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_mac.h" 73 #include "opt_swap.h" 74 #include "opt_vm.h" 75 76 #include <sys/param.h> 77 #include <sys/systm.h> 78 #include <sys/conf.h> 79 #include <sys/kernel.h> 80 #include <sys/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/mac.h> 89 #include <sys/malloc.h> 90 #include <sys/sysctl.h> 91 #include <sys/sysproto.h> 92 #include <sys/blist.h> 93 #include <sys/lock.h> 94 #include <sys/sx.h> 95 #include <sys/vmmeter.h> 96 97 #include <vm/vm.h> 98 #include <vm/pmap.h> 99 #include <vm/vm_map.h> 100 #include <vm/vm_kern.h> 101 #include <vm/vm_object.h> 102 #include <vm/vm_page.h> 103 #include <vm/vm_pager.h> 104 #include <vm/vm_pageout.h> 105 #include <vm/vm_param.h> 106 #include <vm/swap_pager.h> 107 #include <vm/vm_extern.h> 108 #include <vm/uma.h> 109 110 #include <geom/geom.h> 111 112 /* 113 * SWB_NPAGES must be a power of 2. It may be set to 1, 2, 4, 8, or 16 114 * pages per allocation. We recommend you stick with the default of 8. 115 * The 16-page limit is due to the radix code (kern/subr_blist.c). 116 */ 117 #ifndef MAX_PAGEOUT_CLUSTER 118 #define MAX_PAGEOUT_CLUSTER 16 119 #endif 120 121 #if !defined(SWB_NPAGES) 122 #define SWB_NPAGES MAX_PAGEOUT_CLUSTER 123 #endif 124 125 /* 126 * Piecemeal swap metadata structure. Swap is stored in a radix tree. 127 * 128 * If SWB_NPAGES is 8 and sizeof(char *) == sizeof(daddr_t), our radix 129 * is basically 8. Assuming PAGE_SIZE == 4096, one tree level represents 130 * 32K worth of data, two levels represent 256K, three levels represent 131 * 2 MBytes. This is acceptable. 132 * 133 * Overall memory utilization is about the same as the old swap structure. 134 */ 135 #define SWCORRECT(n) (sizeof(void *) * (n) / sizeof(daddr_t)) 136 #define SWAP_META_PAGES (SWB_NPAGES * 2) 137 #define SWAP_META_MASK (SWAP_META_PAGES - 1) 138 139 typedef int32_t swblk_t; /* 140 * swap offset. This is the type used to 141 * address the "virtual swap device" and 142 * therefore the maximum swap space is 143 * 2^32 pages. 144 */ 145 146 struct swdevt; 147 typedef void sw_strategy_t(struct buf *bp, struct swdevt *sw); 148 typedef void sw_close_t(struct thread *td, struct swdevt *sw); 149 150 /* 151 * Swap device table 152 */ 153 struct swdevt { 154 int sw_flags; 155 int sw_nblks; 156 int sw_used; 157 dev_t sw_dev; 158 struct vnode *sw_vp; 159 void *sw_id; 160 swblk_t sw_first; 161 swblk_t sw_end; 162 struct blist *sw_blist; 163 TAILQ_ENTRY(swdevt) sw_list; 164 sw_strategy_t *sw_strategy; 165 sw_close_t *sw_close; 166 }; 167 168 #define SW_CLOSING 0x04 169 170 struct swblock { 171 struct swblock *swb_hnext; 172 vm_object_t swb_object; 173 vm_pindex_t swb_index; 174 int swb_count; 175 daddr_t swb_pages[SWAP_META_PAGES]; 176 }; 177 178 static struct mtx sw_dev_mtx; 179 static TAILQ_HEAD(, swdevt) swtailq = TAILQ_HEAD_INITIALIZER(swtailq); 180 static struct swdevt *swdevhd; /* Allocate from here next */ 181 static int nswapdev; /* Number of swap devices */ 182 int swap_pager_avail; 183 static int swdev_syscall_active = 0; /* serialize swap(on|off) */ 184 185 static void swapdev_strategy(struct buf *, struct swdevt *sw); 186 187 #define SWM_FREE 0x02 /* free, period */ 188 #define SWM_POP 0x04 /* pop out */ 189 190 int swap_pager_full = 2; /* swap space exhaustion (task killing) */ 191 static int swap_pager_almost_full = 1; /* swap space exhaustion (w/hysteresis)*/ 192 static int nsw_rcount; /* free read buffers */ 193 static int nsw_wcount_sync; /* limit write buffers / synchronous */ 194 static int nsw_wcount_async; /* limit write buffers / asynchronous */ 195 static int nsw_wcount_async_max;/* assigned maximum */ 196 static int nsw_cluster_max; /* maximum VOP I/O allowed */ 197 198 static struct swblock **swhash; 199 static int swhash_mask; 200 static struct mtx swhash_mtx; 201 202 static int swap_async_max = 4; /* maximum in-progress async I/O's */ 203 static struct sx sw_alloc_sx; 204 205 206 SYSCTL_INT(_vm, OID_AUTO, swap_async_max, 207 CTLFLAG_RW, &swap_async_max, 0, "Maximum running async swap ops"); 208 209 /* 210 * "named" and "unnamed" anon region objects. Try to reduce the overhead 211 * of searching a named list by hashing it just a little. 212 */ 213 214 #define NOBJLISTS 8 215 216 #define NOBJLIST(handle) \ 217 (&swap_pager_object_list[((int)(intptr_t)handle >> 4) & (NOBJLISTS-1)]) 218 219 static struct mtx sw_alloc_mtx; /* protect list manipulation */ 220 static struct pagerlst swap_pager_object_list[NOBJLISTS]; 221 static uma_zone_t swap_zone; 222 223 /* 224 * pagerops for OBJT_SWAP - "swap pager". Some ops are also global procedure 225 * calls hooked from other parts of the VM system and do not appear here. 226 * (see vm/swap_pager.h). 227 */ 228 static vm_object_t 229 swap_pager_alloc(void *handle, vm_ooffset_t size, 230 vm_prot_t prot, vm_ooffset_t offset); 231 static void swap_pager_dealloc(vm_object_t object); 232 static int swap_pager_getpages(vm_object_t, vm_page_t *, int, int); 233 static void swap_pager_putpages(vm_object_t, vm_page_t *, int, boolean_t, int *); 234 static boolean_t 235 swap_pager_haspage(vm_object_t object, vm_pindex_t pindex, int *before, int *after); 236 static void swap_pager_init(void); 237 static void swap_pager_unswapped(vm_page_t); 238 static void swap_pager_swapoff(struct swdevt *sp, int *sw_used); 239 240 struct pagerops swappagerops = { 241 .pgo_init = swap_pager_init, /* early system initialization of pager */ 242 .pgo_alloc = swap_pager_alloc, /* allocate an OBJT_SWAP object */ 243 .pgo_dealloc = swap_pager_dealloc, /* deallocate an OBJT_SWAP object */ 244 .pgo_getpages = swap_pager_getpages, /* pagein */ 245 .pgo_putpages = swap_pager_putpages, /* pageout */ 246 .pgo_haspage = swap_pager_haspage, /* get backing store status for page */ 247 .pgo_pageunswapped = swap_pager_unswapped, /* remove swap related to page */ 248 }; 249 250 /* 251 * dmmax is in page-sized chunks with the new swap system. It was 252 * dev-bsized chunks in the old. dmmax is always a power of 2. 253 * 254 * swap_*() routines are externally accessible. swp_*() routines are 255 * internal. 256 */ 257 static int dmmax; 258 static int nswap_lowat = 128; /* in pages, swap_pager_almost_full warn */ 259 static int nswap_hiwat = 512; /* in pages, swap_pager_almost_full warn */ 260 261 SYSCTL_INT(_vm, OID_AUTO, dmmax, 262 CTLFLAG_RD, &dmmax, 0, "Maximum size of a swap block"); 263 264 static void swp_sizecheck(void); 265 static void swp_pager_async_iodone(struct buf *bp); 266 static int swapongeom(struct thread *, struct vnode *); 267 static int swaponvp(struct thread *, struct vnode *, u_long); 268 269 /* 270 * Swap bitmap functions 271 */ 272 static void swp_pager_freeswapspace(daddr_t blk, int npages); 273 static daddr_t swp_pager_getswapspace(int npages); 274 275 /* 276 * Metadata functions 277 */ 278 static struct swblock **swp_pager_hash(vm_object_t object, vm_pindex_t index); 279 static void swp_pager_meta_build(vm_object_t, vm_pindex_t, daddr_t); 280 static void swp_pager_meta_free(vm_object_t, vm_pindex_t, daddr_t); 281 static void swp_pager_meta_free_all(vm_object_t); 282 static daddr_t swp_pager_meta_ctl(vm_object_t, vm_pindex_t, int); 283 284 /* 285 * SWP_SIZECHECK() - update swap_pager_full indication 286 * 287 * update the swap_pager_almost_full indication and warn when we are 288 * about to run out of swap space, using lowat/hiwat hysteresis. 289 * 290 * Clear swap_pager_full ( task killing ) indication when lowat is met. 291 * 292 * No restrictions on call 293 * This routine may not block. 294 * This routine must be called at splvm() 295 */ 296 static void 297 swp_sizecheck(void) 298 { 299 300 if (swap_pager_avail < nswap_lowat) { 301 if (swap_pager_almost_full == 0) { 302 printf("swap_pager: out of swap space\n"); 303 swap_pager_almost_full = 1; 304 } 305 } else { 306 swap_pager_full = 0; 307 if (swap_pager_avail > nswap_hiwat) 308 swap_pager_almost_full = 0; 309 } 310 } 311 312 /* 313 * SWP_PAGER_HASH() - hash swap meta data 314 * 315 * This is an helper function which hashes the swapblk given 316 * the object and page index. It returns a pointer to a pointer 317 * to the object, or a pointer to a NULL pointer if it could not 318 * find a swapblk. 319 * 320 * This routine must be called at splvm(). 321 */ 322 static struct swblock ** 323 swp_pager_hash(vm_object_t object, vm_pindex_t index) 324 { 325 struct swblock **pswap; 326 struct swblock *swap; 327 328 index &= ~(vm_pindex_t)SWAP_META_MASK; 329 pswap = &swhash[(index ^ (int)(intptr_t)object) & swhash_mask]; 330 while ((swap = *pswap) != NULL) { 331 if (swap->swb_object == object && 332 swap->swb_index == index 333 ) { 334 break; 335 } 336 pswap = &swap->swb_hnext; 337 } 338 return (pswap); 339 } 340 341 /* 342 * SWAP_PAGER_INIT() - initialize the swap pager! 343 * 344 * Expected to be started from system init. NOTE: This code is run 345 * before much else so be careful what you depend on. Most of the VM 346 * system has yet to be initialized at this point. 347 */ 348 static void 349 swap_pager_init(void) 350 { 351 /* 352 * Initialize object lists 353 */ 354 int i; 355 356 for (i = 0; i < NOBJLISTS; ++i) 357 TAILQ_INIT(&swap_pager_object_list[i]); 358 mtx_init(&sw_alloc_mtx, "swap_pager list", NULL, MTX_DEF); 359 mtx_init(&sw_dev_mtx, "swapdev", NULL, MTX_DEF); 360 361 /* 362 * Device Stripe, in PAGE_SIZE'd blocks 363 */ 364 dmmax = SWB_NPAGES * 2; 365 } 366 367 /* 368 * SWAP_PAGER_SWAP_INIT() - swap pager initialization from pageout process 369 * 370 * Expected to be started from pageout process once, prior to entering 371 * its main loop. 372 */ 373 void 374 swap_pager_swap_init(void) 375 { 376 int n, n2; 377 378 /* 379 * Number of in-transit swap bp operations. Don't 380 * exhaust the pbufs completely. Make sure we 381 * initialize workable values (0 will work for hysteresis 382 * but it isn't very efficient). 383 * 384 * The nsw_cluster_max is constrained by the bp->b_pages[] 385 * array (MAXPHYS/PAGE_SIZE) and our locally defined 386 * MAX_PAGEOUT_CLUSTER. Also be aware that swap ops are 387 * constrained by the swap device interleave stripe size. 388 * 389 * Currently we hardwire nsw_wcount_async to 4. This limit is 390 * designed to prevent other I/O from having high latencies due to 391 * our pageout I/O. The value 4 works well for one or two active swap 392 * devices but is probably a little low if you have more. Even so, 393 * a higher value would probably generate only a limited improvement 394 * with three or four active swap devices since the system does not 395 * typically have to pageout at extreme bandwidths. We will want 396 * at least 2 per swap devices, and 4 is a pretty good value if you 397 * have one NFS swap device due to the command/ack latency over NFS. 398 * So it all works out pretty well. 399 */ 400 nsw_cluster_max = min((MAXPHYS/PAGE_SIZE), MAX_PAGEOUT_CLUSTER); 401 402 mtx_lock(&pbuf_mtx); 403 nsw_rcount = (nswbuf + 1) / 2; 404 nsw_wcount_sync = (nswbuf + 3) / 4; 405 nsw_wcount_async = 4; 406 nsw_wcount_async_max = nsw_wcount_async; 407 mtx_unlock(&pbuf_mtx); 408 409 /* 410 * Initialize our zone. Right now I'm just guessing on the number 411 * we need based on the number of pages in the system. Each swblock 412 * can hold 16 pages, so this is probably overkill. This reservation 413 * is typically limited to around 32MB by default. 414 */ 415 n = cnt.v_page_count / 2; 416 if (maxswzone && n > maxswzone / sizeof(struct swblock)) 417 n = maxswzone / sizeof(struct swblock); 418 n2 = n; 419 swap_zone = uma_zcreate("SWAPMETA", sizeof(struct swblock), NULL, NULL, 420 NULL, NULL, UMA_ALIGN_PTR, UMA_ZONE_NOFREE | UMA_ZONE_VM); 421 do { 422 if (uma_zone_set_obj(swap_zone, NULL, n)) 423 break; 424 /* 425 * if the allocation failed, try a zone two thirds the 426 * size of the previous attempt. 427 */ 428 n -= ((n + 2) / 3); 429 } while (n > 0); 430 if (swap_zone == NULL) 431 panic("failed to create swap_zone."); 432 if (n2 != n) 433 printf("Swap zone entries reduced from %d to %d.\n", n2, n); 434 n2 = n; 435 436 /* 437 * Initialize our meta-data hash table. The swapper does not need to 438 * be quite as efficient as the VM system, so we do not use an 439 * oversized hash table. 440 * 441 * n: size of hash table, must be power of 2 442 * swhash_mask: hash table index mask 443 */ 444 for (n = 1; n < n2 / 8; n *= 2) 445 ; 446 swhash = malloc(sizeof(struct swblock *) * n, M_VMPGDATA, M_WAITOK | M_ZERO); 447 swhash_mask = n - 1; 448 mtx_init(&swhash_mtx, "swap_pager swhash", NULL, MTX_DEF); 449 } 450 451 /* 452 * SWAP_PAGER_ALLOC() - allocate a new OBJT_SWAP VM object and instantiate 453 * its metadata structures. 454 * 455 * This routine is called from the mmap and fork code to create a new 456 * OBJT_SWAP object. We do this by creating an OBJT_DEFAULT object 457 * and then converting it with swp_pager_meta_build(). 458 * 459 * This routine may block in vm_object_allocate() and create a named 460 * object lookup race, so we must interlock. We must also run at 461 * splvm() for the object lookup to handle races with interrupts, but 462 * we do not have to maintain splvm() in between the lookup and the 463 * add because (I believe) it is not possible to attempt to create 464 * a new swap object w/handle when a default object with that handle 465 * already exists. 466 * 467 * MPSAFE 468 */ 469 static vm_object_t 470 swap_pager_alloc(void *handle, vm_ooffset_t size, vm_prot_t prot, 471 vm_ooffset_t offset) 472 { 473 vm_object_t object; 474 vm_pindex_t pindex; 475 476 pindex = OFF_TO_IDX(offset + PAGE_MASK + size); 477 478 if (handle) { 479 mtx_lock(&Giant); 480 /* 481 * Reference existing named region or allocate new one. There 482 * should not be a race here against swp_pager_meta_build() 483 * as called from vm_page_remove() in regards to the lookup 484 * of the handle. 485 */ 486 sx_xlock(&sw_alloc_sx); 487 object = vm_pager_object_lookup(NOBJLIST(handle), handle); 488 489 if (object != NULL) { 490 vm_object_reference(object); 491 } else { 492 object = vm_object_allocate(OBJT_DEFAULT, pindex); 493 object->handle = handle; 494 495 VM_OBJECT_LOCK(object); 496 swp_pager_meta_build(object, 0, SWAPBLK_NONE); 497 VM_OBJECT_UNLOCK(object); 498 } 499 sx_xunlock(&sw_alloc_sx); 500 mtx_unlock(&Giant); 501 } else { 502 object = vm_object_allocate(OBJT_DEFAULT, pindex); 503 504 VM_OBJECT_LOCK(object); 505 swp_pager_meta_build(object, 0, SWAPBLK_NONE); 506 VM_OBJECT_UNLOCK(object); 507 } 508 return (object); 509 } 510 511 /* 512 * SWAP_PAGER_DEALLOC() - remove swap metadata from object 513 * 514 * The swap backing for the object is destroyed. The code is 515 * designed such that we can reinstantiate it later, but this 516 * routine is typically called only when the entire object is 517 * about to be destroyed. 518 * 519 * This routine may block, but no longer does. 520 * 521 * The object must be locked or unreferenceable. 522 */ 523 static void 524 swap_pager_dealloc(vm_object_t object) 525 { 526 int s; 527 528 /* 529 * Remove from list right away so lookups will fail if we block for 530 * pageout completion. 531 */ 532 if (object->handle != NULL) { 533 mtx_lock(&sw_alloc_mtx); 534 TAILQ_REMOVE(NOBJLIST(object->handle), object, pager_object_list); 535 mtx_unlock(&sw_alloc_mtx); 536 } 537 538 VM_OBJECT_LOCK_ASSERT(object, MA_OWNED); 539 vm_object_pip_wait(object, "swpdea"); 540 541 /* 542 * Free all remaining metadata. We only bother to free it from 543 * the swap meta data. We do not attempt to free swapblk's still 544 * associated with vm_page_t's for this object. We do not care 545 * if paging is still in progress on some objects. 546 */ 547 s = splvm(); 548 swp_pager_meta_free_all(object); 549 splx(s); 550 } 551 552 /************************************************************************ 553 * SWAP PAGER BITMAP ROUTINES * 554 ************************************************************************/ 555 556 /* 557 * SWP_PAGER_GETSWAPSPACE() - allocate raw swap space 558 * 559 * Allocate swap for the requested number of pages. The starting 560 * swap block number (a page index) is returned or SWAPBLK_NONE 561 * if the allocation failed. 562 * 563 * Also has the side effect of advising that somebody made a mistake 564 * when they configured swap and didn't configure enough. 565 * 566 * Must be called at splvm() to avoid races with bitmap frees from 567 * vm_page_remove() aka swap_pager_page_removed(). 568 * 569 * This routine may not block 570 * This routine must be called at splvm(). 571 * 572 * We allocate in round-robin fashion from the configured devices. 573 */ 574 static daddr_t 575 swp_pager_getswapspace(int npages) 576 { 577 daddr_t blk; 578 struct swdevt *sp; 579 int i; 580 581 blk = SWAPBLK_NONE; 582 mtx_lock(&sw_dev_mtx); 583 sp = swdevhd; 584 for (i = 0; i < nswapdev; i++) { 585 if (sp == NULL) 586 sp = TAILQ_FIRST(&swtailq); 587 if (!(sp->sw_flags & SW_CLOSING)) { 588 blk = blist_alloc(sp->sw_blist, npages); 589 if (blk != SWAPBLK_NONE) { 590 blk += sp->sw_first; 591 sp->sw_used += npages; 592 swap_pager_avail -= npages; 593 swp_sizecheck(); 594 swdevhd = TAILQ_NEXT(sp, sw_list); 595 goto done; 596 } 597 } 598 sp = TAILQ_NEXT(sp, sw_list); 599 } 600 if (swap_pager_full != 2) { 601 printf("swap_pager_getswapspace(%d): failed\n", npages); 602 swap_pager_full = 2; 603 swap_pager_almost_full = 1; 604 } 605 swdevhd = NULL; 606 done: 607 mtx_unlock(&sw_dev_mtx); 608 return (blk); 609 } 610 611 static struct swdevt * 612 swp_pager_find_dev(daddr_t blk) 613 { 614 struct swdevt *sp; 615 616 mtx_lock(&sw_dev_mtx); 617 TAILQ_FOREACH(sp, &swtailq, sw_list) { 618 if (blk >= sp->sw_first && blk < sp->sw_end) { 619 mtx_unlock(&sw_dev_mtx); 620 return (sp); 621 } 622 } 623 panic("Swapdev not found"); 624 } 625 626 static void 627 swp_pager_strategy(struct buf *bp) 628 { 629 struct swdevt *sp; 630 631 mtx_lock(&sw_dev_mtx); 632 TAILQ_FOREACH(sp, &swtailq, sw_list) { 633 if (bp->b_blkno >= sp->sw_first && bp->b_blkno < sp->sw_end) { 634 mtx_unlock(&sw_dev_mtx); 635 sp->sw_strategy(bp, sp); 636 return; 637 } 638 } 639 panic("Swapdev not found"); 640 } 641 642 643 /* 644 * SWP_PAGER_FREESWAPSPACE() - free raw swap space 645 * 646 * This routine returns the specified swap blocks back to the bitmap. 647 * 648 * Note: This routine may not block (it could in the old swap code), 649 * and through the use of the new blist routines it does not block. 650 * 651 * We must be called at splvm() to avoid races with bitmap frees from 652 * vm_page_remove() aka swap_pager_page_removed(). 653 * 654 * This routine may not block 655 * This routine must be called at splvm(). 656 */ 657 static void 658 swp_pager_freeswapspace(daddr_t blk, int npages) 659 { 660 struct swdevt *sp; 661 662 mtx_lock(&sw_dev_mtx); 663 TAILQ_FOREACH(sp, &swtailq, sw_list) { 664 if (blk >= sp->sw_first && blk < sp->sw_end) { 665 sp->sw_used -= npages; 666 /* 667 * If we are attempting to stop swapping on 668 * this device, we don't want to mark any 669 * blocks free lest they be reused. 670 */ 671 if ((sp->sw_flags & SW_CLOSING) == 0) { 672 blist_free(sp->sw_blist, blk - sp->sw_first, 673 npages); 674 swap_pager_avail += npages; 675 swp_sizecheck(); 676 } 677 mtx_unlock(&sw_dev_mtx); 678 return; 679 } 680 } 681 panic("Swapdev not found"); 682 } 683 684 /* 685 * SWAP_PAGER_FREESPACE() - frees swap blocks associated with a page 686 * range within an object. 687 * 688 * This is a globally accessible routine. 689 * 690 * This routine removes swapblk assignments from swap metadata. 691 * 692 * The external callers of this routine typically have already destroyed 693 * or renamed vm_page_t's associated with this range in the object so 694 * we should be ok. 695 * 696 * This routine may be called at any spl. We up our spl to splvm temporarily 697 * in order to perform the metadata removal. 698 */ 699 void 700 swap_pager_freespace(vm_object_t object, vm_pindex_t start, vm_size_t size) 701 { 702 int s = splvm(); 703 704 VM_OBJECT_LOCK_ASSERT(object, MA_OWNED); 705 swp_pager_meta_free(object, start, size); 706 splx(s); 707 } 708 709 /* 710 * SWAP_PAGER_RESERVE() - reserve swap blocks in object 711 * 712 * Assigns swap blocks to the specified range within the object. The 713 * swap blocks are not zerod. Any previous swap assignment is destroyed. 714 * 715 * Returns 0 on success, -1 on failure. 716 */ 717 int 718 swap_pager_reserve(vm_object_t object, vm_pindex_t start, vm_size_t size) 719 { 720 int s; 721 int n = 0; 722 daddr_t blk = SWAPBLK_NONE; 723 vm_pindex_t beg = start; /* save start index */ 724 725 s = splvm(); 726 VM_OBJECT_LOCK(object); 727 while (size) { 728 if (n == 0) { 729 n = BLIST_MAX_ALLOC; 730 while ((blk = swp_pager_getswapspace(n)) == SWAPBLK_NONE) { 731 n >>= 1; 732 if (n == 0) { 733 swp_pager_meta_free(object, beg, start - beg); 734 VM_OBJECT_UNLOCK(object); 735 splx(s); 736 return (-1); 737 } 738 } 739 } 740 swp_pager_meta_build(object, start, blk); 741 --size; 742 ++start; 743 ++blk; 744 --n; 745 } 746 swp_pager_meta_free(object, start, n); 747 VM_OBJECT_UNLOCK(object); 748 splx(s); 749 return (0); 750 } 751 752 /* 753 * SWAP_PAGER_COPY() - copy blocks from source pager to destination pager 754 * and destroy the source. 755 * 756 * Copy any valid swapblks from the source to the destination. In 757 * cases where both the source and destination have a valid swapblk, 758 * we keep the destination's. 759 * 760 * This routine is allowed to block. It may block allocating metadata 761 * indirectly through swp_pager_meta_build() or if paging is still in 762 * progress on the source. 763 * 764 * This routine can be called at any spl 765 * 766 * XXX vm_page_collapse() kinda expects us not to block because we 767 * supposedly do not need to allocate memory, but for the moment we 768 * *may* have to get a little memory from the zone allocator, but 769 * it is taken from the interrupt memory. We should be ok. 770 * 771 * The source object contains no vm_page_t's (which is just as well) 772 * 773 * The source object is of type OBJT_SWAP. 774 * 775 * The source and destination objects must be locked or 776 * inaccessible (XXX are they ?) 777 */ 778 void 779 swap_pager_copy(vm_object_t srcobject, vm_object_t dstobject, 780 vm_pindex_t offset, int destroysource) 781 { 782 vm_pindex_t i; 783 int s; 784 785 VM_OBJECT_LOCK_ASSERT(srcobject, MA_OWNED); 786 VM_OBJECT_LOCK_ASSERT(dstobject, MA_OWNED); 787 788 s = splvm(); 789 /* 790 * If destroysource is set, we remove the source object from the 791 * swap_pager internal queue now. 792 */ 793 if (destroysource) { 794 if (srcobject->handle != NULL) { 795 mtx_lock(&sw_alloc_mtx); 796 TAILQ_REMOVE( 797 NOBJLIST(srcobject->handle), 798 srcobject, 799 pager_object_list 800 ); 801 mtx_unlock(&sw_alloc_mtx); 802 } 803 } 804 805 /* 806 * transfer source to destination. 807 */ 808 for (i = 0; i < dstobject->size; ++i) { 809 daddr_t dstaddr; 810 811 /* 812 * Locate (without changing) the swapblk on the destination, 813 * unless it is invalid in which case free it silently, or 814 * if the destination is a resident page, in which case the 815 * source is thrown away. 816 */ 817 dstaddr = swp_pager_meta_ctl(dstobject, i, 0); 818 819 if (dstaddr == SWAPBLK_NONE) { 820 /* 821 * Destination has no swapblk and is not resident, 822 * copy source. 823 */ 824 daddr_t srcaddr; 825 826 srcaddr = swp_pager_meta_ctl( 827 srcobject, 828 i + offset, 829 SWM_POP 830 ); 831 832 if (srcaddr != SWAPBLK_NONE) { 833 /* 834 * swp_pager_meta_build() can sleep. 835 */ 836 vm_object_pip_add(srcobject, 1); 837 VM_OBJECT_UNLOCK(srcobject); 838 vm_object_pip_add(dstobject, 1); 839 swp_pager_meta_build(dstobject, i, srcaddr); 840 vm_object_pip_wakeup(dstobject); 841 VM_OBJECT_LOCK(srcobject); 842 vm_object_pip_wakeup(srcobject); 843 } 844 } else { 845 /* 846 * Destination has valid swapblk or it is represented 847 * by a resident page. We destroy the sourceblock. 848 */ 849 850 swp_pager_meta_ctl(srcobject, i + offset, SWM_FREE); 851 } 852 } 853 854 /* 855 * Free left over swap blocks in source. 856 * 857 * We have to revert the type to OBJT_DEFAULT so we do not accidently 858 * double-remove the object from the swap queues. 859 */ 860 if (destroysource) { 861 swp_pager_meta_free_all(srcobject); 862 /* 863 * Reverting the type is not necessary, the caller is going 864 * to destroy srcobject directly, but I'm doing it here 865 * for consistency since we've removed the object from its 866 * queues. 867 */ 868 srcobject->type = OBJT_DEFAULT; 869 } 870 splx(s); 871 } 872 873 /* 874 * SWAP_PAGER_HASPAGE() - determine if we have good backing store for 875 * the requested page. 876 * 877 * We determine whether good backing store exists for the requested 878 * page and return TRUE if it does, FALSE if it doesn't. 879 * 880 * If TRUE, we also try to determine how much valid, contiguous backing 881 * store exists before and after the requested page within a reasonable 882 * distance. We do not try to restrict it to the swap device stripe 883 * (that is handled in getpages/putpages). It probably isn't worth 884 * doing here. 885 */ 886 static boolean_t 887 swap_pager_haspage(vm_object_t object, vm_pindex_t pindex, int *before, int *after) 888 { 889 daddr_t blk0; 890 int s; 891 892 VM_OBJECT_LOCK_ASSERT(object, MA_OWNED); 893 /* 894 * do we have good backing store at the requested index ? 895 */ 896 s = splvm(); 897 blk0 = swp_pager_meta_ctl(object, pindex, 0); 898 899 if (blk0 == SWAPBLK_NONE) { 900 splx(s); 901 if (before) 902 *before = 0; 903 if (after) 904 *after = 0; 905 return (FALSE); 906 } 907 908 /* 909 * find backwards-looking contiguous good backing store 910 */ 911 if (before != NULL) { 912 int i; 913 914 for (i = 1; i < (SWB_NPAGES/2); ++i) { 915 daddr_t blk; 916 917 if (i > pindex) 918 break; 919 blk = swp_pager_meta_ctl(object, pindex - i, 0); 920 if (blk != blk0 - i) 921 break; 922 } 923 *before = (i - 1); 924 } 925 926 /* 927 * find forward-looking contiguous good backing store 928 */ 929 if (after != NULL) { 930 int i; 931 932 for (i = 1; i < (SWB_NPAGES/2); ++i) { 933 daddr_t blk; 934 935 blk = swp_pager_meta_ctl(object, pindex + i, 0); 936 if (blk != blk0 + i) 937 break; 938 } 939 *after = (i - 1); 940 } 941 splx(s); 942 return (TRUE); 943 } 944 945 /* 946 * SWAP_PAGER_PAGE_UNSWAPPED() - remove swap backing store related to page 947 * 948 * This removes any associated swap backing store, whether valid or 949 * not, from the page. 950 * 951 * This routine is typically called when a page is made dirty, at 952 * which point any associated swap can be freed. MADV_FREE also 953 * calls us in a special-case situation 954 * 955 * NOTE!!! If the page is clean and the swap was valid, the caller 956 * should make the page dirty before calling this routine. This routine 957 * does NOT change the m->dirty status of the page. Also: MADV_FREE 958 * depends on it. 959 * 960 * This routine may not block 961 * This routine must be called at splvm() 962 */ 963 static void 964 swap_pager_unswapped(vm_page_t m) 965 { 966 967 VM_OBJECT_LOCK_ASSERT(m->object, MA_OWNED); 968 swp_pager_meta_ctl(m->object, m->pindex, SWM_FREE); 969 } 970 971 /* 972 * SWAP_PAGER_GETPAGES() - bring pages in from swap 973 * 974 * Attempt to retrieve (m, count) pages from backing store, but make 975 * sure we retrieve at least m[reqpage]. We try to load in as large 976 * a chunk surrounding m[reqpage] as is contiguous in swap and which 977 * belongs to the same object. 978 * 979 * The code is designed for asynchronous operation and 980 * immediate-notification of 'reqpage' but tends not to be 981 * used that way. Please do not optimize-out this algorithmic 982 * feature, I intend to improve on it in the future. 983 * 984 * The parent has a single vm_object_pip_add() reference prior to 985 * calling us and we should return with the same. 986 * 987 * The parent has BUSY'd the pages. We should return with 'm' 988 * left busy, but the others adjusted. 989 */ 990 static int 991 swap_pager_getpages(vm_object_t object, vm_page_t *m, int count, int reqpage) 992 { 993 struct buf *bp; 994 vm_page_t mreq; 995 int s; 996 int i; 997 int j; 998 daddr_t blk; 999 1000 mreq = m[reqpage]; 1001 1002 KASSERT(mreq->object == object, 1003 ("swap_pager_getpages: object mismatch %p/%p", 1004 object, mreq->object)); 1005 1006 /* 1007 * Calculate range to retrieve. The pages have already been assigned 1008 * their swapblks. We require a *contiguous* range but we know it to 1009 * not span devices. If we do not supply it, bad things 1010 * happen. Note that blk, iblk & jblk can be SWAPBLK_NONE, but the 1011 * loops are set up such that the case(s) are handled implicitly. 1012 * 1013 * The swp_*() calls must be made at splvm(). vm_page_free() does 1014 * not need to be, but it will go a little faster if it is. 1015 */ 1016 s = splvm(); 1017 blk = swp_pager_meta_ctl(mreq->object, mreq->pindex, 0); 1018 1019 for (i = reqpage - 1; i >= 0; --i) { 1020 daddr_t iblk; 1021 1022 iblk = swp_pager_meta_ctl(m[i]->object, m[i]->pindex, 0); 1023 if (blk != iblk + (reqpage - i)) 1024 break; 1025 } 1026 ++i; 1027 1028 for (j = reqpage + 1; j < count; ++j) { 1029 daddr_t jblk; 1030 1031 jblk = swp_pager_meta_ctl(m[j]->object, m[j]->pindex, 0); 1032 if (blk != jblk - (j - reqpage)) 1033 break; 1034 } 1035 1036 /* 1037 * free pages outside our collection range. Note: we never free 1038 * mreq, it must remain busy throughout. 1039 */ 1040 vm_page_lock_queues(); 1041 { 1042 int k; 1043 1044 for (k = 0; k < i; ++k) 1045 vm_page_free(m[k]); 1046 for (k = j; k < count; ++k) 1047 vm_page_free(m[k]); 1048 } 1049 vm_page_unlock_queues(); 1050 splx(s); 1051 1052 1053 /* 1054 * Return VM_PAGER_FAIL if we have nothing to do. Return mreq 1055 * still busy, but the others unbusied. 1056 */ 1057 if (blk == SWAPBLK_NONE) 1058 return (VM_PAGER_FAIL); 1059 1060 /* 1061 * Getpbuf() can sleep. 1062 */ 1063 VM_OBJECT_UNLOCK(object); 1064 /* 1065 * Get a swap buffer header to perform the IO 1066 */ 1067 bp = getpbuf(&nsw_rcount); 1068 bp->b_flags |= B_PAGING; 1069 1070 /* 1071 * map our page(s) into kva for input 1072 */ 1073 pmap_qenter((vm_offset_t)bp->b_data, m + i, j - i); 1074 1075 bp->b_iocmd = BIO_READ; 1076 bp->b_iodone = swp_pager_async_iodone; 1077 bp->b_rcred = crhold(thread0.td_ucred); 1078 bp->b_wcred = crhold(thread0.td_ucred); 1079 bp->b_blkno = blk - (reqpage - i); 1080 bp->b_bcount = PAGE_SIZE * (j - i); 1081 bp->b_bufsize = PAGE_SIZE * (j - i); 1082 bp->b_pager.pg_reqpage = reqpage - i; 1083 1084 VM_OBJECT_LOCK(object); 1085 vm_page_lock_queues(); 1086 { 1087 int k; 1088 1089 for (k = i; k < j; ++k) { 1090 bp->b_pages[k - i] = m[k]; 1091 vm_page_flag_set(m[k], PG_SWAPINPROG); 1092 } 1093 } 1094 vm_page_unlock_queues(); 1095 VM_OBJECT_UNLOCK(object); 1096 bp->b_npages = j - i; 1097 1098 cnt.v_swapin++; 1099 cnt.v_swappgsin += bp->b_npages; 1100 1101 /* 1102 * We still hold the lock on mreq, and our automatic completion routine 1103 * does not remove it. 1104 */ 1105 VM_OBJECT_LOCK(mreq->object); 1106 vm_object_pip_add(mreq->object, bp->b_npages); 1107 VM_OBJECT_UNLOCK(mreq->object); 1108 1109 /* 1110 * perform the I/O. NOTE!!! bp cannot be considered valid after 1111 * this point because we automatically release it on completion. 1112 * Instead, we look at the one page we are interested in which we 1113 * still hold a lock on even through the I/O completion. 1114 * 1115 * The other pages in our m[] array are also released on completion, 1116 * so we cannot assume they are valid anymore either. 1117 * 1118 * NOTE: b_blkno is destroyed by the call to swapdev_strategy 1119 */ 1120 BUF_KERNPROC(bp); 1121 swp_pager_strategy(bp); 1122 1123 /* 1124 * wait for the page we want to complete. PG_SWAPINPROG is always 1125 * cleared on completion. If an I/O error occurs, SWAPBLK_NONE 1126 * is set in the meta-data. 1127 */ 1128 s = splvm(); 1129 vm_page_lock_queues(); 1130 while ((mreq->flags & PG_SWAPINPROG) != 0) { 1131 vm_page_flag_set(mreq, PG_WANTED | PG_REFERENCED); 1132 cnt.v_intrans++; 1133 if (msleep(mreq, &vm_page_queue_mtx, PSWP, "swread", hz*20)) { 1134 printf("swap_pager: indefinite wait buffer: device: " 1135 "%s, blkno: %jd, size: %ld\n", 1136 bp->b_dev == NULL ? "[NULL]" : devtoname(bp->b_dev), 1137 (intmax_t)bp->b_blkno, bp->b_bcount); 1138 } 1139 } 1140 vm_page_unlock_queues(); 1141 splx(s); 1142 1143 VM_OBJECT_LOCK(mreq->object); 1144 /* 1145 * mreq is left busied after completion, but all the other pages 1146 * are freed. If we had an unrecoverable read error the page will 1147 * not be valid. 1148 */ 1149 if (mreq->valid != VM_PAGE_BITS_ALL) { 1150 return (VM_PAGER_ERROR); 1151 } else { 1152 return (VM_PAGER_OK); 1153 } 1154 1155 /* 1156 * A final note: in a low swap situation, we cannot deallocate swap 1157 * and mark a page dirty here because the caller is likely to mark 1158 * the page clean when we return, causing the page to possibly revert 1159 * to all-zero's later. 1160 */ 1161 } 1162 1163 /* 1164 * swap_pager_putpages: 1165 * 1166 * Assign swap (if necessary) and initiate I/O on the specified pages. 1167 * 1168 * We support both OBJT_DEFAULT and OBJT_SWAP objects. DEFAULT objects 1169 * are automatically converted to SWAP objects. 1170 * 1171 * In a low memory situation we may block in VOP_STRATEGY(), but the new 1172 * vm_page reservation system coupled with properly written VFS devices 1173 * should ensure that no low-memory deadlock occurs. This is an area 1174 * which needs work. 1175 * 1176 * The parent has N vm_object_pip_add() references prior to 1177 * calling us and will remove references for rtvals[] that are 1178 * not set to VM_PAGER_PEND. We need to remove the rest on I/O 1179 * completion. 1180 * 1181 * The parent has soft-busy'd the pages it passes us and will unbusy 1182 * those whos rtvals[] entry is not set to VM_PAGER_PEND on return. 1183 * We need to unbusy the rest on I/O completion. 1184 */ 1185 void 1186 swap_pager_putpages(vm_object_t object, vm_page_t *m, int count, 1187 boolean_t sync, int *rtvals) 1188 { 1189 int i; 1190 int n = 0; 1191 1192 GIANT_REQUIRED; 1193 if (count && m[0]->object != object) { 1194 panic("swap_pager_getpages: object mismatch %p/%p", 1195 object, 1196 m[0]->object 1197 ); 1198 } 1199 1200 /* 1201 * Step 1 1202 * 1203 * Turn object into OBJT_SWAP 1204 * check for bogus sysops 1205 * force sync if not pageout process 1206 */ 1207 if (object->type != OBJT_SWAP) 1208 swp_pager_meta_build(object, 0, SWAPBLK_NONE); 1209 VM_OBJECT_UNLOCK(object); 1210 1211 if (curproc != pageproc) 1212 sync = TRUE; 1213 1214 /* 1215 * Step 2 1216 * 1217 * Update nsw parameters from swap_async_max sysctl values. 1218 * Do not let the sysop crash the machine with bogus numbers. 1219 */ 1220 mtx_lock(&pbuf_mtx); 1221 if (swap_async_max != nsw_wcount_async_max) { 1222 int n; 1223 int s; 1224 1225 /* 1226 * limit range 1227 */ 1228 if ((n = swap_async_max) > nswbuf / 2) 1229 n = nswbuf / 2; 1230 if (n < 1) 1231 n = 1; 1232 swap_async_max = n; 1233 1234 /* 1235 * Adjust difference ( if possible ). If the current async 1236 * count is too low, we may not be able to make the adjustment 1237 * at this time. 1238 */ 1239 s = splvm(); 1240 n -= nsw_wcount_async_max; 1241 if (nsw_wcount_async + n >= 0) { 1242 nsw_wcount_async += n; 1243 nsw_wcount_async_max += n; 1244 wakeup(&nsw_wcount_async); 1245 } 1246 splx(s); 1247 } 1248 mtx_unlock(&pbuf_mtx); 1249 1250 /* 1251 * Step 3 1252 * 1253 * Assign swap blocks and issue I/O. We reallocate swap on the fly. 1254 * The page is left dirty until the pageout operation completes 1255 * successfully. 1256 */ 1257 for (i = 0; i < count; i += n) { 1258 int s; 1259 int j; 1260 struct buf *bp; 1261 daddr_t blk; 1262 1263 /* 1264 * Maximum I/O size is limited by a number of factors. 1265 */ 1266 n = min(BLIST_MAX_ALLOC, count - i); 1267 n = min(n, nsw_cluster_max); 1268 1269 s = splvm(); 1270 1271 /* 1272 * Get biggest block of swap we can. If we fail, fall 1273 * back and try to allocate a smaller block. Don't go 1274 * overboard trying to allocate space if it would overly 1275 * fragment swap. 1276 */ 1277 while ( 1278 (blk = swp_pager_getswapspace(n)) == SWAPBLK_NONE && 1279 n > 4 1280 ) { 1281 n >>= 1; 1282 } 1283 if (blk == SWAPBLK_NONE) { 1284 for (j = 0; j < n; ++j) 1285 rtvals[i+j] = VM_PAGER_FAIL; 1286 splx(s); 1287 continue; 1288 } 1289 1290 /* 1291 * All I/O parameters have been satisfied, build the I/O 1292 * request and assign the swap space. 1293 */ 1294 if (sync == TRUE) { 1295 bp = getpbuf(&nsw_wcount_sync); 1296 } else { 1297 bp = getpbuf(&nsw_wcount_async); 1298 bp->b_flags = B_ASYNC; 1299 } 1300 bp->b_flags |= B_PAGING; 1301 bp->b_iocmd = BIO_WRITE; 1302 1303 pmap_qenter((vm_offset_t)bp->b_data, &m[i], n); 1304 1305 bp->b_rcred = crhold(thread0.td_ucred); 1306 bp->b_wcred = crhold(thread0.td_ucred); 1307 bp->b_bcount = PAGE_SIZE * n; 1308 bp->b_bufsize = PAGE_SIZE * n; 1309 bp->b_blkno = blk; 1310 1311 VM_OBJECT_LOCK(object); 1312 for (j = 0; j < n; ++j) { 1313 vm_page_t mreq = m[i+j]; 1314 1315 swp_pager_meta_build( 1316 mreq->object, 1317 mreq->pindex, 1318 blk + j 1319 ); 1320 vm_page_dirty(mreq); 1321 rtvals[i+j] = VM_PAGER_OK; 1322 1323 vm_page_lock_queues(); 1324 vm_page_flag_set(mreq, PG_SWAPINPROG); 1325 vm_page_unlock_queues(); 1326 bp->b_pages[j] = mreq; 1327 } 1328 VM_OBJECT_UNLOCK(object); 1329 bp->b_npages = n; 1330 /* 1331 * Must set dirty range for NFS to work. 1332 */ 1333 bp->b_dirtyoff = 0; 1334 bp->b_dirtyend = bp->b_bcount; 1335 1336 cnt.v_swapout++; 1337 cnt.v_swappgsout += bp->b_npages; 1338 1339 splx(s); 1340 1341 /* 1342 * asynchronous 1343 * 1344 * NOTE: b_blkno is destroyed by the call to swapdev_strategy 1345 */ 1346 if (sync == FALSE) { 1347 bp->b_iodone = swp_pager_async_iodone; 1348 BUF_KERNPROC(bp); 1349 swp_pager_strategy(bp); 1350 1351 for (j = 0; j < n; ++j) 1352 rtvals[i+j] = VM_PAGER_PEND; 1353 /* restart outter loop */ 1354 continue; 1355 } 1356 1357 /* 1358 * synchronous 1359 * 1360 * NOTE: b_blkno is destroyed by the call to swapdev_strategy 1361 */ 1362 bp->b_iodone = bdone; 1363 swp_pager_strategy(bp); 1364 1365 /* 1366 * Wait for the sync I/O to complete, then update rtvals. 1367 * We just set the rtvals[] to VM_PAGER_PEND so we can call 1368 * our async completion routine at the end, thus avoiding a 1369 * double-free. 1370 */ 1371 s = splbio(); 1372 bwait(bp, PVM, "swwrt"); 1373 for (j = 0; j < n; ++j) 1374 rtvals[i+j] = VM_PAGER_PEND; 1375 /* 1376 * Now that we are through with the bp, we can call the 1377 * normal async completion, which frees everything up. 1378 */ 1379 swp_pager_async_iodone(bp); 1380 splx(s); 1381 } 1382 VM_OBJECT_LOCK(object); 1383 } 1384 1385 /* 1386 * swp_pager_async_iodone: 1387 * 1388 * Completion routine for asynchronous reads and writes from/to swap. 1389 * Also called manually by synchronous code to finish up a bp. 1390 * 1391 * For READ operations, the pages are PG_BUSY'd. For WRITE operations, 1392 * the pages are vm_page_t->busy'd. For READ operations, we PG_BUSY 1393 * unbusy all pages except the 'main' request page. For WRITE 1394 * operations, we vm_page_t->busy'd unbusy all pages ( we can do this 1395 * because we marked them all VM_PAGER_PEND on return from putpages ). 1396 * 1397 * This routine may not block. 1398 * This routine is called at splbio() or better 1399 * 1400 * We up ourselves to splvm() as required for various vm_page related 1401 * calls. 1402 */ 1403 static void 1404 swp_pager_async_iodone(struct buf *bp) 1405 { 1406 int s; 1407 int i; 1408 vm_object_t object = NULL; 1409 1410 bp->b_flags |= B_DONE; 1411 1412 /* 1413 * report error 1414 */ 1415 if (bp->b_ioflags & BIO_ERROR) { 1416 printf( 1417 "swap_pager: I/O error - %s failed; blkno %ld," 1418 "size %ld, error %d\n", 1419 ((bp->b_iocmd == BIO_READ) ? "pagein" : "pageout"), 1420 (long)bp->b_blkno, 1421 (long)bp->b_bcount, 1422 bp->b_error 1423 ); 1424 } 1425 1426 /* 1427 * set object, raise to splvm(). 1428 */ 1429 s = splvm(); 1430 1431 /* 1432 * remove the mapping for kernel virtual 1433 */ 1434 pmap_qremove((vm_offset_t)bp->b_data, bp->b_npages); 1435 1436 if (bp->b_npages) { 1437 object = bp->b_pages[0]->object; 1438 VM_OBJECT_LOCK(object); 1439 } 1440 vm_page_lock_queues(); 1441 /* 1442 * cleanup pages. If an error occurs writing to swap, we are in 1443 * very serious trouble. If it happens to be a disk error, though, 1444 * we may be able to recover by reassigning the swap later on. So 1445 * in this case we remove the m->swapblk assignment for the page 1446 * but do not free it in the rlist. The errornous block(s) are thus 1447 * never reallocated as swap. Redirty the page and continue. 1448 */ 1449 for (i = 0; i < bp->b_npages; ++i) { 1450 vm_page_t m = bp->b_pages[i]; 1451 1452 vm_page_flag_clear(m, PG_SWAPINPROG); 1453 1454 if (bp->b_ioflags & BIO_ERROR) { 1455 /* 1456 * If an error occurs I'd love to throw the swapblk 1457 * away without freeing it back to swapspace, so it 1458 * can never be used again. But I can't from an 1459 * interrupt. 1460 */ 1461 if (bp->b_iocmd == BIO_READ) { 1462 /* 1463 * When reading, reqpage needs to stay 1464 * locked for the parent, but all other 1465 * pages can be freed. We still want to 1466 * wakeup the parent waiting on the page, 1467 * though. ( also: pg_reqpage can be -1 and 1468 * not match anything ). 1469 * 1470 * We have to wake specifically requested pages 1471 * up too because we cleared PG_SWAPINPROG and 1472 * someone may be waiting for that. 1473 * 1474 * NOTE: for reads, m->dirty will probably 1475 * be overridden by the original caller of 1476 * getpages so don't play cute tricks here. 1477 * 1478 * XXX IT IS NOT LEGAL TO FREE THE PAGE HERE 1479 * AS THIS MESSES WITH object->memq, and it is 1480 * not legal to mess with object->memq from an 1481 * interrupt. 1482 */ 1483 m->valid = 0; 1484 if (i != bp->b_pager.pg_reqpage) 1485 vm_page_free(m); 1486 else 1487 vm_page_flash(m); 1488 /* 1489 * If i == bp->b_pager.pg_reqpage, do not wake 1490 * the page up. The caller needs to. 1491 */ 1492 } else { 1493 /* 1494 * If a write error occurs, reactivate page 1495 * so it doesn't clog the inactive list, 1496 * then finish the I/O. 1497 */ 1498 vm_page_dirty(m); 1499 vm_page_activate(m); 1500 vm_page_io_finish(m); 1501 } 1502 } else if (bp->b_iocmd == BIO_READ) { 1503 /* 1504 * For read success, clear dirty bits. Nobody should 1505 * have this page mapped but don't take any chances, 1506 * make sure the pmap modify bits are also cleared. 1507 * 1508 * NOTE: for reads, m->dirty will probably be 1509 * overridden by the original caller of getpages so 1510 * we cannot set them in order to free the underlying 1511 * swap in a low-swap situation. I don't think we'd 1512 * want to do that anyway, but it was an optimization 1513 * that existed in the old swapper for a time before 1514 * it got ripped out due to precisely this problem. 1515 * 1516 * If not the requested page then deactivate it. 1517 * 1518 * Note that the requested page, reqpage, is left 1519 * busied, but we still have to wake it up. The 1520 * other pages are released (unbusied) by 1521 * vm_page_wakeup(). We do not set reqpage's 1522 * valid bits here, it is up to the caller. 1523 */ 1524 pmap_clear_modify(m); 1525 m->valid = VM_PAGE_BITS_ALL; 1526 vm_page_undirty(m); 1527 1528 /* 1529 * We have to wake specifically requested pages 1530 * up too because we cleared PG_SWAPINPROG and 1531 * could be waiting for it in getpages. However, 1532 * be sure to not unbusy getpages specifically 1533 * requested page - getpages expects it to be 1534 * left busy. 1535 */ 1536 if (i != bp->b_pager.pg_reqpage) { 1537 vm_page_deactivate(m); 1538 vm_page_wakeup(m); 1539 } else { 1540 vm_page_flash(m); 1541 } 1542 } else { 1543 /* 1544 * For write success, clear the modify and dirty 1545 * status, then finish the I/O ( which decrements the 1546 * busy count and possibly wakes waiter's up ). 1547 */ 1548 pmap_clear_modify(m); 1549 vm_page_undirty(m); 1550 vm_page_io_finish(m); 1551 if (vm_page_count_severe()) 1552 vm_page_try_to_cache(m); 1553 } 1554 } 1555 vm_page_unlock_queues(); 1556 1557 /* 1558 * adjust pip. NOTE: the original parent may still have its own 1559 * pip refs on the object. 1560 */ 1561 if (object != NULL) { 1562 vm_object_pip_wakeupn(object, bp->b_npages); 1563 VM_OBJECT_UNLOCK(object); 1564 } 1565 1566 /* 1567 * release the physical I/O buffer 1568 */ 1569 relpbuf( 1570 bp, 1571 ((bp->b_iocmd == BIO_READ) ? &nsw_rcount : 1572 ((bp->b_flags & B_ASYNC) ? 1573 &nsw_wcount_async : 1574 &nsw_wcount_sync 1575 ) 1576 ) 1577 ); 1578 splx(s); 1579 } 1580 1581 /* 1582 * swap_pager_isswapped: 1583 * 1584 * Return 1 if at least one page in the given object is paged 1585 * out to the given swap device. 1586 * 1587 * This routine may not block. 1588 */ 1589 int 1590 swap_pager_isswapped(vm_object_t object, struct swdevt *sp) 1591 { 1592 daddr_t index = 0; 1593 int bcount; 1594 int i; 1595 1596 VM_OBJECT_LOCK_ASSERT(object, MA_OWNED); 1597 for (bcount = 0; bcount < object->un_pager.swp.swp_bcount; bcount++) { 1598 struct swblock *swap; 1599 1600 mtx_lock(&swhash_mtx); 1601 if ((swap = *swp_pager_hash(object, index)) != NULL) { 1602 for (i = 0; i < SWAP_META_PAGES; ++i) { 1603 daddr_t v = swap->swb_pages[i]; 1604 if (v == SWAPBLK_NONE) 1605 continue; 1606 if (swp_pager_find_dev(v) == sp) { 1607 mtx_unlock(&swhash_mtx); 1608 return 1; 1609 } 1610 } 1611 } 1612 mtx_unlock(&swhash_mtx); 1613 index += SWAP_META_PAGES; 1614 if (index > 0x20000000) 1615 panic("swap_pager_isswapped: failed to locate all swap meta blocks"); 1616 } 1617 return 0; 1618 } 1619 1620 /* 1621 * SWP_PAGER_FORCE_PAGEIN() - force a swap block to be paged in 1622 * 1623 * This routine dissociates the page at the given index within a 1624 * swap block from its backing store, paging it in if necessary. 1625 * If the page is paged in, it is placed in the inactive queue, 1626 * since it had its backing store ripped out from under it. 1627 * We also attempt to swap in all other pages in the swap block, 1628 * we only guarantee that the one at the specified index is 1629 * paged in. 1630 * 1631 * XXX - The code to page the whole block in doesn't work, so we 1632 * revert to the one-by-one behavior for now. Sigh. 1633 */ 1634 static __inline void 1635 swp_pager_force_pagein(struct swblock *swap, int idx) 1636 { 1637 vm_object_t object; 1638 vm_page_t m; 1639 vm_pindex_t pindex; 1640 1641 object = swap->swb_object; 1642 pindex = swap->swb_index; 1643 mtx_unlock(&swhash_mtx); 1644 1645 VM_OBJECT_LOCK(object); 1646 vm_object_pip_add(object, 1); 1647 m = vm_page_grab(object, pindex + idx, VM_ALLOC_NORMAL|VM_ALLOC_RETRY); 1648 if (m->valid == VM_PAGE_BITS_ALL) { 1649 vm_object_pip_subtract(object, 1); 1650 vm_page_lock_queues(); 1651 vm_page_activate(m); 1652 vm_page_dirty(m); 1653 vm_page_wakeup(m); 1654 vm_page_unlock_queues(); 1655 vm_pager_page_unswapped(m); 1656 VM_OBJECT_UNLOCK(object); 1657 return; 1658 } 1659 1660 if (swap_pager_getpages(object, &m, 1, 0) != 1661 VM_PAGER_OK) 1662 panic("swap_pager_force_pagein: read from swap failed");/*XXX*/ 1663 vm_object_pip_subtract(object, 1); 1664 vm_page_lock_queues(); 1665 vm_page_dirty(m); 1666 vm_page_dontneed(m); 1667 vm_page_wakeup(m); 1668 vm_page_unlock_queues(); 1669 vm_pager_page_unswapped(m); 1670 VM_OBJECT_UNLOCK(object); 1671 } 1672 1673 1674 /* 1675 * swap_pager_swapoff: 1676 * 1677 * Page in all of the pages that have been paged out to the 1678 * given device. The corresponding blocks in the bitmap must be 1679 * marked as allocated and the device must be flagged SW_CLOSING. 1680 * There may be no processes swapped out to the device. 1681 * 1682 * The sw_used parameter points to the field in the swdev structure 1683 * that contains a count of the number of blocks still allocated 1684 * on the device. If we encounter objects with a nonzero pip count 1685 * in our scan, we use this number to determine if we're really done. 1686 * 1687 * This routine may block. 1688 */ 1689 static void 1690 swap_pager_swapoff(struct swdevt *sp, int *sw_used) 1691 { 1692 struct swblock **pswap; 1693 struct swblock *swap; 1694 vm_object_t waitobj; 1695 daddr_t v; 1696 int i, j; 1697 1698 GIANT_REQUIRED; 1699 1700 full_rescan: 1701 waitobj = NULL; 1702 for (i = 0; i <= swhash_mask; i++) { /* '<=' is correct here */ 1703 restart: 1704 pswap = &swhash[i]; 1705 mtx_lock(&swhash_mtx); 1706 while ((swap = *pswap) != NULL) { 1707 for (j = 0; j < SWAP_META_PAGES; ++j) { 1708 v = swap->swb_pages[j]; 1709 if (v != SWAPBLK_NONE && 1710 swp_pager_find_dev(v) == sp) 1711 break; 1712 } 1713 if (j < SWAP_META_PAGES) { 1714 swp_pager_force_pagein(swap, j); 1715 goto restart; 1716 } else if (swap->swb_object->paging_in_progress) { 1717 if (!waitobj) 1718 waitobj = swap->swb_object; 1719 } 1720 pswap = &swap->swb_hnext; 1721 } 1722 mtx_unlock(&swhash_mtx); 1723 } 1724 if (waitobj && *sw_used) { 1725 /* 1726 * We wait on an arbitrary object to clock our rescans 1727 * to the rate of paging completion. 1728 */ 1729 VM_OBJECT_LOCK(waitobj); 1730 vm_object_pip_wait(waitobj, "swpoff"); 1731 VM_OBJECT_UNLOCK(waitobj); 1732 goto full_rescan; 1733 } 1734 if (*sw_used) 1735 panic("swapoff: failed to locate %d swap blocks", *sw_used); 1736 } 1737 1738 /************************************************************************ 1739 * SWAP META DATA * 1740 ************************************************************************ 1741 * 1742 * These routines manipulate the swap metadata stored in the 1743 * OBJT_SWAP object. All swp_*() routines must be called at 1744 * splvm() because swap can be freed up by the low level vm_page 1745 * code which might be called from interrupts beyond what splbio() covers. 1746 * 1747 * Swap metadata is implemented with a global hash and not directly 1748 * linked into the object. Instead the object simply contains 1749 * appropriate tracking counters. 1750 */ 1751 1752 /* 1753 * SWP_PAGER_META_BUILD() - add swap block to swap meta data for object 1754 * 1755 * We first convert the object to a swap object if it is a default 1756 * object. 1757 * 1758 * The specified swapblk is added to the object's swap metadata. If 1759 * the swapblk is not valid, it is freed instead. Any previously 1760 * assigned swapblk is freed. 1761 * 1762 * This routine must be called at splvm(), except when used to convert 1763 * an OBJT_DEFAULT object into an OBJT_SWAP object. 1764 */ 1765 static void 1766 swp_pager_meta_build(vm_object_t object, vm_pindex_t pindex, daddr_t swapblk) 1767 { 1768 struct swblock *swap; 1769 struct swblock **pswap; 1770 int idx; 1771 1772 VM_OBJECT_LOCK_ASSERT(object, MA_OWNED); 1773 /* 1774 * Convert default object to swap object if necessary 1775 */ 1776 if (object->type != OBJT_SWAP) { 1777 object->type = OBJT_SWAP; 1778 object->un_pager.swp.swp_bcount = 0; 1779 1780 if (object->handle != NULL) { 1781 mtx_lock(&sw_alloc_mtx); 1782 TAILQ_INSERT_TAIL( 1783 NOBJLIST(object->handle), 1784 object, 1785 pager_object_list 1786 ); 1787 mtx_unlock(&sw_alloc_mtx); 1788 } 1789 } 1790 1791 /* 1792 * Locate hash entry. If not found create, but if we aren't adding 1793 * anything just return. If we run out of space in the map we wait 1794 * and, since the hash table may have changed, retry. 1795 */ 1796 retry: 1797 mtx_lock(&swhash_mtx); 1798 pswap = swp_pager_hash(object, pindex); 1799 1800 if ((swap = *pswap) == NULL) { 1801 int i; 1802 1803 if (swapblk == SWAPBLK_NONE) 1804 goto done; 1805 1806 swap = *pswap = uma_zalloc(swap_zone, M_NOWAIT); 1807 if (swap == NULL) { 1808 mtx_unlock(&swhash_mtx); 1809 VM_OBJECT_UNLOCK(object); 1810 VM_WAIT; 1811 VM_OBJECT_LOCK(object); 1812 goto retry; 1813 } 1814 1815 swap->swb_hnext = NULL; 1816 swap->swb_object = object; 1817 swap->swb_index = pindex & ~(vm_pindex_t)SWAP_META_MASK; 1818 swap->swb_count = 0; 1819 1820 ++object->un_pager.swp.swp_bcount; 1821 1822 for (i = 0; i < SWAP_META_PAGES; ++i) 1823 swap->swb_pages[i] = SWAPBLK_NONE; 1824 } 1825 1826 /* 1827 * Delete prior contents of metadata 1828 */ 1829 idx = pindex & SWAP_META_MASK; 1830 1831 if (swap->swb_pages[idx] != SWAPBLK_NONE) { 1832 swp_pager_freeswapspace(swap->swb_pages[idx], 1); 1833 --swap->swb_count; 1834 } 1835 1836 /* 1837 * Enter block into metadata 1838 */ 1839 swap->swb_pages[idx] = swapblk; 1840 if (swapblk != SWAPBLK_NONE) 1841 ++swap->swb_count; 1842 done: 1843 mtx_unlock(&swhash_mtx); 1844 } 1845 1846 /* 1847 * SWP_PAGER_META_FREE() - free a range of blocks in the object's swap metadata 1848 * 1849 * The requested range of blocks is freed, with any associated swap 1850 * returned to the swap bitmap. 1851 * 1852 * This routine will free swap metadata structures as they are cleaned 1853 * out. This routine does *NOT* operate on swap metadata associated 1854 * with resident pages. 1855 * 1856 * This routine must be called at splvm() 1857 */ 1858 static void 1859 swp_pager_meta_free(vm_object_t object, vm_pindex_t index, daddr_t count) 1860 { 1861 1862 VM_OBJECT_LOCK_ASSERT(object, MA_OWNED); 1863 if (object->type != OBJT_SWAP) 1864 return; 1865 1866 while (count > 0) { 1867 struct swblock **pswap; 1868 struct swblock *swap; 1869 1870 mtx_lock(&swhash_mtx); 1871 pswap = swp_pager_hash(object, index); 1872 1873 if ((swap = *pswap) != NULL) { 1874 daddr_t v = swap->swb_pages[index & SWAP_META_MASK]; 1875 1876 if (v != SWAPBLK_NONE) { 1877 swp_pager_freeswapspace(v, 1); 1878 swap->swb_pages[index & SWAP_META_MASK] = 1879 SWAPBLK_NONE; 1880 if (--swap->swb_count == 0) { 1881 *pswap = swap->swb_hnext; 1882 uma_zfree(swap_zone, swap); 1883 --object->un_pager.swp.swp_bcount; 1884 } 1885 } 1886 --count; 1887 ++index; 1888 } else { 1889 int n = SWAP_META_PAGES - (index & SWAP_META_MASK); 1890 count -= n; 1891 index += n; 1892 } 1893 mtx_unlock(&swhash_mtx); 1894 } 1895 } 1896 1897 /* 1898 * SWP_PAGER_META_FREE_ALL() - destroy all swap metadata associated with object 1899 * 1900 * This routine locates and destroys all swap metadata associated with 1901 * an object. 1902 * 1903 * This routine must be called at splvm() 1904 */ 1905 static void 1906 swp_pager_meta_free_all(vm_object_t object) 1907 { 1908 daddr_t index = 0; 1909 1910 VM_OBJECT_LOCK_ASSERT(object, MA_OWNED); 1911 if (object->type != OBJT_SWAP) 1912 return; 1913 1914 while (object->un_pager.swp.swp_bcount) { 1915 struct swblock **pswap; 1916 struct swblock *swap; 1917 1918 mtx_lock(&swhash_mtx); 1919 pswap = swp_pager_hash(object, index); 1920 if ((swap = *pswap) != NULL) { 1921 int i; 1922 1923 for (i = 0; i < SWAP_META_PAGES; ++i) { 1924 daddr_t v = swap->swb_pages[i]; 1925 if (v != SWAPBLK_NONE) { 1926 --swap->swb_count; 1927 swp_pager_freeswapspace(v, 1); 1928 } 1929 } 1930 if (swap->swb_count != 0) 1931 panic("swap_pager_meta_free_all: swb_count != 0"); 1932 *pswap = swap->swb_hnext; 1933 uma_zfree(swap_zone, swap); 1934 --object->un_pager.swp.swp_bcount; 1935 } 1936 mtx_unlock(&swhash_mtx); 1937 index += SWAP_META_PAGES; 1938 if (index > 0x20000000) 1939 panic("swp_pager_meta_free_all: failed to locate all swap meta blocks"); 1940 } 1941 } 1942 1943 /* 1944 * SWP_PAGER_METACTL() - misc control of swap and vm_page_t meta data. 1945 * 1946 * This routine is capable of looking up, popping, or freeing 1947 * swapblk assignments in the swap meta data or in the vm_page_t. 1948 * The routine typically returns the swapblk being looked-up, or popped, 1949 * or SWAPBLK_NONE if the block was freed, or SWAPBLK_NONE if the block 1950 * was invalid. This routine will automatically free any invalid 1951 * meta-data swapblks. 1952 * 1953 * It is not possible to store invalid swapblks in the swap meta data 1954 * (other then a literal 'SWAPBLK_NONE'), so we don't bother checking. 1955 * 1956 * When acting on a busy resident page and paging is in progress, we 1957 * have to wait until paging is complete but otherwise can act on the 1958 * busy page. 1959 * 1960 * This routine must be called at splvm(). 1961 * 1962 * SWM_FREE remove and free swap block from metadata 1963 * SWM_POP remove from meta data but do not free.. pop it out 1964 */ 1965 static daddr_t 1966 swp_pager_meta_ctl(vm_object_t object, vm_pindex_t pindex, int flags) 1967 { 1968 struct swblock **pswap; 1969 struct swblock *swap; 1970 daddr_t r1; 1971 int idx; 1972 1973 VM_OBJECT_LOCK_ASSERT(object, MA_OWNED); 1974 /* 1975 * The meta data only exists of the object is OBJT_SWAP 1976 * and even then might not be allocated yet. 1977 */ 1978 if (object->type != OBJT_SWAP) 1979 return (SWAPBLK_NONE); 1980 1981 r1 = SWAPBLK_NONE; 1982 mtx_lock(&swhash_mtx); 1983 pswap = swp_pager_hash(object, pindex); 1984 1985 if ((swap = *pswap) != NULL) { 1986 idx = pindex & SWAP_META_MASK; 1987 r1 = swap->swb_pages[idx]; 1988 1989 if (r1 != SWAPBLK_NONE) { 1990 if (flags & SWM_FREE) { 1991 swp_pager_freeswapspace(r1, 1); 1992 r1 = SWAPBLK_NONE; 1993 } 1994 if (flags & (SWM_FREE|SWM_POP)) { 1995 swap->swb_pages[idx] = SWAPBLK_NONE; 1996 if (--swap->swb_count == 0) { 1997 *pswap = swap->swb_hnext; 1998 uma_zfree(swap_zone, swap); 1999 --object->un_pager.swp.swp_bcount; 2000 } 2001 } 2002 } 2003 } 2004 mtx_unlock(&swhash_mtx); 2005 return (r1); 2006 } 2007 2008 /* 2009 * System call swapon(name) enables swapping on device name, 2010 * which must be in the swdevsw. Return EBUSY 2011 * if already swapping on this device. 2012 */ 2013 #ifndef _SYS_SYSPROTO_H_ 2014 struct swapon_args { 2015 char *name; 2016 }; 2017 #endif 2018 2019 /* 2020 * MPSAFE 2021 */ 2022 /* ARGSUSED */ 2023 int 2024 swapon(struct thread *td, struct swapon_args *uap) 2025 { 2026 struct vattr attr; 2027 struct vnode *vp; 2028 struct nameidata nd; 2029 int error; 2030 2031 mtx_lock(&Giant); 2032 error = suser(td); 2033 if (error) 2034 goto done2; 2035 2036 while (swdev_syscall_active) 2037 tsleep(&swdev_syscall_active, PUSER - 1, "swpon", 0); 2038 swdev_syscall_active = 1; 2039 2040 /* 2041 * Swap metadata may not fit in the KVM if we have physical 2042 * memory of >1GB. 2043 */ 2044 if (swap_zone == NULL) { 2045 error = ENOMEM; 2046 goto done; 2047 } 2048 2049 NDINIT(&nd, LOOKUP, FOLLOW, UIO_USERSPACE, uap->name, td); 2050 error = namei(&nd); 2051 if (error) 2052 goto done; 2053 2054 NDFREE(&nd, NDF_ONLY_PNBUF); 2055 vp = nd.ni_vp; 2056 2057 if (vn_isdisk(vp, &error)) { 2058 error = swapongeom(td, vp); 2059 } else if (vp->v_type == VREG && 2060 (vp->v_mount->mnt_vfc->vfc_flags & VFCF_NETWORK) != 0 && 2061 (error = VOP_GETATTR(vp, &attr, td->td_ucred, td)) == 0) { 2062 /* 2063 * Allow direct swapping to NFS regular files in the same 2064 * way that nfs_mountroot() sets up diskless swapping. 2065 */ 2066 error = swaponvp(td, vp, attr.va_size / DEV_BSIZE); 2067 } 2068 2069 if (error) 2070 vrele(vp); 2071 done: 2072 swdev_syscall_active = 0; 2073 wakeup_one(&swdev_syscall_active); 2074 done2: 2075 mtx_unlock(&Giant); 2076 return (error); 2077 } 2078 2079 static void 2080 swaponsomething(struct vnode *vp, void *id, u_long nblks, sw_strategy_t *strategy, sw_close_t *close, dev_t dev) 2081 { 2082 struct swdevt *sp, *tsp; 2083 swblk_t dvbase; 2084 u_long mblocks; 2085 2086 /* 2087 * If we go beyond this, we get overflows in the radix 2088 * tree bitmap code. 2089 */ 2090 mblocks = 0x40000000 / BLIST_META_RADIX; 2091 if (nblks > mblocks) { 2092 printf("WARNING: reducing size to maximum of %lu blocks per swap unit\n", 2093 mblocks); 2094 nblks = mblocks; 2095 } 2096 /* 2097 * nblks is in DEV_BSIZE'd chunks, convert to PAGE_SIZE'd chunks. 2098 * First chop nblks off to page-align it, then convert. 2099 * 2100 * sw->sw_nblks is in page-sized chunks now too. 2101 */ 2102 nblks &= ~(ctodb(1) - 1); 2103 nblks = dbtoc(nblks); 2104 2105 sp = malloc(sizeof *sp, M_VMPGDATA, M_WAITOK | M_ZERO); 2106 sp->sw_vp = vp; 2107 sp->sw_id = id; 2108 sp->sw_dev = dev; 2109 sp->sw_flags = 0; 2110 sp->sw_nblks = nblks; 2111 sp->sw_used = 0; 2112 sp->sw_strategy = strategy; 2113 sp->sw_close = close; 2114 2115 sp->sw_blist = blist_create(nblks); 2116 /* 2117 * Do not free the first two block in order to avoid overwriting 2118 * any bsd label at the front of the partition 2119 */ 2120 blist_free(sp->sw_blist, 2, nblks - 2); 2121 2122 dvbase = 0; 2123 mtx_lock(&sw_dev_mtx); 2124 TAILQ_FOREACH(tsp, &swtailq, sw_list) { 2125 if (tsp->sw_end >= dvbase) { 2126 /* 2127 * We put one uncovered page between the devices 2128 * in order to definitively prevent any cross-device 2129 * I/O requests 2130 */ 2131 dvbase = tsp->sw_end + 1; 2132 } 2133 } 2134 sp->sw_first = dvbase; 2135 sp->sw_end = dvbase + nblks; 2136 TAILQ_INSERT_TAIL(&swtailq, sp, sw_list); 2137 nswapdev++; 2138 swap_pager_avail += nblks; 2139 swp_sizecheck(); 2140 mtx_unlock(&sw_dev_mtx); 2141 } 2142 2143 /* 2144 * SYSCALL: swapoff(devname) 2145 * 2146 * Disable swapping on the given device. 2147 * 2148 * XXX: Badly designed system call: it should use a device index 2149 * rather than filename as specification. We keep sw_vp around 2150 * only to make this work. 2151 */ 2152 #ifndef _SYS_SYSPROTO_H_ 2153 struct swapoff_args { 2154 char *name; 2155 }; 2156 #endif 2157 2158 /* 2159 * MPSAFE 2160 */ 2161 /* ARGSUSED */ 2162 int 2163 swapoff(struct thread *td, struct swapoff_args *uap) 2164 { 2165 struct vnode *vp; 2166 struct nameidata nd; 2167 struct swdevt *sp; 2168 u_long nblks, dvbase; 2169 int error; 2170 2171 mtx_lock(&Giant); 2172 2173 error = suser(td); 2174 if (error) 2175 goto done2; 2176 2177 while (swdev_syscall_active) 2178 tsleep(&swdev_syscall_active, PUSER - 1, "swpoff", 0); 2179 swdev_syscall_active = 1; 2180 2181 NDINIT(&nd, LOOKUP, FOLLOW, UIO_USERSPACE, uap->name, td); 2182 error = namei(&nd); 2183 if (error) 2184 goto done; 2185 NDFREE(&nd, NDF_ONLY_PNBUF); 2186 vp = nd.ni_vp; 2187 2188 mtx_lock(&sw_dev_mtx); 2189 TAILQ_FOREACH(sp, &swtailq, sw_list) { 2190 if (sp->sw_vp == vp) 2191 goto found; 2192 } 2193 mtx_unlock(&sw_dev_mtx); 2194 error = EINVAL; 2195 goto done; 2196 found: 2197 mtx_unlock(&sw_dev_mtx); 2198 #ifdef MAC 2199 (void) vn_lock(vp, LK_EXCLUSIVE | LK_RETRY, td); 2200 error = mac_check_system_swapoff(td->td_ucred, vp); 2201 (void) VOP_UNLOCK(vp, 0, td); 2202 if (error != 0) 2203 goto done; 2204 #endif 2205 2206 nblks = sp->sw_nblks; 2207 2208 /* 2209 * We can turn off this swap device safely only if the 2210 * available virtual memory in the system will fit the amount 2211 * of data we will have to page back in, plus an epsilon so 2212 * the system doesn't become critically low on swap space. 2213 */ 2214 if (cnt.v_free_count + cnt.v_cache_count + swap_pager_avail < 2215 nblks + nswap_lowat) { 2216 error = ENOMEM; 2217 goto done; 2218 } 2219 2220 /* 2221 * Prevent further allocations on this device. 2222 */ 2223 mtx_lock(&sw_dev_mtx); 2224 sp->sw_flags |= SW_CLOSING; 2225 for (dvbase = 0; dvbase < sp->sw_end; dvbase += dmmax) { 2226 swap_pager_avail -= blist_fill(sp->sw_blist, 2227 dvbase, dmmax); 2228 } 2229 mtx_unlock(&sw_dev_mtx); 2230 2231 /* 2232 * Page in the contents of the device and close it. 2233 */ 2234 #ifndef NO_SWAPPING 2235 vm_proc_swapin_all(sp); 2236 #endif /* !NO_SWAPPING */ 2237 swap_pager_swapoff(sp, &sp->sw_used); 2238 2239 sp->sw_close(td, sp); 2240 sp->sw_id = NULL; 2241 mtx_lock(&sw_dev_mtx); 2242 TAILQ_REMOVE(&swtailq, sp, sw_list); 2243 nswapdev--; 2244 if (nswapdev == 0) { 2245 swap_pager_full = 2; 2246 swap_pager_almost_full = 1; 2247 } 2248 if (swdevhd == sp) 2249 swdevhd = NULL; 2250 mtx_unlock(&sw_dev_mtx); 2251 blist_destroy(sp->sw_blist); 2252 free(sp, M_VMPGDATA); 2253 2254 done: 2255 swdev_syscall_active = 0; 2256 wakeup_one(&swdev_syscall_active); 2257 done2: 2258 mtx_unlock(&Giant); 2259 return (error); 2260 } 2261 2262 void 2263 swap_pager_status(int *total, int *used) 2264 { 2265 struct swdevt *sp; 2266 2267 *total = 0; 2268 *used = 0; 2269 mtx_lock(&sw_dev_mtx); 2270 TAILQ_FOREACH(sp, &swtailq, sw_list) { 2271 *total += sp->sw_nblks; 2272 *used += sp->sw_used; 2273 } 2274 mtx_unlock(&sw_dev_mtx); 2275 } 2276 2277 static int 2278 sysctl_vm_swap_info(SYSCTL_HANDLER_ARGS) 2279 { 2280 int *name = (int *)arg1; 2281 int error, n; 2282 struct xswdev xs; 2283 struct swdevt *sp; 2284 2285 if (arg2 != 1) /* name length */ 2286 return (EINVAL); 2287 2288 n = 0; 2289 mtx_lock(&sw_dev_mtx); 2290 TAILQ_FOREACH(sp, &swtailq, sw_list) { 2291 if (n == *name) { 2292 mtx_unlock(&sw_dev_mtx); 2293 xs.xsw_version = XSWDEV_VERSION; 2294 xs.xsw_dev = sp->sw_dev; 2295 xs.xsw_flags = sp->sw_flags; 2296 xs.xsw_nblks = sp->sw_nblks; 2297 xs.xsw_used = sp->sw_used; 2298 2299 error = SYSCTL_OUT(req, &xs, sizeof(xs)); 2300 return (error); 2301 } 2302 n++; 2303 } 2304 mtx_unlock(&sw_dev_mtx); 2305 return (ENOENT); 2306 } 2307 2308 SYSCTL_INT(_vm, OID_AUTO, nswapdev, CTLFLAG_RD, &nswapdev, 0, 2309 "Number of swap devices"); 2310 SYSCTL_NODE(_vm, OID_AUTO, swap_info, CTLFLAG_RD, sysctl_vm_swap_info, 2311 "Swap statistics by device"); 2312 2313 /* 2314 * vmspace_swap_count() - count the approximate swap useage in pages for a 2315 * vmspace. 2316 * 2317 * The map must be locked. 2318 * 2319 * Swap useage is determined by taking the proportional swap used by 2320 * VM objects backing the VM map. To make up for fractional losses, 2321 * if the VM object has any swap use at all the associated map entries 2322 * count for at least 1 swap page. 2323 */ 2324 int 2325 vmspace_swap_count(struct vmspace *vmspace) 2326 { 2327 vm_map_t map = &vmspace->vm_map; 2328 vm_map_entry_t cur; 2329 int count = 0; 2330 2331 for (cur = map->header.next; cur != &map->header; cur = cur->next) { 2332 vm_object_t object; 2333 2334 if ((cur->eflags & MAP_ENTRY_IS_SUB_MAP) == 0 && 2335 (object = cur->object.vm_object) != NULL) { 2336 VM_OBJECT_LOCK(object); 2337 if (object->type == OBJT_SWAP && 2338 object->un_pager.swp.swp_bcount != 0) { 2339 int n = (cur->end - cur->start) / PAGE_SIZE; 2340 2341 count += object->un_pager.swp.swp_bcount * 2342 SWAP_META_PAGES * n / object->size + 1; 2343 } 2344 VM_OBJECT_UNLOCK(object); 2345 } 2346 } 2347 return (count); 2348 } 2349 2350 /* 2351 * GEOM backend 2352 * 2353 * Swapping onto disk devices. 2354 * 2355 */ 2356 2357 static struct g_class g_swap_class = { 2358 .name = "SWAP", 2359 }; 2360 2361 DECLARE_GEOM_CLASS(g_swap_class, g_class); 2362 2363 2364 static void 2365 swapgeom_done(struct bio *bp2) 2366 { 2367 struct buf *bp; 2368 2369 bp = bp2->bio_caller2; 2370 if (bp2->bio_error) 2371 bp->b_ioflags |= BIO_ERROR; 2372 bufdone(bp); 2373 g_destroy_bio(bp2); 2374 } 2375 2376 static void 2377 swapgeom_strategy(struct buf *bp, struct swdevt *sp) 2378 { 2379 struct bio *bio; 2380 struct g_consumer *cp; 2381 2382 cp = sp->sw_id; 2383 if (cp == NULL) { 2384 bp->b_error = ENXIO; 2385 bp->b_ioflags |= BIO_ERROR; 2386 bufdone(bp); 2387 return; 2388 } 2389 bio = g_clone_bio(&bp->b_io); 2390 if (bio == NULL) { 2391 /* 2392 * XXX: This is better than panicing, but not much better. 2393 * XXX: Somehow this should be retried. A more generic 2394 * XXX: implementation of ENOMEM in geom may be able to cope. 2395 */ 2396 bp->b_error = ENOMEM; 2397 bp->b_ioflags |= BIO_ERROR; 2398 bufdone(bp); 2399 return; 2400 } 2401 bio->bio_caller2 = bp; 2402 bio->bio_offset = (bp->b_blkno - sp->sw_first) * PAGE_SIZE; 2403 bio->bio_length = bp->b_bcount; 2404 bio->bio_done = swapgeom_done; 2405 g_io_request(bio, cp); 2406 return; 2407 } 2408 2409 static void 2410 swapgeom_orphan(struct g_consumer *cp) 2411 { 2412 struct swdevt *sp; 2413 2414 mtx_lock(&sw_dev_mtx); 2415 TAILQ_FOREACH(sp, &swtailq, sw_list) 2416 if (sp->sw_id == cp) 2417 sp->sw_id = NULL; 2418 mtx_unlock(&sw_dev_mtx); 2419 } 2420 2421 static void 2422 swapgeom_close_ev(void *arg, int flags) 2423 { 2424 struct g_consumer *cp; 2425 2426 cp = arg; 2427 g_access(cp, -1, -1, 0); 2428 g_detach(cp); 2429 g_destroy_consumer(cp); 2430 } 2431 2432 static void 2433 swapgeom_close(struct thread *td, struct swdevt *sw) 2434 { 2435 2436 /* XXX: direct call when Giant untangled */ 2437 g_waitfor_event(swapgeom_close_ev, sw->sw_id, M_WAITOK, NULL); 2438 } 2439 2440 2441 struct swh0h0 { 2442 struct cdev *dev; 2443 struct vnode *vp; 2444 int error; 2445 }; 2446 2447 static void 2448 swapongeom_ev(void *arg, int flags) 2449 { 2450 struct swh0h0 *swh; 2451 struct g_provider *pp; 2452 struct g_consumer *cp; 2453 static struct g_geom *gp; 2454 struct swdevt *sp; 2455 u_long nblks; 2456 int error; 2457 2458 swh = arg; 2459 swh->error = 0; 2460 pp = g_dev_getprovider(swh->dev); 2461 if (pp == NULL) { 2462 swh->error = ENODEV; 2463 return; 2464 } 2465 mtx_lock(&sw_dev_mtx); 2466 TAILQ_FOREACH(sp, &swtailq, sw_list) { 2467 cp = sp->sw_id; 2468 if (cp != NULL && cp->provider == pp) { 2469 mtx_unlock(&sw_dev_mtx); 2470 swh->error = EBUSY; 2471 return; 2472 } 2473 } 2474 mtx_unlock(&sw_dev_mtx); 2475 if (gp == NULL) { 2476 gp = g_new_geomf(&g_swap_class, "swap", NULL); 2477 gp->orphan = swapgeom_orphan; 2478 } 2479 cp = g_new_consumer(gp); 2480 g_attach(cp, pp); 2481 /* 2482 * XXX: Everytime you think you can improve the margin for 2483 * footshooting, somebody depends on the ability to do so: 2484 * savecore(8) wants to write to our swapdev so we cannot 2485 * set an exclusive count :-( 2486 */ 2487 error = g_access(cp, 1, 1, 0); 2488 if (error) { 2489 g_detach(cp); 2490 g_destroy_consumer(cp); 2491 swh->error = error; 2492 return; 2493 } 2494 nblks = pp->mediasize / DEV_BSIZE; 2495 swaponsomething(swh->vp, cp, nblks, swapgeom_strategy, 2496 swapgeom_close, dev2udev(swh->dev)); 2497 swh->error = 0; 2498 return; 2499 } 2500 2501 static int 2502 swapongeom(struct thread *td, struct vnode *vp) 2503 { 2504 int error; 2505 struct swh0h0 swh; 2506 2507 vn_lock(vp, LK_EXCLUSIVE | LK_RETRY, td); 2508 2509 swh.dev = vp->v_rdev; 2510 swh.vp = vp; 2511 swh.error = 0; 2512 /* XXX: direct call when Giant untangled */ 2513 error = g_waitfor_event(swapongeom_ev, &swh, M_WAITOK, NULL); 2514 if (!error) 2515 error = swh.error; 2516 VOP_UNLOCK(vp, 0, td); 2517 return (error); 2518 } 2519 2520 /* 2521 * VNODE backend 2522 * 2523 * This is used mainly for network filesystem (read: probably only tested 2524 * with NFS) swapfiles. 2525 * 2526 */ 2527 2528 static void 2529 swapdev_strategy(struct buf *bp, struct swdevt *sp) 2530 { 2531 int s; 2532 struct vnode *vp, *vp2; 2533 2534 bp->b_dev = NULL; 2535 bp->b_blkno = ctodb(bp->b_blkno - sp->sw_first); 2536 2537 vp2 = sp->sw_id; 2538 vhold(vp2); 2539 s = splvm(); 2540 if (bp->b_iocmd == BIO_WRITE) { 2541 vp = bp->b_vp; 2542 if (vp) { 2543 VI_LOCK(vp); 2544 vp->v_numoutput--; 2545 if ((vp->v_iflag & VI_BWAIT) && vp->v_numoutput <= 0) { 2546 vp->v_iflag &= ~VI_BWAIT; 2547 wakeup(&vp->v_numoutput); 2548 } 2549 VI_UNLOCK(vp); 2550 } 2551 VI_LOCK(vp2); 2552 vp2->v_numoutput++; 2553 VI_UNLOCK(vp2); 2554 } 2555 bp->b_vp = vp2; 2556 splx(s); 2557 bp->b_iooffset = dbtob(bp->b_blkno); 2558 VOP_STRATEGY(vp2, bp); 2559 return; 2560 } 2561 2562 static void 2563 swapdev_close(struct thread *td, struct swdevt *sp) 2564 { 2565 2566 VOP_CLOSE(sp->sw_vp, FREAD | FWRITE, td->td_ucred, td); 2567 vrele(sp->sw_vp); 2568 } 2569 2570 2571 static int 2572 swaponvp(struct thread *td, struct vnode *vp, u_long nblks) 2573 { 2574 struct swdevt *sp; 2575 int error; 2576 2577 if (nblks == 0) 2578 return (ENXIO); 2579 mtx_lock(&sw_dev_mtx); 2580 TAILQ_FOREACH(sp, &swtailq, sw_list) { 2581 if (sp->sw_id == vp) { 2582 mtx_unlock(&sw_dev_mtx); 2583 return (EBUSY); 2584 } 2585 } 2586 mtx_unlock(&sw_dev_mtx); 2587 2588 (void) vn_lock(vp, LK_EXCLUSIVE | LK_RETRY, td); 2589 #ifdef MAC 2590 error = mac_check_system_swapon(td->td_ucred, vp); 2591 if (error == 0) 2592 #endif 2593 error = VOP_OPEN(vp, FREAD | FWRITE, td->td_ucred, td, -1); 2594 (void) VOP_UNLOCK(vp, 0, td); 2595 if (error) 2596 return (error); 2597 2598 swaponsomething(vp, vp, nblks, swapdev_strategy, swapdev_close, 2599 NODEV); 2600 return (0); 2601 } 2602