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: src/sys/nfs/nfs_bio.c,v 1.83.2.4 2002/12/29 18:19:53 dillon Exp $ 38 * $DragonFly: src/sys/vfs/nfs/nfs_bio.c,v 1.12 2004/02/13 18:52:35 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 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 61 #include "rpcv2.h" 62 #include "nfsproto.h" 63 #include "nfs.h" 64 #include "nfsmount.h" 65 #include "nqnfs.h" 66 #include "nfsnode.h" 67 68 static struct buf *nfs_getcacheblk (struct vnode *vp, daddr_t bn, int size, 69 struct thread *td); 70 71 extern int nfs_numasync; 72 extern int nfs_pbuf_freecnt; 73 extern struct nfsstats nfsstats; 74 75 /* 76 * Vnode op for VM getpages. 77 */ 78 int 79 nfs_getpages(ap) 80 struct vop_getpages_args /* { 81 struct vnode *a_vp; 82 vm_page_t *a_m; 83 int a_count; 84 int a_reqpage; 85 vm_ooffset_t a_offset; 86 } */ *ap; 87 { 88 struct thread *td = curthread; /* XXX */ 89 int i, error, nextoff, size, toff, count, npages; 90 struct uio uio; 91 struct iovec iov; 92 vm_offset_t kva; 93 struct buf *bp; 94 struct vnode *vp; 95 struct nfsmount *nmp; 96 vm_page_t *pages; 97 98 vp = ap->a_vp; 99 nmp = VFSTONFS(vp->v_mount); 100 pages = ap->a_m; 101 count = ap->a_count; 102 103 if (vp->v_object == NULL) { 104 printf("nfs_getpages: called with non-merged cache vnode??\n"); 105 return VM_PAGER_ERROR; 106 } 107 108 if ((nmp->nm_flag & NFSMNT_NFSV3) != 0 && 109 (nmp->nm_state & NFSSTA_GOTFSINFO) == 0) 110 (void)nfs_fsinfo(nmp, vp, td); 111 112 npages = btoc(count); 113 114 /* 115 * If the requested page is partially valid, just return it and 116 * allow the pager to zero-out the blanks. Partially valid pages 117 * can only occur at the file EOF. 118 */ 119 120 { 121 vm_page_t m = pages[ap->a_reqpage]; 122 123 if (m->valid != 0) { 124 /* handled by vm_fault now */ 125 /* vm_page_zero_invalid(m, TRUE); */ 126 for (i = 0; i < npages; ++i) { 127 if (i != ap->a_reqpage) 128 vnode_pager_freepage(pages[i]); 129 } 130 return(0); 131 } 132 } 133 134 /* 135 * We use only the kva address for the buffer, but this is extremely 136 * convienient and fast. 137 */ 138 bp = getpbuf(&nfs_pbuf_freecnt); 139 140 kva = (vm_offset_t) bp->b_data; 141 pmap_qenter(kva, pages, npages); 142 143 iov.iov_base = (caddr_t) kva; 144 iov.iov_len = count; 145 uio.uio_iov = &iov; 146 uio.uio_iovcnt = 1; 147 uio.uio_offset = IDX_TO_OFF(pages[0]->pindex); 148 uio.uio_resid = count; 149 uio.uio_segflg = UIO_SYSSPACE; 150 uio.uio_rw = UIO_READ; 151 uio.uio_td = td; 152 153 error = nfs_readrpc(vp, &uio); 154 pmap_qremove(kva, npages); 155 156 relpbuf(bp, &nfs_pbuf_freecnt); 157 158 if (error && (uio.uio_resid == count)) { 159 printf("nfs_getpages: error %d\n", error); 160 for (i = 0; i < npages; ++i) { 161 if (i != ap->a_reqpage) 162 vnode_pager_freepage(pages[i]); 163 } 164 return VM_PAGER_ERROR; 165 } 166 167 /* 168 * Calculate the number of bytes read and validate only that number 169 * of bytes. Note that due to pending writes, size may be 0. This 170 * does not mean that the remaining data is invalid! 171 */ 172 173 size = count - uio.uio_resid; 174 175 for (i = 0, toff = 0; i < npages; i++, toff = nextoff) { 176 vm_page_t m; 177 nextoff = toff + PAGE_SIZE; 178 m = pages[i]; 179 180 m->flags &= ~PG_ZERO; 181 182 if (nextoff <= size) { 183 /* 184 * Read operation filled an entire page 185 */ 186 m->valid = VM_PAGE_BITS_ALL; 187 vm_page_undirty(m); 188 } else if (size > toff) { 189 /* 190 * Read operation filled a partial page. 191 */ 192 m->valid = 0; 193 vm_page_set_validclean(m, 0, size - toff); 194 /* handled by vm_fault now */ 195 /* vm_page_zero_invalid(m, TRUE); */ 196 } else { 197 /* 198 * Read operation was short. If no error occured 199 * we may have hit a zero-fill section. We simply 200 * leave valid set to 0. 201 */ 202 ; 203 } 204 if (i != ap->a_reqpage) { 205 /* 206 * Whether or not to leave the page activated is up in 207 * the air, but we should put the page on a page queue 208 * somewhere (it already is in the object). Result: 209 * It appears that emperical results show that 210 * deactivating pages is best. 211 */ 212 213 /* 214 * Just in case someone was asking for this page we 215 * now tell them that it is ok to use. 216 */ 217 if (!error) { 218 if (m->flags & PG_WANTED) 219 vm_page_activate(m); 220 else 221 vm_page_deactivate(m); 222 vm_page_wakeup(m); 223 } else { 224 vnode_pager_freepage(m); 225 } 226 } 227 } 228 return 0; 229 } 230 231 /* 232 * Vnode op for VM putpages. 233 */ 234 int 235 nfs_putpages(ap) 236 struct vop_putpages_args /* { 237 struct vnode *a_vp; 238 vm_page_t *a_m; 239 int a_count; 240 int a_sync; 241 int *a_rtvals; 242 vm_ooffset_t a_offset; 243 } */ *ap; 244 { 245 struct thread *td = curthread; 246 struct uio uio; 247 struct iovec iov; 248 vm_offset_t kva; 249 struct buf *bp; 250 int iomode, must_commit, i, error, npages, count; 251 off_t offset; 252 int *rtvals; 253 struct vnode *vp; 254 struct nfsmount *nmp; 255 struct nfsnode *np; 256 vm_page_t *pages; 257 258 vp = ap->a_vp; 259 np = VTONFS(vp); 260 nmp = VFSTONFS(vp->v_mount); 261 pages = ap->a_m; 262 count = ap->a_count; 263 rtvals = ap->a_rtvals; 264 npages = btoc(count); 265 offset = IDX_TO_OFF(pages[0]->pindex); 266 267 if ((nmp->nm_flag & NFSMNT_NFSV3) != 0 && 268 (nmp->nm_state & NFSSTA_GOTFSINFO) == 0) 269 (void)nfs_fsinfo(nmp, vp, td); 270 271 for (i = 0; i < npages; i++) { 272 rtvals[i] = VM_PAGER_AGAIN; 273 } 274 275 /* 276 * When putting pages, do not extend file past EOF. 277 */ 278 279 if (offset + count > np->n_size) { 280 count = np->n_size - offset; 281 if (count < 0) 282 count = 0; 283 } 284 285 /* 286 * We use only the kva address for the buffer, but this is extremely 287 * convienient and fast. 288 */ 289 bp = getpbuf(&nfs_pbuf_freecnt); 290 291 kva = (vm_offset_t) bp->b_data; 292 pmap_qenter(kva, pages, npages); 293 294 iov.iov_base = (caddr_t) kva; 295 iov.iov_len = count; 296 uio.uio_iov = &iov; 297 uio.uio_iovcnt = 1; 298 uio.uio_offset = offset; 299 uio.uio_resid = count; 300 uio.uio_segflg = UIO_SYSSPACE; 301 uio.uio_rw = UIO_WRITE; 302 uio.uio_td = td; 303 304 if ((ap->a_sync & VM_PAGER_PUT_SYNC) == 0) 305 iomode = NFSV3WRITE_UNSTABLE; 306 else 307 iomode = NFSV3WRITE_FILESYNC; 308 309 error = nfs_writerpc(vp, &uio, &iomode, &must_commit); 310 311 pmap_qremove(kva, npages); 312 relpbuf(bp, &nfs_pbuf_freecnt); 313 314 if (!error) { 315 int nwritten = round_page(count - uio.uio_resid) / PAGE_SIZE; 316 for (i = 0; i < nwritten; i++) { 317 rtvals[i] = VM_PAGER_OK; 318 vm_page_undirty(pages[i]); 319 } 320 if (must_commit) 321 nfs_clearcommit(vp->v_mount); 322 } 323 return rtvals[0]; 324 } 325 326 /* 327 * Vnode op for read using bio 328 */ 329 int 330 nfs_bioread(struct vnode *vp, struct uio *uio, int ioflag) 331 { 332 struct nfsnode *np = VTONFS(vp); 333 int biosize, i; 334 struct buf *bp = 0, *rabp; 335 struct vattr vattr; 336 struct thread *td; 337 struct nfsmount *nmp = VFSTONFS(vp->v_mount); 338 daddr_t lbn, rabn; 339 int bcount; 340 int seqcount; 341 int nra, error = 0, n = 0, on = 0; 342 343 #ifdef DIAGNOSTIC 344 if (uio->uio_rw != UIO_READ) 345 panic("nfs_read mode"); 346 #endif 347 if (uio->uio_resid == 0) 348 return (0); 349 if (uio->uio_offset < 0) /* XXX VDIR cookies can be negative */ 350 return (EINVAL); 351 td = uio->uio_td; 352 353 if ((nmp->nm_flag & NFSMNT_NFSV3) != 0 && 354 (nmp->nm_state & NFSSTA_GOTFSINFO) == 0) 355 (void)nfs_fsinfo(nmp, vp, td); 356 if (vp->v_type != VDIR && 357 (uio->uio_offset + uio->uio_resid) > nmp->nm_maxfilesize) 358 return (EFBIG); 359 biosize = vp->v_mount->mnt_stat.f_iosize; 360 seqcount = (int)((off_t)(ioflag >> IO_SEQSHIFT) * biosize / BKVASIZE); 361 /* 362 * For nfs, cache consistency can only be maintained approximately. 363 * Although RFC1094 does not specify the criteria, the following is 364 * believed to be compatible with the reference port. 365 * For nqnfs, full cache consistency is maintained within the loop. 366 * For nfs: 367 * If the file's modify time on the server has changed since the 368 * last read rpc or you have written to the file, 369 * you may have lost data cache consistency with the 370 * server, so flush all of the file's data out of the cache. 371 * Then force a getattr rpc to ensure that you have up to date 372 * attributes. 373 * NB: This implies that cache data can be read when up to 374 * NFS_ATTRTIMEO seconds out of date. If you find that you need current 375 * attributes this could be forced by setting n_attrstamp to 0 before 376 * the VOP_GETATTR() call. 377 */ 378 if ((nmp->nm_flag & NFSMNT_NQNFS) == 0) { 379 if (np->n_flag & NMODIFIED) { 380 if (vp->v_type != VREG) { 381 if (vp->v_type != VDIR) 382 panic("nfs: bioread, not dir"); 383 nfs_invaldir(vp); 384 error = nfs_vinvalbuf(vp, V_SAVE, td, 1); 385 if (error) 386 return (error); 387 } 388 np->n_attrstamp = 0; 389 error = VOP_GETATTR(vp, &vattr, td); 390 if (error) 391 return (error); 392 np->n_mtime = vattr.va_mtime.tv_sec; 393 } else { 394 error = VOP_GETATTR(vp, &vattr, td); 395 if (error) 396 return (error); 397 if (np->n_mtime != vattr.va_mtime.tv_sec) { 398 if (vp->v_type == VDIR) 399 nfs_invaldir(vp); 400 error = nfs_vinvalbuf(vp, V_SAVE, td, 1); 401 if (error) 402 return (error); 403 np->n_mtime = vattr.va_mtime.tv_sec; 404 } 405 } 406 } 407 do { 408 409 /* 410 * Get a valid lease. If cached data is stale, flush it. 411 */ 412 if (nmp->nm_flag & NFSMNT_NQNFS) { 413 if (NQNFS_CKINVALID(vp, np, ND_READ)) { 414 do { 415 error = nqnfs_getlease(vp, ND_READ, td); 416 } while (error == NQNFS_EXPIRED); 417 if (error) 418 return (error); 419 if (np->n_lrev != np->n_brev || 420 (np->n_flag & NQNFSNONCACHE) || 421 ((np->n_flag & NMODIFIED) && vp->v_type == VDIR)) { 422 if (vp->v_type == VDIR) 423 nfs_invaldir(vp); 424 error = nfs_vinvalbuf(vp, V_SAVE, td, 1); 425 if (error) 426 return (error); 427 np->n_brev = np->n_lrev; 428 } 429 } else if (vp->v_type == VDIR && (np->n_flag & NMODIFIED)) { 430 nfs_invaldir(vp); 431 error = nfs_vinvalbuf(vp, V_SAVE, td, 1); 432 if (error) 433 return (error); 434 } 435 } 436 if (np->n_flag & NQNFSNONCACHE) { 437 switch (vp->v_type) { 438 case VREG: 439 return (nfs_readrpc(vp, uio)); 440 case VLNK: 441 return (nfs_readlinkrpc(vp, uio)); 442 case VDIR: 443 break; 444 default: 445 printf(" NQNFSNONCACHE: type %x unexpected\n", 446 vp->v_type); 447 }; 448 } 449 switch (vp->v_type) { 450 case VREG: 451 nfsstats.biocache_reads++; 452 lbn = uio->uio_offset / biosize; 453 on = uio->uio_offset & (biosize - 1); 454 455 /* 456 * Start the read ahead(s), as required. 457 */ 458 if (nfs_numasync > 0 && nmp->nm_readahead > 0) { 459 for (nra = 0; nra < nmp->nm_readahead && nra < seqcount && 460 (off_t)(lbn + 1 + nra) * biosize < np->n_size; nra++) { 461 rabn = lbn + 1 + nra; 462 if (!incore(vp, rabn)) { 463 rabp = nfs_getcacheblk(vp, rabn, biosize, td); 464 if (!rabp) 465 return (EINTR); 466 if ((rabp->b_flags & (B_CACHE|B_DELWRI)) == 0) { 467 rabp->b_flags |= (B_READ | B_ASYNC); 468 vfs_busy_pages(rabp, 0); 469 if (nfs_asyncio(rabp, td)) { 470 rabp->b_flags |= B_INVAL|B_ERROR; 471 vfs_unbusy_pages(rabp); 472 brelse(rabp); 473 break; 474 } 475 } else { 476 brelse(rabp); 477 } 478 } 479 } 480 } 481 482 /* 483 * Obtain the buffer cache block. Figure out the buffer size 484 * when we are at EOF. If we are modifying the size of the 485 * buffer based on an EOF condition we need to hold 486 * nfs_rslock() through obtaining the buffer to prevent 487 * a potential writer-appender from messing with n_size. 488 * Otherwise we may accidently truncate the buffer and 489 * lose dirty data. 490 * 491 * Note that bcount is *not* DEV_BSIZE aligned. 492 */ 493 494 again: 495 bcount = biosize; 496 if ((off_t)lbn * biosize >= np->n_size) { 497 bcount = 0; 498 } else if ((off_t)(lbn + 1) * biosize > np->n_size) { 499 bcount = np->n_size - (off_t)lbn * biosize; 500 } 501 if (bcount != biosize) { 502 switch(nfs_rslock(np, td)) { 503 case ENOLCK: 504 goto again; 505 /* not reached */ 506 case EINTR: 507 case ERESTART: 508 return(EINTR); 509 /* not reached */ 510 default: 511 break; 512 } 513 } 514 515 bp = nfs_getcacheblk(vp, lbn, bcount, td); 516 517 if (bcount != biosize) 518 nfs_rsunlock(np, td); 519 if (!bp) 520 return (EINTR); 521 522 /* 523 * If B_CACHE is not set, we must issue the read. If this 524 * fails, we return an error. 525 */ 526 527 if ((bp->b_flags & B_CACHE) == 0) { 528 bp->b_flags |= B_READ; 529 vfs_busy_pages(bp, 0); 530 error = nfs_doio(bp, td); 531 if (error) { 532 brelse(bp); 533 return (error); 534 } 535 } 536 537 /* 538 * on is the offset into the current bp. Figure out how many 539 * bytes we can copy out of the bp. Note that bcount is 540 * NOT DEV_BSIZE aligned. 541 * 542 * Then figure out how many bytes we can copy into the uio. 543 */ 544 545 n = 0; 546 if (on < bcount) 547 n = min((unsigned)(bcount - on), uio->uio_resid); 548 break; 549 case VLNK: 550 nfsstats.biocache_readlinks++; 551 bp = nfs_getcacheblk(vp, (daddr_t)0, NFS_MAXPATHLEN, td); 552 if (!bp) 553 return (EINTR); 554 if ((bp->b_flags & B_CACHE) == 0) { 555 bp->b_flags |= B_READ; 556 vfs_busy_pages(bp, 0); 557 error = nfs_doio(bp, td); 558 if (error) { 559 bp->b_flags |= B_ERROR; 560 brelse(bp); 561 return (error); 562 } 563 } 564 n = min(uio->uio_resid, NFS_MAXPATHLEN - bp->b_resid); 565 on = 0; 566 break; 567 case VDIR: 568 nfsstats.biocache_readdirs++; 569 if (np->n_direofoffset 570 && uio->uio_offset >= np->n_direofoffset) { 571 return (0); 572 } 573 lbn = (uoff_t)uio->uio_offset / NFS_DIRBLKSIZ; 574 on = uio->uio_offset & (NFS_DIRBLKSIZ - 1); 575 bp = nfs_getcacheblk(vp, lbn, NFS_DIRBLKSIZ, td); 576 if (!bp) 577 return (EINTR); 578 if ((bp->b_flags & B_CACHE) == 0) { 579 bp->b_flags |= B_READ; 580 vfs_busy_pages(bp, 0); 581 error = nfs_doio(bp, td); 582 if (error) { 583 brelse(bp); 584 } 585 while (error == NFSERR_BAD_COOKIE) { 586 printf("got bad cookie vp %p bp %p\n", vp, bp); 587 nfs_invaldir(vp); 588 error = nfs_vinvalbuf(vp, 0, td, 1); 589 /* 590 * Yuck! The directory has been modified on the 591 * server. The only way to get the block is by 592 * reading from the beginning to get all the 593 * offset cookies. 594 * 595 * Leave the last bp intact unless there is an error. 596 * Loop back up to the while if the error is another 597 * NFSERR_BAD_COOKIE (double yuch!). 598 */ 599 for (i = 0; i <= lbn && !error; i++) { 600 if (np->n_direofoffset 601 && (i * NFS_DIRBLKSIZ) >= np->n_direofoffset) 602 return (0); 603 bp = nfs_getcacheblk(vp, i, NFS_DIRBLKSIZ, td); 604 if (!bp) 605 return (EINTR); 606 if ((bp->b_flags & B_CACHE) == 0) { 607 bp->b_flags |= B_READ; 608 vfs_busy_pages(bp, 0); 609 error = nfs_doio(bp, td); 610 /* 611 * no error + B_INVAL == directory EOF, 612 * use the block. 613 */ 614 if (error == 0 && (bp->b_flags & B_INVAL)) 615 break; 616 } 617 /* 618 * An error will throw away the block and the 619 * for loop will break out. If no error and this 620 * is not the block we want, we throw away the 621 * block and go for the next one via the for loop. 622 */ 623 if (error || i < lbn) 624 brelse(bp); 625 } 626 } 627 /* 628 * The above while is repeated if we hit another cookie 629 * error. If we hit an error and it wasn't a cookie error, 630 * we give up. 631 */ 632 if (error) 633 return (error); 634 } 635 636 /* 637 * If not eof and read aheads are enabled, start one. 638 * (You need the current block first, so that you have the 639 * directory offset cookie of the next block.) 640 */ 641 if (nfs_numasync > 0 && nmp->nm_readahead > 0 && 642 (bp->b_flags & B_INVAL) == 0 && 643 (np->n_direofoffset == 0 || 644 (lbn + 1) * NFS_DIRBLKSIZ < np->n_direofoffset) && 645 !(np->n_flag & NQNFSNONCACHE) && 646 !incore(vp, lbn + 1)) { 647 rabp = nfs_getcacheblk(vp, lbn + 1, NFS_DIRBLKSIZ, td); 648 if (rabp) { 649 if ((rabp->b_flags & (B_CACHE|B_DELWRI)) == 0) { 650 rabp->b_flags |= (B_READ | B_ASYNC); 651 vfs_busy_pages(rabp, 0); 652 if (nfs_asyncio(rabp, td)) { 653 rabp->b_flags |= B_INVAL|B_ERROR; 654 vfs_unbusy_pages(rabp); 655 brelse(rabp); 656 } 657 } else { 658 brelse(rabp); 659 } 660 } 661 } 662 /* 663 * Unlike VREG files, whos buffer size ( bp->b_bcount ) is 664 * chopped for the EOF condition, we cannot tell how large 665 * NFS directories are going to be until we hit EOF. So 666 * an NFS directory buffer is *not* chopped to its EOF. Now, 667 * it just so happens that b_resid will effectively chop it 668 * to EOF. *BUT* this information is lost if the buffer goes 669 * away and is reconstituted into a B_CACHE state ( due to 670 * being VMIO ) later. So we keep track of the directory eof 671 * in np->n_direofoffset and chop it off as an extra step 672 * right here. 673 */ 674 n = lmin(uio->uio_resid, NFS_DIRBLKSIZ - bp->b_resid - on); 675 if (np->n_direofoffset && n > np->n_direofoffset - uio->uio_offset) 676 n = np->n_direofoffset - uio->uio_offset; 677 break; 678 default: 679 printf(" nfs_bioread: type %x unexpected\n",vp->v_type); 680 break; 681 }; 682 683 if (n > 0) { 684 error = uiomove(bp->b_data + on, (int)n, uio); 685 } 686 switch (vp->v_type) { 687 case VREG: 688 break; 689 case VLNK: 690 n = 0; 691 break; 692 case VDIR: 693 /* 694 * Invalidate buffer if caching is disabled, forcing a 695 * re-read from the remote later. 696 */ 697 if (np->n_flag & NQNFSNONCACHE) 698 bp->b_flags |= B_INVAL; 699 break; 700 default: 701 printf(" nfs_bioread: type %x unexpected\n",vp->v_type); 702 } 703 brelse(bp); 704 } while (error == 0 && uio->uio_resid > 0 && n > 0); 705 return (error); 706 } 707 708 /* 709 * Vnode op for write using bio 710 */ 711 int 712 nfs_write(ap) 713 struct vop_write_args /* { 714 struct vnode *a_vp; 715 struct uio *a_uio; 716 int a_ioflag; 717 struct ucred *a_cred; 718 } */ *ap; 719 { 720 int biosize; 721 struct uio *uio = ap->a_uio; 722 struct thread *td = uio->uio_td; 723 struct vnode *vp = ap->a_vp; 724 struct nfsnode *np = VTONFS(vp); 725 int ioflag = ap->a_ioflag; 726 struct buf *bp; 727 struct vattr vattr; 728 struct nfsmount *nmp = VFSTONFS(vp->v_mount); 729 daddr_t lbn; 730 int bcount; 731 int n, on, error = 0, iomode, must_commit; 732 int haverslock = 0; 733 734 #ifdef DIAGNOSTIC 735 if (uio->uio_rw != UIO_WRITE) 736 panic("nfs_write mode"); 737 if (uio->uio_segflg == UIO_USERSPACE && uio->uio_td != curthread) 738 panic("nfs_write proc"); 739 #endif 740 if (vp->v_type != VREG) 741 return (EIO); 742 if (np->n_flag & NWRITEERR) { 743 np->n_flag &= ~NWRITEERR; 744 return (np->n_error); 745 } 746 if ((nmp->nm_flag & NFSMNT_NFSV3) != 0 && 747 (nmp->nm_state & NFSSTA_GOTFSINFO) == 0) 748 (void)nfs_fsinfo(nmp, vp, td); 749 750 /* 751 * Synchronously flush pending buffers if we are in synchronous 752 * mode or if we are appending. 753 */ 754 if (ioflag & (IO_APPEND | IO_SYNC)) { 755 if (np->n_flag & NMODIFIED) { 756 np->n_attrstamp = 0; 757 error = nfs_vinvalbuf(vp, V_SAVE, td, 1); 758 if (error) 759 return (error); 760 } 761 } 762 763 /* 764 * If IO_APPEND then load uio_offset. We restart here if we cannot 765 * get the append lock. 766 */ 767 restart: 768 if (ioflag & IO_APPEND) { 769 np->n_attrstamp = 0; 770 error = VOP_GETATTR(vp, &vattr, td); 771 if (error) 772 return (error); 773 uio->uio_offset = np->n_size; 774 } 775 776 if (uio->uio_offset < 0) 777 return (EINVAL); 778 if ((uio->uio_offset + uio->uio_resid) > nmp->nm_maxfilesize) 779 return (EFBIG); 780 if (uio->uio_resid == 0) 781 return (0); 782 783 /* 784 * We need to obtain the rslock if we intend to modify np->n_size 785 * in order to guarentee the append point with multiple contending 786 * writers, to guarentee that no other appenders modify n_size 787 * while we are trying to obtain a truncated buffer (i.e. to avoid 788 * accidently truncating data written by another appender due to 789 * the race), and to ensure that the buffer is populated prior to 790 * our extending of the file. We hold rslock through the entire 791 * operation. 792 * 793 * Note that we do not synchronize the case where someone truncates 794 * the file while we are appending to it because attempting to lock 795 * this case may deadlock other parts of the system unexpectedly. 796 */ 797 if ((ioflag & IO_APPEND) || 798 uio->uio_offset + uio->uio_resid > np->n_size) { 799 switch(nfs_rslock(np, td)) { 800 case ENOLCK: 801 goto restart; 802 /* not reached */ 803 case EINTR: 804 case ERESTART: 805 return(EINTR); 806 /* not reached */ 807 default: 808 break; 809 } 810 haverslock = 1; 811 } 812 813 /* 814 * Maybe this should be above the vnode op call, but so long as 815 * file servers have no limits, i don't think it matters 816 */ 817 if (td->td_proc && uio->uio_offset + uio->uio_resid > 818 td->td_proc->p_rlimit[RLIMIT_FSIZE].rlim_cur) { 819 psignal(td->td_proc, SIGXFSZ); 820 if (haverslock) 821 nfs_rsunlock(np, td); 822 return (EFBIG); 823 } 824 825 biosize = vp->v_mount->mnt_stat.f_iosize; 826 827 do { 828 /* 829 * Check for a valid write lease. 830 */ 831 if ((nmp->nm_flag & NFSMNT_NQNFS) && 832 NQNFS_CKINVALID(vp, np, ND_WRITE)) { 833 do { 834 error = nqnfs_getlease(vp, ND_WRITE, td); 835 } while (error == NQNFS_EXPIRED); 836 if (error) 837 break; 838 if (np->n_lrev != np->n_brev || 839 (np->n_flag & NQNFSNONCACHE)) { 840 error = nfs_vinvalbuf(vp, V_SAVE, td, 1); 841 if (error) 842 break; 843 np->n_brev = np->n_lrev; 844 } 845 } 846 if ((np->n_flag & NQNFSNONCACHE) && uio->uio_iovcnt == 1) { 847 iomode = NFSV3WRITE_FILESYNC; 848 error = nfs_writerpc(vp, uio, &iomode, &must_commit); 849 if (must_commit) 850 nfs_clearcommit(vp->v_mount); 851 break; 852 } 853 nfsstats.biocache_writes++; 854 lbn = uio->uio_offset / biosize; 855 on = uio->uio_offset & (biosize-1); 856 n = min((unsigned)(biosize - on), uio->uio_resid); 857 again: 858 /* 859 * Handle direct append and file extension cases, calculate 860 * unaligned buffer size. 861 */ 862 863 if (uio->uio_offset == np->n_size && n) { 864 /* 865 * Get the buffer (in its pre-append state to maintain 866 * B_CACHE if it was previously set). Resize the 867 * nfsnode after we have locked the buffer to prevent 868 * readers from reading garbage. 869 */ 870 bcount = on; 871 bp = nfs_getcacheblk(vp, lbn, bcount, td); 872 873 if (bp != NULL) { 874 long save; 875 876 np->n_size = uio->uio_offset + n; 877 np->n_flag |= NMODIFIED; 878 vnode_pager_setsize(vp, np->n_size); 879 880 save = bp->b_flags & B_CACHE; 881 bcount += n; 882 allocbuf(bp, bcount); 883 bp->b_flags |= save; 884 } 885 } else { 886 /* 887 * Obtain the locked cache block first, and then 888 * adjust the file's size as appropriate. 889 */ 890 bcount = on + n; 891 if ((off_t)lbn * biosize + bcount < np->n_size) { 892 if ((off_t)(lbn + 1) * biosize < np->n_size) 893 bcount = biosize; 894 else 895 bcount = np->n_size - (off_t)lbn * biosize; 896 } 897 bp = nfs_getcacheblk(vp, lbn, bcount, td); 898 if (uio->uio_offset + n > np->n_size) { 899 np->n_size = uio->uio_offset + n; 900 np->n_flag |= NMODIFIED; 901 vnode_pager_setsize(vp, np->n_size); 902 } 903 } 904 905 if (!bp) { 906 error = EINTR; 907 break; 908 } 909 910 /* 911 * Issue a READ if B_CACHE is not set. In special-append 912 * mode, B_CACHE is based on the buffer prior to the write 913 * op and is typically set, avoiding the read. If a read 914 * is required in special append mode, the server will 915 * probably send us a short-read since we extended the file 916 * on our end, resulting in b_resid == 0 and, thusly, 917 * B_CACHE getting set. 918 * 919 * We can also avoid issuing the read if the write covers 920 * the entire buffer. We have to make sure the buffer state 921 * is reasonable in this case since we will not be initiating 922 * I/O. See the comments in kern/vfs_bio.c's getblk() for 923 * more information. 924 * 925 * B_CACHE may also be set due to the buffer being cached 926 * normally. 927 */ 928 929 if (on == 0 && n == bcount) { 930 bp->b_flags |= B_CACHE; 931 bp->b_flags &= ~(B_ERROR | B_INVAL); 932 } 933 934 if ((bp->b_flags & B_CACHE) == 0) { 935 bp->b_flags |= B_READ; 936 vfs_busy_pages(bp, 0); 937 error = nfs_doio(bp, td); 938 if (error) { 939 brelse(bp); 940 break; 941 } 942 } 943 if (!bp) { 944 error = EINTR; 945 break; 946 } 947 np->n_flag |= NMODIFIED; 948 949 /* 950 * If dirtyend exceeds file size, chop it down. This should 951 * not normally occur but there is an append race where it 952 * might occur XXX, so we log it. 953 * 954 * If the chopping creates a reverse-indexed or degenerate 955 * situation with dirtyoff/end, we 0 both of them. 956 */ 957 958 if (bp->b_dirtyend > bcount) { 959 printf("NFS append race @%lx:%d\n", 960 (long)bp->b_blkno * DEV_BSIZE, 961 bp->b_dirtyend - bcount); 962 bp->b_dirtyend = bcount; 963 } 964 965 if (bp->b_dirtyoff >= bp->b_dirtyend) 966 bp->b_dirtyoff = bp->b_dirtyend = 0; 967 968 /* 969 * If the new write will leave a contiguous dirty 970 * area, just update the b_dirtyoff and b_dirtyend, 971 * otherwise force a write rpc of the old dirty area. 972 * 973 * While it is possible to merge discontiguous writes due to 974 * our having a B_CACHE buffer ( and thus valid read data 975 * for the hole), we don't because it could lead to 976 * significant cache coherency problems with multiple clients, 977 * especially if locking is implemented later on. 978 * 979 * as an optimization we could theoretically maintain 980 * a linked list of discontinuous areas, but we would still 981 * have to commit them separately so there isn't much 982 * advantage to it except perhaps a bit of asynchronization. 983 */ 984 985 if (bp->b_dirtyend > 0 && 986 (on > bp->b_dirtyend || (on + n) < bp->b_dirtyoff)) { 987 if (VOP_BWRITE(bp->b_vp, bp) == EINTR) { 988 error = EINTR; 989 break; 990 } 991 goto again; 992 } 993 994 /* 995 * Check for valid write lease and get one as required. 996 * In case getblk() and/or bwrite() delayed us. 997 */ 998 if ((nmp->nm_flag & NFSMNT_NQNFS) && 999 NQNFS_CKINVALID(vp, np, ND_WRITE)) { 1000 do { 1001 error = nqnfs_getlease(vp, ND_WRITE, td); 1002 } while (error == NQNFS_EXPIRED); 1003 if (error) { 1004 brelse(bp); 1005 break; 1006 } 1007 if (np->n_lrev != np->n_brev || 1008 (np->n_flag & NQNFSNONCACHE)) { 1009 brelse(bp); 1010 error = nfs_vinvalbuf(vp, V_SAVE, td, 1); 1011 if (error) 1012 break; 1013 np->n_brev = np->n_lrev; 1014 goto again; 1015 } 1016 } 1017 1018 error = uiomove((char *)bp->b_data + on, n, uio); 1019 1020 /* 1021 * Since this block is being modified, it must be written 1022 * again and not just committed. Since write clustering does 1023 * not work for the stage 1 data write, only the stage 2 1024 * commit rpc, we have to clear B_CLUSTEROK as well. 1025 */ 1026 bp->b_flags &= ~(B_NEEDCOMMIT | B_CLUSTEROK); 1027 1028 if (error) { 1029 bp->b_flags |= B_ERROR; 1030 brelse(bp); 1031 break; 1032 } 1033 1034 /* 1035 * Only update dirtyoff/dirtyend if not a degenerate 1036 * condition. 1037 */ 1038 if (n) { 1039 if (bp->b_dirtyend > 0) { 1040 bp->b_dirtyoff = min(on, bp->b_dirtyoff); 1041 bp->b_dirtyend = max((on + n), bp->b_dirtyend); 1042 } else { 1043 bp->b_dirtyoff = on; 1044 bp->b_dirtyend = on + n; 1045 } 1046 vfs_bio_set_validclean(bp, on, n); 1047 } 1048 /* 1049 * If IO_NOWDRAIN then set B_NOWDRAIN (e.g. nfs-backed VN 1050 * filesystem). XXX also use for loopback NFS mounts. 1051 */ 1052 if (ioflag & IO_NOWDRAIN) 1053 bp->b_flags |= B_NOWDRAIN; 1054 1055 /* 1056 * If the lease is non-cachable or IO_SYNC do bwrite(). 1057 * 1058 * IO_INVAL appears to be unused. The idea appears to be 1059 * to turn off caching in this case. Very odd. XXX 1060 */ 1061 if ((np->n_flag & NQNFSNONCACHE) || (ioflag & IO_SYNC)) { 1062 if (ioflag & IO_INVAL) 1063 bp->b_flags |= B_NOCACHE; 1064 error = VOP_BWRITE(bp->b_vp, bp); 1065 if (error) 1066 break; 1067 if (np->n_flag & NQNFSNONCACHE) { 1068 error = nfs_vinvalbuf(vp, V_SAVE, td, 1); 1069 if (error) 1070 break; 1071 } 1072 } else if ((n + on) == biosize && 1073 (nmp->nm_flag & NFSMNT_NQNFS) == 0) { 1074 bp->b_flags |= B_ASYNC; 1075 (void)nfs_writebp(bp, 0, 0); 1076 } else { 1077 bdwrite(bp); 1078 } 1079 } while (uio->uio_resid > 0 && n > 0); 1080 1081 if (haverslock) 1082 nfs_rsunlock(np, td); 1083 1084 return (error); 1085 } 1086 1087 /* 1088 * Get an nfs cache block. 1089 * 1090 * Allocate a new one if the block isn't currently in the cache 1091 * and return the block marked busy. If the calling process is 1092 * interrupted by a signal for an interruptible mount point, return 1093 * NULL. 1094 * 1095 * The caller must carefully deal with the possible B_INVAL state of 1096 * the buffer. nfs_doio() clears B_INVAL (and nfs_asyncio() clears it 1097 * indirectly), so synchronous reads can be issued without worrying about 1098 * the B_INVAL state. We have to be a little more careful when dealing 1099 * with writes (see comments in nfs_write()) when extending a file past 1100 * its EOF. 1101 */ 1102 static struct buf * 1103 nfs_getcacheblk(struct vnode *vp, daddr_t bn, int size, struct thread *td) 1104 { 1105 struct buf *bp; 1106 struct mount *mp; 1107 struct nfsmount *nmp; 1108 1109 mp = vp->v_mount; 1110 nmp = VFSTONFS(mp); 1111 1112 if (nmp->nm_flag & NFSMNT_INT) { 1113 bp = getblk(vp, bn, size, PCATCH, 0); 1114 while (bp == (struct buf *)0) { 1115 if (nfs_sigintr(nmp, (struct nfsreq *)0, td)) 1116 return ((struct buf *)0); 1117 bp = getblk(vp, bn, size, 0, 2 * hz); 1118 } 1119 } else { 1120 bp = getblk(vp, bn, size, 0, 0); 1121 } 1122 1123 if (vp->v_type == VREG) { 1124 int biosize; 1125 1126 biosize = mp->mnt_stat.f_iosize; 1127 bp->b_blkno = bn * (biosize / DEV_BSIZE); 1128 } 1129 return (bp); 1130 } 1131 1132 /* 1133 * Flush and invalidate all dirty buffers. If another process is already 1134 * doing the flush, just wait for completion. 1135 */ 1136 int 1137 nfs_vinvalbuf(struct vnode *vp, int flags, 1138 struct thread *td, int intrflg) 1139 { 1140 struct nfsnode *np = VTONFS(vp); 1141 struct nfsmount *nmp = VFSTONFS(vp->v_mount); 1142 int error = 0, slpflag, slptimeo; 1143 1144 if (vp->v_flag & VXLOCK) { 1145 return (0); 1146 } 1147 1148 if ((nmp->nm_flag & NFSMNT_INT) == 0) 1149 intrflg = 0; 1150 if (intrflg) { 1151 slpflag = PCATCH; 1152 slptimeo = 2 * hz; 1153 } else { 1154 slpflag = 0; 1155 slptimeo = 0; 1156 } 1157 /* 1158 * First wait for any other process doing a flush to complete. 1159 */ 1160 while (np->n_flag & NFLUSHINPROG) { 1161 np->n_flag |= NFLUSHWANT; 1162 error = tsleep((caddr_t)&np->n_flag, 0, "nfsvinval", slptimeo); 1163 if (error && intrflg && nfs_sigintr(nmp, (struct nfsreq *)0, td)) 1164 return (EINTR); 1165 } 1166 1167 /* 1168 * Now, flush as required. 1169 */ 1170 np->n_flag |= NFLUSHINPROG; 1171 error = vinvalbuf(vp, flags, td, slpflag, 0); 1172 while (error) { 1173 if (intrflg && nfs_sigintr(nmp, (struct nfsreq *)0, td)) { 1174 np->n_flag &= ~NFLUSHINPROG; 1175 if (np->n_flag & NFLUSHWANT) { 1176 np->n_flag &= ~NFLUSHWANT; 1177 wakeup((caddr_t)&np->n_flag); 1178 } 1179 return (EINTR); 1180 } 1181 error = vinvalbuf(vp, flags, td, 0, slptimeo); 1182 } 1183 np->n_flag &= ~(NMODIFIED | NFLUSHINPROG); 1184 if (np->n_flag & NFLUSHWANT) { 1185 np->n_flag &= ~NFLUSHWANT; 1186 wakeup((caddr_t)&np->n_flag); 1187 } 1188 return (0); 1189 } 1190 1191 /* 1192 * Initiate asynchronous I/O. Return an error if no nfsiods are available. 1193 * This is mainly to avoid queueing async I/O requests when the nfsiods 1194 * are all hung on a dead server. 1195 * 1196 * Note: nfs_asyncio() does not clear (B_ERROR|B_INVAL) but when the bp 1197 * is eventually dequeued by the async daemon, nfs_doio() *will*. 1198 */ 1199 int 1200 nfs_asyncio(struct buf *bp, struct thread *td) 1201 { 1202 struct nfsmount *nmp; 1203 int i; 1204 int gotiod; 1205 int slpflag = 0; 1206 int slptimeo = 0; 1207 int error; 1208 1209 /* 1210 * If no async daemons then return EIO to force caller to run the rpc 1211 * synchronously. 1212 */ 1213 if (nfs_numasync == 0) 1214 return (EIO); 1215 1216 nmp = VFSTONFS(bp->b_vp->v_mount); 1217 1218 /* 1219 * Commits are usually short and sweet so lets save some cpu and 1220 * leave the async daemons for more important rpc's (such as reads 1221 * and writes). 1222 */ 1223 if ((bp->b_flags & (B_READ|B_NEEDCOMMIT)) == B_NEEDCOMMIT && 1224 (nmp->nm_bufqiods > nfs_numasync / 2)) { 1225 return(EIO); 1226 } 1227 1228 again: 1229 if (nmp->nm_flag & NFSMNT_INT) 1230 slpflag = PCATCH; 1231 gotiod = FALSE; 1232 1233 /* 1234 * Find a free iod to process this request. 1235 */ 1236 for (i = 0; i < NFS_MAXASYNCDAEMON; i++) 1237 if (nfs_iodwant[i]) { 1238 /* 1239 * Found one, so wake it up and tell it which 1240 * mount to process. 1241 */ 1242 NFS_DPF(ASYNCIO, 1243 ("nfs_asyncio: waking iod %d for mount %p\n", 1244 i, nmp)); 1245 nfs_iodwant[i] = NULL; 1246 nfs_iodmount[i] = nmp; 1247 nmp->nm_bufqiods++; 1248 wakeup((caddr_t)&nfs_iodwant[i]); 1249 gotiod = TRUE; 1250 break; 1251 } 1252 1253 /* 1254 * If none are free, we may already have an iod working on this mount 1255 * point. If so, it will process our request. 1256 */ 1257 if (!gotiod) { 1258 if (nmp->nm_bufqiods > 0) { 1259 NFS_DPF(ASYNCIO, 1260 ("nfs_asyncio: %d iods are already processing mount %p\n", 1261 nmp->nm_bufqiods, nmp)); 1262 gotiod = TRUE; 1263 } 1264 } 1265 1266 /* 1267 * If we have an iod which can process the request, then queue 1268 * the buffer. 1269 */ 1270 if (gotiod) { 1271 /* 1272 * Ensure that the queue never grows too large. We still want 1273 * to asynchronize so we block rather then return EIO. 1274 */ 1275 while (nmp->nm_bufqlen >= 2*nfs_numasync) { 1276 NFS_DPF(ASYNCIO, 1277 ("nfs_asyncio: waiting for mount %p queue to drain\n", nmp)); 1278 nmp->nm_bufqwant = TRUE; 1279 error = tsleep(&nmp->nm_bufq, slpflag, 1280 "nfsaio", slptimeo); 1281 if (error) { 1282 if (nfs_sigintr(nmp, NULL, td)) 1283 return (EINTR); 1284 if (slpflag == PCATCH) { 1285 slpflag = 0; 1286 slptimeo = 2 * hz; 1287 } 1288 } 1289 /* 1290 * We might have lost our iod while sleeping, 1291 * so check and loop if nescessary. 1292 */ 1293 if (nmp->nm_bufqiods == 0) { 1294 NFS_DPF(ASYNCIO, 1295 ("nfs_asyncio: no iods after mount %p queue was drained, looping\n", nmp)); 1296 goto again; 1297 } 1298 } 1299 1300 if ((bp->b_flags & B_READ) == 0) 1301 bp->b_flags |= B_WRITEINPROG; 1302 1303 BUF_KERNPROC(bp); 1304 TAILQ_INSERT_TAIL(&nmp->nm_bufq, bp, b_freelist); 1305 nmp->nm_bufqlen++; 1306 return (0); 1307 } 1308 1309 /* 1310 * All the iods are busy on other mounts, so return EIO to 1311 * force the caller to process the i/o synchronously. 1312 */ 1313 NFS_DPF(ASYNCIO, ("nfs_asyncio: no iods available, i/o is synchronous\n")); 1314 return (EIO); 1315 } 1316 1317 /* 1318 * Do an I/O operation to/from a cache block. This may be called 1319 * synchronously or from an nfsiod. 1320 * 1321 * NOTE! TD MIGHT BE NULL 1322 */ 1323 int 1324 nfs_doio(struct buf *bp, struct thread *td) 1325 { 1326 struct uio *uiop; 1327 struct vnode *vp; 1328 struct nfsnode *np; 1329 struct nfsmount *nmp; 1330 int error = 0, iomode, must_commit = 0; 1331 struct uio uio; 1332 struct iovec io; 1333 1334 vp = bp->b_vp; 1335 np = VTONFS(vp); 1336 nmp = VFSTONFS(vp->v_mount); 1337 uiop = &uio; 1338 uiop->uio_iov = &io; 1339 uiop->uio_iovcnt = 1; 1340 uiop->uio_segflg = UIO_SYSSPACE; 1341 uiop->uio_td = td; 1342 1343 /* 1344 * clear B_ERROR and B_INVAL state prior to initiating the I/O. We 1345 * do this here so we do not have to do it in all the code that 1346 * calls us. 1347 */ 1348 bp->b_flags &= ~(B_ERROR | B_INVAL); 1349 1350 KASSERT(!(bp->b_flags & B_DONE), ("nfs_doio: bp %p already marked done", bp)); 1351 1352 /* 1353 * Historically, paging was done with physio, but no more. 1354 */ 1355 if (bp->b_flags & B_PHYS) { 1356 /* 1357 * ...though reading /dev/drum still gets us here. 1358 */ 1359 io.iov_len = uiop->uio_resid = bp->b_bcount; 1360 /* mapping was done by vmapbuf() */ 1361 io.iov_base = bp->b_data; 1362 uiop->uio_offset = ((off_t)bp->b_blkno) * DEV_BSIZE; 1363 if (bp->b_flags & B_READ) { 1364 uiop->uio_rw = UIO_READ; 1365 nfsstats.read_physios++; 1366 error = nfs_readrpc(vp, uiop); 1367 } else { 1368 int com; 1369 1370 iomode = NFSV3WRITE_DATASYNC; 1371 uiop->uio_rw = UIO_WRITE; 1372 nfsstats.write_physios++; 1373 error = nfs_writerpc(vp, uiop, &iomode, &com); 1374 } 1375 if (error) { 1376 bp->b_flags |= B_ERROR; 1377 bp->b_error = error; 1378 } 1379 } else if (bp->b_flags & B_READ) { 1380 io.iov_len = uiop->uio_resid = bp->b_bcount; 1381 io.iov_base = bp->b_data; 1382 uiop->uio_rw = UIO_READ; 1383 1384 switch (vp->v_type) { 1385 case VREG: 1386 uiop->uio_offset = ((off_t)bp->b_blkno) * DEV_BSIZE; 1387 nfsstats.read_bios++; 1388 error = nfs_readrpc(vp, uiop); 1389 1390 if (!error) { 1391 if (uiop->uio_resid) { 1392 /* 1393 * If we had a short read with no error, we must have 1394 * hit a file hole. We should zero-fill the remainder. 1395 * This can also occur if the server hits the file EOF. 1396 * 1397 * Holes used to be able to occur due to pending 1398 * writes, but that is not possible any longer. 1399 */ 1400 int nread = bp->b_bcount - uiop->uio_resid; 1401 int left = uiop->uio_resid; 1402 1403 if (left > 0) 1404 bzero((char *)bp->b_data + nread, left); 1405 uiop->uio_resid = 0; 1406 } 1407 } 1408 if (td && td->td_proc && (vp->v_flag & VTEXT) && 1409 (((nmp->nm_flag & NFSMNT_NQNFS) && 1410 NQNFS_CKINVALID(vp, np, ND_READ) && 1411 np->n_lrev != np->n_brev) || 1412 (!(nmp->nm_flag & NFSMNT_NQNFS) && 1413 np->n_mtime != np->n_vattr.va_mtime.tv_sec))) { 1414 uprintf("Process killed due to text file modification\n"); 1415 psignal(td->td_proc, SIGKILL); 1416 PHOLD(td->td_proc); 1417 } 1418 break; 1419 case VLNK: 1420 uiop->uio_offset = (off_t)0; 1421 nfsstats.readlink_bios++; 1422 error = nfs_readlinkrpc(vp, uiop); 1423 break; 1424 case VDIR: 1425 nfsstats.readdir_bios++; 1426 uiop->uio_offset = ((u_quad_t)bp->b_lblkno) * NFS_DIRBLKSIZ; 1427 if (nmp->nm_flag & NFSMNT_RDIRPLUS) { 1428 error = nfs_readdirplusrpc(vp, uiop); 1429 if (error == NFSERR_NOTSUPP) 1430 nmp->nm_flag &= ~NFSMNT_RDIRPLUS; 1431 } 1432 if ((nmp->nm_flag & NFSMNT_RDIRPLUS) == 0) 1433 error = nfs_readdirrpc(vp, uiop); 1434 /* 1435 * end-of-directory sets B_INVAL but does not generate an 1436 * error. 1437 */ 1438 if (error == 0 && uiop->uio_resid == bp->b_bcount) 1439 bp->b_flags |= B_INVAL; 1440 break; 1441 default: 1442 printf("nfs_doio: type %x unexpected\n",vp->v_type); 1443 break; 1444 }; 1445 if (error) { 1446 bp->b_flags |= B_ERROR; 1447 bp->b_error = error; 1448 } 1449 } else { 1450 /* 1451 * If we only need to commit, try to commit 1452 */ 1453 if (bp->b_flags & B_NEEDCOMMIT) { 1454 int retv; 1455 off_t off; 1456 1457 off = ((u_quad_t)bp->b_blkno) * DEV_BSIZE + bp->b_dirtyoff; 1458 bp->b_flags |= B_WRITEINPROG; 1459 retv = nfs_commit(bp->b_vp, off, 1460 bp->b_dirtyend - bp->b_dirtyoff, td); 1461 bp->b_flags &= ~B_WRITEINPROG; 1462 if (retv == 0) { 1463 bp->b_dirtyoff = bp->b_dirtyend = 0; 1464 bp->b_flags &= ~(B_NEEDCOMMIT | B_CLUSTEROK); 1465 bp->b_resid = 0; 1466 biodone(bp); 1467 return (0); 1468 } 1469 if (retv == NFSERR_STALEWRITEVERF) { 1470 nfs_clearcommit(bp->b_vp->v_mount); 1471 } 1472 } 1473 1474 /* 1475 * Setup for actual write 1476 */ 1477 1478 if ((off_t)bp->b_blkno * DEV_BSIZE + bp->b_dirtyend > np->n_size) 1479 bp->b_dirtyend = np->n_size - (off_t)bp->b_blkno * DEV_BSIZE; 1480 1481 if (bp->b_dirtyend > bp->b_dirtyoff) { 1482 io.iov_len = uiop->uio_resid = bp->b_dirtyend 1483 - bp->b_dirtyoff; 1484 uiop->uio_offset = (off_t)bp->b_blkno * DEV_BSIZE 1485 + bp->b_dirtyoff; 1486 io.iov_base = (char *)bp->b_data + bp->b_dirtyoff; 1487 uiop->uio_rw = UIO_WRITE; 1488 nfsstats.write_bios++; 1489 1490 if ((bp->b_flags & (B_ASYNC | B_NEEDCOMMIT | B_NOCACHE | B_CLUSTER)) == B_ASYNC) 1491 iomode = NFSV3WRITE_UNSTABLE; 1492 else 1493 iomode = NFSV3WRITE_FILESYNC; 1494 1495 bp->b_flags |= B_WRITEINPROG; 1496 error = nfs_writerpc(vp, uiop, &iomode, &must_commit); 1497 1498 /* 1499 * When setting B_NEEDCOMMIT also set B_CLUSTEROK to try 1500 * to cluster the buffers needing commit. This will allow 1501 * the system to submit a single commit rpc for the whole 1502 * cluster. We can do this even if the buffer is not 100% 1503 * dirty (relative to the NFS blocksize), so we optimize the 1504 * append-to-file-case. 1505 * 1506 * (when clearing B_NEEDCOMMIT, B_CLUSTEROK must also be 1507 * cleared because write clustering only works for commit 1508 * rpc's, not for the data portion of the write). 1509 */ 1510 1511 if (!error && iomode == NFSV3WRITE_UNSTABLE) { 1512 bp->b_flags |= B_NEEDCOMMIT; 1513 if (bp->b_dirtyoff == 0 1514 && bp->b_dirtyend == bp->b_bcount) 1515 bp->b_flags |= B_CLUSTEROK; 1516 } else { 1517 bp->b_flags &= ~(B_NEEDCOMMIT | B_CLUSTEROK); 1518 } 1519 bp->b_flags &= ~B_WRITEINPROG; 1520 1521 /* 1522 * For an interrupted write, the buffer is still valid 1523 * and the write hasn't been pushed to the server yet, 1524 * so we can't set B_ERROR and report the interruption 1525 * by setting B_EINTR. For the B_ASYNC case, B_EINTR 1526 * is not relevant, so the rpc attempt is essentially 1527 * a noop. For the case of a V3 write rpc not being 1528 * committed to stable storage, the block is still 1529 * dirty and requires either a commit rpc or another 1530 * write rpc with iomode == NFSV3WRITE_FILESYNC before 1531 * the block is reused. This is indicated by setting 1532 * the B_DELWRI and B_NEEDCOMMIT flags. 1533 * 1534 * If the buffer is marked B_PAGING, it does not reside on 1535 * the vp's paging queues so we cannot call bdirty(). The 1536 * bp in this case is not an NFS cache block so we should 1537 * be safe. XXX 1538 */ 1539 if (error == EINTR 1540 || (!error && (bp->b_flags & B_NEEDCOMMIT))) { 1541 int s; 1542 1543 s = splbio(); 1544 bp->b_flags &= ~(B_INVAL|B_NOCACHE); 1545 if ((bp->b_flags & B_PAGING) == 0) { 1546 bdirty(bp); 1547 bp->b_flags &= ~B_DONE; 1548 } 1549 if (error && (bp->b_flags & B_ASYNC) == 0) 1550 bp->b_flags |= B_EINTR; 1551 splx(s); 1552 } else { 1553 if (error) { 1554 bp->b_flags |= B_ERROR; 1555 bp->b_error = np->n_error = error; 1556 np->n_flag |= NWRITEERR; 1557 } 1558 bp->b_dirtyoff = bp->b_dirtyend = 0; 1559 } 1560 } else { 1561 bp->b_resid = 0; 1562 biodone(bp); 1563 return (0); 1564 } 1565 } 1566 bp->b_resid = uiop->uio_resid; 1567 if (must_commit) 1568 nfs_clearcommit(vp->v_mount); 1569 biodone(bp); 1570 return (error); 1571 } 1572 1573 /* 1574 * Used to aid in handling ftruncate() operations on the NFS client side. 1575 * Truncation creates a number of special problems for NFS. We have to 1576 * throw away VM pages and buffer cache buffers that are beyond EOF, and 1577 * we have to properly handle VM pages or (potentially dirty) buffers 1578 * that straddle the truncation point. 1579 */ 1580 1581 int 1582 nfs_meta_setsize(struct vnode *vp, struct thread *td, u_quad_t nsize) 1583 { 1584 struct nfsnode *np = VTONFS(vp); 1585 u_quad_t tsize = np->n_size; 1586 int biosize = vp->v_mount->mnt_stat.f_iosize; 1587 int error = 0; 1588 1589 np->n_size = nsize; 1590 1591 if (np->n_size < tsize) { 1592 struct buf *bp; 1593 daddr_t lbn; 1594 int bufsize; 1595 1596 /* 1597 * vtruncbuf() doesn't get the buffer overlapping the 1598 * truncation point. We may have a B_DELWRI and/or B_CACHE 1599 * buffer that now needs to be truncated. 1600 */ 1601 error = vtruncbuf(vp, td, nsize, biosize); 1602 lbn = nsize / biosize; 1603 bufsize = nsize & (biosize - 1); 1604 bp = nfs_getcacheblk(vp, lbn, bufsize, td); 1605 if (bp->b_dirtyoff > bp->b_bcount) 1606 bp->b_dirtyoff = bp->b_bcount; 1607 if (bp->b_dirtyend > bp->b_bcount) 1608 bp->b_dirtyend = bp->b_bcount; 1609 bp->b_flags |= B_RELBUF; /* don't leave garbage around */ 1610 brelse(bp); 1611 } else { 1612 vnode_pager_setsize(vp, nsize); 1613 } 1614 return(error); 1615 } 1616 1617