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 1.68 1995/10/29 15:31:13 phk 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/sysproto.h> 39 #include <sys/kernel.h> 40 #include <sys/proc.h> 41 #include <sys/vnode.h> 42 #include <vm/vm.h> 43 #include <vm/vm_kern.h> 44 #include <vm/vm_pageout.h> 45 #include <vm/vm_page.h> 46 #include <vm/vm_object.h> 47 #include <sys/buf.h> 48 #include <sys/mount.h> 49 #include <sys/malloc.h> 50 #include <sys/resourcevar.h> 51 #include <sys/proc.h> 52 53 #include <miscfs/specfs/specdev.h> 54 55 /* 56 * System initialization 57 */ 58 59 static void vfs_update __P((void)); 60 struct proc *updateproc; 61 62 static struct kproc_desc up_kp = { 63 "update", 64 vfs_update, 65 &updateproc 66 }; 67 SYSINIT_KT(update, SI_SUB_KTHREAD_UPDATE, SI_ORDER_FIRST, kproc_start, &up_kp) 68 69 70 struct buf *buf; /* buffer header pool */ 71 struct swqueue bswlist; 72 73 void vm_hold_free_pages(struct buf * bp, vm_offset_t from, vm_offset_t to); 74 void vm_hold_load_pages(struct buf * bp, vm_offset_t from, vm_offset_t to); 75 void vfs_clean_pages(struct buf * bp); 76 static void vfs_setdirty(struct buf *bp); 77 static __inline struct buf * gbincore(struct vnode * vp, daddr_t blkno); 78 79 int needsbuffer; 80 81 /* 82 * Internal update daemon, process 3 83 * The variable vfs_update_wakeup allows for internal syncs. 84 */ 85 int vfs_update_wakeup; 86 87 88 /* 89 * buffers base kva 90 */ 91 caddr_t buffers_kva; 92 93 /* 94 * bogus page -- for I/O to/from partially complete buffers 95 * this is a temporary solution to the problem, but it is not 96 * really that bad. it would be better to split the buffer 97 * for input in the case of buffers partially already in memory, 98 * but the code is intricate enough already. 99 */ 100 vm_page_t bogus_page; 101 vm_offset_t bogus_offset; 102 103 int bufspace, maxbufspace; 104 105 /* 106 * advisory minimum for size of LRU queue or VMIO queue 107 */ 108 int minbuf; 109 110 struct bufhashhdr bufhashtbl[BUFHSZ], invalhash; 111 struct bqueues bufqueues[BUFFER_QUEUES]; 112 113 /* 114 * Initialize buffer headers and related structures. 115 */ 116 void 117 bufinit() 118 { 119 struct buf *bp; 120 int i; 121 122 TAILQ_INIT(&bswlist); 123 LIST_INIT(&invalhash); 124 125 /* first, make a null hash table */ 126 for (i = 0; i < BUFHSZ; i++) 127 LIST_INIT(&bufhashtbl[i]); 128 129 /* next, make a null set of free lists */ 130 for (i = 0; i < BUFFER_QUEUES; i++) 131 TAILQ_INIT(&bufqueues[i]); 132 133 buffers_kva = (caddr_t) kmem_alloc_pageable(buffer_map, MAXBSIZE * nbuf); 134 /* finally, initialize each buffer header and stick on empty q */ 135 for (i = 0; i < nbuf; i++) { 136 bp = &buf[i]; 137 bzero(bp, sizeof *bp); 138 bp->b_flags = B_INVAL; /* we're just an empty header */ 139 bp->b_dev = NODEV; 140 bp->b_rcred = NOCRED; 141 bp->b_wcred = NOCRED; 142 bp->b_qindex = QUEUE_EMPTY; 143 bp->b_vnbufs.le_next = NOLIST; 144 bp->b_data = buffers_kva + i * MAXBSIZE; 145 TAILQ_INSERT_TAIL(&bufqueues[QUEUE_EMPTY], bp, b_freelist); 146 LIST_INSERT_HEAD(&invalhash, bp, b_hash); 147 } 148 /* 149 * maxbufspace is currently calculated to support all filesystem blocks 150 * to be 8K. If you happen to use a 16K filesystem, the size of the buffer 151 * cache is still the same as it would be for 8K filesystems. This 152 * keeps the size of the buffer cache "in check" for big block filesystems. 153 */ 154 minbuf = nbuf / 3; 155 maxbufspace = 2 * (nbuf + 8) * PAGE_SIZE; 156 157 bogus_offset = kmem_alloc_pageable(kernel_map, PAGE_SIZE); 158 bogus_page = vm_page_alloc(kernel_object, 159 bogus_offset - VM_MIN_KERNEL_ADDRESS, VM_ALLOC_NORMAL); 160 161 } 162 163 /* 164 * remove the buffer from the appropriate free list 165 */ 166 void 167 bremfree(struct buf * bp) 168 { 169 int s = splbio(); 170 171 if (bp->b_qindex != QUEUE_NONE) { 172 TAILQ_REMOVE(&bufqueues[bp->b_qindex], bp, b_freelist); 173 bp->b_qindex = QUEUE_NONE; 174 } else { 175 panic("bremfree: removing a buffer when not on a queue"); 176 } 177 splx(s); 178 } 179 180 /* 181 * Get a buffer with the specified data. Look in the cache first. 182 */ 183 int 184 bread(struct vnode * vp, daddr_t blkno, int size, struct ucred * cred, 185 struct buf ** bpp) 186 { 187 struct buf *bp; 188 189 bp = getblk(vp, blkno, size, 0, 0); 190 *bpp = bp; 191 192 /* if not found in cache, do some I/O */ 193 if ((bp->b_flags & B_CACHE) == 0) { 194 if (curproc != NULL) 195 curproc->p_stats->p_ru.ru_inblock++; 196 bp->b_flags |= B_READ; 197 bp->b_flags &= ~(B_DONE | B_ERROR | B_INVAL); 198 if (bp->b_rcred == NOCRED) { 199 if (cred != NOCRED) 200 crhold(cred); 201 bp->b_rcred = cred; 202 } 203 vfs_busy_pages(bp, 0); 204 VOP_STRATEGY(bp); 205 return (biowait(bp)); 206 } 207 return (0); 208 } 209 210 /* 211 * Operates like bread, but also starts asynchronous I/O on 212 * read-ahead blocks. 213 */ 214 int 215 breadn(struct vnode * vp, daddr_t blkno, int size, 216 daddr_t * rablkno, int *rabsize, 217 int cnt, struct ucred * cred, struct buf ** bpp) 218 { 219 struct buf *bp, *rabp; 220 int i; 221 int rv = 0, readwait = 0; 222 223 *bpp = bp = getblk(vp, blkno, size, 0, 0); 224 225 /* if not found in cache, do some I/O */ 226 if ((bp->b_flags & B_CACHE) == 0) { 227 if (curproc != NULL) 228 curproc->p_stats->p_ru.ru_inblock++; 229 bp->b_flags |= B_READ; 230 bp->b_flags &= ~(B_DONE | B_ERROR | B_INVAL); 231 if (bp->b_rcred == NOCRED) { 232 if (cred != NOCRED) 233 crhold(cred); 234 bp->b_rcred = cred; 235 } 236 vfs_busy_pages(bp, 0); 237 VOP_STRATEGY(bp); 238 ++readwait; 239 } 240 for (i = 0; i < cnt; i++, rablkno++, rabsize++) { 241 if (inmem(vp, *rablkno)) 242 continue; 243 rabp = getblk(vp, *rablkno, *rabsize, 0, 0); 244 245 if ((rabp->b_flags & B_CACHE) == 0) { 246 if (curproc != NULL) 247 curproc->p_stats->p_ru.ru_inblock++; 248 rabp->b_flags |= B_READ | B_ASYNC; 249 rabp->b_flags &= ~(B_DONE | B_ERROR | B_INVAL); 250 if (rabp->b_rcred == NOCRED) { 251 if (cred != NOCRED) 252 crhold(cred); 253 rabp->b_rcred = cred; 254 } 255 vfs_busy_pages(rabp, 0); 256 VOP_STRATEGY(rabp); 257 } else { 258 brelse(rabp); 259 } 260 } 261 262 if (readwait) { 263 rv = biowait(bp); 264 } 265 return (rv); 266 } 267 268 /* 269 * Write, release buffer on completion. (Done by iodone 270 * if async.) 271 */ 272 int 273 bwrite(struct buf * bp) 274 { 275 int oldflags = bp->b_flags; 276 277 if (bp->b_flags & B_INVAL) { 278 brelse(bp); 279 return (0); 280 } 281 if (!(bp->b_flags & B_BUSY)) 282 panic("bwrite: buffer is not busy???"); 283 284 bp->b_flags &= ~(B_READ | B_DONE | B_ERROR | B_DELWRI); 285 bp->b_flags |= B_WRITEINPROG; 286 287 if ((oldflags & (B_ASYNC|B_DELWRI)) == (B_ASYNC|B_DELWRI)) { 288 reassignbuf(bp, bp->b_vp); 289 } 290 291 bp->b_vp->v_numoutput++; 292 vfs_busy_pages(bp, 1); 293 if (curproc != NULL) 294 curproc->p_stats->p_ru.ru_oublock++; 295 VOP_STRATEGY(bp); 296 297 if ((oldflags & B_ASYNC) == 0) { 298 int rtval = biowait(bp); 299 300 if (oldflags & B_DELWRI) { 301 reassignbuf(bp, bp->b_vp); 302 } 303 brelse(bp); 304 return (rtval); 305 } 306 return (0); 307 } 308 309 int 310 vn_bwrite(ap) 311 struct vop_bwrite_args *ap; 312 { 313 return (bwrite(ap->a_bp)); 314 } 315 316 /* 317 * Delayed write. (Buffer is marked dirty). 318 */ 319 void 320 bdwrite(struct buf * bp) 321 { 322 323 if ((bp->b_flags & B_BUSY) == 0) { 324 panic("bdwrite: buffer is not busy"); 325 } 326 if (bp->b_flags & B_INVAL) { 327 brelse(bp); 328 return; 329 } 330 if (bp->b_flags & B_TAPE) { 331 bawrite(bp); 332 return; 333 } 334 bp->b_flags &= ~(B_READ|B_RELBUF); 335 if ((bp->b_flags & B_DELWRI) == 0) { 336 bp->b_flags |= B_DONE | B_DELWRI; 337 reassignbuf(bp, bp->b_vp); 338 } 339 340 /* 341 * This bmap keeps the system from needing to do the bmap later, 342 * perhaps when the system is attempting to do a sync. Since it 343 * is likely that the indirect block -- or whatever other datastructure 344 * that the filesystem needs is still in memory now, it is a good 345 * thing to do this. Note also, that if the pageout daemon is 346 * requesting a sync -- there might not be enough memory to do 347 * the bmap then... So, this is important to do. 348 */ 349 if( bp->b_lblkno == bp->b_blkno) { 350 VOP_BMAP(bp->b_vp, bp->b_lblkno, NULL, &bp->b_blkno, NULL, NULL); 351 } 352 353 /* 354 * Set the *dirty* buffer range based upon the VM system dirty pages. 355 */ 356 vfs_setdirty(bp); 357 358 /* 359 * We need to do this here to satisfy the vnode_pager and the 360 * pageout daemon, so that it thinks that the pages have been 361 * "cleaned". Note that since the pages are in a delayed write 362 * buffer -- the VFS layer "will" see that the pages get written 363 * out on the next sync, or perhaps the cluster will be completed. 364 */ 365 vfs_clean_pages(bp); 366 brelse(bp); 367 return; 368 } 369 370 /* 371 * Asynchronous write. 372 * Start output on a buffer, but do not wait for it to complete. 373 * The buffer is released when the output completes. 374 */ 375 void 376 bawrite(struct buf * bp) 377 { 378 bp->b_flags |= B_ASYNC; 379 (void) VOP_BWRITE(bp); 380 } 381 382 /* 383 * Release a buffer. 384 */ 385 void 386 brelse(struct buf * bp) 387 { 388 int s; 389 390 if (bp->b_flags & B_CLUSTER) { 391 relpbuf(bp); 392 return; 393 } 394 /* anyone need a "free" block? */ 395 s = splbio(); 396 397 if (needsbuffer) { 398 needsbuffer = 0; 399 wakeup(&needsbuffer); 400 } 401 402 /* anyone need this block? */ 403 if (bp->b_flags & B_WANTED) { 404 bp->b_flags &= ~(B_WANTED | B_AGE); 405 wakeup(bp); 406 } else if (bp->b_flags & B_VMIO) { 407 bp->b_flags &= ~B_WANTED; 408 wakeup(bp); 409 } 410 if (bp->b_flags & B_LOCKED) 411 bp->b_flags &= ~B_ERROR; 412 413 if ((bp->b_flags & (B_NOCACHE | B_INVAL | B_ERROR)) || 414 (bp->b_bufsize <= 0)) { 415 bp->b_flags |= B_INVAL; 416 bp->b_flags &= ~(B_DELWRI | B_CACHE); 417 if (((bp->b_flags & B_VMIO) == 0) && bp->b_vp) 418 brelvp(bp); 419 } 420 421 /* 422 * VMIO buffer rundown. It is not very necessary to keep a VMIO buffer 423 * constituted, so the B_INVAL flag is used to *invalidate* the buffer, 424 * but the VM object is kept around. The B_NOCACHE flag is used to 425 * invalidate the pages in the VM object. 426 */ 427 if (bp->b_flags & B_VMIO) { 428 vm_offset_t foff; 429 vm_object_t obj; 430 int i, resid; 431 vm_page_t m; 432 struct vnode *vp; 433 int iototal = bp->b_bufsize; 434 435 vp = bp->b_vp; 436 if (!vp) 437 panic("brelse: missing vp"); 438 if (!vp->v_mount) 439 panic("brelse: missing mount info"); 440 441 if (bp->b_npages) { 442 obj = (vm_object_t) vp->v_object; 443 foff = trunc_page(vp->v_mount->mnt_stat.f_iosize * bp->b_lblkno); 444 for (i = 0; i < bp->b_npages; i++) { 445 m = bp->b_pages[i]; 446 if (m == bogus_page) { 447 m = vm_page_lookup(obj, foff); 448 if (!m) { 449 panic("brelse: page missing\n"); 450 } 451 bp->b_pages[i] = m; 452 pmap_qenter(trunc_page(bp->b_data), bp->b_pages, bp->b_npages); 453 } 454 resid = (m->offset + PAGE_SIZE) - foff; 455 if (resid > iototal) 456 resid = iototal; 457 if (resid > 0) { 458 /* 459 * Don't invalidate the page if the local machine has already 460 * modified it. This is the lesser of two evils, and should 461 * be fixed. 462 */ 463 if (bp->b_flags & (B_NOCACHE | B_ERROR)) { 464 vm_page_test_dirty(m); 465 if (m->dirty == 0) { 466 vm_page_set_invalid(m, foff, resid); 467 if (m->valid == 0) 468 vm_page_protect(m, VM_PROT_NONE); 469 } 470 } 471 } 472 foff += resid; 473 iototal -= resid; 474 } 475 } 476 477 if (bp->b_flags & (B_INVAL | B_RELBUF)) { 478 for(i = 0; i < bp->b_npages; i++) { 479 m = bp->b_pages[i]; 480 --m->bmapped; 481 if (m->bmapped == 0) { 482 if (m->flags & PG_WANTED) { 483 wakeup(m); 484 m->flags &= ~PG_WANTED; 485 } 486 if ((m->busy == 0) && ((m->flags & PG_BUSY) == 0)) { 487 vm_page_test_dirty(m); 488 /* 489 * if page isn't valid, no sense in keeping it around 490 */ 491 if (m->valid == 0) { 492 vm_page_protect(m, VM_PROT_NONE); 493 vm_page_free(m); 494 /* 495 * if page isn't dirty and hasn't been referenced by 496 * a process, then cache it 497 */ 498 } else if ((m->dirty & m->valid) == 0 && 499 (m->flags & PG_REFERENCED) == 0 && 500 !pmap_is_referenced(VM_PAGE_TO_PHYS(m))) { 501 vm_page_cache(m); 502 /* 503 * otherwise activate it 504 */ 505 } else if ((m->flags & PG_ACTIVE) == 0) { 506 vm_page_activate(m); 507 m->act_count = 0; 508 } 509 } 510 } 511 } 512 bufspace -= bp->b_bufsize; 513 pmap_qremove(trunc_page((vm_offset_t) bp->b_data), bp->b_npages); 514 bp->b_npages = 0; 515 bp->b_bufsize = 0; 516 bp->b_flags &= ~B_VMIO; 517 if (bp->b_vp) 518 brelvp(bp); 519 } 520 } 521 if (bp->b_qindex != QUEUE_NONE) 522 panic("brelse: free buffer onto another queue???"); 523 524 /* enqueue */ 525 /* buffers with no memory */ 526 if (bp->b_bufsize == 0) { 527 bp->b_qindex = QUEUE_EMPTY; 528 TAILQ_INSERT_TAIL(&bufqueues[QUEUE_EMPTY], bp, b_freelist); 529 LIST_REMOVE(bp, b_hash); 530 LIST_INSERT_HEAD(&invalhash, bp, b_hash); 531 bp->b_dev = NODEV; 532 /* buffers with junk contents */ 533 } else if (bp->b_flags & (B_ERROR | B_INVAL | B_NOCACHE | B_RELBUF)) { 534 bp->b_qindex = QUEUE_AGE; 535 TAILQ_INSERT_HEAD(&bufqueues[QUEUE_AGE], bp, b_freelist); 536 LIST_REMOVE(bp, b_hash); 537 LIST_INSERT_HEAD(&invalhash, bp, b_hash); 538 bp->b_dev = NODEV; 539 /* buffers that are locked */ 540 } else if (bp->b_flags & B_LOCKED) { 541 bp->b_qindex = QUEUE_LOCKED; 542 TAILQ_INSERT_TAIL(&bufqueues[QUEUE_LOCKED], bp, b_freelist); 543 /* buffers with stale but valid contents */ 544 } else if (bp->b_flags & B_AGE) { 545 bp->b_qindex = QUEUE_AGE; 546 TAILQ_INSERT_TAIL(&bufqueues[QUEUE_AGE], bp, b_freelist); 547 /* buffers with valid and quite potentially reuseable contents */ 548 } else { 549 bp->b_qindex = QUEUE_LRU; 550 TAILQ_INSERT_TAIL(&bufqueues[QUEUE_LRU], bp, b_freelist); 551 } 552 553 /* unlock */ 554 bp->b_flags &= ~(B_WANTED | B_BUSY | B_ASYNC | B_NOCACHE | B_AGE | B_RELBUF); 555 splx(s); 556 } 557 558 /* 559 * Check to see if a block is currently memory resident. 560 */ 561 static __inline struct buf * 562 gbincore(struct vnode * vp, daddr_t blkno) 563 { 564 struct buf *bp; 565 struct bufhashhdr *bh; 566 567 bh = BUFHASH(vp, blkno); 568 bp = bh->lh_first; 569 570 /* Search hash chain */ 571 while (bp != NULL) { 572 /* hit */ 573 if (bp->b_vp == vp && bp->b_lblkno == blkno) { 574 break; 575 } 576 bp = bp->b_hash.le_next; 577 } 578 return (bp); 579 } 580 581 /* 582 * this routine implements clustered async writes for 583 * clearing out B_DELWRI buffers... This is much better 584 * than the old way of writing only one buffer at a time. 585 */ 586 void 587 vfs_bio_awrite(struct buf * bp) 588 { 589 int i; 590 daddr_t lblkno = bp->b_lblkno; 591 struct vnode *vp = bp->b_vp; 592 int s; 593 int ncl; 594 struct buf *bpa; 595 596 s = splbio(); 597 if (vp->v_mount && (vp->v_flag & VVMIO) && 598 (bp->b_flags & (B_CLUSTEROK | B_INVAL)) == B_CLUSTEROK) { 599 int size = vp->v_mount->mnt_stat.f_iosize; 600 int maxcl = MAXPHYS / size; 601 602 for (i = 1; i < maxcl; i++) { 603 if ((bpa = gbincore(vp, lblkno + i)) && 604 ((bpa->b_flags & (B_BUSY | B_DELWRI | B_CLUSTEROK | B_INVAL)) == 605 (B_DELWRI | B_CLUSTEROK)) && 606 (bpa->b_bufsize == size)) { 607 if ((bpa->b_blkno == bpa->b_lblkno) || 608 (bpa->b_blkno != bp->b_blkno + (i * size) / DEV_BSIZE)) 609 break; 610 } else { 611 break; 612 } 613 } 614 ncl = i; 615 /* 616 * this is a possible cluster write 617 */ 618 if (ncl != 1) { 619 bremfree(bp); 620 cluster_wbuild(vp, bp, size, lblkno, ncl, -1); 621 splx(s); 622 return; 623 } 624 } 625 /* 626 * default (old) behavior, writing out only one block 627 */ 628 bremfree(bp); 629 bp->b_flags |= B_BUSY | B_ASYNC; 630 (void) VOP_BWRITE(bp); 631 splx(s); 632 } 633 634 635 /* 636 * Find a buffer header which is available for use. 637 */ 638 static struct buf * 639 getnewbuf(int slpflag, int slptimeo, int doingvmio) 640 { 641 struct buf *bp; 642 int s; 643 644 s = splbio(); 645 start: 646 if (bufspace >= maxbufspace) 647 goto trytofreespace; 648 649 /* can we constitute a new buffer? */ 650 if ((bp = bufqueues[QUEUE_EMPTY].tqh_first)) { 651 if (bp->b_qindex != QUEUE_EMPTY) 652 panic("getnewbuf: inconsistent EMPTY queue"); 653 bremfree(bp); 654 goto fillbuf; 655 } 656 trytofreespace: 657 /* 658 * We keep the file I/O from hogging metadata I/O 659 * This is desirable because file data is cached in the 660 * VM/Buffer cache even if a buffer is freed. 661 */ 662 if ((bp = bufqueues[QUEUE_AGE].tqh_first)) { 663 if (bp->b_qindex != QUEUE_AGE) 664 panic("getnewbuf: inconsistent AGE queue"); 665 } else if ((bp = bufqueues[QUEUE_LRU].tqh_first)) { 666 if (bp->b_qindex != QUEUE_LRU) 667 panic("getnewbuf: inconsistent LRU queue"); 668 } 669 if (!bp) { 670 /* wait for a free buffer of any kind */ 671 needsbuffer = 1; 672 tsleep(&needsbuffer, PRIBIO | slpflag, "newbuf", slptimeo); 673 splx(s); 674 return (0); 675 } 676 677 /* if we are a delayed write, convert to an async write */ 678 if ((bp->b_flags & (B_DELWRI | B_INVAL)) == B_DELWRI) { 679 vfs_bio_awrite(bp); 680 if (!slpflag && !slptimeo) { 681 splx(s); 682 return (0); 683 } 684 goto start; 685 } 686 687 if (bp->b_flags & B_WANTED) { 688 bp->b_flags &= ~B_WANTED; 689 wakeup(bp); 690 } 691 bremfree(bp); 692 693 if (bp->b_flags & B_VMIO) { 694 bp->b_flags |= B_RELBUF | B_BUSY | B_DONE; 695 brelse(bp); 696 bremfree(bp); 697 } 698 699 if (bp->b_vp) 700 brelvp(bp); 701 702 /* we are not free, nor do we contain interesting data */ 703 if (bp->b_rcred != NOCRED) 704 crfree(bp->b_rcred); 705 if (bp->b_wcred != NOCRED) 706 crfree(bp->b_wcred); 707 fillbuf: 708 bp->b_flags |= B_BUSY; 709 LIST_REMOVE(bp, b_hash); 710 LIST_INSERT_HEAD(&invalhash, bp, b_hash); 711 splx(s); 712 if (bp->b_bufsize) { 713 allocbuf(bp, 0); 714 } 715 bp->b_flags = B_BUSY; 716 bp->b_dev = NODEV; 717 bp->b_vp = NULL; 718 bp->b_blkno = bp->b_lblkno = 0; 719 bp->b_iodone = 0; 720 bp->b_error = 0; 721 bp->b_resid = 0; 722 bp->b_bcount = 0; 723 bp->b_npages = 0; 724 bp->b_wcred = bp->b_rcred = NOCRED; 725 bp->b_data = buffers_kva + (bp - buf) * MAXBSIZE; 726 bp->b_dirtyoff = bp->b_dirtyend = 0; 727 bp->b_validoff = bp->b_validend = 0; 728 if (bufspace >= maxbufspace) { 729 s = splbio(); 730 bp->b_flags |= B_INVAL; 731 brelse(bp); 732 goto trytofreespace; 733 } 734 return (bp); 735 } 736 737 /* 738 * Check to see if a block is currently memory resident. 739 */ 740 struct buf * 741 incore(struct vnode * vp, daddr_t blkno) 742 { 743 struct buf *bp; 744 struct bufhashhdr *bh; 745 746 int s = splbio(); 747 748 bh = BUFHASH(vp, blkno); 749 bp = bh->lh_first; 750 751 /* Search hash chain */ 752 while (bp != NULL) { 753 /* hit */ 754 if (bp->b_vp == vp && bp->b_lblkno == blkno && 755 (bp->b_flags & B_INVAL) == 0) { 756 break; 757 } 758 bp = bp->b_hash.le_next; 759 } 760 splx(s); 761 return (bp); 762 } 763 764 /* 765 * Returns true if no I/O is needed to access the 766 * associated VM object. This is like incore except 767 * it also hunts around in the VM system for the data. 768 */ 769 770 int 771 inmem(struct vnode * vp, daddr_t blkno) 772 { 773 vm_object_t obj; 774 vm_offset_t off, toff, tinc; 775 vm_page_t m; 776 777 if (incore(vp, blkno)) 778 return 1; 779 if (vp->v_mount == NULL) 780 return 0; 781 if ((vp->v_object == NULL) || (vp->v_flag & VVMIO) == 0) 782 return 0; 783 784 obj = vp->v_object; 785 tinc = PAGE_SIZE; 786 if (tinc > vp->v_mount->mnt_stat.f_iosize) 787 tinc = vp->v_mount->mnt_stat.f_iosize; 788 off = blkno * vp->v_mount->mnt_stat.f_iosize; 789 790 for (toff = 0; toff < vp->v_mount->mnt_stat.f_iosize; toff += tinc) { 791 792 m = vm_page_lookup(obj, trunc_page(toff + off)); 793 if (!m) 794 return 0; 795 if (vm_page_is_valid(m, toff + off, tinc) == 0) 796 return 0; 797 } 798 return 1; 799 } 800 801 /* 802 * now we set the dirty range for the buffer -- 803 * for NFS -- if the file is mapped and pages have 804 * been written to, let it know. We want the 805 * entire range of the buffer to be marked dirty if 806 * any of the pages have been written to for consistancy 807 * with the b_validoff, b_validend set in the nfs write 808 * code, and used by the nfs read code. 809 */ 810 static void 811 vfs_setdirty(struct buf *bp) { 812 int i; 813 vm_object_t object; 814 vm_offset_t boffset, offset; 815 /* 816 * We qualify the scan for modified pages on whether the 817 * object has been flushed yet. The OBJ_WRITEABLE flag 818 * is not cleared simply by protecting pages off. 819 */ 820 if ((bp->b_flags & B_VMIO) && 821 ((object = bp->b_pages[0]->object)->flags & (OBJ_WRITEABLE|OBJ_CLEANING))) { 822 /* 823 * test the pages to see if they have been modified directly 824 * by users through the VM system. 825 */ 826 for (i = 0; i < bp->b_npages; i++) 827 vm_page_test_dirty(bp->b_pages[i]); 828 829 /* 830 * scan forwards for the first page modified 831 */ 832 for (i = 0; i < bp->b_npages; i++) { 833 if (bp->b_pages[i]->dirty) { 834 break; 835 } 836 } 837 boffset = i * PAGE_SIZE; 838 if (boffset < bp->b_dirtyoff) { 839 bp->b_dirtyoff = boffset; 840 } 841 842 /* 843 * scan backwards for the last page modified 844 */ 845 for (i = bp->b_npages - 1; i >= 0; --i) { 846 if (bp->b_pages[i]->dirty) { 847 break; 848 } 849 } 850 boffset = (i + 1) * PAGE_SIZE; 851 offset = boffset + bp->b_pages[0]->offset; 852 if (offset >= object->size) { 853 boffset = object->size - bp->b_pages[0]->offset; 854 } 855 if (bp->b_dirtyend < boffset) { 856 bp->b_dirtyend = boffset; 857 } 858 } 859 } 860 861 /* 862 * Get a block given a specified block and offset into a file/device. 863 */ 864 struct buf * 865 getblk(struct vnode * vp, daddr_t blkno, int size, int slpflag, int slptimeo) 866 { 867 struct buf *bp; 868 int s; 869 struct bufhashhdr *bh; 870 871 s = splbio(); 872 loop: 873 if (bp = gbincore(vp, blkno)) { 874 if (bp->b_flags & (B_BUSY|B_INVAL)) { 875 bp->b_flags |= B_WANTED; 876 if (!tsleep(bp, PRIBIO | slpflag, "getblk", slptimeo)) 877 goto loop; 878 879 splx(s); 880 return (struct buf *) NULL; 881 } 882 bp->b_flags |= B_BUSY | B_CACHE; 883 bremfree(bp); 884 885 /* 886 * check for size inconsistancies (note that they shouldn't happen 887 * but do when filesystems don't handle the size changes correctly.) 888 * We are conservative on metadata and don't just extend the buffer 889 * but write and re-constitute it. 890 */ 891 if (bp->b_bcount != size) { 892 if (bp->b_flags & B_VMIO) { 893 allocbuf(bp, size); 894 } else { 895 bp->b_flags |= B_NOCACHE; 896 VOP_BWRITE(bp); 897 goto loop; 898 } 899 } 900 /* 901 * make sure that all pages in the buffer are valid, if they 902 * aren't, clear the cache flag. 903 * ASSUMPTION: 904 * if the buffer is greater than 1 page in size, it is assumed 905 * that the buffer address starts on a page boundary... 906 */ 907 if (bp->b_flags & B_VMIO) { 908 int szleft, i; 909 szleft = size; 910 for (i=0;i<bp->b_npages;i++) { 911 if (szleft > PAGE_SIZE) { 912 if ((bp->b_pages[i]->valid & VM_PAGE_BITS_ALL) != 913 VM_PAGE_BITS_ALL) { 914 bp->b_flags &= ~(B_CACHE|B_DONE); 915 break; 916 } 917 szleft -= PAGE_SIZE; 918 } else { 919 if (!vm_page_is_valid(bp->b_pages[i], 920 (((vm_offset_t) bp->b_data) & PAGE_MASK), 921 szleft)) { 922 bp->b_flags &= ~(B_CACHE|B_DONE); 923 break; 924 } 925 szleft = 0; 926 } 927 } 928 } 929 splx(s); 930 return (bp); 931 } else { 932 vm_object_t obj; 933 int doingvmio; 934 935 if ((obj = vp->v_object) && (vp->v_flag & VVMIO)) { 936 doingvmio = 1; 937 } else { 938 doingvmio = 0; 939 } 940 if ((bp = getnewbuf(slpflag, slptimeo, doingvmio)) == 0) { 941 if (slpflag || slptimeo) 942 return NULL; 943 goto loop; 944 } 945 946 /* 947 * This code is used to make sure that a buffer is not 948 * created while the getnewbuf routine is blocked. 949 * Normally the vnode is locked so this isn't a problem. 950 * VBLK type I/O requests, however, don't lock the vnode. 951 */ 952 if (!VOP_ISLOCKED(vp) && gbincore(vp, blkno)) { 953 bp->b_flags |= B_INVAL; 954 brelse(bp); 955 goto loop; 956 } 957 958 /* 959 * Insert the buffer into the hash, so that it can 960 * be found by incore. 961 */ 962 bp->b_blkno = bp->b_lblkno = blkno; 963 bgetvp(vp, bp); 964 LIST_REMOVE(bp, b_hash); 965 bh = BUFHASH(vp, blkno); 966 LIST_INSERT_HEAD(bh, bp, b_hash); 967 968 if (doingvmio) { 969 bp->b_flags |= (B_VMIO | B_CACHE); 970 #if defined(VFS_BIO_DEBUG) 971 if (vp->v_type != VREG) 972 printf("getblk: vmioing file type %d???\n", vp->v_type); 973 #endif 974 } else { 975 bp->b_flags &= ~B_VMIO; 976 } 977 splx(s); 978 979 allocbuf(bp, size); 980 return (bp); 981 } 982 } 983 984 /* 985 * Get an empty, disassociated buffer of given size. 986 */ 987 struct buf * 988 geteblk(int size) 989 { 990 struct buf *bp; 991 992 while ((bp = getnewbuf(0, 0, 0)) == 0); 993 allocbuf(bp, size); 994 bp->b_flags |= B_INVAL; 995 return (bp); 996 } 997 998 /* 999 * This code constitutes the buffer memory from either anonymous system 1000 * memory (in the case of non-VMIO operations) or from an associated 1001 * VM object (in the case of VMIO operations). 1002 * 1003 * Note that this code is tricky, and has many complications to resolve 1004 * deadlock or inconsistant data situations. Tread lightly!!! 1005 * 1006 * Modify the length of a buffer's underlying buffer storage without 1007 * destroying information (unless, of course the buffer is shrinking). 1008 */ 1009 int 1010 allocbuf(struct buf * bp, int size) 1011 { 1012 1013 int s; 1014 int newbsize, mbsize; 1015 int i; 1016 1017 if (!(bp->b_flags & B_BUSY)) 1018 panic("allocbuf: buffer not busy"); 1019 1020 if ((bp->b_flags & B_VMIO) == 0) { 1021 /* 1022 * Just get anonymous memory from the kernel 1023 */ 1024 mbsize = ((size + DEV_BSIZE - 1) / DEV_BSIZE) * DEV_BSIZE; 1025 newbsize = round_page(size); 1026 1027 if (newbsize < bp->b_bufsize) { 1028 vm_hold_free_pages( 1029 bp, 1030 (vm_offset_t) bp->b_data + newbsize, 1031 (vm_offset_t) bp->b_data + bp->b_bufsize); 1032 } else if (newbsize > bp->b_bufsize) { 1033 vm_hold_load_pages( 1034 bp, 1035 (vm_offset_t) bp->b_data + bp->b_bufsize, 1036 (vm_offset_t) bp->b_data + newbsize); 1037 } 1038 } else { 1039 vm_page_t m; 1040 int desiredpages; 1041 1042 newbsize = ((size + DEV_BSIZE - 1) / DEV_BSIZE) * DEV_BSIZE; 1043 desiredpages = round_page(newbsize) / PAGE_SIZE; 1044 1045 if (newbsize < bp->b_bufsize) { 1046 if (desiredpages < bp->b_npages) { 1047 pmap_qremove((vm_offset_t) trunc_page(bp->b_data) + 1048 desiredpages * PAGE_SIZE, (bp->b_npages - desiredpages)); 1049 for (i = desiredpages; i < bp->b_npages; i++) { 1050 m = bp->b_pages[i]; 1051 s = splhigh(); 1052 while ((m->flags & PG_BUSY) || (m->busy != 0)) { 1053 m->flags |= PG_WANTED; 1054 tsleep(m, PVM, "biodep", 0); 1055 } 1056 splx(s); 1057 1058 if (m->bmapped == 0) { 1059 printf("allocbuf: bmapped is zero for page %d\n", i); 1060 panic("allocbuf: error"); 1061 } 1062 --m->bmapped; 1063 if (m->bmapped == 0) { 1064 vm_page_protect(m, VM_PROT_NONE); 1065 vm_page_free(m); 1066 } 1067 bp->b_pages[i] = NULL; 1068 } 1069 bp->b_npages = desiredpages; 1070 } 1071 } else if (newbsize > bp->b_bufsize) { 1072 vm_object_t obj; 1073 vm_offset_t tinc, off, toff, objoff; 1074 int pageindex, curbpnpages; 1075 struct vnode *vp; 1076 int bsize; 1077 1078 vp = bp->b_vp; 1079 bsize = vp->v_mount->mnt_stat.f_iosize; 1080 1081 if (bp->b_npages < desiredpages) { 1082 obj = vp->v_object; 1083 tinc = PAGE_SIZE; 1084 if (tinc > bsize) 1085 tinc = bsize; 1086 off = bp->b_lblkno * bsize; 1087 doretry: 1088 curbpnpages = bp->b_npages; 1089 bp->b_flags |= B_CACHE; 1090 for (toff = 0; toff < newbsize; toff += tinc) { 1091 int bytesinpage; 1092 1093 pageindex = toff / PAGE_SIZE; 1094 objoff = trunc_page(toff + off); 1095 if (pageindex < curbpnpages) { 1096 1097 m = bp->b_pages[pageindex]; 1098 if (m->offset != objoff) 1099 panic("allocbuf: page changed offset??!!!?"); 1100 bytesinpage = tinc; 1101 if (tinc > (newbsize - toff)) 1102 bytesinpage = newbsize - toff; 1103 if (!vm_page_is_valid(m, toff + off, bytesinpage)) { 1104 bp->b_flags &= ~B_CACHE; 1105 } 1106 if ((m->flags & PG_ACTIVE) == 0) { 1107 vm_page_activate(m); 1108 m->act_count = 0; 1109 } 1110 continue; 1111 } 1112 m = vm_page_lookup(obj, objoff); 1113 if (!m) { 1114 m = vm_page_alloc(obj, objoff, VM_ALLOC_NORMAL); 1115 if (!m) { 1116 int j; 1117 1118 for (j = bp->b_npages; j < pageindex; j++) { 1119 PAGE_WAKEUP(bp->b_pages[j]); 1120 } 1121 VM_WAIT; 1122 goto doretry; 1123 } 1124 vm_page_activate(m); 1125 m->act_count = 0; 1126 m->valid = 0; 1127 bp->b_flags &= ~B_CACHE; 1128 } else if (m->flags & PG_BUSY) { 1129 int j; 1130 1131 for (j = bp->b_npages; j < pageindex; j++) { 1132 PAGE_WAKEUP(bp->b_pages[j]); 1133 } 1134 1135 s = splbio(); 1136 m->flags |= PG_WANTED; 1137 tsleep(m, PRIBIO, "pgtblk", 0); 1138 splx(s); 1139 1140 goto doretry; 1141 } else { 1142 if ((curproc != pageproc) && 1143 (m->flags & PG_CACHE) && 1144 (cnt.v_free_count + cnt.v_cache_count) < cnt.v_free_min) { 1145 pagedaemon_wakeup(); 1146 } 1147 bytesinpage = tinc; 1148 if (tinc > (newbsize - toff)) 1149 bytesinpage = newbsize - toff; 1150 if (!vm_page_is_valid(m, toff + off, bytesinpage)) { 1151 bp->b_flags &= ~B_CACHE; 1152 } 1153 if ((m->flags & PG_ACTIVE) == 0) { 1154 vm_page_activate(m); 1155 m->act_count = 0; 1156 } 1157 m->flags |= PG_BUSY; 1158 } 1159 bp->b_pages[pageindex] = m; 1160 curbpnpages = pageindex + 1; 1161 } 1162 for (i = bp->b_npages; i < curbpnpages; i++) { 1163 m = bp->b_pages[i]; 1164 m->bmapped++; 1165 PAGE_WAKEUP(m); 1166 } 1167 bp->b_npages = curbpnpages; 1168 bp->b_data = buffers_kva + (bp - buf) * MAXBSIZE; 1169 pmap_qenter((vm_offset_t) bp->b_data, bp->b_pages, bp->b_npages); 1170 bp->b_data += off % PAGE_SIZE; 1171 } 1172 } 1173 } 1174 bufspace += (newbsize - bp->b_bufsize); 1175 bp->b_bufsize = newbsize; 1176 bp->b_bcount = size; 1177 return 1; 1178 } 1179 1180 /* 1181 * Wait for buffer I/O completion, returning error status. 1182 */ 1183 int 1184 biowait(register struct buf * bp) 1185 { 1186 int s; 1187 1188 s = splbio(); 1189 while ((bp->b_flags & B_DONE) == 0) 1190 tsleep(bp, PRIBIO, "biowait", 0); 1191 splx(s); 1192 if (bp->b_flags & B_EINTR) { 1193 bp->b_flags &= ~B_EINTR; 1194 return (EINTR); 1195 } 1196 if (bp->b_flags & B_ERROR) { 1197 return (bp->b_error ? bp->b_error : EIO); 1198 } else { 1199 return (0); 1200 } 1201 } 1202 1203 /* 1204 * Finish I/O on a buffer, calling an optional function. 1205 * This is usually called from interrupt level, so process blocking 1206 * is not *a good idea*. 1207 */ 1208 void 1209 biodone(register struct buf * bp) 1210 { 1211 int s; 1212 1213 s = splbio(); 1214 if (!(bp->b_flags & B_BUSY)) 1215 panic("biodone: buffer not busy"); 1216 1217 if (bp->b_flags & B_DONE) { 1218 splx(s); 1219 printf("biodone: buffer already done\n"); 1220 return; 1221 } 1222 bp->b_flags |= B_DONE; 1223 1224 if ((bp->b_flags & B_READ) == 0) { 1225 vwakeup(bp); 1226 } 1227 #ifdef BOUNCE_BUFFERS 1228 if (bp->b_flags & B_BOUNCE) 1229 vm_bounce_free(bp); 1230 #endif 1231 1232 /* call optional completion function if requested */ 1233 if (bp->b_flags & B_CALL) { 1234 bp->b_flags &= ~B_CALL; 1235 (*bp->b_iodone) (bp); 1236 splx(s); 1237 return; 1238 } 1239 if (bp->b_flags & B_VMIO) { 1240 int i, resid; 1241 vm_offset_t foff; 1242 vm_page_t m; 1243 vm_object_t obj; 1244 int iosize; 1245 struct vnode *vp = bp->b_vp; 1246 1247 foff = vp->v_mount->mnt_stat.f_iosize * bp->b_lblkno; 1248 obj = vp->v_object; 1249 if (!obj) { 1250 panic("biodone: no object"); 1251 } 1252 #if defined(VFS_BIO_DEBUG) 1253 if (obj->paging_in_progress < bp->b_npages) { 1254 printf("biodone: paging in progress(%d) < bp->b_npages(%d)\n", 1255 obj->paging_in_progress, bp->b_npages); 1256 } 1257 #endif 1258 iosize = bp->b_bufsize; 1259 for (i = 0; i < bp->b_npages; i++) { 1260 int bogusflag = 0; 1261 m = bp->b_pages[i]; 1262 if (m == bogus_page) { 1263 bogusflag = 1; 1264 m = vm_page_lookup(obj, foff); 1265 if (!m) { 1266 #if defined(VFS_BIO_DEBUG) 1267 printf("biodone: page disappeared\n"); 1268 #endif 1269 --obj->paging_in_progress; 1270 continue; 1271 } 1272 bp->b_pages[i] = m; 1273 pmap_qenter(trunc_page(bp->b_data), bp->b_pages, bp->b_npages); 1274 } 1275 #if defined(VFS_BIO_DEBUG) 1276 if (trunc_page(foff) != m->offset) { 1277 printf("biodone: foff(%d)/m->offset(%d) mismatch\n", foff, m->offset); 1278 } 1279 #endif 1280 resid = (m->offset + PAGE_SIZE) - foff; 1281 if (resid > iosize) 1282 resid = iosize; 1283 /* 1284 * In the write case, the valid and clean bits are 1285 * already changed correctly, so we only need to do this 1286 * here in the read case. 1287 */ 1288 if ((bp->b_flags & B_READ) && !bogusflag && resid > 0) { 1289 vm_page_set_validclean(m, foff & (PAGE_SIZE-1), resid); 1290 } 1291 1292 /* 1293 * when debugging new filesystems or buffer I/O methods, this 1294 * is the most common error that pops up. if you see this, you 1295 * have not set the page busy flag correctly!!! 1296 */ 1297 if (m->busy == 0) { 1298 printf("biodone: page busy < 0, " 1299 "off: %ld, foff: %ld, " 1300 "resid: %d, index: %d\n", 1301 m->offset, foff, resid, i); 1302 printf(" iosize: %ld, lblkno: %ld, flags: 0x%x, npages: %d\n", 1303 bp->b_vp->v_mount->mnt_stat.f_iosize, 1304 bp->b_lblkno, bp->b_flags, bp->b_npages); 1305 printf(" valid: 0x%x, dirty: 0x%x, mapped: %d\n", 1306 m->valid, m->dirty, m->bmapped); 1307 panic("biodone: page busy < 0\n"); 1308 } 1309 --m->busy; 1310 if ((m->busy == 0) && (m->flags & PG_WANTED)) { 1311 m->flags &= ~PG_WANTED; 1312 wakeup(m); 1313 } 1314 --obj->paging_in_progress; 1315 foff += resid; 1316 iosize -= resid; 1317 } 1318 if (obj && obj->paging_in_progress == 0 && 1319 (obj->flags & OBJ_PIPWNT)) { 1320 obj->flags &= ~OBJ_PIPWNT; 1321 wakeup(obj); 1322 } 1323 } 1324 /* 1325 * For asynchronous completions, release the buffer now. The brelse 1326 * checks for B_WANTED and will do the wakeup there if necessary - so 1327 * no need to do a wakeup here in the async case. 1328 */ 1329 1330 if (bp->b_flags & B_ASYNC) { 1331 brelse(bp); 1332 } else { 1333 bp->b_flags &= ~B_WANTED; 1334 wakeup(bp); 1335 } 1336 splx(s); 1337 } 1338 1339 int 1340 count_lock_queue() 1341 { 1342 int count; 1343 struct buf *bp; 1344 1345 count = 0; 1346 for (bp = bufqueues[QUEUE_LOCKED].tqh_first; 1347 bp != NULL; 1348 bp = bp->b_freelist.tqe_next) 1349 count++; 1350 return (count); 1351 } 1352 1353 int vfs_update_interval = 30; 1354 1355 void 1356 vfs_update() 1357 { 1358 (void) spl0(); 1359 while (1) { 1360 tsleep(&vfs_update_wakeup, PRIBIO, "update", 1361 hz * vfs_update_interval); 1362 vfs_update_wakeup = 0; 1363 sync(curproc, NULL, NULL); 1364 } 1365 } 1366 1367 /* 1368 * This routine is called in lieu of iodone in the case of 1369 * incomplete I/O. This keeps the busy status for pages 1370 * consistant. 1371 */ 1372 void 1373 vfs_unbusy_pages(struct buf * bp) 1374 { 1375 int i; 1376 1377 if (bp->b_flags & B_VMIO) { 1378 struct vnode *vp = bp->b_vp; 1379 vm_object_t obj = vp->v_object; 1380 vm_offset_t foff; 1381 1382 foff = trunc_page(vp->v_mount->mnt_stat.f_iosize * bp->b_lblkno); 1383 1384 for (i = 0; i < bp->b_npages; i++) { 1385 vm_page_t m = bp->b_pages[i]; 1386 1387 if (m == bogus_page) { 1388 m = vm_page_lookup(obj, foff + i * PAGE_SIZE); 1389 if (!m) { 1390 panic("vfs_unbusy_pages: page missing\n"); 1391 } 1392 bp->b_pages[i] = m; 1393 pmap_qenter(trunc_page(bp->b_data), bp->b_pages, bp->b_npages); 1394 } 1395 --obj->paging_in_progress; 1396 --m->busy; 1397 if ((m->busy == 0) && (m->flags & PG_WANTED)) { 1398 m->flags &= ~PG_WANTED; 1399 wakeup(m); 1400 } 1401 } 1402 if (obj->paging_in_progress == 0 && 1403 (obj->flags & OBJ_PIPWNT)) { 1404 obj->flags &= ~OBJ_PIPWNT; 1405 wakeup(obj); 1406 } 1407 } 1408 } 1409 1410 /* 1411 * This routine is called before a device strategy routine. 1412 * It is used to tell the VM system that paging I/O is in 1413 * progress, and treat the pages associated with the buffer 1414 * almost as being PG_BUSY. Also the object paging_in_progress 1415 * flag is handled to make sure that the object doesn't become 1416 * inconsistant. 1417 */ 1418 void 1419 vfs_busy_pages(struct buf * bp, int clear_modify) 1420 { 1421 int i; 1422 1423 if (bp->b_flags & B_VMIO) { 1424 vm_object_t obj = bp->b_vp->v_object; 1425 vm_offset_t foff = bp->b_vp->v_mount->mnt_stat.f_iosize * bp->b_lblkno; 1426 int iocount = bp->b_bufsize; 1427 1428 vfs_setdirty(bp); 1429 for (i = 0; i < bp->b_npages; i++) { 1430 vm_page_t m = bp->b_pages[i]; 1431 int resid = (m->offset + PAGE_SIZE) - foff; 1432 1433 if (resid > iocount) 1434 resid = iocount; 1435 if ((bp->b_flags & B_CLUSTER) == 0) { 1436 obj->paging_in_progress++; 1437 m->busy++; 1438 } 1439 if (clear_modify) { 1440 vm_page_protect(m, VM_PROT_READ); 1441 vm_page_set_validclean(m, 1442 foff & (PAGE_SIZE-1), resid); 1443 } else if (bp->b_bcount >= PAGE_SIZE) { 1444 if (m->valid && (bp->b_flags & B_CACHE) == 0) { 1445 bp->b_pages[i] = bogus_page; 1446 pmap_qenter(trunc_page(bp->b_data), bp->b_pages, bp->b_npages); 1447 } 1448 } 1449 foff += resid; 1450 iocount -= resid; 1451 } 1452 } 1453 } 1454 1455 /* 1456 * Tell the VM system that the pages associated with this buffer 1457 * are clean. This is used for delayed writes where the data is 1458 * going to go to disk eventually without additional VM intevention. 1459 */ 1460 void 1461 vfs_clean_pages(struct buf * bp) 1462 { 1463 int i; 1464 1465 if (bp->b_flags & B_VMIO) { 1466 vm_offset_t foff = 1467 bp->b_vp->v_mount->mnt_stat.f_iosize * bp->b_lblkno; 1468 int iocount = bp->b_bufsize; 1469 1470 for (i = 0; i < bp->b_npages; i++) { 1471 vm_page_t m = bp->b_pages[i]; 1472 int resid = (m->offset + PAGE_SIZE) - foff; 1473 1474 if (resid > iocount) 1475 resid = iocount; 1476 if (resid > 0) { 1477 vm_page_set_validclean(m, 1478 foff & (PAGE_SIZE-1), resid); 1479 } 1480 foff += resid; 1481 iocount -= resid; 1482 } 1483 } 1484 } 1485 1486 void 1487 vfs_bio_clrbuf(struct buf *bp) { 1488 int i; 1489 if( bp->b_flags & B_VMIO) { 1490 if( (bp->b_npages == 1) && (bp->b_bufsize < PAGE_SIZE)) { 1491 int j; 1492 if( bp->b_pages[0]->valid != VM_PAGE_BITS_ALL) { 1493 for(j=0; j < bp->b_bufsize / DEV_BSIZE;j++) { 1494 bzero(bp->b_data + j * DEV_BSIZE, DEV_BSIZE); 1495 } 1496 } 1497 bp->b_resid = 0; 1498 return; 1499 } 1500 for(i=0;i<bp->b_npages;i++) { 1501 if( bp->b_pages[i]->valid == VM_PAGE_BITS_ALL) 1502 continue; 1503 if( bp->b_pages[i]->valid == 0) { 1504 bzero(bp->b_data + i * PAGE_SIZE, PAGE_SIZE); 1505 } else { 1506 int j; 1507 for(j=0;j<PAGE_SIZE/DEV_BSIZE;j++) { 1508 if( (bp->b_pages[i]->valid & (1<<j)) == 0) 1509 bzero(bp->b_data + i * PAGE_SIZE + j * DEV_BSIZE, DEV_BSIZE); 1510 } 1511 } 1512 bp->b_pages[i]->valid = VM_PAGE_BITS_ALL; 1513 } 1514 bp->b_resid = 0; 1515 } else { 1516 clrbuf(bp); 1517 } 1518 } 1519 1520 /* 1521 * vm_hold_load_pages and vm_hold_unload pages get pages into 1522 * a buffers address space. The pages are anonymous and are 1523 * not associated with a file object. 1524 */ 1525 void 1526 vm_hold_load_pages(struct buf * bp, vm_offset_t froma, vm_offset_t toa) 1527 { 1528 vm_offset_t pg; 1529 vm_page_t p; 1530 vm_offset_t from = round_page(froma); 1531 vm_offset_t to = round_page(toa); 1532 1533 for (pg = from; pg < to; pg += PAGE_SIZE) { 1534 1535 tryagain: 1536 1537 p = vm_page_alloc(kernel_object, pg - VM_MIN_KERNEL_ADDRESS, 1538 VM_ALLOC_NORMAL); 1539 if (!p) { 1540 VM_WAIT; 1541 goto tryagain; 1542 } 1543 vm_page_wire(p); 1544 pmap_kenter(pg, VM_PAGE_TO_PHYS(p)); 1545 bp->b_pages[((caddr_t) pg - bp->b_data) / PAGE_SIZE] = p; 1546 PAGE_WAKEUP(p); 1547 bp->b_npages++; 1548 } 1549 } 1550 1551 void 1552 vm_hold_free_pages(struct buf * bp, vm_offset_t froma, vm_offset_t toa) 1553 { 1554 vm_offset_t pg; 1555 vm_page_t p; 1556 vm_offset_t from = round_page(froma); 1557 vm_offset_t to = round_page(toa); 1558 1559 for (pg = from; pg < to; pg += PAGE_SIZE) { 1560 p = bp->b_pages[((caddr_t) pg - bp->b_data) / PAGE_SIZE]; 1561 bp->b_pages[((caddr_t) pg - bp->b_data) / PAGE_SIZE] = 0; 1562 pmap_kremove(pg); 1563 vm_page_free(p); 1564 --bp->b_npages; 1565 } 1566 } 1567