nfs_bio.c revision 1.53
1 /* $NetBSD: nfs_bio.c,v 1.53 2000/09/19 23:26:26 bjh21 Exp $ */ 2 3 /* 4 * Copyright (c) 1989, 1993 5 * The Regents of the University of California. All rights reserved. 6 * 7 * This code is derived from software contributed to Berkeley by 8 * Rick Macklem at The University of Guelph. 9 * 10 * Redistribution and use in source and binary forms, with or without 11 * modification, are permitted provided that the following conditions 12 * are met: 13 * 1. Redistributions of source code must retain the above copyright 14 * notice, this list of conditions and the following disclaimer. 15 * 2. Redistributions in binary form must reproduce the above copyright 16 * notice, this list of conditions and the following disclaimer in the 17 * documentation and/or other materials provided with the distribution. 18 * 3. All advertising materials mentioning features or use of this software 19 * must display the following acknowledgement: 20 * This product includes software developed by the University of 21 * California, Berkeley and its contributors. 22 * 4. Neither the name of the University nor the names of its contributors 23 * may be used to endorse or promote products derived from this software 24 * without specific prior written permission. 25 * 26 * THIS SOFTWARE IS PROVIDED BY THE REGENTS AND CONTRIBUTORS ``AS IS'' AND 27 * ANY EXPRESS OR IMPLIED WARRANTIES, INCLUDING, BUT NOT LIMITED TO, THE 28 * IMPLIED WARRANTIES OF MERCHANTABILITY AND FITNESS FOR A PARTICULAR PURPOSE 29 * ARE DISCLAIMED. IN NO EVENT SHALL THE REGENTS OR CONTRIBUTORS BE LIABLE 30 * FOR ANY DIRECT, INDIRECT, INCIDENTAL, SPECIAL, EXEMPLARY, OR CONSEQUENTIAL 31 * DAMAGES (INCLUDING, BUT NOT LIMITED TO, PROCUREMENT OF SUBSTITUTE GOODS 32 * OR SERVICES; LOSS OF USE, DATA, OR PROFITS; OR BUSINESS INTERRUPTION) 33 * HOWEVER CAUSED AND ON ANY THEORY OF LIABILITY, WHETHER IN CONTRACT, STRICT 34 * LIABILITY, OR TORT (INCLUDING NEGLIGENCE OR OTHERWISE) ARISING IN ANY WAY 35 * OUT OF THE USE OF THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF 36 * SUCH DAMAGE. 37 * 38 * @(#)nfs_bio.c 8.9 (Berkeley) 3/30/95 39 */ 40 41 #include "opt_nfs.h" 42 43 #include <sys/param.h> 44 #include <sys/systm.h> 45 #include <sys/resourcevar.h> 46 #include <sys/signalvar.h> 47 #include <sys/proc.h> 48 #include <sys/buf.h> 49 #include <sys/vnode.h> 50 #include <sys/trace.h> 51 #include <sys/mount.h> 52 #include <sys/kernel.h> 53 #include <sys/namei.h> 54 #include <sys/dirent.h> 55 56 #include <uvm/uvm_extern.h> 57 58 #include <nfs/rpcv2.h> 59 #include <nfs/nfsproto.h> 60 #include <nfs/nfs.h> 61 #include <nfs/nfsmount.h> 62 #include <nfs/nqnfs.h> 63 #include <nfs/nfsnode.h> 64 #include <nfs/nfs_var.h> 65 66 extern int nfs_numasync; 67 extern struct nfsstats nfsstats; 68 69 /* 70 * Vnode op for read using bio 71 * Any similarity to readip() is purely coincidental 72 */ 73 int 74 nfs_bioread(vp, uio, ioflag, cred, cflag) 75 struct vnode *vp; 76 struct uio *uio; 77 int ioflag, cflag; 78 struct ucred *cred; 79 { 80 struct nfsnode *np = VTONFS(vp); 81 int biosize, diff; 82 struct buf *bp = NULL, *rabp; 83 struct vattr vattr; 84 struct proc *p; 85 struct nfsmount *nmp = VFSTONFS(vp->v_mount); 86 struct nfsdircache *ndp = NULL, *nndp = NULL; 87 daddr_t lbn, bn, rabn; 88 caddr_t baddr, ep, edp; 89 int got_buf = 0, nra, error = 0, n = 0, on = 0, not_readin, en, enn; 90 int enough = 0; 91 struct dirent *dp, *pdp; 92 off_t curoff = 0, offdiff; 93 94 #ifdef DIAGNOSTIC 95 if (uio->uio_rw != UIO_READ) 96 panic("nfs_read mode"); 97 #endif 98 if (uio->uio_resid == 0) 99 return (0); 100 if (vp->v_type != VDIR && uio->uio_offset < 0) 101 return (EINVAL); 102 p = uio->uio_procp; 103 #ifndef NFS_V2_ONLY 104 if ((nmp->nm_flag & NFSMNT_NFSV3) && 105 !(nmp->nm_iflag & NFSMNT_GOTFSINFO)) 106 (void)nfs_fsinfo(nmp, vp, cred, p); 107 #endif 108 if (vp->v_type != VDIR && 109 (uio->uio_offset + uio->uio_resid) > nmp->nm_maxfilesize) 110 return (EFBIG); 111 biosize = nmp->nm_rsize; 112 /* 113 * For nfs, cache consistency can only be maintained approximately. 114 * Although RFC1094 does not specify the criteria, the following is 115 * believed to be compatible with the reference port. 116 * For nqnfs, full cache consistency is maintained within the loop. 117 * For nfs: 118 * If the file's modify time on the server has changed since the 119 * last read rpc or you have written to the file, 120 * you may have lost data cache consistency with the 121 * server, so flush all of the file's data out of the cache. 122 * Then force a getattr rpc to ensure that you have up to date 123 * attributes. 124 * NB: This implies that cache data can be read when up to 125 * NFS_ATTRTIMEO seconds out of date. If you find that you need current 126 * attributes this could be forced by setting n_attrstamp to 0 before 127 * the VOP_GETATTR() call. 128 */ 129 if ((nmp->nm_flag & NFSMNT_NQNFS) == 0 && vp->v_type != VLNK) { 130 if (np->n_flag & NMODIFIED) { 131 if (vp->v_type != VREG) { 132 if (vp->v_type != VDIR) 133 panic("nfs: bioread, not dir"); 134 nfs_invaldircache(vp, 0); 135 np->n_direofoffset = 0; 136 error = nfs_vinvalbuf(vp, V_SAVE, cred, p, 1); 137 if (error) 138 return (error); 139 } 140 np->n_attrstamp = 0; 141 error = VOP_GETATTR(vp, &vattr, cred, p); 142 if (error) 143 return (error); 144 np->n_mtime = vattr.va_mtime.tv_sec; 145 } else { 146 error = VOP_GETATTR(vp, &vattr, cred, p); 147 if (error) 148 return (error); 149 if (np->n_mtime != vattr.va_mtime.tv_sec) { 150 if (vp->v_type == VDIR) { 151 nfs_invaldircache(vp, 0); 152 np->n_direofoffset = 0; 153 } 154 error = nfs_vinvalbuf(vp, V_SAVE, cred, p, 1); 155 if (error) 156 return (error); 157 np->n_mtime = vattr.va_mtime.tv_sec; 158 } 159 } 160 } 161 do { 162 163 #ifndef NFS_V2_ONLY 164 /* 165 * Get a valid lease. If cached data is stale, flush it. 166 */ 167 if (nmp->nm_flag & NFSMNT_NQNFS) { 168 if (NQNFS_CKINVALID(vp, np, ND_READ)) { 169 do { 170 error = nqnfs_getlease(vp, ND_READ, cred, p); 171 } while (error == NQNFS_EXPIRED); 172 if (error) 173 return (error); 174 if (np->n_lrev != np->n_brev || 175 (np->n_flag & NQNFSNONCACHE) || 176 ((np->n_flag & NMODIFIED) && vp->v_type == VDIR)) { 177 if (vp->v_type == VDIR) { 178 nfs_invaldircache(vp, 0); 179 np->n_direofoffset = 0; 180 } 181 error = nfs_vinvalbuf(vp, V_SAVE, cred, p, 1); 182 if (error) 183 return (error); 184 np->n_brev = np->n_lrev; 185 } 186 } else if (vp->v_type == VDIR && (np->n_flag & NMODIFIED)) { 187 nfs_invaldircache(vp, 0); 188 error = nfs_vinvalbuf(vp, V_SAVE, cred, p, 1); 189 np->n_direofoffset = 0; 190 if (error) 191 return (error); 192 } 193 } 194 #endif 195 /* 196 * Don't cache symlinks. 197 */ 198 if (np->n_flag & NQNFSNONCACHE 199 || ((vp->v_flag & VROOT) && vp->v_type == VLNK)) { 200 switch (vp->v_type) { 201 case VREG: 202 return (nfs_readrpc(vp, uio, cred)); 203 case VLNK: 204 return (nfs_readlinkrpc(vp, uio, cred)); 205 case VDIR: 206 break; 207 default: 208 printf(" NQNFSNONCACHE: type %x unexpected\n", 209 vp->v_type); 210 }; 211 } 212 baddr = (caddr_t)0; 213 switch (vp->v_type) { 214 case VREG: 215 nfsstats.biocache_reads++; 216 lbn = uio->uio_offset / biosize; 217 on = uio->uio_offset & (biosize - 1); 218 bn = lbn * (biosize / DEV_BSIZE); 219 not_readin = 1; 220 221 /* 222 * Start the read ahead(s), as required. 223 */ 224 if (nfs_numasync > 0 && nmp->nm_readahead > 0 && 225 lbn - 1 == vp->v_lastr) { 226 for (nra = 0; nra < nmp->nm_readahead && 227 (lbn + 1 + nra) * biosize < np->n_size; nra++) { 228 rabn = (lbn + 1 + nra) * (biosize / DEV_BSIZE); 229 if (!incore(vp, rabn)) { 230 rabp = nfs_getcacheblk(vp, rabn, biosize, p); 231 if (!rabp) 232 return (EINTR); 233 if ((rabp->b_flags & (B_DELWRI | B_DONE)) == 0) { 234 rabp->b_flags |= (B_READ | B_ASYNC); 235 if (nfs_asyncio(rabp, cred)) { 236 rabp->b_flags |= B_INVAL; 237 brelse(rabp); 238 } 239 } else 240 brelse(rabp); 241 } 242 } 243 } 244 245 /* 246 * If the block is in the cache and has the required data 247 * in a valid region, just copy it out. 248 * Otherwise, get the block and write back/read in, 249 * as required. 250 */ 251 if ((bp = incore(vp, bn)) && 252 (bp->b_flags & (B_BUSY | B_WRITEINPROG)) == 253 (B_BUSY | B_WRITEINPROG)) 254 got_buf = 0; 255 else { 256 again: 257 bp = nfs_getcacheblk(vp, bn, biosize, p); 258 if (!bp) 259 return (EINTR); 260 got_buf = 1; 261 if ((bp->b_flags & (B_DONE | B_DELWRI)) == 0) { 262 bp->b_flags |= B_READ; 263 not_readin = 0; 264 error = nfs_doio(bp, cred, p); 265 if (error) { 266 brelse(bp); 267 return (error); 268 } 269 } 270 } 271 n = min((unsigned)(biosize - on), uio->uio_resid); 272 offdiff = np->n_size - uio->uio_offset; 273 if (offdiff < (off_t)n) 274 n = (int)offdiff; 275 if (not_readin && n > 0) { 276 if (on < bp->b_validoff || (on + n) > bp->b_validend) { 277 if (!got_buf) { 278 bp = nfs_getcacheblk(vp, bn, biosize, p); 279 if (!bp) 280 return (EINTR); 281 got_buf = 1; 282 } 283 bp->b_flags |= B_INVAFTERWRITE; 284 if (bp->b_dirtyend > 0) { 285 if ((bp->b_flags & B_DELWRI) == 0) 286 panic("nfsbioread"); 287 if (VOP_BWRITE(bp) == EINTR) 288 return (EINTR); 289 } else 290 brelse(bp); 291 goto again; 292 } 293 } 294 vp->v_lastr = lbn; 295 diff = (on >= bp->b_validend) ? 0 : (bp->b_validend - on); 296 if (diff < n) 297 n = diff; 298 break; 299 case VLNK: 300 nfsstats.biocache_readlinks++; 301 bp = nfs_getcacheblk(vp, (daddr_t)0, NFS_MAXPATHLEN, p); 302 if (!bp) 303 return (EINTR); 304 if ((bp->b_flags & B_DONE) == 0) { 305 bp->b_flags |= B_READ; 306 error = nfs_doio(bp, cred, p); 307 if (error) { 308 brelse(bp); 309 return (error); 310 } 311 } 312 n = min(uio->uio_resid, NFS_MAXPATHLEN - bp->b_resid); 313 got_buf = 1; 314 on = 0; 315 break; 316 case VDIR: 317 diragain: 318 nfsstats.biocache_readdirs++; 319 ndp = nfs_searchdircache(vp, uio->uio_offset, 320 (nmp->nm_flag & NFSMNT_XLATECOOKIE), 0); 321 if (!ndp) { 322 /* 323 * We've been handed a cookie that is not 324 * in the cache. If we're not translating 325 * 32 <-> 64, it may be a value that was 326 * flushed out of the cache because it grew 327 * too big. Let the server judge if it's 328 * valid or not. In the translation case, 329 * we have no way of validating this value, 330 * so punt. 331 */ 332 if (nmp->nm_flag & NFSMNT_XLATECOOKIE) 333 return (EINVAL); 334 ndp = nfs_enterdircache(vp, uio->uio_offset, 335 uio->uio_offset, 0, 0); 336 } 337 338 if (uio->uio_offset != 0 && 339 ndp->dc_cookie == np->n_direofoffset) { 340 nfsstats.direofcache_hits++; 341 return (0); 342 } 343 344 bp = nfs_getcacheblk(vp, ndp->dc_blkno, NFS_DIRBLKSIZ, p); 345 if (!bp) 346 return (EINTR); 347 if ((bp->b_flags & B_DONE) == 0) { 348 bp->b_flags |= B_READ; 349 bp->b_dcookie = ndp->dc_blkcookie; 350 error = nfs_doio(bp, cred, p); 351 if (error) { 352 /* 353 * Yuck! The directory has been modified on the 354 * server. Punt and let the userland code 355 * deal with it. 356 */ 357 brelse(bp); 358 if (error == NFSERR_BAD_COOKIE) { 359 nfs_invaldircache(vp, 0); 360 nfs_vinvalbuf(vp, 0, cred, p, 1); 361 error = EINVAL; 362 } 363 return (error); 364 } 365 } 366 367 /* 368 * Just return if we hit EOF right away with this 369 * block. Always check here, because direofoffset 370 * may have been set by an nfsiod since the last 371 * check. 372 */ 373 if (np->n_direofoffset != 0 && 374 ndp->dc_blkcookie == np->n_direofoffset) { 375 brelse(bp); 376 return (0); 377 } 378 379 /* 380 * Find the entry we were looking for in the block. 381 */ 382 383 en = ndp->dc_entry; 384 385 pdp = dp = (struct dirent *)bp->b_data; 386 edp = bp->b_data + bp->b_validend; 387 enn = 0; 388 while (enn < en && (caddr_t)dp < edp) { 389 pdp = dp; 390 dp = (struct dirent *)((caddr_t)dp + dp->d_reclen); 391 enn++; 392 } 393 394 /* 395 * If the entry number was bigger than the number of 396 * entries in the block, or the cookie of the previous 397 * entry doesn't match, the directory cache is 398 * stale. Flush it and try again (i.e. go to 399 * the server). 400 */ 401 if ((caddr_t)dp >= edp || (caddr_t)dp + dp->d_reclen > edp || 402 (en > 0 && NFS_GETCOOKIE(pdp) != ndp->dc_cookie)) { 403 #ifdef DEBUG 404 printf("invalid cache: %p %p %p off %lx %lx\n", 405 pdp, dp, edp, 406 (unsigned long)uio->uio_offset, 407 (unsigned long)NFS_GETCOOKIE(pdp)); 408 #endif 409 brelse(bp); 410 nfs_invaldircache(vp, 0); 411 nfs_vinvalbuf(vp, 0, cred, p, 0); 412 goto diragain; 413 } 414 415 on = (caddr_t)dp - bp->b_data; 416 417 /* 418 * Cache all entries that may be exported to the 419 * user, as they may be thrown back at us. The 420 * NFSBIO_CACHECOOKIES flag indicates that all 421 * entries are being 'exported', so cache them all. 422 */ 423 424 if (en == 0 && pdp == dp) { 425 dp = (struct dirent *) 426 ((caddr_t)dp + dp->d_reclen); 427 enn++; 428 } 429 430 if (uio->uio_resid < (bp->b_validend - on)) { 431 n = uio->uio_resid; 432 enough = 1; 433 } else 434 n = bp->b_validend - on; 435 436 ep = bp->b_data + on + n; 437 438 /* 439 * Find last complete entry to copy, caching entries 440 * (if requested) as we go. 441 */ 442 443 while ((caddr_t)dp < ep && (caddr_t)dp + dp->d_reclen <= ep) { 444 if (cflag & NFSBIO_CACHECOOKIES) { 445 nndp = nfs_enterdircache(vp, NFS_GETCOOKIE(pdp), 446 ndp->dc_blkcookie, enn, bp->b_lblkno); 447 if (nmp->nm_flag & NFSMNT_XLATECOOKIE) { 448 NFS_STASHCOOKIE32(pdp, 449 nndp->dc_cookie32); 450 } 451 } 452 pdp = dp; 453 dp = (struct dirent *)((caddr_t)dp + dp->d_reclen); 454 enn++; 455 } 456 457 /* 458 * If the last requested entry was not the last in the 459 * buffer (happens if NFS_DIRFRAGSIZ < NFS_DIRBLKSIZ), 460 * cache the cookie of the last requested one, and 461 * set of the offset to it. 462 */ 463 464 if ((on + n) < bp->b_validend) { 465 curoff = NFS_GETCOOKIE(pdp); 466 nndp = nfs_enterdircache(vp, curoff, ndp->dc_blkcookie, 467 enn, bp->b_lblkno); 468 if (nmp->nm_flag & NFSMNT_XLATECOOKIE) { 469 NFS_STASHCOOKIE32(pdp, nndp->dc_cookie32); 470 curoff = nndp->dc_cookie32; 471 } 472 } else 473 curoff = bp->b_dcookie; 474 475 /* 476 * Always cache the entry for the next block, 477 * so that readaheads can use it. 478 */ 479 nndp = nfs_enterdircache(vp, bp->b_dcookie, bp->b_dcookie, 0,0); 480 if (nmp->nm_flag & NFSMNT_XLATECOOKIE) { 481 if (curoff == bp->b_dcookie) { 482 NFS_STASHCOOKIE32(pdp, nndp->dc_cookie32); 483 curoff = nndp->dc_cookie32; 484 } 485 } 486 487 n = ((caddr_t)pdp + pdp->d_reclen) - (bp->b_data + on); 488 489 /* 490 * If not eof and read aheads are enabled, start one. 491 * (You need the current block first, so that you have the 492 * directory offset cookie of the next block.) 493 */ 494 if (nfs_numasync > 0 && nmp->nm_readahead > 0 && 495 np->n_direofoffset == 0 && !(np->n_flag & NQNFSNONCACHE)) { 496 rabp = nfs_getcacheblk(vp, nndp->dc_blkno, 497 NFS_DIRBLKSIZ, p); 498 if (rabp) { 499 if ((rabp->b_flags & (B_DONE | B_DELWRI)) == 0) { 500 rabp->b_dcookie = nndp->dc_cookie; 501 rabp->b_flags |= (B_READ | B_ASYNC); 502 if (nfs_asyncio(rabp, cred)) { 503 rabp->b_flags |= B_INVAL; 504 brelse(rabp); 505 } 506 } else 507 brelse(rabp); 508 } 509 } 510 got_buf = 1; 511 break; 512 default: 513 printf(" nfsbioread: type %x unexpected\n",vp->v_type); 514 break; 515 }; 516 517 if (n > 0) { 518 if (!baddr) 519 baddr = bp->b_data; 520 error = uiomove(baddr + on, (int)n, uio); 521 } 522 switch (vp->v_type) { 523 case VREG: 524 break; 525 case VLNK: 526 n = 0; 527 break; 528 case VDIR: 529 if (np->n_flag & NQNFSNONCACHE) 530 bp->b_flags |= B_INVAL; 531 uio->uio_offset = curoff; 532 if (enough) 533 n = 0; 534 break; 535 default: 536 printf(" nfsbioread: type %x unexpected\n",vp->v_type); 537 } 538 if (got_buf) 539 brelse(bp); 540 } while (error == 0 && uio->uio_resid > 0 && n > 0); 541 return (error); 542 } 543 544 /* 545 * Vnode op for write using bio 546 */ 547 int 548 nfs_write(v) 549 void *v; 550 { 551 struct vop_write_args /* { 552 struct vnode *a_vp; 553 struct uio *a_uio; 554 int a_ioflag; 555 struct ucred *a_cred; 556 } */ *ap = v; 557 int biosize; 558 struct uio *uio = ap->a_uio; 559 struct proc *p = uio->uio_procp; 560 struct vnode *vp = ap->a_vp; 561 struct nfsnode *np = VTONFS(vp); 562 struct ucred *cred = ap->a_cred; 563 int ioflag = ap->a_ioflag; 564 struct buf *bp; 565 struct vattr vattr; 566 struct nfsmount *nmp = VFSTONFS(vp->v_mount); 567 daddr_t lbn, bn; 568 int n, on, error = 0, iomode, must_commit; 569 570 #ifdef DIAGNOSTIC 571 if (uio->uio_rw != UIO_WRITE) 572 panic("nfs_write mode"); 573 if (uio->uio_segflg == UIO_USERSPACE && uio->uio_procp != curproc) 574 panic("nfs_write proc"); 575 #endif 576 if (vp->v_type != VREG) 577 return (EIO); 578 if (np->n_flag & NWRITEERR) { 579 np->n_flag &= ~NWRITEERR; 580 return (np->n_error); 581 } 582 #ifndef NFS_V2_ONLY 583 if ((nmp->nm_flag & NFSMNT_NFSV3) && 584 !(nmp->nm_iflag & NFSMNT_GOTFSINFO)) 585 (void)nfs_fsinfo(nmp, vp, cred, p); 586 #endif 587 if (ioflag & (IO_APPEND | IO_SYNC)) { 588 if (np->n_flag & NMODIFIED) { 589 np->n_attrstamp = 0; 590 error = nfs_vinvalbuf(vp, V_SAVE, cred, p, 1); 591 if (error) 592 return (error); 593 } 594 if (ioflag & IO_APPEND) { 595 np->n_attrstamp = 0; 596 error = VOP_GETATTR(vp, &vattr, cred, p); 597 if (error) 598 return (error); 599 uio->uio_offset = np->n_size; 600 } 601 } 602 if (uio->uio_offset < 0) 603 return (EINVAL); 604 if ((uio->uio_offset + uio->uio_resid) > nmp->nm_maxfilesize) 605 return (EFBIG); 606 if (uio->uio_resid == 0) 607 return (0); 608 /* 609 * Maybe this should be above the vnode op call, but so long as 610 * file servers have no limits, i don't think it matters 611 */ 612 if (p && uio->uio_offset + uio->uio_resid > 613 p->p_rlimit[RLIMIT_FSIZE].rlim_cur) { 614 psignal(p, SIGXFSZ); 615 return (EFBIG); 616 } 617 /* 618 * I use nm_rsize, not nm_wsize so that all buffer cache blocks 619 * will be the same size within a filesystem. nfs_writerpc will 620 * still use nm_wsize when sizing the rpc's. 621 */ 622 biosize = nmp->nm_rsize; 623 do { 624 625 /* 626 * XXX make sure we aren't cached in the VM page cache 627 */ 628 (void)uvm_vnp_uncache(vp); 629 630 #ifndef NFS_V2_ONLY 631 /* 632 * Check for a valid write lease. 633 */ 634 if ((nmp->nm_flag & NFSMNT_NQNFS) && 635 NQNFS_CKINVALID(vp, np, ND_WRITE)) { 636 do { 637 error = nqnfs_getlease(vp, ND_WRITE, cred, p); 638 } while (error == NQNFS_EXPIRED); 639 if (error) 640 return (error); 641 if (np->n_lrev != np->n_brev || 642 (np->n_flag & NQNFSNONCACHE)) { 643 error = nfs_vinvalbuf(vp, V_SAVE, cred, p, 1); 644 if (error) 645 return (error); 646 np->n_brev = np->n_lrev; 647 } 648 } 649 #endif 650 if ((np->n_flag & NQNFSNONCACHE) && uio->uio_iovcnt == 1) { 651 iomode = NFSV3WRITE_FILESYNC; 652 error = nfs_writerpc(vp, uio, cred, &iomode, &must_commit); 653 if (must_commit) 654 nfs_clearcommit(vp->v_mount); 655 return (error); 656 } 657 nfsstats.biocache_writes++; 658 lbn = uio->uio_offset / biosize; 659 on = uio->uio_offset & (biosize-1); 660 n = min((unsigned)(biosize - on), uio->uio_resid); 661 bn = lbn * (biosize / DEV_BSIZE); 662 again: 663 bp = nfs_getcacheblk(vp, bn, biosize, p); 664 if (!bp) 665 return (EINTR); 666 if (bp->b_wcred == NOCRED) { 667 crhold(cred); 668 bp->b_wcred = cred; 669 } 670 np->n_flag |= NMODIFIED; 671 if (uio->uio_offset + n > np->n_size) { 672 np->n_size = uio->uio_offset + n; 673 uvm_vnp_setsize(vp, np->n_size); 674 } 675 676 /* 677 * If the new write will leave a contiguous dirty 678 * area, just update the b_dirtyoff and b_dirtyend, 679 * otherwise force a write rpc of the old dirty area. 680 */ 681 if (bp->b_dirtyend > 0 && 682 (on > bp->b_dirtyend || (on + n) < bp->b_dirtyoff)) { 683 bp->b_proc = p; 684 if (VOP_BWRITE(bp) == EINTR) 685 return (EINTR); 686 goto again; 687 } 688 689 #ifndef NFS_V2_ONLY 690 /* 691 * Check for valid write lease and get one as required. 692 * In case getblk() and/or bwrite() delayed us. 693 */ 694 if ((nmp->nm_flag & NFSMNT_NQNFS) && 695 NQNFS_CKINVALID(vp, np, ND_WRITE)) { 696 do { 697 error = nqnfs_getlease(vp, ND_WRITE, cred, p); 698 } while (error == NQNFS_EXPIRED); 699 if (error) { 700 brelse(bp); 701 return (error); 702 } 703 if (np->n_lrev != np->n_brev || 704 (np->n_flag & NQNFSNONCACHE)) { 705 brelse(bp); 706 error = nfs_vinvalbuf(vp, V_SAVE, cred, p, 1); 707 if (error) 708 return (error); 709 np->n_brev = np->n_lrev; 710 goto again; 711 } 712 } 713 #endif 714 error = uiomove((char *)bp->b_data + on, n, uio); 715 if (error) { 716 bp->b_flags |= B_ERROR; 717 brelse(bp); 718 return (error); 719 } 720 if (bp->b_dirtyend > 0) { 721 bp->b_dirtyoff = min(on, bp->b_dirtyoff); 722 bp->b_dirtyend = max((on + n), bp->b_dirtyend); 723 } else { 724 bp->b_dirtyoff = on; 725 bp->b_dirtyend = on + n; 726 } 727 if (bp->b_validend == 0 || bp->b_validend < bp->b_dirtyoff || 728 bp->b_validoff > bp->b_dirtyend) { 729 bp->b_validoff = bp->b_dirtyoff; 730 bp->b_validend = bp->b_dirtyend; 731 } else { 732 bp->b_validoff = min(bp->b_validoff, bp->b_dirtyoff); 733 bp->b_validend = max(bp->b_validend, bp->b_dirtyend); 734 } 735 736 /* 737 * Since this block is being modified, it must be written 738 * again and not just committed. 739 */ 740 if (NFS_ISV3(vp)) { 741 lockmgr(&np->n_commitlock, LK_EXCLUSIVE, NULL); 742 if (bp->b_flags & B_NEEDCOMMIT) { 743 bp->b_flags &= ~B_NEEDCOMMIT; 744 nfs_del_tobecommitted_range(vp, bp); 745 } 746 nfs_del_committed_range(vp, bp); 747 lockmgr(&np->n_commitlock, LK_RELEASE, NULL); 748 } 749 750 /* 751 * If the lease is non-cachable or IO_SYNC do bwrite(). 752 */ 753 if ((np->n_flag & NQNFSNONCACHE) || (ioflag & IO_SYNC)) { 754 bp->b_proc = p; 755 error = VOP_BWRITE(bp); 756 if (error) 757 return (error); 758 if (np->n_flag & NQNFSNONCACHE) { 759 error = nfs_vinvalbuf(vp, V_SAVE, cred, p, 1); 760 if (error) 761 return (error); 762 } 763 } else if ((n + on) == biosize && 764 (nmp->nm_flag & NFSMNT_NQNFS) == 0) { 765 bp->b_proc = (struct proc *)0; 766 bawrite(bp); 767 } else { 768 bdwrite(bp); 769 } 770 } while (uio->uio_resid > 0 && n > 0); 771 return (0); 772 } 773 774 /* 775 * Get an nfs cache block. 776 * Allocate a new one if the block isn't currently in the cache 777 * and return the block marked busy. If the calling process is 778 * interrupted by a signal for an interruptible mount point, return 779 * NULL. 780 */ 781 struct buf * 782 nfs_getcacheblk(vp, bn, size, p) 783 struct vnode *vp; 784 daddr_t bn; 785 int size; 786 struct proc *p; 787 { 788 struct buf *bp; 789 struct nfsmount *nmp = VFSTONFS(vp->v_mount); 790 791 if (nmp->nm_flag & NFSMNT_INT) { 792 bp = getblk(vp, bn, size, PCATCH, 0); 793 while (bp == (struct buf *)0) { 794 if (nfs_sigintr(nmp, (struct nfsreq *)0, p)) 795 return ((struct buf *)0); 796 bp = getblk(vp, bn, size, 0, 2 * hz); 797 } 798 } else 799 bp = getblk(vp, bn, size, 0, 0); 800 return (bp); 801 } 802 803 /* 804 * Flush and invalidate all dirty buffers. If another process is already 805 * doing the flush, just wait for completion. 806 */ 807 int 808 nfs_vinvalbuf(vp, flags, cred, p, intrflg) 809 struct vnode *vp; 810 int flags; 811 struct ucred *cred; 812 struct proc *p; 813 int intrflg; 814 { 815 struct nfsnode *np = VTONFS(vp); 816 struct nfsmount *nmp = VFSTONFS(vp->v_mount); 817 int error = 0, slpflag, slptimeo; 818 819 if ((nmp->nm_flag & NFSMNT_INT) == 0) 820 intrflg = 0; 821 if (intrflg) { 822 slpflag = PCATCH; 823 slptimeo = 2 * hz; 824 } else { 825 slpflag = 0; 826 slptimeo = 0; 827 } 828 /* 829 * First wait for any other process doing a flush to complete. 830 */ 831 while (np->n_flag & NFLUSHINPROG) { 832 np->n_flag |= NFLUSHWANT; 833 error = tsleep((caddr_t)&np->n_flag, PRIBIO + 2, "nfsvinval", 834 slptimeo); 835 if (error && intrflg && nfs_sigintr(nmp, (struct nfsreq *)0, p)) 836 return (EINTR); 837 } 838 839 /* 840 * Now, flush as required. 841 */ 842 np->n_flag |= NFLUSHINPROG; 843 error = vinvalbuf(vp, flags, cred, p, slpflag, 0); 844 while (error) { 845 if (intrflg && nfs_sigintr(nmp, (struct nfsreq *)0, p)) { 846 np->n_flag &= ~NFLUSHINPROG; 847 if (np->n_flag & NFLUSHWANT) { 848 np->n_flag &= ~NFLUSHWANT; 849 wakeup((caddr_t)&np->n_flag); 850 } 851 return (EINTR); 852 } 853 error = vinvalbuf(vp, flags, cred, p, 0, slptimeo); 854 } 855 np->n_flag &= ~(NMODIFIED | NFLUSHINPROG); 856 if (np->n_flag & NFLUSHWANT) { 857 np->n_flag &= ~NFLUSHWANT; 858 wakeup((caddr_t)&np->n_flag); 859 } 860 return (0); 861 } 862 863 /* 864 * Initiate asynchronous I/O. Return an error if no nfsiods are available. 865 * This is mainly to avoid queueing async I/O requests when the nfsiods 866 * are all hung on a dead server. 867 */ 868 int 869 nfs_asyncio(bp, cred) 870 struct buf *bp; 871 struct ucred *cred; 872 { 873 int i; 874 struct nfsmount *nmp; 875 int gotiod, slpflag = 0, slptimeo = 0, error; 876 877 if (nfs_numasync == 0) 878 return (EIO); 879 880 881 nmp = VFSTONFS(bp->b_vp->v_mount); 882 again: 883 if (nmp->nm_flag & NFSMNT_INT) 884 slpflag = PCATCH; 885 gotiod = FALSE; 886 887 /* 888 * Find a free iod to process this request. 889 */ 890 891 for (i = 0; i < NFS_MAXASYNCDAEMON; i++) 892 if (nfs_iodwant[i]) { 893 /* 894 * Found one, so wake it up and tell it which 895 * mount to process. 896 */ 897 nfs_iodwant[i] = (struct proc *)0; 898 nfs_iodmount[i] = nmp; 899 nmp->nm_bufqiods++; 900 wakeup((caddr_t)&nfs_iodwant[i]); 901 gotiod = TRUE; 902 break; 903 } 904 /* 905 * If none are free, we may already have an iod working on this mount 906 * point. If so, it will process our request. 907 */ 908 if (!gotiod && nmp->nm_bufqiods > 0) 909 gotiod = TRUE; 910 911 /* 912 * If we have an iod which can process the request, then queue 913 * the buffer. 914 */ 915 if (gotiod) { 916 /* 917 * Ensure that the queue never grows too large. 918 */ 919 while (nmp->nm_bufqlen >= 2*nfs_numasync) { 920 nmp->nm_bufqwant = TRUE; 921 error = tsleep(&nmp->nm_bufq, slpflag | PRIBIO, 922 "nfsaio", slptimeo); 923 if (error) { 924 if (nfs_sigintr(nmp, NULL, bp->b_proc)) 925 return (EINTR); 926 if (slpflag == PCATCH) { 927 slpflag = 0; 928 slptimeo = 2 * hz; 929 } 930 } 931 /* 932 * We might have lost our iod while sleeping, 933 * so check and loop if nescessary. 934 */ 935 if (nmp->nm_bufqiods == 0) 936 goto again; 937 } 938 939 if (bp->b_flags & B_READ) { 940 if (bp->b_rcred == NOCRED && cred != NOCRED) { 941 crhold(cred); 942 bp->b_rcred = cred; 943 } 944 } else { 945 bp->b_flags |= B_WRITEINPROG; 946 if (bp->b_wcred == NOCRED && cred != NOCRED) { 947 crhold(cred); 948 bp->b_wcred = cred; 949 } 950 } 951 952 TAILQ_INSERT_TAIL(&nmp->nm_bufq, bp, b_freelist); 953 nmp->nm_bufqlen++; 954 return (0); 955 } 956 957 /* 958 * All the iods are busy on other mounts, so return EIO to 959 * force the caller to process the i/o synchronously. 960 */ 961 return (EIO); 962 } 963 964 /* 965 * Do an I/O operation to/from a cache block. This may be called 966 * synchronously or from an nfsiod. 967 */ 968 int 969 nfs_doio(bp, cr, p) 970 struct buf *bp; 971 struct ucred *cr; 972 struct proc *p; 973 { 974 struct uio *uiop; 975 struct vnode *vp; 976 struct nfsnode *np; 977 struct nfsmount *nmp; 978 int error = 0, diff, len, iomode, must_commit = 0, s, retv = 0; 979 int pushedrange; 980 unsigned cnt; 981 struct uio uio; 982 struct iovec io; 983 off_t off; 984 985 vp = bp->b_vp; 986 np = VTONFS(vp); 987 nmp = VFSTONFS(vp->v_mount); 988 uiop = &uio; 989 uiop->uio_iov = &io; 990 uiop->uio_iovcnt = 1; 991 uiop->uio_segflg = UIO_SYSSPACE; 992 uiop->uio_procp = p; 993 994 /* 995 * Historically, paging was done with physio, but no more... 996 */ 997 if (bp->b_flags & B_PHYS) { 998 /* 999 * ...though reading /dev/drum still gets us here. 1000 */ 1001 io.iov_len = uiop->uio_resid = bp->b_bcount; 1002 /* mapping was done by vmapbuf() */ 1003 io.iov_base = bp->b_data; 1004 uiop->uio_offset = ((off_t)bp->b_blkno) * DEV_BSIZE; 1005 if (bp->b_flags & B_READ) { 1006 uiop->uio_rw = UIO_READ; 1007 nfsstats.read_physios++; 1008 error = nfs_readrpc(vp, uiop, cr); 1009 } else { 1010 iomode = NFSV3WRITE_DATASYNC; 1011 uiop->uio_rw = UIO_WRITE; 1012 nfsstats.write_physios++; 1013 error = nfs_writerpc(vp, uiop, cr, &iomode, &must_commit); 1014 } 1015 if (error) { 1016 bp->b_flags |= B_ERROR; 1017 bp->b_error = error; 1018 } 1019 } else if (bp->b_flags & B_READ) { 1020 io.iov_len = uiop->uio_resid = bp->b_bcount; 1021 io.iov_base = bp->b_data; 1022 uiop->uio_rw = UIO_READ; 1023 switch (vp->v_type) { 1024 case VREG: 1025 uiop->uio_offset = ((off_t)bp->b_blkno) * DEV_BSIZE; 1026 nfsstats.read_bios++; 1027 error = nfs_readrpc(vp, uiop, cr); 1028 if (!error) { 1029 bp->b_validoff = 0; 1030 if (uiop->uio_resid) { 1031 /* 1032 * If len > 0, there is a hole in the file and 1033 * no writes after the hole have been pushed to 1034 * the server yet. 1035 * Just zero fill the rest of the valid area. 1036 */ 1037 diff = bp->b_bcount - uiop->uio_resid; 1038 len = np->n_size - (((u_quad_t)bp->b_blkno) * DEV_BSIZE 1039 + diff); 1040 if (len > 0) { 1041 len = min(len, uiop->uio_resid); 1042 memset((char *)bp->b_data + diff, 0, len); 1043 bp->b_validend = diff + len; 1044 } else 1045 bp->b_validend = diff; 1046 } else 1047 bp->b_validend = bp->b_bcount; 1048 } 1049 if (p && (vp->v_flag & VTEXT) && 1050 (((nmp->nm_flag & NFSMNT_NQNFS) && 1051 NQNFS_CKINVALID(vp, np, ND_READ) && 1052 np->n_lrev != np->n_brev) || 1053 (!(nmp->nm_flag & NFSMNT_NQNFS) && 1054 np->n_mtime != np->n_vattr->va_mtime.tv_sec))) { 1055 uprintf("Process killed due to text file modification\n"); 1056 psignal(p, SIGKILL); 1057 p->p_holdcnt++; 1058 } 1059 break; 1060 case VLNK: 1061 uiop->uio_offset = (off_t)0; 1062 nfsstats.readlink_bios++; 1063 error = nfs_readlinkrpc(vp, uiop, cr); 1064 break; 1065 case VDIR: 1066 nfsstats.readdir_bios++; 1067 uiop->uio_offset = bp->b_dcookie; 1068 if (nmp->nm_flag & NFSMNT_RDIRPLUS) { 1069 error = nfs_readdirplusrpc(vp, uiop, cr); 1070 if (error == NFSERR_NOTSUPP) 1071 nmp->nm_flag &= ~NFSMNT_RDIRPLUS; 1072 } 1073 if ((nmp->nm_flag & NFSMNT_RDIRPLUS) == 0) 1074 error = nfs_readdirrpc(vp, uiop, cr); 1075 if (!error) { 1076 bp->b_dcookie = uiop->uio_offset; 1077 bp->b_validoff = 0; 1078 bp->b_validend = bp->b_bcount - uiop->uio_resid; 1079 } 1080 break; 1081 default: 1082 printf("nfs_doio: type %x unexpected\n",vp->v_type); 1083 break; 1084 }; 1085 if (error) { 1086 bp->b_flags |= B_ERROR; 1087 bp->b_error = error; 1088 } 1089 } else { 1090 /* 1091 * If B_NEEDCOMMIT is set, a commit rpc may do the trick. If not 1092 * an actual write will have to be scheduled. 1093 */ 1094 if (bp->b_flags & B_NEEDCOMMIT) { 1095 /* 1096 * If the buffer is in the range that we already committed, 1097 * there's nothing to do. 1098 * 1099 * If it's in the range that we need to commit, push the 1100 * whole range at once. Else only push the buffer. In 1101 * both these cases, acquire the commit lock to avoid 1102 * other processes modifying the range. Normally the 1103 * vnode lock should have handled this, but there are 1104 * no proper vnode locks for NFS yet (XXX). 1105 */ 1106 lockmgr(&np->n_commitlock, LK_EXCLUSIVE, NULL); 1107 if (!(bp->b_flags & B_NEEDCOMMIT)) { 1108 lockmgr(&np->n_commitlock, LK_RELEASE, NULL); 1109 goto dowrite; 1110 } 1111 if (!nfs_in_committed_range(vp, bp)) { 1112 if (nfs_in_tobecommitted_range(vp, bp)) { 1113 pushedrange = 1; 1114 off = np->n_pushlo; 1115 /* XXX will be too big if > 2G buffer cache */ 1116 cnt = np->n_pushhi - np->n_pushlo; 1117 } else { 1118 pushedrange = 0; 1119 off = ((u_quad_t)bp->b_blkno) * DEV_BSIZE; 1120 cnt = bp->b_dirtyend; 1121 } 1122 bp->b_flags |= B_WRITEINPROG; 1123 retv = nfs_commit(bp->b_vp, off, cnt, 1124 bp->b_wcred, bp->b_proc); 1125 1126 bp->b_flags &= ~B_WRITEINPROG; 1127 if (retv == 0) { 1128 if (pushedrange) { 1129 nfs_merge_commit_ranges(vp); 1130 } 1131 else 1132 nfs_add_committed_range(vp, bp); 1133 } 1134 } 1135 lockmgr(&np->n_commitlock, LK_RELEASE, NULL); 1136 if (!retv) { 1137 bp->b_resid = bp->b_dirtyoff = bp->b_dirtyend = 0; 1138 bp->b_flags &= ~B_NEEDCOMMIT; 1139 biodone(bp); 1140 return (0); 1141 } else if (retv == NFSERR_STALEWRITEVERF) 1142 nfs_clearcommit(bp->b_vp->v_mount); 1143 } 1144 dowrite: 1145 io.iov_len = uiop->uio_resid = bp->b_dirtyend 1146 - bp->b_dirtyoff; 1147 uiop->uio_offset = ((off_t)bp->b_blkno) * DEV_BSIZE 1148 + bp->b_dirtyoff; 1149 io.iov_base = (char *)bp->b_data + bp->b_dirtyoff; 1150 uiop->uio_rw = UIO_WRITE; 1151 nfsstats.write_bios++; 1152 if ((bp->b_flags & (B_ASYNC | B_NEEDCOMMIT | B_NOCACHE)) == B_ASYNC) 1153 iomode = NFSV3WRITE_UNSTABLE; 1154 else 1155 iomode = NFSV3WRITE_FILESYNC; 1156 1157 bp->b_flags |= B_WRITEINPROG; 1158 error = nfs_writerpc(vp, uiop, cr, &iomode, &must_commit); 1159 s = splbio(); 1160 if (!error && iomode == NFSV3WRITE_UNSTABLE) { 1161 bp->b_flags |= B_NEEDCOMMIT; 1162 lockmgr(&np->n_commitlock, LK_EXCLUSIVE, NULL); 1163 nfs_add_tobecommitted_range(vp, bp); 1164 lockmgr(&np->n_commitlock, LK_RELEASE, NULL); 1165 } else if (!error && bp->b_flags & B_NEEDCOMMIT) { 1166 bp->b_flags &= ~B_NEEDCOMMIT; 1167 lockmgr(&np->n_commitlock, LK_EXCLUSIVE, NULL); 1168 nfs_del_committed_range(vp, bp); 1169 lockmgr(&np->n_commitlock, LK_RELEASE, NULL); 1170 } 1171 /* XXX the use of NOCACHE is a hack */ 1172 bp->b_flags &= ~(B_WRITEINPROG|B_NOCACHE); 1173 1174 /* 1175 * For an interrupted write, the buffer is still valid and the 1176 * write hasn't been pushed to the server yet, so we can't set 1177 * B_ERROR and report the interruption by setting B_EINTR. For 1178 * the B_ASYNC case, B_EINTR is not relevant, so the rpc attempt 1179 * is essentially a noop. 1180 * For the case of a V3 write rpc not being committed to stable 1181 * storage, the block is still dirty and requires either a commit 1182 * rpc or another write rpc with iomode == NFSV3WRITE_FILESYNC 1183 * before the block is reused. This is indicated by setting the 1184 * B_DELWRI and B_NEEDCOMMIT flags. 1185 */ 1186 if (error == EINTR || (!error && (bp->b_flags & B_NEEDCOMMIT))) { 1187 bp->b_flags |= B_DELWRI; 1188 /* 1189 * A B_ASYNC block still needs to be committed, so put 1190 * it back on the dirty list. 1191 */ 1192 if (bp->b_flags & B_ASYNC) 1193 reassignbuf(bp, vp); 1194 else if (error) 1195 bp->b_flags |= B_EINTR; 1196 } else { 1197 if (error) { 1198 bp->b_flags |= B_ERROR; 1199 bp->b_error = np->n_error = error; 1200 np->n_flag |= NWRITEERR; 1201 } 1202 bp->b_dirtyoff = bp->b_dirtyend = 0; 1203 } 1204 splx(s); 1205 } 1206 bp->b_resid = uiop->uio_resid; 1207 if (must_commit) 1208 nfs_clearcommit(vp->v_mount); 1209 biodone(bp); 1210 return (error); 1211 } 1212
Indexes created Tue Oct 14 11:09:46 GMT 2025