nfs_bio.c revision 125454
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 */ 38 39#include <sys/cdefs.h> 40__FBSDID("$FreeBSD: head/sys/nfsclient/nfs_bio.c 125454 2004-02-04 21:52:57Z jhb $"); 41 42#include <sys/param.h> 43#include <sys/systm.h> 44#include <sys/bio.h> 45#include <sys/buf.h> 46#include <sys/kernel.h> 47#include <sys/mount.h> 48#include <sys/proc.h> 49#include <sys/resourcevar.h> 50#include <sys/signalvar.h> 51#include <sys/vmmeter.h> 52#include <sys/vnode.h> 53 54#include <vm/vm.h> 55#include <vm/vm_extern.h> 56#include <vm/vm_page.h> 57#include <vm/vm_object.h> 58#include <vm/vm_pager.h> 59#include <vm/vnode_pager.h> 60 61#include <rpc/rpcclnt.h> 62 63#include <nfs/rpcv2.h> 64#include <nfs/nfsproto.h> 65#include <nfsclient/nfs.h> 66#include <nfsclient/nfsmount.h> 67#include <nfsclient/nfsnode.h> 68 69#include <nfs4client/nfs4.h> 70 71/* 72 * Just call nfs_writebp() with the force argument set to 1. 73 * 74 * NOTE: B_DONE may or may not be set in a_bp on call. 75 */ 76static int 77nfs4_bwrite(struct buf *bp) 78{ 79 80 return (nfs4_writebp(bp, 1, curthread)); 81} 82 83static int 84nfs_bwrite(struct buf *bp) 85{ 86 87 return (nfs_writebp(bp, 1, curthread)); 88} 89 90struct buf_ops buf_ops_nfs4 = { 91 "buf_ops_nfs4", 92 nfs4_bwrite 93}; 94 95struct buf_ops buf_ops_nfs = { 96 "buf_ops_nfs", 97 nfs_bwrite 98}; 99 100static struct buf *nfs_getcacheblk(struct vnode *vp, daddr_t bn, int size, 101 struct thread *td); 102 103/* 104 * Vnode op for VM getpages. 105 */ 106int 107nfs_getpages(struct vop_getpages_args *ap) 108{ 109 int i, error, nextoff, size, toff, count, npages; 110 struct uio uio; 111 struct iovec iov; 112 vm_offset_t kva; 113 struct buf *bp; 114 struct vnode *vp; 115 struct thread *td; 116 struct ucred *cred; 117 struct nfsmount *nmp; 118 vm_object_t object; 119 vm_page_t *pages; 120 121 GIANT_REQUIRED; 122 123 vp = ap->a_vp; 124 td = curthread; /* XXX */ 125 cred = curthread->td_ucred; /* XXX */ 126 nmp = VFSTONFS(vp->v_mount); 127 pages = ap->a_m; 128 count = ap->a_count; 129 130 if ((object = vp->v_object) == NULL) { 131 printf("nfs_getpages: called with non-merged cache vnode??\n"); 132 return VM_PAGER_ERROR; 133 } 134 135 if ((nmp->nm_flag & NFSMNT_NFSV3) != 0 && 136 (nmp->nm_state & NFSSTA_GOTFSINFO) == 0) { 137 /* We'll never get here for v4, because we always have fsinfo */ 138 (void)nfs_fsinfo(nmp, vp, cred, td); 139 } 140 141 npages = btoc(count); 142 143 /* 144 * If the requested page is partially valid, just return it and 145 * allow the pager to zero-out the blanks. Partially valid pages 146 * can only occur at the file EOF. 147 */ 148 149 { 150 vm_page_t m = pages[ap->a_reqpage]; 151 152 VM_OBJECT_LOCK(object); 153 vm_page_lock_queues(); 154 if (m->valid != 0) { 155 /* handled by vm_fault now */ 156 /* vm_page_zero_invalid(m, TRUE); */ 157 for (i = 0; i < npages; ++i) { 158 if (i != ap->a_reqpage) 159 vm_page_free(pages[i]); 160 } 161 vm_page_unlock_queues(); 162 VM_OBJECT_UNLOCK(object); 163 return(0); 164 } 165 vm_page_unlock_queues(); 166 VM_OBJECT_UNLOCK(object); 167 } 168 169 /* 170 * We use only the kva address for the buffer, but this is extremely 171 * convienient and fast. 172 */ 173 bp = getpbuf(&nfs_pbuf_freecnt); 174 175 kva = (vm_offset_t) bp->b_data; 176 pmap_qenter(kva, pages, npages); 177 cnt.v_vnodein++; 178 cnt.v_vnodepgsin += npages; 179 180 iov.iov_base = (caddr_t) kva; 181 iov.iov_len = count; 182 uio.uio_iov = &iov; 183 uio.uio_iovcnt = 1; 184 uio.uio_offset = IDX_TO_OFF(pages[0]->pindex); 185 uio.uio_resid = count; 186 uio.uio_segflg = UIO_SYSSPACE; 187 uio.uio_rw = UIO_READ; 188 uio.uio_td = td; 189 190 error = (nmp->nm_rpcops->nr_readrpc)(vp, &uio, cred); 191 pmap_qremove(kva, npages); 192 193 relpbuf(bp, &nfs_pbuf_freecnt); 194 195 if (error && (uio.uio_resid == count)) { 196 printf("nfs_getpages: error %d\n", error); 197 VM_OBJECT_LOCK(object); 198 vm_page_lock_queues(); 199 for (i = 0; i < npages; ++i) { 200 if (i != ap->a_reqpage) 201 vm_page_free(pages[i]); 202 } 203 vm_page_unlock_queues(); 204 VM_OBJECT_UNLOCK(object); 205 return VM_PAGER_ERROR; 206 } 207 208 /* 209 * Calculate the number of bytes read and validate only that number 210 * of bytes. Note that due to pending writes, size may be 0. This 211 * does not mean that the remaining data is invalid! 212 */ 213 214 size = count - uio.uio_resid; 215 VM_OBJECT_LOCK(object); 216 vm_page_lock_queues(); 217 for (i = 0, toff = 0; i < npages; i++, toff = nextoff) { 218 vm_page_t m; 219 nextoff = toff + PAGE_SIZE; 220 m = pages[i]; 221 222 m->flags &= ~PG_ZERO; 223 224 if (nextoff <= size) { 225 /* 226 * Read operation filled an entire page 227 */ 228 m->valid = VM_PAGE_BITS_ALL; 229 vm_page_undirty(m); 230 } else if (size > toff) { 231 /* 232 * Read operation filled a partial page. 233 */ 234 m->valid = 0; 235 vm_page_set_validclean(m, 0, size - toff); 236 /* handled by vm_fault now */ 237 /* vm_page_zero_invalid(m, TRUE); */ 238 } else { 239 /* 240 * Read operation was short. If no error occured 241 * we may have hit a zero-fill section. We simply 242 * leave valid set to 0. 243 */ 244 ; 245 } 246 if (i != ap->a_reqpage) { 247 /* 248 * Whether or not to leave the page activated is up in 249 * the air, but we should put the page on a page queue 250 * somewhere (it already is in the object). Result: 251 * It appears that emperical results show that 252 * deactivating pages is best. 253 */ 254 255 /* 256 * Just in case someone was asking for this page we 257 * now tell them that it is ok to use. 258 */ 259 if (!error) { 260 if (m->flags & PG_WANTED) 261 vm_page_activate(m); 262 else 263 vm_page_deactivate(m); 264 vm_page_wakeup(m); 265 } else { 266 vm_page_free(m); 267 } 268 } 269 } 270 vm_page_unlock_queues(); 271 VM_OBJECT_UNLOCK(object); 272 return 0; 273} 274 275/* 276 * Vnode op for VM putpages. 277 */ 278int 279nfs_putpages(struct vop_putpages_args *ap) 280{ 281 struct uio uio; 282 struct iovec iov; 283 vm_offset_t kva; 284 struct buf *bp; 285 int iomode, must_commit, i, error, npages, count; 286 off_t offset; 287 int *rtvals; 288 struct vnode *vp; 289 struct thread *td; 290 struct ucred *cred; 291 struct nfsmount *nmp; 292 struct nfsnode *np; 293 vm_page_t *pages; 294 295 GIANT_REQUIRED; 296 297 vp = ap->a_vp; 298 np = VTONFS(vp); 299 td = curthread; /* XXX */ 300 cred = curthread->td_ucred; /* XXX */ 301 nmp = VFSTONFS(vp->v_mount); 302 pages = ap->a_m; 303 count = ap->a_count; 304 rtvals = ap->a_rtvals; 305 npages = btoc(count); 306 offset = IDX_TO_OFF(pages[0]->pindex); 307 308 if ((nmp->nm_flag & NFSMNT_NFSV3) != 0 && 309 (nmp->nm_state & NFSSTA_GOTFSINFO) == 0) { 310 (void)nfs_fsinfo(nmp, vp, cred, td); 311 } 312 313 for (i = 0; i < npages; i++) 314 rtvals[i] = VM_PAGER_AGAIN; 315 316 /* 317 * When putting pages, do not extend file past EOF. 318 */ 319 320 if (offset + count > np->n_size) { 321 count = np->n_size - offset; 322 if (count < 0) 323 count = 0; 324 } 325 326 /* 327 * We use only the kva address for the buffer, but this is extremely 328 * convienient and fast. 329 */ 330 bp = getpbuf(&nfs_pbuf_freecnt); 331 332 kva = (vm_offset_t) bp->b_data; 333 pmap_qenter(kva, pages, npages); 334 cnt.v_vnodeout++; 335 cnt.v_vnodepgsout += count; 336 337 iov.iov_base = (caddr_t) kva; 338 iov.iov_len = count; 339 uio.uio_iov = &iov; 340 uio.uio_iovcnt = 1; 341 uio.uio_offset = offset; 342 uio.uio_resid = count; 343 uio.uio_segflg = UIO_SYSSPACE; 344 uio.uio_rw = UIO_WRITE; 345 uio.uio_td = td; 346 347 if ((ap->a_sync & VM_PAGER_PUT_SYNC) == 0) 348 iomode = NFSV3WRITE_UNSTABLE; 349 else 350 iomode = NFSV3WRITE_FILESYNC; 351 352 error = (nmp->nm_rpcops->nr_writerpc)(vp, &uio, cred, &iomode, &must_commit); 353 354 pmap_qremove(kva, npages); 355 relpbuf(bp, &nfs_pbuf_freecnt); 356 357 if (!error) { 358 int nwritten = round_page(count - uio.uio_resid) / PAGE_SIZE; 359 for (i = 0; i < nwritten; i++) { 360 rtvals[i] = VM_PAGER_OK; 361 vm_page_undirty(pages[i]); 362 } 363 if (must_commit) { 364 nfs_clearcommit(vp->v_mount); 365 } 366 } 367 return rtvals[0]; 368} 369 370/* 371 * Vnode op for read using bio 372 */ 373int 374nfs_bioread(struct vnode *vp, struct uio *uio, int ioflag, struct ucred *cred) 375{ 376 struct nfsnode *np = VTONFS(vp); 377 int biosize, i; 378 struct buf *bp = 0, *rabp; 379 struct vattr vattr; 380 struct thread *td; 381 struct nfsmount *nmp = VFSTONFS(vp->v_mount); 382 daddr_t lbn, rabn; 383 int bcount; 384 int seqcount; 385 int nra, error = 0, n = 0, on = 0; 386 387#ifdef DIAGNOSTIC 388 if (uio->uio_rw != UIO_READ) 389 panic("nfs_read mode"); 390#endif 391 if (uio->uio_resid == 0) 392 return (0); 393 if (uio->uio_offset < 0) /* XXX VDIR cookies can be negative */ 394 return (EINVAL); 395 td = uio->uio_td; 396 397 if ((nmp->nm_flag & NFSMNT_NFSV3) != 0 && 398 (nmp->nm_state & NFSSTA_GOTFSINFO) == 0) 399 (void)nfs_fsinfo(nmp, vp, cred, td); 400 if (vp->v_type != VDIR && 401 (uio->uio_offset + uio->uio_resid) > nmp->nm_maxfilesize) 402 return (EFBIG); 403 biosize = vp->v_mount->mnt_stat.f_iosize; 404 seqcount = (int)((off_t)(ioflag >> IO_SEQSHIFT) * biosize / BKVASIZE); 405 /* 406 * For nfs, cache consistency can only be maintained approximately. 407 * Although RFC1094 does not specify the criteria, the following is 408 * believed to be compatible with the reference port. 409 * For nfs: 410 * If the file's modify time on the server has changed since the 411 * last read rpc or you have written to the file, 412 * you may have lost data cache consistency with the 413 * server, so flush all of the file's data out of the cache. 414 * Then force a getattr rpc to ensure that you have up to date 415 * attributes. 416 * NB: This implies that cache data can be read when up to 417 * NFS_ATTRTIMEO seconds out of date. If you find that you need current 418 * attributes this could be forced by setting n_attrstamp to 0 before 419 * the VOP_GETATTR() call. 420 */ 421 if (np->n_flag & NMODIFIED) { 422 if (vp->v_type != VREG) { 423 if (vp->v_type != VDIR) 424 panic("nfs: bioread, not dir"); 425 (nmp->nm_rpcops->nr_invaldir)(vp); 426 error = nfs_vinvalbuf(vp, V_SAVE, cred, td, 1); 427 if (error) 428 return (error); 429 } 430 np->n_attrstamp = 0; 431 error = VOP_GETATTR(vp, &vattr, cred, td); 432 if (error) 433 return (error); 434 np->n_mtime = vattr.va_mtime.tv_sec; 435 } else { 436 error = VOP_GETATTR(vp, &vattr, cred, td); 437 if (error) 438 return (error); 439 if (np->n_mtime != vattr.va_mtime.tv_sec) { 440 if (vp->v_type == VDIR) 441 (nmp->nm_rpcops->nr_invaldir)(vp); 442 error = nfs_vinvalbuf(vp, V_SAVE, cred, td, 1); 443 if (error) 444 return (error); 445 np->n_mtime = vattr.va_mtime.tv_sec; 446 } 447 } 448 do { 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 (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) == NULL) { 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_ASYNC; 468 rabp->b_iocmd = BIO_READ; 469 vfs_busy_pages(rabp, 0); 470 if (nfs_asyncio(rabp, cred, td)) { 471 rabp->b_flags |= B_INVAL; 472 rabp->b_ioflags |= BIO_ERROR; 473 vfs_unbusy_pages(rabp); 474 brelse(rabp); 475 break; 476 } 477 } else { 478 brelse(rabp); 479 } 480 } 481 } 482 } 483 484 /* 485 * Obtain the buffer cache block. Figure out the buffer size 486 * when we are at EOF. If we are modifying the size of the 487 * buffer based on an EOF condition we need to hold 488 * nfs_rslock() through obtaining the buffer to prevent 489 * a potential writer-appender from messing with n_size. 490 * Otherwise we may accidently truncate the buffer and 491 * lose dirty data. 492 * 493 * Note that bcount is *not* DEV_BSIZE aligned. 494 */ 495 496again: 497 bcount = biosize; 498 if ((off_t)lbn * biosize >= np->n_size) { 499 bcount = 0; 500 } else if ((off_t)(lbn + 1) * biosize > np->n_size) { 501 bcount = np->n_size - (off_t)lbn * biosize; 502 } 503 if (bcount != biosize) { 504 switch(nfs_rslock(np, td)) { 505 case ENOLCK: 506 goto again; 507 /* not reached */ 508 case EINTR: 509 case ERESTART: 510 return(EINTR); 511 /* not reached */ 512 default: 513 break; 514 } 515 } 516 517 bp = nfs_getcacheblk(vp, lbn, bcount, td); 518 519 if (bcount != biosize) 520 nfs_rsunlock(np, td); 521 if (!bp) 522 return (EINTR); 523 524 /* 525 * If B_CACHE is not set, we must issue the read. If this 526 * fails, we return an error. 527 */ 528 529 if ((bp->b_flags & B_CACHE) == 0) { 530 bp->b_iocmd = BIO_READ; 531 vfs_busy_pages(bp, 0); 532 error = nfs_doio(bp, cred, td); 533 if (error) { 534 brelse(bp); 535 return (error); 536 } 537 } 538 539 /* 540 * on is the offset into the current bp. Figure out how many 541 * bytes we can copy out of the bp. Note that bcount is 542 * NOT DEV_BSIZE aligned. 543 * 544 * Then figure out how many bytes we can copy into the uio. 545 */ 546 547 n = 0; 548 if (on < bcount) 549 n = min((unsigned)(bcount - on), uio->uio_resid); 550 break; 551 case VLNK: 552 nfsstats.biocache_readlinks++; 553 bp = nfs_getcacheblk(vp, (daddr_t)0, NFS_MAXPATHLEN, td); 554 if (!bp) 555 return (EINTR); 556 if ((bp->b_flags & B_CACHE) == 0) { 557 bp->b_iocmd = BIO_READ; 558 vfs_busy_pages(bp, 0); 559 error = nfs_doio(bp, cred, td); 560 if (error) { 561 bp->b_ioflags |= BIO_ERROR; 562 brelse(bp); 563 return (error); 564 } 565 } 566 n = min(uio->uio_resid, NFS_MAXPATHLEN - bp->b_resid); 567 on = 0; 568 break; 569 case VDIR: 570 nfsstats.biocache_readdirs++; 571 if (np->n_direofoffset 572 && uio->uio_offset >= np->n_direofoffset) { 573 return (0); 574 } 575 lbn = (uoff_t)uio->uio_offset / NFS_DIRBLKSIZ; 576 on = uio->uio_offset & (NFS_DIRBLKSIZ - 1); 577 bp = nfs_getcacheblk(vp, lbn, NFS_DIRBLKSIZ, td); 578 if (!bp) 579 return (EINTR); 580 if ((bp->b_flags & B_CACHE) == 0) { 581 bp->b_iocmd = BIO_READ; 582 vfs_busy_pages(bp, 0); 583 error = nfs_doio(bp, cred, td); 584 if (error) { 585 brelse(bp); 586 } 587 while (error == NFSERR_BAD_COOKIE) { 588 printf("got bad cookie vp %p bp %p\n", vp, bp); 589 (nmp->nm_rpcops->nr_invaldir)(vp); 590 error = nfs_vinvalbuf(vp, 0, cred, td, 1); 591 /* 592 * Yuck! The directory has been modified on the 593 * server. The only way to get the block is by 594 * reading from the beginning to get all the 595 * offset cookies. 596 * 597 * Leave the last bp intact unless there is an error. 598 * Loop back up to the while if the error is another 599 * NFSERR_BAD_COOKIE (double yuch!). 600 */ 601 for (i = 0; i <= lbn && !error; i++) { 602 if (np->n_direofoffset 603 && (i * NFS_DIRBLKSIZ) >= np->n_direofoffset) 604 return (0); 605 bp = nfs_getcacheblk(vp, i, NFS_DIRBLKSIZ, td); 606 if (!bp) 607 return (EINTR); 608 if ((bp->b_flags & B_CACHE) == 0) { 609 bp->b_iocmd = BIO_READ; 610 vfs_busy_pages(bp, 0); 611 error = nfs_doio(bp, cred, td); 612 /* 613 * no error + B_INVAL == directory EOF, 614 * use the block. 615 */ 616 if (error == 0 && (bp->b_flags & B_INVAL)) 617 break; 618 } 619 /* 620 * An error will throw away the block and the 621 * for loop will break out. If no error and this 622 * is not the block we want, we throw away the 623 * block and go for the next one via the for loop. 624 */ 625 if (error || i < lbn) 626 brelse(bp); 627 } 628 } 629 /* 630 * The above while is repeated if we hit another cookie 631 * error. If we hit an error and it wasn't a cookie error, 632 * we give up. 633 */ 634 if (error) 635 return (error); 636 } 637 638 /* 639 * If not eof and read aheads are enabled, start one. 640 * (You need the current block first, so that you have the 641 * directory offset cookie of the next block.) 642 */ 643 if (nmp->nm_readahead > 0 && 644 (bp->b_flags & B_INVAL) == 0 && 645 (np->n_direofoffset == 0 || 646 (lbn + 1) * NFS_DIRBLKSIZ < np->n_direofoffset) && 647 incore(vp, lbn + 1) == NULL) { 648 rabp = nfs_getcacheblk(vp, lbn + 1, NFS_DIRBLKSIZ, td); 649 if (rabp) { 650 if ((rabp->b_flags & (B_CACHE|B_DELWRI)) == 0) { 651 rabp->b_flags |= B_ASYNC; 652 rabp->b_iocmd = BIO_READ; 653 vfs_busy_pages(rabp, 0); 654 if (nfs_asyncio(rabp, cred, td)) { 655 rabp->b_flags |= B_INVAL; 656 rabp->b_ioflags |= BIO_ERROR; 657 vfs_unbusy_pages(rabp); 658 brelse(rabp); 659 } 660 } else { 661 brelse(rabp); 662 } 663 } 664 } 665 /* 666 * Unlike VREG files, whos buffer size ( bp->b_bcount ) is 667 * chopped for the EOF condition, we cannot tell how large 668 * NFS directories are going to be until we hit EOF. So 669 * an NFS directory buffer is *not* chopped to its EOF. Now, 670 * it just so happens that b_resid will effectively chop it 671 * to EOF. *BUT* this information is lost if the buffer goes 672 * away and is reconstituted into a B_CACHE state ( due to 673 * being VMIO ) later. So we keep track of the directory eof 674 * in np->n_direofoffset and chop it off as an extra step 675 * right here. 676 */ 677 n = lmin(uio->uio_resid, NFS_DIRBLKSIZ - bp->b_resid - on); 678 if (np->n_direofoffset && n > np->n_direofoffset - uio->uio_offset) 679 n = np->n_direofoffset - uio->uio_offset; 680 break; 681 default: 682 printf(" nfs_bioread: type %x unexpected\n", vp->v_type); 683 break; 684 }; 685 686 if (n > 0) { 687 error = uiomove(bp->b_data + on, (int)n, uio); 688 } 689 switch (vp->v_type) { 690 case VREG: 691 break; 692 case VLNK: 693 n = 0; 694 break; 695 case VDIR: 696 break; 697 default: 698 printf(" nfs_bioread: type %x unexpected\n", vp->v_type); 699 } 700 brelse(bp); 701 } while (error == 0 && uio->uio_resid > 0 && n > 0); 702 return (error); 703} 704 705/* 706 * Vnode op for write using bio 707 */ 708int 709nfs_write(struct vop_write_args *ap) 710{ 711 int biosize; 712 struct uio *uio = ap->a_uio; 713 struct thread *td = uio->uio_td; 714 struct vnode *vp = ap->a_vp; 715 struct nfsnode *np = VTONFS(vp); 716 struct ucred *cred = ap->a_cred; 717 int ioflag = ap->a_ioflag; 718 struct buf *bp; 719 struct vattr vattr; 720 struct nfsmount *nmp = VFSTONFS(vp->v_mount); 721 daddr_t lbn; 722 int bcount; 723 int n, on, error = 0; 724 int haverslock = 0; 725 struct proc *p = td?td->td_proc:NULL; 726 727 GIANT_REQUIRED; 728 729#ifdef DIAGNOSTIC 730 if (uio->uio_rw != UIO_WRITE) 731 panic("nfs_write mode"); 732 if (uio->uio_segflg == UIO_USERSPACE && uio->uio_td != curthread) 733 panic("nfs_write proc"); 734#endif 735 if (vp->v_type != VREG) 736 return (EIO); 737 if (np->n_flag & NWRITEERR) { 738 np->n_flag &= ~NWRITEERR; 739 return (np->n_error); 740 } 741 if ((nmp->nm_flag & NFSMNT_NFSV3) != 0 && 742 (nmp->nm_state & NFSSTA_GOTFSINFO) == 0) 743 (void)nfs_fsinfo(nmp, vp, cred, td); 744 745 /* 746 * Synchronously flush pending buffers if we are in synchronous 747 * mode or if we are appending. 748 */ 749 if (ioflag & (IO_APPEND | IO_SYNC)) { 750 if (np->n_flag & NMODIFIED) { 751 np->n_attrstamp = 0; 752 error = nfs_vinvalbuf(vp, V_SAVE, cred, td, 1); 753 if (error) 754 return (error); 755 } 756 } 757 758 /* 759 * If IO_APPEND then load uio_offset. We restart here if we cannot 760 * get the append lock. 761 */ 762restart: 763 if (ioflag & IO_APPEND) { 764 np->n_attrstamp = 0; 765 error = VOP_GETATTR(vp, &vattr, cred, td); 766 if (error) 767 return (error); 768 uio->uio_offset = np->n_size; 769 } 770 771 if (uio->uio_offset < 0) 772 return (EINVAL); 773 if ((uio->uio_offset + uio->uio_resid) > nmp->nm_maxfilesize) 774 return (EFBIG); 775 if (uio->uio_resid == 0) 776 return (0); 777 778 /* 779 * We need to obtain the rslock if we intend to modify np->n_size 780 * in order to guarentee the append point with multiple contending 781 * writers, to guarentee that no other appenders modify n_size 782 * while we are trying to obtain a truncated buffer (i.e. to avoid 783 * accidently truncating data written by another appender due to 784 * the race), and to ensure that the buffer is populated prior to 785 * our extending of the file. We hold rslock through the entire 786 * operation. 787 * 788 * Note that we do not synchronize the case where someone truncates 789 * the file while we are appending to it because attempting to lock 790 * this case may deadlock other parts of the system unexpectedly. 791 */ 792 if ((ioflag & IO_APPEND) || 793 uio->uio_offset + uio->uio_resid > np->n_size) { 794 switch(nfs_rslock(np, td)) { 795 case ENOLCK: 796 goto restart; 797 /* not reached */ 798 case EINTR: 799 case ERESTART: 800 return(EINTR); 801 /* not reached */ 802 default: 803 break; 804 } 805 haverslock = 1; 806 } 807 808 /* 809 * Maybe this should be above the vnode op call, but so long as 810 * file servers have no limits, i don't think it matters 811 */ 812 if (p != NULL) { 813 PROC_LOCK(p); 814 if (uio->uio_offset + uio->uio_resid > 815 lim_cur(p, RLIMIT_FSIZE)) { 816 psignal(p, SIGXFSZ); 817 PROC_UNLOCK(p); 818 if (haverslock) 819 nfs_rsunlock(np, td); 820 return (EFBIG); 821 } 822 PROC_UNLOCK(p); 823 } 824 825 biosize = vp->v_mount->mnt_stat.f_iosize; 826 827 do { 828 nfsstats.biocache_writes++; 829 lbn = uio->uio_offset / biosize; 830 on = uio->uio_offset & (biosize-1); 831 n = min((unsigned)(biosize - on), uio->uio_resid); 832again: 833 /* 834 * Handle direct append and file extension cases, calculate 835 * unaligned buffer size. 836 */ 837 838 if (uio->uio_offset == np->n_size && n) { 839 /* 840 * Get the buffer (in its pre-append state to maintain 841 * B_CACHE if it was previously set). Resize the 842 * nfsnode after we have locked the buffer to prevent 843 * readers from reading garbage. 844 */ 845 bcount = on; 846 bp = nfs_getcacheblk(vp, lbn, bcount, td); 847 848 if (bp != NULL) { 849 long save; 850 851 np->n_size = uio->uio_offset + n; 852 np->n_flag |= NMODIFIED; 853 vnode_pager_setsize(vp, np->n_size); 854 855 save = bp->b_flags & B_CACHE; 856 bcount += n; 857 allocbuf(bp, bcount); 858 bp->b_flags |= save; 859 bp->b_magic = B_MAGIC_NFS; 860 if ((nmp->nm_flag & NFSMNT_NFSV4) != 0) 861 bp->b_op = &buf_ops_nfs4; 862 else 863 bp->b_op = &buf_ops_nfs; 864 } 865 } else { 866 /* 867 * Obtain the locked cache block first, and then 868 * adjust the file's size as appropriate. 869 */ 870 bcount = on + n; 871 if ((off_t)lbn * biosize + bcount < np->n_size) { 872 if ((off_t)(lbn + 1) * biosize < np->n_size) 873 bcount = biosize; 874 else 875 bcount = np->n_size - (off_t)lbn * biosize; 876 } 877 bp = nfs_getcacheblk(vp, lbn, bcount, td); 878 if (uio->uio_offset + n > np->n_size) { 879 np->n_size = uio->uio_offset + n; 880 np->n_flag |= NMODIFIED; 881 vnode_pager_setsize(vp, np->n_size); 882 } 883 } 884 885 if (!bp) { 886 error = EINTR; 887 break; 888 } 889 890 /* 891 * Issue a READ if B_CACHE is not set. In special-append 892 * mode, B_CACHE is based on the buffer prior to the write 893 * op and is typically set, avoiding the read. If a read 894 * is required in special append mode, the server will 895 * probably send us a short-read since we extended the file 896 * on our end, resulting in b_resid == 0 and, thusly, 897 * B_CACHE getting set. 898 * 899 * We can also avoid issuing the read if the write covers 900 * the entire buffer. We have to make sure the buffer state 901 * is reasonable in this case since we will not be initiating 902 * I/O. See the comments in kern/vfs_bio.c's getblk() for 903 * more information. 904 * 905 * B_CACHE may also be set due to the buffer being cached 906 * normally. 907 */ 908 909 if (on == 0 && n == bcount) { 910 bp->b_flags |= B_CACHE; 911 bp->b_flags &= ~B_INVAL; 912 bp->b_ioflags &= ~BIO_ERROR; 913 } 914 915 if ((bp->b_flags & B_CACHE) == 0) { 916 bp->b_iocmd = BIO_READ; 917 vfs_busy_pages(bp, 0); 918 error = nfs_doio(bp, cred, td); 919 if (error) { 920 brelse(bp); 921 break; 922 } 923 } 924 if (!bp) { 925 error = EINTR; 926 break; 927 } 928 if (bp->b_wcred == NOCRED) 929 bp->b_wcred = crhold(cred); 930 np->n_flag |= NMODIFIED; 931 932 /* 933 * If dirtyend exceeds file size, chop it down. This should 934 * not normally occur but there is an append race where it 935 * might occur XXX, so we log it. 936 * 937 * If the chopping creates a reverse-indexed or degenerate 938 * situation with dirtyoff/end, we 0 both of them. 939 */ 940 941 if (bp->b_dirtyend > bcount) { 942 printf("NFS append race @%lx:%d\n", 943 (long)bp->b_blkno * DEV_BSIZE, 944 bp->b_dirtyend - bcount); 945 bp->b_dirtyend = bcount; 946 } 947 948 if (bp->b_dirtyoff >= bp->b_dirtyend) 949 bp->b_dirtyoff = bp->b_dirtyend = 0; 950 951 /* 952 * If the new write will leave a contiguous dirty 953 * area, just update the b_dirtyoff and b_dirtyend, 954 * otherwise force a write rpc of the old dirty area. 955 * 956 * While it is possible to merge discontiguous writes due to 957 * our having a B_CACHE buffer ( and thus valid read data 958 * for the hole), we don't because it could lead to 959 * significant cache coherency problems with multiple clients, 960 * especially if locking is implemented later on. 961 * 962 * as an optimization we could theoretically maintain 963 * a linked list of discontinuous areas, but we would still 964 * have to commit them separately so there isn't much 965 * advantage to it except perhaps a bit of asynchronization. 966 */ 967 968 if (bp->b_dirtyend > 0 && 969 (on > bp->b_dirtyend || (on + n) < bp->b_dirtyoff)) { 970 if (BUF_WRITE(bp) == EINTR) { 971 error = EINTR; 972 break; 973 } 974 goto again; 975 } 976 977 error = uiomove((char *)bp->b_data + on, n, uio); 978 979 /* 980 * Since this block is being modified, it must be written 981 * again and not just committed. Since write clustering does 982 * not work for the stage 1 data write, only the stage 2 983 * commit rpc, we have to clear B_CLUSTEROK as well. 984 */ 985 bp->b_flags &= ~(B_NEEDCOMMIT | B_CLUSTEROK); 986 987 if (error) { 988 bp->b_ioflags |= BIO_ERROR; 989 brelse(bp); 990 break; 991 } 992 993 /* 994 * Only update dirtyoff/dirtyend if not a degenerate 995 * condition. 996 */ 997 if (n) { 998 if (bp->b_dirtyend > 0) { 999 bp->b_dirtyoff = min(on, bp->b_dirtyoff); 1000 bp->b_dirtyend = max((on + n), bp->b_dirtyend); 1001 } else { 1002 bp->b_dirtyoff = on; 1003 bp->b_dirtyend = on + n; 1004 } 1005 vfs_bio_set_validclean(bp, on, n); 1006 } 1007 1008 /* 1009 * If IO_SYNC do bwrite(). 1010 * 1011 * IO_INVAL appears to be unused. The idea appears to be 1012 * to turn off caching in this case. Very odd. XXX 1013 */ 1014 if ((ioflag & IO_SYNC)) { 1015 if (ioflag & IO_INVAL) 1016 bp->b_flags |= B_NOCACHE; 1017 error = BUF_WRITE(bp); 1018 if (error) 1019 break; 1020 } else if ((n + on) == biosize) { 1021 bp->b_flags |= B_ASYNC; 1022 (void) (nmp->nm_rpcops->nr_writebp)(bp, 0, 0); 1023 } else { 1024 bdwrite(bp); 1025 } 1026 } while (uio->uio_resid > 0 && n > 0); 1027 1028 if (haverslock) 1029 nfs_rsunlock(np, td); 1030 1031 return (error); 1032} 1033 1034/* 1035 * Get an nfs cache block. 1036 * 1037 * Allocate a new one if the block isn't currently in the cache 1038 * and return the block marked busy. If the calling process is 1039 * interrupted by a signal for an interruptible mount point, return 1040 * NULL. 1041 * 1042 * The caller must carefully deal with the possible B_INVAL state of 1043 * the buffer. nfs_doio() clears B_INVAL (and nfs_asyncio() clears it 1044 * indirectly), so synchronous reads can be issued without worrying about 1045 * the B_INVAL state. We have to be a little more careful when dealing 1046 * with writes (see comments in nfs_write()) when extending a file past 1047 * its EOF. 1048 */ 1049static struct buf * 1050nfs_getcacheblk(struct vnode *vp, daddr_t bn, int size, struct thread *td) 1051{ 1052 struct buf *bp; 1053 struct mount *mp; 1054 struct nfsmount *nmp; 1055 1056 mp = vp->v_mount; 1057 nmp = VFSTONFS(mp); 1058 1059 if (nmp->nm_flag & NFSMNT_INT) { 1060 bp = getblk(vp, bn, size, PCATCH, 0, 0); 1061 while (bp == NULL) { 1062 if (nfs_sigintr(nmp, NULL, td)) 1063 return (NULL); 1064 bp = getblk(vp, bn, size, 0, 2 * hz, 0); 1065 } 1066 } else { 1067 bp = getblk(vp, bn, size, 0, 0, 0); 1068 } 1069 1070 if (vp->v_type == VREG) { 1071 int biosize; 1072 1073 biosize = mp->mnt_stat.f_iosize; 1074 bp->b_blkno = bn * (biosize / DEV_BSIZE); 1075 } 1076 return (bp); 1077} 1078 1079/* 1080 * Flush and invalidate all dirty buffers. If another process is already 1081 * doing the flush, just wait for completion. 1082 */ 1083int 1084nfs_vinvalbuf(struct vnode *vp, int flags, struct ucred *cred, 1085 struct thread *td, int intrflg) 1086{ 1087 struct nfsnode *np = VTONFS(vp); 1088 struct nfsmount *nmp = VFSTONFS(vp->v_mount); 1089 int error = 0, slpflag, slptimeo; 1090 1091 ASSERT_VOP_LOCKED(vp, "nfs_vinvalbuf"); 1092 1093 /* 1094 * XXX This check stops us from needlessly doing a vinvalbuf when 1095 * being called through vclean(). It is not clear that this is 1096 * unsafe. 1097 */ 1098 if (vp->v_iflag & VI_XLOCK) 1099 return (0); 1100 1101 if ((nmp->nm_flag & NFSMNT_INT) == 0) 1102 intrflg = 0; 1103 if (intrflg) { 1104 slpflag = PCATCH; 1105 slptimeo = 2 * hz; 1106 } else { 1107 slpflag = 0; 1108 slptimeo = 0; 1109 } 1110 /* 1111 * First wait for any other process doing a flush to complete. 1112 */ 1113 while (np->n_flag & NFLUSHINPROG) { 1114 np->n_flag |= NFLUSHWANT; 1115 error = tsleep(&np->n_flag, PRIBIO + 2, "nfsvinval", 1116 slptimeo); 1117 if (error && intrflg && 1118 nfs_sigintr(nmp, NULL, td)) 1119 return (EINTR); 1120 } 1121 1122 /* 1123 * Now, flush as required. 1124 */ 1125 np->n_flag |= NFLUSHINPROG; 1126 error = vinvalbuf(vp, flags, cred, td, slpflag, 0); 1127 while (error) { 1128 if (intrflg && 1129 nfs_sigintr(nmp, NULL, td)) { 1130 np->n_flag &= ~NFLUSHINPROG; 1131 if (np->n_flag & NFLUSHWANT) { 1132 np->n_flag &= ~NFLUSHWANT; 1133 wakeup(&np->n_flag); 1134 } 1135 return (EINTR); 1136 } 1137 error = vinvalbuf(vp, flags, cred, td, 0, slptimeo); 1138 } 1139 np->n_flag &= ~(NMODIFIED | NFLUSHINPROG); 1140 if (np->n_flag & NFLUSHWANT) { 1141 np->n_flag &= ~NFLUSHWANT; 1142 wakeup(&np->n_flag); 1143 } 1144 return (0); 1145} 1146 1147/* 1148 * Initiate asynchronous I/O. Return an error if no nfsiods are available. 1149 * This is mainly to avoid queueing async I/O requests when the nfsiods 1150 * are all hung on a dead server. 1151 * 1152 * Note: nfs_asyncio() does not clear (BIO_ERROR|B_INVAL) but when the bp 1153 * is eventually dequeued by the async daemon, nfs_doio() *will*. 1154 */ 1155int 1156nfs_asyncio(struct buf *bp, struct ucred *cred, struct thread *td) 1157{ 1158 struct nfsmount *nmp; 1159 int iod; 1160 int gotiod; 1161 int slpflag = 0; 1162 int slptimeo = 0; 1163 int error; 1164 1165 nmp = VFSTONFS(bp->b_vp->v_mount); 1166 1167 /* 1168 * Commits are usually short and sweet so lets save some cpu and 1169 * leave the async daemons for more important rpc's (such as reads 1170 * and writes). 1171 */ 1172 if (bp->b_iocmd == BIO_WRITE && (bp->b_flags & B_NEEDCOMMIT) && 1173 (nmp->nm_bufqiods > nfs_numasync / 2)) { 1174 return(EIO); 1175 } 1176 1177again: 1178 if (nmp->nm_flag & NFSMNT_INT) 1179 slpflag = PCATCH; 1180 gotiod = FALSE; 1181 1182 /* 1183 * Find a free iod to process this request. 1184 */ 1185 for (iod = 0; iod < nfs_numasync; iod++) 1186 if (nfs_iodwant[iod]) { 1187 gotiod = TRUE; 1188 break; 1189 } 1190 1191 /* 1192 * Try to create one if none are free. 1193 */ 1194 if (!gotiod) { 1195 iod = nfs_nfsiodnew(); 1196 if (iod != -1) 1197 gotiod = TRUE; 1198 } 1199 1200 if (gotiod) { 1201 /* 1202 * Found one, so wake it up and tell it which 1203 * mount to process. 1204 */ 1205 NFS_DPF(ASYNCIO, ("nfs_asyncio: waking iod %d for mount %p\n", 1206 iod, nmp)); 1207 nfs_iodwant[iod] = NULL; 1208 nfs_iodmount[iod] = nmp; 1209 nmp->nm_bufqiods++; 1210 wakeup(&nfs_iodwant[iod]); 1211 } 1212 1213 /* 1214 * If none are free, we may already have an iod working on this mount 1215 * point. If so, it will process our request. 1216 */ 1217 if (!gotiod) { 1218 if (nmp->nm_bufqiods > 0) { 1219 NFS_DPF(ASYNCIO, 1220 ("nfs_asyncio: %d iods are already processing mount %p\n", 1221 nmp->nm_bufqiods, nmp)); 1222 gotiod = TRUE; 1223 } 1224 } 1225 1226 /* 1227 * If we have an iod which can process the request, then queue 1228 * the buffer. 1229 */ 1230 if (gotiod) { 1231 /* 1232 * Ensure that the queue never grows too large. We still want 1233 * to asynchronize so we block rather then return EIO. 1234 */ 1235 while (nmp->nm_bufqlen >= 2*nfs_numasync) { 1236 NFS_DPF(ASYNCIO, 1237 ("nfs_asyncio: waiting for mount %p queue to drain\n", nmp)); 1238 nmp->nm_bufqwant = TRUE; 1239 error = tsleep(&nmp->nm_bufq, slpflag | PRIBIO, 1240 "nfsaio", slptimeo); 1241 if (error) { 1242 if (nfs_sigintr(nmp, NULL, td)) 1243 return (EINTR); 1244 if (slpflag == PCATCH) { 1245 slpflag = 0; 1246 slptimeo = 2 * hz; 1247 } 1248 } 1249 /* 1250 * We might have lost our iod while sleeping, 1251 * so check and loop if nescessary. 1252 */ 1253 if (nmp->nm_bufqiods == 0) { 1254 NFS_DPF(ASYNCIO, 1255 ("nfs_asyncio: no iods after mount %p queue was drained, looping\n", nmp)); 1256 goto again; 1257 } 1258 } 1259 1260 if (bp->b_iocmd == BIO_READ) { 1261 if (bp->b_rcred == NOCRED && cred != NOCRED) 1262 bp->b_rcred = crhold(cred); 1263 } else { 1264 bp->b_flags |= B_WRITEINPROG; 1265 if (bp->b_wcred == NOCRED && cred != NOCRED) 1266 bp->b_wcred = crhold(cred); 1267 } 1268 1269 BUF_KERNPROC(bp); 1270 TAILQ_INSERT_TAIL(&nmp->nm_bufq, bp, b_freelist); 1271 nmp->nm_bufqlen++; 1272 return (0); 1273 } 1274 1275 /* 1276 * All the iods are busy on other mounts, so return EIO to 1277 * force the caller to process the i/o synchronously. 1278 */ 1279 NFS_DPF(ASYNCIO, ("nfs_asyncio: no iods available, i/o is synchronous\n")); 1280 return (EIO); 1281} 1282 1283/* 1284 * Do an I/O operation to/from a cache block. This may be called 1285 * synchronously or from an nfsiod. 1286 */ 1287int 1288nfs_doio(struct buf *bp, struct ucred *cr, struct thread *td) 1289{ 1290 struct uio *uiop; 1291 struct vnode *vp; 1292 struct nfsnode *np; 1293 struct nfsmount *nmp; 1294 int error = 0, iomode, must_commit = 0; 1295 struct uio uio; 1296 struct iovec io; 1297 struct proc *p = td ? td->td_proc : NULL; 1298 1299 vp = bp->b_vp; 1300 np = VTONFS(vp); 1301 nmp = VFSTONFS(vp->v_mount); 1302 uiop = &uio; 1303 uiop->uio_iov = &io; 1304 uiop->uio_iovcnt = 1; 1305 uiop->uio_segflg = UIO_SYSSPACE; 1306 uiop->uio_td = td; 1307 1308 /* 1309 * clear BIO_ERROR and B_INVAL state prior to initiating the I/O. We 1310 * do this here so we do not have to do it in all the code that 1311 * calls us. 1312 */ 1313 bp->b_flags &= ~B_INVAL; 1314 bp->b_ioflags &= ~BIO_ERROR; 1315 1316 KASSERT(!(bp->b_flags & B_DONE), ("nfs_doio: bp %p already marked done", bp)); 1317 1318 if (bp->b_iocmd == BIO_READ) { 1319 io.iov_len = uiop->uio_resid = bp->b_bcount; 1320 io.iov_base = bp->b_data; 1321 uiop->uio_rw = UIO_READ; 1322 1323 switch (vp->v_type) { 1324 case VREG: 1325 uiop->uio_offset = ((off_t)bp->b_blkno) * DEV_BSIZE; 1326 nfsstats.read_bios++; 1327 error = (nmp->nm_rpcops->nr_readrpc)(vp, uiop, cr); 1328 1329 if (!error) { 1330 if (uiop->uio_resid) { 1331 /* 1332 * If we had a short read with no error, we must have 1333 * hit a file hole. We should zero-fill the remainder. 1334 * This can also occur if the server hits the file EOF. 1335 * 1336 * Holes used to be able to occur due to pending 1337 * writes, but that is not possible any longer. 1338 */ 1339 int nread = bp->b_bcount - uiop->uio_resid; 1340 int left = uiop->uio_resid; 1341 1342 if (left > 0) 1343 bzero((char *)bp->b_data + nread, left); 1344 uiop->uio_resid = 0; 1345 } 1346 } 1347 /* ASSERT_VOP_LOCKED(vp, "nfs_doio"); */ 1348 if (p && (vp->v_vflag & VV_TEXT) && 1349 (np->n_mtime != np->n_vattr.va_mtime.tv_sec)) { 1350 uprintf("Process killed due to text file modification\n"); 1351 PROC_LOCK(p); 1352 psignal(p, SIGKILL); 1353 _PHOLD(p); 1354 PROC_UNLOCK(p); 1355 } 1356 break; 1357 case VLNK: 1358 uiop->uio_offset = (off_t)0; 1359 nfsstats.readlink_bios++; 1360 error = (nmp->nm_rpcops->nr_readlinkrpc)(vp, uiop, cr); 1361 break; 1362 case VDIR: 1363 nfsstats.readdir_bios++; 1364 uiop->uio_offset = ((u_quad_t)bp->b_lblkno) * NFS_DIRBLKSIZ; 1365 if ((nmp->nm_flag & NFSMNT_NFSV4) != 0) 1366 error = nfs4_readdirrpc(vp, uiop, cr); 1367 else { 1368 if ((nmp->nm_flag & NFSMNT_RDIRPLUS) != 0) { 1369 error = nfs_readdirplusrpc(vp, uiop, cr); 1370 if (error == NFSERR_NOTSUPP) 1371 nmp->nm_flag &= ~NFSMNT_RDIRPLUS; 1372 } 1373 if ((nmp->nm_flag & NFSMNT_RDIRPLUS) == 0) 1374 error = nfs_readdirrpc(vp, uiop, cr); 1375 } 1376 /* 1377 * end-of-directory sets B_INVAL but does not generate an 1378 * error. 1379 */ 1380 if (error == 0 && uiop->uio_resid == bp->b_bcount) 1381 bp->b_flags |= B_INVAL; 1382 break; 1383 default: 1384 printf("nfs_doio: type %x unexpected\n", vp->v_type); 1385 break; 1386 }; 1387 if (error) { 1388 bp->b_ioflags |= BIO_ERROR; 1389 bp->b_error = error; 1390 } 1391 } else { 1392 /* 1393 * If we only need to commit, try to commit 1394 */ 1395 if (bp->b_flags & B_NEEDCOMMIT) { 1396 int retv; 1397 off_t off; 1398 1399 off = ((u_quad_t)bp->b_blkno) * DEV_BSIZE + bp->b_dirtyoff; 1400 bp->b_flags |= B_WRITEINPROG; 1401 retv = (nmp->nm_rpcops->nr_commit)( 1402 bp->b_vp, off, bp->b_dirtyend-bp->b_dirtyoff, 1403 bp->b_wcred, td); 1404 bp->b_flags &= ~B_WRITEINPROG; 1405 if (retv == 0) { 1406 bp->b_dirtyoff = bp->b_dirtyend = 0; 1407 bp->b_flags &= ~(B_NEEDCOMMIT | B_CLUSTEROK); 1408 bp->b_resid = 0; 1409 bufdone(bp); 1410 return (0); 1411 } 1412 if (retv == NFSERR_STALEWRITEVERF) { 1413 nfs_clearcommit(bp->b_vp->v_mount); 1414 } 1415 } 1416 1417 /* 1418 * Setup for actual write 1419 */ 1420 1421 if ((off_t)bp->b_blkno * DEV_BSIZE + bp->b_dirtyend > np->n_size) 1422 bp->b_dirtyend = np->n_size - (off_t)bp->b_blkno * DEV_BSIZE; 1423 1424 if (bp->b_dirtyend > bp->b_dirtyoff) { 1425 io.iov_len = uiop->uio_resid = bp->b_dirtyend 1426 - bp->b_dirtyoff; 1427 uiop->uio_offset = (off_t)bp->b_blkno * DEV_BSIZE 1428 + bp->b_dirtyoff; 1429 io.iov_base = (char *)bp->b_data + bp->b_dirtyoff; 1430 uiop->uio_rw = UIO_WRITE; 1431 nfsstats.write_bios++; 1432 1433 if ((bp->b_flags & (B_ASYNC | B_NEEDCOMMIT | B_NOCACHE | B_CLUSTER)) == B_ASYNC) 1434 iomode = NFSV3WRITE_UNSTABLE; 1435 else 1436 iomode = NFSV3WRITE_FILESYNC; 1437 1438 bp->b_flags |= B_WRITEINPROG; 1439 error = (nmp->nm_rpcops->nr_writerpc)(vp, uiop, cr, &iomode, &must_commit); 1440 1441 /* 1442 * When setting B_NEEDCOMMIT also set B_CLUSTEROK to try 1443 * to cluster the buffers needing commit. This will allow 1444 * the system to submit a single commit rpc for the whole 1445 * cluster. We can do this even if the buffer is not 100% 1446 * dirty (relative to the NFS blocksize), so we optimize the 1447 * append-to-file-case. 1448 * 1449 * (when clearing B_NEEDCOMMIT, B_CLUSTEROK must also be 1450 * cleared because write clustering only works for commit 1451 * rpc's, not for the data portion of the write). 1452 */ 1453 1454 if (!error && iomode == NFSV3WRITE_UNSTABLE) { 1455 bp->b_flags |= B_NEEDCOMMIT; 1456 if (bp->b_dirtyoff == 0 1457 && bp->b_dirtyend == bp->b_bcount) 1458 bp->b_flags |= B_CLUSTEROK; 1459 } else { 1460 bp->b_flags &= ~(B_NEEDCOMMIT | B_CLUSTEROK); 1461 } 1462 bp->b_flags &= ~B_WRITEINPROG; 1463 1464 /* 1465 * For an interrupted write, the buffer is still valid 1466 * and the write hasn't been pushed to the server yet, 1467 * so we can't set BIO_ERROR and report the interruption 1468 * by setting B_EINTR. For the B_ASYNC case, B_EINTR 1469 * is not relevant, so the rpc attempt is essentially 1470 * a noop. For the case of a V3 write rpc not being 1471 * committed to stable storage, the block is still 1472 * dirty and requires either a commit rpc or another 1473 * write rpc with iomode == NFSV3WRITE_FILESYNC before 1474 * the block is reused. This is indicated by setting 1475 * the B_DELWRI and B_NEEDCOMMIT flags. 1476 * 1477 * If the buffer is marked B_PAGING, it does not reside on 1478 * the vp's paging queues so we cannot call bdirty(). The 1479 * bp in this case is not an NFS cache block so we should 1480 * be safe. XXX 1481 */ 1482 if (error == EINTR 1483 || (!error && (bp->b_flags & B_NEEDCOMMIT))) { 1484 int s; 1485 1486 s = splbio(); 1487 bp->b_flags &= ~(B_INVAL|B_NOCACHE); 1488 if ((bp->b_flags & B_PAGING) == 0) { 1489 bdirty(bp); 1490 bp->b_flags &= ~B_DONE; 1491 } 1492 if (error && (bp->b_flags & B_ASYNC) == 0) 1493 bp->b_flags |= B_EINTR; 1494 splx(s); 1495 } else { 1496 if (error) { 1497 bp->b_ioflags |= BIO_ERROR; 1498 bp->b_error = np->n_error = error; 1499 np->n_flag |= NWRITEERR; 1500 } 1501 bp->b_dirtyoff = bp->b_dirtyend = 0; 1502 } 1503 } else { 1504 bp->b_resid = 0; 1505 bufdone(bp); 1506 return (0); 1507 } 1508 } 1509 bp->b_resid = uiop->uio_resid; 1510 if (must_commit) 1511 nfs_clearcommit(vp->v_mount); 1512 bufdone(bp); 1513 return (error); 1514} 1515 1516/* 1517 * Used to aid in handling ftruncate() operations on the NFS client side. 1518 * Truncation creates a number of special problems for NFS. We have to 1519 * throw away VM pages and buffer cache buffers that are beyond EOF, and 1520 * we have to properly handle VM pages or (potentially dirty) buffers 1521 * that straddle the truncation point. 1522 */ 1523 1524int 1525nfs_meta_setsize(struct vnode *vp, struct ucred *cred, struct thread *td, u_quad_t nsize) 1526{ 1527 struct nfsnode *np = VTONFS(vp); 1528 u_quad_t tsize = np->n_size; 1529 int biosize = vp->v_mount->mnt_stat.f_iosize; 1530 int error = 0; 1531 1532 np->n_size = nsize; 1533 1534 if (np->n_size < tsize) { 1535 struct buf *bp; 1536 daddr_t lbn; 1537 int bufsize; 1538 1539 /* 1540 * vtruncbuf() doesn't get the buffer overlapping the 1541 * truncation point. We may have a B_DELWRI and/or B_CACHE 1542 * buffer that now needs to be truncated. 1543 */ 1544 error = vtruncbuf(vp, cred, td, nsize, biosize); 1545 lbn = nsize / biosize; 1546 bufsize = nsize & (biosize - 1); 1547 bp = nfs_getcacheblk(vp, lbn, bufsize, td); 1548 if (bp->b_dirtyoff > bp->b_bcount) 1549 bp->b_dirtyoff = bp->b_bcount; 1550 if (bp->b_dirtyend > bp->b_bcount) 1551 bp->b_dirtyend = bp->b_bcount; 1552 bp->b_flags |= B_RELBUF; /* don't leave garbage around */ 1553 brelse(bp); 1554 } else { 1555 vnode_pager_setsize(vp, nsize); 1556 } 1557 return(error); 1558} 1559 1560