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. Neither the name of the University nor the names of its contributors 54 * may be used to endorse or promote products derived from this software 55 * without specific prior written permission. 56 * 57 * THIS SOFTWARE IS PROVIDED BY THE REGENTS AND CONTRIBUTORS ``AS IS'' AND 58 * ANY EXPRESS OR IMPLIED WARRANTIES, INCLUDING, BUT NOT LIMITED TO, THE 59 * IMPLIED WARRANTIES OF MERCHANTABILITY AND FITNESS FOR A PARTICULAR PURPOSE 60 * ARE DISCLAIMED. IN NO EVENT SHALL THE REGENTS OR CONTRIBUTORS BE LIABLE 61 * FOR ANY DIRECT, INDIRECT, INCIDENTAL, SPECIAL, EXEMPLARY, OR CONSEQUENTIAL 62 * DAMAGES (INCLUDING, BUT NOT LIMITED TO, PROCUREMENT OF SUBSTITUTE GOODS 63 * OR SERVICES; LOSS OF USE, DATA, OR PROFITS; OR BUSINESS INTERRUPTION) 64 * HOWEVER CAUSED AND ON ANY THEORY OF LIABILITY, WHETHER IN CONTRACT, STRICT 65 * LIABILITY, OR TORT (INCLUDING NEGLIGENCE OR OTHERWISE) ARISING IN ANY WAY 66 * OUT OF THE USE OF THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF 67 * SUCH DAMAGE. 68 * 69 * New Swap System 70 * Matthew Dillon 71 * 72 * Radix Bitmap 'blists'. 73 * 74 * - The new swapper uses the new radix bitmap code. This should scale 75 * to arbitrarily small or arbitrarily large swap spaces and an almost 76 * arbitrary degree of fragmentation. 77 * 78 * Features: 79 * 80 * - on the fly reallocation of swap during putpages. The new system 81 * does not try to keep previously allocated swap blocks for dirty 82 * pages. 83 * 84 * - on the fly deallocation of swap 85 * 86 * - No more garbage collection required. Unnecessarily allocated swap 87 * blocks only exist for dirty vm_page_t's now and these are already 88 * cycled (in a high-load system) by the pager. We also do on-the-fly 89 * removal of invalidated swap blocks when a page is destroyed 90 * or renamed. 91 * 92 * from: Utah $Hdr: swap_pager.c 1.4 91/04/30$ 93 * @(#)swap_pager.c 8.9 (Berkeley) 3/21/94 94 * $FreeBSD: src/sys/vm/swap_pager.c,v 1.130.2.12 2002/08/31 21:15:55 dillon Exp $ 95 */ 96 97 #include <sys/param.h> 98 #include <sys/systm.h> 99 #include <sys/conf.h> 100 #include <sys/kernel.h> 101 #include <sys/proc.h> 102 #include <sys/buf.h> 103 #include <sys/vnode.h> 104 #include <sys/malloc.h> 105 #include <sys/vmmeter.h> 106 #include <sys/sysctl.h> 107 #include <sys/blist.h> 108 #include <sys/lock.h> 109 #include <sys/thread2.h> 110 111 #include "opt_swap.h" 112 #include <vm/vm.h> 113 #include <vm/vm_object.h> 114 #include <vm/vm_page.h> 115 #include <vm/vm_pager.h> 116 #include <vm/vm_pageout.h> 117 #include <vm/swap_pager.h> 118 #include <vm/vm_extern.h> 119 #include <vm/vm_zone.h> 120 #include <vm/vnode_pager.h> 121 122 #include <sys/buf2.h> 123 #include <vm/vm_page2.h> 124 125 #ifndef MAX_PAGEOUT_CLUSTER 126 #define MAX_PAGEOUT_CLUSTER SWB_NPAGES 127 #endif 128 129 #define SWM_FREE 0x02 /* free, period */ 130 #define SWM_POP 0x04 /* pop out */ 131 132 #define SWBIO_READ 0x01 133 #define SWBIO_WRITE 0x02 134 #define SWBIO_SYNC 0x04 135 136 struct swfreeinfo { 137 vm_object_t object; 138 vm_pindex_t basei; 139 vm_pindex_t begi; 140 vm_pindex_t endi; /* inclusive */ 141 }; 142 143 struct swswapoffinfo { 144 vm_object_t object; 145 int devidx; 146 int shared; 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 swap_fail_ticks; /* when we became exhausted */ 156 int swap_pager_almost_full; /* swap space exhaustion (w/ hysteresis)*/ 157 int vm_swap_cache_use; 158 int vm_swap_anon_use; 159 static int vm_report_swap_allocs; 160 161 static int nsw_rcount; /* free read buffers */ 162 static int nsw_wcount_sync; /* limit write buffers / synchronous */ 163 static int nsw_wcount_async; /* limit write buffers / asynchronous */ 164 static int nsw_wcount_async_max;/* assigned maximum */ 165 static int nsw_cluster_max; /* maximum VOP I/O allowed */ 166 167 struct blist *swapblist; 168 static int swap_async_max = 4; /* maximum in-progress async I/O's */ 169 static int swap_burst_read = 0; /* allow burst reading */ 170 static swblk_t swapiterator; /* linearize allocations */ 171 172 static struct spinlock swapbp_spin = SPINLOCK_INITIALIZER(&swapbp_spin, "swapbp_spin"); 173 174 /* from vm_swap.c */ 175 extern struct vnode *swapdev_vp; 176 extern struct swdevt *swdevt; 177 extern int nswdev; 178 179 #define BLK2DEVIDX(blk) (nswdev > 1 ? blk / dmmax % nswdev : 0) 180 181 SYSCTL_INT(_vm, OID_AUTO, swap_async_max, 182 CTLFLAG_RW, &swap_async_max, 0, "Maximum running async swap ops"); 183 SYSCTL_INT(_vm, OID_AUTO, swap_burst_read, 184 CTLFLAG_RW, &swap_burst_read, 0, "Allow burst reads for pageins"); 185 186 SYSCTL_INT(_vm, OID_AUTO, swap_cache_use, 187 CTLFLAG_RD, &vm_swap_cache_use, 0, ""); 188 SYSCTL_INT(_vm, OID_AUTO, swap_anon_use, 189 CTLFLAG_RD, &vm_swap_anon_use, 0, ""); 190 SYSCTL_INT(_vm, OID_AUTO, swap_size, 191 CTLFLAG_RD, &vm_swap_size, 0, ""); 192 SYSCTL_INT(_vm, OID_AUTO, report_swap_allocs, 193 CTLFLAG_RW, &vm_report_swap_allocs, 0, ""); 194 195 vm_zone_t swap_zone; 196 197 /* 198 * Red-Black tree for swblock entries 199 * 200 * The caller must hold vm_token 201 */ 202 RB_GENERATE2(swblock_rb_tree, swblock, swb_entry, rb_swblock_compare, 203 vm_pindex_t, swb_index); 204 205 int 206 rb_swblock_compare(struct swblock *swb1, struct swblock *swb2) 207 { 208 if (swb1->swb_index < swb2->swb_index) 209 return(-1); 210 if (swb1->swb_index > swb2->swb_index) 211 return(1); 212 return(0); 213 } 214 215 static 216 int 217 rb_swblock_scancmp(struct swblock *swb, void *data) 218 { 219 struct swfreeinfo *info = data; 220 221 if (swb->swb_index < info->basei) 222 return(-1); 223 if (swb->swb_index > info->endi) 224 return(1); 225 return(0); 226 } 227 228 static 229 int 230 rb_swblock_condcmp(struct swblock *swb, void *data) 231 { 232 struct swfreeinfo *info = data; 233 234 if (swb->swb_index < info->basei) 235 return(-1); 236 return(0); 237 } 238 239 /* 240 * pagerops for OBJT_SWAP - "swap pager". Some ops are also global procedure 241 * calls hooked from other parts of the VM system and do not appear here. 242 * (see vm/swap_pager.h). 243 */ 244 245 static void swap_pager_dealloc (vm_object_t object); 246 static int swap_pager_getpage (vm_object_t, vm_page_t *, int); 247 static void swap_chain_iodone(struct bio *biox); 248 249 struct pagerops swappagerops = { 250 swap_pager_dealloc, /* deallocate an OBJT_SWAP object */ 251 swap_pager_getpage, /* pagein */ 252 swap_pager_putpages, /* pageout */ 253 swap_pager_haspage /* get backing store status for page */ 254 }; 255 256 /* 257 * dmmax is in page-sized chunks with the new swap system. It was 258 * dev-bsized chunks in the old. dmmax is always a power of 2. 259 * 260 * swap_*() routines are externally accessible. swp_*() routines are 261 * internal. 262 */ 263 264 int dmmax; 265 static int dmmax_mask; 266 int nswap_lowat = 128; /* in pages, swap_pager_almost_full warn */ 267 int nswap_hiwat = 512; /* in pages, swap_pager_almost_full warn */ 268 269 static __inline void swp_sizecheck (void); 270 static void swp_pager_async_iodone (struct bio *bio); 271 272 /* 273 * Swap bitmap functions 274 */ 275 276 static __inline void swp_pager_freeswapspace(vm_object_t object, 277 swblk_t blk, int npages); 278 static __inline swblk_t swp_pager_getswapspace(vm_object_t object, int npages); 279 280 /* 281 * Metadata functions 282 */ 283 284 static void swp_pager_meta_convert(vm_object_t); 285 static void swp_pager_meta_build(vm_object_t, vm_pindex_t, swblk_t); 286 static void swp_pager_meta_free(vm_object_t, vm_pindex_t, vm_pindex_t); 287 static void swp_pager_meta_free_all(vm_object_t); 288 static swblk_t swp_pager_meta_ctl(vm_object_t, vm_pindex_t, int); 289 290 /* 291 * SWP_SIZECHECK() - update swap_pager_full indication 292 * 293 * update the swap_pager_almost_full indication and warn when we are 294 * about to run out of swap space, using lowat/hiwat hysteresis. 295 * 296 * Clear swap_pager_full ( task killing ) indication when lowat is met. 297 * 298 * No restrictions on call 299 * This routine may not block. 300 * SMP races are ok. 301 */ 302 static __inline void 303 swp_sizecheck(void) 304 { 305 if (vm_swap_size < nswap_lowat) { 306 if (swap_pager_almost_full == 0) { 307 kprintf("swap_pager: out of swap space\n"); 308 swap_pager_almost_full = 1; 309 swap_fail_ticks = ticks; 310 } 311 } else { 312 swap_pager_full = 0; 313 if (vm_swap_size > nswap_hiwat) 314 swap_pager_almost_full = 0; 315 } 316 } 317 318 /* 319 * SWAP_PAGER_INIT() - initialize the swap pager! 320 * 321 * Expected to be started from system init. NOTE: This code is run 322 * before much else so be careful what you depend on. Most of the VM 323 * system has yet to be initialized at this point. 324 * 325 * Called from the low level boot code only. 326 */ 327 static void 328 swap_pager_init(void *arg __unused) 329 { 330 /* 331 * Device Stripe, in PAGE_SIZE'd blocks 332 */ 333 dmmax = SWB_NPAGES * 2; 334 dmmax_mask = ~(dmmax - 1); 335 } 336 SYSINIT(vm_mem, SI_BOOT1_VM, SI_ORDER_THIRD, swap_pager_init, NULL); 337 338 /* 339 * SWAP_PAGER_SWAP_INIT() - swap pager initialization from pageout process 340 * 341 * Expected to be started from pageout process once, prior to entering 342 * its main loop. 343 * 344 * Called from the low level boot code only. 345 */ 346 void 347 swap_pager_swap_init(void) 348 { 349 int n, n2; 350 351 /* 352 * Number of in-transit swap bp operations. Don't 353 * exhaust the pbufs completely. Make sure we 354 * initialize workable values (0 will work for hysteresis 355 * but it isn't very efficient). 356 * 357 * The nsw_cluster_max is constrained by the number of pages an XIO 358 * holds, i.e., (MAXPHYS/PAGE_SIZE) and our locally defined 359 * MAX_PAGEOUT_CLUSTER. Also be aware that swap ops are 360 * constrained by the swap device interleave stripe size. 361 * 362 * Currently we hardwire nsw_wcount_async to 4. This limit is 363 * designed to prevent other I/O from having high latencies due to 364 * our pageout I/O. The value 4 works well for one or two active swap 365 * devices but is probably a little low if you have more. Even so, 366 * a higher value would probably generate only a limited improvement 367 * with three or four active swap devices since the system does not 368 * typically have to pageout at extreme bandwidths. We will want 369 * at least 2 per swap devices, and 4 is a pretty good value if you 370 * have one NFS swap device due to the command/ack latency over NFS. 371 * So it all works out pretty well. 372 */ 373 374 nsw_cluster_max = min((MAXPHYS/PAGE_SIZE), MAX_PAGEOUT_CLUSTER); 375 376 nsw_rcount = (nswbuf_kva + 1) / 2; 377 nsw_wcount_sync = (nswbuf_kva + 3) / 4; 378 nsw_wcount_async = 4; 379 nsw_wcount_async_max = nsw_wcount_async; 380 381 /* 382 * The zone is dynamically allocated so generally size it to 383 * maxswzone (32MB to 512MB of KVM). Set a minimum size based 384 * on physical memory of around 8x (each swblock can hold 16 pages). 385 * 386 * With the advent of SSDs (vs HDs) the practical (swap:memory) ratio 387 * has increased dramatically. 388 */ 389 n = vmstats.v_page_count / 2; 390 if (maxswzone && n < maxswzone / sizeof(struct swblock)) 391 n = maxswzone / sizeof(struct swblock); 392 n2 = n; 393 394 do { 395 swap_zone = zinit( 396 "SWAPMETA", 397 sizeof(struct swblock), 398 n, 399 ZONE_INTERRUPT, 400 1); 401 if (swap_zone != NULL) 402 break; 403 /* 404 * if the allocation failed, try a zone two thirds the 405 * size of the previous attempt. 406 */ 407 n -= ((n + 2) / 3); 408 } while (n > 0); 409 410 if (swap_zone == NULL) 411 panic("swap_pager_swap_init: swap_zone == NULL"); 412 if (n2 != n) 413 kprintf("Swap zone entries reduced from %d to %d.\n", n2, n); 414 } 415 416 /* 417 * SWAP_PAGER_ALLOC() - allocate a new OBJT_SWAP VM object and instantiate 418 * its metadata structures. 419 * 420 * This routine is called from the mmap and fork code to create a new 421 * OBJT_SWAP object. We do this by creating an OBJT_DEFAULT object 422 * and then converting it with swp_pager_meta_convert(). 423 * 424 * We only support unnamed objects. 425 * 426 * No restrictions. 427 */ 428 vm_object_t 429 swap_pager_alloc(void *handle, off_t size, vm_prot_t prot, off_t offset) 430 { 431 vm_object_t object; 432 433 KKASSERT(handle == NULL); 434 object = vm_object_allocate_hold(OBJT_DEFAULT, 435 OFF_TO_IDX(offset + PAGE_MASK + size)); 436 swp_pager_meta_convert(object); 437 vm_object_drop(object); 438 439 return (object); 440 } 441 442 /* 443 * SWAP_PAGER_DEALLOC() - remove swap metadata from object 444 * 445 * The swap backing for the object is destroyed. The code is 446 * designed such that we can reinstantiate it later, but this 447 * routine is typically called only when the entire object is 448 * about to be destroyed. 449 * 450 * The object must be locked or unreferenceable. 451 * No other requirements. 452 */ 453 static void 454 swap_pager_dealloc(vm_object_t object) 455 { 456 vm_object_hold(object); 457 vm_object_pip_wait(object, "swpdea"); 458 459 /* 460 * Free all remaining metadata. We only bother to free it from 461 * the swap meta data. We do not attempt to free swapblk's still 462 * associated with vm_page_t's for this object. We do not care 463 * if paging is still in progress on some objects. 464 */ 465 swp_pager_meta_free_all(object); 466 vm_object_drop(object); 467 } 468 469 /************************************************************************ 470 * SWAP PAGER BITMAP ROUTINES * 471 ************************************************************************/ 472 473 /* 474 * SWP_PAGER_GETSWAPSPACE() - allocate raw swap space 475 * 476 * Allocate swap for the requested number of pages. The starting 477 * swap block number (a page index) is returned or SWAPBLK_NONE 478 * if the allocation failed. 479 * 480 * Also has the side effect of advising that somebody made a mistake 481 * when they configured swap and didn't configure enough. 482 * 483 * The caller must hold the object. 484 * This routine may not block. 485 */ 486 static __inline swblk_t 487 swp_pager_getswapspace(vm_object_t object, int npages) 488 { 489 swblk_t blk; 490 491 lwkt_gettoken(&vm_token); 492 blk = blist_allocat(swapblist, npages, swapiterator); 493 if (blk == SWAPBLK_NONE) 494 blk = blist_allocat(swapblist, npages, 0); 495 if (blk == SWAPBLK_NONE) { 496 if (swap_pager_full != 2) { 497 if (vm_swap_size == 0) 498 kprintf("Warning: The system would like to " 499 "page to swap but no swap space " 500 "is configured!\n"); 501 else 502 kprintf("swap_pager_getswapspace: " 503 "unable to allocate=%d pages\n", 504 npages); 505 swap_pager_full = 2; 506 if (swap_pager_almost_full == 0) 507 swap_fail_ticks = ticks; 508 swap_pager_almost_full = 1; 509 } 510 } else { 511 /* swapiterator = blk; disable for now, doesn't work well */ 512 swapacctspace(blk, -npages); 513 if (object->type == OBJT_SWAP) 514 vm_swap_anon_use += npages; 515 else 516 vm_swap_cache_use += npages; 517 swp_sizecheck(); 518 } 519 lwkt_reltoken(&vm_token); 520 return(blk); 521 } 522 523 /* 524 * SWP_PAGER_FREESWAPSPACE() - free raw swap space 525 * 526 * This routine returns the specified swap blocks back to the bitmap. 527 * 528 * Note: This routine may not block (it could in the old swap code), 529 * and through the use of the new blist routines it does not block. 530 * 531 * We must be called at splvm() to avoid races with bitmap frees from 532 * vm_page_remove() aka swap_pager_page_removed(). 533 * 534 * This routine may not block. 535 */ 536 537 static __inline void 538 swp_pager_freeswapspace(vm_object_t object, swblk_t blk, int npages) 539 { 540 struct swdevt *sp = &swdevt[BLK2DEVIDX(blk)]; 541 542 lwkt_gettoken(&vm_token); 543 sp->sw_nused -= npages; 544 if (object->type == OBJT_SWAP) 545 vm_swap_anon_use -= npages; 546 else 547 vm_swap_cache_use -= npages; 548 549 if (sp->sw_flags & SW_CLOSING) { 550 lwkt_reltoken(&vm_token); 551 return; 552 } 553 554 blist_free(swapblist, blk, npages); 555 vm_swap_size += npages; 556 swp_sizecheck(); 557 lwkt_reltoken(&vm_token); 558 } 559 560 /* 561 * SWAP_PAGER_FREESPACE() - frees swap blocks associated with a page 562 * range within an object. 563 * 564 * This is a globally accessible routine. 565 * 566 * This routine removes swapblk assignments from swap metadata. 567 * 568 * The external callers of this routine typically have already destroyed 569 * or renamed vm_page_t's associated with this range in the object so 570 * we should be ok. 571 * 572 * No requirements. 573 */ 574 void 575 swap_pager_freespace(vm_object_t object, vm_pindex_t start, vm_pindex_t size) 576 { 577 vm_object_hold(object); 578 swp_pager_meta_free(object, start, size); 579 vm_object_drop(object); 580 } 581 582 /* 583 * No requirements. 584 */ 585 void 586 swap_pager_freespace_all(vm_object_t object) 587 { 588 vm_object_hold(object); 589 swp_pager_meta_free_all(object); 590 vm_object_drop(object); 591 } 592 593 /* 594 * This function conditionally frees swap cache swap starting at 595 * (*basei) in the object. (count) swap blocks will be nominally freed. 596 * The actual number of blocks freed can be more or less than the 597 * requested number. 598 * 599 * This function nominally returns the number of blocks freed. However, 600 * the actual number of blocks freed may be less then the returned value. 601 * If the function is unable to exhaust the object or if it is able to 602 * free (approximately) the requested number of blocks it returns 603 * a value n > count. 604 * 605 * If we exhaust the object we will return a value n <= count. 606 * 607 * The caller must hold the object. 608 * 609 * WARNING! If count == 0 then -1 can be returned as a degenerate case, 610 * callers should always pass a count value > 0. 611 */ 612 static int swap_pager_condfree_callback(struct swblock *swap, void *data); 613 614 int 615 swap_pager_condfree(vm_object_t object, vm_pindex_t *basei, int count) 616 { 617 struct swfreeinfo info; 618 int n; 619 int t; 620 621 ASSERT_LWKT_TOKEN_HELD(vm_object_token(object)); 622 623 info.object = object; 624 info.basei = *basei; /* skip up to this page index */ 625 info.begi = count; /* max swap pages to destroy */ 626 info.endi = count * 8; /* max swblocks to scan */ 627 628 swblock_rb_tree_RB_SCAN(&object->swblock_root, rb_swblock_condcmp, 629 swap_pager_condfree_callback, &info); 630 *basei = info.basei; 631 632 /* 633 * Take the higher difference swblocks vs pages 634 */ 635 n = count - (int)info.begi; 636 t = count * 8 - (int)info.endi; 637 if (n < t) 638 n = t; 639 if (n < 1) 640 n = 1; 641 return(n); 642 } 643 644 /* 645 * The idea is to free whole meta-block to avoid fragmenting 646 * the swap space or disk I/O. We only do this if NO VM pages 647 * are present. 648 * 649 * We do not have to deal with clearing PG_SWAPPED in related VM 650 * pages because there are no related VM pages. 651 * 652 * The caller must hold the object. 653 */ 654 static int 655 swap_pager_condfree_callback(struct swblock *swap, void *data) 656 { 657 struct swfreeinfo *info = data; 658 vm_object_t object = info->object; 659 int i; 660 661 for (i = 0; i < SWAP_META_PAGES; ++i) { 662 if (vm_page_lookup(object, swap->swb_index + i)) 663 break; 664 } 665 info->basei = swap->swb_index + SWAP_META_PAGES; 666 if (i == SWAP_META_PAGES) { 667 info->begi -= swap->swb_count; 668 swap_pager_freespace(object, swap->swb_index, SWAP_META_PAGES); 669 } 670 --info->endi; 671 if ((int)info->begi < 0 || (int)info->endi < 0) 672 return(-1); 673 lwkt_yield(); 674 return(0); 675 } 676 677 /* 678 * Called by vm_page_alloc() when a new VM page is inserted 679 * into a VM object. Checks whether swap has been assigned to 680 * the page and sets PG_SWAPPED as necessary. 681 * 682 * No requirements. 683 */ 684 void 685 swap_pager_page_inserted(vm_page_t m) 686 { 687 if (m->object->swblock_count) { 688 vm_object_hold(m->object); 689 if (swp_pager_meta_ctl(m->object, m->pindex, 0) != SWAPBLK_NONE) 690 vm_page_flag_set(m, PG_SWAPPED); 691 vm_object_drop(m->object); 692 } 693 } 694 695 /* 696 * SWAP_PAGER_RESERVE() - reserve swap blocks in object 697 * 698 * Assigns swap blocks to the specified range within the object. The 699 * swap blocks are not zerod. Any previous swap assignment is destroyed. 700 * 701 * Returns 0 on success, -1 on failure. 702 * 703 * The caller is responsible for avoiding races in the specified range. 704 * No other requirements. 705 */ 706 int 707 swap_pager_reserve(vm_object_t object, vm_pindex_t start, vm_size_t size) 708 { 709 int n = 0; 710 swblk_t blk = SWAPBLK_NONE; 711 vm_pindex_t beg = start; /* save start index */ 712 713 vm_object_hold(object); 714 715 while (size) { 716 if (n == 0) { 717 n = BLIST_MAX_ALLOC; 718 while ((blk = swp_pager_getswapspace(object, n)) == 719 SWAPBLK_NONE) 720 { 721 n >>= 1; 722 if (n == 0) { 723 swp_pager_meta_free(object, beg, 724 start - beg); 725 vm_object_drop(object); 726 return(-1); 727 } 728 } 729 } 730 swp_pager_meta_build(object, start, blk); 731 --size; 732 ++start; 733 ++blk; 734 --n; 735 } 736 swp_pager_meta_free(object, start, n); 737 vm_object_drop(object); 738 return(0); 739 } 740 741 /* 742 * SWAP_PAGER_COPY() - copy blocks from source pager to destination pager 743 * and destroy the source. 744 * 745 * Copy any valid swapblks from the source to the destination. In 746 * cases where both the source and destination have a valid swapblk, 747 * we keep the destination's. 748 * 749 * This routine is allowed to block. It may block allocating metadata 750 * indirectly through swp_pager_meta_build() or if paging is still in 751 * progress on the source. 752 * 753 * XXX vm_page_collapse() kinda expects us not to block because we 754 * supposedly do not need to allocate memory, but for the moment we 755 * *may* have to get a little memory from the zone allocator, but 756 * it is taken from the interrupt memory. We should be ok. 757 * 758 * The source object contains no vm_page_t's (which is just as well) 759 * The source object is of type OBJT_SWAP. 760 * 761 * The source and destination objects must be held by the caller. 762 */ 763 void 764 swap_pager_copy(vm_object_t srcobject, vm_object_t dstobject, 765 vm_pindex_t base_index, int destroysource) 766 { 767 vm_pindex_t i; 768 769 ASSERT_LWKT_TOKEN_HELD(vm_object_token(srcobject)); 770 ASSERT_LWKT_TOKEN_HELD(vm_object_token(dstobject)); 771 772 /* 773 * transfer source to destination. 774 */ 775 for (i = 0; i < dstobject->size; ++i) { 776 swblk_t dstaddr; 777 778 /* 779 * Locate (without changing) the swapblk on the destination, 780 * unless it is invalid in which case free it silently, or 781 * if the destination is a resident page, in which case the 782 * source is thrown away. 783 */ 784 dstaddr = swp_pager_meta_ctl(dstobject, i, 0); 785 786 if (dstaddr == SWAPBLK_NONE) { 787 /* 788 * Destination has no swapblk and is not resident, 789 * copy source. 790 */ 791 swblk_t srcaddr; 792 793 srcaddr = swp_pager_meta_ctl(srcobject, 794 base_index + i, SWM_POP); 795 796 if (srcaddr != SWAPBLK_NONE) 797 swp_pager_meta_build(dstobject, i, srcaddr); 798 } else { 799 /* 800 * Destination has valid swapblk or it is represented 801 * by a resident page. We destroy the sourceblock. 802 */ 803 swp_pager_meta_ctl(srcobject, base_index + i, SWM_FREE); 804 } 805 } 806 807 /* 808 * Free left over swap blocks in source. 809 * 810 * We have to revert the type to OBJT_DEFAULT so we do not accidently 811 * double-remove the object from the swap queues. 812 */ 813 if (destroysource) { 814 /* 815 * Reverting the type is not necessary, the caller is going 816 * to destroy srcobject directly, but I'm doing it here 817 * for consistency since we've removed the object from its 818 * queues. 819 */ 820 swp_pager_meta_free_all(srcobject); 821 if (srcobject->type == OBJT_SWAP) 822 srcobject->type = OBJT_DEFAULT; 823 } 824 } 825 826 /* 827 * SWAP_PAGER_HASPAGE() - determine if we have good backing store for 828 * the requested page. 829 * 830 * We determine whether good backing store exists for the requested 831 * page and return TRUE if it does, FALSE if it doesn't. 832 * 833 * If TRUE, we also try to determine how much valid, contiguous backing 834 * store exists before and after the requested page within a reasonable 835 * distance. We do not try to restrict it to the swap device stripe 836 * (that is handled in getpages/putpages). It probably isn't worth 837 * doing here. 838 * 839 * No requirements. 840 */ 841 boolean_t 842 swap_pager_haspage(vm_object_t object, vm_pindex_t pindex) 843 { 844 swblk_t blk0; 845 846 /* 847 * do we have good backing store at the requested index ? 848 */ 849 vm_object_hold(object); 850 blk0 = swp_pager_meta_ctl(object, pindex, 0); 851 852 if (blk0 == SWAPBLK_NONE) { 853 vm_object_drop(object); 854 return (FALSE); 855 } 856 vm_object_drop(object); 857 return (TRUE); 858 } 859 860 /* 861 * SWAP_PAGER_PAGE_UNSWAPPED() - remove swap backing store related to page 862 * 863 * This removes any associated swap backing store, whether valid or 864 * not, from the page. This operates on any VM object, not just OBJT_SWAP 865 * objects. 866 * 867 * This routine is typically called when a page is made dirty, at 868 * which point any associated swap can be freed. MADV_FREE also 869 * calls us in a special-case situation 870 * 871 * NOTE!!! If the page is clean and the swap was valid, the caller 872 * should make the page dirty before calling this routine. This routine 873 * does NOT change the m->dirty status of the page. Also: MADV_FREE 874 * depends on it. 875 * 876 * The page must be busied or soft-busied. 877 * The caller can hold the object to avoid blocking, else we might block. 878 * No other requirements. 879 */ 880 void 881 swap_pager_unswapped(vm_page_t m) 882 { 883 if (m->flags & PG_SWAPPED) { 884 vm_object_hold(m->object); 885 KKASSERT(m->flags & PG_SWAPPED); 886 swp_pager_meta_ctl(m->object, m->pindex, SWM_FREE); 887 vm_page_flag_clear(m, PG_SWAPPED); 888 vm_object_drop(m->object); 889 } 890 } 891 892 /* 893 * SWAP_PAGER_STRATEGY() - read, write, free blocks 894 * 895 * This implements a VM OBJECT strategy function using swap backing store. 896 * This can operate on any VM OBJECT type, not necessarily just OBJT_SWAP 897 * types. 898 * 899 * This is intended to be a cacheless interface (i.e. caching occurs at 900 * higher levels), and is also used as a swap-based SSD cache for vnode 901 * and device objects. 902 * 903 * All I/O goes directly to and from the swap device. 904 * 905 * We currently attempt to run I/O synchronously or asynchronously as 906 * the caller requests. This isn't perfect because we loose error 907 * sequencing when we run multiple ops in parallel to satisfy a request. 908 * But this is swap, so we let it all hang out. 909 * 910 * No requirements. 911 */ 912 void 913 swap_pager_strategy(vm_object_t object, struct bio *bio) 914 { 915 struct buf *bp = bio->bio_buf; 916 struct bio *nbio; 917 vm_pindex_t start; 918 vm_pindex_t biox_blkno = 0; 919 int count; 920 char *data; 921 struct bio *biox; 922 struct buf *bufx; 923 #if 0 924 struct bio_track *track; 925 #endif 926 927 #if 0 928 /* 929 * tracking for swapdev vnode I/Os 930 */ 931 if (bp->b_cmd == BUF_CMD_READ) 932 track = &swapdev_vp->v_track_read; 933 else 934 track = &swapdev_vp->v_track_write; 935 #endif 936 937 if (bp->b_bcount & PAGE_MASK) { 938 bp->b_error = EINVAL; 939 bp->b_flags |= B_ERROR | B_INVAL; 940 biodone(bio); 941 kprintf("swap_pager_strategy: bp %p offset %lld size %d, " 942 "not page bounded\n", 943 bp, (long long)bio->bio_offset, (int)bp->b_bcount); 944 return; 945 } 946 947 /* 948 * Clear error indication, initialize page index, count, data pointer. 949 */ 950 bp->b_error = 0; 951 bp->b_flags &= ~B_ERROR; 952 bp->b_resid = bp->b_bcount; 953 954 start = (vm_pindex_t)(bio->bio_offset >> PAGE_SHIFT); 955 count = howmany(bp->b_bcount, PAGE_SIZE); 956 data = bp->b_data; 957 958 /* 959 * Deal with BUF_CMD_FREEBLKS 960 */ 961 if (bp->b_cmd == BUF_CMD_FREEBLKS) { 962 /* 963 * FREE PAGE(s) - destroy underlying swap that is no longer 964 * needed. 965 */ 966 vm_object_hold(object); 967 swp_pager_meta_free(object, start, count); 968 vm_object_drop(object); 969 bp->b_resid = 0; 970 biodone(bio); 971 return; 972 } 973 974 /* 975 * We need to be able to create a new cluster of I/O's. We cannot 976 * use the caller fields of the passed bio so push a new one. 977 * 978 * Because nbio is just a placeholder for the cluster links, 979 * we can biodone() the original bio instead of nbio to make 980 * things a bit more efficient. 981 */ 982 nbio = push_bio(bio); 983 nbio->bio_offset = bio->bio_offset; 984 nbio->bio_caller_info1.cluster_head = NULL; 985 nbio->bio_caller_info2.cluster_tail = NULL; 986 987 biox = NULL; 988 bufx = NULL; 989 990 /* 991 * Execute read or write 992 */ 993 vm_object_hold(object); 994 995 while (count > 0) { 996 swblk_t blk; 997 998 /* 999 * Obtain block. If block not found and writing, allocate a 1000 * new block and build it into the object. 1001 */ 1002 blk = swp_pager_meta_ctl(object, start, 0); 1003 if ((blk == SWAPBLK_NONE) && bp->b_cmd != BUF_CMD_READ) { 1004 blk = swp_pager_getswapspace(object, 1); 1005 if (blk == SWAPBLK_NONE) { 1006 bp->b_error = ENOMEM; 1007 bp->b_flags |= B_ERROR; 1008 break; 1009 } 1010 swp_pager_meta_build(object, start, blk); 1011 } 1012 1013 /* 1014 * Do we have to flush our current collection? Yes if: 1015 * 1016 * - no swap block at this index 1017 * - swap block is not contiguous 1018 * - we cross a physical disk boundry in the 1019 * stripe. 1020 */ 1021 if ( 1022 biox && (biox_blkno + btoc(bufx->b_bcount) != blk || 1023 ((biox_blkno ^ blk) & dmmax_mask) 1024 ) 1025 ) { 1026 if (bp->b_cmd == BUF_CMD_READ) { 1027 ++mycpu->gd_cnt.v_swapin; 1028 mycpu->gd_cnt.v_swappgsin += btoc(bufx->b_bcount); 1029 } else { 1030 ++mycpu->gd_cnt.v_swapout; 1031 mycpu->gd_cnt.v_swappgsout += btoc(bufx->b_bcount); 1032 bufx->b_dirtyend = bufx->b_bcount; 1033 } 1034 1035 /* 1036 * Finished with this buf. 1037 */ 1038 KKASSERT(bufx->b_bcount != 0); 1039 if (bufx->b_cmd != BUF_CMD_READ) 1040 bufx->b_dirtyend = bufx->b_bcount; 1041 biox = NULL; 1042 bufx = NULL; 1043 } 1044 1045 /* 1046 * Add new swapblk to biox, instantiating biox if necessary. 1047 * Zero-fill reads are able to take a shortcut. 1048 */ 1049 if (blk == SWAPBLK_NONE) { 1050 /* 1051 * We can only get here if we are reading. Since 1052 * we are at splvm() we can safely modify b_resid, 1053 * even if chain ops are in progress. 1054 */ 1055 bzero(data, PAGE_SIZE); 1056 bp->b_resid -= PAGE_SIZE; 1057 } else { 1058 if (biox == NULL) { 1059 /* XXX chain count > 4, wait to <= 4 */ 1060 1061 bufx = getpbuf(NULL); 1062 biox = &bufx->b_bio1; 1063 cluster_append(nbio, bufx); 1064 bufx->b_flags |= (bp->b_flags & B_ORDERED); 1065 bufx->b_cmd = bp->b_cmd; 1066 biox->bio_done = swap_chain_iodone; 1067 biox->bio_offset = (off_t)blk << PAGE_SHIFT; 1068 biox->bio_caller_info1.cluster_parent = nbio; 1069 biox_blkno = blk; 1070 bufx->b_bcount = 0; 1071 bufx->b_data = data; 1072 } 1073 bufx->b_bcount += PAGE_SIZE; 1074 } 1075 --count; 1076 ++start; 1077 data += PAGE_SIZE; 1078 } 1079 1080 vm_object_drop(object); 1081 1082 /* 1083 * Flush out last buffer 1084 */ 1085 if (biox) { 1086 if (bufx->b_cmd == BUF_CMD_READ) { 1087 ++mycpu->gd_cnt.v_swapin; 1088 mycpu->gd_cnt.v_swappgsin += btoc(bufx->b_bcount); 1089 } else { 1090 ++mycpu->gd_cnt.v_swapout; 1091 mycpu->gd_cnt.v_swappgsout += btoc(bufx->b_bcount); 1092 bufx->b_dirtyend = bufx->b_bcount; 1093 } 1094 KKASSERT(bufx->b_bcount); 1095 if (bufx->b_cmd != BUF_CMD_READ) 1096 bufx->b_dirtyend = bufx->b_bcount; 1097 /* biox, bufx = NULL */ 1098 } 1099 1100 /* 1101 * Now initiate all the I/O. Be careful looping on our chain as 1102 * I/O's may complete while we are still initiating them. 1103 * 1104 * If the request is a 100% sparse read no bios will be present 1105 * and we just biodone() the buffer. 1106 */ 1107 nbio->bio_caller_info2.cluster_tail = NULL; 1108 bufx = nbio->bio_caller_info1.cluster_head; 1109 1110 if (bufx) { 1111 while (bufx) { 1112 biox = &bufx->b_bio1; 1113 BUF_KERNPROC(bufx); 1114 bufx = bufx->b_cluster_next; 1115 vn_strategy(swapdev_vp, biox); 1116 } 1117 } else { 1118 biodone(bio); 1119 } 1120 1121 /* 1122 * Completion of the cluster will also call biodone_chain(nbio). 1123 * We never call biodone(nbio) so we don't have to worry about 1124 * setting up a bio_done callback. It's handled in the sub-IO. 1125 */ 1126 /**/ 1127 } 1128 1129 /* 1130 * biodone callback 1131 * 1132 * No requirements. 1133 */ 1134 static void 1135 swap_chain_iodone(struct bio *biox) 1136 { 1137 struct buf **nextp; 1138 struct buf *bufx; /* chained sub-buffer */ 1139 struct bio *nbio; /* parent nbio with chain glue */ 1140 struct buf *bp; /* original bp associated with nbio */ 1141 int chain_empty; 1142 1143 bufx = biox->bio_buf; 1144 nbio = biox->bio_caller_info1.cluster_parent; 1145 bp = nbio->bio_buf; 1146 1147 /* 1148 * Update the original buffer 1149 */ 1150 KKASSERT(bp != NULL); 1151 if (bufx->b_flags & B_ERROR) { 1152 atomic_set_int(&bufx->b_flags, B_ERROR); 1153 bp->b_error = bufx->b_error; /* race ok */ 1154 } else if (bufx->b_resid != 0) { 1155 atomic_set_int(&bufx->b_flags, B_ERROR); 1156 bp->b_error = EINVAL; /* race ok */ 1157 } else { 1158 atomic_subtract_int(&bp->b_resid, bufx->b_bcount); 1159 } 1160 1161 /* 1162 * Remove us from the chain. 1163 */ 1164 spin_lock(&swapbp_spin); 1165 nextp = &nbio->bio_caller_info1.cluster_head; 1166 while (*nextp != bufx) { 1167 KKASSERT(*nextp != NULL); 1168 nextp = &(*nextp)->b_cluster_next; 1169 } 1170 *nextp = bufx->b_cluster_next; 1171 chain_empty = (nbio->bio_caller_info1.cluster_head == NULL); 1172 spin_unlock(&swapbp_spin); 1173 1174 /* 1175 * Clean up bufx. If the chain is now empty we finish out 1176 * the parent. Note that we may be racing other completions 1177 * so we must use the chain_empty status from above. 1178 */ 1179 if (chain_empty) { 1180 if (bp->b_resid != 0 && !(bp->b_flags & B_ERROR)) { 1181 atomic_set_int(&bp->b_flags, B_ERROR); 1182 bp->b_error = EINVAL; 1183 } 1184 biodone_chain(nbio); 1185 } 1186 relpbuf(bufx, NULL); 1187 } 1188 1189 /* 1190 * SWAP_PAGER_GETPAGES() - bring page in from swap 1191 * 1192 * The requested page may have to be brought in from swap. Calculate the 1193 * swap block and bring in additional pages if possible. All pages must 1194 * have contiguous swap block assignments and reside in the same object. 1195 * 1196 * The caller has a single vm_object_pip_add() reference prior to 1197 * calling us and we should return with the same. 1198 * 1199 * The caller has BUSY'd the page. We should return with (*mpp) left busy, 1200 * and any additinal pages unbusied. 1201 * 1202 * If the caller encounters a PG_RAM page it will pass it to us even though 1203 * it may be valid and dirty. We cannot overwrite the page in this case! 1204 * The case is used to allow us to issue pure read-aheads. 1205 * 1206 * NOTE! XXX This code does not entirely pipeline yet due to the fact that 1207 * the PG_RAM page is validated at the same time as mreq. What we 1208 * really need to do is issue a separate read-ahead pbuf. 1209 * 1210 * No requirements. 1211 */ 1212 static int 1213 swap_pager_getpage(vm_object_t object, vm_page_t *mpp, int seqaccess) 1214 { 1215 struct buf *bp; 1216 struct bio *bio; 1217 vm_page_t mreq; 1218 vm_page_t m; 1219 vm_offset_t kva; 1220 swblk_t blk; 1221 int i; 1222 int j; 1223 int raonly; 1224 int error; 1225 u_int32_t flags; 1226 vm_page_t marray[XIO_INTERNAL_PAGES]; 1227 1228 mreq = *mpp; 1229 1230 vm_object_hold(object); 1231 if (mreq->object != object) { 1232 panic("swap_pager_getpages: object mismatch %p/%p", 1233 object, 1234 mreq->object 1235 ); 1236 } 1237 1238 /* 1239 * We don't want to overwrite a fully valid page as it might be 1240 * dirty. This case can occur when e.g. vm_fault hits a perfectly 1241 * valid page with PG_RAM set. 1242 * 1243 * In this case we see if the next page is a suitable page-in 1244 * candidate and if it is we issue read-ahead. PG_RAM will be 1245 * set on the last page of the read-ahead to continue the pipeline. 1246 */ 1247 if (mreq->valid == VM_PAGE_BITS_ALL) { 1248 if (swap_burst_read == 0 || mreq->pindex + 1 >= object->size) { 1249 vm_object_drop(object); 1250 return(VM_PAGER_OK); 1251 } 1252 blk = swp_pager_meta_ctl(object, mreq->pindex + 1, 0); 1253 if (blk == SWAPBLK_NONE) { 1254 vm_object_drop(object); 1255 return(VM_PAGER_OK); 1256 } 1257 m = vm_page_lookup_busy_try(object, mreq->pindex + 1, 1258 TRUE, &error); 1259 if (error) { 1260 vm_object_drop(object); 1261 return(VM_PAGER_OK); 1262 } else if (m == NULL) { 1263 /* 1264 * Use VM_ALLOC_QUICK to avoid blocking on cache 1265 * page reuse. 1266 */ 1267 m = vm_page_alloc(object, mreq->pindex + 1, 1268 VM_ALLOC_QUICK); 1269 if (m == NULL) { 1270 vm_object_drop(object); 1271 return(VM_PAGER_OK); 1272 } 1273 } else { 1274 if (m->valid) { 1275 vm_page_wakeup(m); 1276 vm_object_drop(object); 1277 return(VM_PAGER_OK); 1278 } 1279 vm_page_unqueue_nowakeup(m); 1280 } 1281 /* page is busy */ 1282 mreq = m; 1283 raonly = 1; 1284 } else { 1285 raonly = 0; 1286 } 1287 1288 /* 1289 * Try to block-read contiguous pages from swap if sequential, 1290 * otherwise just read one page. Contiguous pages from swap must 1291 * reside within a single device stripe because the I/O cannot be 1292 * broken up across multiple stripes. 1293 * 1294 * Note that blk and iblk can be SWAPBLK_NONE but the loop is 1295 * set up such that the case(s) are handled implicitly. 1296 */ 1297 blk = swp_pager_meta_ctl(mreq->object, mreq->pindex, 0); 1298 marray[0] = mreq; 1299 1300 for (i = 1; swap_burst_read && 1301 i < XIO_INTERNAL_PAGES && 1302 mreq->pindex + i < object->size; ++i) { 1303 swblk_t iblk; 1304 1305 iblk = swp_pager_meta_ctl(object, mreq->pindex + i, 0); 1306 if (iblk != blk + i) 1307 break; 1308 if ((blk ^ iblk) & dmmax_mask) 1309 break; 1310 m = vm_page_lookup_busy_try(object, mreq->pindex + i, 1311 TRUE, &error); 1312 if (error) { 1313 break; 1314 } else if (m == NULL) { 1315 /* 1316 * Use VM_ALLOC_QUICK to avoid blocking on cache 1317 * page reuse. 1318 */ 1319 m = vm_page_alloc(object, mreq->pindex + i, 1320 VM_ALLOC_QUICK); 1321 if (m == NULL) 1322 break; 1323 } else { 1324 if (m->valid) { 1325 vm_page_wakeup(m); 1326 break; 1327 } 1328 vm_page_unqueue_nowakeup(m); 1329 } 1330 /* page is busy */ 1331 marray[i] = m; 1332 } 1333 if (i > 1) 1334 vm_page_flag_set(marray[i - 1], PG_RAM); 1335 1336 /* 1337 * If mreq is the requested page and we have nothing to do return 1338 * VM_PAGER_FAIL. If raonly is set mreq is just another read-ahead 1339 * page and must be cleaned up. 1340 */ 1341 if (blk == SWAPBLK_NONE) { 1342 KKASSERT(i == 1); 1343 if (raonly) { 1344 vnode_pager_freepage(mreq); 1345 vm_object_drop(object); 1346 return(VM_PAGER_OK); 1347 } else { 1348 vm_object_drop(object); 1349 return(VM_PAGER_FAIL); 1350 } 1351 } 1352 1353 /* 1354 * map our page(s) into kva for input 1355 */ 1356 bp = getpbuf_kva(&nsw_rcount); 1357 bio = &bp->b_bio1; 1358 kva = (vm_offset_t) bp->b_kvabase; 1359 bcopy(marray, bp->b_xio.xio_pages, i * sizeof(vm_page_t)); 1360 pmap_qenter(kva, bp->b_xio.xio_pages, i); 1361 1362 bp->b_data = (caddr_t)kva; 1363 bp->b_bcount = PAGE_SIZE * i; 1364 bp->b_xio.xio_npages = i; 1365 bio->bio_done = swp_pager_async_iodone; 1366 bio->bio_offset = (off_t)blk << PAGE_SHIFT; 1367 bio->bio_caller_info1.index = SWBIO_READ; 1368 1369 /* 1370 * Set index. If raonly set the index beyond the array so all 1371 * the pages are treated the same, otherwise the original mreq is 1372 * at index 0. 1373 */ 1374 if (raonly) 1375 bio->bio_driver_info = (void *)(intptr_t)i; 1376 else 1377 bio->bio_driver_info = (void *)(intptr_t)0; 1378 1379 for (j = 0; j < i; ++j) 1380 vm_page_flag_set(bp->b_xio.xio_pages[j], PG_SWAPINPROG); 1381 1382 mycpu->gd_cnt.v_swapin++; 1383 mycpu->gd_cnt.v_swappgsin += bp->b_xio.xio_npages; 1384 1385 /* 1386 * We still hold the lock on mreq, and our automatic completion routine 1387 * does not remove it. 1388 */ 1389 vm_object_pip_add(object, bp->b_xio.xio_npages); 1390 1391 /* 1392 * perform the I/O. NOTE!!! bp cannot be considered valid after 1393 * this point because we automatically release it on completion. 1394 * Instead, we look at the one page we are interested in which we 1395 * still hold a lock on even through the I/O completion. 1396 * 1397 * The other pages in our m[] array are also released on completion, 1398 * so we cannot assume they are valid anymore either. 1399 */ 1400 bp->b_cmd = BUF_CMD_READ; 1401 BUF_KERNPROC(bp); 1402 vn_strategy(swapdev_vp, bio); 1403 1404 /* 1405 * Wait for the page we want to complete. PG_SWAPINPROG is always 1406 * cleared on completion. If an I/O error occurs, SWAPBLK_NONE 1407 * is set in the meta-data. 1408 * 1409 * If this is a read-ahead only we return immediately without 1410 * waiting for I/O. 1411 */ 1412 if (raonly) { 1413 vm_object_drop(object); 1414 return(VM_PAGER_OK); 1415 } 1416 1417 /* 1418 * Read-ahead includes originally requested page case. 1419 */ 1420 for (;;) { 1421 flags = mreq->flags; 1422 cpu_ccfence(); 1423 if ((flags & PG_SWAPINPROG) == 0) 1424 break; 1425 tsleep_interlock(mreq, 0); 1426 if (!atomic_cmpset_int(&mreq->flags, flags, 1427 flags | PG_WANTED | PG_REFERENCED)) { 1428 continue; 1429 } 1430 mycpu->gd_cnt.v_intrans++; 1431 if (tsleep(mreq, PINTERLOCKED, "swread", hz*20)) { 1432 kprintf( 1433 "swap_pager: indefinite wait buffer: " 1434 " bp %p offset: %lld, size: %ld\n", 1435 bp, 1436 (long long)bio->bio_offset, 1437 (long)bp->b_bcount 1438 ); 1439 } 1440 } 1441 1442 /* 1443 * mreq is left bussied after completion, but all the other pages 1444 * are freed. If we had an unrecoverable read error the page will 1445 * not be valid. 1446 */ 1447 vm_object_drop(object); 1448 if (mreq->valid != VM_PAGE_BITS_ALL) 1449 return(VM_PAGER_ERROR); 1450 else 1451 return(VM_PAGER_OK); 1452 1453 /* 1454 * A final note: in a low swap situation, we cannot deallocate swap 1455 * and mark a page dirty here because the caller is likely to mark 1456 * the page clean when we return, causing the page to possibly revert 1457 * to all-zero's later. 1458 */ 1459 } 1460 1461 /* 1462 * swap_pager_putpages: 1463 * 1464 * Assign swap (if necessary) and initiate I/O on the specified pages. 1465 * 1466 * We support both OBJT_DEFAULT and OBJT_SWAP objects. DEFAULT objects 1467 * are automatically converted to SWAP objects. 1468 * 1469 * In a low memory situation we may block in vn_strategy(), but the new 1470 * vm_page reservation system coupled with properly written VFS devices 1471 * should ensure that no low-memory deadlock occurs. This is an area 1472 * which needs work. 1473 * 1474 * The parent has N vm_object_pip_add() references prior to 1475 * calling us and will remove references for rtvals[] that are 1476 * not set to VM_PAGER_PEND. We need to remove the rest on I/O 1477 * completion. 1478 * 1479 * The parent has soft-busy'd the pages it passes us and will unbusy 1480 * those whos rtvals[] entry is not set to VM_PAGER_PEND on return. 1481 * We need to unbusy the rest on I/O completion. 1482 * 1483 * No requirements. 1484 */ 1485 void 1486 swap_pager_putpages(vm_object_t object, vm_page_t *m, int count, 1487 int sync, int *rtvals) 1488 { 1489 int i; 1490 int n = 0; 1491 1492 vm_object_hold(object); 1493 1494 if (count && m[0]->object != object) { 1495 panic("swap_pager_getpages: object mismatch %p/%p", 1496 object, 1497 m[0]->object 1498 ); 1499 } 1500 1501 /* 1502 * Step 1 1503 * 1504 * Turn object into OBJT_SWAP 1505 * check for bogus sysops 1506 * force sync if not pageout process 1507 */ 1508 if (object->type == OBJT_DEFAULT) { 1509 if (object->type == OBJT_DEFAULT) 1510 swp_pager_meta_convert(object); 1511 } 1512 1513 if (curthread != pagethread) 1514 sync = TRUE; 1515 1516 /* 1517 * Step 2 1518 * 1519 * Update nsw parameters from swap_async_max sysctl values. 1520 * Do not let the sysop crash the machine with bogus numbers. 1521 */ 1522 if (swap_async_max != nsw_wcount_async_max) { 1523 int n; 1524 1525 /* 1526 * limit range 1527 */ 1528 if ((n = swap_async_max) > nswbuf_kva / 2) 1529 n = nswbuf_kva / 2; 1530 if (n < 1) 1531 n = 1; 1532 swap_async_max = n; 1533 1534 /* 1535 * Adjust difference ( if possible ). If the current async 1536 * count is too low, we may not be able to make the adjustment 1537 * at this time. 1538 * 1539 * vm_token needed for nsw_wcount sleep interlock 1540 */ 1541 lwkt_gettoken(&vm_token); 1542 n -= nsw_wcount_async_max; 1543 if (nsw_wcount_async + n >= 0) { 1544 nsw_wcount_async_max += n; 1545 pbuf_adjcount(&nsw_wcount_async, n); 1546 } 1547 lwkt_reltoken(&vm_token); 1548 } 1549 1550 /* 1551 * Step 3 1552 * 1553 * Assign swap blocks and issue I/O. We reallocate swap on the fly. 1554 * The page is left dirty until the pageout operation completes 1555 * successfully. 1556 */ 1557 1558 for (i = 0; i < count; i += n) { 1559 struct buf *bp; 1560 struct bio *bio; 1561 swblk_t blk; 1562 int j; 1563 1564 /* 1565 * Maximum I/O size is limited by a number of factors. 1566 */ 1567 1568 n = min(BLIST_MAX_ALLOC, count - i); 1569 n = min(n, nsw_cluster_max); 1570 1571 lwkt_gettoken(&vm_token); 1572 1573 /* 1574 * Get biggest block of swap we can. If we fail, fall 1575 * back and try to allocate a smaller block. Don't go 1576 * overboard trying to allocate space if it would overly 1577 * fragment swap. 1578 */ 1579 while ( 1580 (blk = swp_pager_getswapspace(object, n)) == SWAPBLK_NONE && 1581 n > 4 1582 ) { 1583 n >>= 1; 1584 } 1585 if (blk == SWAPBLK_NONE) { 1586 for (j = 0; j < n; ++j) 1587 rtvals[i+j] = VM_PAGER_FAIL; 1588 lwkt_reltoken(&vm_token); 1589 continue; 1590 } 1591 if (vm_report_swap_allocs > 0) { 1592 kprintf("swap_alloc %08jx,%d\n", (intmax_t)blk, n); 1593 --vm_report_swap_allocs; 1594 } 1595 1596 /* 1597 * The I/O we are constructing cannot cross a physical 1598 * disk boundry in the swap stripe. Note: we are still 1599 * at splvm(). 1600 */ 1601 if ((blk ^ (blk + n)) & dmmax_mask) { 1602 j = ((blk + dmmax) & dmmax_mask) - blk; 1603 swp_pager_freeswapspace(object, blk + j, n - j); 1604 n = j; 1605 } 1606 1607 /* 1608 * All I/O parameters have been satisfied, build the I/O 1609 * request and assign the swap space. 1610 */ 1611 if (sync == TRUE) 1612 bp = getpbuf_kva(&nsw_wcount_sync); 1613 else 1614 bp = getpbuf_kva(&nsw_wcount_async); 1615 bio = &bp->b_bio1; 1616 1617 lwkt_reltoken(&vm_token); 1618 1619 pmap_qenter((vm_offset_t)bp->b_data, &m[i], n); 1620 1621 bp->b_bcount = PAGE_SIZE * n; 1622 bio->bio_offset = (off_t)blk << PAGE_SHIFT; 1623 1624 for (j = 0; j < n; ++j) { 1625 vm_page_t mreq = m[i+j]; 1626 1627 swp_pager_meta_build(mreq->object, mreq->pindex, 1628 blk + j); 1629 if (object->type == OBJT_SWAP) 1630 vm_page_dirty(mreq); 1631 rtvals[i+j] = VM_PAGER_OK; 1632 1633 vm_page_flag_set(mreq, PG_SWAPINPROG); 1634 bp->b_xio.xio_pages[j] = mreq; 1635 } 1636 bp->b_xio.xio_npages = n; 1637 1638 mycpu->gd_cnt.v_swapout++; 1639 mycpu->gd_cnt.v_swappgsout += bp->b_xio.xio_npages; 1640 1641 bp->b_dirtyoff = 0; /* req'd for NFS */ 1642 bp->b_dirtyend = bp->b_bcount; /* req'd for NFS */ 1643 bp->b_cmd = BUF_CMD_WRITE; 1644 bio->bio_caller_info1.index = SWBIO_WRITE; 1645 1646 /* 1647 * asynchronous 1648 */ 1649 if (sync == FALSE) { 1650 bio->bio_done = swp_pager_async_iodone; 1651 BUF_KERNPROC(bp); 1652 vn_strategy(swapdev_vp, bio); 1653 1654 for (j = 0; j < n; ++j) 1655 rtvals[i+j] = VM_PAGER_PEND; 1656 continue; 1657 } 1658 1659 /* 1660 * Issue synchrnously. 1661 * 1662 * Wait for the sync I/O to complete, then update rtvals. 1663 * We just set the rtvals[] to VM_PAGER_PEND so we can call 1664 * our async completion routine at the end, thus avoiding a 1665 * double-free. 1666 */ 1667 bio->bio_caller_info1.index |= SWBIO_SYNC; 1668 bio->bio_done = biodone_sync; 1669 bio->bio_flags |= BIO_SYNC; 1670 vn_strategy(swapdev_vp, bio); 1671 biowait(bio, "swwrt"); 1672 1673 for (j = 0; j < n; ++j) 1674 rtvals[i+j] = VM_PAGER_PEND; 1675 1676 /* 1677 * Now that we are through with the bp, we can call the 1678 * normal async completion, which frees everything up. 1679 */ 1680 swp_pager_async_iodone(bio); 1681 } 1682 vm_object_drop(object); 1683 } 1684 1685 /* 1686 * No requirements. 1687 * 1688 * Recalculate the low and high-water marks. 1689 */ 1690 void 1691 swap_pager_newswap(void) 1692 { 1693 if (vm_swap_max) { 1694 nswap_lowat = vm_swap_max * 4 / 100; /* 4% left */ 1695 nswap_hiwat = vm_swap_max * 6 / 100; /* 6% left */ 1696 kprintf("swap low/high-water marks set to %d/%d\n", 1697 nswap_lowat, nswap_hiwat); 1698 } else { 1699 nswap_lowat = 128; 1700 nswap_hiwat = 512; 1701 } 1702 swp_sizecheck(); 1703 } 1704 1705 /* 1706 * swp_pager_async_iodone: 1707 * 1708 * Completion routine for asynchronous reads and writes from/to swap. 1709 * Also called manually by synchronous code to finish up a bp. 1710 * 1711 * For READ operations, the pages are PG_BUSY'd. For WRITE operations, 1712 * the pages are vm_page_t->busy'd. For READ operations, we PG_BUSY 1713 * unbusy all pages except the 'main' request page. For WRITE 1714 * operations, we vm_page_t->busy'd unbusy all pages ( we can do this 1715 * because we marked them all VM_PAGER_PEND on return from putpages ). 1716 * 1717 * This routine may not block. 1718 * 1719 * No requirements. 1720 */ 1721 static void 1722 swp_pager_async_iodone(struct bio *bio) 1723 { 1724 struct buf *bp = bio->bio_buf; 1725 vm_object_t object = NULL; 1726 int i; 1727 int *nswptr; 1728 1729 /* 1730 * report error 1731 */ 1732 if (bp->b_flags & B_ERROR) { 1733 kprintf( 1734 "swap_pager: I/O error - %s failed; offset %lld," 1735 "size %ld, error %d\n", 1736 ((bio->bio_caller_info1.index & SWBIO_READ) ? 1737 "pagein" : "pageout"), 1738 (long long)bio->bio_offset, 1739 (long)bp->b_bcount, 1740 bp->b_error 1741 ); 1742 } 1743 1744 /* 1745 * set object, raise to splvm(). 1746 */ 1747 if (bp->b_xio.xio_npages) 1748 object = bp->b_xio.xio_pages[0]->object; 1749 1750 /* 1751 * remove the mapping for kernel virtual 1752 */ 1753 pmap_qremove((vm_offset_t)bp->b_data, bp->b_xio.xio_npages); 1754 1755 /* 1756 * cleanup pages. If an error occurs writing to swap, we are in 1757 * very serious trouble. If it happens to be a disk error, though, 1758 * we may be able to recover by reassigning the swap later on. So 1759 * in this case we remove the m->swapblk assignment for the page 1760 * but do not free it in the rlist. The errornous block(s) are thus 1761 * never reallocated as swap. Redirty the page and continue. 1762 */ 1763 for (i = 0; i < bp->b_xio.xio_npages; ++i) { 1764 vm_page_t m = bp->b_xio.xio_pages[i]; 1765 1766 if (bp->b_flags & B_ERROR) { 1767 /* 1768 * If an error occurs I'd love to throw the swapblk 1769 * away without freeing it back to swapspace, so it 1770 * can never be used again. But I can't from an 1771 * interrupt. 1772 */ 1773 1774 if (bio->bio_caller_info1.index & SWBIO_READ) { 1775 /* 1776 * When reading, reqpage needs to stay 1777 * locked for the parent, but all other 1778 * pages can be freed. We still want to 1779 * wakeup the parent waiting on the page, 1780 * though. ( also: pg_reqpage can be -1 and 1781 * not match anything ). 1782 * 1783 * We have to wake specifically requested pages 1784 * up too because we cleared PG_SWAPINPROG and 1785 * someone may be waiting for that. 1786 * 1787 * NOTE: for reads, m->dirty will probably 1788 * be overridden by the original caller of 1789 * getpages so don't play cute tricks here. 1790 * 1791 * NOTE: We can't actually free the page from 1792 * here, because this is an interrupt. It 1793 * is not legal to mess with object->memq 1794 * from an interrupt. Deactivate the page 1795 * instead. 1796 */ 1797 1798 m->valid = 0; 1799 vm_page_flag_clear(m, PG_SWAPINPROG); 1800 1801 /* 1802 * bio_driver_info holds the requested page 1803 * index. 1804 */ 1805 if (i != (int)(intptr_t)bio->bio_driver_info) { 1806 vm_page_deactivate(m); 1807 vm_page_wakeup(m); 1808 } else { 1809 vm_page_flash(m); 1810 } 1811 /* 1812 * If i == bp->b_pager.pg_reqpage, do not wake 1813 * the page up. The caller needs to. 1814 */ 1815 } else { 1816 /* 1817 * If a write error occurs remove the swap 1818 * assignment (note that PG_SWAPPED may or 1819 * may not be set depending on prior activity). 1820 * 1821 * Re-dirty OBJT_SWAP pages as there is no 1822 * other backing store, we can't throw the 1823 * page away. 1824 * 1825 * Non-OBJT_SWAP pages (aka swapcache) must 1826 * not be dirtied since they may not have 1827 * been dirty in the first place, and they 1828 * do have backing store (the vnode). 1829 */ 1830 vm_page_busy_wait(m, FALSE, "swadpg"); 1831 swp_pager_meta_ctl(m->object, m->pindex, 1832 SWM_FREE); 1833 vm_page_flag_clear(m, PG_SWAPPED); 1834 if (m->object->type == OBJT_SWAP) { 1835 vm_page_dirty(m); 1836 vm_page_activate(m); 1837 } 1838 vm_page_flag_clear(m, PG_SWAPINPROG); 1839 vm_page_io_finish(m); 1840 vm_page_wakeup(m); 1841 } 1842 } else if (bio->bio_caller_info1.index & SWBIO_READ) { 1843 /* 1844 * NOTE: for reads, m->dirty will probably be 1845 * overridden by the original caller of getpages so 1846 * we cannot set them in order to free the underlying 1847 * swap in a low-swap situation. I don't think we'd 1848 * want to do that anyway, but it was an optimization 1849 * that existed in the old swapper for a time before 1850 * it got ripped out due to precisely this problem. 1851 * 1852 * If not the requested page then deactivate it. 1853 * 1854 * Note that the requested page, reqpage, is left 1855 * busied, but we still have to wake it up. The 1856 * other pages are released (unbusied) by 1857 * vm_page_wakeup(). We do not set reqpage's 1858 * valid bits here, it is up to the caller. 1859 */ 1860 1861 /* 1862 * NOTE: can't call pmap_clear_modify(m) from an 1863 * interrupt thread, the pmap code may have to map 1864 * non-kernel pmaps and currently asserts the case. 1865 */ 1866 /*pmap_clear_modify(m);*/ 1867 m->valid = VM_PAGE_BITS_ALL; 1868 vm_page_undirty(m); 1869 vm_page_flag_clear(m, PG_SWAPINPROG); 1870 vm_page_flag_set(m, PG_SWAPPED); 1871 1872 /* 1873 * We have to wake specifically requested pages 1874 * up too because we cleared PG_SWAPINPROG and 1875 * could be waiting for it in getpages. However, 1876 * be sure to not unbusy getpages specifically 1877 * requested page - getpages expects it to be 1878 * left busy. 1879 * 1880 * bio_driver_info holds the requested page 1881 */ 1882 if (i != (int)(intptr_t)bio->bio_driver_info) { 1883 vm_page_deactivate(m); 1884 vm_page_wakeup(m); 1885 } else { 1886 vm_page_flash(m); 1887 } 1888 } else { 1889 /* 1890 * Mark the page clean but do not mess with the 1891 * pmap-layer's modified state. That state should 1892 * also be clear since the caller protected the 1893 * page VM_PROT_READ, but allow the case. 1894 * 1895 * We are in an interrupt, avoid pmap operations. 1896 * 1897 * If we have a severe page deficit, deactivate the 1898 * page. Do not try to cache it (which would also 1899 * involve a pmap op), because the page might still 1900 * be read-heavy. 1901 * 1902 * When using the swap to cache clean vnode pages 1903 * we do not mess with the page dirty bits. 1904 */ 1905 vm_page_busy_wait(m, FALSE, "swadpg"); 1906 if (m->object->type == OBJT_SWAP) 1907 vm_page_undirty(m); 1908 vm_page_flag_clear(m, PG_SWAPINPROG); 1909 vm_page_flag_set(m, PG_SWAPPED); 1910 if (vm_page_count_severe()) 1911 vm_page_deactivate(m); 1912 #if 0 1913 if (!vm_page_count_severe() || !vm_page_try_to_cache(m)) 1914 vm_page_protect(m, VM_PROT_READ); 1915 #endif 1916 vm_page_io_finish(m); 1917 vm_page_wakeup(m); 1918 } 1919 } 1920 1921 /* 1922 * adjust pip. NOTE: the original parent may still have its own 1923 * pip refs on the object. 1924 */ 1925 1926 if (object) 1927 vm_object_pip_wakeup_n(object, bp->b_xio.xio_npages); 1928 1929 /* 1930 * Release the physical I/O buffer. 1931 * 1932 * NOTE: Due to synchronous operations in the write case b_cmd may 1933 * already be set to BUF_CMD_DONE and BIO_SYNC may have already 1934 * been cleared. 1935 * 1936 * Use vm_token to interlock nsw_rcount/wcount wakeup? 1937 */ 1938 lwkt_gettoken(&vm_token); 1939 if (bio->bio_caller_info1.index & SWBIO_READ) 1940 nswptr = &nsw_rcount; 1941 else if (bio->bio_caller_info1.index & SWBIO_SYNC) 1942 nswptr = &nsw_wcount_sync; 1943 else 1944 nswptr = &nsw_wcount_async; 1945 bp->b_cmd = BUF_CMD_DONE; 1946 relpbuf(bp, nswptr); 1947 lwkt_reltoken(&vm_token); 1948 } 1949 1950 /* 1951 * Fault-in a potentially swapped page and remove the swap reference. 1952 * (used by swapoff code) 1953 * 1954 * object must be held. 1955 */ 1956 static __inline void 1957 swp_pager_fault_page(vm_object_t object, int *sharedp, vm_pindex_t pindex) 1958 { 1959 struct vnode *vp; 1960 vm_page_t m; 1961 int error; 1962 1963 ASSERT_LWKT_TOKEN_HELD(vm_object_token(object)); 1964 1965 if (object->type == OBJT_VNODE) { 1966 /* 1967 * Any swap related to a vnode is due to swapcache. We must 1968 * vget() the vnode in case it is not active (otherwise 1969 * vref() will panic). Calling vm_object_page_remove() will 1970 * ensure that any swap ref is removed interlocked with the 1971 * page. clean_only is set to TRUE so we don't throw away 1972 * dirty pages. 1973 */ 1974 vp = object->handle; 1975 error = vget(vp, LK_SHARED | LK_RETRY | LK_CANRECURSE); 1976 if (error == 0) { 1977 vm_object_page_remove(object, pindex, pindex + 1, TRUE); 1978 vput(vp); 1979 } 1980 } else { 1981 /* 1982 * Otherwise it is a normal OBJT_SWAP object and we can 1983 * fault the page in and remove the swap. 1984 */ 1985 m = vm_fault_object_page(object, IDX_TO_OFF(pindex), 1986 VM_PROT_NONE, 1987 VM_FAULT_DIRTY | VM_FAULT_UNSWAP, 1988 sharedp, &error); 1989 if (m) 1990 vm_page_unhold(m); 1991 } 1992 } 1993 1994 /* 1995 * This removes all swap blocks related to a particular device. We have 1996 * to be careful of ripups during the scan. 1997 */ 1998 static int swp_pager_swapoff_callback(struct swblock *swap, void *data); 1999 2000 int 2001 swap_pager_swapoff(int devidx) 2002 { 2003 struct swswapoffinfo info; 2004 struct vm_object marker; 2005 vm_object_t object; 2006 int n; 2007 2008 bzero(&marker, sizeof(marker)); 2009 marker.type = OBJT_MARKER; 2010 2011 for (n = 0; n < VMOBJ_HSIZE; ++n) { 2012 lwkt_gettoken(&vmobj_tokens[n]); 2013 TAILQ_INSERT_HEAD(&vm_object_lists[n], &marker, object_list); 2014 2015 while ((object = TAILQ_NEXT(&marker, object_list)) != NULL) { 2016 if (object->type == OBJT_MARKER) 2017 goto skip; 2018 if (object->type != OBJT_SWAP && 2019 object->type != OBJT_VNODE) 2020 goto skip; 2021 vm_object_hold(object); 2022 if (object->type != OBJT_SWAP && 2023 object->type != OBJT_VNODE) { 2024 vm_object_drop(object); 2025 goto skip; 2026 } 2027 info.object = object; 2028 info.shared = 0; 2029 info.devidx = devidx; 2030 swblock_rb_tree_RB_SCAN(&object->swblock_root, 2031 NULL, swp_pager_swapoff_callback, 2032 &info); 2033 vm_object_drop(object); 2034 skip: 2035 if (object == TAILQ_NEXT(&marker, object_list)) { 2036 TAILQ_REMOVE(&vm_object_lists[n], 2037 &marker, object_list); 2038 TAILQ_INSERT_AFTER(&vm_object_lists[n], object, 2039 &marker, object_list); 2040 } 2041 } 2042 TAILQ_REMOVE(&vm_object_lists[n], &marker, object_list); 2043 lwkt_reltoken(&vmobj_tokens[n]); 2044 } 2045 2046 /* 2047 * If we fail to locate all swblocks we just fail gracefully and 2048 * do not bother to restore paging on the swap device. If the 2049 * user wants to retry the user can retry. 2050 */ 2051 if (swdevt[devidx].sw_nused) 2052 return (1); 2053 else 2054 return (0); 2055 } 2056 2057 static 2058 int 2059 swp_pager_swapoff_callback(struct swblock *swap, void *data) 2060 { 2061 struct swswapoffinfo *info = data; 2062 vm_object_t object = info->object; 2063 vm_pindex_t index; 2064 swblk_t v; 2065 int i; 2066 2067 index = swap->swb_index; 2068 for (i = 0; i < SWAP_META_PAGES; ++i) { 2069 /* 2070 * Make sure we don't race a dying object. This will 2071 * kill the scan of the object's swap blocks entirely. 2072 */ 2073 if (object->flags & OBJ_DEAD) 2074 return(-1); 2075 2076 /* 2077 * Fault the page, which can obviously block. If the swap 2078 * structure disappears break out. 2079 */ 2080 v = swap->swb_pages[i]; 2081 if (v != SWAPBLK_NONE && BLK2DEVIDX(v) == info->devidx) { 2082 swp_pager_fault_page(object, &info->shared, 2083 swap->swb_index + i); 2084 /* swap ptr might go away */ 2085 if (RB_LOOKUP(swblock_rb_tree, 2086 &object->swblock_root, index) != swap) { 2087 break; 2088 } 2089 } 2090 } 2091 return(0); 2092 } 2093 2094 /************************************************************************ 2095 * SWAP META DATA * 2096 ************************************************************************ 2097 * 2098 * These routines manipulate the swap metadata stored in the 2099 * OBJT_SWAP object. All swp_*() routines must be called at 2100 * splvm() because swap can be freed up by the low level vm_page 2101 * code which might be called from interrupts beyond what splbio() covers. 2102 * 2103 * Swap metadata is implemented with a global hash and not directly 2104 * linked into the object. Instead the object simply contains 2105 * appropriate tracking counters. 2106 */ 2107 2108 /* 2109 * Lookup the swblock containing the specified swap block index. 2110 * 2111 * The caller must hold the object. 2112 */ 2113 static __inline 2114 struct swblock * 2115 swp_pager_lookup(vm_object_t object, vm_pindex_t index) 2116 { 2117 ASSERT_LWKT_TOKEN_HELD(vm_object_token(object)); 2118 index &= ~(vm_pindex_t)SWAP_META_MASK; 2119 return (RB_LOOKUP(swblock_rb_tree, &object->swblock_root, index)); 2120 } 2121 2122 /* 2123 * Remove a swblock from the RB tree. 2124 * 2125 * The caller must hold the object. 2126 */ 2127 static __inline 2128 void 2129 swp_pager_remove(vm_object_t object, struct swblock *swap) 2130 { 2131 ASSERT_LWKT_TOKEN_HELD(vm_object_token(object)); 2132 RB_REMOVE(swblock_rb_tree, &object->swblock_root, swap); 2133 } 2134 2135 /* 2136 * Convert default object to swap object if necessary 2137 * 2138 * The caller must hold the object. 2139 */ 2140 static void 2141 swp_pager_meta_convert(vm_object_t object) 2142 { 2143 if (object->type == OBJT_DEFAULT) { 2144 object->type = OBJT_SWAP; 2145 KKASSERT(object->swblock_count == 0); 2146 } 2147 } 2148 2149 /* 2150 * SWP_PAGER_META_BUILD() - add swap block to swap meta data for object 2151 * 2152 * We first convert the object to a swap object if it is a default 2153 * object. Vnode objects do not need to be converted. 2154 * 2155 * The specified swapblk is added to the object's swap metadata. If 2156 * the swapblk is not valid, it is freed instead. Any previously 2157 * assigned swapblk is freed. 2158 * 2159 * The caller must hold the object. 2160 */ 2161 static void 2162 swp_pager_meta_build(vm_object_t object, vm_pindex_t index, swblk_t swapblk) 2163 { 2164 struct swblock *swap; 2165 struct swblock *oswap; 2166 vm_pindex_t v; 2167 2168 KKASSERT(swapblk != SWAPBLK_NONE); 2169 ASSERT_LWKT_TOKEN_HELD(vm_object_token(object)); 2170 2171 /* 2172 * Convert object if necessary 2173 */ 2174 if (object->type == OBJT_DEFAULT) 2175 swp_pager_meta_convert(object); 2176 2177 /* 2178 * Locate swblock. If not found create, but if we aren't adding 2179 * anything just return. If we run out of space in the map we wait 2180 * and, since the hash table may have changed, retry. 2181 */ 2182 retry: 2183 swap = swp_pager_lookup(object, index); 2184 2185 if (swap == NULL) { 2186 int i; 2187 2188 swap = zalloc(swap_zone); 2189 if (swap == NULL) { 2190 vm_wait(0); 2191 goto retry; 2192 } 2193 swap->swb_index = index & ~(vm_pindex_t)SWAP_META_MASK; 2194 swap->swb_count = 0; 2195 2196 ++object->swblock_count; 2197 2198 for (i = 0; i < SWAP_META_PAGES; ++i) 2199 swap->swb_pages[i] = SWAPBLK_NONE; 2200 oswap = RB_INSERT(swblock_rb_tree, &object->swblock_root, swap); 2201 KKASSERT(oswap == NULL); 2202 } 2203 2204 /* 2205 * Delete prior contents of metadata. 2206 * 2207 * NOTE: Decrement swb_count after the freeing operation (which 2208 * might block) to prevent racing destruction of the swblock. 2209 */ 2210 index &= SWAP_META_MASK; 2211 2212 while ((v = swap->swb_pages[index]) != SWAPBLK_NONE) { 2213 swap->swb_pages[index] = SWAPBLK_NONE; 2214 /* can block */ 2215 swp_pager_freeswapspace(object, v, 1); 2216 --swap->swb_count; 2217 --mycpu->gd_vmtotal.t_vm; 2218 } 2219 2220 /* 2221 * Enter block into metadata 2222 */ 2223 swap->swb_pages[index] = swapblk; 2224 if (swapblk != SWAPBLK_NONE) { 2225 ++swap->swb_count; 2226 ++mycpu->gd_vmtotal.t_vm; 2227 } 2228 } 2229 2230 /* 2231 * SWP_PAGER_META_FREE() - free a range of blocks in the object's swap metadata 2232 * 2233 * The requested range of blocks is freed, with any associated swap 2234 * returned to the swap bitmap. 2235 * 2236 * This routine will free swap metadata structures as they are cleaned 2237 * out. This routine does *NOT* operate on swap metadata associated 2238 * with resident pages. 2239 * 2240 * The caller must hold the object. 2241 */ 2242 static int swp_pager_meta_free_callback(struct swblock *swb, void *data); 2243 2244 static void 2245 swp_pager_meta_free(vm_object_t object, vm_pindex_t index, vm_pindex_t count) 2246 { 2247 struct swfreeinfo info; 2248 2249 ASSERT_LWKT_TOKEN_HELD(vm_object_token(object)); 2250 2251 /* 2252 * Nothing to do 2253 */ 2254 if (object->swblock_count == 0) { 2255 KKASSERT(RB_EMPTY(&object->swblock_root)); 2256 return; 2257 } 2258 if (count == 0) 2259 return; 2260 2261 /* 2262 * Setup for RB tree scan. Note that the pindex range can be huge 2263 * due to the 64 bit page index space so we cannot safely iterate. 2264 */ 2265 info.object = object; 2266 info.basei = index & ~(vm_pindex_t)SWAP_META_MASK; 2267 info.begi = index; 2268 info.endi = index + count - 1; 2269 swblock_rb_tree_RB_SCAN(&object->swblock_root, rb_swblock_scancmp, 2270 swp_pager_meta_free_callback, &info); 2271 } 2272 2273 /* 2274 * The caller must hold the object. 2275 */ 2276 static 2277 int 2278 swp_pager_meta_free_callback(struct swblock *swap, void *data) 2279 { 2280 struct swfreeinfo *info = data; 2281 vm_object_t object = info->object; 2282 int index; 2283 int eindex; 2284 2285 /* 2286 * Figure out the range within the swblock. The wider scan may 2287 * return edge-case swap blocks when the start and/or end points 2288 * are in the middle of a block. 2289 */ 2290 if (swap->swb_index < info->begi) 2291 index = (int)info->begi & SWAP_META_MASK; 2292 else 2293 index = 0; 2294 2295 if (swap->swb_index + SWAP_META_PAGES > info->endi) 2296 eindex = (int)info->endi & SWAP_META_MASK; 2297 else 2298 eindex = SWAP_META_MASK; 2299 2300 /* 2301 * Scan and free the blocks. The loop terminates early 2302 * if (swap) runs out of blocks and could be freed. 2303 * 2304 * NOTE: Decrement swb_count after swp_pager_freeswapspace() 2305 * to deal with a zfree race. 2306 */ 2307 while (index <= eindex) { 2308 swblk_t v = swap->swb_pages[index]; 2309 2310 if (v != SWAPBLK_NONE) { 2311 swap->swb_pages[index] = SWAPBLK_NONE; 2312 /* can block */ 2313 swp_pager_freeswapspace(object, v, 1); 2314 --mycpu->gd_vmtotal.t_vm; 2315 if (--swap->swb_count == 0) { 2316 swp_pager_remove(object, swap); 2317 zfree(swap_zone, swap); 2318 --object->swblock_count; 2319 break; 2320 } 2321 } 2322 ++index; 2323 } 2324 2325 /* swap may be invalid here due to zfree above */ 2326 lwkt_yield(); 2327 2328 return(0); 2329 } 2330 2331 /* 2332 * SWP_PAGER_META_FREE_ALL() - destroy all swap metadata associated with object 2333 * 2334 * This routine locates and destroys all swap metadata associated with 2335 * an object. 2336 * 2337 * NOTE: Decrement swb_count after the freeing operation (which 2338 * might block) to prevent racing destruction of the swblock. 2339 * 2340 * The caller must hold the object. 2341 */ 2342 static void 2343 swp_pager_meta_free_all(vm_object_t object) 2344 { 2345 struct swblock *swap; 2346 int i; 2347 2348 ASSERT_LWKT_TOKEN_HELD(vm_object_token(object)); 2349 2350 while ((swap = RB_ROOT(&object->swblock_root)) != NULL) { 2351 swp_pager_remove(object, swap); 2352 for (i = 0; i < SWAP_META_PAGES; ++i) { 2353 swblk_t v = swap->swb_pages[i]; 2354 if (v != SWAPBLK_NONE) { 2355 /* can block */ 2356 swp_pager_freeswapspace(object, v, 1); 2357 --swap->swb_count; 2358 --mycpu->gd_vmtotal.t_vm; 2359 } 2360 } 2361 if (swap->swb_count != 0) 2362 panic("swap_pager_meta_free_all: swb_count != 0"); 2363 zfree(swap_zone, swap); 2364 --object->swblock_count; 2365 lwkt_yield(); 2366 } 2367 KKASSERT(object->swblock_count == 0); 2368 } 2369 2370 /* 2371 * SWP_PAGER_METACTL() - misc control of swap and vm_page_t meta data. 2372 * 2373 * This routine is capable of looking up, popping, or freeing 2374 * swapblk assignments in the swap meta data or in the vm_page_t. 2375 * The routine typically returns the swapblk being looked-up, or popped, 2376 * or SWAPBLK_NONE if the block was freed, or SWAPBLK_NONE if the block 2377 * was invalid. This routine will automatically free any invalid 2378 * meta-data swapblks. 2379 * 2380 * It is not possible to store invalid swapblks in the swap meta data 2381 * (other then a literal 'SWAPBLK_NONE'), so we don't bother checking. 2382 * 2383 * When acting on a busy resident page and paging is in progress, we 2384 * have to wait until paging is complete but otherwise can act on the 2385 * busy page. 2386 * 2387 * SWM_FREE remove and free swap block from metadata 2388 * SWM_POP remove from meta data but do not free.. pop it out 2389 * 2390 * The caller must hold the object. 2391 */ 2392 static swblk_t 2393 swp_pager_meta_ctl(vm_object_t object, vm_pindex_t index, int flags) 2394 { 2395 struct swblock *swap; 2396 swblk_t r1; 2397 2398 if (object->swblock_count == 0) 2399 return(SWAPBLK_NONE); 2400 2401 r1 = SWAPBLK_NONE; 2402 swap = swp_pager_lookup(object, index); 2403 2404 if (swap != NULL) { 2405 index &= SWAP_META_MASK; 2406 r1 = swap->swb_pages[index]; 2407 2408 if (r1 != SWAPBLK_NONE) { 2409 if (flags & (SWM_FREE|SWM_POP)) { 2410 swap->swb_pages[index] = SWAPBLK_NONE; 2411 --mycpu->gd_vmtotal.t_vm; 2412 if (--swap->swb_count == 0) { 2413 swp_pager_remove(object, swap); 2414 zfree(swap_zone, swap); 2415 --object->swblock_count; 2416 } 2417 } 2418 /* swap ptr may be invalid */ 2419 if (flags & SWM_FREE) { 2420 swp_pager_freeswapspace(object, r1, 1); 2421 r1 = SWAPBLK_NONE; 2422 } 2423 } 2424 /* swap ptr may be invalid */ 2425 } 2426 return(r1); 2427 } 2428