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 1.30 2006/03/27 16:18:39 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 * NQNFS: Full cache coherency is maintained within the loop. 373 * 374 * NFS: If local changes have been made and this is a 375 * directory, the directory must be invalidated and 376 * the attribute cache must be cleared. 377 * 378 * GETATTR is called to synchronize the file size. 379 * 380 * If remote changes are detected local data is flushed 381 * and the cache is invalidated. 382 * 383 * 384 * NOTE: In the normal case the attribute cache is not 385 * cleared which means GETATTR may use cached data and 386 * not immediately detect changes made on the server. 387 */ 388 if ((np->n_flag & NLMODIFIED) && vp->v_type == VDIR) { 389 nfs_invaldir(vp); 390 error = nfs_vinvalbuf(vp, V_SAVE, td, 1); 391 if (error) 392 return (error); 393 np->n_attrstamp = 0; 394 } 395 error = VOP_GETATTR(vp, &vattr, td); 396 if (error) 397 return (error); 398 if (np->n_flag & NRMODIFIED) { 399 if (vp->v_type == VDIR) 400 nfs_invaldir(vp); 401 error = nfs_vinvalbuf(vp, V_SAVE, td, 1); 402 if (error) 403 return (error); 404 np->n_flag &= ~NRMODIFIED; 405 } 406 do { 407 if (np->n_flag & NDONTCACHE) { 408 switch (vp->v_type) { 409 case VREG: 410 return (nfs_readrpc(vp, uio)); 411 case VLNK: 412 return (nfs_readlinkrpc(vp, uio)); 413 case VDIR: 414 break; 415 default: 416 printf(" NDONTCACHE: type %x unexpected\n", vp->v_type); 417 break; 418 }; 419 } 420 switch (vp->v_type) { 421 case VREG: 422 nfsstats.biocache_reads++; 423 lbn = uio->uio_offset / biosize; 424 on = uio->uio_offset & (biosize - 1); 425 loffset = (off_t)lbn * biosize; 426 427 /* 428 * Start the read ahead(s), as required. 429 */ 430 if (nfs_numasync > 0 && nmp->nm_readahead > 0) { 431 for (nra = 0; nra < nmp->nm_readahead && nra < seqcount && 432 (off_t)(lbn + 1 + nra) * biosize < np->n_size; nra++) { 433 rabn = lbn + 1 + nra; 434 raoffset = (off_t)rabn * biosize; 435 if (!findblk(vp, raoffset)) { 436 rabp = nfs_getcacheblk(vp, raoffset, biosize, td); 437 if (!rabp) 438 return (EINTR); 439 if ((rabp->b_flags & (B_CACHE|B_DELWRI)) == 0) { 440 rabp->b_flags |= (B_READ | B_ASYNC); 441 vfs_busy_pages(rabp, 0); 442 if (nfs_asyncio(vp, &rabp->b_bio2, td)) { 443 rabp->b_flags |= B_INVAL|B_ERROR; 444 vfs_unbusy_pages(rabp); 445 brelse(rabp); 446 break; 447 } 448 } else { 449 brelse(rabp); 450 } 451 } 452 } 453 } 454 455 /* 456 * Obtain the buffer cache block. Figure out the buffer size 457 * when we are at EOF. If we are modifying the size of the 458 * buffer based on an EOF condition we need to hold 459 * nfs_rslock() through obtaining the buffer to prevent 460 * a potential writer-appender from messing with n_size. 461 * Otherwise we may accidently truncate the buffer and 462 * lose dirty data. 463 * 464 * Note that bcount is *not* DEV_BSIZE aligned. 465 */ 466 467 again: 468 bcount = biosize; 469 if (loffset >= np->n_size) { 470 bcount = 0; 471 } else if (loffset + biosize > np->n_size) { 472 bcount = np->n_size - loffset; 473 } 474 if (bcount != biosize) { 475 switch(nfs_rslock(np, td)) { 476 case ENOLCK: 477 goto again; 478 /* not reached */ 479 case EINTR: 480 case ERESTART: 481 return(EINTR); 482 /* not reached */ 483 default: 484 break; 485 } 486 } 487 488 bp = nfs_getcacheblk(vp, loffset, bcount, td); 489 490 if (bcount != biosize) 491 nfs_rsunlock(np, td); 492 if (!bp) 493 return (EINTR); 494 495 /* 496 * If B_CACHE is not set, we must issue the read. If this 497 * fails, we return an error. 498 */ 499 500 if ((bp->b_flags & B_CACHE) == 0) { 501 bp->b_flags |= B_READ; 502 vfs_busy_pages(bp, 0); 503 error = nfs_doio(vp, &bp->b_bio2, td); 504 if (error) { 505 brelse(bp); 506 return (error); 507 } 508 } 509 510 /* 511 * on is the offset into the current bp. Figure out how many 512 * bytes we can copy out of the bp. Note that bcount is 513 * NOT DEV_BSIZE aligned. 514 * 515 * Then figure out how many bytes we can copy into the uio. 516 */ 517 518 n = 0; 519 if (on < bcount) 520 n = min((unsigned)(bcount - on), uio->uio_resid); 521 break; 522 case VLNK: 523 nfsstats.biocache_readlinks++; 524 bp = nfs_getcacheblk(vp, (off_t)0, NFS_MAXPATHLEN, td); 525 if (bp == NULL) 526 return (EINTR); 527 if ((bp->b_flags & B_CACHE) == 0) { 528 bp->b_flags |= B_READ; 529 vfs_busy_pages(bp, 0); 530 error = nfs_doio(vp, &bp->b_bio2, td); 531 if (error) { 532 bp->b_flags |= B_ERROR; 533 brelse(bp); 534 return (error); 535 } 536 } 537 n = min(uio->uio_resid, NFS_MAXPATHLEN - bp->b_resid); 538 on = 0; 539 break; 540 case VDIR: 541 nfsstats.biocache_readdirs++; 542 if (np->n_direofoffset 543 && uio->uio_offset >= np->n_direofoffset) { 544 return (0); 545 } 546 lbn = (uoff_t)uio->uio_offset / NFS_DIRBLKSIZ; 547 on = uio->uio_offset & (NFS_DIRBLKSIZ - 1); 548 loffset = uio->uio_offset - on; 549 bp = nfs_getcacheblk(vp, loffset, NFS_DIRBLKSIZ, td); 550 if (bp == NULL) 551 return (EINTR); 552 553 if ((bp->b_flags & B_CACHE) == 0) { 554 bp->b_flags |= B_READ; 555 vfs_busy_pages(bp, 0); 556 error = nfs_doio(vp, &bp->b_bio2, td); 557 if (error) { 558 brelse(bp); 559 } 560 while (error == NFSERR_BAD_COOKIE) { 561 printf("got bad cookie vp %p bp %p\n", vp, bp); 562 nfs_invaldir(vp); 563 error = nfs_vinvalbuf(vp, 0, td, 1); 564 /* 565 * Yuck! The directory has been modified on the 566 * server. The only way to get the block is by 567 * reading from the beginning to get all the 568 * offset cookies. 569 * 570 * Leave the last bp intact unless there is an error. 571 * Loop back up to the while if the error is another 572 * NFSERR_BAD_COOKIE (double yuch!). 573 */ 574 for (i = 0; i <= lbn && !error; i++) { 575 if (np->n_direofoffset 576 && (i * NFS_DIRBLKSIZ) >= np->n_direofoffset) 577 return (0); 578 bp = nfs_getcacheblk(vp, (off_t)i * NFS_DIRBLKSIZ, 579 NFS_DIRBLKSIZ, td); 580 if (!bp) 581 return (EINTR); 582 if ((bp->b_flags & B_CACHE) == 0) { 583 bp->b_flags |= B_READ; 584 vfs_busy_pages(bp, 0); 585 error = nfs_doio(vp, &bp->b_bio2, td); 586 /* 587 * no error + B_INVAL == directory EOF, 588 * use the block. 589 */ 590 if (error == 0 && (bp->b_flags & B_INVAL)) 591 break; 592 } 593 /* 594 * An error will throw away the block and the 595 * for loop will break out. If no error and this 596 * is not the block we want, we throw away the 597 * block and go for the next one via the for loop. 598 */ 599 if (error || i < lbn) 600 brelse(bp); 601 } 602 } 603 /* 604 * The above while is repeated if we hit another cookie 605 * error. If we hit an error and it wasn't a cookie error, 606 * we give up. 607 */ 608 if (error) 609 return (error); 610 } 611 612 /* 613 * If not eof and read aheads are enabled, start one. 614 * (You need the current block first, so that you have the 615 * directory offset cookie of the next block.) 616 */ 617 if (nfs_numasync > 0 && nmp->nm_readahead > 0 && 618 (bp->b_flags & B_INVAL) == 0 && 619 (np->n_direofoffset == 0 || 620 loffset + NFS_DIRBLKSIZ < np->n_direofoffset) && 621 (np->n_flag & NDONTCACHE) == 0 && 622 !findblk(vp, loffset + NFS_DIRBLKSIZ)) { 623 rabp = nfs_getcacheblk(vp, loffset + NFS_DIRBLKSIZ, 624 NFS_DIRBLKSIZ, td); 625 if (rabp) { 626 if ((rabp->b_flags & (B_CACHE|B_DELWRI)) == 0) { 627 rabp->b_flags |= (B_READ | B_ASYNC); 628 vfs_busy_pages(rabp, 0); 629 if (nfs_asyncio(vp, &rabp->b_bio2, td)) { 630 rabp->b_flags |= B_INVAL|B_ERROR; 631 vfs_unbusy_pages(rabp); 632 brelse(rabp); 633 } 634 } else { 635 brelse(rabp); 636 } 637 } 638 } 639 /* 640 * Unlike VREG files, whos buffer size ( bp->b_bcount ) is 641 * chopped for the EOF condition, we cannot tell how large 642 * NFS directories are going to be until we hit EOF. So 643 * an NFS directory buffer is *not* chopped to its EOF. Now, 644 * it just so happens that b_resid will effectively chop it 645 * to EOF. *BUT* this information is lost if the buffer goes 646 * away and is reconstituted into a B_CACHE state ( due to 647 * being VMIO ) later. So we keep track of the directory eof 648 * in np->n_direofoffset and chop it off as an extra step 649 * right here. 650 */ 651 n = lmin(uio->uio_resid, NFS_DIRBLKSIZ - bp->b_resid - on); 652 if (np->n_direofoffset && n > np->n_direofoffset - uio->uio_offset) 653 n = np->n_direofoffset - uio->uio_offset; 654 break; 655 default: 656 printf(" nfs_bioread: type %x unexpected\n",vp->v_type); 657 break; 658 }; 659 660 switch (vp->v_type) { 661 case VREG: 662 if (n > 0) 663 error = uiomove(bp->b_data + on, (int)n, uio); 664 break; 665 case VLNK: 666 if (n > 0) 667 error = uiomove(bp->b_data + on, (int)n, uio); 668 n = 0; 669 break; 670 case VDIR: 671 if (n > 0) { 672 off_t old_off = uio->uio_offset; 673 caddr_t cpos, epos; 674 struct nfs_dirent *dp; 675 676 /* 677 * We are casting cpos to nfs_dirent, it must be 678 * int-aligned. 679 */ 680 if (on & 3) { 681 error = EINVAL; 682 break; 683 } 684 685 cpos = bp->b_data + on; 686 epos = bp->b_data + on + n; 687 while (cpos < epos && error == 0 && uio->uio_resid > 0) { 688 dp = (struct nfs_dirent *)cpos; 689 error = nfs_check_dirent(dp, (int)(epos - cpos)); 690 if (error) 691 break; 692 if (vop_write_dirent(&error, uio, dp->nfs_ino, 693 dp->nfs_type, dp->nfs_namlen, dp->nfs_name)) { 694 break; 695 } 696 cpos += dp->nfs_reclen; 697 } 698 n = 0; 699 if (error == 0) 700 uio->uio_offset = old_off + cpos - bp->b_data - on; 701 } 702 /* 703 * Invalidate buffer if caching is disabled, forcing a 704 * re-read from the remote later. 705 */ 706 if (np->n_flag & NDONTCACHE) 707 bp->b_flags |= B_INVAL; 708 break; 709 default: 710 printf(" nfs_bioread: type %x unexpected\n",vp->v_type); 711 } 712 brelse(bp); 713 } while (error == 0 && uio->uio_resid > 0 && n > 0); 714 return (error); 715 } 716 717 /* 718 * Userland can supply any 'seek' offset when reading a NFS directory. 719 * Validate the structure so we don't panic the kernel. Note that 720 * the element name is nul terminated and the nul is not included 721 * in nfs_namlen. 722 */ 723 static 724 int 725 nfs_check_dirent(struct nfs_dirent *dp, int maxlen) 726 { 727 int nfs_name_off = offsetof(struct nfs_dirent, nfs_name[0]); 728 729 if (nfs_name_off >= maxlen) 730 return (EINVAL); 731 if (dp->nfs_reclen < nfs_name_off || dp->nfs_reclen > maxlen) 732 return (EINVAL); 733 if (nfs_name_off + dp->nfs_namlen >= dp->nfs_reclen) 734 return (EINVAL); 735 if (dp->nfs_reclen & 3) 736 return (EINVAL); 737 return (0); 738 } 739 740 /* 741 * Vnode op for write using bio 742 * 743 * nfs_write(struct vnode *a_vp, struct uio *a_uio, int a_ioflag, 744 * struct ucred *a_cred) 745 */ 746 int 747 nfs_write(struct vop_write_args *ap) 748 { 749 struct uio *uio = ap->a_uio; 750 struct thread *td = uio->uio_td; 751 struct vnode *vp = ap->a_vp; 752 struct nfsnode *np = VTONFS(vp); 753 int ioflag = ap->a_ioflag; 754 struct buf *bp; 755 struct vattr vattr; 756 struct nfsmount *nmp = VFSTONFS(vp->v_mount); 757 daddr_t lbn; 758 off_t loffset; 759 int n, on, error = 0, iomode, must_commit; 760 int haverslock = 0; 761 int bcount; 762 int biosize; 763 764 #ifdef DIAGNOSTIC 765 if (uio->uio_rw != UIO_WRITE) 766 panic("nfs_write mode"); 767 if (uio->uio_segflg == UIO_USERSPACE && uio->uio_td != curthread) 768 panic("nfs_write proc"); 769 #endif 770 if (vp->v_type != VREG) 771 return (EIO); 772 if (np->n_flag & NWRITEERR) { 773 np->n_flag &= ~NWRITEERR; 774 return (np->n_error); 775 } 776 if ((nmp->nm_flag & NFSMNT_NFSV3) != 0 && 777 (nmp->nm_state & NFSSTA_GOTFSINFO) == 0) 778 (void)nfs_fsinfo(nmp, vp, td); 779 780 /* 781 * Synchronously flush pending buffers if we are in synchronous 782 * mode or if we are appending. 783 */ 784 if (ioflag & (IO_APPEND | IO_SYNC)) { 785 if (np->n_flag & NLMODIFIED) { 786 np->n_attrstamp = 0; 787 error = nfs_flush(vp, MNT_WAIT, td, 0); 788 /* error = nfs_vinvalbuf(vp, V_SAVE, td, 1); */ 789 if (error) 790 return (error); 791 } 792 } 793 794 /* 795 * If IO_APPEND then load uio_offset. We restart here if we cannot 796 * get the append lock. 797 */ 798 restart: 799 if (ioflag & IO_APPEND) { 800 np->n_attrstamp = 0; 801 error = VOP_GETATTR(vp, &vattr, td); 802 if (error) 803 return (error); 804 uio->uio_offset = np->n_size; 805 } 806 807 if (uio->uio_offset < 0) 808 return (EINVAL); 809 if ((uio->uio_offset + uio->uio_resid) > nmp->nm_maxfilesize) 810 return (EFBIG); 811 if (uio->uio_resid == 0) 812 return (0); 813 814 /* 815 * We need to obtain the rslock if we intend to modify np->n_size 816 * in order to guarentee the append point with multiple contending 817 * writers, to guarentee that no other appenders modify n_size 818 * while we are trying to obtain a truncated buffer (i.e. to avoid 819 * accidently truncating data written by another appender due to 820 * the race), and to ensure that the buffer is populated prior to 821 * our extending of the file. We hold rslock through the entire 822 * operation. 823 * 824 * Note that we do not synchronize the case where someone truncates 825 * the file while we are appending to it because attempting to lock 826 * this case may deadlock other parts of the system unexpectedly. 827 */ 828 if ((ioflag & IO_APPEND) || 829 uio->uio_offset + uio->uio_resid > np->n_size) { 830 switch(nfs_rslock(np, td)) { 831 case ENOLCK: 832 goto restart; 833 /* not reached */ 834 case EINTR: 835 case ERESTART: 836 return(EINTR); 837 /* not reached */ 838 default: 839 break; 840 } 841 haverslock = 1; 842 } 843 844 /* 845 * Maybe this should be above the vnode op call, but so long as 846 * file servers have no limits, i don't think it matters 847 */ 848 if (td->td_proc && uio->uio_offset + uio->uio_resid > 849 td->td_proc->p_rlimit[RLIMIT_FSIZE].rlim_cur) { 850 psignal(td->td_proc, SIGXFSZ); 851 if (haverslock) 852 nfs_rsunlock(np, td); 853 return (EFBIG); 854 } 855 856 biosize = vp->v_mount->mnt_stat.f_iosize; 857 858 do { 859 if ((np->n_flag & NDONTCACHE) && uio->uio_iovcnt == 1) { 860 iomode = NFSV3WRITE_FILESYNC; 861 error = nfs_writerpc(vp, uio, &iomode, &must_commit); 862 if (must_commit) 863 nfs_clearcommit(vp->v_mount); 864 break; 865 } 866 nfsstats.biocache_writes++; 867 lbn = uio->uio_offset / biosize; 868 on = uio->uio_offset & (biosize-1); 869 loffset = uio->uio_offset - on; 870 n = min((unsigned)(biosize - on), uio->uio_resid); 871 again: 872 /* 873 * Handle direct append and file extension cases, calculate 874 * unaligned buffer size. 875 */ 876 877 if (uio->uio_offset == np->n_size && n) { 878 /* 879 * Get the buffer (in its pre-append state to maintain 880 * B_CACHE if it was previously set). Resize the 881 * nfsnode after we have locked the buffer to prevent 882 * readers from reading garbage. 883 */ 884 bcount = on; 885 bp = nfs_getcacheblk(vp, loffset, bcount, td); 886 887 if (bp != NULL) { 888 long save; 889 890 np->n_size = uio->uio_offset + n; 891 np->n_flag |= NLMODIFIED; 892 vnode_pager_setsize(vp, np->n_size); 893 894 save = bp->b_flags & B_CACHE; 895 bcount += n; 896 allocbuf(bp, bcount); 897 bp->b_flags |= save; 898 } 899 } else { 900 /* 901 * Obtain the locked cache block first, and then 902 * adjust the file's size as appropriate. 903 */ 904 bcount = on + n; 905 if (loffset + bcount < np->n_size) { 906 if (loffset + biosize < np->n_size) 907 bcount = biosize; 908 else 909 bcount = np->n_size - loffset; 910 } 911 bp = nfs_getcacheblk(vp, loffset, bcount, td); 912 if (uio->uio_offset + n > np->n_size) { 913 np->n_size = uio->uio_offset + n; 914 np->n_flag |= NLMODIFIED; 915 vnode_pager_setsize(vp, np->n_size); 916 } 917 } 918 919 if (bp == NULL) { 920 error = EINTR; 921 break; 922 } 923 924 /* 925 * Issue a READ if B_CACHE is not set. In special-append 926 * mode, B_CACHE is based on the buffer prior to the write 927 * op and is typically set, avoiding the read. If a read 928 * is required in special append mode, the server will 929 * probably send us a short-read since we extended the file 930 * on our end, resulting in b_resid == 0 and, thusly, 931 * B_CACHE getting set. 932 * 933 * We can also avoid issuing the read if the write covers 934 * the entire buffer. We have to make sure the buffer state 935 * is reasonable in this case since we will not be initiating 936 * I/O. See the comments in kern/vfs_bio.c's getblk() for 937 * more information. 938 * 939 * B_CACHE may also be set due to the buffer being cached 940 * normally. 941 */ 942 943 if (on == 0 && n == bcount) { 944 bp->b_flags |= B_CACHE; 945 bp->b_flags &= ~(B_ERROR | B_INVAL); 946 } 947 948 if ((bp->b_flags & B_CACHE) == 0) { 949 bp->b_flags |= B_READ; 950 vfs_busy_pages(bp, 0); 951 error = nfs_doio(vp, &bp->b_bio2, td); 952 if (error) { 953 brelse(bp); 954 break; 955 } 956 } 957 if (!bp) { 958 error = EINTR; 959 break; 960 } 961 np->n_flag |= NLMODIFIED; 962 963 /* 964 * If dirtyend exceeds file size, chop it down. This should 965 * not normally occur but there is an append race where it 966 * might occur XXX, so we log it. 967 * 968 * If the chopping creates a reverse-indexed or degenerate 969 * situation with dirtyoff/end, we 0 both of them. 970 */ 971 972 if (bp->b_dirtyend > bcount) { 973 printf("NFS append race @%08llx:%d\n", 974 bp->b_bio2.bio_offset, 975 bp->b_dirtyend - bcount); 976 bp->b_dirtyend = bcount; 977 } 978 979 if (bp->b_dirtyoff >= bp->b_dirtyend) 980 bp->b_dirtyoff = bp->b_dirtyend = 0; 981 982 /* 983 * If the new write will leave a contiguous dirty 984 * area, just update the b_dirtyoff and b_dirtyend, 985 * otherwise force a write rpc of the old dirty area. 986 * 987 * While it is possible to merge discontiguous writes due to 988 * our having a B_CACHE buffer ( and thus valid read data 989 * for the hole), we don't because it could lead to 990 * significant cache coherency problems with multiple clients, 991 * especially if locking is implemented later on. 992 * 993 * as an optimization we could theoretically maintain 994 * a linked list of discontinuous areas, but we would still 995 * have to commit them separately so there isn't much 996 * advantage to it except perhaps a bit of asynchronization. 997 */ 998 999 if (bp->b_dirtyend > 0 && 1000 (on > bp->b_dirtyend || (on + n) < bp->b_dirtyoff)) { 1001 if (VOP_BWRITE(vp, bp) == EINTR) { 1002 error = EINTR; 1003 break; 1004 } 1005 goto again; 1006 } 1007 1008 error = uiomove((char *)bp->b_data + on, n, uio); 1009 1010 /* 1011 * Since this block is being modified, it must be written 1012 * again and not just committed. Since write clustering does 1013 * not work for the stage 1 data write, only the stage 2 1014 * commit rpc, we have to clear B_CLUSTEROK as well. 1015 */ 1016 bp->b_flags &= ~(B_NEEDCOMMIT | B_CLUSTEROK); 1017 1018 if (error) { 1019 bp->b_flags |= B_ERROR; 1020 brelse(bp); 1021 break; 1022 } 1023 1024 /* 1025 * Only update dirtyoff/dirtyend if not a degenerate 1026 * condition. 1027 */ 1028 if (n) { 1029 if (bp->b_dirtyend > 0) { 1030 bp->b_dirtyoff = min(on, bp->b_dirtyoff); 1031 bp->b_dirtyend = max((on + n), bp->b_dirtyend); 1032 } else { 1033 bp->b_dirtyoff = on; 1034 bp->b_dirtyend = on + n; 1035 } 1036 vfs_bio_set_validclean(bp, on, n); 1037 } 1038 /* 1039 * If IO_NOWDRAIN then set B_NOWDRAIN (e.g. nfs-backed VN 1040 * filesystem). XXX also use for loopback NFS mounts. 1041 */ 1042 if (ioflag & IO_NOWDRAIN) 1043 bp->b_flags |= B_NOWDRAIN; 1044 1045 /* 1046 * If the lease is non-cachable or IO_SYNC do bwrite(). 1047 * 1048 * IO_INVAL appears to be unused. The idea appears to be 1049 * to turn off caching in this case. Very odd. XXX 1050 */ 1051 if ((np->n_flag & NDONTCACHE) || (ioflag & IO_SYNC)) { 1052 if (ioflag & IO_INVAL) 1053 bp->b_flags |= B_NOCACHE; 1054 error = VOP_BWRITE(vp, bp); 1055 if (error) 1056 break; 1057 if (np->n_flag & NDONTCACHE) { 1058 error = nfs_vinvalbuf(vp, V_SAVE, td, 1); 1059 if (error) 1060 break; 1061 } 1062 } else if ((n + on) == biosize) { 1063 bp->b_flags |= B_ASYNC; 1064 nfs_writebp(bp, 0, 0); 1065 } else { 1066 bdwrite(bp); 1067 } 1068 } while (uio->uio_resid > 0 && n > 0); 1069 1070 if (haverslock) 1071 nfs_rsunlock(np, td); 1072 1073 return (error); 1074 } 1075 1076 /* 1077 * Get an nfs cache block. 1078 * 1079 * Allocate a new one if the block isn't currently in the cache 1080 * and return the block marked busy. If the calling process is 1081 * interrupted by a signal for an interruptible mount point, return 1082 * NULL. 1083 * 1084 * The caller must carefully deal with the possible B_INVAL state of 1085 * the buffer. nfs_doio() clears B_INVAL (and nfs_asyncio() clears it 1086 * indirectly), so synchronous reads can be issued without worrying about 1087 * the B_INVAL state. We have to be a little more careful when dealing 1088 * with writes (see comments in nfs_write()) when extending a file past 1089 * its EOF. 1090 */ 1091 static struct buf * 1092 nfs_getcacheblk(struct vnode *vp, off_t loffset, int size, struct thread *td) 1093 { 1094 struct buf *bp; 1095 struct mount *mp; 1096 struct nfsmount *nmp; 1097 1098 mp = vp->v_mount; 1099 nmp = VFSTONFS(mp); 1100 1101 if (nmp->nm_flag & NFSMNT_INT) { 1102 bp = getblk(vp, loffset, size, PCATCH, 0); 1103 while (bp == NULL) { 1104 if (nfs_sigintr(nmp, (struct nfsreq *)0, td)) 1105 return (NULL); 1106 bp = getblk(vp, loffset, size, 0, 2 * hz); 1107 } 1108 } else { 1109 bp = getblk(vp, loffset, size, 0, 0); 1110 } 1111 1112 /* 1113 * bio2, the 'device' layer. Since BIOs use 64 bit byte offsets 1114 * now, no translation is necessary. 1115 */ 1116 bp->b_bio2.bio_offset = loffset; 1117 return (bp); 1118 } 1119 1120 /* 1121 * Flush and invalidate all dirty buffers. If another process is already 1122 * doing the flush, just wait for completion. 1123 */ 1124 int 1125 nfs_vinvalbuf(struct vnode *vp, int flags, 1126 struct thread *td, int intrflg) 1127 { 1128 struct nfsnode *np = VTONFS(vp); 1129 struct nfsmount *nmp = VFSTONFS(vp->v_mount); 1130 int error = 0, slpflag, slptimeo; 1131 1132 if (vp->v_flag & VRECLAIMED) 1133 return (0); 1134 1135 if ((nmp->nm_flag & NFSMNT_INT) == 0) 1136 intrflg = 0; 1137 if (intrflg) { 1138 slpflag = PCATCH; 1139 slptimeo = 2 * hz; 1140 } else { 1141 slpflag = 0; 1142 slptimeo = 0; 1143 } 1144 /* 1145 * First wait for any other process doing a flush to complete. 1146 */ 1147 while (np->n_flag & NFLUSHINPROG) { 1148 np->n_flag |= NFLUSHWANT; 1149 error = tsleep((caddr_t)&np->n_flag, 0, "nfsvinval", slptimeo); 1150 if (error && intrflg && nfs_sigintr(nmp, (struct nfsreq *)0, td)) 1151 return (EINTR); 1152 } 1153 1154 /* 1155 * Now, flush as required. 1156 */ 1157 np->n_flag |= NFLUSHINPROG; 1158 error = vinvalbuf(vp, flags, td, slpflag, 0); 1159 while (error) { 1160 if (intrflg && nfs_sigintr(nmp, (struct nfsreq *)0, td)) { 1161 np->n_flag &= ~NFLUSHINPROG; 1162 if (np->n_flag & NFLUSHWANT) { 1163 np->n_flag &= ~NFLUSHWANT; 1164 wakeup((caddr_t)&np->n_flag); 1165 } 1166 return (EINTR); 1167 } 1168 error = vinvalbuf(vp, flags, td, 0, slptimeo); 1169 } 1170 np->n_flag &= ~(NLMODIFIED | NFLUSHINPROG); 1171 if (np->n_flag & NFLUSHWANT) { 1172 np->n_flag &= ~NFLUSHWANT; 1173 wakeup((caddr_t)&np->n_flag); 1174 } 1175 return (0); 1176 } 1177 1178 /* 1179 * Initiate asynchronous I/O. Return an error if no nfsiods are available. 1180 * This is mainly to avoid queueing async I/O requests when the nfsiods 1181 * are all hung on a dead server. 1182 * 1183 * Note: nfs_asyncio() does not clear (B_ERROR|B_INVAL) but when the bp 1184 * is eventually dequeued by the async daemon, nfs_doio() *will*. 1185 */ 1186 int 1187 nfs_asyncio(struct vnode *vp, struct bio *bio, struct thread *td) 1188 { 1189 struct buf *bp = bio->bio_buf; 1190 struct nfsmount *nmp; 1191 int i; 1192 int gotiod; 1193 int slpflag = 0; 1194 int slptimeo = 0; 1195 int error; 1196 1197 /* 1198 * If no async daemons then return EIO to force caller to run the rpc 1199 * synchronously. 1200 */ 1201 if (nfs_numasync == 0) 1202 return (EIO); 1203 1204 KKASSERT(vp->v_tag == VT_NFS); 1205 nmp = VFSTONFS(vp->v_mount); 1206 1207 /* 1208 * Commits are usually short and sweet so lets save some cpu and 1209 * leave the async daemons for more important rpc's (such as reads 1210 * and writes). 1211 */ 1212 if ((bp->b_flags & (B_READ|B_NEEDCOMMIT)) == B_NEEDCOMMIT && 1213 (nmp->nm_bioqiods > nfs_numasync / 2)) { 1214 return(EIO); 1215 } 1216 1217 again: 1218 if (nmp->nm_flag & NFSMNT_INT) 1219 slpflag = PCATCH; 1220 gotiod = FALSE; 1221 1222 /* 1223 * Find a free iod to process this request. 1224 */ 1225 for (i = 0; i < NFS_MAXASYNCDAEMON; i++) 1226 if (nfs_iodwant[i]) { 1227 /* 1228 * Found one, so wake it up and tell it which 1229 * mount to process. 1230 */ 1231 NFS_DPF(ASYNCIO, 1232 ("nfs_asyncio: waking iod %d for mount %p\n", 1233 i, nmp)); 1234 nfs_iodwant[i] = NULL; 1235 nfs_iodmount[i] = nmp; 1236 nmp->nm_bioqiods++; 1237 wakeup((caddr_t)&nfs_iodwant[i]); 1238 gotiod = TRUE; 1239 break; 1240 } 1241 1242 /* 1243 * If none are free, we may already have an iod working on this mount 1244 * point. If so, it will process our request. 1245 */ 1246 if (!gotiod) { 1247 if (nmp->nm_bioqiods > 0) { 1248 NFS_DPF(ASYNCIO, 1249 ("nfs_asyncio: %d iods are already processing mount %p\n", 1250 nmp->nm_bioqiods, nmp)); 1251 gotiod = TRUE; 1252 } 1253 } 1254 1255 /* 1256 * If we have an iod which can process the request, then queue 1257 * the buffer. 1258 */ 1259 if (gotiod) { 1260 /* 1261 * Ensure that the queue never grows too large. We still want 1262 * to asynchronize so we block rather then return EIO. 1263 */ 1264 while (nmp->nm_bioqlen >= 2*nfs_numasync) { 1265 NFS_DPF(ASYNCIO, 1266 ("nfs_asyncio: waiting for mount %p queue to drain\n", nmp)); 1267 nmp->nm_bioqwant = TRUE; 1268 error = tsleep(&nmp->nm_bioq, slpflag, 1269 "nfsaio", slptimeo); 1270 if (error) { 1271 if (nfs_sigintr(nmp, NULL, td)) 1272 return (EINTR); 1273 if (slpflag == PCATCH) { 1274 slpflag = 0; 1275 slptimeo = 2 * hz; 1276 } 1277 } 1278 /* 1279 * We might have lost our iod while sleeping, 1280 * so check and loop if nescessary. 1281 */ 1282 if (nmp->nm_bioqiods == 0) { 1283 NFS_DPF(ASYNCIO, 1284 ("nfs_asyncio: no iods after mount %p queue was drained, looping\n", nmp)); 1285 goto again; 1286 } 1287 } 1288 BUF_KERNPROC(bp); 1289 1290 /* 1291 * The passed bio's buffer is not necessary associated with 1292 * the NFS vnode it is being written to. Store the NFS vnode 1293 * in the BIO driver info. 1294 */ 1295 bio->bio_driver_info = vp; 1296 TAILQ_INSERT_TAIL(&nmp->nm_bioq, bio, bio_act); 1297 nmp->nm_bioqlen++; 1298 return (0); 1299 } 1300 1301 /* 1302 * All the iods are busy on other mounts, so return EIO to 1303 * force the caller to process the i/o synchronously. 1304 */ 1305 NFS_DPF(ASYNCIO, ("nfs_asyncio: no iods available, i/o is synchronous\n")); 1306 return (EIO); 1307 } 1308 1309 /* 1310 * Do an I/O operation to/from a cache block. This may be called 1311 * synchronously or from an nfsiod. The BIO is normalized for DEV_BSIZE. 1312 * 1313 * NOTE! TD MIGHT BE NULL 1314 */ 1315 int 1316 nfs_doio(struct vnode *vp, struct bio *bio, struct thread *td) 1317 { 1318 struct buf *bp = bio->bio_buf; 1319 struct uio *uiop; 1320 struct nfsnode *np; 1321 struct nfsmount *nmp; 1322 int error = 0, iomode, must_commit = 0; 1323 struct uio uio; 1324 struct iovec io; 1325 1326 KKASSERT(vp->v_tag == VT_NFS); 1327 np = VTONFS(vp); 1328 nmp = VFSTONFS(vp->v_mount); 1329 uiop = &uio; 1330 uiop->uio_iov = &io; 1331 uiop->uio_iovcnt = 1; 1332 uiop->uio_segflg = UIO_SYSSPACE; 1333 uiop->uio_td = td; 1334 1335 /* 1336 * clear B_ERROR and B_INVAL state prior to initiating the I/O. We 1337 * do this here so we do not have to do it in all the code that 1338 * calls us. 1339 */ 1340 bp->b_flags &= ~(B_ERROR | B_INVAL); 1341 1342 KASSERT(!(bp->b_flags & B_DONE), ("nfs_doio: bp %p already marked done", bp)); 1343 1344 /* 1345 * Historically, paging was done with physio, but no more. 1346 */ 1347 if (bp->b_flags & B_PHYS) { 1348 /* 1349 * ...though reading /dev/drum still gets us here. 1350 */ 1351 io.iov_len = uiop->uio_resid = bp->b_bcount; 1352 /* mapping was done by vmapbuf() */ 1353 io.iov_base = bp->b_data; 1354 uiop->uio_offset = bio->bio_offset; 1355 if (bp->b_flags & B_READ) { 1356 uiop->uio_rw = UIO_READ; 1357 nfsstats.read_physios++; 1358 error = nfs_readrpc(vp, uiop); 1359 } else { 1360 int com; 1361 1362 iomode = NFSV3WRITE_DATASYNC; 1363 uiop->uio_rw = UIO_WRITE; 1364 nfsstats.write_physios++; 1365 error = nfs_writerpc(vp, uiop, &iomode, &com); 1366 } 1367 if (error) { 1368 bp->b_flags |= B_ERROR; 1369 bp->b_error = error; 1370 } 1371 } else if (bp->b_flags & B_READ) { 1372 io.iov_len = uiop->uio_resid = bp->b_bcount; 1373 io.iov_base = bp->b_data; 1374 uiop->uio_rw = UIO_READ; 1375 1376 switch (vp->v_type) { 1377 case VREG: 1378 uiop->uio_offset = bio->bio_offset; 1379 nfsstats.read_bios++; 1380 error = nfs_readrpc(vp, uiop); 1381 1382 if (!error) { 1383 if (uiop->uio_resid) { 1384 /* 1385 * If we had a short read with no error, we must have 1386 * hit a file hole. We should zero-fill the remainder. 1387 * This can also occur if the server hits the file EOF. 1388 * 1389 * Holes used to be able to occur due to pending 1390 * writes, but that is not possible any longer. 1391 */ 1392 int nread = bp->b_bcount - uiop->uio_resid; 1393 int left = uiop->uio_resid; 1394 1395 if (left > 0) 1396 bzero((char *)bp->b_data + nread, left); 1397 uiop->uio_resid = 0; 1398 } 1399 } 1400 if (td && td->td_proc && (vp->v_flag & VTEXT) && 1401 np->n_mtime != np->n_vattr.va_mtime.tv_sec) { 1402 uprintf("Process killed due to text file modification\n"); 1403 psignal(td->td_proc, SIGKILL); 1404 } 1405 break; 1406 case VLNK: 1407 uiop->uio_offset = 0; 1408 nfsstats.readlink_bios++; 1409 error = nfs_readlinkrpc(vp, uiop); 1410 break; 1411 case VDIR: 1412 nfsstats.readdir_bios++; 1413 uiop->uio_offset = bio->bio_offset; 1414 if (nmp->nm_flag & NFSMNT_RDIRPLUS) { 1415 error = nfs_readdirplusrpc(vp, uiop); 1416 if (error == NFSERR_NOTSUPP) 1417 nmp->nm_flag &= ~NFSMNT_RDIRPLUS; 1418 } 1419 if ((nmp->nm_flag & NFSMNT_RDIRPLUS) == 0) 1420 error = nfs_readdirrpc(vp, uiop); 1421 /* 1422 * end-of-directory sets B_INVAL but does not generate an 1423 * error. 1424 */ 1425 if (error == 0 && uiop->uio_resid == bp->b_bcount) 1426 bp->b_flags |= B_INVAL; 1427 break; 1428 default: 1429 printf("nfs_doio: type %x unexpected\n",vp->v_type); 1430 break; 1431 }; 1432 if (error) { 1433 bp->b_flags |= B_ERROR; 1434 bp->b_error = error; 1435 } 1436 } else { 1437 /* 1438 * If we only need to commit, try to commit 1439 */ 1440 if (bp->b_flags & B_NEEDCOMMIT) { 1441 int retv; 1442 off_t off; 1443 1444 off = bio->bio_offset + bp->b_dirtyoff; 1445 retv = nfs_commit(vp, off, 1446 bp->b_dirtyend - bp->b_dirtyoff, td); 1447 if (retv == 0) { 1448 bp->b_dirtyoff = bp->b_dirtyend = 0; 1449 bp->b_flags &= ~(B_NEEDCOMMIT | B_CLUSTEROK); 1450 bp->b_resid = 0; 1451 biodone(bio); 1452 return (0); 1453 } 1454 if (retv == NFSERR_STALEWRITEVERF) { 1455 nfs_clearcommit(vp->v_mount); 1456 } 1457 } 1458 1459 /* 1460 * Setup for actual write 1461 */ 1462 1463 if (bio->bio_offset + bp->b_dirtyend > np->n_size) 1464 bp->b_dirtyend = np->n_size - bio->bio_offset; 1465 1466 if (bp->b_dirtyend > bp->b_dirtyoff) { 1467 io.iov_len = uiop->uio_resid = bp->b_dirtyend 1468 - bp->b_dirtyoff; 1469 uiop->uio_offset = bio->bio_offset + bp->b_dirtyoff; 1470 io.iov_base = (char *)bp->b_data + bp->b_dirtyoff; 1471 uiop->uio_rw = UIO_WRITE; 1472 nfsstats.write_bios++; 1473 1474 if ((bp->b_flags & (B_ASYNC | B_NEEDCOMMIT | B_NOCACHE | B_CLUSTER)) == B_ASYNC) 1475 iomode = NFSV3WRITE_UNSTABLE; 1476 else 1477 iomode = NFSV3WRITE_FILESYNC; 1478 1479 error = nfs_writerpc(vp, uiop, &iomode, &must_commit); 1480 1481 /* 1482 * When setting B_NEEDCOMMIT also set B_CLUSTEROK to try 1483 * to cluster the buffers needing commit. This will allow 1484 * the system to submit a single commit rpc for the whole 1485 * cluster. We can do this even if the buffer is not 100% 1486 * dirty (relative to the NFS blocksize), so we optimize the 1487 * append-to-file-case. 1488 * 1489 * (when clearing B_NEEDCOMMIT, B_CLUSTEROK must also be 1490 * cleared because write clustering only works for commit 1491 * rpc's, not for the data portion of the write). 1492 */ 1493 1494 if (!error && iomode == NFSV3WRITE_UNSTABLE) { 1495 bp->b_flags |= B_NEEDCOMMIT; 1496 if (bp->b_dirtyoff == 0 1497 && bp->b_dirtyend == bp->b_bcount) 1498 bp->b_flags |= B_CLUSTEROK; 1499 } else { 1500 bp->b_flags &= ~(B_NEEDCOMMIT | B_CLUSTEROK); 1501 } 1502 1503 /* 1504 * For an interrupted write, the buffer is still valid 1505 * and the write hasn't been pushed to the server yet, 1506 * so we can't set B_ERROR and report the interruption 1507 * by setting B_EINTR. For the B_ASYNC case, B_EINTR 1508 * is not relevant, so the rpc attempt is essentially 1509 * a noop. For the case of a V3 write rpc not being 1510 * committed to stable storage, the block is still 1511 * dirty and requires either a commit rpc or another 1512 * write rpc with iomode == NFSV3WRITE_FILESYNC before 1513 * the block is reused. This is indicated by setting 1514 * the B_DELWRI and B_NEEDCOMMIT flags. 1515 * 1516 * If the buffer is marked B_PAGING, it does not reside on 1517 * the vp's paging queues so we cannot call bdirty(). The 1518 * bp in this case is not an NFS cache block so we should 1519 * be safe. XXX 1520 */ 1521 if (error == EINTR 1522 || (!error && (bp->b_flags & B_NEEDCOMMIT))) { 1523 crit_enter(); 1524 bp->b_flags &= ~(B_INVAL|B_NOCACHE); 1525 if ((bp->b_flags & B_PAGING) == 0) { 1526 bdirty(bp); 1527 bp->b_flags &= ~B_DONE; 1528 } 1529 if (error && (bp->b_flags & B_ASYNC) == 0) 1530 bp->b_flags |= B_EINTR; 1531 crit_exit(); 1532 } else { 1533 if (error) { 1534 bp->b_flags |= B_ERROR; 1535 bp->b_error = np->n_error = error; 1536 np->n_flag |= NWRITEERR; 1537 } 1538 bp->b_dirtyoff = bp->b_dirtyend = 0; 1539 } 1540 } else { 1541 bp->b_resid = 0; 1542 biodone(bio); 1543 return (0); 1544 } 1545 } 1546 bp->b_resid = uiop->uio_resid; 1547 if (must_commit) 1548 nfs_clearcommit(vp->v_mount); 1549 biodone(bio); 1550 return (error); 1551 } 1552 1553 /* 1554 * Used to aid in handling ftruncate() operations on the NFS client side. 1555 * Truncation creates a number of special problems for NFS. We have to 1556 * throw away VM pages and buffer cache buffers that are beyond EOF, and 1557 * we have to properly handle VM pages or (potentially dirty) buffers 1558 * that straddle the truncation point. 1559 */ 1560 1561 int 1562 nfs_meta_setsize(struct vnode *vp, struct thread *td, u_quad_t nsize) 1563 { 1564 struct nfsnode *np = VTONFS(vp); 1565 u_quad_t tsize = np->n_size; 1566 int biosize = vp->v_mount->mnt_stat.f_iosize; 1567 int error = 0; 1568 1569 np->n_size = nsize; 1570 1571 if (np->n_size < tsize) { 1572 struct buf *bp; 1573 daddr_t lbn; 1574 off_t loffset; 1575 int bufsize; 1576 1577 /* 1578 * vtruncbuf() doesn't get the buffer overlapping the 1579 * truncation point. We may have a B_DELWRI and/or B_CACHE 1580 * buffer that now needs to be truncated. 1581 */ 1582 error = vtruncbuf(vp, td, nsize, biosize); 1583 lbn = nsize / biosize; 1584 bufsize = nsize & (biosize - 1); 1585 loffset = nsize - bufsize; 1586 bp = nfs_getcacheblk(vp, loffset, bufsize, td); 1587 if (bp->b_dirtyoff > bp->b_bcount) 1588 bp->b_dirtyoff = bp->b_bcount; 1589 if (bp->b_dirtyend > bp->b_bcount) 1590 bp->b_dirtyend = bp->b_bcount; 1591 bp->b_flags |= B_RELBUF; /* don't leave garbage around */ 1592 brelse(bp); 1593 } else { 1594 vnode_pager_setsize(vp, nsize); 1595 } 1596 return(error); 1597 } 1598 1599