vfs_bio.c revision 42007
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 1.187 1998/12/14 21:17:37 dillon Exp $ 15 */ 16 17/* 18 * this file contains a new buffer I/O scheme implementing a coherent 19 * VM object and buffer cache scheme. Pains have been taken to make 20 * sure that the performance degradation associated with schemes such 21 * as this is not realized. 22 * 23 * Author: John S. Dyson 24 * Significant help during the development and debugging phases 25 * had been provided by David Greenman, also of the FreeBSD core team. 26 */ 27 28#define VMIO 29#include <sys/param.h> 30#include <sys/systm.h> 31#include <sys/sysproto.h> 32#include <sys/kernel.h> 33#include <sys/sysctl.h> 34#include <sys/proc.h> 35#include <sys/vnode.h> 36#include <sys/vmmeter.h> 37#include <sys/lock.h> 38#include <miscfs/specfs/specdev.h> 39#include <vm/vm.h> 40#include <vm/vm_param.h> 41#include <vm/vm_prot.h> 42#include <vm/vm_kern.h> 43#include <vm/vm_pageout.h> 44#include <vm/vm_page.h> 45#include <vm/vm_object.h> 46#include <vm/vm_extern.h> 47#include <vm/vm_map.h> 48#include <sys/buf.h> 49#include <sys/mount.h> 50#include <sys/malloc.h> 51#include <sys/resourcevar.h> 52 53static MALLOC_DEFINE(M_BIOBUF, "BIO buffer", "BIO buffer"); 54 55struct bio_ops bioops; /* I/O operation notification */ 56 57#if 0 /* replaced bu sched_sync */ 58static void vfs_update __P((void)); 59static struct proc *updateproc; 60static struct kproc_desc up_kp = { 61 "update", 62 vfs_update, 63 &updateproc 64}; 65SYSINIT_KT(update, SI_SUB_KTHREAD_UPDATE, SI_ORDER_FIRST, kproc_start, &up_kp) 66#endif 67 68struct buf *buf; /* buffer header pool */ 69struct swqueue bswlist; 70 71static void vm_hold_free_pages(struct buf * bp, vm_offset_t from, 72 vm_offset_t to); 73static void vm_hold_load_pages(struct buf * bp, vm_offset_t from, 74 vm_offset_t to); 75static void vfs_buf_set_valid(struct buf *bp, vm_ooffset_t foff, 76 vm_offset_t off, vm_offset_t size, 77 vm_page_t m); 78static void vfs_page_set_valid(struct buf *bp, vm_ooffset_t off, 79 int pageno, vm_page_t m); 80static void vfs_clean_pages(struct buf * bp); 81static void vfs_setdirty(struct buf *bp); 82static void vfs_vmio_release(struct buf *bp); 83static void flushdirtybuffers(int slpflag, int slptimeo); 84 85int needsbuffer; 86 87/* 88 * Internal update daemon, process 3 89 * The variable vfs_update_wakeup allows for internal syncs. 90 */ 91int vfs_update_wakeup; 92 93 94/* 95 * buffers base kva 96 */ 97 98/* 99 * bogus page -- for I/O to/from partially complete buffers 100 * this is a temporary solution to the problem, but it is not 101 * really that bad. it would be better to split the buffer 102 * for input in the case of buffers partially already in memory, 103 * but the code is intricate enough already. 104 */ 105vm_page_t bogus_page; 106static vm_offset_t bogus_offset; 107 108static int bufspace, maxbufspace, vmiospace, maxvmiobufspace, 109 bufmallocspace, maxbufmallocspace; 110int numdirtybuffers; 111static int lodirtybuffers, hidirtybuffers; 112static int numfreebuffers, lofreebuffers, hifreebuffers; 113static int kvafreespace; 114 115SYSCTL_INT(_vfs, OID_AUTO, numdirtybuffers, CTLFLAG_RD, 116 &numdirtybuffers, 0, ""); 117SYSCTL_INT(_vfs, OID_AUTO, lodirtybuffers, CTLFLAG_RW, 118 &lodirtybuffers, 0, ""); 119SYSCTL_INT(_vfs, OID_AUTO, hidirtybuffers, CTLFLAG_RW, 120 &hidirtybuffers, 0, ""); 121SYSCTL_INT(_vfs, OID_AUTO, numfreebuffers, CTLFLAG_RD, 122 &numfreebuffers, 0, ""); 123SYSCTL_INT(_vfs, OID_AUTO, lofreebuffers, CTLFLAG_RW, 124 &lofreebuffers, 0, ""); 125SYSCTL_INT(_vfs, OID_AUTO, hifreebuffers, CTLFLAG_RW, 126 &hifreebuffers, 0, ""); 127SYSCTL_INT(_vfs, OID_AUTO, maxbufspace, CTLFLAG_RW, 128 &maxbufspace, 0, ""); 129SYSCTL_INT(_vfs, OID_AUTO, bufspace, CTLFLAG_RD, 130 &bufspace, 0, ""); 131SYSCTL_INT(_vfs, OID_AUTO, maxvmiobufspace, CTLFLAG_RW, 132 &maxvmiobufspace, 0, ""); 133SYSCTL_INT(_vfs, OID_AUTO, vmiospace, CTLFLAG_RD, 134 &vmiospace, 0, ""); 135SYSCTL_INT(_vfs, OID_AUTO, maxmallocbufspace, CTLFLAG_RW, 136 &maxbufmallocspace, 0, ""); 137SYSCTL_INT(_vfs, OID_AUTO, bufmallocspace, CTLFLAG_RD, 138 &bufmallocspace, 0, ""); 139SYSCTL_INT(_vfs, OID_AUTO, kvafreespace, CTLFLAG_RD, 140 &kvafreespace, 0, ""); 141 142static LIST_HEAD(bufhashhdr, buf) bufhashtbl[BUFHSZ], invalhash; 143struct bqueues bufqueues[BUFFER_QUEUES] = {0}; 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_xflags = 0; 183 LIST_INIT(&bp->b_dep); 184 TAILQ_INSERT_TAIL(&bufqueues[QUEUE_EMPTY], bp, b_freelist); 185 LIST_INSERT_HEAD(&invalhash, bp, b_hash); 186 } 187/* 188 * maxbufspace is currently calculated to support all filesystem blocks 189 * to be 8K. If you happen to use a 16K filesystem, the size of the buffer 190 * cache is still the same as it would be for 8K filesystems. This 191 * keeps the size of the buffer cache "in check" for big block filesystems. 192 */ 193 maxbufspace = (nbuf + 8) * DFLTBSIZE; 194/* 195 * reserve 1/3 of the buffers for metadata (VDIR) which might not be VMIO'ed 196 */ 197 maxvmiobufspace = 2 * maxbufspace / 3; 198/* 199 * Limit the amount of malloc memory since it is wired permanently into 200 * the kernel space. Even though this is accounted for in the buffer 201 * allocation, we don't want the malloced region to grow uncontrolled. 202 * The malloc scheme improves memory utilization significantly on average 203 * (small) directories. 204 */ 205 maxbufmallocspace = maxbufspace / 20; 206 207/* 208 * Remove the probability of deadlock conditions by limiting the 209 * number of dirty buffers. 210 */ 211 hidirtybuffers = nbuf / 8 + 20; 212 lodirtybuffers = nbuf / 16 + 10; 213 numdirtybuffers = 0; 214 lofreebuffers = nbuf / 18 + 5; 215 hifreebuffers = 2 * lofreebuffers; 216 numfreebuffers = nbuf; 217 kvafreespace = 0; 218 219 bogus_offset = kmem_alloc_pageable(kernel_map, PAGE_SIZE); 220 bogus_page = vm_page_alloc(kernel_object, 221 ((bogus_offset - VM_MIN_KERNEL_ADDRESS) >> PAGE_SHIFT), 222 VM_ALLOC_NORMAL); 223 224} 225 226/* 227 * Free the kva allocation for a buffer 228 * Must be called only at splbio or higher, 229 * as this is the only locking for buffer_map. 230 */ 231static void 232bfreekva(struct buf * bp) 233{ 234 if (bp->b_kvasize == 0) 235 return; 236 237 vm_map_delete(buffer_map, 238 (vm_offset_t) bp->b_kvabase, 239 (vm_offset_t) bp->b_kvabase + bp->b_kvasize); 240 241 bp->b_kvasize = 0; 242 243} 244 245/* 246 * remove the buffer from the appropriate free list 247 */ 248void 249bremfree(struct buf * bp) 250{ 251 int s = splbio(); 252 253 if (bp->b_qindex != QUEUE_NONE) { 254 if (bp->b_qindex == QUEUE_EMPTY) { 255 kvafreespace -= bp->b_kvasize; 256 } 257 TAILQ_REMOVE(&bufqueues[bp->b_qindex], bp, b_freelist); 258 bp->b_qindex = QUEUE_NONE; 259 } else { 260#if !defined(MAX_PERF) 261 panic("bremfree: removing a buffer when not on a queue"); 262#endif 263 } 264 if ((bp->b_flags & B_INVAL) || 265 (bp->b_flags & (B_DELWRI|B_LOCKED)) == 0) 266 --numfreebuffers; 267 splx(s); 268} 269 270 271/* 272 * Get a buffer with the specified data. Look in the cache first. 273 */ 274int 275bread(struct vnode * vp, daddr_t blkno, int size, struct ucred * cred, 276 struct buf ** bpp) 277{ 278 struct buf *bp; 279 280 bp = getblk(vp, blkno, size, 0, 0); 281 *bpp = bp; 282 283 /* if not found in cache, do some I/O */ 284 if ((bp->b_flags & B_CACHE) == 0) { 285 if (curproc != NULL) 286 curproc->p_stats->p_ru.ru_inblock++; 287 bp->b_flags |= B_READ; 288 bp->b_flags &= ~(B_DONE | B_ERROR | B_INVAL); 289 if (bp->b_rcred == NOCRED) { 290 if (cred != NOCRED) 291 crhold(cred); 292 bp->b_rcred = cred; 293 } 294 vfs_busy_pages(bp, 0); 295 VOP_STRATEGY(vp, bp); 296 return (biowait(bp)); 297 } 298 return (0); 299} 300 301/* 302 * Operates like bread, but also starts asynchronous I/O on 303 * read-ahead blocks. 304 */ 305int 306breadn(struct vnode * vp, daddr_t blkno, int size, 307 daddr_t * rablkno, int *rabsize, 308 int cnt, struct ucred * cred, struct buf ** bpp) 309{ 310 struct buf *bp, *rabp; 311 int i; 312 int rv = 0, readwait = 0; 313 314 *bpp = bp = getblk(vp, blkno, size, 0, 0); 315 316 /* if not found in cache, do some I/O */ 317 if ((bp->b_flags & B_CACHE) == 0) { 318 if (curproc != NULL) 319 curproc->p_stats->p_ru.ru_inblock++; 320 bp->b_flags |= B_READ; 321 bp->b_flags &= ~(B_DONE | B_ERROR | B_INVAL); 322 if (bp->b_rcred == NOCRED) { 323 if (cred != NOCRED) 324 crhold(cred); 325 bp->b_rcred = cred; 326 } 327 vfs_busy_pages(bp, 0); 328 VOP_STRATEGY(vp, bp); 329 ++readwait; 330 } 331 for (i = 0; i < cnt; i++, rablkno++, rabsize++) { 332 if (inmem(vp, *rablkno)) 333 continue; 334 rabp = getblk(vp, *rablkno, *rabsize, 0, 0); 335 336 if ((rabp->b_flags & B_CACHE) == 0) { 337 if (curproc != NULL) 338 curproc->p_stats->p_ru.ru_inblock++; 339 rabp->b_flags |= B_READ | B_ASYNC; 340 rabp->b_flags &= ~(B_DONE | B_ERROR | B_INVAL); 341 if (rabp->b_rcred == NOCRED) { 342 if (cred != NOCRED) 343 crhold(cred); 344 rabp->b_rcred = cred; 345 } 346 vfs_busy_pages(rabp, 0); 347 VOP_STRATEGY(vp, rabp); 348 } else { 349 brelse(rabp); 350 } 351 } 352 353 if (readwait) { 354 rv = biowait(bp); 355 } 356 return (rv); 357} 358 359/* 360 * Write, release buffer on completion. (Done by iodone 361 * if async.) 362 */ 363int 364bwrite(struct buf * bp) 365{ 366 int oldflags, s; 367 struct vnode *vp; 368 struct mount *mp; 369 370 371 if (bp->b_flags & B_INVAL) { 372 brelse(bp); 373 return (0); 374 } 375 376 oldflags = bp->b_flags; 377 378#if !defined(MAX_PERF) 379 if ((bp->b_flags & B_BUSY) == 0) 380 panic("bwrite: buffer is not busy???"); 381#endif 382 383 bp->b_flags &= ~(B_READ | B_DONE | B_ERROR | B_DELWRI); 384 bp->b_flags |= B_WRITEINPROG; 385 386 s = splbio(); 387 if ((oldflags & B_DELWRI) == B_DELWRI) { 388 --numdirtybuffers; 389 reassignbuf(bp, bp->b_vp); 390 } 391 392 bp->b_vp->v_numoutput++; 393 vfs_busy_pages(bp, 1); 394 if (curproc != NULL) 395 curproc->p_stats->p_ru.ru_oublock++; 396 splx(s); 397 VOP_STRATEGY(bp->b_vp, bp); 398 399 /* 400 * Collect statistics on synchronous and asynchronous writes. 401 * Writes to block devices are charged to their associated 402 * filesystem (if any). 403 */ 404 if ((vp = bp->b_vp) != NULL) { 405 if (vp->v_type == VBLK) 406 mp = vp->v_specmountpoint; 407 else 408 mp = vp->v_mount; 409 if (mp != NULL) 410 if ((oldflags & B_ASYNC) == 0) 411 mp->mnt_stat.f_syncwrites++; 412 else 413 mp->mnt_stat.f_asyncwrites++; 414 } 415 416 if ((oldflags & B_ASYNC) == 0) { 417 int rtval = biowait(bp); 418 brelse(bp); 419 return (rtval); 420 } 421 return (0); 422} 423 424void 425vfs_bio_need_satisfy(void) { 426 ++numfreebuffers; 427 if (!needsbuffer) 428 return; 429 if (numdirtybuffers < lodirtybuffers) { 430 needsbuffer &= ~(VFS_BIO_NEED_ANY | VFS_BIO_NEED_LOWLIMIT); 431 } else { 432 needsbuffer &= ~VFS_BIO_NEED_ANY; 433 } 434 if (numfreebuffers >= hifreebuffers) { 435 needsbuffer &= ~VFS_BIO_NEED_FREE; 436 } 437 wakeup(&needsbuffer); 438} 439 440/* 441 * Delayed write. (Buffer is marked dirty). 442 */ 443void 444bdwrite(struct buf * bp) 445{ 446 struct vnode *vp; 447 448#if !defined(MAX_PERF) 449 if ((bp->b_flags & B_BUSY) == 0) { 450 panic("bdwrite: buffer is not busy"); 451 } 452#endif 453 454 if (bp->b_flags & B_INVAL) { 455 brelse(bp); 456 return; 457 } 458 bp->b_flags &= ~(B_READ|B_RELBUF); 459 if ((bp->b_flags & B_DELWRI) == 0) { 460 bp->b_flags |= B_DONE | B_DELWRI; 461 reassignbuf(bp, bp->b_vp); 462 ++numdirtybuffers; 463 } 464 465 /* 466 * This bmap keeps the system from needing to do the bmap later, 467 * perhaps when the system is attempting to do a sync. Since it 468 * is likely that the indirect block -- or whatever other datastructure 469 * that the filesystem needs is still in memory now, it is a good 470 * thing to do this. Note also, that if the pageout daemon is 471 * requesting a sync -- there might not be enough memory to do 472 * the bmap then... So, this is important to do. 473 */ 474 if (bp->b_lblkno == bp->b_blkno) { 475 VOP_BMAP(bp->b_vp, bp->b_lblkno, NULL, &bp->b_blkno, NULL, NULL); 476 } 477 478 /* 479 * Set the *dirty* buffer range based upon the VM system dirty pages. 480 */ 481 vfs_setdirty(bp); 482 483 /* 484 * We need to do this here to satisfy the vnode_pager and the 485 * pageout daemon, so that it thinks that the pages have been 486 * "cleaned". Note that since the pages are in a delayed write 487 * buffer -- the VFS layer "will" see that the pages get written 488 * out on the next sync, or perhaps the cluster will be completed. 489 */ 490 vfs_clean_pages(bp); 491 bqrelse(bp); 492 493 /* 494 * XXX The soft dependency code is not prepared to 495 * have I/O done when a bdwrite is requested. For 496 * now we just let the write be delayed if it is 497 * requested by the soft dependency code. 498 */ 499 if ((vp = bp->b_vp) && 500 ((vp->v_type == VBLK && vp->v_specmountpoint && 501 (vp->v_specmountpoint->mnt_flag & MNT_SOFTDEP)) || 502 (vp->v_mount && (vp->v_mount->mnt_flag & MNT_SOFTDEP)))) 503 return; 504 505 if (numdirtybuffers >= hidirtybuffers) 506 flushdirtybuffers(0, 0); 507 508 return; 509} 510 511 512/* 513 * Same as first half of bdwrite, mark buffer dirty, but do not release it. 514 * Check how this compares with vfs_setdirty(); XXX [JRE] 515 */ 516void 517bdirty(bp) 518 struct buf *bp; 519{ 520 521 bp->b_flags &= ~(B_READ|B_RELBUF); /* XXX ??? check this */ 522 if ((bp->b_flags & B_DELWRI) == 0) { 523 bp->b_flags |= B_DONE | B_DELWRI; /* why done? XXX JRE */ 524 reassignbuf(bp, bp->b_vp); 525 ++numdirtybuffers; 526 } 527} 528 529/* 530 * Asynchronous write. 531 * Start output on a buffer, but do not wait for it to complete. 532 * The buffer is released when the output completes. 533 */ 534void 535bawrite(struct buf * bp) 536{ 537 bp->b_flags |= B_ASYNC; 538 (void) VOP_BWRITE(bp); 539} 540 541/* 542 * Ordered write. 543 * Start output on a buffer, and flag it so that the device will write 544 * it in the order it was queued. The buffer is released when the output 545 * completes. 546 */ 547int 548bowrite(struct buf * bp) 549{ 550 bp->b_flags |= B_ORDERED|B_ASYNC; 551 return (VOP_BWRITE(bp)); 552} 553 554/* 555 * Release a buffer. 556 */ 557void 558brelse(struct buf * bp) 559{ 560 int s; 561 562 if (bp->b_flags & B_CLUSTER) { 563 relpbuf(bp); 564 return; 565 } 566 567 s = splbio(); 568 569 /* anyone need this block? */ 570 if (bp->b_flags & B_WANTED) { 571 bp->b_flags &= ~(B_WANTED | B_AGE); 572 wakeup(bp); 573 } 574 575 if (bp->b_flags & B_LOCKED) 576 bp->b_flags &= ~B_ERROR; 577 578 if ((bp->b_flags & (B_NOCACHE | B_INVAL | B_ERROR | B_FREEBUF)) || 579 (bp->b_bufsize <= 0)) { 580 bp->b_flags |= B_INVAL; 581 if (LIST_FIRST(&bp->b_dep) != NULL && bioops.io_deallocate) 582 (*bioops.io_deallocate)(bp); 583 if (bp->b_flags & B_DELWRI) 584 --numdirtybuffers; 585 bp->b_flags &= ~(B_DELWRI | B_CACHE | B_FREEBUF); 586 if ((bp->b_flags & B_VMIO) == 0) { 587 if (bp->b_bufsize) 588 allocbuf(bp, 0); 589 if (bp->b_vp) 590 brelvp(bp); 591 } 592 } 593 594 /* 595 * We must clear B_RELBUF if B_DELWRI is set. If vfs_vmio_release() 596 * is called with B_DELWRI set, the underlying pages may wind up 597 * getting freed causing a previous write (bdwrite()) to get 'lost' 598 * because pages associated with a B_DELWRI bp are marked clean. 599 * 600 * We still allow the B_INVAL case to call vfs_vmio_release(), even 601 * if B_DELWRI is set. 602 */ 603 604 if (bp->b_flags & B_DELWRI) 605 bp->b_flags &= ~B_RELBUF; 606 607 /* 608 * VMIO buffer rundown. It is not very necessary to keep a VMIO buffer 609 * constituted, so the B_INVAL flag is used to *invalidate* the buffer, 610 * but the VM object is kept around. The B_NOCACHE flag is used to 611 * invalidate the pages in the VM object. 612 * 613 * If the buffer is a partially filled NFS buffer, keep it 614 * since invalidating it now will lose informatio. The valid 615 * flags in the vm_pages have only DEV_BSIZE resolution but 616 * the b_validoff, b_validend fields have byte resolution. 617 * This can avoid unnecessary re-reads of the buffer. 618 * XXX this seems to cause performance problems. 619 */ 620 if ((bp->b_flags & B_VMIO) 621 && !(bp->b_vp->v_tag == VT_NFS && 622 bp->b_vp->v_type != VBLK && 623 (bp->b_flags & B_DELWRI) != 0) 624#ifdef notdef 625 && (bp->b_vp->v_tag != VT_NFS 626 || bp->b_vp->v_type == VBLK 627 || (bp->b_flags & (B_NOCACHE | B_INVAL | B_ERROR)) 628 || bp->b_validend == 0 629 || (bp->b_validoff == 0 630 && bp->b_validend == bp->b_bufsize)) 631#endif 632 ) { 633 634 int i, j, resid; 635 vm_page_t m; 636 off_t foff; 637 vm_pindex_t poff; 638 vm_object_t obj; 639 struct vnode *vp; 640 641 vp = bp->b_vp; 642 643 resid = bp->b_bufsize; 644 foff = bp->b_offset; 645 646 for (i = 0; i < bp->b_npages; i++) { 647 m = bp->b_pages[i]; 648 vm_page_flag_clear(m, PG_ZERO); 649 if (m == bogus_page) { 650 651 obj = (vm_object_t) vp->v_object; 652 poff = OFF_TO_IDX(bp->b_offset); 653 654 for (j = i; j < bp->b_npages; j++) { 655 m = bp->b_pages[j]; 656 if (m == bogus_page) { 657 m = vm_page_lookup(obj, poff + j); 658#if !defined(MAX_PERF) 659 if (!m) { 660 panic("brelse: page missing\n"); 661 } 662#endif 663 bp->b_pages[j] = m; 664 } 665 } 666 667 if ((bp->b_flags & B_INVAL) == 0) { 668 pmap_qenter(trunc_page((vm_offset_t)bp->b_data), bp->b_pages, bp->b_npages); 669 } 670 } 671 if (bp->b_flags & (B_NOCACHE|B_ERROR)) { 672 int poffset = foff & PAGE_MASK; 673 int presid = resid > (PAGE_SIZE - poffset) ? 674 (PAGE_SIZE - poffset) : resid; 675#ifdef DIAGNOSTIC 676 if (presid < 0) 677 panic("brelse: extra page"); 678#endif 679 vm_page_set_invalid(m, poffset, presid); 680 } 681 resid -= PAGE_SIZE - (foff & PAGE_MASK); 682 foff = (foff + PAGE_SIZE) & ~PAGE_MASK; 683 } 684 685 if (bp->b_flags & (B_INVAL | B_RELBUF)) 686 vfs_vmio_release(bp); 687 688 } else if (bp->b_flags & B_VMIO) { 689 690 if (bp->b_flags & (B_INVAL | B_RELBUF)) 691 vfs_vmio_release(bp); 692 693 } 694 695#if !defined(MAX_PERF) 696 if (bp->b_qindex != QUEUE_NONE) 697 panic("brelse: free buffer onto another queue???"); 698#endif 699 700 /* enqueue */ 701 /* buffers with no memory */ 702 if (bp->b_bufsize == 0) { 703 bp->b_flags |= B_INVAL; 704 bp->b_qindex = QUEUE_EMPTY; 705 TAILQ_INSERT_HEAD(&bufqueues[QUEUE_EMPTY], bp, b_freelist); 706 LIST_REMOVE(bp, b_hash); 707 LIST_INSERT_HEAD(&invalhash, bp, b_hash); 708 bp->b_dev = NODEV; 709 kvafreespace += bp->b_kvasize; 710 711 /* buffers with junk contents */ 712 } else if (bp->b_flags & (B_ERROR | B_INVAL | B_NOCACHE | B_RELBUF)) { 713 bp->b_flags |= B_INVAL; 714 bp->b_qindex = QUEUE_AGE; 715 TAILQ_INSERT_HEAD(&bufqueues[QUEUE_AGE], bp, b_freelist); 716 LIST_REMOVE(bp, b_hash); 717 LIST_INSERT_HEAD(&invalhash, bp, b_hash); 718 bp->b_dev = NODEV; 719 720 /* buffers that are locked */ 721 } else if (bp->b_flags & B_LOCKED) { 722 bp->b_qindex = QUEUE_LOCKED; 723 TAILQ_INSERT_TAIL(&bufqueues[QUEUE_LOCKED], bp, b_freelist); 724 725 /* buffers with stale but valid contents */ 726 } else if (bp->b_flags & B_AGE) { 727 bp->b_qindex = QUEUE_AGE; 728 TAILQ_INSERT_TAIL(&bufqueues[QUEUE_AGE], bp, b_freelist); 729 730 /* buffers with valid and quite potentially reuseable 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_INVAL) || 737 (bp->b_flags & (B_LOCKED|B_DELWRI)) == 0) { 738 if (bp->b_flags & B_DELWRI) { 739 --numdirtybuffers; 740 bp->b_flags &= ~B_DELWRI; 741 } 742 vfs_bio_need_satisfy(); 743 } 744 745 /* unlock */ 746 bp->b_flags &= ~(B_ORDERED | B_WANTED | B_BUSY | 747 B_ASYNC | B_NOCACHE | B_AGE | B_RELBUF); 748 splx(s); 749} 750 751/* 752 * Release a buffer. 753 */ 754void 755bqrelse(struct buf * bp) 756{ 757 int s; 758 759 s = splbio(); 760 761 /* anyone need this block? */ 762 if (bp->b_flags & B_WANTED) { 763 bp->b_flags &= ~(B_WANTED | B_AGE); 764 wakeup(bp); 765 } 766 767#if !defined(MAX_PERF) 768 if (bp->b_qindex != QUEUE_NONE) 769 panic("bqrelse: free buffer onto another queue???"); 770#endif 771 772 if (bp->b_flags & B_LOCKED) { 773 bp->b_flags &= ~B_ERROR; 774 bp->b_qindex = QUEUE_LOCKED; 775 TAILQ_INSERT_TAIL(&bufqueues[QUEUE_LOCKED], bp, b_freelist); 776 /* buffers with stale but valid contents */ 777 } else { 778 bp->b_qindex = QUEUE_LRU; 779 TAILQ_INSERT_TAIL(&bufqueues[QUEUE_LRU], bp, b_freelist); 780 } 781 782 if ((bp->b_flags & (B_LOCKED|B_DELWRI)) == 0) { 783 vfs_bio_need_satisfy(); 784 } 785 786 /* unlock */ 787 bp->b_flags &= ~(B_ORDERED | B_WANTED | B_BUSY | 788 B_ASYNC | B_NOCACHE | B_AGE | B_RELBUF); 789 splx(s); 790} 791 792static void 793vfs_vmio_release(bp) 794 struct buf *bp; 795{ 796 int i, s; 797 vm_page_t m; 798 799 s = splvm(); 800 for (i = 0; i < bp->b_npages; i++) { 801 m = bp->b_pages[i]; 802 bp->b_pages[i] = NULL; 803 /* 804 * In order to keep page LRU ordering consistent, put 805 * everything on the inactive queue. 806 */ 807 vm_page_unwire(m, 0); 808 /* 809 * We don't mess with busy pages, it is 810 * the responsibility of the process that 811 * busied the pages to deal with them. 812 */ 813 if ((m->flags & PG_BUSY) || (m->busy != 0)) 814 continue; 815 816 if (m->wire_count == 0) { 817 vm_page_flag_clear(m, PG_ZERO); 818 /* 819 * Might as well free the page if we can and it has 820 * no valid data. 821 */ 822 if ((bp->b_flags & B_ASYNC) == 0 && !m->valid && m->hold_count == 0) { 823 vm_page_busy(m); 824 vm_page_protect(m, VM_PROT_NONE); 825 vm_page_free(m); 826 } 827 } 828 } 829 splx(s); 830 bufspace -= bp->b_bufsize; 831 vmiospace -= bp->b_bufsize; 832 pmap_qremove(trunc_page((vm_offset_t) bp->b_data), bp->b_npages); 833 bp->b_npages = 0; 834 bp->b_bufsize = 0; 835 bp->b_flags &= ~B_VMIO; 836 if (bp->b_vp) 837 brelvp(bp); 838} 839 840/* 841 * Check to see if a block is currently memory resident. 842 */ 843struct buf * 844gbincore(struct vnode * vp, daddr_t blkno) 845{ 846 struct buf *bp; 847 struct bufhashhdr *bh; 848 849 bh = BUFHASH(vp, blkno); 850 bp = bh->lh_first; 851 852 /* Search hash chain */ 853 while (bp != NULL) { 854 /* hit */ 855 if (bp->b_vp == vp && bp->b_lblkno == blkno && 856 (bp->b_flags & B_INVAL) == 0) { 857 break; 858 } 859 bp = bp->b_hash.le_next; 860 } 861 return (bp); 862} 863 864/* 865 * this routine implements clustered async writes for 866 * clearing out B_DELWRI buffers... This is much better 867 * than the old way of writing only one buffer at a time. 868 */ 869int 870vfs_bio_awrite(struct buf * bp) 871{ 872 int i; 873 daddr_t lblkno = bp->b_lblkno; 874 struct vnode *vp = bp->b_vp; 875 int s; 876 int ncl; 877 struct buf *bpa; 878 int nwritten; 879 int size; 880 int maxcl; 881 882 s = splbio(); 883 /* 884 * right now we support clustered writing only to regular files 885 */ 886 if ((vp->v_type == VREG) && 887 (vp->v_mount != 0) && /* Only on nodes that have the size info */ 888 (bp->b_flags & (B_CLUSTEROK | B_INVAL)) == B_CLUSTEROK) { 889 890 size = vp->v_mount->mnt_stat.f_iosize; 891 maxcl = MAXPHYS / size; 892 893 for (i = 1; i < maxcl; i++) { 894 if ((bpa = gbincore(vp, lblkno + i)) && 895 ((bpa->b_flags & (B_BUSY | B_DELWRI | B_CLUSTEROK | B_INVAL)) == 896 (B_DELWRI | B_CLUSTEROK)) && 897 (bpa->b_bufsize == size)) { 898 if ((bpa->b_blkno == bpa->b_lblkno) || 899 (bpa->b_blkno != bp->b_blkno + ((i * size) >> DEV_BSHIFT))) 900 break; 901 } else { 902 break; 903 } 904 } 905 ncl = i; 906 /* 907 * this is a possible cluster write 908 */ 909 if (ncl != 1) { 910 nwritten = cluster_wbuild(vp, size, lblkno, ncl); 911 splx(s); 912 return nwritten; 913 } 914 } 915 916 bremfree(bp); 917 bp->b_flags |= B_BUSY | B_ASYNC; 918 919 splx(s); 920 /* 921 * default (old) behavior, writing out only one block 922 */ 923 nwritten = bp->b_bufsize; 924 (void) VOP_BWRITE(bp); 925 return nwritten; 926} 927 928 929/* 930 * Find a buffer header which is available for use. 931 */ 932static struct buf * 933getnewbuf(struct vnode *vp, daddr_t blkno, 934 int slpflag, int slptimeo, int size, int maxsize) 935{ 936 struct buf *bp, *bp1; 937 int nbyteswritten = 0; 938 vm_offset_t addr; 939 static int writerecursion = 0; 940 941start: 942 if (bufspace >= maxbufspace) 943 goto trytofreespace; 944 945 /* can we constitute a new buffer? */ 946 if ((bp = TAILQ_FIRST(&bufqueues[QUEUE_EMPTY]))) { 947#if !defined(MAX_PERF) 948 if (bp->b_qindex != QUEUE_EMPTY) 949 panic("getnewbuf: inconsistent EMPTY queue, qindex=%d", 950 bp->b_qindex); 951#endif 952 bp->b_flags |= B_BUSY; 953 bremfree(bp); 954 goto fillbuf; 955 } 956trytofreespace: 957 /* 958 * We keep the file I/O from hogging metadata I/O 959 * This is desirable because file data is cached in the 960 * VM/Buffer cache even if a buffer is freed. 961 */ 962 if ((bp = TAILQ_FIRST(&bufqueues[QUEUE_AGE]))) { 963#if !defined(MAX_PERF) 964 if (bp->b_qindex != QUEUE_AGE) 965 panic("getnewbuf: inconsistent AGE queue, qindex=%d", 966 bp->b_qindex); 967#endif 968 } else if ((bp = TAILQ_FIRST(&bufqueues[QUEUE_LRU]))) { 969#if !defined(MAX_PERF) 970 if (bp->b_qindex != QUEUE_LRU) 971 panic("getnewbuf: inconsistent LRU queue, qindex=%d", 972 bp->b_qindex); 973#endif 974 } 975 if (!bp) { 976 /* wait for a free buffer of any kind */ 977 needsbuffer |= VFS_BIO_NEED_ANY; 978 do 979 tsleep(&needsbuffer, (PRIBIO + 4) | slpflag, "newbuf", 980 slptimeo); 981 while (needsbuffer & VFS_BIO_NEED_ANY); 982 return (0); 983 } 984 985#if defined(DIAGNOSTIC) 986 if (bp->b_flags & B_BUSY) { 987 panic("getnewbuf: busy buffer on free list\n"); 988 } 989#endif 990 991 /* 992 * We are fairly aggressive about freeing VMIO buffers, but since 993 * the buffering is intact without buffer headers, there is not 994 * much loss. We gain by maintaining non-VMIOed metadata in buffers. 995 */ 996 if ((bp->b_qindex == QUEUE_LRU) && (bp->b_usecount > 0)) { 997 if ((bp->b_flags & B_VMIO) == 0 || 998 (vmiospace < maxvmiobufspace)) { 999 --bp->b_usecount; 1000 TAILQ_REMOVE(&bufqueues[QUEUE_LRU], bp, b_freelist); 1001 if (TAILQ_FIRST(&bufqueues[QUEUE_LRU]) != NULL) { 1002 TAILQ_INSERT_TAIL(&bufqueues[QUEUE_LRU], bp, b_freelist); 1003 goto start; 1004 } 1005 TAILQ_INSERT_TAIL(&bufqueues[QUEUE_LRU], bp, b_freelist); 1006 } 1007 } 1008 1009 1010 /* if we are a delayed write, convert to an async write */ 1011 if ((bp->b_flags & (B_DELWRI | B_INVAL)) == B_DELWRI) { 1012 1013 /* 1014 * If our delayed write is likely to be used soon, then 1015 * recycle back onto the LRU queue. 1016 */ 1017 if (vp && (bp->b_vp == vp) && (bp->b_qindex == QUEUE_LRU) && 1018 (bp->b_lblkno >= blkno) && (maxsize > 0)) { 1019 1020 if (bp->b_usecount > 0) { 1021 if (bp->b_lblkno < blkno + (MAXPHYS / maxsize)) { 1022 1023 TAILQ_REMOVE(&bufqueues[QUEUE_LRU], bp, b_freelist); 1024 1025 if (TAILQ_FIRST(&bufqueues[QUEUE_LRU]) != NULL) { 1026 TAILQ_INSERT_TAIL(&bufqueues[QUEUE_LRU], bp, b_freelist); 1027 bp->b_usecount--; 1028 goto start; 1029 } 1030 TAILQ_INSERT_TAIL(&bufqueues[QUEUE_LRU], bp, b_freelist); 1031 } 1032 } 1033 } 1034 1035 /* 1036 * Certain layered filesystems can recursively re-enter the vfs_bio 1037 * code, due to delayed writes. This helps keep the system from 1038 * deadlocking. 1039 */ 1040 if (writerecursion > 0) { 1041 if (writerecursion > 5) { 1042 bp = TAILQ_FIRST(&bufqueues[QUEUE_AGE]); 1043 while (bp) { 1044 if ((bp->b_flags & B_DELWRI) == 0) 1045 break; 1046 bp = TAILQ_NEXT(bp, b_freelist); 1047 } 1048 if (bp == NULL) { 1049 bp = TAILQ_FIRST(&bufqueues[QUEUE_LRU]); 1050 while (bp) { 1051 if ((bp->b_flags & B_DELWRI) == 0) 1052 break; 1053 bp = TAILQ_NEXT(bp, b_freelist); 1054 } 1055 } 1056 if (bp == NULL) 1057 panic("getnewbuf: cannot get buffer, infinite recursion failure"); 1058 } else { 1059 bremfree(bp); 1060 bp->b_flags |= B_BUSY | B_AGE | B_ASYNC; 1061 nbyteswritten += bp->b_bufsize; 1062 ++writerecursion; 1063 VOP_BWRITE(bp); 1064 --writerecursion; 1065 if (!slpflag && !slptimeo) { 1066 return (0); 1067 } 1068 goto start; 1069 } 1070 } else { 1071 ++writerecursion; 1072 nbyteswritten += vfs_bio_awrite(bp); 1073 --writerecursion; 1074 if (!slpflag && !slptimeo) { 1075 return (0); 1076 } 1077 goto start; 1078 } 1079 } 1080 1081 if (bp->b_flags & B_WANTED) { 1082 bp->b_flags &= ~B_WANTED; 1083 wakeup(bp); 1084 } 1085 bremfree(bp); 1086 bp->b_flags |= B_BUSY; 1087 1088 if (bp->b_flags & B_VMIO) { 1089 bp->b_flags &= ~B_ASYNC; 1090 vfs_vmio_release(bp); 1091 } 1092 1093 if (bp->b_vp) 1094 brelvp(bp); 1095 1096fillbuf: 1097 1098 /* we are not free, nor do we contain interesting data */ 1099 if (bp->b_rcred != NOCRED) { 1100 crfree(bp->b_rcred); 1101 bp->b_rcred = NOCRED; 1102 } 1103 if (bp->b_wcred != NOCRED) { 1104 crfree(bp->b_wcred); 1105 bp->b_wcred = NOCRED; 1106 } 1107 if (LIST_FIRST(&bp->b_dep) != NULL && 1108 bioops.io_deallocate) 1109 (*bioops.io_deallocate)(bp); 1110 1111 LIST_REMOVE(bp, b_hash); 1112 LIST_INSERT_HEAD(&invalhash, bp, b_hash); 1113 if (bp->b_bufsize) { 1114 allocbuf(bp, 0); 1115 } 1116 bp->b_flags = B_BUSY; 1117 bp->b_dev = NODEV; 1118 bp->b_vp = NULL; 1119 bp->b_blkno = bp->b_lblkno = 0; 1120 bp->b_offset = NOOFFSET; 1121 bp->b_iodone = 0; 1122 bp->b_error = 0; 1123 bp->b_resid = 0; 1124 bp->b_bcount = 0; 1125 bp->b_npages = 0; 1126 bp->b_dirtyoff = bp->b_dirtyend = 0; 1127 bp->b_validoff = bp->b_validend = 0; 1128 bp->b_usecount = 5; 1129 /* Here, not kern_physio.c, is where this should be done*/ 1130 LIST_INIT(&bp->b_dep); 1131 1132 maxsize = (maxsize + PAGE_MASK) & ~PAGE_MASK; 1133 1134 /* 1135 * we assume that buffer_map is not at address 0 1136 */ 1137 addr = 0; 1138 if (maxsize != bp->b_kvasize) { 1139 bfreekva(bp); 1140 1141findkvaspace: 1142 /* 1143 * See if we have buffer kva space 1144 */ 1145 if (vm_map_findspace(buffer_map, 1146 vm_map_min(buffer_map), maxsize, &addr)) { 1147 if (kvafreespace > 0) { 1148 int totfree = 0, freed; 1149 do { 1150 freed = 0; 1151 for (bp1 = TAILQ_FIRST(&bufqueues[QUEUE_EMPTY]); 1152 bp1 != NULL; bp1 = TAILQ_NEXT(bp1, b_freelist)) { 1153 if (bp1->b_kvasize != 0) { 1154 totfree += bp1->b_kvasize; 1155 freed = bp1->b_kvasize; 1156 bremfree(bp1); 1157 bfreekva(bp1); 1158 brelse(bp1); 1159 break; 1160 } 1161 } 1162 } while (freed); 1163 /* 1164 * if we found free space, then retry with the same buffer. 1165 */ 1166 if (totfree) 1167 goto findkvaspace; 1168 } 1169 bp->b_flags |= B_INVAL; 1170 brelse(bp); 1171 goto trytofreespace; 1172 } 1173 } 1174 1175 /* 1176 * See if we are below are allocated minimum 1177 */ 1178 if (bufspace >= (maxbufspace + nbyteswritten)) { 1179 bp->b_flags |= B_INVAL; 1180 brelse(bp); 1181 goto trytofreespace; 1182 } 1183 1184 /* 1185 * create a map entry for the buffer -- in essence 1186 * reserving the kva space. 1187 */ 1188 if (addr) { 1189 vm_map_insert(buffer_map, NULL, 0, 1190 addr, addr + maxsize, 1191 VM_PROT_ALL, VM_PROT_ALL, MAP_NOFAULT); 1192 1193 bp->b_kvabase = (caddr_t) addr; 1194 bp->b_kvasize = maxsize; 1195 } 1196 bp->b_data = bp->b_kvabase; 1197 1198 return (bp); 1199} 1200 1201static void 1202waitfreebuffers(int slpflag, int slptimeo) { 1203 while (numfreebuffers < hifreebuffers) { 1204 flushdirtybuffers(slpflag, slptimeo); 1205 if (numfreebuffers < hifreebuffers) 1206 break; 1207 needsbuffer |= VFS_BIO_NEED_FREE; 1208 if (tsleep(&needsbuffer, (PRIBIO + 4)|slpflag, "biofre", slptimeo)) 1209 break; 1210 } 1211} 1212 1213static void 1214flushdirtybuffers(int slpflag, int slptimeo) { 1215 int s; 1216 static pid_t flushing = 0; 1217 1218 s = splbio(); 1219 1220 if (flushing) { 1221 if (flushing == curproc->p_pid) { 1222 splx(s); 1223 return; 1224 } 1225 while (flushing) { 1226 if (tsleep(&flushing, (PRIBIO + 4)|slpflag, "biofls", slptimeo)) { 1227 splx(s); 1228 return; 1229 } 1230 } 1231 } 1232 flushing = curproc->p_pid; 1233 1234 while (numdirtybuffers > lodirtybuffers) { 1235 struct buf *bp; 1236 needsbuffer |= VFS_BIO_NEED_LOWLIMIT; 1237 bp = TAILQ_FIRST(&bufqueues[QUEUE_AGE]); 1238 if (bp == NULL) 1239 bp = TAILQ_FIRST(&bufqueues[QUEUE_LRU]); 1240 1241 while (bp && ((bp->b_flags & B_DELWRI) == 0)) { 1242 bp = TAILQ_NEXT(bp, b_freelist); 1243 } 1244 1245 if (bp) { 1246 vfs_bio_awrite(bp); 1247 continue; 1248 } 1249 break; 1250 } 1251 1252 flushing = 0; 1253 wakeup(&flushing); 1254 splx(s); 1255} 1256 1257/* 1258 * Check to see if a block is currently memory resident. 1259 */ 1260struct buf * 1261incore(struct vnode * vp, daddr_t blkno) 1262{ 1263 struct buf *bp; 1264 1265 int s = splbio(); 1266 bp = gbincore(vp, blkno); 1267 splx(s); 1268 return (bp); 1269} 1270 1271/* 1272 * Returns true if no I/O is needed to access the 1273 * associated VM object. This is like incore except 1274 * it also hunts around in the VM system for the data. 1275 */ 1276 1277int 1278inmem(struct vnode * vp, daddr_t blkno) 1279{ 1280 vm_object_t obj; 1281 vm_offset_t toff, tinc, size; 1282 vm_page_t m; 1283 vm_ooffset_t off; 1284 1285 if (incore(vp, blkno)) 1286 return 1; 1287 if (vp->v_mount == NULL) 1288 return 0; 1289 if ((vp->v_object == NULL) || (vp->v_flag & VOBJBUF) == 0) 1290 return 0; 1291 1292 obj = vp->v_object; 1293 size = PAGE_SIZE; 1294 if (size > vp->v_mount->mnt_stat.f_iosize) 1295 size = vp->v_mount->mnt_stat.f_iosize; 1296 off = (vm_ooffset_t)blkno * (vm_ooffset_t)vp->v_mount->mnt_stat.f_iosize; 1297 1298 for (toff = 0; toff < vp->v_mount->mnt_stat.f_iosize; toff += tinc) { 1299 m = vm_page_lookup(obj, OFF_TO_IDX(off + toff)); 1300 if (!m) 1301 return 0; 1302 tinc = size; 1303 if (tinc > PAGE_SIZE - ((toff + off) & PAGE_MASK)) 1304 tinc = PAGE_SIZE - ((toff + off) & PAGE_MASK); 1305 if (vm_page_is_valid(m, 1306 (vm_offset_t) ((toff + off) & PAGE_MASK), tinc) == 0) 1307 return 0; 1308 } 1309 return 1; 1310} 1311 1312/* 1313 * now we set the dirty range for the buffer -- 1314 * for NFS -- if the file is mapped and pages have 1315 * been written to, let it know. We want the 1316 * entire range of the buffer to be marked dirty if 1317 * any of the pages have been written to for consistancy 1318 * with the b_validoff, b_validend set in the nfs write 1319 * code, and used by the nfs read code. 1320 */ 1321static void 1322vfs_setdirty(struct buf *bp) { 1323 int i; 1324 vm_object_t object; 1325 vm_offset_t boffset, offset; 1326 /* 1327 * We qualify the scan for modified pages on whether the 1328 * object has been flushed yet. The OBJ_WRITEABLE flag 1329 * is not cleared simply by protecting pages off. 1330 */ 1331 if ((bp->b_flags & B_VMIO) && 1332 ((object = bp->b_pages[0]->object)->flags & (OBJ_WRITEABLE|OBJ_CLEANING))) { 1333 /* 1334 * test the pages to see if they have been modified directly 1335 * by users through the VM system. 1336 */ 1337 for (i = 0; i < bp->b_npages; i++) { 1338 vm_page_flag_clear(bp->b_pages[i], PG_ZERO); 1339 vm_page_test_dirty(bp->b_pages[i]); 1340 } 1341 1342 /* 1343 * scan forwards for the first page modified 1344 */ 1345 for (i = 0; i < bp->b_npages; i++) { 1346 if (bp->b_pages[i]->dirty) { 1347 break; 1348 } 1349 } 1350 boffset = (i << PAGE_SHIFT) - (bp->b_offset & PAGE_MASK); 1351 if (boffset < bp->b_dirtyoff) { 1352 bp->b_dirtyoff = max(boffset, 0); 1353 } 1354 1355 /* 1356 * scan backwards for the last page modified 1357 */ 1358 for (i = bp->b_npages - 1; i >= 0; --i) { 1359 if (bp->b_pages[i]->dirty) { 1360 break; 1361 } 1362 } 1363 boffset = (i + 1); 1364#if 0 1365 offset = boffset + bp->b_pages[0]->pindex; 1366 if (offset >= object->size) 1367 boffset = object->size - bp->b_pages[0]->pindex; 1368#endif 1369 boffset = (boffset << PAGE_SHIFT) - (bp->b_offset & PAGE_MASK); 1370 if (bp->b_dirtyend < boffset) 1371 bp->b_dirtyend = min(boffset, bp->b_bufsize); 1372 } 1373} 1374 1375/* 1376 * Get a block given a specified block and offset into a file/device. 1377 */ 1378struct buf * 1379getblk(struct vnode * vp, daddr_t blkno, int size, int slpflag, int slptimeo) 1380{ 1381 struct buf *bp; 1382 int i, s; 1383 struct bufhashhdr *bh; 1384 int maxsize; 1385 1386#if !defined(MAX_PERF) 1387 if (size > MAXBSIZE) 1388 panic("getblk: size(%d) > MAXBSIZE(%d)\n", size, MAXBSIZE); 1389#endif 1390 1391 s = splbio(); 1392loop: 1393 if (numfreebuffers < lofreebuffers) { 1394 waitfreebuffers(slpflag, slptimeo); 1395 } 1396 1397 if ((bp = gbincore(vp, blkno))) { 1398 if (bp->b_flags & B_BUSY) { 1399 1400 bp->b_flags |= B_WANTED; 1401 if (bp->b_usecount < BUF_MAXUSE) 1402 ++bp->b_usecount; 1403 1404 if (!tsleep(bp, 1405 (PRIBIO + 4) | slpflag, "getblk", slptimeo)) { 1406 goto loop; 1407 } 1408 1409 splx(s); 1410 return (struct buf *) NULL; 1411 } 1412 bp->b_flags |= B_BUSY | B_CACHE; 1413 bremfree(bp); 1414 1415 /* 1416 * check for size inconsistancies (note that they shouldn't 1417 * happen but do when filesystems don't handle the size changes 1418 * correctly.) We are conservative on metadata and don't just 1419 * extend the buffer but write (if needed) and re-constitute it. 1420 */ 1421 1422 if (bp->b_bcount != size) { 1423 if ((bp->b_flags & B_VMIO) && (size <= bp->b_kvasize)) { 1424 allocbuf(bp, size); 1425 } else { 1426 if (bp->b_flags & B_DELWRI) { 1427 bp->b_flags |= B_NOCACHE; 1428 VOP_BWRITE(bp); 1429 } else { 1430 if ((bp->b_flags & B_VMIO) && 1431 (LIST_FIRST(&bp->b_dep) == NULL)) { 1432 bp->b_flags |= B_RELBUF; 1433 brelse(bp); 1434 } else { 1435 bp->b_flags |= B_NOCACHE; 1436 VOP_BWRITE(bp); 1437 } 1438 } 1439 goto loop; 1440 } 1441 } 1442 1443#ifdef DIAGNOSTIC 1444 if (bp->b_offset == NOOFFSET) 1445 panic("getblk: no buffer offset"); 1446#endif 1447 1448 /* 1449 * Check that the constituted buffer really deserves for the 1450 * B_CACHE bit to be set. B_VMIO type buffers might not 1451 * contain fully valid pages. Normal (old-style) buffers 1452 * should be fully valid. 1453 */ 1454 if (bp->b_flags & B_VMIO) { 1455 int checksize = bp->b_bufsize; 1456 int poffset = bp->b_offset & PAGE_MASK; 1457 int resid; 1458 for (i = 0; i < bp->b_npages; i++) { 1459 resid = (checksize > (PAGE_SIZE - poffset)) ? 1460 (PAGE_SIZE - poffset) : checksize; 1461 if (!vm_page_is_valid(bp->b_pages[i], poffset, resid)) { 1462 bp->b_flags &= ~(B_CACHE | B_DONE); 1463 break; 1464 } 1465 checksize -= resid; 1466 poffset = 0; 1467 } 1468 } 1469 1470 if (bp->b_usecount < BUF_MAXUSE) 1471 ++bp->b_usecount; 1472 splx(s); 1473 return (bp); 1474 } else { 1475 int bsize, maxsize, vmio; 1476 off_t offset; 1477 1478 if (vp->v_type == VBLK) 1479 bsize = DEV_BSIZE; 1480 else if (vp->v_mountedhere) 1481 bsize = vp->v_mountedhere->mnt_stat.f_iosize; 1482 else if (vp->v_mount) 1483 bsize = vp->v_mount->mnt_stat.f_iosize; 1484 else 1485 bsize = size; 1486 1487 offset = (off_t)blkno * bsize; 1488 vmio = (vp->v_object != 0) && (vp->v_flag & VOBJBUF); 1489 maxsize = vmio ? size + (offset & PAGE_MASK) : size; 1490 maxsize = imax(maxsize, bsize); 1491 1492 if ((bp = getnewbuf(vp, blkno, 1493 slpflag, slptimeo, size, maxsize)) == 0) { 1494 if (slpflag || slptimeo) { 1495 splx(s); 1496 return NULL; 1497 } 1498 goto loop; 1499 } 1500 1501 /* 1502 * This code is used to make sure that a buffer is not 1503 * created while the getnewbuf routine is blocked. 1504 * Normally the vnode is locked so this isn't a problem. 1505 * VBLK type I/O requests, however, don't lock the vnode. 1506 */ 1507 if (VOP_ISLOCKED(vp) != LK_EXCLUSIVE && gbincore(vp, blkno)) { 1508 bp->b_flags |= B_INVAL; 1509 brelse(bp); 1510 goto loop; 1511 } 1512 1513 /* 1514 * Insert the buffer into the hash, so that it can 1515 * be found by incore. 1516 */ 1517 bp->b_blkno = bp->b_lblkno = blkno; 1518 bp->b_offset = offset; 1519 1520 bgetvp(vp, bp); 1521 LIST_REMOVE(bp, b_hash); 1522 bh = BUFHASH(vp, blkno); 1523 LIST_INSERT_HEAD(bh, bp, b_hash); 1524 1525 if (vmio) { 1526 bp->b_flags |= (B_VMIO | B_CACHE); 1527#if defined(VFS_BIO_DEBUG) 1528 if (vp->v_type != VREG && vp->v_type != VBLK) 1529 printf("getblk: vmioing file type %d???\n", vp->v_type); 1530#endif 1531 } else { 1532 bp->b_flags &= ~B_VMIO; 1533 } 1534 1535 allocbuf(bp, size); 1536 1537 splx(s); 1538 return (bp); 1539 } 1540} 1541 1542/* 1543 * Get an empty, disassociated buffer of given size. 1544 */ 1545struct buf * 1546geteblk(int size) 1547{ 1548 struct buf *bp; 1549 int s; 1550 1551 s = splbio(); 1552 while ((bp = getnewbuf(0, (daddr_t) 0, 0, 0, size, MAXBSIZE)) == 0); 1553 splx(s); 1554 allocbuf(bp, size); 1555 bp->b_flags |= B_INVAL; /* b_dep cleared by getnewbuf() */ 1556 return (bp); 1557} 1558 1559 1560/* 1561 * This code constitutes the buffer memory from either anonymous system 1562 * memory (in the case of non-VMIO operations) or from an associated 1563 * VM object (in the case of VMIO operations). 1564 * 1565 * Note that this code is tricky, and has many complications to resolve 1566 * deadlock or inconsistant data situations. Tread lightly!!! 1567 * 1568 * Modify the length of a buffer's underlying buffer storage without 1569 * destroying information (unless, of course the buffer is shrinking). 1570 */ 1571int 1572allocbuf(struct buf * bp, int size) 1573{ 1574 1575 int s; 1576 int newbsize, mbsize; 1577 int i; 1578 1579#if !defined(MAX_PERF) 1580 if (!(bp->b_flags & B_BUSY)) 1581 panic("allocbuf: buffer not busy"); 1582 1583 if (bp->b_kvasize < size) 1584 panic("allocbuf: buffer too small"); 1585#endif 1586 1587 if ((bp->b_flags & B_VMIO) == 0) { 1588 caddr_t origbuf; 1589 int origbufsize; 1590 /* 1591 * Just get anonymous memory from the kernel 1592 */ 1593 mbsize = (size + DEV_BSIZE - 1) & ~(DEV_BSIZE - 1); 1594#if !defined(NO_B_MALLOC) 1595 if (bp->b_flags & B_MALLOC) 1596 newbsize = mbsize; 1597 else 1598#endif 1599 newbsize = round_page(size); 1600 1601 if (newbsize < bp->b_bufsize) { 1602#if !defined(NO_B_MALLOC) 1603 /* 1604 * malloced buffers are not shrunk 1605 */ 1606 if (bp->b_flags & B_MALLOC) { 1607 if (newbsize) { 1608 bp->b_bcount = size; 1609 } else { 1610 free(bp->b_data, M_BIOBUF); 1611 bufspace -= bp->b_bufsize; 1612 bufmallocspace -= bp->b_bufsize; 1613 bp->b_data = bp->b_kvabase; 1614 bp->b_bufsize = 0; 1615 bp->b_bcount = 0; 1616 bp->b_flags &= ~B_MALLOC; 1617 } 1618 return 1; 1619 } 1620#endif 1621 vm_hold_free_pages( 1622 bp, 1623 (vm_offset_t) bp->b_data + newbsize, 1624 (vm_offset_t) bp->b_data + bp->b_bufsize); 1625 } else if (newbsize > bp->b_bufsize) { 1626#if !defined(NO_B_MALLOC) 1627 /* 1628 * We only use malloced memory on the first allocation. 1629 * and revert to page-allocated memory when the buffer grows. 1630 */ 1631 if ( (bufmallocspace < maxbufmallocspace) && 1632 (bp->b_bufsize == 0) && 1633 (mbsize <= PAGE_SIZE/2)) { 1634 1635 bp->b_data = malloc(mbsize, M_BIOBUF, M_WAITOK); 1636 bp->b_bufsize = mbsize; 1637 bp->b_bcount = size; 1638 bp->b_flags |= B_MALLOC; 1639 bufspace += mbsize; 1640 bufmallocspace += mbsize; 1641 return 1; 1642 } 1643#endif 1644 origbuf = NULL; 1645 origbufsize = 0; 1646#if !defined(NO_B_MALLOC) 1647 /* 1648 * If the buffer is growing on its other-than-first allocation, 1649 * then we revert to the page-allocation scheme. 1650 */ 1651 if (bp->b_flags & B_MALLOC) { 1652 origbuf = bp->b_data; 1653 origbufsize = bp->b_bufsize; 1654 bp->b_data = bp->b_kvabase; 1655 bufspace -= bp->b_bufsize; 1656 bufmallocspace -= bp->b_bufsize; 1657 bp->b_bufsize = 0; 1658 bp->b_flags &= ~B_MALLOC; 1659 newbsize = round_page(newbsize); 1660 } 1661#endif 1662 vm_hold_load_pages( 1663 bp, 1664 (vm_offset_t) bp->b_data + bp->b_bufsize, 1665 (vm_offset_t) bp->b_data + newbsize); 1666#if !defined(NO_B_MALLOC) 1667 if (origbuf) { 1668 bcopy(origbuf, bp->b_data, origbufsize); 1669 free(origbuf, M_BIOBUF); 1670 } 1671#endif 1672 } 1673 } else { 1674 vm_page_t m; 1675 int desiredpages; 1676 1677 newbsize = (size + DEV_BSIZE - 1) & ~(DEV_BSIZE - 1); 1678 desiredpages = (size == 0) ? 0 : 1679 num_pages((bp->b_offset & PAGE_MASK) + newbsize); 1680 1681#if !defined(NO_B_MALLOC) 1682 if (bp->b_flags & B_MALLOC) 1683 panic("allocbuf: VMIO buffer can't be malloced"); 1684#endif 1685 1686 if (newbsize < bp->b_bufsize) { 1687 if (desiredpages < bp->b_npages) { 1688 for (i = desiredpages; i < bp->b_npages; i++) { 1689 /* 1690 * the page is not freed here -- it 1691 * is the responsibility of vnode_pager_setsize 1692 */ 1693 m = bp->b_pages[i]; 1694#if defined(DIAGNOSTIC) 1695 if (m == bogus_page) 1696 panic("allocbuf: bogus page found"); 1697#endif 1698 vm_page_sleep(m, "biodep", &m->busy); 1699 1700 bp->b_pages[i] = NULL; 1701 vm_page_unwire(m, 0); 1702 } 1703 pmap_qremove((vm_offset_t) trunc_page((vm_offset_t)bp->b_data) + 1704 (desiredpages << PAGE_SHIFT), (bp->b_npages - desiredpages)); 1705 bp->b_npages = desiredpages; 1706 } 1707 } else if (newbsize > bp->b_bufsize) { 1708 vm_object_t obj; 1709 vm_offset_t tinc, toff; 1710 vm_ooffset_t off; 1711 vm_pindex_t objoff; 1712 int pageindex, curbpnpages; 1713 struct vnode *vp; 1714 int bsize; 1715 int orig_validoff = bp->b_validoff; 1716 int orig_validend = bp->b_validend; 1717 1718 vp = bp->b_vp; 1719 1720 if (vp->v_type == VBLK) 1721 bsize = DEV_BSIZE; 1722 else 1723 bsize = vp->v_mount->mnt_stat.f_iosize; 1724 1725 if (bp->b_npages < desiredpages) { 1726 obj = vp->v_object; 1727 tinc = PAGE_SIZE; 1728 1729 off = bp->b_offset; 1730#ifdef DIAGNOSTIC 1731 if (bp->b_offset == NOOFFSET) 1732 panic("allocbuf: no buffer offset"); 1733#endif 1734 1735 curbpnpages = bp->b_npages; 1736 doretry: 1737 bp->b_validoff = orig_validoff; 1738 bp->b_validend = orig_validend; 1739 bp->b_flags |= B_CACHE; 1740 for (toff = 0; toff < newbsize; toff += tinc) { 1741 objoff = OFF_TO_IDX(off + toff); 1742 pageindex = objoff - OFF_TO_IDX(off); 1743 tinc = PAGE_SIZE - ((off + toff) & PAGE_MASK); 1744 if (pageindex < curbpnpages) { 1745 1746 m = bp->b_pages[pageindex]; 1747#ifdef VFS_BIO_DIAG 1748 if (m->pindex != objoff) 1749 panic("allocbuf: page changed offset?!!!?"); 1750#endif 1751 if (tinc > (newbsize - toff)) 1752 tinc = newbsize - toff; 1753 if (bp->b_flags & B_CACHE) 1754 vfs_buf_set_valid(bp, off, toff, tinc, m); 1755 continue; 1756 } 1757 m = vm_page_lookup(obj, objoff); 1758 if (!m) { 1759 m = vm_page_alloc(obj, objoff, VM_ALLOC_NORMAL); 1760 if (!m) { 1761 VM_WAIT; 1762 vm_pageout_deficit += (desiredpages - curbpnpages); 1763 goto doretry; 1764 } 1765 1766 vm_page_wire(m); 1767 vm_page_flag_clear(m, PG_BUSY); 1768 bp->b_flags &= ~B_CACHE; 1769 1770 } else if (m->flags & PG_BUSY) { 1771 s = splvm(); 1772 if (m->flags & PG_BUSY) { 1773 vm_page_flag_set(m, PG_WANTED); 1774 tsleep(m, PVM, "pgtblk", 0); 1775 } 1776 splx(s); 1777 goto doretry; 1778 } else { 1779 if ((curproc != pageproc) && 1780 ((m->queue - m->pc) == PQ_CACHE) && 1781 ((cnt.v_free_count + cnt.v_cache_count) < 1782 (cnt.v_free_min + cnt.v_cache_min))) { 1783 pagedaemon_wakeup(); 1784 } 1785 if (tinc > (newbsize - toff)) 1786 tinc = newbsize - toff; 1787 if (bp->b_flags & B_CACHE) 1788 vfs_buf_set_valid(bp, off, toff, tinc, m); 1789 vm_page_flag_clear(m, PG_ZERO); 1790 vm_page_wire(m); 1791 } 1792 bp->b_pages[pageindex] = m; 1793 curbpnpages = pageindex + 1; 1794 } 1795 if (vp->v_tag == VT_NFS && 1796 vp->v_type != VBLK) { 1797 if (bp->b_dirtyend > 0) { 1798 bp->b_validoff = min(bp->b_validoff, bp->b_dirtyoff); 1799 bp->b_validend = max(bp->b_validend, bp->b_dirtyend); 1800 } 1801 if (bp->b_validend == 0) 1802 bp->b_flags &= ~B_CACHE; 1803 } 1804 bp->b_data = (caddr_t) trunc_page((vm_offset_t)bp->b_data); 1805 bp->b_npages = curbpnpages; 1806 pmap_qenter((vm_offset_t) bp->b_data, 1807 bp->b_pages, bp->b_npages); 1808 ((vm_offset_t) bp->b_data) |= off & PAGE_MASK; 1809 } 1810 } 1811 } 1812 if (bp->b_flags & B_VMIO) 1813 vmiospace += (newbsize - bp->b_bufsize); 1814 bufspace += (newbsize - bp->b_bufsize); 1815 bp->b_bufsize = newbsize; 1816 bp->b_bcount = size; 1817 return 1; 1818} 1819 1820/* 1821 * Wait for buffer I/O completion, returning error status. 1822 */ 1823int 1824biowait(register struct buf * bp) 1825{ 1826 int s; 1827 1828 s = splbio(); 1829 while ((bp->b_flags & B_DONE) == 0) 1830#if defined(NO_SCHEDULE_MODS) 1831 tsleep(bp, PRIBIO, "biowait", 0); 1832#else 1833 if (bp->b_flags & B_READ) 1834 tsleep(bp, PRIBIO, "biord", 0); 1835 else 1836 tsleep(bp, PRIBIO, "biowr", 0); 1837#endif 1838 splx(s); 1839 if (bp->b_flags & B_EINTR) { 1840 bp->b_flags &= ~B_EINTR; 1841 return (EINTR); 1842 } 1843 if (bp->b_flags & B_ERROR) { 1844 return (bp->b_error ? bp->b_error : EIO); 1845 } else { 1846 return (0); 1847 } 1848} 1849 1850/* 1851 * Finish I/O on a buffer, calling an optional function. 1852 * This is usually called from interrupt level, so process blocking 1853 * is not *a good idea*. 1854 */ 1855void 1856biodone(register struct buf * bp) 1857{ 1858 int s; 1859 1860 s = splbio(); 1861 1862#if !defined(MAX_PERF) 1863 if (!(bp->b_flags & B_BUSY)) 1864 panic("biodone: buffer not busy"); 1865#endif 1866 1867 if (bp->b_flags & B_DONE) { 1868 splx(s); 1869#if !defined(MAX_PERF) 1870 printf("biodone: buffer already done\n"); 1871#endif 1872 return; 1873 } 1874 bp->b_flags |= B_DONE; 1875 1876 if (bp->b_flags & B_FREEBUF) { 1877 brelse(bp); 1878 splx(s); 1879 return; 1880 } 1881 1882 if ((bp->b_flags & B_READ) == 0) { 1883 vwakeup(bp); 1884 } 1885 1886 /* call optional completion function if requested */ 1887 if (bp->b_flags & B_CALL) { 1888 bp->b_flags &= ~B_CALL; 1889 (*bp->b_iodone) (bp); 1890 splx(s); 1891 return; 1892 } 1893 if (LIST_FIRST(&bp->b_dep) != NULL && bioops.io_complete) 1894 (*bioops.io_complete)(bp); 1895 1896 if (bp->b_flags & B_VMIO) { 1897 int i, resid; 1898 vm_ooffset_t foff; 1899 vm_page_t m; 1900 vm_object_t obj; 1901 int iosize; 1902 struct vnode *vp = bp->b_vp; 1903 1904 obj = vp->v_object; 1905 1906#if defined(VFS_BIO_DEBUG) 1907 if (vp->v_usecount == 0) { 1908 panic("biodone: zero vnode ref count"); 1909 } 1910 1911 if (vp->v_object == NULL) { 1912 panic("biodone: missing VM object"); 1913 } 1914 1915 if ((vp->v_flag & VOBJBUF) == 0) { 1916 panic("biodone: vnode is not setup for merged cache"); 1917 } 1918#endif 1919 1920 foff = bp->b_offset; 1921#ifdef DIAGNOSTIC 1922 if (bp->b_offset == NOOFFSET) 1923 panic("biodone: no buffer offset"); 1924#endif 1925 1926#if !defined(MAX_PERF) 1927 if (!obj) { 1928 panic("biodone: no object"); 1929 } 1930#endif 1931#if defined(VFS_BIO_DEBUG) 1932 if (obj->paging_in_progress < bp->b_npages) { 1933 printf("biodone: paging in progress(%d) < bp->b_npages(%d)\n", 1934 obj->paging_in_progress, bp->b_npages); 1935 } 1936#endif 1937 iosize = bp->b_bufsize; 1938 for (i = 0; i < bp->b_npages; i++) { 1939 int bogusflag = 0; 1940 m = bp->b_pages[i]; 1941 if (m == bogus_page) { 1942 bogusflag = 1; 1943 m = vm_page_lookup(obj, OFF_TO_IDX(foff)); 1944 if (!m) { 1945#if defined(VFS_BIO_DEBUG) 1946 printf("biodone: page disappeared\n"); 1947#endif 1948 vm_object_pip_subtract(obj, 1); 1949 continue; 1950 } 1951 bp->b_pages[i] = m; 1952 pmap_qenter(trunc_page((vm_offset_t)bp->b_data), bp->b_pages, bp->b_npages); 1953 } 1954#if defined(VFS_BIO_DEBUG) 1955 if (OFF_TO_IDX(foff) != m->pindex) { 1956 printf("biodone: foff(%d)/m->pindex(%d) mismatch\n", foff, m->pindex); 1957 } 1958#endif 1959 resid = IDX_TO_OFF(m->pindex + 1) - foff; 1960 if (resid > iosize) 1961 resid = iosize; 1962 1963 /* 1964 * In the write case, the valid and clean bits are 1965 * already changed correctly, so we only need to do this 1966 * here in the read case. 1967 */ 1968 if ((bp->b_flags & B_READ) && !bogusflag && resid > 0) { 1969 vfs_page_set_valid(bp, foff, i, m); 1970 } 1971 vm_page_flag_clear(m, PG_ZERO); 1972 1973 /* 1974 * when debugging new filesystems or buffer I/O methods, this 1975 * is the most common error that pops up. if you see this, you 1976 * have not set the page busy flag correctly!!! 1977 */ 1978 if (m->busy == 0) { 1979#if !defined(MAX_PERF) 1980 printf("biodone: page busy < 0, " 1981 "pindex: %d, foff: 0x(%x,%x), " 1982 "resid: %d, index: %d\n", 1983 (int) m->pindex, (int)(foff >> 32), 1984 (int) foff & 0xffffffff, resid, i); 1985#endif 1986 if (vp->v_type != VBLK) 1987#if !defined(MAX_PERF) 1988 printf(" iosize: %ld, lblkno: %d, flags: 0x%lx, npages: %d\n", 1989 bp->b_vp->v_mount->mnt_stat.f_iosize, 1990 (int) bp->b_lblkno, 1991 bp->b_flags, bp->b_npages); 1992 else 1993 printf(" VDEV, lblkno: %d, flags: 0x%lx, npages: %d\n", 1994 (int) bp->b_lblkno, 1995 bp->b_flags, bp->b_npages); 1996 printf(" valid: 0x%x, dirty: 0x%x, wired: %d\n", 1997 m->valid, m->dirty, m->wire_count); 1998#endif 1999 panic("biodone: page busy < 0\n"); 2000 } 2001 vm_page_io_finish(m); 2002 vm_object_pip_subtract(obj, 1); 2003 foff += resid; 2004 iosize -= resid; 2005 } 2006 if (obj && 2007 (obj->paging_in_progress == 0) && 2008 (obj->flags & OBJ_PIPWNT)) { 2009 vm_object_clear_flag(obj, OBJ_PIPWNT); 2010 wakeup(obj); 2011 } 2012 } 2013 /* 2014 * For asynchronous completions, release the buffer now. The brelse 2015 * checks for B_WANTED and will do the wakeup there if necessary - so 2016 * no need to do a wakeup here in the async case. 2017 */ 2018 2019 if (bp->b_flags & B_ASYNC) { 2020 if ((bp->b_flags & (B_NOCACHE | B_INVAL | B_ERROR | B_RELBUF)) != 0) 2021 brelse(bp); 2022 else 2023 bqrelse(bp); 2024 } else { 2025 bp->b_flags &= ~B_WANTED; 2026 wakeup(bp); 2027 } 2028 splx(s); 2029} 2030 2031#if 0 /* not with kirks code */ 2032static int vfs_update_interval = 30; 2033 2034static void 2035vfs_update() 2036{ 2037 while (1) { 2038 tsleep(&vfs_update_wakeup, PUSER, "update", 2039 hz * vfs_update_interval); 2040 vfs_update_wakeup = 0; 2041 sync(curproc, NULL); 2042 } 2043} 2044 2045static int 2046sysctl_kern_updateinterval SYSCTL_HANDLER_ARGS 2047{ 2048 int error = sysctl_handle_int(oidp, 2049 oidp->oid_arg1, oidp->oid_arg2, req); 2050 if (!error) 2051 wakeup(&vfs_update_wakeup); 2052 return error; 2053} 2054 2055SYSCTL_PROC(_kern, KERN_UPDATEINTERVAL, update, CTLTYPE_INT|CTLFLAG_RW, 2056 &vfs_update_interval, 0, sysctl_kern_updateinterval, "I", ""); 2057 2058#endif 2059 2060 2061/* 2062 * This routine is called in lieu of iodone in the case of 2063 * incomplete I/O. This keeps the busy status for pages 2064 * consistant. 2065 */ 2066void 2067vfs_unbusy_pages(struct buf * bp) 2068{ 2069 int i; 2070 2071 if (bp->b_flags & B_VMIO) { 2072 struct vnode *vp = bp->b_vp; 2073 vm_object_t obj = vp->v_object; 2074 2075 for (i = 0; i < bp->b_npages; i++) { 2076 vm_page_t m = bp->b_pages[i]; 2077 2078 if (m == bogus_page) { 2079 m = vm_page_lookup(obj, OFF_TO_IDX(bp->b_offset) + i); 2080#if !defined(MAX_PERF) 2081 if (!m) { 2082 panic("vfs_unbusy_pages: page missing\n"); 2083 } 2084#endif 2085 bp->b_pages[i] = m; 2086 pmap_qenter(trunc_page((vm_offset_t)bp->b_data), bp->b_pages, bp->b_npages); 2087 } 2088 vm_object_pip_subtract(obj, 1); 2089 vm_page_flag_clear(m, PG_ZERO); 2090 vm_page_io_finish(m); 2091 } 2092 if (obj->paging_in_progress == 0 && 2093 (obj->flags & OBJ_PIPWNT)) { 2094 vm_object_clear_flag(obj, OBJ_PIPWNT); 2095 wakeup(obj); 2096 } 2097 } 2098} 2099 2100/* 2101 * Set NFS' b_validoff and b_validend fields from the valid bits 2102 * of a page. If the consumer is not NFS, and the page is not 2103 * valid for the entire range, clear the B_CACHE flag to force 2104 * the consumer to re-read the page. 2105 */ 2106static void 2107vfs_buf_set_valid(struct buf *bp, 2108 vm_ooffset_t foff, vm_offset_t off, vm_offset_t size, 2109 vm_page_t m) 2110{ 2111 if (bp->b_vp->v_tag == VT_NFS && bp->b_vp->v_type != VBLK) { 2112 vm_offset_t svalid, evalid; 2113 int validbits = m->valid >> (((foff+off)&PAGE_MASK)/DEV_BSIZE); 2114 2115 /* 2116 * This only bothers with the first valid range in the 2117 * page. 2118 */ 2119 svalid = off; 2120 while (validbits && !(validbits & 1)) { 2121 svalid += DEV_BSIZE; 2122 validbits >>= 1; 2123 } 2124 evalid = svalid; 2125 while (validbits & 1) { 2126 evalid += DEV_BSIZE; 2127 validbits >>= 1; 2128 } 2129 evalid = min(evalid, off + size); 2130 /* 2131 * Make sure this range is contiguous with the range 2132 * built up from previous pages. If not, then we will 2133 * just use the range from the previous pages. 2134 */ 2135 if (svalid == bp->b_validend) { 2136 bp->b_validoff = min(bp->b_validoff, svalid); 2137 bp->b_validend = max(bp->b_validend, evalid); 2138 } 2139 } else if (!vm_page_is_valid(m, 2140 (vm_offset_t) ((foff + off) & PAGE_MASK), 2141 size)) { 2142 bp->b_flags &= ~B_CACHE; 2143 } 2144} 2145 2146/* 2147 * Set the valid bits in a page, taking care of the b_validoff, 2148 * b_validend fields which NFS uses to optimise small reads. Off is 2149 * the offset within the file and pageno is the page index within the buf. 2150 */ 2151static void 2152vfs_page_set_valid(struct buf *bp, vm_ooffset_t off, int pageno, vm_page_t m) 2153{ 2154 struct vnode *vp = bp->b_vp; 2155 vm_ooffset_t soff, eoff; 2156 2157 soff = off; 2158 eoff = (off + PAGE_SIZE) & ~PAGE_MASK; 2159 if (eoff > bp->b_offset + bp->b_bufsize) 2160 eoff = bp->b_offset + bp->b_bufsize; 2161 if (vp->v_tag == VT_NFS && vp->v_type != VBLK) { 2162 vm_ooffset_t sv, ev; 2163 vm_page_set_invalid(m, 2164 (vm_offset_t) (soff & PAGE_MASK), 2165 (vm_offset_t) (eoff - soff)); 2166 sv = (bp->b_offset + bp->b_validoff + DEV_BSIZE - 1) & ~(DEV_BSIZE - 1); 2167 ev = (bp->b_offset + bp->b_validend) & ~(DEV_BSIZE - 1); 2168 soff = qmax(sv, soff); 2169 eoff = qmin(ev, eoff); 2170 } 2171 if (eoff > soff) 2172 vm_page_set_validclean(m, 2173 (vm_offset_t) (soff & PAGE_MASK), 2174 (vm_offset_t) (eoff - soff)); 2175} 2176 2177/* 2178 * This routine is called before a device strategy routine. 2179 * It is used to tell the VM system that paging I/O is in 2180 * progress, and treat the pages associated with the buffer 2181 * almost as being PG_BUSY. Also the object paging_in_progress 2182 * flag is handled to make sure that the object doesn't become 2183 * inconsistant. 2184 */ 2185void 2186vfs_busy_pages(struct buf * bp, int clear_modify) 2187{ 2188 int i, bogus; 2189 2190 if (bp->b_flags & B_VMIO) { 2191 struct vnode *vp = bp->b_vp; 2192 vm_object_t obj = vp->v_object; 2193 vm_ooffset_t foff; 2194 2195 foff = bp->b_offset; 2196#ifdef DIAGNOSTIC 2197 if (bp->b_offset == NOOFFSET) 2198 panic("vfs_busy_pages: no buffer offset"); 2199#endif 2200 2201 vfs_setdirty(bp); 2202 2203retry: 2204 for (i = 0; i < bp->b_npages; i++) { 2205 vm_page_t m = bp->b_pages[i]; 2206 if (vm_page_sleep(m, "vbpage", NULL)) 2207 goto retry; 2208 } 2209 2210 bogus = 0; 2211 for (i = 0; i < bp->b_npages; i++) { 2212 vm_page_t m = bp->b_pages[i]; 2213 2214 vm_page_flag_clear(m, PG_ZERO); 2215 if ((bp->b_flags & B_CLUSTER) == 0) { 2216 vm_object_pip_add(obj, 1); 2217 vm_page_io_start(m); 2218 } 2219 2220 vm_page_protect(m, VM_PROT_NONE); 2221 if (clear_modify) 2222 vfs_page_set_valid(bp, foff, i, m); 2223 else if (m->valid == VM_PAGE_BITS_ALL && 2224 (bp->b_flags & B_CACHE) == 0) { 2225 bp->b_pages[i] = bogus_page; 2226 bogus++; 2227 } 2228 foff = (foff + PAGE_SIZE) & ~PAGE_MASK; 2229 } 2230 if (bogus) 2231 pmap_qenter(trunc_page((vm_offset_t)bp->b_data), bp->b_pages, bp->b_npages); 2232 } 2233} 2234 2235/* 2236 * Tell the VM system that the pages associated with this buffer 2237 * are clean. This is used for delayed writes where the data is 2238 * going to go to disk eventually without additional VM intevention. 2239 */ 2240void 2241vfs_clean_pages(struct buf * bp) 2242{ 2243 int i; 2244 2245 if (bp->b_flags & B_VMIO) { 2246 vm_ooffset_t foff; 2247 foff = bp->b_offset; 2248 2249#ifdef DIAGNOSTIC 2250 if (bp->b_offset == NOOFFSET) 2251 panic("vfs_clean_pages: no buffer offset"); 2252#endif 2253 2254 for (i = 0; i < bp->b_npages; i++) { 2255 vm_page_t m = bp->b_pages[i]; 2256 vfs_page_set_valid(bp, foff, i, m); 2257 foff = (foff + PAGE_SIZE) & ~PAGE_MASK; 2258 } 2259 } 2260} 2261 2262void 2263vfs_bio_clrbuf(struct buf *bp) { 2264 int i, size, mask = 0; 2265 caddr_t sa, ea; 2266 if ((bp->b_flags & (B_VMIO | B_MALLOC)) == B_VMIO) { 2267 if( (bp->b_npages == 1) && (bp->b_bufsize < PAGE_SIZE) && 2268 (bp->b_offset & PAGE_MASK) == 0) { 2269 mask = (1 << (bp->b_bufsize / DEV_BSIZE)) - 1; 2270 if (((bp->b_pages[0]->flags & PG_ZERO) == 0) && 2271 ((bp->b_pages[0]->valid & mask) != mask)) { 2272 bzero(bp->b_data, bp->b_bufsize); 2273 } 2274 bp->b_pages[0]->valid |= mask; 2275 bp->b_resid = 0; 2276 return; 2277 } 2278 ea = sa = bp->b_data; 2279 for(i=0;i<bp->b_npages;i++,sa=ea) { 2280 int j = ((u_long)sa & PAGE_MASK) / DEV_BSIZE; 2281 ea = (caddr_t)trunc_page((vm_offset_t)sa + PAGE_SIZE); 2282 ea = (caddr_t)ulmin((u_long)ea, 2283 (u_long)bp->b_data + bp->b_bufsize); 2284 mask = ((1 << ((ea - sa) / DEV_BSIZE)) - 1) << j; 2285 if ((bp->b_pages[i]->valid & mask) == mask) 2286 continue; 2287 if ((bp->b_pages[i]->valid & mask) == 0) { 2288 if ((bp->b_pages[i]->flags & PG_ZERO) == 0) { 2289 bzero(sa, ea - sa); 2290 } 2291 } else { 2292 for (; sa < ea; sa += DEV_BSIZE, j++) { 2293 if (((bp->b_pages[i]->flags & PG_ZERO) == 0) && 2294 (bp->b_pages[i]->valid & (1<<j)) == 0) 2295 bzero(sa, DEV_BSIZE); 2296 } 2297 } 2298 bp->b_pages[i]->valid |= mask; 2299 vm_page_flag_clear(bp->b_pages[i], PG_ZERO); 2300 } 2301 bp->b_resid = 0; 2302 } else { 2303 clrbuf(bp); 2304 } 2305} 2306 2307/* 2308 * vm_hold_load_pages and vm_hold_unload pages get pages into 2309 * a buffers address space. The pages are anonymous and are 2310 * not associated with a file object. 2311 */ 2312void 2313vm_hold_load_pages(struct buf * bp, vm_offset_t from, vm_offset_t to) 2314{ 2315 vm_offset_t pg; 2316 vm_page_t p; 2317 int index; 2318 2319 to = round_page(to); 2320 from = round_page(from); 2321 index = (from - trunc_page((vm_offset_t)bp->b_data)) >> PAGE_SHIFT; 2322 2323 for (pg = from; pg < to; pg += PAGE_SIZE, index++) { 2324 2325tryagain: 2326 2327 p = vm_page_alloc(kernel_object, 2328 ((pg - VM_MIN_KERNEL_ADDRESS) >> PAGE_SHIFT), 2329 VM_ALLOC_NORMAL); 2330 if (!p) { 2331 vm_pageout_deficit += (to - from) >> PAGE_SHIFT; 2332 VM_WAIT; 2333 goto tryagain; 2334 } 2335 vm_page_wire(p); 2336 p->valid = VM_PAGE_BITS_ALL; 2337 vm_page_flag_clear(p, PG_ZERO); 2338 pmap_kenter(pg, VM_PAGE_TO_PHYS(p)); 2339 bp->b_pages[index] = p; 2340 vm_page_wakeup(p); 2341 } 2342 bp->b_npages = index; 2343} 2344 2345void 2346vm_hold_free_pages(struct buf * bp, vm_offset_t from, vm_offset_t to) 2347{ 2348 vm_offset_t pg; 2349 vm_page_t p; 2350 int index, newnpages; 2351 2352 from = round_page(from); 2353 to = round_page(to); 2354 newnpages = index = (from - trunc_page((vm_offset_t)bp->b_data)) >> PAGE_SHIFT; 2355 2356 for (pg = from; pg < to; pg += PAGE_SIZE, index++) { 2357 p = bp->b_pages[index]; 2358 if (p && (index < bp->b_npages)) { 2359#if !defined(MAX_PERF) 2360 if (p->busy) { 2361 printf("vm_hold_free_pages: blkno: %d, lblkno: %d\n", 2362 bp->b_blkno, bp->b_lblkno); 2363 } 2364#endif 2365 bp->b_pages[index] = NULL; 2366 pmap_kremove(pg); 2367 vm_page_busy(p); 2368 vm_page_unwire(p, 0); 2369 vm_page_free(p); 2370 } 2371 } 2372 bp->b_npages = newnpages; 2373} 2374 2375 2376#include "opt_ddb.h" 2377#ifdef DDB 2378#include <ddb/ddb.h> 2379 2380DB_SHOW_COMMAND(buffer, db_show_buffer) 2381{ 2382 /* get args */ 2383 struct buf *bp = (struct buf *)addr; 2384 2385 if (!have_addr) { 2386 db_printf("usage: show buffer <addr>\n"); 2387 return; 2388 } 2389 2390 db_printf("b_proc = %p,\nb_flags = 0x%b\n", (void *)bp->b_proc, 2391 (u_int)bp->b_flags, PRINT_BUF_FLAGS); 2392 db_printf("b_error = %d, b_bufsize = %ld, b_bcount = %ld, " 2393 "b_resid = %ld\nb_dev = 0x%x, b_data = %p, " 2394 "b_blkno = %d, b_pblkno = %d\n", 2395 bp->b_error, bp->b_bufsize, bp->b_bcount, bp->b_resid, 2396 bp->b_dev, bp->b_data, bp->b_blkno, bp->b_pblkno); 2397 if (bp->b_npages) { 2398 int i; 2399 db_printf("b_npages = %d, pages(OBJ, IDX, PA): ", bp->b_npages); 2400 for (i = 0; i < bp->b_npages; i++) { 2401 vm_page_t m; 2402 m = bp->b_pages[i]; 2403 db_printf("(%p, 0x%lx, 0x%lx)", (void *)m->object, 2404 (u_long)m->pindex, (u_long)VM_PAGE_TO_PHYS(m)); 2405 if ((i + 1) < bp->b_npages) 2406 db_printf(","); 2407 } 2408 db_printf("\n"); 2409 } 2410} 2411#endif /* DDB */ 2412