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 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 struct sx swdev_syscall_lock; /* 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 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 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} 315 316#define SWM_FREE 0x02 /* free, period */ 317#define SWM_POP 0x04 /* pop out */ 318 319int swap_pager_full = 2; /* swap space exhaustion (task killing) */ 320static int swap_pager_almost_full = 1; /* swap space exhaustion (w/hysteresis)*/ 321static int nsw_rcount; /* free read buffers */ 322static int nsw_wcount_sync; /* limit write buffers / synchronous */ 323static int nsw_wcount_async; /* limit write buffers / asynchronous */ 324static int nsw_wcount_async_max;/* assigned maximum */ 325static int nsw_cluster_max; /* maximum VOP I/O allowed */ 326 327static int sysctl_swap_async_max(SYSCTL_HANDLER_ARGS); 328SYSCTL_PROC(_vm, OID_AUTO, swap_async_max, CTLTYPE_INT | CTLFLAG_RW | 329 CTLFLAG_MPSAFE, NULL, 0, sysctl_swap_async_max, "I", 330 "Maximum running async swap ops"); 331 332static struct swblock **swhash; 333static int swhash_mask; 334static struct mtx swhash_mtx; 335 336static struct sx sw_alloc_sx; 337 338/* 339 * "named" and "unnamed" anon region objects. Try to reduce the overhead 340 * of searching a named list by hashing it just a little. 341 */ 342 343#define NOBJLISTS 8 344 345#define NOBJLIST(handle) \ 346 (&swap_pager_object_list[((int)(intptr_t)handle >> 4) & (NOBJLISTS-1)]) 347 348static struct pagerlst swap_pager_object_list[NOBJLISTS]; 349static uma_zone_t swap_zone; 350 351/* 352 * pagerops for OBJT_SWAP - "swap pager". Some ops are also global procedure 353 * calls hooked from other parts of the VM system and do not appear here. 354 * (see vm/swap_pager.h). 355 */ 356static vm_object_t 357 swap_pager_alloc(void *handle, vm_ooffset_t size, 358 vm_prot_t prot, vm_ooffset_t offset, struct ucred *); 359static void swap_pager_dealloc(vm_object_t object); 360static int swap_pager_getpages(vm_object_t, vm_page_t *, int, int *, 361 int *); 362static int swap_pager_getpages_async(vm_object_t, vm_page_t *, int, int *, 363 int *, pgo_getpages_iodone_t, void *); 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_getpages_async = swap_pager_getpages_async, /* pagein (async) */ 377 .pgo_putpages = swap_pager_putpages, /* pageout */ 378 .pgo_haspage = swap_pager_haspage, /* get backing store status for page */ 379 .pgo_pageunswapped = swap_pager_unswapped, /* remove swap related to page */ 380}; 381 382/* 383 * dmmax is in page-sized chunks with the new swap system. It was 384 * dev-bsized chunks in the old. dmmax is always a power of 2. 385 * 386 * swap_*() routines are externally accessible. swp_*() routines are 387 * internal. 388 */ 389static int dmmax; 390static int nswap_lowat = 128; /* in pages, swap_pager_almost_full warn */ 391static int nswap_hiwat = 512; /* in pages, swap_pager_almost_full warn */ 392 393SYSCTL_INT(_vm, OID_AUTO, dmmax, CTLFLAG_RD, &dmmax, 0, 394 "Maximum size of a swap block"); 395 396static void swp_sizecheck(void); 397static void swp_pager_async_iodone(struct buf *bp); 398static int swapongeom(struct vnode *); 399static int swaponvp(struct thread *, struct vnode *, u_long); 400static int swapoff_one(struct swdevt *sp, struct ucred *cred); 401 402/* 403 * Swap bitmap functions 404 */ 405static void swp_pager_freeswapspace(daddr_t blk, int npages); 406static daddr_t swp_pager_getswapspace(int npages); 407 408/* 409 * Metadata functions 410 */ 411static struct swblock **swp_pager_hash(vm_object_t object, vm_pindex_t index); 412static void swp_pager_meta_build(vm_object_t, vm_pindex_t, daddr_t); 413static void swp_pager_meta_free(vm_object_t, vm_pindex_t, daddr_t); 414static void swp_pager_meta_free_all(vm_object_t); 415static daddr_t swp_pager_meta_ctl(vm_object_t, vm_pindex_t, int); 416 417/* 418 * SWP_SIZECHECK() - update swap_pager_full indication 419 * 420 * update the swap_pager_almost_full indication and warn when we are 421 * about to run out of swap space, using lowat/hiwat hysteresis. 422 * 423 * Clear swap_pager_full ( task killing ) indication when lowat is met. 424 * 425 * No restrictions on call 426 * This routine may not block. 427 */ 428static void 429swp_sizecheck(void) 430{ 431 432 if (swap_pager_avail < nswap_lowat) { 433 if (swap_pager_almost_full == 0) { 434 printf("swap_pager: out of swap space\n"); 435 swap_pager_almost_full = 1; 436 } 437 } else { 438 swap_pager_full = 0; 439 if (swap_pager_avail > nswap_hiwat) 440 swap_pager_almost_full = 0; 441 } 442} 443 444/* 445 * SWP_PAGER_HASH() - hash swap meta data 446 * 447 * This is an helper function which hashes the swapblk given 448 * the object and page index. It returns a pointer to a pointer 449 * to the object, or a pointer to a NULL pointer if it could not 450 * find a swapblk. 451 */ 452static struct swblock ** 453swp_pager_hash(vm_object_t object, vm_pindex_t index) 454{ 455 struct swblock **pswap; 456 struct swblock *swap; 457 458 index &= ~(vm_pindex_t)SWAP_META_MASK; 459 pswap = &swhash[(index ^ (int)(intptr_t)object) & swhash_mask]; 460 while ((swap = *pswap) != NULL) { 461 if (swap->swb_object == object && 462 swap->swb_index == index 463 ) { 464 break; 465 } 466 pswap = &swap->swb_hnext; 467 } 468 return (pswap); 469} 470 471/* 472 * SWAP_PAGER_INIT() - initialize the swap pager! 473 * 474 * Expected to be started from system init. NOTE: This code is run 475 * before much else so be careful what you depend on. Most of the VM 476 * system has yet to be initialized at this point. 477 */ 478static void 479swap_pager_init(void) 480{ 481 /* 482 * Initialize object lists 483 */ 484 int i; 485 486 for (i = 0; i < NOBJLISTS; ++i) 487 TAILQ_INIT(&swap_pager_object_list[i]); 488 mtx_init(&sw_dev_mtx, "swapdev", NULL, MTX_DEF); 489 sx_init(&sw_alloc_sx, "swspsx"); 490 sx_init(&swdev_syscall_lock, "swsysc"); 491 492 /* 493 * Device Stripe, in PAGE_SIZE'd blocks 494 */ 495 dmmax = SWB_NPAGES * 2; 496} 497 498/* 499 * SWAP_PAGER_SWAP_INIT() - swap pager initialization from pageout process 500 * 501 * Expected to be started from pageout process once, prior to entering 502 * its main loop. 503 */ 504void 505swap_pager_swap_init(void) 506{ 507 unsigned long n, n2; 508 509 /* 510 * Number of in-transit swap bp operations. Don't 511 * exhaust the pbufs completely. Make sure we 512 * initialize workable values (0 will work for hysteresis 513 * but it isn't very efficient). 514 * 515 * The nsw_cluster_max is constrained by the bp->b_pages[] 516 * array (MAXPHYS/PAGE_SIZE) and our locally defined 517 * MAX_PAGEOUT_CLUSTER. Also be aware that swap ops are 518 * constrained by the swap device interleave stripe size. 519 * 520 * Currently we hardwire nsw_wcount_async to 4. This limit is 521 * designed to prevent other I/O from having high latencies due to 522 * our pageout I/O. The value 4 works well for one or two active swap 523 * devices but is probably a little low if you have more. Even so, 524 * a higher value would probably generate only a limited improvement 525 * with three or four active swap devices since the system does not 526 * typically have to pageout at extreme bandwidths. We will want 527 * at least 2 per swap devices, and 4 is a pretty good value if you 528 * have one NFS swap device due to the command/ack latency over NFS. 529 * So it all works out pretty well. 530 */ 531 nsw_cluster_max = min((MAXPHYS/PAGE_SIZE), MAX_PAGEOUT_CLUSTER); 532 533 mtx_lock(&pbuf_mtx); 534 nsw_rcount = (nswbuf + 1) / 2; 535 nsw_wcount_sync = (nswbuf + 3) / 4; 536 nsw_wcount_async = 4; 537 nsw_wcount_async_max = nsw_wcount_async; 538 mtx_unlock(&pbuf_mtx); 539 540 /* 541 * Initialize our zone. Right now I'm just guessing on the number 542 * we need based on the number of pages in the system. Each swblock 543 * can hold 32 pages, so this is probably overkill. This reservation 544 * is typically limited to around 32MB by default. 545 */ 546 n = vm_cnt.v_page_count / 2; 547 if (maxswzone && n > maxswzone / sizeof(struct swblock)) 548 n = maxswzone / sizeof(struct swblock); 549 n2 = n; 550 swap_zone = uma_zcreate("SWAPMETA", sizeof(struct swblock), NULL, NULL, 551 NULL, NULL, UMA_ALIGN_PTR, UMA_ZONE_NOFREE | UMA_ZONE_VM); 552 if (swap_zone == NULL) 553 panic("failed to create swap_zone."); 554 do { 555 if (uma_zone_reserve_kva(swap_zone, n)) 556 break; 557 /* 558 * if the allocation failed, try a zone two thirds the 559 * size of the previous attempt. 560 */ 561 n -= ((n + 2) / 3); 562 } while (n > 0); 563 if (n2 != n) 564 printf("Swap zone entries reduced from %lu to %lu.\n", n2, n); 565 swap_maxpages = n * SWAP_META_PAGES; 566 swzone = n * sizeof(struct swblock); 567 n2 = n; 568 569 /* 570 * Initialize our meta-data hash table. The swapper does not need to 571 * be quite as efficient as the VM system, so we do not use an 572 * oversized hash table. 573 * 574 * n: size of hash table, must be power of 2 575 * swhash_mask: hash table index mask 576 */ 577 for (n = 1; n < n2 / 8; n *= 2) 578 ; 579 swhash = malloc(sizeof(struct swblock *) * n, M_VMPGDATA, M_WAITOK | M_ZERO); 580 swhash_mask = n - 1; 581 mtx_init(&swhash_mtx, "swap_pager swhash", NULL, MTX_DEF); 582} 583 584static vm_object_t 585swap_pager_alloc_init(void *handle, struct ucred *cred, vm_ooffset_t size, 586 vm_ooffset_t offset) 587{ 588 vm_object_t object; 589 590 if (cred != NULL) { 591 if (!swap_reserve_by_cred(size, cred)) 592 return (NULL); 593 crhold(cred); 594 } 595 object = vm_object_allocate(OBJT_SWAP, OFF_TO_IDX(offset + 596 PAGE_MASK + size)); 597 object->handle = handle; 598 if (cred != NULL) { 599 object->cred = cred; 600 object->charge = size; 601 } 602 object->un_pager.swp.swp_bcount = 0; 603 return (object); 604} 605 606/* 607 * SWAP_PAGER_ALLOC() - allocate a new OBJT_SWAP VM object and instantiate 608 * its metadata structures. 609 * 610 * This routine is called from the mmap and fork code to create a new 611 * OBJT_SWAP object. 612 * 613 * This routine must ensure that no live duplicate is created for 614 * the named object request, which is protected against by 615 * holding the sw_alloc_sx lock in case handle != NULL. 616 */ 617static vm_object_t 618swap_pager_alloc(void *handle, vm_ooffset_t size, vm_prot_t prot, 619 vm_ooffset_t offset, struct ucred *cred) 620{ 621 vm_object_t object; 622 623 if (handle != NULL) { 624 /* 625 * Reference existing named region or allocate new one. There 626 * should not be a race here against swp_pager_meta_build() 627 * as called from vm_page_remove() in regards to the lookup 628 * of the handle. 629 */ 630 sx_xlock(&sw_alloc_sx); 631 object = vm_pager_object_lookup(NOBJLIST(handle), handle); 632 if (object == NULL) { 633 object = swap_pager_alloc_init(handle, cred, size, 634 offset); 635 if (object != NULL) { 636 TAILQ_INSERT_TAIL(NOBJLIST(object->handle), 637 object, pager_object_list); 638 } 639 } 640 sx_xunlock(&sw_alloc_sx); 641 } else { 642 object = swap_pager_alloc_init(handle, cred, size, offset); 643 } 644 return (object); 645} 646 647/* 648 * SWAP_PAGER_DEALLOC() - remove swap metadata from object 649 * 650 * The swap backing for the object is destroyed. The code is 651 * designed such that we can reinstantiate it later, but this 652 * routine is typically called only when the entire object is 653 * about to be destroyed. 654 * 655 * The object must be locked. 656 */ 657static void 658swap_pager_dealloc(vm_object_t object) 659{ 660 661 VM_OBJECT_ASSERT_WLOCKED(object); 662 KASSERT((object->flags & OBJ_DEAD) != 0, ("dealloc of reachable obj")); 663 664 /* 665 * Remove from list right away so lookups will fail if we block for 666 * pageout completion. 667 */ 668 if (object->handle != NULL) { 669 VM_OBJECT_WUNLOCK(object); 670 sx_xlock(&sw_alloc_sx); 671 TAILQ_REMOVE(NOBJLIST(object->handle), object, 672 pager_object_list); 673 sx_xunlock(&sw_alloc_sx); 674 VM_OBJECT_WLOCK(object); 675 } 676 677 vm_object_pip_wait(object, "swpdea"); 678 679 /* 680 * Free all remaining metadata. We only bother to free it from 681 * the swap meta data. We do not attempt to free swapblk's still 682 * associated with vm_page_t's for this object. We do not care 683 * if paging is still in progress on some objects. 684 */ 685 swp_pager_meta_free_all(object); 686 object->handle = NULL; 687 object->type = OBJT_DEAD; 688} 689 690/************************************************************************ 691 * SWAP PAGER BITMAP ROUTINES * 692 ************************************************************************/ 693 694/* 695 * SWP_PAGER_GETSWAPSPACE() - allocate raw swap space 696 * 697 * Allocate swap for the requested number of pages. The starting 698 * swap block number (a page index) is returned or SWAPBLK_NONE 699 * if the allocation failed. 700 * 701 * Also has the side effect of advising that somebody made a mistake 702 * when they configured swap and didn't configure enough. 703 * 704 * This routine may not sleep. 705 * 706 * We allocate in round-robin fashion from the configured devices. 707 */ 708static daddr_t 709swp_pager_getswapspace(int npages) 710{ 711 daddr_t blk; 712 struct swdevt *sp; 713 int i; 714 715 blk = SWAPBLK_NONE; 716 mtx_lock(&sw_dev_mtx); 717 sp = swdevhd; 718 for (i = 0; i < nswapdev; i++) { 719 if (sp == NULL) 720 sp = TAILQ_FIRST(&swtailq); 721 if (!(sp->sw_flags & SW_CLOSING)) { 722 blk = blist_alloc(sp->sw_blist, npages); 723 if (blk != SWAPBLK_NONE) { 724 blk += sp->sw_first; 725 sp->sw_used += npages; 726 swap_pager_avail -= npages; 727 swp_sizecheck(); 728 swdevhd = TAILQ_NEXT(sp, sw_list); 729 goto done; 730 } 731 } 732 sp = TAILQ_NEXT(sp, sw_list); 733 } 734 if (swap_pager_full != 2) { 735 printf("swap_pager_getswapspace(%d): failed\n", npages); 736 swap_pager_full = 2; 737 swap_pager_almost_full = 1; 738 } 739 swdevhd = NULL; 740done: 741 mtx_unlock(&sw_dev_mtx); 742 return (blk); 743} 744 745static int 746swp_pager_isondev(daddr_t blk, struct swdevt *sp) 747{ 748 749 return (blk >= sp->sw_first && blk < sp->sw_end); 750} 751 752static void 753swp_pager_strategy(struct buf *bp) 754{ 755 struct swdevt *sp; 756 757 mtx_lock(&sw_dev_mtx); 758 TAILQ_FOREACH(sp, &swtailq, sw_list) { 759 if (bp->b_blkno >= sp->sw_first && bp->b_blkno < sp->sw_end) { 760 mtx_unlock(&sw_dev_mtx); 761 if ((sp->sw_flags & SW_UNMAPPED) != 0 && 762 unmapped_buf_allowed) { 763 bp->b_data = unmapped_buf; 764 bp->b_offset = 0; 765 } else { 766 pmap_qenter((vm_offset_t)bp->b_data, 767 &bp->b_pages[0], bp->b_bcount / PAGE_SIZE); 768 } 769 sp->sw_strategy(bp, sp); 770 return; 771 } 772 } 773 panic("Swapdev not found"); 774} 775 776 777/* 778 * SWP_PAGER_FREESWAPSPACE() - free raw swap space 779 * 780 * This routine returns the specified swap blocks back to the bitmap. 781 * 782 * This routine may not sleep. 783 */ 784static void 785swp_pager_freeswapspace(daddr_t blk, int npages) 786{ 787 struct swdevt *sp; 788 789 mtx_lock(&sw_dev_mtx); 790 TAILQ_FOREACH(sp, &swtailq, sw_list) { 791 if (blk >= sp->sw_first && blk < sp->sw_end) { 792 sp->sw_used -= npages; 793 /* 794 * If we are attempting to stop swapping on 795 * this device, we don't want to mark any 796 * blocks free lest they be reused. 797 */ 798 if ((sp->sw_flags & SW_CLOSING) == 0) { 799 blist_free(sp->sw_blist, blk - sp->sw_first, 800 npages); 801 swap_pager_avail += npages; 802 swp_sizecheck(); 803 } 804 mtx_unlock(&sw_dev_mtx); 805 return; 806 } 807 } 808 panic("Swapdev not found"); 809} 810 811/* 812 * SWAP_PAGER_FREESPACE() - frees swap blocks associated with a page 813 * range within an object. 814 * 815 * This is a globally accessible routine. 816 * 817 * This routine removes swapblk assignments from swap metadata. 818 * 819 * The external callers of this routine typically have already destroyed 820 * or renamed vm_page_t's associated with this range in the object so 821 * we should be ok. 822 * 823 * The object must be locked. 824 */ 825void 826swap_pager_freespace(vm_object_t object, vm_pindex_t start, vm_size_t size) 827{ 828 829 swp_pager_meta_free(object, start, size); 830} 831 832/* 833 * SWAP_PAGER_RESERVE() - reserve swap blocks in object 834 * 835 * Assigns swap blocks to the specified range within the object. The 836 * swap blocks are not zeroed. Any previous swap assignment is destroyed. 837 * 838 * Returns 0 on success, -1 on failure. 839 */ 840int 841swap_pager_reserve(vm_object_t object, vm_pindex_t start, vm_size_t size) 842{ 843 int n = 0; 844 daddr_t blk = SWAPBLK_NONE; 845 vm_pindex_t beg = start; /* save start index */ 846 847 VM_OBJECT_WLOCK(object); 848 while (size) { 849 if (n == 0) { 850 n = BLIST_MAX_ALLOC; 851 while ((blk = swp_pager_getswapspace(n)) == SWAPBLK_NONE) { 852 n >>= 1; 853 if (n == 0) { 854 swp_pager_meta_free(object, beg, start - beg); 855 VM_OBJECT_WUNLOCK(object); 856 return (-1); 857 } 858 } 859 } 860 swp_pager_meta_build(object, start, blk); 861 --size; 862 ++start; 863 ++blk; 864 --n; 865 } 866 swp_pager_meta_free(object, start, n); 867 VM_OBJECT_WUNLOCK(object); 868 return (0); 869} 870 871/* 872 * SWAP_PAGER_COPY() - copy blocks from source pager to destination pager 873 * and destroy the source. 874 * 875 * Copy any valid swapblks from the source to the destination. In 876 * cases where both the source and destination have a valid swapblk, 877 * we keep the destination's. 878 * 879 * This routine is allowed to sleep. It may sleep allocating metadata 880 * indirectly through swp_pager_meta_build() or if paging is still in 881 * progress on the source. 882 * 883 * The source object contains no vm_page_t's (which is just as well) 884 * 885 * The source object is of type OBJT_SWAP. 886 * 887 * The source and destination objects must be locked. 888 * Both object locks may temporarily be released. 889 */ 890void 891swap_pager_copy(vm_object_t srcobject, vm_object_t dstobject, 892 vm_pindex_t offset, int destroysource) 893{ 894 vm_pindex_t i; 895 896 VM_OBJECT_ASSERT_WLOCKED(srcobject); 897 VM_OBJECT_ASSERT_WLOCKED(dstobject); 898 899 /* 900 * If destroysource is set, we remove the source object from the 901 * swap_pager internal queue now. 902 */ 903 if (destroysource && srcobject->handle != NULL) { 904 vm_object_pip_add(srcobject, 1); 905 VM_OBJECT_WUNLOCK(srcobject); 906 vm_object_pip_add(dstobject, 1); 907 VM_OBJECT_WUNLOCK(dstobject); 908 sx_xlock(&sw_alloc_sx); 909 TAILQ_REMOVE(NOBJLIST(srcobject->handle), srcobject, 910 pager_object_list); 911 sx_xunlock(&sw_alloc_sx); 912 VM_OBJECT_WLOCK(dstobject); 913 vm_object_pip_wakeup(dstobject); 914 VM_OBJECT_WLOCK(srcobject); 915 vm_object_pip_wakeup(srcobject); 916 } 917 918 /* 919 * transfer source to destination. 920 */ 921 for (i = 0; i < dstobject->size; ++i) { 922 daddr_t dstaddr; 923 924 /* 925 * Locate (without changing) the swapblk on the destination, 926 * unless it is invalid in which case free it silently, or 927 * if the destination is a resident page, in which case the 928 * source is thrown away. 929 */ 930 dstaddr = swp_pager_meta_ctl(dstobject, i, 0); 931 932 if (dstaddr == SWAPBLK_NONE) { 933 /* 934 * Destination has no swapblk and is not resident, 935 * copy source. 936 */ 937 daddr_t srcaddr; 938 939 srcaddr = swp_pager_meta_ctl( 940 srcobject, 941 i + offset, 942 SWM_POP 943 ); 944 945 if (srcaddr != SWAPBLK_NONE) { 946 /* 947 * swp_pager_meta_build() can sleep. 948 */ 949 vm_object_pip_add(srcobject, 1); 950 VM_OBJECT_WUNLOCK(srcobject); 951 vm_object_pip_add(dstobject, 1); 952 swp_pager_meta_build(dstobject, i, srcaddr); 953 vm_object_pip_wakeup(dstobject); 954 VM_OBJECT_WLOCK(srcobject); 955 vm_object_pip_wakeup(srcobject); 956 } 957 } else { 958 /* 959 * Destination has valid swapblk or it is represented 960 * by a resident page. We destroy the sourceblock. 961 */ 962 963 swp_pager_meta_ctl(srcobject, i + offset, SWM_FREE); 964 } 965 } 966 967 /* 968 * Free left over swap blocks in source. 969 * 970 * We have to revert the type to OBJT_DEFAULT so we do not accidentally 971 * double-remove the object from the swap queues. 972 */ 973 if (destroysource) { 974 swp_pager_meta_free_all(srcobject); 975 /* 976 * Reverting the type is not necessary, the caller is going 977 * to destroy srcobject directly, but I'm doing it here 978 * for consistency since we've removed the object from its 979 * queues. 980 */ 981 srcobject->type = OBJT_DEFAULT; 982 } 983} 984 985/* 986 * SWAP_PAGER_HASPAGE() - determine if we have good backing store for 987 * the requested page. 988 * 989 * We determine whether good backing store exists for the requested 990 * page and return TRUE if it does, FALSE if it doesn't. 991 * 992 * If TRUE, we also try to determine how much valid, contiguous backing 993 * store exists before and after the requested page. 994 */ 995static boolean_t 996swap_pager_haspage(vm_object_t object, vm_pindex_t pindex, int *before, 997 int *after) 998{ 999 daddr_t blk, blk0; 1000 int i; 1001 1002 VM_OBJECT_ASSERT_LOCKED(object); 1003 1004 /* 1005 * do we have good backing store at the requested index ? 1006 */ 1007 blk0 = swp_pager_meta_ctl(object, pindex, 0); 1008 if (blk0 == SWAPBLK_NONE) { 1009 if (before) 1010 *before = 0; 1011 if (after) 1012 *after = 0; 1013 return (FALSE); 1014 } 1015 1016 /* 1017 * find backwards-looking contiguous good backing store 1018 */ 1019 if (before != NULL) { 1020 for (i = 1; i < SWB_NPAGES; i++) { 1021 if (i > pindex) 1022 break; 1023 blk = swp_pager_meta_ctl(object, pindex - i, 0); 1024 if (blk != blk0 - i) 1025 break; 1026 } 1027 *before = i - 1; 1028 } 1029 1030 /* 1031 * find forward-looking contiguous good backing store 1032 */ 1033 if (after != NULL) { 1034 for (i = 1; i < SWB_NPAGES; i++) { 1035 blk = swp_pager_meta_ctl(object, pindex + i, 0); 1036 if (blk != blk0 + i) 1037 break; 1038 } 1039 *after = i - 1; 1040 } 1041 return (TRUE); 1042} 1043 1044/* 1045 * SWAP_PAGER_PAGE_UNSWAPPED() - remove swap backing store related to page 1046 * 1047 * This removes any associated swap backing store, whether valid or 1048 * not, from the page. 1049 * 1050 * This routine is typically called when a page is made dirty, at 1051 * which point any associated swap can be freed. MADV_FREE also 1052 * calls us in a special-case situation 1053 * 1054 * NOTE!!! If the page is clean and the swap was valid, the caller 1055 * should make the page dirty before calling this routine. This routine 1056 * does NOT change the m->dirty status of the page. Also: MADV_FREE 1057 * depends on it. 1058 * 1059 * This routine may not sleep. 1060 * 1061 * The object containing the page must be locked. 1062 */ 1063static void 1064swap_pager_unswapped(vm_page_t m) 1065{ 1066 1067 swp_pager_meta_ctl(m->object, m->pindex, SWM_FREE); 1068} 1069 1070/* 1071 * swap_pager_getpages() - bring pages in from swap 1072 * 1073 * Attempt to page in the pages in array "m" of length "count". The caller 1074 * may optionally specify that additional pages preceding and succeeding 1075 * the specified range be paged in. The number of such pages is returned 1076 * in the "rbehind" and "rahead" parameters, and they will be in the 1077 * inactive queue upon return. 1078 * 1079 * The pages in "m" must be busied and will remain busied upon return. 1080 */ 1081static int 1082swap_pager_getpages(vm_object_t object, vm_page_t *m, int count, int *rbehind, 1083 int *rahead) 1084{ 1085 struct buf *bp; 1086 vm_page_t mpred, msucc, p; 1087 vm_pindex_t pindex; 1088 daddr_t blk; 1089 int i, j, maxahead, maxbehind, reqcount, shift; 1090 1091 reqcount = count; 1092 1093 VM_OBJECT_WUNLOCK(object); 1094 bp = getpbuf(&nsw_rcount); 1095 VM_OBJECT_WLOCK(object); 1096 1097 if (!swap_pager_haspage(object, m[0]->pindex, &maxbehind, &maxahead)) { 1098 relpbuf(bp, &nsw_rcount); 1099 return (VM_PAGER_FAIL); 1100 } 1101 1102 /* 1103 * Clip the readahead and readbehind ranges to exclude resident pages. 1104 */ 1105 if (rahead != NULL) { 1106 KASSERT(reqcount - 1 <= maxahead, 1107 ("page count %d extends beyond swap block", reqcount)); 1108 *rahead = imin(*rahead, maxahead - (reqcount - 1)); 1109 pindex = m[reqcount - 1]->pindex; 1110 msucc = TAILQ_NEXT(m[reqcount - 1], listq); 1111 if (msucc != NULL && msucc->pindex - pindex - 1 < *rahead) 1112 *rahead = msucc->pindex - pindex - 1; 1113 } 1114 if (rbehind != NULL) { 1115 *rbehind = imin(*rbehind, maxbehind); 1116 pindex = m[0]->pindex; 1117 mpred = TAILQ_PREV(m[0], pglist, listq); 1118 if (mpred != NULL && pindex - mpred->pindex - 1 < *rbehind) 1119 *rbehind = pindex - mpred->pindex - 1; 1120 } 1121 1122 /* 1123 * Allocate readahead and readbehind pages. 1124 */ 1125 shift = rbehind != NULL ? *rbehind : 0; 1126 if (shift != 0) { 1127 for (i = 1; i <= shift; i++) { 1128 p = vm_page_alloc(object, m[0]->pindex - i, 1129 VM_ALLOC_NORMAL | VM_ALLOC_IFNOTCACHED); 1130 if (p == NULL) { 1131 /* Shift allocated pages to the left. */ 1132 for (j = 0; j < i - 1; j++) 1133 bp->b_pages[j] = 1134 bp->b_pages[j + shift - i + 1]; 1135 break; 1136 } 1137 bp->b_pages[shift - i] = p; 1138 } 1139 shift = i - 1; 1140 *rbehind = shift; 1141 } 1142 for (i = 0; i < reqcount; i++) 1143 bp->b_pages[i + shift] = m[i]; 1144 if (rahead != NULL) { 1145 for (i = 0; i < *rahead; i++) { 1146 p = vm_page_alloc(object, 1147 m[reqcount - 1]->pindex + i + 1, 1148 VM_ALLOC_NORMAL | VM_ALLOC_IFNOTCACHED); 1149 if (p == NULL) 1150 break; 1151 bp->b_pages[shift + reqcount + i] = p; 1152 } 1153 *rahead = i; 1154 } 1155 if (rbehind != NULL) 1156 count += *rbehind; 1157 if (rahead != NULL) 1158 count += *rahead; 1159 1160 vm_object_pip_add(object, count); 1161 1162 for (i = 0; i < count; i++) 1163 bp->b_pages[i]->oflags |= VPO_SWAPINPROG; 1164 1165 pindex = bp->b_pages[0]->pindex; 1166 blk = swp_pager_meta_ctl(object, pindex, 0); 1167 KASSERT(blk != SWAPBLK_NONE, 1168 ("no swap blocking containing %p(%jx)", object, (uintmax_t)pindex)); 1169 1170 VM_OBJECT_WUNLOCK(object); 1171 1172 bp->b_flags |= B_PAGING; 1173 bp->b_iocmd = BIO_READ; 1174 bp->b_iodone = swp_pager_async_iodone; 1175 bp->b_rcred = crhold(thread0.td_ucred); 1176 bp->b_wcred = crhold(thread0.td_ucred); 1177 bp->b_blkno = blk; 1178 bp->b_bcount = PAGE_SIZE * count; 1179 bp->b_bufsize = PAGE_SIZE * count; 1180 bp->b_npages = count; 1181 bp->b_pgbefore = rbehind != NULL ? *rbehind : 0; 1182 bp->b_pgafter = rahead != NULL ? *rahead : 0; 1183 1184 PCPU_INC(cnt.v_swapin); 1185 PCPU_ADD(cnt.v_swappgsin, count); 1186 1187 /* 1188 * perform the I/O. NOTE!!! bp cannot be considered valid after 1189 * this point because we automatically release it on completion. 1190 * Instead, we look at the one page we are interested in which we 1191 * still hold a lock on even through the I/O completion. 1192 * 1193 * The other pages in our m[] array are also released on completion, 1194 * so we cannot assume they are valid anymore either. 1195 * 1196 * NOTE: b_blkno is destroyed by the call to swapdev_strategy 1197 */ 1198 BUF_KERNPROC(bp); 1199 swp_pager_strategy(bp); 1200 1201 /* 1202 * Wait for the pages we want to complete. VPO_SWAPINPROG is always 1203 * cleared on completion. If an I/O error occurs, SWAPBLK_NONE 1204 * is set in the metadata for each page in the request. 1205 */ 1206 VM_OBJECT_WLOCK(object); 1207 while ((m[0]->oflags & VPO_SWAPINPROG) != 0) { 1208 m[0]->oflags |= VPO_SWAPSLEEP; 1209 PCPU_INC(cnt.v_intrans); 1210 if (VM_OBJECT_SLEEP(object, &object->paging_in_progress, PSWP, 1211 "swread", hz * 20)) { 1212 printf( 1213"swap_pager: indefinite wait buffer: bufobj: %p, blkno: %jd, size: %ld\n", 1214 bp->b_bufobj, (intmax_t)bp->b_blkno, bp->b_bcount); 1215 } 1216 } 1217 1218 /* 1219 * If we had an unrecoverable read error pages will not be valid. 1220 */ 1221 for (i = 0; i < reqcount; i++) 1222 if (m[i]->valid != VM_PAGE_BITS_ALL) 1223 return (VM_PAGER_ERROR); 1224 1225 return (VM_PAGER_OK); 1226 1227 /* 1228 * A final note: in a low swap situation, we cannot deallocate swap 1229 * and mark a page dirty here because the caller is likely to mark 1230 * the page clean when we return, causing the page to possibly revert 1231 * to all-zero's later. 1232 */ 1233} 1234 1235/* 1236 * swap_pager_getpages_async(): 1237 * 1238 * Right now this is emulation of asynchronous operation on top of 1239 * swap_pager_getpages(). 1240 */ 1241static int 1242swap_pager_getpages_async(vm_object_t object, vm_page_t *m, int count, 1243 int *rbehind, int *rahead, pgo_getpages_iodone_t iodone, void *arg) 1244{ 1245 int r, error; 1246 1247 r = swap_pager_getpages(object, m, count, rbehind, rahead); 1248 VM_OBJECT_WUNLOCK(object); 1249 switch (r) { 1250 case VM_PAGER_OK: 1251 error = 0; 1252 break; 1253 case VM_PAGER_ERROR: 1254 error = EIO; 1255 break; 1256 case VM_PAGER_FAIL: 1257 error = EINVAL; 1258 break; 1259 default: 1260 panic("unhandled swap_pager_getpages() error %d", r); 1261 } 1262 (iodone)(arg, m, count, error); 1263 VM_OBJECT_WLOCK(object); 1264 1265 return (r); 1266} 1267 1268/* 1269 * swap_pager_putpages: 1270 * 1271 * Assign swap (if necessary) and initiate I/O on the specified pages. 1272 * 1273 * We support both OBJT_DEFAULT and OBJT_SWAP objects. DEFAULT objects 1274 * are automatically converted to SWAP objects. 1275 * 1276 * In a low memory situation we may block in VOP_STRATEGY(), but the new 1277 * vm_page reservation system coupled with properly written VFS devices 1278 * should ensure that no low-memory deadlock occurs. This is an area 1279 * which needs work. 1280 * 1281 * The parent has N vm_object_pip_add() references prior to 1282 * calling us and will remove references for rtvals[] that are 1283 * not set to VM_PAGER_PEND. We need to remove the rest on I/O 1284 * completion. 1285 * 1286 * The parent has soft-busy'd the pages it passes us and will unbusy 1287 * those whos rtvals[] entry is not set to VM_PAGER_PEND on return. 1288 * We need to unbusy the rest on I/O completion. 1289 */ 1290static void 1291swap_pager_putpages(vm_object_t object, vm_page_t *m, int count, 1292 int flags, int *rtvals) 1293{ 1294 int i, n; 1295 boolean_t sync; 1296 1297 if (count && m[0]->object != object) { 1298 panic("swap_pager_putpages: object mismatch %p/%p", 1299 object, 1300 m[0]->object 1301 ); 1302 } 1303 1304 /* 1305 * Step 1 1306 * 1307 * Turn object into OBJT_SWAP 1308 * check for bogus sysops 1309 * force sync if not pageout process 1310 */ 1311 if (object->type != OBJT_SWAP) 1312 swp_pager_meta_build(object, 0, SWAPBLK_NONE); 1313 VM_OBJECT_WUNLOCK(object); 1314 1315 n = 0; 1316 if (curproc != pageproc) 1317 sync = TRUE; 1318 else 1319 sync = (flags & VM_PAGER_PUT_SYNC) != 0; 1320 1321 /* 1322 * Step 2 1323 * 1324 * Assign swap blocks and issue I/O. We reallocate swap on the fly. 1325 * The page is left dirty until the pageout operation completes 1326 * successfully. 1327 */ 1328 for (i = 0; i < count; i += n) { 1329 int j; 1330 struct buf *bp; 1331 daddr_t blk; 1332 1333 /* 1334 * Maximum I/O size is limited by a number of factors. 1335 */ 1336 n = min(BLIST_MAX_ALLOC, count - i); 1337 n = min(n, nsw_cluster_max); 1338 1339 /* 1340 * Get biggest block of swap we can. If we fail, fall 1341 * back and try to allocate a smaller block. Don't go 1342 * overboard trying to allocate space if it would overly 1343 * fragment swap. 1344 */ 1345 while ( 1346 (blk = swp_pager_getswapspace(n)) == SWAPBLK_NONE && 1347 n > 4 1348 ) { 1349 n >>= 1; 1350 } 1351 if (blk == SWAPBLK_NONE) { 1352 for (j = 0; j < n; ++j) 1353 rtvals[i+j] = VM_PAGER_FAIL; 1354 continue; 1355 } 1356 1357 /* 1358 * All I/O parameters have been satisfied, build the I/O 1359 * request and assign the swap space. 1360 */ 1361 if (sync == TRUE) { 1362 bp = getpbuf(&nsw_wcount_sync); 1363 } else { 1364 bp = getpbuf(&nsw_wcount_async); 1365 bp->b_flags = B_ASYNC; 1366 } 1367 bp->b_flags |= B_PAGING; 1368 bp->b_iocmd = BIO_WRITE; 1369 1370 bp->b_rcred = crhold(thread0.td_ucred); 1371 bp->b_wcred = crhold(thread0.td_ucred); 1372 bp->b_bcount = PAGE_SIZE * n; 1373 bp->b_bufsize = PAGE_SIZE * n; 1374 bp->b_blkno = blk; 1375 1376 VM_OBJECT_WLOCK(object); 1377 for (j = 0; j < n; ++j) { 1378 vm_page_t mreq = m[i+j]; 1379 1380 swp_pager_meta_build( 1381 mreq->object, 1382 mreq->pindex, 1383 blk + j 1384 ); 1385 vm_page_dirty(mreq); 1386 mreq->oflags |= VPO_SWAPINPROG; 1387 bp->b_pages[j] = mreq; 1388 } 1389 VM_OBJECT_WUNLOCK(object); 1390 bp->b_npages = n; 1391 /* 1392 * Must set dirty range for NFS to work. 1393 */ 1394 bp->b_dirtyoff = 0; 1395 bp->b_dirtyend = bp->b_bcount; 1396 1397 PCPU_INC(cnt.v_swapout); 1398 PCPU_ADD(cnt.v_swappgsout, bp->b_npages); 1399 1400 /* 1401 * We unconditionally set rtvals[] to VM_PAGER_PEND so that we 1402 * can call the async completion routine at the end of a 1403 * synchronous I/O operation. Otherwise, our caller would 1404 * perform duplicate unbusy and wakeup operations on the page 1405 * and object, respectively. 1406 */ 1407 for (j = 0; j < n; j++) 1408 rtvals[i + j] = VM_PAGER_PEND; 1409 1410 /* 1411 * asynchronous 1412 * 1413 * NOTE: b_blkno is destroyed by the call to swapdev_strategy 1414 */ 1415 if (sync == FALSE) { 1416 bp->b_iodone = swp_pager_async_iodone; 1417 BUF_KERNPROC(bp); 1418 swp_pager_strategy(bp); 1419 continue; 1420 } 1421 1422 /* 1423 * synchronous 1424 * 1425 * NOTE: b_blkno is destroyed by the call to swapdev_strategy 1426 */ 1427 bp->b_iodone = bdone; 1428 swp_pager_strategy(bp); 1429 1430 /* 1431 * Wait for the sync I/O to complete. 1432 */ 1433 bwait(bp, PVM, "swwrt"); 1434 1435 /* 1436 * Now that we are through with the bp, we can call the 1437 * normal async completion, which frees everything up. 1438 */ 1439 swp_pager_async_iodone(bp); 1440 } 1441 VM_OBJECT_WLOCK(object); 1442} 1443 1444/* 1445 * swp_pager_async_iodone: 1446 * 1447 * Completion routine for asynchronous reads and writes from/to swap. 1448 * Also called manually by synchronous code to finish up a bp. 1449 * 1450 * This routine may not sleep. 1451 */ 1452static void 1453swp_pager_async_iodone(struct buf *bp) 1454{ 1455 int i; 1456 vm_object_t object = NULL; 1457 1458 /* 1459 * report error 1460 */ 1461 if (bp->b_ioflags & BIO_ERROR) { 1462 printf( 1463 "swap_pager: I/O error - %s failed; blkno %ld," 1464 "size %ld, error %d\n", 1465 ((bp->b_iocmd == BIO_READ) ? "pagein" : "pageout"), 1466 (long)bp->b_blkno, 1467 (long)bp->b_bcount, 1468 bp->b_error 1469 ); 1470 } 1471 1472 /* 1473 * remove the mapping for kernel virtual 1474 */ 1475 if (buf_mapped(bp)) 1476 pmap_qremove((vm_offset_t)bp->b_data, bp->b_npages); 1477 else 1478 bp->b_data = bp->b_kvabase; 1479 1480 if (bp->b_npages) { 1481 object = bp->b_pages[0]->object; 1482 VM_OBJECT_WLOCK(object); 1483 } 1484 1485 /* 1486 * cleanup pages. If an error occurs writing to swap, we are in 1487 * very serious trouble. If it happens to be a disk error, though, 1488 * we may be able to recover by reassigning the swap later on. So 1489 * in this case we remove the m->swapblk assignment for the page 1490 * but do not free it in the rlist. The errornous block(s) are thus 1491 * never reallocated as swap. Redirty the page and continue. 1492 */ 1493 for (i = 0; i < bp->b_npages; ++i) { 1494 vm_page_t m = bp->b_pages[i]; 1495 1496 m->oflags &= ~VPO_SWAPINPROG; 1497 if (m->oflags & VPO_SWAPSLEEP) { 1498 m->oflags &= ~VPO_SWAPSLEEP; 1499 wakeup(&object->paging_in_progress); 1500 } 1501 1502 if (bp->b_ioflags & BIO_ERROR) { 1503 /* 1504 * If an error occurs I'd love to throw the swapblk 1505 * away without freeing it back to swapspace, so it 1506 * can never be used again. But I can't from an 1507 * interrupt. 1508 */ 1509 if (bp->b_iocmd == BIO_READ) { 1510 /* 1511 * NOTE: for reads, m->dirty will probably 1512 * be overridden by the original caller of 1513 * getpages so don't play cute tricks here. 1514 */ 1515 m->valid = 0; 1516 } else { 1517 /* 1518 * If a write error occurs, reactivate page 1519 * so it doesn't clog the inactive list, 1520 * then finish the I/O. 1521 */ 1522 vm_page_dirty(m); 1523 vm_page_lock(m); 1524 vm_page_activate(m); 1525 vm_page_unlock(m); 1526 vm_page_sunbusy(m); 1527 } 1528 } else if (bp->b_iocmd == BIO_READ) { 1529 /* 1530 * NOTE: for reads, m->dirty will probably be 1531 * overridden by the original caller of getpages so 1532 * we cannot set them in order to free the underlying 1533 * swap in a low-swap situation. I don't think we'd 1534 * want to do that anyway, but it was an optimization 1535 * that existed in the old swapper for a time before 1536 * it got ripped out due to precisely this problem. 1537 */ 1538 KASSERT(!pmap_page_is_mapped(m), 1539 ("swp_pager_async_iodone: page %p is mapped", m)); 1540 KASSERT(m->dirty == 0, 1541 ("swp_pager_async_iodone: page %p is dirty", m)); 1542 1543 m->valid = VM_PAGE_BITS_ALL; 1544 if (i < bp->b_pgbefore || 1545 i >= bp->b_npages - bp->b_pgafter) 1546 vm_page_readahead_finish(m); 1547 } else { 1548 /* 1549 * For write success, clear the dirty 1550 * status, then finish the I/O ( which decrements the 1551 * busy count and possibly wakes waiter's up ). 1552 */ 1553 KASSERT(!pmap_page_is_write_mapped(m), 1554 ("swp_pager_async_iodone: page %p is not write" 1555 " protected", m)); 1556 vm_page_undirty(m); 1557 vm_page_sunbusy(m); 1558 if (vm_page_count_severe()) { 1559 vm_page_lock(m); 1560 vm_page_try_to_cache(m); 1561 vm_page_unlock(m); 1562 } 1563 } 1564 } 1565 1566 /* 1567 * adjust pip. NOTE: the original parent may still have its own 1568 * pip refs on the object. 1569 */ 1570 if (object != NULL) { 1571 vm_object_pip_wakeupn(object, bp->b_npages); 1572 VM_OBJECT_WUNLOCK(object); 1573 } 1574 1575 /* 1576 * swapdev_strategy() manually sets b_vp and b_bufobj before calling 1577 * bstrategy(). Set them back to NULL now we're done with it, or we'll 1578 * trigger a KASSERT in relpbuf(). 1579 */ 1580 if (bp->b_vp) { 1581 bp->b_vp = NULL; 1582 bp->b_bufobj = NULL; 1583 } 1584 /* 1585 * release the physical I/O buffer 1586 */ 1587 relpbuf( 1588 bp, 1589 ((bp->b_iocmd == BIO_READ) ? &nsw_rcount : 1590 ((bp->b_flags & B_ASYNC) ? 1591 &nsw_wcount_async : 1592 &nsw_wcount_sync 1593 ) 1594 ) 1595 ); 1596} 1597 1598/* 1599 * swap_pager_isswapped: 1600 * 1601 * Return 1 if at least one page in the given object is paged 1602 * out to the given swap device. 1603 * 1604 * This routine may not sleep. 1605 */ 1606int 1607swap_pager_isswapped(vm_object_t object, struct swdevt *sp) 1608{ 1609 daddr_t index = 0; 1610 int bcount; 1611 int i; 1612 1613 VM_OBJECT_ASSERT_WLOCKED(object); 1614 if (object->type != OBJT_SWAP) 1615 return (0); 1616 1617 mtx_lock(&swhash_mtx); 1618 for (bcount = 0; bcount < object->un_pager.swp.swp_bcount; bcount++) { 1619 struct swblock *swap; 1620 1621 if ((swap = *swp_pager_hash(object, index)) != NULL) { 1622 for (i = 0; i < SWAP_META_PAGES; ++i) { 1623 if (swp_pager_isondev(swap->swb_pages[i], sp)) { 1624 mtx_unlock(&swhash_mtx); 1625 return (1); 1626 } 1627 } 1628 } 1629 index += SWAP_META_PAGES; 1630 } 1631 mtx_unlock(&swhash_mtx); 1632 return (0); 1633} 1634 1635/* 1636 * SWP_PAGER_FORCE_PAGEIN() - force a swap block to be paged in 1637 * 1638 * This routine dissociates the page at the given index within a 1639 * swap block from its backing store, paging it in if necessary. 1640 * If the page is paged in, it is placed in the inactive queue, 1641 * since it had its backing store ripped out from under it. 1642 * We also attempt to swap in all other pages in the swap block, 1643 * we only guarantee that the one at the specified index is 1644 * paged in. 1645 * 1646 * XXX - The code to page the whole block in doesn't work, so we 1647 * revert to the one-by-one behavior for now. Sigh. 1648 */ 1649static inline void 1650swp_pager_force_pagein(vm_object_t object, vm_pindex_t pindex) 1651{ 1652 vm_page_t m; 1653 1654 vm_object_pip_add(object, 1); 1655 m = vm_page_grab(object, pindex, VM_ALLOC_NORMAL); 1656 if (m->valid == VM_PAGE_BITS_ALL) { 1657 vm_object_pip_wakeup(object); 1658 vm_page_dirty(m); 1659 vm_page_lock(m); 1660 vm_page_activate(m); 1661 vm_page_unlock(m); 1662 vm_page_xunbusy(m); 1663 vm_pager_page_unswapped(m); 1664 return; 1665 } 1666 1667 if (swap_pager_getpages(object, &m, 1, NULL, NULL) != VM_PAGER_OK) 1668 panic("swap_pager_force_pagein: read from swap failed");/*XXX*/ 1669 vm_object_pip_wakeup(object); 1670 vm_page_dirty(m); 1671 vm_page_lock(m); 1672 vm_page_deactivate(m); 1673 vm_page_unlock(m); 1674 vm_page_xunbusy(m); 1675 vm_pager_page_unswapped(m); 1676} 1677 1678/* 1679 * swap_pager_swapoff: 1680 * 1681 * Page in all of the pages that have been paged out to the 1682 * given device. The corresponding blocks in the bitmap must be 1683 * marked as allocated and the device must be flagged SW_CLOSING. 1684 * There may be no processes swapped out to the device. 1685 * 1686 * This routine may block. 1687 */ 1688static void 1689swap_pager_swapoff(struct swdevt *sp) 1690{ 1691 struct swblock *swap; 1692 vm_object_t locked_obj, object; 1693 vm_pindex_t pindex; 1694 int i, j, retries; 1695 1696 sx_assert(&swdev_syscall_lock, SA_XLOCKED); 1697 1698 retries = 0; 1699 locked_obj = NULL; 1700full_rescan: 1701 mtx_lock(&swhash_mtx); 1702 for (i = 0; i <= swhash_mask; i++) { /* '<=' is correct here */ 1703restart: 1704 for (swap = swhash[i]; swap != NULL; swap = swap->swb_hnext) { 1705 object = swap->swb_object; 1706 pindex = swap->swb_index; 1707 for (j = 0; j < SWAP_META_PAGES; ++j) { 1708 if (!swp_pager_isondev(swap->swb_pages[j], sp)) 1709 continue; 1710 if (locked_obj != object) { 1711 if (locked_obj != NULL) 1712 VM_OBJECT_WUNLOCK(locked_obj); 1713 locked_obj = object; 1714 if (!VM_OBJECT_TRYWLOCK(object)) { 1715 mtx_unlock(&swhash_mtx); 1716 /* Depends on type-stability. */ 1717 VM_OBJECT_WLOCK(object); 1718 mtx_lock(&swhash_mtx); 1719 goto restart; 1720 } 1721 } 1722 MPASS(locked_obj == object); 1723 mtx_unlock(&swhash_mtx); 1724 swp_pager_force_pagein(object, pindex + j); 1725 mtx_lock(&swhash_mtx); 1726 goto restart; 1727 } 1728 } 1729 } 1730 mtx_unlock(&swhash_mtx); 1731 if (locked_obj != NULL) { 1732 VM_OBJECT_WUNLOCK(locked_obj); 1733 locked_obj = NULL; 1734 } 1735 if (sp->sw_used) { 1736 /* 1737 * Objects may be locked or paging to the device being 1738 * removed, so we will miss their pages and need to 1739 * make another pass. We have marked this device as 1740 * SW_CLOSING, so the activity should finish soon. 1741 */ 1742 retries++; 1743 if (retries > 100) { 1744 panic("swapoff: failed to locate %d swap blocks", 1745 sp->sw_used); 1746 } 1747 pause("swpoff", hz / 20); 1748 goto full_rescan; 1749 } 1750} 1751 1752/************************************************************************ 1753 * SWAP META DATA * 1754 ************************************************************************ 1755 * 1756 * These routines manipulate the swap metadata stored in the 1757 * OBJT_SWAP object. 1758 * 1759 * Swap metadata is implemented with a global hash and not directly 1760 * linked into the object. Instead the object simply contains 1761 * appropriate tracking counters. 1762 */ 1763 1764/* 1765 * SWP_PAGER_META_BUILD() - add swap block to swap meta data for object 1766 * 1767 * We first convert the object to a swap object if it is a default 1768 * object. 1769 * 1770 * The specified swapblk is added to the object's swap metadata. If 1771 * the swapblk is not valid, it is freed instead. Any previously 1772 * assigned swapblk is freed. 1773 */ 1774static void 1775swp_pager_meta_build(vm_object_t object, vm_pindex_t pindex, daddr_t swapblk) 1776{ 1777 static volatile int exhausted; 1778 struct swblock *swap; 1779 struct swblock **pswap; 1780 int idx; 1781 1782 VM_OBJECT_ASSERT_WLOCKED(object); 1783 /* 1784 * Convert default object to swap object if necessary 1785 */ 1786 if (object->type != OBJT_SWAP) { 1787 object->type = OBJT_SWAP; 1788 object->un_pager.swp.swp_bcount = 0; 1789 KASSERT(object->handle == NULL, ("default pager with handle")); 1790 } 1791 1792 /* 1793 * Locate hash entry. If not found create, but if we aren't adding 1794 * anything just return. If we run out of space in the map we wait 1795 * and, since the hash table may have changed, retry. 1796 */ 1797retry: 1798 mtx_lock(&swhash_mtx); 1799 pswap = swp_pager_hash(object, pindex); 1800 1801 if ((swap = *pswap) == NULL) { 1802 int i; 1803 1804 if (swapblk == SWAPBLK_NONE) 1805 goto done; 1806 1807 swap = *pswap = uma_zalloc(swap_zone, M_NOWAIT | 1808 (curproc == pageproc ? M_USE_RESERVE : 0)); 1809 if (swap == NULL) { 1810 mtx_unlock(&swhash_mtx); 1811 VM_OBJECT_WUNLOCK(object); 1812 if (uma_zone_exhausted(swap_zone)) { 1813 if (atomic_cmpset_int(&exhausted, 0, 1)) 1814 printf("swap zone exhausted, " 1815 "increase kern.maxswzone\n"); 1816 vm_pageout_oom(VM_OOM_SWAPZ); 1817 pause("swzonex", 10); 1818 } else 1819 VM_WAIT; 1820 VM_OBJECT_WLOCK(object); 1821 goto retry; 1822 } 1823 1824 if (atomic_cmpset_int(&exhausted, 1, 0)) 1825 printf("swap zone ok\n"); 1826 1827 swap->swb_hnext = NULL; 1828 swap->swb_object = object; 1829 swap->swb_index = pindex & ~(vm_pindex_t)SWAP_META_MASK; 1830 swap->swb_count = 0; 1831 1832 ++object->un_pager.swp.swp_bcount; 1833 1834 for (i = 0; i < SWAP_META_PAGES; ++i) 1835 swap->swb_pages[i] = SWAPBLK_NONE; 1836 } 1837 1838 /* 1839 * Delete prior contents of metadata 1840 */ 1841 idx = pindex & SWAP_META_MASK; 1842 1843 if (swap->swb_pages[idx] != SWAPBLK_NONE) { 1844 swp_pager_freeswapspace(swap->swb_pages[idx], 1); 1845 --swap->swb_count; 1846 } 1847 1848 /* 1849 * Enter block into metadata 1850 */ 1851 swap->swb_pages[idx] = swapblk; 1852 if (swapblk != SWAPBLK_NONE) 1853 ++swap->swb_count; 1854done: 1855 mtx_unlock(&swhash_mtx); 1856} 1857 1858/* 1859 * SWP_PAGER_META_FREE() - free a range of blocks in the object's swap metadata 1860 * 1861 * The requested range of blocks is freed, with any associated swap 1862 * returned to the swap bitmap. 1863 * 1864 * This routine will free swap metadata structures as they are cleaned 1865 * out. This routine does *NOT* operate on swap metadata associated 1866 * with resident pages. 1867 */ 1868static void 1869swp_pager_meta_free(vm_object_t object, vm_pindex_t index, daddr_t count) 1870{ 1871 1872 VM_OBJECT_ASSERT_LOCKED(object); 1873 if (object->type != OBJT_SWAP) 1874 return; 1875 1876 while (count > 0) { 1877 struct swblock **pswap; 1878 struct swblock *swap; 1879 1880 mtx_lock(&swhash_mtx); 1881 pswap = swp_pager_hash(object, index); 1882 1883 if ((swap = *pswap) != NULL) { 1884 daddr_t v = swap->swb_pages[index & SWAP_META_MASK]; 1885 1886 if (v != SWAPBLK_NONE) { 1887 swp_pager_freeswapspace(v, 1); 1888 swap->swb_pages[index & SWAP_META_MASK] = 1889 SWAPBLK_NONE; 1890 if (--swap->swb_count == 0) { 1891 *pswap = swap->swb_hnext; 1892 uma_zfree(swap_zone, swap); 1893 --object->un_pager.swp.swp_bcount; 1894 } 1895 } 1896 --count; 1897 ++index; 1898 } else { 1899 int n = SWAP_META_PAGES - (index & SWAP_META_MASK); 1900 count -= n; 1901 index += n; 1902 } 1903 mtx_unlock(&swhash_mtx); 1904 } 1905} 1906 1907/* 1908 * SWP_PAGER_META_FREE_ALL() - destroy all swap metadata associated with object 1909 * 1910 * This routine locates and destroys all swap metadata associated with 1911 * an object. 1912 */ 1913static void 1914swp_pager_meta_free_all(vm_object_t object) 1915{ 1916 daddr_t index = 0; 1917 1918 VM_OBJECT_ASSERT_WLOCKED(object); 1919 if (object->type != OBJT_SWAP) 1920 return; 1921 1922 while (object->un_pager.swp.swp_bcount) { 1923 struct swblock **pswap; 1924 struct swblock *swap; 1925 1926 mtx_lock(&swhash_mtx); 1927 pswap = swp_pager_hash(object, index); 1928 if ((swap = *pswap) != NULL) { 1929 int i; 1930 1931 for (i = 0; i < SWAP_META_PAGES; ++i) { 1932 daddr_t v = swap->swb_pages[i]; 1933 if (v != SWAPBLK_NONE) { 1934 --swap->swb_count; 1935 swp_pager_freeswapspace(v, 1); 1936 } 1937 } 1938 if (swap->swb_count != 0) 1939 panic("swap_pager_meta_free_all: swb_count != 0"); 1940 *pswap = swap->swb_hnext; 1941 uma_zfree(swap_zone, swap); 1942 --object->un_pager.swp.swp_bcount; 1943 } 1944 mtx_unlock(&swhash_mtx); 1945 index += SWAP_META_PAGES; 1946 } 1947} 1948 1949/* 1950 * SWP_PAGER_METACTL() - misc control of swap and vm_page_t meta data. 1951 * 1952 * This routine is capable of looking up, popping, or freeing 1953 * swapblk assignments in the swap meta data or in the vm_page_t. 1954 * The routine typically returns the swapblk being looked-up, or popped, 1955 * or SWAPBLK_NONE if the block was freed, or SWAPBLK_NONE if the block 1956 * was invalid. This routine will automatically free any invalid 1957 * meta-data swapblks. 1958 * 1959 * It is not possible to store invalid swapblks in the swap meta data 1960 * (other then a literal 'SWAPBLK_NONE'), so we don't bother checking. 1961 * 1962 * When acting on a busy resident page and paging is in progress, we 1963 * have to wait until paging is complete but otherwise can act on the 1964 * busy page. 1965 * 1966 * SWM_FREE remove and free swap block from metadata 1967 * SWM_POP remove from meta data but do not free.. pop it out 1968 */ 1969static daddr_t 1970swp_pager_meta_ctl(vm_object_t object, vm_pindex_t pindex, int flags) 1971{ 1972 struct swblock **pswap; 1973 struct swblock *swap; 1974 daddr_t r1; 1975 int idx; 1976 1977 VM_OBJECT_ASSERT_LOCKED(object); 1978 /* 1979 * The meta data only exists of the object is OBJT_SWAP 1980 * and even then might not be allocated yet. 1981 */ 1982 if (object->type != OBJT_SWAP) 1983 return (SWAPBLK_NONE); 1984 1985 r1 = SWAPBLK_NONE; 1986 mtx_lock(&swhash_mtx); 1987 pswap = swp_pager_hash(object, pindex); 1988 1989 if ((swap = *pswap) != NULL) { 1990 idx = pindex & SWAP_META_MASK; 1991 r1 = swap->swb_pages[idx]; 1992 1993 if (r1 != SWAPBLK_NONE) { 1994 if (flags & SWM_FREE) { 1995 swp_pager_freeswapspace(r1, 1); 1996 r1 = SWAPBLK_NONE; 1997 } 1998 if (flags & (SWM_FREE|SWM_POP)) { 1999 swap->swb_pages[idx] = SWAPBLK_NONE; 2000 if (--swap->swb_count == 0) { 2001 *pswap = swap->swb_hnext; 2002 uma_zfree(swap_zone, swap); 2003 --object->un_pager.swp.swp_bcount; 2004 } 2005 } 2006 } 2007 } 2008 mtx_unlock(&swhash_mtx); 2009 return (r1); 2010} 2011 2012/* 2013 * System call swapon(name) enables swapping on device name, 2014 * which must be in the swdevsw. Return EBUSY 2015 * if already swapping on this device. 2016 */ 2017#ifndef _SYS_SYSPROTO_H_ 2018struct swapon_args { 2019 char *name; 2020}; 2021#endif 2022 2023/* 2024 * MPSAFE 2025 */ 2026/* ARGSUSED */ 2027int 2028sys_swapon(struct thread *td, struct swapon_args *uap) 2029{ 2030 struct vattr attr; 2031 struct vnode *vp; 2032 struct nameidata nd; 2033 int error; 2034 2035 error = priv_check(td, PRIV_SWAPON); 2036 if (error) 2037 return (error); 2038 2039 sx_xlock(&swdev_syscall_lock); 2040 2041 /* 2042 * Swap metadata may not fit in the KVM if we have physical 2043 * memory of >1GB. 2044 */ 2045 if (swap_zone == NULL) { 2046 error = ENOMEM; 2047 goto done; 2048 } 2049 2050 NDINIT(&nd, LOOKUP, ISOPEN | FOLLOW | AUDITVNODE1, UIO_USERSPACE, 2051 uap->name, td); 2052 error = namei(&nd); 2053 if (error) 2054 goto done; 2055 2056 NDFREE(&nd, NDF_ONLY_PNBUF); 2057 vp = nd.ni_vp; 2058 2059 if (vn_isdisk(vp, &error)) { 2060 error = swapongeom(vp); 2061 } else if (vp->v_type == VREG && 2062 (vp->v_mount->mnt_vfc->vfc_flags & VFCF_NETWORK) != 0 && 2063 (error = VOP_GETATTR(vp, &attr, td->td_ucred)) == 0) { 2064 /* 2065 * Allow direct swapping to NFS regular files in the same 2066 * way that nfs_mountroot() sets up diskless swapping. 2067 */ 2068 error = swaponvp(td, vp, attr.va_size / DEV_BSIZE); 2069 } 2070 2071 if (error) 2072 vrele(vp); 2073done: 2074 sx_xunlock(&swdev_syscall_lock); 2075 return (error); 2076} 2077 2078/* 2079 * Check that the total amount of swap currently configured does not 2080 * exceed half the theoretical maximum. If it does, print a warning 2081 * message and return -1; otherwise, return 0. 2082 */ 2083static int 2084swapon_check_swzone(unsigned long npages) 2085{ 2086 unsigned long maxpages; 2087 2088 /* absolute maximum we can handle assuming 100% efficiency */ 2089 maxpages = uma_zone_get_max(swap_zone) * SWAP_META_PAGES; 2090 2091 /* recommend using no more than half that amount */ 2092 if (npages > maxpages / 2) { 2093 printf("warning: total configured swap (%lu pages) " 2094 "exceeds maximum recommended amount (%lu pages).\n", 2095 npages, maxpages / 2); 2096 printf("warning: increase kern.maxswzone " 2097 "or reduce amount of swap.\n"); 2098 return (-1); 2099 } 2100 return (0); 2101} 2102 2103static void 2104swaponsomething(struct vnode *vp, void *id, u_long nblks, 2105 sw_strategy_t *strategy, sw_close_t *close, dev_t dev, int flags) 2106{ 2107 struct swdevt *sp, *tsp; 2108 swblk_t dvbase; 2109 u_long mblocks; 2110 2111 /* 2112 * nblks is in DEV_BSIZE'd chunks, convert to PAGE_SIZE'd chunks. 2113 * First chop nblks off to page-align it, then convert. 2114 * 2115 * sw->sw_nblks is in page-sized chunks now too. 2116 */ 2117 nblks &= ~(ctodb(1) - 1); 2118 nblks = dbtoc(nblks); 2119 2120 /* 2121 * If we go beyond this, we get overflows in the radix 2122 * tree bitmap code. 2123 */ 2124 mblocks = 0x40000000 / BLIST_META_RADIX; 2125 if (nblks > mblocks) { 2126 printf( 2127 "WARNING: reducing swap size to maximum of %luMB per unit\n", 2128 mblocks / 1024 / 1024 * PAGE_SIZE); 2129 nblks = mblocks; 2130 } 2131 2132 sp = malloc(sizeof *sp, M_VMPGDATA, M_WAITOK | M_ZERO); 2133 sp->sw_vp = vp; 2134 sp->sw_id = id; 2135 sp->sw_dev = dev; 2136 sp->sw_flags = 0; 2137 sp->sw_nblks = nblks; 2138 sp->sw_used = 0; 2139 sp->sw_strategy = strategy; 2140 sp->sw_close = close; 2141 sp->sw_flags = flags; 2142 2143 sp->sw_blist = blist_create(nblks, M_WAITOK); 2144 /* 2145 * Do not free the first two block in order to avoid overwriting 2146 * any bsd label at the front of the partition 2147 */ 2148 blist_free(sp->sw_blist, 2, nblks - 2); 2149 2150 dvbase = 0; 2151 mtx_lock(&sw_dev_mtx); 2152 TAILQ_FOREACH(tsp, &swtailq, sw_list) { 2153 if (tsp->sw_end >= dvbase) { 2154 /* 2155 * We put one uncovered page between the devices 2156 * in order to definitively prevent any cross-device 2157 * I/O requests 2158 */ 2159 dvbase = tsp->sw_end + 1; 2160 } 2161 } 2162 sp->sw_first = dvbase; 2163 sp->sw_end = dvbase + nblks; 2164 TAILQ_INSERT_TAIL(&swtailq, sp, sw_list); 2165 nswapdev++; 2166 swap_pager_avail += nblks; 2167 swap_total += (vm_ooffset_t)nblks * PAGE_SIZE; 2168 swapon_check_swzone(swap_total / PAGE_SIZE); 2169 swp_sizecheck(); 2170 mtx_unlock(&sw_dev_mtx); 2171} 2172 2173/* 2174 * SYSCALL: swapoff(devname) 2175 * 2176 * Disable swapping on the given device. 2177 * 2178 * XXX: Badly designed system call: it should use a device index 2179 * rather than filename as specification. We keep sw_vp around 2180 * only to make this work. 2181 */ 2182#ifndef _SYS_SYSPROTO_H_ 2183struct swapoff_args { 2184 char *name; 2185}; 2186#endif 2187 2188/* 2189 * MPSAFE 2190 */ 2191/* ARGSUSED */ 2192int 2193sys_swapoff(struct thread *td, struct swapoff_args *uap) 2194{ 2195 struct vnode *vp; 2196 struct nameidata nd; 2197 struct swdevt *sp; 2198 int error; 2199 2200 error = priv_check(td, PRIV_SWAPOFF); 2201 if (error) 2202 return (error); 2203 2204 sx_xlock(&swdev_syscall_lock); 2205 2206 NDINIT(&nd, LOOKUP, FOLLOW | AUDITVNODE1, UIO_USERSPACE, uap->name, 2207 td); 2208 error = namei(&nd); 2209 if (error) 2210 goto done; 2211 NDFREE(&nd, NDF_ONLY_PNBUF); 2212 vp = nd.ni_vp; 2213 2214 mtx_lock(&sw_dev_mtx); 2215 TAILQ_FOREACH(sp, &swtailq, sw_list) { 2216 if (sp->sw_vp == vp) 2217 break; 2218 } 2219 mtx_unlock(&sw_dev_mtx); 2220 if (sp == NULL) { 2221 error = EINVAL; 2222 goto done; 2223 } 2224 error = swapoff_one(sp, td->td_ucred); 2225done: 2226 sx_xunlock(&swdev_syscall_lock); 2227 return (error); 2228} 2229 2230static int 2231swapoff_one(struct swdevt *sp, struct ucred *cred) 2232{ 2233 u_long nblks, dvbase; 2234#ifdef MAC 2235 int error; 2236#endif 2237 2238 sx_assert(&swdev_syscall_lock, SA_XLOCKED); 2239#ifdef MAC 2240 (void) vn_lock(sp->sw_vp, LK_EXCLUSIVE | LK_RETRY); 2241 error = mac_system_check_swapoff(cred, sp->sw_vp); 2242 (void) VOP_UNLOCK(sp->sw_vp, 0); 2243 if (error != 0) 2244 return (error); 2245#endif 2246 nblks = sp->sw_nblks; 2247 2248 /* 2249 * We can turn off this swap device safely only if the 2250 * available virtual memory in the system will fit the amount 2251 * of data we will have to page back in, plus an epsilon so 2252 * the system doesn't become critically low on swap space. 2253 */ 2254 if (vm_cnt.v_free_count + vm_cnt.v_cache_count + swap_pager_avail < 2255 nblks + nswap_lowat) { 2256 return (ENOMEM); 2257 } 2258 2259 /* 2260 * Prevent further allocations on this device. 2261 */ 2262 mtx_lock(&sw_dev_mtx); 2263 sp->sw_flags |= SW_CLOSING; 2264 for (dvbase = 0; dvbase < sp->sw_end; dvbase += dmmax) { 2265 swap_pager_avail -= blist_fill(sp->sw_blist, 2266 dvbase, dmmax); 2267 } 2268 swap_total -= (vm_ooffset_t)nblks * PAGE_SIZE; 2269 mtx_unlock(&sw_dev_mtx); 2270 2271 /* 2272 * Page in the contents of the device and close it. 2273 */ 2274 swap_pager_swapoff(sp); 2275 2276 sp->sw_close(curthread, sp); 2277 mtx_lock(&sw_dev_mtx); 2278 sp->sw_id = NULL; 2279 TAILQ_REMOVE(&swtailq, sp, sw_list); 2280 nswapdev--; 2281 if (nswapdev == 0) { 2282 swap_pager_full = 2; 2283 swap_pager_almost_full = 1; 2284 } 2285 if (swdevhd == sp) 2286 swdevhd = NULL; 2287 mtx_unlock(&sw_dev_mtx); 2288 blist_destroy(sp->sw_blist); 2289 free(sp, M_VMPGDATA); 2290 return (0); 2291} 2292 2293void 2294swapoff_all(void) 2295{ 2296 struct swdevt *sp, *spt; 2297 const char *devname; 2298 int error; 2299 2300 sx_xlock(&swdev_syscall_lock); 2301 2302 mtx_lock(&sw_dev_mtx); 2303 TAILQ_FOREACH_SAFE(sp, &swtailq, sw_list, spt) { 2304 mtx_unlock(&sw_dev_mtx); 2305 if (vn_isdisk(sp->sw_vp, NULL)) 2306 devname = devtoname(sp->sw_vp->v_rdev); 2307 else 2308 devname = "[file]"; 2309 error = swapoff_one(sp, thread0.td_ucred); 2310 if (error != 0) { 2311 printf("Cannot remove swap device %s (error=%d), " 2312 "skipping.\n", devname, error); 2313 } else if (bootverbose) { 2314 printf("Swap device %s removed.\n", devname); 2315 } 2316 mtx_lock(&sw_dev_mtx); 2317 } 2318 mtx_unlock(&sw_dev_mtx); 2319 2320 sx_xunlock(&swdev_syscall_lock); 2321} 2322 2323void 2324swap_pager_status(int *total, int *used) 2325{ 2326 struct swdevt *sp; 2327 2328 *total = 0; 2329 *used = 0; 2330 mtx_lock(&sw_dev_mtx); 2331 TAILQ_FOREACH(sp, &swtailq, sw_list) { 2332 *total += sp->sw_nblks; 2333 *used += sp->sw_used; 2334 } 2335 mtx_unlock(&sw_dev_mtx); 2336} 2337 2338int 2339swap_dev_info(int name, struct xswdev *xs, char *devname, size_t len) 2340{ 2341 struct swdevt *sp; 2342 const char *tmp_devname; 2343 int error, n; 2344 2345 n = 0; 2346 error = ENOENT; 2347 mtx_lock(&sw_dev_mtx); 2348 TAILQ_FOREACH(sp, &swtailq, sw_list) { 2349 if (n != name) { 2350 n++; 2351 continue; 2352 } 2353 xs->xsw_version = XSWDEV_VERSION; 2354 xs->xsw_dev = sp->sw_dev; 2355 xs->xsw_flags = sp->sw_flags; 2356 xs->xsw_nblks = sp->sw_nblks; 2357 xs->xsw_used = sp->sw_used; 2358 if (devname != NULL) { 2359 if (vn_isdisk(sp->sw_vp, NULL)) 2360 tmp_devname = devtoname(sp->sw_vp->v_rdev); 2361 else 2362 tmp_devname = "[file]"; 2363 strncpy(devname, tmp_devname, len); 2364 } 2365 error = 0; 2366 break; 2367 } 2368 mtx_unlock(&sw_dev_mtx); 2369 return (error); 2370} 2371 2372static int 2373sysctl_vm_swap_info(SYSCTL_HANDLER_ARGS) 2374{ 2375 struct xswdev xs; 2376 int error; 2377 2378 if (arg2 != 1) /* name length */ 2379 return (EINVAL); 2380 error = swap_dev_info(*(int *)arg1, &xs, NULL, 0); 2381 if (error != 0) 2382 return (error); 2383 error = SYSCTL_OUT(req, &xs, sizeof(xs)); 2384 return (error); 2385} 2386 2387SYSCTL_INT(_vm, OID_AUTO, nswapdev, CTLFLAG_RD, &nswapdev, 0, 2388 "Number of swap devices"); 2389SYSCTL_NODE(_vm, OID_AUTO, swap_info, CTLFLAG_RD | CTLFLAG_MPSAFE, 2390 sysctl_vm_swap_info, 2391 "Swap statistics by device"); 2392 2393/* 2394 * vmspace_swap_count() - count the approximate swap usage in pages for a 2395 * vmspace. 2396 * 2397 * The map must be locked. 2398 * 2399 * Swap usage is determined by taking the proportional swap used by 2400 * VM objects backing the VM map. To make up for fractional losses, 2401 * if the VM object has any swap use at all the associated map entries 2402 * count for at least 1 swap page. 2403 */ 2404long 2405vmspace_swap_count(struct vmspace *vmspace) 2406{ 2407 vm_map_t map; 2408 vm_map_entry_t cur; 2409 vm_object_t object; 2410 long count, n; 2411 2412 map = &vmspace->vm_map; 2413 count = 0; 2414 2415 for (cur = map->header.next; cur != &map->header; cur = cur->next) { 2416 if ((cur->eflags & MAP_ENTRY_IS_SUB_MAP) == 0 && 2417 (object = cur->object.vm_object) != NULL) { 2418 VM_OBJECT_WLOCK(object); 2419 if (object->type == OBJT_SWAP && 2420 object->un_pager.swp.swp_bcount != 0) { 2421 n = (cur->end - cur->start) / PAGE_SIZE; 2422 count += object->un_pager.swp.swp_bcount * 2423 SWAP_META_PAGES * n / object->size + 1; 2424 } 2425 VM_OBJECT_WUNLOCK(object); 2426 } 2427 } 2428 return (count); 2429} 2430 2431/* 2432 * GEOM backend 2433 * 2434 * Swapping onto disk devices. 2435 * 2436 */ 2437 2438static g_orphan_t swapgeom_orphan; 2439 2440static struct g_class g_swap_class = { 2441 .name = "SWAP", 2442 .version = G_VERSION, 2443 .orphan = swapgeom_orphan, 2444}; 2445 2446DECLARE_GEOM_CLASS(g_swap_class, g_class); 2447 2448 2449static void 2450swapgeom_close_ev(void *arg, int flags) 2451{ 2452 struct g_consumer *cp; 2453 2454 cp = arg; 2455 g_access(cp, -1, -1, 0); 2456 g_detach(cp); 2457 g_destroy_consumer(cp); 2458} 2459 2460/* 2461 * Add a reference to the g_consumer for an inflight transaction. 2462 */ 2463static void 2464swapgeom_acquire(struct g_consumer *cp) 2465{ 2466 2467 mtx_assert(&sw_dev_mtx, MA_OWNED); 2468 cp->index++; 2469} 2470 2471/* 2472 * Remove a reference from the g_consumer. Post a close event if all 2473 * references go away, since the function might be called from the 2474 * biodone context. 2475 */ 2476static void 2477swapgeom_release(struct g_consumer *cp, struct swdevt *sp) 2478{ 2479 2480 mtx_assert(&sw_dev_mtx, MA_OWNED); 2481 cp->index--; 2482 if (cp->index == 0) { 2483 if (g_post_event(swapgeom_close_ev, cp, M_NOWAIT, NULL) == 0) 2484 sp->sw_id = NULL; 2485 } 2486} 2487 2488static void 2489swapgeom_done(struct bio *bp2) 2490{ 2491 struct swdevt *sp; 2492 struct buf *bp; 2493 struct g_consumer *cp; 2494 2495 bp = bp2->bio_caller2; 2496 cp = bp2->bio_from; 2497 bp->b_ioflags = bp2->bio_flags; 2498 if (bp2->bio_error) 2499 bp->b_ioflags |= BIO_ERROR; 2500 bp->b_resid = bp->b_bcount - bp2->bio_completed; 2501 bp->b_error = bp2->bio_error; 2502 bufdone(bp); 2503 sp = bp2->bio_caller1; 2504 mtx_lock(&sw_dev_mtx); 2505 swapgeom_release(cp, sp); 2506 mtx_unlock(&sw_dev_mtx); 2507 g_destroy_bio(bp2); 2508} 2509 2510static void 2511swapgeom_strategy(struct buf *bp, struct swdevt *sp) 2512{ 2513 struct bio *bio; 2514 struct g_consumer *cp; 2515 2516 mtx_lock(&sw_dev_mtx); 2517 cp = sp->sw_id; 2518 if (cp == NULL) { 2519 mtx_unlock(&sw_dev_mtx); 2520 bp->b_error = ENXIO; 2521 bp->b_ioflags |= BIO_ERROR; 2522 bufdone(bp); 2523 return; 2524 } 2525 swapgeom_acquire(cp); 2526 mtx_unlock(&sw_dev_mtx); 2527 if (bp->b_iocmd == BIO_WRITE) 2528 bio = g_new_bio(); 2529 else 2530 bio = g_alloc_bio(); 2531 if (bio == NULL) { 2532 mtx_lock(&sw_dev_mtx); 2533 swapgeom_release(cp, sp); 2534 mtx_unlock(&sw_dev_mtx); 2535 bp->b_error = ENOMEM; 2536 bp->b_ioflags |= BIO_ERROR; 2537 bufdone(bp); 2538 return; 2539 } 2540 2541 bio->bio_caller1 = sp; 2542 bio->bio_caller2 = bp; 2543 bio->bio_cmd = bp->b_iocmd; 2544 bio->bio_offset = (bp->b_blkno - sp->sw_first) * PAGE_SIZE; 2545 bio->bio_length = bp->b_bcount; 2546 bio->bio_done = swapgeom_done; 2547 if (!buf_mapped(bp)) { 2548 bio->bio_ma = bp->b_pages; 2549 bio->bio_data = unmapped_buf; 2550 bio->bio_ma_offset = (vm_offset_t)bp->b_offset & PAGE_MASK; 2551 bio->bio_ma_n = bp->b_npages; 2552 bio->bio_flags |= BIO_UNMAPPED; 2553 } else { 2554 bio->bio_data = bp->b_data; 2555 bio->bio_ma = NULL; 2556 } 2557 g_io_request(bio, cp); 2558 return; 2559} 2560 2561static void 2562swapgeom_orphan(struct g_consumer *cp) 2563{ 2564 struct swdevt *sp; 2565 int destroy; 2566 2567 mtx_lock(&sw_dev_mtx); 2568 TAILQ_FOREACH(sp, &swtailq, sw_list) { 2569 if (sp->sw_id == cp) { 2570 sp->sw_flags |= SW_CLOSING; 2571 break; 2572 } 2573 } 2574 /* 2575 * Drop reference we were created with. Do directly since we're in a 2576 * special context where we don't have to queue the call to 2577 * swapgeom_close_ev(). 2578 */ 2579 cp->index--; 2580 destroy = ((sp != NULL) && (cp->index == 0)); 2581 if (destroy) 2582 sp->sw_id = NULL; 2583 mtx_unlock(&sw_dev_mtx); 2584 if (destroy) 2585 swapgeom_close_ev(cp, 0); 2586} 2587 2588static void 2589swapgeom_close(struct thread *td, struct swdevt *sw) 2590{ 2591 struct g_consumer *cp; 2592 2593 mtx_lock(&sw_dev_mtx); 2594 cp = sw->sw_id; 2595 sw->sw_id = NULL; 2596 mtx_unlock(&sw_dev_mtx); 2597 2598 /* 2599 * swapgeom_close() may be called from the biodone context, 2600 * where we cannot perform topology changes. Delegate the 2601 * work to the events thread. 2602 */ 2603 if (cp != NULL) 2604 g_waitfor_event(swapgeom_close_ev, cp, M_WAITOK, NULL); 2605} 2606 2607static int 2608swapongeom_locked(struct cdev *dev, struct vnode *vp) 2609{ 2610 struct g_provider *pp; 2611 struct g_consumer *cp; 2612 static struct g_geom *gp; 2613 struct swdevt *sp; 2614 u_long nblks; 2615 int error; 2616 2617 pp = g_dev_getprovider(dev); 2618 if (pp == NULL) 2619 return (ENODEV); 2620 mtx_lock(&sw_dev_mtx); 2621 TAILQ_FOREACH(sp, &swtailq, sw_list) { 2622 cp = sp->sw_id; 2623 if (cp != NULL && cp->provider == pp) { 2624 mtx_unlock(&sw_dev_mtx); 2625 return (EBUSY); 2626 } 2627 } 2628 mtx_unlock(&sw_dev_mtx); 2629 if (gp == NULL) 2630 gp = g_new_geomf(&g_swap_class, "swap"); 2631 cp = g_new_consumer(gp); 2632 cp->index = 1; /* Number of active I/Os, plus one for being active. */ 2633 cp->flags |= G_CF_DIRECT_SEND | G_CF_DIRECT_RECEIVE; 2634 g_attach(cp, pp); 2635 /* 2636 * XXX: Every time you think you can improve the margin for 2637 * footshooting, somebody depends on the ability to do so: 2638 * savecore(8) wants to write to our swapdev so we cannot 2639 * set an exclusive count :-( 2640 */ 2641 error = g_access(cp, 1, 1, 0); 2642 if (error != 0) { 2643 g_detach(cp); 2644 g_destroy_consumer(cp); 2645 return (error); 2646 } 2647 nblks = pp->mediasize / DEV_BSIZE; 2648 swaponsomething(vp, cp, nblks, swapgeom_strategy, 2649 swapgeom_close, dev2udev(dev), 2650 (pp->flags & G_PF_ACCEPT_UNMAPPED) != 0 ? SW_UNMAPPED : 0); 2651 return (0); 2652} 2653 2654static int 2655swapongeom(struct vnode *vp) 2656{ 2657 int error; 2658 2659 vn_lock(vp, LK_EXCLUSIVE | LK_RETRY); 2660 if (vp->v_type != VCHR || (vp->v_iflag & VI_DOOMED) != 0) { 2661 error = ENOENT; 2662 } else { 2663 g_topology_lock(); 2664 error = swapongeom_locked(vp->v_rdev, vp); 2665 g_topology_unlock(); 2666 } 2667 VOP_UNLOCK(vp, 0); 2668 return (error); 2669} 2670 2671/* 2672 * VNODE backend 2673 * 2674 * This is used mainly for network filesystem (read: probably only tested 2675 * with NFS) swapfiles. 2676 * 2677 */ 2678 2679static void 2680swapdev_strategy(struct buf *bp, struct swdevt *sp) 2681{ 2682 struct vnode *vp2; 2683 2684 bp->b_blkno = ctodb(bp->b_blkno - sp->sw_first); 2685 2686 vp2 = sp->sw_id; 2687 vhold(vp2); 2688 if (bp->b_iocmd == BIO_WRITE) { 2689 if (bp->b_bufobj) 2690 bufobj_wdrop(bp->b_bufobj); 2691 bufobj_wref(&vp2->v_bufobj); 2692 } 2693 if (bp->b_bufobj != &vp2->v_bufobj) 2694 bp->b_bufobj = &vp2->v_bufobj; 2695 bp->b_vp = vp2; 2696 bp->b_iooffset = dbtob(bp->b_blkno); 2697 bstrategy(bp); 2698 return; 2699} 2700 2701static void 2702swapdev_close(struct thread *td, struct swdevt *sp) 2703{ 2704 2705 VOP_CLOSE(sp->sw_vp, FREAD | FWRITE, td->td_ucred, td); 2706 vrele(sp->sw_vp); 2707} 2708 2709 2710static int 2711swaponvp(struct thread *td, struct vnode *vp, u_long nblks) 2712{ 2713 struct swdevt *sp; 2714 int error; 2715 2716 if (nblks == 0) 2717 return (ENXIO); 2718 mtx_lock(&sw_dev_mtx); 2719 TAILQ_FOREACH(sp, &swtailq, sw_list) { 2720 if (sp->sw_id == vp) { 2721 mtx_unlock(&sw_dev_mtx); 2722 return (EBUSY); 2723 } 2724 } 2725 mtx_unlock(&sw_dev_mtx); 2726 2727 (void) vn_lock(vp, LK_EXCLUSIVE | LK_RETRY); 2728#ifdef MAC 2729 error = mac_system_check_swapon(td->td_ucred, vp); 2730 if (error == 0) 2731#endif 2732 error = VOP_OPEN(vp, FREAD | FWRITE, td->td_ucred, td, NULL); 2733 (void) VOP_UNLOCK(vp, 0); 2734 if (error) 2735 return (error); 2736 2737 swaponsomething(vp, vp, nblks, swapdev_strategy, swapdev_close, 2738 NODEV, 0); 2739 return (0); 2740} 2741 2742static int 2743sysctl_swap_async_max(SYSCTL_HANDLER_ARGS) 2744{ 2745 int error, new, n; 2746 2747 new = nsw_wcount_async_max; 2748 error = sysctl_handle_int(oidp, &new, 0, req); 2749 if (error != 0 || req->newptr == NULL) 2750 return (error); 2751 2752 if (new > nswbuf / 2 || new < 1) 2753 return (EINVAL); 2754 2755 mtx_lock(&pbuf_mtx); 2756 while (nsw_wcount_async_max != new) { 2757 /* 2758 * Adjust difference. If the current async count is too low, 2759 * we will need to sqeeze our update slowly in. Sleep with a 2760 * higher priority than getpbuf() to finish faster. 2761 */ 2762 n = new - nsw_wcount_async_max; 2763 if (nsw_wcount_async + n >= 0) { 2764 nsw_wcount_async += n; 2765 nsw_wcount_async_max += n; 2766 wakeup(&nsw_wcount_async); 2767 } else { 2768 nsw_wcount_async_max -= nsw_wcount_async; 2769 nsw_wcount_async = 0; 2770 msleep(&nsw_wcount_async, &pbuf_mtx, PSWP, 2771 "swpsysctl", 0); 2772 } 2773 } 2774 mtx_unlock(&pbuf_mtx); 2775 2776 return (0); 2777} 2778