1 /*- 2 * Copyright (c) 1998 Matthew Dillon, 3 * Copyright (c) 1994 John S. Dyson 4 * Copyright (c) 1990 University of Utah. 5 * Copyright (c) 1982, 1986, 1989, 1993 6 * The Regents of the University of California. All rights reserved. 7 * 8 * This code is derived from software contributed to Berkeley by 9 * the Systems Programming Group of the University of Utah Computer 10 * Science Department. 11 * 12 * Redistribution and use in source and binary forms, with or without 13 * modification, are permitted provided that the following conditions 14 * are met: 15 * 1. Redistributions of source code must retain the above copyright 16 * notice, this list of conditions and the following disclaimer. 17 * 2. Redistributions in binary form must reproduce the above copyright 18 * notice, this list of conditions and the following disclaimer in the 19 * documentation and/or other materials provided with the distribution. 20 * 3. All advertising materials mentioning features or use of this software 21 * must display the following acknowledgement: 22 * This product includes software developed by the University of 23 * California, Berkeley and its contributors. 24 * 4. Neither the name of the University nor the names of its contributors 25 * may be used to endorse or promote products derived from this software 26 * without specific prior written permission. 27 * 28 * THIS SOFTWARE IS PROVIDED BY THE REGENTS AND CONTRIBUTORS ``AS IS'' AND 29 * ANY EXPRESS OR IMPLIED WARRANTIES, INCLUDING, BUT NOT LIMITED TO, THE 30 * IMPLIED WARRANTIES OF MERCHANTABILITY AND FITNESS FOR A PARTICULAR PURPOSE 31 * ARE DISCLAIMED. IN NO EVENT SHALL THE REGENTS OR CONTRIBUTORS BE LIABLE 32 * FOR ANY DIRECT, INDIRECT, INCIDENTAL, SPECIAL, EXEMPLARY, OR CONSEQUENTIAL 33 * DAMAGES (INCLUDING, BUT NOT LIMITED TO, PROCUREMENT OF SUBSTITUTE GOODS 34 * OR SERVICES; LOSS OF USE, DATA, OR PROFITS; OR BUSINESS INTERRUPTION) 35 * HOWEVER CAUSED AND ON ANY THEORY OF LIABILITY, WHETHER IN CONTRACT, STRICT 36 * LIABILITY, OR TORT (INCLUDING NEGLIGENCE OR OTHERWISE) ARISING IN ANY WAY 37 * OUT OF THE USE OF THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF 38 * SUCH DAMAGE. 39 * 40 * New Swap System 41 * Matthew Dillon 42 * 43 * Radix Bitmap 'blists'. 44 * 45 * - The new swapper uses the new radix bitmap code. This should scale 46 * to arbitrarily small or arbitrarily large swap spaces and an almost 47 * arbitrary degree of fragmentation. 48 * 49 * Features: 50 * 51 * - on the fly reallocation of swap during putpages. The new system 52 * does not try to keep previously allocated swap blocks for dirty 53 * pages. 54 * 55 * - on the fly deallocation of swap 56 * 57 * - No more garbage collection required. Unnecessarily allocated swap 58 * blocks only exist for dirty vm_page_t's now and these are already 59 * cycled (in a high-load system) by the pager. We also do on-the-fly 60 * removal of invalidated swap blocks when a page is destroyed 61 * or renamed. 62 * 63 * from: Utah $Hdr: swap_pager.c 1.4 91/04/30$ 64 * 65 * @(#)swap_pager.c 8.9 (Berkeley) 3/21/94 66 * @(#)vm_swap.c 8.5 (Berkeley) 2/17/94 67 */ 68 69 #include <sys/cdefs.h> 70 __FBSDID("$FreeBSD$"); 71 72 #include "opt_swap.h" 73 #include "opt_vm.h" 74 75 #include <sys/param.h> 76 #include <sys/systm.h> 77 #include <sys/conf.h> 78 #include <sys/kernel.h> 79 #include <sys/priv.h> 80 #include <sys/proc.h> 81 #include <sys/bio.h> 82 #include <sys/buf.h> 83 #include <sys/disk.h> 84 #include <sys/fcntl.h> 85 #include <sys/mount.h> 86 #include <sys/namei.h> 87 #include <sys/vnode.h> 88 #include <sys/malloc.h> 89 #include <sys/racct.h> 90 #include <sys/resource.h> 91 #include <sys/resourcevar.h> 92 #include <sys/rwlock.h> 93 #include <sys/sysctl.h> 94 #include <sys/sysproto.h> 95 #include <sys/blist.h> 96 #include <sys/lock.h> 97 #include <sys/sx.h> 98 #include <sys/vmmeter.h> 99 100 #include <security/mac/mac_framework.h> 101 102 #include <vm/vm.h> 103 #include <vm/pmap.h> 104 #include <vm/vm_map.h> 105 #include <vm/vm_kern.h> 106 #include <vm/vm_object.h> 107 #include <vm/vm_page.h> 108 #include <vm/vm_pager.h> 109 #include <vm/vm_pageout.h> 110 #include <vm/vm_param.h> 111 #include <vm/swap_pager.h> 112 #include <vm/vm_extern.h> 113 #include <vm/uma.h> 114 115 #include <geom/geom.h> 116 117 /* 118 * SWB_NPAGES must be a power of 2. It may be set to 1, 2, 4, 8, 16 119 * or 32 pages per allocation. 120 * The 32-page limit is due to the radix code (kern/subr_blist.c). 121 */ 122 #ifndef MAX_PAGEOUT_CLUSTER 123 #define MAX_PAGEOUT_CLUSTER 16 124 #endif 125 126 #if !defined(SWB_NPAGES) 127 #define SWB_NPAGES MAX_PAGEOUT_CLUSTER 128 #endif 129 130 /* 131 * The swblock structure maps an object and a small, fixed-size range 132 * of page indices to disk addresses within a swap area. 133 * The collection of these mappings is implemented as a hash table. 134 * Unused disk addresses within a swap area are allocated and managed 135 * using a blist. 136 */ 137 #define SWCORRECT(n) (sizeof(void *) * (n) / sizeof(daddr_t)) 138 #define SWAP_META_PAGES (SWB_NPAGES * 2) 139 #define SWAP_META_MASK (SWAP_META_PAGES - 1) 140 141 struct swblock { 142 struct swblock *swb_hnext; 143 vm_object_t swb_object; 144 vm_pindex_t swb_index; 145 int swb_count; 146 daddr_t swb_pages[SWAP_META_PAGES]; 147 }; 148 149 static MALLOC_DEFINE(M_VMPGDATA, "vm_pgdata", "swap pager private data"); 150 static struct mtx sw_dev_mtx; 151 static TAILQ_HEAD(, swdevt) swtailq = TAILQ_HEAD_INITIALIZER(swtailq); 152 static struct swdevt *swdevhd; /* Allocate from here next */ 153 static int nswapdev; /* Number of swap devices */ 154 int swap_pager_avail; 155 static struct sx swdev_syscall_lock; /* serialize swap(on|off) */ 156 157 static vm_ooffset_t swap_total; 158 SYSCTL_QUAD(_vm, OID_AUTO, swap_total, CTLFLAG_RD, &swap_total, 0, 159 "Total amount of available swap storage."); 160 static vm_ooffset_t swap_reserved; 161 SYSCTL_QUAD(_vm, OID_AUTO, swap_reserved, CTLFLAG_RD, &swap_reserved, 0, 162 "Amount of swap storage needed to back all allocated anonymous memory."); 163 static int overcommit = 0; 164 SYSCTL_INT(_vm, OID_AUTO, overcommit, CTLFLAG_RW, &overcommit, 0, 165 "Configure virtual memory overcommit behavior. See tuning(7) " 166 "for details."); 167 static unsigned long swzone; 168 SYSCTL_ULONG(_vm, OID_AUTO, swzone, CTLFLAG_RD, &swzone, 0, 169 "Actual size of swap metadata zone"); 170 static unsigned long swap_maxpages; 171 SYSCTL_ULONG(_vm, OID_AUTO, swap_maxpages, CTLFLAG_RD, &swap_maxpages, 0, 172 "Maximum amount of swap supported"); 173 174 /* bits from overcommit */ 175 #define SWAP_RESERVE_FORCE_ON (1 << 0) 176 #define SWAP_RESERVE_RLIMIT_ON (1 << 1) 177 #define SWAP_RESERVE_ALLOW_NONWIRED (1 << 2) 178 179 int 180 swap_reserve(vm_ooffset_t incr) 181 { 182 183 return (swap_reserve_by_cred(incr, curthread->td_ucred)); 184 } 185 186 int 187 swap_reserve_by_cred(vm_ooffset_t incr, struct ucred *cred) 188 { 189 vm_ooffset_t r, s; 190 int res, error; 191 static int curfail; 192 static struct timeval lastfail; 193 struct uidinfo *uip; 194 195 uip = cred->cr_ruidinfo; 196 197 if (incr & PAGE_MASK) 198 panic("swap_reserve: & PAGE_MASK"); 199 200 #ifdef RACCT 201 if (racct_enable) { 202 PROC_LOCK(curproc); 203 error = racct_add(curproc, RACCT_SWAP, incr); 204 PROC_UNLOCK(curproc); 205 if (error != 0) 206 return (0); 207 } 208 #endif 209 210 res = 0; 211 mtx_lock(&sw_dev_mtx); 212 r = swap_reserved + incr; 213 if (overcommit & SWAP_RESERVE_ALLOW_NONWIRED) { 214 s = vm_cnt.v_page_count - vm_cnt.v_free_reserved - vm_cnt.v_wire_count; 215 s *= PAGE_SIZE; 216 } else 217 s = 0; 218 s += swap_total; 219 if ((overcommit & SWAP_RESERVE_FORCE_ON) == 0 || r <= s || 220 (error = priv_check(curthread, PRIV_VM_SWAP_NOQUOTA)) == 0) { 221 res = 1; 222 swap_reserved = r; 223 } 224 mtx_unlock(&sw_dev_mtx); 225 226 if (res) { 227 UIDINFO_VMSIZE_LOCK(uip); 228 if ((overcommit & SWAP_RESERVE_RLIMIT_ON) != 0 && 229 uip->ui_vmsize + incr > lim_cur(curthread, RLIMIT_SWAP) && 230 priv_check(curthread, PRIV_VM_SWAP_NORLIMIT)) 231 res = 0; 232 else 233 uip->ui_vmsize += incr; 234 UIDINFO_VMSIZE_UNLOCK(uip); 235 if (!res) { 236 mtx_lock(&sw_dev_mtx); 237 swap_reserved -= incr; 238 mtx_unlock(&sw_dev_mtx); 239 } 240 } 241 if (!res && ppsratecheck(&lastfail, &curfail, 1)) { 242 printf("uid %d, pid %d: swap reservation for %jd bytes failed\n", 243 uip->ui_uid, curproc->p_pid, incr); 244 } 245 246 #ifdef RACCT 247 if (!res) { 248 PROC_LOCK(curproc); 249 racct_sub(curproc, RACCT_SWAP, incr); 250 PROC_UNLOCK(curproc); 251 } 252 #endif 253 254 return (res); 255 } 256 257 void 258 swap_reserve_force(vm_ooffset_t incr) 259 { 260 struct uidinfo *uip; 261 262 mtx_lock(&sw_dev_mtx); 263 swap_reserved += incr; 264 mtx_unlock(&sw_dev_mtx); 265 266 #ifdef RACCT 267 PROC_LOCK(curproc); 268 racct_add_force(curproc, RACCT_SWAP, incr); 269 PROC_UNLOCK(curproc); 270 #endif 271 272 uip = curthread->td_ucred->cr_ruidinfo; 273 PROC_LOCK(curproc); 274 UIDINFO_VMSIZE_LOCK(uip); 275 uip->ui_vmsize += incr; 276 UIDINFO_VMSIZE_UNLOCK(uip); 277 PROC_UNLOCK(curproc); 278 } 279 280 void 281 swap_release(vm_ooffset_t decr) 282 { 283 struct ucred *cred; 284 285 PROC_LOCK(curproc); 286 cred = curthread->td_ucred; 287 swap_release_by_cred(decr, cred); 288 PROC_UNLOCK(curproc); 289 } 290 291 void 292 swap_release_by_cred(vm_ooffset_t decr, struct ucred *cred) 293 { 294 struct uidinfo *uip; 295 296 uip = cred->cr_ruidinfo; 297 298 if (decr & PAGE_MASK) 299 panic("swap_release: & PAGE_MASK"); 300 301 mtx_lock(&sw_dev_mtx); 302 if (swap_reserved < decr) 303 panic("swap_reserved < decr"); 304 swap_reserved -= decr; 305 mtx_unlock(&sw_dev_mtx); 306 307 UIDINFO_VMSIZE_LOCK(uip); 308 if (uip->ui_vmsize < decr) 309 printf("negative vmsize for uid = %d\n", uip->ui_uid); 310 uip->ui_vmsize -= decr; 311 UIDINFO_VMSIZE_UNLOCK(uip); 312 313 racct_sub_cred(cred, RACCT_SWAP, decr); 314 } 315 316 #define SWM_FREE 0x02 /* free, period */ 317 #define SWM_POP 0x04 /* pop out */ 318 319 int swap_pager_full = 2; /* swap space exhaustion (task killing) */ 320 static int swap_pager_almost_full = 1; /* swap space exhaustion (w/hysteresis)*/ 321 static int nsw_rcount; /* free read buffers */ 322 static int nsw_wcount_sync; /* limit write buffers / synchronous */ 323 static int nsw_wcount_async; /* limit write buffers / asynchronous */ 324 static int nsw_wcount_async_max;/* assigned maximum */ 325 static int nsw_cluster_max; /* maximum VOP I/O allowed */ 326 327 static int sysctl_swap_async_max(SYSCTL_HANDLER_ARGS); 328 SYSCTL_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"); 331 332 static struct swblock **swhash; 333 static int swhash_mask; 334 static struct mtx swhash_mtx; 335 336 static 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)]) 347 348 static struct pagerlst swap_pager_object_list[NOBJLISTS]; 349 static 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 */ 356 static vm_object_t 357 swap_pager_alloc(void *handle, vm_ooffset_t size, 358 vm_prot_t prot, vm_ooffset_t offset, struct ucred *); 359 static void swap_pager_dealloc(vm_object_t object); 360 static int swap_pager_getpages(vm_object_t, vm_page_t *, int, int *, 361 int *); 362 static int swap_pager_getpages_async(vm_object_t, vm_page_t *, int, int *, 363 int *, pgo_getpages_iodone_t, void *); 364 static void swap_pager_putpages(vm_object_t, vm_page_t *, int, boolean_t, int *); 365 static boolean_t 366 swap_pager_haspage(vm_object_t object, vm_pindex_t pindex, int *before, int *after); 367 static void swap_pager_init(void); 368 static void swap_pager_unswapped(vm_page_t); 369 static void swap_pager_swapoff(struct swdevt *sp); 370 371 struct 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 */ 389 static int dmmax; 390 static int nswap_lowat = 128; /* in pages, swap_pager_almost_full warn */ 391 static int nswap_hiwat = 512; /* in pages, swap_pager_almost_full warn */ 392 393 SYSCTL_INT(_vm, OID_AUTO, dmmax, 394 CTLFLAG_RD, &dmmax, 0, "Maximum size of a swap block"); 395 396 static void swp_sizecheck(void); 397 static void swp_pager_async_iodone(struct buf *bp); 398 static int swapongeom(struct thread *, struct vnode *); 399 static int swaponvp(struct thread *, struct vnode *, u_long); 400 static int swapoff_one(struct swdevt *sp, struct ucred *cred); 401 402 /* 403 * Swap bitmap functions 404 */ 405 static void swp_pager_freeswapspace(daddr_t blk, int npages); 406 static daddr_t swp_pager_getswapspace(int npages); 407 408 /* 409 * Metadata functions 410 */ 411 static struct swblock **swp_pager_hash(vm_object_t object, vm_pindex_t index); 412 static void swp_pager_meta_build(vm_object_t, vm_pindex_t, daddr_t); 413 static void swp_pager_meta_free(vm_object_t, vm_pindex_t, daddr_t); 414 static void swp_pager_meta_free_all(vm_object_t); 415 static 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 */ 428 static void 429 swp_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 */ 452 static struct swblock ** 453 swp_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 */ 478 static void 479 swap_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 */ 504 void 505 swap_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 } 583 584 static vm_object_t 585 swap_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 */ 617 static vm_object_t 618 swap_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 */ 657 static void 658 swap_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 */ 708 static daddr_t 709 swp_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; 740 done: 741 mtx_unlock(&sw_dev_mtx); 742 return (blk); 743 } 744 745 static int 746 swp_pager_isondev(daddr_t blk, struct swdevt *sp) 747 { 748 749 return (blk >= sp->sw_first && blk < sp->sw_end); 750 } 751 752 static void 753 swp_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 */ 784 static void 785 swp_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 */ 825 void 826 swap_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 */ 840 int 841 swap_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 */ 890 void 891 swap_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 within a reasonable 994 * distance. We do not try to restrict it to the swap device stripe 995 * (that is handled in getpages/putpages). It probably isn't worth 996 * doing here. 997 */ 998 static boolean_t 999 swap_pager_haspage(vm_object_t object, vm_pindex_t pindex, int *before, int *after) 1000 { 1001 daddr_t blk0; 1002 1003 VM_OBJECT_ASSERT_LOCKED(object); 1004 /* 1005 * do we have good backing store at the requested index ? 1006 */ 1007 blk0 = swp_pager_meta_ctl(object, pindex, 0); 1008 1009 if (blk0 == SWAPBLK_NONE) { 1010 if (before) 1011 *before = 0; 1012 if (after) 1013 *after = 0; 1014 return (FALSE); 1015 } 1016 1017 /* 1018 * find backwards-looking contiguous good backing store 1019 */ 1020 if (before != NULL) { 1021 int i; 1022 1023 for (i = 1; i < (SWB_NPAGES/2); ++i) { 1024 daddr_t blk; 1025 1026 if (i > pindex) 1027 break; 1028 blk = swp_pager_meta_ctl(object, pindex - i, 0); 1029 if (blk != blk0 - i) 1030 break; 1031 } 1032 *before = (i - 1); 1033 } 1034 1035 /* 1036 * find forward-looking contiguous good backing store 1037 */ 1038 if (after != NULL) { 1039 int i; 1040 1041 for (i = 1; i < (SWB_NPAGES/2); ++i) { 1042 daddr_t blk; 1043 1044 blk = swp_pager_meta_ctl(object, pindex + i, 0); 1045 if (blk != blk0 + i) 1046 break; 1047 } 1048 *after = (i - 1); 1049 } 1050 return (TRUE); 1051 } 1052 1053 /* 1054 * SWAP_PAGER_PAGE_UNSWAPPED() - remove swap backing store related to page 1055 * 1056 * This removes any associated swap backing store, whether valid or 1057 * not, from the page. 1058 * 1059 * This routine is typically called when a page is made dirty, at 1060 * which point any associated swap can be freed. MADV_FREE also 1061 * calls us in a special-case situation 1062 * 1063 * NOTE!!! If the page is clean and the swap was valid, the caller 1064 * should make the page dirty before calling this routine. This routine 1065 * does NOT change the m->dirty status of the page. Also: MADV_FREE 1066 * depends on it. 1067 * 1068 * This routine may not sleep. 1069 * 1070 * The object containing the page must be locked. 1071 */ 1072 static void 1073 swap_pager_unswapped(vm_page_t m) 1074 { 1075 1076 swp_pager_meta_ctl(m->object, m->pindex, SWM_FREE); 1077 } 1078 1079 /* 1080 * SWAP_PAGER_GETPAGES() - bring pages in from swap 1081 * 1082 * Attempt to retrieve (m, count) pages from backing store, but make 1083 * sure we retrieve at least m[reqpage]. We try to load in as large 1084 * a chunk surrounding m[reqpage] as is contiguous in swap and which 1085 * belongs to the same object. 1086 * 1087 * The code is designed for asynchronous operation and 1088 * immediate-notification of 'reqpage' but tends not to be 1089 * used that way. Please do not optimize-out this algorithmic 1090 * feature, I intend to improve on it in the future. 1091 * 1092 * The parent has a single vm_object_pip_add() reference prior to 1093 * calling us and we should return with the same. 1094 * 1095 * The parent has BUSY'd the pages. We should return with 'm' 1096 * left busy, but the others adjusted. 1097 */ 1098 static int 1099 swap_pager_getpages(vm_object_t object, vm_page_t *m, int count, int *rbehind, 1100 int *rahead) 1101 { 1102 struct buf *bp; 1103 daddr_t blk; 1104 1105 /* 1106 * Calculate range to retrieve. The pages have already been assigned 1107 * their swapblks. We require a *contiguous* range but we know it to 1108 * not span devices. If we do not supply it, bad things 1109 * happen. Note that blk, iblk & jblk can be SWAPBLK_NONE, but the 1110 * loops are set up such that the case(s) are handled implicitly. 1111 * 1112 * The swp_*() calls must be made with the object locked. 1113 */ 1114 blk = swp_pager_meta_ctl(m[0]->object, m[0]->pindex, 0); 1115 1116 if (blk == SWAPBLK_NONE) 1117 return (VM_PAGER_FAIL); 1118 1119 #ifdef INVARIANTS 1120 for (int i = 0; i < count; i++) 1121 KASSERT(blk + i == 1122 swp_pager_meta_ctl(m[i]->object, m[i]->pindex, 0), 1123 ("%s: range is not contiguous", __func__)); 1124 #endif 1125 1126 /* 1127 * Getpbuf() can sleep. 1128 */ 1129 VM_OBJECT_WUNLOCK(object); 1130 /* 1131 * Get a swap buffer header to perform the IO 1132 */ 1133 bp = getpbuf(&nsw_rcount); 1134 bp->b_flags |= B_PAGING; 1135 1136 bp->b_iocmd = BIO_READ; 1137 bp->b_iodone = swp_pager_async_iodone; 1138 bp->b_rcred = crhold(thread0.td_ucred); 1139 bp->b_wcred = crhold(thread0.td_ucred); 1140 bp->b_blkno = blk; 1141 bp->b_bcount = PAGE_SIZE * count; 1142 bp->b_bufsize = PAGE_SIZE * count; 1143 bp->b_npages = count; 1144 1145 VM_OBJECT_WLOCK(object); 1146 for (int i = 0; i < count; i++) { 1147 bp->b_pages[i] = m[i]; 1148 m[i]->oflags |= VPO_SWAPINPROG; 1149 } 1150 1151 PCPU_INC(cnt.v_swapin); 1152 PCPU_ADD(cnt.v_swappgsin, bp->b_npages); 1153 1154 /* 1155 * We still hold the lock on mreq, and our automatic completion routine 1156 * does not remove it. 1157 */ 1158 vm_object_pip_add(object, bp->b_npages); 1159 VM_OBJECT_WUNLOCK(object); 1160 1161 /* 1162 * perform the I/O. NOTE!!! bp cannot be considered valid after 1163 * this point because we automatically release it on completion. 1164 * Instead, we look at the one page we are interested in which we 1165 * still hold a lock on even through the I/O completion. 1166 * 1167 * The other pages in our m[] array are also released on completion, 1168 * so we cannot assume they are valid anymore either. 1169 * 1170 * NOTE: b_blkno is destroyed by the call to swapdev_strategy 1171 */ 1172 BUF_KERNPROC(bp); 1173 swp_pager_strategy(bp); 1174 1175 /* 1176 * wait for the page we want to complete. VPO_SWAPINPROG is always 1177 * cleared on completion. If an I/O error occurs, SWAPBLK_NONE 1178 * is set in the meta-data. 1179 */ 1180 VM_OBJECT_WLOCK(object); 1181 while ((m[0]->oflags & VPO_SWAPINPROG) != 0) { 1182 m[0]->oflags |= VPO_SWAPSLEEP; 1183 PCPU_INC(cnt.v_intrans); 1184 if (VM_OBJECT_SLEEP(object, &object->paging_in_progress, PSWP, 1185 "swread", hz * 20)) { 1186 printf( 1187 "swap_pager: indefinite wait buffer: bufobj: %p, blkno: %jd, size: %ld\n", 1188 bp->b_bufobj, (intmax_t)bp->b_blkno, bp->b_bcount); 1189 } 1190 } 1191 1192 /* 1193 * If we had an unrecoverable read error pages will not be valid. 1194 */ 1195 for (int i = 0; i < count; i++) 1196 if (m[i]->valid != VM_PAGE_BITS_ALL) 1197 return (VM_PAGER_ERROR); 1198 1199 if (rbehind) 1200 *rbehind = 0; 1201 if (rahead) 1202 *rahead = 0; 1203 1204 return (VM_PAGER_OK); 1205 1206 /* 1207 * A final note: in a low swap situation, we cannot deallocate swap 1208 * and mark a page dirty here because the caller is likely to mark 1209 * the page clean when we return, causing the page to possibly revert 1210 * to all-zero's later. 1211 */ 1212 } 1213 1214 /* 1215 * swap_pager_getpages_async(): 1216 * 1217 * Right now this is emulation of asynchronous operation on top of 1218 * swap_pager_getpages(). 1219 */ 1220 static int 1221 swap_pager_getpages_async(vm_object_t object, vm_page_t *m, int count, 1222 int *rbehind, int *rahead, pgo_getpages_iodone_t iodone, void *arg) 1223 { 1224 int r, error; 1225 1226 r = swap_pager_getpages(object, m, count, rbehind, rahead); 1227 VM_OBJECT_WUNLOCK(object); 1228 switch (r) { 1229 case VM_PAGER_OK: 1230 error = 0; 1231 break; 1232 case VM_PAGER_ERROR: 1233 error = EIO; 1234 break; 1235 case VM_PAGER_FAIL: 1236 error = EINVAL; 1237 break; 1238 default: 1239 panic("unhandled swap_pager_getpages() error %d", r); 1240 } 1241 (iodone)(arg, m, count, error); 1242 VM_OBJECT_WLOCK(object); 1243 1244 return (r); 1245 } 1246 1247 /* 1248 * swap_pager_putpages: 1249 * 1250 * Assign swap (if necessary) and initiate I/O on the specified pages. 1251 * 1252 * We support both OBJT_DEFAULT and OBJT_SWAP objects. DEFAULT objects 1253 * are automatically converted to SWAP objects. 1254 * 1255 * In a low memory situation we may block in VOP_STRATEGY(), but the new 1256 * vm_page reservation system coupled with properly written VFS devices 1257 * should ensure that no low-memory deadlock occurs. This is an area 1258 * which needs work. 1259 * 1260 * The parent has N vm_object_pip_add() references prior to 1261 * calling us and will remove references for rtvals[] that are 1262 * not set to VM_PAGER_PEND. We need to remove the rest on I/O 1263 * completion. 1264 * 1265 * The parent has soft-busy'd the pages it passes us and will unbusy 1266 * those whos rtvals[] entry is not set to VM_PAGER_PEND on return. 1267 * We need to unbusy the rest on I/O completion. 1268 */ 1269 static void 1270 swap_pager_putpages(vm_object_t object, vm_page_t *m, int count, 1271 int flags, int *rtvals) 1272 { 1273 int i, n; 1274 boolean_t sync; 1275 1276 if (count && m[0]->object != object) { 1277 panic("swap_pager_putpages: object mismatch %p/%p", 1278 object, 1279 m[0]->object 1280 ); 1281 } 1282 1283 /* 1284 * Step 1 1285 * 1286 * Turn object into OBJT_SWAP 1287 * check for bogus sysops 1288 * force sync if not pageout process 1289 */ 1290 if (object->type != OBJT_SWAP) 1291 swp_pager_meta_build(object, 0, SWAPBLK_NONE); 1292 VM_OBJECT_WUNLOCK(object); 1293 1294 n = 0; 1295 if (curproc != pageproc) 1296 sync = TRUE; 1297 else 1298 sync = (flags & VM_PAGER_PUT_SYNC) != 0; 1299 1300 /* 1301 * Step 2 1302 * 1303 * Assign swap blocks and issue I/O. We reallocate swap on the fly. 1304 * The page is left dirty until the pageout operation completes 1305 * successfully. 1306 */ 1307 for (i = 0; i < count; i += n) { 1308 int j; 1309 struct buf *bp; 1310 daddr_t blk; 1311 1312 /* 1313 * Maximum I/O size is limited by a number of factors. 1314 */ 1315 n = min(BLIST_MAX_ALLOC, count - i); 1316 n = min(n, nsw_cluster_max); 1317 1318 /* 1319 * Get biggest block of swap we can. If we fail, fall 1320 * back and try to allocate a smaller block. Don't go 1321 * overboard trying to allocate space if it would overly 1322 * fragment swap. 1323 */ 1324 while ( 1325 (blk = swp_pager_getswapspace(n)) == SWAPBLK_NONE && 1326 n > 4 1327 ) { 1328 n >>= 1; 1329 } 1330 if (blk == SWAPBLK_NONE) { 1331 for (j = 0; j < n; ++j) 1332 rtvals[i+j] = VM_PAGER_FAIL; 1333 continue; 1334 } 1335 1336 /* 1337 * All I/O parameters have been satisfied, build the I/O 1338 * request and assign the swap space. 1339 */ 1340 if (sync == TRUE) { 1341 bp = getpbuf(&nsw_wcount_sync); 1342 } else { 1343 bp = getpbuf(&nsw_wcount_async); 1344 bp->b_flags = B_ASYNC; 1345 } 1346 bp->b_flags |= B_PAGING; 1347 bp->b_iocmd = BIO_WRITE; 1348 1349 bp->b_rcred = crhold(thread0.td_ucred); 1350 bp->b_wcred = crhold(thread0.td_ucred); 1351 bp->b_bcount = PAGE_SIZE * n; 1352 bp->b_bufsize = PAGE_SIZE * n; 1353 bp->b_blkno = blk; 1354 1355 VM_OBJECT_WLOCK(object); 1356 for (j = 0; j < n; ++j) { 1357 vm_page_t mreq = m[i+j]; 1358 1359 swp_pager_meta_build( 1360 mreq->object, 1361 mreq->pindex, 1362 blk + j 1363 ); 1364 vm_page_dirty(mreq); 1365 mreq->oflags |= VPO_SWAPINPROG; 1366 bp->b_pages[j] = mreq; 1367 } 1368 VM_OBJECT_WUNLOCK(object); 1369 bp->b_npages = n; 1370 /* 1371 * Must set dirty range for NFS to work. 1372 */ 1373 bp->b_dirtyoff = 0; 1374 bp->b_dirtyend = bp->b_bcount; 1375 1376 PCPU_INC(cnt.v_swapout); 1377 PCPU_ADD(cnt.v_swappgsout, bp->b_npages); 1378 1379 /* 1380 * We unconditionally set rtvals[] to VM_PAGER_PEND so that we 1381 * can call the async completion routine at the end of a 1382 * synchronous I/O operation. Otherwise, our caller would 1383 * perform duplicate unbusy and wakeup operations on the page 1384 * and object, respectively. 1385 */ 1386 for (j = 0; j < n; j++) 1387 rtvals[i + j] = VM_PAGER_PEND; 1388 1389 /* 1390 * asynchronous 1391 * 1392 * NOTE: b_blkno is destroyed by the call to swapdev_strategy 1393 */ 1394 if (sync == FALSE) { 1395 bp->b_iodone = swp_pager_async_iodone; 1396 BUF_KERNPROC(bp); 1397 swp_pager_strategy(bp); 1398 continue; 1399 } 1400 1401 /* 1402 * synchronous 1403 * 1404 * NOTE: b_blkno is destroyed by the call to swapdev_strategy 1405 */ 1406 bp->b_iodone = bdone; 1407 swp_pager_strategy(bp); 1408 1409 /* 1410 * Wait for the sync I/O to complete. 1411 */ 1412 bwait(bp, PVM, "swwrt"); 1413 1414 /* 1415 * Now that we are through with the bp, we can call the 1416 * normal async completion, which frees everything up. 1417 */ 1418 swp_pager_async_iodone(bp); 1419 } 1420 VM_OBJECT_WLOCK(object); 1421 } 1422 1423 /* 1424 * swp_pager_async_iodone: 1425 * 1426 * Completion routine for asynchronous reads and writes from/to swap. 1427 * Also called manually by synchronous code to finish up a bp. 1428 * 1429 * This routine may not sleep. 1430 */ 1431 static void 1432 swp_pager_async_iodone(struct buf *bp) 1433 { 1434 int i; 1435 vm_object_t object = NULL; 1436 1437 /* 1438 * report error 1439 */ 1440 if (bp->b_ioflags & BIO_ERROR) { 1441 printf( 1442 "swap_pager: I/O error - %s failed; blkno %ld," 1443 "size %ld, error %d\n", 1444 ((bp->b_iocmd == BIO_READ) ? "pagein" : "pageout"), 1445 (long)bp->b_blkno, 1446 (long)bp->b_bcount, 1447 bp->b_error 1448 ); 1449 } 1450 1451 /* 1452 * remove the mapping for kernel virtual 1453 */ 1454 if (buf_mapped(bp)) 1455 pmap_qremove((vm_offset_t)bp->b_data, bp->b_npages); 1456 else 1457 bp->b_data = bp->b_kvabase; 1458 1459 if (bp->b_npages) { 1460 object = bp->b_pages[0]->object; 1461 VM_OBJECT_WLOCK(object); 1462 } 1463 1464 /* 1465 * cleanup pages. If an error occurs writing to swap, we are in 1466 * very serious trouble. If it happens to be a disk error, though, 1467 * we may be able to recover by reassigning the swap later on. So 1468 * in this case we remove the m->swapblk assignment for the page 1469 * but do not free it in the rlist. The errornous block(s) are thus 1470 * never reallocated as swap. Redirty the page and continue. 1471 */ 1472 for (i = 0; i < bp->b_npages; ++i) { 1473 vm_page_t m = bp->b_pages[i]; 1474 1475 m->oflags &= ~VPO_SWAPINPROG; 1476 if (m->oflags & VPO_SWAPSLEEP) { 1477 m->oflags &= ~VPO_SWAPSLEEP; 1478 wakeup(&object->paging_in_progress); 1479 } 1480 1481 if (bp->b_ioflags & BIO_ERROR) { 1482 /* 1483 * If an error occurs I'd love to throw the swapblk 1484 * away without freeing it back to swapspace, so it 1485 * can never be used again. But I can't from an 1486 * interrupt. 1487 */ 1488 if (bp->b_iocmd == BIO_READ) { 1489 /* 1490 * NOTE: for reads, m->dirty will probably 1491 * be overridden by the original caller of 1492 * getpages so don't play cute tricks here. 1493 */ 1494 m->valid = 0; 1495 } else { 1496 /* 1497 * If a write error occurs, reactivate page 1498 * so it doesn't clog the inactive list, 1499 * then finish the I/O. 1500 */ 1501 vm_page_dirty(m); 1502 vm_page_lock(m); 1503 vm_page_activate(m); 1504 vm_page_unlock(m); 1505 vm_page_sunbusy(m); 1506 } 1507 } else if (bp->b_iocmd == BIO_READ) { 1508 /* 1509 * NOTE: for reads, m->dirty will probably be 1510 * overridden by the original caller of getpages so 1511 * we cannot set them in order to free the underlying 1512 * swap in a low-swap situation. I don't think we'd 1513 * want to do that anyway, but it was an optimization 1514 * that existed in the old swapper for a time before 1515 * it got ripped out due to precisely this problem. 1516 */ 1517 KASSERT(!pmap_page_is_mapped(m), 1518 ("swp_pager_async_iodone: page %p is mapped", m)); 1519 KASSERT(m->dirty == 0, 1520 ("swp_pager_async_iodone: page %p is dirty", m)); 1521 m->valid = VM_PAGE_BITS_ALL; 1522 } else { 1523 /* 1524 * For write success, clear the dirty 1525 * status, then finish the I/O ( which decrements the 1526 * busy count and possibly wakes waiter's up ). 1527 */ 1528 KASSERT(!pmap_page_is_write_mapped(m), 1529 ("swp_pager_async_iodone: page %p is not write" 1530 " protected", m)); 1531 vm_page_undirty(m); 1532 vm_page_sunbusy(m); 1533 if (vm_page_count_severe()) { 1534 vm_page_lock(m); 1535 vm_page_try_to_cache(m); 1536 vm_page_unlock(m); 1537 } 1538 } 1539 } 1540 1541 /* 1542 * adjust pip. NOTE: the original parent may still have its own 1543 * pip refs on the object. 1544 */ 1545 if (object != NULL) { 1546 vm_object_pip_wakeupn(object, bp->b_npages); 1547 VM_OBJECT_WUNLOCK(object); 1548 } 1549 1550 /* 1551 * swapdev_strategy() manually sets b_vp and b_bufobj before calling 1552 * bstrategy(). Set them back to NULL now we're done with it, or we'll 1553 * trigger a KASSERT in relpbuf(). 1554 */ 1555 if (bp->b_vp) { 1556 bp->b_vp = NULL; 1557 bp->b_bufobj = NULL; 1558 } 1559 /* 1560 * release the physical I/O buffer 1561 */ 1562 relpbuf( 1563 bp, 1564 ((bp->b_iocmd == BIO_READ) ? &nsw_rcount : 1565 ((bp->b_flags & B_ASYNC) ? 1566 &nsw_wcount_async : 1567 &nsw_wcount_sync 1568 ) 1569 ) 1570 ); 1571 } 1572 1573 /* 1574 * swap_pager_isswapped: 1575 * 1576 * Return 1 if at least one page in the given object is paged 1577 * out to the given swap device. 1578 * 1579 * This routine may not sleep. 1580 */ 1581 int 1582 swap_pager_isswapped(vm_object_t object, struct swdevt *sp) 1583 { 1584 daddr_t index = 0; 1585 int bcount; 1586 int i; 1587 1588 VM_OBJECT_ASSERT_WLOCKED(object); 1589 if (object->type != OBJT_SWAP) 1590 return (0); 1591 1592 mtx_lock(&swhash_mtx); 1593 for (bcount = 0; bcount < object->un_pager.swp.swp_bcount; bcount++) { 1594 struct swblock *swap; 1595 1596 if ((swap = *swp_pager_hash(object, index)) != NULL) { 1597 for (i = 0; i < SWAP_META_PAGES; ++i) { 1598 if (swp_pager_isondev(swap->swb_pages[i], sp)) { 1599 mtx_unlock(&swhash_mtx); 1600 return (1); 1601 } 1602 } 1603 } 1604 index += SWAP_META_PAGES; 1605 } 1606 mtx_unlock(&swhash_mtx); 1607 return (0); 1608 } 1609 1610 /* 1611 * SWP_PAGER_FORCE_PAGEIN() - force a swap block to be paged in 1612 * 1613 * This routine dissociates the page at the given index within a 1614 * swap block from its backing store, paging it in if necessary. 1615 * If the page is paged in, it is placed in the inactive queue, 1616 * since it had its backing store ripped out from under it. 1617 * We also attempt to swap in all other pages in the swap block, 1618 * we only guarantee that the one at the specified index is 1619 * paged in. 1620 * 1621 * XXX - The code to page the whole block in doesn't work, so we 1622 * revert to the one-by-one behavior for now. Sigh. 1623 */ 1624 static inline void 1625 swp_pager_force_pagein(vm_object_t object, vm_pindex_t pindex) 1626 { 1627 vm_page_t m; 1628 1629 vm_object_pip_add(object, 1); 1630 m = vm_page_grab(object, pindex, VM_ALLOC_NORMAL); 1631 if (m->valid == VM_PAGE_BITS_ALL) { 1632 vm_object_pip_wakeup(object); 1633 vm_page_dirty(m); 1634 vm_page_lock(m); 1635 vm_page_activate(m); 1636 vm_page_unlock(m); 1637 vm_page_xunbusy(m); 1638 vm_pager_page_unswapped(m); 1639 return; 1640 } 1641 1642 if (swap_pager_getpages(object, &m, 1, NULL, NULL) != VM_PAGER_OK) 1643 panic("swap_pager_force_pagein: read from swap failed");/*XXX*/ 1644 vm_object_pip_wakeup(object); 1645 vm_page_dirty(m); 1646 vm_page_lock(m); 1647 vm_page_deactivate(m); 1648 vm_page_unlock(m); 1649 vm_page_xunbusy(m); 1650 vm_pager_page_unswapped(m); 1651 } 1652 1653 /* 1654 * swap_pager_swapoff: 1655 * 1656 * Page in all of the pages that have been paged out to the 1657 * given device. The corresponding blocks in the bitmap must be 1658 * marked as allocated and the device must be flagged SW_CLOSING. 1659 * There may be no processes swapped out to the device. 1660 * 1661 * This routine may block. 1662 */ 1663 static void 1664 swap_pager_swapoff(struct swdevt *sp) 1665 { 1666 struct swblock *swap; 1667 int i, j, retries; 1668 1669 sx_assert(&swdev_syscall_lock, SA_XLOCKED); 1670 1671 retries = 0; 1672 full_rescan: 1673 mtx_lock(&swhash_mtx); 1674 for (i = 0; i <= swhash_mask; i++) { /* '<=' is correct here */ 1675 restart: 1676 for (swap = swhash[i]; swap != NULL; swap = swap->swb_hnext) { 1677 vm_object_t object = swap->swb_object; 1678 vm_pindex_t pindex = swap->swb_index; 1679 for (j = 0; j < SWAP_META_PAGES; ++j) { 1680 if (swp_pager_isondev(swap->swb_pages[j], sp)) { 1681 /* avoid deadlock */ 1682 if (!VM_OBJECT_TRYWLOCK(object)) { 1683 break; 1684 } else { 1685 mtx_unlock(&swhash_mtx); 1686 swp_pager_force_pagein(object, 1687 pindex + j); 1688 VM_OBJECT_WUNLOCK(object); 1689 mtx_lock(&swhash_mtx); 1690 goto restart; 1691 } 1692 } 1693 } 1694 } 1695 } 1696 mtx_unlock(&swhash_mtx); 1697 if (sp->sw_used) { 1698 /* 1699 * Objects may be locked or paging to the device being 1700 * removed, so we will miss their pages and need to 1701 * make another pass. We have marked this device as 1702 * SW_CLOSING, so the activity should finish soon. 1703 */ 1704 retries++; 1705 if (retries > 100) { 1706 panic("swapoff: failed to locate %d swap blocks", 1707 sp->sw_used); 1708 } 1709 pause("swpoff", hz / 20); 1710 goto full_rescan; 1711 } 1712 } 1713 1714 /************************************************************************ 1715 * SWAP META DATA * 1716 ************************************************************************ 1717 * 1718 * These routines manipulate the swap metadata stored in the 1719 * OBJT_SWAP object. 1720 * 1721 * Swap metadata is implemented with a global hash and not directly 1722 * linked into the object. Instead the object simply contains 1723 * appropriate tracking counters. 1724 */ 1725 1726 /* 1727 * SWP_PAGER_META_BUILD() - add swap block to swap meta data for object 1728 * 1729 * We first convert the object to a swap object if it is a default 1730 * object. 1731 * 1732 * The specified swapblk is added to the object's swap metadata. If 1733 * the swapblk is not valid, it is freed instead. Any previously 1734 * assigned swapblk is freed. 1735 */ 1736 static void 1737 swp_pager_meta_build(vm_object_t object, vm_pindex_t pindex, daddr_t swapblk) 1738 { 1739 static volatile int exhausted; 1740 struct swblock *swap; 1741 struct swblock **pswap; 1742 int idx; 1743 1744 VM_OBJECT_ASSERT_WLOCKED(object); 1745 /* 1746 * Convert default object to swap object if necessary 1747 */ 1748 if (object->type != OBJT_SWAP) { 1749 object->type = OBJT_SWAP; 1750 object->un_pager.swp.swp_bcount = 0; 1751 KASSERT(object->handle == NULL, ("default pager with handle")); 1752 } 1753 1754 /* 1755 * Locate hash entry. If not found create, but if we aren't adding 1756 * anything just return. If we run out of space in the map we wait 1757 * and, since the hash table may have changed, retry. 1758 */ 1759 retry: 1760 mtx_lock(&swhash_mtx); 1761 pswap = swp_pager_hash(object, pindex); 1762 1763 if ((swap = *pswap) == NULL) { 1764 int i; 1765 1766 if (swapblk == SWAPBLK_NONE) 1767 goto done; 1768 1769 swap = *pswap = uma_zalloc(swap_zone, M_NOWAIT | 1770 (curproc == pageproc ? M_USE_RESERVE : 0)); 1771 if (swap == NULL) { 1772 mtx_unlock(&swhash_mtx); 1773 VM_OBJECT_WUNLOCK(object); 1774 if (uma_zone_exhausted(swap_zone)) { 1775 if (atomic_cmpset_int(&exhausted, 0, 1)) 1776 printf("swap zone exhausted, " 1777 "increase kern.maxswzone\n"); 1778 vm_pageout_oom(VM_OOM_SWAPZ); 1779 pause("swzonex", 10); 1780 } else 1781 VM_WAIT; 1782 VM_OBJECT_WLOCK(object); 1783 goto retry; 1784 } 1785 1786 if (atomic_cmpset_int(&exhausted, 1, 0)) 1787 printf("swap zone ok\n"); 1788 1789 swap->swb_hnext = NULL; 1790 swap->swb_object = object; 1791 swap->swb_index = pindex & ~(vm_pindex_t)SWAP_META_MASK; 1792 swap->swb_count = 0; 1793 1794 ++object->un_pager.swp.swp_bcount; 1795 1796 for (i = 0; i < SWAP_META_PAGES; ++i) 1797 swap->swb_pages[i] = SWAPBLK_NONE; 1798 } 1799 1800 /* 1801 * Delete prior contents of metadata 1802 */ 1803 idx = pindex & SWAP_META_MASK; 1804 1805 if (swap->swb_pages[idx] != SWAPBLK_NONE) { 1806 swp_pager_freeswapspace(swap->swb_pages[idx], 1); 1807 --swap->swb_count; 1808 } 1809 1810 /* 1811 * Enter block into metadata 1812 */ 1813 swap->swb_pages[idx] = swapblk; 1814 if (swapblk != SWAPBLK_NONE) 1815 ++swap->swb_count; 1816 done: 1817 mtx_unlock(&swhash_mtx); 1818 } 1819 1820 /* 1821 * SWP_PAGER_META_FREE() - free a range of blocks in the object's swap metadata 1822 * 1823 * The requested range of blocks is freed, with any associated swap 1824 * returned to the swap bitmap. 1825 * 1826 * This routine will free swap metadata structures as they are cleaned 1827 * out. This routine does *NOT* operate on swap metadata associated 1828 * with resident pages. 1829 */ 1830 static void 1831 swp_pager_meta_free(vm_object_t object, vm_pindex_t index, daddr_t count) 1832 { 1833 1834 VM_OBJECT_ASSERT_LOCKED(object); 1835 if (object->type != OBJT_SWAP) 1836 return; 1837 1838 while (count > 0) { 1839 struct swblock **pswap; 1840 struct swblock *swap; 1841 1842 mtx_lock(&swhash_mtx); 1843 pswap = swp_pager_hash(object, index); 1844 1845 if ((swap = *pswap) != NULL) { 1846 daddr_t v = swap->swb_pages[index & SWAP_META_MASK]; 1847 1848 if (v != SWAPBLK_NONE) { 1849 swp_pager_freeswapspace(v, 1); 1850 swap->swb_pages[index & SWAP_META_MASK] = 1851 SWAPBLK_NONE; 1852 if (--swap->swb_count == 0) { 1853 *pswap = swap->swb_hnext; 1854 uma_zfree(swap_zone, swap); 1855 --object->un_pager.swp.swp_bcount; 1856 } 1857 } 1858 --count; 1859 ++index; 1860 } else { 1861 int n = SWAP_META_PAGES - (index & SWAP_META_MASK); 1862 count -= n; 1863 index += n; 1864 } 1865 mtx_unlock(&swhash_mtx); 1866 } 1867 } 1868 1869 /* 1870 * SWP_PAGER_META_FREE_ALL() - destroy all swap metadata associated with object 1871 * 1872 * This routine locates and destroys all swap metadata associated with 1873 * an object. 1874 */ 1875 static void 1876 swp_pager_meta_free_all(vm_object_t object) 1877 { 1878 daddr_t index = 0; 1879 1880 VM_OBJECT_ASSERT_WLOCKED(object); 1881 if (object->type != OBJT_SWAP) 1882 return; 1883 1884 while (object->un_pager.swp.swp_bcount) { 1885 struct swblock **pswap; 1886 struct swblock *swap; 1887 1888 mtx_lock(&swhash_mtx); 1889 pswap = swp_pager_hash(object, index); 1890 if ((swap = *pswap) != NULL) { 1891 int i; 1892 1893 for (i = 0; i < SWAP_META_PAGES; ++i) { 1894 daddr_t v = swap->swb_pages[i]; 1895 if (v != SWAPBLK_NONE) { 1896 --swap->swb_count; 1897 swp_pager_freeswapspace(v, 1); 1898 } 1899 } 1900 if (swap->swb_count != 0) 1901 panic("swap_pager_meta_free_all: swb_count != 0"); 1902 *pswap = swap->swb_hnext; 1903 uma_zfree(swap_zone, swap); 1904 --object->un_pager.swp.swp_bcount; 1905 } 1906 mtx_unlock(&swhash_mtx); 1907 index += SWAP_META_PAGES; 1908 } 1909 } 1910 1911 /* 1912 * SWP_PAGER_METACTL() - misc control of swap and vm_page_t meta data. 1913 * 1914 * This routine is capable of looking up, popping, or freeing 1915 * swapblk assignments in the swap meta data or in the vm_page_t. 1916 * The routine typically returns the swapblk being looked-up, or popped, 1917 * or SWAPBLK_NONE if the block was freed, or SWAPBLK_NONE if the block 1918 * was invalid. This routine will automatically free any invalid 1919 * meta-data swapblks. 1920 * 1921 * It is not possible to store invalid swapblks in the swap meta data 1922 * (other then a literal 'SWAPBLK_NONE'), so we don't bother checking. 1923 * 1924 * When acting on a busy resident page and paging is in progress, we 1925 * have to wait until paging is complete but otherwise can act on the 1926 * busy page. 1927 * 1928 * SWM_FREE remove and free swap block from metadata 1929 * SWM_POP remove from meta data but do not free.. pop it out 1930 */ 1931 static daddr_t 1932 swp_pager_meta_ctl(vm_object_t object, vm_pindex_t pindex, int flags) 1933 { 1934 struct swblock **pswap; 1935 struct swblock *swap; 1936 daddr_t r1; 1937 int idx; 1938 1939 VM_OBJECT_ASSERT_LOCKED(object); 1940 /* 1941 * The meta data only exists of the object is OBJT_SWAP 1942 * and even then might not be allocated yet. 1943 */ 1944 if (object->type != OBJT_SWAP) 1945 return (SWAPBLK_NONE); 1946 1947 r1 = SWAPBLK_NONE; 1948 mtx_lock(&swhash_mtx); 1949 pswap = swp_pager_hash(object, pindex); 1950 1951 if ((swap = *pswap) != NULL) { 1952 idx = pindex & SWAP_META_MASK; 1953 r1 = swap->swb_pages[idx]; 1954 1955 if (r1 != SWAPBLK_NONE) { 1956 if (flags & SWM_FREE) { 1957 swp_pager_freeswapspace(r1, 1); 1958 r1 = SWAPBLK_NONE; 1959 } 1960 if (flags & (SWM_FREE|SWM_POP)) { 1961 swap->swb_pages[idx] = SWAPBLK_NONE; 1962 if (--swap->swb_count == 0) { 1963 *pswap = swap->swb_hnext; 1964 uma_zfree(swap_zone, swap); 1965 --object->un_pager.swp.swp_bcount; 1966 } 1967 } 1968 } 1969 } 1970 mtx_unlock(&swhash_mtx); 1971 return (r1); 1972 } 1973 1974 /* 1975 * System call swapon(name) enables swapping on device name, 1976 * which must be in the swdevsw. Return EBUSY 1977 * if already swapping on this device. 1978 */ 1979 #ifndef _SYS_SYSPROTO_H_ 1980 struct swapon_args { 1981 char *name; 1982 }; 1983 #endif 1984 1985 /* 1986 * MPSAFE 1987 */ 1988 /* ARGSUSED */ 1989 int 1990 sys_swapon(struct thread *td, struct swapon_args *uap) 1991 { 1992 struct vattr attr; 1993 struct vnode *vp; 1994 struct nameidata nd; 1995 int error; 1996 1997 error = priv_check(td, PRIV_SWAPON); 1998 if (error) 1999 return (error); 2000 2001 sx_xlock(&swdev_syscall_lock); 2002 2003 /* 2004 * Swap metadata may not fit in the KVM if we have physical 2005 * memory of >1GB. 2006 */ 2007 if (swap_zone == NULL) { 2008 error = ENOMEM; 2009 goto done; 2010 } 2011 2012 NDINIT(&nd, LOOKUP, ISOPEN | FOLLOW | AUDITVNODE1, UIO_USERSPACE, 2013 uap->name, td); 2014 error = namei(&nd); 2015 if (error) 2016 goto done; 2017 2018 NDFREE(&nd, NDF_ONLY_PNBUF); 2019 vp = nd.ni_vp; 2020 2021 if (vn_isdisk(vp, &error)) { 2022 error = swapongeom(td, vp); 2023 } else if (vp->v_type == VREG && 2024 (vp->v_mount->mnt_vfc->vfc_flags & VFCF_NETWORK) != 0 && 2025 (error = VOP_GETATTR(vp, &attr, td->td_ucred)) == 0) { 2026 /* 2027 * Allow direct swapping to NFS regular files in the same 2028 * way that nfs_mountroot() sets up diskless swapping. 2029 */ 2030 error = swaponvp(td, vp, attr.va_size / DEV_BSIZE); 2031 } 2032 2033 if (error) 2034 vrele(vp); 2035 done: 2036 sx_xunlock(&swdev_syscall_lock); 2037 return (error); 2038 } 2039 2040 /* 2041 * Check that the total amount of swap currently configured does not 2042 * exceed half the theoretical maximum. If it does, print a warning 2043 * message and return -1; otherwise, return 0. 2044 */ 2045 static int 2046 swapon_check_swzone(unsigned long npages) 2047 { 2048 unsigned long maxpages; 2049 2050 /* absolute maximum we can handle assuming 100% efficiency */ 2051 maxpages = uma_zone_get_max(swap_zone) * SWAP_META_PAGES; 2052 2053 /* recommend using no more than half that amount */ 2054 if (npages > maxpages / 2) { 2055 printf("warning: total configured swap (%lu pages) " 2056 "exceeds maximum recommended amount (%lu pages).\n", 2057 npages, maxpages / 2); 2058 printf("warning: increase kern.maxswzone " 2059 "or reduce amount of swap.\n"); 2060 return (-1); 2061 } 2062 return (0); 2063 } 2064 2065 static void 2066 swaponsomething(struct vnode *vp, void *id, u_long nblks, 2067 sw_strategy_t *strategy, sw_close_t *close, dev_t dev, int flags) 2068 { 2069 struct swdevt *sp, *tsp; 2070 swblk_t dvbase; 2071 u_long mblocks; 2072 2073 /* 2074 * nblks is in DEV_BSIZE'd chunks, convert to PAGE_SIZE'd chunks. 2075 * First chop nblks off to page-align it, then convert. 2076 * 2077 * sw->sw_nblks is in page-sized chunks now too. 2078 */ 2079 nblks &= ~(ctodb(1) - 1); 2080 nblks = dbtoc(nblks); 2081 2082 /* 2083 * If we go beyond this, we get overflows in the radix 2084 * tree bitmap code. 2085 */ 2086 mblocks = 0x40000000 / BLIST_META_RADIX; 2087 if (nblks > mblocks) { 2088 printf( 2089 "WARNING: reducing swap size to maximum of %luMB per unit\n", 2090 mblocks / 1024 / 1024 * PAGE_SIZE); 2091 nblks = mblocks; 2092 } 2093 2094 sp = malloc(sizeof *sp, M_VMPGDATA, M_WAITOK | M_ZERO); 2095 sp->sw_vp = vp; 2096 sp->sw_id = id; 2097 sp->sw_dev = dev; 2098 sp->sw_flags = 0; 2099 sp->sw_nblks = nblks; 2100 sp->sw_used = 0; 2101 sp->sw_strategy = strategy; 2102 sp->sw_close = close; 2103 sp->sw_flags = flags; 2104 2105 sp->sw_blist = blist_create(nblks, M_WAITOK); 2106 /* 2107 * Do not free the first two block in order to avoid overwriting 2108 * any bsd label at the front of the partition 2109 */ 2110 blist_free(sp->sw_blist, 2, nblks - 2); 2111 2112 dvbase = 0; 2113 mtx_lock(&sw_dev_mtx); 2114 TAILQ_FOREACH(tsp, &swtailq, sw_list) { 2115 if (tsp->sw_end >= dvbase) { 2116 /* 2117 * We put one uncovered page between the devices 2118 * in order to definitively prevent any cross-device 2119 * I/O requests 2120 */ 2121 dvbase = tsp->sw_end + 1; 2122 } 2123 } 2124 sp->sw_first = dvbase; 2125 sp->sw_end = dvbase + nblks; 2126 TAILQ_INSERT_TAIL(&swtailq, sp, sw_list); 2127 nswapdev++; 2128 swap_pager_avail += nblks; 2129 swap_total += (vm_ooffset_t)nblks * PAGE_SIZE; 2130 swapon_check_swzone(swap_total / PAGE_SIZE); 2131 swp_sizecheck(); 2132 mtx_unlock(&sw_dev_mtx); 2133 } 2134 2135 /* 2136 * SYSCALL: swapoff(devname) 2137 * 2138 * Disable swapping on the given device. 2139 * 2140 * XXX: Badly designed system call: it should use a device index 2141 * rather than filename as specification. We keep sw_vp around 2142 * only to make this work. 2143 */ 2144 #ifndef _SYS_SYSPROTO_H_ 2145 struct swapoff_args { 2146 char *name; 2147 }; 2148 #endif 2149 2150 /* 2151 * MPSAFE 2152 */ 2153 /* ARGSUSED */ 2154 int 2155 sys_swapoff(struct thread *td, struct swapoff_args *uap) 2156 { 2157 struct vnode *vp; 2158 struct nameidata nd; 2159 struct swdevt *sp; 2160 int error; 2161 2162 error = priv_check(td, PRIV_SWAPOFF); 2163 if (error) 2164 return (error); 2165 2166 sx_xlock(&swdev_syscall_lock); 2167 2168 NDINIT(&nd, LOOKUP, FOLLOW | AUDITVNODE1, UIO_USERSPACE, uap->name, 2169 td); 2170 error = namei(&nd); 2171 if (error) 2172 goto done; 2173 NDFREE(&nd, NDF_ONLY_PNBUF); 2174 vp = nd.ni_vp; 2175 2176 mtx_lock(&sw_dev_mtx); 2177 TAILQ_FOREACH(sp, &swtailq, sw_list) { 2178 if (sp->sw_vp == vp) 2179 break; 2180 } 2181 mtx_unlock(&sw_dev_mtx); 2182 if (sp == NULL) { 2183 error = EINVAL; 2184 goto done; 2185 } 2186 error = swapoff_one(sp, td->td_ucred); 2187 done: 2188 sx_xunlock(&swdev_syscall_lock); 2189 return (error); 2190 } 2191 2192 static int 2193 swapoff_one(struct swdevt *sp, struct ucred *cred) 2194 { 2195 u_long nblks, dvbase; 2196 #ifdef MAC 2197 int error; 2198 #endif 2199 2200 sx_assert(&swdev_syscall_lock, SA_XLOCKED); 2201 #ifdef MAC 2202 (void) vn_lock(sp->sw_vp, LK_EXCLUSIVE | LK_RETRY); 2203 error = mac_system_check_swapoff(cred, sp->sw_vp); 2204 (void) VOP_UNLOCK(sp->sw_vp, 0); 2205 if (error != 0) 2206 return (error); 2207 #endif 2208 nblks = sp->sw_nblks; 2209 2210 /* 2211 * We can turn off this swap device safely only if the 2212 * available virtual memory in the system will fit the amount 2213 * of data we will have to page back in, plus an epsilon so 2214 * the system doesn't become critically low on swap space. 2215 */ 2216 if (vm_cnt.v_free_count + vm_cnt.v_cache_count + swap_pager_avail < 2217 nblks + nswap_lowat) { 2218 return (ENOMEM); 2219 } 2220 2221 /* 2222 * Prevent further allocations on this device. 2223 */ 2224 mtx_lock(&sw_dev_mtx); 2225 sp->sw_flags |= SW_CLOSING; 2226 for (dvbase = 0; dvbase < sp->sw_end; dvbase += dmmax) { 2227 swap_pager_avail -= blist_fill(sp->sw_blist, 2228 dvbase, dmmax); 2229 } 2230 swap_total -= (vm_ooffset_t)nblks * PAGE_SIZE; 2231 mtx_unlock(&sw_dev_mtx); 2232 2233 /* 2234 * Page in the contents of the device and close it. 2235 */ 2236 swap_pager_swapoff(sp); 2237 2238 sp->sw_close(curthread, sp); 2239 mtx_lock(&sw_dev_mtx); 2240 sp->sw_id = NULL; 2241 TAILQ_REMOVE(&swtailq, sp, sw_list); 2242 nswapdev--; 2243 if (nswapdev == 0) { 2244 swap_pager_full = 2; 2245 swap_pager_almost_full = 1; 2246 } 2247 if (swdevhd == sp) 2248 swdevhd = NULL; 2249 mtx_unlock(&sw_dev_mtx); 2250 blist_destroy(sp->sw_blist); 2251 free(sp, M_VMPGDATA); 2252 return (0); 2253 } 2254 2255 void 2256 swapoff_all(void) 2257 { 2258 struct swdevt *sp, *spt; 2259 const char *devname; 2260 int error; 2261 2262 sx_xlock(&swdev_syscall_lock); 2263 2264 mtx_lock(&sw_dev_mtx); 2265 TAILQ_FOREACH_SAFE(sp, &swtailq, sw_list, spt) { 2266 mtx_unlock(&sw_dev_mtx); 2267 if (vn_isdisk(sp->sw_vp, NULL)) 2268 devname = devtoname(sp->sw_vp->v_rdev); 2269 else 2270 devname = "[file]"; 2271 error = swapoff_one(sp, thread0.td_ucred); 2272 if (error != 0) { 2273 printf("Cannot remove swap device %s (error=%d), " 2274 "skipping.\n", devname, error); 2275 } else if (bootverbose) { 2276 printf("Swap device %s removed.\n", devname); 2277 } 2278 mtx_lock(&sw_dev_mtx); 2279 } 2280 mtx_unlock(&sw_dev_mtx); 2281 2282 sx_xunlock(&swdev_syscall_lock); 2283 } 2284 2285 void 2286 swap_pager_status(int *total, int *used) 2287 { 2288 struct swdevt *sp; 2289 2290 *total = 0; 2291 *used = 0; 2292 mtx_lock(&sw_dev_mtx); 2293 TAILQ_FOREACH(sp, &swtailq, sw_list) { 2294 *total += sp->sw_nblks; 2295 *used += sp->sw_used; 2296 } 2297 mtx_unlock(&sw_dev_mtx); 2298 } 2299 2300 int 2301 swap_dev_info(int name, struct xswdev *xs, char *devname, size_t len) 2302 { 2303 struct swdevt *sp; 2304 const char *tmp_devname; 2305 int error, n; 2306 2307 n = 0; 2308 error = ENOENT; 2309 mtx_lock(&sw_dev_mtx); 2310 TAILQ_FOREACH(sp, &swtailq, sw_list) { 2311 if (n != name) { 2312 n++; 2313 continue; 2314 } 2315 xs->xsw_version = XSWDEV_VERSION; 2316 xs->xsw_dev = sp->sw_dev; 2317 xs->xsw_flags = sp->sw_flags; 2318 xs->xsw_nblks = sp->sw_nblks; 2319 xs->xsw_used = sp->sw_used; 2320 if (devname != NULL) { 2321 if (vn_isdisk(sp->sw_vp, NULL)) 2322 tmp_devname = devtoname(sp->sw_vp->v_rdev); 2323 else 2324 tmp_devname = "[file]"; 2325 strncpy(devname, tmp_devname, len); 2326 } 2327 error = 0; 2328 break; 2329 } 2330 mtx_unlock(&sw_dev_mtx); 2331 return (error); 2332 } 2333 2334 static int 2335 sysctl_vm_swap_info(SYSCTL_HANDLER_ARGS) 2336 { 2337 struct xswdev xs; 2338 int error; 2339 2340 if (arg2 != 1) /* name length */ 2341 return (EINVAL); 2342 error = swap_dev_info(*(int *)arg1, &xs, NULL, 0); 2343 if (error != 0) 2344 return (error); 2345 error = SYSCTL_OUT(req, &xs, sizeof(xs)); 2346 return (error); 2347 } 2348 2349 SYSCTL_INT(_vm, OID_AUTO, nswapdev, CTLFLAG_RD, &nswapdev, 0, 2350 "Number of swap devices"); 2351 SYSCTL_NODE(_vm, OID_AUTO, swap_info, CTLFLAG_RD | CTLFLAG_MPSAFE, 2352 sysctl_vm_swap_info, 2353 "Swap statistics by device"); 2354 2355 /* 2356 * vmspace_swap_count() - count the approximate swap usage in pages for a 2357 * vmspace. 2358 * 2359 * The map must be locked. 2360 * 2361 * Swap usage is determined by taking the proportional swap used by 2362 * VM objects backing the VM map. To make up for fractional losses, 2363 * if the VM object has any swap use at all the associated map entries 2364 * count for at least 1 swap page. 2365 */ 2366 long 2367 vmspace_swap_count(struct vmspace *vmspace) 2368 { 2369 vm_map_t map; 2370 vm_map_entry_t cur; 2371 vm_object_t object; 2372 long count, n; 2373 2374 map = &vmspace->vm_map; 2375 count = 0; 2376 2377 for (cur = map->header.next; cur != &map->header; cur = cur->next) { 2378 if ((cur->eflags & MAP_ENTRY_IS_SUB_MAP) == 0 && 2379 (object = cur->object.vm_object) != NULL) { 2380 VM_OBJECT_WLOCK(object); 2381 if (object->type == OBJT_SWAP && 2382 object->un_pager.swp.swp_bcount != 0) { 2383 n = (cur->end - cur->start) / PAGE_SIZE; 2384 count += object->un_pager.swp.swp_bcount * 2385 SWAP_META_PAGES * n / object->size + 1; 2386 } 2387 VM_OBJECT_WUNLOCK(object); 2388 } 2389 } 2390 return (count); 2391 } 2392 2393 /* 2394 * GEOM backend 2395 * 2396 * Swapping onto disk devices. 2397 * 2398 */ 2399 2400 static g_orphan_t swapgeom_orphan; 2401 2402 static struct g_class g_swap_class = { 2403 .name = "SWAP", 2404 .version = G_VERSION, 2405 .orphan = swapgeom_orphan, 2406 }; 2407 2408 DECLARE_GEOM_CLASS(g_swap_class, g_class); 2409 2410 2411 static void 2412 swapgeom_close_ev(void *arg, int flags) 2413 { 2414 struct g_consumer *cp; 2415 2416 cp = arg; 2417 g_access(cp, -1, -1, 0); 2418 g_detach(cp); 2419 g_destroy_consumer(cp); 2420 } 2421 2422 /* 2423 * Add a reference to the g_consumer for an inflight transaction. 2424 */ 2425 static void 2426 swapgeom_acquire(struct g_consumer *cp) 2427 { 2428 2429 mtx_assert(&sw_dev_mtx, MA_OWNED); 2430 cp->index++; 2431 } 2432 2433 /* 2434 * Remove a reference from the g_consumer. Post a close event if 2435 * all referneces go away. 2436 */ 2437 static void 2438 swapgeom_release(struct g_consumer *cp, struct swdevt *sp) 2439 { 2440 2441 mtx_assert(&sw_dev_mtx, MA_OWNED); 2442 cp->index--; 2443 if (cp->index == 0) { 2444 if (g_post_event(swapgeom_close_ev, cp, M_NOWAIT, NULL) == 0) 2445 sp->sw_id = NULL; 2446 } 2447 } 2448 2449 static void 2450 swapgeom_done(struct bio *bp2) 2451 { 2452 struct swdevt *sp; 2453 struct buf *bp; 2454 struct g_consumer *cp; 2455 2456 bp = bp2->bio_caller2; 2457 cp = bp2->bio_from; 2458 bp->b_ioflags = bp2->bio_flags; 2459 if (bp2->bio_error) 2460 bp->b_ioflags |= BIO_ERROR; 2461 bp->b_resid = bp->b_bcount - bp2->bio_completed; 2462 bp->b_error = bp2->bio_error; 2463 bufdone(bp); 2464 sp = bp2->bio_caller1; 2465 mtx_lock(&sw_dev_mtx); 2466 swapgeom_release(cp, sp); 2467 mtx_unlock(&sw_dev_mtx); 2468 g_destroy_bio(bp2); 2469 } 2470 2471 static void 2472 swapgeom_strategy(struct buf *bp, struct swdevt *sp) 2473 { 2474 struct bio *bio; 2475 struct g_consumer *cp; 2476 2477 mtx_lock(&sw_dev_mtx); 2478 cp = sp->sw_id; 2479 if (cp == NULL) { 2480 mtx_unlock(&sw_dev_mtx); 2481 bp->b_error = ENXIO; 2482 bp->b_ioflags |= BIO_ERROR; 2483 bufdone(bp); 2484 return; 2485 } 2486 swapgeom_acquire(cp); 2487 mtx_unlock(&sw_dev_mtx); 2488 if (bp->b_iocmd == BIO_WRITE) 2489 bio = g_new_bio(); 2490 else 2491 bio = g_alloc_bio(); 2492 if (bio == NULL) { 2493 mtx_lock(&sw_dev_mtx); 2494 swapgeom_release(cp, sp); 2495 mtx_unlock(&sw_dev_mtx); 2496 bp->b_error = ENOMEM; 2497 bp->b_ioflags |= BIO_ERROR; 2498 bufdone(bp); 2499 return; 2500 } 2501 2502 bio->bio_caller1 = sp; 2503 bio->bio_caller2 = bp; 2504 bio->bio_cmd = bp->b_iocmd; 2505 bio->bio_offset = (bp->b_blkno - sp->sw_first) * PAGE_SIZE; 2506 bio->bio_length = bp->b_bcount; 2507 bio->bio_done = swapgeom_done; 2508 if (!buf_mapped(bp)) { 2509 bio->bio_ma = bp->b_pages; 2510 bio->bio_data = unmapped_buf; 2511 bio->bio_ma_offset = (vm_offset_t)bp->b_offset & PAGE_MASK; 2512 bio->bio_ma_n = bp->b_npages; 2513 bio->bio_flags |= BIO_UNMAPPED; 2514 } else { 2515 bio->bio_data = bp->b_data; 2516 bio->bio_ma = NULL; 2517 } 2518 g_io_request(bio, cp); 2519 return; 2520 } 2521 2522 static void 2523 swapgeom_orphan(struct g_consumer *cp) 2524 { 2525 struct swdevt *sp; 2526 int destroy; 2527 2528 mtx_lock(&sw_dev_mtx); 2529 TAILQ_FOREACH(sp, &swtailq, sw_list) { 2530 if (sp->sw_id == cp) { 2531 sp->sw_flags |= SW_CLOSING; 2532 break; 2533 } 2534 } 2535 /* 2536 * Drop reference we were created with. Do directly since we're in a 2537 * special context where we don't have to queue the call to 2538 * swapgeom_close_ev(). 2539 */ 2540 cp->index--; 2541 destroy = ((sp != NULL) && (cp->index == 0)); 2542 if (destroy) 2543 sp->sw_id = NULL; 2544 mtx_unlock(&sw_dev_mtx); 2545 if (destroy) 2546 swapgeom_close_ev(cp, 0); 2547 } 2548 2549 static void 2550 swapgeom_close(struct thread *td, struct swdevt *sw) 2551 { 2552 struct g_consumer *cp; 2553 2554 mtx_lock(&sw_dev_mtx); 2555 cp = sw->sw_id; 2556 sw->sw_id = NULL; 2557 mtx_unlock(&sw_dev_mtx); 2558 /* XXX: direct call when Giant untangled */ 2559 if (cp != NULL) 2560 g_waitfor_event(swapgeom_close_ev, cp, M_WAITOK, NULL); 2561 } 2562 2563 2564 struct swh0h0 { 2565 struct cdev *dev; 2566 struct vnode *vp; 2567 int error; 2568 }; 2569 2570 static void 2571 swapongeom_ev(void *arg, int flags) 2572 { 2573 struct swh0h0 *swh; 2574 struct g_provider *pp; 2575 struct g_consumer *cp; 2576 static struct g_geom *gp; 2577 struct swdevt *sp; 2578 u_long nblks; 2579 int error; 2580 2581 swh = arg; 2582 swh->error = 0; 2583 pp = g_dev_getprovider(swh->dev); 2584 if (pp == NULL) { 2585 swh->error = ENODEV; 2586 return; 2587 } 2588 mtx_lock(&sw_dev_mtx); 2589 TAILQ_FOREACH(sp, &swtailq, sw_list) { 2590 cp = sp->sw_id; 2591 if (cp != NULL && cp->provider == pp) { 2592 mtx_unlock(&sw_dev_mtx); 2593 swh->error = EBUSY; 2594 return; 2595 } 2596 } 2597 mtx_unlock(&sw_dev_mtx); 2598 if (gp == NULL) 2599 gp = g_new_geomf(&g_swap_class, "swap"); 2600 cp = g_new_consumer(gp); 2601 cp->index = 1; /* Number of active I/Os, plus one for being active. */ 2602 cp->flags |= G_CF_DIRECT_SEND | G_CF_DIRECT_RECEIVE; 2603 g_attach(cp, pp); 2604 /* 2605 * XXX: Every time you think you can improve the margin for 2606 * footshooting, somebody depends on the ability to do so: 2607 * savecore(8) wants to write to our swapdev so we cannot 2608 * set an exclusive count :-( 2609 */ 2610 error = g_access(cp, 1, 1, 0); 2611 if (error) { 2612 g_detach(cp); 2613 g_destroy_consumer(cp); 2614 swh->error = error; 2615 return; 2616 } 2617 nblks = pp->mediasize / DEV_BSIZE; 2618 swaponsomething(swh->vp, cp, nblks, swapgeom_strategy, 2619 swapgeom_close, dev2udev(swh->dev), 2620 (pp->flags & G_PF_ACCEPT_UNMAPPED) != 0 ? SW_UNMAPPED : 0); 2621 swh->error = 0; 2622 } 2623 2624 static int 2625 swapongeom(struct thread *td, struct vnode *vp) 2626 { 2627 int error; 2628 struct swh0h0 swh; 2629 2630 vn_lock(vp, LK_EXCLUSIVE | LK_RETRY); 2631 2632 swh.dev = vp->v_rdev; 2633 swh.vp = vp; 2634 swh.error = 0; 2635 /* XXX: direct call when Giant untangled */ 2636 error = g_waitfor_event(swapongeom_ev, &swh, M_WAITOK, NULL); 2637 if (!error) 2638 error = swh.error; 2639 VOP_UNLOCK(vp, 0); 2640 return (error); 2641 } 2642 2643 /* 2644 * VNODE backend 2645 * 2646 * This is used mainly for network filesystem (read: probably only tested 2647 * with NFS) swapfiles. 2648 * 2649 */ 2650 2651 static void 2652 swapdev_strategy(struct buf *bp, struct swdevt *sp) 2653 { 2654 struct vnode *vp2; 2655 2656 bp->b_blkno = ctodb(bp->b_blkno - sp->sw_first); 2657 2658 vp2 = sp->sw_id; 2659 vhold(vp2); 2660 if (bp->b_iocmd == BIO_WRITE) { 2661 if (bp->b_bufobj) 2662 bufobj_wdrop(bp->b_bufobj); 2663 bufobj_wref(&vp2->v_bufobj); 2664 } 2665 if (bp->b_bufobj != &vp2->v_bufobj) 2666 bp->b_bufobj = &vp2->v_bufobj; 2667 bp->b_vp = vp2; 2668 bp->b_iooffset = dbtob(bp->b_blkno); 2669 bstrategy(bp); 2670 return; 2671 } 2672 2673 static void 2674 swapdev_close(struct thread *td, struct swdevt *sp) 2675 { 2676 2677 VOP_CLOSE(sp->sw_vp, FREAD | FWRITE, td->td_ucred, td); 2678 vrele(sp->sw_vp); 2679 } 2680 2681 2682 static int 2683 swaponvp(struct thread *td, struct vnode *vp, u_long nblks) 2684 { 2685 struct swdevt *sp; 2686 int error; 2687 2688 if (nblks == 0) 2689 return (ENXIO); 2690 mtx_lock(&sw_dev_mtx); 2691 TAILQ_FOREACH(sp, &swtailq, sw_list) { 2692 if (sp->sw_id == vp) { 2693 mtx_unlock(&sw_dev_mtx); 2694 return (EBUSY); 2695 } 2696 } 2697 mtx_unlock(&sw_dev_mtx); 2698 2699 (void) vn_lock(vp, LK_EXCLUSIVE | LK_RETRY); 2700 #ifdef MAC 2701 error = mac_system_check_swapon(td->td_ucred, vp); 2702 if (error == 0) 2703 #endif 2704 error = VOP_OPEN(vp, FREAD | FWRITE, td->td_ucred, td, NULL); 2705 (void) VOP_UNLOCK(vp, 0); 2706 if (error) 2707 return (error); 2708 2709 swaponsomething(vp, vp, nblks, swapdev_strategy, swapdev_close, 2710 NODEV, 0); 2711 return (0); 2712 } 2713 2714 static int 2715 sysctl_swap_async_max(SYSCTL_HANDLER_ARGS) 2716 { 2717 int error, new, n; 2718 2719 new = nsw_wcount_async_max; 2720 error = sysctl_handle_int(oidp, &new, 0, req); 2721 if (error != 0 || req->newptr == NULL) 2722 return (error); 2723 2724 if (new > nswbuf / 2 || new < 1) 2725 return (EINVAL); 2726 2727 mtx_lock(&pbuf_mtx); 2728 while (nsw_wcount_async_max != new) { 2729 /* 2730 * Adjust difference. If the current async count is too low, 2731 * we will need to sqeeze our update slowly in. Sleep with a 2732 * higher priority than getpbuf() to finish faster. 2733 */ 2734 n = new - nsw_wcount_async_max; 2735 if (nsw_wcount_async + n >= 0) { 2736 nsw_wcount_async += n; 2737 nsw_wcount_async_max += n; 2738 wakeup(&nsw_wcount_async); 2739 } else { 2740 nsw_wcount_async_max -= nsw_wcount_async; 2741 nsw_wcount_async = 0; 2742 msleep(&nsw_wcount_async, &pbuf_mtx, PSWP, 2743 "swpsysctl", 0); 2744 } 2745 } 2746 mtx_unlock(&pbuf_mtx); 2747 2748 return (0); 2749 } 2750