1 /* 2 * Copyright (c) 1994,1997 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. Absolutely no warranty of function or purpose is made by the author 12 * John S. Dyson. 13 * 14 * $Id: vfs_bio.c,v 1.189 1998/12/22 18:57:30 dillon Exp $ 15 */ 16 17 /* 18 * this file contains a new buffer I/O scheme implementing a coherent 19 * VM object and buffer cache scheme. Pains have been taken to make 20 * sure that the performance degradation associated with schemes such 21 * as this is not realized. 22 * 23 * Author: John S. Dyson 24 * Significant help during the development and debugging phases 25 * had been provided by David Greenman, also of the FreeBSD core team. 26 * 27 * see man buf(9) for more info. 28 */ 29 30 #define VMIO 31 #include <sys/param.h> 32 #include <sys/systm.h> 33 #include <sys/sysproto.h> 34 #include <sys/kernel.h> 35 #include <sys/sysctl.h> 36 #include <sys/proc.h> 37 #include <sys/vnode.h> 38 #include <sys/vmmeter.h> 39 #include <sys/lock.h> 40 #include <miscfs/specfs/specdev.h> 41 #include <vm/vm.h> 42 #include <vm/vm_param.h> 43 #include <vm/vm_prot.h> 44 #include <vm/vm_kern.h> 45 #include <vm/vm_pageout.h> 46 #include <vm/vm_page.h> 47 #include <vm/vm_object.h> 48 #include <vm/vm_extern.h> 49 #include <vm/vm_map.h> 50 #include <sys/buf.h> 51 #include <sys/mount.h> 52 #include <sys/malloc.h> 53 #include <sys/resourcevar.h> 54 55 static MALLOC_DEFINE(M_BIOBUF, "BIO buffer", "BIO buffer"); 56 57 struct bio_ops bioops; /* I/O operation notification */ 58 59 #if 0 /* replaced bu sched_sync */ 60 static void vfs_update __P((void)); 61 static struct proc *updateproc; 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 #endif 69 70 struct buf *buf; /* buffer header pool */ 71 struct swqueue bswlist; 72 73 static void vm_hold_free_pages(struct buf * bp, vm_offset_t from, 74 vm_offset_t to); 75 static void vm_hold_load_pages(struct buf * bp, vm_offset_t from, 76 vm_offset_t to); 77 static void vfs_buf_set_valid(struct buf *bp, vm_ooffset_t foff, 78 vm_offset_t off, vm_offset_t size, 79 vm_page_t m); 80 static void vfs_page_set_valid(struct buf *bp, vm_ooffset_t off, 81 int pageno, vm_page_t m); 82 static void vfs_clean_pages(struct buf * bp); 83 static void vfs_setdirty(struct buf *bp); 84 static void vfs_vmio_release(struct buf *bp); 85 static void flushdirtybuffers(int slpflag, int slptimeo); 86 87 int needsbuffer; 88 89 /* 90 * Internal update daemon, process 3 91 * The variable vfs_update_wakeup allows for internal syncs. 92 */ 93 int vfs_update_wakeup; 94 95 96 /* 97 * buffers base kva 98 */ 99 100 /* 101 * bogus page -- for I/O to/from partially complete buffers 102 * this is a temporary solution to the problem, but it is not 103 * really that bad. it would be better to split the buffer 104 * for input in the case of buffers partially already in memory, 105 * but the code is intricate enough already. 106 */ 107 vm_page_t bogus_page; 108 static vm_offset_t bogus_offset; 109 110 static int bufspace, maxbufspace, vmiospace, maxvmiobufspace, 111 bufmallocspace, maxbufmallocspace; 112 int numdirtybuffers; 113 static int lodirtybuffers, hidirtybuffers; 114 static int numfreebuffers, lofreebuffers, hifreebuffers; 115 static int kvafreespace; 116 117 SYSCTL_INT(_vfs, OID_AUTO, numdirtybuffers, CTLFLAG_RD, 118 &numdirtybuffers, 0, ""); 119 SYSCTL_INT(_vfs, OID_AUTO, lodirtybuffers, CTLFLAG_RW, 120 &lodirtybuffers, 0, ""); 121 SYSCTL_INT(_vfs, OID_AUTO, hidirtybuffers, CTLFLAG_RW, 122 &hidirtybuffers, 0, ""); 123 SYSCTL_INT(_vfs, OID_AUTO, numfreebuffers, CTLFLAG_RD, 124 &numfreebuffers, 0, ""); 125 SYSCTL_INT(_vfs, OID_AUTO, lofreebuffers, CTLFLAG_RW, 126 &lofreebuffers, 0, ""); 127 SYSCTL_INT(_vfs, OID_AUTO, hifreebuffers, CTLFLAG_RW, 128 &hifreebuffers, 0, ""); 129 SYSCTL_INT(_vfs, OID_AUTO, maxbufspace, CTLFLAG_RW, 130 &maxbufspace, 0, ""); 131 SYSCTL_INT(_vfs, OID_AUTO, bufspace, CTLFLAG_RD, 132 &bufspace, 0, ""); 133 SYSCTL_INT(_vfs, OID_AUTO, maxvmiobufspace, CTLFLAG_RW, 134 &maxvmiobufspace, 0, ""); 135 SYSCTL_INT(_vfs, OID_AUTO, vmiospace, CTLFLAG_RD, 136 &vmiospace, 0, ""); 137 SYSCTL_INT(_vfs, OID_AUTO, maxmallocbufspace, CTLFLAG_RW, 138 &maxbufmallocspace, 0, ""); 139 SYSCTL_INT(_vfs, OID_AUTO, bufmallocspace, CTLFLAG_RD, 140 &bufmallocspace, 0, ""); 141 SYSCTL_INT(_vfs, OID_AUTO, kvafreespace, CTLFLAG_RD, 142 &kvafreespace, 0, ""); 143 144 static LIST_HEAD(bufhashhdr, buf) bufhashtbl[BUFHSZ], invalhash; 145 struct bqueues bufqueues[BUFFER_QUEUES] = {0}; 146 147 extern int vm_swap_size; 148 149 #define BUF_MAXUSE 24 150 151 #define VFS_BIO_NEED_ANY 1 152 #define VFS_BIO_NEED_LOWLIMIT 2 153 #define VFS_BIO_NEED_FREE 4 154 155 /* 156 * Initialize buffer headers and related structures. 157 */ 158 void 159 bufinit() 160 { 161 struct buf *bp; 162 int i; 163 164 TAILQ_INIT(&bswlist); 165 LIST_INIT(&invalhash); 166 167 /* first, make a null hash table */ 168 for (i = 0; i < BUFHSZ; i++) 169 LIST_INIT(&bufhashtbl[i]); 170 171 /* next, make a null set of free lists */ 172 for (i = 0; i < BUFFER_QUEUES; i++) 173 TAILQ_INIT(&bufqueues[i]); 174 175 /* finally, initialize each buffer header and stick on empty q */ 176 for (i = 0; i < nbuf; i++) { 177 bp = &buf[i]; 178 bzero(bp, sizeof *bp); 179 bp->b_flags = B_INVAL; /* we're just an empty header */ 180 bp->b_dev = NODEV; 181 bp->b_rcred = NOCRED; 182 bp->b_wcred = NOCRED; 183 bp->b_qindex = QUEUE_EMPTY; 184 bp->b_xflags = 0; 185 LIST_INIT(&bp->b_dep); 186 TAILQ_INSERT_TAIL(&bufqueues[QUEUE_EMPTY], bp, b_freelist); 187 LIST_INSERT_HEAD(&invalhash, bp, b_hash); 188 } 189 /* 190 * maxbufspace is currently calculated to support all filesystem blocks 191 * to be 8K. If you happen to use a 16K filesystem, the size of the buffer 192 * cache is still the same as it would be for 8K filesystems. This 193 * keeps the size of the buffer cache "in check" for big block filesystems. 194 */ 195 maxbufspace = (nbuf + 8) * DFLTBSIZE; 196 /* 197 * reserve 1/3 of the buffers for metadata (VDIR) which might not be VMIO'ed 198 */ 199 maxvmiobufspace = 2 * maxbufspace / 3; 200 /* 201 * Limit the amount of malloc memory since it is wired permanently into 202 * the kernel space. Even though this is accounted for in the buffer 203 * allocation, we don't want the malloced region to grow uncontrolled. 204 * The malloc scheme improves memory utilization significantly on average 205 * (small) directories. 206 */ 207 maxbufmallocspace = maxbufspace / 20; 208 209 /* 210 * Remove the probability of deadlock conditions by limiting the 211 * number of dirty buffers. 212 */ 213 hidirtybuffers = nbuf / 8 + 20; 214 lodirtybuffers = nbuf / 16 + 10; 215 numdirtybuffers = 0; 216 lofreebuffers = nbuf / 18 + 5; 217 hifreebuffers = 2 * lofreebuffers; 218 numfreebuffers = nbuf; 219 kvafreespace = 0; 220 221 bogus_offset = kmem_alloc_pageable(kernel_map, PAGE_SIZE); 222 bogus_page = vm_page_alloc(kernel_object, 223 ((bogus_offset - VM_MIN_KERNEL_ADDRESS) >> PAGE_SHIFT), 224 VM_ALLOC_NORMAL); 225 226 } 227 228 /* 229 * Free the kva allocation for a buffer 230 * Must be called only at splbio or higher, 231 * as this is the only locking for buffer_map. 232 */ 233 static void 234 bfreekva(struct buf * bp) 235 { 236 if (bp->b_kvasize == 0) 237 return; 238 239 vm_map_delete(buffer_map, 240 (vm_offset_t) bp->b_kvabase, 241 (vm_offset_t) bp->b_kvabase + bp->b_kvasize); 242 243 bp->b_kvasize = 0; 244 245 } 246 247 /* 248 * remove the buffer from the appropriate free list 249 */ 250 void 251 bremfree(struct buf * bp) 252 { 253 int s = splbio(); 254 255 if (bp->b_qindex != QUEUE_NONE) { 256 if (bp->b_qindex == QUEUE_EMPTY) { 257 kvafreespace -= bp->b_kvasize; 258 } 259 TAILQ_REMOVE(&bufqueues[bp->b_qindex], bp, b_freelist); 260 bp->b_qindex = QUEUE_NONE; 261 } else { 262 #if !defined(MAX_PERF) 263 panic("bremfree: removing a buffer when not on a queue"); 264 #endif 265 } 266 if ((bp->b_flags & B_INVAL) || 267 (bp->b_flags & (B_DELWRI|B_LOCKED)) == 0) 268 --numfreebuffers; 269 splx(s); 270 } 271 272 273 /* 274 * Get a buffer with the specified data. Look in the cache first. 275 */ 276 int 277 bread(struct vnode * vp, daddr_t blkno, int size, struct ucred * cred, 278 struct buf ** bpp) 279 { 280 struct buf *bp; 281 282 bp = getblk(vp, blkno, size, 0, 0); 283 *bpp = bp; 284 285 /* if not found in cache, do some I/O */ 286 if ((bp->b_flags & B_CACHE) == 0) { 287 if (curproc != NULL) 288 curproc->p_stats->p_ru.ru_inblock++; 289 bp->b_flags |= B_READ; 290 bp->b_flags &= ~(B_DONE | B_ERROR | B_INVAL); 291 if (bp->b_rcred == NOCRED) { 292 if (cred != NOCRED) 293 crhold(cred); 294 bp->b_rcred = cred; 295 } 296 vfs_busy_pages(bp, 0); 297 VOP_STRATEGY(vp, bp); 298 return (biowait(bp)); 299 } 300 return (0); 301 } 302 303 /* 304 * Operates like bread, but also starts asynchronous I/O on 305 * read-ahead blocks. 306 */ 307 int 308 breadn(struct vnode * vp, daddr_t blkno, int size, 309 daddr_t * rablkno, int *rabsize, 310 int cnt, struct ucred * cred, struct buf ** bpp) 311 { 312 struct buf *bp, *rabp; 313 int i; 314 int rv = 0, readwait = 0; 315 316 *bpp = bp = getblk(vp, blkno, size, 0, 0); 317 318 /* if not found in cache, do some I/O */ 319 if ((bp->b_flags & B_CACHE) == 0) { 320 if (curproc != NULL) 321 curproc->p_stats->p_ru.ru_inblock++; 322 bp->b_flags |= B_READ; 323 bp->b_flags &= ~(B_DONE | B_ERROR | B_INVAL); 324 if (bp->b_rcred == NOCRED) { 325 if (cred != NOCRED) 326 crhold(cred); 327 bp->b_rcred = cred; 328 } 329 vfs_busy_pages(bp, 0); 330 VOP_STRATEGY(vp, bp); 331 ++readwait; 332 } 333 for (i = 0; i < cnt; i++, rablkno++, rabsize++) { 334 if (inmem(vp, *rablkno)) 335 continue; 336 rabp = getblk(vp, *rablkno, *rabsize, 0, 0); 337 338 if ((rabp->b_flags & B_CACHE) == 0) { 339 if (curproc != NULL) 340 curproc->p_stats->p_ru.ru_inblock++; 341 rabp->b_flags |= B_READ | B_ASYNC; 342 rabp->b_flags &= ~(B_DONE | B_ERROR | B_INVAL); 343 if (rabp->b_rcred == NOCRED) { 344 if (cred != NOCRED) 345 crhold(cred); 346 rabp->b_rcred = cred; 347 } 348 vfs_busy_pages(rabp, 0); 349 VOP_STRATEGY(vp, rabp); 350 } else { 351 brelse(rabp); 352 } 353 } 354 355 if (readwait) { 356 rv = biowait(bp); 357 } 358 return (rv); 359 } 360 361 /* 362 * Write, release buffer on completion. (Done by iodone 363 * if async.) 364 */ 365 int 366 bwrite(struct buf * bp) 367 { 368 int oldflags, s; 369 struct vnode *vp; 370 struct mount *mp; 371 372 373 if (bp->b_flags & B_INVAL) { 374 brelse(bp); 375 return (0); 376 } 377 378 oldflags = bp->b_flags; 379 380 #if !defined(MAX_PERF) 381 if ((bp->b_flags & B_BUSY) == 0) 382 panic("bwrite: buffer is not busy???"); 383 #endif 384 385 bp->b_flags &= ~(B_READ | B_DONE | B_ERROR | B_DELWRI); 386 bp->b_flags |= B_WRITEINPROG; 387 388 s = splbio(); 389 if ((oldflags & B_DELWRI) == B_DELWRI) { 390 --numdirtybuffers; 391 reassignbuf(bp, bp->b_vp); 392 } 393 394 bp->b_vp->v_numoutput++; 395 vfs_busy_pages(bp, 1); 396 if (curproc != NULL) 397 curproc->p_stats->p_ru.ru_oublock++; 398 splx(s); 399 VOP_STRATEGY(bp->b_vp, bp); 400 401 /* 402 * Collect statistics on synchronous and asynchronous writes. 403 * Writes to block devices are charged to their associated 404 * filesystem (if any). 405 */ 406 if ((vp = bp->b_vp) != NULL) { 407 if (vp->v_type == VBLK) 408 mp = vp->v_specmountpoint; 409 else 410 mp = vp->v_mount; 411 if (mp != NULL) 412 if ((oldflags & B_ASYNC) == 0) 413 mp->mnt_stat.f_syncwrites++; 414 else 415 mp->mnt_stat.f_asyncwrites++; 416 } 417 418 if ((oldflags & B_ASYNC) == 0) { 419 int rtval = biowait(bp); 420 brelse(bp); 421 return (rtval); 422 } 423 return (0); 424 } 425 426 void 427 vfs_bio_need_satisfy(void) { 428 ++numfreebuffers; 429 if (!needsbuffer) 430 return; 431 if (numdirtybuffers < lodirtybuffers) { 432 needsbuffer &= ~(VFS_BIO_NEED_ANY | VFS_BIO_NEED_LOWLIMIT); 433 } else { 434 needsbuffer &= ~VFS_BIO_NEED_ANY; 435 } 436 if (numfreebuffers >= hifreebuffers) { 437 needsbuffer &= ~VFS_BIO_NEED_FREE; 438 } 439 wakeup(&needsbuffer); 440 } 441 442 /* 443 * Delayed write. (Buffer is marked dirty). 444 */ 445 void 446 bdwrite(struct buf * bp) 447 { 448 struct vnode *vp; 449 450 #if !defined(MAX_PERF) 451 if ((bp->b_flags & B_BUSY) == 0) { 452 panic("bdwrite: buffer is not busy"); 453 } 454 #endif 455 456 if (bp->b_flags & B_INVAL) { 457 brelse(bp); 458 return; 459 } 460 bp->b_flags &= ~(B_READ|B_RELBUF); 461 if ((bp->b_flags & B_DELWRI) == 0) { 462 bp->b_flags |= B_DONE | B_DELWRI; 463 reassignbuf(bp, bp->b_vp); 464 ++numdirtybuffers; 465 } 466 467 /* 468 * This bmap keeps the system from needing to do the bmap later, 469 * perhaps when the system is attempting to do a sync. Since it 470 * is likely that the indirect block -- or whatever other datastructure 471 * that the filesystem needs is still in memory now, it is a good 472 * thing to do this. Note also, that if the pageout daemon is 473 * requesting a sync -- there might not be enough memory to do 474 * the bmap then... So, this is important to do. 475 */ 476 if (bp->b_lblkno == bp->b_blkno) { 477 VOP_BMAP(bp->b_vp, bp->b_lblkno, NULL, &bp->b_blkno, NULL, NULL); 478 } 479 480 /* 481 * Set the *dirty* buffer range based upon the VM system dirty pages. 482 */ 483 vfs_setdirty(bp); 484 485 /* 486 * We need to do this here to satisfy the vnode_pager and the 487 * pageout daemon, so that it thinks that the pages have been 488 * "cleaned". Note that since the pages are in a delayed write 489 * buffer -- the VFS layer "will" see that the pages get written 490 * out on the next sync, or perhaps the cluster will be completed. 491 */ 492 vfs_clean_pages(bp); 493 bqrelse(bp); 494 495 /* 496 * XXX The soft dependency code is not prepared to 497 * have I/O done when a bdwrite is requested. For 498 * now we just let the write be delayed if it is 499 * requested by the soft dependency code. 500 */ 501 if ((vp = bp->b_vp) && 502 ((vp->v_type == VBLK && vp->v_specmountpoint && 503 (vp->v_specmountpoint->mnt_flag & MNT_SOFTDEP)) || 504 (vp->v_mount && (vp->v_mount->mnt_flag & MNT_SOFTDEP)))) 505 return; 506 507 if (numdirtybuffers >= hidirtybuffers) 508 flushdirtybuffers(0, 0); 509 510 return; 511 } 512 513 514 /* 515 * Same as first half of bdwrite, mark buffer dirty, but do not release it. 516 * Check how this compares with vfs_setdirty(); XXX [JRE] 517 */ 518 void 519 bdirty(bp) 520 struct buf *bp; 521 { 522 523 bp->b_flags &= ~(B_READ|B_RELBUF); /* XXX ??? check this */ 524 if ((bp->b_flags & B_DELWRI) == 0) { 525 bp->b_flags |= B_DONE | B_DELWRI; /* why done? XXX JRE */ 526 reassignbuf(bp, bp->b_vp); 527 ++numdirtybuffers; 528 } 529 } 530 531 /* 532 * Asynchronous write. 533 * Start output on a buffer, but do not wait for it to complete. 534 * The buffer is released when the output completes. 535 */ 536 void 537 bawrite(struct buf * bp) 538 { 539 bp->b_flags |= B_ASYNC; 540 (void) VOP_BWRITE(bp); 541 } 542 543 /* 544 * Ordered write. 545 * Start output on a buffer, and flag it so that the device will write 546 * it in the order it was queued. The buffer is released when the output 547 * completes. 548 */ 549 int 550 bowrite(struct buf * bp) 551 { 552 bp->b_flags |= B_ORDERED|B_ASYNC; 553 return (VOP_BWRITE(bp)); 554 } 555 556 /* 557 * Release a buffer. 558 */ 559 void 560 brelse(struct buf * bp) 561 { 562 int s; 563 564 if (bp->b_flags & B_CLUSTER) { 565 relpbuf(bp); 566 return; 567 } 568 569 s = splbio(); 570 571 /* anyone need this block? */ 572 if (bp->b_flags & B_WANTED) { 573 bp->b_flags &= ~(B_WANTED | B_AGE); 574 wakeup(bp); 575 } 576 577 if (bp->b_flags & B_LOCKED) 578 bp->b_flags &= ~B_ERROR; 579 580 if ((bp->b_flags & (B_NOCACHE | B_INVAL | B_ERROR | B_FREEBUF)) || 581 (bp->b_bufsize <= 0)) { 582 bp->b_flags |= B_INVAL; 583 if (LIST_FIRST(&bp->b_dep) != NULL && bioops.io_deallocate) 584 (*bioops.io_deallocate)(bp); 585 if (bp->b_flags & B_DELWRI) 586 --numdirtybuffers; 587 bp->b_flags &= ~(B_DELWRI | B_CACHE | B_FREEBUF); 588 if ((bp->b_flags & B_VMIO) == 0) { 589 if (bp->b_bufsize) 590 allocbuf(bp, 0); 591 if (bp->b_vp) 592 brelvp(bp); 593 } 594 } 595 596 /* 597 * We must clear B_RELBUF if B_DELWRI is set. If vfs_vmio_release() 598 * is called with B_DELWRI set, the underlying pages may wind up 599 * getting freed causing a previous write (bdwrite()) to get 'lost' 600 * because pages associated with a B_DELWRI bp are marked clean. 601 * 602 * We still allow the B_INVAL case to call vfs_vmio_release(), even 603 * if B_DELWRI is set. 604 */ 605 606 if (bp->b_flags & B_DELWRI) 607 bp->b_flags &= ~B_RELBUF; 608 609 /* 610 * VMIO buffer rundown. It is not very necessary to keep a VMIO buffer 611 * constituted, so the B_INVAL flag is used to *invalidate* the buffer, 612 * but the VM object is kept around. The B_NOCACHE flag is used to 613 * invalidate the pages in the VM object. 614 * 615 * The b_{validoff,validend,dirtyoff,dirtyend} values are relative 616 * to b_offset and currently have byte granularity, whereas the 617 * valid flags in the vm_pages have only DEV_BSIZE resolution. 618 * The byte resolution fields are used to avoid unnecessary re-reads 619 * of the buffer but the code really needs to be genericized so 620 * other filesystem modules can take advantage of these fields. 621 * 622 * XXX this seems to cause performance problems. 623 */ 624 if ((bp->b_flags & B_VMIO) 625 && !(bp->b_vp->v_tag == VT_NFS && 626 bp->b_vp->v_type != VBLK && 627 (bp->b_flags & B_DELWRI) != 0) 628 #ifdef notdef 629 && (bp->b_vp->v_tag != VT_NFS 630 || bp->b_vp->v_type == VBLK 631 || (bp->b_flags & (B_NOCACHE | B_INVAL | B_ERROR)) 632 || bp->b_validend == 0 633 || (bp->b_validoff == 0 634 && bp->b_validend == bp->b_bufsize)) 635 #endif 636 ) { 637 638 int i, j, resid; 639 vm_page_t m; 640 off_t foff; 641 vm_pindex_t poff; 642 vm_object_t obj; 643 struct vnode *vp; 644 645 vp = bp->b_vp; 646 647 /* 648 * Get the base offset and length of the buffer. Note that 649 * for block sizes that are less then PAGE_SIZE, the b_data 650 * base of the buffer does not represent exactly b_offset and 651 * neither b_offset nor b_size are necessarily page aligned. 652 * Instead, the starting position of b_offset is: 653 * 654 * b_data + (b_offset & PAGE_MASK) 655 * 656 * block sizes less then DEV_BSIZE (usually 512) are not 657 * supported due to the page granularity bits (m->valid, 658 * m->dirty, etc...). 659 * 660 * See man buf(9) for more information 661 */ 662 663 resid = bp->b_bufsize; 664 foff = bp->b_offset; 665 666 for (i = 0; i < bp->b_npages; i++) { 667 m = bp->b_pages[i]; 668 vm_page_flag_clear(m, PG_ZERO); 669 if (m == bogus_page) { 670 671 obj = (vm_object_t) vp->v_object; 672 poff = OFF_TO_IDX(bp->b_offset); 673 674 for (j = i; j < bp->b_npages; j++) { 675 m = bp->b_pages[j]; 676 if (m == bogus_page) { 677 m = vm_page_lookup(obj, poff + j); 678 #if !defined(MAX_PERF) 679 if (!m) { 680 panic("brelse: page missing\n"); 681 } 682 #endif 683 bp->b_pages[j] = m; 684 } 685 } 686 687 if ((bp->b_flags & B_INVAL) == 0) { 688 pmap_qenter(trunc_page((vm_offset_t)bp->b_data), bp->b_pages, bp->b_npages); 689 } 690 } 691 if (bp->b_flags & (B_NOCACHE|B_ERROR)) { 692 int poffset = foff & PAGE_MASK; 693 int presid = resid > (PAGE_SIZE - poffset) ? 694 (PAGE_SIZE - poffset) : resid; 695 KASSERT(presid >= 0, ("brelse: extra page")); 696 vm_page_set_invalid(m, poffset, presid); 697 } 698 resid -= PAGE_SIZE - (foff & PAGE_MASK); 699 foff = (foff + PAGE_SIZE) & ~PAGE_MASK; 700 } 701 702 if (bp->b_flags & (B_INVAL | B_RELBUF)) 703 vfs_vmio_release(bp); 704 705 } else if (bp->b_flags & B_VMIO) { 706 707 if (bp->b_flags & (B_INVAL | B_RELBUF)) 708 vfs_vmio_release(bp); 709 710 } 711 712 #if !defined(MAX_PERF) 713 if (bp->b_qindex != QUEUE_NONE) 714 panic("brelse: free buffer onto another queue???"); 715 #endif 716 717 /* enqueue */ 718 /* buffers with no memory */ 719 if (bp->b_bufsize == 0) { 720 bp->b_flags |= B_INVAL; 721 bp->b_qindex = QUEUE_EMPTY; 722 TAILQ_INSERT_HEAD(&bufqueues[QUEUE_EMPTY], bp, b_freelist); 723 LIST_REMOVE(bp, b_hash); 724 LIST_INSERT_HEAD(&invalhash, bp, b_hash); 725 bp->b_dev = NODEV; 726 kvafreespace += bp->b_kvasize; 727 728 /* buffers with junk contents */ 729 } else if (bp->b_flags & (B_ERROR | B_INVAL | B_NOCACHE | B_RELBUF)) { 730 bp->b_flags |= B_INVAL; 731 bp->b_qindex = QUEUE_AGE; 732 TAILQ_INSERT_HEAD(&bufqueues[QUEUE_AGE], bp, b_freelist); 733 LIST_REMOVE(bp, b_hash); 734 LIST_INSERT_HEAD(&invalhash, bp, b_hash); 735 bp->b_dev = NODEV; 736 737 /* buffers that are locked */ 738 } else if (bp->b_flags & B_LOCKED) { 739 bp->b_qindex = QUEUE_LOCKED; 740 TAILQ_INSERT_TAIL(&bufqueues[QUEUE_LOCKED], bp, b_freelist); 741 742 /* buffers with stale but valid contents */ 743 } else if (bp->b_flags & B_AGE) { 744 bp->b_qindex = QUEUE_AGE; 745 TAILQ_INSERT_TAIL(&bufqueues[QUEUE_AGE], bp, b_freelist); 746 747 /* buffers with valid and quite potentially reuseable contents */ 748 } else { 749 bp->b_qindex = QUEUE_LRU; 750 TAILQ_INSERT_TAIL(&bufqueues[QUEUE_LRU], bp, b_freelist); 751 } 752 753 if ((bp->b_flags & B_INVAL) || 754 (bp->b_flags & (B_LOCKED|B_DELWRI)) == 0) { 755 if (bp->b_flags & B_DELWRI) { 756 --numdirtybuffers; 757 bp->b_flags &= ~B_DELWRI; 758 } 759 vfs_bio_need_satisfy(); 760 } 761 762 /* unlock */ 763 bp->b_flags &= ~(B_ORDERED | B_WANTED | B_BUSY | 764 B_ASYNC | B_NOCACHE | B_AGE | B_RELBUF); 765 splx(s); 766 } 767 768 /* 769 * Release a buffer. 770 */ 771 void 772 bqrelse(struct buf * bp) 773 { 774 int s; 775 776 s = splbio(); 777 778 /* anyone need this block? */ 779 if (bp->b_flags & B_WANTED) { 780 bp->b_flags &= ~(B_WANTED | B_AGE); 781 wakeup(bp); 782 } 783 784 #if !defined(MAX_PERF) 785 if (bp->b_qindex != QUEUE_NONE) 786 panic("bqrelse: free buffer onto another queue???"); 787 #endif 788 789 if (bp->b_flags & B_LOCKED) { 790 bp->b_flags &= ~B_ERROR; 791 bp->b_qindex = QUEUE_LOCKED; 792 TAILQ_INSERT_TAIL(&bufqueues[QUEUE_LOCKED], bp, b_freelist); 793 /* buffers with stale but valid contents */ 794 } else { 795 bp->b_qindex = QUEUE_LRU; 796 TAILQ_INSERT_TAIL(&bufqueues[QUEUE_LRU], bp, b_freelist); 797 } 798 799 if ((bp->b_flags & (B_LOCKED|B_DELWRI)) == 0) { 800 vfs_bio_need_satisfy(); 801 } 802 803 /* unlock */ 804 bp->b_flags &= ~(B_ORDERED | B_WANTED | B_BUSY | 805 B_ASYNC | B_NOCACHE | B_AGE | B_RELBUF); 806 splx(s); 807 } 808 809 static void 810 vfs_vmio_release(bp) 811 struct buf *bp; 812 { 813 int i, s; 814 vm_page_t m; 815 816 s = splvm(); 817 for (i = 0; i < bp->b_npages; i++) { 818 m = bp->b_pages[i]; 819 bp->b_pages[i] = NULL; 820 /* 821 * In order to keep page LRU ordering consistent, put 822 * everything on the inactive queue. 823 */ 824 vm_page_unwire(m, 0); 825 /* 826 * We don't mess with busy pages, it is 827 * the responsibility of the process that 828 * busied the pages to deal with them. 829 */ 830 if ((m->flags & PG_BUSY) || (m->busy != 0)) 831 continue; 832 833 if (m->wire_count == 0) { 834 vm_page_flag_clear(m, PG_ZERO); 835 /* 836 * Might as well free the page if we can and it has 837 * no valid data. 838 */ 839 if ((bp->b_flags & B_ASYNC) == 0 && !m->valid && m->hold_count == 0) { 840 vm_page_busy(m); 841 vm_page_protect(m, VM_PROT_NONE); 842 vm_page_free(m); 843 } 844 } 845 } 846 splx(s); 847 bufspace -= bp->b_bufsize; 848 vmiospace -= bp->b_bufsize; 849 pmap_qremove(trunc_page((vm_offset_t) bp->b_data), bp->b_npages); 850 bp->b_npages = 0; 851 bp->b_bufsize = 0; 852 bp->b_flags &= ~B_VMIO; 853 if (bp->b_vp) 854 brelvp(bp); 855 } 856 857 /* 858 * Check to see if a block is currently memory resident. 859 */ 860 struct buf * 861 gbincore(struct vnode * vp, daddr_t blkno) 862 { 863 struct buf *bp; 864 struct bufhashhdr *bh; 865 866 bh = BUFHASH(vp, blkno); 867 bp = bh->lh_first; 868 869 /* Search hash chain */ 870 while (bp != NULL) { 871 /* hit */ 872 if (bp->b_vp == vp && bp->b_lblkno == blkno && 873 (bp->b_flags & B_INVAL) == 0) { 874 break; 875 } 876 bp = bp->b_hash.le_next; 877 } 878 return (bp); 879 } 880 881 /* 882 * this routine implements clustered async writes for 883 * clearing out B_DELWRI buffers... This is much better 884 * than the old way of writing only one buffer at a time. 885 */ 886 int 887 vfs_bio_awrite(struct buf * bp) 888 { 889 int i; 890 daddr_t lblkno = bp->b_lblkno; 891 struct vnode *vp = bp->b_vp; 892 int s; 893 int ncl; 894 struct buf *bpa; 895 int nwritten; 896 int size; 897 int maxcl; 898 899 s = splbio(); 900 /* 901 * right now we support clustered writing only to regular files 902 */ 903 if ((vp->v_type == VREG) && 904 (vp->v_mount != 0) && /* Only on nodes that have the size info */ 905 (bp->b_flags & (B_CLUSTEROK | B_INVAL)) == B_CLUSTEROK) { 906 907 size = vp->v_mount->mnt_stat.f_iosize; 908 maxcl = MAXPHYS / size; 909 910 for (i = 1; i < maxcl; i++) { 911 if ((bpa = gbincore(vp, lblkno + i)) && 912 ((bpa->b_flags & (B_BUSY | B_DELWRI | B_CLUSTEROK | B_INVAL)) == 913 (B_DELWRI | B_CLUSTEROK)) && 914 (bpa->b_bufsize == size)) { 915 if ((bpa->b_blkno == bpa->b_lblkno) || 916 (bpa->b_blkno != bp->b_blkno + ((i * size) >> DEV_BSHIFT))) 917 break; 918 } else { 919 break; 920 } 921 } 922 ncl = i; 923 /* 924 * this is a possible cluster write 925 */ 926 if (ncl != 1) { 927 nwritten = cluster_wbuild(vp, size, lblkno, ncl); 928 splx(s); 929 return nwritten; 930 } 931 } 932 933 bremfree(bp); 934 bp->b_flags |= B_BUSY | B_ASYNC; 935 936 splx(s); 937 /* 938 * default (old) behavior, writing out only one block 939 */ 940 nwritten = bp->b_bufsize; 941 (void) VOP_BWRITE(bp); 942 return nwritten; 943 } 944 945 946 /* 947 * Find a buffer header which is available for use. 948 */ 949 static struct buf * 950 getnewbuf(struct vnode *vp, daddr_t blkno, 951 int slpflag, int slptimeo, int size, int maxsize) 952 { 953 struct buf *bp, *bp1; 954 int nbyteswritten = 0; 955 vm_offset_t addr; 956 static int writerecursion = 0; 957 958 start: 959 if (bufspace >= maxbufspace) 960 goto trytofreespace; 961 962 /* can we constitute a new buffer? */ 963 if ((bp = TAILQ_FIRST(&bufqueues[QUEUE_EMPTY]))) { 964 #if !defined(MAX_PERF) 965 if (bp->b_qindex != QUEUE_EMPTY) 966 panic("getnewbuf: inconsistent EMPTY queue, qindex=%d", 967 bp->b_qindex); 968 #endif 969 bp->b_flags |= B_BUSY; 970 bremfree(bp); 971 goto fillbuf; 972 } 973 trytofreespace: 974 /* 975 * We keep the file I/O from hogging metadata I/O 976 * This is desirable because file data is cached in the 977 * VM/Buffer cache even if a buffer is freed. 978 */ 979 if ((bp = TAILQ_FIRST(&bufqueues[QUEUE_AGE]))) { 980 #if !defined(MAX_PERF) 981 if (bp->b_qindex != QUEUE_AGE) 982 panic("getnewbuf: inconsistent AGE queue, qindex=%d", 983 bp->b_qindex); 984 #endif 985 } else if ((bp = TAILQ_FIRST(&bufqueues[QUEUE_LRU]))) { 986 #if !defined(MAX_PERF) 987 if (bp->b_qindex != QUEUE_LRU) 988 panic("getnewbuf: inconsistent LRU queue, qindex=%d", 989 bp->b_qindex); 990 #endif 991 } 992 if (!bp) { 993 /* wait for a free buffer of any kind */ 994 needsbuffer |= VFS_BIO_NEED_ANY; 995 do 996 tsleep(&needsbuffer, (PRIBIO + 4) | slpflag, "newbuf", 997 slptimeo); 998 while (needsbuffer & VFS_BIO_NEED_ANY); 999 return (0); 1000 } 1001 1002 KASSERT(!(bp->b_flags & B_BUSY), 1003 ("getnewbuf: busy buffer on free list\n")); 1004 1005 /* 1006 * We are fairly aggressive about freeing VMIO buffers, but since 1007 * the buffering is intact without buffer headers, there is not 1008 * much loss. We gain by maintaining non-VMIOed metadata in buffers. 1009 */ 1010 if ((bp->b_qindex == QUEUE_LRU) && (bp->b_usecount > 0)) { 1011 if ((bp->b_flags & B_VMIO) == 0 || 1012 (vmiospace < maxvmiobufspace)) { 1013 --bp->b_usecount; 1014 TAILQ_REMOVE(&bufqueues[QUEUE_LRU], bp, b_freelist); 1015 if (TAILQ_FIRST(&bufqueues[QUEUE_LRU]) != NULL) { 1016 TAILQ_INSERT_TAIL(&bufqueues[QUEUE_LRU], bp, b_freelist); 1017 goto start; 1018 } 1019 TAILQ_INSERT_TAIL(&bufqueues[QUEUE_LRU], bp, b_freelist); 1020 } 1021 } 1022 1023 1024 /* if we are a delayed write, convert to an async write */ 1025 if ((bp->b_flags & (B_DELWRI | B_INVAL)) == B_DELWRI) { 1026 1027 /* 1028 * If our delayed write is likely to be used soon, then 1029 * recycle back onto the LRU queue. 1030 */ 1031 if (vp && (bp->b_vp == vp) && (bp->b_qindex == QUEUE_LRU) && 1032 (bp->b_lblkno >= blkno) && (maxsize > 0)) { 1033 1034 if (bp->b_usecount > 0) { 1035 if (bp->b_lblkno < blkno + (MAXPHYS / maxsize)) { 1036 1037 TAILQ_REMOVE(&bufqueues[QUEUE_LRU], bp, b_freelist); 1038 1039 if (TAILQ_FIRST(&bufqueues[QUEUE_LRU]) != NULL) { 1040 TAILQ_INSERT_TAIL(&bufqueues[QUEUE_LRU], bp, b_freelist); 1041 bp->b_usecount--; 1042 goto start; 1043 } 1044 TAILQ_INSERT_TAIL(&bufqueues[QUEUE_LRU], bp, b_freelist); 1045 } 1046 } 1047 } 1048 1049 /* 1050 * Certain layered filesystems can recursively re-enter the vfs_bio 1051 * code, due to delayed writes. This helps keep the system from 1052 * deadlocking. 1053 */ 1054 if (writerecursion > 0) { 1055 if (writerecursion > 5) { 1056 bp = TAILQ_FIRST(&bufqueues[QUEUE_AGE]); 1057 while (bp) { 1058 if ((bp->b_flags & B_DELWRI) == 0) 1059 break; 1060 bp = TAILQ_NEXT(bp, b_freelist); 1061 } 1062 if (bp == NULL) { 1063 bp = TAILQ_FIRST(&bufqueues[QUEUE_LRU]); 1064 while (bp) { 1065 if ((bp->b_flags & B_DELWRI) == 0) 1066 break; 1067 bp = TAILQ_NEXT(bp, b_freelist); 1068 } 1069 } 1070 if (bp == NULL) 1071 panic("getnewbuf: cannot get buffer, infinite recursion failure"); 1072 } else { 1073 bremfree(bp); 1074 bp->b_flags |= B_BUSY | B_AGE | B_ASYNC; 1075 nbyteswritten += bp->b_bufsize; 1076 ++writerecursion; 1077 VOP_BWRITE(bp); 1078 --writerecursion; 1079 if (!slpflag && !slptimeo) { 1080 return (0); 1081 } 1082 goto start; 1083 } 1084 } else { 1085 ++writerecursion; 1086 nbyteswritten += vfs_bio_awrite(bp); 1087 --writerecursion; 1088 if (!slpflag && !slptimeo) { 1089 return (0); 1090 } 1091 goto start; 1092 } 1093 } 1094 1095 if (bp->b_flags & B_WANTED) { 1096 bp->b_flags &= ~B_WANTED; 1097 wakeup(bp); 1098 } 1099 bremfree(bp); 1100 bp->b_flags |= B_BUSY; 1101 1102 if (bp->b_flags & B_VMIO) { 1103 bp->b_flags &= ~B_ASYNC; 1104 vfs_vmio_release(bp); 1105 } 1106 1107 if (bp->b_vp) 1108 brelvp(bp); 1109 1110 fillbuf: 1111 1112 /* we are not free, nor do we contain interesting data */ 1113 if (bp->b_rcred != NOCRED) { 1114 crfree(bp->b_rcred); 1115 bp->b_rcred = NOCRED; 1116 } 1117 if (bp->b_wcred != NOCRED) { 1118 crfree(bp->b_wcred); 1119 bp->b_wcred = NOCRED; 1120 } 1121 if (LIST_FIRST(&bp->b_dep) != NULL && 1122 bioops.io_deallocate) 1123 (*bioops.io_deallocate)(bp); 1124 1125 LIST_REMOVE(bp, b_hash); 1126 LIST_INSERT_HEAD(&invalhash, bp, b_hash); 1127 if (bp->b_bufsize) { 1128 allocbuf(bp, 0); 1129 } 1130 bp->b_flags = B_BUSY; 1131 bp->b_dev = NODEV; 1132 bp->b_vp = NULL; 1133 bp->b_blkno = bp->b_lblkno = 0; 1134 bp->b_offset = NOOFFSET; 1135 bp->b_iodone = 0; 1136 bp->b_error = 0; 1137 bp->b_resid = 0; 1138 bp->b_bcount = 0; 1139 bp->b_npages = 0; 1140 bp->b_dirtyoff = bp->b_dirtyend = 0; 1141 bp->b_validoff = bp->b_validend = 0; 1142 bp->b_usecount = 5; 1143 /* Here, not kern_physio.c, is where this should be done*/ 1144 LIST_INIT(&bp->b_dep); 1145 1146 maxsize = (maxsize + PAGE_MASK) & ~PAGE_MASK; 1147 1148 /* 1149 * we assume that buffer_map is not at address 0 1150 */ 1151 addr = 0; 1152 if (maxsize != bp->b_kvasize) { 1153 bfreekva(bp); 1154 1155 findkvaspace: 1156 /* 1157 * See if we have buffer kva space 1158 */ 1159 if (vm_map_findspace(buffer_map, 1160 vm_map_min(buffer_map), maxsize, &addr)) { 1161 if (kvafreespace > 0) { 1162 int totfree = 0, freed; 1163 do { 1164 freed = 0; 1165 for (bp1 = TAILQ_FIRST(&bufqueues[QUEUE_EMPTY]); 1166 bp1 != NULL; bp1 = TAILQ_NEXT(bp1, b_freelist)) { 1167 if (bp1->b_kvasize != 0) { 1168 totfree += bp1->b_kvasize; 1169 freed = bp1->b_kvasize; 1170 bremfree(bp1); 1171 bfreekva(bp1); 1172 brelse(bp1); 1173 break; 1174 } 1175 } 1176 } while (freed); 1177 /* 1178 * if we found free space, then retry with the same buffer. 1179 */ 1180 if (totfree) 1181 goto findkvaspace; 1182 } 1183 bp->b_flags |= B_INVAL; 1184 brelse(bp); 1185 goto trytofreespace; 1186 } 1187 } 1188 1189 /* 1190 * See if we are below are allocated minimum 1191 */ 1192 if (bufspace >= (maxbufspace + nbyteswritten)) { 1193 bp->b_flags |= B_INVAL; 1194 brelse(bp); 1195 goto trytofreespace; 1196 } 1197 1198 /* 1199 * create a map entry for the buffer -- in essence 1200 * reserving the kva space. 1201 */ 1202 if (addr) { 1203 vm_map_insert(buffer_map, NULL, 0, 1204 addr, addr + maxsize, 1205 VM_PROT_ALL, VM_PROT_ALL, MAP_NOFAULT); 1206 1207 bp->b_kvabase = (caddr_t) addr; 1208 bp->b_kvasize = maxsize; 1209 } 1210 bp->b_data = bp->b_kvabase; 1211 1212 return (bp); 1213 } 1214 1215 static void 1216 waitfreebuffers(int slpflag, int slptimeo) { 1217 while (numfreebuffers < hifreebuffers) { 1218 flushdirtybuffers(slpflag, slptimeo); 1219 if (numfreebuffers < hifreebuffers) 1220 break; 1221 needsbuffer |= VFS_BIO_NEED_FREE; 1222 if (tsleep(&needsbuffer, (PRIBIO + 4)|slpflag, "biofre", slptimeo)) 1223 break; 1224 } 1225 } 1226 1227 static void 1228 flushdirtybuffers(int slpflag, int slptimeo) { 1229 int s; 1230 static pid_t flushing = 0; 1231 1232 s = splbio(); 1233 1234 if (flushing) { 1235 if (flushing == curproc->p_pid) { 1236 splx(s); 1237 return; 1238 } 1239 while (flushing) { 1240 if (tsleep(&flushing, (PRIBIO + 4)|slpflag, "biofls", slptimeo)) { 1241 splx(s); 1242 return; 1243 } 1244 } 1245 } 1246 flushing = curproc->p_pid; 1247 1248 while (numdirtybuffers > lodirtybuffers) { 1249 struct buf *bp; 1250 needsbuffer |= VFS_BIO_NEED_LOWLIMIT; 1251 bp = TAILQ_FIRST(&bufqueues[QUEUE_AGE]); 1252 if (bp == NULL) 1253 bp = TAILQ_FIRST(&bufqueues[QUEUE_LRU]); 1254 1255 while (bp && ((bp->b_flags & B_DELWRI) == 0)) { 1256 bp = TAILQ_NEXT(bp, b_freelist); 1257 } 1258 1259 if (bp) { 1260 vfs_bio_awrite(bp); 1261 continue; 1262 } 1263 break; 1264 } 1265 1266 flushing = 0; 1267 wakeup(&flushing); 1268 splx(s); 1269 } 1270 1271 /* 1272 * Check to see if a block is currently memory resident. 1273 */ 1274 struct buf * 1275 incore(struct vnode * vp, daddr_t blkno) 1276 { 1277 struct buf *bp; 1278 1279 int s = splbio(); 1280 bp = gbincore(vp, blkno); 1281 splx(s); 1282 return (bp); 1283 } 1284 1285 /* 1286 * Returns true if no I/O is needed to access the 1287 * associated VM object. This is like incore except 1288 * it also hunts around in the VM system for the data. 1289 */ 1290 1291 int 1292 inmem(struct vnode * vp, daddr_t blkno) 1293 { 1294 vm_object_t obj; 1295 vm_offset_t toff, tinc, size; 1296 vm_page_t m; 1297 vm_ooffset_t off; 1298 1299 if (incore(vp, blkno)) 1300 return 1; 1301 if (vp->v_mount == NULL) 1302 return 0; 1303 if ((vp->v_object == NULL) || (vp->v_flag & VOBJBUF) == 0) 1304 return 0; 1305 1306 obj = vp->v_object; 1307 size = PAGE_SIZE; 1308 if (size > vp->v_mount->mnt_stat.f_iosize) 1309 size = vp->v_mount->mnt_stat.f_iosize; 1310 off = (vm_ooffset_t)blkno * (vm_ooffset_t)vp->v_mount->mnt_stat.f_iosize; 1311 1312 for (toff = 0; toff < vp->v_mount->mnt_stat.f_iosize; toff += tinc) { 1313 m = vm_page_lookup(obj, OFF_TO_IDX(off + toff)); 1314 if (!m) 1315 return 0; 1316 tinc = size; 1317 if (tinc > PAGE_SIZE - ((toff + off) & PAGE_MASK)) 1318 tinc = PAGE_SIZE - ((toff + off) & PAGE_MASK); 1319 if (vm_page_is_valid(m, 1320 (vm_offset_t) ((toff + off) & PAGE_MASK), tinc) == 0) 1321 return 0; 1322 } 1323 return 1; 1324 } 1325 1326 /* 1327 * now we set the dirty range for the buffer -- 1328 * for NFS -- if the file is mapped and pages have 1329 * been written to, let it know. We want the 1330 * entire range of the buffer to be marked dirty if 1331 * any of the pages have been written to for consistancy 1332 * with the b_validoff, b_validend set in the nfs write 1333 * code, and used by the nfs read code. 1334 */ 1335 static void 1336 vfs_setdirty(struct buf *bp) { 1337 int i; 1338 vm_object_t object; 1339 vm_offset_t boffset, offset; 1340 /* 1341 * We qualify the scan for modified pages on whether the 1342 * object has been flushed yet. The OBJ_WRITEABLE flag 1343 * is not cleared simply by protecting pages off. 1344 */ 1345 if ((bp->b_flags & B_VMIO) && 1346 ((object = bp->b_pages[0]->object)->flags & (OBJ_WRITEABLE|OBJ_CLEANING))) { 1347 /* 1348 * test the pages to see if they have been modified directly 1349 * by users through the VM system. 1350 */ 1351 for (i = 0; i < bp->b_npages; i++) { 1352 vm_page_flag_clear(bp->b_pages[i], PG_ZERO); 1353 vm_page_test_dirty(bp->b_pages[i]); 1354 } 1355 1356 /* 1357 * scan forwards for the first page modified 1358 */ 1359 for (i = 0; i < bp->b_npages; i++) { 1360 if (bp->b_pages[i]->dirty) { 1361 break; 1362 } 1363 } 1364 boffset = (i << PAGE_SHIFT) - (bp->b_offset & PAGE_MASK); 1365 if (boffset < bp->b_dirtyoff) { 1366 bp->b_dirtyoff = max(boffset, 0); 1367 } 1368 1369 /* 1370 * scan backwards for the last page modified 1371 */ 1372 for (i = bp->b_npages - 1; i >= 0; --i) { 1373 if (bp->b_pages[i]->dirty) { 1374 break; 1375 } 1376 } 1377 boffset = (i + 1); 1378 #if 0 1379 offset = boffset + bp->b_pages[0]->pindex; 1380 if (offset >= object->size) 1381 boffset = object->size - bp->b_pages[0]->pindex; 1382 #endif 1383 boffset = (boffset << PAGE_SHIFT) - (bp->b_offset & PAGE_MASK); 1384 if (bp->b_dirtyend < boffset) 1385 bp->b_dirtyend = min(boffset, bp->b_bufsize); 1386 } 1387 } 1388 1389 /* 1390 * Get a block given a specified block and offset into a file/device. 1391 */ 1392 struct buf * 1393 getblk(struct vnode * vp, daddr_t blkno, int size, int slpflag, int slptimeo) 1394 { 1395 struct buf *bp; 1396 int i, s; 1397 struct bufhashhdr *bh; 1398 int maxsize; 1399 1400 #if !defined(MAX_PERF) 1401 if (size > MAXBSIZE) 1402 panic("getblk: size(%d) > MAXBSIZE(%d)\n", size, MAXBSIZE); 1403 #endif 1404 1405 s = splbio(); 1406 loop: 1407 if (numfreebuffers < lofreebuffers) { 1408 waitfreebuffers(slpflag, slptimeo); 1409 } 1410 1411 if ((bp = gbincore(vp, blkno))) { 1412 if (bp->b_flags & B_BUSY) { 1413 1414 bp->b_flags |= B_WANTED; 1415 if (bp->b_usecount < BUF_MAXUSE) 1416 ++bp->b_usecount; 1417 1418 if (!tsleep(bp, 1419 (PRIBIO + 4) | slpflag, "getblk", slptimeo)) { 1420 goto loop; 1421 } 1422 1423 splx(s); 1424 return (struct buf *) NULL; 1425 } 1426 bp->b_flags |= B_BUSY | B_CACHE; 1427 bremfree(bp); 1428 1429 /* 1430 * check for size inconsistancies (note that they shouldn't 1431 * happen but do when filesystems don't handle the size changes 1432 * correctly.) We are conservative on metadata and don't just 1433 * extend the buffer but write (if needed) and re-constitute it. 1434 */ 1435 1436 if (bp->b_bcount != size) { 1437 if ((bp->b_flags & B_VMIO) && (size <= bp->b_kvasize)) { 1438 allocbuf(bp, size); 1439 } else { 1440 if (bp->b_flags & B_DELWRI) { 1441 bp->b_flags |= B_NOCACHE; 1442 VOP_BWRITE(bp); 1443 } else { 1444 if ((bp->b_flags & B_VMIO) && 1445 (LIST_FIRST(&bp->b_dep) == NULL)) { 1446 bp->b_flags |= B_RELBUF; 1447 brelse(bp); 1448 } else { 1449 bp->b_flags |= B_NOCACHE; 1450 VOP_BWRITE(bp); 1451 } 1452 } 1453 goto loop; 1454 } 1455 } 1456 1457 KASSERT(bp->b_offset != NOOFFSET, 1458 ("getblk: no buffer offset")); 1459 1460 /* 1461 * Check that the constituted buffer really deserves for the 1462 * B_CACHE bit to be set. B_VMIO type buffers might not 1463 * contain fully valid pages. Normal (old-style) buffers 1464 * should be fully valid. 1465 */ 1466 if (bp->b_flags & B_VMIO) { 1467 int checksize = bp->b_bufsize; 1468 int poffset = bp->b_offset & PAGE_MASK; 1469 int resid; 1470 for (i = 0; i < bp->b_npages; i++) { 1471 resid = (checksize > (PAGE_SIZE - poffset)) ? 1472 (PAGE_SIZE - poffset) : checksize; 1473 if (!vm_page_is_valid(bp->b_pages[i], poffset, resid)) { 1474 bp->b_flags &= ~(B_CACHE | B_DONE); 1475 break; 1476 } 1477 checksize -= resid; 1478 poffset = 0; 1479 } 1480 } 1481 1482 if (bp->b_usecount < BUF_MAXUSE) 1483 ++bp->b_usecount; 1484 splx(s); 1485 return (bp); 1486 } else { 1487 int bsize, maxsize, vmio; 1488 off_t offset; 1489 1490 if (vp->v_type == VBLK) 1491 bsize = DEV_BSIZE; 1492 else if (vp->v_mountedhere) 1493 bsize = vp->v_mountedhere->mnt_stat.f_iosize; 1494 else if (vp->v_mount) 1495 bsize = vp->v_mount->mnt_stat.f_iosize; 1496 else 1497 bsize = size; 1498 1499 offset = (off_t)blkno * bsize; 1500 vmio = (vp->v_object != 0) && (vp->v_flag & VOBJBUF); 1501 maxsize = vmio ? size + (offset & PAGE_MASK) : size; 1502 maxsize = imax(maxsize, bsize); 1503 1504 if ((bp = getnewbuf(vp, blkno, 1505 slpflag, slptimeo, size, maxsize)) == 0) { 1506 if (slpflag || slptimeo) { 1507 splx(s); 1508 return NULL; 1509 } 1510 goto loop; 1511 } 1512 1513 /* 1514 * This code is used to make sure that a buffer is not 1515 * created while the getnewbuf routine is blocked. 1516 * Normally the vnode is locked so this isn't a problem. 1517 * VBLK type I/O requests, however, don't lock the vnode. 1518 */ 1519 if (VOP_ISLOCKED(vp) != LK_EXCLUSIVE && gbincore(vp, blkno)) { 1520 bp->b_flags |= B_INVAL; 1521 brelse(bp); 1522 goto loop; 1523 } 1524 1525 /* 1526 * Insert the buffer into the hash, so that it can 1527 * be found by incore. 1528 */ 1529 bp->b_blkno = bp->b_lblkno = blkno; 1530 bp->b_offset = offset; 1531 1532 bgetvp(vp, bp); 1533 LIST_REMOVE(bp, b_hash); 1534 bh = BUFHASH(vp, blkno); 1535 LIST_INSERT_HEAD(bh, bp, b_hash); 1536 1537 if (vmio) { 1538 bp->b_flags |= (B_VMIO | B_CACHE); 1539 #if defined(VFS_BIO_DEBUG) 1540 if (vp->v_type != VREG && vp->v_type != VBLK) 1541 printf("getblk: vmioing file type %d???\n", vp->v_type); 1542 #endif 1543 } else { 1544 bp->b_flags &= ~B_VMIO; 1545 } 1546 1547 allocbuf(bp, size); 1548 1549 splx(s); 1550 return (bp); 1551 } 1552 } 1553 1554 /* 1555 * Get an empty, disassociated buffer of given size. 1556 */ 1557 struct buf * 1558 geteblk(int size) 1559 { 1560 struct buf *bp; 1561 int s; 1562 1563 s = splbio(); 1564 while ((bp = getnewbuf(0, (daddr_t) 0, 0, 0, size, MAXBSIZE)) == 0); 1565 splx(s); 1566 allocbuf(bp, size); 1567 bp->b_flags |= B_INVAL; /* b_dep cleared by getnewbuf() */ 1568 return (bp); 1569 } 1570 1571 1572 /* 1573 * This code constitutes the buffer memory from either anonymous system 1574 * memory (in the case of non-VMIO operations) or from an associated 1575 * VM object (in the case of VMIO operations). 1576 * 1577 * Note that this code is tricky, and has many complications to resolve 1578 * deadlock or inconsistant data situations. Tread lightly!!! 1579 * 1580 * Modify the length of a buffer's underlying buffer storage without 1581 * destroying information (unless, of course the buffer is shrinking). 1582 */ 1583 int 1584 allocbuf(struct buf * bp, int size) 1585 { 1586 1587 int s; 1588 int newbsize, mbsize; 1589 int i; 1590 1591 #if !defined(MAX_PERF) 1592 if (!(bp->b_flags & B_BUSY)) 1593 panic("allocbuf: buffer not busy"); 1594 1595 if (bp->b_kvasize < size) 1596 panic("allocbuf: buffer too small"); 1597 #endif 1598 1599 if ((bp->b_flags & B_VMIO) == 0) { 1600 caddr_t origbuf; 1601 int origbufsize; 1602 /* 1603 * Just get anonymous memory from the kernel 1604 */ 1605 mbsize = (size + DEV_BSIZE - 1) & ~(DEV_BSIZE - 1); 1606 #if !defined(NO_B_MALLOC) 1607 if (bp->b_flags & B_MALLOC) 1608 newbsize = mbsize; 1609 else 1610 #endif 1611 newbsize = round_page(size); 1612 1613 if (newbsize < bp->b_bufsize) { 1614 #if !defined(NO_B_MALLOC) 1615 /* 1616 * malloced buffers are not shrunk 1617 */ 1618 if (bp->b_flags & B_MALLOC) { 1619 if (newbsize) { 1620 bp->b_bcount = size; 1621 } else { 1622 free(bp->b_data, M_BIOBUF); 1623 bufspace -= bp->b_bufsize; 1624 bufmallocspace -= bp->b_bufsize; 1625 bp->b_data = bp->b_kvabase; 1626 bp->b_bufsize = 0; 1627 bp->b_bcount = 0; 1628 bp->b_flags &= ~B_MALLOC; 1629 } 1630 return 1; 1631 } 1632 #endif 1633 vm_hold_free_pages( 1634 bp, 1635 (vm_offset_t) bp->b_data + newbsize, 1636 (vm_offset_t) bp->b_data + bp->b_bufsize); 1637 } else if (newbsize > bp->b_bufsize) { 1638 #if !defined(NO_B_MALLOC) 1639 /* 1640 * We only use malloced memory on the first allocation. 1641 * and revert to page-allocated memory when the buffer grows. 1642 */ 1643 if ( (bufmallocspace < maxbufmallocspace) && 1644 (bp->b_bufsize == 0) && 1645 (mbsize <= PAGE_SIZE/2)) { 1646 1647 bp->b_data = malloc(mbsize, M_BIOBUF, M_WAITOK); 1648 bp->b_bufsize = mbsize; 1649 bp->b_bcount = size; 1650 bp->b_flags |= B_MALLOC; 1651 bufspace += mbsize; 1652 bufmallocspace += mbsize; 1653 return 1; 1654 } 1655 #endif 1656 origbuf = NULL; 1657 origbufsize = 0; 1658 #if !defined(NO_B_MALLOC) 1659 /* 1660 * If the buffer is growing on its other-than-first allocation, 1661 * then we revert to the page-allocation scheme. 1662 */ 1663 if (bp->b_flags & B_MALLOC) { 1664 origbuf = bp->b_data; 1665 origbufsize = bp->b_bufsize; 1666 bp->b_data = bp->b_kvabase; 1667 bufspace -= bp->b_bufsize; 1668 bufmallocspace -= bp->b_bufsize; 1669 bp->b_bufsize = 0; 1670 bp->b_flags &= ~B_MALLOC; 1671 newbsize = round_page(newbsize); 1672 } 1673 #endif 1674 vm_hold_load_pages( 1675 bp, 1676 (vm_offset_t) bp->b_data + bp->b_bufsize, 1677 (vm_offset_t) bp->b_data + newbsize); 1678 #if !defined(NO_B_MALLOC) 1679 if (origbuf) { 1680 bcopy(origbuf, bp->b_data, origbufsize); 1681 free(origbuf, M_BIOBUF); 1682 } 1683 #endif 1684 } 1685 } else { 1686 vm_page_t m; 1687 int desiredpages; 1688 1689 newbsize = (size + DEV_BSIZE - 1) & ~(DEV_BSIZE - 1); 1690 desiredpages = (size == 0) ? 0 : 1691 num_pages((bp->b_offset & PAGE_MASK) + newbsize); 1692 1693 #if !defined(NO_B_MALLOC) 1694 if (bp->b_flags & B_MALLOC) 1695 panic("allocbuf: VMIO buffer can't be malloced"); 1696 #endif 1697 1698 if (newbsize < bp->b_bufsize) { 1699 if (desiredpages < bp->b_npages) { 1700 for (i = desiredpages; i < bp->b_npages; i++) { 1701 /* 1702 * the page is not freed here -- it 1703 * is the responsibility of vnode_pager_setsize 1704 */ 1705 m = bp->b_pages[i]; 1706 KASSERT(m != bogus_page, 1707 ("allocbuf: bogus page found")); 1708 vm_page_sleep(m, "biodep", &m->busy); 1709 1710 bp->b_pages[i] = NULL; 1711 vm_page_unwire(m, 0); 1712 } 1713 pmap_qremove((vm_offset_t) trunc_page((vm_offset_t)bp->b_data) + 1714 (desiredpages << PAGE_SHIFT), (bp->b_npages - desiredpages)); 1715 bp->b_npages = desiredpages; 1716 } 1717 } else if (newbsize > bp->b_bufsize) { 1718 vm_object_t obj; 1719 vm_offset_t tinc, toff; 1720 vm_ooffset_t off; 1721 vm_pindex_t objoff; 1722 int pageindex, curbpnpages; 1723 struct vnode *vp; 1724 int bsize; 1725 int orig_validoff = bp->b_validoff; 1726 int orig_validend = bp->b_validend; 1727 1728 vp = bp->b_vp; 1729 1730 if (vp->v_type == VBLK) 1731 bsize = DEV_BSIZE; 1732 else 1733 bsize = vp->v_mount->mnt_stat.f_iosize; 1734 1735 if (bp->b_npages < desiredpages) { 1736 obj = vp->v_object; 1737 tinc = PAGE_SIZE; 1738 1739 off = bp->b_offset; 1740 KASSERT(bp->b_offset != NOOFFSET, 1741 ("allocbuf: no buffer offset")); 1742 1743 curbpnpages = bp->b_npages; 1744 doretry: 1745 bp->b_validoff = orig_validoff; 1746 bp->b_validend = orig_validend; 1747 bp->b_flags |= B_CACHE; 1748 for (toff = 0; toff < newbsize; toff += tinc) { 1749 objoff = OFF_TO_IDX(off + toff); 1750 pageindex = objoff - OFF_TO_IDX(off); 1751 tinc = PAGE_SIZE - ((off + toff) & PAGE_MASK); 1752 if (pageindex < curbpnpages) { 1753 1754 m = bp->b_pages[pageindex]; 1755 #ifdef VFS_BIO_DIAG 1756 if (m->pindex != objoff) 1757 panic("allocbuf: page changed offset?!!!?"); 1758 #endif 1759 if (tinc > (newbsize - toff)) 1760 tinc = newbsize - toff; 1761 if (bp->b_flags & B_CACHE) 1762 vfs_buf_set_valid(bp, off, toff, tinc, m); 1763 continue; 1764 } 1765 m = vm_page_lookup(obj, objoff); 1766 if (!m) { 1767 m = vm_page_alloc(obj, objoff, VM_ALLOC_NORMAL); 1768 if (!m) { 1769 VM_WAIT; 1770 vm_pageout_deficit += (desiredpages - curbpnpages); 1771 goto doretry; 1772 } 1773 1774 vm_page_wire(m); 1775 vm_page_flag_clear(m, PG_BUSY); 1776 bp->b_flags &= ~B_CACHE; 1777 1778 } else if (m->flags & PG_BUSY) { 1779 s = splvm(); 1780 if (m->flags & PG_BUSY) { 1781 vm_page_flag_set(m, PG_WANTED); 1782 tsleep(m, PVM, "pgtblk", 0); 1783 } 1784 splx(s); 1785 goto doretry; 1786 } else { 1787 if ((curproc != pageproc) && 1788 ((m->queue - m->pc) == PQ_CACHE) && 1789 ((cnt.v_free_count + cnt.v_cache_count) < 1790 (cnt.v_free_min + cnt.v_cache_min))) { 1791 pagedaemon_wakeup(); 1792 } 1793 if (tinc > (newbsize - toff)) 1794 tinc = newbsize - toff; 1795 if (bp->b_flags & B_CACHE) 1796 vfs_buf_set_valid(bp, off, toff, tinc, m); 1797 vm_page_flag_clear(m, PG_ZERO); 1798 vm_page_wire(m); 1799 } 1800 bp->b_pages[pageindex] = m; 1801 curbpnpages = pageindex + 1; 1802 } 1803 if (vp->v_tag == VT_NFS && 1804 vp->v_type != VBLK) { 1805 if (bp->b_dirtyend > 0) { 1806 bp->b_validoff = min(bp->b_validoff, bp->b_dirtyoff); 1807 bp->b_validend = max(bp->b_validend, bp->b_dirtyend); 1808 } 1809 if (bp->b_validend == 0) 1810 bp->b_flags &= ~B_CACHE; 1811 } 1812 bp->b_data = (caddr_t) trunc_page((vm_offset_t)bp->b_data); 1813 bp->b_npages = curbpnpages; 1814 pmap_qenter((vm_offset_t) bp->b_data, 1815 bp->b_pages, bp->b_npages); 1816 ((vm_offset_t) bp->b_data) |= off & PAGE_MASK; 1817 } 1818 } 1819 } 1820 if (bp->b_flags & B_VMIO) 1821 vmiospace += (newbsize - bp->b_bufsize); 1822 bufspace += (newbsize - bp->b_bufsize); 1823 bp->b_bufsize = newbsize; 1824 bp->b_bcount = size; 1825 return 1; 1826 } 1827 1828 /* 1829 * Wait for buffer I/O completion, returning error status. 1830 */ 1831 int 1832 biowait(register struct buf * bp) 1833 { 1834 int s; 1835 1836 s = splbio(); 1837 while ((bp->b_flags & B_DONE) == 0) 1838 #if defined(NO_SCHEDULE_MODS) 1839 tsleep(bp, PRIBIO, "biowait", 0); 1840 #else 1841 if (bp->b_flags & B_READ) 1842 tsleep(bp, PRIBIO, "biord", 0); 1843 else 1844 tsleep(bp, PRIBIO, "biowr", 0); 1845 #endif 1846 splx(s); 1847 if (bp->b_flags & B_EINTR) { 1848 bp->b_flags &= ~B_EINTR; 1849 return (EINTR); 1850 } 1851 if (bp->b_flags & B_ERROR) { 1852 return (bp->b_error ? bp->b_error : EIO); 1853 } else { 1854 return (0); 1855 } 1856 } 1857 1858 /* 1859 * Finish I/O on a buffer, calling an optional function. 1860 * This is usually called from interrupt level, so process blocking 1861 * is not *a good idea*. 1862 */ 1863 void 1864 biodone(register struct buf * bp) 1865 { 1866 int s; 1867 1868 s = splbio(); 1869 1870 #if !defined(MAX_PERF) 1871 if (!(bp->b_flags & B_BUSY)) 1872 panic("biodone: buffer not busy"); 1873 #endif 1874 1875 if (bp->b_flags & B_DONE) { 1876 splx(s); 1877 #if !defined(MAX_PERF) 1878 printf("biodone: buffer already done\n"); 1879 #endif 1880 return; 1881 } 1882 bp->b_flags |= B_DONE; 1883 1884 if (bp->b_flags & B_FREEBUF) { 1885 brelse(bp); 1886 splx(s); 1887 return; 1888 } 1889 1890 if ((bp->b_flags & B_READ) == 0) { 1891 vwakeup(bp); 1892 } 1893 1894 /* call optional completion function if requested */ 1895 if (bp->b_flags & B_CALL) { 1896 bp->b_flags &= ~B_CALL; 1897 (*bp->b_iodone) (bp); 1898 splx(s); 1899 return; 1900 } 1901 if (LIST_FIRST(&bp->b_dep) != NULL && bioops.io_complete) 1902 (*bioops.io_complete)(bp); 1903 1904 if (bp->b_flags & B_VMIO) { 1905 int i, resid; 1906 vm_ooffset_t foff; 1907 vm_page_t m; 1908 vm_object_t obj; 1909 int iosize; 1910 struct vnode *vp = bp->b_vp; 1911 1912 obj = vp->v_object; 1913 1914 #if defined(VFS_BIO_DEBUG) 1915 if (vp->v_usecount == 0) { 1916 panic("biodone: zero vnode ref count"); 1917 } 1918 1919 if (vp->v_object == NULL) { 1920 panic("biodone: missing VM object"); 1921 } 1922 1923 if ((vp->v_flag & VOBJBUF) == 0) { 1924 panic("biodone: vnode is not setup for merged cache"); 1925 } 1926 #endif 1927 1928 foff = bp->b_offset; 1929 KASSERT(bp->b_offset != NOOFFSET, 1930 ("biodone: no buffer offset")); 1931 1932 #if !defined(MAX_PERF) 1933 if (!obj) { 1934 panic("biodone: no object"); 1935 } 1936 #endif 1937 #if defined(VFS_BIO_DEBUG) 1938 if (obj->paging_in_progress < bp->b_npages) { 1939 printf("biodone: paging in progress(%d) < bp->b_npages(%d)\n", 1940 obj->paging_in_progress, bp->b_npages); 1941 } 1942 #endif 1943 iosize = bp->b_bufsize; 1944 for (i = 0; i < bp->b_npages; i++) { 1945 int bogusflag = 0; 1946 m = bp->b_pages[i]; 1947 if (m == bogus_page) { 1948 bogusflag = 1; 1949 m = vm_page_lookup(obj, OFF_TO_IDX(foff)); 1950 if (!m) { 1951 #if defined(VFS_BIO_DEBUG) 1952 printf("biodone: page disappeared\n"); 1953 #endif 1954 vm_object_pip_subtract(obj, 1); 1955 continue; 1956 } 1957 bp->b_pages[i] = m; 1958 pmap_qenter(trunc_page((vm_offset_t)bp->b_data), bp->b_pages, bp->b_npages); 1959 } 1960 #if defined(VFS_BIO_DEBUG) 1961 if (OFF_TO_IDX(foff) != m->pindex) { 1962 printf("biodone: foff(%d)/m->pindex(%d) mismatch\n", foff, m->pindex); 1963 } 1964 #endif 1965 resid = IDX_TO_OFF(m->pindex + 1) - foff; 1966 if (resid > iosize) 1967 resid = iosize; 1968 1969 /* 1970 * In the write case, the valid and clean bits are 1971 * already changed correctly, so we only need to do this 1972 * here in the read case. 1973 */ 1974 if ((bp->b_flags & B_READ) && !bogusflag && resid > 0) { 1975 vfs_page_set_valid(bp, foff, i, m); 1976 } 1977 vm_page_flag_clear(m, PG_ZERO); 1978 1979 /* 1980 * when debugging new filesystems or buffer I/O methods, this 1981 * is the most common error that pops up. if you see this, you 1982 * have not set the page busy flag correctly!!! 1983 */ 1984 if (m->busy == 0) { 1985 #if !defined(MAX_PERF) 1986 printf("biodone: page busy < 0, " 1987 "pindex: %d, foff: 0x(%x,%x), " 1988 "resid: %d, index: %d\n", 1989 (int) m->pindex, (int)(foff >> 32), 1990 (int) foff & 0xffffffff, resid, i); 1991 #endif 1992 if (vp->v_type != VBLK) 1993 #if !defined(MAX_PERF) 1994 printf(" iosize: %ld, lblkno: %d, flags: 0x%lx, npages: %d\n", 1995 bp->b_vp->v_mount->mnt_stat.f_iosize, 1996 (int) bp->b_lblkno, 1997 bp->b_flags, bp->b_npages); 1998 else 1999 printf(" VDEV, lblkno: %d, flags: 0x%lx, npages: %d\n", 2000 (int) bp->b_lblkno, 2001 bp->b_flags, bp->b_npages); 2002 printf(" valid: 0x%x, dirty: 0x%x, wired: %d\n", 2003 m->valid, m->dirty, m->wire_count); 2004 #endif 2005 panic("biodone: page busy < 0\n"); 2006 } 2007 vm_page_io_finish(m); 2008 vm_object_pip_subtract(obj, 1); 2009 foff += resid; 2010 iosize -= resid; 2011 } 2012 if (obj && 2013 (obj->paging_in_progress == 0) && 2014 (obj->flags & OBJ_PIPWNT)) { 2015 vm_object_clear_flag(obj, OBJ_PIPWNT); 2016 wakeup(obj); 2017 } 2018 } 2019 /* 2020 * For asynchronous completions, release the buffer now. The brelse 2021 * checks for B_WANTED and will do the wakeup there if necessary - so 2022 * no need to do a wakeup here in the async case. 2023 */ 2024 2025 if (bp->b_flags & B_ASYNC) { 2026 if ((bp->b_flags & (B_NOCACHE | B_INVAL | B_ERROR | B_RELBUF)) != 0) 2027 brelse(bp); 2028 else 2029 bqrelse(bp); 2030 } else { 2031 bp->b_flags &= ~B_WANTED; 2032 wakeup(bp); 2033 } 2034 splx(s); 2035 } 2036 2037 #if 0 /* not with kirks code */ 2038 static int vfs_update_interval = 30; 2039 2040 static void 2041 vfs_update() 2042 { 2043 while (1) { 2044 tsleep(&vfs_update_wakeup, PUSER, "update", 2045 hz * vfs_update_interval); 2046 vfs_update_wakeup = 0; 2047 sync(curproc, NULL); 2048 } 2049 } 2050 2051 static int 2052 sysctl_kern_updateinterval SYSCTL_HANDLER_ARGS 2053 { 2054 int error = sysctl_handle_int(oidp, 2055 oidp->oid_arg1, oidp->oid_arg2, req); 2056 if (!error) 2057 wakeup(&vfs_update_wakeup); 2058 return error; 2059 } 2060 2061 SYSCTL_PROC(_kern, KERN_UPDATEINTERVAL, update, CTLTYPE_INT|CTLFLAG_RW, 2062 &vfs_update_interval, 0, sysctl_kern_updateinterval, "I", ""); 2063 2064 #endif 2065 2066 2067 /* 2068 * This routine is called in lieu of iodone in the case of 2069 * incomplete I/O. This keeps the busy status for pages 2070 * consistant. 2071 */ 2072 void 2073 vfs_unbusy_pages(struct buf * bp) 2074 { 2075 int i; 2076 2077 if (bp->b_flags & B_VMIO) { 2078 struct vnode *vp = bp->b_vp; 2079 vm_object_t obj = vp->v_object; 2080 2081 for (i = 0; i < bp->b_npages; i++) { 2082 vm_page_t m = bp->b_pages[i]; 2083 2084 if (m == bogus_page) { 2085 m = vm_page_lookup(obj, OFF_TO_IDX(bp->b_offset) + i); 2086 #if !defined(MAX_PERF) 2087 if (!m) { 2088 panic("vfs_unbusy_pages: page missing\n"); 2089 } 2090 #endif 2091 bp->b_pages[i] = m; 2092 pmap_qenter(trunc_page((vm_offset_t)bp->b_data), bp->b_pages, bp->b_npages); 2093 } 2094 vm_object_pip_subtract(obj, 1); 2095 vm_page_flag_clear(m, PG_ZERO); 2096 vm_page_io_finish(m); 2097 } 2098 if (obj->paging_in_progress == 0 && 2099 (obj->flags & OBJ_PIPWNT)) { 2100 vm_object_clear_flag(obj, OBJ_PIPWNT); 2101 wakeup(obj); 2102 } 2103 } 2104 } 2105 2106 /* 2107 * Set NFS' b_validoff and b_validend fields from the valid bits 2108 * of a page. If the consumer is not NFS, and the page is not 2109 * valid for the entire range, clear the B_CACHE flag to force 2110 * the consumer to re-read the page. 2111 */ 2112 static void 2113 vfs_buf_set_valid(struct buf *bp, 2114 vm_ooffset_t foff, vm_offset_t off, vm_offset_t size, 2115 vm_page_t m) 2116 { 2117 if (bp->b_vp->v_tag == VT_NFS && bp->b_vp->v_type != VBLK) { 2118 vm_offset_t svalid, evalid; 2119 int validbits = m->valid >> (((foff+off)&PAGE_MASK)/DEV_BSIZE); 2120 2121 /* 2122 * This only bothers with the first valid range in the 2123 * page. 2124 */ 2125 svalid = off; 2126 while (validbits && !(validbits & 1)) { 2127 svalid += DEV_BSIZE; 2128 validbits >>= 1; 2129 } 2130 evalid = svalid; 2131 while (validbits & 1) { 2132 evalid += DEV_BSIZE; 2133 validbits >>= 1; 2134 } 2135 evalid = min(evalid, off + size); 2136 /* 2137 * Make sure this range is contiguous with the range 2138 * built up from previous pages. If not, then we will 2139 * just use the range from the previous pages. 2140 */ 2141 if (svalid == bp->b_validend) { 2142 bp->b_validoff = min(bp->b_validoff, svalid); 2143 bp->b_validend = max(bp->b_validend, evalid); 2144 } 2145 } else if (!vm_page_is_valid(m, 2146 (vm_offset_t) ((foff + off) & PAGE_MASK), 2147 size)) { 2148 bp->b_flags &= ~B_CACHE; 2149 } 2150 } 2151 2152 /* 2153 * Set the valid bits in a page, taking care of the b_validoff, 2154 * b_validend fields which NFS uses to optimise small reads. Off is 2155 * the offset within the file and pageno is the page index within the buf. 2156 */ 2157 static void 2158 vfs_page_set_valid(struct buf *bp, vm_ooffset_t off, int pageno, vm_page_t m) 2159 { 2160 struct vnode *vp = bp->b_vp; 2161 vm_ooffset_t soff, eoff; 2162 2163 soff = off; 2164 eoff = (off + PAGE_SIZE) & ~PAGE_MASK; 2165 if (eoff > bp->b_offset + bp->b_bufsize) 2166 eoff = bp->b_offset + bp->b_bufsize; 2167 if (vp->v_tag == VT_NFS && vp->v_type != VBLK) { 2168 vm_ooffset_t sv, ev; 2169 vm_page_set_invalid(m, 2170 (vm_offset_t) (soff & PAGE_MASK), 2171 (vm_offset_t) (eoff - soff)); 2172 sv = (bp->b_offset + bp->b_validoff + DEV_BSIZE - 1) & ~(DEV_BSIZE - 1); 2173 ev = (bp->b_offset + bp->b_validend) & ~(DEV_BSIZE - 1); 2174 soff = qmax(sv, soff); 2175 eoff = qmin(ev, eoff); 2176 } 2177 if (eoff > soff) 2178 vm_page_set_validclean(m, 2179 (vm_offset_t) (soff & PAGE_MASK), 2180 (vm_offset_t) (eoff - soff)); 2181 } 2182 2183 /* 2184 * This routine is called before a device strategy routine. 2185 * It is used to tell the VM system that paging I/O is in 2186 * progress, and treat the pages associated with the buffer 2187 * almost as being PG_BUSY. Also the object paging_in_progress 2188 * flag is handled to make sure that the object doesn't become 2189 * inconsistant. 2190 */ 2191 void 2192 vfs_busy_pages(struct buf * bp, int clear_modify) 2193 { 2194 int i, bogus; 2195 2196 if (bp->b_flags & B_VMIO) { 2197 struct vnode *vp = bp->b_vp; 2198 vm_object_t obj = vp->v_object; 2199 vm_ooffset_t foff; 2200 2201 foff = bp->b_offset; 2202 KASSERT(bp->b_offset != NOOFFSET, 2203 ("vfs_busy_pages: no buffer offset")); 2204 2205 vfs_setdirty(bp); 2206 2207 retry: 2208 for (i = 0; i < bp->b_npages; i++) { 2209 vm_page_t m = bp->b_pages[i]; 2210 if (vm_page_sleep(m, "vbpage", NULL)) 2211 goto retry; 2212 } 2213 2214 bogus = 0; 2215 for (i = 0; i < bp->b_npages; i++) { 2216 vm_page_t m = bp->b_pages[i]; 2217 2218 vm_page_flag_clear(m, PG_ZERO); 2219 if ((bp->b_flags & B_CLUSTER) == 0) { 2220 vm_object_pip_add(obj, 1); 2221 vm_page_io_start(m); 2222 } 2223 2224 vm_page_protect(m, VM_PROT_NONE); 2225 if (clear_modify) 2226 vfs_page_set_valid(bp, foff, i, m); 2227 else if (m->valid == VM_PAGE_BITS_ALL && 2228 (bp->b_flags & B_CACHE) == 0) { 2229 bp->b_pages[i] = bogus_page; 2230 bogus++; 2231 } 2232 foff = (foff + PAGE_SIZE) & ~PAGE_MASK; 2233 } 2234 if (bogus) 2235 pmap_qenter(trunc_page((vm_offset_t)bp->b_data), bp->b_pages, bp->b_npages); 2236 } 2237 } 2238 2239 /* 2240 * Tell the VM system that the pages associated with this buffer 2241 * are clean. This is used for delayed writes where the data is 2242 * going to go to disk eventually without additional VM intevention. 2243 */ 2244 void 2245 vfs_clean_pages(struct buf * bp) 2246 { 2247 int i; 2248 2249 if (bp->b_flags & B_VMIO) { 2250 vm_ooffset_t foff; 2251 foff = bp->b_offset; 2252 2253 KASSERT(bp->b_offset != NOOFFSET, 2254 ("vfs_clean_pages: no buffer offset")); 2255 2256 for (i = 0; i < bp->b_npages; i++) { 2257 vm_page_t m = bp->b_pages[i]; 2258 vfs_page_set_valid(bp, foff, i, m); 2259 foff = (foff + PAGE_SIZE) & ~PAGE_MASK; 2260 } 2261 } 2262 } 2263 2264 void 2265 vfs_bio_clrbuf(struct buf *bp) { 2266 int i, size, mask = 0; 2267 caddr_t sa, ea; 2268 if ((bp->b_flags & (B_VMIO | B_MALLOC)) == B_VMIO) { 2269 if( (bp->b_npages == 1) && (bp->b_bufsize < PAGE_SIZE) && 2270 (bp->b_offset & PAGE_MASK) == 0) { 2271 mask = (1 << (bp->b_bufsize / DEV_BSIZE)) - 1; 2272 if (((bp->b_pages[0]->flags & PG_ZERO) == 0) && 2273 ((bp->b_pages[0]->valid & mask) != mask)) { 2274 bzero(bp->b_data, bp->b_bufsize); 2275 } 2276 bp->b_pages[0]->valid |= mask; 2277 bp->b_resid = 0; 2278 return; 2279 } 2280 ea = sa = bp->b_data; 2281 for(i=0;i<bp->b_npages;i++,sa=ea) { 2282 int j = ((u_long)sa & PAGE_MASK) / DEV_BSIZE; 2283 ea = (caddr_t)trunc_page((vm_offset_t)sa + PAGE_SIZE); 2284 ea = (caddr_t)ulmin((u_long)ea, 2285 (u_long)bp->b_data + bp->b_bufsize); 2286 mask = ((1 << ((ea - sa) / DEV_BSIZE)) - 1) << j; 2287 if ((bp->b_pages[i]->valid & mask) == mask) 2288 continue; 2289 if ((bp->b_pages[i]->valid & mask) == 0) { 2290 if ((bp->b_pages[i]->flags & PG_ZERO) == 0) { 2291 bzero(sa, ea - sa); 2292 } 2293 } else { 2294 for (; sa < ea; sa += DEV_BSIZE, j++) { 2295 if (((bp->b_pages[i]->flags & PG_ZERO) == 0) && 2296 (bp->b_pages[i]->valid & (1<<j)) == 0) 2297 bzero(sa, DEV_BSIZE); 2298 } 2299 } 2300 bp->b_pages[i]->valid |= mask; 2301 vm_page_flag_clear(bp->b_pages[i], PG_ZERO); 2302 } 2303 bp->b_resid = 0; 2304 } else { 2305 clrbuf(bp); 2306 } 2307 } 2308 2309 /* 2310 * vm_hold_load_pages and vm_hold_unload pages get pages into 2311 * a buffers address space. The pages are anonymous and are 2312 * not associated with a file object. 2313 */ 2314 void 2315 vm_hold_load_pages(struct buf * bp, vm_offset_t from, vm_offset_t to) 2316 { 2317 vm_offset_t pg; 2318 vm_page_t p; 2319 int index; 2320 2321 to = round_page(to); 2322 from = round_page(from); 2323 index = (from - trunc_page((vm_offset_t)bp->b_data)) >> PAGE_SHIFT; 2324 2325 for (pg = from; pg < to; pg += PAGE_SIZE, index++) { 2326 2327 tryagain: 2328 2329 p = vm_page_alloc(kernel_object, 2330 ((pg - VM_MIN_KERNEL_ADDRESS) >> PAGE_SHIFT), 2331 VM_ALLOC_NORMAL); 2332 if (!p) { 2333 vm_pageout_deficit += (to - from) >> PAGE_SHIFT; 2334 VM_WAIT; 2335 goto tryagain; 2336 } 2337 vm_page_wire(p); 2338 p->valid = VM_PAGE_BITS_ALL; 2339 vm_page_flag_clear(p, PG_ZERO); 2340 pmap_kenter(pg, VM_PAGE_TO_PHYS(p)); 2341 bp->b_pages[index] = p; 2342 vm_page_wakeup(p); 2343 } 2344 bp->b_npages = index; 2345 } 2346 2347 void 2348 vm_hold_free_pages(struct buf * bp, vm_offset_t from, vm_offset_t to) 2349 { 2350 vm_offset_t pg; 2351 vm_page_t p; 2352 int index, newnpages; 2353 2354 from = round_page(from); 2355 to = round_page(to); 2356 newnpages = index = (from - trunc_page((vm_offset_t)bp->b_data)) >> PAGE_SHIFT; 2357 2358 for (pg = from; pg < to; pg += PAGE_SIZE, index++) { 2359 p = bp->b_pages[index]; 2360 if (p && (index < bp->b_npages)) { 2361 #if !defined(MAX_PERF) 2362 if (p->busy) { 2363 printf("vm_hold_free_pages: blkno: %d, lblkno: %d\n", 2364 bp->b_blkno, bp->b_lblkno); 2365 } 2366 #endif 2367 bp->b_pages[index] = NULL; 2368 pmap_kremove(pg); 2369 vm_page_busy(p); 2370 vm_page_unwire(p, 0); 2371 vm_page_free(p); 2372 } 2373 } 2374 bp->b_npages = newnpages; 2375 } 2376 2377 2378 #include "opt_ddb.h" 2379 #ifdef DDB 2380 #include <ddb/ddb.h> 2381 2382 DB_SHOW_COMMAND(buffer, db_show_buffer) 2383 { 2384 /* get args */ 2385 struct buf *bp = (struct buf *)addr; 2386 2387 if (!have_addr) { 2388 db_printf("usage: show buffer <addr>\n"); 2389 return; 2390 } 2391 2392 db_printf("b_proc = %p,\nb_flags = 0x%b\n", (void *)bp->b_proc, 2393 (u_int)bp->b_flags, PRINT_BUF_FLAGS); 2394 db_printf("b_error = %d, b_bufsize = %ld, b_bcount = %ld, " 2395 "b_resid = %ld\nb_dev = 0x%x, b_data = %p, " 2396 "b_blkno = %d, b_pblkno = %d\n", 2397 bp->b_error, bp->b_bufsize, bp->b_bcount, bp->b_resid, 2398 bp->b_dev, bp->b_data, bp->b_blkno, bp->b_pblkno); 2399 if (bp->b_npages) { 2400 int i; 2401 db_printf("b_npages = %d, pages(OBJ, IDX, PA): ", bp->b_npages); 2402 for (i = 0; i < bp->b_npages; i++) { 2403 vm_page_t m; 2404 m = bp->b_pages[i]; 2405 db_printf("(%p, 0x%lx, 0x%lx)", (void *)m->object, 2406 (u_long)m->pindex, (u_long)VM_PAGE_TO_PHYS(m)); 2407 if ((i + 1) < bp->b_npages) 2408 db_printf(","); 2409 } 2410 db_printf("\n"); 2411 } 2412 } 2413 #endif /* DDB */ 2414