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 */ 39 40 41 #include <sys/param.h> 42 #include <sys/systm.h> 43 #include <sys/resourcevar.h> 44 #include <sys/signalvar.h> 45 #include <sys/proc.h> 46 #include <sys/buf.h> 47 #include <sys/vnode.h> 48 #include <sys/mount.h> 49 #include <sys/kernel.h> 50 #include <sys/mbuf.h> 51 52 #include <vm/vm.h> 53 #include <vm/vm_extern.h> 54 #include <vm/vm_page.h> 55 #include <vm/vm_object.h> 56 #include <vm/vm_pager.h> 57 #include <vm/vnode_pager.h> 58 59 #include <sys/buf2.h> 60 #include <sys/thread2.h> 61 #include <vm/vm_page2.h> 62 63 #include "rpcv2.h" 64 #include "nfsproto.h" 65 #include "nfs.h" 66 #include "nfsmount.h" 67 #include "nfsnode.h" 68 #include "xdr_subs.h" 69 #include "nfsm_subs.h" 70 71 72 static struct buf *nfs_getcacheblk(struct vnode *vp, off_t loffset, 73 int size, struct thread *td); 74 static int nfs_check_dirent(struct nfs_dirent *dp, int maxlen); 75 static void nfsiodone_sync(struct bio *bio); 76 static void nfs_readrpc_bio_done(nfsm_info_t info); 77 static void nfs_writerpc_bio_done(nfsm_info_t info); 78 static void nfs_commitrpc_bio_done(nfsm_info_t info); 79 80 /* 81 * Vnode op for read using bio 82 */ 83 int 84 nfs_bioread(struct vnode *vp, struct uio *uio, int ioflag) 85 { 86 struct nfsnode *np = VTONFS(vp); 87 int biosize, i; 88 struct buf *bp, *rabp; 89 struct vattr vattr; 90 struct thread *td; 91 struct nfsmount *nmp = VFSTONFS(vp->v_mount); 92 off_t lbn, rabn; 93 off_t raoffset; 94 off_t loffset; 95 int seqcount; 96 int nra, error = 0; 97 int boff = 0; 98 size_t n; 99 100 #ifdef DIAGNOSTIC 101 if (uio->uio_rw != UIO_READ) 102 panic("nfs_read mode"); 103 #endif 104 if (uio->uio_resid == 0) 105 return (0); 106 if (uio->uio_offset < 0) /* XXX VDIR cookies can be negative */ 107 return (EINVAL); 108 td = uio->uio_td; 109 110 if ((nmp->nm_flag & NFSMNT_NFSV3) != 0 && 111 (nmp->nm_state & NFSSTA_GOTFSINFO) == 0) 112 (void)nfs_fsinfo(nmp, vp, td); 113 if (vp->v_type != VDIR && 114 (uio->uio_offset + uio->uio_resid) > nmp->nm_maxfilesize) 115 return (EFBIG); 116 biosize = vp->v_mount->mnt_stat.f_iosize; 117 seqcount = (int)((off_t)(ioflag >> IO_SEQSHIFT) * biosize / BKVASIZE); 118 119 /* 120 * For nfs, cache consistency can only be maintained approximately. 121 * Although RFC1094 does not specify the criteria, the following is 122 * believed to be compatible with the reference port. 123 * 124 * NFS: If local changes have been made and this is a 125 * directory, the directory must be invalidated and 126 * the attribute cache must be cleared. 127 * 128 * GETATTR is called to synchronize the file size. 129 * 130 * If remote changes are detected local data is flushed 131 * and the cache is invalidated. 132 * 133 * NOTE: In the normal case the attribute cache is not 134 * cleared which means GETATTR may use cached data and 135 * not immediately detect changes made on the server. 136 */ 137 if ((np->n_flag & NLMODIFIED) && vp->v_type == VDIR) { 138 nfs_invaldir(vp); 139 error = nfs_vinvalbuf(vp, V_SAVE, 1); 140 if (error) 141 return (error); 142 np->n_attrstamp = 0; 143 } 144 error = VOP_GETATTR(vp, &vattr); 145 if (error) 146 return (error); 147 148 /* 149 * This can deadlock getpages/putpages for regular 150 * files. Only do it for directories. 151 */ 152 if (np->n_flag & NRMODIFIED) { 153 if (vp->v_type == VDIR) { 154 nfs_invaldir(vp); 155 error = nfs_vinvalbuf(vp, V_SAVE, 1); 156 if (error) 157 return (error); 158 np->n_flag &= ~NRMODIFIED; 159 } 160 } 161 162 /* 163 * Loop until uio exhausted or we hit EOF 164 */ 165 do { 166 bp = NULL; 167 168 switch (vp->v_type) { 169 case VREG: 170 nfsstats.biocache_reads++; 171 lbn = uio->uio_offset / biosize; 172 boff = uio->uio_offset & (biosize - 1); 173 loffset = lbn * biosize; 174 175 /* 176 * Start the read ahead(s), as required. 177 */ 178 if (nmp->nm_readahead > 0 && nfs_asyncok(nmp)) { 179 for (nra = 0; nra < nmp->nm_readahead && nra < seqcount && 180 (off_t)(lbn + 1 + nra) * biosize < np->n_size; nra++) { 181 rabn = lbn + 1 + nra; 182 raoffset = rabn * biosize; 183 if (findblk(vp, raoffset, FINDBLK_TEST) == NULL) { 184 rabp = nfs_getcacheblk(vp, raoffset, biosize, td); 185 if (!rabp) 186 return (EINTR); 187 if ((rabp->b_flags & (B_CACHE|B_DELWRI)) == 0) { 188 rabp->b_cmd = BUF_CMD_READ; 189 vfs_busy_pages(vp, rabp); 190 nfs_asyncio(vp, &rabp->b_bio2); 191 } else { 192 brelse(rabp); 193 } 194 } 195 } 196 } 197 198 /* 199 * Obtain the buffer cache block. Figure out the buffer size 200 * when we are at EOF. If we are modifying the size of the 201 * buffer based on an EOF condition we need to hold 202 * nfs_rslock() through obtaining the buffer to prevent 203 * a potential writer-appender from messing with n_size. 204 * Otherwise we may accidently truncate the buffer and 205 * lose dirty data. 206 * 207 * Note that bcount is *not* DEV_BSIZE aligned. 208 */ 209 if (loffset + boff >= np->n_size) { 210 n = 0; 211 break; 212 } 213 bp = nfs_getcacheblk(vp, loffset, biosize, td); 214 215 if (bp == NULL) 216 return (EINTR); 217 218 /* 219 * If B_CACHE is not set, we must issue the read. If this 220 * fails, we return an error. 221 */ 222 if ((bp->b_flags & B_CACHE) == 0) { 223 bp->b_cmd = BUF_CMD_READ; 224 bp->b_bio2.bio_done = nfsiodone_sync; 225 bp->b_bio2.bio_flags |= BIO_SYNC; 226 vfs_busy_pages(vp, bp); 227 error = nfs_doio(vp, &bp->b_bio2, td); 228 if (error) { 229 brelse(bp); 230 return (error); 231 } 232 } 233 234 /* 235 * on is the offset into the current bp. Figure out how many 236 * bytes we can copy out of the bp. Note that bcount is 237 * NOT DEV_BSIZE aligned. 238 * 239 * Then figure out how many bytes we can copy into the uio. 240 */ 241 n = biosize - boff; 242 if (n > uio->uio_resid) 243 n = uio->uio_resid; 244 if (loffset + boff + n > np->n_size) 245 n = np->n_size - loffset - boff; 246 break; 247 case VLNK: 248 biosize = min(NFS_MAXPATHLEN, np->n_size); 249 nfsstats.biocache_readlinks++; 250 bp = nfs_getcacheblk(vp, (off_t)0, biosize, td); 251 if (bp == NULL) 252 return (EINTR); 253 if ((bp->b_flags & B_CACHE) == 0) { 254 bp->b_cmd = BUF_CMD_READ; 255 bp->b_bio2.bio_done = nfsiodone_sync; 256 bp->b_bio2.bio_flags |= BIO_SYNC; 257 vfs_busy_pages(vp, bp); 258 error = nfs_doio(vp, &bp->b_bio2, td); 259 if (error) { 260 bp->b_flags |= B_ERROR | B_INVAL; 261 brelse(bp); 262 return (error); 263 } 264 } 265 n = szmin(uio->uio_resid, (size_t)bp->b_bcount - bp->b_resid); 266 boff = 0; 267 break; 268 case VDIR: 269 nfsstats.biocache_readdirs++; 270 if (np->n_direofoffset && 271 uio->uio_offset >= np->n_direofoffset 272 ) { 273 return (0); 274 } 275 lbn = (uoff_t)uio->uio_offset / NFS_DIRBLKSIZ; 276 boff = uio->uio_offset & (NFS_DIRBLKSIZ - 1); 277 loffset = uio->uio_offset - boff; 278 bp = nfs_getcacheblk(vp, loffset, NFS_DIRBLKSIZ, td); 279 if (bp == NULL) 280 return (EINTR); 281 282 if ((bp->b_flags & B_CACHE) == 0) { 283 bp->b_cmd = BUF_CMD_READ; 284 bp->b_bio2.bio_done = nfsiodone_sync; 285 bp->b_bio2.bio_flags |= BIO_SYNC; 286 vfs_busy_pages(vp, bp); 287 error = nfs_doio(vp, &bp->b_bio2, td); 288 if (error) 289 brelse(bp); 290 while (error == NFSERR_BAD_COOKIE) { 291 kprintf("got bad cookie vp %p bp %p\n", vp, bp); 292 nfs_invaldir(vp); 293 error = nfs_vinvalbuf(vp, 0, 1); 294 /* 295 * Yuck! The directory has been modified on the 296 * server. The only way to get the block is by 297 * reading from the beginning to get all the 298 * offset cookies. 299 * 300 * Leave the last bp intact unless there is an error. 301 * Loop back up to the while if the error is another 302 * NFSERR_BAD_COOKIE (double yuch!). 303 */ 304 for (i = 0; i <= lbn && !error; i++) { 305 if (np->n_direofoffset 306 && (i * NFS_DIRBLKSIZ) >= np->n_direofoffset) 307 return (0); 308 bp = nfs_getcacheblk(vp, (off_t)i * NFS_DIRBLKSIZ, 309 NFS_DIRBLKSIZ, td); 310 if (!bp) 311 return (EINTR); 312 if ((bp->b_flags & B_CACHE) == 0) { 313 bp->b_cmd = BUF_CMD_READ; 314 bp->b_bio2.bio_done = nfsiodone_sync; 315 bp->b_bio2.bio_flags |= BIO_SYNC; 316 vfs_busy_pages(vp, bp); 317 error = nfs_doio(vp, &bp->b_bio2, td); 318 /* 319 * no error + B_INVAL == directory EOF, 320 * use the block. 321 */ 322 if (error == 0 && (bp->b_flags & B_INVAL)) 323 break; 324 } 325 /* 326 * An error will throw away the block and the 327 * for loop will break out. If no error and this 328 * is not the block we want, we throw away the 329 * block and go for the next one via the for loop. 330 */ 331 if (error || i < lbn) 332 brelse(bp); 333 } 334 } 335 /* 336 * The above while is repeated if we hit another cookie 337 * error. If we hit an error and it wasn't a cookie error, 338 * we give up. 339 */ 340 if (error) 341 return (error); 342 } 343 344 /* 345 * If not eof and read aheads are enabled, start one. 346 * (You need the current block first, so that you have the 347 * directory offset cookie of the next block.) 348 */ 349 if (nmp->nm_readahead > 0 && nfs_asyncok(nmp) && 350 (bp->b_flags & B_INVAL) == 0 && 351 (np->n_direofoffset == 0 || 352 loffset + NFS_DIRBLKSIZ < np->n_direofoffset) && 353 findblk(vp, loffset + NFS_DIRBLKSIZ, FINDBLK_TEST) == NULL 354 ) { 355 rabp = nfs_getcacheblk(vp, loffset + NFS_DIRBLKSIZ, 356 NFS_DIRBLKSIZ, td); 357 if (rabp) { 358 if ((rabp->b_flags & (B_CACHE|B_DELWRI)) == 0) { 359 rabp->b_cmd = BUF_CMD_READ; 360 vfs_busy_pages(vp, rabp); 361 nfs_asyncio(vp, &rabp->b_bio2); 362 } else { 363 brelse(rabp); 364 } 365 } 366 } 367 /* 368 * Unlike VREG files, whos buffer size ( bp->b_bcount ) is 369 * chopped for the EOF condition, we cannot tell how large 370 * NFS directories are going to be until we hit EOF. So 371 * an NFS directory buffer is *not* chopped to its EOF. Now, 372 * it just so happens that b_resid will effectively chop it 373 * to EOF. *BUT* this information is lost if the buffer goes 374 * away and is reconstituted into a B_CACHE state ( due to 375 * being VMIO ) later. So we keep track of the directory eof 376 * in np->n_direofoffset and chop it off as an extra step 377 * right here. 378 * 379 * NOTE: boff could already be beyond EOF. 380 */ 381 if ((size_t)boff > NFS_DIRBLKSIZ - bp->b_resid) { 382 n = 0; 383 } else { 384 n = szmin(uio->uio_resid, 385 NFS_DIRBLKSIZ - bp->b_resid - (size_t)boff); 386 } 387 if (np->n_direofoffset && 388 n > (size_t)(np->n_direofoffset - uio->uio_offset)) { 389 n = (size_t)(np->n_direofoffset - uio->uio_offset); 390 } 391 break; 392 default: 393 kprintf(" nfs_bioread: type %x unexpected\n",vp->v_type); 394 n = 0; 395 break; 396 }; 397 398 switch (vp->v_type) { 399 case VREG: 400 if (n > 0) 401 error = uiomovebp(bp, bp->b_data + boff, n, uio); 402 break; 403 case VLNK: 404 if (n > 0) 405 error = uiomovebp(bp, bp->b_data + boff, n, uio); 406 n = 0; 407 break; 408 case VDIR: 409 if (n > 0) { 410 off_t old_off = uio->uio_offset; 411 caddr_t cpos, epos; 412 struct nfs_dirent *dp; 413 414 /* 415 * We are casting cpos to nfs_dirent, it must be 416 * int-aligned. 417 */ 418 if (boff & 3) { 419 error = EINVAL; 420 break; 421 } 422 423 cpos = bp->b_data + boff; 424 epos = bp->b_data + boff + n; 425 while (cpos < epos && error == 0 && uio->uio_resid > 0) { 426 dp = (struct nfs_dirent *)cpos; 427 error = nfs_check_dirent(dp, (int)(epos - cpos)); 428 if (error) 429 break; 430 if (vop_write_dirent(&error, uio, dp->nfs_ino, 431 dp->nfs_type, dp->nfs_namlen, dp->nfs_name)) { 432 break; 433 } 434 cpos += dp->nfs_reclen; 435 } 436 n = 0; 437 if (error == 0) { 438 uio->uio_offset = old_off + cpos - 439 bp->b_data - boff; 440 } 441 } 442 break; 443 default: 444 kprintf(" nfs_bioread: type %x unexpected\n",vp->v_type); 445 } 446 if (bp) 447 brelse(bp); 448 } while (error == 0 && uio->uio_resid > 0 && n > 0); 449 return (error); 450 } 451 452 /* 453 * Userland can supply any 'seek' offset when reading a NFS directory. 454 * Validate the structure so we don't panic the kernel. Note that 455 * the element name is nul terminated and the nul is not included 456 * in nfs_namlen. 457 */ 458 static 459 int 460 nfs_check_dirent(struct nfs_dirent *dp, int maxlen) 461 { 462 int nfs_name_off = offsetof(struct nfs_dirent, nfs_name[0]); 463 464 if (nfs_name_off >= maxlen) 465 return (EINVAL); 466 if (dp->nfs_reclen < nfs_name_off || dp->nfs_reclen > maxlen) 467 return (EINVAL); 468 if (nfs_name_off + dp->nfs_namlen >= dp->nfs_reclen) 469 return (EINVAL); 470 if (dp->nfs_reclen & 3) 471 return (EINVAL); 472 return (0); 473 } 474 475 /* 476 * Vnode op for write using bio 477 * 478 * nfs_write(struct vnode *a_vp, struct uio *a_uio, int a_ioflag, 479 * struct ucred *a_cred) 480 */ 481 int 482 nfs_write(struct vop_write_args *ap) 483 { 484 struct uio *uio = ap->a_uio; 485 struct thread *td = uio->uio_td; 486 struct vnode *vp = ap->a_vp; 487 struct nfsnode *np = VTONFS(vp); 488 int ioflag = ap->a_ioflag; 489 struct buf *bp; 490 struct vattr vattr; 491 struct nfsmount *nmp = VFSTONFS(vp->v_mount); 492 off_t loffset; 493 int boff, bytes; 494 int error = 0; 495 int haverslock = 0; 496 int bcount; 497 int biosize; 498 int trivial; 499 500 #ifdef DIAGNOSTIC 501 if (uio->uio_rw != UIO_WRITE) 502 panic("nfs_write mode"); 503 if (uio->uio_segflg == UIO_USERSPACE && uio->uio_td != curthread) 504 panic("nfs_write proc"); 505 #endif 506 if (vp->v_type != VREG) 507 return (EIO); 508 509 lwkt_gettoken(&nmp->nm_token); 510 511 if (np->n_flag & NWRITEERR) { 512 np->n_flag &= ~NWRITEERR; 513 lwkt_reltoken(&nmp->nm_token); 514 return (np->n_error); 515 } 516 if ((nmp->nm_flag & NFSMNT_NFSV3) != 0 && 517 (nmp->nm_state & NFSSTA_GOTFSINFO) == 0) { 518 (void)nfs_fsinfo(nmp, vp, td); 519 } 520 521 /* 522 * Synchronously flush pending buffers if we are in synchronous 523 * mode or if we are appending. 524 */ 525 if (ioflag & (IO_APPEND | IO_SYNC)) { 526 if (np->n_flag & NLMODIFIED) { 527 np->n_attrstamp = 0; 528 error = nfs_flush(vp, MNT_WAIT, td, 0); 529 /* error = nfs_vinvalbuf(vp, V_SAVE, 1); */ 530 if (error) 531 goto done; 532 } 533 } 534 535 /* 536 * If IO_APPEND then load uio_offset. We restart here if we cannot 537 * get the append lock. 538 */ 539 restart: 540 if (ioflag & IO_APPEND) { 541 np->n_attrstamp = 0; 542 error = VOP_GETATTR(vp, &vattr); 543 if (error) 544 goto done; 545 uio->uio_offset = np->n_size; 546 } 547 548 if (uio->uio_offset < 0) { 549 error = EINVAL; 550 goto done; 551 } 552 if ((uio->uio_offset + uio->uio_resid) > nmp->nm_maxfilesize) { 553 error = EFBIG; 554 goto done; 555 } 556 if (uio->uio_resid == 0) { 557 error = 0; 558 goto done; 559 } 560 561 /* 562 * We need to obtain the rslock if we intend to modify np->n_size 563 * in order to guarentee the append point with multiple contending 564 * writers, to guarentee that no other appenders modify n_size 565 * while we are trying to obtain a truncated buffer (i.e. to avoid 566 * accidently truncating data written by another appender due to 567 * the race), and to ensure that the buffer is populated prior to 568 * our extending of the file. We hold rslock through the entire 569 * operation. 570 * 571 * Note that we do not synchronize the case where someone truncates 572 * the file while we are appending to it because attempting to lock 573 * this case may deadlock other parts of the system unexpectedly. 574 */ 575 if ((ioflag & IO_APPEND) || 576 uio->uio_offset + uio->uio_resid > np->n_size) { 577 switch(nfs_rslock(np)) { 578 case ENOLCK: 579 goto restart; 580 /* not reached */ 581 case EINTR: 582 case ERESTART: 583 error = EINTR; 584 goto done; 585 /* not reached */ 586 default: 587 break; 588 } 589 haverslock = 1; 590 } 591 592 /* 593 * Maybe this should be above the vnode op call, but so long as 594 * file servers have no limits, i don't think it matters 595 */ 596 if (td && td->td_proc && uio->uio_offset + uio->uio_resid > 597 td->td_proc->p_rlimit[RLIMIT_FSIZE].rlim_cur) { 598 lwpsignal(td->td_proc, td->td_lwp, SIGXFSZ); 599 if (haverslock) 600 nfs_rsunlock(np); 601 error = EFBIG; 602 goto done; 603 } 604 605 biosize = vp->v_mount->mnt_stat.f_iosize; 606 607 do { 608 nfsstats.biocache_writes++; 609 boff = uio->uio_offset & (biosize-1); 610 loffset = uio->uio_offset - boff; 611 bytes = (int)szmin((unsigned)(biosize - boff), uio->uio_resid); 612 again: 613 /* 614 * Handle direct append and file extension cases, calculate 615 * unaligned buffer size. When extending B_CACHE will be 616 * set if possible. See UIO_NOCOPY note below. 617 */ 618 if (uio->uio_offset + bytes > np->n_size) { 619 np->n_flag |= NLMODIFIED; 620 trivial = (uio->uio_segflg != UIO_NOCOPY && 621 uio->uio_offset <= np->n_size); 622 nfs_meta_setsize(vp, td, uio->uio_offset + bytes, 623 trivial); 624 } 625 bp = nfs_getcacheblk(vp, loffset, biosize, td); 626 if (bp == NULL) { 627 error = EINTR; 628 break; 629 } 630 631 /* 632 * Actual bytes in buffer which we care about 633 */ 634 if (loffset + biosize < np->n_size) 635 bcount = biosize; 636 else 637 bcount = (int)(np->n_size - loffset); 638 639 /* 640 * Avoid a read by setting B_CACHE where the data we 641 * intend to write covers the entire buffer. Note 642 * that the buffer may have been set to B_CACHE by 643 * nfs_meta_setsize() above or otherwise inherited the 644 * flag, but if B_CACHE isn't set the buffer may be 645 * uninitialized and must be zero'd to accomodate 646 * future seek+write's. 647 * 648 * See the comments in kern/vfs_bio.c's getblk() for 649 * more information. 650 * 651 * When doing a UIO_NOCOPY write the buffer is not 652 * overwritten and we cannot just set B_CACHE unconditionally 653 * for full-block writes. 654 */ 655 if (boff == 0 && bytes == biosize && 656 uio->uio_segflg != UIO_NOCOPY) { 657 bp->b_flags |= B_CACHE; 658 bp->b_flags &= ~(B_ERROR | B_INVAL); 659 } 660 661 /* 662 * b_resid may be set due to file EOF if we extended out. 663 * The NFS bio code will zero the difference anyway so 664 * just acknowledged the fact and set b_resid to 0. 665 */ 666 if ((bp->b_flags & B_CACHE) == 0) { 667 bp->b_cmd = BUF_CMD_READ; 668 bp->b_bio2.bio_done = nfsiodone_sync; 669 bp->b_bio2.bio_flags |= BIO_SYNC; 670 vfs_busy_pages(vp, bp); 671 error = nfs_doio(vp, &bp->b_bio2, td); 672 if (error) { 673 brelse(bp); 674 break; 675 } 676 bp->b_resid = 0; 677 } 678 np->n_flag |= NLMODIFIED; 679 680 /* 681 * If dirtyend exceeds file size, chop it down. This should 682 * not normally occur but there is an append race where it 683 * might occur XXX, so we log it. 684 * 685 * If the chopping creates a reverse-indexed or degenerate 686 * situation with dirtyoff/end, we 0 both of them. 687 */ 688 if (bp->b_dirtyend > bcount) { 689 kprintf("NFS append race @%08llx:%d\n", 690 (long long)bp->b_bio2.bio_offset, 691 bp->b_dirtyend - bcount); 692 bp->b_dirtyend = bcount; 693 } 694 695 if (bp->b_dirtyoff >= bp->b_dirtyend) 696 bp->b_dirtyoff = bp->b_dirtyend = 0; 697 698 /* 699 * If the new write will leave a contiguous dirty 700 * area, just update the b_dirtyoff and b_dirtyend, 701 * otherwise force a write rpc of the old dirty area. 702 * 703 * While it is possible to merge discontiguous writes due to 704 * our having a B_CACHE buffer ( and thus valid read data 705 * for the hole), we don't because it could lead to 706 * significant cache coherency problems with multiple clients, 707 * especially if locking is implemented later on. 708 * 709 * as an optimization we could theoretically maintain 710 * a linked list of discontinuous areas, but we would still 711 * have to commit them separately so there isn't much 712 * advantage to it except perhaps a bit of asynchronization. 713 */ 714 if (bp->b_dirtyend > 0 && 715 (boff > bp->b_dirtyend || 716 (boff + bytes) < bp->b_dirtyoff) 717 ) { 718 if (bwrite(bp) == EINTR) { 719 error = EINTR; 720 break; 721 } 722 goto again; 723 } 724 725 error = uiomovebp(bp, bp->b_data + boff, bytes, uio); 726 727 /* 728 * Since this block is being modified, it must be written 729 * again and not just committed. Since write clustering does 730 * not work for the stage 1 data write, only the stage 2 731 * commit rpc, we have to clear B_CLUSTEROK as well. 732 */ 733 bp->b_flags &= ~(B_NEEDCOMMIT | B_CLUSTEROK); 734 735 if (error) { 736 brelse(bp); 737 break; 738 } 739 740 /* 741 * Only update dirtyoff/dirtyend if not a degenerate 742 * condition. 743 * 744 * The underlying VM pages have been marked valid by 745 * virtue of acquiring the bp. Because the entire buffer 746 * is marked dirty we do not have to worry about cleaning 747 * out the related dirty bits (and wouldn't really know 748 * how to deal with byte ranges anyway) 749 */ 750 if (bytes) { 751 if (bp->b_dirtyend > 0) { 752 bp->b_dirtyoff = imin(boff, bp->b_dirtyoff); 753 bp->b_dirtyend = imax(boff + bytes, 754 bp->b_dirtyend); 755 } else { 756 bp->b_dirtyoff = boff; 757 bp->b_dirtyend = boff + bytes; 758 } 759 } 760 761 /* 762 * If the lease is non-cachable or IO_SYNC do bwrite(). 763 * 764 * IO_INVAL appears to be unused. The idea appears to be 765 * to turn off caching in this case. Very odd. XXX 766 * 767 * If nfs_async is set bawrite() will use an unstable write 768 * (build dirty bufs on the server), so we might as well 769 * push it out with bawrite(). If nfs_async is not set we 770 * use bdwrite() to cache dirty bufs on the client. 771 */ 772 if (ioflag & IO_SYNC) { 773 if (ioflag & IO_INVAL) 774 bp->b_flags |= B_NOCACHE; 775 error = bwrite(bp); 776 if (error) 777 break; 778 } else if (boff + bytes == biosize && nfs_async) { 779 bawrite(bp); 780 } else { 781 bdwrite(bp); 782 } 783 } while (uio->uio_resid > 0 && bytes > 0); 784 785 if (haverslock) 786 nfs_rsunlock(np); 787 788 done: 789 lwkt_reltoken(&nmp->nm_token); 790 return (error); 791 } 792 793 /* 794 * Get an nfs cache block. 795 * 796 * Allocate a new one if the block isn't currently in the cache 797 * and return the block marked busy. If the calling process is 798 * interrupted by a signal for an interruptible mount point, return 799 * NULL. 800 * 801 * The caller must carefully deal with the possible B_INVAL state of 802 * the buffer. nfs_startio() clears B_INVAL (and nfs_asyncio() clears it 803 * indirectly), so synchronous reads can be issued without worrying about 804 * the B_INVAL state. We have to be a little more careful when dealing 805 * with writes (see comments in nfs_write()) when extending a file past 806 * its EOF. 807 */ 808 static struct buf * 809 nfs_getcacheblk(struct vnode *vp, off_t loffset, int size, struct thread *td) 810 { 811 struct buf *bp; 812 struct mount *mp; 813 struct nfsmount *nmp; 814 815 mp = vp->v_mount; 816 nmp = VFSTONFS(mp); 817 818 if (nmp->nm_flag & NFSMNT_INT) { 819 bp = getblk(vp, loffset, size, GETBLK_PCATCH, 0); 820 while (bp == NULL) { 821 if (nfs_sigintr(nmp, NULL, td)) 822 return (NULL); 823 bp = getblk(vp, loffset, size, 0, 2 * hz); 824 } 825 } else { 826 bp = getblk(vp, loffset, size, 0, 0); 827 } 828 829 /* 830 * bio2, the 'device' layer. Since BIOs use 64 bit byte offsets 831 * now, no translation is necessary. 832 */ 833 bp->b_bio2.bio_offset = loffset; 834 return (bp); 835 } 836 837 /* 838 * Flush and invalidate all dirty buffers. If another process is already 839 * doing the flush, just wait for completion. 840 */ 841 int 842 nfs_vinvalbuf(struct vnode *vp, int flags, int intrflg) 843 { 844 struct nfsnode *np = VTONFS(vp); 845 struct nfsmount *nmp = VFSTONFS(vp->v_mount); 846 int error = 0, slpflag, slptimeo; 847 thread_t td = curthread; 848 849 if (vp->v_flag & VRECLAIMED) 850 return (0); 851 852 if ((nmp->nm_flag & NFSMNT_INT) == 0) 853 intrflg = 0; 854 if (intrflg) { 855 slpflag = PCATCH; 856 slptimeo = 2 * hz; 857 } else { 858 slpflag = 0; 859 slptimeo = 0; 860 } 861 /* 862 * First wait for any other process doing a flush to complete. 863 */ 864 while (np->n_flag & NFLUSHINPROG) { 865 np->n_flag |= NFLUSHWANT; 866 error = tsleep((caddr_t)&np->n_flag, 0, "nfsvinval", slptimeo); 867 if (error && intrflg && nfs_sigintr(nmp, NULL, td)) 868 return (EINTR); 869 } 870 871 /* 872 * Now, flush as required. 873 */ 874 np->n_flag |= NFLUSHINPROG; 875 error = vinvalbuf(vp, flags, slpflag, 0); 876 while (error) { 877 if (intrflg && nfs_sigintr(nmp, NULL, td)) { 878 np->n_flag &= ~NFLUSHINPROG; 879 if (np->n_flag & NFLUSHWANT) { 880 np->n_flag &= ~NFLUSHWANT; 881 wakeup((caddr_t)&np->n_flag); 882 } 883 return (EINTR); 884 } 885 error = vinvalbuf(vp, flags, 0, slptimeo); 886 } 887 np->n_flag &= ~(NLMODIFIED | NFLUSHINPROG); 888 if (np->n_flag & NFLUSHWANT) { 889 np->n_flag &= ~NFLUSHWANT; 890 wakeup((caddr_t)&np->n_flag); 891 } 892 return (0); 893 } 894 895 /* 896 * Return true (non-zero) if the txthread and rxthread are operational 897 * and we do not already have too many not-yet-started BIO's built up. 898 */ 899 int 900 nfs_asyncok(struct nfsmount *nmp) 901 { 902 return (nmp->nm_bioqlen < nfs_maxasyncbio && 903 nmp->nm_bioqlen < nmp->nm_maxasync_scaled / NFS_ASYSCALE && 904 nmp->nm_rxstate <= NFSSVC_PENDING && 905 nmp->nm_txstate <= NFSSVC_PENDING); 906 } 907 908 /* 909 * The read-ahead code calls this to queue a bio to the txthread. 910 * 911 * We don't touch the bio otherwise... that is, we do not even 912 * construct or send the initial rpc. The txthread will do it 913 * for us. 914 * 915 * NOTE! nm_bioqlen is not decremented until the request completes, 916 * so it does not reflect the number of bio's on bioq. 917 */ 918 void 919 nfs_asyncio(struct vnode *vp, struct bio *bio) 920 { 921 struct buf *bp = bio->bio_buf; 922 struct nfsmount *nmp = VFSTONFS(vp->v_mount); 923 924 KKASSERT(vp->v_tag == VT_NFS); 925 BUF_KERNPROC(bp); 926 927 /* 928 * Shortcut swap cache (not done automatically because we are not 929 * using bread()). 930 */ 931 if (vn_cache_strategy(vp, bio)) 932 return; 933 934 bio->bio_driver_info = vp; 935 crit_enter(); 936 TAILQ_INSERT_TAIL(&nmp->nm_bioq, bio, bio_act); 937 atomic_add_int(&nmp->nm_bioqlen, 1); 938 crit_exit(); 939 nfssvc_iod_writer_wakeup(nmp); 940 } 941 942 /* 943 * nfs_doio() - Execute a BIO operation synchronously. The BIO will be 944 * completed and its error returned. The caller is responsible 945 * for brelse()ing it. ONLY USE FOR BIO_SYNC IOs! Otherwise 946 * our error probe will be against an invalid pointer. 947 * 948 * nfs_startio()- Execute a BIO operation assynchronously. 949 * 950 * NOTE: nfs_asyncio() is used to initiate an asynchronous BIO operation, 951 * which basically just queues it to the txthread. nfs_startio() 952 * actually initiates the I/O AFTER it has gotten to the txthread. 953 * 954 * NOTE: td might be NULL. 955 * 956 * NOTE: Caller has already busied the I/O. 957 */ 958 void 959 nfs_startio(struct vnode *vp, struct bio *bio, struct thread *td) 960 { 961 struct buf *bp = bio->bio_buf; 962 963 KKASSERT(vp->v_tag == VT_NFS); 964 965 /* 966 * clear B_ERROR and B_INVAL state prior to initiating the I/O. We 967 * do this here so we do not have to do it in all the code that 968 * calls us. 969 */ 970 bp->b_flags &= ~(B_ERROR | B_INVAL); 971 972 KASSERT(bp->b_cmd != BUF_CMD_DONE, 973 ("nfs_doio: bp %p already marked done!", bp)); 974 975 if (bp->b_cmd == BUF_CMD_READ) { 976 switch (vp->v_type) { 977 case VREG: 978 nfsstats.read_bios++; 979 nfs_readrpc_bio(vp, bio); 980 break; 981 case VLNK: 982 #if 0 983 bio->bio_offset = 0; 984 nfsstats.readlink_bios++; 985 nfs_readlinkrpc_bio(vp, bio); 986 #else 987 nfs_doio(vp, bio, td); 988 #endif 989 break; 990 case VDIR: 991 /* 992 * NOTE: If nfs_readdirplusrpc_bio() is requested but 993 * not supported, it will chain to 994 * nfs_readdirrpc_bio(). 995 */ 996 #if 0 997 nfsstats.readdir_bios++; 998 uiop->uio_offset = bio->bio_offset; 999 if (nmp->nm_flag & NFSMNT_RDIRPLUS) 1000 nfs_readdirplusrpc_bio(vp, bio); 1001 else 1002 nfs_readdirrpc_bio(vp, bio); 1003 #else 1004 nfs_doio(vp, bio, td); 1005 #endif 1006 break; 1007 default: 1008 kprintf("nfs_doio: type %x unexpected\n",vp->v_type); 1009 bp->b_flags |= B_ERROR; 1010 bp->b_error = EINVAL; 1011 biodone(bio); 1012 break; 1013 } 1014 } else { 1015 /* 1016 * If we only need to commit, try to commit. If this fails 1017 * it will chain through to the write. Basically all the logic 1018 * in nfs_doio() is replicated. 1019 */ 1020 KKASSERT(bp->b_cmd == BUF_CMD_WRITE); 1021 if (bp->b_flags & B_NEEDCOMMIT) 1022 nfs_commitrpc_bio(vp, bio); 1023 else 1024 nfs_writerpc_bio(vp, bio); 1025 } 1026 } 1027 1028 int 1029 nfs_doio(struct vnode *vp, struct bio *bio, struct thread *td) 1030 { 1031 struct buf *bp = bio->bio_buf; 1032 struct uio *uiop; 1033 struct nfsnode *np; 1034 struct nfsmount *nmp; 1035 int error = 0; 1036 int iomode, must_commit; 1037 size_t n; 1038 struct uio uio; 1039 struct iovec io; 1040 1041 #if 0 1042 /* 1043 * Shortcut swap cache (not done automatically because we are not 1044 * using bread()). 1045 * 1046 * XXX The biowait is a hack until we can figure out how to stop a 1047 * biodone chain when a middle element is BIO_SYNC. BIO_SYNC is 1048 * set so the bp shouldn't get ripped out from under us. The only 1049 * use-cases are fully synchronous I/O cases. 1050 * 1051 * XXX This is having problems, give up for now. 1052 */ 1053 if (vn_cache_strategy(vp, bio)) { 1054 error = biowait(&bio->bio_buf->b_bio1, "nfsrsw"); 1055 return (error); 1056 } 1057 #endif 1058 1059 KKASSERT(vp->v_tag == VT_NFS); 1060 np = VTONFS(vp); 1061 nmp = VFSTONFS(vp->v_mount); 1062 uiop = &uio; 1063 uiop->uio_iov = &io; 1064 uiop->uio_iovcnt = 1; 1065 uiop->uio_segflg = UIO_SYSSPACE; 1066 uiop->uio_td = td; 1067 1068 /* 1069 * clear B_ERROR and B_INVAL state prior to initiating the I/O. We 1070 * do this here so we do not have to do it in all the code that 1071 * calls us. 1072 */ 1073 bp->b_flags &= ~(B_ERROR | B_INVAL); 1074 1075 KASSERT(bp->b_cmd != BUF_CMD_DONE, 1076 ("nfs_doio: bp %p already marked done!", bp)); 1077 1078 if (bp->b_cmd == BUF_CMD_READ) { 1079 io.iov_len = uiop->uio_resid = (size_t)bp->b_bcount; 1080 io.iov_base = bp->b_data; 1081 uiop->uio_rw = UIO_READ; 1082 1083 switch (vp->v_type) { 1084 case VREG: 1085 /* 1086 * When reading from a regular file zero-fill any residual. 1087 * Note that this residual has nothing to do with NFS short 1088 * reads, which nfs_readrpc_uio() will handle for us. 1089 * 1090 * We have to do this because when we are write extending 1091 * a file the server may not have the same notion of 1092 * filesize as we do. Our BIOs should already be sized 1093 * (b_bcount) to account for the file EOF. 1094 */ 1095 nfsstats.read_bios++; 1096 uiop->uio_offset = bio->bio_offset; 1097 error = nfs_readrpc_uio(vp, uiop); 1098 if (error == 0 && uiop->uio_resid) { 1099 n = (size_t)bp->b_bcount - uiop->uio_resid; 1100 bzero(bp->b_data + n, bp->b_bcount - n); 1101 uiop->uio_resid = 0; 1102 } 1103 if (td && td->td_proc && (vp->v_flag & VTEXT) && 1104 np->n_mtime != np->n_vattr.va_mtime.tv_sec) { 1105 uprintf("Process killed due to text file modification\n"); 1106 ksignal(td->td_proc, SIGKILL); 1107 } 1108 break; 1109 case VLNK: 1110 uiop->uio_offset = 0; 1111 nfsstats.readlink_bios++; 1112 error = nfs_readlinkrpc_uio(vp, uiop); 1113 break; 1114 case VDIR: 1115 nfsstats.readdir_bios++; 1116 uiop->uio_offset = bio->bio_offset; 1117 if (nmp->nm_flag & NFSMNT_RDIRPLUS) { 1118 error = nfs_readdirplusrpc_uio(vp, uiop); 1119 if (error == NFSERR_NOTSUPP) 1120 nmp->nm_flag &= ~NFSMNT_RDIRPLUS; 1121 } 1122 if ((nmp->nm_flag & NFSMNT_RDIRPLUS) == 0) 1123 error = nfs_readdirrpc_uio(vp, uiop); 1124 /* 1125 * end-of-directory sets B_INVAL but does not generate an 1126 * error. 1127 */ 1128 if (error == 0 && uiop->uio_resid == bp->b_bcount) 1129 bp->b_flags |= B_INVAL; 1130 break; 1131 default: 1132 kprintf("nfs_doio: type %x unexpected\n",vp->v_type); 1133 break; 1134 }; 1135 if (error) { 1136 bp->b_flags |= B_ERROR; 1137 bp->b_error = error; 1138 } 1139 bp->b_resid = uiop->uio_resid; 1140 } else { 1141 /* 1142 * If we only need to commit, try to commit. 1143 * 1144 * NOTE: The I/O has already been staged for the write and 1145 * its pages busied, so b_dirtyoff/end is valid. 1146 */ 1147 KKASSERT(bp->b_cmd == BUF_CMD_WRITE); 1148 if (bp->b_flags & B_NEEDCOMMIT) { 1149 int retv; 1150 off_t off; 1151 1152 off = bio->bio_offset + bp->b_dirtyoff; 1153 retv = nfs_commitrpc_uio(vp, off, 1154 bp->b_dirtyend - bp->b_dirtyoff, 1155 td); 1156 if (retv == 0) { 1157 bp->b_dirtyoff = bp->b_dirtyend = 0; 1158 bp->b_flags &= ~(B_NEEDCOMMIT | B_CLUSTEROK); 1159 bp->b_resid = 0; 1160 biodone(bio); 1161 return(0); 1162 } 1163 if (retv == NFSERR_STALEWRITEVERF) { 1164 nfs_clearcommit(vp->v_mount); 1165 } 1166 } 1167 1168 /* 1169 * Setup for actual write 1170 */ 1171 if (bio->bio_offset + bp->b_dirtyend > np->n_size) 1172 bp->b_dirtyend = np->n_size - bio->bio_offset; 1173 1174 if (bp->b_dirtyend > bp->b_dirtyoff) { 1175 io.iov_len = uiop->uio_resid = bp->b_dirtyend 1176 - bp->b_dirtyoff; 1177 uiop->uio_offset = bio->bio_offset + bp->b_dirtyoff; 1178 io.iov_base = (char *)bp->b_data + bp->b_dirtyoff; 1179 uiop->uio_rw = UIO_WRITE; 1180 nfsstats.write_bios++; 1181 1182 if ((bp->b_flags & (B_NEEDCOMMIT | B_NOCACHE | B_CLUSTER)) == 0) 1183 iomode = NFSV3WRITE_UNSTABLE; 1184 else 1185 iomode = NFSV3WRITE_FILESYNC; 1186 1187 must_commit = 0; 1188 error = nfs_writerpc_uio(vp, uiop, &iomode, &must_commit); 1189 1190 /* 1191 * We no longer try to use kern/vfs_bio's cluster code to 1192 * cluster commits, so B_CLUSTEROK is no longer set with 1193 * B_NEEDCOMMIT. The problem is that a vfs_busy_pages() 1194 * may have to clear B_NEEDCOMMIT if it finds underlying 1195 * pages have been redirtied through a memory mapping 1196 * and doing this on a clustered bp will probably cause 1197 * a panic, plus the flag in the underlying NFS bufs 1198 * making up the cluster bp will not be properly cleared. 1199 */ 1200 if (!error && iomode == NFSV3WRITE_UNSTABLE) { 1201 bp->b_flags |= B_NEEDCOMMIT; 1202 #if 0 1203 /* XXX do not enable commit clustering */ 1204 if (bp->b_dirtyoff == 0 1205 && bp->b_dirtyend == bp->b_bcount) 1206 bp->b_flags |= B_CLUSTEROK; 1207 #endif 1208 } else { 1209 bp->b_flags &= ~(B_NEEDCOMMIT | B_CLUSTEROK); 1210 } 1211 1212 /* 1213 * For an interrupted write, the buffer is still valid 1214 * and the write hasn't been pushed to the server yet, 1215 * so we can't set B_ERROR and report the interruption 1216 * by setting B_EINTR. For the async case, B_EINTR 1217 * is not relevant, so the rpc attempt is essentially 1218 * a noop. For the case of a V3 write rpc not being 1219 * committed to stable storage, the block is still 1220 * dirty and requires either a commit rpc or another 1221 * write rpc with iomode == NFSV3WRITE_FILESYNC before 1222 * the block is reused. This is indicated by setting 1223 * the B_DELWRI and B_NEEDCOMMIT flags. 1224 * 1225 * If the buffer is marked B_PAGING, it does not reside on 1226 * the vp's paging queues so we cannot call bdirty(). The 1227 * bp in this case is not an NFS cache block so we should 1228 * be safe. XXX 1229 */ 1230 if (error == EINTR 1231 || (!error && (bp->b_flags & B_NEEDCOMMIT))) { 1232 crit_enter(); 1233 bp->b_flags &= ~(B_INVAL|B_NOCACHE); 1234 if ((bp->b_flags & B_PAGING) == 0) 1235 bdirty(bp); 1236 if (error) 1237 bp->b_flags |= B_EINTR; 1238 crit_exit(); 1239 } else { 1240 if (error) { 1241 bp->b_flags |= B_ERROR; 1242 bp->b_error = np->n_error = error; 1243 np->n_flag |= NWRITEERR; 1244 } 1245 bp->b_dirtyoff = bp->b_dirtyend = 0; 1246 } 1247 if (must_commit) 1248 nfs_clearcommit(vp->v_mount); 1249 bp->b_resid = uiop->uio_resid; 1250 } else { 1251 bp->b_resid = 0; 1252 } 1253 } 1254 1255 /* 1256 * I/O was run synchronously, biodone() it and calculate the 1257 * error to return. 1258 */ 1259 biodone(bio); 1260 KKASSERT(bp->b_cmd == BUF_CMD_DONE); 1261 if (bp->b_flags & B_EINTR) 1262 return (EINTR); 1263 if (bp->b_flags & B_ERROR) 1264 return (bp->b_error ? bp->b_error : EIO); 1265 return (0); 1266 } 1267 1268 /* 1269 * Handle all truncation, write-extend, and ftruncate()-extend operations 1270 * on the NFS lcient side. 1271 * 1272 * We use the new API in kern/vfs_vm.c to perform these operations in a 1273 * VM-friendly way. With this API VM pages are properly zerod and pages 1274 * still mapped into the buffer straddling EOF are not invalidated. 1275 */ 1276 int 1277 nfs_meta_setsize(struct vnode *vp, struct thread *td, off_t nsize, int trivial) 1278 { 1279 struct nfsnode *np = VTONFS(vp); 1280 off_t osize; 1281 int biosize = vp->v_mount->mnt_stat.f_iosize; 1282 int error; 1283 1284 osize = np->n_size; 1285 np->n_size = nsize; 1286 1287 if (nsize < osize) { 1288 error = nvtruncbuf(vp, nsize, biosize, -1, 0); 1289 } else { 1290 error = nvextendbuf(vp, osize, nsize, 1291 biosize, biosize, -1, -1, 1292 trivial); 1293 } 1294 return(error); 1295 } 1296 1297 /* 1298 * Synchronous completion for nfs_doio. Call bpdone() with elseit=FALSE. 1299 * Caller is responsible for brelse()'ing the bp. 1300 */ 1301 static void 1302 nfsiodone_sync(struct bio *bio) 1303 { 1304 bio->bio_flags = 0; 1305 bpdone(bio->bio_buf, 0); 1306 } 1307 1308 /* 1309 * nfs read rpc - BIO version 1310 */ 1311 void 1312 nfs_readrpc_bio(struct vnode *vp, struct bio *bio) 1313 { 1314 struct buf *bp = bio->bio_buf; 1315 u_int32_t *tl; 1316 struct nfsmount *nmp; 1317 int error = 0, len, tsiz; 1318 struct nfsm_info *info; 1319 1320 info = kmalloc(sizeof(*info), M_NFSREQ, M_WAITOK); 1321 info->mrep = NULL; 1322 info->v3 = NFS_ISV3(vp); 1323 1324 nmp = VFSTONFS(vp->v_mount); 1325 tsiz = bp->b_bcount; 1326 KKASSERT(tsiz <= nmp->nm_rsize); 1327 if (bio->bio_offset + tsiz > nmp->nm_maxfilesize) { 1328 error = EFBIG; 1329 goto nfsmout; 1330 } 1331 nfsstats.rpccnt[NFSPROC_READ]++; 1332 len = tsiz; 1333 nfsm_reqhead(info, vp, NFSPROC_READ, 1334 NFSX_FH(info->v3) + NFSX_UNSIGNED * 3); 1335 ERROROUT(nfsm_fhtom(info, vp)); 1336 tl = nfsm_build(info, NFSX_UNSIGNED * 3); 1337 if (info->v3) { 1338 txdr_hyper(bio->bio_offset, tl); 1339 *(tl + 2) = txdr_unsigned(len); 1340 } else { 1341 *tl++ = txdr_unsigned(bio->bio_offset); 1342 *tl++ = txdr_unsigned(len); 1343 *tl = 0; 1344 } 1345 info->bio = bio; 1346 info->done = nfs_readrpc_bio_done; 1347 nfsm_request_bio(info, vp, NFSPROC_READ, NULL, 1348 nfs_vpcred(vp, ND_READ)); 1349 return; 1350 nfsmout: 1351 kfree(info, M_NFSREQ); 1352 bp->b_error = error; 1353 bp->b_flags |= B_ERROR; 1354 biodone(bio); 1355 } 1356 1357 static void 1358 nfs_readrpc_bio_done(nfsm_info_t info) 1359 { 1360 struct nfsmount *nmp = VFSTONFS(info->vp->v_mount); 1361 struct bio *bio = info->bio; 1362 struct buf *bp = bio->bio_buf; 1363 u_int32_t *tl; 1364 int attrflag; 1365 int retlen; 1366 int eof; 1367 int error = 0; 1368 1369 KKASSERT(info->state == NFSM_STATE_DONE); 1370 1371 lwkt_gettoken(&nmp->nm_token); 1372 1373 if (info->v3) { 1374 ERROROUT(nfsm_postop_attr(info, info->vp, &attrflag, 1375 NFS_LATTR_NOSHRINK)); 1376 NULLOUT(tl = nfsm_dissect(info, 2 * NFSX_UNSIGNED)); 1377 eof = fxdr_unsigned(int, *(tl + 1)); 1378 } else { 1379 ERROROUT(nfsm_loadattr(info, info->vp, NULL)); 1380 eof = 0; 1381 } 1382 NEGATIVEOUT(retlen = nfsm_strsiz(info, nmp->nm_rsize)); 1383 ERROROUT(nfsm_mtobio(info, bio, retlen)); 1384 m_freem(info->mrep); 1385 info->mrep = NULL; 1386 1387 /* 1388 * No error occured, if retlen is less then bcount and no EOF 1389 * and NFSv3 a zero-fill short read occured. 1390 * 1391 * For NFSv2 a short-read indicates EOF. 1392 */ 1393 if (retlen < bp->b_bcount && info->v3 && eof == 0) { 1394 bzero(bp->b_data + retlen, bp->b_bcount - retlen); 1395 retlen = bp->b_bcount; 1396 } 1397 1398 /* 1399 * If we hit an EOF we still zero-fill, but return the expected 1400 * b_resid anyway. This should normally not occur since async 1401 * BIOs are not used for read-before-write case. Races against 1402 * the server can cause it though and we don't want to leave 1403 * garbage in the buffer. 1404 */ 1405 if (retlen < bp->b_bcount) { 1406 bzero(bp->b_data + retlen, bp->b_bcount - retlen); 1407 } 1408 bp->b_resid = 0; 1409 /* bp->b_resid = bp->b_bcount - retlen; */ 1410 nfsmout: 1411 lwkt_reltoken(&nmp->nm_token); 1412 kfree(info, M_NFSREQ); 1413 if (error) { 1414 bp->b_error = error; 1415 bp->b_flags |= B_ERROR; 1416 } 1417 biodone(bio); 1418 } 1419 1420 /* 1421 * nfs write call - BIO version 1422 * 1423 * NOTE: Caller has already busied the I/O. 1424 */ 1425 void 1426 nfs_writerpc_bio(struct vnode *vp, struct bio *bio) 1427 { 1428 struct nfsmount *nmp = VFSTONFS(vp->v_mount); 1429 struct nfsnode *np = VTONFS(vp); 1430 struct buf *bp = bio->bio_buf; 1431 u_int32_t *tl; 1432 int len; 1433 int iomode; 1434 int error = 0; 1435 struct nfsm_info *info; 1436 off_t offset; 1437 1438 /* 1439 * Setup for actual write. Just clean up the bio if there 1440 * is nothing to do. b_dirtyoff/end have already been staged 1441 * by the bp's pages getting busied. 1442 */ 1443 if (bio->bio_offset + bp->b_dirtyend > np->n_size) 1444 bp->b_dirtyend = np->n_size - bio->bio_offset; 1445 1446 if (bp->b_dirtyend <= bp->b_dirtyoff) { 1447 bp->b_resid = 0; 1448 biodone(bio); 1449 return; 1450 } 1451 len = bp->b_dirtyend - bp->b_dirtyoff; 1452 offset = bio->bio_offset + bp->b_dirtyoff; 1453 if (offset + len > nmp->nm_maxfilesize) { 1454 bp->b_flags |= B_ERROR; 1455 bp->b_error = EFBIG; 1456 biodone(bio); 1457 return; 1458 } 1459 bp->b_resid = len; 1460 nfsstats.write_bios++; 1461 1462 info = kmalloc(sizeof(*info), M_NFSREQ, M_WAITOK); 1463 info->mrep = NULL; 1464 info->v3 = NFS_ISV3(vp); 1465 info->info_writerpc.must_commit = 0; 1466 if ((bp->b_flags & (B_NEEDCOMMIT | B_NOCACHE | B_CLUSTER)) == 0) 1467 iomode = NFSV3WRITE_UNSTABLE; 1468 else 1469 iomode = NFSV3WRITE_FILESYNC; 1470 1471 KKASSERT(len <= nmp->nm_wsize); 1472 1473 nfsstats.rpccnt[NFSPROC_WRITE]++; 1474 nfsm_reqhead(info, vp, NFSPROC_WRITE, 1475 NFSX_FH(info->v3) + 5 * NFSX_UNSIGNED + nfsm_rndup(len)); 1476 ERROROUT(nfsm_fhtom(info, vp)); 1477 if (info->v3) { 1478 tl = nfsm_build(info, 5 * NFSX_UNSIGNED); 1479 txdr_hyper(offset, tl); 1480 tl += 2; 1481 *tl++ = txdr_unsigned(len); 1482 *tl++ = txdr_unsigned(iomode); 1483 *tl = txdr_unsigned(len); 1484 } else { 1485 u_int32_t x; 1486 1487 tl = nfsm_build(info, 4 * NFSX_UNSIGNED); 1488 /* Set both "begin" and "current" to non-garbage. */ 1489 x = txdr_unsigned((u_int32_t)offset); 1490 *tl++ = x; /* "begin offset" */ 1491 *tl++ = x; /* "current offset" */ 1492 x = txdr_unsigned(len); 1493 *tl++ = x; /* total to this offset */ 1494 *tl = x; /* size of this write */ 1495 } 1496 ERROROUT(nfsm_biotom(info, bio, bp->b_dirtyoff, len)); 1497 info->bio = bio; 1498 info->done = nfs_writerpc_bio_done; 1499 nfsm_request_bio(info, vp, NFSPROC_WRITE, NULL, 1500 nfs_vpcred(vp, ND_WRITE)); 1501 return; 1502 nfsmout: 1503 kfree(info, M_NFSREQ); 1504 bp->b_error = error; 1505 bp->b_flags |= B_ERROR; 1506 biodone(bio); 1507 } 1508 1509 static void 1510 nfs_writerpc_bio_done(nfsm_info_t info) 1511 { 1512 struct nfsmount *nmp = VFSTONFS(info->vp->v_mount); 1513 struct nfsnode *np = VTONFS(info->vp); 1514 struct bio *bio = info->bio; 1515 struct buf *bp = bio->bio_buf; 1516 int wccflag = NFSV3_WCCRATTR; 1517 int iomode = NFSV3WRITE_FILESYNC; 1518 int commit; 1519 int rlen; 1520 int error; 1521 int len = bp->b_resid; /* b_resid was set to shortened length */ 1522 u_int32_t *tl; 1523 1524 lwkt_gettoken(&nmp->nm_token); 1525 1526 if (info->v3) { 1527 /* 1528 * The write RPC returns a before and after mtime. The 1529 * nfsm_wcc_data() macro checks the before n_mtime 1530 * against the before time and stores the after time 1531 * in the nfsnode's cached vattr and n_mtime field. 1532 * The NRMODIFIED bit will be set if the before 1533 * time did not match the original mtime. 1534 */ 1535 wccflag = NFSV3_WCCCHK; 1536 ERROROUT(nfsm_wcc_data(info, info->vp, &wccflag)); 1537 if (error == 0) { 1538 NULLOUT(tl = nfsm_dissect(info, 2 * NFSX_UNSIGNED + NFSX_V3WRITEVERF)); 1539 rlen = fxdr_unsigned(int, *tl++); 1540 if (rlen == 0) { 1541 error = NFSERR_IO; 1542 m_freem(info->mrep); 1543 info->mrep = NULL; 1544 goto nfsmout; 1545 } else if (rlen < len) { 1546 #if 0 1547 /* 1548 * XXX what do we do here? 1549 */ 1550 backup = len - rlen; 1551 uiop->uio_iov->iov_base = (char *)uiop->uio_iov->iov_base - backup; 1552 uiop->uio_iov->iov_len += backup; 1553 uiop->uio_offset -= backup; 1554 uiop->uio_resid += backup; 1555 len = rlen; 1556 #endif 1557 } 1558 commit = fxdr_unsigned(int, *tl++); 1559 1560 /* 1561 * Return the lowest committment level 1562 * obtained by any of the RPCs. 1563 */ 1564 if (iomode == NFSV3WRITE_FILESYNC) 1565 iomode = commit; 1566 else if (iomode == NFSV3WRITE_DATASYNC && 1567 commit == NFSV3WRITE_UNSTABLE) 1568 iomode = commit; 1569 if ((nmp->nm_state & NFSSTA_HASWRITEVERF) == 0){ 1570 bcopy(tl, (caddr_t)nmp->nm_verf, NFSX_V3WRITEVERF); 1571 nmp->nm_state |= NFSSTA_HASWRITEVERF; 1572 } else if (bcmp(tl, nmp->nm_verf, NFSX_V3WRITEVERF)) { 1573 info->info_writerpc.must_commit = 1; 1574 bcopy(tl, (caddr_t)nmp->nm_verf, NFSX_V3WRITEVERF); 1575 } 1576 } 1577 } else { 1578 ERROROUT(nfsm_loadattr(info, info->vp, NULL)); 1579 } 1580 m_freem(info->mrep); 1581 info->mrep = NULL; 1582 len = 0; 1583 nfsmout: 1584 if (info->vp->v_mount->mnt_flag & MNT_ASYNC) 1585 iomode = NFSV3WRITE_FILESYNC; 1586 bp->b_resid = len; 1587 1588 /* 1589 * End of RPC. Now clean up the bp. 1590 * 1591 * We no longer enable write clustering for commit operations, 1592 * See around line 1157 for a more detailed comment. 1593 */ 1594 if (!error && iomode == NFSV3WRITE_UNSTABLE) { 1595 bp->b_flags |= B_NEEDCOMMIT; 1596 #if 0 1597 /* XXX do not enable commit clustering */ 1598 if (bp->b_dirtyoff == 0 && bp->b_dirtyend == bp->b_bcount) 1599 bp->b_flags |= B_CLUSTEROK; 1600 #endif 1601 } else { 1602 bp->b_flags &= ~(B_NEEDCOMMIT | B_CLUSTEROK); 1603 } 1604 1605 /* 1606 * For an interrupted write, the buffer is still valid 1607 * and the write hasn't been pushed to the server yet, 1608 * so we can't set B_ERROR and report the interruption 1609 * by setting B_EINTR. For the async case, B_EINTR 1610 * is not relevant, so the rpc attempt is essentially 1611 * a noop. For the case of a V3 write rpc not being 1612 * committed to stable storage, the block is still 1613 * dirty and requires either a commit rpc or another 1614 * write rpc with iomode == NFSV3WRITE_FILESYNC before 1615 * the block is reused. This is indicated by setting 1616 * the B_DELWRI and B_NEEDCOMMIT flags. 1617 * 1618 * If the buffer is marked B_PAGING, it does not reside on 1619 * the vp's paging queues so we cannot call bdirty(). The 1620 * bp in this case is not an NFS cache block so we should 1621 * be safe. XXX 1622 */ 1623 if (error == EINTR || (!error && (bp->b_flags & B_NEEDCOMMIT))) { 1624 crit_enter(); 1625 bp->b_flags &= ~(B_INVAL|B_NOCACHE); 1626 if ((bp->b_flags & B_PAGING) == 0) 1627 bdirty(bp); 1628 if (error) 1629 bp->b_flags |= B_EINTR; 1630 crit_exit(); 1631 } else { 1632 if (error) { 1633 bp->b_flags |= B_ERROR; 1634 bp->b_error = np->n_error = error; 1635 np->n_flag |= NWRITEERR; 1636 } 1637 bp->b_dirtyoff = bp->b_dirtyend = 0; 1638 } 1639 if (info->info_writerpc.must_commit) 1640 nfs_clearcommit(info->vp->v_mount); 1641 lwkt_reltoken(&nmp->nm_token); 1642 1643 kfree(info, M_NFSREQ); 1644 if (error) { 1645 bp->b_flags |= B_ERROR; 1646 bp->b_error = error; 1647 } 1648 biodone(bio); 1649 } 1650 1651 /* 1652 * Nfs Version 3 commit rpc - BIO version 1653 * 1654 * This function issues the commit rpc and will chain to a write 1655 * rpc if necessary. 1656 */ 1657 void 1658 nfs_commitrpc_bio(struct vnode *vp, struct bio *bio) 1659 { 1660 struct nfsmount *nmp = VFSTONFS(vp->v_mount); 1661 struct buf *bp = bio->bio_buf; 1662 struct nfsm_info *info; 1663 int error = 0; 1664 u_int32_t *tl; 1665 1666 if ((nmp->nm_state & NFSSTA_HASWRITEVERF) == 0) { 1667 bp->b_dirtyoff = bp->b_dirtyend = 0; 1668 bp->b_flags &= ~(B_NEEDCOMMIT | B_CLUSTEROK); 1669 bp->b_resid = 0; 1670 biodone(bio); 1671 return; 1672 } 1673 1674 info = kmalloc(sizeof(*info), M_NFSREQ, M_WAITOK); 1675 info->mrep = NULL; 1676 info->v3 = 1; 1677 1678 nfsstats.rpccnt[NFSPROC_COMMIT]++; 1679 nfsm_reqhead(info, vp, NFSPROC_COMMIT, NFSX_FH(1)); 1680 ERROROUT(nfsm_fhtom(info, vp)); 1681 tl = nfsm_build(info, 3 * NFSX_UNSIGNED); 1682 txdr_hyper(bio->bio_offset + bp->b_dirtyoff, tl); 1683 tl += 2; 1684 *tl = txdr_unsigned(bp->b_dirtyend - bp->b_dirtyoff); 1685 info->bio = bio; 1686 info->done = nfs_commitrpc_bio_done; 1687 nfsm_request_bio(info, vp, NFSPROC_COMMIT, NULL, 1688 nfs_vpcred(vp, ND_WRITE)); 1689 return; 1690 nfsmout: 1691 /* 1692 * Chain to write RPC on (early) error 1693 */ 1694 kfree(info, M_NFSREQ); 1695 nfs_writerpc_bio(vp, bio); 1696 } 1697 1698 static void 1699 nfs_commitrpc_bio_done(nfsm_info_t info) 1700 { 1701 struct nfsmount *nmp = VFSTONFS(info->vp->v_mount); 1702 struct bio *bio = info->bio; 1703 struct buf *bp = bio->bio_buf; 1704 u_int32_t *tl; 1705 int wccflag = NFSV3_WCCRATTR; 1706 int error = 0; 1707 1708 lwkt_gettoken(&nmp->nm_token); 1709 1710 ERROROUT(nfsm_wcc_data(info, info->vp, &wccflag)); 1711 if (error == 0) { 1712 NULLOUT(tl = nfsm_dissect(info, NFSX_V3WRITEVERF)); 1713 if (bcmp(nmp->nm_verf, tl, NFSX_V3WRITEVERF)) { 1714 bcopy(tl, nmp->nm_verf, NFSX_V3WRITEVERF); 1715 error = NFSERR_STALEWRITEVERF; 1716 } 1717 } 1718 m_freem(info->mrep); 1719 info->mrep = NULL; 1720 1721 /* 1722 * On completion we must chain to a write bio if an 1723 * error occurred. 1724 */ 1725 nfsmout: 1726 if (error == 0) { 1727 bp->b_dirtyoff = bp->b_dirtyend = 0; 1728 bp->b_flags &= ~(B_NEEDCOMMIT | B_CLUSTEROK); 1729 bp->b_resid = 0; 1730 biodone(bio); 1731 } else { 1732 nfs_writerpc_bio(info->vp, bio); 1733 } 1734 kfree(info, M_NFSREQ); 1735 lwkt_reltoken(&nmp->nm_token); 1736 } 1737 1738