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