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