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 8.9 (Berkeley) 3/30/95 37 * $FreeBSD: /repoman/r/ncvs/src/sys/nfsclient/nfs_bio.c,v 1.130 2004/04/14 23:23:55 peadar Exp $ 38 * $DragonFly: src/sys/vfs/nfs/nfs_bio.c,v 1.31 2006/04/10 17:46:44 dillon Exp $ 39 */ 40 41 42 #include <sys/param.h> 43 #include <sys/systm.h> 44 #include <sys/resourcevar.h> 45 #include <sys/signalvar.h> 46 #include <sys/proc.h> 47 #include <sys/buf.h> 48 #include <sys/vnode.h> 49 #include <sys/mount.h> 50 #include <sys/kernel.h> 51 #include <sys/buf2.h> 52 #include <sys/msfbuf.h> 53 54 #include <vm/vm.h> 55 #include <vm/vm_extern.h> 56 #include <vm/vm_page.h> 57 #include <vm/vm_object.h> 58 #include <vm/vm_pager.h> 59 #include <vm/vnode_pager.h> 60 61 #include <sys/thread2.h> 62 63 #include "rpcv2.h" 64 #include "nfsproto.h" 65 #include "nfs.h" 66 #include "nfsmount.h" 67 #include "nfsnode.h" 68 69 static struct buf *nfs_getcacheblk(struct vnode *vp, off_t loffset, 70 int size, struct thread *td); 71 static int nfs_check_dirent(struct nfs_dirent *dp, int maxlen); 72 73 extern int nfs_numasync; 74 extern int nfs_pbuf_freecnt; 75 extern struct nfsstats nfsstats; 76 77 /* 78 * Vnode op for VM getpages. 79 * 80 * nfs_getpages(struct vnode *a_vp, vm_page_t *a_m, int a_count, 81 * int a_reqpage, vm_ooffset_t a_offset) 82 */ 83 int 84 nfs_getpages(struct vop_getpages_args *ap) 85 { 86 struct thread *td = curthread; /* XXX */ 87 int i, error, nextoff, size, toff, count, npages; 88 struct uio uio; 89 struct iovec iov; 90 char *kva; 91 struct vnode *vp; 92 struct nfsmount *nmp; 93 vm_page_t *pages; 94 vm_page_t m; 95 struct msf_buf *msf; 96 97 vp = ap->a_vp; 98 nmp = VFSTONFS(vp->v_mount); 99 pages = ap->a_m; 100 count = ap->a_count; 101 102 if (vp->v_object == NULL) { 103 printf("nfs_getpages: called with non-merged cache vnode??\n"); 104 return VM_PAGER_ERROR; 105 } 106 107 if ((nmp->nm_flag & NFSMNT_NFSV3) != 0 && 108 (nmp->nm_state & NFSSTA_GOTFSINFO) == 0) 109 (void)nfs_fsinfo(nmp, vp, td); 110 111 npages = btoc(count); 112 113 /* 114 * NOTE that partially valid pages may occur in cases other 115 * then file EOF, such as when a file is partially written and 116 * ftruncate()-extended to a larger size. It is also possible 117 * for the valid bits to be set on garbage beyond the file EOF and 118 * clear in the area before EOF (e.g. m->valid == 0xfc), which can 119 * occur due to vtruncbuf() and the buffer cache's handling of 120 * pages which 'straddle' buffers or when b_bufsize is not a 121 * multiple of PAGE_SIZE.... the buffer cache cannot normally 122 * clear the extra bits. This kind of situation occurs when you 123 * make a small write() (m->valid == 0x03) and then mmap() and 124 * fault in the buffer(m->valid = 0xFF). When NFS flushes the 125 * buffer (vinvalbuf() m->valid = 0xFC) we are left with a mess. 126 * 127 * This is combined with the possibility that the pages are partially 128 * dirty or that there is a buffer backing the pages that is dirty 129 * (even if m->dirty is 0). 130 * 131 * To solve this problem several hacks have been made: (1) NFS 132 * guarentees that the IO block size is a multiple of PAGE_SIZE and 133 * (2) The buffer cache, when invalidating an NFS buffer, will 134 * disregard the buffer's fragmentory b_bufsize and invalidate 135 * the whole page rather then just the piece the buffer owns. 136 * 137 * This allows us to assume that a partially valid page found here 138 * is fully valid (vm_fault will zero'd out areas of the page not 139 * marked as valid). 140 */ 141 m = pages[ap->a_reqpage]; 142 if (m->valid != 0) { 143 for (i = 0; i < npages; ++i) { 144 if (i != ap->a_reqpage) 145 vnode_pager_freepage(pages[i]); 146 } 147 return(0); 148 } 149 150 /* 151 * Use an MSF_BUF as a medium to retrieve data from the pages. 152 */ 153 msf_map_pagelist(&msf, pages, npages, 0); 154 KKASSERT(msf); 155 kva = msf_buf_kva(msf); 156 157 iov.iov_base = kva; 158 iov.iov_len = count; 159 uio.uio_iov = &iov; 160 uio.uio_iovcnt = 1; 161 uio.uio_offset = IDX_TO_OFF(pages[0]->pindex); 162 uio.uio_resid = count; 163 uio.uio_segflg = UIO_SYSSPACE; 164 uio.uio_rw = UIO_READ; 165 uio.uio_td = td; 166 167 error = nfs_readrpc(vp, &uio); 168 msf_buf_free(msf); 169 170 if (error && (uio.uio_resid == count)) { 171 printf("nfs_getpages: error %d\n", error); 172 for (i = 0; i < npages; ++i) { 173 if (i != ap->a_reqpage) 174 vnode_pager_freepage(pages[i]); 175 } 176 return VM_PAGER_ERROR; 177 } 178 179 /* 180 * Calculate the number of bytes read and validate only that number 181 * of bytes. Note that due to pending writes, size may be 0. This 182 * does not mean that the remaining data is invalid! 183 */ 184 185 size = count - uio.uio_resid; 186 187 for (i = 0, toff = 0; i < npages; i++, toff = nextoff) { 188 nextoff = toff + PAGE_SIZE; 189 m = pages[i]; 190 191 m->flags &= ~PG_ZERO; 192 193 if (nextoff <= size) { 194 /* 195 * Read operation filled an entire page 196 */ 197 m->valid = VM_PAGE_BITS_ALL; 198 vm_page_undirty(m); 199 } else if (size > toff) { 200 /* 201 * Read operation filled a partial page. 202 */ 203 m->valid = 0; 204 vm_page_set_validclean(m, 0, size - toff); 205 /* handled by vm_fault now */ 206 /* vm_page_zero_invalid(m, TRUE); */ 207 } else { 208 /* 209 * Read operation was short. If no error occured 210 * we may have hit a zero-fill section. We simply 211 * leave valid set to 0. 212 */ 213 ; 214 } 215 if (i != ap->a_reqpage) { 216 /* 217 * Whether or not to leave the page activated is up in 218 * the air, but we should put the page on a page queue 219 * somewhere (it already is in the object). Result: 220 * It appears that emperical results show that 221 * deactivating pages is best. 222 */ 223 224 /* 225 * Just in case someone was asking for this page we 226 * now tell them that it is ok to use. 227 */ 228 if (!error) { 229 if (m->flags & PG_WANTED) 230 vm_page_activate(m); 231 else 232 vm_page_deactivate(m); 233 vm_page_wakeup(m); 234 } else { 235 vnode_pager_freepage(m); 236 } 237 } 238 } 239 return 0; 240 } 241 242 /* 243 * Vnode op for VM putpages. 244 * 245 * nfs_putpages(struct vnode *a_vp, vm_page_t *a_m, int a_count, int a_sync, 246 * int *a_rtvals, vm_ooffset_t a_offset) 247 */ 248 int 249 nfs_putpages(struct vop_putpages_args *ap) 250 { 251 struct thread *td = curthread; 252 struct uio uio; 253 struct iovec iov; 254 char *kva; 255 int iomode, must_commit, i, error, npages, count; 256 off_t offset; 257 int *rtvals; 258 struct vnode *vp; 259 struct nfsmount *nmp; 260 struct nfsnode *np; 261 vm_page_t *pages; 262 struct msf_buf *msf; 263 264 vp = ap->a_vp; 265 np = VTONFS(vp); 266 nmp = VFSTONFS(vp->v_mount); 267 pages = ap->a_m; 268 count = ap->a_count; 269 rtvals = ap->a_rtvals; 270 npages = btoc(count); 271 offset = IDX_TO_OFF(pages[0]->pindex); 272 273 if ((nmp->nm_flag & NFSMNT_NFSV3) != 0 && 274 (nmp->nm_state & NFSSTA_GOTFSINFO) == 0) 275 (void)nfs_fsinfo(nmp, vp, td); 276 277 for (i = 0; i < npages; i++) { 278 rtvals[i] = VM_PAGER_AGAIN; 279 } 280 281 /* 282 * When putting pages, do not extend file past EOF. 283 */ 284 285 if (offset + count > np->n_size) { 286 count = np->n_size - offset; 287 if (count < 0) 288 count = 0; 289 } 290 291 /* 292 * Use an MSF_BUF as a medium to retrieve data from the pages. 293 */ 294 msf_map_pagelist(&msf, pages, npages, 0); 295 KKASSERT(msf); 296 kva = msf_buf_kva(msf); 297 298 iov.iov_base = kva; 299 iov.iov_len = count; 300 uio.uio_iov = &iov; 301 uio.uio_iovcnt = 1; 302 uio.uio_offset = offset; 303 uio.uio_resid = count; 304 uio.uio_segflg = UIO_SYSSPACE; 305 uio.uio_rw = UIO_WRITE; 306 uio.uio_td = td; 307 308 if ((ap->a_sync & VM_PAGER_PUT_SYNC) == 0) 309 iomode = NFSV3WRITE_UNSTABLE; 310 else 311 iomode = NFSV3WRITE_FILESYNC; 312 313 error = nfs_writerpc(vp, &uio, &iomode, &must_commit); 314 315 msf_buf_free(msf); 316 317 if (!error) { 318 int nwritten = round_page(count - uio.uio_resid) / PAGE_SIZE; 319 for (i = 0; i < nwritten; i++) { 320 rtvals[i] = VM_PAGER_OK; 321 vm_page_undirty(pages[i]); 322 } 323 if (must_commit) 324 nfs_clearcommit(vp->v_mount); 325 } 326 return rtvals[0]; 327 } 328 329 /* 330 * Vnode op for read using bio 331 */ 332 int 333 nfs_bioread(struct vnode *vp, struct uio *uio, int ioflag) 334 { 335 struct nfsnode *np = VTONFS(vp); 336 int biosize, i; 337 struct buf *bp = 0, *rabp; 338 struct vattr vattr; 339 struct thread *td; 340 struct nfsmount *nmp = VFSTONFS(vp->v_mount); 341 daddr_t lbn, rabn; 342 off_t raoffset; 343 off_t loffset; 344 int bcount; 345 int seqcount; 346 int nra, error = 0, n = 0, on = 0; 347 348 #ifdef DIAGNOSTIC 349 if (uio->uio_rw != UIO_READ) 350 panic("nfs_read mode"); 351 #endif 352 if (uio->uio_resid == 0) 353 return (0); 354 if (uio->uio_offset < 0) /* XXX VDIR cookies can be negative */ 355 return (EINVAL); 356 td = uio->uio_td; 357 358 if ((nmp->nm_flag & NFSMNT_NFSV3) != 0 && 359 (nmp->nm_state & NFSSTA_GOTFSINFO) == 0) 360 (void)nfs_fsinfo(nmp, vp, td); 361 if (vp->v_type != VDIR && 362 (uio->uio_offset + uio->uio_resid) > nmp->nm_maxfilesize) 363 return (EFBIG); 364 biosize = vp->v_mount->mnt_stat.f_iosize; 365 seqcount = (int)((off_t)(ioflag >> IO_SEQSHIFT) * biosize / BKVASIZE); 366 367 /* 368 * For nfs, cache consistency can only be maintained approximately. 369 * Although RFC1094 does not specify the criteria, the following is 370 * believed to be compatible with the reference port. 371 * 372 * NFS: If local changes have been made and this is a 373 * directory, the directory must be invalidated and 374 * the attribute cache must be cleared. 375 * 376 * GETATTR is called to synchronize the file size. 377 * 378 * If remote changes are detected local data is flushed 379 * and the cache is invalidated. 380 * 381 * NOTE: In the normal case the attribute cache is not 382 * cleared which means GETATTR may use cached data and 383 * not immediately detect changes made on the server. 384 */ 385 if ((np->n_flag & NLMODIFIED) && vp->v_type == VDIR) { 386 nfs_invaldir(vp); 387 error = nfs_vinvalbuf(vp, V_SAVE, td, 1); 388 if (error) 389 return (error); 390 np->n_attrstamp = 0; 391 } 392 error = VOP_GETATTR(vp, &vattr, td); 393 if (error) 394 return (error); 395 if (np->n_flag & NRMODIFIED) { 396 if (vp->v_type == VDIR) 397 nfs_invaldir(vp); 398 error = nfs_vinvalbuf(vp, V_SAVE, td, 1); 399 if (error) 400 return (error); 401 np->n_flag &= ~NRMODIFIED; 402 } 403 do { 404 if (np->n_flag & NDONTCACHE) { 405 switch (vp->v_type) { 406 case VREG: 407 return (nfs_readrpc(vp, uio)); 408 case VLNK: 409 return (nfs_readlinkrpc(vp, uio)); 410 case VDIR: 411 break; 412 default: 413 printf(" NDONTCACHE: type %x unexpected\n", vp->v_type); 414 break; 415 }; 416 } 417 switch (vp->v_type) { 418 case VREG: 419 nfsstats.biocache_reads++; 420 lbn = uio->uio_offset / biosize; 421 on = uio->uio_offset & (biosize - 1); 422 loffset = (off_t)lbn * biosize; 423 424 /* 425 * Start the read ahead(s), as required. 426 */ 427 if (nfs_numasync > 0 && nmp->nm_readahead > 0) { 428 for (nra = 0; nra < nmp->nm_readahead && nra < seqcount && 429 (off_t)(lbn + 1 + nra) * biosize < np->n_size; nra++) { 430 rabn = lbn + 1 + nra; 431 raoffset = (off_t)rabn * biosize; 432 if (!findblk(vp, raoffset)) { 433 rabp = nfs_getcacheblk(vp, raoffset, biosize, td); 434 if (!rabp) 435 return (EINTR); 436 if ((rabp->b_flags & (B_CACHE|B_DELWRI)) == 0) { 437 rabp->b_flags |= (B_READ | B_ASYNC); 438 vfs_busy_pages(rabp, 0); 439 if (nfs_asyncio(vp, &rabp->b_bio2, td)) { 440 rabp->b_flags |= B_INVAL|B_ERROR; 441 vfs_unbusy_pages(rabp); 442 brelse(rabp); 443 break; 444 } 445 } else { 446 brelse(rabp); 447 } 448 } 449 } 450 } 451 452 /* 453 * Obtain the buffer cache block. Figure out the buffer size 454 * when we are at EOF. If we are modifying the size of the 455 * buffer based on an EOF condition we need to hold 456 * nfs_rslock() through obtaining the buffer to prevent 457 * a potential writer-appender from messing with n_size. 458 * Otherwise we may accidently truncate the buffer and 459 * lose dirty data. 460 * 461 * Note that bcount is *not* DEV_BSIZE aligned. 462 */ 463 464 again: 465 bcount = biosize; 466 if (loffset >= np->n_size) { 467 bcount = 0; 468 } else if (loffset + biosize > np->n_size) { 469 bcount = np->n_size - loffset; 470 } 471 if (bcount != biosize) { 472 switch(nfs_rslock(np, td)) { 473 case ENOLCK: 474 goto again; 475 /* not reached */ 476 case EINTR: 477 case ERESTART: 478 return(EINTR); 479 /* not reached */ 480 default: 481 break; 482 } 483 } 484 485 bp = nfs_getcacheblk(vp, loffset, bcount, td); 486 487 if (bcount != biosize) 488 nfs_rsunlock(np, td); 489 if (!bp) 490 return (EINTR); 491 492 /* 493 * If B_CACHE is not set, we must issue the read. If this 494 * fails, we return an error. 495 */ 496 497 if ((bp->b_flags & B_CACHE) == 0) { 498 bp->b_flags |= B_READ; 499 vfs_busy_pages(bp, 0); 500 error = nfs_doio(vp, &bp->b_bio2, td); 501 if (error) { 502 brelse(bp); 503 return (error); 504 } 505 } 506 507 /* 508 * on is the offset into the current bp. Figure out how many 509 * bytes we can copy out of the bp. Note that bcount is 510 * NOT DEV_BSIZE aligned. 511 * 512 * Then figure out how many bytes we can copy into the uio. 513 */ 514 515 n = 0; 516 if (on < bcount) 517 n = min((unsigned)(bcount - on), uio->uio_resid); 518 break; 519 case VLNK: 520 biosize = min(NFS_MAXPATHLEN, np->n_size); 521 nfsstats.biocache_readlinks++; 522 bp = nfs_getcacheblk(vp, (off_t)0, biosize, td); 523 if (bp == NULL) 524 return (EINTR); 525 if ((bp->b_flags & B_CACHE) == 0) { 526 bp->b_flags |= B_READ; 527 vfs_busy_pages(bp, 0); 528 error = nfs_doio(vp, &bp->b_bio2, td); 529 if (error) { 530 bp->b_flags |= B_ERROR; 531 brelse(bp); 532 return (error); 533 } 534 } 535 n = min(uio->uio_resid, bp->b_bcount - bp->b_resid); 536 on = 0; 537 break; 538 case VDIR: 539 nfsstats.biocache_readdirs++; 540 if (np->n_direofoffset 541 && uio->uio_offset >= np->n_direofoffset) { 542 return (0); 543 } 544 lbn = (uoff_t)uio->uio_offset / NFS_DIRBLKSIZ; 545 on = uio->uio_offset & (NFS_DIRBLKSIZ - 1); 546 loffset = uio->uio_offset - on; 547 bp = nfs_getcacheblk(vp, loffset, NFS_DIRBLKSIZ, td); 548 if (bp == NULL) 549 return (EINTR); 550 551 if ((bp->b_flags & B_CACHE) == 0) { 552 bp->b_flags |= B_READ; 553 vfs_busy_pages(bp, 0); 554 error = nfs_doio(vp, &bp->b_bio2, td); 555 if (error) { 556 brelse(bp); 557 } 558 while (error == NFSERR_BAD_COOKIE) { 559 printf("got bad cookie vp %p bp %p\n", vp, bp); 560 nfs_invaldir(vp); 561 error = nfs_vinvalbuf(vp, 0, td, 1); 562 /* 563 * Yuck! The directory has been modified on the 564 * server. The only way to get the block is by 565 * reading from the beginning to get all the 566 * offset cookies. 567 * 568 * Leave the last bp intact unless there is an error. 569 * Loop back up to the while if the error is another 570 * NFSERR_BAD_COOKIE (double yuch!). 571 */ 572 for (i = 0; i <= lbn && !error; i++) { 573 if (np->n_direofoffset 574 && (i * NFS_DIRBLKSIZ) >= np->n_direofoffset) 575 return (0); 576 bp = nfs_getcacheblk(vp, (off_t)i * NFS_DIRBLKSIZ, 577 NFS_DIRBLKSIZ, td); 578 if (!bp) 579 return (EINTR); 580 if ((bp->b_flags & B_CACHE) == 0) { 581 bp->b_flags |= B_READ; 582 vfs_busy_pages(bp, 0); 583 error = nfs_doio(vp, &bp->b_bio2, td); 584 /* 585 * no error + B_INVAL == directory EOF, 586 * use the block. 587 */ 588 if (error == 0 && (bp->b_flags & B_INVAL)) 589 break; 590 } 591 /* 592 * An error will throw away the block and the 593 * for loop will break out. If no error and this 594 * is not the block we want, we throw away the 595 * block and go for the next one via the for loop. 596 */ 597 if (error || i < lbn) 598 brelse(bp); 599 } 600 } 601 /* 602 * The above while is repeated if we hit another cookie 603 * error. If we hit an error and it wasn't a cookie error, 604 * we give up. 605 */ 606 if (error) 607 return (error); 608 } 609 610 /* 611 * If not eof and read aheads are enabled, start one. 612 * (You need the current block first, so that you have the 613 * directory offset cookie of the next block.) 614 */ 615 if (nfs_numasync > 0 && nmp->nm_readahead > 0 && 616 (bp->b_flags & B_INVAL) == 0 && 617 (np->n_direofoffset == 0 || 618 loffset + NFS_DIRBLKSIZ < np->n_direofoffset) && 619 (np->n_flag & NDONTCACHE) == 0 && 620 !findblk(vp, loffset + NFS_DIRBLKSIZ)) { 621 rabp = nfs_getcacheblk(vp, loffset + NFS_DIRBLKSIZ, 622 NFS_DIRBLKSIZ, td); 623 if (rabp) { 624 if ((rabp->b_flags & (B_CACHE|B_DELWRI)) == 0) { 625 rabp->b_flags |= (B_READ | B_ASYNC); 626 vfs_busy_pages(rabp, 0); 627 if (nfs_asyncio(vp, &rabp->b_bio2, td)) { 628 rabp->b_flags |= B_INVAL|B_ERROR; 629 vfs_unbusy_pages(rabp); 630 brelse(rabp); 631 } 632 } else { 633 brelse(rabp); 634 } 635 } 636 } 637 /* 638 * Unlike VREG files, whos buffer size ( bp->b_bcount ) is 639 * chopped for the EOF condition, we cannot tell how large 640 * NFS directories are going to be until we hit EOF. So 641 * an NFS directory buffer is *not* chopped to its EOF. Now, 642 * it just so happens that b_resid will effectively chop it 643 * to EOF. *BUT* this information is lost if the buffer goes 644 * away and is reconstituted into a B_CACHE state ( due to 645 * being VMIO ) later. So we keep track of the directory eof 646 * in np->n_direofoffset and chop it off as an extra step 647 * right here. 648 */ 649 n = lmin(uio->uio_resid, NFS_DIRBLKSIZ - bp->b_resid - on); 650 if (np->n_direofoffset && n > np->n_direofoffset - uio->uio_offset) 651 n = np->n_direofoffset - uio->uio_offset; 652 break; 653 default: 654 printf(" nfs_bioread: type %x unexpected\n",vp->v_type); 655 break; 656 }; 657 658 switch (vp->v_type) { 659 case VREG: 660 if (n > 0) 661 error = uiomove(bp->b_data + on, (int)n, uio); 662 break; 663 case VLNK: 664 if (n > 0) 665 error = uiomove(bp->b_data + on, (int)n, uio); 666 n = 0; 667 break; 668 case VDIR: 669 if (n > 0) { 670 off_t old_off = uio->uio_offset; 671 caddr_t cpos, epos; 672 struct nfs_dirent *dp; 673 674 /* 675 * We are casting cpos to nfs_dirent, it must be 676 * int-aligned. 677 */ 678 if (on & 3) { 679 error = EINVAL; 680 break; 681 } 682 683 cpos = bp->b_data + on; 684 epos = bp->b_data + on + n; 685 while (cpos < epos && error == 0 && uio->uio_resid > 0) { 686 dp = (struct nfs_dirent *)cpos; 687 error = nfs_check_dirent(dp, (int)(epos - cpos)); 688 if (error) 689 break; 690 if (vop_write_dirent(&error, uio, dp->nfs_ino, 691 dp->nfs_type, dp->nfs_namlen, dp->nfs_name)) { 692 break; 693 } 694 cpos += dp->nfs_reclen; 695 } 696 n = 0; 697 if (error == 0) 698 uio->uio_offset = old_off + cpos - bp->b_data - on; 699 } 700 /* 701 * Invalidate buffer if caching is disabled, forcing a 702 * re-read from the remote later. 703 */ 704 if (np->n_flag & NDONTCACHE) 705 bp->b_flags |= B_INVAL; 706 break; 707 default: 708 printf(" nfs_bioread: type %x unexpected\n",vp->v_type); 709 } 710 brelse(bp); 711 } while (error == 0 && uio->uio_resid > 0 && n > 0); 712 return (error); 713 } 714 715 /* 716 * Userland can supply any 'seek' offset when reading a NFS directory. 717 * Validate the structure so we don't panic the kernel. Note that 718 * the element name is nul terminated and the nul is not included 719 * in nfs_namlen. 720 */ 721 static 722 int 723 nfs_check_dirent(struct nfs_dirent *dp, int maxlen) 724 { 725 int nfs_name_off = offsetof(struct nfs_dirent, nfs_name[0]); 726 727 if (nfs_name_off >= maxlen) 728 return (EINVAL); 729 if (dp->nfs_reclen < nfs_name_off || dp->nfs_reclen > maxlen) 730 return (EINVAL); 731 if (nfs_name_off + dp->nfs_namlen >= dp->nfs_reclen) 732 return (EINVAL); 733 if (dp->nfs_reclen & 3) 734 return (EINVAL); 735 return (0); 736 } 737 738 /* 739 * Vnode op for write using bio 740 * 741 * nfs_write(struct vnode *a_vp, struct uio *a_uio, int a_ioflag, 742 * struct ucred *a_cred) 743 */ 744 int 745 nfs_write(struct vop_write_args *ap) 746 { 747 struct uio *uio = ap->a_uio; 748 struct thread *td = uio->uio_td; 749 struct vnode *vp = ap->a_vp; 750 struct nfsnode *np = VTONFS(vp); 751 int ioflag = ap->a_ioflag; 752 struct buf *bp; 753 struct vattr vattr; 754 struct nfsmount *nmp = VFSTONFS(vp->v_mount); 755 daddr_t lbn; 756 off_t loffset; 757 int n, on, error = 0, iomode, must_commit; 758 int haverslock = 0; 759 int bcount; 760 int biosize; 761 762 #ifdef DIAGNOSTIC 763 if (uio->uio_rw != UIO_WRITE) 764 panic("nfs_write mode"); 765 if (uio->uio_segflg == UIO_USERSPACE && uio->uio_td != curthread) 766 panic("nfs_write proc"); 767 #endif 768 if (vp->v_type != VREG) 769 return (EIO); 770 if (np->n_flag & NWRITEERR) { 771 np->n_flag &= ~NWRITEERR; 772 return (np->n_error); 773 } 774 if ((nmp->nm_flag & NFSMNT_NFSV3) != 0 && 775 (nmp->nm_state & NFSSTA_GOTFSINFO) == 0) 776 (void)nfs_fsinfo(nmp, vp, td); 777 778 /* 779 * Synchronously flush pending buffers if we are in synchronous 780 * mode or if we are appending. 781 */ 782 if (ioflag & (IO_APPEND | IO_SYNC)) { 783 if (np->n_flag & NLMODIFIED) { 784 np->n_attrstamp = 0; 785 error = nfs_flush(vp, MNT_WAIT, td, 0); 786 /* error = nfs_vinvalbuf(vp, V_SAVE, td, 1); */ 787 if (error) 788 return (error); 789 } 790 } 791 792 /* 793 * If IO_APPEND then load uio_offset. We restart here if we cannot 794 * get the append lock. 795 */ 796 restart: 797 if (ioflag & IO_APPEND) { 798 np->n_attrstamp = 0; 799 error = VOP_GETATTR(vp, &vattr, td); 800 if (error) 801 return (error); 802 uio->uio_offset = np->n_size; 803 } 804 805 if (uio->uio_offset < 0) 806 return (EINVAL); 807 if ((uio->uio_offset + uio->uio_resid) > nmp->nm_maxfilesize) 808 return (EFBIG); 809 if (uio->uio_resid == 0) 810 return (0); 811 812 /* 813 * We need to obtain the rslock if we intend to modify np->n_size 814 * in order to guarentee the append point with multiple contending 815 * writers, to guarentee that no other appenders modify n_size 816 * while we are trying to obtain a truncated buffer (i.e. to avoid 817 * accidently truncating data written by another appender due to 818 * the race), and to ensure that the buffer is populated prior to 819 * our extending of the file. We hold rslock through the entire 820 * operation. 821 * 822 * Note that we do not synchronize the case where someone truncates 823 * the file while we are appending to it because attempting to lock 824 * this case may deadlock other parts of the system unexpectedly. 825 */ 826 if ((ioflag & IO_APPEND) || 827 uio->uio_offset + uio->uio_resid > np->n_size) { 828 switch(nfs_rslock(np, td)) { 829 case ENOLCK: 830 goto restart; 831 /* not reached */ 832 case EINTR: 833 case ERESTART: 834 return(EINTR); 835 /* not reached */ 836 default: 837 break; 838 } 839 haverslock = 1; 840 } 841 842 /* 843 * Maybe this should be above the vnode op call, but so long as 844 * file servers have no limits, i don't think it matters 845 */ 846 if (td->td_proc && uio->uio_offset + uio->uio_resid > 847 td->td_proc->p_rlimit[RLIMIT_FSIZE].rlim_cur) { 848 psignal(td->td_proc, SIGXFSZ); 849 if (haverslock) 850 nfs_rsunlock(np, td); 851 return (EFBIG); 852 } 853 854 biosize = vp->v_mount->mnt_stat.f_iosize; 855 856 do { 857 if ((np->n_flag & NDONTCACHE) && uio->uio_iovcnt == 1) { 858 iomode = NFSV3WRITE_FILESYNC; 859 error = nfs_writerpc(vp, uio, &iomode, &must_commit); 860 if (must_commit) 861 nfs_clearcommit(vp->v_mount); 862 break; 863 } 864 nfsstats.biocache_writes++; 865 lbn = uio->uio_offset / biosize; 866 on = uio->uio_offset & (biosize-1); 867 loffset = uio->uio_offset - on; 868 n = min((unsigned)(biosize - on), uio->uio_resid); 869 again: 870 /* 871 * Handle direct append and file extension cases, calculate 872 * unaligned buffer size. 873 */ 874 875 if (uio->uio_offset == np->n_size && n) { 876 /* 877 * Get the buffer (in its pre-append state to maintain 878 * B_CACHE if it was previously set). Resize the 879 * nfsnode after we have locked the buffer to prevent 880 * readers from reading garbage. 881 */ 882 bcount = on; 883 bp = nfs_getcacheblk(vp, loffset, bcount, td); 884 885 if (bp != NULL) { 886 long save; 887 888 np->n_size = uio->uio_offset + n; 889 np->n_flag |= NLMODIFIED; 890 vnode_pager_setsize(vp, np->n_size); 891 892 save = bp->b_flags & B_CACHE; 893 bcount += n; 894 allocbuf(bp, bcount); 895 bp->b_flags |= save; 896 } 897 } else { 898 /* 899 * Obtain the locked cache block first, and then 900 * adjust the file's size as appropriate. 901 */ 902 bcount = on + n; 903 if (loffset + bcount < np->n_size) { 904 if (loffset + biosize < np->n_size) 905 bcount = biosize; 906 else 907 bcount = np->n_size - loffset; 908 } 909 bp = nfs_getcacheblk(vp, loffset, bcount, td); 910 if (uio->uio_offset + n > np->n_size) { 911 np->n_size = uio->uio_offset + n; 912 np->n_flag |= NLMODIFIED; 913 vnode_pager_setsize(vp, np->n_size); 914 } 915 } 916 917 if (bp == NULL) { 918 error = EINTR; 919 break; 920 } 921 922 /* 923 * Issue a READ if B_CACHE is not set. In special-append 924 * mode, B_CACHE is based on the buffer prior to the write 925 * op and is typically set, avoiding the read. If a read 926 * is required in special append mode, the server will 927 * probably send us a short-read since we extended the file 928 * on our end, resulting in b_resid == 0 and, thusly, 929 * B_CACHE getting set. 930 * 931 * We can also avoid issuing the read if the write covers 932 * the entire buffer. We have to make sure the buffer state 933 * is reasonable in this case since we will not be initiating 934 * I/O. See the comments in kern/vfs_bio.c's getblk() for 935 * more information. 936 * 937 * B_CACHE may also be set due to the buffer being cached 938 * normally. 939 */ 940 941 if (on == 0 && n == bcount) { 942 bp->b_flags |= B_CACHE; 943 bp->b_flags &= ~(B_ERROR | B_INVAL); 944 } 945 946 if ((bp->b_flags & B_CACHE) == 0) { 947 bp->b_flags |= B_READ; 948 vfs_busy_pages(bp, 0); 949 error = nfs_doio(vp, &bp->b_bio2, td); 950 if (error) { 951 brelse(bp); 952 break; 953 } 954 } 955 if (!bp) { 956 error = EINTR; 957 break; 958 } 959 np->n_flag |= NLMODIFIED; 960 961 /* 962 * If dirtyend exceeds file size, chop it down. This should 963 * not normally occur but there is an append race where it 964 * might occur XXX, so we log it. 965 * 966 * If the chopping creates a reverse-indexed or degenerate 967 * situation with dirtyoff/end, we 0 both of them. 968 */ 969 970 if (bp->b_dirtyend > bcount) { 971 printf("NFS append race @%08llx:%d\n", 972 bp->b_bio2.bio_offset, 973 bp->b_dirtyend - bcount); 974 bp->b_dirtyend = bcount; 975 } 976 977 if (bp->b_dirtyoff >= bp->b_dirtyend) 978 bp->b_dirtyoff = bp->b_dirtyend = 0; 979 980 /* 981 * If the new write will leave a contiguous dirty 982 * area, just update the b_dirtyoff and b_dirtyend, 983 * otherwise force a write rpc of the old dirty area. 984 * 985 * While it is possible to merge discontiguous writes due to 986 * our having a B_CACHE buffer ( and thus valid read data 987 * for the hole), we don't because it could lead to 988 * significant cache coherency problems with multiple clients, 989 * especially if locking is implemented later on. 990 * 991 * as an optimization we could theoretically maintain 992 * a linked list of discontinuous areas, but we would still 993 * have to commit them separately so there isn't much 994 * advantage to it except perhaps a bit of asynchronization. 995 */ 996 997 if (bp->b_dirtyend > 0 && 998 (on > bp->b_dirtyend || (on + n) < bp->b_dirtyoff)) { 999 if (VOP_BWRITE(vp, bp) == EINTR) { 1000 error = EINTR; 1001 break; 1002 } 1003 goto again; 1004 } 1005 1006 error = uiomove((char *)bp->b_data + on, n, uio); 1007 1008 /* 1009 * Since this block is being modified, it must be written 1010 * again and not just committed. Since write clustering does 1011 * not work for the stage 1 data write, only the stage 2 1012 * commit rpc, we have to clear B_CLUSTEROK as well. 1013 */ 1014 bp->b_flags &= ~(B_NEEDCOMMIT | B_CLUSTEROK); 1015 1016 if (error) { 1017 bp->b_flags |= B_ERROR; 1018 brelse(bp); 1019 break; 1020 } 1021 1022 /* 1023 * Only update dirtyoff/dirtyend if not a degenerate 1024 * condition. 1025 */ 1026 if (n) { 1027 if (bp->b_dirtyend > 0) { 1028 bp->b_dirtyoff = min(on, bp->b_dirtyoff); 1029 bp->b_dirtyend = max((on + n), bp->b_dirtyend); 1030 } else { 1031 bp->b_dirtyoff = on; 1032 bp->b_dirtyend = on + n; 1033 } 1034 vfs_bio_set_validclean(bp, on, n); 1035 } 1036 /* 1037 * If IO_NOWDRAIN then set B_NOWDRAIN (e.g. nfs-backed VN 1038 * filesystem). XXX also use for loopback NFS mounts. 1039 */ 1040 if (ioflag & IO_NOWDRAIN) 1041 bp->b_flags |= B_NOWDRAIN; 1042 1043 /* 1044 * If the lease is non-cachable or IO_SYNC do bwrite(). 1045 * 1046 * IO_INVAL appears to be unused. The idea appears to be 1047 * to turn off caching in this case. Very odd. XXX 1048 */ 1049 if ((np->n_flag & NDONTCACHE) || (ioflag & IO_SYNC)) { 1050 if (ioflag & IO_INVAL) 1051 bp->b_flags |= B_NOCACHE; 1052 error = VOP_BWRITE(vp, bp); 1053 if (error) 1054 break; 1055 if (np->n_flag & NDONTCACHE) { 1056 error = nfs_vinvalbuf(vp, V_SAVE, td, 1); 1057 if (error) 1058 break; 1059 } 1060 } else if ((n + on) == biosize) { 1061 bp->b_flags |= B_ASYNC; 1062 nfs_writebp(bp, 0, 0); 1063 } else { 1064 bdwrite(bp); 1065 } 1066 } while (uio->uio_resid > 0 && n > 0); 1067 1068 if (haverslock) 1069 nfs_rsunlock(np, td); 1070 1071 return (error); 1072 } 1073 1074 /* 1075 * Get an nfs cache block. 1076 * 1077 * Allocate a new one if the block isn't currently in the cache 1078 * and return the block marked busy. If the calling process is 1079 * interrupted by a signal for an interruptible mount point, return 1080 * NULL. 1081 * 1082 * The caller must carefully deal with the possible B_INVAL state of 1083 * the buffer. nfs_doio() clears B_INVAL (and nfs_asyncio() clears it 1084 * indirectly), so synchronous reads can be issued without worrying about 1085 * the B_INVAL state. We have to be a little more careful when dealing 1086 * with writes (see comments in nfs_write()) when extending a file past 1087 * its EOF. 1088 */ 1089 static struct buf * 1090 nfs_getcacheblk(struct vnode *vp, off_t loffset, int size, struct thread *td) 1091 { 1092 struct buf *bp; 1093 struct mount *mp; 1094 struct nfsmount *nmp; 1095 1096 mp = vp->v_mount; 1097 nmp = VFSTONFS(mp); 1098 1099 if (nmp->nm_flag & NFSMNT_INT) { 1100 bp = getblk(vp, loffset, size, PCATCH, 0); 1101 while (bp == NULL) { 1102 if (nfs_sigintr(nmp, (struct nfsreq *)0, td)) 1103 return (NULL); 1104 bp = getblk(vp, loffset, size, 0, 2 * hz); 1105 } 1106 } else { 1107 bp = getblk(vp, loffset, size, 0, 0); 1108 } 1109 1110 /* 1111 * bio2, the 'device' layer. Since BIOs use 64 bit byte offsets 1112 * now, no translation is necessary. 1113 */ 1114 bp->b_bio2.bio_offset = loffset; 1115 return (bp); 1116 } 1117 1118 /* 1119 * Flush and invalidate all dirty buffers. If another process is already 1120 * doing the flush, just wait for completion. 1121 */ 1122 int 1123 nfs_vinvalbuf(struct vnode *vp, int flags, 1124 struct thread *td, int intrflg) 1125 { 1126 struct nfsnode *np = VTONFS(vp); 1127 struct nfsmount *nmp = VFSTONFS(vp->v_mount); 1128 int error = 0, slpflag, slptimeo; 1129 1130 if (vp->v_flag & VRECLAIMED) 1131 return (0); 1132 1133 if ((nmp->nm_flag & NFSMNT_INT) == 0) 1134 intrflg = 0; 1135 if (intrflg) { 1136 slpflag = PCATCH; 1137 slptimeo = 2 * hz; 1138 } else { 1139 slpflag = 0; 1140 slptimeo = 0; 1141 } 1142 /* 1143 * First wait for any other process doing a flush to complete. 1144 */ 1145 while (np->n_flag & NFLUSHINPROG) { 1146 np->n_flag |= NFLUSHWANT; 1147 error = tsleep((caddr_t)&np->n_flag, 0, "nfsvinval", slptimeo); 1148 if (error && intrflg && nfs_sigintr(nmp, (struct nfsreq *)0, td)) 1149 return (EINTR); 1150 } 1151 1152 /* 1153 * Now, flush as required. 1154 */ 1155 np->n_flag |= NFLUSHINPROG; 1156 error = vinvalbuf(vp, flags, td, slpflag, 0); 1157 while (error) { 1158 if (intrflg && nfs_sigintr(nmp, (struct nfsreq *)0, td)) { 1159 np->n_flag &= ~NFLUSHINPROG; 1160 if (np->n_flag & NFLUSHWANT) { 1161 np->n_flag &= ~NFLUSHWANT; 1162 wakeup((caddr_t)&np->n_flag); 1163 } 1164 return (EINTR); 1165 } 1166 error = vinvalbuf(vp, flags, td, 0, slptimeo); 1167 } 1168 np->n_flag &= ~(NLMODIFIED | NFLUSHINPROG); 1169 if (np->n_flag & NFLUSHWANT) { 1170 np->n_flag &= ~NFLUSHWANT; 1171 wakeup((caddr_t)&np->n_flag); 1172 } 1173 return (0); 1174 } 1175 1176 /* 1177 * Initiate asynchronous I/O. Return an error if no nfsiods are available. 1178 * This is mainly to avoid queueing async I/O requests when the nfsiods 1179 * are all hung on a dead server. 1180 * 1181 * Note: nfs_asyncio() does not clear (B_ERROR|B_INVAL) but when the bp 1182 * is eventually dequeued by the async daemon, nfs_doio() *will*. 1183 */ 1184 int 1185 nfs_asyncio(struct vnode *vp, struct bio *bio, struct thread *td) 1186 { 1187 struct buf *bp = bio->bio_buf; 1188 struct nfsmount *nmp; 1189 int i; 1190 int gotiod; 1191 int slpflag = 0; 1192 int slptimeo = 0; 1193 int error; 1194 1195 /* 1196 * If no async daemons then return EIO to force caller to run the rpc 1197 * synchronously. 1198 */ 1199 if (nfs_numasync == 0) 1200 return (EIO); 1201 1202 KKASSERT(vp->v_tag == VT_NFS); 1203 nmp = VFSTONFS(vp->v_mount); 1204 1205 /* 1206 * Commits are usually short and sweet so lets save some cpu and 1207 * leave the async daemons for more important rpc's (such as reads 1208 * and writes). 1209 */ 1210 if ((bp->b_flags & (B_READ|B_NEEDCOMMIT)) == B_NEEDCOMMIT && 1211 (nmp->nm_bioqiods > nfs_numasync / 2)) { 1212 return(EIO); 1213 } 1214 1215 again: 1216 if (nmp->nm_flag & NFSMNT_INT) 1217 slpflag = PCATCH; 1218 gotiod = FALSE; 1219 1220 /* 1221 * Find a free iod to process this request. 1222 */ 1223 for (i = 0; i < NFS_MAXASYNCDAEMON; i++) 1224 if (nfs_iodwant[i]) { 1225 /* 1226 * Found one, so wake it up and tell it which 1227 * mount to process. 1228 */ 1229 NFS_DPF(ASYNCIO, 1230 ("nfs_asyncio: waking iod %d for mount %p\n", 1231 i, nmp)); 1232 nfs_iodwant[i] = NULL; 1233 nfs_iodmount[i] = nmp; 1234 nmp->nm_bioqiods++; 1235 wakeup((caddr_t)&nfs_iodwant[i]); 1236 gotiod = TRUE; 1237 break; 1238 } 1239 1240 /* 1241 * If none are free, we may already have an iod working on this mount 1242 * point. If so, it will process our request. 1243 */ 1244 if (!gotiod) { 1245 if (nmp->nm_bioqiods > 0) { 1246 NFS_DPF(ASYNCIO, 1247 ("nfs_asyncio: %d iods are already processing mount %p\n", 1248 nmp->nm_bioqiods, nmp)); 1249 gotiod = TRUE; 1250 } 1251 } 1252 1253 /* 1254 * If we have an iod which can process the request, then queue 1255 * the buffer. 1256 */ 1257 if (gotiod) { 1258 /* 1259 * Ensure that the queue never grows too large. We still want 1260 * to asynchronize so we block rather then return EIO. 1261 */ 1262 while (nmp->nm_bioqlen >= 2*nfs_numasync) { 1263 NFS_DPF(ASYNCIO, 1264 ("nfs_asyncio: waiting for mount %p queue to drain\n", nmp)); 1265 nmp->nm_bioqwant = TRUE; 1266 error = tsleep(&nmp->nm_bioq, slpflag, 1267 "nfsaio", slptimeo); 1268 if (error) { 1269 if (nfs_sigintr(nmp, NULL, td)) 1270 return (EINTR); 1271 if (slpflag == PCATCH) { 1272 slpflag = 0; 1273 slptimeo = 2 * hz; 1274 } 1275 } 1276 /* 1277 * We might have lost our iod while sleeping, 1278 * so check and loop if nescessary. 1279 */ 1280 if (nmp->nm_bioqiods == 0) { 1281 NFS_DPF(ASYNCIO, 1282 ("nfs_asyncio: no iods after mount %p queue was drained, looping\n", nmp)); 1283 goto again; 1284 } 1285 } 1286 BUF_KERNPROC(bp); 1287 1288 /* 1289 * The passed bio's buffer is not necessary associated with 1290 * the NFS vnode it is being written to. Store the NFS vnode 1291 * in the BIO driver info. 1292 */ 1293 bio->bio_driver_info = vp; 1294 TAILQ_INSERT_TAIL(&nmp->nm_bioq, bio, bio_act); 1295 nmp->nm_bioqlen++; 1296 return (0); 1297 } 1298 1299 /* 1300 * All the iods are busy on other mounts, so return EIO to 1301 * force the caller to process the i/o synchronously. 1302 */ 1303 NFS_DPF(ASYNCIO, ("nfs_asyncio: no iods available, i/o is synchronous\n")); 1304 return (EIO); 1305 } 1306 1307 /* 1308 * Do an I/O operation to/from a cache block. This may be called 1309 * synchronously or from an nfsiod. The BIO is normalized for DEV_BSIZE. 1310 * 1311 * NOTE! TD MIGHT BE NULL 1312 */ 1313 int 1314 nfs_doio(struct vnode *vp, struct bio *bio, struct thread *td) 1315 { 1316 struct buf *bp = bio->bio_buf; 1317 struct uio *uiop; 1318 struct nfsnode *np; 1319 struct nfsmount *nmp; 1320 int error = 0, iomode, must_commit = 0; 1321 struct uio uio; 1322 struct iovec io; 1323 1324 KKASSERT(vp->v_tag == VT_NFS); 1325 np = VTONFS(vp); 1326 nmp = VFSTONFS(vp->v_mount); 1327 uiop = &uio; 1328 uiop->uio_iov = &io; 1329 uiop->uio_iovcnt = 1; 1330 uiop->uio_segflg = UIO_SYSSPACE; 1331 uiop->uio_td = td; 1332 1333 /* 1334 * clear B_ERROR and B_INVAL state prior to initiating the I/O. We 1335 * do this here so we do not have to do it in all the code that 1336 * calls us. 1337 */ 1338 bp->b_flags &= ~(B_ERROR | B_INVAL); 1339 1340 KASSERT(!(bp->b_flags & B_DONE), ("nfs_doio: bp %p already marked done", bp)); 1341 1342 /* 1343 * Historically, paging was done with physio, but no more. 1344 */ 1345 if (bp->b_flags & B_PHYS) { 1346 /* 1347 * ...though reading /dev/drum still gets us here. 1348 */ 1349 io.iov_len = uiop->uio_resid = bp->b_bcount; 1350 /* mapping was done by vmapbuf() */ 1351 io.iov_base = bp->b_data; 1352 uiop->uio_offset = bio->bio_offset; 1353 if (bp->b_flags & B_READ) { 1354 uiop->uio_rw = UIO_READ; 1355 nfsstats.read_physios++; 1356 error = nfs_readrpc(vp, uiop); 1357 } else { 1358 int com; 1359 1360 iomode = NFSV3WRITE_DATASYNC; 1361 uiop->uio_rw = UIO_WRITE; 1362 nfsstats.write_physios++; 1363 error = nfs_writerpc(vp, uiop, &iomode, &com); 1364 } 1365 if (error) { 1366 bp->b_flags |= B_ERROR; 1367 bp->b_error = error; 1368 } 1369 } else if (bp->b_flags & B_READ) { 1370 io.iov_len = uiop->uio_resid = bp->b_bcount; 1371 io.iov_base = bp->b_data; 1372 uiop->uio_rw = UIO_READ; 1373 1374 switch (vp->v_type) { 1375 case VREG: 1376 uiop->uio_offset = bio->bio_offset; 1377 nfsstats.read_bios++; 1378 error = nfs_readrpc(vp, uiop); 1379 1380 if (!error) { 1381 if (uiop->uio_resid) { 1382 /* 1383 * If we had a short read with no error, we must have 1384 * hit a file hole. We should zero-fill the remainder. 1385 * This can also occur if the server hits the file EOF. 1386 * 1387 * Holes used to be able to occur due to pending 1388 * writes, but that is not possible any longer. 1389 */ 1390 int nread = bp->b_bcount - uiop->uio_resid; 1391 int left = uiop->uio_resid; 1392 1393 if (left > 0) 1394 bzero((char *)bp->b_data + nread, left); 1395 uiop->uio_resid = 0; 1396 } 1397 } 1398 if (td && td->td_proc && (vp->v_flag & VTEXT) && 1399 np->n_mtime != np->n_vattr.va_mtime.tv_sec) { 1400 uprintf("Process killed due to text file modification\n"); 1401 psignal(td->td_proc, SIGKILL); 1402 } 1403 break; 1404 case VLNK: 1405 uiop->uio_offset = 0; 1406 nfsstats.readlink_bios++; 1407 error = nfs_readlinkrpc(vp, uiop); 1408 break; 1409 case VDIR: 1410 nfsstats.readdir_bios++; 1411 uiop->uio_offset = bio->bio_offset; 1412 if (nmp->nm_flag & NFSMNT_RDIRPLUS) { 1413 error = nfs_readdirplusrpc(vp, uiop); 1414 if (error == NFSERR_NOTSUPP) 1415 nmp->nm_flag &= ~NFSMNT_RDIRPLUS; 1416 } 1417 if ((nmp->nm_flag & NFSMNT_RDIRPLUS) == 0) 1418 error = nfs_readdirrpc(vp, uiop); 1419 /* 1420 * end-of-directory sets B_INVAL but does not generate an 1421 * error. 1422 */ 1423 if (error == 0 && uiop->uio_resid == bp->b_bcount) 1424 bp->b_flags |= B_INVAL; 1425 break; 1426 default: 1427 printf("nfs_doio: type %x unexpected\n",vp->v_type); 1428 break; 1429 }; 1430 if (error) { 1431 bp->b_flags |= B_ERROR; 1432 bp->b_error = error; 1433 } 1434 } else { 1435 /* 1436 * If we only need to commit, try to commit 1437 */ 1438 if (bp->b_flags & B_NEEDCOMMIT) { 1439 int retv; 1440 off_t off; 1441 1442 off = bio->bio_offset + bp->b_dirtyoff; 1443 retv = nfs_commit(vp, off, 1444 bp->b_dirtyend - bp->b_dirtyoff, td); 1445 if (retv == 0) { 1446 bp->b_dirtyoff = bp->b_dirtyend = 0; 1447 bp->b_flags &= ~(B_NEEDCOMMIT | B_CLUSTEROK); 1448 bp->b_resid = 0; 1449 biodone(bio); 1450 return (0); 1451 } 1452 if (retv == NFSERR_STALEWRITEVERF) { 1453 nfs_clearcommit(vp->v_mount); 1454 } 1455 } 1456 1457 /* 1458 * Setup for actual write 1459 */ 1460 1461 if (bio->bio_offset + bp->b_dirtyend > np->n_size) 1462 bp->b_dirtyend = np->n_size - bio->bio_offset; 1463 1464 if (bp->b_dirtyend > bp->b_dirtyoff) { 1465 io.iov_len = uiop->uio_resid = bp->b_dirtyend 1466 - bp->b_dirtyoff; 1467 uiop->uio_offset = bio->bio_offset + bp->b_dirtyoff; 1468 io.iov_base = (char *)bp->b_data + bp->b_dirtyoff; 1469 uiop->uio_rw = UIO_WRITE; 1470 nfsstats.write_bios++; 1471 1472 if ((bp->b_flags & (B_ASYNC | B_NEEDCOMMIT | B_NOCACHE | B_CLUSTER)) == B_ASYNC) 1473 iomode = NFSV3WRITE_UNSTABLE; 1474 else 1475 iomode = NFSV3WRITE_FILESYNC; 1476 1477 error = nfs_writerpc(vp, uiop, &iomode, &must_commit); 1478 1479 /* 1480 * When setting B_NEEDCOMMIT also set B_CLUSTEROK to try 1481 * to cluster the buffers needing commit. This will allow 1482 * the system to submit a single commit rpc for the whole 1483 * cluster. We can do this even if the buffer is not 100% 1484 * dirty (relative to the NFS blocksize), so we optimize the 1485 * append-to-file-case. 1486 * 1487 * (when clearing B_NEEDCOMMIT, B_CLUSTEROK must also be 1488 * cleared because write clustering only works for commit 1489 * rpc's, not for the data portion of the write). 1490 */ 1491 1492 if (!error && iomode == NFSV3WRITE_UNSTABLE) { 1493 bp->b_flags |= B_NEEDCOMMIT; 1494 if (bp->b_dirtyoff == 0 1495 && bp->b_dirtyend == bp->b_bcount) 1496 bp->b_flags |= B_CLUSTEROK; 1497 } else { 1498 bp->b_flags &= ~(B_NEEDCOMMIT | B_CLUSTEROK); 1499 } 1500 1501 /* 1502 * For an interrupted write, the buffer is still valid 1503 * and the write hasn't been pushed to the server yet, 1504 * so we can't set B_ERROR and report the interruption 1505 * by setting B_EINTR. For the B_ASYNC case, B_EINTR 1506 * is not relevant, so the rpc attempt is essentially 1507 * a noop. For the case of a V3 write rpc not being 1508 * committed to stable storage, the block is still 1509 * dirty and requires either a commit rpc or another 1510 * write rpc with iomode == NFSV3WRITE_FILESYNC before 1511 * the block is reused. This is indicated by setting 1512 * the B_DELWRI and B_NEEDCOMMIT flags. 1513 * 1514 * If the buffer is marked B_PAGING, it does not reside on 1515 * the vp's paging queues so we cannot call bdirty(). The 1516 * bp in this case is not an NFS cache block so we should 1517 * be safe. XXX 1518 */ 1519 if (error == EINTR 1520 || (!error && (bp->b_flags & B_NEEDCOMMIT))) { 1521 crit_enter(); 1522 bp->b_flags &= ~(B_INVAL|B_NOCACHE); 1523 if ((bp->b_flags & B_PAGING) == 0) { 1524 bdirty(bp); 1525 bp->b_flags &= ~B_DONE; 1526 } 1527 if (error && (bp->b_flags & B_ASYNC) == 0) 1528 bp->b_flags |= B_EINTR; 1529 crit_exit(); 1530 } else { 1531 if (error) { 1532 bp->b_flags |= B_ERROR; 1533 bp->b_error = np->n_error = error; 1534 np->n_flag |= NWRITEERR; 1535 } 1536 bp->b_dirtyoff = bp->b_dirtyend = 0; 1537 } 1538 } else { 1539 bp->b_resid = 0; 1540 biodone(bio); 1541 return (0); 1542 } 1543 } 1544 bp->b_resid = uiop->uio_resid; 1545 if (must_commit) 1546 nfs_clearcommit(vp->v_mount); 1547 biodone(bio); 1548 return (error); 1549 } 1550 1551 /* 1552 * Used to aid in handling ftruncate() operations on the NFS client side. 1553 * Truncation creates a number of special problems for NFS. We have to 1554 * throw away VM pages and buffer cache buffers that are beyond EOF, and 1555 * we have to properly handle VM pages or (potentially dirty) buffers 1556 * that straddle the truncation point. 1557 */ 1558 1559 int 1560 nfs_meta_setsize(struct vnode *vp, struct thread *td, u_quad_t nsize) 1561 { 1562 struct nfsnode *np = VTONFS(vp); 1563 u_quad_t tsize = np->n_size; 1564 int biosize = vp->v_mount->mnt_stat.f_iosize; 1565 int error = 0; 1566 1567 np->n_size = nsize; 1568 1569 if (np->n_size < tsize) { 1570 struct buf *bp; 1571 daddr_t lbn; 1572 off_t loffset; 1573 int bufsize; 1574 1575 /* 1576 * vtruncbuf() doesn't get the buffer overlapping the 1577 * truncation point. We may have a B_DELWRI and/or B_CACHE 1578 * buffer that now needs to be truncated. 1579 */ 1580 error = vtruncbuf(vp, td, nsize, biosize); 1581 lbn = nsize / biosize; 1582 bufsize = nsize & (biosize - 1); 1583 loffset = nsize - bufsize; 1584 bp = nfs_getcacheblk(vp, loffset, bufsize, td); 1585 if (bp->b_dirtyoff > bp->b_bcount) 1586 bp->b_dirtyoff = bp->b_bcount; 1587 if (bp->b_dirtyend > bp->b_bcount) 1588 bp->b_dirtyend = bp->b_bcount; 1589 bp->b_flags |= B_RELBUF; /* don't leave garbage around */ 1590 brelse(bp); 1591 } else { 1592 vnode_pager_setsize(vp, nsize); 1593 } 1594 return(error); 1595 } 1596 1597