vfs_bio.c revision 33134
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 1.148 1998/02/04 22:32:39 eivind 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 28#include "opt_bounce.h" 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 <vm/vm.h> 41#include <vm/vm_param.h> 42#include <vm/vm_prot.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 <vm/vm_extern.h> 48#include <vm/vm_map.h> 49#include <sys/buf.h> 50#include <sys/mount.h> 51#include <sys/malloc.h> 52#include <sys/resourcevar.h> 53 54static MALLOC_DEFINE(M_BIOBUF, "BIO buffer", "BIO buffer"); 55 56static void vfs_update __P((void)); 57static struct proc *updateproc; 58static struct kproc_desc up_kp = { 59 "update", 60 vfs_update, 61 &updateproc 62}; 63SYSINIT_KT(update, SI_SUB_KTHREAD_UPDATE, SI_ORDER_FIRST, kproc_start, &up_kp) 64 65struct buf *buf; /* buffer header pool */ 66struct swqueue bswlist; 67 68int count_lock_queue __P((void)); 69static void vm_hold_free_pages(struct buf * bp, vm_offset_t from, 70 vm_offset_t to); 71static void vm_hold_load_pages(struct buf * bp, vm_offset_t from, 72 vm_offset_t to); 73static void vfs_buf_set_valid(struct buf *bp, vm_ooffset_t foff, 74 vm_offset_t off, vm_offset_t size, 75 vm_page_t m); 76static void vfs_page_set_valid(struct buf *bp, vm_ooffset_t off, 77 int pageno, vm_page_t m); 78static void vfs_clean_pages(struct buf * bp); 79static void vfs_setdirty(struct buf *bp); 80static void vfs_vmio_release(struct buf *bp); 81static void flushdirtybuffers(int slpflag, int slptimeo); 82 83int needsbuffer; 84 85/* 86 * Internal update daemon, process 3 87 * The variable vfs_update_wakeup allows for internal syncs. 88 */ 89int vfs_update_wakeup; 90 91 92/* 93 * buffers base kva 94 */ 95 96/* 97 * bogus page -- for I/O to/from partially complete buffers 98 * this is a temporary solution to the problem, but it is not 99 * really that bad. it would be better to split the buffer 100 * for input in the case of buffers partially already in memory, 101 * but the code is intricate enough already. 102 */ 103vm_page_t bogus_page; 104static vm_offset_t bogus_offset; 105 106static int bufspace, maxbufspace, vmiospace, maxvmiobufspace, 107 bufmallocspace, maxbufmallocspace; 108int numdirtybuffers, lodirtybuffers, hidirtybuffers; 109static int numfreebuffers, lofreebuffers, hifreebuffers; 110static int kvafreespace; 111 112SYSCTL_INT(_vfs, OID_AUTO, numdirtybuffers, CTLFLAG_RD, 113 &numdirtybuffers, 0, ""); 114SYSCTL_INT(_vfs, OID_AUTO, lodirtybuffers, CTLFLAG_RW, 115 &lodirtybuffers, 0, ""); 116SYSCTL_INT(_vfs, OID_AUTO, hidirtybuffers, CTLFLAG_RW, 117 &hidirtybuffers, 0, ""); 118SYSCTL_INT(_vfs, OID_AUTO, numfreebuffers, CTLFLAG_RD, 119 &numfreebuffers, 0, ""); 120SYSCTL_INT(_vfs, OID_AUTO, lofreebuffers, CTLFLAG_RW, 121 &lofreebuffers, 0, ""); 122SYSCTL_INT(_vfs, OID_AUTO, hifreebuffers, CTLFLAG_RW, 123 &hifreebuffers, 0, ""); 124SYSCTL_INT(_vfs, OID_AUTO, maxbufspace, CTLFLAG_RW, 125 &maxbufspace, 0, ""); 126SYSCTL_INT(_vfs, OID_AUTO, bufspace, CTLFLAG_RD, 127 &bufspace, 0, ""); 128SYSCTL_INT(_vfs, OID_AUTO, maxvmiobufspace, CTLFLAG_RW, 129 &maxvmiobufspace, 0, ""); 130SYSCTL_INT(_vfs, OID_AUTO, vmiospace, CTLFLAG_RD, 131 &vmiospace, 0, ""); 132SYSCTL_INT(_vfs, OID_AUTO, maxmallocbufspace, CTLFLAG_RW, 133 &maxbufmallocspace, 0, ""); 134SYSCTL_INT(_vfs, OID_AUTO, bufmallocspace, CTLFLAG_RD, 135 &bufmallocspace, 0, ""); 136SYSCTL_INT(_vfs, OID_AUTO, kvafreespace, CTLFLAG_RD, 137 &kvafreespace, 0, ""); 138 139static LIST_HEAD(bufhashhdr, buf) bufhashtbl[BUFHSZ], invalhash; 140static TAILQ_HEAD(bqueues, buf) bufqueues[BUFFER_QUEUES]; 141 142extern int vm_swap_size; 143 144#define BUF_MAXUSE 24 145 146#define VFS_BIO_NEED_ANY 1 147#define VFS_BIO_NEED_LOWLIMIT 2 148#define VFS_BIO_NEED_FREE 4 149 150/* 151 * Initialize buffer headers and related structures. 152 */ 153void 154bufinit() 155{ 156 struct buf *bp; 157 int i; 158 159 TAILQ_INIT(&bswlist); 160 LIST_INIT(&invalhash); 161 162 /* first, make a null hash table */ 163 for (i = 0; i < BUFHSZ; i++) 164 LIST_INIT(&bufhashtbl[i]); 165 166 /* next, make a null set of free lists */ 167 for (i = 0; i < BUFFER_QUEUES; i++) 168 TAILQ_INIT(&bufqueues[i]); 169 170 /* finally, initialize each buffer header and stick on empty q */ 171 for (i = 0; i < nbuf; i++) { 172 bp = &buf[i]; 173 bzero(bp, sizeof *bp); 174 bp->b_flags = B_INVAL; /* we're just an empty header */ 175 bp->b_dev = NODEV; 176 bp->b_rcred = NOCRED; 177 bp->b_wcred = NOCRED; 178 bp->b_qindex = QUEUE_EMPTY; 179 bp->b_vnbufs.le_next = NOLIST; 180 bp->b_generation = 0; 181 TAILQ_INSERT_TAIL(&bufqueues[QUEUE_EMPTY], bp, b_freelist); 182 LIST_INSERT_HEAD(&invalhash, bp, b_hash); 183 } 184/* 185 * maxbufspace is currently calculated to support all filesystem blocks 186 * to be 8K. If you happen to use a 16K filesystem, the size of the buffer 187 * cache is still the same as it would be for 8K filesystems. This 188 * keeps the size of the buffer cache "in check" for big block filesystems. 189 */ 190 maxbufspace = (nbuf + 8) * DFLTBSIZE; 191/* 192 * reserve 1/3 of the buffers for metadata (VDIR) which might not be VMIO'ed 193 */ 194 maxvmiobufspace = 2 * maxbufspace / 3; 195/* 196 * Limit the amount of malloc memory since it is wired permanently into 197 * the kernel space. Even though this is accounted for in the buffer 198 * allocation, we don't want the malloced region to grow uncontrolled. 199 * The malloc scheme improves memory utilization significantly on average 200 * (small) directories. 201 */ 202 maxbufmallocspace = maxbufspace / 20; 203 204/* 205 * Remove the probability of deadlock conditions by limiting the 206 * number of dirty buffers. 207 */ 208 hidirtybuffers = nbuf / 8 + 20; 209 lodirtybuffers = nbuf / 16 + 10; 210 numdirtybuffers = 0; 211 lofreebuffers = nbuf / 18 + 5; 212 hifreebuffers = 2 * lofreebuffers; 213 numfreebuffers = nbuf; 214 kvafreespace = 0; 215 216 bogus_offset = kmem_alloc_pageable(kernel_map, PAGE_SIZE); 217 bogus_page = vm_page_alloc(kernel_object, 218 ((bogus_offset - VM_MIN_KERNEL_ADDRESS) >> PAGE_SHIFT), 219 VM_ALLOC_NORMAL); 220 221} 222 223/* 224 * Free the kva allocation for a buffer 225 * Must be called only at splbio or higher, 226 * as this is the only locking for buffer_map. 227 */ 228static void 229bfreekva(struct buf * bp) 230{ 231 if (bp->b_kvasize == 0) 232 return; 233 234 vm_map_delete(buffer_map, 235 (vm_offset_t) bp->b_kvabase, 236 (vm_offset_t) bp->b_kvabase + bp->b_kvasize); 237 238 bp->b_kvasize = 0; 239 240} 241 242/* 243 * remove the buffer from the appropriate free list 244 */ 245void 246bremfree(struct buf * bp) 247{ 248 int s = splbio(); 249 250 if (bp->b_qindex != QUEUE_NONE) { 251 if (bp->b_qindex == QUEUE_EMPTY) { 252 kvafreespace -= bp->b_kvasize; 253 } 254 TAILQ_REMOVE(&bufqueues[bp->b_qindex], bp, b_freelist); 255 bp->b_qindex = QUEUE_NONE; 256 } else { 257#if !defined(MAX_PERF) 258 panic("bremfree: removing a buffer when not on a queue"); 259#endif 260 } 261 if ((bp->b_flags & B_INVAL) || 262 (bp->b_flags & (B_DELWRI|B_LOCKED)) == 0) 263 --numfreebuffers; 264 splx(s); 265} 266 267 268/* 269 * Get a buffer with the specified data. Look in the cache first. 270 */ 271int 272bread(struct vnode * vp, daddr_t blkno, int size, struct ucred * cred, 273 struct buf ** bpp) 274{ 275 struct buf *bp; 276 277 bp = getblk(vp, blkno, size, 0, 0); 278 *bpp = bp; 279 280 /* if not found in cache, do some I/O */ 281 if ((bp->b_flags & B_CACHE) == 0) { 282 if (curproc != NULL) 283 curproc->p_stats->p_ru.ru_inblock++; 284 bp->b_flags |= B_READ; 285 bp->b_flags &= ~(B_DONE | B_ERROR | B_INVAL); 286 if (bp->b_rcred == NOCRED) { 287 if (cred != NOCRED) 288 crhold(cred); 289 bp->b_rcred = cred; 290 } 291 vfs_busy_pages(bp, 0); 292 VOP_STRATEGY(bp); 293 return (biowait(bp)); 294 } 295 return (0); 296} 297 298/* 299 * Operates like bread, but also starts asynchronous I/O on 300 * read-ahead blocks. 301 */ 302int 303breadn(struct vnode * vp, daddr_t blkno, int size, 304 daddr_t * rablkno, int *rabsize, 305 int cnt, struct ucred * cred, struct buf ** bpp) 306{ 307 struct buf *bp, *rabp; 308 int i; 309 int rv = 0, readwait = 0; 310 311 *bpp = bp = getblk(vp, blkno, size, 0, 0); 312 313 /* if not found in cache, do some I/O */ 314 if ((bp->b_flags & B_CACHE) == 0) { 315 if (curproc != NULL) 316 curproc->p_stats->p_ru.ru_inblock++; 317 bp->b_flags |= B_READ; 318 bp->b_flags &= ~(B_DONE | B_ERROR | B_INVAL); 319 if (bp->b_rcred == NOCRED) { 320 if (cred != NOCRED) 321 crhold(cred); 322 bp->b_rcred = cred; 323 } 324 vfs_busy_pages(bp, 0); 325 VOP_STRATEGY(bp); 326 ++readwait; 327 } 328 for (i = 0; i < cnt; i++, rablkno++, rabsize++) { 329 if (inmem(vp, *rablkno)) 330 continue; 331 rabp = getblk(vp, *rablkno, *rabsize, 0, 0); 332 333 if ((rabp->b_flags & B_CACHE) == 0) { 334 if (curproc != NULL) 335 curproc->p_stats->p_ru.ru_inblock++; 336 rabp->b_flags |= B_READ | B_ASYNC; 337 rabp->b_flags &= ~(B_DONE | B_ERROR | B_INVAL); 338 if (rabp->b_rcred == NOCRED) { 339 if (cred != NOCRED) 340 crhold(cred); 341 rabp->b_rcred = cred; 342 } 343 vfs_busy_pages(rabp, 0); 344 VOP_STRATEGY(rabp); 345 } else { 346 brelse(rabp); 347 } 348 } 349 350 if (readwait) { 351 rv = biowait(bp); 352 } 353 return (rv); 354} 355 356/* 357 * Write, release buffer on completion. (Done by iodone 358 * if async.) 359 */ 360int 361bwrite(struct buf * bp) 362{ 363 int oldflags = bp->b_flags; 364 365 if (bp->b_flags & B_INVAL) { 366 brelse(bp); 367 return (0); 368 } 369#if !defined(MAX_PERF) 370 if (!(bp->b_flags & B_BUSY)) 371 panic("bwrite: buffer is not busy???"); 372#endif 373 374 bp->b_flags &= ~(B_READ | B_DONE | B_ERROR | B_DELWRI); 375 bp->b_flags |= B_WRITEINPROG; 376 377 if ((oldflags & B_DELWRI) == B_DELWRI) { 378 --numdirtybuffers; 379 reassignbuf(bp, bp->b_vp); 380 } 381 382 bp->b_vp->v_numoutput++; 383 vfs_busy_pages(bp, 1); 384 if (curproc != NULL) 385 curproc->p_stats->p_ru.ru_oublock++; 386 VOP_STRATEGY(bp); 387 388 if ((oldflags & B_ASYNC) == 0) { 389 int rtval = biowait(bp); 390 391 if (oldflags & B_DELWRI) { 392 reassignbuf(bp, bp->b_vp); 393 } 394 brelse(bp); 395 return (rtval); 396 } 397 return (0); 398} 399 400inline void 401vfs_bio_need_satisfy(void) { 402 ++numfreebuffers; 403 if (!needsbuffer) 404 return; 405 if (numdirtybuffers < lodirtybuffers) { 406 needsbuffer &= ~(VFS_BIO_NEED_ANY | VFS_BIO_NEED_LOWLIMIT); 407 } else { 408 needsbuffer &= ~VFS_BIO_NEED_ANY; 409 } 410 if (numfreebuffers >= hifreebuffers) { 411 needsbuffer &= ~VFS_BIO_NEED_FREE; 412 } 413 wakeup(&needsbuffer); 414} 415 416/* 417 * Delayed write. (Buffer is marked dirty). 418 */ 419void 420bdwrite(struct buf * bp) 421{ 422 423#if !defined(MAX_PERF) 424 if ((bp->b_flags & B_BUSY) == 0) { 425 panic("bdwrite: buffer is not busy"); 426 } 427#endif 428 429 if (bp->b_flags & B_INVAL) { 430 brelse(bp); 431 return; 432 } 433 if (bp->b_flags & B_TAPE) { 434 bawrite(bp); 435 return; 436 } 437 bp->b_flags &= ~(B_READ|B_RELBUF); 438 if ((bp->b_flags & B_DELWRI) == 0) { 439 bp->b_flags |= B_DONE | B_DELWRI; 440 reassignbuf(bp, bp->b_vp); 441 ++numdirtybuffers; 442 } 443 444 /* 445 * This bmap keeps the system from needing to do the bmap later, 446 * perhaps when the system is attempting to do a sync. Since it 447 * is likely that the indirect block -- or whatever other datastructure 448 * that the filesystem needs is still in memory now, it is a good 449 * thing to do this. Note also, that if the pageout daemon is 450 * requesting a sync -- there might not be enough memory to do 451 * the bmap then... So, this is important to do. 452 */ 453 if (bp->b_lblkno == bp->b_blkno) { 454 VOP_BMAP(bp->b_vp, bp->b_lblkno, NULL, &bp->b_blkno, NULL, NULL); 455 } 456 457 /* 458 * Set the *dirty* buffer range based upon the VM system dirty pages. 459 */ 460 vfs_setdirty(bp); 461 462 /* 463 * We need to do this here to satisfy the vnode_pager and the 464 * pageout daemon, so that it thinks that the pages have been 465 * "cleaned". Note that since the pages are in a delayed write 466 * buffer -- the VFS layer "will" see that the pages get written 467 * out on the next sync, or perhaps the cluster will be completed. 468 */ 469 vfs_clean_pages(bp); 470 bqrelse(bp); 471 472 if (numdirtybuffers >= hidirtybuffers) 473 flushdirtybuffers(0, 0); 474 475 return; 476} 477 478/* 479 * Asynchronous write. 480 * Start output on a buffer, but do not wait for it to complete. 481 * The buffer is released when the output completes. 482 */ 483void 484bawrite(struct buf * bp) 485{ 486 bp->b_flags |= B_ASYNC; 487 (void) VOP_BWRITE(bp); 488} 489 490/* 491 * Ordered write. 492 * Start output on a buffer, but only wait for it to complete if the 493 * output device cannot guarantee ordering in some other way. Devices 494 * that can perform asynchronous ordered writes will set the B_ASYNC 495 * flag in their strategy routine. 496 * The buffer is released when the output completes. 497 */ 498int 499bowrite(struct buf * bp) 500{ 501 /* 502 * XXX Add in B_ASYNC once the SCSI 503 * layer can deal with ordered 504 * writes properly. 505 */ 506 bp->b_flags |= B_ORDERED; 507 return (VOP_BWRITE(bp)); 508} 509 510/* 511 * Release a buffer. 512 */ 513void 514brelse(struct buf * bp) 515{ 516 int s; 517 518 if (bp->b_flags & B_CLUSTER) { 519 relpbuf(bp); 520 return; 521 } 522 /* anyone need a "free" block? */ 523 s = splbio(); 524 525 /* anyone need this block? */ 526 if (bp->b_flags & B_WANTED) { 527 bp->b_flags &= ~(B_WANTED | B_AGE); 528 wakeup(bp); 529 } 530 531 if (bp->b_flags & B_LOCKED) 532 bp->b_flags &= ~B_ERROR; 533 534 if ((bp->b_flags & (B_NOCACHE | B_INVAL | B_ERROR)) || 535 (bp->b_bufsize <= 0)) { 536 bp->b_flags |= B_INVAL; 537 if (bp->b_flags & B_DELWRI) 538 --numdirtybuffers; 539 bp->b_flags &= ~(B_DELWRI | B_CACHE); 540 if (((bp->b_flags & B_VMIO) == 0) && bp->b_vp) { 541 if (bp->b_bufsize) 542 allocbuf(bp, 0); 543 brelvp(bp); 544 } 545 } 546 547 /* 548 * VMIO buffer rundown. It is not very necessary to keep a VMIO buffer 549 * constituted, so the B_INVAL flag is used to *invalidate* the buffer, 550 * but the VM object is kept around. The B_NOCACHE flag is used to 551 * invalidate the pages in the VM object. 552 * 553 * If the buffer is a partially filled NFS buffer, keep it 554 * since invalidating it now will lose informatio. The valid 555 * flags in the vm_pages have only DEV_BSIZE resolution but 556 * the b_validoff, b_validend fields have byte resolution. 557 * This can avoid unnecessary re-reads of the buffer. 558 * XXX this seems to cause performance problems. 559 */ 560 if ((bp->b_flags & B_VMIO) 561 && !(bp->b_vp->v_tag == VT_NFS && 562 bp->b_vp->v_type != VBLK && 563 (bp->b_flags & B_DELWRI) != 0) 564#ifdef notdef 565 && (bp->b_vp->v_tag != VT_NFS 566 || bp->b_vp->v_type == VBLK 567 || (bp->b_flags & (B_NOCACHE | B_INVAL | B_ERROR)) 568 || bp->b_validend == 0 569 || (bp->b_validoff == 0 570 && bp->b_validend == bp->b_bufsize)) 571#endif 572 ) { 573 vm_ooffset_t foff; 574 vm_object_t obj; 575 int i, resid; 576 vm_page_t m; 577 struct vnode *vp; 578 int iototal = bp->b_bufsize; 579 580 vp = bp->b_vp; 581 582#if !defined(MAX_PERF) 583 if (!vp) 584 panic("brelse: missing vp"); 585#endif 586 587 if (bp->b_npages) { 588 vm_pindex_t poff; 589 obj = (vm_object_t) vp->v_object; 590 if (vp->v_type == VBLK) 591 foff = ((vm_ooffset_t) bp->b_lblkno) << DEV_BSHIFT; 592 else 593 foff = (vm_ooffset_t) vp->v_mount->mnt_stat.f_iosize * bp->b_lblkno; 594 poff = OFF_TO_IDX(foff); 595 for (i = 0; i < bp->b_npages; i++) { 596 m = bp->b_pages[i]; 597 if (m == bogus_page) { 598 m = vm_page_lookup(obj, poff + i); 599#if !defined(MAX_PERF) 600 if (!m) { 601 panic("brelse: page missing\n"); 602 } 603#endif 604 bp->b_pages[i] = m; 605 pmap_qenter(trunc_page(bp->b_data), 606 bp->b_pages, bp->b_npages); 607 } 608 resid = IDX_TO_OFF(m->pindex+1) - foff; 609 if (resid > iototal) 610 resid = iototal; 611 if (resid > 0) { 612 /* 613 * Don't invalidate the page if the local machine has already 614 * modified it. This is the lesser of two evils, and should 615 * be fixed. 616 */ 617 if (bp->b_flags & (B_NOCACHE | B_ERROR)) { 618 vm_page_test_dirty(m); 619 if (m->dirty == 0) { 620 vm_page_set_invalid(m, (vm_offset_t) foff, resid); 621 if (m->valid == 0) 622 vm_page_protect(m, VM_PROT_NONE); 623 } 624 } 625 if (resid >= PAGE_SIZE) { 626 if ((m->valid & VM_PAGE_BITS_ALL) != VM_PAGE_BITS_ALL) { 627 bp->b_flags |= B_INVAL; 628 } 629 } else { 630 if (!vm_page_is_valid(m, 631 (((vm_offset_t) bp->b_data) & PAGE_MASK), resid)) { 632 bp->b_flags |= B_INVAL; 633 } 634 } 635 } 636 foff += resid; 637 iototal -= resid; 638 } 639 } 640 if (bp->b_flags & (B_INVAL | B_RELBUF)) 641 vfs_vmio_release(bp); 642 } else if (bp->b_flags & B_VMIO) { 643 if (bp->b_flags && (B_INVAL | B_RELBUF)) 644 vfs_vmio_release(bp); 645 } 646 647#if !defined(MAX_PERF) 648 if (bp->b_qindex != QUEUE_NONE) 649 panic("brelse: free buffer onto another queue???"); 650#endif 651 652 /* enqueue */ 653 /* buffers with no memory */ 654 if (bp->b_bufsize == 0) { 655 bp->b_flags |= B_INVAL; 656 bp->b_qindex = QUEUE_EMPTY; 657 TAILQ_INSERT_HEAD(&bufqueues[QUEUE_EMPTY], bp, b_freelist); 658 LIST_REMOVE(bp, b_hash); 659 LIST_INSERT_HEAD(&invalhash, bp, b_hash); 660 bp->b_dev = NODEV; 661 kvafreespace += bp->b_kvasize; 662 bp->b_generation++; 663 664 /* buffers with junk contents */ 665 } else if (bp->b_flags & (B_ERROR | B_INVAL | B_NOCACHE | B_RELBUF)) { 666 bp->b_flags |= B_INVAL; 667 bp->b_qindex = QUEUE_AGE; 668 TAILQ_INSERT_HEAD(&bufqueues[QUEUE_AGE], bp, b_freelist); 669 LIST_REMOVE(bp, b_hash); 670 LIST_INSERT_HEAD(&invalhash, bp, b_hash); 671 bp->b_dev = NODEV; 672 bp->b_generation++; 673 674 /* buffers that are locked */ 675 } else if (bp->b_flags & B_LOCKED) { 676 bp->b_qindex = QUEUE_LOCKED; 677 TAILQ_INSERT_TAIL(&bufqueues[QUEUE_LOCKED], bp, b_freelist); 678 679 /* buffers with stale but valid contents */ 680 } else if (bp->b_flags & B_AGE) { 681 bp->b_qindex = QUEUE_AGE; 682 TAILQ_INSERT_TAIL(&bufqueues[QUEUE_AGE], bp, b_freelist); 683 684 /* buffers with valid and quite potentially reuseable contents */ 685 } else { 686 bp->b_qindex = QUEUE_LRU; 687 TAILQ_INSERT_TAIL(&bufqueues[QUEUE_LRU], bp, b_freelist); 688 } 689 690 if ((bp->b_flags & B_INVAL) || 691 (bp->b_flags & (B_LOCKED|B_DELWRI)) == 0) { 692 if (bp->b_flags & B_DELWRI) { 693 --numdirtybuffers; 694 bp->b_flags &= ~B_DELWRI; 695 } 696 vfs_bio_need_satisfy(); 697 } 698 699 /* unlock */ 700 bp->b_flags &= ~(B_ORDERED | B_WANTED | B_BUSY | 701 B_ASYNC | B_NOCACHE | B_AGE | B_RELBUF); 702 splx(s); 703} 704 705/* 706 * Release a buffer. 707 */ 708void 709bqrelse(struct buf * bp) 710{ 711 int s; 712 713 s = splbio(); 714 715 /* anyone need this block? */ 716 if (bp->b_flags & B_WANTED) { 717 bp->b_flags &= ~(B_WANTED | B_AGE); 718 wakeup(bp); 719 } 720 721#if !defined(MAX_PERF) 722 if (bp->b_qindex != QUEUE_NONE) 723 panic("bqrelse: free buffer onto another queue???"); 724#endif 725 726 if (bp->b_flags & B_LOCKED) { 727 bp->b_flags &= ~B_ERROR; 728 bp->b_qindex = QUEUE_LOCKED; 729 TAILQ_INSERT_TAIL(&bufqueues[QUEUE_LOCKED], bp, b_freelist); 730 /* buffers with stale but valid contents */ 731 } else { 732 bp->b_qindex = QUEUE_LRU; 733 TAILQ_INSERT_TAIL(&bufqueues[QUEUE_LRU], bp, b_freelist); 734 } 735 736 if ((bp->b_flags & (B_LOCKED|B_DELWRI)) == 0) { 737 vfs_bio_need_satisfy(); 738 } 739 740 /* unlock */ 741 bp->b_flags &= ~(B_ORDERED | B_WANTED | B_BUSY | 742 B_ASYNC | B_NOCACHE | B_AGE | B_RELBUF); 743 splx(s); 744} 745 746static void 747vfs_vmio_release(bp) 748 struct buf *bp; 749{ 750 int i; 751 vm_page_t m; 752 753 for (i = 0; i < bp->b_npages; i++) { 754 m = bp->b_pages[i]; 755 bp->b_pages[i] = NULL; 756 vm_page_unwire(m); 757 /* 758 * We don't mess with busy pages, it is 759 * the responsibility of the process that 760 * busied the pages to deal with them. 761 */ 762 if ((m->flags & PG_BUSY) || (m->busy != 0)) 763 continue; 764 765 if (m->wire_count == 0) { 766 767 if (m->flags & PG_WANTED) { 768 m->flags &= ~PG_WANTED; 769 wakeup(m); 770 } 771 772 /* 773 * If this is an async free -- we cannot place 774 * pages onto the cache queue. If it is an 775 * async free, then we don't modify any queues. 776 * This is probably in error (for perf reasons), 777 * and we will eventually need to build 778 * a more complete infrastructure to support I/O 779 * rundown. 780 */ 781 if ((bp->b_flags & B_ASYNC) == 0) { 782 783 /* 784 * In the case of sync buffer frees, we can do pretty much 785 * anything to any of the memory queues. Specifically, 786 * the cache queue is okay to be modified. 787 */ 788 if (m->valid) { 789 if(m->dirty == 0) 790 vm_page_test_dirty(m); 791 /* 792 * this keeps pressure off of the process memory 793 */ 794 if (m->dirty == 0 && m->hold_count == 0) 795 vm_page_cache(m); 796 else 797 vm_page_deactivate(m); 798 } else if (m->hold_count == 0) { 799 m->flags |= PG_BUSY; 800 vm_page_protect(m, VM_PROT_NONE); 801 vm_page_free(m); 802 } 803 } else { 804 /* 805 * If async, then at least we clear the 806 * act_count. 807 */ 808 m->act_count = 0; 809 } 810 } 811 } 812 bufspace -= bp->b_bufsize; 813 vmiospace -= bp->b_bufsize; 814 pmap_qremove(trunc_page((vm_offset_t) bp->b_data), bp->b_npages); 815 bp->b_npages = 0; 816 bp->b_bufsize = 0; 817 bp->b_flags &= ~B_VMIO; 818 if (bp->b_vp) 819 brelvp(bp); 820} 821 822/* 823 * Check to see if a block is currently memory resident. 824 */ 825struct buf * 826gbincore(struct vnode * vp, daddr_t blkno) 827{ 828 struct buf *bp; 829 struct bufhashhdr *bh; 830 831 bh = BUFHASH(vp, blkno); 832 bp = bh->lh_first; 833 834 /* Search hash chain */ 835 while (bp != NULL) { 836 /* hit */ 837 if (bp->b_vp == vp && bp->b_lblkno == blkno && 838 (bp->b_flags & B_INVAL) == 0) { 839 break; 840 } 841 bp = bp->b_hash.le_next; 842 } 843 return (bp); 844} 845 846/* 847 * this routine implements clustered async writes for 848 * clearing out B_DELWRI buffers... This is much better 849 * than the old way of writing only one buffer at a time. 850 */ 851int 852vfs_bio_awrite(struct buf * bp) 853{ 854 int i; 855 daddr_t lblkno = bp->b_lblkno; 856 struct vnode *vp = bp->b_vp; 857 int s; 858 int ncl; 859 struct buf *bpa; 860 int nwritten; 861 int size; 862 int maxcl; 863 864 s = splbio(); 865 /* 866 * right now we support clustered writing only to regular files 867 */ 868 if ((vp->v_type == VREG) && 869 (vp->v_mount != 0) && /* Only on nodes that have the size info */ 870 (bp->b_flags & (B_CLUSTEROK | B_INVAL)) == B_CLUSTEROK) { 871 872 size = vp->v_mount->mnt_stat.f_iosize; 873 maxcl = MAXPHYS / size; 874 875 for (i = 1; i < maxcl; i++) { 876 if ((bpa = gbincore(vp, lblkno + i)) && 877 ((bpa->b_flags & (B_BUSY | B_DELWRI | B_CLUSTEROK | B_INVAL)) == 878 (B_DELWRI | B_CLUSTEROK)) && 879 (bpa->b_bufsize == size)) { 880 if ((bpa->b_blkno == bpa->b_lblkno) || 881 (bpa->b_blkno != bp->b_blkno + ((i * size) >> DEV_BSHIFT))) 882 break; 883 } else { 884 break; 885 } 886 } 887 ncl = i; 888 /* 889 * this is a possible cluster write 890 */ 891 if (ncl != 1) { 892 nwritten = cluster_wbuild(vp, size, lblkno, ncl); 893 splx(s); 894 return nwritten; 895 } 896 } 897#if 0 898 else if ((vp->v_flag & VOBJBUF) && (vp->v_type == VBLK) && 899 ((size = bp->b_bufsize) >= PAGE_SIZE)) { 900 maxcl = MAXPHYS / size; 901 for (i = 1; i < maxcl; i++) { 902 if ((bpa = gbincore(vp, lblkno + i)) && 903 ((bpa->b_flags & (B_BUSY | B_DELWRI | B_CLUSTEROK | B_INVAL)) == 904 (B_DELWRI | B_CLUSTEROK)) && 905 (bpa->b_bufsize == size)) { 906 if (bpa->b_blkno != 907 bp->b_blkno + ((i * size) >> DEV_BSHIFT)) 908 break; 909 } else { 910 break; 911 } 912 } 913 ncl = i; 914 /* 915 * this is a possible cluster write 916 */ 917 if (ncl != 1) { 918 nwritten = cluster_wbuild(vp, size, lblkno, ncl); 919 splx(s); 920 return nwritten; 921 } 922 } 923#endif 924 925 bremfree(bp); 926 splx(s); 927 /* 928 * default (old) behavior, writing out only one block 929 */ 930 bp->b_flags |= B_BUSY | B_ASYNC; 931 nwritten = bp->b_bufsize; 932 (void) VOP_BWRITE(bp); 933 return nwritten; 934} 935 936 937/* 938 * Find a buffer header which is available for use. 939 */ 940static struct buf * 941getnewbuf(struct vnode *vp, daddr_t blkno, 942 int slpflag, int slptimeo, int size, int maxsize) 943{ 944 struct buf *bp, *bp1; 945 int nbyteswritten = 0; 946 vm_offset_t addr; 947 static int writerecursion = 0; 948 949start: 950 if (bufspace >= maxbufspace) 951 goto trytofreespace; 952 953 /* can we constitute a new buffer? */ 954 if ((bp = TAILQ_FIRST(&bufqueues[QUEUE_EMPTY]))) { 955#if !defined(MAX_PERF) 956 if (bp->b_qindex != QUEUE_EMPTY) 957 panic("getnewbuf: inconsistent EMPTY queue, qindex=%d", 958 bp->b_qindex); 959#endif 960 bp->b_flags |= B_BUSY; 961 bremfree(bp); 962 goto fillbuf; 963 } 964trytofreespace: 965 /* 966 * We keep the file I/O from hogging metadata I/O 967 * This is desirable because file data is cached in the 968 * VM/Buffer cache even if a buffer is freed. 969 */ 970 if ((bp = TAILQ_FIRST(&bufqueues[QUEUE_AGE]))) { 971#if !defined(MAX_PERF) 972 if (bp->b_qindex != QUEUE_AGE) 973 panic("getnewbuf: inconsistent AGE queue, qindex=%d", 974 bp->b_qindex); 975#endif 976 } else if ((bp = TAILQ_FIRST(&bufqueues[QUEUE_LRU]))) { 977#if !defined(MAX_PERF) 978 if (bp->b_qindex != QUEUE_LRU) 979 panic("getnewbuf: inconsistent LRU queue, qindex=%d", 980 bp->b_qindex); 981#endif 982 } 983 if (!bp) { 984 /* wait for a free buffer of any kind */ 985 needsbuffer |= VFS_BIO_NEED_ANY; 986 do 987 tsleep(&needsbuffer, (PRIBIO + 1) | slpflag, "newbuf", 988 slptimeo); 989 while (needsbuffer & VFS_BIO_NEED_ANY); 990 return (0); 991 } 992 993#if defined(DIAGNOSTIC) 994 if (bp->b_flags & B_BUSY) { 995 panic("getnewbuf: busy buffer on free list\n"); 996 } 997#endif 998 999 /* 1000 * We are fairly aggressive about freeing VMIO buffers, but since 1001 * the buffering is intact without buffer headers, there is not 1002 * much loss. We gain by maintaining non-VMIOed metadata in buffers. 1003 */ 1004 if ((bp->b_qindex == QUEUE_LRU) && (bp->b_usecount > 0)) { 1005 if ((bp->b_flags & B_VMIO) == 0 || 1006 (vmiospace < maxvmiobufspace)) { 1007 --bp->b_usecount; 1008 TAILQ_REMOVE(&bufqueues[QUEUE_LRU], bp, b_freelist); 1009 if (TAILQ_FIRST(&bufqueues[QUEUE_LRU]) != NULL) { 1010 TAILQ_INSERT_TAIL(&bufqueues[QUEUE_LRU], bp, b_freelist); 1011 goto start; 1012 } 1013 TAILQ_INSERT_TAIL(&bufqueues[QUEUE_LRU], bp, b_freelist); 1014 } 1015 } 1016 1017 1018 /* if we are a delayed write, convert to an async write */ 1019 if ((bp->b_flags & (B_DELWRI | B_INVAL)) == B_DELWRI) { 1020 1021 /* 1022 * If our delayed write is likely to be used soon, then 1023 * recycle back onto the LRU queue. 1024 */ 1025 if (vp && (bp->b_vp == vp) && (bp->b_qindex == QUEUE_LRU) && 1026 (bp->b_lblkno >= blkno) && (maxsize > 0)) { 1027 1028 if (bp->b_usecount > 0) { 1029 if (bp->b_lblkno < blkno + (MAXPHYS / maxsize)) { 1030 1031 TAILQ_REMOVE(&bufqueues[QUEUE_LRU], bp, b_freelist); 1032 1033 if (TAILQ_FIRST(&bufqueues[QUEUE_LRU]) != NULL) { 1034 TAILQ_INSERT_TAIL(&bufqueues[QUEUE_LRU], bp, b_freelist); 1035 bp->b_usecount--; 1036 goto start; 1037 } 1038 TAILQ_INSERT_TAIL(&bufqueues[QUEUE_LRU], bp, b_freelist); 1039 } 1040 } 1041 } 1042 1043 /* 1044 * Certain layered filesystems can recursively re-enter the vfs_bio 1045 * code, due to delayed writes. This helps keep the system from 1046 * deadlocking. 1047 */ 1048 if (writerecursion > 0) { 1049 bp = TAILQ_FIRST(&bufqueues[QUEUE_AGE]); 1050 while (bp) { 1051 if ((bp->b_flags & B_DELWRI) == 0) 1052 break; 1053 bp = TAILQ_NEXT(bp, b_freelist); 1054 } 1055 if (bp == NULL) { 1056 bp = TAILQ_FIRST(&bufqueues[QUEUE_LRU]); 1057 while (bp) { 1058 if ((bp->b_flags & B_DELWRI) == 0) 1059 break; 1060 bp = TAILQ_NEXT(bp, b_freelist); 1061 } 1062 } 1063 if (bp == NULL) 1064 panic("getnewbuf: cannot get buffer, infinite recursion failure"); 1065 } else { 1066 ++writerecursion; 1067 nbyteswritten += vfs_bio_awrite(bp); 1068 --writerecursion; 1069 if (!slpflag && !slptimeo) { 1070 return (0); 1071 } 1072 goto start; 1073 } 1074 } 1075 1076 if (bp->b_flags & B_WANTED) { 1077 bp->b_flags &= ~B_WANTED; 1078 wakeup(bp); 1079 } 1080 bremfree(bp); 1081 bp->b_flags |= B_BUSY; 1082 1083 if (bp->b_flags & B_VMIO) { 1084 bp->b_flags &= ~B_ASYNC; 1085 vfs_vmio_release(bp); 1086 } 1087 1088 if (bp->b_vp) 1089 brelvp(bp); 1090 1091fillbuf: 1092 bp->b_generation++; 1093 1094 /* we are not free, nor do we contain interesting data */ 1095 if (bp->b_rcred != NOCRED) { 1096 crfree(bp->b_rcred); 1097 bp->b_rcred = NOCRED; 1098 } 1099 if (bp->b_wcred != NOCRED) { 1100 crfree(bp->b_wcred); 1101 bp->b_wcred = NOCRED; 1102 } 1103 1104 LIST_REMOVE(bp, b_hash); 1105 LIST_INSERT_HEAD(&invalhash, bp, b_hash); 1106 if (bp->b_bufsize) { 1107 allocbuf(bp, 0); 1108 } 1109 bp->b_flags = B_BUSY; 1110 bp->b_dev = NODEV; 1111 bp->b_vp = NULL; 1112 bp->b_blkno = bp->b_lblkno = 0; 1113 bp->b_iodone = 0; 1114 bp->b_error = 0; 1115 bp->b_resid = 0; 1116 bp->b_bcount = 0; 1117 bp->b_npages = 0; 1118 bp->b_dirtyoff = bp->b_dirtyend = 0; 1119 bp->b_validoff = bp->b_validend = 0; 1120 bp->b_usecount = 5; 1121 1122 maxsize = (maxsize + PAGE_MASK) & ~PAGE_MASK; 1123 1124 /* 1125 * we assume that buffer_map is not at address 0 1126 */ 1127 addr = 0; 1128 if (maxsize != bp->b_kvasize) { 1129 bfreekva(bp); 1130 1131findkvaspace: 1132 /* 1133 * See if we have buffer kva space 1134 */ 1135 if (vm_map_findspace(buffer_map, 1136 vm_map_min(buffer_map), maxsize, &addr)) { 1137 if (kvafreespace > 0) { 1138 int totfree = 0, freed; 1139 do { 1140 freed = 0; 1141 for (bp1 = TAILQ_FIRST(&bufqueues[QUEUE_EMPTY]); 1142 bp1 != NULL; bp1 = TAILQ_NEXT(bp1, b_freelist)) { 1143 if (bp1->b_kvasize != 0) { 1144 totfree += bp1->b_kvasize; 1145 freed = bp1->b_kvasize; 1146 bremfree(bp1); 1147 bfreekva(bp1); 1148 brelse(bp1); 1149 break; 1150 } 1151 } 1152 } while (freed); 1153 /* 1154 * if we found free space, then retry with the same buffer. 1155 */ 1156 if (totfree) 1157 goto findkvaspace; 1158 } 1159 bp->b_flags |= B_INVAL; 1160 brelse(bp); 1161 goto trytofreespace; 1162 } 1163 } 1164 1165 /* 1166 * See if we are below are allocated minimum 1167 */ 1168 if (bufspace >= (maxbufspace + nbyteswritten)) { 1169 bp->b_flags |= B_INVAL; 1170 brelse(bp); 1171 goto trytofreespace; 1172 } 1173 1174 /* 1175 * create a map entry for the buffer -- in essence 1176 * reserving the kva space. 1177 */ 1178 if (addr) { 1179 vm_map_insert(buffer_map, NULL, 0, 1180 addr, addr + maxsize, 1181 VM_PROT_ALL, VM_PROT_ALL, MAP_NOFAULT); 1182 1183 bp->b_kvabase = (caddr_t) addr; 1184 bp->b_kvasize = maxsize; 1185 } 1186 bp->b_data = bp->b_kvabase; 1187 1188 return (bp); 1189} 1190 1191static void 1192waitfreebuffers(int slpflag, int slptimeo) { 1193 while (numfreebuffers < hifreebuffers) { 1194 flushdirtybuffers(slpflag, slptimeo); 1195 if (numfreebuffers < hifreebuffers) 1196 break; 1197 needsbuffer |= VFS_BIO_NEED_FREE; 1198 if (tsleep(&needsbuffer, PRIBIO|slpflag, "biofre", slptimeo)) 1199 break; 1200 } 1201} 1202 1203static void 1204flushdirtybuffers(int slpflag, int slptimeo) { 1205 int s; 1206 static pid_t flushing = 0; 1207 1208 s = splbio(); 1209 1210 if (flushing) { 1211 if (flushing == curproc->p_pid) { 1212 splx(s); 1213 return; 1214 } 1215 while (flushing) { 1216 if (tsleep(&flushing, PRIBIO|slpflag, "biofls", slptimeo)) { 1217 splx(s); 1218 return; 1219 } 1220 } 1221 } 1222 flushing = curproc->p_pid; 1223 1224 while (numdirtybuffers > lodirtybuffers) { 1225 struct buf *bp; 1226 needsbuffer |= VFS_BIO_NEED_LOWLIMIT; 1227 bp = TAILQ_FIRST(&bufqueues[QUEUE_AGE]); 1228 if (bp == NULL) 1229 bp = TAILQ_FIRST(&bufqueues[QUEUE_LRU]); 1230 1231 while (bp && ((bp->b_flags & B_DELWRI) == 0)) { 1232 bp = TAILQ_NEXT(bp, b_freelist); 1233 } 1234 1235 if (bp) { 1236 vfs_bio_awrite(bp); 1237 continue; 1238 } 1239 break; 1240 } 1241 1242 flushing = 0; 1243 wakeup(&flushing); 1244 splx(s); 1245} 1246 1247/* 1248 * Check to see if a block is currently memory resident. 1249 */ 1250struct buf * 1251incore(struct vnode * vp, daddr_t blkno) 1252{ 1253 struct buf *bp; 1254 1255 int s = splbio(); 1256 bp = gbincore(vp, blkno); 1257 splx(s); 1258 return (bp); 1259} 1260 1261/* 1262 * Returns true if no I/O is needed to access the 1263 * associated VM object. This is like incore except 1264 * it also hunts around in the VM system for the data. 1265 */ 1266 1267int 1268inmem(struct vnode * vp, daddr_t blkno) 1269{ 1270 vm_object_t obj; 1271 vm_offset_t toff, tinc; 1272 vm_page_t m; 1273 vm_ooffset_t off; 1274 1275 if (incore(vp, blkno)) 1276 return 1; 1277 if (vp->v_mount == NULL) 1278 return 0; 1279 if ((vp->v_object == NULL) || (vp->v_flag & VOBJBUF) == 0) 1280 return 0; 1281 1282 obj = vp->v_object; 1283 tinc = PAGE_SIZE; 1284 if (tinc > vp->v_mount->mnt_stat.f_iosize) 1285 tinc = vp->v_mount->mnt_stat.f_iosize; 1286 off = blkno * vp->v_mount->mnt_stat.f_iosize; 1287 1288 for (toff = 0; toff < vp->v_mount->mnt_stat.f_iosize; toff += tinc) { 1289 1290 m = vm_page_lookup(obj, OFF_TO_IDX(off + toff)); 1291 if (!m) 1292 return 0; 1293 if (vm_page_is_valid(m, (vm_offset_t) (toff + off), tinc) == 0) 1294 return 0; 1295 } 1296 return 1; 1297} 1298 1299/* 1300 * now we set the dirty range for the buffer -- 1301 * for NFS -- if the file is mapped and pages have 1302 * been written to, let it know. We want the 1303 * entire range of the buffer to be marked dirty if 1304 * any of the pages have been written to for consistancy 1305 * with the b_validoff, b_validend set in the nfs write 1306 * code, and used by the nfs read code. 1307 */ 1308static void 1309vfs_setdirty(struct buf *bp) { 1310 int i; 1311 vm_object_t object; 1312 vm_offset_t boffset, offset; 1313 /* 1314 * We qualify the scan for modified pages on whether the 1315 * object has been flushed yet. The OBJ_WRITEABLE flag 1316 * is not cleared simply by protecting pages off. 1317 */ 1318 if ((bp->b_flags & B_VMIO) && 1319 ((object = bp->b_pages[0]->object)->flags & (OBJ_WRITEABLE|OBJ_CLEANING))) { 1320 /* 1321 * test the pages to see if they have been modified directly 1322 * by users through the VM system. 1323 */ 1324 for (i = 0; i < bp->b_npages; i++) 1325 vm_page_test_dirty(bp->b_pages[i]); 1326 1327 /* 1328 * scan forwards for the first page modified 1329 */ 1330 for (i = 0; i < bp->b_npages; i++) { 1331 if (bp->b_pages[i]->dirty) { 1332 break; 1333 } 1334 } 1335 boffset = (i << PAGE_SHIFT); 1336 if (boffset < bp->b_dirtyoff) { 1337 bp->b_dirtyoff = boffset; 1338 } 1339 1340 /* 1341 * scan backwards for the last page modified 1342 */ 1343 for (i = bp->b_npages - 1; i >= 0; --i) { 1344 if (bp->b_pages[i]->dirty) { 1345 break; 1346 } 1347 } 1348 boffset = (i + 1); 1349 offset = boffset + bp->b_pages[0]->pindex; 1350 if (offset >= object->size) 1351 boffset = object->size - bp->b_pages[0]->pindex; 1352 if (bp->b_dirtyend < (boffset << PAGE_SHIFT)) 1353 bp->b_dirtyend = (boffset << PAGE_SHIFT); 1354 } 1355} 1356 1357/* 1358 * Get a block given a specified block and offset into a file/device. 1359 */ 1360struct buf * 1361getblk(struct vnode * vp, daddr_t blkno, int size, int slpflag, int slptimeo) 1362{ 1363 struct buf *bp; 1364 int s; 1365 struct bufhashhdr *bh; 1366 int maxsize; 1367 int generation; 1368 1369 if (vp->v_mount) { 1370 maxsize = vp->v_mount->mnt_stat.f_iosize; 1371 /* 1372 * This happens on mount points. 1373 */ 1374 if (maxsize < size) 1375 maxsize = size; 1376 } else { 1377 maxsize = size; 1378 } 1379 1380#if !defined(MAX_PERF) 1381 if (size > MAXBSIZE) 1382 panic("getblk: size(%d) > MAXBSIZE(%d)\n", size, MAXBSIZE); 1383#endif 1384 1385 s = splbio(); 1386loop: 1387 if (numfreebuffers < lofreebuffers) { 1388 waitfreebuffers(slpflag, slptimeo); 1389 } 1390 1391 if ((bp = gbincore(vp, blkno))) { 1392loop1: 1393 generation = bp->b_generation; 1394 if (bp->b_flags & B_BUSY) { 1395 bp->b_flags |= B_WANTED; 1396 if (bp->b_usecount < BUF_MAXUSE) 1397 ++bp->b_usecount; 1398 if (!tsleep(bp, 1399 (PRIBIO + 1) | slpflag, "getblk", slptimeo)) { 1400 if (bp->b_generation != generation) 1401 goto loop; 1402 goto loop1; 1403 } else { 1404 splx(s); 1405 return (struct buf *) NULL; 1406 } 1407 } 1408 bp->b_flags |= B_BUSY | B_CACHE; 1409 bremfree(bp); 1410 1411 /* 1412 * check for size inconsistancies (note that they shouldn't 1413 * happen but do when filesystems don't handle the size changes 1414 * correctly.) We are conservative on metadata and don't just 1415 * extend the buffer but write and re-constitute it. 1416 */ 1417 1418 if (bp->b_bcount != size) { 1419 bp->b_generation++; 1420 if ((bp->b_flags & B_VMIO) && (size <= bp->b_kvasize)) { 1421 allocbuf(bp, size); 1422 } else { 1423 bp->b_flags |= B_NOCACHE; 1424 VOP_BWRITE(bp); 1425 goto loop; 1426 } 1427 } 1428 1429 if (bp->b_usecount < BUF_MAXUSE) 1430 ++bp->b_usecount; 1431 splx(s); 1432 return (bp); 1433 } else { 1434 vm_object_t obj; 1435 1436 if ((bp = getnewbuf(vp, blkno, 1437 slpflag, slptimeo, size, maxsize)) == 0) { 1438 if (slpflag || slptimeo) { 1439 splx(s); 1440 return NULL; 1441 } 1442 goto loop; 1443 } 1444 1445 /* 1446 * This code is used to make sure that a buffer is not 1447 * created while the getnewbuf routine is blocked. 1448 * Normally the vnode is locked so this isn't a problem. 1449 * VBLK type I/O requests, however, don't lock the vnode. 1450 */ 1451 if (!VOP_ISLOCKED(vp) && gbincore(vp, blkno)) { 1452 bp->b_flags |= B_INVAL; 1453 brelse(bp); 1454 goto loop; 1455 } 1456 1457 /* 1458 * Insert the buffer into the hash, so that it can 1459 * be found by incore. 1460 */ 1461 bp->b_blkno = bp->b_lblkno = blkno; 1462 bgetvp(vp, bp); 1463 LIST_REMOVE(bp, b_hash); 1464 bh = BUFHASH(vp, blkno); 1465 LIST_INSERT_HEAD(bh, bp, b_hash); 1466 1467 if ((obj = vp->v_object) && (vp->v_flag & VOBJBUF)) { 1468 bp->b_flags |= (B_VMIO | B_CACHE); 1469#if defined(VFS_BIO_DEBUG) 1470 if (vp->v_type != VREG && vp->v_type != VBLK) 1471 printf("getblk: vmioing file type %d???\n", vp->v_type); 1472#endif 1473 } else { 1474 bp->b_flags &= ~B_VMIO; 1475 } 1476 splx(s); 1477 1478 allocbuf(bp, size); 1479#ifdef PC98 1480 /* 1481 * 1024byte/sector support 1482 */ 1483#define B_XXX2 0x8000000 1484 if (vp->v_flag & 0x10000) bp->b_flags |= B_XXX2; 1485#endif 1486 return (bp); 1487 } 1488} 1489 1490/* 1491 * Get an empty, disassociated buffer of given size. 1492 */ 1493struct buf * 1494geteblk(int size) 1495{ 1496 struct buf *bp; 1497 int s; 1498 1499 s = splbio(); 1500 while ((bp = getnewbuf(0, (daddr_t) 0, 0, 0, size, MAXBSIZE)) == 0); 1501 splx(s); 1502 allocbuf(bp, size); 1503 bp->b_flags |= B_INVAL; 1504 return (bp); 1505} 1506 1507 1508/* 1509 * This code constitutes the buffer memory from either anonymous system 1510 * memory (in the case of non-VMIO operations) or from an associated 1511 * VM object (in the case of VMIO operations). 1512 * 1513 * Note that this code is tricky, and has many complications to resolve 1514 * deadlock or inconsistant data situations. Tread lightly!!! 1515 * 1516 * Modify the length of a buffer's underlying buffer storage without 1517 * destroying information (unless, of course the buffer is shrinking). 1518 */ 1519int 1520allocbuf(struct buf * bp, int size) 1521{ 1522 1523 int s; 1524 int newbsize, mbsize; 1525 int i; 1526 1527#if !defined(MAX_PERF) 1528 if (!(bp->b_flags & B_BUSY)) 1529 panic("allocbuf: buffer not busy"); 1530 1531 if (bp->b_kvasize < size) 1532 panic("allocbuf: buffer too small"); 1533#endif 1534 1535 if ((bp->b_flags & B_VMIO) == 0) { 1536 caddr_t origbuf; 1537 int origbufsize; 1538 /* 1539 * Just get anonymous memory from the kernel 1540 */ 1541 mbsize = (size + DEV_BSIZE - 1) & ~(DEV_BSIZE - 1); 1542#if !defined(NO_B_MALLOC) 1543 if (bp->b_flags & B_MALLOC) 1544 newbsize = mbsize; 1545 else 1546#endif 1547 newbsize = round_page(size); 1548 1549 if (newbsize < bp->b_bufsize) { 1550#if !defined(NO_B_MALLOC) 1551 /* 1552 * malloced buffers are not shrunk 1553 */ 1554 if (bp->b_flags & B_MALLOC) { 1555 if (newbsize) { 1556 bp->b_bcount = size; 1557 } else { 1558 free(bp->b_data, M_BIOBUF); 1559 bufspace -= bp->b_bufsize; 1560 bufmallocspace -= bp->b_bufsize; 1561 bp->b_data = bp->b_kvabase; 1562 bp->b_bufsize = 0; 1563 bp->b_bcount = 0; 1564 bp->b_flags &= ~B_MALLOC; 1565 } 1566 return 1; 1567 } 1568#endif 1569 vm_hold_free_pages( 1570 bp, 1571 (vm_offset_t) bp->b_data + newbsize, 1572 (vm_offset_t) bp->b_data + bp->b_bufsize); 1573 } else if (newbsize > bp->b_bufsize) { 1574#if !defined(NO_B_MALLOC) 1575 /* 1576 * We only use malloced memory on the first allocation. 1577 * and revert to page-allocated memory when the buffer grows. 1578 */ 1579 if ( (bufmallocspace < maxbufmallocspace) && 1580 (bp->b_bufsize == 0) && 1581 (mbsize <= PAGE_SIZE/2)) { 1582 1583 bp->b_data = malloc(mbsize, M_BIOBUF, M_WAITOK); 1584 bp->b_bufsize = mbsize; 1585 bp->b_bcount = size; 1586 bp->b_flags |= B_MALLOC; 1587 bufspace += mbsize; 1588 bufmallocspace += mbsize; 1589 return 1; 1590 } 1591#endif 1592 origbuf = NULL; 1593 origbufsize = 0; 1594#if !defined(NO_B_MALLOC) 1595 /* 1596 * If the buffer is growing on it's other-than-first allocation, 1597 * then we revert to the page-allocation scheme. 1598 */ 1599 if (bp->b_flags & B_MALLOC) { 1600 origbuf = bp->b_data; 1601 origbufsize = bp->b_bufsize; 1602 bp->b_data = bp->b_kvabase; 1603 bufspace -= bp->b_bufsize; 1604 bufmallocspace -= bp->b_bufsize; 1605 bp->b_bufsize = 0; 1606 bp->b_flags &= ~B_MALLOC; 1607 newbsize = round_page(newbsize); 1608 } 1609#endif 1610 vm_hold_load_pages( 1611 bp, 1612 (vm_offset_t) bp->b_data + bp->b_bufsize, 1613 (vm_offset_t) bp->b_data + newbsize); 1614#if !defined(NO_B_MALLOC) 1615 if (origbuf) { 1616 bcopy(origbuf, bp->b_data, origbufsize); 1617 free(origbuf, M_BIOBUF); 1618 } 1619#endif 1620 } 1621 } else { 1622 vm_page_t m; 1623 int desiredpages; 1624 1625 newbsize = (size + DEV_BSIZE - 1) & ~(DEV_BSIZE - 1); 1626 desiredpages = (round_page(newbsize) >> PAGE_SHIFT); 1627 1628#if !defined(NO_B_MALLOC) 1629 if (bp->b_flags & B_MALLOC) 1630 panic("allocbuf: VMIO buffer can't be malloced"); 1631#endif 1632 1633 if (newbsize < bp->b_bufsize) { 1634 if (desiredpages < bp->b_npages) { 1635 for (i = desiredpages; i < bp->b_npages; i++) { 1636 /* 1637 * the page is not freed here -- it 1638 * is the responsibility of vnode_pager_setsize 1639 */ 1640 m = bp->b_pages[i]; 1641#if defined(DIAGNOSTIC) 1642 if (m == bogus_page) 1643 panic("allocbuf: bogus page found"); 1644#endif 1645 s = splvm(); 1646 while ((m->flags & PG_BUSY) || (m->busy != 0)) { 1647 m->flags |= PG_WANTED; 1648 tsleep(m, PVM, "biodep", 0); 1649 } 1650 splx(s); 1651 1652 bp->b_pages[i] = NULL; 1653 vm_page_unwire(m); 1654 } 1655 pmap_qremove((vm_offset_t) trunc_page(bp->b_data) + 1656 (desiredpages << PAGE_SHIFT), (bp->b_npages - desiredpages)); 1657 bp->b_npages = desiredpages; 1658 } 1659 } else if (newbsize > bp->b_bufsize) { 1660 vm_object_t obj; 1661 vm_offset_t tinc, toff; 1662 vm_ooffset_t off; 1663 vm_pindex_t objoff; 1664 int pageindex, curbpnpages; 1665 struct vnode *vp; 1666 int bsize; 1667 1668 vp = bp->b_vp; 1669 1670 if (vp->v_type == VBLK) 1671 bsize = DEV_BSIZE; 1672 else 1673 bsize = vp->v_mount->mnt_stat.f_iosize; 1674 1675 if (bp->b_npages < desiredpages) { 1676 obj = vp->v_object; 1677 tinc = PAGE_SIZE; 1678 if (tinc > bsize) 1679 tinc = bsize; 1680 off = (vm_ooffset_t) bp->b_lblkno * bsize; 1681 curbpnpages = bp->b_npages; 1682 doretry: 1683 bp->b_flags |= B_CACHE; 1684 bp->b_validoff = bp->b_validend = 0; 1685 for (toff = 0; toff < newbsize; toff += tinc) { 1686 int bytesinpage; 1687 1688 pageindex = toff >> PAGE_SHIFT; 1689 objoff = OFF_TO_IDX(off + toff); 1690 if (pageindex < curbpnpages) { 1691 1692 m = bp->b_pages[pageindex]; 1693#ifdef VFS_BIO_DIAG 1694 if (m->pindex != objoff) 1695 panic("allocbuf: page changed offset??!!!?"); 1696#endif 1697 bytesinpage = tinc; 1698 if (tinc > (newbsize - toff)) 1699 bytesinpage = newbsize - toff; 1700 if (bp->b_flags & B_CACHE) 1701 vfs_buf_set_valid(bp, off, toff, bytesinpage, m); 1702 continue; 1703 } 1704 m = vm_page_lookup(obj, objoff); 1705 if (!m) { 1706 m = vm_page_alloc(obj, objoff, VM_ALLOC_NORMAL); 1707 if (!m) { 1708 VM_WAIT; 1709 vm_pageout_deficit += (desiredpages - bp->b_npages); 1710 goto doretry; 1711 } 1712 /* 1713 * Normally it is unwise to clear PG_BUSY without 1714 * PAGE_WAKEUP -- but it is okay here, as there is 1715 * no chance for blocking between here and vm_page_alloc 1716 */ 1717 m->flags &= ~PG_BUSY; 1718 vm_page_wire(m); 1719 bp->b_flags &= ~B_CACHE; 1720 } else if (m->flags & PG_BUSY) { 1721 s = splvm(); 1722 if (m->flags & PG_BUSY) { 1723 m->flags |= PG_WANTED; 1724 tsleep(m, PVM, "pgtblk", 0); 1725 } 1726 splx(s); 1727 goto doretry; 1728 } else { 1729 if ((curproc != pageproc) && 1730 ((m->queue - m->pc) == PQ_CACHE) && 1731 ((cnt.v_free_count + cnt.v_cache_count) < 1732 (cnt.v_free_min + cnt.v_cache_min))) { 1733 pagedaemon_wakeup(); 1734 } 1735 bytesinpage = tinc; 1736 if (tinc > (newbsize - toff)) 1737 bytesinpage = newbsize - toff; 1738 if (bp->b_flags & B_CACHE) 1739 vfs_buf_set_valid(bp, off, toff, bytesinpage, m); 1740 vm_page_wire(m); 1741 } 1742 bp->b_pages[pageindex] = m; 1743 curbpnpages = pageindex + 1; 1744 } 1745 if (vp->v_tag == VT_NFS && 1746 vp->v_type != VBLK) { 1747 if (bp->b_dirtyend > 0) { 1748 bp->b_validoff = min(bp->b_validoff, bp->b_dirtyoff); 1749 bp->b_validend = max(bp->b_validend, bp->b_dirtyend); 1750 } 1751 if (bp->b_validend == 0) 1752 bp->b_flags &= ~B_CACHE; 1753 } 1754 bp->b_data = (caddr_t) trunc_page(bp->b_data); 1755 bp->b_npages = curbpnpages; 1756 pmap_qenter((vm_offset_t) bp->b_data, 1757 bp->b_pages, bp->b_npages); 1758 ((vm_offset_t) bp->b_data) |= off & PAGE_MASK; 1759 } 1760 } 1761 } 1762 if (bp->b_flags & B_VMIO) 1763 vmiospace += (newbsize - bp->b_bufsize); 1764 bufspace += (newbsize - bp->b_bufsize); 1765 bp->b_bufsize = newbsize; 1766 bp->b_bcount = size; 1767 return 1; 1768} 1769 1770/* 1771 * Wait for buffer I/O completion, returning error status. 1772 */ 1773int 1774biowait(register struct buf * bp) 1775{ 1776 int s; 1777 1778 s = splbio(); 1779 while ((bp->b_flags & B_DONE) == 0) 1780#if defined(NO_SCHEDULE_MODS) 1781 tsleep(bp, PRIBIO, "biowait", 0); 1782#else 1783 if (bp->b_flags & B_READ) 1784 tsleep(bp, PRIBIO, "biord", 0); 1785 else 1786 tsleep(bp, curproc->p_usrpri, "biowr", 0); 1787#endif 1788 splx(s); 1789 if (bp->b_flags & B_EINTR) { 1790 bp->b_flags &= ~B_EINTR; 1791 return (EINTR); 1792 } 1793 if (bp->b_flags & B_ERROR) { 1794 return (bp->b_error ? bp->b_error : EIO); 1795 } else { 1796 return (0); 1797 } 1798} 1799 1800/* 1801 * Finish I/O on a buffer, calling an optional function. 1802 * This is usually called from interrupt level, so process blocking 1803 * is not *a good idea*. 1804 */ 1805void 1806biodone(register struct buf * bp) 1807{ 1808 int s; 1809 1810 s = splbio(); 1811 1812#if !defined(MAX_PERF) 1813 if (!(bp->b_flags & B_BUSY)) 1814 panic("biodone: buffer not busy"); 1815#endif 1816 1817 if (bp->b_flags & B_DONE) { 1818 splx(s); 1819#if !defined(MAX_PERF) 1820 printf("biodone: buffer already done\n"); 1821#endif 1822 return; 1823 } 1824 bp->b_flags |= B_DONE; 1825 1826 if ((bp->b_flags & B_READ) == 0) { 1827 vwakeup(bp); 1828 } 1829#ifdef BOUNCE_BUFFERS 1830 if (bp->b_flags & B_BOUNCE) 1831 vm_bounce_free(bp); 1832#endif 1833 1834 /* call optional completion function if requested */ 1835 if (bp->b_flags & B_CALL) { 1836 bp->b_flags &= ~B_CALL; 1837 (*bp->b_iodone) (bp); 1838 splx(s); 1839 return; 1840 } 1841 if (bp->b_flags & B_VMIO) { 1842 int i, resid; 1843 vm_ooffset_t foff; 1844 vm_page_t m; 1845 vm_object_t obj; 1846 int iosize; 1847 struct vnode *vp = bp->b_vp; 1848 1849 obj = vp->v_object; 1850 1851#if defined(VFS_BIO_DEBUG) 1852 if (vp->v_usecount == 0) { 1853 panic("biodone: zero vnode ref count"); 1854 } 1855 1856 if (vp->v_object == NULL) { 1857 panic("biodone: missing VM object"); 1858 } 1859 1860 if ((vp->v_flag & VOBJBUF) == 0) { 1861 panic("biodone: vnode is not setup for merged cache"); 1862 } 1863#endif 1864 1865 if (vp->v_type == VBLK) 1866 foff = (vm_ooffset_t) DEV_BSIZE * bp->b_lblkno; 1867 else 1868 foff = (vm_ooffset_t) vp->v_mount->mnt_stat.f_iosize * bp->b_lblkno; 1869#if !defined(MAX_PERF) 1870 if (!obj) { 1871 panic("biodone: no object"); 1872 } 1873#endif 1874#if defined(VFS_BIO_DEBUG) 1875 if (obj->paging_in_progress < bp->b_npages) { 1876 printf("biodone: paging in progress(%d) < bp->b_npages(%d)\n", 1877 obj->paging_in_progress, bp->b_npages); 1878 } 1879#endif 1880 iosize = bp->b_bufsize; 1881 for (i = 0; i < bp->b_npages; i++) { 1882 int bogusflag = 0; 1883 m = bp->b_pages[i]; 1884 if (m == bogus_page) { 1885 bogusflag = 1; 1886 m = vm_page_lookup(obj, OFF_TO_IDX(foff)); 1887 if (!m) { 1888#if defined(VFS_BIO_DEBUG) 1889 printf("biodone: page disappeared\n"); 1890#endif 1891 --obj->paging_in_progress; 1892 continue; 1893 } 1894 bp->b_pages[i] = m; 1895 pmap_qenter(trunc_page(bp->b_data), bp->b_pages, bp->b_npages); 1896 } 1897#if defined(VFS_BIO_DEBUG) 1898 if (OFF_TO_IDX(foff) != m->pindex) { 1899 printf("biodone: foff(%d)/m->pindex(%d) mismatch\n", foff, m->pindex); 1900 } 1901#endif 1902 resid = IDX_TO_OFF(m->pindex + 1) - foff; 1903 if (resid > iosize) 1904 resid = iosize; 1905 /* 1906 * In the write case, the valid and clean bits are 1907 * already changed correctly, so we only need to do this 1908 * here in the read case. 1909 */ 1910 if ((bp->b_flags & B_READ) && !bogusflag && resid > 0) { 1911 vfs_page_set_valid(bp, foff, i, m); 1912 } 1913 1914 /* 1915 * when debugging new filesystems or buffer I/O methods, this 1916 * is the most common error that pops up. if you see this, you 1917 * have not set the page busy flag correctly!!! 1918 */ 1919 if (m->busy == 0) { 1920#if !defined(MAX_PERF) 1921 printf("biodone: page busy < 0, " 1922 "pindex: %d, foff: 0x(%x,%x), " 1923 "resid: %d, index: %d\n", 1924 (int) m->pindex, (int)(foff >> 32), 1925 (int) foff & 0xffffffff, resid, i); 1926#endif 1927 if (vp->v_type != VBLK) 1928#if !defined(MAX_PERF) 1929 printf(" iosize: %ld, lblkno: %d, flags: 0x%lx, npages: %d\n", 1930 bp->b_vp->v_mount->mnt_stat.f_iosize, 1931 (int) bp->b_lblkno, 1932 bp->b_flags, bp->b_npages); 1933 else 1934 printf(" VDEV, lblkno: %d, flags: 0x%lx, npages: %d\n", 1935 (int) bp->b_lblkno, 1936 bp->b_flags, bp->b_npages); 1937 printf(" valid: 0x%x, dirty: 0x%x, wired: %d\n", 1938 m->valid, m->dirty, m->wire_count); 1939#endif 1940 panic("biodone: page busy < 0\n"); 1941 } 1942 --m->busy; 1943 if ((m->busy == 0) && (m->flags & PG_WANTED)) { 1944 m->flags &= ~PG_WANTED; 1945 wakeup(m); 1946 } 1947 --obj->paging_in_progress; 1948 foff += resid; 1949 iosize -= resid; 1950 } 1951 if (obj && obj->paging_in_progress == 0 && 1952 (obj->flags & OBJ_PIPWNT)) { 1953 obj->flags &= ~OBJ_PIPWNT; 1954 wakeup(obj); 1955 } 1956 } 1957 /* 1958 * For asynchronous completions, release the buffer now. The brelse 1959 * checks for B_WANTED and will do the wakeup there if necessary - so 1960 * no need to do a wakeup here in the async case. 1961 */ 1962 1963 if (bp->b_flags & B_ASYNC) { 1964 if ((bp->b_flags & (B_NOCACHE | B_INVAL | B_ERROR | B_RELBUF)) != 0) 1965 brelse(bp); 1966 else 1967 bqrelse(bp); 1968 } else { 1969 bp->b_flags &= ~B_WANTED; 1970 wakeup(bp); 1971 } 1972 splx(s); 1973} 1974 1975int 1976count_lock_queue() 1977{ 1978 int count; 1979 struct buf *bp; 1980 1981 count = 0; 1982 for (bp = TAILQ_FIRST(&bufqueues[QUEUE_LOCKED]); 1983 bp != NULL; 1984 bp = TAILQ_NEXT(bp, b_freelist)) 1985 count++; 1986 return (count); 1987} 1988 1989int vfs_update_interval = 30; 1990 1991static void 1992vfs_update() 1993{ 1994 while (1) { 1995 tsleep(&vfs_update_wakeup, PUSER, "update", 1996 hz * vfs_update_interval); 1997 vfs_update_wakeup = 0; 1998 sync(curproc, NULL); 1999 } 2000} 2001 2002static int 2003sysctl_kern_updateinterval SYSCTL_HANDLER_ARGS 2004{ 2005 int error = sysctl_handle_int(oidp, 2006 oidp->oid_arg1, oidp->oid_arg2, req); 2007 if (!error) 2008 wakeup(&vfs_update_wakeup); 2009 return error; 2010} 2011 2012SYSCTL_PROC(_kern, KERN_UPDATEINTERVAL, update, CTLTYPE_INT|CTLFLAG_RW, 2013 &vfs_update_interval, 0, sysctl_kern_updateinterval, "I", ""); 2014 2015 2016/* 2017 * This routine is called in lieu of iodone in the case of 2018 * incomplete I/O. This keeps the busy status for pages 2019 * consistant. 2020 */ 2021void 2022vfs_unbusy_pages(struct buf * bp) 2023{ 2024 int i; 2025 2026 if (bp->b_flags & B_VMIO) { 2027 struct vnode *vp = bp->b_vp; 2028 vm_object_t obj = vp->v_object; 2029 vm_ooffset_t foff; 2030 2031 foff = (vm_ooffset_t) vp->v_mount->mnt_stat.f_iosize * bp->b_lblkno; 2032 2033 for (i = 0; i < bp->b_npages; i++) { 2034 vm_page_t m = bp->b_pages[i]; 2035 2036 if (m == bogus_page) { 2037 m = vm_page_lookup(obj, OFF_TO_IDX(foff) + i); 2038#if !defined(MAX_PERF) 2039 if (!m) { 2040 panic("vfs_unbusy_pages: page missing\n"); 2041 } 2042#endif 2043 bp->b_pages[i] = m; 2044 pmap_qenter(trunc_page(bp->b_data), bp->b_pages, bp->b_npages); 2045 } 2046 --obj->paging_in_progress; 2047 --m->busy; 2048 if ((m->busy == 0) && (m->flags & PG_WANTED)) { 2049 m->flags &= ~PG_WANTED; 2050 wakeup(m); 2051 } 2052 } 2053 if (obj->paging_in_progress == 0 && 2054 (obj->flags & OBJ_PIPWNT)) { 2055 obj->flags &= ~OBJ_PIPWNT; 2056 wakeup(obj); 2057 } 2058 } 2059} 2060 2061/* 2062 * Set NFS' b_validoff and b_validend fields from the valid bits 2063 * of a page. If the consumer is not NFS, and the page is not 2064 * valid for the entire range, clear the B_CACHE flag to force 2065 * the consumer to re-read the page. 2066 */ 2067static void 2068vfs_buf_set_valid(struct buf *bp, 2069 vm_ooffset_t foff, vm_offset_t off, vm_offset_t size, 2070 vm_page_t m) 2071{ 2072 if (bp->b_vp->v_tag == VT_NFS && bp->b_vp->v_type != VBLK) { 2073 vm_offset_t svalid, evalid; 2074 int validbits = m->valid; 2075 2076 /* 2077 * This only bothers with the first valid range in the 2078 * page. 2079 */ 2080 svalid = off; 2081 while (validbits && !(validbits & 1)) { 2082 svalid += DEV_BSIZE; 2083 validbits >>= 1; 2084 } 2085 evalid = svalid; 2086 while (validbits & 1) { 2087 evalid += DEV_BSIZE; 2088 validbits >>= 1; 2089 } 2090 /* 2091 * Make sure this range is contiguous with the range 2092 * built up from previous pages. If not, then we will 2093 * just use the range from the previous pages. 2094 */ 2095 if (svalid == bp->b_validend) { 2096 bp->b_validoff = min(bp->b_validoff, svalid); 2097 bp->b_validend = max(bp->b_validend, evalid); 2098 } 2099 } else if (!vm_page_is_valid(m, 2100 (vm_offset_t) ((foff + off) & PAGE_MASK), 2101 size)) { 2102 bp->b_flags &= ~B_CACHE; 2103 } 2104} 2105 2106/* 2107 * Set the valid bits in a page, taking care of the b_validoff, 2108 * b_validend fields which NFS uses to optimise small reads. Off is 2109 * the offset within the file and pageno is the page index within the buf. 2110 */ 2111static void 2112vfs_page_set_valid(struct buf *bp, vm_ooffset_t off, int pageno, vm_page_t m) 2113{ 2114 struct vnode *vp = bp->b_vp; 2115 vm_ooffset_t soff, eoff; 2116 2117 soff = off; 2118 eoff = off + min(PAGE_SIZE, bp->b_bufsize); 2119 vm_page_set_invalid(m, 2120 (vm_offset_t) (soff & PAGE_MASK), 2121 (vm_offset_t) (eoff - soff)); 2122 if (vp->v_tag == VT_NFS && vp->v_type != VBLK) { 2123 vm_ooffset_t sv, ev; 2124 off = off - pageno * PAGE_SIZE; 2125 sv = off + ((bp->b_validoff + DEV_BSIZE - 1) & ~(DEV_BSIZE - 1)); 2126 ev = off + (bp->b_validend & ~(DEV_BSIZE - 1)); 2127 soff = max(sv, soff); 2128 eoff = min(ev, eoff); 2129 } 2130 if (eoff > soff) 2131 vm_page_set_validclean(m, 2132 (vm_offset_t) (soff & PAGE_MASK), 2133 (vm_offset_t) (eoff - soff)); 2134} 2135 2136/* 2137 * This routine is called before a device strategy routine. 2138 * It is used to tell the VM system that paging I/O is in 2139 * progress, and treat the pages associated with the buffer 2140 * almost as being PG_BUSY. Also the object paging_in_progress 2141 * flag is handled to make sure that the object doesn't become 2142 * inconsistant. 2143 */ 2144void 2145vfs_busy_pages(struct buf * bp, int clear_modify) 2146{ 2147 int i,s; 2148 2149 if (bp->b_flags & B_VMIO) { 2150 struct vnode *vp = bp->b_vp; 2151 vm_object_t obj = vp->v_object; 2152 vm_ooffset_t foff; 2153 2154 if (vp->v_type == VBLK) 2155 foff = (vm_ooffset_t) DEV_BSIZE * bp->b_lblkno; 2156 else 2157 foff = (vm_ooffset_t) vp->v_mount->mnt_stat.f_iosize * bp->b_lblkno; 2158 2159 vfs_setdirty(bp); 2160 2161retry: 2162 for (i = 0; i < bp->b_npages; i++) { 2163 vm_page_t m = bp->b_pages[i]; 2164 2165 if (m && (m->flags & PG_BUSY)) { 2166 s = splvm(); 2167 while (m->flags & PG_BUSY) { 2168 m->flags |= PG_WANTED; 2169 tsleep(m, PVM, "vbpage", 0); 2170 } 2171 splx(s); 2172 goto retry; 2173 } 2174 } 2175 2176 for (i = 0; i < bp->b_npages; i++, foff += PAGE_SIZE) { 2177 vm_page_t m = bp->b_pages[i]; 2178 2179 if ((bp->b_flags & B_CLUSTER) == 0) { 2180 obj->paging_in_progress++; 2181 m->busy++; 2182 } 2183 2184 vm_page_protect(m, VM_PROT_NONE); 2185 if (clear_modify) 2186 vfs_page_set_valid(bp, foff, i, m); 2187 else if (bp->b_bcount >= PAGE_SIZE) { 2188 if (m->valid && (bp->b_flags & B_CACHE) == 0) { 2189 bp->b_pages[i] = bogus_page; 2190 pmap_qenter(trunc_page(bp->b_data), bp->b_pages, bp->b_npages); 2191 } 2192 } 2193 } 2194 } 2195} 2196 2197/* 2198 * Tell the VM system that the pages associated with this buffer 2199 * are clean. This is used for delayed writes where the data is 2200 * going to go to disk eventually without additional VM intevention. 2201 */ 2202void 2203vfs_clean_pages(struct buf * bp) 2204{ 2205 int i; 2206 2207 if (bp->b_flags & B_VMIO) { 2208 struct vnode *vp = bp->b_vp; 2209 vm_ooffset_t foff; 2210 2211 if (vp->v_type == VBLK) 2212 foff = (vm_ooffset_t) DEV_BSIZE * bp->b_lblkno; 2213 else 2214 foff = (vm_ooffset_t) vp->v_mount->mnt_stat.f_iosize * bp->b_lblkno; 2215 for (i = 0; i < bp->b_npages; i++, foff += PAGE_SIZE) { 2216 vm_page_t m = bp->b_pages[i]; 2217 2218 vfs_page_set_valid(bp, foff, i, m); 2219 } 2220 } 2221} 2222 2223void 2224vfs_bio_clrbuf(struct buf *bp) { 2225 int i; 2226 if( bp->b_flags & B_VMIO) { 2227 if( (bp->b_npages == 1) && (bp->b_bufsize < PAGE_SIZE)) { 2228 int mask; 2229 mask = 0; 2230 for(i=0;i<bp->b_bufsize;i+=DEV_BSIZE) 2231 mask |= (1 << (i/DEV_BSIZE)); 2232 if( bp->b_pages[0]->valid != mask) { 2233 bzero(bp->b_data, bp->b_bufsize); 2234 } 2235 bp->b_pages[0]->valid = mask; 2236 bp->b_resid = 0; 2237 return; 2238 } 2239 for(i=0;i<bp->b_npages;i++) { 2240 if( bp->b_pages[i]->valid == VM_PAGE_BITS_ALL) 2241 continue; 2242 if( bp->b_pages[i]->valid == 0) { 2243 if ((bp->b_pages[i]->flags & PG_ZERO) == 0) { 2244 bzero(bp->b_data + (i << PAGE_SHIFT), PAGE_SIZE); 2245 } 2246 } else { 2247 int j; 2248 for(j=0;j<PAGE_SIZE/DEV_BSIZE;j++) { 2249 if( (bp->b_pages[i]->valid & (1<<j)) == 0) 2250 bzero(bp->b_data + (i << PAGE_SHIFT) + j * DEV_BSIZE, DEV_BSIZE); 2251 } 2252 } 2253 /* bp->b_pages[i]->valid = VM_PAGE_BITS_ALL; */ 2254 } 2255 bp->b_resid = 0; 2256 } else { 2257 clrbuf(bp); 2258 } 2259} 2260 2261/* 2262 * vm_hold_load_pages and vm_hold_unload pages get pages into 2263 * a buffers address space. The pages are anonymous and are 2264 * not associated with a file object. 2265 */ 2266void 2267vm_hold_load_pages(struct buf * bp, vm_offset_t from, vm_offset_t to) 2268{ 2269 vm_offset_t pg; 2270 vm_page_t p; 2271 int index; 2272 2273 to = round_page(to); 2274 from = round_page(from); 2275 index = (from - trunc_page(bp->b_data)) >> PAGE_SHIFT; 2276 2277 for (pg = from; pg < to; pg += PAGE_SIZE, index++) { 2278 2279tryagain: 2280 2281 p = vm_page_alloc(kernel_object, 2282 ((pg - VM_MIN_KERNEL_ADDRESS) >> PAGE_SHIFT), 2283 VM_ALLOC_NORMAL); 2284 if (!p) { 2285 vm_pageout_deficit += (to - from) >> PAGE_SHIFT; 2286 VM_WAIT; 2287 goto tryagain; 2288 } 2289 vm_page_wire(p); 2290 pmap_kenter(pg, VM_PAGE_TO_PHYS(p)); 2291 bp->b_pages[index] = p; 2292 PAGE_WAKEUP(p); 2293 } 2294 bp->b_npages = index; 2295} 2296 2297void 2298vm_hold_free_pages(struct buf * bp, vm_offset_t from, vm_offset_t to) 2299{ 2300 vm_offset_t pg; 2301 vm_page_t p; 2302 int index, newnpages; 2303 2304 from = round_page(from); 2305 to = round_page(to); 2306 newnpages = index = (from - trunc_page(bp->b_data)) >> PAGE_SHIFT; 2307 2308 for (pg = from; pg < to; pg += PAGE_SIZE, index++) { 2309 p = bp->b_pages[index]; 2310 if (p && (index < bp->b_npages)) { 2311#if !defined(MAX_PERF) 2312 if (p->busy) { 2313 printf("vm_hold_free_pages: blkno: %d, lblkno: %d\n", 2314 bp->b_blkno, bp->b_lblkno); 2315 } 2316#endif 2317 bp->b_pages[index] = NULL; 2318 pmap_kremove(pg); 2319 p->flags |= PG_BUSY; 2320 vm_page_unwire(p); 2321 vm_page_free(p); 2322 } 2323 } 2324 bp->b_npages = newnpages; 2325} 2326 2327 2328#include "opt_ddb.h" 2329#ifdef DDB 2330#include <ddb/ddb.h> 2331 2332DB_SHOW_COMMAND(buffer, db_show_buffer) 2333{ 2334 /* get args */ 2335 struct buf *bp = (struct buf *)addr; 2336 2337 if (!have_addr) { 2338 db_printf("usage: show buffer <addr>\n"); 2339 return; 2340 } 2341 2342 db_printf("b_proc = %p,\nb_flags = 0x%b\n", (void *)bp->b_proc, 2343 bp->b_flags, "\20\40bounce\37cluster\36vmio\35ram\34ordered" 2344 "\33paging\32xxx\31writeinprog\30wanted\27relbuf\26tape" 2345 "\25read\24raw\23phys\22clusterok\21malloc\20nocache" 2346 "\17locked\16inval\15gathered\14error\13eintr\12done\11dirty" 2347 "\10delwri\7call\6cache\5busy\4bad\3async\2needcommit\1age"); 2348 db_printf("b_error = %d, b_bufsize = %ld, b_bcount = %ld, " 2349 "b_resid = %ld\nb_dev = 0x%x, b_data = %p, " 2350 "b_blkno = %d, b_pblkno = %d\n", 2351 bp->b_error, bp->b_bufsize, bp->b_bcount, bp->b_resid, 2352 bp->b_dev, bp->b_data, bp->b_blkno, bp->b_pblkno); 2353 if (bp->b_npages) { 2354 int i; 2355 db_printf("b_npages = %d, pages(OBJ, IDX, PA): ", bp->b_npages); 2356 for (i = 0; i < bp->b_npages; i++) { 2357 vm_page_t m; 2358 m = bp->b_pages[i]; 2359 db_printf("(0x%x, 0x%x, 0x%x)", m->object, m->pindex, 2360 VM_PAGE_TO_PHYS(m)); 2361 if ((i + 1) < bp->b_npages) 2362 db_printf(","); 2363 } 2364 db_printf("\n"); 2365 } 2366} 2367#endif /* DDB */ 2368