vfs_bio.c revision 9708
1/* 2 * Copyright (c) 1994 John S. Dyson 3 * All rights reserved. 4 * 5 * Redistribution and use in source and binary forms, with or without 6 * modification, are permitted provided that the following conditions 7 * are met: 8 * 1. Redistributions of source code must retain the above copyright 9 * notice immediately at the beginning of the file, without modification, 10 * this list of conditions, and the following disclaimer. 11 * 2. Redistributions in binary form must reproduce the above copyright 12 * notice, this list of conditions and the following disclaimer in the 13 * documentation and/or other materials provided with the distribution. 14 * 3. Absolutely no warranty of function or purpose is made by the author 15 * John S. Dyson. 16 * 4. This work was done expressly for inclusion into FreeBSD. Other use 17 * is allowed if this notation is included. 18 * 5. Modifications may be freely made to this file if the above conditions 19 * are met. 20 * 21 * $Id: vfs_bio.c,v 1.54 1995/07/25 05:03:06 davidg Exp $ 22 */ 23 24/* 25 * this file contains a new buffer I/O scheme implementing a coherent 26 * VM object and buffer cache scheme. Pains have been taken to make 27 * sure that the performance degradation associated with schemes such 28 * as this is not realized. 29 * 30 * Author: John S. Dyson 31 * Significant help during the development and debugging phases 32 * had been provided by David Greenman, also of the FreeBSD core team. 33 */ 34 35#define VMIO 36#include <sys/param.h> 37#include <sys/systm.h> 38#include <sys/kernel.h> 39#include <sys/proc.h> 40#include <sys/vnode.h> 41#include <vm/vm.h> 42#include <vm/vm_kern.h> 43#include <vm/vm_pageout.h> 44#include <vm/vm_page.h> 45#include <vm/vm_object.h> 46#include <sys/buf.h> 47#include <sys/mount.h> 48#include <sys/malloc.h> 49#include <sys/resourcevar.h> 50#include <sys/proc.h> 51 52#include <miscfs/specfs/specdev.h> 53 54struct buf *buf; /* buffer header pool */ 55int nbuf; /* number of buffer headers calculated 56 * elsewhere */ 57struct swqueue bswlist; 58 59void vm_hold_free_pages(struct buf * bp, vm_offset_t from, vm_offset_t to); 60void vm_hold_load_pages(struct buf * bp, vm_offset_t from, vm_offset_t to); 61void vfs_clean_pages(struct buf * bp); 62static void vfs_setdirty(struct buf *bp); 63 64int needsbuffer; 65 66/* 67 * Internal update daemon, process 3 68 * The variable vfs_update_wakeup allows for internal syncs. 69 */ 70int vfs_update_wakeup; 71 72 73/* 74 * buffers base kva 75 */ 76caddr_t buffers_kva; 77 78/* 79 * bogus page -- for I/O to/from partially complete buffers 80 * this is a temporary solution to the problem, but it is not 81 * really that bad. it would be better to split the buffer 82 * for input in the case of buffers partially already in memory, 83 * but the code is intricate enough already. 84 */ 85vm_page_t bogus_page; 86vm_offset_t bogus_offset; 87 88int bufspace, maxbufspace; 89 90/* 91 * advisory minimum for size of LRU queue or VMIO queue 92 */ 93int minbuf; 94 95/* 96 * Initialize buffer headers and related structures. 97 */ 98void 99bufinit() 100{ 101 struct buf *bp; 102 int i; 103 104 TAILQ_INIT(&bswlist); 105 LIST_INIT(&invalhash); 106 107 /* first, make a null hash table */ 108 for (i = 0; i < BUFHSZ; i++) 109 LIST_INIT(&bufhashtbl[i]); 110 111 /* next, make a null set of free lists */ 112 for (i = 0; i < BUFFER_QUEUES; i++) 113 TAILQ_INIT(&bufqueues[i]); 114 115 buffers_kva = (caddr_t) kmem_alloc_pageable(buffer_map, MAXBSIZE * nbuf); 116 /* finally, initialize each buffer header and stick on empty q */ 117 for (i = 0; i < nbuf; i++) { 118 bp = &buf[i]; 119 bzero(bp, sizeof *bp); 120 bp->b_flags = B_INVAL; /* we're just an empty header */ 121 bp->b_dev = NODEV; 122 bp->b_rcred = NOCRED; 123 bp->b_wcred = NOCRED; 124 bp->b_qindex = QUEUE_EMPTY; 125 bp->b_vnbufs.le_next = NOLIST; 126 bp->b_data = buffers_kva + i * MAXBSIZE; 127 TAILQ_INSERT_TAIL(&bufqueues[QUEUE_EMPTY], bp, b_freelist); 128 LIST_INSERT_HEAD(&invalhash, bp, b_hash); 129 } 130/* 131 * maxbufspace is currently calculated to support all filesystem blocks 132 * to be 8K. If you happen to use a 16K filesystem, the size of the buffer 133 * cache is still the same as it would be for 8K filesystems. This 134 * keeps the size of the buffer cache "in check" for big block filesystems. 135 */ 136 minbuf = nbuf / 3; 137 maxbufspace = 2 * (nbuf + 8) * PAGE_SIZE; 138 139 bogus_offset = kmem_alloc_pageable(kernel_map, PAGE_SIZE); 140 bogus_page = vm_page_alloc(kernel_object, 141 bogus_offset - VM_MIN_KERNEL_ADDRESS, VM_ALLOC_NORMAL); 142 143} 144 145/* 146 * remove the buffer from the appropriate free list 147 */ 148void 149bremfree(struct buf * bp) 150{ 151 int s = splbio(); 152 153 if (bp->b_qindex != QUEUE_NONE) { 154 TAILQ_REMOVE(&bufqueues[bp->b_qindex], bp, b_freelist); 155 bp->b_qindex = QUEUE_NONE; 156 } else { 157 panic("bremfree: removing a buffer when not on a queue"); 158 } 159 splx(s); 160} 161 162/* 163 * Get a buffer with the specified data. Look in the cache first. 164 */ 165int 166bread(struct vnode * vp, daddr_t blkno, int size, struct ucred * cred, 167 struct buf ** bpp) 168{ 169 struct buf *bp; 170 171 bp = getblk(vp, blkno, size, 0, 0); 172 *bpp = bp; 173 174 /* if not found in cache, do some I/O */ 175 if ((bp->b_flags & B_CACHE) == 0) { 176 if (curproc != NULL) 177 curproc->p_stats->p_ru.ru_inblock++; 178 bp->b_flags |= B_READ; 179 bp->b_flags &= ~(B_DONE | B_ERROR | B_INVAL); 180 if (bp->b_rcred == NOCRED) { 181 if (cred != NOCRED) 182 crhold(cred); 183 bp->b_rcred = cred; 184 } 185 vfs_busy_pages(bp, 0); 186 VOP_STRATEGY(bp); 187 return (biowait(bp)); 188 } 189 return (0); 190} 191 192/* 193 * Operates like bread, but also starts asynchronous I/O on 194 * read-ahead blocks. 195 */ 196int 197breadn(struct vnode * vp, daddr_t blkno, int size, 198 daddr_t * rablkno, int *rabsize, 199 int cnt, struct ucred * cred, struct buf ** bpp) 200{ 201 struct buf *bp, *rabp; 202 int i; 203 int rv = 0, readwait = 0; 204 205 *bpp = bp = getblk(vp, blkno, size, 0, 0); 206 207 /* if not found in cache, do some I/O */ 208 if ((bp->b_flags & B_CACHE) == 0) { 209 if (curproc != NULL) 210 curproc->p_stats->p_ru.ru_inblock++; 211 bp->b_flags |= B_READ; 212 bp->b_flags &= ~(B_DONE | B_ERROR | B_INVAL); 213 if (bp->b_rcred == NOCRED) { 214 if (cred != NOCRED) 215 crhold(cred); 216 bp->b_rcred = cred; 217 } 218 vfs_busy_pages(bp, 0); 219 VOP_STRATEGY(bp); 220 ++readwait; 221 } 222 for (i = 0; i < cnt; i++, rablkno++, rabsize++) { 223 if (inmem(vp, *rablkno)) 224 continue; 225 rabp = getblk(vp, *rablkno, *rabsize, 0, 0); 226 227 if ((rabp->b_flags & B_CACHE) == 0) { 228 if (curproc != NULL) 229 curproc->p_stats->p_ru.ru_inblock++; 230 rabp->b_flags |= B_READ | B_ASYNC; 231 rabp->b_flags &= ~(B_DONE | B_ERROR | B_INVAL); 232 if (rabp->b_rcred == NOCRED) { 233 if (cred != NOCRED) 234 crhold(cred); 235 rabp->b_rcred = cred; 236 } 237 vfs_busy_pages(rabp, 0); 238 VOP_STRATEGY(rabp); 239 } else { 240 brelse(rabp); 241 } 242 } 243 244 if (readwait) { 245 rv = biowait(bp); 246 } 247 return (rv); 248} 249 250/* 251 * Write, release buffer on completion. (Done by iodone 252 * if async.) 253 */ 254int 255bwrite(struct buf * bp) 256{ 257 int oldflags = bp->b_flags; 258 259 if (bp->b_flags & B_INVAL) { 260 brelse(bp); 261 return (0); 262 } 263 if (!(bp->b_flags & B_BUSY)) 264 panic("bwrite: buffer is not busy???"); 265 266 bp->b_flags &= ~(B_READ | B_DONE | B_ERROR | B_DELWRI); 267 bp->b_flags |= B_WRITEINPROG; 268 269 if ((oldflags & (B_ASYNC|B_DELWRI)) == (B_ASYNC|B_DELWRI)) { 270 reassignbuf(bp, bp->b_vp); 271 } 272 273 bp->b_vp->v_numoutput++; 274 vfs_busy_pages(bp, 1); 275 if (curproc != NULL) 276 curproc->p_stats->p_ru.ru_oublock++; 277 VOP_STRATEGY(bp); 278 279 if ((oldflags & B_ASYNC) == 0) { 280 int rtval = biowait(bp); 281 282 if (oldflags & B_DELWRI) { 283 reassignbuf(bp, bp->b_vp); 284 } 285 brelse(bp); 286 return (rtval); 287 } 288 return (0); 289} 290 291int 292vn_bwrite(ap) 293 struct vop_bwrite_args *ap; 294{ 295 return (bwrite(ap->a_bp)); 296} 297 298/* 299 * Delayed write. (Buffer is marked dirty). 300 */ 301void 302bdwrite(struct buf * bp) 303{ 304 305 if ((bp->b_flags & B_BUSY) == 0) { 306 panic("bdwrite: buffer is not busy"); 307 } 308 if (bp->b_flags & B_INVAL) { 309 brelse(bp); 310 return; 311 } 312 if (bp->b_flags & B_TAPE) { 313 bawrite(bp); 314 return; 315 } 316 bp->b_flags &= ~(B_READ|B_RELBUF); 317 if ((bp->b_flags & B_DELWRI) == 0) { 318 bp->b_flags |= B_DONE | B_DELWRI; 319 reassignbuf(bp, bp->b_vp); 320 } 321 322 /* 323 * This bmap keeps the system from needing to do the bmap later, 324 * perhaps when the system is attempting to do a sync. Since it 325 * is likely that the indirect block -- or whatever other datastructure 326 * that the filesystem needs is still in memory now, it is a good 327 * thing to do this. Note also, that if the pageout daemon is 328 * requesting a sync -- there might not be enough memory to do 329 * the bmap then... So, this is important to do. 330 */ 331 if( bp->b_lblkno == bp->b_blkno) { 332 VOP_BMAP(bp->b_vp, bp->b_lblkno, NULL, &bp->b_blkno, NULL); 333 } 334 335 /* 336 * Set the *dirty* buffer range based upon the VM system dirty pages. 337 */ 338 vfs_setdirty(bp); 339 340 /* 341 * We need to do this here to satisfy the vnode_pager and the 342 * pageout daemon, so that it thinks that the pages have been 343 * "cleaned". Note that since the pages are in a delayed write 344 * buffer -- the VFS layer "will" see that the pages get written 345 * out on the next sync, or perhaps the cluster will be completed. 346 */ 347 vfs_clean_pages(bp); 348 brelse(bp); 349 return; 350} 351 352/* 353 * Asynchronous write. 354 * Start output on a buffer, but do not wait for it to complete. 355 * The buffer is released when the output completes. 356 */ 357void 358bawrite(struct buf * bp) 359{ 360 bp->b_flags |= B_ASYNC; 361 (void) VOP_BWRITE(bp); 362} 363 364/* 365 * Release a buffer. 366 */ 367void 368brelse(struct buf * bp) 369{ 370 int s; 371 372 if (bp->b_flags & B_CLUSTER) { 373 relpbuf(bp); 374 return; 375 } 376 /* anyone need a "free" block? */ 377 s = splbio(); 378 379 if (needsbuffer) { 380 needsbuffer = 0; 381 wakeup(&needsbuffer); 382 } 383 384 /* anyone need this block? */ 385 if (bp->b_flags & B_WANTED) { 386 bp->b_flags &= ~(B_WANTED | B_AGE); 387 wakeup(bp); 388 } else if (bp->b_flags & B_VMIO) { 389 bp->b_flags &= ~B_WANTED; 390 wakeup(bp); 391 } 392 if (bp->b_flags & B_LOCKED) 393 bp->b_flags &= ~B_ERROR; 394 395 if ((bp->b_flags & (B_NOCACHE | B_INVAL | B_ERROR)) || 396 (bp->b_bufsize <= 0)) { 397 bp->b_flags |= B_INVAL; 398 bp->b_flags &= ~(B_DELWRI | B_CACHE); 399 if (((bp->b_flags & B_VMIO) == 0) && bp->b_vp) 400 brelvp(bp); 401 } 402 403 /* 404 * VMIO buffer rundown. It is not very necessary to keep a VMIO buffer 405 * constituted, so the B_INVAL flag is used to *invalidate* the buffer, 406 * but the VM object is kept around. The B_NOCACHE flag is used to 407 * invalidate the pages in the VM object. 408 */ 409 if (bp->b_flags & B_VMIO) { 410 vm_offset_t foff; 411 vm_object_t obj; 412 int i, resid; 413 vm_page_t m; 414 int iototal = bp->b_bufsize; 415 416 foff = 0; 417 obj = 0; 418 if (bp->b_npages) { 419 if (bp->b_vp && bp->b_vp->v_mount) { 420 foff = bp->b_vp->v_mount->mnt_stat.f_iosize * bp->b_lblkno; 421 } else { 422 /* 423 * vnode pointer has been ripped away -- 424 * probably file gone... 425 */ 426 foff = bp->b_pages[0]->offset; 427 } 428 } 429 for (i = 0; i < bp->b_npages; i++) { 430 m = bp->b_pages[i]; 431 if (m == bogus_page) { 432 m = vm_page_lookup(obj, foff); 433 if (!m) { 434 panic("brelse: page missing\n"); 435 } 436 bp->b_pages[i] = m; 437 pmap_qenter(trunc_page(bp->b_data), bp->b_pages, bp->b_npages); 438 } 439 resid = (m->offset + PAGE_SIZE) - foff; 440 if (resid > iototal) 441 resid = iototal; 442 if (resid > 0) { 443 /* 444 * Don't invalidate the page if the local machine has already 445 * modified it. This is the lesser of two evils, and should 446 * be fixed. 447 */ 448 if (bp->b_flags & (B_NOCACHE | B_ERROR)) { 449 vm_page_test_dirty(m); 450 if (m->dirty == 0) { 451 vm_page_set_invalid(m, foff, resid); 452 if (m->valid == 0) 453 vm_page_protect(m, VM_PROT_NONE); 454 } 455 } 456 } 457 foff += resid; 458 iototal -= resid; 459 } 460 461 if (bp->b_flags & (B_INVAL | B_RELBUF)) { 462 for(i=0;i<bp->b_npages;i++) { 463 m = bp->b_pages[i]; 464 --m->bmapped; 465 if (m->bmapped == 0) { 466 if (m->flags & PG_WANTED) { 467 wakeup(m); 468 m->flags &= ~PG_WANTED; 469 } 470 vm_page_test_dirty(m); 471 if ((m->dirty & m->valid) == 0 && 472 (m->flags & PG_REFERENCED) == 0 && 473 !pmap_is_referenced(VM_PAGE_TO_PHYS(m))) { 474 vm_page_cache(m); 475 } else if ((m->flags & PG_ACTIVE) == 0) { 476 vm_page_activate(m); 477 m->act_count = 0; 478 } 479 } 480 } 481 bufspace -= bp->b_bufsize; 482 pmap_qremove(trunc_page((vm_offset_t) bp->b_data), bp->b_npages); 483 bp->b_npages = 0; 484 bp->b_bufsize = 0; 485 bp->b_flags &= ~B_VMIO; 486 if (bp->b_vp) 487 brelvp(bp); 488 } 489 } 490 if (bp->b_qindex != QUEUE_NONE) 491 panic("brelse: free buffer onto another queue???"); 492 493 /* enqueue */ 494 /* buffers with no memory */ 495 if (bp->b_bufsize == 0) { 496 bp->b_qindex = QUEUE_EMPTY; 497 TAILQ_INSERT_TAIL(&bufqueues[QUEUE_EMPTY], bp, b_freelist); 498 LIST_REMOVE(bp, b_hash); 499 LIST_INSERT_HEAD(&invalhash, bp, b_hash); 500 bp->b_dev = NODEV; 501 /* buffers with junk contents */ 502 } else if (bp->b_flags & (B_ERROR | B_INVAL | B_NOCACHE | B_RELBUF)) { 503 bp->b_qindex = QUEUE_AGE; 504 TAILQ_INSERT_HEAD(&bufqueues[QUEUE_AGE], bp, b_freelist); 505 LIST_REMOVE(bp, b_hash); 506 LIST_INSERT_HEAD(&invalhash, bp, b_hash); 507 bp->b_dev = NODEV; 508 /* buffers that are locked */ 509 } else if (bp->b_flags & B_LOCKED) { 510 bp->b_qindex = QUEUE_LOCKED; 511 TAILQ_INSERT_TAIL(&bufqueues[QUEUE_LOCKED], bp, b_freelist); 512 /* buffers with stale but valid contents */ 513 } else if (bp->b_flags & B_AGE) { 514 bp->b_qindex = QUEUE_AGE; 515 TAILQ_INSERT_TAIL(&bufqueues[QUEUE_AGE], bp, b_freelist); 516 /* buffers with valid and quite potentially reuseable contents */ 517 } else { 518 bp->b_qindex = QUEUE_LRU; 519 TAILQ_INSERT_TAIL(&bufqueues[QUEUE_LRU], bp, b_freelist); 520 } 521 522 /* unlock */ 523 bp->b_flags &= ~(B_WANTED | B_BUSY | B_ASYNC | B_NOCACHE | B_AGE | B_RELBUF); 524 splx(s); 525} 526 527/* 528 * this routine implements clustered async writes for 529 * clearing out B_DELWRI buffers... This is much better 530 * than the old way of writing only one buffer at a time. 531 */ 532void 533vfs_bio_awrite(struct buf * bp) 534{ 535 int i; 536 daddr_t lblkno = bp->b_lblkno; 537 struct vnode *vp = bp->b_vp; 538 int s; 539 int ncl; 540 struct buf *bpa; 541 542 s = splbio(); 543 if( vp->v_mount && (vp->v_flag & VVMIO) && 544 (bp->b_flags & (B_CLUSTEROK | B_INVAL)) == B_CLUSTEROK) { 545 int size = vp->v_mount->mnt_stat.f_iosize; 546 547 for (i = 1; i < MAXPHYS / size; i++) { 548 if ((bpa = incore(vp, lblkno + i)) && 549 ((bpa->b_flags & (B_BUSY | B_DELWRI | B_CLUSTEROK | B_INVAL)) == 550 (B_DELWRI | B_CLUSTEROK)) && 551 (bpa->b_bufsize == size)) { 552 if ((bpa->b_blkno == bpa->b_lblkno) || 553 (bpa->b_blkno != bp->b_blkno + (i * size) / DEV_BSIZE)) 554 break; 555 } else { 556 break; 557 } 558 } 559 ncl = i; 560 /* 561 * this is a possible cluster write 562 */ 563 if (ncl != 1) { 564 bremfree(bp); 565 cluster_wbuild(vp, bp, size, lblkno, ncl, -1); 566 splx(s); 567 return; 568 } 569 } 570 /* 571 * default (old) behavior, writing out only one block 572 */ 573 bremfree(bp); 574 bp->b_flags |= B_BUSY | B_ASYNC; 575 (void) VOP_BWRITE(bp); 576 splx(s); 577} 578 579 580/* 581 * Find a buffer header which is available for use. 582 */ 583static struct buf * 584getnewbuf(int slpflag, int slptimeo, int doingvmio) 585{ 586 struct buf *bp; 587 int s; 588 int firstbp = 1; 589 590 s = splbio(); 591start: 592 if (bufspace >= maxbufspace) 593 goto trytofreespace; 594 595 /* can we constitute a new buffer? */ 596 if ((bp = bufqueues[QUEUE_EMPTY].tqh_first)) { 597 if (bp->b_qindex != QUEUE_EMPTY) 598 panic("getnewbuf: inconsistent EMPTY queue"); 599 bremfree(bp); 600 goto fillbuf; 601 } 602trytofreespace: 603 /* 604 * We keep the file I/O from hogging metadata I/O 605 * This is desirable because file data is cached in the 606 * VM/Buffer cache even if a buffer is freed. 607 */ 608 if ((bp = bufqueues[QUEUE_AGE].tqh_first)) { 609 if (bp->b_qindex != QUEUE_AGE) 610 panic("getnewbuf: inconsistent AGE queue"); 611 } else if ((bp = bufqueues[QUEUE_LRU].tqh_first)) { 612 if (bp->b_qindex != QUEUE_LRU) 613 panic("getnewbuf: inconsistent LRU queue"); 614 } 615 if (!bp) { 616 /* wait for a free buffer of any kind */ 617 needsbuffer = 1; 618 tsleep(&needsbuffer, PRIBIO | slpflag, "newbuf", slptimeo); 619 splx(s); 620 return (0); 621 } 622 623 /* if we are a delayed write, convert to an async write */ 624 if ((bp->b_flags & (B_DELWRI | B_INVAL)) == B_DELWRI) { 625 vfs_bio_awrite(bp); 626 if (!slpflag && !slptimeo) { 627 splx(s); 628 return (0); 629 } 630 goto start; 631 } 632 633 if (bp->b_flags & B_WANTED) { 634 bp->b_flags &= ~B_WANTED; 635 wakeup(bp); 636 } 637 bremfree(bp); 638 639 if (bp->b_flags & B_VMIO) { 640 bp->b_flags |= B_RELBUF | B_BUSY | B_DONE; 641 brelse(bp); 642 bremfree(bp); 643 } 644 645 if (bp->b_vp) 646 brelvp(bp); 647 648 /* we are not free, nor do we contain interesting data */ 649 if (bp->b_rcred != NOCRED) 650 crfree(bp->b_rcred); 651 if (bp->b_wcred != NOCRED) 652 crfree(bp->b_wcred); 653fillbuf: 654 bp->b_flags |= B_BUSY; 655 LIST_REMOVE(bp, b_hash); 656 LIST_INSERT_HEAD(&invalhash, bp, b_hash); 657 splx(s); 658 if (bp->b_bufsize) { 659 allocbuf(bp, 0); 660 } 661 bp->b_flags = B_BUSY; 662 bp->b_dev = NODEV; 663 bp->b_vp = NULL; 664 bp->b_blkno = bp->b_lblkno = 0; 665 bp->b_iodone = 0; 666 bp->b_error = 0; 667 bp->b_resid = 0; 668 bp->b_bcount = 0; 669 bp->b_npages = 0; 670 bp->b_wcred = bp->b_rcred = NOCRED; 671 bp->b_data = buffers_kva + (bp - buf) * MAXBSIZE; 672 bp->b_dirtyoff = bp->b_dirtyend = 0; 673 bp->b_validoff = bp->b_validend = 0; 674 if (bufspace >= maxbufspace) { 675 s = splbio(); 676 bp->b_flags |= B_INVAL; 677 brelse(bp); 678 goto trytofreespace; 679 } 680 return (bp); 681} 682 683/* 684 * Check to see if a block is currently memory resident. 685 */ 686struct buf * 687incore(struct vnode * vp, daddr_t blkno) 688{ 689 struct buf *bp; 690 struct bufhashhdr *bh; 691 692 int s = splbio(); 693 694 bh = BUFHASH(vp, blkno); 695 bp = bh->lh_first; 696 697 /* Search hash chain */ 698 while (bp) { 699 /* hit */ 700 if (bp->b_lblkno == blkno && bp->b_vp == vp && 701 (bp->b_flags & B_INVAL) == 0) { 702 splx(s); 703 return (bp); 704 } 705 bp = bp->b_hash.le_next; 706 } 707 splx(s); 708 709 return (0); 710} 711 712/* 713 * Returns true if no I/O is needed to access the 714 * associated VM object. This is like incore except 715 * it also hunts around in the VM system for the data. 716 */ 717 718int 719inmem(struct vnode * vp, daddr_t blkno) 720{ 721 vm_object_t obj; 722 vm_offset_t off, toff, tinc; 723 vm_page_t m; 724 725 if (incore(vp, blkno)) 726 return 1; 727 if (vp->v_mount == 0) 728 return 0; 729 if ((vp->v_object == 0) || (vp->v_flag & VVMIO) == 0) 730 return 0; 731 732 obj = vp->v_object; 733 tinc = PAGE_SIZE; 734 if (tinc > vp->v_mount->mnt_stat.f_iosize) 735 tinc = vp->v_mount->mnt_stat.f_iosize; 736 off = blkno * vp->v_mount->mnt_stat.f_iosize; 737 738 for (toff = 0; toff < vp->v_mount->mnt_stat.f_iosize; toff += tinc) { 739 int mask; 740 741 m = vm_page_lookup(obj, trunc_page(toff + off)); 742 if (!m) 743 return 0; 744 if (vm_page_is_valid(m, toff + off, tinc) == 0) 745 return 0; 746 } 747 return 1; 748} 749 750/* 751 * now we set the dirty range for the buffer -- 752 * for NFS -- if the file is mapped and pages have 753 * been written to, let it know. We want the 754 * entire range of the buffer to be marked dirty if 755 * any of the pages have been written to for consistancy 756 * with the b_validoff, b_validend set in the nfs write 757 * code, and used by the nfs read code. 758 */ 759static void 760vfs_setdirty(struct buf *bp) { 761 int i; 762 vm_object_t object; 763 vm_offset_t boffset, offset; 764 /* 765 * We qualify the scan for modified pages on whether the 766 * object has been flushed yet. The OBJ_WRITEABLE flag 767 * is not cleared simply by protecting pages off. 768 */ 769 if ((bp->b_flags & B_VMIO) && 770 ((object = bp->b_pages[0]->object)->flags & OBJ_WRITEABLE)) { 771 /* 772 * test the pages to see if they have been modified directly 773 * by users through the VM system. 774 */ 775 for (i = 0; i < bp->b_npages; i++) 776 vm_page_test_dirty(bp->b_pages[i]); 777 778 /* 779 * scan forwards for the first page modified 780 */ 781 for (i = 0; i < bp->b_npages; i++) { 782 if (bp->b_pages[i]->dirty) { 783 break; 784 } 785 } 786 boffset = i * PAGE_SIZE; 787 if (boffset < bp->b_dirtyoff) { 788 bp->b_dirtyoff = boffset; 789 } 790 791 /* 792 * scan backwards for the last page modified 793 */ 794 for (i = bp->b_npages - 1; i >= 0; --i) { 795 if (bp->b_pages[i]->dirty) { 796 break; 797 } 798 } 799 boffset = (i + 1) * PAGE_SIZE; 800 offset = boffset + bp->b_pages[0]->offset; 801 if (offset >= object->size) { 802 boffset = object->size - bp->b_pages[0]->offset; 803 } 804 if (bp->b_dirtyend < boffset) { 805 bp->b_dirtyend = boffset; 806 } 807 } 808} 809 810/* 811 * Get a block given a specified block and offset into a file/device. 812 */ 813struct buf * 814getblk(struct vnode * vp, daddr_t blkno, int size, int slpflag, int slptimeo) 815{ 816 struct buf *bp; 817 int s; 818 struct bufhashhdr *bh; 819 vm_offset_t off; 820 int nleft; 821 822 s = splbio(); 823loop: 824 if (bp = incore(vp, blkno)) { 825 if (bp->b_flags & B_BUSY) { 826 bp->b_flags |= B_WANTED; 827 if (!tsleep(bp, PRIBIO | slpflag, "getblk", slptimeo)) 828 goto loop; 829 830 splx(s); 831 return (struct buf *) NULL; 832 } 833 bp->b_flags |= B_BUSY | B_CACHE; 834 bremfree(bp); 835 /* 836 * check for size inconsistancies 837 */ 838 if (bp->b_bcount != size) { 839 if (bp->b_flags & B_VMIO) { 840 allocbuf(bp, size); 841 } else { 842 bp->b_flags |= B_NOCACHE; 843 VOP_BWRITE(bp); 844 goto loop; 845 } 846 } 847 splx(s); 848 return (bp); 849 } else { 850 vm_object_t obj; 851 int doingvmio; 852 853 if ((obj = vp->v_object) && (vp->v_flag & VVMIO)) { 854 doingvmio = 1; 855 } else { 856 doingvmio = 0; 857 } 858 if ((bp = getnewbuf(slpflag, slptimeo, doingvmio)) == 0) { 859 if (slpflag || slptimeo) 860 return NULL; 861 goto loop; 862 } 863 864 /* 865 * This code is used to make sure that a buffer is not 866 * created while the getnewbuf routine is blocked. 867 * Normally the vnode is locked so this isn't a problem. 868 * VBLK type I/O requests, however, don't lock the vnode. 869 */ 870 if (!VOP_ISLOCKED(vp) && incore(vp, blkno)) { 871 bp->b_flags |= B_INVAL; 872 brelse(bp); 873 goto loop; 874 } 875 876 /* 877 * Insert the buffer into the hash, so that it can 878 * be found by incore. 879 */ 880 bp->b_blkno = bp->b_lblkno = blkno; 881 bgetvp(vp, bp); 882 LIST_REMOVE(bp, b_hash); 883 bh = BUFHASH(vp, blkno); 884 LIST_INSERT_HEAD(bh, bp, b_hash); 885 886 if (doingvmio) { 887 bp->b_flags |= (B_VMIO | B_CACHE); 888#if defined(VFS_BIO_DEBUG) 889 if (vp->v_type != VREG) 890 printf("getblk: vmioing file type %d???\n", vp->v_type); 891#endif 892 } else { 893 bp->b_flags &= ~B_VMIO; 894 } 895 splx(s); 896 897 allocbuf(bp, size); 898 return (bp); 899 } 900} 901 902/* 903 * Get an empty, disassociated buffer of given size. 904 */ 905struct buf * 906geteblk(int size) 907{ 908 struct buf *bp; 909 910 while ((bp = getnewbuf(0, 0, 0)) == 0); 911 allocbuf(bp, size); 912 bp->b_flags |= B_INVAL; 913 return (bp); 914} 915 916/* 917 * This code constitutes the buffer memory from either anonymous system 918 * memory (in the case of non-VMIO operations) or from an associated 919 * VM object (in the case of VMIO operations). 920 * 921 * Note that this code is tricky, and has many complications to resolve 922 * deadlock or inconsistant data situations. Tread lightly!!! 923 * 924 * Modify the length of a buffer's underlying buffer storage without 925 * destroying information (unless, of course the buffer is shrinking). 926 */ 927int 928allocbuf(struct buf * bp, int size) 929{ 930 931 int s; 932 int newbsize, mbsize; 933 int i; 934 935 if (!(bp->b_flags & B_BUSY)) 936 panic("allocbuf: buffer not busy"); 937 938 if ((bp->b_flags & B_VMIO) == 0) { 939 /* 940 * Just get anonymous memory from the kernel 941 */ 942 mbsize = ((size + DEV_BSIZE - 1) / DEV_BSIZE) * DEV_BSIZE; 943 newbsize = round_page(size); 944 945 if (newbsize < bp->b_bufsize) { 946 vm_hold_free_pages( 947 bp, 948 (vm_offset_t) bp->b_data + newbsize, 949 (vm_offset_t) bp->b_data + bp->b_bufsize); 950 } else if (newbsize > bp->b_bufsize) { 951 vm_hold_load_pages( 952 bp, 953 (vm_offset_t) bp->b_data + bp->b_bufsize, 954 (vm_offset_t) bp->b_data + newbsize); 955 } 956 } else { 957 vm_page_t m; 958 int desiredpages; 959 960 newbsize = ((size + DEV_BSIZE - 1) / DEV_BSIZE) * DEV_BSIZE; 961 desiredpages = round_page(newbsize) / PAGE_SIZE; 962 963 if (newbsize < bp->b_bufsize) { 964 if (desiredpages < bp->b_npages) { 965 pmap_qremove((vm_offset_t) trunc_page(bp->b_data) + 966 desiredpages * PAGE_SIZE, (bp->b_npages - desiredpages)); 967 for (i = desiredpages; i < bp->b_npages; i++) { 968 m = bp->b_pages[i]; 969 s = splhigh(); 970 while ((m->flags & PG_BUSY) || (m->busy != 0)) { 971 m->flags |= PG_WANTED; 972 tsleep(m, PVM, "biodep", 0); 973 } 974 splx(s); 975 976 if (m->bmapped == 0) { 977 printf("allocbuf: bmapped is zero for page %d\n", i); 978 panic("allocbuf: error"); 979 } 980 --m->bmapped; 981 if (m->bmapped == 0) { 982 vm_page_protect(m, VM_PROT_NONE); 983 vm_page_free(m); 984 } 985 bp->b_pages[i] = NULL; 986 } 987 bp->b_npages = desiredpages; 988 } 989 } else if (newbsize > bp->b_bufsize) { 990 vm_object_t obj; 991 vm_offset_t tinc, off, toff, objoff; 992 int pageindex, curbpnpages; 993 struct vnode *vp; 994 int bsize; 995 996 vp = bp->b_vp; 997 bsize = vp->v_mount->mnt_stat.f_iosize; 998 999 if (bp->b_npages < desiredpages) { 1000 obj = vp->v_object; 1001 tinc = PAGE_SIZE; 1002 if (tinc > bsize) 1003 tinc = bsize; 1004 off = bp->b_lblkno * bsize; 1005 doretry: 1006 curbpnpages = bp->b_npages; 1007 bp->b_flags |= B_CACHE; 1008 for (toff = 0; toff < newbsize; toff += tinc) { 1009 int mask; 1010 int bytesinpage; 1011 1012 pageindex = toff / PAGE_SIZE; 1013 objoff = trunc_page(toff + off); 1014 if (pageindex < curbpnpages) { 1015 int pb; 1016 1017 m = bp->b_pages[pageindex]; 1018 if (m->offset != objoff) 1019 panic("allocbuf: page changed offset??!!!?"); 1020 bytesinpage = tinc; 1021 if (tinc > (newbsize - toff)) 1022 bytesinpage = newbsize - toff; 1023 if (!vm_page_is_valid(m, toff + off, bytesinpage)) { 1024 bp->b_flags &= ~B_CACHE; 1025 } 1026 if ((m->flags & PG_ACTIVE) == 0) { 1027 vm_page_activate(m); 1028 m->act_count = 0; 1029 } 1030 continue; 1031 } 1032 m = vm_page_lookup(obj, objoff); 1033 if (!m) { 1034 m = vm_page_alloc(obj, objoff, VM_ALLOC_NORMAL); 1035 if (!m) { 1036 int j; 1037 1038 for (j = bp->b_npages; j < pageindex; j++) { 1039 PAGE_WAKEUP(bp->b_pages[j]); 1040 } 1041 VM_WAIT; 1042 goto doretry; 1043 } 1044 vm_page_activate(m); 1045 m->act_count = 0; 1046 m->valid = 0; 1047 bp->b_flags &= ~B_CACHE; 1048 } else if (m->flags & PG_BUSY) { 1049 int j; 1050 1051 for (j = bp->b_npages; j < pageindex; j++) { 1052 PAGE_WAKEUP(bp->b_pages[j]); 1053 } 1054 1055 s = splbio(); 1056 m->flags |= PG_WANTED; 1057 tsleep(m, PRIBIO, "pgtblk", 0); 1058 splx(s); 1059 1060 goto doretry; 1061 } else { 1062 int pb; 1063 if ((curproc != pageproc) && 1064 (m->flags & PG_CACHE) && 1065 (cnt.v_free_count + cnt.v_cache_count) < cnt.v_free_min) { 1066 pagedaemon_wakeup(); 1067 } 1068 bytesinpage = tinc; 1069 if (tinc > (newbsize - toff)) 1070 bytesinpage = newbsize - toff; 1071 if (!vm_page_is_valid(m, toff + off, bytesinpage)) { 1072 bp->b_flags &= ~B_CACHE; 1073 } 1074 if ((m->flags & PG_ACTIVE) == 0) { 1075 vm_page_activate(m); 1076 m->act_count = 0; 1077 } 1078 m->flags |= PG_BUSY; 1079 } 1080 bp->b_pages[pageindex] = m; 1081 curbpnpages = pageindex + 1; 1082 } 1083 for (i = bp->b_npages; i < curbpnpages; i++) { 1084 m = bp->b_pages[i]; 1085 m->bmapped++; 1086 PAGE_WAKEUP(m); 1087 } 1088 bp->b_npages = curbpnpages; 1089 bp->b_data = buffers_kva + (bp - buf) * MAXBSIZE; 1090 pmap_qenter((vm_offset_t) bp->b_data, bp->b_pages, bp->b_npages); 1091 bp->b_data += off % PAGE_SIZE; 1092 } 1093 } 1094 } 1095 bufspace += (newbsize - bp->b_bufsize); 1096 bp->b_bufsize = newbsize; 1097 bp->b_bcount = size; 1098 return 1; 1099} 1100 1101/* 1102 * Wait for buffer I/O completion, returning error status. 1103 */ 1104int 1105biowait(register struct buf * bp) 1106{ 1107 int s; 1108 1109 s = splbio(); 1110 while ((bp->b_flags & B_DONE) == 0) 1111 tsleep(bp, PRIBIO, "biowait", 0); 1112 splx(s); 1113 if (bp->b_flags & B_EINTR) { 1114 bp->b_flags &= ~B_EINTR; 1115 return (EINTR); 1116 } 1117 if (bp->b_flags & B_ERROR) { 1118 return (bp->b_error ? bp->b_error : EIO); 1119 } else { 1120 return (0); 1121 } 1122} 1123 1124/* 1125 * Finish I/O on a buffer, calling an optional function. 1126 * This is usually called from interrupt level, so process blocking 1127 * is not *a good idea*. 1128 */ 1129void 1130biodone(register struct buf * bp) 1131{ 1132 int s; 1133 1134 s = splbio(); 1135 if (!(bp->b_flags & B_BUSY)) 1136 panic("biodone: buffer not busy"); 1137 1138 if (bp->b_flags & B_DONE) { 1139 splx(s); 1140 printf("biodone: buffer already done\n"); 1141 return; 1142 } 1143 bp->b_flags |= B_DONE; 1144 1145 if ((bp->b_flags & B_READ) == 0) { 1146 struct vnode *vp = bp->b_vp; 1147 vwakeup(bp); 1148 } 1149#ifdef BOUNCE_BUFFERS 1150 if (bp->b_flags & B_BOUNCE) 1151 vm_bounce_free(bp); 1152#endif 1153 1154 /* call optional completion function if requested */ 1155 if (bp->b_flags & B_CALL) { 1156 bp->b_flags &= ~B_CALL; 1157 (*bp->b_iodone) (bp); 1158 splx(s); 1159 return; 1160 } 1161 if (bp->b_flags & B_VMIO) { 1162 int i, resid; 1163 vm_offset_t foff; 1164 vm_page_t m; 1165 vm_object_t obj; 1166 int iosize; 1167 struct vnode *vp = bp->b_vp; 1168 1169 foff = vp->v_mount->mnt_stat.f_iosize * bp->b_lblkno; 1170 obj = vp->v_object; 1171 if (!obj) { 1172 panic("biodone: no object"); 1173 } 1174#if defined(VFS_BIO_DEBUG) 1175 if (obj->paging_in_progress < bp->b_npages) { 1176 printf("biodone: paging in progress(%d) < bp->b_npages(%d)\n", 1177 obj->paging_in_progress, bp->b_npages); 1178 } 1179#endif 1180 iosize = bp->b_bufsize; 1181 for (i = 0; i < bp->b_npages; i++) { 1182 int bogusflag = 0; 1183 m = bp->b_pages[i]; 1184 if (m == bogus_page) { 1185 bogusflag = 1; 1186 m = vm_page_lookup(obj, foff); 1187 if (!m) { 1188#if defined(VFS_BIO_DEBUG) 1189 printf("biodone: page disappeared\n"); 1190#endif 1191 --obj->paging_in_progress; 1192 continue; 1193 } 1194 bp->b_pages[i] = m; 1195 pmap_qenter(trunc_page(bp->b_data), bp->b_pages, bp->b_npages); 1196 } 1197#if defined(VFS_BIO_DEBUG) 1198 if (trunc_page(foff) != m->offset) { 1199 printf("biodone: foff(%d)/m->offset(%d) mismatch\n", foff, m->offset); 1200 } 1201#endif 1202 resid = (m->offset + PAGE_SIZE) - foff; 1203 if (resid > iosize) 1204 resid = iosize; 1205 /* 1206 * In the write case, the valid and clean bits are 1207 * already changed correctly, so we only need to do this 1208 * here in the read case. 1209 */ 1210 if ((bp->b_flags & B_READ) && !bogusflag && resid > 0) { 1211 vm_page_set_valid(m, foff & (PAGE_SIZE-1), resid); 1212 vm_page_set_clean(m, foff & (PAGE_SIZE-1), resid); 1213 } 1214 1215 /* 1216 * when debugging new filesystems or buffer I/O methods, this 1217 * is the most common error that pops up. if you see this, you 1218 * have not set the page busy flag correctly!!! 1219 */ 1220 if (m->busy == 0) { 1221 printf("biodone: page busy < 0, " 1222 "off: %ld, foff: %ld, " 1223 "resid: %d, index: %d\n", 1224 m->offset, foff, resid, i); 1225 printf(" iosize: %ld, lblkno: %ld, flags: 0x%x, npages: %d\n", 1226 bp->b_vp->v_mount->mnt_stat.f_iosize, 1227 bp->b_lblkno, bp->b_flags, bp->b_npages); 1228 printf(" valid: 0x%x, dirty: 0x%x, mapped: %d\n", 1229 m->valid, m->dirty, m->bmapped); 1230 panic("biodone: page busy < 0\n"); 1231 } 1232 --m->busy; 1233 if ((m->busy == 0) && (m->flags & PG_WANTED)) { 1234 m->flags &= ~PG_WANTED; 1235 wakeup(m); 1236 } 1237 --obj->paging_in_progress; 1238 foff += resid; 1239 iosize -= resid; 1240 } 1241 if (obj && obj->paging_in_progress == 0 && 1242 (obj->flags & OBJ_PIPWNT)) { 1243 obj->flags &= ~OBJ_PIPWNT; 1244 wakeup(obj); 1245 } 1246 } 1247 /* 1248 * For asynchronous completions, release the buffer now. The brelse 1249 * checks for B_WANTED and will do the wakeup there if necessary - so 1250 * no need to do a wakeup here in the async case. 1251 */ 1252 1253 if (bp->b_flags & B_ASYNC) { 1254 brelse(bp); 1255 } else { 1256 bp->b_flags &= ~B_WANTED; 1257 wakeup(bp); 1258 } 1259 splx(s); 1260} 1261 1262int 1263count_lock_queue() 1264{ 1265 int count; 1266 struct buf *bp; 1267 1268 count = 0; 1269 for (bp = bufqueues[QUEUE_LOCKED].tqh_first; 1270 bp != NULL; 1271 bp = bp->b_freelist.tqe_next) 1272 count++; 1273 return (count); 1274} 1275 1276int vfs_update_interval = 30; 1277 1278void 1279vfs_update() 1280{ 1281 (void) spl0(); 1282 while (1) { 1283 tsleep(&vfs_update_wakeup, PRIBIO, "update", 1284 hz * vfs_update_interval); 1285 vfs_update_wakeup = 0; 1286 sync(curproc, NULL, NULL); 1287 } 1288} 1289 1290/* 1291 * This routine is called in lieu of iodone in the case of 1292 * incomplete I/O. This keeps the busy status for pages 1293 * consistant. 1294 */ 1295void 1296vfs_unbusy_pages(struct buf * bp) 1297{ 1298 int i; 1299 1300 if (bp->b_flags & B_VMIO) { 1301 struct vnode *vp = bp->b_vp; 1302 vm_object_t obj = vp->v_object; 1303 vm_offset_t foff; 1304 1305 foff = trunc_page(vp->v_mount->mnt_stat.f_iosize * bp->b_lblkno); 1306 1307 for (i = 0; i < bp->b_npages; i++) { 1308 vm_page_t m = bp->b_pages[i]; 1309 1310 if (m == bogus_page) { 1311 m = vm_page_lookup(obj, foff + i * PAGE_SIZE); 1312 if (!m) { 1313 panic("vfs_unbusy_pages: page missing\n"); 1314 } 1315 bp->b_pages[i] = m; 1316 pmap_qenter(trunc_page(bp->b_data), bp->b_pages, bp->b_npages); 1317 } 1318 --obj->paging_in_progress; 1319 --m->busy; 1320 if ((m->busy == 0) && (m->flags & PG_WANTED)) { 1321 m->flags &= ~PG_WANTED; 1322 wakeup(m); 1323 } 1324 } 1325 if (obj->paging_in_progress == 0 && 1326 (obj->flags & OBJ_PIPWNT)) { 1327 obj->flags &= ~OBJ_PIPWNT; 1328 wakeup(obj); 1329 } 1330 } 1331} 1332 1333/* 1334 * This routine is called before a device strategy routine. 1335 * It is used to tell the VM system that paging I/O is in 1336 * progress, and treat the pages associated with the buffer 1337 * almost as being PG_BUSY. Also the object paging_in_progress 1338 * flag is handled to make sure that the object doesn't become 1339 * inconsistant. 1340 */ 1341void 1342vfs_busy_pages(struct buf * bp, int clear_modify) 1343{ 1344 int i; 1345 1346 if (bp->b_flags & B_VMIO) { 1347 vm_object_t obj = bp->b_vp->v_object; 1348 vm_offset_t foff = bp->b_vp->v_mount->mnt_stat.f_iosize * bp->b_lblkno; 1349 int iocount = bp->b_bufsize; 1350 1351 vfs_setdirty(bp); 1352 for (i = 0; i < bp->b_npages; i++) { 1353 vm_page_t m = bp->b_pages[i]; 1354 int resid = (m->offset + PAGE_SIZE) - foff; 1355 1356 if (resid > iocount) 1357 resid = iocount; 1358 obj->paging_in_progress++; 1359 m->busy++; 1360 if (clear_modify) { 1361 vm_page_protect(m, VM_PROT_READ); 1362 vm_page_set_valid(m, 1363 foff & (PAGE_SIZE-1), resid); 1364 vm_page_set_clean(m, 1365 foff & (PAGE_SIZE-1), resid); 1366 } else if (bp->b_bcount >= PAGE_SIZE) { 1367 if (m->valid && (bp->b_flags & B_CACHE) == 0) { 1368 bp->b_pages[i] = bogus_page; 1369 pmap_qenter(trunc_page(bp->b_data), bp->b_pages, bp->b_npages); 1370 } 1371 } 1372 foff += resid; 1373 iocount -= resid; 1374 } 1375 } 1376} 1377 1378/* 1379 * Tell the VM system that the pages associated with this buffer 1380 * are clean. This is used for delayed writes where the data is 1381 * going to go to disk eventually without additional VM intevention. 1382 */ 1383void 1384vfs_clean_pages(struct buf * bp) 1385{ 1386 int i; 1387 1388 if (bp->b_flags & B_VMIO) { 1389 vm_offset_t foff = 1390 bp->b_vp->v_mount->mnt_stat.f_iosize * bp->b_lblkno; 1391 int iocount = bp->b_bufsize; 1392 1393 for (i = 0; i < bp->b_npages; i++) { 1394 vm_page_t m = bp->b_pages[i]; 1395 int resid = (m->offset + PAGE_SIZE) - foff; 1396 1397 if (resid > iocount) 1398 resid = iocount; 1399 if (resid > 0) { 1400 vm_page_set_valid(m, 1401 foff & (PAGE_SIZE-1), resid); 1402 vm_page_set_clean(m, 1403 foff & (PAGE_SIZE-1), resid); 1404 } 1405 foff += resid; 1406 iocount -= resid; 1407 } 1408 } 1409} 1410 1411void 1412vfs_bio_clrbuf(struct buf *bp) { 1413 int i; 1414 if( bp->b_flags & B_VMIO) { 1415 if( (bp->b_npages == 1) && (bp->b_bufsize < PAGE_SIZE)) { 1416 int j; 1417 if( bp->b_pages[0]->valid != VM_PAGE_BITS_ALL) { 1418 for(j=0; j < bp->b_bufsize / DEV_BSIZE;j++) { 1419 bzero(bp->b_data + j * DEV_BSIZE, DEV_BSIZE); 1420 } 1421 } 1422 bp->b_resid = 0; 1423 return; 1424 } 1425 for(i=0;i<bp->b_npages;i++) { 1426 if( bp->b_pages[i]->valid == VM_PAGE_BITS_ALL) 1427 continue; 1428 if( bp->b_pages[i]->valid == 0) { 1429 bzero(bp->b_data + i * PAGE_SIZE, PAGE_SIZE); 1430 } else { 1431 int j; 1432 for(j=0;j<PAGE_SIZE/DEV_BSIZE;j++) { 1433 if( (bp->b_pages[i]->valid & (1<<j)) == 0) 1434 bzero(bp->b_data + i * PAGE_SIZE + j * DEV_BSIZE, DEV_BSIZE); 1435 } 1436 } 1437 bp->b_pages[i]->valid = VM_PAGE_BITS_ALL; 1438 } 1439 bp->b_resid = 0; 1440 } else { 1441 clrbuf(bp); 1442 } 1443} 1444 1445/* 1446 * vm_hold_load_pages and vm_hold_unload pages get pages into 1447 * a buffers address space. The pages are anonymous and are 1448 * not associated with a file object. 1449 */ 1450void 1451vm_hold_load_pages(struct buf * bp, vm_offset_t froma, vm_offset_t toa) 1452{ 1453 vm_offset_t pg; 1454 vm_page_t p; 1455 vm_offset_t from = round_page(froma); 1456 vm_offset_t to = round_page(toa); 1457 1458 for (pg = from; pg < to; pg += PAGE_SIZE) { 1459 1460tryagain: 1461 1462 p = vm_page_alloc(kernel_object, pg - VM_MIN_KERNEL_ADDRESS, 1463 VM_ALLOC_NORMAL); 1464 if (!p) { 1465 VM_WAIT; 1466 goto tryagain; 1467 } 1468 vm_page_wire(p); 1469 pmap_kenter(pg, VM_PAGE_TO_PHYS(p)); 1470 bp->b_pages[((caddr_t) pg - bp->b_data) / PAGE_SIZE] = p; 1471 PAGE_WAKEUP(p); 1472 bp->b_npages++; 1473 } 1474} 1475 1476void 1477vm_hold_free_pages(struct buf * bp, vm_offset_t froma, vm_offset_t toa) 1478{ 1479 vm_offset_t pg; 1480 vm_page_t p; 1481 vm_offset_t from = round_page(froma); 1482 vm_offset_t to = round_page(toa); 1483 1484 for (pg = from; pg < to; pg += PAGE_SIZE) { 1485 p = bp->b_pages[((caddr_t) pg - bp->b_data) / PAGE_SIZE]; 1486 bp->b_pages[((caddr_t) pg - bp->b_data) / PAGE_SIZE] = 0; 1487 pmap_kremove(pg); 1488 vm_page_free(p); 1489 --bp->b_npages; 1490 } 1491} 1492