nfs_bio.c revision 12911
1/* 2 * Copyright (c) 1989, 1993 3 * The Regents of the University of California. All rights reserved. 4 * 5 * This code is derived from software contributed to Berkeley by 6 * Rick Macklem at The University of Guelph. 7 * 8 * Redistribution and use in source and binary forms, with or without 9 * modification, are permitted provided that the following conditions 10 * are met: 11 * 1. Redistributions of source code must retain the above copyright 12 * notice, this list of conditions and the following disclaimer. 13 * 2. Redistributions in binary form must reproduce the above copyright 14 * notice, this list of conditions and the following disclaimer in the 15 * documentation and/or other materials provided with the distribution. 16 * 3. All advertising materials mentioning features or use of this software 17 * must display the following acknowledgement: 18 * This product includes software developed by the University of 19 * California, Berkeley and its contributors. 20 * 4. Neither the name of the University nor the names of its contributors 21 * may be used to endorse or promote products derived from this software 22 * without specific prior written permission. 23 * 24 * THIS SOFTWARE IS PROVIDED BY THE REGENTS AND CONTRIBUTORS ``AS IS'' AND 25 * ANY EXPRESS OR IMPLIED WARRANTIES, INCLUDING, BUT NOT LIMITED TO, THE 26 * IMPLIED WARRANTIES OF MERCHANTABILITY AND FITNESS FOR A PARTICULAR PURPOSE 27 * ARE DISCLAIMED. IN NO EVENT SHALL THE REGENTS OR CONTRIBUTORS BE LIABLE 28 * FOR ANY DIRECT, INDIRECT, INCIDENTAL, SPECIAL, EXEMPLARY, OR CONSEQUENTIAL 29 * DAMAGES (INCLUDING, BUT NOT LIMITED TO, PROCUREMENT OF SUBSTITUTE GOODS 30 * OR SERVICES; LOSS OF USE, DATA, OR PROFITS; OR BUSINESS INTERRUPTION) 31 * HOWEVER CAUSED AND ON ANY THEORY OF LIABILITY, WHETHER IN CONTRACT, STRICT 32 * LIABILITY, OR TORT (INCLUDING NEGLIGENCE OR OTHERWISE) ARISING IN ANY WAY 33 * OUT OF THE USE OF THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF 34 * SUCH DAMAGE. 35 * 36 * @(#)nfs_bio.c 8.5 (Berkeley) 1/4/94 37 * $Id: nfs_bio.c,v 1.20 1995/12/07 12:47:23 davidg Exp $ 38 */ 39 40#include <sys/param.h> 41#include <sys/systm.h> 42#include <sys/resourcevar.h> 43#include <sys/signalvar.h> 44#include <sys/proc.h> 45#include <sys/buf.h> 46#include <sys/vnode.h> 47#include <sys/mount.h> 48#include <sys/kernel.h> 49 50#include <vm/vm.h> 51#include <vm/vm_param.h> 52#include <vm/vm_extern.h> 53 54#include <nfs/rpcv2.h> 55#include <nfs/nfsproto.h> 56#include <nfs/nfs.h> 57#include <nfs/nfsmount.h> 58#include <nfs/nqnfs.h> 59#include <nfs/nfsnode.h> 60 61static struct buf *nfs_getcacheblk __P((struct vnode *vp, daddr_t bn, int size, 62 struct proc *p)); 63 64extern struct proc *nfs_iodwant[NFS_MAXASYNCDAEMON]; 65extern int nfs_numasync; 66extern struct nfsstats nfsstats; 67 68/* 69 * Ifdefs for FreeBSD-current's merged VM/buffer cache. It is unfortunate 70 * that this isn't done inside getblk() and brelse() so these calls 71 * wouldn't need to be here. 72 */ 73#ifdef B_VMIO 74#define vnode_pager_uncache(vp) 75#else 76#define vfs_busy_pages(bp, f) 77#define vfs_unbusy_pages(bp) 78#define vfs_dirty_pages(bp) 79#endif 80 81/* 82 * Vnode op for read using bio 83 * Any similarity to readip() is purely coincidental 84 */ 85int 86nfs_bioread(vp, uio, ioflag, cred) 87 register struct vnode *vp; 88 register struct uio *uio; 89 int ioflag; 90 struct ucred *cred; 91{ 92 register struct nfsnode *np = VTONFS(vp); 93 register int biosize, diff, i; 94 struct buf *bp = 0, *rabp; 95 struct vattr vattr; 96 struct proc *p; 97 struct nfsmount *nmp = VFSTONFS(vp->v_mount); 98 daddr_t lbn, rabn; 99 int bufsize; 100 int nra, error = 0, n = 0, on = 0, not_readin; 101 102#ifdef DIAGNOSTIC 103 if (uio->uio_rw != UIO_READ) 104 panic("nfs_read mode"); 105#endif 106 if (uio->uio_resid == 0) 107 return (0); 108 if (uio->uio_offset < 0) 109 return (EINVAL); 110 p = uio->uio_procp; 111 if ((nmp->nm_flag & (NFSMNT_NFSV3 | NFSMNT_GOTFSINFO)) == NFSMNT_NFSV3) 112 (void)nfs_fsinfo(nmp, vp, cred, p); 113 biosize = vp->v_mount->mnt_stat.f_iosize; 114 /* 115 * For nfs, cache consistency can only be maintained approximately. 116 * Although RFC1094 does not specify the criteria, the following is 117 * believed to be compatible with the reference port. 118 * For nqnfs, full cache consistency is maintained within the loop. 119 * For nfs: 120 * If the file's modify time on the server has changed since the 121 * last read rpc or you have written to the file, 122 * you may have lost data cache consistency with the 123 * server, so flush all of the file's data out of the cache. 124 * Then force a getattr rpc to ensure that you have up to date 125 * attributes. 126 * NB: This implies that cache data can be read when up to 127 * NFS_ATTRTIMEO seconds out of date. If you find that you need current 128 * attributes this could be forced by setting n_attrstamp to 0 before 129 * the VOP_GETATTR() call. 130 */ 131 if ((nmp->nm_flag & NFSMNT_NQNFS) == 0) { 132 if (np->n_flag & NMODIFIED) { 133 if (vp->v_type != VREG) { 134 if (vp->v_type != VDIR) 135 panic("nfs: bioread, not dir"); 136 nfs_invaldir(vp); 137 error = nfs_vinvalbuf(vp, V_SAVE, cred, p, 1); 138 if (error) 139 return (error); 140 } 141 np->n_attrstamp = 0; 142 error = VOP_GETATTR(vp, &vattr, cred, p); 143 if (error) 144 return (error); 145 np->n_mtime = vattr.va_mtime.ts_sec; 146 } else { 147 error = VOP_GETATTR(vp, &vattr, cred, p); 148 if (error) 149 return (error); 150 if (np->n_mtime != vattr.va_mtime.ts_sec) { 151 if (vp->v_type == VDIR) 152 nfs_invaldir(vp); 153 error = nfs_vinvalbuf(vp, V_SAVE, cred, p, 1); 154 if (error) 155 return (error); 156 np->n_mtime = vattr.va_mtime.ts_sec; 157 } 158 } 159 } 160 do { 161 162 /* 163 * Get a valid lease. If cached data is stale, flush it. 164 */ 165 if (nmp->nm_flag & NFSMNT_NQNFS) { 166 if (NQNFS_CKINVALID(vp, np, ND_READ)) { 167 do { 168 error = nqnfs_getlease(vp, ND_READ, cred, p); 169 } while (error == NQNFS_EXPIRED); 170 if (error) 171 return (error); 172 if (np->n_lrev != np->n_brev || 173 (np->n_flag & NQNFSNONCACHE) || 174 ((np->n_flag & NMODIFIED) && vp->v_type == VDIR)) { 175 if (vp->v_type == VDIR) 176 nfs_invaldir(vp); 177 error = nfs_vinvalbuf(vp, V_SAVE, cred, p, 1); 178 if (error) 179 return (error); 180 np->n_brev = np->n_lrev; 181 } 182 } else if (vp->v_type == VDIR && (np->n_flag & NMODIFIED)) { 183 nfs_invaldir(vp); 184 error = nfs_vinvalbuf(vp, V_SAVE, cred, p, 1); 185 if (error) 186 return (error); 187 } 188 } 189 if (np->n_flag & NQNFSNONCACHE) { 190 switch (vp->v_type) { 191 case VREG: 192 return (nfs_readrpc(vp, uio, cred)); 193 case VLNK: 194 return (nfs_readlinkrpc(vp, uio, cred)); 195 case VDIR: 196 break; 197 default: 198 printf(" NQNFSNONCACHE: type %x unexpected\n", 199 vp->v_type); 200 }; 201 } 202 switch (vp->v_type) { 203 case VREG: 204 nfsstats.biocache_reads++; 205 lbn = uio->uio_offset / biosize; 206 on = uio->uio_offset & (biosize - 1); 207 not_readin = 1; 208 209 /* 210 * Start the read ahead(s), as required. 211 */ 212 if (nfs_numasync > 0 && nmp->nm_readahead > 0) { 213 for (nra = 0; nra < nmp->nm_readahead && 214 (lbn + 1 + nra) * biosize < np->n_size; nra++) { 215 rabn = lbn + 1 + nra; 216 if (!incore(vp, rabn)) { 217 rabp = nfs_getcacheblk(vp, rabn, biosize, p); 218 if (!rabp) 219 return (EINTR); 220 if ((rabp->b_flags & (B_CACHE|B_DELWRI)) == 0) { 221 rabp->b_flags |= (B_READ | B_ASYNC); 222 vfs_busy_pages(rabp, 0); 223 if (nfs_asyncio(rabp, cred)) { 224 rabp->b_flags |= B_INVAL|B_ERROR; 225 vfs_unbusy_pages(rabp); 226 brelse(rabp); 227 } 228 } else { 229 brelse(rabp); 230 } 231 } 232 } 233 } 234 235 /* 236 * If the block is in the cache and has the required data 237 * in a valid region, just copy it out. 238 * Otherwise, get the block and write back/read in, 239 * as required. 240 */ 241again: 242 bufsize = biosize; 243 if ((lbn + 1) * biosize > np->n_size) { 244 bufsize = np->n_size - lbn * biosize; 245 bufsize = (bufsize + DEV_BSIZE - 1) & ~(DEV_BSIZE - 1); 246 } 247 bp = nfs_getcacheblk(vp, lbn, bufsize, p); 248 if (!bp) 249 return (EINTR); 250 if ((bp->b_flags & B_CACHE) == 0) { 251 bp->b_flags |= B_READ; 252 not_readin = 0; 253 vfs_busy_pages(bp, 0); 254 error = nfs_doio(bp, cred, p); 255 if (error) { 256 brelse(bp); 257 return (error); 258 } 259 } 260 if (bufsize > on) { 261 n = min((unsigned)(bufsize - on), uio->uio_resid); 262 } else { 263 n = 0; 264 } 265 diff = np->n_size - uio->uio_offset; 266 if (diff < n) 267 n = diff; 268 if (not_readin && n > 0) { 269 if (on < bp->b_validoff || (on + n) > bp->b_validend) { 270 bp->b_flags |= B_NOCACHE; 271 if (bp->b_dirtyend > 0) { 272 if ((bp->b_flags & B_DELWRI) == 0) 273 panic("nfsbioread"); 274 if (VOP_BWRITE(bp) == EINTR) 275 return (EINTR); 276 } else 277 brelse(bp); 278 goto again; 279 } 280 } 281 vp->v_lastr = lbn; 282 diff = (on >= bp->b_validend) ? 0 : (bp->b_validend - on); 283 if (diff < n) 284 n = diff; 285 break; 286 case VLNK: 287 nfsstats.biocache_readlinks++; 288 bp = nfs_getcacheblk(vp, (daddr_t)0, NFS_MAXPATHLEN, p); 289 if (!bp) 290 return (EINTR); 291 if ((bp->b_flags & B_CACHE) == 0) { 292 bp->b_flags |= B_READ; 293 vfs_busy_pages(bp, 0); 294 error = nfs_doio(bp, cred, p); 295 if (error) { 296 bp->b_flags |= B_ERROR; 297 brelse(bp); 298 return (error); 299 } 300 } 301 n = min(uio->uio_resid, NFS_MAXPATHLEN - bp->b_resid); 302 on = 0; 303 break; 304 case VDIR: 305 nfsstats.biocache_readdirs++; 306 lbn = uio->uio_offset / NFS_DIRBLKSIZ; 307 on = uio->uio_offset & (NFS_DIRBLKSIZ - 1); 308 bp = nfs_getcacheblk(vp, lbn, NFS_DIRBLKSIZ, p); 309 if (!bp) 310 return (EINTR); 311 if ((bp->b_flags & B_CACHE) == 0) { 312 bp->b_flags |= B_READ; 313 vfs_busy_pages(bp, 0); 314 error = nfs_doio(bp, cred, p); 315 if (error) { 316 brelse(bp); 317 while (error == NFSERR_BAD_COOKIE) { 318 nfs_invaldir(vp); 319 error = nfs_vinvalbuf(vp, 0, cred, p, 1); 320 /* 321 * Yuck! The directory has been modified on the 322 * server. The only way to get the block is by 323 * reading from the beginning to get all the 324 * offset cookies. 325 */ 326 for (i = 0; i <= lbn && !error; i++) { 327 bp = nfs_getcacheblk(vp, i, NFS_DIRBLKSIZ, p); 328 if (!bp) 329 return (EINTR); 330 if ((bp->b_flags & B_DONE) == 0) { 331 bp->b_flags |= B_READ; 332 vfs_busy_pages(bp, 0); 333 error = nfs_doio(bp, cred, p); 334 if (error) 335 brelse(bp); 336 } 337 } 338 } 339 if (error) 340 return (error); 341 } 342 } 343 344 /* 345 * If not eof and read aheads are enabled, start one. 346 * (You need the current block first, so that you have the 347 * directory offset cookie of the next block.) 348 */ 349 if (nfs_numasync > 0 && nmp->nm_readahead > 0 && 350 (np->n_direofoffset == 0 || 351 (lbn + 1) * NFS_DIRBLKSIZ < np->n_direofoffset) && 352 !(np->n_flag & NQNFSNONCACHE) && 353 !incore(vp, lbn + 1)) { 354 rabp = nfs_getcacheblk(vp, lbn + 1, NFS_DIRBLKSIZ, p); 355 if (rabp) { 356 if ((rabp->b_flags & (B_CACHE|B_DELWRI)) == 0) { 357 rabp->b_flags |= (B_READ | B_ASYNC); 358 vfs_busy_pages(rabp, 0); 359 if (nfs_asyncio(rabp, cred)) { 360 rabp->b_flags |= B_INVAL|B_ERROR; 361 vfs_unbusy_pages(rabp); 362 brelse(rabp); 363 } 364 } else { 365 brelse(rabp); 366 } 367 } 368 } 369 n = min(uio->uio_resid, NFS_DIRBLKSIZ - bp->b_resid - on); 370 break; 371 default: 372 printf(" nfs_bioread: type %x unexpected\n",vp->v_type); 373 break; 374 }; 375 376 if (n > 0) { 377 error = uiomove(bp->b_data + on, (int)n, uio); 378 } 379 switch (vp->v_type) { 380 case VREG: 381 break; 382 case VLNK: 383 n = 0; 384 break; 385 case VDIR: 386 if (np->n_flag & NQNFSNONCACHE) 387 bp->b_flags |= B_INVAL; 388 break; 389 default: 390 printf(" nfs_bioread: type %x unexpected\n",vp->v_type); 391 } 392 brelse(bp); 393 } while (error == 0 && uio->uio_resid > 0 && n > 0); 394 return (error); 395} 396 397/* 398 * Vnode op for write using bio 399 */ 400int 401nfs_write(ap) 402 struct vop_write_args /* { 403 struct vnode *a_vp; 404 struct uio *a_uio; 405 int a_ioflag; 406 struct ucred *a_cred; 407 } */ *ap; 408{ 409 register int biosize; 410 register struct uio *uio = ap->a_uio; 411 struct proc *p = uio->uio_procp; 412 register struct vnode *vp = ap->a_vp; 413 struct nfsnode *np = VTONFS(vp); 414 register struct ucred *cred = ap->a_cred; 415 int ioflag = ap->a_ioflag; 416 struct buf *bp; 417 struct vattr vattr; 418 struct nfsmount *nmp = VFSTONFS(vp->v_mount); 419 daddr_t lbn; 420 int bufsize; 421 int n, on, error = 0, iomode, must_commit; 422 423#ifdef DIAGNOSTIC 424 if (uio->uio_rw != UIO_WRITE) 425 panic("nfs_write mode"); 426 if (uio->uio_segflg == UIO_USERSPACE && uio->uio_procp != curproc) 427 panic("nfs_write proc"); 428#endif 429 if (vp->v_type != VREG) 430 return (EIO); 431 if (np->n_flag & NWRITEERR) { 432 np->n_flag &= ~NWRITEERR; 433 return (np->n_error); 434 } 435 if ((nmp->nm_flag & (NFSMNT_NFSV3 | NFSMNT_GOTFSINFO)) == NFSMNT_NFSV3) 436 (void)nfs_fsinfo(nmp, vp, cred, p); 437 if (ioflag & (IO_APPEND | IO_SYNC)) { 438 if (np->n_flag & NMODIFIED) { 439 np->n_attrstamp = 0; 440 error = nfs_vinvalbuf(vp, V_SAVE, cred, p, 1); 441 if (error) 442 return (error); 443 } 444 if (ioflag & IO_APPEND) { 445 np->n_attrstamp = 0; 446 error = VOP_GETATTR(vp, &vattr, cred, p); 447 if (error) 448 return (error); 449 uio->uio_offset = np->n_size; 450 } 451 } 452 if (uio->uio_offset < 0) 453 return (EINVAL); 454 if (uio->uio_resid == 0) 455 return (0); 456 /* 457 * Maybe this should be above the vnode op call, but so long as 458 * file servers have no limits, i don't think it matters 459 */ 460 if (p && uio->uio_offset + uio->uio_resid > 461 p->p_rlimit[RLIMIT_FSIZE].rlim_cur) { 462 psignal(p, SIGXFSZ); 463 return (EFBIG); 464 } 465 /* 466 * I use nm_rsize, not nm_wsize so that all buffer cache blocks 467 * will be the same size within a filesystem. nfs_writerpc will 468 * still use nm_wsize when sizing the rpc's. 469 */ 470 biosize = vp->v_mount->mnt_stat.f_iosize; 471 do { 472 473 /* 474 * XXX make sure we aren't cached in the VM page cache 475 */ 476 /* 477 * Check for a valid write lease. 478 */ 479 if ((nmp->nm_flag & NFSMNT_NQNFS) && 480 NQNFS_CKINVALID(vp, np, ND_WRITE)) { 481 do { 482 error = nqnfs_getlease(vp, ND_WRITE, cred, p); 483 } while (error == NQNFS_EXPIRED); 484 if (error) 485 return (error); 486 if (np->n_lrev != np->n_brev || 487 (np->n_flag & NQNFSNONCACHE)) { 488 error = nfs_vinvalbuf(vp, V_SAVE, cred, p, 1); 489 if (error) 490 return (error); 491 np->n_brev = np->n_lrev; 492 } 493 } 494 if ((np->n_flag & NQNFSNONCACHE) && uio->uio_iovcnt == 1) { 495 iomode = NFSV3WRITE_FILESYNC; 496 error = nfs_writerpc(vp, uio, cred, &iomode, &must_commit); 497 if (must_commit) 498 nfs_clearcommit(vp->v_mount); 499 return (error); 500 } 501 nfsstats.biocache_writes++; 502 lbn = uio->uio_offset / biosize; 503 on = uio->uio_offset & (biosize-1); 504 n = min((unsigned)(biosize - on), uio->uio_resid); 505again: 506 if (uio->uio_offset + n > np->n_size) { 507 np->n_size = uio->uio_offset + n; 508 vnode_pager_setsize(vp, (u_long)np->n_size); 509 } 510 bufsize = biosize; 511 if ((lbn + 1) * biosize > np->n_size) { 512 bufsize = np->n_size - lbn * biosize; 513 bufsize = (bufsize + DEV_BSIZE - 1) & ~(DEV_BSIZE - 1); 514 } 515 bp = nfs_getcacheblk(vp, lbn, bufsize, p); 516 if (!bp) 517 return (EINTR); 518 if (bp->b_wcred == NOCRED) { 519 crhold(cred); 520 bp->b_wcred = cred; 521 } 522 np->n_flag |= NMODIFIED; 523 524 if ((bp->b_blkno * DEV_BSIZE) + bp->b_dirtyend > np->n_size) { 525 bp->b_dirtyend = np->n_size - (bp->b_blkno * DEV_BSIZE); 526 } 527 528 /* 529 * If the new write will leave a contiguous dirty 530 * area, just update the b_dirtyoff and b_dirtyend, 531 * otherwise force a write rpc of the old dirty area. 532 */ 533 if (bp->b_dirtyend > 0 && 534 (on > bp->b_dirtyend || (on + n) < bp->b_dirtyoff)) { 535 bp->b_proc = p; 536 if (VOP_BWRITE(bp) == EINTR) 537 return (EINTR); 538 goto again; 539 } 540 541 /* 542 * Check for valid write lease and get one as required. 543 * In case getblk() and/or bwrite() delayed us. 544 */ 545 if ((nmp->nm_flag & NFSMNT_NQNFS) && 546 NQNFS_CKINVALID(vp, np, ND_WRITE)) { 547 do { 548 error = nqnfs_getlease(vp, ND_WRITE, cred, p); 549 } while (error == NQNFS_EXPIRED); 550 if (error) { 551 brelse(bp); 552 return (error); 553 } 554 if (np->n_lrev != np->n_brev || 555 (np->n_flag & NQNFSNONCACHE)) { 556 brelse(bp); 557 error = nfs_vinvalbuf(vp, V_SAVE, cred, p, 1); 558 if (error) 559 return (error); 560 np->n_brev = np->n_lrev; 561 goto again; 562 } 563 } 564 error = uiomove((char *)bp->b_data + on, n, uio); 565 if (error) { 566 bp->b_flags |= B_ERROR; 567 brelse(bp); 568 return (error); 569 } 570 if (bp->b_dirtyend > 0) { 571 bp->b_dirtyoff = min(on, bp->b_dirtyoff); 572 bp->b_dirtyend = max((on + n), bp->b_dirtyend); 573 } else { 574 bp->b_dirtyoff = on; 575 bp->b_dirtyend = on + n; 576 } 577 if (bp->b_validend == 0 || bp->b_validend < bp->b_dirtyoff || 578 bp->b_validoff > bp->b_dirtyend) { 579 bp->b_validoff = bp->b_dirtyoff; 580 bp->b_validend = bp->b_dirtyend; 581 } else { 582 bp->b_validoff = min(bp->b_validoff, bp->b_dirtyoff); 583 bp->b_validend = max(bp->b_validend, bp->b_dirtyend); 584 } 585 /* 586 * If the lease is non-cachable or IO_SYNC do bwrite(). 587 */ 588 if ((np->n_flag & NQNFSNONCACHE) || (ioflag & IO_SYNC)) { 589 bp->b_proc = p; 590 error = VOP_BWRITE(bp); 591 if (error) 592 return (error); 593 if (np->n_flag & NQNFSNONCACHE) { 594 error = nfs_vinvalbuf(vp, V_SAVE, cred, p, 1); 595 if (error) 596 return (error); 597 } 598 } else if ((n + on) == biosize && 599 (nmp->nm_flag & NFSMNT_NQNFS) == 0) { 600 bp->b_proc = (struct proc *)0; 601 bp->b_flags |= B_ASYNC; 602 (void)nfs_writebp(bp, 0); 603 } else 604 bdwrite(bp); 605 } while (uio->uio_resid > 0 && n > 0); 606 return (0); 607} 608 609/* 610 * Get an nfs cache block. 611 * Allocate a new one if the block isn't currently in the cache 612 * and return the block marked busy. If the calling process is 613 * interrupted by a signal for an interruptible mount point, return 614 * NULL. 615 */ 616static struct buf * 617nfs_getcacheblk(vp, bn, size, p) 618 struct vnode *vp; 619 daddr_t bn; 620 int size; 621 struct proc *p; 622{ 623 register struct buf *bp; 624 struct nfsmount *nmp = VFSTONFS(vp->v_mount); 625 int biosize = vp->v_mount->mnt_stat.f_iosize; 626 627 if (nmp->nm_flag & NFSMNT_INT) { 628 bp = getblk(vp, bn, size, PCATCH, 0); 629 while (bp == (struct buf *)0) { 630 if (nfs_sigintr(nmp, (struct nfsreq *)0, p)) 631 return ((struct buf *)0); 632 bp = getblk(vp, bn, size, 0, 2 * hz); 633 } 634 } else 635 bp = getblk(vp, bn, size, 0, 0); 636 637 if( vp->v_type == VREG) 638 bp->b_blkno = (bn * biosize) / DEV_BSIZE; 639 640 return (bp); 641} 642 643/* 644 * Flush and invalidate all dirty buffers. If another process is already 645 * doing the flush, just wait for completion. 646 */ 647int 648nfs_vinvalbuf(vp, flags, cred, p, intrflg) 649 struct vnode *vp; 650 int flags; 651 struct ucred *cred; 652 struct proc *p; 653 int intrflg; 654{ 655 register struct nfsnode *np = VTONFS(vp); 656 struct nfsmount *nmp = VFSTONFS(vp->v_mount); 657 int error = 0, slpflag, slptimeo; 658 659 if ((nmp->nm_flag & NFSMNT_INT) == 0) 660 intrflg = 0; 661 if (intrflg) { 662 slpflag = PCATCH; 663 slptimeo = 2 * hz; 664 } else { 665 slpflag = 0; 666 slptimeo = 0; 667 } 668 /* 669 * First wait for any other process doing a flush to complete. 670 */ 671 while (np->n_flag & NFLUSHINPROG) { 672 np->n_flag |= NFLUSHWANT; 673 error = tsleep((caddr_t)&np->n_flag, PRIBIO + 2, "nfsvinval", 674 slptimeo); 675 if (error && intrflg && nfs_sigintr(nmp, (struct nfsreq *)0, p)) 676 return (EINTR); 677 } 678 679 /* 680 * Now, flush as required. 681 */ 682 np->n_flag |= NFLUSHINPROG; 683 error = vinvalbuf(vp, flags, cred, p, slpflag, 0); 684 while (error) { 685 if (intrflg && nfs_sigintr(nmp, (struct nfsreq *)0, p)) { 686 np->n_flag &= ~NFLUSHINPROG; 687 if (np->n_flag & NFLUSHWANT) { 688 np->n_flag &= ~NFLUSHWANT; 689 wakeup((caddr_t)&np->n_flag); 690 } 691 return (EINTR); 692 } 693 error = vinvalbuf(vp, flags, cred, p, 0, slptimeo); 694 } 695 np->n_flag &= ~(NMODIFIED | NFLUSHINPROG); 696 if (np->n_flag & NFLUSHWANT) { 697 np->n_flag &= ~NFLUSHWANT; 698 wakeup((caddr_t)&np->n_flag); 699 } 700 return (0); 701} 702 703/* 704 * Initiate asynchronous I/O. Return an error if no nfsiods are available. 705 * This is mainly to avoid queueing async I/O requests when the nfsiods 706 * are all hung on a dead server. 707 */ 708int 709nfs_asyncio(bp, cred) 710 register struct buf *bp; 711 struct ucred *cred; 712{ 713 register int i; 714 715 if (nfs_numasync == 0) 716 return (EIO); 717 for (i = 0; i < NFS_MAXASYNCDAEMON; i++) 718 if (nfs_iodwant[i]) { 719 if (bp->b_flags & B_READ) { 720 if (bp->b_rcred == NOCRED && cred != NOCRED) { 721 crhold(cred); 722 bp->b_rcred = cred; 723 } 724 } else { 725 bp->b_flags |= B_WRITEINPROG; 726 if (bp->b_wcred == NOCRED && cred != NOCRED) { 727 crhold(cred); 728 bp->b_wcred = cred; 729 } 730 } 731 732 TAILQ_INSERT_TAIL(&nfs_bufq, bp, b_freelist); 733 nfs_iodwant[i] = (struct proc *)0; 734 wakeup((caddr_t)&nfs_iodwant[i]); 735 return (0); 736 } 737 738 /* 739 * If it is a read or a write already marked B_WRITEINPROG or B_NOCACHE 740 * return EIO so the process will call nfs_doio() and do it 741 * synchronously. 742 */ 743 if (bp->b_flags & (B_READ | B_WRITEINPROG | B_NOCACHE)) 744 return (EIO); 745 746 /* 747 * Just turn the async write into a delayed write, instead of 748 * doing in synchronously. Hopefully, at least one of the nfsiods 749 * is currently doing a write for this file and will pick up the 750 * delayed writes before going back to sleep. 751 */ 752 bp->b_flags |= B_DELWRI; 753 reassignbuf(bp, bp->b_vp); 754 biodone(bp); 755 return (0); 756} 757 758/* 759 * Do an I/O operation to/from a cache block. This may be called 760 * synchronously or from an nfsiod. 761 */ 762int 763nfs_doio(bp, cr, p) 764 register struct buf *bp; 765 struct ucred *cr; 766 struct proc *p; 767{ 768 register struct uio *uiop; 769 register struct vnode *vp; 770 struct nfsnode *np; 771 struct nfsmount *nmp; 772 int error = 0, diff, len, iomode, must_commit = 0; 773 struct uio uio; 774 struct iovec io; 775 776 vp = bp->b_vp; 777 np = VTONFS(vp); 778 nmp = VFSTONFS(vp->v_mount); 779 uiop = &uio; 780 uiop->uio_iov = &io; 781 uiop->uio_iovcnt = 1; 782 uiop->uio_segflg = UIO_SYSSPACE; 783 uiop->uio_procp = p; 784 785 /* 786 * Historically, paging was done with physio, but no more. 787 */ 788 if (bp->b_flags & B_PHYS) { 789 /* 790 * ...though reading /dev/drum still gets us here. 791 */ 792 io.iov_len = uiop->uio_resid = bp->b_bcount; 793 /* mapping was done by vmapbuf() */ 794 io.iov_base = bp->b_data; 795 uiop->uio_offset = ((off_t)bp->b_blkno) * DEV_BSIZE; 796 if (bp->b_flags & B_READ) { 797 uiop->uio_rw = UIO_READ; 798 nfsstats.read_physios++; 799 error = nfs_readrpc(vp, uiop, cr); 800 } else { 801 int com; 802 803 iomode = NFSV3WRITE_DATASYNC; 804 uiop->uio_rw = UIO_WRITE; 805 nfsstats.write_physios++; 806 error = nfs_writerpc(vp, uiop, cr, &iomode, &com); 807 } 808 if (error) { 809 bp->b_flags |= B_ERROR; 810 bp->b_error = error; 811 } 812 } else if (bp->b_flags & B_READ) { 813 io.iov_len = uiop->uio_resid = bp->b_bcount; 814 io.iov_base = bp->b_data; 815 uiop->uio_rw = UIO_READ; 816 switch (vp->v_type) { 817 case VREG: 818 uiop->uio_offset = ((off_t)bp->b_blkno) * DEV_BSIZE; 819 nfsstats.read_bios++; 820 error = nfs_readrpc(vp, uiop, cr); 821 if (!error) { 822 bp->b_validoff = 0; 823 if (uiop->uio_resid) { 824 /* 825 * If len > 0, there is a hole in the file and 826 * no writes after the hole have been pushed to 827 * the server yet. 828 * Just zero fill the rest of the valid area. 829 */ 830 diff = bp->b_bcount - uiop->uio_resid; 831 len = np->n_size - (((u_quad_t)bp->b_blkno) * DEV_BSIZE 832 + diff); 833 if (len > 0) { 834 len = min(len, uiop->uio_resid); 835 bzero((char *)bp->b_data + diff, len); 836 bp->b_validend = diff + len; 837 } else 838 bp->b_validend = diff; 839 } else 840 bp->b_validend = bp->b_bcount; 841 } 842 if (p && (vp->v_flag & VTEXT) && 843 (((nmp->nm_flag & NFSMNT_NQNFS) && 844 NQNFS_CKINVALID(vp, np, ND_READ) && 845 np->n_lrev != np->n_brev) || 846 (!(nmp->nm_flag & NFSMNT_NQNFS) && 847 np->n_mtime != np->n_vattr.va_mtime.ts_sec))) { 848 uprintf("Process killed due to text file modification\n"); 849 psignal(p, SIGKILL); 850#ifdef __NetBSD__ 851 p->p_holdcnt++; 852#else 853 p->p_flag |= P_NOSWAP; 854#endif 855 } 856 break; 857 case VLNK: 858 uiop->uio_offset = (off_t)0; 859 nfsstats.readlink_bios++; 860 error = nfs_readlinkrpc(vp, uiop, cr); 861 break; 862 case VDIR: 863 nfsstats.readdir_bios++; 864 uiop->uio_offset = ((u_quad_t)bp->b_lblkno) * NFS_DIRBLKSIZ; 865 if (nmp->nm_flag & NFSMNT_RDIRPLUS) { 866 error = nfs_readdirplusrpc(vp, uiop, cr); 867 if (error == NFSERR_NOTSUPP) 868 nmp->nm_flag &= ~NFSMNT_RDIRPLUS; 869 } 870 if ((nmp->nm_flag & NFSMNT_RDIRPLUS) == 0) 871 error = nfs_readdirrpc(vp, uiop, cr); 872 break; 873 default: 874 printf("nfs_doio: type %x unexpected\n",vp->v_type); 875 break; 876 }; 877 if (error) { 878 bp->b_flags |= B_ERROR; 879 bp->b_error = error; 880 } 881 } else { 882 if (((bp->b_blkno * DEV_BSIZE) + bp->b_dirtyend) > np->n_size) 883 bp->b_dirtyend = np->n_size - (bp->b_blkno * DEV_BSIZE); 884 885 if (bp->b_dirtyend > bp->b_dirtyoff) { 886 io.iov_len = uiop->uio_resid = bp->b_dirtyend 887 - bp->b_dirtyoff; 888 uiop->uio_offset = ((off_t)bp->b_blkno) * DEV_BSIZE 889 + bp->b_dirtyoff; 890 io.iov_base = (char *)bp->b_data + bp->b_dirtyoff; 891 uiop->uio_rw = UIO_WRITE; 892 nfsstats.write_bios++; 893 if ((bp->b_flags & (B_ASYNC | B_NEEDCOMMIT | B_NOCACHE)) == B_ASYNC) 894 iomode = NFSV3WRITE_UNSTABLE; 895 else 896 iomode = NFSV3WRITE_FILESYNC; 897 bp->b_flags |= B_WRITEINPROG; 898 error = nfs_writerpc(vp, uiop, cr, &iomode, &must_commit); 899 if (!error && iomode == NFSV3WRITE_UNSTABLE) 900 bp->b_flags |= B_NEEDCOMMIT; 901 else 902 bp->b_flags &= ~B_NEEDCOMMIT; 903 bp->b_flags &= ~B_WRITEINPROG; 904 905 /* 906 * For an interrupted write, the buffer is still valid 907 * and the write hasn't been pushed to the server yet, 908 * so we can't set B_ERROR and report the interruption 909 * by setting B_EINTR. For the B_ASYNC case, B_EINTR 910 * is not relevant, so the rpc attempt is essentially 911 * a noop. For the case of a V3 write rpc not being 912 * committed to stable storage, the block is still 913 * dirty and requires either a commit rpc or another 914 * write rpc with iomode == NFSV3WRITE_FILESYNC before 915 * the block is reused. This is indicated by setting 916 * the B_DELWRI and B_NEEDCOMMIT flags. 917 */ 918 if (error == EINTR 919 || (!error && (bp->b_flags & B_NEEDCOMMIT))) { 920 bp->b_flags &= ~(B_INVAL|B_NOCACHE); 921 bp->b_flags |= B_DELWRI; 922 923 /* 924 * Since for the B_ASYNC case, nfs_bwrite() has reassigned the 925 * buffer to the clean list, we have to reassign it back to the 926 * dirty one. Ugh. 927 */ 928 if (bp->b_flags & B_ASYNC) 929 reassignbuf(bp, vp); 930 else 931 bp->b_flags |= B_EINTR; 932 } else { 933 if (error) { 934 bp->b_flags |= B_ERROR; 935 bp->b_error = np->n_error = error; 936 np->n_flag |= NWRITEERR; 937 } 938 bp->b_dirtyoff = bp->b_dirtyend = 0; 939 } 940 } else { 941 bp->b_resid = 0; 942 biodone(bp); 943 return (0); 944 } 945 } 946 bp->b_resid = uiop->uio_resid; 947 if (must_commit) 948 nfs_clearcommit(vp->v_mount); 949 biodone(bp); 950 return (error); 951} 952