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_swap.h" 73 #include "opt_vm.h" 74 75 #include <sys/param.h> 76 #include <sys/systm.h> 77 #include <sys/conf.h> 78 #include <sys/kernel.h> 79 #include <sys/priv.h> 80 #include <sys/proc.h> 81 #include <sys/bio.h> 82 #include <sys/buf.h> 83 #include <sys/disk.h> 84 #include <sys/fcntl.h> 85 #include <sys/mount.h> 86 #include <sys/namei.h> 87 #include <sys/vnode.h> 88 #include <sys/malloc.h> 89 #include <sys/racct.h> 90 #include <sys/resource.h> 91 #include <sys/resourcevar.h> 92 #include <sys/rwlock.h> 93 #include <sys/sysctl.h> 94 #include <sys/sysproto.h> 95 #include <sys/blist.h> 96 #include <sys/lock.h> 97 #include <sys/sx.h> 98 #include <sys/vmmeter.h> 99 100 #include <security/mac/mac_framework.h> 101 102 #include <vm/vm.h> 103 #include <vm/pmap.h> 104 #include <vm/vm_map.h> 105 #include <vm/vm_kern.h> 106 #include <vm/vm_object.h> 107 #include <vm/vm_page.h> 108 #include <vm/vm_pager.h> 109 #include <vm/vm_pageout.h> 110 #include <vm/vm_param.h> 111 #include <vm/swap_pager.h> 112 #include <vm/vm_extern.h> 113 #include <vm/uma.h> 114 115 #include <geom/geom.h> 116 117 /* 118 * SWB_NPAGES must be a power of 2. It may be set to 1, 2, 4, 8, 16 119 * or 32 pages per allocation. 120 * The 32-page limit is due to the radix code (kern/subr_blist.c). 121 */ 122 #ifndef MAX_PAGEOUT_CLUSTER 123 #define MAX_PAGEOUT_CLUSTER 16 124 #endif 125 126 #if !defined(SWB_NPAGES) 127 #define SWB_NPAGES MAX_PAGEOUT_CLUSTER 128 #endif 129 130 /* 131 * The swblock structure maps an object and a small, fixed-size range 132 * of page indices to disk addresses within a swap area. 133 * The collection of these mappings is implemented as a hash table. 134 * Unused disk addresses within a swap area are allocated and managed 135 * using a blist. 136 */ 137 #define SWCORRECT(n) (sizeof(void *) * (n) / sizeof(daddr_t)) 138 #define SWAP_META_PAGES (SWB_NPAGES * 2) 139 #define SWAP_META_MASK (SWAP_META_PAGES - 1) 140 141 struct swblock { 142 struct swblock *swb_hnext; 143 vm_object_t swb_object; 144 vm_pindex_t swb_index; 145 int swb_count; 146 daddr_t swb_pages[SWAP_META_PAGES]; 147 }; 148 149 static MALLOC_DEFINE(M_VMPGDATA, "vm_pgdata", "swap pager private data"); 150 static struct mtx sw_dev_mtx; 151 static TAILQ_HEAD(, swdevt) swtailq = TAILQ_HEAD_INITIALIZER(swtailq); 152 static struct swdevt *swdevhd; /* Allocate from here next */ 153 static int nswapdev; /* Number of swap devices */ 154 int swap_pager_avail; 155 static int swdev_syscall_active = 0; /* serialize swap(on|off) */ 156 157 static vm_ooffset_t swap_total; 158 SYSCTL_QUAD(_vm, OID_AUTO, swap_total, CTLFLAG_RD, &swap_total, 0, 159 "Total amount of available swap storage."); 160 static vm_ooffset_t swap_reserved; 161 SYSCTL_QUAD(_vm, OID_AUTO, swap_reserved, CTLFLAG_RD, &swap_reserved, 0, 162 "Amount of swap storage needed to back all allocated anonymous memory."); 163 static int overcommit = 0; 164 SYSCTL_INT(_vm, OID_AUTO, overcommit, CTLFLAG_RW, &overcommit, 0, 165 "Configure virtual memory overcommit behavior. See tuning(7) " 166 "for details."); 167 static unsigned long swzone; 168 SYSCTL_ULONG(_vm, OID_AUTO, swzone, CTLFLAG_RD, &swzone, 0, 169 "Actual size of swap metadata zone"); 170 static unsigned long swap_maxpages; 171 SYSCTL_ULONG(_vm, OID_AUTO, swap_maxpages, CTLFLAG_RD, &swap_maxpages, 0, 172 "Maximum amount of swap supported"); 173 174 /* bits from overcommit */ 175 #define SWAP_RESERVE_FORCE_ON (1 << 0) 176 #define SWAP_RESERVE_RLIMIT_ON (1 << 1) 177 #define SWAP_RESERVE_ALLOW_NONWIRED (1 << 2) 178 179 int 180 swap_reserve(vm_ooffset_t incr) 181 { 182 183 return (swap_reserve_by_cred(incr, curthread->td_ucred)); 184 } 185 186 int 187 swap_reserve_by_cred(vm_ooffset_t incr, struct ucred *cred) 188 { 189 vm_ooffset_t r, s; 190 int res, error; 191 static int curfail; 192 static struct timeval lastfail; 193 struct uidinfo *uip; 194 195 uip = cred->cr_ruidinfo; 196 197 if (incr & PAGE_MASK) 198 panic("swap_reserve: & PAGE_MASK"); 199 200 #ifdef RACCT 201 if (racct_enable) { 202 PROC_LOCK(curproc); 203 error = racct_add(curproc, RACCT_SWAP, incr); 204 PROC_UNLOCK(curproc); 205 if (error != 0) 206 return (0); 207 } 208 #endif 209 210 res = 0; 211 mtx_lock(&sw_dev_mtx); 212 r = swap_reserved + incr; 213 if (overcommit & SWAP_RESERVE_ALLOW_NONWIRED) { 214 s = vm_cnt.v_page_count - vm_cnt.v_free_reserved - vm_cnt.v_wire_count; 215 s *= PAGE_SIZE; 216 } else 217 s = 0; 218 s += swap_total; 219 if ((overcommit & SWAP_RESERVE_FORCE_ON) == 0 || r <= s || 220 (error = priv_check(curthread, PRIV_VM_SWAP_NOQUOTA)) == 0) { 221 res = 1; 222 swap_reserved = r; 223 } 224 mtx_unlock(&sw_dev_mtx); 225 226 if (res) { 227 UIDINFO_VMSIZE_LOCK(uip); 228 if ((overcommit & SWAP_RESERVE_RLIMIT_ON) != 0 && 229 uip->ui_vmsize + incr > lim_cur(curthread, RLIMIT_SWAP) && 230 priv_check(curthread, PRIV_VM_SWAP_NORLIMIT)) 231 res = 0; 232 else 233 uip->ui_vmsize += incr; 234 UIDINFO_VMSIZE_UNLOCK(uip); 235 if (!res) { 236 mtx_lock(&sw_dev_mtx); 237 swap_reserved -= incr; 238 mtx_unlock(&sw_dev_mtx); 239 } 240 } 241 if (!res && ppsratecheck(&lastfail, &curfail, 1)) { 242 printf("uid %d, pid %d: swap reservation for %jd bytes failed\n", 243 uip->ui_uid, curproc->p_pid, incr); 244 } 245 246 #ifdef RACCT 247 if (!res) { 248 PROC_LOCK(curproc); 249 racct_sub(curproc, RACCT_SWAP, incr); 250 PROC_UNLOCK(curproc); 251 } 252 #endif 253 254 return (res); 255 } 256 257 void 258 swap_reserve_force(vm_ooffset_t incr) 259 { 260 struct uidinfo *uip; 261 262 mtx_lock(&sw_dev_mtx); 263 swap_reserved += incr; 264 mtx_unlock(&sw_dev_mtx); 265 266 #ifdef RACCT 267 PROC_LOCK(curproc); 268 racct_add_force(curproc, RACCT_SWAP, incr); 269 PROC_UNLOCK(curproc); 270 #endif 271 272 uip = curthread->td_ucred->cr_ruidinfo; 273 PROC_LOCK(curproc); 274 UIDINFO_VMSIZE_LOCK(uip); 275 uip->ui_vmsize += incr; 276 UIDINFO_VMSIZE_UNLOCK(uip); 277 PROC_UNLOCK(curproc); 278 } 279 280 void 281 swap_release(vm_ooffset_t decr) 282 { 283 struct ucred *cred; 284 285 PROC_LOCK(curproc); 286 cred = curthread->td_ucred; 287 swap_release_by_cred(decr, cred); 288 PROC_UNLOCK(curproc); 289 } 290 291 void 292 swap_release_by_cred(vm_ooffset_t decr, struct ucred *cred) 293 { 294 struct uidinfo *uip; 295 296 uip = cred->cr_ruidinfo; 297 298 if (decr & PAGE_MASK) 299 panic("swap_release: & PAGE_MASK"); 300 301 mtx_lock(&sw_dev_mtx); 302 if (swap_reserved < decr) 303 panic("swap_reserved < decr"); 304 swap_reserved -= decr; 305 mtx_unlock(&sw_dev_mtx); 306 307 UIDINFO_VMSIZE_LOCK(uip); 308 if (uip->ui_vmsize < decr) 309 printf("negative vmsize for uid = %d\n", uip->ui_uid); 310 uip->ui_vmsize -= decr; 311 UIDINFO_VMSIZE_UNLOCK(uip); 312 313 racct_sub_cred(cred, RACCT_SWAP, decr); 314 } 315 316 static void swapdev_strategy(struct buf *, struct swdevt *sw); 317 318 #define SWM_FREE 0x02 /* free, period */ 319 #define SWM_POP 0x04 /* pop out */ 320 321 int swap_pager_full = 2; /* swap space exhaustion (task killing) */ 322 static int swap_pager_almost_full = 1; /* swap space exhaustion (w/hysteresis)*/ 323 static int nsw_rcount; /* free read buffers */ 324 static int nsw_wcount_sync; /* limit write buffers / synchronous */ 325 static int nsw_wcount_async; /* limit write buffers / asynchronous */ 326 static int nsw_wcount_async_max;/* assigned maximum */ 327 static int nsw_cluster_max; /* maximum VOP I/O allowed */ 328 329 static int sysctl_swap_async_max(SYSCTL_HANDLER_ARGS); 330 SYSCTL_PROC(_vm, OID_AUTO, swap_async_max, CTLTYPE_INT | CTLFLAG_RW, 331 NULL, 0, sysctl_swap_async_max, "I", "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 mtx sw_alloc_mtx; /* protect list manipulation */ 350 static struct pagerlst swap_pager_object_list[NOBJLISTS]; 351 static uma_zone_t swap_zone; 352 353 /* 354 * pagerops for OBJT_SWAP - "swap pager". Some ops are also global procedure 355 * calls hooked from other parts of the VM system and do not appear here. 356 * (see vm/swap_pager.h). 357 */ 358 static vm_object_t 359 swap_pager_alloc(void *handle, vm_ooffset_t size, 360 vm_prot_t prot, vm_ooffset_t offset, struct ucred *); 361 static void swap_pager_dealloc(vm_object_t object); 362 static int swap_pager_getpages(vm_object_t, vm_page_t *, int, int); 363 static int swap_pager_getpages_async(vm_object_t, vm_page_t *, int, int, 364 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, 395 CTLFLAG_RD, &dmmax, 0, "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 thread *, 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, daddr_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 static void 419 swp_pager_free_nrpage(vm_page_t m) 420 { 421 422 vm_page_lock(m); 423 if (m->wire_count == 0) 424 vm_page_free(m); 425 vm_page_unlock(m); 426 } 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 * SWP_PAGER_HASH() - hash swap meta data 457 * 458 * This is an helper function which hashes the swapblk given 459 * the object and page index. It returns a pointer to a pointer 460 * to the object, or a pointer to a NULL pointer if it could not 461 * find a swapblk. 462 */ 463 static struct swblock ** 464 swp_pager_hash(vm_object_t object, vm_pindex_t index) 465 { 466 struct swblock **pswap; 467 struct swblock *swap; 468 469 index &= ~(vm_pindex_t)SWAP_META_MASK; 470 pswap = &swhash[(index ^ (int)(intptr_t)object) & swhash_mask]; 471 while ((swap = *pswap) != NULL) { 472 if (swap->swb_object == object && 473 swap->swb_index == index 474 ) { 475 break; 476 } 477 pswap = &swap->swb_hnext; 478 } 479 return (pswap); 480 } 481 482 /* 483 * SWAP_PAGER_INIT() - initialize the swap pager! 484 * 485 * Expected to be started from system init. NOTE: This code is run 486 * before much else so be careful what you depend on. Most of the VM 487 * system has yet to be initialized at this point. 488 */ 489 static void 490 swap_pager_init(void) 491 { 492 /* 493 * Initialize object lists 494 */ 495 int i; 496 497 for (i = 0; i < NOBJLISTS; ++i) 498 TAILQ_INIT(&swap_pager_object_list[i]); 499 mtx_init(&sw_alloc_mtx, "swap_pager list", NULL, MTX_DEF); 500 mtx_init(&sw_dev_mtx, "swapdev", NULL, MTX_DEF); 501 502 /* 503 * Device Stripe, in PAGE_SIZE'd blocks 504 */ 505 dmmax = SWB_NPAGES * 2; 506 } 507 508 /* 509 * SWAP_PAGER_SWAP_INIT() - swap pager initialization from pageout process 510 * 511 * Expected to be started from pageout process once, prior to entering 512 * its main loop. 513 */ 514 void 515 swap_pager_swap_init(void) 516 { 517 unsigned long n, n2; 518 519 /* 520 * Number of in-transit swap bp operations. Don't 521 * exhaust the pbufs completely. Make sure we 522 * initialize workable values (0 will work for hysteresis 523 * but it isn't very efficient). 524 * 525 * The nsw_cluster_max is constrained by the bp->b_pages[] 526 * array (MAXPHYS/PAGE_SIZE) and our locally defined 527 * MAX_PAGEOUT_CLUSTER. Also be aware that swap ops are 528 * constrained by the swap device interleave stripe size. 529 * 530 * Currently we hardwire nsw_wcount_async to 4. This limit is 531 * designed to prevent other I/O from having high latencies due to 532 * our pageout I/O. The value 4 works well for one or two active swap 533 * devices but is probably a little low if you have more. Even so, 534 * a higher value would probably generate only a limited improvement 535 * with three or four active swap devices since the system does not 536 * typically have to pageout at extreme bandwidths. We will want 537 * at least 2 per swap devices, and 4 is a pretty good value if you 538 * have one NFS swap device due to the command/ack latency over NFS. 539 * So it all works out pretty well. 540 */ 541 nsw_cluster_max = min((MAXPHYS/PAGE_SIZE), MAX_PAGEOUT_CLUSTER); 542 543 mtx_lock(&pbuf_mtx); 544 nsw_rcount = (nswbuf + 1) / 2; 545 nsw_wcount_sync = (nswbuf + 3) / 4; 546 nsw_wcount_async = 4; 547 nsw_wcount_async_max = nsw_wcount_async; 548 mtx_unlock(&pbuf_mtx); 549 550 /* 551 * Initialize our zone. Right now I'm just guessing on the number 552 * we need based on the number of pages in the system. Each swblock 553 * can hold 32 pages, so this is probably overkill. This reservation 554 * is typically limited to around 32MB by default. 555 */ 556 n = vm_cnt.v_page_count / 2; 557 if (maxswzone && n > maxswzone / sizeof(struct swblock)) 558 n = maxswzone / sizeof(struct swblock); 559 n2 = n; 560 swap_zone = uma_zcreate("SWAPMETA", sizeof(struct swblock), NULL, NULL, 561 NULL, NULL, UMA_ALIGN_PTR, UMA_ZONE_NOFREE | UMA_ZONE_VM); 562 if (swap_zone == NULL) 563 panic("failed to create swap_zone."); 564 do { 565 if (uma_zone_reserve_kva(swap_zone, n)) 566 break; 567 /* 568 * if the allocation failed, try a zone two thirds the 569 * size of the previous attempt. 570 */ 571 n -= ((n + 2) / 3); 572 } while (n > 0); 573 if (n2 != n) 574 printf("Swap zone entries reduced from %lu to %lu.\n", n2, n); 575 swap_maxpages = n * SWAP_META_PAGES; 576 swzone = n * sizeof(struct swblock); 577 n2 = n; 578 579 /* 580 * Initialize our meta-data hash table. The swapper does not need to 581 * be quite as efficient as the VM system, so we do not use an 582 * oversized hash table. 583 * 584 * n: size of hash table, must be power of 2 585 * swhash_mask: hash table index mask 586 */ 587 for (n = 1; n < n2 / 8; n *= 2) 588 ; 589 swhash = malloc(sizeof(struct swblock *) * n, M_VMPGDATA, M_WAITOK | M_ZERO); 590 swhash_mask = n - 1; 591 mtx_init(&swhash_mtx, "swap_pager swhash", NULL, MTX_DEF); 592 } 593 594 /* 595 * SWAP_PAGER_ALLOC() - allocate a new OBJT_SWAP VM object and instantiate 596 * its metadata structures. 597 * 598 * This routine is called from the mmap and fork code to create a new 599 * OBJT_SWAP object. We do this by creating an OBJT_DEFAULT object 600 * and then converting it with swp_pager_meta_build(). 601 * 602 * This routine may block in vm_object_allocate() and create a named 603 * object lookup race, so we must interlock. 604 * 605 * MPSAFE 606 */ 607 static vm_object_t 608 swap_pager_alloc(void *handle, vm_ooffset_t size, vm_prot_t prot, 609 vm_ooffset_t offset, struct ucred *cred) 610 { 611 vm_object_t object; 612 vm_pindex_t pindex; 613 614 pindex = OFF_TO_IDX(offset + PAGE_MASK + size); 615 if (handle) { 616 mtx_lock(&Giant); 617 /* 618 * Reference existing named region or allocate new one. There 619 * should not be a race here against swp_pager_meta_build() 620 * as called from vm_page_remove() in regards to the lookup 621 * of the handle. 622 */ 623 sx_xlock(&sw_alloc_sx); 624 object = vm_pager_object_lookup(NOBJLIST(handle), handle); 625 if (object == NULL) { 626 if (cred != NULL) { 627 if (!swap_reserve_by_cred(size, cred)) { 628 sx_xunlock(&sw_alloc_sx); 629 mtx_unlock(&Giant); 630 return (NULL); 631 } 632 crhold(cred); 633 } 634 object = vm_object_allocate(OBJT_DEFAULT, pindex); 635 VM_OBJECT_WLOCK(object); 636 object->handle = handle; 637 if (cred != NULL) { 638 object->cred = cred; 639 object->charge = size; 640 } 641 swp_pager_meta_build(object, 0, SWAPBLK_NONE); 642 VM_OBJECT_WUNLOCK(object); 643 } 644 sx_xunlock(&sw_alloc_sx); 645 mtx_unlock(&Giant); 646 } else { 647 if (cred != NULL) { 648 if (!swap_reserve_by_cred(size, cred)) 649 return (NULL); 650 crhold(cred); 651 } 652 object = vm_object_allocate(OBJT_DEFAULT, pindex); 653 VM_OBJECT_WLOCK(object); 654 if (cred != NULL) { 655 object->cred = cred; 656 object->charge = size; 657 } 658 swp_pager_meta_build(object, 0, SWAPBLK_NONE); 659 VM_OBJECT_WUNLOCK(object); 660 } 661 return (object); 662 } 663 664 /* 665 * SWAP_PAGER_DEALLOC() - remove swap metadata from object 666 * 667 * The swap backing for the object is destroyed. The code is 668 * designed such that we can reinstantiate it later, but this 669 * routine is typically called only when the entire object is 670 * about to be destroyed. 671 * 672 * The object must be locked. 673 */ 674 static void 675 swap_pager_dealloc(vm_object_t object) 676 { 677 678 /* 679 * Remove from list right away so lookups will fail if we block for 680 * pageout completion. 681 */ 682 if (object->handle != NULL) { 683 mtx_lock(&sw_alloc_mtx); 684 TAILQ_REMOVE(NOBJLIST(object->handle), object, pager_object_list); 685 mtx_unlock(&sw_alloc_mtx); 686 } 687 688 VM_OBJECT_ASSERT_WLOCKED(object); 689 vm_object_pip_wait(object, "swpdea"); 690 691 /* 692 * Free all remaining metadata. We only bother to free it from 693 * the swap meta data. We do not attempt to free swapblk's still 694 * associated with vm_page_t's for this object. We do not care 695 * if paging is still in progress on some objects. 696 */ 697 swp_pager_meta_free_all(object); 698 object->handle = NULL; 699 object->type = OBJT_DEAD; 700 } 701 702 /************************************************************************ 703 * SWAP PAGER BITMAP ROUTINES * 704 ************************************************************************/ 705 706 /* 707 * SWP_PAGER_GETSWAPSPACE() - allocate raw swap space 708 * 709 * Allocate swap for the requested number of pages. The starting 710 * swap block number (a page index) is returned or SWAPBLK_NONE 711 * if the allocation failed. 712 * 713 * Also has the side effect of advising that somebody made a mistake 714 * when they configured swap and didn't configure enough. 715 * 716 * This routine may not sleep. 717 * 718 * We allocate in round-robin fashion from the configured devices. 719 */ 720 static daddr_t 721 swp_pager_getswapspace(int npages) 722 { 723 daddr_t blk; 724 struct swdevt *sp; 725 int i; 726 727 blk = SWAPBLK_NONE; 728 mtx_lock(&sw_dev_mtx); 729 sp = swdevhd; 730 for (i = 0; i < nswapdev; i++) { 731 if (sp == NULL) 732 sp = TAILQ_FIRST(&swtailq); 733 if (!(sp->sw_flags & SW_CLOSING)) { 734 blk = blist_alloc(sp->sw_blist, npages); 735 if (blk != SWAPBLK_NONE) { 736 blk += sp->sw_first; 737 sp->sw_used += npages; 738 swap_pager_avail -= npages; 739 swp_sizecheck(); 740 swdevhd = TAILQ_NEXT(sp, sw_list); 741 goto done; 742 } 743 } 744 sp = TAILQ_NEXT(sp, sw_list); 745 } 746 if (swap_pager_full != 2) { 747 printf("swap_pager_getswapspace(%d): failed\n", npages); 748 swap_pager_full = 2; 749 swap_pager_almost_full = 1; 750 } 751 swdevhd = NULL; 752 done: 753 mtx_unlock(&sw_dev_mtx); 754 return (blk); 755 } 756 757 static int 758 swp_pager_isondev(daddr_t blk, struct swdevt *sp) 759 { 760 761 return (blk >= sp->sw_first && blk < sp->sw_end); 762 } 763 764 static void 765 swp_pager_strategy(struct buf *bp) 766 { 767 struct swdevt *sp; 768 769 mtx_lock(&sw_dev_mtx); 770 TAILQ_FOREACH(sp, &swtailq, sw_list) { 771 if (bp->b_blkno >= sp->sw_first && bp->b_blkno < sp->sw_end) { 772 mtx_unlock(&sw_dev_mtx); 773 if ((sp->sw_flags & SW_UNMAPPED) != 0 && 774 unmapped_buf_allowed) { 775 bp->b_kvaalloc = bp->b_data; 776 bp->b_data = unmapped_buf; 777 bp->b_kvabase = unmapped_buf; 778 bp->b_offset = 0; 779 bp->b_flags |= B_UNMAPPED; 780 } else { 781 pmap_qenter((vm_offset_t)bp->b_data, 782 &bp->b_pages[0], bp->b_bcount / PAGE_SIZE); 783 } 784 sp->sw_strategy(bp, sp); 785 return; 786 } 787 } 788 panic("Swapdev not found"); 789 } 790 791 792 /* 793 * SWP_PAGER_FREESWAPSPACE() - free raw swap space 794 * 795 * This routine returns the specified swap blocks back to the bitmap. 796 * 797 * This routine may not sleep. 798 */ 799 static void 800 swp_pager_freeswapspace(daddr_t blk, int npages) 801 { 802 struct swdevt *sp; 803 804 mtx_lock(&sw_dev_mtx); 805 TAILQ_FOREACH(sp, &swtailq, sw_list) { 806 if (blk >= sp->sw_first && blk < sp->sw_end) { 807 sp->sw_used -= npages; 808 /* 809 * If we are attempting to stop swapping on 810 * this device, we don't want to mark any 811 * blocks free lest they be reused. 812 */ 813 if ((sp->sw_flags & SW_CLOSING) == 0) { 814 blist_free(sp->sw_blist, blk - sp->sw_first, 815 npages); 816 swap_pager_avail += npages; 817 swp_sizecheck(); 818 } 819 mtx_unlock(&sw_dev_mtx); 820 return; 821 } 822 } 823 panic("Swapdev not found"); 824 } 825 826 /* 827 * SWAP_PAGER_FREESPACE() - frees swap blocks associated with a page 828 * range within an object. 829 * 830 * This is a globally accessible routine. 831 * 832 * This routine removes swapblk assignments from swap metadata. 833 * 834 * The external callers of this routine typically have already destroyed 835 * or renamed vm_page_t's associated with this range in the object so 836 * we should be ok. 837 * 838 * The object must be locked. 839 */ 840 void 841 swap_pager_freespace(vm_object_t object, vm_pindex_t start, vm_size_t size) 842 { 843 844 swp_pager_meta_free(object, start, size); 845 } 846 847 /* 848 * SWAP_PAGER_RESERVE() - reserve swap blocks in object 849 * 850 * Assigns swap blocks to the specified range within the object. The 851 * swap blocks are not zeroed. Any previous swap assignment is destroyed. 852 * 853 * Returns 0 on success, -1 on failure. 854 */ 855 int 856 swap_pager_reserve(vm_object_t object, vm_pindex_t start, vm_size_t size) 857 { 858 int n = 0; 859 daddr_t blk = SWAPBLK_NONE; 860 vm_pindex_t beg = start; /* save start index */ 861 862 VM_OBJECT_WLOCK(object); 863 while (size) { 864 if (n == 0) { 865 n = BLIST_MAX_ALLOC; 866 while ((blk = swp_pager_getswapspace(n)) == SWAPBLK_NONE) { 867 n >>= 1; 868 if (n == 0) { 869 swp_pager_meta_free(object, beg, start - beg); 870 VM_OBJECT_WUNLOCK(object); 871 return (-1); 872 } 873 } 874 } 875 swp_pager_meta_build(object, start, blk); 876 --size; 877 ++start; 878 ++blk; 879 --n; 880 } 881 swp_pager_meta_free(object, start, n); 882 VM_OBJECT_WUNLOCK(object); 883 return (0); 884 } 885 886 /* 887 * SWAP_PAGER_COPY() - copy blocks from source pager to destination pager 888 * and destroy the source. 889 * 890 * Copy any valid swapblks from the source to the destination. In 891 * cases where both the source and destination have a valid swapblk, 892 * we keep the destination's. 893 * 894 * This routine is allowed to sleep. It may sleep allocating metadata 895 * indirectly through swp_pager_meta_build() or if paging is still in 896 * progress on the source. 897 * 898 * The source object contains no vm_page_t's (which is just as well) 899 * 900 * The source object is of type OBJT_SWAP. 901 * 902 * The source and destination objects must be locked. 903 * Both object locks may temporarily be released. 904 */ 905 void 906 swap_pager_copy(vm_object_t srcobject, vm_object_t dstobject, 907 vm_pindex_t offset, int destroysource) 908 { 909 vm_pindex_t i; 910 911 VM_OBJECT_ASSERT_WLOCKED(srcobject); 912 VM_OBJECT_ASSERT_WLOCKED(dstobject); 913 914 /* 915 * If destroysource is set, we remove the source object from the 916 * swap_pager internal queue now. 917 */ 918 if (destroysource) { 919 if (srcobject->handle != NULL) { 920 mtx_lock(&sw_alloc_mtx); 921 TAILQ_REMOVE( 922 NOBJLIST(srcobject->handle), 923 srcobject, 924 pager_object_list 925 ); 926 mtx_unlock(&sw_alloc_mtx); 927 } 928 } 929 930 /* 931 * transfer source to destination. 932 */ 933 for (i = 0; i < dstobject->size; ++i) { 934 daddr_t dstaddr; 935 936 /* 937 * Locate (without changing) the swapblk on the destination, 938 * unless it is invalid in which case free it silently, or 939 * if the destination is a resident page, in which case the 940 * source is thrown away. 941 */ 942 dstaddr = swp_pager_meta_ctl(dstobject, i, 0); 943 944 if (dstaddr == SWAPBLK_NONE) { 945 /* 946 * Destination has no swapblk and is not resident, 947 * copy source. 948 */ 949 daddr_t srcaddr; 950 951 srcaddr = swp_pager_meta_ctl( 952 srcobject, 953 i + offset, 954 SWM_POP 955 ); 956 957 if (srcaddr != SWAPBLK_NONE) { 958 /* 959 * swp_pager_meta_build() can sleep. 960 */ 961 vm_object_pip_add(srcobject, 1); 962 VM_OBJECT_WUNLOCK(srcobject); 963 vm_object_pip_add(dstobject, 1); 964 swp_pager_meta_build(dstobject, i, srcaddr); 965 vm_object_pip_wakeup(dstobject); 966 VM_OBJECT_WLOCK(srcobject); 967 vm_object_pip_wakeup(srcobject); 968 } 969 } else { 970 /* 971 * Destination has valid swapblk or it is represented 972 * by a resident page. We destroy the sourceblock. 973 */ 974 975 swp_pager_meta_ctl(srcobject, i + offset, SWM_FREE); 976 } 977 } 978 979 /* 980 * Free left over swap blocks in source. 981 * 982 * We have to revert the type to OBJT_DEFAULT so we do not accidently 983 * double-remove the object from the swap queues. 984 */ 985 if (destroysource) { 986 swp_pager_meta_free_all(srcobject); 987 /* 988 * Reverting the type is not necessary, the caller is going 989 * to destroy srcobject directly, but I'm doing it here 990 * for consistency since we've removed the object from its 991 * queues. 992 */ 993 srcobject->type = OBJT_DEFAULT; 994 } 995 } 996 997 /* 998 * SWAP_PAGER_HASPAGE() - determine if we have good backing store for 999 * the requested page. 1000 * 1001 * We determine whether good backing store exists for the requested 1002 * page and return TRUE if it does, FALSE if it doesn't. 1003 * 1004 * If TRUE, we also try to determine how much valid, contiguous backing 1005 * store exists before and after the requested page within a reasonable 1006 * distance. We do not try to restrict it to the swap device stripe 1007 * (that is handled in getpages/putpages). It probably isn't worth 1008 * doing here. 1009 */ 1010 static boolean_t 1011 swap_pager_haspage(vm_object_t object, vm_pindex_t pindex, int *before, int *after) 1012 { 1013 daddr_t blk0; 1014 1015 VM_OBJECT_ASSERT_LOCKED(object); 1016 /* 1017 * do we have good backing store at the requested index ? 1018 */ 1019 blk0 = swp_pager_meta_ctl(object, pindex, 0); 1020 1021 if (blk0 == SWAPBLK_NONE) { 1022 if (before) 1023 *before = 0; 1024 if (after) 1025 *after = 0; 1026 return (FALSE); 1027 } 1028 1029 /* 1030 * find backwards-looking contiguous good backing store 1031 */ 1032 if (before != NULL) { 1033 int i; 1034 1035 for (i = 1; i < (SWB_NPAGES/2); ++i) { 1036 daddr_t blk; 1037 1038 if (i > pindex) 1039 break; 1040 blk = swp_pager_meta_ctl(object, pindex - i, 0); 1041 if (blk != blk0 - i) 1042 break; 1043 } 1044 *before = (i - 1); 1045 } 1046 1047 /* 1048 * find forward-looking contiguous good backing store 1049 */ 1050 if (after != NULL) { 1051 int i; 1052 1053 for (i = 1; i < (SWB_NPAGES/2); ++i) { 1054 daddr_t blk; 1055 1056 blk = swp_pager_meta_ctl(object, pindex + i, 0); 1057 if (blk != blk0 + i) 1058 break; 1059 } 1060 *after = (i - 1); 1061 } 1062 return (TRUE); 1063 } 1064 1065 /* 1066 * SWAP_PAGER_PAGE_UNSWAPPED() - remove swap backing store related to page 1067 * 1068 * This removes any associated swap backing store, whether valid or 1069 * not, from the page. 1070 * 1071 * This routine is typically called when a page is made dirty, at 1072 * which point any associated swap can be freed. MADV_FREE also 1073 * calls us in a special-case situation 1074 * 1075 * NOTE!!! If the page is clean and the swap was valid, the caller 1076 * should make the page dirty before calling this routine. This routine 1077 * does NOT change the m->dirty status of the page. Also: MADV_FREE 1078 * depends on it. 1079 * 1080 * This routine may not sleep. 1081 * 1082 * The object containing the page must be locked. 1083 */ 1084 static void 1085 swap_pager_unswapped(vm_page_t m) 1086 { 1087 1088 swp_pager_meta_ctl(m->object, m->pindex, SWM_FREE); 1089 } 1090 1091 /* 1092 * SWAP_PAGER_GETPAGES() - bring pages in from swap 1093 * 1094 * Attempt to retrieve (m, count) pages from backing store, but make 1095 * sure we retrieve at least m[reqpage]. We try to load in as large 1096 * a chunk surrounding m[reqpage] as is contiguous in swap and which 1097 * belongs to the same object. 1098 * 1099 * The code is designed for asynchronous operation and 1100 * immediate-notification of 'reqpage' but tends not to be 1101 * used that way. Please do not optimize-out this algorithmic 1102 * feature, I intend to improve on it in the future. 1103 * 1104 * The parent has a single vm_object_pip_add() reference prior to 1105 * calling us and we should return with the same. 1106 * 1107 * The parent has BUSY'd the pages. We should return with 'm' 1108 * left busy, but the others adjusted. 1109 */ 1110 static int 1111 swap_pager_getpages(vm_object_t object, vm_page_t *m, int count, int reqpage) 1112 { 1113 struct buf *bp; 1114 vm_page_t mreq; 1115 int i; 1116 int j; 1117 daddr_t blk; 1118 1119 mreq = m[reqpage]; 1120 1121 /* 1122 * Calculate range to retrieve. The pages have already been assigned 1123 * their swapblks. We require a *contiguous* range but we know it to 1124 * not span devices. If we do not supply it, bad things 1125 * happen. Note that blk, iblk & jblk can be SWAPBLK_NONE, but the 1126 * loops are set up such that the case(s) are handled implicitly. 1127 * 1128 * The swp_*() calls must be made with the object locked. 1129 */ 1130 blk = swp_pager_meta_ctl(mreq->object, mreq->pindex, 0); 1131 1132 for (i = reqpage - 1; i >= 0; --i) { 1133 daddr_t iblk; 1134 1135 iblk = swp_pager_meta_ctl(m[i]->object, m[i]->pindex, 0); 1136 if (blk != iblk + (reqpage - i)) 1137 break; 1138 } 1139 ++i; 1140 1141 for (j = reqpage + 1; j < count; ++j) { 1142 daddr_t jblk; 1143 1144 jblk = swp_pager_meta_ctl(m[j]->object, m[j]->pindex, 0); 1145 if (blk != jblk - (j - reqpage)) 1146 break; 1147 } 1148 1149 /* 1150 * free pages outside our collection range. Note: we never free 1151 * mreq, it must remain busy throughout. 1152 */ 1153 if (0 < i || j < count) { 1154 int k; 1155 1156 for (k = 0; k < i; ++k) 1157 swp_pager_free_nrpage(m[k]); 1158 for (k = j; k < count; ++k) 1159 swp_pager_free_nrpage(m[k]); 1160 } 1161 1162 /* 1163 * Return VM_PAGER_FAIL if we have nothing to do. Return mreq 1164 * still busy, but the others unbusied. 1165 */ 1166 if (blk == SWAPBLK_NONE) 1167 return (VM_PAGER_FAIL); 1168 1169 /* 1170 * Getpbuf() can sleep. 1171 */ 1172 VM_OBJECT_WUNLOCK(object); 1173 /* 1174 * Get a swap buffer header to perform the IO 1175 */ 1176 bp = getpbuf(&nsw_rcount); 1177 bp->b_flags |= B_PAGING; 1178 1179 bp->b_iocmd = BIO_READ; 1180 bp->b_iodone = swp_pager_async_iodone; 1181 bp->b_rcred = crhold(thread0.td_ucred); 1182 bp->b_wcred = crhold(thread0.td_ucred); 1183 bp->b_blkno = blk - (reqpage - i); 1184 bp->b_bcount = PAGE_SIZE * (j - i); 1185 bp->b_bufsize = PAGE_SIZE * (j - i); 1186 bp->b_pager.pg_reqpage = reqpage - i; 1187 1188 VM_OBJECT_WLOCK(object); 1189 { 1190 int k; 1191 1192 for (k = i; k < j; ++k) { 1193 bp->b_pages[k - i] = m[k]; 1194 m[k]->oflags |= VPO_SWAPINPROG; 1195 } 1196 } 1197 bp->b_npages = j - i; 1198 1199 PCPU_INC(cnt.v_swapin); 1200 PCPU_ADD(cnt.v_swappgsin, bp->b_npages); 1201 1202 /* 1203 * We still hold the lock on mreq, and our automatic completion routine 1204 * does not remove it. 1205 */ 1206 vm_object_pip_add(object, bp->b_npages); 1207 VM_OBJECT_WUNLOCK(object); 1208 1209 /* 1210 * perform the I/O. NOTE!!! bp cannot be considered valid after 1211 * this point because we automatically release it on completion. 1212 * Instead, we look at the one page we are interested in which we 1213 * still hold a lock on even through the I/O completion. 1214 * 1215 * The other pages in our m[] array are also released on completion, 1216 * so we cannot assume they are valid anymore either. 1217 * 1218 * NOTE: b_blkno is destroyed by the call to swapdev_strategy 1219 */ 1220 BUF_KERNPROC(bp); 1221 swp_pager_strategy(bp); 1222 1223 /* 1224 * wait for the page we want to complete. VPO_SWAPINPROG is always 1225 * cleared on completion. If an I/O error occurs, SWAPBLK_NONE 1226 * is set in the meta-data. 1227 */ 1228 VM_OBJECT_WLOCK(object); 1229 while ((mreq->oflags & VPO_SWAPINPROG) != 0) { 1230 mreq->oflags |= VPO_SWAPSLEEP; 1231 PCPU_INC(cnt.v_intrans); 1232 if (VM_OBJECT_SLEEP(object, &object->paging_in_progress, PSWP, 1233 "swread", hz * 20)) { 1234 printf( 1235 "swap_pager: indefinite wait buffer: bufobj: %p, blkno: %jd, size: %ld\n", 1236 bp->b_bufobj, (intmax_t)bp->b_blkno, bp->b_bcount); 1237 } 1238 } 1239 1240 /* 1241 * mreq is left busied after completion, but all the other pages 1242 * are freed. If we had an unrecoverable read error the page will 1243 * not be valid. 1244 */ 1245 if (mreq->valid != VM_PAGE_BITS_ALL) { 1246 return (VM_PAGER_ERROR); 1247 } else { 1248 return (VM_PAGER_OK); 1249 } 1250 1251 /* 1252 * A final note: in a low swap situation, we cannot deallocate swap 1253 * and mark a page dirty here because the caller is likely to mark 1254 * the page clean when we return, causing the page to possibly revert 1255 * to all-zero's later. 1256 */ 1257 } 1258 1259 /* 1260 * swap_pager_getpages_async(): 1261 * 1262 * Right now this is emulation of asynchronous operation on top of 1263 * swap_pager_getpages(). 1264 */ 1265 static int 1266 swap_pager_getpages_async(vm_object_t object, vm_page_t *m, int count, 1267 int reqpage, pgo_getpages_iodone_t iodone, void *arg) 1268 { 1269 int r, error; 1270 1271 r = swap_pager_getpages(object, m, count, reqpage); 1272 VM_OBJECT_WUNLOCK(object); 1273 switch (r) { 1274 case VM_PAGER_OK: 1275 error = 0; 1276 break; 1277 case VM_PAGER_ERROR: 1278 error = EIO; 1279 break; 1280 case VM_PAGER_FAIL: 1281 error = EINVAL; 1282 break; 1283 default: 1284 panic("unhandled swap_pager_getpages() error %d", r); 1285 } 1286 (iodone)(arg, m, count, error); 1287 VM_OBJECT_WLOCK(object); 1288 1289 return (r); 1290 } 1291 1292 /* 1293 * swap_pager_putpages: 1294 * 1295 * Assign swap (if necessary) and initiate I/O on the specified pages. 1296 * 1297 * We support both OBJT_DEFAULT and OBJT_SWAP objects. DEFAULT objects 1298 * are automatically converted to SWAP objects. 1299 * 1300 * In a low memory situation we may block in VOP_STRATEGY(), but the new 1301 * vm_page reservation system coupled with properly written VFS devices 1302 * should ensure that no low-memory deadlock occurs. This is an area 1303 * which needs work. 1304 * 1305 * The parent has N vm_object_pip_add() references prior to 1306 * calling us and will remove references for rtvals[] that are 1307 * not set to VM_PAGER_PEND. We need to remove the rest on I/O 1308 * completion. 1309 * 1310 * The parent has soft-busy'd the pages it passes us and will unbusy 1311 * those whos rtvals[] entry is not set to VM_PAGER_PEND on return. 1312 * We need to unbusy the rest on I/O completion. 1313 */ 1314 void 1315 swap_pager_putpages(vm_object_t object, vm_page_t *m, int count, 1316 int flags, int *rtvals) 1317 { 1318 int i, n; 1319 boolean_t sync; 1320 1321 if (count && m[0]->object != object) { 1322 panic("swap_pager_putpages: object mismatch %p/%p", 1323 object, 1324 m[0]->object 1325 ); 1326 } 1327 1328 /* 1329 * Step 1 1330 * 1331 * Turn object into OBJT_SWAP 1332 * check for bogus sysops 1333 * force sync if not pageout process 1334 */ 1335 if (object->type != OBJT_SWAP) 1336 swp_pager_meta_build(object, 0, SWAPBLK_NONE); 1337 VM_OBJECT_WUNLOCK(object); 1338 1339 n = 0; 1340 if (curproc != pageproc) 1341 sync = TRUE; 1342 else 1343 sync = (flags & VM_PAGER_PUT_SYNC) != 0; 1344 1345 /* 1346 * Step 2 1347 * 1348 * Assign swap blocks and issue I/O. We reallocate swap on the fly. 1349 * The page is left dirty until the pageout operation completes 1350 * successfully. 1351 */ 1352 for (i = 0; i < count; i += n) { 1353 int j; 1354 struct buf *bp; 1355 daddr_t blk; 1356 1357 /* 1358 * Maximum I/O size is limited by a number of factors. 1359 */ 1360 n = min(BLIST_MAX_ALLOC, count - i); 1361 n = min(n, nsw_cluster_max); 1362 1363 /* 1364 * Get biggest block of swap we can. If we fail, fall 1365 * back and try to allocate a smaller block. Don't go 1366 * overboard trying to allocate space if it would overly 1367 * fragment swap. 1368 */ 1369 while ( 1370 (blk = swp_pager_getswapspace(n)) == SWAPBLK_NONE && 1371 n > 4 1372 ) { 1373 n >>= 1; 1374 } 1375 if (blk == SWAPBLK_NONE) { 1376 for (j = 0; j < n; ++j) 1377 rtvals[i+j] = VM_PAGER_FAIL; 1378 continue; 1379 } 1380 1381 /* 1382 * All I/O parameters have been satisfied, build the I/O 1383 * request and assign the swap space. 1384 */ 1385 if (sync == TRUE) { 1386 bp = getpbuf(&nsw_wcount_sync); 1387 } else { 1388 bp = getpbuf(&nsw_wcount_async); 1389 bp->b_flags = B_ASYNC; 1390 } 1391 bp->b_flags |= B_PAGING; 1392 bp->b_iocmd = BIO_WRITE; 1393 1394 bp->b_rcred = crhold(thread0.td_ucred); 1395 bp->b_wcred = crhold(thread0.td_ucred); 1396 bp->b_bcount = PAGE_SIZE * n; 1397 bp->b_bufsize = PAGE_SIZE * n; 1398 bp->b_blkno = blk; 1399 1400 VM_OBJECT_WLOCK(object); 1401 for (j = 0; j < n; ++j) { 1402 vm_page_t mreq = m[i+j]; 1403 1404 swp_pager_meta_build( 1405 mreq->object, 1406 mreq->pindex, 1407 blk + j 1408 ); 1409 vm_page_dirty(mreq); 1410 rtvals[i+j] = VM_PAGER_OK; 1411 1412 mreq->oflags |= VPO_SWAPINPROG; 1413 bp->b_pages[j] = mreq; 1414 } 1415 VM_OBJECT_WUNLOCK(object); 1416 bp->b_npages = n; 1417 /* 1418 * Must set dirty range for NFS to work. 1419 */ 1420 bp->b_dirtyoff = 0; 1421 bp->b_dirtyend = bp->b_bcount; 1422 1423 PCPU_INC(cnt.v_swapout); 1424 PCPU_ADD(cnt.v_swappgsout, bp->b_npages); 1425 1426 /* 1427 * asynchronous 1428 * 1429 * NOTE: b_blkno is destroyed by the call to swapdev_strategy 1430 */ 1431 if (sync == FALSE) { 1432 bp->b_iodone = swp_pager_async_iodone; 1433 BUF_KERNPROC(bp); 1434 swp_pager_strategy(bp); 1435 1436 for (j = 0; j < n; ++j) 1437 rtvals[i+j] = VM_PAGER_PEND; 1438 /* restart outter loop */ 1439 continue; 1440 } 1441 1442 /* 1443 * synchronous 1444 * 1445 * NOTE: b_blkno is destroyed by the call to swapdev_strategy 1446 */ 1447 bp->b_iodone = bdone; 1448 swp_pager_strategy(bp); 1449 1450 /* 1451 * Wait for the sync I/O to complete, then update rtvals. 1452 * We just set the rtvals[] to VM_PAGER_PEND so we can call 1453 * our async completion routine at the end, thus avoiding a 1454 * double-free. 1455 */ 1456 bwait(bp, PVM, "swwrt"); 1457 for (j = 0; j < n; ++j) 1458 rtvals[i+j] = VM_PAGER_PEND; 1459 /* 1460 * Now that we are through with the bp, we can call the 1461 * normal async completion, which frees everything up. 1462 */ 1463 swp_pager_async_iodone(bp); 1464 } 1465 VM_OBJECT_WLOCK(object); 1466 } 1467 1468 /* 1469 * swp_pager_async_iodone: 1470 * 1471 * Completion routine for asynchronous reads and writes from/to swap. 1472 * Also called manually by synchronous code to finish up a bp. 1473 * 1474 * This routine may not sleep. 1475 */ 1476 static void 1477 swp_pager_async_iodone(struct buf *bp) 1478 { 1479 int i; 1480 vm_object_t object = NULL; 1481 1482 /* 1483 * report error 1484 */ 1485 if (bp->b_ioflags & BIO_ERROR) { 1486 printf( 1487 "swap_pager: I/O error - %s failed; blkno %ld," 1488 "size %ld, error %d\n", 1489 ((bp->b_iocmd == BIO_READ) ? "pagein" : "pageout"), 1490 (long)bp->b_blkno, 1491 (long)bp->b_bcount, 1492 bp->b_error 1493 ); 1494 } 1495 1496 /* 1497 * remove the mapping for kernel virtual 1498 */ 1499 if ((bp->b_flags & B_UNMAPPED) != 0) { 1500 bp->b_data = bp->b_kvaalloc; 1501 bp->b_kvabase = bp->b_kvaalloc; 1502 bp->b_flags &= ~B_UNMAPPED; 1503 } else 1504 pmap_qremove((vm_offset_t)bp->b_data, bp->b_npages); 1505 1506 if (bp->b_npages) { 1507 object = bp->b_pages[0]->object; 1508 VM_OBJECT_WLOCK(object); 1509 } 1510 1511 /* 1512 * cleanup pages. If an error occurs writing to swap, we are in 1513 * very serious trouble. If it happens to be a disk error, though, 1514 * we may be able to recover by reassigning the swap later on. So 1515 * in this case we remove the m->swapblk assignment for the page 1516 * but do not free it in the rlist. The errornous block(s) are thus 1517 * never reallocated as swap. Redirty the page and continue. 1518 */ 1519 for (i = 0; i < bp->b_npages; ++i) { 1520 vm_page_t m = bp->b_pages[i]; 1521 1522 m->oflags &= ~VPO_SWAPINPROG; 1523 if (m->oflags & VPO_SWAPSLEEP) { 1524 m->oflags &= ~VPO_SWAPSLEEP; 1525 wakeup(&object->paging_in_progress); 1526 } 1527 1528 if (bp->b_ioflags & BIO_ERROR) { 1529 /* 1530 * If an error occurs I'd love to throw the swapblk 1531 * away without freeing it back to swapspace, so it 1532 * can never be used again. But I can't from an 1533 * interrupt. 1534 */ 1535 if (bp->b_iocmd == BIO_READ) { 1536 /* 1537 * When reading, reqpage needs to stay 1538 * locked for the parent, but all other 1539 * pages can be freed. We still want to 1540 * wakeup the parent waiting on the page, 1541 * though. ( also: pg_reqpage can be -1 and 1542 * not match anything ). 1543 * 1544 * We have to wake specifically requested pages 1545 * up too because we cleared VPO_SWAPINPROG and 1546 * someone may be waiting for that. 1547 * 1548 * NOTE: for reads, m->dirty will probably 1549 * be overridden by the original caller of 1550 * getpages so don't play cute tricks here. 1551 */ 1552 m->valid = 0; 1553 if (i != bp->b_pager.pg_reqpage) 1554 swp_pager_free_nrpage(m); 1555 else { 1556 vm_page_lock(m); 1557 vm_page_flash(m); 1558 vm_page_unlock(m); 1559 } 1560 /* 1561 * If i == bp->b_pager.pg_reqpage, do not wake 1562 * the page up. The caller needs to. 1563 */ 1564 } else { 1565 /* 1566 * If a write error occurs, reactivate page 1567 * so it doesn't clog the inactive list, 1568 * then finish the I/O. 1569 */ 1570 vm_page_dirty(m); 1571 vm_page_lock(m); 1572 vm_page_activate(m); 1573 vm_page_unlock(m); 1574 vm_page_sunbusy(m); 1575 } 1576 } else if (bp->b_iocmd == BIO_READ) { 1577 /* 1578 * NOTE: for reads, m->dirty will probably be 1579 * overridden by the original caller of getpages so 1580 * we cannot set them in order to free the underlying 1581 * swap in a low-swap situation. I don't think we'd 1582 * want to do that anyway, but it was an optimization 1583 * that existed in the old swapper for a time before 1584 * it got ripped out due to precisely this problem. 1585 * 1586 * If not the requested page then deactivate it. 1587 * 1588 * Note that the requested page, reqpage, is left 1589 * busied, but we still have to wake it up. The 1590 * other pages are released (unbusied) by 1591 * vm_page_xunbusy(). 1592 */ 1593 KASSERT(!pmap_page_is_mapped(m), 1594 ("swp_pager_async_iodone: page %p is mapped", m)); 1595 m->valid = VM_PAGE_BITS_ALL; 1596 KASSERT(m->dirty == 0, 1597 ("swp_pager_async_iodone: page %p is dirty", m)); 1598 1599 /* 1600 * We have to wake specifically requested pages 1601 * up too because we cleared VPO_SWAPINPROG and 1602 * could be waiting for it in getpages. However, 1603 * be sure to not unbusy getpages specifically 1604 * requested page - getpages expects it to be 1605 * left busy. 1606 */ 1607 if (i != bp->b_pager.pg_reqpage) { 1608 vm_page_lock(m); 1609 vm_page_deactivate(m); 1610 vm_page_unlock(m); 1611 vm_page_xunbusy(m); 1612 } else { 1613 vm_page_lock(m); 1614 vm_page_flash(m); 1615 vm_page_unlock(m); 1616 } 1617 } else { 1618 /* 1619 * For write success, clear the dirty 1620 * status, then finish the I/O ( which decrements the 1621 * busy count and possibly wakes waiter's up ). 1622 */ 1623 KASSERT(!pmap_page_is_write_mapped(m), 1624 ("swp_pager_async_iodone: page %p is not write" 1625 " protected", m)); 1626 vm_page_undirty(m); 1627 vm_page_sunbusy(m); 1628 if (vm_page_count_severe()) { 1629 vm_page_lock(m); 1630 vm_page_try_to_cache(m); 1631 vm_page_unlock(m); 1632 } 1633 } 1634 } 1635 1636 /* 1637 * adjust pip. NOTE: the original parent may still have its own 1638 * pip refs on the object. 1639 */ 1640 if (object != NULL) { 1641 vm_object_pip_wakeupn(object, bp->b_npages); 1642 VM_OBJECT_WUNLOCK(object); 1643 } 1644 1645 /* 1646 * swapdev_strategy() manually sets b_vp and b_bufobj before calling 1647 * bstrategy(). Set them back to NULL now we're done with it, or we'll 1648 * trigger a KASSERT in relpbuf(). 1649 */ 1650 if (bp->b_vp) { 1651 bp->b_vp = NULL; 1652 bp->b_bufobj = NULL; 1653 } 1654 /* 1655 * release the physical I/O buffer 1656 */ 1657 relpbuf( 1658 bp, 1659 ((bp->b_iocmd == BIO_READ) ? &nsw_rcount : 1660 ((bp->b_flags & B_ASYNC) ? 1661 &nsw_wcount_async : 1662 &nsw_wcount_sync 1663 ) 1664 ) 1665 ); 1666 } 1667 1668 /* 1669 * swap_pager_isswapped: 1670 * 1671 * Return 1 if at least one page in the given object is paged 1672 * out to the given swap device. 1673 * 1674 * This routine may not sleep. 1675 */ 1676 int 1677 swap_pager_isswapped(vm_object_t object, struct swdevt *sp) 1678 { 1679 daddr_t index = 0; 1680 int bcount; 1681 int i; 1682 1683 VM_OBJECT_ASSERT_WLOCKED(object); 1684 if (object->type != OBJT_SWAP) 1685 return (0); 1686 1687 mtx_lock(&swhash_mtx); 1688 for (bcount = 0; bcount < object->un_pager.swp.swp_bcount; bcount++) { 1689 struct swblock *swap; 1690 1691 if ((swap = *swp_pager_hash(object, index)) != NULL) { 1692 for (i = 0; i < SWAP_META_PAGES; ++i) { 1693 if (swp_pager_isondev(swap->swb_pages[i], sp)) { 1694 mtx_unlock(&swhash_mtx); 1695 return (1); 1696 } 1697 } 1698 } 1699 index += SWAP_META_PAGES; 1700 } 1701 mtx_unlock(&swhash_mtx); 1702 return (0); 1703 } 1704 1705 /* 1706 * SWP_PAGER_FORCE_PAGEIN() - force a swap block to be paged in 1707 * 1708 * This routine dissociates the page at the given index within a 1709 * swap block from its backing store, paging it in if necessary. 1710 * If the page is paged in, it is placed in the inactive queue, 1711 * since it had its backing store ripped out from under it. 1712 * We also attempt to swap in all other pages in the swap block, 1713 * we only guarantee that the one at the specified index is 1714 * paged in. 1715 * 1716 * XXX - The code to page the whole block in doesn't work, so we 1717 * revert to the one-by-one behavior for now. Sigh. 1718 */ 1719 static inline void 1720 swp_pager_force_pagein(vm_object_t object, vm_pindex_t pindex) 1721 { 1722 vm_page_t m; 1723 1724 vm_object_pip_add(object, 1); 1725 m = vm_page_grab(object, pindex, VM_ALLOC_NORMAL); 1726 if (m->valid == VM_PAGE_BITS_ALL) { 1727 vm_object_pip_wakeup(object); 1728 vm_page_dirty(m); 1729 vm_page_lock(m); 1730 vm_page_activate(m); 1731 vm_page_unlock(m); 1732 vm_page_xunbusy(m); 1733 vm_pager_page_unswapped(m); 1734 return; 1735 } 1736 1737 if (swap_pager_getpages(object, &m, 1, 0) != VM_PAGER_OK) 1738 panic("swap_pager_force_pagein: read from swap failed");/*XXX*/ 1739 vm_object_pip_wakeup(object); 1740 vm_page_dirty(m); 1741 vm_page_lock(m); 1742 vm_page_deactivate(m); 1743 vm_page_unlock(m); 1744 vm_page_xunbusy(m); 1745 vm_pager_page_unswapped(m); 1746 } 1747 1748 /* 1749 * swap_pager_swapoff: 1750 * 1751 * Page in all of the pages that have been paged out to the 1752 * given device. The corresponding blocks in the bitmap must be 1753 * marked as allocated and the device must be flagged SW_CLOSING. 1754 * There may be no processes swapped out to the device. 1755 * 1756 * This routine may block. 1757 */ 1758 static void 1759 swap_pager_swapoff(struct swdevt *sp) 1760 { 1761 struct swblock *swap; 1762 int i, j, retries; 1763 1764 GIANT_REQUIRED; 1765 1766 retries = 0; 1767 full_rescan: 1768 mtx_lock(&swhash_mtx); 1769 for (i = 0; i <= swhash_mask; i++) { /* '<=' is correct here */ 1770 restart: 1771 for (swap = swhash[i]; swap != NULL; swap = swap->swb_hnext) { 1772 vm_object_t object = swap->swb_object; 1773 vm_pindex_t pindex = swap->swb_index; 1774 for (j = 0; j < SWAP_META_PAGES; ++j) { 1775 if (swp_pager_isondev(swap->swb_pages[j], sp)) { 1776 /* avoid deadlock */ 1777 if (!VM_OBJECT_TRYWLOCK(object)) { 1778 break; 1779 } else { 1780 mtx_unlock(&swhash_mtx); 1781 swp_pager_force_pagein(object, 1782 pindex + j); 1783 VM_OBJECT_WUNLOCK(object); 1784 mtx_lock(&swhash_mtx); 1785 goto restart; 1786 } 1787 } 1788 } 1789 } 1790 } 1791 mtx_unlock(&swhash_mtx); 1792 if (sp->sw_used) { 1793 /* 1794 * Objects may be locked or paging to the device being 1795 * removed, so we will miss their pages and need to 1796 * make another pass. We have marked this device as 1797 * SW_CLOSING, so the activity should finish soon. 1798 */ 1799 retries++; 1800 if (retries > 100) { 1801 panic("swapoff: failed to locate %d swap blocks", 1802 sp->sw_used); 1803 } 1804 pause("swpoff", hz / 20); 1805 goto full_rescan; 1806 } 1807 } 1808 1809 /************************************************************************ 1810 * SWAP META DATA * 1811 ************************************************************************ 1812 * 1813 * These routines manipulate the swap metadata stored in the 1814 * OBJT_SWAP object. 1815 * 1816 * Swap metadata is implemented with a global hash and not directly 1817 * linked into the object. Instead the object simply contains 1818 * appropriate tracking counters. 1819 */ 1820 1821 /* 1822 * SWP_PAGER_META_BUILD() - add swap block to swap meta data for object 1823 * 1824 * We first convert the object to a swap object if it is a default 1825 * object. 1826 * 1827 * The specified swapblk is added to the object's swap metadata. If 1828 * the swapblk is not valid, it is freed instead. Any previously 1829 * assigned swapblk is freed. 1830 */ 1831 static void 1832 swp_pager_meta_build(vm_object_t object, vm_pindex_t pindex, daddr_t swapblk) 1833 { 1834 static volatile int exhausted; 1835 struct swblock *swap; 1836 struct swblock **pswap; 1837 int idx; 1838 1839 VM_OBJECT_ASSERT_WLOCKED(object); 1840 /* 1841 * Convert default object to swap object if necessary 1842 */ 1843 if (object->type != OBJT_SWAP) { 1844 object->type = OBJT_SWAP; 1845 object->un_pager.swp.swp_bcount = 0; 1846 1847 if (object->handle != NULL) { 1848 mtx_lock(&sw_alloc_mtx); 1849 TAILQ_INSERT_TAIL( 1850 NOBJLIST(object->handle), 1851 object, 1852 pager_object_list 1853 ); 1854 mtx_unlock(&sw_alloc_mtx); 1855 } 1856 } 1857 1858 /* 1859 * Locate hash entry. If not found create, but if we aren't adding 1860 * anything just return. If we run out of space in the map we wait 1861 * and, since the hash table may have changed, retry. 1862 */ 1863 retry: 1864 mtx_lock(&swhash_mtx); 1865 pswap = swp_pager_hash(object, pindex); 1866 1867 if ((swap = *pswap) == NULL) { 1868 int i; 1869 1870 if (swapblk == SWAPBLK_NONE) 1871 goto done; 1872 1873 swap = *pswap = uma_zalloc(swap_zone, M_NOWAIT | 1874 (curproc == pageproc ? M_USE_RESERVE : 0)); 1875 if (swap == NULL) { 1876 mtx_unlock(&swhash_mtx); 1877 VM_OBJECT_WUNLOCK(object); 1878 if (uma_zone_exhausted(swap_zone)) { 1879 if (atomic_cmpset_int(&exhausted, 0, 1)) 1880 printf("swap zone exhausted, " 1881 "increase kern.maxswzone\n"); 1882 vm_pageout_oom(VM_OOM_SWAPZ); 1883 pause("swzonex", 10); 1884 } else 1885 VM_WAIT; 1886 VM_OBJECT_WLOCK(object); 1887 goto retry; 1888 } 1889 1890 if (atomic_cmpset_int(&exhausted, 1, 0)) 1891 printf("swap zone ok\n"); 1892 1893 swap->swb_hnext = NULL; 1894 swap->swb_object = object; 1895 swap->swb_index = pindex & ~(vm_pindex_t)SWAP_META_MASK; 1896 swap->swb_count = 0; 1897 1898 ++object->un_pager.swp.swp_bcount; 1899 1900 for (i = 0; i < SWAP_META_PAGES; ++i) 1901 swap->swb_pages[i] = SWAPBLK_NONE; 1902 } 1903 1904 /* 1905 * Delete prior contents of metadata 1906 */ 1907 idx = pindex & SWAP_META_MASK; 1908 1909 if (swap->swb_pages[idx] != SWAPBLK_NONE) { 1910 swp_pager_freeswapspace(swap->swb_pages[idx], 1); 1911 --swap->swb_count; 1912 } 1913 1914 /* 1915 * Enter block into metadata 1916 */ 1917 swap->swb_pages[idx] = swapblk; 1918 if (swapblk != SWAPBLK_NONE) 1919 ++swap->swb_count; 1920 done: 1921 mtx_unlock(&swhash_mtx); 1922 } 1923 1924 /* 1925 * SWP_PAGER_META_FREE() - free a range of blocks in the object's swap metadata 1926 * 1927 * The requested range of blocks is freed, with any associated swap 1928 * returned to the swap bitmap. 1929 * 1930 * This routine will free swap metadata structures as they are cleaned 1931 * out. This routine does *NOT* operate on swap metadata associated 1932 * with resident pages. 1933 */ 1934 static void 1935 swp_pager_meta_free(vm_object_t object, vm_pindex_t index, daddr_t count) 1936 { 1937 1938 VM_OBJECT_ASSERT_LOCKED(object); 1939 if (object->type != OBJT_SWAP) 1940 return; 1941 1942 while (count > 0) { 1943 struct swblock **pswap; 1944 struct swblock *swap; 1945 1946 mtx_lock(&swhash_mtx); 1947 pswap = swp_pager_hash(object, index); 1948 1949 if ((swap = *pswap) != NULL) { 1950 daddr_t v = swap->swb_pages[index & SWAP_META_MASK]; 1951 1952 if (v != SWAPBLK_NONE) { 1953 swp_pager_freeswapspace(v, 1); 1954 swap->swb_pages[index & SWAP_META_MASK] = 1955 SWAPBLK_NONE; 1956 if (--swap->swb_count == 0) { 1957 *pswap = swap->swb_hnext; 1958 uma_zfree(swap_zone, swap); 1959 --object->un_pager.swp.swp_bcount; 1960 } 1961 } 1962 --count; 1963 ++index; 1964 } else { 1965 int n = SWAP_META_PAGES - (index & SWAP_META_MASK); 1966 count -= n; 1967 index += n; 1968 } 1969 mtx_unlock(&swhash_mtx); 1970 } 1971 } 1972 1973 /* 1974 * SWP_PAGER_META_FREE_ALL() - destroy all swap metadata associated with object 1975 * 1976 * This routine locates and destroys all swap metadata associated with 1977 * an object. 1978 */ 1979 static void 1980 swp_pager_meta_free_all(vm_object_t object) 1981 { 1982 daddr_t index = 0; 1983 1984 VM_OBJECT_ASSERT_WLOCKED(object); 1985 if (object->type != OBJT_SWAP) 1986 return; 1987 1988 while (object->un_pager.swp.swp_bcount) { 1989 struct swblock **pswap; 1990 struct swblock *swap; 1991 1992 mtx_lock(&swhash_mtx); 1993 pswap = swp_pager_hash(object, index); 1994 if ((swap = *pswap) != NULL) { 1995 int i; 1996 1997 for (i = 0; i < SWAP_META_PAGES; ++i) { 1998 daddr_t v = swap->swb_pages[i]; 1999 if (v != SWAPBLK_NONE) { 2000 --swap->swb_count; 2001 swp_pager_freeswapspace(v, 1); 2002 } 2003 } 2004 if (swap->swb_count != 0) 2005 panic("swap_pager_meta_free_all: swb_count != 0"); 2006 *pswap = swap->swb_hnext; 2007 uma_zfree(swap_zone, swap); 2008 --object->un_pager.swp.swp_bcount; 2009 } 2010 mtx_unlock(&swhash_mtx); 2011 index += SWAP_META_PAGES; 2012 } 2013 } 2014 2015 /* 2016 * SWP_PAGER_METACTL() - misc control of swap and vm_page_t meta data. 2017 * 2018 * This routine is capable of looking up, popping, or freeing 2019 * swapblk assignments in the swap meta data or in the vm_page_t. 2020 * The routine typically returns the swapblk being looked-up, or popped, 2021 * or SWAPBLK_NONE if the block was freed, or SWAPBLK_NONE if the block 2022 * was invalid. This routine will automatically free any invalid 2023 * meta-data swapblks. 2024 * 2025 * It is not possible to store invalid swapblks in the swap meta data 2026 * (other then a literal 'SWAPBLK_NONE'), so we don't bother checking. 2027 * 2028 * When acting on a busy resident page and paging is in progress, we 2029 * have to wait until paging is complete but otherwise can act on the 2030 * busy page. 2031 * 2032 * SWM_FREE remove and free swap block from metadata 2033 * SWM_POP remove from meta data but do not free.. pop it out 2034 */ 2035 static daddr_t 2036 swp_pager_meta_ctl(vm_object_t object, vm_pindex_t pindex, int flags) 2037 { 2038 struct swblock **pswap; 2039 struct swblock *swap; 2040 daddr_t r1; 2041 int idx; 2042 2043 VM_OBJECT_ASSERT_LOCKED(object); 2044 /* 2045 * The meta data only exists of the object is OBJT_SWAP 2046 * and even then might not be allocated yet. 2047 */ 2048 if (object->type != OBJT_SWAP) 2049 return (SWAPBLK_NONE); 2050 2051 r1 = SWAPBLK_NONE; 2052 mtx_lock(&swhash_mtx); 2053 pswap = swp_pager_hash(object, pindex); 2054 2055 if ((swap = *pswap) != NULL) { 2056 idx = pindex & SWAP_META_MASK; 2057 r1 = swap->swb_pages[idx]; 2058 2059 if (r1 != SWAPBLK_NONE) { 2060 if (flags & SWM_FREE) { 2061 swp_pager_freeswapspace(r1, 1); 2062 r1 = SWAPBLK_NONE; 2063 } 2064 if (flags & (SWM_FREE|SWM_POP)) { 2065 swap->swb_pages[idx] = SWAPBLK_NONE; 2066 if (--swap->swb_count == 0) { 2067 *pswap = swap->swb_hnext; 2068 uma_zfree(swap_zone, swap); 2069 --object->un_pager.swp.swp_bcount; 2070 } 2071 } 2072 } 2073 } 2074 mtx_unlock(&swhash_mtx); 2075 return (r1); 2076 } 2077 2078 /* 2079 * System call swapon(name) enables swapping on device name, 2080 * which must be in the swdevsw. Return EBUSY 2081 * if already swapping on this device. 2082 */ 2083 #ifndef _SYS_SYSPROTO_H_ 2084 struct swapon_args { 2085 char *name; 2086 }; 2087 #endif 2088 2089 /* 2090 * MPSAFE 2091 */ 2092 /* ARGSUSED */ 2093 int 2094 sys_swapon(struct thread *td, struct swapon_args *uap) 2095 { 2096 struct vattr attr; 2097 struct vnode *vp; 2098 struct nameidata nd; 2099 int error; 2100 2101 error = priv_check(td, PRIV_SWAPON); 2102 if (error) 2103 return (error); 2104 2105 mtx_lock(&Giant); 2106 while (swdev_syscall_active) 2107 tsleep(&swdev_syscall_active, PUSER - 1, "swpon", 0); 2108 swdev_syscall_active = 1; 2109 2110 /* 2111 * Swap metadata may not fit in the KVM if we have physical 2112 * memory of >1GB. 2113 */ 2114 if (swap_zone == NULL) { 2115 error = ENOMEM; 2116 goto done; 2117 } 2118 2119 NDINIT(&nd, LOOKUP, ISOPEN | FOLLOW | AUDITVNODE1, UIO_USERSPACE, 2120 uap->name, td); 2121 error = namei(&nd); 2122 if (error) 2123 goto done; 2124 2125 NDFREE(&nd, NDF_ONLY_PNBUF); 2126 vp = nd.ni_vp; 2127 2128 if (vn_isdisk(vp, &error)) { 2129 error = swapongeom(td, vp); 2130 } else if (vp->v_type == VREG && 2131 (vp->v_mount->mnt_vfc->vfc_flags & VFCF_NETWORK) != 0 && 2132 (error = VOP_GETATTR(vp, &attr, td->td_ucred)) == 0) { 2133 /* 2134 * Allow direct swapping to NFS regular files in the same 2135 * way that nfs_mountroot() sets up diskless swapping. 2136 */ 2137 error = swaponvp(td, vp, attr.va_size / DEV_BSIZE); 2138 } 2139 2140 if (error) 2141 vrele(vp); 2142 done: 2143 swdev_syscall_active = 0; 2144 wakeup_one(&swdev_syscall_active); 2145 mtx_unlock(&Giant); 2146 return (error); 2147 } 2148 2149 /* 2150 * Check that the total amount of swap currently configured does not 2151 * exceed half the theoretical maximum. If it does, print a warning 2152 * message and return -1; otherwise, return 0. 2153 */ 2154 static int 2155 swapon_check_swzone(unsigned long npages) 2156 { 2157 unsigned long maxpages; 2158 2159 /* absolute maximum we can handle assuming 100% efficiency */ 2160 maxpages = uma_zone_get_max(swap_zone) * SWAP_META_PAGES; 2161 2162 /* recommend using no more than half that amount */ 2163 if (npages > maxpages / 2) { 2164 printf("warning: total configured swap (%lu pages) " 2165 "exceeds maximum recommended amount (%lu pages).\n", 2166 npages, maxpages / 2); 2167 printf("warning: increase kern.maxswzone " 2168 "or reduce amount of swap.\n"); 2169 return (-1); 2170 } 2171 return (0); 2172 } 2173 2174 static void 2175 swaponsomething(struct vnode *vp, void *id, u_long nblks, 2176 sw_strategy_t *strategy, sw_close_t *close, dev_t dev, int flags) 2177 { 2178 struct swdevt *sp, *tsp; 2179 swblk_t dvbase; 2180 u_long mblocks; 2181 2182 /* 2183 * nblks is in DEV_BSIZE'd chunks, convert to PAGE_SIZE'd chunks. 2184 * First chop nblks off to page-align it, then convert. 2185 * 2186 * sw->sw_nblks is in page-sized chunks now too. 2187 */ 2188 nblks &= ~(ctodb(1) - 1); 2189 nblks = dbtoc(nblks); 2190 2191 /* 2192 * If we go beyond this, we get overflows in the radix 2193 * tree bitmap code. 2194 */ 2195 mblocks = 0x40000000 / BLIST_META_RADIX; 2196 if (nblks > mblocks) { 2197 printf( 2198 "WARNING: reducing swap size to maximum of %luMB per unit\n", 2199 mblocks / 1024 / 1024 * PAGE_SIZE); 2200 nblks = mblocks; 2201 } 2202 2203 sp = malloc(sizeof *sp, M_VMPGDATA, M_WAITOK | M_ZERO); 2204 sp->sw_vp = vp; 2205 sp->sw_id = id; 2206 sp->sw_dev = dev; 2207 sp->sw_flags = 0; 2208 sp->sw_nblks = nblks; 2209 sp->sw_used = 0; 2210 sp->sw_strategy = strategy; 2211 sp->sw_close = close; 2212 sp->sw_flags = flags; 2213 2214 sp->sw_blist = blist_create(nblks, M_WAITOK); 2215 /* 2216 * Do not free the first two block in order to avoid overwriting 2217 * any bsd label at the front of the partition 2218 */ 2219 blist_free(sp->sw_blist, 2, nblks - 2); 2220 2221 dvbase = 0; 2222 mtx_lock(&sw_dev_mtx); 2223 TAILQ_FOREACH(tsp, &swtailq, sw_list) { 2224 if (tsp->sw_end >= dvbase) { 2225 /* 2226 * We put one uncovered page between the devices 2227 * in order to definitively prevent any cross-device 2228 * I/O requests 2229 */ 2230 dvbase = tsp->sw_end + 1; 2231 } 2232 } 2233 sp->sw_first = dvbase; 2234 sp->sw_end = dvbase + nblks; 2235 TAILQ_INSERT_TAIL(&swtailq, sp, sw_list); 2236 nswapdev++; 2237 swap_pager_avail += nblks; 2238 swap_total += (vm_ooffset_t)nblks * PAGE_SIZE; 2239 swapon_check_swzone(swap_total / PAGE_SIZE); 2240 swp_sizecheck(); 2241 mtx_unlock(&sw_dev_mtx); 2242 } 2243 2244 /* 2245 * SYSCALL: swapoff(devname) 2246 * 2247 * Disable swapping on the given device. 2248 * 2249 * XXX: Badly designed system call: it should use a device index 2250 * rather than filename as specification. We keep sw_vp around 2251 * only to make this work. 2252 */ 2253 #ifndef _SYS_SYSPROTO_H_ 2254 struct swapoff_args { 2255 char *name; 2256 }; 2257 #endif 2258 2259 /* 2260 * MPSAFE 2261 */ 2262 /* ARGSUSED */ 2263 int 2264 sys_swapoff(struct thread *td, struct swapoff_args *uap) 2265 { 2266 struct vnode *vp; 2267 struct nameidata nd; 2268 struct swdevt *sp; 2269 int error; 2270 2271 error = priv_check(td, PRIV_SWAPOFF); 2272 if (error) 2273 return (error); 2274 2275 mtx_lock(&Giant); 2276 while (swdev_syscall_active) 2277 tsleep(&swdev_syscall_active, PUSER - 1, "swpoff", 0); 2278 swdev_syscall_active = 1; 2279 2280 NDINIT(&nd, LOOKUP, FOLLOW | AUDITVNODE1, UIO_USERSPACE, uap->name, 2281 td); 2282 error = namei(&nd); 2283 if (error) 2284 goto done; 2285 NDFREE(&nd, NDF_ONLY_PNBUF); 2286 vp = nd.ni_vp; 2287 2288 mtx_lock(&sw_dev_mtx); 2289 TAILQ_FOREACH(sp, &swtailq, sw_list) { 2290 if (sp->sw_vp == vp) 2291 break; 2292 } 2293 mtx_unlock(&sw_dev_mtx); 2294 if (sp == NULL) { 2295 error = EINVAL; 2296 goto done; 2297 } 2298 error = swapoff_one(sp, td->td_ucred); 2299 done: 2300 swdev_syscall_active = 0; 2301 wakeup_one(&swdev_syscall_active); 2302 mtx_unlock(&Giant); 2303 return (error); 2304 } 2305 2306 static int 2307 swapoff_one(struct swdevt *sp, struct ucred *cred) 2308 { 2309 u_long nblks, dvbase; 2310 #ifdef MAC 2311 int error; 2312 #endif 2313 2314 mtx_assert(&Giant, MA_OWNED); 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_cnt.v_free_count + vm_cnt.v_cache_count + swap_pager_avail < 2331 nblks + nswap_lowat) { 2332 return (ENOMEM); 2333 } 2334 2335 /* 2336 * Prevent further allocations on this device. 2337 */ 2338 mtx_lock(&sw_dev_mtx); 2339 sp->sw_flags |= SW_CLOSING; 2340 for (dvbase = 0; dvbase < sp->sw_end; dvbase += dmmax) { 2341 swap_pager_avail -= blist_fill(sp->sw_blist, 2342 dvbase, dmmax); 2343 } 2344 swap_total -= (vm_ooffset_t)nblks * PAGE_SIZE; 2345 mtx_unlock(&sw_dev_mtx); 2346 2347 /* 2348 * Page in the contents of the device and close it. 2349 */ 2350 swap_pager_swapoff(sp); 2351 2352 sp->sw_close(curthread, sp); 2353 sp->sw_id = NULL; 2354 mtx_lock(&sw_dev_mtx); 2355 TAILQ_REMOVE(&swtailq, sp, sw_list); 2356 nswapdev--; 2357 if (nswapdev == 0) { 2358 swap_pager_full = 2; 2359 swap_pager_almost_full = 1; 2360 } 2361 if (swdevhd == sp) 2362 swdevhd = NULL; 2363 mtx_unlock(&sw_dev_mtx); 2364 blist_destroy(sp->sw_blist); 2365 free(sp, M_VMPGDATA); 2366 return (0); 2367 } 2368 2369 void 2370 swapoff_all(void) 2371 { 2372 struct swdevt *sp, *spt; 2373 const char *devname; 2374 int error; 2375 2376 mtx_lock(&Giant); 2377 while (swdev_syscall_active) 2378 tsleep(&swdev_syscall_active, PUSER - 1, "swpoff", 0); 2379 swdev_syscall_active = 1; 2380 2381 mtx_lock(&sw_dev_mtx); 2382 TAILQ_FOREACH_SAFE(sp, &swtailq, sw_list, spt) { 2383 mtx_unlock(&sw_dev_mtx); 2384 if (vn_isdisk(sp->sw_vp, NULL)) 2385 devname = devtoname(sp->sw_vp->v_rdev); 2386 else 2387 devname = "[file]"; 2388 error = swapoff_one(sp, thread0.td_ucred); 2389 if (error != 0) { 2390 printf("Cannot remove swap device %s (error=%d), " 2391 "skipping.\n", devname, error); 2392 } else if (bootverbose) { 2393 printf("Swap device %s removed.\n", devname); 2394 } 2395 mtx_lock(&sw_dev_mtx); 2396 } 2397 mtx_unlock(&sw_dev_mtx); 2398 2399 swdev_syscall_active = 0; 2400 wakeup_one(&swdev_syscall_active); 2401 mtx_unlock(&Giant); 2402 } 2403 2404 void 2405 swap_pager_status(int *total, int *used) 2406 { 2407 struct swdevt *sp; 2408 2409 *total = 0; 2410 *used = 0; 2411 mtx_lock(&sw_dev_mtx); 2412 TAILQ_FOREACH(sp, &swtailq, sw_list) { 2413 *total += sp->sw_nblks; 2414 *used += sp->sw_used; 2415 } 2416 mtx_unlock(&sw_dev_mtx); 2417 } 2418 2419 int 2420 swap_dev_info(int name, struct xswdev *xs, char *devname, size_t len) 2421 { 2422 struct swdevt *sp; 2423 const char *tmp_devname; 2424 int error, n; 2425 2426 n = 0; 2427 error = ENOENT; 2428 mtx_lock(&sw_dev_mtx); 2429 TAILQ_FOREACH(sp, &swtailq, sw_list) { 2430 if (n != name) { 2431 n++; 2432 continue; 2433 } 2434 xs->xsw_version = XSWDEV_VERSION; 2435 xs->xsw_dev = sp->sw_dev; 2436 xs->xsw_flags = sp->sw_flags; 2437 xs->xsw_nblks = sp->sw_nblks; 2438 xs->xsw_used = sp->sw_used; 2439 if (devname != NULL) { 2440 if (vn_isdisk(sp->sw_vp, NULL)) 2441 tmp_devname = devtoname(sp->sw_vp->v_rdev); 2442 else 2443 tmp_devname = "[file]"; 2444 strncpy(devname, tmp_devname, len); 2445 } 2446 error = 0; 2447 break; 2448 } 2449 mtx_unlock(&sw_dev_mtx); 2450 return (error); 2451 } 2452 2453 static int 2454 sysctl_vm_swap_info(SYSCTL_HANDLER_ARGS) 2455 { 2456 struct xswdev xs; 2457 int error; 2458 2459 if (arg2 != 1) /* name length */ 2460 return (EINVAL); 2461 error = swap_dev_info(*(int *)arg1, &xs, NULL, 0); 2462 if (error != 0) 2463 return (error); 2464 error = SYSCTL_OUT(req, &xs, sizeof(xs)); 2465 return (error); 2466 } 2467 2468 SYSCTL_INT(_vm, OID_AUTO, nswapdev, CTLFLAG_RD, &nswapdev, 0, 2469 "Number of swap devices"); 2470 SYSCTL_NODE(_vm, OID_AUTO, swap_info, CTLFLAG_RD, sysctl_vm_swap_info, 2471 "Swap statistics by device"); 2472 2473 /* 2474 * vmspace_swap_count() - count the approximate swap usage in pages for a 2475 * vmspace. 2476 * 2477 * The map must be locked. 2478 * 2479 * Swap usage is determined by taking the proportional swap used by 2480 * VM objects backing the VM map. To make up for fractional losses, 2481 * if the VM object has any swap use at all the associated map entries 2482 * count for at least 1 swap page. 2483 */ 2484 long 2485 vmspace_swap_count(struct vmspace *vmspace) 2486 { 2487 vm_map_t map; 2488 vm_map_entry_t cur; 2489 vm_object_t object; 2490 long count, n; 2491 2492 map = &vmspace->vm_map; 2493 count = 0; 2494 2495 for (cur = map->header.next; cur != &map->header; cur = cur->next) { 2496 if ((cur->eflags & MAP_ENTRY_IS_SUB_MAP) == 0 && 2497 (object = cur->object.vm_object) != NULL) { 2498 VM_OBJECT_WLOCK(object); 2499 if (object->type == OBJT_SWAP && 2500 object->un_pager.swp.swp_bcount != 0) { 2501 n = (cur->end - cur->start) / PAGE_SIZE; 2502 count += object->un_pager.swp.swp_bcount * 2503 SWAP_META_PAGES * n / object->size + 1; 2504 } 2505 VM_OBJECT_WUNLOCK(object); 2506 } 2507 } 2508 return (count); 2509 } 2510 2511 /* 2512 * GEOM backend 2513 * 2514 * Swapping onto disk devices. 2515 * 2516 */ 2517 2518 static g_orphan_t swapgeom_orphan; 2519 2520 static struct g_class g_swap_class = { 2521 .name = "SWAP", 2522 .version = G_VERSION, 2523 .orphan = swapgeom_orphan, 2524 }; 2525 2526 DECLARE_GEOM_CLASS(g_swap_class, g_class); 2527 2528 2529 static void 2530 swapgeom_close_ev(void *arg, int flags) 2531 { 2532 struct g_consumer *cp; 2533 2534 cp = arg; 2535 g_access(cp, -1, -1, 0); 2536 g_detach(cp); 2537 g_destroy_consumer(cp); 2538 } 2539 2540 static void 2541 swapgeom_done(struct bio *bp2) 2542 { 2543 struct swdevt *sp; 2544 struct buf *bp; 2545 struct g_consumer *cp; 2546 2547 bp = bp2->bio_caller2; 2548 cp = bp2->bio_from; 2549 bp->b_ioflags = bp2->bio_flags; 2550 if (bp2->bio_error) 2551 bp->b_ioflags |= BIO_ERROR; 2552 bp->b_resid = bp->b_bcount - bp2->bio_completed; 2553 bp->b_error = bp2->bio_error; 2554 bufdone(bp); 2555 mtx_lock(&sw_dev_mtx); 2556 if ((--cp->index) == 0 && cp->private) { 2557 if (g_post_event(swapgeom_close_ev, cp, M_NOWAIT, NULL) == 0) { 2558 sp = bp2->bio_caller1; 2559 sp->sw_id = NULL; 2560 } 2561 } 2562 mtx_unlock(&sw_dev_mtx); 2563 g_destroy_bio(bp2); 2564 } 2565 2566 static void 2567 swapgeom_strategy(struct buf *bp, struct swdevt *sp) 2568 { 2569 struct bio *bio; 2570 struct g_consumer *cp; 2571 2572 mtx_lock(&sw_dev_mtx); 2573 cp = sp->sw_id; 2574 if (cp == NULL) { 2575 mtx_unlock(&sw_dev_mtx); 2576 bp->b_error = ENXIO; 2577 bp->b_ioflags |= BIO_ERROR; 2578 bufdone(bp); 2579 return; 2580 } 2581 cp->index++; 2582 mtx_unlock(&sw_dev_mtx); 2583 if (bp->b_iocmd == BIO_WRITE) 2584 bio = g_new_bio(); 2585 else 2586 bio = g_alloc_bio(); 2587 if (bio == NULL) { 2588 bp->b_error = ENOMEM; 2589 bp->b_ioflags |= BIO_ERROR; 2590 bufdone(bp); 2591 return; 2592 } 2593 2594 bio->bio_caller1 = sp; 2595 bio->bio_caller2 = bp; 2596 bio->bio_cmd = bp->b_iocmd; 2597 bio->bio_offset = (bp->b_blkno - sp->sw_first) * PAGE_SIZE; 2598 bio->bio_length = bp->b_bcount; 2599 bio->bio_done = swapgeom_done; 2600 if ((bp->b_flags & B_UNMAPPED) != 0) { 2601 bio->bio_ma = bp->b_pages; 2602 bio->bio_data = unmapped_buf; 2603 bio->bio_ma_offset = (vm_offset_t)bp->b_offset & PAGE_MASK; 2604 bio->bio_ma_n = bp->b_npages; 2605 bio->bio_flags |= BIO_UNMAPPED; 2606 } else { 2607 bio->bio_data = bp->b_data; 2608 bio->bio_ma = NULL; 2609 } 2610 g_io_request(bio, cp); 2611 return; 2612 } 2613 2614 static void 2615 swapgeom_orphan(struct g_consumer *cp) 2616 { 2617 struct swdevt *sp; 2618 int destroy; 2619 2620 mtx_lock(&sw_dev_mtx); 2621 TAILQ_FOREACH(sp, &swtailq, sw_list) { 2622 if (sp->sw_id == cp) { 2623 sp->sw_flags |= SW_CLOSING; 2624 break; 2625 } 2626 } 2627 cp->private = (void *)(uintptr_t)1; 2628 destroy = ((sp != NULL) && (cp->index == 0)); 2629 if (destroy) 2630 sp->sw_id = NULL; 2631 mtx_unlock(&sw_dev_mtx); 2632 if (destroy) 2633 swapgeom_close_ev(cp, 0); 2634 } 2635 2636 static void 2637 swapgeom_close(struct thread *td, struct swdevt *sw) 2638 { 2639 struct g_consumer *cp; 2640 2641 mtx_lock(&sw_dev_mtx); 2642 cp = sw->sw_id; 2643 sw->sw_id = NULL; 2644 mtx_unlock(&sw_dev_mtx); 2645 /* XXX: direct call when Giant untangled */ 2646 if (cp != NULL) 2647 g_waitfor_event(swapgeom_close_ev, cp, M_WAITOK, NULL); 2648 } 2649 2650 2651 struct swh0h0 { 2652 struct cdev *dev; 2653 struct vnode *vp; 2654 int error; 2655 }; 2656 2657 static void 2658 swapongeom_ev(void *arg, int flags) 2659 { 2660 struct swh0h0 *swh; 2661 struct g_provider *pp; 2662 struct g_consumer *cp; 2663 static struct g_geom *gp; 2664 struct swdevt *sp; 2665 u_long nblks; 2666 int error; 2667 2668 swh = arg; 2669 swh->error = 0; 2670 pp = g_dev_getprovider(swh->dev); 2671 if (pp == NULL) { 2672 swh->error = ENODEV; 2673 return; 2674 } 2675 mtx_lock(&sw_dev_mtx); 2676 TAILQ_FOREACH(sp, &swtailq, sw_list) { 2677 cp = sp->sw_id; 2678 if (cp != NULL && cp->provider == pp) { 2679 mtx_unlock(&sw_dev_mtx); 2680 swh->error = EBUSY; 2681 return; 2682 } 2683 } 2684 mtx_unlock(&sw_dev_mtx); 2685 if (gp == NULL) 2686 gp = g_new_geomf(&g_swap_class, "swap"); 2687 cp = g_new_consumer(gp); 2688 cp->index = 0; /* Number of active I/Os. */ 2689 cp->private = NULL; /* Orphanization flag */ 2690 g_attach(cp, pp); 2691 /* 2692 * XXX: Everytime you think you can improve the margin for 2693 * footshooting, somebody depends on the ability to do so: 2694 * savecore(8) wants to write to our swapdev so we cannot 2695 * set an exclusive count :-( 2696 */ 2697 error = g_access(cp, 1, 1, 0); 2698 if (error) { 2699 g_detach(cp); 2700 g_destroy_consumer(cp); 2701 swh->error = error; 2702 return; 2703 } 2704 nblks = pp->mediasize / DEV_BSIZE; 2705 swaponsomething(swh->vp, cp, nblks, swapgeom_strategy, 2706 swapgeom_close, dev2udev(swh->dev), 2707 (pp->flags & G_PF_ACCEPT_UNMAPPED) != 0 ? SW_UNMAPPED : 0); 2708 swh->error = 0; 2709 } 2710 2711 static int 2712 swapongeom(struct thread *td, struct vnode *vp) 2713 { 2714 int error; 2715 struct swh0h0 swh; 2716 2717 vn_lock(vp, LK_EXCLUSIVE | LK_RETRY); 2718 2719 swh.dev = vp->v_rdev; 2720 swh.vp = vp; 2721 swh.error = 0; 2722 /* XXX: direct call when Giant untangled */ 2723 error = g_waitfor_event(swapongeom_ev, &swh, M_WAITOK, NULL); 2724 if (!error) 2725 error = swh.error; 2726 VOP_UNLOCK(vp, 0); 2727 return (error); 2728 } 2729 2730 /* 2731 * VNODE backend 2732 * 2733 * This is used mainly for network filesystem (read: probably only tested 2734 * with NFS) swapfiles. 2735 * 2736 */ 2737 2738 static void 2739 swapdev_strategy(struct buf *bp, struct swdevt *sp) 2740 { 2741 struct vnode *vp2; 2742 2743 bp->b_blkno = ctodb(bp->b_blkno - sp->sw_first); 2744 2745 vp2 = sp->sw_id; 2746 vhold(vp2); 2747 if (bp->b_iocmd == BIO_WRITE) { 2748 if (bp->b_bufobj) 2749 bufobj_wdrop(bp->b_bufobj); 2750 bufobj_wref(&vp2->v_bufobj); 2751 } 2752 if (bp->b_bufobj != &vp2->v_bufobj) 2753 bp->b_bufobj = &vp2->v_bufobj; 2754 bp->b_vp = vp2; 2755 bp->b_iooffset = dbtob(bp->b_blkno); 2756 bstrategy(bp); 2757 return; 2758 } 2759 2760 static void 2761 swapdev_close(struct thread *td, struct swdevt *sp) 2762 { 2763 2764 VOP_CLOSE(sp->sw_vp, FREAD | FWRITE, td->td_ucred, td); 2765 vrele(sp->sw_vp); 2766 } 2767 2768 2769 static int 2770 swaponvp(struct thread *td, struct vnode *vp, u_long nblks) 2771 { 2772 struct swdevt *sp; 2773 int error; 2774 2775 if (nblks == 0) 2776 return (ENXIO); 2777 mtx_lock(&sw_dev_mtx); 2778 TAILQ_FOREACH(sp, &swtailq, sw_list) { 2779 if (sp->sw_id == vp) { 2780 mtx_unlock(&sw_dev_mtx); 2781 return (EBUSY); 2782 } 2783 } 2784 mtx_unlock(&sw_dev_mtx); 2785 2786 (void) vn_lock(vp, LK_EXCLUSIVE | LK_RETRY); 2787 #ifdef MAC 2788 error = mac_system_check_swapon(td->td_ucred, vp); 2789 if (error == 0) 2790 #endif 2791 error = VOP_OPEN(vp, FREAD | FWRITE, td->td_ucred, td, NULL); 2792 (void) VOP_UNLOCK(vp, 0); 2793 if (error) 2794 return (error); 2795 2796 swaponsomething(vp, vp, nblks, swapdev_strategy, swapdev_close, 2797 NODEV, 0); 2798 return (0); 2799 } 2800 2801 static int 2802 sysctl_swap_async_max(SYSCTL_HANDLER_ARGS) 2803 { 2804 int error, new, n; 2805 2806 new = nsw_wcount_async_max; 2807 error = sysctl_handle_int(oidp, &new, 0, req); 2808 if (error != 0 || req->newptr == NULL) 2809 return (error); 2810 2811 if (new > nswbuf / 2 || new < 1) 2812 return (EINVAL); 2813 2814 mtx_lock(&pbuf_mtx); 2815 while (nsw_wcount_async_max != new) { 2816 /* 2817 * Adjust difference. If the current async count is too low, 2818 * we will need to sqeeze our update slowly in. Sleep with a 2819 * higher priority than getpbuf() to finish faster. 2820 */ 2821 n = new - nsw_wcount_async_max; 2822 if (nsw_wcount_async + n >= 0) { 2823 nsw_wcount_async += n; 2824 nsw_wcount_async_max += n; 2825 wakeup(&nsw_wcount_async); 2826 } else { 2827 nsw_wcount_async_max -= nsw_wcount_async; 2828 nsw_wcount_async = 0; 2829 msleep(&nsw_wcount_async, &pbuf_mtx, PSWP, 2830 "swpsysctl", 0); 2831 } 2832 } 2833 mtx_unlock(&pbuf_mtx); 2834 2835 return (0); 2836 } 2837