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