swap_pager.c revision 284100
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: stable/10/sys/vm/swap_pager.c 284100 2015-06-06 20:37:40Z jhb $"); 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 141struct 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 149static MALLOC_DEFINE(M_VMPGDATA, "vm_pgdata", "swap pager private data"); 150static struct mtx sw_dev_mtx; 151static TAILQ_HEAD(, swdevt) swtailq = TAILQ_HEAD_INITIALIZER(swtailq); 152static struct swdevt *swdevhd; /* Allocate from here next */ 153static int nswapdev; /* Number of swap devices */ 154int swap_pager_avail; 155static int swdev_syscall_active = 0; /* serialize swap(on|off) */ 156 157static vm_ooffset_t swap_total; 158SYSCTL_QUAD(_vm, OID_AUTO, swap_total, CTLFLAG_RD, &swap_total, 0, 159 "Total amount of available swap storage."); 160static vm_ooffset_t swap_reserved; 161SYSCTL_QUAD(_vm, OID_AUTO, swap_reserved, CTLFLAG_RD, &swap_reserved, 0, 162 "Amount of swap storage needed to back all allocated anonymous memory."); 163static int overcommit = 0; 164SYSCTL_INT(_vm, OID_AUTO, overcommit, CTLFLAG_RW, &overcommit, 0, 165 "Configure virtual memory overcommit behavior. See tuning(7) " 166 "for details."); 167static unsigned long swzone; 168SYSCTL_ULONG(_vm, OID_AUTO, swzone, CTLFLAG_RD, &swzone, 0, 169 "Actual size of swap metadata zone"); 170static unsigned long swap_maxpages; 171SYSCTL_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 179int 180swap_reserve(vm_ooffset_t incr) 181{ 182 183 return (swap_reserve_by_cred(incr, curthread->td_ucred)); 184} 185 186int 187swap_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 PROC_LOCK(curproc); 202 error = racct_add(curproc, RACCT_SWAP, incr); 203 PROC_UNLOCK(curproc); 204 if (error != 0) 205 return (0); 206#endif 207 208 res = 0; 209 mtx_lock(&sw_dev_mtx); 210 r = swap_reserved + incr; 211 if (overcommit & SWAP_RESERVE_ALLOW_NONWIRED) { 212 s = cnt.v_page_count - cnt.v_free_reserved - cnt.v_wire_count; 213 s *= PAGE_SIZE; 214 } else 215 s = 0; 216 s += swap_total; 217 if ((overcommit & SWAP_RESERVE_FORCE_ON) == 0 || r <= s || 218 (error = priv_check(curthread, PRIV_VM_SWAP_NOQUOTA)) == 0) { 219 res = 1; 220 swap_reserved = r; 221 } 222 mtx_unlock(&sw_dev_mtx); 223 224 if (res) { 225 PROC_LOCK(curproc); 226 UIDINFO_VMSIZE_LOCK(uip); 227 if ((overcommit & SWAP_RESERVE_RLIMIT_ON) != 0 && 228 uip->ui_vmsize + incr > lim_cur(curproc, RLIMIT_SWAP) && 229 priv_check(curthread, PRIV_VM_SWAP_NORLIMIT)) 230 res = 0; 231 else 232 uip->ui_vmsize += incr; 233 UIDINFO_VMSIZE_UNLOCK(uip); 234 PROC_UNLOCK(curproc); 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 257void 258swap_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 280void 281swap_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 291void 292swap_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 316static void swapdev_strategy(struct buf *, struct swdevt *sw); 317 318#define SWM_FREE 0x02 /* free, period */ 319#define SWM_POP 0x04 /* pop out */ 320 321int swap_pager_full = 2; /* swap space exhaustion (task killing) */ 322static int swap_pager_almost_full = 1; /* swap space exhaustion (w/hysteresis)*/ 323static int nsw_rcount; /* free read buffers */ 324static int nsw_wcount_sync; /* limit write buffers / synchronous */ 325static int nsw_wcount_async; /* limit write buffers / asynchronous */ 326static int nsw_wcount_async_max;/* assigned maximum */ 327static int nsw_cluster_max; /* maximum VOP I/O allowed */ 328 329static struct swblock **swhash; 330static int swhash_mask; 331static struct mtx swhash_mtx; 332 333static int swap_async_max = 4; /* maximum in-progress async I/O's */ 334static struct sx sw_alloc_sx; 335 336 337SYSCTL_INT(_vm, OID_AUTO, swap_async_max, 338 CTLFLAG_RW, &swap_async_max, 0, "Maximum running async swap ops"); 339 340/* 341 * "named" and "unnamed" anon region objects. Try to reduce the overhead 342 * of searching a named list by hashing it just a little. 343 */ 344 345#define NOBJLISTS 8 346 347#define NOBJLIST(handle) \ 348 (&swap_pager_object_list[((int)(intptr_t)handle >> 4) & (NOBJLISTS-1)]) 349 350static struct mtx sw_alloc_mtx; /* protect list manipulation */ 351static struct pagerlst swap_pager_object_list[NOBJLISTS]; 352static uma_zone_t swap_zone; 353 354/* 355 * pagerops for OBJT_SWAP - "swap pager". Some ops are also global procedure 356 * calls hooked from other parts of the VM system and do not appear here. 357 * (see vm/swap_pager.h). 358 */ 359static vm_object_t 360 swap_pager_alloc(void *handle, vm_ooffset_t size, 361 vm_prot_t prot, vm_ooffset_t offset, struct ucred *); 362static void swap_pager_dealloc(vm_object_t object); 363static int swap_pager_getpages(vm_object_t, vm_page_t *, int, int); 364static void swap_pager_putpages(vm_object_t, vm_page_t *, int, boolean_t, int *); 365static boolean_t 366 swap_pager_haspage(vm_object_t object, vm_pindex_t pindex, int *before, int *after); 367static void swap_pager_init(void); 368static void swap_pager_unswapped(vm_page_t); 369static void swap_pager_swapoff(struct swdevt *sp); 370 371struct pagerops swappagerops = { 372 .pgo_init = swap_pager_init, /* early system initialization of pager */ 373 .pgo_alloc = swap_pager_alloc, /* allocate an OBJT_SWAP object */ 374 .pgo_dealloc = swap_pager_dealloc, /* deallocate an OBJT_SWAP object */ 375 .pgo_getpages = swap_pager_getpages, /* pagein */ 376 .pgo_putpages = swap_pager_putpages, /* pageout */ 377 .pgo_haspage = swap_pager_haspage, /* get backing store status for page */ 378 .pgo_pageunswapped = swap_pager_unswapped, /* remove swap related to page */ 379}; 380 381/* 382 * dmmax is in page-sized chunks with the new swap system. It was 383 * dev-bsized chunks in the old. dmmax is always a power of 2. 384 * 385 * swap_*() routines are externally accessible. swp_*() routines are 386 * internal. 387 */ 388static int dmmax; 389static int nswap_lowat = 128; /* in pages, swap_pager_almost_full warn */ 390static int nswap_hiwat = 512; /* in pages, swap_pager_almost_full warn */ 391 392SYSCTL_INT(_vm, OID_AUTO, dmmax, 393 CTLFLAG_RD, &dmmax, 0, "Maximum size of a swap block"); 394 395static void swp_sizecheck(void); 396static void swp_pager_async_iodone(struct buf *bp); 397static int swapongeom(struct thread *, struct vnode *); 398static int swaponvp(struct thread *, struct vnode *, u_long); 399static int swapoff_one(struct swdevt *sp, struct ucred *cred); 400 401/* 402 * Swap bitmap functions 403 */ 404static void swp_pager_freeswapspace(daddr_t blk, int npages); 405static daddr_t swp_pager_getswapspace(int npages); 406 407/* 408 * Metadata functions 409 */ 410static struct swblock **swp_pager_hash(vm_object_t object, vm_pindex_t index); 411static void swp_pager_meta_build(vm_object_t, vm_pindex_t, daddr_t); 412static void swp_pager_meta_free(vm_object_t, vm_pindex_t, daddr_t); 413static void swp_pager_meta_free_all(vm_object_t); 414static daddr_t swp_pager_meta_ctl(vm_object_t, vm_pindex_t, int); 415 416static void 417swp_pager_free_nrpage(vm_page_t m) 418{ 419 420 vm_page_lock(m); 421 if (m->wire_count == 0) 422 vm_page_free(m); 423 vm_page_unlock(m); 424} 425 426/* 427 * SWP_SIZECHECK() - update swap_pager_full indication 428 * 429 * update the swap_pager_almost_full indication and warn when we are 430 * about to run out of swap space, using lowat/hiwat hysteresis. 431 * 432 * Clear swap_pager_full ( task killing ) indication when lowat is met. 433 * 434 * No restrictions on call 435 * This routine may not block. 436 */ 437static void 438swp_sizecheck(void) 439{ 440 441 if (swap_pager_avail < nswap_lowat) { 442 if (swap_pager_almost_full == 0) { 443 printf("swap_pager: out of swap space\n"); 444 swap_pager_almost_full = 1; 445 } 446 } else { 447 swap_pager_full = 0; 448 if (swap_pager_avail > nswap_hiwat) 449 swap_pager_almost_full = 0; 450 } 451} 452 453/* 454 * SWP_PAGER_HASH() - hash swap meta data 455 * 456 * This is an helper function which hashes the swapblk given 457 * the object and page index. It returns a pointer to a pointer 458 * to the object, or a pointer to a NULL pointer if it could not 459 * find a swapblk. 460 */ 461static struct swblock ** 462swp_pager_hash(vm_object_t object, vm_pindex_t index) 463{ 464 struct swblock **pswap; 465 struct swblock *swap; 466 467 index &= ~(vm_pindex_t)SWAP_META_MASK; 468 pswap = &swhash[(index ^ (int)(intptr_t)object) & swhash_mask]; 469 while ((swap = *pswap) != NULL) { 470 if (swap->swb_object == object && 471 swap->swb_index == index 472 ) { 473 break; 474 } 475 pswap = &swap->swb_hnext; 476 } 477 return (pswap); 478} 479 480/* 481 * SWAP_PAGER_INIT() - initialize the swap pager! 482 * 483 * Expected to be started from system init. NOTE: This code is run 484 * before much else so be careful what you depend on. Most of the VM 485 * system has yet to be initialized at this point. 486 */ 487static void 488swap_pager_init(void) 489{ 490 /* 491 * Initialize object lists 492 */ 493 int i; 494 495 for (i = 0; i < NOBJLISTS; ++i) 496 TAILQ_INIT(&swap_pager_object_list[i]); 497 mtx_init(&sw_alloc_mtx, "swap_pager list", NULL, MTX_DEF); 498 mtx_init(&sw_dev_mtx, "swapdev", NULL, MTX_DEF); 499 500 /* 501 * Device Stripe, in PAGE_SIZE'd blocks 502 */ 503 dmmax = SWB_NPAGES * 2; 504} 505 506/* 507 * SWAP_PAGER_SWAP_INIT() - swap pager initialization from pageout process 508 * 509 * Expected to be started from pageout process once, prior to entering 510 * its main loop. 511 */ 512void 513swap_pager_swap_init(void) 514{ 515 unsigned long n, n2; 516 517 /* 518 * Number of in-transit swap bp operations. Don't 519 * exhaust the pbufs completely. Make sure we 520 * initialize workable values (0 will work for hysteresis 521 * but it isn't very efficient). 522 * 523 * The nsw_cluster_max is constrained by the bp->b_pages[] 524 * array (MAXPHYS/PAGE_SIZE) and our locally defined 525 * MAX_PAGEOUT_CLUSTER. Also be aware that swap ops are 526 * constrained by the swap device interleave stripe size. 527 * 528 * Currently we hardwire nsw_wcount_async to 4. This limit is 529 * designed to prevent other I/O from having high latencies due to 530 * our pageout I/O. The value 4 works well for one or two active swap 531 * devices but is probably a little low if you have more. Even so, 532 * a higher value would probably generate only a limited improvement 533 * with three or four active swap devices since the system does not 534 * typically have to pageout at extreme bandwidths. We will want 535 * at least 2 per swap devices, and 4 is a pretty good value if you 536 * have one NFS swap device due to the command/ack latency over NFS. 537 * So it all works out pretty well. 538 */ 539 nsw_cluster_max = min((MAXPHYS/PAGE_SIZE), MAX_PAGEOUT_CLUSTER); 540 541 mtx_lock(&pbuf_mtx); 542 nsw_rcount = (nswbuf + 1) / 2; 543 nsw_wcount_sync = (nswbuf + 3) / 4; 544 nsw_wcount_async = 4; 545 nsw_wcount_async_max = nsw_wcount_async; 546 mtx_unlock(&pbuf_mtx); 547 548 /* 549 * Initialize our zone. Right now I'm just guessing on the number 550 * we need based on the number of pages in the system. Each swblock 551 * can hold 32 pages, so this is probably overkill. This reservation 552 * is typically limited to around 32MB by default. 553 */ 554 n = cnt.v_page_count / 2; 555 if (maxswzone && n > maxswzone / sizeof(struct swblock)) 556 n = maxswzone / sizeof(struct swblock); 557 n2 = n; 558 swap_zone = uma_zcreate("SWAPMETA", sizeof(struct swblock), NULL, NULL, 559 NULL, NULL, UMA_ALIGN_PTR, UMA_ZONE_NOFREE | UMA_ZONE_VM); 560 if (swap_zone == NULL) 561 panic("failed to create swap_zone."); 562 do { 563 if (uma_zone_reserve_kva(swap_zone, n)) 564 break; 565 /* 566 * if the allocation failed, try a zone two thirds the 567 * size of the previous attempt. 568 */ 569 n -= ((n + 2) / 3); 570 } while (n > 0); 571 if (n2 != n) 572 printf("Swap zone entries reduced from %lu to %lu.\n", n2, n); 573 swap_maxpages = n * SWAP_META_PAGES; 574 swzone = n * sizeof(struct swblock); 575 n2 = n; 576 577 /* 578 * Initialize our meta-data hash table. The swapper does not need to 579 * be quite as efficient as the VM system, so we do not use an 580 * oversized hash table. 581 * 582 * n: size of hash table, must be power of 2 583 * swhash_mask: hash table index mask 584 */ 585 for (n = 1; n < n2 / 8; n *= 2) 586 ; 587 swhash = malloc(sizeof(struct swblock *) * n, M_VMPGDATA, M_WAITOK | M_ZERO); 588 swhash_mask = n - 1; 589 mtx_init(&swhash_mtx, "swap_pager swhash", NULL, MTX_DEF); 590} 591 592/* 593 * SWAP_PAGER_ALLOC() - allocate a new OBJT_SWAP VM object and instantiate 594 * its metadata structures. 595 * 596 * This routine is called from the mmap and fork code to create a new 597 * OBJT_SWAP object. We do this by creating an OBJT_DEFAULT object 598 * and then converting it with swp_pager_meta_build(). 599 * 600 * This routine may block in vm_object_allocate() and create a named 601 * object lookup race, so we must interlock. 602 * 603 * MPSAFE 604 */ 605static vm_object_t 606swap_pager_alloc(void *handle, vm_ooffset_t size, vm_prot_t prot, 607 vm_ooffset_t offset, struct ucred *cred) 608{ 609 vm_object_t object; 610 vm_pindex_t pindex; 611 612 pindex = OFF_TO_IDX(offset + PAGE_MASK + size); 613 if (handle) { 614 mtx_lock(&Giant); 615 /* 616 * Reference existing named region or allocate new one. There 617 * should not be a race here against swp_pager_meta_build() 618 * as called from vm_page_remove() in regards to the lookup 619 * of the handle. 620 */ 621 sx_xlock(&sw_alloc_sx); 622 object = vm_pager_object_lookup(NOBJLIST(handle), handle); 623 if (object == NULL) { 624 if (cred != NULL) { 625 if (!swap_reserve_by_cred(size, cred)) { 626 sx_xunlock(&sw_alloc_sx); 627 mtx_unlock(&Giant); 628 return (NULL); 629 } 630 crhold(cred); 631 } 632 object = vm_object_allocate(OBJT_DEFAULT, pindex); 633 VM_OBJECT_WLOCK(object); 634 object->handle = handle; 635 if (cred != NULL) { 636 object->cred = cred; 637 object->charge = size; 638 } 639 swp_pager_meta_build(object, 0, SWAPBLK_NONE); 640 VM_OBJECT_WUNLOCK(object); 641 } 642 sx_xunlock(&sw_alloc_sx); 643 mtx_unlock(&Giant); 644 } else { 645 if (cred != NULL) { 646 if (!swap_reserve_by_cred(size, cred)) 647 return (NULL); 648 crhold(cred); 649 } 650 object = vm_object_allocate(OBJT_DEFAULT, pindex); 651 VM_OBJECT_WLOCK(object); 652 if (cred != NULL) { 653 object->cred = cred; 654 object->charge = size; 655 } 656 swp_pager_meta_build(object, 0, SWAPBLK_NONE); 657 VM_OBJECT_WUNLOCK(object); 658 } 659 return (object); 660} 661 662/* 663 * SWAP_PAGER_DEALLOC() - remove swap metadata from object 664 * 665 * The swap backing for the object is destroyed. The code is 666 * designed such that we can reinstantiate it later, but this 667 * routine is typically called only when the entire object is 668 * about to be destroyed. 669 * 670 * The object must be locked. 671 */ 672static void 673swap_pager_dealloc(vm_object_t object) 674{ 675 676 /* 677 * Remove from list right away so lookups will fail if we block for 678 * pageout completion. 679 */ 680 if (object->handle != NULL) { 681 mtx_lock(&sw_alloc_mtx); 682 TAILQ_REMOVE(NOBJLIST(object->handle), object, pager_object_list); 683 mtx_unlock(&sw_alloc_mtx); 684 } 685 686 VM_OBJECT_ASSERT_WLOCKED(object); 687 vm_object_pip_wait(object, "swpdea"); 688 689 /* 690 * Free all remaining metadata. We only bother to free it from 691 * the swap meta data. We do not attempt to free swapblk's still 692 * associated with vm_page_t's for this object. We do not care 693 * if paging is still in progress on some objects. 694 */ 695 swp_pager_meta_free_all(object); 696 object->handle = NULL; 697 object->type = OBJT_DEAD; 698} 699 700/************************************************************************ 701 * SWAP PAGER BITMAP ROUTINES * 702 ************************************************************************/ 703 704/* 705 * SWP_PAGER_GETSWAPSPACE() - allocate raw swap space 706 * 707 * Allocate swap for the requested number of pages. The starting 708 * swap block number (a page index) is returned or SWAPBLK_NONE 709 * if the allocation failed. 710 * 711 * Also has the side effect of advising that somebody made a mistake 712 * when they configured swap and didn't configure enough. 713 * 714 * This routine may not sleep. 715 * 716 * We allocate in round-robin fashion from the configured devices. 717 */ 718static daddr_t 719swp_pager_getswapspace(int npages) 720{ 721 daddr_t blk; 722 struct swdevt *sp; 723 int i; 724 725 blk = SWAPBLK_NONE; 726 mtx_lock(&sw_dev_mtx); 727 sp = swdevhd; 728 for (i = 0; i < nswapdev; i++) { 729 if (sp == NULL) 730 sp = TAILQ_FIRST(&swtailq); 731 if (!(sp->sw_flags & SW_CLOSING)) { 732 blk = blist_alloc(sp->sw_blist, npages); 733 if (blk != SWAPBLK_NONE) { 734 blk += sp->sw_first; 735 sp->sw_used += npages; 736 swap_pager_avail -= npages; 737 swp_sizecheck(); 738 swdevhd = TAILQ_NEXT(sp, sw_list); 739 goto done; 740 } 741 } 742 sp = TAILQ_NEXT(sp, sw_list); 743 } 744 if (swap_pager_full != 2) { 745 printf("swap_pager_getswapspace(%d): failed\n", npages); 746 swap_pager_full = 2; 747 swap_pager_almost_full = 1; 748 } 749 swdevhd = NULL; 750done: 751 mtx_unlock(&sw_dev_mtx); 752 return (blk); 753} 754 755static int 756swp_pager_isondev(daddr_t blk, struct swdevt *sp) 757{ 758 759 return (blk >= sp->sw_first && blk < sp->sw_end); 760} 761 762static void 763swp_pager_strategy(struct buf *bp) 764{ 765 struct swdevt *sp; 766 767 mtx_lock(&sw_dev_mtx); 768 TAILQ_FOREACH(sp, &swtailq, sw_list) { 769 if (bp->b_blkno >= sp->sw_first && bp->b_blkno < sp->sw_end) { 770 mtx_unlock(&sw_dev_mtx); 771 if ((sp->sw_flags & SW_UNMAPPED) != 0 && 772 unmapped_buf_allowed) { 773 bp->b_kvaalloc = bp->b_data; 774 bp->b_data = unmapped_buf; 775 bp->b_kvabase = unmapped_buf; 776 bp->b_offset = 0; 777 bp->b_flags |= B_UNMAPPED; 778 } else { 779 pmap_qenter((vm_offset_t)bp->b_data, 780 &bp->b_pages[0], bp->b_bcount / PAGE_SIZE); 781 } 782 sp->sw_strategy(bp, sp); 783 return; 784 } 785 } 786 panic("Swapdev not found"); 787} 788 789 790/* 791 * SWP_PAGER_FREESWAPSPACE() - free raw swap space 792 * 793 * This routine returns the specified swap blocks back to the bitmap. 794 * 795 * This routine may not sleep. 796 */ 797static void 798swp_pager_freeswapspace(daddr_t blk, int npages) 799{ 800 struct swdevt *sp; 801 802 mtx_lock(&sw_dev_mtx); 803 TAILQ_FOREACH(sp, &swtailq, sw_list) { 804 if (blk >= sp->sw_first && blk < sp->sw_end) { 805 sp->sw_used -= npages; 806 /* 807 * If we are attempting to stop swapping on 808 * this device, we don't want to mark any 809 * blocks free lest they be reused. 810 */ 811 if ((sp->sw_flags & SW_CLOSING) == 0) { 812 blist_free(sp->sw_blist, blk - sp->sw_first, 813 npages); 814 swap_pager_avail += npages; 815 swp_sizecheck(); 816 } 817 mtx_unlock(&sw_dev_mtx); 818 return; 819 } 820 } 821 panic("Swapdev not found"); 822} 823 824/* 825 * SWAP_PAGER_FREESPACE() - frees swap blocks associated with a page 826 * range within an object. 827 * 828 * This is a globally accessible routine. 829 * 830 * This routine removes swapblk assignments from swap metadata. 831 * 832 * The external callers of this routine typically have already destroyed 833 * or renamed vm_page_t's associated with this range in the object so 834 * we should be ok. 835 * 836 * The object must be locked. 837 */ 838void 839swap_pager_freespace(vm_object_t object, vm_pindex_t start, vm_size_t size) 840{ 841 842 swp_pager_meta_free(object, start, size); 843} 844 845/* 846 * SWAP_PAGER_RESERVE() - reserve swap blocks in object 847 * 848 * Assigns swap blocks to the specified range within the object. The 849 * swap blocks are not zeroed. Any previous swap assignment is destroyed. 850 * 851 * Returns 0 on success, -1 on failure. 852 */ 853int 854swap_pager_reserve(vm_object_t object, vm_pindex_t start, vm_size_t size) 855{ 856 int n = 0; 857 daddr_t blk = SWAPBLK_NONE; 858 vm_pindex_t beg = start; /* save start index */ 859 860 VM_OBJECT_WLOCK(object); 861 while (size) { 862 if (n == 0) { 863 n = BLIST_MAX_ALLOC; 864 while ((blk = swp_pager_getswapspace(n)) == SWAPBLK_NONE) { 865 n >>= 1; 866 if (n == 0) { 867 swp_pager_meta_free(object, beg, start - beg); 868 VM_OBJECT_WUNLOCK(object); 869 return (-1); 870 } 871 } 872 } 873 swp_pager_meta_build(object, start, blk); 874 --size; 875 ++start; 876 ++blk; 877 --n; 878 } 879 swp_pager_meta_free(object, start, n); 880 VM_OBJECT_WUNLOCK(object); 881 return (0); 882} 883 884/* 885 * SWAP_PAGER_COPY() - copy blocks from source pager to destination pager 886 * and destroy the source. 887 * 888 * Copy any valid swapblks from the source to the destination. In 889 * cases where both the source and destination have a valid swapblk, 890 * we keep the destination's. 891 * 892 * This routine is allowed to sleep. It may sleep allocating metadata 893 * indirectly through swp_pager_meta_build() or if paging is still in 894 * progress on the source. 895 * 896 * The source object contains no vm_page_t's (which is just as well) 897 * 898 * The source object is of type OBJT_SWAP. 899 * 900 * The source and destination objects must be locked. 901 * Both object locks may temporarily be released. 902 */ 903void 904swap_pager_copy(vm_object_t srcobject, vm_object_t dstobject, 905 vm_pindex_t offset, int destroysource) 906{ 907 vm_pindex_t i; 908 909 VM_OBJECT_ASSERT_WLOCKED(srcobject); 910 VM_OBJECT_ASSERT_WLOCKED(dstobject); 911 912 /* 913 * If destroysource is set, we remove the source object from the 914 * swap_pager internal queue now. 915 */ 916 if (destroysource) { 917 if (srcobject->handle != NULL) { 918 mtx_lock(&sw_alloc_mtx); 919 TAILQ_REMOVE( 920 NOBJLIST(srcobject->handle), 921 srcobject, 922 pager_object_list 923 ); 924 mtx_unlock(&sw_alloc_mtx); 925 } 926 } 927 928 /* 929 * transfer source to destination. 930 */ 931 for (i = 0; i < dstobject->size; ++i) { 932 daddr_t dstaddr; 933 934 /* 935 * Locate (without changing) the swapblk on the destination, 936 * unless it is invalid in which case free it silently, or 937 * if the destination is a resident page, in which case the 938 * source is thrown away. 939 */ 940 dstaddr = swp_pager_meta_ctl(dstobject, i, 0); 941 942 if (dstaddr == SWAPBLK_NONE) { 943 /* 944 * Destination has no swapblk and is not resident, 945 * copy source. 946 */ 947 daddr_t srcaddr; 948 949 srcaddr = swp_pager_meta_ctl( 950 srcobject, 951 i + offset, 952 SWM_POP 953 ); 954 955 if (srcaddr != SWAPBLK_NONE) { 956 /* 957 * swp_pager_meta_build() can sleep. 958 */ 959 vm_object_pip_add(srcobject, 1); 960 VM_OBJECT_WUNLOCK(srcobject); 961 vm_object_pip_add(dstobject, 1); 962 swp_pager_meta_build(dstobject, i, srcaddr); 963 vm_object_pip_wakeup(dstobject); 964 VM_OBJECT_WLOCK(srcobject); 965 vm_object_pip_wakeup(srcobject); 966 } 967 } else { 968 /* 969 * Destination has valid swapblk or it is represented 970 * by a resident page. We destroy the sourceblock. 971 */ 972 973 swp_pager_meta_ctl(srcobject, i + offset, SWM_FREE); 974 } 975 } 976 977 /* 978 * Free left over swap blocks in source. 979 * 980 * We have to revert the type to OBJT_DEFAULT so we do not accidently 981 * double-remove the object from the swap queues. 982 */ 983 if (destroysource) { 984 swp_pager_meta_free_all(srcobject); 985 /* 986 * Reverting the type is not necessary, the caller is going 987 * to destroy srcobject directly, but I'm doing it here 988 * for consistency since we've removed the object from its 989 * queues. 990 */ 991 srcobject->type = OBJT_DEFAULT; 992 } 993} 994 995/* 996 * SWAP_PAGER_HASPAGE() - determine if we have good backing store for 997 * the requested page. 998 * 999 * We determine whether good backing store exists for the requested 1000 * page and return TRUE if it does, FALSE if it doesn't. 1001 * 1002 * If TRUE, we also try to determine how much valid, contiguous backing 1003 * store exists before and after the requested page within a reasonable 1004 * distance. We do not try to restrict it to the swap device stripe 1005 * (that is handled in getpages/putpages). It probably isn't worth 1006 * doing here. 1007 */ 1008static boolean_t 1009swap_pager_haspage(vm_object_t object, vm_pindex_t pindex, int *before, int *after) 1010{ 1011 daddr_t blk0; 1012 1013 VM_OBJECT_ASSERT_LOCKED(object); 1014 /* 1015 * do we have good backing store at the requested index ? 1016 */ 1017 blk0 = swp_pager_meta_ctl(object, pindex, 0); 1018 1019 if (blk0 == SWAPBLK_NONE) { 1020 if (before) 1021 *before = 0; 1022 if (after) 1023 *after = 0; 1024 return (FALSE); 1025 } 1026 1027 /* 1028 * find backwards-looking contiguous good backing store 1029 */ 1030 if (before != NULL) { 1031 int i; 1032 1033 for (i = 1; i < (SWB_NPAGES/2); ++i) { 1034 daddr_t blk; 1035 1036 if (i > pindex) 1037 break; 1038 blk = swp_pager_meta_ctl(object, pindex - i, 0); 1039 if (blk != blk0 - i) 1040 break; 1041 } 1042 *before = (i - 1); 1043 } 1044 1045 /* 1046 * find forward-looking contiguous good backing store 1047 */ 1048 if (after != NULL) { 1049 int i; 1050 1051 for (i = 1; i < (SWB_NPAGES/2); ++i) { 1052 daddr_t blk; 1053 1054 blk = swp_pager_meta_ctl(object, pindex + i, 0); 1055 if (blk != blk0 + i) 1056 break; 1057 } 1058 *after = (i - 1); 1059 } 1060 return (TRUE); 1061} 1062 1063/* 1064 * SWAP_PAGER_PAGE_UNSWAPPED() - remove swap backing store related to page 1065 * 1066 * This removes any associated swap backing store, whether valid or 1067 * not, from the page. 1068 * 1069 * This routine is typically called when a page is made dirty, at 1070 * which point any associated swap can be freed. MADV_FREE also 1071 * calls us in a special-case situation 1072 * 1073 * NOTE!!! If the page is clean and the swap was valid, the caller 1074 * should make the page dirty before calling this routine. This routine 1075 * does NOT change the m->dirty status of the page. Also: MADV_FREE 1076 * depends on it. 1077 * 1078 * This routine may not sleep. 1079 * 1080 * The object containing the page must be locked. 1081 */ 1082static void 1083swap_pager_unswapped(vm_page_t m) 1084{ 1085 1086 swp_pager_meta_ctl(m->object, m->pindex, SWM_FREE); 1087} 1088 1089/* 1090 * SWAP_PAGER_GETPAGES() - bring pages in from swap 1091 * 1092 * Attempt to retrieve (m, count) pages from backing store, but make 1093 * sure we retrieve at least m[reqpage]. We try to load in as large 1094 * a chunk surrounding m[reqpage] as is contiguous in swap and which 1095 * belongs to the same object. 1096 * 1097 * The code is designed for asynchronous operation and 1098 * immediate-notification of 'reqpage' but tends not to be 1099 * used that way. Please do not optimize-out this algorithmic 1100 * feature, I intend to improve on it in the future. 1101 * 1102 * The parent has a single vm_object_pip_add() reference prior to 1103 * calling us and we should return with the same. 1104 * 1105 * The parent has BUSY'd the pages. We should return with 'm' 1106 * left busy, but the others adjusted. 1107 */ 1108static int 1109swap_pager_getpages(vm_object_t object, vm_page_t *m, int count, int reqpage) 1110{ 1111 struct buf *bp; 1112 vm_page_t mreq; 1113 int i; 1114 int j; 1115 daddr_t blk; 1116 1117 mreq = m[reqpage]; 1118 1119 KASSERT(mreq->object == object, 1120 ("swap_pager_getpages: object mismatch %p/%p", 1121 object, mreq->object)); 1122 1123 /* 1124 * Calculate range to retrieve. The pages have already been assigned 1125 * their swapblks. We require a *contiguous* range but we know it to 1126 * not span devices. If we do not supply it, bad things 1127 * happen. Note that blk, iblk & jblk can be SWAPBLK_NONE, but the 1128 * loops are set up such that the case(s) are handled implicitly. 1129 * 1130 * The swp_*() calls must be made with the object locked. 1131 */ 1132 blk = swp_pager_meta_ctl(mreq->object, mreq->pindex, 0); 1133 1134 for (i = reqpage - 1; i >= 0; --i) { 1135 daddr_t iblk; 1136 1137 iblk = swp_pager_meta_ctl(m[i]->object, m[i]->pindex, 0); 1138 if (blk != iblk + (reqpage - i)) 1139 break; 1140 } 1141 ++i; 1142 1143 for (j = reqpage + 1; j < count; ++j) { 1144 daddr_t jblk; 1145 1146 jblk = swp_pager_meta_ctl(m[j]->object, m[j]->pindex, 0); 1147 if (blk != jblk - (j - reqpage)) 1148 break; 1149 } 1150 1151 /* 1152 * free pages outside our collection range. Note: we never free 1153 * mreq, it must remain busy throughout. 1154 */ 1155 if (0 < i || j < count) { 1156 int k; 1157 1158 for (k = 0; k < i; ++k) 1159 swp_pager_free_nrpage(m[k]); 1160 for (k = j; k < count; ++k) 1161 swp_pager_free_nrpage(m[k]); 1162 } 1163 1164 /* 1165 * Return VM_PAGER_FAIL if we have nothing to do. Return mreq 1166 * still busy, but the others unbusied. 1167 */ 1168 if (blk == SWAPBLK_NONE) 1169 return (VM_PAGER_FAIL); 1170 1171 /* 1172 * Getpbuf() can sleep. 1173 */ 1174 VM_OBJECT_WUNLOCK(object); 1175 /* 1176 * Get a swap buffer header to perform the IO 1177 */ 1178 bp = getpbuf(&nsw_rcount); 1179 bp->b_flags |= B_PAGING; 1180 1181 bp->b_iocmd = BIO_READ; 1182 bp->b_iodone = swp_pager_async_iodone; 1183 bp->b_rcred = crhold(thread0.td_ucred); 1184 bp->b_wcred = crhold(thread0.td_ucred); 1185 bp->b_blkno = blk - (reqpage - i); 1186 bp->b_bcount = PAGE_SIZE * (j - i); 1187 bp->b_bufsize = PAGE_SIZE * (j - i); 1188 bp->b_pager.pg_reqpage = reqpage - i; 1189 1190 VM_OBJECT_WLOCK(object); 1191 { 1192 int k; 1193 1194 for (k = i; k < j; ++k) { 1195 bp->b_pages[k - i] = m[k]; 1196 m[k]->oflags |= VPO_SWAPINPROG; 1197 } 1198 } 1199 bp->b_npages = j - i; 1200 1201 PCPU_INC(cnt.v_swapin); 1202 PCPU_ADD(cnt.v_swappgsin, bp->b_npages); 1203 1204 /* 1205 * We still hold the lock on mreq, and our automatic completion routine 1206 * does not remove it. 1207 */ 1208 vm_object_pip_add(object, bp->b_npages); 1209 VM_OBJECT_WUNLOCK(object); 1210 1211 /* 1212 * perform the I/O. NOTE!!! bp cannot be considered valid after 1213 * this point because we automatically release it on completion. 1214 * Instead, we look at the one page we are interested in which we 1215 * still hold a lock on even through the I/O completion. 1216 * 1217 * The other pages in our m[] array are also released on completion, 1218 * so we cannot assume they are valid anymore either. 1219 * 1220 * NOTE: b_blkno is destroyed by the call to swapdev_strategy 1221 */ 1222 BUF_KERNPROC(bp); 1223 swp_pager_strategy(bp); 1224 1225 /* 1226 * wait for the page we want to complete. VPO_SWAPINPROG is always 1227 * cleared on completion. If an I/O error occurs, SWAPBLK_NONE 1228 * is set in the meta-data. 1229 */ 1230 VM_OBJECT_WLOCK(object); 1231 while ((mreq->oflags & VPO_SWAPINPROG) != 0) { 1232 mreq->oflags |= VPO_SWAPSLEEP; 1233 PCPU_INC(cnt.v_intrans); 1234 if (VM_OBJECT_SLEEP(object, &object->paging_in_progress, PSWP, 1235 "swread", hz * 20)) { 1236 printf( 1237"swap_pager: indefinite wait buffer: bufobj: %p, blkno: %jd, size: %ld\n", 1238 bp->b_bufobj, (intmax_t)bp->b_blkno, bp->b_bcount); 1239 } 1240 } 1241 1242 /* 1243 * mreq is left busied after completion, but all the other pages 1244 * are freed. If we had an unrecoverable read error the page will 1245 * not be valid. 1246 */ 1247 if (mreq->valid != VM_PAGE_BITS_ALL) { 1248 return (VM_PAGER_ERROR); 1249 } else { 1250 return (VM_PAGER_OK); 1251 } 1252 1253 /* 1254 * A final note: in a low swap situation, we cannot deallocate swap 1255 * and mark a page dirty here because the caller is likely to mark 1256 * the page clean when we return, causing the page to possibly revert 1257 * to all-zero's later. 1258 */ 1259} 1260 1261/* 1262 * swap_pager_putpages: 1263 * 1264 * Assign swap (if necessary) and initiate I/O on the specified pages. 1265 * 1266 * We support both OBJT_DEFAULT and OBJT_SWAP objects. DEFAULT objects 1267 * are automatically converted to SWAP objects. 1268 * 1269 * In a low memory situation we may block in VOP_STRATEGY(), but the new 1270 * vm_page reservation system coupled with properly written VFS devices 1271 * should ensure that no low-memory deadlock occurs. This is an area 1272 * which needs work. 1273 * 1274 * The parent has N vm_object_pip_add() references prior to 1275 * calling us and will remove references for rtvals[] that are 1276 * not set to VM_PAGER_PEND. We need to remove the rest on I/O 1277 * completion. 1278 * 1279 * The parent has soft-busy'd the pages it passes us and will unbusy 1280 * those whos rtvals[] entry is not set to VM_PAGER_PEND on return. 1281 * We need to unbusy the rest on I/O completion. 1282 */ 1283void 1284swap_pager_putpages(vm_object_t object, vm_page_t *m, int count, 1285 int flags, int *rtvals) 1286{ 1287 int i, n; 1288 boolean_t sync; 1289 1290 if (count && m[0]->object != object) { 1291 panic("swap_pager_putpages: object mismatch %p/%p", 1292 object, 1293 m[0]->object 1294 ); 1295 } 1296 1297 /* 1298 * Step 1 1299 * 1300 * Turn object into OBJT_SWAP 1301 * check for bogus sysops 1302 * force sync if not pageout process 1303 */ 1304 if (object->type != OBJT_SWAP) 1305 swp_pager_meta_build(object, 0, SWAPBLK_NONE); 1306 VM_OBJECT_WUNLOCK(object); 1307 1308 n = 0; 1309 if (curproc != pageproc) 1310 sync = TRUE; 1311 else 1312 sync = (flags & VM_PAGER_PUT_SYNC) != 0; 1313 1314 /* 1315 * Step 2 1316 * 1317 * Update nsw parameters from swap_async_max sysctl values. 1318 * Do not let the sysop crash the machine with bogus numbers. 1319 */ 1320 mtx_lock(&pbuf_mtx); 1321 if (swap_async_max != nsw_wcount_async_max) { 1322 int n; 1323 1324 /* 1325 * limit range 1326 */ 1327 if ((n = swap_async_max) > nswbuf / 2) 1328 n = nswbuf / 2; 1329 if (n < 1) 1330 n = 1; 1331 swap_async_max = n; 1332 1333 /* 1334 * Adjust difference ( if possible ). If the current async 1335 * count is too low, we may not be able to make the adjustment 1336 * at this time. 1337 */ 1338 n -= nsw_wcount_async_max; 1339 if (nsw_wcount_async + n >= 0) { 1340 nsw_wcount_async += n; 1341 nsw_wcount_async_max += n; 1342 wakeup(&nsw_wcount_async); 1343 } 1344 } 1345 mtx_unlock(&pbuf_mtx); 1346 1347 /* 1348 * Step 3 1349 * 1350 * Assign swap blocks and issue I/O. We reallocate swap on the fly. 1351 * The page is left dirty until the pageout operation completes 1352 * successfully. 1353 */ 1354 for (i = 0; i < count; i += n) { 1355 int j; 1356 struct buf *bp; 1357 daddr_t blk; 1358 1359 /* 1360 * Maximum I/O size is limited by a number of factors. 1361 */ 1362 n = min(BLIST_MAX_ALLOC, count - i); 1363 n = min(n, nsw_cluster_max); 1364 1365 /* 1366 * Get biggest block of swap we can. If we fail, fall 1367 * back and try to allocate a smaller block. Don't go 1368 * overboard trying to allocate space if it would overly 1369 * fragment swap. 1370 */ 1371 while ( 1372 (blk = swp_pager_getswapspace(n)) == SWAPBLK_NONE && 1373 n > 4 1374 ) { 1375 n >>= 1; 1376 } 1377 if (blk == SWAPBLK_NONE) { 1378 for (j = 0; j < n; ++j) 1379 rtvals[i+j] = VM_PAGER_FAIL; 1380 continue; 1381 } 1382 1383 /* 1384 * All I/O parameters have been satisfied, build the I/O 1385 * request and assign the swap space. 1386 */ 1387 if (sync == TRUE) { 1388 bp = getpbuf(&nsw_wcount_sync); 1389 } else { 1390 bp = getpbuf(&nsw_wcount_async); 1391 bp->b_flags = B_ASYNC; 1392 } 1393 bp->b_flags |= B_PAGING; 1394 bp->b_iocmd = BIO_WRITE; 1395 1396 bp->b_rcred = crhold(thread0.td_ucred); 1397 bp->b_wcred = crhold(thread0.td_ucred); 1398 bp->b_bcount = PAGE_SIZE * n; 1399 bp->b_bufsize = PAGE_SIZE * n; 1400 bp->b_blkno = blk; 1401 1402 VM_OBJECT_WLOCK(object); 1403 for (j = 0; j < n; ++j) { 1404 vm_page_t mreq = m[i+j]; 1405 1406 swp_pager_meta_build( 1407 mreq->object, 1408 mreq->pindex, 1409 blk + j 1410 ); 1411 vm_page_dirty(mreq); 1412 rtvals[i+j] = VM_PAGER_OK; 1413 1414 mreq->oflags |= VPO_SWAPINPROG; 1415 bp->b_pages[j] = mreq; 1416 } 1417 VM_OBJECT_WUNLOCK(object); 1418 bp->b_npages = n; 1419 /* 1420 * Must set dirty range for NFS to work. 1421 */ 1422 bp->b_dirtyoff = 0; 1423 bp->b_dirtyend = bp->b_bcount; 1424 1425 PCPU_INC(cnt.v_swapout); 1426 PCPU_ADD(cnt.v_swappgsout, bp->b_npages); 1427 1428 /* 1429 * asynchronous 1430 * 1431 * NOTE: b_blkno is destroyed by the call to swapdev_strategy 1432 */ 1433 if (sync == FALSE) { 1434 bp->b_iodone = swp_pager_async_iodone; 1435 BUF_KERNPROC(bp); 1436 swp_pager_strategy(bp); 1437 1438 for (j = 0; j < n; ++j) 1439 rtvals[i+j] = VM_PAGER_PEND; 1440 /* restart outter loop */ 1441 continue; 1442 } 1443 1444 /* 1445 * synchronous 1446 * 1447 * NOTE: b_blkno is destroyed by the call to swapdev_strategy 1448 */ 1449 bp->b_iodone = bdone; 1450 swp_pager_strategy(bp); 1451 1452 /* 1453 * Wait for the sync I/O to complete, then update rtvals. 1454 * We just set the rtvals[] to VM_PAGER_PEND so we can call 1455 * our async completion routine at the end, thus avoiding a 1456 * double-free. 1457 */ 1458 bwait(bp, PVM, "swwrt"); 1459 for (j = 0; j < n; ++j) 1460 rtvals[i+j] = VM_PAGER_PEND; 1461 /* 1462 * Now that we are through with the bp, we can call the 1463 * normal async completion, which frees everything up. 1464 */ 1465 swp_pager_async_iodone(bp); 1466 } 1467 VM_OBJECT_WLOCK(object); 1468} 1469 1470/* 1471 * swp_pager_async_iodone: 1472 * 1473 * Completion routine for asynchronous reads and writes from/to swap. 1474 * Also called manually by synchronous code to finish up a bp. 1475 * 1476 * This routine may not sleep. 1477 */ 1478static void 1479swp_pager_async_iodone(struct buf *bp) 1480{ 1481 int i; 1482 vm_object_t object = NULL; 1483 1484 /* 1485 * report error 1486 */ 1487 if (bp->b_ioflags & BIO_ERROR) { 1488 printf( 1489 "swap_pager: I/O error - %s failed; blkno %ld," 1490 "size %ld, error %d\n", 1491 ((bp->b_iocmd == BIO_READ) ? "pagein" : "pageout"), 1492 (long)bp->b_blkno, 1493 (long)bp->b_bcount, 1494 bp->b_error 1495 ); 1496 } 1497 1498 /* 1499 * remove the mapping for kernel virtual 1500 */ 1501 if ((bp->b_flags & B_UNMAPPED) != 0) { 1502 bp->b_data = bp->b_kvaalloc; 1503 bp->b_kvabase = bp->b_kvaalloc; 1504 bp->b_flags &= ~B_UNMAPPED; 1505 } else 1506 pmap_qremove((vm_offset_t)bp->b_data, bp->b_npages); 1507 1508 if (bp->b_npages) { 1509 object = bp->b_pages[0]->object; 1510 VM_OBJECT_WLOCK(object); 1511 } 1512 1513 /* 1514 * cleanup pages. If an error occurs writing to swap, we are in 1515 * very serious trouble. If it happens to be a disk error, though, 1516 * we may be able to recover by reassigning the swap later on. So 1517 * in this case we remove the m->swapblk assignment for the page 1518 * but do not free it in the rlist. The errornous block(s) are thus 1519 * never reallocated as swap. Redirty the page and continue. 1520 */ 1521 for (i = 0; i < bp->b_npages; ++i) { 1522 vm_page_t m = bp->b_pages[i]; 1523 1524 m->oflags &= ~VPO_SWAPINPROG; 1525 if (m->oflags & VPO_SWAPSLEEP) { 1526 m->oflags &= ~VPO_SWAPSLEEP; 1527 wakeup(&object->paging_in_progress); 1528 } 1529 1530 if (bp->b_ioflags & BIO_ERROR) { 1531 /* 1532 * If an error occurs I'd love to throw the swapblk 1533 * away without freeing it back to swapspace, so it 1534 * can never be used again. But I can't from an 1535 * interrupt. 1536 */ 1537 if (bp->b_iocmd == BIO_READ) { 1538 /* 1539 * When reading, reqpage needs to stay 1540 * locked for the parent, but all other 1541 * pages can be freed. We still want to 1542 * wakeup the parent waiting on the page, 1543 * though. ( also: pg_reqpage can be -1 and 1544 * not match anything ). 1545 * 1546 * We have to wake specifically requested pages 1547 * up too because we cleared VPO_SWAPINPROG and 1548 * someone may be waiting for that. 1549 * 1550 * NOTE: for reads, m->dirty will probably 1551 * be overridden by the original caller of 1552 * getpages so don't play cute tricks here. 1553 */ 1554 m->valid = 0; 1555 if (i != bp->b_pager.pg_reqpage) 1556 swp_pager_free_nrpage(m); 1557 else { 1558 vm_page_lock(m); 1559 vm_page_flash(m); 1560 vm_page_unlock(m); 1561 } 1562 /* 1563 * If i == bp->b_pager.pg_reqpage, do not wake 1564 * the page up. The caller needs to. 1565 */ 1566 } else { 1567 /* 1568 * If a write error occurs, reactivate page 1569 * so it doesn't clog the inactive list, 1570 * then finish the I/O. 1571 */ 1572 vm_page_dirty(m); 1573 vm_page_lock(m); 1574 vm_page_activate(m); 1575 vm_page_unlock(m); 1576 vm_page_sunbusy(m); 1577 } 1578 } else if (bp->b_iocmd == BIO_READ) { 1579 /* 1580 * NOTE: for reads, m->dirty will probably be 1581 * overridden by the original caller of getpages so 1582 * we cannot set them in order to free the underlying 1583 * swap in a low-swap situation. I don't think we'd 1584 * want to do that anyway, but it was an optimization 1585 * that existed in the old swapper for a time before 1586 * it got ripped out due to precisely this problem. 1587 * 1588 * If not the requested page then deactivate it. 1589 * 1590 * Note that the requested page, reqpage, is left 1591 * busied, but we still have to wake it up. The 1592 * other pages are released (unbusied) by 1593 * vm_page_xunbusy(). 1594 */ 1595 KASSERT(!pmap_page_is_mapped(m), 1596 ("swp_pager_async_iodone: page %p is mapped", m)); 1597 m->valid = VM_PAGE_BITS_ALL; 1598 KASSERT(m->dirty == 0, 1599 ("swp_pager_async_iodone: page %p is dirty", m)); 1600 1601 /* 1602 * We have to wake specifically requested pages 1603 * up too because we cleared VPO_SWAPINPROG and 1604 * could be waiting for it in getpages. However, 1605 * be sure to not unbusy getpages specifically 1606 * requested page - getpages expects it to be 1607 * left busy. 1608 */ 1609 if (i != bp->b_pager.pg_reqpage) { 1610 vm_page_lock(m); 1611 vm_page_deactivate(m); 1612 vm_page_unlock(m); 1613 vm_page_xunbusy(m); 1614 } else { 1615 vm_page_lock(m); 1616 vm_page_flash(m); 1617 vm_page_unlock(m); 1618 } 1619 } else { 1620 /* 1621 * For write success, clear the dirty 1622 * status, then finish the I/O ( which decrements the 1623 * busy count and possibly wakes waiter's up ). 1624 */ 1625 KASSERT(!pmap_page_is_write_mapped(m), 1626 ("swp_pager_async_iodone: page %p is not write" 1627 " protected", m)); 1628 vm_page_undirty(m); 1629 vm_page_sunbusy(m); 1630 if (vm_page_count_severe()) { 1631 vm_page_lock(m); 1632 vm_page_try_to_cache(m); 1633 vm_page_unlock(m); 1634 } 1635 } 1636 } 1637 1638 /* 1639 * adjust pip. NOTE: the original parent may still have its own 1640 * pip refs on the object. 1641 */ 1642 if (object != NULL) { 1643 vm_object_pip_wakeupn(object, bp->b_npages); 1644 VM_OBJECT_WUNLOCK(object); 1645 } 1646 1647 /* 1648 * swapdev_strategy() manually sets b_vp and b_bufobj before calling 1649 * bstrategy(). Set them back to NULL now we're done with it, or we'll 1650 * trigger a KASSERT in relpbuf(). 1651 */ 1652 if (bp->b_vp) { 1653 bp->b_vp = NULL; 1654 bp->b_bufobj = NULL; 1655 } 1656 /* 1657 * release the physical I/O buffer 1658 */ 1659 relpbuf( 1660 bp, 1661 ((bp->b_iocmd == BIO_READ) ? &nsw_rcount : 1662 ((bp->b_flags & B_ASYNC) ? 1663 &nsw_wcount_async : 1664 &nsw_wcount_sync 1665 ) 1666 ) 1667 ); 1668} 1669 1670/* 1671 * swap_pager_isswapped: 1672 * 1673 * Return 1 if at least one page in the given object is paged 1674 * out to the given swap device. 1675 * 1676 * This routine may not sleep. 1677 */ 1678int 1679swap_pager_isswapped(vm_object_t object, struct swdevt *sp) 1680{ 1681 daddr_t index = 0; 1682 int bcount; 1683 int i; 1684 1685 VM_OBJECT_ASSERT_WLOCKED(object); 1686 if (object->type != OBJT_SWAP) 1687 return (0); 1688 1689 mtx_lock(&swhash_mtx); 1690 for (bcount = 0; bcount < object->un_pager.swp.swp_bcount; bcount++) { 1691 struct swblock *swap; 1692 1693 if ((swap = *swp_pager_hash(object, index)) != NULL) { 1694 for (i = 0; i < SWAP_META_PAGES; ++i) { 1695 if (swp_pager_isondev(swap->swb_pages[i], sp)) { 1696 mtx_unlock(&swhash_mtx); 1697 return (1); 1698 } 1699 } 1700 } 1701 index += SWAP_META_PAGES; 1702 } 1703 mtx_unlock(&swhash_mtx); 1704 return (0); 1705} 1706 1707/* 1708 * SWP_PAGER_FORCE_PAGEIN() - force a swap block to be paged in 1709 * 1710 * This routine dissociates the page at the given index within a 1711 * swap block from its backing store, paging it in if necessary. 1712 * If the page is paged in, it is placed in the inactive queue, 1713 * since it had its backing store ripped out from under it. 1714 * We also attempt to swap in all other pages in the swap block, 1715 * we only guarantee that the one at the specified index is 1716 * paged in. 1717 * 1718 * XXX - The code to page the whole block in doesn't work, so we 1719 * revert to the one-by-one behavior for now. Sigh. 1720 */ 1721static inline void 1722swp_pager_force_pagein(vm_object_t object, vm_pindex_t pindex) 1723{ 1724 vm_page_t m; 1725 1726 vm_object_pip_add(object, 1); 1727 m = vm_page_grab(object, pindex, VM_ALLOC_NORMAL); 1728 if (m->valid == VM_PAGE_BITS_ALL) { 1729 vm_object_pip_subtract(object, 1); 1730 vm_page_dirty(m); 1731 vm_page_lock(m); 1732 vm_page_activate(m); 1733 vm_page_unlock(m); 1734 vm_page_xunbusy(m); 1735 vm_pager_page_unswapped(m); 1736 return; 1737 } 1738 1739 if (swap_pager_getpages(object, &m, 1, 0) != VM_PAGER_OK) 1740 panic("swap_pager_force_pagein: read from swap failed");/*XXX*/ 1741 vm_object_pip_subtract(object, 1); 1742 vm_page_dirty(m); 1743 vm_page_lock(m); 1744 vm_page_deactivate(m); 1745 vm_page_unlock(m); 1746 vm_page_xunbusy(m); 1747 vm_pager_page_unswapped(m); 1748} 1749 1750/* 1751 * swap_pager_swapoff: 1752 * 1753 * Page in all of the pages that have been paged out to the 1754 * given device. The corresponding blocks in the bitmap must be 1755 * marked as allocated and the device must be flagged SW_CLOSING. 1756 * There may be no processes swapped out to the device. 1757 * 1758 * This routine may block. 1759 */ 1760static void 1761swap_pager_swapoff(struct swdevt *sp) 1762{ 1763 struct swblock *swap; 1764 int i, j, retries; 1765 1766 GIANT_REQUIRED; 1767 1768 retries = 0; 1769full_rescan: 1770 mtx_lock(&swhash_mtx); 1771 for (i = 0; i <= swhash_mask; i++) { /* '<=' is correct here */ 1772restart: 1773 for (swap = swhash[i]; swap != NULL; swap = swap->swb_hnext) { 1774 vm_object_t object = swap->swb_object; 1775 vm_pindex_t pindex = swap->swb_index; 1776 for (j = 0; j < SWAP_META_PAGES; ++j) { 1777 if (swp_pager_isondev(swap->swb_pages[j], sp)) { 1778 /* avoid deadlock */ 1779 if (!VM_OBJECT_TRYWLOCK(object)) { 1780 break; 1781 } else { 1782 mtx_unlock(&swhash_mtx); 1783 swp_pager_force_pagein(object, 1784 pindex + j); 1785 VM_OBJECT_WUNLOCK(object); 1786 mtx_lock(&swhash_mtx); 1787 goto restart; 1788 } 1789 } 1790 } 1791 } 1792 } 1793 mtx_unlock(&swhash_mtx); 1794 if (sp->sw_used) { 1795 /* 1796 * Objects may be locked or paging to the device being 1797 * removed, so we will miss their pages and need to 1798 * make another pass. We have marked this device as 1799 * SW_CLOSING, so the activity should finish soon. 1800 */ 1801 retries++; 1802 if (retries > 100) { 1803 panic("swapoff: failed to locate %d swap blocks", 1804 sp->sw_used); 1805 } 1806 pause("swpoff", hz / 20); 1807 goto full_rescan; 1808 } 1809} 1810 1811/************************************************************************ 1812 * SWAP META DATA * 1813 ************************************************************************ 1814 * 1815 * These routines manipulate the swap metadata stored in the 1816 * OBJT_SWAP object. 1817 * 1818 * Swap metadata is implemented with a global hash and not directly 1819 * linked into the object. Instead the object simply contains 1820 * appropriate tracking counters. 1821 */ 1822 1823/* 1824 * SWP_PAGER_META_BUILD() - add swap block to swap meta data for object 1825 * 1826 * We first convert the object to a swap object if it is a default 1827 * object. 1828 * 1829 * The specified swapblk is added to the object's swap metadata. If 1830 * the swapblk is not valid, it is freed instead. Any previously 1831 * assigned swapblk is freed. 1832 */ 1833static void 1834swp_pager_meta_build(vm_object_t object, vm_pindex_t pindex, daddr_t swapblk) 1835{ 1836 static volatile int exhausted; 1837 struct swblock *swap; 1838 struct swblock **pswap; 1839 int idx; 1840 1841 VM_OBJECT_ASSERT_WLOCKED(object); 1842 /* 1843 * Convert default object to swap object if necessary 1844 */ 1845 if (object->type != OBJT_SWAP) { 1846 object->type = OBJT_SWAP; 1847 object->un_pager.swp.swp_bcount = 0; 1848 1849 if (object->handle != NULL) { 1850 mtx_lock(&sw_alloc_mtx); 1851 TAILQ_INSERT_TAIL( 1852 NOBJLIST(object->handle), 1853 object, 1854 pager_object_list 1855 ); 1856 mtx_unlock(&sw_alloc_mtx); 1857 } 1858 } 1859 1860 /* 1861 * Locate hash entry. If not found create, but if we aren't adding 1862 * anything just return. If we run out of space in the map we wait 1863 * and, since the hash table may have changed, retry. 1864 */ 1865retry: 1866 mtx_lock(&swhash_mtx); 1867 pswap = swp_pager_hash(object, pindex); 1868 1869 if ((swap = *pswap) == NULL) { 1870 int i; 1871 1872 if (swapblk == SWAPBLK_NONE) 1873 goto done; 1874 1875 swap = *pswap = uma_zalloc(swap_zone, M_NOWAIT | 1876 (curproc == pageproc ? M_USE_RESERVE : 0)); 1877 if (swap == NULL) { 1878 mtx_unlock(&swhash_mtx); 1879 VM_OBJECT_WUNLOCK(object); 1880 if (uma_zone_exhausted(swap_zone)) { 1881 if (atomic_cmpset_int(&exhausted, 0, 1)) 1882 printf("swap zone exhausted, " 1883 "increase kern.maxswzone\n"); 1884 vm_pageout_oom(VM_OOM_SWAPZ); 1885 pause("swzonex", 10); 1886 } else 1887 VM_WAIT; 1888 VM_OBJECT_WLOCK(object); 1889 goto retry; 1890 } 1891 1892 if (atomic_cmpset_int(&exhausted, 1, 0)) 1893 printf("swap zone ok\n"); 1894 1895 swap->swb_hnext = NULL; 1896 swap->swb_object = object; 1897 swap->swb_index = pindex & ~(vm_pindex_t)SWAP_META_MASK; 1898 swap->swb_count = 0; 1899 1900 ++object->un_pager.swp.swp_bcount; 1901 1902 for (i = 0; i < SWAP_META_PAGES; ++i) 1903 swap->swb_pages[i] = SWAPBLK_NONE; 1904 } 1905 1906 /* 1907 * Delete prior contents of metadata 1908 */ 1909 idx = pindex & SWAP_META_MASK; 1910 1911 if (swap->swb_pages[idx] != SWAPBLK_NONE) { 1912 swp_pager_freeswapspace(swap->swb_pages[idx], 1); 1913 --swap->swb_count; 1914 } 1915 1916 /* 1917 * Enter block into metadata 1918 */ 1919 swap->swb_pages[idx] = swapblk; 1920 if (swapblk != SWAPBLK_NONE) 1921 ++swap->swb_count; 1922done: 1923 mtx_unlock(&swhash_mtx); 1924} 1925 1926/* 1927 * SWP_PAGER_META_FREE() - free a range of blocks in the object's swap metadata 1928 * 1929 * The requested range of blocks is freed, with any associated swap 1930 * returned to the swap bitmap. 1931 * 1932 * This routine will free swap metadata structures as they are cleaned 1933 * out. This routine does *NOT* operate on swap metadata associated 1934 * with resident pages. 1935 */ 1936static void 1937swp_pager_meta_free(vm_object_t object, vm_pindex_t index, daddr_t count) 1938{ 1939 1940 VM_OBJECT_ASSERT_LOCKED(object); 1941 if (object->type != OBJT_SWAP) 1942 return; 1943 1944 while (count > 0) { 1945 struct swblock **pswap; 1946 struct swblock *swap; 1947 1948 mtx_lock(&swhash_mtx); 1949 pswap = swp_pager_hash(object, index); 1950 1951 if ((swap = *pswap) != NULL) { 1952 daddr_t v = swap->swb_pages[index & SWAP_META_MASK]; 1953 1954 if (v != SWAPBLK_NONE) { 1955 swp_pager_freeswapspace(v, 1); 1956 swap->swb_pages[index & SWAP_META_MASK] = 1957 SWAPBLK_NONE; 1958 if (--swap->swb_count == 0) { 1959 *pswap = swap->swb_hnext; 1960 uma_zfree(swap_zone, swap); 1961 --object->un_pager.swp.swp_bcount; 1962 } 1963 } 1964 --count; 1965 ++index; 1966 } else { 1967 int n = SWAP_META_PAGES - (index & SWAP_META_MASK); 1968 count -= n; 1969 index += n; 1970 } 1971 mtx_unlock(&swhash_mtx); 1972 } 1973} 1974 1975/* 1976 * SWP_PAGER_META_FREE_ALL() - destroy all swap metadata associated with object 1977 * 1978 * This routine locates and destroys all swap metadata associated with 1979 * an object. 1980 */ 1981static void 1982swp_pager_meta_free_all(vm_object_t object) 1983{ 1984 daddr_t index = 0; 1985 1986 VM_OBJECT_ASSERT_WLOCKED(object); 1987 if (object->type != OBJT_SWAP) 1988 return; 1989 1990 while (object->un_pager.swp.swp_bcount) { 1991 struct swblock **pswap; 1992 struct swblock *swap; 1993 1994 mtx_lock(&swhash_mtx); 1995 pswap = swp_pager_hash(object, index); 1996 if ((swap = *pswap) != NULL) { 1997 int i; 1998 1999 for (i = 0; i < SWAP_META_PAGES; ++i) { 2000 daddr_t v = swap->swb_pages[i]; 2001 if (v != SWAPBLK_NONE) { 2002 --swap->swb_count; 2003 swp_pager_freeswapspace(v, 1); 2004 } 2005 } 2006 if (swap->swb_count != 0) 2007 panic("swap_pager_meta_free_all: swb_count != 0"); 2008 *pswap = swap->swb_hnext; 2009 uma_zfree(swap_zone, swap); 2010 --object->un_pager.swp.swp_bcount; 2011 } 2012 mtx_unlock(&swhash_mtx); 2013 index += SWAP_META_PAGES; 2014 } 2015} 2016 2017/* 2018 * SWP_PAGER_METACTL() - misc control of swap and vm_page_t meta data. 2019 * 2020 * This routine is capable of looking up, popping, or freeing 2021 * swapblk assignments in the swap meta data or in the vm_page_t. 2022 * The routine typically returns the swapblk being looked-up, or popped, 2023 * or SWAPBLK_NONE if the block was freed, or SWAPBLK_NONE if the block 2024 * was invalid. This routine will automatically free any invalid 2025 * meta-data swapblks. 2026 * 2027 * It is not possible to store invalid swapblks in the swap meta data 2028 * (other then a literal 'SWAPBLK_NONE'), so we don't bother checking. 2029 * 2030 * When acting on a busy resident page and paging is in progress, we 2031 * have to wait until paging is complete but otherwise can act on the 2032 * busy page. 2033 * 2034 * SWM_FREE remove and free swap block from metadata 2035 * SWM_POP remove from meta data but do not free.. pop it out 2036 */ 2037static daddr_t 2038swp_pager_meta_ctl(vm_object_t object, vm_pindex_t pindex, int flags) 2039{ 2040 struct swblock **pswap; 2041 struct swblock *swap; 2042 daddr_t r1; 2043 int idx; 2044 2045 VM_OBJECT_ASSERT_LOCKED(object); 2046 /* 2047 * The meta data only exists of the object is OBJT_SWAP 2048 * and even then might not be allocated yet. 2049 */ 2050 if (object->type != OBJT_SWAP) 2051 return (SWAPBLK_NONE); 2052 2053 r1 = SWAPBLK_NONE; 2054 mtx_lock(&swhash_mtx); 2055 pswap = swp_pager_hash(object, pindex); 2056 2057 if ((swap = *pswap) != NULL) { 2058 idx = pindex & SWAP_META_MASK; 2059 r1 = swap->swb_pages[idx]; 2060 2061 if (r1 != SWAPBLK_NONE) { 2062 if (flags & SWM_FREE) { 2063 swp_pager_freeswapspace(r1, 1); 2064 r1 = SWAPBLK_NONE; 2065 } 2066 if (flags & (SWM_FREE|SWM_POP)) { 2067 swap->swb_pages[idx] = SWAPBLK_NONE; 2068 if (--swap->swb_count == 0) { 2069 *pswap = swap->swb_hnext; 2070 uma_zfree(swap_zone, swap); 2071 --object->un_pager.swp.swp_bcount; 2072 } 2073 } 2074 } 2075 } 2076 mtx_unlock(&swhash_mtx); 2077 return (r1); 2078} 2079 2080/* 2081 * System call swapon(name) enables swapping on device name, 2082 * which must be in the swdevsw. Return EBUSY 2083 * if already swapping on this device. 2084 */ 2085#ifndef _SYS_SYSPROTO_H_ 2086struct swapon_args { 2087 char *name; 2088}; 2089#endif 2090 2091/* 2092 * MPSAFE 2093 */ 2094/* ARGSUSED */ 2095int 2096sys_swapon(struct thread *td, struct swapon_args *uap) 2097{ 2098 struct vattr attr; 2099 struct vnode *vp; 2100 struct nameidata nd; 2101 int error; 2102 2103 error = priv_check(td, PRIV_SWAPON); 2104 if (error) 2105 return (error); 2106 2107 mtx_lock(&Giant); 2108 while (swdev_syscall_active) 2109 tsleep(&swdev_syscall_active, PUSER - 1, "swpon", 0); 2110 swdev_syscall_active = 1; 2111 2112 /* 2113 * Swap metadata may not fit in the KVM if we have physical 2114 * memory of >1GB. 2115 */ 2116 if (swap_zone == NULL) { 2117 error = ENOMEM; 2118 goto done; 2119 } 2120 2121 NDINIT(&nd, LOOKUP, ISOPEN | FOLLOW | AUDITVNODE1, UIO_USERSPACE, 2122 uap->name, td); 2123 error = namei(&nd); 2124 if (error) 2125 goto done; 2126 2127 NDFREE(&nd, NDF_ONLY_PNBUF); 2128 vp = nd.ni_vp; 2129 2130 if (vn_isdisk(vp, &error)) { 2131 error = swapongeom(td, vp); 2132 } else if (vp->v_type == VREG && 2133 (vp->v_mount->mnt_vfc->vfc_flags & VFCF_NETWORK) != 0 && 2134 (error = VOP_GETATTR(vp, &attr, td->td_ucred)) == 0) { 2135 /* 2136 * Allow direct swapping to NFS regular files in the same 2137 * way that nfs_mountroot() sets up diskless swapping. 2138 */ 2139 error = swaponvp(td, vp, attr.va_size / DEV_BSIZE); 2140 } 2141 2142 if (error) 2143 vrele(vp); 2144done: 2145 swdev_syscall_active = 0; 2146 wakeup_one(&swdev_syscall_active); 2147 mtx_unlock(&Giant); 2148 return (error); 2149} 2150 2151/* 2152 * Check that the total amount of swap currently configured does not 2153 * exceed half the theoretical maximum. If it does, print a warning 2154 * message and return -1; otherwise, return 0. 2155 */ 2156static int 2157swapon_check_swzone(unsigned long npages) 2158{ 2159 unsigned long maxpages; 2160 2161 /* absolute maximum we can handle assuming 100% efficiency */ 2162 maxpages = uma_zone_get_max(swap_zone) * SWAP_META_PAGES; 2163 2164 /* recommend using no more than half that amount */ 2165 if (npages > maxpages / 2) { 2166 printf("warning: total configured swap (%lu pages) " 2167 "exceeds maximum recommended amount (%lu pages).\n", 2168 npages, maxpages / 2); 2169 printf("warning: increase kern.maxswzone " 2170 "or reduce amount of swap.\n"); 2171 return (-1); 2172 } 2173 return (0); 2174} 2175 2176static void 2177swaponsomething(struct vnode *vp, void *id, u_long nblks, 2178 sw_strategy_t *strategy, sw_close_t *close, dev_t dev, int flags) 2179{ 2180 struct swdevt *sp, *tsp; 2181 swblk_t dvbase; 2182 u_long mblocks; 2183 2184 /* 2185 * nblks is in DEV_BSIZE'd chunks, convert to PAGE_SIZE'd chunks. 2186 * First chop nblks off to page-align it, then convert. 2187 * 2188 * sw->sw_nblks is in page-sized chunks now too. 2189 */ 2190 nblks &= ~(ctodb(1) - 1); 2191 nblks = dbtoc(nblks); 2192 2193 /* 2194 * If we go beyond this, we get overflows in the radix 2195 * tree bitmap code. 2196 */ 2197 mblocks = 0x40000000 / BLIST_META_RADIX; 2198 if (nblks > mblocks) { 2199 printf( 2200 "WARNING: reducing swap size to maximum of %luMB per unit\n", 2201 mblocks / 1024 / 1024 * PAGE_SIZE); 2202 nblks = mblocks; 2203 } 2204 2205 sp = malloc(sizeof *sp, M_VMPGDATA, M_WAITOK | M_ZERO); 2206 sp->sw_vp = vp; 2207 sp->sw_id = id; 2208 sp->sw_dev = dev; 2209 sp->sw_flags = 0; 2210 sp->sw_nblks = nblks; 2211 sp->sw_used = 0; 2212 sp->sw_strategy = strategy; 2213 sp->sw_close = close; 2214 sp->sw_flags = flags; 2215 2216 sp->sw_blist = blist_create(nblks, M_WAITOK); 2217 /* 2218 * Do not free the first two block in order to avoid overwriting 2219 * any bsd label at the front of the partition 2220 */ 2221 blist_free(sp->sw_blist, 2, nblks - 2); 2222 2223 dvbase = 0; 2224 mtx_lock(&sw_dev_mtx); 2225 TAILQ_FOREACH(tsp, &swtailq, sw_list) { 2226 if (tsp->sw_end >= dvbase) { 2227 /* 2228 * We put one uncovered page between the devices 2229 * in order to definitively prevent any cross-device 2230 * I/O requests 2231 */ 2232 dvbase = tsp->sw_end + 1; 2233 } 2234 } 2235 sp->sw_first = dvbase; 2236 sp->sw_end = dvbase + nblks; 2237 TAILQ_INSERT_TAIL(&swtailq, sp, sw_list); 2238 nswapdev++; 2239 swap_pager_avail += nblks; 2240 swap_total += (vm_ooffset_t)nblks * PAGE_SIZE; 2241 swapon_check_swzone(swap_total / PAGE_SIZE); 2242 swp_sizecheck(); 2243 mtx_unlock(&sw_dev_mtx); 2244} 2245 2246/* 2247 * SYSCALL: swapoff(devname) 2248 * 2249 * Disable swapping on the given device. 2250 * 2251 * XXX: Badly designed system call: it should use a device index 2252 * rather than filename as specification. We keep sw_vp around 2253 * only to make this work. 2254 */ 2255#ifndef _SYS_SYSPROTO_H_ 2256struct swapoff_args { 2257 char *name; 2258}; 2259#endif 2260 2261/* 2262 * MPSAFE 2263 */ 2264/* ARGSUSED */ 2265int 2266sys_swapoff(struct thread *td, struct swapoff_args *uap) 2267{ 2268 struct vnode *vp; 2269 struct nameidata nd; 2270 struct swdevt *sp; 2271 int error; 2272 2273 error = priv_check(td, PRIV_SWAPOFF); 2274 if (error) 2275 return (error); 2276 2277 mtx_lock(&Giant); 2278 while (swdev_syscall_active) 2279 tsleep(&swdev_syscall_active, PUSER - 1, "swpoff", 0); 2280 swdev_syscall_active = 1; 2281 2282 NDINIT(&nd, LOOKUP, FOLLOW | AUDITVNODE1, UIO_USERSPACE, uap->name, 2283 td); 2284 error = namei(&nd); 2285 if (error) 2286 goto done; 2287 NDFREE(&nd, NDF_ONLY_PNBUF); 2288 vp = nd.ni_vp; 2289 2290 mtx_lock(&sw_dev_mtx); 2291 TAILQ_FOREACH(sp, &swtailq, sw_list) { 2292 if (sp->sw_vp == vp) 2293 break; 2294 } 2295 mtx_unlock(&sw_dev_mtx); 2296 if (sp == NULL) { 2297 error = EINVAL; 2298 goto done; 2299 } 2300 error = swapoff_one(sp, td->td_ucred); 2301done: 2302 swdev_syscall_active = 0; 2303 wakeup_one(&swdev_syscall_active); 2304 mtx_unlock(&Giant); 2305 return (error); 2306} 2307 2308static int 2309swapoff_one(struct swdevt *sp, struct ucred *cred) 2310{ 2311 u_long nblks, dvbase; 2312#ifdef MAC 2313 int error; 2314#endif 2315 2316 mtx_assert(&Giant, MA_OWNED); 2317#ifdef MAC 2318 (void) vn_lock(sp->sw_vp, LK_EXCLUSIVE | LK_RETRY); 2319 error = mac_system_check_swapoff(cred, sp->sw_vp); 2320 (void) VOP_UNLOCK(sp->sw_vp, 0); 2321 if (error != 0) 2322 return (error); 2323#endif 2324 nblks = sp->sw_nblks; 2325 2326 /* 2327 * We can turn off this swap device safely only if the 2328 * available virtual memory in the system will fit the amount 2329 * of data we will have to page back in, plus an epsilon so 2330 * the system doesn't become critically low on swap space. 2331 */ 2332 if (cnt.v_free_count + cnt.v_cache_count + swap_pager_avail < 2333 nblks + nswap_lowat) { 2334 return (ENOMEM); 2335 } 2336 2337 /* 2338 * Prevent further allocations on this device. 2339 */ 2340 mtx_lock(&sw_dev_mtx); 2341 sp->sw_flags |= SW_CLOSING; 2342 for (dvbase = 0; dvbase < sp->sw_end; dvbase += dmmax) { 2343 swap_pager_avail -= blist_fill(sp->sw_blist, 2344 dvbase, dmmax); 2345 } 2346 swap_total -= (vm_ooffset_t)nblks * PAGE_SIZE; 2347 mtx_unlock(&sw_dev_mtx); 2348 2349 /* 2350 * Page in the contents of the device and close it. 2351 */ 2352 swap_pager_swapoff(sp); 2353 2354 sp->sw_close(curthread, sp); 2355 sp->sw_id = NULL; 2356 mtx_lock(&sw_dev_mtx); 2357 TAILQ_REMOVE(&swtailq, sp, sw_list); 2358 nswapdev--; 2359 if (nswapdev == 0) { 2360 swap_pager_full = 2; 2361 swap_pager_almost_full = 1; 2362 } 2363 if (swdevhd == sp) 2364 swdevhd = NULL; 2365 mtx_unlock(&sw_dev_mtx); 2366 blist_destroy(sp->sw_blist); 2367 free(sp, M_VMPGDATA); 2368 return (0); 2369} 2370 2371void 2372swapoff_all(void) 2373{ 2374 struct swdevt *sp, *spt; 2375 const char *devname; 2376 int error; 2377 2378 mtx_lock(&Giant); 2379 while (swdev_syscall_active) 2380 tsleep(&swdev_syscall_active, PUSER - 1, "swpoff", 0); 2381 swdev_syscall_active = 1; 2382 2383 mtx_lock(&sw_dev_mtx); 2384 TAILQ_FOREACH_SAFE(sp, &swtailq, sw_list, spt) { 2385 mtx_unlock(&sw_dev_mtx); 2386 if (vn_isdisk(sp->sw_vp, NULL)) 2387 devname = devtoname(sp->sw_vp->v_rdev); 2388 else 2389 devname = "[file]"; 2390 error = swapoff_one(sp, thread0.td_ucred); 2391 if (error != 0) { 2392 printf("Cannot remove swap device %s (error=%d), " 2393 "skipping.\n", devname, error); 2394 } else if (bootverbose) { 2395 printf("Swap device %s removed.\n", devname); 2396 } 2397 mtx_lock(&sw_dev_mtx); 2398 } 2399 mtx_unlock(&sw_dev_mtx); 2400 2401 swdev_syscall_active = 0; 2402 wakeup_one(&swdev_syscall_active); 2403 mtx_unlock(&Giant); 2404} 2405 2406void 2407swap_pager_status(int *total, int *used) 2408{ 2409 struct swdevt *sp; 2410 2411 *total = 0; 2412 *used = 0; 2413 mtx_lock(&sw_dev_mtx); 2414 TAILQ_FOREACH(sp, &swtailq, sw_list) { 2415 *total += sp->sw_nblks; 2416 *used += sp->sw_used; 2417 } 2418 mtx_unlock(&sw_dev_mtx); 2419} 2420 2421int 2422swap_dev_info(int name, struct xswdev *xs, char *devname, size_t len) 2423{ 2424 struct swdevt *sp; 2425 const char *tmp_devname; 2426 int error, n; 2427 2428 n = 0; 2429 error = ENOENT; 2430 mtx_lock(&sw_dev_mtx); 2431 TAILQ_FOREACH(sp, &swtailq, sw_list) { 2432 if (n != name) { 2433 n++; 2434 continue; 2435 } 2436 xs->xsw_version = XSWDEV_VERSION; 2437 xs->xsw_dev = sp->sw_dev; 2438 xs->xsw_flags = sp->sw_flags; 2439 xs->xsw_nblks = sp->sw_nblks; 2440 xs->xsw_used = sp->sw_used; 2441 if (devname != NULL) { 2442 if (vn_isdisk(sp->sw_vp, NULL)) 2443 tmp_devname = devtoname(sp->sw_vp->v_rdev); 2444 else 2445 tmp_devname = "[file]"; 2446 strncpy(devname, tmp_devname, len); 2447 } 2448 error = 0; 2449 break; 2450 } 2451 mtx_unlock(&sw_dev_mtx); 2452 return (error); 2453} 2454 2455static int 2456sysctl_vm_swap_info(SYSCTL_HANDLER_ARGS) 2457{ 2458 struct xswdev xs; 2459 int error; 2460 2461 if (arg2 != 1) /* name length */ 2462 return (EINVAL); 2463 error = swap_dev_info(*(int *)arg1, &xs, NULL, 0); 2464 if (error != 0) 2465 return (error); 2466 error = SYSCTL_OUT(req, &xs, sizeof(xs)); 2467 return (error); 2468} 2469 2470SYSCTL_INT(_vm, OID_AUTO, nswapdev, CTLFLAG_RD, &nswapdev, 0, 2471 "Number of swap devices"); 2472SYSCTL_NODE(_vm, OID_AUTO, swap_info, CTLFLAG_RD, sysctl_vm_swap_info, 2473 "Swap statistics by device"); 2474 2475/* 2476 * vmspace_swap_count() - count the approximate swap usage in pages for a 2477 * vmspace. 2478 * 2479 * The map must be locked. 2480 * 2481 * Swap usage is determined by taking the proportional swap used by 2482 * VM objects backing the VM map. To make up for fractional losses, 2483 * if the VM object has any swap use at all the associated map entries 2484 * count for at least 1 swap page. 2485 */ 2486long 2487vmspace_swap_count(struct vmspace *vmspace) 2488{ 2489 vm_map_t map; 2490 vm_map_entry_t cur; 2491 vm_object_t object; 2492 long count, n; 2493 2494 map = &vmspace->vm_map; 2495 count = 0; 2496 2497 for (cur = map->header.next; cur != &map->header; cur = cur->next) { 2498 if ((cur->eflags & MAP_ENTRY_IS_SUB_MAP) == 0 && 2499 (object = cur->object.vm_object) != NULL) { 2500 VM_OBJECT_WLOCK(object); 2501 if (object->type == OBJT_SWAP && 2502 object->un_pager.swp.swp_bcount != 0) { 2503 n = (cur->end - cur->start) / PAGE_SIZE; 2504 count += object->un_pager.swp.swp_bcount * 2505 SWAP_META_PAGES * n / object->size + 1; 2506 } 2507 VM_OBJECT_WUNLOCK(object); 2508 } 2509 } 2510 return (count); 2511} 2512 2513/* 2514 * GEOM backend 2515 * 2516 * Swapping onto disk devices. 2517 * 2518 */ 2519 2520static g_orphan_t swapgeom_orphan; 2521 2522static struct g_class g_swap_class = { 2523 .name = "SWAP", 2524 .version = G_VERSION, 2525 .orphan = swapgeom_orphan, 2526}; 2527 2528DECLARE_GEOM_CLASS(g_swap_class, g_class); 2529 2530 2531static void 2532swapgeom_close_ev(void *arg, int flags) 2533{ 2534 struct g_consumer *cp; 2535 2536 cp = arg; 2537 g_access(cp, -1, -1, 0); 2538 g_detach(cp); 2539 g_destroy_consumer(cp); 2540} 2541 2542static void 2543swapgeom_done(struct bio *bp2) 2544{ 2545 struct swdevt *sp; 2546 struct buf *bp; 2547 struct g_consumer *cp; 2548 int destroy; 2549 2550 bp = bp2->bio_caller2; 2551 cp = bp2->bio_from; 2552 bp->b_ioflags = bp2->bio_flags; 2553 if (bp2->bio_error) 2554 bp->b_ioflags |= BIO_ERROR; 2555 bp->b_resid = bp->b_bcount - bp2->bio_completed; 2556 bp->b_error = bp2->bio_error; 2557 bufdone(bp); 2558 mtx_lock(&sw_dev_mtx); 2559 destroy = ((--cp->index) == 0 && cp->private); 2560 if (destroy) { 2561 sp = bp2->bio_caller1; 2562 sp->sw_id = NULL; 2563 } 2564 mtx_unlock(&sw_dev_mtx); 2565 g_destroy_bio(bp2); 2566 if (destroy) 2567 g_waitfor_event(swapgeom_close_ev, cp, M_WAITOK, NULL); 2568} 2569 2570static void 2571swapgeom_strategy(struct buf *bp, struct swdevt *sp) 2572{ 2573 struct bio *bio; 2574 struct g_consumer *cp; 2575 2576 mtx_lock(&sw_dev_mtx); 2577 cp = sp->sw_id; 2578 if (cp == NULL) { 2579 mtx_unlock(&sw_dev_mtx); 2580 bp->b_error = ENXIO; 2581 bp->b_ioflags |= BIO_ERROR; 2582 bufdone(bp); 2583 return; 2584 } 2585 cp->index++; 2586 mtx_unlock(&sw_dev_mtx); 2587 if (bp->b_iocmd == BIO_WRITE) 2588 bio = g_new_bio(); 2589 else 2590 bio = g_alloc_bio(); 2591 if (bio == NULL) { 2592 bp->b_error = ENOMEM; 2593 bp->b_ioflags |= BIO_ERROR; 2594 bufdone(bp); 2595 return; 2596 } 2597 2598 bio->bio_caller1 = sp; 2599 bio->bio_caller2 = bp; 2600 bio->bio_cmd = bp->b_iocmd; 2601 bio->bio_offset = (bp->b_blkno - sp->sw_first) * PAGE_SIZE; 2602 bio->bio_length = bp->b_bcount; 2603 bio->bio_done = swapgeom_done; 2604 if ((bp->b_flags & B_UNMAPPED) != 0) { 2605 bio->bio_ma = bp->b_pages; 2606 bio->bio_data = unmapped_buf; 2607 bio->bio_ma_offset = (vm_offset_t)bp->b_offset & PAGE_MASK; 2608 bio->bio_ma_n = bp->b_npages; 2609 bio->bio_flags |= BIO_UNMAPPED; 2610 } else { 2611 bio->bio_data = bp->b_data; 2612 bio->bio_ma = NULL; 2613 } 2614 g_io_request(bio, cp); 2615 return; 2616} 2617 2618static void 2619swapgeom_orphan(struct g_consumer *cp) 2620{ 2621 struct swdevt *sp; 2622 int destroy; 2623 2624 mtx_lock(&sw_dev_mtx); 2625 TAILQ_FOREACH(sp, &swtailq, sw_list) { 2626 if (sp->sw_id == cp) { 2627 sp->sw_flags |= SW_CLOSING; 2628 break; 2629 } 2630 } 2631 cp->private = (void *)(uintptr_t)1; 2632 destroy = ((sp != NULL) && (cp->index == 0)); 2633 if (destroy) 2634 sp->sw_id = NULL; 2635 mtx_unlock(&sw_dev_mtx); 2636 if (destroy) 2637 swapgeom_close_ev(cp, 0); 2638} 2639 2640static void 2641swapgeom_close(struct thread *td, struct swdevt *sw) 2642{ 2643 struct g_consumer *cp; 2644 2645 mtx_lock(&sw_dev_mtx); 2646 cp = sw->sw_id; 2647 sw->sw_id = NULL; 2648 mtx_unlock(&sw_dev_mtx); 2649 /* XXX: direct call when Giant untangled */ 2650 if (cp != NULL) 2651 g_waitfor_event(swapgeom_close_ev, cp, M_WAITOK, NULL); 2652} 2653 2654 2655struct swh0h0 { 2656 struct cdev *dev; 2657 struct vnode *vp; 2658 int error; 2659}; 2660 2661static void 2662swapongeom_ev(void *arg, int flags) 2663{ 2664 struct swh0h0 *swh; 2665 struct g_provider *pp; 2666 struct g_consumer *cp; 2667 static struct g_geom *gp; 2668 struct swdevt *sp; 2669 u_long nblks; 2670 int error; 2671 2672 swh = arg; 2673 swh->error = 0; 2674 pp = g_dev_getprovider(swh->dev); 2675 if (pp == NULL) { 2676 swh->error = ENODEV; 2677 return; 2678 } 2679 mtx_lock(&sw_dev_mtx); 2680 TAILQ_FOREACH(sp, &swtailq, sw_list) { 2681 cp = sp->sw_id; 2682 if (cp != NULL && cp->provider == pp) { 2683 mtx_unlock(&sw_dev_mtx); 2684 swh->error = EBUSY; 2685 return; 2686 } 2687 } 2688 mtx_unlock(&sw_dev_mtx); 2689 if (gp == NULL) 2690 gp = g_new_geomf(&g_swap_class, "swap"); 2691 cp = g_new_consumer(gp); 2692 cp->index = 0; /* Number of active I/Os. */ 2693 cp->private = NULL; /* Orphanization flag */ 2694 g_attach(cp, pp); 2695 /* 2696 * XXX: Everytime you think you can improve the margin for 2697 * footshooting, somebody depends on the ability to do so: 2698 * savecore(8) wants to write to our swapdev so we cannot 2699 * set an exclusive count :-( 2700 */ 2701 error = g_access(cp, 1, 1, 0); 2702 if (error) { 2703 g_detach(cp); 2704 g_destroy_consumer(cp); 2705 swh->error = error; 2706 return; 2707 } 2708 nblks = pp->mediasize / DEV_BSIZE; 2709 swaponsomething(swh->vp, cp, nblks, swapgeom_strategy, 2710 swapgeom_close, dev2udev(swh->dev), 2711 (pp->flags & G_PF_ACCEPT_UNMAPPED) != 0 ? SW_UNMAPPED : 0); 2712 swh->error = 0; 2713} 2714 2715static int 2716swapongeom(struct thread *td, struct vnode *vp) 2717{ 2718 int error; 2719 struct swh0h0 swh; 2720 2721 vn_lock(vp, LK_EXCLUSIVE | LK_RETRY); 2722 2723 swh.dev = vp->v_rdev; 2724 swh.vp = vp; 2725 swh.error = 0; 2726 /* XXX: direct call when Giant untangled */ 2727 error = g_waitfor_event(swapongeom_ev, &swh, M_WAITOK, NULL); 2728 if (!error) 2729 error = swh.error; 2730 VOP_UNLOCK(vp, 0); 2731 return (error); 2732} 2733 2734/* 2735 * VNODE backend 2736 * 2737 * This is used mainly for network filesystem (read: probably only tested 2738 * with NFS) swapfiles. 2739 * 2740 */ 2741 2742static void 2743swapdev_strategy(struct buf *bp, struct swdevt *sp) 2744{ 2745 struct vnode *vp2; 2746 2747 bp->b_blkno = ctodb(bp->b_blkno - sp->sw_first); 2748 2749 vp2 = sp->sw_id; 2750 vhold(vp2); 2751 if (bp->b_iocmd == BIO_WRITE) { 2752 if (bp->b_bufobj) 2753 bufobj_wdrop(bp->b_bufobj); 2754 bufobj_wref(&vp2->v_bufobj); 2755 } 2756 if (bp->b_bufobj != &vp2->v_bufobj) 2757 bp->b_bufobj = &vp2->v_bufobj; 2758 bp->b_vp = vp2; 2759 bp->b_iooffset = dbtob(bp->b_blkno); 2760 bstrategy(bp); 2761 return; 2762} 2763 2764static void 2765swapdev_close(struct thread *td, struct swdevt *sp) 2766{ 2767 2768 VOP_CLOSE(sp->sw_vp, FREAD | FWRITE, td->td_ucred, td); 2769 vrele(sp->sw_vp); 2770} 2771 2772 2773static int 2774swaponvp(struct thread *td, struct vnode *vp, u_long nblks) 2775{ 2776 struct swdevt *sp; 2777 int error; 2778 2779 if (nblks == 0) 2780 return (ENXIO); 2781 mtx_lock(&sw_dev_mtx); 2782 TAILQ_FOREACH(sp, &swtailq, sw_list) { 2783 if (sp->sw_id == vp) { 2784 mtx_unlock(&sw_dev_mtx); 2785 return (EBUSY); 2786 } 2787 } 2788 mtx_unlock(&sw_dev_mtx); 2789 2790 (void) vn_lock(vp, LK_EXCLUSIVE | LK_RETRY); 2791#ifdef MAC 2792 error = mac_system_check_swapon(td->td_ucred, vp); 2793 if (error == 0) 2794#endif 2795 error = VOP_OPEN(vp, FREAD | FWRITE, td->td_ucred, td, NULL); 2796 (void) VOP_UNLOCK(vp, 0); 2797 if (error) 2798 return (error); 2799 2800 swaponsomething(vp, vp, nblks, swapdev_strategy, swapdev_close, 2801 NODEV, 0); 2802 return (0); 2803} 2804