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 1.11 2003/08/20 09:56:33 rob 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 ucred *cred; 255 struct nfsmount *nmp; 256 struct nfsnode *np; 257 vm_page_t *pages; 258 259 KKASSERT(td->td_proc); 260 cred = td->td_proc->p_ucred; 261 262 vp = ap->a_vp; 263 np = VTONFS(vp); 264 nmp = VFSTONFS(vp->v_mount); 265 pages = ap->a_m; 266 count = ap->a_count; 267 rtvals = ap->a_rtvals; 268 npages = btoc(count); 269 offset = IDX_TO_OFF(pages[0]->pindex); 270 271 if ((nmp->nm_flag & NFSMNT_NFSV3) != 0 && 272 (nmp->nm_state & NFSSTA_GOTFSINFO) == 0) 273 (void)nfs_fsinfo(nmp, vp, td); 274 275 for (i = 0; i < npages; i++) { 276 rtvals[i] = VM_PAGER_AGAIN; 277 } 278 279 /* 280 * When putting pages, do not extend file past EOF. 281 */ 282 283 if (offset + count > np->n_size) { 284 count = np->n_size - offset; 285 if (count < 0) 286 count = 0; 287 } 288 289 /* 290 * We use only the kva address for the buffer, but this is extremely 291 * convienient and fast. 292 */ 293 bp = getpbuf(&nfs_pbuf_freecnt); 294 295 kva = (vm_offset_t) bp->b_data; 296 pmap_qenter(kva, pages, npages); 297 298 iov.iov_base = (caddr_t) kva; 299 iov.iov_len = count; 300 uio.uio_iov = &iov; 301 uio.uio_iovcnt = 1; 302 uio.uio_offset = offset; 303 uio.uio_resid = count; 304 uio.uio_segflg = UIO_SYSSPACE; 305 uio.uio_rw = UIO_WRITE; 306 uio.uio_td = td; 307 308 if ((ap->a_sync & VM_PAGER_PUT_SYNC) == 0) 309 iomode = NFSV3WRITE_UNSTABLE; 310 else 311 iomode = NFSV3WRITE_FILESYNC; 312 313 error = nfs_writerpc(vp, &uio, &iomode, &must_commit); 314 315 pmap_qremove(kva, npages); 316 relpbuf(bp, &nfs_pbuf_freecnt); 317 318 if (!error) { 319 int nwritten = round_page(count - uio.uio_resid) / PAGE_SIZE; 320 for (i = 0; i < nwritten; i++) { 321 rtvals[i] = VM_PAGER_OK; 322 vm_page_undirty(pages[i]); 323 } 324 if (must_commit) 325 nfs_clearcommit(vp->v_mount); 326 } 327 return rtvals[0]; 328 } 329 330 /* 331 * Vnode op for read using bio 332 */ 333 int 334 nfs_bioread(struct vnode *vp, struct uio *uio, int ioflag) 335 { 336 struct nfsnode *np = VTONFS(vp); 337 int biosize, i; 338 struct buf *bp = 0, *rabp; 339 struct vattr vattr; 340 struct thread *td; 341 struct nfsmount *nmp = VFSTONFS(vp->v_mount); 342 daddr_t lbn, rabn; 343 int bcount; 344 int seqcount; 345 int nra, error = 0, n = 0, on = 0; 346 347 #ifdef DIAGNOSTIC 348 if (uio->uio_rw != UIO_READ) 349 panic("nfs_read mode"); 350 #endif 351 if (uio->uio_resid == 0) 352 return (0); 353 if (uio->uio_offset < 0) /* XXX VDIR cookies can be negative */ 354 return (EINVAL); 355 td = uio->uio_td; 356 357 if ((nmp->nm_flag & NFSMNT_NFSV3) != 0 && 358 (nmp->nm_state & NFSSTA_GOTFSINFO) == 0) 359 (void)nfs_fsinfo(nmp, vp, td); 360 if (vp->v_type != VDIR && 361 (uio->uio_offset + uio->uio_resid) > nmp->nm_maxfilesize) 362 return (EFBIG); 363 biosize = vp->v_mount->mnt_stat.f_iosize; 364 seqcount = (int)((off_t)(ioflag >> IO_SEQSHIFT) * biosize / BKVASIZE); 365 /* 366 * For nfs, cache consistency can only be maintained approximately. 367 * Although RFC1094 does not specify the criteria, the following is 368 * believed to be compatible with the reference port. 369 * For nqnfs, full cache consistency is maintained within the loop. 370 * For nfs: 371 * If the file's modify time on the server has changed since the 372 * last read rpc or you have written to the file, 373 * you may have lost data cache consistency with the 374 * server, so flush all of the file's data out of the cache. 375 * Then force a getattr rpc to ensure that you have up to date 376 * attributes. 377 * NB: This implies that cache data can be read when up to 378 * NFS_ATTRTIMEO seconds out of date. If you find that you need current 379 * attributes this could be forced by setting n_attrstamp to 0 before 380 * the VOP_GETATTR() call. 381 */ 382 if ((nmp->nm_flag & NFSMNT_NQNFS) == 0) { 383 if (np->n_flag & NMODIFIED) { 384 if (vp->v_type != VREG) { 385 if (vp->v_type != VDIR) 386 panic("nfs: bioread, not dir"); 387 nfs_invaldir(vp); 388 error = nfs_vinvalbuf(vp, V_SAVE, td, 1); 389 if (error) 390 return (error); 391 } 392 np->n_attrstamp = 0; 393 error = VOP_GETATTR(vp, &vattr, td); 394 if (error) 395 return (error); 396 np->n_mtime = vattr.va_mtime.tv_sec; 397 } else { 398 error = VOP_GETATTR(vp, &vattr, td); 399 if (error) 400 return (error); 401 if (np->n_mtime != vattr.va_mtime.tv_sec) { 402 if (vp->v_type == VDIR) 403 nfs_invaldir(vp); 404 error = nfs_vinvalbuf(vp, V_SAVE, td, 1); 405 if (error) 406 return (error); 407 np->n_mtime = vattr.va_mtime.tv_sec; 408 } 409 } 410 } 411 do { 412 413 /* 414 * Get a valid lease. If cached data is stale, flush it. 415 */ 416 if (nmp->nm_flag & NFSMNT_NQNFS) { 417 if (NQNFS_CKINVALID(vp, np, ND_READ)) { 418 do { 419 error = nqnfs_getlease(vp, ND_READ, td); 420 } while (error == NQNFS_EXPIRED); 421 if (error) 422 return (error); 423 if (np->n_lrev != np->n_brev || 424 (np->n_flag & NQNFSNONCACHE) || 425 ((np->n_flag & NMODIFIED) && vp->v_type == VDIR)) { 426 if (vp->v_type == VDIR) 427 nfs_invaldir(vp); 428 error = nfs_vinvalbuf(vp, V_SAVE, td, 1); 429 if (error) 430 return (error); 431 np->n_brev = np->n_lrev; 432 } 433 } else if (vp->v_type == VDIR && (np->n_flag & NMODIFIED)) { 434 nfs_invaldir(vp); 435 error = nfs_vinvalbuf(vp, V_SAVE, td, 1); 436 if (error) 437 return (error); 438 } 439 } 440 if (np->n_flag & NQNFSNONCACHE) { 441 switch (vp->v_type) { 442 case VREG: 443 return (nfs_readrpc(vp, uio)); 444 case VLNK: 445 return (nfs_readlinkrpc(vp, uio)); 446 case VDIR: 447 break; 448 default: 449 printf(" NQNFSNONCACHE: type %x unexpected\n", 450 vp->v_type); 451 }; 452 } 453 switch (vp->v_type) { 454 case VREG: 455 nfsstats.biocache_reads++; 456 lbn = uio->uio_offset / biosize; 457 on = uio->uio_offset & (biosize - 1); 458 459 /* 460 * Start the read ahead(s), as required. 461 */ 462 if (nfs_numasync > 0 && nmp->nm_readahead > 0) { 463 for (nra = 0; nra < nmp->nm_readahead && nra < seqcount && 464 (off_t)(lbn + 1 + nra) * biosize < np->n_size; nra++) { 465 rabn = lbn + 1 + nra; 466 if (!incore(vp, rabn)) { 467 rabp = nfs_getcacheblk(vp, rabn, biosize, td); 468 if (!rabp) 469 return (EINTR); 470 if ((rabp->b_flags & (B_CACHE|B_DELWRI)) == 0) { 471 rabp->b_flags |= (B_READ | B_ASYNC); 472 vfs_busy_pages(rabp, 0); 473 if (nfs_asyncio(rabp, td)) { 474 rabp->b_flags |= B_INVAL|B_ERROR; 475 vfs_unbusy_pages(rabp); 476 brelse(rabp); 477 break; 478 } 479 } else { 480 brelse(rabp); 481 } 482 } 483 } 484 } 485 486 /* 487 * Obtain the buffer cache block. Figure out the buffer size 488 * when we are at EOF. If we are modifying the size of the 489 * buffer based on an EOF condition we need to hold 490 * nfs_rslock() through obtaining the buffer to prevent 491 * a potential writer-appender from messing with n_size. 492 * Otherwise we may accidently truncate the buffer and 493 * lose dirty data. 494 * 495 * Note that bcount is *not* DEV_BSIZE aligned. 496 */ 497 498 again: 499 bcount = biosize; 500 if ((off_t)lbn * biosize >= np->n_size) { 501 bcount = 0; 502 } else if ((off_t)(lbn + 1) * biosize > np->n_size) { 503 bcount = np->n_size - (off_t)lbn * biosize; 504 } 505 if (bcount != biosize) { 506 switch(nfs_rslock(np, td)) { 507 case ENOLCK: 508 goto again; 509 /* not reached */ 510 case EINTR: 511 case ERESTART: 512 return(EINTR); 513 /* not reached */ 514 default: 515 break; 516 } 517 } 518 519 bp = nfs_getcacheblk(vp, lbn, bcount, td); 520 521 if (bcount != biosize) 522 nfs_rsunlock(np, td); 523 if (!bp) 524 return (EINTR); 525 526 /* 527 * If B_CACHE is not set, we must issue the read. If this 528 * fails, we return an error. 529 */ 530 531 if ((bp->b_flags & B_CACHE) == 0) { 532 bp->b_flags |= B_READ; 533 vfs_busy_pages(bp, 0); 534 error = nfs_doio(bp, td); 535 if (error) { 536 brelse(bp); 537 return (error); 538 } 539 } 540 541 /* 542 * on is the offset into the current bp. Figure out how many 543 * bytes we can copy out of the bp. Note that bcount is 544 * NOT DEV_BSIZE aligned. 545 * 546 * Then figure out how many bytes we can copy into the uio. 547 */ 548 549 n = 0; 550 if (on < bcount) 551 n = min((unsigned)(bcount - on), uio->uio_resid); 552 break; 553 case VLNK: 554 nfsstats.biocache_readlinks++; 555 bp = nfs_getcacheblk(vp, (daddr_t)0, NFS_MAXPATHLEN, td); 556 if (!bp) 557 return (EINTR); 558 if ((bp->b_flags & B_CACHE) == 0) { 559 bp->b_flags |= B_READ; 560 vfs_busy_pages(bp, 0); 561 error = nfs_doio(bp, td); 562 if (error) { 563 bp->b_flags |= B_ERROR; 564 brelse(bp); 565 return (error); 566 } 567 } 568 n = min(uio->uio_resid, NFS_MAXPATHLEN - bp->b_resid); 569 on = 0; 570 break; 571 case VDIR: 572 nfsstats.biocache_readdirs++; 573 if (np->n_direofoffset 574 && uio->uio_offset >= np->n_direofoffset) { 575 return (0); 576 } 577 lbn = (uoff_t)uio->uio_offset / NFS_DIRBLKSIZ; 578 on = uio->uio_offset & (NFS_DIRBLKSIZ - 1); 579 bp = nfs_getcacheblk(vp, lbn, NFS_DIRBLKSIZ, td); 580 if (!bp) 581 return (EINTR); 582 if ((bp->b_flags & B_CACHE) == 0) { 583 bp->b_flags |= B_READ; 584 vfs_busy_pages(bp, 0); 585 error = nfs_doio(bp, td); 586 if (error) { 587 brelse(bp); 588 } 589 while (error == NFSERR_BAD_COOKIE) { 590 printf("got bad cookie vp %p bp %p\n", vp, bp); 591 nfs_invaldir(vp); 592 error = nfs_vinvalbuf(vp, 0, td, 1); 593 /* 594 * Yuck! The directory has been modified on the 595 * server. The only way to get the block is by 596 * reading from the beginning to get all the 597 * offset cookies. 598 * 599 * Leave the last bp intact unless there is an error. 600 * Loop back up to the while if the error is another 601 * NFSERR_BAD_COOKIE (double yuch!). 602 */ 603 for (i = 0; i <= lbn && !error; i++) { 604 if (np->n_direofoffset 605 && (i * NFS_DIRBLKSIZ) >= np->n_direofoffset) 606 return (0); 607 bp = nfs_getcacheblk(vp, i, NFS_DIRBLKSIZ, td); 608 if (!bp) 609 return (EINTR); 610 if ((bp->b_flags & B_CACHE) == 0) { 611 bp->b_flags |= B_READ; 612 vfs_busy_pages(bp, 0); 613 error = nfs_doio(bp, td); 614 /* 615 * no error + B_INVAL == directory EOF, 616 * use the block. 617 */ 618 if (error == 0 && (bp->b_flags & B_INVAL)) 619 break; 620 } 621 /* 622 * An error will throw away the block and the 623 * for loop will break out. If no error and this 624 * is not the block we want, we throw away the 625 * block and go for the next one via the for loop. 626 */ 627 if (error || i < lbn) 628 brelse(bp); 629 } 630 } 631 /* 632 * The above while is repeated if we hit another cookie 633 * error. If we hit an error and it wasn't a cookie error, 634 * we give up. 635 */ 636 if (error) 637 return (error); 638 } 639 640 /* 641 * If not eof and read aheads are enabled, start one. 642 * (You need the current block first, so that you have the 643 * directory offset cookie of the next block.) 644 */ 645 if (nfs_numasync > 0 && nmp->nm_readahead > 0 && 646 (bp->b_flags & B_INVAL) == 0 && 647 (np->n_direofoffset == 0 || 648 (lbn + 1) * NFS_DIRBLKSIZ < np->n_direofoffset) && 649 !(np->n_flag & NQNFSNONCACHE) && 650 !incore(vp, lbn + 1)) { 651 rabp = nfs_getcacheblk(vp, lbn + 1, NFS_DIRBLKSIZ, td); 652 if (rabp) { 653 if ((rabp->b_flags & (B_CACHE|B_DELWRI)) == 0) { 654 rabp->b_flags |= (B_READ | B_ASYNC); 655 vfs_busy_pages(rabp, 0); 656 if (nfs_asyncio(rabp, td)) { 657 rabp->b_flags |= B_INVAL|B_ERROR; 658 vfs_unbusy_pages(rabp); 659 brelse(rabp); 660 } 661 } else { 662 brelse(rabp); 663 } 664 } 665 } 666 /* 667 * Unlike VREG files, whos buffer size ( bp->b_bcount ) is 668 * chopped for the EOF condition, we cannot tell how large 669 * NFS directories are going to be until we hit EOF. So 670 * an NFS directory buffer is *not* chopped to its EOF. Now, 671 * it just so happens that b_resid will effectively chop it 672 * to EOF. *BUT* this information is lost if the buffer goes 673 * away and is reconstituted into a B_CACHE state ( due to 674 * being VMIO ) later. So we keep track of the directory eof 675 * in np->n_direofoffset and chop it off as an extra step 676 * right here. 677 */ 678 n = lmin(uio->uio_resid, NFS_DIRBLKSIZ - bp->b_resid - on); 679 if (np->n_direofoffset && n > np->n_direofoffset - uio->uio_offset) 680 n = np->n_direofoffset - uio->uio_offset; 681 break; 682 default: 683 printf(" nfs_bioread: type %x unexpected\n",vp->v_type); 684 break; 685 }; 686 687 if (n > 0) { 688 error = uiomove(bp->b_data + on, (int)n, uio); 689 } 690 switch (vp->v_type) { 691 case VREG: 692 break; 693 case VLNK: 694 n = 0; 695 break; 696 case VDIR: 697 /* 698 * Invalidate buffer if caching is disabled, forcing a 699 * re-read from the remote later. 700 */ 701 if (np->n_flag & NQNFSNONCACHE) 702 bp->b_flags |= B_INVAL; 703 break; 704 default: 705 printf(" nfs_bioread: type %x unexpected\n",vp->v_type); 706 } 707 brelse(bp); 708 } while (error == 0 && uio->uio_resid > 0 && n > 0); 709 return (error); 710 } 711 712 /* 713 * Vnode op for write using bio 714 */ 715 int 716 nfs_write(ap) 717 struct vop_write_args /* { 718 struct vnode *a_vp; 719 struct uio *a_uio; 720 int a_ioflag; 721 struct ucred *a_cred; 722 } */ *ap; 723 { 724 int biosize; 725 struct uio *uio = ap->a_uio; 726 struct thread *td = uio->uio_td; 727 struct vnode *vp = ap->a_vp; 728 struct nfsnode *np = VTONFS(vp); 729 int ioflag = ap->a_ioflag; 730 struct buf *bp; 731 struct vattr vattr; 732 struct nfsmount *nmp = VFSTONFS(vp->v_mount); 733 daddr_t lbn; 734 int bcount; 735 int n, on, error = 0, iomode, must_commit; 736 int haverslock = 0; 737 738 #ifdef DIAGNOSTIC 739 if (uio->uio_rw != UIO_WRITE) 740 panic("nfs_write mode"); 741 if (uio->uio_segflg == UIO_USERSPACE && uio->uio_td != curthread) 742 panic("nfs_write proc"); 743 #endif 744 if (vp->v_type != VREG) 745 return (EIO); 746 if (np->n_flag & NWRITEERR) { 747 np->n_flag &= ~NWRITEERR; 748 return (np->n_error); 749 } 750 if ((nmp->nm_flag & NFSMNT_NFSV3) != 0 && 751 (nmp->nm_state & NFSSTA_GOTFSINFO) == 0) 752 (void)nfs_fsinfo(nmp, vp, td); 753 754 /* 755 * Synchronously flush pending buffers if we are in synchronous 756 * mode or if we are appending. 757 */ 758 if (ioflag & (IO_APPEND | IO_SYNC)) { 759 if (np->n_flag & NMODIFIED) { 760 np->n_attrstamp = 0; 761 error = nfs_vinvalbuf(vp, V_SAVE, td, 1); 762 if (error) 763 return (error); 764 } 765 } 766 767 /* 768 * If IO_APPEND then load uio_offset. We restart here if we cannot 769 * get the append lock. 770 */ 771 restart: 772 if (ioflag & IO_APPEND) { 773 np->n_attrstamp = 0; 774 error = VOP_GETATTR(vp, &vattr, td); 775 if (error) 776 return (error); 777 uio->uio_offset = np->n_size; 778 } 779 780 if (uio->uio_offset < 0) 781 return (EINVAL); 782 if ((uio->uio_offset + uio->uio_resid) > nmp->nm_maxfilesize) 783 return (EFBIG); 784 if (uio->uio_resid == 0) 785 return (0); 786 787 /* 788 * We need to obtain the rslock if we intend to modify np->n_size 789 * in order to guarentee the append point with multiple contending 790 * writers, to guarentee that no other appenders modify n_size 791 * while we are trying to obtain a truncated buffer (i.e. to avoid 792 * accidently truncating data written by another appender due to 793 * the race), and to ensure that the buffer is populated prior to 794 * our extending of the file. We hold rslock through the entire 795 * operation. 796 * 797 * Note that we do not synchronize the case where someone truncates 798 * the file while we are appending to it because attempting to lock 799 * this case may deadlock other parts of the system unexpectedly. 800 */ 801 if ((ioflag & IO_APPEND) || 802 uio->uio_offset + uio->uio_resid > np->n_size) { 803 switch(nfs_rslock(np, td)) { 804 case ENOLCK: 805 goto restart; 806 /* not reached */ 807 case EINTR: 808 case ERESTART: 809 return(EINTR); 810 /* not reached */ 811 default: 812 break; 813 } 814 haverslock = 1; 815 } 816 817 /* 818 * Maybe this should be above the vnode op call, but so long as 819 * file servers have no limits, i don't think it matters 820 */ 821 if (td->td_proc && uio->uio_offset + uio->uio_resid > 822 td->td_proc->p_rlimit[RLIMIT_FSIZE].rlim_cur) { 823 psignal(td->td_proc, SIGXFSZ); 824 if (haverslock) 825 nfs_rsunlock(np, td); 826 return (EFBIG); 827 } 828 829 biosize = vp->v_mount->mnt_stat.f_iosize; 830 831 do { 832 /* 833 * Check for a valid write lease. 834 */ 835 if ((nmp->nm_flag & NFSMNT_NQNFS) && 836 NQNFS_CKINVALID(vp, np, ND_WRITE)) { 837 do { 838 error = nqnfs_getlease(vp, ND_WRITE, td); 839 } while (error == NQNFS_EXPIRED); 840 if (error) 841 break; 842 if (np->n_lrev != np->n_brev || 843 (np->n_flag & NQNFSNONCACHE)) { 844 error = nfs_vinvalbuf(vp, V_SAVE, td, 1); 845 if (error) 846 break; 847 np->n_brev = np->n_lrev; 848 } 849 } 850 if ((np->n_flag & NQNFSNONCACHE) && uio->uio_iovcnt == 1) { 851 iomode = NFSV3WRITE_FILESYNC; 852 error = nfs_writerpc(vp, uio, &iomode, &must_commit); 853 if (must_commit) 854 nfs_clearcommit(vp->v_mount); 855 break; 856 } 857 nfsstats.biocache_writes++; 858 lbn = uio->uio_offset / biosize; 859 on = uio->uio_offset & (biosize-1); 860 n = min((unsigned)(biosize - on), uio->uio_resid); 861 again: 862 /* 863 * Handle direct append and file extension cases, calculate 864 * unaligned buffer size. 865 */ 866 867 if (uio->uio_offset == np->n_size && n) { 868 /* 869 * Get the buffer (in its pre-append state to maintain 870 * B_CACHE if it was previously set). Resize the 871 * nfsnode after we have locked the buffer to prevent 872 * readers from reading garbage. 873 */ 874 bcount = on; 875 bp = nfs_getcacheblk(vp, lbn, bcount, td); 876 877 if (bp != NULL) { 878 long save; 879 880 np->n_size = uio->uio_offset + n; 881 np->n_flag |= NMODIFIED; 882 vnode_pager_setsize(vp, np->n_size); 883 884 save = bp->b_flags & B_CACHE; 885 bcount += n; 886 allocbuf(bp, bcount); 887 bp->b_flags |= save; 888 } 889 } else { 890 /* 891 * Obtain the locked cache block first, and then 892 * adjust the file's size as appropriate. 893 */ 894 bcount = on + n; 895 if ((off_t)lbn * biosize + bcount < np->n_size) { 896 if ((off_t)(lbn + 1) * biosize < np->n_size) 897 bcount = biosize; 898 else 899 bcount = np->n_size - (off_t)lbn * biosize; 900 } 901 bp = nfs_getcacheblk(vp, lbn, bcount, td); 902 if (uio->uio_offset + n > np->n_size) { 903 np->n_size = uio->uio_offset + n; 904 np->n_flag |= NMODIFIED; 905 vnode_pager_setsize(vp, np->n_size); 906 } 907 } 908 909 if (!bp) { 910 error = EINTR; 911 break; 912 } 913 914 /* 915 * Issue a READ if B_CACHE is not set. In special-append 916 * mode, B_CACHE is based on the buffer prior to the write 917 * op and is typically set, avoiding the read. If a read 918 * is required in special append mode, the server will 919 * probably send us a short-read since we extended the file 920 * on our end, resulting in b_resid == 0 and, thusly, 921 * B_CACHE getting set. 922 * 923 * We can also avoid issuing the read if the write covers 924 * the entire buffer. We have to make sure the buffer state 925 * is reasonable in this case since we will not be initiating 926 * I/O. See the comments in kern/vfs_bio.c's getblk() for 927 * more information. 928 * 929 * B_CACHE may also be set due to the buffer being cached 930 * normally. 931 */ 932 933 if (on == 0 && n == bcount) { 934 bp->b_flags |= B_CACHE; 935 bp->b_flags &= ~(B_ERROR | B_INVAL); 936 } 937 938 if ((bp->b_flags & B_CACHE) == 0) { 939 bp->b_flags |= B_READ; 940 vfs_busy_pages(bp, 0); 941 error = nfs_doio(bp, td); 942 if (error) { 943 brelse(bp); 944 break; 945 } 946 } 947 if (!bp) { 948 error = EINTR; 949 break; 950 } 951 np->n_flag |= NMODIFIED; 952 953 /* 954 * If dirtyend exceeds file size, chop it down. This should 955 * not normally occur but there is an append race where it 956 * might occur XXX, so we log it. 957 * 958 * If the chopping creates a reverse-indexed or degenerate 959 * situation with dirtyoff/end, we 0 both of them. 960 */ 961 962 if (bp->b_dirtyend > bcount) { 963 printf("NFS append race @%lx:%d\n", 964 (long)bp->b_blkno * DEV_BSIZE, 965 bp->b_dirtyend - bcount); 966 bp->b_dirtyend = bcount; 967 } 968 969 if (bp->b_dirtyoff >= bp->b_dirtyend) 970 bp->b_dirtyoff = bp->b_dirtyend = 0; 971 972 /* 973 * If the new write will leave a contiguous dirty 974 * area, just update the b_dirtyoff and b_dirtyend, 975 * otherwise force a write rpc of the old dirty area. 976 * 977 * While it is possible to merge discontiguous writes due to 978 * our having a B_CACHE buffer ( and thus valid read data 979 * for the hole), we don't because it could lead to 980 * significant cache coherency problems with multiple clients, 981 * especially if locking is implemented later on. 982 * 983 * as an optimization we could theoretically maintain 984 * a linked list of discontinuous areas, but we would still 985 * have to commit them separately so there isn't much 986 * advantage to it except perhaps a bit of asynchronization. 987 */ 988 989 if (bp->b_dirtyend > 0 && 990 (on > bp->b_dirtyend || (on + n) < bp->b_dirtyoff)) { 991 if (VOP_BWRITE(bp->b_vp, bp) == EINTR) { 992 error = EINTR; 993 break; 994 } 995 goto again; 996 } 997 998 /* 999 * Check for valid write lease and get one as required. 1000 * In case getblk() and/or bwrite() delayed us. 1001 */ 1002 if ((nmp->nm_flag & NFSMNT_NQNFS) && 1003 NQNFS_CKINVALID(vp, np, ND_WRITE)) { 1004 do { 1005 error = nqnfs_getlease(vp, ND_WRITE, td); 1006 } while (error == NQNFS_EXPIRED); 1007 if (error) { 1008 brelse(bp); 1009 break; 1010 } 1011 if (np->n_lrev != np->n_brev || 1012 (np->n_flag & NQNFSNONCACHE)) { 1013 brelse(bp); 1014 error = nfs_vinvalbuf(vp, V_SAVE, td, 1); 1015 if (error) 1016 break; 1017 np->n_brev = np->n_lrev; 1018 goto again; 1019 } 1020 } 1021 1022 error = uiomove((char *)bp->b_data + on, n, uio); 1023 1024 /* 1025 * Since this block is being modified, it must be written 1026 * again and not just committed. Since write clustering does 1027 * not work for the stage 1 data write, only the stage 2 1028 * commit rpc, we have to clear B_CLUSTEROK as well. 1029 */ 1030 bp->b_flags &= ~(B_NEEDCOMMIT | B_CLUSTEROK); 1031 1032 if (error) { 1033 bp->b_flags |= B_ERROR; 1034 brelse(bp); 1035 break; 1036 } 1037 1038 /* 1039 * Only update dirtyoff/dirtyend if not a degenerate 1040 * condition. 1041 */ 1042 if (n) { 1043 if (bp->b_dirtyend > 0) { 1044 bp->b_dirtyoff = min(on, bp->b_dirtyoff); 1045 bp->b_dirtyend = max((on + n), bp->b_dirtyend); 1046 } else { 1047 bp->b_dirtyoff = on; 1048 bp->b_dirtyend = on + n; 1049 } 1050 vfs_bio_set_validclean(bp, on, n); 1051 } 1052 /* 1053 * If IO_NOWDRAIN then set B_NOWDRAIN (e.g. nfs-backed VN 1054 * filesystem). XXX also use for loopback NFS mounts. 1055 */ 1056 if (ioflag & IO_NOWDRAIN) 1057 bp->b_flags |= B_NOWDRAIN; 1058 1059 /* 1060 * If the lease is non-cachable or IO_SYNC do bwrite(). 1061 * 1062 * IO_INVAL appears to be unused. The idea appears to be 1063 * to turn off caching in this case. Very odd. XXX 1064 */ 1065 if ((np->n_flag & NQNFSNONCACHE) || (ioflag & IO_SYNC)) { 1066 if (ioflag & IO_INVAL) 1067 bp->b_flags |= B_NOCACHE; 1068 error = VOP_BWRITE(bp->b_vp, bp); 1069 if (error) 1070 break; 1071 if (np->n_flag & NQNFSNONCACHE) { 1072 error = nfs_vinvalbuf(vp, V_SAVE, td, 1); 1073 if (error) 1074 break; 1075 } 1076 } else if ((n + on) == biosize && 1077 (nmp->nm_flag & NFSMNT_NQNFS) == 0) { 1078 bp->b_flags |= B_ASYNC; 1079 (void)nfs_writebp(bp, 0, 0); 1080 } else { 1081 bdwrite(bp); 1082 } 1083 } while (uio->uio_resid > 0 && n > 0); 1084 1085 if (haverslock) 1086 nfs_rsunlock(np, td); 1087 1088 return (error); 1089 } 1090 1091 /* 1092 * Get an nfs cache block. 1093 * 1094 * Allocate a new one if the block isn't currently in the cache 1095 * and return the block marked busy. If the calling process is 1096 * interrupted by a signal for an interruptible mount point, return 1097 * NULL. 1098 * 1099 * The caller must carefully deal with the possible B_INVAL state of 1100 * the buffer. nfs_doio() clears B_INVAL (and nfs_asyncio() clears it 1101 * indirectly), so synchronous reads can be issued without worrying about 1102 * the B_INVAL state. We have to be a little more careful when dealing 1103 * with writes (see comments in nfs_write()) when extending a file past 1104 * its EOF. 1105 */ 1106 static struct buf * 1107 nfs_getcacheblk(struct vnode *vp, daddr_t bn, int size, struct thread *td) 1108 { 1109 struct buf *bp; 1110 struct mount *mp; 1111 struct nfsmount *nmp; 1112 1113 mp = vp->v_mount; 1114 nmp = VFSTONFS(mp); 1115 1116 if (nmp->nm_flag & NFSMNT_INT) { 1117 bp = getblk(vp, bn, size, PCATCH, 0); 1118 while (bp == (struct buf *)0) { 1119 if (nfs_sigintr(nmp, (struct nfsreq *)0, td)) 1120 return ((struct buf *)0); 1121 bp = getblk(vp, bn, size, 0, 2 * hz); 1122 } 1123 } else { 1124 bp = getblk(vp, bn, size, 0, 0); 1125 } 1126 1127 if (vp->v_type == VREG) { 1128 int biosize; 1129 1130 biosize = mp->mnt_stat.f_iosize; 1131 bp->b_blkno = bn * (biosize / DEV_BSIZE); 1132 } 1133 return (bp); 1134 } 1135 1136 /* 1137 * Flush and invalidate all dirty buffers. If another process is already 1138 * doing the flush, just wait for completion. 1139 */ 1140 int 1141 nfs_vinvalbuf(struct vnode *vp, int flags, 1142 struct thread *td, int intrflg) 1143 { 1144 struct nfsnode *np = VTONFS(vp); 1145 struct nfsmount *nmp = VFSTONFS(vp->v_mount); 1146 int error = 0, slpflag, slptimeo; 1147 1148 if (vp->v_flag & VXLOCK) { 1149 return (0); 1150 } 1151 1152 if ((nmp->nm_flag & NFSMNT_INT) == 0) 1153 intrflg = 0; 1154 if (intrflg) { 1155 slpflag = PCATCH; 1156 slptimeo = 2 * hz; 1157 } else { 1158 slpflag = 0; 1159 slptimeo = 0; 1160 } 1161 /* 1162 * First wait for any other process doing a flush to complete. 1163 */ 1164 while (np->n_flag & NFLUSHINPROG) { 1165 np->n_flag |= NFLUSHWANT; 1166 error = tsleep((caddr_t)&np->n_flag, 0, "nfsvinval", slptimeo); 1167 if (error && intrflg && nfs_sigintr(nmp, (struct nfsreq *)0, td)) 1168 return (EINTR); 1169 } 1170 1171 /* 1172 * Now, flush as required. 1173 */ 1174 np->n_flag |= NFLUSHINPROG; 1175 error = vinvalbuf(vp, flags, td, slpflag, 0); 1176 while (error) { 1177 if (intrflg && nfs_sigintr(nmp, (struct nfsreq *)0, td)) { 1178 np->n_flag &= ~NFLUSHINPROG; 1179 if (np->n_flag & NFLUSHWANT) { 1180 np->n_flag &= ~NFLUSHWANT; 1181 wakeup((caddr_t)&np->n_flag); 1182 } 1183 return (EINTR); 1184 } 1185 error = vinvalbuf(vp, flags, td, 0, slptimeo); 1186 } 1187 np->n_flag &= ~(NMODIFIED | NFLUSHINPROG); 1188 if (np->n_flag & NFLUSHWANT) { 1189 np->n_flag &= ~NFLUSHWANT; 1190 wakeup((caddr_t)&np->n_flag); 1191 } 1192 return (0); 1193 } 1194 1195 /* 1196 * Initiate asynchronous I/O. Return an error if no nfsiods are available. 1197 * This is mainly to avoid queueing async I/O requests when the nfsiods 1198 * are all hung on a dead server. 1199 * 1200 * Note: nfs_asyncio() does not clear (B_ERROR|B_INVAL) but when the bp 1201 * is eventually dequeued by the async daemon, nfs_doio() *will*. 1202 */ 1203 int 1204 nfs_asyncio(struct buf *bp, struct thread *td) 1205 { 1206 struct nfsmount *nmp; 1207 int i; 1208 int gotiod; 1209 int slpflag = 0; 1210 int slptimeo = 0; 1211 int error; 1212 1213 /* 1214 * If no async daemons then return EIO to force caller to run the rpc 1215 * synchronously. 1216 */ 1217 if (nfs_numasync == 0) 1218 return (EIO); 1219 1220 nmp = VFSTONFS(bp->b_vp->v_mount); 1221 1222 /* 1223 * Commits are usually short and sweet so lets save some cpu and 1224 * leave the async daemons for more important rpc's (such as reads 1225 * and writes). 1226 */ 1227 if ((bp->b_flags & (B_READ|B_NEEDCOMMIT)) == B_NEEDCOMMIT && 1228 (nmp->nm_bufqiods > nfs_numasync / 2)) { 1229 return(EIO); 1230 } 1231 1232 again: 1233 if (nmp->nm_flag & NFSMNT_INT) 1234 slpflag = PCATCH; 1235 gotiod = FALSE; 1236 1237 /* 1238 * Find a free iod to process this request. 1239 */ 1240 for (i = 0; i < NFS_MAXASYNCDAEMON; i++) 1241 if (nfs_iodwant[i]) { 1242 /* 1243 * Found one, so wake it up and tell it which 1244 * mount to process. 1245 */ 1246 NFS_DPF(ASYNCIO, 1247 ("nfs_asyncio: waking iod %d for mount %p\n", 1248 i, nmp)); 1249 nfs_iodwant[i] = NULL; 1250 nfs_iodmount[i] = nmp; 1251 nmp->nm_bufqiods++; 1252 wakeup((caddr_t)&nfs_iodwant[i]); 1253 gotiod = TRUE; 1254 break; 1255 } 1256 1257 /* 1258 * If none are free, we may already have an iod working on this mount 1259 * point. If so, it will process our request. 1260 */ 1261 if (!gotiod) { 1262 if (nmp->nm_bufqiods > 0) { 1263 NFS_DPF(ASYNCIO, 1264 ("nfs_asyncio: %d iods are already processing mount %p\n", 1265 nmp->nm_bufqiods, nmp)); 1266 gotiod = TRUE; 1267 } 1268 } 1269 1270 /* 1271 * If we have an iod which can process the request, then queue 1272 * the buffer. 1273 */ 1274 if (gotiod) { 1275 /* 1276 * Ensure that the queue never grows too large. We still want 1277 * to asynchronize so we block rather then return EIO. 1278 */ 1279 while (nmp->nm_bufqlen >= 2*nfs_numasync) { 1280 NFS_DPF(ASYNCIO, 1281 ("nfs_asyncio: waiting for mount %p queue to drain\n", nmp)); 1282 nmp->nm_bufqwant = TRUE; 1283 error = tsleep(&nmp->nm_bufq, slpflag, 1284 "nfsaio", slptimeo); 1285 if (error) { 1286 if (nfs_sigintr(nmp, NULL, td)) 1287 return (EINTR); 1288 if (slpflag == PCATCH) { 1289 slpflag = 0; 1290 slptimeo = 2 * hz; 1291 } 1292 } 1293 /* 1294 * We might have lost our iod while sleeping, 1295 * so check and loop if nescessary. 1296 */ 1297 if (nmp->nm_bufqiods == 0) { 1298 NFS_DPF(ASYNCIO, 1299 ("nfs_asyncio: no iods after mount %p queue was drained, looping\n", nmp)); 1300 goto again; 1301 } 1302 } 1303 1304 if ((bp->b_flags & B_READ) == 0) 1305 bp->b_flags |= B_WRITEINPROG; 1306 1307 BUF_KERNPROC(bp); 1308 TAILQ_INSERT_TAIL(&nmp->nm_bufq, bp, b_freelist); 1309 nmp->nm_bufqlen++; 1310 return (0); 1311 } 1312 1313 /* 1314 * All the iods are busy on other mounts, so return EIO to 1315 * force the caller to process the i/o synchronously. 1316 */ 1317 NFS_DPF(ASYNCIO, ("nfs_asyncio: no iods available, i/o is synchronous\n")); 1318 return (EIO); 1319 } 1320 1321 /* 1322 * Do an I/O operation to/from a cache block. This may be called 1323 * synchronously or from an nfsiod. 1324 * 1325 * NOTE! TD MIGHT BE NULL 1326 */ 1327 int 1328 nfs_doio(struct buf *bp, struct thread *td) 1329 { 1330 struct uio *uiop; 1331 struct vnode *vp; 1332 struct nfsnode *np; 1333 struct nfsmount *nmp; 1334 int error = 0, iomode, must_commit = 0; 1335 struct uio uio; 1336 struct iovec io; 1337 1338 vp = bp->b_vp; 1339 np = VTONFS(vp); 1340 nmp = VFSTONFS(vp->v_mount); 1341 uiop = &uio; 1342 uiop->uio_iov = &io; 1343 uiop->uio_iovcnt = 1; 1344 uiop->uio_segflg = UIO_SYSSPACE; 1345 uiop->uio_td = td; 1346 1347 /* 1348 * clear B_ERROR and B_INVAL state prior to initiating the I/O. We 1349 * do this here so we do not have to do it in all the code that 1350 * calls us. 1351 */ 1352 bp->b_flags &= ~(B_ERROR | B_INVAL); 1353 1354 KASSERT(!(bp->b_flags & B_DONE), ("nfs_doio: bp %p already marked done", bp)); 1355 1356 /* 1357 * Historically, paging was done with physio, but no more. 1358 */ 1359 if (bp->b_flags & B_PHYS) { 1360 /* 1361 * ...though reading /dev/drum still gets us here. 1362 */ 1363 io.iov_len = uiop->uio_resid = bp->b_bcount; 1364 /* mapping was done by vmapbuf() */ 1365 io.iov_base = bp->b_data; 1366 uiop->uio_offset = ((off_t)bp->b_blkno) * DEV_BSIZE; 1367 if (bp->b_flags & B_READ) { 1368 uiop->uio_rw = UIO_READ; 1369 nfsstats.read_physios++; 1370 error = nfs_readrpc(vp, uiop); 1371 } else { 1372 int com; 1373 1374 iomode = NFSV3WRITE_DATASYNC; 1375 uiop->uio_rw = UIO_WRITE; 1376 nfsstats.write_physios++; 1377 error = nfs_writerpc(vp, uiop, &iomode, &com); 1378 } 1379 if (error) { 1380 bp->b_flags |= B_ERROR; 1381 bp->b_error = error; 1382 } 1383 } else if (bp->b_flags & B_READ) { 1384 io.iov_len = uiop->uio_resid = bp->b_bcount; 1385 io.iov_base = bp->b_data; 1386 uiop->uio_rw = UIO_READ; 1387 1388 switch (vp->v_type) { 1389 case VREG: 1390 uiop->uio_offset = ((off_t)bp->b_blkno) * DEV_BSIZE; 1391 nfsstats.read_bios++; 1392 error = nfs_readrpc(vp, uiop); 1393 1394 if (!error) { 1395 if (uiop->uio_resid) { 1396 /* 1397 * If we had a short read with no error, we must have 1398 * hit a file hole. We should zero-fill the remainder. 1399 * This can also occur if the server hits the file EOF. 1400 * 1401 * Holes used to be able to occur due to pending 1402 * writes, but that is not possible any longer. 1403 */ 1404 int nread = bp->b_bcount - uiop->uio_resid; 1405 int left = uiop->uio_resid; 1406 1407 if (left > 0) 1408 bzero((char *)bp->b_data + nread, left); 1409 uiop->uio_resid = 0; 1410 } 1411 } 1412 if (td && td->td_proc && (vp->v_flag & VTEXT) && 1413 (((nmp->nm_flag & NFSMNT_NQNFS) && 1414 NQNFS_CKINVALID(vp, np, ND_READ) && 1415 np->n_lrev != np->n_brev) || 1416 (!(nmp->nm_flag & NFSMNT_NQNFS) && 1417 np->n_mtime != np->n_vattr.va_mtime.tv_sec))) { 1418 uprintf("Process killed due to text file modification\n"); 1419 psignal(td->td_proc, SIGKILL); 1420 PHOLD(td->td_proc); 1421 } 1422 break; 1423 case VLNK: 1424 uiop->uio_offset = (off_t)0; 1425 nfsstats.readlink_bios++; 1426 error = nfs_readlinkrpc(vp, uiop); 1427 break; 1428 case VDIR: 1429 nfsstats.readdir_bios++; 1430 uiop->uio_offset = ((u_quad_t)bp->b_lblkno) * NFS_DIRBLKSIZ; 1431 if (nmp->nm_flag & NFSMNT_RDIRPLUS) { 1432 error = nfs_readdirplusrpc(vp, uiop); 1433 if (error == NFSERR_NOTSUPP) 1434 nmp->nm_flag &= ~NFSMNT_RDIRPLUS; 1435 } 1436 if ((nmp->nm_flag & NFSMNT_RDIRPLUS) == 0) 1437 error = nfs_readdirrpc(vp, uiop); 1438 /* 1439 * end-of-directory sets B_INVAL but does not generate an 1440 * error. 1441 */ 1442 if (error == 0 && uiop->uio_resid == bp->b_bcount) 1443 bp->b_flags |= B_INVAL; 1444 break; 1445 default: 1446 printf("nfs_doio: type %x unexpected\n",vp->v_type); 1447 break; 1448 }; 1449 if (error) { 1450 bp->b_flags |= B_ERROR; 1451 bp->b_error = error; 1452 } 1453 } else { 1454 /* 1455 * If we only need to commit, try to commit 1456 */ 1457 if (bp->b_flags & B_NEEDCOMMIT) { 1458 int retv; 1459 off_t off; 1460 1461 off = ((u_quad_t)bp->b_blkno) * DEV_BSIZE + bp->b_dirtyoff; 1462 bp->b_flags |= B_WRITEINPROG; 1463 retv = nfs_commit(bp->b_vp, off, 1464 bp->b_dirtyend - bp->b_dirtyoff, td); 1465 bp->b_flags &= ~B_WRITEINPROG; 1466 if (retv == 0) { 1467 bp->b_dirtyoff = bp->b_dirtyend = 0; 1468 bp->b_flags &= ~(B_NEEDCOMMIT | B_CLUSTEROK); 1469 bp->b_resid = 0; 1470 biodone(bp); 1471 return (0); 1472 } 1473 if (retv == NFSERR_STALEWRITEVERF) { 1474 nfs_clearcommit(bp->b_vp->v_mount); 1475 } 1476 } 1477 1478 /* 1479 * Setup for actual write 1480 */ 1481 1482 if ((off_t)bp->b_blkno * DEV_BSIZE + bp->b_dirtyend > np->n_size) 1483 bp->b_dirtyend = np->n_size - (off_t)bp->b_blkno * DEV_BSIZE; 1484 1485 if (bp->b_dirtyend > bp->b_dirtyoff) { 1486 io.iov_len = uiop->uio_resid = bp->b_dirtyend 1487 - bp->b_dirtyoff; 1488 uiop->uio_offset = (off_t)bp->b_blkno * DEV_BSIZE 1489 + bp->b_dirtyoff; 1490 io.iov_base = (char *)bp->b_data + bp->b_dirtyoff; 1491 uiop->uio_rw = UIO_WRITE; 1492 nfsstats.write_bios++; 1493 1494 if ((bp->b_flags & (B_ASYNC | B_NEEDCOMMIT | B_NOCACHE | B_CLUSTER)) == B_ASYNC) 1495 iomode = NFSV3WRITE_UNSTABLE; 1496 else 1497 iomode = NFSV3WRITE_FILESYNC; 1498 1499 bp->b_flags |= B_WRITEINPROG; 1500 error = nfs_writerpc(vp, uiop, &iomode, &must_commit); 1501 1502 /* 1503 * When setting B_NEEDCOMMIT also set B_CLUSTEROK to try 1504 * to cluster the buffers needing commit. This will allow 1505 * the system to submit a single commit rpc for the whole 1506 * cluster. We can do this even if the buffer is not 100% 1507 * dirty (relative to the NFS blocksize), so we optimize the 1508 * append-to-file-case. 1509 * 1510 * (when clearing B_NEEDCOMMIT, B_CLUSTEROK must also be 1511 * cleared because write clustering only works for commit 1512 * rpc's, not for the data portion of the write). 1513 */ 1514 1515 if (!error && iomode == NFSV3WRITE_UNSTABLE) { 1516 bp->b_flags |= B_NEEDCOMMIT; 1517 if (bp->b_dirtyoff == 0 1518 && bp->b_dirtyend == bp->b_bcount) 1519 bp->b_flags |= B_CLUSTEROK; 1520 } else { 1521 bp->b_flags &= ~(B_NEEDCOMMIT | B_CLUSTEROK); 1522 } 1523 bp->b_flags &= ~B_WRITEINPROG; 1524 1525 /* 1526 * For an interrupted write, the buffer is still valid 1527 * and the write hasn't been pushed to the server yet, 1528 * so we can't set B_ERROR and report the interruption 1529 * by setting B_EINTR. For the B_ASYNC case, B_EINTR 1530 * is not relevant, so the rpc attempt is essentially 1531 * a noop. For the case of a V3 write rpc not being 1532 * committed to stable storage, the block is still 1533 * dirty and requires either a commit rpc or another 1534 * write rpc with iomode == NFSV3WRITE_FILESYNC before 1535 * the block is reused. This is indicated by setting 1536 * the B_DELWRI and B_NEEDCOMMIT flags. 1537 * 1538 * If the buffer is marked B_PAGING, it does not reside on 1539 * the vp's paging queues so we cannot call bdirty(). The 1540 * bp in this case is not an NFS cache block so we should 1541 * be safe. XXX 1542 */ 1543 if (error == EINTR 1544 || (!error && (bp->b_flags & B_NEEDCOMMIT))) { 1545 int s; 1546 1547 s = splbio(); 1548 bp->b_flags &= ~(B_INVAL|B_NOCACHE); 1549 if ((bp->b_flags & B_PAGING) == 0) { 1550 bdirty(bp); 1551 bp->b_flags &= ~B_DONE; 1552 } 1553 if (error && (bp->b_flags & B_ASYNC) == 0) 1554 bp->b_flags |= B_EINTR; 1555 splx(s); 1556 } else { 1557 if (error) { 1558 bp->b_flags |= B_ERROR; 1559 bp->b_error = np->n_error = error; 1560 np->n_flag |= NWRITEERR; 1561 } 1562 bp->b_dirtyoff = bp->b_dirtyend = 0; 1563 } 1564 } else { 1565 bp->b_resid = 0; 1566 biodone(bp); 1567 return (0); 1568 } 1569 } 1570 bp->b_resid = uiop->uio_resid; 1571 if (must_commit) 1572 nfs_clearcommit(vp->v_mount); 1573 biodone(bp); 1574 return (error); 1575 } 1576 1577 /* 1578 * Used to aid in handling ftruncate() operations on the NFS client side. 1579 * Truncation creates a number of special problems for NFS. We have to 1580 * throw away VM pages and buffer cache buffers that are beyond EOF, and 1581 * we have to properly handle VM pages or (potentially dirty) buffers 1582 * that straddle the truncation point. 1583 */ 1584 1585 int 1586 nfs_meta_setsize(struct vnode *vp, struct thread *td, u_quad_t nsize) 1587 { 1588 struct nfsnode *np = VTONFS(vp); 1589 u_quad_t tsize = np->n_size; 1590 int biosize = vp->v_mount->mnt_stat.f_iosize; 1591 int error = 0; 1592 1593 np->n_size = nsize; 1594 1595 if (np->n_size < tsize) { 1596 struct buf *bp; 1597 daddr_t lbn; 1598 int bufsize; 1599 1600 /* 1601 * vtruncbuf() doesn't get the buffer overlapping the 1602 * truncation point. We may have a B_DELWRI and/or B_CACHE 1603 * buffer that now needs to be truncated. 1604 */ 1605 error = vtruncbuf(vp, td, nsize, biosize); 1606 lbn = nsize / biosize; 1607 bufsize = nsize & (biosize - 1); 1608 bp = nfs_getcacheblk(vp, lbn, bufsize, td); 1609 if (bp->b_dirtyoff > bp->b_bcount) 1610 bp->b_dirtyoff = bp->b_bcount; 1611 if (bp->b_dirtyend > bp->b_bcount) 1612 bp->b_dirtyend = bp->b_bcount; 1613 bp->b_flags |= B_RELBUF; /* don't leave garbage around */ 1614 brelse(bp); 1615 } else { 1616 vnode_pager_setsize(vp, nsize); 1617 } 1618 return(error); 1619 } 1620 1621