_count() - count the approximate swap usage in pages for a 2468 * vmspace. 2469 * 2470 * The map must be locked. 2471 * 2472 * Swap usage is determined by taking the proportional swap used by 2473 * VM objects backing the VM map. To make up for fractional losses, 2474 * if the VM object has any swap use at all the associated map entries 2475 * count for at least 1 swap page. 2476 */ 2477 long 2478 vmspace_swap_count(struct vmspace *vmspace) 2479 { 2480 vm_map_t map; 2481 vm_map_entry_t cur; 2482 vm_object_t object; 2483 long count, n; 2484 2485 map = &vmspace->vm_map; 2486 count = 0; 2487 2488 for (cur = map->header.next; cur != &map->header; cur = cur->next) { 2489 if ((cur->eflags & MAP_ENTRY_IS_SUB_MAP) == 0 && 2490 (object = cur->object.vm_object) != NULL) { 2491 VM_OBJECT_WLOCK(object); 2492 if (object->type == OBJT_SWAP && 2493 object->un_pager.swp.swp_bcount != 0) { 2494 n = (cur->end - cur->start) / PAGE_SIZE; 2495 count += object->un_pager.swp.swp_bcount * 2496 SWAP_META_PAGES * n / object->size + 1; 2497 } 2498 VM_OBJECT_WUNLOCK(object); 2499 } 2500 } 2501 return (count); 2502 } 2503 2504 /* 2505 * GEOM backend 2506 * 2507 * Swapping onto disk devices. 2508 * 2509 */ 2510 2511 static g_orphan_t swapgeom_orphan; 2512 2513 static struct g_class g_swap_class = { 2514 .name = "SWAP", 2515 .version = G_VERSION, 2516 .orphan = swapgeom_orphan, 2517 }; 2518 2519 DECLARE_GEOM_CLASS(g_swap_class, g_class); 2520 2521 2522 static void 2523 swapgeom_close_ev(void *arg, int flags) 2524 { 2525 struct g_consumer *cp; 2526 2527 cp = arg; 2528 g_access(cp, -1, -1, 0); 2529 g_detach(cp); 2530 g_destroy_consumer(cp); 2531 } 2532 2533 /* 2534 * Add a reference to the g_consumer for an inflight transaction. 2535 */ 2536 static void 2537 swapgeom_acquire(struct g_consumer *cp) 2538 { 2539 2540 mtx_assert(&sw_dev_mtx, MA_OWNED); 2541 cp->index++; 2542 } 2543 2544 /* 2545 * Remove a reference from the g_consumer. Post a close event if all 2546 * references go away, since the function might be called from the 2547 * biodone context. 2548 */ 2549 static void 2550 swapgeom_release(struct g_consumer *cp, struct swdevt *sp) 2551 { 2552 2553 mtx_assert(&sw_dev_mtx, MA_OWNED); 2554 cp->index--; 2555 if (cp->index == 0) { 2556 if (g_post_event(swapgeom_close_ev, cp, M_NOWAIT, NULL) == 0) 2557 sp->sw_id = NULL; 2558 } 2559 } 2560 2561 static void 2562 swapgeom_done(struct bio *bp2) 2563 { 2564 struct swdevt *sp; 2565 struct buf *bp; 2566 struct g_consumer *cp; 2567 2568 bp = bp2->bio_caller2; 2569 cp = bp2->bio_from; 2570 bp->b_ioflags = bp2->bio_flags; 2571 if (bp2->bio_error) 2572 bp->b_ioflags |= BIO_ERROR; 2573 bp->b_resid = bp->b_bcount - bp2->bio_completed; 2574 bp->b_error = bp2->bio_error; 2575 bufdone(bp); 2576 sp = bp2->bio_caller1; 2577 mtx_lock(&sw_dev_mtx); 2578 swapgeom_release(cp, sp); 2579 mtx_unlock(&sw_dev_mtx); 2580 g_destroy_bio(bp2); 2581 } 2582 2583 static void 2584 swapgeom_strategy(struct buf *bp, struct swdevt *sp) 2585 { 2586 struct bio *bio; 2587 struct g_consumer *cp; 2588 2589 mtx_lock(&sw_dev_mtx); 2590 cp = sp->sw_id; 2591 if (cp == NULL) { 2592 mtx_unlock(&sw_dev_mtx); 2593 bp->b_error = ENXIO; 2594 bp->b_ioflags |= BIO_ERROR; 2595 bufdone(bp); 2596 return; 2597 } 2598 swapgeom_acquire(cp); 2599 mtx_unlock(&sw_dev_mtx); 2600 if (bp->b_iocmd == BIO_WRITE) 2601 bio = g_new_bio(); 2602 else 2603 bio = g_alloc_bio(); 2604 if (bio == NULL) { 2605 mtx_lock(&sw_dev_mtx); 2606 swapgeom_release(cp, sp); 2607 mtx_unlock(&sw_dev_mtx); 2608 bp->b_error = ENOMEM; 2609 bp->b_ioflags |= BIO_ERROR; 2610 bufdone(bp); 2611 return; 2612 } 2613 2614 bio->bio_caller1 = sp; 2615 bio->bio_caller2 = bp; 2616 bio->bio_cmd = bp->b_iocmd; 2617 bio->bio_offset = (bp->b_blkno - sp->sw_first) * PAGE_SIZE; 2618 bio->bio_length = bp->b_bcount; 2619 bio->bio_done = swapgeom_done; 2620 if (!buf_mapped(bp)) { 2621 bio->bio_ma = bp->b_pages; 2622 bio->bio_data = unmapped_buf; 2623 bio->bio_ma_offset = (vm_offset_t)bp->b_offset & PAGE_MASK; 2624 bio->bio_ma_n = bp->b_npages; 2625 bio->bio_flags |= BIO_UNMAPPED; 2626 } else { 2627 bio->bio_data = bp->b_data; 2628 bio->bio_ma = NULL; 2629 } 2630 g_io_request(bio, cp); 2631 return; 2632 } 2633 2634 static void 2635 swapgeom_orphan(struct g_consumer *cp) 2636 { 2637 struct swdevt *sp; 2638 int destroy; 2639 2640 mtx_lock(&sw_dev_mtx); 2641 TAILQ_FOREACH(sp, &swtailq, sw_list) { 2642 if (sp->sw_id == cp) { 2643 sp->sw_flags |= SW_CLOSING; 2644 break; 2645 } 2646 } 2647 /* 2648 * Drop reference we were created with. Do directly since we're in a 2649 * special context where we don't have to queue the call to 2650 * swapgeom_close_ev(). 2651 */ 2652 cp->index--; 2653 destroy = ((sp != NULL) && (cp->index == 0)); 2654 if (destroy) 2655 sp->sw_id = NULL; 2656 mtx_unlock(&sw_dev_mtx); 2657 if (destroy) 2658 swapgeom_close_ev(cp, 0); 2659 } 2660 2661 static void 2662 swapgeom_close(struct thread *td, struct swdevt *sw) 2663 { 2664 struct g_consumer *cp; 2665 2666 mtx_lock(&sw_dev_mtx); 2667 cp = sw->sw_id; 2668 sw->sw_id = NULL; 2669 mtx_unlock(&sw_dev_mtx); 2670 2671 /* 2672 * swapgeom_close() may be called from the biodone context, 2673 * where we cannot perform topology changes. Delegate the 2674 * work to the events thread. 2675 */ 2676 if (cp != NULL) 2677 g_waitfor_event(swapgeom_close_ev, cp, M_WAITOK, NULL); 2678 } 2679 2680 static int 2681 swapongeom_locked(struct cdev *dev, struct vnode *vp) 2682 { 2683 struct g_provider *pp; 2684 struct g_consumer *cp; 2685 static struct g_geom *gp; 2686 struct swdevt *sp; 2687 u_long nblks; 2688 int error; 2689 2690 pp = g_dev_getprovider(dev); 2691 if (pp == NULL) 2692 return (ENODEV); 2693 mtx_lock(&sw_dev_mtx); 2694 TAILQ_FOREACH(sp, &swtailq, sw_list) { 2695 cp = sp->sw_id; 2696 if (cp != NULL && cp->provider == pp) { 2697 mtx_unlock(&sw_dev_mtx); 2698 return (EBUSY); 2699 } 2700 } 2701 mtx_unlock(&sw_dev_mtx); 2702 if (gp == NULL) 2703 gp = g_new_geomf(&g_swap_class, "swap"); 2704 cp = g_new_consumer(gp); 2705 cp->index = 1; /* Number of active I/Os, plus one for being active. */ 2706 cp->flags |= G_CF_DIRECT_SEND | G_CF_DIRECT_RECEIVE; 2707 g_attach(cp, pp); 2708 /* 2709 * XXX: Every time you think you can improve the margin for 2710 * footshooting, somebody depends on the ability to do so: 2711 * savecore(8) wants to write to our swapdev so we cannot 2712 * set an exclusive count :-( 2713 */ 2714 error = g_access(cp, 1, 1, 0); 2715 if (error != 0) { 2716 g_detach(cp); 2717 g_destroy_consumer(cp); 2718 return (error); 2719 } 2720 nblks = pp->mediasize / DEV_BSIZE; 2721 swaponsomething(vp, cp, nblks, swapgeom_strategy, 2722 swapgeom_close, dev2udev(dev), 2723 (pp->flags & G_PF_ACCEPT_UNMAPPED) != 0 ? SW_UNMAPPED : 0); 2724 return (0); 2725 } 2726 2727 static int 2728 swapongeom(struct vnode *vp) 2729 { 2730 int error; 2731 2732 vn_lock(vp, LK_EXCLUSIVE | LK_RETRY); 2733 if (vp->v_type != VCHR || (vp->v_iflag & VI_DOOMED) != 0) { 2734 error = ENOENT; 2735 } else { 2736 g_topology_lock(); 2737 error = swapongeom_locked(vp->v_rdev, vp); 2738 g_topology_unlock(); 2739 } 2740 VOP_UNLOCK(vp, 0); 2741 return (error); 2742 } 2743 2744 /* 2745 * VNODE backend 2746 * 2747 * This is used mainly for network filesystem (read: probably only tested 2748 * with NFS) swapfiles. 2749 * 2750 */ 2751 2752 static void 2753 swapdev_strategy(struct buf *bp, struct swdevt *sp) 2754 { 2755 struct vnode *vp2; 2756 2757 bp->b_blkno = ctodb(bp->b_blkno - sp->sw_first); 2758 2759 vp2 = sp->sw_id; 2760 vhold(vp2); 2761 if (bp->b_iocmd == BIO_WRITE) { 2762 if (bp->b_bufobj) 2763 bufobj_wdrop(bp->b_bufobj); 2764 bufobj_wref(&vp2->v_bufobj); 2765 } 2766 if (bp->b_bufobj != &vp2->v_bufobj) 2767 bp->b_bufobj = &vp2->v_bufobj; 2768 bp->b_vp = vp2; 2769 bp->b_iooffset = dbtob(bp->b_blkno); 2770 bstrategy(bp); 2771 return; 2772 } 2773 2774 static void 2775 swapdev_close(struct thread *td, struct swdevt *sp) 2776 { 2777 2778 VOP_CLOSE(sp->sw_vp, FREAD | FWRITE, td->td_ucred, td); 2779 vrele(sp->sw_vp); 2780 } 2781 2782 2783 static int 2784 swaponvp(struct thread *td, struct vnode *vp, u_long nblks) 2785 { 2786 struct swdevt *sp; 2787 int error; 2788 2789 if (nblks == 0) 2790 return (ENXIO); 2791 mtx_lock(&sw_dev_mtx); 2792 TAILQ_FOREACH(sp, &swtailq, sw_list) { 2793 if (sp->sw_id == vp) { 2794 mtx_unlock(&sw_dev_mtx); 2795 return (EBUSY); 2796 } 2797 } 2798 mtx_unlock(&sw_dev_mtx); 2799 2800 (void) vn_lock(vp, LK_EXCLUSIVE | LK_RETRY); 2801 #ifdef MAC 2802 error = mac_system_check_swapon(td->td_ucred, vp); 2803 if (error == 0) 2804 #endif 2805 error = VOP_OPEN(vp, FREAD | FWRITE, td->td_ucred, td, NULL); 2806 (void) VOP_UNLOCK(vp, 0); 2807 if (error) 2808 return (error); 2809 2810 swaponsomething(vp, vp, nblks, swapdev_strategy, swapdev_close, 2811 NODEV, 0); 2812 return (0); 2813 } 2814 2815 static int 2816 sysctl_swap_async_max(SYSCTL_HANDLER_ARGS) 2817 { 2818 int error, new, n; 2819 2820 new = nsw_wcount_async_max; 2821 error = sysctl_handle_int(oidp, &new, 0, req); 2822 if (error != 0 || req->newptr == NULL) 2823 return (error); 2824 2825 if (new > nswbuf / 2 || new < 1) 2826 return (EINVAL); 2827 2828 mtx_lock(&pbuf_mtx); 2829 while (nsw_wcount_async_max != new) { 2830 /* 2831 * Adjust difference. If the current async count is too low, 2832 * we will need to sqeeze our update slowly in. Sleep with a 2833 * higher priority than getpbuf() to finish faster. 2834 */ 2835 n = new - nsw_wcount_async_max; 2836 if (nsw_wcount_async + n >= 0) { 2837 nsw_wcount_async += n; 2838 nsw_wcount_async_max += n; 2839 wakeup(&nsw_wcount_async); 2840 } else { 2841 nsw_wcount_async_max -= nsw_wcount_async; 2842 nsw_wcount_async = 0; 2843 msleep(&nsw_wcount_async, &pbuf_mtx, PSWP, 2844 "swpsysctl", 0); 2845 } 2846 } 2847 mtx_unlock(&pbuf_mtx); 2848 2849 return (0); 2850 } 2851