1 /* $NetBSD: nfs_bio.c,v 1.49 2000/05/18 08:34:26 pk 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 <sys/param.h> 42 #include <sys/systm.h> 43 #include <sys/resourcevar.h> 44 #include <sys/signalvar.h> 45 #include <sys/proc.h> 46 #include <sys/buf.h> 47 #include <sys/vnode.h> 48 #include <sys/trace.h> 49 #include <sys/mount.h> 50 #include <sys/kernel.h> 51 #include <sys/namei.h> 52 #include <sys/dirent.h> 53 54 #include <vm/vm.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 if ((nmp->nm_flag & NFSMNT_NFSV3) && 104 !(nmp->nm_iflag & NFSMNT_GOTFSINFO)) 105 (void)nfs_fsinfo(nmp, vp, cred, p); 106 if (vp->v_type != VDIR && 107 (uio->uio_offset + uio->uio_resid) > nmp->nm_maxfilesize) 108 return (EFBIG); 109 biosize = nmp->nm_rsize; 110 /* 111 * For nfs, cache consistency can only be maintained approximately. 112 * Although RFC1094 does not specify the criteria, the following is 113 * believed to be compatible with the reference port. 114 * For nqnfs, full cache consistency is maintained within the loop. 115 * For nfs: 116 * If the file's modify time on the server has changed since the 117 * last read rpc or you have written to the file, 118 * you may have lost data cache consistency with the 119 * server, so flush all of the file's data out of the cache. 120 * Then force a getattr rpc to ensure that you have up to date 121 * attributes. 122 * NB: This implies that cache data can be read when up to 123 * NFS_ATTRTIMEO seconds out of date. If you find that you need current 124 * attributes this could be forced by setting n_attrstamp to 0 before 125 * the VOP_GETATTR() call. 126 */ 127 if ((nmp->nm_flag & NFSMNT_NQNFS) == 0 && vp->v_type != VLNK) { 128 if (np->n_flag & NMODIFIED) { 129 if (vp->v_type != VREG) { 130 if (vp->v_type != VDIR) 131 panic("nfs: bioread, not dir"); 132 nfs_invaldircache(vp, 0); 133 np->n_direofoffset = 0; 134 error = nfs_vinvalbuf(vp, V_SAVE, cred, p, 1); 135 if (error) 136 return (error); 137 } 138 np->n_attrstamp = 0; 139 error = VOP_GETATTR(vp, &vattr, cred, p); 140 if (error) 141 return (error); 142 np->n_mtime = vattr.va_mtime.tv_sec; 143 } else { 144 error = VOP_GETATTR(vp, &vattr, cred, p); 145 if (error) 146 return (error); 147 if (np->n_mtime != vattr.va_mtime.tv_sec) { 148 if (vp->v_type == VDIR) { 149 nfs_invaldircache(vp, 0); 150 np->n_direofoffset = 0; 151 } 152 error = nfs_vinvalbuf(vp, V_SAVE, cred, p, 1); 153 if (error) 154 return (error); 155 np->n_mtime = vattr.va_mtime.tv_sec; 156 } 157 } 158 } 159 do { 160 161 /* 162 * Get a valid lease. If cached data is stale, flush it. 163 */ 164 if (nmp->nm_flag & NFSMNT_NQNFS) { 165 if (NQNFS_CKINVALID(vp, np, ND_READ)) { 166 do { 167 error = nqnfs_getlease(vp, ND_READ, cred, p); 168 } while (error == NQNFS_EXPIRED); 169 if (error) 170 return (error); 171 if (np->n_lrev != np->n_brev || 172 (np->n_flag & NQNFSNONCACHE) || 173 ((np->n_flag & NMODIFIED) && vp->v_type == VDIR)) { 174 if (vp->v_type == VDIR) { 175 nfs_invaldircache(vp, 0); 176 np->n_direofoffset = 0; 177 } 178 error = nfs_vinvalbuf(vp, V_SAVE, cred, p, 1); 179 if (error) 180 return (error); 181 np->n_brev = np->n_lrev; 182 } 183 } else if (vp->v_type == VDIR && (np->n_flag & NMODIFIED)) { 184 nfs_invaldircache(vp, 0); 185 error = nfs_vinvalbuf(vp, V_SAVE, cred, p, 1); 186 np->n_direofoffset = 0; 187 if (error) 188 return (error); 189 } 190 } 191 /* 192 * Don't cache symlinks. 193 */ 194 if (np->n_flag & NQNFSNONCACHE 195 || ((vp->v_flag & VROOT) && vp->v_type == VLNK)) { 196 switch (vp->v_type) { 197 case VREG: 198 return (nfs_readrpc(vp, uio, cred)); 199 case VLNK: 200 return (nfs_readlinkrpc(vp, uio, cred)); 201 case VDIR: 202 break; 203 default: 204 printf(" NQNFSNONCACHE: type %x unexpected\n", 205 vp->v_type); 206 }; 207 } 208 baddr = (caddr_t)0; 209 switch (vp->v_type) { 210 case VREG: 211 nfsstats.biocache_reads++; 212 lbn = uio->uio_offset / biosize; 213 on = uio->uio_offset & (biosize - 1); 214 bn = lbn * (biosize / DEV_BSIZE); 215 not_readin = 1; 216 217 /* 218 * Start the read ahead(s), as required. 219 */ 220 if (nfs_numasync > 0 && nmp->nm_readahead > 0 && 221 lbn - 1 == vp->v_lastr) { 222 for (nra = 0; nra < nmp->nm_readahead && 223 (lbn + 1 + nra) * biosize < np->n_size; nra++) { 224 rabn = (lbn + 1 + nra) * (biosize / DEV_BSIZE); 225 if (!incore(vp, rabn)) { 226 rabp = nfs_getcacheblk(vp, rabn, biosize, p); 227 if (!rabp) 228 return (EINTR); 229 if ((rabp->b_flags & (B_DELWRI | B_DONE)) == 0) { 230 rabp->b_flags |= (B_READ | B_ASYNC); 231 if (nfs_asyncio(rabp, cred)) { 232 rabp->b_flags |= B_INVAL; 233 brelse(rabp); 234 } 235 } else 236 brelse(rabp); 237 } 238 } 239 } 240 241 /* 242 * If the block is in the cache and has the required data 243 * in a valid region, just copy it out. 244 * Otherwise, get the block and write back/read in, 245 * as required. 246 */ 247 if ((bp = incore(vp, bn)) && 248 (bp->b_flags & (B_BUSY | B_WRITEINPROG)) == 249 (B_BUSY | B_WRITEINPROG)) 250 got_buf = 0; 251 else { 252 again: 253 bp = nfs_getcacheblk(vp, bn, biosize, p); 254 if (!bp) 255 return (EINTR); 256 got_buf = 1; 257 if ((bp->b_flags & (B_DONE | B_DELWRI)) == 0) { 258 bp->b_flags |= B_READ; 259 not_readin = 0; 260 error = nfs_doio(bp, cred, p); 261 if (error) { 262 brelse(bp); 263 return (error); 264 } 265 } 266 } 267 n = min((unsigned)(biosize - on), uio->uio_resid); 268 offdiff = np->n_size - uio->uio_offset; 269 if (offdiff < (off_t)n) 270 n = (int)offdiff; 271 if (not_readin && n > 0) { 272 if (on < bp->b_validoff || (on + n) > bp->b_validend) { 273 if (!got_buf) { 274 bp = nfs_getcacheblk(vp, bn, biosize, p); 275 if (!bp) 276 return (EINTR); 277 got_buf = 1; 278 } 279 bp->b_flags |= B_INVAFTERWRITE; 280 if (bp->b_dirtyend > 0) { 281 if ((bp->b_flags & B_DELWRI) == 0) 282 panic("nfsbioread"); 283 if (VOP_BWRITE(bp) == EINTR) 284 return (EINTR); 285 } else 286 brelse(bp); 287 goto again; 288 } 289 } 290 vp->v_lastr = lbn; 291 diff = (on >= bp->b_validend) ? 0 : (bp->b_validend - on); 292 if (diff < n) 293 n = diff; 294 break; 295 case VLNK: 296 nfsstats.biocache_readlinks++; 297 bp = nfs_getcacheblk(vp, (daddr_t)0, NFS_MAXPATHLEN, p); 298 if (!bp) 299 return (EINTR); 300 if ((bp->b_flags & B_DONE) == 0) { 301 bp->b_flags |= B_READ; 302 error = nfs_doio(bp, cred, p); 303 if (error) { 304 brelse(bp); 305 return (error); 306 } 307 } 308 n = min(uio->uio_resid, NFS_MAXPATHLEN - bp->b_resid); 309 got_buf = 1; 310 on = 0; 311 break; 312 case VDIR: 313 diragain: 314 nfsstats.biocache_readdirs++; 315 ndp = nfs_searchdircache(vp, uio->uio_offset, 316 (nmp->nm_flag & NFSMNT_XLATECOOKIE), 0); 317 if (!ndp) { 318 /* 319 * We've been handed a cookie that is not 320 * in the cache. If we're not translating 321 * 32 <-> 64, it may be a value that was 322 * flushed out of the cache because it grew 323 * too big. Let the server judge if it's 324 * valid or not. In the translation case, 325 * we have no way of validating this value, 326 * so punt. 327 */ 328 if (nmp->nm_flag & NFSMNT_XLATECOOKIE) 329 return (EINVAL); 330 ndp = nfs_enterdircache(vp, uio->uio_offset, 331 uio->uio_offset, 0, 0); 332 } 333 334 if (uio->uio_offset != 0 && 335 ndp->dc_cookie == np->n_direofoffset) { 336 nfsstats.direofcache_hits++; 337 return (0); 338 } 339 340 bp = nfs_getcacheblk(vp, ndp->dc_blkno, NFS_DIRBLKSIZ, p); 341 if (!bp) 342 return (EINTR); 343 if ((bp->b_flags & B_DONE) == 0) { 344 bp->b_flags |= B_READ; 345 bp->b_dcookie = ndp->dc_blkcookie; 346 error = nfs_doio(bp, cred, p); 347 if (error) { 348 /* 349 * Yuck! The directory has been modified on the 350 * server. Punt and let the userland code 351 * deal with it. 352 */ 353 brelse(bp); 354 if (error == NFSERR_BAD_COOKIE) { 355 nfs_invaldircache(vp, 0); 356 nfs_vinvalbuf(vp, 0, cred, p, 1); 357 error = EINVAL; 358 } 359 return (error); 360 } 361 } 362 363 /* 364 * Just return if we hit EOF right away with this 365 * block. Always check here, because direofoffset 366 * may have been set by an nfsiod since the last 367 * check. 368 */ 369 if (np->n_direofoffset != 0 && 370 ndp->dc_blkcookie == np->n_direofoffset) { 371 brelse(bp); 372 return (0); 373 } 374 375 /* 376 * Find the entry we were looking for in the block. 377 */ 378 379 en = ndp->dc_entry; 380 381 pdp = dp = (struct dirent *)bp->b_data; 382 edp = bp->b_data + bp->b_validend; 383 enn = 0; 384 while (enn < en && (caddr_t)dp < edp) { 385 pdp = dp; 386 dp = (struct dirent *)((caddr_t)dp + dp->d_reclen); 387 enn++; 388 } 389 390 /* 391 * If the entry number was bigger than the number of 392 * entries in the block, or the cookie of the previous 393 * entry doesn't match, the directory cache is 394 * stale. Flush it and try again (i.e. go to 395 * the server). 396 */ 397 if ((caddr_t)dp >= edp || (caddr_t)dp + dp->d_reclen > edp || 398 (en > 0 && NFS_GETCOOKIE(pdp) != ndp->dc_cookie)) { 399 #ifdef DEBUG 400 printf("invalid cache: %p %p %p off %lx %lx\n", 401 pdp, dp, edp, 402 (unsigned long)uio->uio_offset, 403 (unsigned long)NFS_GETCOOKIE(pdp)); 404 #endif 405 brelse(bp); 406 nfs_invaldircache(vp, 0); 407 nfs_vinvalbuf(vp, 0, cred, p, 0); 408 goto diragain; 409 } 410 411 on = (caddr_t)dp - bp->b_data; 412 413 /* 414 * Cache all entries that may be exported to the 415 * user, as they may be thrown back at us. The 416 * NFSBIO_CACHECOOKIES flag indicates that all 417 * entries are being 'exported', so cache them all. 418 */ 419 420 if (en == 0 && pdp == dp) { 421 dp = (struct dirent *) 422 ((caddr_t)dp + dp->d_reclen); 423 enn++; 424 } 425 426 if (uio->uio_resid < (bp->b_validend - on)) { 427 n = uio->uio_resid; 428 enough = 1; 429 } else 430 n = bp->b_validend - on; 431 432 ep = bp->b_data + on + n; 433 434 /* 435 * Find last complete entry to copy, caching entries 436 * (if requested) as we go. 437 */ 438 439 while ((caddr_t)dp < ep && (caddr_t)dp + dp->d_reclen <= ep) { 440 if (cflag & NFSBIO_CACHECOOKIES) { 441 nndp = nfs_enterdircache(vp, NFS_GETCOOKIE(pdp), 442 ndp->dc_blkcookie, enn, bp->b_lblkno); 443 if (nmp->nm_flag & NFSMNT_XLATECOOKIE) { 444 NFS_STASHCOOKIE32(pdp, 445 nndp->dc_cookie32); 446 } 447 } 448 pdp = dp; 449 dp = (struct dirent *)((caddr_t)dp + dp->d_reclen); 450 enn++; 451 } 452 453 /* 454 * If the last requested entry was not the last in the 455 * buffer (happens if NFS_DIRFRAGSIZ < NFS_DIRBLKSIZ), 456 * cache the cookie of the last requested one, and 457 * set of the offset to it. 458 */ 459 460 if ((on + n) < bp->b_validend) { 461 curoff = NFS_GETCOOKIE(pdp); 462 nndp = nfs_enterdircache(vp, curoff, ndp->dc_blkcookie, 463 enn, bp->b_lblkno); 464 if (nmp->nm_flag & NFSMNT_XLATECOOKIE) { 465 NFS_STASHCOOKIE32(pdp, nndp->dc_cookie32); 466 curoff = nndp->dc_cookie32; 467 } 468 } else 469 curoff = bp->b_dcookie; 470 471 /* 472 * Always cache the entry for the next block, 473 * so that readaheads can use it. 474 */ 475 nndp = nfs_enterdircache(vp, bp->b_dcookie, bp->b_dcookie, 0,0); 476 if (nmp->nm_flag & NFSMNT_XLATECOOKIE) { 477 if (curoff == bp->b_dcookie) { 478 NFS_STASHCOOKIE32(pdp, nndp->dc_cookie32); 479 curoff = nndp->dc_cookie32; 480 } 481 } 482 483 n = ((caddr_t)pdp + pdp->d_reclen) - (bp->b_data + on); 484 485 /* 486 * If not eof and read aheads are enabled, start one. 487 * (You need the current block first, so that you have the 488 * directory offset cookie of the next block.) 489 */ 490 if (nfs_numasync > 0 && nmp->nm_readahead > 0 && 491 np->n_direofoffset == 0 && !(np->n_flag & NQNFSNONCACHE)) { 492 rabp = nfs_getcacheblk(vp, nndp->dc_blkno, 493 NFS_DIRBLKSIZ, p); 494 if (rabp) { 495 if ((rabp->b_flags & (B_DONE | B_DELWRI)) == 0) { 496 rabp->b_dcookie = nndp->dc_cookie; 497 rabp->b_flags |= (B_READ | B_ASYNC); 498 if (nfs_asyncio(rabp, cred)) { 499 rabp->b_flags |= B_INVAL; 500 brelse(rabp); 501 } 502 } else 503 brelse(rabp); 504 } 505 } 506 got_buf = 1; 507 break; 508 default: 509 printf(" nfsbioread: type %x unexpected\n",vp->v_type); 510 break; 511 }; 512 513 if (n > 0) { 514 if (!baddr) 515 baddr = bp->b_data; 516 error = uiomove(baddr + on, (int)n, uio); 517 } 518 switch (vp->v_type) { 519 case VREG: 520 break; 521 case VLNK: 522 n = 0; 523 break; 524 case VDIR: 525 if (np->n_flag & NQNFSNONCACHE) 526 bp->b_flags |= B_INVAL; 527 uio->uio_offset = curoff; 528 if (enough) 529 n = 0; 530 break; 531 default: 532 printf(" nfsbioread: type %x unexpected\n",vp->v_type); 533 } 534 if (got_buf) 535 brelse(bp); 536 } while (error == 0 && uio->uio_resid > 0 && n > 0); 537 return (error); 538 } 539 540 /* 541 * Vnode op for write using bio 542 */ 543 int 544 nfs_write(v) 545 void *v; 546 { 547 struct vop_write_args /* { 548 struct vnode *a_vp; 549 struct uio *a_uio; 550 int a_ioflag; 551 struct ucred *a_cred; 552 } */ *ap = v; 553 int biosize; 554 struct uio *uio = ap->a_uio; 555 struct proc *p = uio->uio_procp; 556 struct vnode *vp = ap->a_vp; 557 struct nfsnode *np = VTONFS(vp); 558 struct ucred *cred = ap->a_cred; 559 int ioflag = ap->a_ioflag; 560 struct buf *bp; 561 struct vattr vattr; 562 struct nfsmount *nmp = VFSTONFS(vp->v_mount); 563 daddr_t lbn, bn; 564 int n, on, error = 0, iomode, must_commit; 565 566 #ifdef DIAGNOSTIC 567 if (uio->uio_rw != UIO_WRITE) 568 panic("nfs_write mode"); 569 if (uio->uio_segflg == UIO_USERSPACE && uio->uio_procp != curproc) 570 panic("nfs_write proc"); 571 #endif 572 if (vp->v_type != VREG) 573 return (EIO); 574 if (np->n_flag & NWRITEERR) { 575 np->n_flag &= ~NWRITEERR; 576 return (np->n_error); 577 } 578 if ((nmp->nm_flag & NFSMNT_NFSV3) && 579 !(nmp->nm_iflag & NFSMNT_GOTFSINFO)) 580 (void)nfs_fsinfo(nmp, vp, cred, p); 581 if (ioflag & (IO_APPEND | IO_SYNC)) { 582 if (np->n_flag & NMODIFIED) { 583 np->n_attrstamp = 0; 584 error = nfs_vinvalbuf(vp, V_SAVE, cred, p, 1); 585 if (error) 586 return (error); 587 } 588 if (ioflag & IO_APPEND) { 589 np->n_attrstamp = 0; 590 error = VOP_GETATTR(vp, &vattr, cred, p); 591 if (error) 592 return (error); 593 uio->uio_offset = np->n_size; 594 } 595 } 596 if (uio->uio_offset < 0) 597 return (EINVAL); 598 if ((uio->uio_offset + uio->uio_resid) > nmp->nm_maxfilesize) 599 return (EFBIG); 600 if (uio->uio_resid == 0) 601 return (0); 602 /* 603 * Maybe this should be above the vnode op call, but so long as 604 * file servers have no limits, i don't think it matters 605 */ 606 if (p && uio->uio_offset + uio->uio_resid > 607 p->p_rlimit[RLIMIT_FSIZE].rlim_cur) { 608 psignal(p, SIGXFSZ); 609 return (EFBIG); 610 } 611 /* 612 * I use nm_rsize, not nm_wsize so that all buffer cache blocks 613 * will be the same size within a filesystem. nfs_writerpc will 614 * still use nm_wsize when sizing the rpc's. 615 */ 616 biosize = nmp->nm_rsize; 617 do { 618 619 /* 620 * XXX make sure we aren't cached in the VM page cache 621 */ 622 (void)uvm_vnp_uncache(vp); 623 624 /* 625 * Check for a valid write lease. 626 */ 627 if ((nmp->nm_flag & NFSMNT_NQNFS) && 628 NQNFS_CKINVALID(vp, np, ND_WRITE)) { 629 do { 630 error = nqnfs_getlease(vp, ND_WRITE, cred, p); 631 } while (error == NQNFS_EXPIRED); 632 if (error) 633 return (error); 634 if (np->n_lrev != np->n_brev || 635 (np->n_flag & NQNFSNONCACHE)) { 636 error = nfs_vinvalbuf(vp, V_SAVE, cred, p, 1); 637 if (error) 638 return (error); 639 np->n_brev = np->n_lrev; 640 } 641 } 642 if ((np->n_flag & NQNFSNONCACHE) && uio->uio_iovcnt == 1) { 643 iomode = NFSV3WRITE_FILESYNC; 644 error = nfs_writerpc(vp, uio, cred, &iomode, &must_commit); 645 if (must_commit) 646 nfs_clearcommit(vp->v_mount); 647 return (error); 648 } 649 nfsstats.biocache_writes++; 650 lbn = uio->uio_offset / biosize; 651 on = uio->uio_offset & (biosize-1); 652 n = min((unsigned)(biosize - on), uio->uio_resid); 653 bn = lbn * (biosize / DEV_BSIZE); 654 again: 655 bp = nfs_getcacheblk(vp, bn, biosize, p); 656 if (!bp) 657 return (EINTR); 658 if (bp->b_wcred == NOCRED) { 659 crhold(cred); 660 bp->b_wcred = cred; 661 } 662 np->n_flag |= NMODIFIED; 663 if (uio->uio_offset + n > np->n_size) { 664 np->n_size = uio->uio_offset + n; 665 uvm_vnp_setsize(vp, np->n_size); 666 } 667 668 /* 669 * If the new write will leave a contiguous dirty 670 * area, just update the b_dirtyoff and b_dirtyend, 671 * otherwise force a write rpc of the old dirty area. 672 */ 673 if (bp->b_dirtyend > 0 && 674 (on > bp->b_dirtyend || (on + n) < bp->b_dirtyoff)) { 675 bp->b_proc = p; 676 if (VOP_BWRITE(bp) == EINTR) 677 return (EINTR); 678 goto again; 679 } 680 681 /* 682 * Check for valid write lease and get one as required. 683 * In case getblk() and/or bwrite() delayed us. 684 */ 685 if ((nmp->nm_flag & NFSMNT_NQNFS) && 686 NQNFS_CKINVALID(vp, np, ND_WRITE)) { 687 do { 688 error = nqnfs_getlease(vp, ND_WRITE, cred, p); 689 } while (error == NQNFS_EXPIRED); 690 if (error) { 691 brelse(bp); 692 return (error); 693 } 694 if (np->n_lrev != np->n_brev || 695 (np->n_flag & NQNFSNONCACHE)) { 696 brelse(bp); 697 error = nfs_vinvalbuf(vp, V_SAVE, cred, p, 1); 698 if (error) 699 return (error); 700 np->n_brev = np->n_lrev; 701 goto again; 702 } 703 } 704 error = uiomove((char *)bp->b_data + on, n, uio); 705 if (error) { 706 bp->b_flags |= B_ERROR; 707 brelse(bp); 708 return (error); 709 } 710 if (bp->b_dirtyend > 0) { 711 bp->b_dirtyoff = min(on, bp->b_dirtyoff); 712 bp->b_dirtyend = max((on + n), bp->b_dirtyend); 713 } else { 714 bp->b_dirtyoff = on; 715 bp->b_dirtyend = on + n; 716 } 717 if (bp->b_validend == 0 || bp->b_validend < bp->b_dirtyoff || 718 bp->b_validoff > bp->b_dirtyend) { 719 bp->b_validoff = bp->b_dirtyoff; 720 bp->b_validend = bp->b_dirtyend; 721 } else { 722 bp->b_validoff = min(bp->b_validoff, bp->b_dirtyoff); 723 bp->b_validend = max(bp->b_validend, bp->b_dirtyend); 724 } 725 726 /* 727 * Since this block is being modified, it must be written 728 * again and not just committed. 729 */ 730 bp->b_flags &= ~B_NEEDCOMMIT; 731 732 /* 733 * If the lease is non-cachable or IO_SYNC do bwrite(). 734 */ 735 if ((np->n_flag & NQNFSNONCACHE) || (ioflag & IO_SYNC)) { 736 bp->b_proc = p; 737 error = VOP_BWRITE(bp); 738 if (error) 739 return (error); 740 if (np->n_flag & NQNFSNONCACHE) { 741 error = nfs_vinvalbuf(vp, V_SAVE, cred, p, 1); 742 if (error) 743 return (error); 744 } 745 } else if ((n + on) == biosize && 746 (nmp->nm_flag & NFSMNT_NQNFS) == 0) { 747 bp->b_proc = (struct proc *)0; 748 bp->b_flags |= B_ASYNC; 749 (void)nfs_writebp(bp, 0); 750 } else { 751 bdwrite(bp); 752 } 753 } while (uio->uio_resid > 0 && n > 0); 754 return (0); 755 } 756 757 /* 758 * Get an nfs cache block. 759 * Allocate a new one if the block isn't currently in the cache 760 * and return the block marked busy. If the calling process is 761 * interrupted by a signal for an interruptible mount point, return 762 * NULL. 763 */ 764 struct buf * 765 nfs_getcacheblk(vp, bn, size, p) 766 struct vnode *vp; 767 daddr_t bn; 768 int size; 769 struct proc *p; 770 { 771 struct buf *bp; 772 struct nfsmount *nmp = VFSTONFS(vp->v_mount); 773 774 if (nmp->nm_flag & NFSMNT_INT) { 775 bp = getblk(vp, bn, size, PCATCH, 0); 776 while (bp == (struct buf *)0) { 777 if (nfs_sigintr(nmp, (struct nfsreq *)0, p)) 778 return ((struct buf *)0); 779 bp = getblk(vp, bn, size, 0, 2 * hz); 780 } 781 } else 782 bp = getblk(vp, bn, size, 0, 0); 783 return (bp); 784 } 785 786 /* 787 * Flush and invalidate all dirty buffers. If another process is already 788 * doing the flush, just wait for completion. 789 */ 790 int 791 nfs_vinvalbuf(vp, flags, cred, p, intrflg) 792 struct vnode *vp; 793 int flags; 794 struct ucred *cred; 795 struct proc *p; 796 int intrflg; 797 { 798 struct nfsnode *np = VTONFS(vp); 799 struct nfsmount *nmp = VFSTONFS(vp->v_mount); 800 int error = 0, slpflag, slptimeo; 801 802 if ((nmp->nm_flag & NFSMNT_INT) == 0) 803 intrflg = 0; 804 if (intrflg) { 805 slpflag = PCATCH; 806 slptimeo = 2 * hz; 807 } else { 808 slpflag = 0; 809 slptimeo = 0; 810 } 811 /* 812 * First wait for any other process doing a flush to complete. 813 */ 814 while (np->n_flag & NFLUSHINPROG) { 815 np->n_flag |= NFLUSHWANT; 816 error = tsleep((caddr_t)&np->n_flag, PRIBIO + 2, "nfsvinval", 817 slptimeo); 818 if (error && intrflg && nfs_sigintr(nmp, (struct nfsreq *)0, p)) 819 return (EINTR); 820 } 821 822 /* 823 * Now, flush as required. 824 */ 825 np->n_flag |= NFLUSHINPROG; 826 error = vinvalbuf(vp, flags, cred, p, slpflag, 0); 827 while (error) { 828 if (intrflg && nfs_sigintr(nmp, (struct nfsreq *)0, p)) { 829 np->n_flag &= ~NFLUSHINPROG; 830 if (np->n_flag & NFLUSHWANT) { 831 np->n_flag &= ~NFLUSHWANT; 832 wakeup((caddr_t)&np->n_flag); 833 } 834 return (EINTR); 835 } 836 error = vinvalbuf(vp, flags, cred, p, 0, slptimeo); 837 } 838 np->n_flag &= ~(NMODIFIED | NFLUSHINPROG); 839 if (np->n_flag & NFLUSHWANT) { 840 np->n_flag &= ~NFLUSHWANT; 841 wakeup((caddr_t)&np->n_flag); 842 } 843 return (0); 844 } 845 846 /* 847 * Initiate asynchronous I/O. Return an error if no nfsiods are available. 848 * This is mainly to avoid queueing async I/O requests when the nfsiods 849 * are all hung on a dead server. 850 */ 851 int 852 nfs_asyncio(bp, cred) 853 struct buf *bp; 854 struct ucred *cred; 855 { 856 int i; 857 struct nfsmount *nmp; 858 int gotiod, slpflag = 0, slptimeo = 0, error; 859 860 if (nfs_numasync == 0) 861 return (EIO); 862 863 864 nmp = VFSTONFS(bp->b_vp->v_mount); 865 again: 866 if (nmp->nm_flag & NFSMNT_INT) 867 slpflag = PCATCH; 868 gotiod = FALSE; 869 870 /* 871 * Find a free iod to process this request. 872 */ 873 874 for (i = 0; i < NFS_MAXASYNCDAEMON; i++) 875 if (nfs_iodwant[i]) { 876 /* 877 * Found one, so wake it up and tell it which 878 * mount to process. 879 */ 880 nfs_iodwant[i] = (struct proc *)0; 881 nfs_iodmount[i] = nmp; 882 nmp->nm_bufqiods++; 883 wakeup((caddr_t)&nfs_iodwant[i]); 884 gotiod = TRUE; 885 break; 886 } 887 /* 888 * If none are free, we may already have an iod working on this mount 889 * point. If so, it will process our request. 890 */ 891 if (!gotiod && nmp->nm_bufqiods > 0) 892 gotiod = TRUE; 893 894 /* 895 * If we have an iod which can process the request, then queue 896 * the buffer. 897 */ 898 if (gotiod) { 899 /* 900 * Ensure that the queue never grows too large. 901 */ 902 while (nmp->nm_bufqlen >= 2*nfs_numasync) { 903 nmp->nm_bufqwant = TRUE; 904 error = tsleep(&nmp->nm_bufq, slpflag | PRIBIO, 905 "nfsaio", slptimeo); 906 if (error) { 907 if (nfs_sigintr(nmp, NULL, bp->b_proc)) 908 return (EINTR); 909 if (slpflag == PCATCH) { 910 slpflag = 0; 911 slptimeo = 2 * hz; 912 } 913 } 914 /* 915 * We might have lost our iod while sleeping, 916 * so check and loop if nescessary. 917 */ 918 if (nmp->nm_bufqiods == 0) 919 goto again; 920 } 921 922 if (bp->b_flags & B_READ) { 923 if (bp->b_rcred == NOCRED && cred != NOCRED) { 924 crhold(cred); 925 bp->b_rcred = cred; 926 } 927 } else { 928 bp->b_flags |= B_WRITEINPROG; 929 if (bp->b_wcred == NOCRED && cred != NOCRED) { 930 crhold(cred); 931 bp->b_wcred = cred; 932 } 933 } 934 935 TAILQ_INSERT_TAIL(&nmp->nm_bufq, bp, b_freelist); 936 nmp->nm_bufqlen++; 937 return (0); 938 } 939 940 /* 941 * All the iods are busy on other mounts, so return EIO to 942 * force the caller to process the i/o synchronously. 943 */ 944 return (EIO); 945 } 946 947 /* 948 * Do an I/O operation to/from a cache block. This may be called 949 * synchronously or from an nfsiod. 950 */ 951 int 952 nfs_doio(bp, cr, p) 953 struct buf *bp; 954 struct ucred *cr; 955 struct proc *p; 956 { 957 struct uio *uiop; 958 struct vnode *vp; 959 struct nfsnode *np; 960 struct nfsmount *nmp; 961 int error = 0, diff, len, iomode, must_commit = 0, s; 962 struct uio uio; 963 struct iovec io; 964 965 vp = bp->b_vp; 966 np = VTONFS(vp); 967 nmp = VFSTONFS(vp->v_mount); 968 uiop = &uio; 969 uiop->uio_iov = &io; 970 uiop->uio_iovcnt = 1; 971 uiop->uio_segflg = UIO_SYSSPACE; 972 uiop->uio_procp = p; 973 974 /* 975 * Historically, paging was done with physio, but no more... 976 */ 977 if (bp->b_flags & B_PHYS) { 978 /* 979 * ...though reading /dev/drum still gets us here. 980 */ 981 io.iov_len = uiop->uio_resid = bp->b_bcount; 982 /* mapping was done by vmapbuf() */ 983 io.iov_base = bp->b_data; 984 uiop->uio_offset = ((off_t)bp->b_blkno) * DEV_BSIZE; 985 if (bp->b_flags & B_READ) { 986 uiop->uio_rw = UIO_READ; 987 nfsstats.read_physios++; 988 error = nfs_readrpc(vp, uiop, cr); 989 } else { 990 iomode = NFSV3WRITE_DATASYNC; 991 uiop->uio_rw = UIO_WRITE; 992 nfsstats.write_physios++; 993 error = nfs_writerpc(vp, uiop, cr, &iomode, &must_commit); 994 } 995 if (error) { 996 bp->b_flags |= B_ERROR; 997 bp->b_error = error; 998 } 999 } else if (bp->b_flags & B_READ) { 1000 io.iov_len = uiop->uio_resid = bp->b_bcount; 1001 io.iov_base = bp->b_data; 1002 uiop->uio_rw = UIO_READ; 1003 switch (vp->v_type) { 1004 case VREG: 1005 uiop->uio_offset = ((off_t)bp->b_blkno) * DEV_BSIZE; 1006 nfsstats.read_bios++; 1007 error = nfs_readrpc(vp, uiop, cr); 1008 if (!error) { 1009 bp->b_validoff = 0; 1010 if (uiop->uio_resid) { 1011 /* 1012 * If len > 0, there is a hole in the file and 1013 * no writes after the hole have been pushed to 1014 * the server yet. 1015 * Just zero fill the rest of the valid area. 1016 */ 1017 diff = bp->b_bcount - uiop->uio_resid; 1018 len = np->n_size - (((u_quad_t)bp->b_blkno) * DEV_BSIZE 1019 + diff); 1020 if (len > 0) { 1021 len = min(len, uiop->uio_resid); 1022 memset((char *)bp->b_data + diff, 0, len); 1023 bp->b_validend = diff + len; 1024 } else 1025 bp->b_validend = diff; 1026 } else 1027 bp->b_validend = bp->b_bcount; 1028 } 1029 if (p && (vp->v_flag & VTEXT) && 1030 (((nmp->nm_flag & NFSMNT_NQNFS) && 1031 NQNFS_CKINVALID(vp, np, ND_READ) && 1032 np->n_lrev != np->n_brev) || 1033 (!(nmp->nm_flag & NFSMNT_NQNFS) && 1034 np->n_mtime != np->n_vattr->va_mtime.tv_sec))) { 1035 uprintf("Process killed due to text file modification\n"); 1036 psignal(p, SIGKILL); 1037 p->p_holdcnt++; 1038 } 1039 break; 1040 case VLNK: 1041 uiop->uio_offset = (off_t)0; 1042 nfsstats.readlink_bios++; 1043 error = nfs_readlinkrpc(vp, uiop, cr); 1044 break; 1045 case VDIR: 1046 nfsstats.readdir_bios++; 1047 uiop->uio_offset = bp->b_dcookie; 1048 if (nmp->nm_flag & NFSMNT_RDIRPLUS) { 1049 error = nfs_readdirplusrpc(vp, uiop, cr); 1050 if (error == NFSERR_NOTSUPP) 1051 nmp->nm_flag &= ~NFSMNT_RDIRPLUS; 1052 } 1053 if ((nmp->nm_flag & NFSMNT_RDIRPLUS) == 0) 1054 error = nfs_readdirrpc(vp, uiop, cr); 1055 if (!error) { 1056 bp->b_dcookie = uiop->uio_offset; 1057 bp->b_validoff = 0; 1058 bp->b_validend = bp->b_bcount - uiop->uio_resid; 1059 } 1060 break; 1061 default: 1062 printf("nfs_doio: type %x unexpected\n",vp->v_type); 1063 break; 1064 }; 1065 if (error) { 1066 bp->b_flags |= B_ERROR; 1067 bp->b_error = error; 1068 } 1069 } else { 1070 io.iov_len = uiop->uio_resid = bp->b_dirtyend 1071 - bp->b_dirtyoff; 1072 uiop->uio_offset = ((off_t)bp->b_blkno) * DEV_BSIZE 1073 + bp->b_dirtyoff; 1074 io.iov_base = (char *)bp->b_data + bp->b_dirtyoff; 1075 uiop->uio_rw = UIO_WRITE; 1076 nfsstats.write_bios++; 1077 if ((bp->b_flags & (B_ASYNC | B_NEEDCOMMIT | B_NOCACHE)) == B_ASYNC) 1078 iomode = NFSV3WRITE_UNSTABLE; 1079 else 1080 iomode = NFSV3WRITE_FILESYNC; 1081 bp->b_flags |= B_WRITEINPROG; 1082 #ifdef fvdl_debug 1083 printf("nfs_doio(%p): bp %p doff %d dend %d\n", 1084 vp, bp, bp->b_dirtyoff, bp->b_dirtyend); 1085 #endif 1086 error = nfs_writerpc(vp, uiop, cr, &iomode, &must_commit); 1087 s = splbio(); 1088 if (!error && iomode == NFSV3WRITE_UNSTABLE) 1089 bp->b_flags |= B_NEEDCOMMIT; 1090 else 1091 bp->b_flags &= ~B_NEEDCOMMIT; 1092 bp->b_flags &= ~B_WRITEINPROG; 1093 1094 /* 1095 * For an interrupted write, the buffer is still valid and the 1096 * write hasn't been pushed to the server yet, so we can't set 1097 * B_ERROR and report the interruption by setting B_EINTR. For 1098 * the B_ASYNC case, B_EINTR is not relevant, so the rpc attempt 1099 * is essentially a noop. 1100 * For the case of a V3 write rpc not being committed to stable 1101 * storage, the block is still dirty and requires either a commit 1102 * rpc or another write rpc with iomode == NFSV3WRITE_FILESYNC 1103 * before the block is reused. This is indicated by setting the 1104 * B_DELWRI and B_NEEDCOMMIT flags. 1105 */ 1106 if (error == EINTR || (!error && (bp->b_flags & B_NEEDCOMMIT))) { 1107 bp->b_flags |= B_DELWRI; 1108 1109 /* 1110 * Since for the B_ASYNC case, nfs_bwrite() has reassigned the 1111 * buffer to the clean list, we have to reassign it back to the 1112 * dirty one. Ugh. 1113 */ 1114 if (bp->b_flags & B_ASYNC) { 1115 reassignbuf(bp, vp); 1116 } else if (error) 1117 bp->b_flags |= B_EINTR; 1118 } else { 1119 if (error) { 1120 bp->b_flags |= B_ERROR; 1121 bp->b_error = np->n_error = error; 1122 np->n_flag |= NWRITEERR; 1123 } 1124 bp->b_dirtyoff = bp->b_dirtyend = 0; 1125 } 1126 splx(s); 1127 } 1128 bp->b_resid = uiop->uio_resid; 1129 if (must_commit) 1130 nfs_clearcommit(vp->v_mount); 1131 biodone(bp); 1132 return (error); 1133 } 1134