vfs_bio.c revision 41589
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 1.185 1998/11/18 09:00:47 dg 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 vm_page_set_invalid(m, poffset, presid); 676 } 677 resid -= PAGE_SIZE; 678 } 679 680 if (bp->b_flags & (B_INVAL | B_RELBUF)) 681 vfs_vmio_release(bp); 682 683 } else if (bp->b_flags & B_VMIO) { 684 685 if (bp->b_flags & (B_INVAL | B_RELBUF)) 686 vfs_vmio_release(bp); 687 688 } 689 690#if !defined(MAX_PERF) 691 if (bp->b_qindex != QUEUE_NONE) 692 panic("brelse: free buffer onto another queue???"); 693#endif 694 695 /* enqueue */ 696 /* buffers with no memory */ 697 if (bp->b_bufsize == 0) { 698 bp->b_flags |= B_INVAL; 699 bp->b_qindex = QUEUE_EMPTY; 700 TAILQ_INSERT_HEAD(&bufqueues[QUEUE_EMPTY], bp, b_freelist); 701 LIST_REMOVE(bp, b_hash); 702 LIST_INSERT_HEAD(&invalhash, bp, b_hash); 703 bp->b_dev = NODEV; 704 kvafreespace += bp->b_kvasize; 705 706 /* buffers with junk contents */ 707 } else if (bp->b_flags & (B_ERROR | B_INVAL | B_NOCACHE | B_RELBUF)) { 708 bp->b_flags |= B_INVAL; 709 bp->b_qindex = QUEUE_AGE; 710 TAILQ_INSERT_HEAD(&bufqueues[QUEUE_AGE], bp, b_freelist); 711 LIST_REMOVE(bp, b_hash); 712 LIST_INSERT_HEAD(&invalhash, bp, b_hash); 713 bp->b_dev = NODEV; 714 715 /* buffers that are locked */ 716 } else if (bp->b_flags & B_LOCKED) { 717 bp->b_qindex = QUEUE_LOCKED; 718 TAILQ_INSERT_TAIL(&bufqueues[QUEUE_LOCKED], bp, b_freelist); 719 720 /* buffers with stale but valid contents */ 721 } else if (bp->b_flags & B_AGE) { 722 bp->b_qindex = QUEUE_AGE; 723 TAILQ_INSERT_TAIL(&bufqueues[QUEUE_AGE], bp, b_freelist); 724 725 /* buffers with valid and quite potentially reuseable contents */ 726 } else { 727 bp->b_qindex = QUEUE_LRU; 728 TAILQ_INSERT_TAIL(&bufqueues[QUEUE_LRU], bp, b_freelist); 729 } 730 731 if ((bp->b_flags & B_INVAL) || 732 (bp->b_flags & (B_LOCKED|B_DELWRI)) == 0) { 733 if (bp->b_flags & B_DELWRI) { 734 --numdirtybuffers; 735 bp->b_flags &= ~B_DELWRI; 736 } 737 vfs_bio_need_satisfy(); 738 } 739 740 /* unlock */ 741 bp->b_flags &= ~(B_ORDERED | B_WANTED | B_BUSY | 742 B_ASYNC | B_NOCACHE | B_AGE | B_RELBUF); 743 splx(s); 744} 745 746/* 747 * Release a buffer. 748 */ 749void 750bqrelse(struct buf * bp) 751{ 752 int s; 753 754 s = splbio(); 755 756 /* anyone need this block? */ 757 if (bp->b_flags & B_WANTED) { 758 bp->b_flags &= ~(B_WANTED | B_AGE); 759 wakeup(bp); 760 } 761 762#if !defined(MAX_PERF) 763 if (bp->b_qindex != QUEUE_NONE) 764 panic("bqrelse: free buffer onto another queue???"); 765#endif 766 767 if (bp->b_flags & B_LOCKED) { 768 bp->b_flags &= ~B_ERROR; 769 bp->b_qindex = QUEUE_LOCKED; 770 TAILQ_INSERT_TAIL(&bufqueues[QUEUE_LOCKED], bp, b_freelist); 771 /* buffers with stale but valid contents */ 772 } else { 773 bp->b_qindex = QUEUE_LRU; 774 TAILQ_INSERT_TAIL(&bufqueues[QUEUE_LRU], bp, b_freelist); 775 } 776 777 if ((bp->b_flags & (B_LOCKED|B_DELWRI)) == 0) { 778 vfs_bio_need_satisfy(); 779 } 780 781 /* unlock */ 782 bp->b_flags &= ~(B_ORDERED | B_WANTED | B_BUSY | 783 B_ASYNC | B_NOCACHE | B_AGE | B_RELBUF); 784 splx(s); 785} 786 787static void 788vfs_vmio_release(bp) 789 struct buf *bp; 790{ 791 int i, s; 792 vm_page_t m; 793 794 s = splvm(); 795 for (i = 0; i < bp->b_npages; i++) { 796 m = bp->b_pages[i]; 797 bp->b_pages[i] = NULL; 798 /* 799 * In order to keep page LRU ordering consistent, put 800 * everything on the inactive queue. 801 */ 802 vm_page_unwire(m, 0); 803 /* 804 * We don't mess with busy pages, it is 805 * the responsibility of the process that 806 * busied the pages to deal with them. 807 */ 808 if ((m->flags & PG_BUSY) || (m->busy != 0)) 809 continue; 810 811 if (m->wire_count == 0) { 812 vm_page_flag_clear(m, PG_ZERO); 813 /* 814 * Might as well free the page if we can and it has 815 * no valid data. 816 */ 817 if ((bp->b_flags & B_ASYNC) == 0 && !m->valid && m->hold_count == 0) { 818 vm_page_busy(m); 819 vm_page_protect(m, VM_PROT_NONE); 820 vm_page_free(m); 821 } 822 } 823 } 824 splx(s); 825 bufspace -= bp->b_bufsize; 826 vmiospace -= bp->b_bufsize; 827 pmap_qremove(trunc_page((vm_offset_t) bp->b_data), bp->b_npages); 828 bp->b_npages = 0; 829 bp->b_bufsize = 0; 830 bp->b_flags &= ~B_VMIO; 831 if (bp->b_vp) 832 brelvp(bp); 833} 834 835/* 836 * Check to see if a block is currently memory resident. 837 */ 838struct buf * 839gbincore(struct vnode * vp, daddr_t blkno) 840{ 841 struct buf *bp; 842 struct bufhashhdr *bh; 843 844 bh = BUFHASH(vp, blkno); 845 bp = bh->lh_first; 846 847 /* Search hash chain */ 848 while (bp != NULL) { 849 /* hit */ 850 if (bp->b_vp == vp && bp->b_lblkno == blkno && 851 (bp->b_flags & B_INVAL) == 0) { 852 break; 853 } 854 bp = bp->b_hash.le_next; 855 } 856 return (bp); 857} 858 859/* 860 * this routine implements clustered async writes for 861 * clearing out B_DELWRI buffers... This is much better 862 * than the old way of writing only one buffer at a time. 863 */ 864int 865vfs_bio_awrite(struct buf * bp) 866{ 867 int i; 868 daddr_t lblkno = bp->b_lblkno; 869 struct vnode *vp = bp->b_vp; 870 int s; 871 int ncl; 872 struct buf *bpa; 873 int nwritten; 874 int size; 875 int maxcl; 876 877 s = splbio(); 878 /* 879 * right now we support clustered writing only to regular files 880 */ 881 if ((vp->v_type == VREG) && 882 (vp->v_mount != 0) && /* Only on nodes that have the size info */ 883 (bp->b_flags & (B_CLUSTEROK | B_INVAL)) == B_CLUSTEROK) { 884 885 size = vp->v_mount->mnt_stat.f_iosize; 886 maxcl = MAXPHYS / size; 887 888 for (i = 1; i < maxcl; i++) { 889 if ((bpa = gbincore(vp, lblkno + i)) && 890 ((bpa->b_flags & (B_BUSY | B_DELWRI | B_CLUSTEROK | B_INVAL)) == 891 (B_DELWRI | B_CLUSTEROK)) && 892 (bpa->b_bufsize == size)) { 893 if ((bpa->b_blkno == bpa->b_lblkno) || 894 (bpa->b_blkno != bp->b_blkno + ((i * size) >> DEV_BSHIFT))) 895 break; 896 } else { 897 break; 898 } 899 } 900 ncl = i; 901 /* 902 * this is a possible cluster write 903 */ 904 if (ncl != 1) { 905 nwritten = cluster_wbuild(vp, size, lblkno, ncl); 906 splx(s); 907 return nwritten; 908 } 909 } 910 911 bremfree(bp); 912 bp->b_flags |= B_BUSY | B_ASYNC; 913 914 splx(s); 915 /* 916 * default (old) behavior, writing out only one block 917 */ 918 nwritten = bp->b_bufsize; 919 (void) VOP_BWRITE(bp); 920 return nwritten; 921} 922 923 924/* 925 * Find a buffer header which is available for use. 926 */ 927static struct buf * 928getnewbuf(struct vnode *vp, daddr_t blkno, 929 int slpflag, int slptimeo, int size, int maxsize) 930{ 931 struct buf *bp, *bp1; 932 int nbyteswritten = 0; 933 vm_offset_t addr; 934 static int writerecursion = 0; 935 936start: 937 if (bufspace >= maxbufspace) 938 goto trytofreespace; 939 940 /* can we constitute a new buffer? */ 941 if ((bp = TAILQ_FIRST(&bufqueues[QUEUE_EMPTY]))) { 942#if !defined(MAX_PERF) 943 if (bp->b_qindex != QUEUE_EMPTY) 944 panic("getnewbuf: inconsistent EMPTY queue, qindex=%d", 945 bp->b_qindex); 946#endif 947 bp->b_flags |= B_BUSY; 948 bremfree(bp); 949 goto fillbuf; 950 } 951trytofreespace: 952 /* 953 * We keep the file I/O from hogging metadata I/O 954 * This is desirable because file data is cached in the 955 * VM/Buffer cache even if a buffer is freed. 956 */ 957 if ((bp = TAILQ_FIRST(&bufqueues[QUEUE_AGE]))) { 958#if !defined(MAX_PERF) 959 if (bp->b_qindex != QUEUE_AGE) 960 panic("getnewbuf: inconsistent AGE queue, qindex=%d", 961 bp->b_qindex); 962#endif 963 } else if ((bp = TAILQ_FIRST(&bufqueues[QUEUE_LRU]))) { 964#if !defined(MAX_PERF) 965 if (bp->b_qindex != QUEUE_LRU) 966 panic("getnewbuf: inconsistent LRU queue, qindex=%d", 967 bp->b_qindex); 968#endif 969 } 970 if (!bp) { 971 /* wait for a free buffer of any kind */ 972 needsbuffer |= VFS_BIO_NEED_ANY; 973 do 974 tsleep(&needsbuffer, (PRIBIO + 4) | slpflag, "newbuf", 975 slptimeo); 976 while (needsbuffer & VFS_BIO_NEED_ANY); 977 return (0); 978 } 979 980#if defined(DIAGNOSTIC) 981 if (bp->b_flags & B_BUSY) { 982 panic("getnewbuf: busy buffer on free list\n"); 983 } 984#endif 985 986 /* 987 * We are fairly aggressive about freeing VMIO buffers, but since 988 * the buffering is intact without buffer headers, there is not 989 * much loss. We gain by maintaining non-VMIOed metadata in buffers. 990 */ 991 if ((bp->b_qindex == QUEUE_LRU) && (bp->b_usecount > 0)) { 992 if ((bp->b_flags & B_VMIO) == 0 || 993 (vmiospace < maxvmiobufspace)) { 994 --bp->b_usecount; 995 TAILQ_REMOVE(&bufqueues[QUEUE_LRU], bp, b_freelist); 996 if (TAILQ_FIRST(&bufqueues[QUEUE_LRU]) != NULL) { 997 TAILQ_INSERT_TAIL(&bufqueues[QUEUE_LRU], bp, b_freelist); 998 goto start; 999 } 1000 TAILQ_INSERT_TAIL(&bufqueues[QUEUE_LRU], bp, b_freelist); 1001 } 1002 } 1003 1004 1005 /* if we are a delayed write, convert to an async write */ 1006 if ((bp->b_flags & (B_DELWRI | B_INVAL)) == B_DELWRI) { 1007 1008 /* 1009 * If our delayed write is likely to be used soon, then 1010 * recycle back onto the LRU queue. 1011 */ 1012 if (vp && (bp->b_vp == vp) && (bp->b_qindex == QUEUE_LRU) && 1013 (bp->b_lblkno >= blkno) && (maxsize > 0)) { 1014 1015 if (bp->b_usecount > 0) { 1016 if (bp->b_lblkno < blkno + (MAXPHYS / maxsize)) { 1017 1018 TAILQ_REMOVE(&bufqueues[QUEUE_LRU], bp, b_freelist); 1019 1020 if (TAILQ_FIRST(&bufqueues[QUEUE_LRU]) != NULL) { 1021 TAILQ_INSERT_TAIL(&bufqueues[QUEUE_LRU], bp, b_freelist); 1022 bp->b_usecount--; 1023 goto start; 1024 } 1025 TAILQ_INSERT_TAIL(&bufqueues[QUEUE_LRU], bp, b_freelist); 1026 } 1027 } 1028 } 1029 1030 /* 1031 * Certain layered filesystems can recursively re-enter the vfs_bio 1032 * code, due to delayed writes. This helps keep the system from 1033 * deadlocking. 1034 */ 1035 if (writerecursion > 0) { 1036 if (writerecursion > 5) { 1037 bp = TAILQ_FIRST(&bufqueues[QUEUE_AGE]); 1038 while (bp) { 1039 if ((bp->b_flags & B_DELWRI) == 0) 1040 break; 1041 bp = TAILQ_NEXT(bp, b_freelist); 1042 } 1043 if (bp == NULL) { 1044 bp = TAILQ_FIRST(&bufqueues[QUEUE_LRU]); 1045 while (bp) { 1046 if ((bp->b_flags & B_DELWRI) == 0) 1047 break; 1048 bp = TAILQ_NEXT(bp, b_freelist); 1049 } 1050 } 1051 if (bp == NULL) 1052 panic("getnewbuf: cannot get buffer, infinite recursion failure"); 1053 } else { 1054 bremfree(bp); 1055 bp->b_flags |= B_BUSY | B_AGE | B_ASYNC; 1056 nbyteswritten += bp->b_bufsize; 1057 ++writerecursion; 1058 VOP_BWRITE(bp); 1059 --writerecursion; 1060 if (!slpflag && !slptimeo) { 1061 return (0); 1062 } 1063 goto start; 1064 } 1065 } else { 1066 ++writerecursion; 1067 nbyteswritten += vfs_bio_awrite(bp); 1068 --writerecursion; 1069 if (!slpflag && !slptimeo) { 1070 return (0); 1071 } 1072 goto start; 1073 } 1074 } 1075 1076 if (bp->b_flags & B_WANTED) { 1077 bp->b_flags &= ~B_WANTED; 1078 wakeup(bp); 1079 } 1080 bremfree(bp); 1081 bp->b_flags |= B_BUSY; 1082 1083 if (bp->b_flags & B_VMIO) { 1084 bp->b_flags &= ~B_ASYNC; 1085 vfs_vmio_release(bp); 1086 } 1087 1088 if (bp->b_vp) 1089 brelvp(bp); 1090 1091fillbuf: 1092 1093 /* we are not free, nor do we contain interesting data */ 1094 if (bp->b_rcred != NOCRED) { 1095 crfree(bp->b_rcred); 1096 bp->b_rcred = NOCRED; 1097 } 1098 if (bp->b_wcred != NOCRED) { 1099 crfree(bp->b_wcred); 1100 bp->b_wcred = NOCRED; 1101 } 1102 if (LIST_FIRST(&bp->b_dep) != NULL && 1103 bioops.io_deallocate) 1104 (*bioops.io_deallocate)(bp); 1105 1106 LIST_REMOVE(bp, b_hash); 1107 LIST_INSERT_HEAD(&invalhash, bp, b_hash); 1108 if (bp->b_bufsize) { 1109 allocbuf(bp, 0); 1110 } 1111 bp->b_flags = B_BUSY; 1112 bp->b_dev = NODEV; 1113 bp->b_vp = NULL; 1114 bp->b_blkno = bp->b_lblkno = 0; 1115 bp->b_offset = NOOFFSET; 1116 bp->b_iodone = 0; 1117 bp->b_error = 0; 1118 bp->b_resid = 0; 1119 bp->b_bcount = 0; 1120 bp->b_npages = 0; 1121 bp->b_dirtyoff = bp->b_dirtyend = 0; 1122 bp->b_validoff = bp->b_validend = 0; 1123 bp->b_usecount = 5; 1124 /* Here, not kern_physio.c, is where this should be done*/ 1125 LIST_INIT(&bp->b_dep); 1126 1127 maxsize = (maxsize + PAGE_MASK) & ~PAGE_MASK; 1128 1129 /* 1130 * we assume that buffer_map is not at address 0 1131 */ 1132 addr = 0; 1133 if (maxsize != bp->b_kvasize) { 1134 bfreekva(bp); 1135 1136findkvaspace: 1137 /* 1138 * See if we have buffer kva space 1139 */ 1140 if (vm_map_findspace(buffer_map, 1141 vm_map_min(buffer_map), maxsize, &addr)) { 1142 if (kvafreespace > 0) { 1143 int totfree = 0, freed; 1144 do { 1145 freed = 0; 1146 for (bp1 = TAILQ_FIRST(&bufqueues[QUEUE_EMPTY]); 1147 bp1 != NULL; bp1 = TAILQ_NEXT(bp1, b_freelist)) { 1148 if (bp1->b_kvasize != 0) { 1149 totfree += bp1->b_kvasize; 1150 freed = bp1->b_kvasize; 1151 bremfree(bp1); 1152 bfreekva(bp1); 1153 brelse(bp1); 1154 break; 1155 } 1156 } 1157 } while (freed); 1158 /* 1159 * if we found free space, then retry with the same buffer. 1160 */ 1161 if (totfree) 1162 goto findkvaspace; 1163 } 1164 bp->b_flags |= B_INVAL; 1165 brelse(bp); 1166 goto trytofreespace; 1167 } 1168 } 1169 1170 /* 1171 * See if we are below are allocated minimum 1172 */ 1173 if (bufspace >= (maxbufspace + nbyteswritten)) { 1174 bp->b_flags |= B_INVAL; 1175 brelse(bp); 1176 goto trytofreespace; 1177 } 1178 1179 /* 1180 * create a map entry for the buffer -- in essence 1181 * reserving the kva space. 1182 */ 1183 if (addr) { 1184 vm_map_insert(buffer_map, NULL, 0, 1185 addr, addr + maxsize, 1186 VM_PROT_ALL, VM_PROT_ALL, MAP_NOFAULT); 1187 1188 bp->b_kvabase = (caddr_t) addr; 1189 bp->b_kvasize = maxsize; 1190 } 1191 bp->b_data = bp->b_kvabase; 1192 1193 return (bp); 1194} 1195 1196static void 1197waitfreebuffers(int slpflag, int slptimeo) { 1198 while (numfreebuffers < hifreebuffers) { 1199 flushdirtybuffers(slpflag, slptimeo); 1200 if (numfreebuffers < hifreebuffers) 1201 break; 1202 needsbuffer |= VFS_BIO_NEED_FREE; 1203 if (tsleep(&needsbuffer, (PRIBIO + 4)|slpflag, "biofre", slptimeo)) 1204 break; 1205 } 1206} 1207 1208static void 1209flushdirtybuffers(int slpflag, int slptimeo) { 1210 int s; 1211 static pid_t flushing = 0; 1212 1213 s = splbio(); 1214 1215 if (flushing) { 1216 if (flushing == curproc->p_pid) { 1217 splx(s); 1218 return; 1219 } 1220 while (flushing) { 1221 if (tsleep(&flushing, (PRIBIO + 4)|slpflag, "biofls", slptimeo)) { 1222 splx(s); 1223 return; 1224 } 1225 } 1226 } 1227 flushing = curproc->p_pid; 1228 1229 while (numdirtybuffers > lodirtybuffers) { 1230 struct buf *bp; 1231 needsbuffer |= VFS_BIO_NEED_LOWLIMIT; 1232 bp = TAILQ_FIRST(&bufqueues[QUEUE_AGE]); 1233 if (bp == NULL) 1234 bp = TAILQ_FIRST(&bufqueues[QUEUE_LRU]); 1235 1236 while (bp && ((bp->b_flags & B_DELWRI) == 0)) { 1237 bp = TAILQ_NEXT(bp, b_freelist); 1238 } 1239 1240 if (bp) { 1241 vfs_bio_awrite(bp); 1242 continue; 1243 } 1244 break; 1245 } 1246 1247 flushing = 0; 1248 wakeup(&flushing); 1249 splx(s); 1250} 1251 1252/* 1253 * Check to see if a block is currently memory resident. 1254 */ 1255struct buf * 1256incore(struct vnode * vp, daddr_t blkno) 1257{ 1258 struct buf *bp; 1259 1260 int s = splbio(); 1261 bp = gbincore(vp, blkno); 1262 splx(s); 1263 return (bp); 1264} 1265 1266/* 1267 * Returns true if no I/O is needed to access the 1268 * associated VM object. This is like incore except 1269 * it also hunts around in the VM system for the data. 1270 */ 1271 1272int 1273inmem(struct vnode * vp, daddr_t blkno) 1274{ 1275 vm_object_t obj; 1276 vm_offset_t toff, tinc; 1277 vm_page_t m; 1278 vm_ooffset_t off; 1279 1280 if (incore(vp, blkno)) 1281 return 1; 1282 if (vp->v_mount == NULL) 1283 return 0; 1284 if ((vp->v_object == NULL) || (vp->v_flag & VOBJBUF) == 0) 1285 return 0; 1286 1287 obj = vp->v_object; 1288 tinc = PAGE_SIZE; 1289 if (tinc > vp->v_mount->mnt_stat.f_iosize) 1290 tinc = vp->v_mount->mnt_stat.f_iosize; 1291 off = blkno * vp->v_mount->mnt_stat.f_iosize; 1292 1293 for (toff = 0; toff < vp->v_mount->mnt_stat.f_iosize; toff += tinc) { 1294 1295 m = vm_page_lookup(obj, OFF_TO_IDX(off + toff)); 1296 if (!m) 1297 return 0; 1298 if (vm_page_is_valid(m, 1299 (vm_offset_t) ((toff + off) & PAGE_MASK), tinc) == 0) 1300 return 0; 1301 } 1302 return 1; 1303} 1304 1305/* 1306 * now we set the dirty range for the buffer -- 1307 * for NFS -- if the file is mapped and pages have 1308 * been written to, let it know. We want the 1309 * entire range of the buffer to be marked dirty if 1310 * any of the pages have been written to for consistancy 1311 * with the b_validoff, b_validend set in the nfs write 1312 * code, and used by the nfs read code. 1313 */ 1314static void 1315vfs_setdirty(struct buf *bp) { 1316 int i; 1317 vm_object_t object; 1318 vm_offset_t boffset, offset; 1319 /* 1320 * We qualify the scan for modified pages on whether the 1321 * object has been flushed yet. The OBJ_WRITEABLE flag 1322 * is not cleared simply by protecting pages off. 1323 */ 1324 if ((bp->b_flags & B_VMIO) && 1325 ((object = bp->b_pages[0]->object)->flags & (OBJ_WRITEABLE|OBJ_CLEANING))) { 1326 /* 1327 * test the pages to see if they have been modified directly 1328 * by users through the VM system. 1329 */ 1330 for (i = 0; i < bp->b_npages; i++) { 1331 vm_page_flag_clear(bp->b_pages[i], PG_ZERO); 1332 vm_page_test_dirty(bp->b_pages[i]); 1333 } 1334 1335 /* 1336 * scan forwards for the first page modified 1337 */ 1338 for (i = 0; i < bp->b_npages; i++) { 1339 if (bp->b_pages[i]->dirty) { 1340 break; 1341 } 1342 } 1343 boffset = (i << PAGE_SHIFT); 1344 if (boffset < bp->b_dirtyoff) { 1345 bp->b_dirtyoff = boffset; 1346 } 1347 1348 /* 1349 * scan backwards for the last page modified 1350 */ 1351 for (i = bp->b_npages - 1; i >= 0; --i) { 1352 if (bp->b_pages[i]->dirty) { 1353 break; 1354 } 1355 } 1356 boffset = (i + 1); 1357 offset = boffset + bp->b_pages[0]->pindex; 1358 if (offset >= object->size) 1359 boffset = object->size - bp->b_pages[0]->pindex; 1360 if (bp->b_dirtyend < (boffset << PAGE_SHIFT)) 1361 bp->b_dirtyend = (boffset << PAGE_SHIFT); 1362 } 1363} 1364 1365/* 1366 * Get a block given a specified block and offset into a file/device. 1367 */ 1368struct buf * 1369getblk(struct vnode * vp, daddr_t blkno, int size, int slpflag, int slptimeo) 1370{ 1371 struct buf *bp; 1372 int i, s; 1373 struct bufhashhdr *bh; 1374 int maxsize; 1375 int checksize; 1376 1377 if (vp->v_mount) { 1378 maxsize = vp->v_mount->mnt_stat.f_iosize; 1379 /* 1380 * This happens on mount points. 1381 */ 1382 if (maxsize < size) 1383 maxsize = size; 1384 } else { 1385 maxsize = size; 1386 } 1387 1388#if !defined(MAX_PERF) 1389 if (size > MAXBSIZE) 1390 panic("getblk: size(%d) > MAXBSIZE(%d)\n", size, MAXBSIZE); 1391#endif 1392 1393 s = splbio(); 1394loop: 1395 if (numfreebuffers < lofreebuffers) { 1396 waitfreebuffers(slpflag, slptimeo); 1397 } 1398 1399 if ((bp = gbincore(vp, blkno))) { 1400 if (bp->b_flags & B_BUSY) { 1401 1402 bp->b_flags |= B_WANTED; 1403 if (bp->b_usecount < BUF_MAXUSE) 1404 ++bp->b_usecount; 1405 1406 if (!tsleep(bp, 1407 (PRIBIO + 4) | slpflag, "getblk", slptimeo)) { 1408 goto loop; 1409 } 1410 1411 splx(s); 1412 return (struct buf *) NULL; 1413 } 1414 bp->b_flags |= B_BUSY | B_CACHE; 1415 bremfree(bp); 1416 1417 /* 1418 * check for size inconsistancies (note that they shouldn't 1419 * happen but do when filesystems don't handle the size changes 1420 * correctly.) We are conservative on metadata and don't just 1421 * extend the buffer but write (if needed) and re-constitute it. 1422 */ 1423 1424 if (bp->b_bcount != size) { 1425 if ((bp->b_flags & B_VMIO) && (size <= bp->b_kvasize)) { 1426 allocbuf(bp, size); 1427 } else { 1428 if (bp->b_flags & B_DELWRI) { 1429 bp->b_flags |= B_NOCACHE; 1430 VOP_BWRITE(bp); 1431 } else { 1432 if ((bp->b_flags & B_VMIO) && 1433 (LIST_FIRST(&bp->b_dep) == NULL)) { 1434 bp->b_flags |= B_RELBUF; 1435 brelse(bp); 1436 } else { 1437 bp->b_flags |= B_NOCACHE; 1438 VOP_BWRITE(bp); 1439 } 1440 } 1441 goto loop; 1442 } 1443 } 1444 1445#ifdef DIAGNOSTIC 1446 if (bp->b_offset == NOOFFSET) 1447 panic("getblk: no buffer offset"); 1448#endif 1449 1450 /* 1451 * Check that the constituted buffer really deserves for the 1452 * B_CACHE bit to be set. B_VMIO type buffers might not 1453 * contain fully valid pages. Normal (old-style) buffers 1454 * should be fully valid. 1455 */ 1456 if (bp->b_flags & B_VMIO) { 1457 checksize = bp->b_bufsize; 1458 for (i = 0; i < bp->b_npages; i++) { 1459 int resid; 1460 int poffset; 1461 poffset = bp->b_offset & PAGE_MASK; 1462 resid = (checksize > (PAGE_SIZE - poffset)) ? 1463 (PAGE_SIZE - poffset) : checksize; 1464 if (!vm_page_is_valid(bp->b_pages[i], poffset, resid)) { 1465 bp->b_flags &= ~(B_CACHE | B_DONE); 1466 break; 1467 } 1468 checksize -= resid; 1469 } 1470 } 1471 1472 if (bp->b_usecount < BUF_MAXUSE) 1473 ++bp->b_usecount; 1474 splx(s); 1475 return (bp); 1476 } else { 1477 vm_object_t obj; 1478 1479 if ((bp = getnewbuf(vp, blkno, 1480 slpflag, slptimeo, size, maxsize)) == 0) { 1481 if (slpflag || slptimeo) { 1482 splx(s); 1483 return NULL; 1484 } 1485 goto loop; 1486 } 1487 1488 /* 1489 * This code is used to make sure that a buffer is not 1490 * created while the getnewbuf routine is blocked. 1491 * Normally the vnode is locked so this isn't a problem. 1492 * VBLK type I/O requests, however, don't lock the vnode. 1493 */ 1494 if (VOP_ISLOCKED(vp) != LK_EXCLUSIVE && gbincore(vp, blkno)) { 1495 bp->b_flags |= B_INVAL; 1496 brelse(bp); 1497 goto loop; 1498 } 1499 1500 /* 1501 * Insert the buffer into the hash, so that it can 1502 * be found by incore. 1503 */ 1504 bp->b_blkno = bp->b_lblkno = blkno; 1505 1506 if (vp->v_type != VBLK) 1507 bp->b_offset = (off_t) blkno * maxsize; 1508 else 1509 bp->b_offset = (off_t) blkno * DEV_BSIZE; 1510 1511 bgetvp(vp, bp); 1512 LIST_REMOVE(bp, b_hash); 1513 bh = BUFHASH(vp, blkno); 1514 LIST_INSERT_HEAD(bh, bp, b_hash); 1515 1516 if ((obj = vp->v_object) && (vp->v_flag & VOBJBUF)) { 1517 bp->b_flags |= (B_VMIO | B_CACHE); 1518#if defined(VFS_BIO_DEBUG) 1519 if (vp->v_type != VREG && vp->v_type != VBLK) 1520 printf("getblk: vmioing file type %d???\n", vp->v_type); 1521#endif 1522 } else { 1523 bp->b_flags &= ~B_VMIO; 1524 } 1525 1526 allocbuf(bp, size); 1527 1528 splx(s); 1529 return (bp); 1530 } 1531} 1532 1533/* 1534 * Get an empty, disassociated buffer of given size. 1535 */ 1536struct buf * 1537geteblk(int size) 1538{ 1539 struct buf *bp; 1540 int s; 1541 1542 s = splbio(); 1543 while ((bp = getnewbuf(0, (daddr_t) 0, 0, 0, size, MAXBSIZE)) == 0); 1544 splx(s); 1545 allocbuf(bp, size); 1546 bp->b_flags |= B_INVAL; /* b_dep cleared by getnewbuf() */ 1547 return (bp); 1548} 1549 1550 1551/* 1552 * This code constitutes the buffer memory from either anonymous system 1553 * memory (in the case of non-VMIO operations) or from an associated 1554 * VM object (in the case of VMIO operations). 1555 * 1556 * Note that this code is tricky, and has many complications to resolve 1557 * deadlock or inconsistant data situations. Tread lightly!!! 1558 * 1559 * Modify the length of a buffer's underlying buffer storage without 1560 * destroying information (unless, of course the buffer is shrinking). 1561 */ 1562int 1563allocbuf(struct buf * bp, int size) 1564{ 1565 1566 int s; 1567 int newbsize, mbsize; 1568 int i; 1569 1570#if !defined(MAX_PERF) 1571 if (!(bp->b_flags & B_BUSY)) 1572 panic("allocbuf: buffer not busy"); 1573 1574 if (bp->b_kvasize < size) 1575 panic("allocbuf: buffer too small"); 1576#endif 1577 1578 if ((bp->b_flags & B_VMIO) == 0) { 1579 caddr_t origbuf; 1580 int origbufsize; 1581 /* 1582 * Just get anonymous memory from the kernel 1583 */ 1584 mbsize = (size + DEV_BSIZE - 1) & ~(DEV_BSIZE - 1); 1585#if !defined(NO_B_MALLOC) 1586 if (bp->b_flags & B_MALLOC) 1587 newbsize = mbsize; 1588 else 1589#endif 1590 newbsize = round_page(size); 1591 1592 if (newbsize < bp->b_bufsize) { 1593#if !defined(NO_B_MALLOC) 1594 /* 1595 * malloced buffers are not shrunk 1596 */ 1597 if (bp->b_flags & B_MALLOC) { 1598 if (newbsize) { 1599 bp->b_bcount = size; 1600 } else { 1601 free(bp->b_data, M_BIOBUF); 1602 bufspace -= bp->b_bufsize; 1603 bufmallocspace -= bp->b_bufsize; 1604 bp->b_data = bp->b_kvabase; 1605 bp->b_bufsize = 0; 1606 bp->b_bcount = 0; 1607 bp->b_flags &= ~B_MALLOC; 1608 } 1609 return 1; 1610 } 1611#endif 1612 vm_hold_free_pages( 1613 bp, 1614 (vm_offset_t) bp->b_data + newbsize, 1615 (vm_offset_t) bp->b_data + bp->b_bufsize); 1616 } else if (newbsize > bp->b_bufsize) { 1617#if !defined(NO_B_MALLOC) 1618 /* 1619 * We only use malloced memory on the first allocation. 1620 * and revert to page-allocated memory when the buffer grows. 1621 */ 1622 if ( (bufmallocspace < maxbufmallocspace) && 1623 (bp->b_bufsize == 0) && 1624 (mbsize <= PAGE_SIZE/2)) { 1625 1626 bp->b_data = malloc(mbsize, M_BIOBUF, M_WAITOK); 1627 bp->b_bufsize = mbsize; 1628 bp->b_bcount = size; 1629 bp->b_flags |= B_MALLOC; 1630 bufspace += mbsize; 1631 bufmallocspace += mbsize; 1632 return 1; 1633 } 1634#endif 1635 origbuf = NULL; 1636 origbufsize = 0; 1637#if !defined(NO_B_MALLOC) 1638 /* 1639 * If the buffer is growing on its other-than-first allocation, 1640 * then we revert to the page-allocation scheme. 1641 */ 1642 if (bp->b_flags & B_MALLOC) { 1643 origbuf = bp->b_data; 1644 origbufsize = bp->b_bufsize; 1645 bp->b_data = bp->b_kvabase; 1646 bufspace -= bp->b_bufsize; 1647 bufmallocspace -= bp->b_bufsize; 1648 bp->b_bufsize = 0; 1649 bp->b_flags &= ~B_MALLOC; 1650 newbsize = round_page(newbsize); 1651 } 1652#endif 1653 vm_hold_load_pages( 1654 bp, 1655 (vm_offset_t) bp->b_data + bp->b_bufsize, 1656 (vm_offset_t) bp->b_data + newbsize); 1657#if !defined(NO_B_MALLOC) 1658 if (origbuf) { 1659 bcopy(origbuf, bp->b_data, origbufsize); 1660 free(origbuf, M_BIOBUF); 1661 } 1662#endif 1663 } 1664 } else { 1665 vm_page_t m; 1666 int desiredpages; 1667 1668 newbsize = (size + DEV_BSIZE - 1) & ~(DEV_BSIZE - 1); 1669 desiredpages = (round_page(newbsize) >> PAGE_SHIFT); 1670 1671#if !defined(NO_B_MALLOC) 1672 if (bp->b_flags & B_MALLOC) 1673 panic("allocbuf: VMIO buffer can't be malloced"); 1674#endif 1675 1676 if (newbsize < bp->b_bufsize) { 1677 if (desiredpages < bp->b_npages) { 1678 for (i = desiredpages; i < bp->b_npages; i++) { 1679 /* 1680 * the page is not freed here -- it 1681 * is the responsibility of vnode_pager_setsize 1682 */ 1683 m = bp->b_pages[i]; 1684#if defined(DIAGNOSTIC) 1685 if (m == bogus_page) 1686 panic("allocbuf: bogus page found"); 1687#endif 1688 vm_page_sleep(m, "biodep", &m->busy); 1689 1690 bp->b_pages[i] = NULL; 1691 vm_page_unwire(m, 0); 1692 } 1693 pmap_qremove((vm_offset_t) trunc_page((vm_offset_t)bp->b_data) + 1694 (desiredpages << PAGE_SHIFT), (bp->b_npages - desiredpages)); 1695 bp->b_npages = desiredpages; 1696 } 1697 } else if (newbsize > bp->b_bufsize) { 1698 vm_object_t obj; 1699 vm_offset_t tinc, toff; 1700 vm_ooffset_t off; 1701 vm_pindex_t objoff; 1702 int pageindex, curbpnpages; 1703 struct vnode *vp; 1704 int bsize; 1705 int orig_validoff = bp->b_validoff; 1706 int orig_validend = bp->b_validend; 1707 1708 vp = bp->b_vp; 1709 1710 if (vp->v_type == VBLK) 1711 bsize = DEV_BSIZE; 1712 else 1713 bsize = vp->v_mount->mnt_stat.f_iosize; 1714 1715 if (bp->b_npages < desiredpages) { 1716 obj = vp->v_object; 1717 tinc = PAGE_SIZE; 1718 if (tinc > bsize) 1719 tinc = bsize; 1720 1721 off = bp->b_offset; 1722#ifdef DIAGNOSTIC 1723 if (bp->b_offset == NOOFFSET) 1724 panic("allocbuf: no buffer offset"); 1725#endif 1726 1727 curbpnpages = bp->b_npages; 1728 doretry: 1729 bp->b_validoff = orig_validoff; 1730 bp->b_validend = orig_validend; 1731 bp->b_flags |= B_CACHE; 1732 for (toff = 0; toff < newbsize; toff += tinc) { 1733 int bytesinpage; 1734 1735 pageindex = toff >> PAGE_SHIFT; 1736 objoff = OFF_TO_IDX(off + toff); 1737 if (pageindex < curbpnpages) { 1738 1739 m = bp->b_pages[pageindex]; 1740#ifdef VFS_BIO_DIAG 1741 if (m->pindex != objoff) 1742 panic("allocbuf: page changed offset?!!!?"); 1743#endif 1744 bytesinpage = tinc; 1745 if (tinc > (newbsize - toff)) 1746 bytesinpage = newbsize - toff; 1747 if (bp->b_flags & B_CACHE) 1748 vfs_buf_set_valid(bp, off, toff, bytesinpage, m); 1749 continue; 1750 } 1751 m = vm_page_lookup(obj, objoff); 1752 if (!m) { 1753 m = vm_page_alloc(obj, objoff, VM_ALLOC_NORMAL); 1754 if (!m) { 1755 VM_WAIT; 1756 vm_pageout_deficit += (desiredpages - bp->b_npages); 1757 goto doretry; 1758 } 1759 1760 vm_page_wire(m); 1761 vm_page_flag_clear(m, PG_BUSY); 1762 bp->b_flags &= ~B_CACHE; 1763 1764 } else if (m->flags & PG_BUSY) { 1765 s = splvm(); 1766 if (m->flags & PG_BUSY) { 1767 vm_page_flag_set(m, PG_WANTED); 1768 tsleep(m, PVM, "pgtblk", 0); 1769 } 1770 splx(s); 1771 goto doretry; 1772 } else { 1773 if ((curproc != pageproc) && 1774 ((m->queue - m->pc) == PQ_CACHE) && 1775 ((cnt.v_free_count + cnt.v_cache_count) < 1776 (cnt.v_free_min + cnt.v_cache_min))) { 1777 pagedaemon_wakeup(); 1778 } 1779 bytesinpage = tinc; 1780 if (tinc > (newbsize - toff)) 1781 bytesinpage = newbsize - toff; 1782 if (bp->b_flags & B_CACHE) 1783 vfs_buf_set_valid(bp, off, toff, bytesinpage, m); 1784 vm_page_flag_clear(m, PG_ZERO); 1785 vm_page_wire(m); 1786 } 1787 bp->b_pages[pageindex] = m; 1788 curbpnpages = pageindex + 1; 1789 } 1790 if (vp->v_tag == VT_NFS && 1791 vp->v_type != VBLK) { 1792 if (bp->b_dirtyend > 0) { 1793 bp->b_validoff = min(bp->b_validoff, bp->b_dirtyoff); 1794 bp->b_validend = max(bp->b_validend, bp->b_dirtyend); 1795 } 1796 if (bp->b_validend == 0) 1797 bp->b_flags &= ~B_CACHE; 1798 } 1799 bp->b_data = (caddr_t) trunc_page((vm_offset_t)bp->b_data); 1800 bp->b_npages = curbpnpages; 1801 pmap_qenter((vm_offset_t) bp->b_data, 1802 bp->b_pages, bp->b_npages); 1803 ((vm_offset_t) bp->b_data) |= off & PAGE_MASK; 1804 } 1805 } 1806 } 1807 if (bp->b_flags & B_VMIO) 1808 vmiospace += (newbsize - bp->b_bufsize); 1809 bufspace += (newbsize - bp->b_bufsize); 1810 bp->b_bufsize = newbsize; 1811 bp->b_bcount = size; 1812 return 1; 1813} 1814 1815/* 1816 * Wait for buffer I/O completion, returning error status. 1817 */ 1818int 1819biowait(register struct buf * bp) 1820{ 1821 int s; 1822 1823 s = splbio(); 1824 while ((bp->b_flags & B_DONE) == 0) 1825#if defined(NO_SCHEDULE_MODS) 1826 tsleep(bp, PRIBIO, "biowait", 0); 1827#else 1828 if (bp->b_flags & B_READ) 1829 tsleep(bp, PRIBIO, "biord", 0); 1830 else 1831 tsleep(bp, PRIBIO, "biowr", 0); 1832#endif 1833 splx(s); 1834 if (bp->b_flags & B_EINTR) { 1835 bp->b_flags &= ~B_EINTR; 1836 return (EINTR); 1837 } 1838 if (bp->b_flags & B_ERROR) { 1839 return (bp->b_error ? bp->b_error : EIO); 1840 } else { 1841 return (0); 1842 } 1843} 1844 1845/* 1846 * Finish I/O on a buffer, calling an optional function. 1847 * This is usually called from interrupt level, so process blocking 1848 * is not *a good idea*. 1849 */ 1850void 1851biodone(register struct buf * bp) 1852{ 1853 int s; 1854 1855 s = splbio(); 1856 1857#if !defined(MAX_PERF) 1858 if (!(bp->b_flags & B_BUSY)) 1859 panic("biodone: buffer not busy"); 1860#endif 1861 1862 if (bp->b_flags & B_DONE) { 1863 splx(s); 1864#if !defined(MAX_PERF) 1865 printf("biodone: buffer already done\n"); 1866#endif 1867 return; 1868 } 1869 bp->b_flags |= B_DONE; 1870 1871 if (bp->b_flags & B_FREEBUF) { 1872 brelse(bp); 1873 splx(s); 1874 return; 1875 } 1876 1877 if ((bp->b_flags & B_READ) == 0) { 1878 vwakeup(bp); 1879 } 1880 1881 /* call optional completion function if requested */ 1882 if (bp->b_flags & B_CALL) { 1883 bp->b_flags &= ~B_CALL; 1884 (*bp->b_iodone) (bp); 1885 splx(s); 1886 return; 1887 } 1888 if (LIST_FIRST(&bp->b_dep) != NULL && bioops.io_complete) 1889 (*bioops.io_complete)(bp); 1890 1891 if (bp->b_flags & B_VMIO) { 1892 int i, resid; 1893 vm_ooffset_t foff; 1894 vm_page_t m; 1895 vm_object_t obj; 1896 int iosize; 1897 struct vnode *vp = bp->b_vp; 1898 1899 obj = vp->v_object; 1900 1901#if defined(VFS_BIO_DEBUG) 1902 if (vp->v_usecount == 0) { 1903 panic("biodone: zero vnode ref count"); 1904 } 1905 1906 if (vp->v_object == NULL) { 1907 panic("biodone: missing VM object"); 1908 } 1909 1910 if ((vp->v_flag & VOBJBUF) == 0) { 1911 panic("biodone: vnode is not setup for merged cache"); 1912 } 1913#endif 1914 1915 foff = bp->b_offset; 1916#ifdef DIAGNOSTIC 1917 if (bp->b_offset == NOOFFSET) 1918 panic("biodone: no buffer offset"); 1919#endif 1920 1921#if !defined(MAX_PERF) 1922 if (!obj) { 1923 panic("biodone: no object"); 1924 } 1925#endif 1926#if defined(VFS_BIO_DEBUG) 1927 if (obj->paging_in_progress < bp->b_npages) { 1928 printf("biodone: paging in progress(%d) < bp->b_npages(%d)\n", 1929 obj->paging_in_progress, bp->b_npages); 1930 } 1931#endif 1932 iosize = bp->b_bufsize; 1933 for (i = 0; i < bp->b_npages; i++) { 1934 int bogusflag = 0; 1935 m = bp->b_pages[i]; 1936 if (m == bogus_page) { 1937 bogusflag = 1; 1938 m = vm_page_lookup(obj, OFF_TO_IDX(foff)); 1939 if (!m) { 1940#if defined(VFS_BIO_DEBUG) 1941 printf("biodone: page disappeared\n"); 1942#endif 1943 vm_object_pip_subtract(obj, 1); 1944 continue; 1945 } 1946 bp->b_pages[i] = m; 1947 pmap_qenter(trunc_page((vm_offset_t)bp->b_data), bp->b_pages, bp->b_npages); 1948 } 1949#if defined(VFS_BIO_DEBUG) 1950 if (OFF_TO_IDX(foff) != m->pindex) { 1951 printf("biodone: foff(%d)/m->pindex(%d) mismatch\n", foff, m->pindex); 1952 } 1953#endif 1954 resid = IDX_TO_OFF(m->pindex + 1) - foff; 1955 if (resid > iosize) 1956 resid = iosize; 1957 1958 /* 1959 * In the write case, the valid and clean bits are 1960 * already changed correctly, so we only need to do this 1961 * here in the read case. 1962 */ 1963 if ((bp->b_flags & B_READ) && !bogusflag && resid > 0) { 1964 vfs_page_set_valid(bp, foff, i, m); 1965 } 1966 vm_page_flag_clear(m, PG_ZERO); 1967 1968 /* 1969 * when debugging new filesystems or buffer I/O methods, this 1970 * is the most common error that pops up. if you see this, you 1971 * have not set the page busy flag correctly!!! 1972 */ 1973 if (m->busy == 0) { 1974#if !defined(MAX_PERF) 1975 printf("biodone: page busy < 0, " 1976 "pindex: %d, foff: 0x(%x,%x), " 1977 "resid: %d, index: %d\n", 1978 (int) m->pindex, (int)(foff >> 32), 1979 (int) foff & 0xffffffff, resid, i); 1980#endif 1981 if (vp->v_type != VBLK) 1982#if !defined(MAX_PERF) 1983 printf(" iosize: %ld, lblkno: %d, flags: 0x%lx, npages: %d\n", 1984 bp->b_vp->v_mount->mnt_stat.f_iosize, 1985 (int) bp->b_lblkno, 1986 bp->b_flags, bp->b_npages); 1987 else 1988 printf(" VDEV, lblkno: %d, flags: 0x%lx, npages: %d\n", 1989 (int) bp->b_lblkno, 1990 bp->b_flags, bp->b_npages); 1991 printf(" valid: 0x%x, dirty: 0x%x, wired: %d\n", 1992 m->valid, m->dirty, m->wire_count); 1993#endif 1994 panic("biodone: page busy < 0\n"); 1995 } 1996 vm_page_io_finish(m); 1997 vm_object_pip_subtract(obj, 1); 1998 foff += resid; 1999 iosize -= resid; 2000 } 2001 if (obj && 2002 (obj->paging_in_progress == 0) && 2003 (obj->flags & OBJ_PIPWNT)) { 2004 vm_object_clear_flag(obj, OBJ_PIPWNT); 2005 wakeup(obj); 2006 } 2007 } 2008 /* 2009 * For asynchronous completions, release the buffer now. The brelse 2010 * checks for B_WANTED and will do the wakeup there if necessary - so 2011 * no need to do a wakeup here in the async case. 2012 */ 2013 2014 if (bp->b_flags & B_ASYNC) { 2015 if ((bp->b_flags & (B_NOCACHE | B_INVAL | B_ERROR | B_RELBUF)) != 0) 2016 brelse(bp); 2017 else 2018 bqrelse(bp); 2019 } else { 2020 bp->b_flags &= ~B_WANTED; 2021 wakeup(bp); 2022 } 2023 splx(s); 2024} 2025 2026#if 0 /* not with kirks code */ 2027static int vfs_update_interval = 30; 2028 2029static void 2030vfs_update() 2031{ 2032 while (1) { 2033 tsleep(&vfs_update_wakeup, PUSER, "update", 2034 hz * vfs_update_interval); 2035 vfs_update_wakeup = 0; 2036 sync(curproc, NULL); 2037 } 2038} 2039 2040static int 2041sysctl_kern_updateinterval SYSCTL_HANDLER_ARGS 2042{ 2043 int error = sysctl_handle_int(oidp, 2044 oidp->oid_arg1, oidp->oid_arg2, req); 2045 if (!error) 2046 wakeup(&vfs_update_wakeup); 2047 return error; 2048} 2049 2050SYSCTL_PROC(_kern, KERN_UPDATEINTERVAL, update, CTLTYPE_INT|CTLFLAG_RW, 2051 &vfs_update_interval, 0, sysctl_kern_updateinterval, "I", ""); 2052 2053#endif 2054 2055 2056/* 2057 * This routine is called in lieu of iodone in the case of 2058 * incomplete I/O. This keeps the busy status for pages 2059 * consistant. 2060 */ 2061void 2062vfs_unbusy_pages(struct buf * bp) 2063{ 2064 int i; 2065 2066 if (bp->b_flags & B_VMIO) { 2067 struct vnode *vp = bp->b_vp; 2068 vm_object_t obj = vp->v_object; 2069 2070 for (i = 0; i < bp->b_npages; i++) { 2071 vm_page_t m = bp->b_pages[i]; 2072 2073 if (m == bogus_page) { 2074 m = vm_page_lookup(obj, OFF_TO_IDX(bp->b_offset) + i); 2075#if !defined(MAX_PERF) 2076 if (!m) { 2077 panic("vfs_unbusy_pages: page missing\n"); 2078 } 2079#endif 2080 bp->b_pages[i] = m; 2081 pmap_qenter(trunc_page((vm_offset_t)bp->b_data), bp->b_pages, bp->b_npages); 2082 } 2083 vm_object_pip_subtract(obj, 1); 2084 vm_page_flag_clear(m, PG_ZERO); 2085 vm_page_io_finish(m); 2086 } 2087 if (obj->paging_in_progress == 0 && 2088 (obj->flags & OBJ_PIPWNT)) { 2089 vm_object_clear_flag(obj, OBJ_PIPWNT); 2090 wakeup(obj); 2091 } 2092 } 2093} 2094 2095/* 2096 * Set NFS' b_validoff and b_validend fields from the valid bits 2097 * of a page. If the consumer is not NFS, and the page is not 2098 * valid for the entire range, clear the B_CACHE flag to force 2099 * the consumer to re-read the page. 2100 */ 2101static void 2102vfs_buf_set_valid(struct buf *bp, 2103 vm_ooffset_t foff, vm_offset_t off, vm_offset_t size, 2104 vm_page_t m) 2105{ 2106 if (bp->b_vp->v_tag == VT_NFS && bp->b_vp->v_type != VBLK) { 2107 vm_offset_t svalid, evalid; 2108 int validbits = m->valid; 2109 2110 /* 2111 * This only bothers with the first valid range in the 2112 * page. 2113 */ 2114 svalid = off; 2115 while (validbits && !(validbits & 1)) { 2116 svalid += DEV_BSIZE; 2117 validbits >>= 1; 2118 } 2119 evalid = svalid; 2120 while (validbits & 1) { 2121 evalid += DEV_BSIZE; 2122 validbits >>= 1; 2123 } 2124 /* 2125 * Make sure this range is contiguous with the range 2126 * built up from previous pages. If not, then we will 2127 * just use the range from the previous pages. 2128 */ 2129 if (svalid == bp->b_validend) { 2130 bp->b_validoff = min(bp->b_validoff, svalid); 2131 bp->b_validend = max(bp->b_validend, evalid); 2132 } 2133 } else if (!vm_page_is_valid(m, 2134 (vm_offset_t) ((foff + off) & PAGE_MASK), 2135 size)) { 2136 bp->b_flags &= ~B_CACHE; 2137 } 2138} 2139 2140/* 2141 * Set the valid bits in a page, taking care of the b_validoff, 2142 * b_validend fields which NFS uses to optimise small reads. Off is 2143 * the offset within the file and pageno is the page index within the buf. 2144 */ 2145static void 2146vfs_page_set_valid(struct buf *bp, vm_ooffset_t off, int pageno, vm_page_t m) 2147{ 2148 struct vnode *vp = bp->b_vp; 2149 vm_ooffset_t soff, eoff; 2150 2151 soff = off; 2152 eoff = off + min(PAGE_SIZE, bp->b_bufsize); 2153 if (vp->v_tag == VT_NFS && vp->v_type != VBLK) { 2154 vm_ooffset_t sv, ev; 2155 vm_page_set_invalid(m, 2156 (vm_offset_t) (soff & PAGE_MASK), 2157 (vm_offset_t) (eoff - soff)); 2158 off = off - pageno * PAGE_SIZE; 2159 sv = off + ((bp->b_validoff + DEV_BSIZE - 1) & ~(DEV_BSIZE - 1)); 2160 ev = off + ((bp->b_validend + DEV_BSIZE - 1) & ~(DEV_BSIZE - 1)); 2161 soff = qmax(sv, soff); 2162 eoff = qmin(ev, eoff); 2163 } 2164 if (eoff > soff) 2165 vm_page_set_validclean(m, 2166 (vm_offset_t) (soff & PAGE_MASK), 2167 (vm_offset_t) (eoff - soff)); 2168} 2169 2170/* 2171 * This routine is called before a device strategy routine. 2172 * It is used to tell the VM system that paging I/O is in 2173 * progress, and treat the pages associated with the buffer 2174 * almost as being PG_BUSY. Also the object paging_in_progress 2175 * flag is handled to make sure that the object doesn't become 2176 * inconsistant. 2177 */ 2178void 2179vfs_busy_pages(struct buf * bp, int clear_modify) 2180{ 2181 int i; 2182 2183 if (bp->b_flags & B_VMIO) { 2184 struct vnode *vp = bp->b_vp; 2185 vm_object_t obj = vp->v_object; 2186 vm_ooffset_t foff; 2187 2188 foff = bp->b_offset; 2189#ifdef DIAGNOSTIC 2190 if (bp->b_offset == NOOFFSET) 2191 panic("vfs_busy_pages: no buffer offset"); 2192#endif 2193 2194 vfs_setdirty(bp); 2195 2196retry: 2197 for (i = 0; i < bp->b_npages; i++) { 2198 vm_page_t m = bp->b_pages[i]; 2199 if (vm_page_sleep(m, "vbpage", NULL)) 2200 goto retry; 2201 } 2202 2203 for (i = 0; i < bp->b_npages; i++, foff += PAGE_SIZE) { 2204 vm_page_t m = bp->b_pages[i]; 2205 2206 vm_page_flag_clear(m, PG_ZERO); 2207 if ((bp->b_flags & B_CLUSTER) == 0) { 2208 vm_object_pip_add(obj, 1); 2209 vm_page_io_start(m); 2210 } 2211 2212 vm_page_protect(m, VM_PROT_NONE); 2213 if (clear_modify) 2214 vfs_page_set_valid(bp, foff, i, m); 2215 else if (bp->b_bcount >= PAGE_SIZE) { 2216 if (m->valid && (bp->b_flags & B_CACHE) == 0) { 2217 bp->b_pages[i] = bogus_page; 2218 pmap_qenter(trunc_page((vm_offset_t)bp->b_data), bp->b_pages, bp->b_npages); 2219 } 2220 } 2221 } 2222 } 2223} 2224 2225/* 2226 * Tell the VM system that the pages associated with this buffer 2227 * are clean. This is used for delayed writes where the data is 2228 * going to go to disk eventually without additional VM intevention. 2229 */ 2230void 2231vfs_clean_pages(struct buf * bp) 2232{ 2233 int i; 2234 2235 if (bp->b_flags & B_VMIO) { 2236 vm_ooffset_t foff; 2237 foff = bp->b_offset; 2238 2239#ifdef DIAGNOSTIC 2240 if (bp->b_offset == NOOFFSET) 2241 panic("vfs_clean_pages: no buffer offset"); 2242#endif 2243 2244 for (i = 0; i < bp->b_npages; i++, foff += PAGE_SIZE) { 2245 vm_page_t m = bp->b_pages[i]; 2246 vfs_page_set_valid(bp, foff, i, m); 2247 } 2248 } 2249} 2250 2251void 2252vfs_bio_clrbuf(struct buf *bp) { 2253 int i; 2254 if ((bp->b_flags & (B_VMIO | B_MALLOC)) == B_VMIO) { 2255 if( (bp->b_npages == 1) && (bp->b_bufsize < PAGE_SIZE)) { 2256 int mask; 2257 mask = 0; 2258 for(i=0;i<bp->b_bufsize;i+=DEV_BSIZE) 2259 mask |= (1 << (i/DEV_BSIZE)); 2260 if(((bp->b_pages[0]->flags & PG_ZERO) == 0) && 2261 (bp->b_pages[0]->valid != mask)) { 2262 bzero(bp->b_data, bp->b_bufsize); 2263 } 2264 bp->b_pages[0]->valid = mask; 2265 bp->b_resid = 0; 2266 return; 2267 } 2268 for(i=0;i<bp->b_npages;i++) { 2269 if( bp->b_pages[i]->valid == VM_PAGE_BITS_ALL) 2270 continue; 2271 if( bp->b_pages[i]->valid == 0) { 2272 if ((bp->b_pages[i]->flags & PG_ZERO) == 0) { 2273 bzero(bp->b_data + (i << PAGE_SHIFT), PAGE_SIZE); 2274 } 2275 } else { 2276 int j; 2277 for(j=0;j<PAGE_SIZE/DEV_BSIZE;j++) { 2278 if (((bp->b_pages[i]->flags & PG_ZERO) == 0) && 2279 (bp->b_pages[i]->valid & (1<<j)) == 0) 2280 bzero(bp->b_data + (i << PAGE_SHIFT) + j * DEV_BSIZE, DEV_BSIZE); 2281 } 2282 } 2283 bp->b_pages[i]->valid = VM_PAGE_BITS_ALL; 2284 vm_page_flag_clear(bp->b_pages[i], PG_ZERO); 2285 } 2286 bp->b_resid = 0; 2287 } else { 2288 clrbuf(bp); 2289 } 2290} 2291 2292/* 2293 * vm_hold_load_pages and vm_hold_unload pages get pages into 2294 * a buffers address space. The pages are anonymous and are 2295 * not associated with a file object. 2296 */ 2297void 2298vm_hold_load_pages(struct buf * bp, vm_offset_t from, vm_offset_t to) 2299{ 2300 vm_offset_t pg; 2301 vm_page_t p; 2302 int index; 2303 2304 to = round_page(to); 2305 from = round_page(from); 2306 index = (from - trunc_page((vm_offset_t)bp->b_data)) >> PAGE_SHIFT; 2307 2308 for (pg = from; pg < to; pg += PAGE_SIZE, index++) { 2309 2310tryagain: 2311 2312 p = vm_page_alloc(kernel_object, 2313 ((pg - VM_MIN_KERNEL_ADDRESS) >> PAGE_SHIFT), 2314 VM_ALLOC_NORMAL); 2315 if (!p) { 2316 vm_pageout_deficit += (to - from) >> PAGE_SHIFT; 2317 VM_WAIT; 2318 goto tryagain; 2319 } 2320 vm_page_wire(p); 2321 p->valid = VM_PAGE_BITS_ALL; 2322 vm_page_flag_clear(p, PG_ZERO); 2323 pmap_kenter(pg, VM_PAGE_TO_PHYS(p)); 2324 bp->b_pages[index] = p; 2325 vm_page_wakeup(p); 2326 } 2327 bp->b_npages = index; 2328} 2329 2330void 2331vm_hold_free_pages(struct buf * bp, vm_offset_t from, vm_offset_t to) 2332{ 2333 vm_offset_t pg; 2334 vm_page_t p; 2335 int index, newnpages; 2336 2337 from = round_page(from); 2338 to = round_page(to); 2339 newnpages = index = (from - trunc_page((vm_offset_t)bp->b_data)) >> PAGE_SHIFT; 2340 2341 for (pg = from; pg < to; pg += PAGE_SIZE, index++) { 2342 p = bp->b_pages[index]; 2343 if (p && (index < bp->b_npages)) { 2344#if !defined(MAX_PERF) 2345 if (p->busy) { 2346 printf("vm_hold_free_pages: blkno: %d, lblkno: %d\n", 2347 bp->b_blkno, bp->b_lblkno); 2348 } 2349#endif 2350 bp->b_pages[index] = NULL; 2351 pmap_kremove(pg); 2352 vm_page_busy(p); 2353 vm_page_unwire(p, 0); 2354 vm_page_free(p); 2355 } 2356 } 2357 bp->b_npages = newnpages; 2358} 2359 2360 2361#include "opt_ddb.h" 2362#ifdef DDB 2363#include <ddb/ddb.h> 2364 2365DB_SHOW_COMMAND(buffer, db_show_buffer) 2366{ 2367 /* get args */ 2368 struct buf *bp = (struct buf *)addr; 2369 2370 if (!have_addr) { 2371 db_printf("usage: show buffer <addr>\n"); 2372 return; 2373 } 2374 2375 db_printf("b_proc = %p,\nb_flags = 0x%b\n", (void *)bp->b_proc, 2376 (u_int)bp->b_flags, PRINT_BUF_FLAGS); 2377 db_printf("b_error = %d, b_bufsize = %ld, b_bcount = %ld, " 2378 "b_resid = %ld\nb_dev = 0x%x, b_data = %p, " 2379 "b_blkno = %d, b_pblkno = %d\n", 2380 bp->b_error, bp->b_bufsize, bp->b_bcount, bp->b_resid, 2381 bp->b_dev, bp->b_data, bp->b_blkno, bp->b_pblkno); 2382 if (bp->b_npages) { 2383 int i; 2384 db_printf("b_npages = %d, pages(OBJ, IDX, PA): ", bp->b_npages); 2385 for (i = 0; i < bp->b_npages; i++) { 2386 vm_page_t m; 2387 m = bp->b_pages[i]; 2388 db_printf("(%p, 0x%lx, 0x%lx)", (void *)m->object, 2389 (u_long)m->pindex, (u_long)VM_PAGE_TO_PHYS(m)); 2390 if ((i + 1) < bp->b_npages) 2391 db_printf(","); 2392 } 2393 db_printf("\n"); 2394 } 2395} 2396#endif /* DDB */ 2397