vfs_bio.c revision 31380
1/* 2 * Copyright (c) 1994 John S. Dyson 3 * All rights reserved. 4 * 5 * Redistribution and use in source and binary forms, with or without 6 * modification, are permitted provided that the following conditions 7 * are met: 8 * 1. Redistributions of source code must retain the above copyright 9 * notice immediately at the beginning of the file, without modification, 10 * this list of conditions, and the following disclaimer. 11 * 2. Redistributions in binary form must reproduce the above copyright 12 * notice, this list of conditions and the following disclaimer in the 13 * documentation and/or other materials provided with the distribution. 14 * 3. Absolutely no warranty of function or purpose is made by the author 15 * John S. Dyson. 16 * 4. This work was done expressly for inclusion into FreeBSD. Other use 17 * is allowed if this notation is included. 18 * 5. Modifications may be freely made to this file if the above conditions 19 * are met. 20 * 21 * $Id: vfs_bio.c,v 1.134 1997/11/07 08:53:04 phk Exp $ 22 */ 23 24/* 25 * this file contains a new buffer I/O scheme implementing a coherent 26 * VM object and buffer cache scheme. Pains have been taken to make 27 * sure that the performance degradation associated with schemes such 28 * as this is not realized. 29 * 30 * Author: John S. Dyson 31 * Significant help during the development and debugging phases 32 * had been provided by David Greenman, also of the FreeBSD core team. 33 */ 34 35#include "opt_bounce.h" 36 37#define VMIO 38#include <sys/param.h> 39#include <sys/systm.h> 40#include <sys/sysproto.h> 41#include <sys/kernel.h> 42#include <sys/sysctl.h> 43#include <sys/proc.h> 44#include <sys/vnode.h> 45#include <sys/vmmeter.h> 46#include <vm/vm.h> 47#include <vm/vm_param.h> 48#include <vm/vm_prot.h> 49#include <vm/vm_kern.h> 50#include <vm/vm_pageout.h> 51#include <vm/vm_page.h> 52#include <vm/vm_object.h> 53#include <vm/vm_extern.h> 54#include <vm/vm_map.h> 55#include <sys/buf.h> 56#include <sys/mount.h> 57#include <sys/malloc.h> 58#include <sys/resourcevar.h> 59 60static MALLOC_DEFINE(M_BIOBUF, "BIO buffer", "BIO buffer"); 61 62static void vfs_update __P((void)); 63static struct proc *updateproc; 64static struct kproc_desc up_kp = { 65 "update", 66 vfs_update, 67 &updateproc 68}; 69SYSINIT_KT(update, SI_SUB_KTHREAD_UPDATE, SI_ORDER_FIRST, kproc_start, &up_kp) 70 71struct buf *buf; /* buffer header pool */ 72struct swqueue bswlist; 73 74int count_lock_queue __P((void)); 75static void vm_hold_free_pages(struct buf * bp, vm_offset_t from, 76 vm_offset_t to); 77static void vm_hold_load_pages(struct buf * bp, vm_offset_t from, 78 vm_offset_t to); 79static void vfs_buf_set_valid(struct buf *bp, vm_ooffset_t foff, 80 vm_offset_t off, vm_offset_t size, 81 vm_page_t m); 82static void vfs_page_set_valid(struct buf *bp, vm_ooffset_t off, 83 int pageno, vm_page_t m); 84static void vfs_clean_pages(struct buf * bp); 85static void vfs_setdirty(struct buf *bp); 86static void vfs_vmio_release(struct buf *bp); 87static void flushdirtybuffers(int slpflag, int slptimeo); 88 89int needsbuffer; 90 91/* 92 * Internal update daemon, process 3 93 * The variable vfs_update_wakeup allows for internal syncs. 94 */ 95int vfs_update_wakeup; 96 97 98/* 99 * buffers base kva 100 */ 101 102/* 103 * bogus page -- for I/O to/from partially complete buffers 104 * this is a temporary solution to the problem, but it is not 105 * really that bad. it would be better to split the buffer 106 * for input in the case of buffers partially already in memory, 107 * but the code is intricate enough already. 108 */ 109vm_page_t bogus_page; 110static vm_offset_t bogus_offset; 111 112static int bufspace, maxbufspace, vmiospace, maxvmiobufspace, 113 bufmallocspace, maxbufmallocspace; 114int numdirtybuffers, lodirtybuffers, hidirtybuffers; 115static int numfreebuffers, lofreebuffers, hifreebuffers; 116 117SYSCTL_INT(_vfs, OID_AUTO, numdirtybuffers, CTLFLAG_RD, 118 &numdirtybuffers, 0, ""); 119SYSCTL_INT(_vfs, OID_AUTO, lodirtybuffers, CTLFLAG_RW, 120 &lodirtybuffers, 0, ""); 121SYSCTL_INT(_vfs, OID_AUTO, hidirtybuffers, CTLFLAG_RW, 122 &hidirtybuffers, 0, ""); 123SYSCTL_INT(_vfs, OID_AUTO, numfreebuffers, CTLFLAG_RD, 124 &numfreebuffers, 0, ""); 125SYSCTL_INT(_vfs, OID_AUTO, lofreebuffers, CTLFLAG_RW, 126 &lofreebuffers, 0, ""); 127SYSCTL_INT(_vfs, OID_AUTO, hifreebuffers, CTLFLAG_RW, 128 &hifreebuffers, 0, ""); 129SYSCTL_INT(_vfs, OID_AUTO, maxbufspace, CTLFLAG_RW, 130 &maxbufspace, 0, ""); 131SYSCTL_INT(_vfs, OID_AUTO, bufspace, CTLFLAG_RD, 132 &bufspace, 0, ""); 133SYSCTL_INT(_vfs, OID_AUTO, maxvmiobufspace, CTLFLAG_RW, 134 &maxvmiobufspace, 0, ""); 135SYSCTL_INT(_vfs, OID_AUTO, vmiospace, CTLFLAG_RD, 136 &vmiospace, 0, ""); 137SYSCTL_INT(_vfs, OID_AUTO, maxmallocbufspace, CTLFLAG_RW, 138 &maxbufmallocspace, 0, ""); 139SYSCTL_INT(_vfs, OID_AUTO, bufmallocspace, CTLFLAG_RD, 140 &bufmallocspace, 0, ""); 141 142static LIST_HEAD(bufhashhdr, buf) bufhashtbl[BUFHSZ], invalhash; 143static TAILQ_HEAD(bqueues, buf) bufqueues[BUFFER_QUEUES]; 144 145extern int vm_swap_size; 146 147#define BUF_MAXUSE 24 148 149#define VFS_BIO_NEED_ANY 1 150#define VFS_BIO_NEED_LOWLIMIT 2 151#define VFS_BIO_NEED_FREE 4 152 153/* 154 * Initialize buffer headers and related structures. 155 */ 156void 157bufinit() 158{ 159 struct buf *bp; 160 int i; 161 162 TAILQ_INIT(&bswlist); 163 LIST_INIT(&invalhash); 164 165 /* first, make a null hash table */ 166 for (i = 0; i < BUFHSZ; i++) 167 LIST_INIT(&bufhashtbl[i]); 168 169 /* next, make a null set of free lists */ 170 for (i = 0; i < BUFFER_QUEUES; i++) 171 TAILQ_INIT(&bufqueues[i]); 172 173 /* finally, initialize each buffer header and stick on empty q */ 174 for (i = 0; i < nbuf; i++) { 175 bp = &buf[i]; 176 bzero(bp, sizeof *bp); 177 bp->b_flags = B_INVAL; /* we're just an empty header */ 178 bp->b_dev = NODEV; 179 bp->b_rcred = NOCRED; 180 bp->b_wcred = NOCRED; 181 bp->b_qindex = QUEUE_EMPTY; 182 bp->b_vnbufs.le_next = NOLIST; 183 TAILQ_INSERT_TAIL(&bufqueues[QUEUE_EMPTY], bp, b_freelist); 184 LIST_INSERT_HEAD(&invalhash, bp, b_hash); 185 } 186/* 187 * maxbufspace is currently calculated to support all filesystem blocks 188 * to be 8K. If you happen to use a 16K filesystem, the size of the buffer 189 * cache is still the same as it would be for 8K filesystems. This 190 * keeps the size of the buffer cache "in check" for big block filesystems. 191 */ 192 maxbufspace = (nbuf + 8) * DFLTBSIZE; 193/* 194 * reserve 1/3 of the buffers for metadata (VDIR) which might not be VMIO'ed 195 */ 196 maxvmiobufspace = 2 * maxbufspace / 3; 197/* 198 * Limit the amount of malloc memory since it is wired permanently into 199 * the kernel space. Even though this is accounted for in the buffer 200 * allocation, we don't want the malloced region to grow uncontrolled. 201 * The malloc scheme improves memory utilization significantly on average 202 * (small) directories. 203 */ 204 maxbufmallocspace = maxbufspace / 20; 205 206/* 207 * Remove the probability of deadlock conditions by limiting the 208 * number of dirty buffers. 209 */ 210 hidirtybuffers = nbuf / 6 + 20; 211 lodirtybuffers = nbuf / 12 + 10; 212 numdirtybuffers = 0; 213 lofreebuffers = nbuf / 18 + 5; 214 hifreebuffers = 2 * lofreebuffers; 215 numfreebuffers = nbuf; 216 217 bogus_offset = kmem_alloc_pageable(kernel_map, PAGE_SIZE); 218 bogus_page = vm_page_alloc(kernel_object, 219 ((bogus_offset - VM_MIN_KERNEL_ADDRESS) >> PAGE_SHIFT), 220 VM_ALLOC_NORMAL); 221 222} 223 224/* 225 * Free the kva allocation for a buffer 226 * Must be called only at splbio or higher, 227 * as this is the only locking for buffer_map. 228 */ 229static void 230bfreekva(struct buf * bp) 231{ 232 if (bp->b_kvasize == 0) 233 return; 234 235 vm_map_delete(buffer_map, 236 (vm_offset_t) bp->b_kvabase, 237 (vm_offset_t) bp->b_kvabase + bp->b_kvasize); 238 239 bp->b_kvasize = 0; 240 241} 242 243/* 244 * remove the buffer from the appropriate free list 245 */ 246void 247bremfree(struct buf * bp) 248{ 249 int s = splbio(); 250 251 if (bp->b_qindex != QUEUE_NONE) { 252 TAILQ_REMOVE(&bufqueues[bp->b_qindex], bp, b_freelist); 253 bp->b_qindex = QUEUE_NONE; 254 } else { 255#if !defined(MAX_PERF) 256 panic("bremfree: removing a buffer when not on a queue"); 257#endif 258 } 259 if ((bp->b_flags & B_INVAL) || 260 (bp->b_flags & (B_DELWRI|B_LOCKED)) == 0) 261 --numfreebuffers; 262 splx(s); 263} 264 265/* 266 * Get a buffer with the specified data. Look in the cache first. 267 */ 268int 269bread(struct vnode * vp, daddr_t blkno, int size, struct ucred * cred, 270 struct buf ** bpp) 271{ 272 struct buf *bp; 273 274 bp = getblk(vp, blkno, size, 0, 0); 275 *bpp = bp; 276 277 /* if not found in cache, do some I/O */ 278 if ((bp->b_flags & B_CACHE) == 0) { 279 if (curproc != NULL) 280 curproc->p_stats->p_ru.ru_inblock++; 281 bp->b_flags |= B_READ; 282 bp->b_flags &= ~(B_DONE | B_ERROR | B_INVAL); 283 if (bp->b_rcred == NOCRED) { 284 if (cred != NOCRED) 285 crhold(cred); 286 bp->b_rcred = cred; 287 } 288 vfs_busy_pages(bp, 0); 289 VOP_STRATEGY(bp); 290 return (biowait(bp)); 291 } 292 return (0); 293} 294 295/* 296 * Operates like bread, but also starts asynchronous I/O on 297 * read-ahead blocks. 298 */ 299int 300breadn(struct vnode * vp, daddr_t blkno, int size, 301 daddr_t * rablkno, int *rabsize, 302 int cnt, struct ucred * cred, struct buf ** bpp) 303{ 304 struct buf *bp, *rabp; 305 int i; 306 int rv = 0, readwait = 0; 307 308 *bpp = bp = getblk(vp, blkno, size, 0, 0); 309 310 /* if not found in cache, do some I/O */ 311 if ((bp->b_flags & B_CACHE) == 0) { 312 if (curproc != NULL) 313 curproc->p_stats->p_ru.ru_inblock++; 314 bp->b_flags |= B_READ; 315 bp->b_flags &= ~(B_DONE | B_ERROR | B_INVAL); 316 if (bp->b_rcred == NOCRED) { 317 if (cred != NOCRED) 318 crhold(cred); 319 bp->b_rcred = cred; 320 } 321 vfs_busy_pages(bp, 0); 322 VOP_STRATEGY(bp); 323 ++readwait; 324 } 325 for (i = 0; i < cnt; i++, rablkno++, rabsize++) { 326 if (inmem(vp, *rablkno)) 327 continue; 328 rabp = getblk(vp, *rablkno, *rabsize, 0, 0); 329 330 if ((rabp->b_flags & B_CACHE) == 0) { 331 if (curproc != NULL) 332 curproc->p_stats->p_ru.ru_inblock++; 333 rabp->b_flags |= B_READ | B_ASYNC; 334 rabp->b_flags &= ~(B_DONE | B_ERROR | B_INVAL); 335 if (rabp->b_rcred == NOCRED) { 336 if (cred != NOCRED) 337 crhold(cred); 338 rabp->b_rcred = cred; 339 } 340 vfs_busy_pages(rabp, 0); 341 VOP_STRATEGY(rabp); 342 } else { 343 brelse(rabp); 344 } 345 } 346 347 if (readwait) { 348 rv = biowait(bp); 349 } 350 return (rv); 351} 352 353/* 354 * Write, release buffer on completion. (Done by iodone 355 * if async.) 356 */ 357int 358bwrite(struct buf * bp) 359{ 360 int oldflags = bp->b_flags; 361 362 if (bp->b_flags & B_INVAL) { 363 brelse(bp); 364 return (0); 365 } 366#if !defined(MAX_PERF) 367 if (!(bp->b_flags & B_BUSY)) 368 panic("bwrite: buffer is not busy???"); 369#endif 370 371 bp->b_flags &= ~(B_READ | B_DONE | B_ERROR | B_DELWRI); 372 bp->b_flags |= B_WRITEINPROG; 373 374 if ((oldflags & B_DELWRI) == B_DELWRI) { 375 --numdirtybuffers; 376 reassignbuf(bp, bp->b_vp); 377 } 378 379 bp->b_vp->v_numoutput++; 380 vfs_busy_pages(bp, 1); 381 if (curproc != NULL) 382 curproc->p_stats->p_ru.ru_oublock++; 383 VOP_STRATEGY(bp); 384 385 if ((oldflags & B_ASYNC) == 0) { 386 int rtval = biowait(bp); 387 388 if (oldflags & B_DELWRI) { 389 reassignbuf(bp, bp->b_vp); 390 } 391 brelse(bp); 392 return (rtval); 393 } 394 return (0); 395} 396 397void 398vfs_bio_need_satisfy(void) { 399 ++numfreebuffers; 400 if (!needsbuffer) 401 return; 402 if (numdirtybuffers < lodirtybuffers) { 403 needsbuffer &= ~(VFS_BIO_NEED_ANY | VFS_BIO_NEED_LOWLIMIT); 404 } else { 405 needsbuffer &= ~VFS_BIO_NEED_ANY; 406 } 407 if (numfreebuffers >= hifreebuffers) { 408 needsbuffer &= ~VFS_BIO_NEED_FREE; 409 } 410 wakeup(&needsbuffer); 411} 412 413/* 414 * Delayed write. (Buffer is marked dirty). 415 */ 416void 417bdwrite(struct buf * bp) 418{ 419 420#if !defined(MAX_PERF) 421 if ((bp->b_flags & B_BUSY) == 0) { 422 panic("bdwrite: buffer is not busy"); 423 } 424#endif 425 426 if (bp->b_flags & B_INVAL) { 427 brelse(bp); 428 return; 429 } 430 if (bp->b_flags & B_TAPE) { 431 bawrite(bp); 432 return; 433 } 434 bp->b_flags &= ~(B_READ|B_RELBUF); 435 if ((bp->b_flags & B_DELWRI) == 0) { 436 bp->b_flags |= B_DONE | B_DELWRI; 437 reassignbuf(bp, bp->b_vp); 438 ++numdirtybuffers; 439 } 440 441 /* 442 * This bmap keeps the system from needing to do the bmap later, 443 * perhaps when the system is attempting to do a sync. Since it 444 * is likely that the indirect block -- or whatever other datastructure 445 * that the filesystem needs is still in memory now, it is a good 446 * thing to do this. Note also, that if the pageout daemon is 447 * requesting a sync -- there might not be enough memory to do 448 * the bmap then... So, this is important to do. 449 */ 450 if (bp->b_lblkno == bp->b_blkno) { 451 VOP_BMAP(bp->b_vp, bp->b_lblkno, NULL, &bp->b_blkno, NULL, NULL); 452 } 453 454 /* 455 * Set the *dirty* buffer range based upon the VM system dirty pages. 456 */ 457 vfs_setdirty(bp); 458 459 /* 460 * We need to do this here to satisfy the vnode_pager and the 461 * pageout daemon, so that it thinks that the pages have been 462 * "cleaned". Note that since the pages are in a delayed write 463 * buffer -- the VFS layer "will" see that the pages get written 464 * out on the next sync, or perhaps the cluster will be completed. 465 */ 466 vfs_clean_pages(bp); 467 bqrelse(bp); 468 469 if (numdirtybuffers >= hidirtybuffers) 470 flushdirtybuffers(0, 0); 471 472 return; 473} 474 475/* 476 * Asynchronous write. 477 * Start output on a buffer, but do not wait for it to complete. 478 * The buffer is released when the output completes. 479 */ 480void 481bawrite(struct buf * bp) 482{ 483 bp->b_flags |= B_ASYNC; 484 (void) VOP_BWRITE(bp); 485} 486 487/* 488 * Ordered write. 489 * Start output on a buffer, but only wait for it to complete if the 490 * output device cannot guarantee ordering in some other way. Devices 491 * that can perform asynchronous ordered writes will set the B_ASYNC 492 * flag in their strategy routine. 493 * The buffer is released when the output completes. 494 */ 495int 496bowrite(struct buf * bp) 497{ 498 /* 499 * XXX Add in B_ASYNC once the SCSI 500 * layer can deal with ordered 501 * writes properly. 502 */ 503 bp->b_flags |= B_ORDERED; 504 return (VOP_BWRITE(bp)); 505} 506 507/* 508 * Release a buffer. 509 */ 510void 511brelse(struct buf * bp) 512{ 513 int s; 514 515 if (bp->b_flags & B_CLUSTER) { 516 relpbuf(bp); 517 return; 518 } 519 /* anyone need a "free" block? */ 520 s = splbio(); 521 522 /* anyone need this block? */ 523 if (bp->b_flags & B_WANTED) { 524 bp->b_flags &= ~(B_WANTED | B_AGE); 525 wakeup(bp); 526 } 527 528 if (bp->b_flags & B_LOCKED) 529 bp->b_flags &= ~B_ERROR; 530 531 if ((bp->b_flags & (B_NOCACHE | B_INVAL | B_ERROR)) || 532 (bp->b_bufsize <= 0)) { 533 bp->b_flags |= B_INVAL; 534 if (bp->b_flags & B_DELWRI) 535 --numdirtybuffers; 536 bp->b_flags &= ~(B_DELWRI | B_CACHE); 537 if (((bp->b_flags & B_VMIO) == 0) && bp->b_vp) { 538 if (bp->b_bufsize) 539 allocbuf(bp, 0); 540 brelvp(bp); 541 } 542 } 543 544 /* 545 * VMIO buffer rundown. It is not very necessary to keep a VMIO buffer 546 * constituted, so the B_INVAL flag is used to *invalidate* the buffer, 547 * but the VM object is kept around. The B_NOCACHE flag is used to 548 * invalidate the pages in the VM object. 549 * 550 * If the buffer is a partially filled NFS buffer, keep it 551 * since invalidating it now will lose informatio. The valid 552 * flags in the vm_pages have only DEV_BSIZE resolution but 553 * the b_validoff, b_validend fields have byte resolution. 554 * This can avoid unnecessary re-reads of the buffer. 555 * XXX this seems to cause performance problems. 556 */ 557 if ((bp->b_flags & B_VMIO) 558 && !(bp->b_vp->v_tag == VT_NFS && 559 bp->b_vp->v_type != VBLK && 560 (bp->b_flags & B_DELWRI) != 0) 561#ifdef notdef 562 && (bp->b_vp->v_tag != VT_NFS 563 || bp->b_vp->v_type == VBLK 564 || (bp->b_flags & (B_NOCACHE | B_INVAL | B_ERROR)) 565 || bp->b_validend == 0 566 || (bp->b_validoff == 0 567 && bp->b_validend == bp->b_bufsize)) 568#endif 569 ) { 570 vm_ooffset_t foff; 571 vm_object_t obj; 572 int i, resid; 573 vm_page_t m; 574 struct vnode *vp; 575 int iototal = bp->b_bufsize; 576 577 vp = bp->b_vp; 578 579#if !defined(MAX_PERF) 580 if (!vp) 581 panic("brelse: missing vp"); 582#endif 583 584 if (bp->b_npages) { 585 vm_pindex_t poff; 586 obj = (vm_object_t) vp->v_object; 587 if (vp->v_type == VBLK) 588 foff = ((vm_ooffset_t) bp->b_lblkno) << DEV_BSHIFT; 589 else 590 foff = (vm_ooffset_t) vp->v_mount->mnt_stat.f_iosize * bp->b_lblkno; 591 poff = OFF_TO_IDX(foff); 592 for (i = 0; i < bp->b_npages; i++) { 593 m = bp->b_pages[i]; 594 if (m == bogus_page) { 595 m = vm_page_lookup(obj, poff + i); 596#if !defined(MAX_PERF) 597 if (!m) { 598 panic("brelse: page missing\n"); 599 } 600#endif 601 bp->b_pages[i] = m; 602 pmap_qenter(trunc_page(bp->b_data), 603 bp->b_pages, bp->b_npages); 604 } 605 resid = IDX_TO_OFF(m->pindex+1) - foff; 606 if (resid > iototal) 607 resid = iototal; 608 if (resid > 0) { 609 /* 610 * Don't invalidate the page if the local machine has already 611 * modified it. This is the lesser of two evils, and should 612 * be fixed. 613 */ 614 if (bp->b_flags & (B_NOCACHE | B_ERROR)) { 615 vm_page_test_dirty(m); 616 if (m->dirty == 0) { 617 vm_page_set_invalid(m, (vm_offset_t) foff, resid); 618 if (m->valid == 0) 619 vm_page_protect(m, VM_PROT_NONE); 620 } 621 } 622 if (resid >= PAGE_SIZE) { 623 if ((m->valid & VM_PAGE_BITS_ALL) != VM_PAGE_BITS_ALL) { 624 bp->b_flags |= B_INVAL; 625 } 626 } else { 627 if (!vm_page_is_valid(m, 628 (((vm_offset_t) bp->b_data) & PAGE_MASK), resid)) { 629 bp->b_flags |= B_INVAL; 630 } 631 } 632 } 633 foff += resid; 634 iototal -= resid; 635 } 636 } 637 if (bp->b_flags & (B_INVAL | B_RELBUF)) 638 vfs_vmio_release(bp); 639 } 640#if !defined(MAX_PERF) 641 if (bp->b_qindex != QUEUE_NONE) 642 panic("brelse: free buffer onto another queue???"); 643#endif 644 645 /* enqueue */ 646 /* buffers with no memory */ 647 if (bp->b_bufsize == 0) { 648 bp->b_flags |= B_INVAL; 649 bp->b_qindex = QUEUE_EMPTY; 650 TAILQ_INSERT_HEAD(&bufqueues[QUEUE_EMPTY], bp, b_freelist); 651 LIST_REMOVE(bp, b_hash); 652 LIST_INSERT_HEAD(&invalhash, bp, b_hash); 653 bp->b_dev = NODEV; 654 655 /* buffers with junk contents */ 656 } else if (bp->b_flags & (B_ERROR | B_INVAL | B_NOCACHE | B_RELBUF)) { 657 bp->b_flags |= B_INVAL; 658 bp->b_qindex = QUEUE_AGE; 659 TAILQ_INSERT_HEAD(&bufqueues[QUEUE_AGE], bp, b_freelist); 660 LIST_REMOVE(bp, b_hash); 661 LIST_INSERT_HEAD(&invalhash, bp, b_hash); 662 bp->b_dev = NODEV; 663 664 /* buffers that are locked */ 665 } else if (bp->b_flags & B_LOCKED) { 666 bp->b_qindex = QUEUE_LOCKED; 667 TAILQ_INSERT_TAIL(&bufqueues[QUEUE_LOCKED], bp, b_freelist); 668 669 /* buffers with stale but valid contents */ 670 } else if (bp->b_flags & B_AGE) { 671 bp->b_qindex = QUEUE_AGE; 672 TAILQ_INSERT_TAIL(&bufqueues[QUEUE_AGE], bp, b_freelist); 673 674 /* buffers with valid and quite potentially reuseable contents */ 675 } else { 676 bp->b_qindex = QUEUE_LRU; 677 TAILQ_INSERT_TAIL(&bufqueues[QUEUE_LRU], bp, b_freelist); 678 } 679 680 if ((bp->b_flags & B_INVAL) || 681 (bp->b_flags & (B_LOCKED|B_DELWRI)) == 0) { 682 if (bp->b_flags & B_DELWRI) { 683 --numdirtybuffers; 684 bp->b_flags &= ~B_DELWRI; 685 } 686 vfs_bio_need_satisfy(); 687 } 688 689 /* unlock */ 690 bp->b_flags &= ~(B_ORDERED | B_WANTED | B_BUSY | 691 B_ASYNC | B_NOCACHE | B_AGE | B_RELBUF); 692 splx(s); 693} 694 695/* 696 * Release a buffer. 697 */ 698void 699bqrelse(struct buf * bp) 700{ 701 int s; 702 703 s = splbio(); 704 705 /* anyone need this block? */ 706 if (bp->b_flags & B_WANTED) { 707 bp->b_flags &= ~(B_WANTED | B_AGE); 708 wakeup(bp); 709 } 710 711#if !defined(MAX_PERF) 712 if (bp->b_qindex != QUEUE_NONE) 713 panic("bqrelse: free buffer onto another queue???"); 714#endif 715 716 if (bp->b_flags & B_LOCKED) { 717 bp->b_flags &= ~B_ERROR; 718 bp->b_qindex = QUEUE_LOCKED; 719 TAILQ_INSERT_TAIL(&bufqueues[QUEUE_LOCKED], bp, b_freelist); 720 /* buffers with stale but valid contents */ 721 } else { 722 bp->b_qindex = QUEUE_LRU; 723 TAILQ_INSERT_TAIL(&bufqueues[QUEUE_LRU], bp, b_freelist); 724 } 725 726 if ((bp->b_flags & (B_LOCKED|B_DELWRI)) == 0) { 727 vfs_bio_need_satisfy(); 728 } 729 730 /* unlock */ 731 bp->b_flags &= ~(B_ORDERED | B_WANTED | B_BUSY | 732 B_ASYNC | B_NOCACHE | B_AGE | B_RELBUF); 733 splx(s); 734} 735 736static void 737vfs_vmio_release(bp) 738 struct buf *bp; 739{ 740 int i; 741 vm_page_t m; 742 743 for (i = 0; i < bp->b_npages; i++) { 744 m = bp->b_pages[i]; 745 bp->b_pages[i] = NULL; 746 vm_page_unwire(m); 747 /* 748 * We don't mess with busy pages, it is 749 * the responsibility of the process that 750 * busied the pages to deal with them. 751 */ 752 if ((m->flags & PG_BUSY) || (m->busy != 0)) 753 continue; 754 755 if (m->wire_count == 0) { 756 757 if (m->flags & PG_WANTED) { 758 m->flags &= ~PG_WANTED; 759 wakeup(m); 760 } 761 762 /* 763 * If this is an async free -- we cannot place 764 * pages onto the cache queue. If it is an 765 * async free, then we don't modify any queues. 766 * This is probably in error (for perf reasons), 767 * and we will eventually need to build 768 * a more complete infrastructure to support I/O 769 * rundown. 770 */ 771 if ((bp->b_flags & B_ASYNC) == 0) { 772 773 /* 774 * In the case of sync buffer frees, we can do pretty much 775 * anything to any of the memory queues. Specifically, 776 * the cache queue is okay to be modified. 777 */ 778 if (m->valid) { 779 if(m->dirty == 0) 780 vm_page_test_dirty(m); 781 /* 782 * this keeps pressure off of the process memory 783 */ 784 if (m->dirty == 0 && m->hold_count == 0) 785 vm_page_cache(m); 786 else 787 vm_page_deactivate(m); 788 } else if (m->hold_count == 0) { 789 vm_page_protect(m, VM_PROT_NONE); 790 vm_page_free(m); 791 } 792 } else { 793 /* 794 * If async, then at least we clear the 795 * act_count. 796 */ 797 m->act_count = 0; 798 } 799 } 800 } 801 bufspace -= bp->b_bufsize; 802 vmiospace -= bp->b_bufsize; 803 pmap_qremove(trunc_page((vm_offset_t) bp->b_data), bp->b_npages); 804 bp->b_npages = 0; 805 bp->b_bufsize = 0; 806 bp->b_flags &= ~B_VMIO; 807 if (bp->b_vp) 808 brelvp(bp); 809} 810 811/* 812 * Check to see if a block is currently memory resident. 813 */ 814struct buf * 815gbincore(struct vnode * vp, daddr_t blkno) 816{ 817 struct buf *bp; 818 struct bufhashhdr *bh; 819 820 bh = BUFHASH(vp, blkno); 821 bp = bh->lh_first; 822 823 /* Search hash chain */ 824 while (bp != NULL) { 825 /* hit */ 826 if (bp->b_vp == vp && bp->b_lblkno == blkno && 827 (bp->b_flags & B_INVAL) == 0) { 828 break; 829 } 830 bp = bp->b_hash.le_next; 831 } 832 return (bp); 833} 834 835/* 836 * this routine implements clustered async writes for 837 * clearing out B_DELWRI buffers... This is much better 838 * than the old way of writing only one buffer at a time. 839 */ 840int 841vfs_bio_awrite(struct buf * bp) 842{ 843 int i; 844 daddr_t lblkno = bp->b_lblkno; 845 struct vnode *vp = bp->b_vp; 846 int s; 847 int ncl; 848 struct buf *bpa; 849 int nwritten; 850 851 s = splbio(); 852 /* 853 * right now we support clustered writing only to regular files 854 */ 855 if ((vp->v_type == VREG) && 856 (vp->v_mount != 0) && /* Only on nodes that have the size info */ 857 (bp->b_flags & (B_CLUSTEROK | B_INVAL)) == B_CLUSTEROK) { 858 int size; 859 int maxcl; 860 861 size = vp->v_mount->mnt_stat.f_iosize; 862 maxcl = MAXPHYS / size; 863 864 for (i = 1; i < maxcl; i++) { 865 if ((bpa = gbincore(vp, lblkno + i)) && 866 ((bpa->b_flags & (B_BUSY | B_DELWRI | B_CLUSTEROK | B_INVAL)) == 867 (B_DELWRI | B_CLUSTEROK)) && 868 (bpa->b_bufsize == size)) { 869 if ((bpa->b_blkno == bpa->b_lblkno) || 870 (bpa->b_blkno != bp->b_blkno + ((i * size) >> DEV_BSHIFT))) 871 break; 872 } else { 873 break; 874 } 875 } 876 ncl = i; 877 /* 878 * this is a possible cluster write 879 */ 880 if (ncl != 1) { 881 nwritten = cluster_wbuild(vp, size, lblkno, ncl); 882 splx(s); 883 return nwritten; 884 } 885 } 886 bremfree(bp); 887 splx(s); 888 /* 889 * default (old) behavior, writing out only one block 890 */ 891 bp->b_flags |= B_BUSY | B_ASYNC; 892 nwritten = bp->b_bufsize; 893 (void) VOP_BWRITE(bp); 894 return nwritten; 895} 896 897 898/* 899 * Find a buffer header which is available for use. 900 */ 901static struct buf * 902getnewbuf(struct vnode *vp, int slpflag, int slptimeo, int size, int maxsize) 903{ 904 struct buf *bp; 905 int nbyteswritten = 0; 906 vm_offset_t addr; 907 static int writerecursion = 0; 908 909start: 910 if (bufspace >= maxbufspace) 911 goto trytofreespace; 912 913 /* can we constitute a new buffer? */ 914 if ((bp = TAILQ_FIRST(&bufqueues[QUEUE_EMPTY]))) { 915#if !defined(MAX_PERF) 916 if (bp->b_qindex != QUEUE_EMPTY) 917 panic("getnewbuf: inconsistent EMPTY queue, qindex=%d", 918 bp->b_qindex); 919#endif 920 bp->b_flags |= B_BUSY; 921 bremfree(bp); 922 goto fillbuf; 923 } 924trytofreespace: 925 /* 926 * We keep the file I/O from hogging metadata I/O 927 * This is desirable because file data is cached in the 928 * VM/Buffer cache even if a buffer is freed. 929 */ 930 if ((bp = TAILQ_FIRST(&bufqueues[QUEUE_AGE]))) { 931#if !defined(MAX_PERF) 932 if (bp->b_qindex != QUEUE_AGE) 933 panic("getnewbuf: inconsistent AGE queue, qindex=%d", 934 bp->b_qindex); 935#endif 936 } else if ((bp = TAILQ_FIRST(&bufqueues[QUEUE_LRU]))) { 937#if !defined(MAX_PERF) 938 if (bp->b_qindex != QUEUE_LRU) 939 panic("getnewbuf: inconsistent LRU queue, qindex=%d", 940 bp->b_qindex); 941#endif 942 } 943 if (!bp) { 944 /* wait for a free buffer of any kind */ 945 needsbuffer |= VFS_BIO_NEED_ANY; 946 do 947 tsleep(&needsbuffer, (PRIBIO + 1) | slpflag, "newbuf", 948 slptimeo); 949 while (needsbuffer & VFS_BIO_NEED_ANY); 950 return (0); 951 } 952 953#if defined(DIAGNOSTIC) 954 if (bp->b_flags & B_BUSY) { 955 panic("getnewbuf: busy buffer on free list\n"); 956 } 957#endif 958 959 /* 960 * We are fairly aggressive about freeing VMIO buffers, but since 961 * the buffering is intact without buffer headers, there is not 962 * much loss. We gain by maintaining non-VMIOed metadata in buffers. 963 */ 964 if ((bp->b_qindex == QUEUE_LRU) && (bp->b_usecount > 0)) { 965 if ((bp->b_flags & B_VMIO) == 0 || 966 (vmiospace < maxvmiobufspace)) { 967 --bp->b_usecount; 968 TAILQ_REMOVE(&bufqueues[QUEUE_LRU], bp, b_freelist); 969 if (TAILQ_FIRST(&bufqueues[QUEUE_LRU]) != NULL) { 970 TAILQ_INSERT_TAIL(&bufqueues[QUEUE_LRU], bp, b_freelist); 971 goto start; 972 } 973 TAILQ_INSERT_TAIL(&bufqueues[QUEUE_LRU], bp, b_freelist); 974 } 975 } 976 977 978 /* if we are a delayed write, convert to an async write */ 979 if ((bp->b_flags & (B_DELWRI | B_INVAL)) == B_DELWRI) { 980 981 if (writerecursion > 0) { 982 bp = TAILQ_FIRST(&bufqueues[QUEUE_AGE]); 983 while (bp) { 984 if ((bp->b_flags & B_DELWRI) == 0) 985 break; 986 bp = TAILQ_NEXT(bp, b_freelist); 987 } 988 if (bp == NULL) { 989 bp = TAILQ_FIRST(&bufqueues[QUEUE_LRU]); 990 while (bp) { 991 if ((bp->b_flags & B_DELWRI) == 0) 992 break; 993 bp = TAILQ_NEXT(bp, b_freelist); 994 } 995 } 996 if (bp == NULL) 997 panic("getnewbuf: cannot get buffer, infinite recursion failure"); 998 } else { 999 ++writerecursion; 1000 nbyteswritten += vfs_bio_awrite(bp); 1001 --writerecursion; 1002 if (!slpflag && !slptimeo) { 1003 return (0); 1004 } 1005 goto start; 1006 } 1007 } 1008 1009 if (bp->b_flags & B_WANTED) { 1010 bp->b_flags &= ~B_WANTED; 1011 wakeup(bp); 1012 } 1013 bremfree(bp); 1014 bp->b_flags |= B_BUSY; 1015 1016 if (bp->b_flags & B_VMIO) { 1017 bp->b_flags &= ~B_ASYNC; 1018 vfs_vmio_release(bp); 1019 } 1020 1021 if (bp->b_vp) 1022 brelvp(bp); 1023 1024fillbuf: 1025 /* we are not free, nor do we contain interesting data */ 1026 if (bp->b_rcred != NOCRED) { 1027 crfree(bp->b_rcred); 1028 bp->b_rcred = NOCRED; 1029 } 1030 if (bp->b_wcred != NOCRED) { 1031 crfree(bp->b_wcred); 1032 bp->b_wcred = NOCRED; 1033 } 1034 1035 LIST_REMOVE(bp, b_hash); 1036 LIST_INSERT_HEAD(&invalhash, bp, b_hash); 1037 if (bp->b_bufsize) { 1038 allocbuf(bp, 0); 1039 } 1040 bp->b_flags = B_BUSY; 1041 bp->b_dev = NODEV; 1042 bp->b_vp = NULL; 1043 bp->b_blkno = bp->b_lblkno = 0; 1044 bp->b_iodone = 0; 1045 bp->b_error = 0; 1046 bp->b_resid = 0; 1047 bp->b_bcount = 0; 1048 bp->b_npages = 0; 1049 bp->b_dirtyoff = bp->b_dirtyend = 0; 1050 bp->b_validoff = bp->b_validend = 0; 1051 bp->b_usecount = 4; 1052 1053 maxsize = (maxsize + PAGE_MASK) & ~PAGE_MASK; 1054 1055 /* 1056 * we assume that buffer_map is not at address 0 1057 */ 1058 addr = 0; 1059 if (maxsize != bp->b_kvasize) { 1060 bfreekva(bp); 1061 1062 /* 1063 * See if we have buffer kva space 1064 */ 1065 if (vm_map_findspace(buffer_map, 1066 vm_map_min(buffer_map), maxsize, &addr)) { 1067 bp->b_flags |= B_INVAL; 1068 brelse(bp); 1069 goto trytofreespace; 1070 } 1071 } 1072 1073 /* 1074 * See if we are below are allocated minimum 1075 */ 1076 if (bufspace >= (maxbufspace + nbyteswritten)) { 1077 bp->b_flags |= B_INVAL; 1078 brelse(bp); 1079 goto trytofreespace; 1080 } 1081 1082 /* 1083 * create a map entry for the buffer -- in essence 1084 * reserving the kva space. 1085 */ 1086 if (addr) { 1087 vm_map_insert(buffer_map, NULL, 0, 1088 addr, addr + maxsize, 1089 VM_PROT_ALL, VM_PROT_ALL, MAP_NOFAULT); 1090 1091 bp->b_kvabase = (caddr_t) addr; 1092 bp->b_kvasize = maxsize; 1093 } 1094 bp->b_data = bp->b_kvabase; 1095 1096 return (bp); 1097} 1098 1099static void 1100waitfreebuffers(int slpflag, int slptimeo) { 1101 while (numfreebuffers < hifreebuffers) { 1102 flushdirtybuffers(slpflag, slptimeo); 1103 if (numfreebuffers < hifreebuffers) 1104 break; 1105 needsbuffer |= VFS_BIO_NEED_FREE; 1106 if (tsleep(&needsbuffer, PRIBIO|slpflag, "biofre", slptimeo)) 1107 break; 1108 } 1109} 1110 1111static void 1112flushdirtybuffers(int slpflag, int slptimeo) { 1113 int s; 1114 static pid_t flushing = 0; 1115 1116 s = splbio(); 1117 1118 if (flushing) { 1119 if (flushing == curproc->p_pid) { 1120 splx(s); 1121 return; 1122 } 1123 while (flushing) { 1124 if (tsleep(&flushing, PRIBIO|slpflag, "biofls", slptimeo)) { 1125 splx(s); 1126 return; 1127 } 1128 } 1129 } 1130 flushing = curproc->p_pid; 1131 1132 while (numdirtybuffers > lodirtybuffers) { 1133 struct buf *bp; 1134 needsbuffer |= VFS_BIO_NEED_LOWLIMIT; 1135 bp = TAILQ_FIRST(&bufqueues[QUEUE_AGE]); 1136 if (bp == NULL) 1137 bp = TAILQ_FIRST(&bufqueues[QUEUE_LRU]); 1138 1139 while (bp && ((bp->b_flags & B_DELWRI) == 0)) { 1140 bp = TAILQ_NEXT(bp, b_freelist); 1141 } 1142 1143 if (bp) { 1144 splx(s); 1145 vfs_bio_awrite(bp); 1146 s = splbio(); 1147 continue; 1148 } 1149 break; 1150 } 1151 1152 flushing = 0; 1153 wakeup(&flushing); 1154 splx(s); 1155} 1156 1157/* 1158 * Check to see if a block is currently memory resident. 1159 */ 1160struct buf * 1161incore(struct vnode * vp, daddr_t blkno) 1162{ 1163 struct buf *bp; 1164 1165 int s = splbio(); 1166 bp = gbincore(vp, blkno); 1167 splx(s); 1168 return (bp); 1169} 1170 1171/* 1172 * Returns true if no I/O is needed to access the 1173 * associated VM object. This is like incore except 1174 * it also hunts around in the VM system for the data. 1175 */ 1176 1177int 1178inmem(struct vnode * vp, daddr_t blkno) 1179{ 1180 vm_object_t obj; 1181 vm_offset_t toff, tinc; 1182 vm_page_t m; 1183 vm_ooffset_t off; 1184 1185 if (incore(vp, blkno)) 1186 return 1; 1187 if (vp->v_mount == NULL) 1188 return 0; 1189 if ((vp->v_object == NULL) || (vp->v_flag & VVMIO) == 0) 1190 return 0; 1191 1192 obj = vp->v_object; 1193 tinc = PAGE_SIZE; 1194 if (tinc > vp->v_mount->mnt_stat.f_iosize) 1195 tinc = vp->v_mount->mnt_stat.f_iosize; 1196 off = blkno * vp->v_mount->mnt_stat.f_iosize; 1197 1198 for (toff = 0; toff < vp->v_mount->mnt_stat.f_iosize; toff += tinc) { 1199 1200 m = vm_page_lookup(obj, OFF_TO_IDX(off + toff)); 1201 if (!m) 1202 return 0; 1203 if (vm_page_is_valid(m, (vm_offset_t) (toff + off), tinc) == 0) 1204 return 0; 1205 } 1206 return 1; 1207} 1208 1209/* 1210 * now we set the dirty range for the buffer -- 1211 * for NFS -- if the file is mapped and pages have 1212 * been written to, let it know. We want the 1213 * entire range of the buffer to be marked dirty if 1214 * any of the pages have been written to for consistancy 1215 * with the b_validoff, b_validend set in the nfs write 1216 * code, and used by the nfs read code. 1217 */ 1218static void 1219vfs_setdirty(struct buf *bp) { 1220 int i; 1221 vm_object_t object; 1222 vm_offset_t boffset, offset; 1223 /* 1224 * We qualify the scan for modified pages on whether the 1225 * object has been flushed yet. The OBJ_WRITEABLE flag 1226 * is not cleared simply by protecting pages off. 1227 */ 1228 if ((bp->b_flags & B_VMIO) && 1229 ((object = bp->b_pages[0]->object)->flags & (OBJ_WRITEABLE|OBJ_CLEANING))) { 1230 /* 1231 * test the pages to see if they have been modified directly 1232 * by users through the VM system. 1233 */ 1234 for (i = 0; i < bp->b_npages; i++) 1235 vm_page_test_dirty(bp->b_pages[i]); 1236 1237 /* 1238 * scan forwards for the first page modified 1239 */ 1240 for (i = 0; i < bp->b_npages; i++) { 1241 if (bp->b_pages[i]->dirty) { 1242 break; 1243 } 1244 } 1245 boffset = (i << PAGE_SHIFT); 1246 if (boffset < bp->b_dirtyoff) { 1247 bp->b_dirtyoff = boffset; 1248 } 1249 1250 /* 1251 * scan backwards for the last page modified 1252 */ 1253 for (i = bp->b_npages - 1; i >= 0; --i) { 1254 if (bp->b_pages[i]->dirty) { 1255 break; 1256 } 1257 } 1258 boffset = (i + 1); 1259 offset = boffset + bp->b_pages[0]->pindex; 1260 if (offset >= object->size) 1261 boffset = object->size - bp->b_pages[0]->pindex; 1262 if (bp->b_dirtyend < (boffset << PAGE_SHIFT)) 1263 bp->b_dirtyend = (boffset << PAGE_SHIFT); 1264 } 1265} 1266 1267/* 1268 * Get a block given a specified block and offset into a file/device. 1269 */ 1270struct buf * 1271getblk(struct vnode * vp, daddr_t blkno, int size, int slpflag, int slptimeo) 1272{ 1273 struct buf *bp; 1274 int s; 1275 struct bufhashhdr *bh; 1276 int maxsize; 1277 1278 if (vp->v_mount) { 1279 maxsize = vp->v_mount->mnt_stat.f_iosize; 1280 /* 1281 * This happens on mount points. 1282 */ 1283 if (maxsize < size) 1284 maxsize = size; 1285 } else { 1286 maxsize = size; 1287 } 1288 1289#if !defined(MAX_PERF) 1290 if (size > MAXBSIZE) 1291 panic("getblk: size(%d) > MAXBSIZE(%d)\n", size, MAXBSIZE); 1292#endif 1293 1294 s = splbio(); 1295loop: 1296 if (numfreebuffers < lofreebuffers) { 1297 waitfreebuffers(slpflag, slptimeo); 1298 } 1299 1300 if ((bp = gbincore(vp, blkno))) { 1301 if (bp->b_flags & B_BUSY) { 1302 bp->b_flags |= B_WANTED; 1303 if (bp->b_usecount < BUF_MAXUSE) 1304 ++bp->b_usecount; 1305 if (!tsleep(bp, 1306 (PRIBIO + 1) | slpflag, "getblk", slptimeo)) 1307 goto loop; 1308 1309 splx(s); 1310 return (struct buf *) NULL; 1311 } 1312 bp->b_flags |= B_BUSY | B_CACHE; 1313 bremfree(bp); 1314 1315 /* 1316 * check for size inconsistancies (note that they shouldn't 1317 * happen but do when filesystems don't handle the size changes 1318 * correctly.) We are conservative on metadata and don't just 1319 * extend the buffer but write and re-constitute it. 1320 */ 1321 1322 if (bp->b_bcount != size) { 1323 if ((bp->b_flags & B_VMIO) && (size <= bp->b_kvasize)) { 1324 allocbuf(bp, size); 1325 } else { 1326 bp->b_flags |= B_NOCACHE; 1327 VOP_BWRITE(bp); 1328 goto loop; 1329 } 1330 } 1331 1332 if (bp->b_usecount < BUF_MAXUSE) 1333 ++bp->b_usecount; 1334 splx(s); 1335 return (bp); 1336 } else { 1337 vm_object_t obj; 1338 1339 if ((bp = getnewbuf(vp, slpflag, slptimeo, size, maxsize)) == 0) { 1340 if (slpflag || slptimeo) { 1341 splx(s); 1342 return NULL; 1343 } 1344 goto loop; 1345 } 1346 1347 /* 1348 * This code is used to make sure that a buffer is not 1349 * created while the getnewbuf routine is blocked. 1350 * Normally the vnode is locked so this isn't a problem. 1351 * VBLK type I/O requests, however, don't lock the vnode. 1352 */ 1353 if (!VOP_ISLOCKED(vp) && gbincore(vp, blkno)) { 1354 bp->b_flags |= B_INVAL; 1355 brelse(bp); 1356 goto loop; 1357 } 1358 1359 /* 1360 * Insert the buffer into the hash, so that it can 1361 * be found by incore. 1362 */ 1363 bp->b_blkno = bp->b_lblkno = blkno; 1364 bgetvp(vp, bp); 1365 LIST_REMOVE(bp, b_hash); 1366 bh = BUFHASH(vp, blkno); 1367 LIST_INSERT_HEAD(bh, bp, b_hash); 1368 1369 if ((obj = vp->v_object) && (vp->v_flag & VVMIO)) { 1370 bp->b_flags |= (B_VMIO | B_CACHE); 1371#if defined(VFS_BIO_DEBUG) 1372 if (vp->v_type != VREG && vp->v_type != VBLK) 1373 printf("getblk: vmioing file type %d???\n", vp->v_type); 1374#endif 1375 } else { 1376 bp->b_flags &= ~B_VMIO; 1377 } 1378 splx(s); 1379 1380 allocbuf(bp, size); 1381#ifdef PC98 1382 /* 1383 * 1024byte/sector support 1384 */ 1385#define B_XXX2 0x8000000 1386 if (vp->v_flag & 0x10000) bp->b_flags |= B_XXX2; 1387#endif 1388 return (bp); 1389 } 1390} 1391 1392/* 1393 * Get an empty, disassociated buffer of given size. 1394 */ 1395struct buf * 1396geteblk(int size) 1397{ 1398 struct buf *bp; 1399 int s; 1400 1401 s = splbio(); 1402 while ((bp = getnewbuf(0, 0, 0, size, MAXBSIZE)) == 0); 1403 splx(s); 1404 allocbuf(bp, size); 1405 bp->b_flags |= B_INVAL; 1406 return (bp); 1407} 1408 1409 1410/* 1411 * This code constitutes the buffer memory from either anonymous system 1412 * memory (in the case of non-VMIO operations) or from an associated 1413 * VM object (in the case of VMIO operations). 1414 * 1415 * Note that this code is tricky, and has many complications to resolve 1416 * deadlock or inconsistant data situations. Tread lightly!!! 1417 * 1418 * Modify the length of a buffer's underlying buffer storage without 1419 * destroying information (unless, of course the buffer is shrinking). 1420 */ 1421int 1422allocbuf(struct buf * bp, int size) 1423{ 1424 1425 int s; 1426 int newbsize, mbsize; 1427 int i; 1428 1429#if !defined(MAX_PERF) 1430 if (!(bp->b_flags & B_BUSY)) 1431 panic("allocbuf: buffer not busy"); 1432 1433 if (bp->b_kvasize < size) 1434 panic("allocbuf: buffer too small"); 1435#endif 1436 1437 if ((bp->b_flags & B_VMIO) == 0) { 1438 caddr_t origbuf; 1439 int origbufsize; 1440 /* 1441 * Just get anonymous memory from the kernel 1442 */ 1443 mbsize = (size + DEV_BSIZE - 1) & ~(DEV_BSIZE - 1); 1444#if !defined(NO_B_MALLOC) 1445 if (bp->b_flags & B_MALLOC) 1446 newbsize = mbsize; 1447 else 1448#endif 1449 newbsize = round_page(size); 1450 1451 if (newbsize < bp->b_bufsize) { 1452#if !defined(NO_B_MALLOC) 1453 /* 1454 * malloced buffers are not shrunk 1455 */ 1456 if (bp->b_flags & B_MALLOC) { 1457 if (newbsize) { 1458 bp->b_bcount = size; 1459 } else { 1460 free(bp->b_data, M_BIOBUF); 1461 bufspace -= bp->b_bufsize; 1462 bufmallocspace -= bp->b_bufsize; 1463 bp->b_data = bp->b_kvabase; 1464 bp->b_bufsize = 0; 1465 bp->b_bcount = 0; 1466 bp->b_flags &= ~B_MALLOC; 1467 } 1468 return 1; 1469 } 1470#endif 1471 vm_hold_free_pages( 1472 bp, 1473 (vm_offset_t) bp->b_data + newbsize, 1474 (vm_offset_t) bp->b_data + bp->b_bufsize); 1475 } else if (newbsize > bp->b_bufsize) { 1476#if !defined(NO_B_MALLOC) 1477 /* 1478 * We only use malloced memory on the first allocation. 1479 * and revert to page-allocated memory when the buffer grows. 1480 */ 1481 if ( (bufmallocspace < maxbufmallocspace) && 1482 (bp->b_bufsize == 0) && 1483 (mbsize <= PAGE_SIZE/2)) { 1484 1485 bp->b_data = malloc(mbsize, M_BIOBUF, M_WAITOK); 1486 bp->b_bufsize = mbsize; 1487 bp->b_bcount = size; 1488 bp->b_flags |= B_MALLOC; 1489 bufspace += mbsize; 1490 bufmallocspace += mbsize; 1491 return 1; 1492 } 1493#endif 1494 origbuf = NULL; 1495 origbufsize = 0; 1496#if !defined(NO_B_MALLOC) 1497 /* 1498 * If the buffer is growing on it's other-than-first allocation, 1499 * then we revert to the page-allocation scheme. 1500 */ 1501 if (bp->b_flags & B_MALLOC) { 1502 origbuf = bp->b_data; 1503 origbufsize = bp->b_bufsize; 1504 bp->b_data = bp->b_kvabase; 1505 bufspace -= bp->b_bufsize; 1506 bufmallocspace -= bp->b_bufsize; 1507 bp->b_bufsize = 0; 1508 bp->b_flags &= ~B_MALLOC; 1509 newbsize = round_page(newbsize); 1510 } 1511#endif 1512 vm_hold_load_pages( 1513 bp, 1514 (vm_offset_t) bp->b_data + bp->b_bufsize, 1515 (vm_offset_t) bp->b_data + newbsize); 1516#if !defined(NO_B_MALLOC) 1517 if (origbuf) { 1518 bcopy(origbuf, bp->b_data, origbufsize); 1519 free(origbuf, M_BIOBUF); 1520 } 1521#endif 1522 } 1523 } else { 1524 vm_page_t m; 1525 int desiredpages; 1526 1527 newbsize = (size + DEV_BSIZE - 1) & ~(DEV_BSIZE - 1); 1528 desiredpages = (round_page(newbsize) >> PAGE_SHIFT); 1529 1530#if !defined(NO_B_MALLOC) 1531 if (bp->b_flags & B_MALLOC) 1532 panic("allocbuf: VMIO buffer can't be malloced"); 1533#endif 1534 1535 if (newbsize < bp->b_bufsize) { 1536 if (desiredpages < bp->b_npages) { 1537 for (i = desiredpages; i < bp->b_npages; i++) { 1538 /* 1539 * the page is not freed here -- it 1540 * is the responsibility of vnode_pager_setsize 1541 */ 1542 m = bp->b_pages[i]; 1543#if defined(DIAGNOSTIC) 1544 if (m == bogus_page) 1545 panic("allocbuf: bogus page found"); 1546#endif 1547 s = splvm(); 1548 while ((m->flags & PG_BUSY) || (m->busy != 0)) { 1549 m->flags |= PG_WANTED; 1550 tsleep(m, PVM, "biodep", 0); 1551 } 1552 splx(s); 1553 1554 bp->b_pages[i] = NULL; 1555 vm_page_unwire(m); 1556 } 1557 pmap_qremove((vm_offset_t) trunc_page(bp->b_data) + 1558 (desiredpages << PAGE_SHIFT), (bp->b_npages - desiredpages)); 1559 bp->b_npages = desiredpages; 1560 } 1561 } else if (newbsize > bp->b_bufsize) { 1562 vm_object_t obj; 1563 vm_offset_t tinc, toff; 1564 vm_ooffset_t off; 1565 vm_pindex_t objoff; 1566 int pageindex, curbpnpages; 1567 struct vnode *vp; 1568 int bsize; 1569 1570 vp = bp->b_vp; 1571 1572 if (vp->v_type == VBLK) 1573 bsize = DEV_BSIZE; 1574 else 1575 bsize = vp->v_mount->mnt_stat.f_iosize; 1576 1577 if (bp->b_npages < desiredpages) { 1578 obj = vp->v_object; 1579 tinc = PAGE_SIZE; 1580 if (tinc > bsize) 1581 tinc = bsize; 1582 off = (vm_ooffset_t) bp->b_lblkno * bsize; 1583 curbpnpages = bp->b_npages; 1584 doretry: 1585 bp->b_flags |= B_CACHE; 1586 bp->b_validoff = bp->b_validend = 0; 1587 for (toff = 0; toff < newbsize; toff += tinc) { 1588 int bytesinpage; 1589 1590 pageindex = toff >> PAGE_SHIFT; 1591 objoff = OFF_TO_IDX(off + toff); 1592 if (pageindex < curbpnpages) { 1593 1594 m = bp->b_pages[pageindex]; 1595#ifdef VFS_BIO_DIAG 1596 if (m->pindex != objoff) 1597 panic("allocbuf: page changed offset??!!!?"); 1598#endif 1599 bytesinpage = tinc; 1600 if (tinc > (newbsize - toff)) 1601 bytesinpage = newbsize - toff; 1602 if (bp->b_flags & B_CACHE) 1603 vfs_buf_set_valid(bp, off, toff, bytesinpage, m); 1604 continue; 1605 } 1606 m = vm_page_lookup(obj, objoff); 1607 if (!m) { 1608 m = vm_page_alloc(obj, objoff, VM_ALLOC_NORMAL); 1609 if (!m) { 1610 VM_WAIT; 1611 goto doretry; 1612 } 1613 /* 1614 * Normally it is unwise to clear PG_BUSY without 1615 * PAGE_WAKEUP -- but it is okay here, as there is 1616 * no chance for blocking between here and vm_page_alloc 1617 */ 1618 m->flags &= ~PG_BUSY; 1619 vm_page_wire(m); 1620 bp->b_flags &= ~B_CACHE; 1621 } else if (m->flags & PG_BUSY) { 1622 s = splvm(); 1623 if (m->flags & PG_BUSY) { 1624 m->flags |= PG_WANTED; 1625 tsleep(m, PVM, "pgtblk", 0); 1626 } 1627 splx(s); 1628 goto doretry; 1629 } else { 1630 if ((curproc != pageproc) && 1631 ((m->queue - m->pc) == PQ_CACHE) && 1632 ((cnt.v_free_count + cnt.v_cache_count) < 1633 (cnt.v_free_min + cnt.v_cache_min))) { 1634 pagedaemon_wakeup(); 1635 } 1636 bytesinpage = tinc; 1637 if (tinc > (newbsize - toff)) 1638 bytesinpage = newbsize - toff; 1639 if (bp->b_flags & B_CACHE) 1640 vfs_buf_set_valid(bp, off, toff, bytesinpage, m); 1641 vm_page_wire(m); 1642 } 1643 bp->b_pages[pageindex] = m; 1644 curbpnpages = pageindex + 1; 1645 } 1646 if (vp->v_tag == VT_NFS && 1647 vp->v_type != VBLK) { 1648 if (bp->b_dirtyend > 0) { 1649 bp->b_validoff = min(bp->b_validoff, bp->b_dirtyoff); 1650 bp->b_validend = max(bp->b_validend, bp->b_dirtyend); 1651 } 1652 if (bp->b_validend == 0) 1653 bp->b_flags &= ~B_CACHE; 1654 } 1655 bp->b_data = (caddr_t) trunc_page(bp->b_data); 1656 bp->b_npages = curbpnpages; 1657 pmap_qenter((vm_offset_t) bp->b_data, 1658 bp->b_pages, bp->b_npages); 1659 ((vm_offset_t) bp->b_data) |= off & PAGE_MASK; 1660 } 1661 } 1662 } 1663 if (bp->b_flags & B_VMIO) 1664 vmiospace += (newbsize - bp->b_bufsize); 1665 bufspace += (newbsize - bp->b_bufsize); 1666 bp->b_bufsize = newbsize; 1667 bp->b_bcount = size; 1668 return 1; 1669} 1670 1671/* 1672 * Wait for buffer I/O completion, returning error status. 1673 */ 1674int 1675biowait(register struct buf * bp) 1676{ 1677 int s; 1678 1679 s = splbio(); 1680 while ((bp->b_flags & B_DONE) == 0) 1681#if defined(NO_SCHEDULE_MODS) 1682 tsleep(bp, PRIBIO, "biowait", 0); 1683#else 1684 tsleep(bp, curproc->p_usrpri, "biowait", 0); 1685#endif 1686 splx(s); 1687 if (bp->b_flags & B_EINTR) { 1688 bp->b_flags &= ~B_EINTR; 1689 return (EINTR); 1690 } 1691 if (bp->b_flags & B_ERROR) { 1692 return (bp->b_error ? bp->b_error : EIO); 1693 } else { 1694 return (0); 1695 } 1696} 1697 1698/* 1699 * Finish I/O on a buffer, calling an optional function. 1700 * This is usually called from interrupt level, so process blocking 1701 * is not *a good idea*. 1702 */ 1703void 1704biodone(register struct buf * bp) 1705{ 1706 int s; 1707 1708 s = splbio(); 1709 1710#if !defined(MAX_PERF) 1711 if (!(bp->b_flags & B_BUSY)) 1712 panic("biodone: buffer not busy"); 1713#endif 1714 1715 if (bp->b_flags & B_DONE) { 1716 splx(s); 1717#if !defined(MAX_PERF) 1718 printf("biodone: buffer already done\n"); 1719#endif 1720 return; 1721 } 1722 bp->b_flags |= B_DONE; 1723 1724 if ((bp->b_flags & B_READ) == 0) { 1725 vwakeup(bp); 1726 } 1727#ifdef BOUNCE_BUFFERS 1728 if (bp->b_flags & B_BOUNCE) 1729 vm_bounce_free(bp); 1730#endif 1731 1732 /* call optional completion function if requested */ 1733 if (bp->b_flags & B_CALL) { 1734 bp->b_flags &= ~B_CALL; 1735 (*bp->b_iodone) (bp); 1736 splx(s); 1737 return; 1738 } 1739 if (bp->b_flags & B_VMIO) { 1740 int i, resid; 1741 vm_ooffset_t foff; 1742 vm_page_t m; 1743 vm_object_t obj; 1744 int iosize; 1745 struct vnode *vp = bp->b_vp; 1746 1747 obj = vp->v_object; 1748 1749#if defined(VFS_BIO_DEBUG) 1750 if (vp->v_usecount == 0) { 1751 panic("biodone: zero vnode ref count"); 1752 } 1753 1754 if (vp->v_object == NULL) { 1755 panic("biodone: missing VM object"); 1756 } 1757 1758 if ((vp->v_flag & VVMIO) == 0) { 1759 panic("biodone: vnode is not setup for merged cache"); 1760 } 1761#endif 1762 1763 if (vp->v_type == VBLK) 1764 foff = (vm_ooffset_t) DEV_BSIZE * bp->b_lblkno; 1765 else 1766 foff = (vm_ooffset_t) vp->v_mount->mnt_stat.f_iosize * bp->b_lblkno; 1767#if !defined(MAX_PERF) 1768 if (!obj) { 1769 panic("biodone: no object"); 1770 } 1771#endif 1772#if defined(VFS_BIO_DEBUG) 1773 if (obj->paging_in_progress < bp->b_npages) { 1774 printf("biodone: paging in progress(%d) < bp->b_npages(%d)\n", 1775 obj->paging_in_progress, bp->b_npages); 1776 } 1777#endif 1778 iosize = bp->b_bufsize; 1779 for (i = 0; i < bp->b_npages; i++) { 1780 int bogusflag = 0; 1781 m = bp->b_pages[i]; 1782 if (m == bogus_page) { 1783 bogusflag = 1; 1784 m = vm_page_lookup(obj, OFF_TO_IDX(foff)); 1785 if (!m) { 1786#if defined(VFS_BIO_DEBUG) 1787 printf("biodone: page disappeared\n"); 1788#endif 1789 --obj->paging_in_progress; 1790 continue; 1791 } 1792 bp->b_pages[i] = m; 1793 pmap_qenter(trunc_page(bp->b_data), bp->b_pages, bp->b_npages); 1794 } 1795#if defined(VFS_BIO_DEBUG) 1796 if (OFF_TO_IDX(foff) != m->pindex) { 1797 printf("biodone: foff(%d)/m->pindex(%d) mismatch\n", foff, m->pindex); 1798 } 1799#endif 1800 resid = IDX_TO_OFF(m->pindex + 1) - foff; 1801 if (resid > iosize) 1802 resid = iosize; 1803 /* 1804 * In the write case, the valid and clean bits are 1805 * already changed correctly, so we only need to do this 1806 * here in the read case. 1807 */ 1808 if ((bp->b_flags & B_READ) && !bogusflag && resid > 0) { 1809 vfs_page_set_valid(bp, foff, i, m); 1810 } 1811 1812 /* 1813 * when debugging new filesystems or buffer I/O methods, this 1814 * is the most common error that pops up. if you see this, you 1815 * have not set the page busy flag correctly!!! 1816 */ 1817 if (m->busy == 0) { 1818#if !defined(MAX_PERF) 1819 printf("biodone: page busy < 0, " 1820 "pindex: %d, foff: 0x(%x,%x), " 1821 "resid: %d, index: %d\n", 1822 (int) m->pindex, (int)(foff >> 32), 1823 (int) foff & 0xffffffff, resid, i); 1824#endif 1825 if (vp->v_type != VBLK) 1826#if !defined(MAX_PERF) 1827 printf(" iosize: %ld, lblkno: %d, flags: 0x%lx, npages: %d\n", 1828 bp->b_vp->v_mount->mnt_stat.f_iosize, 1829 (int) bp->b_lblkno, 1830 bp->b_flags, bp->b_npages); 1831 else 1832 printf(" VDEV, lblkno: %d, flags: 0x%lx, npages: %d\n", 1833 (int) bp->b_lblkno, 1834 bp->b_flags, bp->b_npages); 1835 printf(" valid: 0x%x, dirty: 0x%x, wired: %d\n", 1836 m->valid, m->dirty, m->wire_count); 1837#endif 1838 panic("biodone: page busy < 0\n"); 1839 } 1840 --m->busy; 1841 if ((m->busy == 0) && (m->flags & PG_WANTED)) { 1842 m->flags &= ~PG_WANTED; 1843 wakeup(m); 1844 } 1845 --obj->paging_in_progress; 1846 foff += resid; 1847 iosize -= resid; 1848 } 1849 if (obj && obj->paging_in_progress == 0 && 1850 (obj->flags & OBJ_PIPWNT)) { 1851 obj->flags &= ~OBJ_PIPWNT; 1852 wakeup(obj); 1853 } 1854 } 1855 /* 1856 * For asynchronous completions, release the buffer now. The brelse 1857 * checks for B_WANTED and will do the wakeup there if necessary - so 1858 * no need to do a wakeup here in the async case. 1859 */ 1860 1861 if (bp->b_flags & B_ASYNC) { 1862 if ((bp->b_flags & (B_NOCACHE | B_INVAL | B_ERROR | B_RELBUF)) != 0) 1863 brelse(bp); 1864 else 1865 bqrelse(bp); 1866 } else { 1867 bp->b_flags &= ~B_WANTED; 1868 wakeup(bp); 1869 } 1870 splx(s); 1871} 1872 1873int 1874count_lock_queue() 1875{ 1876 int count; 1877 struct buf *bp; 1878 1879 count = 0; 1880 for (bp = TAILQ_FIRST(&bufqueues[QUEUE_LOCKED]); 1881 bp != NULL; 1882 bp = TAILQ_NEXT(bp, b_freelist)) 1883 count++; 1884 return (count); 1885} 1886 1887int vfs_update_interval = 30; 1888 1889static void 1890vfs_update() 1891{ 1892 while (1) { 1893 tsleep(&vfs_update_wakeup, PUSER, "update", 1894 hz * vfs_update_interval); 1895 vfs_update_wakeup = 0; 1896 sync(curproc, NULL); 1897 } 1898} 1899 1900static int 1901sysctl_kern_updateinterval SYSCTL_HANDLER_ARGS 1902{ 1903 int error = sysctl_handle_int(oidp, 1904 oidp->oid_arg1, oidp->oid_arg2, req); 1905 if (!error) 1906 wakeup(&vfs_update_wakeup); 1907 return error; 1908} 1909 1910SYSCTL_PROC(_kern, KERN_UPDATEINTERVAL, update, CTLTYPE_INT|CTLFLAG_RW, 1911 &vfs_update_interval, 0, sysctl_kern_updateinterval, "I", ""); 1912 1913 1914/* 1915 * This routine is called in lieu of iodone in the case of 1916 * incomplete I/O. This keeps the busy status for pages 1917 * consistant. 1918 */ 1919void 1920vfs_unbusy_pages(struct buf * bp) 1921{ 1922 int i; 1923 1924 if (bp->b_flags & B_VMIO) { 1925 struct vnode *vp = bp->b_vp; 1926 vm_object_t obj = vp->v_object; 1927 vm_ooffset_t foff; 1928 1929 foff = (vm_ooffset_t) vp->v_mount->mnt_stat.f_iosize * bp->b_lblkno; 1930 1931 for (i = 0; i < bp->b_npages; i++) { 1932 vm_page_t m = bp->b_pages[i]; 1933 1934 if (m == bogus_page) { 1935 m = vm_page_lookup(obj, OFF_TO_IDX(foff) + i); 1936#if !defined(MAX_PERF) 1937 if (!m) { 1938 panic("vfs_unbusy_pages: page missing\n"); 1939 } 1940#endif 1941 bp->b_pages[i] = m; 1942 pmap_qenter(trunc_page(bp->b_data), bp->b_pages, bp->b_npages); 1943 } 1944 --obj->paging_in_progress; 1945 --m->busy; 1946 if ((m->busy == 0) && (m->flags & PG_WANTED)) { 1947 m->flags &= ~PG_WANTED; 1948 wakeup(m); 1949 } 1950 } 1951 if (obj->paging_in_progress == 0 && 1952 (obj->flags & OBJ_PIPWNT)) { 1953 obj->flags &= ~OBJ_PIPWNT; 1954 wakeup(obj); 1955 } 1956 } 1957} 1958 1959/* 1960 * Set NFS' b_validoff and b_validend fields from the valid bits 1961 * of a page. If the consumer is not NFS, and the page is not 1962 * valid for the entire range, clear the B_CACHE flag to force 1963 * the consumer to re-read the page. 1964 */ 1965static void 1966vfs_buf_set_valid(struct buf *bp, 1967 vm_ooffset_t foff, vm_offset_t off, vm_offset_t size, 1968 vm_page_t m) 1969{ 1970 if (bp->b_vp->v_tag == VT_NFS && bp->b_vp->v_type != VBLK) { 1971 vm_offset_t svalid, evalid; 1972 int validbits = m->valid; 1973 1974 /* 1975 * This only bothers with the first valid range in the 1976 * page. 1977 */ 1978 svalid = off; 1979 while (validbits && !(validbits & 1)) { 1980 svalid += DEV_BSIZE; 1981 validbits >>= 1; 1982 } 1983 evalid = svalid; 1984 while (validbits & 1) { 1985 evalid += DEV_BSIZE; 1986 validbits >>= 1; 1987 } 1988 /* 1989 * Make sure this range is contiguous with the range 1990 * built up from previous pages. If not, then we will 1991 * just use the range from the previous pages. 1992 */ 1993 if (svalid == bp->b_validend) { 1994 bp->b_validoff = min(bp->b_validoff, svalid); 1995 bp->b_validend = max(bp->b_validend, evalid); 1996 } 1997 } else if (!vm_page_is_valid(m, 1998 (vm_offset_t) ((foff + off) & PAGE_MASK), 1999 size)) { 2000 bp->b_flags &= ~B_CACHE; 2001 } 2002} 2003 2004/* 2005 * Set the valid bits in a page, taking care of the b_validoff, 2006 * b_validend fields which NFS uses to optimise small reads. Off is 2007 * the offset within the file and pageno is the page index within the buf. 2008 */ 2009static void 2010vfs_page_set_valid(struct buf *bp, vm_ooffset_t off, int pageno, vm_page_t m) 2011{ 2012 struct vnode *vp = bp->b_vp; 2013 vm_ooffset_t soff, eoff; 2014 2015 soff = off; 2016 eoff = off + min(PAGE_SIZE, bp->b_bufsize); 2017 vm_page_set_invalid(m, 2018 (vm_offset_t) (soff & PAGE_MASK), 2019 (vm_offset_t) (eoff - soff)); 2020 if (vp->v_tag == VT_NFS && vp->v_type != VBLK) { 2021 vm_ooffset_t sv, ev; 2022 off = off - pageno * PAGE_SIZE; 2023 sv = off + ((bp->b_validoff + DEV_BSIZE - 1) & ~(DEV_BSIZE - 1)); 2024 ev = off + (bp->b_validend & ~(DEV_BSIZE - 1)); 2025 soff = max(sv, soff); 2026 eoff = min(ev, eoff); 2027 } 2028 if (eoff > soff) 2029 vm_page_set_validclean(m, 2030 (vm_offset_t) (soff & PAGE_MASK), 2031 (vm_offset_t) (eoff - soff)); 2032} 2033 2034/* 2035 * This routine is called before a device strategy routine. 2036 * It is used to tell the VM system that paging I/O is in 2037 * progress, and treat the pages associated with the buffer 2038 * almost as being PG_BUSY. Also the object paging_in_progress 2039 * flag is handled to make sure that the object doesn't become 2040 * inconsistant. 2041 */ 2042void 2043vfs_busy_pages(struct buf * bp, int clear_modify) 2044{ 2045 int i; 2046 2047 if (bp->b_flags & B_VMIO) { 2048 struct vnode *vp = bp->b_vp; 2049 vm_object_t obj = vp->v_object; 2050 vm_ooffset_t foff; 2051 2052 if (vp->v_type == VBLK) 2053 foff = (vm_ooffset_t) DEV_BSIZE * bp->b_lblkno; 2054 else 2055 foff = (vm_ooffset_t) vp->v_mount->mnt_stat.f_iosize * bp->b_lblkno; 2056 vfs_setdirty(bp); 2057 for (i = 0; i < bp->b_npages; i++, foff += PAGE_SIZE) { 2058 vm_page_t m = bp->b_pages[i]; 2059 2060 if ((bp->b_flags & B_CLUSTER) == 0) { 2061 obj->paging_in_progress++; 2062 m->busy++; 2063 } 2064 vm_page_protect(m, VM_PROT_NONE); 2065 if (clear_modify) 2066 vfs_page_set_valid(bp, foff, i, m); 2067 else if (bp->b_bcount >= PAGE_SIZE) { 2068 if (m->valid && (bp->b_flags & B_CACHE) == 0) { 2069 bp->b_pages[i] = bogus_page; 2070 pmap_qenter(trunc_page(bp->b_data), bp->b_pages, bp->b_npages); 2071 } 2072 } 2073 } 2074 } 2075} 2076 2077/* 2078 * Tell the VM system that the pages associated with this buffer 2079 * are clean. This is used for delayed writes where the data is 2080 * going to go to disk eventually without additional VM intevention. 2081 */ 2082void 2083vfs_clean_pages(struct buf * bp) 2084{ 2085 int i; 2086 2087 if (bp->b_flags & B_VMIO) { 2088 struct vnode *vp = bp->b_vp; 2089 vm_ooffset_t foff; 2090 2091 if (vp->v_type == VBLK) 2092 foff = (vm_ooffset_t) DEV_BSIZE * bp->b_lblkno; 2093 else 2094 foff = (vm_ooffset_t) vp->v_mount->mnt_stat.f_iosize * bp->b_lblkno; 2095 for (i = 0; i < bp->b_npages; i++, foff += PAGE_SIZE) { 2096 vm_page_t m = bp->b_pages[i]; 2097 2098 vfs_page_set_valid(bp, foff, i, m); 2099 } 2100 } 2101} 2102 2103void 2104vfs_bio_clrbuf(struct buf *bp) { 2105 int i; 2106 if( bp->b_flags & B_VMIO) { 2107 if( (bp->b_npages == 1) && (bp->b_bufsize < PAGE_SIZE)) { 2108 int mask; 2109 mask = 0; 2110 for(i=0;i<bp->b_bufsize;i+=DEV_BSIZE) 2111 mask |= (1 << (i/DEV_BSIZE)); 2112 if( bp->b_pages[0]->valid != mask) { 2113 bzero(bp->b_data, bp->b_bufsize); 2114 } 2115 bp->b_pages[0]->valid = mask; 2116 bp->b_resid = 0; 2117 return; 2118 } 2119 for(i=0;i<bp->b_npages;i++) { 2120 if( bp->b_pages[i]->valid == VM_PAGE_BITS_ALL) 2121 continue; 2122 if( bp->b_pages[i]->valid == 0) { 2123 if ((bp->b_pages[i]->flags & PG_ZERO) == 0) { 2124 bzero(bp->b_data + (i << PAGE_SHIFT), PAGE_SIZE); 2125 } 2126 } else { 2127 int j; 2128 for(j=0;j<PAGE_SIZE/DEV_BSIZE;j++) { 2129 if( (bp->b_pages[i]->valid & (1<<j)) == 0) 2130 bzero(bp->b_data + (i << PAGE_SHIFT) + j * DEV_BSIZE, DEV_BSIZE); 2131 } 2132 } 2133 /* bp->b_pages[i]->valid = VM_PAGE_BITS_ALL; */ 2134 } 2135 bp->b_resid = 0; 2136 } else { 2137 clrbuf(bp); 2138 } 2139} 2140 2141/* 2142 * vm_hold_load_pages and vm_hold_unload pages get pages into 2143 * a buffers address space. The pages are anonymous and are 2144 * not associated with a file object. 2145 */ 2146void 2147vm_hold_load_pages(struct buf * bp, vm_offset_t from, vm_offset_t to) 2148{ 2149 vm_offset_t pg; 2150 vm_page_t p; 2151 int index; 2152 2153 to = round_page(to); 2154 from = round_page(from); 2155 index = (from - trunc_page(bp->b_data)) >> PAGE_SHIFT; 2156 2157 for (pg = from; pg < to; pg += PAGE_SIZE, index++) { 2158 2159tryagain: 2160 2161 p = vm_page_alloc(kernel_object, 2162 ((pg - VM_MIN_KERNEL_ADDRESS) >> PAGE_SHIFT), 2163 VM_ALLOC_NORMAL); 2164 if (!p) { 2165 VM_WAIT; 2166 goto tryagain; 2167 } 2168 vm_page_wire(p); 2169 pmap_kenter(pg, VM_PAGE_TO_PHYS(p)); 2170 bp->b_pages[index] = p; 2171 PAGE_WAKEUP(p); 2172 } 2173 bp->b_npages = index; 2174} 2175 2176void 2177vm_hold_free_pages(struct buf * bp, vm_offset_t from, vm_offset_t to) 2178{ 2179 vm_offset_t pg; 2180 vm_page_t p; 2181 int index, newnpages; 2182 2183 from = round_page(from); 2184 to = round_page(to); 2185 newnpages = index = (from - trunc_page(bp->b_data)) >> PAGE_SHIFT; 2186 2187 for (pg = from; pg < to; pg += PAGE_SIZE, index++) { 2188 p = bp->b_pages[index]; 2189 if (p && (index < bp->b_npages)) { 2190#if !defined(MAX_PERF) 2191 if (p->busy) { 2192 printf("vm_hold_free_pages: blkno: %d, lblkno: %d\n", 2193 bp->b_blkno, bp->b_lblkno); 2194 } 2195#endif 2196 bp->b_pages[index] = NULL; 2197 pmap_kremove(pg); 2198 vm_page_unwire(p); 2199 vm_page_free(p); 2200 } 2201 } 2202 bp->b_npages = newnpages; 2203} 2204 2205 2206#include "opt_ddb.h" 2207#ifdef DDB 2208#include <ddb/ddb.h> 2209 2210DB_SHOW_COMMAND(buffer, db_show_buffer) 2211{ 2212 /* get args */ 2213 struct buf *bp = (struct buf *)addr; 2214 2215 if (!have_addr) { 2216 db_printf("usage: show buffer <addr>\n"); 2217 return; 2218 } 2219 2220 db_printf("b_proc = %p,\nb_flags = 0x%b\n", (void *)bp->b_proc, 2221 bp->b_flags, "\20\40bounce\37cluster\36vmio\35ram\34ordered" 2222 "\33paging\32xxx\31writeinprog\30wanted\27relbuf\26tape" 2223 "\25read\24raw\23phys\22clusterok\21malloc\20nocache" 2224 "\17locked\16inval\15gathered\14error\13eintr\12done\11dirty" 2225 "\10delwri\7call\6cache\5busy\4bad\3async\2needcommit\1age"); 2226 db_printf("b_error = %d, b_bufsize = %ld, b_bcount = %ld, " 2227 "b_resid = %ld\nb_dev = 0x%x, b_un.b_addr = %p, " 2228 "b_blkno = %d, b_pblkno = %d\n", 2229 bp->b_error, bp->b_bufsize, bp->b_bcount, bp->b_resid, 2230 bp->b_dev, bp->b_un.b_addr, bp->b_blkno, bp->b_pblkno); 2231 if (bp->b_npages) { 2232 int i; 2233 db_printf("b_npages = %d, pages(OBJ, IDX, PA): ", bp->b_npages); 2234 for (i = 0; i < bp->b_npages; i++) { 2235 vm_page_t m; 2236 m = bp->b_pages[i]; 2237 db_printf("(0x%x, 0x%x, 0x%x)", m->object, m->pindex, 2238 VM_PAGE_TO_PHYS(m)); 2239 if ((i + 1) < bp->b_npages) 2240 db_printf(","); 2241 } 2242 db_printf("\n"); 2243 } 2244} 2245#endif /* DDB */ 2246