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