1 /* 2 * (MPSAFE) 3 * 4 * Copyright (c) 1998-2010 The DragonFly Project. All rights reserved. 5 * 6 * This code is derived from software contributed to The DragonFly Project 7 * by Matthew Dillon <dillon@backplane.com> 8 * 9 * Redistribution and use in source and binary forms, with or without 10 * modification, are permitted provided that the following conditions 11 * are met: 12 * 13 * 1. Redistributions of source code must retain the above copyright 14 * notice, this list of conditions and the following disclaimer. 15 * 2. Redistributions in binary form must reproduce the above copyright 16 * notice, this list of conditions and the following disclaimer in 17 * the documentation and/or other materials provided with the 18 * distribution. 19 * 3. Neither the name of The DragonFly Project nor the names of its 20 * contributors may be used to endorse or promote products derived 21 * from this software without specific, prior written permission. 22 * 23 * THIS SOFTWARE IS PROVIDED BY THE COPYRIGHT HOLDERS AND CONTRIBUTORS 24 * ``AS IS'' AND ANY EXPRESS OR IMPLIED WARRANTIES, INCLUDING, BUT NOT 25 * LIMITED TO, THE IMPLIED WARRANTIES OF MERCHANTABILITY AND FITNESS 26 * FOR A PARTICULAR PURPOSE ARE DISCLAIMED. IN NO EVENT SHALL THE 27 * COPYRIGHT HOLDERS OR CONTRIBUTORS BE LIABLE FOR ANY DIRECT, INDIRECT, 28 * INCIDENTAL, SPECIAL, EXEMPLARY OR CONSEQUENTIAL DAMAGES (INCLUDING, 29 * BUT NOT LIMITED TO, PROCUREMENT OF SUBSTITUTE GOODS OR SERVICES; 30 * LOSS OF USE, DATA, OR PROFITS; OR BUSINESS INTERRUPTION) HOWEVER CAUSED 31 * AND ON ANY THEORY OF LIABILITY, WHETHER IN CONTRACT, STRICT LIABILITY, 32 * OR TORT (INCLUDING NEGLIGENCE OR OTHERWISE) ARISING IN ANY WAY OUT 33 * OF THE USE OF THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF 34 * SUCH DAMAGE. 35 * 36 * Copyright (c) 1994 John S. Dyson 37 * Copyright (c) 1990 University of Utah. 38 * Copyright (c) 1991, 1993 39 * The Regents of the University of California. All rights reserved. 40 * 41 * This code is derived from software contributed to Berkeley by 42 * the Systems Programming Group of the University of Utah Computer 43 * Science Department. 44 * 45 * Redistribution and use in source and binary forms, with or without 46 * modification, are permitted provided that the following conditions 47 * are met: 48 * 1. Redistributions of source code must retain the above copyright 49 * notice, this list of conditions and the following disclaimer. 50 * 2. Redistributions in binary form must reproduce the above copyright 51 * notice, this list of conditions and the following disclaimer in the 52 * documentation and/or other materials provided with the distribution. 53 * 3. All advertising materials mentioning features or use of this software 54 * must display the following acknowledgement: 55 * This product includes software developed by the University of 56 * California, Berkeley and its contributors. 57 * 4. Neither the name of the University nor the names of its contributors 58 * may be used to endorse or promote products derived from this software 59 * without specific prior written permission. 60 * 61 * THIS SOFTWARE IS PROVIDED BY THE REGENTS AND CONTRIBUTORS ``AS IS'' AND 62 * ANY EXPRESS OR IMPLIED WARRANTIES, INCLUDING, BUT NOT LIMITED TO, THE 63 * IMPLIED WARRANTIES OF MERCHANTABILITY AND FITNESS FOR A PARTICULAR PURPOSE 64 * ARE DISCLAIMED. IN NO EVENT SHALL THE REGENTS OR CONTRIBUTORS BE LIABLE 65 * FOR ANY DIRECT, INDIRECT, INCIDENTAL, SPECIAL, EXEMPLARY, OR CONSEQUENTIAL 66 * DAMAGES (INCLUDING, BUT NOT LIMITED TO, PROCUREMENT OF SUBSTITUTE GOODS 67 * OR SERVICES; LOSS OF USE, DATA, OR PROFITS; OR BUSINESS INTERRUPTION) 68 * HOWEVER CAUSED AND ON ANY THEORY OF LIABILITY, WHETHER IN CONTRACT, STRICT 69 * LIABILITY, OR TORT (INCLUDING NEGLIGENCE OR OTHERWISE) ARISING IN ANY WAY 70 * OUT OF THE USE OF THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF 71 * SUCH DAMAGE. 72 * 73 * New Swap System 74 * Matthew Dillon 75 * 76 * Radix Bitmap 'blists'. 77 * 78 * - The new swapper uses the new radix bitmap code. This should scale 79 * to arbitrarily small or arbitrarily large swap spaces and an almost 80 * arbitrary degree of fragmentation. 81 * 82 * Features: 83 * 84 * - on the fly reallocation of swap during putpages. The new system 85 * does not try to keep previously allocated swap blocks for dirty 86 * pages. 87 * 88 * - on the fly deallocation of swap 89 * 90 * - No more garbage collection required. Unnecessarily allocated swap 91 * blocks only exist for dirty vm_page_t's now and these are already 92 * cycled (in a high-load system) by the pager. We also do on-the-fly 93 * removal of invalidated swap blocks when a page is destroyed 94 * or renamed. 95 * 96 * from: Utah $Hdr: swap_pager.c 1.4 91/04/30$ 97 * @(#)swap_pager.c 8.9 (Berkeley) 3/21/94 98 * $FreeBSD: src/sys/vm/swap_pager.c,v 1.130.2.12 2002/08/31 21:15:55 dillon Exp $ 99 */ 100 101 #include <sys/param.h> 102 #include <sys/systm.h> 103 #include <sys/conf.h> 104 #include <sys/kernel.h> 105 #include <sys/proc.h> 106 #include <sys/buf.h> 107 #include <sys/vnode.h> 108 #include <sys/malloc.h> 109 #include <sys/vmmeter.h> 110 #include <sys/sysctl.h> 111 #include <sys/blist.h> 112 #include <sys/lock.h> 113 #include <sys/thread2.h> 114 115 #ifndef MAX_PAGEOUT_CLUSTER 116 #define MAX_PAGEOUT_CLUSTER 16 117 #endif 118 119 #define SWB_NPAGES MAX_PAGEOUT_CLUSTER 120 121 #include "opt_swap.h" 122 #include <vm/vm.h> 123 #include <vm/vm_object.h> 124 #include <vm/vm_page.h> 125 #include <vm/vm_pager.h> 126 #include <vm/vm_pageout.h> 127 #include <vm/swap_pager.h> 128 #include <vm/vm_extern.h> 129 #include <vm/vm_zone.h> 130 #include <vm/vnode_pager.h> 131 132 #include <sys/buf2.h> 133 #include <vm/vm_page2.h> 134 135 #define SWM_FREE 0x02 /* free, period */ 136 #define SWM_POP 0x04 /* pop out */ 137 138 #define SWBIO_READ 0x01 139 #define SWBIO_WRITE 0x02 140 #define SWBIO_SYNC 0x04 141 142 struct swfreeinfo { 143 vm_object_t object; 144 vm_pindex_t basei; 145 vm_pindex_t begi; 146 vm_pindex_t endi; /* inclusive */ 147 }; 148 149 /* 150 * vm_swap_size is in page-sized chunks now. It was DEV_BSIZE'd chunks 151 * in the old system. 152 */ 153 154 int swap_pager_full; /* swap space exhaustion (task killing) */ 155 int vm_swap_cache_use; 156 int vm_swap_anon_use; 157 158 static int swap_pager_almost_full; /* swap space exhaustion (w/ hysteresis)*/ 159 static int nsw_rcount; /* free read buffers */ 160 static int nsw_wcount_sync; /* limit write buffers / synchronous */ 161 static int nsw_wcount_async; /* limit write buffers / asynchronous */ 162 static int nsw_wcount_async_max;/* assigned maximum */ 163 static int nsw_cluster_max; /* maximum VOP I/O allowed */ 164 165 struct blist *swapblist; 166 static int swap_async_max = 4; /* maximum in-progress async I/O's */ 167 static int swap_burst_read = 0; /* allow burst reading */ 168 169 /* from vm_swap.c */ 170 extern struct vnode *swapdev_vp; 171 extern struct swdevt *swdevt; 172 extern int nswdev; 173 174 #define BLK2DEVIDX(blk) (nswdev > 1 ? blk / dmmax % nswdev : 0) 175 176 SYSCTL_INT(_vm, OID_AUTO, swap_async_max, 177 CTLFLAG_RW, &swap_async_max, 0, "Maximum running async swap ops"); 178 SYSCTL_INT(_vm, OID_AUTO, swap_burst_read, 179 CTLFLAG_RW, &swap_burst_read, 0, "Allow burst reads for pageins"); 180 181 SYSCTL_INT(_vm, OID_AUTO, swap_cache_use, 182 CTLFLAG_RD, &vm_swap_cache_use, 0, ""); 183 SYSCTL_INT(_vm, OID_AUTO, swap_anon_use, 184 CTLFLAG_RD, &vm_swap_anon_use, 0, ""); 185 SYSCTL_INT(_vm, OID_AUTO, swap_size, 186 CTLFLAG_RD, &vm_swap_size, 0, ""); 187 188 vm_zone_t swap_zone; 189 190 /* 191 * Red-Black tree for swblock entries 192 * 193 * The caller must hold vm_token 194 */ 195 RB_GENERATE2(swblock_rb_tree, swblock, swb_entry, rb_swblock_compare, 196 vm_pindex_t, swb_index); 197 198 int 199 rb_swblock_compare(struct swblock *swb1, struct swblock *swb2) 200 { 201 if (swb1->swb_index < swb2->swb_index) 202 return(-1); 203 if (swb1->swb_index > swb2->swb_index) 204 return(1); 205 return(0); 206 } 207 208 static 209 int 210 rb_swblock_scancmp(struct swblock *swb, void *data) 211 { 212 struct swfreeinfo *info = data; 213 214 if (swb->swb_index < info->basei) 215 return(-1); 216 if (swb->swb_index > info->endi) 217 return(1); 218 return(0); 219 } 220 221 static 222 int 223 rb_swblock_condcmp(struct swblock *swb, void *data) 224 { 225 struct swfreeinfo *info = data; 226 227 if (swb->swb_index < info->basei) 228 return(-1); 229 return(0); 230 } 231 232 /* 233 * pagerops for OBJT_SWAP - "swap pager". Some ops are also global procedure 234 * calls hooked from other parts of the VM system and do not appear here. 235 * (see vm/swap_pager.h). 236 */ 237 238 static void swap_pager_dealloc (vm_object_t object); 239 static int swap_pager_getpage (vm_object_t, vm_page_t *, int); 240 static void swap_chain_iodone(struct bio *biox); 241 242 struct pagerops swappagerops = { 243 swap_pager_dealloc, /* deallocate an OBJT_SWAP object */ 244 swap_pager_getpage, /* pagein */ 245 swap_pager_putpages, /* pageout */ 246 swap_pager_haspage /* get backing store status for page */ 247 }; 248 249 /* 250 * dmmax is in page-sized chunks with the new swap system. It was 251 * dev-bsized chunks in the old. dmmax is always a power of 2. 252 * 253 * swap_*() routines are externally accessible. swp_*() routines are 254 * internal. 255 */ 256 257 int dmmax; 258 static int dmmax_mask; 259 int nswap_lowat = 128; /* in pages, swap_pager_almost_full warn */ 260 int nswap_hiwat = 512; /* in pages, swap_pager_almost_full warn */ 261 262 static __inline void swp_sizecheck (void); 263 static void swp_pager_async_iodone (struct bio *bio); 264 265 /* 266 * Swap bitmap functions 267 */ 268 269 static __inline void swp_pager_freeswapspace(vm_object_t object, 270 swblk_t blk, int npages); 271 static __inline swblk_t swp_pager_getswapspace(vm_object_t object, int npages); 272 273 /* 274 * Metadata functions 275 */ 276 277 static void swp_pager_meta_convert(vm_object_t); 278 static void swp_pager_meta_build(vm_object_t, vm_pindex_t, swblk_t); 279 static void swp_pager_meta_free(vm_object_t, vm_pindex_t, vm_pindex_t); 280 static void swp_pager_meta_free_all(vm_object_t); 281 static swblk_t swp_pager_meta_ctl(vm_object_t, vm_pindex_t, int); 282 283 /* 284 * SWP_SIZECHECK() - update swap_pager_full indication 285 * 286 * update the swap_pager_almost_full indication and warn when we are 287 * about to run out of swap space, using lowat/hiwat hysteresis. 288 * 289 * Clear swap_pager_full ( task killing ) indication when lowat is met. 290 * 291 * No restrictions on call 292 * This routine may not block. 293 * SMP races are ok. 294 */ 295 static __inline void 296 swp_sizecheck(void) 297 { 298 if (vm_swap_size < nswap_lowat) { 299 if (swap_pager_almost_full == 0) { 300 kprintf("swap_pager: out of swap space\n"); 301 swap_pager_almost_full = 1; 302 } 303 } else { 304 swap_pager_full = 0; 305 if (vm_swap_size > nswap_hiwat) 306 swap_pager_almost_full = 0; 307 } 308 } 309 310 /* 311 * SWAP_PAGER_INIT() - initialize the swap pager! 312 * 313 * Expected to be started from system init. NOTE: This code is run 314 * before much else so be careful what you depend on. Most of the VM 315 * system has yet to be initialized at this point. 316 * 317 * Called from the low level boot code only. 318 */ 319 static void 320 swap_pager_init(void *arg __unused) 321 { 322 /* 323 * Device Stripe, in PAGE_SIZE'd blocks 324 */ 325 dmmax = SWB_NPAGES * 2; 326 dmmax_mask = ~(dmmax - 1); 327 } 328 SYSINIT(vm_mem, SI_BOOT1_VM, SI_ORDER_THIRD, swap_pager_init, NULL) 329 330 /* 331 * SWAP_PAGER_SWAP_INIT() - swap pager initialization from pageout process 332 * 333 * Expected to be started from pageout process once, prior to entering 334 * its main loop. 335 * 336 * Called from the low level boot code only. 337 */ 338 void 339 swap_pager_swap_init(void) 340 { 341 int n, n2; 342 343 /* 344 * Number of in-transit swap bp operations. Don't 345 * exhaust the pbufs completely. Make sure we 346 * initialize workable values (0 will work for hysteresis 347 * but it isn't very efficient). 348 * 349 * The nsw_cluster_max is constrained by the number of pages an XIO 350 * holds, i.e., (MAXPHYS/PAGE_SIZE) and our locally defined 351 * MAX_PAGEOUT_CLUSTER. Also be aware that swap ops are 352 * constrained by the swap device interleave stripe size. 353 * 354 * Currently we hardwire nsw_wcount_async to 4. This limit is 355 * designed to prevent other I/O from having high latencies due to 356 * our pageout I/O. The value 4 works well for one or two active swap 357 * devices but is probably a little low if you have more. Even so, 358 * a higher value would probably generate only a limited improvement 359 * with three or four active swap devices since the system does not 360 * typically have to pageout at extreme bandwidths. We will want 361 * at least 2 per swap devices, and 4 is a pretty good value if you 362 * have one NFS swap device due to the command/ack latency over NFS. 363 * So it all works out pretty well. 364 */ 365 366 nsw_cluster_max = min((MAXPHYS/PAGE_SIZE), MAX_PAGEOUT_CLUSTER); 367 368 nsw_rcount = (nswbuf + 1) / 2; 369 nsw_wcount_sync = (nswbuf + 3) / 4; 370 nsw_wcount_async = 4; 371 nsw_wcount_async_max = nsw_wcount_async; 372 373 /* 374 * The zone is dynamically allocated so generally size it to 375 * maxswzone (32MB to 512MB of KVM). Set a minimum size based 376 * on physical memory of around 8x (each swblock can hold 16 pages). 377 * 378 * With the advent of SSDs (vs HDs) the practical (swap:memory) ratio 379 * has increased dramatically. 380 */ 381 n = vmstats.v_page_count / 2; 382 if (maxswzone && n < maxswzone / sizeof(struct swblock)) 383 n = maxswzone / sizeof(struct swblock); 384 n2 = n; 385 386 do { 387 swap_zone = zinit( 388 "SWAPMETA", 389 sizeof(struct swblock), 390 n, 391 ZONE_INTERRUPT, 392 1); 393 if (swap_zone != NULL) 394 break; 395 /* 396 * if the allocation failed, try a zone two thirds the 397 * size of the previous attempt. 398 */ 399 n -= ((n + 2) / 3); 400 } while (n > 0); 401 402 if (swap_zone == NULL) 403 panic("swap_pager_swap_init: swap_zone == NULL"); 404 if (n2 != n) 405 kprintf("Swap zone entries reduced from %d to %d.\n", n2, n); 406 } 407 408 /* 409 * SWAP_PAGER_ALLOC() - allocate a new OBJT_SWAP VM object and instantiate 410 * its metadata structures. 411 * 412 * This routine is called from the mmap and fork code to create a new 413 * OBJT_SWAP object. We do this by creating an OBJT_DEFAULT object 414 * and then converting it with swp_pager_meta_convert(). 415 * 416 * We only support unnamed objects. 417 * 418 * No restrictions. 419 */ 420 vm_object_t 421 swap_pager_alloc(void *handle, off_t size, vm_prot_t prot, off_t offset) 422 { 423 vm_object_t object; 424 425 KKASSERT(handle == NULL); 426 lwkt_gettoken(&vm_token); 427 object = vm_object_allocate(OBJT_DEFAULT, 428 OFF_TO_IDX(offset + PAGE_MASK + size)); 429 swp_pager_meta_convert(object); 430 lwkt_reltoken(&vm_token); 431 432 return (object); 433 } 434 435 /* 436 * SWAP_PAGER_DEALLOC() - remove swap metadata from object 437 * 438 * The swap backing for the object is destroyed. The code is 439 * designed such that we can reinstantiate it later, but this 440 * routine is typically called only when the entire object is 441 * about to be destroyed. 442 * 443 * The object must be locked or unreferenceable. 444 * No other requirements. 445 */ 446 static void 447 swap_pager_dealloc(vm_object_t object) 448 { 449 lwkt_gettoken(&vm_token); 450 vm_object_pip_wait(object, "swpdea"); 451 452 /* 453 * Free all remaining metadata. We only bother to free it from 454 * the swap meta data. We do not attempt to free swapblk's still 455 * associated with vm_page_t's for this object. We do not care 456 * if paging is still in progress on some objects. 457 */ 458 crit_enter(); 459 swp_pager_meta_free_all(object); 460 crit_exit(); 461 lwkt_reltoken(&vm_token); 462 } 463 464 /************************************************************************ 465 * SWAP PAGER BITMAP ROUTINES * 466 ************************************************************************/ 467 468 /* 469 * SWP_PAGER_GETSWAPSPACE() - allocate raw swap space 470 * 471 * Allocate swap for the requested number of pages. The starting 472 * swap block number (a page index) is returned or SWAPBLK_NONE 473 * if the allocation failed. 474 * 475 * Also has the side effect of advising that somebody made a mistake 476 * when they configured swap and didn't configure enough. 477 * 478 * The caller must hold vm_token. 479 * This routine may not block. 480 * 481 * NOTE: vm_token must be held to avoid races with bitmap frees from 482 * vm_page_remove() via swap_pager_page_removed(). 483 */ 484 static __inline swblk_t 485 swp_pager_getswapspace(vm_object_t object, int npages) 486 { 487 swblk_t blk; 488 489 ASSERT_LWKT_TOKEN_HELD(&vm_token); 490 491 if ((blk = blist_alloc(swapblist, npages)) == SWAPBLK_NONE) { 492 if (swap_pager_full != 2) { 493 kprintf("swap_pager_getswapspace: failed\n"); 494 swap_pager_full = 2; 495 swap_pager_almost_full = 1; 496 } 497 } else { 498 swapacctspace(blk, -npages); 499 if (object->type == OBJT_SWAP) 500 vm_swap_anon_use += npages; 501 else 502 vm_swap_cache_use += npages; 503 swp_sizecheck(); 504 } 505 return(blk); 506 } 507 508 /* 509 * SWP_PAGER_FREESWAPSPACE() - free raw swap space 510 * 511 * This routine returns the specified swap blocks back to the bitmap. 512 * 513 * Note: This routine may not block (it could in the old swap code), 514 * and through the use of the new blist routines it does not block. 515 * 516 * We must be called at splvm() to avoid races with bitmap frees from 517 * vm_page_remove() aka swap_pager_page_removed(). 518 * 519 * The caller must hold vm_token. 520 * This routine may not block. 521 */ 522 523 static __inline void 524 swp_pager_freeswapspace(vm_object_t object, swblk_t blk, int npages) 525 { 526 struct swdevt *sp = &swdevt[BLK2DEVIDX(blk)]; 527 528 sp->sw_nused -= npages; 529 if (object->type == OBJT_SWAP) 530 vm_swap_anon_use -= npages; 531 else 532 vm_swap_cache_use -= npages; 533 534 if (sp->sw_flags & SW_CLOSING) 535 return; 536 537 blist_free(swapblist, blk, npages); 538 vm_swap_size += npages; 539 swp_sizecheck(); 540 } 541 542 /* 543 * SWAP_PAGER_FREESPACE() - frees swap blocks associated with a page 544 * range within an object. 545 * 546 * This is a globally accessible routine. 547 * 548 * This routine removes swapblk assignments from swap metadata. 549 * 550 * The external callers of this routine typically have already destroyed 551 * or renamed vm_page_t's associated with this range in the object so 552 * we should be ok. 553 * 554 * No requirements. 555 */ 556 void 557 swap_pager_freespace(vm_object_t object, vm_pindex_t start, vm_pindex_t size) 558 { 559 crit_enter(); 560 lwkt_gettoken(&vm_token); 561 swp_pager_meta_free(object, start, size); 562 lwkt_reltoken(&vm_token); 563 crit_exit(); 564 } 565 566 /* 567 * No requirements. 568 */ 569 void 570 swap_pager_freespace_all(vm_object_t object) 571 { 572 crit_enter(); 573 lwkt_gettoken(&vm_token); 574 swp_pager_meta_free_all(object); 575 lwkt_reltoken(&vm_token); 576 crit_exit(); 577 } 578 579 /* 580 * This function conditionally frees swap cache swap starting at 581 * (*basei) in the object. (count) swap blocks will be nominally freed. 582 * The actual number of blocks freed can be more or less than the 583 * requested number. 584 * 585 * This function nominally returns the number of blocks freed. However, 586 * the actual number of blocks freed may be less then the returned value. 587 * If the function is unable to exhaust the object or if it is able to 588 * free (approximately) the requested number of blocks it returns 589 * a value n > count. 590 * 591 * If we exhaust the object we will return a value n <= count. 592 * 593 * The caller must hold vm_token. 594 */ 595 static int swap_pager_condfree_callback(struct swblock *swap, void *data); 596 597 int 598 swap_pager_condfree(vm_object_t object, vm_pindex_t *basei, int count) 599 { 600 struct swfreeinfo info; 601 602 ASSERT_LWKT_TOKEN_HELD(&vm_token); 603 604 info.object = object; 605 info.basei = *basei; /* skip up to this page index */ 606 info.begi = count; /* max swap pages to destroy */ 607 info.endi = count * 8; /* max swblocks to scan */ 608 609 swblock_rb_tree_RB_SCAN(&object->swblock_root, rb_swblock_condcmp, 610 swap_pager_condfree_callback, &info); 611 *basei = info.basei; 612 if (info.endi < 0 && info.begi <= count) 613 info.begi = count + 1; 614 return(count - (int)info.begi); 615 } 616 617 /* 618 * The idea is to free whole meta-block to avoid fragmenting 619 * the swap space or disk I/O. We only do this if NO VM pages 620 * are present. 621 * 622 * We do not have to deal with clearing PG_SWAPPED in related VM 623 * pages because there are no related VM pages. 624 * 625 * The caller must hold vm_token. 626 */ 627 static int 628 swap_pager_condfree_callback(struct swblock *swap, void *data) 629 { 630 struct swfreeinfo *info = data; 631 vm_object_t object = info->object; 632 int i; 633 634 for (i = 0; i < SWAP_META_PAGES; ++i) { 635 if (vm_page_lookup(object, swap->swb_index + i)) 636 break; 637 } 638 info->basei = swap->swb_index + SWAP_META_PAGES; 639 if (i == SWAP_META_PAGES) { 640 info->begi -= swap->swb_count; 641 swap_pager_freespace(object, swap->swb_index, SWAP_META_PAGES); 642 } 643 --info->endi; 644 if ((int)info->begi < 0 || (int)info->endi < 0) 645 return(-1); 646 return(0); 647 } 648 649 /* 650 * Called by vm_page_alloc() when a new VM page is inserted 651 * into a VM object. Checks whether swap has been assigned to 652 * the page and sets PG_SWAPPED as necessary. 653 * 654 * No requirements. 655 */ 656 void 657 swap_pager_page_inserted(vm_page_t m) 658 { 659 if (m->object->swblock_count) { 660 crit_enter(); 661 lwkt_gettoken(&vm_token); 662 if (swp_pager_meta_ctl(m->object, m->pindex, 0) != SWAPBLK_NONE) 663 vm_page_flag_set(m, PG_SWAPPED); 664 lwkt_reltoken(&vm_token); 665 crit_exit(); 666 } 667 } 668 669 /* 670 * SWAP_PAGER_RESERVE() - reserve swap blocks in object 671 * 672 * Assigns swap blocks to the specified range within the object. The 673 * swap blocks are not zerod. Any previous swap assignment is destroyed. 674 * 675 * Returns 0 on success, -1 on failure. 676 * 677 * The caller is responsible for avoiding races in the specified range. 678 * No other requirements. 679 */ 680 int 681 swap_pager_reserve(vm_object_t object, vm_pindex_t start, vm_size_t size) 682 { 683 int n = 0; 684 swblk_t blk = SWAPBLK_NONE; 685 vm_pindex_t beg = start; /* save start index */ 686 687 crit_enter(); 688 lwkt_gettoken(&vm_token); 689 while (size) { 690 if (n == 0) { 691 n = BLIST_MAX_ALLOC; 692 while ((blk = swp_pager_getswapspace(object, n)) == 693 SWAPBLK_NONE) 694 { 695 n >>= 1; 696 if (n == 0) { 697 swp_pager_meta_free(object, beg, 698 start - beg); 699 lwkt_reltoken(&vm_token); 700 crit_exit(); 701 return(-1); 702 } 703 } 704 } 705 swp_pager_meta_build(object, start, blk); 706 --size; 707 ++start; 708 ++blk; 709 --n; 710 } 711 swp_pager_meta_free(object, start, n); 712 lwkt_reltoken(&vm_token); 713 crit_exit(); 714 return(0); 715 } 716 717 /* 718 * SWAP_PAGER_COPY() - copy blocks from source pager to destination pager 719 * and destroy the source. 720 * 721 * Copy any valid swapblks from the source to the destination. In 722 * cases where both the source and destination have a valid swapblk, 723 * we keep the destination's. 724 * 725 * This routine is allowed to block. It may block allocating metadata 726 * indirectly through swp_pager_meta_build() or if paging is still in 727 * progress on the source. 728 * 729 * This routine can be called at any spl 730 * 731 * XXX vm_page_collapse() kinda expects us not to block because we 732 * supposedly do not need to allocate memory, but for the moment we 733 * *may* have to get a little memory from the zone allocator, but 734 * it is taken from the interrupt memory. We should be ok. 735 * 736 * The source object contains no vm_page_t's (which is just as well) 737 * 738 * The source object is of type OBJT_SWAP. 739 * 740 * The source and destination objects must be locked or 741 * inaccessible (XXX are they ?) 742 * 743 * The caller must hold vm_token. 744 */ 745 void 746 swap_pager_copy(vm_object_t srcobject, vm_object_t dstobject, 747 vm_pindex_t base_index, int destroysource) 748 { 749 vm_pindex_t i; 750 751 ASSERT_LWKT_TOKEN_HELD(&vm_token); 752 crit_enter(); 753 754 /* 755 * transfer source to destination. 756 */ 757 for (i = 0; i < dstobject->size; ++i) { 758 swblk_t dstaddr; 759 760 /* 761 * Locate (without changing) the swapblk on the destination, 762 * unless it is invalid in which case free it silently, or 763 * if the destination is a resident page, in which case the 764 * source is thrown away. 765 */ 766 dstaddr = swp_pager_meta_ctl(dstobject, i, 0); 767 768 if (dstaddr == SWAPBLK_NONE) { 769 /* 770 * Destination has no swapblk and is not resident, 771 * copy source. 772 */ 773 swblk_t srcaddr; 774 775 srcaddr = swp_pager_meta_ctl(srcobject, 776 base_index + i, SWM_POP); 777 778 if (srcaddr != SWAPBLK_NONE) 779 swp_pager_meta_build(dstobject, i, srcaddr); 780 } else { 781 /* 782 * Destination has valid swapblk or it is represented 783 * by a resident page. We destroy the sourceblock. 784 */ 785 swp_pager_meta_ctl(srcobject, base_index + i, SWM_FREE); 786 } 787 } 788 789 /* 790 * Free left over swap blocks in source. 791 * 792 * We have to revert the type to OBJT_DEFAULT so we do not accidently 793 * double-remove the object from the swap queues. 794 */ 795 if (destroysource) { 796 /* 797 * Reverting the type is not necessary, the caller is going 798 * to destroy srcobject directly, but I'm doing it here 799 * for consistency since we've removed the object from its 800 * queues. 801 */ 802 swp_pager_meta_free_all(srcobject); 803 if (srcobject->type == OBJT_SWAP) 804 srcobject->type = OBJT_DEFAULT; 805 } 806 crit_exit(); 807 } 808 809 /* 810 * SWAP_PAGER_HASPAGE() - determine if we have good backing store for 811 * the requested page. 812 * 813 * We determine whether good backing store exists for the requested 814 * page and return TRUE if it does, FALSE if it doesn't. 815 * 816 * If TRUE, we also try to determine how much valid, contiguous backing 817 * store exists before and after the requested page within a reasonable 818 * distance. We do not try to restrict it to the swap device stripe 819 * (that is handled in getpages/putpages). It probably isn't worth 820 * doing here. 821 * 822 * No requirements. 823 */ 824 boolean_t 825 swap_pager_haspage(vm_object_t object, vm_pindex_t pindex) 826 { 827 swblk_t blk0; 828 829 /* 830 * do we have good backing store at the requested index ? 831 */ 832 833 crit_enter(); 834 lwkt_gettoken(&vm_token); 835 blk0 = swp_pager_meta_ctl(object, pindex, 0); 836 837 if (blk0 == SWAPBLK_NONE) { 838 lwkt_reltoken(&vm_token); 839 crit_exit(); 840 return (FALSE); 841 } 842 lwkt_reltoken(&vm_token); 843 crit_exit(); 844 return (TRUE); 845 } 846 847 /* 848 * SWAP_PAGER_PAGE_UNSWAPPED() - remove swap backing store related to page 849 * 850 * This removes any associated swap backing store, whether valid or 851 * not, from the page. This operates on any VM object, not just OBJT_SWAP 852 * objects. 853 * 854 * This routine is typically called when a page is made dirty, at 855 * which point any associated swap can be freed. MADV_FREE also 856 * calls us in a special-case situation 857 * 858 * NOTE!!! If the page is clean and the swap was valid, the caller 859 * should make the page dirty before calling this routine. This routine 860 * does NOT change the m->dirty status of the page. Also: MADV_FREE 861 * depends on it. 862 * 863 * The page must be busied or soft-busied. 864 * The caller must hold vm_token if the caller does not wish to block here. 865 * No other requirements. 866 */ 867 void 868 swap_pager_unswapped(vm_page_t m) 869 { 870 if (m->flags & PG_SWAPPED) { 871 crit_enter(); 872 lwkt_gettoken(&vm_token); 873 KKASSERT(m->flags & PG_SWAPPED); 874 swp_pager_meta_ctl(m->object, m->pindex, SWM_FREE); 875 vm_page_flag_clear(m, PG_SWAPPED); 876 lwkt_reltoken(&vm_token); 877 crit_exit(); 878 } 879 } 880 881 /* 882 * SWAP_PAGER_STRATEGY() - read, write, free blocks 883 * 884 * This implements a VM OBJECT strategy function using swap backing store. 885 * This can operate on any VM OBJECT type, not necessarily just OBJT_SWAP 886 * types. 887 * 888 * This is intended to be a cacheless interface (i.e. caching occurs at 889 * higher levels), and is also used as a swap-based SSD cache for vnode 890 * and device objects. 891 * 892 * All I/O goes directly to and from the swap device. 893 * 894 * We currently attempt to run I/O synchronously or asynchronously as 895 * the caller requests. This isn't perfect because we loose error 896 * sequencing when we run multiple ops in parallel to satisfy a request. 897 * But this is swap, so we let it all hang out. 898 * 899 * No requirements. 900 */ 901 void 902 swap_pager_strategy(vm_object_t object, struct bio *bio) 903 { 904 struct buf *bp = bio->bio_buf; 905 struct bio *nbio; 906 vm_pindex_t start; 907 vm_pindex_t biox_blkno = 0; 908 int count; 909 char *data; 910 struct bio *biox; 911 struct buf *bufx; 912 struct bio_track *track; 913 914 /* 915 * tracking for swapdev vnode I/Os 916 */ 917 if (bp->b_cmd == BUF_CMD_READ) 918 track = &swapdev_vp->v_track_read; 919 else 920 track = &swapdev_vp->v_track_write; 921 922 if (bp->b_bcount & PAGE_MASK) { 923 bp->b_error = EINVAL; 924 bp->b_flags |= B_ERROR | B_INVAL; 925 biodone(bio); 926 kprintf("swap_pager_strategy: bp %p offset %lld size %d, " 927 "not page bounded\n", 928 bp, (long long)bio->bio_offset, (int)bp->b_bcount); 929 return; 930 } 931 932 /* 933 * Clear error indication, initialize page index, count, data pointer. 934 */ 935 bp->b_error = 0; 936 bp->b_flags &= ~B_ERROR; 937 bp->b_resid = bp->b_bcount; 938 939 start = (vm_pindex_t)(bio->bio_offset >> PAGE_SHIFT); 940 count = howmany(bp->b_bcount, PAGE_SIZE); 941 data = bp->b_data; 942 943 /* 944 * Deal with BUF_CMD_FREEBLKS 945 */ 946 if (bp->b_cmd == BUF_CMD_FREEBLKS) { 947 /* 948 * FREE PAGE(s) - destroy underlying swap that is no longer 949 * needed. 950 */ 951 crit_enter(); 952 lwkt_gettoken(&vm_token); 953 swp_pager_meta_free(object, start, count); 954 lwkt_reltoken(&vm_token); 955 crit_exit(); 956 bp->b_resid = 0; 957 biodone(bio); 958 return; 959 } 960 961 /* 962 * We need to be able to create a new cluster of I/O's. We cannot 963 * use the caller fields of the passed bio so push a new one. 964 * 965 * Because nbio is just a placeholder for the cluster links, 966 * we can biodone() the original bio instead of nbio to make 967 * things a bit more efficient. 968 */ 969 nbio = push_bio(bio); 970 nbio->bio_offset = bio->bio_offset; 971 nbio->bio_caller_info1.cluster_head = NULL; 972 nbio->bio_caller_info2.cluster_tail = NULL; 973 974 biox = NULL; 975 bufx = NULL; 976 977 /* 978 * Execute read or write 979 */ 980 crit_enter(); 981 lwkt_gettoken(&vm_token); 982 while (count > 0) { 983 swblk_t blk; 984 985 /* 986 * Obtain block. If block not found and writing, allocate a 987 * new block and build it into the object. 988 */ 989 blk = swp_pager_meta_ctl(object, start, 0); 990 if ((blk == SWAPBLK_NONE) && bp->b_cmd != BUF_CMD_READ) { 991 blk = swp_pager_getswapspace(object, 1); 992 if (blk == SWAPBLK_NONE) { 993 bp->b_error = ENOMEM; 994 bp->b_flags |= B_ERROR; 995 break; 996 } 997 swp_pager_meta_build(object, start, blk); 998 } 999 1000 /* 1001 * Do we have to flush our current collection? Yes if: 1002 * 1003 * - no swap block at this index 1004 * - swap block is not contiguous 1005 * - we cross a physical disk boundry in the 1006 * stripe. 1007 */ 1008 if ( 1009 biox && (biox_blkno + btoc(bufx->b_bcount) != blk || 1010 ((biox_blkno ^ blk) & dmmax_mask) 1011 ) 1012 ) { 1013 if (bp->b_cmd == BUF_CMD_READ) { 1014 ++mycpu->gd_cnt.v_swapin; 1015 mycpu->gd_cnt.v_swappgsin += btoc(bufx->b_bcount); 1016 } else { 1017 ++mycpu->gd_cnt.v_swapout; 1018 mycpu->gd_cnt.v_swappgsout += btoc(bufx->b_bcount); 1019 bufx->b_dirtyend = bufx->b_bcount; 1020 } 1021 1022 /* 1023 * Finished with this buf. 1024 */ 1025 KKASSERT(bufx->b_bcount != 0); 1026 if (bufx->b_cmd != BUF_CMD_READ) 1027 bufx->b_dirtyend = bufx->b_bcount; 1028 biox = NULL; 1029 bufx = NULL; 1030 } 1031 1032 /* 1033 * Add new swapblk to biox, instantiating biox if necessary. 1034 * Zero-fill reads are able to take a shortcut. 1035 */ 1036 if (blk == SWAPBLK_NONE) { 1037 /* 1038 * We can only get here if we are reading. Since 1039 * we are at splvm() we can safely modify b_resid, 1040 * even if chain ops are in progress. 1041 */ 1042 bzero(data, PAGE_SIZE); 1043 bp->b_resid -= PAGE_SIZE; 1044 } else { 1045 if (biox == NULL) { 1046 /* XXX chain count > 4, wait to <= 4 */ 1047 1048 bufx = getpbuf(NULL); 1049 biox = &bufx->b_bio1; 1050 cluster_append(nbio, bufx); 1051 bufx->b_flags |= (bufx->b_flags & B_ORDERED); 1052 bufx->b_cmd = bp->b_cmd; 1053 biox->bio_done = swap_chain_iodone; 1054 biox->bio_offset = (off_t)blk << PAGE_SHIFT; 1055 biox->bio_caller_info1.cluster_parent = nbio; 1056 biox_blkno = blk; 1057 bufx->b_bcount = 0; 1058 bufx->b_data = data; 1059 } 1060 bufx->b_bcount += PAGE_SIZE; 1061 } 1062 --count; 1063 ++start; 1064 data += PAGE_SIZE; 1065 } 1066 lwkt_reltoken(&vm_token); 1067 crit_exit(); 1068 1069 /* 1070 * Flush out last buffer 1071 */ 1072 if (biox) { 1073 if (bufx->b_cmd == BUF_CMD_READ) { 1074 ++mycpu->gd_cnt.v_swapin; 1075 mycpu->gd_cnt.v_swappgsin += btoc(bufx->b_bcount); 1076 } else { 1077 ++mycpu->gd_cnt.v_swapout; 1078 mycpu->gd_cnt.v_swappgsout += btoc(bufx->b_bcount); 1079 bufx->b_dirtyend = bufx->b_bcount; 1080 } 1081 KKASSERT(bufx->b_bcount); 1082 if (bufx->b_cmd != BUF_CMD_READ) 1083 bufx->b_dirtyend = bufx->b_bcount; 1084 /* biox, bufx = NULL */ 1085 } 1086 1087 /* 1088 * Now initiate all the I/O. Be careful looping on our chain as 1089 * I/O's may complete while we are still initiating them. 1090 * 1091 * If the request is a 100% sparse read no bios will be present 1092 * and we just biodone() the buffer. 1093 */ 1094 nbio->bio_caller_info2.cluster_tail = NULL; 1095 bufx = nbio->bio_caller_info1.cluster_head; 1096 1097 if (bufx) { 1098 while (bufx) { 1099 biox = &bufx->b_bio1; 1100 BUF_KERNPROC(bufx); 1101 bufx = bufx->b_cluster_next; 1102 vn_strategy(swapdev_vp, biox); 1103 } 1104 } else { 1105 biodone(bio); 1106 } 1107 1108 /* 1109 * Completion of the cluster will also call biodone_chain(nbio). 1110 * We never call biodone(nbio) so we don't have to worry about 1111 * setting up a bio_done callback. It's handled in the sub-IO. 1112 */ 1113 /**/ 1114 } 1115 1116 /* 1117 * biodone callback 1118 * 1119 * No requirements. 1120 */ 1121 static void 1122 swap_chain_iodone(struct bio *biox) 1123 { 1124 struct buf **nextp; 1125 struct buf *bufx; /* chained sub-buffer */ 1126 struct bio *nbio; /* parent nbio with chain glue */ 1127 struct buf *bp; /* original bp associated with nbio */ 1128 int chain_empty; 1129 1130 bufx = biox->bio_buf; 1131 nbio = biox->bio_caller_info1.cluster_parent; 1132 bp = nbio->bio_buf; 1133 1134 /* 1135 * Update the original buffer 1136 */ 1137 KKASSERT(bp != NULL); 1138 if (bufx->b_flags & B_ERROR) { 1139 atomic_set_int(&bufx->b_flags, B_ERROR); 1140 bp->b_error = bufx->b_error; /* race ok */ 1141 } else if (bufx->b_resid != 0) { 1142 atomic_set_int(&bufx->b_flags, B_ERROR); 1143 bp->b_error = EINVAL; /* race ok */ 1144 } else { 1145 atomic_subtract_int(&bp->b_resid, bufx->b_bcount); 1146 } 1147 1148 /* 1149 * Remove us from the chain. 1150 */ 1151 spin_lock(&bp->b_lock.lk_spinlock); 1152 nextp = &nbio->bio_caller_info1.cluster_head; 1153 while (*nextp != bufx) { 1154 KKASSERT(*nextp != NULL); 1155 nextp = &(*nextp)->b_cluster_next; 1156 } 1157 *nextp = bufx->b_cluster_next; 1158 chain_empty = (nbio->bio_caller_info1.cluster_head == NULL); 1159 spin_unlock(&bp->b_lock.lk_spinlock); 1160 1161 /* 1162 * Clean up bufx. If the chain is now empty we finish out 1163 * the parent. Note that we may be racing other completions 1164 * so we must use the chain_empty status from above. 1165 */ 1166 if (chain_empty) { 1167 if (bp->b_resid != 0 && !(bp->b_flags & B_ERROR)) { 1168 atomic_set_int(&bp->b_flags, B_ERROR); 1169 bp->b_error = EINVAL; 1170 } 1171 biodone_chain(nbio); 1172 } 1173 relpbuf(bufx, NULL); 1174 } 1175 1176 /* 1177 * SWAP_PAGER_GETPAGES() - bring page in from swap 1178 * 1179 * The requested page may have to be brought in from swap. Calculate the 1180 * swap block and bring in additional pages if possible. All pages must 1181 * have contiguous swap block assignments and reside in the same object. 1182 * 1183 * The caller has a single vm_object_pip_add() reference prior to 1184 * calling us and we should return with the same. 1185 * 1186 * The caller has BUSY'd the page. We should return with (*mpp) left busy, 1187 * and any additinal pages unbusied. 1188 * 1189 * If the caller encounters a PG_RAM page it will pass it to us even though 1190 * it may be valid and dirty. We cannot overwrite the page in this case! 1191 * The case is used to allow us to issue pure read-aheads. 1192 * 1193 * NOTE! XXX This code does not entirely pipeline yet due to the fact that 1194 * the PG_RAM page is validated at the same time as mreq. What we 1195 * really need to do is issue a separate read-ahead pbuf. 1196 * 1197 * No requirements. 1198 */ 1199 static int 1200 swap_pager_getpage(vm_object_t object, vm_page_t *mpp, int seqaccess) 1201 { 1202 struct buf *bp; 1203 struct bio *bio; 1204 vm_page_t mreq; 1205 vm_page_t m; 1206 vm_offset_t kva; 1207 swblk_t blk; 1208 int i; 1209 int j; 1210 int raonly; 1211 vm_page_t marray[XIO_INTERNAL_PAGES]; 1212 1213 mreq = *mpp; 1214 1215 if (mreq->object != object) { 1216 panic("swap_pager_getpages: object mismatch %p/%p", 1217 object, 1218 mreq->object 1219 ); 1220 } 1221 1222 /* 1223 * We don't want to overwrite a fully valid page as it might be 1224 * dirty. This case can occur when e.g. vm_fault hits a perfectly 1225 * valid page with PG_RAM set. 1226 * 1227 * In this case we see if the next page is a suitable page-in 1228 * candidate and if it is we issue read-ahead. PG_RAM will be 1229 * set on the last page of the read-ahead to continue the pipeline. 1230 */ 1231 if (mreq->valid == VM_PAGE_BITS_ALL) { 1232 if (swap_burst_read == 0 || mreq->pindex + 1 >= object->size) 1233 return(VM_PAGER_OK); 1234 crit_enter(); 1235 lwkt_gettoken(&vm_token); 1236 blk = swp_pager_meta_ctl(object, mreq->pindex + 1, 0); 1237 if (blk == SWAPBLK_NONE) { 1238 lwkt_reltoken(&vm_token); 1239 crit_exit(); 1240 return(VM_PAGER_OK); 1241 } 1242 m = vm_page_lookup(object, mreq->pindex + 1); 1243 if (m == NULL) { 1244 m = vm_page_alloc(object, mreq->pindex + 1, 1245 VM_ALLOC_QUICK); 1246 if (m == NULL) { 1247 lwkt_reltoken(&vm_token); 1248 crit_exit(); 1249 return(VM_PAGER_OK); 1250 } 1251 } else { 1252 if ((m->flags & PG_BUSY) || m->busy || m->valid) { 1253 lwkt_reltoken(&vm_token); 1254 crit_exit(); 1255 return(VM_PAGER_OK); 1256 } 1257 vm_page_unqueue_nowakeup(m); 1258 vm_page_busy(m); 1259 } 1260 mreq = m; 1261 raonly = 1; 1262 lwkt_reltoken(&vm_token); 1263 crit_exit(); 1264 } else { 1265 raonly = 0; 1266 } 1267 1268 /* 1269 * Try to block-read contiguous pages from swap if sequential, 1270 * otherwise just read one page. Contiguous pages from swap must 1271 * reside within a single device stripe because the I/O cannot be 1272 * broken up across multiple stripes. 1273 * 1274 * Note that blk and iblk can be SWAPBLK_NONE but the loop is 1275 * set up such that the case(s) are handled implicitly. 1276 */ 1277 crit_enter(); 1278 lwkt_gettoken(&vm_token); 1279 blk = swp_pager_meta_ctl(mreq->object, mreq->pindex, 0); 1280 marray[0] = mreq; 1281 1282 for (i = 1; swap_burst_read && 1283 i < XIO_INTERNAL_PAGES && 1284 mreq->pindex + i < object->size; ++i) { 1285 swblk_t iblk; 1286 1287 iblk = swp_pager_meta_ctl(object, mreq->pindex + i, 0); 1288 if (iblk != blk + i) 1289 break; 1290 if ((blk ^ iblk) & dmmax_mask) 1291 break; 1292 m = vm_page_lookup(object, mreq->pindex + i); 1293 if (m == NULL) { 1294 m = vm_page_alloc(object, mreq->pindex + i, 1295 VM_ALLOC_QUICK); 1296 if (m == NULL) 1297 break; 1298 } else { 1299 if ((m->flags & PG_BUSY) || m->busy || m->valid) 1300 break; 1301 vm_page_unqueue_nowakeup(m); 1302 vm_page_busy(m); 1303 } 1304 marray[i] = m; 1305 } 1306 if (i > 1) 1307 vm_page_flag_set(marray[i - 1], PG_RAM); 1308 1309 lwkt_reltoken(&vm_token); 1310 crit_exit(); 1311 1312 /* 1313 * If mreq is the requested page and we have nothing to do return 1314 * VM_PAGER_FAIL. If raonly is set mreq is just another read-ahead 1315 * page and must be cleaned up. 1316 */ 1317 if (blk == SWAPBLK_NONE) { 1318 KKASSERT(i == 1); 1319 if (raonly) { 1320 vnode_pager_freepage(mreq); 1321 return(VM_PAGER_OK); 1322 } else { 1323 return(VM_PAGER_FAIL); 1324 } 1325 } 1326 1327 /* 1328 * map our page(s) into kva for input 1329 */ 1330 bp = getpbuf_kva(&nsw_rcount); 1331 bio = &bp->b_bio1; 1332 kva = (vm_offset_t) bp->b_kvabase; 1333 bcopy(marray, bp->b_xio.xio_pages, i * sizeof(vm_page_t)); 1334 pmap_qenter(kva, bp->b_xio.xio_pages, i); 1335 1336 bp->b_data = (caddr_t)kva; 1337 bp->b_bcount = PAGE_SIZE * i; 1338 bp->b_xio.xio_npages = i; 1339 bio->bio_done = swp_pager_async_iodone; 1340 bio->bio_offset = (off_t)blk << PAGE_SHIFT; 1341 bio->bio_caller_info1.index = SWBIO_READ; 1342 1343 /* 1344 * Set index. If raonly set the index beyond the array so all 1345 * the pages are treated the same, otherwise the original mreq is 1346 * at index 0. 1347 */ 1348 if (raonly) 1349 bio->bio_driver_info = (void *)(intptr_t)i; 1350 else 1351 bio->bio_driver_info = (void *)(intptr_t)0; 1352 1353 for (j = 0; j < i; ++j) 1354 vm_page_flag_set(bp->b_xio.xio_pages[j], PG_SWAPINPROG); 1355 1356 mycpu->gd_cnt.v_swapin++; 1357 mycpu->gd_cnt.v_swappgsin += bp->b_xio.xio_npages; 1358 1359 /* 1360 * We still hold the lock on mreq, and our automatic completion routine 1361 * does not remove it. 1362 */ 1363 vm_object_pip_add(object, bp->b_xio.xio_npages); 1364 1365 /* 1366 * perform the I/O. NOTE!!! bp cannot be considered valid after 1367 * this point because we automatically release it on completion. 1368 * Instead, we look at the one page we are interested in which we 1369 * still hold a lock on even through the I/O completion. 1370 * 1371 * The other pages in our m[] array are also released on completion, 1372 * so we cannot assume they are valid anymore either. 1373 */ 1374 bp->b_cmd = BUF_CMD_READ; 1375 BUF_KERNPROC(bp); 1376 vn_strategy(swapdev_vp, bio); 1377 1378 /* 1379 * Wait for the page we want to complete. PG_SWAPINPROG is always 1380 * cleared on completion. If an I/O error occurs, SWAPBLK_NONE 1381 * is set in the meta-data. 1382 * 1383 * If this is a read-ahead only we return immediately without 1384 * waiting for I/O. 1385 */ 1386 if (raonly) 1387 return(VM_PAGER_OK); 1388 1389 /* 1390 * Read-ahead includes originally requested page case. 1391 */ 1392 crit_enter(); 1393 lwkt_gettoken(&vm_token); 1394 while ((mreq->flags & PG_SWAPINPROG) != 0) { 1395 vm_page_flag_set(mreq, PG_WANTED | PG_REFERENCED); 1396 mycpu->gd_cnt.v_intrans++; 1397 if (tsleep(mreq, 0, "swread", hz*20)) { 1398 kprintf( 1399 "swap_pager: indefinite wait buffer: " 1400 " offset: %lld, size: %ld\n", 1401 (long long)bio->bio_offset, 1402 (long)bp->b_bcount 1403 ); 1404 } 1405 } 1406 lwkt_reltoken(&vm_token); 1407 crit_exit(); 1408 1409 /* 1410 * mreq is left bussied after completion, but all the other pages 1411 * are freed. If we had an unrecoverable read error the page will 1412 * not be valid. 1413 */ 1414 if (mreq->valid != VM_PAGE_BITS_ALL) 1415 return(VM_PAGER_ERROR); 1416 else 1417 return(VM_PAGER_OK); 1418 1419 /* 1420 * A final note: in a low swap situation, we cannot deallocate swap 1421 * and mark a page dirty here because the caller is likely to mark 1422 * the page clean when we return, causing the page to possibly revert 1423 * to all-zero's later. 1424 */ 1425 } 1426 1427 /* 1428 * swap_pager_putpages: 1429 * 1430 * Assign swap (if necessary) and initiate I/O on the specified pages. 1431 * 1432 * We support both OBJT_DEFAULT and OBJT_SWAP objects. DEFAULT objects 1433 * are automatically converted to SWAP objects. 1434 * 1435 * In a low memory situation we may block in vn_strategy(), but the new 1436 * vm_page reservation system coupled with properly written VFS devices 1437 * should ensure that no low-memory deadlock occurs. This is an area 1438 * which needs work. 1439 * 1440 * The parent has N vm_object_pip_add() references prior to 1441 * calling us and will remove references for rtvals[] that are 1442 * not set to VM_PAGER_PEND. We need to remove the rest on I/O 1443 * completion. 1444 * 1445 * The parent has soft-busy'd the pages it passes us and will unbusy 1446 * those whos rtvals[] entry is not set to VM_PAGER_PEND on return. 1447 * We need to unbusy the rest on I/O completion. 1448 * 1449 * No requirements. 1450 */ 1451 void 1452 swap_pager_putpages(vm_object_t object, vm_page_t *m, int count, 1453 boolean_t sync, int *rtvals) 1454 { 1455 int i; 1456 int n = 0; 1457 1458 if (count && m[0]->object != object) { 1459 panic("swap_pager_getpages: object mismatch %p/%p", 1460 object, 1461 m[0]->object 1462 ); 1463 } 1464 1465 /* 1466 * Step 1 1467 * 1468 * Turn object into OBJT_SWAP 1469 * check for bogus sysops 1470 * force sync if not pageout process 1471 */ 1472 if (object->type == OBJT_DEFAULT) { 1473 lwkt_gettoken(&vm_token); 1474 if (object->type == OBJT_DEFAULT) 1475 swp_pager_meta_convert(object); 1476 lwkt_reltoken(&vm_token); 1477 } 1478 1479 if (curthread != pagethread) 1480 sync = TRUE; 1481 1482 /* 1483 * Step 2 1484 * 1485 * Update nsw parameters from swap_async_max sysctl values. 1486 * Do not let the sysop crash the machine with bogus numbers. 1487 */ 1488 1489 if (swap_async_max != nsw_wcount_async_max) { 1490 int n; 1491 1492 /* 1493 * limit range 1494 */ 1495 if ((n = swap_async_max) > nswbuf / 2) 1496 n = nswbuf / 2; 1497 if (n < 1) 1498 n = 1; 1499 swap_async_max = n; 1500 1501 /* 1502 * Adjust difference ( if possible ). If the current async 1503 * count is too low, we may not be able to make the adjustment 1504 * at this time. 1505 */ 1506 crit_enter(); 1507 lwkt_gettoken(&vm_token); 1508 n -= nsw_wcount_async_max; 1509 if (nsw_wcount_async + n >= 0) { 1510 nsw_wcount_async += n; 1511 nsw_wcount_async_max += n; 1512 wakeup(&nsw_wcount_async); 1513 } 1514 lwkt_reltoken(&vm_token); 1515 crit_exit(); 1516 } 1517 1518 /* 1519 * Step 3 1520 * 1521 * Assign swap blocks and issue I/O. We reallocate swap on the fly. 1522 * The page is left dirty until the pageout operation completes 1523 * successfully. 1524 */ 1525 1526 for (i = 0; i < count; i += n) { 1527 struct buf *bp; 1528 struct bio *bio; 1529 swblk_t blk; 1530 int j; 1531 1532 /* 1533 * Maximum I/O size is limited by a number of factors. 1534 */ 1535 1536 n = min(BLIST_MAX_ALLOC, count - i); 1537 n = min(n, nsw_cluster_max); 1538 1539 crit_enter(); 1540 lwkt_gettoken(&vm_token); 1541 1542 /* 1543 * Get biggest block of swap we can. If we fail, fall 1544 * back and try to allocate a smaller block. Don't go 1545 * overboard trying to allocate space if it would overly 1546 * fragment swap. 1547 */ 1548 while ( 1549 (blk = swp_pager_getswapspace(object, n)) == SWAPBLK_NONE && 1550 n > 4 1551 ) { 1552 n >>= 1; 1553 } 1554 if (blk == SWAPBLK_NONE) { 1555 for (j = 0; j < n; ++j) 1556 rtvals[i+j] = VM_PAGER_FAIL; 1557 lwkt_reltoken(&vm_token); 1558 crit_exit(); 1559 continue; 1560 } 1561 1562 /* 1563 * The I/O we are constructing cannot cross a physical 1564 * disk boundry in the swap stripe. Note: we are still 1565 * at splvm(). 1566 */ 1567 if ((blk ^ (blk + n)) & dmmax_mask) { 1568 j = ((blk + dmmax) & dmmax_mask) - blk; 1569 swp_pager_freeswapspace(object, blk + j, n - j); 1570 n = j; 1571 } 1572 1573 /* 1574 * All I/O parameters have been satisfied, build the I/O 1575 * request and assign the swap space. 1576 */ 1577 if (sync == TRUE) 1578 bp = getpbuf_kva(&nsw_wcount_sync); 1579 else 1580 bp = getpbuf_kva(&nsw_wcount_async); 1581 bio = &bp->b_bio1; 1582 1583 pmap_qenter((vm_offset_t)bp->b_data, &m[i], n); 1584 1585 bp->b_bcount = PAGE_SIZE * n; 1586 bio->bio_offset = (off_t)blk << PAGE_SHIFT; 1587 1588 for (j = 0; j < n; ++j) { 1589 vm_page_t mreq = m[i+j]; 1590 1591 swp_pager_meta_build(mreq->object, mreq->pindex, 1592 blk + j); 1593 if (object->type == OBJT_SWAP) 1594 vm_page_dirty(mreq); 1595 rtvals[i+j] = VM_PAGER_OK; 1596 1597 vm_page_flag_set(mreq, PG_SWAPINPROG); 1598 bp->b_xio.xio_pages[j] = mreq; 1599 } 1600 bp->b_xio.xio_npages = n; 1601 1602 mycpu->gd_cnt.v_swapout++; 1603 mycpu->gd_cnt.v_swappgsout += bp->b_xio.xio_npages; 1604 1605 lwkt_reltoken(&vm_token); 1606 crit_exit(); 1607 1608 bp->b_dirtyoff = 0; /* req'd for NFS */ 1609 bp->b_dirtyend = bp->b_bcount; /* req'd for NFS */ 1610 bp->b_cmd = BUF_CMD_WRITE; 1611 bio->bio_caller_info1.index = SWBIO_WRITE; 1612 1613 /* 1614 * asynchronous 1615 */ 1616 if (sync == FALSE) { 1617 bio->bio_done = swp_pager_async_iodone; 1618 BUF_KERNPROC(bp); 1619 vn_strategy(swapdev_vp, bio); 1620 1621 for (j = 0; j < n; ++j) 1622 rtvals[i+j] = VM_PAGER_PEND; 1623 continue; 1624 } 1625 1626 /* 1627 * Issue synchrnously. 1628 * 1629 * Wait for the sync I/O to complete, then update rtvals. 1630 * We just set the rtvals[] to VM_PAGER_PEND so we can call 1631 * our async completion routine at the end, thus avoiding a 1632 * double-free. 1633 */ 1634 bio->bio_caller_info1.index |= SWBIO_SYNC; 1635 bio->bio_done = biodone_sync; 1636 bio->bio_flags |= BIO_SYNC; 1637 vn_strategy(swapdev_vp, bio); 1638 biowait(bio, "swwrt"); 1639 1640 for (j = 0; j < n; ++j) 1641 rtvals[i+j] = VM_PAGER_PEND; 1642 1643 /* 1644 * Now that we are through with the bp, we can call the 1645 * normal async completion, which frees everything up. 1646 */ 1647 swp_pager_async_iodone(bio); 1648 } 1649 } 1650 1651 /* 1652 * No requirements. 1653 */ 1654 void 1655 swap_pager_newswap(void) 1656 { 1657 swp_sizecheck(); 1658 } 1659 1660 /* 1661 * swp_pager_async_iodone: 1662 * 1663 * Completion routine for asynchronous reads and writes from/to swap. 1664 * Also called manually by synchronous code to finish up a bp. 1665 * 1666 * For READ operations, the pages are PG_BUSY'd. For WRITE operations, 1667 * the pages are vm_page_t->busy'd. For READ operations, we PG_BUSY 1668 * unbusy all pages except the 'main' request page. For WRITE 1669 * operations, we vm_page_t->busy'd unbusy all pages ( we can do this 1670 * because we marked them all VM_PAGER_PEND on return from putpages ). 1671 * 1672 * This routine may not block. 1673 * 1674 * No requirements. 1675 */ 1676 static void 1677 swp_pager_async_iodone(struct bio *bio) 1678 { 1679 struct buf *bp = bio->bio_buf; 1680 vm_object_t object = NULL; 1681 int i; 1682 int *nswptr; 1683 1684 /* 1685 * report error 1686 */ 1687 if (bp->b_flags & B_ERROR) { 1688 kprintf( 1689 "swap_pager: I/O error - %s failed; offset %lld," 1690 "size %ld, error %d\n", 1691 ((bio->bio_caller_info1.index & SWBIO_READ) ? 1692 "pagein" : "pageout"), 1693 (long long)bio->bio_offset, 1694 (long)bp->b_bcount, 1695 bp->b_error 1696 ); 1697 } 1698 1699 /* 1700 * set object, raise to splvm(). 1701 */ 1702 if (bp->b_xio.xio_npages) 1703 object = bp->b_xio.xio_pages[0]->object; 1704 crit_enter(); 1705 lwkt_gettoken(&vm_token); 1706 1707 /* 1708 * remove the mapping for kernel virtual 1709 */ 1710 pmap_qremove((vm_offset_t)bp->b_data, bp->b_xio.xio_npages); 1711 1712 /* 1713 * cleanup pages. If an error occurs writing to swap, we are in 1714 * very serious trouble. If it happens to be a disk error, though, 1715 * we may be able to recover by reassigning the swap later on. So 1716 * in this case we remove the m->swapblk assignment for the page 1717 * but do not free it in the rlist. The errornous block(s) are thus 1718 * never reallocated as swap. Redirty the page and continue. 1719 */ 1720 for (i = 0; i < bp->b_xio.xio_npages; ++i) { 1721 vm_page_t m = bp->b_xio.xio_pages[i]; 1722 1723 if (bp->b_flags & B_ERROR) { 1724 /* 1725 * If an error occurs I'd love to throw the swapblk 1726 * away without freeing it back to swapspace, so it 1727 * can never be used again. But I can't from an 1728 * interrupt. 1729 */ 1730 1731 if (bio->bio_caller_info1.index & SWBIO_READ) { 1732 /* 1733 * When reading, reqpage needs to stay 1734 * locked for the parent, but all other 1735 * pages can be freed. We still want to 1736 * wakeup the parent waiting on the page, 1737 * though. ( also: pg_reqpage can be -1 and 1738 * not match anything ). 1739 * 1740 * We have to wake specifically requested pages 1741 * up too because we cleared PG_SWAPINPROG and 1742 * someone may be waiting for that. 1743 * 1744 * NOTE: for reads, m->dirty will probably 1745 * be overridden by the original caller of 1746 * getpages so don't play cute tricks here. 1747 * 1748 * NOTE: We can't actually free the page from 1749 * here, because this is an interrupt. It 1750 * is not legal to mess with object->memq 1751 * from an interrupt. Deactivate the page 1752 * instead. 1753 */ 1754 1755 m->valid = 0; 1756 vm_page_flag_clear(m, PG_ZERO); 1757 vm_page_flag_clear(m, PG_SWAPINPROG); 1758 1759 /* 1760 * bio_driver_info holds the requested page 1761 * index. 1762 */ 1763 if (i != (int)(intptr_t)bio->bio_driver_info) { 1764 vm_page_deactivate(m); 1765 vm_page_wakeup(m); 1766 } else { 1767 vm_page_flash(m); 1768 } 1769 /* 1770 * If i == bp->b_pager.pg_reqpage, do not wake 1771 * the page up. The caller needs to. 1772 */ 1773 } else { 1774 /* 1775 * If a write error occurs remove the swap 1776 * assignment (note that PG_SWAPPED may or 1777 * may not be set depending on prior activity). 1778 * 1779 * Re-dirty OBJT_SWAP pages as there is no 1780 * other backing store, we can't throw the 1781 * page away. 1782 * 1783 * Non-OBJT_SWAP pages (aka swapcache) must 1784 * not be dirtied since they may not have 1785 * been dirty in the first place, and they 1786 * do have backing store (the vnode). 1787 */ 1788 swp_pager_meta_ctl(m->object, m->pindex, 1789 SWM_FREE); 1790 vm_page_flag_clear(m, PG_SWAPPED); 1791 if (m->object->type == OBJT_SWAP) { 1792 vm_page_dirty(m); 1793 vm_page_activate(m); 1794 } 1795 vm_page_flag_clear(m, PG_SWAPINPROG); 1796 vm_page_io_finish(m); 1797 } 1798 } else if (bio->bio_caller_info1.index & SWBIO_READ) { 1799 /* 1800 * NOTE: for reads, m->dirty will probably be 1801 * overridden by the original caller of getpages so 1802 * we cannot set them in order to free the underlying 1803 * swap in a low-swap situation. I don't think we'd 1804 * want to do that anyway, but it was an optimization 1805 * that existed in the old swapper for a time before 1806 * it got ripped out due to precisely this problem. 1807 * 1808 * clear PG_ZERO in page. 1809 * 1810 * If not the requested page then deactivate it. 1811 * 1812 * Note that the requested page, reqpage, is left 1813 * busied, but we still have to wake it up. The 1814 * other pages are released (unbusied) by 1815 * vm_page_wakeup(). We do not set reqpage's 1816 * valid bits here, it is up to the caller. 1817 */ 1818 1819 /* 1820 * NOTE: can't call pmap_clear_modify(m) from an 1821 * interrupt thread, the pmap code may have to map 1822 * non-kernel pmaps and currently asserts the case. 1823 */ 1824 /*pmap_clear_modify(m);*/ 1825 m->valid = VM_PAGE_BITS_ALL; 1826 vm_page_undirty(m); 1827 vm_page_flag_clear(m, PG_ZERO | PG_SWAPINPROG); 1828 vm_page_flag_set(m, PG_SWAPPED); 1829 1830 /* 1831 * We have to wake specifically requested pages 1832 * up too because we cleared PG_SWAPINPROG and 1833 * could be waiting for it in getpages. However, 1834 * be sure to not unbusy getpages specifically 1835 * requested page - getpages expects it to be 1836 * left busy. 1837 * 1838 * bio_driver_info holds the requested page 1839 */ 1840 if (i != (int)(intptr_t)bio->bio_driver_info) { 1841 vm_page_deactivate(m); 1842 vm_page_wakeup(m); 1843 } else { 1844 vm_page_flash(m); 1845 } 1846 } else { 1847 /* 1848 * Mark the page clean but do not mess with the 1849 * pmap-layer's modified state. That state should 1850 * also be clear since the caller protected the 1851 * page VM_PROT_READ, but allow the case. 1852 * 1853 * We are in an interrupt, avoid pmap operations. 1854 * 1855 * If we have a severe page deficit, deactivate the 1856 * page. Do not try to cache it (which would also 1857 * involve a pmap op), because the page might still 1858 * be read-heavy. 1859 * 1860 * When using the swap to cache clean vnode pages 1861 * we do not mess with the page dirty bits. 1862 */ 1863 if (m->object->type == OBJT_SWAP) 1864 vm_page_undirty(m); 1865 vm_page_flag_clear(m, PG_SWAPINPROG); 1866 vm_page_flag_set(m, PG_SWAPPED); 1867 vm_page_io_finish(m); 1868 if (vm_page_count_severe()) 1869 vm_page_deactivate(m); 1870 #if 0 1871 if (!vm_page_count_severe() || !vm_page_try_to_cache(m)) 1872 vm_page_protect(m, VM_PROT_READ); 1873 #endif 1874 } 1875 } 1876 1877 /* 1878 * adjust pip. NOTE: the original parent may still have its own 1879 * pip refs on the object. 1880 */ 1881 1882 if (object) 1883 vm_object_pip_wakeupn(object, bp->b_xio.xio_npages); 1884 1885 /* 1886 * Release the physical I/O buffer. 1887 * 1888 * NOTE: Due to synchronous operations in the write case b_cmd may 1889 * already be set to BUF_CMD_DONE and BIO_SYNC may have already 1890 * been cleared. 1891 */ 1892 if (bio->bio_caller_info1.index & SWBIO_READ) 1893 nswptr = &nsw_rcount; 1894 else if (bio->bio_caller_info1.index & SWBIO_SYNC) 1895 nswptr = &nsw_wcount_sync; 1896 else 1897 nswptr = &nsw_wcount_async; 1898 bp->b_cmd = BUF_CMD_DONE; 1899 relpbuf(bp, nswptr); 1900 lwkt_reltoken(&vm_token); 1901 crit_exit(); 1902 } 1903 1904 /* 1905 * Fault-in a potentially swapped page and remove the swap reference. 1906 */ 1907 static __inline void 1908 swp_pager_fault_page(vm_object_t object, vm_pindex_t pindex) 1909 { 1910 struct vnode *vp; 1911 vm_page_t m; 1912 int error; 1913 1914 if (object->type == OBJT_VNODE) { 1915 /* 1916 * Any swap related to a vnode is due to swapcache. We must 1917 * vget() the vnode in case it is not active (otherwise 1918 * vref() will panic). Calling vm_object_page_remove() will 1919 * ensure that any swap ref is removed interlocked with the 1920 * page. clean_only is set to TRUE so we don't throw away 1921 * dirty pages. 1922 */ 1923 vp = object->handle; 1924 error = vget(vp, LK_SHARED | LK_RETRY | LK_CANRECURSE); 1925 if (error == 0) { 1926 vm_object_page_remove(object, pindex, pindex + 1, TRUE); 1927 vput(vp); 1928 } 1929 } else { 1930 /* 1931 * Otherwise it is a normal OBJT_SWAP object and we can 1932 * fault the page in and remove the swap. 1933 */ 1934 m = vm_fault_object_page(object, IDX_TO_OFF(pindex), 1935 VM_PROT_NONE, 1936 VM_FAULT_DIRTY | VM_FAULT_UNSWAP, 1937 &error); 1938 if (m) 1939 vm_page_unhold(m); 1940 } 1941 } 1942 1943 int 1944 swap_pager_swapoff(int devidx) 1945 { 1946 vm_object_t object; 1947 struct swblock *swap; 1948 swblk_t v; 1949 int i; 1950 1951 lwkt_gettoken(&vm_token); 1952 lwkt_gettoken(&vmobj_token); 1953 rescan: 1954 TAILQ_FOREACH(object, &vm_object_list, object_list) { 1955 if (object->type == OBJT_SWAP || object->type == OBJT_VNODE) { 1956 RB_FOREACH(swap, swblock_rb_tree, &object->swblock_root) { 1957 for (i = 0; i < SWAP_META_PAGES; ++i) { 1958 v = swap->swb_pages[i]; 1959 if (v != SWAPBLK_NONE && 1960 BLK2DEVIDX(v) == devidx) { 1961 swp_pager_fault_page( 1962 object, 1963 swap->swb_index + i); 1964 goto rescan; 1965 } 1966 } 1967 } 1968 } 1969 } 1970 lwkt_reltoken(&vmobj_token); 1971 lwkt_reltoken(&vm_token); 1972 1973 /* 1974 * If we fail to locate all swblocks we just fail gracefully and 1975 * do not bother to restore paging on the swap device. If the 1976 * user wants to retry the user can retry. 1977 */ 1978 if (swdevt[devidx].sw_nused) 1979 return (1); 1980 else 1981 return (0); 1982 } 1983 1984 /************************************************************************ 1985 * SWAP META DATA * 1986 ************************************************************************ 1987 * 1988 * These routines manipulate the swap metadata stored in the 1989 * OBJT_SWAP object. All swp_*() routines must be called at 1990 * splvm() because swap can be freed up by the low level vm_page 1991 * code which might be called from interrupts beyond what splbio() covers. 1992 * 1993 * Swap metadata is implemented with a global hash and not directly 1994 * linked into the object. Instead the object simply contains 1995 * appropriate tracking counters. 1996 */ 1997 1998 /* 1999 * Lookup the swblock containing the specified swap block index. 2000 * 2001 * The caller must hold vm_token. 2002 */ 2003 static __inline 2004 struct swblock * 2005 swp_pager_lookup(vm_object_t object, vm_pindex_t index) 2006 { 2007 index &= ~SWAP_META_MASK; 2008 return (RB_LOOKUP(swblock_rb_tree, &object->swblock_root, index)); 2009 } 2010 2011 /* 2012 * Remove a swblock from the RB tree. 2013 * 2014 * The caller must hold vm_token. 2015 */ 2016 static __inline 2017 void 2018 swp_pager_remove(vm_object_t object, struct swblock *swap) 2019 { 2020 RB_REMOVE(swblock_rb_tree, &object->swblock_root, swap); 2021 } 2022 2023 /* 2024 * Convert default object to swap object if necessary 2025 * 2026 * The caller must hold vm_token. 2027 */ 2028 static void 2029 swp_pager_meta_convert(vm_object_t object) 2030 { 2031 if (object->type == OBJT_DEFAULT) { 2032 object->type = OBJT_SWAP; 2033 KKASSERT(object->swblock_count == 0); 2034 } 2035 } 2036 2037 /* 2038 * SWP_PAGER_META_BUILD() - add swap block to swap meta data for object 2039 * 2040 * We first convert the object to a swap object if it is a default 2041 * object. Vnode objects do not need to be converted. 2042 * 2043 * The specified swapblk is added to the object's swap metadata. If 2044 * the swapblk is not valid, it is freed instead. Any previously 2045 * assigned swapblk is freed. 2046 * 2047 * The caller must hold vm_token. 2048 */ 2049 static void 2050 swp_pager_meta_build(vm_object_t object, vm_pindex_t index, swblk_t swapblk) 2051 { 2052 struct swblock *swap; 2053 struct swblock *oswap; 2054 2055 KKASSERT(swapblk != SWAPBLK_NONE); 2056 2057 /* 2058 * Convert object if necessary 2059 */ 2060 if (object->type == OBJT_DEFAULT) 2061 swp_pager_meta_convert(object); 2062 2063 /* 2064 * Locate swblock. If not found create, but if we aren't adding 2065 * anything just return. If we run out of space in the map we wait 2066 * and, since the hash table may have changed, retry. 2067 */ 2068 retry: 2069 swap = swp_pager_lookup(object, index); 2070 2071 if (swap == NULL) { 2072 int i; 2073 2074 swap = zalloc(swap_zone); 2075 if (swap == NULL) { 2076 vm_wait(0); 2077 goto retry; 2078 } 2079 swap->swb_index = index & ~SWAP_META_MASK; 2080 swap->swb_count = 0; 2081 2082 ++object->swblock_count; 2083 2084 for (i = 0; i < SWAP_META_PAGES; ++i) 2085 swap->swb_pages[i] = SWAPBLK_NONE; 2086 oswap = RB_INSERT(swblock_rb_tree, &object->swblock_root, swap); 2087 KKASSERT(oswap == NULL); 2088 } 2089 2090 /* 2091 * Delete prior contents of metadata 2092 */ 2093 2094 index &= SWAP_META_MASK; 2095 2096 if (swap->swb_pages[index] != SWAPBLK_NONE) { 2097 swp_pager_freeswapspace(object, swap->swb_pages[index], 1); 2098 --swap->swb_count; 2099 } 2100 2101 /* 2102 * Enter block into metadata 2103 */ 2104 swap->swb_pages[index] = swapblk; 2105 if (swapblk != SWAPBLK_NONE) 2106 ++swap->swb_count; 2107 } 2108 2109 /* 2110 * SWP_PAGER_META_FREE() - free a range of blocks in the object's swap metadata 2111 * 2112 * The requested range of blocks is freed, with any associated swap 2113 * returned to the swap bitmap. 2114 * 2115 * This routine will free swap metadata structures as they are cleaned 2116 * out. This routine does *NOT* operate on swap metadata associated 2117 * with resident pages. 2118 * 2119 * The caller must hold vm_token. 2120 */ 2121 static int swp_pager_meta_free_callback(struct swblock *swb, void *data); 2122 2123 static void 2124 swp_pager_meta_free(vm_object_t object, vm_pindex_t index, vm_pindex_t count) 2125 { 2126 struct swfreeinfo info; 2127 2128 /* 2129 * Nothing to do 2130 */ 2131 if (object->swblock_count == 0) { 2132 KKASSERT(RB_EMPTY(&object->swblock_root)); 2133 return; 2134 } 2135 if (count == 0) 2136 return; 2137 2138 /* 2139 * Setup for RB tree scan. Note that the pindex range can be huge 2140 * due to the 64 bit page index space so we cannot safely iterate. 2141 */ 2142 info.object = object; 2143 info.basei = index & ~SWAP_META_MASK; 2144 info.begi = index; 2145 info.endi = index + count - 1; 2146 swblock_rb_tree_RB_SCAN(&object->swblock_root, rb_swblock_scancmp, 2147 swp_pager_meta_free_callback, &info); 2148 } 2149 2150 /* 2151 * The caller must hold vm_token. 2152 */ 2153 static 2154 int 2155 swp_pager_meta_free_callback(struct swblock *swap, void *data) 2156 { 2157 struct swfreeinfo *info = data; 2158 vm_object_t object = info->object; 2159 int index; 2160 int eindex; 2161 2162 /* 2163 * Figure out the range within the swblock. The wider scan may 2164 * return edge-case swap blocks when the start and/or end points 2165 * are in the middle of a block. 2166 */ 2167 if (swap->swb_index < info->begi) 2168 index = (int)info->begi & SWAP_META_MASK; 2169 else 2170 index = 0; 2171 2172 if (swap->swb_index + SWAP_META_PAGES > info->endi) 2173 eindex = (int)info->endi & SWAP_META_MASK; 2174 else 2175 eindex = SWAP_META_MASK; 2176 2177 /* 2178 * Scan and free the blocks. The loop terminates early 2179 * if (swap) runs out of blocks and could be freed. 2180 */ 2181 while (index <= eindex) { 2182 swblk_t v = swap->swb_pages[index]; 2183 2184 if (v != SWAPBLK_NONE) { 2185 swp_pager_freeswapspace(object, v, 1); 2186 swap->swb_pages[index] = SWAPBLK_NONE; 2187 if (--swap->swb_count == 0) { 2188 swp_pager_remove(object, swap); 2189 zfree(swap_zone, swap); 2190 --object->swblock_count; 2191 break; 2192 } 2193 } 2194 ++index; 2195 } 2196 /* swap may be invalid here due to zfree above */ 2197 return(0); 2198 } 2199 2200 /* 2201 * SWP_PAGER_META_FREE_ALL() - destroy all swap metadata associated with object 2202 * 2203 * This routine locates and destroys all swap metadata associated with 2204 * an object. 2205 * 2206 * The caller must hold vm_token. 2207 */ 2208 static void 2209 swp_pager_meta_free_all(vm_object_t object) 2210 { 2211 struct swblock *swap; 2212 int i; 2213 2214 while ((swap = RB_ROOT(&object->swblock_root)) != NULL) { 2215 swp_pager_remove(object, swap); 2216 for (i = 0; i < SWAP_META_PAGES; ++i) { 2217 swblk_t v = swap->swb_pages[i]; 2218 if (v != SWAPBLK_NONE) { 2219 --swap->swb_count; 2220 swp_pager_freeswapspace(object, v, 1); 2221 } 2222 } 2223 if (swap->swb_count != 0) 2224 panic("swap_pager_meta_free_all: swb_count != 0"); 2225 zfree(swap_zone, swap); 2226 --object->swblock_count; 2227 } 2228 KKASSERT(object->swblock_count == 0); 2229 } 2230 2231 /* 2232 * SWP_PAGER_METACTL() - misc control of swap and vm_page_t meta data. 2233 * 2234 * This routine is capable of looking up, popping, or freeing 2235 * swapblk assignments in the swap meta data or in the vm_page_t. 2236 * The routine typically returns the swapblk being looked-up, or popped, 2237 * or SWAPBLK_NONE if the block was freed, or SWAPBLK_NONE if the block 2238 * was invalid. This routine will automatically free any invalid 2239 * meta-data swapblks. 2240 * 2241 * It is not possible to store invalid swapblks in the swap meta data 2242 * (other then a literal 'SWAPBLK_NONE'), so we don't bother checking. 2243 * 2244 * When acting on a busy resident page and paging is in progress, we 2245 * have to wait until paging is complete but otherwise can act on the 2246 * busy page. 2247 * 2248 * SWM_FREE remove and free swap block from metadata 2249 * SWM_POP remove from meta data but do not free.. pop it out 2250 * 2251 * The caller must hold vm_token. 2252 */ 2253 static swblk_t 2254 swp_pager_meta_ctl(vm_object_t object, vm_pindex_t index, int flags) 2255 { 2256 struct swblock *swap; 2257 swblk_t r1; 2258 2259 if (object->swblock_count == 0) 2260 return(SWAPBLK_NONE); 2261 2262 r1 = SWAPBLK_NONE; 2263 swap = swp_pager_lookup(object, index); 2264 2265 if (swap != NULL) { 2266 index &= SWAP_META_MASK; 2267 r1 = swap->swb_pages[index]; 2268 2269 if (r1 != SWAPBLK_NONE) { 2270 if (flags & SWM_FREE) { 2271 swp_pager_freeswapspace(object, r1, 1); 2272 r1 = SWAPBLK_NONE; 2273 } 2274 if (flags & (SWM_FREE|SWM_POP)) { 2275 swap->swb_pages[index] = SWAPBLK_NONE; 2276 if (--swap->swb_count == 0) { 2277 swp_pager_remove(object, swap); 2278 zfree(swap_zone, swap); 2279 --object->swblock_count; 2280 } 2281 } 2282 } 2283 } 2284 return(r1); 2285 } 2286