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 1.45 2008/07/18 00:09:39 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/mbuf.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/buf2.h> 62 #include <sys/thread2.h> 63 #include <vm/vm_page2.h> 64 65 #include "rpcv2.h" 66 #include "nfsproto.h" 67 #include "nfs.h" 68 #include "nfsmount.h" 69 #include "nfsnode.h" 70 #include "xdr_subs.h" 71 #include "nfsm_subs.h" 72 73 74 static struct buf *nfs_getcacheblk(struct vnode *vp, off_t loffset, 75 int size, struct thread *td); 76 static int nfs_check_dirent(struct nfs_dirent *dp, int maxlen); 77 static void nfsiodone_sync(struct bio *bio); 78 static void nfs_readrpc_bio_done(nfsm_info_t info); 79 static void nfs_writerpc_bio_done(nfsm_info_t info); 80 static void nfs_commitrpc_bio_done(nfsm_info_t info); 81 82 /* 83 * Vnode op for VM getpages. 84 * 85 * nfs_getpages(struct vnode *a_vp, vm_page_t *a_m, int a_count, 86 * int a_reqpage, vm_ooffset_t a_offset) 87 */ 88 int 89 nfs_getpages(struct vop_getpages_args *ap) 90 { 91 struct thread *td = curthread; /* XXX */ 92 int i, error, nextoff, size, toff, count, npages; 93 struct uio uio; 94 struct iovec iov; 95 char *kva; 96 struct vnode *vp; 97 struct nfsmount *nmp; 98 vm_page_t *pages; 99 vm_page_t m; 100 struct msf_buf *msf; 101 102 vp = ap->a_vp; 103 nmp = VFSTONFS(vp->v_mount); 104 pages = ap->a_m; 105 count = ap->a_count; 106 107 if (vp->v_object == NULL) { 108 kprintf("nfs_getpages: called with non-merged cache vnode??\n"); 109 return VM_PAGER_ERROR; 110 } 111 112 if ((nmp->nm_flag & NFSMNT_NFSV3) != 0 && 113 (nmp->nm_state & NFSSTA_GOTFSINFO) == 0) 114 (void)nfs_fsinfo(nmp, vp, td); 115 116 npages = btoc(count); 117 118 /* 119 * NOTE that partially valid pages may occur in cases other 120 * then file EOF, such as when a file is partially written and 121 * ftruncate()-extended to a larger size. It is also possible 122 * for the valid bits to be set on garbage beyond the file EOF and 123 * clear in the area before EOF (e.g. m->valid == 0xfc), which can 124 * occur due to vtruncbuf() and the buffer cache's handling of 125 * pages which 'straddle' buffers or when b_bufsize is not a 126 * multiple of PAGE_SIZE.... the buffer cache cannot normally 127 * clear the extra bits. This kind of situation occurs when you 128 * make a small write() (m->valid == 0x03) and then mmap() and 129 * fault in the buffer(m->valid = 0xFF). When NFS flushes the 130 * buffer (vinvalbuf() m->valid = 0xFC) we are left with a mess. 131 * 132 * This is combined with the possibility that the pages are partially 133 * dirty or that there is a buffer backing the pages that is dirty 134 * (even if m->dirty is 0). 135 * 136 * To solve this problem several hacks have been made: (1) NFS 137 * guarentees that the IO block size is a multiple of PAGE_SIZE and 138 * (2) The buffer cache, when invalidating an NFS buffer, will 139 * disregard the buffer's fragmentory b_bufsize and invalidate 140 * the whole page rather then just the piece the buffer owns. 141 * 142 * This allows us to assume that a partially valid page found here 143 * is fully valid (vm_fault will zero'd out areas of the page not 144 * marked as valid). 145 */ 146 m = pages[ap->a_reqpage]; 147 if (m->valid != 0) { 148 for (i = 0; i < npages; ++i) { 149 if (i != ap->a_reqpage) 150 vnode_pager_freepage(pages[i]); 151 } 152 return(0); 153 } 154 155 /* 156 * Use an MSF_BUF as a medium to retrieve data from the pages. 157 */ 158 msf_map_pagelist(&msf, pages, npages, 0); 159 KKASSERT(msf); 160 kva = msf_buf_kva(msf); 161 162 iov.iov_base = kva; 163 iov.iov_len = count; 164 uio.uio_iov = &iov; 165 uio.uio_iovcnt = 1; 166 uio.uio_offset = IDX_TO_OFF(pages[0]->pindex); 167 uio.uio_resid = count; 168 uio.uio_segflg = UIO_SYSSPACE; 169 uio.uio_rw = UIO_READ; 170 uio.uio_td = td; 171 172 error = nfs_readrpc_uio(vp, &uio); 173 msf_buf_free(msf); 174 175 if (error && ((int)uio.uio_resid == count)) { 176 kprintf("nfs_getpages: error %d\n", error); 177 for (i = 0; i < npages; ++i) { 178 if (i != ap->a_reqpage) 179 vnode_pager_freepage(pages[i]); 180 } 181 return VM_PAGER_ERROR; 182 } 183 184 /* 185 * Calculate the number of bytes read and validate only that number 186 * of bytes. Note that due to pending writes, size may be 0. This 187 * does not mean that the remaining data is invalid! 188 */ 189 190 size = count - (int)uio.uio_resid; 191 192 for (i = 0, toff = 0; i < npages; i++, toff = nextoff) { 193 nextoff = toff + PAGE_SIZE; 194 m = pages[i]; 195 196 m->flags &= ~PG_ZERO; 197 198 /* 199 * NOTE: vm_page_undirty/clear_dirty etc do not clear the 200 * pmap modified bit. 201 */ 202 if (nextoff <= size) { 203 /* 204 * Read operation filled an entire page 205 */ 206 m->valid = VM_PAGE_BITS_ALL; 207 vm_page_undirty(m); 208 } else if (size > toff) { 209 /* 210 * Read operation filled a partial page. 211 */ 212 m->valid = 0; 213 vm_page_set_valid(m, 0, size - toff); 214 vm_page_clear_dirty_end_nonincl(m, 0, size - toff); 215 /* handled by vm_fault now */ 216 /* vm_page_zero_invalid(m, TRUE); */ 217 } else { 218 /* 219 * Read operation was short. If no error occured 220 * we may have hit a zero-fill section. We simply 221 * leave valid set to 0. 222 */ 223 ; 224 } 225 if (i != ap->a_reqpage) { 226 /* 227 * Whether or not to leave the page activated is up in 228 * the air, but we should put the page on a page queue 229 * somewhere (it already is in the object). Result: 230 * It appears that emperical results show that 231 * deactivating pages is best. 232 */ 233 234 /* 235 * Just in case someone was asking for this page we 236 * now tell them that it is ok to use. 237 */ 238 if (!error) { 239 if (m->flags & PG_WANTED) 240 vm_page_activate(m); 241 else 242 vm_page_deactivate(m); 243 vm_page_wakeup(m); 244 } else { 245 vnode_pager_freepage(m); 246 } 247 } 248 } 249 return 0; 250 } 251 252 /* 253 * Vnode op for VM putpages. 254 * 255 * The pmap modified bit was cleared prior to the putpages and probably 256 * couldn't get set again until after our I/O completed, since the page 257 * should not be mapped. But don't count on it. The m->dirty bits must 258 * be completely cleared when we finish even if the count is truncated. 259 * 260 * nfs_putpages(struct vnode *a_vp, vm_page_t *a_m, int a_count, int a_sync, 261 * int *a_rtvals, vm_ooffset_t a_offset) 262 */ 263 int 264 nfs_putpages(struct vop_putpages_args *ap) 265 { 266 struct thread *td = curthread; 267 struct uio uio; 268 struct iovec iov; 269 char *kva; 270 int iomode, must_commit, i, error, npages, count; 271 off_t offset; 272 int *rtvals; 273 struct vnode *vp; 274 struct nfsmount *nmp; 275 struct nfsnode *np; 276 vm_page_t *pages; 277 struct msf_buf *msf; 278 279 vp = ap->a_vp; 280 np = VTONFS(vp); 281 nmp = VFSTONFS(vp->v_mount); 282 pages = ap->a_m; 283 count = ap->a_count; 284 rtvals = ap->a_rtvals; 285 npages = btoc(count); 286 offset = IDX_TO_OFF(pages[0]->pindex); 287 288 if ((nmp->nm_flag & NFSMNT_NFSV3) != 0 && 289 (nmp->nm_state & NFSSTA_GOTFSINFO) == 0) 290 (void)nfs_fsinfo(nmp, vp, td); 291 292 for (i = 0; i < npages; i++) { 293 rtvals[i] = VM_PAGER_AGAIN; 294 } 295 296 /* 297 * When putting pages, do not extend file past EOF. 298 */ 299 300 if (offset + count > np->n_size) { 301 count = np->n_size - offset; 302 if (count < 0) 303 count = 0; 304 } 305 306 /* 307 * Use an MSF_BUF as a medium to retrieve data from the pages. 308 */ 309 msf_map_pagelist(&msf, pages, npages, 0); 310 KKASSERT(msf); 311 kva = msf_buf_kva(msf); 312 313 iov.iov_base = kva; 314 iov.iov_len = count; 315 uio.uio_iov = &iov; 316 uio.uio_iovcnt = 1; 317 uio.uio_offset = offset; 318 uio.uio_resid = (size_t)count; 319 uio.uio_segflg = UIO_SYSSPACE; 320 uio.uio_rw = UIO_WRITE; 321 uio.uio_td = td; 322 323 if ((ap->a_sync & VM_PAGER_PUT_SYNC) == 0) 324 iomode = NFSV3WRITE_UNSTABLE; 325 else 326 iomode = NFSV3WRITE_FILESYNC; 327 328 error = nfs_writerpc_uio(vp, &uio, &iomode, &must_commit); 329 330 msf_buf_free(msf); 331 332 if (error == 0) { 333 int nwritten; 334 335 nwritten = round_page(count - (int)uio.uio_resid) / PAGE_SIZE; 336 for (i = 0; i < nwritten; i++) { 337 rtvals[i] = VM_PAGER_OK; 338 vm_page_undirty(pages[i]); 339 } 340 if (must_commit) 341 nfs_clearcommit(vp->v_mount); 342 } 343 return rtvals[0]; 344 } 345 346 /* 347 * Vnode op for read using bio 348 */ 349 int 350 nfs_bioread(struct vnode *vp, struct uio *uio, int ioflag) 351 { 352 struct nfsnode *np = VTONFS(vp); 353 int biosize, i; 354 struct buf *bp, *rabp; 355 struct vattr vattr; 356 struct thread *td; 357 struct nfsmount *nmp = VFSTONFS(vp->v_mount); 358 off_t lbn, rabn; 359 off_t raoffset; 360 off_t loffset; 361 int seqcount; 362 int nra, error = 0; 363 int boff = 0; 364 size_t n; 365 366 #ifdef DIAGNOSTIC 367 if (uio->uio_rw != UIO_READ) 368 panic("nfs_read mode"); 369 #endif 370 if (uio->uio_resid == 0) 371 return (0); 372 if (uio->uio_offset < 0) /* XXX VDIR cookies can be negative */ 373 return (EINVAL); 374 td = uio->uio_td; 375 376 if ((nmp->nm_flag & NFSMNT_NFSV3) != 0 && 377 (nmp->nm_state & NFSSTA_GOTFSINFO) == 0) 378 (void)nfs_fsinfo(nmp, vp, td); 379 if (vp->v_type != VDIR && 380 (uio->uio_offset + uio->uio_resid) > nmp->nm_maxfilesize) 381 return (EFBIG); 382 biosize = vp->v_mount->mnt_stat.f_iosize; 383 seqcount = (int)((off_t)(ioflag >> IO_SEQSHIFT) * biosize / BKVASIZE); 384 385 /* 386 * For nfs, cache consistency can only be maintained approximately. 387 * Although RFC1094 does not specify the criteria, the following is 388 * believed to be compatible with the reference port. 389 * 390 * NFS: If local changes have been made and this is a 391 * directory, the directory must be invalidated and 392 * the attribute cache must be cleared. 393 * 394 * GETATTR is called to synchronize the file size. 395 * 396 * If remote changes are detected local data is flushed 397 * and the cache is invalidated. 398 * 399 * NOTE: In the normal case the attribute cache is not 400 * cleared which means GETATTR may use cached data and 401 * not immediately detect changes made on the server. 402 */ 403 if ((np->n_flag & NLMODIFIED) && vp->v_type == VDIR) { 404 nfs_invaldir(vp); 405 error = nfs_vinvalbuf(vp, V_SAVE, 1); 406 if (error) 407 return (error); 408 np->n_attrstamp = 0; 409 } 410 error = VOP_GETATTR(vp, &vattr); 411 if (error) 412 return (error); 413 if (np->n_flag & NRMODIFIED) { 414 if (vp->v_type == VDIR) 415 nfs_invaldir(vp); 416 error = nfs_vinvalbuf(vp, V_SAVE, 1); 417 if (error) 418 return (error); 419 np->n_flag &= ~NRMODIFIED; 420 } 421 422 /* 423 * Loop until uio exhausted or we hit EOF 424 */ 425 do { 426 bp = NULL; 427 428 switch (vp->v_type) { 429 case VREG: 430 nfsstats.biocache_reads++; 431 lbn = uio->uio_offset / biosize; 432 boff = uio->uio_offset & (biosize - 1); 433 loffset = (off_t)lbn * biosize; 434 435 /* 436 * Start the read ahead(s), as required. 437 */ 438 if (nmp->nm_readahead > 0 && nfs_asyncok(nmp)) { 439 for (nra = 0; nra < nmp->nm_readahead && nra < seqcount && 440 (off_t)(lbn + 1 + nra) * biosize < np->n_size; nra++) { 441 rabn = lbn + 1 + nra; 442 raoffset = (off_t)rabn * biosize; 443 if (findblk(vp, raoffset, FINDBLK_TEST) == NULL) { 444 rabp = nfs_getcacheblk(vp, raoffset, biosize, td); 445 if (!rabp) 446 return (EINTR); 447 if ((rabp->b_flags & (B_CACHE|B_DELWRI)) == 0) { 448 rabp->b_cmd = BUF_CMD_READ; 449 vfs_busy_pages(vp, rabp); 450 nfs_asyncio(vp, &rabp->b_bio2); 451 } else { 452 brelse(rabp); 453 } 454 } 455 } 456 } 457 458 /* 459 * Obtain the buffer cache block. Figure out the buffer size 460 * when we are at EOF. If we are modifying the size of the 461 * buffer based on an EOF condition we need to hold 462 * nfs_rslock() through obtaining the buffer to prevent 463 * a potential writer-appender from messing with n_size. 464 * Otherwise we may accidently truncate the buffer and 465 * lose dirty data. 466 * 467 * Note that bcount is *not* DEV_BSIZE aligned. 468 */ 469 if (loffset + boff >= np->n_size) { 470 n = 0; 471 break; 472 } 473 bp = nfs_getcacheblk(vp, loffset, biosize, td); 474 475 if (bp == NULL) 476 return (EINTR); 477 478 /* 479 * If B_CACHE is not set, we must issue the read. If this 480 * fails, we return an error. 481 */ 482 if ((bp->b_flags & B_CACHE) == 0) { 483 bp->b_cmd = BUF_CMD_READ; 484 bp->b_bio2.bio_done = nfsiodone_sync; 485 bp->b_bio2.bio_flags |= BIO_SYNC; 486 vfs_busy_pages(vp, bp); 487 error = nfs_doio(vp, &bp->b_bio2, td); 488 if (error) { 489 brelse(bp); 490 return (error); 491 } 492 } 493 494 /* 495 * on is the offset into the current bp. Figure out how many 496 * bytes we can copy out of the bp. Note that bcount is 497 * NOT DEV_BSIZE aligned. 498 * 499 * Then figure out how many bytes we can copy into the uio. 500 */ 501 n = biosize - boff; 502 if (n > uio->uio_resid) 503 n = uio->uio_resid; 504 if (loffset + boff + n > np->n_size) 505 n = np->n_size - loffset - boff; 506 break; 507 case VLNK: 508 biosize = min(NFS_MAXPATHLEN, np->n_size); 509 nfsstats.biocache_readlinks++; 510 bp = nfs_getcacheblk(vp, (off_t)0, biosize, td); 511 if (bp == NULL) 512 return (EINTR); 513 if ((bp->b_flags & B_CACHE) == 0) { 514 bp->b_cmd = BUF_CMD_READ; 515 bp->b_bio2.bio_done = nfsiodone_sync; 516 bp->b_bio2.bio_flags |= BIO_SYNC; 517 vfs_busy_pages(vp, bp); 518 error = nfs_doio(vp, &bp->b_bio2, td); 519 if (error) { 520 bp->b_flags |= B_ERROR | B_INVAL; 521 brelse(bp); 522 return (error); 523 } 524 } 525 n = szmin(uio->uio_resid, (size_t)bp->b_bcount - bp->b_resid); 526 boff = 0; 527 break; 528 case VDIR: 529 nfsstats.biocache_readdirs++; 530 if (np->n_direofoffset && 531 uio->uio_offset >= np->n_direofoffset 532 ) { 533 return (0); 534 } 535 lbn = (uoff_t)uio->uio_offset / NFS_DIRBLKSIZ; 536 boff = uio->uio_offset & (NFS_DIRBLKSIZ - 1); 537 loffset = uio->uio_offset - boff; 538 bp = nfs_getcacheblk(vp, loffset, NFS_DIRBLKSIZ, td); 539 if (bp == NULL) 540 return (EINTR); 541 542 if ((bp->b_flags & B_CACHE) == 0) { 543 bp->b_cmd = BUF_CMD_READ; 544 bp->b_bio2.bio_done = nfsiodone_sync; 545 bp->b_bio2.bio_flags |= BIO_SYNC; 546 vfs_busy_pages(vp, bp); 547 error = nfs_doio(vp, &bp->b_bio2, td); 548 if (error) 549 brelse(bp); 550 while (error == NFSERR_BAD_COOKIE) { 551 kprintf("got bad cookie vp %p bp %p\n", vp, bp); 552 nfs_invaldir(vp); 553 error = nfs_vinvalbuf(vp, 0, 1); 554 /* 555 * Yuck! The directory has been modified on the 556 * server. The only way to get the block is by 557 * reading from the beginning to get all the 558 * offset cookies. 559 * 560 * Leave the last bp intact unless there is an error. 561 * Loop back up to the while if the error is another 562 * NFSERR_BAD_COOKIE (double yuch!). 563 */ 564 for (i = 0; i <= lbn && !error; i++) { 565 if (np->n_direofoffset 566 && (i * NFS_DIRBLKSIZ) >= np->n_direofoffset) 567 return (0); 568 bp = nfs_getcacheblk(vp, (off_t)i * NFS_DIRBLKSIZ, 569 NFS_DIRBLKSIZ, td); 570 if (!bp) 571 return (EINTR); 572 if ((bp->b_flags & B_CACHE) == 0) { 573 bp->b_cmd = BUF_CMD_READ; 574 bp->b_bio2.bio_done = nfsiodone_sync; 575 bp->b_bio2.bio_flags |= BIO_SYNC; 576 vfs_busy_pages(vp, bp); 577 error = nfs_doio(vp, &bp->b_bio2, td); 578 /* 579 * no error + B_INVAL == directory EOF, 580 * use the block. 581 */ 582 if (error == 0 && (bp->b_flags & B_INVAL)) 583 break; 584 } 585 /* 586 * An error will throw away the block and the 587 * for loop will break out. If no error and this 588 * is not the block we want, we throw away the 589 * block and go for the next one via the for loop. 590 */ 591 if (error || i < lbn) 592 brelse(bp); 593 } 594 } 595 /* 596 * The above while is repeated if we hit another cookie 597 * error. If we hit an error and it wasn't a cookie error, 598 * we give up. 599 */ 600 if (error) 601 return (error); 602 } 603 604 /* 605 * If not eof and read aheads are enabled, start one. 606 * (You need the current block first, so that you have the 607 * directory offset cookie of the next block.) 608 */ 609 if (nmp->nm_readahead > 0 && nfs_asyncok(nmp) && 610 (bp->b_flags & B_INVAL) == 0 && 611 (np->n_direofoffset == 0 || 612 loffset + NFS_DIRBLKSIZ < np->n_direofoffset) && 613 findblk(vp, loffset + NFS_DIRBLKSIZ, FINDBLK_TEST) == NULL 614 ) { 615 rabp = nfs_getcacheblk(vp, loffset + NFS_DIRBLKSIZ, 616 NFS_DIRBLKSIZ, td); 617 if (rabp) { 618 if ((rabp->b_flags & (B_CACHE|B_DELWRI)) == 0) { 619 rabp->b_cmd = BUF_CMD_READ; 620 vfs_busy_pages(vp, rabp); 621 nfs_asyncio(vp, &rabp->b_bio2); 622 } else { 623 brelse(rabp); 624 } 625 } 626 } 627 /* 628 * Unlike VREG files, whos buffer size ( bp->b_bcount ) is 629 * chopped for the EOF condition, we cannot tell how large 630 * NFS directories are going to be until we hit EOF. So 631 * an NFS directory buffer is *not* chopped to its EOF. Now, 632 * it just so happens that b_resid will effectively chop it 633 * to EOF. *BUT* this information is lost if the buffer goes 634 * away and is reconstituted into a B_CACHE state ( due to 635 * being VMIO ) later. So we keep track of the directory eof 636 * in np->n_direofoffset and chop it off as an extra step 637 * right here. 638 */ 639 n = szmin(uio->uio_resid, 640 NFS_DIRBLKSIZ - bp->b_resid - (size_t)boff); 641 if (np->n_direofoffset && 642 n > (size_t)(np->n_direofoffset - uio->uio_offset)) { 643 n = (size_t)(np->n_direofoffset - uio->uio_offset); 644 } 645 break; 646 default: 647 kprintf(" nfs_bioread: type %x unexpected\n",vp->v_type); 648 n = 0; 649 break; 650 }; 651 652 switch (vp->v_type) { 653 case VREG: 654 if (n > 0) 655 error = uiomove(bp->b_data + boff, n, uio); 656 break; 657 case VLNK: 658 if (n > 0) 659 error = uiomove(bp->b_data + boff, n, uio); 660 n = 0; 661 break; 662 case VDIR: 663 if (n > 0) { 664 off_t old_off = uio->uio_offset; 665 caddr_t cpos, epos; 666 struct nfs_dirent *dp; 667 668 /* 669 * We are casting cpos to nfs_dirent, it must be 670 * int-aligned. 671 */ 672 if (boff & 3) { 673 error = EINVAL; 674 break; 675 } 676 677 cpos = bp->b_data + boff; 678 epos = bp->b_data + boff + n; 679 while (cpos < epos && error == 0 && uio->uio_resid > 0) { 680 dp = (struct nfs_dirent *)cpos; 681 error = nfs_check_dirent(dp, (int)(epos - cpos)); 682 if (error) 683 break; 684 if (vop_write_dirent(&error, uio, dp->nfs_ino, 685 dp->nfs_type, dp->nfs_namlen, dp->nfs_name)) { 686 break; 687 } 688 cpos += dp->nfs_reclen; 689 } 690 n = 0; 691 if (error == 0) { 692 uio->uio_offset = old_off + cpos - 693 bp->b_data - boff; 694 } 695 } 696 break; 697 default: 698 kprintf(" nfs_bioread: type %x unexpected\n",vp->v_type); 699 } 700 if (bp) 701 brelse(bp); 702 } while (error == 0 && uio->uio_resid > 0 && n > 0); 703 return (error); 704 } 705 706 /* 707 * Userland can supply any 'seek' offset when reading a NFS directory. 708 * Validate the structure so we don't panic the kernel. Note that 709 * the element name is nul terminated and the nul is not included 710 * in nfs_namlen. 711 */ 712 static 713 int 714 nfs_check_dirent(struct nfs_dirent *dp, int maxlen) 715 { 716 int nfs_name_off = offsetof(struct nfs_dirent, nfs_name[0]); 717 718 if (nfs_name_off >= maxlen) 719 return (EINVAL); 720 if (dp->nfs_reclen < nfs_name_off || dp->nfs_reclen > maxlen) 721 return (EINVAL); 722 if (nfs_name_off + dp->nfs_namlen >= dp->nfs_reclen) 723 return (EINVAL); 724 if (dp->nfs_reclen & 3) 725 return (EINVAL); 726 return (0); 727 } 728 729 /* 730 * Vnode op for write using bio 731 * 732 * nfs_write(struct vnode *a_vp, struct uio *a_uio, int a_ioflag, 733 * struct ucred *a_cred) 734 */ 735 int 736 nfs_write(struct vop_write_args *ap) 737 { 738 struct uio *uio = ap->a_uio; 739 struct thread *td = uio->uio_td; 740 struct vnode *vp = ap->a_vp; 741 struct nfsnode *np = VTONFS(vp); 742 int ioflag = ap->a_ioflag; 743 struct buf *bp; 744 struct vattr vattr; 745 struct nfsmount *nmp = VFSTONFS(vp->v_mount); 746 off_t loffset; 747 int boff, bytes; 748 int error = 0; 749 int haverslock = 0; 750 int bcount; 751 int biosize; 752 753 #ifdef DIAGNOSTIC 754 if (uio->uio_rw != UIO_WRITE) 755 panic("nfs_write mode"); 756 if (uio->uio_segflg == UIO_USERSPACE && uio->uio_td != curthread) 757 panic("nfs_write proc"); 758 #endif 759 if (vp->v_type != VREG) 760 return (EIO); 761 if (np->n_flag & NWRITEERR) { 762 np->n_flag &= ~NWRITEERR; 763 return (np->n_error); 764 } 765 if ((nmp->nm_flag & NFSMNT_NFSV3) != 0 && 766 (nmp->nm_state & NFSSTA_GOTFSINFO) == 0) 767 (void)nfs_fsinfo(nmp, vp, td); 768 769 /* 770 * Synchronously flush pending buffers if we are in synchronous 771 * mode or if we are appending. 772 */ 773 if (ioflag & (IO_APPEND | IO_SYNC)) { 774 if (np->n_flag & NLMODIFIED) { 775 np->n_attrstamp = 0; 776 error = nfs_flush(vp, MNT_WAIT, td, 0); 777 /* error = nfs_vinvalbuf(vp, V_SAVE, 1); */ 778 if (error) 779 return (error); 780 } 781 } 782 783 /* 784 * If IO_APPEND then load uio_offset. We restart here if we cannot 785 * get the append lock. 786 */ 787 restart: 788 if (ioflag & IO_APPEND) { 789 np->n_attrstamp = 0; 790 error = VOP_GETATTR(vp, &vattr); 791 if (error) 792 return (error); 793 uio->uio_offset = np->n_size; 794 } 795 796 if (uio->uio_offset < 0) 797 return (EINVAL); 798 if ((uio->uio_offset + uio->uio_resid) > nmp->nm_maxfilesize) 799 return (EFBIG); 800 if (uio->uio_resid == 0) 801 return (0); 802 803 /* 804 * We need to obtain the rslock if we intend to modify np->n_size 805 * in order to guarentee the append point with multiple contending 806 * writers, to guarentee that no other appenders modify n_size 807 * while we are trying to obtain a truncated buffer (i.e. to avoid 808 * accidently truncating data written by another appender due to 809 * the race), and to ensure that the buffer is populated prior to 810 * our extending of the file. We hold rslock through the entire 811 * operation. 812 * 813 * Note that we do not synchronize the case where someone truncates 814 * the file while we are appending to it because attempting to lock 815 * this case may deadlock other parts of the system unexpectedly. 816 */ 817 if ((ioflag & IO_APPEND) || 818 uio->uio_offset + uio->uio_resid > np->n_size) { 819 switch(nfs_rslock(np)) { 820 case ENOLCK: 821 goto restart; 822 /* not reached */ 823 case EINTR: 824 case ERESTART: 825 return(EINTR); 826 /* not reached */ 827 default: 828 break; 829 } 830 haverslock = 1; 831 } 832 833 /* 834 * Maybe this should be above the vnode op call, but so long as 835 * file servers have no limits, i don't think it matters 836 */ 837 if (td->td_proc && uio->uio_offset + uio->uio_resid > 838 td->td_proc->p_rlimit[RLIMIT_FSIZE].rlim_cur) { 839 lwpsignal(td->td_proc, td->td_lwp, SIGXFSZ); 840 if (haverslock) 841 nfs_rsunlock(np); 842 return (EFBIG); 843 } 844 845 biosize = vp->v_mount->mnt_stat.f_iosize; 846 847 do { 848 nfsstats.biocache_writes++; 849 boff = uio->uio_offset & (biosize-1); 850 loffset = uio->uio_offset - boff; 851 bytes = (int)szmin((unsigned)(biosize - boff), uio->uio_resid); 852 again: 853 /* 854 * Handle direct append and file extension cases, calculate 855 * unaligned buffer size. When extending B_CACHE will be 856 * set if possible. See UIO_NOCOPY note below. 857 */ 858 if (uio->uio_offset + bytes > np->n_size) { 859 np->n_flag |= NLMODIFIED; 860 bp = nfs_meta_setsize(vp, td, loffset, boff, bytes); 861 } else { 862 bp = nfs_getcacheblk(vp, loffset, biosize, td); 863 } 864 if (bp == NULL) { 865 error = EINTR; 866 break; 867 } 868 869 /* 870 * Actual bytes in buffer which we care about 871 */ 872 if (loffset + biosize < np->n_size) 873 bcount = biosize; 874 else 875 bcount = (int)(np->n_size - loffset); 876 877 /* 878 * Avoid a read by setting B_CACHE where the data we 879 * intend to write covers the entire buffer. Note 880 * that the buffer may have been set to B_CACHE by 881 * nfs_meta_setsize() above or otherwise inherited the 882 * flag, but if B_CACHE isn't set the buffer may be 883 * uninitialized and must be zero'd to accomodate 884 * future seek+write's. 885 * 886 * See the comments in kern/vfs_bio.c's getblk() for 887 * more information. 888 * 889 * When doing a UIO_NOCOPY write the buffer is not 890 * overwritten and we cannot just set B_CACHE unconditionally 891 * for full-block writes. 892 */ 893 if (boff == 0 && bytes == biosize && 894 uio->uio_segflg != UIO_NOCOPY) { 895 bp->b_flags |= B_CACHE; 896 bp->b_flags &= ~(B_ERROR | B_INVAL); 897 } 898 899 /* 900 * b_resid may be set due to file EOF if we extended out. 901 * The NFS bio code will zero the difference anyway so 902 * just acknowledged the fact and set b_resid to 0. 903 */ 904 if ((bp->b_flags & B_CACHE) == 0) { 905 bp->b_cmd = BUF_CMD_READ; 906 bp->b_bio2.bio_done = nfsiodone_sync; 907 bp->b_bio2.bio_flags |= BIO_SYNC; 908 vfs_busy_pages(vp, bp); 909 error = nfs_doio(vp, &bp->b_bio2, td); 910 if (error) { 911 brelse(bp); 912 break; 913 } 914 bp->b_resid = 0; 915 } 916 np->n_flag |= NLMODIFIED; 917 918 /* 919 * If dirtyend exceeds file size, chop it down. This should 920 * not normally occur but there is an append race where it 921 * might occur XXX, so we log it. 922 * 923 * If the chopping creates a reverse-indexed or degenerate 924 * situation with dirtyoff/end, we 0 both of them. 925 */ 926 if (bp->b_dirtyend > bcount) { 927 kprintf("NFS append race @%08llx:%d\n", 928 (long long)bp->b_bio2.bio_offset, 929 bp->b_dirtyend - bcount); 930 bp->b_dirtyend = bcount; 931 } 932 933 if (bp->b_dirtyoff >= bp->b_dirtyend) 934 bp->b_dirtyoff = bp->b_dirtyend = 0; 935 936 /* 937 * If the new write will leave a contiguous dirty 938 * area, just update the b_dirtyoff and b_dirtyend, 939 * otherwise force a write rpc of the old dirty area. 940 * 941 * While it is possible to merge discontiguous writes due to 942 * our having a B_CACHE buffer ( and thus valid read data 943 * for the hole), we don't because it could lead to 944 * significant cache coherency problems with multiple clients, 945 * especially if locking is implemented later on. 946 * 947 * as an optimization we could theoretically maintain 948 * a linked list of discontinuous areas, but we would still 949 * have to commit them separately so there isn't much 950 * advantage to it except perhaps a bit of asynchronization. 951 */ 952 if (bp->b_dirtyend > 0 && 953 (boff > bp->b_dirtyend || 954 (boff + bytes) < bp->b_dirtyoff) 955 ) { 956 if (bwrite(bp) == EINTR) { 957 error = EINTR; 958 break; 959 } 960 goto again; 961 } 962 963 error = uiomove(bp->b_data + boff, bytes, uio); 964 965 /* 966 * Since this block is being modified, it must be written 967 * again and not just committed. Since write clustering does 968 * not work for the stage 1 data write, only the stage 2 969 * commit rpc, we have to clear B_CLUSTEROK as well. 970 */ 971 bp->b_flags &= ~(B_NEEDCOMMIT | B_CLUSTEROK); 972 973 if (error) { 974 brelse(bp); 975 break; 976 } 977 978 /* 979 * Only update dirtyoff/dirtyend if not a degenerate 980 * condition. 981 * 982 * The underlying VM pages have been marked valid by 983 * virtue of acquiring the bp. Because the entire buffer 984 * is marked dirty we do not have to worry about cleaning 985 * out the related dirty bits (and wouldn't really know 986 * how to deal with byte ranges anyway) 987 */ 988 if (bytes) { 989 if (bp->b_dirtyend > 0) { 990 bp->b_dirtyoff = imin(boff, bp->b_dirtyoff); 991 bp->b_dirtyend = imax(boff + bytes, 992 bp->b_dirtyend); 993 } else { 994 bp->b_dirtyoff = boff; 995 bp->b_dirtyend = boff + bytes; 996 } 997 } 998 999 /* 1000 * If the lease is non-cachable or IO_SYNC do bwrite(). 1001 * 1002 * IO_INVAL appears to be unused. The idea appears to be 1003 * to turn off caching in this case. Very odd. XXX 1004 * 1005 * If nfs_async is set bawrite() will use an unstable write 1006 * (build dirty bufs on the server), so we might as well 1007 * push it out with bawrite(). If nfs_async is not set we 1008 * use bdwrite() to cache dirty bufs on the client. 1009 */ 1010 if (ioflag & IO_SYNC) { 1011 if (ioflag & IO_INVAL) 1012 bp->b_flags |= B_NOCACHE; 1013 error = bwrite(bp); 1014 if (error) 1015 break; 1016 } else if (boff + bytes == biosize && nfs_async) { 1017 bawrite(bp); 1018 } else { 1019 bdwrite(bp); 1020 } 1021 } while (uio->uio_resid > 0 && bytes > 0); 1022 1023 if (haverslock) 1024 nfs_rsunlock(np); 1025 1026 return (error); 1027 } 1028 1029 /* 1030 * Get an nfs cache block. 1031 * 1032 * Allocate a new one if the block isn't currently in the cache 1033 * and return the block marked busy. If the calling process is 1034 * interrupted by a signal for an interruptible mount point, return 1035 * NULL. 1036 * 1037 * The caller must carefully deal with the possible B_INVAL state of 1038 * the buffer. nfs_startio() clears B_INVAL (and nfs_asyncio() clears it 1039 * indirectly), so synchronous reads can be issued without worrying about 1040 * the B_INVAL state. We have to be a little more careful when dealing 1041 * with writes (see comments in nfs_write()) when extending a file past 1042 * its EOF. 1043 */ 1044 static struct buf * 1045 nfs_getcacheblk(struct vnode *vp, off_t loffset, int size, struct thread *td) 1046 { 1047 struct buf *bp; 1048 struct mount *mp; 1049 struct nfsmount *nmp; 1050 1051 mp = vp->v_mount; 1052 nmp = VFSTONFS(mp); 1053 1054 if (nmp->nm_flag & NFSMNT_INT) { 1055 bp = getblk(vp, loffset, size, GETBLK_PCATCH, 0); 1056 while (bp == NULL) { 1057 if (nfs_sigintr(nmp, NULL, td)) 1058 return (NULL); 1059 bp = getblk(vp, loffset, size, 0, 2 * hz); 1060 } 1061 } else { 1062 bp = getblk(vp, loffset, size, 0, 0); 1063 } 1064 1065 /* 1066 * bio2, the 'device' layer. Since BIOs use 64 bit byte offsets 1067 * now, no translation is necessary. 1068 */ 1069 bp->b_bio2.bio_offset = loffset; 1070 return (bp); 1071 } 1072 1073 /* 1074 * Flush and invalidate all dirty buffers. If another process is already 1075 * doing the flush, just wait for completion. 1076 */ 1077 int 1078 nfs_vinvalbuf(struct vnode *vp, int flags, int intrflg) 1079 { 1080 struct nfsnode *np = VTONFS(vp); 1081 struct nfsmount *nmp = VFSTONFS(vp->v_mount); 1082 int error = 0, slpflag, slptimeo; 1083 thread_t td = curthread; 1084 1085 if (vp->v_flag & VRECLAIMED) 1086 return (0); 1087 1088 if ((nmp->nm_flag & NFSMNT_INT) == 0) 1089 intrflg = 0; 1090 if (intrflg) { 1091 slpflag = PCATCH; 1092 slptimeo = 2 * hz; 1093 } else { 1094 slpflag = 0; 1095 slptimeo = 0; 1096 } 1097 /* 1098 * First wait for any other process doing a flush to complete. 1099 */ 1100 while (np->n_flag & NFLUSHINPROG) { 1101 np->n_flag |= NFLUSHWANT; 1102 error = tsleep((caddr_t)&np->n_flag, 0, "nfsvinval", slptimeo); 1103 if (error && intrflg && nfs_sigintr(nmp, NULL, td)) 1104 return (EINTR); 1105 } 1106 1107 /* 1108 * Now, flush as required. 1109 */ 1110 np->n_flag |= NFLUSHINPROG; 1111 error = vinvalbuf(vp, flags, slpflag, 0); 1112 while (error) { 1113 if (intrflg && nfs_sigintr(nmp, NULL, td)) { 1114 np->n_flag &= ~NFLUSHINPROG; 1115 if (np->n_flag & NFLUSHWANT) { 1116 np->n_flag &= ~NFLUSHWANT; 1117 wakeup((caddr_t)&np->n_flag); 1118 } 1119 return (EINTR); 1120 } 1121 error = vinvalbuf(vp, flags, 0, slptimeo); 1122 } 1123 np->n_flag &= ~(NLMODIFIED | NFLUSHINPROG); 1124 if (np->n_flag & NFLUSHWANT) { 1125 np->n_flag &= ~NFLUSHWANT; 1126 wakeup((caddr_t)&np->n_flag); 1127 } 1128 return (0); 1129 } 1130 1131 /* 1132 * Return true (non-zero) if the txthread and rxthread are operational 1133 * and we do not already have too many not-yet-started BIO's built up. 1134 */ 1135 int 1136 nfs_asyncok(struct nfsmount *nmp) 1137 { 1138 return (nmp->nm_bioqlen < nfs_maxasyncbio && 1139 nmp->nm_bioqlen < nmp->nm_maxasync_scaled / NFS_ASYSCALE && 1140 nmp->nm_rxstate <= NFSSVC_PENDING && 1141 nmp->nm_txstate <= NFSSVC_PENDING); 1142 } 1143 1144 /* 1145 * The read-ahead code calls this to queue a bio to the txthread. 1146 * 1147 * We don't touch the bio otherwise... that is, we do not even 1148 * construct or send the initial rpc. The txthread will do it 1149 * for us. 1150 * 1151 * NOTE! nm_bioqlen is not decremented until the request completes, 1152 * so it does not reflect the number of bio's on bioq. 1153 */ 1154 void 1155 nfs_asyncio(struct vnode *vp, struct bio *bio) 1156 { 1157 struct buf *bp = bio->bio_buf; 1158 struct nfsmount *nmp = VFSTONFS(vp->v_mount); 1159 1160 KKASSERT(vp->v_tag == VT_NFS); 1161 BUF_KERNPROC(bp); 1162 bio->bio_driver_info = vp; 1163 crit_enter(); 1164 TAILQ_INSERT_TAIL(&nmp->nm_bioq, bio, bio_act); 1165 atomic_add_int(&nmp->nm_bioqlen, 1); 1166 crit_exit(); 1167 nfssvc_iod_writer_wakeup(nmp); 1168 } 1169 1170 /* 1171 * nfs_dio() - Execute a BIO operation synchronously. The BIO will be 1172 * completed and its error returned. The caller is responsible 1173 * for brelse()ing it. ONLY USE FOR BIO_SYNC IOs! Otherwise 1174 * our error probe will be against an invalid pointer. 1175 * 1176 * nfs_startio()- Execute a BIO operation assynchronously. 1177 * 1178 * NOTE: nfs_asyncio() is used to initiate an asynchronous BIO operation, 1179 * which basically just queues it to the txthread. nfs_startio() 1180 * actually initiates the I/O AFTER it has gotten to the txthread. 1181 * 1182 * NOTE: td might be NULL. 1183 * 1184 * NOTE: Caller has already busied the I/O. 1185 */ 1186 void 1187 nfs_startio(struct vnode *vp, struct bio *bio, struct thread *td) 1188 { 1189 struct buf *bp = bio->bio_buf; 1190 struct nfsnode *np; 1191 struct nfsmount *nmp; 1192 1193 KKASSERT(vp->v_tag == VT_NFS); 1194 np = VTONFS(vp); 1195 nmp = VFSTONFS(vp->v_mount); 1196 1197 /* 1198 * clear B_ERROR and B_INVAL state prior to initiating the I/O. We 1199 * do this here so we do not have to do it in all the code that 1200 * calls us. 1201 */ 1202 bp->b_flags &= ~(B_ERROR | B_INVAL); 1203 1204 KASSERT(bp->b_cmd != BUF_CMD_DONE, 1205 ("nfs_doio: bp %p already marked done!", bp)); 1206 1207 if (bp->b_cmd == BUF_CMD_READ) { 1208 switch (vp->v_type) { 1209 case VREG: 1210 nfsstats.read_bios++; 1211 nfs_readrpc_bio(vp, bio); 1212 break; 1213 case VLNK: 1214 #if 0 1215 bio->bio_offset = 0; 1216 nfsstats.readlink_bios++; 1217 nfs_readlinkrpc_bio(vp, bio); 1218 #else 1219 nfs_doio(vp, bio, td); 1220 #endif 1221 break; 1222 case VDIR: 1223 /* 1224 * NOTE: If nfs_readdirplusrpc_bio() is requested but 1225 * not supported, it will chain to 1226 * nfs_readdirrpc_bio(). 1227 */ 1228 #if 0 1229 nfsstats.readdir_bios++; 1230 uiop->uio_offset = bio->bio_offset; 1231 if (nmp->nm_flag & NFSMNT_RDIRPLUS) 1232 nfs_readdirplusrpc_bio(vp, bio); 1233 else 1234 nfs_readdirrpc_bio(vp, bio); 1235 #else 1236 nfs_doio(vp, bio, td); 1237 #endif 1238 break; 1239 default: 1240 kprintf("nfs_doio: type %x unexpected\n",vp->v_type); 1241 bp->b_flags |= B_ERROR; 1242 bp->b_error = EINVAL; 1243 biodone(bio); 1244 break; 1245 } 1246 } else { 1247 /* 1248 * If we only need to commit, try to commit. If this fails 1249 * it will chain through to the write. Basically all the logic 1250 * in nfs_doio() is replicated. 1251 */ 1252 KKASSERT(bp->b_cmd == BUF_CMD_WRITE); 1253 if (bp->b_flags & B_NEEDCOMMIT) 1254 nfs_commitrpc_bio(vp, bio); 1255 else 1256 nfs_writerpc_bio(vp, bio); 1257 } 1258 } 1259 1260 int 1261 nfs_doio(struct vnode *vp, struct bio *bio, struct thread *td) 1262 { 1263 struct buf *bp = bio->bio_buf; 1264 struct uio *uiop; 1265 struct nfsnode *np; 1266 struct nfsmount *nmp; 1267 int error = 0; 1268 int iomode, must_commit; 1269 size_t n; 1270 struct uio uio; 1271 struct iovec io; 1272 1273 KKASSERT(vp->v_tag == VT_NFS); 1274 np = VTONFS(vp); 1275 nmp = VFSTONFS(vp->v_mount); 1276 uiop = &uio; 1277 uiop->uio_iov = &io; 1278 uiop->uio_iovcnt = 1; 1279 uiop->uio_segflg = UIO_SYSSPACE; 1280 uiop->uio_td = td; 1281 1282 /* 1283 * clear B_ERROR and B_INVAL state prior to initiating the I/O. We 1284 * do this here so we do not have to do it in all the code that 1285 * calls us. 1286 */ 1287 bp->b_flags &= ~(B_ERROR | B_INVAL); 1288 1289 KASSERT(bp->b_cmd != BUF_CMD_DONE, 1290 ("nfs_doio: bp %p already marked done!", bp)); 1291 1292 if (bp->b_cmd == BUF_CMD_READ) { 1293 io.iov_len = uiop->uio_resid = (size_t)bp->b_bcount; 1294 io.iov_base = bp->b_data; 1295 uiop->uio_rw = UIO_READ; 1296 1297 switch (vp->v_type) { 1298 case VREG: 1299 /* 1300 * When reading from a regular file zero-fill any residual. 1301 * Note that this residual has nothing to do with NFS short 1302 * reads, which nfs_readrpc_uio() will handle for us. 1303 * 1304 * We have to do this because when we are write extending 1305 * a file the server may not have the same notion of 1306 * filesize as we do. Our BIOs should already be sized 1307 * (b_bcount) to account for the file EOF. 1308 */ 1309 nfsstats.read_bios++; 1310 uiop->uio_offset = bio->bio_offset; 1311 error = nfs_readrpc_uio(vp, uiop); 1312 if (error == 0 && uiop->uio_resid) { 1313 n = (size_t)bp->b_bcount - uiop->uio_resid; 1314 bzero(bp->b_data + n, bp->b_bcount - n); 1315 uiop->uio_resid = 0; 1316 } 1317 if (td && td->td_proc && (vp->v_flag & VTEXT) && 1318 np->n_mtime != np->n_vattr.va_mtime.tv_sec) { 1319 uprintf("Process killed due to text file modification\n"); 1320 ksignal(td->td_proc, SIGKILL); 1321 } 1322 break; 1323 case VLNK: 1324 uiop->uio_offset = 0; 1325 nfsstats.readlink_bios++; 1326 error = nfs_readlinkrpc_uio(vp, uiop); 1327 break; 1328 case VDIR: 1329 nfsstats.readdir_bios++; 1330 uiop->uio_offset = bio->bio_offset; 1331 if (nmp->nm_flag & NFSMNT_RDIRPLUS) { 1332 error = nfs_readdirplusrpc_uio(vp, uiop); 1333 if (error == NFSERR_NOTSUPP) 1334 nmp->nm_flag &= ~NFSMNT_RDIRPLUS; 1335 } 1336 if ((nmp->nm_flag & NFSMNT_RDIRPLUS) == 0) 1337 error = nfs_readdirrpc_uio(vp, uiop); 1338 /* 1339 * end-of-directory sets B_INVAL but does not generate an 1340 * error. 1341 */ 1342 if (error == 0 && uiop->uio_resid == bp->b_bcount) 1343 bp->b_flags |= B_INVAL; 1344 break; 1345 default: 1346 kprintf("nfs_doio: type %x unexpected\n",vp->v_type); 1347 break; 1348 }; 1349 if (error) { 1350 bp->b_flags |= B_ERROR; 1351 bp->b_error = error; 1352 } 1353 bp->b_resid = uiop->uio_resid; 1354 } else { 1355 /* 1356 * If we only need to commit, try to commit. 1357 * 1358 * NOTE: The I/O has already been staged for the write and 1359 * its pages busied, so b_dirtyoff/end is valid. 1360 */ 1361 KKASSERT(bp->b_cmd == BUF_CMD_WRITE); 1362 if (bp->b_flags & B_NEEDCOMMIT) { 1363 int retv; 1364 off_t off; 1365 1366 off = bio->bio_offset + bp->b_dirtyoff; 1367 retv = nfs_commitrpc_uio(vp, off, 1368 bp->b_dirtyend - bp->b_dirtyoff, 1369 td); 1370 if (retv == 0) { 1371 bp->b_dirtyoff = bp->b_dirtyend = 0; 1372 bp->b_flags &= ~(B_NEEDCOMMIT | B_CLUSTEROK); 1373 bp->b_resid = 0; 1374 biodone(bio); 1375 return(0); 1376 } 1377 if (retv == NFSERR_STALEWRITEVERF) { 1378 nfs_clearcommit(vp->v_mount); 1379 } 1380 } 1381 1382 /* 1383 * Setup for actual write 1384 */ 1385 if (bio->bio_offset + bp->b_dirtyend > np->n_size) 1386 bp->b_dirtyend = np->n_size - bio->bio_offset; 1387 1388 if (bp->b_dirtyend > bp->b_dirtyoff) { 1389 io.iov_len = uiop->uio_resid = bp->b_dirtyend 1390 - bp->b_dirtyoff; 1391 uiop->uio_offset = bio->bio_offset + bp->b_dirtyoff; 1392 io.iov_base = (char *)bp->b_data + bp->b_dirtyoff; 1393 uiop->uio_rw = UIO_WRITE; 1394 nfsstats.write_bios++; 1395 1396 if ((bp->b_flags & (B_NEEDCOMMIT | B_NOCACHE | B_CLUSTER)) == 0) 1397 iomode = NFSV3WRITE_UNSTABLE; 1398 else 1399 iomode = NFSV3WRITE_FILESYNC; 1400 1401 must_commit = 0; 1402 error = nfs_writerpc_uio(vp, uiop, &iomode, &must_commit); 1403 1404 /* 1405 * When setting B_NEEDCOMMIT also set B_CLUSTEROK to try 1406 * to cluster the buffers needing commit. This will allow 1407 * the system to submit a single commit rpc for the whole 1408 * cluster. We can do this even if the buffer is not 100% 1409 * dirty (relative to the NFS blocksize), so we optimize the 1410 * append-to-file-case. 1411 * 1412 * (when clearing B_NEEDCOMMIT, B_CLUSTEROK must also be 1413 * cleared because write clustering only works for commit 1414 * rpc's, not for the data portion of the write). 1415 */ 1416 1417 if (!error && iomode == NFSV3WRITE_UNSTABLE) { 1418 bp->b_flags |= B_NEEDCOMMIT; 1419 if (bp->b_dirtyoff == 0 1420 && bp->b_dirtyend == bp->b_bcount) 1421 bp->b_flags |= B_CLUSTEROK; 1422 } else { 1423 bp->b_flags &= ~(B_NEEDCOMMIT | B_CLUSTEROK); 1424 } 1425 1426 /* 1427 * For an interrupted write, the buffer is still valid 1428 * and the write hasn't been pushed to the server yet, 1429 * so we can't set B_ERROR and report the interruption 1430 * by setting B_EINTR. For the async case, B_EINTR 1431 * is not relevant, so the rpc attempt is essentially 1432 * a noop. For the case of a V3 write rpc not being 1433 * committed to stable storage, the block is still 1434 * dirty and requires either a commit rpc or another 1435 * write rpc with iomode == NFSV3WRITE_FILESYNC before 1436 * the block is reused. This is indicated by setting 1437 * the B_DELWRI and B_NEEDCOMMIT flags. 1438 * 1439 * If the buffer is marked B_PAGING, it does not reside on 1440 * the vp's paging queues so we cannot call bdirty(). The 1441 * bp in this case is not an NFS cache block so we should 1442 * be safe. XXX 1443 */ 1444 if (error == EINTR 1445 || (!error && (bp->b_flags & B_NEEDCOMMIT))) { 1446 crit_enter(); 1447 bp->b_flags &= ~(B_INVAL|B_NOCACHE); 1448 if ((bp->b_flags & B_PAGING) == 0) 1449 bdirty(bp); 1450 if (error) 1451 bp->b_flags |= B_EINTR; 1452 crit_exit(); 1453 } else { 1454 if (error) { 1455 bp->b_flags |= B_ERROR; 1456 bp->b_error = np->n_error = error; 1457 np->n_flag |= NWRITEERR; 1458 } 1459 bp->b_dirtyoff = bp->b_dirtyend = 0; 1460 } 1461 if (must_commit) 1462 nfs_clearcommit(vp->v_mount); 1463 bp->b_resid = uiop->uio_resid; 1464 } else { 1465 bp->b_resid = 0; 1466 } 1467 } 1468 1469 /* 1470 * I/O was run synchronously, biodone() it and calculate the 1471 * error to return. 1472 */ 1473 biodone(bio); 1474 KKASSERT(bp->b_cmd == BUF_CMD_DONE); 1475 if (bp->b_flags & B_EINTR) 1476 return (EINTR); 1477 if (bp->b_flags & B_ERROR) 1478 return (bp->b_error ? bp->b_error : EIO); 1479 return (0); 1480 } 1481 1482 /* 1483 * Used to aid in handling ftruncate() and non-trivial write-extend 1484 * operations on the NFS client side. Note that trivial write-extend 1485 * operations (appending with no write hole) are handled by nfs_write() 1486 * directly to avoid silly flushes. 1487 * 1488 * Truncation creates a number of special problems for NFS. We have to 1489 * throw away VM pages and buffer cache buffers that are beyond EOF, and 1490 * we have to properly handle VM pages or (potentially dirty) buffers 1491 * that straddle the truncation point. 1492 * 1493 * File extension no longer has an issue now that the buffer size is 1494 * fixed. When extending the intended overwrite area is specified 1495 * by (boff, bytes). This function uses the parameters to determine 1496 * what areas must be zerod. If there are no gaps we set B_CACHE. 1497 */ 1498 struct buf * 1499 nfs_meta_setsize(struct vnode *vp, struct thread *td, off_t nbase, 1500 int boff, int bytes) 1501 { 1502 1503 struct nfsnode *np = VTONFS(vp); 1504 off_t osize = np->n_size; 1505 off_t nsize; 1506 int biosize = vp->v_mount->mnt_stat.f_iosize; 1507 int error = 0; 1508 struct buf *bp; 1509 1510 nsize = nbase + boff + bytes; 1511 np->n_size = nsize; 1512 1513 if (nsize < osize) { 1514 /* 1515 * vtruncbuf() doesn't get the buffer overlapping the 1516 * truncation point, but it will invalidate pages in 1517 * that buffer and zero the appropriate byte range in 1518 * the page straddling EOF. 1519 */ 1520 error = vtruncbuf(vp, nsize, biosize); 1521 1522 /* 1523 * NFS doesn't do a good job tracking changes in the EOF 1524 * so it may not revisit the buffer if the file is extended. 1525 * 1526 * After truncating just clear B_CACHE on the buffer 1527 * straddling EOF. If the buffer is dirty then clean 1528 * out the portion beyond the file EOF. 1529 */ 1530 if (error) { 1531 bp = NULL; 1532 } else { 1533 bp = nfs_getcacheblk(vp, nbase, biosize, td); 1534 if (bp->b_flags & B_DELWRI) { 1535 if (bp->b_dirtyoff > bp->b_bcount) 1536 bp->b_dirtyoff = bp->b_bcount; 1537 if (bp->b_dirtyend > bp->b_bcount) 1538 bp->b_dirtyend = bp->b_bcount; 1539 boff = (int)nsize & (biosize - 1); 1540 bzero(bp->b_data + boff, biosize - boff); 1541 } else if (nsize != nbase) { 1542 boff = (int)nsize & (biosize - 1); 1543 bzero(bp->b_data + boff, biosize - boff); 1544 } 1545 } 1546 } else { 1547 /* 1548 * The newly expanded portions of the buffer should already 1549 * be zero'd out if B_CACHE is set. If B_CACHE is not 1550 * set and the buffer is beyond osize we can safely zero it 1551 * and set B_CACHE to avoid issuing unnecessary degenerate 1552 * read rpcs. 1553 * 1554 * Don't do this if the caller is going to overwrite the 1555 * entire buffer anyway (and also don't set B_CACHE!). 1556 * This allows the caller to optimize the operation. 1557 */ 1558 KKASSERT(nsize >= 0); 1559 vnode_pager_setsize(vp, (vm_ooffset_t)nsize); 1560 1561 bp = nfs_getcacheblk(vp, nbase, biosize, td); 1562 if ((bp->b_flags & B_CACHE) == 0 && nbase >= osize && 1563 !(boff == 0 && bytes == biosize) 1564 ) { 1565 bzero(bp->b_data, biosize); 1566 bp->b_flags |= B_CACHE; 1567 bp->b_flags &= ~(B_ERROR | B_INVAL); 1568 } 1569 } 1570 return(bp); 1571 } 1572 1573 /* 1574 * Synchronous completion for nfs_doio. Call bpdone() with elseit=FALSE. 1575 * Caller is responsible for brelse()'ing the bp. 1576 */ 1577 static void 1578 nfsiodone_sync(struct bio *bio) 1579 { 1580 bio->bio_flags = 0; 1581 bpdone(bio->bio_buf, 0); 1582 } 1583 1584 /* 1585 * nfs read rpc - BIO version 1586 */ 1587 void 1588 nfs_readrpc_bio(struct vnode *vp, struct bio *bio) 1589 { 1590 struct buf *bp = bio->bio_buf; 1591 u_int32_t *tl; 1592 struct nfsmount *nmp; 1593 int error = 0, len, tsiz; 1594 struct nfsm_info *info; 1595 1596 info = kmalloc(sizeof(*info), M_NFSREQ, M_WAITOK); 1597 info->mrep = NULL; 1598 info->v3 = NFS_ISV3(vp); 1599 1600 nmp = VFSTONFS(vp->v_mount); 1601 tsiz = bp->b_bcount; 1602 KKASSERT(tsiz <= nmp->nm_rsize); 1603 if (bio->bio_offset + tsiz > nmp->nm_maxfilesize) { 1604 error = EFBIG; 1605 goto nfsmout; 1606 } 1607 nfsstats.rpccnt[NFSPROC_READ]++; 1608 len = tsiz; 1609 nfsm_reqhead(info, vp, NFSPROC_READ, 1610 NFSX_FH(info->v3) + NFSX_UNSIGNED * 3); 1611 ERROROUT(nfsm_fhtom(info, vp)); 1612 tl = nfsm_build(info, NFSX_UNSIGNED * 3); 1613 if (info->v3) { 1614 txdr_hyper(bio->bio_offset, tl); 1615 *(tl + 2) = txdr_unsigned(len); 1616 } else { 1617 *tl++ = txdr_unsigned(bio->bio_offset); 1618 *tl++ = txdr_unsigned(len); 1619 *tl = 0; 1620 } 1621 info->bio = bio; 1622 info->done = nfs_readrpc_bio_done; 1623 nfsm_request_bio(info, vp, NFSPROC_READ, NULL, 1624 nfs_vpcred(vp, ND_READ)); 1625 return; 1626 nfsmout: 1627 kfree(info, M_NFSREQ); 1628 bp->b_error = error; 1629 bp->b_flags |= B_ERROR; 1630 biodone(bio); 1631 } 1632 1633 static void 1634 nfs_readrpc_bio_done(nfsm_info_t info) 1635 { 1636 struct nfsmount *nmp = VFSTONFS(info->vp->v_mount); 1637 struct bio *bio = info->bio; 1638 struct buf *bp = bio->bio_buf; 1639 u_int32_t *tl; 1640 int attrflag; 1641 int retlen; 1642 int eof; 1643 int error = 0; 1644 1645 KKASSERT(info->state == NFSM_STATE_DONE); 1646 1647 if (info->v3) { 1648 ERROROUT(nfsm_postop_attr(info, info->vp, &attrflag, 1649 NFS_LATTR_NOSHRINK)); 1650 NULLOUT(tl = nfsm_dissect(info, 2 * NFSX_UNSIGNED)); 1651 eof = fxdr_unsigned(int, *(tl + 1)); 1652 } else { 1653 ERROROUT(nfsm_loadattr(info, info->vp, NULL)); 1654 eof = 0; 1655 } 1656 NEGATIVEOUT(retlen = nfsm_strsiz(info, nmp->nm_rsize)); 1657 ERROROUT(nfsm_mtobio(info, bio, retlen)); 1658 m_freem(info->mrep); 1659 info->mrep = NULL; 1660 1661 /* 1662 * No error occured, if retlen is less then bcount and no EOF 1663 * and NFSv3 a zero-fill short read occured. 1664 * 1665 * For NFSv2 a short-read indicates EOF. 1666 */ 1667 if (retlen < bp->b_bcount && info->v3 && eof == 0) { 1668 bzero(bp->b_data + retlen, bp->b_bcount - retlen); 1669 retlen = bp->b_bcount; 1670 } 1671 1672 /* 1673 * If we hit an EOF we still zero-fill, but return the expected 1674 * b_resid anyway. This should normally not occur since async 1675 * BIOs are not used for read-before-write case. Races against 1676 * the server can cause it though and we don't want to leave 1677 * garbage in the buffer. 1678 */ 1679 if (retlen < bp->b_bcount) { 1680 bzero(bp->b_data + retlen, bp->b_bcount - retlen); 1681 } 1682 bp->b_resid = 0; 1683 /* bp->b_resid = bp->b_bcount - retlen; */ 1684 nfsmout: 1685 kfree(info, M_NFSREQ); 1686 if (error) { 1687 bp->b_error = error; 1688 bp->b_flags |= B_ERROR; 1689 } 1690 biodone(bio); 1691 } 1692 1693 /* 1694 * nfs write call - BIO version 1695 * 1696 * NOTE: Caller has already busied the I/O. 1697 */ 1698 void 1699 nfs_writerpc_bio(struct vnode *vp, struct bio *bio) 1700 { 1701 struct nfsmount *nmp = VFSTONFS(vp->v_mount); 1702 struct nfsnode *np = VTONFS(vp); 1703 struct buf *bp = bio->bio_buf; 1704 u_int32_t *tl; 1705 int len; 1706 int iomode; 1707 int error = 0; 1708 struct nfsm_info *info; 1709 off_t offset; 1710 1711 /* 1712 * Setup for actual write. Just clean up the bio if there 1713 * is nothing to do. b_dirtyoff/end have already been staged 1714 * by the bp's pages getting busied. 1715 */ 1716 if (bio->bio_offset + bp->b_dirtyend > np->n_size) 1717 bp->b_dirtyend = np->n_size - bio->bio_offset; 1718 1719 if (bp->b_dirtyend <= bp->b_dirtyoff) { 1720 bp->b_resid = 0; 1721 biodone(bio); 1722 return; 1723 } 1724 len = bp->b_dirtyend - bp->b_dirtyoff; 1725 offset = bio->bio_offset + bp->b_dirtyoff; 1726 if (offset + len > nmp->nm_maxfilesize) { 1727 bp->b_flags |= B_ERROR; 1728 bp->b_error = EFBIG; 1729 biodone(bio); 1730 return; 1731 } 1732 bp->b_resid = len; 1733 nfsstats.write_bios++; 1734 1735 info = kmalloc(sizeof(*info), M_NFSREQ, M_WAITOK); 1736 info->mrep = NULL; 1737 info->v3 = NFS_ISV3(vp); 1738 info->info_writerpc.must_commit = 0; 1739 if ((bp->b_flags & (B_NEEDCOMMIT | B_NOCACHE | B_CLUSTER)) == 0) 1740 iomode = NFSV3WRITE_UNSTABLE; 1741 else 1742 iomode = NFSV3WRITE_FILESYNC; 1743 1744 KKASSERT(len <= nmp->nm_wsize); 1745 1746 nfsstats.rpccnt[NFSPROC_WRITE]++; 1747 nfsm_reqhead(info, vp, NFSPROC_WRITE, 1748 NFSX_FH(info->v3) + 5 * NFSX_UNSIGNED + nfsm_rndup(len)); 1749 ERROROUT(nfsm_fhtom(info, vp)); 1750 if (info->v3) { 1751 tl = nfsm_build(info, 5 * NFSX_UNSIGNED); 1752 txdr_hyper(offset, tl); 1753 tl += 2; 1754 *tl++ = txdr_unsigned(len); 1755 *tl++ = txdr_unsigned(iomode); 1756 *tl = txdr_unsigned(len); 1757 } else { 1758 u_int32_t x; 1759 1760 tl = nfsm_build(info, 4 * NFSX_UNSIGNED); 1761 /* Set both "begin" and "current" to non-garbage. */ 1762 x = txdr_unsigned((u_int32_t)offset); 1763 *tl++ = x; /* "begin offset" */ 1764 *tl++ = x; /* "current offset" */ 1765 x = txdr_unsigned(len); 1766 *tl++ = x; /* total to this offset */ 1767 *tl = x; /* size of this write */ 1768 } 1769 ERROROUT(nfsm_biotom(info, bio, bp->b_dirtyoff, len)); 1770 info->bio = bio; 1771 info->done = nfs_writerpc_bio_done; 1772 nfsm_request_bio(info, vp, NFSPROC_WRITE, NULL, 1773 nfs_vpcred(vp, ND_WRITE)); 1774 return; 1775 nfsmout: 1776 kfree(info, M_NFSREQ); 1777 bp->b_error = error; 1778 bp->b_flags |= B_ERROR; 1779 biodone(bio); 1780 } 1781 1782 static void 1783 nfs_writerpc_bio_done(nfsm_info_t info) 1784 { 1785 struct nfsmount *nmp = VFSTONFS(info->vp->v_mount); 1786 struct nfsnode *np = VTONFS(info->vp); 1787 struct bio *bio = info->bio; 1788 struct buf *bp = bio->bio_buf; 1789 int wccflag = NFSV3_WCCRATTR; 1790 int iomode = NFSV3WRITE_FILESYNC; 1791 int commit; 1792 int rlen; 1793 int error; 1794 int len = bp->b_resid; /* b_resid was set to shortened length */ 1795 u_int32_t *tl; 1796 1797 if (info->v3) { 1798 /* 1799 * The write RPC returns a before and after mtime. The 1800 * nfsm_wcc_data() macro checks the before n_mtime 1801 * against the before time and stores the after time 1802 * in the nfsnode's cached vattr and n_mtime field. 1803 * The NRMODIFIED bit will be set if the before 1804 * time did not match the original mtime. 1805 */ 1806 wccflag = NFSV3_WCCCHK; 1807 ERROROUT(nfsm_wcc_data(info, info->vp, &wccflag)); 1808 if (error == 0) { 1809 NULLOUT(tl = nfsm_dissect(info, 2 * NFSX_UNSIGNED + NFSX_V3WRITEVERF)); 1810 rlen = fxdr_unsigned(int, *tl++); 1811 if (rlen == 0) { 1812 error = NFSERR_IO; 1813 m_freem(info->mrep); 1814 info->mrep = NULL; 1815 goto nfsmout; 1816 } else if (rlen < len) { 1817 #if 0 1818 /* 1819 * XXX what do we do here? 1820 */ 1821 backup = len - rlen; 1822 uiop->uio_iov->iov_base = (char *)uiop->uio_iov->iov_base - backup; 1823 uiop->uio_iov->iov_len += backup; 1824 uiop->uio_offset -= backup; 1825 uiop->uio_resid += backup; 1826 len = rlen; 1827 #endif 1828 } 1829 commit = fxdr_unsigned(int, *tl++); 1830 1831 /* 1832 * Return the lowest committment level 1833 * obtained by any of the RPCs. 1834 */ 1835 if (iomode == NFSV3WRITE_FILESYNC) 1836 iomode = commit; 1837 else if (iomode == NFSV3WRITE_DATASYNC && 1838 commit == NFSV3WRITE_UNSTABLE) 1839 iomode = commit; 1840 if ((nmp->nm_state & NFSSTA_HASWRITEVERF) == 0){ 1841 bcopy(tl, (caddr_t)nmp->nm_verf, NFSX_V3WRITEVERF); 1842 nmp->nm_state |= NFSSTA_HASWRITEVERF; 1843 } else if (bcmp(tl, nmp->nm_verf, NFSX_V3WRITEVERF)) { 1844 info->info_writerpc.must_commit = 1; 1845 bcopy(tl, (caddr_t)nmp->nm_verf, NFSX_V3WRITEVERF); 1846 } 1847 } 1848 } else { 1849 ERROROUT(nfsm_loadattr(info, info->vp, NULL)); 1850 } 1851 m_freem(info->mrep); 1852 info->mrep = NULL; 1853 len = 0; 1854 nfsmout: 1855 if (info->vp->v_mount->mnt_flag & MNT_ASYNC) 1856 iomode = NFSV3WRITE_FILESYNC; 1857 bp->b_resid = len; 1858 1859 /* 1860 * End of RPC. Now clean up the bp. 1861 * 1862 * When setting B_NEEDCOMMIT also set B_CLUSTEROK to try 1863 * to cluster the buffers needing commit. This will allow 1864 * the system to submit a single commit rpc for the whole 1865 * cluster. We can do this even if the buffer is not 100% 1866 * dirty (relative to the NFS blocksize), so we optimize the 1867 * append-to-file-case. 1868 * 1869 * (when clearing B_NEEDCOMMIT, B_CLUSTEROK must also be 1870 * cleared because write clustering only works for commit 1871 * rpc's, not for the data portion of the write). 1872 */ 1873 if (!error && iomode == NFSV3WRITE_UNSTABLE) { 1874 bp->b_flags |= B_NEEDCOMMIT; 1875 if (bp->b_dirtyoff == 0 && bp->b_dirtyend == bp->b_bcount) 1876 bp->b_flags |= B_CLUSTEROK; 1877 } else { 1878 bp->b_flags &= ~(B_NEEDCOMMIT | B_CLUSTEROK); 1879 } 1880 1881 /* 1882 * For an interrupted write, the buffer is still valid 1883 * and the write hasn't been pushed to the server yet, 1884 * so we can't set B_ERROR and report the interruption 1885 * by setting B_EINTR. For the async case, B_EINTR 1886 * is not relevant, so the rpc attempt is essentially 1887 * a noop. For the case of a V3 write rpc not being 1888 * committed to stable storage, the block is still 1889 * dirty and requires either a commit rpc or another 1890 * write rpc with iomode == NFSV3WRITE_FILESYNC before 1891 * the block is reused. This is indicated by setting 1892 * the B_DELWRI and B_NEEDCOMMIT flags. 1893 * 1894 * If the buffer is marked B_PAGING, it does not reside on 1895 * the vp's paging queues so we cannot call bdirty(). The 1896 * bp in this case is not an NFS cache block so we should 1897 * be safe. XXX 1898 */ 1899 if (error == EINTR || (!error && (bp->b_flags & B_NEEDCOMMIT))) { 1900 crit_enter(); 1901 bp->b_flags &= ~(B_INVAL|B_NOCACHE); 1902 if ((bp->b_flags & B_PAGING) == 0) 1903 bdirty(bp); 1904 if (error) 1905 bp->b_flags |= B_EINTR; 1906 crit_exit(); 1907 } else { 1908 if (error) { 1909 bp->b_flags |= B_ERROR; 1910 bp->b_error = np->n_error = error; 1911 np->n_flag |= NWRITEERR; 1912 } 1913 bp->b_dirtyoff = bp->b_dirtyend = 0; 1914 } 1915 if (info->info_writerpc.must_commit) 1916 nfs_clearcommit(info->vp->v_mount); 1917 kfree(info, M_NFSREQ); 1918 if (error) { 1919 bp->b_flags |= B_ERROR; 1920 bp->b_error = error; 1921 } 1922 biodone(bio); 1923 } 1924 1925 /* 1926 * Nfs Version 3 commit rpc - BIO version 1927 * 1928 * This function issues the commit rpc and will chain to a write 1929 * rpc if necessary. 1930 */ 1931 void 1932 nfs_commitrpc_bio(struct vnode *vp, struct bio *bio) 1933 { 1934 struct nfsmount *nmp = VFSTONFS(vp->v_mount); 1935 struct buf *bp = bio->bio_buf; 1936 struct nfsm_info *info; 1937 int error = 0; 1938 u_int32_t *tl; 1939 1940 if ((nmp->nm_state & NFSSTA_HASWRITEVERF) == 0) { 1941 bp->b_dirtyoff = bp->b_dirtyend = 0; 1942 bp->b_flags &= ~(B_NEEDCOMMIT | B_CLUSTEROK); 1943 bp->b_resid = 0; 1944 biodone(bio); 1945 return; 1946 } 1947 1948 info = kmalloc(sizeof(*info), M_NFSREQ, M_WAITOK); 1949 info->mrep = NULL; 1950 info->v3 = 1; 1951 1952 nfsstats.rpccnt[NFSPROC_COMMIT]++; 1953 nfsm_reqhead(info, vp, NFSPROC_COMMIT, NFSX_FH(1)); 1954 ERROROUT(nfsm_fhtom(info, vp)); 1955 tl = nfsm_build(info, 3 * NFSX_UNSIGNED); 1956 txdr_hyper(bio->bio_offset + bp->b_dirtyoff, tl); 1957 tl += 2; 1958 *tl = txdr_unsigned(bp->b_dirtyend - bp->b_dirtyoff); 1959 info->bio = bio; 1960 info->done = nfs_commitrpc_bio_done; 1961 nfsm_request_bio(info, vp, NFSPROC_COMMIT, NULL, 1962 nfs_vpcred(vp, ND_WRITE)); 1963 return; 1964 nfsmout: 1965 /* 1966 * Chain to write RPC on (early) error 1967 */ 1968 kfree(info, M_NFSREQ); 1969 nfs_writerpc_bio(vp, bio); 1970 } 1971 1972 static void 1973 nfs_commitrpc_bio_done(nfsm_info_t info) 1974 { 1975 struct nfsmount *nmp = VFSTONFS(info->vp->v_mount); 1976 struct bio *bio = info->bio; 1977 struct buf *bp = bio->bio_buf; 1978 u_int32_t *tl; 1979 int wccflag = NFSV3_WCCRATTR; 1980 int error = 0; 1981 1982 ERROROUT(nfsm_wcc_data(info, info->vp, &wccflag)); 1983 if (error == 0) { 1984 NULLOUT(tl = nfsm_dissect(info, NFSX_V3WRITEVERF)); 1985 if (bcmp(nmp->nm_verf, tl, NFSX_V3WRITEVERF)) { 1986 bcopy(tl, nmp->nm_verf, NFSX_V3WRITEVERF); 1987 error = NFSERR_STALEWRITEVERF; 1988 } 1989 } 1990 m_freem(info->mrep); 1991 info->mrep = NULL; 1992 1993 /* 1994 * On completion we must chain to a write bio if an 1995 * error occurred. 1996 */ 1997 nfsmout: 1998 kfree(info, M_NFSREQ); 1999 if (error == 0) { 2000 bp->b_dirtyoff = bp->b_dirtyend = 0; 2001 bp->b_flags &= ~(B_NEEDCOMMIT | B_CLUSTEROK); 2002 bp->b_resid = 0; 2003 biodone(bio); 2004 } else { 2005 nfs_writerpc_bio(info->vp, bio); 2006 } 2007 } 2008 2009