1 /* 2 * Copyright (c) 1998,2004 The DragonFly Project. All rights reserved. 3 * 4 * This code is derived from software contributed to The DragonFly Project 5 * by Matthew Dillon <dillon@backplane.com> 6 * 7 * Redistribution and use in source and binary forms, with or without 8 * modification, are permitted provided that the following conditions 9 * are met: 10 * 11 * 1. Redistributions of source code must retain the above copyright 12 * notice, this list of conditions and the following disclaimer. 13 * 2. Redistributions in binary form must reproduce the above copyright 14 * notice, this list of conditions and the following disclaimer in 15 * the documentation and/or other materials provided with the 16 * distribution. 17 * 3. Neither the name of The DragonFly Project nor the names of its 18 * contributors may be used to endorse or promote products derived 19 * from this software without specific, prior written permission. 20 * 21 * THIS SOFTWARE IS PROVIDED BY THE COPYRIGHT HOLDERS AND CONTRIBUTORS 22 * ``AS IS'' AND ANY EXPRESS OR IMPLIED WARRANTIES, INCLUDING, BUT NOT 23 * LIMITED TO, THE IMPLIED WARRANTIES OF MERCHANTABILITY AND FITNESS 24 * FOR A PARTICULAR PURPOSE ARE DISCLAIMED. IN NO EVENT SHALL THE 25 * COPYRIGHT HOLDERS OR CONTRIBUTORS BE LIABLE FOR ANY DIRECT, INDIRECT, 26 * INCIDENTAL, SPECIAL, EXEMPLARY OR CONSEQUENTIAL DAMAGES (INCLUDING, 27 * BUT NOT LIMITED TO, PROCUREMENT OF SUBSTITUTE GOODS OR SERVICES; 28 * LOSS OF USE, DATA, OR PROFITS; OR BUSINESS INTERRUPTION) HOWEVER CAUSED 29 * AND ON ANY THEORY OF LIABILITY, WHETHER IN CONTRACT, STRICT LIABILITY, 30 * OR TORT (INCLUDING NEGLIGENCE OR OTHERWISE) ARISING IN ANY WAY OUT 31 * OF THE USE OF THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF 32 * SUCH DAMAGE. 33 * 34 * Copyright (c) 1994 John S. Dyson 35 * Copyright (c) 1990 University of Utah. 36 * Copyright (c) 1991, 1993 37 * The Regents of the University of California. All rights reserved. 38 * 39 * This code is derived from software contributed to Berkeley by 40 * the Systems Programming Group of the University of Utah Computer 41 * Science Department. 42 * 43 * Redistribution and use in source and binary forms, with or without 44 * modification, are permitted provided that the following conditions 45 * are met: 46 * 1. Redistributions of source code must retain the above copyright 47 * notice, this list of conditions and the following disclaimer. 48 * 2. Redistributions in binary form must reproduce the above copyright 49 * notice, this list of conditions and the following disclaimer in the 50 * documentation and/or other materials provided with the distribution. 51 * 3. All advertising materials mentioning features or use of this software 52 * must display the following acknowledgement: 53 * This product includes software developed by the University of 54 * California, Berkeley and its contributors. 55 * 4. Neither the name of the University nor the names of its contributors 56 * may be used to endorse or promote products derived from this software 57 * without specific prior written permission. 58 * 59 * THIS SOFTWARE IS PROVIDED BY THE REGENTS AND CONTRIBUTORS ``AS IS'' AND 60 * ANY EXPRESS OR IMPLIED WARRANTIES, INCLUDING, BUT NOT LIMITED TO, THE 61 * IMPLIED WARRANTIES OF MERCHANTABILITY AND FITNESS FOR A PARTICULAR PURPOSE 62 * ARE DISCLAIMED. IN NO EVENT SHALL THE REGENTS OR CONTRIBUTORS BE LIABLE 63 * FOR ANY DIRECT, INDIRECT, INCIDENTAL, SPECIAL, EXEMPLARY, OR CONSEQUENTIAL 64 * DAMAGES (INCLUDING, BUT NOT LIMITED TO, PROCUREMENT OF SUBSTITUTE GOODS 65 * OR SERVICES; LOSS OF USE, DATA, OR PROFITS; OR BUSINESS INTERRUPTION) 66 * HOWEVER CAUSED AND ON ANY THEORY OF LIABILITY, WHETHER IN CONTRACT, STRICT 67 * LIABILITY, OR TORT (INCLUDING NEGLIGENCE OR OTHERWISE) ARISING IN ANY WAY 68 * OUT OF THE USE OF THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF 69 * SUCH DAMAGE. 70 * 71 * New Swap System 72 * Matthew Dillon 73 * 74 * Radix Bitmap 'blists'. 75 * 76 * - The new swapper uses the new radix bitmap code. This should scale 77 * to arbitrarily small or arbitrarily large swap spaces and an almost 78 * arbitrary degree of fragmentation. 79 * 80 * Features: 81 * 82 * - on the fly reallocation of swap during putpages. The new system 83 * does not try to keep previously allocated swap blocks for dirty 84 * pages. 85 * 86 * - on the fly deallocation of swap 87 * 88 * - No more garbage collection required. Unnecessarily allocated swap 89 * blocks only exist for dirty vm_page_t's now and these are already 90 * cycled (in a high-load system) by the pager. We also do on-the-fly 91 * removal of invalidated swap blocks when a page is destroyed 92 * or renamed. 93 * 94 * from: Utah $Hdr: swap_pager.c 1.4 91/04/30$ 95 * 96 * @(#)swap_pager.c 8.9 (Berkeley) 3/21/94 97 * 98 * $FreeBSD: src/sys/vm/swap_pager.c,v 1.130.2.12 2002/08/31 21:15:55 dillon Exp $ 99 * $DragonFly: src/sys/vm/swap_pager.c,v 1.32 2008/07/01 02:02:56 dillon Exp $ 100 */ 101 102 #include <sys/param.h> 103 #include <sys/systm.h> 104 #include <sys/conf.h> 105 #include <sys/kernel.h> 106 #include <sys/proc.h> 107 #include <sys/buf.h> 108 #include <sys/vnode.h> 109 #include <sys/malloc.h> 110 #include <sys/vmmeter.h> 111 #include <sys/sysctl.h> 112 #include <sys/blist.h> 113 #include <sys/lock.h> 114 #include <sys/thread2.h> 115 116 #ifndef MAX_PAGEOUT_CLUSTER 117 #define MAX_PAGEOUT_CLUSTER 16 118 #endif 119 120 #define SWB_NPAGES MAX_PAGEOUT_CLUSTER 121 122 #include "opt_swap.h" 123 #include <vm/vm.h> 124 #include <vm/vm_object.h> 125 #include <vm/vm_page.h> 126 #include <vm/vm_pager.h> 127 #include <vm/vm_pageout.h> 128 #include <vm/swap_pager.h> 129 #include <vm/vm_extern.h> 130 #include <vm/vm_zone.h> 131 132 #include <sys/buf2.h> 133 #include <vm/vm_page2.h> 134 135 #define SWM_FREE 0x02 /* free, period */ 136 #define SWM_POP 0x04 /* pop out */ 137 138 #define SWBIO_READ 0x01 139 #define SWBIO_WRITE 0x02 140 #define SWBIO_SYNC 0x04 141 142 /* 143 * vm_swap_size is in page-sized chunks now. It was DEV_BSIZE'd chunks 144 * in the old system. 145 */ 146 147 extern int vm_swap_size; /* number of free swap blocks, in pages */ 148 149 int swap_pager_full; /* swap space exhaustion (task killing) */ 150 static int swap_pager_almost_full; /* swap space exhaustion (w/ hysteresis)*/ 151 static int nsw_rcount; /* free read buffers */ 152 static int nsw_wcount_sync; /* limit write buffers / synchronous */ 153 static int nsw_wcount_async; /* limit write buffers / asynchronous */ 154 static int nsw_wcount_async_max;/* assigned maximum */ 155 static int nsw_cluster_max; /* maximum VOP I/O allowed */ 156 static int sw_alloc_interlock; /* swap pager allocation interlock */ 157 158 struct blist *swapblist; 159 static struct swblock **swhash; 160 static int swhash_mask; 161 static int swap_async_max = 4; /* maximum in-progress async I/O's */ 162 163 extern struct vnode *swapdev_vp; /* from vm_swap.c */ 164 165 SYSCTL_INT(_vm, OID_AUTO, swap_async_max, 166 CTLFLAG_RW, &swap_async_max, 0, "Maximum running async swap ops"); 167 168 /* 169 * "named" and "unnamed" anon region objects. Try to reduce the overhead 170 * of searching a named list by hashing it just a little. 171 */ 172 173 #define NOBJLISTS 8 174 175 #define NOBJLIST(handle) \ 176 (&swap_pager_object_list[((int)(intptr_t)handle >> 4) & (NOBJLISTS-1)]) 177 178 static struct pagerlst swap_pager_object_list[NOBJLISTS]; 179 struct pagerlst swap_pager_un_object_list; 180 vm_zone_t swap_zone; 181 182 /* 183 * pagerops for OBJT_SWAP - "swap pager". Some ops are also global procedure 184 * calls hooked from other parts of the VM system and do not appear here. 185 * (see vm/swap_pager.h). 186 */ 187 188 static vm_object_t 189 swap_pager_alloc (void *handle, off_t size, 190 vm_prot_t prot, off_t offset); 191 static void swap_pager_dealloc (vm_object_t object); 192 static int swap_pager_getpages (vm_object_t, vm_page_t *, int, int); 193 static void swap_pager_init (void); 194 static void swap_pager_unswapped (vm_page_t); 195 static void swap_pager_strategy (vm_object_t, struct bio *); 196 static void swap_chain_iodone(struct bio *biox); 197 198 struct pagerops swappagerops = { 199 swap_pager_init, /* early system initialization of pager */ 200 swap_pager_alloc, /* allocate an OBJT_SWAP object */ 201 swap_pager_dealloc, /* deallocate an OBJT_SWAP object */ 202 swap_pager_getpages, /* pagein */ 203 swap_pager_putpages, /* pageout */ 204 swap_pager_haspage, /* get backing store status for page */ 205 swap_pager_unswapped, /* remove swap related to page */ 206 swap_pager_strategy /* pager strategy call */ 207 }; 208 209 /* 210 * dmmax is in page-sized chunks with the new swap system. It was 211 * dev-bsized chunks in the old. dmmax is always a power of 2. 212 * 213 * swap_*() routines are externally accessible. swp_*() routines are 214 * internal. 215 */ 216 217 int dmmax; 218 static int dmmax_mask; 219 int nswap_lowat = 128; /* in pages, swap_pager_almost_full warn */ 220 int nswap_hiwat = 512; /* in pages, swap_pager_almost_full warn */ 221 222 static __inline void swp_sizecheck (void); 223 static void swp_pager_async_iodone (struct bio *bio); 224 225 /* 226 * Swap bitmap functions 227 */ 228 229 static __inline void swp_pager_freeswapspace (daddr_t blk, int npages); 230 static __inline daddr_t swp_pager_getswapspace (int npages); 231 232 /* 233 * Metadata functions 234 */ 235 236 static void swp_pager_meta_build (vm_object_t, vm_pindex_t, daddr_t); 237 static void swp_pager_meta_free (vm_object_t, vm_pindex_t, daddr_t); 238 static void swp_pager_meta_free_all (vm_object_t); 239 static daddr_t swp_pager_meta_ctl (vm_object_t, vm_pindex_t, int); 240 241 /* 242 * SWP_SIZECHECK() - update swap_pager_full indication 243 * 244 * update the swap_pager_almost_full indication and warn when we are 245 * about to run out of swap space, using lowat/hiwat hysteresis. 246 * 247 * Clear swap_pager_full ( task killing ) indication when lowat is met. 248 * 249 * No restrictions on call 250 * This routine may not block. 251 * This routine must be called at splvm() 252 */ 253 254 static __inline void 255 swp_sizecheck(void) 256 { 257 if (vm_swap_size < nswap_lowat) { 258 if (swap_pager_almost_full == 0) { 259 kprintf("swap_pager: out of swap space\n"); 260 swap_pager_almost_full = 1; 261 } 262 } else { 263 swap_pager_full = 0; 264 if (vm_swap_size > nswap_hiwat) 265 swap_pager_almost_full = 0; 266 } 267 } 268 269 /* 270 * SWAP_PAGER_INIT() - initialize the swap pager! 271 * 272 * Expected to be started from system init. NOTE: This code is run 273 * before much else so be careful what you depend on. Most of the VM 274 * system has yet to be initialized at this point. 275 */ 276 277 static void 278 swap_pager_init(void) 279 { 280 /* 281 * Initialize object lists 282 */ 283 int i; 284 285 for (i = 0; i < NOBJLISTS; ++i) 286 TAILQ_INIT(&swap_pager_object_list[i]); 287 TAILQ_INIT(&swap_pager_un_object_list); 288 289 /* 290 * Device Stripe, in PAGE_SIZE'd blocks 291 */ 292 293 dmmax = SWB_NPAGES * 2; 294 dmmax_mask = ~(dmmax - 1); 295 } 296 297 /* 298 * SWAP_PAGER_SWAP_INIT() - swap pager initialization from pageout process 299 * 300 * Expected to be started from pageout process once, prior to entering 301 * its main loop. 302 */ 303 304 void 305 swap_pager_swap_init(void) 306 { 307 int n, n2; 308 309 /* 310 * Number of in-transit swap bp operations. Don't 311 * exhaust the pbufs completely. Make sure we 312 * initialize workable values (0 will work for hysteresis 313 * but it isn't very efficient). 314 * 315 * The nsw_cluster_max is constrained by the number of pages an XIO 316 * holds, i.e., (MAXPHYS/PAGE_SIZE) and our locally defined 317 * MAX_PAGEOUT_CLUSTER. Also be aware that swap ops are 318 * constrained by the swap device interleave stripe size. 319 * 320 * Currently we hardwire nsw_wcount_async to 4. This limit is 321 * designed to prevent other I/O from having high latencies due to 322 * our pageout I/O. The value 4 works well for one or two active swap 323 * devices but is probably a little low if you have more. Even so, 324 * a higher value would probably generate only a limited improvement 325 * with three or four active swap devices since the system does not 326 * typically have to pageout at extreme bandwidths. We will want 327 * at least 2 per swap devices, and 4 is a pretty good value if you 328 * have one NFS swap device due to the command/ack latency over NFS. 329 * So it all works out pretty well. 330 */ 331 332 nsw_cluster_max = min((MAXPHYS/PAGE_SIZE), MAX_PAGEOUT_CLUSTER); 333 334 nsw_rcount = (nswbuf + 1) / 2; 335 nsw_wcount_sync = (nswbuf + 3) / 4; 336 nsw_wcount_async = 4; 337 nsw_wcount_async_max = nsw_wcount_async; 338 339 /* 340 * The zone is dynamically allocated so generally size it to 341 * maxswzone (32MB to 512MB of KVM). Set a minimum size based 342 * on physical memory of around 8x (each swblock can hold 16 pages). 343 * 344 * With the advent of SSDs (vs HDs) the practical (swap:memory) ratio 345 * has increased dramatically. 346 */ 347 n = vmstats.v_page_count / 2; 348 if (maxswzone && n < maxswzone / sizeof(struct swblock)) 349 n = maxswzone / sizeof(struct swblock); 350 n2 = n; 351 352 do { 353 swap_zone = zinit( 354 "SWAPMETA", 355 sizeof(struct swblock), 356 n, 357 ZONE_INTERRUPT, 358 1); 359 if (swap_zone != NULL) 360 break; 361 /* 362 * if the allocation failed, try a zone two thirds the 363 * size of the previous attempt. 364 */ 365 n -= ((n + 2) / 3); 366 } while (n > 0); 367 368 if (swap_zone == NULL) 369 panic("swap_pager_swap_init: swap_zone == NULL"); 370 if (n2 != n) 371 kprintf("Swap zone entries reduced from %d to %d.\n", n2, n); 372 n2 = n; 373 374 /* 375 * Initialize our meta-data hash table. The swapper does not need to 376 * be quite as efficient as the VM system, so we do not use an 377 * oversized hash table. 378 * 379 * n: size of hash table, must be power of 2 380 * swhash_mask: hash table index mask 381 */ 382 383 for (n = 1; n < n2 / 8; n *= 2) 384 ; 385 386 swhash = kmalloc(sizeof(struct swblock *) * n, M_VMPGDATA, 387 M_WAITOK | M_ZERO); 388 389 swhash_mask = n - 1; 390 } 391 392 /* 393 * SWAP_PAGER_ALLOC() - allocate a new OBJT_SWAP VM object and instantiate 394 * its metadata structures. 395 * 396 * This routine is called from the mmap and fork code to create a new 397 * OBJT_SWAP object. We do this by creating an OBJT_DEFAULT object 398 * and then converting it with swp_pager_meta_build(). 399 * 400 * This routine may block in vm_object_allocate() and create a named 401 * object lookup race, so we must interlock. We must also run at 402 * splvm() for the object lookup to handle races with interrupts, but 403 * we do not have to maintain splvm() in between the lookup and the 404 * add because (I believe) it is not possible to attempt to create 405 * a new swap object w/handle when a default object with that handle 406 * already exists. 407 */ 408 409 static vm_object_t 410 swap_pager_alloc(void *handle, off_t size, vm_prot_t prot, off_t offset) 411 { 412 vm_object_t object; 413 414 if (handle) { 415 /* 416 * Reference existing named region or allocate new one. There 417 * should not be a race here against swp_pager_meta_build() 418 * as called from vm_page_remove() in regards to the lookup 419 * of the handle. 420 */ 421 422 while (sw_alloc_interlock) { 423 sw_alloc_interlock = -1; 424 tsleep(&sw_alloc_interlock, 0, "swpalc", 0); 425 } 426 sw_alloc_interlock = 1; 427 428 object = vm_pager_object_lookup(NOBJLIST(handle), handle); 429 430 if (object != NULL) { 431 vm_object_reference(object); 432 } else { 433 object = vm_object_allocate(OBJT_DEFAULT, 434 OFF_TO_IDX(offset + PAGE_MASK + size)); 435 object->handle = handle; 436 437 swp_pager_meta_build(object, 0, SWAPBLK_NONE); 438 } 439 440 if (sw_alloc_interlock < 0) 441 wakeup(&sw_alloc_interlock); 442 443 sw_alloc_interlock = 0; 444 } else { 445 object = vm_object_allocate(OBJT_DEFAULT, 446 OFF_TO_IDX(offset + PAGE_MASK + size)); 447 448 swp_pager_meta_build(object, 0, SWAPBLK_NONE); 449 } 450 451 return (object); 452 } 453 454 /* 455 * SWAP_PAGER_DEALLOC() - remove swap metadata from object 456 * 457 * The swap backing for the object is destroyed. The code is 458 * designed such that we can reinstantiate it later, but this 459 * routine is typically called only when the entire object is 460 * about to be destroyed. 461 * 462 * This routine may block, but no longer does. 463 * 464 * The object must be locked or unreferenceable. 465 */ 466 467 static void 468 swap_pager_dealloc(vm_object_t object) 469 { 470 /* 471 * Remove from list right away so lookups will fail if we block for 472 * pageout completion. 473 */ 474 475 if (object->handle == NULL) { 476 TAILQ_REMOVE(&swap_pager_un_object_list, object, pager_object_list); 477 } else { 478 TAILQ_REMOVE(NOBJLIST(object->handle), object, pager_object_list); 479 } 480 481 vm_object_pip_wait(object, "swpdea"); 482 483 /* 484 * Free all remaining metadata. We only bother to free it from 485 * the swap meta data. We do not attempt to free swapblk's still 486 * associated with vm_page_t's for this object. We do not care 487 * if paging is still in progress on some objects. 488 */ 489 crit_enter(); 490 swp_pager_meta_free_all(object); 491 crit_exit(); 492 } 493 494 /************************************************************************ 495 * SWAP PAGER BITMAP ROUTINES * 496 ************************************************************************/ 497 498 /* 499 * SWP_PAGER_GETSWAPSPACE() - allocate raw swap space 500 * 501 * Allocate swap for the requested number of pages. The starting 502 * swap block number (a page index) is returned or SWAPBLK_NONE 503 * if the allocation failed. 504 * 505 * Also has the side effect of advising that somebody made a mistake 506 * when they configured swap and didn't configure enough. 507 * 508 * Must be called at splvm() to avoid races with bitmap frees from 509 * vm_page_remove() aka swap_pager_page_removed(). 510 * 511 * This routine may not block 512 * This routine must be called at splvm(). 513 */ 514 515 static __inline daddr_t 516 swp_pager_getswapspace(int npages) 517 { 518 daddr_t blk; 519 520 if ((blk = blist_alloc(swapblist, npages)) == SWAPBLK_NONE) { 521 if (swap_pager_full != 2) { 522 kprintf("swap_pager_getswapspace: failed\n"); 523 swap_pager_full = 2; 524 swap_pager_almost_full = 1; 525 } 526 } else { 527 vm_swap_size -= npages; 528 swp_sizecheck(); 529 } 530 return(blk); 531 } 532 533 /* 534 * SWP_PAGER_FREESWAPSPACE() - free raw swap space 535 * 536 * This routine returns the specified swap blocks back to the bitmap. 537 * 538 * Note: This routine may not block (it could in the old swap code), 539 * and through the use of the new blist routines it does not block. 540 * 541 * We must be called at splvm() to avoid races with bitmap frees from 542 * vm_page_remove() aka swap_pager_page_removed(). 543 * 544 * This routine may not block 545 * This routine must be called at splvm(). 546 */ 547 548 static __inline void 549 swp_pager_freeswapspace(daddr_t blk, int npages) 550 { 551 blist_free(swapblist, blk, npages); 552 vm_swap_size += npages; 553 swp_sizecheck(); 554 } 555 556 /* 557 * SWAP_PAGER_FREESPACE() - frees swap blocks associated with a page 558 * range within an object. 559 * 560 * This is a globally accessible routine. 561 * 562 * This routine removes swapblk assignments from swap metadata. 563 * 564 * The external callers of this routine typically have already destroyed 565 * or renamed vm_page_t's associated with this range in the object so 566 * we should be ok. 567 * 568 * This routine may be called at any spl. We up our spl to splvm temporarily 569 * in order to perform the metadata removal. 570 */ 571 572 void 573 swap_pager_freespace(vm_object_t object, vm_pindex_t start, vm_size_t size) 574 { 575 crit_enter(); 576 swp_pager_meta_free(object, start, size); 577 crit_exit(); 578 } 579 580 /* 581 * SWAP_PAGER_RESERVE() - reserve swap blocks in object 582 * 583 * Assigns swap blocks to the specified range within the object. The 584 * swap blocks are not zerod. Any previous swap assignment is destroyed. 585 * 586 * Returns 0 on success, -1 on failure. 587 */ 588 589 int 590 swap_pager_reserve(vm_object_t object, vm_pindex_t start, vm_size_t size) 591 { 592 int n = 0; 593 daddr_t blk = SWAPBLK_NONE; 594 vm_pindex_t beg = start; /* save start index */ 595 596 crit_enter(); 597 while (size) { 598 if (n == 0) { 599 n = BLIST_MAX_ALLOC; 600 while ((blk = swp_pager_getswapspace(n)) == SWAPBLK_NONE) { 601 n >>= 1; 602 if (n == 0) { 603 swp_pager_meta_free(object, beg, start - beg); 604 crit_exit(); 605 return(-1); 606 } 607 } 608 } 609 swp_pager_meta_build(object, start, blk); 610 --size; 611 ++start; 612 ++blk; 613 --n; 614 } 615 swp_pager_meta_free(object, start, n); 616 crit_exit(); 617 return(0); 618 } 619 620 /* 621 * SWAP_PAGER_COPY() - copy blocks from source pager to destination pager 622 * and destroy the source. 623 * 624 * Copy any valid swapblks from the source to the destination. In 625 * cases where both the source and destination have a valid swapblk, 626 * we keep the destination's. 627 * 628 * This routine is allowed to block. It may block allocating metadata 629 * indirectly through swp_pager_meta_build() or if paging is still in 630 * progress on the source. 631 * 632 * This routine can be called at any spl 633 * 634 * XXX vm_page_collapse() kinda expects us not to block because we 635 * supposedly do not need to allocate memory, but for the moment we 636 * *may* have to get a little memory from the zone allocator, but 637 * it is taken from the interrupt memory. We should be ok. 638 * 639 * The source object contains no vm_page_t's (which is just as well) 640 * 641 * The source object is of type OBJT_SWAP. 642 * 643 * The source and destination objects must be locked or 644 * inaccessible (XXX are they ?) 645 */ 646 647 void 648 swap_pager_copy(vm_object_t srcobject, vm_object_t dstobject, 649 vm_pindex_t offset, int destroysource) 650 { 651 vm_pindex_t i; 652 653 crit_enter(); 654 655 /* 656 * If destroysource is set, we remove the source object from the 657 * swap_pager internal queue now. 658 */ 659 660 if (destroysource) { 661 if (srcobject->handle == NULL) { 662 TAILQ_REMOVE( 663 &swap_pager_un_object_list, 664 srcobject, 665 pager_object_list 666 ); 667 } else { 668 TAILQ_REMOVE( 669 NOBJLIST(srcobject->handle), 670 srcobject, 671 pager_object_list 672 ); 673 } 674 } 675 676 /* 677 * transfer source to destination. 678 */ 679 680 for (i = 0; i < dstobject->size; ++i) { 681 daddr_t dstaddr; 682 683 /* 684 * Locate (without changing) the swapblk on the destination, 685 * unless it is invalid in which case free it silently, or 686 * if the destination is a resident page, in which case the 687 * source is thrown away. 688 */ 689 690 dstaddr = swp_pager_meta_ctl(dstobject, i, 0); 691 692 if (dstaddr == SWAPBLK_NONE) { 693 /* 694 * Destination has no swapblk and is not resident, 695 * copy source. 696 */ 697 daddr_t srcaddr; 698 699 srcaddr = swp_pager_meta_ctl( 700 srcobject, 701 i + offset, 702 SWM_POP 703 ); 704 705 if (srcaddr != SWAPBLK_NONE) 706 swp_pager_meta_build(dstobject, i, srcaddr); 707 } else { 708 /* 709 * Destination has valid swapblk or it is represented 710 * by a resident page. We destroy the sourceblock. 711 */ 712 713 swp_pager_meta_ctl(srcobject, i + offset, SWM_FREE); 714 } 715 } 716 717 /* 718 * Free left over swap blocks in source. 719 * 720 * We have to revert the type to OBJT_DEFAULT so we do not accidently 721 * double-remove the object from the swap queues. 722 */ 723 724 if (destroysource) { 725 swp_pager_meta_free_all(srcobject); 726 /* 727 * Reverting the type is not necessary, the caller is going 728 * to destroy srcobject directly, but I'm doing it here 729 * for consistency since we've removed the object from its 730 * queues. 731 */ 732 srcobject->type = OBJT_DEFAULT; 733 } 734 crit_exit(); 735 } 736 737 /* 738 * SWAP_PAGER_HASPAGE() - determine if we have good backing store for 739 * the requested page. 740 * 741 * We determine whether good backing store exists for the requested 742 * page and return TRUE if it does, FALSE if it doesn't. 743 * 744 * If TRUE, we also try to determine how much valid, contiguous backing 745 * store exists before and after the requested page within a reasonable 746 * distance. We do not try to restrict it to the swap device stripe 747 * (that is handled in getpages/putpages). It probably isn't worth 748 * doing here. 749 */ 750 751 boolean_t 752 swap_pager_haspage(vm_object_t object, vm_pindex_t pindex, int *before, 753 int *after) 754 { 755 daddr_t blk0; 756 757 /* 758 * do we have good backing store at the requested index ? 759 */ 760 761 crit_enter(); 762 blk0 = swp_pager_meta_ctl(object, pindex, 0); 763 764 if (blk0 == SWAPBLK_NONE) { 765 crit_exit(); 766 if (before) 767 *before = 0; 768 if (after) 769 *after = 0; 770 return (FALSE); 771 } 772 773 /* 774 * find backwards-looking contiguous good backing store 775 */ 776 777 if (before != NULL) { 778 int i; 779 780 for (i = 1; i < (SWB_NPAGES/2); ++i) { 781 daddr_t blk; 782 783 if (i > pindex) 784 break; 785 blk = swp_pager_meta_ctl(object, pindex - i, 0); 786 if (blk != blk0 - i) 787 break; 788 } 789 *before = (i - 1); 790 } 791 792 /* 793 * find forward-looking contiguous good backing store 794 */ 795 796 if (after != NULL) { 797 int i; 798 799 for (i = 1; i < (SWB_NPAGES/2); ++i) { 800 daddr_t blk; 801 802 blk = swp_pager_meta_ctl(object, pindex + i, 0); 803 if (blk != blk0 + i) 804 break; 805 } 806 *after = (i - 1); 807 } 808 crit_exit(); 809 return (TRUE); 810 } 811 812 /* 813 * SWAP_PAGER_PAGE_UNSWAPPED() - remove swap backing store related to page 814 * 815 * This removes any associated swap backing store, whether valid or 816 * not, from the page. 817 * 818 * This routine is typically called when a page is made dirty, at 819 * which point any associated swap can be freed. MADV_FREE also 820 * calls us in a special-case situation 821 * 822 * NOTE!!! If the page is clean and the swap was valid, the caller 823 * should make the page dirty before calling this routine. This routine 824 * does NOT change the m->dirty status of the page. Also: MADV_FREE 825 * depends on it. 826 * 827 * This routine may not block 828 * This routine must be called at splvm() 829 */ 830 831 static void 832 swap_pager_unswapped(vm_page_t m) 833 { 834 swp_pager_meta_ctl(m->object, m->pindex, SWM_FREE); 835 } 836 837 /* 838 * SWAP_PAGER_STRATEGY() - read, write, free blocks 839 * 840 * This implements the vm_pager_strategy() interface to swap and allows 841 * other parts of the system to directly access swap as backing store 842 * through vm_objects of type OBJT_SWAP. This is intended to be a 843 * cacheless interface ( i.e. caching occurs at higher levels ). 844 * Therefore we do not maintain any resident pages. All I/O goes 845 * directly to and from the swap device. 846 * 847 * We currently attempt to run I/O synchronously or asynchronously as 848 * the caller requests. This isn't perfect because we loose error 849 * sequencing when we run multiple ops in parallel to satisfy a request. 850 * But this is swap, so we let it all hang out. 851 */ 852 853 static void 854 swap_pager_strategy(vm_object_t object, struct bio *bio) 855 { 856 struct buf *bp = bio->bio_buf; 857 struct bio *nbio; 858 vm_pindex_t start; 859 vm_pindex_t biox_blkno = 0; 860 int count; 861 char *data; 862 struct bio *biox; 863 struct buf *bufx; 864 struct bio_track *track; 865 866 /* 867 * tracking for swapdev vnode I/Os 868 */ 869 if (bp->b_cmd == BUF_CMD_READ) 870 track = &swapdev_vp->v_track_read; 871 else 872 track = &swapdev_vp->v_track_write; 873 874 if (bp->b_bcount & PAGE_MASK) { 875 bp->b_error = EINVAL; 876 bp->b_flags |= B_ERROR | B_INVAL; 877 biodone(bio); 878 kprintf("swap_pager_strategy: bp %p offset %lld size %d, " 879 "not page bounded\n", 880 bp, (long long)bio->bio_offset, (int)bp->b_bcount); 881 return; 882 } 883 884 /* 885 * Clear error indication, initialize page index, count, data pointer. 886 */ 887 bp->b_error = 0; 888 bp->b_flags &= ~B_ERROR; 889 bp->b_resid = bp->b_bcount; 890 891 start = (vm_pindex_t)(bio->bio_offset >> PAGE_SHIFT); 892 count = howmany(bp->b_bcount, PAGE_SIZE); 893 data = bp->b_data; 894 895 /* 896 * Deal with BUF_CMD_FREEBLKS 897 */ 898 if (bp->b_cmd == BUF_CMD_FREEBLKS) { 899 /* 900 * FREE PAGE(s) - destroy underlying swap that is no longer 901 * needed. 902 */ 903 swp_pager_meta_free(object, start, count); 904 bp->b_resid = 0; 905 biodone(bio); 906 return; 907 } 908 909 /* 910 * We need to be able to create a new cluster of I/O's. We cannot 911 * use the caller fields of the passed bio so push a new one. 912 * 913 * Because nbio is just a placeholder for the cluster links, 914 * we can biodone() the original bio instead of nbio to make 915 * things a bit more efficient. 916 */ 917 nbio = push_bio(bio); 918 nbio->bio_offset = bio->bio_offset; 919 nbio->bio_caller_info1.cluster_head = NULL; 920 nbio->bio_caller_info2.cluster_tail = NULL; 921 922 biox = NULL; 923 bufx = NULL; 924 925 /* 926 * Execute read or write 927 */ 928 while (count > 0) { 929 daddr_t blk; 930 931 /* 932 * Obtain block. If block not found and writing, allocate a 933 * new block and build it into the object. 934 */ 935 blk = swp_pager_meta_ctl(object, start, 0); 936 if ((blk == SWAPBLK_NONE) && bp->b_cmd != BUF_CMD_READ) { 937 blk = swp_pager_getswapspace(1); 938 if (blk == SWAPBLK_NONE) { 939 bp->b_error = ENOMEM; 940 bp->b_flags |= B_ERROR; 941 break; 942 } 943 swp_pager_meta_build(object, start, blk); 944 } 945 946 /* 947 * Do we have to flush our current collection? Yes if: 948 * 949 * - no swap block at this index 950 * - swap block is not contiguous 951 * - we cross a physical disk boundry in the 952 * stripe. 953 */ 954 if ( 955 biox && (biox_blkno + btoc(bufx->b_bcount) != blk || 956 ((biox_blkno ^ blk) & dmmax_mask) 957 ) 958 ) { 959 if (bp->b_cmd == BUF_CMD_READ) { 960 ++mycpu->gd_cnt.v_swapin; 961 mycpu->gd_cnt.v_swappgsin += btoc(bufx->b_bcount); 962 } else { 963 ++mycpu->gd_cnt.v_swapout; 964 mycpu->gd_cnt.v_swappgsout += btoc(bufx->b_bcount); 965 bufx->b_dirtyend = bufx->b_bcount; 966 } 967 968 /* 969 * Finished with this buf. 970 */ 971 KKASSERT(bufx->b_bcount != 0); 972 if (bufx->b_cmd != BUF_CMD_READ) 973 bufx->b_dirtyend = bufx->b_bcount; 974 biox = NULL; 975 bufx = NULL; 976 } 977 978 /* 979 * Add new swapblk to biox, instantiating biox if necessary. 980 * Zero-fill reads are able to take a shortcut. 981 */ 982 if (blk == SWAPBLK_NONE) { 983 /* 984 * We can only get here if we are reading. Since 985 * we are at splvm() we can safely modify b_resid, 986 * even if chain ops are in progress. 987 */ 988 bzero(data, PAGE_SIZE); 989 bp->b_resid -= PAGE_SIZE; 990 } else { 991 if (biox == NULL) { 992 /* XXX chain count > 4, wait to <= 4 */ 993 994 bufx = getpbuf(NULL); 995 biox = &bufx->b_bio1; 996 cluster_append(nbio, bufx); 997 bufx->b_flags |= (bufx->b_flags & B_ORDERED); 998 bufx->b_cmd = bp->b_cmd; 999 biox->bio_done = swap_chain_iodone; 1000 biox->bio_offset = (off_t)blk << PAGE_SHIFT; 1001 biox->bio_caller_info1.cluster_parent = nbio; 1002 biox_blkno = blk; 1003 bufx->b_bcount = 0; 1004 bufx->b_data = data; 1005 } 1006 bufx->b_bcount += PAGE_SIZE; 1007 } 1008 --count; 1009 ++start; 1010 data += PAGE_SIZE; 1011 } 1012 1013 /* 1014 * Flush out last buffer 1015 */ 1016 if (biox) { 1017 if (bufx->b_cmd == BUF_CMD_READ) { 1018 ++mycpu->gd_cnt.v_swapin; 1019 mycpu->gd_cnt.v_swappgsin += btoc(bufx->b_bcount); 1020 } else { 1021 ++mycpu->gd_cnt.v_swapout; 1022 mycpu->gd_cnt.v_swappgsout += btoc(bufx->b_bcount); 1023 bufx->b_dirtyend = bufx->b_bcount; 1024 } 1025 KKASSERT(bufx->b_bcount); 1026 if (bufx->b_cmd != BUF_CMD_READ) 1027 bufx->b_dirtyend = bufx->b_bcount; 1028 /* biox, bufx = NULL */ 1029 } 1030 1031 /* 1032 * Now initiate all the I/O. Be careful looping on our chain as 1033 * I/O's may complete while we are still initiating them. 1034 */ 1035 nbio->bio_caller_info2.cluster_tail = NULL; 1036 bufx = nbio->bio_caller_info1.cluster_head; 1037 1038 while (bufx) { 1039 biox = &bufx->b_bio1; 1040 BUF_KERNPROC(bufx); 1041 bufx = bufx->b_cluster_next; 1042 vn_strategy(swapdev_vp, biox); 1043 } 1044 1045 /* 1046 * Completion of the cluster will also call biodone_chain(nbio). 1047 * We never call biodone(nbio) so we don't have to worry about 1048 * setting up a bio_done callback. It's handled in the sub-IO. 1049 */ 1050 /**/ 1051 } 1052 1053 static void 1054 swap_chain_iodone(struct bio *biox) 1055 { 1056 struct buf **nextp; 1057 struct buf *bufx; /* chained sub-buffer */ 1058 struct bio *nbio; /* parent nbio with chain glue */ 1059 struct buf *bp; /* original bp associated with nbio */ 1060 int chain_empty; 1061 1062 bufx = biox->bio_buf; 1063 nbio = biox->bio_caller_info1.cluster_parent; 1064 bp = nbio->bio_buf; 1065 1066 /* 1067 * Update the original buffer 1068 */ 1069 KKASSERT(bp != NULL); 1070 if (bufx->b_flags & B_ERROR) { 1071 atomic_set_int(&bufx->b_flags, B_ERROR); 1072 bp->b_error = bufx->b_error; 1073 } else if (bufx->b_resid != 0) { 1074 atomic_set_int(&bufx->b_flags, B_ERROR); 1075 bp->b_error = EINVAL; 1076 } else { 1077 atomic_subtract_int(&bp->b_resid, bufx->b_bcount); 1078 } 1079 1080 /* 1081 * Remove us from the chain. 1082 */ 1083 spin_lock_wr(&bp->b_lock.lk_spinlock); 1084 nextp = &nbio->bio_caller_info1.cluster_head; 1085 while (*nextp != bufx) { 1086 KKASSERT(*nextp != NULL); 1087 nextp = &(*nextp)->b_cluster_next; 1088 } 1089 *nextp = bufx->b_cluster_next; 1090 chain_empty = (nbio->bio_caller_info1.cluster_head == NULL); 1091 spin_unlock_wr(&bp->b_lock.lk_spinlock); 1092 1093 /* 1094 * Clean up bufx. If the chain is now empty we finish out 1095 * the parent. Note that we may be racing other completions 1096 * so we must use the chain_empty status from above. 1097 */ 1098 if (chain_empty) { 1099 if (bp->b_resid != 0 && !(bp->b_flags & B_ERROR)) { 1100 atomic_set_int(&bp->b_flags, B_ERROR); 1101 bp->b_error = EINVAL; 1102 } 1103 biodone_chain(nbio); 1104 } 1105 relpbuf(bufx, NULL); 1106 } 1107 1108 /* 1109 * SWAP_PAGER_GETPAGES() - bring pages in from swap 1110 * 1111 * Attempt to retrieve (m, count) pages from backing store, but make 1112 * sure we retrieve at least m[reqpage]. We try to load in as large 1113 * a chunk surrounding m[reqpage] as is contiguous in swap and which 1114 * belongs to the same object. 1115 * 1116 * The code is designed for asynchronous operation and 1117 * immediate-notification of 'reqpage' but tends not to be 1118 * used that way. Please do not optimize-out this algorithmic 1119 * feature, I intend to improve on it in the future. 1120 * 1121 * The parent has a single vm_object_pip_add() reference prior to 1122 * calling us and we should return with the same. 1123 * 1124 * The parent has BUSY'd the pages. We should return with 'm' 1125 * left busy, but the others adjusted. 1126 */ 1127 1128 static int 1129 swap_pager_getpages(vm_object_t object, vm_page_t *m, int count, int reqpage) 1130 { 1131 struct buf *bp; 1132 struct bio *bio; 1133 vm_page_t mreq; 1134 int i; 1135 int j; 1136 daddr_t blk; 1137 vm_offset_t kva; 1138 vm_pindex_t lastpindex; 1139 1140 mreq = m[reqpage]; 1141 1142 if (mreq->object != object) { 1143 panic("swap_pager_getpages: object mismatch %p/%p", 1144 object, 1145 mreq->object 1146 ); 1147 } 1148 1149 /* 1150 * Calculate range to retrieve. The pages have already been assigned 1151 * their swapblks. We require a *contiguous* range that falls entirely 1152 * within a single device stripe. If we do not supply it, bad things 1153 * happen. Note that blk, iblk & jblk can be SWAPBLK_NONE, but the 1154 * loops are set up such that the case(s) are handled implicitly. 1155 * 1156 * The swp_*() calls must be made at splvm(). vm_page_free() does 1157 * not need to be, but it will go a little faster if it is. 1158 */ 1159 crit_enter(); 1160 blk = swp_pager_meta_ctl(mreq->object, mreq->pindex, 0); 1161 1162 for (i = reqpage - 1; i >= 0; --i) { 1163 daddr_t iblk; 1164 1165 iblk = swp_pager_meta_ctl(m[i]->object, m[i]->pindex, 0); 1166 if (blk != iblk + (reqpage - i)) 1167 break; 1168 if ((blk ^ iblk) & dmmax_mask) 1169 break; 1170 } 1171 ++i; 1172 1173 for (j = reqpage + 1; j < count; ++j) { 1174 daddr_t jblk; 1175 1176 jblk = swp_pager_meta_ctl(m[j]->object, m[j]->pindex, 0); 1177 if (blk != jblk - (j - reqpage)) 1178 break; 1179 if ((blk ^ jblk) & dmmax_mask) 1180 break; 1181 } 1182 1183 /* 1184 * free pages outside our collection range. Note: we never free 1185 * mreq, it must remain busy throughout. 1186 */ 1187 1188 { 1189 int k; 1190 1191 for (k = 0; k < i; ++k) 1192 vm_page_free(m[k]); 1193 for (k = j; k < count; ++k) 1194 vm_page_free(m[k]); 1195 } 1196 crit_exit(); 1197 1198 1199 /* 1200 * Return VM_PAGER_FAIL if we have nothing to do. Return mreq 1201 * still busy, but the others unbusied. 1202 */ 1203 1204 if (blk == SWAPBLK_NONE) 1205 return(VM_PAGER_FAIL); 1206 1207 /* 1208 * Get a swap buffer header to perform the IO 1209 */ 1210 1211 bp = getpbuf(&nsw_rcount); 1212 bio = &bp->b_bio1; 1213 kva = (vm_offset_t) bp->b_data; 1214 1215 /* 1216 * map our page(s) into kva for input 1217 */ 1218 1219 pmap_qenter(kva, m + i, j - i); 1220 1221 bp->b_data = (caddr_t) kva; 1222 bp->b_bcount = PAGE_SIZE * (j - i); 1223 bio->bio_done = swp_pager_async_iodone; 1224 bio->bio_offset = (off_t)(blk - (reqpage - i)) << PAGE_SHIFT; 1225 bio->bio_driver_info = (void *)(intptr_t)(reqpage - i); 1226 bio->bio_caller_info1.index = SWBIO_READ; 1227 1228 { 1229 int k; 1230 1231 for (k = i; k < j; ++k) { 1232 bp->b_xio.xio_pages[k - i] = m[k]; 1233 vm_page_flag_set(m[k], PG_SWAPINPROG); 1234 } 1235 } 1236 bp->b_xio.xio_npages = j - i; 1237 1238 mycpu->gd_cnt.v_swapin++; 1239 mycpu->gd_cnt.v_swappgsin += bp->b_xio.xio_npages; 1240 1241 /* 1242 * We still hold the lock on mreq, and our automatic completion routine 1243 * does not remove it. 1244 */ 1245 1246 vm_object_pip_add(mreq->object, bp->b_xio.xio_npages); 1247 lastpindex = m[j-1]->pindex; 1248 1249 /* 1250 * perform the I/O. NOTE!!! bp cannot be considered valid after 1251 * this point because we automatically release it on completion. 1252 * Instead, we look at the one page we are interested in which we 1253 * still hold a lock on even through the I/O completion. 1254 * 1255 * The other pages in our m[] array are also released on completion, 1256 * so we cannot assume they are valid anymore either. 1257 */ 1258 1259 bp->b_cmd = BUF_CMD_READ; 1260 BUF_KERNPROC(bp); 1261 vn_strategy(swapdev_vp, bio); 1262 1263 /* 1264 * wait for the page we want to complete. PG_SWAPINPROG is always 1265 * cleared on completion. If an I/O error occurs, SWAPBLK_NONE 1266 * is set in the meta-data. 1267 */ 1268 1269 crit_enter(); 1270 1271 while ((mreq->flags & PG_SWAPINPROG) != 0) { 1272 vm_page_flag_set(mreq, PG_WANTED | PG_REFERENCED); 1273 mycpu->gd_cnt.v_intrans++; 1274 if (tsleep(mreq, 0, "swread", hz*20)) { 1275 kprintf( 1276 "swap_pager: indefinite wait buffer: " 1277 " offset: %lld, size: %ld\n", 1278 (long long)bio->bio_offset, 1279 (long)bp->b_bcount 1280 ); 1281 } 1282 } 1283 1284 crit_exit(); 1285 1286 /* 1287 * mreq is left bussied after completion, but all the other pages 1288 * are freed. If we had an unrecoverable read error the page will 1289 * not be valid. 1290 */ 1291 1292 if (mreq->valid != VM_PAGE_BITS_ALL) { 1293 return(VM_PAGER_ERROR); 1294 } else { 1295 return(VM_PAGER_OK); 1296 } 1297 1298 /* 1299 * A final note: in a low swap situation, we cannot deallocate swap 1300 * and mark a page dirty here because the caller is likely to mark 1301 * the page clean when we return, causing the page to possibly revert 1302 * to all-zero's later. 1303 */ 1304 } 1305 1306 /* 1307 * swap_pager_putpages: 1308 * 1309 * Assign swap (if necessary) and initiate I/O on the specified pages. 1310 * 1311 * We support both OBJT_DEFAULT and OBJT_SWAP objects. DEFAULT objects 1312 * are automatically converted to SWAP objects. 1313 * 1314 * In a low memory situation we may block in vn_strategy(), but the new 1315 * vm_page reservation system coupled with properly written VFS devices 1316 * should ensure that no low-memory deadlock occurs. This is an area 1317 * which needs work. 1318 * 1319 * The parent has N vm_object_pip_add() references prior to 1320 * calling us and will remove references for rtvals[] that are 1321 * not set to VM_PAGER_PEND. We need to remove the rest on I/O 1322 * completion. 1323 * 1324 * The parent has soft-busy'd the pages it passes us and will unbusy 1325 * those whos rtvals[] entry is not set to VM_PAGER_PEND on return. 1326 * We need to unbusy the rest on I/O completion. 1327 */ 1328 void 1329 swap_pager_putpages(vm_object_t object, vm_page_t *m, int count, 1330 boolean_t sync, int *rtvals) 1331 { 1332 int i; 1333 int n = 0; 1334 1335 if (count && m[0]->object != object) { 1336 panic("swap_pager_getpages: object mismatch %p/%p", 1337 object, 1338 m[0]->object 1339 ); 1340 } 1341 1342 /* 1343 * Step 1 1344 * 1345 * Turn object into OBJT_SWAP 1346 * check for bogus sysops 1347 * force sync if not pageout process 1348 */ 1349 1350 if (object->type != OBJT_SWAP) 1351 swp_pager_meta_build(object, 0, SWAPBLK_NONE); 1352 1353 if (curthread != pagethread) 1354 sync = TRUE; 1355 1356 /* 1357 * Step 2 1358 * 1359 * Update nsw parameters from swap_async_max sysctl values. 1360 * Do not let the sysop crash the machine with bogus numbers. 1361 */ 1362 1363 if (swap_async_max != nsw_wcount_async_max) { 1364 int n; 1365 1366 /* 1367 * limit range 1368 */ 1369 if ((n = swap_async_max) > nswbuf / 2) 1370 n = nswbuf / 2; 1371 if (n < 1) 1372 n = 1; 1373 swap_async_max = n; 1374 1375 /* 1376 * Adjust difference ( if possible ). If the current async 1377 * count is too low, we may not be able to make the adjustment 1378 * at this time. 1379 */ 1380 crit_enter(); 1381 n -= nsw_wcount_async_max; 1382 if (nsw_wcount_async + n >= 0) { 1383 nsw_wcount_async += n; 1384 nsw_wcount_async_max += n; 1385 wakeup(&nsw_wcount_async); 1386 } 1387 crit_exit(); 1388 } 1389 1390 /* 1391 * Step 3 1392 * 1393 * Assign swap blocks and issue I/O. We reallocate swap on the fly. 1394 * The page is left dirty until the pageout operation completes 1395 * successfully. 1396 */ 1397 1398 for (i = 0; i < count; i += n) { 1399 struct buf *bp; 1400 struct bio *bio; 1401 daddr_t blk; 1402 int j; 1403 1404 /* 1405 * Maximum I/O size is limited by a number of factors. 1406 */ 1407 1408 n = min(BLIST_MAX_ALLOC, count - i); 1409 n = min(n, nsw_cluster_max); 1410 1411 crit_enter(); 1412 1413 /* 1414 * Get biggest block of swap we can. If we fail, fall 1415 * back and try to allocate a smaller block. Don't go 1416 * overboard trying to allocate space if it would overly 1417 * fragment swap. 1418 */ 1419 while ( 1420 (blk = swp_pager_getswapspace(n)) == SWAPBLK_NONE && 1421 n > 4 1422 ) { 1423 n >>= 1; 1424 } 1425 if (blk == SWAPBLK_NONE) { 1426 for (j = 0; j < n; ++j) 1427 rtvals[i+j] = VM_PAGER_FAIL; 1428 crit_exit(); 1429 continue; 1430 } 1431 1432 /* 1433 * The I/O we are constructing cannot cross a physical 1434 * disk boundry in the swap stripe. Note: we are still 1435 * at splvm(). 1436 */ 1437 if ((blk ^ (blk + n)) & dmmax_mask) { 1438 j = ((blk + dmmax) & dmmax_mask) - blk; 1439 swp_pager_freeswapspace(blk + j, n - j); 1440 n = j; 1441 } 1442 1443 /* 1444 * All I/O parameters have been satisfied, build the I/O 1445 * request and assign the swap space. 1446 */ 1447 1448 if (sync == TRUE) 1449 bp = getpbuf(&nsw_wcount_sync); 1450 else 1451 bp = getpbuf(&nsw_wcount_async); 1452 bio = &bp->b_bio1; 1453 1454 pmap_qenter((vm_offset_t)bp->b_data, &m[i], n); 1455 1456 bp->b_bcount = PAGE_SIZE * n; 1457 bio->bio_offset = (off_t)blk << PAGE_SHIFT; 1458 1459 for (j = 0; j < n; ++j) { 1460 vm_page_t mreq = m[i+j]; 1461 1462 swp_pager_meta_build( 1463 mreq->object, 1464 mreq->pindex, 1465 blk + j 1466 ); 1467 vm_page_dirty(mreq); 1468 rtvals[i+j] = VM_PAGER_OK; 1469 1470 vm_page_flag_set(mreq, PG_SWAPINPROG); 1471 bp->b_xio.xio_pages[j] = mreq; 1472 } 1473 bp->b_xio.xio_npages = n; 1474 1475 mycpu->gd_cnt.v_swapout++; 1476 mycpu->gd_cnt.v_swappgsout += bp->b_xio.xio_npages; 1477 1478 crit_exit(); 1479 1480 bp->b_dirtyoff = 0; /* req'd for NFS */ 1481 bp->b_dirtyend = bp->b_bcount; /* req'd for NFS */ 1482 bp->b_cmd = BUF_CMD_WRITE; 1483 bio->bio_caller_info1.index = SWBIO_WRITE; 1484 1485 /* 1486 * asynchronous 1487 */ 1488 if (sync == FALSE) { 1489 bio->bio_done = swp_pager_async_iodone; 1490 BUF_KERNPROC(bp); 1491 vn_strategy(swapdev_vp, bio); 1492 1493 for (j = 0; j < n; ++j) 1494 rtvals[i+j] = VM_PAGER_PEND; 1495 continue; 1496 } 1497 1498 /* 1499 * Issue synchrnously. 1500 * 1501 * Wait for the sync I/O to complete, then update rtvals. 1502 * We just set the rtvals[] to VM_PAGER_PEND so we can call 1503 * our async completion routine at the end, thus avoiding a 1504 * double-free. 1505 */ 1506 bio->bio_caller_info1.index |= SWBIO_SYNC; 1507 bio->bio_done = biodone_sync; 1508 bio->bio_flags |= BIO_SYNC; 1509 vn_strategy(swapdev_vp, bio); 1510 biowait(bio, "swwrt"); 1511 1512 for (j = 0; j < n; ++j) 1513 rtvals[i+j] = VM_PAGER_PEND; 1514 1515 /* 1516 * Now that we are through with the bp, we can call the 1517 * normal async completion, which frees everything up. 1518 */ 1519 swp_pager_async_iodone(bio); 1520 } 1521 } 1522 1523 void 1524 swap_pager_newswap(void) 1525 { 1526 swp_sizecheck(); 1527 } 1528 1529 /* 1530 * swp_pager_async_iodone: 1531 * 1532 * Completion routine for asynchronous reads and writes from/to swap. 1533 * Also called manually by synchronous code to finish up a bp. 1534 * 1535 * For READ operations, the pages are PG_BUSY'd. For WRITE operations, 1536 * the pages are vm_page_t->busy'd. For READ operations, we PG_BUSY 1537 * unbusy all pages except the 'main' request page. For WRITE 1538 * operations, we vm_page_t->busy'd unbusy all pages ( we can do this 1539 * because we marked them all VM_PAGER_PEND on return from putpages ). 1540 * 1541 * This routine may not block. 1542 */ 1543 static void 1544 swp_pager_async_iodone(struct bio *bio) 1545 { 1546 struct buf *bp = bio->bio_buf; 1547 vm_object_t object = NULL; 1548 int i; 1549 int *nswptr; 1550 1551 /* 1552 * report error 1553 */ 1554 if (bp->b_flags & B_ERROR) { 1555 kprintf( 1556 "swap_pager: I/O error - %s failed; offset %lld," 1557 "size %ld, error %d\n", 1558 ((bio->bio_caller_info1.index & SWBIO_READ) ? 1559 "pagein" : "pageout"), 1560 (long long)bio->bio_offset, 1561 (long)bp->b_bcount, 1562 bp->b_error 1563 ); 1564 } 1565 1566 /* 1567 * set object, raise to splvm(). 1568 */ 1569 if (bp->b_xio.xio_npages) 1570 object = bp->b_xio.xio_pages[0]->object; 1571 crit_enter(); 1572 1573 /* 1574 * remove the mapping for kernel virtual 1575 */ 1576 pmap_qremove((vm_offset_t)bp->b_data, bp->b_xio.xio_npages); 1577 1578 /* 1579 * cleanup pages. If an error occurs writing to swap, we are in 1580 * very serious trouble. If it happens to be a disk error, though, 1581 * we may be able to recover by reassigning the swap later on. So 1582 * in this case we remove the m->swapblk assignment for the page 1583 * but do not free it in the rlist. The errornous block(s) are thus 1584 * never reallocated as swap. Redirty the page and continue. 1585 */ 1586 for (i = 0; i < bp->b_xio.xio_npages; ++i) { 1587 vm_page_t m = bp->b_xio.xio_pages[i]; 1588 1589 vm_page_flag_clear(m, PG_SWAPINPROG); 1590 1591 if (bp->b_flags & B_ERROR) { 1592 /* 1593 * If an error occurs I'd love to throw the swapblk 1594 * away without freeing it back to swapspace, so it 1595 * can never be used again. But I can't from an 1596 * interrupt. 1597 */ 1598 1599 if (bio->bio_caller_info1.index & SWBIO_READ) { 1600 /* 1601 * When reading, reqpage needs to stay 1602 * locked for the parent, but all other 1603 * pages can be freed. We still want to 1604 * wakeup the parent waiting on the page, 1605 * though. ( also: pg_reqpage can be -1 and 1606 * not match anything ). 1607 * 1608 * We have to wake specifically requested pages 1609 * up too because we cleared PG_SWAPINPROG and 1610 * someone may be waiting for that. 1611 * 1612 * NOTE: for reads, m->dirty will probably 1613 * be overridden by the original caller of 1614 * getpages so don't play cute tricks here. 1615 * 1616 * NOTE: We can't actually free the page from 1617 * here, because this is an interrupt. It 1618 * is not legal to mess with object->memq 1619 * from an interrupt. Deactivate the page 1620 * instead. 1621 */ 1622 1623 m->valid = 0; 1624 vm_page_flag_clear(m, PG_ZERO); 1625 1626 /* 1627 * bio_driver_info holds the requested page 1628 * index. 1629 */ 1630 if (i != (int)(intptr_t)bio->bio_driver_info) { 1631 vm_page_deactivate(m); 1632 vm_page_wakeup(m); 1633 } else { 1634 vm_page_flash(m); 1635 } 1636 /* 1637 * If i == bp->b_pager.pg_reqpage, do not wake 1638 * the page up. The caller needs to. 1639 */ 1640 } else { 1641 /* 1642 * If a write error occurs, reactivate page 1643 * so it doesn't clog the inactive list, 1644 * then finish the I/O. 1645 */ 1646 vm_page_dirty(m); 1647 vm_page_activate(m); 1648 vm_page_io_finish(m); 1649 } 1650 } else if (bio->bio_caller_info1.index & SWBIO_READ) { 1651 /* 1652 * NOTE: for reads, m->dirty will probably be 1653 * overridden by the original caller of getpages so 1654 * we cannot set them in order to free the underlying 1655 * swap in a low-swap situation. I don't think we'd 1656 * want to do that anyway, but it was an optimization 1657 * that existed in the old swapper for a time before 1658 * it got ripped out due to precisely this problem. 1659 * 1660 * clear PG_ZERO in page. 1661 * 1662 * If not the requested page then deactivate it. 1663 * 1664 * Note that the requested page, reqpage, is left 1665 * busied, but we still have to wake it up. The 1666 * other pages are released (unbusied) by 1667 * vm_page_wakeup(). We do not set reqpage's 1668 * valid bits here, it is up to the caller. 1669 */ 1670 1671 /* 1672 * NOTE: can't call pmap_clear_modify(m) from an 1673 * interrupt thread, the pmap code may have to map 1674 * non-kernel pmaps and currently asserts the case. 1675 */ 1676 /*pmap_clear_modify(m);*/ 1677 m->valid = VM_PAGE_BITS_ALL; 1678 vm_page_undirty(m); 1679 vm_page_flag_clear(m, PG_ZERO); 1680 1681 /* 1682 * We have to wake specifically requested pages 1683 * up too because we cleared PG_SWAPINPROG and 1684 * could be waiting for it in getpages. However, 1685 * be sure to not unbusy getpages specifically 1686 * requested page - getpages expects it to be 1687 * left busy. 1688 * 1689 * bio_driver_info holds the requested page 1690 */ 1691 if (i != (int)(intptr_t)bio->bio_driver_info) { 1692 vm_page_deactivate(m); 1693 vm_page_wakeup(m); 1694 } else { 1695 vm_page_flash(m); 1696 } 1697 } else { 1698 /* 1699 * Mark the page clean but do not mess with the 1700 * pmap-layer's modified state. That state should 1701 * also be clear since the caller protected the 1702 * page VM_PROT_READ, but allow the case. 1703 * 1704 * We are in an interrupt, avoid pmap operations. 1705 * 1706 * If we have a severe page deficit, deactivate the 1707 * page. Do not try to cache it (which would also 1708 * involve a pmap op), because the page might still 1709 * be read-heavy. 1710 */ 1711 vm_page_undirty(m); 1712 vm_page_io_finish(m); 1713 if (vm_page_count_severe()) 1714 vm_page_deactivate(m); 1715 #if 0 1716 if (!vm_page_count_severe() || !vm_page_try_to_cache(m)) 1717 vm_page_protect(m, VM_PROT_READ); 1718 #endif 1719 } 1720 } 1721 1722 /* 1723 * adjust pip. NOTE: the original parent may still have its own 1724 * pip refs on the object. 1725 */ 1726 1727 if (object) 1728 vm_object_pip_wakeupn(object, bp->b_xio.xio_npages); 1729 1730 /* 1731 * Release the physical I/O buffer. 1732 * 1733 * NOTE: Due to synchronous operations in the write case b_cmd may 1734 * already be set to BUF_CMD_DONE and BIO_SYNC may have already 1735 * been cleared. 1736 */ 1737 if (bio->bio_caller_info1.index & SWBIO_READ) 1738 nswptr = &nsw_rcount; 1739 else if (bio->bio_caller_info1.index & SWBIO_SYNC) 1740 nswptr = &nsw_wcount_sync; 1741 else 1742 nswptr = &nsw_wcount_async; 1743 bp->b_cmd = BUF_CMD_DONE; 1744 relpbuf(bp, nswptr); 1745 crit_exit(); 1746 } 1747 1748 /************************************************************************ 1749 * SWAP META DATA * 1750 ************************************************************************ 1751 * 1752 * These routines manipulate the swap metadata stored in the 1753 * OBJT_SWAP object. All swp_*() routines must be called at 1754 * splvm() because swap can be freed up by the low level vm_page 1755 * code which might be called from interrupts beyond what splbio() covers. 1756 * 1757 * Swap metadata is implemented with a global hash and not directly 1758 * linked into the object. Instead the object simply contains 1759 * appropriate tracking counters. 1760 */ 1761 1762 /* 1763 * SWP_PAGER_HASH() - hash swap meta data 1764 * 1765 * This is an inline helper function which hashes the swapblk given 1766 * the object and page index. It returns a pointer to a pointer 1767 * to the object, or a pointer to a NULL pointer if it could not 1768 * find a swapblk. 1769 * 1770 * This routine must be called at splvm(). 1771 */ 1772 1773 static __inline struct swblock ** 1774 swp_pager_hash(vm_object_t object, vm_pindex_t index) 1775 { 1776 struct swblock **pswap; 1777 struct swblock *swap; 1778 1779 index &= ~SWAP_META_MASK; 1780 pswap = &swhash[(index ^ (int)(intptr_t)object) & swhash_mask]; 1781 1782 while ((swap = *pswap) != NULL) { 1783 if (swap->swb_object == object && 1784 swap->swb_index == index 1785 ) { 1786 break; 1787 } 1788 pswap = &swap->swb_hnext; 1789 } 1790 return(pswap); 1791 } 1792 1793 /* 1794 * SWP_PAGER_META_BUILD() - add swap block to swap meta data for object 1795 * 1796 * We first convert the object to a swap object if it is a default 1797 * object. 1798 * 1799 * The specified swapblk is added to the object's swap metadata. If 1800 * the swapblk is not valid, it is freed instead. Any previously 1801 * assigned swapblk is freed. 1802 * 1803 * This routine must be called at splvm(), except when used to convert 1804 * an OBJT_DEFAULT object into an OBJT_SWAP object. 1805 1806 */ 1807 1808 static void 1809 swp_pager_meta_build( 1810 vm_object_t object, 1811 vm_pindex_t index, 1812 daddr_t swapblk 1813 ) { 1814 struct swblock *swap; 1815 struct swblock **pswap; 1816 1817 /* 1818 * Convert default object to swap object if necessary 1819 */ 1820 1821 if (object->type != OBJT_SWAP) { 1822 object->type = OBJT_SWAP; 1823 object->un_pager.swp.swp_bcount = 0; 1824 1825 if (object->handle != NULL) { 1826 TAILQ_INSERT_TAIL( 1827 NOBJLIST(object->handle), 1828 object, 1829 pager_object_list 1830 ); 1831 } else { 1832 TAILQ_INSERT_TAIL( 1833 &swap_pager_un_object_list, 1834 object, 1835 pager_object_list 1836 ); 1837 } 1838 } 1839 1840 /* 1841 * Locate hash entry. If not found create, but if we aren't adding 1842 * anything just return. If we run out of space in the map we wait 1843 * and, since the hash table may have changed, retry. 1844 */ 1845 1846 retry: 1847 pswap = swp_pager_hash(object, index); 1848 1849 if ((swap = *pswap) == NULL) { 1850 int i; 1851 1852 if (swapblk == SWAPBLK_NONE) 1853 return; 1854 1855 swap = *pswap = zalloc(swap_zone); 1856 if (swap == NULL) { 1857 vm_wait(0); 1858 goto retry; 1859 } 1860 swap->swb_hnext = NULL; 1861 swap->swb_object = object; 1862 swap->swb_index = index & ~SWAP_META_MASK; 1863 swap->swb_count = 0; 1864 1865 ++object->un_pager.swp.swp_bcount; 1866 1867 for (i = 0; i < SWAP_META_PAGES; ++i) 1868 swap->swb_pages[i] = SWAPBLK_NONE; 1869 } 1870 1871 /* 1872 * Delete prior contents of metadata 1873 */ 1874 1875 index &= SWAP_META_MASK; 1876 1877 if (swap->swb_pages[index] != SWAPBLK_NONE) { 1878 swp_pager_freeswapspace(swap->swb_pages[index], 1); 1879 --swap->swb_count; 1880 } 1881 1882 /* 1883 * Enter block into metadata 1884 */ 1885 1886 swap->swb_pages[index] = swapblk; 1887 if (swapblk != SWAPBLK_NONE) 1888 ++swap->swb_count; 1889 } 1890 1891 /* 1892 * SWP_PAGER_META_FREE() - free a range of blocks in the object's swap metadata 1893 * 1894 * The requested range of blocks is freed, with any associated swap 1895 * returned to the swap bitmap. 1896 * 1897 * This routine will free swap metadata structures as they are cleaned 1898 * out. This routine does *NOT* operate on swap metadata associated 1899 * with resident pages. 1900 * 1901 * This routine must be called at splvm() 1902 */ 1903 1904 static void 1905 swp_pager_meta_free(vm_object_t object, vm_pindex_t index, daddr_t count) 1906 { 1907 if (object->type != OBJT_SWAP) 1908 return; 1909 1910 while (count > 0) { 1911 struct swblock **pswap; 1912 struct swblock *swap; 1913 1914 pswap = swp_pager_hash(object, index); 1915 1916 if ((swap = *pswap) != NULL) { 1917 daddr_t v = swap->swb_pages[index & SWAP_META_MASK]; 1918 1919 if (v != SWAPBLK_NONE) { 1920 swp_pager_freeswapspace(v, 1); 1921 swap->swb_pages[index & SWAP_META_MASK] = 1922 SWAPBLK_NONE; 1923 if (--swap->swb_count == 0) { 1924 *pswap = swap->swb_hnext; 1925 zfree(swap_zone, swap); 1926 --object->un_pager.swp.swp_bcount; 1927 } 1928 } 1929 --count; 1930 ++index; 1931 } else { 1932 int n = SWAP_META_PAGES - (index & SWAP_META_MASK); 1933 count -= n; 1934 index += n; 1935 } 1936 } 1937 } 1938 1939 /* 1940 * SWP_PAGER_META_FREE_ALL() - destroy all swap metadata associated with object 1941 * 1942 * This routine locates and destroys all swap metadata associated with 1943 * an object. 1944 * 1945 * This routine must be called at splvm() 1946 */ 1947 1948 static void 1949 swp_pager_meta_free_all(vm_object_t object) 1950 { 1951 daddr_t index = 0; 1952 1953 if (object->type != OBJT_SWAP) 1954 return; 1955 1956 while (object->un_pager.swp.swp_bcount) { 1957 struct swblock **pswap; 1958 struct swblock *swap; 1959 1960 pswap = swp_pager_hash(object, index); 1961 if ((swap = *pswap) != NULL) { 1962 int i; 1963 1964 for (i = 0; i < SWAP_META_PAGES; ++i) { 1965 daddr_t v = swap->swb_pages[i]; 1966 if (v != SWAPBLK_NONE) { 1967 --swap->swb_count; 1968 swp_pager_freeswapspace(v, 1); 1969 } 1970 } 1971 if (swap->swb_count != 0) 1972 panic("swap_pager_meta_free_all: swb_count != 0"); 1973 *pswap = swap->swb_hnext; 1974 zfree(swap_zone, swap); 1975 --object->un_pager.swp.swp_bcount; 1976 } 1977 index += SWAP_META_PAGES; 1978 if (index > 0x20000000) 1979 panic("swp_pager_meta_free_all: failed to locate all swap meta blocks"); 1980 } 1981 } 1982 1983 /* 1984 * SWP_PAGER_METACTL() - misc control of swap and vm_page_t meta data. 1985 * 1986 * This routine is capable of looking up, popping, or freeing 1987 * swapblk assignments in the swap meta data or in the vm_page_t. 1988 * The routine typically returns the swapblk being looked-up, or popped, 1989 * or SWAPBLK_NONE if the block was freed, or SWAPBLK_NONE if the block 1990 * was invalid. This routine will automatically free any invalid 1991 * meta-data swapblks. 1992 * 1993 * It is not possible to store invalid swapblks in the swap meta data 1994 * (other then a literal 'SWAPBLK_NONE'), so we don't bother checking. 1995 * 1996 * When acting on a busy resident page and paging is in progress, we 1997 * have to wait until paging is complete but otherwise can act on the 1998 * busy page. 1999 * 2000 * This routine must be called at splvm(). 2001 * 2002 * SWM_FREE remove and free swap block from metadata 2003 * SWM_POP remove from meta data but do not free.. pop it out 2004 */ 2005 2006 static daddr_t 2007 swp_pager_meta_ctl( 2008 vm_object_t object, 2009 vm_pindex_t index, 2010 int flags 2011 ) { 2012 struct swblock **pswap; 2013 struct swblock *swap; 2014 daddr_t r1; 2015 2016 /* 2017 * The meta data only exists of the object is OBJT_SWAP 2018 * and even then might not be allocated yet. 2019 */ 2020 2021 if (object->type != OBJT_SWAP) 2022 return(SWAPBLK_NONE); 2023 2024 r1 = SWAPBLK_NONE; 2025 pswap = swp_pager_hash(object, index); 2026 2027 if ((swap = *pswap) != NULL) { 2028 index &= SWAP_META_MASK; 2029 r1 = swap->swb_pages[index]; 2030 2031 if (r1 != SWAPBLK_NONE) { 2032 if (flags & SWM_FREE) { 2033 swp_pager_freeswapspace(r1, 1); 2034 r1 = SWAPBLK_NONE; 2035 } 2036 if (flags & (SWM_FREE|SWM_POP)) { 2037 swap->swb_pages[index] = SWAPBLK_NONE; 2038 if (--swap->swb_count == 0) { 2039 *pswap = swap->swb_hnext; 2040 zfree(swap_zone, swap); 2041 --object->un_pager.swp.swp_bcount; 2042 } 2043 } 2044 } 2045 } 2046 return(r1); 2047 } 2048